Final Program and Abstracts - SIAM: Society for … Address DoubleTree by Hilton Hotel Philadelphia...

This conference is sponsored by the SIAM Activity Group on Imaging Science

The SIAM Activity Group on Imaging Science brings together SIAM members and other scientists and engineers with an interest in the mathematical and

computational aspects of imaging.

The reconstruction, enhancement, segmentation, analysis, registration, compression, representation, and tracking of two and three dimensional images are vital to many areas of science, medicine, and

engineering. As a result, increasingly sophisticated mathematical, statistical, and computational methods are being employed in these research areas, which may be referred to as “imaging science.” These techniques include transform and orthogonal series methods, nonlinear optimization, numerical linear algebra, integral equations, partial differential equations, Bayesian and other statistical inverse

estimation methods, operator theory, differential geometry, information theory, interpolation and approximation, inverse problems, computer graphics and vision, stochastic processes, and others.

The activity group organizes the biennial SIAM Conference on Imaging Science, awards the SIAG on Imaging Science Prize every two years to the authors of the best paper on mathematical and

computational aspects of imaging, and maintains a website, a member directory, and an electronic mailing list.

Final Program and Abstracts

Society for Industrial and Applied Mathematics

3600 Market Street, 6th Floor

Philadelphia, PA 19104-2688 USA

Telephone: +1-215-382-9800 Fax: +1-215-386-7999

Conference E-mail: [email protected]

Conference Web: www.siam.org/meetings/

Membership and Customer Service: (800) 447-7426 (US & Canada) or +1-215-382-9800 (worldwide)

www.siam.org/meetings/is12

2 2012SIAMConferenceonImagingScience

Table of Contents

Program-at-a-Glance ......Fold out section

General Information ...............................2

Get-togethers ..........................................4

Invited Plenary Presentations ...............7

Minitutorials ...........................................5

Poster Session ......................................23

Prize Lecture ..........................................8

Program Schedule ..................................9

Abstracts ..............................................49

Speaker and Organizer Index ............117

Conference Budget .....Inside Back cover

Hotel Meeting Room Map .....Back cover

Organizing Committee Co-ChairsPeter Kuchment Texas A&M University, USA

Jennifer MuellerColorado State University, USA

Organizing CommitteeSimon Arridge University College London, United Kingdom

Peter Burgholzer Research Center for Non Destructive Testing

GmbH, Linz, Austria

Anthony DavisJet Propulsion Laboratory, USA

Maarten de HoopPurdue University, USA

Misha KilmerTufts University, USA

Robert KrugerOptoSonics, Inc., USA

Jin Keun SeoYonsei University, Korea

Samuli SiltanenUniversity of Helsinki, Finland

SIAM Registration Desk The SIAM registration desk is located in Overture - 3rd Floor, and is open during the following times:

Saturday, May 19

4:00 PM - 7:00 PM

Sunday, May 20

7:00 AM - 5:00 PM

Monday, May 21

7:45 AM - 5:00 PM

Tuesday, May 22

7:45 AM - 5:00 PM

Hotel Address DoubleTree by Hilton Hotel Philadelphia Center City

237 South Broad Street

Philadelphia, Pennsylvania 19107-5686

Hotel Telephone NumberTo reach an attendee or leave a message, call +1-215-893-1600. The hotel operator can either connect you with the SIAM registration desk or to the attendee’s room. Messages taken at the SIAM registration desk will be posted to the message board located in the registration area.

Hotel Check-in and Check-out TimesCheck-in time is 4:00 PM and check-out time is 12:00 PM.

Child CareFor local child care information, please contact the concierge at the Doubletree (+1-215-893-1600).

Corporate Members and AffiliatesSIAM corporate members provide their employees with knowledge about, access to, and contacts in the applied mathematics and computational sciences community through their membership benefits. Corporate membership is more than just a bundle of tangible products and services; it is an expression of support for SIAM and its programs. SIAM is pleased to acknowledge its corporate members and sponsors. In recognition of their support, non-member attendees who are employed by the following organizations are entitled to the SIAM member registration rate.

Corporate Institutional MembersThe Aerospace Corporation

Air Force Office of Scientific Research

AT&T Laboratories - Research

Bechtel Marine Propulsion Laboratory

The Boeing Company

CEA/DAM

DSTO- Defence Science and Technology Organisation

Hewlett-Packard

IBM Corporation

IDA Center for Communications Research, La Jolla

IDA Center for Communications Research, Princeton

Institute for Defense Analyses, Center for Computing Sciences

Lawrence Berkeley National Laboratory

Lockheed Martin

Mathematical Sciences Research Institute

Max-Planck-Institute for Dynamics of Complex Technical Systems

Mentor Graphics

The MITRE Corporation

National Institute of Standards and Technology (NIST)

National Security Agency (DIRNSA)

2012SIAMConferenceonImagingScience 3

Naval Surface Warfare Center, Dahlgren Division

NEC Laboratories America, Inc.

Oak Ridge National Laboratory, managed by UT-Battelle for the Department of Energy

PETROLEO BRASILEIRO S.A. – PETROBRAS

Philips Research

Sandia National Laboratories

Schlumberger-Doll Research

Tech X Corporation

Texas Instruments Incorporated

U.S. Army Corps of Engineers, Engineer Research and Development Center

United States Department of Energy

*List current March 2012

Funding AgencySIAM and the Conference Organizing Committee wish to extend their thanks and appreciation to the U.S. National Science Foundation for its support of this conference.

Leading the applied mathematics community . . .

Join SIAM and save!SIAM members save up to $130 on full registration for the SIAM Conference on Imaging Science (IS12)! Join your peers in supporting the premier professional society for applied mathematicians and computational scientists. SIAM members receive subscriptions to SIAM Review and SIAM News and enjoy substantial discounts on SIAM books, journal subscriptions, and conference registrations.

If you are not a SIAM member and paid the Non-Member or Non-Member

Mini Speaker/Organizer rate to attend the conference, you can apply the difference between what you paid and what a member would have paid ($130 for a Non-Member and $65 for a Non-Member Mini Speaker/Organizer) towards a SIAM membership. Contact SIAM Customer Service for details or join at the conference registration desk.

If you are a SIAM member, it only costs $10 to join the SIAM Activity Group on Imaging Science (SIAG/IS). As a SIAG/IS member, you are eligible for an additional $10 discount on this conference, so if you paid the SIAM member rate to attend the conference, you might be eligible for a free SIAG/IS membership. Check at the registration desk.

Free Student Memberships are available to students who attend an institution that is an Academic Member of SIAM, are members of Student Chapters of SIAM, or are nominated by a Regular Member of SIAM.

Join onsite at the registration desk, go to www.siam.org/joinsiam to join online or download an application form, or contact SIAM Customer Service

Telephone: +1-215-382-9800 (worldwide); or

800-447-7426 (U.S. and Canada only)

Fax: +1-215-386-7999

E-mail: [email protected]

Postal mail:

Society for Industrial and Applied Mathematics

3600 Market Street, 6th Floor

Philadelphia, PA 19104-2688 USA

Standard Audio/Visual Set-Up in Meeting Rooms SIAM does not provide computers for any speaker. When giving an electronic presentation, speakers must provide their

own computers. SIAM is not responsible for the safety and security of speakers’ computers.

The Plenary Session Room will have two (2) overhead projectors, two (2) screens and one (1) data projector. Cables or adaptors for Apple computers are not supplied, as they vary for each model. Please bring your own cable/adaptor if using a Mac computer.

All other concurrent/breakout rooms will have one (1) screen and one (1) data projector. Cables or adaptors for Apple computers are not supplied, as they vary for each model. Please bring your own cable/adaptor if using a Mac computer. Overhead projectors will be provided only when requested.

If you have questions regarding availability of equipment in the meeting room of your presentation, or to request an overhead projector for your session, please see a SIAM staff member at the registration desk.

E-mail AccessSIAM attendees booked within the SIAM room block will have complimentary wired Internet access in their guest rooms. Upon accessing Internet in the guest rooms, individuals must accept charges to be billed to the hotel room, however, the fee will be reversed upon check-out. Internet charges will not appear on the final room bill. If sharing a guest room, please be aware the hotel is only able to waive the fee for one laptop.

Complimentary wireless Internet access in the hotel lobby and restaurant will also be available. Additionally, wireless Internet access in the meeting rooms will be provided to SIAM attendees, the cost for which is incorporated into the conference registration fee.

SIAM will also provide a limited number of email stations for attendees during registration hours.


Registration Fee Includes• Admission to all technical sessions

• Business Meeting (open to SIAG/IS members)

• Coffee breaks daily

• Room set-ups and audio/visual equipment

• Welcome Reception and Poster Session

Job PostingsPlease check with the SIAM registration desk regarding the availability of job postings or visit http://jobs.siam.org.

Important Notice to Poster PresentersThe poster session is scheduled for Sunday, May 20 at 8:00 PM. Poster presenters are requested to set up their poster material on the provided 4’ x 6’ poster boards in the Ormandy West Room between the hours of 3:00 PM and 8:00 PM. All materials must be posted by Sunday, May 20 at 8:00 PM, the official start time of the session. Posters will remain on display until 10:00 PM. Poster displays must be removed by 10:00 PM. Posters remaining after this time will be discarded. SIAM is not responsible for discarded posters.

SIAM Books and JournalsDisplay copies of books and complimentary copies of journals are available on site. SIAM books are available at a discounted price during the conference. If a SIAM books representative is not available, completed order forms and payment (credit cards are preferred) may be taken to the SIAM registration desk. The books table will close at 1:00 PM on Tuesday, May 22.

Table Top DisplayIOP Publishing

Maney Publishing

SIAM

Springer

Name BadgesA space for emergency contact information is provided on the back of your name badge. Help us help you in the event of an emergency!

Comments?Comments about SIAM meetings are encouraged! Please send to:

Sven Leyffer, SIAM Vice President for Programs ([email protected])

Get-togethers • Welcome Reception and Poster

Session

Sunday, May 20 8:00 PM – 10:00 PM

• Business Meeting (open to SIAG/IS members)

Monday, May 21 8:00 PM – 8:45 PM Complimentary beer and wine will be served.

Please NoteSIAM is not responsible for the safety and security of attendees’ computers. Do not leave your laptop computers unattended. Please remember to turn off your cell phones, pagers, etc. during sessions.

Recording of PresentationsAudio and video recording of presentations at SIAM meetings is prohibited without the written permission of the presenter and SIAM.

Social MediaSIAM is promoting the use of social media, such as Facebook and Twitter, in order to enhance scientific discussion at its meetings and enable attendees to connect with each other prior to, during and after conferences. If you are tweeting about a conference, please use the designated hashtag to enable other attendees to keep up with the Twitter conversation and to allow better archiving of our conference discussions. The hashtag for this meeting is #SIAMIS12.

SAVE THE DATE!The 2014 SIAM Conference on

Imaging Science will be held

May 12-14, 2014, at Hong Kong

Baptist University, Hong Kong.


Minitutorials** All Minitutorials will take place in the Symphony Ballroom - 3rd Floor**

Sunday,May20

MT1 Harry Potter’s Cloak via Transformation Optics9:30 AM - 11:30 AMWe will explain, in a non-technical fashion, the basic ideas of what has been called transformation optics which has been used to make objects invisible to detection by electromagnetic waves, acoustic waves and other types of waves. Transformation optics uses the invariance properties under transformations of Maxwell’s equations in the case of electromagnetic waves. The medium parameters needed for perfect cloaking are singular at the interface of the region to the cloaked. We will also consider regularization procedures of perfect cloaking that lead to non-singular medium parameters and achieve approximate cloaking.

Organizer:

Gunther Uhlmann, University of California, Irvine, USA and University of Washington, USA

Speakers:


Ting Zhou, Massachusetts Institute of Technology, USA

Monday,May21

MT2 Mathematics and Science of 3D Electron Microscope Imaging9:30 AM - 11:30 AMThe two parts of this minitutorial deal with mathematics for Electron Tomography (ET) and Single Particle (SP). The one on ET surveys reconstruction methods for ET with emphasis on difficulties and their mathematical consequences. It also includes the model for image formation in electron microscopy, which is the same in SP. The minitutorial on SP focuses on presenting a novel reconstruction method, focusing on resolving the angular reconstitution problem. The method in question is based on a novel framework for non-linear optimization, which in turn is based on representation theoretic ideas - where the solution is an object of a category.

Organizers:

Ozan Öktem, KTH - Royal Institute of Technology, Sweden

Eric Todd Quinto, Tufts University, USA

Speakers:

Ronny Hadani, University of Texas at Austin, USA



SIAM Activity Group on Imaging Science (SIAG/IS) www.siam.org/activity/imaging

A GREAT WAY TO GET invOlvEd!

Collaborate and interact with mathematicians and applied scientists whose work involves imaging science.

ACTIVITES INCLUDE:• Special sessions at SIAM Annual Meetings• Biennial conference• SIAG on Imaging Science Prize• Website

BENEFITS OF SIAG/IS mEmBErShIp:• Listing in the SIAG’s online-only membership directory • Additional $10 discount on registration at the SIAM Conference on Imaging Science (excludes student)

• Electronic communications about recent developments in your specialty• Eligibility for candidacy for SIAG/IS office• Participation in the selection of SIAG/IS officers

ELIGIBILITY:• Be a current SIAM member

COST:• $10 per year• Student SIAM members can join 2 activity groups for free!

2012-2013 SIAG/IS OFFICErS:• Chair: Christine De Mol, Universite Libre du Bruxelles • Vice Chair: Naoki Saito, University of California• Program Director: Fadil Santosa, University of Minnesota, Twin Cities• Secretary: Justin Romberg, Georgia Institute of Technology

TO JOIN:

SIAG/IS: my.siam.org/forms/join_siag.htm

SIAm: www.siam.org/joinsiam

May 20 - 22, 2012DoubleTree by Hilton Hotel Philadelphia, Philadelphia, Pennsylvania, USA


Invited Plenary Speakers** All Invited Plenary Presentations will take place in the Symphony Ballroom - 3rd Floor **

Sunday, May 20

8:15 AM - 9:00 AM

IP1MagneticResonanceElastography

Richard L. Ehman,Mayo Clinic, USA

Monday, May 21

8:15 AM - 9:00 AM

IP2 TargetedSeismicImaging

Alison Malcolm,Massachusetts Institute of Technology, USA

1:00 PM - 1:45 PM

IP3ThreeDimensionalStructureDeterminationofMacromoleculesbyCryo-electronMicroscopyfromtheAppliedMathPerspective

Amit Singer,Princeton University, USA

Tuesday, May 22

8:15 AM - 9:00 AM

IP4DivideandConquer:UsingSpectralMethodsonPartitionedImages

Dianne P. O’Leary, University of Maryland, College Park, USA

1:00 PM - 1:45 PM

IP5DesigninImaging–FromCompressivetoComprehensiveSensing

Lior Horesh, IBM T.J. Watson Research Center, USA


Prize Lecture** The Prize Lecture will take place in the Symphony Ballroom - 3rd Floor **

Sunday, May 20

1:00 PM - 1:45 PM

SIAGonImagingSciencesPrizeLecture

Recipient Information and Title TBD


IS12 Program


To order, shop online at www.siam.org/catalog. Use your credit card (AMEX, MasterCard, and VISA) by phone: +1-215-382-9800 worldwide or toll free at 800-447-SIAM in USA and Canada or fax: +1-215-386-7999. Or send check or money order in US dollars to: SIAM, Dept. BKIS12, 3600 Market Street, 6th Floor, Philadelphia, PA 19104-2688 USA.

FAIR: Flexible Algorithms for Image RegistrationJan ModersitzkiThis overview of state-of-the-art registration techniques from theory to practice includes numerous exercises designed to enhance readers’ understanding of the principles and mechanisms of the described techniques. It also provides, via a supplementary Web page, free access to FAIR.m, a package that is based on the MATLAB® software environment, which enables readers to experiment with the proposed algorithms and explore the presented examples in more depth.2009 · xxii + 189 pages · Softcover · ISBN 978-0-898716-90-0 List $69.00 · Attendee $55.20 · SIAM Member $48.30 · Order Code FA06

Fundamentals of Radar ImagingMargaret Cheney and Brett Borden This book includes a description of how a radar system works, together with the relevant mathematics; theory that guides the choice of radar waveforms; derivation of the fundamentals of scattering theory; derivation and discussion of the image formation process; and a long list of current open problems. Applied mathematicians will want this book because it explains the basics of radar imaging, provides a foundation for understanding the engineering literature, and gives references for many of the open problems.2009 · xxiv + 140 pages · Softcover · ISBN 978-0-898716-77-1 List $59.00 · Attendee $47.20 · SIAM/CBMS $41.30 · Order Code CB79

Deblurring Images: Matrices, Spectra, and FilteringPer Christian Hansen, James G. Nagy, and Dianne P. O’Leary“At a very affordable and introductory level, the book has successfully integrated an emerging important application, mathematical analysis, efficient algorithms, and practical software implementation into a coherent and very appealing subject. …[there are] color plates for most figures, which further enhances the physical and informative appeal of the book.” — Jackie Shen, Professor of Mathematics, University of Minnesota

2006 · xiv+130 pages · Softcover · ISBN 978-0-898716-18-4 List $68.00 · Attendee $54.40 · SIAM Member $47.60 · Order Code FA03

Society for induStrial and applied MatheMaticS

Image Processing and Analysis: Variational, PDE, Wavelet, and Stochastic MethodsTony F. Chan and Jianhong (Jackie) ShenThis book develops the mathematical foundation of modern image processing and low-level computer vision, and presents a general framework from the analysis of image structures and patterns to their processing. The core mathematical and computational ingredients of several important image processing tasks are investigated. The book bridges contemporary mathematics with state-of-the-art methodologies in modern image processing while organizing the vast contemporary literature into a coherent and logical structure.2005 · xxii + 400 pages · Softcover · ISBN 978-0-898715-89-7 List $81.00 · Attendee $64.80 · SIAM Member $56.70 · Order Code OT94

Introduction to the Mathematics of Medical Imaging, Second EditionCharles L. EpsteinThis textbook provides a firm foundation in the mathematical tools used to model the measurements and derive the reconstruction algorithms used in most imaging modalities in current use. In the process, it also covers many important analytic concepts and techniques used in Fourier analysis, integral equations, sampling theory, and noise analysis. The book uses X-ray computed tomography as a “pedagogical machine” to illustrate important ideas and incorporates extensive discussions of background material making the more advanced mathematical topics accessible to readers with a less formal mathematical education. The mathematical concepts are illuminated with over 200 illustrations and numerous exercises.2007 · xxxiv + 761 pages · Softcover · ISBN 978-0-898716-42-9 List $102.00 · Attendee $81.60 · SIAM Member $71.40 · Order Code OT102

Check out

www.siam.org/ebooks

Nonmembers must use code “BKIS12” to get

Special Conference Attendee Price. Expires 6-22-12.

SIAM BOOKSSIAM BOOKSCheck out these and other books at the SIAM booth.

Conference attendees receive discounts on all displayed titles.


Saturday, May 19

Registration4:00 PM-7:00 PMRoom:Overture - 3rd Floor

Sunday,May20Registration7:00 AM-5:00 PMRoom:Overture - 3rd Floor

Welcoming Remarks8:00 AM-8:15 AMRoom:Symphony Ballroom - 3rd Floor

IP1Magnetic Resonance Elastography8:15 AM-9:00 AMRoom:Symphony Ballroom - 3rd Floor

Chair: Peter Kuchment, Texas A&M University, USA

Many disease processes cause profound changes in the mechanical properties of tissues, yet none of the conventional medical imaging techniques such as CT, MRI, and ultrasound are capable of quantitatively delineating these properties. Magnetic Resonance Elastography (MRE) is an emerging diagnostic imaging technology that employs a novel MRI-based technique to visualize propagating acoustic shear waves in the body. Inversion algorithms are used to process these wave images to generate quantitative maps of tissue mechanical properties such as stiffness, viscosity, and anisotropy. The most important clinical application of MRE currently to evaluate chronic liver disease, where it can eliminate the need for an invasive biopsy. However, there are many other potential applications. This presentation will review the rationale and physical basis of MRE and the basic approaches used for data analysis. Areas of opportunity will be identified where more advanced inversion algorithms could create new applications or substantially contribute to the clinical value of the technology.

RichardL.EhmanMayo Clinic, USA

Sunday,May20Coffee Break9:00 AM-9:30 AMRoom:Orchestra - 2nd Floor

MT1Harry Potter’s Cloak via Transformation Optics9:30 AM-11:30 AMRoom:Symphony Ballroom - 3rd Floor

Chair: Gunther Uhlmann, University of California, Irvine, USA and University of Washington, USA

We will explain, in a non-technical fashion, the basic ideas of what has been called transformation optics which has been used to make objects invisible to detection by electromagnetic waves, acoustic waves and other types of waves. Transformation optics uses the invariance properties under transformations of Maxwell’s equations in the case of electromagnetic waves. The medium parameters needed for perfect cloaking are singular at the interface of the region to the cloaked. We will also consider regularization procedures of perfect cloaking that lead to non-singular medium parameters and achieve approximate cloaking.

Speakers:


Ting Zhou, Massachusetts Institute of Technology, USA


Sunday,May20

MS2Advances in Sparse Recovery Algorithms - Part I of II9:30 AM-11:30 AMRoom:Assembly E - 5th Floor

For Part 2 see MS12 Nonsmooth optimization plays a central role in sparse recovery. The recent theory of compressed sensing has revitalized the interest in such optimization problems. Sparse recovery algorithms have an extensive literature, all of which can in essence be categorized into three main classes: convex (geometric), greedy (combinatorial), and probabilistic (variational/Belief Propagation) algorithms. The focus of this minisymposium is to give an overview of recent algorithmic advances in sparse recovery. We compare and contrast the advantages and limitations of these seemingly different approaches, and identify potential new research avenues for their interactions. Session will gather talks by leading experts in the field.

Organizer:VolkanCevherÉcole Polytechnique Fédérale de Lausanne, Switzerland

Organizer:JalalFadiliUniversité de Caen, France

9:30-9:55 Graphical Models and Message Passing Algorithms to Solve Large Scale Regularized Regression ProblemsAndrea Montanari, Stanford University, USA

10:00-10:25 Recovering Cosparse Vectors using Convex vs Greedy AlgorithmsRemi Gribonval and Sangnam Nam, IRISA-

INRIA, France; Mike Davies, University of Edinburgh, United Kingdom; Michael Elad, Technion, Israel

10:30-10:55 Simple and Practical Algorithm for Sparse Fourier TransformHaitham Hassanieh, Massachusetts Institute

of Technology, USA; Piotr Indyk, Massachusetts Institute of Technology, USA; Dina Katabi, Massachusetts Institute of Technology, USA

11:00-11:25 Hard Thresholding with Norm ConstraintsVolkan Cevher, École Polytechnique Fédérale

de Lausanne, Switzerland

Sunday,May20

MS1Analysis of Tomographic Images for Clinics: Recent Advances and Challenges - Part I of II9:30 AM-11:30 AMRoom:Assembly C - 5th Floor

For Part 2 see MS11 Analysis of tomographic images is of central importance for diagnostic and therapeutic applications in various clinical contexts. Distinct characteristics between tissues, along with different image properties and qualities among imaging modalities, have posed great challenges to the analysis tasks. Therefore, new ideas, adaptive models, and special techniques are in high demand. This minisymposium aims to bring together mathematicians, medical physicists, radiologists, and engineers, especially young scientists, to exchange the most recent advances in tomographic image analysis. We hope this minisymposium can provide a forum to stimulate discussions and establish collaborations between mathematicians and clinic scientists for further developments in this emerging research field.

Organizer:XiaojingYeGeorgia Institute of Technology, USA

Organizer:XunJiaUniversity of California, San Diego, USA

9:30-9:55 Applications of Tomographical Images in Cancer Radiotherapy and Cone Beam CT Image Enhancement Using Prior Nonlocal Means MethodXun Jia, University of California, San Diego,

USA

10:00-10:25 The Block DROP and CAV Algorithms for Compressed Sensing Based TomographyJiehua Zhu and Xiezhang Li, Georgia

Southern University, USA

10:30-10:55 Trabecular Texture Analysis in Dental Cone Beam CTHaibin Ling, Xiong Yang, Jie Yang, and

Fangfang Xie, Temple University, USA; Yong Xu, South China University of Technology, China; Vasileios Megalooikonomou, Temple University, USA

11:00-11:25 Joint CT/CBCT Deformable Registration and Cone Beam CT Image EnhancementYifei Lou, Georgia Institute of Technology,

USA; Xun Jia, University of California, San Diego, USA; Tianye Niu, Patricio Vela, and Lei Zhu, Georgia Institute of Technology, USA; Steve Jiang, University of California, San Diego, USA; Allen Tannenbaum, Boston University, USA

continued in next column


Sunday,May20

MS3Novel Tomographic Imaging Techniques in Medicine9:30 AM-11:30 AMRoom:Concerto A - 3rd Floor

Tomographic image restoration is a widely applied modality in medical imaging. Research in this area requires a multidisciplinary approach incorporating mathematical and numerical analysis, imaging processing methods, inverse problems, image reconstruction algorithms, and experimental techniques. A deep understanding of underlying physical phenomena and implementation details of reconstruction algorithms are prerequisites for recovering tomographic images with practical significance and value. In this minisymposium novel techniques in tomographic imaging are presented. These range from methods-based aspects such as sparse recovery techniques, topological sensitivities and advanced computational tools to modelling issues in medical imaging of electrical, optical, and mechanical properties of tissue.

Organizer:MichaelHintermuellerHumboldt University Berlin, Germany

9:30-9:55 Hybrid Techniques on Electrical Tissue Property ImagingJin Keun Seo, Yonsei University, South Korea

10:00-10:25 Conductivity Imaging from Minimal Current Density DataAlexandru Tamasan, University of Central

Florida, USA

10:30-10:55 Anisotropic Elastic Moduli Reconstruction of Transversely Isotropic Material in Magnetic Resonance ElastographyOhin Kwon, Konkuk University, Korea

11:00-11:25 A Shape and Topology Optimization Method for the Resolution of Inverse Problems in TomographyAntoine Laurain, Humboldt University Berlin,

Germany

Sunday,May20

MS4Advances in 2D and 3D Hyperspectral Imaging - Part I of II9:30 AM-10:30 AMRoom:Concerto B - 3rd Floor

For Part 2 see MS14 This symposium addresses methods for hyperspectral image data as smoothing, clustering, segmentation and registration. In this context we consider images with potentially arbitrarily many color channels. The challenges in this area are edge preserving denoising of this type of images and dealing with data which is not sampled on regular grids as is the case after registration. Especcially applications in mass spectrometrie imaging shall be focussed.

Organizer:StefanSchifflerUniversity of Bremen, Germany

Organizer:JanStrehlowFraunhofer MEVIS, Germany

9:30-9:55 TV-Minimization for Color Images on Non-regular GridsStefan Schiffler, University of Bremen,

Germany

10:00-10:25 Efficient Algorithms for Photoacoustic ImagingStefan Kunis, University of Osnabrueck,

Germany

Sunday,May20

MS5Current Developments and Challenges in Imaging Through Turbulence - Part I of II9:30 AM-11:30 AMRoom:Aria A - 3rd Floor

For Part 2 see MS15 The video sequence captured in a long-range system, such as surveillance, is often corrupted by the atmospheric turbulence degradation. Each frame suffers from geometric distortion and focal blur. To process such data arises a large amount of challenges in image processing and computer vision, such as to real-time enhancement, object detection and recognition. This minisymposium is intended to present the state-of-the-art methods in reconstructing a video or a latent image with higher resolution, which include non-rigid registration, image fusion and blind deconvolution. The richness of its applications would draw attention to many researchers working on imaging science.

Organizer:YifeiLouGeorgia Institute of Technology, USA

Organizer:SungHaKangGeorgia Institute of Technology, USA

9:30-9:55 The Past and Future of Imaging Through TurbulenceArjuna Flenner, Naval Air Weapons Station,

USA

10:00-10:25 Removing Atmospheric TurbulenceXiang Zhu and Peyman Milanfar, University

of California, Santa Cruz, USA

10:30-10:55 Turbulence Restoration: From Stabilization to Atmospheric DeblurringJerome Gilles, and Stanley J. Osher,

University of California, Los Angeles, USA

11:00-11:25 A Dynamic Texture Model for Imaging Through TurbulenceMario Micheli, Université Paris Descartes,

France; Yifei Lou, Georgia Institute of Technology, USA; Stefano Soatto, and Andrea L. Bertozzi, University of California, Los Angeles, USA


Sunday,May20

MS6Advances in Nonsmooth and Nonconvex Minimization for Imaging - Part I of II9:30 AM-11:30 AMRoom:Maestro A - 4th Floor

For Part 2 see MS16 Optimization plays a central role in modern image processing. While convex optimization has received considerable interest in the field, nonsmooth nonconvex optimization remains much less developed. Yet, there are many problems in image processing that involve nonconvex and nonsmooth functionals to minimize, e.g. in inverse problems. The minisymposium will give an overview of recent advances on both theoretical and algorithmic aspects of minimizing nonsmooth nonconvex objective functionals. Topics that will be covered range from well-posedness, to characterization of the minimizers and algorithms to reach them. The minisymposium will bring together recognized experts in optimization theory and imaging sciences.

Organizer:MilaNikolovaENS Cachan, France


9:30-9:55 Nonconvex Nonsmooth Optimization via Gradient SamplingFrank E. Curtis, and Xiaocun Que, Lehigh

University, USA

10:00-10:25 On the Evaluation Complexity of Nonsmooth Composite Function Minimization with Applications to Nonconvex Nonlinear ProgrammingCoralia Cartis, University of Edinburgh,

United Kingdom; Nicholas I.M. Gould, Rutherford Appleton Laboratory, United Kingdom; Philippe L. Toint, University of Namur, Belgium

10:30-10:55 High-Dimensional Covariance Estimation under Sparse Kronecker Product StructureAlfred O. Hero, The University of Michigan,

Ann Arbor, USA; Theodoris Tsiligkardis, University of Michigan, USA

11:00-11:25 Wasserstein Barycenter: Global Minimizers and AlgorithmRabin Julien, University of Caen, France

Sunday,May20

MS8Generalized Radon Transform and Microlocal Analysis9:30 AM-11:30 AMRoom:Minuet - 4th Floor

Generalized Radon transforms arise in a wide range of imaging aplications. This includes synthetic aperture radar, sonar, seismic imaging, X-Ray CT and electron microscopy to mention a few. This minisymposium will bring together researchers working on variety of imaging applications where the underlying imaging models are generalized Radon transforms and present the most recent results on the role of microlocal analysis in image formation.

Organizer:BirsenYaziciRensselaer Polytechnic Institute, USA

9:30-9:55 Synthetic Aperture Imaging of Moving Targets using Ultra-narrowband waveformsBirsen Yazici and Ling Wang, Rensselaer

Polytechnic Institute, USA

10:00-10:25 Microlocal Analysis of an Ultrasound Transform with Circular Source and Receiver TrajectoriesGaik Ambartsoumian, University of Texas at

Arlington, USA; Venky P. Krishnan, Tata Institute of Fundamental Research, India; Todd Quinto, Tufts University, USA

10:30-10:55 Inversion of the V-Line Radon Transform in a Disc and Its Applications in ImagingGaik Ambartsoumian, University of Texas at

Arlington, USA

11:00-11:25 Microlocal Analysis of Common Midpoint SAR ImagingRaluca Felea, Rochester Institute of

Technology, USA

Sunday,May20

MS7Mathematical Modeling of Textures - Part I of II9:30 AM-11:30 AMRoom:Maestro B - 4th Floor

For Part 2 see MS17 Texture is a key component of natural images. Developing efficient models for textures is the key to go beyond the state of the art for several problems in image processing (compression, super-resolution), computer graphics (static and dynamic textures synthesis), computer vision (segmentation and retrieval) and psychophysics (visual perception). This requires to advance the research front in both statistical and deterministic models of natural images. This minisymposium features talks from world renown experts in the field of mathematical texture modeling.

Organizer:GabrielPeyréUniversité Paris Dauphine, France

Organizer:Jean-FrancoisAujolIMB, CNRS, Université Bordeaux 1, France

9:30-9:55 Denoising Textures with a Multiplicative Regularization MethodXavier Bresson, City University of Hong

Kong, Hong Kong

10:00-10:25 A Level-set Approach to Image-Texture Segmentation: Old and NewJing Yuan, University of Western Ontario,

Canada

10:30-10:55 Sparse Modeling of Random Phase TexturesLionel Moisan, Universite Paris 5, France

11:00-11:25 Optimal Transport Mixing of Gaussian Texture ModelsGabriel Peyré, Université Paris Dauphine,

France


Sunday,May20

CP3Contributed Session 39:30 AM-11:30 AMRoom:Rhapsody - 4th Floor

Chair: Carlos A. Ramirez, University of Texas, El Paso, USA

9:30-9:45 A New Variational Model for Removal of Both Additive and Multiplicative Noise and a Fast Algorithm for Its Numerical ApproximationNoppadol Chumchob, Silpakorn University,

Thailand and Centre for Mathematical Imaging Techniques, United Kingdom

9:50-10:05 Solving Highly Ill-Conditioned Blind Deconvolution ProblemsPaul Shearer and Anna Gilbert, University of

Michigan, USA

10:10-10:25 An l1 Minimization Algorithm with Applications in Image ProcessingCarlos A. Ramirez and Miguel Argaez,

University of Texas, El Paso, USA

10:30-10:45 Corrected Diffusion Approximation in Layered TissuesShelley B. Rohde and Arnold D. Kim,

University of California, Merced, USA

10:50-11:05 Iteratively Reweighted Least Squares for L1 Problems: Boring But Still EffectivePaul Rodriguez, Pontifical Catholic

University of Peru, Peru; Brendt Wohlberg, Los Alamos National Laboratory, USA

11:10-11:25 An Adaptive Spectrally Weighted Structure Tensor for Hyperspectral Imagery Based on the Heat OperatorMaider J. Marin-Mcgee and Miguel

Velez-Reyes, University of Puerto Rico, Mayaguez, Puerto Rico

Lunch Break11:30 AM-1:00 PMAttendees on their own

SIAG/IS Prize Lecture1:00 PM-1:45 PMRoom:Symphony Ballroom - 3rd Floor

Sunday,May20Intermission1:45 PM-2:00 PM

MS10Inverse Problems and Image Analysis in Remote Sensing Science - Part I of IV2:00 PM-4:00 PMRoom:Symphony Ballroom - 3rd Floor

For Part 2 see MS19 Remote sensing starts with the acquisition of information about an object or scene from a distance. These data can be spectral, multiangle, and polarimetric, and are generated using a variety of sensors, including optical, infrared, microwave, and radar. There is a broad range of remote sensing problems, solutions of which are immensely important for scientific understanding, and there has been a rapid development of robust computational approaches to address them. This interdisciplinary minisymposium will bring together researchers from different areas to represent the diversity of analytical and computational approaches for solving a variety of inverse problems in remote sensing science.

Organizer:IgorYanovskyJet Propulsion Laboratory, California Institute of Technology, USA

Organizer:AnthonyB.DavisCalifornia Institute of Technology, USA

Organizer:LuminitaA.VeseUniversity of California, Los Angeles, USA

2:00-2:25 Variational Methods for Remote Sensing ApplicationsIgor Yanovsky, Jet Propulsion Laboratory,

California Institute of Technology, USA

2:30-2:55 Radar ImagingMargaret Cheney, Rensselaer Polytechnic

Institute and Naval Postgraduate School, USA

3:00-3:25 Edge-based Image Reconstruction, Hierarchical Feature Extraction, and Analysis - A Structural ApproachLakshman Prasad, Los Alamos National

Laboratory, USA

3:30-3:55 Contrast Invariant and Affine Sub-Pixel Optical-Flow for Remote Sensing ApplicationsNeus Sabater, California Institute of

Technology, USA



Sunday,May20

MS11Analysis of Tomographic Images for Clinics: Recent Advances and Challenges - Part II of II2:00 PM-4:00 PMRoom:Assembly C - 5th Floor

For Part 1 see MS1 Analysis of tomographic images is of central importance for diagnostic and therapeutic applications in various clinical contexts. Distinct characteristics between tissues, along with different image properties and qualities among imaging modalities, have posed great challenges to the analysis tasks. Therefore, new ideas, adaptive models, and special techniques are in high demand. This minisymposium aims to bring together mathematicians, medical physicists, radiologists, and engineers, especially young scientists, to exchange the most recent advances in tomographic image analysis. We hope this minisymposium can provide a forum to stimulate discussions and establish collaborations between mathematicians and clinic scientists for further developments in this emerging research field.

Organizer:XiaojingYeGeorgia Institute of Technology, USA

Organizer:XunJiaUniversity of California, San Diego, USA

2:00-2:25 Temporally Consistent Segmentation of Longitudinal MRI Data and Related IssuesChunming Li, Vanderbilt University, USA

2:30-2:55 Marginal Space Learning for Efficient Detection and Segmentation of Anatomical Structures in Medical ImagingYefeng Zheng, Siemens Corporation

Research, Germany

Sunday,May20

MS12Advances in Sparse Recovery Algorithms - Part II of II2:00 PM-3:30 PMRoom:Assembly E - 5th Floor

For Part 1 see MS2 Nonsmooth optimization plays a central role in sparse recovery. The recent theory of compressed sensing has revitalized the interest in such optimization problems. Sparse recovery algorithms have an extensive literature, all of which can in essence be categorized into three main classes: convex (geometric), greedy (combinatorial), and probabilistic (variational/Belief Propagation) algorithms. The focus of this minisymposium is to give an overview of recent algorithmic advances in sparse recovery. We compare and contrast the advantages and limitations of these seemingly different approaches, and identify potential new research avenues for their interactions. Session will gather talks by leading experts in the field.

Organizer:VolkanCevherÉcole Polytechnique Fédérale de Lausanne, Switzerland


2:00-2:25 Generalized Forward-Backward Splitting for Sparse RecoveryJalal Fadili, Université de Caen, France;

Gabriel Peyre, Université Paris Dauphine, France; Hugo Raguet, Universite Paris Dauphine and CNRS, France

2:30-2:55 General Purpose First-order Methods for Convex MinimizationStephen Becker, California Institute of

Technology, USA; Emmanuel Candes, Stanford University, USA; Michael C. Grant, California Institute of Technology, USA; Jalal Fadili, Université de Caen, France

3:00-3:25 Constructing Test Instances for Sparse Recovery AlgorithmsChristian Kruschel, and Dirk Lorenz,

Technische Universitaet Braunschweig, Germany

3:00-3:25 An Efficient Algorithm for Multiphase Image Segmentation with Intensity Bias CorrectionHaili Zhang and Yunmei Chen, University

of Florida, USA; Xiaojing Ye, Georgia Institute of Technology, USA

3:30-3:55 A Geodesic Active Contour Based Model for Short Axis Cardiac-MR Image SegmentationSung Ha Kang, Georgia Institute of

Technology, USA; Wei Zhu, University of Alabama, USA; George Biros, Georgia Institute of Technology, USA



Sunday,May20

MS13Computational Photography2:00 PM-3:30 PMRoom:Concerto A - 3rd Floor

Recent advances in digital imaging technology and in mathematical signal processing have created new possibilities for photography. Examples include massively detailed gigapixel images, improved scene understanding via finely time-resolved detection of photons, selective focusing of photographs after the images have been recorded, and capturing huge dynamical ranges of illumination by combining different exposures. This minisymposium collects together some of these exciting new directions of photographic imaging.

Organizer:SamuliSiltanenUniversity of Helsinki, Finland

2:00-2:25 Sensing Videos with the Single Pixel CameraRichard G. Baraniuk, Rice University, USA

2:30-2:55 Infinite PhotographyTapio Helin, Matti Lassas, and Samuli

Siltanen, University of Helsinki, Finland

3:00-3:25 Femto-Photography: Time Resolved Imaging for Looking Around CornersRamesh Raskar, Massachusetts Institute of

Technology, USA

Sunday,May20

MS15Current Developments and Challenges in Imaging Through Turbulence - Part II of II2:00 PM-3:30 PMRoom:Aria A - 3rd Floor

For Part 1 see MS5 The video sequence captured in a long-range system, such as surveillance, is often corrupted by the atmospheric turbulence degradation. Each frame suffers from geometric distortion and focal blur. To process such data arises a large amount of challenges in image processing and computer vision, such as to real-time enhancement, object detection and recognition. This minisymposium is intended to present the state-of-the-art methods in reconstructing a video or a latent image with higher resolution, which include non-rigid registration, image fusion and blind deconvolution. The richness of its applications would draw attention to many researchers working on imaging science.

Organizer:YifeiLouGeorgia Institute of Technology, USA

Organizer:SungHaKangGeorgia Institute of Technology, USA

2:00-2:25 On-the-Fly Turbulence Effect Mitigation in Long-Range Surveillance Applications Based on Lucky Region FusionMathieu Aubailly, University of Maryland,

USA

2:30-2:55 Video Restoration of Turbulence DistortionYifei Lou and Sung Ha Kang, Georgia Institute

of Technology, USA; Stefano Soatto, and Andrea L. Bertozzi, University of California, Los Angeles, USA

3:00-3:25 Non Rigid Geometric Distortions Correction -- Application to Atmospheric Turbulence StabilizationYu Mao, University of Minnesota, USA;

Jerome Gilles, University of California, Los Angeles, USA

Sunday,May20

MS14Advances in 2D and 3D Hyperspectral Imaging - Part II of II2:00 PM-4:00 PMRoom:Concerto B - 3rd Floor

For Part 1 see MS4 This symposium addresses methods for hyperspectral image data as smoothing, clustering, segmentation and registration. In this context we consider images with potentially arbitrarily many color channels. The challanges in this area are edge preserving denoising of this type of images and dealing with data which is not sampled on regular grids as is the case after registration. Especially applications in mass spectrometrie imaging shall be focussed.

Organizer:StefanSchifflerUniversity of Bremen, Germany

Organizer:JanStrehlowFraunhofer MEVIS, Germany

2:00-2:25 Registration of MALDI Imaging DataStefan Heldmann, Judith Berger, Jan Strehlow,

and Stefan Wirtz, Fraunhofer MEVIS, Germany

2:30-2:55 Generating 3D MALDI Imaging DataJan Strehlow, Judith Berger, Stefan Heldmann,

and Stefan Wirtz, Fraunhofer MEVIS, Germany

3:00-3:25 Including Spatial Similarities into Hierarchical Clustering of Hyperspectral Terahertz ImagesHenrike Stephani, Fraunhofer Institute for

Industrial Mathematics, Germany; Karin Wiesauer and Stefan Katletz, RECENDT, Linz, Austria; Daniel Molter, Fraunhofer Institute for Physical Measurement Techniques, Germany; Bettina Heise, Johannes Kepler University, Austria

3:30-3:55 Artifact-free Decompression of Transform-coded Multi-channel Images with TGVMartin Holler and Kristian Bredies, University

of Graz, Austria


Sunday,May20

MS18Mathematical Challenges in 4D Imaging - Part I of III2:00 PM-4:00 PMRoom:Minuet - 4th Floor

For Part 2 see MS27 Modeling and understanding of complex systems in biology, medicine and geophysics require data that provide a solid resolution of spatial and temporal phenomena. Typically, the analysis of this data is memory consuming and computationally intense. With the increase of computing power and the invention of sophisticated numerical tools these large scale 3D and even 4D imaging problems are becoming tractable. In this minisymposium we aim to provide an overview on state-of-the-art imaging techniques from various fields which are mutually connected by similar technologies. The goal is to report on current computational approaches and challenges. Moreover, we aim to identify synergies between these fields.

Organizer:JanModersitzkiUniversity of Lübeck, Germany

Organizer:StefanHeldmannFraunhofer MEVIS, Germany

Organizer:NathanD.CahillRochester Institute of Technology, USA

2:00-2:25 4D Inverse Problems: Optimal Transport meets Sparsity and Low RankChristoph Brune, Hao Gao, and Stanley

J. Osher, University of California, Los Angeles, USA

2:30-2:55 Which Geometric Structure for the Statistical Analysis of Longitudinal Deformations?Xavier Pennec, INRIA, France

3:00-3:25 Motion Compensation for Dynamic Contrast Enhanced MRIStefan Heldmann, Fraunhofer MEVIS,

Germany; Stephen Keeling, University of Graz, Austria; Jan Modersitzki, University of Lübeck, Germany; Lars Ruthotto, University of Muenster, Germany

3:30-3:55 A Spatio-temporal Motion Correction Scheme for Cardiac PETLars Ruthotto, University of Muenster,

Germany; Jan Modersitzki, University of Lübeck, Germany

Sunday,May20

MS16Advances in Nonsmooth and Nonconvex Minimization for Imaging - Part II of II2:00 PM-4:00 PMRoom:Maestro A - 4th Floor

For Part 1 see MS6 Optimization plays a central role in modern image processing. While convex optimization has received considerable interest in the field, nonsmooth nonconvex optimization remains much less developed. Yet, there are many problems in image processing that involve nonconvex and nonsmooth functionals to minimize, e.g. in inverse problems. The minisymposium will give an overview of recent advances on both theoretical and algorithmic aspects of minimizing nonsmooth nonconvex objective functionals. Topics that will be covered range from well-posedness, to characterization of the minimizers and algorithms to reach them. The minisymposium will bring together recognized experts in optimization theory and imaging sciences.

Organizer:MilaNikolovaENS Cachan, France


2:00-2:25 On Local Linear Convergence of Elementary Algorithms with Sparsity ConstraintsRussell Luke, University of Goettingen,

Germany

2:30-2:55 Lp(Ω)Optimization with p∈ [0,1)Karl Kunisch, Universität Graz, Austria;

Kazufumi Ito, North Carolina State University, USA

3:00-3:25 Sparse Approximation via Penalty Decomposition MethodsZhaosong Lu, Simon Fraser University,

Canada

3:30-3:55 Symbology-based Algorithms for Robust Bar Code RecoveryRachel Ward, University of Texas at Austin,

USA

Sunday,May20

MS17Mathematical Modeling of Textures - Part II of II2:00 PM-3:30 PMRoom:Maestro B - 4th Floor

For Part 1 see MS7 Texture is a key component of natural images. Developing efficient models for textures is the key to go beyond the state of the art for several problems in image processing (compression, super-resolution), computer graphics (static and dynamic textures synthesis), computer vision (segmentation and retrieval) and psychophysics (visual perception). This requires to advance the research front in both statistical and deterministic models of natural images. This minisymposium features talks from world renown experts in the field of mathematical texture modeling.


Organizer:Jean-FrancoisAujolIMB, CNRS, Université Bordeaux 1, France

2:00-2:25 Texture, Structure and Visual MatchingStefano Soatto, University of California, Los

Angeles, USA

2:30-2:55 Maximum Entropy Texture Modeling Based on a Physiological Front EndEero P. Simoncelli, Courant Institute

of Mathematical Sciences, New York University, USA

3:00-3:25 Linking a Texture Model to Visual Processing in CortexJeremy Freeman, New York University, USA


Sunday,May20

MS20Hybrid Methods in Biomedical Imaging - Part I of II4:30 PM-6:30 PMRoom:Assembly C - 5th Floor

For Part 2 see MS37 Many of classical tomography modalitites either have insufficient contrast with respect to certain features of medical interest (such as cancerous tumors lying within soft tissues), or are unstable and cannot deliver high resolution necessary to recover such features. In recent years several novel hybrid techniques were introduced; these modalities combine two (or even three) different types of physical fields to yield both high contrast and high resolution required by today’s medicine. The present minisymposium brings together leading experts on photoacoustic, thermoacoustic, acousto-optic and acousto-electric tomographies.

Organizer:LeonidA.KunyanskyUniversity of Arizona, USA

4:30-4:55 Acousto-optic Imaging and Related Inverse ProblemsJohn Schotland, University of Michigan, USA

5:00-5:25 Electro-Acoustic Imaging MethodsYves Capdeboscq, Oxford University, United

Kingdom

5:30-5:55 Stabilizing Inverse Problems by Internal DataDustin Steinhauer, Texas A&M University,

USA

6:00-6:25 Joint Estimation of Wave Speed and Absorption Density Functions with Photoacoustic MeasurementOtmar Scherzer, University of Vienna,

Austria; Andreas Kirsch, Karlsruhe Institute of Technology, Germany

Sunday,May20

CP1Contributed Session 12:00 PM-4:00 PMRoom:Rhapsody - 4th Floor

Chair: Glenn Easley, System Planning Corporation, USA

2:00-2:15 Learning Sparsifying Transforms for Signal and Image ProcessingSaiprasad Ravishankar and Yoram Bresler,

University of Illinois, USA

2:20-2:35 A Compressive Sensing Synthetic Aperture Radar Method Incorporating SpeckleGlenn Easley, System Planning Corporation,

USA; Vishal Patel, University of Maryland, USA; Rama Chellappa, University of Maryland, College Park, USA

2:40-2:55 Image Compression With Nearest-Neighbor TransformsAliaksei Sandryhaila, and Jose M.F. Moura,

Carnegie Mellon University, USA

3:00-3:15 Robust High Frequency Information Guided Compressive Sensing ReconstructionWeihong Guo andJing Qin, Case Western

Reserve University, USA

3:20-3:35 Adapted Curvelet Sparse Regularization in Limited Angle TomographyJürgen Frikel, Helmholtz Zentrum

München, Germany

3:40-3:55 Sparse Shape ReconstructionAlireza Aghasi and Eric Miller, Tufts

University, USA; Justin Romberg, Georgia Institute of Technology, USA

Coffee Break4:00 PM-4:30 PMRoom:Orchestra - 2nd Floor

Sunday,May20

MS19Inverse Problems and Image Analysis in Remote Sensing Science - Part II of IV4:30 PM-6:30 PMRoom:Symphony Ballroom - 3rd Floor

For Part 1 see MS10 For Part 3 see MS32 Remote sensing starts with the acquisition of information about an object or scene from a distance. These data can be spectral, multiangle, and polarimetric, and are generated using a variety of sensors, including optical, infrared, microwave, and radar. There is a broad range of remote sensing problems, solutions of which are immensely important for scientific understanding, and there has been a rapid development of robust computational approaches to address them. This interdisciplinary minisymposium will bring together researchers from different areas to represent the diversity of analytical and computational approaches for solving a variety of inverse problems in remote sensing science.




4:30-4:55 Photon State Space and Radiative Transfer: The Physics of Imaging with Remote SensorsAnthony B. Davis, California Institute of

Technology, USA

5:00-5:25 Sensitivity Studies of Optical Scanning Polarimeter Retrievals of Aerosol and Cloud Optical PropertiesKirk Knobelspiesse, Brian Cairns, and

Michael Mishchenko, NASA Goddard Space Flight Center, USA

5:30-5:55 Spectral Invariance in Atmospheric RadiationAlexander Marshak, NASA Goddard Space

Flight Center, USA; Yuri Knyazikhin, Boston University, USA

6:00-6:25 Atmospheric Radiative Effects and their Correction in Landsat ImagesAlexei Lyapustin, NASA Goddard Space

Flight Center, USA


Sunday,May20

MS21Seismic Imaging and Inversion - Part I of III4:30 PM-6:30 PMRoom:Assembly E - 5th Floor

For Part 2 see MS29 Inferring Earth structure from seismic data is challenging because of the mult-scale, heterogeneous nature of the Earth. Further challenges lie in choosing how to acquire, represent and handle the large seismic data sets. The seismic inverse problem itself is generally non-convex and largely ill-posed, limiting the use of conventional optimization methods. As data are collected in regions of increasingly complicated geology, moving beyond linearized approximations is key to obtaining realistic, multi-scale models. To handle the large data sets intelligently, methods for sparse representation of wave fields are necessary to efficiently model and represent data, as well as to form images.

Organizer:AlisonMalcolmMassachusetts Institute of Technology, USA

Organizer:LaurentDemanetMassachusetts Institute of Technology, USA

Organizer:IvanVasconcelosSchlumberger Cambridge Research, United Kingdom

4:30-4:55 Do High Frequencies Contain Information about Low Frequencies?Laurent Demanet, Massachusetts Institute of

Technology, USA

5:00-5:25 Nonlinear Imaging and Inversion of Seismic Wavefields Using InterferometryIvan Vasconcelos, Schlumberger Cambridge

Research, United Kingdom; Clement Fleury, Colorado School of Mines, USA; Daniel Macedo, University of Campinas, Brazil

5:30-5:55 Modeling and Inverting Seismic P-S mode Conversions from Anelastic TargetsKris Innanen, University of Calgary, Canada

6:00-6:25 Interferometric Seismic Imaging by Sparse InversionJoost van der Neut, Delft University of

Technology, Netherlands; Tristan van Leeuwen andFelix J. Herrmann, University of British Columbia, Canada; Kees Wapenaar, Delft University of Technology, Netherlands

Sunday,May20

MS22Functional Analysis and Accurate Numerical Methods in Image Processing4:30 PM-6:30 PMRoom:Concerto A - 3rd Floor

In this minisymposium, methods in the calculus of variation and partial differential equations are used to understand mathematical properties in image denoising and colorization and to improve numerical simulations. In image colorization, existence of minimizers, reconstructibility, and faithfulness of color image restoration model are discussed. In image denoising, stability and consistency properties of various variational models are considered and fast, efficient, and accurate numerical methods for the Fenchel pre-dual in the total variation model are proposed. Comparison with well-known various efficient schemes for the Rudin-Osher-Fatemi model in terms of accuracy and efficiency will be undertaken.

