TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 15, No. 2, August 201
5, pp. 301 ~
312
DOI: 10.115
9
1
/telkomni
ka.
v
15i2.837
3
301
Re
cei
v
ed Ap
ril 17, 2015; Revi
sed
Jul
y
2, 2015; Accept
ed Jul
y
18, 2
015
Recognition of a Face in a Mixed Document
Lhouss
a
ine Bouho
u
1
, Rachid El A
y
ac
hi*
2
, Mohamed Fakir
3
, Mohamed Ou
k
essou
4
1,2,
3
Computer S
c
ienc
es Dep
a
rtement, F
a
cult
y of Science a
n
d
T
e
chnolo
g
y
,
Sultan Mo
ul
a
y
Sliman
e Un
iver
sit
y
, Morocco
4
Mathematics
Dep
a
rtement, F
a
cult
y
of Scie
nce an
d T
e
chnolo
g
y
,
Sultan Mo
ul
a
y
Sliman
e Un
iver
sit
y
, Morocco
*Corres
p
o
ndi
n
g
author, em
ail
:
lhoussb
ou
h@
ya
ho
o.fr
1
, rachid.elay
a
chi@usms.ma
2
, fakfad@y
a
hoo.fr
3
,
ouk_m
oham
ed
@
y
ah
oo.fr
4
A
b
st
r
a
ct
F
a
ce reco
gniti
on is th
e fiel
d
of great i
n
tere
st in
the d
o
m
ai
ne of res
earch
for several
ap
plicati
o
n
s
such as
bi
o
m
etry ide
n
tificati
on, surve
ill
anc
e, an
d h
u
m
an
-mac
hin
e
i
n
ter
a
ction. T
h
is
p
aper
exp
o
ses
a
system
of face recognitio
n. This system
exploits an
im
age
docum
ent text embedding a c
o
lor hum
an fac
e
imag
e. Initial
l
y, the syste
m
, in
its phas
e of e
x
traction, exp
l
o
i
tis the h
o
ri
z
o
n
t
al an
d vertic
al
histogr
a
m
of t
h
e
document, det
ects the image which c
ontains the hum
a
n face
. The second task of t
he system
consist
s
of
detectin
g
th
e i
n
clu
ded
face
i
n
oth
e
r to
dete
r
mi
ne, w
i
th
th
e
hel
p
of inv
a
ria
n
ts mo
ments, t
he c
haracter
i
stics
of the face. T
h
e third and l
a
s
t
task of the system is
to determine, via the
same i
n
vari
an
ts moments, t
h
e
character
i
stics of each fac
e
s
t
ored i
n
a d
a
ta
base i
n
or
der to co
mp
are the
m
by
means
o
f
a classificati
o
n
tool (Ne
u
ral N
e
tw
orks and K nearest ne
ig
h
bors) w
i
th the one deter
min
ed in the sec
ond task for th
e
purp
o
se
of tak
i
ng
the
dec
isi
on
of i
dentific
ation
in
that
d
a
tabas
e, of t
h
e
most s
i
mil
a
r
face t
o
th
e o
n
e
detected i
n
the
input i
m
a
ge.
Ke
y
w
ords
:
his
t
ogra
m
, hu
mo
me
nts, Lege
nd
re mo
ment
s, n
earest ne
ig
hbo
r, neural n
e
tw
orks.
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
The use of biometri
cs h
a
s
two trend
s, the firs
t is to facilitate the lifestyle and the se
con
d
is to p
r
event
fraud.
Ho
wever, face re
co
gnition i
s
on
e
of the biom
e
t
ric te
chni
que
s. The
difficul
t
y
of face reco
gnition by compute
r
de
p
end
s on
wh
ether the
co
ndition
s for
acq
u
isitio
n. In a
controlled
en
vironme
n
t, paramete
r
s
su
ch as the
ba
ckgroun
d, the dire
ction
a
nd
intensity of light
sou
r
ces, the
angle of the
shot, and di
stance from
the ca
mera to the subj
ect
are pa
ram
e
ters
controlled
by the syste
m
. In an un
co
ntrolled e
n
vi
ron
m
ent, a se
rie
s
of pretre
atments a
r
e
often
requi
re
d p
r
ior to the
recog
n
ition itself. It must firs
t d
e
tect the
prese
n
ce
or ab
sen
c
e
of the fa
ce
in
the image. T
he face
sh
ou
ld then be
se
gmented.
Fin
a
lly, if we are
workin
g on
a video st
rea
m
,
the system
m
u
st tra
ck th
e face from on
e
image
to the
next. The pro
b
lem of face reco
gnition
ca
n
be formulate
d
a
s
follo
ws:
Given o
ne
o
r
mo
re im
ag
es
of a fa
ce,
the task i
s
t
o
find o
r
ve
ri
fy
a
person
'
s ide
n
tity by com
parin
g hi
s fa
ce to
the
e
n
tire fa
ce
im
age
s
store
d
in a
data
b
a
se.
