视频数据库的聚类索引方法  

Cluster-based Index Method for Video Database

作　　者：施智平[1] 胡宏[1] 李清勇[2] 史俊[1] 史忠植[1] 

机构地区：[1]中国科学院计算技术研究所智能信息处理重点实验室,北京100080 [2]北京交通大学计算机与信息技术学院,北京100044

出　　处：《计算机学报》2007年第3期397-404,共8页Chinese Journal of Computers

基　　金：国家自然科学基金(60435010;90104021;60675010);国家"八六三"高技术研究发展计划项目基金(2006AA01Z128);国家"九七三"重点基础研究发展规划项目基金(2003C6317004);北京市自然科学基金(4052025)资助

摘　　要：理想的视频库组织方法应该把语义相关并且特征相似的视频的特征向量相邻存储.针对大规模视频库的特点,在语义监督下基于低层视觉特征对视频库进行层次聚类划分,当一个聚类中只包含一个语义类别的视频时,为这个聚类建立索引项,每个聚类所包含的原始特征数据在磁盘上连续存储.统计低层特征和高层特征的概率联系,构造Bayes分类器.查询时对用户的查询范例,首先确定最可能的候选聚类,然后在候选聚类范围内查询相似视频片段.实验结果表明,文中的方法不仅提高了检索速度而且提高了检索的语义敏感度.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, the authors hierarchically cluster feature vectors in video database supervised by video semantic classes until every cluster just contains such videos that belong to the same semantic class. The clusters here are called as index clusters. An index entry is created for an index cluster and a Bayes classifier is built with probability relationship between low-level features and the semantic class. For a given query, the first phase computes the distances between the query example and each cluster index and returns the clusters with the smallest distance, here namely candidate clusters; then the second phase retrieves the original feature vectors within the candidate clusters to gain the approximate nearest neighbors. The proposed method speeds up searching and improves retrieval semantic sensitivity.

关 键 词：视频库 高维索引 视频语义分类 聚类 视频检索 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]