Organizer:JooyoungHahnUniversity of Graz, Austria

4:30-4:55 Exact Reconstruction of Damaged Color Images using a Total Variation ModelIrene Fonseca and Giovanni Leoni, Carnegie

Mellon University, USA; Francesco Maggi, University of Florence, Italy; Massimiliano Morini, Universita degli Studi di Parma, Italy

5:00-5:25 Variational Models for Denoising of Images and their Functional Analytic PropertiesBarbara Zwicknagl, Carnegie Mellon

University, USA; Rustum Choksi, McGill University, Canada; Irene Fonseca, Carnegie Mellon University, USA

5:30-5:55 Accurate Numerical Schemes in Image DenoisingJooyoung Hahn and Carlos N. Rautenberg,

University of Graz, Austria; Michael Hintermueller, Humboldt University Berlin, Germany

6:00-6:25 Fast Algorithms for Second Order TV and Fourth Order Mean Curvature DenoisingKe Chen, University of Liverpool, United

Kingdom; Carlos Brito, Universidad Autónoma de Yucatán, Mexico; Li Sun, Lanzhou University, China

Sunday,May20

MS23Efficient Optimization Algorithms and their Application to Image Analysis - Part I of II4:30 PM-6:30 PMRoom:Concerto B - 3rd Floor

For Part 2 see MS31 Many problems in image analysis, including image denoising, image deblurring, image reconstruction, and compressed sensing, are eventually reduced to optimization problems. This minisymposium explores recently proposed optimization algorithms with application to image analysis.

Organizer:YunmeiChenUniversity of Florida, USA

Organizer:WilliamHagerUniversity of Florida, USA

Organizer:MaryamYashtiniUniversity of Florida, USA

4:30-4:55 Bregmanized Operator Splitting with Variable Stepsize: Convergence Analysis and MRI ApplicationMaryam Yashtini and William Hager,

University of Florida, USA; Xiaojing Ye, Georgia Institute of Technology, USA; Yunmei Chen, University of Florida, USA

5:00-5:25 Title Not Available at Time of PublicationStanley J. Osher, University of California,

Los Angeles, USA

5:30-5:55 Smoothed Sparse OptimizationMing-Jun Lai, University of Georgia, USA;

Wotao Yin, Rice University, USA

6:00-6:25 Iterative Regularization of Inverse Problems by Efficient Inexact Uzawa MethodsChristoph Brune, University of California,

Los Angeles, USA; Klaus Frick, University of Goettingen, Germany; Martin Burger, University of Muenster, Germany


Sunday,May20

MS25Advances in Operator Splitting Algorithms for Nonsmooth Optimization - Part I of II4:30 PM-6:00 PMRoom:Maestro A - 4th Floor

For Part 2 see MS33 Nonsmooth optimization is at the heart of many problems in modern image processing. It turns out that many of these nonsmooth objective functionals have a structure that can be exploited to design fast, effective and provably convergent splitting algorithms. The goal of these algorithms is to achieve full splitting where each part in the objective is used separately. The minisymposium will give an overview of recent advances in operator splitting for nonsmooth optimization, both in the convex and nonconvex case. The minisymposium will bring together recognized experts in optimization theory and imaging sciences.



Organizer:LaurentCondatUniversity of Caen, France

4:30-4:55 Smoothing and First Order Methods: A Unified FrameworkMarc Teboulle, Tel Aviv University, Israel;

Amir Beck, Technion - Israel Institute of Technology, Israel

5:00-5:25 Implementation of a Block Decomposition Algorithm for Solving Large-scale Conic Optimization ProblemsRenato Monteiro and Camilo Ortiz, Georgia

Institute of Technology, USA; Benar F. Svaiter, IMPA, Brazil

5:30-5:55 Primal-dual Splitting, Recent Improvements and VariantsThomas Pock, Graz University of

Technology, Austria; Chambolle Antonin, Ecole Polytechnique, France

Sunday,May20

MS24CANCELLED4:30 PM-6:30 PM

Sunday,May20

MS26Recent Advances in Patch-based Image Processing - Part I of III4:30 PM-6:30 PMRoom:Maestro B - 4th Floor

For Part 2 see MS34 The past few years have witnessed the emergence of a series of papers that tackle various image processing tasks in a locally adaptive, patch-based manner. These methods serve various applications, such as denoising, deblurring, inpainting, image decomposition, segmentation, super-resolution reconstruction, and more. As the name suggests, these techniques operate locally in the image, applying the same process for every pixel by manipulating small patches of pixels. In this minisymposium we intend to gather the leading researchers in this maturing arena to present a series of talks that will expose the current state of knowledge in this field.


Organizer:MichaelEladTechnion, Israel

Organizer:PeymanMilanfarUniversity of California, Santa Cruz, USA

4:30-4:55 The Pat(c)h Ahead: A Wide-Angle View of Image FilteringPeyman Milanfar, University of California,

Santa Cruz, USA

5:00-5:25 Poisson Noise Reduction with Non-local PCAJoseph Salmon, Duke University, USA;

Charles-Alban Deledalle, Université Paris Dauphine, France; Rebecca Willett and Zachary T. Harmany, Duke University, USA

5:30-5:55 Getting It Right: Parameter Selection For Non-Local Means Using SureDimitri Van De Ville, and Michel Kocher,

École Polytechnique Fédérale de Lausanne, Switzerland

6:00-6:25 Aggregation Methods for Optical Flow ComputationCharles Kervrann, INRIA, France


Sunday,May20

MS27Mathematical Challenges in 4D Imaging - Part II of III4:30 PM-6:30 PMRoom:Minuet - 4th Floor

For Part 1 see MS18 For Part 3 see MS35 Modeling and understanding of complex systems in biology, medicine and geophysics require data that provide a solid resolution of spatial and temporal phenomena. Typically, the analysis of this data is memory consuming and computationally intense. With the increase of computing power and the invention of sophisticated numerical tools these large scale 3D and even 4D imaging problems are becoming tractable. In this minisymposium we aim to provide an overview on state-of-the-art imaging techniques from various fields which are mutually connected by similar technologies. The goal is to report on current computational approaches and challenges. Moreover, we aim to identify synergies between these fields.




4:30-4:55 Numerical Techniques for Structural and Joint InversionEldad Haber, Emory University, USA; Michal

Holtzman Gazit, University of British Columbia, Canada

5:00-5:25 Fast Algorithms for 4D Biophysically-constrained Image RegistrationGeorge Biros, University of Texas at Austin,

USA

5:30-5:55 Large scale Imaging on Current Many-Core PlatformsHarald Koestler, University of Erlangen-

Nuremberg, Germany

6:00-6:25 Image-Based Modelling of Brain Tumour Progression: From Individualization to Priors for Non-Rigid Image RegistrationAndreas Mang and Thorsten M. Buzug,

University of Luebeck, Germany

Sunday,May20

MS67Sparse and Redundant Representations for Image Reconstruction and Geometry Extraction4:30 PM-6:30 PMRoom:Aria A - 3rd Floor

Sparse and redundant representations are powerful and popular mathematical tools for image denoising, reconstruction, sampling as well as feature extraction, classification etc. This minisymposium brings together researchers to discuss the recent advances in using wavelet frames, shearlets, and redundant dictionaries to reconstruct images and to extract geometrices.

Organizer:WeihongGuoCase Western Reserve University, USA

4:30-4:55 Shearlets and Geometry ExtractionGitta Kutyniok, Technische Universitaet

Berlin, Germany; David L. Donoho, Stanford University, USA; Wang-Q Lim, Technische Universitaet Berlin, Germany

5:00-5:25 Box Spline Wavelet Frames for Image Edge DetectionWeihong Guo, Case Western Reserve

University, USA; Ming-Jun Lai, University of Georgia, USA

5:30-5:55 The Analysis Co-Sparse Model: Pursuit, Dictionary Learning, and BeyondMichael Elad, Ron Rubinstein, and Tomer

Faktor, Technion, Israel

6:00-6:25 Sparse Approximation by Wavelet Frames with Applications to Image RestorationBin Dong, University of Arizona, USA

Sunday,May20


Chair: To Be Determined

4:30-4:45 Reconstruction from Spherical Mean Data by Summability MethodsFrank Filbir, Helmholtz Zentrum München,

Germany

4:50-5:05 3D Isar Image Reconstruction of Targets Through Filtered Back ProjectionZhijun Qiao, Jaime Lopez, and Tim Ray,

University of Texas - Pan American, USA

5:10-5:25 Actin Filament Tracking in Electron Tomograms of negatively stained Lamellipodia using the Localized Radon TransformChristoph Winkler, Johann Radon Institute for

Computational and Applied Mathematics, Austria; Marlene Vinzenz and Victor Small, Austrian Academy of Sciences, Austria; Christian Schmeiser, University of Vienna, Austria

5:30-5:45 Attenuation Compensation in Ultrasound ImagingJue Wang, Union College, USA; Yongjian Yu,

InfiMed Inc., USA

5:50-6:05 Automatic MRI Scan Positioning for Feeding Arteries of the BrainYan Cao, University of Texas, Dallas, USA;

Peiying Liu and Hanzhang Lu, University of Texas Southwestern Medical Center at Dallas, USA

6:10-6:25 Restricted Convergence of Projected Landweber Iteration in Banach Spaces for Nonlinear Inverse ProblemsLingyun Qiu and Maarten V. de Hoop,

Purdue University, USA; Otmar Scherzer, University of Vienna, Austria

Dinner Break6:30 PM-8:00 PMAttendees on their own


Monday,May21

IP2Targeted Seismic Imaging8:15 AM-9:00 AMRoom:Symphony Ballroom - 3rd Floor

Chair: Samuli Siltanen, University of Helsinki, Finland

The seismic imaging problem, which is typically formulated as an inverse problem for the wavespeed in the acoustic wave equation, involves the estimation of both the smooth and oscillatory part of this wavespeed. For both of these problems, a large data set is typically recorded and subjected to extensive processing to estimate both parts of the wavespeed throughout the entire region the data are sensitive to. For many applications, however, only a part of this image is of interest. I will describe techniques for enhancing part of the image through appropriate data choices and regularization.

AlisonMalcolmMassachusetts Institute of Technology, USA

Coffee Break9:00 AM-9:30 AMRoom:Orchestra - 2nd Floor

Monday, May 21

Registration7:45 AM-5:00 PMRoom:Overture - 3rd Floor

Remarks8:10 AM-8:15 AMRoom:Symphony Ballroom - 3rd Floor

Sunday,May20

PP1Welcome Reception and Poster Session8:00 PM-10:00 PMRoom:Ormandy West - Lobby Level

Automated Parameter Selection Tool for Solution to Ill-Posed ProblemsBrianna Cash, University of Maryland,

College Park, USA

Phase Retrieval with Random IlluminationWenjing Liao and Albert Fannjiang,

University of California, Davis, USA

A Generalized Simulated Annealing Approach to Image ReconstructionMarc C. Robini, INSA Lyon, France

Image Reconstruction with Multiresolution Edge-Continuation and Non-Quadratic Data FidelityMarc C. Robini and Pierre-Jean Viverge,

INSA Lyon, France

Direct Reconstruction from Spherical Mean DataRuben Seyfried, Helmholtz Zentrum

München, Germany

A Nonconvex TVq-Model in Image RestorationMichael Hintermüller, Humboldt University

Berlin, Germany; Tao Wu, University of Graz, Austria


Monday,May21

MT2Mathematics and Science of 3D Electron Microscope Imaging9:30 AM-11:30 AMRoom:Symphony Ballroom - 3rd Floor

Chair: Ozan Öktem, KTH Royal Institute of Technology, Sweden

Chair: Eric Todd Quinto, Tufts University, USA

The two parts of this minitutorial deal with mathematics for Electron Tomography (ET) and Single Particle (SP). The one on ET surveys reconstruction methods for ET with emphasis on difficulties and their mathematical consequences. It also includes the model for image formation in electron microscopy, which is the same in SP. The mini tutorial on SP focuses on presenting a novel reconstruction method, focusing on resolving the angular reconstitution problem. The method in question is based on a novel framework for non-linear optimization, which in turn is based on representation theoretic ideas - where the solution is an object of a category.

Speakers:

Ronny Hadani, University of Texas at Austin, USA


Monday,May21

MS28Mathematics of Medical Imaging and Shape Analysis - Part I of III9:30 AM-11:30 AMRoom:Assembly C - 5th Floor

For Part 2 see MS46 Medical imaging is the technique used to create images of human body seeking to reveal and examine diseases. The major challenge is to reconstruct high quality images using very limited data. Shape analysis, on the other hand, is a rapidly rising area whose aim is to provide intrinsic geometric descriptors of biological structures that can be used to detect and quantify diseases. Imaging is the process of acquiring data, while shape analysis is to analyze the acquired data. This minisymposium is to promote interactions between the two subjects that, we hope, can stimulate further developments of each subject.

Organizer:BinDongUniversity of Arizona, USA

Organizer:RongjieLaiUniversity of Southern California, USA

Organizer:DavidGuState University of New York, Stony Brook, USA

Organizer:RonaldLokMingLuiChinese University of Hong Kong, Hong Kong

9:30-9:55 Laplace-Beltrami Eigen-Geometry and Applications to 3D Medical ImagingRongjie Lai, University of Southern

California, USA

10:00-10:25 Respiratory Signal Extraction from X-ray Projection Images in 4D Cone Beam CTXun Jia, Hao Yan and Xiaoyu Wang,

University of California, San Diego, USA; Wotao Yin, Rice University, USA; Steve Jiang, University of California, San Diego, USA

10:30-10:55 Rosian Noise Removal on HARDI Data in Medical ImagingMelissa Tong, University of California, Los

Angeles, USA; Yunho Kim, University of California, Irvine, USA; Luminita A. Vese, University of California, Los Angeles, USA

11:00-11:25 Real-Time Anatomy Tracking with Low-SNR MRIDan Ruan, University of California, Los Angeles, USA



Monday,May21

MS31Efficient Optimization Algorithms and their Application to Image Analysis - Part II of II9:30 AM-11:30 AMRoom:Concerto B - 3rd Floor

For Part 1 see MS23 Many problems in image analysis, including image denoising, image deblurring, image reconstruction, and compressed sensing, are eventually reduced to optimization problems. This minisymposium explores recently proposed optimization algorithms with application to image analysis.

Organizer:YunmeiChenUniversity of Florida, USA

Organizer:WilliamHagerUniversity of Florida, USA

Organizer:MaryamYashtiniUniversity of Florida, USA

9:30-9:55 A Primal Dual Method for Nonconvex Problems and ApplicationsXiaoqun Zhang, Shanghai Jiaotong University,

China; Ernie Esser, University of California, Irvine, USA

10:00-10:25 Nonsmooth Image Reconstruction Algorithms and Their Applications in Partially Parallel MR ImagingXiaojing Ye, Georgia Institute of Technology,

USA; Yunmei Chen and William Hager, University of Florida, USA

10:30-10:55 A Primal Dual Method for Solving a Convex Model for DOAS AnalysisErnie Esser and Jack Xin, University of

California, Irvine, USA

11:00-11:25 Title Not Available at Time of PublicationJing Yuan, University of Western Ontario,

Canada

Monday,May21

MS30Low Rank Modeling and its Applications to Imaging - Part I of II9:30 AM-11:30 AMRoom:Concerto A - 3rd Floor

For Part 2 see MS52 Low rank modeling is arguably the most ubiquitous paradigm of data analysis and related fields. Despite the many existing standard techniques and their useful applications, there has been an overflow of recent developments with fundamentally novel viewpoints and new kinds of applications. These include the completion of missing values of low-rank matrices, robust low-rank models with sparse corruptions, mixtures of low-rank models, structured dictionary learning and selective sampling of hierarchically-structured high-rank matrices. The goal of this workshop is to present some of the recent theoretical progress of low rank modeling and its important applications to imaging and computer vision.

Organizer:GiladLermanUniversity of Minnesota, USA

9:30-9:55 PCA with Outliers and Missing DataSujay Sanghavi, University of Texas at Austin,

USA

10:00-10:25 A Novel M-Estimator for Robust PCATeng Zhang, and Gilad Lerman, University of

Minnesota, USA

10:30-10:55 Online Subspace Estimation and Tracking from Incomplete and Corrupted DataLaura Balzano, University of Wisconsin,

USA; Jun He, Nanjing University of Science & Technology, China; Arthur Szlam, New York University, USA; Brian Eriksson, Boston University, USA; Benjamin Recht and Robert Nowak, University of Wisconsin, USA

11:00-11:25 On the Use of Low-rank Regularization for Learning Graphical Models with Missing DataPradeep Ravikumar, University of Texas,

Austin, USA

Monday,May21

MS29Seismic Imaging and Inversion - Part II of III9:30 AM-11:30 AMRoom:Assembly E - 5th Floor

For Part 1 see MS21 For Part 3 see MS38 Inferring Earth structure from seismic data is challenging because of the mult-scale, heterogeneous nature of the Earth. Further challenges lie in choosing how to acquire, represent and handle the large seismic data sets. The seismic inverse problem itself is generally non-convex and largely ill-posed, limiting the use of conventional optimization methods. As data are collected in regions of increasingly complicated geology, moving beyond linearized approximations is key to obtaining realistic, multi-scale models. To handle the large data sets intelligently, methods for sparse representation of wave fields are necessary to efficiently model and represent data, as well as to form images.



9:30-9:55 Velocity ContinuationSergey Fomel, University of Texas at Austin,

USA

10:00-10:25 Title Not Available at Time of PublicationFons ten Kroode, Shell International

Exploration & Production B.V., Netherlands

10:30-10:55 Applying Gauss-Newton and exact Newton methods to Full Waveform InversionJean Virieux, ISTerre, University Joseph

Fourier, France; Ludovic Metivier, Romain Brossier, and Stephane Operto, Grenoble Institut of Technology, France

11:00-11:25 Time-Lapse Image-Domain Wavefield TomographyJeffrey C. Shragge and David Lumley,

University of Western Australia, Australia


Monday,May21

MS33Advances in Operator Splitting Algorithms for Nonsmooth Optimization - Part II of II9:30 AM-11:30 AMRoom:Maestro A - 4th Floor

For Part 1 see MS25 Nonsmooth optimization is at the heart of many problems in modern image processing. It turns out that many of these nonsmooth objective functionals have a structure that can be exploited to design fast, effective and provably convergent splitting algorithms. The goal of these algorithms is to achieve full splitting where each part in the objective is used separately. The minisymposium will give an overview of recent advances in operator splitting for nonsmooth optimization, both in the convex and nonconvex case. The minisymposium will bring together recognized experts in optimization theory and imaging sciences.



Organizer:LaurentCondatUniversity of Caen, France

9:30-9:55 High-order Methods for Sparse Signal RecoveryTom Goldstein, Stanford University, USA

10:00-10:25 Forward-backward Splitting and Proximal Alternating Methods for Nonconvex Nonsmooth Tame ProblemsJérôme Bolte, Université Toulouse I, France

10:30-10:55 A Memory Gradient Algorithm for Nonconvex Regularization with Applications to Image RestorationEmilie Chouzenoux, Université Paris-Est

Marne-la-Vallée, France

11:00-11:25 Combinatorial Selection and Least Absolute Shrinkage via the CLASH AlgorithmVolkan Cevher, École Polytechnique Fédérale

de Lausanne, Switzerland

10:30-10:55 Texture Adaptive Image Restoration Using Fractional Order RegularizationFiorella Sgallari, University of Bologna,

Italy; Raymond H. Chan, Chinese University of Hong Kong, Hong Kong; Alessandro Lanza and Serena Morigi, University of Bologna, Italy

11:00-11:25 Patch-based Locally Optimal DenoisingPeyman Milanfar and Priyam Chatterjee,

University of California, Santa Cruz, USA

Monday,May21

MS32Inverse Problems and Image Analysis in Remote Sensing Science - Part III of IV9:30 AM-11:30 AMRoom:Aria A - 3rd Floor

For Part 2 see MS19 For Part 4 see MS41 Remote sensing starts with the acquisition of information about an object or scene from a distance. These data can be spectral, multiangle, and polarimetric, and are generated using a variety of sensors, including optical, infrared, microwave, and radar. There is a broad range of remote sensing problems, solutions of which are immensely important for scientific understanding, and there has been a rapid development of robust computational approaches to address them. This interdisciplinary minisymposium will bring together researchers from different areas to represent the diversity of analytical and computational approaches for solving a variety of inverse problems in remote sensing science.




9:30-9:55 Inverse Transport Problems in Remote Sensing ApplicationsGuillaume Bal, Columbia University,

USA; Anthony B. Davis, California Institute of Technology, USA; Ian Langmore, Columbia University, USA; Youssef M. Marzouk, Massachusetts Institute of Technology, USA

10:00-10:25 Edge-preserving Super-resolution via Blind Deconvolution and UpsamplingAntonio Marquina, University of Valencia,

Spain



Monday,May21

MS44Some Strategies to Improve Resolution in Electrical Impedance Tomography9:30 AM-11:30 AMRoom:Rhapsody - 4th Floor

It is well known that absolute Electrical Impedance Tomography (EIT) medical images need improvement in terms of spatial resolution and accuracy of the electrical properties. This minisymposium is intended to make public some of the ideas that may improve spatial resolution and accuracy of absolute images. The Approximation Error Theory, anatomy-based priors, and the combination of ultrasound and EIT are examples of recently explored strategies.

Organizer:RaulG.LimaUniversity of Sao Paulo, Brazil

9:30-9:55 Electrical Impedance Tomography and Ultrasound: Some ExperiencesAlex Hartov and Ryan Halter, Dartmouth

College, USA

10:00-10:25 An Unscented Kalman Filter with an Approximation Error Model for Electrical Impedance TomographyJari Kaipio, Kuopio University, Finland;

Raul Lima, University of Sao Paulo, Brazil

10:30-10:55 Prior Models of the Human Chest in Electrical Impedance TomographyFernando S. Moura, University of Sao

Paulo, Brazil

11:00-11:25 Regional Lung Perfusion Estimated by Electrical Impedance TomographyMarcelo Amato and Eduardo Leite Costa,

University of Sao Paulo, Brazil


Monday,May21

MS35Mathematical Challenges in 4D Imaging - Part III of III9:30 AM-11:30 AMRoom:Minuet - 4th Floor

For Part 2 see MS27 Modeling and understanding of complex systems in biology, medicine and geophysics require data that provide a solid resolution of spatial and temporal phenomena. Typically, the analysis of this data is memory consuming and computationally intense. With the increase of computing power and the invention of sophisticated numerical tools these large scale 3D and even 4D imaging problems are becoming tractable. In this minisymposium we aim to provide an overview on state-of-the-art imaging techniques from various fields which are mutually connected by similar technologies. The goal is to report on current computational approaches and challenges. Moreover, we aim to identify synergies between these fields.




9:30-9:55 Matrix Factorization Techniques for Multivariate Imaging AnalysisJames Gee, University of Pennsylvania, USA

10:00-10:25 Free Form Deformations with Dense Control Point Spacing for Image RegistrationNathan D. Cahill, Rochester Institute of

Technology, USA

10:30-10:55 4D Image Based Parameter Estimation of Vascular PropertiesLuca Bertagna, Emory University, USA;

Mauro Perego, Florida State University, USA; Alessandro Veneziani, Emory University, USA

11:00-11:25 Image Processing with Stochastic PDEsTobias Preusser, University of Bremen,

Germany; Torben Pätz, Fraunhofer MEVIS, Germany

Monday,May21

MS34Recent Advances in Patch-based Image Processing - Part II of III9:30 AM-11:30 AMRoom:Maestro B - 4th Floor

For Part 1 see MS26 For Part 3 see MS43 The past few years have witnessed the emergence of a series of papers that tackle various image processing tasks in a locally adaptive, patch-based manner. These methods serve various applications, such as denoising, deblurring, inpainting, image decomposition, segmentation, super-resolution reconstruction, and more. As the name suggests, these techniques operate locally in the image, applying the same process for every pixel by manipulating small patches of pixels. In this minisymposium we intend to gather the leading researchers in this maturing arena to present a series of talks that will expose the current state of knowledge in this field.




9:30-9:55 A Generalized Tree-Based Wavelet Transform and Its Application to Patch-Based Image DenoisingMichael Elad, Technion, Israel

10:00-10:25 Analysis of Image Patches: A Unified Geometric PerspectiveFrancois G. Meyer, University of Colorado,

Boulder, USA

10:30-10:55 Patch Based Image Denoising With and Without Dictionary LearningLei Zhang, Hong Kong Polytechnic

University, China

11:00-11:25 Matching By Tone MappingYacov Hel-Or, The Interdisciplinary Center,

Israel


straightforwardly performed. More challenging is the problem of single-particle reconstruction (SPR) where a three-dimensional density map is to be obtained from images of individual molecules present in random positions and orientations in the ice layer. Because it does not require the formation of crystalline arrays of macromolecules, SPR is a very powerful and general technique, which has been successfully used for 3D structure determination of many protein molecules and large complexes. I will introduce the main theoretical and computational (rather than experimental) challenges posed by SPR, and will focus on our (*) recent progress for obtaining three-dimensional structures, for image classification and for denoising. From the mathematical perspective, we combine ideas from different areas, such as, spectral graph theory, dimensionality reduction, representation theory and semidefinite programming. This conference also hosts a minisymposium on cryo-EM. (*) This is a joint project with Fred Sigworth (Yale), Yoel Shkolnisky (Tel Aviv), Ronny Hadani (UT Austin), and Shamgar Gurevich (Wisconsin-Madison).

AmitSingerPrinceton University, USA

Intermission1:45 PM-2:00 PM

Monday,May21

IP3Three Dimensional Structure Determination of Macromolecules by Cryo-electron Microscopy from the Applied Math Perspective1:00 PM-1:45 PMRoom:Symphony Ballroom - 3rd Floor

Chair: Simon Arridge, University College London, United Kingdom

Cryo-electron microscopy is a technique by which biological macromolecules are imaged in an electron microscope. The molecules are rapidly frozen in a thin layer of vitreous ice, trapping them in a nearly-physiological state. Cryo-EM images, however, have very low contrast, due to the absence of heavy-metal stains or other contrast enhancements, and have very high noise due to the small electron doses that can be applied to the specimen. Thus, to obtain a reliable three-dimensional density map of a macromolecule, the information from thousands of images of identical molecules must be combined. When the molecules are arrayed in a crystal, the necessary signal-averaging of noisy images is

Monday,May21

PD1Forward Looking Panel Discussion12:00 PM-12:45 PMRoom:Symphony Ballroom - 3rd Floor

Chair: Gunther Uhlmann, University of California, Irvine, USA and University of Washington, USA

Panelists:

Guillaume BalColumbia University, USA

Margaret CheneyRensselaer Polytechnic Institute and Naval

Postgraduate School, USA

Alison MalcolmMassachusetts Institute of Technology, USA

Fadil SantosaUniversity of Minnesota, USA

Amit SingerPrinceton University, USA


Monday,May21

MS37Hybrid Methods in Biomedical Imaging - Part II of II2:00 PM-4:00 PMRoom:Assembly C - 5th Floor

For Part 1 see MS20 Many of classical tomography modalitites either have insufficient contrast with respect to certain features of medical interest (such as cancerous tumors lying within soft tissues), or are unstable and cannot deliver high resolution necessary to recover such features. In recent years several novel hybrid techniques were introduced; these modalities combine two (or even three) different types of physical fields to yield both high contrast and high resolution required by today’s medicine. The present minisymposium brings together leading experts on photoacoustic, thermoacoustic, acousto-optic and acousto-electric tomographies.

Organizer:LeonidA.KunyanskyUniversity of Arizona, USA

2:00-2:25 Noise in Photoacoustic Tomography: From Point-like Detectors to Large integrating Detectors and its Influence on Spatial Resolution and SensitivityPeter Burgholzer, Research Center for Non

Destructive Testing GmbH, Linz, Austria

2:30-2:55 Compressed Sensing and Photo-Acoustic TomographySimon Arridge, University College London,

United Kingdom

3:00-3:25 Gradient-based Inversion for 3D Quantitative Photoacoustic ImagingTeedah Saratoon and Ben Cox, University

College London, United Kingdom; Tanja Tarvainen, University of Eastern Finland, Finland; Simon Arridge, University College London, United Kingdom

3:30-3:55 Experimental Translation of Image Reconstruction Methods for Three-Dimensional Optoacoustic TomographyMark Anastasio, and Kun Wang, Washington

University in St. Louis, USA; Sergey Ermilov, Richard Su, and Alexander Oraevsky, TomoWave Laboratories, USA

Monday,May21

MS36Three-Dimensional Macromolecular Structure Determination Using Cryo Electron Microscopy - Part I of II2:00 PM-4:00 PMRoom:Symphony Ballroom - 3rd Floor

For Part 2 see MS45 The cryo-electron microscopy (EM) reconstruction problem is to find the three-dimensional structure of a macromolecule given noisy samples of its two-dimensional projection images at unknown random directions and positions. Cryo-EM images have very low contrast, due to the absence of heavy-metal stains or other contrast enhancements, and have very high noise due to the small electron doses that can be applied to the specimen. This minisymposium is concerned with algorithms, computational tools and mathematical theory necessary for solving the cryo-EM problem.

Organizer:AmitSingerPrinceton University, USA

Organizer:RonnyHadaniUniversity of Texas at Austin, USA

2:00-2:25 Progress with Classification of Cryo-EM Data by Manifold Embedding and Other TechniquesJoachim Frank, Robert Langlois, and Hstau

Liao, Columbia University, USA; Chun Hong Yoon, University of Wisconsin, USA; Dimitris Giannakis, New York University, USA; Peter Schwander and Abbas Ourmazd, University of Wisconsin, USA

2:30-2:55 Resampling-based Assessment of Variability in 3D Reconstructions of Biological MoleculesPawel A. Penczek, University of Texas,

Houston, USA

3:00-3:25 Viewing Direction Estimation in cryo-EM Using SynchronizationYoel Shkolnisky, Tel Aviv University, Israel;

Amit Singer, Princeton University, USA

3:30-3:55 A Fourier-based Approach for Iterative 3D Reconstruction from Cryo-EM ImagesLanhui Wang, Princeton University, USA;

Yoel Shkolnisky, Tel Aviv University, Israel; Amit Singer, Princeton University, USA

Monday,May21

MS9X-Ray CT Image Reconstruction2:00 PM-4:00 PMRoom:Minuet - 4th Floor

With the emergence of compressive sensing theory, there is a renewed interest in exploring edge preserving, statistical and optimization based approaches in X-Ray CT image reconstruction. On the other hand, analytic image reconstruction methods are still the benchmark methods implemented in X-Ray CT machines due to their many advantages including computational efficiency and well understood predictable performance. This workshop will bring together researchers working on X-Ray CT image reconstruction and present the most recent advances in analytic and optimization based approaches thereby identify new directions and challenges in X-Ray CT image reconstruction.

Organizer:BirsenYaziciRensselaer Polytechnic Institute, USA

Organizer:EricMillerTufts University, USA

2:00-2:25 Tensor-based Formulation for Spectral Computed TomographyOguz Semerci, Ning Hao, Misha E. Kilmer,

and Eric Miller, Tufts University, USA

2:30-2:55 The Hilbert Transform in Analytic Cone-beam Image Reconstruction: A Visual TourJed Pack, GE Global Research, USA

3:00-3:25 A Fourier Integral Operator Model for Conebeam X-ray CT and its InversionBirsen Yazici and Zhengmin Li, Rensselaer

Polytechnic Institute, USA

3:30-3:55 Recent Advances in Image Reconstruction from Interior DataAlexander Katsevich and Alexander Tovbis,

University of Central Florida, USA


Monday,May21

MS40Higher-Order Models for Image Restoration - Part I of II2:00 PM-4:00 PMRoom:Concerto B - 3rd Floor

For Part 2 see MS49 Non-smooth variational problems and PDEs based on higher-order models gain increasing attention in various image restoration tasks including image denoising, decomposition and image inpainting. Inspired by the famous ROF (Rudin, Osher and Fatemi) model for image denoising, non-smooth penalties are used in these approaches which are in particular aware of higher-order smoothness. While still aiming to preserve sharp features (like edges) in the image, higher regularity is imposed, where necessary, to eliminate the well-known staircasing artifact present in solutions of the ROF model. In this minisymposium we want to bring together different recent higher- order imaging approaches and discuss their analytic properties and their efficient numerical solution.

Organizer:CarolaB.SchoenliebUniversity of Cambridge, United Kingdom

Organizer:KristianBrediesUniversity of Graz, Austria

2:00-2:25 Image Restoration with Total Generalised VariationKristian Bredies, University of Graz, Austria

2:30-2:55 A Variational Model for Functions of Bounded Hessian - Theory, Numerics and ApplicationsKostas Papafitsoros, University of Cambridge,

United Kingdom

3:00-3:25 Numerical Algorithms for Euler’s Elastica Models in Image ProcessingJooyoung Hahn, University of Graz, Austria;

Xue-cheng Tai, University of Bergen, Norway, and Nanyang Technological University, Singapore; Jason Ginmo Chung, Nanyang Technological University, Singapore; Yu Wang, Technion, Israel; Yuping Duan, Nanyang Technological University, Singapore

3:30-3:55 Infimal Convolution Regularizations with Discrete l1-type FunctionalsTanja Teuber, University of Kaiserslautern,

Germany

Monday,May21

MS39Models and Applications in Compressive Imaging - Part I of III2:00 PM-4:00 PMRoom:Concerto A - 3rd Floor

For Part 2 see MS48 Compressive Sensing is a relatively new field in mathematics whereby one can collect far fewer samples of a signal than traditional sampling theory would imply while still capturing the essential information content in the underlying signal. In compressive imaging applications, this could result in dramatically new paradigms in image collection as well as algorithms for image exploitation. This minisymposium will address some of the latest research in the underlying sensing or measurement models, algorithms for image and video reconstruction, adaptation, and algorithms for image/video exploitation directly from the compressive measurements. We will tie the theory to practical applications of interest.

Organizer:RobertR.MuiseLockheed-Martin, USA

Organizer:JustinRombergGeorgia Institute of Technology, USA

2:00-2:25 Noval Measurements in Compressive ImagingMark Neifeld, DARPA/DSO, USA

2:30-2:55 Code Design for Target Tracking and DiscriminationRonald Coifman, Yale University, USA

3:00-3:25 Stability and Robustness of Weak Orthogonal Matching PursuitsSimon Foucart, Drexel University, USA

3:30-3:55 Adaptive Compressive Imaging Using Sparse Hierarchical Learned DictionariesJarvis Haupt, University of Minnesota, USA

Monday,May21

MS38Seismic Imaging and Inversion - Part III of III2:00 PM-4:00 PMRoom:Assembly E - 5th Floor

For Part 2 see MS29 Inferring Earth structure from seismic data is challenging because of the mult-scale, heterogeneous nature of the Earth. Further challenges lie in choosing how to acquire, represent and handle the large seismic data sets. The seismic inverse problem itself is generally non-convex and largely ill-posed, limiting the use of conventional optimization methods. As data are collected in regions of increasingly complicated geology, moving beyond linearized approximations is key to obtaining realistic, multi-scale models. To handle the large data sets intelligently, methods for sparse representation of wave fields are necessary to efficiently model and represent data, as well as to form images.



2:00-2:25 Resolution Analysis in Full Waveform InversionAndreas Fichtner and Jeannot Trampert,

Utrecht University, The Netherlands

2:30-2:55 Seismic Full Waveform Inversion, Recent Works and Research DirectionsHenri Calandra, Total CSTGF, France;

Changsoo Shin, Seoul National University, Korea; Christian Rivera, Total E&P, USA; Herve Chauris, Ecole des Mines de Paris, France; Alexander Aravkin, University of British Columbia, Canada

3:00-3:25 Dreamlet as a Physical Wavelet for Seismic Data Compression and ImagingRu-Shan Wu, University of California, Santa

Cruz, USA

3:30-3:55 Migration Based on Two-way Depth ExtrapolationGregory Beylkin, University of Colorado,

USA


Monday,May21

MS43Recent Advances in Patch-based Image Processing - Part III of III2:00 PM-4:00 PMRoom:Maestro B - 4th Floor

For Part 2 see MS34 The past few years have witnessed the emergence of a series of papers that tackle various image processing tasks in a locally adaptive, patch-based manner. These methods serve various applications, such as denoising, deblurring, inpainting, image decomposition, segmentation, super-resolution reconstruction, and more. As the name suggests, these techniques operate locally in the image, applying the same process for every pixel by manipulating small patches of pixels. In this minisymposium we intend to gather the leading researchers in this maturing arena to present a series of talks that will expose the current state of knowledge in this field.




2:00-2:25 Optimal Transport Over the Space of PatchesGabriel Peyré, Université Paris Dauphine,

France

2:30-2:55 Learning to Sense GMMsGuillermo Sapiro, University of Minnesota,

Minneapolis, USA

3:00-3:25 Patch Foveation in Nonlocal ImagingAlessandro Foi, Tampere University of

Technology, Finland; Giacomo Boracchi, Politecnico di Milano, Italy

3:30-3:55 Analysis of a Variational Framework for Exemplar Based Image InpaintingGabriele Facciolo, ENS Cachan, France

Monday,May21

MS42Deterministic and Probabilistic Methods for Inverse Scattering and Impedance Tomography - Part I of II2:00 PM-4:00 PMRoom:Maestro A - 4th Floor

For Part 2 see MS51 Inverse scattering and impedance tomography have been a very active field of applied mathematics in recent years. New trends have emerged that have allowed to obtain further insights and encouraging results for well established and fascinating inverse problems. The minisymposium focuses on recent developments and innovative contributions in this direction, considering theoretical results and numerical algorithms as well as their application to specific real world problems. Both deterministic and probabilistic formulations will be presented.

Organizer:RolandGriesmaierUniversity of Mainz, Germany

Organizer:NuuttiHyvonenAalto University, Finland

2:00-2:25 Inverse Scattering Problem for Inhomogeneous Media Containing ObstaclesFioralba Cakoni, University of Delaware,

USA; Anne Cossonniere, CERFACS, France; Houssem Haddar, Ecole Polytechnique, France

2:30-2:55 A Mixed Formulation of the Quasi-reversibility Method to Solve Elliptic Cauchy Problems in 2d and 3dJeremi Darde, Antti Hannukainen, and

Nuutti Hyvonen, Aalto University, Finland

3:00-3:25 Inverse Source Problems and the Windowed Fourier TransformRoland Griesmaier, University of Mainz,

Germany; Martin Hanke, Johannes Gutenberg-Universität, Mainz, Germany; Thorsten Raasch, University of Mainz, Germany

3:30-3:55 A Monotony Based Method for Electrical Impedance TomographyMarcel Ullrich and Bastian Harrach,

Universität Würzburg, Germany

Monday,May21

MS41Inverse Problems and Image Analysis in Remote Sensing Science - Part IV of IV2:00 PM-4:00 PMRoom:Aria A - 3rd Floor

For Part 3 see MS32 Remote sensing starts with the acquisition of information about an object or scene from a distance. These data can be spectral, multiangle, and polarimetric, and are generated using a variety of sensors, including optical, infrared, microwave, and radar. There is a broad range of remote sensing problems, solutions of which are immensely important for scientific understanding, and there has been a rapid development of robust computational approaches to address them. This interdisciplinary minisymposium will bring together researchers from different areas to represent the diversity of analytical and computational approaches for solving a variety of inverse problems in remote sensing science.




2:00-2:25 Structured Models for Inverse ProblemsGuillermo Sapiro, University of Minnesota,

Minneapolis, USA

2:30-2:55 L1 based Template Matching and its Application to Hyper Spectral ImagingStanley J. Osher, University of California,

Los Angeles, USA

3:00-3:25 The Adaptive Inverse Scale Space Method for Hyperspectral UnmixingMichael Moeller, Martin Burger, and

Martin Benning, University of Muenster, Germany; Stanley J. Osher, University of California, Los Angeles, USA

3:30-3:55 Local Robust Principal Components for Nonlinear DatasetsJames Theiler, Los Alamos National

Laboratory, USA; Brendt Wohlberg and Rick Chartrand, Los Alamos National Laboratory, USA


Monday,May21

MS46Mathematics of Medical Imaging and Shape Analysis - Part II of III4:30 PM-6:30 PMRoom:Assembly C - 5th Floor

For Part 1 see MS28 For Part 3 see MS55 Medical imaging is the technique used to create images of human body seeking to reveal and examine diseases. The major challenge is to reconstruct high quality images using very limited data. Shape analysis, on the other hand, is a rapidly rising area whose aim is to provide intrinsic geometric descriptors of biological structures that can be used to detect and quantify diseases. Imaging is the process of acquiring data, while shape analysis is to analyze the acquired data. This minisymposium is to promote interactions between the two subjects that, we hope, can stimulate further developments of each subject.





4:30-4:55 Surface Ricci Flow and its ApplicationsDavid Gu, State University of New York,

Stony Brook, USA

5:00-5:25 Robust Principal Component Analysis-based Four-dimensional Computed TomographyHao Gao, University of Los Angeles, USA;

Jianfeng Cai, University of Iowa, USA; Zuowei Shen, University of Wisconsin, USA; Hongkai Zhao, University of California, Irvine, USA

Monday,May21

MS45Three-Dimensional Macromolecular Structure Determination Using Cryo Electron Microscopy - Part II of II4:30 PM-6:30 PMRoom:Symphony Ballroom - 3rd Floor

For Part 1 see MS36 The cryo-electron microscopy (EM) reconstruction problem is to find the three-dimensional structure of a macromolecule given noisy samples of its two-dimensional projection images at unknown random directions and positions. Cryo-EM images have very low contrast, due to the absence of heavy-metal stains or other contrast enhancements, and have very high noise due to the small electron doses that can be applied to the specimen. This minisymposium is concerned with algorithms, computational tools and mathematical theory necessary for solving the cryo-EM problem.

Organizer:AmitSingerPrinceton University, USA

Organizer:RonnyHadaniUniversity of Texas at Austin, USA

4:30-4:55 Title Not Available at Time of PublicationFred Sigworth, Yale University, USA

5:00-5:25 Spherically Constrained Reconstruction in Cryo-EMHemant Tagare, Andrew Barthel, and Fred

Sigworth, Yale University, USA

5:30-5:55 New Rotationally Invariant Viewing Direction Classification Method for Cryo-EM Single Particle ReconstructionZhizhen Zhao, and Amit Singer, Princeton

University, USA

6:00-6:25 Detecting Molecular Symmetry and Shape Using Cryo-electron Microscopy, Geometry, and Group Representation TheoryShamgar Gurevich, University of Wisconsin,

Madison, USA; Ronny Hadani, University of Texas at Austin, USA; Amit Singer, Princeton University, USA

Monday,May21


Chair: Hayden Schaeffer, University of California, Los Angeles, USA

2:00-2:15 Image Recognition Via Moment MethodShanshan Huang and Mireille Boutin,

Purdue University, USA

2:20-2:35 An Algorithm for Shape Detection in Inverse ProblemsGunay Dogan, National Institute of

Standards and Technology, USA

2:40-2:55 Segmentation Based on a Limiting Mumford and Shah FunctionalHayden Schaeffer, and Luminita A. Vese,

University of California, Los Angeles, USA

3:00-3:15 Reconstruction of Images: Computation of Confidence IntervalsViktoria Taroudaki, University of Maryland,

USA; Dianne P. O’Leary, University of Maryland, College Park, USA

3:20-3:35 Restoration of Images Corrupted by Impulse Noise Using Blind Inpainting and l0 NormMing Yan, University of California, Los

Angeles, USA

3:40-3:55 Volterra Equations and Image Filtering and ProcessingEduardo Cuesta, University of Valladolid,

Spain; Mokhtar Kirane and Salman Malik, Universitè de la Rochelle, France


continued on next page


Monday,May21

MS48Models and Applications in Compressive Imaging - Part II of III4:30 PM-6:30 PMRoom:Concerto A - 3rd Floor

For Part 1 see MS39 For Part 3 see MS57 Compressive Sensing is a relatively new field in mathematics whereby one can collect far fewer samples of a signal than traditional sampling theory would imply while still capturing the essential information content in the underlying signal. In compressive imaging applications, this could result in dramatically new paradigms in image collection as well as algorithms for image exploitation. This minisymposium will address some of the latest research in the underlying sensing or measurement models, algorithms for image and video reconstruction, adaptation, and algorithms for image/video exploitation directly from the compressive measurements. We will tie the theory to practical applications of interest.



4:30-4:55 Efficient 2-D Deconvolution for Filters Yielding Triangular StructureJustin Romberg, Georgia Institute of

Technology, USA

5:00-5:25 GMM for Audio AnalysisGuillermo Sapiro, University of Minnesota,

Minneapolis, USA

5:30-5:55 Adaptive Compressive Imaging: Information-theoretic DesignAmit Ashok, University of Arizona, USA

6:00-6:25 Managing Model Complexity in High Frame Rate Compressive Video SensingMichael B. Wakin, Colorado School of

Mines, USA

Monday,May21

MS47Sparse Models for the Analysis and Exploitation of Imagery, Video and Geospatial Data4:30 PM-6:30 PMRoom:Assembly E - 5th Floor

New developments in data acquisition systems are producing inexpensive sensors that can collect significant amounts of data at very high rates. Thus the storing, transmitting, exploiting and visualizing of these massive datasets are taxing computer memory and computational time. Novel theoretical advancements in mathematics and statistics such as Compressive Sensing, Sparse Representations and Matrix Completion have led to algorithms that can efficiently analyze and exploit these large datasets. This minisymposium explores new computational approaches employed for applications such as data reduction, noise reduction, segmentation, activity-based learning and real-time user assisted image analysis.

Organizer:EdwardH.BoschNational Geospatial-Intelligence Agency, USA

Organizer:JohnGreerNational Geospatial-Intelligence Agency, USA

4:30-4:55 Sparse Modeling for Human Activity Analysis in Motion ImageryAlexey Castrodad, University of Minnesota,

USA; Guillermo Sapiro, University of Minnesota, Minneapolis, USA

5:00-5:25 A Convex Non-Negative Matrix Factorization and Dimensionality Reduction on Physical SpaceStanley J. Osher, University of California, Los

Angeles, USA

5:30-5:55 Evaluating Compressively Sensed Hyperspectral ImageryJohn Greer, National Geospatial-Intelligence

Agency, USA

6:00-6:15 Construction of Frames for Image Analysis and RepresentationEdward H. Bosch, National Geospatial-

Intelligence Agency, USA

6:15-6:30 Frame Theory and Sparse Deterministic Representations for Image ApplicationsWojtek Czaja, University of Maryland, College

Park, USA

5:30-5:55 Solving Pdes on Pont CloudsHongkai Zhao, University of California,

Irvine, USA; Rongjie Lai, University of Southern California, USA; Jian Liang and Alvin Wong, University of California, Irvine, USA

6:00-6:25 Fast Multiscale Reconstruction of Images from Compressed Sensing DataTom Goldstein, Stanford University, USA;

Judah Jacobson, University of California, Los Angeles, USA


Monday,May21

MS51Deterministic and Probabilistic Methods for Inverse Scattering and Impedance Tomography - Part II of II4:30 PM-6:00 PMRoom:Maestro A - 4th Floor

For Part 1 see MS42 Inverse scattering and impedance tomography have been a very active field of applied mathematics in recent years. New trends have emerged that have allowed to obtain further insights and encouraging results for well established and fascinating inverse problems. The minisymposium focuses on recent developments and innovative contributions in this direction, considering theoretical results and numerical algorithms as well as their application to specific real world problems. Both deterministic and probabilistic formulations will be presented.

Organizer:RolandGriesmaierUniversity of Mainz, Germany

Organizer:NuuttiHyvonenAalto University, Finland

4:30-4:55 Feasibility of Electrical Impedance Tomography Imaging for Non-destructive Testing of ConcreteAku Seppanen and Kimmo Karhunen,

University of Eastern Finland, Finland; Nuutti Hyvonen, Aalto University, Finland; Jari P Kaipio, University of Eastern Finland, Finland and University of Auckland, New Zealand

5:00-5:25 Simultaneous Reconstruction of Outer Boundary Shape and Conductivity Distribution in Electrical Impedance TomographyStratos Staboulis and Nuutti Hyvonen,

Aalto University, Finland; Aku Seppanen, University of Eastern Finland, Finland; Jeremi Darde, Aalto University, Finland

5:30-5:55 A Feynman-Kac-type Formula for Impedance TomographyMartin Simon and Martin Hanke, University

of Mainz, Germany; Petteri Piiroinen, University of Helsinki, Finland

Monday,May21

MS50Radar Imaging - Part I of II4:30 PM-6:30 PMRoom:Aria A - 3rd Floor

For Part 2 see MS59 Radar imaging is a technology that has been developed, very successfully, within the engineering community during the last 50 years. Radar systems on satellites now make beautiful images of regions of our earth and of other planets such as Venus. One of the key components of this impressive technology is mathematics, and many of the open problems are mathematical ones. This minisymposium will address some recent mathematical advances in the field.

Organizer:MargaretCheneyRensselaer Polytechnic Institute and Naval Postgraduate School, USA

Organizer:MatthewFerraraMatrix Research, Inc., USA

4:30-4:55 Multiple-Input Multiple-Output Synthetic Aperture Radar ImagingJeffrey Krolik and Li Li, Duke University,

USA

5:00-5:25 Polarimetric Synthetic-aperture Inversion for Extended Targets in ClutterKaitlyn Voccola, Air Force Research

Laboratory, USA; Margaret Cheney, Rensselaer Polytechnic Institute and Naval Postgraduate School, USA; Birsen Yazici, Rensselaer Polytechnic Institute, USA

5:30-5:55 Reducing the Ionospheric Distortions of Spaceborne SAR Images with the Help of Dual Carrier ProbingMikhail Gilman, Erick Smith, and Semyon

V. Tsynkov, North Carolina State University, USA

6:00-6:25 Estimating Size, Shape, and Motion via Near-field Invariants of Radar SignalsMark Stuff, Michigan Tech Research

Institute, USA

Monday,May21

MS49Higher-Order Models for Image Restoration - Part II of II4:30 PM-6:30 PMRoom:Concerto B - 3rd Floor

For Part 1 see MS40 Non-smooth variational problems and PDEs based on higher-order models gain increasing attention in various image restoration tasks including image denoising, decomposition and image inpainting. Inspired by the famous ROF (Rudin, Osher and Fatemi) model for image denoising, non-smooth penalties are used in these approaches which are in particular aware of higher-order smoothness. While still aiming to preserve sharp features (like edges) in the image, higher regularity is imposed, where necessary, to eliminate the well-known staircasing artifact present in solutions of the ROF model. In this minisymposium we want to bring together different recent higher-order imaging approaches and discuss their analytic properties and their efficient numerical solution.

Organizer:CarolaB.SchoenliebUniversity of Cambridge, United Kingdom

Organizer:KristianBrediesUniversity of Graz, Austria

4:30-4:55 A Convex, Lower Semi-continuous Approximation of the Euler-elastica FunctionalKristian Bredies, University of Graz,

Austria; Thomas Pock, Graz University of Technology, Austria; Benedikt K. Wirth, Courant Institute of Mathematical Sciences, New York University, USA

5:00-5:25 The H1 Norm of Tubular Neighborhoods of Curves and its Relation to the Elastica EnergyYves van Gennip, University of California,

Los Angeles, USA

5:30-5:55 Approximate Envelope Minimization for Higher-Order ModelsThomas Pock, Graz University of Technology,

Austria

6:00-6:25 High Order Models and Fast Algorithms for Fairing Variational Implicit SurfacesCarlos Brito, Universidad Autónoma de

Yucatán, Mexico; Ke Chen, University of Liverpool, United Kingdom


Monday,May21


Chair: Thomas Höft, Tufts University, USA

4:30-4:45 Reservoir Model Improvement by a Better Integration of 4D Seismic Data Using Image Processing MethodsRatiba Derfoul and Sebastien Da Veiga, IFP

Energies Nouvelles, France; Christian Gout and Carole Le Guyader, INSA Rouen, France

4:50-5:05 An Inverse Problem Approach to High Dynamic Range PhotographyThomas Höft, Tufts University, USA

5:10-5:25 Seismic Full-Waveform Inversion Using Sources and Receivers Compression SchemeAria Abubakar, Tarek Habashy, and

Guangdong Pan, Schlumberger-Doll Research, USA

5:30-5:45 A One Dimensional Algorithm for Seismic Imaging and Inversion: Theoretical Development and Numerical TestsBogdan G. Nita, Montclair State University,

USA; Ashley Ciesla, Brookedale Community College, USA; Christopher Smith, The College of New Jersey, USA

5:50-6:05 Compressing Aerial Images Using a Critical Representation of Laplacian PyramidFang Zheng and Youngmi Hur, Johns

Hopkins University, USA

Dinner Break6:30 PM-8:00 PMAttendees on their own

SIAG/IS Business Meeting8:00 PM-8:45 PMRoom:Symphony Ballroom - 3rd Floor

Complimentary beer and wine will be served.

Monday,May21

MS53Algorithms for Diffractive Imaging - Part I of II4:30 PM-6:30 PMRoom:Minuet - 4th Floor

For Part 2 see MS62 The latest advances in X-ray science have made it possible to deduce three dimension (3D) structures of a variety of objects from a collection of 2D diffraction patterns. In order to perform such a reconstruction one must be able to assemble 2D diffraction images in 3D to obtain a 3D diffraction pattern. A phase retrieval algorithm must then be applied to recover the structure of the object in real space. This minisymposium highlights the recent progress in both areas and challenges that require further research effort.

