Images, integral to digital media, are large-ly the reason why the World Wide Web’s

popularity quickly surpassed that of other Internet-based systems such as Usenet News and Gopher.Images, however, present designers with problemsassociated with creation, processing, recognition, stor-age, retrieval, and transmission. Specifically, we mustconsider how to describe properties, or extract features,of an image; how to classify objects in the image; and

how to compare objects in differentimages. These tasks require a vari-ety of analysis techniques, withwhich we can then enable comput-ers to understand image featuresand contents. Table 1 summarizesthe basic image analysis techniquesand their applications.

Traditional image analysis appli-cations include robot vision, auto-mated product inspection, andoptical character recognition.1 Thetask of image analysis systems—toaccurately describe a given image—is almost effortless for humans butdifficult for computers to performadequately. Nevertheless, manyapplications have successfullyapplied image analysis underrestricted conditions.

In this broad-based overview, wereview image analysis for threeclasses of digital media applications:

■ traditional signal processing related procedures,including image enhancement, restoration and com-pression;

■ relatively new media creation and data retrieval prob-lems, including computer animation and multimediadata retrieval; and

■ pattern recognition, including document imaging andelectronic books.

We discuss specific problem areas, and techniques thataddress some of these, and briefly identify opportuni-ties for future research.

Signal processing applicationsLow-level image analysis procedures treat an image as

a two-dimensional signal. A video sequence can be con-sidered a time-varying 2D, or a 3D, signal. The opera-tions to analyze the sequence include enhancement,filtering, restoration, transformation, reconstructionand compression, which are traditional image process-ing problems. Most techniques are developed as anextension of one-dimensional signal processing meth-ods. Some image processing algorithms such as edgedetection, a form of 2D filtering, are developed specifi-cally for 2D applications. In this section, we considerthree problems of image processing most relevant tomedia applications.

Image enhancement and restorationFor gray-image enhancement, we commonly use meth-

ods such as histogram equalization and filtering toimprove image contrast and reduce noise.1 A color imagecomplicates enhancement because we must retain thecolor appearance and consider both luminance contrastand color contrast. To solve the problem, we can scale theR(x, y), G(x, y), and B(x, y) components, where x and yrepresent the pixel position, using the same factor.14 Thenew color components become k(x, y)R(x, y), k(x, y)G(x,y), and k(x, y)B(x, y), respectively. Function k(x, y) canbe determined from the spatial luminance and color infor-mation of the image.14 Figure 1 shows an example.

Color image enhancement procedures are more use-ful when combined with image segmentation in imageand video editing and composition applications. Forexample, to enhance a selected object or area, we needautomated boundary detection of the object. Automat-ed detection in a complex scene requires sophisticatedimage edge detection, segmentation, and data cluster-ing algorithms. That is, we must integrate low- and high-level image analysis techniques to improve the overallsystem performance.

Image restoration’s task is to correct distortions inthe image, such as noise and artifact reduction andde-blurring. To solve a restoration problem, we mustbuild a mathematical model describing the image for-mation process and then find the inverse of the modelto determine the source image. Unfortunately, theproblem is usually ill-posed and we often have to

0272-1716/01/$10.00 © 2001 IEEE

Tutorial

18 January/February 2001

Image analysis plays an

important role in many

digital-media-related

applications, as this

overview of analysis

techniques explains.

Although researchers have

developed a variety of

algorithms to solve many

problems, numerous

challenges remain.

Hong YanCity University of Hong Kong

Image Analysis forDigital MediaApplications

reconstruct the image from incom-plete information.

Maximum entropy2 and projec-tion onto convex sets (POCS)3,15 aretwo effective algorithms for solvingrestoration problems. Figure 2shows an example of magnetic res-onance image (MRI) restorationfrom motion-corrupted data. Imagerestoration is application depen-dent, and existing methods are usu-ally iterative. It remains an activeresearch problem to find robust, sta-ble, and fast algorithms to solveproblems in this area.

