Indonesi
an
Journa
l
of El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
13
,
No.
1
,
Jan
uar
y
201
9
,
pp.
1
9
9
~
2
04
I
SS
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
3
.i
1
.pp
1
99
-
2
04
199
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Colored
facial im
age r
esto
ration b
y s
imi
l
arity enh
an
ce
d
imp
licative fuz
zy ass
oc
i
ation mem
ory
Kwan
g
B
aek Ki
m
1
,
D
oo He
on
Song
2
1
Depa
rtment of
Com
pute
r
Engi
n
ee
ring
,
Sil
la Uni
ver
sit
y
,
Busan
4
6958,
Kore
a
2
Depa
rtment of
Com
pute
r
Gam
es,
Yong
-
In
Song
Dam
Coll
ege,
Y
ong
-
in
17145,
K
ore
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Sep
1
2
, 201
8
Re
vised N
ov
9
, 2018
Accepte
d Nov
19
, 201
8
Im
age
restor
at
io
n
ref
ers
to
the
rec
over
y
of
an
under
l
y
ing
imag
e
from
an
observa
ti
on
tha
t
has
bee
n
cor
ru
pte
d
b
y
var
ious
t
y
p
es
of
noise
.
In
a
digita
l
fore
nsic
softwar
e,
such
image
r
estora
t
ion
proc
e
ss
should
be
no
ise
-
tolera
nt
,
robust,
fast,
an
d
sca
la
bl
e.
In
thi
s
pape
r,
w
e
apply
impli
c
at
iv
e
fuz
z
y
associa
t
i
on
m
e
m
ory
struct
ur
e
in
co
lore
d
f
a
ci
a
l
image
rest
ora
ti
on
wi
th
enha
nc
ed
sim
il
a
rity
m
ea
sure
inv
olve
d
in
output
computar
ion.
Th
e
eff
i
cac
y
i
f
the
proposed
fuz
z
y
associative
m
emor
y
m
odel
is
ver
ified
b
y
the
e
xper
iment
in
tha
t
it
was
95%
succ
essful
(w
it
h
ze
ro
m
ea
n
square
err
or)
out
of
20
te
sted
images.
Ke
yw
or
ds:
Faci
al
i
m
age
Fu
zzy
ass
ociat
ive m
e
m
or
y
Fu
zzy
sim
il
arity
Im
age r
est
or
at
i
on
Me
an
s
qu
a
re e
rror
Copyright
©
201
9
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed.
Corres
pond
in
g
Aut
h
or
:
Kw
a
ng Bae
k Kim
,
Dep
a
rtm
ent o
f C
om
pu
te
r
E
ng
i
neer
i
ng,
Sil
la
U
niv
e
rsit
y,
Busan 4
6958,
Korea.
Em
a
il
:
gb
kim
@sil
la
.ac.kr
1.
INTROD
U
CTION
Im
age
resto
rati
on
proce
ss
is
an
im
po
rtant
s
te
p
us
ed
in
de
no
isi
ng
that
re
fer
s
to
t
he
rec
ov
e
ry
of
a
n
unde
rly
ing
im
age
from
an
ob
ser
vatio
n
that
has
bee
n
co
rrup
te
d
by
noise
.
Ther
e
a
re
va
rio
us
source
s
le
t
the
dig
it
al
i
m
ages
to
be
c
orrupt
ed
by
po
or
c
on
t
rast
an
d
noise
.
T
hese
s
ources
i
nclu
de
i
m
age
trans
m
issi
on
,
acqu
isi
ti
on,
c
o
m
pr
ession,
qu
antiz
at
ion
,
il
lu
m
inati
on
co
nd
it
ion
s,
m
al
fu
nc
ti
on
in
g
inst
rum
ents,
il
l
po
s
it
ion
s
and
m
or
e.
T
he
ap
plica
ti
on
area
of
s
uch
im
age
de
no
isi
ng
i
nclud
e
s
ge
ner
al
obj
ect
rec
ogniti
on,
dig
it
al
entertai
nm
ent,
m
edica
l
i
m
age
under
s
ta
nd
in
g
an
d
re
m
ote
sensing
i
m
aging
[
1]
an
d
we
are
espe
ci
al
ly
interest
ed
i
n
t
he
resto
rati
on
of
c
orrupted
fac
ia
l
i
m
age
for
dig
it
al
f
or
e
ns
ic
ap
plica
ti
on
s
[
2].
