关键词: 深度图匹配, 图匹配问题, 计算机视觉, 组合优化, 深度学习

Abstract: Abstract: Deep graph matching models always use graph convolution network to learn the feature of nodes in the stage of node feature extraction. However, graph convolution network has limited learning ability for node features, so it affects the distinguishability of node features which makes a bad result of similarity measure and leads to loss accuracy. To solve this problem, this paper designed a deep graph matching model via self-attention network (GSAN-GM), which used self-attention network (GSAN) based on spatial encoder to learn the features of nodes in the model phase of nodes feature extraction. The principle of GSAN was through a spatial encoder and a self-attention module to learn spatial features of nodes and the relationship between all nodes to enhance the node features. In order to reduce the accuracy loss caused by relaxing graph matching problem, in this paper, the problem was modelled as an integer linear programming problem that considers constraints of structure matching and utilized an efficient combinatorial optimization solver to compute the local optimal solution of graph matching problem. The accuracy of our model was compared with the state-of-the-art methods on Willow Object and Pascal VOC datasets. Experimental results show that GSAN-GM has better performance and achieves the best accuracy on matching tasks of multiple categories of images.

Key words: Keywords: deep graph matching, graph matching problem, computer vision, combinatorial optimization, deep learning