Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
4, N
o
. 4
,
A
ugu
st
2014
, pp
. 53
9
~
54
7
I
S
SN
: 208
8-8
7
0
8
5
39
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Gender Classi
fication Usi
n
g Hy
brid of Gabor Filters and
Binary F
e
atures of an Im
age
Mahb
oobeh Naz
a
rloo
1
, Ebr
a
him P
a
rcham*
2
, Rez
a
Ak
bar Po
ura
n
i
3
1
Ele
c
tri
cal
and
C
o
m
puter Engin
e
ering Dep
a
rtm
e
n
t
, Qa
zvin
Univer
s
i
t
y
T
e
hran
, Ir
an
2
Electrical
and Computer
Engin
eer
ing
De
partm
e
nt,
T
e
hra
n
S
c
ien
c
e
& R
e
s
earch
Univers
i
t
y
Tehr
an,
Iran
3
Ele
c
tri
cal
and
C
o
m
puter Engin
e
ering Dep
a
rtm
e
n
t
,
Tabr
iz S
c
ien
c
e
& R
e
s
ear
ch Uni
v
ers
i
t
y
Tabri
z
,
Ir
an
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
Ma
r 6, 2014
Rev
i
sed
Jun
11,
201
4
Accepte
d J
u
l
4, 2014
F
ace is
one of
t
h
e m
o
s
t
im
portant biom
etri
c of
hum
an and cont
ains
lots
of
us
eful inform
at
ion s
u
ch
as
gender,
ag
e,
r
ace
and
iden
ti
t
y
. Gend
er
classification is
ver
y
eas
y
for
hum
an but it considers a challenge for
computers. Gen
d
er cla
ssificatio
n
through f
a
ce images has r
e
cently
been
considered so much. Gender r
e
cognition ca
n be
useful in interaction between
hum
an and co
m
puter like
id
entif
ying
indiv
i
dual’s iden
tit
y.
It is a
l
so
applicable in TV networks in order to
stud
y
the rate of view
ers. Various
algorithm
s
hav
e
been d
e
s
i
gned
f
o
r this
is
s
u
e
and
ea
ch of
them
ha
s
unravel
ed
that
to some ex
tent. The last ob
tain
ed rate to
id
entif
y
gender
w
a
s through
arti
cle writ
ten
b
y
M
o
zaffari
who
obtained mean rate of 83% for
identif
ication
.
It is the proposed me
thod of th
e present stud
y
which has
brought identif
ic
ation rat
e
to 92.5. in this m
e
thod we draw out f
ace fe
atur
es
based on Gabor
filte
rs and loc
a
l
binar
y
pa
tterns
.
These fe
atur
es a
r
e resistan
t
against noise an
d the
y
sel
e
c
t
proper feat
ur
es agai
nst bottlene
ck of im
ages. In
order to obtain
a proper classification,
we use self-organized
map (SOM)
(ty
p
e of ar
tificial neur
al network).
This neur
al network finds the proper
weights
for each
gender with ver
y
li
ttl
e error
.
Obtain
ed res
u
lts
ar
e com
p
are
d
with existing d
a
tasets and ther
efore,
superiority
of th
e proposed method
would be
evid
en
t.
Keyword:
Gabor filter
Gende
r
classifi
cation
Geom
etric features
Self-orga
n
ized feature
m
a
p
Self
-o
rg
an
ized
map
Copyright ©
201
4 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Ebra
him
Parch
a
m
Electrical and
Com
puter E
ngi
neeri
n
g De
part
ment, Tehr
a
n
Science & Rese
arch Un
iv
ersity Teh
r
an
, Iran
Em
a
il: eb
rah
i
m
p
arch
am
@g
mail.co
m
1.
INTRODUCTION
One
of the im
portant iss
u
es
whic
h are
propose
d be
side identification sys
t
e
m
s based
on
face im
ages
is determ
ining indi
viduals’
gende
r
through thei
r
face i
m
ag
e. Until now,
propose
d
m
e
thods
to
re
cognize
indivi
duals
by
face im
age ha
ve
had di
ffe
re
nt algorithm
s
.
Most of t
h
e methods
propose
d
to rec
o
gnize
a face
are base
d on le
arni
ng process and
using seve
ral im
ages fo
r each indi
vidua
l
such as
neura
l
network and
SVM.
Using learni
ng-base
d
m
e
thods of rec
o
gnition re
quire
s serie
s
with seve
ral
diffe
re
nt i
m
ages for each individual
whi
c
h pre
p
a
r
i
n
g t
h
em
woul
d
be di
f
f
i
c
ul
t
.
Fr
om
ot
her si
de,
i
t
i
s
not
possi
bl
e t
o
pr
ovi
de several
i
m
ages from
one
by real applications such as
Vide
o Surveillance.
There
f
ore, t
oday, single-sample face recognition
m
e
t
hods
ha
ve
been
cha
n
ged
t
o
a
n
i
m
port
a
nt
researc
h
i
s
s
u
e.
