Asymmetric CycleGAN for Unpaired Image-to-Image Translation Based on Dual Attention Module

Numerous studies have been conducted on image style transfer. With the advancement of artificial intelligence, the application of convolutional neural networks and Generative Adversarial Networks (GAN) has garnered considerable attention in this field. Traditional CycleGAN is limited in extracting the features of different images well and effectively for asymmetric translation due to its own completely symmetric structure. In this paper, we propose an asymmetric CycleGAN model for spring and winter image style transfer. Our method modifies one of the generators by adding dual attention module, which can capture global features through spatial and channel dimensions so as to obtain rich information of images more deeply. The model is experimentally validated to be more capable of color recognition than the original CycleGAN.