Organizer:ChaoYangLawrence Berkeley National Laboratory, USA

4:30-4:55 Algorithms for Single Molecule Diffractive ImagingChao Yang, Lawrence Berkeley National

Laboratory, USA

5:00-5:25 Comparison of Expectation-maximization and Graph-embedding Algorithms for Single Particle ImagingVeit Elser, Cornell University, USA

5:30-5:55 Computational Challenges for Biological Structure Determination Using X-ray DiffractionNicholas Sauter, and Ralf Grosse-Kunstleve,

Lawrence Berkeley National Laboratory, USA

6:00-6:25 Structure and Dynamics from Manifold Symmetries of Image FormationDimitris Giannakis, New York University,

USA

Monday,May21

MS52Low Rank Modeling and its Applications to Imaging - Part II of II4:30 PM-6:30 PMRoom:Maestro B - 4th Floor

For Part 1 see MS30 Low rank modeling is arguably the most ubiquitous paradigm of data analysis and related fields. Despite the many existing standard techniques and their useful applications, there has been an overflow of recent developments with fundamentally novel viewpoints and new kinds of applications. These include the completion of missing values of low-rank matrices, robust low-rank models with sparse corruptions, mixtures of low-rank models, structured dictionary learning and selective sampling of hierarchically-structured high-rank matrices. The goal of this workshop is to present some of the recent theoretical progress of low rank modeling and its important applications to imaging and computer vision.

Organizer:GiladLermanUniversity of Minnesota, USA

4:30-4:55 Dictionary Learning by L1 MinimizationJohn Wright, Columbia University, USA

5:00-5:25 Robust Locally Linear Analysis with Applications to Image Denoising and Blind InpaintingYi Wang, University of Minnesota, USA;

Arthur Szlam, New York University, USA; Gilad Lerman, University of Minnesota, USA

5:30-5:55 On Collective Matrix FactorizationMing Yuan, Georgia Institute of Technology,

USA

6:00-6:25 Efficient Spectral Clustering using Selective SimilaritiesAarti Singh, Carnegie Mellon University,

USA


Tuesday,May22

MS54Electron Tomography: Mathematics and Science - Part I of II9:30 AM-11:30 AMRoom:Symphony Ballroom - 3rd Floor

For Part 2 see MS65 Understanding the structural conformation of proteins and macromolecular assemblies is closely related to understanding their function within biological processes in the cell. Knowledge of these structures provides clues when developing therapeutic interventions for many diseases. Electron tomography is the only technique available to study structures of individual molecules in their natural environment. The reconstruction problem is however severely ill-posed and data are noisy, so advanced mathematics is needed. This workshop will bring together leaders at this interdisciplinary interface of image processing and stimulate new partnerships to address computational problems at this exciting frontier of cell biology.

Organizer:EricToddQuintoTufts University, USA

Organizer:OzanOktemKTH Stockholm, Sweden

9:30-9:55 Local Algorithms in Electron TomographyEric Todd Quinto, Tufts University, USA;

Ozan Oktem, KTH Stockholm, Sweden; Hans Rullgard, Stockholm University, Sweden

10:00-10:25 ET as an Inverse Problem with an Unbounded OperatorHolger Kohr, Universitaet des Saarlandes,

Germany

10:30-10:55 Computational Methods in CryoEMHans Hebert, Karolinska Institutet,

Sweden; Philip Koeck, Royal Institute of Technology, Sweden

11:00-11:25 Detection of Macromolecular Structures in Tomograms Using Reduced Representation TemplatesNeils Volkmann, Burnham Medical Research

Institute, USA

Tuesday,May22

IP4Divide and Conquer: Using Spectral Methods on Partitioned Images8:15 AM-9:00 AMRoom:Symphony Ballroom - 3rd Floor

Chair: Peter Burgholzer, Research Center for Non Destructive Testing GmbH, Linz, Austria

Spectral filtering methods such as Tikhonov regularization are powerful tools in deblurring images, but their usefulness is limited by two factors: the expense of the computation and the lack of flexibility in the filter function. This talk focuses on overcoming these limitations by partitioning the data in order to better process each segment. The data division can take several forms: for example, partitioning by spectral frequency, by input channel, or by subimages. Benefits of a divide and conquer approach to imaging can include smaller error, better conditioning of subproblems, and faster computations. We give examples using various data partitionings.

Speaker:DianneP.O’LearyUniversity of Maryland, College Park, USA

Co-authors:GlennEasleySystem Planning Corporation, USAJulianneChungUniversity of Texas at Arlington, USA

Coffee Break9:00 AM-9:30 AMRoom:Orchestra - 2nd Floor

Tuesday, May 22

Registration7:45 AM-5:00 PMRoom:Overture - 3rd Floor

Closing Remarks8:10 AM-8:15 AMRoom:Symphony Ballroom - 3rd Floor


Tuesday,May22

MS57Models and Applications in Compressive Imaging - Part III of III9:30 AM-11:30 AMRoom:Concerto A - 3rd Floor

For Part 2 see MS48 Compressive Sensing is a relatively new field in mathematics whereby one can collect far fewer samples of a signal than traditional sampling theory would imply while still capturing the essential information content in the underlying signal. In compressive imaging applications, this could result in dramatically new paradigms in image collection as well as algorithms for image exploitation. This minisymposium will address some of the latest research in the underlying sensing or measurement models, algorithms for image and video reconstruction, adaptation, and algorithms for image/video exploitation directly from the compressive measurements. We will tie the theory to practical applications of interest.



9:30-9:55 Recovering Sparse and Low Rank Matrices from Compressive MeasurementsRichard G. Baraniuk and Aswin

Sankaranarayanan, Rice University, USA

10:00-10:25 On the Power of Adaptivity in Sparse RecoveryPiotr Indyk, Massachusetts Institute of

Technology, USA

10:30-10:55 Spectral Super-Resolution of Hypercritical ImagesChris Rozell, Georgia Institute of Technology,

USA

11:00-11:25 Optimal Measurement Kernels for Dictionary Image PriorsRobert R. Muise, Lockheed-Martin, USA

Tuesday,May22

MS56New Directions in X-ray CT Reconstruction: Limited, Subsampled, Dynamic Data - Part I of II9:30 AM-11:30 AMRoom:Assembly E - 5th Floor

For Part 2 see MS66 Recently, limited data and subsampled X-ray CT received a lot of attention, in particular we mention breast tomosynthesis, dental CT and dynamic imaging. Somewhat parallel developments have been made in security applications, where cost-effectiveness and speed have been the major driving forces. This minisymposium addresses the image reconstruction problem for a wide range of limited data problems, the scope ranging from integration of prior information, non-linear methods, new cone beam geometries to sparsity enhanced image reconstruction as means to reduce the number of measurements and improving the quality of the reconstruction.

Organizer:MartaBetckeUniversity College London, United Kingdom


9:30-9:55 Approximation Error Approach in Local TomographyVille P. Kolehmainen, University of Kuopio,

Finland; Jari P Kaipio, University of Eastern Finland, Finland and University of Auckland, New Zealand

10:00-10:25 Adaptive Frequency-domain Regularization for Sparse-data TomographyMartti Kalke, University of Helsinki, Finland

10:30-10:55 Polyenergetic X-Ray Reconstruction Algorithms for Breast ImagingJames G. Nagy, Ioannis Sechopoulos and

Veronica M. Bustamante, Emory University, USA; Julianne Chung, University of Texas at Arlington, USA; Steve Feng, Emory University, USA

11:00-11:25 Multi-Sheet Rebinning Methods for Reconstruction from Asymmetrically Truncated Cone Beam ProjectionsMarta Betcke, University College London,

United Kingdom; Bill Lionheart, University of Manchester, United Kingdom

Tuesday,May22

MS55Mathematics of Medical Imaging and Shape Analysis - Part III of III9:30 AM-11:30 AMRoom:Assembly C - 5th Floor

For Part 2 see MS46 Medical imaging is the technique used to create images of human body seeking to reveal and examine diseases. The major challenge is to reconstruct high quality images using very limited data. Shape analysis, on the other hand, is a rapidly rising area whose aim is to provide intrinsic geometric descriptors of biological structures that can be used to detect and quantify diseases. Imaging is the process of acquiring data, while shape analysis is to analyze the acquired data. This minisymposium is to promote interactions between the two subjects that, we hope, can stimulate further developments of each subject.





9:30-9:55 Medical Morphometry and Computer Visions Using Quasi-Conformal Teichmuller TheoryRonald Lok Ming Lui, Chinese University of

Hong Kong, Hong Kong

10:00-10:25 Reconstruction of Binary Function from Incomplete Frequency InformationYu Mao, University of Minnesota, USA

10:30-10:55 Edge Detection and Image Reconstruction Using Area Preserving FlowsCatherine M. Kublik, University of Texas,

Austin, USA; Selim Esedoglu, University of Michigan, Ann Arbor, USA

11:00-11:25 Applications of Computational Quasiconformal Geometry on Medical Morphometry and Computer GraphicsAlvin Tsz Wai Wong, University of

California, Irvine, USA


Tuesday,May22

MS60Electrical Impedance Tomography: Algorithms and Applications - Part I of II9:30 AM-11:30 AMRoom:Maestro A - 4th Floor

For Part 2 see MS70 Electrical impedance tomography (EIT) is an imaging method where one probes an unknown conductive body using electric currents and aims to recover the inner conductivity distribution of the body from the data. The image formation task is an ill-posed and nonlinear inverse problem, requiring high-quality instrumentation and carefully regularized computational inversion algorithms. Applications of EIT include medical imaging, subsurface prospecting and non- destructive testing. This minisymposium presents recent progress of EIT imaging from the point of view of applications and computation.

Organizer:JenniferL.MuellerColorado State University, USA


9:30-9:55 Regularized D-bar Method for Electrical Impedance TomographyJennifer L. Mueller, Colorado State

University, USA; Kim Knudsen, Technical University of Denmark, Denmark; Matti Lassas and Samuli Siltanen, University of Helsinki, Finland

10:00-10:25 Electrical Impedance Tomography and the Noise SubspaceArmin Lechleiter, University of Bremen,

Germany

10:30-10:55 Direct Reconstruction in 3D Electrical Impedance Tomography Using a Nonlinear Fourier TransformKim Knudsen, Technical University of

Denmark, Denmark

11:00-11:25 Direct Reconstructions in EIT and Heat ProbingJanne P. Tamminen, Tampere University of

Technology, Finland

Tuesday,May22

MS59Radar Imaging - Part II of II9:30 AM-11:30 AMRoom:Aria A - 3rd Floor

For Part 1 see MS50 Radar imaging is a technology that has been developed, very successfully, within the engineering community during the last 50 years. Radar systems on satellites now make beautiful images of regions of our earth and of other planets such as Venus. One of the key components of this impressive technology is mathematics, and many of the open problems are mathematical ones. This minisymposium will address some recent mathematical advances in the field.


Organizer:MatthewFerraraMatrix Research, Inc., USA

9:30-9:55 Overview of Multi-baseline Radar Interferometry and Tomography and Applications PotentialScott Hensley, Jet Propulsion Laboratory,

California Institute of Technology, USA`

10:00-10:25 Stochastic and Deterministic Methods for Analysis of Oscillatory Radar SignalsMasoud Farshchian, Naval Research

Laboratory, USA

10:30-10:55 Mathematical Problems in Ultrawideband RadarEric Mokole, Naval Research Laboratory,

USA

11:00-11:25 Clutter Effects in ImagingKnut Solna, University of California, Irvine,

USA

Tuesday,May22

MS58Multi-frame Image and Video Processing - Part I of II9:30 AM-11:30 AMRoom:Concerto B - 3rd Floor

For Part 2 see MS68 The output of a given image processing task is often improved when several image inputs are considered. These situations can be quite diverse in application, spanning high dynamic range imaging, pan sharpening, medical image fusion, and denoising or analyzing video data (to name a few). The common thread is that all of these techniques involve merging information from multiple frames, and often their mathematical approaches transfer across applications. This minisymposium addresses work that uses multiple image frames to enhance or analyze image or video data.

Organizer:StaceyE.LevineDuquesne University, USA

Organizer:MarceloBertalmíoUniversitat Pompeu Fabra, Spain

9:30-9:55 A Variational Approach for the Fusion of Exposure Bracketed PairsMarcelo Bertalmío, Universitat Pompeu

Fabra, Spain; Stacey Levine, Duquesne University, USA

10:00-10:25 Optimal Estimators for HDR ReconstructionJulie Delon, Aguerrebere Cecilia, and

Gousseau Yann, Telecom Paris, France

10:30-10:55 Blind Subpixel Point Spread Function Estimation from Scaled Image PairsMauricio Delbracio, CMLA, ENS de

Cachan, France;Andrés Almansa, Telecom Paris, France; Pablo Muse, Universidad de la República, Uruguay; Jean Michel Morel, CMLA, ENS de Cachan, France

11:00-11:25 Burst DenoisingToni Buades, Paris Descartes, France; Yifei

Lou, Georgia Institute of Technology, USA; Jean-Michel Morel, Universitat de les Illes Balears, Spain; Zhongwei Tang, Ecole Normale Superieur de Cachan, France


Tuesday,May22

MS63Statistical Multiscale Methods in Imaging9:30 AM-11:30 AMRoom:Rhapsody - 4th Floor

In statistical imaging two seemingly different fundamental approaches are well established. These are multiscale methods, such as thresholding in wavelet or curvelet frames, on the one hand and variational methods, where e.g. estimators are defined as solutions of suitable optimization problems, on the other hand. This minisymposium aims for highlighting the recent achievements in connecting these two paradigms which, amongst others, include fully data driven and spatially adaptive image reconstruction methods.

Organizer:AxelMunkUniversity of Goettingen, Germany

Organizer:MarkusHaltmeierMax Planck Institute for Biophysical Chemistry, Germany

9:30-9:55 Statistical Extreme Value Theory Meets Multiscale ImagingMarkus Haltmeier, Max Planck Institute for

Biophysical Chemistry, Germany

10:00-10:25 Statistical Multiscale Methods for Biophotonic ImagingHannes Sieling, Klaus Frick, and Munk Axel,

University of Goettingen, Germany

10:30-10:55 Spectral Density Estimation for Unevenly Spaced Time SeriesIvan Mizera, University of Alberta, Canada

11:00-11:25 Geometric Multiscale Analysis and InpaintingEmily King, Universitaet Bonn, Germany;

Gitta Kutyniok, Wang-Q Lim, and Xiaosheng Zhuang, Technische Universitaet Berlin, Germany


Tuesday,May22

MS62Algorithms for Diffractive Imaging - Part II of II9:30 AM-11:30 AMRoom:Minuet - 4th Floor

For Part 1 see MS53 The latest advances in imaging science have made it possible to deduce three dimension (3D) structures of a variety of objects from a collection of 2D diffraction patterns. In order to perform such a reconstruction one must be able to assemble 2D diffraction images in 3D to obtain a 3D diffraction pattern. A phase retrieval algorithm must then be applied to recover the structure of the object in real space. This minisymposium highlights the recent progress in both areas and challenges that require further research effort.

Organizer:ChaoYangLawrence Berkeley National Laboratory, USA

Organizer:StefanoMarchesiniLawrence Berkeley National Laboratory, USA

9:30-9:55 Title Not Available at Time of PublicationStefano Marchesini, Lawrence Berkeley

National Laboratory, USA

10:00-10:25 Projection Methods for Low-count Diffractive Imaging in the Near and Far FieldsRussell Luke, University of Goettingen,

Germany

10:30-10:55 Real-time Ptychographic X-ray Image ReconstructionFilipe Maia, Lawrence Berkeley National

Laboratory, USA

11:00-11:25 PhaseLift: Exact and Stable Phase Retrieval Via Convex OptimizationVladislav Y. Voroninski, University of

California, Berkeley, USA; Emmanuel Candes, Stanford University, USA; Thomas Strohmer, University of California, Davis, USA

Tuesday,May22

MS61Unconventional Regularization Methods - Part I of II9:30 AM-11:30 AMRoom:Maestro B - 4th Floor

For Part 2 see MS71 Though regularization theory already goes back to the work of Tikhonov in the sixties, it is still a very active field, since the vast amount of different inverse problems constantly requires or at least benefits from more involved and better adapted regularization methods. However, the desired features of the regularized solution often ask for more complicated structures in the regularization method such as e.g. non-smoothness, non-convexity, or non-locality of the regularization terms, which not only lead to theoretically, but also to practically challenging problems. We hope to cover at least some of the new developments in this area in this minisymposium.

Organizer:ThomasFidlerUniversity of Vienna, Austria

Organizer:PeterElbauRICAM, Austrian Academy of Sciences, Austria

9:30-9:55 Variations on Lasso RegularizationChristine De Mol, Université Libre de

Bruxelles, Belgium

10:00-10:25 Linear Ill-posed Operator Equations with Poisson DataPaul Eggermont and Vincent LaRiccia,

University of Delaware, USA; Zuhair Nashed, University of Central Florida, USA

10:30-10:55 Source Identification from Line Integral Measurements and Simple Atmospheric ModelsSelim Esedoglu, University of Michigan,

Ann Arbor, USA

11:00-11:25 The Minimizers of Least Squares Regularized with l0 Norm: Uniqueness of the Global MinimizerMila Nikolova, ENS Cachan, France


Tuesday,May22

MS65Electron Tomography: Mathematics and Science - Part II of II2:00 PM-4:00 PMRoom:Symphony Ballroom - 3rd Floor

For Part 1 see MS54 Understanding the structural conformation of proteins and macromolecular assemblies is closely related to understanding their function within biological processes in the cell. Knowledge of these structures provides clues when developing therapeutic interventions for many diseases. Electron tomography is the only technique available to study structures of individual molecules in their natural environment. The reconstruction problem is however severely ill-posed and data are noisy, so advanced mathematics is needed. This workshop will bring together leaders at this interdisciplinary interface of image processing and stimulate new partnerships to address computational problems at this exciting frontier of cell biology.

Organizer:EricToddQuintoTufts University, USA

Organizer:OzanOktemKTH Stockholm, Sweden

2:00-2:25 Making Electron Tomography and Template Matching PracticalRemco Schoenmakers, FEI Company, The

Netherlands

2:30-2:55 Nonlinear Approximations as a Tool in TomographyMatt Reynolds, Lucas Monzon, and Gregory

Beylkin, University of Colorado at Boulder, USA

3:00-3:25 Mumford Shah Methods in Limited Data TomographyWolfgang Ring, University of Graz, Austria

3:30-3:55 Variational Methods in Molecular Electron TomographyChandrajit Bajaj, University of Texas at

Austin, USA

Tuesday,May22

MS64Stochastic Methods and Image Processing2:00 PM-4:00 PMRoom:Assembly C - 5th Floor

Stochastic methods for image analysis is a rapidly growing field that provides a general framework for a large class of imaging problems. Furthermore, imaging problem formulation within a Bayesian framework has recently witnessed a dramatic increase in popularity and utility. Optimal estimators of these models, such as Maximum a Posteriori and Minimum Mean Square Errors estimators, are generally difficult to compute and several techniques, which generally rely on Monte Carlo Markov Chain methods, have recently come to the forefront of computational research. This workshop covers both the Bayesian modeling formulation as well as efficient computational techniques for the estimators.

Organizer:ArjunaFlennerNaval Air Weapons Station, USA

2:00-2:25 Analysis of Bayesian Problems Using Non-Asyptotic Probability Tail BoundsGary Hewer, Naval Air Warfare Center,

Weapons Division, USA; Arjuna Flenner, Naval Air Weapons Station, USA

2:30-2:55 Non-parametric Bayesian Modeling in Image ProcessingLarry Carin, Duke University, USA

3:00-3:25 Total Variation Denoising Using Posterior ExpectationCecile Louchet, University of Orléans,

France

3:30-3:55 On Perfect Sampling of Some Markovian EnergiesJerome Darbon, ENS Cachan, France; Marc

Sigelle, Télécom ParisTech, France

Tuesday,May22

IP5Design in Imaging – From Compressive to Comprehensive Sensing1:00 PM-1:45 PMRoom:Symphony Ballroom - 3rd Floor

Chair: Jin Keun Seo, Yonsei University, South Korea

In the quest for improving imaging fidelity, vast attention has been devoted to effective solution of ill-posed problems under various regularization configurations. Nevertheless, issues such as determining optimal configurations for data acquisition or more generally any other controllable parameters of the apparatus and process are frequently overlooked. While design for well-posed problems has been extensively studied in the past, insufficient consideration has been given to its ill-posed counterpart. This is markedly in contrast to the fact that many real-life problems are of such nature. In this talk I shall describe some of the intrinsic difficulties associated with design for ill-posed inverse problems, lay out a coherent formulation to address them and finally demonstrate the importance of such designs for various imaging problems.

LiorHoreshIBM T.J. Watson Research Center, USA

Intermission1:45 PM-2:00 PM


Tuesday,May22

MS68Multi-frame Image and Video Processing - Part II of II2:00 PM-4:00 PMRoom:Concerto B - 3rd Floor

For Part 1 see MS58 The output of a given image processing task is often improved when several image inputs are considered. These situations can be quite diverse in application, spanning high dynamic range imaging, pan sharpening, medical image fusion, and denoising or analyzing video data (to name a few). The common thread is that all of these techniques involve merging information from multiple frames, and often their mathematical approaches transfer across applications. This minisymposium addresses work that uses multiple image frames to enhance or analyze image or video data.

Organizer:StaceyE.LevineDuquesne University, USA

Organizer:MarceloBertalmíoUniversitat Pompeu Fabra, Spain

2:00-2:25 Learning Human Activities and Interactions from VideoGuillermo Sapiro, University of Minnesota,

Minneapolis, USA

2:30-2:55 Variational Imaging on Manifolds and Some ApplicationsOtmar Scherzer and Nicolas Thorstensen,

University of Vienna, Austria

3:00-3:25 Exact Histogram Specifcation for Digital Images Using a Variational ApproachMila Nikolova, ENS Cachan, France;

Raymond H. Chan, Chinese University of Hong Kong, Hong Kong; You-Wei Wen, Kunming University of Science and Technology, China

3:30-3:55 Restoration of Photographies of PaintingsToni Buades, Paris Descartes, France; Gloria

Haro, Universitat Pompeu Fabra; Jean-Michel Morel, ENS Cachan, France

3:00-3:25 Proximal Algorithms and CT: New Results on 3D Real Datas and Color CTYannick Boursier, Mathieu Dupont, and

Sandrine Anthoine, Aix-Marseille Université - CNRS, France; Jean-Francois Aujol, IMB, CNRS, Université Bordeaux 1, France; Clothilde Melot, Aix-Marseille Université - CNRS, France

3:30-3:55 Low Rank Matrix Decomposition/Factorization Methods for 4D CtHongkai Zhao, University of California,

Irvine, USA; Jianfeng Cai, University of Iowa, USA; Hao Gao, University of California, Los Angeles, USA; Xun Jia and Steve Jiang, University of California, San Diego, USA; Zuowei Shen, University of Wisconsin, USA

Tuesday,May22

MS66New Directions in X-ray CT Reconstruction: Limited, Subsampled, Dynamic Data - Part II of II2:00 PM-4:00 PMRoom:Assembly E - 5th Floor

For Part 1 see MS56 Recently, limited data and subsampled X-ray CT received a lot of attention, in particular we mention breast tomosynthesis, dental CT and dynamic imaging. Somewhat parallel developments have been made in security applications, where cost-effectiveness and speed have been the major driving forces. This minisymposium addresses the image reconstruction problem for a wide range of limited data problems, the scope ranging from integration of prior information, non-linear methods, new cone beam geometries to sparsity enhanced image reconstruction as means to reduce the number of measurements and improving the quality of the reconstruction.

Organizer:MartaBetckeUniversity College London, United Kingdom


2:00-2:25 A Survey of CT Applications Enabled by Sparsity-exploiting Image ReconstructionEmil Sidky, University of Chicago, USA;

Jakob H. Joergensen, Technical University of Denmark, Denmark; Xiaochuan Pan, University of Chicago, USA

2:30-2:55 Connecting Object Sparsity and Number of Measurements in Total Variation-regularized X-ray CTJakob H. Joergensen, Technical University

of Denmark, Denmark; Emil Y. Sidky, University of Chicago, USA; Per Christian Hansen, Technical University of Denmark, Denmark; Xiaochuan Pan, University of Chicago, USA



Tuesday,May22

MS73Splitting-Based L1 Minimization with Applications in Medical Imaging, Electrochemical Impedance Spectroscopy, and Seismic Imaging2:00 PM-4:00 PMRoom:Aria A - 3rd Floor

L1 type of minimization such as total-variation regularization, sparse sensing play an important role in inverse problems. In this minisymposium, we want to bring together different large-scale computational applications, which employ L1 related regularization. In particular, we will cover applications including spiral MRI and ultrasound image reconstruction from medical imaging, estimation of polarization resistances of electrochemical impedance spectroscopy, and full-waveform inversion for seismic imaging. Computational issues such as operator splitting minimization methods, the selection of the regularization parameter, cost and performance analysis, and convergence will also be addressed and discussed.

Organizer:YouzuoLinLos Alamos National Laboratory, USA

Organizer:WolfgangStefanUniversity of Texas M. D. Anderson Cancer Center, USA

2:00-2:25 Solution of Ill-posed Inverse Problems Pertaining to Signal Restoration: Total Variation Parameter EstimationRosemary A. Renaut, Arizona State

University, USA

2:30-2:55 L1/TV Models and Its VariantsLixin Shen, Syracuse University, USA

Tuesday,May22

MS71Unconventional Regularization Methods - Part II of II2:00 PM-4:00 PMRoom:Maestro B - 4th Floor

For Part 1 see MS61 Though regularization theory already goes back to the work of Tikhonov in the sixties, it is still a very active field, since the vast amount of different inverse problems constantly requires or at least benefits from more involved and better adapted regularization methods. However, the desired features of the regularized solution often ask for more complicated structures in the regularization method such as e.g. non-smoothness, non-convexity, or non-locality of the regularization terms, which not only lead to theoretically, but also to practically challenging problems. We hope to cover at least some of the new developments in this area in this minisymposium.

Organizer:ThomasFidlerUniversity of Vienna, Austria

Organizer:PeterElbauRICAM, Austrian Academy of Sciences, Austria

2:00-2:25 Convergence of Variational Regularization Methods for Imaging on Riemannian ManifoldsNicolas Thorstensen, University of Vienna,

Austria

2:30-2:55 L-infinity Regularized Models for Segmentation, Cartoon-texture Decomposition, and Image RestorationLuminita A. Vese, University of California,

Los Angeles, USA

3:00-3:25 Morozov’s Principle for the Augmented Lagrangian MethodElena Resmerita, University Klagenfurt,

Austria

3:30-3:55 Regularization Based on Integral InvariantsThomas Fidler, University of Vienna, Austria

Tuesday,May22

MS70Electrical Impedance Tomography: Algorithms and Applications - Part II of II2:00 PM-4:00 PMRoom:Maestro A - 4th Floor

For Part 1 see MS60 Electrical impedance tomography (EIT) is an imaging method where one probes an unknown conductive body using electric currents and aims to recover the inner conductivity distribution of the body from the data. The image formation task is an ill-posed and nonlinear inverse problem, requiring high-quality instrumentation and carefully regularized computational inversion algorithms. Applications of EIT include medical imaging, subsurface prospecting and non- destructive testing. This minisymposium presents recent progress of EIT imaging from the point of view of applications and computation.

Organizer:JenniferL.MuellerColorado State University, USA


2:00-2:25 Imaging of Non-stationary Flows in Industrial Process TomographyAku Seppanen and Antti Lipponen, University

of Eastern Finland, Finland; Jari P Kaipio, University of Eastern Finland, Finland and University of Auckland, New Zealand

2:30-2:55 Problems in Electrical Impedance TomographyDavid Isaacson, Rensselaer Polytechnic

Institute, USA

3:00-3:25 Estimation of Spatial Resolution Improvement by the use of an Anatomy Based Prior of Pig Thoraxes in Electrical Impedance TomographyRaul G. Lima, University of Sao Paulo, Brazil

3:30-3:55 A Direct D-bar Method for the Reconstruction of Complex Admittivities from EIT DataSarah Hamilton, Colorado State University,

USA; Natalia Herrera, University of Sao Paulo, Brazil; Jennifer L. Mueller, Colorado State University, USA; Alan Von Herrmann, Colby College, USA



Tuesday,May22


Chair: Julianne Chung, University of Texas at Arlington, USA

2:00-2:15 A Variational Infinite Perimeter Model for Image Selective SegmentationLavdie Rada and Ke Chen, University of

Liverpool, United Kingdom

2:20-2:35 A Robust Approach to Detect Faces from Still ImagesQinghan Xiao, Defence R&D Canada,

Canada; Patrick Laytner and Chrisford Ling, University of Waterloo, Canada

2:40-2:55 An Optimal Approach for Image Deblurring and Denoising with Data Acquisition ErrorsJulianne Chung, University of Texas at

Arlington, USA; Matthias Chung, Texas State University, USA; Dianne P. O’Leary, University of Maryland, College Park, USA

3:00-3:15 Image Segmentation and Inpainting Techniques for Reconstruction of Incomplete Cell PathsJustin Wan, University of Waterloo, Canada

3:20-3:35 Pde Transforms and Non-Smooth Data: Singularity Detection, Segmentation and DenoisingRishu Saxena, Arizona State University, USA;

Siyang Yang, Michigan State University, USA

Tuesday,May22

CP6Contributed Session 62:00 PM-3:40 PMRoom:Minuet - 4th Floor

Chair: Cheng Guan Koay, University of Wisconsin, Madison, USA

2:00-2:15 Optimal Acquisition Design for Diffusion MRI: A Case of Searching for a Needle in the UniverseCheng Guan Koay, University of Wisconsin,

Madison, USA; Evren Özarslan, National Institutes of Health, USA; M. Elizabeth Meyerand, University of Wisconsin, Madison, USA

2:20-2:35 Simultaneous Image Reconstruction and Sensitivity Map Estimation in Partially Parallel MR ImagingMeng Liu and Yuyuan Ouyang, University

of Florida, USA; Xiaojing Ye, Georgia Institute of Technology, USA; Yunmei Chen, University of Florida, USA; Feng Huang, Invivo Corporation, USA

2:40-2:55 Precision Limits for Source Parameter Estimation in Slitless SpectrometryFigen S. Oktem and Farzad Kamalabadi,

University of Illinois at Urbana-Champaign, USA; Joseph M. Davila, NASA Goddard Space Flight Center, USA

3:00-3:15 Sparse Reconstruction Applied to Tomographic SAR InversionXiao Xiang Zhu, Technische Universitaet

Muenchen, Germany; Richard Bamler, German Aerospace Center, Oberpfaffenhofen, Germany

3:20-3:35 Bandwidth-Efficient Remote Visualization on Mobile DevicesSebastian Ritterbusch, Vincent Heuveline,

Roman Reiner, and Andreas Helfrich-Schkarbanenko, Karlsruhe Institute of Technology, Germany

Tuesday,May22

MS73Splitting-Based L1 Minimization with Applications in Medical Imaging, Electrochemical Impedance Spectroscopy, and Seismic Imaging2:00 PM-4:00 PMcontinued

3:00-3:25 A New Paradigm for Fast MRI Reconstruction with TV Regularization from Arbitrary Trajectories though K-SpaceWolfgang Stefan, University of Texas M. D.

Anderson Cancer Center, USA; Ken-Pin Hwang, GE Healthcare, USA; R. Jason Stafford and John D Hazle, University of Texas M. D. Anderson Cancer Center, USA

3:30-3:55 Edge-Preserving Full-Waveform Inversion with a Modified Total-Variation Regularization SchemeYouzuo Lin, Lianjie Huang, and Zhigang

Zhang, Los Alamos National Laboratory, USA


Tuesday,May22

MS74Inverse Problems and Statistical Learning in Imaging Applications4:30 PM-6:30 PMRoom:Concerto A - 3rd Floor

Computing reliable solutions to ill-posed inverse problems is important in many imaging applications (e.g., image deblurring, medical image reconstruction, computer vision, and geophysical imaging). In recent years, there has been significant progress in using statistical learning and modeling techniques for image processing and understanding. Speakers in this minisymposium will discuss novel methods, computational challenges, and recent results in the area of inverse problems and statistical learning techniques.

Organizer:JulianneChungUniversity of Texas at Arlington, USA

Organizer:MatthiasChungTexas State University, USA

4:30-4:55 Designing Optimal Spectral Filters for Inverse ProblemsMatthias Chung, Texas State University, USA

5:00-5:25 Fast Composite Splitting Algorithm for Linear Composite RegularizationJunzhou Huang, University of Texas at

Arlington, USA

5:30-5:55 Imaging the Earth’s Conductivity and ChangeabilityEldad Haber, Emory University, USA

6:00-6:25 Model-based Synthetic Aperture Radar ImagingVishal Patel, University of Maryland, USA

Tuesday,May22

MS72Publishing Reproducible Science and Software4:30 PM-6:30 PMRoom:Symphony Ballroom - 3rd Floor

Reproducibility, the ability of an experiment or study to be independently and accurately reproduced or replicated, is gaining momentum as a desirable principle of the scientific method for computational sciences. In order to become more common, this methodology needs so be implanted at the publication level with new contents, tools, criteria and journals. Software and data must become first-class scholarship, together with the traditional article content. The symposium gathers four speakers with an experience in publishing reproducible research and building and managing systems for publishing reproducible research. They will present their experience and the recent developments, current issues and future perspectives of their respective projects.

Organizer:NicolasLimareENS Cachan, France

4:30-4:55 Image Processing On Line: Experiences in Reproducible Science and SoftwareNicolas Limare, ENS Cachan, France; Toni

Buades, Paris Descartes, France

5:00-5:25 Open Science: When Open Source Meets Open AccessJulien Jomier and Luis Ibanez, Kitware,

Incorporated, USA

5:30-5:55 Open Research Computation: Supporting Reproducible Research Through Open Source SoftwareCameron Neylon, Science and Technology

Facilities Council, United Kingdom

6:00-6:25 Reproducible Research in Signal Processing: How to Increase ImpactMartin Vetterli, École Polytechnique

Fédérale de Lausanne, Switzerland; Patrick Vandewalle, ReproducibleResearch.org, USA

Tuesday,May22

MS69 CANCELLED 2:00 PM-4:00 PM



Tuesday,May22

MS77Binary Constrained Variational Models and Algorithms in Image Processing4:30 PM-6:30 PMRoom:Maestro A - 4th Floor

Binary constrained variational models are commonly used in several image processing areas, such as text restoration, image segmentation, multilabelling, and so on. Because of nonconvexity, it is also a big challenge to solve these models. This minisymposium aims at bringing together researchers in this area to present a series of talks on modeling, theoretical analysis, efficient numerical algorithms and applications.

Organizer:YiqiuDongHelmholtz Zentrum München, Germany

4:30-4:55 A Convex Shape Prior and Shape Matching Based on the Gromov-Wasserstein DistanceChristoph Schnoerr, University of Heidelberg,

Germany

5:00-5:25 A Binary Constrained General Model in Image ProcessingYiqiu Dong, Helmholtz Zentrum München,

Germany

5:30-5:55 Flow-Maximizing joint with Message-PassingJing Yuan, University of Western Ontario,

Canada

6:00-6:25 Optimality of Relaxations for Integer-Constrained ProblemsJan Lellmann, University of Cambridge,

United Kingdom

Tuesday,May22

MS76Radar Signal Processing4:30 PM-6:30 PMRoom:Aria A - 3rd Floor

Radar signal processing is a rich source of interesting mathematical problems, which arise from a need to understand the structure and properties of the various abstract levels in the radar processing chain. The problems range from those associated with the sampling, detection, and reconstruction of electromagnetic signals to the high-order analyses required to fully extract and understand the underlying information content in the sensed signals. This minisymposium, which will feature some of the latest work in this area, aims to highlight the inter-relationships between the various aspects of radar signal processing.


Organizer:RaghuRajNaval Research Laboratory, USA

4:30-4:55 Information Theory in Radar Imaging, Detection and ClassificationRaghu Raj, Naval Research Laboratory, USA

5:00-5:25 Radar Imaging and Feature Detection of Biological MediaAnalee Miranda, Air Force Research

Laboratory, USA

5:30-5:55 Compressive Illumination Waveforms for High Resolution Radar SensingEmre Ertin, Ohio State University, USA

6:00-6:25 A Butterfly Algorithm for SAR ImagingLaurent Demanet, Massachusetts Institute of

Technology, USA

Tuesday,May22

MS75The Beltrami Framework and its Applications4:30 PM-6:30 PMRoom:Concerto B - 3rd Floor

The Beltrami framework was introduced by Sochen, Kimmel and Malladi for image enhancement, 15 years ago. This is a pure geometric framework as images, shapes, vector fields are viewed as geodesics or harmonic maps in this approach. The Beltrami framework provides important properties, such as parametrization invariance and easy work on any Riemannian surfaces (e.g. brain cortical surface, multi-scale image, omni-directional image), so as to design efficient algorithms for geometric imaging problems. In this minisymposium, we will discuss recent advances of this framework and explore its potentialities towards a unifying variational method for geometric problems in imaging and related fields.

Organizer:XavierBressonCity University of Hong Kong, Hong Kong

Organizer:DominiqueZossoÉcole Polytechnique Fédérale de Lausanne, Switzerland

Organizer:Jean-PhilippeThiranÉcole Polytechnique Fédérale de Lausanne, Switzerland

4:30-4:55 Unifying Beltrami and Mumford-Shah FrameworksNir Sochen, Tel Aviv University, Israel

5:00-5:25 Invariant Beltrami-Operators - A Computational StudyRon Kimmel, Technion Israel Institute of

Technology, Israel

5:30-5:55 Diffusion Descriptors for 3D Shape AnalysisMichael M. Bronstein, Università della

Svizzera Italiana, Switzerland

6:00-6:25 Geodesic Active Fields - A Geometric Framework for Image RegistrationDominique Zosso, École Polytechnique

Fédérale de Lausanne, Switzerland; Xavier Bresson, City University of Hong Kong, Hong Kong; Jean-Philippe Thiran, École Polytechnique Fédérale de Lausanne, Switzerland


Tuesday,May22

CP10Contributed Session 104:30 PM-6:10 PMRoom:Minuet - 4th Floor

Chair: Mi Wang, University of Leeds, United Kingdom

4:30-4:45 A 3D Model-Based Inversion Algorithm for Electromagnetic DataMaokun Li, Aria Abubakar, and Tarek

Habashy, Schlumberger-Doll Research, USA

4:50-5:05 Uniqueness of the Circular Area Invariant for Graphs of Periodic FunctionsJeff Calder, University of Michigan, USA;

Selim Esedoglu, University of Michigan, Ann Arbor, USA

5:10-5:25 Non-Structural Imaging Methods for Characterisation of Nano-Particle SuspensionMi Wang, University of Leeds, United

Kingdom

5:30-5:45 Nonlocal Pdes-Based Morphology on Weighted Graphs for Image and Data ProcessingOlivier Lezoray and Elmoataz Abderrahim,

Université de Caen, France; Vinh Thong Ta, Universite de Bordeaux I, France

5:50-6:05 Tug of War game and Nonlocal Infinity Laplacian Equation with Application in Image Processing and Machine LearningAbderrahim Elmoataz and Olivier Lezoray,

Université de Caen, France

Tuesday,May22

CP9Contributed Session 94:30 PM-6:30 PMRoom:Maestro B - 4th Floor

Chair: Peter Blomgren, San Diego State University, USA

4:30-4:45 Introduction of Non-Linear Elasticity Models for Characterization of Shape and Deformation Statistics: Application to Contractility Assessment of Isolated Adult CardiocytesPeter Blomgren, Carlos Bazan, Trevor

Hawkins, David Torres, and Paul Paolini, San Diego State University, USA

4:50-5:05 Dynamic Multi-Channel MRI Reconstruction Via Low n-Rank Tensor PursuitJoshua D. Trzasko and Armando Manduca,

Mayo Clinic, USA

5:10-5:25 Magnetic Resonance Elastography of Brain in TbiCorina Drapaca, Brian Johnson and Thomas

Neuberger, Pennsylvania State University, USA

5:30-5:45 Computational Methods for Inverting the Soft X-Ray TransformJoanna Klukowska and Gabor T. Herman,

City University of New York, USA

5:50-6:05 Inverse Problems for Helical Objects: A Forward ModelQiu Wang and Peter C. Doerschuk, Cornell

University, USA

6:10-6:25 Fast Algorithm for Large-scale Linear Inversion with an Application in CO2 MonitoringSivaram Ambikasaran, Stanford University,

USA

Tuesday,May22

CP8Contributed Session 84:30 PM-6:10 PMRoom:Assembly C - 5th Floor

Chair: To Be Determined

4:30-4:45 Regularized Nonlinear Kaczmarz’ Algorithm Applied on Inverse Scattering for the Full Maxwell EquationsAref Lakhal, University of Saarland, Germany

4:50-5:05 Local Regularization for Inverse Problems in ImagingAndreas Langer, University of Graz,

Austria; Michael Hintermueller, Humboldt University Berlin, Germany

5:10-5:25 Noise-Robust Forward/Backward Compressive Music in Joint Sparse Recovery Problems Using Subspace Fitting CriterionJong Min Kim, Ok Kyun Lee, and Jong Chul

Ye, KAIST, Korea

5:30-5:45 Tracking Shape Deformations in Images: A Statistical Geometric ApproachValentina Staneva and Laurent Younes, Johns

Hopkins University, USA

5:50-6:05 Total Variation Based Multi-scale Analysis with Application on Fundus Image AnalysisYan Wang and Triet Le, University of

Pennsylvania, USA


Tuesday,May22


Chair: Ning Hao, Tufts University, USA

4:30-4:45 New Nonnegative Tensor Factorizations (NTF) with Anderson AccelerationNing Hao and Misha E. Kilmer, Tufts

University, USA

4:50-5:05 Discrete Composite Wavelet TransformsGlenn Easley, System Planning Corporation,

USA; Demetrio Labate, University of Houston, USA; Vishal Patel, University of Maryland, USA

5:10-5:25 Anisotropic Triangulations for Image ApproximationsLaurent Demaret, Helmholtz Zentrum

München, Germany

5:30-5:45 Analysis and Comparative Evaluation of Methods for Phase RetrievalFigen S. Oktem and Richard E. Blahut,

University of Illinois at Urbana-Champaign, USA

5:50-6:05 Seismogram Registration and Application to Nonlinear Seismic InversionHyoungsu Baek and Laurent Demanet,

Massachusetts Institute of Technology, USA

6:10-6:25 Scattered Light in the Lunar Reconnaissance Orbiter Wide Angled Camera Images – A Preliminary StudyPrasun Mahanti, Mark Robinson, and David

Humm, LROC Science Operations Center, USA


Notes


IS12 Abstracts

Abstracts are printed as submitted by the author.

50 IS12 Abstracts

IP1Magnetic Resonance Elastography

Many disease processes cause profound changes in the me-chanical properties of tissues, yet none of the conventionalmedical imaging techniques such as CT, MRI, and ultra-sound are capable of quantitatively delineating these prop-erties. Magnetic Resonance Elastography (MRE) is anemerging diagnostic imaging technology that employs anovel MRI-based technique to visualize propagating acous-tic shear waves in the body. Inversion algorithms are usedto process these wave images to generate quantitative mapsof tissue mechanical properties such as stiffness, viscosity,and anisotropy. The most important clinical applicationof MRE currently to evaluate chronic liver disease, whereit can eliminate the need for an invasive biopsy. However,there are many other potential applications. This presen-tation will review the rationale and physical basis of MREand the basic approaches used for data analysis. Areas ofopportunity will be identified where more advanced inver-sion algorithms could create new applications or substan-tially contribute to the clinical value of the technology.

Richard L. EhmanDepartment of RadiologyMayo [email protected]

IP2Targeted Seismic Imaging

The seismic imaging problem, which is typically formulatedas an inverse problem for the wavespeed in the acousticwave equation, involves the estimation of both the smoothand oscillatory part of this wavespeed. For both of theseproblems, a large data set is typically recorded and sub-jected to extensive processing to estimate both parts ofthe wavespeed throughout the entire region the data aresensitive to. For many applications, however, only a partof this image is of interest. I will describe techniques for en-hancing part of the image through appropriate data choicesand regularization.

Alison [email protected]

IP3Three Dimensional Structure Determination ofMacromolecules by Cryo-electron Microscopy fromthe Applied Math Perspective

Cryo-electron microscopy is a technique by which biologicalmacromolecules are imaged in an electron microscope. Themolecules are rapidly frozen in a thin layer of vitreous ice,trapping them in a nearly-physiological state. Cryo-EMimages, however, have very low contrast, due to the absenceof heavy-metal stains or other contrast enhancements, andhave very high noise due to the small electron doses thatcan be applied to the specimen. Thus, to obtain a reliablethree-dimensional density map of a macromolecule, the in-formation from thousands of images of identical moleculesmust be combined. When the molecules are arrayed ina crystal, the necessary signal-averaging of noisy images isstraightforwardly performed. More challenging is the prob-lem of single-particle reconstruction (SPR) where a three-dimensional density map is to be obtained from images ofindividual molecules present in random positions and ori-entations in the ice layer. Because it does not require the

formation of crystalline arrays of macromolecules, SPR isa very powerful and general technique, which has been suc-cessfully used for 3D structure determination of many pro-tein molecules and large complexes. I will introduce themain theoretical and computational (rather than experi-mental) challenges posed by SPR, and will focus on our(*) recent progress for obtaining three-dimensional struc-tures, for image classification and for denoising. From themathematical perspective, we combine ideas from differentareas, such as, spectral graph theory, dimensionality reduc-tion, representation theory and semidefinite programming.This conference also hosts a mini-symposium on cryo-EM.(*) This is a joint project with Fred Sigworth (Yale), YoelShkolnisky (Tel Aviv), Ronny Hadani (UT Austin), andShamgar Gurevich (Wisconsin-Madison).

Amit SingerPrinceton [email protected]

IP4Divide and Conquer: Using Spectral Methods onPartitioned Images

Spectral filtering methods such as Tikhonov regularizationare powerful tools in deblurring images, but their usefulnessis limited by two factors: the expense of the computationand the lack of flexibility in the filter function. This talkfocuses on overcoming these limitations by partitioning thedata in order to better process each segment. The datadivision can take several forms: for example, partitioningby spectral frequency, by input channel, or by subimages.Benefits of a divide and conquer approach to imaging caninclude smaller error, better conditioning of subproblems,and faster computations. We give examples using variousdata partitionings.

Dianne P. O’LearyUniversity of Maryland, College ParkDepartment of Computer [email protected]

Julianne ChungUniversity of Texas at [email protected]

Glenn EasleySystems Planning [email protected]

IP5Design in Imaging From Compressive to Compre-hensive Sensing

In the quest for improving imaging fidelity, vast attentionhas been devoted to effective solution of ill-posed problemsunder various regularization configurations. Nevertheless,issues such as determining optimal configurations for dataacquisition or more generally any other controllable pa-rameters of the apparatus and process are frequently over-looked. While design for well-posed problems has been ex-tensively studied in the past, insufficient consideration hasbeen given to its ill-posed counterpart. This is markedlyin contrast to the fact that many real-life problems are ofsuch nature. In this talk I shall describe some of the intrin-sic difficulties associated with design for ill-posed inverseproblems, lay out a coherent formulation to address themand finally demonstrate the importance of such designs for

IS12 Abstracts 51

various imaging problems.

Lior HoreshBusiness Analytics and Mathematical SciencesIBM TJ Watson Research [email protected]

CP1Sparse Shape Reconstruction

Given a collection (dictionary) of shapes, our problem ofinterest is choosing the right elements and “appropriately’combining them together to form a desired shape. A para-metric level set approach is used to express an arbitraryshape as an interaction of various “sub-shapes’. The appro-priate elements are chosen based on a sparsity constraintto use the least number of sub-shapes in the dictionary.The validity of the technique is tested through various ex-amples.

Alireza Aghasi, Eric MillerTufts [email protected], [email protected]

Justin RombergSchool of ECEGeorgia [email protected]

CP1A Compressive Sensing Synthetic Aperture RadarMethod Incorporating Speckle

Many techniques involving compressive sensing for Syn-thetic Aperture Radar (SAR) have been suggested, yetnone of these adequately deal with the fact that SAR im-ages contain a multiplicative component known as speckle.Speckle makes the compressibility and interpretation ofSAR images difficult. In this work, we investigate a morecomplete formulation of the problem of reconstructing thepiecewise smooth and texture component (speckle) directlyfrom a set of compressed measurements.


Vishal PatelUniversity of [email protected]

Rama ChellappaUniversity of Maryland, College [email protected]

CP1Adapted Curvelet Sparse Regularization in Lim-ited Angle Tomography

We investigate the reconstruction problem of limited angletomography which is known to be severely ill-posed. Tostabilize the reconstruction we present an adapted curveletsparse regularization method. This method is based on thecharacterization of curvelet coefficients, that can be stablyreconstructed from a limited angular range data. We showthat the dimension of the problem can be significantly re-duced in the curvelet domain, leading to a considerablespeedup of the reconstruction algorithm. In numerical ex-

periments, we show the practical relevance of these results.

Jurgen FrikelInstitute of Biomathematics and BiometryHelmholtz Zentrum [email protected]

CP1Robust High Frequency Information Guided Com-pressive Sensing Reconstruction

In compressive sensing, accurate image reconstruction fromvery few noisy data is challenging. We extract high fre-quency sparse features directly from measurements and usethem as a priori to guide further reconstruction. Variouscontinuation strategies and non-negative Garotte shrinkageare applied to enhance the performance and the iterativeapproach suppresses the high frequency errors caused by in-sufficient data. The proposed model recovers images withrich geometrical information both efficiently and robustlyusing less data.

Weihong GuoDepartment of Mathematics, Case Western [email protected]

Jing QinCase Western Reserve [email protected]

CP1Learning Sparsifying Transforms for Signal and Im-age Processing

We propose novel problem formulations and algorithmsfor learning sparsifying transforms from data. The algo-rithms are insensitive to initialization, and produce well-conditoned transforms that have much lower sparsificationerrors than analytical transforms such as wavelets. Thesetransforms can also be used to construct equivalent syn-thesis dictionaries that provide similar/lower data fittingerrors than the popular K-SVD algorithm, at 1/10th ofthe computational cost. Applied to image denoising thenew sparsifying transforms perform better than overcom-plete K-SVD dictionaries.

Saiprasad Ravishankar, Yoram BreslerDepartment of Electrical and Computer Engineering andtheCoordinated Science Laboratory, University of [email protected], [email protected]

CP1Image Compression With Nearest-Neighbor Trans-forms

We propose a graph-based algorithm to image modelingand compression. Our approach constructs a nearest-neighbor graph representation of images by associating pix-els with nodes and determining optimal edge weights fromthe structure of the images. The constructed graph thenleads to the identification of an orthonormal basis for theexpansion of the images and efficient image compression.

Aliaksei SandryhailaCarnegie Mellon [email protected]

52 IS12 Abstracts

Jose M.F. MouraDept. of Electrical and Computer EngineeringCarnegie Mellon [email protected]

CP2Automatic MRI Scan Positioning for Feeding Ar-teries of the Brain

Blood supply is very important to the daily function ofthe brain. It is possible to monitor the cerebral blood flow(CBF) noninvasively using phase-contrast (PC) MRI. ThePC scans need to be positioned perpendicular to feedingarteries at the entry point of skull. Given large variabilityin arterial trajectory across individuals, this task presentsa major challenge to the operator and contributes signif-icantly to measurement noise. In this project, we devel-oped an automatic MRI scan positioning method for feed-ing arteries of the brain. The four major feeding arterieswere first segmented out from the image and then the ax-ial representations of the vertebral arteries were extracted.Optimal positioning for further MRI scans were then de-termined based on the geometric properties of the arteries.We compared the results to those obtained with manualpositioning by an experienced operator. The findings sug-gest that the automatic PC positioning algorithm providesmeasures that are at least as accurate and precise as anexperienced operator.

Yan CaoThe University of Texas at DallasDepartment of Mathematical [email protected]

Peiying Liu, Hanzhang LuUT Southwestern Medical [email protected],[email protected]

CP2Reconstruction from Spherical Mean Data bySummability Methods

The reconstruction from spherical mean data plays an im-portant role in thermoacoustic and photoacoustic tomog-raphy. We consider a modification of a summability typeapproximative reconstruction. Among the consequences ofthis approach we present a inversion formula of the spher-ical mean Radon transform.

Frank FilbirInstitute of Biomathematics and BiometryHelmholtz Center [email protected]

CP23D Isar Image Reconstruction of Targets ThroughFiltered Back Projection

In this talk, an inversion scheme for near-field inverse syn-thetic aperture radar (ISAR) data is derived for both twoand three dimensions from a scalar wave equation model.The proposed data inversion scheme motivates the use ofa filtered back projection (FBP) imaging algorithm. Thepaper provides a derivation of the the general imaging fil-ter needed for FBP, which will be shown to reduce to a

familiar result for near-field ISAR imaging.

Zhijun QiaoDepartment of [email protected]

Jaime Lopez, Tim [email protected], [email protected]

CP2Restricted Convergence of Projected LandweberIteration in Banach Spaces for Nonlinear InverseProblems

We consider a class of inverse problems defined by a nonlin-ear map from parameter or model functions to data. Theparameter functions and data are contained in certain Ba-nach spaces. We employ a projected Landweber iterationwith posterior stepsize to give explicitly approximated se-quences of the solution to the inverse problem. A restrictedconvergence result is established given a Lipschitz type sta-bility condition on a closed, convex subset of the parameterfunction space.

Lingyun Qiu, Maarten V. de HoopPurdue [email protected], [email protected]

Otmar ScherzerComputational Science CenterUniversity [email protected]

CP2Attenuation Compensation in Ultrasound Imaging

Ultrasound B-scan exhibits shadowing and enhancementartifacts due to acoustic wave propagation and spatiallyvarying attenuation across tissue layers. Estimation of lo-cal attenuation coefficients is important for clinical diagno-sis and analysis. We present the mathematical frameworkof a novel joint estimation method for attenuation com-pensation and artifact reduction in pulse-echo ultrasound.Spatial resolution and speckle patterns are retained. Ourresults give higher quality attenuation compensation com-pared to several existing techniques using B-mode or RFimages.

