一种基于二分图最优匹配的镜头检索方法

Shot retrieval plays a critical role in content-based video retrieval. Motivated by the theory of optimal matching in bipartite graph, we propose a novel approach based on the Kuhn-Munkres algorithm for shot retrieval. In contrast to existing algorithms, the proposed approach emphasizes one-to-one mapping among frames between two shots for effective similarity measure. A weighted bipartite graph is constructed to model the similarity between two shots: every vertex in a bipartite graph represents one frame in a shot, and the weight of every edge represents the similarity value for a pair of frames between two shots. Then Kuhn-Munkres algorithm is employed to compute the maximum weight of a constructed bipartite graph as the similarity value between two shots by guaranteeing the one-to-one mapping among frames. To improve the speed efficiency, we also propose two improved algorithms. Experimental results indicate that the proposed approach achieves superior performance than some existing methods.