【作者】 王波；

【导师】 孔祥维； 尤新刚；

【作者基本信息】 大连理工大学 ， 信号与信息处理， 2010， 博士

【摘要】 随着人类社会进入数字时代,数码相机的普及,简单实用的图像编辑软件的广泛流传,使得制作篡改伪造数字照片图像的门槛不断降低。近年来在政治、经济、新闻等各个领域发生的,由篡改伪造数字图像频繁引发的纠纷和争端,也引起了人们对于如何保证数字时代多媒体内容真实性的深深忧虑。于是,关心数字图像完整性和真实性的数字图像取证近年来成为了学术界的一个研究热点,而其中仅利用数字图像本身特性进行取证分析的数字图像被动盲取证更是成为重中之重。本文针对数字图像取证研究中仍然存在的诸多关键问题,以数字图像的相机成像引入特征为切入点,从数字照片图像的相机来源鉴别、拼接伪造图像的定位检测以及模糊篡改伪造操作的定位检测三个方面,对数字图像被动盲取证的理论和方法进行了研究,主要工作和成果总结如下：在数字照片图像的相机来源鉴别方面,首先分析了现有的数字图像相机来源鉴别模型和方法的局限性,着重指出现有方法在相机样本较多的情况下检测准确率低,以及无法对训练模型外型号的数码相机拍摄图像进行来源检测的两个主要问题,并分析了其原因所在。在此基础上从数码相机成像过程的视角出发,利用成像过程中的关键步骤CFA插值对数码照片图像进行来源鉴别。提出了一种基于协方差矩阵的CFA插值系数估计方法。以此为特征向量,提出了基于多类分类器的数字图像相机来源鉴别,以及基于一类和多类分类器联合的数字图像相机来源鉴别方法。在拼接伪造图像的定位研究方面,对目前的拼接图像检测方法进行了分析。在此基础上,给出了基于原图估计的拼接伪造图像定位模型。通过对图像进行原图估计,恢复和重建图像中可能因拼接操作而被破坏的像素邻域连续性和相关性,进而对拼接区域进行像素级定位。在此基础上,提出了基于“软成像”的原图估计框架,并给出了CFA插值、JPEG压缩和自适应CFA插值三种拼接伪造图像像素级定位检测方法。在模糊篡改伪造的定位检测方面,首先指出由于数码相机的成像机理,使得数码照片图像的局部邻域存在色彩一致性,并进一步分析了模糊篡改伪造操作对这种成像引入色彩一致性特征的影响。在HSI色彩空间的色调域中,提出采用异常色调来对图像的局部邻域色彩一致性进行量化描述。进而提出了基于异常色调率和基于色调特征的两种模糊篡改定位检测方法。更多还原

【Abstract】 The popularization of digital camera and the wide spread of sophisticated image processing software, make image forgery creation easier and easier in the digital era. In recent years, people show deep worries about the authenticity of multimedia, because of the frequent disputes caused by the falsification of digital images in various areas such as politics, economy and news. As a result, digital image forensics, which concerns the integrality and authenticity of images, receives considerable attention these years. Consequently, the passive and blind digital image forensics, which only uses the image for analysis, is becoming a hot topic for the academic research. Aiming at the various problems in this field, a study on passive and blind digital image forensics has been made on the aspects of source camera identification of digital images, splicing localization, and blur localization, based on the features introduced by image formation. The main contribution of this dissertation is as follows:As to the source camera identification, an analysis on the limitation of current model for source camera identification is presented firstly, and then two problems are pointed out. One is the decreasing accuracy of the source camera identification as the number of the camera models increasing. The other is that the existing methods are incapable of the outlier models of the training model. Under the perspective of digital image formation in digital camera, CFA interpolation, the key operation in image formation, is utilized for source camera identification. A method for estimating CFA interpolation coefficients is proposed based on covariance matrix. Furthermore, two algorithms, which are based on multi-class classifier and the combination of multi-class and one-class classifier, for source camera identification are presented respectively using the CFA interpolation coefficients as feature vector.As to the study of splicing localization, the existing methods of splicing detection are reviewed and analyzed. Furthermore, a general model of splicing localization for spliced forgeries is proposed. By estimating the original image counterpart, the coherence and consistency in the neighborhood of the pixels, which may destroyed by splicing, are recovered for localizing the tampered area. Based on this theory, a framework using’soft imaging’ is presented for original image counterpart estimation, and three approaches of soft imaging, which are CFA interpolation, JPEG compression and adaptive CFA interpolation, are proposed respectively, to localize the splicing area in pixel-level. As to the study of blur localization, an observation is firstly pointed out that the coherence and consistency exist in color space in the neighborhood of the pixels introduced by the image formation, and also a theoretical analysis of the effects caused by blur tampering is made. On the basis of these work, the abnormal hue in HSI color model is proposed to describe the coherence of the color space in the neighborhood of pixels. Moreover, two methods based on abnormal hue and hue characteristics are presented to localize the blur tampering.更多还原