A new general framework for shot boundary detection and key-frame extraction

Abstract

Video shot boundary detection is an important step in many video applications. Since the rapid development of video editing technology, especially, the extensive use of sub-window in news video, the original method of video segmentation cannot efficiently detect the video shot boundary caused by special video technique. In this paper, previous temporal multi-resolution analysis (TMRA) work was extended by first using SVM (Supported Vector Machines) classify the video frames within a sliding window into normal frames, gradual transition frames and CUT frames, then clustering the classified frames into different shot categories. The experimental result on ground truth, which has about 21 hours (10,250 shots) news video clip, shows that the new framework has relatively good accuracy for the detection of shot boundaries. It basically resolves the difficulties of shot boundaries detection caused by sub-window technique in video. The framework also greatly improves accuracy of gradual transitions of shot.