【作者】 刘洋；

【导师】 毛建旭；

【作者基本信息】 湖南大学 ， 电路与系统， 2008， 硕士

【摘要】 随着多媒体技术和网络技术的飞速发展,数字视频的获取和传播变得越来越容易,已经逐渐成为人类信息传播的主要载体之一。在视频信息高度膨胀的今天,随之而来的问题就是对海量视频的高效检索和浏览。传统的视频检索通过对视频以手工的方法添加文字标识符的方式进行检索,这种检索方式工作量巨大、效率很低,而且受主观因素的影响,因此不能满足实际使用的需要。基于内容的视频检索技术借助计算机对视频进行从低层到高层的处理、分析和理解的过程获取其内容并根据内容进行检索,克服了传统的基于文本检索方式的不足,已成为多媒体信息检索领域的研究热点。本文首先分析总结了视频检索技术的理论框架和研究现状,然后对该领域中的视频镜头分割、关键帧提取、镜头聚类等关键技术进行了深入的研究和探索。视频镜头分割是进行视频处理的第一步,本文在总结现有镜头分割方法的基础上,研究了基于互信息量的视频镜头分割方法。设计并实现了一种基于双滑动窗口的镜头切变检测算法,算法通过计算视频帧间的互信息量作为衡量两帧相似度的依据,采用双滑动窗口方法找出相邻帧间互信息量的局部极值用于确定切变镜头的边界。针对运动和闪光对镜头检测的干扰,提出了一种基于图像分块的互信息量镜头切变检测算法,算法以互信息量作为评价帧间差异的准则,通过把帧图像分块,然后分别计算相邻两帧对应子图像块之间的互信息量,再进行反比例变换后累加,利用自适应阈值方法找出帧间差的局部极值,从而找出切变镜头的边界。研究并实现了一种基于互信息量的镜头渐变检测算法,算法利用不同帧间距的非相邻帧间互信息量差值检测渐变镜头边界。实验结果表明,本文所提出的视频镜头分割方法指标明确、算法简单,对切变和渐变达到了较高的查全率和准确率。视频关键帧的提取是基于内容的视频检索技术的关键步骤之一,本文首先研究了关键帧提取技术的原理和主要方法,然后将互信息量引入关键帧提取中,提出了一种基于互信息量的关键帧提取算法,算法针对镜头内互信息量的变化,通过计算帧间差的标准差来判断镜头内连续帧的相似性,并对相似性较高的连续帧提取一帧作为关键帧。实验结果表明,使用本文算法提取的关键帧可以准确地反映镜头内容,较好地得到了真正意义上的关键帧。镜头聚类作为一种从视频内容低层特征到高级抽象的桥梁,在基于内容的视频检索系统中起着很关键的作用。本文首先研究了聚类分析的原则和特点,并简单分析了该领域存在的主要算法,然后提出了一种基于关键帧颜色特征的视频镜头聚类算法,算法将关键帧的颜色特征作为聚类依据,运用改进的K均值聚类算法对镜头进行聚类,并对聚类结果进行优化。实验结果表明,本文所提出的算法具有较高的准确率和效率,增加了聚类结果的稳定性,达到了令人满意的效果。更多还原

【Abstract】 With the rapid development of the multimedia technology and network technology, the access and dissemination of digital video has became more easily.The video has became one of the main carrier of information dissemination. In the days of highly inflated video information, the attendant problem is efficient retrieval and browsing to the massive video. The traditional way of video retrieval searches in the way that adds the text identifier using manual methods. The workload of this method is large and the efficiency is low. Also the performance of this methods is affected by the subjective factor, so it can’t meet the actual needs.The Content-Based Video Retrieval which get the content of the video employing the method of processing, analysis and understanding to the video from low to high-level by computer and searches video accordance with it has became one of hot issue in the field of multimedia information retrieval.This paper summarized and analysised the the theoretical framework and the study status of video retrieval firstly,then conducted a in-depth study and exploration to some key Techniques in this field, including shot segmentation、key frame extraction and shot clustering. The shot segmentation is the first step to the video retrieval.On the basis of concluding the existing algorithm, the paper did some research into shot segmentation based on mutual information. The paper designed and implemented a Cut Transition algorithm based on dual-sliding window, which calculate the Mutual information among frames to determine the similitude of two frames and use the dual-sliding window method to find the local extremum of neighbor frame mutual information in order to locate the boundary of cut shots.The paper proposed a mutual information Cut Transition detecting algorithm based on image block against the interference of the movement and flash in the detection of Cut Transition.The mutual information is used as the criteria for evaluation of differences between frames.The algorithm partition the image first and then calculate the mutual information of the corresponding sub-image between the adjacent image.Then the algorithm conduct a inverse proportion transformation to the value of mutual information and do a accumulation to all values in the whole image.At last the algorithm use adaptive threshold method to identify the local extreme of the difference between frames to find the boundary of cut shots.The paper research and achieve a shot Gradual Transition algorithm which the difference of different frame spacing of non-neighbor frame mutual information is used to inspect the boundary of gradual transition shots. Experiment results show that the algorithm is simple、clear indicators and achieve a high recall rate and accuracy.The key frame extraction is the key step of content-based video retrieval. The paper first conduct a research on the principle and the main methods of key frame extraction technology and thenr made a introduction of mutual information to the key frame extraction and advised a mutual information based key frame extraction algorithm.The algorithm calculate the standard deviation of the difference between frames aim at the variety of mutual information inside the shot to determine the similarity of consecutive frames from which extract a key frame. The experiment results showe that the use of this method can accurately reflect the contents of the shots and get a real sense of the key frame better.The shot clustering plays a very important role in Content-Based Video Retrieval system as a bridge from the low-level characteristics of video content to the the high abstract of video content.The paper first do some research to the rinciples and characteristics of clustering analysis and analysised the major algorithms in the areas simplely. After that the paper suggested a shot cluster algorithm based the color characteristics of the key frame.The algorithm use an improved K-means clustering algorithm for shot clustering according as color characteristics of the key frame and optimize the clustering results. Experimental results show that the method achieved the high accuracy rate and efficiency, increase the stability of the Cluster results, obtained the satisfactory results.更多还原