This image showcases a research poster presented at the 2024 CVPR (Conference on Computer Vision and Pattern Recognition) held in Seattle, WA. The research, conducted by Yuchuan Tian, Hanting Chen, Chao Xu, and Yunhe Wang from the National Key Lab of General AI at Peking University and Huawei Noah’s Ark Lab, focuses on "Breaking Rigidity in Super-Resolution" using Image Processing GNNs (Graph Neural Networks).

The poster is titled "The Lopsided Nature of SR" and addresses the problem of Super-Resolution (SR) in image restoration, where the task is to convert a low-resolution (LR) image into a high-resolution (HR) image. It highlights how SR problems are lopsided, with high-frequency areas requiring more restoration efforts compared to low-frequency areas.

Key components of the research include:
- Calculation of the difference between HR and LR.
- Existing SR solutions' rigidity, which includes spatial boundaries and the neglect of SR’s lopsided nature.
- Leveraging graph flexibility to improve SR by allowing nodes to flexibly adapt during the processing.
- The proposed architecture details and experimental results demonstrating the efficacy of their method.

Various diagrams and experimental results support their findings, and the poster is marked as an "Award Candidate," indicating its notable contribution to the field. The vibrant and comprehensive design highlights the complexity and innovation of the presented research. In the background, the venue is equipped with modern lighting, providing a conducive atmosphere for the conference. Poster number 191 is visible on the display board.
Text transcribed from the image:
(xdg
WT
CVPR
SEATTLE, WA JUNE 17-21, 2024
Award
Candidate
Image Processing GNN
Breaking Rigidity in Super-Resolution
Yuchuan Tian, Hanting Chen, Chao Xu, Yunhe Wang*
The Lopsided Nature of SR
Super-Resolution (SR): an image restoration task
Receive a low-resolution (LR) image, and
Return a high resolution (HR) image
•
We calculate the difference of HR & LR:
北京大学
191
PEKING
National Key Lab of General AI, SIST, Peking University Huawei Noah's Ark Lab
Leveraging Graph Flexibility in SR
We hope to leverage degree flexibility of graphs
• DF:=
F-Fatal indicates High-Freq parts
Image Processing GNN (IPG) Architecture
One block shares a pair of graphs
Loal Graph
--Cabal Graph
Experiments
-
=
SR is a lopsided restoration problem
•
High-Frequency: more restoration efforts
Low-Frequency: minimal modifications
Convolution
Self-Attention
KNN Graph
•
High-Freq parts have higher node degrees
Pixels as nodes is the most flexible:
SR x4
CAT-A
31.08
Set5
09052
Set14
B100
07960 27.90
ART (4)
33.04 09051
0755
•
Patch nodes could cause misalignments
However, pixel space is large
GRL-822
HAT (5
32.90 09009 29.14
30.04
27.57 07510
0.7950 2256 02407 2251
Urban100 Manga109
DEN
IPG (Dan)
29.23 07973 28.00 07317 2797
33.15 09062 2924 07973 27.99
07319
28.13 092 32.53
• It's impractical to search through all pixels
(
•
Existing SR Solutions are Rigid
Existing SR measures suffer rigidity in that:
Spatially bounded within a fixed boundary
Number of connected pixels fixed
The lopsided nature of SR is neglected
Space flexibility is maintained via sampling
.
Local sampling (M) + Global Sampling (R)
Search & gather global & local node info. in
an alternative manner
Ling
A
CHAN
Full Connect
KNY
x
31.09 29.19
Thresholding
3134 2936
2025
Deg-Flex Set Set 4 than 1001
W 2924 2010
Our Flex-Degree
20 strategy performs
Detail (Ours)
35 224
28.13
even better than
Table 9. Comparison of degree-dexible graphs against plain KNN Full-Connect
graphs in IPG. SR4 results are imported
CFAT: Unleashing T
Abhisek R
Problem Definition
Contribution and Formulati
999 99
91998-99
Implications
Computational Cost
DA