SA-GAN: Chinese Character Style Transfer Based on Skeleton and Attention Model.

The Chinese character transfer task must meet two requirements: the transfer image should retain the content structure information of the original Chinese character as much as possible, and present different reference styles. Some of the earlier methods required training with large amounts of paired data, which was a time-consuming task. The existing method follows the normal form of style-content disentanglement, and realizes style transfer by combining the reference Chinese character style. This method is easy to cause the problem of missing stroke content and inaccurate overall style transfer. To address these issues, a generation network based on the Chinese character skeleton and attention model is proposed. To further ensure the completeness of the content of the converted Chinese characters, a more efficient upsampling module is introduced to improve the quality of the converted Chinese characters. Through extensive experiments, it is shown that the model requires only one reference Chinese character to produce higher quality Chinese character images than the current state-of-the-art methods.