TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 10, Octobe
r 20
14, pp. 7280
~ 728
6
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.645
5
7280
Re
cei
v
ed
Jun
e
1, 2014; Re
vised July 4,
2014; Accept
ed Jul
y
29, 2
014
Local Standard Deviation Based Imag
e Quality
Metrics for JPEG Compressed
Images
Aksh
a
y
Gor
e
*
1
, H.K.Kan
sal
2
, Sa
v
i
ta
Gupta
3
1
Departme
n
t of Electronics a
n
d
Commu
nicati
on, Cha
n
d
i
gar
h Univ
ersit
y
, P
unj
ab, Indi
a
2
Departme
n
t of Mechan
ical E
ngi
neer
in
g,
Panja
b
Univ
ersit
y
, Chand
ig
arh, Indi
a
3
Departme
n
t of Computer Sci
ence a
nd En
gi
neer
ing, Pa
nja
b
Univ
ersit
y
, In
dia
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: aksha
y
g
o
re
@live.com
A
b
st
r
a
ct
In this p
a
p
e
r, w
e
addr
ess th
e F
u
ll-
Refere
n
c
e (F
R) Image
Quality M
e
tric (IQM) to assess t
h
e
qua
lity of JPE
G
-coded
imag
es an
d w
e
pr
esent a
new
effective an
d
e
fficient IQA mo
de
l, call
ed
Loca
l
Standar
d Dev
i
ation B
a
sed I
m
age Qu
ality (L
SDBIQ). T
he
appro
a
ch is
bas
ed o
n
the co
mparis
on of the
l
o
ca
l
standar
d dev
ia
tion of tw
o ima
ges. T
he pro
p
o
s
ed
metrics
is t
e
sted o
n
four
w
e
ll-k
now
n d
a
tabas
es ava
ila
b
l
e
in the liter
ature
(T
ID2013, T
I
D200
8,
LIVE an
d CSIQ). Experimenta
l
result
s
show
that the prop
osed
metrics
outperfor
m
s
o
t
her
mo
dels
f
o
r the
ass
e
ss
me
nt of
i
m
a
g
e
q
ual
ity a
n
d
hav
e v
e
ry l
o
w
co
mputati
ona
l
complexity.
Ke
y
w
ords
:
JPEG2000, h
u
m
a
n
visu
al sys
tem (HVS), loc
a
l
stand
ard d
e
v
iatio
n
, image
qua
lity assess
me
nt
(IQA)
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
As Hi
gh
-Re
s
olution di
gital
image
s th
at
are
used i
n
i
m
age
proce
s
sing
technolo
g
ies, te
nd
to be of larg
e
size
s a
nd th
ereby
con
s
u
m
ing lar
ge
storag
e spa
c
e,
large tran
sm
issi
on ba
nd
wi
dth
,
and l
ong
tra
n
smi
ssi
on ti
mes. T
h
e
r
efo
r
e, ima
ge
co
mpre
ssion
is re
quired
bef
ore
sto
r
ag
e
and
transmissio
n. JPEG
and
JPEG 200
0 a
r
e two
impo
rta
n
t tech
niqu
es used fo
r im
a
ge
comp
re
ssi
on.
J
PEG
imag
e
c
o
mpr
e
ss
io
n s
t
a
n
d
a
r
d
us
e
D
i
s
c
r
e
te
C
o
sine
Tran
sform (DCT
).
Th
e DCT i
s
a f
a
st
transfo
rm. It is a widely
use
d
and ro
bust meth
o
d
for image compressio
n. It has excell
en
t
comp
actio
n
for hig
h
ly co
rrelated d
a
ta. JPEG2000 i
s
t
he late
st ima
ge comp
re
ssi
on sta
nda
rd t
hat
comp
re
sse
s
and de
co
mpresse
s
the im
age
s u
s
ing
wavelet tran
sfo
r
mation.
