A Semantic Content Mining Algorithm For Video Summarization

With the development of multimedia and video database, the technology of content-based video retrieval is receiving increasing attention to structurally analyzing and managing the non-structural video data. Nowadays, video summarization, as an important content-based video retrieval technology and retrieval tool, has been researched by lots of people.Firstly, we introduce the definition, classification and application of video summarization. A video summarization is a sequence of moving images, extracted from a longer video, much shorter than the original, and preserving the essential information of the original. According to the purpose of a video summarization, it can be defined in structure and content and it can be divided into different types.Secondly, we review the techniques concerning key frame-based video summarization. A hierarchical structure of video streams is proposed to effectively build a video summarization. This video organization schema arranges video into four layers: video event, scene, shot and image frame. After the video is segment into video shots or clips, extracting the salient images (key frames) from each shot or clip is very important to get a video summarization. Depending on the content complexity of the shot, one or more key frames can be extracted from a single shot by use of shot boundaries, visually perceptual features, feature spaces, and clusters methods. The selected representative frames with minimal visual content redundancy are accepted in video summarization.Thirdly, in order to hierarchically organize the video, we test the effect of preprocessing methods by video boundary detecting, scene segmenting, and support vector clustering. After video is transformed into a relational dataset of key frame classes, we subsequently introduce video mining strategies to explore special patterns from large video sets. While an association-based video summarization scheme that mines sequential associations from video data for summary creation has received increasing attention, little work has been devoted to the problem of semantic contents in huge video databases which raises other interesting difficulties. A good video summarization may describe what is important in a video, but rather what distinguishes this video-sectionfrom the others.Therefore, in the last part of this thesis, we propose a solution for a new research area of video mining, and develop a novel video summarization approach that mines subject key frames by an algorithm based on vector space model. We use TF/IDF formula to construct knowledge-based key frames to support efficient video summarization. Using subject key frames detection algorithm, we keep the pertinent key frames that distinguish one scene to others and remove the visual-content redundancy from video content. The corresponding summary is finally obtained by assembling them by their original temporal order. Experiments are conducted to evaluate the effectiveness of our proposed approach with summary compression ratio and content coverage. The experimental results have verified that meaningful news video summaries will be generated.