Learning kernel subspace for face recognition

Kernel subspace methods are effective for face recognition. However, their final solutions need to take advantage of all training examples, such that their runs in real-world applications with large scale training data sets are time-consuming. This paper proposes to apply radial basis function neural network (RBFNN) to learn the feature extraction process of kernel subspace methods, specifically, Kernel Principle Component Analysis (KPCA) and two-phase Kernel Linear Discriminant Analysis (KLDA or KPCA+LDA), in order to improve the running efficiency of testing phase of kernel-based face recognition system. Experimental results based on three small face benchmark data sets including ORL, YALE and UMIST as well as two larger benchmark data sets including AR and FERET, show that the proposed method can reach approximately the recognition accuracy of the original KPCA or the two-phase KLDA by using less vectors in its final solution than those of the two original kernel methods. The proposed method can be applied to real-time or online face recognition systems.