Unsupervised image style transformation of generative adversarial networks based on cyclic consistency

Image style transformation is a fascinating and challenging problem in the field of computer vision. The performance of existing image style transfer algorithms is suppressed to some extent due to the lack of sufficient paired datasets. In order to reduce dependence on paired datasets and reduce model structural risk, this paper proposes an unsupervised image style transformation algorithm based on cycle consistency generative adversarial networks. To solve the instability problem of model training, a Spectrum Normalization method is constructed in the discriminator of cycle generative adversarial networks. To improve the performance of the model generator, a Res2Net residual module using the group convolution method is constructed in the network. The network layer can increase the receptive field range of the network layer without increasing the calculation load, and better realize the extraction of multi-scale features. Ablation experiments and comparative experiments are conducted on four datasets with different artistic styles, the experimental results show that the method in this paper has achieved good results in PSNR, SSIM, FID, and KID evaluation metrics, which proves the robustness of the algorithm.