The Research Of Image Registration Based On Feature Point Matching

In a different time, using different sensors, on different perspectives and on different shooting condition, Image registration matches the geometric relationship of the two or more images. Its main purpose is to be removed or inhibited the geometric inconsistencies of reference images and moving image registration, including translation, rotation, scaling deformation. Image registration is widely required in remote sensing, in medicine, in 3D reconstruction,and so on.Feature-based image registration method has two main difficulties: the diversity of feature extraction and similarity computation complexity. In this paper, in view of these two issues, the main job is as the following:Firstly, we introduced three kinds of common methods of interpolation in digital image processing, analyzed the advantages and disadvantages of these methods, and emphatically discussed the geometry transform principle of digital image in space. Meanwhile, we gave numerical simulations of geometry transform of image by Matlab.Secondly, we introduced characteristics of digital image on point, line and region, traditional corner detector principle about Moravec, Harris and SUSAN, analyzed and expounded the algorithm principles and procedures of SIFT, SURF and PCA-SIFT, and compared the detection performance of inspection operators for feature points. From Matlab simulation of feature points inspection of SIFT, SURF and PCA-SIFT, we found that the foregoing three kinds of feature point matching algorithm can not work out the matching problems of image with reversal and self similarity. In this paper, we gave a iterative orderly and k-neighbor algorithm, which realize matching of feature point of digital image. Digital simulations show that this algorithm can solve above problems.Thirdly, we introduced the definition and properties of Fourier transform and applications of cross-correlation in image registration, proposed a kind of image registration method based on feature-point detection, gave the procedure of the method, meanwhile, we verify the method by Matlab. Experiments show that the method has high accuracy on translation and rotation registration, and strong performance in resisting noise.Finally, this work is summarized this paper, and look to the future in the field of image registration.