Sim
il
arly
,
i
m
age
denoisin
g
is
a
n
inv
e
rse
pro
blem
of
any
im
ag
e
corr
up
ti
on p
r
ocess
a
nd
ca
n
be
viewe
d
as
a
f
il
te
ring
syst
em
with
v
isi
ble
f
uzzine
ss
as
the
existe
nce
of
f
uzzine
ss
in
t
he
im
age
sign
al
a
nd
c
on
ta
m
inate
d
sign
al
t
hu
s
m
any
fu
zzy
log
ic
base
d
a
ppr
oac
hes have
sh
ow
n
th
ei
r
str
eng
t
hs
i
n
this
dom
ai
n
[3
]
.
In
this
pa
per,
we
are
especia
ll
y
interest
ed
in
the
resto
rati
on
of
co
rru
pted
fa
ci
al
i
m
age
fr
om
ro
ug
h
i
m
ages
li
ke
c
asual
sm
artphon
e
ph
otogra
phs
a
nd
im
ages
from
cl
os
ed
ci
rcu
it
te
le
vision
(CCT
V)
c
a
m
era.
In
rece
nt
ye
ars
,
CC
TV
has
be
en
widely
us
e
d
fo
r
recordi
ng
rand
om
scenes
that
are
easy
t
o
identify
s
us
pe
ct
ed
crim
inals
and
it
is
even
we
ll
visible
at
night.
Digital
f
or
e
ns
ic
co
ncerns
with
an
a
uto
m
at
ed
face
rec
ogniti
on
scenari
o
th
at
involves
com
par
in
g
degrade
d
facial
phot
ogra
phs
of
sub
j
ect
s
a
gainst
t
heir
high
-
res
olu
ti
on
counter
par
ts
[
4
].
Howe
ver
,
pr
ob
le
m
of
te
n
e
ncou
ntere
d
in
forensi
c
face
r
ecognit
ion
in
volves
l
ow
-
res
ol
ution
face
im
ages
that
hav
e
bee
n
fa
xed,
pri
nte
d,
or
hea
vily
com
pr
esse
d
[
5].
T
he
refor
e
,
it
is
nec
essary
to
resto
r
e
the
dam
aged
colo
r
i
m
ages
fr
om
CC
TV
or
sm
artphon
e
t
hat
ha
s
been
a
n
im
po
rta
nt
subsyst
em
of
any
de
ve
lop
e
d
dig
it
al
fo
re
ns
ic
softwa
re.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
1
9
9
–
2
0
4
200
Wh
il
e previ
ous
att
e
m
pts o
n
the syst
e
m
atic i
m
age r
est
or
at
i
on
from
d
egr
a
ded
im
age in
d
igit
al
f
or
e
ns
ic
so
ft
war
e
a
re
usual
ly
com
par
ing
degra
ded
i
m
ages
with
hi
gh
-
re
so
l
ute
saved
m
od
el
s
of
ex
-
c
onvicts
[
4
-
5],
there
is
a
gro
wing
us
a
ge
of
low
r
esol
ute
colo
r
im
ages
in
repo
rting
s
uspic
io
us
s
ubj
e
ct
of
s
uch
inc
idents
.
Th
us
,
we
ne
ed
fast
but
rob
us
t
aut
oma
ti
c
facial
im
age
de
no
isi
ng
/
reconstr
ucting
m
et
ho
do
l
ogy
[
2].
A
c
on
si
der
a
ble
num
ber
of
pra
ct
ic
al
app
li
cat
ion
s
inclu
ding
dig
it
al
al
bu
m
,
su
r
veill
ances
vi
de
os
proces
sing
an
d
per
s
onal
au
t
he
ntica
ti
on
i
nvol
ve
the
r
est
or
at
i
on of
blurre
d f
aces [
6].
W
it
h
that
pur
po
s
e,
we
ta
ke
associat
ive
m
e
m
or
y
appro
ac
hes
as
our
e
ngine
f
or
im
age
restor
at
io
n.
Since
it
is
ai
m
ed
at
st
or
i
ng
a
nd
recall
in
g
as
so
ci
at
ion
s
am
on
g
patte
r
ns
(
da
ta
)
[
7].
Am
on
g
m
any
m
odel
s
of
associat
ive
m
e
m
or
y,
Fu
zzy
Associ
at
ive
m
e
m
or
y
(F
AM
)
[8
]
is
on
e
of
th
e
su
ccessf
ul
im
ple
m
entat
ion
of
suc
h
structu
res
us
i
ng
a
f
uzzy
Hebbian
le
arn
i
ng
r
ule
in
te
rm
s
of
m
ax
-
m
in
or
m
ax
-
pro
duct
c
om
po
sit
ion
s
f
or
the
synthesis
of
it
s
wei
gh
t
m
at
rix
W
bu
t
it
al
so
has
sim
il
ar
low
capaci
ty
pro
blem
[9
]
.
FAM
s
po
sses
s
i
m
portant
adv
a
ntage
s
inc
lud
in
g
noise
tolera
nce,
unli
m
it
ed
storag
e,
and
one
pas
s
converge
nce
though.
