Abstract: In order to improve the performances of bag-of-features（Bo F） in face recognition, a face recognition algorithm based on the fusion of multiple features under the framework of bag-of-features was proposed in this paper. This method firstly extracted different kinds of local features in face images to learn a corresponding over complete visual dictionary in advance. Then it mapped each local feature into a high dimensional mid-level semantic space, and employed spatial pyramid matching（SPM） to pool local coding features. Different kinds of features were concatenated as the final representation of images and classified by trained linear SVM. Experimental results on several benchmark datasets show that our method is more robust to position variations, expression changes and occlusion and can effectively solve the small training size problem.