An automatic detection method for road damage in UAV images based on multi-level perception and feature aggregation

Road infrastructure monitoring is critical for transportation safety and maintenance efficiency. However, existing road damage detection methods face challenges in multi-scale target recognition within complex environments, primarily attributed to their inadequate detection accuracy when processing drone-captured images. To overcome these limitations, this paper introduces the MFD-YOLO model. To tackle the insufficient detection accuracy of existing methods under complex backgrounds, this study designs two key structures: (1) A DS-MDPB backbone network that integrates three feature focusing mechanisms—MANet, SimAM, and BRA—to enhance feature extraction capabilities for pavement cracks and irregular distresses. (2) An MFDPN architecture that employs bidirectional feature propagation and dynamic weight allocation strategies, optimizing feature fusion through DG-C and RE-C modules. Experimental results demonstrate that on the RDD2022_China_Drone dataset, the model achieves an mAP50 of 73.1 %, reflecting a 5.5 % relative improvement over the baseline, and a 30.03 % relative enhancement in AP_s. The model also exhibits consistent performance on UAV-PDD2023 and self-constructed datasets, demonstrating cross-scenario generalization capability. This method maintains high efficiency (FPS > 300) while improving detection accuracy for road damage in complex environments, providing reliable technical support for intelligent road maintenance.