Face recognition using ICA and class information

In this paper we modify the basic ICA algorithm by utilizing class information, and then apply it to the field of face recognition. Firstly, we address the face representation used in ICA and adopt an architecture whose outputs are independent or as independent as possible. Secondly, we present our modified ICA algorithm. By minimizing the distance within class we derive a simpler and faster algorithm, which is more suit for recognition. Three public available databases (UMIST, ORL and Yale University) are selected to evaluate the recognition performance and computational cost. Experiments show that the results are encouraging.