IEEE Computer Graphics and Applications 19

Table 1. Three levels—low, intermediate, and high—of basic image analysis methods and their applications. Thisclassification is approximate, as one method can apply to different task levels.

Type Techniques and Applications

Image processing (low-level analysis) Transformations: filtering, feature extraction, enhancement, compression1

Maximum entropy method (MEM): de-convolution, super resolution, reconstruction2

Projection onto convex sets (POCS): reconstruction, de-convolution, filter design3

Fractals: compression, object matching4,5

Feature extraction (intermediate-level analysis) Thresholding: object extraction from background6

Edge detection: boundary detection1

Thinning: skeletonization1,7

Morphological operations: noise removal, object extraction1

Snakes: boundary detection, object tracking8

Self-organizing maps (SOMs): segmentation, clustering9

Fuzzy c-means algorithms (FCMs): clustering9

Morphing: animation through shape deformation10

Object recognition and matching (high-level analysis) Bayes theory: classification11

Neural net classifiers: object recognition, segmentation11

Fuzzy classifiers: object recognition, rule-based systems9

Hidden Markov models (HMMs): speech and handwriting recognition12

Graph matching: structural matching1

Hough transform: known shape detection1

Shape from shading: finding 3D shapes from 2D imagesRelaxation labeling: object matching13

1 Example of image enhancement:(a) the original color image and (b) the enhanced image. Fine fea-tures of the trees and flowersbecome more visible after theenhancement.

2 An exampleof magneticresonanceimage (MRI)reconstructedfrom data cor-rupted bymotion. Left:the image withso-called ghostartifacts frommotion. Right:image restoredusing the pro-jection ontoconvex sets(POCS) method.

(a) (b)

Image and video compressionResearchers have extensively studied the compres-

sion of sound, image, and video data in signal process-ing. JPEG and MPEG standards recommend the discretecosine transform (DCT) and wavelet transform forimage and video compression. These methods divide animage into small blocks and transform each block toanother domain to reduce spatial correlation, thenquantize the transform coefficients to minimize theamount of data required for image reconstruction.1

When the compression ratio is high, boundaries

between the blocks become visible.Researchers have investigated sev-eral methods to solve this problem.Figure 3 shows the result of block-ing artifact reduction using awavelet-transform-based method.

The DCT and wavelet transformsdon’t consider high-level image con-tents, such as the shapes of imageobjects, so they provide only mod-erate compression. Fractal- andsegmentation-based compressionmethods overcome this problem.Fractal methods derive from the col-lage theorem, which states that animage can be reconstructed usingthe iterated function systems (IFS),which map one area to anotherstarting from a random initializa-tion.4,5 Unfortunately, arbitraryshapes cannot be matched to gener-ate the IFS codes automatically.Instead, areas of regular shapes,such as squares and triangles, are

used for matching in an image.4 Figure 4 shows a com-pressed color image using the IFS of square areas. Geo-metric mappings, such as nonlinear transforms, improvethe fractal-based method’s compression efficiency.5

Segmentation-based image compression (or second-generation) methods provide a very high compressionratio by dividing an image into small homogeneousregions that can be compressed efficiently.16 However,applying this method to a wide class of natural imagesrequires more research. MPEG-4 introduces image seg-mentation to separate the objects and the scene in animage, which allows user interactions with differentvideo data components.17

MPEG-4 also accepts synthetic images. For example,to show a face in a video, we need only face definitionparameters (FDPs) to describe the face shape and faceanimation parameters (FAPs) to describe facial expres-sions. Because the data needed are the parameters only,not the actual image frame sequence, we can minimizethe bit rate. One way to generate personalized FDPs is toanalyze the person’s face in photographs.18 This requireskey feature point extraction and matching as well as tex-ture mapping. Making the procedure automatic is achallenge.

