Water leakage in metro tunnels is a critical safety indicator, necessitating regular inspections to avert catastrophic failures.Deep learning-based computer vision is currently utilized to detect water leakage in metro tunnels.However, challenges like large model parameters, low detection accuracy, and poor robustness in complex environments persist.
Therefore, this study proposes a novel lightweight semantic segmentation model.An adaptive information extraction network, integrating spatial and channel squeeze-and-excitation offworld drum pads mechanisms, is adopted in the encoder to enhance critical feature representation and accelerate inference.Additionally, a multiscale and lightweight feature enhancement module is introduced to capture global contextual information while reducing the number of parameters.
The decoder employs a sequential up-sampling strategy, complemented by skip connections, to integrate multiscale features effectively and enhance gradient flow.The experimental results show that the proposed model has better segmentation performance than other compared models, which is very the gel bottle audrey suitable for detecting metro tunnel leakage in complex environments.