Scholarlink

<p>Defect detection in printed circuit board (PCB) is critical to quality control in their production process. Conventional detection methods rely on a large number of annotated images, while the acquisition of defect samples is time-consuming and labor-intensive. Few-shot object detection detects novel classes with a few instances and attracts an increased interest. We propose a dynamic selective regulation fusion method (DSRF) for few-shot PCB surface defect detection. Concretely, we design a selective feature enhancement (SFE) module that focuses on foreground information while suppressing irrelevant background details, enabling more effective utilization of useful information in defective samples. Additionally, we introduce a channel regulation transformation (CRT) module that enhances the detection capability of the network by constructing inter-channel relationships and capturing key information about tiny defects. In order to overcome the shortcomings of existing meta-learning methods in query image perception, we introduce a dynamic information fusion (DIF) module to effectively integrate the information in the query branch into the support branch, and to improve the feature expression ability of the support features. Experiments on the PCB datasets demonstrate that our method significantly outperforms state-of-the-art baselines across different sample settings. Our method not only achieves significant improvement in accuracy, but also demonstrates superiority in handling few-shot data, providing an effective solution for PCB defect detection. Code is available at <a href="https://github.com/lydcv/DSRF">https://github.com/lydcv/DSRF</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DSRF: few-shot PCB surface defect detection via dynamic selective regulation fusion

  • 期刊论文

  • 发表日期：

    2025年2月22日

查看全文
  • Yudong Li,
  • Shaoqing Wang,
  • Zihao Jing,
  • Jinghua Zheng,
  • Xiaobo Han,
  • Xiao Zheng,
  • Fuzhen Sun

摘要

Defect detection in printed circuit board (PCB) is critical to quality control in their production process. Conventional detection methods rely on a large number of annotated images, while the acquisition of defect samples is time-consuming and labor-intensive. Few-shot object detection detects novel classes with a few instances and attracts an increased interest. We propose a dynamic selective regulation fusion method (DSRF) for few-shot PCB surface defect detection. Concretely, we design a selective feature enhancement (SFE) module that focuses on foreground information while suppressing irrelevant background details, enabling more effective utilization of useful information in defective samples. Additionally, we introduce a channel regulation transformation (CRT) module that enhances the detection capability of the network by constructing inter-channel relationships and capturing key information about tiny defects. In order to overcome the shortcomings of existing meta-learning methods in query image perception, we introduce a dynamic information fusion (DIF) module to effectively integrate the information in the query branch into the support branch, and to improve the feature expression ability of the support features. Experiments on the PCB datasets demonstrate that our method significantly outperforms state-of-the-art baselines across different sample settings. Our method not only achieves significant improvement in accuracy, but also demonstrates superiority in handling few-shot data, providing an effective solution for PCB defect detection. Code is available at https://github.com/lydcv/DSRF.