Appearance-based face recognition under small sample size scenario
Jie Wang · 2007
Appearance-based approach is one of the most successful solutions to the face recognition (FR) problem, dominating the literature in the past decade. Its effectiveness, however, is limited by the technical difficulties pertinent to FR applications. It is well-known that the appearance of face object often exhibits considerable variations in illumination, expression, pose and other factors, resulting in a complex distribution of face patterns. At the same time, image samples available for each subject are limited which makes the characterization of face object properties a difficult task. This is the so-called small sample size (SSS) problem that exists in many real-world FR. applications such as surveillance photo identification and forensic identification. In this research, efforts have been focused on the development of appearance-based FR algorithms under the SSS scenario. By combining the strengths of the advanced kernel machine technique and the quadratic discriminant learning solution, a novel nonlinear FR algorithm, denoted as kernel quadratic discriminant analysis, is first proposed to address the problems associated with small sample size and complex face pattern distribution. The proposed method provides a general pattern recognition framework from which a set of novel or traditional learning algorithms can be derived. Following that, an automatic Gaussian kernel parameter determination procedure has been developed to constitute a meaningful contribution to kernel-based learning algorithms. In order to address the one-training-sample (OTS) problem, an extreme SSS case, a novel concept, denoted as generic learning, is introduced. Within the generic learning framework, state-of-the-art FR solutions are extensively studied and evaluated. In addition, a regularized feature selection algorithm is developed within the generic learning framework to make the unsupervised eigenface approach more attractive to classification tasks. Extensive experimentation has been conducted on the well-known data sets such as the FERET, AR and PIE databases, to evaluate the methods presented in this dissertation.