Feature Matching Algorithm Based On Points In Image Registration System

Image registration is a key and fundamental technology in image processing and pattern recognition, and it is widely used in many areas such as medical image processing, computer vision and remote sensing. Along with the development of science and technology, the requirement for the accuracy, efficiency and robust of image registration methods has increased. Among the image registration methods, the one based on features is capable of quickly registering images of different nature and efficiently handling larger and/or more complex misalignments or variations between scenes. Meanwhile, feature matching is the most crucial step in image registration. This dissertation concentrates on the feature matching methods based on corner points so as to match features accurately and efficiently. The main research work in the dissertation is as follows: 1. Two main corner detection methods that are Harris and Susan are introduced and analyzed. The methods are tested on images in different conditions. The experiments illustrate that Harris corner detection method has superior performance in computation speed and the change of intense and scale of images, whereas Susan corner detection method is good at handling the rotated images and noise images, and has the good positioning accuracy. 2. An approximately global optimization method based on ICP is introduced by using a new energy function and a new representation about point sets. The method is tested on a set of unstructured points and the points detected in natural images. The results of experiments indicate the introduced method can match feature better than ICP. 3. Two improved strategies based on WGTM algorithm are proposed to remove the outliers produced in the process of matching features: weighting the two graphs generated by the corresponding points in sense and reference images; saving the inliers which are determined at iterations. The two strategies are tested on the images on three different conditions. The experiments demonstrate that the proposed strategies have a better performance in accuracy and precision.