Abstract: In view of the defects resulting from the combination of traditional method of feature extraction and BP neural network,this paper presented a new classification model “PCABP network”.Firstly,the PCA eigenvector is used to initialize the initial-layer weight matrix of the PCABP network,thus,the initial-layer of the new classification model “PCABP Network” replaces the role of PCA in the function of feature extraction.Secondly,in the training process,with the application of GHA and GD algorithm,dynamic adjustment to the projection weights matrix of the initial layer has been achieved,and accordingly,the PCA eigenvector has been optimized.This method optimizes “category separation” and “feature extraction” from source samples,finds out the best connection point between sample dimension reduction and classification,and replaces the traditional recognition pattern “firstly separate feature extraction,then classification by the use of classifier”.The experiment based on FERET face library verifies the effectiveness of this method.

Key words: BP beural betwork,PCA,GHA,RPROP,Gender classification,Mixed neural network