Efficient key-frame extraction using density-estimation-based nonparametric clustering

Key-frame extraction has been recognized the important research issue in content-based video analysis. In this paper, an efficient key-frame extraction approach is presented to represent the video content, which provides the capability of browsing digital video sequences more efficiently. Firstly, each video sequence is partitioned into shots by applying a shot-cut detection algorithm. Within each shot, candidate frames are extracted based on maximizing Kullback-Leibler divergence criterion for locating a candidate frame set. Secondly, we construct multidimensional feature vector using color and texture information extracted from the low and high frequency bands of wavelet transforms. Key-frame extraction is accomplished by density-estimation-based nonparametric clustering, the mean shift method, which can efficiently analyze complex multimodal feature space and delineate arbitrarily shaped clusters in it. An experimental system has been build up. Experiments verify the effectiveness of the proposed approach.