SVD of Shot Boundary Detection Based on Accumulative Difference

Shots are the basic unit of video. Shot boundary detection is a pre-task for video semantic analysis and video retrieval. In this paper we propose a new and effective shot boundary detection (SBD) method to detect cut transition (CT) and gradual transition (GT). By introducing the selection mechanism of candidate segment, a large number of non-boundary frames are avoided. We extract the hue-saturation-value (HSV) features of each frame and use singular value decomposition (SVD) to reduce the dimension of the feature space to save computing overhead. In the SBD stage, we propose a scheme based on accumulative distance. By calculating the maximum distance and accumulative distance of the candidate segment, it can be determined whether the shot boundary is included and the type of the boundary. The maximum distance can directly determine the location of the CTs, and the GTs are defined using inverted triangle pattern matching. Experiments on TRECVID 2007 test data and 100 dual-shots movie clips. Compared with the latest solution, the solution has an excellent accuracy and universally applicable for different types of video.