Research On Image Registration Algorithm Based On Point Features

Image registration is one of the key technologies in the field of computer vision,which is widely used in medical diagnosis,unmanned driving,3D reconstruction and other fields.However,image registration still faces many challenges,including efficiency,robustness and accuracy.Point features have the advantages of high stability and small amount of calculation.Therefore,the registration based on point features has become a commonly used registration algorithm.In this thesis,the point feature-based registration and its sub-problems corner detection are carried out in-depth research,and two improved registration algorithms based on point features and a new corner detection algorithm are proposed.This thesis mainly includes the following three parts:1.An image registration algorithm based on salient region is proposed.The traditional point-based registration algorithm has poor registration result in the case of complex background and high repetition of background heterotexture.To solve this problem,an algorithm that uses segmentation results to extract significant regions and eliminate the interference of complex background is proposed.HRNet algorithm is used to segment the image,and the background area and noise area are removed according to the segmentation results.Then the transformation model is obtained by extracting and matching the feature points from the remaining salient region images.The initial image is transformed using the obtained transformation model.Experiments show that the algorithm can effectively improve the registration performance.2.An improved RANSAC algorithm based on k-nearest neighbor similarity is proposed.RANSAC,the most commonly used model parameter estimation algorithm,has a sharp decline in registration efficiency and accuracy when there are many wrong matching point pairs.In this thesis,the k-nearest neighbor matching similarity algorithm is proposed by using the position,direction and scale information of feature points.By this algorithm,some mismatched points are removed.Then the sampling points set and the model quality evaluation points set are obtained by the RANSAC algorithm with two thresholds.A more accurate set of matching feature points is obtained by using a small threshold to estimate model parameters,which significantly reduces the number of iterations.Experiments show that the improved RANSAC based on k-nearest neighbor similarity improves the accuracy and efficiency of registration.3.An image corner detection and classification algorithm based on robust local descriptor is proposed.According to the characteristics of corner region,b-AG2 D is proposed,which indirectly detects corners by detecting radial-edges.The b-AG2 D has two characteristics.The first is to move the center of the detector to a relatively stable support region to reduce the adverse impact of the center region.Second,a new anisotropic Gaussian kernel is proposed,which is obtained by inserting a mean kernel function into the anisotropic Gaussian kernel function forming Gaussian-like kernel function,which enhances the stability and anisotropy of the detector.In order to improve the ability of discriminating edge points and corner points,a new corner measure,fake-residual area,is proposed.The method consists of five steps: First,the contour extraction and connection are performed.Second,the b-AG2 D is applied to extract a set of radial-edges in a region surround a contour point.The local descriptor is efficiently described by a set of radial-edges.Third,a thresholding method is utilized to pick up potential corners from the LCD.Fourth,non-maximum suppression is applied to determine true corners.Fifth,the number of the strengths that are greater than the threshold value is used to classify corner types.Experiments show that the corner detection algorithm in this thesis can improve the corner detection rate and classification accuracy.