LarGAN: A Label Auto-Rescaling Generation Adversarial Network for Rare Surface Defects

Abstract

The emergence of single-image generation (SIG) has opened up new possibilities for generative models, making it feasible to generate small datasets that were previously impractical. This paper presents LarGAN, a generative model designed specifically for generating images of rare defects, such as casting slabs, and explores its utility in the context of data augmentation and defect detection tasks. LarGAN model leverages a progressive training framework and an adaptive label auto-scaling method to produce defect images that closely resemble the input image, using only a single defect image as input. The results of the experiments demonstrate that LarGAN outperforms other single-image generative models in terms of both image quality and diversity. Moreover, the generated data can supplement the missing independent data distribution in the original dataset, rendering it particularly suitable for data augmentation and defect detection tasks, particularly when the availability of sample data is limited. Additionally, the experimental results indicate that the data generated by LarGAN can effectively augment the feature space of the original dataset, thereby improving the accuracy and generalization performance of the detection model. As such, this study provides a valuable generation method for detection models utilized in industrial contexts that require substantial amounts of data.