Jue WangUnion [email protected]

Yongjian YuInfiMed [email protected]

CP2Actin Filament Tracking in Electron Tomograms ofnegatively stained Lamellipodia using the LocalizedRadon Transform

The aim of this work was to develop a protocol for auto-mated tracking of actin filaments in electron tomograms oflamellipodia embedded in negative stain. We show that alocalized version of the Radon transform for the detectionof filament directions enables three-dimensional visualiza-tions of filament network architecture, facilitating extrac-

IS12 Abstracts 53

tion of statistical information including orientation profiles.We discuss the requirements for parameter selection set bythe raw image data in the context of other, similar trackingprotocols.

Christoph WinklerJohann Radon Institute for Computational and AppliedMathematics, Austrian Academy of [email protected]

Marlene Vinzenz, Victor SmallInstitute of Molecular BiotechnologyAustrian Academy of [email protected],[email protected]

Christian SchmeiserUniversity of [email protected]

CP3A New Variational Model for Removal of Both Ad-ditive and Multiplicative Noise and a Fast Algo-rithm for Its Numerical Approximation

Variational image denoising models for additive and mul-tiplicative noise removal from digital images are rarely en-countered in the literature. To this end, this paper pro-poses a new variational model and its fast numerical algo-rithm. Numerical tests using both synthetic and realisticimages not only confirm that our new model delivers qual-ity results but also that the proposed numerical algorithmallows a very fast and accurate numerical solution of theproposed variational model.

Noppadol ChumchobCentre for Mathematical Imaging Techniques andDepartment of Mathematical Sciences, University [email protected]

CP3An Adaptive Spectrally Weighted Structure Ten-sor for Hyperspectral Imagery Based on the HeatOperator

A Tensor Nonlinear Anisotropic Diffusion (TAND) with aweighted Structure Tensor (ST) for Hyperspectral Images(HSI) is proposed. The weights are based on the heat op-erator. This is possible since neighboring spectral bandsin HSI are highly correlated, as are the bands of its gradi-ent. Comparisons with TAND using the classical ST showthat the heat weighting helps TAND to better discriminateedges. Experiments with HSI of different spatial character-istics are presented.

Maider J. Marin-Mcgee, Miguel Velez-ReyesLaboratory for Applied Remote Sensing and ImageProcessingUniversity of Puerto Rico at [email protected], [email protected]

CP3An 1 Minimization Algorithm with Applicationsin Image Processing

In this work, we consider a homotopic principle for solvinglarge scale 1 unconstrained minimization problems. Our

approach consists in solving a sequence of relaxed uncon-strained problems depending on a positive regularizationparameter that converges to zero. The optimality condi-tions of each sub-problem are characterized through a fixedpoint equation, where a preconditioned conjugate gradientalgorithm is applied to solve a sequence of resulting linearsystems. Numerical experiments are conducted showingthat our algorithm compare favorably with state-of-the-art solvers. Moreover, we present a set of image processingapplications including image denoising, image deblurring,image separation and image inpainting.

Carlos A. RamirezThe University of Texas at El [email protected]

Miguel ArgaezUniversity of Texas at El [email protected]

CP3Iteratively Reweighted Least Squares for L1 Prob-lems: Boring But Still Effective

Split Bregman (SB) or Alternating Direction Method ofMultipliers (ADMM) methods have become very popularfor 1 norm optimization problems, including Total Varia-tion and Basis Pursuit Denoising. It appears to be gen-erally assumed that they deliver much better computa-tional performance than older methods such as IterativelyReweighted Least Squares (IRLS). We show, however, thatIRLS type methods are computationally competitive withSB/ADMM methods for a variety of problems, and in somecases outperform them.

Paul RodriguezPontifical Catholic University of [email protected]

Brendt WohlbergLos Alamos National [email protected]

CP3Corrected Diffusion Approximation in Layered Tis-sues

I will present the corrected diffusion approximation for athin beam incident on a layered medium. This model im-proves upon the diffusion approximation for small sourcedetector separation distances without the long computa-tion times of the radiative transport equation. It also pro-vides a means to study oblique incidence and compute thefull spatial and angular dependence of the radiance.

Shelley B. RohdeUC [email protected]

Arnold D. KimUniversity of California, [email protected]

CP3Solving Highly Ill-Conditioned Blind Deconvolu-tion Problems

Some blind deconvolution problems have high-quality an-

54 IS12 Abstracts

swers, but cannot be solved efficiently by popular optimiza-tion techniques such as alternating minimization, quasi-Newton, and conjugate gradients. We pose scientificallyimportant model problems whose highly ill-conditionedHessians grind these methods to a standstill. A new Schurcomplement technique to efficiently invert these Hessians isproposed, and we use this technique to power a large-scaletrust region Newton method that can quickly solve highlyill-conditioned problems.

Paul ShearerUniversity of [email protected]

Anna GilbertDepartment of MathematicsUniversity of [email protected]

CP4Volterra Equations and Image Filtering and Pro-cessing

Fractional time derivatives allow to consider linear mod-els intermediates between pure difussive (heat equation)and conservative (wave equation) models. These modelsenjoy intermediate properties which lead to formulate lin-ear Volterra equations based models which can be under-stood as a generalization of classical fractional derivatives.The main contribution of our work is the model we pro-pose allows to hadle the difussion by means of viscosity pa-rameters in the context well posed linear PDEs. To makethe high quality of results clear, practical experiments areshown.

Eduardo CuestaUniversity of [email protected]

Mokhtar Kirane, Salman MalikUniversite de La [email protected], [email protected]

CP4An Algorithm for Shape Detection in Inverse Prob-lems

In this work, we propose a computational method to recon-struct shapes, i.e. regions or boundaries, in inverse prob-lems. Our method is general in the sense that it can be usedin conjunction with different inverse problems. However, itis designed to work for approximately piecewise constantfunctions. The method is based on a shape optimizationformulation of the inverse problem. In contrast with re-cent work in this area, the geometry is modeled explicitlyas a set of polygonal curves, not as a level set function.This approach allows flexibility in modeling and geometricregularization, and also results in computational efficiency,which will be illustrated with several examples.

Gunay DoganTheiss ResearchNational Institute of Standards and [email protected]

CP4Image Recognition Via Moment Method

We will introduce the ”shape dictionary”, which is settedup by extracting a set of moment features from the mo-ments we define. The shape features can help us recogniz-ing letters and images.

Shanshan Huang, Mireille BoutinSchool of ECEPurdue [email protected], [email protected]

CP4Segmentation Based on a Limiting Mumford andShah Functional

We propose an energy minimization model for segmenta-tion based on a limiting of the Mumford and Shah func-tional. Our proposed model is only depended on the edgeset and is formulated using the level set framework withan extension for free endpoints/crack tips (open edges).The model is able to detect both open and closed edges.The algorithm is tested on both synthetic and real images,converging within a few iterations in some examples.

Hayden SchaefferUCLA Math [email protected]

Luminita A. VeseUniversity of California, Los AngelesDepartment of [email protected]

CP4Reconstruction of Images: Computation of Confi-dence Intervals

Computation of confidence intervals that bound the pixelvalues of an image is an important but not trivial prob-lem. We use methods that reduce the computational costfor large images and parallelize them to reduce the run-ning time. We experiment with different divisions of theimage into sub-images and several boundary conditions.Ways to better use prior information for the boundary ofthe sub-images to improve the resulting intervals are alsoconsidered.

Viktoria TaroudakiGraduate Student,AMSC Graduate Program,University of [email protected]


CP4Restoration of Images Corrupted by Impulse NoiseUsing Blind Inpainting and 0 Norm

We study the problem of image restoration of observed im-ages corrupted by impulse noise and other types of noise(e.g. zero-mean Gaussian white noise). Since the pixelsdamaged by impulse noise contain no information aboutthe true image, these damaged pixels can also be consid-

IS12 Abstracts 55

ered as missing information. If the pixels corrupted byimpulse noise are known, then the image restoration prob-lem becomes a standard image inpainting problem. How-ever, the set of damaged pixels is usually unknown, thushow to find this set correctly is a very important problem.We proposed a method that can simultaneously find thedamaged pixels and restore the image. This method canalso be applied to situations where the damaged pixels arenot randomly chosen, but follow some unknown procedure.By iteratively restoring the image and updating the set ofdamaged pixels, this method has better performance thanother methods, as shown in the experiments.

Ming YanUniversity of California, Los AngelesDepartment of [email protected]

CP5Seismic Full-Waveform Inversion Using Sourcesand Receivers Compression Scheme

We present a source-receiver compression approach for re-ducing the computational time and memory usage of theseismic full-waveform inversions. By detecting and quan-tifying the extent of redundancy in the data, we assemblea reduced set of simultaneous sources and receivers thatare weighted sums of the physical sources and receiversemployed in the survey. Because the number of these si-multaneous sources and receivers can be significantly lessthan those of the physical sources and receivers, the com-putational time and memory usage of an inversion methodcan be tremendously reduced. The scheme is based ondecomposing the data into their principal components us-ing a singular-value decomposition approach and the datareduction is done through the elimination of the smalleigenvalues. Consequently this will suppress the effect ofnoise in the data. Moreover, taking advantage of the re-dundancy in the data, this compression scheme effectivelystacks the redundant data resulting in an increased signal-to-noise ratio. We present inversion results for well-knowntwo-dimensional and three-dimensional models for surfaceseismic measurements. We show that this approach hasthe potential of significantly reducing both computationaltime and memory usage of the Gauss-Newton method by1-2 orders of magnitude.

Aria AbubakarSchlumberger Doll [email protected]

Tarek Habashy, Guangdong PanSchlumberger-Doll [email protected], [email protected]

CP5Reservoir Model Improvement by a Better Integra-tion of 4D Seismic Data Using Image ProcessingMethods

Calibrating reservoir models to flow data involves historymatching processes. Model parameters are iteratively ad-justed to minimize the misfit between the real data and thecorresponding simulated responses. The current formula-tion used to quantify the seismic data mismatch is neitherrepresentative of the difference between two images, nor ofmatching quality. We describe an alternative formulation,using filtering, segmentation and image matching meth-ods, to extract relevant information from seismic images

and compute their dissimilarity.

Ratiba DerfoulIFP Energies nouvellesLaboratoire de Mathematiques de l’INSA, INSA [email protected]

Sebastien Da VeigaIFP Energies [email protected]

Christian GoutINSA RouenLab. de Mathematiques de l’[email protected]

Carole Le GuyaderINSA Rouen, [email protected]

CP5An Inverse Problem Approach to High DynamicRange Photography

We describe an inverse problem for determining the re-sponse function of a camera system from multiple pho-tographs of a scene captured with different exposure times.The irradiance map (amount of light entering the inputof the camera) of a natural scene spans up to 8–10 or-ders of magnitude but typical digital cameras record onlytwo or three orders of magnitude. Recovering the responsefunction of the camera allows one to combine low dynamicrange images into a single image with the original dynamicrange. We present preliminary results and analysis.

Thomas HoftTufts [email protected]

CP5A One Dimensional Algorithm for Seismic Imag-ing and Inversion: Theoretical Development andNumerical Tests

We present an inverse scattering method for geophysicalimaging and amplitude correction from measured data. Noknowledge about the medium under investigation is as-sumed. Analytic and numerical one dimensional examplesshow excellent results in finding both the location of inter-faces and the amplitude of acoustic reflections. Our testsinclude different number of layers, high/low contrasts, ve-locity inversions and noisy data.

Bogdan G. NitaMontclair State [email protected]

Ashley CieslaBrookedale Community [email protected]

Christopher SmithThe College of New [email protected]

CP5Compressing Aerial Images Using a Critical Rep-

56 IS12 Abstracts

resentation of Laplacian Pyramid

An extremely simple scheme for transforming a Laplacianpyramid (LP) to be non-redundant is introduced recently.The resulting process provides a critical multiresolutionrepresentation of data, and its analysis/synthesis opera-tors form a wavelet filter bank. In this talk, we brieflyreview the new scheme and then discuss the resulting sys-tem’s very fast, LP-like, decomposition/reconstruction al-gorithms. We also present that the new system shows con-sistently more favorable results than other existing systemsin compressing aerial images.

Fang Zheng, Youngmi HurJohns Hopkins [email protected], [email protected]

CP6Optimal Acquisition Design for Diffusion MRI: ACase of Searching for a Needle in the Universe

Diffusion MRI is a promising technique that allows sci-entists to investigate noninvasively the structural connec-tivity of the human brain. The focus of this work willbe on sparse and optimal acquisition design for diffusionMRI. We propose a novel optimality criterion for sparsemultiple-shell acquisition and quasi-multiple-shell designsin diffusion MRI and an effective semi-stochastic and mod-erately greedy search strategy to locate the optimum de-sign configuration among more than 1027 to 10232 distinctconfigurations.

Cheng Guan KoaySection on Tissue Biophysics and Biomimetics, NICHDNational Institutes of [email protected]

Evren OzarslanNational Institutes of [email protected]

M. Elizabeth MeyerandDepartment of Medical PhysicsUniversity of [email protected]

CP6Simultaneous Image Reconstruction and Sensitiv-ity Map Estimation in Partially Parallel MR Imag-ing

We develop a variational model and numerical algorithmfor simultaneous sensitivity map estimation and image re-construction in partially parallel MR imaging with signif-icantly undersampled data. The objective functional in-cludes the data fidelity term and various regularizationson the underlying image and sensitivity maps. The al-gorithm decouples the problem into easier sub-problemsinvolving variable splitting techniques and treats the TVterm by PDHG scheme. Experimental results are pre-sented to demonstrate the effectiveness of the proposedmethod.

Meng LiuDepartment of MathematicsUniversity of [email protected]

Yuyuan Ouyang

Department of Mathematics, University of [email protected]

Xiaojing YeSchool of MathematicsGeorgia [email protected]

Yunmei ChenUniversity of [email protected]

Feng HuangInvivo Corporation, Philips HealthCare, Gainesville, [email protected]

CP6Precision Limits for Source Parameter Estimationin Slitless Spectrometry

Multi-order slitless imaging spectrometers have recentlybeen proposed for observing dynamic events in the solaratmosphere. We analyze the problem of estimating thesource parameters from such spectrometer measurements.We derive analytical lower bounds on the precision of es-timates, via the application of Cramer-Rao bounds. Wecompare the derived bounds with numerically simulatederrors. Furthermore, we explore the optimized instrumentrequirements including the resolution, signal-to-noise ratio,and measurement configuration.

Figen S. Oktem, Farzad KamalabadiUniversity of Illinois at [email protected], [email protected]

Joseph M. DavilaNASA Goddard Space Flight Center, [email protected]

CP6Bandwidth-Efficient Remote Visualization on Mo-bile Devices

Mobile devices, such as tablets, are well suited for mobileanalysis of numerical simulations, but traditional remotevisualization methods are not usable for high-latency low-bandwidth networks. We introduce a mathematical modelfor the visualization and interaction error based on the ca-pabilities of current mobiles devices. We then derive theo-retic upper and lower error-bounds for minimal bandwidth-consumption independent of 3D scene size and complexity.Finally, we propose a bandwidth-optimized remote visual-ization algorithm satisfying the error-bounds.

Sebastian RitterbuschKarlsruhe Institute of Technology (KIT)Engineering Mathematics and Computing Lab (EMCL)[email protected]

Vincent HeuvelineEngineering Mathematics and Computing Lab (EMCL)Karlsruhe Institute of Technology (KIT)[email protected]

Roman Reiner, Andreas Helfrich-SchkarbanenkoKarlsruhe Institute of Technology (KIT)Engineering Mathematics and Computing Lab (EMCL)

IS12 Abstracts 57

[email protected],[email protected]

CP6Sparse Reconstruction Applied to TomographicSAR Inversion

The highly anisotropic 3D resolution element of modernspaceborne multi-baseline SAR systems allow us to exploitsparsity for superresolution tomographic SAR inversion.This is particularly true for mapping of urban infrastruc-ture. The proposed algorithm SL1MMER is shown to beefficient. Its superresolution power is derived by simula-tions and demonstrated with real TerraSAR-X data. 3Dreconstructions of buildings are presented including theirseasonal thermal deformations at resolution and coveragelevels not possible so far.

Xiao Xiang ZhuGerman Aerospace Center (DLR)Remote Sensing Technology [email protected]

Richard BamlerGerman Aerospace Center (DLR)Remote Sensing Technology Institute (IMF)[email protected]

CP7Mass Transport Problem and Its Application inPoint Set Matching Problems

We study one point-set matching method based on Monge-Kantorvich mass transport problems and its application inmatching two sets of pulmonary vascular tree branch pointswhose displacement is caused by the lung volume changes.Perfect match performances verdict the effectiveness. Weprovide theoretical analysis to explain these results: thecurl-cardinality relation. It says that given point cardinal-ity n, when the maximum curl does not exceed C/n, perfectmatches can be obtained. (joint work with Ching-Long Linand I-Liang Chern).

Pengwen ChenNational Taiwan [email protected]

CP7An Optimal Approach for Image Deblurring andDenoising with Data Acquisition Errors

Data acquisition errors due to dead pixels or other hard-ware defects can cause undesirable artifacts in imaging ap-plications. Compensating for these defects typically re-quires knowledge such as a defective pixel map, which canbe difficult or costly to obtain and which is not necessarilystatic. However, recent calibration data is readily availablein many applications. In this talk, we discuss an optimalcombined approach for image deconvolution and denoising,which uses calibration data and minimizes the empiricalBayes risk. We show that our approach is able to recon-struct missing information better than standard filteringapproaches and is robust even in the presence of a largenumber of defects.


Matthias ChungDepartment of MathematicsTexas State [email protected]


CP7A Robust Approach to Detect Faces from Still Im-ages

Haar-like features have been widely adopted for frontal facedetection. However, the algorithm has challenges in de-tecting dark coloured faces or faces captured under poorillumination. To overcome these limitations, we present anew method of face detection in this paper. Experimentalresults from 9,883 positive images and 10,349 negative im-ages show a considerable improvement in face detection hitrates without a significant change in false acceptance ratescompared with the OpenCVs Haar detection algorithm.

Qinghan XiaoDefence R&D Canada - [email protected]

Patrick Laytner, Chrisford LingUniversity of [email protected], [email protected]

CP7A Variational Infinite Perimeter Model for ImageSelective Segmentation

In this talk we first present a dual level-set selective seg-mentation (DLSS) model with a combination of the edgedetection and geometric distance functions, which can dif-ferentiate two objects having similar or identical intensi-ties. The DLSS model uses a combination of two level-sets: a global level-set which segments all boundaries, anda local level-set which evolves and finds the boundary ofthe object closest to the geometric constraints (markers).For this model, a H∞ Hausdorff measure (extending theclassical notion of length) has been used as the penaliza-tion term. Then we show our new infinite perimeter duallevel-set selective segmentation (IPDLSS) model which canselect an object with oscillatory boundaries and corneringeffect. For the IPDLSS model, the incorporation of the L∈Lebesgue measure of the ε-neighborhood of the contour asthe penalization term is essential. Test results will showthat IPDLSS leads to better results in cornering and withoscillatory boundaries than DLSS.

Lavdie RadaUniversity of LiverpoolCentre for Mathematical Imaging Techniques [email protected]

Ke ChenUniversity of [email protected]

CP7Pde Transforms and Non-Smooth Data: Singular-

58 IS12 Abstracts

ity Detection, Segmentation and Denoising

The use of PDE Transforms for processing piecewisesmooth data potentially contaminated with noise is ex-plored. PDE transforms rely on coupled PDEs to decom-pose data into functional frequency modes useful for dualtime-frequency analysis and data processing. Modes gener-ated by the PDE transform provide a useful analysis toolfor recovering various types of edges on different scales.High order PDEs help in resolving singularities of differentdegrees of severity as well as modes with adjacent frequen-cies in continuous regions.

Rishu SaxenaArizona State [email protected]

Siyang YangMichigan State [email protected]

CP7Image Segmentation and Inpainting Techniques forReconstruction of Incomplete Cell Paths

One challenge of segmenting cells from fluorescence mi-croscopy is that cells frequently disappear, resulting in bro-ken cell paths in a 3D image volume. In this talk, wepresent a segmentation and an inpainting model that canreconstruct incomplete cell paths. The key idea of the seg-mentation model is to perform 2D segmentation in a 3Dframework. We will also discuss the use of inpainting tech-niques which have the advantage of not just obtaining thecell contours but also the cell interiors. Numerical resultsof fluorescent live cell images will be given.

Justin WanUniversity of WaterlooDepartment of Computer [email protected]

CP8Noise-Robust Forward/Backward CompressiveMusic in Joint Sparse Recovery Problems UsingSubspace Fitting Criterion

We study a multiple measurement vector (MMV) problemwhere multiple signals share common sparse support setand are sampled by a common sensing matrix. Althoughwe can expect that joint sparsity can improve the recoveryperformance over single measurement vector(SMV) prob-lem, compressive sensing (CS) algorithms for MMV exhibitperformance saturation as the number of multiple signalsincreases. Recently, to overcome this drawbacks of CS ap-proaches, our group proposed so-called compressive MU-SIC (CS-MUSIC) algorithm, which optimally combines CSwith MUSIC using a generalized MUSIC criterion so that itoutperforms all the conventional MMV algorithms. How-ever, the existing CS-MUSIC does not fully exploit addi-tional sampling more than necessary in terms of restrictedisometry property condition. Hence, in this talk, we in-troduce a novel subspace fitting criterion that extends thegeneralized MUSIC criterion to deal with such situation sothat it exhibits near-optimal behaviours. In addition, thesubspace fitting criterion leads us two alternative forms ofCS-MUSIC algorithm with forward and backward supportselection, which significantly improves its noise robustness.Finally, we will illustrate several practical applications ofthe forward/backward CS-MUSIC such as dynamic track-

ing of time-varying signals and etc.

Jong Min Kim, Ok Kyun Lee, Jong Chul [email protected], [email protected],[email protected]

CP8Regularized Nonlinear Kaczmarz’ Algorithm Ap-plied on Inverse Scattering for the Full MaxwellEquations

We consider the inverse scattering problem for the fulltime-harmonic Maxwell equations using near field measure-ments. Based on a rigorous analysis, we recast the inter-twined vector equations in the full Maxwell system intodecoupled scalar inverse problems. To reconstruct the con-trast function, we apply an adapted version of the Kacz-marz’ algorithm to non-linear problems with thoroughlycomputed starting value. We regularize using the approx-imate inverse, where we apply analytically pre-computedreconstruction kernels at each iteration step. Numericalexperiments in 3D using real data show the efficiency andstability of the algorithm.

Aref LakhalInstitute of Applied Mathematics,University of Saarland, [email protected]

CP8Local Regularization for Inverse Problems in Imag-ing

In applications in image processing the recovery of the un-known “ideal’ image is often an ill-posed inverse problem.Thus regularization techniques are required to solve thisproblem. In order to enhance image details while pre-serving homogenous regions, the regularization parametershould be based on local image features. Therefore wediscuss a model that utilizes a locally dependent regular-ization parameter for general inverse problems in imageprocessing and we show convincing numerical examples inimage restoration.

Andreas LangerJohann Radon Institute for Computational andand Applied Mathematics (RICAM)[email protected]

Michael HintermuellerHumboldt-University of [email protected]

CP8Tracking Shape Deformations in Images: A Statis-tical Geometric Approach

We describe a statistical framework for tracking deformableobjects in an image sequence under the constraint of pre-serving their topology. We represent the object’s boundaryby a manifold-valued random variable and we introduce astochastic model for its evolution which ensures the in-duced deformations are smooth and invertible. We esti-mate the shape of the object from each image frame byrestricting to low dimensional deformations which can beinterpreted as random steps along geodesic paths in sub-

IS12 Abstracts 59

Riemannian geometry.

Valentina StanevaJohns Hopkins [email protected]

Laurent YounesApplied Mathematics and [email protected]

CP8Total Variation Based Multi-scale Analysis withApplication on Fundus Image Analysis

Total variation minimization has been widely used in im-age analysis. Previous studies showed that this model isrelated to the local scale features in images. We will givea detailed comparison study of the local scale informationrevealed by different total variation models include TV-L2, TV-L1 and nonlinear diffusion. In addition, we willcombine the advantages of different models, and show anapplication on fundus images analysis for blood vessels anddrusen structures segmentation.

Yan WangUniversity of [email protected]

Triet LeDepartment of MathematicsUniversity of [email protected]

CP9Fast Algorithm for Large-scale Linear Inversionwith an Application in CO2 Monitoring

Large scale inverse modeling is being widely used in sub-surface problems, where direct measurements of parame-ters are expensive and sometimes impossible. Subsurfacemedia are inherently heterogeneous in many ways. Fur-thermore, techniques such as hydraulic tomography andelectric resistivity tomography allow the collection of moreindirect measurements, and at the same time, there is anincreased appreciation of the value of detailed characteriza-tion of the subsurface media. For instance in CO2 seques-tration projects, in-order to predict events like CO2 leak-age, it is imperative to model fine scale features. This re-sults in solving for a large number of unknowns. Employingconventional numerical algorithms to find these unknownsis computationally expensive. The need for fast, efficientlarge scale inversion techniques is hence of much practicalinterest. In the current work, we use a Bayesian formu-lation to solve a large-scale tomography problem with anapplication to real-time monitoring at a CO2 sequestrationsite. The Bayesian formulation has two significant steps tocompute the unknowns. The first step involves forminga linear system and solving the linear system. The sec-ond step involves finding the unknowns from the solutionobtained by solving the linear system. The fast novel algo-rithm, which exploits the underlying sparsity coupled withfast multipole method, accelerates both these steps. Us-ing our new algorithm, we estimate 100 million unknownsin real-time at every time instant. This fast new algo-rithm, based on sparsity and the fast multipole method,has an asymptotic complexity of O(

√) for the first step

and O() for the second step, where m is the number of un-

knowns. This provides an enormous scale up compared tothe conventional method which has a complexity of O(∈)for both steps. The new algorithm is fairly general and canbe extended to other similar tomography problems as well.

Sivaram AmbikasaranInstitute for Computational and MathematicalEngineeringStanford [email protected]

CP9Introduction of Non-Linear Elasticity Models forCharacterization of Shape and Deformation Statis-tics: Application to Contractility Assessment ofIsolated Adult Cardiocytes

We explore cardiocyte contractility assessment based onbiomechanical properties, energy conservation, and infor-mation content. Contraction is defined by the energy-distance between the shapes of the contracting cell, ob-tained by warping one shape into another. Relevant resultsare obtained by employing a non-linear features fidelity,and a regularization term. The responses in isolated adultrat cardiocytes are computed and contrasted with previ-ously established measures. Qualitative and quantitativeagreement for frequency, pacing, and overall behavior isexcellent.

Peter BlomgrenSan Diego State UniversityDepartment of Mathematics and [email protected]

Carlos BazanSan Diego State UniversityComputational Science Research [email protected]

Trevor HawkinsSan Diego State UniversityDepartment of Mathematics and [email protected]

David Torres, Paul PaoliniSan Diego State UniversityComputational Science Research [email protected], [email protected]

CP9Magnetic Resonance Elastography of Brain in Tbi

Magnetic Resonance Elastography (MRE) is a non-invasiveimaging technique that combines Magnetic ResonanceImaging and wave mechanics to estimate tissue stiffnessin vivo. Recent studies show that MRE could aid in thediagnosis of diseases and tumor detection. In particular,MRE estimated decreased brain stiffness in multiple scle-rosis, normal pressure hydrocephalus, and Alzheimers dis-ease. We will use MRE and brain pulsations to estimatebrain stiffness in TBI and possible correlations to loss ofbrain function.

Corina DrapacaDepartment of Engineering Science and MechanicsPennsylvania State [email protected]

60 IS12 Abstracts

Brian JohnsonDepartment of KinesiologyPennsylvania State [email protected]

Thomas NeubergerHuck Institutes - Life SciencesPennsylvania State [email protected]

CP9Computational Methods for Inverting the Soft X-Ray Transform

We present formulas for inverting the soft x-ray transform.This solves the inverse problem of reconstruction of 3Dstructures from 2D projections obtained by a soft x-raymicroscope. This problem arose from experimental cellu-lar biology. Previously-proposed approaches to solving itlacked mathematical rigour. We demonstrate, using biolog-ical data, why an exact mathematical inversion approachwill result in more efficacious reconstructions.This work issupported by NSF grant DMS-1114901.

Joanna Klukowska, Gabor T. HermanThe Graduate Center, [email protected], [email protected]

CP9Dynamic Multi-Channel MRI Reconstruction ViaLow n-Rank Tensor Pursuit

In this work, we propose a low n-rank tensor pursuit strat-egy for accelerated dynamic multi-channel MRI reconstruc-tion. We show that several contemporary low-rank matrixpursuit methods for MRI reconstruction arise as special in-stances of the proposed model, which is based on weightedtensor trace norm minimization. We generalize and com-pare these methods for dynamic multi-channel MRI re-construction, investigate hybridization strategies for simul-taneously exploiting redundancies in multiple dimensions,and demonstrate the different approaches for cardiac MRI.

Joshua D. TrzaskoMayo [email protected]

Armando ManducaCollege of MedicineMayo [email protected]

CP9Inverse Problems for Helical Objects: A ForwardModel

Many biological macromolecular complexes have helicalsymmetry. Cryo electron microscopy provides 2-D pro-jection images with low SNR and unknown projection di-rections. By focusing on the axial and rotational shiftsbetween adjacent motifs of the helix rather than on theperiod, number of motifs per period, and number of turnsper period, a forward model is described which avoids inte-ger programming and shortens computation time in inversecalculations. The method is demonstrated on Tobacco Mo-

saic Virus.

Qiu WangCornell [email protected]

Peter C. DoerschukCornell UniversityBiomedical Engineering and Electrical and [email protected]

CP10Uniqueness of the Circular Area Invariant forGraphs of Periodic Functions

The representation of curves by integral invariants has be-come an important step in shape recognition and classi-fication due primarily to their robustness with respect tonoise. However, for many integral invariants, the questionof uniqueness of representation has not been addressed. Inthis work, we study the case of the circular area invariantand show that the representation is unique for graphs ofperiodic functions and derive a stability estimate on thereconstruction map.

Jeff CalderDepartment of MathematicsUniversity of [email protected]

Selim EsedogluUniversity of [email protected]

CP10Tug of War game and Nonlocal Infinity LaplacianEquation with Application in Image Processing andMachine Learning

Certain PDEs such as infinity Laplacian Equation can beobtained as limits of values of tug-of-war games when theparameter that controls the length of the possible move-ments goes to zero. In this presentation we consider anonlocal version of the game on weighted Graphs, and wewill show that the value function of this game correspondto a nonlocal infinity Laplacian on Graphs. We will pro-vide applications for interpolation data on graphs in imageprocessing and machine learning.

Abderrahim ElmoatazUniversitee de Caen [email protected]

Olivier LezorayUniversite de Caen [email protected]

CP10Nonlocal Pdes-Based Morphology on WeightedGraphs for Image and Data Processing

Mathematical morphology (MM) offers a wide range ofoperators to address various image processing problems.These operators can be defined in terms of algebraic (dis-crete) sets or as partial differential equations (PDEs). Weintroduce a nonlocal PDEs-based morphological frameworkdefined on weighted graphs. This formulation introduces

IS12 Abstracts 61

nonlocal patch-based configurations for image processingand extends PDEs-based approach to the processing of ar-bitrary data such as nonuniform high dimensional data.

Olivier Lezoray, Elmoataz AbderrahimUniversite de Caen [email protected],[email protected]

Vinh Thong TaUniversite de [email protected]

CP10A 3D Model-Based Inversion Algorithm for Elec-tromagnetic Data

In this approach, the models are described by points in 3Dand the radial basis functions are used as the interpolationfunctions for connecting these points. This renders the sur-face of the target intrinsically smooth. The multiplicativeL2-norm and weighted L2-norm regularization schemes areemployed in the inversion to constrain the curvature and tofurther smooth the surface. The inversion results demon-strate that both the shapes and material properties of thetarget are well reconstructed.

Maokun LiSchlumberger-Doll [email protected]

Aria AbubakarSchlumberger Doll [email protected]

Tarek HabashySchlumberger-Doll [email protected]

CP10Non-Structural Imaging Methods for Characterisa-tion of Nano-Particle Suspension

Under an external alternating force on a colloid, e.g. pres-sure or electric field, a particle vibration potential will begenerated, which forms the secondary current flow in ad-dition to the ionic current flow. This paper addresses thelinkage between the elementary analyses of colloids in ki-netic model and the electric field solutions in static modelfor analysing nano-particle characteristics in a slab modeland in spatial distribution, providing potential methods fornon-structural imaging.

Mi WangUniversity of [email protected]

CP11Seismogram Registration and Application to Non-linear Seismic Inversion

Seismograms from seismic surveys contain subsurface ve-locity and structure information. We show that optimalmapping between a seismogram and corresponding syn-thetic seismogram can be obtained and effectively used forfull waveform inversion. Highly non-convex image registra-tion problems due to the oscillatory wavelets are solved ina multiscale manner. Once found, mappings warp the seis-

mogram so that the least square misfit optimization avoidslocal minima. Successful application to seismic inversionproblems are demonstrated.

Hyoungsu BaekMassachusetts Institute of TechnologyMathematics [email protected]

Laurent DemanetMathematics, [email protected]

CP11Anisotropic Triangulations for Image Approxima-tions

Adaptive triangulations offer a powerful framework forimage approximation. In this talk, we deal with edge-preserving representations: this requires the design ofanisotropic triangulations. We present a method basedon Delaunay triangulations: related approximation theory,algorithmic issues, application to compression and connec-tions with local metrics are discussed.

Laurent DemaretInstitute of Biomathematics and BiometryHelmholtz Zentrum Munchen, [email protected]

CP11Discrete Composite Wavelet Transforms

Composite wavelets provide a general framework for theconstruction of waveforms defined not only at various scalesand locations, as traditional wavelets, but also at variousorientations and with different scaling factors in each coor-dinate. In this work, we further investigate the construc-tions derived from this approach to develop critically sam-pled and nonsubsampled composite wavelets for the pur-pose of image coding and denoising. The demonstrationswill show consistent improvements upon competing state-of-the-art methods.


Demetrio LabateUniversity of [email protected]


CP11New Nonnegative Tensor Factorizations (NTF)with Anderson Acceleration

Traditionally, NTF represents a tensor as a sum of nonneg-ative outer products of vectors. Here we explore a noveltensor representation and approximate a nonnegative ten-sor as a sum of structured tensors named PCTs. Thismethod allows the optimization to be achieved through a2-way alternation instead of 3-way. Additionally, we adoptthe Anderson acceleration during the iteration. As an ex-ample, we test the NMF, traditional NTF and our novel

62 IS12 Abstracts

NTF on the CBCL dataset.

Ning HaoTufts UniversityDepartment of [email protected]

Misha E. KilmerTufts [email protected]

CP11Scattered Light in the Lunar Reconnaissance Or-biter Wide Angled Camera Images A PreliminaryStudy

The Wide Angle Camera (WAC) onboard the Lunar Re-connaissance Orbiter spacecraft is mapping the completeMoon nearly every month with images at a scale of 100meters/pixel in seven color bands. The WAC has largefield-of-view and this makes the analysis of the contribu-tion of wavelength dependant scattered light towards im-aged scene brightness extremely important to future quan-titative studies. A preliminary analysis is performed in thiswork and results and future directions are discussed.

Prasun Mahanti, Mark Robinson, David HummLROC Science Operations [email protected], [email protected], [email protected]

CP11Analysis and Comparative Evaluation of Methodsfor Phase Retrieval

Phase-retrieval is the recovery of signals from Fouriermagnitude with wide applications in imaging. To datethere are not many comparative studies of phase-retrievalalgorithms. We analyze and evaluate Schulz-Snyderand Fienup algorithms, the only general-purpose phase-retrieval algorithms, both theoretically and numerically.We review their recent re-interpretations and perform anexperimental evaluation in terms of reliability, convergenceproperties, and noise sensitivity. Additional comparisonis done against algorithms for blind-deconvolution, whichphase-retrieval is special case of.

Figen S. Oktem, Richard E. BlahutUniversity of Illinois at [email protected], [email protected]

MS1Applications of Tomographical Images in CancerRadiotherapy and Cone Beam CT Image Enhance-ment Using Prior Nonlocal Means Method

Cone beam computed tomography (CBCT) plays an im-portant role in for cancer radiotherapy. In this talk, I willfirst discuss the applications of CBCT in radiotherapy andpresent some challenging yet clinically important problems,including reconstruction and enhancement, CT/CBCT de-formable registration, and segmentation. I will also presentour recent work on CBCT image enhancement where a pa-tient CT is utilized to enhance the CBCT quality using aprior non-local means method.

Xun JiaDepartment of Radiation OncologyUniversity of California San Diego

[email protected]

MS1Trabecular Texture Analysis in Dental Cone BeamCT

It is known that variation of trabecular texture patternsinside oral cavity is correlated to bone density loss, whichcan leads serious diseases such as osteoporosis. Automaticanalysis of such texture patterns provides a potential so-lution toward osteoporosis screening, and at a low costsince the data can be obtained during routine dental ex-aminations. We evaluate several modern texture descrip-tors on trabecular texture in three dimensional dental conebeam computational tomography (CBCT). The study isperformed on a dataset containing CBCT volumes from96 subjects, which are further divided into four gender-age subgroups. For each volume, eight regions-of-interest(ROIs) containing trabecular bone structures are croppedby a dentist. After that, five types of texture descrip-tors, including the icosahedron HOG, the standard frac-tal dimension (FD), the multi-fractal spectrum (MFS), themean intensity, and the intensity histogram are extractedfrom these ROIs. Statistical analysis are then performedon all these features to determine whether they are signif-icantly different between different age-gender groups. Theresults show that many features are significantly correlatedwith the age-gender separation with p-value<0.05. Fur-thermore, we feed the features with low p-values into thesupport vector machine and have observed significant im-provement in classification accuracy.

Haibin LingDept. of Computer and Information Sciences, [email protected]

Xiong Yang, Jie Yang, Fangfang XieTemple [email protected], [email protected],[email protected]

Yong XuSouth China University of [email protected]

Vasileios MegalooikonomouTemple [email protected]

MS1Joint CT/CBCT Deformable Registration andCone Beam CT Image Enhancement

In this paper, we propose an algorithm that simultaneouslyperforms the registration of CT and conebeam CT (CBCT)and image enhancement of CBCT. In particular, we adapta viscous fluid model which naturally incorporates two in-gredients, a similarity measure for registration and an in-tensity correction term for image enhancement. By addingthis term back to CBCT, we can obtain an intensity cor-rected CBCT with better image quality. Furthermore, wefind that including such intensity correction term can inturn help to improve the registration results. To reach theclinical goal of minimal processing time, the algorithm isimplemented on a graphic processing unit (GPU) platform.The robustness of our algorithm is illustrated using two pa-

IS12 Abstracts 63

tient datasets of head-and-neck and pelvis respectively

Yifei LouSchool of Electrical and Computer EngineeringGeorgia Institute of [email protected]

Xun JiaDepartment of Radiation OncologyUniversity of California San [email protected]

Tianye NiuSchool of Mechanical EngineeringGeorgia Institute of [email protected]

Patricio VelaSchool of Electrical and Computer EngineeringGeorgia Institute of [email protected]

Lei ZhuSchool of Mechanical EngineeringGeorgia Institute of [email protected]

Steve [email protected]

Allen TannenbaumECE & BMEBoston [email protected]

MS1The Block DROP and CAV Algorithms for Com-pressed Sensing Based Tomography

In this talk, we will introduce a block diagonally-relaxedorthogonal projectionalgorithm and a block component av-eraging algorithm for computedtomography image recon-struction in the compressed sensing framework andderivethe convergence. Numerical experiments are shown to il-lustrate the convergence of the new algorithms.

Jiehua ZhuGeorgia Southern [email protected]

Xiezhang LiGeorgia Southern UniversityDepartment of Mathematical [email protected]

MS2Hard Thresholding with Norm Constraints

We introduce a new sparse recovery paradigm, where ef-ficient algorithms from combinatorial and convex opti-mization interface for interpretable and model-based so-lutions. A highlight is the introduction of a new algo-rithmic definition of union-of-subspaces models, which wedub as the Polynomial time Modular Approximation Prop-erty (PMAP). PMAP rigorously identifies tractable modelswithin the generic union-of-subspaces models, extendingthe reach of the model-based sparse recovery to a broader

set of applications. Synthetic and real data experimentsillustrate that our approach can significantly enhance theperformance of both hard thresholding methods and con-vex solvers in sparse recovery.

Volkan [email protected]

MS2Recovering Cosparse Vectors using Convex vsGreedy Algorithms

Analysis cosparse models are an alternative to standardsparse models, where an analysis operator multiplies thesignal to obtain sparse coefficients. In this model the goodold (synthesis) dictionary is no longer exploited and ismade redundant. After highlighting the main differencesbetween the models, empirical and theoretical performancecomparisons will be provided between convex cosparse re-construction algorithms and a recently introduced cosparseanalogue to greedy sparse approximation algorithms, calledthe Greedy Analysis Pursuit. Imaging and acoustic appli-cations will be discussed.

Remi GribonvalProjet METISS, IRISA-INRIA35042 Rennes cedex, [email protected]

Sangnam [email protected]

Mike DaviesUniversity of [email protected]

Michael EladComputer Science [email protected]

MS2Simple and Practical Algorithm for Sparse FourierTransform

The Fast Fourier Transform (FFT) is one of the most fun-damental numerical algorithms. It computes the DiscreteFourier Transform (DFT) of an n-dimensional signal in O(nlog n) time. The algorithm plays an important role in manyareas. In many applications (e.g., audio, image or videocompression), most of the Fourier coefficients of a signal are”small” or equal to zero, i.e., the output of the transform is(approximately) sparse. In this case, there are algorithmsthat enable computing the non-zero coefficients faster thanthe FFT. However, in practice, the exponents in the run-time of these algorithms and their complex structure havelimited their applicability to only very sparse signals. Inthis talk, I will describe a new set of algorithms for sparseFourier Transform. Their key feature is simplicity, whichleads to efficient running time with low overhead, both intheory and in practice. In particular, we can achieve aruntime of O(k log n), where k is the number of non-zeroFourier coefficients of the signal. This improves over theruntime of the FFT for any k ¡¡n.

Haitham HassaniehMIT

64 IS12 Abstracts

[email protected]

Piotr IndykMassachusetts Institute of [email protected]

Dina Katabi, Eric [email protected], [email protected]

MS2Graphical Models and Message Passing Algorithmsto Solve Large Scale Regularized Regression Prob-lems

This work surveys recent work in applying ideas fromgraphical models and message passing algorithms to solvelarge scale regularized regression problems. In particular,the focus is on compressed sensing reconstruction via 1penalized least-squares (known as LASSO or BPDN). Wediscuss how to derive fast approximate message passing al-gorithms to solve this problem. Surprisingly, the analysisof such algorithms allows to prove exact high-dimensionallimit results for the LASSO risk.

Andrea MontanariStanford [email protected]

MS3Anisotropic Elastic Moduli Reconstruction ofTransversely Isotropic Material in Magnetic Res-onance Elastography

Magnetic resonance elastography (MRE) is an elastic tissueproperty imaging modality in which phase-contrast basedMRI imaging technique is used to measure internal dis-placement induced by a harmonically oscillating mechani-cal vibration. Since soft tissues like skeletal muscles showanisotropic behavior, we develop a new method for theanisotropic elastic property imaging. In this talk, we con-siders the reconstruction in a transversely isotropic modelwhich is the simplest case of anisotropy, based on an ex-plicit representation formula using the Newtonian potentialof the measured displacement.

Ohin KwonKonkuk [email protected]

MS3A Shape and Topology Optimization Method forthe Resolution of Inverse Problems in Tomography

We propose a shape optimization approach for the resolu-tion of inverse problems in tomography, namely ElectricalImpedance Tomography and Fluorescence Diffuse OpticalTomography. These problems are severely ill-posed, andwe regularize them by assuming that the functions to bereconstructed are piecewise constants. Then the problemessentially boils down to determining the shapes of somehidden inclusions. The sensitivity of the objective with re-spect to small boundary and topological perturbations ofthe inclusions is analysed.

Antoine LaurainHumboldt-University of [email protected]

MS3Hybrid Techniques on Electrical Tissue PropertyImaging

Noting that MRI scanner can noninvasively measure mag-netic fields inside the human body in a form of cross-sectional image, impedance imaging methods using MRIhave been lately proposed. MREIT performs conductivityimaging below 1 kHz, whereas MREPT produces both con-ductivity and permittivity images at the Larmor frequencybased on B1-mapping techniques. In this talk, we reviewthe latest techniques in MREIT and MREPT.

Jin Keun [email protected]

MS3Conductivity Imaging from Minimal Current Den-sity Data

Recently has been shown that the magnitude of one cur-rent density field suffices to recover the electrical conduc-tivity inside a body. This interior data can currently beobtained from Magnetic Resonance data in the presence ofdirect/very low frequency current applied at the boundaryof the object. One of the methods reduces the problem tominimizing a weighted gradient in the space of functionsof bounded variation. In this talk I will present recentprogress on this minimization problem. The results havebeen obtained jointly with A. Nachman and A. Moradifam.

Alexandru TamasanUniversity of Central [email protected]

MS4Efficient Algorithms for Photoacoustic Imaging

Advances in photoacoustic imaging (PAI) come with theurgent need for fast reconstruction algorithms whose com-putational complexity scales up to logarithmic factors lin-ear in the problem size. The well known fast Fourier trans-form (FFT) is such an algorithm and found its applica-tion as nonequispaced variant e.g. in magnetic resonanceimaging. As part of an iterative reconstruction scheme,we discuss a specific four-dimensional sparse FFT for thenecessary computation of spherical mean values in three-dimensional PAI.

Stefan KunisUniversity of [email protected]

MS4TV-Minimization for Color Images on Non-regularGrids

Registering slice images of a 3D object in order to recon-struct the object results in a reconstruction where the pix-els are no longer on a regular grid. This is even worse withelastic registration. In applications as MALDI imaging itis not suitable to ”resample” the reconstruction hence weneed to adjust image processing algorithms to non-regulargrids. This results in some problems e.g. for TV-based de-noising methods. In this talk we will present our adaption

IS12 Abstracts 65

of the Chambolle method.

Stefan SchifflerUniversity of [email protected]

MS5The Past and Future of Imaging Through Turbu-lence

The normal state of the atmosphere is a turbulent state dueto wind and thermal currents. The atmospheric effects arenoticeable, but not a serious complication for many com-mon imaging situations. For example, atmospheric turbu-lence effects are rarely seen in vacation photos. When vi-sual aids such as zoom lenses are used, or when the ambienttemperature gradient is large, the atmospheric turbulenceproblems are highlighted. Therefore, atmospheric turbu-lence was initially addressed for improving imaging fromtelescopes. These visual aids, however, are increasinglyused in imaging sensors for persistent surveillance and ar-tificial sight for unmanned vehicles. For this reason, thereis a need to improve atmospheric turbulence corrections toaid other problems such as target recognition, autonomousnavigation, and automated image understanding. We willdiscuss some of models of atmospheric turbulence as well asthe inherent time and length scales. Furthermore, we willdiscuss the performance metrics that are necessary to en-able target recognition, navigation and image understand-ing applications.

Arjuna FlennerNaval Air Weapons [email protected]

MS5Turbulence Restoration: From Stabilization to At-mospheric Deblurring

Atmospheric turbulence affects long range imaging systemsand can be modeled as a composition of geometric distor-tions and blur. First, we propose to combine a deforma-tion field estimation and some regularization schemes inorder to retrieve a stabilized image. Secondly, we adressthe deblurring problem and efficiently solve it by using ananalytical atmospheric kernel. Many experiments on realimages will illustrate the performances of our approach.

Jerome GillesUniversity of California Los Angeles (UCLA)Department of [email protected]

Stanley J. OsherUniversity of CaliforniaDepartment of [email protected]

MS5A Dynamic Texture Model for Imaging ThroughTurbulence

We address the problem of recovering an image from a se-quence of distorted versions of it, where the distortion iscaused by what is commonly referred to as ground-levelturbulence. In mathematical terms, such distortion can beseen as the cumulative effect of a time-dependent anisopla-natic (space-variant) blur and a time-dependent deforma-

tion of the image domain. We introduce a dynamic modelthat is statistical in nature, rather than physical, for thegeneration of turbulence based on the dynamic texture ap-proach. We expand the model to include the unknownimage as part of the unobserved state and apply Kalmanfiltering to estimate such state. This operation yields ablurry image where the blurring kernel is, in fact, close tobeing space-invariant. Applying blind nonlocal Total Vari-ation (NL-TV) deconvolution yields a crisp final result.

Mario MicheliLaboratoire de Mathematiques AppliqueesUniversite Paris [email protected]


Stefano SoattoUCLAComputer Science [email protected]

Andrea L. BertozziUCLA Department of [email protected]

MS5Removing Atmospheric Turbulence

A new approach is proposed, capable of restoring a sin-gle high-quality image from a given image sequence dis-torted by atmospheric turbulence. This approach reducesthe space and time-varying deblurring problem to a shiftinvariant one. It first registers each frame to suppress ge-ometric deformation through non-rigid registration. Next,a temporal regression process is carried out to produce animage from the registered frames, which can be viewed asbeing convolved with a space invariant near-diffraction-limited blur. Finally, a blind deconvolution algorithm isimplemented to deblur the fused image, generating a highquality output. Experiments using real data illustrate thatthis approach can effectively alleviate blur and distortions,recover details of the scene, and significantly improve visualquality.

Xiang ZhuUniversity of California, Santa [email protected]

Peyman MilanfarEE DepartmentUniversity of California, Santa [email protected]

MS6On the Evaluation Complexity of Nonsmooth Com-posite Function Minimization with Applications toNonconvex Nonlinear Programming

We estimate the worst-case complexity of minimizingan unconstrained, nonconvex composite objective with astructured nonsmooth term by means of some first-ordermethods. We find that it is unaffected by the nonsmooth-ness of the objective in that a first-order trust-region orquadratic regularization method applied to it takes at most

66 IS12 Abstracts

O(ε−∈) function-evaluations to reduce the size of a first-order criticality measure below ε. These results allow usto analyze the evaluation complexity of nonlinear non-convex constrained programming by means of a penaltymethod and a short-step target-tracking type algorithm.We find that the complexity of reaching within ε of aKKT point is at most O(ε−∈) problem-evaluations, whichis the same in order as the function-evaluation complex-ity of steepest-descent methods applied to unconstrained,nonconvex smooth optimization.

Coralia CartisUniversity of EdinburghSchool of [email protected]

Nicholas I.M. GouldRutherford Appleton [email protected]

Philippe L. TointUniversity of Namur, [email protected]

MS6Nonconvex Nonsmooth Optimization via GradientSampling

We present an algorithm for general-purpose nonconvexnonsmooth optimization. The methodology assumes thatthe problem functions are locally Lipschitz and continu-ously differentiable over dense subsets of Rn. It is ap-plicable in unconstrained and constrained environments,where in the latter setting the constraint functions canalso be nonconvex and nonsmooth. The algorithm is basedon a process of sampling gradients of the problem func-tions randomly in an epsilon-neighborhood of a given so-lution estimate, solving a quadratic optimization subprob-lem, and then performing a line search. Global convergenceguarantees state that, with probability one, the algorithmconverges to a stationary point of the optimization prob-lem and numerical results indicate that the method is effi-cient on a wide range of problems. It is expected that themethodology can easily be tailored to specific applications.

Frank E. CurtisIndustrial and Systems EngineeringLehigh [email protected]

Xiaocun QueLehigh [email protected]

MS6High-Dimensional Covariance Estimation underSparse Kronecker Product Structure

We consider the problem of structured covariance matrixestimation when the covariance factors as a kronecker prod-uct of low dimensional matrices and the observations areGaussian distributed. This constitutes a new approach toestimating sparse graphical models for imaging, bioinfor-matics, finance, and other applications. The problem isformulated as a variational optimization of a Gaussian log-likelihood function with additive l1 penalty on non-sparsityof the kronecker product matrices. The dual formulation

of this problem motivates an iterative algorithm based ona block coordinate-descent approach. Although the varia-tional problem is nonconvex and nonsmooth, we show thatthe proposed algorithm converges to a local maximum un-der relatively mild assumptions. We give a tight rate ofFrobenius norm convergence as the dimensions of the co-variance matrix go to infinity. This rate is an order ofmagnitude better than the state-of-the art unstructuredgraphical lasso (Glasso) algorithm for estimation of gen-eral unstructured covariance matrices.

Alfred O. HeroThe University of MichiganDept of Elect Eng/Comp [email protected]

Theodoris TsiligkardisUniversity of [email protected]

MS6Wasserstein Barycenter: Global Minimizers andAlgorithm

This paper proposes a new definition of the averaging ofdiscrete probability distributions as a barycenter over the-Wasserstein space. Replacing the Wasserstein original met-ric by a sliced approximation over 1D distributions allowsus to use a fast stochastic gradient descent algorithm. Thisnew notion of barycenter of probabilities is likely to findapplications in computer vision where one wants to averagefeatures defined as distributions.

Rabin JulienUniversity of [email protected]

MS7Denoising Textures with a Multiplicative Regular-ization Method

Many problems in scientific computing can be cast as in-verse problems. Unfortunately, most inverse problems areill-posed, leading to unstable solutions. The most com-mon approach to have well-posed inverse problems is touse the Tikhonovs additive regularization method datedback to 1943, which adds a regularization function to thedata function. In this talk, we will introduce an alternateregularization method for inverse problems, which basi-cally considers a multiplicative approach between regular-ization functional and data-oriented functional. The mainimprovements of this approach compared to Tikhonov’sregularization method are the geometric invariance (impor-tant for medical and modern imaging techniques s.a. omni-directional imaging) and the data adaptive regularization(process that can locally adapt its strength to the data).Experiments show that the proposed denoising methodprovides promising results for piecewise smooth images andtextures.

