To investigate the validity of Support Vector Machine(SVM) generalization error bounds as the feature selection criterion,a novel framework of facial feature selection based on Filter and Wrapper approaches was proposed.By taking a Wavelet Transformation(WT) and Kernel Principal Component Analysis(KPCA) as a Filter approach,the Vapnik-Chervonenkis(VC) Leave-one-Out(LOO) error bound was minimized.Then,the span bound of support vector was regarded as the feature selection criterion of Wrapper approach.Finally,Recursive Feature Elimination(RFE) search strategy was used for searching optimum facial subset.The experiments on UMIST face database were executed by the proposed method.The experimental results indicate that the facial feature dimensions can be reduced to 80 and 70,respectively,while both of the classification accuracies remain over 94%,so the proposed feature selection criterion and search strategy are effective for facial feature selection.