Using Convolutional Neural Networks to detect defects in manufacturing

Abstract

The manufacturing industry is undergoing a technical revolution thanks to advanced develop- ments in deep learning (DL) techniques. This study makes an important contribution to the field of automatic defect detection by applying CNNs, focusing on the YOLOv8 and YOLOv9 models. These two models were designed and trained on a custom dataset of PCB images with the aim of detecting defects with high accuracy and efficiency. The results showed significant superiority of both models over traditional methods in the defect detection task. Moreover, the use of “transfer learning” technology has proven to be very effective in accelerating the training process and significantly improving the performance of the two models while relying on a smaller amount of training data. Through the experiments that we conducted, we recorded a difference in the results obtained for the two models according to the standards. In terms of accuracy, they achieve approximately the same percentage, but in terms of the time taken for training and classification, the YOLOv8 model is better than YOLOv9. As for recall, the YOLOv9 model is better than YOLOv8. In each case, the criterion must be determined. The appropriate trial will determine which of the two models is most effective for detecting defects and practical assistance in manufacturing. CNNs, especially YOLOv8 and YOLOv9, have been able to achieve high accuracy in de- tecting PCB defects. The use of transfer learning also contributed to accelerating the training process and im- proving the performance of the two models.