ProCoS: Degradation-Robust Manipulation Localization for Special Equipment Inspection Images

Manipulation localization in Special Equipment Inspection (SEI) images is crucial for trustworthy industrial supervision. However, existing image manipulation localization models struggle to handle the compound degradation disturbances commonly encountered in SEI scenarios. To address this issue, we propose a Prototypical Contrastive-Synergistic Network (ProCoS). The proposed model enhances consistent discrimination between clean and degraded observations through the Degradation-Resilient Synergistic Learning architecture, while introducing a Prototypical Contrastive Learning mechanism to improve stable representation under compound degradation. Furthermore, we construct SEI-Asym, a manipulation localization dataset for SEI scenarios, and establish a compound degradation evaluation protocol based on orthogonal experimental design. Experimental results show that ProCoS achieves an F1 of 0.6531, an AUC of 0.9670, and an IoU of 0.5485 on SEI-Asym, and reduces ΔF1¯ and (ΔF1)max under compound degradation to 0.0873 and 0.1966, respectively. The proposed model provides an effective technical pathway for trustworthy perception, anomaly discrimination, and industrial supervision based on SEI images.