Wav
e
let tran
sform-
based im
age
com
p
re
ssion
algo
rithms a
llow im
age
s t
o
be
retain
e
d
witho
u
t mu
ch di
sto
r
tion
or
loss when
compa
r
ed to
JPEG, and
h
ence are re
cogni
zed
as
a
sup
e
rio
r
me
thod. Ho
wev
e
r,
comp
re
ssion
leads to lo
ss of sp
atial and spe
c
tral
feature
s
of the image a
nd may lead
to
errone
ou
s re
sults. Thu
s
, there i
s
a ne
ed for
image
quality asse
ssment (IQA) of compre
ssed
image
s at v
a
riou
s
co
mpression
sta
g
e
s
. Lo
ssy im
age
com
p
ression
techniq
ues allo
w
high
comp
re
ssion
rates, but onl
y at
the cost
of so
me perceived de
gra
dation in ima
ge quality. For
lossy JPEG
comp
re
ssed i
m
age
s, the
main di
storti
o
n
that might
be introdu
ce
d is
blurring
and
ringin
g
. Therefore, it become impe
rati
ve to
develop a quality assessme
nt method that can
evaluate
perceptual im
age
quality a
s
g
ood
as hum
a
n
subje
c
tive
evaluation. T
h
is
ne
ce
ssita
tes
the develo
p
m
ent of obj
ecti
ve IQA app
ro
ach
e
s th
at can auto
m
atically pr
e
d
ict p
e
rceived
JPE
G
-
comp
re
ssed i
m
age qu
ality [1-5].
2.
Image Quality
Assessments
IQA techniqu
es can be div
i
ded into two
grou
ps, nam
ely subje
c
tive and obje
c
tive, which
are di
scusse
d in the following.
2.1. Subjectiv
e
The be
st way for asse
ssi
ng the q
uality of an image i
s
the subje
c
tive qualit
y
measurement
reco
mmen
d
a
tions given
by the ITU [6], which con
s
i
s
ts of Differen
c
e M
ean
Opinio
n Sco
r
e (DM
O
S) from a numb
e
r of expert ob
serve
r
s by lookin
g at ima
ge. Ho
wever,
for
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Local Standa
rd De
viatio
n Based Im
age
Quality Metri
cs fo
r JPEG
Com
p
re
ssed
… (Aksha
y G
o
re
)
7281
most appli
c
at
ions the DM
OS method i
s
inconv
enie
n
t beca
u
se DMOS evalu
a
tion is slo
w
and
co
stly, since i
t
employs a g
r
oup of pe
opl
e in the evalu
a
tion pro
c
e
ss [7].
2.2. Objectiv
e
In orde
r to solve this pro
b
lem i.e. the
need for p
e
ople in the e
v
aluation pro
c
e
ss, an
obje
c
tive app
roa
c
h i
s
requi
red. Su
ch o
b
j
e
ctive
qu
ality asse
ssm
ent
system ha
s g
r
eat potential i
n
a wid
e
ra
nge
of appli
c
atio
n enviro
n
me
nts. Usua
lly t
he obj
ective i
m
age q
uality approa
che
s
can
be catego
rized into three
grou
ps
dep
endin
g
on th
e availability of the origi
n
a
l
image. (1
)
Full
Referen
c
e (F
R) met
hod
s
perfo
rm a di
rect compa
r
ison between t
he imag
e un
der te
st and
a
referen
c
e or
origin
al imag
e. (2) No Ref
e
ren
c
e
(NR)
metrics, are a
pplied when t
he origi
nal im
age
is unavail
abl
e. (3)
Redu
ced Refe
ren
c
e
(RR)
met
r
ics lie between
FR and
NR
metrics an
d are
desi
gne
d to
predi
ct im
ag
e qu
ality with
only p
a
rtial
i
n
formatio
n a
bout the
refe
ren
c
e
imag
e
[2].
Focu
sin
g
on
FR metrics,
the method
s ca
n be
targeted to esti
mate the pre
s
en
ce of JP
EG
-
comp
re
ssed i
m
age
s.
3. R
e
lated
Work
The conventi
onal pixel
-
ba
sed m
e
tri
cs
su
ch
a
s
Me
a
n
Squa
re Error (MSE
), Signal-to
-
Noi
s
e
Ratio
(SNR)
and P
eak Si
gnal
-to
-
Noi
s
e
Ratio
(PSNR) a
r
e
most
widely
use
d
in im
a
ge
pro
c
e
ssi
ng
a
s
the
s
e
metri
c
s are
simpl
e
to calc
ulat
e and
ea
sy to use. Howe
ver, these pi
xel-
based m
e
tri
c
s d
o
n
o
t correlates well
wi
th huma
n
su
bjective
eval
uation, a
nd
rese
arche
r
s h
a
ve
been
devotin
g much effort
s in d
e
velopi
ng adva
n
ced
Huma
n Visu
al System (HVS) IQA mod
e
ls
[8, 9]. Re
ce
ntly, Wang
et
al
propo
sed
Struct
u
r
al
Similarity Index (SSIM) based o
n
the
assumptio
n
that HVS is hi
ghly accu
sto
m
ed to extr
a
c
t stru
ctu
r
al i
n
formatio
n from an ima
g
e
[1].
