Few-Shot Defect Image Generation via Defect-Aware Feature Manipulation

The performances of defect inspection have been severely hindered by insufficient images in industries, which can be alleviated generating more samples as data augmentation. We propose the first image generation method challenging few-shot cases. Given just a handful and relatively defect-free ones, our goal is to augment dataset with new images. Our consists two training stages. First, we train data-efficient StyleGAN2 on backbone. Second, attach defect-aware residual blocks backbone, learn produce reasonable masks accordingly manipulate features within masked regions added modules limited Extensive experiments MVTec AD not only validate effectiveness realistic diverse images, but also manifest benefits it brings downstream tasks. Codes are available at https://github.com/Ldhlwh/DFMGAN.