Return to Home Page

This book is a collection of 31 scientific papers organized in four main sections:  “Pattern recognition and Machine Intelligence”, “Computer Vision and Image Processing”, “Face Recognition and Forensics” and “Biometrics Authentication”. These chapters cover a broad domain making the book appealing to a large group of specialists. The difficulty ranges from beginner to advanced through the chapters of the book. There are simple introductory overview chapters but also chapters with solid theoretical support and extensive experimental results.

In the first part, to get a basic introduction to evolutionary algorithms, we recommend the first chapter. Chapter 4 is a good paper for anyone interested in image fusion. Chapter 5 is an interesting application of pattern recognition to the wireless sensor selection problem.

In the second part, the reader (recommended for students) could find a fascinating article on human extremity detection in Chapter 10, an introduction to ensemble learning in Chapter 11, and an image rendering system based on depth information in Chapter 12.

The third part “Face recognition and Forensics” starts with an intriguing experiment on gender and race identification, using 3D data. You have to read also the comparison of computer versus human classification.  If you are interested in face recognition you should certainly not miss Chapter 14. It is a well written article covering existing methods. There is a good short overview of existing methods, a theoretical proof of their method and impressive experimental results. Chapter 18 has another very good paper on face recognition. This is a comprehensive overview with good theoretical background and experimental comparisons.  For those interested in fingerprints, Chapter 17 gives an extensive overview of state-of-the-art methods for fingerprint identification. Chapter 22 presents a method for detection of image forgery and hidden content in JPEG images and wave files.

The fourth part has nine chapters on aspects of biometric identification. An overview of biometric authentication systems is given in Chapters 23 and 28. Chapter 24 presents a method for online handwritten Chinese characters recognition.  Automatic signature verification is discussed in Chapters 26, 30 and 31. A good overview on iris recognition is given in Chapter 28.

As a collection of independent articles written by various authors,  there is some repetition of topics in the book. The disadvantage of this is that many introductions and short reviews refer to the same field. On the other hand, this is also an advantage:  one can see different views on the same algorithms and different interpretations of similar results.

A future edition of the book could consider removing some of the chapters (i.e., 7, 8, 23, and 29), correcting the English errors, and printing result graphs in color to improve readability.

We give a more detailed review per chapter below:

Part I “Pattern recognition and Machine Intelligence”

1. A Review of Applications of Evolutionary Algorithms in Pattern Recognition

This chapter is a review of evolutionary algorithms applied to pattern recognition. It gives a good introduction to evolutionary algorithms. Firstly, evolutionary algorithms are exemplified on k-means clustering and ellipse fitting problems. Then the use of EA is exemplified in all three aspects of pattern recognition: segmentation, feature selection, and classification.

This chapter gives the reader a good insight into evolutionary algorithms, advocating for their advantages in pattern recognition problems. For more detailed information, the reader is presented with an extensive list of references.

2. Pattern Discovery and Recognition in Sequences

This chapter begins by presenting  a brief review of the challenging task of pattern discovery with applications from natural language processing to DNA analysis. Then the authors introduce their pattern discovery framework.

3. A Hybrid Method of Tone Assessment  for  Mandarin CALL System

This chapter describes a new method for spoken language tone assessments.  Since language tone is more important in Mandarin than it is in European languages, their system implements tone recognition for Mandarin.

4. Fusion with Infrared Images for an Improved Performance and Perception

This chapter reviews existing methods for image fusion at three different levels: pixel, feature, and decision. The review is specifically oriented for fusion of infrared-images with visible light images. Some applications of image fusion are presented at the end.

5. Feature selection and ranking for pattern classification in wireless sensor networks

In this chapter feature selection is applied to a specific problem. In order to reduce the power consumption and the lifetime of a network of wireless sensors, the number of sensors should be reduced.  Here they use a k-nearest neighbor classifier to select the features (sensors). Their proposed system was implemented and tested for accuracy and lifetime in a set of five experiments.

