Discriminant analysis algorithms for face recognition
Pong Chi Yuen, Jian Huang · 2006
Face Recognition research started in the late 70s and a number of algorithms/systems have been developed in the last decade. Among various algorithms, appearance-based approach is one of the promising approaches in face recognition. Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA) are two most popular feature extraction and dimension reduction techniques used in the appearance-based approach. From classification point of view, generally, LDA-based algorithms perform better than that of PCA-based algorithms. However, there are two major problems when applying LDA in face recognition. The first one is the small sample size (S3) problem. It occurs when the image dimension is larger than the number of training samples. In turn, the within-class scatter matrix becomes singular. The second is the complicated and nonlinear image distributions when the images are captured under different poses and illuminations. In this thesis, discriminant analysis algorithms are designed and proposed, from the subspace approach and the regularization approach, to overcome these two limitations. In the subspace approach, a new Subspace-LDA (SLDA) method is proposed to solve the S3 problem. Compared with the existing LDA-based face recognition methods for solving S3 problem, the proposed SLDA method is more efficient and gives better performance. By applying the kernel trick in the SLDA method, a Kernel Subspace-LDA (KSLDA) method is proposed to solve the nonlinear face image distribution problem. One of the crucial factors in the Kernel approach is the determination of the kernel parameters which highly affect the generalization performance of the kernel-based learning methods. To further improve the performance of our proposed KSLDA method, an automatic parameter estimation algorithm based on eigenvalue stability, namely the Eigenvalue Stability Bounded Margin Maximization (ESBMM), is developed. This thesis also explores the regularization approach to solve the above-mentioned two problems. By applying the kernel trick into one-parameter regularized discriminant analysis (RDA) method, a kernel-based one-parameter regularized Fisher discriminant (K1PRFD) method is proposed. A gradient decent-based algorithm is also developed to find the optimal regularization parameter. In the final part, to utilize different advantages of different classifiers, a weighted combination scheme is proposed to integrate different kinds of face recognition algorithms. (Abstract shortened by UMI.)