Image Registration Method Based On The Angular Point And Scale Invariant Features Transform

Image registration, that is, the same scene taken at different times, different perspectives or different sensors, two or more images according to some similarity measure, carries on the spatial transformation processing, make the images corresponding to the geometry. In image stitching, medical diagnosis, computer vision, remote sensing image processing and other fields, image registration techniques is one of the essential basic steps, and get the result of registration will directly affect the subsequent image processing. As a result, image registration have been focused on and emphasized. Based on image registration process using the different image information, registration method can be divided into three categories:the registration method based on gray level information, the registration method based on transform domain and registration method based on feature. Which based on the characteristics of the method now is the most common method in image registration, its biggest advantage is able to through the analysis of the point of image features to analyze the whole image, which greatly reduces the computational complexity in the process of image processing. So this paper mainly study the registration method based on feature points.This paper introduces the commonly used corner detection algorithm, and mainly introduces the Harris corner detection operator and Susan corner detection operator, and the principle of the normalized cross correlation (NCC) as a similarity criterion, combined with the two operators to achieve image feature point extraction and matching.Based on scale invariant feature transform (SIFT) matching method is one of the main method of image registration. After a great deal of experiments proved that SIFT operator has the invariance to the brightness changes, rotation changes, scale changes of the images, and has a certain robustness also to noise, affine transformation and perspective changes. Feature points of SIFT operator has128dimensional feature vector descriptor, the feature vector has a wealth of information and unique features, can achieve precise match very well. But the feature points extracted by this algorithm are too many, the dimension of feature vector descriptor is too high, the registration time is too long, and there are still a small amount of unstable points in the large number of feature points extracted, these points can make the registration precision and efficiency is reduced. To solve these problems, this article improves SIFT algorithm from the following several aspects.First, this article introduces Harris corner detection operator in the image registration based on SIFT algorithm, calculates the Harris corner response for these feature points which extracted by SIFT operator, eliminates those characteristics points of low response, in this way, the feature points after screening are more stable, more representative image information, and reduce the huge amount of calculation in the process of feature points feature vector description.Second, in SIFT operator the128-dimensional feature descriptor is reduced to24-dimensional in this article. In this way, reduce a lot of time in the matching process, and accelerate the registration velocity.Third, due to although the speed of unidirectional matching algorithm is faster, but the matching efficiency is not high, and in the matching results there still appear many wrong matching, therefore, change the matching strategy to bidirectional matching, the matching algorithm although increased the match time, but it greatly improves the matching efficiency, basically no mismatch in the matching results.Experimental results show that the improved algorithm eliminates a lot of SIFT feature points which are not unique and feature points are left mostly gathered in the outline of the image, and the number less than half of the original number of feature points. Because of the improved algorithm eliminated most of the feature points, and simplified the feature descriptors, thus greatly reduced the first half of running time of the registration process; With bidirectional matching method, although the second half of the running time of registration process will be increased a small amount, but on the whole, many registration time is shortened, and is very well to improve the matching accuracy and efficiency.