Media creation and data retrievalTwo relatively new applications of image analysis are

computer animation and media data retrieval. In bothcases, unsolved problems are accurate image-contentanalysis and pattern matching, recognition, andmodeling.

Computer animationImage analysis is increasingly useful for computer-

generated scene images, character animation, and vir-tual reality applications. The analysis of hand drawings,photographs, video, and motion data is useful in digitalmedia applications for two reasons: to let the computer

Tutorial

20 January/February 2001

4 Example ofimage compres-sion usingfractals: (a) theoriginal imageand (b) thecompressedimage. Thecompressionratio is 20.07and the peaksignal-to-noiseratio (PSNR) is31.21 dB.

3 Blocking artifact removal fromhighly compressed images. (a) Acompressed color image using theblock discrete cosine transform(BDCT). It has a compression ratioof 77 and peak signal-to-noise ratio(PSNR) of 23.98 dB. (b) Result ofblocking artifact removal usingwavelet transforms. This image hasa PSNR of 25.43 dB.

(a)

(b)

(a)

(b)

perform labor-intensive tasks and todevelop graphics models of naturalobjects based on the featuresextracted from the images of realobjects.

2D animation. In cartoonmovie production, animators drawkey frames and then draw many in-betweens—frames that interpolatethe shapes in the key frames to showmoving objects, a tedious and time-consuming process. The computercan generate in-betweens automat-ically or semiautomatically by ana-lyzing key frames.

Figure 5 shows a cartoon imageprocessing system we recentlydeveloped. Two image analysis tasksare involved. Vectorization, ex-plained later, converts scannedbitmap images to lines, which canthen be edited, stored, and re-used easily. The secondtask implements pattern matching, which identifies andmatches the same parts of an object in different frames,then interpolates them to generate the in-betweens. Ina method we have developed recently, relaxation label-ing and graph matching algorithms solve this problem.

Figure 6 shows an example of in-between frames gen-erated from two key frames. The matching procedurecan make mistakes if one frame is substantially differ-ent from the other. For example, an open mouth canhave many more features than a closed mouth. Anima-tors solve these problems based on rules, library pat-terns, and knowledge. Programming a computer toperform the same operations will be a challenge.

3D animation. Computer animation’s most diffi-cult and exciting task is probably 3D shape and motionmodeling of humans.19,20 Current computer modelingtechniques always result in a stiff human image, evenwhen the person is smiling. It’s long been the objectiveof computer animation research to make the humanfacial expressions look natural. One solution is to modelthe human head’s physical structureand muscle movement, as Figure 7shows. Another promising methodto obtain more natural-lookingfacial appearances is to identifyclues from a real object’s image orvideo and use the features in graph-ics models.

Image analysis techniques aid inprocessing data from motion-cap-turing devices, 3D scanners, andvideo images. For example, one wayto create human face models is byanalyzing two orthogonal head pic-tures of a person, then matchingcorresponding key points and map-ping them to a generic mesh.18 Wecan make the head model more

IEEE Computer Graphics and Applications 21

Hand-drawnkey frames

In-betweenframes

Rasterimages

VectorgraphicsMatchingMatched

pairsInterpolation

Vectorization

EditingCorrection Rules

Digitization(scanning)

5 A cartoon image-processing system. Image analysis techniques solve hand-drawing vector-ization and pattern-matching problems.

6 Key frames(dark lines) andin-betweens(gray lines)generated usingour method.The matchingbetween thetwo key framesis done auto-matically.

7 Facial expres-sions generatedusing anabstract musclemodel.

realistic by mapping texture images to the mesh. Wecan also build a 3D face model with data from laserscanners.

In all these methods, to automate the process, wemust use sophisticated feature extraction algorithms tolocate and match the positions of nose, eyes, and ears—for example, in the images.21 Figure 8 shows 3D headimages reconstructed from 3D laser-scanned range andreflectance data.

