The Research Of Non-rigid Registration Algorithms Based On Image Characteristics And Optical Flow

Non-rigid image registration technology is an important research director in digital image processing,which has been widely applied in remote sensing image analysis and medical image analysis,etc.Classical non-rigid registration algorithms can be divided into registration algorithms based on characteristics and registration algorithms based on gray.Even though the computation time of the registration algorithms based on characteristics is very short,but the registration accuracy is low.While the registration accuracy of the registration algorithm based on gray is high,but the registration time is long.In order to deal with the problems that the classical algorithms could not satisfy the need of registration time and registration accuracy together,several kinds of related non-rigid registration algorithms are proposed based on deep research in the paper:non-rigid registration algorithm based on circle descriptor feature(CDF),non-rigid registration algorithm based on dynamic driving force Demons(DDFD)and non-rigid registration algorithms based on image characteristics and optical flow.Non-rigid registration algorithm based on circle descriptor feature(CDF),in which the feature points are extracted from the float image and the reference image,the circle descriptor is used to keep the rotation invariance of the registration,so the registration efficiency is accelerated.Non-rigid registration algorithm based on dynamic driving force Demons(DDFD),in which the driving force is changed by introducing the driving force coefficient,so the problems of registration time and registration accuracy can be solved effectively.Non-rigid registration algorithms based on image characteristics and optical flow,in which firstly the feature points are extracted from float image and reference image by using registration algorithm based on image characteristics,and these extracted feature points are used to get a coarse registered image(feature level registration).Secondly,on the basis of the coarse registered image,optical-flow algorithms are used to accurate registration(pixel level registration).Finally,based on the idea:from coarse to fine,from feature level to pixel level,five non-rigid registration algorithms based on image characteristics and optical flow(SIFT-Demons,SURF-Demons,SIFT-TV-L1,SURF-TV-L1 and CDF-DDFD)are proposed.In order to prove the effectiveness of the proposed non-rigid registration algorithms,checkboard image,natural image,colored image and medical image sequence are used to conduct registration experiments.The experimental results are as follows.Compared with the classical SIFT and SURF algorithms,the registration time of CDF algorithm is reduced under the guarantee of registration accuracy;DDFD algorithm has obvious advantage in both registration accuracy and registration time compared with classical Demons,Active Demons and TV-L1 algorithms;non-rigid registration algorithms based on image characteristics and optical flow are better than classical algorithms(SIFT?SURF?Demons?Active Demons and TV-L1)on both registration efficiency and registration accuracy,which also keep the adaptability for large-deformation images and the rotation invariance.