Scholarlink

<p>Printed circuit boards (PCBs) are critical to the quality and reliability of electronic products, necessitating accurate surface defect detection. However, conventional methods are often limited by the scarcity of training data and background noise. To address these challenges, we propose a novel unsupervised defect detection framework, referred to as DDMF, which integrates conditional denoising diffusion and multi-scale feature fusion. First, a pre- and postprocessing strategy is employed to suppress background interference and recover coordinate information. Then, a Feature Modulation Block and a two-dimensional normalizing flow module (FastFlow) are introduced to better match the network to the distribution of normal data. Finally, a dynamic implicit conditioning strategy is designed to suppress the reconstruction of anomalous regions and accelerate inference. Experimental results demonstrate that DDMF achieves high detection accuracy while meeting real-time requirements using only 50 training images. Specifically, it attains image-level area under the receiver operating characteristic curve (AUROC) scores of 100.0% and 99.6%, and pixel-level AUROC scores of 97.5% and 97.7% on the laboratory-collected PCB and MVTec Anomaly Detection (MVTec AD) datasets, respectively.</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

DDMF: a PCB surface defect detection model based on conditional denoising diffusion and multiscale feature fusion

  • 期刊论文

  • 发表日期：

    2025年9月20日

查看全文
  • Wanyu Deng,
  • Luyao Yan,
  • Chenming Wang

摘要

Printed circuit boards (PCBs) are critical to the quality and reliability of electronic products, necessitating accurate surface defect detection. However, conventional methods are often limited by the scarcity of training data and background noise. To address these challenges, we propose a novel unsupervised defect detection framework, referred to as DDMF, which integrates conditional denoising diffusion and multi-scale feature fusion. First, a pre- and postprocessing strategy is employed to suppress background interference and recover coordinate information. Then, a Feature Modulation Block and a two-dimensional normalizing flow module (FastFlow) are introduced to better match the network to the distribution of normal data. Finally, a dynamic implicit conditioning strategy is designed to suppress the reconstruction of anomalous regions and accelerate inference. Experimental results demonstrate that DDMF achieves high detection accuracy while meeting real-time requirements using only 50 training images. Specifically, it attains image-level area under the receiver operating characteristic curve (AUROC) scores of 100.0% and 99.6%, and pixel-level AUROC scores of 97.5% and 97.7% on the laboratory-collected PCB and MVTec Anomaly Detection (MVTec AD) datasets, respectively.