Synergistic context edge-enhanced HRNet for steel bridge corrosion segmentation

Accurate segmentation of bridge corrosion is critical for structural health monitoring and maintenance decision-making. However, existing semantic segmentation methods often suffer from inter-class ambiguity among corrosion severity classes and blurred boundaries of corrosion regions when confronted with complex corrosion morphologies and visually cluttered backgrounds. To address these challenges, this study proposes a Synergistic Context Edge-Enhanced HRNet (SCE-HRNet). The proposed model embeds a Pyramid Contextual Attention (PCA) hierarchy into the backbone to enhance multi-scale contextual feature interaction, thereby improving the discrimination of corrosion regions under complex environmental conditions and mitigating inter-class ambiguity between similar corrosion severity classes. In addition, an Edge Enhancement Module (EEM) is introduced to explicitly sharpen blurred boundaries and recover fine structural details. Experimental results on the Corrosion Condition Rating Database (CRD) and the Corrosion Condition State Semantic Segmentation Dataset (CSSD) demonstrate that SCE-HRNet achieves mIoU improvements of 2.72% and 2.68%, respectively, over HRNet and consistently outperforms several mainstream segmentation models. However, the integration of contextual attention and edge refinement introduces additional computational overhead in terms of FLOPs and parameters. This creates a trade-off that limits the model’s strictly real-time applicability on resource-constrained edge devices for field inspections, suggesting a need for future model pruning or quantization. Moreover, this study is specifically optimized for steel bridge corrosion, and its generalizability to non-steel structures or other material degradation types remains a primary limitation that requires further evaluation with more diverse cross-domain datasets. Overall, the results validate the effectiveness of incorporating contextual modeling and boundary-aware mechanisms into high-resolution architectures for improving corrosion segmentation accuracy and boundary quality, and provide valuable methodological insights for future model optimization toward practical inspection applications.