6. Principles and applications of RIDED-2D – a robust edge detection method in range images

This chapter presents a computationally efficient edge detection method from range images. The range line image is filtered first for noise, and then edges are detected with a rule based algorithm.  Integrating the denoising and edge detection in one algorithm, they obtain an efficient system able to process one scan line in 0.1 milliseconds with over 98% accuracy.

PART II. Computer Vision and Image Processing

7. Lens Shading Correction for Dirt Detection

This article presents a method for automatic lens dirt detection on inexpensive camera modules.

8. Using Prototype-Based Classification for Automatic Knowledge Acquisition

This is a “trial and error” method to group some data. There is no learning or generalization power. They apply this to mitochondrial movement classification.

9. Tracking Deformable Objects With Evolving Templates for Real-Time Machine Vision

The paper addresses the most important issues of object tracking: real-time performance, matching, template updating, occlusion, and multiple objects tracking by using a contour-based tracking approach. Their A* search method is about 4 times faster than blind search template matching while providing the same accuracy.

10. Human Extremity Detection for Action Recognition

This chapter proposes an approximation for human action recognition. Instead of tracking all internal body joints, they show that positions of body extremities alone are enough for an excellent approximation of body motion. Since the inner joints are more difficult to automatically detect from video, without special markers, this method is an important step forward towards automatic human action recognition. They use either contour or image patches to find extremities.  On the two public datasets:  Weizmann and Tower, they get 93.6% and 86.7% accuracy, respectively.  Tower data set is more difficult since the humans have a strong shadow and image resolution is quite low. Compared to the precise extremities method, their probable extremities method gets better results: 95.7% on Weizmann and 98.3% on Tower data.

11. Ensemble Learning for Object Recognition and Tracking

This chapter is a review of complex classification methods consisting of a set of simple classifiers. This decomposition is called ensemble learning. Different classifiers can be generated by different initializations, different training sets, or different classifiers trained over same feature set and training data. Two main classes of ensemble learning are discussed here, the random subspace method and the boosting method. They are exemplified on face recognition and object tracking, respectively.

12. Depth Image Based Rendering

This chapter presents a method for automatic image rendering based on depth information from video images.  The human brain generates depth information from images captured by two eyes. This chapter introduces a method that generates the second image feed from a 2D video and depth information. Rendering issues are tackled here: disocclusion, imperfect depth maps, ghosting, and cardboard effect.  They propose a new standard for 3D video representation: “video -plus-depth” (instead of current system “video-plus-video”) with advantages of a smaller data rate and the possibility to represent more views with adaptable rendering parameters.

Part III: Face Recognition and Forensics

13. Gender and Race Identification by Man and Machine

This chapter presents an interesting experiment on gender and race identification. They use an existing 3D face dataset with accompanied ground-truth and standard classification methods. Their goal is to prove experimentally that profile contour and the color information makes a difference in gender and race identification. They also compare computer results with classifications made by a group of human subjects.  They use shape matching for profile contour and SVM for frontal face classification. They conclude that the profile contour of a face can be used as a cue for gender and race identification, although the classification results are poor. Color provides very little information and therefore it is not recommended to be used for gender and race identification. Surprisingly, humans could not identify the gender and race without error, and furthermore the computer outperforms humans on this classification.

14. Common Vector Based Face Recognition Algorithm

This chapter starts with a well-written, short review of existing methods for face recognition. Then theoretical proof of their new method for face recognition based on common vector is presented.  They show impressive experimental results compared to standard methods of PCA, KPCA and LDA. For instance, on the FERET database, their recognition rate is 71.9% versus 42.3% obtained by PCA.

15. A Look at Eye Detection for Unconstrained Environments

Here are presented two methods for precise eye detection in unconstrained images. One is based on PCA and learning; the other one is based on filter correlation using Fourier transform. They compare their methods to a leading commercial application on images captured in low light conditions, blurred images, and low resolution images. In all these cases, their method outperforms the commercial application.

