Single Image Super-Resolution Based on Wasserstein GANs

In this paper, a novel single image super-resolution method unifying deep residual network and Wasserstein generative adversarial nets is proposed aiming at generating a photo-realistic image with finer texture details. Specifically, we construct a framework consisting of a generator that recovers a high-resolution image with an input low-resolution image and a discriminator that tries to distinguish the recovered image from the real image. The competing of the generator and discriminator drives the generator to produce images that are highly similar to real images. Meanwhile, we define a new loss function by taking both the pixel-wise error and the abstract feature difference into account to force the generator to converge towards a better solution approximating the distribution of real images. Experimental results indicate the effectiveness and robustness of the proposed method for single image super-resolution.