TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 2743 ~ 2
7
5
2
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4368
2743
Re
cei
v
ed Au
g 19, 2013; Revi
sed O
c
t 23
, 2013; Accep
t
ed No
v 16, 2
013
An Ada
p
tive Detection Method of Multiple Faces
Wei Li
Chin
a W
e
st No
rmal Univ
ersit
y
, No. 1 Shida R
oad, Com
puter
School, Na
nch
ong, Ch
in
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: nos03
6@1
6
3
.
com
A
b
st
r
a
ct
T
he ap
pe
ara
n
c
e
of
mu
ltipl
e
fa
ces is i
n
flu
enc
ed
by
ab
nor
ma
l exp
o
sur
e
, int
e
rferin
g b
a
ckgr
oun
ds or
fake objects gr
eatly in the c
o
lor face image.
A m
u
lt
iple-face detection
m
e
t
hod
bas
ed on the adaptive dual
skin mode
l an
d
impr
oved fu
zzy C-me
an clust
e
rin
g
w
a
s
presented i
n
this study. F
i
rst an adaptiv
e skin-co
lor
mo
de
l an
d a
n
ada
ptive ski
n-
prob
abi
lity
mo
del w
e
re
bui
lt
to acq
u
ire t
he
skin lik
eli
h
o
od
for clusteri
ng, th
e
ada
ptive
in
itial
clusteri
ng c
ent
ers,
an
d th
e
ad
aptive
clust
e
rin
g
w
e
ig
hts. T
h
e
n
the
ski
n-lik
eli
hoo
d i
m
ag
e w
a
s
seg
m
e
n
ted dy
na
mic
a
lly by i
m
pr
ove
d
fu
zz
y
C-me
an cl
usterin
g
. F
i
nally t
he
multi
p
le-fac
e targets w
e
r
e
distin
guis
h
e
d
and
extracted
by jo
intly
usi
ng the
effe
cti
v
e are
a
s, circ
umfer
enc
es a
nd circ
ular
ities
of
conn
ected targ
ets. Experi
m
e
n
t
show
ed that the al
gor
ith
m
h
ad g
ood r
e
su
lt
s and h
i
gh s
p
e
ed, accur
a
cy, an
d
ada
ptab
ility of face detecti
on.
Ke
y
w
ords
: ski
n mo
de
l, face detectio
n
, fu
zzy clusterin
g
, connecte
d co
mp
one
nt lab
e
ll
ing
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
The fa
ce recognition i
s
a
n
active
subj
ect in the fiel
ds of
com
put
er visio
n
an
d
pattern
recognitio
n
, whi
c
h ha
s a
wide
rang
e o
f
potential ap
plicatio
ns. Th
e face recog
n
ition is an
a
c
tive
subj
ect i
n
th
e field
s
of
compute
r
visi
on a
nd
patte
rn
re
cog
n
itio
n, whi
c
h
ha
s a
wide
ra
ng
e of
potential ap
pl
ication
s
[1]. On the matte
r of sin
g
le-fa
c
e dete
c
tion
unde
r the co
ndition
s of si
mple
backg
rou
nd
and
unifo
rm
exposure,
th
e face
dete
c
tion m
e
thod
b
a
se
d o
n
skin
col
o
r mo
del
can
achi
eve g
ood
re
sult
s b
e
ca
use
of the
co
mpactn
ess of
skin
col
o
r cl
usteri
ng
dist
ri
bution [2
-5].
For
fast multiple-face dete
c
tio
n
unde
r co
m
p
lex bac
kg
ro
und
s, the sta
b
ility, accura
cy, adaptabili
ty,
anti-interference, and applicable
conditions of the existed de
tection
method still had some
limitations.
Ruiz-del
-Solar
and Q
u
inte
ro
s [6] ma
de a
study of p
r
ep
roce
ssing
app
roa
c
he
s, Zh
a
n
g
et al. [2] reali
z
ed th
e
singl
e-face d
e
tecti
on in
com
p
le
x backg
rou
n
d
,
Shih et al. [
3
] reali
z
e
d
th
e
multiple-fa
c
e
detectio
n
, but
the method
wa
s greatly
a
ffected by illu
mination, ima
g
ing devi
c
e,
and
the interferin
g backg
rou
n
d
s or fa
ke o
b
ject
s. Wan
g
and Li [4] combine
d
the Gau
ssi
an mi
xture
model
and t
e
mplate
mat
c
hin
g
, but t
he pa
ram
e
te
r e
s
timation
of mod
e
l a
nd the te
mp
late
matchin
g
co
st too much time of calcula
t
ion. Hsu
et al
. [5] used the ellipse skin m
odel to se
gm
ent
the skin
regi
o
n
s, but the
m
e
thod n
eed li
ghting
co
mp
e
n
satio
n
, and t
o
ce
nsus l
o
ts
of ski
n sampl
e
s
for calculatin
g the fixed m
odel p
a
ra
met
e
rs that
may
be not
suitabl
e for the
ce
rt
ain inp
u
t ima
ge.
Its anti-interfe
r
en
ce an
d the
sco
pe of app
lication a
r
e li
mited.
Some pe
ople
[7-10] p
r
ovi
ded the
met
hod b
a
s
ed o
n
AdaBoo
st
or n
eural net
work, b
u
t
their meth
od
s requi
red t
r
aining
sam
p
l
e
s, the a
c
cu
racy
of dete
c
tion
wa
s g
r
eatly affecte
d
b
y
training
sam
p
les a
nd te
st image
s, an
d the traini
n
g
or le
arni
ng
wa
s time-co
n
sumi
ng. Fa
cing
these
proble
m
s, for fa
ce
image
s
with
abno
rmal
e
x
posu
r
e a
n
d
interferi
ng b
a
ckgroun
ds,
this
study p
r
o
p
o
s
ed a
fa
st ad
aptive multipl
e
-face
det
e
c
tion m
e
thod.
To e
n
sure
th
e fast
spee
d
,
adapta
b
ility,
and high
er d
e
tection a
c
cu
racy, this
met
hod nee
d not
preprocess, train cla
s
sifie
r
s,
and ado
pt fixed statisti
cal para
m
et
ers that depen
d o
n
sampl
e
set
s
. It is realize
d
base
d
on a self-
establi
s
h
ed d
ual skin m
o
d
e
l, whi
c
h
con
t
ains
a
self-d
efined YIC
skin mo
del a
n
d a self-defin
ed
YIQ skin
-
pro
bability mode
l. Based on the YIC ski
n
model, it acq
u
ired the ski
n
likeliho
od a
n
d
adaptive initi
a
l cl
uste
ring
cente
r
s. B
a
sed o
n
the YI
Q skin
-p
rob
a
b
ility model,
it acq
u
ire
d
t
h
e
adaptive
clu
s
terin
g
weig
hts. Th
en th
e fuzzy
C-m
ean
s
(FCM) clu
s
te
ring
wa
s im
prove
d
to
segm
ent the
skin
regi
o
n
s fa
st and
dynamica
lly. After skin
segm
entati
on, the are
a
s,
circumfe
ren
c
es, and
circu
l
arities of the
conn
ecte
d target
s we
re
use
d
jointly for finishi
ng the
multiple-fa
c
e detectio
n
fina
lly.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2743 – 2
752
2744
2. The Algori
t
hm Flo
w
The multiple
-face d
e
tectio
n
flow is depi
ct
ed in Figu
re 1
,
as the following stag
es:
(a) Input imag
e
(
T
w
o-
perso
n),
(b) Skin-lik
eli
h
ood im
age,
(c) Histogram
of skin
likel
ih
ood,
(d) Skin-pr
oba
bilit
y
imag
e,
(e) Skin-segmented
imag
e,
(f) Connecte
d regi
on
imag
e,
(g) F
a
ce-segm
ented
imag
e,
(h) Face-ex
t
rac
t
ed image.
