【作者】 文军；

【导师】 吴玲达；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2008， 博士

【摘要】 新闻报道是信息的重要载体,用户对新闻报道关注的重点是部分特定新闻事件,因此迫切需要能够自动实现基于新闻事件的新闻报道跟踪。目前主要在文本领域开展了新闻报道话题探测与跟踪研究,与文本媒体类型相比,新闻视频面临结构复杂,媒体模态多样等一系列问题,要在不同来源新闻视频中实现新闻事件各个报道内容的跟踪面临很多困难。根据新闻视频结构特点,可以把视频划分为帧、镜头、故事单元、视频四个层次。与新闻事件密切相关的层次是故事单元,因此在新闻视频数据库中研究识别和跟踪报道相同新闻事件故事单元的相关技术成为当前新闻视频研究领域的前沿课题。本文对这个具有重要理论意义和广阔应用前景的课题进行了探索和研究,旨在解决新闻视频故事单元跟踪研究中的部分关键技术,为新闻视频数据库基于新闻事件的信息分析和利用提供可行的解决途径。本文首先建立一个新闻视频故事单元跟踪研究的框架,在此基础上重点研究了故事单元分割、故事单元关联分析、故事单元线程化跟踪等关键技术,通过实验验证了研究的可行性和算法效率。论文的主要贡献体现在以下几个方面:1、提出了新闻视频故事单元跟踪研究的技术框架。首先对研究中涉及的概念和关键术语进行了阐述,然后研究了新闻视频文件和故事单元描述模型,提出了新闻视频数据库的“故事单元空间”表示方式,为开展故事单元跟踪研究提供了理论基础。在此基础上提出了新闻视频故事单元跟踪研究的技术框架,探讨了研究实现的技术途径和部分关键技术,明确了研究的主要任务。2、提出和改进了新闻视频故事单元分割方法。通过对新闻视频故事单元编辑模式的分析,提出了一种有效的视频、音频特征候选分割点选择策略,其中突出研究了自适应的播音员镜头探测方法;同时,研究了不同的集合运算方法来融合分析不同类型的视频特征候选分割点与音频特征候选分割点,对不同来源的新闻视频都可以有效实现故事单元分割。3、提出了新闻视频故事单元关联分析方法。分析了相似关键帧与故事单元关联分析的内在联系及各种领域知识;研究了关联分析子数据库构建策略和局部关键点精减策略,在本质上提高了关键帧匹配分析速度;提出了一种利用局部关键点匹配技术的层次化过滤方法快速有效的识别相似关键帧;提出了基于相似关键帧和关联关系传递性的故事单元关联分析技术。4、提出了新闻视频故事单元“多线程”跟踪方法。为体现新闻事件报道的“多线程”属性,首先提出了一种融合各个语义层次、各种模态信息的故事单元相似度计算方法,方法结合新闻视频和故事单元的描述模型,重点研究了底层视觉特征中的局部特征相似度计算方法、中层语义概念中基于关键帧场景信息的相似度计算方法、高层语义的文本相似度计算方法以及相似度融合方法;在此基础上,研究了图论知识对于故事单元跟踪研究的有效性,提出了利用有向图理论对故事单元之间的相似关系进行“多线程”跟踪的方法。5、设计和实现了一个新闻视频故事单元跟踪系统。详细描述了NStoryThread系统的设计思路和各功能模块,并介绍了原型系统的实现,为研究的应用提供了基础。综上所述,本文的主要研究集中在新闻视频故事单元跟踪系统方法的关键技术上,如:新闻视频故事单元分割、故事单元关联分析和故事单元跟踪等,并对各关键技术进行了实验验证。这些研究不仅对新闻视频的分析和挖掘技术具有积极的影响,同时也对多媒体情报分析技术具有显著的理论和实践意义。更多还原

【Abstract】 The news report is an important information carrier. Users pay attention to the news based on some specific events. Therefore a kind of intelligent services, which can automatically analyzing and tracking the news are in urgent need. The research of event-based news tracking is developed in text. Compared with the report based on text media, news video is faced with some problems, such as complexity of structure and multiplicity of media modals. Tracking reports in news video across different event-based sources is a challenging work.News video can be represented by a hierarchical structure consisting of 4 levels: frame, shot, story and video, in which story is the unit relating news events. Based on news video database, the research on the technique of identifying and tracking the stories which report the same news event is becoming the frontier topic in news video research field. Therefore, this thesis explores the topic on event-based news video story tracking technology, which is a research issue with great significance in theory and wide perspective in application. The goal of this thesis is to find a possible way to solve the problems of analyzing and utilizing information in news video database based on event by probing into the key technologies of tracking news stories.Firstly, the architecture of tracking news stories is proposed in this thesis. Secondly, the related key techniques of story segmenting, story correlation analyzing and story tracking are discussed. Feasibility and effect of these techniques are validated by experiments. The original contributions of this thesis can be described as follows:1. A frame of event-based news video story tracking is proposed. Firstly, relevant concepts and terms are defined. Then, describing modals of news video and story are investigated. And, this thesis proposes a pattern named“story space”for describing news video database. These works provide theoretical basis for tracking news stories. On this basis, this thesis proposes technology frame of tracking news stories, which discusses the approaches and key techniques to realize event-based news story tracking, and points out the problems this thesis concentrates on.2. A news video story segmentation method is proposed and improved. In view of characters of news video’s edition, this thesis presents a novel strategy for selecting video and audio candidate points as segmentation boundaries, in which an adaptive method to detect anchorperson shot is studied prominently. Different set operating methods are developed to fuse diverse modal candidate points and get story boundaries efficiently.3. News video story correlation analyzing method is proposed. This thesis investigates internal relations and domain knowledge between near duplicate keyframes and correlation analysis. To increase matching speed of keyframes essentially, approach for constructing sub-database and pruning local keypoints is studied. Then a hierarchical approach for identifying near duplicate keyframes based on matching local keypoints is proposed. Finally, this thesis presents a method to identify correlations of related stories based on near duplicate keyframes and transitivity of correlations.4. A method for tracking news video stories with“multithreading”is proposed. In order to incarnate“multithreading”of news event reporting, a method for calculating similarity of news stories is presented in detail. It fuses information of all semantic levels and all media modals, in which methods to calculate similarity of local feature in lower visual feature, similarity of keyframe scene class in middle caption, similarity of text in high-level semantic and fusion strategy of these similarities are researched prominently.Then, this thesis studies validity of graph theory for tracking news stories and proposes an approach for tracking the similarity of news stories with“multithreading”based on digraph.5. A system for tracking news stories is designed and implemented. The design idea and each functional module of system NStoryThread are described in detail, and the implementation of prototype system is also presented, which provides a support to the applications of the frame and relevant methods.As a general, the thesis focuses on the key techniques of tracking news video stories, such as story segmentation, story correlation analysis, story tracking and so on, and each method is validated by experiments. The achievements of this thesis promote the development of news video analyzing and data mining, and also have great theoretic and realistic significance in multimedia information analysis.更多还原