The Research On Image Matching Based On Feature Point

Image registration, is a process to match two or more images of the same scene taken at different times, from different viewpoint, or by different sensors, with the development of the computer vision, which is widely required in remote sensing, in medicine, in 3D reconstruction,and in computer vision (target location, automatic quality control), and so on. The methods of image registration can be classified into two categories: the intensity-based matching approach and the feature-based matching approach. However, as one of the most common methods, the largest advantage of the feature-based approach is its ability of translating the analysis of the whole image into its features that contains the feature point, the feature curve, etc. And as a result, it speeds up the image process. After many years of research, the feature-based image registration technology has made some achievements, the majority of the registration methods consist of four steps: image acquisition, feature detection, feature matching and image resampling and transformation.This paper is arranged according to the above four steps. Firstly, recent research, characteristics and application areas of image registration are discribed, thereby demonstrating its wide application prospect. Then image acquisition in two ways by adopting different hardware equipment is discussed, as well as common image transform model. Finally, there are two aspects which are we focused on: (1) The process of creating SIFT (Scale Invariant Feature Transform) descriptors and edge detection technology involved are emphatically given by studying edge and point features, Then, in order to improve edge detection result on the low SNR (Signal to Noise Ratio) image, a mathematical morphology of multi-structure and multi-scale element and Canny algorithm is introduced. So the performance of the details and anti-noise ability are strengthened, the convenience for the following steps such as feature extract and object recognition is provided; (2) Based on the SIFT feature extract, an algorithm using network flow to obtain optimal feature matching-Min cost K flow Problem is put forward, which utilizes the direction and gradient information of SIFT and the improved matching cost function for measuring the similarity on the network flow, with Minimum Cost Maximum Flow algorithm to derive the global optimal matching, under the arc ratio of the matching measure and the bi-directional check constraints to remove pseudo matching. Eventually, optimal matching will be achieved. As a result, about 14 percent improvement can be obtained. Experments on the test image sets demonstrate high accuracy, robustness, and more application.