An algorithm of face alignment and recognition by sparse and low rank decomposition

Abstract

Due to Sparse Representation Classification basing face recognition algorithm is easily disturbed by occlusion, noise and misaligned, it is difficult to obtain a perfect performance, for this, a new method of face alignment and recognition based on sparse and low rank matrix decomposition is proposed in this paper. First the training matrix is decomposed into a low rank matrix which is the clean face image and a sparse error matrix representing the noise, occlusion and other errors. Then a transformation matrix factor is utilized in the optimization model, which can be computed while decomposing the training matrix, realizing the auto face alignment in X-Y plane. Last, the low rank matrix is used as the face training data to be classified by sparse representation method. Experimental results show that the recognition rate of our method can perform equivalently with the newest LR-SRC method on face database which is contaminated by occlusion and noise but the posture is aligned, and increase by 1.92% to 97.28% when the training database is corrupted by noise, light condition, occlusion and misaligned.