【作者】 吴琼；

【导师】 李国辉；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2008， 博士

【摘要】 随着现代数字技术的发展和图像处理工具性能的日益强大,数字图像极易被篡改并使得人眼难以觉察出伪造的痕迹,因此迫切需要对图像的真实性进行鉴别。传统的数字签名和数字水印鉴别技术都需要内容提供方对图像进行预处理(提取签名或嵌入水印),从而导致两者的应用受到了限制。针对上述应用需求和技术需求,本课题研究一种刚刚兴起的鉴别技术—数字图像盲取证。数字图像盲取证是一种不依赖任何预签名提取或预嵌入信息来鉴别图像真伪和来源的技术,它有着广泛的应用前景,正逐步成为多媒体安全领域新的研究热点。本文围绕数字图像盲取证中的真实性检测问题,采用理论分析、算法设计和实验验证相结合的方法,对数字图像盲取证的理论和技术进行了深入研究。论文的主要研究内容及创新点如下:数字图像盲取证系统框架的研究。数字图像盲取证技术的发展处于初级阶段,目前尚未形成统一的、成熟的系统框架。本文在总结现有研究成果的基础上,提出了一个数字图像盲取证的基本框架,它包括图像建模、特征提取与特征分析、算法设计、测试与验证、图像篡改检测定位与图像分类、相关的图像源特性以及图像数据库等主要组成部分。该框架为本文的研究提供了理论指导,对今后的盲取证工作也具有一定的借鉴意义。复制-粘贴伪造图像的盲取证方法研究。针对在同一幅图像中复制部分特定区域来覆盖伪造目标区域的复制-粘贴伪造类型,本文提出了一种基于小波和奇异值分解的图像盲取证算法。该算法将复制-粘贴伪造检测问题转化为相似块对的匹配问题,利用小波变换提取图像的近似分量作为分析对象,并对其进行滑窗分块操作,对图像块进行奇异值分解和量化,然后对由所有块的量化奇异值特征组成的特征矩阵按行进行字典排序,最后结合相似图像块对的偏移频率信息,对复制-粘贴伪造区域进行检测和定位。实验结果表明,该算法能够有效地检测并定位出图像的复制-粘贴篡改区域,对JPEG压缩和高斯噪声具有较好的鲁棒性,并且具有较高的检测效率。纹理合成修复伪造图像的盲取证方法研究。针对利用纹理合成图像修复技术进行图像篡改的修复伪造类型,本文首次提出了一种基于零值连通和模糊隶属度的图像盲取证算法。该算法利用零值连通特征来刻画修复伪造图像中异常的块对相似性,然后引入模糊理论中的隶属函数,将这种块对的相似性转换成待检测块属于篡改块的隶属度,最后通过截集划分,对伪造区域进行检测和定位。实验结果表明,该算法能够对多种修复方法生成的伪造图像进行有效的检测并对篡改区域进行准确的定位,同时对JPEG压缩和高斯噪声具有一定的鲁棒性。拼接伪造图像盲取证方法的研究。针对从一幅或多幅图像中剪切部分区域,将其拼接到另一幅图像中并进行必要的后处理的拼接伪造类型,本文提出了一种基于自然图像统计特性的图像盲取证算法。该算法将拼接图像检测问题看作是一个两类模式识别问题,从分析自然图像的统计特性出发,利用广义高斯分布模型对图像小波细节子带系数的统计分布进行建模,提取模型参数及模型预测误差作为特征;同时利用马尔可夫链对图像离散余弦变换系数之间的相关性进行建模,提取模型的状态转移概率矩阵作为特征;然后将两部分特征合并形成图像的统计特征向量,采用支持向量机实现了对自然图像和拼接图像的有效分类。利用该算法对三个公开的拼接图像数据库进行了测试,实验结果表明算法具有较高的分类准确率,从而验证了特征的有效性。综上所述,本文的主要工作集中在数字图像盲取证的系统框架和方法的研究上,在理论和应用上都取得了一定的成果,这些成果将对多媒体盲取证技术的发展产生积极的推动作用。更多还原

