RDD-DETR: lightweight road damage detection with hierarchical feature interaction

Abstract Timely and precisely detection of road damage is crucial for ensuring traffic safety and reducing maintenance costs. However, conventional detection methods frequently encounter challenges in accurately identifying small and multi-scale damages within complex road environments. To address these issues, we propose a fast road damage detection model named RDD-DETR based on RT-DETR with fusing mutli-scale features. First, we design a Reparameterized Faster Block (RFBlock) to strengthen the backbone network’s feature extraction capability while substantially reducing computational complexity. Second, To achieve High-Low Frequency Intra-scale Feature Interaction (HLFFI), Hilo attention is introduced in the intra-scale feature interaction. It effectively captures local details and global semantics while minimizing information loss during feature extraction by separating high-frequency and low-frequency features of the image. Third, we propose a Multi-scale Damage Fusion (MDF) module to adaptively integrate features across varying scales, enhancing the model’s robustness to road damage scale variations. Additionally, Channel Knowledge Distillation (CWD) is employed to enable self-distillation, enhancing detection accuracy without increasing model parameters or computational complexity. Extensive experiments on Peking University’s SVRDD dataset demonstrate that RDD-DETR achieves SOTA performance, i.e., a 2.3% improvement in mAP50, a 2.4% increase in F1 score, an 18.5% reduction in parameters, and an 8.8% decrease in FLOPs. Generalization tests on the public UAPD dataset further validate the model’s efficacy, with RDD-DETR surpassing the original RT-DETR by 3.3% in mAP50. These results underscore the proposed model’s superior accuracy, efficiency, and adaptability for road damage detection in diverse scenarios.