【作者】 师鸣若；

【导师】 胡涛； 刘昉；

【作者基本信息】 西安理工大学 ， 信号与信息处理， 2003， 硕士

【摘要】 随着视频信息的大规模引入，基于关键字的传统检索技术已不能适应人们的需要，近年来，基于内容的检索成为研究的热点。本课题分析了现有的视频检索理论框架，并对传统的文字识别算法加以改进，实现了视频帧中复杂背景下的实时文字检测与数字识别。对视频中文字的丰富的语义信息加以分析，在镜头分割、场景聚类的基础上形成镜头描述的特征空间，建立了以文字信息为主要特征的视频检索系统。 本课题以篮球比赛的视频片断为例，主要进行了以下几个方面的研究：1)针对视频中的“闪动”和“急跳”现象，提出了“二级镜头分割算法”，解决了镜头边缘的误识别问题。2)对所检测出的视频片断或整个视频流进行播放。3)基本解决了复杂背景下文字检测与定位问题，提高了数字的识别速度与正确率。4)文字区域的检测分为固定区域和非固定区域的文字检测，本文采用“双阈值窗口检测法”算法讨论了对固定区域的探测。5)使用改进了的模板匹配方法实现对印刷体文字进行识别，在传统的模板匹配基础上，提出了“二值化掩码模板”和“三灰度加权匹配”的算法。 本系统以Microsoft Window 2000 professional为操作平台，采用Microsoft Visual Studio C++6.0为工具开发出"NBA篮球比赛视频检索系统”(简称NBVRS)。该系统采用模块化设计，层次分明，界面友好，识别正确率高，且基本满足了用户 西安理工大学硕士学位论文实时性检索的要求。同时，该软件还具有视频的括放功能。最后对系统所采用的算法进行了实验验证，并指出了后续研究的方向。 本课题在原有的视频检索框架上对视频片段的语义内容加深了理解，基本达到依靠比分来进行播放的目的，虽然与实用阶段还有距离，但对于其它研究有着借鉴意义，并具有良好的应用前景。更多还原

【Abstract】 As the video information introduced cosmically, the traditional retrieval technology based on key words can not achieve the satisfying goals, instead, content-based video retrieval system has been the focus of study on video in recent years. The subject makes an analysis on the theory frame of nowadays video retrieval, add the improvements to the traditional algorithms of printed character recognition, and realize the real-time text detection and character recognition on the complex background. The paper Analyzes the fruitful information of the text in video, develops the symbolic space to describe the shot based on the video segmentation and scenes grouping, and establishes the video retrieval system characterized the text information.The study on query system using the basketball video fragments as an example is involved in such areas: first, in the condition of "video-flash" and "jump" phenomenon, the "two-level video segmentation algorithm" solves the mistaken detection of the shot edge; second, the part or the whole video as the result of detection can be well played; third, the text on the complex background can be detected and fixed, not only the speed but also the right proportion have been improved; forth, the text can be divided into "regular detection" and "irregular detection" . The "double thresholdwindow detection algorithm " is applied in the former, and another uses the progress of "crude detection algorithm " and "delicate detection" ; fifth, using the improved templates matching method to realize the printed character recognition. "binary mask templates matching" and "tri-grayscale matching" algorithms are introduced."NBA video retrieval system" has been developed using the Microsoft Visual Studio C++6. 0 under "Microsoft Window 2000 professional" . The system is designed on the module structure, so the arrangement is clearly demarcated, and the interface is handled friendly. It can realize highly right rate of recognition, and satisfied the user’s real-time retrieval. At last, the algorithm is verified, and the later research directions are also pointed out.更多还原