Data Augmentation Using Generative Adversarial Networks for Surface Defect Classification in Cold-roll Steel Strip

With the recent increase in demand for steel products, the need for high-quality steel has intensified. Cold-rolled steel strip often exhibit various surface defects due to numerous roll pressing processes. Some of these surface defects can lead to critical issues in steel product manufacturing. Thus, developing a high-performance model for classifying steel sheet defects is crucial for producing and maintaining high-quality metal products. In this paper, we propose a data augmentation model using Generative Adversarial Networks (GANs) to generate images of defects for classes with insufficient data. Additionally, we introduce a pre-processing method designed to minimize the loss of defect information when resizing images. Our experimental results demonstrate that applying the proposed method significantly improves performance, evidenced by an increase of up to 6% in the classification F1-score.