Additional inf
o
rmatio
n, su
ch as ra
ce, ag
e, gende
r,
or
spe
e
ch ca
n b
e
use
d
to red
u
ce the
sea
r
ch
spa
c
e.
In ge
neral, a
fa
ce
re
cog
n
ition system us
ual
ly con
s
ist
s
of
seve
ral
step
s to
whi
c
h
o
u
r
system, Nowadays Biom
e
t
ric sy
stem
s
are in
crea
sin
g
ly used. Th
e
use
of the comp
uter a
n
d
its
ability to process and st
ore the data l
ed to the creation of
com
puteri
z
ed biometric sy
stems.
There are
se
veral uniq
ue
physi
cal c
haracteri
stics of an individual
su
ch a
s
finge
rpri
nt face, voice
recognitio
n
, DNA, etc. wh
ich explain
s
the diversit
y of systems ap
plying biomet
rics, as we take
into account.
Recently, the face re
co
g
n
ition is
attra
c
ting mo
re a
nd more attention. Secu
rity
remai
n
s the p
r
imary area o
f
application.
In this ar
ea of
the face reco
gnition is re
spon
sible for the
identificatio
n
and a
u
thenti
c
ation. Yan
g
and Al hav
e pro
p
o
s
ed
a cla
s
sificati
on of metho
d
s of
fac
i
al loc
a
liz
a
tion [1]:
a) "Kn
o
wle
d
g
e
-Ba
s
ed
Met
hod
s": The
s
e
metho
d
s to reco
gni
ze
different
point
s
o
f
intere
st
that make up
a face
an
d rel
a
tionship
s
be
tween th
em.
I
n
Chi
ang
et a
l
., the probl
e
m
in this type
o
f
method is tha
t
it is difficult to uniqu
ely de
fine a face [2].
b) "Feature in
variant app
ro
ach
e
s": They are to
use th
e element
s invariant to cha
nge
s in
illumination,
orientatio
n or expr
e
ssi
on
su
ch a
s
texture o
r
the
si
gnature color of the ski
n
for
detectio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 301 –
312
302
c) "Templ
ate
matching
m
e
thod
s": Feat
ure
s
m
odel
s of
a whol
e
fa
ce
or sub-part of the
face
(mo
u
th, eyes, n
o
se
)
are
created.
The lo
catio
n
is the
n
ba
se
d on th
e
correlation of th
e
s
e
model
s with the ca
ndid
a
te
s.
d) "Appea
ran
c
e-ba
sed met
hod
s”: The
s
e
are method
s that use the same p
r
in
cipl
e whi
c
h
wa
s introd
uced in the previous p
a
ra
g
r
aph ex
c
ept
that they have t
he advant
age to run v
e
r
y
quickly but require a long
period of tra
i
ning.
The m
e
thod
s in this cate
gory h
a
ve sho
w
n g
ood
results ag
ain
s
t 3
othe
r m
e
thod
s in
clu
d
e
am
ong
the
m
the
metho
d
ba
se
d
on
neural
netwo
rks
Ro
wley et al.,
the meth
od
of Schn
eide
rman a
nd Ka
n
ade
ba
sed
o
n
a
naive Ba
yes
cla
ssifie
r
and
the famous Vi
ola and
Jon
e
s
algo
rithm o
peratin
g in re
al time [3-6].
The meth
od
of Viola & Jo
nes i
s
a mo
re
efficient met
hod at
pre
s
e
n
t. This all
o
ws to
scan
an ima
ge, u
s
i
ng a
dete
c
tio
n
wi
ndo
w
of i
n
itial si
ze
by
24px 2
4px a
n
d
whet
he
r a
face
is presen
t.
Whe
n
the im
age h
a
s
bee
n
com
p
letely traversed,
the
wind
ow
si
ze i
s
in
cre
a
sed a
nd the
scanni
ng
is re
sta
r
ted,
until the wi
nd
ow
size ma
kes the im
age.
Incre
a
si
ng th
e win
d
o
w
si
ze take
s pl
ace
b
y
a multiplicatio
n factor of 1.2
5
.
As to sca
n
it, it con
s
ist
s
in
shifting the
wi
ndo
w by on
e
pixel. This
shi
ft can be
ch
a
nged to
spe
ed up th
e
pro
c
e
ss, b
u
t a shift of one
pixel ens
ures
maximum
a
c
curacy whi
c
h differentiate
s it
from others
[
7
]:
a) The u
s
e of
integral ima
g
e
s.
b) Sele
ction b
y
boosting fe
ature
s
.
c) Th
e co
mbi
nation bo
oste
d ca
scade of
cla
ssifie
r
s.
Selection
by
boo
sting i
s
to
re
place a
sin
g
le
cl
a
ssif
i
e
r
"st
r
on
g" sev
e
ral
"
w
ea
k" cla
ssif
i
e
r
s
in
cascad
e. Ho
wever, wit
h
a singl
e cla
ssifie
r
calle
d
"stron
g" shou
ld expe
ct that
the
classifie
r
is
analyzed the
entire wi
ndo
w to determine
wheth
e
r a face is pre
s
e
n
t in the image o
r
not.
