Abstract: Tunnel seepage poses a significant threat to the durability and operational safety of shield tunnels, while manual inspection and visible-light imaging struggle to provide reliable detection. Therefore, a mobile infrared acquisition system consisting of twelve electronically focusing thermal cameras is developed for synchronized, high-resolution imaging of tunnel linings. To reduce labeling cost and address the inconsistency of existing labeling methods, an efficient HSV-based automatic annotation method is proposed, which leverages the stable color-temperature characteristics of infrared images to generate consistent seepage masks with higher efficiency. Compared with manual and SAM-assisted labeling methods, the HSV annotation method achieves higher consistency with lower manual correction cost for tunnel infrared images. A lightweight Unet model is further constructed by integrating MobileNet-based depthwise separable convolutions, full-scale padding-preserved convolutions, and optimized skip-connection pruning strategies, which substantially reduces the number of parameters by 87.3% while maintaining pixel-level accuracy. Experiments conducted on 199 real tunnel infrared images and an augmented dataset of 1 000 samples show that the proposed model achieves 99.76% accuracy, 98.23% IoU, and 22.4 FPS, outperforming TR-Unet in both accuracy and inference speed. A prototype inspection system combining the proposed model and the multi-camera platform achieves 17.8 FPS in real tunnel tests, demonstrating the practicality and effectiveness of the proposed method for intelligent tunnel seepage detection.

Key words: tunnel leakage, infrared imaging, intelligent recognition, semantic segmentation, Unet model