Scholarlink

<p>In recent years, the defect detection for printed circuit board (PCB) has encountered a significant challenge: the disproportionate ratio between defect areas and background regions adversely impacts the efficacy of computer vision detection methodologies. To address this issue, a dual-domain high-frequency adaptive enhanced pooling method (DHAEP) is proposed in this paper, aimed at detecting defects and damages in bare plates. Initially, a multi-scale high-frequency feature fusion network is designed to extract features from images, thereby mitigating the loss of high-frequency information associated with small targets. Subsequently, we introduce local dynamic range adjustments and tunable coefficients in the frequency domain to facilitate dynamic enhancement of multi-scale high-frequency information. Following this, the feature information from both spatial and frequency domains is integrated to achieve comprehensive multi-level and multidirectional dual-domain feature fusion. Finally, in the public PCB defect dataset as well experimental results derived from extensive industrial PCB image datasets demonstrate that our proposed method outperforms other pooling techniques in terms of detection accuracy.</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

Pcb defect detection based on a novel dual-domain high-frequency adaptive enhanced pooling method

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Kun Li,
  • Xu Zhong,
  • Ying Han

摘要

In recent years, the defect detection for printed circuit board (PCB) has encountered a significant challenge: the disproportionate ratio between defect areas and background regions adversely impacts the efficacy of computer vision detection methodologies. To address this issue, a dual-domain high-frequency adaptive enhanced pooling method (DHAEP) is proposed in this paper, aimed at detecting defects and damages in bare plates. Initially, a multi-scale high-frequency feature fusion network is designed to extract features from images, thereby mitigating the loss of high-frequency information associated with small targets. Subsequently, we introduce local dynamic range adjustments and tunable coefficients in the frequency domain to facilitate dynamic enhancement of multi-scale high-frequency information. Following this, the feature information from both spatial and frequency domains is integrated to achieve comprehensive multi-level and multidirectional dual-domain feature fusion. Finally, in the public PCB defect dataset as well experimental results derived from extensive industrial PCB image datasets demonstrate that our proposed method outperforms other pooling techniques in terms of detection accuracy.