The Style-transfer Of Chinese Character Based On Deep Learning

The font design of Chinses character is a difficult job due to the huge amount of Chinese characters.Designing a new font of Chinese character require a lot of manpower and time costs to draw each character.So,it is necessary to find a method to help the font designing for Chinese character.This article will use the deep learning method to achieve the style transfer task for Chinese character,which will help to improve the font design efficiency.Deep learning is a machine learning methods which simulates the connection among human brain neurons.With the dramatic jump in computer computing capability,deep learning methods have proven to have excellent data processing capability in many areas,such as image processing,natural language processing,automatic driving,and so on.This article tries to take the advantage of the powerful data processing capability of the deep learning method to achieve the style transfer of Chinese character.The goal is essentially a kind of image processing task that needs to generate the required image according to the style and the content featureThis paper focuses on the problem of Chinese font style migration,and adopts deep learning methods.(1)In order to analyze the Chinese character font style,we should determine the nature of the task,difficulties and concept framework to solve the problems.(2)Using the traditional convolutional neural network model structure to perform the Chinese character style migration task.Different regularization algorithm,such as batch normalization,drop out,etc,have been used to train the model.(3)Improve the model according to the shortness of the structure of convolutional neural network.Using the structure of encoder-decoder to redesign the model.Adding the bridge structure between encoder and decoder to promote the transfer of information,and embedding labels to the code to control the class of generated image.(4)Reconstruct the model according to the variational Auto-Encoder and constrain the distribution of the code,so that the model can generate image with Gaussian noise.In addition,the model could achieve smooth transition between different style.