MGSLU-Net: a lightweight network for efficient detection of water leakage in subway tunnel linings

MGSLU-Net: a lightweight network for efficient detection of water leakage in subway tunnel linings

연구 분야: Networking

학회: The Visual Computer

The occurrence of water leakage in the linings of subway tunnels represents a considerable threat to the safety and operational integrity of the tunnel infrastructure. Conventional detection techniques depend significantly on manual examination, which is labor-intensive, subjective, and susceptible to inaccuracies. To address these challenges, this paper presents a lightweight segmentation model, MGSLU-net, which is based on improvements to the U-net architecture. By employing MobileNetV3-Large as the foundational feature extraction network, MGSLU-net significantly decreases the quantity of model parameters and computational complexity while preserving good segmentation accuracy. Moreover, the incorporation of the GSConv module and the mixed local channel attention mechanism enhances the model’s capacity to discern pertinent features of water leakage and mitigate interference from extraneous background information. Meanwhile, a class-balanced loss function is put forth as a means of addressing the issue of significant class imbalance present within the water leakage dataset. The experimental results indicate that MGSLU-net attains state-of-the-art performance in terms of mean intersection over union (83.98%), mean pixel accuracy (91.5%), and accuracy rate (98.06%), with a minimal number of parameters (2.844M) and low computational demand (4.286 GFLOPs). The model operates at a high frame rate of 81.84 f/s, rendering it ideal for real-time detection of water leakage in subway tunnels. The proposed MGSLU-net model presents a viable approach for the efficient and precise automatic detection of water leakage in subway tunnels.