After the
gre
a
t su
cce
s
s of
SSIM, numb
e
r
of IQA
metr
ics
ha
ve
bee
n
de
ve
lo
pe
d w
i
th a
tte
mp
t to
mimic
the HV
S.
Howeve
r, until
now not even
a si
ngle IQA metric
can completel
y
mimic HVS
for
evaluation p
u
r
po
se. A com
p
reh
e
n
s
ive e
v
aluation an
d
su
rvey of FR-IQA is av
aila
ble in [10-13]. It
is
still a
chall
engin
g
ta
sk t
o
a
c
hieve
10
0% co
ns
i
s
ten
c
y a
huma
n
l
i
ke
perce
ptio
n in IQA
und
er
different ci
rcumstan
ce
s.
Therefore, th
e obje
c
tive o
f
this re
sea
r
ch wo
rk i
s
to
develop
su
ch
a
quality assessment
metri
c
for
JPEG-compresse
d i
m
ages, which will
work
effectively and
effic
i
ently.
In practi
ce, a
n
IQA mo
del
sho
u
ld
be
no
t only effectiv
e but
also eff
i
cient.
Unfort
unately,
accuracy an
d efficien
cy are difficult
to ac
hieve
simultan
eou
sl
y, and most previou
s
IQA
algorith
m
s ca
n rea
c
h only
one of the two goal
s. To
filling this need, in this pape
r we ha
ve
con
s
id
ere
d
the case
of JPEG-di
s
torte
d
image
s, an
d we h
a
ve p
r
opo
se
d a m
odel ba
se
d o
n
a
local
stand
ard deviation of
an image cal
l
ed a
LSDBIQ
model.
The pap
er i
s
organi
ze
d a
s
follows: in Section IV the prop
osed
LSDBIQ tech
nique i
s
pre
s
ente
d
. In Section V the experim
e
n
tal resu
lts (i
n terms of p
e
rform
a
n
c
e compa
r
ison a
nd
efficien
cy evaluation) a
r
e
compa
r
ed. Fin
a
lly
, the concl
u
sio
n
s a
r
e d
r
awn in Se
ctio
n VI.
4. Proposed
T
e
chnique
A sche
m
atic
overview of the LSDBIQ a
ppro
a
ch
prop
ose
d
here, is sho
w
n in Fig
u
re 1.
Figure 1. Overview of ou
r L
S
DBIQ Model
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 728
0
– 7286
7282
4.1.
Local Stand
a
rd Dev
i
ation of an Imag
e
Measures of
local
stand
ard deviation h
a
ve
been
wi
dely use
d
in
image p
r
o
c
e
s
sing fo
r
texture me
asure
s
and
stu
d
ies of
spati
a
l imag
e
st
ru
cture
[14,
15]
. To
cal
c
ulat
e lo
cal
stan
d
a
rd
deviation of a
n
image
I
, a local
stan
dard
deviation filter (std
filt) i
s
a
v
ailable in M
A
TLAB software
[16]. This too
l
perfo
rm
s a l
o
cal
stand
ard
deviation f
ilter
on
a ra
ste
r
imag
e,
i.e. it calculates t
h
e
stand
ard
de
viation withi
n
a nei
ghb
ourin
g a
r
ea
aro
und
ea
ch g
r
id
cell.
A
local
sta
ndard
deviation (
σ
) can b
e
used to empha
si
ze
the local
stru
cture in a
n
im
age an
d defin
ed as:
∑
(
1
)
Whe
r
e
is sta
ndard deviati
on and
N
is
number of pixels
.
The local sta
ndard deviati
on of the refe
ren
c
e
(I
r) an
d
distorted
(Id) images i
s
de
fined as:
Ir = stdfilt (Ir)
Id =
s
t
dfilt (Id)
With the help
of Ir and Id stand
ard d
e
viat
ion map
s
, we defin
e the Local Quali
t
y Map
(LQM
) bet
we
en two imag
e
s
Ir and Id a
s
:
(2)
Whe
r
e T i
s
a
small
po
sitive co
nsta
nt to stabili
ze the
re
sult an
d its p
r
op
ose
d
v
a
lue i
s
0.00
1
0
.
From Equ
a
tio
n
(2), if Ir and
Id are equal,
then LSM will
achieve the
maximum value 1.
5.
