MAMQ-Net: A Robust Framework Leveraging Multi-level Tamper-aware Queries and Complementary Representations of Tampering Features for Progressive Image Forgery Localization

Image Forgery Localization (IFL) aims to identify and localize the tampered regions within edited images. Many studies employ a dual-branch backbone to extract tampering features from dual modalities, followed by feature fusion at the final stage. In this process, the extraction and fusion of dual-modality features is relatively independent, which fails to fully leverage the complementarity between different modalities and thus diminishes sensitivity to tampering artifacts. Inspired by the way humans continuously integrate multi-faceted knowledge to understand the world, we propose MAMQ-Net, which contains a novel Multi-stage Alternating Feature Extraction and Interaction architecture. At each stage, we deeply explore the intrinsic relationships and mappings between different modality features. Feature extraction and interaction are performed alternately, constructing complementary dual-modality tampering feature representations and enhancing sensitivity to tampering artifacts. Additionally, we introduce a lightweight, Query-driven Multi-level Feature Decoding. This mechanism progressively aggregates key information from multi-level dual-modality tampering features through multiple sets of learnable tamper-aware queries, effectively filtering out irrelevant features. Finally, multi-level queries are used to refine discriminative features, enabling precise localization of tampered regions. Extensive experiments demonstrate that our framework outperforms current state-of-the-art models in localization accuracy and robustness across multiple public datasets, achieving a favorable balance between performance and efficiency.