StrokeGAN: Reducing Mode Collapse in Chinese Font Generation via Stroke Encoding

The generation of stylish Chinese fonts is an important problem involved in many applications. Most existing methods are based on the deep generative models, particularly, adversarial networks (GAN) models. However, these models may suffer from mode collapse issue, which significantly degrades diversity and quality generated results. In this paper, we introduce a one-bit stroke encoding to capture key information characters then incorporate it into CycleGAN, popular model for font generation. As result propose efficient method called StrokeGAN, mainly motivated by observation that contains amount characters. order reconstruct associated characters, stroke-encoding reconstruction loss imposed discriminator. Equipped with such loss, issue CycleGAN can be alleviated, improved preservation strokes effectiveness StrokeGAN demonstrated series tasks over nine datasets different fonts. numerical results demonstrate generally outperforms state-of-the-art terms content recognition accuracies, as well certain error, also generates more realistic