Research On Calligraphic Character Generation Based On Generative Adversarial Network

Calligraphy,as a treasure of Chinese culture,can enhance the beauty of the text itself and make the text information more easily conveyed and expressed.It is an arduous task to design Chinese characters,and the design of calligraphy and printing fonts requires more careful design by font designers who are proficient in calligraphy theory and techniques.Therefore,many scholars studied Chinese font generation based on traditional statistical methods,but the effect was not satisfactory.In recent years,with the development of deep learning,the effect of using deep neural network to generate images is obviously better than that of traditional methods.Since generative adversarial network has a good effect in image generation,this thesis converts characters into pictures and studies the generation of Chinese characters through deep neural network,and does the following work:(1)the existing networks in the generation of the ancient calligraphy font when still use the simplified method and influence the calligraphy text repair work and study,this article is based on convolutional neural network to design a font style classification networks FCNN,It can identify the font style and type,and then convert the characters that need to be converted into traditional characters by querying the simple and complex comparison table,thus solving the problem that the existing network generates calligraphy fonts into simplified characters.(2)The same Chinese character can appear in a variety of font styles,while the conventional Chinese character generation model can only train and generate fonts of one style at a time.To generate characters of different styles,the model needs to be retrained.In reality,font designers are familiar with a variety of font styles on the basis of mastering the structure and principle of characters.Based on this characteristic,the MFGAN model designed in this thesis trains multiple font styles at the same time to generate multiple font styles at one time.(3)When the model learns multiple font styles at the same time,the styles will be confused and mixed together,and the font classification loss function will be increased to maintain the style of the font itself.(4)Interpolate between the transitions of different font styles to generate new fonts outside of the training font styles.Through the experiment,we can see that the character images generated by the model are clear and the strokes are smooth,which verifies the effect of the model.