摘要
针对深度学习算法模型中参数多、计算量大、复杂度高的问题,特别是图像处理算法对系统硬件平台要求高,导致其难以在小型移动终端设备上部署应用等问题,在传统的YOLOv4算法模型基础上提出了一种轻量级卷积网络的改进方法,并将其应用于输电线绝缘子视觉检测。利用Ghostnet轻量化模型对YOLOv4的主干网络进行改进,并对改进网络的计算复杂度进行分析;同时,提出了基于改进算法的绝缘子视觉检测流程。搭建了系统检测算法仿真平台,对比了所提算法与常规YOLOv4算法、Mobilenetv3-YOLOv4算法和Ghostnet-YOLOv4算法用于绝缘子检测时的不同效果。实验结果表明,相比其他算法,Ghostnet-YOLOv4算法不仅能够较为精准地检测到绝缘子串,而且在保持检测精度的情况下有效降低了计算量。
Addressing the problems of many parameters,large computation and high complexity in the deep learning algorithm model,especially the image processing algorithm requires high system hardware platform,which makes it difficult to deploy applications on small mobile terminal devices,an improved method of lightweight convolutional network is proposed on the basis of the traditional YOLOv4 algorithm model and applied to transmission line insulator visual inspection.The backbone network of YOLOv4 is improved using the Ghostnet lightweight model,and the computational complexity of the improved network is analyzed.Meanwhile,the insulator visual inspection process based on the improved algorithm is proposed.A system detection algorithm simulation platform is built to compare the different effects of applying the conventional YOLOv4 algorithm,Mobilenetv3-YOLOv4 algorithm and Ghostnet-YOLOv4 algorithm to insulator detection.The experimental results show that the Ghostnet-YOLOv4 algorithm can not only detect insulator strings more accurately than other algorithms,but also effectively reduce the computational effort while maintaining the proper detection accuracy.
作者
严宇
张宏伟
肖奕
江维
YAN Yu;ZHANG Hongwei;XIAO Yi;JIANG Wei(State Grid Hunan Electric Power Company Extra High Voltage Substation Company,Changsha 420100,China;School of Mechanical Engineering and Automation,Wuhan Textile University,Wuhan 430073,China)
出处
《电力科学与工程》
2022年第12期38-46,共9页
Electric Power Science and Engineering
基金
国网湖南省电力有限公司科技项目(5216A322000B)。
关键词
输电线路
绝缘子
图像处理
视觉检测
深度学习
transmission line
insulator
image processing
visual inspection
deep learning
作者简介
严宇(1986-),男,高级工程师,主要研究方向为电力机器人;通信作者:江维(1983-),男,讲师,主要研究方向为电力机器人。
