DSMUNet: A Lightweight Model for Road Crack Segmentation

As the most common safety hazard among road surface distresses, pavement cracks require fine-grained segmentation to support subsequent maintenance decision-making. However, existing methods find it difficult to cope with complex environmental interference and multi-morphology cracks while maintaining low computational cost. To address the above issues, this study proposes a lightweight crack segmentation model, DSMUNet. Based on the U-shaped encoder–decoder framework, depthwise separable convolution is adopted to replace traditional convolution to reduce computational cost, and the SGE spatial group enhancement mechanism is introduced at the encoder side to emphasize crack features and suppress interference from non-crack textures. In addition, a new multi-scale feature fusion module is proposed to improve the overall connectivity of cracks during the feature fusion stage, thereby further improving model performance on the basis of model lightweighting. The experimental results show that DSMUNet contains only 0.55 M parameters and requires 21.708 GFLOPs, with an average inference latency of 5.42 ms, while achieving Dice/mIoU scores of 78.10%/69.05% on the private dataset and 87.12%/79.57% on the Crack500 dataset, respectively. These results demonstrate the advantages of this study in terms of resource consumption as well as its excellent segmentation performance, providing a more efficient implementation scheme for pavement crack segmentation in resource-constrained environments.