Study Of Image Registration Based On Feature

Image registration, which is the first issue should be settled in image data amalgamation and applied in the aspects of remote sensing image processing, medical image analysis and so on, is an significant task in the fields of computer vision and mode identification. There will be relative move, rotation and different proportion zoom between the same image data gained at different time using different sensor and imaging mode, so we must settle the image registration first of all.We seek a kind of space transformation for the space coherence achieving by corresponding point of the same objectives which represents the two images. The result of registration should make all the points or practical applied points at least in both images achieve matching. This is the course of image registration.The current image registration techniques can be divided into manual registration and automated registration. Automatic registration does not need manual intervention, and it is the ultimate goal of image registration technique. In this paper, the automatic image registration is studied. The current method of automatic registration can be divided into two major categories, area-based and feature-based. The methods of feature of registration are the focus in the article. In this paper, we complete the following works.1. Study the concepts, principles and common methods of feature-based registration. Then we present a new algorithm for image registration which is based on the features of patches and chain code, improving the traditional method which is based on chain code. This registration method can improve the registration accuracy.2. We study several extraction operators of feature points. Analyze performance of Harris operator, SUSAN operator, SIFT operator, by comparing the detection efficiency, rotation invariance, scale invariance and the performance of anti-noise. And we give some conclusions according to the analysis. Also, we analyze several matching strategies which are based on feature points. The matching strategies contain cross-correlation, virtual triangle, matching points supporting intensity, RANSAC strategy and so on. Then we study the performance of these matching strategies by the matching condition, stability, complexity, etc. We summarize the advantage and shortcoming of each matching strategy.3. According to the analysis above, we present three registration methods which are based on Harris, based on SUSAN and based on SIFT respectively. The experiment results demonstrate that these registrations can work well when they are used in remote sensing image and video image.4. At last, we implement the manual registration. Manual registration can apply to any kind image, but have to select matching points by persons. It is hard to deal with large images.