Multimedia data retrievalBecause of the rapidly swelling

volume of digital media data, theindexing and retrieval of databaseitems assumes greater impor-tance.22 Traditionally, databasequeries are textual, which is some-what useful for image retrieval. Forexample, to buy a coffeemaker witha specific color, shape, and size, wecan search through a database ofcoffeemakers using a particular setof parameters. One problem is howto specify in words the exact shapeand size, and match the user’srequirement with the shapes in the

database. If we are given only the coffeemaker-compo-nent images, we must analyze these images and extractthe shapes to do the matching.

For image retrieval, we can also submit an objectimage or sketch as a database query to find similarobjects. Such retrievals require us to match the inputimage or drawing with the database images. In somecases, we can match them based on statistical informa-tion, such as color histograms and image transform coef-ficients. However, when structural image informationis required, we must analyze the image contents anddeal with complicated pattern classification and match-ing problems.9,11,23

Document processing and characterrecognition

Printed and handwritten materials, such as newspa-pers, books, and office documents, have historicallyserved as one of the most useful forms of media infor-mation exchange. The aim of document imaging sys-tems is to convert paper documents to an electronicformat for more efficient document storage, indexing,and distribution. Although many new documents cannow be directly generated as computer files, we mustdigitize and analyze a huge number of existing paperdocuments, in addition to new ones being producedevery day.

Document imagingAlthough it’s a simple procedure to scan a document,

it can be complex to analyze its contents, as we explainhere.

Image binarization. Most document images arestored in binary (black-and-white) format, which weobtain by thresholding a gray-scale image.6 If the gray-scale image has a high contrast, a global threshold is suf-ficient to separate the image background fromforeground. However, an image containing a varying,textured background requires more sophisticated bina-rization, in which case we must determine the thresholdvalue locally, even at each individual pixel.6

Layout analysis. Layout analysis, which separatesimage text, diagrams, and pictures, must precede doc-ument content recognition and analysis because text

Tutorial

22 January/February 2001

8 Examples of3D head modelscreated basedon laser-scannedrange andreflectance data.

9 Documentlayout analysis:(a) the originaldocumentimage, (b) images anddiagramsextracted fromthe documentimage, and (c) text in differ-ent orientationsextracted fromthe documentimage.

(a)

(b)

(c)

and graphics require different analysis techniques.12,24

If text lines are horizontally oriented, and the text anddiagram blocks are rectangular, we can use pixel pro-jection histograms to extract text and diagram areas.This method works only if the document can bedeskewed first.25A more complicated document struc-ture requires additional analysis features, such as pixelrun lengths, connected components, thinned whitespace, texture, and k-nearest neighbors.

The layout analysis problem is essentially classifyingeach pixel or region. Figure 9 shows segmentationresults of a document that contains pictures, diagrams,and text in different orientations.26 Once these compo-nents are separated, they can be passed to the next doc-ument processing system for compression orrecognition.

Layout analysis also aids document retrieval. Forexample, even one scanned newspaper page can holdtoo much data to download from the Internet. If devel-opers can analyze the page structure, extract each arti-cle, and index the page using the article titles, users canbrowse the page and download articles only as needed.The solution to this complicated problem requires lay-out analysis combined with character recognition.

Character recognition. A well-defined techniquewith many applications, optical character recognitionhas long been one of the most actively researched top-ics in pattern recognition.27 A related problem is signa-ture verification,28 in which we must identify the writerrather than the input sample contents. Figure 10 showsthe components of a general OCR system. The prepro-cessing procedure usually isolates and normalizes thecharacter image and removes noise. Then the featureextraction procedure computes a set of features x = [x1, x2, ..., xN] from the bitmap of the input character andpasses them to the classifier. The classification resultsare modified according to additional information suchas a word’s spelling. The classifier is considered a func-tion f(x), which is produced from training samples. Thefeatures x represent statistical properties of the inputcharacter, such as character pixel distribution, and char-acter structure, such as stroke positions, directions, andtheir relations.

