Image Tampering Detection Based on Multi-Scale Adaptive Dense Residual Feature Fusion Network

Rapid advances in generative AI and image editing have fueled widespread image forgery, challenging authenticity and security. Traditional splice detection suffers from high false positives, while deep learning models lack precise localization and segmentation of tampered regions. To address these limitations, this paper proposes a multi-scale adaptive dense residual feature fusion Network (ADRF-Net), which integrates a multi-scale feature fusion and an attention dense module. The fusion module aggregates and refines features across spatial resolutions, concurrently capturing subtle local tampering traces and global contextual patterns to enable pixel-precise localization. The weighted attention dense Module synergizes residual learning with local similarity-based attention mechanisms, which amplify discriminative feature learning for tampering artifacts. The framework enhances tampering trace preservation through context-aware refinement modules and optimizes gradient propagation across encoder-decoder layers. Experiments on CASIA datasets demonstrate that ADRF-Net achieves an 8.1% higher precision than state-of-the-art counterparts, while simultaneously excelling in pixel-level localization of tampered regions.