|
|
|
| Research on tunnel crack detection based on improved DeepLabV3+ and percolation algorithm |
| CHEN Zhangxin1,2，WANG Gang1*，LI Wenfeng3，LI Ke3，JIANG Song1，LIU Tingfang1
|
| (1. Fujian University of Technology, Fujian 350118, Fuzhou, China; 2. The fifth construction Co. Ltd. of CCCC Fouth Harbor Engineering Co. Ltd., Fujian 350008, Fuzhou, China; 3. China Merchants Chongqing Communications Technology Research & Design Institute Co. Ltd., Chongqing 400074, China) |
|
|
|
|
Abstract To address the problem of poor edge localization in deep learning-based methods and low efficiency of traditional percolation algorithms for tunnel crack detection, a two-stage approach that integrates an improved DeepLabV3+ with a skeleton-guided percolation algorithm was proposed.In the first stage, an enhanced DeepLabV3+ model—incorporating a CBAM attention module, an optimized ASPP module, and a Dice loss function—was developed to achieve high-recall crack pre-segmentation. In the second stage, a skeleton-guided percolation growth strategy combined with morphological constraints was applied to refine crack edges and measure crack widths.A tunnel crack dataset containing 20 604 pixel-level annotated images was constructed, covering various tunnel lining regions (crown, haunch, sidewall) and surface conditions (dry, wet, stained).Experimental results on this dataset showed that the pre-segmentation module achieved an accuracy of 90.1% and a recall of 86.7%.The improved percolation algorithm increased the precision to 98.5% while maintaining high recall, and improved computational efficiency by approximately 20 times.Engineering validation demonstrated a detection rate exceeding 84% for cracks wider than 0.1?mm, with a mean absolute error of less than 0.3?mm.The proposed method effectively balanced detection accuracy and computational efficiency, providing a feasible solution for automated tunnel lining crack detection.
|
|
Received: 15 January 2026
Published: 21 May 2026
|
|
|
|
|
|
|
|
