Reproduce “MVSS-Net: Multi-View Multi-Scale Supervised Networks for Image Manipulation Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chengbo Dong, Xinru Chen, Ruohan Hu, Juan Cao, Xirong Li
MVSS-Net++ introduces a multi-view multi-scale supervised network for image manipulation detection, leveraging both noise inconsistencies and tampering boundary artifacts to extract semantic-agnostic features. The model achieves superior generalization, demonstrating a mean pixel-level F1 of 38.7% (compared to 21.4% for the best baseline) and an image-level F1 of 68.0% across diverse datasets, significantly reducing false alarms on authentic images.
Key insights
- Jointly exploiting both noise inconsistencies and explicit tampering boundary artifacts provides more generalizable, semantic-agnostic features for manipulation detection.
- Multi-scale supervision, particularly integrating an image-level classification loss, is crucial for improving model specificity and reducing false alarms on authentic images.
- ConvGeM effectively addresses limitations of standard pooling methods by allowing more pixels to contribute to image-level predictions while incorporating spatial correlation through a decaying convolutional block.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.