Discriminative Moment Feature Descriptors for Face Recognition

Introduction

Success of any face recognition system depends extensively on the discriminative competence of features extracted to represent facial images. In this regard, several approaches have been reported in literature which can be categorized into structural and statistical methods. Structural techniques emphasize on individual face features such as eyes, nose and mouth or on facial distances (Brunelli & Poggio, 1993; Chellappa & Malsburg, 1992; Cox, Ghosn, & Yiaios, 1996; Kanade, 1973; Lades et al., 1993; Manjunath et al., 1992). Statistical approaches focus on the statistical distribution of the pixels and include methods such as those based on subspace (Bartlett, Movellan, & Sejnowski, 2002; Belhumeur, Hespanha, & Kriegman, 1996; Liu, Huang, Lu, & Ma, 2002; Martin, 2006; Turk & Pentland, 1991), histograms (Ahonen, Deniz, Bueno, Salido, & Torre, 2011; Hadid & Pietikainen, 2004), filters (Bhuiyan & Liu; 2007; Struc, Gajsek, & Pavešić, 2009), transforms (Hafed & Levine, 2001; Spies & Ricketts, 2000) and moments (Arnold, Madasu, Boles, & Yarlagadda, 2007; Foon, Pang, Jin, & Ling, 2003; Haddadnia, Faez, & Ahmadi, 2003; Pang, Teoh, & Ngo, 2006; Rani, 2012; Saradha & Annadurai, 2005; Singh, Mittal, & Walia, 2011; Singh, Walia, & Mittal, 2011, 2012). Feature extraction techniques generally follow two approaches for invariant face representation. In the first one, the actual images having noise factors of illumination or pose are corrected to make the standard images, followed by extraction of features. In the second approach, features invariant to these factors are extracted directly from the images. Moments-based methods, which are centred on the latter approach, have been extensively explored for face recognition in earlier studies owing to their invariance and efficient image reconstruction abilities. The magnitudes of these moments extracted at some order are used as invariant image descriptors. These methods possess minimum information redundancy, are robust to noise, invariant to rotation and can be made translation and scale invariant through proper normalization. Zernike moment (ZM) is considered the most successful among them, with high efficacy and promising results.