Research On Unsupervised Multi-style Chinese Character Generation Algorithm Based On GAN

Chinese characters are an important carrier for transmitting and expressing information.All kinds of fonts are widely used in press and publication,radio,film and television,advertising design and other industries.Font design has a broad market.The structure of Chinese characters is complex and numerous.The minimum character coding standard to be followed in font design is GB2312.It can basically solve the problem of using computers to process Chinese characters,which is composed of 6763 Chinese characters.Therefore,Chinese font design and font library creation is a huge workload and many repetitive projects.At present,the advanced Chinese character generation method is mainly based on the generation adversarial network,which is optimized to unsupervised learning and multi domain transformation.Most methods only meet one of them,and there are some problems,such as the need for additional information guidance,complex multi-stage model training,poor visual quality of generated results,and unable to deal with the deformation of stroke position or writing mode in complex fonts.Aiming at the above problems,this paper takes unsupervised multi style Chinese character generation as the research content,and puts forward two unsupervised multi style Chinese character generation methods based on generation adversarial network with the goal of model simplification and quality optimization.The main work is as follows:1)An unsupervised multi-style Chinese character generation algorithm based on feature reuse is designed.Based on Star GAN v2,this paper establishes a single-order training model,further improves the Chinese character generation algorithm(Ada IN-SC and CodeConsistence Loss GAN,ACCL-GAN)for simple fonts by using the loss of style code consistency and the adaptive instances skip connections.First,replace the loss of style diversity with the loss of style code consistency to solve the problem of unstable results and incorrect structure.Then,adaptive instance normalized skip connections are added to the codec structure of the generator,the features of the down-sampling process are cached and stylized,and then passed to the up-sampling process to improve the loss of stroke details through feature reuse.Experiments were carried out on three simple font unpaired datasets,and several advanced methods of Chinese character generation and image translation were compared.The results show that ACCL-GAN can generate Chinese characters with more consistent content,more similar style and closer visual quality to the true image.2)An unsupervised multi-style Chinese character generation algorithm is designed to enhance shape distortion.Based on ACCL-GAN,a Chinese character generation algorithm(Dilated Res Block and MS-SSIM GAN,DRMS-GAN)for complex Fonts is obtained by using dilated convolution residual blocks and multiscale structural similarity loss to enhance the deformation of the network.First,the dilated convolution residuals block is used in the discriminator to obtain a larger perception field and a more context-aware generator through antagonistic training.Then,by introducing the loss of multiscale structure similarity,the geometric differences are identified by area statistics to guide the network to better handle shape changes.The performance of the network is discussed on three complex font unpaired datasets and compared with several advanced methods of Chinese character generation and image translation.The results show that DRMS-GAN has good ability of processing Chinese character structure deformation and can produce excellent results in visual quality.