Inverse problems in earth observation and cartography Shape modelling via higher-order active...

34
Inverse problems in earth observation and cartography Shape modelling via higher-order active contours and phase fields Ian Jermyn Josiane Zerubia Marie Rochery Peter Horvath Ting Peng Aymen El Ghoul

Transcript of Inverse problems in earth observation and cartography Shape modelling via higher-order active...

Inverse problems in earth observation and cartography

Shape modelling via higher-order active

contours and phase fields

Ian JermynJosiane Zerubia

Marie RocheryPeter Horvath

Ting PengAymen El Ghoul

2

Overview

Problem: entity extraction from (remote sensing) images. Need for prior ‘shape’ knowledge.

Modelling prior shape knowledge: higher-order active contours (HOACs). Two examples: networks (roads) and circles (tree crowns). Difficulties.

Phase fields: What are they and why use them? Phase field HOACs. Two examples: networks (roads) and circles (tree crowns).

Future.

3

Problem: entity extraction

Ubiquitous in image processing and computer vision: Find in the image the region occupied by particular

entities. E.g. for remote sensing: road network, tree crowns,…

4

Calculate a MAP estimate of the region:

In practice, minimize an energy:

Problem formulation

R = argmaxR

P(RjI ;K )

P(RjI ;K ) / P(I jR;K )P(RjK )

R = argminR

E (R;I )

E (R;I ) = ¡ lnP(I jR;K ) ¡ lnP(RjK )

= E I (I ;R) +EG (R) +const

5

Building EG: active contours

A region is represented by its boundary, R = [], the ‘contour’.

Standard prior energies: Length of R and area of R. Single integrals over the

region boundary: Short range dependencies. Boundary smoothness.

Insufficient prior knowledge.

RR

S1

R2

6

Difficulties

(Remote sensing) images are complex.

Regions of interest distinguished by their shape. But topology can be non-trivial, and

unknown a priori. Strong prior information about the

region needed, without constraining the topology.

7

Building a better EG: higher-order active contours

Introduce prior knowledge via long-range dependencies between tuples of points.

How? Multiple integrals over the contour. E.g. Euclidean invariant two-point term:

Dependent°(t); _°(t)°(t0); _°(t0)

E (°) = ¡ZZ

dt dt0 _°(t) ¢_°(t0) ª (j°(t) ¡ °(t0)j)

8

Prior for networks

Gradient descent with this energy. A perturbed circle evolves towards a structure

composed of arms joining at junctions.

EG (R) =¸L(@R) +®A(R)

¡¯2

ZZdt dt0 _°(t) ¢_°(t0) ª (j°(t) ¡ °(t0)j)

r

r)

9

Prior for a gas of circles

The same energy EG can model a ‘gas of circles’ for certain parameter ranges. Which ranges?

A circle should be a stable configuration of the energy (local minimum). Stability analysis.

10

Phase diagrams

Circle Bar

11

Optimization problem

Minimize E(R, I) = EI(I, R) + EG(R). Algorithm: gradient descent using distance

function level sets. But the gradient of the HOAC term is non-local,

and requires: The extraction of the contour; Many integrations around the contour; Velocity extension.

12

Example I: HOAC results

13

Example I: HOAC results

14

Example II: HOAC results

15

Example II: HOAC results

16

Problems with HOACs

Modelling: Space of regions is complicated to express in the

contour representation. Probabilistic formulation is difficult.

Parameter and model learning are hampered.

Algorithm: Not enough topological freedom. Gradient descent is complex to implement for

higher-order terms. Slow.

Solution: ‘phase fields’.

17

Phase fields

Phase fields are a level set representation (z() = {x : (x) > z}), but the functions, , are unconstrained.

How do we know we are modelling regions?

R

R = 1

R = -1

E0(Á) =Z

­d2x

½D2r Á¢r Á+ ¸(

14Á4 ¡

12Á2)

¾

ÁR = arg minÁ: ³ (Á)=R

E0(R)

18

One can show that

R is a minimum for fixed R. Thus gradient descent with E0 mimics

gradient descent with L: ‘valley following’.

Can also add odd potential term to mimic:

Relation to active contours

R1

R2R3

E0(ÁR ) ' ¸C L(@R)

E0(ÁR ) ' ¸C L(@R) +®CA(R)

19

Why use them?

