Road Defect Detection Model Based on Improved YOLO11

Current road defect detection models are difficult to deploy in practice due to their high complexity and computational requirements. Therefore, an improved model based on YOLO11 is proposed. Firstly, the backbone network is optimized through the use of receptive field attention convolution and a coordinated attention mechanism, which dynamically adjusts the receptive field and enhances feature capture capabilities. Secondly, the neck network incorporates the MARA feature fusion structure, combining MAFPN and CARAFE modules to effectively fuse multi-scale features and improve the detection capability for small targets. Finally, the detection head employs an SEAM to adapt to complex scene detection. Experimental results demonstrate that, on the RDD2022 dataset, the proposed model achieves a 3.1% improvement in mAP50 compared to the baseline model.