Dual Discriminator Generative Adversarial Network for Single Image Super-Resolution

Abstract

Single image super-resolution(SISR) is to reconstruct a high resolution(HR) image from a single low resolution(LR) image. In this paper, with generative adversarial networks(GAN) model as the basic component, we build a dual discriminator super-resolution reconstruction network(DDSRRN) to improve the quality of image super-resolution reconstruction. By adding another discriminator based on GAN, we combine the Kullback-Leibler(KL) with reverse KL divergence to make a unified objective function to train the two discriminators, and by using the complementary statistical characteristics of these two divergences, the prediction density is effectively dispersed in multi-mode, which can avoid collapse of the network model during the reconstruction process and improve the stability of model training. We build the content loss function using the Charbonnier loss and use the intermediate features information of the two discriminators to build the perceptual loss function and style loss function. The experimental results show that the proposed method has sharp edges and rich details in subjective vision, and obtains better subjective visual evaluation and objective quantitative evaluation.