In the image, a person is standing in front of a poster displayed on a wall. The poster appears to be showcasing the enhancement of the quality of images, possibly along with text describing the process or results. The person, who seems to be looking at the poster, might be interested in learning more about the image enhancement process or comparing the before and after results.
Text transcribed from the image:
DIV2K dataset
Flickr2K dataset
北京航空航天大學
BEIHANG UNIVERSITY
Enhancing Quality of Compressed Images by Mitigating
Enhancement Bias Towards Compression Domain
Qunliang Xing, Mai Xu, Shengxi Li, Xin Deng, Meisong Zheng, Huaida Liu, Ying Chen
Enh. domain (ESRGAN)
Motivation
5.93
Biased
(-45.33%)
-8.84
9.73
Comp. domain
Raw domain
Raw
Enhanced
Compressed
Enh. domain (Real-ESRGAN)
IR
6.13
Biased
(-78.36%)
10.58
Underlying causal gra
enhanced image I com
(PGM), comprising a the
Correlation between IR
9.73
Comp. domain
Raw domain
Realism score: 0.85
0.28 (-0.57)
0.87 (+0.02)
Enh. domain (ESRGAN)
6.35
Biased
(-28.99%)
8.75
11.18
Comp. domain
Raw domain
Enh. domain (Real-ESRGAN)
7.23
Biased -13.24
(-98.43%)
11.18
Comp. domain
Raw domain
FID scores between enhanced, compressed, and raw images: For
existing methods, enhanced images are more aligned with compressed
images than with raw images.
0.13 (-0.53)
0.85 (+0.19)
Realism score: 0.66
Discriminator scores of SR methods (e.g., Real-ESRGAN): Despite
the presence of compression artifacts, existing methods perceive the
compression image as more realistic than the raw image.
enhanced images that
P(IR, IE IC)=p(IE
Bayesian D-separatic
independent given Ic.
➤ Method #1: We dep
the compression im-
Specifically, we prow
of compressed and
Data.
F
C
Results
Enh. domain
Ours with [44]
•
Enh. domain [44]
6.13
Biased
10.58
9.36 Debiased 8.77
9.73
9.73
Compressed
Real-ESRGAN [44]
Ours with [44]
Raw
Comp. domain
Raw domain
Comp. domain
Raw domain
DIV2K
Flickr2K
Providing the discrimin
and enhanced/raw ima
> Method #2: We inc
that promotes a larg
compression domai