TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 13, No. 2, Februa
ry 20
15, pp. 314 ~ 320
DOI: 10.115
9
1
/telkomni
ka.
v
13i2.700
1
314
Re
cei
v
ed
No
vem
ber 8, 20
14; Re
vised Janua
ry 2, 20
1
5
; Acce
pted Janua
ry 2
0
, 20
15
Filtering Based Illumination Normalization Techniques
for Face Recognition
Sasan Kara
mizadeh*
1
, Shahidan M abdullah
2
, Sey
e
d Mohammad Cher
ag
hi
3
,
Mazd
akZam
ani
4
1, 2,
4
Advanced
Informatics Schoo
l (AIS), Universiti T
e
knolo
g
i Mal
a
ysia, 54
100 Ku
al
a Lum
pur, Mala
ys
ia
3
Mala
ysia-J
ap
a
n
Internatio
na
l Institute of
T
e
chno
log
y
, U
n
ive
r
siti T
e
knologi
Mala
ysi
a
,
541
00 Ku
ala L
u
mpur, Mal
a
ysi
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: Ksasan2
@liv
e.utm.m
y
A
b
st
r
a
ct
T
he
mai
n
ch
all
eng
e ex
peri
e
n
c
ed by
the
pre
s
ent
face r
e
co
gniti
on tec
hni
q
ues a
nd s
m
oot
h filter
s
are th
e d
i
fficult
y in
man
agi
ng
illu
min
a
tion. Th
e d
i
fferenc
es i
n
face
i
m
a
ges
that are
cre
a
te
d by
il
lu
mi
natio
n
are n
o
rmal
l
y b
i
gg
er co
mp
are
d
to the d
i
ffere
nces i
n
inter-
p
e
rson th
at is u
t
ili
z
e
d t
o
differ
entiate
ide
n
titi
es.
How
e
ver, face
recog
n
iti
on
ov
er il
lu
mi
natio
n
has
mor
e
us
es
in
a l
o
t of a
p
p
licatio
ns th
at d
eal w
i
th s
ubj
ec
ts
that are not co
oper
ative w
her
e the hig
hest potenti
a
l of the
face rec
ogn
iti
on as a no
n-in
trusive bi
ometric
feature can
be
executed a
n
d
utili
z
e
d. A lot
of w
o
rk has been p
u
t in
to th
e researc
h
an
d deve
l
op
ment
of
illu
min
a
tion
an
d face rec
o
g
n
i
t
ion i
n
the
pre
s
ent era
an
d a
lot of critica
l
meth
ods
hav
e
bee
n intr
oduc
ed.
Neverth
e
less,
there ar
e so
me c
onc
erns
w
i
th face
rec
ogn
ition
an
d i
llu
mi
nati
on th
at requ
ire furt
he
r
consi
derati
ons
w
h
ich
incl
ud
e
the
defic
ienc
i
e
s i
n
co
mp
re
h
end
ing
the
su
b-spac
es i
n
ill
umin
ation
p
i
ctures,
prob
le
ms w
i
th intractabi
lity in face mod
e
ll
in
g and co
mplic
ate
d
mec
h
a
n
is
ms
of face surface
reflections.
Ke
y
w
ords
: ill
u
m
i
nati
on, face recog
n
itio
n,
techni
ques, strate
gies, filters
Copy
right
©
2015 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
Face
recognit
i
on usin
g different illumin
a
t
ion is
a critical challe
nge f
o
r appli
c
ation
s
use
d
in real time.
Many illumination filtering techni
ques are used and
introduc
ed
by researchers to
manag
e th
e i
s
sue
at h
and
. Neve
rthele
s
s, the
pr
esen
t
app
roa
c
h
e
s
are
b
o
th archaic
an
d do not
inclu
de th
e
cruci
a
l
analysi
s
of
pe
rform
ance of
illum
i
nation filte
r
in
g techniq
u
e
s
[1, 2]. F
a
ce
recognitio
n
tech
niqu
es
at pre
s
ent i
n
clude a
gr
o
u
p
of functio
nalities th
at ca
rry out t
h
e
norm
a
lization
of illuminati
on an
d thu
s
, handl
e the
main
challe
n
ge with th
e f
a
ce
re
co
gnition
s
y
s
t
em [3].