A ca
scad
e of
cla
ssifie
r
s
wh
ose
sele
ction
crite
r
ia i
s
l
e
ss seve
re
may
repla
c
e th
e p
r
evious
stron
g
cl
assifi
er. Co
nsequ
e
n
tly, as so
on
as o
ne
step
s
believe
s that
there i
s
no fa
ce, the
wind
ow
is reje
cted a
n
d
the algorith
m
pro
c
ee
ds a
fter whi
c
h sav
e
s con
s
ide
r
a
b
le time.
A method
s
b
a
se
d o
n
n
eural net
works
can b
e
su
mm
a
r
ize
d
with the
use of
a
cla
s
sifier to
two outp
u
ts
repre
s
e
n
ting t
he p
r
e
s
en
ce
or a
b
sen
c
e o
f
the obje
c
t sought in
a
su
bregi
on of th
e
image
[8]. Th
e ba
si
c p
r
in
ci
ple i
s
to
sca
n
the im
age
with attention
fixed dim
e
n
s
io
ns
and
reali
z
e
the detectio
n
of the sub
-
picture win
d
o
w
.
There are oth
e
r meth
ods t
hat are i
n
tere
sted in
the
ch
ara
c
teri
stic p
o
ints of the fa
ce
su
ch
as the n
o
se, mouth and
eyes [8]. Th
e relative po
sition
s of different p
a
rts
o
f
the face are
detecte
d afte
r bein
g
stu
d
i
ed. Ho
weve
r it there'
s difficulty encou
ntered i
n
this approa
ch i
s
to
transl
a
te by strict rule
s to define ho
w th
e rese
a
r
cher
rep
r
e
s
ent
s the face. If these rul
e
s a
r
e t
o
o
spe
c
ific, they
do not reco
gnize some f
a
ce
s. Othe
rwise, they cau
s
e fal
s
e ala
r
ms. Ho
weve
r, it
allows a qui
ck focu
s on th
e intere
sting
parts of the i
m
age [9, 10].
The propo
se
d approa
ch i
n
this pape
r is base
d
o
n
the use o
f
two methods: The
detectio
n
me
thod for rem
o
ving skin se
ction
s
not
ha
ving the colo
r of the skin
and that of the
template mat
c
hin
g
for rem
o
ving se
ction
s
do not have
the facial fea
t
ures [8].
This pa
pe
r is
orga
nized a
s
follows:
a) Sectio
n pixels re
presenti
ng skin from
an input imag
e.
b) Filterin
g skin area
s.
c) Fa
ce d
e
tection by template matchin
g
"template matchin
g
".
Several stu
d
i
e
s have
bee
n based o
n
the detec
t
i
o
n
of
skin.
I
n
most
c
a
s
e
s,
ski
n is
rep
r
e
s
ente
d
by a portion
of a particul
a
r colo
r spa
c
e
.
Using the b
ound
arie
s of this regi
on a
s
threshold val
ues in an im
age, it is possible to
extra
c
t the pixels who
s
e colo
r can be li
kene
d to
that of the ski
n [11].
This is
done i
n
real time [1
2]:
a) The lo
catio
n
of all the face
s in the visual field,
b) Co
ntinue
d face
s ca
pture
the visual field to another,
c) Fa
cial
reco
gnition,
d) A backu
p of new faces i
n
a databa
se
e) Co
nst
r
u
c
tion of the ord
e
r of moveme
nt of the head
to fix a particular pe
rson.
The m
e
thod
of face
re
co
g
n
ition is to
co
nver
t ima
ges into lo
cal
bin
a
ry patte
rn
s,
divided
into seve
ral
sub
-
regio
n
s t
o
determine
the identit
y of a face
by com
pari
ng t
heir hi
stog
ra
ms
sub
r
egi
on
s.
The method of template or template m
a
tchi
ng, utilizes a compari
s
on of the int
ensity of
the pixels bet
wee
n
a prede
fined templat
e
and several
sub
-re
gion
s
of the image
to be analyze
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition o
f
a Face in a
Mixed
Do
cum
ent (Lho
ussai
ne Bouho
u)
303
This i
s
in
p
r
a
c
tice
to pe
rfo
r
m multipl
e
scan
s
cove
rin
g
the e
n
tire
area
of the i
m
age. T
he
most
con
d
u
c
ive to human fa
ce
s area
s will
therefo
r
e
e
a
sily identifie
d by the mi
nimum di
sta
n
ce
betwe
en the template an
d the und
erlying
image [8].
The L1 no
rm (Manath
a
n
distan
ce
) and t
he L2
norm (E
uclid
ean dist
ance
)
of the
stand
ard
s
a
r
e
used in o
ne
distan
ce p
o
ssible in this me
thod.
