基于支持向量机的数字水印技术研究-智能与分布计算实验室

基于支持向量机的数字水印技术研究

姓名	李春花
论文答辩日期	2006.11.08
论文提交日期	2006.11.22
论文级别	博士
中文题名	基于支持向量机的数字水印技术研究
英文题名	Research on the Digital Watermarking Technique Based on the Support Vector Machine
导师1	卢正鼎
导师2
中文关键词	数字水印;支持向量机;人眼视觉系统;模糊聚类分析;小波系数方向树;图像内部的自相似性
英文关键词	digital watermarking;support vector machine;human visual system;fuzzy clustering analysis;wavelet coefficient direction-tree;similarity in an image
中文文摘	计算机网络技术和多媒体处理技术的迅猛发展，使得多媒体信息的安全问题成为目前一个相当重要而又富有挑战性的研究课题。数字水印技术就是在这种背景下产生的并很快获得了业界的广泛重视，成为当前发展最为迅速的领域之一。因此，针对不同的应用领域，研究性能良好的数字水印系统有着重要的理论和现实意义。将一种新的机器学习方法??支持向量机（SVM）引入数字水印领域，以期望最大限度地改善水印系统的综合性能，同时也为支持向量机在图像处理和信息安全领域中的新的应用进行有益的探索。在对支持向量机理论分析的基础上，针对目前数字水印技术的研究现状和存在的一些不足，分析了支持向量机在数字图像水印中可能潜在的一些应用，针对其中的一些应用进行了深入的研究和探索。研究了数字水印领域中支持向量机的参数选择问题，提出一种变尺度混沌优化SVM模型参数的算法，并给出了设定模型参数初始值范围的方法。该算法将SVM模型参数的选择看作是参数的组合优化，通过建立合理的优化目标函数，采用变尺度混沌优化算法来搜索最优目标函数值。为提高搜索效率，算法根据寻优过程中得到的临时最优解，不断缩小优化变量的搜索空间。在此算法基础上，通过大量实验，分析了回归支持向量机（SVR）模型参数对数字图像水印性能的影响，得出了纹理复杂程度不同的图像的比较理想的SVR学习参数范围。针对空域水印算法普遍较差问题，结合支持向量机优良的学习性能，提出一种基于支持向量回归的空域盲水印嵌入算法。该算法根据空域图像邻域像素的灰度值之间具有很强的相关性这一特点，运用回归支持向量机建立图像中邻域像素之间的内在关系模型，通过调整关系模型的输出值与目标值之间的大小关系来隐藏水印信息。提取水印时，不需要原始载体图像和水印图像，只需要根据水印嵌入位置的密钥就可以通过关系模型恢复出水印。实验结果表明了此算法的有效性。根据SVM与人眼视觉系统在自学习、泛化和非线性逼近等方面具有极大的相似性，结合图像的局部相关性特性，提出一种基于模糊支持向量机的自适应水印算法。该算法利用SVM来模拟人眼视觉系统特征，构造了以信息熵、亮度、对比度和纹理掩蔽值四个分量组成的特征向量的一些样本，从而为空域图像像素建立分类模型，根据此模型自适应地确定水印的最佳嵌入位置和嵌入强度。在利用SVM建立分类模型时，根据人类视觉的模糊特性，提出一种基于支持向量机的模糊多分类方法（FMSVC），运用FMSVC对图像像素进行模糊分类，并采用无监督的模糊聚类分析方法为有监督的支持向量机构造训练样本。实验表明了此算法的有效性。结合小波变换的多分辨率特点和支持向量机在理论上和学习上的优势，研究了小波域中基于SVM的水印算法，提出了小波域中基于SVM方向树模型的鲁棒水印算法和半脆弱水印算法。首先根据小波变换空频局域性特点，给出了小波系数方向树的概念，然后运用支持向量机建立了方向树上根节点与其子孙节点之间的依赖关系模型（即方向树模型），根据此模型设计了两种水印算法。其中，基于方向树模型的鲁棒水印算法将图像的空域和变换域相结合，采用模糊聚类分析的方法从空域中选取合适的水印嵌入位置，并映射到小波变换域的相应子带区域，从而自适应地确定水印嵌入的位置，水印嵌入的强度由嵌入位置的隶属度决定。而基于方向树模型的半脆弱水印算法通过密钥随机选择水印的嵌入位置，如果不知道模型参数和密钥，很难检测出水印。由于SVM模型捆绑了方向树上的小波系数之间的关系，对图像中任何一点的修改都会影响到水印位的正确恢复，因此要想绕过水印而对图像进行篡改有着很高的难度。算法通过一个滑动窗口对中值滤波后的篡改信息矩阵进行扫描，计算各滑动窗口的局部篡改率，根据最大局部篡改率来判断是常规操作还是恶意的篡改。实验表明了两种算法的有效性。
英文文摘	With the rapid development of computer, network and multimedia techniques, the multimedia security becomes a quite important and challenging research topic. Digital watermarking technique is a new method for digital content protection such as copyright protection, content authentication, transaction tracking, copy control, broadcasting monitoring and secret communication, and becomes an active research area in the field of media security. A kind of new machine learning method, namely support vector machine (SVM), is hopefully introduced to improve the watermark performance as better as possible. Meanwhile, it can also extend the new application of SVM in the filed of image processing and information security. Some potential schemes are analyzed in detail about using SVM to enhance the performance of image watermarking after analyzing the theory of SVM and the deficiency of current image watermarking, parts of schemes are studied deeply in this dissertation. Optimal SVM parameters selection used in the field of digital watermarking is studied, and a mutative scale chaos optimization algorithm is proposed based on the chaos variables which considered the selection problem of SVM parameters as a compound optimization problem by searching optimal objective