Defect Detection in Printed Circuit Boards Using Semi-Supervised Learning.

Area of Science:

• Semiconductor manufacturing
• Artificial intelligence
• Computer vision

Background:

• Automated defect inspection is crucial for semiconductor manufacturing to ensure high-quality printed circuit boards (PCBs).
• Traditional inspection methods are often manual, leading to increased labor costs and time consumption.
• Existing automated systems can be limited by the need for extensive labeled data.

Purpose of the Study:

• To develop a semi-supervised learning (SSL) model for efficient and accurate PCB defect detection.
• To reduce the reliance on large amounts of manually labeled data in automated inspection.
• To enhance the robustness and generalization capabilities of defect detection models.

Main Methods:

• Developed a semi-supervised learning (SSL) model named PCB_SS.
• Trained the model using both labeled and unlabeled PCB images with data augmentation.
• Compared the performance of PCB_SS against a fully supervised model (PCB_FS) and other classifiers.

Main Results:

• The PCB_SS model demonstrated superior performance compared to the fully supervised PCB_FS model.
• PCB_SS exhibited greater robustness with limited or incorrectly labeled data, maintaining stable accuracy.
• The inclusion of unlabeled data improved the deep learning model's generalization and defect detection performance.

Conclusions:

• The proposed PCB_SS model offers a rapid and accurate automated solution for PCB defect inspection.
• This SSL approach effectively alleviates the burden of manual data labeling in semiconductor manufacturing.
• The method provides a more resilient and efficient alternative for defect detection systems.