Research Of Key Technologies On Video Retrieval Based On Bof Model

The amount of traditional broadcast video and newborn online video has grown exponentially in recent years. It becomes a hot research area of multimedia technology on how to retrieve the video which meet the user needs efficiently and effectively. There are two traditional video retrieval methods including the text-based and the content-based retrieval, both of which are unable to retrieve the video shots in a semantic way. With the emergence of content-based image retrieval technology and the development of machine learning, semantic-based video retrieval has seen promising progress in recent years.The thesis conducts a formal study of the semantic-based video retrieval, mainly focuses on shot boundary detection and bag-of-features model and implements a prototype system based on bag-of-features. The prototype first segments the video into shots those are easy to browse and index using shot boundary detection. After that, the prototype uses the key frames extracted from shots to construct visual vocabularies. Finally, at the retrieval stage, the prototype utilizes the invert file and weighting scheme to improve the retrieval speed and accuracy. The work and innovation of the thesis can be listed as following:(1) The thesis proposes a novel shot boundary detection method based on sub-block weighted histogram and adaptive threshold which shows good segmentation performance for shot transition types. Besides, the prototype can retrieve shots by using shot boundary detection.(2) The prototype constructs visual vocabularies through different feature sampling strategies. Result shows that dense sampling features outperform affine covariant features when the video is mainly constructed of scene-oriented frames.(3) The retrieved shots are displayed in a HTML5webpage which can be played on demand real-time, which provides a new solution for web-based video shot retrieval.