Face recognition based on warping-robust elastic bunch of graph matching and projection-based discriminative feature analysis

Abstract

Although face recognition is a task that human beings perform effortlessly in their daily lives, it is not a simple job for a machine because the variation in facial pose, illumination, and facial expression cause serious nonlinearity of face manifolds. The high dimensionality of facial patterns and the small number of training samples for machine learning are also annoying problems which should be taken into account. In our research, the face recognition process is analyzed into four steps; face alignment, low-level feature extraction, discriminative feature selection, and classification. In this thesis, advanced methods for each step are proposed to solve the previously mentioned problems. In alignment step, enhanced graph matching method is proposed to elastically match a face to a multimodal face model describing the general human face. Our proposed robust jet and warping robust matching cost function is more effective to minimize nonlinearity in a face pattern than the conventional graph matching method. In low-level feature extraction step, directionally classified Eigen-block method is introduced as an alternative to the widely used Gabor wavelet transformation. The proposed method is superior to the conventional approaches in not only discriminativeness but also compactness of extracted features. In third discriminative feature selection step, to consistently find discriminative feature subspace, traditionally used Fisher’s criterion is modified by the span-based class modeling, which is effective in face recognition with small number of samples. Then to overcome the nonlinearity in face pattern distribution, we develop a kernel-based version of the proposed feature selection method. In last classification step, we consider the nearest subspace classification method as a better solution than the widely used nearest neighbor classification method in small sample size case. Moreover the kernel-based nearest subspace classification method is proposed to pr...