Character Style Transfer Algorithm Based On CycleGAN And Its Application

At present,Generative Adversarial Networks(GAN)has become a popular model of in-depth learning.GAN has achieved amazing results in image generation,image editing,representation learning and image style transfer and so on.However,at present,the research on the image style transfer focuses on the transfer of oil painting,landscape painting and other image styles,and there is a lack of corresponding research on the transfer of calligraphy character style.This paper extends the technique of style transfer to the style of calligraphy fonts,and proposes a method based on generative antagonism network(GAN).It uses GAN to learn the mapping between Chinese famous calligraphy fonts and common fonts,and then automatically generates any common fonts into corresponding calligraphy fonts.At the same time,handwritten character recognition is a research topic of pattern recognition,at present,the performance of recognition algorithm based on deep learning is much better than traditional methods,but the deep learning algorithm of handwriting font recognition is limited by the difficulty of collecting a large amount of handwriting font data,and the accuracy of handwriting font recognition still has room for improvement.For this reason,using GAN to generate a large number of training data sets for handwritten character recognition with different styles to improve the accuracy of handwritten character recognition.The main innovations of this paper are as follows:(1)A new font style transfer algorithm based on generative adversarial networks is proposed.It is an improved CycleGAN model which combines U-net and Resnet.The network layer concat structure in U-net can overcome the lack of strokes in the process of font style transfer,and Resnet can improve the network depth to extract deeper features without worrying about gradient dispersion.(2)A transfer learning method for handwritten fonts is proposed to solve the difficulty of collecting a large amount of training data for current handwritten character recognition algorithms based on deep network.First,a network model is trained by using above-mentioned character style transfer algorithm to convert printed characters(regular scripts)into handwritten characters.Then,the dataset is switched to another style of handwritten characters,and a few parameters are trained on the basis of the pre-training model to change the data distribution of the network output to fit the second style of handwritten characters,so that a large number of handwritten characters with the second style can be generated from the printed fonts using the second model.With the large number of handwritten characters dataset generated,corresponding recognition models are trained for handwritten recognition.