Content-Based Video Scene Segmentation

In recent years, along with the development of multimedia technology, the popularity of videos is exploding and even cram our daily life. It is an urgent problem in the domain of computer intelligence that how to manage and retrieve useful information effectively from the universe of information. By the way, the video scene segmentation as a basis for classification and identification of the multimedia information, plays an important role in video data understanding. The traditional method is based on the underlying characteristics of the video for scene segmentation, without considering the content of semantic information, having low accuracy. The video scene senmentation method considering video content builds semantic features to scene segmentation by training relational model between underlying features and semantic information. But the semantic gap is so big that the video scene segmentation technology still has great challenges.In this paper, we analyze the video scene senmentation method based on the video content. And the major research is how to break up a video into a certain fragment. The main contents include:Firstly, we improve the detection methods of video scene boundary. According to the local motion in videos, on the basis of traditional difference description of adjacent frames, the researcher removes grid blocks which are most disturbed by the local motion. And the research studies the adaptive threshold selecting. Meanwhile, the method of video shot clustering is improved by lens sliding window in consideration of time as well as finishing the video scene boundary detection combined with the development model of the shot in scene.Afterwards, an algorithm is proposed to measure the complexity of videos. First of all, the researcher extracts gray uniformity and edge feature based on the background and the distribution of objectives in video frame. Then, six-parameter model is polished up according to the camera actual motion. It estimates background motion model and builds motion vector field of objects for extracting speed and direction entropy feature. The last step is to extract eye visual concentration feature as the auxiliary information to build a new feature space, and training SVR fitter to get the measured value of the complexity for video frame.Meanwhile, an algorithm to measure the excellence of videos is designed. Step by step, the paper extracts shot intension by analyzing the shot cut situation. Then the visual excitement, the type of camera motion and dense optical flow features are picked up to describe respectively the global motion, camera motion and objects’ motion situations and reflect excellence of video. In the end, the local bright differences and color energy build high lightness feature space and train SVR fitter to get the measured value of the high lightness for video shot.Finally, we propose an algorithm to segment video scene on the basis of videos’ complexity and high lightness. Firstly, it is essential to generate the complexity curve and the high lightness curve. And the action is based on a sliding window which detects the troughs in the complexity curve and crest in the high lightness curve. Then, the highlights are located according to the film editing principles and segment. They adapt with the duration of users’ requirement. As the simple video clip, the researcher selects two sides of the trough to cut the video.To evaluate the effectiveness of algorithm, we construct a video database and then evaluates it. The result shows that the method is valid. It can get highlights and simple video clips.