基于特征点的抗几何攻击的图像盲水印技术研究-智能与分布计算实验室

基于特征点的抗几何攻击的图像盲水印技术研究

姓名	余艳玮
论文答辩日期	2007.11.07
论文提交日期	2007.11.09
论文级别	博士
中文题名	基于特征点的抗几何攻击的图像盲水印技术研究
英文题名	Geometrical Invariant Image Blind Watermarking Techniques based on Feature Points
导师1	卢正鼎
导师2	凌贺飞
中文关键词	数字水印;图像水印;几何攻击;第二代水印;特征;同步
英文关键词	Digital watermarking;Image watermarking;Geometrical attacks;Two Generation watermarking;feature;synchronizatio
中文文摘	目前，人们不但可以通过互联网和CD-ROM方便快捷地获得多媒体信息，还可以得到与原始数据完全相同的复制品，由此引发的盗版问题和版权纷争已成为日益严重的社会问题。数字水印技术作为解决这类问题的一种潜在的有效技术，正受到国内外学者的广泛关注。由于图像信息在多媒体信息中的广泛性和重要地位，以及其研究过程所具有的代表性，因此这里主要以数字图像水印为研究对象。数字水印作为版权保护的手段，必然会受到各种形式的攻击，因此用于版权保护的数字水印应该具有鲁棒性。然而在抗几何攻击方面，目前仍然没有令人十分满意的水印方案面世，尤其在盲水印应用中。在目前已经提出的抗几何攻击的盲水印方案中，第二代水印，即基于图像内容的局部水印方案，有望能解决鲁棒盲水印中的抗几何攻击难题。基于内容的局部水印方案的基本思想是：利用图像中稳定的特征点标识水印嵌入的位置，并在与每个特征点对应的局部区域中独立地嵌入水印，从而当只有部分图像时，仍能通过这些特征点来定位和检测水印，因而可有效地抵抗裁剪攻击。基于局部内容的局部水印（也称为基于特征的局部水印），按逻辑可划分为三个模块：特征点的检测、用于水印嵌入和水印检测的局部区域的构造和各个局部区域的水印的嵌入和检测。整个水印系统的鲁棒性与特征点的鲁棒性、基于特征点构造的局部区域的稳定性以及局部区域上的水印算法的鲁棒性密切相关，缺一不可。因此，分别针对这三个功能模块展开深入研究。目前已有的基于特征的水印算法都是直接将模式识别领域中的特征用来同步水印。但是由于水印应用和模式识别领域的需求不一致，模式识别领域中的特征并不一定适用于水印应用。因此，提出了一种基于圆盘平均亮度适用于水印应用的旋转缩放不变特征（Disk Based Rotation and Scale Invariant Feature，简称为DRSIF）的提取算法。对于图像上的每一点，以该点为圆心构造一个内环半径呈等比数列变化、且内外环半径比固定的同心圆环序列，计算各个圆环的内外环的绝对平均亮度差，并将内外环的绝对平均亮度差的最大值（且为极大值）作为该点的响应值，将内外环的绝对平均亮度差取得最大值（且为极大值）、且最靠近圆心的圆环的内环半径作为该点的特征半径。检测图像上响应值取得局部极值的点以作为DRSIF点。从理论角度分析DRSIF提取算法的时间复杂度，从实验角度验证DRSIF点的鲁棒性，并与模式识别领域中最具代表性的两种尺度不变特征点进行比较。实验表明，DRSIF点对于水印应用中的常见攻击具有鲁棒性，因此适用于水印应用中。特征点对于几何的协变性并不能保证剖分得到的局部区域也具有几何协变性。基于特征点构造的局部区域的几何协变性还依赖于具体的剖分方式，即局部区域的构造方式。分别详细阐述了如何基于DRSIF点来构造局部水印嵌入区域和检测区域。基于原始图象上的DRSIF点构造出圆形区域，并从中选择熵值比较大、鲁棒的、且互不交叠的区域作为水印嵌入区域。基于待检测图象上的DRSIF点构造出圆形区域，并从中选择熵值比较大的区域作为候选水印检测区域。嵌入的水印以及图像所遭受的失真都可能使得图像的内容发生轻微的变化，因此待测图像上水印嵌入区域的构造过程很难准确再现。提出了分区策略来提高水印检测过程中水印嵌入区域的命中率，以加快检测到嵌有水印区域的速度。通过实验验证了基于DRSIF点构造的水印嵌入区域的鲁棒性和水印检测区域构造过程中分区策略的有效性。在前面的研究基础上，提出了两套完整的基于DRSIF点的局部水印方案：1）基于DRSIF点的空域局部水印方案；2）基于DRSIF点和不变域的局部水印方案。第一个方案中，利用区域的矩计算水印嵌入区域的特征角度来反映图像的旋转角度，以使得区域具有旋转不变性。第二个方案将水印重复地嵌入到各个水印嵌入区域的旋转缩放（简称RS）不变域中，通过将几何不变域与局部水印思想相结合，互补长短，不仅可以抵抗全局RST攻击，还可以抵抗裁剪等局部几何攻击。通过实验验证了这两套方案的鲁棒性。且与同类算法相比，提出的两套方案的性能优于同类算法。提出的两套局部水印方案相比，变换域局部水印方案的鲁棒性比空域局部水印方案略差，这是由于前者对于特征点的精确再现要求比后者高得多。另外，前者的水印嵌入容量比后者小得多。因此，必须要权衡鲁棒性和水印容量，结合具体的应用需求，选取适当的局部水印嵌入和检测算法。
英文文摘	Nowadays, through Internet and CD-ROM, the digital copies with the same quality as original data can be got much more easily and quickly. But the consequential problems such as copyright piracy, illegal use and copyright dissension have increasingly become a serious social issue. As a potential effective solution to the aforementioned problems, digital watermarking technology has received extensive attention. Because of the universality and important role of the digital image in the multimedia data, and it representative research process, the digital image watermarking is mainly taken as the research object. As a means of copyright protection, the digital watermarking will suffer various attacks, thus the watermarking for copyright protection should have the robustness. However, the robustness of the existing watermarking schemes against the geometrical attacks is still not enough, especially in the blind watermarking applications. The second generation watermarking, namely the content-based localized watermarking scheme, is promising to resist to the geometrical attacks. The basic idea of the content-based localized watermarking scheme is that, the robust feature point of the image is used to locate the watermark embedding and the watermark is independently embedded into the corresponding patch to each feature point so that the watermark can still be located and detected by the feature points of the part image, and thus can resist to the cropping attacks effectively. The content-based localized watermarking, also called the feature-based localized watermarking can logically divided into three parts: the detection of feature points, the construction of the elementary patches used to embed and detect the watermark, and the watermarking algorithm of elementary patches. The robustness of the whole watermarking scheme depends on the robustness of the feature points, the robustness of the constructed patches based on the feature point, and the robustness of the watermarking algorithm of the patch. Most of the current