VarSR: Variational Super-Resolution Network for Very Low Resolution Images

As is well known, single image super-resolution (SR) is an ill-posed problem where multiple high resolution (HR) images can be matched to one low resolution (LR) image due to the difference in their representation capabilities. Such many-to-one nature is particularly magnified when super-resolving with large upscaling factors from very low dimensional domains such as 8 × 8 resolution where detailed information of HR is hardly discovered. Most existing methods are optimized for deterministic generation of SR images under pre-defined objectives such as pixel-level reconstruction and thus limited to the one-to-one correspondence between LR and SR images against the nature. In this paper, we propose VarSR, Variational Super Resolution Network, that matches latent distributions of LR and HR images to recover the missing details. Specifically, we draw samples from the learned common latent distribution of LR and HR to generate diverse SR images as the many-to-one relationship. Experimental results validate that our method can produce more accurate and perceptually plausible SR images from very low resolutions compared to the deterministic techniques.