A survey of style transfer based on generative adversarial network

Abstract

Generative adversarial network (GAN) has become a hot research topic in the field of image processing. As an unsupervised training model, GAN has been widely used in the field of computer vision, especially in image style transfer. The purpose of the GAN is to make the generator generate a false image, and the discriminator cannot tell whether the input image is the real image or the generated image. Compared with traditional network models, GAN model has these advantages in image style transfer: GAN is composed of two different networks, and the loss function is automatically learned by playing games with each other. GAN belongs to unsupervised training and does not need to annotate the data set, which saves a lot of work. In this paper, improved GAN models related to image style migration are summarized. Firstly, the principle and method of image style transfer based on convolutional neural network are introduced. Secondly, the status, principle and prospect of GAN are introduced, and the causes of gradient disappearance and mode collapse of GAN are analyzed in detail. On this basis, the principles, advantages and disadvantages of CGAN, DCGAN, CycleGAN and StarGAN V2 network models are introduced. Finally, it summarizes the current problems and future research directions of style transfer based on GAN.