【作者】 佘堃；

【导师】 周明天；

【作者基本信息】 电子科技大学 ， 计算机应用技术， 2006， 博士

【摘要】 数字水印是一种保护数字材料版权的方案。自从上个世纪90年代以来，越来越多的人工作在这个充满趣味、挑战和机会的研究领域。在本论文中，引入了LCNN(Lagrange约束神经网络)方法，模仿哺乳动物双眼系统比单眼更清晰的观测过程，以期获得更清晰的数字水印。经典LCNN(CLCNN)效率低，容易碰到“病态”矩阵问题。本文仔细研究了LCNN的学习矩阵的逼近过程，发现了Lagrange约束入与有监督学习目标min(As-X)之间的关系，引入x的逼近加速度替换入病态学习逼近的方案，更彻底地阐述了大脑有监督学习和无监督学习的“你中有我、我中有你”的关系。正是通过对这种并行学习的认识，本文提出了4类自适应LCNN(ALCNN)快速算法，为欠定稀疏分析(m＜n)和超超定(m＞＞n)分析的快速、并行实现，提供了新的思路和解决办法。本文在讨论多种主流数字水印技术(基于小波和基于独立成分)的基础上，自主设计了2种小波域自适应水印嵌入方案，一种是二值鲁棒水印，一种是鲁棒灰度水印。进一步，利用LCNN和双正交的小波基，设计了一个多分辨率子带分解与独立成分分析融合(MSD-ICA)的灰度数字水印嵌入方案，实验表明，对于较高强度的噪声、压缩等攻击，它的提取精度提高到90％以上，这对受环境影响较大的弱图像和弱水印的提取和验证是非常有用的。最后，针对安全中间件技术在电子政务印章系统中的应用项目，本文设计了一个基于双脆弱水印的电子印章公文保护方案，并申请了发明专利。在创新方面，归纳如下：(1)提出了一种根据宿主信号亮度变化自适应调整嵌入参数的灰度水印算法：针对灰度水印，自动调整插入和提取乘法系数。该算法对灰度水印和Cox的乘法、加法嵌入规则是有效的。能抗击水印游戏竞技中的较恶劣噪声、JPEG压缩等攻击。实际上，该算法对于一些背景亮度敏感的图片，效果也不错。相关内容发表在[134]。(2)提出了一种根据二值水印关键点的自适应调整的鲁棒参考水印算法：对著名的鲁棒参考水印算法(RRW)进行分析，针对其对纹理平滑部分的不适应的缺陷，提出了根据二值水印边界的关键点来选择不同长度网格(Grid)的量化指数算法，实验表明在所考虑的情况下该算法优于RRW算法，适应面更广。相关内容已经发表在[135][136]。(3)提出了4种快速、高效的自适应Lagrange约束神经网络(ALCNN)算法：重新认识了Lagrange函数约束项系数入的物理意义——有监督学习的加速度。所设计的4种算法的学习矩阵收敛都是O(n)，其中自适应调整λ算法的独立成分求解复杂度也是O(n)，使它具有更广阔的应用空间，并奠定了ICA大规模并行处理的基础；其它3种算法的独立成分求解复杂度都是O(n~2)，可以利用它们学习矩阵收敛的优势，在识别、追踪方面发挥作用。而带检验的ALCNN算法，则充分发挥了无监督学习、有监督学习的融合优势。这些工作经过总结，已递交到“电子科技大学学报”上。(4)设计了基于双脆弱水印的电子印章公文保护算法：将文档Hash值作为水印数据嵌入到电子印章里，加盖到电子文档中，使电子印章与文档相关联，以保护水印-文档的唯一性，具有数字签名性质。电子印章在单独存放时，用单位的标识作为脆弱水印嵌入，以保证印章的完整性。上述工作成果已集成到安全中间件体系架构中，并在“电子政务安全服务平台”项目中获得了应用。相关内容已经申请发明专利。专利申请号：200510021291．3，公开号：1725244。(5)提出了基于LCNN和小波技术嵌入灰度水印到彩色宿主图像的算法。使用LCNN的ICA技术，将灰度水印嵌入到彩色图像中。该方案对弱水印的提取效果较好，在好的不可感知性(高PSNR)的情况下，能获得较好的水印相似性NC。相关内容已经发表在[226][227]。更多还原

