Text, Speech, and Vision for Video Segmentation: The...
Transcript of Text, Speech, and Vision for Video Segmentation: The...
AbstractWe describe three technologies involved in
creating a digital video library suitable for full-content search and retrieval. Image processinganalyzes scenes, speech processing transcribesthe audio signal, and natural language process-ing determines word relevance. The integrationof these technologies enables us to include vastamounts of video data in the library.
1 IntroductionThe Informedia Digital Video Library Project
at Carnegie Mellon University is creating a digitallibrary of text, images, videos and audio data availablefor full content retrieval [Stevens94][Christel94]. Theinitial testbed will be installed in several K-12 schoolsand students will use the system to explore multi-mediadata for educational purposes. The Informedia systemfor video libraries goes far beyond the current paradigmof video-on-demand by retrieving a short video para-graph in response to the user’s query. The project can bedivided into two phases: library creation and libraryexploration (See Figure 1).
1.1 Library creationThe Informedia project is creating intelligent,
automatic mechanisms for populating a video libraryand allowing for its full-content and knowledge-basedsearch and segment retrieval. The material is obtainedfrom video assets of WQED/Pittsburgh as well as theBritish Open University video courses.The project usesthe Sphinx-II speech recognition system to transcribeand align narratives and dialogues automatically. Theresulting transcript is then processed through methodsof natural language understanding to extract subjectivedescriptions and mark relevant key words. Acoustic sig-nal analysis identifies potential segment boundaries of“paragraph” size. Within a paragraph, scenes are iso-lated and clustered into video segments through the useof various image understanding techniques. These com-ponents are described in Figure 2.
1.2 Library explorationUsers are able to explore the Informedia library
through an interface that allows them to search using
typed or spoken natural language queries, select relevantdocuments retrieved from the library and display thematerial on their PC workstations. The library retrievalsystem can effectively process spoken queries anddeliver relevant video data in a compact format, basedon information embedded with the video during librarycreation. Video and other data may be explored in depthfor related content. During retrieval based on keywordsearches by a user, only the relevant video segments aredisplayed. Prototype exploration systems have beenimplemented on both Macintosh and PC platforms.
In this paper we will focus on the library cre-ation aspect of the Informedia Project. In particular, we
Indexed
StorePresentation
Video Segments
TV Footage
Extra Footage
New Video Footage
Raw Videoaudio video
VideoDatabase
OFFLINE
ONLINE
CREATION
EXPLORATION
IndexedTranscript of
Text
SegmentedDescribed
Video
IndexedTranscript of
Text
SegmentedDescribed
Video
Video Library
Distribution or Sale to Users
Visual, Spoken, and NaturalLanguage Query
Interactive VideoSearch
Speech & LanguageInterpretation and Indexing
Video Segmentationand Description
Figure 1: Overview of the Informedia DigitalVideo Library System
Text, Speech, and Vision for Video Segmentation:The InformediaTM Project
Alexander G. Hauptmann Michael A. SmithSchool of Computer Science Dept. of Electrical and Computer EngineeringCarnegie Mellon University Carnegie Mellon University
Pittsburgh, PA 15213 Pittsburgh, PA [email protected] [email protected]
will describe how to segment a video meaningfullyusing the integration of different technologies. Throughthe combined efforts of Carnegie Mellon’s speech,image and natural language processing groups, this sys-tem provides a robust tool for segmenting many types ofvideo data in order to utilize them within a digital videolibrary.
2 Video SegmentationGenerally, the videos in the Informedia Library
are full one-hour feature broadcast videos based on edu-cational documentaries. To allow efficient access to therelevant content of the videos, we need to separate theminto small pieces. To answer a user query by showing anhour long video is rarely a reasonable response.
The Informedia library creation phase usesthree different levels of segmentation for a video. Thefirst and generally largest segment shows a “video para-graph”, which consists of a series of related scenes witha common content. The second level of segmentationidentifies a single scene on the video within the videoparagraph. Finally, within a single scene we also need tobe able to select a representative frame icon for staticdisplays.
2.1 Video paragraphsWhen a user receives the response to a query,
the system needs to determine how much content andcontext to display. Where should the video clip start andwhere does it end? The answer to this is partly deter-mined by the content of the user query. But the answer isalso dependent on natural segments within the videowhich we call “video paragraphs”. In the ideal case, avideo paragraph starts at the natural boundary of the rel-evant content and ends wherever the video moves to adifferent context.
