Transactions on Machine Design Zetao Jiang, Chuan Guo 4 1 2016 http://www.dline.info/tmd/fulltext/v4n1/tmdv4n1_2.pdf Aiming at long registration time and mismatch problems resulted by images that exist local region similar or generated SIFT vector similar in the traditional SIFT registration method, a fine registration method based on geometric constraints of block and parallel architecture is put forward. This method using SIFT algorithm parallelly extract feature points to calculate initial transform matrix, in order to provide geometric constraints for block and fine registration, after segment the overlapping region into several blocks, we do blockwise SIFT matching using parallel architecture to achieve fine registration. During the fine registration, we use affine invariance of Mahalanobis distance to screen feature points and eliminate duplicate matches and mismatches. The experimental results show that this method eliminates the mismatch generated by same local features, registration accuracy and speed has improved, the proposed approach has practical value.