Gend
er classificatio
n
is th
e mo
st i
m
p
o
r
tan
t
task
in
app
licatio
n
s
su
ch
as mo
n
itoring
,
b
u
sin
e
ss pro
f
ile
o
f
regu
latory issu
es an
d
so
on
. Th
e
p
r
esen
t article h
a
s p
r
op
osed
a
n
e
w m
e
t
h
od
to
esti
m
a
te
g
e
nd
er. In
ad
ditio
n,
it is a ne
w method to classi
fy ge
nder
bas
e
d
on ext
r
acting
s
w
itching features
a
n
d bi
nary patterns of
each
im
age and
bas
e
d on
weights of Sel
f
-orga
n
ized m
a
p. In
t
h
e propose
d
m
e
thod,
we use general
obtained faces
to
apply face
im
a
g
e of eac
h
gender a
v
e
r
agely.
Input im
age is com
p
ared
by
input im
age of
a m
a
n and a
wom
a
n.
Local appare
nt differe
nces a
r
e then
defi
ne
d. In order to increase accur
acy of
gende
r
classification, the
propose
d
syste
m
can extract the m
o
st
signifi
cant ra
dius
of
a face
whic
h is
in fact t
h
e
dis
tinguis
h
ing
fea
t
ures
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
53
9
–
54
7
54
0
of eac
h
face ba
sed
on LB
P and Ga
bor
features and by se
gm
enting face
based
on t
r
iangular m
e
thod a
n
d it can
obt
ai
n a
bet
t
e
r
resul
t
i
n
c
o
m
p
are
wi
t
h
ot
he
r segm
ent
a
t
i
o
n
s
(S
qua
re
or ci
rcul
ar
).
Aft
e
r e
x
t
r
act
i
n
g bi
nar
y
and
Gab
o
r
feat
u
r
es
and
wi
t
h
c
o
n
s
i
d
er t
o
t
h
e di
ff
erences
bet
w
e
e
n ge
ne
ral
im
ages
of eac
h ge
nde
r,
we
pr
oce
e
d t
o
cl
ust
e
r t
h
em
.
Thi
s
cl
ust
e
ri
ng
i
s
cond
uct
e
d base
d o
n
sel
f
-
o
r
g
a
n
i
zed m
a
p and fi
nal
l
y
, g
e
nde
r i
s
rec
o
g
n
i
z
e
d
b
a
sed
o
n
lo
catio
n of t
h
e im
ag
e in
clusters.
2.
A REVIEW ON
T
H
E
PREVIOUS ATTE
MPTS
Vari
ous m
e
thods ha
ve bee
n
propose
d
to classify gender
until
now.
In
addition, the
r
e
are several
m
e
t
hods
t
o
e
x
t
r
act
feat
ures
i
n
o
r
de
r t
o
cl
assi
fy
ge
n
d
er
so
t
h
at
t
h
ey
can
be
di
vi
de
d i
n
t
o
se
veral
a
p
pr
oac
h
es. I
n
t
h
e
p
r
ese
n
t
art
i
cl
e,
we ha
ve u
s
ed pr
o
p
o
s
ed
m
e
t
hod o
f
Dr.
M
o
zaf
fari
whi
c
h
i
s
base
d on
usi
n
g ge
neral
i
m
ages
o
f
m
a
n
an
d
wo
m
a
n
an
d
u
tilizin
g
av
erag
e
o
f
th
ese im
ag
es an
d
ap
p
l
yin
g
Eu
clid
ean
d
i
stan
ce to
reco
gn
ize
gen
d
e
r
.
2.
1 Appe
ar
an
ce-b
a
sed Ap
p
r
oac
hes
Features a
r
e e
x
tracted
from
gene
ral data
of a face in
a
ppe
arance
-ba
s
ed a
p
proaches
. At
first, in t
h
i
s
m
e
thod
we locate data of the im
age successively in a
row or colum
n
.
We the
n
use s
t
atistical
m
e
thod to
red
u
ce
di
m
e
nsi
o
n
an
d
dat
a
res
o
l
u
t
i
o
n.
Fo
r e
x
am
pl
e, usi
n
g
I
C
A, L
D
A,
PC
A a
n
d
… m
e
t
hods
ca
n
be
not
ed.
2.
2 Ge
ome
t
ri
c
Fea
t
ures
-b
as
ed A
ppr
oac
he
s
In t
h
ese a
p
proaches, face im
age is
divi
ded
into
diffe
re
nt a
r
eas s
u
c
h
as ey
es, m
outh, nos
e
and …
and
geom
etric features
of eac
h areas of the fac
e
such as
length of the
nose, distance between eyes and
… are
det
e
rm
i
n
ed. I
n
fact
, i
n
t
h
ese m
e
t
hods
, sha
p
e and l
o
cat
i
o
n
of fac
e
com
p
o
n
ent
s
s
u
ch as
m
out
h, ey
es, ey
ebr
o
ws
and nose
are
determined a
n
d
f
eature
’
s
vect
or is e
x
tracted fro
m
all areas of the
face.
2.
