LDP: A Lightweight Denoising Plugin Enhancing Generalization in Single-Image Super-Resolution

Junyi He, Xiaobin Zhu, Hongyang Zhou, Jingyan Qin, Xu-Cheng Yin

Keywords: low-level vision, Single Image Super-Resolution, degradation model

TL;DR: we propose LDP, a lightweight denoising autoencoder plug-in. It improves SR model generalization via LR-prediction–based cyclic regularization.

Abstract:

Current single-image super-resolution (SISR) models struggle to generalize to real-world degradations. To address this challenge, we propose LDP, an innovative lightweight denoising autoencoder~(DAE) plug-in. It improves the generalization ability of SR models via low-resolution (LR) images prediction-based cyclic regularization. LDP models the SISR degradation process within the DAE framework. It leverages a property of diffusion models, where after noise is added, high-resolution (HR) images and LR features become aligned, so that denoising noisy HR features is equivalent to denoising noisy LR features. During the corruption process, noise is added independently to each HR patch. During the denoising process, a convolutional denoiser uses learned filters to approximate blur kernels.
In addition, LR degradation is used to distinguish different LR from the same HR. LDP can be applied to SR models in two modes: as a training loss to improve reconstruction quality, or as an inference post-processing step to correct artifacts. Extensive experiments demonstrate that LDP substantially improves the generalization of existing SR models to unseen degradations.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 8948