Xavier BressonUCLADepartment of [email protected]

IS12 Abstracts 67

MS7Sparse Modeling of Random Phase Textures

Randomizing the Fourier phases of an image is a simplealgorithm that permits to reproduce well several kinds ofmicrotextures. This Random Phase Noise (RPN) model isalso interesting from a mathematical viewpoint, in partic-ular because of its link with stationnary Gaussian RandomFields. We describe a way to represent a RPN by a sparseimage called ”reduced texton”, that solves an optimizationproblem, and show its usefulness both for texture analysisand the synthesis of arbitrary large texture samples. (jointwork with A.Desolneux and S.Ronsin)

Lionel MoisanParis [email protected]

MS7Optimal Transport Mixing of Gaussian TextureModels

In this talk I will tackle the problem of mixing color texturemodels learned from an input dataset. I focus on station-ary Gaussian texture models, also known as spot noises.I derive the barycenter and geodesic path between mod-els according to optimal transport. This allows the userto navigate inside the set of texture models, and performtexture synthesis from the obtained interpolated models.Numerical examples on a library of exemplars show theability of our method to generate arbitrary interpolationsamong unstructured natural textures. This is a joint workwith Sira Ferradans, Gui-Song Xia and Jean-Francois Au-jol.

Gabriel PeyreCeremade, Universite Paris [email protected]

MS7A Level-set Approach to Image-Texture Segmenta-tion: Old and New

In this talk we present an ’old’ and ’new’ level-set methodto image texture segmentation: the ’old’ idea of level sets,the ’new’ insight with helps of recent convex optimizationdevelopments in image processing. It leads to a fast andreliable approach to computing segments based on texturefeatures: the inherent obstacles with the classical level-setsare solved under the new variational perspective.

Jing YuanDepartment of Computer ScienceUniversity of Western [email protected]

MS8Inversion of the V-Line Radon Transform in a Discand Its Applications in Imaging

Several novel imaging modalities proposed during the lastcouple of years are based on a mathematical model, whichuses the V-line Radon transform (VRT). This transform,sometimes also called broken-ray Radon transform, inte-grates a function along V-shaped piecewise linear trajec-tories composed of two intervals in the plane with a com-mon endpoint. Image reconstruction problems in all thosemodalities require inversion of the V-line Radon transform.

While there are ample results about inversion of the regu-lar Radon transform integrating along straight lines, verylittle is known for the case of VRT. In this talk we discussan exact inversion formula for the V-line Radon transformof functions supported in a disc of arbitrary radius. Ourmethod is based on the classical filtered back-projectioninversion formula of the regular Radon transform, and hassimilar features in terms of stability, speed, and accuracy.We illustrate our results by presenting various numericalsimulations. The talk also discusses the interior problemof VRT, and some other open problems.

Gaik AmbartsoumianMathematics DepartmentTexas A&M [email protected]

MS8Microlocal Analysis of Common Midpoint SARImaging

In Synthetic Aperture Radar (SAR) imaging, a plane car-rying an antenna moves along a path. The antenna emitselectromagnetic waves which scatter off the ground and aredetected back with the same antenna. The received signalsare used to produce an image of the ground. In commonmidpoint SAR imaging, the source and receiver move alonga line at equal speeds away from a common midpoint. Weanalyze the microlocal properties of the forward operatorF which is a Fourier integral operator with singularities.To recover the image we study the normal operator F ∗Fand show that F ∗F is a sum of four operators belonging toIp,l(∆, Ci), where Ci are canonical graphs representing theartifacts which appear by backprojecting the data. This isjoint work with G. Ambartsoumian, V. Krishnan, C. Nolanand T. Quinto.

Raluca FeleaDepartment of MathematicsRochester Institute of [email protected]

MS8Microlocal Analysis of an Ultrasound Transformwith Circular Source and Receiver Trajectories

We consider a generalized Radon transform that comes upin ultrasound reflection tomography. In our model, the ul-trasound emitter and receiver move at a constant distanceapart along a circle. We analyze the microlocal proper-ties of the transform R that arises from this model. As aconsequence, we show that for distributions with supportsufficiently inside the circle, the normal operator R∗R (R∗is the L2 adjoint of R) is an elliptic pseudodifferential op-erator of order −1 and hence all the singularities of suchdistributions can be recovered.

Gaik AmbartsoumianMathematics DepartmentTexas A&M [email protected]

Venky P. KrishnanDepartment of MathematicsTufts [email protected]

Todd QuintoTufts University

68 IS12 Abstracts

[email protected]

MS8Synthetic Aperture Imaging of Moving Targets us-ing Ultra-narrowband waveforms

We present a novel method of imaging multiple movingtargets using a SAR system transmitting ultra-narrowbandcontinuouswaveforms. Ourmethod comprises of a new for-wardmodel that relates the velocity as well as reflectiv-ity information at each location to a correlated receivedsignal; and a novel image formation method based onfiltered-backprojection and image-contrast optimization.The method results in well-focused reflectivity images ofmoving targets and their velocity estimates regardless ofthe target location, speed, and velocity direction.

Birsen YaziciDepartment of Electrical, Computer and SystemsEngineeringRensselaer Polytechnic [email protected]

Ling [email protected]

MS9Recent Advances in Image Reconstruction from In-terior Data

The interior problem of tomography does not have a uniquesolution. When additional information about the regionof interest is available, the uniqueness and stability can berecovered. In this talk we will summarize recent progress intheoretical understanding of the interior problem and somenew algorithms for its numerical solution. Our startingpoint is the Gelfand-Graev formula, which establishes arelation between the ray transform of a function and itsHilbert transform along lines.

Alexander KatsevichMathematics DepartmentUniversity of Central [email protected]

Alexander TovbisUniversity of Central FloridaDepartment of [email protected]

MS9The Hilbert Transform in Analytic Cone-beam Im-age Reconstruction: A Visual Tour

Several of the advances that have occurred in the past10 years in analytic cone-beam image reconstruction havecentered around the Hilbert transform. These advancesinclude Katsevich’s theoretically exact image reconstruc-tion framework, the derivative backprojection (also calledbackprojection-filtration) approach, and several others. Iwill present a visual framework that I find useful for de-signing reconstruction algorithms and show how each ofthese approaches can be understood in this framework. Iwill also touch briefly on some of the interesting tradeoffsthat are made where theory meets practice in CT.

Jed Pack

GE Global [email protected]

MS9Tensor-based Formulation for Spectral ComputedTomography

Utilizing photon counting detector technology spectralcomputed tomography (CT) provides energy-selective mea-surements and reconstructions. In this work, we interpretthe spectral CT data as a multidimensional array (tensor)and apply decomposition techniques to reduce the dimen-sionality of the linear inverse problem. Furthermore, weinvestigate tensor-based formulation of the multi-spectralreconstruction of linear attenuation coefficient and regular-ization techniques such as spectral regularization based onlow rank assumptions for the case of densely binned energyspectra. We compare our method filtered back projection(FBP) reconstructions for each energy bin to demonstratethe advantages of the proposed formulation.

Oguz SemerciTufts [email protected]

Ning HaoTufts UniversityDepartment of [email protected]

Misha E. Kilmer, Eric MillerTufts [email protected], [email protected]

MS9A Fourier Integral Operator Model for ConebeamX-ray CT and its Inversion

This talk presents an analytic model in the form of aFourier integral operator and its inversion for X-ray CTimage reconstruction. The model can accommodate sys-tem related parameters and arbitrary source trajectories.We present a comparison of the inversion formula with theexisting ones and demonstrate its performance in numeri-cal simulations.


Zhengmin [email protected]

MS10Radar Imaging

This talk will address current mathematical work in radarimaging.

Margaret CheneyRensselaer Polytechnic InstDept of Mathematical [email protected]

IS12 Abstracts 69

MS10Edge-based Image Reconstruction, HierarchicalFeature Extraction, and Analysis - A StructuralApproach

Remote sensing image analysis has largely focused on spec-tral analysis. Modern sensors reveal spatial detail that callsfor exploitation of structural information. Moreover, to-days volume of imagery requires faster exploitation thatleverages pixel redundancy for data reduction. We presenta perceptually motivated approach to feature extractionthat that utilizes edge information to recast images interms of polygonal features. This representation can beused for automating object, anomaly, and change detec-tion in remote sensing imagery.

Lakshman PrasadLos Alamos National [email protected]

MS10Contrast Invariant and Affine Sub-Pixel Optical-Flow for Remote Sensing Applications

CIAO (Contrast Invariant and Affine Optical Flow) isa dense area-based subpixel image matching algorithm.CIAO does not force the estimated disparity field betweentwo images to be smooth and allows for contrast andbrightness changes. The proposed model considers localaffine displacements and it is robust to drastic changes inthe images’ content. CIAO proves particularly useful toextract high quality, high accuracy, and high density dis-parity maps from pairs of stereoscopic images.

Neus SabaterCMLA - ENS [email protected]

MS10Variational Methods for Remote Sensing Applica-tions

The development of a variety of novel sensors for captur-ing remote sensing and atmospheric data gave rise to manynew applications and an enormous need for solving newinverse problems arising in these fields. This talk will de-scribe some of the applications and variational approachesfor solving associated inverse problems. In particular, de-composition of different layers of clouds, reconstruction ofhurricane images, and data fusion of satellite records ac-quired by multiple instruments will be discussed.

Igor YanovskyJet Propulsion Laboratory, California Institute ofTechnologUniversity of California, Los [email protected]

MS11Temporally Consistent Segmentation of Longitudi-nal MRI Data and Related Issues

Abstract not available at time of publication.

Chunming LiVanderbilt [email protected]

MS11An Efficient Algorithm for Multiphase Image Seg-mentation with Intensity Bias Correction

We present an efficient algorithm for multiphase imagesegmentation in the presence of strong noise and inten-sity inhomogeneity. The problem is formalized as a min-max optimization problem that consists both primal anddual variables. We use the primal dual hybrid gradient(PDHG) algorithm to alternately solve for the optimal so-lutions. The proposed algorithm is quite efficient in thatall the subproblems have closed form solutions. Moreover,the computational complexity is shown to be linear withrespect to the size of the image. Numerical experimentson various images demonstrated that our algorithm out-performs recently developed methods in terms of efficiencyand accuracy.

Haili Zhang, Yunmei ChenDepartment of MathematicsUniversity of [email protected], [email protected]


MS11Marginal Space Learning for Efficient Detectionand Segmentation of Anatomical Structures inMedical Imaging

Recently, we proposed marginal space learning (MSL) as ageneric approach for automatic detection and segmentationof 2D/3D anatomical structures in many medical imagingmodalities. To accurately localize a 3D object, we need toestimate nine pose parameters (three for position, three fororientation, and three for anisotropic scaling). Instead ofexhaustively searching the original nine-dimensional poseparameter space, only low-dimensional marginal spaces aresearched in MSL to improve the detection speed. In thistalk, I will present MSL in detail, followed by some re-cent developments, e.g., constrained MSL, MSL for non-rigid shape detection, and hierarchical MSL. Live demoson various applications of MSL will be shown.

Yefeng ZhengSiemens Corporate [email protected]

MS11A Geodesic Active Contour Based Model for ShortAxis Cardiac-MR Image Segmentation

We propose a novel geodesic active contour based varia-tional model that uses two level set functions to segmentthe right/left ventricles and the epicardium in short axisMR images respectively. To relax the restriction on thechoice of initial contours, we develop a new edge function.We also propose an iterative method to minimize the en-ergy of our model. Experimental results are presented tovalidate the effectiveness of the proposed model.

Sung Ha KangGeorgia Inst. of [email protected]

70 IS12 Abstracts

Wei ZhuMathematics, University of [email protected]

George BirosBiomedical Engineering, Georgia Institute of [email protected]

MS12General Purpose First-order Methods for ConvexMinimization

Interior-point methods handle convex programs regard-less of smoothness or constraints, but have difficulty withlarge problems. Researchers are turning to simpler meth-ods such as (projected) gradient-descent, but this requiressmoothness and the ability to efficiently project. To ad-dress these issues, we introduce a framework and softwarepackage called TFOCS (Becker,Grant,Candes), and com-pare with recent splitting methods (Chambolle,Pock). Wealso present work on improving convergence speed by usingnon-diagonal preconditioners (Becker,Fadili).

Stephen BeckerCalifornia Institute of [email protected]

Emmanuel CandesStanford UniversityDepartments of Mathematics and of [email protected]

Michael C. GrantCVX ResearchCalifornia Institute of [email protected]

Jalal FadiliCNRS, ENSICAEN-Univ. of Caen, [email protected]

MS12Generalized Forward-Backward Splitting forSparse Recovery

This work introduces the generalized forward-backwardsplitting algorithm for minimizing convex functions of theform F +

∑n

i=1 Gi, where F has a Lipschitz-continuousgradient and the Gi’s are simple in the sense that theirMoreau proximity operators are easy to compute. Whilethe forward-backward algorithm cannot deal with morethan n = 1 non-smooth function, our method generalizesit to the case of arbitrary n. Our method makes an ex-plicit use of the regularity of F in the forward step, andthe proximity operators of the Gi’s are applied in parallelin the backward step. This allows the generalized forwardbackward to efficiently address an important class of con-vex problems. We prove its convergence in infinite dimen-sion, and its robustness to errors on the computation ofthe proximity operators and of the gradient of F . Exam-ples on sparse recovery and inverse problems demonstratethe advantage of the proposed method in comparison toother splitting algorithms.

Jalal FadiliCNRS, ENSICAEN-Univ. of Caen, [email protected]

Gabriel PeyreCEREMADE, Universite Paris [email protected]

Hugo RaguetCeremade CNRS-Universite [email protected]

MS12Constructing Test Instances for Sparse RecoveryAlgorithms

Since the number of available algorithms solving the Ba-sis Pursuit Denoising problem rises, it is desirable to cre-ate test instances containing a matrix, a right side anda (given) solution, which can be recovered exactly. Toguarantee the existence of such a solution, the so-calledsource condition is sufficient and necessary. In this talk, wepresent strategies to construct test instances with differentadditional properties (e.g. a favorable dual certificate).

Christian Kruschel, Dirk LorenzTU [email protected],[email protected]

MS13Sensing Videos with the Single Pixel Camera

Recent progress in compressive sensing (CS) has enabledthe construction of computational imaging devices thatsense parsimoniously at the information rate of the sceneunder view rather than its Nyquist rate. One such CS de-vice is the single pixel camera that randomly multiplexeslight spatio-temporally onto a single sensor, which enablesit to operate efficiently in wavelength regimes (such as deepinto the infrared) that require exotic detectors. In this talk,we report on recent work to efficiently sense and recovervideo from CS measurements. Our first approach relieson recovering foreground innovations from a video by per-forming background subtraction directly on compressivemeasurements. Our second approach addresses the com-pressive video sensing problem by regularizing the scenewith a dynamical systems prior. The predictive/generativemodels associated with dynamical systems are employed toprovide significant spatio-temporal tradeoffs at sensing. Fi-nally, our third approach highlights some of the recent workon simultaneously sensing the scene as well as its motionas objects in the scene articulate in front of the camera.Together, these methods provide a wide range of tradeoffson parsimony in sensing and accuracy of reconstructed sig-nals. We demonstrate the effectiveness of our techniqueson simulated and real single pixel camera data.

Richard G. BaraniukRice UniversityElectrical and Computer Engineering [email protected]

MS13Infinite Photography

We introduce a new mathematical image model using thephotographic process as the starting point. Images are rep-resented as infinite sequences of photons allowing analysisat arbitrarily high resolution. We show that the resultingspace has a metric structure and is intimately connectedwith bounded Borel measures. Furthermore, novel compu-

IS12 Abstracts 71

tational image processing algorithms based on the modelare illustrated. This is joint work with Matti Lassas andSamuli Siltanen.

Tapio HelinUniversity of [email protected]

Matti LassasUniversity of [email protected]

Samuli SiltanenUniversity of [email protected]

MS13Femto-Photography: Time Resolved Imaging forLooking Around Corners

Can we look around corners beyond the line of sight? Ourgoal is to exploit the finite speed of light to improve imagecapture and scene understanding. New theoretical analy-sis coupled with emerging ultra-high-speed imaging tech-niques can lead to a new source of computational visualperception. We are developing the theoretical foundationfor sensing and reasoning using Femto-photography andtransient light transport, and experimenting with scenariosin which transient reasoning exposes scene properties thatare beyond the reach of traditional computer vision. (Jointwork with a large team, see http://raskar.info/femto)

Ramesh RaskarMassachusetts Institute of [email protected]

MS14Registration of MALDI Imaging Data

Registration is central in lifting existing 2D MALDI imag-ing to a 3D image modality. In the talk we present anapproached based on nonlinear image registration for thereconstruction of 3D MALDI images from a stack of 2DMALDI slices. Therefore we first apply elastic serial regis-tration resulting a stack of aligned 2D images. Afterwards,we perform 3D co-registration with an initially acquiredMR image to restore the object’s shape.

Stefan Heldmann, Judith Berger, Jan StrehlowFraunhofer [email protected],[email protected],[email protected]

Stefan WirtzFraunhofer MEVISInstitute for Medical Image [email protected]

MS14Artifact-free Decompression of Transform-codedMulti-channel Images with TGV

The problem of artifact-free decompression of transform-coded multi-channel images, such as JPEG compressedcolor images, is addressed. This is done by formulating aconstrained optimization problem involving a convex indi-

cator function as data fidelity- and the recently introducedTotal Generalized Variation (TGV) functional as regular-ization term. After a discussion of the continuous model,the application to JPEG decompression is considered, in-cluding practical aspects such as numerical realization andfast implementation on the GPU.

Martin HollerUniversity of [email protected]

Kristian BrediesUniversity of GrazInstitute of Mathematics and Scientific [email protected]

MS14Including Spatial Similarities into HierarchicalClustering of Hyperspectral Terahertz Images

Hierarchical clustering (HC) is well suited for data ex-ploration. For computational feasibility and flexibility ofmethods we combine the phase I preclustering of Karypis’Chameleon with agglomerative HC. Spatial similarity isintegrated by slicewise image processing or direct inclu-sion into the similarity matrix: Precluster members areassumed to be independent Gaussian distributed and theiroverlap is calculated. A high overlap increases the totalsimilarity. Results on hyperspectral Terahertz images showimproved coherence of segmented objects.

Henrike StephaniFraunhofer Institute for Industrial MathematicsKaiserslautern, [email protected]

Karin Wiesauer, Stefan KatletzRECENDT GmbHLinz, [email protected], [email protected]

Daniel MolterFraunhofer Institute for Physical MeasurementTechniquesKaiserslautern, [email protected]

Bettina HeiseJohannes Kepler University, CDL MS-MACHLinz, [email protected]

MS14Generating 3D MALDI Imaging Data

MALDI imaging is a technique which records mass spec-trums of tissue in two dimensions. We set up a processingpipline for generating three-dimensional (3D) MALDI-dataenabling for 3D-clustering. We present a fully-automaticpipeline, regarding detection and segmentation of tissue onthe carrier, pre-alignment, rigid reconstruction and imagefusion. The whole processes brings MALDI spectrum dataand histologic images together.

Jan Strehlow, Judith Berger, Stefan HeldmannFraunhofer [email protected],[email protected],[email protected]

72 IS12 Abstracts

Stefan WirtzFraunhofer MEVISInstitute for Medical Image [email protected]

MS15On-the-Fly Turbulence Effect Mitigation in Long-Range Surveillance Applications Based on LuckyRegion Fusion

We present a technique and experimental results for mit-igation of the turbulence effect in long-range imaging sce-narios. Our approach is based on the selection and fusionof image lucky regions. Lucky region selection is performedbased on the local image quality and allows fusion of a di-versity of images including different distortions, exposuresand spectral ranges.

Mathieu AubaillyIntelligent Optics LaboratoryUniversity of [email protected]

MS15Video Restoration of Turbulence Distortion

When the video is taken from a long range system, at-mospheric turbulence can corrupt the video sequence andan object in a distance can look distorted in the video se-quence. Blurring and diffeomorphism are couple of themain effect of atmospheric turbulence, and we proposemethods to stabilize the video sequence and give a goodreference latent image. We reconstruct a new video se-quence using Sobolev gradient deblurring with the tempo-ral smoothing, and one latent image is found further utiliz-ing the lucky-region method. With these methods, withoutany prior knowledge, the video sequence is stabilized whilekeeping sharp details, and the latent image shows moreconsistent straight edges.


Sung Ha KangGeorgia Inst. of [email protected]


Andrea L. BertozziUCLA Department of [email protected]

MS15Non Rigid Geometric Distortions Correction – Ap-plication to Atmospheric Turbulence Stabilization

A novel approach is presented to recover image degradedby atmospheric turbulence. Given a sequence of framesaffected by turbulence, we construct a variational modelto characterize the static image. The optimization problemis solved by the Bregman iteration and operator splitting

method. Our algorithm is simple and efficient, and can beeasily generalized for different scenarios.

Yu MaoInstitute for Mathematics and Its ApplicationsUniversity of [email protected]

Jerome GillesUniversity of California Los Angeles (UCLA)Department of [email protected]

MS16Lp(Ω)−Optimization with p ∈ [0, 1)

For p ∈ [0, 1) we consider the problem

inf J(u) =12|Ku− f|2Y + β Np(u) over u ∈ C (P)

where K is a bounded, possibly compact operator withvalues in a Hilbert space Y , C a closed convex set, and

Np(f) =∫

S

|f(x)|p dx for p ∈ (0, 1) and N0(f) = measx ∈ Ω : f(x) = 0.

We show existence for an L2-regularized version of (P ) andderive optimality conditions. Further we propose a pri-mal dual active set method and we analyse a monotonescheme for obtaining numerical solutions. Applications in-clude mathematical imaging with sparsity constraints.

Karl KunischKarl-Franzens University GrazInstitute of Mathematics and Scientific [email protected]

Kazufumi ItoNorth Carolina State UniversityDepartment of [email protected]

MS16Sparse Approximation via Penalty DecompositionMethods

In this talk we consider sparse approximation problems,that is, general l0 minimization problems with the l0-“norm’ of a vector being a part of constraints or objectivefunction. In particular, we first study the first-order op-timality conditions for these problems. We then proposepenalty decomposition (PD) methods for solving them inwhich a sequence of penalty subproblems are solved by ablock coordinate descent (BCD) method. Under some suit-able assumptions, we establish that any accumulation pointof the sequence generated by the PD methods satisfies thefirst-order optimality conditions of the problems. Further-more, for the problems in which the l0 part is the onlynonconvex part, we show that such an accumulation pointis a local minimizer of the problems. In addition, we showthat any accumulation point of the sequence generated bythe BCD method is a saddle point of the penalty subprob-lem. Moreover, for the problems in which the l0 part is theonly nonconvex part, we establish that such an accumula-tion point is a local minimizer of the penalty subproblem.Finally, we test the performance of our PD methods by ap-plying them to sparse logistic regression, sparse inverse co-variance selection, and compressed sensing problems. The

IS12 Abstracts 73

computational results demonstrate that our methods gen-erally outperform the existing methods in terms of solutionquality and/or speed.

Zhaosong LuDepartment of MathematicsSimon Fraser [email protected]

MS16On Local Linear Convergence of Elementary Algo-rithms with Sparsity Constraints

In sparsity optimization it is usually the sparsity of the sig-nal or matrix that is minimized subject to some (usuallylinear) constraint. The naive formulation of the probleminvolves the 0 function, and is avoided since the resultingoptimization problem has no known polynomial-time algo-rithms for its solution. Most practical numerical techniquesinvolve relaxations to convex or quasi-convex surrogates.Surprisingly, however, sets built upon the 0 function aremuch more regular than would be expected. This regular-ity allows us to prove local linear convergence of elementaryalgorithms even for nonlinear problems with sparsity con-straints without resorting to convex relaxations or even tothe usual mutual coherence or restricted isometry condi-tions.

Russell LukeDepartment of MathematicsUniversity of [email protected]

MS16Symbology-based Algorithms for Robust Bar CodeRecovery

UPC bar codes can be characterized as sparse represen-tations with respect to a certain symbology basis. We ex-ploit this low-dimensional structure and introduce a greedybar code reconstruction algorithm which can recover UPCbar codes from very noisy measurements and inaccurateparameter information. Extensions to general bar codes,radio-frequency identification, and text denoising will bediscussed.

Rachel WardDepartment of MathematicsUniversity of Texas at [email protected]

MS17Linking a Texture Model to Visual Processing inCortex

Many mathematical models of texture employ a hierarchy,with simpler elements at one stage combined into morecomplex features at later stages. It is also well known thatthe cortical areas of the human visual system are orga-nized into such a cascade. I will describe work that com-bines texture synthesis and human perception to forge alink between these two forms of hierarchy. Our model usesa cascade of biologically-plausible computations to capturea rich variety of features found in naturalistic texture im-ages. To link the model to perception, we exploit the factthat visual signals are not encoded uniformly across thevisual field; rather, neurons pool signals locally, in regionsthat grow with distance from the center of gaze. We de-velop a synthesis algorithm to generate images that are

matched to the locally-pooled textural features of an origi-nal image. Such synthetic images are increasingly distortedor jumbled as the regions grow large, but so long as theregions are matched to those of cortical neurons, these dis-tortions are invisible – the images are ”metamers”. Weuse perceptual experiments to quantify these effects, linkthem to different stages of cortical processing, and explaina variety of perceptual deficits that arise in the peripheryof the visual field.

Jeremy [email protected]

MS17Maximum Entropy Texture Modeling Based on aPhysiological Front End

I’ll describe our work in developing implicit models forboth visual and auditory textures based on cascades oflocal measurements. Specifically, input signals are decom-posed through application of bandpass linear filterbanks,interdigitated with rectifying nonlinearities, roughly anal-ogous to the processing found in biological visual and au-ditory systems. The representational capabilities of thesemeasurements may be studied by synthesizing novel sig-nals that yield identical sets of measurements. We findthat they are remarkably powerful in capturing structuralproperties of complex real-world textures.

Eero P. SimoncelliCourant Institute of Mathematical SciencesNew York [email protected]

MS17Texture, Structure and Visual Matching

I will describe a notion of Information for the purpose ofdecision and control tasks, as opposed to data transmissionand storage tasks implicit in Communication Theory. It isrooted in ideas of J. J. Gibson, and is specific to classesof tasks and nuisance factors affecting the data formationprocess. When such nuisances involve scaling and occlusionphenomena, as in most imaging modalities, the “Informa-tion Gap’ between the maximal invariants and the min-imal sufficient statistics can only be closed by exercisingcontrol on the sensing process. Thus, senging, control andinformation are inextricably tied. This has consequencesin understanding the so-called “signal-to-symbol barrier’problem, as well as in the analysis and design of activesensing systems. I will show applications in vision-basedcontrol, navigation, 3-D reconstruction and rendering, aswell as detection, localization, recognition and categoriza-tion of objects and scenes in live video.


MS184D Inverse Problems: Optimal Transport meetsSparsity and Low Rank

The aim of this talk is to study models and efficient algo-rithms for 4D reconstruction in biomedical imaging. Stan-dard reconstruction methods do not incorporate time de-

74 IS12 Abstracts

pendent information or kinetics. This can lead to deficientaccuracy at object boundaries, e.g. at cardiac walls. Wediscuss constrained optimization models combining ideas ofinverse problems, spatio-temporal regularization and opti-mal transport. The main emphasis lies on 4D image de-composition via sparsity and low rank. This allows forbetter reconstructions in difficult MRI imaging sequencesor video analysis with undesired background motion.

Christoph BruneDepartment of MathematicsUniversity of California Los [email protected]

Hao GaoUniversity of Los [email protected]


MS18Motion Compensation for Dynamic Contrast En-hanced MRI

In this work we present a novel approach for motion cor-rection for dynamic enhanced MRI sequences. The ideafor our method is twofold. First, we build on image regis-tration techniques for correcting motion artifacts from thedata. Second, we jointly estimate contrast uptake undercertain model assumptions. In a first approach, we makethe generic assumption the contrast uptake is a smoothprocess over time. In a second approach, we explicitlymodel the uptake as gamma function.

Stefan HeldmannFraunhofer [email protected]

Stephen KeelingInstitute for Mathematics and Scientific ComputingUniversity of [email protected]

Jan ModersitzkiUniversity of LubeckInstitute of Mathematics and Image [email protected]

Lars RuthottoInstitute of Mathematics and Image [email protected]

MS18Which Geometric Structure for the StatisticalAnalysis of Longitudinal Deformations?

Longitudinal morphological deformations can be evaluatedusing non-rigid registration for each subject but require theparallel transport to a common reference for longitudinalgroup-wise analysis. We have recently proposed a discreteconstruction of parallel transport only based on geodesics,the Schild’s Ladder procedure. It was shown to be im-plemented extremely efficiently with great stability in theframework of diffeomorphisms parameterized by stationaryvelocity fields (SVF). Along with the efficient image reg-

istration methods based on SVFs (DARTEL, log-demons,etc), this constitute an interesting practical alternative toRiemannian-based methods (LDDMM). We show that thetheoretical bases of this SVF framework in finite dimen-sion can be based on the canonical Cartan connections.This raises the question of the minimal geometric struc-ture needed to define interesting statistics on manifolds:can we remove the metric to generalize (part of) the statis-tical setting from Riemannian manifolds to affine connec-tion spaces?

Xavier PennecINRIA Sophia-Antipoolis [email protected]

MS18A Spatio-temporal Motion Correction Scheme forCardiac PET

Positron Emission Tomography (PET) is a nuclear imagingtechnique of increasing importance e.g. in cardiovascularinvestigations. However, cardiac and respiratory motion ofthe patient degrade the image quality due to acquisitiontimes in the order of minutes. Reconstructions withoutmotion compensation are prone to spatial blurring and af-fected attenuation correction. These effects can be reducedby gating, motion correction and finally summation of thetransformed images. Gating schemes that are based on afiner grouping into phases of the respiratory and cardiaccycle are desirable in this application as each individualimage shows less motion artifacts. However, the gated im-ages are also based on a smaller fraction of counts andthus regularization becomes essential. In this work, wepresent a spatial-temporal approach to cardiac and res-piratory motion correction in gated 3D PET. It is basedon a mass-preserving and hyperelastic image registrationscheme, which acknowledges the density property of PETand guarantees diffeomor- phic solutions. To improve therobustness against noise, we introduce additional smooth-ing in time and jointly optimize over all transformations.

Lars RuthottoInstitute of Mathematics and Image [email protected]

Jan ModersitzkiUniversity of LubeckInstitute of Mathematics and Image [email protected]

MS19Photon State Space and Radiative Transfer: ThePhysics of Imaging with Remote Sensors

Overhead sensors ultimately count photons, thus samplingtheir few state-space dimensions: energy (wavelength), mo-mentum (viewing direction), spin (polarization). Physics-based retrievals of geophysical quantities from remote-sensing data therefore use at best multi-spectral/multi-angle radio-polarimetric input. However, implicit here isthat just one spatial sample (pixel) is processed at a time,and this is indeed how Earth-observation missions operate.Although much can be extracted at pixel-scales, new fron-tiers will be multi-pixel algorithms, and deeper exploitationof pulsed sources.

Anthony B. DavisJet Propulsion Laboratory

IS12 Abstracts 75

California Institute of [email protected]

MS19Sensitivity Studies of Optical Scanning PolarimeterRetrievals of Aerosol and Cloud Optical Properties

Scanning polarimeters are capable of simultaneously re-trieving optical properties of both aerosols and clouds inpartially cloudy pixels. This has been shown previously forsimulations with one dimensional radiative transfer mod-els, whose speed is required for operational retrievals. How-ever, neglecting the impacts of three dimensional scatteringwill degrade retrieval accuracy. We investigated the signifi-cance of this degradation on retrieval accuracy to select theideal operational retrieval strategy for scanning polarime-ters.

Kirk Knobelspiesse, Brian Cairns, Michael MishchenkoNASA Goddard Institute for Space [email protected], [email protected],[email protected]

MS19Atmospheric Radiative Effects and their Correc-tion in Landsat Images

The high resolution images of the Earth surface obtainedfrom satellites contain distortions caused by atmosphericscattering and absorption of solar radiation. We will dis-cuss the atmospheric radiative effects and an accuratemethod of atmospheric correction over non-Lambertianspatially variable surface in application to the Landsat im-agery. The approach uses formalism of the atmosphericoptical transfer function implemented in radiative transferalgorithm SHARM-3D.

Alexei LyapustinNASA Goddard Space Flight [email protected]

MS19Spectral Invariance in Atmospheric Radiation

Certain algebraic combinations of single-scattering albedoand solar radiation reflected from, or transmitted throughcloudy atmospheres do not vary with wavelength. Theseare called spectrally-invariant relationships; they are theconsequence of the fact that in cloud-dominated atmo-spheres the total extinction and total scattering phase func-tion are only weakly sensitive to wavelength. We will dis-cuss the physics behind this phenomenon, its mathematicalbasis, and possible applications to remote sensing and cli-mate.

Alexander MarshakNASA Goddard Space Flight [email protected]

Yuri KnyazikhinBoston [email protected]

MS20Electro-Acoustic Imaging Methods

In this talk, I will discuss recent progresses in electro-acoustic imaging, based on the idea of imaging by mod-

ification: a controlled modification of the internal physicalproperties, by an exterior stimulation, allows to extract lo-calized internal information on the unperturbed medium.I will conclude by a remark on the benefits of many re-dundant measurements for a microwave imaging problem.

Yves CapdeboscqMathematical Institute, University of [email protected]

MS20Joint Estimation of Wave Speed and AbsorptionDensity Functions with Photoacoustic Measure-ment

We present an approach for simultaneous identification ofthe absorption density and the speed of sound by photoa-coustic measurements. Experimentally our approach canbe realized with sectional photoacoustic experiments. Themathematical model for such an experiment is developedand exact reconstruction formulas for both parameters arepresented.


Andreas KirschKarlsruhe Institute for [email protected]

MS20Acousto-optic Imaging and Related Inverse Prob-lems

A tomographic method to reconstruct the optical proper-ties of a highly-scattering medium from incoherent acousto-optic measurements will be discussed. The method is basedon the solution to an inverse problem for the diffusion equa-tion and makes use of the principle of interior control ofboundary measurements by an external wave field.

John SchotlandUniversity of [email protected]

MS20Stabilizing Inverse Problems by Internal Data

Several hybrid imaging methods (e.g. those combiningelectrical impedance or optical imaging with acoustics) en-able one to obtain internal information. This informationstabilizes the exponentially unstable modalities of opticaland electrical impedance tomography. A number of man-ifestations of this effect have been studied previously. Iwill discuss a simple, general approach that shows whenand why internal information stabilizes the reconstruction.(This talk presents joint work with Peter Kuchment.)

Dustin SteinhauerTexas A&M [email protected]

MS21Do High Frequencies Contain Information about

76 IS12 Abstracts

Low Frequencies?

Data usually come in a high frequency band in wave-basedimaging, yet one often wishes to determine large-scale fea-tures of the model that predicted them. When is this pos-sible? Both the specifics of wave propagation and signalstructure matter in trying to deal with this multifacetedquestion. I report on some recent progress with Paul Handand Hyoungsu Baek. The answers are not always pretty.


MS21Interferometric Seismic Imaging by Sparse Inver-sion

Seismic interferometry complements conventional imagingby using multidimensional cross-correlations of recordedwavefields. Unfortunately, its performance may be im-pacted by practicalities such as noise, limited aperture, andincomplete data. We adress these issues by combining in-terferometric deconvolution with transform-domain spar-sity promotion and imaging. This leads to a formulationthat deals with data imperfections. Our preliminary find-ings show that sparsity promotion can lead to a significantimprovement of the image quality under a salt flank.

Joost van der NeutCITG, Delft University of [email protected]

Tristan van LeeuwenDepartment of Earth and Ocean ScienceUniversity of British Columbia, Vancouver BC [email protected]

Felix J. HerrmannSeismic Laboratory for Imaging and ModelingThe University of British [email protected]

Kees WapenaarCITG, Delft University of Technologyc.p.a.wapenaar@tudelft

MS21Modeling and Inverting Seismic P-S mode Conver-sions from Anelastic Targets

Truncated series expansions of modified Knott-Zoeppritzequations accurately reproduce pre-critical seismic ampli-tudes associated with anelastic P-S mode conversions. Theleading order contribution includes a correction for QS con-trasts. Amplitude-variation-with-frequency data from suchconversions can be used to directly determine target QS

values. It may be established: (1) that a P-S mode conver-sion occurs at an interface across which QS only varies; (2)that a QP contrast causes a negligible conversion, but sig-nificantly influences those caused by QS; (3) that the recip-rocal quality factors are proportional to the frequency ratesof change of the anelastic RPP , RPS and RSS; and (4) thatin the elastic limit and for large contrasts a VP /VS ratio of 2takes on special importance, decoupling density variationsand velocity variations across the reflecting boundary.

Kris InnanenUniversity of Calgary

[email protected]

MS21Nonlinear Imaging and Inversion of Seismic Wave-fields Using Interferometry

Using recent advances in seismic interferometry and scat-tering reciprocity, extended images are described as fieldsthat are reconstructed inside the subsurface where no realobservations exist, using observed boundary data and an apriori model. This definition of a seismic image provides afew advantages over more conventional definitions based onlinearized back-projections: they account for the full non-linear nature of seismic data, and can thus handle multiple-scattered waves as well as nonlinear finite-frequency ampli-tude effects. Since the extended images are by definitionsubsurface fields, they allow the design of novel image-domain objective functions that use extended-image an-nihilators explicitly defined in terms of partial differentialequations (PDEs) for scattered fields. These PDE-basedannihilators provide a flexible and comprehensive descrip-tion of the model-dependent behavior of extended images:this allows for several unique features, including the designof single-frequency annihilators that allow for multi-scaleinversions using extended images. The extended-imageframework allows for nonlinear, finite-frequency inversionswhere both the extended images as well as the annihilatorsthemselves are model dependent. Finally, we propose aninversion metric that jointly uses data- and image-domainmisfit, and show that extended images provide additionalinformation and potentially reduce ambiguity in the com-putation of model gradients.

Ivan [email protected]

Clement FleuryColorado School of [email protected]

Daniel MacedoUniversity of [email protected]

MS22Fast Algorithms for Second Order TV and FourthOrder Mean Curvature Denoising

Image denoising is a fundamental task, studied by vari-ous researchers with a large range of variational models.The most well-known are the ROF model of Rudin-Osher-Fatimi 1992 and the PM model by Perona-Malik 1990..While the indirect method of a split-Bregman implemen-tation for the total variation model is effective, we reviewthe fast and direct methods for the primal and dual mod-els. The mean curvature model provides a powerful alter-native to the TV as it can both preserve both edges andsmooth regions. But the corresponding fourth order PDEis more challenging to solve than the TV case. We discussrecent algorithmic developments in new fix-point methodsand multigrid methods. Joint work with C Brito, L Sun,Y Q Dong, M Hintermuller, J Savage, F L Yang, B Yu.


Carlos Brito

IS12 Abstracts 77

Universidad Autonoma de [email protected]

Li SunLanzhou [email protected]

MS22Exact Reconstruction of Damaged Color Imagesusing a Total Variation Model

Variational methods based on the minimization of certainenergy functionals have been successfully employed to treata fairly general class of image restoration problems. Theunderlying theoretical challenges are common to the vari-ational formulation of problems in other areas (e.g. mate-rials science). Here first order RGB variational problemsfor recolorization will be analyzed.

Irene FonsecaCarnegie Mellon UniversityCenter for Nonlinear [email protected]

Giovanni LeoniCarnegie Mellon [email protected]

Francesco MaggiUniversity of Florence, [email protected]

Massimiliano MoriniUniversita degli Studi di ParmaParma, [email protected]

MS22Accurate Numerical Schemes in Image Denoising

In this talk, we review well-known numerical schemes tosolve the Rudin-Osher-Fatemi (ROF) model and discussabout difficulties to obtain accurate discretization for ob-taining a piecewise constant image. Based on analyticalview point of ROF model in an infinite dimension, weuse the pre-dual formulation to accurately achieve a so-lution in BV space, which is very robust to rotational in-variance. We also demonstrate comparison with variousefficient schemes.

Jooyoung HahnMathematics and Scientific ComputingUniversity of Graz, [email protected]

Carlos N. RautenbergDepartment of Mathematics and Scientific ComputingKarl-Franzens University of [email protected]

Michael HintermuellerHumboldt-University of [email protected]

MS22Variational Models for Denoising of Images andtheir Functional Analytic Properties

Variational models have proven useful for the denoising ofnatural images. They usually consist of a fidelity term anda regularization term. The two terms are weighted by a reg-ularization parameter. Examples include the Rudin-Osher-Fatemi model and higher-order models. We compare recov-ery and stability properties of various variational models.This includes the impact of noise in the data, consistencyproperties of the models and the numerical computabilityof approximate solutions.

Barbara ZwicknaglCarnegie-Mellon [email protected]

Rustum ChoksiDepartment of MathematicsMcGill [email protected]

Irene FonsecaCarnegie Mellon UniversityCenter for Nonlinear [email protected]

MS23Iterative Regularization of Inverse Problems by Ef-ficient Inexact Uzawa Methods

The aim of this talk is to study convergence properties ofinexact Uzawa methods applied to linear ill-posed prob-lems. Augmented Lagrangian and Uzawa-type methodshave recently received increasing attention as algorithmsfor edge-preserving or sparsity-promoting regularizationsin image analysis. The inversion of forward operators canbe avoided in inexact Uzawa approaches. The efficiency interms of iteration number and computation time is relatedto splitting and preconditioning techniques. We study po-tential applications of the algorithm in biomedical imageanalysis.

Christoph BruneDepartment of MathematicsUniversity of California Los [email protected]

Klaus FrickUniversity of [email protected]

Martin BurgerUniversity of MuensterMuenster, [email protected]

MS23Title Not Available at Time of Publication



78 IS12 Abstracts

MS23Bregmanized Operator Splitting with VariableStepsize: Convergence Analysis and MRI Appli-cation

This paper develops a Bregman operator splitting algo-rithm with variable stepsize (BOSVS) for solving problemsof the form minφ(Bu) + 1/2‖Au − f‖2

2, where φ may benonsmooth. The original Bregman Operator Splitting al-gorithm employed a fixed stepsize, while BOSVS uses a linesearch to achieve better efficiency. These schemes are appli-cable to total variation (TV)-based image reconstruction.The stepsize rule starts with a Barzilai-Borwein (BB) step,and increases the nominal step until a termination condi-tion is satisfied. The stepsize rule is related to the schemeused in SpaRSA (Sparse Reconstruction by Separable Ap-proximation). Global convergence of the proposed BOSVSalgorithm to a solution of the optimization problem is es-tablished. BOSVS is compared with other operator split-ting schemes using partially parallel magnetic resonanceimage reconstruction problems.

Maryam YashtiniUniversity of [email protected]

William HagerDepartment of MathematicsUniversity of [email protected]


Yunmei ChenUniversity of [email protected]

MS23Smoothed Sparse Optimization

It is well known that minimizing ‖x‖1 subject to Ax = btends to give a sparse solution while minimizing ‖x‖2

2 givea dense solution. We show that minimizing the augmentedL1 objective ‖x‖1 + 1

2α‖x‖2

2, where 12α

is an appropri-ately small weight, can efficiently recover sparse vectorswith provable recovery guarantees that are similar to thoseknown for pure L1 minimization. For compressive sensing,the above augmented L1 minimization has exact and stablerecovery guarantees given in terms of the null-space prop-erty, restricted isometry property, spherical section prop-erty, and RIPless property of the sensing matrix A. Most ofthese properties also hold for minimizing ‖X‖∗+ 1

2α‖X‖2

F ,for recovering low-rank matrices. Moreover, we show thatthe dual problems of various augmented L1 minimizationproblems, after simplification, are convex, unconstrained,and differentiable; hence, they enjoy a rich set of classicaltechniques such as gradient descent, line search, Barzilai-Borwein steps, quasi-Newton methods, and Nesterovs ac-celeration, which do not directly apply to L1 minimizationor its dual. We show that the gradient descent iteration isglobally linearly convergent, and we give an explicit rate.This is the first global linear convergence result among thegradient-based algorithms for sparse optimization.

Ming-Jun LaiUniversity of Georgia

Department of [email protected]

Wotao YinRice [email protected]

MS25Implementation of a Block Decomposition Algo-rithm for Solving Large-scale Conic OptimizationProblems

A recent work by Monteiro and Svaiter studied the itera-tion complexity of block decomposition methods for solvingmonotone variational inequalities and convex optimizationproblems. In this talk we review these methods and theircorresponding complexity bounds. We also report very en-couraging computational results comparing our methodswith the second order algorithm SDPNAL (X. Zhao, D.Sun, and K. Toh) and the boundary point method intro-duced by J. Povh, F. Rendl, and A. Wiegele. The resultsobtained on a varied collection of large-scale conic prob-lems consisting of both nonnegative vector and/or positivesemidefinite matrix variables are quite promising.

Renato Monteiro, Camillo OrtizGeorgia Institute of [email protected], [email protected]

Benar F. SvaiterIMPA, [email protected]

MS25Primal-dual Splitting, Recent Improvements andVariants

In this work, first-order primal-dual algorithms are studiedto solve a certain class of convex optimization problemswith known saddle-point structure.

minx∈X

maxy∈Y

〈Kx, y〉 + G(x)− F ∗(y),

where X and Y are finite-dimensional vector spacesequipped with standard inner products 〈·, ·〉, K : X → Yis a bounded linear operator and G and F are lsc properconvex functions with known structure. We also discusssome variants of the basic algorithm, as well as recent im-provements based on preconditioning techniques.

Thomas PockInstitute for Computer Graphics and VisionGraz University of [email protected]

Chambolle AntoninEcole Polytechnique, [email protected]

MS25Smoothing and First Order Methods: A UnifiedFramework

We propose a unifying framework that combines smooth-ing approximation with fast first order algorithms for solv-ing nonsmooth convex minimization problems. We provethat independently of the structure of the convex nons-mooth function involved, and of the given fast first order

IS12 Abstracts 79

iterative scheme, it is always possible to improve the com-plexity rate and reach an O(ε−1) efficiency estimate forthe original nonsmooth problem, by solving an adequatelysmoothed approximation counterpart. Our approach relieson the combination of the notion of smoothable functionsthat we introduce with a natural extension of the Moreau-infimal convolution technique along with its connection tothe smoothing mechanism via asymptotic (recession) func-tions. This allows for clarification and unification of severalissues on the design, analysis, and the potential applica-tions of smoothing methods when combined with fast firstorder algorithms.

Marc TeboulleTel Aviv UniversitySchool of Mathematical [email protected]

Amir BeckTECHNION - Israel Institute of [email protected]

MS26Aggregation Methods for Optical Flow Computa-tion

Global variational methods for optical flow estimation suf-fer from an over-smoothing effect due to the use of coarse-to-fine schemes. We propose a semi-local estimation frame-work designed to integrate and improve any variationalmethod. The idea is to implicitly segment the minimizationdomain into coherently moving patches. First, semi-localvariational estimations are performed in overlapping squarepatches. Then, a global discrete optimization, based on anaggregation scheme and not prone to over-smoothing, se-lects for each pixel the optimal motion vector from the onesestimated at the preceding stage. The overall computationframework is simple and can be straightforwardly paral-lelized. Experiments demonstrate that this novel approachyields better results than the baseline global variationalmethod: more accurate registration is globally achievedand motion discontinuities are sharpened.

Charles KervrannINRIA Rennes - Bretagne Atlantique / MIA - INRARennes, [email protected]

MS26The Pat(c)h Ahead: A Wide-Angle View of ImageFiltering

Filtering 2- and 3-D data (i.e. images and video) is fun-damental and common to many fields. From graphics,machine vision and imaging, to applied mathematics andstatistics, important innovations have been introduced inthe past decade which have advanced the state of the artin applications such as denoising and deblurring. Whilethe impact of these contributions has been significant, lit-tle has been said to illuminate the relationship betweenthe theoretical foundations, and the practical implementa-tions. Furthermore, new algorithms and results continueto be produced, but generally without reference to fun-damental statistical performance bounds. In this talk, Iwill present a wide-angle view of filtering, from the prac-tical to the theoretical, and discuss performance analysisand bounds, indicating perhaps where the biggest (if any)improvements can be expected. Furthermore, I will de-scribe various broadly applicable techniques for improving

the performance of any given denoising algorithm in themean-squared error sense, without assuming prior informa-tion, and based on the given noisy data alone.


MS26Poisson Noise Reduction with Non-local PCA

Photon limitations arise in spectral imaging, nuclearmedicine, astronomy and night vision. The Poisson dis-tribution used to model this noise has variance equal to itsmean so blind application of standard noise removals meth-ods yields significant artifacts. Recently, overcomplete dic-tionaries combined with sparse learning techniques havebecome extremely popular in image reconstruction. Theaim of the present work is to demonstrate that for the taskof image denoising, nearly state-of-the-art results can beachieved using small dictionaries only, provided that theyare learned directly from the noisy image. To this end,we introduce patch-based denoising algorithms which per-forms an adaptation of PCA (Principal Component Anal-ysis) for Poisson noise. We carry out a comprehensive em-pirical evaluation of the performance of our algorithms interms of accuracy when the photon count is very low. Theresults reveal that, despite its simplicity, PCA-flavored de-noising appears to be competitive with other state-of-the-art denoising algorithms. Joint work with C-A. Deledalle,R. Willet and Z. Harmany.

Joseph [email protected]

Charles-Alban DeledalleUniversite Paris Dauphine, [email protected]

Rebecca WillettDuke [email protected]

Zachary T. HarmanyDepartment of Electrical and Computer EngineeringDuke [email protected]

MS26Getting It Right: Parameter Selection ForNon-Local Means Using Sure

Non-local means (NLM) is an effective denoising methodthat applies adaptive averaging based on similarity be-tween neighborhoods in the image. An attractive way toboth improve and speed-up NLM is by first performinga linear projection of the neighborhood. One particularexample is to use principal components analysis (PCA)to perform dimensionality reduction. However, tuning thevarious parameters of the algorithm is a difficult problemdue to their data dependency; e.g., width of the smooth-ing kernel, size of the neighborhood, dimensionality of thesubspace when using PCA. To tackle this problem, we havederived an explicit analytical expression for Stein’s unbi-ased risk estimate (SURE) in the context of NLM withlinear projection of the neighborhoods. The SURE-NLMallows to monitor the MSE for restoration of an image cor-

80 IS12 Abstracts

rupted by additive white Gaussian noise without knowl-edge of the noise-free signal. Moreover, the SURE comeswith low computational cost. The SURE-NLM can thenbe used to optimize the parameters by combining it witha search algorithm in the parameter space. We propose analternative based on the principle of linear expansions ofmultiple NLMs, each with a fixed parameter set, for whichthe optimal weights can be found by solving a linear sys-tem of equations. The experimental results demonstratethe accuracy of the SURE and its successful application totune the parameters for NLM.

Dimitri Van De VilleBiomedical Imaging GroupEcole Polytechnique Federale de [email protected]

Michel KocherEcole Polytechnique Federale de [email protected]

MS27Fast Algorithms for 4D Biophysically-constrainedImage Registration

In my talk, I will discuss constrained variational meth-ods for image registration. Variational methods have beenquite successful in medical imaging, in particular in reg-istration and segmentation. Examples include snakes, de-formable registration, and level set methods. Biophysicallyconstrained variational methods are image processing tech-niques that use prior knowledge related to the physiology(mechanics, biochemistry, electrophysiology) of the imagedtissue. The target application I consider in this talk is animage registration problem of MR images of patients withbrain tumors. The goal of the image analysis is the char-acterization of the tumor aggressiveness and the construc-tion of statistical atlases. The biophysical constraints arerelated to conservation laws that model soft tissue defor-mations and tumor growth. I will describe the details ofthe algorithms used in the analysis and present results onsynthetic and clinical datasets.

George BirosUniversity of Texas at Austinbiros@ices. utexas.edu

MS27Numerical Techniques for Structural and Joint In-version

When recovering a model function from noisy incompletedata, the problem is typically ill-posed, i.e. the solution isnot unique and discontinuous with respect to data pertur-bations. Here, we propose a novel method for incorporat-ing structural information into the model. We assume thatthe structure s(x) is aligned with the model m(x), and dis-cuss the incorporation of structure in inverse problems andjoint inversion, while using stable and efficient numericalimplementation of such algorithms.

Eldad Haber, Michal Holtzman GazitDepartment of MathematicsThe University of British [email protected], [email protected]

MS27Large scale Imaging on Current Many-Core Plat-

forms

Today it is possible to solve medium size 3D or 4D imagingproblems on modern computer architectures like GPUs ina few seconds. For the class of variational methods thatrequire often the solution of a global, nonlinear optimiza-tion problem we show efficient parallel solution strategies,where a core part are numerical solvers like multigrid. Wehave incorporated imaging applications like image registra-tion in our WaLBerla framework and show results on 3Dmedical data sets.

Harald KoestlerUniversity of [email protected]

MS27Image-Based Modelling of Brain Tumour Progres-sion: From Individualization to Priors for Non-Rigid Image Registration

In the present talk we will discuss modelling brain tumourprogression on a tissue level. Such models have not onlybeen designed to enable prediction of growth patterns in in-dividual patients but also have evolved to a powerful tool toassist non-rigid image registration. We will discuss differ-ent flavours of models against such an area of application.As for the plain modelling approaches we will in particularfocus on challenges posed to model individualization.

