DSCP-UNet: A Tunnel Crack Segmentation Algorithm Based on Lightweight Diminutive Size and Colossal Perception

To effectively enhance the ability to detect surface crack defects in tunnels under tunnel conditions, this paper proposes a new tunnel crack segmentation algorithm using light-weight diminutive size and colossal perception UNet (DSCP-UNet). This method integrates a diminutive group convolution (DGConv) module and a self-adaptive pooling (SAPool) module into the UNet architecture, resulting in a segmentation network that balances lightweight design with strong perceptual capability. In the constructed DSCP-UNet model, the conventional convolutional backbone is replaced by DGConv, significantly reducing the number of model parameters. Aconvolutional block attention module (CBAM) is introduced to enhance the model’s focus on crack features, forming an efficient crack recognition network. The pyramid attention with sampling group (PASG) module is also investigated to improve the model’s ability to discriminate multi-scale crack features. Using collected tunnel crack images, the proposed DSCP-UNet achieves a model size of only 3.85 MB and demonstrates superior performance in tunnel crack segmentation tasks. Based on the collected images of surface cracks in the tunnel, we verified the rationality of the proposed algorithm. Meanwhile, by comparing with other models, it was found that the algorithm proposed in this paper has high robustness.