Face Discrimination Using the Orientation and Size Recognition Characteristics of the Spreading Associative Neural Network

Abstract

With the rapid progress of the information society, biometrics identification technology has been developed for security applications. Biometrics is person authentication using physical features such as the face, fingerprints, irises, etc. It is advantageous for psychological resistance to be minimized; verification using the face is uninvasive compared with fingerprints. Such security systems using remote monitoring are in demand in customs house and airports, etc. Recently, face recognition by the on-line processing of facial images has been widely applied in various fields and evaluated the face recognition performance using large scale database (Phillips et al., 2007). The representative face recognition method is classified into two categories. The first is a feature-based approach which uses feature vectors created with complex Gabor wavelet coefficients at each node (Wiskott et al., 1997). The second is the holistic or pattern (template) matching approach. The well-known example for the latter is the approach using eigenfaces which are obtained from principal component analysis of a large number of either full face images (Turk & Pentland, 1991) or local feature images of the face, e.g. eyebrow, eye, nose, cheek, mouth, etc (Penev & Atick, 1996). In both approaches, the conventional nearest neighbor algorithm or neural network is used for face classification. In personal authentication using facial images, it is a common problem to realize robust recognition independent of variations of illumination, orientation, size, pose, and expression, etc. The various methods of orientation recognition for facial image were proposed (Wong et al., 2001; Wu et al., 2006; Su, 2000). The orientation of facial image obtained by these methods is used for the orientation correction before the face (shape) recognition process. On the other hand, the face size is usually normalized by using the information of distance between eyes or face width. Recently, a rotation and size spreading associative neural network (RS-SAN net) was developed based on space and 3-D shape recognition systems in the brain (Nakamura & Miyamoto, 2001). Using RS-SAN net, a personal authentication method, which was not influenced by the orientation and size changes was proposed. The RS-SAN net correctly recognized face shape, orientation and size, regardless of the input orientation and size, once facial images were learned (Nakamura & Miyamoto, 2001; Nakamura & Takano, 2006). However, the face shape recognition performance of the RS-SAN net was slightly low compared with other face recognition methods. In this chapter, we introduce a novel face recognition method using the characteristics of orientation and size recognition for decreasing false acceptance. Section 2 and 3 describe the outline of the rotation and size spreading associative neural network (RS-SAN net). 10