16. Kernel Methods for Facial Image Preprocessing

In this chapter are shown impressive results of kernel PCA applied to preprocessing of face images. In particular, outstanding results are shown on image denoising, occlusion recovery, illumination normalization, and facial expression normalization.  Kernel methods are normally used in classification, this chapter shows their strength in the less tackled domain of image preprocessing.

17. Fingerprint Identification – Ideas, Influences, and Trends of New Age

Here is an extensive review of state-of-the-art methods for fingerprint identification. Almost 200 papers on fingerprints are reviewed here. They track the first use of fingerprints back to 1684. Then major fingerprint publications from the last 20 years are synthesized in a good fingerprint identification presentation.  Several methods from the literature of preprocessing, minutiae extraction, and texture-based feature extraction are described here briefly. At the end of this review the authors identify a few open problems in fingerprint identification: better accuracy, identification in low quality images, detection from overlapped finger print images, and identification speed in large databases. 

18. Subspaces Versus Submanifolds – A Comparative Study of Face Recognition

This is a theoretical and experimental comparison of subspaces versus submanifolds methods used in face recognition.  First, a brief review of subspace based face recognition algorithms is given: PCA and LDA. Then, submanifold based algorithms for face recognition are discussed.  Four main groups of manifold learning algorithms are identified: global methods, global alignment of local models, local methods, and extensions. Out of each group, several representative methods are presented briefly. At the end, experiments are performed on the CMU-PIE and FERET face databases. Very valuable analysis of results and conclusions make this chapter one of our favorites.

19. Linear and Nonlinear Feature Extraction for Face Recognition

This chapter discusses the limitations of linear methods for face recognition: dimensionality problem, small sample size problem, and nonlinear problem. After a short presentation of classical methods (PCA,LDA, Fisherface, and direct LDA), they introduce the theoretical base of a new method for regularized LDA called here 2SRDA. They test 2SRDA against Fisherface, DLDA, and Huang’s method on the ORL and FERET databases. 2SRDA has better recognition rates than other methods and it also has a smaller computational complexity.

Since variations of pose and illumination are too difficult for linear classifiers, they present further a kernel method based on Mercer kernel theory. Experimental results show its robustness on pose and illumination variations on the FERET and CMU-PIE datasets.

20. Facial Occlusion Reconstruction Using Direct Combined Model

Here is presented a new method for reconstructing the partially occluded face image.  After a brief review of existing methods, the authors introduce their Direct Combined Model (DCM), which basically combines the shape and texture into one Eigen space in order to maximize their covariance. Then a DCM transform is derived from the learned combined model. Their experiments on a face dataset manually labeled and manually occluded shows similar results to other existing methods, while DCM is more robust and can handle larger occlusion areas.

21. Generative Models and Probability Evaluation for Forensic Evidence

This chapter describes a generative model with applications in forensics. Defined here are three metrics based on the probability of random correspondence. These three metrics are applied to several modalities:  birthdays, human heights, and fingerprints using ridge flow and minutiae.  The fingerprint minutiae are modeled as mixtures of Gaussian and von Mises distributions. For evaluation they use the NIST special database 4. Their estimated probability is very close to the empirical results, concluding that generative model offers a reasonable and accurate fingerprint representation.

22. Feature Mining and Pattern Recognition in Multimedia Forensics – Detection of JPEG Image Based Steganography, Double-Compression, Interpolation and WAV Audio Based Steganography

After a short introduction and review of existing methods, this chapter presents a method for detection of steganography based on Markov analysis of the DCT coefficients.  Their study shows that embedding of the secret information changes the neighboring joint distribution. They also show that Markov analysis alone is not enough to detect steganography properly.

For image forgery, they analyze the statistical properties of DCT coefficients to detect double compression and interpolation.

For steganalysis of audio files, they present a new derivative-based method exploiting the Mel-cepstrum coefficients and  Markov transition features.

Experimental results on a set of 5000 images shows good results on detection of hidden data, double compression, and jpeg resampling. The audio steganalysis was tested on a set of over 25,000 audio sequences. Results indicate significant improvements over the previous methods.

