The performance of sparse representation based face recognition (SRFC) can be effectively improved by embedding a priori of similarity information. Aiming at expressions variations, partial occlusions and disguise in the uncontrolled face images, SRFC by embedding maximum block similarity information is proposed. Firstly, the training samples and query samples are divided into multiple non-overlapping blocks in the same way. Secondly, the similarities of corresponding blocks between the query samples and the training samples are calculated. Then, the maximum value is extracted to measure the similarity of inter-images. Finally, the extracted maximum block similarity information is embedded into sparse representation stage. Experimental results on AR face databases show that the proposed method achieves better recognition performance compared with those based on embedding global similarity, especially when both training images and query images contain expression, occlusions or disguise.