2. Filtering of Illumination
normalization Face recognition
We h
a
ve expl
ained
som
e
o
f
the filtering
of
illumination
norm
a
lizatio
n face
re
co
gn
ition in
this pap
er.
2.1. Single Scale Re
tinex
or (SSR) alg
o
rithm
Whe
n
the SSR scale lo
wers, it improves
the lo
cal cont
radi
cti
on and
offers a bette
r
robu
st comp
ression
rang
e. However, it has
som
e
drawb
a
cks su
ch as the the
halo artifa
ct. On
the cont
rary, whe
n
the SSR scale
rises, the feat
ures of the color
con
s
tan
c
y would imp
r
ove
at
the same tim
e
. Neve
rthele
ss, it i
s
n
o
t a
b
le to
cont
ra
ct the ro
bu
st scale
of an
im
age
well
eno
ugh
by not taking
into con
s
ide
r
ation the
ch
ara
c
teri
stics
of the image
[4], since the
image
s’ rati
o of
comp
re
ssed robu
st scale
s
varies [5].
Single-S
c
ale Retinex
The SSR form is defined
according to (1),
R(x
,
y
)
= lo
gI(x
,y
)-log(
F(x
.
y
)
*
I
(x
,y
)),
(1)
Whe
r
e R(x, y) is the Retinex output; I (x,
y)
is th
e image inte
nsity; ‘*’ reflects the
operation of convolution; a
nd F(x,
y) is a
Gaussia
n
function
ality.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Filtering Ba
se
d Illum
i
nation Norm
alizatio
n
Tech
niqu
es for Face
… (Sasan Ka
ram
i
zad
e
h)
315
F(x, y) =
K
•e-(x2+
y
2), integral (int
eg
ral (F
(x
, y
)
, dx
, Dy=1,
(2)
Figure1. Illust
ration o
r
iginal
images a
r
e i
n
the uppe
r rows, Images
pro
c
e
s
sed by
SSR - the
reflec
tion func
tionalities
are in the lower rows
2.2. Multi Scale Retin
ex
or (MSR) alg
orithm
The
algo
rithm
for
Multi S
c
al
e Retinex o
r
MSR
i
s
exten
ded f
r
om
the
algorith
m of
the SSR
as
su
gge
sted
[6].The th
eory of retin
e
x h
a
s th
e
assum
p
tion that
the
pe
rception
of
col
o
r is
strict
ly
depe
ndent o
n
the human
vision syste
m
’s neu
ral st
ructu
r
e. Ref
e
ren
c
e [7, 8] introdu
ced t
h
e
retinex m
ode
l for li
ghtne
ss
com
putatio
n. In o
r
de
r t
o
imp
r
ove th
e ima
g
e
s
with conditio
n
s
of
compl
e
x lighti
ngs, the
theo
ry of retin
e
x, whi
c
h
i
s
de
riv
ed from
the
system of hu
m
an visu
al, ha
s
been
utilized
for improvem
ents in im
age
contrastin
g. Referen
c
e [7,
9] first intro
d
u
ce
d the theo
ry
of retin
e
x. Used
La
nd’
s t
heory
to
cre
a
te the SS
R [8] and
MS
R [6]. In
gen
e
ral, th
e MS
R i
s
su
ccessful in
improving lo
cal c
ontra
stin
g and comp
ression of a robu
st rang
e. In rece
nt times,
resea
r
che
s
h
a
ve introdu
ce
d method
s to enha
nce t
he MSR ba
sed o
n
the co
rre
cti
on of colo
r [1
0],
natural im
pre
ssi
on
s [11], and halo effe
ct [12].