3
Model-b
a
s
e
d Appr
oac
he
s
In m
odel
-
base
d ap
p
r
oac
h
es, t
h
e ba
sed m
ode
l
uses f
r
om
dat
a
of
di
f
f
ere
n
t
c
o
m
pone
nt
s o
f
t
h
e face
. I
n
fact, these approache
s
aim
s
to cons
truct a model of hum
a
n face base
d on
face changes and features, the
r
efore
this m
odel is a
b
le to
receive
face c
h
anges.
From
thes
e
kinds
of m
e
thods,
Elastic Bunch Gra
p
h Model and
Act
i
v
e Ap
pea
r
ance
M
odel
ca
n be not
e
d
.
2.
4 H
y
bri
d
A
p
pro
a
ches
These a
p
pr
oac
h
es c
o
m
b
i
n
e t
w
o
o
r
m
o
re a
p
pr
oac
h
es m
e
nt
i
one
d a
b
ove
.
3.
THE PROPOSED
METHOD
The
base
of
our m
e
thod is
using si
ngle im
age of
a
n
individual’s
face a
nd
ge
nde
r clas
sification is
base
d on feat
ures
of
Gabor filter and LB
P utilizing
triangular se
gm
entation of the
face. The m
e
th
od
of
ex
tracting
featu
r
es is
b
a
sed
o
n
Gabo
r filter and
l
o
cal
b
i
nary feat
u
r
es and
classification cond
u
c
ted
b
a
sed
on
self-organize
d
map.
T
h
is neural
ne
t
w
ork c
o
ntains
50 i
n
ternal layer s
o
t
h
at accepts im
a
g
es
of
dim
e
ns
ions
64×
6
4
as i
n
pu
t
.
In
or
der t
o
t
e
st
t
h
i
s
m
e
t
hod, i
m
ages of
AR
dat
a
base h
a
ve bee
n
use
d
.
Di
agr
a
m
of g
e
nd
e
r
classification s
y
ste
m
is shown in Figure
1.
3.
1 E
x
tr
acti
n
g
Fea
t
ures
Algo
rith
m
o
f
g
e
nd
er classifi
catio
n
is b
a
sed
o
n
two
feat
u
r
es of
Gab
o
r filter an
d Local Bin
a
ry
Pattern
s.
3.
1.
1 Appl
yi
n
g
G
a
b
o
r Fi
l
t
e
r
After e
x
tracting features of a face and pre
p
roces
sing
stage of the face,
it is
turn to a
pply Ga
bor
coefficients
on area
of the
fac
e
. T
h
is
area
of
the face
is crea
ted ba
sed on
t
r
i
a
ng
ul
ar se
gm
ent
a
t
i
on
o
n
t
h
e
face
as Figu
re 2.
Gab
o
r
filter is as fo
llo
w. Log
a
rith
m
i
c wa
v
e
let o
f
Gabo
r
h
a
s b
e
en
k
nown as th
e m
o
st effectiv
e
m
e
t
hod t
o
se
g
m
ent
a t
e
xt
ure
so t
h
at
i
t
can easi
l
y
separat
e
dat
a
rel
a
t
e
d
t
o
t
h
e t
e
xt
ur
e
from
i
n
t
e
rm
edi
a
t
e
frequency
bands and apply it fo
r al
gorithm
of
segm
enting. As extractin
g
features to recognize a
face t
h
rough
Gabor filter co
n
t
ribu
tes to
go
od
resu
lts, tech
n
i
c
o
f
f
eature ex
traction
based
on
Gabor filter h
a
s been
u
s
ed
h
e
re. Relatio
n
1
sh
ows Gabo
r filter relation
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Gend
er
Cla
ssifica
tio
n
Usi
n
g Hyb
r
i
d
o
f
Gabo
r
Filters
a
n
d
Bin
a
r
y Fea
t
u
r
es o
f
an
Ima
g
e
(Eb
r
a
h
i
m Pa
rch
a
m)
54
1
Fi
gu
re
1.
B
l
oc
k
of
t
h
e
pr
o
p
o
s
ed
gen
d
e
r
cl
ass
i
fi
cat
i
on sy
st
e
m
di
agram
or u
s
i
n
g
f
r
om
ful
l
-
vi
ew i
m
age o
f
a face
Fi
gu
re
2.
Tri
a
n
gul
a
r
se
gm
ent
a
t
i
on
of
an
i
m
ag
e
W
x,
y,
θ,
λ,
φ,
σ,
γ
e
x
p
c
o
s
2
π
x
x
cos
θ
y
s
i
n
θ
y
x
sin
θ
y
cos
θ
(1
)
(x
,y
) are
spat
i
a
l
i
nde
x an
d
σ
,
λ
,
γ
,Ø
,m
are
param
e
t
e
rs
of wavel
e
t
.
B
y
t
h
e hel
p
of Wi
skot
t
,
we use sim
i
l
a
r
param
e
ters.
,
∑∑
.
(2
)
cos
.
2.
2
1
.
cos
.
2.
2
1
.
,
.
