ScholarWorks@Y-Scholar Hub

초록

Image Manipulation Detection and Localization (IMDL) aims to identify tampered images and their altered regions. Existing RGB-centered approaches often overemphasize RGB information while overlooking complementary insights from noise-view modalities. This reliance on RGB limits their ability to detect subtle manipulation traces. To overcome these challenges, we propose Multi-Noise-View Fusion (MNVFusion), a framework that balances the contributions of RGB and noise-view modalities using a multi-branch encoder structure. MNVFusion incorporates the Multi-Branch Channel Mixing Module (MB-CMM), enabling efficient channel-wise fusion to integrate diverse modality features. Additionally, we introduce Fixed GeM, a training-free image-level detection module that enhances overall efficiency through fixed operations on localization maps. Experiments on six benchmark datasets show that MNVFusion delivers state-of-the-art performance in both detection and localization tasks.