EF-RT-DETR: A Efficient Focused Real-Time DETR Model for Pavement Diseases Detection

Pavement disease classification and localisation is the key to intelligent road health monitoring. Aiming at the problems of complex road background, various shapes of disease objects and high hardware resource requirements. An efficient focusing real-time DETR pavement diseases detection model (EF-RT-DETR) is proposed.Firstly, based on RT-DETR, a DBBNCSPELAN backbone network is designed for capturing fine features. Secondly, the feature focusing module is used to replace the AIFI module to effectively reduce the noise impact and improve the local detail processing. Finally, the Efficient Multi-Feature Fusion module is used in the feature fusion stage to facilitate the interaction between local and global features. On RDD2022, the experimental results show that EF-RT-DETR improves 9.5% and 7.1% relative to RT-DETR on mAP$50 and mAP50:95, respectively, with a 34.7% reduction in computation and a 22.0% reduction in the number of parameters, which achieves real-time and accurate detection of pavement diseases. Mathematics Subject Classification (2020) 68T07 · 68U10 · 68T05