Research Of Font Style Transfer Based On Generative Adversarial Network

As continuously developing of computer science and pursuing of artistic beauty,the font style transfer technology assisted with computer is of significance.As one of the mainstream art-design technologies,font style transfer technology is widely used in graphic design and magazine typography.It uses a series of editing methods to render ordinary fonts into stylized fonts with various artistic effects,which makes the original monotonous fonts present complex and changeable text effects,and thus becomes more ornamental and practical in related fields.The conventional font style transfer method uses the image editing software such as Photoshop to manually finish the operation of style transfer.The limitations of this method such as complicated software operation,tedious steps in making stylized fonts and incapability of generating stylized fonts in batches lead to high design cost and low efficiency.In order to overcome these shortcomings,some scholars propose a statistics-based font style transfer method,which uses the spatial distribution of the rendering effects as constraints to guide the texture synthesis of the font.However,the texture effect generated by this method is limited to the font skeleton and is not universal.Therefore,how to provide an efficient font style transfer method which is not limited to the text skeleton is an urgent problem.In recent years,with the rapid development of deep learning,the feature extraction algorithm based on deep neural network has made rapid progress in the field of image style transfer,which enables us to use the deep learning method to complete the style transfer of ordinary fonts.In this paper,we first use conditional generative adversarial network to focus on the research of font style transfer.Then we find that the stylized fonts obtained by the stylized model have the defects of poor and blurred texture effect.In order to overcome the shortcomings of the above font style transfer methods,this paper proposes the solutions are as follows:(1)A font style transfer method based on improved conditional generative adversarial network is proposed.Firstly,a small number of handmade datasets are put into the network for training a stylized model,and then we use the model to implement semi-automatic batch style transfer processing,which effectively solves the problems of time-consuming and low efficiency.The experimental results show that although the texture effects generated by this method are varied and are not limited to the text skeleton,they still suffer from the defect of blur.(2)A two-stage font style transfer method based on adversarial learning is proposed.The method consists of two steps,style transfer processing and sharpness processing.First,we build a stylized network model to render various effects on input fonts.Then,we build a sharpening network model to sharpen the generated stylized fonts,which overcomes the defect of fuzzy fonts generation in a single stylized network.The experimental results show that,our method achieves high performance for generating more clear stylized fonts,whose texture details are relatively rich,and has a higher practical value.