Research On Image Registration Method With Stabilization And Affine Invariance

Image registration is a fundamental problem in image processing, which is a process to match two or more images of the same scene taken at different times, from different viewpoints, or by different sensors. Image registration is essential precondition for many image analysis tasks and applications such as pattern recognition, change detection and 3D reconstruction. Its abroad applications and complexity make image registration to be one of hot issues in image processing.The edge-based image registration method is common used technique in the image registration. There are two tasks which need to be handled during an image registration process. They are the feature extraction and correspondence establishment. In this thesis several problems are discussed: At the feature extraction phase, corner, as an important local feature, not only holding the basilica information of the image, but also reducing the computation effectively, has been the important image feature in image understanding and pattern recognition. So in this thesis, the feature points here are the corners of the different parts which have been detected. In this precondition, we study how pick out the points which are more stable and more effective. For this purpose, we introduce the concept of condition number, which give a quantitative analysis to corners'stabilization. Furthermore, it can overcome errors introduced by noise and ill-conditioning, which advances the performance of registration methods. At the correspondence establishment choice phase, incorrect match points often make the common match methods failed. So we choose the RANSAC method which behaves excellent in this aspect. After considering the incorrect match points and eliminating the influence, RANSAC method can distinguish them, and then compute the optimum transform according to transform estimation function. According to the RANSAC method, we lay emphasis on convex hull's affine invariance and local controllability. Then, we propose a convex hull RANSAC-like method, which induce the computation, improve the speed, and make the registration more efficiency.At last, experiments and analysis are given to show the algorithms we proposed have a good stabilization, invariance and application performance.