Industrial Defect Image Generation Based on Improved StyleGAN2

The performance of defect detection in industrial scenarios based on traditional deep learning methods depends on a large number of training data. However, defect images in industrial scenarios are scarce and difficult to collect. In order to solve this issue, a defect image generation algorithm for industrial scenes based on improved StyleGAN2 is proposed. Firstly, we simplify the mapping and synthesis networks to reduce the model parameters. Secondly, we employ a data augmentation mechanism in discriminator. The results of experiments on the MVTec AD dataset prove that the proposed algorithm can effectively generate defect images with limited training data in industrial scenarios. At the same time, compared with other method, the quantitative evaluation indexes FID (Frechet-inception Distance) has been significantly optimized, indicating that the quality of defect image synthesis has been substantially improved.