Automatic tunnel lining crack evaluation and measurement using deep learning

• Published in: Tunnelling and underground space technology Vol. 124; p. 104472
• Main Authors: Dang, L. Minh, Wang, Hanxiang, Li, Yanfen, Park, Yesul, Oh, Chanmi, Nguyen, Tan N., Moon, Hyeonjoon
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 01.06.2022; Elsevier BV
• Subjects: Coders; Crack propagation; Deep learning; Flaw detection; Image resolution; Image segmentation; Inspection; Lining crack; Measurement; Measuring; Passageways; Segmentation; Structural stability; Tunnel; Tunnel linings; Tunnels; U-Net; Underground construction; Deep learning; Lining crack; Tunnel; Segmentation; U-Net; Measuring
• ISSN: 0886-7798; 1878-4364
• DOI: 10.1016/j.tust.2022.104472

[Display omitted] •A massive tunnel crack segmentation dataset with over 170,000 images.•A lining crack segmentation model integrates the ResNet152 into the U-Net encoder.•Automated measurement of segmented cracks.•Robust against noise in challenging tunnel environment. A tunnel is an imperative underground passageway that supports fast and uninterrupted transportation. Over time, various factors, such as ageing, topographical changes, and excessive force, slowly affect the tunnel's internal structure, which causes tunnel defects that can reduce the structure's stability and eventually lead to enormous damage. Therefore, the tunnels need to be checked regularly to detect and fix the cracks promptly. Earlier inspection approaches mainly relied on the operators who directly observed videos to detect the cracks and determine their seriousness, which is laborious, error-prone, and tedious. This research suggests a deep learning-based tunnel lining crack segmentation framework for tunnel images taken by high-resolution cameras. The primary contributions are (1) a lining crack segmentation framework, which is motivated by U-Net architecture, where the encoder is replaced by a ResNet-152 model, (2) the automated measurement of the segmented cracks, which include length, thickness, and type, and (3) a huge lining crack segmentation database. The experimental results showed that the framework obtained comparable performance compared to existing crack segmentation models and supported the automated measurement of the segmented cracks.