关键词: 角点, 相关向量机, 稀疏光流, 滚动字幕, 上下文特征

Abstract:

A moving caption detection method based on relevance vector machine (RVM) and the context of moving caption is proposed. Harris corner detector is used to determine caption region of video keyframes, and then the sparse optical flow field is obtained from Horn-Schunck(HS) optical flow algorithm, meanwhile, the motion and static text features is extracted respectively as well. A spatial-temporal context relationship among multiple text frames is described by features cascading. Finally, the relevance vector is learned and a two-class classifier is constructed. Experiments show that the performance of the proposed method is better than the existing four approaches, and supports vector machine-based algorithm.

Key words: corner, relevance vector machine, sparse optical flow, moving caption, context feature