Qualit
y
Scor
e Measur
e
m
e
nt
We h
a
ve ap
p
lied ou
r qu
ality sco
re m
e
a
s
urem
ent met
hod to LSM
value
s
u
s
ing
standard
deviation. Th
e prop
osed m
e
trics is
calle
d as LSDBIQ
and defin
e as:
LSDBIQ
∑
/
(
3
)
Whe
r
e
is
:
∑
(4)
Whe
r
e N i
s
n
u
mbe
r
of pixels in the imag
e.
Values
of ob
jective LSDB
IQ and hu
m
an su
bje
c
tive
Differen
c
e
Mean O
p
inio
n Sco
r
es
(
DM
OS) sco
r
e also me
asu
r
es di
sto
r
tion,
lo
wer the val
ue better will
be the image
quality.
6. Experiment
Resul
t
6.1. Demons
tra
t
i
v
e
Results
Figure 2 sh
o
w
s
some
rep
r
esentative result
s from the CSIQ dat
aba
se where
Flowe
r
image with
different levels
of JPEG200
0
comp
re
ssi
on
are compa
r
e
d
. The su
bje
c
tive rating
s
of
quality in term of DM
OS a
r
e al
so
shown for
com
p
a
r
i
s
on. A
s
can
b
e
seen i
n
Fig
u
re
2, from
(a
) to
(f), the level
of JPEG20
00
comp
re
ssion
distorti
o
n
in
crea
se
s an
d so doe
s the
DMOS su
bje
c
tive
rating
s of qu
ality. This make
s the LSDBIQ can
pre
d
i
ct quality of these image
s in a manne
r that
is highly correlated with th
e s
ubj
ective ratings of qu
al
ity.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Local Standa
rd De
viatio
n Based Im
age
Quality Metri
cs fo
r JPEG
Com
p
re
ssed
… (Aksha
y G
o
re
)
7283
(a)
(b)
(c
)
(d)
(e)
(f)
Figure 2. Co
mpari
s
o
n
bet
wee
n
LSDBI
Q
and DMOS
as a su
bje
c
tive quality indicator.
Note th
at
like DM
OS, JPEG-IQA is a
distortion in
d
e
x (a lowe
r DMOS/JPEG-I
QA value me
ans hi
ghe
r
quality). (a)
O
r
iginal im
age
Flowe
r
(DMO
S= 0; LSDBIQ=0
), (b
) Image Flower, its JPEG200
0
comp
re
ssion
Level 1 (DM
O
S=0.0
28; L
S
DBIQ
=0.0
0
70), (c) Imag
e Flowe
r
, its JPEG20
00
comp
re
ssion
Level 2 (DM
O
S=0.1
43; L
S
DBIQ=0.0
2
35), (d
) Imag
e Flowe
r
, its JPEG20
00
comp
re
ssion
Level 3 (DM
O
S=0.4
60; L
S
DBIQ=0.0
7
38), (e
) Imag
e Flowe
r
, its JPEG20
00
comp
re
ssion
Level 4 (DM
O
S=0.7
79; L
S
DBIQ
=0.1
6
30). (f) Imag
e
Flowe
r
, its JPEG2000
comp
re
ssion
Level 5 (DM
O
S=0.9
26; L
S
DBIQ=0.2
2
91).
7. Perf
ormance
C
o
mparison
The pe
rform
a
nce of LSDB
IQ is comp
ared wi
th existi
ng Image Qu
ality Metrics
(IQMs)
inclu
d
ing G
M
SD [17], FSIM [9], SSIM
[1] and PSNR model
s. All the IQMs
wo
ul
d be validate
d
on
four pu
blicly
available im
a
ge data
b
a
s
e
s
that
in
clud
e JPEG2
000
-co
m
presse
d
image
s sub
s
ets:
Tampe
r
e Ima
ge Data
ba
se
2008 (TID2
0
1
3
) [18], Tamp
ere Imag
e Da
tabase 200
8 (TID200
8) [19],
Labo
rato
ry for Ima
ge a
n
d
Video E
n
ginee
ring
(LI
VE) [20] an
d Cate
go
rica
l Image
Qu
ality
Datab
a
se (CSIQ) [21]. We summ
ari
z
e
in Table 1 th
ese d
a
taba
se
s only for wh
at con
c
e
r
ned
the
J
PEG c
o
mpress
ion.
