摘要
为了提高输电线路的巡检效率,保证输电线的分割精度和速度,本文提出基于改进U-Net的轻量级网络GU-Net。首先,以U-Net网络为基础,在编码器部分引入轻量化主干提取网络Ghost-Net;然后采用双线性插值方法完成上采样,并利用深度可分离卷积代替部分普通卷积;最后在训练过程引入多损失函数以解决输电线和背景像素占比不平衡问题,并采用迁移学习策略训练模型。在E-Wire输电线数据集上测试,GU-Net网络的MIoU和F1-score分别为80.04%和87.77%,与现有的轻量化输电线语义分割网络Wire-Detection相比分别提升了4.26%和2.96%,且分割速度几乎没有损失,参数量约是它的20%。实验结果表明,本文提出的算法能够实现快速高效、轻量化地分割出复杂图像中的输电线。
To improve the inspection efficiency of transmission lines and ensure the segmentation accuracy and speed of transmission lines,this paper proposes GU-Net,a lightweight network based on improved U-Net.Firstly,based on the U-Net network,the lightweight trunk extraction network Ghost-Net is introduced in the encoder part;then a bilinear interpolation method to complete the up-sampling and use the depth-separable convolution to replace part of the ordinary convolution;finally,introduce multiple loss functions in the training process to solve the imbalance between the transmission line and the background pixel occupancy,and train the model with a migration learning strategy.Tested on the E-Wire transmission line dataset,the MIoU and F1-score of the GU-Net network are 80.04% and 87.77%,respectively,which are 4.26% and 2.96% better than Wire-Detection,an existing semantic segmentation network for lightweight transmission lines,with almost no loss in the segmentation speed,and the number of references is about 20% of it.The experimental results show that the algorithm proposed in this paper can achieve fast,efficient and lightweight segmentation of transmission lines in complex images.
作者
胡冠华
张永雷
申立群
Hu Guanhua;Zhang Yonglei;Shen Liqun(School of Mechanical Engineering&Automation,Northeastern University,Shenyang 110819,China;School of Instrumentation Science and Engineering,Harbin Institute of Technology,Harbin 150001,China)
出处
《电子测量与仪器学报》
CSCD
北大核心
2024年第1期211-218,共8页
Journal of Electronic Measurement and Instrumentation
基金
国家电网有限公司总部科技项目(52170218000G)资助。
关键词
深度学习
语义分割
编解码网络
轻量化
输电线路
deep learning
semantic segmentation
codec networks
light weighting
power line
作者简介
胡冠华,现为东北大学本科生,主要研究方向为人工智能、机器视觉等。E-mail:hu_guanhua@163.com;通信作者:张永雷,2023年于东北林业大学获得学士学位,现为哈尔滨工业大学硕士研究生,主要研究方向为自动检测与图像处理。E-mail:zyl200109@yeah.net;申立群,2002年于哈尔滨工业大学获得学士学位,2004年于哈尔滨工业大学获得硕士学位,2008年于哈尔滨工业大学获得博士学位,现为哈尔滨工业大学副教授,主要研究方向为自动检测与控制技术。E-mail:shenliqun@hit.edu.cn。
