数字图像篡改定位若干关键性问题研究

Nowadays, digital image has become one of the most important carriers for exchanging information. With the rapid development on various image editing software, however, image tampering without leaving any visual artifacts became more and more easier, which would inevitably lead to some negative influences on individual privacy, legal forensics, scientific development and/or national security. Digital image forensic is now facing the severe challenge. Most of the existing literatures on image forensics are mainly focused on tampering detection, just a few works have been proposed for image forgery localization. This project will investigate several critical issues on digital image forgery localization as follows: ..1.Adaptive feature extraction and optimization for image forgery localization. Conventional localization methods usually divide the test image into rectangular regions for feature extraction without considering image contents and/or human visual attention. In this project, we try to analyze different sizes and shapes of the test regions for feature extraction, and extract the corresponding features based on their image contents adaptively...2.Multi-source information fusion for image forgery localization. Since several different forensic approaches would be adopted for image forgery localization, it is an important issue to fuse their detection results to generate the final localization result. In this project, we try to design some new fusion methods based on the properties of different forensic approaches, and further develop a universal fusion framework for different fusion cases. ..3.Refinement of image forgery localization results based on image semantic. Typically, image tampering in practice would modifies some semantic information within an image. It is expected that image semantics would provide some helpful information for forgery localization, which has not been well considered in previous literatures. In this project, we try to propose some novel methods to improve the conventional forgery localization methods based on image semantic analysis and understanding. ..4.Robustness analysis against JPEG compression for image forgery localization. JPEG compression is the most popular compression scheme for digital image. However, most previous works for forgery localization assume that the tampered image are stored in original uncompressed format, which may not be suitable for real applications. In this project, we try to propose some robust methods for forgery localization via analyzing various artifacts introduced by JPEG compression and their impacts on the existing localization methods. ..5.Image forgery localization with advanced machine learning techniques. Recently, some advanced machine learning techniques, such as deep learning, transfer learning, have made great progress in many research areas. However, their applications in image forgery localization have not been well studied. In this project, we try to develop a suitable deep learning framework for image forgery localization, and try to apply transfer learning to solve the cover-source mismatch problem and adapt some useful visual models for forgery localization.

如今，数字图像已经成为人们获取及传递信息的最主要载体之一。然而，随着各种图像处理工具的迅速发展，对数字图像篡改且不留下任何视觉上痕迹变得越来越容易，这将不可避免地对我们个人隐私、法律取证、科学发展乃至国家安全带来一定的负面影响。数字图像取证正面临着严峻的挑战。目前大部分取证方法考察的是篡改检测，而有关篡改定位的报道则相对较少。本项目主要围绕着图像篡改定位中的若干关键性问题进行展开，主要研究内容包括以下五个方面：1.自适应图像篡改定位特征的设计及优化； 2.篡改定位中的多源信息融合；3.结合图像语义信息的篡改定位优化方法； 4.篡改定位方法的鲁棒性分析与优化； 5.高级机器学习方法在篡改定位中的应用。拟通过本项目的研究，进一步完善数字图像篡改定位的理论和方法，并为数字图像取证实际应用提供技术支撑。