GN-CycleGAN: artistic style transfer with gradient normalization and cycle generative adversarial networks
Qinghong Shao, Xin Chen, Ruocheng Su, Yongheng Zhao, Yaqi Ba · 2023
Image style transfer is a technique for generating images with the help of computers. It can transfer the style features of image B to image A to generate image A with style B, which can not only meet the individual needs of users, but also meet the needs of users. The problem of insufficient image data for a certain style. In recent years, a series of models such as CycleGAN have achieved great success in the field of domain style transfer, but their training is unstable, and the problems of blurred, unstable, and inefficient generated images that are prone to mode collapse have not been solved, this paper proposes the GN-CycleGAN model for this problem. In this paper, a GN-CycleGAN model is proposed to solve this problem. By replacing the instance normalization in the discriminator with the gradient normalization, the discriminator space can be smoother without affecting the performance of the discriminator, so as to overcome the training instability of CycleGAN caused by steep gradient space. The experimental results show that, compared with CycleGAN, the loss function of GN-CycleGAN model can converge quickly under the same training time,, and it performs better on the monet2photo dataset. The generated images are in PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity) It is better than CycleGAN in two indicators.