Uncertainty Aware Implicit Image Function for Arbitrary-Scale Super-Resolution

Abstract

The most integral step in computer vision and image processing tasks is the representation of images. Recently, continuous image parameterization using Implicit Neural Representations (INR) has shown a great advantage over discrete representations due to their spatial invariance. This property immediately finds application in the context of single-image super-resolution (SISR) at an arbitrary scale. However, most super-resolution models, including the INR-based Local Implicit Image Function (LIIF), produce only a single output, failing to address the ill-posedness of SISR. Moreover, these models tend to optimize a mean-squared-error (MSE) based loss function which causes blurring and structural distortion in regions exhibiting a high degree of variance (details). Our work proposes a novel uncertainty-aware self-supervised methodology (U-LIIF) that extends on LIIF, to reduce the blurriness and deals with the ill-posedness of SISR. Our U-LIIF does not require any re-training and yields diversified high-resolution images by leveraging model uncertainty. The efficacy of the proposed method is validated by substantial experiments on various benchmark datasets.