【作者】 韩冰；

【导师】 姬红兵； 高新波；

【作者基本信息】 西安电子科技大学 ， 模式识别与智能系统， 2006， 博士

【摘要】 基于内容的视频检索是近年来国际上在信息检索领域最前沿和最活跃的研究方向之一。由于其在信息检索和数据挖掘领域已经展现出有效而广阔的理论和应用前景,从而备受国内外研究学者和专家的高度关注。本论文针对视频检索中的关键技术――镜头边界检测,以粗糙集、聚类分析等软计算方法为基础,研究并提出了一些实时有效的新方法,利用视频的多条检索线索实现了新闻条目的分割。主要研究工作如下:第一章介绍了本文研究的背景和意义,概述了基于内容的视频检索和镜头边界检测的研究进展及现状,分析了镜头边界检测研究中存在的问题。第二章首先简述了软计算的定义及分类,重点讨论了本文中用到的几种软计算方法以及与粗糙集的融合,分析并比较了其性能。第三章分析了视频特征、镜头过渡类型和粗糙集条件属性、决策属性之间的关系,提出了一种基于粗糙集和模糊聚类的镜头边界检测方法,成功地应用于新闻视频片断中突变、渐变和无场景镜头边界的检测。针对粗糙集在处理连续数据的时需进行数据离散化的缺陷,引入了模糊粗糙集概念,给出了模糊粗糙算子定义。根据视频数据量大的特点,提出了“先训练再检测”的具体检测流程以及算法中的门限选择方法,提高了自动化检测程度。第四章研究了传统粗糙集对于数据准确性要求过于苛严的问题,并分析了模糊c-峰值算法的局限性,从而引入变精度粗糙集,提出了变精度粗糙集特征加权聚类。针对数值型和混合型数据,建立了适用于混合数据聚类且具有特征加权的目标函数,提出了基于变精度模糊粗糙集和变精度粗糙集的特征加权新算法,给出了镜头边界检测的算法流程,并在不同的数据集和视频数据上对算法进行了验证。第五章根据仿生模式识别与传统模式识别的区别,结合仿生模式识别的基本特性,给出了仿生模式识别实现的一个基本方式――高维空间几何理论。分析了多权值神经网络与高维空间闭合超曲面,提出了一种基于仿生模式识别的镜头边界检测方法。该方法利用双权值神经网络来实现,为镜头边界检测方法研究开辟了一条十分有效的途径。第六章将镜头边界检测与其它检索线索相结合应用于视频检索――新闻条目分割中。在分析了多媒体的交叉索引技术以及结合了新闻视频特殊性的基础上,给出了结合音频检测和分类,镜头边界检测、主持人帧的检测和字幕识别的新闻条目联合分割方法,实现了基于内容的新闻检索的自动化,充实了基于内容更多还原

【Abstract】 In recent years, the content-based video retrieval is becoming one of the most advanced and active research topics in the field of the information indexing in the world. Due to the foreground of wide theory and application shown in the fields of information indexing and data mining, the content-based video indexing has been highly regarded by researchers and experts all over the world. This paper focuses on the key technique—shot boundary detection. Based on the soft computation method, such as rough sets, clustering analysis and so on, some effective methods have been proposed, and the segmentation of news items has been implemented according to the several indexing clues. The main research fruits achieved in this thesis are given as follows:Chapter 1 introduces the background and significance of our work firstly. Then, a brief review of the advances and status for the content-based video indexing and shot boundary detection is presented and the research orientation of this paper is point out by analyzing the major defects in shot boundary detection.In Chapter 2, a brief introduction to the definitions and classifications of soft computations is made, with the emphasis on several soft computation methods used in this thesis. Then, the method combining with rough sets used in this dissertation to other mathematical tools or soft computation methods are discuss in detail with the comparison of their performances.Analyzing the relationship between video features, the types of shot transitions and conditional attributes of rough sets, decision attributes of rough sets, the rough sets and fuzzy clustering based method for shot boundary detection is proposed in Chapter 3 and has been successfully applied into the segmentation of cut transition, gradual transition and no transition in news video. For the drawbacks of rough sets dealing with data, the rough-fuzzy sets is introduced and the definition of rough-fuzzy operator is given. According to the feature of video in big data, the detailed procedure of“training before detection”and the automatic threshold method are brought foreword, which improves the automatization degree.Firstly, it is studied that exacting strict problems for data veracity in rough sets and the limitation of fuzzy c-modes is analyzed in Chapter 4, the variable precision rough sets is introduced and the feature weighted clustering method based on the更多还原