Research On Image Mathching Algorithm Based On Feature Point Extracting And Point Set Matching

With the rapid development of artificial intelligence technology,computer vision science has also gained new opportunities and extensive development.As the basis of computer vision and many artificial intelligence applications,image matching technology has attracted the attention of researchers and companies both at home and abroad.Image matching has been developed for many years and various image matching algorithms have been proposed,including the feature extraction algorithm and feature description method under the whole matching framework.This paper summarizes and introduces this field deeply and compares the features and applicabilities of different typical algorithms through experiments.However,the problem of mismatching usually exists in image matching.Searching for the corresponding matching relationship of reliable features is an important and challenging task in computer vision.Matching problems are inherently combinatorial in nature,making the matching space that contains all possible matches larger.Even if you do not consider the outlier,a simple N point to N point matching problem will bring N!permutation.The common approach to solve this problem is to reduce the possible matching set by constructing a hypothesis correspondence group by introducing similarity constraints,which can only match points with similar descriptors.In this way,the matching problem comes down to determining the correctness of each match under the hypothetical set.In the past few decades,researchers have proposed a large number of robust estimation methods to solve the mismatching problem.However,in the face of many practical problems,it is still very challenging to design an effective algorithm to adapt to the problem.First of all,applying only local descriptors inevitably leads to a large number of false matches,especially when the images to be matched have low quality,occlusion and repetitive structures.Secondly,the theoretical transformation model between two images in the real world is very diverse,so designing a common algorithm is very difficult.Finally,the higher computational cost,especially in the face of non-rigid changes,limits its application to real tasks,especially real-time tasks.Based on these analyzes,this paper designs a fast and effective algorithm for mismatching rejection.For a pair or a group of images of the same scene or object,the jagged distance between the two sets of feature points may change significantly with changes of viewing angles or non-rigidities.However,spatial neighborhood relationships between feature points that represent the topological structure of the image scene are often preserved due to some physical constraints.Based on this fact,we can establish a mathematical model that makes the internal correspondence have similar local neighborhood structure information.This modeling approach has the generalized nature of dealing with problems related to rigid and non-rigid transformations between two images.We derive the closed solution of the model and find that it has the time complexity of linear logarithm.Based on the theoretical modeling and complexity analysis,this paper designs several validity verification experiments.Experiments on the Mikolajczyk dataset and other complex real images show that the proposed method can handle the large number of mismatching rejection problem under most kinds of image transformations,simultaneously with fast running speed.It can identify singular points(outliers)in more than 1000 hypothetical matches in a few milliseconds.Therefore,this method can provide a quick initialization for many complex and specific matching problems,it can also be applied directly to some matching problems,meeting real-time application needs.