Figure 2. Show origi
nal ima
ges a
r
e in th
e
upper
ro
ws, Image
s
proce
s
sed by MSR – the
reflectio
n
fun
c
tionalitie
s are in the lowe
r rows
2.3. Homomorphic Filteri
ng-ba
sed
No
rmalization (HOM
O) Algo
rithm
Homo
morphi
c filtering o
r
HOM
O
is a p
opula
r
tech
ni
que for n
o
rm
alizatio
n wh
e
r
eby the
image i
nput i
s
firstly ch
an
ged into
the
necessa
ry
lo
garithm
and
after that into
the do
main
for
freque
ncy [2
0]. After that, the comp
onent
s of hi
gh freq
uen
cy are focused on a
n
d
the
comp
one
nts
of lower freq
uen
cy are d
e
crea
sed.
Fi
nally, the image is
chan
ged ba
ck to the
spatio
n dom
a
i
n utilizing th
e
inverse Fou
r
i
e
r tra
n
sfo
r
ma
tion and h
e
n
c
e retri
e
ving th
e expone
ntia
l
outcom
e
[21].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 314 – 320
316
Figure 3. Illustration ori
g
ina
l
images a
r
e i
n
the uppe
r rows, Images
pro
c
e
s
sed by
HOMO - the
reflectio
n
fun
c
tionalitie
s are in the lowe
r rows
The component of illumination is located clos
e to the cent
ral two-dim
e
nsional Fourier
that cha
nge
s
the image’
s l
ogarith
m
. Ho
wever, the
co
mpone
nt of reflectan
c
e i
s
situated fu
rth
e
r in
the outer
re
gion of the
Fouri
e
r
spe
c
t
r
um’
s
tw
o di
mensi
on. In
gene
ral, the
illumination
a
n
d
comp
one
nt of reflectan
c
e
are not di
re
ctly s
eparate
d
in Fourie
r’s
spa
c
e [14]. An approprai
ate
divide for the
comp
one
nts i
s
d
epe
nde
nt
on the
suit
abl
e pa
ram
e
ters of filter
of the
homo
m
orphi
c.
Normally, the parameters are
adjusted accordingly by utiliz
i
ng
the experiences of
the
researcher. Adelmann [14] in his research utiliz
ed homomorphi
c filtering
with various parameter
setting
s and
the most
suitable setting
wa
s ch
osen
f
o
r mo
re p
r
o
c
essing
wo
rks. Refere
nce [15]
cho
s
e
a
static g
r
ou
p of va
lues for th
e h
o
momo
rphi
c
para
m
eters
a
c
cordi
ng to
th
eir exp
e
rie
n
ces
in utilizi
ng the algorithm
f
o
r l
o
oking for the
homom
orphi
c filteri
ng.
According to [16], there
are
variou
s value
s
for thi
s
kind
of filter wh
ere they
have t
e
sted th
e pa
rameters a
nd t
he mo
st suita
b
le
is sele
cted. Method
s
for sele
cting pa
rameters
in
al
l the techniq
ues
dep
end
on thei
r data
s
et.
Neverth
e
le
ss,
[17] sugge
st
ed that a sel
e
cti
on meth
od
for ch
oo
sing t
he hom
omorphic p
a
ra
met
e
r
sho
u
ld
be m
a
de a
c
cording
to the gl
obal
contrast
facto
r
s (G
CF
) [1
8], whi
c
h i
s
not
relia
nt o
n
a
n
y
databa
se
s.
2.4. Isotropi
c Diffusion-b
ased Normal
ization
Algor
ithm
The al
gorithm known as the isotropi
c diffu
sion-based normali
z
at
ion
or IS
utilizes
isotro
pic
sm
o
o
thenin
g
of th
e image
to ca
lculate th
e fu
nction
ality of the lumina
nce. It reflects t
h
e
simplified
r-variant
of the
techni
que
of
anisotr
opi
c
di
ffusion-ba
sed
normali
zatio
n
a
s
sug
g
e
s
ted
by Brajovic a
nd Gro
s
s [19]
.
The al
gorithm for
anisotropic diffusion (AD)
i
s
popular for the feature of illumination
invariant
rem
o
val of the fa
ce im
age. Th
e functio
n
of t
he alg
o
rithm f
o
r the
ani
sotropic
diffusio
n
is
depe
ndent o
n
the con
d
u
c
tion function
a
lity and measur
e of disco
n
tinuity [20].
