Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 4
,
A
ugu
st
2016
, pp
. 16
10
~
1
616
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
4.9
761
1
610
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
Face Recognition with Modul
ar Two Dimensional PCA under
Uncontrolled Illumination Variations
Venk
atr
a
m
a
p
h
anik
umar
S,
K. V. Krishn
a
Kish
ore
Department o
f
C
S
&E, Vign
an’s
Foundation for
Scie
n
c
e, Techno
log
y
and R
e
sear
ch, Guntur
, India
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Dec 21, 2015
Rev
i
sed
Jun
10,
201
6
Accepted
Jun 26, 2016
Person authen
ticaton using f
a
ces became
one of
the most popular secu
r
ity
approach
es for the last thr
ee d
e
cades.
From th
e liter
a
ture it is
found that
perofrm
ance of
m
o
st of the m
e
t
hods used in
recognition w
a
s li
m
ited due t
o
uncontrolled co
nditions like illu
mination
and pose variations. In this work,
to address the limitations of uncont
rolled environment, Modular
two-
dimensional Principle Component Analy
s
is
(M2D-PCA) is prop
osed. In
this
approach
, th
e in
put im
age is par
tition
e
d into fou
r
equal segm
ent
s
and the
n
Histogram
Equa
liz
ation
is appl
ie
d to redu
ce
illu
m
i
nation im
pac
t
caused du
e
to var
y
ing
ligh
t
ening
conditio
n
s.
Then M2D
-
PCA algorithm is applied
paral
l
el on e
ach
s
e
gm
ent and then all f
eatur
es
e
x
trac
ted from
the s
e
gm
ents
are fus
e
d
with
wieghted s
u
m
m
a
tion
.
Exp
e
rim
e
nts
are
carr
i
ed o
u
t on ben
c
h
m
a
rk datase
ts l
i
ke ex
tend
ed Yale d
a
tabase B
,
ORL and AR
datab
a
se.
Results of th
e p
r
oposed approach produced
goo
d recogn
ition
rate with
low
com
putation
a
l
ti
m
e
against
var
i
o
u
s illum
i
na
tion
e
nvironm
ents.
Keyword:
Face Recognition
H
i
stog
r
a
m
Eq
ualizatio
n
PCA
LDA
M2
D-
PCA
Copyright ©
201
6 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
:
K.
V.
Kriah
n
a
K
i
sho
r
e,
Depa
rt
m
e
nt
of
C
S
&
E,
VFSTR
Un
iv
ersity,
Gu
nt
u
r
,
I
ndi
a.
Em
a
il: k
i
sh
o
r
ek
vk_
1@yahoo
.co
m
1.
INTRODUCTION
Pers
on a
u
t
h
e
n
t
i
cat
on usi
ng
fa
ces i
s
one
of t
h
e
m
o
st
po
pul
a
r
bi
om
et
ri
c based secu
ri
t
y
sy
stem
s used i
n
identification and veri
fication.
Biom
etric
recognition sy
ste
m
using fac
e
tr
ait is
to recognize pe
rs
on faces
ag
ain
s
t in app
licatio
n
s
lik
e cred
it card
v
e
rification
,
crimin
al id
en
tificatio
n
and
su
rv
eillan
ce sy
ste
m
s.
Perform
a
n
ce o
f
ex
isting
system
s
is lo
w in
reco
gn
itio
n rate
d
u
e
t
o
m
u
ltip
le
v
a
riation
s
(Pose and
Illu
m
i
n
a
tio
n)
[1, 4].
Face re
cognition beca
me a diffi
cult
problem
due to si
m
ilar shape
of
t
h
e faces and m
u
ltiple variations
am
ong t
h
e i
m
ages o
f
t
h
e s
a
m
e
perso
n
.
A bi
om
et
ri
c appl
i
cat
i
o
n i
s
m
a
i
n
l
y
of t
h
r
ee pha
ses;
t
h
e
s
e are
Enr
o
l
m
ent
,
Veri
fi
cat
i
on an
d I
d
ent
i
f
i
cat
i
o
n. In t
h
e en
r
o
l
m
ent
phase
, t
h
e sy
st
em
ext
r
act
s i
n
f
o
rm
at
i
on about
t
h
e
pers
o
n
t
o
be i
d
ent
i
f
i
e
d by
m
e
asuri
ng s
o
m
e
charact
eri
s
t
i
c
s.