This
method
c
o
ns
is
ts
in performing the following s
t
eps
:
a)
Creation
o
f
the face m
odel
: Th
e mo
del will
be
a
gray level
im
age. Fig
u
re
1
sh
ows
example
s
of model
s [13]:
Figure 1. Examples of mo
d
e
ls of face
s,
(c) templ
a
te u
s
ed by this m
e
thod [13]
b) Search fo
r face
s in skin re
gion
s
: "Template M
a
tchin
g
" sin
c
e any image
and a
pred
efined m
odel, determi
ning the de
gree of membe
r
shi
p
(simila
ri
ty) of this model to the image [14].
c) Simila
rity measures
: T
here
are sev
e
ral
ways to
meas
ure the
simila
rity betwee
n
two
image
s. Th
e f
i
rst m
e
a
s
u
r
e
applie
d by thi
s
m
e
thod
is a
sim
p
le
difference of
sta
n
d
a
rd
data
by t
he
first formul
a for the L1 no
rm, and the se
con
d
formul
a for the L2 no
rm.
Norm L1 :
∑
|
,
,
|
,
(1)
Norm L2 :
∑
,
,
,
(2)
d) Ap
plication
of the
metho
d
of the
"Tem
pl
ate Mat
c
hin
g
" whi
c
h
requ
ires the
pa
ssage
by
the followin
g
step
s:
1) Extraction
of the image gray level,
2) No
rmali
z
at
ion of the seg
m
ent,
3) Co
mpa
r
iso
n
of the segm
ent with the model,
4) De
ci
sion.
A method
of l
o
catin
g
an
d e
x
tracting fa
ci
al f
eatures u
s
ing both
shap
e an
d colo
r [15] was
prop
osed by
Sobottka a
n
d
Pitas [16]. This meth
od
consi
s
ts in th
e
segm
entatio
n of the imag
e in
the HSV
colo
r spa
c
e to l
o
cate the regio
n
of the sk
in. A
nd by a
pplyin
g
a regio
n
g
r
owin
g alg
o
rit
h
m
with a
coa
r
se
re
solution, t
he conn
ecte
d
com
pon
ents are th
en d
e
termin
ed. Fo
r
each conn
ect
ed
component, the algorithm
adjusts an
ellipse in
order to determi
ne the
candi
date area
which
corre
s
p
ond
s t
o
the face. Fi
nally, a more
detailed
anal
ysis of featu
r
es
within thi
s
regio
n
lea
d
s to
the con
c
lu
sio
n
about the p
r
esen
ce of a face o
r
not [17
]
.
In [18, 19]
a
method
is u
s
ed to
cla
s
sify t
he pixel
s
of skin
col
o
r b
a
se
d o
n
a
G
aussia
n
model fo
r the
ski
n colo
r. A set ofelevn l
o
we
r-or
d
e
r
g
eometri
cal m
o
ment i
s
calculated u
s
in
g the
Fouri
e
r tra
n
sf
orm an
d the radial Melli
n tran
sform
to chara
c
te
rize the sh
ape of "cluste
r
s" in the
binari
z
e
d
image. To dete
c
t the face region, a
neu
ral network i
s
traine
d usi
ng the extra
c
ts
geomet
rical moment
s [17].
In this pape
r Figure 2 illust
rat
ed the recognition
syste
m
adopted.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 301 –
312
304
Figure 2. Face recognitio
n
system
2. Extrac
tion
T
h
e
e
x
tr
ac
tion
s
t
e
p
is
the k
e
y s
t
ep
in
t
he process,
becau
se t
he perfo
rman
ce of
the
whol
e sy
stem
depe
nd
s on
i
t. In this step
also
kn
own a
s
ind
e
xing o
r
modelin
g, is
extracted
fro
m
the face ima
ge info
rmatio
n that
can
m
odel th
e fa
ce
of a
pe
rson
by a m
easurement ve
ctor that
cha
r
a
c
teri
ze
s (feature ve
ctor or
sign
ature).
The ju
dici
ou
s choi
ce
of the
extractio
n
m
e
thod i
s
extre
m
ely impo
rta
n
t insofa
r
as t
he next
step
(cl
a
ssifi
c
ation)
will be made only
on the
basi
s of the values of th
ese vectors. Extract
i
on
method
s are nume
r
ou
s, a
m
ong them,
Zerni
k
e
mo
m
ents, Hu a
nd
Lege
ndre are
mentioned.
2.1. Zernike
Moments
Zerni
k
e
mom
ents
are ofte
n u
s
ed
to
ca
p
t
ure th
e gl
ob
al featu
r
e
s
of
an im
age
re
cognition
and im
age
a
nalysi
s
. Intro
duced fo
r the
first time
in
comp
uter visi
on by T
eag
u
e
[20], this shape
descri
p
tor
ha
s p
r
oven it
s superi
o
ri
ty ove
r
othe
r fun
c
tions
mome
nts in its a
b
ility descri
p
tion a
n
d
robu
stne
ss t
o
noi
se
an
d
disto
r
tion. V
e
ry re
cently, many
re
sea
r
ch
er
s have looked at
th
ese
moment
s, ma
inly to optimize comp
utatio
n time and im
prove a
c
cura
cy.
