由上下文和相关向量决策的滚动字幕检测算法

Moving Caption Detection Using Context and Relevance Vector Discriminant Analysis

提出了一种基于时空上下文特征和相关向量机的视频滚动字幕检测算法.可检测视频关键帧中的角点,并估计出角点上的稀疏光流;在对光流场优化的基础上,提出一种新的融合静态和动态特性的滚动字幕统计描述方法,进而结合多个关键帧特征建立起滚动字幕的时空上下文联系;引入相关向量机进行决策.实验结果表明,该算法优于现有4种典型方法,综合性能也略好于基于支持向量机的方法.

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.