An
im
portant
pro
per
ty
,
deci
ding
FA
M
perform
ance,
is
t
he
abili
ty
to
captu
re
co
nte
nt
of
each
patte
r
n,
an
d
ass
ocia
ti
on
of
patte
rn
s
[1
0].
Im
plica
t
ive
Fu
zzy
Associ
at
ive
Mem
or
ie
s
(I
F
AM)
is
a
sing
le
-
la
ye
r
feedfo
r
ward
f
uzzy
neural
netw
orks
e
quipp
e
d
with
ne
uron
s
that
c
om
pu
te
the
m
axim
um
of
a
t
-
nor
m
.
IF
AM
ex
hi
bits
excell
ent
t
olera
nce
with
re
sp
e
ct
to
ei
ther
po
sit
iv
e
or
ne
gative
no
ise
with
optim
al
abso
lute
stora
ge
ca
pacit
y
[11].
Wh
il
e
FA
Ms
m
igh
t
be
us
e
d
as
a
powerfu
l
too
l
f
or
im
pl
e
m
enting
f
uzz
y
ru
le
-
based
s
yst
e
m
s,
the
insigh
t
that
F
A
Ms
are
cl
os
el
y
relat
ed
to
m
at
he
m
at
i
cal
m
or
pholog
y
has
le
d
t
o
the
dev
el
op
m
ent
of
ne
w
f
uzzy
m
or
phol
og
ic
al
associat
ive
m
e
m
or
y
m
od
el
s
and
IF
AM
is
a
good
e
xam
ple
of
them
wh
ic
h
is
al
so
m
a
them
a
ti
cal
l
y
s
ound
[12
-
13]
. Som
e IF
AM
m
od
el
s
exh
i
bited t
heir usef
uln
e
ss in
im
age r
est
orat
ion p
r
ob
le
m
s [
14, 1
5].
In
t
his
pa
pe
r,
we
propose
a
c
olored
facial
im
age
resto
rati
on
m
echan
ism
unde
r
I
FA
M
s
tructu
re
with
e
m
ph
asi
zi
ng
the
si
m
il
arity
m
easur
e
of
the
patte
rn
s
bei
ng
c
on
si
der
e
d.
Pr
evi
ously
,
FA
M
struct
ure
was
su
ccess
fu
ll
y
a
pp
li
ed
to
our
a
pp
li
cat
io
n
do
m
ai
n
with
gr
ey
m
od
el
[16].
H
ow
e
ve
r,
t
he
ge
ner
al
FA
M
str
uctu
re
te
nd
s
t
o
ha
ve
t
oo
m
any
0’s
in
relat
ion
with
t
he
c
onnecti
on
stren
gth
m
at
rix
and
the
t
hr
es
hold
due
to
it
s
m
ax
-
C
op
e
rato
r
usa
ge
.
This
F
AM
c
har
act
erist
ic
s
r
esults
in
f
re
quent
incom
plete
i
m
age
restor
a
ti
on
in
real
w
orl
d
app
li
cat
io
ns
.
T
hu
s
,
we
a
d
op
t
IF
AM
m
od
el
instea
d
of
FA
M
an
d
le
t
t
he
m
od
el
work
directl
y
from
the
col
or
e
d
i
m
age f
or
fast
scal
abili
ty
co
nc
ern
s
.
2.
IMAGE
RES
TORA
TI
ON
WITH P
ROP
OSED IFA
M
STRU
T
U
RE
An
ass
ociat
ive
m
e
m
or
y
par
ad
igm
m
a
y
be
form
ulate
d
as
an
inp
ut
–
outp
ut
s
yst
e
m
that
is
a
ble
to
s
tore
diff
e
re
nt
patte
rn
s
pairs.
F
A
M
m
od
el
s
are
cl
assifi
ed
into
tw
o
cat
eg
or
ie
s
–
aut
o
-
as
s
ociat
ive
and
heter
o
-
associat
ive.
He
te
ro
-
a
sso
ci
at
iv
e
FA
M
m
od
el
has
sim
il
ar
struc
ture
with
Bi
directi
on
al
A
ss
oc
ia
ti
ve
Mem
ory
[17]
that
has
ty
pica
l
struct
ur
e
sho
wn
as
Fig
ur
e
1.
I
n
t
his
sc
hem
e,
the
retrie
ve
d
patte
r
n
is
dif
f
eren
t
from
the
input
patte
rn not
on
l
y i
n
co
ntent
bu
t possi
bly al
so
in ty
pe
a
nd form
at
.
