Research On Video Scene Segmentation Algorithm

With the rapid development of Internet technology, the video information isgrowing exponentially. How to effectively classify and manage massive video data,how to quickly retrieve the video information of interest to users, has become a realproblem urgently to be solved in today’s society. As is known, the video data is anunstructured stream media, and contains a wealth of semantic information, so it isnecessary for video structured handing. Video scene segmentation is not only one ofthe key technologies of video structured, but also the important bridge across the“semantic gap” to achieve the semantic level retrieval.In the analysis of the video content, the video data is divided into four levelsaccording to the particle size of the video content, wherein the scene layer is usedas the object of study in this thesis. Video data is unstructured in form, but there is astrong logical relationship in content. Therefore, it can be divided into a hierarchyof video content in order to solve unstructured initial of the video data. In addition,the scene is as the smallest semantic unit of the video retrieval in this thesis, so notonly can significantly reduce the scope of video retrieval, and the search results aremore accord with the people’s thinking habits.In the video data preprocessing, this thesis uses multi-modality theory toextract the underlying feature of video data. Then, it uses SimFusion algorithm tomeasure the similarity relationship between the video shot, and LPP method toreduce the dimension of high-dimensional vectors in this thesis. In view of themulti-modality selection, integration and collaboration play an important role in getthe high-level semantics, this thesis select a variety of images, audio and textfeatures to be fused in order to fully mining the TAC character between thedifferent modality. Obviously, if the high-dimensional data is used directly as inputsamples for training and classification, it will result in “curse of dimensionality”.So, LPP is adopted in this thesis, which purpose is to map the high-dimensionalfeature vectors to low-dimensional semantic subspace. This will effectivelycompress the video data and reduce the amount of data storage space.In the video semantic concept detection, this thesis uses the SVMclassification ideas to build semantic extraction model. SVM can not only solvenonlinear problems, but also has good generalization ability. Without priorknowledge of the sample distribution, it can take advantage of the limited sample toclassify. So, this thesis uses SVM to detect video semantic concept, selection ofradial basis function to calculate the optimal classification surface, selection of thehigh-level semantics defined by TRECVID as the semantic concepts of sample set.Then, this thesis uses the LIBSVM software tools to construct a number of semanticclassifiers, and extracts the semantic concept vectors of test set. In the video scene segmentation algorithm based on semantic concepts, thefirst frame and last frame of video shot are regard as the key frames of the shot.And, different scene set are divided by overlapped shot linked algorithm in th isthesis. From the experimental results and analysis, the video scene segmentationalgorithm proposed in this thesis can detect the different semantic concepts in videoshot, and the detection results of different semantic concept is not the same,depending on the complexity and interaction of semantic concepts. In addition, theproposed algorithm recall, precision and M values are higher than the STGimproved algorithm proposed in the literature [15], wherein the maximum value ofM can reach83.4%, mainly due to the proposed algorithm fully consider theinherent character of multi-modality video data.