Abstract
Accurate quantification and analysis of steel corrosion is crucial for reliability assessment studies of in-service reinforced concrete structures. However, the pixel-level cross-sectional data provided by X-ray computed tomography (XCT) proves difficult to quantify, especially for the amorphous corrosion products filled in mortar, due to the absence of robust feature extraction methods. In this study, multiple deep learning models were trained to automatically identify corrosion products from a large number of XCT images. The database comprised XCT images obtained from a RC component subjected to chloride-rich environment for four years. The results indicate that deep learning models can segment different regions of XCT images with high accuracy. Among the models, the K-Net model performed the best on this dataset, achieving an accuracy of 94.60%, and a mean Precision (mPrecision) of 88.21%. This advance makes it possible to automatically extract parameters that characterise steel corrosion and to assess the damage to RC structures caused by corrosion.
Keywords
steel corrosion, XCT, deep learning algorithm, semantic segmentation.
Date of Version
2025
DOI
10.5703/1288284318131
Recommended Citation
Zhao, Yuxi; Wang, Yuzhou; Miao, Xu; and Wan, Hanjun, "Feature Extraction of Steel Corrosion based on XCT Scanning and Deep Learning Model" (2025). International Conference on Durability of Concrete Structures. 2.
https://docs.lib.purdue.edu/icdcs/2025/keynote/2
Feature Extraction of Steel Corrosion based on XCT Scanning and Deep Learning Model
Accurate quantification and analysis of steel corrosion is crucial for reliability assessment studies of in-service reinforced concrete structures. However, the pixel-level cross-sectional data provided by X-ray computed tomography (XCT) proves difficult to quantify, especially for the amorphous corrosion products filled in mortar, due to the absence of robust feature extraction methods. In this study, multiple deep learning models were trained to automatically identify corrosion products from a large number of XCT images. The database comprised XCT images obtained from a RC component subjected to chloride-rich environment for four years. The results indicate that deep learning models can segment different regions of XCT images with high accuracy. Among the models, the K-Net model performed the best on this dataset, achieving an accuracy of 94.60%, and a mean Precision (mPrecision) of 88.21%. This advance makes it possible to automatically extract parameters that characterise steel corrosion and to assess the damage to RC structures caused by corrosion.
