Scholarlink

<p>Printed circuit board (PCB) defect detection is essential for industrial quality control, yet it remains challenging due to the tiny size and low contrast of defects, as well as complex backgrounds. Existing detectors still struggle with the loss of fine defect details, inefficient global context modeling in repetitive PCB background patterns, and suboptimal multi-scale feature fusion under defect-size variations. To address these issues, we propose HTA-Det, a hybrid CNN–Transformer detector with three synergistic components. First, we propose a detail-preserving backbone that uses cascaded RepNCSP blocks for progressive feature refinement to mitigate the early-stage loss of fine-grained details crucial for tiny defects. The aggregated features are then passed through an attention refinement stage to selectively enhance defect-relevant features. Second, a hybrid polarity-aware linear-attention encoder, enhanced by three submodules, is developed for efficient global context modeling. Finally, an adaptive fusion strategy is designed to learn content-aware and scale-specific weights for feature maps from different levels. Experiments on the PKU-Market-PCB and DeepPCB datasets show that HTA-Det achieves mAP<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <mmultiscripts> <mrow /> <mn>50</mn> <mrow /> </mmultiscripts> </math></EquationSource> </InlineEquation> scores of 99.3% and 99.1%, respectively, with only 9.2M parameters. These results indicate that HTA-Det provides an effective and efficient solution for PCB defect detection.</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

Hybrid CNN-transformer with adaptive multi-scale fusion for tiny PCB defect identification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Yunfei Yan,
  • Gengliang Chen,
  • Ye Gu,
  • Xiaojiang Peng

摘要

Printed circuit board (PCB) defect detection is essential for industrial quality control, yet it remains challenging due to the tiny size and low contrast of defects, as well as complex backgrounds. Existing detectors still struggle with the loss of fine defect details, inefficient global context modeling in repetitive PCB background patterns, and suboptimal multi-scale feature fusion under defect-size variations. To address these issues, we propose HTA-Det, a hybrid CNN–Transformer detector with three synergistic components. First, we propose a detail-preserving backbone that uses cascaded RepNCSP blocks for progressive feature refinement to mitigate the early-stage loss of fine-grained details crucial for tiny defects. The aggregated features are then passed through an attention refinement stage to selectively enhance defect-relevant features. Second, a hybrid polarity-aware linear-attention encoder, enhanced by three submodules, is developed for efficient global context modeling. Finally, an adaptive fusion strategy is designed to learn content-aware and scale-specific weights for feature maps from different levels. Experiments on the PKU-Market-PCB and DeepPCB datasets show that HTA-Det achieves mAP \(_{50}\) 50 scores of 99.3% and 99.1%, respectively, with only 9.2M parameters. These results indicate that HTA-Det provides an effective and efficient solution for PCB defect detection.