Deep Learning-Based Segmentation and Quantification of Metro Shield Tunnel Lining Cracks

Abstract

To address the issues of insufficient segmentation accuracy and high computational complexity in crack detection of metro shield tunnel linings, a lightweight instance segmentation model named YOLO-LCD based on YOLO11 is proposed. The model introduces StarNet into the backbone to reduce computational cost, constructs a C3k2-S module in the neck to enhance crack feature representation, and replaces traditional concatenation with a Modulation Fusion Module (MFM) to improve multi-scale feature fusion efficiency. Meanwhile, a Lightweight Shared Convolution Detection head (LSCD) is adopted to reduce parameter redundancy. Furthermore, a complete crack extraction and quantification framework is developed by integrating image enhancement, Frangi filtering, and skeleton extraction, enabling automatic measurement of crack length, width, and orientation. Experimental results demonstrate that the proposed model significantly reduces parameters and computational cost while maintaining competitive segmentation accuracy, achieving a favorable balance between accuracy and efficiency and meeting the practical requirements of metro tunnel inspection.