Abstract: For bridging semantic gap between low-level features and user’s information needs in the semantic-based video retrieval system, this paper proposes a temporal-probabilistic hypergraph model which fuses temporal factor in the process of construction, and then introduces a video multi-semantic annotation framework based on temporal-probabilistic hypergraph model, which is named TPH-VMLAF. The TPH-VMLAF combines the temporal consistency property of video data, and annotates video segments with semantic keywords via temporal-probabilistic hypergraph based multi-label semi-supervised learning approach. The proposed framework addresses two important issues simultaneously in the process of video annotation, the insufficiency of labeled videos and the multiple labeling issues. Experimental results indicate that this method improves the precision of annotation and shows good performance.

Key words: video annotation, multi-semantic annotation, temporal-probabilistic hypergraph, temporal consistency, semi-supervised learning

摘要： 在基于语义的视频检索系统中，为了弥补视频底层特征与高层用户需求之间的差异，提出了时序概率超图模型。它将时间序列因素融入到模型的构建中，在此基础上提出了一种基于时序概率超图模型的视频多语义标注框架（TPH-VMLAF）。该框架结合视频时间相关性，通过使用基于时序概率超图的镜头多标签半监督分类学习算法对视频镜头进行多语义标注。标注过程中同时解决了已标注视频数据不足和多语义标注的问题。实验结果表明，该框架提高了标注的精确度，表现出了良好的性能。

关键词: 视频标注, 多语义标注, 时序概率超图, 时间相关性, 半监督学习