文章摘要
刘思言,王博,高昆仑,等.基于R-FCN的航拍巡检图像目标检测方法[J].电力系统自动化. DOI: 10.7500/AEPS20180921005.
LIU Siyan,Wang Bo,GAO Kunlun, et al.Object Detection Method for Inspection Aerial Image of Transmission Line based on Region-based Fully Convolutional Network[J].Automation of Electric Power Systems. DOI: 10.7500/AEPS20180921005.
基于R-FCN的航拍巡检图像目标检测方法
Object Detection Method for Inspection Aerial Image of Transmission Line based on Region-based Fully Convolutional Network
DOI:10.7500/AEPS20180921005
关键词: 深度学习  R-FCN  目标检测  航拍巡检  缺陷识别
KeyWords: deep learning  R-FCN  object detection  aerial image inspection  fault detection
上网日期:2019-05-15
基金项目:国家电网公司科技项目(合同编号SGGR0000JSJS1800569)
作者单位E-mail
刘思言 全球能源互联网研究院有限公司 liusiyan@geiri.sgcc.com.cn 
王博 全球能源互联网研究院有限公司 flish_wang@sina.com 
高昆仑 全球能源互联网研究院有限公司 gkl@geiri.sgcc.com.cn 
王岳 全球能源互联网研究院有限公司 wangyue@geiri.sgcc.com.cn 
高畅 全球能源互联网研究院有限公司 gaochang@geiri.sgcc.com.cn 
陈江琦 全球能源互联网研究院有限公司 chenjiangqi@geiri.sgcc.com.cn 
摘要:
      航拍巡检是输电线路巡检的主要方式之一,目前的航拍巡检方式效率较低,受巡检员主观因素影响大,亟需一种智能检测算法自动定位并识别输电线路巡检图片中的故障,基于深度学习的航拍巡检图像目标检测技术作为一种可能的解决方案,得到了广泛关注。提出了一种利用基于区域的全卷积网络(Region-based Fully Convolutional Network, R-FCN)的航拍巡检图像目标检测方法,并利用在线困难样本挖掘(Online Hard Example Mining, OHEM)、样本优化、软性非极大值抑制(Soft Non-Maximum Suppression, Soft-NMS)等改进方法进行优化。实验证明,本文所提方法与当前已有方法相比,具有目标定位准确、平均准确率高、单模型可同时检测目标种类多等明显优势。
Abstract:
      Aerial photography is one of the main means of transmission line inspection. In consideration of inefficient and subjective factors of manual aerial image inspection, there is an urgent need for an intelligent detection algorithm to locate and to identify faults in the transmission line inspection pictures. As a possible solution, object detection based on deep learning has received extensive attention. An object detection method utilizing region-based fully convolutional network (R-FCN) is proposed. Online hard example mining (OHEM), sample adjusting, and soft non-maximum suppression (Soft-NMS) are adopted to improve performance of the proposed algorithm. The experiment results show that comparing with existing methods the proposed method has obvious advantages on localization, average precision, and variety of detectable objects.
查看全文   查看/发表评论  下载PDF阅读器