Multiscale Segmentation and Quantitative Grading Detection of Subway Tunnel Crack Images

While crack detection technologies for subway tunnels have diversified, systematic width‐based classification remains underexplored. We developed a tunnel inspection system with high‐resolution area array cameras, enabling high‐definition crack imaging. Systematically classified by millimeter‐scale criteria and annotated with pixel‐level precision, the dataset established a multiscale crack database encompassing diverse dimensional features. Building on this, an encoder–decoder neural network optimized for multiscale crack segmentation was proposed, achieving enhanced recognition accuracy across crack dimensions. Experimental results demonstrated the method’s significant advantages over conventional models. The method achieved 91.8% pixel accuracy (PA) for cracks with widths of 10 pixels or more, 80.05% PA for cracks spanning 6–9 pixels, and maintained 61.94% PA for cracks as narrow as 1‐2 pixels. These metrics underscore the framework’s robustness in supporting precision maintenance protocols for underground infrastructure.