Article Preview
TopIntroduction
In the last few decades, face recognition has received extensive attention from research community due to its vital applications in security and surveillance systems. Several face recognition algorithms have been proposed in literature (Zhao, Chellappa, Phillips, & Rosenfeld, 2003). However, the performance of most of the algorithms deteriorates in uncontrolled environment as these are sensitive to change in lighting, pose, occlusion and other conditions. In particular, pose variation has been identified as one of most challenging problem in face recognition (Zhao, Chellappa, Phillips, & Rosenfeld, 2003; Chen, Tan, Zhou, & Zhang, 2006). It has generated interest among the machine learning and pattern recognition community. Many pose invariant face recognition algorithms have been proposed in literature (Zhang & Gao, 2009) to handle this problem. These face recognition algorithms have been classified broadly into three categories: (i) General algorithms, (ii) 3D approaches, and (iii) 2D approaches.
General algorithms do not have any specific strategy to address pose variation but are developed to handle all image variations equally (e.g., age, illumination, expression, and pose etc.). General algorithms are further divided into two subcategories: (a) holistic and (b) local. Holistic approaches use the whole face image for feature extraction. Examples of such approaches are: Principal component analysis (PCA) (Turk & Pentland, 1991), Linear (Fisher) discriminant analysis (LDA) and its variants (Belhumeur, Hespanha, & Kriegman, 1997), Directional corner point (Gao & Qi, 2005) and Line edge map (Gao & Leung, 2002) etc. Local approaches make use of local regions from face images such as eyes, nose, mouth etc. for feature extraction. Elastic bunch graph matching (Wiskott, Fellous, Kruger, & von der Malsburg, 1997), Local binary patterns (Ahonen, Hadid, & Pietikainen, 2006) and Modular PCA (Pentland, Moghaddam, & Starner, 1994) etc. are the approaches in this category. Although General approaches are simple and fast but cannot handle larger pose variation and also sensitive to pixel misalignment.
Face recognition with assistance of 3D models has become one of the successful approaches to handle pose variation. 3D approaches try to capture the variations in image due to pose variations in 3D space rather than limiting them within the image plane. The 3D face models used in face recognition can be divided into three categories (i) generic 3D shape models (Gao, Leung, Wang, & Hui, 2001; Liu, & Chen, 2005; Zhang, Gao, & Leung, 2006), (ii) personalised 3D scans (Ishiyama, Hamanaka, & Sakamoto, 2005; Kakadiaris Passalis, Toderici, Murtuza, Lu, Karampatziakis, & Theoharis, 2007; Wang, & Chua, 2006), and (iii) personalised 3D models reconstructed from 2D images (Blanz, & Vetter, 2003; Castillo & Jacobs, 2007; Georghiades, Belhumeur, & Kriegman, 2001; Jiang, Hu, Yan, Zhang, & Gao, 2005). 3D approaches generally perform better at the expense of significantly increased implementation complexity or even computation time, thus outweighing their advantages for real time face recognition.