关键词: 神经网络, 径向基, 主元分析法, 线性判别式

Abstract: This paper uses the PCA method to extract facial features further mapped to the Fisher optimal sub-space，in this sub-space，between-class distribution within the same distribution ratio of the maximum，then puts forward a novel supervised clustering method，uses limited training data information to select the RBF structure and initial parameters.Finally，a hybrid learning algorithm to train the RBF neural networks，makes the gradient descent optimization algorithm greatly reduce the search space dimension.In ORL database the simulation results demonstrate that this method both in the classification error rate or the efficiency in the study can show excellent performance.

Key words: neural networks, radial basis, principal component analysis, linear discriminant