基于篡改评价模型的图像真伪盲鉴别方法研究

The issues of image authenticity have infiltrated politics, economy, civilization, military affairs and so on, which results in a lot of negative effects. Nowadays, accurate blind identifications of digital images depend on the knowledge of the used tampering method. However there is little work on developing model for evaluating which tampering method has been used on certain images. Thus this project seeks to develop (i) an evaluation model for analyzing which tampering method is used on the image and (ii) identification algorithms for identifying which part of the image is being tampered. Evaluation model of image tampering will be established to analyze the tampering operations qualitatively. As for identification algorithms, blind identification algorithm will be researched and realized for recaptured images, which is based on the model of imaging process and the model of physical characteristics. For spliced images, we will research on the blind identification algorithm based on the model of the consistence of natural image features. For copy-paste images, the feature vectors of SIFT key points will be constructed, which are made up of the marked graph, and then the blind identification algorithm robust to rotation and scale will be researched. Finally, we will combine the evaluation model of image tampering and the blind identification algorithms, so the tampered regions of an image can be located accurately after the qualitative analysis of tampering methods. The technology indicators of this project are expected to: the accuracy rate of evaluation model can reach up to 80%; the average accuracy rate of our identification algorithms can be more than 85%. This project can not only open new perspectives and introduce new theoretical framework for the research of blind identification of digital images, but also improve its practicality in the field of judicature and forensic.

图像真伪问题已渗透到政治、经济、军事、文化等各大领域，严重影响了人们正常的生活和生产秩序。目前数字图像盲鉴别算法需已知篡改手段，仍处于零散技术的研究阶段。基于此，本项目深入研究基于篡改评价模型的图像盲鉴别方法及应用。建立基于自然图像特征的篡改评价模型，定性分析篡改手段。针对重获图像，研究基于成像过程模型和物理特征模型的盲鉴别算法；针对拼接篡改图像，研究基于自然图像特征一致性模型的盲鉴别算法；针对复制粘贴篡改图像，构造SIFT关键点标记图特征向量，研究具有旋转、缩放不变性的盲鉴别算法。将图像篡改评价模型与盲鉴别算法相结合，在定性分析篡改手段的基础上，精确定位篡改区域。 本项目的预期技术指标为：图像篡改手段定性判断的准确率达80%，盲鉴别算法的平均正检率达85%。本项目不仅为数字图像盲鉴别技术的研究开辟了新的视角，引入了新的理论框架，而且提高了数字图像盲鉴别技术在司法取证等领域的实用性

本项目研究了基于篡改评价模型的图像盲鉴别技术的基础理论与方法，其主要成果包括：. 1. 为了实现图像篡改手段的定性分析，提取图像的LTP三值模式特征、LBP纹理特征和WLD局部特征，构造了基于图像统计特征的篡改评价模型，能够在盲环境下实现对自然图像、计算机生成图像、复制-粘贴图像、拼接篡改图像、重获图像的多分类。. 2. 针对重获图像的盲鉴别，一方面分析其与真实图像成像过程中所产生的差异，定义局部平面线性点，利用这些点的性质计算特征向量。另一方面，分析重获图像与自然图像的统计特征差异，提取小波域特征和噪声特征。分别利用SVM分类器实现了重获图像的盲鉴别。. 3. 对于拼接篡改图像的盲鉴别，为了提高算法的检测率并降低计算复杂度，从以下三个方面提取图像的统计特征：(1) 利用视觉注意模型获取图像的关键特征片段，并提取其扩展的DCT域的HMM特征；(2) 对原始图像、预测误差图像和基于离散小波变换的重构图像，建立模糊游程特征向量；(3) 计算基于差分分块离散余弦变换矩阵的灰度共生矩阵(GLCM)，提取其二阶矩、熵、对比度、自相关、逆矩和逆差矩六种纹理特征。. 4. 对于复制粘贴图像的盲鉴别，从特征点匹配的角度，在SIFT关键点邻域内构造基于最大视角的标记图特征向量，实现了复制粘贴篡改的盲检测。为了降低误匹配，在高斯尺度空间提取oFAST特征点和ORB特征，取得了较好的检测效果。为了提高算法对光照变换的鲁棒性，利用混合灰度序模式(MIOP)特征描述高斯差分区域，实现了鲁棒性较好的盲鉴别算法。从图像块匹配的角度，定义图像的灰度级别和灰度结构，以此定位可疑块对，并利用对数极坐标变换，将复制-粘贴区域的旋转、缩放变换转换成平移变换，实现了具有几何不变性的盲鉴别算法。另外，针对隐蔽性复制粘贴篡改, 将彩色信息与LBP纹理特征的融合，并提取图像块的GLCM，实现了复制-粘贴区域的精确定位。. 5. 结合以上评价模型及相关算法，建立了图像篡改盲鉴别系统的原型。. 本项目的研究结果为数字图像盲鉴别领域的研究提供新思路和理论依据。