基于元学习的变电设备小样本缺陷图像检测
DOI:
CSTR:
作者:
作者单位:

作者简介:

通讯作者:

中图分类号:

TP391. 4 TH701

基金项目:

江苏省自然科学基金 (BK20231427)、国家自然科学基金 (92066106)、东南大学“至善青年学者”支持计划(中央高校基本科研业务费) (2242022R40022)项目资助


Meta-learning-based few-shot image detection of defects in substation equipment
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    缺陷图像检测是变电设备运维的重要技术手段。 然而由于缺陷样本的稀缺,传统的基于海量数据训练的深度学习模型 在实际应用中面临小样本缺陷检测的挑战。 为此,本文引入元学习思想,提出一种面向变电设备小样本缺陷图像检测的深度学 习模型。 该模型的核心是前端网络权重的优化和面向小样本测试任务的模型微调。 前者通过基于语义信息的任务生成策略, 使模型能够快速适应新任务;后者则通过基于元学习的网络优化方法对模型进行微调,使模型能够在新任务上获得优异性能。 实验结果表明,本文提出的改进方法可以使模型的综合检测精度提升 7. 33% ,新增类别的检测精度提升 11. 48% ,显著改善了模 型对小样本缺陷和新增类别缺陷的检测性能。

    Abstract:

    Defect image detection is an important technical tool for substation equipment operation and maintenance. However, due to the scarcity of defect samples, the traditional deep learning model based on massive data training faces the challenge of few-shot defect detection in practical applications. Therefore, this article introduces the idea of meta-learning and proposes a deep learning model for few-shot defect image detection of substation equipment. The core of the model is the optimization of front-end network weights and model fine-tuning for few-shot testing tasks. The former enables the model to quickly adapt to new tasks through a task generation strategy based on semantic information, while the latter fine-tune the model through a network optimization method based on meta-learning. Therefore, the model can obtain excellent performance on new tasks. The experimental results show that the improved method can enhance the model′s overall detection accuracy by 7. 33% and the detection accuracy of the new categories by 11. 48% , which significantly improves the detection performance on few-shot defects and defects of new categories.

    参考文献
    相似文献
    引证文献
引用本文

仲林林,吴 奇,叶俊杰,高丙团.基于元学习的变电设备小样本缺陷图像检测[J].仪器仪表学报,2024,45(10):154-167

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:
  • 最后修改日期:
  • 录用日期:
  • 在线发布日期: 2025-01-03
  • 出版日期:
文章二维码