상호증류를 활용한 SRGAN의 성능 개선

Abstract

This paper proposes a method to improve the performance of SRGAN, a generative adversarial network (GAN)-based super-resolution neural network model, by applying mutual distillation. The proposed method adds a discriminator for mutual distillation to SRGAN, allowing the generator of each SRGAN to indirectly share the knowledge of the other SRGANs’ generators through adversarial learning with the discriminators added to the other SRGANs. Experiments using two SRGANs were conducted to analyze the performance of the proposed method, and the experiments confirmed that SRGANs produced super-resolution images with improved qualitative and quantitative image quality by applying mutual distillation. In addition, it was confirmed that the performance varies depending on the amount of knowledge transferred through mutual distillation.