Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Correction Errors via Frequency-Self Attention for Blind Image Super-Resolution

Haochen Sun, Yan Yuan, Lijuan Su, Haotian Shao

TL;DR

A novel blind SR approach that focuses on Learning Correction Errors (LCE) that employs a lightweight Corrector to obtain a corrected low-resolution (CLR) image and proposes a new Frequency-Self Attention block (FSAB) that enhances the global information utilization ability of Transformer.

Abstract

Previous approaches for blind image super-resolution (SR) have relied on degradation estimation to restore high-resolution (HR) images from their low-resolution (LR) counterparts. However, accurate degradation estimation poses significant challenges. The SR model's incompatibility with degradation estimation methods, particularly the Correction Filter, may significantly impair performance as a result of correction errors. In this paper, we introduce a novel blind SR approach that focuses on Learning Correction Errors (LCE). Our method employs a lightweight Corrector to obtain a corrected low-resolution (CLR) image. Subsequently, within an SR network, we jointly optimize SR performance by utilizing both the original LR image and the frequency learning of the CLR image. Additionally, we propose a new Frequency-Self Attention block (FSAB) that enhances the global information utilization ability of Transformer. This block integrates both self-attention and frequency spatial attention mechanisms. Extensive ablation and comparison experiments conducted across various settings demonstrate the superiority of our method in terms of visual quality and accuracy. Our approach effectively addresses the challenges associated with degradation estimation and correction errors, paving the way for more accurate blind image SR.

Learning Correction Errors via Frequency-Self Attention for Blind Image Super-Resolution

TL;DR

A novel blind SR approach that focuses on Learning Correction Errors (LCE) that employs a lightweight Corrector to obtain a corrected low-resolution (CLR) image and proposes a new Frequency-Self Attention block (FSAB) that enhances the global information utilization ability of Transformer.

Abstract

Previous approaches for blind image super-resolution (SR) have relied on degradation estimation to restore high-resolution (HR) images from their low-resolution (LR) counterparts. However, accurate degradation estimation poses significant challenges. The SR model's incompatibility with degradation estimation methods, particularly the Correction Filter, may significantly impair performance as a result of correction errors. In this paper, we introduce a novel blind SR approach that focuses on Learning Correction Errors (LCE). Our method employs a lightweight Corrector to obtain a corrected low-resolution (CLR) image. Subsequently, within an SR network, we jointly optimize SR performance by utilizing both the original LR image and the frequency learning of the CLR image. Additionally, we propose a new Frequency-Self Attention block (FSAB) that enhances the global information utilization ability of Transformer. This block integrates both self-attention and frequency spatial attention mechanisms. Extensive ablation and comparison experiments conducted across various settings demonstrate the superiority of our method in terms of visual quality and accuracy. Our approach effectively addresses the challenges associated with degradation estimation and correction errors, paving the way for more accurate blind image SR.
Paper Structure (25 sections, 15 equations, 7 figures, 8 tables)

This paper contains 25 sections, 15 equations, 7 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Deep Networks for Non-blind SR
  4. Deep Networks for Blind SR
  5. Frequency Learning
  6. Methods
  7. Motivation
  8. Formulation
  9. Architecture
  10. Corrector and CLR Feature Extractor.
  11. Super Resolver.
  12. Frequency Attention Block (FAB).
  13. Experiments and Discussion
  14. Experiment Setup and Datasets
  15. Ablation Study
  16. ...and 10 more sections

Figures (7)

  • Figure 1: Difference between previous methods using Correction Filter and our method. (a) Previous methods are incompatible between models. (b) Our method achieves compatibility by learning correction errors.
  • Figure 2: Analysis of correction errors.
  • Figure 3: The overall architecture of the method LCE, consisting of a lightweight Corrector, a Feature Extractor for CLR images, and a Super Resolver.
  • Figure 4: Comparison of features extracted by different methods.
  • Figure 5: Visual results of img_005 and img_004 in Urban100 Urban100, with scale factor 4 and isotropic kernel width 2.4.
  • ...and 2 more figures