Research On Registration Algorithms Based On Global And Local Structural Similarity Measures

Point set registration is a hot and an important direction in the field of image processing and computer vision.It has important application value in medical image analysis,target recognition and tracking.The goal of point set registration is to restore the spatial transformation from the template point set to the target point set,so that the two point sets are aligned.The process is mainly divided into two steps,namely determining the corresponding relationship and solving the space transformation.They are prerequisites for each other,alternate iterations until the registration task is completed.Due to the existence of non-rigid deformation,noise,outliers,rotation and other interference factors in the point set,the accuracy and efficiency of the point set registration algorithm are all restricted to a certain extent.This thesis takes the probabilistic model matching method as the basic framework,researches and proposes a more robust and efficient point set registration algorithm.The main work is as follows:1.The existence of outliers will not only increase the difficulty of finding an accurate correspondence and the registration error,but also increase the computational complexity of the algorithm.A robust point set registration algorithm based on multi-scale sliding windows is proposed in this thesis.First,the template point set is scaled to construct multi-scale sliding windows with different scales.The two-dimensional and three-dimensional point sets construct rectangular windows and cuboid windows,respectively.Second,the sliding window traverses the target point set in equidistant steps,and the Hausdorff distance between the point sets in the overlapping area is calculated.Then the target subset with the highest similarity is located and segmented,and make the spatial position of two target objects as close as possible.At the same time,most outliers are screened out,which provides good initial conditions for the follow-up fine registration task.The experimental results show that the algorithm is robust when there are a large number of outliers in the point set.2.In order to deal with the problem that the Coherent Point Drift algorithm cannot handle large-angle rotations,a robust point set registration algorithm based on rough matching of outer contours,which adopts a matching strategy from rigid to non-rigid and from coarse to fine is proposed in this thesis.First,the minimum volume enclosing ellipsoid model of the point set is constructed.By solving the convex optimization problem,the external shape and spatial position of the ellipse or ellipsoid are determined.The angle between two ellipses or two ellipsoids can be calculated according to the corresponding principal axis vector and the rotation matrix and translation vector that “align” two ellipses or two ellipsoids are solved.Then,the corresponding rigid transformation is performed on the template point set,so that the directions of the targets in two points are kept as consistent as possible,which provide a good initial parameter for CPD.Experimental results show that the algorithm can deal with degradation problems such as large-angle rotation and improve the registration accuracy.3.Due to the complexity of image transformation,it is difficult to achieve the ideal registration effect by relying on only one feature.The traditional point set registration algorithm only considers the global distance between two point sets to determine the corresponding probability matrix,and the relationship between each point and the rest of the point set is ignored.A fast point set registration algorithm that combines features and spatial information is proposed in this thesis.First,a simple feature descriptor based on the anisotropic Gaussian kernel is constructed to fully explore the structual features of the point set itself.Through the spatial position information of the fusion point and the feature information obtained based on the anisotropic Gaussian kernel,a more accurate correspondence can be quickly evaluated.This method determines a better balance point between the feature description ability and time.While ensuring the accuracy of registration,the convergence speed of the algorithm is greatly improved,the registration time is shortened,and the registration efficiency is effectively improved.