Improved Kernel CCA: A Novel method for Face Recognition

It is known to all that obtaining an effectual feature representation is of paramount importance to face recognition. In this paper, the latest feature extraction method based on KCCA is introduced. However, in the training stage of the standard KCCA-based extractor, it requires to store and manipulate the kernel matrix, the size of which is square of the number of samples. When the sample numbers become large, the calculation of eigenvalues and eigenvectors will be time-consuming. In order to enhance the extraction efficiency, this paper proposes to utilize a feature vector selection (FVS) scheme based on geometrical consideration. The algorithm can select a subset of samples whose mappings in feature space are sufficient to represent all of the data in feature space as a linear combination of them. Hence, this will largely reduce the computational complexity of KCCA. Furthermore, the framework of KCCA plus SVDD-based classifier used in face recognition is also proposed. Both the theoretical analysis and the experiment results demonstrate the competitiveness and efficiency of the proposed method compared to the conventional KCCA-based methods.