Edge-Enhanced Multi-Scale Fusion for Pixel-Level Image Tampering Localization

Image manipulation localization has recently advanced through the integration of multi-scale feature representations and edge‑aware supervisory mechanisms. However, existing methods often struggle to simultaneously preserve global semantic consistency and fine‑grained boundary precision, particularly in challenging low‑texture or visually cluttered regions. To address these issues, we propose MEP‑Net, a lightweight yet highly effective manipulation localization framework that incorporates Multi-branch Feature Enhancement Fusion Module, edge positioning supervision, and multi‑scale prediction fusion. Through comprehensive experiments and ablation studies, MEP‑Net demonstrates competitive performance against both CNN‑based and recent Transformer‑based detectors, while maintaining strong boundary accuracy and computational efficiency.