Automatic Chinese Character Generation Method Based On Generative Adversarial Networks

As the cultural treasure and spiritual wealth of the Chinese nation,Chinese characters are an important part of the Chinese nation’s culture.Chinese characters have developed a variety of fonts in the long history,creating the diversity and artistry of Chinese characters,and many calligraphers have left many precious copybooks of calligraphy.However,with the passage of time,some calligraphy works have certain deficiencies because they are not well preserved.The automatic generation of Chinese characters will provide some potential solution to this problem.In addition,the task of automatic Chinese character generation can also be used in many practical applications such as calligraphy creation,artistic font generation,handwritten font generation,and signature verification.Different from other languages,the structure of Chinese characters is more complex and numerous,which bring challenge to the generation of Chinese characters.At this stage,Chinese character generation methods can be divided into two types: the first method is mainly based on the local structural features of Chinese characters such as strokes and radicals,while using traditional machine learning methods to automatically generate Chinese character fonts;the second method is mainly based on depth generative model,especially the generative adversarial networks,realizes the automatic generation of end-to-end Chinese fonts.The first type of method is mainly due to the feature engineering that usually requires manual intervention,which requires a lot of manpower and material resources.Although the second type of method can achieve end-to-end training,it still has deficiencies such as mode collapse and lack of guidance in feature extraction,resulting in the performance of generating Chinese characters still needs to be improved.Aiming at the limitation of existing methods,this paper proposes three effective Chinese character font generation models.The main contributions of this article are as follows:(1)Aiming at the single-style Chinese character generation model based on supervised learning,this paper uses the Cycle generative adversarial networks as the baseline model,and provides different Chinese character models for the generative adversarial networks model(GAN)by embedding the stroke information of the Chinese characters,thereby alleviating the mode collapse of GAN in the training process,and improving the quality of the generated Chinese characters.In addition to Cycle generative adversarial networks,the stroke encoding method proposed in this article can also be embedded in other Chinese character generation models,and can also be used in other character generation tasks such as Korean and Japanese.(2)Aiming at the single-style Chinese character generation model based on selfsupervised learning,this paper designs an auxiliary self-supervised task based on "Tianzige",which guides the generation model focus on the structural characteristics of Chinese characters by embedding the geometric transformation of "Tianzige",such as upper and lower structure,left and right structure,etc.This method can help the generation model pay attention to the structure information of Chinese characters without adding additional network structure and artificial labels,and can effectively migrate to other generation models.(3)Aiming at the task of generating multi-style Chinese characters,this paper is based on the "star " generative adversarial networks(Star GAN)to realize the mutual translation between various styles of Chinese characters.In this way,you only need to train one generator at a time to generate multiple styles of Chinese characters.Aiming at the potential mode collapse problem of the generative adversarial network,this paper introduces diversity regularization,which can effectively alleviate the mode collapse problem.In order to verify the effectiveness of the above model,this paper also produced a dataset of various style fonts including the Handwritten Chinese Characters dataset(CASIA-HWDB).This paper designs a series of experiments based on these datasets.The experimental results show that the method proposed in this paper has very good effects.