Neverth
e
le
ss,
in
the conve
n
tio
nal algo
rithm
of the anisot
r
opi
c di
ffusio
n
, the discont
inuity measu
r
ement no
rma
l
ly
take
s on the i
n
-ho
m
og
eneit
y
or spa
c
e g
r
adient [21, 35
].
In the model
kno
w
n
as th
e
Lambe
rtian
convex
surfa
c
e, the fa
ce i
m
age i
s
reflected a
s
reveale
d
belo
w
:
I(
x,
y)
=
R
(
x,
y)
L(
x,
y)
.
(
3
)
Acco
rdi
ng to
(3
),
R(x, y) is u
s
ually t
he
re
fle
c
tion
of the
sce
n
e
an
d L
(x,
y) u
s
ually
represent
s the illumination. Thus, the normali
zation
of illumination for veri
fying the face can be
achi
eved by.
Cal
c
ulation t
he illumi
nation L(x, y).
The L(x,
y) cannot be calcul
ated from the I(x, y)
because of t
he ill- posed
position.
It is comm
only assumed that
R(x,
y) differs
quicker than
the
L(x, y). The
r
efore, in m
o
st of the app
roca
he
s, R(
x, y) is
retrieve
d by co
nsi
d
e
r
ing th
e varia
n
ce
betwe
en the i
m
age
s’s I(x, y) logarithm
and its sm
oot
henin
g
versi
o
n that is the approximatio
n of
L(x, y). The logarith
m
ic fu
nction
ality gets rid of
the n
o
ise
s
an
d ca
use
s
the me
asu
r
em
ent to be
easily do
ne.
This cl
assifi
cation i
s
kno
w
n a
s
the g
eneri
c
q
uotie
nt image. Th
e fundam
ent
a
l
diagram of the method for
gene
ric q
uot
i
ent image is
revealed in Fi
gure 4.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Filtering Ba
se
d Illum
i
nation Norm
alizatio
n
Tech
niqu
es for Face
… (Sasan Ka
ram
i
zad
eh)
317
Figure 4. Basic diag
ram of
gene
ric q
uoti
ent image me
thod
The mai
n
hi
ghlight of g
eneri
c
q
uotie
nt image i
s
on the m
e
thod to
cal
c
u
l
ate the
illumination L
(x, y) using i
m
age I (x, y).
Figure 5. Illustration ori
g
ina
l
images a
r
e i
n
the uppe
r rows, Images
pro
c
e
s
sed by
Isotropi
c
Diffusio
n
-b
ased - the refle
c
tion function
a
lities are in th
e lowe
r ro
ws.
2.5. Adap
tiv
e
Non
-
Loc
al Means
Bas
e
d
Techniqu
e
for Normalization
The te
chni
qu
e kno
w
n
as the Aad
aptive
No
n-L
o
cal
Mean
s b
a
se
d no
rmali
z
ati
on
(ANL
)
wa
s intro
d
u
c
ed by
ˇ
Stru
c an
d Pave
ˇ
si
ˇ
c [22, 34].
This te
chni
que utilizes t
he algo
rithm
for
adaptive no
n-local me
an
s
de-n
o
isi
ng to
measure t
he function
ality of luminan
ce
and to cal
c
ul
ate
the refle
c
tan
c
e after that. T
he Adaptive
Non
Lo
cal M
ean
s Filter
wi
th a Mixture
of Wavelet i
s
just
like the NLM filtering ho
wev
e
r, the smo
o
thenin
g
param
eter is ad
apte
d
locally a
s
e
x
presse
d in:
(
4
)
Whe
r
eby th
e
distan
ce
is
m
easure
d
fro
m
a vo
lum
e
R
that is in
putte
d a
s
the
su
btractio
n
of the origin
al noisy volu
me u and th
e low pa
ss filtered volum
e
(
)
. It was di
scovere
d
experim
entall
y
that the distan
ce
minim
a
lly requi
red
in this situ
ation is a
bout t
he sa
me a
s
s2
becau
se of th
e rem
o
ved in
formation
of low fr
equ
ency
and the
ap
plying of the mi
nimal op
erato
r
[23, 24].