Figure
1. Heter
o
-
ass
ociat
ive
f
uzzy ass
ociat
iv
e m
e
m
or
y st
ru
ct
ur
e
In
this
pa
per,
we
a
dopt
a
uto
-
associat
ive
FAM
m
od
el
that
r
et
rieves
a
pre
vi
ou
sly
st
or
e
d
pa
tt
ern
wh
ic
h
m
os
t cl
os
el
y resem
bles the curr
e
nt p
at
te
r
n. F
ig
ure
2
s
hows
a ty
pical
stru
ct
ur
e
of FAM m
od
el
.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Colore
d
faci
al
image rest
orati
on b
y
simil
ar
it
y en
hance
d
im
pl
ic
ative
fuzzy
as
s
ociatio
n…
(
Kw
ang
B
aek K
im
)
201
Figure
2.
A
uto
-
asso
ci
at
ive fu
zzy
asso
ci
at
ive
m
e
m
or
y st
ru
ct
ur
e
The
gen
e
rali
zed
F
AM
(
GFAM)
[
18]
,
ca
n
be
desc
ribe
d
in
te
rm
s
of
the
fo
ll
ow
i
ng
relat
io
ns
hi
p
betwee
n
a
n
in
put patt
er
n
x
∈
[
0,
1]
n
an
d
t
he
c
orres
pondin
g o
utput patt
er
n
y
∈
[
0,
1]
m
as s
how
n
i
n form
ul
a (
1)
.
Y
=
(
W
°
)
°
(1)
wh
e
re
Y
denot
e
the
outp
ut
a
nd
W
de
no
t
e
t
he
c
onnecti
on
stren
gth
m
at
rix
with
t
hr
es
ho
ld
θ.
Θ
is
def
i
ned
as
E
quat
ion (
2).
θ
=
⋀
=
,
0
≤
≤
1
(2)
wh
e
re
y
k
de
note
s the
outp
ut of
k
th
le
ar
ne
d pa
tt
ern
(
m
axim
u
m
p
patte
rn
s
av
ai
la
ble).
In
E
quat
io
n
(1),
t
he
sym
bo
l
◦
T
de
no
te
s
t
he
m
ax
-
C
pr
oduct
w
her
e
C
is
a
t
-
norm
that
sat
isfie
s
op
e
rato
r
associat
ivit
y,
c
omm
uta
ti
vity
,
and
m
on
ot
on
e
non
-
decr
easi
ng
[11].
And
F
or
A
∈
[0,
1]
mxp
and
B
∈
[
0,
1]
pxn
then
t
he
m
ax
-
C pr
oduct C
=
A ◦
B is
d
e
fine
d
as
E
quat
io
n (
3)
=
⋁
(
,
)
∀
=
,
…
,
,
∀
=
1
,
…
,
=
1
(3)
Howe
ver,
du
ring
the
im
age
r
est
or
at
io
n
proc
ess
with
t
his
G
FA
M
str
uctu
re
,
the
r
es
ult
ou
t
pu
t
obta
ined
from
the
E
qu
at
ion
(1)
fall
s
int
o
0’s
m
or
e
tha
n
de
sired
due
t
o
the
m
ax
-
C
operat
or
c
har
act
erist
ic
s
as
sho
wn
i
n
E
quat
ion (
2). T
hu
s
, we a
dopt
I
FA
M st
r
uct
ur
e
w
it
h
c
ol
ore
d p
ixel co
ntr
ol as
sh
ow
n
in
E
qu
a
ti
on
(4).
W
=
(
Y
°
)
,
=
(
°
)
⋁
(4)
wh
e
re
x
rgb
de
note
the
col
or
inf
or
m
at
ion
of
the
pix
el
.
I
n
de
ci
din
g
W,
we
ta
ke
Mi
n
op
e
rati
on
an
d
the
ou
t
pu
t
Y
is c
om
pu
te
d by
Ma
x o
per
at
i
on.
Ba
sed
on
this
IF
AM
st
ru
ct
ure,
we
a
dd
patte
rn
sim
il
arity
t
o
com
pu
te
the
final
outp
ut.
I
f
there
is
a
no
ise
in
t
he
in
put patt
er
n, the
r
est
or
at
io
n proc
ess is ba
sed
on
t
he
sim
il
arity m
easur
e s
how
n
as
E
quat
io
n (
5).
simi
larity
=
‖
∧
′
‖
‖
′
‖
(5)
w
he
re
S a
nd S’ d
e
no
te
t
he
st
ored
patte
rn an
d i
nput p
at
te
r
n w
it
h
noise
in
r
es
pecti
vely
.
Th
us
, t
he final
ou
t
pu
t
Y
is c
om
pu
te
d
by E
qu
at
ion
(6).
