Aerial Road Disease Detection Model Based on SD-YOLOv11

QIAN Canglong; TANG Wenbing

doi:10.13501/j.cnki.42-1908/n.2025.06.002

Abstract：To address the challenges of small object detection and complex background interference in unmanned aerial vehicle(UAV)aerial images,a spatial reconstruction convolution and dynamic upsampling-you only look once version 11(SD-YOLOv11)model was proposed for road surface disease detection.Firstly,the model was built upon the YOLOv11 nano(YOLOv11n)architecture,and a spatial and channel reconstruction convolution(SCConv)module was introduced.This reduced redundant information in both the spatial and channel dimensions of the feature maps,thereby improving the quality of feature extraction.Secondly,during the multi-scale fusion stage,a dynamic upsampler(DySample)was employed to enhance the model's learning capacity and significantly improve small target detection performance.Finally,the″squeeze-and-excitation″attention(SEAttention)mechanism was integrated to reduce background interference and increase the accuracy of distress detection.The results demonstrated that compared with the YOLOv11n model,the SD-YOLOv11 model exhibited improvements of 8.2%and 15.0%in mean average precision at intersection-over-union thresholds of 0.50 and 0.50～0.95,respectively,while computational complexity and parameter count were optimized.The SD-YOLOv11 model not only improved the accuracy of pavement distress detection but also demonstrated advantages in reducing false detection due to complex backgrounds,offering an efficient solution for accurately detecting aerial road distresses.