Researchers have studied many methods for classifi-er design, including nearest-neighbor classifiers, Bayesclassifiers, neural nets, fuzzy classifiers, Markov mod-els and classification trees, and combinations of these.9,11

The training and classification times can differ signifi-cantly for different classifiers, but if designed correctly,classifiers can have a similar recognition performancefor a given set of features because the classification erroris bounded by the Bayes error.

Computers are humans’ equal at recognizing isolat-ed and clean-printed characters, but are much less ablethan humans in recognizing variable, degraded input.Research problems include recognition of connected orbroken characters, of free handwriting or drawing, andof different classifier combinations.

Line vectorization. Usually called thinning, linevectorization is useful for extracting skeletons of char-

acters and for analyzing mechanical drawings, maps,and cartoon images.7 In traditional thinning algorithms,an “onion peeling” procedure removes redundant pix-els in a line and makes it only one pixel thick. Thismethod often produces artifacts across line cross sec-tions and is sensitive to noise. In more sophisticatedmethods, we analyze the line boundary and the cross-section structure to generate a more accurate represen-tation.7 Heavy noise is an unsolved problem; a relatedproblem is how to cluster scattered data that have a lineshape. Clustering algorithms can be used to extract cir-cles, ellipses, and straight lines, but more flexible tech-niques are needed to deal with more general shapes.

Document image compression. Widely used forcompressing binary images, the CCITT Group 4 methodwas developed based on run-length coding to minimizeredundancy within a line and correlation to the previousline. This method provides lossless compression. A newstandard, JBIG2, proposed by the Joint Bi-level ImageExperts Group, is emerging that provides compressionbased on the symbol-matching concept.29 For example,a document may contain the letter “A” in many places.To place the letter correctly in the document recon-structed after compression, we need only to store onebitmap of “A” and code its positions. The originalbitmaps of the letter may be somewhat different in dif-ferent places, but they can all be replaced by the samerepresentative bitmap without changing its meaning.30

As long as all symbols are matched correctly, the com-pression can be made content-lossless, although it’slossy at the pixel level.

Figure 11 (next page) shows a document image andthe associated symbols needed to represent the printedand handwritten text in the document. A symbol-match-ing-based compression method provides a compressionratio several times higher than the CCITT Group 4

IEEE Computer Graphics and Applications 23

Input

Output

Preprocessing

Featureextraction

Classification

Postprocessing

x = [x1, x2, …, xN]

Structural and statistical features

y = f(x)98 as text output

Dictionary, rules, contextCorrection of mistakes

Output is “98 or even“year 1998 according to context

““

10 Flow dia-gram of a pat-tern recognitionsystem. Theexample showsconnectedhandwrittencharacter recog-nition.

method and is useful for developing new fax machine,document scanner, and other document-imaging sys-tem applications. Current research problems includeextracting and matching the symbols in an image effi-ciently, quickly, and accurately.

Electronic booksElectronic books have many advantages over paper

books. For example, e-books can present multimediadata and can be interactive. Also, they are environmen-tally friendly; require much less storage; and are easilyindexed, searched, and distributed.

An e-book usually consists of two parts: a data file thatcontains the book’s contents and a reader that makes the

presentation. The reader can usual-ly be used to read many differentbooks. Some e-books can be read ona general-purpose computer inwhich the reader is just a computerprogram. There are also specializedhardware e-book readers, rangingfrom small palm devices to specialmultiscreen notebook computers.Some readers allow user interactionsand written and voice annotations.

Our work on e-books for educa-tional purposes is inspired by theobservation that many mathemati-cal operations are easily explainedwith animation. The author’s use ofcomputer movies in undergraduateteaching has been very well receivedby students. For example, animationmovies effectively show signal con-volution, as Figure 12 shows, and thedecomposition of a rotation aroundthe axis in an arbitrary direction (seehttp://www.HyperAcademy.com).

