| Feature point correspondence is the basis of many work in computer vision and image processing,the matching precision has an important influence on the quality of the following work.However,high precision feature point matching is a very difficult task,which has attracted much attention from many researchers.At present,the mainstream methods of feature point correspondence are based on feature descriptors.First,the initial correspondence of feature points is established on the description space by using the nearest neighbor method,and then the wrong matching point pairs are removed by using RANSAC algorithm.When the scene is complex or the texture is similar,the result is not good.To solve this problem,we propose a new feature point correspondence method,which is a hybrid method combining cross validation and multiple homography transformation.First,we detect the SIFT feature points and use brute-force method to get preliminary feature point correspondence.Then we use standard RANSAC algorithm to select a set of corresponding points with high credibility.We train a more credible homography transformation from this set by cross validation.The transformation and the Euclidean distance of descriptor space is applied to filter all the remaining feature points.Point pairs that meet the condition can be added to high credible corresponding point set as supplement.For all the remaining feature points that do not enter the corresponding point set,we repeat the above process until the selected feature points are less.We have done a lot of experiments on the existing test picture library and the picture set taken by ourselves,and the results show that the multi-homography transformation in this paper can effectively improve the matching accuracy and the number of corresponding points.The performance is better than the classical RANSAC matching algorithm.In addition,the comparison with the latest matching algorithm shows that our method has advantages in corresponding accuracy. |
PDF Full Text Request