《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (10): 3363-3370.DOI: 10.11772/j.issn.1001-9081.2024091322

• 前沿与综合应用 • 上一篇    

基于迁移学习的小样本绝缘子缺陷检测方法

张红1,2, 谢慷慷1(), 宁霞1, 宋婉莹1,2   

  1. 1.西安科技大学 通信与信息工程学院,西安 710054
    2.西安科技大学 西安市网络融合通信重点实验室,西安 710054
  • 收稿日期:2024-09-13 修回日期:2024-11-28 接受日期:2024-12-02 发布日期:2025-03-14 出版日期:2025-10-10
  • 通讯作者: 谢慷慷
  • 作者简介:张红(1972—),女,陕西西安人,副教授,博士,主要研究方向:信号与信息处理、计算机视觉、机器学习
    谢慷慷(1999—),男,河南周口人,硕士研究生,主要研究方向:深度学习、计算机视觉、小样本目标检测 Email:2501534647@qq.com
    宁霞(1998—),女,陕西汉中人,硕士研究生,主要研究方向:深度学习、计算机视觉
    宋婉莹(1988—),女,陕西西安人,副教授,博士,主要研究方向:SAR图像分类、机器学习、模式识别。
  • 基金资助:
    国家自然科学基金资助项目(61901358)

Few-shot insulator defect detection method based on transfer learning

Hong ZHANG1,2, Kangkang XIE1(), Xia NING1, Wanying SONG1,2   

  1. 1.College of Communication and Information Engineering,Xi’an University of Science and Technology,Xi’an Shaanxi 710054,China
    2.Xi’an Laboratory of Network Convergence Communication,Xi’an University of Science and Technology,Xi’an Shaanxi 710054,China
  • Received:2024-09-13 Revised:2024-11-28 Accepted:2024-12-02 Online:2025-03-14 Published:2025-10-10
  • Contact: Kangkang XIE
  • About author:ZHANG Hong, born in 1972, Ph. D., associate professor. Her research interests include signal and information processing, computer vision, machine learning.
    XIE Kangkang, born in 1999, M. S. candidate. His research interests include deep learning, computer vision, few-shot object detection.
    NING Xia, born in 1998, M. S. candidate. Her research interests include deep learning, computer vision.
    SONG Wanying, born in 1988, Ph. D., associate professor. Her research interests include SAR image classification, machine learning, pattern recognition.
  • Supported by:
    National Natural Science Foundation of China(61901358)

摘要:

针对深度学习缺陷检测方法需要大量标注样本训练,而绝缘子缺陷样本获取困难的问题,提出一种基于迁移学习的小样本绝缘子缺陷检测方法。首先,在主干网络加入高效多尺度注意力(EMA)机制,以增强模型对目标特征的表征能力;其次,构建分层采样的区域建议网络(RPN)在特征金字塔中均匀选择锚框,提高模型对不同尺度下新类对象的捕获能力;最后,设计解耦分类头,并通过正负两个头分别处理正负样本,从而使模型可以更有效地适应新类对象。实验结果表明,与基线方法TFA(Two-stage Fine-tuning Approach)相比,在公共数据集PASCAL VOC上,所提方法对新类的平均精度均值(mAP)(交并比(IoU)为0.5)平均提升了9.5个百分点;在绝缘子缺陷数据集上,所提方法在1-shot、5-shot、10-shot、20-shot和30-shot检测任务中的mAP50分别提高了15.8、12.2、17.4、7.3和7.1个百分点。

关键词: 绝缘子缺陷, 小样本目标检测, 迁移学习, 注意力机制, 解耦分类头

Abstract:

In order to solve the problem that deep learning defect detection methods require many labeled samples for training and insulator defect samples are difficult to obtain, a few-shot insulator defect detection method based on transfer learning was proposed. Firstly, an Efficient Multi-scale Attention (EMA) mechanism was added to the backbone network to enhance the model’s ability to represent target features. Secondly, a hierarchical sampling-based Region Proposal Network (RPN) was constructed to select anchors in the feature pyramid uniformly, thereby improving the model’s ability to capture new class objects at different scales. Finally, the classification heads were decoupled, and the positive and negative samples were processed by the positive and negative heads, respectively, so that the model was able to adapt to the new class of objects more effectively. Experimental results show that compared with the baseline method TFA (Two-stage Fine-tuning Approach), on public dataset PASCAL VOC, the proposed method improves the mean Average Precision (mAP) (with IoU (Intersection over Union) of 0.5) of new class by 9.5 percentage points on average; on the insulator defect dataset, the proposed method has the mAP50 in detection tasks of 1-shot, 5-shot, 10-shot, 20-shot, and 30-shot increased by 15.8, 12.2, 17.4, 7.3 and 7.1 percentage points, respectively.

Key words: insulator defect, few-shot object detection, transfer learning, attention mechanism, decoupled classification head

中图分类号: