Scholarlink

<p>Printed Circuit Board (PCB) is an important part of modern electronic products. However, due to the large number of electronic components and their different shapes, detecting defects on PCB involves many small targets, which may lead to high missed detection or false detection of small target objects and low detection accuracy in complex backgrounds. We introduce an improved network ACSD-YOLO based on You Only Look Once version 10 (YOLOv10) to solve the problems. First, the backbone network replaces the C2f component with Adaptive Channel Gating Block (ACGBlock) to improve its ability to model shallow texture features and enhance the performance of the model with high missed detection or false detection. Secondly, the neck network is changed by adding the Adaptive Dynamic Scale-P2 (ADS-P2) feature layer, which helps to extract the positioning details and accurate position information of small targets and improve the accuracy of the model in complex backgrounds. Finally, a new prediction head named Deep Hierarchical Dynamic Convolutional Network version 4 (DHDCNv4) is designed to further improve the detection accuracy while reducing model parameters and latency. Tests on the PCB dataset show that, compared with YOLOv10, with the coordinated effect of the three components, the new ACSD-YOLO network achieves 94.9% mean average precision (mAP50), 95.3% accuracy, and 53.8% mAP50-95. We confirm the accuracy and effectiveness of our method on the DeepPCB dataset.</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

A study on optimization of PCB defect detection based on ACSD-YOLO

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Qinsheng Du,
  • Chao Shen,
  • Shiyan Zhang,
  • Ningbo Zhang,
  • Zuosheng Du,
  • Xin Guo,
  • Jian Zhao

摘要

Printed Circuit Board (PCB) is an important part of modern electronic products. However, due to the large number of electronic components and their different shapes, detecting defects on PCB involves many small targets, which may lead to high missed detection or false detection of small target objects and low detection accuracy in complex backgrounds. We introduce an improved network ACSD-YOLO based on You Only Look Once version 10 (YOLOv10) to solve the problems. First, the backbone network replaces the C2f component with Adaptive Channel Gating Block (ACGBlock) to improve its ability to model shallow texture features and enhance the performance of the model with high missed detection or false detection. Secondly, the neck network is changed by adding the Adaptive Dynamic Scale-P2 (ADS-P2) feature layer, which helps to extract the positioning details and accurate position information of small targets and improve the accuracy of the model in complex backgrounds. Finally, a new prediction head named Deep Hierarchical Dynamic Convolutional Network version 4 (DHDCNv4) is designed to further improve the detection accuracy while reducing model parameters and latency. Tests on the PCB dataset show that, compared with YOLOv10, with the coordinated effect of the three components, the new ACSD-YOLO network achieves 94.9% mean average precision (mAP50), 95.3% accuracy, and 53.8% mAP50-95. We confirm the accuracy and effectiveness of our method on the DeepPCB dataset.