In terms of image analysis,research is needed in two mainareas. One is to improve the readerdisplay, including fonts, layout, anduser interface. Ordinary computerscreens still have much lower reso-lution than laser printers, and thetouch and feel of e-books are differ-ent from the paper books people areused to. The reading of e-booksmust be made more comfortable.

Research is also needed to createmore intelligent, easy-to-use au-thoring tools, especially for educa-tional material. For e-books tobecome popular, we need manyauthors to write them. Simple stat-ic material like scanned pages willattract few users. However, it’s dif-ficult and time-consuming to createe-books of interactive mathemati-cal and technical material, whichdiscourages authors. One solutionwould be authoring software that

accepts handwriting. It should let authors write anequation and draw a diagram, and then have a com-puter recognize and change them to font symbols andwell-formatted diagrams.

The recognition need not be perfect. For example, itcould provide the user with several choices or templates.It should also allow the user to create animation to relateequations and variables to diagrams and show thechange of equations so that a concept or procedure canbe described clearly. Essentially, the authoring tool mustbe simple and intuitive to use, and be able to handle allthe user’s work.

Tutorial

24 January/February 2001

11 Documentimage compres-sion: (a) theoriginal docu-ment and (b)extracted sym-bols to repre-sent printedand handwrit-ten text in thedocument.

(a)

(b)

DiscussionImage filtering, enhancement, transformation, com-

pression, and restoration algorithms are relatively wellunderstood and well developed. These techniques canhave many useful applications in Internet and mobilecommunications systems. For example, vectorized car-toon images are well suited for mobile displays. Hand-writing analysis and recognition in e-books can be usedfor processing handwritten notes and drawings. In thenext few years, we shall be able to use more and moreInternet- and mobile-based imaging and video servicesin education, business, and entertainment. Futureresearch needed in these areas includes faster hardwareimplementation, for example, for image and video cod-ing; efficient and secure data transmission throughInternet and wireless systems; and development ofrobust and stable algorithms for image restoration.

The success of new industrial products in digitalmedia will also depend on user acceptance. For exam-ple, e-books can potentially have a very large market,but it is not known whether the users can switch frompaper books to e-books. As discussed above, authors’acceptance is also crucial. At the same time, the studyof user requirements might also lead to many interest-ing, practical research problems.

Feature extraction and object recognition are key butlargely unsolved image analysis problems. Current com-puter algorithms are far less reliable than humans. Theproblems include these: Computer models are too sen-sitive to noise and image distortions; learning algo-rithms are too slow and require too many trainingsamples; and low- and high-level feature extraction andobject classification procedures are mainly separateprocesses. Although researchers have tried to solve theproblems based on structural and statistical methods,symbolic and numerical formulations, and on variousclues from physiological findings of the brain, the recog-nition problem remains challenging.

Despite the difficulties in searching for a theoreticalsolution to the object recognition problem, it’s still pos-sible to build useful systems for practical applications.We can improve system performance by putting morerestrictive conditions on input images, having well-defined objectives, using many expert rules, and com-bining different classifiers. All these strategies can workwell in a handwriting recognition system. For example,we can achieve a high recognition rate for well-isolatedcharacters by integrating several classifiers.9

In research, useful ideas can be learned from a dif-ferent field. For example, the hidden Markov model(HMM), which has been used successfully for speechrecognition, can also be useful to solve image recogni-tion problems. Image analysis and computer graphicshave till now been studied mainly in separate fields. Themerging of techniques from other fields can lead to newalgorithms and applications. ■

AcknowledgmentOur work on cartoon image analysis, handwriting

recognition, and document image compression is sup-ported by several grants from the Australian ResearchCouncil. Our work on Electronic Books is supported by

an information technology grant from the University ofSydney and a teaching development grant from the CityUniversity of Hong Kong. Many research staff and post-graduate students of the Image Processing Lab at the Uni-versity of Sydney and the Signal Processing Lab at theCity University of Hong Kong have worked on these pro-jects and provided the images presented in this article.