Figure 1. The
Flow of Dete
ction Algo
rith
m and Involved Image in
Every Stage
Stage 1
:
Mult
iple-fa
c
e ima
ge acqui
sition
.
Stage 2
:
Est
ablish the ad
aptive YIC skin-colo
r
mod
e
l for multipl
e
-face input i
m
age, ba
sed
on
whi
c
h calcula
t
e the skin
-
co
lor likeli
hoo
d and a
c
qui
re the skin-li
k
eli
h
ood imag
e.
Stage 3:
Ba
sed on th
e ski
n
-likeliho
od i
m
age, a
u
tom
a
ti
cally search its smoot
h
e
d gray histo
g
ram,
and acqui
re the initial fuzzy clustering centers.
Stage 4
:
Est
ablish the ad
aptive YIQ skin-p
rob
abilit
y model for mu
ltiple-face inp
u
t image, based
on whi
c
h
cal
c
ulate the skin
proba
bility and acquire the ski
n-p
r
o
bab
ility image.
Stage
5:
Regard the ski
n
probability
as the adaptive fuzzy
clus
teri
ng
weights. Usi
n
g the
adaptive initi
a
l clu
s
te
ring
cente
r
s an
d fuzzy clu
s
te
ri
n
g
wei
ghts,
se
gment the
ski
n
targ
ets of t
he
ski
n-li
kelih
oo
d image
fast
and dyn
a
mically by impro
v
ed FCM
clu
s
terin
g
, and
acq
u
ire th
e
skin-
segm
ented i
m
age.
Stage 6:
La
bel the
co
nn
ected
targ
ets (the fa
ce
candid
a
tes) in
the skin
-se
g
mented
ima
ge.
Cal
c
ulate the
geometri
c feature
s
(the e
ffective area
s, circu
m
fer
e
n
c
e
s
, and ci
rcularitie
s) of the
con
n
e
c
ted targets.
Stage 7
:
By jointly usin
g a
nd limiting th
e geom
etri
c feature
s
, rem
o
ve the noi
ses a
nd the
n
on-
face re
gion
s to disting
u
ish and a
c
qui
re the real fa
ce region
s.
Stage 8
:
Out
put the extracted face targe
t
s.
3. Skin Region Segmentation
3.1. The Dua
l
Skin Model
for Multiple
-face Input Im
age
Variety of color
spa
c
e
s
applie
d to
probl
em
s o
f
skin
-
colo
r
detectio
n
[1
1, 12] or
segm
entation
method [13], such a
s
normalize
d
RGB, YCbCr, YIQ
,
HSV, and TSL. The YCb
C
r
colo
r spa
c
e u
s
ed wid
e
ly
in
kind
s of
existing
skin
-colo
r
mo
dels,
su
ch as Ga
ussia
n
mod
e
l [3, 4
,
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Dete
ction Me
thod of Multiple Face
s (Wei
Li)
2745
12], and elli
pse m
odel [
2
, 5]. To se
gment the
multiple-fa
c
e
targets effe
ctively unde
r th
e
con
d
ition
s
of compl
e
x ba
ckgrou
nd
s, fake
obje
c
ts
, or
a
bnormal expo
sure, a reliabl
e skin mo
del i
s
need
ed. Co
n
s
ide
r
ing n
o
t only the model
can refle
c
t t
he cha
r
a
c
teri
st
ic of distrib
u
tion of ski
n col
o
r,
but also it has good
stabili
ty, ant
i-interference and
adaptability, th
is study planed to establi
s
h
an
adaptive d
ual
ski
n-col
o
r m
odel. Th
e e
s
tablished m
o
d
e
l fuse
d feat
ure
s
of m
u
ltiple comp
one
nts
and
com
b
ine
d
the imp
r
ov
ed fuzzy
clu
s
terin
g
, which incre
a
sed
the sp
eed
a
nd a
c
curacy
of
clu
s
terin
g
to realize the skin segm
entati
on. Con
c
e
r
ni
ng the dual skin-col
or mo
d
e
l, the first model
wa
s u
s
e
d
to
cal
c
ul
ate th
e skin
likelih
ood
and
a
c
quire
the
ski
n
-likeliho
od
gray im
age
for
clu
s
terin
g
, an
alyzing, and
obtainin
g
the initial cl
uste
ri
ng cente
r
s. The se
con
d
model wa
s u
s
e
d
to
acq
u
ire the n
e
ce
ssary info
rmation
su
ch
as the
clu
s
t
e
ring
weig
hts that can en
han
ce the effect
and a
b
ility of clu
s
terin
g
. By analysi
s
, I compon
ent
in
YIQ colo
r
sp
ace
cove
re
d the rang
e of ski
n
colo
r, and
re
pre
s
ent
s the
tone ra
nge from bisque
to
blue-gre
en.
The Cb, Cg,
and
Cr a
r
e th
ree
chromin
a
n
c
e
comp
onent
s indepe
nde
n
t
from Y luminan
ce
co
mpone
nt. If
the influen
ce
o
f
luminan
ce
(Y
)
coul
d b
e
d
epre
s
s, b
e
sid
e
s I
compo
n
ent an
d diffe
ren
c
e
s
of three
ch
romin
a
n
ce
comp
one
nts
coul
d be
co
mbined
in a
n
adaptive
ski
n
mod
e
l, it is very ben
eficial for im
prov
ing
anti-inte
rfere
n
ce an
d ada
ptability of segmentati
on.
Based o
n
this idea, an ad
aptive YIC model
fused multipl
e
comp
one
nts of Y, I, Cb, Cg, and
Cr was built (Eq
u
a
t
ion (2)).
(a)
(b)
(c
)
(d)
(a) Origin
al im
age 1,
(b) likel
ih
ood i
m
age 1,
(c) histogram o
f
(b)(burrs),
(d) smoothe
d h
i
stogram
of (c)(t
w
o
peak
s),
(e)
(f)
(g)
(h)
(e) Origin
al im
age 2,
(f) likelih
oo
d im
age 2,
(g) histogr
am o
f
(f)(burrs),
(h) smoothe
d h
i
stogram
of (g)(three pe
aks),
(i)
(j)
(k
)
(i) skin-lik
eli
h
o
od
distrib
u
tion of (
a
),
(j) skin-probabilit
y
distrib
u
tion of (
a
),
(k) Comparis
on of
distrib
u
tion of (i
) and (j),
(l)
(m)
(n)
(l) skin-lik
eli
h
o
od
distrib
u
tion of (
e
),
(m) skin-probabilit
y
distrib
u
tion of (
e
),
(n) Comparison of
distrib
u
tion of (l
) and
(m).
