Browse by author
Lookup NU author(s): Zhen Chee, Professor Cheng ChinORCiD, Hao Chen, Dr Zi Jie ChoongORCiD, Dr Jun Jie ChongORCiD
Full text for this publication is not currently held within this repository. Alternative links are provided below where available.
Despite significant progress in corrosion monitoring, accurately targeting critical areas remains a persistent challenge due to irregular textures and environmental variability, which limit the effectiveness of traditional transfer learning approaches. To address this, this study explores the potential of optimised pool-based active learning to enhance corrosion detection. Pool-based active learning prioritises high-value samples, improving segmentation performance while reducing annotation costs by focusing on refining sample selection for corrosion-specific features rather than generic image uncertainty. Two distinct datasets were used to validate the segmentation model rigorously. The first is a laboratory-controlled dataset featuring standardised corrosion samples with precise ground-truth annotations, and the second is a site-realistic dataset captured under real-world environmental conditions. The laboratory experiments were conducted first to validate the methodology under controlled conditions, ensuring accurate segmentation against well-defined corrosion samples, before progressing to the site dataset. Experimental results demonstrate that the DeepLabv3 + model with an EfficientNet backbone, train with batch size of 16 and 50 epochs with an 80% train, 10% validation and 10% test dataset split using the Bayesian Active Learning by Disagreement (BALD) method, achieves 98% ± 0.16% pixel accuracy in controlled laboratory conditions and 87.8% ± 0.98% pixel accuracy on real-world on-site images. Furthermore, the on-site model demonstrated robust segmentation capabilities with a mean Intersection over Union (IoU) of 86.7% ± 0.28%, under challenging conditions. The findings underscore the strengths and trade-offs of active learning in corrosion detection. Future work would explore further optimisation methods to balance accuracy, efficiency, and scalability across diverse operating conditions.
Author(s): Chee ZQ, Chin CS, Chen H, Choong ZJ, Chong JJ, Canturri C, Portafaix T, Tan SJ
Publication type: Article
Publication status: Published
Journal: Advanced Engineering Informatics
Year: 2026
Volume: 71
Issue: Part C
Print publication date: 31/01/2026
Online publication date: 31/01/2026
Acceptance date: 17/01/2026
ISSN (electronic): 1474-0346
Publisher: Elsevier
URL: https://doi.org/10.1016/j.aei.2026.104366
DOI: 10.1016/j.aei.2026.104366
Altmetrics provided by Altmetric
