Learning to localize image forgery using boundary-preserving mask R-CNN

Digital image manipulation is a growing concern in multimedia security. Many existing forgery detection methods struggle with accurately localizing manipulated regions-especially near boundaries-and often fail to generalize across different manipulation types. To address these challenges, we propose a novel Boundary-Preserving Mask R-CNN framework that enhances detection precision by incorporating boundary-aware features. The model integrates channel attention mechanisms to better capture detailed spatial information and leverages frequency domain features to improve robustness. We evaluated the framework on six diverse benchmark datasets-CASIA V2, Columbia, Carvalho, CoMoFoD, MICC-F220, and CG-1050-covering splicing, copy-move, and compositing manipulations. Extensive preprocessing ensured uniform input, and pixel-level segmentation enabled accurate region detection. Our method demonstrated strong performance across multiple metrics, including accuracy, precision, recall, F1-score, IoU, and AUC. These results highlight its potential as a reliable tool for digital forensics and investigative applications.