Optimizing Deep Learning Inference to Detect PCB Soldering Defects

Rapid and accurate defect detection in printed circuit board (PCB) manufacturing plays a vital role in the quality control (QC) of consumer electronic products. Automated visual inspection is gaining increased popularity owing to low production cost and high portability. Prior works rely on a combination of image subtraction and processing techniques whereas the recent works utilize the generalization capability of deep learning. The former operates poorly in a mismatch condition of image pairs whereas the latter consumes higher computation time and processing power. In this paper, we propose a YOLOv4-based PCB detection framework which detects six types of defects. To attain production-ready capability, we optimize the models using the Intel OpenVINO toolkit. Experiment results reveal that the optimized models yield good accuracy for defect detection with significant execution speed on low-power embedded platform.