| The style transfer of Chinese fonts has always been a hot issue in the field of image style transfer.It is a branch of the field of image translation.Different from the style transfer of text fonts in other countries,the complex structure of Chinese characters,its very artistic and delicate brushwork,varied styles and huge number are all great challenges for the relevant research in this field.The traditional Chinese font style research focuses on the stroke details and stroke texture of the font style.So the traditional method has low efficiency and huge workload,which cannot meet the needs of industrialization and commercialization.Since the 21 st century,with the exponential improvement of computer computing power and the emergence of more and more large data sets,deep learning has become increasingly popular.Researchers have begun to apply deep learning algorithms to the field of font style transfer gradually.In 2017,the cyclic consistency generative adversarial network(Cycle GAN)proposed by Zhu et al.realized the first application of style transfer on unpaired images,achieving remarkable results and providing new ideas for the research of font style transfer.However,the current field is still at its exploration stage,the related study of Chinese font style transfer is less.This paper proposes an end-to-end model based on Cycle GAN,focusing on one of the difficulties in current Chinese font style transfer research-the style transfer of calligraphy fonts.According to its font characteristics,two innovative improvements are made to the original model architecture: First,we use multi-level pooling in downsampling process;the second is to add residual branches to the residual structure of the original Cycle GAN.The results have proved that the improved model can not only keep more information of style feature,and improve the results of adversarial training,but also greatly enhance the effect of style transfer.Compared with the original Cycle GAN,The PSNR value of the picture is significantly improved,our improved model can converge faster.At the end of this paper,we put forward some thoughts about the research results,and further improvement strategies for the model are proposed for future implementation. |
