The RBF neural network for classification is applied in face recognition. With two important criterion for estimating the initial width of RBF unit

the width can control the generalization ability of RBF neural network classifier. PCA method to the training sample set the projection to the face space

to reduce dimension. On the basis of the PCA transform

an optimal subspace classification makes the distance between the classes to maximize the ratio of the distance using FLD method. Simulation is conducted on the ORL database

and its results show that the algorithm is efficiency and effectiveness.