RELEVANCE FEEDBACK OF VIDEO RETRIEVAL

Content based video retrieval is an important research direction attracting many researchers' interest. The commonly used retrieval method is content based video retrieval by example video clip. But the problem of how to define whether two videos are similar is still an obstacle to further application. And furthermore, since different persons may pay attention to different aspects of the same video for its complex content, adopting feedback advice into similarity model computation in order to reflect users' emphasis and refine the query results is important. Integrating all the visual judgement factors used in real life by ordinary people, a video similarity model is introduced to support the similarity measurement between videos from multi levels(such as shot and video level) and manifold angles(such as time spanning, order, continuity and so on). And on the basis of video similarity model, a novel method is proposed to process video retrieval feedback from various granularity, such as shot and video level, by adjusting all the factors used in computation. This progress is dynamically done. In this way, the retrieval result can be optimized flexibly according to different user's need.