首页|基于元度量学习的小样本输电线路图像部件缺陷分类方法

基于元度量学习的小样本输电线路图像部件缺陷分类方法

扫码查看
原文链接
  • 万方数据
  • 维普
对巡检图像缺陷准确分类是输电线路自动巡检领域中的关键技术之一.针对因缺陷类别图片数量少而导致传统深度学习方法容易出现过拟合与精度低的问题,提出了一种基于元度量学习的小样本输电线路图像部件缺陷分类方法.首先,搭建了基于小样本学习的图像分类网络,采用基于深度残差结构的卷积层来增强网络特征表达能力.然后,在度量模块中引入基于皮尔森相似度的k-近邻算法与局部特征描述符重加权机制,以提高网络分类能力.最后,为验证所提方法的有效性,利用巡检图像构成的数据集对本文方法和其他基于元度量学习的小样本分类方法进行实验对比分析.结果表明:该文提出的方法在分类性能上有明显优势.同时,本文算法的平均准确率在每类缺陷的测试样本仅有15张图片的情况下达到80.24%.
Few-shot Power Transmission Line Image Component Defect Classification Method Based on Meta Metric Learning
Accurately classifying defects in inspection images is a crucial technology in the field of automatic inspection of transmission lines.To address the problem of overfitting and low accuracy due to the small number of defect category images in traditional deep learning techniques,we proposed a few-shot classification method for transmission line image component defects based on meta-metric learning.Firstly,we built a few-shot learning-based image classification network using convolutional layers based on deep residual structures that enhance feature representation capability.Then,a Pear-son similarity-based k-nearest neighbor algorithm and a local feature descriptor re-weighting mechanism were introduced in the metric module to improve the classification ability of the network.Finally,to validate the effectiveness of the pro-posed method,experiments were conducted to comparatively analyze this method and other few-shot classification methods based on meta metric learning after using a dataset composed of inspection images.The results indicate that the method proposed in this paper has significant advantages in classification performance.Meanwhile,the average accuracy of the algorithm in each defect category reaches 80.24%with only 15 test images per category.

few-shot classificationmeta-metric learningPearson similaritylocal descriptor re-weightingtransmission line images

董超、张珂、谢志远、石超君、王宁、赵振兵

展开 >

华北电力大学电子与通信工程系,保定 071003

华北电力大学河北省电力物联网技术重点实验室,保定 071003

河北省互感器技术创新中心,保定 071003

中国南方电网有限责任公司超高压输电公司,广州 510000

展开 >

小样本分类 元度量学习 皮尔森相似度 局部描述符重加权 输电线路图像

国家自然科学基金国家自然科学基金国家自然科学基金中央高校基本科研业务费专项资金中央高校基本科研业务费专项资金河北省教育厅在读研究生创新能力培养项目

6187118262076093622060952020MS0992022MS078CXZZBS2022150

2024

10.13336/j.1003-6520.hve.20230202

高电压技术
中国电力科学研究院　中国电机工程学会

高电压技术

CSTPCD北大核心
影响因子：2.32
ISSN：1003-6520
年,卷(期)：2024.50(9)
  • 20