Research On The Optimization Algorithm Of SIFT Feature Point Matching And Its Multi-objective Matching Application

Image matching technology has been widely used in key fields such as remote sensing satellite image processing,medical image diagnosis,and autonomous driving.Image matching algorithms mainly include matching algorithm based on gray correlation and matching algorithm based on features at present.Feature-based image matching algorithms are widely used because of their good stability to image scaling,rotation,and affine transformation.Among them,the SIFT operator of scale invariance is used in many fields because of its advantages such as good discrimination,fast speed,and strong scalability.When using SIFT for multi-target matching applications,the number of detection targets needs to be determined according to the initial matching feature point set,and the initial matching feature point set is clustered using a clustering algorithm to achieve target feature point pair separation,based on each separated Target feature point set,remove the mismatched feature points to achieve target matching,but there are a large number of mismatches in the SIFT matching results,and it is difficult to adaptively determine the number of detection targets based on the initial matching feature point set;so in order to solve the above problem and achieve the practical application of multi-object matching,this article is based on SIFT to research,the main content is as follows:(1)When using the ratio test method for matching,the matching result of SIFT is greatly affected by the threshold.There are a large number of mismatched feature points in the obtained matching points,and the number and accuracy cannot be considered at the same time.The algorithm of SIFT feature point matching optimization is proposed.The algorithm first uses a high threshold ratio test algorithm to remove the initial matching point set coarsely,and then sorts according to the matching accuracy,and selects the N matching feature points with the best matching accuracy as the construction of local coordinates.The basic point set of the system uses the voting mechanism to filter out the three feature point pairs with the best matching from the basic point set,construct the local coordinate system in the two images,and calculate coordinates of the quantized linear primitive in the local coordinate system.through the similarity between the coordinate values to eliminate mismatches to achieve fine rejection;in order to verify the performance of the algorithm,based on the Mikolajczyk datasets,the algorithm is compared with the high threshold(0.8)ratio test+RANSAC algorithm,for rigid transformation,scale transformation,perspective transformation and linear transformation,the results show the algorithm can take into account both the number and accuracy of matching,and obtain a feature point set with high matching degree.(2)In order to realize the application of feature-based matching algorithm in multi-object matching,a multi-object matching algorithm based on adaptive K-means clustering is proposed.The algorithm uses the advantages of the density peak clustering algorithm(DPC)can calculate the number of clusters adaptively,merges it with the K-means clustering algorithm for improving the K-means clustering algorithm,and uses the improved K-means algorithm clusters the initial matching feature point sets to achieve the separation of multiple matching target feature point pairs.Based on the separated feature point sets,the optimization algorithm of SIFT feature point matching is used to extract high matching feature points Set,calculate the transformation model between the set of points,so as to achieve accurate matching of the target.Experimental results show that the algorithm has strong robustness to the matching of a single target,with scale,rotation invariance and strong anti-interference ability;moreover,it can be adaptively determined the number of targets according to the initial matching feature point set when performing multi-target matching,and can be matched matching even if targets are in different states.