Video Shot Segmentation And The Application In Video Retrieval

With the development of multimedia technology and Web technology,more and more information is stored,transmitted and expressed by the form of video.How to retrieve useful or interesting videos quickly and easily becomes a research hotspot in recent years.Content-based video analysis and retrieval system is an effective way to achieve this goal.In accordance with the top-down structure,video can be divided into four levels of video,scene,shot and frame by the use of video structure technology.In these four levels,shot is considered as the most suitable level for content-based video retrieval.So the content-based video analysis and retrieval system is divided into three steps generally.First,shot boundary is detected to segment video into shots;second,key frames are extracted for the simple expression of content of the shot;finally,the video information is tagged,that is,using dynamic and static characteristics to classify and mark the shots for retrieving easily.The research of this paper is about two aspects.One side is shot boundary detection,dividing video into shots;another side is shot classification based content.For shot boundary detection,this paper presents a neural networks based approach of shot boundary detection using multi-video features.Two approaches,based on feature differences between two adjacent frames and shifting window respectively,were employed to detect abrupt transition,and motion information was used to reduce the influence of strong movement of objects.The fusion and voting techniques were exploited in the final decision stage.In gradual change detection,three patterns of the variance curve of intensity during the period of dissolves were distinguished using three neural networks,respectively.Then the interference was eliminated according to the characteristics of linear increasing or decreasing of the mean value of intensity during dissolve interval.The experimental results carried out on TRECVID database indicate that the proposed approach works well in detecting shot boundary measured by both recall and precision,and it is also robust to motion and flash light.From the perspective of semantic,shot can be divided into three levels as type, event and object.Correspondingly,shot classification can be divided into type classification,event classification and object classification.In this paper,we classify shots in the level of type in connection with six types of shots which are shots of news, landscape,fighting,basketball games,football games,and Ping-Pong games.By comparing the differences of low-level features from these six types of shots,we extract motion information,color,and pixel-differences as features for our experiment.Then, these features were fused as a set of input property,and the types of the shot as an output property,to train C4.5 decision tree.Through the training of C4.5 decision tree,a bridge was build to eliminate the gap from low-level features to high-level semantic,and a good performance of shot classification was achieved.