Abstract: In order to better detect the occlusion of the face shape, this paper propose a sparse coding-based method for facial feature point detection. This method adopts the idea of cascade regression, and uses the reconstruction model based on sparse constraint to iteratively find the positions of facial feature point. Firstly, a sparse dictionary is obtained during training stage, which is global and universal. Then, using the support vector regression and combining the texture information in the local region of the facial feature points, the method can reconstruct differential shape for each iteration. In order to verify the feasibility and effectiveness, this method is tested on three public face datasets, and compares with two methods. The experimental results show that our proposed method is feasible and the recognition rate of occluded face feature points is high.