1 of 10

Yulun Zhang1, Kai Zhang1, Zheng Chen2, Yawei Li1, Radu Timofte1,3

​

1 ETH Zürich, 2 Shanghai Jiao Tong University, 3 University of Würzburg

NTIRE 2023 Challenge on Image Super-Resolution (×4):

Methods and Results

2 of 10

Introduction

  • Challenge
    • The task of image super-resolution (SR) is to generate a high-resolution (HR) output from a corresponding low-resolution (LR) input by leveraging prior information from paired LR-HR images.

​

  • Motivation
    • To provide an overview of the new trends and advances in super-resolution areas.
    • To obtain a network design/solution capable of producing high-quality results with the best performance (e.g., PSNR).

1

3 of 10

Overview

  • Phase
    • Development and validation phase.
    • Testing phase.
  • Dataset
    • The challenge does not specify training datasets (except the test data: DIV2K testing dataset).
    • Some datasets are recommended, such as DIV2K, Flickr2K, and LSDIR.
  • Track
    • The ranking of participating teams is determined based on the Peak Signal-to-Noise Ratio (PSNR) value on the DIV2K testing dataset.
  • Evaluation
    • The standard metrics: PSNR and SSIM are used in the challenge.

​

​

​

2

4 of 10

Overview

  • Participant
    • 33 registered participants, out of which 15 teams submitted valid entries.

​

  • Teams

​

​

​

​

3

Index

Team

Leader

1

ZZPM

Junpei Zhang

2

Graphene

Huaibo Huang

3

IPLAB

Yajun Qiu

4

SRC-B

Dafeng Zhang

5

LDCC

TaeHyung Kim

Index

Team

Leader

6

NTU607-SR

Zhi-Kai Huang

7

Swin2SR

Ui-Jin Choi

8

IKLAB-TUK

Sunder Ali Khowaja

9

GarasSjtu

Garas Gendy

10

LVGroup HFUT

Zhao Zhang

Index

Team

Leader

11

AhRightRightRight

Jiayu Wei

12

helloooo

Xinyi Wang

13

chaobaer

Jun Cao

14

Alpha

Jianbin Zheng

15

SVNIT NTNU

Anjali Sarvaiya

5 of 10

Overview

  • Results
    • There are 15 teams submitted valid.
    • PSNR/SSIM results are measured on the DIV2K testing dataset.
    • The ZZPM team achieves the highest overall ranking in this Challenge.
    • The average PSNR�value of the top five teams is above 31 dB.

​

3

6 of 10

Methods and Teams

  • Winner: ZZPM

4

  • Multiple data augmentation methods are used to train five state-of-the-art methods.
  • The final image is obtained by fusing the outputs of all methods in Post-processing part.

7 of 10

Methods and Teams

5

  • 2nd Place: Graphene
  • Transformer with Cross-Scale Attention (CSA) and Wavelet Hallucination (WH).
  • CSA is enhanced with depth-wise convolution.
  • WH apply the Haar wavelet.

8 of 10

Methods and Teams

6

  • The Attention Retractable Frequency Transformer (ARFT).
  • Two attention strategies: D-MSA and S-MSA.
  • The FEB network is intended to capture the long-range context in the frequency domain.
  • Progressive Training Strategy is used.
  • 3rd Place: IPLAB

9 of 10

Conclusion

​

​

​

​

​

​

7

  • The challenge provides several new trends and advances in super-resolution areas at the method and data aspects.

​

  • Method: Adopting and modifying the Transformer architecture is the mainstream technology. Global information is crucial for SR.

​

  • Data: More training data and better image augmentation methods could effectively improve performance.�

10 of 10

Thanks

The code and models are available at: https://github.com/zhengchen1999/NTIRE2023_ImageSR_x4