| In the inforamtion era, news video, which is one of the most representative multimedia information source, will emerge on a large scale and play an important role in the future battlefield. Thus it becomes urgent to extract and manage valuable intelligence from large quantities of news video data sets in the field of information engineering. The traditional data mining technique does well in knowledge discovering from structure and semistructure data sets, but it is difficult to apply the relevant techniques to the unstructured multimedia data directly. The goal of this thesis is to find a possible way to solve the above problems by probing into the technology of news video data mining.Firstly, the thesis proposes a news video mining architecture. Secondly, the thesis discusses the related techniques of structure mining, semantic content mining and news video summarization. Finally, it validates the effects of the proposed mining architecture and methods by designing and implementing a news video mining system supporting intelligence analysis. The original contributions of this thesis include the following:A novel news video mining architecture, which includes conceptional architecture and technical architecture, is proposed. In the conceptional architecture, the thesis defines the conceptions and tasks of news video mining, and presents a hierarchical framework including two hierarchies called low-level mining and high-level mining respectively. In the technical architecture, the thesis proposes a technical framework and some techniques related. It also discusses the applications of news video mining in the field of intelligence analysis.In order to extract the hierarchical and the narrative structure of news videos, the thesis puts forwards some low-level news video structure mining methods, including the shot detection method based on background tracking, the anchorperson shot detection method called AnchorClu and the news story detection method called SATS, which fuses multi-features.By analyzing two typical semantic objects, namely headings and faces, the thesis proposes some low-level news video semantic mining methods. Such as the method of extracting texts from news videos, which includes the process of heading event detection, heading area detection and text recognition, etc. It also improves the methods of face detection based on skin color and netural network, and face labeling based on multi-feature fusion.Considering the lack of a unified generation model of news video summarization, the thesis proposes an extensible model called EDU, which is the abbreviation of Entity-Description-Utility. The proposed model includes the process of entity extraction, description generation, and utility computation, etc. Based on this model, methods of news story summarization and news topic summarization are proposed. And two novel visualization methods called time-tendency graph and time-space distribution graph are designed.To provide convenient multimedia intelligence services in the strategic decision simulation environment and to validate the effects of news video mining methods, the thesis designs and implements a news video mining system supporting intellignce analysis. As a result, a feasible way is presented to solve the problems of multimedia intelligence analysis.In a word, the thesis proposes the conceptions, architecture and methods of news video mining, and implements the mining process from low-level to high-level by designing the news video mining system supporting intelligence analysis. The achievements of this thesis have great theoretic and realistic significance in deepening the technology of multimedia content analysis and multimedia data mining, and also promote the development of multimedia intelligence analysis. |