Veri
fi
cat
i
o
n o
f
t
h
e pers
on i
s
kn
o
w
n as o
n
e
t
o
one
m
a
t
c
hi
ng a
n
d i
t
sim
p
l
y
answe
r
s
whet
her
t
h
e
pers
o
n
i
s
ge
nui
ne
or
n
o
t
.
Ide
n
t
i
fi
cat
i
on i
s
o
n
e
-
t
o
-m
any
m
a
t
c
hi
n
g
,
in
wh
ich
th
e syste
m
selects t
h
e
b
e
st t
h
at match
e
s th
e
test sam
p
le. Th
ere are
so
m
e
k
e
y
p
a
ram
e
ters th
at are
use
d
to m
easure the system
perform
a
nce such a
s
re
c
o
g
n
i
t
i
on rat
e
a
nd t
i
m
e
co
m
p
l
e
xi
t
y
. There a
r
e
m
a
ny
approaches in
the past th
at le
ads to the de
velopm
ent of successful
face
recognition syste
m
s. They include:
PCA [1], L
D
A [2], 2D-PC
A
[3], F
ace
Analysis for Commercial Entities
(FACE
)
[4], and
Support Vector
Machine
(SVM) [5]. Hence
,
the a
b
ove
m
e
thods a
r
e s
u
cce
ssful t
o
a ce
rta
i
n exte
nt,
but t
h
ey are lim
ited
to the
g
a
llery of im
ag
es th
at are tak
e
n
in con
t
ro
lled
en
v
i
ron
m
en
ts.
Thi
s
pa
per i
n
t
r
o
duces a Fa
ce R
ecog
n
i
t
i
on Sy
st
em
t
h
at
wor
k
s ef
fi
ci
ent
l
y
unde
r u
n
co
nt
r
o
l
l
e
d
en
v
i
ron
m
en
t wh
ich ov
erco
m
e
th
e effects of
v
a
ri
n
g
illu
m
i
n
a
tio
n
.
In th
is
work, a wei
g
h
t
ed
fu
si
on
ap
pro
ach i
s
use
d
t
o
fu
se ext
r
act
ed f
aci
al
feat
ure
s
fr
om
m
odul
es of fac
e
. Ob
ject
i
v
e o
f
t
h
i
s
m
e
t
hod i
s
defi
ne
d t
o
e
n
hance
the efficiency
of the
syste
m
by ex
t
r
act
i
n
g t
h
e feat
ures
pa
ral
l
e
l
s
on se
ge
men
t
s o
f
faces in
recogn
izing
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
ce Recogn
itio
n with
Mo
du
l
a
r Tw
o
Dimensio
na
l PCA
und
er
Uncon
t
ro
ll
ed
.... (Ven
ka
t
r
a
m
ap
han
ikuma
r
S
)
1
611
persons unde
r varying illumi
nations. In this approac
h
fi
rst face i
m
age
has to be porti
one
d into four equa
l
p
a
rts. Th
en
p
r
e-pro
c
essing
is d
o
n
e
on
all p
a
rtitio
n
s
in
d
e
p
e
nd
en
tly. Th
en
2D-PC
A
algo
rit
h
m
with
map
red
u
ce
app
r
oach i
s
us
ed t
o
e
x
t
r
act
f
eat
ures
o
n
eac
h o
f
t
h
e
i
n
di
vi
dual
p
o
rt
i
o
ns.
Sco
r
e n
o
rm
al
izat
i
on i
s
per
f
o
r
m
e
d on
the ext
r
acted
feature set t
o
bring all
the
feat
ures
into com
m
on scale. Propos
ed m
odel
i
s
eval
uat
e
d
on
e
x
t
e
n
d
ed
Yale face d
a
tab
a
se B
un
d
e
r
varian
t illu
m
i
n
a
t
i
o
n
con
d
ition
s
.
