Data augmentation for solving industrial recognition tasks with underrepresented defect classes

Abstract

This paper discusses neural network-based data augmentation to increase the performance of neural networks in classification of datasets with underrepresented defect classes. The performance of deep neural networks suffers from an inhomogeneous class distribution in recognition tasks. In particular, applications of deep neural networks to solve quality assurance tasks in industrial production suffer from such unbalanced class distributions. In order to train deep learning networks, a large amount of data is needed to avoid overfitting and to give the network a good generalisation ability. Therefore, a large amount of defect class objects is needed. However, when it comes to producing defect classes, obtaining a dataset for training can be costly. To reduce this costs, artificial intelligence in the form of Generative Adversarial Networks (GANs) can be used to generate images without producing real objects of defect classes. This allows a cost-effective solution for any kind of underrepresented classes. However, the focus of this work is on defect classes. In this paper a comparison of GANs for data augmentation with classical data augmentation methods for simulating images of defect classes in an industrial context is presented. The results show the positive effect of both, classical and GAN-based data augmentation. By applying both methods parallel the best results for defect-class recognition tasks of datasets with underrepresented classes can be achieved.