Surface Feature and Defect Detection Method for Shield Tunnel Based on Deep Learning

Abstract

Surface defects in the segmental lining of shield tunnels, such as water leakage and damage, pose significant threats to safety. Currently, manual inspection methods are inefficient and inaccurate. Most artificial intelligence techniques for detecting tunnel features and surface defects face challenges, including poor data quality and high computational costs in real-world settings. This paper introduces an automated system for tunnel information acquisition and defect detection, offering a comprehensive solution for identifying surface features and defects. An intelligent tunnel inspection vehicle was designed for automatic image acquisition, and a preprocessing method combining adaptive local tone mapping (ALTM) with contrast-limited adaptive histogram equalization (CLAHE) was used to improve image illumination and contrast. An enhanced deep-learning method based on segmenting objects by locations version 2 (SOLOv2) was proposed, which incorporates an F-ResNeSt backbone with a focus structure from split-attention networks with 101 layers (ResNeSt101), and an improved bi-directional feature pyramid network (BIFPN) with a convolutional block attention module (CBAM) in the feature fusion module. Applied to the Xiangya Road Tunnel, the method proves to be efficient, lightweight, and accurate, offering novel approaches for detecting tunnel surface features and defects.