Dakshina Ranjan Kisku

Design of Certain Robust Unimodal and Multimodal Biometric Systems for Human Identification

Human physiological or behavioral characteristics are available uniquely to each individual as biometrics evidence and the biometrics system can be able to verify or identify a person correctly. Reliable person recognition is an important problem in diverse fields. Biometrics recognition based on distinctive personal traits, has the potential to become an irreplaceable part of many identification systems. Security is the one of the major issues by means of individual’s self protection as well as protection of important documents from malicious users in today’s techno-savvy world. Security threats often occur as information thefts in terms of vulnerability of data by stealing security enabled information, identity thefts of individuals in terms of thefts of various ID cards information, cross-border vulnerability by suspected terrorists or illegal immigrants, etc. As a result, there is a need of identification/verification of the person and also of the things that may cause the security susceptible to attack. Therefore, it is necessary to build
an automatic system based on physiological (i.e., face, hand geometry and palmprint, iris, retina, ear, fingerprint and DNA) or behavioral characteristics (signature, keystroke and voice) that provide accurate biometric authentication or identification of people and save a country and its people from cross-border terrorism and counterfeiting of identities.
The thesis proposes some efficient and robust unimodal and multimodal biometric systems for human identification/verification. In case of unimodal system, it considers face and ear recognition systems. In the field of multimodal biometric system, it considers face, palmprint, fingerprint and ear evidences for fusion at feature extraction level, match score level and low level/sensor level. It also proposes multi-classifier and multi-algorithmic fusion strategy for developing a system using face and signature evidences. In unibiometrics contributions, the thesis presents three systems on face recognition and one system on ear identification. The first face recognition technique is developed based on SIFT features and graph matching topology. The graph-based face matching has three matching constraints, namely Gallery Image based Match Constraint, Reduced Point
based Match Constraint and Regular Grid based Match Constraint. Apart from the graph based face recognition, the thesis discusses a robust and cost effective face recognition technique using SIFT features that are extracted from face images. In this regard, face matching technique, based on local and global information and their fusion are proposed. In the local and global matching strategy, SIFT keypoint features are extracted from face
images in the areas corresponding to facial landmarks, such as the eyes, nose and mouth and matching is performed. Further, the matching scores obtained from the local and global strategies are further fused together using the Dempster-Shafer decision theory. The third face recognition technique is based on the Dempster-Shafer decision theory based fusion of relaxation graphs drawn on invariant SIFT features points detected from
salient landmark regions such as both eyes, mouth of a face. In the group of unimodal biometrics system along with the face recognition techniques, SIFT descriptor based an ear recognition system is also presented using Gaussian Mixture Model (GMM) and K-L divergence. Furthermore, the ear recognition system uses color segmented slice regions idea to increase the overall performance of the system. In multibiometrics applications, five multibiometric systems are presented. Such as feature level fusion of fingerprint and ear biometrics with Doddington’s concept, match score level fusion of face and ear biometrics using belief theory, Low level fusion of multispectral palm images using wavelet decomposition for person authentication using ant colony optimization, biometrics evidence fusion using face and palm print image
fusion with wavelet decomposition and monotonic-decreasing relational graph, and feature level fusion of face and palmprint biometrics by isomorphic graph-based k-medoids partitioning. Besides, the match score level fusion of face and ear biometrics using belief theory, and multispectral palm image fusion at low level, all three remaining methods have used SIFT features as a key and invariant feature descriptor.
The last contribution deals with two multi-algorithms fusion based biometrics systems. In this context, two methods have been proposed. Such as robust multi-camera view face recognition and an offline signature identification system. The multi-camera view face recognition system uses Gabor filter bank for facial features characterization and fusion of reduced eigenface and canonical face vectors using weighted mean fusion rule. On the
other hand, offline signature identification system that uses Support Vector Machines (SVM) to fuse multiple matchers by means of three distance similarity measures. The offline signature identification also uses global and local features as signature features. Experimental results obtained in case of unimodal face recognition algorithms demonstrate the effectiveness and potential for the human recognition systems which achieve more than 90% recognition accuracy while the algorithms use invariant SIFT features, various graph topologies, facial landmarks and fusion strategy. Even in few cases, more than 98% recognition accuracy is obtained with minimum FAR. On the other hand, ear identification achieves more than 98% identification accuracy when color segmentation idea of ear image and Dempster-Shafer theory based fusion rule is applied together. Without using color segmentation, it achieves about ~ 93% accuracy. Apart from these unibiometrics systems, proposed multibiometrics systems also achieve high
accuracies as ~ 99% recognition rate in case of two feature level fusion approaches such as fusion of face and palmprint, and fusion of fingerprint and ear biometrics. Also in case of multisensor based evidence fusion of face and palmprint biometrics, more than 98% recognition accuracy is obtained. However, score level fusion of face and ear biometrics determined ~ 96% recognition accuracy and multispectral palm image fusion algorithm shows robust performance labeled with more than 96% recognition accuracy and 6.25% determined as total error while SVM with RBF kernel function is used. Whereas in multi-algorithms fusion based biometrics, multi-camera view face recognition attains 98% and 96% recognition accuracies while SVM with RBF and linear kernels are used. The identification rate for the offline signature system is obtained as 97.17%. The
experimental results determine with each of these biometric systems not only show robust and efficient performance, but also show cost effectiveness for human identification/verification in a few unibiometrics and multibiometrics systems.