IEEE Access (Jan 2024)

PCB-DETR: A Detection Network of PCB Surface Defect With Spatial Attention Offset Module

  • Qifeng Li,
  • Lihao Wu,
  • Huangpei Xiao,
  • Chuangmian Huang

DOI
https://doi.org/10.1109/ACCESS.2024.3486176
Journal volume & issue
Vol. 12
pp. 158436 – 158445

Abstract

Read online

The paper proposes PCB-DETR, a novel detection network for PCB surface defect identification, which enhances the performance of small defect detection through improvements to the Deformable-DETR model. Traditional supervised learning methods demonstrate strong performance in PCB surface defect detection but require large labeled datasets, which are often costly and impractical to obtain. PCB-DETR addresses this challenge by integrating EfficientNet as the backbone and introducing a spatial attention offset module that dynamically adjusts attention maps to enhance detection accuracy for small and scattered defects.The proposed model is evaluated on the PKU-Market-PCB dataset, consisting of six common PCB defect types. Experimental results demonstrate that PCB-DETR outperforms several mainstream object detection models, including Faster R-CNN, YOLOv5, YOLOv7, and the original Deformable-DETR, achieving higher mean Average Precision (mAP) and recall. The incorporation of spatial attention offset module and efficient feature extraction techniques allows PCB-DETR to excel in detecting small and complex defects while maintaining low computational overhead, making it suitable for real-time industrial applications. The study concludes that, compared to existing methods, PCB-DETR improves accuracy and the ability to determine whether defects are present in images, providing a reliable solution for small defect detection in complex manufacturing environments.

Keywords