Reference-based Image Super-Resolution by Dual-Variational AutoEncoder

Due to severe information loss of low-resolution images, the development of single-image super-resolution methods is limited. Recently, the reference-based image super-resolution methods, which super-resolve the low-resolution inputs with the guidance of high-resolution reference images are emerging. In this paper, we design a Dual-Variational AutoEncoder (DVAE) for reference-based image super-resolution task, which can learn the high-frequency information and latent distribution of the high-resolution reference images as priors to improve the restoration quality of image super-resolution. Moreover, a hierarchical variational autoencoder strategy is exploited to further study latent space. Complementary to a quantitative evaluation, we demonstrate the effectiveness of the proposed approach.