This ap
proa
ch is simpl
e
a
nd ha
s two critical
advanta
ges. It lets the same p
a
tch
e
s with a
simila
r
stru
ct
ure
to b
e
di
scovered
b
u
t wi
th a va
ried
m
ean l
e
vel
co
mpen
sating
f
o
r th
e inte
nsit
y in
the homog
en
eity that normally exists in the
data [25, 26]. However, the n
o
ise vari
an
ce’s
overe
s
timatio
n
will
be mi
nimal in
situ
ations t
hat h
a
ve with
uni
que p
a
tch
e
s in the
se
arch
databa
se [34
]. Therefore, the adapt
ive
filtering will e
s
tabli
s
h the
para
m
eter h
2
similar to th
e
minimal di
sta
n
ce
cal
c
ulatio
n as ref
e
rred
to in Equation
(4).
2
min(d(R
i
, R
j
))
ia
ndR
(
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 314 – 320
318
Figure 6: Illustration ori
g
ina
l
images a
r
e i
n
the uppe
r rows, Images
pro
c
e
s
sed by
ANL- the
reflec
tion func
tionalities
are in the lower row
3. Comparis
ons
Table 1 demonst
r
ates
pros and cons of illu
mination filters.
In Table 1 different
techni
que
s are comp
ared i
n
term of perf
o
rma
n
ce and
accuracy.
Table 1. Advantage a
nd di
sadva
n
taag
e of
illumination
filters in face
Recognitio
n
ALG
O
RITHM
ADVANTAG
E
DISADVANTA
GE
REFERENCE
SSR
1. The scale of a
n
SSR increases, its
color reliability
f
e
atures improve.
1.Includes halo artifacts
2. SSRs
have differentd
yn
amic
range compression. Characteristics
according tothe
scale.
3.The SSR do
es not rungo
od tona
l
execution
[5,27]
MSR
1. Works
effectively
w
i
thimages that
are gra
y
scale.
1.Histogram equ
alization isut
ilize
d
to improve the co
lorof the images.
This ma
y
r
esult i
n
a change in
the color scale causing the
artifacts and having
animbalance in
the color ofthe i
m
ages.
[28,28]
HOM
O
1.The feat
ures cause the linkw
ith
the
low
f
r
equenc
y of
the image
w
i
th
illumination and the
high frequenc
w
i
t
h
reflection.
1. Discrete feature in attractive
aspects bet
w
eencontiguo
us
discrete scalest
hat ma
y
n
ot be
found at theo
utp
ut.
[16,31]
I
S
O
T
RO
PI
C
1.Is able to prese
r
ve edges ofimag
e
and is able to red
u
cenoise at the same
time.
1.It is insensit
ive
toorientation and
sy
mm
etric,resulting in blurred
edges.
[21,32]
Adaptive
non-local
1. The applications that
includesegmentation, Relaxomet
r
y,
ortractog
raph
y m
i
ght make useof the
improved data th
at iscreatedafter
appl
y
i
ng
theproposed
filte
r
ing.
1. This technique does notreduce
the difficulty
of th
ealgorithm
significantly
while onl
y
decreasing slightly
t
he
accurac
y
offiltering.
[26,33]
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Filtering Ba
se
d Illum
i
nation Norm
alizatio
n
Tech
niqu
es for Face
… (Sasan Ka
ram
i
zad
eh)
319
4. Conclusio
n
Given the
cri
t
ical challen
g
e
in face re
c
ognition, researche
r
s hav
e used a
n
e
x
tensive
manne
r of illumination va
riation
s
in the work
s of pattern reco
gnition and
comp
uter vision.
Several te
ch
nique
s
were
portrayed th
at tolera
te
d
and/or comp
ensated the
variation
s
in
the
image cau
s
e
d
by chang
es of illumination. Neverthel
e
ss, gai
ning ill
umination in f
a
ce recogniti
on
is still
a hu
ge
chall
enge
tha
t
need
s
conti
nuou
s effo
rt
a
nd attention.
This
study h
a
d
ca
rri
ed o
u
t
an
extensive survey and inclu
ded techniq
u
e
s that have rece
ntly introd
uce
d
.
