Image Semantic Correspondence Algorithm Based On Feature Matching Enhancement

In the era of big data,there are more and more image data including the same object in multi-sensor,multi-mode,multi-view and multi-time points.How to match them to make use of complementary information has always been the focus of research.In the task of image semantic correspondence,there is a greater challenge that the images that need to be matched contain different backgrounds and different instances of the same object.Because there are great differences between different instances of paired images,cluttered background and occlusion in the semantic correspondence task,the extracted features may not pay attention to the objects that need to be matched,which leads to the low quality of similar matching between features.therefore,how to improve the quality of related images is worth studying.Based on this,the following researches have been carried out in this thesis:In the first part,a threshold segmentation and data enhancement algorithm based on histogram is proposed,which improves the quality of feature matching correlation graph.Firstly,by analyzing the histogram distribution of correlation graph,the decision boundary for reasonably dividing similar values is selected.Secondly,different data enhancement functions are designed for different decision intervals,so as to expand the focus of semantic flow in the input displacement estimation stage.To verify the effectiveness of the above methods,based on SCOT(Semantic Correspondence as an Optimal Transport Problem)model,Taking the percentage of key points PCK(Percentage of Correct Keypoints)as the quantitative evaluation index,it is verified on PF-PASCAL and PF-WILLOW data sets.The experimental results show that the proposed data enhancement algorithm improves the overall accuracy.In the second part,we propose to use spatial autocorrelation to cluster the correlation graph,and take the local spatial autocorrelation test value as the threshold division criterion,and then carry out data enhancement to improve the quality of the correlation graph.Firstly,the spatial weight matrix is constructed,then the local spatial autocorrelation is calculated,the spatial cold hot spot analysis of the correlation graph is carried out,then the threshold is divided according to the test value in the cold hot spot analysis,and the data of the related graph is enhanced based on the first part.Taking the PF-PASCAL data set as the experimental data,the experimental results show that the accuracy is improved,which confirms the stability and effectiveness of considering spatial correlation.