Research On The Method For Video Scene Segmentation

With the rapid development of multimedia technology and Internet, more and more video materials could be reached. In video processing and analysis, the hot issues are how to save and manage videos effectively, pick their segments the user needs quickly and spend the least time in understanding their main contents, etc.Video scene segmentation is an important and hot technology in video analysis. It uses the shot as its research object and clusters those shots that have related contents and are adjacent in time into the same scene. Therefore, a complete video can be segmented into a certain number of logical story units that have semantic meaning. Firstly, after summarizing and analyzing the definitions of scene in the existing researches, a more practical definition of scene and a method about how to segment a video into scenes manually are obtained. Next, because traditional methods for shot boundary detection use neighboring frame difference to detect abrupt changes and non-neighboring frame difference to detect gradual changes, which is very complex and too time-consuming, a new method for shot segmentation is proposed, which can obtain the shot boundaries directly by increasing the difference between frame difference at shot boundary and at non-boundary. Moreover, it can effectively eliminate the impact of flashlights.Then, because traditional methods for shot feature extraction usually extract shot key frames to describe the shot, which doesn't take temporal content of the shot into consideration, a method is proposed that temporal content of the shot is obtained by tracking dominant colors and its space information is expressed by mass-center moving. Dominant colors and mass-center are combined to measure the shot similarity. Finally, in order to decrease the time spending in segmentation and overcome the impact of over-segmentation, an improved algorithm of bi-direction sliding window for scene segmentation is proposed. The first scene segmentation results are obtained by linking similar and contiguous shots based on forward and backward sliding the window. And scene segmentation is finished by processing those scenes that contains only one or two shots. The experimental results show that the proposed method is feasible and effective.