An efficient scene break detection based on linear prediction

Detection of edits in a video sequence is the first step for video analysis, which segments a video into its basic components. In this paper, we propose a novel and efficient approach to scene break detection, which can detect cuts, dissolves, and fades reliably. Our approach is based on the temporal linear prediction of the frames. For the frames in a video shot, a frame can be predicted from its previous frames. If the prediction error is high, a cut should happen. For dissolves and fades, the prediction coefficients are constants, and these kinds of gradual transitions can be detected by comparing the prediction errors of two different linear predictions. Experimental results show that our algorithm can achieve high precision even if a video contains object motion and camera motion, and is able to detect and classify cuts, dissolves and fades in real time.