Cuan Font Generation Model Based On Dense Adaptive Generative Adversarial Network

The style transfer technology for Chinese characters has always been a research hotspot in the field of image processing and computer graphics.Most of the traditional Chinese character style transfer techniques synthesize glyphs through stroke recombination,which is time-consuming and labor-intensive due to the high manual participation.In recent years,scholars have begun to use methods based on convolutional neural networks and generative adversarial networks to generate Chinese characters.Such methods can achieve good results in the generation of sans-serif fonts,but with the development and progress of the times,people’s demand for personalized handwritten fonts is increasing.Compared with sans-serif fonts,It is more difficult to generate.As a typical handwritten serif font,Cuan font originated in Qujing,Yunnan,and are now widely used in various places on the streets of Yunnan.It is of great significance to use generative adversarial networks to generate Cuan font.Cuan font have a specific calligraphic style,and the glyph structure is complex and diverse,and the existing models cannot generate them effectively.During the generation process,serious artifacts are prone to appear at the bends of the strokes,accompanied by blurring and distortion,and the phenomenon of pixel sticking between some independent strokes.In view of the shortcomings of existing work,this paper proposes a Cuan font generation model based on dense adaptive generative adversarial network.The main work is as follows:(1)Aiming at the problem of serious artifacts in the overall generation of handwritten serif fonts,this paper proposes a generation model for cursive characters based on dense adaptive generative adversarial networks.Specifically,this paper combines adaptive instance normalization with dense convolutional neural network to obtain dense adaptive convolution block,which is placed between the encoder and decoder,which can strengthen the extraction of features and deepen the learning of the network.ability to reduce the generation of artifacts.(2)Aiming at the problem of ambiguity and distortion in details such as radicals in the generation effect of handwritten serif fonts,this paper proposes an edge loss function to constrain the glyph structure,which can retain as much detail information in the structure of Chinese characters as possible.(3)In order to further make the generated image as close as possible to the real image,in order to further make the generated image as close as possible to the real image,this paper uses the difference between the generated image and the real image in order to solve the problem that the glyph structure is not kept in place and some independent strokes appear in the generation effect of handwritten serif font.A perceptual loss is added in between,which further improves the performance of the generated image.In addition,this paper uses pixel filtering to beautify the glyph after the model generates the image,so that the generated image maintains a certain sharpness.By comparing with other works in the field of font style transfer,this paper effectively verifies the excellent performance of the Cuan font generation model.In addition,this paper also quantitatively compares the effect of model generation through objective evaluation indicators such as SSIM and Euclidean distance,which further proves the advanced nature of the model in this paper.The effectiveness of each improvement in this paper is verified by ablation experiments.The rationality and scientificity of each parameter setting in this paper are verified by parameter analysis experiments.