Enhancing PCB Quality Control with Deep Learning Based Defect Detection

Ensuring high-quality PCBs is crucial for reliable electronic devices. Traditional inspection methods are often time-consuming and prone to errors, leading to the need for more efficient and accurate defect detection techniques. This research explores the use of deep learning models to enhance PCB quality control by automating the detection of defects such as missing holes, mouse bites, short circuits, and spurious copper. The research employed three object detection models: YOLOv8n, Faster R-CNN R50 FPN, and RetinaNet R50 FPN to detect and classify PCB defects. Data augmentation techniques such as horizontal/vertical flipping, grayscale conversion, hue adjustment, noise addition, saturation, blur, and brightness adjustments were applied to further improve model performance. Through comprehensive training on a labeled PCB dataset, the models were evaluated based on metrics such as mAP (mean Average Precision) and total loss. The results showed that YOLOv8n achieved the highest accuracy with a mAP@50 of 89.3% and mAP@50-95 of 45%, along with a superior inference speed of 5.9ms. Faster R-CNN R50 FPN followed with a mAP@50 of 85.9% and mAP@50-95 of 41.9%, and RetinaNet R50 FPN achieved a mAP@50 of 71.6% and mAP@50-95 of 34.5%. The results underscore YOLOv8n's suitability for real-time defect detection due to its combination of high accuracy and fast processing speed, while the other models, although improved after augmentation enhancement, did not match its performance. This research demonstrates the potential of deep learning models for automating PCB defect detection with high accuracy and efficiency, focusing on ordinary single-layered PCB surfaces. While effective, it does not address defect detection in multi-layered PCBs and faces challenges like defect complexity, data imbalance, and limited samples. Future work can explore multi-layered PCBs, improve model robustness and generalization, employ advanced augmentation techniques, and enhance model interpretability for practical reliability.