Research On Retinal Image Registration In Color Fundus Images

With the rapid development of computer science,image registration technology has gradually attracted wide attention from all walks of life,and has been widely researched and applied in many fields,such as computer vision,medical diagnosis,automatic tracking and positioning,face recognition,and image 3D reconstruction.Image registration is the process of matching and superimposing the same content of two or more images of the same scene.This article focuses on techniques for the registration of fundus retinal images that are important in practical medical clinical practice.The fundus retinal image is an important basis for the diagnosis of diabetes,glaucoma,hypertension,coronary heart disease and other diseases.If two different time or different modes of fundus retinal images of the same patient are registered,the doctor can provide a variety of complementary information of the diseased tissue or organ,thereby providing a more comprehensive basis for the doctor to diagnose the condition.Although the researchers have done a lot of work,the technique of retinal image registration in the fundus still has a lot of room for improvement.Firstly,this paper introduces the background and significance of retinal image registration research,and analyzes the research status of fundus retinal image registration.Then it conducts in-depth research on some key technologies in fundus retinal image registration and proposes some new ideas and algorithms.The main work and innovations of this paper are as follows:1、Aiming at the problem that the registration algorithm has a long running time in the face of large feature search space,this paper proposes a color fundus retinal image registration algorithm based on bifurcation point and SURF algorithm.The algorithm consists of three parts:feature screening extraction,feature matching,and estimation transformation model.Firstly,we use the eight-neighbor method and the SURF algorithm to detect the bifurcation points and the SURF key points,respectively.Secondly,the SURF points are searched in a rectangular template region centered on the bifurcation points.Then,the extracted features are matched by Euclidean distance,and the 30 matching point pairs with the best similarity are used to estimate the transformation model parameters.The experimental results show that the proposed algorithm can quickly and effectively achieve retinal image registration and reduce a large number ofunnecessary searches.2、In order to remove the mismatching relationship more effectively and avoid manual operation,an image registration algorithm from coarse matching of Euclidean distance to RANSAC fine matching is proposed.Firstly,the algorithm uses the Euclidean distance for initial matching and sets the reserved value.Secondly,the RANSAC algorithm is used to perform fine matching,and the mismatching relationship of the reserved matching pairs is removed.Then,the remaining correct matching point pairs are used to estimate the transformation model parameters.The algorithm is validated on the public fundus database FIRE.Experiments show that the algorithm can effectively remove the error matching,extract more correct matching point pairs,and improve the matching precision.