【作者】 胡媛媛；

【导师】 牛夏牧；

【作者基本信息】 哈尔滨工业大学 ， 信息安全， 2011， 博士

【摘要】 图像作为视觉的基本组成部分,是信息传播和知识积累过程中不可或缺的载体。然而,数字图像处理技术的进步给图像带来的诸多安全问题已经影响到公共安全等许多重要领域,使得针对图像的认证成为一个亟待解决的问题。图像的安全问题关键在于图像在应用过程中的基于内容的真实性和完整性,传统的数据认证和加密手段很难满足其应用需求。感知哈希基于媒体的感知内容,将媒体唯一地映射为一段数字摘要,具有对内容保持操作的鲁棒性和对内容篡改的区分性,并满足单向性、篡改定位性等要求,为媒体内容的识别和认证提供了一种有效的解决手段。图像的识别和认证本质上关注的是图像所传达的感知信息而非图像载体本身。现有图像哈希算法的特征提取至今停留在数据表达层面,没有上升到人类视觉系统的感知层面,从而导致现有图像哈希算法针对不同应用对象的鲁棒性、区分性和失真环境下的篡改定位性较差。本文从人类视觉系统的视觉模型出发,分析了结构化视觉模型与人眼视觉主观感受出现差异的原因,提出一种融合人类视觉系统的基本感知特性与整体结构化感知特性的视觉模型,进而分别针对原始域和变换域两类应用对象,基于该视觉模型,提出了一组能够实现良好鲁棒性、区分性和篡改定位性的图像哈希算法,为进一步增强图像哈希算法在失真环境下的篡改定位性,提出了一种通用的篡改定位模型,并将其应用到经典的图像哈希算法和设计新哈希算法上。本文的主要研究工作和创新点在于:(1)通过对人类视觉系统的基本构造和特性,以及传统视觉模型的分析,针对现有结构化模型没有考虑人类视觉系统亮度适应性和对比度掩蔽对视觉感知的影响这一局限性,将基于对视觉信息从人眼到大脑各个信息处理阶段模拟的传统视觉模型与人眼对图像的整体感受的结构化模型相结合,提出了一种新的人类视觉系统模型,实现了与人类视觉感知的高度一致性,并为哈希算法感知特征的提取提供了通用性的理论指导。(2)基于该视觉模型,提出了两种原始域的图像哈希算法。迭代尺度交互几何模型的原始域图像哈希算法通过与人类视觉反应最接近的Gabor滤波器提取人类视觉反应最为敏感的图像边缘信息,然后利用对旋转鲁棒的Radon变换将图像的边缘信息一维化生成哈希序列,从而获得了良好的鲁棒性、区分性和篡改定位性。该算法不受图像编码标准的约束,适用于各类图像。原始域环形模糊匹配图像哈希算法通过对特征点环形模糊匹配,有效提高了基于特征点的哈希算法的旋转鲁棒性。(3)基于该视觉模型,提出了两种变换域的图像哈希算法。基于DCT域的图像哈希算法通过直接从JPEG编码图像熵解码后的码流中提取感知不变特征,避免了DCT反变换,提高了哈希生成的效率,使得算法在保持了良好区分性的同时,不仅对高压缩率JPEG编码、直方图均衡化等内容保持操作非常鲁棒,而且还具备现有基于DCT域的图像哈希算法所缺乏的轻微旋转鲁棒性。基于DWT域的图像哈希算法通过提取DWT系数矩阵子带内的重要系数为感知不变特征,符合JPEG2000码流渐进式特点,实现了算法良好的鲁棒性、区分性和篡改定位性。(4)基于该视觉模型,提出了一种通用的篡改定位模型。图像在传输或存储的过程中,将不可避免地引入内容不变失真。现有哈希算法的篡改定位方案不能有效区分篡改和内容失真,从而导致他们的篡改定位性能随着图像失真程度的加大而迅速下降。本文提出的模型通过控制提取特征的鲁棒性,有效解决了提取特征鲁棒性与区分性之间的均衡性问题。基于该模型,现有图像哈希算法在失真环境下的篡改定位精度能够得到有效提高。更多还原