To provide a
compl
e
te e
v
aluation of
each IQM
s
, five commo
nl
y used
perfo
rman
ce
correl
ation co
efficient are
employed a
s
sugg
es
te
d in Video Qual
ity Experts group [22]. The
s
e
five perform
ance metri
cs are th
e Sp
earm
an
Ran
k
-Ord
er Co
rrelation C
oeff
i
cient (S
RO
CC),
Kendall Ra
nk-Order Correl
ation
Coeffici
ent (K
RO
CC), Pearson
Li
near Correlat
ion Coeffici
e
n
t
(PLCC),
Root
Mean Squa
red Error (RM
SE) and Outl
i
e
r Ratio (O
R). In addition, we cho
s
e a five-
para
m
eter lo
gistic fun
c
tion
for nonline
a
r
mappin
g
as
sugge
sted in V
Q
EQ [22].
(
5
)
Whe
r
e
x
den
ote the obje
c
tive sco
re an
d
G(x)
d
enot
es the p
r
edi
cted subj
ectiv
e
DMOS
sco
r
e.
The five para
m
eters are e
s
timated by fitting t
he functi
on to the subj
ective and o
b
j
ective data.
The Ta
ble 1 l
i
sts the S
R
O
CC, KROCC,
PL
CC,
RMS
E
and O
R
re
sults
of LSDBIQ and
other fou
r
IQ
Ms on th
e TID20
13, TID2
008, LI
VE an
d CSIQ data
bases. Th
e b
e
st one
metri
c
prod
uci
ng th
e greate
s
t correlation
s
fo
r each
datab
ase a
r
e ma
rked in bol
dfa
c
e. The Ta
bl
e
1
sho
w
th
at th
e LS
DBIQ p
e
rform
s
be
st
(effe
ctive)
on all
data
b
a
se
in
term
s of
correlat
ion
coeffici
ents.
In order to provide a visua
l
compa
r
ison
of
the fi
ve I
Q
Ms
(PS
NR,
SSIM,
FSIM
, GMSD
and LS
DBIQ), there
scatter plot
s of su
bjective
DM
OS (
G(
x)
)
v
e
rsus obje
c
ti
ve DMOS
score
s
obtaine
d on L
I
VE database
are sh
own in
Figure 3, wh
ere ea
ch p
o
in
t repre
s
e
n
ts
one test imag
e.
The
curve
s
shown in Fi
gure 3. ar
e o
b
tai
ned by
a no
nl
inear fitting function
acco
rd
ing to [23]. O
n
comp
ari
s
o
n
with oth
e
r
scatter plot
s, L
S
DBIQ’s
poi
n
t
s a
r
e m
o
re
close
to n
onlin
ear fitting
curve,
whi
c
h mea
n
s
that LSDBIQ correl
ates
we
ll with subj
ect
i
ve DMOS score.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 728
0
– 7286
7284
8. Efficiency
Ev
a
l
uation
At last, let us discu
ss the
comp
utation
comp
lexity of LSDBIQ wit
h
different IQMs
, We
thus an
alyze
the comp
uta
t
ional co
mple
xity
of LSDBIQ, and then
comp
are the
comp
eting IQ
A
model
s in terms of run
n
ing
time.
Suppo
se that
an ima
ge h
a
s
M*
N pixel
s
. The m
a
in o
p
e
ration
s in
th
e propo
sed
L
S
DBIQ
model in
clud
e cal
c
ulatin
g image lo
cal
standard dev
i
a
tion, there b
y
produ
cing l
o
cal q
uality map
and
quality score. Ove
r
all
,
it requi
re
s f
i
ve lines
cod
e
with
9 M*
N multipli
cati
ons an
d 8
M
*
N
addition
s to y
i
eld the fin
a
l
quality sco
r
e.
The
r
efor
e, computation
complexity of
LSDBIQ i
s
v
e
ry
low a
s
com
p
a
r
e GMS
D
, SSIM and FSIM.
In re
al tim
e
imag
e-pro
c
essing
a
ppli
c
atio
n
runni
ng time
of
IQMs
be
com
e
cru
c
ial.
We thus eval
uated the
running time
of
each four IQM
s
on a Tosh
i
b
a Satellite P
C
with Intel
Core
i3 CP
U a
nd
8GB RAM
a
nd
comp
are
d
with LS
DBIQ. The S
o
ftware pl
atform wa
s
MAT
L
AB
R2012. Table III lis
t the
running time of
t
he four IQMs
on an image of
s
i
ze 512
˟
512
taken
from
CSIQ database. Cl
early, from T
able III, apar
t from
PSNR, the
LSDB
IQ takes only
0.0193
se
con
d
to proce
s
s an ima
ge, whi
c
h is
1.088 time
s f
a
ster tha
n
G
M
SD, 28.44 times faste
r
than
FSIM, 3.53 ti
mes faste
r
th
an SSIM. Cl
e
a
rly, on
e ve
ry attractive
a
d
vantage
of
LSDBIQ i
s
th
eir
efficien
cy co
mpared with
other maj
o
r IQA m
odel
s such PSNR, SSIM, FSIM an
d GMSD etc.