2.2 Individual scenesSegment breaks produced by image processing
are examined along with the boundaries identified bythe speech and natural language processing of the tran-script, and an improved set of segment boundaries areheuristically derived to partition the video paragraphsinto scenes.
All frames from each new scene will be used toselect the frame icon. This technique will allow for theinclusion of all relevant image information in the videoand the elimination of redundant data.
Scene Isolation
Representative Frame Icon
- Image Analysis
- Keyword Search
Speech Transcript- Keywords
Video Paragraph - Speech SNR
Figure 2: Combined technology to select representative frame (icon)
they
are
toug
hth
eyar
ede
man
ding
they
are
the
jury
ever
yto
ym
anuf
actu
rer
hope
sto pl
ease
the
crea
tors
of the
new
toy
knex
have
rece
ived
a larg
eam
ount
of
2.3 Frame iconsFor the purposes of static displays, the most
characteristic frame of a scene is included in static (non-animated) representations of the user’s selection. A sin-gle frame is displayed as the representative for thewhole video segment. This is used in an outlined displayshowing the results of a user query.
Showing frame icons allows the user to simul-taneously look at a static representation of multiplevideo paragraphs and to obtain some information abouttheir content and possible relevance to the user’s query,before selecting any one paragraph for playback. Frameicons are also important as encapsulations of the videoparagraph for printed reports and viewgraphs.
In order to create these various levels of seg-mentation, we integrate a number of different technolo-gies which will be described in the next section.
3 Component TechnologiesThere are 3 broad categories of technologies
we can bring to bear to the problem of identifying videosegments from broadcast video materials.
a. Text processing looks at the textual (ASCII)representation of the words that were spoken,as well as other annotations derived from thetranscript, the production notes or the close-captioning that may be available.
b. Speech signal analysis provides the basis foranalyzing the audio component of the material.
c. Image analysis looks at the images in thevideo-only portion.
Currently in the library creation phase of theInformedia Digital Video Library the following specificapproaches are used to create segmentation information.
3.1 Text AnalysisText analysis can work on an existing ASCII
transcript to help segment the text into paragraphs. Ananalysis of keyword prominence allows us to identifyimportant sections in the transcript [Mauldin 89]. Othermore sophisticated language based criteria are underinvestigation. The notion of semantic connectionsbetween text portions might be exploited for segmenta-tion as well. Currently we use two main techniques innatural language analysis.
a. If we have a complete time aligned transcriptavailable from the close-captioning or througha human generated transcription, we canexploit natural “structural” text markers suchas punctuation to identify segments of videoparagraph granularity
b. To identify and rank the contents of varioussegments, we use the well-known technique ofTF/IDF (term frequency/inverse document fre-quency) to identify critical keywords and theirrelative importance for the video document[Salton83].
3.2 Speech AnalysisSpeech analysis operates only on the audio
portion of the video. Using speech recognition we canobtain a transcript, although it may contain errors. Wecan also detect transitions between speakers and topicswhich are usually marked by silence or low energy areasin the acoustic signal.
RecognitionTo transcribe the content of the video material,
we recognize spoken words with the Sphinx-II speechrecognizer. The CMU Sphinx-II system uses semi-con-tinuous Hidden Markov Models to model context-dependent phones (triphones), including between wordcontext [Hwang94]. The recognizer processes an utter-ance in 3 steps: It makes a forward time synchronouspass using full between word models, Viterbi scoringand a trigram language model. This produces a word lat-tice where words may have only one begin time but sev-eral end times. The recognizer then makes a backwardpass which uses the end times from the words in the firstpass and produces a second lattice which contains multi-ple begin times for words. An A* algorithm is used togenerate the best hypothesis from these two lattices. Thelanguage model consists of words (with probabilities),bigrams/trigrams which are word pairs/triplets with con-ditional probabilities for the last word given the previ-ous word(s). The language model was constructed froma corpus of news stories from the Wall Street Journalfrom 1989 to 1994 and the Associated Press news ser-vice stories from 1988 to 1990. Only trigrams that wereencountered more than once were included in the model,but all bigrams and the most frequent 58800 words inthe corpus were included [Rudnicky95].
Processing the video tape using the speech rec-ognition system gives us a transcript. This transcriptcontains errors, which depending on the quality of thetape and the subject matter, currently range from 20% to70% word error rate.
