基于GOLD-YOLO改进YOLOv5模型道路病害检测研究  

Research on road disease detection based on GOLD-YOLO improved YOLOv5 model

作　　者：陈飞宇 张应迁 吴嘉懿 李睿鑫 彭良吉 Chen Feiyu;Zhang Yingqian;Wu Jiayi;Li Ruixin;Peng Liangji(School of Mechanical Engineering,Sichuan University of Science and Engineering,Zigong 643000,China;School of Civil Engineering,Sichuan University of Science and Engineering,Zigong 643000,China;School of Education and Psychological Science,Sichuan University of Science and Engineering,Zigong 643000,China)

机构地区：[1]四川轻化工大学机械工程学院,自贡643000 [2]四川轻化工大学土木工程学院,自贡643000 [3]四川轻化工大学教育与心理科学学院,自贡643000

出　　处：《现代计算机》2024年第19期7-12,共6页Modern Computer

基　　金：四川省科技厅重点项目(2023ZHCG0020);四川轻化工大学大学生创新创业项目(S202210622028)。

摘　　要：随着机动车数量的增加和道路负荷的增大,道路病害问题日益严重,需要及时发现和识别各类病害以保障道路交通安全,但传统YOLOv5采用的FPN信息融合方式可能导致信息损失。因此,结合华为GOLD-YOLO中的Gather-and-Distribute模块对传统YOLOv5的特征融合模块进行改进。实验结果表明,优化后的YOLOv5算法在训练模型时收敛所需的迭代次数有了显著降低,从590次减少到366次,大大提高了训练速度。同时,总体的mAP@0.5也从原来的87.4%提升到了88.7%。With the increase in the number of motor vehicles and road loads,the problem of road diseases is becoming increasingly serious.It is necessary to timely detect and identify various diseases to ensure road traffic safety.However,the FPN information fusion method adopted by the traditional YOLOv5 may lead to information loss.Therefore,the traditional YOLOv5 feature fusion module is improved by combining the Gather and Distribution module in Huawei GOLD-YOLO.The experimental results show that the optimized YOLOv5 algorithm significantly reduces the number of iterations required for convergence during model training,from 590 to 366,greatly improving training speed.Meanwhile,overall mAP@0.5 It has also increased from 87.4%to 88.7%.

关 键 词：YOLOv5 道路病害 目标检测 信息融合 信息损失 网络改性 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程] U418.6[交通运输工程—道路与铁道工程]