Referen
ces
[1]
Manso
o
r, Naf
e
es, AKM Muz
a
hid
u
l Isl
a
m, a
n
d
M. Ab
dur
Ra
zzak.
Multi
p
le
descri
p
tion
di
sc
r
e
te
c
o
si
ne
transform-
bas
e
d
imag
e cod
i
ng
using D
C
coef
fici
ent rel
o
catio
n
and AC c
oefficient i
n
terpo
l
at
ion
.
Jour
na
l
of Electronic I
m
a
g
i
ng.
20
13; 22(2): 02
30
30-
023
03
0.
[2]
X T
an, B T
r
ig
gs. Enh
anc
ed
loc
a
l te
xt
ure
feature
se
ts fo
r fa
ce re
cogn
i
t
i
o
n
un
de
r diffi
cu
l
t
l
i
g
h
t
i
ng
cond
itions. Ima
ge Process
i
ng,
IEEE Transactions
o
n
. 201
0; 19(6): 16
35-
16
50.
[
3
]
S
K
a
r
a
m
i
z
a
d
e
h
,
S
M
A
b
d
u
l
l
a
h
,
M
Z
a
m
a
n
i
.
A
n
O
v
e
r
v
i
e
w
o
f
H
o
l
i
s
t
i
c
F
a
c
e
R
e
c
o
g
n
i
t
i
o
n
.
IJRCCT.
201
3; 2(9): 738
-741.
[4]
Z
U
Rahma
n
, DJ Jobson, GA W
oode
ll. Multi-
scale reti
ne
x fo
r color ima
ge e
nha
nceme
n
t. 1003-
100
6.
[5]
CY Jang, J H
y
un, S Cho, HS
Kim, YH Kim.
Adaptiv
e Sel
e
c
t
ion of W
e
ig
hts in Multi-sca
le
Retin
e
x usin
g
Illum
i
nation and Object Edges
. IPCV. 2012.
[6]
DJ Jobson, ZU Rahman, GA
Woodell. A multisca
ler
e
tinex
for bridging th
e gap bet
w
een color images
and the
huma
n
observati
on of
scenes
. Im
age Processing, IEEE Transactions
on. 19
97; 6(
7): 965-9
76.
[7]
JA F
r
ankle, JJ McCan
n
. Method an
d ap
parat
us
for lightn
e
ss
imagi
ng. Goog
le Patents. 19
8
3
.
[8]
F
Ciurea, B F
u
nt, “T
uning reti
ne
x par
ameter
s,” Journal
of Electron
ic Imagi
ng, vo
l. 13, n
o
. 1, pp. 58-6
4
,
200
4.
[9] IR
T
e
rol-Villal
obos.
Mu
ltisca
le i
m
a
ge
enh
ance
m
ent an
d
seg
m
e
n
tation
base
d
on
morph
o
lo
gic
a
l
conn
ected c
o
n
t
rast map
p
i
ngs
. MICAI 2004: Advanc
es in A
r
tificial Intel
l
i
g
e
n
ce. Spri
nger.
200
4; 66
2
-
671.
[10]
DH Cho
i
, IH Jang, MH Kim, NC Kim.
Color
ima
ge e
nha
nc
ement bas
ed o
n
sing
le-sca
le
retinex w
i
th a
JND-bas
ed n
o
n
lin
ear filter
. Illumination.
20
07
; 1
:
1
.
[11]
M Herscovitz, O Yadid-Pecht.
A modifi
ed
Multi Scal
e R
e
tinex a
l
g
o
rith
m w
i
th an i
m
prove
d
glo
b
a
l
impressi
on
of b
r
ightn
e
ss for w
i
de dyn
a
mic
rang
e pictur
es
. Machin
e Visi
on an
d Ap
plic
ations. 2
004
;
15(4): 22
0-2
2
8
.
[12]
B Sun, W
Chen, H Li, W
T
ao, J Li.
Modified l
u
min
ance b
a
se
d ada
ptive MS
R
. 116-1
20.
[1
3
]
H
G
Ad
el
ma
nn
.