2.
PROP
OSE
D
METHO
D
2.
1.
Sys
t
em Fr
ame
w
ork
The m
a
in aim
of the propo
se
d work is to e
n
hance t
h
e efficiency
of
recognition system
for the facial
i
m
ag
es th
at are tak
e
n und
er v
a
ring
illu
m
i
n
a
ti
o
n
cond
itio
n
s
.
M2
D-PCA is a p
a
rallel alg
o
ri
th
m
with
lin
e
b
a
sed
lo
cal
characteristic approac
h
. T
h
e
arc
h
itecture
of
robust face rec
o
gnition
unde
r illu
m
i
nation changes is gi
ve
n in Figure 1.
The input im
a
g
e is partitione
d into
fo
ur e
q
ual
pa
rt
i
t
i
ons by
c
o
nsi
d
eri
ng t
h
e m
e
an pi
xel
o
f
t
h
e i
m
age, i
.
e.,
LU
,
LL, R
U
an
d R
L
im
ages. T
h
e
n
, al
l
th
e fo
ur eq
u
a
l
seg
m
en
ts o
f
t
h
e im
ag
e will u
n
d
e
rgo
Histog
ram
Eq
u
a
lizatio
n
(HE)
p
h
a
se o
f
m
a
p
p
i
n
g
as p
e
r
m
a
pred
uce a
p
p
r
oac
h
.
I
n
next
pha
se, M
2
D
-
P
C
A i
s
ap
pl
i
e
d
t
o
e
x
tract the
fe
atures. T
h
en, the
norm
alization is
applied t
o
get
norm
alized features set. The
n
the f
eat
ures
are fused
(re
duced) with t
h
e
weighte
d
sum
m
ation
o
p
e
r
a
tion
.
Near
est n
e
i
g
hbou
r
classif
i
er
i
s
u
s
ed
f
o
r
i
d
en
tif
icatio
n
of p
e
rson
s. Map
r
edu
ce ap
pr
oach
is
i
m
p
l
e
m
en
ted
with
a cl
u
s
ter of
4
CPUs
with
3
2
GB R
A
M.
.
Fi
gu
re
1.
A
r
chi
t
ect
ure
of t
h
e
p
r
o
p
o
sed
m
e
t
hod
wi
t
h
m
a
p re
d
u
ce a
p
p
r
oach
2.
2.
Histog
ra
m Equa
liza
t
io
n
In i
m
age proc
essi
ng
, o
b
ject
i
v
e o
f
t
h
e hi
st
o
g
ram
equal
i
zat
i
on i
s
co
nt
rast
enha
ncem
ent
of t
h
e
gi
ve
n
i
m
ag
e. Histogram
Eq
u
a
lizati
o
n read
ju
sts t
h
e o
r
i
g
in
al
h
i
stog
ram
to
i
m
p
r
ov
e
q
u
a
lity o
f
the i
m
ag
e b
y
ch
an
g
i
ng
p
i
x
e
l
g
r
ay levels. By th
e un
ifo
r
m
d
i
stribu
tio
n of
p
i
x
e
l
in
ten
s
ity values, lin
ear cu
mu
lativ
e
h
i
sto
g
ra
m
is
g
e
n
e
rated
[6
].
Fo
r an
im
ag
e I(x, y) with
‘K’
n
u
m
b
e
r
o
f
d
i
screte g
r
ay
v
a
lu
es, prob
ab
ility o
f
o
c
cu
rren
ce of g
r
ay
lev
e
l l is d
e
fin
e
d
as:
(1
)
whe
r
e
k =
0,
1
…
l
-
1
gr
ay
scal
es an
d N i
s
t
h
e num
ber
of
pi
xel
s
o
f
a
n
i
m
age.
New
i
n
t
e
ns
i
t
y
gray
-l
evel
of t
h
e
pi
xel
i
s
defi
ne
d as
I
out.