In
w
h
ich:
√
0
1
(3
)
In
o
r
d
e
r to
extract featu
r
es
o
f
a set
o
f
i
m
ag
es, Gabo
r filters with
5 sp
atial freq
u
e
n
c
ies and
8
o
r
ien
t
atio
n
s
of sep
a
rate ang
l
e is u
s
ed
wh
ich
create 4
0
d
i
fferen
t
Gab
o
r filters. It h
a
s sh
own
in
fi
g
u
re 3
.
Th
ese
feature
s
are
s
p
atially im
p
l
e
m
ented
on
face i
m
ages.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
53
9
–
54
7
54
2
Fig
u
re 3
.
Gabor
filter with
5
sp
atial
frequ
en
cies and
7
sep
a
rate o
r
ien
t
atio
n
Fo
r inpu
t im
ag
e of fi
g
u
re
4
,
Gab
o
r
filter is i
m
p
l
e
m
en
ted
as th
e
fo
llowing
stag
es:
Fig
u
re
4
.
In
pu
t
i
m
ag
e to
Gab
o
r filter
Gabor co
efficien
t with d
i
fferen
t
ang
l
e
o
f
ro
t
a
tio
n
is as th
e fo
llo
wi
n
g
fi
g
u
re:
Fig
u
re 5
.
Gabor
filter with
ro
tatio
n
an
g
l
es
of 1
4
4
,
0, 3
6
, 7
2
, 1
0
8
d
e
g
r
ees
resp
ectiv
ely
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Gend
er
Cla
ssifica
tio
n
Usi
n
g Hyb
r
i
d
o
f
Gabo
r
Filters
a
n
d
Bin
a
r
y Fea
t
u
r
es o
f
an
Ima
g
e
(Eb
r
a
h
i
m Pa
rch
a
m)
54
3
After creating
Gabor co
efficien
t and
m
u
ltip
lyin
g
b
y
inp
u
t
i
m
ag
e, fi
g
u
re
6
is ob
tain
ed
:
Fig
u
re 6
.
Ou
tpu
t
ob
tain
ed
from
i
m
p
l
e
m
en
tin
g
Gabor
filter with
d
i
fferen
t
an
g
l
es
on
i
n
pu
t
i
m
ag
e
As it could
be
seen in the
figure 6, these
fi
gures s
h
ow
good
sam
p
les of a fa
ce and ot
her shapes
of t
h
e
face so that it enables
us to take di
ffere
n
t texture from
diffe
rent angles
and e
x
tract loc
a
tions of eyes,
m
outh
and
fore
head.
Propose
d m
e
th
od
of t
h
e prese
n
t article is
used to e
x
tract the
related weight
s of eac
h face i
m
age
th
ro
ugh
g
e
n
e
ral av
erag
e
o
f
Gab
o
r
filter of each
ind
i
v
i
d
u
al.
3.
1.
2 L
o
cal
B
i
nar
y
P
a
t
t
erns
LB
P ope
rat
o
r t
a
gs l
a
bel
t
o
pi
xel
s
of a
n
i
m
age t
h
r
o
ug
h t
h
r
e
sh
ol
di
n
g
a ne
i
g
h
b
o
r
ho
o
d
of
3×3 i
n
eac
h
pi
xel
s
o
t
h
at
i
t
i
s
res
u
l
t
e
d i
n
a
bi
na
ry
n
u
m
b
er.
LB
P
feat
ure
h
a
s bee
n
fre
q
u
e
n
t
l
y
use
d
t
o
cl
assi
fy
ge
n
d
er.
1
0
(4
)
=
∑
2
7
0
(5
)
In
relatio
ns
4
an
d 5, f_
c is th
e
v
a
lu
e
of p
i
x
e
l cen
ter an
d
〖
f
〗
_p
i
s
t
h
e
val
u
e
o
f
ce
nt
er pi
xel
f
o
r
a
nei
g
hb
o
r
h
o
o
d
.
The
val
u
e
of
LB
P i
n
cent
e
r
pi
xel
i
s
o
b
t
a
i
n
ed
fr
om
rel
a
t
i
on
5
.
Sy
m
bol
of
〖
LB
P
〗
_(
P,
R
)
^
u
i
s
use
d
f
o
r
uni
fo
r
m
LB
P oper
a
t
o
r
whi
c
h u
s
ag
e of LB
P
o
p
er
at
or f
o
r nei
g
h
b
o
r
h
oo
d o
f
P
p
o
i
n
t
s
i
s
sam
p
l
e
d o
n
a
circle of
ra
diu
s
R.
Fi
gu
re
7.
Usa
g
e o
f
LB
P
o
p
era
t
or i
n
nei
g
hb
or
ho
o
d
(1
,
8)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
53
9
–
54
7
54
4
Fi
gu
re
8.
Im
pl
em
ent
i
ng LB
P
ope
rat
o
r
on
fac
e
im
ages wi
t
h
param
e
t
e
rs R
=
1 &
P=
8
3.