Y
=
′
∨
∨
(6)
In
order
to
e
va
luate
the
corre
ct
ness
of
resto
red
pix
el
,
we
t
ake
m
ean
sq
ua
re
e
rror
(MSE)
m
easur
e
as
sh
ow
n
in
E
qu
a
ti
on
(7).
MSE
=
∑
(
−
̂
)
2
=
1
−
2
(7)
wh
e
re
y
1
de
no
t
es the
or
i
gin
al
pix
el
a
nd
̂
is t
he
d
e
gr
a
de
d pix
el
. Th
e
d
e
gree
of freed
om
is
n
-
2.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
1
9
9
–
2
0
4
202
3.
E
X
PERI
MEN
T
The
pro
posed
m
et
ho
d
was
i
m
ple
m
ented
in
C#
under
Vis
ual
Stud
i
o
2017
en
vironm
ent
with
In
te
l
(R
)
Du
al
C
or
e
(T
M)
i3
-
5005
U
CPU
@
2.0
G
Hz
a
nd
4
GB
RAM
PC.
We
te
ste
d
20
fa
ci
al
i
m
ages
an
d
s
om
e
su
ccess
fu
l a
nd
f
ai
le
d
e
xam
ple
s ar
e
show
n i
n Fi
gure
3
a
nd Fi
g
ure
4 res
pecti
vely
.
(a)
O
rigin
al
Im
age
(b)
Dam
aged
i
nput
(c
)
Re
st
or
e
d res
ult
Figure
3. S
ucc
essfu
l
im
age r
est
or
at
io
n by th
e pro
posed
I
F
AM (
MS
E =
0.0)
(a)
O
rig
inal
Im
age (
b) D
am
aged in
pu
t
(c)
Re
store
d
re
su
lt
Figure
4. Fail
ed
im
age r
est
or
at
ion
by the
pr
op
os
e
d IF
AM
(MSE =
0.06)
As
s
how
n
in
Figure
3,
we
m
ake
so
m
e
dam
age
on
t
he
ori
gin
al
im
age
(F
ig
ure
3(
a)
)
and
m
ake
the
inout
li
ke
Fig
ure
3(b
)
that
is
an
extr
em
e
exa
m
ple
of
stst
em
at
ic
delet
ion
of
non
-
ne
glisa
ble
pa
rt
of
t
he
i
m
age.
The
pro
posed
IF
AM
was
s
uc
cessf
ul
to
res
tor
e
t
he
lo
st
pa
rt
as
s
h
own
i
n
Fi
gure
3(
c
)
whose
MSE
va
lue
conve
rg
e
rs
to
zero.
That
r
esult
is
act
ually
par
tl
y
helped
by
the
s
ymm
et
rici
t
y
of
the
facial
im
age.
Howe
ver, f
ace
recog
niti
on
fro
m
su
ch
ki
nd of
p
a
rtly
tor
n p
hoto
gr
a
ph
occur
s in
t
he
real
w
orl
d
a
ppli
cat
ion
.
In
Fi
gure
4,
t
hough,
we
ha
ve
the
only
fail
ed
restor
at
io
n
case
from
ou
r
e
xp
e
rim
ent.
In
that
case,
there
exists
no
n
-
ne
glisable
siz
e
of
m
os
ai
c
ar
ea
in
the
dam
a
ged
in
put
(F
ig
ur
e
4(
b)).
I
n
suc
h
a
case,
the
pi
xel
in
the
m
os
ai
c
area
on
ly
ha
d
the
aver
a
ge
d
col
or
inf
orm
atio
n
thu
s
it
wa
s
not
su
f
fici
ent
f
or
IF
AM
’s
Ma
x
-
Mi
n
op
e
rati
on
to
re
store
the
lost
inf
or
m
at
ion
.
Th
us
,
that
la
ck
of
inform
ation
re
su
lt
s
in
the
fail
ed
case
as
sh
own
in
Figure
4(c)
where
MSE
is
0.
06.
Sti
ll
,
in
our
exp
e
rim
ent,
M
SE
co
nv
e
rg
e
s
to
zer
o
in
the
re
st
19
s
yst
em
a
tical
ly
or
ta
ndom
ly
da
m
aged
phot
ogra
phs.
For
t
he
rec
ord,
the
a
ver
a
ge
M
SE
usi
ng
the
grey
m
od
e
IFAM
[
16
]
was
0.21 on t
he
sa
m
e set of
e
xam
ples th
us
t
he pr
opos
e
d
m
et
ho
d sh
ows a
big i
m
pr
ov
em
ent.
Othe
r
tha
n
tha
t
on
e
cas
e,
the
eff
ic
acy
of
th
e
pro
po
se
d
m
et
ho
d
ca
n
be
dem
on
strat
ed
as
show
n
i
n
Figure
5
t
hat
directl
y
com
par
es
the
res
ult
with
the
pr
e
vious
gre
y
GFAM
st
ruct
ur
e
us
e
d
i
n
[
16
]
.