In
p
ut
/
Test
Im
a
g
e
Partitio
n
Left
U
pp
er
I
m
a
g
e
Partitio
n
Left
Lowe
r
Im
a
g
e
Partitio
n
Ri
g
h
t
U
pp
er
I
m
a
g
e
Partitio
n
Ri
g
h
t
Lowe
r
Im
a
g
e
Histo
g
ram
E
q
u
a
lizatio
n
Histo
g
ram
E
q
u
a
lizatio
n
Histo
g
ram
E
q
u
a
lizatio
n
Histo
g
ram
E
q
u
a
lizatio
n
2D
-
P
C
A
2D
-
P
C
A
2D
-
P
C
A
2D
-
P
C
A
N
o
r
m
a
lizat
io
n
We
i
g
ht
ed
F
u
si
on
N
orm
a
liza
tio
n
N
orm
a
liza
tio
n
N
orm
a
liza
tio
n
Reco
g
n
ition
/
Verification
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
16
10
–
1
616
1
612
(2
)
Out
put
val
u
es are i
n
t
h
e ran
g
e
of [0
, 1]
. To
t
r
ansf
orm
t
h
e
pi
xel
val
u
e
s
i
n
t
o
t
h
e ori
g
i
n
al
dom
ai
n, i
t
will
b
e
resized b
y
K
−
1
value
.
Figure
2, s
h
ows the facial i
m
ages
be
fo
re a
n
d
aft
e
r
hi
st
og
r
a
m
equal
i
zat
i
o
n.
Fig
u
re
2
.
An
orig
in
al im
ag
e an
d its h
i
stogram
an
d
lin
ear
h
i
sto
g
ram
eq
u
a
lizatio
n
2.
3.
M
2
D-
PCA
2.
3.
1.
Principle Compone
n
t Anal
ysis (PCA)
PCA is one of the classical feat
ure ext
r
act
i
on t
e
c
hni
ques
,
used t
o
e
x
t
r
ac
t
t
h
e gl
obal
fe
at
ures w
h
i
c
h
m
a
y
or m
a
y
not
b
e
nat
u
ral
l
y
un
de
rst
a
n
d
a
b
l
e
. P
r
i
n
ci
pal
com
pone
nt
s
g
e
nerat
e
d
fr
om
t
h
e ei
gen
v
ect
ors
are
m
o
stly associated with t
r
aining
im
ages. The
face im
a
g
es c
o
m
puted with eige
nve
c
tors a
r
e called as
Eigenfaces
which appears
like ghost f
aces
. Num
b
er of images taken i
n
the trai
ning set is ‘M’.
Out
of
extracted ei
ge
nfaces
‘K’ m
o
st significant
Eigenfaces ar
e
used in e
n
c
o
ding the va
ria
tion in face i
m
ages.
Train
i
ng
set wi
th
M sam
p
les with
each
sam
p
le si
ze is
N*N, so to
tal size
of m
a
trix
is M x N
2
.
1.
Ave
r
a
g
e
of t
h
e
t
r
ai
ni
n
g
sam
p
l
e
s i
s
de
fi
ne
d as
‘
Ψ
’
Ψ
= (1/M)
whe
r
e
‘G’ is t
h
e im
age vect
or.
2.
Mean s
h
ould be subtracted from
the tr
ai
ni
n
g
sam
p
l
e
s and i
t
i
s
de
fi
ne
d as
Φ
i =
Γ
i
–
Ψ
3.
Com
pute Cova
riance m
a
trix
C = A
T
*A, where A=
[
Φ
1
,…,
Φ
M
]
4.
The ei
genvectors
of
covaria
n
c
e
m
a
trix are
U
i
= A * V
i
, U
i
re
sem
b
le facial images called a
s
Eige
nfaces
.
5.
Each
face im
a
g
e is
proj
ected into face
s
p
ace
Ω
k
= U
T
(
Γ
k
–
Ψ
)
6.
The
n
, probe
image ‘
Γ
’ is
proje
c
ted into s
ubs
pace to c
o
m
pute a vect
or ‘
Ω
’
Ω
= U
T
(
Γ
–
Ψ
)
7.