2 Co
nstr
ucti
on of
Ge
neral
Face
I
m
a
g
e
As t
h
e
p
r
o
p
o
se
d al
g
o
ri
t
h
m
i
s
not
base
d
on t
r
ai
ni
ng
p
r
oc
ed
u
r
e, a
ge
neral
i
m
age i
s
use
d
t
o
rec
o
gni
ze
gen
d
e
r
of i
n
p
u
t
im
age. Ge
ner
a
l
im
age of
m
a
n a
n
d
w
o
m
a
n i
s
o
b
t
a
i
n
e
d
t
h
ro
ug
h l
i
n
ea
r c
o
m
b
i
n
at
i
o
n
of
i
m
ages
.
Th
e
fo
llowing
figu
re is a samp
le of
ge
neral
i
m
ages o
f
m
a
n and
w
o
m
a
n.
Fi
gu
re
9.
A
sa
m
p
l
e
of ge
ner
a
l
im
ages of
m
a
n a
n
d
w
o
m
a
n.
(a):
AR
dat
a
ba
se, (
b
):
Et
h
n
i
c
dat
a
base
3.
3 Se
gme
n
t
a
t
i
on
Segm
enting fa
ce im
age into several a
r
eas c
ont
ribute
s to
ex
tract d
i
stin
ctiv
e feat
ures
between m
e
n
and
w
o
m
e
n. B
a
sed o
n
fi
g
u
re
2, eac
h i
m
age of
dat
a
base
is seg
m
en
ted
in
8
section
s
so
t
h
at it h
a
s 9
sectio
ns
co
m
p
u
tin
g
wit
h
th
e m
a
in
i
m
a
g
e.
3.4 Determini
ng Features
of
Face
Al
l
of t
h
e feat
ures
of fac
e
d
o
n
o
t
hav
e
t
h
e
sam
e
im
port
a
nce i
n
ge
n
d
er
cl
assi
fi
cat
i
on.
In
or
der t
o
determ
ine i
m
p
o
rta
n
t features of f
ace, input i
m
age is
filte
red through
using Sobel and L
OG ope
rator.
Two
out
put im
ages have
diffe
r
ent i
n
form
ation obt
a
ined
by c
o
m
b
ining wa
velet.
Som
e
feat
ures
of
face like
eyes and
m
o
u
t
h
wh
ich is m
o
re im
p
o
r
tan
t
in g
e
nd
er
classificati
o
n
will b
e
app
a
ren
t
in con
s
idered
im
ag
es. Fi
gu
re 10
shows
detection of
face features.
Fi
gu
re
1
0
.
Det
ect
i
on
of
face
f
eat
ures
(a):
I
n
p
u
t
i
m
age. (b
):
I
m
pl
em
ent
i
ng L
o
g
act
o
r
.
(c):
I
m
pl
em
ent
i
ng S
obel
act
or.
(
d
):
C
o
m
b
i
n
i
n
g act
ors
o
f
L
o
g
an
d
So
be
l
3.
5 Com
p
u
t
i
n
g
Wei
g
h
t
s
Im
age entropy is used to det
e
rm
ine which
section
of the
propose
d
im
age that has endure
d feat
ure
ex
traction
h
a
s m
o
re
weigh
t
in
p
r
o
cess of gen
d
e
r rec
ogn
itio
n. Th
ose sect
io
n
s
with
h
i
gher en
trop
y, it
mean
s
havi
ng
m
o
re w
e
i
ght
, a
r
e c
o
nsi
d
ere
d
as
m
o
re i
m
port
a
nt
bl
ock
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
Gend
er
Cla
ssifica
tio
n
Usi
n
g Hyb
r
i
d
o
f
Gabo
r
Filters a
n
d
Bin
a
r
y
Fea
t
u
r
es
o
f
an
Ima
g
e
(Eb
r
a
h
i
m
Pa
rch
a
m)
54
5
3.
6 L
e
arni
n
g
Al
g
o
ri
thm
o
f
Sel
f
-
o
rg
ani
z
ed M
a
ps
Learni
ng al
go
ri
t
h
m
of sel
f-o
rga
n
i
zed m
a
ps i
s
a t
y
pe of u
n
s
u
per
v
i
s
ed l
ear
ni
n
g
.
B
a
si
cal
l
y
,
uns
u
p
er
vi
se
d l
earni
ng al
g
o
r
i
t
h
m
can be speci
fi
ed by
fi
rs
t
orde
r eq
uat
i
ons
. The
s
e eq
uat
i
o
n
s
desc
ri
be
h
o
w
n
e
two
r
k
weights b
eco
m
e
co
mp
atib
le to
tim
e
o
r
rep
e
tition
of d
i
screte m
o
d
e
. In
ord
e
r to
ad
ju
st
weigh
t
s,
scale
of si
m
i
l
a
ri
ty
or shari
ng
pat
t
e
r
n
i
s
use
d
t
o
co
nd
uct
l
earni
n
g
pr
ocess
whi
c
h
l
eads us t
o
s
o
m
e
correl
a
t
i
on
fo
rm
s,
clu
s
tering
o
r
co
m
p
etitiv
e b
e
hav
i
ors
o
f
t
h
e network. Gen
e
rally, learn
i
ng
alg
o
rith
m
o
f
self-o
rg
an
ized
m
a
p
s
are
base
d
on
sel
ect
i
ng
wi
n
n
i
n
g
ne
ur
o
n
a
n
d
m
ovem
e
nt
of
t
h
e m
e
nt
i
one
d
ne
ur
o
n
an
d s
o
m
e
of i
t
s nei
g
h
b
o
rs
t
o
war
d
th
e con
s
i
d
ered
in
pu
t d
a
ta.
