摘要
路面巡查是道路养护的重要一环。传统的人工巡查存在主观性强、耗时耗力和安全性差等问题。近年来,基于深度学习的路面病害智能检测技术取得显著的进展,其中云边协同方案因其传输与计算成本的均衡性优点日益成为主流。针对车载巡查系统边缘侧硬件的内存和计算能力有限的情况,研究了轻量化网络结构与模型压缩策略,提出一种改进通道结构化剪枝算法,构建基于YOLOv5模型结构化剪枝的轻量化深度学习网络模型。试验证明,优化剪枝后的模型能够获得与原模型相当的检测性能,可在平均准确率mAP损失1.3%的情况下,降低模型约71%的规模,并提高36.3%的速度,为边缘侧路面病害智能检测提供有效支持。
Regular inspection is an important part of roadway maintenance.Traditional manual inspection has many problems,such as strong subjectivity,time-consuming and labor-intensive,and poor security.In recent years,the intelligent detection technology of pavement diseases based on deep learning has made significant progress,among which the cloud edge cooperation scheme has increasingly become the mainstream due to its advantages of balanced transmission and computing costs.In view of the limited memory and computing power of the edge hardware of the vehicle inspection system,the lightweight network structure and model compression strategy is studied,an improved channel structured pruning algorithm is proposed,and a lightweight deep learning network model based on the structured pruning of YOLOv5 model is constructed.The test shows that the optimized pruning model can achieve the same detection performance as the original model,reduce the scale of the model by about 71%and increase the speed by 36.3%under the average accuracy loss of 1.3%of mAP,and provide effective support for the intelligent detection of pavement diseases on the edge side.
作者
阚倩
孟安鑫
陈李沐
KAN Qian;MENG Anxin;CHEN Limu(School of Electronic Information and Communication,Huazhong University of Science and Technology,Wuhan 430074,China;Shenzhen Urban Transport Planning Center Co.,Ltd,Shenzhen 518000,China)
出处
《交通与运输》
2023年第S01期195-200,208,共7页
Traffic & Transportation
基金
深圳市技术攻关面上项目(编号:JSGG20210802093207022)
关键词
日常巡查
病害检测
人工智能
边缘计算
结构化剪枝
模型压缩
Regular inspection
Diseases detection
Artificial intelligence
Edge compute
Structured pruning
Model compression
