Research On Non-rigid Image Registration Algorithms Using Optical Flow Field

Image registration aims to find the spatial transformation relationship between two or more images obtained under different conditions,and then realize the spatial one-to-one mapping of each point.Compared to the rigid images,the non-rigid ones have many characters,such as severe local deformation,non-linear transformation,and non-uniform transformation models.Due to these reasons,non-rigid image registration is always a research hotspot and difficult point in the field of computer vision,and represents the study direction in the future.Firstly,this thesis analyzes the research background and significance of non-rigid image registration,and summarizes the research status at home and abroad.Then the basic theoretical knowledge of image registration and the commonly-used non-rigid image registration methods are introduced as well as the quality evaluation methods of image registration.Secondly,this thesis studies the non-rigid image registration algorithm using optical flow field model in detail.Aiming at the problem of details lost and lacking of precision on non-rigid images registration using traditional optical flow models,an improved optical flow field model is proposed.To avoid blurring and preserve the edge and detail features of images,a new adaptive diffusion regular term is proposed to replace the original uniform diffusion one.Besides,a non-local smoothness term,which contains neighborhood information,is introduced to remove optical flow outliers.Moreover,a weight function combining image structure and optical flow information is added to reduce the details lost caused by over-smoothing and then improve the robustness.Compared with the traditional methods,experimental results show the robustness and the ideal registration effects of the proposed method on different types of non-rigid images.Since the traditional optical flow methods always fail in registering large displacement deformations,a coarse-to-fine two-step registration method is proposed.At first,image segmentation technique and SIFT feature matching are utilized to achieve the local coarse image registration as well as obtain the initial displacement field,and then the precise registration can be realized using optical flow method.By combining the gray-scale conservation with the gradient one,the data item can be obtained to reduce the influence of illumination and noise.Besides,the adaptive regular term is combined to avoid over-smoothing and preserve image edges.Experimental results demonstrate the effectiveness of the proposed method.Finally,this thesis summarizes our works and discusses the foreground in the future.