Object detection based on computer vision is essential to accelerate the production of electronic products. However, the automatic detection of defects on PCB surfaces is still a challenging task. Despite the existence of several computer vision-based detectors that address these issues, current detectors face challenges in achieving high detection accuracy and speed. For the training and testing of the neural network, three metrics were considered to evaluate the detection results: precision, recall and mAP, and for the classification the average accuracy was considered. The objective is to propose an approach to detect and classify three categories of solder spheres, in the soldering process of silicon wafers on BGA contained in PCB substrates, combining the YOLOv5 model and a CNN. The experimental results show that the detector achieved considerable performance, scoring a mAP@50 of 92.6% for the YOLOv5 model and an average accuracy of 97.87% for the CNN model.
|