3.
6.
1 E
a
rl
y
St
age
:
In th
is stag
e,
weigh
t
of each
n
e
uron is created
based on the
pre
v
ious
stage of weight
ext
r
acti
on
th
ro
ugh
non
-
l
i
n
ear
f
eatur
es
.
In
t
h
e prese
n
t
art
i
c
l
e
,
wei
g
ht
s
are
i
m
pl
em
ent
e
d base
d on separat
i
n
g
l
o
w
-
l
e
vel
feat
ure
s
an
d al
so ext
r
a
c
t
i
on
of a
d
j
u
st
e
d
Ga
bo
r an
d L
o
cal
B
i
nary
feat
ur
es and a
n
i
n
p
u
t
pat
t
e
rn
of
g
e
neral
i
m
ag
es to
th
e network.
3.
6.
2
S
p
eci
f
y
i
n
g the Wi
nni
n
g
Neur
on:
In t
h
i
s
st
a
g
e, w
i
nni
n
g
neu
r
o
n
i
s
speci
fi
ed
bas
e
d o
n
si
m
i
l
a
ri
ty
of net
w
o
r
k
.
Di
ffe
re
nt
sim
i
lari
t
y
cri
t
e
ri
a
can
b
e
app
lied in
self-o
rg
an
i
zed
m
a
ps but
t
h
e m
o
st
com
m
on c
r
i
t
e
ri
o
n
w
h
i
c
h i
s
a
ppl
i
e
d
i
n
t
h
ese
net
w
or
ks i
s
Eu
clid
ean
d
i
stan
ce.
Th
e relatio
n of Eu
clid
ean
sim
ilarit
y
is as fo
llo
w:
‖
‖
(6
)
No
w,
at
t
h
e
sa
m
e
t
i
m
e
, i
nput
i
s
com
p
are
d
wi
t
h
al
l
e
x
i
s
t
i
n
g el
em
ent
s
i
n
si
de t
h
e
net
w
or
k
.
W
i
nni
n
g
neu
r
on
i
s
a
neu
r
on
wi
t
h
t
h
e m
i
nim
u
m
di
st
ance am
ong
al
l
refe
rence
pat
t
erns
o
f
i
n
p
u
t
d
a
t
a
.
‖
‖
m
i
n
‖
‖
(7
)
So t
h
at m
c
i
s
the
wi
n
n
i
n
g
ne
ur
o
n
a
n
d
m
r
i
s
ref
e
re
nce
vect
ors
.
A sam
p
l
e
of
sel
ect
i
n
g
wi
nni
ng
neu
r
o
n
a
m
ong
refe
rence
patte
rns
is s
h
o
w
n i
n
fig
u
re
1
1
.
T
h
ese
win
n
in
g
n
e
uro
n
s
id
en
tify i
m
ag
es with
th
e sam
e
co
n
t
en
t and
adj
u
st
t
h
ei
r nei
g
h
b
o
ri
ng
ne
ur
o
n
s
i
n
or
der
t
o
o
b
t
a
i
n
bet
t
e
r res
u
l
t
s
.
Fi
gu
re 1
1
. Sel
ect
i
ng wi
n
n
i
n
g neu
r
on
am
ong refe
rence
pat
t
e
rns
i
n
s
e
l
f
-
o
r
g
a
n
i
zed
m
a
ps
3.
6.
3 Determi
n
i
n
g Nei
g
hb
or
i
n
g Neur
ons
Aft
e
r s
p
eci
fy
i
n
g wi
n
n
i
n
g ne
u
r
o
n
, a set
of ne
i
g
h
b
o
r
i
n
g ne
ur
ons
of t
h
e wi
n
n
i
n
g ne
ur
on
w
h
i
c
h sh
o
u
l
d
be c
h
an
ge
d i
n
val
u
e a
r
e
det
e
r
m
i
n
ed. C
h
an
gi
ng
rel
a
t
e
d
val
u
es t
o
nei
g
hb
o
r
i
n
g
ne
ur
o
n
s i
s
g
e
neral
l
y
d
o
n
e i
n
t
w
o
ways: In
th
e
first m
e
th
o
d
, a sp
ecified
n
e
ig
hbo
ri
n
g
rad
i
us is selected
aroun
d
t
h
e
winn
ing
n
e
uron
. In
th
is
m
e
t
hod, al
l
ne
ur
o
n
s o
f
t
h
e n
e
t
w
o
r
k
whi
c
h are i
n
t
h
e spec
i
f
i
e
d di
st
ance
of t
h
e
wi
n
n
i
n
g
neu
r
o
n
m
ove t
o
wa
r
d
in
pu
t with a co
n
s
tan
t
co
efficien
t.
