UFG-Net: Uncertainty and frequency guided network for image forgery localization

In Image Forgery Localization (IFL) tasks, post-processing of manipulated images often leaves subtle traces, increasing model uncertainty in predicting tampered areas. Current methods fail to leverage these uncertainties to extract key features effectively. Additionally, important information in the frequency domain is often overlooked. However, existing methods that merely combine frequency and spatial domain features can lead to data redundancy and feature misalignment across domains. To address these challenges, we propose the Uncertainty and Frequency Guided Network (UFG-Net), which integrates uncertainty information during training and utilizes frequency domain insights to enhance feature extraction in the spatial domain. In addition, the Uncertainty and Frequency Guidance (UFG) module is introduced to optimize feature extraction using a multi-scale approach with dual attention mechanisms, while the Forgery Inference (FI) module is designed to enhance final localization precision. Extensive experiments across five diverse datasets demonstrate UFG-Net's exceptional capabilities in IFL. Notably, on the IMD dataset, UFG-Net achieved improvements of 1.8% in AUC, 1.7% in F1-score, and 2 % in IoU, outperforming the second-best models across all metrics. Moreover, cross-dataset experiments and robustness tests further underscore UFG-Net's outstanding generalization and robustness.