Automatic detection of tunnel lining cracks has been widely studied for infrastructure safety, but reliable recognition in real operating tunnels remains challenging. Visual interference from glare, shadows, seepage traces, segment joints, and textured backgrounds can obscure fine or discontinuous cracks, making missed detection a major practical risk.

In this work, we propose CSwinUKAN, an enhanced U-Net architecture for tunnel lining crack segmentation. The model combines a pretrained Swin Transformer to capture long-range crack context, a Tokenized Kolmogorov–Arnold Network module to strengthen nonlinear feature representation, and Swin-based decoding to improve crack continuity recovery. Focal Loss is further adopted to address the imbalance between crack and background pixels and reduce false negatives.

Validated on an integrated dataset of 10,422 images from public crack datasets and metro shield-tunnel images from Ningbo and Shanghai in China, CSwinUKAN outperformed U-Net, Swin-U-Net, U-KAN, and YOLO-based baselines, achieving stronger segmentation performance and recall. This work provides a promising approach for high-accuracy tunnel crack inspection and lays a foundation for future lightweight and real-time deployment. The work entitled “CSwinUKAN: An Enhanced U-Net Architecture with Swin-Transformer and Kolmogorov–Arnold Network Modification for Tunnel Lining Cracks Segmentation” was published in the Journal of Railway Science and Technology (published in June, 2026).
doi:10.1016/j.jrst.2026.04.001