Abstract: Aiming at  the problem of poor effect of the style transfer of existing image ink painting, we proposed a new data enhanced cycle generative adversarial network (GAN) for the style transfer of ink painting of unpaired natural landscape photos. Firstly, the binary synthesizer and discriminator structure was designed to effectively improve the mapping constraints of one-way GAN models. Secondly, we used multiple loss functions to optimize the model, introduced total variational loss and identity mapping loss, and designed a new cyclic consistency loss function combined with multi-scale structural similarity to better capture the characteristics of traditional ink painting. Finally, data enhancement techniques were used to increase the amount and variety of real and generated data to improve generator performance. The comparative experimental results show that this method can effectively transfer natural landscape images to traditional ink painting style images.

Key words:  , CycleGAN, style transfer of image ink painting, loss function, data enhancement