Automated Corrosion Detection in ASTM A572 Grade 50 W-Beams Using U-Net and Attention U-Net: A Comparative Analysis for Semantic Segmentation

Corrosion in steel structures represents a critical challenge in civil engineering, with economic losses exceeding 2.5 trillion dollars annually worldwide. Traditional visual inspection methods present significant limitations, including subjectivity, dependence on inspector experience, and difficulty accessing critical structural elements. Automated detection systems using deep learning offer a promising solution to enhance inspection objectivity and efficiency. This study presents a rigorous comparative analysis between U-Net and attention U-Net architectures for automated corrosion detection in ASTM A572 Grade 50 W-beams. A comprehensive data set was systematically developed comprising 217 original high-resolution images ( 512 × 512 pixels ) captured under controlled laboratory conditions from W-beam specimens with varying corrosion severity levels (6 months to 5 years of atmospheric exposure). After data augmentation, the final data set contained 414 images for model training and evaluation. Manual annotations were performed by three structural pathology specialists with interannotator agreement of κ = 0.87 . The attention U-Net architecture demonstrated statistically superior performance across all metrics: mean IoU of 0.775 ± 0.089 versus 0.693 ± 0.078 for U-Net (improvement of 11.8%, p < 0.001 , Cohen’s d = 0.98 ); Dice coefficient of 0.741 ± 0.067 versus 0.678 ± 0.071 (9.3% improvement); and F1-score of 0.823 ± 0.054 versus 0.751 ± 0.063 (9.6% improvement). False positive rate was reduced by 46% (12.8% versus 23.4%). For pitting corrosion detection, attention U-Net achieved 87.3% detection rate compared to 71.2% for classical U-Net. These results conclusively demonstrate that incorporating attention mechanisms substantially improves automated corrosion detection capability, offering a practical tool for nondestructive inspection of ASTM A572 Grade 50 steel structures with accuracy approaching interexpert agreement levels [0.80–0.85 intersection over union (IoU)].