Complementary attention feature disentanglement network for image tampering detection and localization

In the image tampering detection and localization task, the existing methods face significant challenges in disentangling tampering artifacts and interfering semantic information, which seriously restricts the robustness and versatility of the model. To address this problem, we propose a complementary attention feature disentanglement network. We innovatively adopted multiview feature extraction and a complementary attention disentanglement representation learning strategy to efficiently disentangle tampering artifacts and content information. Specifically, the two-stream feature extractor captures the heterogeneous features of the RGB visual stream and the Spatial Rich Model noise stream in parallel. Using the hierarchical progressive strategy and the hybriy atrous pyramid enhancement module, the multilevel fusion encoder realizes the effective fusion of low, middle, and high cross-modal information, which significantly improves the richness of features and the expression ability of tampered information. The complementary attention feature disentanglement module generates the complementary attention map through the collaborative operation of spatial and channel dual attention mechanisms to strengthen the artifact features and content features. The decision module uses four binary detectors for image-level tamper detection and locates the tampered region precisely based on the artifact features and the ground truth mask. We conducted extensive experiments on six representative benchmark datasets. Experimental results show that complementary attention feature disentanglement network exhibits superior robustness and adaptability in postprocessing and across datasets compared with the current state-of-the-art methods.