A Deep Learning-Driven Framework for Real-Time Recognition and Quantification of Subway Tunnel Surface Defects Using High-Resolution Imaging

Abstract Subway tunnels, with their complex underground environment, frequently develop surface defects like cracks and water leakage over time. Traditional manual inspection methods are inadequate for current accuracy and efficiency needs. This paper introduces a deep learning approach to detect and assess these defects in subway tunnels, aiming for real-time detection and quantitative assessment. It proposes a data collection strategy for rapid detection and detailed exploration in subway shield tunnels. A multi-category dataset of tunnel defects was created, alongside an automatic identification method using the YOLO v7 algorithm for operational tunnels. This method, validated against Qingdao Metro’s manual inspection records, markedly reduces manual inspection costs during tunnel operation. Research indicates a strong correlation between image inspection equipment efficiency, defect detection accuracy, and actual project needs in subway tunnels. Image quality is positively linked to tunnel illumination intensity. A balance between inspection speed and focal length is crucial for image size and precision. Lowering the confidence threshold from 0.4 to 0.2 increases detection rates for cracks and water leakage by 20.60 and 4.06%, respectively, minimizing defect oversight. This study presents algorithms, frameworks, and methods for real-time quantification to enhance tunnel operation, maintenance, and manual processing.