DefectGAN: Generate Industrial Defects from Extremely Few Samples

Learning-based defect inspection systems are becoming more and more influential in the industry. However, these approaches are usually restricted by the insufficiency of training data, especially the defect samples. It is natural to generate “artificial defect samples”. In this paper, a novel approach to generate high-quality industrial defect images from extremely few samples is proposed. We design a generative adversarial network (GAN) for this task. Mode collapse is a common problem of GANs when training data is insufficient, thus, fully convolutional network and Markovian discriminator are applied to eliminate this problem. In order to generate images of high resolution, we also propose a novel sampling strategy to balance memory consumption and training quality. Experiments illustrate that our model is capable of generating high-resolution industrial defect images from extremely few samples of various categories. With the generated “artificial samples”, the classification and detection models outperform superior accuracy compared to the same models with the original rare samples. Thus, this approach can increase the intelligence of defect inspection systems.