A review of cutting-edge aspects of image quality evaluation (2022)

Image Quality Assessment (IQA) mainly includes two parts. One is subjective quality assessment, that is, people's subjective perception evaluation of pictures. It is generally described quantitatively using MOS or DMOS indicators. The way to obtain it is through subjective psychology. Experiment, establish an image quality evaluation database. According to different distortion types, experimental methods, etc., the types of IQA database are very rich; the other is objective quality evaluation, the evaluation subject is a machine, usually a classifier or regressor is designed to evaluate the image , an image quality evaluation algorithm based on deep learning that has been widely studied at present.

This article mainly summarizes the following two aspects:

• Deep-learning based IQA model
• IQA database

paper

The objective image quality assessment algorithm is mainly divided into two parts. One is NR-IQA, or No-reference IQA, which is a no-reference image quality assessment algorithm; the other is FR-IQA, which is Full-Reference IQA, which is a full-reference image quality assessment. The following is a summary of these two parts

NR-IQA

Below I list the currently common NR-IQA algorithms in chronological order, as shown in the following table:

Algorithm name	Publication time/team	code	type	Google Scholar	Main idea
NIKE	SPL2012	—	Tradition	–
BRISQUE	TIP2012	—	Tradition	–
ILNIQE	TIP2015	—	Tradition	–
HOSA	TIP2016	—	Tradition	–
NRQM (Ma)	CVIU2017	Matlab	Tradition	–
CNNIKA	CVPR2014	—	DL	–
dipIQ	TIP2017	—	DL	–
MIND	TIP2017	—	DL	–
RankIQA	ICCV2017	—	DL	–
WHAT	TIP2018	—	DL	–
WaDIQaM (the deep)	TIP2018	—	DL	–
BPSQM	CVPR2018	—	DL	–
HIQA	CVPR2018	—	DL	–
PI	2018 PIRM Challenge	—	DL	–
PQR	TIP2019	—	DL	–
SFA	TMM2019	—	DL	–
DBCNN	TCSVT2020	—	DL	–
GIQA	ECCV2020	—	DL	–
Meta-IQA	CVPR2020	—	DL	–
HyperIQA	CVPR2020	—	DL	–
UNIQUE	TIP2021	—	DL	–
CKDN	ICCV2021	—	DL	–
MUSIC	ICCV2021	—	DL	–
KonIQ++	BMVC2021	—	DL	–

FR-IQA

Algorithm name	issuing time	type	Google Scholar	code	Main idea
CVRKD	AAAI2022	NAR	A	Official
IQT	CVPRW2021	FR	C	PyTorch	Transformer
A-DISTS	ACMM2021	FR	A	Official
DISTS	TPAMI2021	FR	T	Official
JND-SalCAR	TCSVT2020	FR	T
QADS	TIP2019	FR	T	Project
LPIPS	CVPR2018	FR	C	Project
PieAPP	CVPR2018	FR	C	Project
WaDIQaM	TIP2018	NR/FR	T	Official
FSIM	TIP2011	FR	T	Project
VIF/IFC	TIP2006	FR	OF	Project
MS-SSIM	pdf	FR		Project
YES	TIP2004	FR	T	Project
PSNR	–	FR

other

IQA database

IQA databases are mainly divided into two categories, the mainstream IQA database and other eye movement and aesthetics-related databases.

• Mainstream databases mainly include: single distortion database, mixed distortion database, different photographic equipment database and real distortion database
• Other databases mainly include: eye movement database and aesthetics database

Mainstream IQA database

The following is the mainstream IQA database. I will explain it in more detail later. For now, I will summarize the main information.

Name database	issuing time	type	Google Scholar	Link	illustrate	Remark
PaQ-2-PiQ	CVPR2020	real distortion	CV	Official github	40k, 120k patches	4M
SPAQ	CVPR2020	–	CVPR2020	Offical github	11k (smartphone)
KonIQ-10k	TIP2020	–	TIP2020	Project	10k from YFCC100M	1.2M
CLIVE	TIP2016	–	TIP2016	Project	1200	350k
PIPAL	ECCV2020	–	ECCV2020	Project	250	1.13M
KADIS-700k	arXiv	–	arXiv	Project	140k pristine / 700k distorted	30 ratings (DCRs) per image.
KADID-10k	QoMEX2019	–	QoMEX2019	Project	81	10k distortions
Waterloo-Exp	TIP2017	–	TIP2017	Project	4744	94k distortions
MDID	PR2017	–	PR2017	—	20	1600 distortions
TIME2013	SP2015	–	SP2015	Project	25	3000 distortions
LIVEMD	ACSSC2012	–	ACSSC2012	Project	15 pristine images	two successive distortions
CSIQ	JEI2010	–	sada	—	30	866 distortions
TID2008	2008	–	2009	Project	25	1700 distortions
A57	2007	–	2005	—	10	185 distortions
LIVE	TIP2006	–	TIP2006	Project	29 images, 780 synthetic distortions
IVC	2005	–	2005	—	10	185 distortions

美学和眼动等数据库