The distance
between ‘
Ω
’ to
each class is
re
prese
n
ted as
Ω
k
= U
T
(
Γ
k
–
Ψ
)
Є
k
2
= ||
Ω
-
Ω
k
||
2
8.
A di
st
ance
t
h
re
shol
d,
Ө
c
, is
Ө
c
= 1/
2 m
a
x(j
,
k
)
(|
|
Ω
j
-
Ω
k
||)
Each im
age in the trai
ning s
e
t is a weight
ed linea
r c
o
mbinatio
n of ba
sis
faces
c
o
mpute
d
. Each
ei
gen
v
al
ue
re
p
r
esent
s
t
h
e am
ou
nt
of
vari
a
n
ce t
h
at
has
be
en ca
pt
ur
ed
b
y
one c
o
m
p
o
n
e
nt
. T
h
e
num
ber
o
f
Prin
ci
p
a
l co
mp
on
en
ts ob
tained
is eq
ual to
th
e to
tal
n
u
m
b
er sa
m
p
les
in
th
e train
i
ng
set. First prin
cip
l
e
com
pone
nt
hol
ds t
h
e hi
g
h
est
vari
a
n
ce am
ong ot
he
r pri
n
ci
p
l
e co
m
pone
nt
s.
Second
pri
n
ci
pl
e com
pone
nt
hol
d
s
next
t
o
fi
rst
PC
i
n
rep
r
ese
n
t
i
n
g vari
a
n
ce.
2.
3.
2.
Mo
dul
ar
T
w
o
Di
mensi
on
al
PCA
Gen
e
rally in
PCA, th
e 2D i
m
ag
e
m
a
trices are tra
n
s
f
ormed int
o
one
d
i
men
s
io
n
a
l either co
lu
m
n
or
row
vector. C
o
m
p
ared to c
onventional
PCA,
in
2D
-
P
CA
[3
],
[7
],
[8
] th
e co
m
p
u
t
atio
n
o
f
Co
v
a
r
i
an
ce m
a
tr
ix
is
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
ce Recogn
itio
n with
Mo
du
l
a
r Tw
o
Dimensio
na
l PCA
und
er
Uncon
t
ro
ll
ed
.... (Ven
ka
t
r
a
m
ap
han
ikuma
r
S
)
1
613
si
m
p
ler an
d
so th
e ti
m
e
co
m
p
lex
ity in
b
o
t
h
t
r
ain
i
ng
an
d
testin
g
will b
e
red
u
c
ed
.
To
im
p
r
ov
e th
e
p
e
rforman
ce
o
f
t
h
e m
o
d
a
l ag
ain
s
t
v
a
riation
s
in illu
min
a
tio
n
and
p
o
s
e, t
h
e im
ag
e is p
a
rtitio
n
e
d in
to
fo
ur eq
u
a
l
segmen
ts
su
ch
as
LL,
LU
, R
U
an
d RL
as show
n in
Fig
u
r
e
3
.
Let us i
ndicate
an im
age with
m
r
o
w
s
an
d n co
lu
m
n
s.
(3
)
Here
‘O’ is a
projected feat
ure vector
of
an i
m
age A. T
h
e c
ova
riance
m
a
tr
ix G
t
i
s
de
fi
ne
d
as,
(4
)
Figure
3
(
a)
Orig
i
n
al Im
ag
e
b) Partitio
n
e
d
Im
ag
es (LU, R
U
,
LL, RL)
Here E
is an
expectation a
n
d
is the average of the
gal
l
ery f
ace datas
e
t. Princi
pal
com
pone
nt
vect
o
r
s ca
n be
sh
ow
n as,
V
= [o
1
; o
2
; ::; o
S
]
.
W
h
e
r
e,
‘V
’
is den
o
ted a
s
s
a
m
p
le im
age feature
vector
matrix.
Distance am
ong two
feature
vectors
are
calculated with
t
h
e nearest ne
ig
hbo
rho
o
d
classif
i
er
.
