Video Shot Annotation Based on Hypergraph Random Walk Algorithm

Abstract

We introduced a semi-supervised learning framework based on hyper graph random walk algorithm (HRWA) for content-based video-shot annotation, in which a probabilistic hyper graph is used to represent the relevance relationship among the vertices (video shots). First, based on the similarity matrix computed from attribute values of video shots, a label random walk graph is built on the training set. Second, the random walk processing is carried on a hyper graph system when an unlabeled data arrives and a probability distribution among all labels is obtained. Then, the label category of each shot is determined according to the threshold. Finally, the effectiveness of HRWA is demonstrated by experiments on news video database of TRECVID 2007.