Research And Implementation Of Image Registration Based On Feature

Image registration is a fundamental technology in digital image processing and patternrecognition, it has been widely used in many fields, such as image mosaics, imagesuper-resolution reconstruction and image fusion. The emphasis of image registrationresearch is how to improve the efficiency, precision and automation of registration. Thefeature-based image registration method has been widely used because it has thecharacteristics of less calculation and strong robustness.This thesis thoroughly studies both feature-based automatic image registration and fastimage registration, the main work is as follows:Firstly, the performance of four classic feature extraction algorithms(SIFT、SURF、ORB、BRISK) are compared and then the advantages and disadvantages of these algorithmsare analyzed through experiment. This research provide important references for selectingfeature extractor in pratical application.Secondly, in order to solve the problems caused when BRISK applied to imageregistration, an image registration method based on improved BRISK is proposed. First of all,initial global threshold is set according to image complexity and image keypoints are detected,the image is divided into several uniform sub-image blocks and the range of keypointsnumber of sub-image block is set, keypoints are added or deleted based on each sub-imageblock’s keypoints number, the distance between keypoints is limited; Keypoints are matchedand RANSAC is used to remove the mismatches and the transformation matrix is calculated.The experiment results demonstrate that the proposed method can set threshold adaptivelyand increase the degree of automation of image registration and has high precision at thesame time.Finally, an image registration method for image sequences is proposed. AGAST is usedto detect image corners and limit the number of corners, binary descriptors are createefficiently by comparing the sum of pixels intensity in random image patch. In the process of feature matching, a search region is set for every corner, which reduce mismatches situationefficiency. The experiment results show that this method improves the speed of registrationfor image sequences and has high precision.