Butterw
orth e
quati
ons
for h
o
m
o
m
orp
h
ic fi
lterin
g of
i
m
a
g
e
s
. Comp
uters i
n
Bi
olo
g
y
an
d
Medici
ne. 19
98
; 28(2): 169-1
8
1
.
[14]
K Delac, M Grgic, T Kos.
Sub-image
ho
mo
morph
i
c filterin
g techn
i
qu
e for i
m
pr
ovin
g facia
l
identific
atio
n
und
er difficult il
lu
min
a
tio
n
con
d
itio
ns
. 21-23.
[15]
CN F
an, F
Y
Z
hang.
Ho
momorp
hic filter
ing b
a
se
d ill
u
m
i
nati
on nor
mali
z
a
ti
on
meth
od for fac
e
recog
n
itio
n
. Pa
ttern Recog
n
iti
on Letters. 20
1
1
; 32(10): 1
468
-147
9.
[16]
K Baek, Y Cha
ng, D Kim, Y Ki
m, B Lee, H C
hun
g, Y Han,
H Hah
n
.
F
a
ce regi
on d
e
tectio
n usin
g DC
T
and h
o
m
o
m
orp
h
ic filter
. 7-12.
[17] K
Matkovi
ć
, L
Neum
ann, A N
euma
nn, T
Psi
k
, W
Purgathofer.
Global c
ont
rast factor-a n
e
w
appro
a
c
h
to ima
ge co
ntrast
. 159-16
7.
[18]
R Gross, V Br
ajovic.
An
i
m
a
ge
pre
p
rocess
i
ng
alg
o
rith
m fo
r ill
u
m
in
atio
n
in
varia
n
t face
re
cogn
ition.
10-
18.
[19]
H W
ang, SZ
Li, Y W
ang.
F
a
ce recog
n
itio
n u
nder vary
ing l
i
ghtin
g
con
d
itio
ns usin
g self q
uotie
nt imag
e
.
819-
824.
[20]
W
Li,
T
Kuang, W
Gong.
Anisotrop
i
c d
i
ffusion a
l
g
o
rit
h
m
base
d
on
w
eber local
descri
p
tor for
illu
min
a
tion i
n
v
a
ria
n
t face veri
fication
. Mach
i
ne Visi
on a
nd
Appl
icatio
ns. 2
014; 25(
4): 997
-100
6.
[21]
V Štruc, N Paveši
ć
. Illumi
nati
on inv
a
ria
n
t face recog
n
itio
n
b
y
n
on-l
o
cal sm
oothi
ng: Spri
ng
er. 2009.
[22]
Karamiz
ade
h, Sasan, Sh
ahi
d
an M Abd
u
ll
ah
, Maz
dak Z
a
mani, Atab
akKh
erikh
ah. Patter
n
Reco
gniti
o
n
T
e
chniques:
Studies o
n
Appro
p
riate C
l
a
ssifica
tio
n
s. Advanc
ed Com
puter an
d Co
mmunicati
o
n
Engi
neer
in
g T
e
chno
log
y
.
Spri
nger Intern
atio
nal Pu
blis
hi
ng
. 201
5; 791-
799.
[23]
JV Manj
ón, P
Cou
pé, L M
a
rtí
‐
Bonmatí, DL
Collins, M
Robles. Adaptive
non
‐
l
o
cal
mea
n
s
den
ois
i
ng
of
MR ima
ges
w
i
t
h
sp
atial
l
y
var
y
ing
no
ise
lev
e
l
s
.
Journ
a
l
of M
agn
etic R
e
so
n
ance
Imagi
ng
. 201
0;
3
1
(1):
192-
203.
[24]
N W
i
est-Daes
slé, S Prima, P Coup
é, SP Morrisse
y, C
Barill
ot. Ricia
n
noise r
e
mova
l b
y
no
n-loc
a
l
means filteri
n
g
for lo
w
si
gna
l
-
to-nois
e
ratio MRI:
applicati
o
ns to DT
-MRI.
Medical Imag
e Comp
uti
n
g
and C
o
mput
er-Assisted Interv
entio
n–MICCAI
.
Springer.
20
0
8
: 171-1
79.
