Image Forgery Localization via Cross-Modality Trust Gate and Semantic Decoupled Amplification Decoder
Year: 2026
Type: conference-paper
Abstract: Image forgery localization is challenging due to diverse tampering techniques and subtle artifacts along manipulation boundaries. Existing multi-modality methods typically fuse RGB and noise features via naive concatenation or pooling, ignoring their pixel-wise reliability variance. Moreover, conventional decoders struggle with ambiguous edge semantics. To address this, we propose a novel network featuring a Cross-Modality Trust Gate (CMTG) and a Semantic Decoupled Amplification Decoder (SDAD). CMTG adaptively evaluates the spatial reliability of RGB and Noiseprint++ features for dynamic modal arbitration. During decoding, SDAD explicitly disentangles deep features into foreground, background, and boundary branches. A Discrepancy Amplification Unit (DAU) and an Interaction Module are then designed to recover high-frequency traces and enable cross-branch refinement. Furthermore, a Supervised Contrastive Loss is introduced to separate authentic and forged representations in the embedding space. Extensive experiments demonstrate our method's superiority in both localization accuracy and generalization ability.
Authors: Zhibin Li, Yifang Chen, Haoran Shen, Fei Shang
Institution: Guangdong Polytechnic Normal University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204160380