Research On Algorithms Of Multi-sensor Images Matching

Multi-sensor images registration techniques is a very challenging problem in the field of computer vision.This thesis dedicate to the matching algorithm of multi-sensor images,and mainly choose the infrared images and visible images as experimental subjects.The experimental results show that these two methods we proposed which named registration based on contour feature and registration based on grandient direction distribution fields can accomplish matching process correctly. To sum up, the main contributions of this thesis are as follow:(1) This thesis introduces the region gradient and adaptive weight to improve the traditional CV model. The extracted contours based on the improved model are more satisfactory than previous ones.Then we use the ESD distance and Hausdorff distance to calculate the similarity of the images of contour. The results of matching show that our method is more robust and accurate.(2) This is the first time to introduce distribution fields to the field of multi-sensor images registration and this thesis propose a new method named Multi-sensor images registration based on grandient direction distribution fields. First, we choose grandient direction to replace grayscale intensity for constructing the DF.Secondly, we define the dominant direction and use it to solve the problem of rotation invariant. Thirdly, we introduce Chi-square distance as similarity measure to calculate the similarity of two column vectors.The Chi-square distance improves the robustness of this method. Finally, we use hill-climbing to satisfy the real-time demand and improve the speed greatly. Experimental results show that this method can solve the problem of rotation accurately. And our method is very robust, fast and adaptive.