In
t
h
e seco
nd
m
e
th
o
d
,
all ex
istin
g
n
e
u
r
on
s in th
e
network
m
o
v
e
to
ward
i
n
p
u
t
wi
t
h
une
qual
coe
ffi
ci
en
t
.
Thi
s
u
n
e
qual
coe
ffi
ci
ent
ha
s t
h
e
m
a
xim
u
m
val
u
e i
n
t
h
e
wi
n
n
i
n
g
ne
ur
o
n
a
n
d i
t
decrease
s
whe
n
recedes
fr
om
the wining ne
uron.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 4
,
N
o
. 4
,
Au
gu
st 2
014
:
53
9
–
54
7
54
6
3.
6.
4 Wei
g
h
t
s
Mo
di
fi
ca
ti
on
At
t
h
e end
,
w
e
i
ght
s rel
a
t
e
d t
o
t
h
e wi
n
n
i
n
g neu
r
on an
d i
t
s
nei
g
h
b
o
rs s
h
o
u
l
d
be m
odi
fi
ed base
d
o
n
in
pu
t of
n
e
twor
k. Th
ese ch
ang
e
s ar
e co
ndu
cted
b
a
sed on
t
h
e r
e
latio
n 8
:
1
.
(8
)
So t
h
at
x i
s
i
nput
vect
or i
n
t
i
m
e t
,
m
r
is r
th
referen
ce p
a
ttern
in
ti
m
e
t,
α
is rat
e
of l
earni
ng i
n
t
i
m
e
t
and
i
s
nei
g
h
b
o
r
h
oo
d
fu
nct
i
o
n
whi
c
h i
s
defi
ned
ba
sed
o
n
Ker
n
el
f
u
nct
i
o
n as
f
o
l
l
o
w:
e
x
p
‖
‖
2
2
2
(9
)
So t
h
at
r
e
p
r
esen
ts w
i
nn
ing
neu
r
on
an
d
is its n
e
ighb
or’s
referen
ce
p
a
ttern
and
σ
i
s
radi
us of
K
e
rn
el fu
n
c
ti
on
in
tim
e t. r
e
su
lt o
f
t
h
e above issu
es is m
o
d
i
f
i
catio
n
of
w
e
ig
h
t
s and
m
o
v
m
en
t o
f
th
e m
e
n
tio
n
e
d
neu
r
ons t
o
wa
r
d
t
r
ai
ni
ng sam
p
l
e
. It
i
s
a pa
r
a
m
e
t
e
r whi
c
h
i
s
appl
i
e
d i
n
o
r
de
r t
o
c
ont
rol
con
v
e
r
ge
nce
of t
h
e
algorithm
and it is
depe
nde
d
on re
petition. It is necessar
ily according to t
as a stead
y decrease. Un
supervised
t
r
ai
ni
n
g
i
s
ge
n
e
ral
l
y
and
i
n
e
v
i
t
a
bl
y
m
o
re com
p
l
i
cat
ed t
h
an s
u
per
v
i
s
ed
m
e
t
hod,
t
h
e
r
ef
ore
,
i
t
re
q
u
i
r
es
m
o
re
ti
m
e
to
learn
train
i
n
g
p
a
ttern
s.
3.7
Clustering & Ge
nder
Re
cognition
After ex
tracting
i
m
ag
es’ featu
r
es an
d
d
e
termin
in
g
weig
ht of the conside
r
ed im
age, we cluster inputs
of feat
ure e
x
t
r
act
i
on (Sel
f
-
o
r
g
ani
z
e
d
m
a
p) and t
h
e
n
deal
wi
t
h
rec
o
g
n
i
z
i
ng
gen
d
e
r
of t
h
e i
ndi
vi
d
u
al
b
a
sed o
n
lo
catio
n
of th
e
i
m
ag
e in
th
e d
e
termin
ed
clusters.
4.
Experimental results
In
order to e
v
a
l
uate the proposed m
e
thod,
we used
AR data
base related to
face im
ages. AR database
are com
posed
of 5
6
di
f
f
ere
n
t
wom
e
n and 7
0
m
e
n (t
ot
al
l
y
126
peo
p
l
e
).
W
i
t
h
rega
rd t
o
t
h
e
pro
p
o
se
d al
go
ri
t
h
m
and
t
h
e m
e
t
h
o
d
of
D
r
. M
o
zaf
fari
’s
art
i
c
l
e
,
re
sul
t
s
o
f
t
a
bl
e 1
sho
w
su
peri
ori
t
y
of t
h
e
pr
op
o
s
ed m
e
t
hod
.
Tabl
e
1.
Ge
ner
a
l
resul
t
s
of
co
m
p
ari
ng t
h
e
p
r
op
ose
d
m
e
t
hod
an
d
Dr
. M
o
za
f
f
ari
’
art
i
c
l
e
General
percent
Per
cent of wo
m
a
n
recognition
Per
cent of
m
a
n
recognition
Na
m
e
of
the
M
e
thod
83.
7
86.
50
80.
9
M
o
z
a
ffa
r
i
92.
5
92.
5
92.
5
T
h
e pr
oposed
m
e
thod
5.
