Research On Point Set Registration Based On Global Structure And Local Structure

With the further development of society and the progress of science and technology,digitalization,intelligence and informationization have become the main development trends.The use of computer is no longer limited to numerical calculation or executing simple commands.Seeing is the ability that humans are born with.Humans can quickly find the information they want in complex scenes and can recognize deformed objects.With the development of artificial intelligence,researchers are trying to give machines the ability to see.This is machine vision technology,and this research field is called computer vision,such as shape recognition,image classification,target detection,and so on.The registration of graphics or images is a fundamental and important issue in the field of computer vision,which plays a crucial part in medical image analysis,three-dimensional surface reconstruction,shape recognition,cultural relics protection,and so on.The goal of registration is to obtain the corresponding relationships between the two graphics or two images to be registered,and to obtain a certain spatial transformation.Because points are easy to extract,easy to represent,and convenient to store information,point-based registration has been widely used.This paper focus on the point set registration.Local structure is very important for point set registration.Many methods achieve point set registration using local structure.In order to make full use of the local structure,this paper further digs into the information contained in the local structure to improve the accuracy of the registration method.Specifically,the following work has been done.In this paper,a matching method combining feature information and spatial location information is designed.The feature information and spatial location information of points are extracted to create a mixed dissimilarity matrix.In order to better describe the feature of points,this paper improves the traditional shape context descriptor(SC).Secondly,the global and local location information of points are extracted to describe the dissimilarities of points.The combination of feature information and spatial location information can more accurately describe the dissimilarity between points.In the matching stage,the objective function is solved and the final binary solution is obtained by using Hungarian method.Tests on three common datasets show that the proposed method is superior to other advanced methods.This paper proposes a registration method based on weighted feature descriptor and local structure constraint.The improved SC descriptor introduced in the previous work has achieved good performance in almost all cases,but there are still some shortcomings.In order to further improve the performance of descriptor,this paper proposes a weighted SC.Weighted SC gives different weight values to points,which are used to represent the contribution of points to the bin.In the registration process,while retaining the global structure constraint,local structure constraint is added to improve the accuracy of the registration.In this paper,experiments on four commonly used datasets and comparisons with other advanced methods prove the effectiveness of the proposed method.This paper proposes a registration method based on neighborhood information support.In the previous work,the initial correspondences are obtained by comparing the feature descriptors.However,due to the existence of noise points and outlier points,the correspondences only by comparing the feature descriptors may be not so accurate.Since the initial correspondences will have a direct impact on the accuracy of the registration,this paper extracts the neighborhood information and uses the support of the neighborhood to calculate the initial correspondences.The neighborhood information contains two parts.The first part contains the matched neighbors by comparing feature descriptors.The second part contains the neighboring point pairs with the most similar spatial relative position.Finally,the algorithm is optimized iteratively in the framework of EM method.In this paper,five general datasets are tested,and the results are compared and analyzed with the existing methods,and the robustness and superiority of the method are proved.