Automatic PCB Sample Generation and Defect Detection Based on ControlNet and Swin Transformer.

Area of Science:

• Manufacturing Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Printed Circuit Board (PCB) fabrication requires efficient and accurate detection of soldering defects on surface-mounted components.
• Existing methods may struggle with multitarget detection and dataset limitations.

Purpose of the Study:

• To propose a novel method for improving the efficiency and accuracy of multitarget soldering defect detection on PCBs.
• To enhance sample dataset generation and defect feature extraction for improved localization and segmentation.

Main Methods:

• Utilized Stable Diffusion Model and ControlNet for automatic generation of non-independent samples from industrial images.
• Integrated Swin Transformer as a backbone within Cascade Mask R-CNN for enhanced defect feature extraction.
• Developed a method for simultaneous, large-area inspection of all components, rather than individual segmentation.

Main Results:

• Successfully scaled non-independent sample datasets, enabling high-quality dataset generation.
• Achieved accurate recognition of targets and defect types during multitarget PCB inspection.
• Demonstrated improvements in Average Recall (AR) by 2.8% and mean Average Precision (mAP) by 1.9% compared to other models.

Conclusions:

• The proposed method effectively addresses challenges in PCB soldering defect detection.
• The combination of advanced sample generation and Swin Transformer-based detection significantly improves inspection performance.
• This approach offers a robust solution for high-quality, accurate multitarget defect analysis in PCB manufacturing.