Learning Chinese Character One-To--Many Style Transformation And Generation By Generative Adversarial Networks

With the continuous development and advancement of information technology,peo-ple nowadays have developed ever-glowing reliance on computers.All kinds of document and files are gradually transformed from handwritten version to printed version.People need to transform different styles of fonts every day,but the current style transformation can be applied to text files only.Consequently,it is troublesome to transform font styles in scanned documents.Besides,Chinese characters have very complex structure – each Chinese character contains radicals and a variety of strokes.Different radicals and differ-ent strokes can make a variety of Chinese characters based on different arrangements and combinations.As a result,the total number of Chinese characters is very large.Therefore,it's very troublesome for Chinese character style artists to design a new set of font style for Chinese characters,because they need to design font styles of characters which are used daily by most of the population.Therefore,researchers gradually focus on the generation of Chinese characters and transformation of font styles.At present,most of the research on the generation of Chinese characters and the transformation of font styles are one-to-one.One of the disadvantage is that,they cannot transform to multiple styles of fonts at the same time.On the other hand,they cannot specify the font to be transformed due to restrictions of the model.Besides,there are de-fects in font imitation – oftentimes transformed font pictures are not clear enough,strokes are blurred,and strokes are not smooth enough.Based on the above defects,this paper presents a method of Chinese character font style transformation and generation based on the Generative Adversarial Networks.Unlike traditional models that use Generative Adversarial Networks to generate Chinese characters,we have improved the overall net-work structure,added a new style assignment mechanism to the Generative Adversarial Networks,and added a loss of classification function and loss of semantic consistency function to constrain the optimization of our network parameters.Our model can generate different font styles that are specified by users with the font style assignment mechanism(one-to-many).Meanwhile,our model can create a new font by combining the features of different font styles.We have done a lot of experiments on commonly used Chinese character datasets.Compared with previous experiments on Chinese characters genera-tion and migration models,our model produces clearer pictures of Chinese character fonts,smoother strokes,and better reflects the style characteristics of each font.Besides,in our model,users can specify the fonts they want to generate based on their needs.In addition to their desired font style,we can also generate new fonts in the meantime,which could not be done in previous models.Last but not least,during our experiment,we added bal-ance analysis of hyperparameters and detailed comparison of picture quality,which fully shows that our method is effective.