摘要
针对路面损伤实时检测时,检测计算成本高、算法体积大等问题,以YOLOv8n的网络结构为基础,设计了一种面向路面损伤轻量化检测算法AUG-YOLOv8-D。该算法采用Adown模块替代YOLOv8n中卷积模块,引入UIB模块优化C2f模块,在骨干网络末端引入多模态注意力机制;然后改进头部并引入Ghost模块构建Ghost-Head检测头;最后通过通道级知识蒸馏优化检测算法性能。为验证改进有效性,实验数据集选用GRDDC2022中的China_DroneBike图像集。AUG-YOLOv8-D与YOLOv8n相比,mAP50,mAP50-90和FPS分别提高0.4%,1%和10.26%,Params,GFLOPs和MS分别减少54.84%,48.15%和50.79%。AUG-YOLOv8-D在计算性能和成本之间取得很好平衡,降低了边缘设备部署的计算要求,为路面损伤智能检测提供了技术参考。
To address high-cost detection,large-sized algorithms in real-time road-damage detection,a lightweight algorithm,AUG-YOLOv8-D,is designed based on YOLOv8n.It replaces YOLOv8n's convolutional module with Adown,optimizes C2f with UIB,adds multi-modal attention to the backbone's end,improves the head with a Ghost-Head using Ghost modules,and enhances performance via channel-level knowledge distillation.Using the China_DroneBike dataset from GRDDC2022,AUG-YOLOv8-D outperforms YOLOv8n:mAP50,mAP50-90 up by 0.4%and 1%respectively,FPS up by 10.26%,while Params,GFLOPs,and MS down by 54.84%,48.15%,and 50.79%respectively.AUG-YOLOv8-D balances performance and cost,eases edge-device computing,offering a technical reference for road-damage intelligent detection.
作者
邵军
潘道远
高清振
龚智强
SHAO Jun;PAN Daoyuan;GAO Qingzhen;GONG Zhiqiang(School of Mechanical and Automotive Engineering,Anhui Polytechnic University,Wuhu Anhui 241000,China;Intelligent Manufacturing and Elevator College,Huzhou Vocational&Technical College,Huzhou Zhejiang 313000,China)
出处
《佳木斯大学学报(自然科学版)》
2025年第7期5-8,69,共5页
Journal of Jiamusi University:Natural Science Edition
基金
国家自然科学基金项目(51805001)
浙江省教育厅科研项目资助(Y202352845)
湖州市自然科学基金(2024YZ55)。
