Synthesizing Industrial Defect Images Under Data Imbalance

Defect detection is a crucial technology in the industry that enhances production efficiency within the manufacturing sector. However, obtaining a balanced dataset with sufficient samples of both normal and defect is often challenging and time-consuming. Constructing an unbalanced dataset skewed toward normal samples results in decreased performance and reduced generalization of trained models. Therefore, building an appropriate dataset is essential for effectively training deep models. In this study, we propose a defect image augmentation technique based on generative adversarial networks (GANs), dubbed SyNDGAN, to address the challenges of unbalanced datasets encountered in real-world manufacturing scenarios. Specifically, our SyNDGAN synthesizes defect samples from normal images with given segmentation maps which contain the defect types and location of the defect. We validate our method by utilizing manufacturer data which considers the industrial scenario, with limited data. In our experiments, the proposed method shows superior quality compared to other methods both quantitatively and qualitatively. Furthermore, we demonstrate that synthesized data helps to improve the defect recognition performance, which can be utilized in real-world scenarios.