Chinese Calligraphy Font Generation Method Based On Generative Adversarial Network

Chinese calligraphy is the oldest script still in circulation in the world,and as a unique artistic expression of Chinese characters,it contains the Chinese people’s perceptions and thoughts on nature,society,life,and carries the cultural accumulation of the last 5,000 years of China.The generation of calligraphic fonts has become the mainstream method.Before the emergence of artificial intelligence technology,domestic researchers used traditional computer graphics as the basis for restoring famous calligraphic fonts,but such generation methods were rigid and lacked intelligence,so there were problems such as low generation efficiency and inability to accurately restore the font structure,which restricted the development and application of calligraphic font generation.Excellent performance has been achieved in the direction of human faces,animation,and landscape images,but due to the large number,diverse styles,and complex structure of Chinese characters themselves,direct use of generative models that perform well on other types of images is not able to achieve ideal generative results.To address the problems of blurred complex structure,inconsistent detail restoration,and loss of stylistic features in the generation of Chinese calligraphic font images by existing models,this paper conducts a study on the generation of Chinese calligraphic font images based on generative adversarial networks,and the main research work includes the following three parts:(1)To address the performance of generative adversarial network models for calligraphic font images.three novel and comprehensive generative adversarial network models,Pix2Pix,CycleGAN and StyleGAN,were selected for the generation of Chinese calligraphic font images and analyzed for their suitability as a base model for subsequent research work.The experimental results show that the generative adversarial network models have the potential to generate images of calligraphic fonts,but since the above models do not have suitable structural design for the characteristics of calligraphic fonts,the generated font images have shortcomings in terms of skeleton structure and style features,and this paper specifies the direction of the subsequent research work to solve these problems.(2)To address the problem that Chinese calligraphic fonts are prone to structural disorder and ambiguity when generating images,this paper proposes a PSCGAN model based on part structure consistency to strengthen the structural constraints on calligraphic fonts.Inspired by the "field grid" in calligraphy practice,we design a part structure consistency module to randomly select parts among standard style fonts and calligraphic style fonts,and when the selected parts of the two styles of fonts are in the same relative position,the structure of the fonts in this part should be as similar as possible,and vice versa,the structure of the fonts in this part should be as different as possible.And according to this module,we design the network parameters of the part structure consistency loss function optimization model.The experimental results show that PSCGAN can accurately preserve the structural features of fonts,and the calligraphic font generation effect is significantly better than the existing font generation models.Finally,the residual network structure and the part structure consistency loss function of the generator in this chapter are ablated to reflect the importance of the residual network and the part structure consistency loss function for PSCGAN.(3)To address the problem that the Chinese calligraphic font generation model can only learn one calligraphic style at a time and the calligraphic style is easily lost,this paper proposes a multi-style font generation adversarial network model MDFGAN based on the idea of feature decoupling,and designs mutually independent structural feature encoder and style feature encoder for extracting structural and stylistic features of calligraphic fonts,respectively,and feeds the extracted two types of feature vectors into the extracted feature vectors are fed into the image reduction decoder to obtain the target calligraphic fonts.The trained style feature encoder can extract the style features of multiple calligraphic fonts at the same time and generate multiple styles of calligraphic fonts with the structural feature information.The experimental results show that the MDFGAN model can generate multiple styles of calligraphic fonts simultaneously,and the styles of different fonts do not affect each other,and the final calligraphic fonts are better than the existing font generation models.