【Abstract】 Images, as an essential carrier of vision information, are indispensable in the process of information transfer and knowledge accumulation. However, many major fields such as public security have been suffered threaten from the security of images brought by the development of digitial image processing techniques, which made image authentication techniques to an urgently issue to be solved.The key issues of the security of images lie on the authenticity and integrity during the use. These requirments are no longer satisfied by the conventional methods of authentication and encryption. Perceptual hashing also refers as robust hashing and fingerprinting, maps digital multimedia data into a compact digital digest based on their perceptual contents. Multimedia with different contents would be mapped into distinct hashing sequence, while multimedia with same contents would be mapped into same hashing sequence values. Robustness to content-preserving manipulations, sensitivity to malicious tampering and security are the basic criteria of perceptual hashing. Perceptual hashing supplies an efficient solution to image identification and authentication. It has a wide application in content-based identification, retrieval, tampering localization and etc. However, feature extraction of existing hashing algorithms concentrates on the data expression instead of the perception of human vision system, which degrades their robustness to content-preserving manipulations, sensibility to content changes, and tampering localization in distorted images.In this dissertation, the reasons why the structural statistics based human man vision (HVS) model could not correspond well with the judgments of human observers are analyzed. A novel HVS model which fuses the classical HVS model and the structural statistics based HVS model is proposed. Then, in the guidelines of the fusion models of human vision system, hashing algorithms based on pixel domain, DCT and DWT domain are propsed. In addition, a universal tampering localization model to improve the performance of existing hashing algorithms on tampering localization in distorted images is brought out.The major works of this dissertation are described as follow:Fistly, based on the classical HVS models which imitate the information processing from human eyes to human brains and the model which measures the human observers’ feeling by structural statistics, the HVS fusion model is proposed. Experimental results show that, this fusion model is consistent with the judgements of human observers, serving as a feasible theroretical guideline for the extraction of robust features.Secondly, image hashing algorithms in the pixel domain are proposed. By Gabor filter, whose response is in accordance with HVS, the two dimensional (2D) edge features are extracted. Then, by Radon transform, which has high robustness to rotation, 2D edge features are transformed into one dimensional (1D) features. Last, by quantization of 1D features, hashing sequence is generated. Experimental results show that the above proposed algorithm is very robust to common content-preserving manipulations, sensitive to content changes, and has the performance on tampering localization. Since features are extracted in the pixel domain, the prposed hashing algorithms could be applied to images regardless of their storing formats. In addition, an hashing algorithms based on the fuzzy distance matching is proposed to improve the robustness of rotation for existing hashing algorithms with low level features.Thirdly, image hashing algorithms in DCT and DWT domains are proposed respectively. For the hashing algorithm in DCT domain, robust features are extracted from the bit stream which is entropy decoded for JPEG compressed images. Experimental results show that the proposed algorithm is not only robust to JPEG recompression, histogram equalization and some other content-preserving manipulations, but also robust to mild rotation attack, which is scarce for most existing hashing algorithms in DCT domain. For the hashing algorithm in DWT domain, significant bits in intra scales instead of inter scales are extracted, which is in accordance with the scalable property of JPEG2000 bits steam. Experimental results demonstrate the proposed algorithm excellant performance.Fourthly, a novel image hashing model for existing hashing algorithms to improve their performance on tampering localization in distorted images is proposed in this dissertation. Many image hashing algorithms have been proposed to detect the malicious tampering for content authentication. However, their tampering localization performance degrades dramatically on images with content-preserving distortion, as these algorithms cannot distinguish the malicious tampering from content-preserving distortion. In this framwork, high precision of tampering localization in distorted images is achieved by controlling the robustness of extracted features. By experimenting with classical image hashing algorithms, the correctness of the proposed model is proved.更多还原