摘要
针对路面缺陷数据尺寸差距较大,数据类别之间的距离较近,背景复杂,漏检误检率高等问题,提出一种基于改进YOLOv8s的路面缺陷检测算法。在主干使用感受野块来全面感知输入数据的内容,充分提取上下文信息,并引入注意力机制,以关注网络有用信息,抑制无用信息;在颈部使用DAMO-YOLO的高效重参数化广义特征金字塔网络(RepGFPN),将高级语义信息和低级空间信息进行充分交互,传递有效的信息,提高检测精度;在颈部使用轻量级的组合混合卷积(GSConv)替换常规卷积,并且引入到C2f模块中,在降低参数量的同时保持检测精度。算法在RDD2022数据集上进行验证,实验结果表明,改进后的YOLOv8s平均检测精度(mAP@0.5)达到78.1%,相比于原模型提高了3.5%,参数量降低了24%,满足路面缺陷检测在精度和速度上的要求。
Aiming at issues such as large dimensional gaps in road surface defect data,close distances between data categories,complex backgrounds,and high rates of missed and false detections,an improved road surface defect detection algorithm based on YOLOv8s is proposed.Receptive field blocks are used in the backbone to comprehensively perceive the content of input data,extract contextual information,and introduce attention mechanisms to focus on useful network information while suppressing irrelevant information.In the neck,RepGFPN from DAMO-YOLO is used to adequately interact between high⁃level semantic information and low⁃level spatial information,facilitating effective information transmission and enhancing detection accuracy.Lightweight Grouped Shuffle Convolution(GSConv)is employed in the neck to replace conventional convolutions and is integrated into the C2f module to reduce parameter count while maintaining detection accuracy.The algorithm is validated on the RDD2022 dataset,with experimental results showing that the improved YOLOv8s achieves an average precision of 78.1%(mAP@0.5),a 3.5%improvement over the original model,with a 24%reduction in parameters,meeting the requirements of road surface defect detection in terms of both accuracy and speed.
作者
朱灵茜
于泳波
毛健
李庆党
孙振
ZHU Lingxi;YU Yongbo;MAO Jian;LI Qingdang;SUN Zhen(School of Data Science,Qingdao University of Science and Technology,Qingdao 266061,China;College of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China;Qingdao Central Hospital,University of Health and Rehabilitation Sciences,Qingdao 266000,China;College of Sino-German Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China)
出处
《电子设计工程》
2026年第1期150-154,共5页
Electronic Design Engineering
基金
山东省重大创新工程(2017CXGC0607)
山东省泰山学者项目(tshw2015042)。
