Lightweight instance segmentation for rapid leakage detection in shield tunnel linings under extreme low-light conditions

Abstract

Rapid and accurate water leakage detection and segmentation is essential for ensuring the structural safety of subway tunnels. This paper simulates the extreme low-light conditions inside the tunnel from multiple perspectives. By employing inverse operations, pseudo-RAW format data are generated, providing more original features and avoiding the complex computations associated with traditional image enhancement and denoising algorithms. A lightweight instance segmentation network is optimised and designed, incorporating a multi-stage star-shaped backbone to improve feature extraction in dark environments, and serial-parallel structured detection-segmentation heads are used to accelerate segmentation speed. Experiment results demonstrate that the optimised model, using pseudo-RAW data, achieves a segmentation precision of 84.4 % in leakage instance segmentation under low-light conditions, with a model size of only 2.7 M. The proposed method closely aligns with real-world engineering environments, providing a low-cost and efficient solution for leakage monitoring in subway shield tunnels.