A Multi-Path Inpainting Forensics Network Based on Frequency Attention and Boundary Guidance

Abstract

With the continuous advancement of image-editing technologies, it is particularly important to develop image forensics methods for digital information security. In this study, a deep neural network called multi-path inpainting forensics network (MPIF-Net) was developed to locate the inpainted regions in an image. The interaction of shallow and deep features between different paths was established, which not only preserved detailed information but also allowed for the further mining of deep features. Meanwhile, an improved residual dense block was employed as the deep feature extraction module of each path, which can enhance the feature extraction ability of the model by introducing a frequency domain attention mechanism. In addition, a boundary guidance module was constructed to alleviate the prediction distortion in the boundaries of the inpainted region. Finally, extensive experimental results regarding various deep inpainting datasets demonstrated that the proposed network can accurately locate inpainted regions, exhibit excellent generalization and robustness, and verify the effectiveness of the designed module.