【作者】 白志杰；

【导师】 李弼程；

【作者基本信息】 解放军信息工程大学 ， 军事情报学， 2009， 硕士

【摘要】 新闻视频主题追踪技术可以用来监控各个电视台的新闻报道,检测相同的主题,并追踪事件的发展,对重大事件前因后果和来龙去脉进行陈述。该技术可以提高视频信息的利用价值,满足相关单位和个人的需求,在信息安全、数字图书馆等领域都有广阔的应用前景。本文研究新闻视频主题追踪技术,主要取得了如下3个方面的研究成果:(1)研究了视频结构分析技术,提出了一种基于全局统计的关键帧合成方法。不同于传统的关键帧提取的方法,该方法从合成的角度出发,利用镜头中所有帧在相同位置像素值出现的概率来合成关键帧。实验结果表明,该方法充分考虑了镜头内所有帧的空间和时间信息,与传统的抽取方法相比,能够较好地表示镜头信息。(2)研究了近似重复帧检测技术,提出一种基于BIC(Bayesian Information Criterion)的新闻视频近似重复帧检测方法。该方法先对帧进行关键点检测,提取关键点区域特征组成特征值序列;再使用BIC对两个帧的特征值序列进行判决,来确定是否近似重复。实验结果表明,新方法的召回率和准确率优于传统方法。(3)研究了主题追踪技术,提出了一种基于多模态特征匹配融合的主题追踪方法。该方法利用全局特征和局部特征分别进行帧间相似度匹配,对匹配结果使用线性加权方法进行融合,根据帧的相似度得分进行故事排序,进而进行示例库扩展和故事库扩展,来完成主题追踪。实验结果表明,该方法与单独使用全局特征或局部特征相比,性能有较大提升。更多还原

【Abstract】 News video topic tracking techniques can be used to scout news reports in every news station, detect the same topic, track the develop of event, present the cause-and-effect generating of event. They can advance the video information value in use and satisfy the demand of some units and individual with a wide application foreground in information security, video library, etal. In this paper, we have studied news video topic tracking techniques and obtained the following three achievements mainly.(1) We have researched keyframe extraction technology and proposed a synthesizing method based upon global statistic. Unlike traditional methods, we obtain a synthesized frame according to the possibilities that pixel values appear at the same position of every frame. Experimental results indicate that this method has sufficiently considered both spacial and temporal information of every frame in the shot and has better shot information representation than conventional methods.(2) We have studied the near duplicate keyframe detection of news video, and a BIC (Bayesian Information Criterion) based method is proposed. Firstly, corner points of the frame are detected and feature vectors are extracted at these points; then whether two frames are near-duplicate or not can be judged by comparing their feature vectors using BIC. Experimental results show that this method, which doesn’t need threshold value and machine learning, has better recall and precision than conventional methods.(3) We have researched topic tracking techniques of news video, and proposed a multimodal feature matching fusion method to track topic. In this method, both low-level features and near-duplicate features are used in keyframes similarity detection. We use a linear weighted fusion method for combining the relevance outcome, get story order from frames similarity score, and then extend query corpus and story corpus to accomplish tracking. Experiment results prove that this method has more powerful tracking ability than use global features or local features alone.更多还原