Facial Image Analysis Using Subspace Segregation Based on Class Information

摘要

Analysis and classification of facial images have been a challenging topic in the field of pattern recognition and computer vision. In order to get efficient features from raw facial images, a large number of feature extraction methods have been developed. Still, the necessity of more sophisticated feature extraction method has been increasing as the classification purposes of facial images are diversified. In this paper, we propose a method for segregating facial image space into two subspaces according to a given purpose of classification. From raw input data, we first find a subspace representing noise features which should be removed for widening class discrepancy. By segregating the noise subspace, we can obtain a residual subspace which includes essential information for the given classification task. We then apply some conventional feature extraction method such as PCA and ICA to the residual subspace so as to obtain some efficient features. Through computational experiments on various facial image classification tasks - individual identification, pose detection, and expression recognition - , we confirm that the proposed method can find an optimized subspace and features for each specific classification task.