面向图像迁移伪造的检测方法研究

The rapid development of multimedia technologies brings people a large amount of image integrity problems. Recently, image transfer techniques have achieved significant progresses. Unfortunately, there exists no research about the detection of image transfer based forgeries. Under current circumstances, if these fake transferred images are examined by experts’ eyes, the process will be time-consuming and thus not practical to the large-scale examination of images. Meanwhile, traditional fake image detection approaches are not generally applicable to this new kind of image forgeries. Therefore, to fill the void in this area, we propose to study the fake transferred images, including transferred scene images, transferred facial images and transferred artistic images, and develop specific detection schemes based on color, texture/structure and style cues. In this project, we will propose three detection techniques: an end-to-end fake scene image detection network via optimal color space transform and joint signal processing of 1D/2D prior knowledges, a fake facial image detection approach by analyzing the regional prior knowledges and modelling the texture/structure distortions, a fake artistic image detection method by extracting the deep style features and analyzing the color/texture prior knowledges. This project will provide specific solutions for detecting the fake transferred images, providing efficient approaches to detect potential fake transferred images in the cyber space, and laying a solid foundation for future researches.

随着多媒体技术的飞速发展，衍生出了大量有关图像可信度的安全问题。近年来，图像迁移伪造技术取得了突破性的进展，却并未有相应的检测方法被研究，而现有条件下如以专业人士来进行肉眼检测则过于耗时耗力，不适用于数据量较大的情况，传统的伪造图像检测方法又不普遍适用于检测此类新伪造图像。因此，为了填补领域的空白，本项目拟针对迁移伪造图像，具体研究其中的三种典型图像-场景迁移图像、人脸迁移图像、风格迁移图像的检测问题，基于颜色、纹理/结构、风格三种线索，研究检测算法群：1）基于检测颜色空间优化变换与1D/2D先验知识联合处理的场景迁移图像检测；2）基于区域性先验知识分析与纹理/结构失真建模的人脸迁移图像检测；3）基于深度风格提取与纹理/色彩先验知识分析的风格迁移图像检测。本项目研究成果将为迁移伪造图像的检测问题提供具体的解决方案，为网络空间中潜在的该类图像提供有效的检测手段，也为后续的检测研究奠定了基础。

随着多媒体技术的飞速发展，衍生出了大量有关图像可信度的安全问题。近年来，图像迁移伪造技术取得了突破性的进展，却并未有相应的检测方法被研究，而现有条件下如以专业人士来进行肉眼检测则过于耗时耗力，不适用于数据量较大的情况，传统的伪造图像检测方法又不普遍适用于检测此类新伪造图像。鉴于此，项目针对迁移伪造图像的检测问题，具体研究其中的三种典型图像-场景迁移图像、人脸迁移图像、风格迁移图像的检测及相关领域进行研究，提出了面向场景迁移图像、人脸迁移图像、风格迁移图像的伪造检测算法群、面向图像拼接伪造的篡改区域定位算法及多项相关领域算法，突破了风格迁移艺术伪造图像的检测关键技术，显著提升了场景染色伪造图像、人脸迁移伪造图像的检测精度与图像拼接伪造区域定位的正确率。项目团队累计发表学术论文18篇，其中IEEE/ACM Trans或CCF-A类论文12篇，申请中国发明专利4项，较好地完成了预期研究任务。本项目的研究成果为迁移伪造图像的检测问题提供了具体的解决方案，为网络空间中潜在的该类图像提供了有效的检测手段，也为后续的检测研究奠定了基础。