Shot boundary detection using correlation based spectral residual saliency map

Information present in videos are used by variety of applications like surveillance, intelligent business analysis, high tech education etc. To harness this vital information segmenting the video to its basic architectural unit is required. This paper proposes an approach for shot identification based on spectral residual. The spectral residual of a video frame is obtained by analysing the log spectrum of the frame. The ability of spectral residual to represent the innovative or novelty part of the frame and to provide consolidated representation of the scene makes it promising for video segmentation. The spatial domain representation of spectral residual gives the saliency map. At the shot boundaries there will be variations in the innovative regions of the adjacent frames. Using the correlation between the saliency map of adjacent frames shot boundaries are detected. Experiments conducted on videos from TRECVID 2001 and TRECVID 2007 test dataset show that the proposed approach is simple, fast, reliable and robust for shot boundary detection.