Abstract: We define the correlative vector and effective supplement for classifiers combination, and bring forward a new combination rule, ie, maximal effective supplement rule. We do orthogonal wavelet transform of a face image, get its four sub-images of different frequency bands, then respectively extract their singular value features. We find in experiment that the correlation within these four feature groups and the correlation of their sorted results are all small, and that the combination results are obviously superior to the classification results of singular value features of the initial images and superior to the commonly used mark-counting combing method. We combine the classifiers with linear weights and use genetic algorithm to train the confident weights of every classifier.