【作者】 吴国民；

【导师】 潘云鹤； 庄越挺；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2007， 博士

【摘要】 计算机运算速度和存储容量的不断提高，网络技术、数字化技术的迅猛发展，使得图像、视频、音频、三维图形等多媒体信息日益丰富，其传播日益容易。目前，几乎所有人类知识都可以数字化并在网络上传播。 然而，大量的媒体信息鱼目混杂，基于互联网的信息盗版十分猖獗，如何规范互联网信息的传播秩序?如何约束信息源的非法传播?本文研究海量环境中的多媒体信息安全、多媒体盲水印的嵌入和提取、基于视觉特征的视频水印技术以及音频水印技术研究。主要完成如下工作： 1．提出了一个面向搜索引擎的海量多媒体信息安全系统结构，并给出了原型系统实现。目前无处不在的搜索引擎是多媒体盗版和非法传播的主要途径，本方案通过基于数字水印技术的互联网信息发布、面向搜索引擎的海量多媒体监控、和海量信息安全认证三个部分创建一个中央权威机构。将传统水印技术和搜索引擎技术相结合，通过创建中央权威认证库、海量多媒体信息特征库、海量多媒体源数据库以及引擎跟踪数据库，对基于搜索引擎的海量信息传播行为进行监督。通过提供明确的非法信息来源来约束搜索引擎对非法信息的传播，并为正版多媒体信息通过搜索引擎的有偿传播提供了一条新的路径，为搜索引擎的信息传播行为开辟了一个新的方向。 2．实现了一个面向搜索引擎的海量多媒体信息监控系统，系统采用Tracker和Crawlers分工合作的方式，使海量多媒体的监控的系统复杂度大大降低。系统通过人工规则设定，区域划分，搜集本地化实现面向引擎的海量信息跟踪与监控。 3．提出了一个多媒体盲水印技术以实现上述普遍适用的音、视频水印算法。针对传统盲水印算法的不足，结合海量应用对水印技术的基本要求，提出了能量跨度函数来实现盲水印嵌入。以此为基础，根据零水印技术的基本原理并利用媒体信息的丰富特征，给出了一个适合于多媒体水印技术海量应用的盲水印嵌入和提取模型。实验结果表明，该模型具有很高的可行性，针对海量的多媒体信息处理，这一模型具备较高的使用价值。 4．提出了一个基于特征建模的自适应视频水印算法。该算法根据人类视觉系统(HVS)特性，对DCT子块进行特征分类。将分类得到的视频帧纹理、亮度、边缘信息以及运动特征来计算水印嵌入的伪装系数，实现了水印嵌入自适应于视频的视觉特征；针对视频数据的海量应用，考虑算法鲁棒性需求，提出了一个独立于视觉特征的统计模型。该统计模型进一步细化视觉伪装模型所描述的子块之间的特征差异。实验结果表明该算法具有很好的更多还原

【Abstract】 With the fast developments of networking technologies and multimedia technologies, there is an increasing demand of information security protection and multimedia authentication technologies. Nowerdays, the copyright protection is in great demand, especially in the case of search engine, video ordering and online music. This thesis mainly focuses on the magnitude watermarking and its implementation and the main contributions of this thesis are as follows:1. Considering the illegal downloading and spreading of digital media, a framework which works finely to protect the digital media copyrights under the massive circumstance. The framework combines the traditional copyright protection and search engine technologies and it consists of video-audio watermarking embedding, digital media retrieval and media authentication. The video-audio watermarking embedding algorithm is composed of video and audio watermarking embedding algorithms which can be widely used; both the content-based retrieval and text-based retrieval algorithms consititute the digital media retrieval algorithms; the media authentication mainly validates media watermarking and other information related to the internet.2. A monitor system of mass multimedia is implemented. The goal of the system is to monitor and correct the behavior of search engine.3. A blind watermarking technology is proposed in this paper to facilitate video-audio watermarking embedding and security authentication of the framework. Put the blemishes of traditional copyright protection technologies and the principles of zero watermarking into consideration, an energy span fuction is proposed. Based on the function, a watermarking embedding and detection model which can deal whith the massive digital media is proposed. The model can make use of abundant information in digital media themselves and experiments show that the model is robust and the performance is remarkable when dealing with massive multimedia objects.4. According to the blind watermarking technology, we present a novel watermarking algorithm of MPEG video based on visual features. An effective mixed feature model of the block is delicately designed by texture, luminance and motion features of the video. According to the perceptual features derived from the masking behavior of the human perceptual model, a feature statistics model is introduced to adjust the extracted block feature, and then the copyright protection is embedded into the block according to the modified block feature. Experimental results show that the watermarked video is highly imperceptible robust.更多还原