function. In order to improve the efficiency, searching range of variables is shrunk continually during optimization according to the temporary optimal result. Based on the above result, the influence of support vector regression (SVR) parameters on the image watermarking performance is analyzed, and the ideal value range of SVR parameters is given respectively for different images. A blind spatial domain watermarking scheme based on support vector regression is proposed which considering the correlation among neighboring pixels in an image. It uses SVR to learn the relation between the central pixel and its neighboring pixels by choosing limited training samples and suitable SVR learning parameters, then, a bit of the watermark is embedded or extracted based on this relation model. The presented scheme can extract the watermark without the help of the original cover image and watermark image. Experimental results show the effectiveness of the proposed scheme. An adaptive spatial domain image watermarking algorithm based on fuzzy multi- classification support vector machine (FMSVC) is proposed. According to the very close similarity between SVM and human visual system (HVS) in self-learning, generalization and non-linear approximation, SVM is used to classify the pixels based on the characteristic of entropy, luminance, contrast and texture. Sequentially, the watermark embedding locations and strength are identified adaptively. The notable characteristic of the scheme is that, a kind of fuzzy multi-classification method based on SVM is presented which adopts fuzzy c-mean clustering algorithm, a kind of un-supervisory machine learning method, to construct training samples for FMSVC. Experimental results show the effectiveness of the proposed scheme. Two kinds of watermarking schemes, namely robust and semi-fragile watermarking, are proposed based on the coefficient direction-tree structure in the discrete wavelet transform domain. First, the concept of wavelet coefficient direction-tree is defined, then the inherent relation between root node and its offspring in the direction-tree is discovered and modeled using SVM, finally the above two watermarking scheme are designed based on the direction-tree model. The proposed robust watermarking scheme combines the spatial domain and transform domain, and selects the watermark embedding locations adaptively from the spatial domain by using fuzzy clustering method. The proposed semi-fragile watermarking scheme chooses the watermark embedding locations randomly with the secret key. So, the watermark is hardly detected without the knowledge of SVM parameters and secret key. Specially, since SVM binds the relation among wavelet coefficients, any modification will affect the recover of watermark bit. In this scheme, a tempering information matrix (TIM) and a local tempering appraising function (TAFL) is defined respectively. TAFL is calculated with a sliding widow which scans the TIM operated by medium filtering, and the maximal TAFL, namely TAFM is used to distinguish the common image operations (such as JPEG compression, filtering, noise) or the malicious tempering operations. Experimental results show the effectiveness of two proposed scheme.