Research on the Algorithm of Art Style Transfer of Xin'an Painting School

Abstract

Xin‘an Painting School plays an important role in the history of Chinese painting. It takes Huizhou landscape as the creative theme and has a unique artistic style. However, the current art style transfer field does not concern about this very regional characteristics of painting school. Therefore, we propose an improved CycleGAN to realize the transfer of Xin'an painting style. Firstly, DenseNet is introduced to alleviate the gradient vanishing problem and optimize the content and style features transfer between the layers of neural network. Secondly, group normalization is used to reduce the calculation error and keep the network training process stable. Finally, the least square loss is introduced in the adversarial losses, and the identity loss is introduced to obtain the feature of the target image as much as possible, which constrains the arbitrary transformation of the feature of the input image. The experiment shows that the generated pictures have a good artistic style of Xin’ an Painting School.