Automated detection system of metro tunnel lining crack using dynamic snake convolution

Inspecting defects in tunnel linings is a crucial part of tunnel maintenance work. Traditional tunnel inspection methods are generally inefficient, making it difficult to complete intensive inspection tasks and provide detailed characteristic data of cracks within the limited maintenance time. In this research, a deep learning–driven automatic inspection method was developed to evaluate the structural health of metro tunnel linings and deliver quantitative data on lining cracks. The proposed framework encompasses (1) addressing the balance between detection speed and imaging accuracy, developing a front-end inspection vehicle—the metro tunnel defect detection system—which efficiently and rapidly captures high-resolution images of the lining surface, alongside establishing a CNN-based classifier for fast classification of crack images. (2) Considering the slender morphological characteristics of cracks, the DSC_CrackU model was developed for crack segmentation. This model introduces dynamic snake convolution (DSConv) and achieves fine segmentation of cracks by adaptively adjusting the shape of convolution kernels. Meanwhile, it integrates multidimensional feature information by means of the feature fusion module and utilizes the self-efficient channel module to enhance sensitivity to crack regions. Results show that the algorithm uses fewer computational parameters while maintaining excellent performance in other metrics. (3) We propose a quantitative characterization algorithm based on DSC_CrackU recognition outcomes, which maps pixel-dimensional features of cracks to the physical dimension, thereby forging a connection between the theoretical framework and engineering standards. Field application tests across multiple tunnels validated the technical feasibility of the proposed framework for engineering applications.