Abstract: In order to obtain higher accuracy of face recognition, it could meet the realtime requirement of face recognition, we proposed a face recognition algorithm based on optimal discriminant feature and relevance vector machine. Firstly, wavelet transform was used to denoise face image, and multi direction and multiscale Gabor features of face were extracted. Secondly, kernel principal component analysis was used to screen Gabor features of faces to find the optimal discriminant feature which had a great influence on face recognition results, the number of features was effectively reduced, and redundant information among features was removed. Finally, relevance vector machine was used to learn the optimal discriminant feature vectors and establish multiclassifier for face recognition, and standard face database was used to carried out experiments to test performance compared with the classical face recognition algorithms. The experimental results show that the average face recognition rate of the proposed algorithm is greatly improved, and the average face recognition time is less than that of the classical face recognition algorithms.

Key words: feature dimension reduction, relevance vector machine, face classifier, face image, optimal discriminant feature