An improved SIFT method is proposed. Firstly, the step of describing key points changes the rectangular region of the descriptor into a circular region. By calculating the weighted gradient orientation histogram for each partitioning, a descriptor with 96 dimensions is composed. Secondly, we emphasize that a smaller nearest neighbor ratio threshold leads to a highly accurate matching probability for the matched point but yields a low number of matching points. The matching points are ranked by the distance threshold of the nearest neighbor to the second nearest neighbor. We finally simplify RANSAC by establishing a new dataset based on the optimal matched points. The experimental results demonstrate that our approach enhances computation efficiency (about 10 times) and slightly improves accuracy than other algorithm. When the matching feature points increases, our method has more advantages in matching speed and higher matching accuracy than RANSAC-SIFT.