Diverse adversarial network for image super-resolution

Abstract

Recently, there is a fast growth in Generative adversarial network and many works have appeared focusing not only images but also videos. Despite of remarkable success of GAN in image super resolution, it suffers from the major problem of poor perceptual quality. While employing a GAN for super resolution, it tends to generate over-smoothed images that lacks high frequency textures and do not look natural. We propose an intuitive generalization to Generative Adversarial Network and its conditional variations to address the problem of image super-resolution and improves the test quality of images. DGAN is a diverse GAN architecture incorporating multiple generators and a single discriminator. The main intuition is to employ multiple generators, instead of using a single one as in the original GAN. To enforce that multiple generators produce diverse samples, the discriminator trains a loss function to distinguish between real and fake samples by designed margins, and multiple generators alternately produce realistic samples by minimizing their losses. In fact, this paper addresses 2 main challenges; recovering realistic texture low resolution images and speed up the training process. We perform extensive experiments and compare the proposed model with other variants of GAN to demonstrate the efficiency and stability of the proposed model in both quantitative and qualitative benchmarks.