Research On Image Registration On Improved SIFT Algorithm

As one of the important fields in image processing realms, image registration has become the research topic of object detection and recognition, 3D reconstruction,information retrieval and other image analysis fields. Image registration methods based on feature contexts extract feature information. Then, theses methods implement the process of matching via the relationship between feature contents. By virtue of better robustness, methods on feature have been widely used in various complicated applications. Among these algorithms, scale space methods have attracted more researchers. Sift algorithm has become one of the most widely used methods in image registration. However, this method still has some drawbacks, for example, complicated process of descriptor construction, low repeatability, orientation affected by discrete degree of histogram, and ignoring of local feature information.As for these problems, this paper does research and improvements methods on sift.Meanwhile, lots of experiments have been used to demonstrate the effectiveness and feasibility of our methods. Our research work will be exhibited as follows:(1) As for the complicated process of feature descriptor construction, an improved sift method based on fast feature descriptor is proposed. Firstly, this approach constructs descriptor based on local contrast context. Then, geometric location information is adopted to do the process of matching.(2) As for accuracy and repeatability in reality scene registration, an improved matching algorithm on multiple strategies is presented in this paper. This approach discards mismatched points via geometric feature consistency constraint and data clustering.(3) In the process of traditional sift algorithm, the feature orientation is affected by the discrete degree. Besides, traditional process of feature descriptor construction does not make full use of local feature information. As for these problems, a new algorithm based on statistics of feature distribution and consistency constraint will be presented in this paper. Firstly, this approach obtains principal orientation via remapping the gradient space. Then, this method implements the process of feature descriptor construction via statistics of feature distribution. Finally, a new matching rule on consistency constraint is adopted in feature matching section.In order to test the performances of our methods, we do lots of experiments. The experimental results demonstrate the feasibility and robustness of these proposed approaches.