PART IV BIOMETRIC AUTHENTICATION

23. Biometric Authentication

As an overview of biometrics methods, this is a good beginning of Part IV. Here are presented basic principles and operations of fingerprints, face, iris, voiceprint, and vein recognition systems. Here are also discussed issues of biometric standardization and certification.

24. Radical-Based Hybrid Statistical-Structural Approach for Online Handwritten Chinese Character Recognition

Here is presented a framework for handwritten Chinese character recognition. Many Chinese characters share common substructures called radicals. This method combines the statistical and structural approaches for radical-based character recognition. There are two main classes of radicals: special and non-special. Binary classifiers, one for each class of special radical, are used for detection of special radicals. For non-special radicals (left-right, up-down, and single-element) detection they use two methods: sequential and hierarchical.  The experimental results show that the independent modules of special and non-special radical recognition perform comparably to holistic character recognition. But the integrated radical-based, whole-character recognition is slightly lower than that of holistic recognition.

25. Current Trends in Multimodal Biometric System – Rank Level Fusion

This chapter is an overview of existing multimodal biometric systems and the current trends in multimodal biometric fusion. Multimodal biometric systems have the advantage of improved recognition accuracy over unimodal systems. One important issue in these systems is how to combine the extra information of individual systems, called fusion. There are two major classes of fusion methods:  pre and post matching.

26. Off-line Signature Verification by Matching with 3D Reference Knowledge Image - From Research to Actual Application

This is one of the best chapters in this book. The authors review existing methods and introduce a method for off-line verification of signatures. In signature verification there are three types of images used: binary, thinned, and high pressure region images. From these images, several features are extracted to be used in classification: slopes, slants, horizontal and vertical projections, etc.  3D Reference Knowledge Image (3DRKI), which adds another dimension to the feature space, is introduced. This is simply built by superimposing the binary genuine samples centered on their gravity center. The third dimension is the number of occurrences of pixels from different samples. Global and local feature extraction is presented in details together with distance metrics. Experiments performed using 3DRKI with three image types show outstanding results compared to other methods. We also  find their comments on the transition from research to a real commercial application interesting.

27. Unified Entropy Theory and Maximum Discrimination on Pattern Recognition

This chapter presents the unified entropy theory on pattern recognition.  Based on this theory, the authors prove theoretically and practically, with experiments on Chinese character recognition, that mutual information is directly proportional to accuracy in patter recognition. 

28. Fundamentals of Biometrics – Hand Written Signature and Iris

This chapter is a well-written, extensive review of handwritten signature and iris recognition methods. All aspects of offline and online handwritten signature analysis systems are discussed: the methodology used, preprocessing, feature detection, classification, acquisition, existing databases, and commercial systems. The iris-based biometrics is discussed in detail, from iris anatomy to feature extraction as well as preprocessing and classification methods. This chapter is an excellent start for researchers interested in handwritten signature analysis or iris recognition.

29. Recent Trends in Iris Recognition

This is a short review of iris recognition methods, tackling the basic modules of current systems. It mentions briefly the performance measures and discusses some limitations of current techniques.

While we cannot argue about the correctness of its content we did not find this chapter useful considering that the previous chapter contains a more extensive overview of iris recognition systems.

30. Using Multisets of Features and Interactive Feature Selection to Get Best Qualitative Performance for Automatic Signature Verification

This chapter presents a method for automatic (offline) signature verification where multiple sets of features are used, in contrast to other current methods that use one set of features. Features used in this research are: percentage of slant measured globally and locally on six regions of the signature, the coordinates of gravity center, the width and height, the baseline and the percentage of signature pixels from the total area. Multiple sets of features are selected by the “circulant matrix” technique briefly described here. Experiment results show an important performance improvement over the method using one set of features. An interactive tool for feature selection proven useful in this research is also discussed.

31. Fourier Transform in Numeral Recognition and Signature Verification

This chapter starts with an introduction to the Fourier transform and discrete Fourier transform and their properties. Then it is shown how DFT can be used for handwritten numeral recognition and for online signature verification. This chapter is missing a comparison with other methods and experimental results.