多谱黎曼流形协同的润饰、变造数字图像定量鉴别算法研究

In latest years the technology of image editing and computer graphics has greatly prompted the application of digital image in virtual reality, digital entertainment, animation, film production, and social network etc., and provided the technical support for image retouch or alteration. It is considerately significant to identify the retouch/alteration image in fields of state security, internet, news, publishing, photography, law testimony and trust document.. There exists the inconsistency in the retouch/alteration image, including illumination discrepancy, prospective unmatch and geometry inconsistency. The retouch/alteration manipulation makes such inconsistency more subtle, more difficult to perceive. Moreover, such inconsistency makes the forgery image detection more complicated than ever.. The proposal covers the investigation on feature descriptor, the efficiency improvement on suspicious region identification, the quantitative identification on retouch/alteration image, and the appraisal on the authentication result.. The feature descriptor oriented for retouch/alteration detection would quantitatively and qualitatively represent the relative attribute and indirect information for providing the descriptor with strong subtle variation differentiation.. The proposal concentrates on the improvement to retouch/alteration identification, would evaluate the image data by multi-criteria, project the data under the constraint of mean overlap density, and store the data into multiple hash bins. The consistent appearance and context patches would be searched for locating the suspicious forgery region. The multiple hashing schema should be advantageous for providing the rich candidate patch set, and the successive Walsh-Hadamard projection would be helpful for collecting the resemblance patch with consistent appearance and context.. The proposal investigate on in Riemannian manifold space the approach of retouch/alteration representation, distance measure and parameter determination. In Riemannian manifold space, the retouch/alteration images show the isometry with the source one, the exploitation for the isometry is advantages for revealing the retouch/alteration variation. The proposal presents the reversely determination the concerned parameters from distance. The novelty is the quantitative identification on the retouch/alteration image.. An appraisal architecture for retouch/alteration image detection would be investigated by evaluating various visual features with compare the image rendered by the resolved parameters and the original one.

互联网+时代数字媒体文档的非授权篡改问题比以往更加严峻，鉴别润饰、变造图像在国家安全、互联网、新闻出版、摄影、法律证据、可信文档等方面非常迫切。润饰、变造图像破坏了图像内容的一致性，其定量鉴别是一个挑战性问题。申请书主要研究润饰、变造图像的定量鉴别问题，主要贡献是研究多谱黎曼流形协同的润饰、变造图像定量鉴别算法，包括以下四个研究点，1）调和编码多频谱的绝对和相对属性及直接和间接信息以提高描述子区分润饰、变造细微变化的能力；2）研究多测度度量、多投影和多哈希表存贮的耦合优化策略，以空间和内容约束搜索达到区域精准定位；3）研究润饰、变造图像的黎曼流形同态性、簇距离及度量问题，从度量距离解析反演润饰、变造参数；4）以众包方式和机器学习方法构建合理可行的多视觉特征结果评估体系。研究内容给出微尺度信息的高区分描述子、改善区域定位准确性和效率、提出定量鉴别的解析反演方法并展示结果的可信度。

互联网+时代数字媒体文档的非授权篡改问题比以往更加严峻，鉴别润饰、变造图像问题变得非常突出和迫切。. 项目研究给出描述润饰/变造图像区域微尺度信息的高区分描述子、研究润饰/变造的解析反演式定量鉴别方法以及润饰/变造数据集的生成方法。. 课题组在实验中发现候选图块实质承载着模板的某种变换，运用最优传输理论对模板和候选图块进行建模表示其间的非线性变换。提出了一种基于沃瑟斯坦Wasserstein重心坐标的图像描述子，并用于变造图像重复区域盲检测。结果表明，提出的基于Wasserstein直方图欧氏度量的变造图像重复区域盲检测算法能够检测出复制粘贴的伪造区域。. 针对复制粘贴常伴随着边缘模糊变造图像，提出了基于暗通道的边缘模糊复制粘贴篡改区域鉴定算法。依据边缘模糊区域的暗通道非零像素数高于其他区域，快速定位至粘贴物体的边缘模糊区域；在希尔伯特空间描述区域集，通过区域间的归一化相关系数确定复制源区域。. 提出了单张图像的L2型滤波核的微观多频谱分解方法，用主向量分析法提取影响突出的子频谱段滤波，利用谱理论扩展频谱至连续无限尺度空间，以多项式基线性调和表示子频谱段滤波，在子频谱段分析滤波的作用效果、子频谱段调谐、滤波核反演。. 项目利用对抗生成网络GAN生成新润饰/变造数据，一定程度解决了润饰/变造图像数据集量小、种类少、尺寸小的问题。研究了应用CycleGAN进行图像风格迁移的算法；提出了几何和纹理相协调的3D建筑物生成方法。. 发表SCI论文20篇，CCF A类论文2篇，EI论文15篇，中文核心期刊论文10篇，申请发明专利3项，培养了14名数字媒体取证方向的研究生。. 数字图像的润饰/变造定量鉴定是一个道高一尺魔高一丈的哲学辩证问题，润饰/变造图像的定量鉴别具有挑战性。项目以高斯模糊算子的定量鉴定给出了解析反演的理论探索；在神经认知层面探索了润饰/变造图像的定量鉴定机理；探索了基于大数据利用深度学习润饰/变造的本征性问题。