CO
NCL
USI
O
N
Reco
gn
izing
gen
d
e
r is on
e of th
e im
p
o
r
t issu
es in
statistic an
d
p
s
ycho
l
o
g
y
wh
ich
is n
ecessary in
co
m
p
u
t
er system
s wh
en
su
rv
eyed
d
a
ta are increased. Th
e
presen
t article uses Gab
o
r
filter and
Lo
cal Bin
a
ry
Pattern
s to
ex
t
r
act d
i
stin
ctiv
e featur
es
of
face im
ag
es wh
i
c
h
are
n
o
t
sens
itiv
e to
n
o
i
se of th
e im
ag
es an
d are
also
ab
le to sep
a
rate
p
a
ttern
s with
h
i
g
h
d
i
fferen
ce. Th
is syste
m
co
n
t
ributes to
g
e
nd
er
reco
gn
itio
n thro
ugh
clu
s
tering
n
e
ural n
e
two
r
k
(self-org
an
ized
map
)
so
t
h
at th
e ob
tain
ed
resu
lts sho
w
t
h
e su
p
e
riority o
f
th
e
m
e
t
hod.
REFERE
NC
ES
[1]
C. Shan, S. Gong, and P.W
.
McOwan, “
L
ear
nin
g
gender from
hum
an gaits and faces”
,
IEEE Co
nf. on Advan
c
ed
Video
and Signa
l Based
Surveillance
, pp.505–51
0, 2007
.
[2]
V. Thom
as
, N.
V. Chawla, K
.
W
.
Bow
y
er
, and
P
.
J
.
F
l
y
n
n
,
“
L
e
a
rning to pr
edic
t
gender from
iris
im
ages
”,
IEEE Int.
onference on
Biometrics: Theo
r
y
, Applications, and
Systems
, pp. 1–5, 2007.
[3]
F. Hing C. Tiv
i
ve and A
.
Bo
uzerdoum, “A Shunting
Inhibitor
y
Convolutio
nal Neural Network for Gender
Cla
ssifica
tion”
,
The 18th In
t. Co
nference on
Pattern Recognition
, 2006.
[4]
M. Toews and T. Arbel, “Detection, lo
calization and sex classification of fa
ces fr
om arbitrar
y
viewpoints and under
occlus
ion
”
,
IEEE transaction on
PAMI
, Vol 31,
No.9, 2009
.
[5]
C. Ben Abdelk
ader, and
P. Griff
i
n, “
A Local
Reg
i
on-based Appro
a
ch to Gender C
l
assificat
ion Fro
m
Face Images
”,
Proc. of
the IE
E
E
Confer
ence on
Com
puter Visio
n
and Patt
ern Recognition
,
pp
.52
-
56 ,2005
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Gend
er
Cla
ssifica
tio
n
Usi
n
g Hyb
r
i
d
o
f
Gabo
r
Filters
a
n
d
Bin
a
r
y Fea
t
u
r
es o
f
an
Ima
g
e
(Eb
r
a
h
i
m Pa
rch
a
m)
54
7
[6]
Y. Fang, and
Z. Wang, “
Improving LBP f
e
ature
s
for gender cl
assification
”, Pr
oc. of th
e Int.
Conf on Wavelet
Analy
s
is
and Pattern R
ecognition
,
pp
. 373-377
. 2
008.
[7]
J. Hadadnia, M. Ahmadi, a
nd K. Faez, “An Efficien
t F
eatur
e Ex
traction Method
with Pseudo-Zernike Moment in
RBF
Neural Network-Based Hum
a
n F
ace Reco
gnition S
y
st
em
”,
EURASIP Jour
nal on Applied
Signal Processing
,
(9), pp
.890–901, 2003.
[8]
G. Amay
eh, G
.
Bebis, and
M
.
Nicoles
c
u
,
“
G
ender cl
as
s
i
fica
tio
n from
hand s
h
ape”
,
IEEE Co
mputer vision a
nd
pattern recognition workshop, C
VPR
, pp
.23-28,2
008.
[9]
E. Mak
i
nen
and
R. Raisamo, “Evaluati
on of g
e
n
d
er classificatio
n methods with
automatically
detected and
align
e
d
faces
”
,
In
IE
EE
Tran. on Pat
t
ern
Anal
ysis and M
a
chine Intellig
en
ce
, Vol 30(3)
, pp
. 541--547
, 2008
.
[10]
A
.
M
.
M
a
rt
iniz
,
R. Ben
a
ven
t
e,
“
T
he A
R
f
a
c
e
d
a
t
a
bas
e
”
,
Technical Report
, 24, C
V
C, 1998.
[11]
Saeed Mozaffar
i
, Ham
i
d Behravan, Roholl
a
h Akbari, “
G
ende
r Cl
assifica
tion using Single Frontal Im
age per Person:
Com
b
ination
of
Appearanc
e
and Geom
etric
b
a
s
e
d
F
eatur
es
”,
2010
Internat
ional C
onference on
Pa
ttern
Recogn
itio
n
,
pp.1192-1195.
Evaluation Warning : The document was created with Spire.PDF for Python.