Figure 2. The
Analysis of the Adaptiv
e Skin Likeliho
od and Ski
n
Proba
bility
As Figu
re 2 showed, the Fi
gure
2(a
)
is a
multiple-fa
c
e
colo
r image
with und
erex
posure,
the Figure 2(e) is a m
u
lti-f
a
ce
colo
r ima
ge
with
comp
licated b
a
ckg
r
oun
ds
and u
neven expo
sure.
Based
o
n
th
e built
YIC
model, th
e a
c
qui
red
n
o
rm
alize
d
a
dapti
v
e likeli
hoo
d
gray ima
ge,
the
likeliho
od
gray histog
ram
,
and the
smoothed
hi
stogra
m
a
r
e
showed in
Fi
gure
2(b)-(d) and
Figure 2(f
)
-(h
)
, the skin
-likelihoo
d dist
ri
bution im
a
ge
are
sho
w
n i
n
Figure 2(i
)
a
n
d
2(l
)
. From t
h
e
distrib
u
tion, t
he no
rmali
z
e
d
skin
-likeli
h
ood valu
es
based o
n
th
e YIC mo
del
can
refle
c
t
the
comp
actn
ess and
sta
b
ility of re
al
ski
n
distri
bution
well. Mo
re
over the
smoot
hed
histo
g
ra
m
approximatel
y showed the
double
-
pe
ak
curve o
r
three-p
e
a
k
cu
rv
e. Th
is indi
ca
ted that most
o
f
the pixel values of no
rmali
z
ed li
kelih
oo
d image
co
n
c
entrated
on two to thre
e main gray levels,
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2743 – 2
752
2746
as
sh
owed
in
Figu
re
2(d)
and
(h
).
Note
that the
r
e
are two
p
e
a
k
s (F
0
and F
2
) in
Figure 2
(
d
)
, a
nd
there a
r
e th
re
e pea
ks
(F
0,
F
1,
and F
2
) in F
i
gure
2(h
)
. Th
e sha
r
p
-
poi
nted maximum
pea
k F
0
, which
rep
r
e
s
ente
d
the num
ber
of pixels in ba
ckgroun
d re
gi
ons, i
s
nea
r to 0 gray level or in the ran
g
e
of low gray levels. The possible s
harp-p
o
inted
se
con
d
maximum
p
eak F
1
, w
h
ic
h r
e
pr
es
en
te
d
th
e
numbe
r of pixels in no
n-skin or fake-ski
n
region
s,
is lo
cated in the
range of lo
w g
r
ay levels. Th
e
flat pea
k F
2
,
whi
c
h
rep
r
e
s
ented th
e n
u
m
ber of pixel
s
in
skin
re
gi
ons, i
s
l
o
cate
d in th
e rang
e of
middle o
r
hig
h
gray levels.
Con
s
id
erin
g
that the F
C
M clu
s
te
ring
algo
rith
m i
s
se
nsitive to
the initial
clusteri
ng
cente
r
s an
d t
he
spe
ed a
n
d
effect
of th
e cl
uste
ring
maybe u
n
sta
b
le b
e
cau
s
e
of the differe
nce
of
initial clu
s
teri
ng cente
r
s, t
h
is
study im
proved
the F
C
M
clu
s
terin
g
to re
alize the dynami
c
ski
n
regio
n
segm
entation. Th
e
main id
ea
o
f
improved
F
C
M al
gorith
m
is a
s
follo
ws: ba
sed
on
the
traditional
F
C
M algo
rithm, t
he g
r
ay level
s
of
pea
ks in
ski
n-li
kelih
oo
d hi
stogram
wa
s rega
rd
ed
as
the initial fu
zzy cl
uste
ring
ce
nters to
decrea
s
e
the
invalid
com
putation
of cl
usteri
ng, a
n
d
to
greatly im
pro
v
e the
sp
eed
, effect a
nd
adapta
b
ility
o
f
clu
s
teri
ng. I
n
ad
dition, a
nother a
dapti
v
e
YIQ skin-pro
bability mod
e
l
(Equ
ation
(5
))
wa
s
esta
bli
s
he
d. The
a
c
quire
d
ski
n p
r
obability valu
es
based o
n
th
e YIQ model
reflecte
d th
e deg
ree
of ski
n proba
bili
ty. The skin
prob
ability was
rega
rd
ed as
the adaptive weig
hts of fuzzy cl
uste
ri
n
g
to update the clu
s
teri
ng
centers, and
to
achi
eve the
purp
o
ses
of improvin
g
the sp
eed
of clu
s
teri
ng an
d a
c
q
u
iring th
e b
e
tter
segm
entation
.
The
built a
d
a
p
tive YIQ ski
n
-p
rob
ability
model
co
mbi
ned th
e diffe
rences bet
we
en I, Q,
and Y compo
nents, a
nd th
e differen
c
e
betwe
en I co
mpone
nt and
its expectati
on (
)
, as
sho
w
n
in Equatio
n
(5). Th
e di
stri
bution
of the
ski
n p
r
ob
abilit
y based
on th
e YIQ mo
del i
s
a
pproximat
ely
clo
s
e to th
e d
i
stributio
n of t
he n
o
rmali
z
e
d
skin
li
keli
ho
od b
a
se
d o
n
the YIC m
o
d
e
l, as
sh
own
in
Figure 2. To
the skin ta
rgets in
Figu
re 2(a)
a
nd
(e), the Fi
gure 2(i
)
an
d (l
) sho
w
th
e skin-
likeliho
od di
stribution b
a
sed on YIC m
odel, the Fi
g
u
re 2
(
j) a
nd
(m)
sho
w
the
skin
-
p
r
ob
abil
i
ty
distrib
u
tion b
a
se
d on YI
Q mod
e
l, the Figu
re
2
(
k) and
(n
)
show th
e fitness of the t
w
o
distrib
u
tion
s.
As mention
e
d
before, for
the sub
s
e
que
nt clus
te
ring
segm
entation
,
the initial clusteri
ng
cente
r
s
wa
s acq
u
ire
d
by sea
r
ching th
e ski
n-
li
kelih
ood histo
g
ra
m of the YIC
model, the ski
n
prob
ability of
the YIQ model was rega
rd
ed as the cl
u
s
terin
g
wei
g
h
t
s in this stud
y. This method
actually
comb
ined the t
w
o
kind
s of a
dap
tive model
s,
whi
c
h h
ad th
e goo
d ability of descri
p
tio
n
to
the clu
s
teri
ng
distrib
u
tion o
f
skin targets.
