Scholarlink

<p>In modern industrial settings, the identification of chips on PCB boards is crucial for quality control and efficiency. However, achieving both speed and accuracy in chip detection remains a significant challenge. To address this issue, we propose the YOLO-GSG deep network model, which incorporates several novel modifications to the standard YOLO architecture. The key innovations include the replacement of the ELAN module with the C3Ghostnet module in the backbone network, improving feature extraction and reducing model complexity, and the introduction of the SE attention mechanism to minimize feature loss. Additionally, the GSnet module and GSConv convolution are integrated into the neck network to enhance feature fusion. The experimental results indicate that the YOLO-GSG algorithm achieves a mAP of 99.014%, with precision and recall improvements of 1.080% and 1.446% over the baseline YOLOv7 model. Additionally, the improved model has 24.478M parameters, 61.4 GFLOPs, and a model size of 50.8&#xa0;MB. The model achieves an FPS of 231.55, representing a 12.8% speedup over the baseline. These results indicate that the YOLO-GSG model offers a superior balance of speed and accuracy for chip identification in industrial applications. This study contributes to the advancement of deep learning applications in industrial environments, providing a more efficient and effective tool for quality control in PCB manufacturing.</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

Real-time recognition method for PCB chip targets based on YOLO-GSG

  • 期刊论文

  • 发表日期：

    2025年1月12日

查看全文
  • Zeang Yue,
  • Xun Li,
  • Huilong Zhou,
  • Gaopin Wang,
  • Wenjie Wang

摘要

In modern industrial settings, the identification of chips on PCB boards is crucial for quality control and efficiency. However, achieving both speed and accuracy in chip detection remains a significant challenge. To address this issue, we propose the YOLO-GSG deep network model, which incorporates several novel modifications to the standard YOLO architecture. The key innovations include the replacement of the ELAN module with the C3Ghostnet module in the backbone network, improving feature extraction and reducing model complexity, and the introduction of the SE attention mechanism to minimize feature loss. Additionally, the GSnet module and GSConv convolution are integrated into the neck network to enhance feature fusion. The experimental results indicate that the YOLO-GSG algorithm achieves a mAP of 99.014%, with precision and recall improvements of 1.080% and 1.446% over the baseline YOLOv7 model. Additionally, the improved model has 24.478M parameters, 61.4 GFLOPs, and a model size of 50.8 MB. The model achieves an FPS of 231.55, representing a 12.8% speedup over the baseline. These results indicate that the YOLO-GSG model offers a superior balance of speed and accuracy for chip identification in industrial applications. This study contributes to the advancement of deep learning applications in industrial environments, providing a more efficient and effective tool for quality control in PCB manufacturing.