Table 1. Data
bases that Contain jpe
g
-di
s
tortion Ima
g
e
s
Datab
ase
Referen
ce
Images
Images
Consi
d
er
Distor
tio
n
Co
ns
ider
Obser
v
e
r
TID2013
25 500
JPEG
compression
JPEG2000 comp
ression
JPEG transmission erro
rs
JPEG2000 tra
n
smission errors
971
TID2008
25 400
JPEG
compression
JPEG2000 comp
ression
JPEG transmission erro
rs
JPEG2000 tra
n
smission errors
838
LIVE
29 344
JPEG2000 comp
ression
JPEG
20
CSIQ
30 300
JPEG2000 comp
ression
Motion JPEG
compr
e
ssion
35
Table 2. Perf
orma
nce of the Propo
se
d L
S
DB
IQ and the other F
our Competin
g IQA Model
s
Interms of SRC, PCC, KRO
CC, RMSE a
nd or on the
4 Datab
a
ses
D
A
T
A
B
A
S
E
Metrics
LSDBI
Q
G
M
SD
FSIM
SSIM
PSNR
TID2013
(500)
PLCC
0.9167
0.9161
0.9004
0.8714
0.8683
SROCC
0.9084
0.9060
0.8929
0.9194
0.8713
KROCC
0.7285
0.7309
0.6973
0.7381
0.6726
RMSE 0.5718
0.5737
0.6226
2.1614
0.7097
OR
0.0740
0.0700
0.1060
0.1700
0.1040
TID2008
(400)
PLCC
0.8926
0.8796
0.8712
0.8738
0.7918
SROCC
0.9084
0.8977
0.8850
0.8968
0.8145
KROCC
0.7291
0.7152
0.6865
0.7088
0.6068
RMSE 0.7291
0.7462
0.7702
0.7629
0.9581
OR
0.1225
0.1300
0.1225
0.1225
0.1375
LIVE
(344)
PLCC
0.9810
0.9776
0.9389
0.9624
0.8701
SROCC
0.9789
0.9747
0.9357
0.9627
0.8718
KROCC
0.8662
0.8572
0.7762
0.8249
0.6810
RMSE 5.6267
6.1042
9.9855
7.8739
14.2944
OR
0.0523
0.0523
0.0959
0.0872
0.1279
CSIQ
(300)
PLCC
0.9575
0.9541
0.8570
0.9447
0.9039
SROCC
0.9488
0.9347
0.8472
0.9311
0.8957
KROCC
0.7956
0.7667
0.6512
0.7638
0.7043
RMSE 0.0898
0.0933
0.1605
0.1022
0.1332
OR
0.0867
0.0833
0.1000
0.0767
0.0900
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Local Standa
rd De
viatio
n Based Im
age
Quality Metri
cs fo
r JPEG
Com
p
re
ssed
… (Aksha
y G
o
re
)
7285
Table 3. Ru
n
n
ing Time of the Com
p
etin
g IQA Models
IQ
A
M
odels
LSDBI
Q
G
MSD
F
SIM
SSIM
PSNR
Run
n
in
g ti
me (s
)
0.0193
0.0210
0.5490
0.0683
0.0095
(a)
(b)
(c
)
(d)
(e)
Figure 3. Sca
tter Plots of five IQMs on L
I
VE
Database. (a) PSNR, (b
) SSIM, (c) FSIM, (d)
GMSD, (e
) L
S
DBIQ
9. Conclu
sion
In this paper,
we have con
s
ide
r
ed the case of JPEG
-comp
r
e
s
sed i
m
age
s
, and we have
prop
osed
a
FR-IQA
mod
e
l ba
sed
on
a lo
cal
stan
dard
deviatio
n
in a
n
ima
g
e
. Experime
n
t
al
results
sho
w
s that the prop
ose
d
ESDBI
Q mod
e
l pe
rform
s
bette
r i
n
term
s of bo
th accu
ra
cy a
n
d
efficien
cy. Th
e propo
se
d n
e
w m
odel i
s
straig
htfo
rward and
can b
e
easily
gene
ralize
d
to oth
e
r
types of local
feature. Fu
rther
work in
cl
ude
s
extendi
ng the propo
sed al
go
rithm
to assess ot
her
kind
s of disto
r
tion.