Vp
= [o
(p
)
1
;
o(
p)
2
; ::; o
(
p
)
S
] an
d Vq
= [o
(q
)
1
;
o(
q)
2
; :::; o
(
q
)
S
]
i
s
gi
ve
n
as,
(5
)
2D
-PC
A
i
s
ap
pl
i
e
d o
n
eac
h m
odul
ari
zed fa
ce segm
ent
s
and t
h
en t
h
e fea
t
ure set
s
o
f
al
l
m
odul
es are
neede
d
to be
norm
alized. Norm
alization proces
s bri
ngs
the feature set
into a common scale. Similarly
n
o
rm
aliza
tio
n
is ap
p
lied
fo
r test i
m
ag
es als
o
. Z Score No
rmalizat
io
n
is u
s
ed
to
g
e
n
e
rate th
e si
milarity
sco
r
e
of each m
odul
arized
face se
gm
en
ts of traini
ng sam
p
les.
Z =
(6
)
whe
r
e
rep
r
esen
ts th
e arith
m
e
tic
m
ean
and
r
e
prese
n
t
s
t
h
e st
anda
r
d
devi
at
i
o
n
.
2.
4.
Weigh
t
ed F
u
s
i
on
In t
h
i
s
p
h
ase t
h
e feat
u
r
es ext
r
act
ed f
r
om
l
e
f
t
uppe
r, ri
ght
u
ppe
r, l
e
ft
l
o
we
r, ri
g
h
t
l
o
we
r s
e
gm
ent
s
are
integrate
d
to im
prove accura
cy of th
e face recognition. In this work wei
ghte
d
sum
m
ation m
e
thod is used to
fuse t
h
e res
u
lts of
norm
alized
feature
s
. In a face im
age,
m
o
re discrim
i
native inform
ation is availa
ble in uppe
r
part
t
h
a
n
l
o
wer
part
. B
a
se
d o
n
t
h
i
s
vi
ew,
wei
ght
s a
r
e assi
g
n
e
d t
o
t
h
e m
o
d
u
l
es.
W
e
i
g
ht
ed
sum
m
at
i
on i
s
gi
ve
n
as
(7
)
3.
E
X
PERI
MEN
T
AL RES
U
L
T
S
Per
f
or
m
a
n
ce of
th
e pr
opo
sed m
e
th
o
d
M2D-
PCA is evaluated
on
v
a
r
i
ous stand
a
rd
d
a
t
a
b
a
ses and
com
p
ared t
h
e
recognition rate
s with existing
m
e
thods
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
16
10
–
1
616
1
614
3.
1.
Face Databas
e
Extende
d
Yale face
database
B com
p
rises of 2,414
face images
from
38
differe
nt s
u
bject
s ha
ving
64
illu
m
i
n
a
tio
n
variatio
n
s
are t
a
k
e
n
for exp
e
rim
e
n
t
an
alysi
s
. Th
e
Yale datab
a
se co
m
p
rised
o
f
10
d
i
fferen
t
i
ndi
vi
dual
s
f
r
o
m
t
h
e ori
g
i
n
al
dat
a
base
an
d
2
8
di
ffe
rent
i
ndi
vi
d
u
al
s o
f
e
x
t
e
nde
d
Yal
e
dat
a
base B
.
Fe
w i
m
ages
f
r
o
m
Y
a
le d
a
tab
a
se B ar
e show
n in
Figu
r
e
4.
Fi
gu
re
4.
Sam
p
l
e
im
ages fr
om
Ext
e
nde
d
Yal
e
dat
a
ba
se B
ORL Databas
e
[9] consists of 40 distinct classes each
with
10 sam
p
les i.e., 400 im
ages are available.
Im
ag
es are acq
u
i
red
d
u
ring
d
i
fferen
t
ti
m
e
s
with
d
i
ff
eren
t
illu
m
i
n
a
tio
n
an
d
exp
r
essi
o
n
v
a
riation
s
.
ORL also
consists of
vari
ations i
n
facial
expre
ssi
ons
.
E
ach sam
p
le is of size
92*112.
