电工技术学报  2022, Vol. 37 Issue (9): 2230-2240    DOI: 10.19595/j.cnki.1000-6753.tces.211849
“电力装备可靠性与智能化”专题(特约主编:李奎教授 牛峰教授) |
基于改进生成对抗网络的无人机电力杆塔巡检图像异常检测
仲林林1, 胡霞2, 刘柯妤2
1.东南大学电气工程学院 南京 210096;
2.东南大学-蒙纳士大学苏州联合研究生院(东南大学) 苏州 215123
Power Tower Anomaly Detection from Unmanned Aerial Vehicles Inspection Images Based on Improved Generative Adversarial Network
Zhong Linlin1, Hu Xia2, Liu Keyu2
1. School of Electrical Engineering Southeast University Nanjing 210096 China;
2. SEU-Monash Joint Graduate School (Suzhou) Southeast University Suzhou 215123 China
全文: PDF (20235 KB)   HTML
输出: BibTeX | EndNote (RIS)      
摘要 无人机电力线路巡检拍摄的杆塔图像背景复杂且正负样本极不均衡,严重影响了电力杆塔异常检测的准确性。该文提出一种基于压缩激活改进的快速异常检测生成对抗网络(SE-f-AnoGAN),可提高复杂背景下无人机电力杆塔巡检图像异常检测的精度。首先,在f-AnoGAN编码器中引入压缩激活网络(SENet),提取图像中的显著性信息。然后,将生成对抗网络的无监督学习和二分类器的有监督学习有机结合,实现前者特征提取优势和后者判别优势的互补。在此基础上,借助基于迁移学习的优化训练策略进一步有效提升模型在大规模数据集上的泛化性能。实验结果显示,总体样本的检测准确率为95.74%,正负样本的召回率分别达到96.05%和95.36%,证明了SE-f-AnoGAN在异常检测中的有效性。
服务
把本文推荐给朋友
加入我的书架
加入引用管理器
E-mail Alert
RSS
作者相关文章
仲林林
胡霞
刘柯妤
关键词 电力杆塔无人机巡检异常检测生成对抗网络迁移学习不平衡样本    
Abstract:Power tower inspection images captured by unmanned aerial vehicles(UAVs)have complex background, and the normal and abnormal samples are extremely unbalanced, which significantly affect the accuracy of anomaly detection for power towers. In order to improve the anomaly detection accuracy in the complex environment, this paper proposes an anomaly detection method for UAV power tower inspection images based on Squeeze-and-Excitation improved fast unsupervised anomaly detection with generative adversarial network(SE-f-AnoGAN). Firstly, the Squeeze-and-Excitation Network(SENet) is added to the encoder in f-AnoGAN to extract the saliency information of images. Next, a combination of unsupervised learning by generative adversarial network and supervised learning by binary classifier is proposed, which consolidates the advantages of feature extractor and discriminator respectively. On this basis, the optimization training strategy based on transfer learning is used to further effectively improve the generalization of the model on the large datasets. The experimental results show that the accuracy rate of overall samples is 95.74% and the recall rates of positive and negative samples reach 96.05% and 95.36% respectively, which demonstrates the effectiveness of SE-f-AnoGAN in anomaly detection.
Key wordsPower tower    unmanned aerial vehicle (UAV) inspection    anomaly detection    generative adversarial network    transfer learning    unbalanced dataset   
收稿日期: 2021-11-13     
PACS: TM75  
  TP39  
基金资助:江苏省科协青年科技人才托举工程(2021031)、东南大学“至善青年学者”支持计划和中央高校基本科研业务费专项资金资助项目
通讯作者: 仲林林 男,1990年生,副研究员,博士生导师,研究方向为高电压技术、放电等离子体技术、人工智能技术。E-mail:linlin@seu.edu.cn   
作者简介: 胡 霞 女,1998年生,硕士研究生,研究方向为目标检测与异常检测。E-mail:huxia@seu.edu.cn
引用本文:   
仲林林, 胡霞, 刘柯妤. 基于改进生成对抗网络的无人机电力杆塔巡检图像异常检测[J]. 电工技术学报, 2022, 37(9): 2230-2240. Zhong Linlin, Hu Xia, Liu Keyu. Power Tower Anomaly Detection from Unmanned Aerial Vehicles Inspection Images Based on Improved Generative Adversarial Network. Transactions of China Electrotechnical Society, 2022, 37(9): 2230-2240.
链接本文:  
https://dgjsxb.ces-transaction.com/CN/10.19595/j.cnki.1000-6753.tces.211849          https://dgjsxb.ces-transaction.com/CN/Y2022/V37/I9/2230