The style translation of Chinese calligraphy image is a challenging problem: changing the style and layout of strokes while retaining the content attributes like overall structure and combination of radicals. The existing calligraphy character generation methods, such as brush modeling, skeleton rendering, strokes extracting and assembling, etc., are ineffective and difficult to apply to multi-chirography styles and numerous data. The recent neural style transfer methods for general image-to-image style translation tasks can not apply to our problem due to the different meanings of the "style". The GAN-based methods demonstrating good results require image-pairs for training, which is hard to collect in our task. Therefore, in this paper we propose a novel GAN-based model, called CalliGAN, for the multi-chirography Chinese calligraphy image translation. In CalliGAN, We present a joint optimization method which only requires unpaired multiple chirography sets for training. In our experiment, we build a chirography style dataset called Chiro-4 and then compare our method with various general translation methods. The experiment results demonstrate the validity of our method on calligraphy style translation task.
展开▼
