Research On Video Scene Structure Analysis And Scene Recognition Technique

Digital multimedia resources are increasingly becoming an import part inour daily cultural life. Structural analysis and scene recognition of a large numberof film and television video is the base of the content-based video retrievaltechnique. Scene boundary detection can divide video into scene semantic units,and automatically marked scene tags of video clips can provide video semanticcontent.Because of the computationally intensive problems in scene similarity graphalgorithm for video scene boundary detection, a sliding shot window is used tocluster similar shots satisfying temporal locality, at the same time, scenedevelopment model is used to detect scene boundary by merging staggered lensand similar adjacent shot class, resulting in a video scene structural units.Meanwhile, using only visual feature similarity of video shots is likely to causeover-segmentation problems, especially in scenes with much movement. To solvethis problem, movement information of video shot which uses visual changesbetween frames to measure the motion of the shot is also taken into account inshot similarity calculation, the use of visual similarity and motion information ifform of weighted sum can effectively solve over-segmentation problem in fights,chase scenes.The traditional image and video representation based on the underlyingfeature, often fails to deal with high level semantic tasks due to lack ofsemantic information. Objects contained in the image are very important visualsemantic contents, using Object Bank representation of image can achieve goodresults in the image visual recognition task, such as scene recognition. Thisarticle extends Object Bank approach to video feature representation, usingrecognition result of a set of common object detectors in the video frame torepresent scene of video. Max-pooling and mean-pooling of statistics informationfrom object recognition result in per key frames of video clip is also appliedbetween key frames to represent video feature, then the SVM classifier is usedfor training and recognition of video scenes. Object Bank based video featurerepresentation achieves good result in the street, bedroom, dining room, living room scene recognition experiment.Through technique of video scene boundary detection and scene recognition,the scene structure and scene semantic content of the video can be extracted,which provide the basis for video content retrieval based on scene. Meanwhile,the use of Object Bank video representation contains object semantic informationof video, can at the same time provide support for object-based video contentretrieval.