References1. R.C. Gonzalez and R.E. Woods, Digital Image Processing,

Addison-Wesley, Reading, Mass., 1992.2. Maximum Entropy and Bayesian Methods, J. Skilling, ed.,

Kluwer Academic, Dordrecht, the Netherlands, 1989.3. H. Stark and Y. Yang, Vector Space Projections, John Wiley

and Sons, New York, 1998.4. A.E. Jacquin, “Image Coding Based on a Fractal Theory of

Iterated Contractive Image Transformations,” IEEE Trans.Image Processing, vol. 1, no. 1, Jan. 1992, pp. 18-30.

5. D. Popescu and H. Yan, “A Non-Linear Model for FractalImage Coding,” IEEE Trans. Image Processing, vol. 6, no. 3,1997, pp. 373-382.

6. Y. Yang and H. Yan, “An Adaptive Logical Method for Bina-rization of Degraded Document Images,” Pattern Recogni-

IEEE Computer Graphics and Applications 25

12 Two framesof an animationmovie show theoperations ofsignal convolu-tion. View thismovie athttp://www.HyperAcademy.com orhttp://www.hy8.com.

f(x)

x1

1

g(x)

x1

1

f(x)

x1

1

g(x)

x1−1

1

f(a)g(1.5 − a)

x1 1.50.5−1

f(a)g(1.5 − a)

0.5

f(x)*g(x)

f g f a g a da( . ) * ( . ) ( ) ( . ) .1 5 1 5 1 5 0 125= − =−∞

∞

∫

1 2x

(a)

(b)

tion, vol. 33, no. 5, 2000, pp. 787-807.7. H.H. Chang and H. Yan, “Analysis of Stroke Structure of

Handwritten Chinese Characters,” IEEE Trans. Systems,Man and Cybernetics, vol. 29, no. 1, 1999, pp. 47-61.

8. Y. Zhu and H. Yan, “Computerized Tumor Boundary Detec-tion Using a Hopfield Network,” IEEE Trans. Medical Imag-ing, vol. 16, no. 1, 1997, pp. 55-67.

9. Z. Chi, H. Yan, and T.D. Pham, Fuzzy Algorithms, with Appli-cations to Image Processing and Pattern Recognition, WorldScientific Publishing, Singapore, 1996.

10. J. Gomes et al., Warping and Morphing of Graphics Objects,Morgan Kaufmann, San Francisco, 1999.

11. J. Schurmann, Pattern Classification, A Unified View of Sta-tistical and Neural Approaches, John Wiley and Sons, NewYork, 1996.

12. Document Image Analysis, L. O’Gorman and R. Kasturi, eds.,IEEE Computer Soc. Press, Los Alamitos, Calif., 1995.

13. R. Klette, K. Schluns, and A. Koschan, Computer Vision,Three Dimensional Data from Images, Springer-Verlag, Sin-gapore, 1998.

14. N. Liu and H. Yan, “Improved Method for Color ImageEnhancement Based on Luminance and Color Contrast,”J. Electronic Imaging, vol. 3, no. 2, 1994, pp. 190-197.

15. M. Hedley, H. Yan, and D. Rosenfeld, “Motion Artifact Cor-rection in MRI using Generalized Projections,” IEEE Trans.Medical Imaging, vol. 10, no. 1, 1991, pp. 40-46.

16. T. Ebrahimi and M. Kunt, “Visual Data Compression forMultimedia Applications,” Proc. IEEE, vol. 86, no. 6, June1998, pp. 1109-1125.

17. G.A. Abrantes and F. Pereira, “MPEG-4 Facial AnimationTechnology: Survey, Implementation, and Results,” IEEETrans. Circuits and Systems for Video Tech., vol. 9, no. 2, Mar.1999, pp. 290-305.

