Efficient domain adaptation network for corrosion segmentation in infrastructure: from laboratory to real-world images
Authors:
Open Access
- 5 September 2025
- journal article
- research article
- Published by Springer Nature in Journal of Civil Structural Health Monitoring
- Vol. 15 (8) , 3535-3552
Abstract
Corrosion in infrastructures, such as locks, dams, and facilities, poses significant risks to structural integrity and drives up maintenance costs. Accurate detection and segmentation of corrosion from inspection imagery are critical for proactive structural health monitoring (SHM) and maintenance decision making. Traditional methods for corrosion assessment rely heavily on manual inspection, which are time-consuming, inconsistent, and often ineffective under diverse environmental conditions. Recent advancements in deep learning have improved automated image-based corrosion detection. However, many existing approaches struggle with generalization across domains due to variability in corrosion appearance, surface materials, and imaging conditions (e.g., laboratory versus field). To address this challenge, this paper proposes a deep learning-based corrosion detection framework that integrates domain adaptation to enable robust segmentation performance across heterogeneous datasets. The proposed method consists of four key components: (1) a geometric structure preservation (GSP) module to retain local topological relationships across domains, (2) a singular value decomposition (SVD)-based local discrepancy (SLD) module that leverages SVD to align local feature subspaces at a fine-grained level, (3) a global consistency alignment (GCA) module using maximum mean idscrepancy (MMD) to reconcile distributional shifts between source and target domains, and (4) an auxiliary domain adversarial neural network (DANN) component to further encourage domain-invariant feature learning. These modules are supported by a proposed lightweight segmentation backbone, the efficient segmentation network (EffSegNet), which enables efficient corrosion pattern recognition. The proposed framework was implemented and tested on corrosion imagery from real-world infrastructure assets. The method achieved high segmentation accuracy and demonstrated strong generalization capabilities across domain-shifted datasets, indicating its potential to support scalable, efficient, and automated corrosion assessment for SHM applications.Keywords
Funding Information
- U.S. Army Corps of Engineers (PE 0603119A)
This publication has 55 references indexed in Scilit:
- A pencil lead break-triggered, adversarial autoencoder-based approach for rapid and robust rail damage detectionEngineering Applications of Artificial Intelligence, 2025
- Creating a first-pass algorithm for corrosion assessment in bridge inspections using machine learning and UAV-collected imagery dataStructure and Infrastructure Engineering, 2025
- Revolutionizing structural health monitoring in marine environment with internet of things: a comprehensive reviewInnovative Infrastructure Solutions, 2025
- Automated Image Segmentation and Processing Pipeline Applied to X‐Ray Computed Tomography Studies of Pitting Corrosion in Aluminum WiresAdvanced Engineering Materials, 2025
- Segmentation of Bridge Components from Various Real Scene Inspection Images2024
- HOMDA: High-Order Moment-Based Domain Alignment for unsupervised domain adaptationKnowledge-Based Systems, 2022
- LMISA: A lightweight multi-modality image segmentation network via domain adaptation using gradient magnitude and shape constraintMedical Image Analysis, 2022
- Geometric Unsupervised Domain Adaptation for Semantic Segmentation2021
- Learning adaptive geometry for unsupervised domain adaptationPattern Recognition, 2021
- Definition and Importance of Corrosion2008