Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

JPEG Artifact Correction using Denoising Diffusion Restoration Models

Bahjat Kawar, Jiaming Song, Stefano Ermon, Michael Elad

TL;DR

The paper tackles JPEG artifact correction as a non-linear inverse problem and extends Denoising Diffusion Restoration Models by substituting the linear operator with JPEG encoding/decoding, effectively creating a problem-agnostic restoration framework. It demonstrates that the generalized DDRM update can leverage unconditional diffusion priors for JPEG restoration, achieving competitive performance against a specialized GAN baseline and better generalization at low quality factors. The approach also generalizes to non-linear tasks such as image dequantization without retraining, highlighting broad applicability. Overall, the work shows that diffusion priors can efficiently handle a wider class of non-linear inverse problems in practical image restoration settings.

Abstract

Diffusion models can be used as learned priors for solving various inverse problems. However, most existing approaches are restricted to linear inverse problems, limiting their applicability to more general cases. In this paper, we build upon Denoising Diffusion Restoration Models (DDRM) and propose a method for solving some non-linear inverse problems. We leverage the pseudo-inverse operator used in DDRM and generalize this concept for other measurement operators, which allows us to use pre-trained unconditional diffusion models for applications such as JPEG artifact correction. We empirically demonstrate the effectiveness of our approach across various quality factors, attaining performance levels that are on par with state-of-the-art methods trained specifically for the JPEG restoration task.

JPEG Artifact Correction using Denoising Diffusion Restoration Models

TL;DR

The paper tackles JPEG artifact correction as a non-linear inverse problem and extends Denoising Diffusion Restoration Models by substituting the linear operator with JPEG encoding/decoding, effectively creating a problem-agnostic restoration framework. It demonstrates that the generalized DDRM update can leverage unconditional diffusion priors for JPEG restoration, achieving competitive performance against a specialized GAN baseline and better generalization at low quality factors. The approach also generalizes to non-linear tasks such as image dequantization without retraining, highlighting broad applicability. Overall, the work shows that diffusion priors can efficiently handle a wider class of non-linear inverse problems in practical image restoration settings.

Abstract

Diffusion models can be used as learned priors for solving various inverse problems. However, most existing approaches are restricted to linear inverse problems, limiting their applicability to more general cases. In this paper, we build upon Denoising Diffusion Restoration Models (DDRM) and propose a method for solving some non-linear inverse problems. We leverage the pseudo-inverse operator used in DDRM and generalize this concept for other measurement operators, which allows us to use pre-trained unconditional diffusion models for applications such as JPEG artifact correction. We empirically demonstrate the effectiveness of our approach across various quality factors, attaining performance levels that are on par with state-of-the-art methods trained specifically for the JPEG restoration task.
Paper Structure (9 sections, 7 equations, 4 figures, 1 table)

This paper contains 9 sections, 7 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Linear Inverse Problems.
  4. Denoising Diffusion Restoration Models (DDRM).
  5. JPEG.
  6. JPEG Artifact Correction with DDRM
  7. Experimental Results
  8. Conclusion
  9. Additional Visual Results

Figures (4)

  • Figure 1: Pairs of JPEG images and restorations using our method. Best viewed zoomed in.
  • Figure 2: Pairs of quantized ($9$ bits per color) and restored images using our method. Best viewed zoomed in.
  • Figure 3: Rate-distortion curves for standard JPEG (blue) and our method (green).
  • Figure 4: Triplets of original (ground-truth), JPEG compressed, and restored images. Across different quality factors (QF), our method successfully corrects artifacts of JPEG compression. Images are accompanied by a zoomed-in area in the bottom right corner to highlight specific artifact removals.