RDL-YOLO: an efficient algorithm for detecting road surface defects

Road surface defect detection is essential for intelligent transportation, autonomous driving safety, and preventive road maintenance. However, accurately identifying potholes and protrusions under complex illumination, irregular morphology, and multi-scale variations remains challenging, especially for lightweight models deployed on edge devices. This paper proposes RDL-YOLO, an efficient and lightweight detection framework built upon the YOLO11n architecture. The model integrates the C3k2_RVB and C3k2_REVB modules to enhance multi-scale feature extraction and improve robustness under low-light conditions. A shared detection head LSCD_LQE is introduced to jointly optimize classification and localization accuracy by incorporating learnable quality estimation. The DySample module enhances adaptive feature upsampling and spatial detail restoration, while DWConv reduces computational overhead to achieve real-time inference on resource-constrained platforms. Experimental results on a road-defect dataset demonstrate that RDL-YOLO achieves notable performance gains. Precision increases by 0.63 percent, recall increases by 4.24 percent, the F1-score improves by 1.82 percent, and mAP at 0.5 improves by 1.76 percent compared with the baseline YOLO11n model. Meanwhile, computational complexity is significantly reduced, with GFLOPs decreasing by 44 percent and parameter count decreasing by 25 percent. Cross-dataset generalization tests further verify the robustness and adaptability of the proposed model in diverse road conditions. The results confirm that RDL-YOLO provides an effective and lightweight solution for real-time road defect detection and intelligent vehicle hazard avoidance.