【作者】 赵亚琴；

【导师】 周献中；

【作者基本信息】 南京理工大学 ， 控制科学与工程， 2007， 博士

【摘要】 近年来，随着视频信息资源的日益丰富，从大量的视频数据中检索感兴趣的视频片段已经成为目前视频检索研究的热点。人们期望在不久的将来像文本检索那样，由用户向系统提交查询的视频内容，系统自动地返回符合查询需要的结果。在这种背景下，基于内容的视频检索应运而生，它综合了图像处理、模式识别、计算机视觉、图像理解等领域的知识，具有广阔的应用前景。本文以广告类视频、球拍类体育比赛视频和新闻视频为研究对象，主要研究了这三类视频的片段检索方法。提出了一种新的广告类视频片段检索方法。在定义查询片段镜头的相似镜头集的基础上，给出了两个片段一一对应相似镜头数目的计算方法，以排除伪相似片段，进而定义了片段的匹配函数；采用滑动镜头窗从连续的视频流中自动分割出多个相似片段；将相似镜头集映射为相似镜头矩阵，利用矩阵的特性来计算影响片段相似程度的各个因子，建立了相似片段的排序模型。与同样功能的其他方法相比，在确保较高的检索精度的前提下，提高了检索速度。融合视觉特征、运动特征和音频特征，提出了一种球拍类体育比赛视频的Rally精彩片段检索方法。提出了一种基于运动特征和主颜色特征的自动选择Global View Shots类的参考关键帧的方法，以便将镜头划分为Global View Shots类和Non-Global View Shots类；SVM用于识别镜头的重要音频类型；将音频信息应用到镜头类，建立了感兴趣事件检测的相应规则；利用回放的Rally片段和Rally精彩程度模型，从球拍类比赛中检索出符合多数人心理的Rally精彩片段。提出了一种基于视觉特征和语义信息的新闻视频片段检索方法。提出了一种新闻视频主题字幕文本检测方法，进而讨论了一种通过主题字幕文本和静音片段进行新闻视频故事分段的方法；基于低层视觉特征和主题字幕文本获得的语义信息建立了新闻片段的相似度模型，并应用相关反馈技术从新闻节目中检索出最符合用户需求的新闻故事片段。设计并初步实现了一个基于内容的视频片段检索实验平台，验证了上述片段检索方法的性能。更多还原

【Abstract】 Rapid growth of video resources leads to an urgent demand for effectively retrievinginteresting video clip from huge amounts of video data in recent years. As text retrieval,video retrieval is also expected that a retrieval system can automatically return results inaccord with query demand when users submit query video content. Content-based videoretrieval techniques have emerged as the times requires, which combine image processing,pattern recognition, computer vision and image understanding, etc. And it has wideapplication foreground.This paper mainly presents content-based clip retrieval methods of three video genres,commercial video, racquet sports video and news video.A new commercial video clip retrieval method is presented, in which the similar shotset of shots of query clip is defined to compute the number of one-one similar shotsbetween two video clips, on the basis of this, the definition of matching function of twovideo clips is given. Afterwards, similar commercial clips are automatically segmentedfrom continuous video databases using sliding shot window according to their matchingdegrees. Furthermore, the similar shot sets are mapped into one similar shot matrix tocompute various factors for similarity ranking of the similar clips by characters of thematrix. Experimental results showed that the proposed method could effectively andefficiently retrieve and rank similar video clips.This paper presents a new audiovisual integration scheme for retrieving rally highlightfrom racquet sports video. Motion features and dominant color are applied to classify shotsinto two classes, global view shots and non-global view shots. At the same time, importantauditory features including both ball hitting and applause are detected by using SVM.Afterwards, audio features are applied into shot classes for identifying interesting eventswith strong semantic meaning. Finally, rally highlights are retrieved by rally clips replayedand exciting degree model of a rally from racquet sports video.This paper presents a scheme of retrieving news video clips based on low-level visualfeature and high-level semantic information. The semantic information is obtained fromtopic caption text. In the method, news video is first segmented into a series of news storyclips on the basis of topic caption text and silence clip. Moreover, the segmented story clips are annotated by topic caption text. Afterwards, a similarity model of two news clipsis established based on topic caption text and low-level visual features. Furthermore,relevant feedback technique is applied to retrieve similar news story clips, which haverelative topic, or similar vision, or relative topic and similar vision with query news clips.A content-based video clip retrieval platform was designed and implemented tovalidate the performance of the above clip retrieval methods for three video genres.更多还原