The Research On Non-rigid Image Registration Based On B-Splines

Image registration is a fundemental tool in image processing, which is widely used in the field of weather, remote sensing, computer vision, pattern recognition, etc. Image registration is the process of geometrically aligning two or more images of the same scene taken by different sensors, or from different viewpoints, at different times. The goal of image registration is to determine the spatial transformation that maps points from one image to homologous points on an object in the second image. For instance, satellite remote sensing image fusion and medical multimode image fusion are built on the basis of image registration. Image-guided radiotherapy and the check of lesions change are also based on image registration.This thesis introduces relevant research background of image registration and reviews the current registration methods. After the introduction of the framework of image registration, we compare the merits and demerits of feature-based and intensity-based registration algorithms, and then figure out the importance of elastic image registration in various fields. Therefore, this thesis is mainly foused on elastic image registration.We focuse on B-Spline-based elastic image registration and the principle and merits of B-Spline have been introduced in this thesis. In order to overcome the problem of feature points mismatch and highly local deformation respectively, main contributions of this thesis are as follows.1. We proposed an image registration method based on advanced SIFT and B-Spline. To address the situation of the mismatch of feature points extracted by SIFT algorithm, a threshold analysis method based on SD (the square of distance) has been introduced, which assures the accuracy and robustness of feature-basd registration. What's more, we adopt B-spline as interpolation method and use the distance of matched points as similarity measure. The experimental results show that the proposed method can overcome the problem of point mismatch effectively. This method can achieve great registration results on the images without sharp deformation effectively.2. A new elastic image registraion approach is presented for the images which contain highly local deformation, using hierarchical B-Spline Free-Form Deformation (FFD) and Overlapped Block Check (OBC) method. At first, we try to align the highly local deformation only by hierarchical B-Spline FFD, but the results are unsatisfactory. In our proposed method, we adopt big grid hierarchical B-Spline FFD for global elastic registration, and then divide the rough registered image and fixed image into a series of corresponding overlapped block pairs. We determine the pairs which need further registration by the difference images and then use small grid hierarchical B-Spline FFD. In addition, to avoid boundary effects, we propose a strategy to dominate the motion of control points near the boundaries. The purpose of using Overlapped Block Check is to assure that all the area can be aligned without the influence of grid control. The experimental results certify that the different spacing sizes of control points influence the accuracy of image registration in different aspects, and prove that the proposed method is more effective than the conventional hierarchical B-spline FFD method though difference images of registered images and fixed images and similarity measures in terms of MI, NMI, SAD, etc.