We d
enote
b
y
Zerni
k
e mo
ments a
se
ri
es of
cal
c
ulat
ions
used to
transfo
rm a
n
image
into a vector
of real co
mpo
nents rep
r
e
s
e
n
tative mome
nts Zij [21].
The m
a
in
ad
vantage th
at
this Z
e
rni
k
e
moment
s a
r
e
insen
s
itive to tra
n
slatio
n, rotation
and scali
ng which p
r
e
s
e
r
ve
s the inform
ation co
ntaine
d
in the image
s [21].
Mathemati
c
al
ly, Zernike m
o
ments a
r
e d
e
fined with a
per o
r
de
r and
repetition q o
n
:
D =
,
|0
∞
,
|
|
,
|
|
:
∬
|
∗
,
,
(
3
)
Whe
r
e
V
∗
denotes the compl
e
x conju
gate
of
V
himself de
fined as:
,
.
(4)
And,
∑
!
!
!
!
|
|
|
|
(5)
From the Eq
uation (1) a
n
d
(2),
we can
expr
e
ss Z
e
rnike m
o
ment
s of a rotate
d
image of
an angl
e
α
around its o
r
igi
n
in polar
coo
r
dinate
s
:
∝
∝
(6)
The Eq
uation
(3
) p
r
ove
s
t
he inva
rian
ce
rotation
of
Zerni
k
e
mom
ents m
odul
e
sin
c
e |
Z
pq
e
iq
α
| =.
|
Z
pq
| Thanks to the ortho
gonality pro
p
e
rty,
image reco
nstructio
n
can be
sim
p
ly
expre
s
sed a
s
the sum of e
a
ch b
a
si
c fun
c
tion we
ighte
d
by the corre
s
po
ndin
g
Zernike mo
ment:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition o
f
a Face in a
Mixed
Do
cum
ent (Lho
ussai
ne Bouho
u)
305
,
~
∑∑
,
,
∈
(7)
2.2. Hu Moments
In [22], from
the g
eomet
rical
mom
ent
s,
Hu [23]
p
r
opo
se
d a
set of seven
moment
invariant
s to
transl
a
tion,
rotation an
d scaling.
T
hey are
wid
e
ly used in
the literature
for
descri
b
ing fo
rms for a cl
assificatio
n
or i
ndexing,
but
are quite
sen
s
itive to noise. Moreove
r
, this
family of descriptors is neit
her o
r
thog
on
al nor compl
e
te.
The form
ula
of the mome
nt of Hu m
pq
orde
r p
+ q o
f
a distrib
u
tio
n
functio
n
f (x, y) is as
follows
:
,
∞
∞
∞
∞
(8)
For a digital i
m
age g (x, y) of size M * N,
the above formula be
com
e
s:
∑∑
,
(9)
The ce
ntral m
o
ment
μ
pq
is given by:
∑∑
ẋ
ẏ
)
,
a
ẋ
,
ẏ
=
(10)
The no
rmali
z
ed ce
ntral mo
ment is defin
ed:
ƞ
,
,
2,3,
…
(11)
2.3. Legendr
e Moments
Lege
ndre mo
ments have
been introdu
ced by
Tea
g
ue [24]. The
y
have been
used in
many a
pplica
t
ions
of p
a
ttern
re
cog
n
itio
n [25].The
times for a
discrete
L
ege
nd
re im
age
M
x N
pixels
with int
ens
ity func
tion f (x, y) is
[2
6]:
L
ƛ
∑∑
P
x
P
y
fx,
y
(12)
Whe
r
e
ƛ
∗
and
denote the normali
ze
d pixel coordi
nates in the range [1, 1],
whi
c
h are giv
en
by:
(13)
is
the
-o
rd
er Leg
end
er
polynomial d
e
fined by:
∑
!
!
!
(14)
And, the recu
rre
nt formula
of the Legen
d
r
e polyno
m
ial
s
is:
,
1
(15)
In whi
c
h
wo
rk th
e recurrent form
ula
is u
s
e
d
for
cal
c
ulatin
g L
egen
dre
poly
nomial
s
in orde
r to increa
se the comp
utation
spe
ed.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 301 –
312
306
3. Classifica
tion
The cla
s
sifica
tion is the assignm
ent of a spe
c
if
ic cl
a
ss o
r
not
a giv
en f
a
ce cla
s
s:
cla
ss
here
is
a pe
rson with fa
ce
i
m
age
s in th
e
databa
se
and
the non
-cla
ss in the
case
whe
r
e th
e face
belon
gs a
n
y cla
ss. Thi
s
all
o
catio
n
req
u
ires the intro
d
u
c
tion of a me
asu
r
e of simil
a
rity.
