Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (8): 221-229.DOI: 10.3778/j.issn.1002-8331.2010-0033

• Graphics and Image Processing • Previous Articles     Next Articles

Personalized Handwritten Chinese Character Generation Method for Unsupervised Image Translation

LU Peng, CHEN Jinyu, ZOU Guoliang, WAN Ying, ZHENG Zongsheng, WANG Zhenhua   

  1. School of Information, Shanghai Ocean University, Shanghai 201306, China
  • Online:2022-04-15 Published:2022-04-15



  1. 上海海洋大学 信息学院,上海 201306

Abstract: Chinese characters have a large number of characters, most of the researches focus on its recognition and classification. There are few researches on automatic generation of Chinese characters, especially in the absence of a large number of paired characters datasets. This model uses the unpaired Chinese characters datasets to learn the characteristics of personalized handwritten Chinese characters and transform the formula of generating Chinese handwritten characters into a problem of mapping from the existing standard printing font to the personalized handwritten Chinese characters style. Based on the unsupervised image translation model, attention mechanism and adaptive normalization layer are added to enhance the content and the style of personalized Chinese characters image generation and the ability of the discriminator network is improved by modifying the loss function. Experiments are conducted on the CASIA-HWDB handwritten Chinese character dataset and the Lantingxu calligraphy dataset, and the effectiveness of the method is verified by comparing the evaluation indexes of the content accuracy and the style differences.

Key words: generative adversarial network, image-to-image translation, unsupervised learning, generation of handwritten Chinese characters

摘要: 由于汉字拥有大量的字符,大多数对汉字的研究主要集中在汉字的识别和分类问题上,对于生成汉字的研究较少,尤其是在没有大量配对的汉字数据集的情况下。该模型使用内容和风格样式都不匹配的汉字数据集,将生成个性化手写汉字的过程公式化为一个从现有的标准印刷字体到个性化手写汉字样式映射的问题。在基于无监督学习的图像翻译模型的基础上,利用注意力机制和自适应标准化层来增强个性化汉字生成的内容和风格,并且通过改进损失函数提高了判别器网络的判别能力。在CASIA-HWDB手写汉字数据集和兰亭序书法数据集上进行了实验,通过对比内容准确性和风格差异性的评价指标,验证了该方法的有效性。

关键词: 生成对抗网络, 图像翻译, 无监督学习, 手写汉字生成