According to the problem that conventional subspace linear analysis method can t describe the man-ifold structure of the sample accurately，based on the theory of canonical component analysis algorithm and localitypreserving projections algorithm，we combine two kinds of algorithms． The article puts forward restricted- class lo-cality preserving canonical component analysis algorithm． The algorithm distinguishes the sample information andkeeps the local information structure of the sample by introducing the class information． The algorithm achieves thelevel of maximum related on the base of making a distinction of the sample class information． It as well as makesthe feature projections have no relativity． This algorithm significantly improves the algorithm s recognition rate． Thealgorithm respectively does the experiment on the Yale face and AR face database． The recognition rate can reach to98%．