Combined weighted eigenface and BP-based networks for face recognition

Computational complexity of K-L transform is the bottleneck of traditional eigenface algorithm. Test face image was divided into 9 sub blocks to reduce dimensions and computational complexity. Different weights were given to different parts of image according to their importance at recognition stage. So importance of human face key parts was enhanced. Within-class average face was adopted instead of mix average face, because it could keep with-class information and enlarge differences between classes. Adaptive learning step BP-based network was adopted as classifier. Experiments on ORL and Yale face database show that the recognition rate reaches 95.62% and 93.33%. The increases are 7.74% and 14.28% respectively compared with traditional algorithm. Analysis demonstrates the proposed algorithm in this paper has less computational complexity, higher recognition rate, and more robust than traditional algorithm.