Research And Application Of Key Technology On LIRE-based Video Retrieval

With the application and development of the Internet, the number of videos on the Internet is growing exponentially, and the demand of video retrieval is also increasing rapidly. Retrieving video data effectively to obtain useful information has become an important subject in the field of computer science. As a result, the Content-based Video Retrieval(CBVR) emerged. Recently, the research of CBVR technology has made great achievements in many aspects. However, there are still some problems remaining unresolved as follows:(1)The precision of shot boundary detection needs to be further improved. Though there are some algorithms for shot boundary detection, the precision of those algorithms still can be improved.(2)The process of key frame extraction can easily extract unnecessary key frames, and its efficiency is also not satisfactory.(3)The problem of selecting appropriate features for better performance remains unsolved.To solve the problems above, some improvement methods have been proposed in this paper. The main contributions of this paper can be summarized as follows:(1) An algorithm to identify the video shot boundaries is improved by using Text Tiling method. Segmenting the video through a sliding window, each frame is projected onto a low-dimensional feature space with Principal Components Analysis(PCA) and the distance between two consecutive frames is computed in the new space. And then shot boundaries are detected by measuring the depth score between adjacent windows. With the TREC-2001 video as testing data, experiments on shot boundary detection show that the proposed algorithm has reached the average recall of 89% and the average precision of 96.5%.(2)An algorithm to improve the algorithm of key frame extraction based on content analysis is proposed by calculating the distance between consecutive frames in each sliding window amid the shot boundary detection process, and avoid the problem of selecting too many unnecessary key frames suffered by the original algorithm, as reduces the time consumption and improves the efficiency of key frame extraction.(3)The Borda voting method is used for multi-feature search result fusion that has a much more stable performance than the methods using one single feature.(4)A prototype system of CBVR based on LIRE image retrieval library is established. In the system the aforementioned algorithms for feature extraction, index creation, retrieval with different features and multi-feature search results fusion is implemented.