Partial least squares (PLS) algorithm has been widely used in face recognition recent years, but the improved algorithm of PLS did not consider both the constraint of nonnegative and sparsity to improve the recognition accuracy and robustness. To take over these disadvantages, the paper proposes a novel approach to extract the facial features called Two-dimension Nonnegative Sparse Partial Least Squares (2DNSPLS). The main idea of the approach is grabbing the local features via adding the constraint of nonnegative and sparse to 2DPLS, which make the approach gain not only the advantages of 2DPLS, incorporating both inherent structure and category information of images, but also the local features, having nonnegative interpretability and sparsity. For evaluating the approach’s performance, a series of experiments are conducted on two famous face image databases Yale and PIE face databases, which demonstrate that the proposed approach outperforms the state-of-art algorithms and has good robustness to occlusion.