Unsupervised image style transfer based on U-GAT-IT

Abstract

Thanks to the rapid development of Generative Adversarial Networks (GAN) in recent years, great progress has been made in image translation using GAN. Image-to-image transformation is one of the important applications in the field of computer vision, and its scope includes image inpainting, image colorization, super-resolution and image style transfer. In recent years, there are many classic research on GAN-based image transformation, such as CycleGAN, UNIT, AGGAN, etc. This paper focuses on researching and testing the U-GAT-IT unsupervised image style transfer method. The authors introduced a new attention module and a new learnable normalization function (AdaLIN), which enables flexible control of the amount of change in shape and texture during image conversion. This paper uses a new dataset for testing and verification, trying to achieve bidirectional conversion between real face photos and sketches, and qualitatively and quantitatively analyzes the method by calculating PSNR and SSIM.