SBDNet: A deep learning-based method for the segmentation and quantification of fatigue cracks in steel bridges

Abstract

Employing deep learning to automate the processing of fatigue crack images in steel structure bridges is a cutting-edge research frontier in damage assessment and safe operation. However, existing methods lack pixel-level segmentation accuracy and quantitative metrics for model generalization. This paper introduces Steel Bridge Damage Networks (SBDNet), which is designed for high-precision pixel-level segmentation and quantification of fatigue cracks. We established metrics to measure domain differences between training and test sets and validated SBDNet on a fatigue crack image dataset, comparing its performance with state-of-the-art models. Results show that SBDNet achieves an average IoU of 76.8% and a crack geometric quantification error of less than 3%, exhibiting robust generalization. The proposed method enhances damage detection efficiency and provides quantitative references for maintenance decision-making.