An efficient graph theoretic approach to video scene clustering

In this paper, a graph theoretic approach is presented for video scene clustering in sports and news videos. In the approach, video shots are grouped into clusters of similar scenes based on shot color attributes. First, the similarity between video shots is measured by shot color histogram intersection. To obtain scene likeness matrix in a maximum-a-posterior (MAP) probability manner, a thresholding method is proposed on shot similarity matrix. Then, a graph is constructed based on the scene likeness that similar shots have an edge between them. Based on the constructed graph, a graph partitioning method is proposed to cluster video shots into different scenes such that the connectivity of video shots within one cluster is higher than that between different clusters. The advantage of the graph partitioning method is that the cluster number need not to be known as a prior. The graph partitioning method is compared with conventional k-means clustering method; in which the cluster number is determined by the cluster validity measure. The main contributions of the paper lie in the formulation of video scene clustering to a graph partitioning problem, and the comparison with the conventional k-means clustering method. Experimental results are presented to show the effectiveness and efficiency of the proposed graph theoretic approach.