Researchers have introduced an enhanced deep learning architecture named CSwinUKAN to advance automated tunnel lining crack segmentation and significantly reduce missed detections during safety-critical infrastructure inspections. While automatic detection of tunnel lining fractures has been widely studied to protect infrastructure safety, reliable recognition in real operating tunnels remains difficult due to environmental complexities. The newly developed model integrates a pretrained Swin Transformer and a Tokenized Kolmogorov–Arnold Network within an enhanced U-Net framework to resolve these structural monitoring challenges. In experimental evaluations, the proposed architecture outperformed multiple established baseline models, delivering stronger overall segmentation performance and notably higher recall.
Visual interference from glare, shadows, and seepage traces can obscure fine or discontinuous cracks. In addition, segment joints and textured backgrounds can obscure fine or discontinuous cracks. This visual interference makes missed detection a major practical risk.
To overcome these visual impediments, the proposed CSwinUKAN model incorporates specialized algorithmic modifications into an enhanced U-Net segmentation framework. The model combines a pretrained Swin Transformer to capture long-range crack context. The model also incorporates a Tokenized Kolmogorov–Arnold Network module to strengthen nonlinear feature representation. Furthermore, Swin-based decoding is utilized to improve crack continuity recovery, while Focal Loss is adopted during training to address the extreme pixel imbalance between narrow crack targets and massive background areas, thereby reducing false negatives.
The experimental research team validated the segmentation capabilities of CSwinUKAN on an integrated dataset containing 10,422 inspection images. This extensive image collection combined publicly accessible crack datasets with real-world metro shield-tunnel imagery collected from operating transit systems in Ningbo and Shanghai, China. In head-to-head performance evaluations, CSwinUKAN outperformed multiple established deep learning baselines, including standard U-Net, Swin-U-Net, U-KAN, and YOLO-based segmentation models. The experimental results demonstrated that the model achieved stronger segmentation performance and recall.
By delivering higher recall and mitigating missed detections, the CSwinUKAN architecture provides a promising approach for high-accuracy tunnel crack inspection across civil infrastructure networks. The work also lays a foundation for future lightweight and real-time deployment. The complete peer-reviewed study, titled “CSwinUKAN: An Enhanced U-Net Architecture with Swin-Transformer and Kolmogorov–Arnold Network Modification for Tunnel Lining Cracks Segmentation,” was officially published in the Journal of Railway Science and Technology in June 2026.
Subject of Research: Deep learning architecture for tunnel lining crack segmentation
Article Title: Advancing tunnel crack detection with AI for safer infrastructure inspection and monitoring
Article References: Advancing tunnel crack detection with AI for safer infrastructure inspection and monitoring. Original publication (Bibliographic details unavailable.)
Image Credits: AI Generated
DOI: Not provided
Keywords: CSwinUKAN, tunnel lining cracks, crack segmentation, Swin Transformer, Kolmogorov–Arnold Network, Tokenized KAN, U-Net architecture, Focal Loss, infrastructure inspection, metro shield tunnels, Ningbo metro, Shanghai metro, visual interference, crack continuity recovery, missed detection reduction, Journal of Railway Science and Technology, long-range crack context
News Source: Denise Maddox. (November 12, 2026). Enhanced U-Net Architecture Improves Crack Detection in Operational Railway Tunnels. Scienmag.



