Design And Mplementation Of Handwritten Font Style Migration Model Based On Generative Adversarial Network

Writing is an important means of recording and transmitting information in human history.It plays an important role in human history for thousands of years.The research on characters is also rich and diverse,and the style transfer of fonts is one of them.In the new era,people have higher and higher requirements for aesthetics.In many cases,they don’t want to just use monotonous printing,and then designing a set of fonts requires a lot of work by professionals.The style migration of fonts can improve the efficiency of font design,and can also be used in current instant chat tools to let everyone have their own unique fonts,Improve the cold characteristics of the unified font of the machine and bring more temperature to our daily communication.Both the traditional font style migration method based on font decomposition and reorganization and the recent font style migration method based on neural network have made some progress.At the same time,there are many problems.The generation quality of font image is poor and the requirements for source font and target font are high.In short,there is still a long way to go for font style migration,which needs more in-depth research.The font style migration model studied in this paper is also based on image-to-image font style migration.The mapping between different domains of the image is completed through neural network,so as to achieve the purpose of font style migration.The essence is image-to-image conversion.The main work of this paper is as follows:1.A one to many font style transfer model is proposed to effectively solve the problem of over fitting in the training process of one to one font style transfer model.One to many font style migration model and the ordinary deep learning style font style migration model do not require the corresponding correspondence between target text and input text,and can be trained asymmetrically.2.Improve the generation model by adding HDC model and attention mechanism on the basis of RESNET.RESNET can improve the fitting accuracy of the generation model,and HDC expands the field of vision of a single convolution kernel.The superiority of HDC model over traditional convolutional neural network is verified on the experimental data set.The training time of adding HDC model is significantly reduced.3.The content discriminator and the style discriminator perform their respective duties and cooperate organically to optimize the quality of font images generated by the generation model.The content discriminator and style discriminator can generate the text of the target font in the training process by discriminating the input content of different dimensions.4.Based on the one to many font style migration model proposed in this paper,a complete font style migration system is designed and implemented on the basis of NoesisGui framework.