| Summary: | 碩士 === 逢甲大學 === 電機工程所 === 91 === In order to support the new functionalities of the multimedia application, the development of techniques for fast and efficient analysis of video streams is essential. Partitioning a video sequence into shots is the first step toward video structure parsing and contend-based video indexing and retrieval. Given that video is often stored efficiently in compressed domain, the costly overhead of decompression can be reduced by analyzing the compressed data directly. In this research, we propose an object-based algorithm for detecting shot-boundary that is directly applicable to MPEG-2 compressed domain.
Inspired by object-based video compressed standard, we segment video frame into two objects: foreground object and background object. The foreground object that is influenced either by fast global motion or large scaled object motion is unreliable for shot-boundary detection. Therefore, the proposed algorithm first extract background region and then detect video cuts, which include hard cuts, fades and dissolves by utilizing the compressed domain data in these regions. To detect hard cuts in video, we evaluate the DC block difference and further calculate the cumulative DC block difference for the determination of the candidates of gradual transitions: fades and dissolves. By observing the temporal curve of DC value around edge, we find that dissolves always occur in the down-concave regions and fades happen in down hill to zero or upper hill from zero. We will propose some methods to detect these effects and further study the usefulness of other compressed domain data to support the detection of shots.
|
|---|