3.1. K-Near
e
s
t Neig
hbor
K-Ne
are
s
t
Ne
ighbo
r
(K-NN) is pa
rt of
su
per
vised l
earning
whi
c
h
h
a
s
bee
n
u
s
e
d
in ma
ny
appli
c
ation
s
inclu
d
ing dat
a
mining, statistica
l pattern
reco
gnition a
n
d
image p
r
o
c
essing.
The
al
gorith
m
doe
s not construct a cla
ssifi
cati
on
mo
del but it i
s
b
a
se
d on
valu
es fo
und
in the sto
r
ag
e or m
e
mo
ry. To identify the cl
ass
of a
n
input, the a
l
gorithm
sel
e
cts the
cla
s
s
to
whi
c
h the maj
o
rity of the k
nearest nei
gh
bors belo
ng e
n
try.
The K-nn al
gorithm is
co
nsid
ere
d
one
of the simplest machine
learni
ng algo
rithms.
Ho
wever, it is co
mputat
io
nally expen
sive espe
cially
when the
size of the training set be
co
me
s
large
whi
c
h would cau
s
e th
e cla
ssifi
catio
n
task to be v
e
ry slo
w
.
The followi
ng Figure 3 illustrates
the modeling of this
algorithm.
Figure 3. K-NN
3.2. Neural Net
w
o
r
k
s
It is a set of interconn
ecte
d neu
ron
s
for formal resol
u
tion of co
mp
lex probl
ems
su
ch a
s
pattern recog
n
ition or natu
r
al langu
age p
r
ocessin
g
, by adjustin
g
wei
ghts in a lea
r
ning ph
ase.
The
ope
ratio
n
of th
e n
eural net
wo
rk ba
sed
on
the
fu
nctioni
ng
of b
i
ologi
cal
neu
r
ons an
d
is e
m
bo
died
i
n
a
form
of
a
co
mpute
r
al
gorithm
.
N
eu
r
a
l net
wo
rk
s i
s
kno
w
n
f
o
r
it
s
st
re
ngt
h i
n
t
he
field of face reco
gnition a
p
p
roa
c
h.
3.2.1. Biological Neuron
s
In very
simp
listic, a
ne
uron i
s
a
biol
ogica
l
cell th
at is ch
aract
e
rized
by
synap
s
e
s
,
dend
rites, ax
ons a
nd core.
The biologi
ca
l
living cell
s are
neu
ro
ns whi
c
h
a
r
e in
te
r
c
o
n
ne
c
t
ed
b
y
lin
ks
c
a
lled
a
x
o
n
s
[27] leading the elect
r
ic
si
gnal
s
from the output of a neuro
n
to
the input (syn
a
p
se
) of anoth
e
r
neuron.
3.2.2. Artifici
al Neuron
A neural net
work con
s
ist
s
of simple ele
m
ents, ope
rat
i
ng in parallel
,
which were inspi
r
ed
by the bi
olo
g
ical
ne
rvou
s
system. It
co
nsi
s
ts of
a
weig
hted
dire
cted
g
r
aph
wh
ose
node
s
symboli
z
e
th
e form
al n
e
u
r
ons.
The
s
e
h
a
ve an
a
c
tivation fun
c
tion t
o
influe
nce ot
her ne
uro
n
s i
n
the netwo
rk. Con
n
e
c
tion
s betwe
en ne
urons, the
so
-c
alled syn
a
ptic links, p
r
o
pag
ate the activity
of neuro
n
s
wi
th characte
ri
stic
weig
hting
of the conn
ection.
It may cause a neural network fo
r a spe
c
ific j
ob (fo
r
e
x
ample, cha
r
acter
re
cognit
i
on) by
adju
s
ting the
values of the
con
n
e
c
tion
s (or
wei
ght) b
e
twee
n the ele
m
ents (neu
ro
ns).
The formal n
euro
n
is an e
l
ementa
r
y unit of a neural netwo
rk. It perform
s the
weig
hted
sum of its inp
u
ts, as the va
lue of this su
mmation
is
compa
r
ed
with a thresh
old. The output of
the
neuron is:
∑
.
(16)
Y
(17)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition o
f
a Face in a
Mixed
Do
cum
ent (Lho
ussai
ne Bouho
u)
307
Whe
r
e,
: element of the input vect
or,
i=
1...n: entries
formal neuron,
y: output,
: weighting p
a
ram
e
ters,
f: ac
tivation f
unc
tion.
4. Results a
nd Analy
s
is
The
obje
c
tive of this
se
ction i
s
to p
r
e
s
ent
th
e
re
co
gnition
sy
ste
m
(who
se
st
eps a
r
e
sho
w
n
in
Fig
u
re
4
ado
pte
d
for the
re
cognition
of a
face
in
a te
xt docu
m
ent,
and
the
re
sults
found in ea
ch
phase of the system.
