Image Style Transfer based on Generative Adversarial Network

The traditional style transfer based on GAN model is limited by paired images. CycleGAN solves this problem effectively, but its structure is complex and training time-consuming. This paper proposed a style transfer model based on generative adversarial network, which abandons the redundant structure of two GAN models trained by CycleGAN, and trains only one generator and one discriminator without pairing image samples. And the semantic content between the input image and the generated image is constrained by the VGG network feature map. In order to accelerate the convergence of the model, a pretraining stage is introduced. The experimental results show that the proposed model is effective than CycleGAN, and outperforms state-of-the-art methods.