3rd IMA Conference on Inverse Problems from Theory to Application
Date:
Image super-resolution is an ill-posed inverse problem in the sense that diverse high- resolution candidates are plausible solutions for each single low-resolution image. In this work we propose to make use of deep hierarchical variational autoencoders (VAE) to produce diverse super resolution. Hierarchical VAEs have shown impressive results for the task of high-resolution image synthesis and provide a strong prior image model via a self- organized hierarchy of latent variables. We find that these structured latent variables are related with the image information at different scales. Based on this observation, we show that pretrainded hierarchical VAEs can be repurposed to perform diverse super-resolution.