On
e
can
clea
rl
y see
that the
prop
os
ed
m
et
ho
d
s
uccess
fu
ll
y
resto
res
the
pre
viously
f
ai
le
d
no
ise
.
(a)
G
ray GFA
M [16]
(b)
P
rop
os
e
d
Figure
5. Com
par
i
ng Facial
im
age r
est
orat
ion wit
h p
re
vious
gr
ey
m
od
e
I
FA
M
[16]
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Colore
d
faci
al
image rest
orati
on b
y
simil
ar
it
y en
hance
d
im
pl
ic
ative
fuzzy
as
s
ociatio
n…
(
Kw
ang
B
aek K
im
)
203
The
reas
on
of
f
ai
lure
by
the
pr
evio
us
at
te
m
pt
sh
ow
n
i
n
Fi
g
ure
5(a)
is
that
as
m
entioned
i
n
S
ect
ion
2,
E
quat
ion
(
1)
c
onve
rg
es
to
ze
r
o
m
or
e
of
te
n
w
hen
the
obj
e
ct
was
t
o
o
sm
alle
r
c
om
par
ed
with
the
bac
kgr
ou
nd
in
the
process
of
Ma
x
-
Mi
n
operati
on.
The
pro
posed
sim
il
a
rity
enh
a
nced
IF
AM
overc
om
es
that
pr
obl
e
m
as
sh
ow
n
in
Fi
gur
e
5(b
).
4.
CONCL
US
I
O
N
In
this
paper,
we
e
xten
d
t
he
IFAM
m
od
el
wir
h
e
nh
a
nce
d
sim
il
arit
y
m
e
asur
e
to
t
he
c
olored
facial
i
m
age
resto
rati
on
case
s.
Vari
ou
s
ty
pe
s
of
FA
M
m
od
el
s
hav
e
bee
n
wi
dely
ap
plied
i
n
m
any
eng
i
ne
erin
g
app
li
cat
io
ns
i
n
the
la
st
tw
o
decad
e
s
an
d
the
re
a
re
di
ff
ere
nt
m
od
el
s
to
a
dap
t
t
he
ap
plica
ti
on
areas’
char
act
e
risti
cs
for
ta
sk
s
s
uc
h
as
est
i
m
ation
,
pr
e
dicti
on
a
nd
infer
e
nce.
W
hat
we
are
i
nterested
in
us
in
g
t
his
associat
ive
m
em
iry
m
od
el
is
to
de
velo
p
a
su
bsy
ste
m
in
a
dig
it
al
f
oren
sic
syst
e
m
with
a
n
a
uto
m
at
e
d
face
recog
niti
on
s
c
enar
i
o
that
i
nvol
ves
c
om
par
ing
de
gr
a
d
e
d
facial
phot
ogr
aphs
of
sub
j
ec
ts
against
t
hei
r
high
-
reso
l
ution
c
ou
nter
par
ts
f
ro
m
casual
sm
artph
one
photogra
ph
s
a
nd
cl
os
e
d
ci
rcu
it
te
le
vision
(CCT
V)
c
a
m
era.
Also
,
we
nee
d
a
fast,
rob
us
t,
and
scal
a
ble
m
et
hod
to
do
it
.
IF
AM
m
od
el
s
are
stron
gly
noise
tolerant
but
have
zero
-
co
nver
ge
nce
prob
le
m
in
Mi
n
-
Ma
x
op
erati
on.
Ou
t
si
m
il
arity
enh
a
nc
ed
I
FA
M
str
uc
ture
was
desi
gn
e
d
t
o
avo
i
d
t
hat
zer
o
-
c
onve
rg
e
nce
pro
blem
and
the
pr
opos
e
d
m
et
ho
d
was
hi
gh
ly
s
uccess
ful
(19
out
of
20
te
st
e
d
cases o
r 9
5%
s
uccess
rate) i
n exp
e
rim
ent.
REFERE
NCE
S
[1]
R.
Yan,
et
a
l
.
,
“
Nonloca
l
hi
era
r
c
hic
a
l
dic
t
iona
r
y
le
arn
ing
using
wave
l
et
s
for
image
denoi
sing,
”
IE
EE
Tr
ansacti
ons
on
Image Proce
s
sing
,
vol
.
22
,
no
.
12,
pp.
4689
-
46
98,
2013
.
[2]
K.
B.
Kim
and
D.
H.
Song,
“
Faci
a
l
Im
age
Deno
ising
from
Degra
ded
Roug
h
Cas
ual
Photographs
using
Hopfiel
d
Neura
l
N
et
work,
”
Int
ernati
onal
I
nformation
Institute
(
Tokyo)
.
