AsphaltCrackNet: A Novel Architecture for Classifying Cracks in Asphalt Pavement

Public roads and highways span hundreds of thousands of kilometers, and asphalt is the most widely used material for road construction. However, increasing traffic loads make it prone to wear and cracking, requiring efficient, scalable monitoring and maintenance strategies. To this extent, we propose AsphaltCrackNet, a novel lightweight deep learning architecture for real-time asphalt crack classification. Based on an enhanced U-Net architecture, it integrates multi-scale feature extraction, cross-attention modules, and attention gates to increase performance and interpretability. Evaluated using a five-fold cross-validation, AsphaltCrackNet achieves an accuracy of 99.18%, outperforming previous approaches. In addition, we visualize the classification decision, demonstrating that the models’ decisions are focused on image areas containing asphalt cracks. Lastly, we highlight the suitability of our lightweight architecture for deployment in embedded systems on vehicles or drones, with applications extending to both civilian and military infrastructure monitoring.