existing feature-based watermarking algorithms directly use the feature from the pattern recognition fields to synchronize the watermark. However, the requirements in the watermarking applications and the pattern recognition fields are not identical and the feature used for the pattern recognition fields is not necessarily suit for the watermarking applications. Thus, based on the mean luminance of the disk, a rotation and scale invariant feature （shortly named DRSIF）extraction algorithm used for the watermarking applications is proposed. For each point of the image, centered at which, firstly a series of concentric rings are generated and the absolute difference between the mean luminance of the inner ring and the mean luminance of the outer ring is computed for each ring when the inner radiuses of the rings come from a discrete geometric progression and the ratio of the outer radius to the inner radius of each ring is fixed, then the maximum of absolute mean luminance difference sequence is taken as the response value of the observed point, and the inner radius of the ring which attains the maximum of absolute mean luminance difference sequence is taken as the characteristic radius of the observed point. Finally the points whose response value attains the local maximum are selected as DRSIF points. In comparison to the two traditional scale invariant feature points, the time complexity of the DRSIF extractor is analyzed and the robustness of the DRSIF pints are verified. The experimental results show that the DRSIF points are robust against common attacks in the watermarking applications and are suit for the watermarking applications. The feature points are utilized to divide the image into the elementary patches used to embed and detect the watermark before the watermark is embedded and detected. However, the geometrical invariability of the feature points can not necessarily ensure that the constructed patches have the same geometrical invariability. The geometrical invariability of the patch also depends on the specific division, namely the way to construct the patch. How to construct the circular patches used to embed and detect the watermark is respectively elaborated. A partitioning strategy is proposed to improve the hit ratio of the embedding patches during the watermarking detection so that the speed to find the watermarked patch is greatly increased. Through experiments, the robustness of the elementary patch used to embed the watermark is verified and the proposed partitioning strategy during the watermarking detection is proved to be effective. Based on the aforementioned research achievements, two complete DRSIF-based localized watermarking schemes are proposed: one is DRSIF-based localized watermarking scheme in the spatial domain, and the other is localized watermarking scheme based on DRSIF points and invariant domain. In the former scheme, the moment of the patch is used to compute the characteristic angle of the embedding patch, which can reflect the rotation distortions suffered by the image, so that the rotation invariant of the patch is attained. In the latter scheme, the watermark is independently embedded into the rotation and scale （shortly named RS）invariant domain of the embedding patch. By combining the RS invariant domain with the idea of the localized watermarking, the watermark can resist to not only the whole geometrical attacks, but also the cropping attacks and such local geometrical attacks. Through experiments, the two watermarking schemes are proved to be robust. In comparison to the other algorithms, the performances of the proposed schemes perform better. The proposed two localized watermarking schemes are further compared. The robustness of the localized watermarking scheme based on DRSIF points and invariant domain is better than the other scheme because the latter requires more accurate repeatability of the feature points. In addition, the watermark volume of the localized watermarking scheme based on DRSIF points is less than the other. Thus, In the concrete application, the watermark embedding algorithms and detection algorithms are carefully selected by considering the requirement of the robustness and the watermark volume.