Andreas MangUniversity of [email protected]

Thorsten M. BuzugUniversity of LuebeckInstitute of Medical [email protected]

MS28Respiratory Signal Extraction from X-ray Projec-tion Images in 4D Cone Beam CT

Patient respiratory signal is useful for the reconstructionof 4D cone beam CT. This talk will present a method toextract this signal from x-ray projection images. We no-ticed that these projections correspond to points along ahelical trajectory on a torus surface. Principle componentanalysis is utilized on projections close to each other to finda signal corresponds to these projections. Global signal isobtained by combining all local signals.


Hao Yan, Xiaoyu WangDepartment of Radiation [email protected], [email protected]

Wotao YinRice [email protected]


IS12 Abstracts 81

MS28Rosian Noise Removal on HARDI Data in MedicalImaging

High Angular Resolution Diffusion Imaging (HARDI) dataof the brain contain noise that prevents us from accuratelyextracting fiber pathways in the brain. In this work, wepresent a variational denoising model to denoise HARDIdata corrupted by Rician noise. Our proposed model con-siders the Rician distribution of noise explicitly suggestedby the underlying structure. Existence of a minimizeris discussed and numerical experiments are performed onthree types of data: 2D synthetic data, 3D diffusion-weighted Magnetic Resonance Imaging (DW-MRI) data ofa hardware phantom containing synthetic fibers, and 3Dreal HARDI data. Experiments show that our model iseffective for denoising HARDI-type data while preservingimportant aspects of the fiber pathways such as fractionalanisotropy (FA) and the orientation distribution functions(ODF).

Melissa TongUniversity of California, Los [email protected]

Yunho KimDepartment of MathematicsUniversity of California, [email protected]


MS28Laplace-Beltrami Eigen-Geometry and Applica-tions to 3D Medical Imaging

Rapid development of 3D data acquisition technologiesstimulates researches on 3D surface analysis. Intrinsic de-scriptors of 3D surfaces are crucial to either process or ana-lyze surfaces. In this talk, I will present our recent work on3D surfaces analysis by using Laplace-Beltrami (LB) eigen-system. The intrinsically defined LB operator provides usa powerful tool to study surface geometry through its LBeigen-system. By combining with other variational PDEson surfaces, I will show our results on skeleton construc-tion, feature extraction, pattern identification and surfacemapping in 3D brain imaging by using LB eigen-geometry.The nature of LB eigen-system guarantee that our methodsare robust to surfaces rotation and translation variations.

Rongjie LaiMathematics, University of Southern [email protected]

MS28Conformal Metric Optimization onSurface (CMOS) for Deformation and Mapping inLaplace-Beltrami Embedding Space

In this work we develop a novel technique for surface de-formation and mapping in the high-dimensional Laplace-Beltrami embedding space. The key idea of our work isto realize surface deformation in the embedding space viaoptimization of a conformal metric on the surface.

Yonggang Shi

LONI, CCBUCLA School of [email protected]

Rongjie LaiMathematics, University of Southern [email protected]

Arthur TogaUCLALaboratory of Neuro [email protected]

MS29Velocity Continuation

Seismic reflection imaging, also known as seismic migra-tion, requires estimation of velocity parameters. Velocitycontinuation is a process of seismic image transformationwith changes in migration velocity. In the simplest caseof zero-offset continuation, the process can be describedby a hyperbolic PDE, extended recently to the 3-D mul-tiazimuth case. In more complex cases (prestack velocitycontinuation in common-offset, common-shot or common-angle domain), the process requires pseudo-differential op-erators. I will describe novel extensions of the velocitycontinuations theory to a number of important practicalcases as well as novel numerical methods for implement-ing velocity continuation operators in practice. A lowrankapproximation of the wave extrapolation symbol leads toefficient spectral and finite-difference methods for practicalvelocity continuation.

Sergey FomelUniversity of Texas at [email protected]

MS29Time-Lapse Image-Domain Wavefield Tomography

Time-lapse seismic analysis is very successful at captur-ing detailed reservoir changes (e.g., pressure, saturation,fluid flow). Conventional 4D analyses involve linearizedinversion about constant baseline models. However, lin-earity is violated where reservoirs are significantly al-tered by production/injection that generate complex 4Dcoda. These situations necessitate more robust depth-based wave-equation imaging and velocity analyses. Wedemonstrate this approach in a synthetic CO2 geoseques-tration experiment by inverting for slowness perturbationsintroduced by injecting a thin CO2 layer.

Jeffrey C. Shragge, David LumleyUniversity of Western [email protected], [email protected]

MS29Applying Gauss-Newton and exact Newton meth-ods to Full Waveform Inversion

In the past years, Full Waveform Inversion (FWI) has beenmainly based upon preconditioned non-linear conjugategradient methods. We investigate here the application ofGauss-Newton and exact Newton methods to FWI, in aNewton-Krylov matrix free framework. This is achievedthrough second order adjoint state method formula for thecomputation of Hessian-vector products. Then two ques-tions are considered: How to find adaptive stopping crite-

82 IS12 Abstracts

rion for the inner Krylov solver? How to efficiently precon-dition the inner linear systems.

Jean VirieuxISTerre, Universiy Joseph Fourier, [email protected]

Ludovic Metivier, Romain Brossier, Stephane OpertoGrenobleludovic mtivier ¡[email protected], brossierromain ¡[email protected]., operto stphane¡[email protected]¿



Fons ten [email protected]

MS30Online Subspace Estimation and Tracking from In-complete and Corrupted Data

Low-dimensional linear subspace approximations to high-dimensional data are useful for handling problems of esti-mation, detection and prediction, with applications such asnetwork monitoring, collaborative filtering, object trackingin computer vision, and environmental sensing. Corruptedand missing data are the norm in many high-dimensionalsituations, not only because of errors and failures, but be-cause it may be impossible to collect all the interestingmeasurements or impractical to compute in real-time us-ing all the data available. Recently developed algorithmsfor ”Matrix Completion” and ”Robust PCA” have offeredtools to find low-dimensional approximations to data evenwhen data are missing and corrupted. However, these al-gorithms operate on all the available data at once andassume that the underlying low-dimensional subspace re-mains constant. In this talk I will describe two alternatives,a subspace tracking algorithm called GROUSE (Grassman-nian Rank-one Update Subspace Estimation) and its ro-bust counterpart GRASTA (Grassmannian Robust Adap-tive Subspace Tracking Algorithm). Both are incremen-tal algorithms that process one incomplete and/or cor-rupted measurement vector at a time. Thus GROUSEand GRASTA operate with considerably less computationthan the other algorithms, while at the same time allowingflexibility for real-time applications such as tracking andanomaly detection. I will present these algorithms, discusstheir relationship to LMS subspace tracking algorithms fa-miliar to those in signal processing, and show their appli-cation to matrix completion, robust PCA, and backgroundand foreground separation in video surveillance.

Laura Balzano3610 Engineering HallUniversity of [email protected]

Jun HeNanjing University of Information Science and [email protected]

Arthur SzlamDepartment of MathematicsNYU

[email protected]

Brian ErikssonBoston [email protected]

Benjamin RechtUniversity of Wisconsin – [email protected]

Robert NowakUniversity of WisconsinElectrical and Computer [email protected]

MS30On the Use of Low-rank Regularization for Learn-ing Graphical Models with Missing Data

Gaussian graphical models have been quite popular formodeling joint distributions over real-valued random vari-ables, where the association structure among the randomvariables is also of interest. They are specified by an undi-rected graph, and represent the set of multivariate Gaus-sian distributions whose inverse covariance matrices havesupport matching the undirected graph. Recovering thisunderlying graph structure when given limited samples isthus an important task in many of the applications of Gaus-sian graphical models. In many settings however, and inbiological settings in particular, the data is not fully ob-served: samples have incomplete and potentially differingsets of covariates. In this talk, we consider the task of re-covering graphical model structure under such missing datasettings. We propose a novel M-estimator with regulariza-tion corresponding to sums of structure-inducing functions,where each such function encourages heterogeneous struc-ture corresponding to the sum of a low-rank and a sparsematrix. We provide strong statistical and empirical guar-antees for the performance of this M-estimator.

Pradeep RavikumarDepartment of Computer SciencesUniversity of Texas, [email protected]

MS30PCA with Outliers and Missing Data

In the standard form, PCA involves organizing data into amatrix where columns represent the points, and rows thefeatures. We consider PCA in a challenging setting wherethere are outliers (i.e. some columns are completely arbi-trary), and most data is missing (i.e. most elements of thematrix are not observed). We propose a recovery methodbased on convex optimization, and provide analytical guar-antees on when it is able to both (a) recover the low-rankmatrix, and (b) identify the outliers. These tasks are chal-lenging for other methods, because they are coupled andhave to be performed jointly.

Sujay SanghaviElectrical and Computer EnigneeringUniversity of Texas at [email protected]

IS12 Abstracts 83

MS30A Novel M-Estimator for Robust PCA

We formulate a convex minimization to robustly recovera subspace from a contaminated data set, partially sam-pled around it, and propose a fast iterative algorithm toachieve the corresponding minimum. We establish exactrecovery by this minimizer, quantify the effect of noise andregularization, explain how to take advantage of a knownintrinsic dimension and establish linear convergence of theiterative algorithm. Our minimizer is an M-estimator. Wedemonstrate its significance by adapting it to formulatea convex minimization equivalent to the non-convex totalleast squares (which is solved by PCA). We compare ourmethod with many other algorithms for Robust PCA onsynthetic and real data sets and demonstrate state-of-the-art speed and accuracy.

Teng ZhangInstitute for Mathematics and its ApplicationsUniversity of [email protected]

Gilad LermanDepartment of MathematicsUniversity of [email protected]

MS31A Primal Dual Method for Solving a Convex Modelfor DOAS Analysis

We propose a variational model for differential optical ab-sorption spectroscopy (DOAS) analysis that differs fromprevious approaches in that it is a convex model which alsocouples the estimation of fitting coefficients, backgroundsubtraction, wavelength misalignment and noise. DOAS isa technique for using Beer’s Law to estimate concentrationsof gases by fitting a combination of their characteristic ab-sorption spectra to light intensity measurements taken overa range of wavelengths. We show how to efficiently solveour model using primal dual methods. Additionally, weuse this example to illustrate the importance of precondi-tioning for improving the rate of convergence.

Ernie EsserUniversity of California IrvineUC Irvine Math [email protected]

Jack XinDepartment of [email protected]

MS31Nonsmooth Image Reconstruction Algorithms andTheir Applications in Partially Parallel MR Imag-ing

Image reconstructions with nonsmooth regularization (e.g.total variation) are capable to recover intrinsic sparsityof images for well preserved edges etc. We introduce aseries of numerical algorithms that can effectively solvethese nonsmooth minimization problems, which may in-volve much general data fitting terms with arbitrary sens-ing matrices and non-Gaussian noises. Results on clinicalpartially parallel MRI are shown to demonstrate their ef-fectiveness. Convergence analysis and extensions to non-

convex regularization will be discussed.


Yunmei Chen, William HagerDepartment of MathematicsUniversity of [email protected], [email protected]



Jing YuanDepartment of Computer ScienceUniversity of Western [email protected]

MS31A Primal Dual Method for Nonconvex Problemsand Applications

We show the primal dual method studied in E. Esser,X. Zhang, and T. F. Chan, A General Framework for aClass of First Order Primal- Dual Algorithms for Con-vex Optimization in Imaging Science, SIAM J. ImagingSci. Volume 3, Issue 4, pp. 1015-1046, 2010, which isan efficient method for minimizing large non-differentiableconvex functions, can also be a practical method for mini-mizing certain large non-differentiable nonconvex functionsfor which the nonconvex component is smooth and can bemade strongly convex by the addition of a quadratic. Wefocus on applications where we want to increase the spar-sity of a variable that is constrained to be nonnegative andsum to one. Instead we choose to encourage sparsity byadding a concave quadratic function to an otherwise con-vex function. In particular, we show how to apply thisstrategy to multiphase segmentation and nonlocal patch-based inpainting problems.

Xiaoqun ZhangShanghai Jiaotong [email protected]

Ernie EsserUniversity of California IrvineUC Irvine Math [email protected]

MS32Inverse Transport Problems in Remote SensingApplications

We consider the reconstruction of inclusions in a scatter-ing atmosphere from remote measurements of light, whosepropagation is modeled by a transport equation. This ill-posed problem is solved using a Bayesian framework toprescribe measurement errors and prior assumptions. Theposterior distribution is sampled by using a novel MarkovChain Monte Carlo called MC3. We use a Monte Carlopath recycling strategy to significantly speed up costlyforward transport simulations. Numerical simulations are

84 IS12 Abstracts

presented.

Guillaume BalColumbia UniversityDepartment of Applied Physics and Applied [email protected]

Anthony B. DavisJet Propulsion LaboratoryCalifornia Institute of [email protected]

Ian LangmoreColumbia [email protected]

Youssef M. MarzoukMassachusetts Institute of [email protected]

MS32Edge-preserving Super-resolution via Blind Decon-volution and Upsampling

We consider the super-resolution model introduced by S.J.Osher and this author in J. Sci. Comput., 2008 consist-ing of the solution of a total-variation based variationalproblem that solves a linear degradation model involvinga convolution operator and a down- sampling operator.In this research work we explore different edge preservingup/down-sampling operators with different orders of spa-tial accuracy and different convolution operators to removemotion blur from degraded low resolved images. Some nu-merical examples are provided to show the features of theproposed algorithm.

Antonio MarquinaUniversity of Valencia, [email protected]

MS32Patch-based Locally Optimal Denoising

We formulate the fundamental limits of image denoisingusing a statistical framework. Next, we propose a practi-cal algorithm with the explicit aim of achieving the bound.The proposed method is a patch-based Wiener filter thattakes advantage of both geometrically and photometricallysimilar patches. The resultant approach has a solid sta-tistical foundation while producing denoising results thatare comparable to or exceeding the current state-of-the-art,both visually and quantitatively.


Priyam ChatterjeeUniversity of California, Santa [email protected]

MS32Texture Adaptive Image Restoration Using Frac-tional Order Regularization

We present an adaptive strategy for the restoration of im-ages contaminated by blur and noise that allows to preserve

textures. Total Variation regularization has good perfor-mance in noise removal and edge preservation but lacksin texture restoration. According to a texture detectionstrategy, we apply fractional-integer order diffusion. Thiscan be related to a weighted norm regularization. A fastalgorithm based on the half-quadratic technique is used.Numerical results show the effectiveness of our strategy.

Fiorella SgallariUniversity of BolognaDepartment of [email protected]

Raymond H. ChanThe Chinese Univ of Hong KongDepartment of [email protected]

Alessandro Lanza, Serena MorigiUniversity of Bologna, [email protected], [email protected]

MS33Forward-backward Splitting and Proximal Alter-nating Methods for Nonconvex Nonsmooth TameProblems

In view of the minimization of a nonsmooth nonconvexfunction f , we present an abstract convergence result fordescent methods satisfying a sufficient-decrease assump-tion, and allowing a relative error tolerance. Our resultguarantees the convergence of bounded sequences, underthe assumption that the function f satisfies the Kurdyka-Lojasiewicz inequality. This general result allows to covera wide range of problems, including nonsmooth semi-algebraic (i.e. polynomial-like) or tame minimization. Thespecialization of our result to different kinds of structuredproblems provides several new convergence results for in-exact versions of the forward-backward splitting algorithm,gradient projection method, some proximal-like Gauss-Seidel methods, etc.. Our results are illustrated throughfeasibility problems, or iterative thresholding proceduresfor compressive sensing.

Jerome BolteUniversite Toulouse [email protected]

MS33Combinatorial Selection and Least AbsoluteShrinkage via the CLASH Algorithm

The least absolute shrinkage and selection operator (Lasso)for linear regression exploits the geometric interplay of the2-data error objective and the 1-norm constraint to arbi-trarily select sparse models. Guiding this uninformed se-lection process with sparsity models has been precisely thecenter of attention over the last decade in order to improvelearning performance. To this end, we alter the selectionprocess of Lasso in this paper to explicitly leverage combi-natorial sparsity models (CSMs) via the combinatorial se-lection and least absolute shrinkage (CLASH) algorithm. Ahighlight is the introduction of a new algorithmic definitionof CSMs, which we dub as the Polynomial time Modularε-Approximation Property (PMAPε). PMAPε enables usto determine the impact of approximate combinatorial pro-jections within CLASH. We then provide concrete guide-lines how to leverage sets with PMAPε within CLASH, andcharacterize CLASH’s estimation guarantees as a function

IS12 Abstracts 85

of ε as well as the set restricted isometry constants of theregression matrix. Finally, we present experimental resultsusing both simulated and real world data to demonstratethe effectiveness of CLASH.

Volkan [email protected]

MS33A Memory Gradient Algorithm for Noncon-vex Regularization with Applications to ImageRestoration

A memory gradient algorithm for non-convex regulariza-tion with applications to image restoration. Abstract: Inthis work, we consider a class of non-convex differentiablepenalty functions for image recovery problems. The reg-ularization is applied to a linear transform of the targetimage. As special cases, it includes edge preserving mea-sures or frame analysis potentials. We develop a newMajorize-Minimize memory gradient algorithm for solvingthe non-convex minimization problem. The use of proxi-mal tools allows us to consider the case of non-differentiableterms arising in the objective function. The convergenceof the algorithm is investigated by using recent results innon-convex optimization. The fast convergence proper-ties of the proposed optimization algorithm are illustratedthrough image processing examples.

Emilie ChouzenouxUniversite Paris-Est [email protected]

MS33High-order Methods for Sparse Signal Recovery

Forward-backward splitting is a common way of solvingnon-differentiable problems. In this talk, we introduce anacceleration step that can be added to forward-backwardsplitting which implicitly performs conjugate gradient ac-celeration. In most situations, this method finds the ex-act solution to an L1 regularized minimization in a finitenumber of steps. This is done by exploiting the fact that,when recovering a sparse signal, the basis pursuit problemis equivalent to a quadratic minimization problem involv-ing a low-rank matrix. The novel method takes advan-tage of this property by applying a conjugate gradient par-tan scheme to the underlying low-rank problem. This newalgorithm has several advantages over other basis-pursuitschemes. In particular, it requires no time-step parametersas input, and the speed of the algorithm is almost com-pletely insensitive to the condition number of the problem.This allows for the fast solution of basis pursuit problemsinvolving convolution, orthogonal, and random matrices.

Tom GoldsteinStanford [email protected]

MS34A Generalized Tree-Based Wavelet Transform andIts Application to Patch-Based Image Denoising

What if we take all the overlapping patches from a givenimage and organize them to create the shortest path by us-ing their mutual distances? This suggests a reordering ofthe image pixels in a way that creates a maximal 1D reg-

ularity. Could we repeat this process in several ’scales’?What could we do with such a construction? In this talkwe consider a wider perspective of the above line of ques-tions: We introduce a wavelet transform that is meant fordata organized as a connected-graph or as a cloud of high-dimensional points. The proposed transform constructs atree that applies a 1D wavelet decomposition filters, cou-pled with a pre-reordering of the input, so as to best spar-sify the given data. We adopt this transform to image pro-cessing tasks by considering the image as a graph, whereevery patch is a node, and vertices are obtained by Eu-clidean distances between corresponding patches. State-of-the-art image denoising results are obtained with theproposed scheme. Joint work with Idan Ram and IsraelCohen, Electrical Engineering Department - Technion.


MS34Matching By Tone Mapping

A fast pattern matching scheme termed Matching by ToneMapping (MTM) is introduced which allows matching un-der non-linear tone mappings. We exploit the recentlyintroduced Slice Transform to implement a fast compu-tational scheme requiring computational time similar tothe fast implementation of Normalized Cross Correlation(NCC). In fact, the MTM measure can be viewed as ageneralization of the NCC for non-linear mappings and ac-tually reduces to NCC when mappings are restricted to belinear. The MTM is shown to be invariant to non-lineartone mappings, and is empirically shown to be highly dis-criminative and robust to noise.

Yacov Hel-OrSchool of Computer ScienceThe Interdisciplinary Center, [email protected]

MS34Analysis of Image Patches: A Unified GeometricPerspective

Recent work in computer vision and image processing in-dicates that the elusive quest for the ”universal” trans-form has been replaced by a fresh perspective. Indeed,researchers have recently proposed to represent images asa ”collage” of small patches. The patches can be shapedinto square blocks or into optimized contours that mimicthe parts of a jigsaw puzzle. These patch-based appear-ance models are generative statistical models that can belearned from an image or a set of images. All these patch-based methods implicitly take advantage of the followingfact: the dataset of patches, whether it is aggregated froma single image, or a library of images, is a smooth and lowdimensional structure. In this work, we propose a unify-ing geometric theory to explain the success of patch-basedmethod in image processing. In addition, we explain howgeometric information such as tangent plane and curvaturein patch-space can be used to process images in an optimalmanner. This is a joint work with D.N. Kaslovsky, and B.Wohlberg.

Francois G. MeyerUniversity of Colorado at BoulderUSA

86 IS12 Abstracts

[email protected]

MS34Patch Based Image Denoising With and WithoutDictionary Learning

Patch based image processing has proved to be very effec-tive for denoising. By exploiting the image nonlocal re-dundancy, some representative patch based methods suchas BM3D have achieved state-of-the-art denoising results.In this talk, we will discuss the role of dictionary learning inpatch-based image denoising (PID), and compare the de-noising performance with and without dictionary learning.In the case of PID with dictionary learning, the dictionarycan be pre-learned from clean example images, or it can belearned online from the noisy image, while the dictionarycan be orthogonal or non-orthogonal. In the case of PIDwithout dictionary learning, an analytically pre-designeddictionary such as DCT bases and wavelet/curvelet basescan be used, or the noisy patches themselves are directlyused as a dictionary. In this talk, we will present extensiveexperiments to compare the various schemes, analyze theirpros and cons, and discuss the future development of PID.

Lei ZhangHong Kong Polytechnic [email protected]

MS35Free Form Deformations with Dense Control PointSpacing for Image Registration

B-spline free form deformations (FFDs) have long beenused in image registration because they provide a natu-rally intuitive parameterization in terms of control pointdisplacements. However, the typical formulation of imageregistration as an optimization problem limits the densitywith which contol points can be defined. In this talk, weshow how formulating the image registration problem interms of an approximate solution to the Euler-Lagrangeequations over the B-spline basis enables FFD-based im-age registration that is computationally efficient when thecontrol point spacing approaches the density of the voxelspacing.

Nathan D. CahillSchool of Mathematical SciencesRochester Institute of [email protected]

MS35Matrix Factorization Techniques for MultivariateImaging Analysis

Contemporary neuroscientists typically collect a wealth ofmeasurements on relatively few subjects. Studies are there-fore usually underpowered. There is a pressing need forautomated tools that are able to identify a reduced set ofmultidimensional salient features that drive the neurobio-logical mechanisms of disease, behavior and development.This talk will discuss some multivariate methods for com-puting the statistical comparisons commonly needed bypopulation studies in medical imaging.

James GeeUniversity of PennsylvaniaDepartment of [email protected]

MS35Image Processing with Stochastic PDEs

We discuss methods based on stochastic PDEs (SPDEs)for the processing of images and image sequences with un-certain gray values resulting from measurement errors andnoise. Our approach yields a reliable precision estimate forthe image processing results. The ansatz space for such3D and 4D images identifies gray values with random vari-ables. The generalization of classical PDE image process-ing in this context leads to SPDEs which we discretize withthe generalized polynomial chaos expansion and the gener-alized spectral decomposition method.

Tobias PreusserFraunhofer MEVIS, [email protected]

Torben PatzFraunhofer [email protected]

MS354D Image Based Parameter Estimation of VascularProperties

Non-invasive estimation of some properties of arterial tis-sues such as compliance can be obtained by medical images,solving inverse fluid-structure interaction (FSI) problems.After the 4D registration of time snapshots providing thedisplacement field of the wall, the parameter of interestminimizes the misfit between data and simulation underthe constraint of the FSI problem, whose effective numer-ical solution raises several challenges. We address in par-ticular the model reduction for the FSI problem.

Luca BertagnaDept. Math & CSEmory [email protected]

Mauro PeregoDepartment of Scientific ComputingFlorida State [email protected]

Alessandro VenezianiMathCS, Emory University, Atlanta, [email protected]

MS36Progress with Classification of Cryo-EM Data byManifold Embedding and Other Techniques

Single-particle data acquired by low-dose cryo-EM ofmacromolecules often exhibit heterogeneity, due to the co-existence of different states. In our efforts to develop andapply techniques of unsupervised classification, we are fol-lowing three paths: (i) the maximum-likelihood method,(ii) the bootstrap reconstructions method, and (iii) man-ifold embedding. For manifold embedding, encouragingresults have been obtained with simulated data, but ex-perimental cryo-EM data still face difficulties due to thelarge size of the parameter space and the need for its finesampling.

Joachim Frank, Robert Langlois, Hstau LiaoColumbia [email protected], [email protected],

IS12 Abstracts 87

[email protected]

Chun Hong YoonUniversity of [email protected]

Dimitris GiannakisNew York [email protected]

Peter Schwander, Abbas OurmazdUniversity of [email protected], [email protected]

MS36Resampling-based Assessment of Variability in 3DReconstructions of Biological Molecules

In cryo-electron microscopy biological macromolecules arecaptured in their native environment and their structure ispreserved. It is thus possible to study dynamical effects ofdifferent functional states. These effects can be quantita-tively characterized by the variance/covariance matrix ofthe 3D mass distribution of the structure. Nevertheless, asthe data is available as projections of the macromoleculeformed in the electron microscope, we address the problemusing a stratified resampling technique.

Pawel A. PenczekUniversity of Texas - Houston Medical [email protected]

MS36Viewing Direction Estimation in cryo-EM UsingSynchronization

We present a noise-robust synchronization-based algorithmfor determining the unknown viewing directions of cryo-EM images. Information about the spatial relation be-tween pairs of images is first extracted using common linesbetween triplets of images, and then encoded in a spe-cially designed matrix. In the noise-free case, this matrixis shown to have rank 3, and its non-trivial eigenspace isshown to reveal the projection orientations of all images.

Yoel ShkolniskyTel Aviv [email protected]


MS36A Fourier-based Approach for Iterative 3D Recon-struction from Cryo-EM Images

In this talk, we focus on the 3D reconstruction problemfrom cryo-EM images assuming an existing estimate fortheir orientations and positions. We propose a fast andaccurate Fourier-based Iterative Reconstruction Method(FIRM) that exploits the Toeplitz structure of the compo-sition of a forward-projector and its back-projector. FIRMis fast and accurate compared with current methods. More-over, FIRM combines images from different defocus groupssimultaneously.

Lanhui Wang

Princeton UniversityThe Program in Applied and Computational [email protected]

Yoel ShkolniskyTel Aviv [email protected]


MS37Experimental Translation of Image ReconstructionMethods for Three-Dimensional Optoacoustic To-mography

In this work we implement and investigate the use of math-ematically exact and iterative image reconstruction meth-ods in 3D optoacoustic tomography (OAT). Accurate im-age models that incorporate the transducer response areutilized and the reconstruction algorithms are implementedusing graphics processing units (GPUs) to alleviate thecomputational burden involved. The algorithms are ap-plied to experimentally measured data acquired by use ofa 3D OAT scanner and are evaluated by use of task-specificimage quality metrics.

Mark Anastasio, Kun WangWashington University in St. [email protected], [email protected]

Sergey Ermilov, Richard Su, Alexander OraevskyTomoWave [email protected], [email protected],[email protected]

MS37Compressed Sensing and Photo-Acoustic Tomogra-phy

Photo-Acoustic Tomography (PAT) with dense arraysof detectors allows the reconstruction of high-resolutionthree-dimensional images of acoustic pressure correspond-ing to absorption of light in tissue. Current systems arequite slow to acquire a full data set, so dynamic studiesare limited. In this talk I will discuss a new approach foracquiring a compressed set of data and thus to allow feasi-ble reconstruction of four-dimensional images.

Simon ArridgeUniversity College [email protected]

MS37Noise in Photoacoustic Tomography: From Point-like Detectors to Large integrating Detectors andits Influence on Spatial Resolution and Sensitivity

For photoacoustic tomography point-like detectors are notideal at all: because of the small detector volume the ther-modynamic fluctuations (= noise) get high and the signalamplitude is low, which results in a bad signal-to-noise ra-tio. But for a bigger detector volume the signal is averagedover the whole detector volume, which results in blurringof the reconstructed image. To describe this trade-off be-tween spatial resolution and sensitivity for a varying detec-tor volume in a quantitative way the pressure is described

88 IS12 Abstracts

by a stochastic process.

Peter BurgholzerRECENDT, Linz, [email protected]

MS37Gradient-based Inversion for 3D Quantitative Pho-toacoustic Imaging

The main aim in quantitative photoacoustics is to recoveraccurate estimates of the chromophore concentration dis-tributions from photoacoustic images, thereby broadeningthe range of clinical and preclinical applications of pho-toacoustic imaging. The images are dependent on the tis-sue optical properties and the distribution of light, both ofwhich are dependent on the tissue chromophores, and theproblem is generally nonlinear, ill-posed and large-scale.Here we discuss the recent progress on solving this opticalinverse problem.

Teedah Saratoon, Ben CoxUniversity College [email protected], [email protected]

Tanja TarvainenDepartment of Applied PhysicsUniversity of Eastern [email protected]

Simon ArridgeUniversity College [email protected]

MS38Migration Based on Two-way Depth Extrapolation

We discuss a new family of methods for migration and ve-locity analysis based on two-way depth extrapolation ofseismic data. By using spectral projectors to avo id ill-posedness of two-way depth extrapolation, we maintainquality of imaging of steep features in arbitrary compli-cated background medium. Our method allows us to down-ward continue temporal frequency bands independently,thus opening the door for effective velocity analysis.

Gregory BeylkinUniversity of [email protected]

MS38Seismic Full Waveform Inversion, Recent Worksand Research Directions

Seismic Full Wave Inversion (FWI) is very promising for es-timating subsurface structure for O&G industry, but manyquestions are still open such as how to handle features un-explained by the forward model, the domain of integration,the choice of the initial model, the lack of low frequenciesin the data... In this presentation we will present somerecent work done in FWI and some research directions weare exploring for solving these challenging problems.

Henri CalandraTotalCSTJF, Geophysical Operations & Technology, R&[email protected]

Changsoo [email protected]

Christian RiveraTOTAL E&P [email protected]

Herve ChaurisEcole des Mines [email protected]

Alexander [email protected]

MS38Resolution Analysis in Full Waveform Inversion

We propose a new method for the quantitative resolutionanalysis in full seismic waveform inversion that overcomesthe limitations of synthetic inversions while being computa-tionally more efficient and applicable to any misfit measure.The method rests on a parametrised Hessian of the misfitfunctional that allows to extract various resolution prox-ies, including direction-dependent resolution lengths andthe posterior covariance. We illustrate our method with afull waveform inversion for Europe and western Asia.

Andreas Fichtner, Jeannot TrampertUtrecht [email protected], [email protected]

MS38Dreamlet as a Physical Wavelet for Seismic DataCompression and Imaging

Seismic data are special data sets and they cannot fill the4-D space-time in ar bitrary ways. The time-space distri-butions must observe causality which is roote d from thewave equation. Wave solutions can only exist on the lightcone in the 4D Fourier space. Dreamlet is a type of physi-cal wavelet defined on an observat ion plane (earth surfaceor a plane at depth z during extrapolation). Causality(or dispersion relation) built into the wavelet (dreamlet)and propagator is a d istinctive feature of dreamlet. Wewill discuss the dreamlet decomposition of se ismic dataand dreamlet imaging in the compressed domain by surveysinking migra tion.

Ru-Shan WuUniversity of California, Santa [email protected]

MS39Code Design for Target Tracking and Discrimina-tion

Patch based de-noising algorithms and patch manifoldsmoothing have emerged as efficient de-noising methods.This talk provides a new insight on these methods, suchas the Non Local Means or the Image Graph De-noising,by showing its use for filtering a selected pattern. Appli-cations to target tracking in compressively sensed imageryis discussed.

Ronald CoifmanYale University

IS12 Abstracts 89

Department of Computer [email protected]

MS39Stability and Robustness of Weak OrthogonalMatching Pursuits

A recent result establishing, under restricted isometry con-ditions, the success of sparse recovery via OrthogonalMatching Pursuit using a number of iterations propor-tional to the sparsity level is extended to Weak OrthogonalMatching Pursuits. The new result also applies to a PureOrthogonal Matching Pursuit, where the index appendedto the support at each iteration is chosen to maximize thesubsequent decrease of the squared-norm of the residualvector.

Simon FoucartDrexel [email protected]

MS39Adaptive Compressive Imaging Using Sparse Hier-archical Learned Dictionaries

Recent results in compressive sensing (CS) have estab-lished that many high dimensional objects can be accu-rately recovered from a relatively small number of linearprojection observations, provided that the objects possessa sparse representation in some basis. Subsequent effortshave shown that the performance of CS can be improvedby exploiting structure that may exist in the sparse repre-sentation of the object being acquired (structured spar-sity), or by some form of online focusing of measure-ments throughout the sensing process (adaptivity). Herewe discuss a powerful hybrid of these two ideas. We de-scribe a simple adaptive sensing procedure for acquiringobjects that exhibit structured sparsity characterized bylow-degree graph-based coefficient dependencies, and wedemonstrate that representations exhibiting this form ofstructured sparsity can be learned from collections of train-ing data using recently developed techniques from sparsehierarchical dictionary learning. The combination of theseideas results in an effective and resource-efficient adaptiveCS procedure, which we demonstrate in the context of com-pressive imaging.

Jarvis HauptUniversity of [email protected]

MS39Noval Measurements in Compressive Imaging

Compressive measurement focuses on making rela-tively few information-rich measurements, rather thanmany information-poor measurements; exploiting theprior knowledge that natural images are nearly al-ways sparse/compressible in some domain. Knowledge-Enhanced Compressive Measurements amplify the bene-fits of compressive measurement by incorporating into themeasurement process additional prior knowledge concern-ing (a) signal classes, (b) task requirements, and (c) adap-tation. Incorporation of signal priors can be used to ensurethat measurements do not waste resources measuring some-thing that we already know; whereas, the inclusion of taskpriors facilitates extraction of only that information mostimportant to the exploitation task. Adaptation promotes

an increasingly efficient measurement process, incorporat-ing knowledge from earlier experience.

Mark NeifeldDARPA - [email protected]

MS40Image Restoration with Total Generalised Varia-tion

We study the recently introduced total generalised varia-tion functional (TGV) of arbitrary order which constitutesa well-suited convex model for piecewise smooth images.After recalling basic properties, well-posedness of TGV-regularised ill-posed inverse problems is discussed, general-ising the results already obtained for the second order case.In particular, we examine spaces of symmetric tensor fieldsof bounded deformation which play a crucial rule for deriv-ing these existence results. Finally, some extensions of theTGV functional are proposed and illustrated by numericalexamples.

Kristian BrediesUniversity of GrazInstitute of Mathematics and Scientific [email protected]

MS40Numerical Algorithms for Eulers Elastica Modelsin Image Processing

We introduce recently developed fast and efficient numeri-cal algorithms to solve Euler’s elastica models with applica-tions to variational image denoising, image inpainting, andimage zooming. Minimization problems are reformulatedas constrained problems by a variable splitting methodand relaxation technique. The proposed constrained min-imization problems are solved by using an augmented La-grangian approach. Numerical tests are demonstrated toshow an efficiency of proposed method.

Jooyoung HahnMathematics and Scientific ComputingUniversity of Graz, [email protected]

Xue-cheng TaiDept. of Mathematics, University of Bergen, Norway andDiv. of Math. Sciences, Nanyang Tech. [email protected]

Jason Ginmo ChungDiv. of Math. Sciences, Nanyang Tech. [email protected]

Yu WangTechnion, [email protected]

Yuping DuanDiv. of Math. Sciences, Nanyang Tech. [email protected]

MS40A Variational Model for Functions of Bounded Hes-

90 IS12 Abstracts

sian - Theory, Numerics and Applications

We formulate a minimisation problem in the space of func-tions of bounded Hessian. We prove existence and unique-ness of solutions through a relaxation process and we im-plement our method numerically using a split Bregmantechnique. Our results suggest improvement in compari-son to total variation denoising and deblurring in terms ofthe staicasing effect and creation of artifacts in the restoredimage.

Kostas PapafitsorosDAMTP, University of [email protected]

MS40Infimal Convolution Regularizations with Discretel1-type Functionals

As first demonstrated by Chambolle and Lions, the stair-casing effect of the Rudin-Osher-Fatemi model can be re-duced by using infimal convolutions of functionals contain-ing higher order derivatives. In this talk, we examine amodification of such infimal convolutions in a general dis-crete setting. For the special case of finite difference matri-ces, we show the relation of our approach to the continuoustotal generalized variation recently developed by Bredies,Kunisch and Pock. Moreover, we present splitting methodsto compute the minimizers of the l22 – (modified) infimalconvolution functionals, which are superior to previouslyapplied second order cone programming methods. Finally,the differences between the ordinary and the modified infi-mal convolution approach are illustrated by numerical ex-amples.This is joint work with Simon Setzer (Saarland University)and Gabriele Steidl (University of Kaiserslautern).

Tanja TeuberUniversity of Kaiserslautern, GermanyDept. of Mathematics and Computer [email protected]

MS41The Adaptive Inverse Scale Space Method for Hy-perspectral Unmixing

Hyperspectral unmixing describes the technique of findingthe abundances (fractions) of certain endmembers (spec-tral signatures of materials) in a hyperspectral image. Wewill show that the so called inverse scale space flow yieldsimproved results for the recovery of sparse abundance mapsin comparison to 1 regularization. With the means of con-vex analysis we will develop a very efficient and adaptivenumerical method to compute the solution to this flow ex-actly without discretization.

Michael MoellerUniversity of Munster, [email protected]

Martin BurgerInstitute for Computational and Applied MathematicsUniversity of Muenster, [email protected]

Martin BenningInstitute for Computational and Applied MathematicsUniversity of [email protected]


MS41L1 based Template Matching and its Applicationto Hyper Spectral Imaging



MS41Structured Models for Inverse Problems

In this work we will describe advances in the area of hy-perspectral image analysis with sparse modeling tools. Inparticular, models for sub-pixel classification and unmixingfrom partial data are discussed, and results are presentedboth for urban and non-urban data. This is joint work withAlexey Castrodad, Zhengming Xing, John Greer, EdwardBosch, and Lawrence Carin.

Guillermo SapiroUniversity of MinnesotaDept Electrical & Computer [email protected]

MS41Local Robust Principal Components for NonlinearDatasets

A robust variant of Principal Components Analysis decom-poses a data matrix into low rank and sparse components,the former representing a low-dimensional linear model ofthe data, and the latter representing deviations from themodel. This decomposition can be useful, but is not ap-propriate when the data are not well modeled by a singlelow-dimensional subspace. We construct a new decompo-sition corresponding to a more general underlying modelconsisting of a union of low-dimensional subspaces.

James TheilerSpace and Remote Sensing SciencesLos Alamos National [email protected]

Brendt Wohlberg, Rick ChartrandLos Alamos National [email protected], [email protected]

MS42Inverse Scattering Problem for InhomogeneousMedia Containing Obstacles

We consider the scattering problem for an inhomogeneous(possibly anisotropic) dielectric media containing obstaclesinside. Such a problem arises in non-destructive testing ofinhomogeneous materials using electromagnetic interroga-tion. The goal is to obtain estimates on the material prop-erties of the inhomogeneity and/or the geometrical prop-erties of embedded obstacles. More specifically, we willdiscuss associated eigenvalue problems and show that thecorresponding eigenvalues can be computed from the scat-

IS12 Abstracts 91

tering data and provide information on the inhomogeneity-obstacle structure.

Fioralba CakoniUniversity of [email protected]

Anne CossonniereCERFACS, [email protected]

Houssem HaddarEcole [email protected]

MS42A Mixed Formulation of the Quasi-reversibilityMethod to Solve Elliptic Cauchy Problems in 2dand 3d

We propose a new mixed formulation of the quasi-reversibility method to solve bidimensional and tridimen-sional Cauchy problems. Unlike the standard formula-tion of quasi-reversibility, this new formulation can be dis-cretized using standard finite element spaces, which easesits numerical implementation. We illustrate the functional-ity of the method with the help of numerical experiments.

Jeremi DardeAalto [email protected]

Antti HannukainenAalto [email protected]

Nuutti HyvonenAalto UniversityDepartment of Mathematics and Systems [email protected]

MS42Inverse Source Problems and the WindowedFourier Transform

The reconstruction of time-harmonic acoustic or electro-magnetic sources from measurements of the far field of thecorresponding radiated wave is an ill-posed inverse prob-lem with fascinating applications. Based on the observa-tion that the windowed Fourier transform of the radiatedfar field is related to an exponential Radon transform ofa smoothed approximation of the source, we present a fil-tered backprojection algorithm to recover information onthe unknown source. We discuss this algorithm and presentnumerical results.

Roland GriesmaierInstitute of MathematicsUniversity of [email protected]

Martin HankeJoh. Gutenberg-UniversitatMainz, [email protected]

Thorsten RaaschUniversity of Mainz

[email protected]

MS42A Monotony Based Method for ElectricalImpedance Tomography

The aim of Electrical Impedance Tomography is to re-construct the spatially perturbed conductivity of an ob-ject from current-voltage measurements on its boundary.We consider the localization of inhomogeneous conduct-ing areas surrounded by a constant background conduc-tivity. The introduced reconstruction method bases on awell-known monotony relation combined with a conceptof localized potentials. We specially concentrate on fastimplementations that stem from the recent result that lin-earization does not lead to shape errors.

Marcel Ullrich, Bastian HarrachUniversitat [email protected],[email protected]

MS43Analysis of a Variational Framework for ExemplarBased Image Inpainting

We discuss a variational framework for exemplar based im-age inpainting, in which both the reconstructed image uand the non-local weights w are treated as variables andupdated using the variational formulation. This approachpermits to draw relations with some existing inpaintingschemes, in addition to leading to novel ones. We carryout an analysis of two of these schemes providing existenceof the solution, regularity of the minima and convergenceof the alternating scheme for the variables (u,w).

Gabriele FaccioloENS [email protected]

MS43Patch Foveation in Nonlocal Imaging

Patch-based nonlocal imaging methods rely on the assump-tion that natural images contain a large number of mutu-ally similar patches at different locations within the image.Patch similarity is typically assessed through the Euclideandistance of the pixel intensities and therefore depends onthe patch size: while large patches guarantee stability withrespect to degradations such as noise, the mutual similar-ity that can be verified between pairs of patches tends toreduce as the patch size grows. Thus, a windowed Eu-clidean distance is commonly used to balance these twoconflicting aspects, assigning lower weights to pixels farfrom the patch center. We propose patch foveation as analternative to windowing in nonlocal imaging. Foveationis performed by a spatially variant blur operator, char-acterized by point-spread functions having bandwidth de-creasing with the spatial distance from the patch center.Patch similarity is thus assessed by the Euclidean distanceof foveated patches, leading to the concept of foveated self-similarity. In contrast with the conventional windowing,which is only spatially selective and attenuates sharp de-tails and smooth areas in equal way, patch foveation isselective in both space and frequency. In particular, wepresent an explicit construction of a patch-foveation oper-ator that, given an arbitrary windowing kernel, replaces thecorresponding windowing operator providing equivalent at-

92 IS12 Abstracts

tenuation of i.i.d. Gaussian noise, yet giving full weights toflat regions. Examples of this special form of self-similarityare shown for a number of imaging applications, with par-ticular emphasis on image filtering, for which we demon-strate that foveated self-similarity is a more effective regu-larity assumption than the windowed self-similarity in as-sessing the patch similarity in nonlocal means denoising.

Alessandro FoiDepartment of Signal ProcessingTampere University of Technology, [email protected]

Giacomo BoracchiPolitecnico di Milano, ItalyDipartimento di Elettronica e [email protected]

MS43Optimal Transport Over the Space of Patches

In this talk, I will propose a new model for static and dy-namic textures as a statistical distribution over the spaceof patches. This set of patches is described as a discretizedmanifold using a nearest neighbor graph. A library of tex-tures defines a collection of high dimensional distributions.It is possible to navigate over the set of distributions usingthe machinery of optimal transport on the graph manifold.I will show several applications such as patch-based synthe-sis using optimal transport projection and texture mixingusing geodesic transport interpolation.

Gabriel PeyreCeremade, Universite Paris [email protected]

MS43Learning to Sense GMMs

In this work we describe methodologies to design the sens-ing matrix for efficient coding and detection of signalsthat follow a Gaussian Mixture Model. We describe bothoff-line and on-line techniques, in particular, two-step ap-proaches that first detect the correct model and then opti-mally sense for it. Joint work with J. Duarte-Carvajalino,G. Yu, and L. Carin.


MS44Regional Lung Perfusion Estimated by ElectricalImpedance Tomography

In order to visualize blood perfusion from electricalimpedance tomographic images, it is necessary to separateperfusion information from images that also contain venti-lation and noise. Clinically, it must be done in real time.This separation is performed through real-time projectioninto a subspace of principal components of perfusion, re-cently estimated from the same individual. Some clinicalexamples are presented.

Marcelo Amato, Eduardo Leite CostaUniversity of Sao [email protected], [email protected]

MS44Electrical Impedance Tomography and Ultrasound:Some Experiences


Alex HartovDartmouth [email protected]

Ryan HalterDartmouth [email protected]

MS44An Unscented Kalman Filter with an Approxima-tion Error Model for Electrical Impedance Tomog-raphy

Kalman filters are suitable for Electrical impedance tomog-raphy dynamic imaging. Both evolution and observationmodels are required and Kalman filters performance de-pends on the quality of these models. In many applications,accurate models are unfeasible for the inverse problem andapproximated models must be employed. To improve theaccuracy of approximated models, the approximation er-ror method can be used for estimating the statistics of theresulting errors. Once these statistics are estimated, the in-verse problem can be solved leading to better performance.

Jari KaipioKuopio [email protected]

MS44Prior Models of the Human Chest in ElectricalImpedance Tomography

Prior information is important for solving inverse problemsin many fields. Often, priors are available as subjectiveknowledge and translating these into classical regulariza-tion schemes can be cumbersome. Bayesian methods pro-vide a statistical framework for inverse problems, statingthe problem as finding the posterior probability densityfunction conditioned to the acquired measurements. Un-der this framework, adding subjective prior information isoften more straightforward. In this talk, prior models ofthe human chest for EIT are presented and included in aBayesian algorithm.

Fernando S. MouraUniversity of Sao [email protected]

MS45Detecting Molecular Symmetry and Shape UsingCryo-electron Microscopy, Geometry, and GroupRepresentation Theory

I will explain an algorithm that analyzes the noisy Radonimages of a molecule obtained using cryo-electron micro-scope. The output of the algorithm is the symmetry andshape of the molecule. This algorithm should be an alter-native to existing algorithms that are not robust to noise,use some a priori assumptions on the shape of the molecule,or assume our ability to crystallize the molecule. The ideasbehind the algorithm, use geometric and group represen-

IS12 Abstracts 93

tation theoretic tools.

Shamgar GurevichUniversity of [email protected]

Ronny HadaniUniversity of Texas at [email protected]




Fred SigworthYale [email protected]

MS45Spherically Constrained Reconstruction in Cryo-EM

In this talk, I will consider the problem of reconstructingtransmembrane proteins after they are embedded in a syn-thesized sphere. Such spherically constrained reconstruc-tion offers the advantage of limiting the degrees of freedomin reconstruction and of reconstructing the embedded stateof the protein.

Hemant Tagare, Andrew Barthel, Fred SigworthYale [email protected], [email protected],[email protected]

MS45New Rotationally Invariant Viewing DirectionClassification Method for Cryo-EM Single ParticleReconstruction

One of the challenges in Single Particle Reconstruction is toidentify images of the same view and bring them into reg-ister without prior knowledge of the molecule. We proposea new rotationally invariant viewing direction classificationmethod to deal with this problem. In this method, imagesof the same view are identified by direct cross-correlation ofthe bispectrum-like rotationally invariant features and thein-plane rotational alignment is done only among those ofsimilar viewing direction.

Zhizhen Zhao, Amit SingerPrinceton [email protected], [email protected]

MS46Robust Principal Component Analysis-based Four-dimensional Computed Tomography

The Robust Principle Component Analysis model to4DCT. Instead of viewing 4D object as a temporal col-lection of three-dimensional (3D) images and looking forlocal coherence in time or space independently, we per-cept it as a mixture of low-rank matrix and sparse matrix

and explore the maximum temporal coherence of spatialstructure among phases. Here the low-rank matrix corre-sponds to the “stationary’ background or reference statein space over time, while the sparse matrix stands for the“moving’ or “changing’ components, (under some propersparsifying transform,) e.g., tumors or organs in motion.Together with a dynamic data acquisition procedure, ourmethod can potentially maintain the similar reconstruc-tion accuracy while using fewer projections of the data,as this dynamic scheme can reduce the total number ofmeasurements and hence the radiation dose by acquiringcomplementary data in different phases without redundantmeasurements of the common background structure.

Hao GaoUniversity of Los [email protected]

Jianfeng CaiDepartment of MathematicsUniversity of [email protected]

Zuowei ShenUniversity of WisconsinComputer Science [email protected]

Hongkai ZhaoUniversity of California, IrvineDepartment of [email protected]

MS46Fast Multiscale Reconstruction of Images fromCompressed Sensing Data

In this paper, we introduce multi-scale methods for recon-struction of sparse signals and images from measurementsin a transform domain (i.e. the signal is sampled in theFourier, Hadamard, or wavelet domain). We will show howa decomposition can be used to form ”coarse-grid” versionsof large-scale problems. These reduced formulations of theproblem are then used to craft multi-scale solvers for sig-nal reconstruction problems. These multi-scale solvers areunconditionally stable, and are much more efficient thanconventional first-order methods.

Tom GoldsteinStanford [email protected]

Judah JacobsonUCLA [email protected]

MS46Surface Ricci Flow and its Applications

Hamilton’s Ricci flow deforms the Riemannian metric pro-protional to the curvature, such that the curvature evolveslike a heat diffusion process, and eventually curvature be-comes constant everywhere. In this talk, the discrete the-ories and computational algorithms for surface Ricci flowwill be explained in details. Its applications in medicalimaging, computer vision, geometric modeling and wire-less sensor networks will be briefly introduced.

David Gu

94 IS12 Abstracts

State University of New York at Stony [email protected]

MS46Solving Pdes on Pont Clouds

We present an efficient and accurate numerical method forsolving PDEs on point clouds directly without a surfacereconstruction. The method only need a local least squareapproximation of the surface in a local coordinate systemto learn the local metric to discretize differential operators.We will show computation of eigenvalues and eigenfunc-tions of Laplace-Beltrami operator and other applications.


Rongjie LaiUniversity of Southern [email protected]

Jian Liang, Alvin WongUC [email protected], [email protected]

MS47Construction of Frames for Image Analysis andRepresentation

We propose a new method for constructing tight framesfrom a set of orthonormal vectors via upsampling and cir-cular convolution. When the vectors satisfy an orthonor-mal wavelet condition the technique effectively character-izes variation information and removes polynomial trend in1D and 2D data. We extend the applications of this newtechnique to image edge detection, image representationand reconstruction, and data reduction.

Edward H. [email protected]

MS47Sparse Modeling for Human Activity Analysis inMotion Imagery

An efficient sparse modeling pipeline for the recognitionof human actions for action-based summarization of videois here developed. Spatio-temporal features that charac-terize local changes in the image are first extracted. Thisis followed by the learning of a class-structured dictionaryencoding the individual actions of interest. Recognitionis then based on a tradeoff of sparsity and quality of re-construction, where the labels assigned to localized videoframes come from the optimal sparse linear combination ofthe learned basis vectors (action primitives) representingthe actions. A low computational cost deep-layer modellearning the interclass correlations of the data is added forincreasing discriminative power. In spite of its simplicity,the method is capable of modeling complex human inter-actions, outperforming previously reported results.

Alexey CastrodadNGA and the University of [email protected]


MS47Frame Theory and Sparse Deterministic Represen-tations for Image Applications

We show in this work novel Frame Theoretic techniquesand algorithms that exploit geometric information forphysical meaningful sparse deterministic representations ofhigh dimensional hyperspectral data. The proposed theo-retical developments are employed for the purposes of pixelclassification.

Wojtek CzajaUniversity of MarylandCollege [email protected]

MS47Evaluating Compressively Sensed HyperspectralImagery

We introduce a new family of algorithms for construct-ing HSI from CS measurements. These algorithms com-bine spatial Total Variation (TV) with smoothing in thespectral dimension. We examine models for three differ-ent CS sensors: the Coded Aperture Snapshot SpectralImager-Single Disperser (CASSI-SD) [Wagadarikar et al.]and Dual Disperser (CASSI-DD) [Gehm et al.] cameras,and a random sensing model closer to CS theory, but notnecessarily implementable with existing technology. Wesimulate the capture of remotely sensed images by apply-ing the sensor forward models to well-known HSI scenes anAVIRIS image of Cuprite, Nevada and the HYMAP Urbanimage.

John GreerNational Geospatial-Intelligence [email protected]

MS47A Convex Non-Negative Matrix Factorization andDimensionality Reduction on Physical Space

A convex framework for factoring a data matrix X intoa non-negative product AS, with a sparse coefficient ma-trix S is developed. The columns of the dictionary matrixcoincide with certain columns of X, thereby guaranteeinga physically meaningful dictionary and dimensionality re-duction. We focus on applications to hyperspectral andabundances identification.