Info
rm
ati
on
,
vol
.
20
,
no.
4ª,
pp
.
2513
-
2518,
2017
.
[3]
I.
Irum
,
et
al
.
,
“
A
Revi
ew
of
Ima
ge
Denoising
Methods,
”
Jour
nal
of
Engi
ne
eri
ng
Sci
ence
&
T
ec
hnolog
y
Review
,
vol.
8
,
no
.
5
,
pp
.
41
-
48,
2015
.
[4]
T.
Bourl
ai
,
et
al
.
,
“
Restori
ng
deg
rad
ed
fa
ce
images:
A
ca
se
stud
y
in
m
at
chi
ng
f
ax
ed,
prin
te
d
,
and
sca
nned
photos
.
IEE
E
Tr
ansacti
o
ns on
Informatio
n
Forensic
s and
Sec
urit
y
,
vol. 6,
no.
2
,
pp
.
371
-
3
84,
2011
.
[5]
A.
K.
Jain,
et
al
.
,
“
Face
m
at
c
hing
and
ret
ri
e
val
in
fore
nsics
appl
icati
ons.
I
EE
E
multi
med
i
a
,
vol.
19,
no.
1,
pp.
2
-
10
,
2012
.
[6]
F.
Xin
,
et
al
.
,
“
F
ac
e
image
r
estor
at
ion
base
d
on
st
at
isti
cal
prior
an
d
image
blur
m
ea
sure,
”
In
Mul
tim
edi
a
and
Ex
po,
2003.
ICME'03
.
Proce
ed
ings.
20
03
Inte
rnat
ional
Confe
renc
e
on
.
I
EE
E
,
vol
.
3
,
no
.
III
-
297,
2003
Jul
.
[7]
B.
Kos
ko,
“
Ada
pti
ve
bidi
r
ec
t
ion
al
associa
t
ive m
emories,
”
Appl. O
pt
.
,
vol. 26, no. 23, pp. 4947
-
49
60,
1987
.
[8]
B.
Kos
ko,
Neur
al
net
works
and
fuzzy
systems:
a
dynamic
al
systems
approach
t
o
machine
int
e
lligence
,
Pren
ti
c
e
Hall
.
1992.
[9]
P.
Sus
sner
and
M.
E.
Val
le,
Fu
zz
y
ass
oci
ati
v
e
memor
ie
s
and
t
hei
r
rela
ti
onship
to
math
emati
ca
l
mor
phology
,
I
n
Skow
ron
A,
Pedr
y
c
z
W
,
Krein
ovic
h
V.
Edi
tor
s
Handbook
of
Granul
ar
Com
puti
ng,
New
York.
John
W
il
e
y
&
Son
s.
,
pp.
1
-
41,
2008.
[10]
T.
H.
Nong
and
T.
K.
Dang
,
“
Im
proving
learni
ng
rule
for
fuz
z
y
a
ss
oci
at
iv
e
m
emor
y
with
combina
ti
on
of
con
te
nt
and
associ
at
ion
,
”
Neurocomput
i
ng
,
vol
.
149
,
pp
.
59
-
64,
2015
.
[11]
P.
Sus
sner
and
M.
E.
Valle,
“
Im
pli
ca
ti
v
e
fu
zzy
associa
t
ive
m
emories,
”
IE
EE
T
rans
act
ions
on
Fuzzy
Syste
ms
,
vol.
14,
no
.
6
,
pp
.
79
3
-
807,
2006
.
[12]
M.
E.
Valle
and
P.
Sus
sner,
“
A
g
ene
ra
l
fra
m
ework
for
fuz
z
y
m
or
phologi
c
al
assoc
ia
ti
v
e
m
emories.”
Fuzzy
sets
and
syst
ems
,
vol
.
15
9,
no
.
7
,
pp
.
747
-
768,
2008
.
[13]
M.
Vajgl
and
I.
Perfil
j
eva,
“
Autoassociative
Fu
zzy
Im
pli
c
at
iv
e
Mem
ory
on
the
Plat
form
of
Fuzz
y
Preorde
r
,
”
In
IFSA
-
EUSFLAT
,
vol. 30, pp. 159
8
-
1603,
2015
.
[14]
M.
E.
Valle
and
A.
C.
de
Souza,
“
On
the
rec
all
ca
pab
il
i
t
y
of
r
ec
urre
n
t
exponentia
l
fu
zzy
associative
m
emorie
s
base
d
on
sim
ilar
ity
m
ea
sures
,
”
M
athware
and
So
f
t
Computing
Ma
gazine
,
vo
l. 22,
pp.