Complex topologies are easily represented. Representation space is linear:

can be expressed, e.g., in wavelet basis for multiscale analysis of shape.

Probabilistic formulation (relatively) simple. Gradient descent is based solely on the PDE

arising from the energy functional: No reinitialization or ad hoc regularization. Implementation is simple and the algorithm is fast.

20

Why use them?

Neutral initialization: No initial region. No bias towards “interior” or

“exterior”. Greater topological freedom:

Can change number of connected components and number of handles without splitting or wrapping.

+

21

How to write HOACs as phase fields?

Use that rR is zero except near R, where it is proportional to the normal vector.

One can show that

EQ (R) = ¡¯C2

ZZdt dt0 _°(t) ¢GC (°(t);°(t0)) ¢_°(t0)

EN L (Á) = ¡¯2

ZZ

­ 2d2x d2x0r Á(x) ¢G(x;x0) ¢r Á(x0)

EN L (ÁR ;¯ ;G) ' EQ (R;¯C ;G)

22

Phase fields : likelihood energies EI

r : normal vector to the contour. r¢r : boundary indicator. (1 § )/2 : characteristic function of the region

(+) or its complement (-). Using these elements, one can construct the

equivalents of active contour and HOAC likelihoods.

23

Optimization problem

Minimize

Algorithm: gradient descent, but…

E (R;I ) = E I (I ;R) +EG (R)

= E I (I ;R) +E0(R) +EN L (R)

24

Major advantage of phase fields for HOAC energies

Whereas, due to the multiple integrals, HOAC terms require Contour extraction, contour integrations, and force

extension, Phase field HOACs require only a

convolution:

±EN L±Á(x)

= ¯Z

­d2x0r 2G(x ¡ x0)Á(x0)

25

Phase field HOACs: results

26

Phase field HOACs: VHR image results

27

Phase field HOACs: tree crown results

28

Phase field HOACs: results

29

Future

New prior models Directed and rectilinear networks (rivers, big

cities…). ‘Gas of rectangles’; Controlled perturbed circle.

Multiscale models; New algorithms (multiscale, stochastic,…); Parameter estimation; Higher-dimensions; …

30

Thank you

31

Stability analysis

EC;g(°0+ ±°) =

"

¸C2¼r0+ ®C¼r20 ¡

¯C22¼

Z 2¼

0F00dp

#

+a0

"

¸C2¼+ ®C2¼r0 ¡¯C24¼

Z 2¼

0F10dp

#

+X

kjakj

2"

¸C¼r0k2+ ®C¼

¡¯C24¼

( Ã

2Z 2¼

0F20dp+

Z 2¼

0F21e

ir0kpdp

!

+ k

Ã

2ir0Z 2¼

0F23e

ir0kpdp

!

+k2Ã

r20

Z 2¼

0F24e

ir0kpdp

! ) #

32

HOACs: EI

Linear term that favours large gradients normal to contour.

Quadratic term that favours pairs of points with tangents and image gradients parallel or anti-parallel.

E I ;1(°) =Z

S1dt _°(t) £ r I (°(t))

tangent vectors

gradients

weighting function

E I ;2(°) = ¡ZZ

dt dt0 _° ¢_°0(r I ¢r I 0) ª (j°(t) ¡ °(t0)j)

34

Nonlinearity in the model, not the representation: example

Probability distribution P on R2. P pushes forward to Q on S1. If P is strongly peaked at r0 then

Q() ' P(r0, ). Gradient descent with –ln(P) on R2

mimics gradient descent with -ln(Q) on S1 (‘valley following’).

P (r;µ) = d2x e¡ E (r;µ)

£ exp ¡ ( r4

4 ¡ r22 )

35

Turing stability

To avoid decay of interior and exterior, we require stability of functions § = § 1: 2E/2 must be positive definite (E = E0 + ENL).

For prior terms, this is diagonal in the Fourier basis:

Gives condition on parameters. Better result would be existence and

uniqueness of R for ‘any’ R.

±2E±Á(k0)±Á(k)

(Á§ ) = ±(k;k0)nk2(D ¡ ¯G(k)) +2(¸ ¨ ®)

o