【Abstract】 With the development of modern digital technology and the availability of increasingly powerful image processing tools, digital images are easy to be manipulated without leaving obvious visual traces of having been tampered, so there is an urgent need to identify the authenticity of images. The applications of traditional digital signature and watermarking authentication technologies are limited, because they require the content providers to pre-process the images, such as extracting signature or embedding watermark. Due to the requirements of application and technology mentioned above, this thesis focuses on blind digital image forensics, which is an emerging authentication technology. As a technology of detecting image authenticity and source without relying on any pre-extracted or pre-embedded information, blind digital image forensics is becoming a new hotspot with broad prospect in the multimedia security area.This thesis makes an in-depth research on the authenticity detection problem of blind digital image forensics by applying the combined methods of theory analysis, algorithm design and experiment validation. The main contents and innovations are as follows:Firstly, the basic framework of digital image forensics is studied. As the research of blind digital image forensics technology is still in its infancy, there is no unified and mature architecture. Based on the recent developments in this field, a basic framework of digital image forensics is proposed, which consists of image modeling, feature extraction and analysis, algorithm design, test and verification, forgery area localization and image classification, image source characteristics as well as image database. The framework provides a theoretical guidance for the research of this paper, and has some reference to the future work of digital forensics.Secondly, a blind forensic approach for detecting copy-paste images is studied. The copy-paste forgery is to copy a particular part of a digital image and to cover another part of the same image. A blind image forensic algorithm based on wavelet and singular value decomposition is proposed to detect the specific forgery. In this algorithm, the copy-paste forgery detection is translated into a matching problem of similar block pairs. The wavelet transform is applied to extract the approximate component of the image, on which the sliding window operation is used. Then the singular value decomposition and quantization are adopted to extract characteristics of the fixed-size image blocks. The quantized singular value vectors are lexicographically sorted and the copy-paste forgery regions are localized by detecting all neighborhood vectors. The experimental results demonstrate that the proposed approach can detect and localize the copy-paste forgery regions accurately, and has good robustness to JPEG compression and Gaussian noise. In addition, the efficiency of our approach is improved significantly.Thirdly, a blind forensic approach for detecting inpainted image based on texture synthesis is studied. The technique of image inpainting can be used to remove objects from an image and play visual tricks. As a first attempt, a blind image forensic algorithm based on zero-connectivity feature and fuzzy membership is proposed to detect the specific forgery. Zero-connectivity labeling is applied on block pairs to yield matching degree feature of all blocks in the region of suspicion, and fuzzy memberships of these blocks are then computed by constructing a membership function. Then the tampered regions are identified by a cut set. The experimental results show that the proposed approach can effectively detect the inpainted images, which are generated by a variety of image inpainting methods, and it can localize the tampered region accurately. Furthermore, the approach has robustness to JPEG compression and Gaussian noise to a certain extent.Finally, a blind forensic approach for detecting spliced image is studied. Image splicing is a process of cropping and pasting regions from different images to form another image with necessary post-processing. A blind image forensic algorithm based on statistical characteristics of natural images is proposed to detect this specific forgery. In the algorithm, the splicing detection can be treated as a two-class pattern recognition problem. On the one hand, the generalized Gaussian distribution is adopted to model the statistical distribution of wavelet details subbands of images, and the model parameters and prediction error of each wavelet details subband are extracted as features. On the other hand, the Markov chain is applied to model the correlation of discrete cosine transform coefficients, and the state transition probability matrix is extracted as feature. Then the two kinds of features are combined to form natural image statistical feature vector, which is used to distinguish natural images from spliced images using support vector machine. The proposed algorithm is tested on three public image splicing detection databases, and the experimental results show that the algorithm has a high classification accuracy, which verifies the effectiveness of the proposed features.In conclusion, this thesis mainly focuses on the research of the systematic framework and methods for blind digital image forensics. Some achievements have been made in theory and applications. These achievements will play a positive role in promoting the development of blind multimedia forensics.更多还原