A Study on Image Generation by Combining GAN and Cut-and-Paste Method for Appearance Inspection by Machine Learning

Abstract

This study proposes a novel data augmentation method to improve the accuracy of appearance inspection by machine learning, by combining GANs and the cut-and-paste method. When employing this method and training the model with the augmented dataset, the defect detection ratio improved by 2.0pt compared to the case without augmentation. Moreover, the system developed in this study enables automatic annotation of each training data, unlike the conventional supervised learning method that requires manual annotation for each data. This automated system effectively reduces training setup time, human workload, and human variability.