Resampling Consensus Constrainted by Local Structure Similarity

At present, the matching method based on local graph structure has become a popular method for image matching and mismatch elimination and can achieve good results in many scenes. However, in some scenes with numerous repeated structures, the effect of the matching method based on local graph structure will be greatly reduced. Therefore, this paper proposes a method of eliminating mismatching based on local graph structure similarity combined with resampling to obtain the optimal sampling model. The method firstly obtains the initial matching point pairs based on SIFT and KNN, and then obtains K nearest neighbor nodes of each feature point, and then calculates the local nearest neighbor structural similarity score of each pair of feature matching points. Finally, the optimal transformation matrix model is calculated based on some matching points with the highest similarity score, and the mismatching is detected and removed by the optimal transformation matrix. The experimental results on the dataset Hpatches show that compared with the LGSC algorithm, the precision is increased by more than 7% on average, and the F -value is increased by 3% on average.