As t
he distri
butional p
a
ttern an
d gray l
e
vels range
of
the skin li
keli
hood an
d the
skin proba
bi
lity were cl
o
s
e to each other ap
proxim
ately, the initial
clu
s
terin
g
ce
nters
(came f
r
om skin li
keli
hood
) an
d the pro
babili
stic weight
s (ski
n
prob
ability)
wa
s
use
d
for fu
zzy cluste
ring n
o
t only avoid
ed the di
verg
ence and
un
controlla
bility of clust
e
rin
g
, but
also
de
cre
a
sed the m
a
ssi
v
e cal
c
ulatio
n
for invalid
cl
usteri
ng. By this
way, the
sea
r
ching fo
r the
ski
n target
s b
y
the adaptive clu
s
terin
g
q
u
ickly and a
c
curately can
be reali
z
e
d
.
3.2. The Improv
ed Fuzz
y
C-me
ans Clu
s
ter us
ed fo
r
Skin Segmenta
tion
Based o
n
the adaptive dual skin m
odel,
the ski
n segm
entati
on can b
e
realized
according to the followi
ng steps:
Step 1:
A
c
qui
re the Y, I, Q, Cb, Cr, an
d Cg compo
n
e
n
ts of input image by Equ
a
tion (1
).
(1)
Step 2:
E
s
ta
blish th
e a
d
a
p
tive YIC ski
n
mod
e
l u
s
in
g Y, I, Cb, Cr, and
Cg
com
pone
nts,
based on
whi
c
h calculate the
ada
ptive skin likelihoo
d (P
i,j
) by Equation (2
).
(2)
Whe
r
e, i and
j are the ho
rizontal and vert
ical
coordinat
es of every pi
xel in input image.
I
R
G
B
Cr
Cg
Cb
Q
I
Y
500
.
0
4187
.
0
0813
.
0
316
.
0
500
.
0
184
.
0
1687
.
0
3313
.
0
500
.
0
212
.
0
523
.
0
311
.
0
0.596
0.274
0.322
299
.
0
0.578
114
.
0
128
128
128
0
0
128
)
2
/
)
((
)
(
)
(
3
2
3
2
3
2
,
Y
I
Cr
Cg
Cb
Y
I
Cr
Cg
Y
I
Cr
Cb
P
j
i
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Dete
ction Me
thod of Multiple Face
s (Wei
Li)
2747
Step 3:
Acqu
ire the likeli
h
ood gray ima
ge (T
p
) u
s
ing
the ski
n likeli
hood (P
i, j
) by
Eq. (3),
norm
a
lize th
e gray levels to the
rang
e
of [0 25
5], a
nd a
c
qui
re th
e no
rmali
z
ed
likelih
ood
gray
image (T
p
'
)
b
y
Eq. (4), as sho
w
n in Fig
u
re 2
(
b), 2
(
f)
and Figu
re 3
(
d).
(3)
0
)
'
'
(
if
)
'
'
(
)
'
'
(
255
else
0
'
'
m
m
m
m
m
in
ax
in
ax
in
P
P
P
P
P
P
P
(4)
Whe
r
e, i a
nd
j are th
e ho
ri
zontal
and ve
rtical
coo
r
di
n
a
tes of eve
r
y
pixel, P' and
P'' are t
he g
r
a
y
values of no
n
-
no
rmali
z
ed a
nd normali
ze
d likelih
ood i
m
age (T
p
and
T
p
').
Step 4:
Acq
u
ire the
one
-dimen
sion
al
histog
ram
of norm
a
lized l
i
kelih
ood
gra
y
image
(T
p
'),
as sho
w
n i
n
Fi
gure
2(c) an
d
2(g), ave
r
ag
e t
he th
ree
nei
ghbo
rho
o
d
s
of gray level
s
to
smooth
the
h
i
stogram
burr, and
acqui
re
the
smo
o
th
e
d
hi
stog
ram,
as
sh
own in
Figure 2
(
d
)
a
nd
2(h
)
. In the range of [50 2
50], sea
r
ch
the po
sition
s of gray levels (T
1
and T
2
),
whi
c
h the pe
aks
(F
1
an
d F
2
) in his
t
ogram
were located in. If the T
1
i
s
not foun
d, th
ere
are o
n
ly the two p
e
a
k
s F
0
and F
2
, s
o
T
1
=0. Re
gard the T
1
and T
2
as the gray levels of initial FCM clu
s
te
ring
centers.
Step 5:
Calculate I and Q para
m
eters a
nd their
expe
ctation
s
(
and
) requi
re
d by YIQ
ski
n-p
r
o
babili
ty model. Usi
ng I and Q co
mpone
nts an
d their expe
ctations, a
c
qui
re the differen
c
e
(W'
)
between
I an
d its dyn
a
mic expe
cta
t
ion (
)
(i
ndi
ca
ting the
skin
prob
ability, a
s
the
a
daptiv
e
weig
hts
of fuzzy cl
uste
ri
ng). T
hat m
ean
s, calc
ula
t
e the a
dapti
v
e clu
s
teri
ng
wei
ghts W'
by
Equation (5) t
o
acq
u
ire th
e weig
hts gray image T
w
, normalize the
we
ights to the ra
nge of [0 255
]
by Eq. (6) to
acq
u
ire th
e n
o
rmali
z
e
d
ad
aptive clu
s
teri
ng wei
ghts
gray image T
w
' (Figu
r
e 3
(
e
))
for
updatin
g the FCM cl
uste
ri
ng ce
nters.
(5)
0
)
'
'
(
if
)
'
'
(
)
'
'
(
255
else
0
'
'
m
m
m
m
m
in
ax
in
ax
in
W
W
W
W
W
W
W
(6)
Whe
r
e,
W'
a
nd
W''
are th
e no
n-normal
i
zed
an
d
normalize
d
adap
tive clu
s
teri
n
g
weight
s. T
w
and
T
w
' are the clusteri
ng weig
hts gray ima
g
e
s.
Step 6:
Seg
m
ent the no
rmalize
d
likeli
hood im
age
(T
p
') a
nd a
c
q
u
ire the
bina
ry ski
n-
segm
ented i
m
age T
b
by improve
d
FCM clust
e
rin
g
.
3.3. The Improv
ed Fuzz
y
C-me
ans Clu
s
ter us
ed fo
r
Skin Segmenta
tion
Use normalized adaptive li
kelih
ood ima
ge (T
p
'), initial
cluste
ring
ce
nters
(T
1
and
T
2
), and
adaptive cl
ustering
weig
ht (W'
') to imp
r
o
v
e traditi
onal fuzzy C-m
e
a
n
s cluste
r.
Th
e
pro
c
e
dures
of
improve
d
FCM clust
e
rin
g
algorith
m
are
descri
bed a
s
follows:
Step 1:
T
he
numbe
r
of sa
mples (n) i
n
t
he initial
sa
m
p
le
set ({x
i
(i=1,2,…,n)}) is
the total
numbe
r of pi
xels in the n
o
rmali
z
e
d
ad
aptive likeliho
od image T
p
',
that is, n=h×w (h an
d w a
r
e
height a
nd wi
dth of T
p
'). S
e
t the numb
e
r
of clu
s
te
rin
g
ce
nters (c=2) an
d the
weighe
d expon
ent
(b=3) that ca
n control the fuzzy degree
of cluste
ring.