Figure 4. Re
cognition
syste
m
The d
o
cume
nts p
r
e
s
ente
d
in Fig
u
re
5
show
t
w
o
exa
m
ples treate
d
by ou
r sy
ste
m
usi
n
g
an interfa
c
e
develop
ed in
MATLAB, wh
ich offers
different me
nu t
o
cho
o
se, for each te
st, both
the descri
p
tor to be used in
extraction le
vel and the
cl
assifier to use in the cla
ssi
fication level.
(a)
(b)
Figure 5. Examples of do
cuments p
r
o
c
e
s
sed by our
system
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 301 –
312
308
4.1. Data
Ba
se
There
was a databa
se of
20
faces (Fig
ure
6
)
spread
over 5
different cla
s
se
s o
f
people
with 4 fa
ce
varied, d
epe
nding
on th
e facial
expression
s of t
h
e pe
rson a
n
d
the dista
n
ce
cha
r
a
c
teri
zin
g
the image o
f
the same pe
rso
n
at the time from hi
s g
r
asp.
Figure 6. Dat
aba
se u
s
ed b
y
the develop
ed syste
m
This data
b
a
s
e is tran
sform
ed, durin
g the
tr
eatment, two referen
c
e b
a
se
s, one for
the Hu
Moment
s de
scripto
r
, and the other by th
e Lege
ndre moment
s de
scripto
r
.
4.2. Face Re
cognition Sy
stem
To describe t
h
is sy
stem, we pre
s
ent he
re different ph
ase
s
.
Phase 1: Sel
ection o
f
a d
o
cumen
t
Dete
cting a f
a
ce in
a do
cument is
as
easy
without
hypothe
s
is. T
h
is i
s
why o
u
r
syste
m
will co
nsi
der t
he hypothe
si
s of a text docume
n
t
conta
i
ning a si
ngle
image (witho
ut the pre
s
en
ce,
in the pi
ctu
r
e
,
other fig
u
re
s that
ca
n di
sru
p
t
sy
stem
perfo
rma
n
ce
). So
we
coll
ected
differe
nt
document gl
ad both text
and an imag
e containi
ng
a
face to sel
e
ct (Fig
ure 7
)
one of these
document
s
will be the main elem
ent of
whi
c
h
came
from the sy
stem m
u
st
extract the face
portion to be
sub
s
eq
uentl
y
compa
r
ed
with those
in
the databa
se to identify the perso
n in
the
input image.
Figure 7. Example of sele
cting a docum
ent
Phase 2: Ima
g
e extr
a
c
t
ion
This ph
ase was carried o
u
t in two step
s:
a) Step
1:
Th
e verti
c
al
hist
ogra
m
of
the
doc
ument
ha
s all
o
wed
us
to dete
rmine
the first
and last refe
ren
c
e
s
and
pixel (vertica
lly) of the im
age in the do
cume
nt, as illustrate
d in
Table 1, to d
e
termin
e the extraction of t
he first or
i
g
in
al document
wa
s inde
ed a
b
le to extract,
o
f
the document
, by these two part unlimit
ed point
s
an
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition o
f
a Face in a
Mixed
Do
cum
ent (Lho
ussai
ne Bouho
u)
309
b) Step
2: b
a
s
ed
on
the
p
o
rtion
of mat
e
rial
extra
c
te
d in
step
1, t
he h
o
ri
zontal
histog
ram
has
allo
wed
us to
determi
ne the first a
nd la
st refe
rence pixel
and
(horizontally) in the
image
of the
do
cum
ent a
s
it i
s
sho
w
n in
T
abl
e 1
to dete
r
mine
the extra
c
tio
n
of th
e
se
cond
extraction
p
r
e
v
ious i
nde
ed
it has b
een
e
x
tract
ed,
of the d
o
cument,
by the
s
e
two
pa
rt unlimite
d
point
and
.
Table 1. Step
s of extractin
g
the image o
f
the docume
n
t
This ph
ase al
lows us, therefore, a
s
sh
o
w
n in Figu
re
8, to define the image a
r
ea
contai
ning th
e face, from the initial do
cu
ment for use in the followin
g
pha
se
s.
Figure 8. Extraction of the
document im
age
Phase 3: F
a
c
e
dete
ction i
m
age
This
p
h
a
s
e starts with
a pretreatment on
the extra
c
ted image i
n
the previo
us pha
se
contai
ning th
e face
dete
c
ting to dete
r
mi
ne the
cont
o
u
r of the fa
ce
followe
d by a pro
c
e
s
sing
for
locatin
g
the f
a
ce
portio
n
o
f
the non-fa
ce part
(F
igu
r
e 9) to extra
c
t only the fa
ce o
n
which
the
system
will build in the ne
xt phase.
