An intelligent tunnel lining crack inspection framework via deep learning segmentation and geometric feature quantification

This is the authors' abstract. We don't add key points for this paper.

Automated crack inspection is essential for structural health monitoring of tunnel infrastructure, yet most existing methods focus mainly on pixel-level segmentation with limited support for geometric interpretation and engineering assessment. This study proposes an integrated tunnel lining crack inspection framework combining deep learning-based segmentation, geometry-aware feature characterization, and interactive visualization. A multi-scale crack perception module based on FCN and FPN, together with a lightweight PANet-inspired feature-propagation strategy, is adopted as the segmentation component of the integrated framework. Six feature extraction backbones are evaluated under the same framework, and ResNet101 is selected as the benchmark configuration because of its favorable balance between segmentation accuracy and computational efficiency. On the tunnel lining crack dataset, the ResNet101-based model achieves an IoU of 51.32 % and an F1-score of 65.92 %. After independent retraining on the Crack500 pavement crack dataset, the same architecture achieves an mIoU of 76.08 %, providing evidence of cross-scenario adaptability after dataset-specific retraining without scenario-specific architectural redesign. Based on the binary crack maps, geometry-aware refinement enables crack-region decomposition, geometry-derived width characterization, direction identification, and engineering-based condition grading. The developed intelligent crack segmentation system (ICSS) further integrates segmentation, feature quantification, condition assessment, and visualization into a unified workflow, providing a practical basis for tunnel lining crack inspection and subsequent application-oriented specialization.