Article Preview
TopIntroduction
Face recognition offers numerous applications including authentication, access control, surveillance and human computer interaction. It has been an active research area over the past decades, but is still a challenging problem as the human face can undergo wide variations such as those due to pose, illumination, age and occlusion. Among the three steps in face recognition – face segmentation, feature extraction and classification, feature extraction is considered crucial as it ultimately governs the recognition accuracy. The feature extraction algorithm needs to possess good representative and discriminative capability, should be computationally efficient and also robust to facial variations.
Feature extraction algorithms reported in the literature are either based on global approach or local approach. Global methods extract features from the whole face image and thus, deal with the complete facial information. Some popular global methods include Eigenfaces (Turk & Pentland, 1993; Martin, 2006; Yang, Zhang, Frangi, & Yang, 2004), Fisherfaces (Belhumeur, Hespanha, & Kriegman, 1996; Liu, Huang, Lu, & Ma, 2002; Li & Yuan, 2005), Independent Component Analysis (Bartlett, Movellan, & Sejnowski, 2002; Liu et al., 2005) and orthogonal rotation invariant moments (Singh, Mittal, & Walia, 2011; Singh, Walia, & Mittal, 2012; Haddadnia, Faez, Ahmadi, 2003; Pang, Teoh, & Ngo, 2006; Foon, Pang, Jin, & Ling, 2004; Arnold, Madasu, Boles, & Yarlagadda, 2007; Singh, Walia, & Mittal, 2011). These methods are widely used as they are simple, independent of any geometrical structures and easy to implement but exhibit two drawbacks – firstly, these are less comprehensive as they do not focus on precise details which is crucial for face representation and secondly, they are also affected by face variations. In contrast, local methods focus on image sub-regions and possess the capability to represent even the minute facial details. They also exhibit better invariance to light, pose and expression changes and thus, are being intensively explored for face identification. The widely used local face descriptors are Gabor Filters, Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG) and Scale Invariant Feature Transform (SIFT). Gabor Filters can efficiently represent the facial lines or edges and are invariant to scale and orientation (Struc, Gajsek, & Pavešić, 2009; Bhuiyan, & Liu, 2007). The main difficulty is, however, their high computational complexity. SIFT is another prominent descriptor in this category which can extract rotation invariant features but SIFT features are susceptible to light variations (Lowe, 2004; Soyel, & Demirel, 2010). In particular, LBP operator has proved to be the powerful and most successful descriptor and has been applied in numerous state-of-the-art face recognition systems (Ahonen, Hadid, & Pietikainen, 2004). It captures the texture information and represents local regions of the face efficiently by comparing each pixel with its neighboring pixels. The two most important benefits of LBP are its computational simplicity and its tolerance to monotonic illumination changes. However, the downside of the LBP descriptors is that the size of feature vector produced has high dimensionality. To address this issue, several variations of LBPs have been proposed in the literature which significantly reduce the dimensionality of the LBP feature vector (Huan, Shan, Ardebilian, Wang, & Chen, 2011). Recently, a variant of LBP, orthogonal combination of local binary patterns (OC-LBP) has been proposed for image description which generates less dimensional LBP features while still maintaining the discrimination power (Zhu, Biichot, & Chen, 2013). This method, in contrast to other LBP variations, has achieved accuracy improvement of up to 5% on standard texture classification datasets.