| Along with the development of technologies such as computer processing performance, digital devices, multimedia and database, and the improvement of internet infrastructure, mass of video information are produced rapidly. Therefore, facing ocean volume of video information, how to classify and organize these contents effectively to realize quick searching has become an urgent problem.For the moment, video retrieval technologies could be divided in three categories: the first is based on meta information of video depictions; the second relies on content and images analyzing called low-level feature retrieval; the third on the basis of semantic analyzing. Nevertheless, none of these methods could step over semantic gaps and extract effective video semantic features. In the circumstance of no evident breakthrough machine learning, the most practical video indexes are also based on manual annotation from the commercial aspect. As the creation of video description interface—MPEG-7, depicting video meta information, low feature and advanced semantic feature with a structured form has become possible.In order to realize the mass video search supporting semantic, this dissertation designs and realizes an MPEG-7 based and distributed video retrieval system—Xunet. Main research contents are as following:1) The system develop an improved video semantic processing tool which could extract videos' low-level feature automatically, complete automatic or semi-automatic text annotation and manual semantic graph annotation, then generate MPEG-7 description files containing semantic strings.2) We propose a semantic graph retrieval mechanism. It utilizes semantic graph generated from manual annotation to describe video contents. Subsequently, semantic graph could be serialized to level-0, level-1 and level-2 semantic strings in the form of keywords which built as inverted indexes. On the other hand, we use semantic graph to express users' query intentions and multi-level semantic strings in forms of keywords as level-0, level-1, level-2 are matched with query semantic strings to realize semantic search.3) We propose a Chinese natural language retrieval mechanism. It processes and analyzes text annotation information through Chinese natural language parsing to generate semantic graphs and establish inverted indexes with semantic graph. From another intention, users' query intention could be represented by Chinese natural language and be proceeded and transformed to semantic graphs. Thus query intention could be matched with indexed information through natural language parsing.4) We propose a semantic expansion mechanism. During the query procedure, both synonyms' and hypernyms' expansions are adopted when parsing query keywords set.5) We reconstruct Lucene's rating formula to adapt distributed indexes establishment.At present, Xunet has been developed, the system provides various web query interfaces, including keywords semantic expansion query, semantic graph query and natural language query. Users could find out semantic similar videos and segments promptly. Results browsing and video on demand could be served as well. Owing to adopting a distributed architecture, the system is highly scalable and supports mass video information index and retrieval. |
