AFCMS-Net: Adaptive feature coupling and multi-level supervision network for effective image forgery localization

With the proliferation of forged images on the internet, the development of effective methods for localizing image forgery has become a research topic of increasing interest. Although deep learning-based models have generally demonstrated good performance, most focus only on Convolutional Neural Network (CNN)-based local information and ignore feature purification. This leads to feature redundancy and a lack of global context, resulting in inaccurate localization of tampered regions. In addition, they often overlook the importance of modeling the correlations between tampered and real regions in an image; thus, the extracted features lack discrimination. To address these challenges, we propose a novel method for improving the localization accuracy of multi-scale tampered regions by fusing multi-scale local and global critical correlations and enhancing feature discrimination. First, we use a multi-level Transformer to establish long-range dependencies between different regions in the images. This global information extraction capability contributes to a more comprehensive localization of tampered regions in an image. Second, we designed an Adaptive Selection and Interaction Aggregation (ASIA) module, which adaptively aggregates multi-level features and captures multi-scale hierarchical dependencies. This can enhance the representation of critical information and improve the accuracy of multi-scale tampered region localization. Third, the proposed Cross-domain Coupling Guide Refinement (CCGR) module can achieve complementary fusion and correlation enhancement of the local features and global representations based on the importance of different information. This fusion mechanism allows the model to consider both local details and global structure when locating tampered regions, which helps to improve localization accuracy and stability. Furthermore, we propose a multi-level intermediate supervision mechanism to compare the similarities and differences between tampered and real regions in forged images at multi-level features. It can learn more discriminative representations, thereby effectively enhancing the robustness of the model. Comprehensive experiments demonstrate that the proposed method has superior performance compared to most advanced techniques. It exhibits remarkable performance in localizing multi-scale tampered regions, even for post-processing and Online Social Network (OSN) attack images. The code of our proposed method can be available at https://github.com/SwallowIsXYZ/0617_AFCMS-Net