Ackn
o
w
l
e
dg
ements
The autho
rs woul
d like to than
ks
Dr. Hu
a-Wen
Chan
g from the Zh
eng
zho
u
Univ
ersity of
Light Indu
stry
, Zheng
zho
u
, Chin
a for
co
nstru
c
tive
pie
c
e
s
of advi
c
e
that have prompted u
s
fo
r
our research.
Referen
ces
[1]
Z
.
W
ang, A.C Bovik, H.R Shei
kh, and E.P Si
monce
lli,
“Imag
e
Qualit
y Asse
ssment: F
r
om Error Visib
ili
t
y
to Structural Si
milarit
y
”, IEEE T
r
ansactions o
f
Image Proces
sing, vol. 13,
p
p
. 1-12, April 2
004.
[2]
Z
.
W
ang, and
A. C. Bovik, “Moder
n Image Q
ua
lit
y Ass
e
ssment S
y
nthes
is
Lectures
on Im
age”, Vi
deo,
and Mu
ltimed
ia
Processin
g
, vol.
2, no.1, pp. 1-15
6, 200
6.
[3]
Zhou Wang, “A
pplic
atio
ns of
Objective Imag
e Qualit
y
Ass
e
ssment
Metho
d
s”, IEEE Signal Processi
n
g
Magaz
ine [1
42
], NO
V 2011
[4]
Z
hou W
a
n
g
a
n
d
Ala
n
C. Bovi
k, “A Univers
a
l
Im
age Q
u
a
lit
y Inde
x,” IEEE Sign
al Proc
ess
i
ng
Letters,
vol. 9, no. 3, pp.
81-84, Mar. 200
2.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 728
0
– 7286
7286
[5]
M Chamb
ah, S Ouni, M. Herbi
n
, E. Z
agroub
a,
“T
o
w
ard
s
an Automati
c Subjectiv
e
i
m
age Qua
lit
y
Assessment S
y
stem”, SPIE-IS&T
/Vol 724
2
724
2. M.
Youn
g, T
he
T
e
chnical W
r
iters Ha
ndb
ook. Mil
l
Vall
e
y
, CA: Un
i
v
ersit
y
Sci
enc
e
,
1989.
[6]
Recomme
nd
ati
on 5
00-
11: m
e
thod
ol
og
y for
the sub
j
ectiv
e
assessm
ent
of the q
ual
it
y for televis
i
o
n
pictures, IT
U-R
Rec. BT
.500, 200
2
[7]
Z
.
W
ang, A. C. Bovik, an
d L.
Lu, “W
h
y
Is Im
age
Qu
alit
y As
sessment So
Difficult?”, IEE
E
Internatio
na
l
Confer
ence
on
Acoustics, Speech,
an
d Sig
n
a
l Process
i
ng,
200
2.
[8]
B. Girod, “W
h
a
t'
s
w
r
on
g
w
i
t
h
me
an-sq
uar
ed
erro
r?
’’, D
i
g
ital Im
ages
a
nd H
u
ma
n Vis
i
on. MIT
press,
199
3.
[9]
Lin Z
h
a
ng, Le
i
Z
hang, X. Mo
u and D. Z
h
a
n
g
, “F
SI
M: A Feature Sim
ilar
i
t
y
Inde
x for Image Qua
lit
y
Assessment,” IEEE
T
r
ans. Image Proc
ess., vol.
20, no. 8, pp. 237
8-2
386,
Aug. 201
1.
[10]
Lih
uo
He, F
e
i
Gao, W
e
il
on
g Ho
u &
Le
i
Hao,
"Ob
j
ectiv
e
ima
g
e
qu
ali
t
y
ass
e
ssme
n
t: a surv
e
y
"
,
Internatio
na
l Journ
a
l of Com
puter Mathem
a
t
ics,Jun 201
3.
[11]
Lin
Z
han
g, L
e
i
Z
han
g,
Xu
an
qin
Mou,
Davi
d Z
h
a
ng: A
c
o
mpre
hens
ive
eval
uatio
n
of full r
e
fere
nc
e
imag
e qua
lit
y
a
ssessment al
g
o
rithms. ICIP 2012: 14
77-
148
0.