【Abstract】 Digital watermarking is one solution to protect copyright of digital materials. Since 1990s, more and more people focused on this interesting, full of challenges and opportunities field. In this dissertation, LCNN(Lagrange Constraints Neural Network) was introduced to imitate 2-eye system of mammals for getting the clearer watermarks in a noised environment.Classical LCNN(CLCNN) fell short of lower effective and ill-conditioned matrix. In this dissertation, LCNN was carefully investigated to learn the approximate procedure of the learning matrix, and found out the relationship between Lagrange constraintλand supervised learning target min(As-x), so the approximate acceleration of x,which is a supervised learning gradient,was supposed to replaceλwhich should be obtained by an unsupervised method, and the relationship——"And and Or Logistic"——between supervised and unsupervised learning, was depicted more deeply. Based on this kind of parallel learning, 4 types of Adaptive LCNN(ALCNN) algorithm were discovered, new idea and solution were emerged for quick and parallel analysis of underdetermined and overdetermined matrices.Several mainstream watermarking technologies,which are based on wavelet and independent components, were discussed, and 2 adaptive embedding solutions were designed, one is a robust 2-value watermarking, and another is a robust grey watermarking. Furthermore, based on the fusion of multiple resolutions subband decomposition and LCNN independent components, LCNN and biorthogonal wavelet basis were employed for a grey watermarking.The test results showed that the decoding accuracy had been improved to over 90% for higher noise, compression, etc. attacks. This is very useful to the extract and verification of weak watermarked image or weak watermarks, which will be varied with the environment. At the end of dissertation, a double-fragile-watermark solution was proposed in a security middleware project of E-government sealing system for protecting the documents with an electronic seal and an invented patent has been hold on this achievement.Innovations of this dissertation were depicted as follows: (1) Proposing a grey watermarking algorithm with adaptively adjusting embedded references, according to the luminance of host image, and has been published in a paper [134].(2) Proposing an adaptive robust reference watermarking(RRW) algorithm based on on the keypoints of 2-value watermarks.According to the shortages the famous RRW for texture smooth, the solution based on the keypoints, which were the edges of 2-value watermarks, was proposed. The relevant QIM can be chosed from different scale grid with the keypoints and achieve a better result than that standard RRW in some experiments. This achievements have been published in papers [135][136](3) Proposing 4-type rapid, effective ALCNN Algorithms. The coefficientλof Lagrange function constraints was re-known as an acceleration of that supervised learning, and 4-type of adaptive LCNN algorithm were designed. We proved that all learning matrices of the 4-type algorithm converged in O(n) time, and the adaptiveλlearning algorithm converged also in O(n) for solving independent components(IC), which means this algorithm will be used in wider applications and lays the foundation for parellel ICA, and the rests(the other 3 types of algorithm) converged in O(n2) for ICs, which can be used in the recognition,and tracing, while the algorithm of ALCNNs with checking showed the advances of fusing the unsupervised and supervised learning. All works of the achievements have been in the paper submitted to "Journal of the University of Electronic Science and Technology of China".(4) A double-fragile-watermarking algorithm for protecting electronic documents with electronic seals was designed. The Hash value of the document was used as watermarks to embed into an electronic seal and then the seal was stamped over the document. This process combined the seal with the document and guarrented the uniqueness of the pair of watermark-document, with the characteristic of digital signature. When the electronic seal was stored standalone, the tag of unit can be embedded as a fragile watermark in order to assure the integration of the seal. This achievement has been integrated into security middleware architecture and used in the project "Secure Service Platform of E-government", now, a patent has been applied with the application number 200510021291.3, and publication NO.1725244.(5) Based on LCNN and wavelet, a solution with embedding the grey watermarks into color host image was designed: this solution was particularly useful to extract the weak watermarks, and the good comparability(NC) between original and estimative watermarks can be obtained when the imperception is higher(higher PSNR). This achievement has been published in papers.更多还原