Step 2:
Initi
a
lize the
sample
set {x
i
(i=1,2,…,n
)}
containe
d n
sample
s. Initialize th
e
clu
s
t
e
rin
g
ce
nt
ers
(C
j
(j=1,2,…,c)) and t
he memb
ership functio
n
(
μ
ij
(i=1,2,…,n;
j=1,2,…,c)),
and
the
μ
ij
represents the probability of the fi
rst i sam
p
le (x
i
) belong
s to
the class j.
Step 3:
A
c
co
rding
to the
i
m
prove
d
me
mbershi
p
fun
c
tion
(Equ
ation (11
))
of F
C
M, u
s
e
the curre
n
t cl
usteri
ng cent
ers
(C
j
) to cal
c
ulate the val
ue (
μ
ij
') of me
mbershi
p
deg
ree. Accordin
g
to the com
p
utational formula of cl
ustering
c
ente
r
(Equatio
n (10)), use th
e ne
w value
of
membe
r
ship degree (
μ
ij
') t
o
upd
ate the
clu
s
teri
ng
ce
nters (C
j
). A
c
cording
to th
e Equatio
n (10)
and
(11
)
, ma
ke th
e
clu
s
tering
cente
r
s a
pproxim
ate th
e target p
o
siti
on from the
i
n
itial po
sition
by
contin
uou
s iteration.
For tradition
al FCM
clu
s
t
e
ring
algo
rith
m,
the com
p
utational formula of me
mbershi
p
degree is d
e
scrib
ed by Eq. (9), and the
comp
utati
ona
l formula of cl
usteri
ng cent
ers i
s
de
scrib
e
d
by Equation
(10
)
. The Eq
uation (9) a
n
d
(10
)
a
r
e a
c
qui
red
by limiting the to
tal membe
r
ship
255
'
j
i,
P
P
I
Q
I
)
(
'
Y
I
Q
I
I
I
W
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2743 – 2
752
2748
degree
(Equ
ation (7)) of
sampl
e
s to e
v
ery cla
s
s in
data
set a
nd
finding th
e m
i
nimum
of value
function (E
qu
ation (8
)).
(7)
(8)
c
1,2,...,
j
,
n
1,2,...,
i
1
1
1
2
2
c
k
b
k
i
b
j
i
ij
C
x
C
x
(9)
c
1,2,...,
j
,
1
1
n
i
b
ij
n
i
i
b
ij
j
x
C
(10
)
Based o
n
tra
d
itional FCM,
this study introdu
ce
d the gray value of pixels in normalize
d
clu
s
terin
g
we
ights im
age
T
w
' (Figure
3
(
e)) into
com
putational
fo
rmula of me
m
bership
deg
ree.
That mean
s t
he ada
ptive ski
n probabili
ty was re
gar
ded a
s
ada
ptive cluste
ring
weight (W''
)
to
improve the f
o
rmul
a of membershi
p
deg
ree, as
sho
w
n in Equation
(11
)
.
c
1,2,...,
j
,
n
1,2,...,
i
,
1
1
'
1
'
'
2
'
'
2
c
k
b
k
i
b
j
i
ij
i
i
W
C
x
W
C
x
(11
)
Step 4:
Afte
r ea
ch of up
dating the cl
usteri
ng c
ent
ers i
n
the proce
s
s of interactio
n,
cal
c
ulate
d
th
e value
fun
c
tion by E
quatio
n (8),
a
nd det
ermin
ed wh
ether
the ch
ang
e
of
the
curre
n
t
value relative
to the last va
lue of value func
tio
n
wa
s less than the thre
shol
d valu
e (K=1
×10
-9
).
If
the cha
nge
wa
s less tha
n
threshold,
whi
c
h
re
presented the al
gorithm
wa
s convergent,
so
stopp
ed the clu
s
terin
g
, or else retu
rn
ed to cy
clica
lly execute the E
quation
(11) an
d (1
0),
contin
ued
the
execution
an
d sto
ppe
d th
e cl
uste
rin
g
u
n
til the
ch
ang
e of val
ue fu
nction
was le
ss
than the thre
shold value.
The
ski
n-seg
m
ented im
ag
e (T
b
) segme
n
ted by the
improve
d
F
C
M is
sho
w
n
i
n
Figu
re
3(f). Compari
ng the segmented re
sults
based on the
Ga
ussi
an and Ellipse mo
del in Figure 3(b)
and (c), the e
rro
r se
gme
n
tation wa
s ob
viously red
u
ced in Figu
re
3(f). Accordin
g to the stan
dard
targets, th
e
averag
e seg
m
entation a
c
curacy of
th
e improved
FCM
clu
s
teri
ng to the ta
rgets
rea
c
he
d more than 95%.
4. Multiple Faces
Detec
t
ion
As sho
w
n in
Figure 3(f), in
the skin
-
se
g
m
ented imag
e segm
ented
by the improved FCM
clu
s
terin
g
, th
ere
still were
some
po
ssibl
e
noi
se
s
a
nd non-skin or
fa
lse-skin non
-face
targ
ets. To
extract the
re
al face ta
rg
ets, this
study
comb
i
ned
used an
d limite
d
the ge
omet
ric fe
ature
s
(the
effective
a
r
e
a
s, circumfe
rences, and
circul
aritie
s) of
co
nne
cted
target
s to
re
move noi
se
s and
non-fa
ce ta
rg
ets. By the statistics a
nd
analysi
s
of
th
e face o
b
je
cts (with ba
ckg
r
oun
ds
or fa
ke
targets,
non
-simple
face
) i
n
the multipl
e
-face
set,
it is foun
d that,
the multiple
-face ta
rg
et often
wa
s sm
aller i
n
si
ze than t
he sin
g
le
-face targe
t, but
wa
s larger th
an the noi
se.
After norm
a
lize
d
zoomi
ng th
e
multiple-fa
c
e
image to
the
size of
1
5
0
×
150 pixel
s
, th
e mo
st of eff
e
ctive a
r
ea
s
of
multiple-fa
c
e
targets
we
re in the rang
e of 0.5%
~90%
of the total
area
s (Eq
uati
on (12
)). If the
effective area
ratio of a conne
cted targ
et was
big
g
e
r
than 90%, the target maybe a singl
e face.
