【作者】 曾高荣；

【导师】 裘正定；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2010， 博士

【摘要】 数字水印的鲁棒性是水印技术实用化的一个重要指标。与加密技术不同,数字水印技术并不严格要求内容无失真传输,只要人们在视觉或听觉上感知无变化,允许内容使用时和原始内容存在一定程度的失真。这就要求隐藏在数字内容中的水印信息在传输和处理后能幸存(被有效检测),这种幸存能力称为数字水印的鲁棒性。论文结合鲁棒水印的研究需求,从理论和算法设计两个方面开展研究。理论上要设计符合鲁棒性要求的水印算法必须首先确定鲁棒性的描述方式和评价标准。论文在这一领域的主要研究成果如下：(1)将数字水印看作一个通信过程,将参与水印系统的各方看作随机统计量,以接收的数字内容和水印信息之间的互信息为代价函数,建立水印系统鲁棒性描述和度量的一般模型——互信息度量模型。该模型可作为一般水印系统鲁棒性分析的框架。(2)论文在互信息度量模型框架下,按照数字水印透明性和鲁棒性的要求,对具有典型代表意义的扩频水印和量化水印进行性能分析。i.在扩频水印分析中,从一般加性扩频方案和改进的扩频方案两个方面,分析推导了度量鲁棒性的互信息函数求解方式,仿真计算了高斯噪声和JPEG压缩情景下鲁棒性函数随水印噪声比或压缩强度的变化关系,并通过实验测量和统计误比特率来验证鲁棒性评测结果的有效性。实验结果表明互信息函数评测的鲁棒性能与实验误码率评估的结果很好地匹配,且水印噪声比越大,互信息函数值越大,水印系统越鲁棒,相应的误码率越小。ii.针对量化水印范式,论文分析了量化索引调制(Quantization Index Modulation:QIM)水印和失真补偿量化索引调制(Distortion Compensated Quantization Index Modulation:DC-QIM)水印之间性能差异,运用互信息度量模型评测它们在不同条件下的鲁棒性,并通过仿真实验检验该模型在高斯噪声和均匀噪声以及JPEG压缩条件下的鲁棒性估计,实验结果表明,互信息度量模型评测的鲁棒性随水印噪声比增大而增大,且与实验检测的误码率之间是匹配的,两者对水印系统可靠性的评测结果是一致的。(3)算法研究方面,在DCT域水印的嵌入策略分析中,以互信息度量模型对鲁棒性评测结论为指导,确定水印嵌入位置的选择方法,指出在兼顾透明性和鲁棒性要求的前提下DCT域DC系数是一种较优的嵌入位置。实验仿真表明,基于DC系数的水印方案和低频、高频AC系数相比,具有较好的鲁棒性,且其PSNR或SNR是可接受的感知范围。(4)论文以互信息度量模型对鲁棒性的评测结论为指导,结合独立分量分析(Independent Component Analysis:ICA)方法提出一种DCT域上的图像水印算法。在算法实现时,以视觉掩蔽门限系数为参数调制水印的嵌入强度,以公开传播渠道获取的公共参照图像,作为ICA算法的一个参考输入,提取水印。在鲁棒性测试实验中,算法表现出较好的鲁棒性。由于参与ICA运算的公共参照图像由密钥控制生成,无需作为隐秘参数,因此不占用水印系统的传输信道,从这个意义上认为是个盲检测的水印算法。更多还原

【Abstract】 Robustness is one of the most important requirements when digital watermarking is applied. Different from encryption, digital watermarking doesn’t strictly require distortion-free transmission. Before people’s visual or auditory signals come into being, the host data can tolerate some distortion from signal processing. This requires the embedded watermark can be effectively detected after some signal processing have taken place. The detected ability is known as robustness of digital watermarking. To meet the R&D requirements of robust watermarking, the theory about robustness and algorithm design of watermarking are studied.In order to design the robust watermarking algorithm, the description method and evaluation criteria must be created beforehand. In this regard the main results and contributions of this dissertation are as follows:1. Digital watermarking is viewed as a communication process. The paties participating in the watermarking system are seen as random statistic. A mutual information function between the watermak and the observation is defined as a criterion measuring the robustness of watermarking algorithm. The mutual information metric model can be used as a general framework of watermarking system robustness analysis.2. In the framework of mutual information metric model, robustness of spread spectrum watermarking and quantization watermarking are analyzed with the consideration of transparency and robustness requirements.(1) Spread-spectrum watermarking is divided into simple additive spread-spectrum watermarking and improved spread-spectrum watermarking. The calculation formulas of mutual information function are derived to evaluate the robustness of algorithm. In the experiment, spread spectrum watermarking is implemented in discrete cosine transform (DCT) and the statistic Bit Error Rate (BER) is derived against Gaussian distribution noise and JPEG compression. Experiment results show that evaluation conclusion of mutual information method is in accord with empiric BER. The greater the mutual information function is, the more robust watermarking system is, and the smaller corresponding BER is.(2) The performance analysis of another watermarking paradigm, quantization watermarking is given, which includes Quantization Index Modulation (QIM) watermarking and Distortion Compensated QIM (DC-QIM) watermarking. With the use of mutual information measurement model, their robustnesses are analyzed. Simulation experiments are done to validate the model in Gaussian noise, uniform noise and JPEG compression. Evaluation of the robustness based on the mutual information metric model is matched with the empiric BER. Two evaluation results for reliability of system are consistent.3. When digital watermarking is applied, system performance constraints and convenient implementation scheme are the important consideration. In this dissertation watermark embedding position is selected in DCT domain and realization methods of blind watermarking are studied. Based on the result of mutual information measurement, a method to determine the optimal embedding location is given. In view of robustness, compared with low frequency AC coefficient and high frequency AC coefficient, DC coefficient is a better location. And its PSNR or SNR is acceptable limitation of perception.4. Considering the convenience, the blind watermark detection is a good idea. Taking advantage of Independent Component Analysis’s (ICA’s) blind source separation characteristics, a blind watermark extraction algorithm is designed in DCT domain. In order to obtain the larger WNR, the visual masking threshold parameter is used to modulate the watermark embedding strength. When the original image and the original watermark are unavailable, a legal user can select the watermarked image and the public reference image as inputs to ICA, and then extract the watermark by ICA algorithm. The public reference image is generated by the key image and original image, but it can be delivered by public transmission channel. So it is not Hidding parameter and does not share the watermarking channel. In this sense, it is a blind watermarking sheme.更多还原