Abstract:  Sparse representation based classifier (SRC) has been proved to be an effective classifier. However, it is well known that the overcomplete basis for computing the test sample’s sparse representation is constructed by all
training data. When the size of training data is very large, the computational complexity becomes the bottle neck of SRC. As a result，this restricts SRC to be applied in many real-time recognition problems.To address this problem,
an efficient and effective SRC approach is proposed in this paper.The SRC method is based on the assumption that any test sample lies in the subspace spanned by the training samples belonging to the same class. So the construction
of the over-complete dictionary of the proposed method is based on the following the fact that the samples used for the sparse representation of the test sample should be only samples in the classes which the test sample may belong
to rather than all training samples. So, a more compact over-complete dictionary is built for computing the sparse representation of the test sample, which greatly reduces the computational complexity. Specifically, for each test
sample，the k-nearest samples arc first computed using k-nearest neighbor orδ nearest neighbor; then the set of  classes the k-nearest samples belong to is defined as the neighbor class of the test sample and used to determine the
most likely classes the test sample belong to. Finally, all samples of the neighbor classes of the test sample arc used to construct the overcomplete set for the computation of the sparse representation.The experiment results on ORL
and FERET face databases show that the proposed method not only largely reduces the computational complexity, but also increases the performance of recognition as the influence from the noise classes is eliminated.