If the effective are
a
ratio (
)
of a co
nne
ct
ed ta
rg
et wa
s smalle
r tha
n
0.5%, the target mayb
e
the noise or
non-fa
ce
sma
ll regio
n
. Vital few targ
ets
were the tiny face
s, whi
c
h
can b
e
seen
as
the noi
se
be
cause it
did
no
t have
enou
g
h
info
rmat
ion
to
fini
sh su
b
s
eq
uent processing, su
ch
as
the facial fea
t
ures
extracti
on. The mo
st
of circ
umferences (Equat
ion
(1
3))
of the re
al multiple-
face targ
ets were in th
e rang
e of 10~225
00 pi
xels (The b
ound
ary len
g
th of the
whol
e
norm
a
lized
zoomed i
m
ag
e
wa
s 22
500
pixels. Th
e m
o
st of
circumf
e
ren
c
e
s
of th
e noi
se o
r
n
o
n
-
face small re
gion
s wa
s
smaller tha
n
1
0
pixels). Th
us, this
study
removed th
e
noise
s a
nd
no
n
-
c
j
n
i
ij
11
1
c
j
n
i
j
i
b
ij
C
x
J
11
2
i
S
S
K
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Dete
ction Me
thod of Multiple Face
s (Wei
Li)
2749
face
small t
a
rget
s by joi
n
tly limiting the e
ffective
area
(Equ
ation (1
2)) and
circumfe
ren
c
e
(Equatio
n (13
)). Besid
e
s,
according to the facial
ge
ometri
c cha
r
acteri
stic
s of oval, it used the
value inte
rval (Equatio
n (1
5))
of the
circula
r
ity
factor (Equatio
n (1
4))
of conn
e
c
ted ta
rget
s to
disting
u
ish th
e face
and
n
on-fa
ce la
rg
e
r
targ
ets.
Accordingly the
non-fa
ce l
a
rger
regi
on
s that
wa
s clo
s
e in
size but different in sha
pe to the non-fa
ce regio
n
s
can
be remove
d.
(12
)
(13
)
(14
)
(15
)
Whe
r
e, S is t
o
tal are
a
of the wh
ole ima
ge,
K is se
ria
l
numbe
r of the co
nne
cted
target, S
k
,
F
k
,
and C
k
a
r
e th
e effective area, circumfe
rence, ci
rcula
r
ity of
the first K conne
cted t
a
rget.
The ste
p
s of
multiple-fa
c
e
detectio
n
are
descri
bed a
s
follows:
Step 1:
Fast
label
the co
nne
cted regi
ons
of the
skin-segm
ented
image
(T
b
), cal
c
ulate
effective area
(S
k
) and ci
rcumfere
nce (F
k
) of the conn
ected ta
rgets.
Step 2:
Rem
o
ve the n
o
ises
or
som
e
n
on-fa
ce
small
regi
on
s by
combine
d
limit
ing the
effective area
s and the
circumfere
nces.
Step 3:
Cal
c
ulate circul
ari
t
y factors (C
k
). By limitin
g the ran
ge
of circula
r
ity factors,
remove the n
on-fa
ce la
rge
r
target
s and
acq
u
ire the
re
al face targ
ets.
By limiting the ge
ometri
c feature
s
, th
e
a
c
qui
red multiple-fa
c
e
se
gmente
d
image
i
s
sho
w
n in Fig
u
re 3
(
g). Con
t
rasting Fi
gure 3(f), it
is easy to find that, the non-fa
ce skin
regio
n
s
(e.g. the pal
ms, arm
s
, an
d legs), the p
o
ssible
ski
n-similar regio
n
s, and the noi
se
s in the fa
ce
can
d
idate
s
were removed.
5. Results a
nd Discu
ssi
on
5.1. Experimental Results and An
aly
s
is
The platform
of experime
n
ts is: P4 2.
10G
Hz
CPU,
2G memo
ry, WinXP, VC6.0. The
prop
osed
me
thod
wa
s a
p
p
lied in
the
experim
ents on
two
multiple
-face te
st
set
s
. The
built te
st
set 1
co
ntain
s
20
0 ima
g
e
s
, whi
c
h
con
s
iste
d of
1
0
0
two-perso
n,
50 thr
ee-pe
rson, and
50
4
-
or-
more
-pe
r
son
image
s. Ta
bl
e 1
sum
m
ari
z
es th
e fa
ce
e
x
traction
re
su
lts ba
se
d o
n
t
e
st
set 1,
whi
c
h
contai
n 662 f
a
ce
s with va
ri
ations. Th
e Detection
Rate
(DR) is
de
fi
n
ed as the
rati
o of the num
ber
of corre
c
tly detected fa
ce
s to the total numbe
r of faces in all im
ag
es. The
(FP
R
) is d
e
fi
ned a
s
the ratio of the numb
e
r o
f
detected false po
siti
ves
to the total n
u
mbe
r
of faces. The Fal
s
e
Pos
i
tive Rate (FNR) is
de
fi
ned as the ra
tio of the number of
false
negatives to the total number
of face
s. The
built test set
2 co
ntain
s
30
0 image
s, whi
c
h
con
s
i
s
ted
of 150 ima
g
e
s
with a
bno
rmal
exposure
an
d 150 ima
g
e
s
with inte
rfering
ba
ckg
round
s o
r
fake obje
c
ts. T
able 2 list
s
the
perfo
rman
ce
comp
ari
s
on
of propo
se
d method a
nd othe
r me
thods b
a
sed
on test set 2.
Comp
ari
ng wi
th the other
method
s, there is n
o
ne
ed t
o
pre
p
rocess,
train cl
assifie
r
s, an
d an
alyze
lots of statist
i
cal sampl
e
s for acquiri
n
g
fixed
para
m
eters of m
odel o
r
interval threshold
in
prop
osed m
e
thod. It acq
u
ired skin li
kelih
ood a
nd initia
l clu
s
terin
g
ce
nters by the b
u
ilt YIC mode
l,
and a
c
qui
red
adaptive fuzzy clu
s
terin
g
weight
s by
the built YIQ model. Fro
m
Table 2, the
detectio
n
rat
e
of prop
ose
d
method i
s
highe
r than
t
hat of other
method
s, but
the false po
sitive
rate of pro
p
o
s
ed meth
od i
s
lower than t
hat of other
method
s.
Table 1. Perf
orma
nce Co
mpari
s
o
n
of 2-
pe
rson, 3-p
e
r
so
n, and 4
-
o
r-mo
r
e
-
pe
rso
n
No. of Images
No. of Faces
N
o. of CD
No. of FP
No. of FN
DR
FPR
F
NR
2-person
100
200
197
9
3
98.5
4
.5
1.5
3-person
50
150
146
11
5
97.3
7
.5
3.3
4-or-
m
ore
-
perso
n 50
312
304
26
15
97.4
8
.3
4.8
Total 200
662
647
46
23
97.7
6
.9
3.5
)
90%
S
5%
(
S
S
k
2500)
2
0
1
(
F
K
S
F
C
k
k
k
/
2
)
20
5
(
C
k
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2743 – 2
752
2750
Table 2. Perf
orma
nce Co
mpari
s
o
n
s of
Propo
se
d Me
thod and oth
e
r
Method
s
Detection Metho
d
s
Test Set / Faces
Abnormal Exp
o
sure/457
Interfering Backg
rounds/472
DR
FPR
DR
FPR
Detection using Gaussian Model [3]
84.6
14.5
85.5
15.8
Detection using Ellipse Model [2
5]
84.7
13.9
86.2
14.3
Detection using Adaboost [7 9]
85.6
15.6
85.7
16.2
Detection using Neural Net
w
ork
[8 9]
84.3
6.9
86.3
7.6
Detection using Template Matchi
ng [4]
94.3
7.3
95.8
8.1
Detection using Proposed Metho
d
98.2
6.5
97.5
7.2
CD: Corre
c
t Dete
ction, FP: Fals
e Positi
ve, FN: False
Negative,
DR(%): Detec
t
ion Rate, FPR(%): Fals
e Pos
i
tive Rate, FNR(%): Fa
ls
e Negative
Rate.
