Research On Chinese Font Style Transfer Based On Generative Adversarial Network

Chinese fonts are an important carrier to show the characteristics of Chinese history and culture,and have rich artistic charm and humanistic heritage.It carries the functions of historical inheritance,emotional expression,and artistic dissemination.Therefore,designing a set of Chinese fonts is conducive to the dissemination of Chinese culture,and can also meet the individual needs of individuals for the font library.Because Chinese fonts vary in form,in different application scenarios,Chinese fonts are mainly divided into printed and handwritten fonts,which need to be designed by professional designers word by word.Therefore,designing a set of fonts is a time-consuming and laborious work.Traditional geometric modeling methods are inefficient.With the birth of generative adversarial networks,many studies have gradually regarded the process of font design as a process of image style transfer.This has a new inspiration for the study of font generation.The use of arbitrary fonts to generate fonts with a specified style has extremely high practical significance for the protection of Chinese historical relics and the promotion of Chinese civilization and history.The style transfer of fonts is to achieve image-to-image translation between multiple fonts.However,in the current field,there are problems such as low definition,structural dislocation,and insufficient style diversity in Chinese font generation.Fonts and unpaired fonts are researched and analyzed.The specific contents are as follows:(1)A style transfer model for paired Chinese fonts.Based on the conditional generative adversarial network,the model realizes font generation and style transfer in the case of font pairing.The internal structure of the generator and discriminator of the model and the optimization objectives of each part have been improved to improve the aesthetics and recognition of the generated fonts.(2)A reconstruction loss constraint generator is introduced to improve the font generation quality during style transfer.(3)A style transfer model for unpaired Chinese fonts.The model is based on the recurrent generative adversarial network,which realizes the generation and style transfer of fonts when the fonts are not paired,so that the training samples are no longer limited by the need for pairing.The model uses two sets of generators and discriminators at the same time,and uses two adversarial losses to constrain the two sets of generators and discriminators respectively,and uses cycle consistency loss to optimize the generator,so that the structure of the generated font is stable and the strokes are clear.(4)The encoder-decoder model based on the full residual structure strengthens the feature multiplexing under different scale transformations and reduces the information loss during the encoding and decoding process.