Chinese Typography Transfer Model Based on Generative Adversarial Network

Digital calligraphy has grown into a distinctive cross-research field after decades of research and development. With the help of deep learning technology, the fonts that have not been written can be generated in the calligrapher's style based on the works retained. This can provide more samples for appreciation and promote the art of calligraphy. In addition, many inscriptions, rubbings and inks are badly damaged. This method can restore the missing characters in them, which is conducive to the preservation of cultural relics. In this paper, we propose an unsupervised Chinese typography transfer model based on generative adversarial network. The main contributions are as following: the generator network is optimized by introducing DenseNet module between the encoding and decoding network, which can improve the ability of font feature extraction and transformation; at the end of discriminator network, we add two convolutional layer branches to improve the quality of the generated image. Without preparing paired data, this model can realize style transfer between any two fonts by one training process. Experiment shows that this model works well in printed font transferring and calligraphy transferring.