On Generation Method Of Multi-Style Chinese Characters Based On StarGAN

As an important carrier of the Chinese culture,Chinese character has diverse writing styles,leading to unique aesthetic effects.The damaged literature with missing part is an important part of the Chinese culture,thus its restoration is very valuable,The automatic generation of Chinese characters provides a potential solution to this problem.The automatic generation of Chinese characters which focuses on how to generate Chinese characters of a specified style.The main challenge of problem is the large number of Chinese characters and the complexity of some Chinese characters.The current popular methods of Chinese character automatic generation are based on paired data,that is,the method of supervised learning.In order to get an effective machine learning model,it is essential to establish a large-scale paired training set,which will result in a large amount of human,material and time costs.To address this problem,some recent research work has achieved the automatic generation of Chinese characters based on unpaired training sets by means of the recently emerging Generative Adversarial Network(GAN)technology.Most of these research work focus on the automatic generation of single-style Chinese characters.However,the applications such as calligraphy appreciation involve automatic generation of multi-style Chinese.Therefore,how to realize the automatic generation of multi-style Chinese characters is a very meaningful research topic.Unlike the automatic generation of single-style Chinese characters,the automatic generation of multi-style Chinese characters involves conversion among multiple styles.The existing methods are less efficient in this task of multi-style Chinese character automatic generation,mainly due to the following two aspects.First,in order to learn all mappings between k different font fields,the existing methods of single-style Chinese characters generation must train a generator between every two font fields,resulting in the need to train O(k~2)generators.Second,each generator can only use two training samples of the font field for training,but not the entire training data,which to a certain extent brings about the waste of data resources.Aiming at some shortcomings of existing methods,this paper proposes an effective multi-style Chinese character generation method.In order to overcome the difficulty of lacking the paired training set,this paper uses the current popular GAN to achieve unsupervised training.Based on a star-shape GAN model(StarGAN),this paper deals with the mutual conversion between different feature domains to realize the conversion between multi-attribute styles.Different from the single-style Chinese character generation method,what the Chinese character generation method based on StarGAN model proposed in this paper only needs is to train a generator to realize automatic conversion between multiple style fields,thus alleviating the problem of poor scalability of the single-style Chinese character generation method while dealing with multiple font fields in some extent.To verify the effectiveness of the proposed method,this paper firstly establishes a corresponding data set based on Chinese handwritten database(CASIA-HWDB),in which the data set contains about 30,000pictures.Secondly,different network structures and training techniques in the StarGAN model are considered,including(a)ResNet,(b)ResNet+Spectral Normalization and(c)DenseNet.The experimental results show that the proposed method can effectively generate high-quality multi-style Chinese characters under the above three different conditions.Meanwhile,the comparisons of the three different network scenarios show that the residual network and spectral normalization in the StarGAN model training have a generally better performance.