18. W.-S. Lee and N. Magnenat-Thalmann, “Fast Head Model-ing for Animation,” Image and Vision Computing, vol. 18,2000, pp. 355-364.

19. F.I. Parke and K. Waters, Computer Facial Animation, A.K.Peters, Wellesley, Mass., 1996.

20. S.K. Karunaratne and H. Yan, “An Emotional Viseme Com-piler for Facial Animation,” Proc. Fifth Int’l Symp. SignalProcessing and Its Applications, Queensland University ofTechnology, Brisbane, Australia, Aug. 1999, pp. 459-462.

21. K.M. Lam and H. Yan, “An Analytic-to-Holistic Approachto Face Recognition Based on a Single Frontal View,” IEEETrans. Pattern Analysis and Machine Intelligence, vol. 20,no. 7, 1998, pp. 673-686.

22. Y. Rui et al., “Relevance Feedback: A Power Tool for Inter-active Content-based Image Retrieval,” IEEE Trans. Circuitsand Systems for Video Tech., vol. 8, no. 5, Sept. 1998, pp.644-655.

23. T. Tan and H. Yan, “Face Recognition by Fractal Transfor-mations,” Proc. IEEE Int’l Conf. Acoustics, Speech, and SignalProcessing, IEEE Press, Piscataway, N.J., Mar. 1999, pp.3537-3540.

24. A.K. Jain and B. Yu, “Document Representation and ItsApplication to Page Decomposition,” IEEE Trans. PatternAnalysis Machine Intelligence, vol. 20, no. 3, Mar. 1998, pp.294-308.

25. H. Yan, “Skew Correction of Document Images Using Inter-line Cross-Correlation,” CVGIP: Graphical Models and ImageProcessing, Vol. 55, no. 6, 1993, pp. 538-543.

26. P.E. Mitchell and H. Yan, “Document Page SegmentationBased on Pattern Spread Analysis,” Optical Eng., vol. 39,no. 3, 2000, pp. 724-734.

27. J. Hu and H. Yan, “A Model-Based Segmentation Methodfor Handwritten Numeral Strings,” Computer Vision andImage Understanding, vol. 70, no. 3, 1998, pp. 383-403.

28. K. Huang and H. Yan, “Off-Line Signature VerificationBased on Geometric Feature Extraction and Neural Net-work Classification,” Pattern Recognition, vol. 30, no. 1,1997, pp. 9-17.

29. P.G. Howard et al., “The Emerging JBIG2 Standard,” IEEETrans. Circuits and Systems for Video Tech., vol. 8, no. 7, Nov.1998, pp. 838-848.

30. Y. Yang, H. Yan, and D. Yu, “Content Lossless Compressionof Document Images Based on Structural Analysis and Pat-tern Matching,” Pattern Recognition, vol. 33, no. 8, 2000,pp. 1277-1293.

Hong Yan is a professor of comput-er engineering in the ElectronicEngineering Dept. at the City Uni-versity of Hong Kong. His researchinterests include computer anima-tion, signal and image processing,pattern recognition, and neural and

fuzzy algorithms. Yan received a BE from Nanking Insti-tute of Posts and Telecommunications in 1982, an MSEfrom the University of Michigan in 1984, and a PhD fromYale University in 1989, all in electrical engineering. He isa fellow of the International Association for Pattern Recog-nition; a fellow of the Institution of Engineers, Australia;a senior member of the IEEE; and a member of the Inter-national Society for Optical Engineers, the InternationalNeural Network Society, and the International Society forMagnetic Resonance in Medicine.

Yan is currently on leave from the School of Electricaland Information Engineering, University of Sydney, NSW2006, Australia.

Contact him at Dept. of Electronic Eng., City Univ. ofHong Kong, 83 Tat Chee Ave., Kowloon, Hong Kong;h.yan@cityu.edu.hk.

Tutorial

26 January/February 2001