Acoustic SegmentationTo detect breaks between utterances we use a
modification of Signal to Noise ratio (SNR) techniqueswhich compute signal power. This algorithm computesthe power of digitized speech samples where Si is a pre-
emphasized sample of speech within a frame of 20 mil-liseconds. A low power level indicates that there is little
Power1n---
Si2
∑⋅
log=
active speech occurring in this frame (low energy). Seg-mentation breaks between utterances are set at the mini-mum power as averaged over a 1 second window. Toprevent unusually long segments, we force the system toplace at least one break within 30 seconds.
3.3 Image Analysis Image analysis is primarily used for the identi-
fication of breaks between scenes and the identificationof a single static frame icon that is representative of ascene.
Histogram AnalysisVideo is segmented into scenes through the use
of comparative difference measures [Zhang93]. Imageswith small histogram disparity are considered to be rela-tively equivalent. By detecting significant changes inthe weighted color histogram of each successive frame,image sequences can be separated into individualscenes. A comparison between the cumulative distribu-tions is used as a difference measure. The histogram dif-ference plot is shown in the bottom graph of Figure 3.
This result is passed through a high pass filter to furtherisolate peaks and an empirical threshold is used to selectonly those regions where scene breaks occur. To makethe analysis more robust, we examine the individualimages in tiled subwindows. This reduces the noise inour difference data and compensates for motionbetween frames. The images are initially subsampled toprovide an efficient means of computation. Using onlythe histogram difference, we have achieved 90% accu-racy on a test set of roughly 200,000 video images (2hours).
Optical FlowOne important method of visual segmentation
and description is based on interpreting camera motion
0 20 40 60 80 100 1200
10
20
30
40
50Motion Vector Confidence Measure
0 20 40 60 80 100 1200
50
100
150
Frames
Histogram Difference Analysis
Figure 3: Scene segmentation and motionvector error.
[Akutsu94]. We can interpret camera motion as a pan orzoom by examining the geometric properties of theoptical flow vectors. Using the Lucas-Kanade gradientdescent method for optical flow, we can track individualregions from one frame to the next [Lucas81]. Bymeasuring the velocity that individual regions show overtime, a motion representation of the scene is created.Figure 4 shows examples of the optical flow analysis fordifferent types of camera motion. Drastic changes in thisflow describe random motion, and therefore, new scenes.These changes will also occur during gradual transitionsbetween images such as fades or special effects.
Only regions of low ambiguity are selected fortracking. Trackable regions are found by searching theentire image for subwindows whose gradient derivativesexhibit relatively similar eigenvalues. In order toaccurately track a region over large areas, a multi-resolution structure is used. With this structure we cantrack regions across many pixels and reduce the timeneeded for computation. When optical flow is minimalthe frames are suitable for an iconic framerepresentation. Since we are primarily interested indistinguishing static frames from motion frames, it wassufficient to track only the top 30 regions.
These techniques work well when scenechanges are abrupt, however, camera motion and grad-ual changes can severely affect the accuracy of the sys-tem. The first graph in Figure 3 shows the optical flowerror for a given sequence. When changes are gradual,we combine the optical flow results with histogramanalysis. This allows for segmentation under conditionsthat do not involve drastic changes in image content anddetection accuracy as high as 95%.
Figure 4: Camera motion analysis using opticalflow. Flow vectors are amplified forvisibility.
I0 I1 I2 Flow
4 Technology SynthesisWe now describe how we integrate the differ-
ent component technologies. In our early work on theInformedia digital video library, all segmentation wasdone by hand. We have now moved to a procedurewhere the segmentation boundaries are suggested by thesystem, but adjusted and verified by a person supervis-ing the digital video library creation process. Eventuallywe will transition from computer-assisted procedures tofully automatic video segmentation, as the algorithmsdescribed above become better tested and more robust.
Our current library creation process starts witha raw digitized video tape. The audio portion is fedthrough the speech analysis routines, which produces atranscript of the spoken text. The speech signal is alsoanalyzed for low energy sections that indicate acoustic“paragraphs” through silence. This is the first pass atsegmentation. If a close-caption transcript is available,we use that instead of the speech recognition output,since it is less errorful.
The transcript is processed by the natural lan-guage system and important keywords are identified.Using the results returned from the image analysis, we
then match the acoustic paragraph to the nearest scenebreak. This gives us an appropriate video paragraph clipin response to a user’s request.
The keywords and their corresponding para-graph locations in the video are indexed in the informe-dia library catalogue. To obtain video clips suitable forviewers, we first search for keywords from the userquery in the recognition transcript. When we find amatch, the surrounding video paragraph is returned.
