Non-Rigid Image Registration Using Improved Optical Flow Field

The purpose of image registration is to find the mapping of two or more images of the same object,and then to achieve the automatic alignment justification of the images on physical coordinates.Compared to the deformation of rigid objects,that of non-rigid objects has the characteristics of severe local deformation,serious non-linear characteristics and inconsistent transformation model.Due to these reasons,the non-rigid image registration has always been the research focus in computer vision field.It is of important research significance and practical value in pattern recognition,medical image processing,robot vision,remote sensing image analysis,image fusion,etc.Firstly,this paper analyzes the background of non-rigid image registration and its research status.Then,the basic principle and classification of non-rigid image registration are introduced.Moreover,the algorithms based on optical flow field are emphatically introduced.At last,several commonly used image registration evaluation criteria are given.The traditional coarse-to-fine pyramid algorithm cannot effectively deal with the large deformation displacements which widely exist in non-rigid images.To solve this problem,an initial displacement field estimation method is proposed to correct the original optical flow field,and then to obtain the final optical flow field using iteration.At the same time,a gradient conservation is added to the data term to improve the robustness to the illumination change.A weight function is added to the regularization term to avoid over-smoothness.Moreover,an edge conservation term is added to the original energy function to preserve the detail features of image.Simulation results demonstrate the effectiveness of the proposed method.Finally,an improved optical flow algorithm based on feature vector is proposed.In order to realize the automatic registration of large displacement deformation and preserve the details of images,the traditional energy function is improved by replacing the assumption of brightness consistency constraint in the traditional optical flow model with that of feature vector conservation.In addition,by processing the displacement fields obtained by iteration with bilateral filtering in different layers,the noise can be removed while preserving the image edge information.Experimental results show that the proposed algorithm is robust and can achieve better registration results for complex non-rigid deformations.