Experimental study demonstrates accurate instance segmentation of water leakages in subway tunnels, indicating enhanced automated monitoring for railway safety.
Key Points
To develop an automated deep learning framework capable of accurately segmenting water leakages in shield subway tunnels using mobile laser scanning intensity images.
Integrated a Res2Net residual network module into a cascade architecture to extract multiscale geometric features of water leakages.
Expanded the receptive field of network layers using graded residual connections within individual blocks.
Tested the framework on five water leakage datasets converted from point clouds captured by a custom mobile laser scanning system.
The expanded receptive fields enabled effective extraction of complex geometric characteristics of moisture under tunnel disturbance conditions.
Comparative experimental evaluations across five tunnel datasets demonstrated superior segmentation performance over conventional baseline models.