MS48Adaptive Compressive Imaging: Information-theoretic Design

We describe an information-theoretic design framework forcompressive measurement design that exploits the addi-tional prior information beyond signal/image sparsity. Theperformance of information-optimal compressive measure-

IS12 Abstracts 95

ments is compared with random projections based mea-surements for different system parameters (e.g. signal tonoise ratio) and prior models. We also consider a mixtureof Gaussian signal prior for adaptive compressive imagingdesign and quantify the performance improvement with re-spect to a static measurement scheme.

Amit AshokThe University of [email protected]

MS48Efficient 2-D Deconvolution for Filters YieldingTriangular Structure

In this work, we will show that certain type of 2D de-convolution operators can be inverted in O(Nlog2N) time.While such “superfast” solvers have been developed re-cently for Toeplitz matrices (1D convolution), it has provendifficult to date to extend these algorithms to multi-levelToeplitz structure (i.e. convolution in multiple dimen-sions). Our algorithms are fast and exact (requiring justa few FFTs if the kernel is known in advance), and workon 2-level Toeplitz matrices that are triangular on one orboth of their levels.

Justin RombergSchool of ECEGeorgia [email protected]

MS48GMM for Audio Analysis

A framework for representing quasi-harmonic signals, andits application to score-informed single channel musical in-struments separation, is introduced in this paper. In theproposed approach, the signals pitch and spectral envelopeare modeled separately. The model combines parametricfilters enforcing an harmonic structure in the representa-tion, with Gaussian modeling for representing the spec-tral envelope. The estimation of the signals model is castas an inverse problem efficiently solved via a maximum aposteriori expectation-maximization algorithm. The rela-tion of the proposed framework with common non-negativefactorization methods is also discussed. The algorithm isevaluated with both real and synthetic instruments mix-tures, and comparisons with recently proposed techniquesare presented. Extensions to speaker separation are pre-sented as well.


MS48Managing Model Complexity in High Frame RateCompressive Video Sensing

In compressive streaming measurement systems such as the”single-pixel” camera, video reconstruction can be particu-larly daunting, because one may record as little as one mea-surement per time instant. We describe a framework formanaging the complexity of representing such high framerate videos. We explain how this framework can be incor-porated into compressive sensing reconstruction, and wediscuss the implications and tradeoffs associated with spa-tial and temporal resolution and the speed of moving ob-

jects.

Michael B. WakinDivision of EngineeringColorado School of [email protected]

MS49High Order Models and Fast Algorithms for FairingVariational Implicit Surfaces

We present two numerical algorithms to solve three relatedcurvature-based high order models for surface fairing. Thefirst, is a stabilized fixed-point algorithm already tested inimage denoising and now adapted to solving 3D models.The second, is a modification of a two-step method used tosolve one of the models and here generalised to solve theother two presenting the new algorithm in an unified nu-merical framework applicable to all three models. Detailsfrom http://www.liv.ac.uk/cmit

Carlos BritoUniversidad Autonoma de [email protected]


MS49Approximate Envelope Minimization for Higher-Order Models

We propose a general method for minimizing a non-convexfunction, that can be written as the sum of simple func-tions. The method minimizes a convex function that isobtained by constructing a local convex envelope of theoriginal non-convex function. The obtained approximateconvex envelope can be written as a linear program incor-porating local polytopes. For numerical solution we adopta recently proposed pre-conditioned first-order primal-dualalgorithm. We show results on classical energy minimiza-tion problems including elastica minimization and stereousing a second order smoothness term.


MS49A Convex, Lower Semi-continuous Approximationof the Euler-elastica Functional

We propose a convex lower semi-continuous approxima-tion of Euler’s elastica energy to be used for regulariza-tion in image processing. The approximation is relatedto the convex relaxation, which however seems difficult toobtain (unless the total variation part of the elastica en-ergy is neglected in which case the convex relaxation inter-estingly reduces to constantly zero). Our approximationarises via so-called functional lifting of the image gradientinto a space which has the normal and curvature of theimage level lines as additional coordinates, and it can beexpressed as a linear program.

Kristian BrediesUniversity of Graz

96 IS12 Abstracts

Institute of Mathematics and Scientific [email protected]


Benedikt K. WirthRheinische Friedrich-Wilhelms-Universitat [email protected]

MS49The H−1 Norm of Tubular Neighborhoods ofCurves and its Relation to the Elastica Energy

The H−1 norm appears frequently in variational modelsfor pattern formation as well as in some models for imageanalysis. Because patterns and images often have struc-tures that are localized around curves, it is of interest toknow how the H−1 norm interacts with the geometry of thecurve. With this goal in mind in this talk we will study

(a rescaled version of) the H−1 norm of the constant func-tion 1 on tubular neighborhoods of curves. We prove twoasymptotic expansions in the width of the neighborhood,that show that at lowest order the H−1 norm depends onthe length of the curve, at the next order the open endsof the curve (which are absent for a closed curve) con-tribute, and the curvature squared appears in the thirdlowest order term. This last term implies a relation withthe elastica functional. One of the asymptotic results is

a pointwise asymptotic expansion for a fixed smooth openor closed curve, the other result is a Gamma-convergenceresult where the (closed) curves along the sequence are al-lowed to vary. This is joint work with Mark A. Peletier

done while both authors were at Eindhoven University ofTechnology.

Yves van GennipDepartment of MathematicsSimon FRaser [email protected]

MS50Multiple-Input Multiple-Output Synthetic Aper-ture Radar Imaging

In this paper, multi-input multi-output (MIMO) forwardlooking synthetic aperture radar (FL-SAR) is developedfor imaging from a moving ground vehicle in multipath en-vironments. Synthetic aperture radar, commonly deployedon airborne platforms, is currently precluded on surfacevehicles due to multipath imaging artifacts in complex ter-rain. In this paper, MIMO methods are used to improveSAR images by suppressing directions of departure whichwould otherwise be multipath scattered and added to di-rect path returns.

Jeffrey KrolikElectrical & Computer EngineeringDuke [email protected]

Li LiDuke Universityn/a

MS50Estimating Size, Shape, and Motion via Near-fieldInvariants of Radar Signals

Understanding range dependent signals (from radar, seis-mic, etc) leads to Euclidean geometry problems, which re-quire the determination of geometric configurations frompartial information. Methods of attack include coordinatebased approaches, invariant equations, and matrix com-pletions. The best methods with respect to computationalcomplexity and sensitivity to noise, seem to be not yet dis-covered. Insights are being gained from experiments withalgebraic solutions of invariant equations, numerical opti-mization methods, matrix completion methods, and fixedpoint iterations.

Mark StuffMichigan Tech Research [email protected]

MS50Reducing the Ionospheric Distortions of Space-borne SAR Images with the Help of Dual CarrierProbing

Spaceborne SAR images deteriorate because of the mis-match between the received signal (affected by the iono-spheric dispersion) and the matched filter (taken as if thepropagation was unobstructed). The filter can be adjustedby including the parameters of the ionosphere, which, inturn, can be obtained by probing the terrain on two car-rier frequencies. We study robustness of this approach, andshow that the quality of the images improves after the filterhas been corrected.

Mikhail Gilman, Erick SmithNorth Carolina State [email protected], [email protected]

Semyon V. TsynkovDepartment of MathematicsNorth Carolina State [email protected]

MS50Polarimetric Synthetic-aperture Inversion for Ex-tended Targets in Clutter

We present an analytic inversion scheme for polarimetricsynthetic-aperture radar in the case of an extended tar-get embedded in clutter. We use microlocal analysis ina statistical setting to develop filtered-backprojection-typereconstruction methods. We model scatterers as dipoles;a scattering matrix thus characterizes scatterers. We in-clude directional scattering assumptions to distinguish anextended target from clutter. We choose the backprojec-tion filter that minimizes the mean-square error betweenthe image and the target scattering matrix.

Kaitlyn VoccolaNational Research CouncilAir Force Research [email protected]

Margaret CheneyRensselaer Polytechnic InstDept of Mathematical [email protected]

IS12 Abstracts 97


MS51Approximate Marginalization Over Uncertaintiesin Electrical Impedance Tomography

Electrical impedance tomography, like other inverse prob-lems, tolerates model uncertainties poorly. In biomedicalproblems, for example, the boundary of the domain is time-varying, and constructing the computational model usinga nominal geometry typically results in meaningless con-ductivity estimates. Similarly, truncation of the computa-tional domain usually has the same effect. We consider aBayesian inversion approach which models the uncertain-ties as stochastic entities and uses a computationally effi-cient marginalization approach to recover from the uncer-tainties.

Jari P KaipioDepartment of Applied Physics, University of EasternFinlandDepartment of Mathematics, University of [email protected]

MS51Feasibility of Electrical Impedance TomographyImaging for Non-destructive Testing of Concrete

Electrical Impedance Tomography (EIT) is an imagingmodality in which the internal three-dimensional conduc-tivity distribution of the target is reconstructed on the ba-sis of electrical measurements acquired from the surface ofthe target. We have recently studied the feasibility of EITfor non-destructive testing of concrete. Especially we haveconsidered the localization of reinforcing bars and the iden-tification of cracks in concrete. The results indicate thatEIT holds potential for the assessment of concrete struc-tures.

Aku Seppanen, Kimmo KarhunenDepartment of Applied PhysicsUniversity of Eastern [email protected], [email protected]



MS51A Feynman-Kac-type Formula for Impedance To-mography

In this talk we consider current-to-voltage maps arising inthe complete electrode model of impedance tomography.A generalization of the classical Feynman-Kac formula isestablished. Furthermore, we discuss the numerical compu-tation of current-to-voltage maps by a Monte Carlo method

derived from this representation formula. As an applica-tion we study the reconstruction problem in the Bayesianframework using a probabilistic forward model.

Martin SimonJohannes Gutenberg Universiat [email protected]

Martin HankeUniversity of [email protected]

Petteri PiiroinenUniversity of [email protected]

MS51Simultaneous Reconstruction of Outer BoundaryShape and Conductivity Distribution in ElectricalImpedance Tomography

The aim of electrical impedance tomography (EIT) is to re-construct the admittance distribution inside a body fromboundary measurements of current and voltage. In thiswork, the need for prior geometric information on themeasurement setting of EIT is relaxed by introducing aNewton-type output least squares algorithm that recon-structs the admittance distribution and the object shape si-multaneously. The functionality of the technique is demon-strated via numerical tests with experimental data.

Stratos StaboulisAalto [email protected]


Aku SeppanenDepartment of Applied PhysicsUniversity of Eastern [email protected]

Jeremi DardeAalto [email protected]

MS52Efficient Spectral Clustering using Selective Simi-larities

Spectral clustering has been very successful at segment-ing images based on non-local similarities between pixels.However in many applications, e.g. multi-scale clusteringof fiber-tracks in the brain using high-resolution DSI im-ages, computing similarities between all pixels can be veryinefficient. We show that eigenvectors of hierarchically-structured high-rank similarity matrices can be recoveredby selectively sampling only O(N log N) entries of the N xN matrix. This leads to very efficient hierarchical spectralclustering algorithms.

Aarti SinghMachine Learning DepartmentCarnegie Mellon [email protected]

98 IS12 Abstracts

MS52Robust Locally Linear Analysis with Applicationsto Image Denoising and Blind Inpainting

I will talk about the problems of denoising images cor-rupted by impulsive noise and blind inpainting (i.e., in-painting when the deteriorated region is unknown). Ourbasic approach is to model the set of patches of pixels in animage as a union of low-dimensional subspaces, corruptedby sparse but perhaps large magnitude noise. For this pur-pose, we develop a robust and iterative method for singlesubspace modeling and extend it to an iterative algorithmfor modeling multiple subspaces. I will also cover the con-vergence for the algorithm and demonstrate state of theart performance of our method for both imaging problems.

Yi WangSchool of MathematicsUniversity of [email protected]

Arthur SzlamDepartment of [email protected]

Gilad LermanDepartment of MathematicsUniversity of [email protected]

MS52Dictionary Learning by L1 Minimization

The idea that many important classes of signals can bewell-represented by linear combinations of a small set ofatoms selected from a given dictionary has had dramaticimpact on the theory and practice of signal processing. Forpractical problems in which an appropriate sparsifying dic-tionary is not known ahead of time, a very popular andsuccessful heuristic is to search for a dictionary that mini-mizes an appropriate sparsity surrogate over a given set ofsample data. While this idea is appealing, the behavior ofthese algorithms is largely a mystery; although there is abody of empirical evidence suggesting they do learn veryeffective representations, there is little theory to guaran-tee when they will behave correctly, or when the learneddictionary can be expected to generalize. In this talk, wedescribe several steps toward such a theory. We show thatunder mild hypotheses, the dictionary learning problem islocally well-posed: the desired solution is indeed a localminimum of the L1 norm. Namely, if A is an incoherent(and possibly overcomplete) dictionary, and the coefficientsX follow a random sparse model, then with high probability(A;X) is a local minimum of the L1 norm over the mani-fold of factorizations (A,X) satisfying AX = Y , providedthe number of samples is sufficiently large. For overcom-plete A, this is the first result showing that the dictionarylearning problem is locally solvable. Our analysis draws ontools developed for the problem of completing a low-rankmatrix from a small subset of its entries. We also discussseveral restricted situations in which the problem can besolved globally by efficient algorithms. We motivate theseproblems with applications examples from image process-ing and computer vision, such as image super-resolutionand hyperspectral image acquisition. We also draw con-nections to related problems in sparse vector and low-rank

matrix recovery.

John WrightDepartment of Electrical EngineeringColumbia [email protected]

MS52On Collective Matrix Factorization

The problem of collective matrix factorization naturallyarises in a number of situations such as relational learningand protein-protein interaction prediction. Motivated bythese applications, we investigate trace norm type of pe-nalization to simultaneously learn a collection of low rankmatrices from noisy observations on a few entries. We de-velop efficient algorithms to compute and establish errorbounds for the estimate. Numerical examples are also in-cluded to demonstrate its merits.

Ming YuanSchool of Industrial and Systems EngineeringGeorgia Institute of [email protected]

MS53Comparison of Expectation-maximization andGraph-embedding Algorithms for Single ParticleImaging

Single-particle imaging with pulsed x-ray sources is facedwith two problems: low signal and missing (orientation)information. One proposed reconstruction algorithm, theexpectation-maximization method, builds a model fromscratch and incrementally refines it to increase the prob-ability of its consistency with the data. The alternativegraph-embedding method constructs a graph whose edgesare defined by pairs of similar data, and then embeds thegraph into the space of orientations using the low ordereigenfunctions of the graph/space. In this talk I reviewthese methods and contrast their performance in the limitof low signal.

Veit ElserDepartment of PhysicsCornell [email protected]

MS53Structure and Dynamics from Manifold Symme-tries of Image Formation

We present a technique to solve the orientation recoveryproblem in ultra-low signal image and diffraction applica-tions based on the differential-geometric properties of therotation group SO(3). The method employs the leadingeigenfunctions of the Laplace-Beltrami operator on SO(3),evaluated via sparse graph-theoretic algorithms, to make aconsistent orientation assignment to each snapshot in thedata set. We demonstrate efficient reconstruction of molec-ular structure from single-molecule diffraction patterns andimages, and dynamically-evolving objects.

Dimitris GiannakisNew York [email protected]

IS12 Abstracts 99

MS53Computational Challenges for Biological StructureDetermination Using X-ray Diffraction

Computational methods are a key success factor in theuse of X-ray crystallography for discovering the structuresof large biomolecular complexes. Routine steps, calcu-lated on small workstations, include derivation of Fourierphases to map the electron density, and improvement ofatomic models using stereochemical knowledge. Treat-ment of raw diffraction images requires multiprocessing ap-proaches, particularly at modern femtosecond X-ray free-electron lasers, where data are acquired at rates exceeding1 GB/s and whole datasets reach 100 TB.

Nicholas SauterLawrence Berkeley National [email protected]

MS53Algorithms for Single Molecule Diffractive Imaging

The recent advances in X-ray science have made it possibleto elucidate the 3D structure of molecules using 2D diffrac-tion patterns of isolated molecules of the same type. Toaccomplish this task, we must first obtain the 3D diffrac-tion pattern of the molecule. The reconstruction problemcan be formulated as a nonlinear least squares problem ora maximum likelihood estimation problem. I will reviewand compare a number of different algorithms for solvingthese problems.

Chao YangLawrence Berkeley National [email protected]

MS54Computational Methods in CryoEM

Cryo electron microscopy (cryoEM) provides structural in-formation spanning from the atomic to the cellular leveland is valuable for insights into the chemistry, molecularbiology, and sub-cellular context of macromolecules andlarger assemblies. CryoEM methodology can be dividedinto electron crystallography, single particle cryoEM andelectron tomography. Several recent computational de-velopments are focused on obtaining reliable 3-D startingmodels and the fact that the complexes being investigatedoften show structural heterogeneity that reflects biologicaldynamics.

Hans HebertDepartment of Biosciences and NutritionKarolinska Institute, [email protected]

Philip KoeckRoyal Institute of [email protected]

MS54ET as an Inverse Problem with an Unbounded Op-erator

We investigate the general setting of a linear operator

A : D(A) ⊃ X −→ Y (1)

mapping a dense subset of a Hilbert space X to a Hilbertspace Y. The concept of the approximate inverse (Louis /Maass 1990) is shown to be easily adapted to this situation,and a natural extension to feature reconstruction (Louis2011) is presented.This method is applied to electron tomography, resultingin a fast WBP-type algorithm and a convergence theoremcharacterizing the best possible reconstruction.

Holger KohrAngewandte MathematikUniversitaet des [email protected]

MS54Local Algorithms in Electron Tomography

The authors will present derivative backprojection algo-rithms for electron microscope tomography (ET). Thesealgorithms are local in the sense they need only limited an-gle region of interest data. This is exactly the data givenin ET. The first algorithm is joint work with Ozan Oktemand is for linear electron paths. The second is joint workwith Hans Rullgard, and it is for curved electron paths,which occur in electron microscopes with larger fields.

Eric Todd QuintoTufts [email protected]

Ozan OktemKTH - Royal Institute of [email protected]

Hans RullgardStockholm University and [email protected]

MS54Detection of Macromolecular Structures in Tomo-grams Using Reduced Representation Templates

Electron Tomography (ET) is useful for elucidating the3D-architecture of cellular and sub-cellular systems at res-olution of about 5 nm. One difficulty in ET is to assignmolecular components to densities within the reconstruc-tion. We will discuss detection of structures in tomogramsusing reduced representation templates. These consist ofsmall point sets capturing the characteristics of the under-lying structure. This approach is computationally feasibleand useful for structures with higher order such as filamentsand bundles.

Neils VolkmannBioinformatics and Systems [email protected]

MS55Edge Detection and Image Reconstruction UsingArea Preserving Flows

We present the application of area preserving curvaturemotion to certain inverse problems that arise in medicalimaging, such as tomographic inversion. In particular, wecompare the common choices of perimeter penalty as aregularization term, for example as in the Mumford-Shah

100 IS12 Abstracts

functional, with area preserving flows, and show that theuse of area preserving curvature motion abates some re-construction artifacts such as shrinkage of object contoursdue to perimeter penalties. Numerical experiments alsosuggest that area preserving flows may be less sensitive toparameters than Mumford-Shah based flows.

Catherine M. KublikUniversity of Texas, [email protected]


MS55Medical Morphometry and Computer Visions Us-ing Quasi-Conformal Teichmuller Theory

In medical imaging and computer visions, studying the ge-ometry and the deformation of 2D/3D shapes are of utmostimportance. For example, comparing shape differences be-tween anatomical structures (such as human brains) is cru-cial for disease analysis. Quasi-conformal Teichmuller The-ory, which studies geometric deformations between shapes,is an ideal tool for this purpose. In practice, 2D/3D shapesare usually represented discretely by triangulation meshes.We therefore need to have a discrete version of Quasi-conformal theory on discrete meshes. In this talk, I willfirstly talk about how can develop a discrete analogue ofQC geometry on meshes. I will then talk about how com-putational QC geometry can be applied to medical imag-ing and computer visions applications. This is a joint workwith Tony F. Chan, Alvin Wong and S-T Yau

Ronald Lok Ming LuiDepartment of MathematicsThe Chinese University of Hong [email protected]

MS55Reconstruction of Binary Function from Incom-plete Frequency Information

Binary function is a class of important function that ap-pears in many applications e.g. image segmentation, barcode recognition, shape detection and so on. Most stud-ies on reconstruction of binary function are based on thenonconvex double-well potential or total variation. In thisresearch we proved that under certain conditions the binaryfunction can be reconstructed from partial frequency infor-mation by using only simple linear programming, which isfar more efficient.

Yu MaoInstitute for Mathematics and Its ApplicationsUniversity of [email protected]

MS55Applications of Computational Quasiconformal Ge-ometry on Medical Morphometry and ComputerGraphics

Conformal mappings have been widely applied in medicalimaging and computer graphics, such as in brain registra-tion and texture mapping, where the mappings are con-structed to be as conformal as possible to reduce geometric

distortions. A direct generalization of conformal mappingsis quasiconformal mappings, where the mappings are al-lowed to have bounded conformality distortions. In thistalk, we explore how the theories of quasiconformal map-pings and their computations can be applied to areas whereconformal mappings are used traditionally. These includesregistration of biological surfaces, shape analysis, medicalmorphometry, compression and refinement of texture map-pings, and the inpainting of surface diffeomorphisms.

Alvin Tsz Wai WongDepartment of MathematicsUniversity of California, [email protected]

MS56Multi-Sheet Rebinning Methods for Reconstruc-tion from Asymmetrically Truncated Cone BeamProjections

We present a family of multi-sheet rebinning methods forreconstruction from axially asymmetrically truncated conebeam X-ray projections, as acquired by a new type of multi-source scanners. In the sense of integral geometry thedata is severely limited, nonetheless, the proposed meth-ods achieve a reconstruction of quality similar to that ofthe complete data, due to an optimal combination of ana-lytical and numerical techniques.

Marta BetckeUniversity of College London, [email protected]

Bill LionheartUniversity of Manchesterbill lionheart [email protected]

MS56Adaptive Frequency-domain Regularization forSparse-data Tomography

A novel reconstruction technique for sparse tomographicimaging is introduced. This method applies a spatiallyvarying constrained least square filter with regularizationmethod based on total variation. The WIRT reconstruc-tion is implemented in the frequency domain, where theinformation based on measurements and regularization canbe treated separately. This algorithm is selectively apply-ing regularization in the frequency regions where the fre-quency component values cannot be defined by the mea-surements, which enables computationally more effectiveiteration procedure than the conventional iterative meth-ods.

Martti KalkePaloDex [email protected]

MS56Approximation Error Approach in Local Tomogra-phy

Recently, the approximation error approach has been pro-posed for handling uncertainty and model reduction relatederrors in inverse problems. In this approach, approximatemarginalization of these errors is carried out before the esti-mation of the interesting variables. In this talk, we describe

IS12 Abstracts 101

the adaptation of the approximation error approach to themarginalization of the local tomography problem with re-spect the unknown x-ray attenuation density outside theregion-of-interest.

Ville P. KolehmainenDepartment of Applied PhysicsUniversity of Eastern [email protected]


MS56Polyenergetic X-Ray Reconstruction Algorithmsfor Breast Imaging

The U.S. Food and Drug Administration recently approvedthe first 3-D tomosynthesis breast imaging system. Thetechnique uses standard mammography hardware to obtainlimited angle tomography data. To reduce beam-hardeningartifacts, and to obtain quantitative information about ma-terials in the breast, we use an image reconstruction ap-proach that incorporates the polyenergetic nature of the x-ray beam. In this talk we describe efficient computationalapproaches to solve the resulting large-scale nonlinear in-verse problem.

James G. NagyEmory UniversityDepartment of Math and Computer [email protected]

Ioannis SechopoulosDepartment of Radiology and Winship Cancer InstituteEmory [email protected]

Veronica M. BustamanteEmory [email protected]


Steve FengDepartment of RadiologyEmory Universitysteve si [email protected]

MS57Recovering Sparse and Low Rank Matrices fromCompressive Measurements

We consider the problem of recovering a matrix M thatis the sum of a low-rank matrix L and a sparse matrixS from a small set of linear measurements of the formy=A(M) = A(L+S). This model subsumes three importantclasses of signal recovery problems: compressive sensing,affine rank minimization, and robust principal componentanalysis. We propose a natural optimization problem forsignal recovery under this model and develop a new greedyalgorithm called SpaRCS to solve it. SpaRCS inherits anumber of desirable properties from the state-of-the-art

CoSaMP and ADMiRA algorithms, including exponentialconvergence and efficient implementation. Simulation re-sults with video compressive sensing, hyperspectral imag-ing, and robust matrix completion data sets demonstrateboth the accuracy and efficacy of the algorithm.

Richard G. BaraniukRice UniversityElectrical and Computer Engineering [email protected]

MS57On the Power of Adaptivity in Sparse Recovery

We consider the problem of recovering the (approximately)best k-sparse approximation x* of an n-dimensional vectorx from linear measurements of x. It is known that this taskcan be accomplished using m=O(k log (n/k)) non-adaptivemeasurements and that this bound is tight. In this talk weshow that if one is allowed to perform measurements thatare *adaptive*, then the number of measurements can beconsiderably (sometimes exponentially) reduced comparedto the non-adaptive case. We discuss implications of ourresults to compressive sensing and data stream computing.

Piotr IndykMassachusetts Institute of [email protected]

MS57Optimal Measurement Kernels for Dictionary Im-age Priors

We assume a dictionary which represents the canonical di-rections which represent image patches and wish to recon-struct imagery from compressive measurements. We derivea measurement criteria based upon optimizing the separa-tion of these canonical image patch directions such thatany reconstruction algorithm has better performance. Wealso adapt the measurement paradigm to the constraintsof real hardware.

Robert R. MuiseLockheed [email protected]

MS57Spectral Super-Resolution of Hypercritical Images

Hyperspectral imagery (HSI) is a remote sensing modalitythat yields highly detailed spectral descriptions of the ter-rain, often measuring 100+ spectral bands. HSI imagers,however, are relatively expensive and slow compared tomore common multi-spectral imagers (MSI) measuring 8spectral bands. In this work we propose recovering HSI-level spectral resolution from MSI-level imagery via sparseencoding in a learned spectral dictionary of spectral signa-tures. Learning a HSI dictionary under a sparsity modelcaptures the rich higher-order statistics in HSI, and thisinformation can be leveraged in a linear inverse problem tosuper-resolve HSI-level resolution from MSI-level imagery.We show that the per-pixel spectra can be recovered to highaccuracy even when the scene has changed substantiallydue to seasonal variations, indicating that this procedurecan be used to reduce the cost of obtaining high-resolutionspectral imagery.

Chris Rozell

102 IS12 Abstracts

Georgia Institute of [email protected]

MS58Blind Subpixel Point Spread Function Estimationfrom Scaled Image Pairs

We introduce a blind algorithm for subpixel estimationof the PSF of a digital camera from aliased photographs(necessary assumption, since most cameras sample eachcolor channel far below the Nyquist critical rate). Thenumerical procedure simply uses two fronto-parallel pho-tographs of any planar textured scene at different dis-tances. Hence knowledge of a known calibration grid (com-mon in single-image-based methods) becomes unnecessary.An experimental evaluation shows that the proposed al-gorithm reaches the accuracy levels of the best non-blindstate-of-the-art methods.

Mauricio DelbracioCMLA - ENS Cachan (France) &Universidad de la Republica (Uruguay)[email protected]

Andres AlmansaTelecom ParisTech & [email protected]

Pablo MuseFacultad de IngenieriaUniversidad de la Republica (Uruguay)[email protected]

Jean Michel MorelCMLA (CNRS UMR 8536)ENS Cachan (France)[email protected]

MS58A Variational Approach for the Fusion of ExposureBracketed Pairs

We propose a variational method for automatically recom-bining an exposure bracketed pair of images into a singlepicture that reflects the best properties of each one: in gen-eral, good brightness and colour information are retainedfrom longer exposure settings while sharp details are ob-tained from the shorter ones. The method is able to han-dle camera and subject motion as well as noise, and resultscompare favourably with the state of the art.

Marcelo BertalmıoUniversitat Pompeu FabraBarcelona, [email protected]

Stacey LevineDuquesne [email protected]

MS58Optimal Estimators for HDR Reconstruction

There are physical limitations on the maximal variationsof luminosity that a camera sensor can measure. Theselimitations are due to the fact that each sensor has a finitecapacity : above a given quantity of incident photons, thesensor saturates. This phenomenon is well known by pho-

tographers and is the reason of backlighting in images. Inorder to get details simultaneously in shadows and high-lights, a possibility is to mix several images of the samescene taken with different exposure times. This can bedone directly on RAW images, to avoid information lossdue to quantization and gamma correction. Assume im-ages have been perfectly registered. For each sensor cell ofthe camera, we aim at recovering the irradiance reachingthe cell from the input measurements obtained with dif-ferent exposure times. Knowing the physical model of thesensor, this can be seen as a statistical estimation problem.Several solutions have been proposed in the literature, us-ing for instance heuristics [Debevec, Malik, 97], or a finestochastic modeling of the camera noise [Granados et al,2010]. In this work, we follow this last trend and inves-tigate the optimality of different statistical estimators ofthe irradiance reaching the sensor. This question can betreated independently on each sensor cell, or more glob-ally on the image, for instance by only considering linearestimators.

Julie Delon, Aguerrebere Cecilia, Gousseau YannLTCI - Telecom ParisTech - [email protected], [email protected],[email protected]

MS58Burst Denoising

Taking photographs under low light conditions with ahand-held camera is problematic. A long exposure timecan cause motion blur due to the camera shaking and ashort exposure time gives a noisy image. We consider thenew technical possibility offered by cameras that take im-age bursts. Each image of the burst is sharp but noisy.In this preliminary investigation, we explore a strategy toefficiently denoise multi-images or video. The proposedalgorithm is a complex image processing chain involvingaccurate registration, video equalization, noise estimationand the use of state-of-the-art denoising methods. Yet, weshow that this complex chain may become risk free thanksto a key feature: the noise model can be estimated ac-curately from the image burst. Preliminary tests will bepresented. On the technical side, the method can alreadybe used to estimate a non parametric camera noise modelfrom any image burst.

Toni BuadesCNRS Paris Descartes – Universitat Illes [email protected]


Jean-Michel MorelUniversitat de les Illes Balears, Departament [email protected]

Zhongwei TangEcole Normale Superieure de Cachan, [email protected]

MS59Stochastic and Deterministic Methods for Analysis

IS12 Abstracts 103

of Oscillatory Radar Signals

We examine two different paradigms to analyze the fluc-tuation rates and oscillatory behavior of radar signals.The first method, which was recently proposed by us, isa stochastic method that relies on the level crossing rate(LCR), average out duration (AOD) and average surge du-ration (ASD). This method requires the joint distributionof a stochastic process and its time derivative, the gen-eral solution of which is an open problem and unknown.After defining the concept of this approach and relatingit to radar terminology, we derive some new analyticalresults for a specific stochastic process. The accuracy ofour derived equations are demonstrated vis--vis real radarsignals. The second approach which we describe is thetunable Q-factor wavelet transform (TQWT) which is wellsuited to fast algorithms for sparsity-based inverse prob-lems due to being a Parseval frame, easily invertible andefficiently implementable via radix-2 FFTs. The TQWT,which depends on the two main parameters of the Q-factorand asymptotic redundancy, when tuned, matches the os-cillatory behavior of the wavelet transform. We demon-strate how this approach, which differs from the Fourierand Wavelet transform, decomposes a signal into a high-resonance and low-resonance component, and how it canbe used to distinguish two radar signals. Application ofthis approach to real radar signals is also demonstrated.

Masoud FarshchianNaval Research [email protected]

MS59Overview of Multi-baseline Radar Interferometryand Tomography and Applications Potential


Scott HensleyJet Propulsion [email protected]

MS59Mathematical Problems in Ultrawideband Radar

The very broad instantaneous bandwidth of ultrawideband(UWB) short-pulse radar permits more highly resolvedimaging than traditional radar. Some mathematical prob-lems and issues associated with the underlying electromag-netic theory and signal processing of an UWB radar arediscussed: the fundamental notion of where an antennaradiates; what constitutes a reasonable representation ofan UWB waveform; the appropriateness of the far-field ap-proximation; and the concept of sub-wavelength resolution.

Eric MokoleNaval Research [email protected]

MS59Clutter Effects in Imaging


Knut SolnaUniversity of California at [email protected]

MS60Direct Reconstruction in 3D Electrical ImpedanceTomography Using a Nonlinear Fourier Transform

The Calderon problem is the mathematical formulation ofElectrical Impedance Tomography. In 3D the problem wassolved in the mid 1980’s using a non-linear Fourier trans-form, and only recently a numerical implementation of thereconstruction algorithm has been obtained. In this talkthe algorithm and its complete implementation and regu-larization will be presented. The implementation will beevaluated on several examples and the results will be com-pared to the results of a simpler linearization algorithm.

Kim KnudsenDepartment of MathematicsTechnical University of [email protected]

MS60Electrical Impedance Tomography and the NoiseSubspace

Noise subspace approaches like the MUSIC algorithm arepopular approaches to inverse problems in scattering the-ory. We show that an analysis of the MUSIC algorithmoriginating from inverse scattering can be carried over tothe impedance tomography problem to detect (small or ex-tended) inclusions.

Armin LechleiterUniversity of Karlsruhe, Karlsruhe, [email protected]

MS60Regularized D-bar Method for ElectricalImpedance Tomography

Imaging via electrical impedance tomography (EIT) is anonlinear ill-posed inverse problem. A non-iterative EITimaging algorithm is presented based on the use of a non-linear Fourier transform. Regularization of the method isprovided by nonlinear low-pass filtering, where the cutofffrequency is explicitly determined from the noise amplitudein the measured data. Results from numerical and exper-imental data are presented, suggesting that the method isuseful for imaging the heart and lungs of a living patient.

Jennifer L. MuellerColorado State UniversityDept of [email protected]

Kim KnudsenDepartment of MathematicsTechnical University of [email protected]

Matti LassasUniversity of [email protected]

Samuli SiltanenUniversity of [email protected]

104 IS12 Abstracts

MS60Direct Reconstructions in EIT and Heat Probing

Complex Geometric Optics solutions are used in direct re-constructions methods in EIT and heat probing. Numer-ical examples are presented using a boundary correctingprocedure in the D-bar method, and a novelle heat prob-ing method in 1D, 2D and with moving inclusion of heatconductivity. The results show that the theory involvedcan improve reconstructions in the case of EIT and can benumerically implemented in heat probing.

Janne P. TamminenTampere University of [email protected]

MS61Variations on Lasso Regularization

We will provide an overview of sparsity-based regulariza-tion methods for inverse imaging problems and of itera-tive algorithms allowing to compute the corresponding so-lutions. The emphasis will be put on some recent devel-opments such as the use of mixed penalties, the quest foraccelerating strategies and the solution of blind (or my-opic) inverse problems where the operator to be inverted isalso unknown (or partially known). Some open challengingquestions will also be discussed.

Christine De MolUniversite Libre de [email protected]

MS61Linear Ill-posed Operator Equations with PoissonData

We consider moment discretization for ill-posed problemswith discrete Poisson data by means of penalized max-imum likelihood estimation. Several computational ap-proaches are discussed. Simplest of all is to just con-sider the plain least-squares problem and use “optimally’stopped conjugate gradients type methods. Next is theplain Expectation-Maximization (EM) algorithm for theunpenalized maximum likelihood problem (which has to“optimally’ stopped as well, although it is not so clearhow). There is also a nonlinearly smoothed EM algorithmfor a specific penalized version of the maximum likelihoodproblem, depending on a smoothing (regularization) pa-rameter. This smoothed version of the EM algorithms maybe run until convergence, but the smoothing parameterneeds to be chosen “optimally’. However, these EM al-gorithm are computationally very expensive. A final ap-proach is to approximate the Poisson data with Gaus-sian data with the means equal to the variances. Thisis asymptotically equivalent to an iteratively reweightedleast-squares problem, so conjugate gradient type methodscan then be brought to bear on its solution.

Paul Eggermont, Vincent LaRicciaUniversity of [email protected], [email protected]

Zuhair NashedUniversity of Central [email protected]

MS61Source Identification from Line Integral Measure-ments and Simple Atmospheric Models



MS61The Minimizers of Least Squares Regularized with0 norm: Uniqueness of the Global Minimizer

Data d are generated using an M×N real-valued arbitrarymatrix A (e.g. a dictionary) with M < N . We providean in-depth analysis of the (local and global) minimizersof an objective function F combining a quadratic data-fidelity term and an 0 penalty applied to each entry of thesought after solution, weighted by a regularization param-eter β > 0. For several decades, this objective focuses aceaseless effort to conceive algorithms approaching a goodminimizer. Our theoretical contributions, summarized be-low, shed new light on the existing algorithms and can helpthe conception of innovative numerical schemes. To solvethe normal equation associated with any M -row subma-trix of A is equivalent to compute a local minimizer u ofF. (Local) minimizers u of F are strict if and only if thesubmatrix, composed of those columns of A whose indexesform the support of u, has full column rank. An outcomeis that strict local minimizers of F are easily computedwithout knowing the value of β. Each strict local mini-mizer is linear in data. The global minimizers of F arealways strict. They are studied in more details under the(standard) assumption that rank(A) = M < N. The globalminimizers with M -length support are seen to be imprac-tical. Given d, critical values βK for any K ≤ M − 1 areexhibited such that if β > βK , all global minimizers of Fare K-sparse. An assumption on A is adopted and provento be generically true. Then for generically all data d, theobjective F has a unique global minimizer and the latteris K-sparse for K ≤ M − 1. Instructive small-size (5× 10)numerical illustrations confirm the main theoretical results.

Mila NikolovaCentre de Mathematiques et de Leurs Applications(CMLA)ENS de Cachan, [email protected]

MS62Projection Methods for Low-count DiffractiveImaging in the Near and Far Fields

With few exceptions, the most prevalent practical meth-ods for solving phase retrieval problems arising in imagingare projection methods. Surprisingly, convergence of eventhe simplest algorithms has never been proved in settingsappropriate for phase retrieval. We prove convergence ofa very elementary algorithm for some (not all) phase re-trieval problems and outline the extension of these resultsto more sophisticated algorithms and settings, in particularfor low-count data where Poisson noise is significant andunavoidable. We demonstrate the theory and algorithmson experimental data from low-count photonic imaging ex-periments in the near and far-fields.

Russell Luke

IS12 Abstracts 105

Department of MathematicsUniversity of [email protected]

MS62Real-time Ptychographic X-ray Image Reconstruc-tion

Ptychography is a recent imaging technique which com-bines the ease of use of a scanning transmission X-ray mi-croscope with the high resolution of coherent diffractiveimaging methods. Its robustness, when compared to singleparticle diffractive imaging, make it ideally suited to be-come a routine experimental method. We present the firstreal-time soft X-ray ptychography framework. Its ease ofuse and immediate feedback are finally turning diffractionimaging into a useable routine technique.

Filipe MaiaNERSCLawrence Berkeley National [email protected]



Stefano MarchesiniLawrence Berkeley National [email protected]

MS62PhaseLift: Exact and Stable Phase Retrieval ViaConvex Optimization

PhaseLift is a novel methodology for phase retrieval, com-bining structured illuminations with convex programmingto recover phase from intensity measurements. Along withempirical demonstration that any complex-valued objectcan be recovered from the magnitude of few diffracted pat-terns by solving a convex optimization problem, stablywith respect to noise, we exhibit masks which yield thesignal uniquely and under further assumptions show thatrecovery via convex optimization is exact and stable withvery high probability.

Vladislav Y. VoroninskiUC Berkeley Mathematics [email protected]

Emmanuel CandesStanford UniversityDepartments of Mathematics and of [email protected]

Thomas StrohmerUniversity of California,DavisDepartment of [email protected]

MS63Statistical Extreme Value Theory Meets MultiscaleImaging


Markus Haltmeier

Research Group Statistical Inverse Problems inBiophysicsMax Planck Institute for Biophysical [email protected]

MS63Geometric Multiscale Analysis and Inpainting

One possible model for images is the cartoon model assum-ing that images are governed by edges. Shearlets have beenintroduced as an image processing methodology designedto sparsely approximate cartoon-like images, while ensur-ing a unified treatment of the continuum and digital world.In this talk, we will first discuss novel results on compactlysupported shearlets. Then we will focus on the applicationof shearlets to the task of inpainting, providing theoreticaland numerical results.

Emily KingUniversitaet [email protected]

Gitta KutyniokTechnische Universitat [email protected]

Wang-Q Lim, Xiaosheng ZhuangTechnische Universitaet [email protected], [email protected]

MS63Spectral Density Estimation for Unevenly SpacedTime Series

Regularization approaches for the estimation of the spec-tral density of unevenly spaced time series are proposedand studied; both L2 and L1 options are considered, in thepenalty as well as in the infidelity term. The feasibility ofthe methods depends on the modern algorithms of convexoptimization.

Ivan MizeraDepartment of Mathematical and Statistical SciencesUniversity of Alberta, [email protected]

MS63Statistical Multiscale Methods for BiophotonicImaging.

We present a locally adaptive estimation method of a one ortwo-dimensional regression function for Poisson data. Ourapproach is built on a multiresolution likelihood statisticwhich is not a Gaussian approximation but taylor suitedfor Poisson data and consequently particularly suited forlow count observations, as in photonic spectroscopy. Wefurther provide asymptotic results and illustrate the capa-bility of the proposed method by various examples fromimaging and signal detection.

Hannes SielingInstitute of Mathematical StochasticsUniversity of [email protected]

Klaus FrickUniversity of [email protected]

106 IS12 Abstracts

Munk AxelInstitute of Mathematical StochasticsUniversity of [email protected]

MS64Non-parametric Bayesian Modeling in Image Pro-cessing

A new nonparametric Bayesian model is developed to in-tegrate dictionary learning and topic model into a unifiedframework. The model is employed to analyze partiallyannotated images, with the dictionary learning performeddirectly on image patches or on associated extracted fea-tures. A hierarchical representation of the image patchesand of the text is manifested via a generalization of thenested Chinese restaurant process. Efficient inference isperformed with a Gibbs-slice sampler, and encouraging re-sults are reported on widely used datasets.

Larry CarinElectrical & Computer EngineeringDuke [email protected]

MS64On Perfect Sampling of Some Markovian Energies

In this talk, we consider the sampling of Markov RandomFields using coupled Markov Chain to draw samples fromthe stationary distribution. We propose an algorithm thatperforms perfect sampling, i.e., after a finite number ofsteps the sample is exactly generated according to the tar-get distribution. We give a sufficient condition (that isessentially of combinatorial nature) for the applicability ofthe method.

Jerome DarbonCMLA, ENS Cachan, FranceMathematics Department, UCLA [email protected]

MS64Analysis of Bayesian Problems Using Non-Asyptotic Probability Tail Bounds

An approach to determine if a learned Bayesian dictionaryis appropriate for images not used during the training stageis demonstrated, i.e. the concepts in the new images havedrifted. The procedure relies on probability tail bounds,and a new result on obtaining probability tail bounds isproven. In particular, tail bounds for the exponential fam-ily of distributions are demonstrated. The tail bounds areapplied to SIFT learned dictionaries to discover when con-cept drift occurs.

Arjuna FlennerNaval Air Weapons [email protected]

Gary HewerNAWC Weapons [email protected]

MS64Total Variation Denoising Using Posterior Expec-

tation

Total variation image denoising can be interpreted as aMaximum A Posteriori estimate. This maximization as-pect is partly responsible for the staircasing effect, i.e.the outbreak of quasi-constant regions separated by sharpedges. In this talk we transpose this denoising methodinto an estimation based on the posterior expectation. Wedemonstrate that the denoised images do not suffer fromthe staircasing effect. Practical computation of TV-LSE,and a proposed MCMC algorithm is analyzed.

Cecile LouchetMathematics DepartmentUniversity of Orleans, [email protected]

MS65Variational Methods in Molecular Electron Tomog-raphy

I shall present a variational approach for reconstructing athree dimensional density function from a tilt series of twodimensional electron microscopy (EM) images. By min-imizing an energy functional consisting of a data fidelityterm and a choice of regularization terms, an L2-gradientflow is derived and efficiently solved. An iterative solutionfor the flow is constructed using tensor product B-spline fi-nite elements in the spatial direction and an explicit Eulerscheme in the temporal direction. Experimental tests showthis variational method is efficient and effective and pro-vides higher resolution 3D EM reconstructions, comparedto several existing methods.

Chandrajit BajajComputer Science, ICES, CVCThe University of Texas - [email protected]

MS65Nonlinear Approximations as a Tool in Tomogra-phy

Rapid changes in the medium (e.g. boundaries) are the keyfeatures of interest in tomographic images. However, theyare difficult to reconstruct accurately since the measureddata are bandlimited and contaminated by noise. We usenear optimal rational approximations in order to capturerapid changes in measured quantities. Effectively, we makeuse of the fact that such nonlinear approximations are notsubject to the usual (Nyquist) sampling requirements. Weshow that, when used within our tomographic reconstruc-tion algorithm, the approach results in improved resolutionand noise reduction.

Matt ReynoldsUniversity of Colorado, [email protected]

Lucas MonzonUniversity of Colorado at [email protected]

Gregory BeylkinUniv. of Colorado, [email protected]

IS12 Abstracts 107

MS65Mumford Shah Methods in Limited Data Tomog-raphy

We introduce Mumford-Shah energy functionals for the re-trieval of geometric and physical information from highlynoise-corrupted EMT data. The introduction of appro-priate homogeneity measures reduces ill-posedness of theinverse problem an — at the same time — reconstructsgeometric information of shape and location of the probedirectly from the data. The issue of regularization of thecontour variable by area penalization in the Mumford-Shahapproach is addressed both theoretically and numericallyvia a semi-implicit numerical approach for the iterative up-date of the shape variable. Numerical experiments compar-ing the Mumford-Shah method with other techniques arepresented.

Wolfgang RingInstitut fuer MathematikUniversitaet [email protected]

MS65Making Electron Tomography and TemplateMatching Practical

Acquisition of electron tomographic data sets has becomealmost fully automated nowadays. The bottleneck in thethroughput chain is now clearly in the data analysis partof the tomography workflow, both in preprosessing (align-ment of the tilt series) and in the further analysis using e.g.template matching of the acquired tomograms. In this pre-sentation we present how high performance computing canbe used to take over the role of the human operator in theseareas.

Remco SchoenmakersSenior Software ScientistFEI [email protected]

MS66Proximal Algorithms and CT: New Results on 3DReal Datas and Color CT

The CPPM in Marseille developed recently a new demon-strator of small animal CT-scan PIXSCAN2 and an hybridPET/CT Scan ClearPET/XPAD. The last one is designedto record images for both modalities CT and PET simul-taneously. Moreover the use of a new generation of detec-tors which are not affected by dark noise leads to modelthe measurements with pure Poisson noise. Proximal al-gorithms can be used to solve the inverse problem in CTand they demonstrate on real data very good reconstruc-tion properties at low radiation dose in 2D as well as inthe context of CT spectral acquisitions. In this talk wewill study their properties in the 3D case.

Yannick BoursierAix-Marseille Univ., CPPM, CNRS/IN2P3Marseille, [email protected]

Mathieu DupontAix-Marseille Univ., CPPM, CNRS/IN2P3,Marseille, [email protected]

Sandrine AnthoineAix-Marseille Univ., LATP, CNRSMarseille, [email protected]

Jean-Francois AujolIMB, CNRS, Universite Bordeaux [email protected]

Clothilde MelotAix-Marseille Univ., LATP, CNRSMarseille, [email protected]

MS66Connecting Object Sparsity and Number of Mea-surements in Total Variation-regularized X-ray CT

In X-ray computed tomography (CT), total variation(TV)-regularized image reconstruction shows potential forlow-dose imaging from few-view projection data. How-ever, it remains unclear precisely for which images TV-regularized reconstruction can be expected to be accu-rate. We discuss approaches for determining samplingconditions ensuring recovery of the original image usingTV-regularized reconstruction, including empirical phan-tom reconstruction studies, in an attempt to link objectsparsity and the sufficient number of samples.

Jakob H. JoergensenTechnical University of [email protected]

Emil Y. SidkyUniversity of ChicagoDept. of [email protected]

Per Christian HansenTechnical University of DenmarkInformatics and Mathematical [email protected]

Xiaochuan PanThe University of ChicagoDepartment of [email protected]

MS66A Survey of CT Applications Enabled by Sparsity-exploiting Image Reconstruction

Since 2005, when Candes, Romberg, and Tao launched thesubject dubbed ”Compressive Sensing” (CS), we have beenstudying the optimization problems and sparsity-exploitingideas of this field for the purpose of application to imagereconstruction in Computed Tomography (CT). From theapplications perspective, the importance of the various CSconcepts take on a different weighting than the theoreti-cal investigations. This talk will survey the various appli-cations of exploiting sparsity in X-ray based image recon-struction. We will cover, specifically, sparse-view sampling,limited-angular range scanning, digital breast tomosynthe-sis, and low-intensity X-ray illumination. In keeping withthe theme of explaining the difference between practice andtheory, we will discuss the role of convergence of the variousCS optimization problems and suitability of the CT-modelsystem matrices for sparsity-exploiting image reconstruc-

108 IS12 Abstracts

tion algorithms.

Emil SidkyDepartment of RadiologyThe University of [email protected]

Jakob H. JoergensenTechnical University of [email protected]

Xiaochuan PanThe University of ChicagoDepartment of [email protected]

MS66Low Rank Matrix Decomposition/FactorizationMethods for 4D Ct

We propose to use matrix decomposition/factorizationmethod to reconstruct 4D CT images. The key point isto utilize redundancy in different projections for the com-mon background and global time correlation such as pe-riodicity to improve the image reconstruction and reducethe dose needed. I will discuss two low rank matrix models,one based on matrix decomposition and the other one onmatrix factorization for 4D CT image reconstruction.


Jianfeng CaiDepartment of MathematicsUniversity of [email protected]

Hao [email protected]



Zuowei ShenUniversity of WisconsinComputer Science [email protected]

MS67Sparse Approximation by Wavelet Frames with Ap-plications to Image Restoration

Wavelet frames are known to be able to sparsely approxi-mate piecewise smooth functions, which has recently beenused for image restoration problems including image de-blurring, inpainting and medical imaging. In the first halfof this talk, I will start with a brief review of some waveletframe based models and their connections to variationalmodels. Then I will present a model (as well as some fast

algorithms) that penalizes the 0-norm of the frame coeffi-cients, instead of the commonly used 1-norm. Numericalexperiments show that using the 0-norm has advantagesfor some specific type of images. The second half of the talkis focused on the applications of wavelet frame based mod-els to x-ray based conventional CT image reconstruction,as well as spectral CT image reconstruction.

Bin DongThe University of [email protected]

MS67The Analysis Co-Sparse Model: Pursuit, Dictio-nary Learning, and Beyond

The synthesis-based sparse representation model for signalshas drawn a considerable interest in the past decade. Sucha model assumes that the signal of interest can be decom-posed as a linear combination of a few atoms from a givendictionary. In this talk we concentrate on an alternative,analysis-based model, where an analysis operator (here-after referred to as the Analysis Dictionary) multiplies thesignal, leading to a sparse outcome. Our goal is to learn theanalysis dictionary from a set of signal examples, and theapproach taken is parallel and similar to the one adopted bythe K-SVD algorithm that serves the corresponding prob-lem in the synthesis model. We present the development ofthe algorithm steps, which include a tailored pursuit algo-rithm termed Backward Greedy algorithm and a penaltyfunction for the dictionary update stage. We demonstrateits effectiveness in several experiments, treating syntheticdata and real images, showing a successful and meaningfulrecovery of the analysis dictionary.


Ron RubinsteinCS department, [email protected]

Tomer FaktorTechnion - Israel Institute of [email protected]

MS67Shearlets and Geometry Extraction

One main problem in imaging science is the extraction ofgeometrically distinct features from a mixture of those; anexample being the extraction of spikes (pointlike features)and dendrites (curvelike features) from neurobiological im-ages. For this, we will exploit sparsity methodologies byusing a combined dictionary of wavelets and shearlets. Wewill present a theoretical analysis, an efficient algorithm aswell as numerical experiments for the extraction of generalpoint- and curvilinear features.

Gitta KutyniokTechnische Universitat [email protected]

David L. DonohoStanford UniversityDepartment of [email protected]

IS12 Abstracts 109

Wang-Q LimTechnische Universitaet [email protected]

MS67Box Spline Wavelet Frames for Image Edge Detec-tion

We will explain a bivariate box spline based on 8 directions(1,0), (0,1), (1,1), (1,-1), (2,1),(1,2), (-2,1),(-1,2) and con-struction of tight wavelet frames. These frame frameletsare used for image edge detection. Comparison with otherbox spline wavelet frames will be shown to demonstratethat this wavelet frame works very well. Effectiveness onedge detection is also compared with standard wavelets,shearlets, classic Canny and segmentation based methods.

Weihong GuoDepartment of Mathematics, Case Western [email protected]

Ming-Jun LaiUniversity of GeorgiaDepartment of [email protected]

MS68Restoration of Photographies of Paintings

In this talk, we will review the main limitations of photog-raphy in terms of image quality as low signal noise ratio,blur, compression artifacts. Other limitations are intro-duced specifically for photographies of paintings as no con-trolled illumination, no flash, highlights from the varnish,medium, or glass in front of the painting. We will discusshow to remove or reduce these limitations when we disposeof a series of photographies of the same painting.


