<p>Even though the current deep learning approaches for image tampering localization (ITL) have shown excellent results, they still encounter some challenges: poor localization results for small tampered regions; limited ability to recognize imperceptible tampering traces and insufficiently accurate localization of tampered edges; a tendency to focus on the image's foreground while ignoring its background. This paper suggests an image tampering localization network based on mask co-guidance and progressive edge (PEMG-Net) to overcome these challenges. In the two branches encoding stage, PEMG-Net extracts multi-scale semantic features from RGB image and inconsistent noise features through constrained convolution layer. We design a cross-attention feature fusion module (CFFM) for integrating semantic and noise features to obtain a more discriminative feature representation, fully utilizing the complementary advantages of features from different domains. To obtain detailed edge information, we employ an edge reconstruction module (ERM). The multi-scale fusion features are then guided by the edge guidance module (EGM) using the edge information to further optimize the tampered features. To achieve the prediction in the decoding stage, we take a step-by-step approach from high-level to low-level and introduce a progressive mask guidance module (PMGM). By using gradually improved masks and mask guidance strategies, PMGM deeply analyzes foreground and background, global and local features, and achieves high-precision localization of tampered areas. To validate the effectiveness of PEMG-Net, we conducted extensive experiments on five benchmark datasets. The experimental results show that PEMG-Net performs better than the state-of-the-art (SoTA) techniques in terms of both robustness and localization accuracy based on widely-used evaluation metrics.</p>