open access

https://hdl.handle.net/2115/100450

Active Semi-Supervised Learning for Scalable Corrosion Detection in Steel Structures

英語

Structural health monitoring

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

Detection of corrosion in steel structures is essential for reliable monitoring of structural health, but the creation of large, high quality labelled data sets remains costly and labor-intensive. This study introduces an active semi-supervised learning framework to efficiently scale corrosion detection by combining uncertainty-based active sample selection and semi-supervised learning strategies using large-scale unlabeled field images. The proposed pipeline will iteratively query only the most informative samples for expert annotations, refining the pseudo-labels by regularizing consistency and adaptive confidence filtering. The experimental results for steel corrosion image datasets show that the ASSL approach achieves performance comparable to fully supervised models with significantly fewer labelled samples, demonstrating improved efficiency, generalization capability and robustness to the variability of the real world. This work highlights ASSL as a practical route to cost-effective and scalable corrosion detection in the monitoring of civil infrastructures.

会議発表論文

Version of Record