UMI-Net: Unveiling Multi-Domain Inconsistency Learning with Structure Guidance for Image Manipulation Localization
Year: 2026
Type: article
Abstract: The growing demand for digital image authenticity verification has driven advances in Image Manipulation Localization (IML). Although existing methods aim to capture tampering-induced inconsistency features, such cues are often subtle. They can be obscured by representations dominated by high-level semantic information, thereby limiting the model’s ability to detect fine-grained anomalies. To address this challenge, we propose UMI-Net, an inconsistency-driven framework for image manipulation localization. Specifically, UMI-Net comprises two core components: an Inconsistency-Aware Encoder (IAE) and a Structure-Constrained Decoder (SCD). The IAE employs a Dynamic Frequency-Domain Gating Module (DFG) and a Cross-Domain Inconsistency Interaction Module (CII) to capture cross-domain inconsistency responses under semantic structure constraints, thereby learning robust inconsistency representations. Subsequently, the SCD introduces global constraints through a Multi-Scale Structure Prior Extraction Module (MSPE) and progressively refines the localization results using a Structure-Aware Refinement Module (SARM), effectively filtering out background noise and reconstructing clear and continuous tampering boundaries. Extensive experiments on multiple benchmark datasets demonstrate that UMI-Net outperforms existing methods in localization accuracy and robustness in most cases, while also showing promising potential for applications in image authenticity verification and multimedia forensics.
Source: Electronics
Institution: Anhui University of Science and Technology
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7234395045
