<p>The automated detection of road cracks holds immense practical significance for ensuring traffic safety. Deep learning-based semantic segmentation models have achieved state-of-the-art results in the domain of road crack detection. However, the requirement for high detection accuracy often leads these models to exhibit elevated complexity, resulting in large model sizes that render them unsuitable for deployment on mobile devices. To address this challenge, we propose a lightweight road surface crack segmentation model dubbed GTRS-Net, which is built upon improvements to the U-Net architecture. Firstly, we design a novel Ghost Tiered Fused (GTF) Module that focuses on extracting more comprehensive crack feature information, enabling more effective feature fusion. Secondly, the GTF module is integrated into a new lightweight encoder–decoder architecture, the Ghost Tiered Reconstructive (GTR) Module, which can enhance the extraction of contextual feature information while significantly reducing computational complexity. Finally, in the decoding stage of the model, we introduce the SE (Squeeze and Excitation) attention mechanism that can effectively suppress non-crack features, allowing our model to extract more detailed edge and contextual information during the decoding process. The proposed GTRS-Net model has been comprehensively evaluated on three publicly available road crack datasets: DeepCrack, Crack760, and Crack850. The experimental results demonstrate that, in comparison to other semantic segmentation algorithms, the GTRS-Net model achieves both a lower parameter count and higher segmentation accuracy, rendering it more valuable for practical applications.</p>