Loupe: image forgery detection and localization based on a multilevel feature collaboration bridge
TL;DR
This paper proposes an image forgery detection and localization method based on Multilevel Feature Collaboration Bridge (Loupe), using Vision Transformer as the backbone to extract multi-level features and designs a Feature Collaboration Bridge Network to realize efficient information interaction between image-level classification and pixel-level segmentation branches.
Abstract
The popularity of digital image editing techniques has made image forgery detection increasingly important. Existing methods suffer from insufficient robustness and generalization ability, as well as poor synergy between detection and localization tasks. This paper proposes an image forgery detection and localization method based on Multilevel Feature Collaboration Bridge (Loupe). Using Vision Transformer as the backbone to extract multi-level features, we design a Feature Collaboration Bridge Network to realize efficient information interaction between image-level classification and pixel-level segmentation branches via feature concatenation, attention fusion and gating mechanism. The model has only 6.6M parameters, with F1-score of 0.9110 and AUC of 0.9623, outperforming mainstream baseline models. The effectiveness of each component is verified by ablation experiments. Extensive tests under compression and noise interference verify its excellent robustness and generalization, making it suitable for complex real-world scenarios.