Organization And Retrieval In Large-Scale Video Database

With the Internet popularizing and developing of multimedia technology, people face explosively increasing vision information. Content-based video retrieval becomes very active research area for its wide application foreground. In this paper, we mainly research large-scale video database organization and retrieval so as to enable more effectively storing, organizing large collections of videos and fleetly accurately retrieving video clips from it. The research contents include high dimension feature vectors index, semantic video classification, relevance feedback, similarity measure of video clips and video content representation. We summarized the main works and contributions as follows:1. Video clip retrieval method based on merging key frame sequence: Because there are at least a pair of similar key frames between a pair of similar video clips, we firstly find the candidate video clips,each including at least a key frame similar to example clip's one, then, calculate the similarity between each of candidate clips and the example. As a consequence, the measure of similarity about irrelevant clips is avoided. A joint distribution histogram fusing multi-features is proposed to represent video content, and the joint distribution histogram fusing color and texture features is used to detecting sub-shot. Each sub-shot is represented by a key frame. When retrieving, the similar key frames that are similar to key frames of the example clip are found, then, the similar key frames that are temporally sequential are merged into clips, i.e. candidate video clips. These key frames of above candidate clips are relevantly matched to example's ones by many-to-many relationship. Redundant matches are removed to form optimal key frame match sequences. General similarity measure of clips is calculated taking into visual similarity and temporal similarity account. The proposed method accords with human visual character and has low expense of computing.2. Clustering index method supervised by semantic class and Bayes-based semantic video classification method: A perfect video database organization should be that video feature vectors, which are not only semantic relevant but also their visual features themselves similar, are stored continuously. According to large-scale video database character, we hierarchically cluster feature vectors in video database supervised by video semantic class until every cluster just contains such videos that belong to the same semantic class. The cluster here is called as index cluster. An index entry is created for an index cluster and a Bayes classifier is built with probability relationship between low-level features and semantic class. The Bayes classifiers can also classify other videos semantically. For a given query, semantic class of example video...