Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
Eine Plattform für die Wissenschaft: Bauingenieurwesen, Architektur und Urbanistik
An Optimized Dual-View Snake Unet Model for Tunnel Lining Crack Detection
The prompt and accurate detection of tunnel lining cracks is essential for maintaining the safety and reliability of tunnels. Deep learning-based approaches have significantly advanced automated crack detection, delivering improved efficiency and precision in tunnel inspection. Nevertheless, the intricate characteristics of cracks, manifesting as fine, elongated, and irregular structures, pose substantial challenges for deep learning-based semantic segmentation networks, hindering their ability to achieve comprehensive and accurate identification. Aiming to tackle these challenges, this paper proposes a novel dual-view snake Unet (DSUnet) model, which integrates a hybrid snake cascading (HSC) module and a Haar wavelet downsampling (HWD) operation. The HSC module enhances the network’s capability of extracting tunnel lining cracks by synergistically combining features derived from standard convolutions and bidirectional dynamic snake convolutions, thereby capturing intricate geometric and contextual information. Meanwhile, the HWD operation facilitates the preservation of critical spatial information by performing multi-scale feature refinement, which effectively reduces segmentation uncertainty. Experimental results demonstrate the proposed DSUnet achieves a mean Dice coefficient (MDice) of 71.8% and a mean intersection over union (MIoU) of 77.4%. Compared to the baseline Unet model, DSUnet delivers improvements of 1.3% in MDice and 0.6% in MIoU, respectively. Additionally, the proposed model consistently outperforms several state-of-the-art semantic segmentation networks, highlighting its robustness and accuracy in detecting tunnel lining cracks. These findings position DSUnet as a promising tool for automated tunnel inspection, contributing to improved safety and operational reliability.
An Optimized Dual-View Snake Unet Model for Tunnel Lining Crack Detection
The prompt and accurate detection of tunnel lining cracks is essential for maintaining the safety and reliability of tunnels. Deep learning-based approaches have significantly advanced automated crack detection, delivering improved efficiency and precision in tunnel inspection. Nevertheless, the intricate characteristics of cracks, manifesting as fine, elongated, and irregular structures, pose substantial challenges for deep learning-based semantic segmentation networks, hindering their ability to achieve comprehensive and accurate identification. Aiming to tackle these challenges, this paper proposes a novel dual-view snake Unet (DSUnet) model, which integrates a hybrid snake cascading (HSC) module and a Haar wavelet downsampling (HWD) operation. The HSC module enhances the network’s capability of extracting tunnel lining cracks by synergistically combining features derived from standard convolutions and bidirectional dynamic snake convolutions, thereby capturing intricate geometric and contextual information. Meanwhile, the HWD operation facilitates the preservation of critical spatial information by performing multi-scale feature refinement, which effectively reduces segmentation uncertainty. Experimental results demonstrate the proposed DSUnet achieves a mean Dice coefficient (MDice) of 71.8% and a mean intersection over union (MIoU) of 77.4%. Compared to the baseline Unet model, DSUnet delivers improvements of 1.3% in MDice and 0.6% in MIoU, respectively. Additionally, the proposed model consistently outperforms several state-of-the-art semantic segmentation networks, highlighting its robustness and accuracy in detecting tunnel lining cracks. These findings position DSUnet as a promising tool for automated tunnel inspection, contributing to improved safety and operational reliability.
An Optimized Dual-View Snake Unet Model for Tunnel Lining Crack Detection
Baoxian Li (Autor:in) / Hao Xu (Autor:in) / Xin Jin (Autor:in) / Huaizhi Zhang (Autor:in) / Shuo Jin (Autor:in) / Qianyu Chen (Autor:in) / Fengyuan Wu (Autor:in)
2025
Aufsatz (Zeitschrift)
Elektronische Ressource
Unbekannt
Metadata by DOAJ is licensed under CC BY-SA 1.0
| Engineering Index Backfile | 1912
Tunnel lining crack detection model based on improved YOLOv5
| Elsevier | 2024
Tunnel lining crack detection model based on improved YOLOv5
| Elsevier | 2024
UNet-based model for crack detection integrating visual explanations
| Elsevier | 2021
Tunnel Lining Improvement by Optimized Anchoring
| British Library Conference Proceedings | 1996