Publication: Detection of defect regions in solder mask peel off images using you-only-look-once convolutional neural network
Loading...
Date
2022-07-01
Authors
Heng, Poh Yee
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Cosmetics defect such as solder mask peel off on a circuit board can cause critical financial losses for the industry. This defect can be identified either manually through inspection from human or automatically by using a computer vision system. However, the manual inspection brought high false-call rate and high over-reject rate. Thus, a better algorithm is required to these issues. Driven by the evolution of AI, deep learning technology with the object detection technique is very essential for the defect detection in the industries, especially semiconductor manufacturing sector. In this research, a convolutional neural network (CNN) model with You-Only-Look-Once-small (YOLOS) network will be implemented into the solder mask peel off detection from the images provided by Western Digital Malaysia. Each image is cropped into region of interest (ROI) and labelled with no-defect or defect. All the images are then saved in COCO dataset format for further processing such as image pre-processing, training, validating, and testing processes. Finally, YOLOS network is trained to detect the solder mask in circuit board. As a result, the best model performed by YOLO-v5 with nearest-exact activation function in 50 epochs, with the performance achieved 98.70% of average precision, 89.60% of average recall, and 93.93% of F1-score. With this good performance, the YOLO-v5 provides a possible and reliable solution for the development of automated defect detection in solder mask peel off.