Chinese character style transfer based on improved StarGAN v2 network

Chinese character generation has attracted a lot of attention due to its wide range of applications. Mainstream methods for generating Chinese character fonts are mainly based on generative adversarial networks, however, the structure of Chinese characters is more complex than other fonts and the problem of font structure change and style loss occurs when generating complex fonts and the mainstream methods require paired datasets, which is difficult and time-consuming to collect paired datasets. This paper proposes Trans-StarGAN v2 network for the above problems, which is based on StarGAN v2, introduces the Transformer structure for spatial feature extraction and channel feature extraction, which improves the feature extraction and generation ability of the network, and secondly, introduces the perceptual loss to strengthen the model training process. The experimental results show that compared with other Chinese character generation networks, the proposed network can generate multiple styles of fonts at the same time, improve the quality of the generated characters, preserve the structure of the fonts and make the style more complete in the face of complex fonts, and improve the FID and LPIPS indexes of the generated Chinese character content. Copyright © The Author(s) 2024. Published by Inderscience Publishers Ltd. This is an Open Access Article distributed under the CC BY license. (http://creativecommons.org/licenses/by/4.0/)