Gloria HaroUniversitat Pompeu [email protected]

Jean-Michel MorelENS [email protected]

MS68Exact Histogram Specifcation for Digital ImagesUsing a Variational Approach

Histogram specification refers to the problem of transform-ing an input quantized (digital) image to an output (alsodigital) image that follows a prescribed histogram. Thisis an ill-posed problem for digital images since there canbe numerous pixels with the same intensity level, and therecan be numerous ways of assigning them new intensity levelaccording to the prescribed histogram. Classical histogrammodification methods being designed for real-valued im-ages, map these pixels into the same intensity level: thenthe output image deviates arbitrarily from the prescribedhistogram. So as to satisfy the prescribed histogram, allpixels of the input image must be rearranged in a strictly

ordered way, using some auxiliary attributes, while pre-serving the specific features of the input image. This is acrucial challenge in exact histogram specification for digi-tal images. We propose a new method that efficiently pro-vides a strict ordering for all pixel values. It is based on awell designed variational approach. Noticing that the in-put digital image contains quantization noise, we minimizean objective function whose solution is a real-valued imagewith reduced quantization noise. We show that all the pix-els of this real-valued can be ordered in a strict way. Thentransforming the latter image into another digital imagesatisfying a specified histogram is an easy task. This pro-vides a new tool to restore image sequences corrupted byflicker (unnatural temporal fluctuations of intensity, dueto variations in exposure time or inhomogeneous ageing ordegradation of the film support).

Mila NikolovaCentre de Mathematiques et de Leurs Applications(CMLA)ENS de Cachan, [email protected]

Raymond H. ChanThe Chinese Univ of Hong KongDepartment of [email protected]

You-Wei WenFaculty of ScienceKunming University of Science and Technology, [email protected]

MS68Learning Human Activities and Interactions fromVideo

In this talk, joint work with Alexey Castrodad, we willdescribe our advances in human activity recognition fromvideo. We will illustrate how very simple models achievestate-of-the-art results. Such models are based on dictio-nary learning and sparse coding, embedded in a hierarchi-cal structure.


MS68Variational Imaging on Manifolds and Some Appli-cations

In this talk we consider the problem of optical flow onmanifolds. We model appropriate equations describing themovements on manifold. Moreover, we show some partialdifferential equations for zooming of data on manifolds.


Nicolas ThorstensenUniversity of [email protected]

110 IS12 Abstracts

MS70A Direct D-bar Method for the Reconstruction ofComplex Admittivities from EIT Data

A direct reconstruction algorithm for complex conductivi-ties in W 2,∞(), where is a bounded, simply connected Lip-schitz domain in 2, is presented. The algorithm constitutesthe first D-bar method for the reconstruction of conductiv-ities and permittivities in two dimensions. Reconstructionsof numerically simulated chest phantoms with discontinu-ities at the organ boundaries are presented.

Sarah HamiltonColorado State [email protected]

Natalia HerreraUniversity of Sao [email protected]

Jennifer L. MuellerColorado State UniversityDept of [email protected]

Alan Von HerrmannColby [email protected]

MS70Problems in Electrical Impedance Tomography

In electrical impedance tomography one tries to recon-struct and display the electrical properties of the interior ofa body from measurements made on the surface or outsideof the body. One of the reasons that EIT may be useful indiagnosing diseases is that tissues have differing conductiv-ity depending on their state. For example lungs filled withair have a lower conductivity than lungs depleted of air.For this reason imaging the lungs conductivity may helpto diagnose problems with ventilation , or lung function.Similarly hearts filled with blood have a higher conductiv-ity than when emptied so that impedance imaging of theheart may help diagnose heart disease. Some breast can-cers seem to have different impedance properties from be-nign lesions. This suggests that impedance imaging mightbe helpful in diagnosing breast cancer. In this talk it willbe explained how impedance imaging can be seen as aninverse boundary value problem for Maxwell’s equations.Problems arising in the design and application of EIT sys-tems for medical imaging will be described. Movies andimages from our EIT systems will be shown.

David [email protected]

MS70Estimation of Spatial Resolution Improvement bythe use of an Anatomy Based Prior of Pig Thoraxesin Electrical Impedance Tomography

The inverse problem associated to the Electrical ImpedanceTomography (EIT) requires regularization, which can beinterpreted as prior information. Anatomy based priorsfor medical and veterinary images generated by EIT are ex-pected to improve spatial resolution. In the present work,the improvement of the spatial resolution is estimated. The

anatomy based prior is based on in vivo tissue impedi-tivity measurements and CT-Scan images of pig thoraxes.These in vivo measurements were necessary to capture vari-ations caused by perfusion of blood, muscle contractionand, sometimes, ventilation. A detailed numerical modelof a pig thorax was developed. Objects of simple geome-try were inserted in the numerical thorax model and theresulting EIT images were used to approximately quantifythe spatial resolution.

Raul G. LimaUniversity of So [email protected]

MS70Imaging of Non-stationary Flows in Industrial Pro-cess Tomography

In this paper, we consider the problem of imaging fluidsunder non-stationary flow fields. As the imaging modal-ity, we use the Electrical Impedance Tomography (EIT).We model the evolutions of targets with the Navier-Stokesequations and the convection-diffusion equation, and esti-mate the concentration distribution and the velocity fieldbased on EIT data using the state-estimation. Numeri-cal results suggest that tracking of complex flow fields ispossible when suitable flow models are utilized in state es-timation.

Aku Seppanen, Antti LipponenDepartment of Applied PhysicsUniversity of Eastern [email protected], [email protected]


MS71Regularization Based on Integral Invariants

We investigate the applicability of integral invariants asgeometrical shape descriptors in the context of ill-posedinverse problems. We propose the use of a Tikhonov func-tional, where the penalty term is based on the difffferenceof integral invariants. As a case example, we consider theproblem of inverting the Radon transform of an objectwith only limited data available.

Thomas FidlerUniversity of ViennaComputational Science [email protected]

MS71Morozov’s Principle for the Augmented LagrangianMethod

The Augmented Lagragian Method received much atten-tion recently (also under the name Bregman iteration), asan approach for regularizing inverse problems. This workshows convergence and convergence rates for this methodwhen a special a posteriori rule, namely Morozov’s discrep-ancy principle, is chosen as a stopping criterion. As poten-tial fields of application we study implications of these re-sults for particular examples in imaging, that is total vari-

IS12 Abstracts 111

ation regularization as well as q penalties with q ∈ [1, 2].

Elena ResmeritaAlpen-Adria University [email protected]

MS71Convergence of Variational Regularization Meth-ods for Imaging on Riemannian Manifolds

We consider abstract operator equations Fu = y, where Fis a compact linear operator between Hilbert spaces U andV , which are function spaces on closed, fin te dimensionalRiemannian manifolds, respectively. This setting is of in-terest in numerous applications such as Computer Visionand non-destructive evaluation. In this work, we studythe approximation of the solution of the ill-posed opera-tor equation with Tikhonov type regularization methods.We state well-posedness, stability, convergence, and con-vergence rates of the regularization methods. Moreover,we study in detail the numerical analysis and the numeri-cal implementation. Finally, we provide for three differentinverse problems numerical experiments.

Nicolas ThorstensenUniversity of [email protected]

MS71L-infinity Regularized Models for Segmenta-tion, Cartoon-texture Decomposition, and ImageRestoration

We propose a cartoon-texture separation using the pre-dualto the space of Lipschitz functions. Specifically, we decom-pose an image into the sum of a cartoon part, in BV , anda texture component in the pre-dual of W 1,∞. This de-composition is used to analyze the various features in theimage, remove noise from cartoon images, and also to helpbetter reconstruct the texture when degraded by (knownand semi-blind) blur. The algorithm is tested on both syn-thetic and real images, with various implementations basedon duality and on projectors.


MS72Open Science: When Open Source Meets Open Ac-cess

In the past decade, Open Source software has become moreand more prominent and pervasive in the domain of scien-tific research. In parallel, scientific communities have cometo the conclusion that open-access is essential to the futureof science. The Insight Journal is an open-access infras-tructure for collecting, validating and disseminating opensource code, data and knowledge to the scientific commu-nity; founded on the principle of enabling reproducibilityverification as a routine scientific practice.

Julien Jomier, Luis [email protected], [email protected]

MS72Image Processing On Line: Experiences in Repro-ducible Science and Software

Image Processing On Line (IPOL) publishes image algo-rithms with an attention to reproducibility: the descrip-tion and the implementation are peer-reviewed and pub-lished under free licenses, the software must be reusable,and a web test and archive interface allows extensive ex-ploration of the algorithm. With different publication pro-cedures and tools, new constraints, and new benefits forresearchers, the notions of quality and usefulness of a re-search article are reevaluated. Image Processing On Line.http://www.ipol.im/. ISSN:2105-1232

Nicolas LimareENS [email protected]


MS72Open Research Computation: Supporting Repro-ducible Research Through Open Source Software

The communication of research in a fashion which makes itreproducible remains a significant challenge. While compu-tational research is an area where high standards of repro-ducibility are in principle possible adherence to best prac-tice is rare. There is no real incentive for producing andsharing high quality software products related to research.Open Research Computation aims to develop incentives forresearchers first to publish the software products of theirresearch in a way that maximises the potential for their useand re-use through providing a traditional journal formatthat will develop a high profile.

Cameron [email protected]

MS72Reproducible Research in Signal Processing: Howto Increase Impact

Worries about the reproducibility of research results dateback to Descartes’ work Discourse on (Scientific) Method.However, in computational sciences, new approaches to re-producibility are required. We give an overview of our per-sonal experiences with reproducible research in the field ofsignal and image processing, and issues we ran into. Wealso present results from our reproducibility study on imageprocessing papers. Finally, we give indications of increasedimpact for papers with reproducible results.

Martin VetterliAudiovisual communications laboratoryEcole Polytechnique Federale de [email protected]

Patrick [email protected]

MS73Edge-Preserving Full-Waveform Inversion with a

112 IS12 Abstracts

Modified Total-Variation Regularization Scheme

Full-waveform inversion using acoustic or elastic waves hasbecome more and more feasible with increasing computingpower. The full-waveform inversion is usually ill-posed andrequires computational methodologies to stabilize the in-version. We develop a full-waveform inversion method witha modified total-variation regularization scheme to preserveedges in tomography reconstructions. We apply the newmethod to acoustic/elastic reflection data for geophysicalimaging and demonstrate the improved tomography capa-bility of the method.

Youzuo Lin, Lianjie Huang, Zhigang ZhangLos Alamos National [email protected], [email protected], [email protected]

MS73Solution of Ill-posed Inverse Problems Pertainingto Signal Restoration: Total Variation ParameterEstimation

Spectral decomposition assists understanding the solutionof ill-posed inverse problems. In the talk I consider applica-tion of standard approaches from Tikhonov regularizationfor finding appropriate regularization parameters in the to-tal variation augmented Lagrangian implementations. Theanalysis of the solution uses the SVD and GSVD, andassists with understanding how to pick relevant regular-ization parameters for the total variation implementation.Application to solid oxide fuel cells will also be discussed.

Rosemary A. RenautArizona State UniversitySchool of Mathematical and Statistical [email protected]

MS73L1/TV Models and Its Variants

A proximity operator framework for studying the L1/TVimage denoising model will be presented. The solutions ofthe L1/TV model are identified as fixed points of a nonlin-ear mapping expressed in terms of the proximity operatorof the L1-norm or L2-norm. This formulation naturallyleads to fixed-point algorithms for the numerical treatmentof the model.

Lixin ShenMathematics Department, Syracuse [email protected]

MS73A New Paradigm for Fast MRI Reconstructionwith TV Regularization from Arbitrary Trajecto-ries though K-Space

A novel, fast generalized non-Cartesian MRI reconstruc-tion technique is presented. The approach computes anuncorrected non-uniform Fourier Transform from the k-space samples and generates a convolution kernel to correcttrajectory induced errors. An iterative total variation reg-ularized deconvolution is then performed on a Cartesiangrid using fast Fourier Transforms to remove artifacts andblurring. The technique is demonstrated on synthesizedMRI data for comparison to Cartesian sampling for sev-eral k-space trajectories, including spiral and rosette.

Wolfgang Stefan

The University of Texas MD [email protected]

Ken-Pin Hwang2Applied Science Laboratory, GE [email protected]

R. Jason Stafford, John D HazleThe University of Texas MD [email protected], [email protected]

MS74Designing Optimal Spectral Filters for InverseProblems

Filtering suppresses the amplification of errors when com-puting solutions to ill-posed inverse problems; however, se-lecting good regularization parameters is often expensive.In many applications, data are available from calibrationexperiments. Here, we describe how to use such data topre-compute optimal filters. We formulate the problemin an empirical Bayes risk minimization framework. Nu-merical examples from image deconvolution illustrate thatour proposed filters perform consistently better than well-established filtering methods.

Matthias ChungDepartment of MathematicsTexas State [email protected]

MS74Imaging the Earths Conductivity and Changeabil-ity

Conductivity and chargeability of earth’s materials aregood proxies to mineralization and to hydrocarbons. Imag-ing these properties is therefore important in detecting andmanaging natural resources. In this talk we discuss theimaging of these properties under various conditions. Wewill show cases that traditional techniques fail and discussa new technique we have develop to overcome many of theexisting imaging roadblocks

Eldad HaberDepartment of MathematicsThe University of British [email protected]

MS74Fast Composite Splitting Algorithm for LinearComposite Regularization

In this talk, we consider the minimization of a smoothconvex function regularized by the composite prior mod-els. This problem is generally difficult to solve even ifeach subproblem regularized by one prior model is con-vex and simple. In this paper, we present two algorithmsto effectively solve it. First, the original problem is de-composed into multiple simpler subproblems. Then, thesesubproblems are efficiently solved by existing techniques inparallel. Finally, the result of the original problem is ob-tained by averaging solutions of subproblems in an iterativeframework. The proposed composite splitting algorithmsare applied to the compressed MR image reconstructionand low-rank tensor completion. Numerous experimentsdemonstrate the superior performance of the proposed al-gorithms in terms of both accuracy and computation com-

IS12 Abstracts 113

plexity.

Junzhou HuangDepartment of Computer Science and EngineeringUniversity of Texas at [email protected]

MS74Model-based Synthetic Aperture Radar Imaging

There are many significant advantages such as reducing therequired antenna size and reducing the needed number ofmeasurements that are possible when Compressive Sensing(CS) techniques are used for imaging using Synthetic Aper-ture Radar (SAR). CS deals with reconstructing imagesthat are sampled below the Nyquist rate provided certainconditions are met. These suggested techniques while pow-erful have, however, not adequately dealt with the fact thatSAR images contain a multiplicative component known asspeckle. Speckle makes the compressibility and interpreta-tion of SAR images difficult. In this talk, I will present amore complete formulation of the problem of reconstruct-ing the piecewise smooth and texture component (speckle)directly from a set of compressive measurements.


MS75Diffusion Descriptors for 3D Shape Analysis

Large databases of 3D models available in public domainhave created the demand for shape search and retrieval al-gorithms capable of finding similar shapes in the same waya search engine responds to text quires. Since many shapesmanifest rich variability, shape retrieval is often requiredto be invariant to different classes of transformations andshape variations. One of the most challenging settings inthe case of non-rigid shapes, in which the class of trans-formations may be very wide due to the capability of suchshapes to bend and assume different forms. In this talk,we will explore approaches to 3D shape retrieval analogousto feature-based representations popular in the computervision community. We will show how to construct invariantlocal feature descriptors based on heat diffusion and com-pute them efficiently using the Laplace-Beltrami operatorin order to represent 3D shapes as collections of geometric”words” and ”expressions” and how to adopt methods em-ployed in search engines for efficient indexing and search ofshapes.

Michael M. BronsteinInstitute of Computational Science Faculty of InformaticsUniversita della Svizzera italiana, USI, [email protected]

MS75Invariant Beltrami-Operators - A ComputationalStudy

In this talk we modify the metric in the Laplace-Beltramioperator so that it captures the task at hand. We show howvarious classical solutions in image processing and analy-sis could be formulated as equations in which the laplace-Beltrami, where the metric is carefully selected, plays amajor role. The next step is extending the Beltrami frame-work to surface processing and analysis. We deal with var-ious groups of invariants from which the right metric is

extracted, plugged into the operator, and demonstrate therobustness and efficiency of the resulting numerical solvers.

Ron KimmelTechnion, Haifa, [email protected]

MS75Unifying Beltrami and Mumford-Shah Frameworks

We present a unifying generalization of the Mumford-Shahfunctional, in the Ambrosio-Tortorelli set up, and the Bel-trami framework. The generalization of the Ambrosio-Tortorelli is in using a diffusion tensor as an indicator ofthe edge set instead of a function. The generalization ofthe Beltrami framework is in adding a penalty term onthe metric such that it is defined dynamically from min-imization of the functional. We show that we are able,in this way, to have the benefits of true anisotropic dif-fusion together with a dynamically tuned metric/diffusiontensor. The functional is naturally defined in terms of thevielbein-the metrics square root. Preliminary results showimprovement on both the Beltrami flow and the Mumford-Shah flow.

Nir SochenDepartment of Applied MathematicsTel Aviv University, [email protected]

MS75Geodesic Active Fields - A Geometric Frameworkfor Image Registration

Image registration is an ill-posed inverse problem whereone looks for the underlying deformation field that bestmaps one image onto another. Here, we propose to embedthe deformation field in a weighted minimal surface prob-lem. The energy of the deformation field is measured withthe Beltrami energy, weighted by local image mismatch.Minimizing this energy drives the deformation field towarda minimal surface, while being attracted by the solution ofthe registration problem.

Dominique ZossoEcole Polytechnique Federale de Lausanne (EPFL)Signal Processing Laboratory (LTS5)[email protected]

Xavier BressonUCLADepartment of [email protected]

Jean-Philippe ThiranSignal Processing Laboratory LTS5Ecole Polytechnique Federale de Lausanne (EPFL)[email protected]

MS76A Butterfly Algorithm for SAR Imaging

SAR image formation involves computing an oscillatory in-tegral which looks like a distortion of a Fourier transform.FFT-based algorithms exist and are fast, but they tend tobe inaccurate. I present an algorithmic alternative to theFFT, called the butterfly algorithm, to compute the SARimaging integral to within prescribed accuracy, in complex-

114 IS12 Abstracts

ity O(N log N) where N is the number of pixels. Joint workwith Matthew Ferrara, Nicholas Maxwell, Jack Poulson,and Lexing Ying.


MS76Compressive Illumination Waveforms for High Res-olution Radar Sensing

Compressive sensing algorithms guarantee stable recoveryof sparse signals from sub-Nyquist samples provided thatmutual coherence between the columns of the sensing oper-ator is low. For standard radar sensors with match filteringon receive the mutual coherence of the sensing matrix canbe linked to the maximum side lobe level of the Woodwardambiguity function. In this talk we introduce of a novelradar sensor with wideband illumination and sub-Nyquistsampling on receive. We show that for this system the mu-tual coherence of the sensing operator can be controlledthrough random wideband illumination to enable high res-olution range imaging from sub-Nyquist samples. Exper-imental results for this compressive sensing strategy willbe given based on low-rate sampling of staggered multifre-quency linear FM signals.

Emre ErtinThe Ohio State [email protected]

MS76Radar Imaging and Feature Detection of BiologicalMedia

In this paper we derive a method to reconstruct an im-age and detect features from far field scattering data ofhuman beings. We utilize physics and math-based theoryto determine an appropriate model for the scattering data.We subsequently use this data model to find the appropri-ate algorithm that will reconstruct an image of the humanbeing and detect features such as head size, arm, and leglength.

Analee MirandaAir Force Research [email protected]

MS76Information Theory in Radar Imaging, Detectionand Classification

We discuss some of our recent results in radar signal pro-cessing ranging from waveform design for radar imaging,discovery of a novel statistical characterization of sea-clutter with implications to detection theoretic algorithms,and recent results in ATR (automatic target recognition)based on novel feature selection methodologies coupledwith probabilistic graphical approaches to classification.We attempt to place this in context of the larger panoramaof activities taking place within radar signal processing.

Raghu RajNaval Research [email protected]

MS77A Binary Constrained General Model in ImageProcessing

In this talk, we consider a class of problems given by abinary constrained model. The model covers several ap-plications in image processing, such as the total variationmodel for binary image restoration, Mumford-Shah modelfor image segmentation, multilabelling problem, etc. In or-der to solve this model, we introduce its exact relaxationand give an efficient algorithm based on the Fenchel dual-ity technique. Numerical results are presented in the endof the talk.

Yiqiu DongInstitute of Biomathematics and BiometryHelmholtz Zentrum [email protected]

MS77Optimality of Relaxations for Integer-ConstrainedProblems

Variational convex relaxations can be used to compute ap-proximate minimizers of optimal partitioning and multi-class labeling problems on continuous domains. We presentrecent developments on proving a priori upper bounds forthe objective of the relaxed problem, and on the numericalsolution of such problems.

Jan LellmannUniversity of [email protected]

MS77A Convex Shape Prior and Shape Matching Basedon the Gromov-Wasserstein Distance

We present a novel convex shape prior functional with po-tential for application in variational image segmentation.Based on the Gromov-Wasserstein distance an approxima-tion is derived that takes the form of an optimal trans-port problem with relaxed constraints. The approach in-herits the ability to incorporate vast classes of geometricinvariances beyond rigid isometries and to process addi-tional (non-geometric) feature information. The resultingfunctional can be minimized by standard linear program-ming methods and yields a unique assignment of a givenshape template to a given image. Numerical experimentsare reported that illustrate key aspects of the approach,and open problems are outlined.

Christoph SchnoerrUniversity of [email protected]

MS77Flow-Maximizing joint with Message-Passing

Computing ’cuts’ in the spatially continuous setting pro-poses a series of challenging binary constrained optimiza-tion problems in image processing. By exploring theirprimal-dual formulations and some specified Bregman di-vergence, a family of efficient and interesting algorithmscan be discovered with potential connections to the exist-ing two approaches: max-flow and message-passing.

Jing YuanDepartment of Computer Science

IS12 Abstracts 115

University of Western [email protected]

PP1Automated Parameter Selection Tool for Solutionto Ill-Posed Problems

In many ill-posed problems it can be assumed that the er-ror in the data is dominated by noise which is independentidentically normally distributed. Given this assumptionthe residual should also be normally distributed with simi-lar mean and variance. This idea has been used to developthree statistical diagnostic tests to constrain the region ofplausible solutions. This poster demonstrates a softwarepackage that aids in the generation of a range of plausibleregularization parameters.

Brianna CashUniversity of Maryland, College ParkApplied Math, Scientific Computing (AMSC)[email protected]

PP1A Short Time Existence/uniqueness Result fora Nonlocal Topology-Preserving SegmentationModel

Motivated by a prior applied work of Vese and Le Guyaderdedicated to segmentation under topological constraints,we derive a slightly modified model phrased as a functionalminimization problem, and propose to study it from a the-oretical viewpoint. The mathematical model leads to asecond order nonlinear PDE with a singularity at Du = 0and containing a nonlocal term. A suitable setting is thusthe one of the viscosity solution theory and, in this frame-work, we establish a short time existence/uniqueness resultas well as a Lipschitz regularity result for the solution.

Carole Le GuyaderINSA Rouen, [email protected]

Nicolas ForcadelCEREMADE, Universite [email protected]

PP1Phase Retrieval with Random Illumination

Obtaining an image from the magnitude of its Fouriertransform is an important problem in physical sciences.We will show that the reconstruction of phase retrieval isgreatly enhanced by random illumination, which can be re-alized by random phase/magnitude modulators or randommasks.

Wenjing Liao, Albert FannjiangUC [email protected], [email protected]

PP1A Generalized Simulated Annealing Approach toImage Reconstruction

It is common pratice to speed-up simulated annealing byallowing the cost function and/or the candidate-solutiongeneration mechanism to vary with temperature. We de-rive simple sufficient conditions for the global convergence

of such generalized simulated annealing algorithms. Theseconditions are surprisingly weak; in particular, they do notinvolve the variations of the cost function with tempera-ture. We show that our results can be successfully appliedto image reconstruction problems involving challenging op-timization tasks.

Marc C. RobiniCREATISINSA [email protected]

PP1Image Reconstruction with Multiresolution Edge-Continuation and Non-Quadratic Data Fidelity

The standard approach to image reconstruction, whichconsists in minimizing a cost function combining data-fidelity with edge-preserving regularization, tends to pro-duce noisy object boundaries and to create a staircase ef-fect. We propose to incorporate the smoothness of theedge field via an additional penalty term defined in a mul-tiresolution domain. We provide an efficient half-quadraticalgorithm to solve the resulting optimization problem, in-cluding the case when the data-fidelity is non-quadraticand the cost function non-convex.

Marc C. RobiniCREATISINSA [email protected]

Pierre-Jean VivergeDepartment of Electrical EngineeringINSA [email protected]

PP1Direct reconstruction from spherical mean data

Circular and spherical mean data arises in various mod-els of thermoacoustic and photoacoustic tomography whichare rapidly developing modalities for in vivo imaging. Wedescribe how summability respectively kernel based meth-ods can be applied in order to come up with new recon-struction techniques. We will show how suitable kernelscan be constructed. A detailed description for the mostimportant bivariate and trivariate case will be provided,leading to effective recovery of image and volume data fromspherical mean values, respectively. Finally, we present ourimplementation and some numerical examples.

Ruben SeyfriedHelmholtz Zentrum [email protected]

PP1A Nonconvex TVq-Model in Image Restoration

A nonconvex variational model is introduced which con-tains q-norm, q ∈ (0, 1), of image gradient as regular-ization. Such a regularization is a nonconvex compromisebetween support minimization and convex total-variationmodel. In finite-dimensional setting, existence of mini-mizer is proven, a semismooth Newton solver is introduced,and its global and locally superlinear convergence is estab-lished. The potential indefiniteness of Hessian is handledby a trust-region-based regularization scheme. Finally, the

116 IS12 Abstracts

associated model in function space is discussed.

Michael HintermullerDepartment of Mathematics, Humboldt-University [email protected]

Tao WuInstitute for Mathematics and Scientific ComputingKarl-Franzens-University of [email protected]


IS12 Speaker and Organizer Index

117

Italicizednamesindicatesessionorganizers.


DDarbon, Jerome, MS64, 3:30 Tue

Darde, Jeremi, MS42, 2:30 Mon

Davis, Anthony B., MS10, 2:00 Sun



Davis, Anthony B., MS32, 9:30 Mon

Davis, Anthony B., MS41, 2:00 Mon

De Mol, Christine, MS61, 9:30 Tue

Delon, Julie, MS58, 10:00 Tue

Demanet, Laurent, MS21, 4:30 Sun

Demanet, Laurent, MS21, 4:30 Sun

Demanet, Laurent, MS29, 9:30 Mon

Demanet, Laurent, MS38, 2:00 Mon

Demanet, Laurent, MS76, 6:00 Tue

Demaret, Laurent, CP11, 5:10 Tue

Derfoul, Ratiba, CP5, 4:30 Mon

Dogan, Gunay, CP4, 2:20 Mon

Dong, Bin, MS67, 6:00 Sun

Dong, Bin, MS28, 9:30 Mon

Dong, Bin, MS46, 4:30 Mon

Dong, Bin, MS55, 9:30 Tue

Dong, Yiqiu, MS77, 4:30 Tue

Dong, Yiqiu, MS77, 5:00 Tue

Drapaca, Corina, CP9, 5:10 Tue

EEasley, Glenn, CP1, 2:20 Sun

Easley, Glenn, CP11, 4:50 Tue

Eggermont, Paul, MS61, 10:00 Tue

Ehman, Richard L., IP1, 8:15 Sun

Elad, Michael, MS26, 4:30 Sun

Elad, Michael, MS67, 5:30 Sun

Elad, Michael, MS34, 9:30 Mon



Elbau, Peter, MS61, 9:30 Tue

Elbau, Peter, MS71, 2:00 Tue

Elser, Veit, MS53, 5:00 Mon

Ertin, Emre, MS76, 5:30 Tue

Esedoglu, Selim, MS61, 10:30 Tue

Esser, Ernie, MS31, 10:30 Mon


AAbubakar, Aria, CP5, 5:10 Mon

Aghasi, Alireza, CP1, 3:40 Sun

Almansa, Andrés, MS58, 10:30 Tue

Ambartsoumian, Gaik, MS8, 10:30 Sun

Ambikasaran, Sivaram, CP9, 6:10 Tue

Anastasio, Mark, MS37, 3:30 Mon

Arridge, Simon, MS37, 2:30 Mon

Ashok, Amit, MS48, 5:30 Mon

Aubailly, Mathieu, MS15, 2:00 Sun

Aujol, Jean-Francois, MS7, 9:30 Sun

Aujol, Jean-Francois, MS17, 2:00 Sun

BBaek, Hyoungsu, CP11, 5:50 Tue

Bajaj, Chandrajit, MS65, 3:30 Tue

Bal, Guillaume, MS32, 9:30 Mon

Bal, Guillaume, PD1, 12:00 Mon

Balzano, Laura, MS30, 10:30 Mon

Baraniuk, Richard G., MS13, 2:00 Sun

Becker, Stephen, MS12, 2:30 Sun

Bertalmío, Marcelo, MS58, 9:30 Tue



Betcke, Marta, MS56, 9:30 Tue



Beylkin, Gregory, MS38, 3:30 Mon

Beylkin, Gregory, MS65, 2:30 Tue

Biros, George, MS27, 5:00 Sun

Blomgren, Peter, CP9, 4:30 Tue

Bolte, Jérôme, MS33, 10:00 Mon

Bosch, Edward H., MS47, 4:30 Mon

Bosch, Edward H., MS47, 6:00 Mon

Boursier, Yannick, MS66, 3:00 Tue

Bredies, Kristian, MS40, 2:00 Mon



Bresson, Xavier, MS7, 9:30 Sun

Bresson, Xavier, MS75, 4:30 Tue

Brito, Carlos, MS49, 6:00 Mon

Bronstein, Michael M., MS75, 5:30 Tue

Brune, Christoph, MS18, 2:00 Sun

Brune, Christoph, MS23, 6:00 Sun

Buades, Toni, MS68, 3:30 Tue

Burgholzer, Peter, MS37, 2:00 Mon

CCahill, Nathan D., MS18, 2:00 Sun

Cahill, Nathan D., MS27, 4:30 Sun

Cahill, Nathan D., MS35, 9:30 Mon

Cahill, Nathan D., MS35, 10:00 Mon

Cai, Jianfeng, MS46, 5:00 Mon

Cakoni, Fioralba, MS42, 2:00 Mon

Calandra, Henri, MS38, 2:30 Mon

Calder, Jeff, CP10, 4:50 Tue

Cao, Yan, CP2, 5:50 Sun

Capdeboscq, Yves, MS20, 5:00 Sun

Carin, Larry, MS64, 2:30 Tue

Cartis, Coralia, MS6, 10:00 Sun

Cash, Brianna, PP1, 8:00 Sun

Castrodad, Alexey, MS47, 4:30 Mon

Cevher, Volkan, MS2, 9:30 Sun



Cevher, Volkan, MS33, 11:00 Mon

Chen, Ke, MS22, 6:00 Sun

Chen, Yunmei, MS23, 4:30 Sun

Chen, Yunmei, MS31, 9:30 Mon

Cheney, Margaret, MS10, 2:30 Sun

Cheney, Margaret, PD1, 12:00 Mon

Cheney, Margaret, MS50, 4:30 Mon

Cheney, Margaret, MS59, 9:30 Tue

Cheney, Margaret, MS76, 4:30 Tue

Chouzenoux, Emilie, MS33, 10:30 Mon

Chumchob, Noppadol, CP3, 9:30 Sun

Chung, Julianne, CP7, 2:40 Tue

Chung, Julianne, MS74, 4:30 Tue

Chung, Matthias, MS74, 4:30 Tue

Chung, Matthias, MS74, 4:30 Tue

Coifman, Ronald, MS39, 2:30 Mon

Condat, Laurent, MS25, 4:30 Sun

Condat, Laurent, MS33, 9:30 Mon

Costa, Eduardo Leite, MS44, 11:00 Mon

Cuesta, Eduardo, CP4, 3:40 Mon

Curtis, Frank E., MS6, 9:30 Sun

Czaja, Wojtek, MS47, 6:15 Mon

118


FFacciolo, Gabriele, MS43, 3:30 Mon

Fadili, Jalal, MS2, 9:30 Sun






Fadili, Jalal, MS33, 9:30 Mon

Farshchian, Masoud, MS59, 10:00 Tue

Felea, Raluca, MS8, 11:00 Sun

Ferrara, Matthew, MS50, 4:30 Mon

Ferrara, Matthew, MS59, 9:30 Tue

Fichtner, Andreas, MS38, 2:00 Mon

Fidler, Thomas, MS61, 9:30 Tue



Filbir, Frank, CP2, 4:30 Sun

Flenner, Arjuna, MS5, 9:30 Sun

Flenner, Arjuna, MS64, 2:00 Tue

Flenner, Arjuna, MS64, 2:00 Tue

Foi, Alessandro, MS43, 3:00 Mon

Fomel, Sergey, MS29, 9:30 Mon

Fonseca, Irene, MS22, 4:30 Sun

Foucart, Simon, MS39, 3:00 Mon

Freeman, Jeremy, MS17, 3:00 Sun

Frikel, Jürgen, CP1, 3:20 Sun

GGee, James, MS35, 9:30 Mon

Giannakis, Dimitris, MS53, 6:00 Mon

Gilles, Jerome, MS5, 10:30 Sun

Gilman, Mikhail, MS50, 5:30 Mon

Goldstein, Tom, MS33, 9:30 Mon

Goldstein, Tom, MS46, 6:00 Mon

Greer, John, MS47, 4:30 Mon

Greer, John, MS47, 5:30 Mon

Gribonval, Remi, MS2, 10:00 Sun

Griesmaier, Roland, MS42, 2:00 Mon



Gross-Kunstleve, Ralf, MS53, 5:30 Mon

Gu, David, MS28, 9:30 Mon



Gu, David, MS55, 9:30 Tue

Guo, Weihong, MS67, 4:30 Sun

Gurevich, Shamgar, MS45, 6:00 Mon

HHöft, Thomas, CP5, 4:50 Mon

Haber, Eldad, MS74, 5:30 Tue

Hadani, Ronny, MT2, 9:30 Mon

Hadani, Ronny, MS36, 2:00 Mon

Hadani, Ronny, MS45, 4:30 Mon

Hager, William, MS23, 4:30 Sun

Hager, William, MS31, 9:30 Mon

Hahn, Jooyoung, MS22, 4:30 Sun

Hahn, Jooyoung, MS22, 5:30 Sun

Hahn, Jooyoung, MS40, 3:00 Mon

Haltmeier, Markus, MS63, 9:30 Tue

Haltmeier, Markus, MS63, 9:30 Tue

Hamilton, Sarah, MS70, 3:30 Tue

Hartov, Alex, MS44, 9:30 Mon

Hao, Ning, CP11, 4:30 Tue

Haupt, Jarvis, MS39, 3:30 Mon

Hebert, Hans, MS54, 10:30 Tue

Heldmann, Stefan, MS18, 2:00 Sun




Heldmann, Stefan, MS35, 9:30 Mon

Helin, Tapio, MS13, 2:30 Sun

Hel-Or, Yacov, MS34, 11:00 Mon

Hensley, Scott, MS59, 9:30 Tue

Hero, Alfred O., MS6, 10:30 Sun

Herrmann, Felix J., MS21, 6:00 Sun

Hintermueller, Michael, MS3, 9:30 Sun

Holler, Martin, MS14, 3:30 Sun

Holtzman Gazit, Michal, MS27, 4:30 Sun

Horesh, Lior, IP5, 1:00 Tue

Huang, Junzhou, MS74, 5:00 Tue

Huang, Shanshan, CP4, 2:00 Mon

Hyvonen, Nuutti, MS42, 2:00 Mon

Hyvonen, Nuutti, MS51, 4:30 Mon

IIndyk, Piotr, MS2, 10:30 Sun

Indyk, Piotr, MS57, 10:00 Tue

Innanen, Kris, MS21, 5:30 Sun

Isaacson, David, MS70, 2:30 Tue

JJia, Xun, MS1, 9:30 Sun

Jia, Xun, MS1, 9:30 Sun

Jia, Xun, MS11, 2:00 Sun

Jia, Xun, MS28, 10:00 Mon

Joergensen, Jakob H., MS66, 2:30 Tue

Jomier, Julien, MS72, 5:00 Tue

Julien, Rabin, MS6, 11:00 Sun

KKalke, Martti, MS56, 10:00 Tue

Kang, Sung Ha, MS5, 9:30 Sun

Kang, Sung Ha, MS15, 2:00 Sun

Katsevich, Alexander, MS9, 3:30 Mon

Kervrann, Charles, MS26, 6:00 Sun

Kim, Jong Min, CP8, 5:10 Tue

Kim, Yunho, MS28, 10:30 Mon

Kimmel, Ron, MS75, 5:00 Tue

Klukowska, Joanna, CP9, 5:30 Tue

Knobelspiesse, Kirk, MS19, 5:00 Sun

Knudsen, Kim, MS60, 10:30 Tue

Koay, Cheng Guan, CP6, 2:00 Tue

Koestler, Harald, MS27, 5:30 Sun

Kohr, Holger, MS54, 10:00 Tue

Kolehmainen, Ville P., MS56, 9:30 Tue

Krishnan, Venky P., MS8, 10:00 Sun

Krolik, Jeffrey, MS50, 4:30 Mon

Kruschel, Christian, MS12, 3:00 Sun

Kublik, Catherine M., MS55, 10:30 Tue

Kunis, Stefan, MS4, 10:00 Sun

Kunisch, Karl, MS16, 2:30 Sun

Kunyansky, Leonid A., MS20, 4:30 Sun

Kunyansky, Leonid A., MS37, 2:00 Mon

Kutyniok, Gitta, MS67, 4:30 Sun

Kutyniok, Gitta, MS63, 11:00 Tue

Kwon, Ohin, MS3, 10:30 Sun


119


LLai, Ming-Jun, MS67, 5:00 Sun

Lai, Rongjie, MS28, 9:30 Mon



Lai, Rongjie, MS55, 9:30 Tue

Lakhal, Aref, CP8, 4:30 Tue

Langer, Andreas, CP8, 4:50 Tue

Langlois, Joachim, MS36, 2:00 Mon

Laurain, Antoine, MS3, 11:00 Sun

Laytner, Patrick, CP7, 2:20 Tue

Lechleiter, Armin, MS60, 10:00 Tue

Lellmann, Jan, MS77, 6:00 Tue

Lerman, Gilad, MS30, 9:30 Mon

Lerman, Gilad, MS52, 4:30 Mon

Levine, Stacey E., MS58, 9:30 Tue

Levine, Stacey E., MS68, 2:00 Tue

Lewis, Matthew A., MS24, 4:30 Sun

Lezoray, Olivier, CP10, 5:30 Tue

Lezoray, Olivier, CP10, 5:50 Tue

Li, Chunming, MS11, 2:00 Sun

Li, Maokun, CP10, 4:30 Tue

Liao, Wenjing, PP1, 8:00 Sun

Lima, Raul G., MS44, 9:30 Mon

Lima, Raul G., MS44, 10:00 Mon

Lima, Raul G., MS70, 3:00 Tue

Limare, Nicolas, MS72, 4:30 Tue

Limare, Nicolas, MS72, 4:30 Tue

Lin, Youzuo, MS73, 2:00 Tue

Lin, Youzuo, MS73, 3:30 Tue

Ling, Haibin, MS1, 10:30 Sun

Liu, Meng, CP6, 2:20 Tue

Lou, Yifei, MS5, 9:30 Sun




Lou, Yifei, MS58, 11:00 Tue

Louchet, Cecile, MS64, 3:00 Tue

Lu, Zhaosong, MS16, 3:00 Sun

Lui, Ronald Lok Ming, MS28, 9:30 Mon

Lui, Ronald Lok Ming, MS46, 4:30 Mon

Lui, Ronald Lok Ming, MS55, 9:30 Tue

Lui, Ronald Lok Ming, MS55, 9:30 Tue

Luke, Russell, MS16, 2:00 Sun

Luke, Russell, MS62, 10:00 Tue

Lyapustin, Alexei, MS19, 6:00 Sun

MMahanti, Prasun, CP11, 6:10 Tue

Maia, Filipe, MS62, 10:30 Tue

Malcolm, Alison, IP2, 8:15 Mon

Malcolm, Alison, MS21, 4:30 Sun

Malcolm, Alison, MS29, 9:30 Mon

Malcolm, Alison, PD1, 12:00 Mon

Malcolm, Alison, MS38, 2:00 Mon

Mang, Andreas, MS27, 6:00 Sun

Mao, Yu, MS15, 3:00 Sun

Mao, Yu, MS55, 10:00 Tue

Marchesini, Stefano, MS62, 9:30 Tue

Marchesini, Stefano, MS62, 9:30 Tue

Marin-Mcgee, Maider J., CP3, 11:10 Sun

Marquina, Antonio, MS32, 10:00 Mon

Marshak, Alexander, MS19, 5:30 Sun

Meyer, Francois G., MS34, 10:00 Mon

Micheli, Mario, MS5, 11:00 Sun

Milanfar, Peyman, MS26, 4:30 Sun

Milanfar, Peyman, MS26, 4:30 Sun

Milanfar, Peyman, MS34, 9:30 Mon



Miller, Eric, MS9, 2:00 Mon

Miranda, Analee, MS76, 5:00 Tue

Mizera, Ivan, MS63, 10:30 Tue

Modersitzki, Jan, MS18, 2:00 Sun

Modersitzki, Jan, MS27, 4:30 Sun

Modersitzki, Jan, MS35, 9:30 Mon

Moeller, Michael, MS41, 3:00 Mon

Moisan, Lionel, MS7, 10:30 Sun

Mokole, Eric, MS59, 10:30 Tue

Montanari, Andrea, MS2, 9:30 Sun

Moura, Fernando S., MS44, 10:30 Mon

Mueller, Jennifer L., MS60, 9:30 Tue



Muise, Robert R., MS39, 2:00 Mon

Muise, Robert R., MS48, 4:30 Mon

Muise, Robert R., MS57, 9:30 Tue

Muise, Robert R., MS57, 11:00 Tue

Munk, Axel, MS63, 9:30 Tue

NNagy, James G., MS56, 10:30 Tue

Neifeld, Mark, MS39, 2:00 Mon

Neylon, Cameron, MS72, 5:30 Tue

Nikolova, Mila, MS6, 9:30 Sun

Nikolova, Mila, MS16, 2:00 Sun

Nikolova, Mila, MS61, 11:00 Tue

Nikolova, Mila, MS68, 3:00 Tue

Nita, Bogdan G., CP5, 5:30 Mon

OOktem, Figen S., CP6, 2:40 Tue

Oktem, Figen S., CP11, 5:30 Tue

Oktem, Ozan, MS54, 9:30 Tue

Oktem, Ozan, MS65, 2:00 Tue

O’Leary, Dianne P., IP4, 8:15 Tue

Ortiz, Camillo, MS25, 5:00 Sun

Osher, Stanley J., MS23, 5:00 Sun

Osher, Stanley J., MS41, 2:30 Mon

Osher, Stanley J., MS47, 5:00 Mon

ÖÖktem, Ozan, MT2, 9:30 Mon

PPack, Jed, MS9, 2:30 Mon

Papafitsoros, Kostas, MS40, 2:30 Mon

Patel, Vishal, MS74, 6:00 Tue

Penczek, Pawel A., MS36, 2:30 Mon

Pennec, Xavier, MS18, 2:30 Sun

Peyré, Gabriel, MS7, 9:30 Sun





Peyré, Gabriel, MS33, 9:30 Mon




Pock, Thomas, MS25, 5:30 Sun

Pock, Thomas, MS49, 5:30 Mon

Prasad, Lakshman, MS10, 3:00 Sun

Preusser, Tobias, MS35, 11:00 Mon


120


QQiao, Zhijun, CP2, 4:50 Sun

Qin, Jing, CP1, 3:00 Sun

Qiu, Lingyun, CP2, 6:10 Sun

Quinto, Eric Todd, MS54, 9:30 Tue



RRada, Lavdie, CP7, 2:00 Tue

Raj, Raghu, MS76, 4:30 Tue

Raj, Raghu, MS76, 4:30 Tue

Ramirez, Carlos A., CP3, 10:10 Sun

Raskar, Ramesh, MS13, 3:00 Sun

Ravikumar, Pradeep, MS30, 11:00 Mon

Ravishankar, Saiprasad, CP1, 2:00 Sun

Renaut, Rosemary A., MS73, 2:00 Tue

Resmerita, Elena, MS71, 3:00 Tue

Ring, Wolfgang, MS65, 3:00 Tue

Ritterbusch, Sebastian, CP6, 3:20 Tue

Robini, Marc C., PP1, 8:00 Sun

Robini, Marc C., PP1, 8:00 Sun

Rodriguez, Paul, CP3, 10:50 Sun

Rohde, Shelley B., CP3, 10:30 Sun

Romberg, Justin, MS39, 2:00 Mon



Romberg, Justin, MS57, 9:30 Tue

Rozell, Chris, MS57, 10:30 Tue

Ruan, Dan, MS28, 11:00 Mon

Ruthotto, Lars, MS18, 3:30 Sun

SSabater, Neus, MS10, 3:30 Sun

Salmon, Joseph, MS26, 5:00 Sun

Sandryhaila, Aliaksei, CP1, 2:40 Sun

Sanghavi, Sujay, MS30, 9:30 Mon

Sankaranarayanan, Aswin, MS57, 9:30 Tue

Santosa, Fadil, PD1, 12:00 Mon

Sapiro, Guillermo, MS41, 2:00 Mon



Sapiro, Guillermo, MS68, 2:00 Tue

Saratoon, Teedah, MS37, 3:00 Mon

Saxena, Rishu, CP7, 3:20 Tue

Schaeffer, Hayden, CP4, 2:40 Mon

Scherzer, Otmar, MS20, 6:00 Sun

Scherzer, Otmar, MS68, 2:30 Tue

Schiffler, Stefan, MS4, 9:30 Sun



Schnoerr, Christoph, MS77, 4:30 Tue

Schoenlieb, Carola B., MS40, 2:00 Mon

Schoenlieb, Carola B., MS49, 4:30 Mon

Schoenmakers, Remco, MS65, 2:00 Tue

Schotland, John, MS20, 4:30 Sun

Semerci, Oguz, MS9, 2:00 Mon

Seo, Jin Keun, MS3, 9:30 Sun

Seppanen, Aku, MS51, 4:30 Mon

Seppanen, Aku, MS70, 2:00 Tue

Seyfried, Ruben, PP1, 8:00 Sun

Sgallari, Fiorella, MS32, 10:30 Mon

Shearer, Paul, CP3, 9:50 Sun

Shen, Lixin, MS73, 2:30 Tue

Shkolnisky, Yoel, MS36, 3:00 Mon

Shragge, Jeffrey C., MS29, 11:00 Mon

Sidky, Emil, MS66, 2:00 Tue

Sieling, Hannes, MS63, 10:00 Tue

Sigworth, Fred, MS45, 4:30 Mon

Siltanen, Samuli, MS13, 2:00 Sun

Siltanen, Samuli, MS56, 9:30 Tue




Simon, Martin, MS51, 5:30 Mon

Simoncelli, Eero P., MS17, 2:30 Sun

Singer, Amit, PD1, 12:00 Mon

Singer, Amit, IP3, 1:00 Mon

Singer, Amit, MS36, 2:00 Mon

Singer, Amit, MS45, 4:30 Mon

Singh, Aarti, MS52, 6:00 Mon

Slavine, Nikolai, MS24, 4:30 Sun

Soatto, Stefano, MS17, 2:00 Sun

Sochen, Nir, MS75, 4:30 Tue

Solna, Knut, MS59, 11:00 Tue

Staboulis, Stratos, MS51, 5:00 Mon

Staneva, Valentina, CP8, 5:30 Tue

Stefan, Wolfgang, MS73, 2:00 Tue

Stefan, Wolfgang, MS73, 3:00 Tue

Steinhauer, Dustin, MS20, 5:30 Sun

Stephani, Henrike, MS14, 3:00 Sun

Strehlow, Jan, MS4, 9:30 Sun



Stuff, Mark, MS50, 6:00 Mon

TTagare, Hemant, MS45, 5:00 Mon

Tamasan, Alexandru, MS3, 10:00 Sun

Tamminen, Janne P., MS60, 11:00 Tue

Taroudaki, Viktoria, CP4, 3:00 Mon

Teboulle, Marc, MS25, 4:30 Sun

ten Kroode, Fons, MS29, 10:00 Mon

Teuber, Tanja, MS40, 3:30 Mon

Theiler, James, MS41, 3:30 Mon

Thiran, Jean-Philippe, MS75, 4:30 Tue

Thorstensen, Nicolas, MS71, 2:00 Tue

Trzasko, Joshua D., CP9, 4:50 Tue

UUhlmann, Gunther, MT1, 9:30 Sun

Ullrich, Marcel, MS42, 3:30 Mon

VVan De Ville, Dimitri, MS26, 5:30 Sun

van Gennip, Yves, MS49, 5:00 Mon

Vandewalle, Patrick, MS72, 6:00 Tue

Vasconcelos, Ivan, MS21, 4:30 Sun

Vasconcelos, Ivan, MS21, 5:00 Sun

Veneziani, Alessandro, MS35, 10:30 Mon

Vese, Luminita A., MS10, 2:00 Sun

Vese, Luminita A., MS19, 4:30 Sun

Vese, Luminita A., MS32, 9:30 Mon

Vese, Luminita A., MS41, 2:00 Mon

Vese, Luminita A., MS71, 2:30 Tue

Virieux, Jean, MS29, 10:30 Mon

Voccola, Kaitlyn, MS50, 5:00 Mon

Volkmann, Neils, MS54, 11:00 Tue


121


Voroninski, Vladislav Y., MS62, 11:00 Tue

WWakin, Michael B., MS48, 6:00 Mon

Wan, Justin, CP7, 3:00 Tue

Wang, Jue, CP2, 5:30 Sun

Wang, Lanhui, MS36, 3:30 Mon

Wang, Mi, CP10, 5:10 Tue

Wang, Qiu, CP9, 5:50 Tue

Wang, Yan, CP8, 5:50 Tue

Wang, Yi, MS52, 5:00 Mon

Ward, Rachel, MS16, 3:30 Sun

Winkler, Christoph, CP2, 5:10 Sun

Wirth, Benedikt K., MS49, 4:30 Mon

Wong, Alvin Tsz Wai, MS55, 11:00 Tue

Wright, John, MS52, 4:30 Mon

Wu, Ru-Shan, MS38, 3:00 Mon

Wu, Tao, PP1, 8:00 Sun

YYan, Ming, CP4, 3:20 Mon

Yang, Chao, MS53, 4:30 Mon

Yang, Chao, MS53, 4:30 Mon

Yang, Chao, MS62, 9:30 Tue

Yanovsky, Igor, MS10, 2:00 Sun



Yanovsky, Igor, MS32, 9:30 Mon

Yanovsky, Igor, MS41, 2:00 Mon

Yashtini, Maryam, MS23, 4:30 Sun

Yashtini, Maryam, MS23, 4:30 Sun

Yashtini, Maryam, MS31, 9:30 Mon

Yazici, Birsen, MS8, 9:30 Sun

Yazici, Birsen, MS8, 9:30 Sun

Yazici, Birsen, MS9, 2:00 Mon

Yazici, Birsen, MS9, 3:00 Mon

Ye, Xiaojing, MS1, 9:30 Sun

Ye, Xiaojing, MS11, 2:00 Sun

Ye, Xiaojing, MS31, 10:00 Mon

Yin, Wotao, MS23, 5:30 Sun

Yuan, Jing, MS7, 10:00 Sun

Yuan, Jing, MS31, 11:00 Mon

Yuan, Jing, MS77, 5:30 Tue

Yuan, Ming, MS52, 5:30 Mon

ZZhang, Haili, MS11, 3:00 Sun

Zhang, Lei, MS34, 10:30 Mon

Zhang, Teng, MS30, 10:00 Mon

Zhang, Xiaoqun, MS31, 9:30 Mon

Zhao, Hongkai, MS46, 5:30 Mon

Zhao, Hongkai, MS66, 3:30 Tue

Zhao, Zhizhen, MS45, 5:30 Mon

Zheng, Fang, CP5, 5:50 Mon

Zheng, Yefeng, MS11, 2:30 Sun

Zhou, Ting, MT1, 9:30 Sun

Zhu, Jiehua, MS1, 10:00 Sun

Zhu, Wei, MS11, 3:30 Sun

Zhu, Xiang, MS5, 10:00 Sun

Zhu, Xiao Xiang, CP6, 3:00 Tue

Zosso, Dominique, MS75, 4:30 Tue

Zosso, Dominique, MS75, 6:00 Tue

Zwicknagl, Barbara, MS22, 5:00 Sun

122


IS12 Budget

51

Revenue

Registration $102,135

Total $102,135

Direct Expenses

Printing $2,400

Organizing Committee $2,100

Invited Speaker $12,750

Food and Beverage $15,000

Telecomm $4,300

AV and Equipment (rental) $13,700

Room (rental) $0

Advertising $5,400

Conference Staff Labor $23,600

Other (supplies, staff travel, freight, misc.) $1,600

Total Direct Expenses: $80,850

Support Services: *

Services covered by Revenue $21,285

Services covered by SIAM $39,022

Total Support Services: $60,307

Total Expenses: $141,157

* Support services includes customer service, accounting, computer support, shipping, marketing and other SIAM support staff. It also includes a share of

the computer systems and general items (building expenses in the SIAM HQ).

Conference BudgetSIAM Conference on Imaging Science

May 20-22, 2012Philadelphia, PA

Expected Paid Attendance: 300

151123


Doubletree by Hilton Hotel Philadelphia Center City Map

FSC logo text box indicating size & layout of logo. Conlins to insert logo.

Final Program and Abstracts - SIAM: Society for … Address DoubleTree by Hilton Hotel Philadelphia...

Documents

Transcript of Final Program and Abstracts - SIAM: Society for … Address DoubleTree by Hilton Hotel Philadelphia...