Study On The Video Content Organization And Indexing

With the increasing amount of digital video data, the demands of retrieving interested video segments from video data sets has lead to research and development in the area of content based video retrieval (CBVR). In CBVR, the retrieval objects are not original video data, but the data which describe the content of the video data. The MPEG-7, which formally called "Multimedia Content Description Interface", specifies a standard way of describing various types of audiovisual information, and is of benefit to the transmission and exchange of the content describing information. Because of the complexity and variety of video data, there are many kinds of textual information and high dimensional feature information to describe the content of video data. How to organize and manage the describing information of video data for retrieval application has become an attractive topic of CBVR. So, a general organization model for video content describing information and an indexing structure for multi-features are presented.To satisfy the demand of video retrieval, a general video organization model which constructed by four layers of content unit, namely video, episode, scene and shot layer, based on MPEG-7 standard is proposed. In the organization model, the video and episode layer only include textual information, and low level features only describe the scene and shot layer. A retrieval object model which is based on the video organization model is used to describe video retrieval objects. The textual information and low level feature information is separated indexing in general video retrieval framework for retriebal by keywords application and retrieval by sample application.In the framework of generate and manage video content, content describing data are stored in video describing database which is a XML-enabled database. The existed MPEG-7 document is parsed, information which used to retrieval, including structure information, feature information and semantic information are stored in relation table, and other information are stored as XML segment. To obtain sufficient information for retrieval application, feature extracting and structure analyse are performed to original video data.In the application of video retrieval by sample, video data is described by multiple features, and the weights of these features are changed in different queries. An indexing structure called multi-feature index tree (MFI-Tree) and an aggressive decided distance for kNN search algorithm (ADD-kNN) are proposed. MFI-Tree employs tree structure which is benefit for browsing application, and indexing multiple high-dimensional features based on a uniform similarity distance function. Division algorithm is important for the building and updating of MFI-Tree. Here, a new division algorithm which obtains several separately subsets and vitural object which is used to represent the center point of the set are employed. To support K Nearest neighbors (KNN) queries, ADD-kNN directly search cluster nodes in the last level of MFI-Tree to reduce the high-dimensional effect. And, ADD-kNN is an efficient filter-and-refine approach which fast decreases the filtering value to avoid accessing data regions without objects belonging to the result-set. Experimental results demonstrate that the MFI-Tree and ADD-kNN algorithm has the advantage over sequential scan in performance.The AnyVideoStdio shows some example for MPEG-7 document creation using feature extracting and structure analyse processing. To help users obtain sample which describe the need of users easily, a video retrieve interface based on image sample is mentioned. In the interface, an image edit processing, including cuting and stitching, is used to create sample image. The interface is benefit for users to create useful sample, but increases the difficulty of operation.There are many research issues exist in the CBVR. There are many kinds of low lever features for video content description, and these features are high-dimension mostly. So how to find effective features for video retrieval application is an important area in CBVR. As for video structure analyse, shot detection arithmetics are ripe on the base of existing research, but because of semantic gap, scene clusting arithmetics which are related semantic are used to domain video based on domain knowledge. As for video indexing structures which are not satisfy the demand of video retrieval engine, how to reduce the dimension of feature effectively to improve the efficiency of retrieval is an important research issue in the future. Furthermore, video retrieval interface which help users to describe the retrieval demand should be further expanded.