Improved image registration based on SIFT features

SIFT (Scale-invariant feature detection) feature has been applied on image registration. However, how to achieve an ideal matching result and reduce the matching time are the most important steps that we study in our work. The original SIFT algorithm is famous for its abundant feature points, but the final keypoints are so excessive that the matching speed is very slow at the next step of searching for homonymy point-pairs. In this paper, we analyze the performance of SIFT and conquer its deficiencies applying RANSAC arithmetic and Least Squares Method in order to reach a perfect robustness and precision. Experiments with real-world scenes demonstrate that the method can reach a better precision and robustness, which outperforms previously proposed schemes. Compared with conventional localization algorithm, this method makes the precision more stable, which reaches 0.01 pixel, and also reduce the time of image registration.