Study On Discontinuity And Threshold In Shot Boundary Detection Problem

With digital technology widely used in video, videos could be repeatedlyshot by digital video camera, easily edited through computer software,permanently stored by CD-ROM, rapidly transmitted through the network andvividly projected to the digital screen. Great digitalization makes it happen.However, the explosive growth of the videos (online videos especially),coupled with the video content nature, non-level and non-structural, causesvideo classification, indexing, archiving and searching a tremendous duty. Onthe other hand, YouTube and other video sites have attracted millions ofInternet viewers around the world, It's reported that in every minute about 20-hour long videos are uploaded to the YouTube, of which there is some of sexand bloody scenes. How to identify the illegal content precisely andcompletely from those upload videos has become a heavy burden for thevideo website staff.In this field, content-based video retrieval research has been raised by theacademia. Similar with the method in text content analysis, we first format thevideo according to the content, achieve high-level semantics (content)analysis and extraction and then do video classification, indexing, archivingetc. This paper studies video shot boundary detection which is the most basicwork in video content formation.In this paper, shot boundary detection problem is divided into threeindependent sub-problems: feature extraction, discontinuity definition andthreshold calculation to determine. In each sub-problem, I summarize theprevious research methods, the results achieved and the drawback as well. Tosolve some of the mentioned shortcomings, the following two methods are put forward:1. In the MPEG compressed domain, the method selects the DCcoefficients of I-frame, does sub-block histogram statistics and calculates thediscontinuity by comparison. The shot boundary is determined withχ2 thehelp of adaptive threshold which is based on time factor and motion differencefactor. This method is time-efficient for no video decoding process.2. In the non-compressed domain, the method sets the variable-sizewindow and calculates the discontinuity based on the sub-block histogramdifference between the first frame and the last one of the window. The windowspan is adjusted according to the current discontinuity. The method decodesonly a small number of frames but extract more abundant features.To test two-factor adaptive threshold and sub-block histogram, wechoose three representative videos. Results prove their superiority. In thispaper, two kinds of shot boundary detection algorithm are evaluated by theauthoritative TRECVid 2001 video set. The result verifies that the raisedmethods excel a number of similar algorithms in overall performance.