Content-based Video Retrieval Research On Shot Detection And Key Frame Extraction

Content-based video retrieval (CBVR) has become the focus of the field of multi-mediaresearch in recent years. The core technologies of the research in this area is shot detectionand key frame extraction. It deals with the issues of high-tech in many fields such as imageprocessing, pattern recognition, machine vision and artificial intelligence. It plays animportant practical significance for representation, organization, management and store ofmulti-media data.This paper describes the basic theories and key technologies of shot detection and keyframe extraction in CBVR. It focuses on the technique of key frame extraction based onclustering.In the shot detection, first, an overview and summary was provided for major currentshot detection methods, and analyzed their advantages and disadvantages. Then a shotdetection algorithm based on double threshold was used according to the characteristics ofshot change and the advantages and disadvantages of existing algorithms. The algorithmbehaves as the different threshold respective on abrupt change and gradual change. So it maybe employed to detect abrupt change and gradual change effectively, and reduce the impact ofmovement and noise. The experimental results show that the algorithm is simple andeffective.In the key frame extraction, the key frame extraction based on unsupervised clusteringhas been study on at first. Compared to other clustering as K-mean clustering, the advantageof unsupervised clustering shows that it not needs to set the center and number of cluster inadvance. But the shortage is that the threshold of cluster must be set manually. In view of theno-automatic setting, this paper presents an improved algorithm for adaptive threshold whichcan be implemented easily. It can not only automatically set different thresholds by different types of video, but also chose the number of key frames adaptively. Experimental results of alot of videos with different types demonstrate the good performance of the proposedalgorithm.