Deep Learning–Based Intelligent Detection of Shield Tunnel Lining Defects: A Comprehensive Review of Image Classification, Object Detection, and Semantic Segmentation

Shield tunnels are crucial for urban transit systems, yet they are susceptible to defects like cracks and leaks that compromise structural integrity and safety. Traditional inspection methods are inefficient and subjective. This paper provides a comprehensive review of deep learning–based intelligent detection methods for shield tunnel lining defects, with a focus on image classification, object detection, and semantic segmentation. This paper also systematically evaluates the end‐to‐end pipeline, including image acquisition technologies and preprocessing techniques optimized for complex tunnel environments. Feature extraction frameworks, such as convolutional neural networks (CNNs) and vision transformers (ViTs), are compared for their handling of noise and illumination changes. Multitask recognition methods, including single‐stage and two‐stage detectors, are benchmarked for real‐time localization and precision. Semantic segmentation networks are assessed for pixel‐level defect delineation. Our analysis reveals that transfer learning improves small‐dataset performance, attention mechanisms reduce background interference, and lightweight designs lower computational burdens. Despite challenges like small‐defect sensitivity and speed‐accuracy trade‐offs, solutions like Segment Anything Model (SAM)–guided 3D reconstruction and infrared–visible fusion are proposed. This work establishes a foundational framework for DL–driven tunnel inspection, providing guidelines for model selection, optimization, and deployment.