[25]
L Yan
g
, R Parton, G Ball, Z
Qiu, AH Green
a
w
a
y
,
I Dav
i
s, W
Lu. An ada
p
t
ive non-
loc
a
l
means filter for
den
oisi
ng l
i
ve-
c
ell im
ages
an
d improv
ing
pa
rticle detecti
on.
Journa
l of structural bi
olo
g
y
. 201
0; 172(
3)
:
233-
243.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 13, No. 2, Februa
ry 2015 : 314 – 320
320
[26]
L W
u
, P Z
hou, X Xu.
An Illu
mi
nati
on Invari
ant F
a
ce Reco
gniti
on Sche
me to Combi
n
in
g Nor
m
al
i
z
e
d
Structural Des
c
riptor w
i
th Sin
g
le Sca
l
e Reti
n
e
x
. Biometric R
e
cog
n
itio
n. Spr
i
ng
er. 201
3; 34
-42.
[27]
Q Meng, D Bian, M G
uo, F Lu, D Liu. Improve
d
multi
-
scale
retin
e
x
alg
o
rithm for medic
a
l ima
g
e
enh
anc
ement. Information En
gin
eeri
ng a
nd
Appl
icatio
ns.
Sprin
ger
. 20
12; 930-
937.
[28]
L Yong, YP
Xia
n
.
Multi-di
mens
ion
a
l multi-sc
al
e imag
e en
han
cement al
gorit
hm
. V3-
165-V
3
-168.
[29]
Manso
o
r, Nafees, F
a
touNd
i
a
y
e, S
hah
in A
k
ter Cho
w
d
hur
y, M Abdur R
a
zzak.
Multip
l
e
descri
p
tio
n
imag
e co
din
g
: A low
co
mpl
e
x
i
ty appr
oac
h fo
r lossy n
e
tw
orks
.
T
E
NCON 2
008-
200
8 IEE
E
Reg
i
on
1
0
Confer
ence. IE
EE. 2008; 1-5.
[30]
AK Sao, B Yegna
nara
y
a
n
a
. On t
he use of phas
e of the F
ouri
e
r trans
for
m
for face recogn
ition
und
e
r
variati
ons in i
l
l
u
min
a
tion.
Sig
nal, i
m
a
ge a
n
d
video pr
ocessi
ng
. 201
0; 4(3): 353-
358
[31]
W GE,
Gj Li,
Yq CHENG, C X
U
E, M ZHU.
F
a
ce ima
ge il
lu
min
a
tio
n
pro
c
essin
g
base
d
on improve
d
Retin
e
x
. Opt. Precisio
n Eng. 2
010; 18(
4): 101
1-10
20.
[32]
T
T
a
sdizen. Pri
n
cip
a
l comp
on
ents for
non-l
o
cal mea
n
s ima
ge de
no
isin
g. 1728-
173
1.
[33]
Karamiz
ade
h,
Sasan, S
h
a
h
id
an M A
b
d
u
ll
ah
, M
ehran
Ha
li
mi, Jafar Sh
a
y
an, Mo
hamma
d jav
a
d
R
aj
abi
.
Advanta
ge a
n
d
Draw
back of Supp
or
t Vector Machi
ne F
uncti
ona
lity
.
[34]
Qiang, So
ng, Z
hang H
a
i-F
e
ng. Re
se
arch
on App
licati
o
n
of Sinterin
g Basicit
y
of Base
d on Vari
o
u
s
Intelli
gent A
l
g
o
rithms.
T
E
LK
OMNIKA Indo
nesi
an J
ourn
a
l
of Electric
al
Engi
neer
in
g.
2
014; 12(1
1
):
772
8-77
37.
[35]
Adha
ni, Gita, AgusBu
on
o, AkhmadF
a
q
ih.
Optimiza
tio
n
of Support Vect
or Regress
i
o
n
usin
g Genetic
Algorit
hm a
n
d
Particle
S
w
ar
m Optimizatio
n
for R
a
infa
ll
Predicti
on
in
Dr
y
Seas
on.
TE
L
K
OM
N
I
KA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2014; 1
2
(11):
791
2-79
19.
Evaluation Warning : The document was created with Spire.PDF for Python.