摘要： 在图匹配模型中权重的设置对匹配性能有很大影响，但直接计算的权重往往不符合匹配图像的实际情况。为此，参照二次分配问题的图匹配学习思想，给出一阶和二阶最大权对集模型的权重学习计算方法。一阶最大权对集模型直接采用图像特征点作为图的顶点，而二阶最大权对集模型则采用某些特征点之间的连接边作为顶点，2个模型都可以通过Kuhn-Munkras算法求解。一阶最大权对集模型在本质上等价于二次分配问题的线性情况。在CMU House数据库上的图像匹配实验结果表明，二阶最大权对集模型优于一阶最大权对集模型，且两者在学习计算时的性能也优于直接计算的情况。

关键词: 图像匹配, 权重学习, 最大权对集, Kuhn-Munkras算法, Delaunay三角化, 二次分配

Abstract: Weight setting has a great impact on performance of graph matching models. Weights by direct calculation often produce unsatisfactory correspondences between real images. Based on the idea of learning graph matching for quadratic assignment problems, this paper considers weight learning method for first-order and second-order maximum weight matching models. In a first-order maximum weight matching model, image feature points are regarded as vertices of a bipartite graph, whereas in a second-order maximum weight matching model, edges connecting two feature points are viewed as vertices. Both of the first-order and second-order models can be solved by the Kuhn-Munkras algorithm. The first-order maximum weight matching model is essentially equivalent to the linear quadratic assignment problem. Experimental results on the CMU House database show that the second-order maximum weight matching model can totally outperform the first-order maximum weight matching model, and both of them perform better in the case of weight leaning than direct calculation.

Key words: image matching, weight learning, maximum weight matching, Kuhn-Munkras algorithm, Delaunay triangulation, quadratic assignment

中图分类号: 

• TP391.41