Statistical model-based video segmentation and its application to very low bit-rate video coding

Abstract

With the fast growth of video resources, efficient video classification and management are becoming more and more important. Video partitioning is a key issue in video classification. The video partitioning involves the detection of boundaries between uninterrupted segments (video shots) of scenes. Shot boundaries can be classified into two categories, gradual transition and instantaneous change (called camera break). Detection of a gradual transition is considered to be difficult. Block-based image comparison was proposed to detect shot boundaries. Unfortunately, if the differences of the corresponding blocks in the images are measured by gray levels, the method will make false alarms when the gray level change suddenly due to reasons other than shot shifts such as illumination variation which is common in news video. What is more, the step-variable algorithm can not distinguish wipe and dissolve. It is likely that step-variable algorithm will make false alarms of gradual transitions. In this paper, the proposed algorithm named MVS (Model-based Video Segmentation) can distinguish illumination variation from camera breaks as well as wipe from dissolve. Moreover, the positions of the gradual transition are located correctly. MVS is especially efficient in detecting wipes. Experimental results are reported in the paper to validate the proposed method.