Image Registration Based On SIFT Feature Points Screening

LImage registration is the basic task of image stitching. Image registration algorithm is mainly applied to a number of different areas, including computer vision, pattern recognition, remote sensing and medicine and so on. It is of great significance and wide application. Image registration technology is development increasingly. Since there is not a image registration technique can be widely applicable to all situations, it’s needed to develop different image registration techniques for applications of different fields.Firstly, this thesis introduces the history and development of image registra-tion. This thesis respectively introduces the basic theory and method of image registration algorithm traditional gray-based algorithms and feature-based algo-rithms.Then, this thesis elaborates the image registration based on SIFT features.In this thesis, we firstly introduces the SIFT feature point detection and match-ing algorithm, and classical approach feature point selection (RANSAC). And it discuss the advantage and disadvantage RANSAC algorithm. SITF algorithm is more widely applied, but it produces a lot of mismatching points. For this rea-son, we propose a method based on the characteristics of the homograph matrix which was presented to eliminate mismatching points. SIFT algorithm is used to extract the feature points and get initial matching. Then, preliminary screen-ing is conducted and the benchmark homograph matrix is calculated by using similar triangles. Furthermore, we set a threshold value to eliminate matching point pairs that don’t meet the threshold and get accurate corresponding point pairs. Through the contrast experiment with both RANSAC algorithm and im-proved RANSAC algorithm combined with the Euclidean distance, it is verified the algorithm has higher correct matching rate.The general image registration is carried out on the whole image transfor-m without considering the characteristics of the local. As a result, it will pro-duce poor quality results in some image registration problems. This thesis imple-ments and analyses two algorithms based on local information, include content-preserving warping and Moving DLT. Image grid processing is convenient to image processing.In the end of this thesis, a summary of the whole dissertation is given, and future research work is prospected.