Natural and Realistic Single Image Super-Resolution with Explicit Natural Manifold Discrimination

Abstract

Recently, many convolutional neural networksfor single image super-resolution (SISR) have been proposed, which focus on reconstructingthe high-resolutionimages in terms of objective distortion measures. However, the networks trainedwith objective lossfunctions generallyfail to reconstructthe realisticfine textures and details that are essential for better perceptual quality. Recovering the realistic details remains a challengingproblem, and only a few works have been proposed which aim at increasing the perceptual quality by generating enhanced textures. However the generatedfake details often make undesirableartifacts and the overall image looks somewhat unnatural. Therefore, in thispaper,we present a new approachto reconstructingrealistic super-resolved images with high perceptualquality, while maintainingthe naturalnessof the result. In particular we focus on the domain priorpropertiesof SISR problem. Specifically, we define the naturalnesspriorin the lowlevel domain and constrain the output image in the natural manifold, which eventually generates more natural and realistic images. Our results show better naturalness compared to the recent super-resolution algorithms including perception-orientedones.