Deep learning-based YOLO for crack segmentation and measurement in metro tunnels

Abstract

To address the increasing issue of cracks in metro shield tunnels, this paper proposes the YOLOv8-GSD model, which integrates DySnakeConv, BiLevelRoutingAttention, and the Gather-and-Distribute Mechanism with the YOLOv8 algorithm. This model is designed for detecting and segmenting cracks in tunnel linings and employs a pixel grouping method to measure crack length and width. Using a real crack dataset from a subway section in Suzhou, China, comparative experiments with YOLOv8x, BlendMask, SOLOv2, and YOLACT demonstrate that YOLOv8-GSD excels in segmentation performance (AP of 82.4 %) and accuracy (IoU of 0.812). The measured crack dimensions show an error within 5 % compared to actual values, confirming the model's effectiveness. These results highlight the potential of YOLOv8-GSD for enhancing the maintenance and safety of metro tunnels.