齐冬莲^1,,,
钱佳莹¹,
闫云凤¹,
曾晓红²
1.浙江大学 杭州 310027
2.西南交通大学 成都 611756
基金项目:浙江省重点研发计划(2019C01001)，国家青年科学基金(62001416)，中央高校基本科研业务费专项基金(2018FZA122)

详细信息

作者简介:齐冬莲：女，1973年生，教授，研究方向为控制理论与控制工程、电气工程
钱佳莹：女，1996年生，硕士生，研究方向为图像处理、深度学习目标检测
闫云凤：女，1988年生，博士，研究方向为神经网络深度学习、图像处理
曾晓红：女，1964年生，高级工程师，研究方向为铁道电气化与自动化

通讯作者:齐冬莲　qidl@zju.edu.cn

中图分类号:TN911.73; TP391.4

计量

文章访问数:267
HTML全文浏览量:104
PDF下载量:29
被引次数:0

出版历程

收稿日期:2020-05-08
修回日期:2021-02-21
网络出版日期:2021-03-30
刊出日期:2021-07-10

A Multi-scale Detection Method for Dropper States in High-speed Railway Contact Network Based on RefineDet Network and Hough Transform

Donglian QI^1,,,
Jiaying QIAN¹,
Yunfeng YAN¹,
Xiaohong ZENG²
1. Zhejiang University, Hangzhou 310027, China
2. Southwest Jiaotong University, Chengdu 611756, China
Funds:The Key Research and Development Plan of Zhejiang Province (2019C01001), The National Youth Science Fund Project (62001416), The Fundamental Research Funds for the Central Universities (2018FZA122)


摘要
摘要:针对高速铁路接触网吊弦的状态检测问题，该文提出一种基于RefineDet网络和霍夫变换的吊弦多尺度定位与识别方法。通过设计RefineDet网络的粗调和精调模块对吊弦整体结构进行定位，采用霍夫变换锁定吊弦中部吊悬线所在直线，并利用旋转因子沿直线方向提取吊悬线区域；以吊悬线区域代替吊弦结构整体区域送入分类网络进行训练，通过所建立的多尺度吊弦状态检测模型，实现吊弦状态的精确识别。实验结果表明，吊弦定位模型的准确率达95.3%以上；霍夫变换可排除无效区域对吊弦状态识别的干扰，提高分类网络的训练速度，吊弦状态识别模型准确率达97.5%以上。
关键词:目标检测/
深度学习/
接触网4C/
缺陷分析/
霍夫变换
Abstract:In order to solve the problems of detection and state analysis of high-speed railway catenary droppers, this paper proposes a multi-scale detection method for dropper states based on Refinedet network and Hough transform. First, the positioning result of droppers through Refinedet network is obtained, and Hough transform is used to locate where the dropper line is; Then the surrounding area of the dropper line is extracted with a ralated twiddle factor. Those extracted areas, replacing the results of detection net, are fed into classification network for training the final dropper state analysis mode. Experiments show that the accuracy of dropper detection model is over 95.3%, and the dropper state analysis model can eliminate the impact of meaningless area pixels while accelerating training process, the final state analysis model achieves a high accuracy over 97.5%.
Key words:Target detection/
Deep learning/
Contact network 4C/
Defect analysis/
Hough transform



PDF全文下载地址:

https://jeit.ac.cn/article/exportPdf?id=da422a18-d480-42fa-958f-f3b8b879f014