Sparse Representation And Discriminant Analysis Techniques In Face Recognition

This thesis studies and enriches two topics in face recognition, i.e., the application of sparse representation techniques in face recognition, and linear discriminant anlaysis and its extensions for various applications.Sparse representation technique stems from the signal processing community, whose basic idea is to select from a large number of elementary signals to linearly reconstruct a target signal, in order to achieve low compression rate. For the past three years, this idea has been employed to do face recognition and brought to it a breakthrough, as well as new challenges. In this thesis, we study and enrich the applications of sparse representation in face classification and feature extraction. Current sparse representation (SR) based classification schemes does not take full advantage of the property of SR. In section 2.1, we discuss the interpretation of SR from the perspective of classificaiton, novely introduce a intepretable SR, i.e., fused SR, to do face recognition, and design a simple but effective classification rule. Regarding feature extraction, current methods do not consider the difference between representative powers of intra-class data and inter-class data. In section 2.2, we consider such difference and develop a novel feature extraction method that tends to maximize representative errors of inter-class data and simultaneously minimize representative errors of intra-class data.In Chapter 3, we study linear discriminant analysis (LDA) under different problems. LDA is a classic statistical feature extraction technique, which aims at obtaining discirminant features via linear transform. In this thesis, we develop LDA from the following two aspects: 1) In section 3.2., we try to solve the class imbalance problem encountered in class specific LDA; 2) in section 3.3, we propose a novel multi-modal biometrics technqiues based on subclass discriminant analysis to do face and palmprint recogntion.Experiments are conducted on several databases to validate the effectiveness of all proposed methods and results are reported.