Abstract: Still-to-Video face recognition is an identification technology, whose gallery contains still face images and the query is a video with low quality. With the purpose of overcoming the difficulties of face alignment and motion blur, this paper proposed an improved approach for face recognition in videos based on sparse representation. According to the gradient variance, the proposed approach enabled face alignment in videos. Moreover, motion blur was solved by the dictionary which was constructed by multi-scale filtering. Further, using the cross correlation coefficient between frames, key-frames in video sequences were selected. Experimental results show that the proposed approach outperforms other several approaches, such as neural networks and support vector machines.