Abstract: CASPCM model is proposed to make up the disadvantages of ASPCM model while dealing with faces with large angles.The training samples are grouped according to their distances to model centers and a local ASPCM model is constructed for each group.Synthesis result of CASPCM model is obtained by averaging results of the local ASPCM models with appropriate weights.Gradient-descent algorithm is used to iteratively improve estimate of the head pose.Accuracy and generalization are used to gauge analysis and synthesis abilities of the model.Experimental results show that the two abilities of CASPCM model are both superior to ASPCM model;recognition ratio of CASPCM model is 7% higher than ASPCM model.