数字图像和三维几何模型水印技术研究

【作者】 李黎；

【导师】 潘志庚； 张大鹏；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2004， 博士

【摘要】 伴随着网络技术和多媒体技术的飞速发展，如何保护多媒体信息的安全成为国际上研究的热门课题。本论文是关于数字图像和三维模型水印技术的一些研究，研究成果包括： 提出了改进的盲水印提取模型：在基于频域的盲水印算法中，传统水印提取模型由于阈值不能自适应而影响水印提取效率。为了解决这一问题，我们将嵌入水印图像中的水印信息看作一种噪声，在提取水印前，利用降噪方法得到原始图像的一个逼近，通过和待检测图像的比较，得到含有水印信息的噪声，再将这种噪声信息进行放大，最后利用局部最大的编码原则将水印恢复出来；将改进的水印提取模型应用于基于Fourier-Mellin变换的盲水印算法，解决了由于插值和多变换造成误差积累而造成很难提取水印这一问题。实验证明该算法有很强的强壮性，比如抗几何变换、加噪等。 提出了构造和提取“零水印”的算法：“零水印”是利用图像的重要特征来构造水印信息，而不修改原始图像的任何数据。但由于“零水印”和传统的水印相比，没有嵌入强度，因此如何构造一个强壮的“零水印”是一个较难的问题。本文“零水印”构造算法是基于FFT和LPM来提取几何不变量，并选取一部分作为零水印，实验结果表明在经历了旋转变换、平移变换、缩放、JPEG压缩、裁剪、涂改和锐化等攻击仍然能将“零水印”提取出来。 提出了稳健性好的图像盲水印算法：白色像素较多的图像加入水印后会产生大量溢出误差，从而很难将水印提取出来或者对图像质量造成严重下降。针对这一问题，本算法基于Fourier变换，将原始图像进行分类，对不同的图像嵌入强度不同的水印，将水印信息嵌入频域幅值谱中不同的位置。实验表明该算法稳健性较好，对裁剪、噪声、压缩、涂改、平移抵抗性较强，并且算法复杂度较低，实用性较强。 提出了基于球面参数化的三维网格模型水印算法：利用网格全局球面参数化的方法得到网格的球面调和分析系数，将水印嵌入调和分析系数中，然后进行相应的逆变换得到嵌入水印的网格。在提取水印过程中，将网格水印信息进行放大以利于水印的提取，该算法还解决了由于采样和嵌入水印而造成的网格变形问题。球面调和分析保证本算法是将水印嵌入在真正Fourie义上的频率，对噪声、裁剪、滤波和增强的攻击有很强的强壮性；而且对噪声、裁剪、理想滤波、增强、相似变换、几何变换、网格重采样等攻击不需要网格对准和重采样。 提出了三维网格模型盲水印算法：利用Fourier加法性质将水印间接嵌入网格模型顶点坐标上，即先将水印嵌入一个稳定图像中，然后将嵌入水印后的图像和原始图像进行比较得到水印信息，最后将水印信息嵌入网格模型顶点坐标上得到嵌入水印的三维网格模型。实验结果证明算法有较强的强壮性，并且实现提取浙江大学博士学位论文水印不需要任何辅助信息和预处理。 提出了细分曲面和NURBS曲面水印算法:细分曲面水印算法是基于网格水印算法，将初始网格迭代三次以使网格的顶点数足够多，然后取出一部分顶点坐标用来嵌入水印，嵌入水印后的网格作为新的初始网格。NURBS曲面水印算法是保持模型形状完全不变、而对模型数据文件的大小做轻微改变，将水印信息嵌入原始节点中的相对距离中。更多还原

【Abstract】 The multimedia information security problems including copyrights piracy emerged as side effects over network. In this dissertation, we study the technologies of digital image watermarking and 3D model watermarking. The main contributions of this dissertation are:A Novel Detection Strategy In Blind Image Watermarking AlgorithmGenerally, the traditional detection strategy corresponding to embedment model I=I+ aW has its limitation. First, we treat the watermark information as noise for an image and expand the noise before detection. Then, we find the corresponding data in the frequency-domain and adjust them according to our relative minmum principle. Finally, we recover the watermark according to encodement. We also apply the novel detection strategy into the image watermarking algorithm based on Fourier-Mellin transformation, which is trong to geometric invariants. The original image was not required for extracting the embedded marks.A Method to Construct "Zero-Watermarks" "Zero-Watermark" is a new concept that donot change the original image during watermark embedment. It is difficulit to construct a strong "Zero-Watermark" because "Zero-Watermark" is weaker than the traditional watermark. The geometric invariant which is produced by Fourier-Mellin is computed and some of them are taken as "Zero-Watermarks". The experimental results show that our "Zero-Watermark" is effective and robust.A Valid Blind Watermarking Algorithm On Image The images that have a lot of white pixels usually can produce a lot of overflows after being embedded the watermark. The algorithm embeds the watermarking signal modulated on some fix positions that keep some distance from each other. Some decimals are added to the image-signed and different parameters are mbedded for different images. The experiments show that it is transparent, robust, accurate and cost-efficient.A 3D Mesh Watermarking Algorithm based on Spherical ParameterizationFirst, we transformed the coordinate signals of the 3D triangular mesh into spherical signals using a global spherical parameterization and an even sampling scheme. Then, spherical harmonic transformation is used to generate some data for embedding watermarks. As a result, the watermarks can be embedded in the Fourier-frequency domain of the original mesh. Experimental results showed that our watermarking algorithm is robust since watermarks can be extracted without mesh alignment or re-meshing under a variety of attacks. The processing is needed when the watermarked mesh goes through vertexes reordering and simplification.A Public Mesh Watermarking Algorithm The watermarked mesh is obtained by using a public image watermarking algorithm to embed a watermark in a gray image. First, the watermarked image minus the original image is the watermark information. According to the addition property of the Fourier transform, a change of spatial domain will cause a change in the frequency domain, which will allow watermarks to be extracted from the watermark information. Then, the watermark information is scaled down and embedded in one part of the x-coordinate of the original mesh. Thus, the watermarked mesh is obtained. Experimental results prove that the algorithm is resistant to a variety of attacks.Watermarking Subdivision Surfaces and NURBS Surfaces For subdivision surface, some 3D mesh watermarking algorithm is used and the watermarked subdivision surface is obtained. For NURBS surface, the watermarks are encoded and embedded into some intervals of the knots of the oringnal NURBS suface. The shape of the oringnal NURBS surface is not changed after being watermarked.更多还原