Optimized deep learning for steel bridge bolt corrosion detection and classification

Abstract

Corrosion of steel bridge bolts poses a major threat to structural integrity. As a key component, bridge bolts are susceptible to corrosion and detachment, necessitating close monitoring during maintenance. However, the vast number of corroded bolts presents challenges. Conventional detection relies on subjective visual inspection without unified assessment criteria. Fast and accurate localization and classification of bolt corrosion levels thus became a research priority for bridge maintenance. This study puts forward a new two-stage corrosion detection method, integrating Deep Convolutional Generative Adversarial Network (DCGAN) and improved You Only Look Once version 5 (YOLOv5) networks, which enables rapid categorization of bridge bolts by different corrosion extents, providing an intelligent solution for bridge maintenance. First, in the image preprocessing stage, a DCGAN with least squares loss generates simulated corroded bolt images through adversarial learning to expand the dataset. Second, three corrosion levels are defined based on visual features and percentage of surface rust, with this classification scheme applied in dataset annotation. Finally, the annotated images are input into an enhanced YOLOv5 model named YOLOv5-SE for corrosion identification. To improve accuracy, Squeeze-and-Excitation blocks are added after each residual block in the YOLOv5 backbone. Comparative experiments prove YOLOv5-SE's superior performance over YOLOv5 and other models like Faster R-CNN and VGG16. A pytorch system enables model deployment for corrosion detection. The integration of DCGAN simulation and attention-optimized YOLOv5-SE enables precise classification and localization of corroded bolts. This study represented a significant step towards intelligent upgrades in steel bridge maintenance.