Conditional TransGAN-Based Data Augmentation for PCB Electronic Component Inspection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Automated PCB inspection is crucial for electronic product quality assurance.
• Challenges in PCB inspection include small component sizes, complex backgrounds, and limited training data for deep learning models.
• Accurate identification and classification of PCB components and defects are essential for efficient quality control.

Purpose of the Study:

• To propose a generative model, conditional TransGAN (cTransGAN), for data augmentation in PCB inspection tasks.
• To enhance the quantity and diversity of training datasets for improved accuracy in electronic component recognition and defect detection.
• To validate the effectiveness of cTransGAN in improving object detection algorithms for PCB analysis.

Main Methods:

• Developed conditional TransGAN (cTransGAN), integrating conditional GAN and TransGAN architectures.
• Utilized cTransGAN for data augmentation to generate high-quality synthetic images conditioned on class embeddings.
• Conducted extensive experiments on two datasets (PCB component and defect detection) and evaluated Faster R-CNN, YOLO V3, and SCNet algorithms.

Main Results:

• cTransGAN effectively increased the quality and diversity of the training set.
• The proposed method led to superior performance in both PCB component recognition and defect detection tasks.
• Ablation studies confirmed the positive impact of cTransGAN on object detection algorithm performance.

Conclusions:

• Conditional TransGAN (cTransGAN) is a viable solution for addressing data scarcity in PCB inspection.
• Data augmentation using cTransGAN significantly improves the accuracy and robustness of deep learning models for PCB analysis.
• The open-sourced project encourages further research and development in automated PCB quality inspection.