Discriminant Subspace Analysis for Uncertain Situation in Facial Recognition

Facial analysis and recognition have received substential attention from researchers in biometrics, pattern recognition, and computer vision communities. They have a large number of applications, such as security, communication, and entertainment. Although a great deal of efforts has been devoted to automated face recognition systems, it still remains a challenging uncertainty problem. This is because human facial appearance has potentially of very large intra-subject variations of head pose, illumination, facial expression, occlusion due to other objects or accessories, facial hair and aging. These misleading variations may cause classifiers to degrade generalization performance. It is important for face recognition systems to employ an effective feature extraction scheme to enhance separability between pattern classes which should maintain and enhance features of the input data that make distinct pattern classes separable (Jan, 2004). In general, there exist a number of different feature extraction methods. The most common feature extraction methods are subspace analysis methods such as principle component analysis (PCA) (Kirby & Sirovich, 1990) (Jolliffe, 1986) (Turk & Pentland, 1991b), kernel principle component analysis (KPCA) (Schölkopf et al., 1998) (Kim et al., 2002) (all of which extract the most informative features and reduce the feature dimensionality), Fisher’s linear discriminant analysis (FLD) (Duda et al., 2000) (Belhumeur et al., 1997), and kernel Fisher’s discriminant analysis (KFLD) (Mika et al., 1999) (Scholkopf & Smola, 2002) (which discriminate different patterns; that is, they minimize the intra-class pattern compactness while enhancing the extra-class separability). The discriminant analysis is necessary because the patterns may overlap in decision space. Recently, Lu et al. (Lu et al., 2003) stated that PCA and LDA are the most widely used conventional tools for dimensionality reduction and feature extraction in the appearancebased face recognition. However, because facial features are naturally non-linear and the inherent linear nature of PCA and LDA, there are some limitations when applying these methods to the facial data distribution (Bichsel & Pentland, 1994) (Lu et al., 2003). To overcome such problems, nonlinear methods can be applied to better construct the most discriminative subspace. In real world applications, overlapping classes and various environmental variations can significantly impact face recognition accuracy and robustness. Such misleading information make Machine Learning difficult in modelling facial data. According to Adini et al. (Adini et al., 1997), it is desirable to have a recognition system which is able to recognize a face insensitive to these within-personal variations. O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg