Full-process computer vision framework for tunnel lining crack detection: From system development to engineering application

Automated tunnel lining crack detection remains challenging for full-section inspection and spatial localization. This paper introduces a software system integrating computer vision, deep learning, and dynamic visualization. Scale-Invariant Feature Transform (SIFT)-based stitching enables continuous full-section image reconstruction, while a generic 2D-to-3D mapping strategy achieves macro-level spatial localization and refinement of cracks. The framework integrates image stitching, crack recognition, quantitative characterization, multidimensional visualization, interactive refinement, and automated reporting within a closed-loop workflow, supporting the complete process from image acquisition to engineering output. Field tests demonstrated image registration accuracy of approximately 3 pixels and detection rates ranging from 82.5% to 85.7% under complex conditions. Human-in-the-loop refinement effectively reduced false positives and restored crack continuity, improving precision and recall by more than 0.7 and 0.025, respectively. The 2D/3D visualization intuitively reveals the spatial distribution of tunnel cracks, underscoring the system's engineering value in advancing automated digital lifecycle maintenance of tunnels.