[12]
D. M. Chan
dl
e
r
, M. M. Alam,
and
T
.
D. Phan "S
eve
n
ch
a
llen
ges f
o
r ima
ge q
u
a
lit
y rese
arch," Proc.
SPIE Human V
i
sion
and El
ect
r
onic Imag
i
ng
XI
X, 901
40
2 (F
ebru
a
r
y
25, 20
14)
[13]
D. M. Chandl
er, "Seven C
hall
e
n
ges i
n
Im
age Qua
lit
y Assessment: Past, Present, and F
u
tur
e
Rese
arch," ISRN Sig
nal Pro
c
essin
g
, vol. 2013, Article ID
905
68
5, 53 pa
ges, 201
3.
[14]
Gonzal
ez, R. C., and W
i
ntz,
P., 1987. Dig
ital Image Proc
e
ssing, Add
i
so
n
-
W
e
sle
y
, Me
nl
o Park, CA, p.
414.
[15]
W
oodcock, C. E., and Strahle
r
, A.
H., 1987. T
he factor of scale i
n
re
mote
sensi
ng, Remo
te Sensin
g o
f
Enviro
nment, 2
1
, pp. 311-
332.
[16]
MAT
L
AB versi
on 7.10.
0. Nati
ck, Massachus
etts:
T
he MathW
o
rks Inc., 2010.
[17]
W
u
feng
Xue,
L
e
i Z
h
ang,
“Gra
die
n
t Mag
n
itu
d
e
Simi
la
r
i
t
y
De
viatio
n: A H
i
g
h
l
y
Effici
ent P
e
rc
eptua
l Ima
g
e
Qualit
y
Inde
x
”
,
IEEE
T
r
ansactions On Image
Pr
ocessi
ng, Vo
l. 23, No. 2, Februar
y
2
0
1
4
.
[18]
N. Ponomar
en
ko, O. Ieremeiev, V. Lukin, K. Egiaza
ri
an,
L. Jin, J. Astola,
B. Vozel,
K. Chehd
i, M.
Carli, F
.
Battisti, C.-C. Jay Kuo, Col
o
r Image
D
a
tab
a
s
e T
I
D2013:
Pecul
i
ariti
e
s a
nd Prel
imin
ar
y
Results, Proce
edi
ngs of 4th E
u
ro
p
i
an W
o
rks
hop o
n
Visu
al I
n
formatio
n
Pro
c
essin
g
EUVIP201
3, Paris
,
F
r
ance, June
1
0
-12, 20
13, pp.
106-1
11.
[19]
N. Ponom
are
n
k
o, V. Luki
n
, A. Z
e
lensk
y
, K. E
g
iaz
a
ria
n
, M. C
a
rli, F
.
Battisti, "T
ID2008 - A D
a
tabas
e fo
r
Evalu
a
tion
of
F
u
ll-R
e
feren
c
e Visu
al Q
ualit
y Ass
e
ss
ment Metrics
"
, Advanc
es
of Mod
e
r
n
Radi
oe
lectro
ni
cs, Vol.
10, pp. 30-45, 20
09.
[20]
H. R. Sheikh,
Z
.
W
ang, L.
Cormack, an
d A. C. Bovik
.
(2005) “L
ive
Image Qua
lit
y Assessm
ent
Database Release 2",http://live.ec
e.ute
x
as.e
du/rese
arch/q
u
a
lit
y
.
[21]
Eric C. L
a
rso
n
an
d D
a
mo
n
M. Cha
ndl
er, “Mo
st ap
pare
n
t distorti
on: full-refer
enc
e i
m
age
qu
ali
t
y
assessme
n
t an
d the role of strateg
y
,” J.
Electron. Imagin
g
1
9
, 0110
06, Jan
07, 201
0.
[22]
VQEG, “F
inal r
eport from th
e
vide
o q
ual
it
y e
x
p
e
rts
gro
u
p
o
n
the v
a
li
dati
o
n
of ob
jectiv
e m
ode
ls of vi
de
o
qua
lit
y
assess
ment”, http://
w
w
w
.
v
q
e
g
.org, 2
000.
[23]
A. J. Bell an
d
T
.
J. Sejno
w
sk
i, “An inform
ati
on-ma
xim
i
zati
o
n
ap
proac
h to
blin
d se
par
atio
n an
d bl
in
d
deconvolution,” Neural
Comput., vol.7, no.
6, pp.
112
9–1159, 1995.
Evaluation Warning : The document was created with Spire.PDF for Python.