.png){kind=icon}
.png){kind=icon}
Return
Tunnel lining crack segmentation using attention mechanisms and conditional GAN data augmentation
.png){kind=icon}
PRE
AI
A
C
X
Z
DOI：10.1016/j.tust.2026.108024.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate segmentation of tunnel lining cracks is essential for assessing the long-term safety and serviceability of underground infrastructure. Deep learning has emerged as an effective approach for automated crack segmentation. However, two challenges are frequently encountered in practice: the difficulty of detecting thin and low-contrast crack structures on complex tunnel surfaces, and the limited availability of annotated crack images required for training data-driven models. To address these challenges, this study proposes a crack segmentation framework that combines network architecture improvement with dataset augmentation. Specifically, the Convolutional Block Attention Module (CBAM) is integrated into a baseline U-Net model to help the network focus more effectively on crack-related features by refining both channel and spatial information. In addition, a Pix2Pix-based conditional generative adversarial network (cGAN) is employed to translate diversified crack masks into realistic tunnel lining crack images, thereby enriching the training dataset while preserving image–mask correspondence and reducing the need for additional manual annotation. Results demonstrate that both architectural refinement and dataset enrichment contribute to improved segmentation performance. Compared with the baseline U-Net trained on the original dataset, the CBAM-enhanced U-Net trained on the cGAN-enriched dataset achieves an average improvement of 3.13 percentage points in Intersection over Union (IoU) and 3.10 percentage points in F1-score on the real image test set. The proposed framework provides an effective solution for improving crack segmentation performance under limited data conditions and contributes to the development of more reliable automated tunnel inspection systems for infrastructure maintenance.
Keywords:
Tunnel lining cracks
Crack segmentation
U-Net
Attention mechanism
Conditional generative adversarial network
Journal
IF:
7.4
Papers:
6.8K
Citations:
3.5W
Organization
Cited Papers