Research On Generating Chinese Calligraphy Characters Based On Generative Adversarial Networks

According to Chinese legend,when Cangjie created Chinese characters,corn were falling from the sky like rain and the fairy kept crying in the night in fear.Since then,Chinese calligraphy boarded the stage of history.After 5,000 years of historical and cultural changes,the Chinese calligraphy characters have evolved from Oracle bone script,Chinese bronze inscriptions,big seal character,small seal character to cursive script,regular script and running script in the late Eastern Han Dynasty.These Chinese calligraphy styles are not only important carriers of past history,but also an important proof of the rich creativity and imagination of Chinese people.However,Chinese calligraphy that uses rice paper and steles as carriers are easy to be destroyed and disappeared,which influence the effective inheritance of Chinese calligraphy.So it is meaningful to use modern technology to virtual restore and supplements Chinese calligraphy.Since the 21 st century,the rapid development of artificial intelligence technology provides necessary conditions for the virtual restoration and supplement of Chinese calligraphy.Traditional Chinese calligraphy characters generative methods and current artificial intelligence image generative models are discussed in the thesis,analyzed the pros and cons of them,and proposed two types of Chinese calligraphy generative models and one Chinese calligraphy style fusion model.The main innovative work of this thesis is as follows:1.A generative model of Chinese calligraphy characters based on paired data.Owing to the traditional Chinese calligraphy character generative models always generate bad results and low efficiency.A Chinese calligraphy generative model based on paired data is proposed in this thesis.The model is based on a Generative Adversarial Network framework,added Auto-Encoder,optimized the structure of neural networks,introduced the Batch Normalization layer,Adam optimizer and related loss functions.After training on datasets by paired Chinese calligraphy,for the calligraphy style or detail information,the calligraphy samples generated by the model in batches are highly consistent with the authentic samples,which proves the authenticity and reliability of the model.2.A generative model of Chinese calligraphy characters based on unpaired data.A generative model of Chinese calligraphy character based on paired data requires paired dataset as input data,which restriction is too strict.A Chinese calligraphy characters generative model based on unpaired data is proposed in this thesis.The model is based on the Generative Adversarial Network style translation framework,added Distribution Transform module,added fusion structure,optimized the structure of neural networks,added sampling module,introduced a loss function of WGAN and related optimizer.After training on datasets by unpaired Chinese calligraphy,for the calligraphy style or detail information,the calligraphy samples generated by the model in batches are highly consistent with the authentic samples.In addition,the model is compared with various generative models through objective evaluation named SSIM,which proves the authenticity,effectiveness and reliability of the model.3.Chinese calligraphy characters style fusion model.In order to help calligraphy lovers and researchers better understand different Chinese calligraphy styles of different calligraphers,and generate calligraphy samples with calligraphy styles in common.A Chinese calligraphy character style fusion model based on Generative Adversarial Networks is proposed in this thesis.The model consists of a Generative Adversarial Network,multiple loss functions and neural network modules,introduced style interpolation variables,loss function of WGAN and related optimizer.After training on datasets by paired Chinese calligraphy,the samples generated in batches of the model are effectively compatible with different calligraphy styles,retaining the details of them,which proves the authenticity and reliability of the model.