摘要
为了解决道路裂缝自动化检测中目标分布尺度广、特征复杂多样以及需要处理大批量数据集的问题,提出一种基于YOLOv8n的轻量化道路裂缝检测算法GCW-YOLO。首先,将全局注意力机制引入到主干网络中,以增强道路裂缝特征的提取和融合能力;其次,采用Wise-IoU损失函数更换原本的损失函数来获得更好的特征聚焦,降低预测中的特征与分类损失;最后,将轻量化网络结构GhostNet引入残差计算模块,用于提高模型的特征提取效率,降低计算复杂度。实验在自制的高速公路裂缝病害数据集共计15 116张图片上进行训练与预测,并在公开数据集上验证算法的泛化性能。实验结果显示,所提算法平均精度均值达到63.5%,相较于原始模型提高6.0%,算法在空间和时间效率上分别提升3.0%和8.5%,检测速度达到250 f/s。对比实验结果表明,GCW-YOLO算法综合了轻量化和检测精度,并表现出良好的泛化性,在道路养护方面具有良好的实用价值和推广前景。
In view of the wide object distribution scale,complex and diverse features and the demand of dealing with a large number of datasets in automatic road crack detection,a lightweight road crack detection algorithm GCW-YOLO based on YOLOv8n is proposed.Firstly,the global attention mechanism is introduced into the backbone network to enhance the ability to extract and fuse road crack features first,and then the original loss function is replaced with Wise-IoU loss function to get better feature focus and reduce the loss of features and classification in prediction.Finally,the lightweight network structure GhostNet is introduced into the C2f module to improve the feature extraction efficiency of the model and reduce the computational complexity.Experiments were conducted on a self-made expressway crack disease dataset with a total of 15116 images,and the generalization performance of the algorithm was verified on public datasets.Experimental results show that the mean average precision(mAP)of the proposed algorithm reaches 63.5%,which is improved by 6.0%in comparison with that of the original model,its spatial and temporal efficiency is improved by 3.0%and 8.5%,respectively,and its detection speed reaches 250 f/s.The comparative experimental results show that the GCW-YOLO algorithm combines the advantages of lightweight and high detection accuracy,and shows good generalization,so it has good practical value and promotion prospect in road maintenance.
作者
吐尔逊·买买提
邱建卓
朱兴林
徐粒
TURSUN Mamat;QIU Jianzhuo;ZHU Xinglin;XU Li(College of Transportation and Logistics Engineering,Xinjiang Agricultural University,Urumqi 830052,China;Engineering Research Center for Intelligent Transportation,Xinjiang Agricultural University,Urumqi 830052,China)
出处
《现代电子技术》
北大核心
2025年第13期11-19,共9页
Modern Electronics Technique
基金
国家自然科学基金项目(51768071)
新疆交通投资(集团)有限责任公司科技项目:基于沥青路面病害数据及深度学习的病害智慧识别系统研究(JCZXXJAU2023001)。
