Abstract: This paper make full use of dictionary learning and sparse representation for signal reconstruction and classification, and present a two-step dictionary learning method and apply it to robust face pose recognition. The proposed approach models the appearance of face images from the subspace via K-SVD that learns the sub-dictionary from a set of images. A combination of the trained sub-dictionaries of all pose classes are used as an over-complete dictionary. Finally, the Gabor features are extracted for sparse representation and classification. In order to improve the classification ability, we put forward a two-step dictionary learning method, and carry out dictionary learning with label constraints in the second step. Additionally, in order to improve robustness against face occlusion, we introduce a pose occlusion dictionary to code the occluded portions of face images. Several experiments were performed on XJTU, PIE, and CAS-PEAL-R1 databases. Recognition results show that the proposed method can achieve a recognition rate of about 95% under illumination, noise, and occlusion variations. It can satisfy the requirements of practical applications.