Semi-supervised method for automated detection and quantitative assessment of corrosion states in structural members

Abstract

Accurate detection and comprehensive assessment of corrosion states are essential for bridge safety and durability. Deep learning-based semantic segmentation methods show significant potential for corrosion detection. However, supervised methods confront substantial challenges in labor-intensive annotation and limited datasets. To address these challenges, a semi-supervised method for corrosion state segmentation (Model A) and structural member segmentation (Model B) is proposed. It adopts the weak-to-strong semi-supervised framework with SE attention and a random cut strategy, outperforming supervised methods with only 40 % labeled corrosion and 20 % labeled member images. New evaluation metrics are established to evaluate the integrated results of Model A and Model B. A smartphone-based mobile detection platform is developed to achieve automatic corrosion detection and quantitative assessments. The proposed method achieves high accuracy with limited manual annotations, offering an advanced and intelligent solution for detecting, quantifying, and managing corrosion states on bridge structural members.