基于Cascade R-CNN的输电线路关键部件识别  

Identification of Transmission Line Key Components Based on Cascade R-CNN

作　　者：李峻宇 马海霞 罗永超 LI Junyu;MA Haixia;LUO Yongchao(School of Electrical Engineering,Guangzhou City University of Technology,Guangzhou 510800,China;School of Electronic and Information Engineering,Guangzhou City University of Technology,Guangzhou 510800,China)

机构地区：[1]广州城市理工学院电气工程学院,广东广州510800 [2]广州城市理工学院电子信息工程学院,广东广州510800

出　　处：《微型电脑应用》2023年第5期32-35,共4页Microcomputer Applications

基　　金：广东省制造装备数字化重点实验室开放课题(2020B1212060014)。

摘　　要：针对无人机在输电线路关键部件巡检图像中目标多,且尺度相差较大,导致识别率低的问题,提出使用深度学习目标检测算法Cascade R-CNN进行识别。由于输电线路关键部件无公开数据集,采集输电线路关键部件图片,并将图片数据的分辨率进行统一,使用LabelImg软件进行标注,制作一个符合训练要求的数据集。基于Paddle框架进行模型搭建,使用PaddleDetection训练工具进行全流程训练调优,进行测试。实验结果表明,基于Cascade R-CNN算法的模型在进行目标尺寸相差较大的多目标检测时,在测试集上的精度(mAP)可以达到91.39%,检测效果较好。Aiming at the problems of low recognition rate due to multiple targets and large difference in inspection images of key components of UAV transmission lines,a deep learning target detection algorithm,Cascade R-CNN,is proposed.Since there is no public data set for the key components of transmission lines,pictures of the key components of transmission lines are collected,and the resolution of the picture data is unified.LabelImg software is used to mark the data,and a data set that meets the training requirements is made.The model is built based on Paddle framework,and the PaddleDetection training tool is used for training and tuning of the whole process.The test is carried out.The experimental results show that when the model based on Cascade R-CNN algorithm performs multi-target detection with large target size difference,the precision(mAP)of the test set can reach 91.39%,and the detection effect is good.

关 键 词：输电线路 Cascade R-CNN 深度学习 目标检测 

分 类 号：TP37[自动化与计算机技术—计算机系统结构]