(a)
(b)
(c
)
(d)
(e)
(f)
(g)
Figure 3. The
Results of Skin Segm
enta
t
ion
and Fa
ce
Detectio
n wit
h
Propo
se
d Algorithm
(a) Multipl
e
-fac
e inp
u
t imag
es, (b) Skin-segm
ented im
ag
es of (a) based
on
CbCr Gaussi
a
n
mode
l, (c)
Skin-se
g
mente
d
imag
es of (a) based o
n
Cb
C
r
e
llips
e
mod
e
l,
(d) Skin-like
l
i
h
ood im
ages
ba
sed on
prop
osed YIC
mode
l, (e) Skin
-prob
abi
lit
y
ima
ges
bas
ed o
n
prop
osed YIQ mode
l, (f) Skin-segme
n
ted
imag
es base
d
on pro
pos
ed m
ode
l (YIC and
YIQ dual
skin
mode
l), (g) F
a
ce-detecte
d ima
ges bas
ed o
n
prop
osed m
e
th
od.
The exp
e
rim
ental exam
pl
es
with
different
con
d
itio
ns (Lo
w
exposure, co
m
p
licate
d
backg
rou
n
d
s
and gla
s
ses, compli
cate
d backg
rou
n
d
s
and
fa
ke
obj
ect,
hig
h
exp
o
su
re and
fa
ke
obje
c
t) are
sh
own in Fig
u
re
3 and Figure
4. Figur
e 3 shows the re
sults of
skin
se
gmentation a
nd
face d
e
tectio
n with p
r
op
o
s
ed
algo
rith
m. Figure
4
sho
w
s re
sult
s of
mult
iple
-
f
ace ex
t
r
a
c
t
i
o
n
.
In
pro
c
e
s
s of
de
tection, the
scalin
g
strateg
y
is u
s
e
d
to i
m
age
no
rmali
z
ation
for han
dling ve
ry la
rge
image, which is up to 4 milli
on pixels.
Ho
wever,
the m
e
mory
con
s
u
m
ing of dete
c
tion wa
s only
8-
12M, and th
e
averag
e tim
e
co
nsuming
wa
s 70
~8
0m
s (1
2.5~14fp
s
). Th
rou
gh e
x
perime
n
ts, it is
clea
r that, for poo
r-qual
ity color fa
ce imag
e with abn
orm
a
l exposure
and interfe
r
ing
backg
rou
n
d
s
, the p
r
o
pose
d
alg
o
rithm
had
sati
sf
act
o
ry
spee
d, a
c
cura
cy
and
ada
ptability of
detectio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
An Adaptive
Dete
ction Me
thod of Multiple Face
s (Wei
Li)
2751
(a)
(b)
(c
)
(d)
(e)
(f)
(g)
(h)
Figure 4. The
Examples of Multiple-fa
c
e
Extraction
(a) T
w
o-
pers
o
n
image (
w
it
h lo
w
e
x
pos
ure), (b) T
h
ree-perso
n imag
e (
w
ith c
o
mplic
ated b
a
c
k
grou
nds
and g
l
ass
e
s), (c)
T
h
ree-pers
o
n imag
e (
w
ith c
o
mplic
ated b
a
c
k
grou
nds a
nd fake ob
ject), (d)
T
h
ree-
perso
n imag
e (w
it
h hi
gh e
x
po
sure an
d fake
obj
ect), (e
)F
ace-e
x
tracted im
age of (a
), (f) Face-e
xtracte
d
image of (b), (g) Face-ex
t
ract
ed image of
(c),
(h) Face-ex
t
ra
c
t
ed imag
e of (d
).
5.2. Discussi
on of Propo
sed Algori
t
h
m
The main feat
ure
s
of this al
gorithm a
r
e a
s
follows:
1. There
wa
s no ne
ed to prep
ro
ce
ss and train the cla
s
sifier.
By combini
ng the
luminan
ce
(Y
), the differe
nce
s
of the
chromin
a
n
c
e
(Cb,
Cg, an
d Cr) an
d I comp
one
nt, this
algorith
m
est
ablished the
adaptive mul
t
iple-com
p
o
n
ent YIC skin
model. The
calculated skin
likeliho
o
d
ba
sed o
n
YIC mo
del
can
reflect the
clu
s
terin
g
cha
r
a
c
teri
stics of
skin
col
o
r, a
nd
help
to
resi
stant to lu
minan
ce vari
ations a
nd ba
ckgro
und
s int
e
rferi
ng.
2.
Usi
ng
th
e distrib
u
tion chara
c
te
risti
c
of
norm
a
lized
ski
n li
keliho
o
d
, the ad
apti
v
e initia
l
clu
s
terin
g
ce
nters of F
C
M
was a
c
qui
re
d by t
he automatic hi
stog
ram an
alysis and sea
r
chi
ng.
This d
e
crea
sed the invali
d cal
c
ulatio
n
and the
cl
u
s
terin
g
in
stab
ility cause
d
b
y
the unsuita
ble
initial cluste
ri
ng cente
r
s, which hel
ped t
o
improv
e th
e spee
d, accura
cy and ad
aptability of fuzz
y
clu
s
terin
g
se
gmentation.
3. Combini
n
g
I, Q compo
nents a
nd th
eir expe
ctatio
ns, this alg
o
rithm establi
s
hed the
adaptive YIQ skin
-
proba
bil
i
ty model. The skin p
r
ob
ab
ility based on
YIQ model was re
garded
as
the dynami
c
clu
s
terin
g
we
ights to
red
u
c
e the
inva
lid
cal
c
ulatio
n o
f
cluste
rin
g
a
nd p
r
omote
the
clu
s
terin
g
co
nverge
nce a
s
so
on a
s
po
ssible.
Usi
ng t
he initial
clu
s
t
e
ring
centers
and th
e dyna
mic
clu
s
terin
g
we
ights, the im
proved
co
mp
utational
formula of me
m
bership
deg
ree was
pro
p
o
se
d.
The imp
r
ove
d
FCM
clu
s
te
ring
segm
ent
ation wa
s b
e
nefit for improving t
he spe
ed, accu
ra
cy and
adapta
b
ility of segmentatio
n.
4. Jointly u
s
ing an
d limit
ing three fa
ctor
s
(the ef
fective area,
circumfe
re
n
c
e, an
d
circula
r
ity) of
con
n
e
c
ted
targets i
n
skin-s
e
g
me
nte
d
image, th
e non
-face region
s in fa
ce
can
d
idate
s
were removed
fast, which in
cre
a
sed the correct dete
c
ti
on rate.
