End-to-End Image Super-Resolution via Generative Adversarial Network

Abstract

Image super-resolution is to extract information from single or more low- resolution images and use this information to get the corresponding high-resolution image. The ability to capture high texture details of low-resolution images is one of the impressive advantages of generative adversarial networks (GANs). This paper mainly studies the image reconstruction method based on a single low-resolution image, and builds an end-to-end image Super-Resolution method via generative adversarial networks to improve the image resolution. Convolution neural networks use the mean square error (MSE) as loss function. For such loss function, structural similarity (SSIM) and peak signal to noise ratio (PSNR) can be obtained, but high texture details of original images are often lost. The generative adversarial network mainly contains two parts: generator network and discriminator network. Adversarial loss uses a discriminator network to push our solution to a natural image manifold, which is trained to distinguish super-resolution images from photo-realistic images. Additionally, content loss of our network was indicated by perceptual similarity rather than similarity of pixel-wised space.