Utilizing YOLO Models for Real-World Scenarios: Assessing Novel Mixed Defect Detection Dataset in PCBs

Abstract

In the domain of printed circuit board (PCB) defect detection and classification, the availability of diverse and comprehensive datasets is fundamental for developing effective detection models. However, existing datasets often lack comprehensive labeling and focus on specific defect types, limiting their applicability to real-world scenarios. To address this gap, we introduce a new dataset named ‘dataset for Mixed Defect Detection in PCB’ (MDD_PCB), which includes intentionally induced mixed PCB defects to provide a more realistic representation of practical scenarios. We evaluate the MDD_PCB dataset using YOLO models and implement it successfully for real-time inference on Jetson Nano, achieving enhanced detection capabilities. Our optimized YOLOv5n model trained on the MDD_PCB dataset achieves impressive metrics (accuracy 93%, precision 95%, recall 96%, mAP 95%, F1-score 94%) with a detection speed of 120.69 frames per second (FPS). Real-time deployment on the Jetson Nano demonstrates practical usability with a detection speed of 30 frames per second (FPS). These results underscore the significance of the diverse dataset proposed, which contributes to robust detection solutions and advances in PCB defect detection methodologies.

Keywords

• YOLO models
• PCB defect detection
• classification
• printed circuit boards (PCBs)
• PCB dataset
• NVIDIA Jetson nano