Corrosion detection and area measurement of steel structural components based on semantic segmentation

Abstract

Corrosion of steel structures has become a critical factor affecting structural safety and service life, highlighting the urgent need for advanced technologies to achieve accurate corrosion detection and quantitative area assessment. This study proposed a deep learning-based intelligent detection and area measurement approach for steel structure corrosion. A total of eight models from three algorithmic families including You Only Look Once (YOLO) series, Mask R-convolutional neural network, and You Only Look At CoefficienTs were trained for pixel-level semantic segmentation of corrosion regions. To enhance the detection accuracy and robustness of small-scale corrosion areas, the YOLOv11 and YOLOv12 models were improved by introducing the hierarchical feature pyramid network feature fusion module and the convolutional block attention module (CBAM) attention mechanism, respectively. The optimal model was then identified through comparative analysis, and corrosion area quantification was performed based on semantic segmentation results. Finally, the optimal model and the area measurement algorithm were integrated to achieve intelligent corrosion detection and area evaluation for steel structures. The results demonstrate that the improved YOLO models achieve significantly better overall performance, making them well suited to corrosion detection tasks. In particular, the YOLOv12 model with the CBAM attention mechanism exhibits the best performance in detecting and segmenting corrosion regions, especially for small-scale corrosion.