6. Conclusio
n
This stu
d
y p
r
ovided
an
ad
aptive an
d fa
st dete
c
ti
on
a
l
gorithm
of m
u
ltiple-fa
c
e t
a
rgets.
It
establi
s
h
ed two ada
ptive ski
n model
s, based on wh
i
c
h it acqui
red
the skin
-
likelihood ima
ge for
clu
s
terin
g
, the adaptive ini
t
ial clust
e
rin
g
cente
r
s and fuzzy clu
s
teri
ng
weights. Then
it
ad
opt
ed
the improve
d
FCM cl
uste
ri
ng to segm
en
t and to acqui
re the skin ta
rgets. Fin
a
lly it distingui
she
d
the face ta
rg
ets by j
o
intly
limiting the
areas,
ci
rcumfe
ren
c
e
s
, an
d
circula
r
itie
s o
f
the conn
ect
ed
targets.
The
experi
m
enta
l
re
sults sho
w
that, fo
r t
he multipl
e
-f
ace
imag
es
with inte
rferi
ng
backg
rou
n
d
s
, fake
obj
ect
s
, an
d a
bno
rmal expo
su
re, this algo
ri
thm ha
s
sati
sfacto
ry
spe
ed,
accuracy, an
d adapta
b
ility of face detection.
Ackn
o
w
l
e
dg
ements
This
wo
rk
wa
s supp
orted
i
n
pa
rt by the
Si
chua
n Prov
incial
De
part
m
ent of Scie
nce
and
Tech
nolo
g
y Suppo
rting P
r
oje
c
t (No. 2
012GZ
002
0),
the Natu
ral
Scien
c
e Ke
y Found
ation
of
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 2743 – 2
752
2752
Sichua
n’s Province Edu
c
ation De
part
m
ent (No.
11ZA041
), the Key Project of Educat
ion
Dep
a
rtme
nt of Sichuan P
r
ovince (No.
13ZA00
15), the Scientific
Re
sea
r
ch Fo
undatio
n of the
Educatio
n De
partme
n
t of Sichu
an Provin
ce of Chi
na (No. 13ZB0
01
2).
Referen
ces
[1]
Abate AF
, Nap
p
i M, Riccio D,
Sabatin
o G. 2D and 3
D
F
a
ce Reco
gniti
on: A Surve
y
.
Patt. Recog. Lett
.
200
7; 28: 188
5
-
190
6.
[2]
Z
hang DZ
, W
u
BY, Sun JB, Liao QL.
A Face Detection Metho
d
b
a
sed o
n
Skin
Color Mod
e
l
.
Procee
din
g
s of
the 11th Joi
n
t Confer
ence
on
Info
rmation Sc
ienc
es. Shenz
hen, Ch
in
a. 20
08; pp: 1-5.
[3]
Shih FY, Ch
e
ng S
X
, C
hua
n
g
CF, Wang
PSP. Ex
tr
actin
g
Faces a
nd
Facial Fe
ature
s
from Col
o
r
Images
. Int. J.
Pattern Re
cognit. Artif.
Intell
. 200
8; 22: 515-
534.
[4]
W
ang Z
,
Li S.
F
a
ce Reco
gniti
on usi
ng Skin
Co
lor Se
gme
n
tation a
nd T
e
mplate Match
i
ng
Algorithms
.
Inf. T
e
chnol. J.
2011; 1
0
: 230
8-23
14.
[5]
Hsu RL, Abdel-Mottaleb M, Jain
AK. Face Detectio
n in C
o
lor Imag
es.
T
r
ans. Pattern Anal. Mach.
Intell.
200
2; 24
: 696-70
6.
[6]
Ruiz-d
el-S
olar
J,
Quinteros
J. Illuminati
o
n
Compe
n
sati
o
n
an
d Norm
ali
z
ation i
n
Ei
ge
nspac
e-bas
e
d
F
a
ce Rec
ogn
iti
on: A Comp
ar
ative Stud
y
of Different Pre-
p
r
ocessi
ng Ap
pr
oach
e
s
. Pattern Recognit.
Lett.
2008; 29:
196
6-19
79.
[7]
Yang M, Cr
e
n
sha
w
J, Agu
s
tine B, Mare
ache
n R, W
u
Y. AdaBoost
-
base
d
F
a
ce
Detectio
n for
Embed
de
d S
y
s
t
ems.
Comp
ut. Visio
n
Ima
ge U
ndersta
ndi
ng
.
201
0; 114: 11
1
6
-11
25.
[8]
Yang J, Liu C,
Z
hang L. Col
o
r space n
o
rmalizat
i
on: Enh
a
n
cin
g
the Disc
r
iminati
ng Po
wer of Color
Spaces for F
a
c
e
Reco
gniti
on.
Pattern Recognit.
2010; 4
3
: 1454-
146
6.
[9]
W
ang Z
,
L
i
T
.
A F
a
ce
Detecti
on Syste
m
b
a
s
ed Sk
in
Co
lor
an
d N
eur
al
N
e
tw
ork
. Procee
din
g
s of t
h
e
Internatio
na
l
C
onfere
n
ce on
Comp
uter
Sci
e
nce and
Soft
w
a
re E
ngi
ne
erin
g. W
uha
n, Ch
i
na. 2
008;
pp:
961-
964.
[10]
Guo JM, L
i
n
C
C
, W
u
MF
, Ch
ang
CH,
Le
e
H. Co
mp
l
e
xi
ty
R
e
du
ce
d Fa
ce D
e
te
ctio
n u
s
ing
Prob
ab
i
l
i
t
y
-
based Face M
a
sk Pre
fi
lter
ing
and P
i
xe
l-bas
ed H
i
erarc
h
ica
l
-feature Ad
ab
o
o
sting.
IEEE Signal Proces
s
Lett.
2011; 18:
447-
450.
[11]
Chav
es-Gonza
l
ez JM, V
ega-
Rodri
g
u
e
z MA,
Gomez-
Pu
lid
o
JA, Sanc
hez-
perez
JM. D
e
tecting
Skin
i
n
F
a
ce Reco
gniti
on S
y
stems: A Colo
ur Spac
es
Stud
y
.
Di
gital Sign
al
Process
i
ng.
20
10; 20:
806-
823.
[12]
C Lin, Qin
X, Z
hu GL, W
e
i JH, Lin C. F
a
ce
Detectio
n Al
gorithm Bas
e
d
on Multi-or
ien
t
ation Gab
o
r
F
ilters and F
e
ature F
u
sio
n
.
T
E
LKOMNIKA Indon
esia
n
Journ
a
l
of Ele
c
trical Eng
i
ne
erin
g.
201
3;
11(1
0
):
59
86-
599
4.
[13]
W
ang YT
, Li W
B
, Pang S,
Kan JM. Se
gm
entatio
n
Meth
o
d
of L
i
ng
w
u
Lo
ng Ju
ju
bes B
a
sed o
n
L
*
a
*
b
*
Color S
pace.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(9): 5
344~
5
3
5
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.