StyleGANs and Transfer Learning for Generating Synthetic Images in Industrial Applications

Deep learning applications on computer vision involve the use of large-volume and representative data to obtain state-of-the-art results due massive number parameters optimise in deep models. However, are limited with asymmetric distributions industrial rare cases, legal restrictions, high image-acquisition costs. Data augmentation based generative adversarial networks, such as StyleGAN, has arisen a way create training symmetric that may improve generalisation capability built StyleGAN generates highly realistic images variety domains strategy but requires large amount build image generators. Thus, transfer conjunction models used small datasets. there no reports impact pre-trained models, using learning. In this paper, we evaluate model different application domains—training paintings, portraits, Pokémon, bedrooms, cats—to generate target levels content variability: bean seeds (low variability), faces subjects between 5 19 years old (medium charcoal (high variability). We first version publicly available The Fréchet Inception Distance was for evaluating quality synthetic images. found produced good images, being an alternative generating evaluated domains.