迁移学习用于牵引变电所视频多目标识别研究  

Research on Multi-Target Recognition of Traction Substation Video Based on Transfer Learning

作　　者：完颜幸幸 傅钦翠 吉鑫[1] WANYAN Xingxing;FU Qincui;JI Xin(School of Electrical Engineering and Automation,East China Jiaotong University,Nanchang 330013,China;Institute of Traffic Information and Control,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学电气与自动化工程学院,南昌330013 [2]华东交通大学交通信息及控制研究所,南昌330013

出　　处：《计算机工程与应用》2019年第24期196-201,共6页Computer Engineering and Applications

基　　金：2017年度江西省教育厅科学技术研究项目（No.GJJ170375）;江西省2018年度科技计划项目（No.20181BBE58011）

摘　　要：为实现牵引变电所视频图像的多目标识别,为牵引变电所的远程智能巡检提供技术支持。基于迁移学习的理论研究,利用SSD(Single Shot Multibox Detector)和YOLOv2(You Only Look Once v2)模型,实现牵引变电所视频图像中高压开关柜的仪表、分合指示灯状态以及隔离开关的分合状态的自动识别。利用TensorFlow平台实现的多目标识别方法识别速度快而且鲁棒性好。In order to realize video image recognition of traction substation and provide technical support for remote intelligent inspection of traction substation.Based on the theory of transfer learning,the SSD(Single Shot Multibox Detector)and YOLOv2(You Only Look Once v2)models are used to realize automatic recognition of meters,split indicator lamp status of High Voltage Switchgear and disconnector status in the video image of traction substation.The multi-target recognition method implemented by TensorFlow platform is fast and robust.

关 键 词：牵引变电所 迁移学习 目标检测 YOLOv2模型 SSD模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]