Subspace Methods for Face Recognition: Singularity, Regularization, and Robustness

Abstract

Face recognition has been an important issue in computer vision and pattern recognition over the last several decades (Zhao et al., 2003). While human can recognize faces easily, automated face recognition remains a great challenge in computer-based automated recognition research. One difficulty in face recognition is how to handle the variations in expression, pose and illumination when only a limited number of training samples are available. Currently,face recognition methods can be grouped into three categories, feature-based, holistic-based, and hybrid approaches (Zhao et al., 2003). Feature-based approaches, which extract local features such as the locations and local statistics of the eyes, nose, and mouth, had been investigated in the beginning of the face recognition research (Kanade, 1973). Recently, with the introduction of elastic bunch graph matching (Wiskott, 1997) and local binary pattern (Timo, 2004), local feature-based approaches have shown promising results in face recognition. Holistic-based approaches extract a holistic representation of the whole face region, and have robust recognition performance under noise, blurring, and partial occlusion. After the introduction of Eigenfaces (Turk & Pentland, 1991) and Fisherfaces (Belhumeur et al., 1997), holistic-based approaches were extensively studied and widely applied to face recognition. Motivated by human perception system, hybrid approaches use both local feature and the whole face region for face recognition, and thus are expected to be potentially effective in improving recognition accuracy. In holistic-based face recognition, feature extraction is fundamental, which can be revealed from three aspects. First, the input facial image is high dimensional and most current recognition approaches suffer from the “curse of dimensionality” problem. Thus a feature extraction step is necessary. Second, facial image usually contains less discriminative or unfavorable information for recognition (e.g., illumination). By making use of feature extraction, this information can be efficiently suppressed while retaining discriminative information. Third, feature extraction can greatly reduce the dimensionality of facial image, and this reduces the system’s memory and computational requirements. Subspace method, which aims to reduce the dimension of the data while retaining the statistical separation property between distinct classes, has been a natural choice for facial feature extraction. Face images, however, are generally high dimensional and their withinclass variations is much larger than the between-class variations, which will cause the serious performance degradation of classical subspace methods. By far, various subspace methods have been proposed and applied to face recognition. O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg