CSwinUKAN: An enhanced U-Net architecture with swin-transformer and Kolmogorov–Arnold network modification for tunnel lining cracks segmentation

Concrete is the most widely utilized material for tunnel linings. However, concrete structures are prone to cracking due to various environmental factors or actions. The formation and propagation of cracks deteriorate lining stiffness and durability, endangering tunnel safety. Therefore, concrete crack detection is always an important topic for structural health monitoring in tunnels. Traditional manual inspection methods are labor-intensive and prone to human error, while Image Processing Technique-based methods are unable to handle complex crack morphologies and rely on inefficient manual feature engineering for parameter optimization. Consequently, Deep Learning models capable of autonomously updating their parameters via backpropagation have been widely adopted for the crack segmentation task. Despite the recent breakthrough on advanced deep learning architectures, existing segmentation models struggle in mitigating false negatives. Hence, a pioneering concrete crack segmentation model, CSwinUKAN, is proposed in this paper, which combines the Kolmogorov Arnold Network and Swin Transformer in the U-shape Encoder–Decoder backbone, enabling the precise disentanglement of subtle, continuous semantic features from the concrete crack through the sliding window self-attention mechanism and a learnable activation function. Meanwhile, an original proprietary dataset comprising 444 annotated samples of cracked and intact segmental linings was developed using in-situ captures from the Ningbo and Shanghai Metro tunnels. The proposed CSwinUKAN was subsequently trained and validated on a comprehensive repository of 10,422 samples, which integrates developed dataset with 8 open-source concrete crack datasets. The model performance in terms of mIoU and F1 score was compared to four baselines, namely U-Net, Swin-U-Net, U-KAN, and YOLO26, indicating that the proposed model achieves a significant improvement, further confirming its robust potential for practical deployment in real-world applications.