Research On Fast Retrieval Of Large Scale Web Video Contents

With the fast development of computer networks and digital multimedia technology, Internet video application becomes popular, the amount of video clips over the Internet is very huge, and how to effectively discover video contents, retrieval video clips and processing video data becomes a challenging task in both research and industry.State-of-art Internet video retrieval systems are based text such as the title and surrounding description text of video. Although it is fast, there are many problems in it, such as the incompletion of indexing, the incorrect of description text, which lead to the poor result in the retrieval result. These systems also ignore the content of videos, which leads to much duplication in the retrieval result. According to these problems, this paper focuses and studies the key issues in Internet video retrieval system, and research on the data mining methods of video contents in order to improve the result.The main contributions are illustrated as follows: 1. We propose a semantic structure based internet video discovery and information method. Based on the text method of video retrieval, we study the discovery method for internet video container WebPages and the video description text extraction algorithm. First, WebPages are covert into semantic structural representation, then we propose a tree based similarity algorithm to identify video container WebPages and extract video description text.2. We propose the model of video content similarity based on SIFT feature and time series feature. First, we propose a frame matching strategy based on SIFT feature, to speed up the matching process, LSH hashing is employed, the matching precision is improved by the feature. Then we use RANSAC to filter the matching sequence and effectively remove the outliers, which combines the time series feature. The method improves the precision of similarity measurement and is effective in near identical video detection and elimination.3. A video search result optimization strategy is proposed based on affinity propagation clustering. There are many duplicates or near-duplicates in the video retrieval result, which affects the quality of results and user experience. In this paper a clustering based on affinity propagation is proposed the group the similar video clips. The method improves the quality of video search result and optimizes the video search system.4. We propose a distributed architecture to compute similarity between video clips in large scale environment. State-of-art video processing methods are only fit for small amount of videos, but not for large-scale environment. According to this problem, we first propose a video feature that can be used in large scale environment, and then fast similarity computation framework based on Map/Reduce distributed computing is also proposed, which improves the performance of computation and is highly extensible.