This paper addresses face recognition problem in a more challenging scenario where the training and test samples are both subject to the visual variations of poses, expressions and misalignments. We employ dense Scale-invariant feature transform (SIFT) feature matching as a generic transformation to roughly align training samples; and then identify input facial images via an improved sparse representation model based on the aligned training samples. Compared with previous methods, the extensive experimental results demonstrate the effectiveness of our method for the task of face recognition on three benchmark datasets.