33
-
39
,
2015
.
[15]
G.
Ta
n
aka
and
K.
Aihar
a
,
“
Com
ple
x
-
val
ued
m
ult
istate
associa
t
ive
m
emor
y
wi
t
h
nonli
ne
ar
m
ultile
v
el
func
ti
ons
for
gr
ay
-
le
v
el i
m
age re
constru
ct
io
n,
”
IEEE
Tr
ansactions
on
Neural
N
et
works
,
vo
l. 20, no. 9, pp. 1463
-
1473,
2009
.
[16]
J.
H.
Le
e
and
D.
H.
Song,
“
A
F
ast
Scal
ab
le
Im
a
ge
Restora
t
ion
b
ase
d
on
Fuzz
y
As
socia
ti
ve
Me
m
ory
Stru
ct
ure
,
”
Inte
rnational
Inf
orm
ati
on
Instit
ut
e
(
Tokyo
)
.
Information
,
vo
l. 20,
no.
1B,
pp.
543
-
548,
2017
.
[17]
D.
H.
Song
,
et
a
l
.
,
“
Act
ivi
t
y
ce
n
te
red
d
esign
of
sm
art
phone
user
int
erf
ace:
L
ea
r
ning
app
ex
ec
ut
i
on
pat
t
ern
s
with
neur
al ne
twork
m
odel
,
”
Inte
rnat
ional
Journal
of
Smar
t
Hom
e
,
vol
.
8
,
no
.
2
,
pp
.
10
1
-
106,
2014
.
[18]
D.
Chung
and
T
.
L
ee,
“
On
fuz
z
y
associative
m
e
m
ory
with
m
ult
i
ple
-
rul
e
storag
e
ca
pa
ci
t
y
,
”
IE
EE
Tr
ansacti
ons
on
Fuzzy
Syste
ms
,
v
ol.
4
,
pp
.
375
-
38
4,
1996
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
13
, N
o.
1
,
Ja
nu
a
ry 20
19
:
1
9
9
–
2
0
4
204
BIOGR
AP
HI
ES OF
A
UTH
ORS
Kw
ang
Bae
k
Kim
rec
ei
v
ed
his
M.S.
and
Ph.
D.
degr
e
es
from
the
Depa
rtm
ent
of
Com
puter
Scie
nc
e,
Pus
an
Nati
ona
l
Univer
sit
y
,
Busan
,
Kor
ea
,
in
1993
an
d
1999,
resp
ec
t
ivel
y
.
From
1997
to
the
pre
s
ent
,
he
is
a
profe
ss
or
a
t
the
Depa
r
tment
of
Com
pute
r
Engi
ne
eri
ng
,
S
il
la
Univer
si
t
y
,
Korea
.
He
is
cu
rre
ntly
an
associ
at
e
ed
it
or
for
Journal
of
Inte
l
li
g
enc
e
and
Inform
at
ion
S
y
st
ems
and
Th
e
Journal
of
Inform
at
ion
and
Com
m
unicati
on
Conv
erg
e
nce
Engi
ne
eri
ng
.
His
rese
arch
int
er
ests
include
fuz
z
y
cl
ust
eri
ng
and
fuz
z
y
cont
r
ol
s
y
st
em,
da
ta
m
ini
ng,
image
p
roc
essing,
and
bioi
nform
at
i
cs.
Doo
Heon
Song
rec
e
ive
d
his
B.
S.
degr
ee
in
St
atistic
s
&
Com
pute
r
Scie
n
ce
from
Seoul
Nati
ona
l
Univer
sit
y
,
Kor
e
a
and
M.S.
degr
ee
in
Com
pute
r
Scie
nc
e
from
th
e
Korea
Adv
anced
Instit
u
te
of
Scie
nc
e
and
Tec
hnolog
y
in
1983
.
He
rec
e
ive
d
hi
s
Ph.D.
Cert
ifi
c
at
e
in
Com
pute
r
Scie
nce
from
the
Univer
sit
y
o
f
Cal
ifornia
at
I
rvi
ne
in
1994.
He
has
bee
n
a
profe
ss
or
at
the
Depa
rtment
of
Com
pute
r
Gam
es,
SongD
am Col
le
ge
,
Korea
,
sinc
e
1997.
He
has
serve
d
as
an
associate
editor
for
Journal
of
Multi
m
edi
a
Signal
Proce
ss
ing
and
Inform
at
ion
Hiding
and
The
Journal
of
Inform
at
ion
and
Com
m
unic
at
ion
Converge
n
c
e
Engi
ne
eri
ng
.
His re
sea
rch
top
ics
inc
lude
ar
ti
fi
cial
intelligence,
vide
o
g
ame
desi
gn
& cul
tur
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.