For a static icon representative of a video clip,we place most emphasis on the image data. The para-graph is determined by the transcript and keywords.Within the paragraph the most prominent keywordsidentify the most prominent scene. The scene bound-aries are determined by the image analysis of color his-togram differences and optical flow analysis. Figure 5shows the integration of technologies used by the sys-tem.
5 ConclusionWe are currently using these techniques to cre-
ate digital video library collections suitable for full con-tent retrieval. While some of the steps are not yet fully
Scenes
Histogram Scene Analysis
Figure 5: Analysis of scene changes in video and audio signal
desp
itehe
roic
effo
rts
man
y of the
wor
lds
wild
crea
ture
sar
edo
omed the
loss of
spec
ies
isno
wth
esa
me
asw
hen
the
grea
tdi
nosa
urs
beco
me
extin
ctw
illth
ese
crea
ture
sbe
com
eth
edi
nosa
urs of our
time
toda
ym
anki
nd isch
angi
ng the
entir
efa
ce ofpl
anet
eart
h ...
Audio Segments and Text
0 1 2 3 4 5 6 7 8 9 10
x 105
−2
1.5
−1
0.5
0
0.5
1
1.5x 10
4
Samples
Audio Signal
integrated, each one has been shown to work indepen-dently, and several of the techniques are fully integratedwithin the informedia system.
Another use of the combined technologies willbe the development of the video skim [Smith95]. Byonly presenting significant regions, a short synopsis ofthe video paragraph can be used as a preview for theactual segment.
The Informedia Project will establish an on-line digital video library consisting of over 1000 hoursof video material. In order to be able to process this vol-ume of data, practical, effective and efficient tools areessential. We have outlined a practical set of techniquesfor video segmentation that allows us to automaticallyprocess the volume of data required.
6 References[Akutsu94] Akutsu, A. and Tonomura, Y. “Video
Tomography: An efficient method forCamerawork Extraction and MotionAnalysis,” Proc of ACM Multimedia‘94, Oct. 15-20, 1994, San Francisco,CA, pp. 349-356.
[Christel94] Christel, M., Stevens, S., & Wactlar, H.“Informedia Digital Video Library,”Proceedings of the Second ACM Inter-national Conference on Multimedia,Video Program. New York: ACM, Octo-ber, 1994, pp. 480-481.
[Hwang94] Hwang, M., Rosenfeld, R., Thayer, E.,Mosur, R., Chase, L., Weide, R., Huang,X., Alleva, F., “Improving Speech Rec-ognition Performance via Phone-Depen-dent VQ Codebooks and AdaptiveLanguage Models in SPHINX-II.”ICASSP-94, vol. I, pp. 549-552.
[Lucas 81] Lucas, B.D., Kanade, T. “An IterativeTechnique of Image Registration and ItsApplication to Stereo,”Proc. 7th Inter-national Joint Conference on ArtificialIntelligence, pp. 674-679, August 1981.
[Mauldin89] Mauldin, M. “Information Retrieval byText Skimming,” PhD Thesis, CarnegieMellon University. August 1989.Revised edition published as “Concep-tual Information Retrieval: A CaseStudy in Adaptive Partial Parsing, Klu-wer Press, September 1991.
[Rudnicky95] Rudnicky, A., “Language Modelingwith Limited Domain Data,” Proceedingof the 1995 ARPA Workshop on SpokenLanguage Technology, in press.
[Salton83] Salton, G., McGill, M.J. “Introductionto Modern Information Retrieval,”McGraw-Hill, New York, McGraw-HillComputer Science Series, 1983.
[Stevens94] Stevens, S., Christel, M., Wactlar, H.“Informedia: Improving Access to Digi-tal Video”. Interactions1 (October1994), pp. 67-71.
[Zhang93] Zhang, H., Kankanhalli, A., and Smo-liar, S. “Automatic partitioning of full-motion video,” Multimedia Systems(1993) 1, pp. 10-28.
[Smith95] Smith, M., Kanade, T., “Video Skim-ming for Quick Browsing Based onAudio and Image Characterization,” CSTechnical Report, Carnegie Mellon Uni-versity, Summer 1995.
7 AcknowledgmentThe authors would like to thank Howard Wact-
lar and the other members of the Informedia Project fortheir valuable discussions and contributions. This workis partially funded by the National Science Foundation,the National Space and Aeronautics Administration, andthe Advanced Research Projects Agency.