Figure 9. Face Dete
ction
Phase 4: Ex
traction o
f
fa
cial features
Wa
s u
s
ed in
this pha
se, th
e time and th
e moment of
Hu Le
gen
dre
(detaile
d in
Cha
p
ter
2) to obtain, on the one h
and, the faci
al featur
e
s
d
e
tected in th
e previou
s
p
hase, and create,
More
over, for each mo
me
nt, a refe
ren
c
e datab
ase,
for all th
e ima
ges
of the d
a
t
abase. Each
of
these
ba
se
wi
ll be op
erat
e
in the next p
hase de
pen
di
ng on th
e
ch
oice
of cla
s
si
fier u
s
ed i
n
this
pha
se.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 15, No. 2, August 2015 : 301 –
312
310
Phase 5:
Cla
ssifica
tion
Neu
r
al net
work, first req
u
ire
s
an initi
a
l training
step whi
c
h wil
l
be defined
in the
con
s
tru
c
tion
of the multilayer st
ru
cture of t
he neu
ral net
wo
rk
(the numb
e
r
of inputs
of the
netwo
rk d
epe
nds o
n
the si
ze of the feature ve
cto
r
, the numbe
r cell
s of the hidde
n layer and t
h
e
numbe
r
of d
e
sired
output
), for ou
r
sy
stem, the
s
e
inputs 7
indi
cate
s
cha
r
a
c
teristics of
Hu
moment
s, the
cell
s 10
and
hidde
n layer
20 corre
s
po
n
d
ing to the
n
u
mbe
r
of out
puts in th
e b
a
se
face data
Fig
u
re 1
0
. Lea
rn
ing end
s
whe
n
the algo
rith
m rea
c
he
s th
e stop
crite
r
io
n defined
by our
system in 10
00 iteration
s
.
Figure 10. Phase le
arni
ng
neural net
Once the
ne
twork h
a
s
accepta
b
le p
e
rf
orma
nce, it can
be
used
for recogniti
on of a
possibl
e pha
se in a secon
d
step, this p
hase of test
in
g, by Hu Mo
ments of
the
facial feature
s
on
the ima
ge t
hose of
the
output
of t
he n
e
two
r
k t
o
ma
ke
a
d
e
ci
sion
on
t
he
re
sult of
the
cla
ssif
i
cat
i
on.
We also u
s
e
d
the de
script
or Le
gen
dre
Moment
s to o
r
de
r 3, whi
c
h
allows u
s
to h
a
ve 10
feature
s
for e
a
ch fa
ce of the datab
ase (and th
e st
ru
cture
of the n
eural
netwo
rk, in this ca
se,
is
10 entrie
s
ind
i
cating
cha
r
a
c
teri
stics of L
egen
dr
e m
o
ments, 10
cel
l
s hidd
en lay
e
r and 2
0
out
puts
corre
s
p
ondin
g
to the num
ber of fa
ce
s
in the dat
a
b
a
se
) that will
be teste
d
with those
of the
detecte
d face
to be there too, a deci
s
io
n on t
he cla
s
sif
i
cat
i
o
n
re
su
lt
s.
Experimental
results obtain
ed by th
e u
s
e
of
Hu m
o
me
nts a
nd
mom
ents
of the
Le
gend
re
at
extractio
n
pha
se attribu
t
e,
and neu
ral
net
wo
rks
and th
e K-ne
are
s
t nei
ghb
ors (k-nn
)
at
the
cla
ssifi
cation
pha
se a
r
e ill
ustrate
d
in
T
able 2
sh
ows the re
co
gniti
on rate (RR) and E
rro
r
Rate
(ER)
cal
c
ulat
ed for differe
nt appro
a
che
s
to a databa
se of 20 face
s.
Table 2. Experime
n
tal Re
sults
Descr
iptor
Neural
Net
w
orks
K-NN
Euclidean Manhattan
R.R
E.R
R.R
E.R
R.R
E.R
Hu Moments
70%
30%
65 %
35%
60 %
40%
Legendre
Mome
nts
72%
28%
68%
32%
65%
35%
Several
simil
a
rities have
b
een
pro
p
o
s
e
d
du
ring
the
experim
ents for th
e
cla
ssifi
cation
by
the K-n
earest neighb
ors a
nd only
simil
a
rities wi
th
th
e Eucli
dea
n
distan
ce
and
the Ma
nhatt
an
distan
ce
were sele
cted for the continu
a
tion of our p
r
oj
ect.
We did
som
e
tests o
n
the image
s we
hav
e taken
a perso
nal p
hoto-p
a
irs an
d other
download
ed f
r
om th
e
Net.
The te
sts are
pe
rforme
d o
n
faces that
come
in
different
colo
rs a
n
d
different ori
e
n
t
ations an
d scale
s
. We n
o
te in:
a) Figu
re 5(a) positive rec
ognition be
ca
use, for ea
ch
of these thre
e ca
se
s, the deci
s
io
n
is co
rrect be
cause ea
ch pe
rso
n
to be ide
n
tified is in the databa
se.
b) Figu
re 5(b), gain positive
reco
gnition b
e
ca
use
in this case the de
cision is al
so correct
that the perso
n to be identifi
ed is not in the datab
ase.
Evaluation Warning : The document was created with Spire.PDF for Python.