Research On Image Registration And Application Based On Point Feature

In the field of digital image processing, image registration is a very important branch, it solves registration problem of two or more images with relative rotation, translation and scaling, the images got form the same scene at different time, different angles use the same or different optical equipment.The purpose of image registration is to match the images and removing distortion between the images. At present, the image registration techniques have been used in many fields, such as medicine, remote sensing, military application and soon.At present, the image registration based on point feature has become a research focus in the field of image registration, feature point extraction and matching is the two important parts in image registration, the main research contents of this article is extraction and matching feature points, aiming at the shortcomings of the classical algorithm to do the corresponding improvement, and realize join together images with the improved algorithm at last.In this paper, the main research work can be divided into the following several aspects:1. Build the model of image registration through mathematical means, classify the image registration from the method and application two aspects, and divided the image registration into four steps in detail.2. Make a detailed description with the process of feature points extraction. Detailed introduce three classic feature points extraction operator:Harris operator, SIFT operator, SURF operator, analysis the performance of the three kinds of operators and made a simple comparison. Given the SIFT operator’s high precision and good robustness, selecting SIFT operator as the main research content; SIFT algorithm’s shortage is the128d feature descriptor, it lead to a lot of storage space toke up and long operation time used, aiming at the shortcomings of the algorithm, this paper puts forward the improved SIFT algorithm, the improved algorithm described the dimensions of the feature descriptor from128d down to48d, at the same time, feature descriptor describes range from16*16up to24*24. Experiments show that the improved algorithm not only reduce the occupied space and operation time, also improve the matching precision in a certain degree.3. Introduce the process of feature points matching. Firstly, describe the principle of classical Kd-tree query method and the build process, and raise the BBF query method through the deficiency of the algorithm, BBF query method not only improve the matching efficiency but also extend the application scope; Secondly, introduce the ratio of purification method to remove false matching points; Finally, in view of the purified by still more image matching point, this paper puts forward the Kendall coefficient constraint, the constraint can effectively eliminate false matching points but will increase the operation time, so it used with certain restrictions.4. Combined the improved SIFT algorithm with the BBF query method and practical application in the joining together of two images. Simple introduction RANSAC algorithm and linear gradient fusion method used in splicing process, and give the splicing image working sketch at last.Through the different types of image registration and the final image stitching in this article, prove that the image registration algorithm based on point feature has good effect of registration, and have a good prospect in the field of image matching, image super-resolution reconstruction and so on.