Defect detection of printed circuit board assembly based on YOLOv5

Area of Science:

• Computer Vision
• Manufacturing Technology
• Artificial Intelligence

Background:

• Printed Circuit Board Assembly (PCBA) defect detection is vital for manufacturing efficiency.
• Existing methods may lack the required accuracy and speed for real-time applications.
• The YOLOv5 architecture offers a strong foundation for object detection tasks.

Purpose of the Study:

• To propose an improved YOLOv5-based method, PCBA-YOLO, for enhanced PCBA defect detection.
• To increase the accuracy and efficiency of automated defect identification in PCBA manufacturing.
• To introduce a new dataset, PCBA-DET, for training and evaluating PCBA defect detection models.

Main Methods:

• Modification of the YOLOv5 neck network by replacing the spatial pyramid pooling module with a cross-stage partial structure for multi-scale feature capture.
• Integration of large kernel convolution in the backbone network to expand the effective receptive field while minimizing computational cost.
• Utilization of an SIoU loss function to accelerate model convergence by incorporating angular cost.

Main Results:

• The proposed PCBA-YOLO method demonstrates lower loss values and higher accuracy on PCB defect datasets.
• Achieved a mean average precision (mAP) of 97.3% on the newly created PCBA-DET dataset.
• Realized a real-time detection speed of 322.6 frames per second, balancing accuracy and model size.

Conclusions:

• PCBA-YOLO significantly improves PCBA defect detection accuracy and efficiency.
• The method offers a competitive solution compared to other YOLO series detection networks.
• The developed PCBA-DET dataset and source code are publicly available for further research.