Figure
5. Sam
p
le im
ages of ORL Face
Database
AR face
database [10] consis
ts of
126 distinct
classes
of uni
que pe
ople
,
each
with 26 sam
p
les
i.e.,
3
276
im
ag
es are th
ere i
n
th
e
d
a
tab
a
se. Im
ag
es are acq
u
i
sitio
n
e
d
d
u
ring
d
i
fferen
t
ligh
t
en
i
n
g
con
d
ition
s
with
v
a
r
i
ed
ex
pr
essi
o
n
s
. Each
samp
le is
o
f
size 57
6*7
68
.
Fi
gu
re
6.
Sam
p
l
e
Faci
al
Vari
a
t
i
on i
n
AR
Fac
e
Dat
a
ba
se
3.
2.
Perfor
mance Evaluati
on
To
validate the
perform
ance
of the
proposed
m
e
thod, the e
x
tende
d
Yale da
tabase is pa
rtitione
d i
n
to
training set and test set. He
re five im
ages per eac
h
s
u
bject are use
d
as
the trai
n
i
ng
set an
d
th
e
remain
ing
fifty-ni
ne im
a
g
es
per each subject are
use
d
to ev
al
uate the performance of
propos
ed m
e
thod. T
w
o
expe
rim
e
nts have bee
n
c
o
nducted. One is e
v
aluation
of pe
rform
a
nce
on whole
f
ace with
the propose
d
M2D-
PCA.
Ot
her is
evaluation
of perform
a
nce on
partiti
one
d
segments of faces
with
the
proposed M
2
D-PC
A.
3.
2.
1.
Experiment 1
In th
is exp
e
rimen
t
an
alysis, wh
o
l
e im
ag
e sam
p
les
from
t
h
e ext
e
nde
d
Yal
e
dat
a
base B
a
r
e
co
nsi
d
e
r
e
d
d
u
ring
th
e train
i
ng
and
th
e testin
g
ph
ase.
Th
is exp
e
rim
e
n
t
atio
n
m
a
in
ly
carried
ou
t to id
en
tify wh
et
h
e
r th
e
p
r
op
o
s
ed
M
2
D-PC
A m
e
th
od
is rob
u
st to
illu
m
i
n
a
tio
n
v
a
riation
s
o
r
no
t.
We con
s
id
ered so
m
e
o
t
h
e
r
ap
pro
ach
es like PCA and
LDA fo
r co
m
p
arativ
e an
alysis. Recog
n
ition
rate for t
h
ese ap
pro
ach
es on th
e
raw
face data
base i
s
24.32%
for PCA an
d
26.28% for L
DA a
n
d 29.21% for
M2
D-PCA. After applying histogram
equal
i
zat
i
o
n o
n
t
r
ai
ni
n
g
sam
p
l
e
s, rec
o
gni
t
i
on
rat
e
s are i
m
pr
o
v
ed t
o
t
h
e
r
a
t
e
s of 4
9
.
2
1%
for
PC
A an
d
56
.1
2%
for LDA an
d
6
2
.24
%
for M
2
D-PC
A. M2D-PC
A ap
pro
a
ch
yield
e
d
b
e
t
t
er recogn
itio
n resu
lts co
m
p
ared
t
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Fa
ce Recogn
itio
n with
Mo
du
l
a
r Tw
o
Dimensio
na
l PCA
und
er
Uncon
t
ro
ll
ed
.... (Ven
ka
t
r
a
m
ap
han
ikuma
r
S
)
1
615
othe
r m
e
thods. This
proves t
h
at M2D-PCA is robust
agai
nst all the
different illum
i
nation
variations. The
reco
g
n
i
t
i
on
res
u
l
t
s
are
gi
ve
n
i
n
Ta
bl
e
1.
Tabl
e
1.
Ex
pe
ri
m
e
nt
al
resul
t
s
f
o
r
w
hol
e
i
m
ages
usi
n
g
Yal
e
dat
a
ba
se B
I
m
age Type
Recognition m
odels
PCA
LDA
M2
D-PCA
Raw
24.
32%
26.
28%
29.
21%
Histogr
am
49.
21%
56.
12%
62.
24%
Perform
a
nce of the
propose
d
m
e
thod is also
eval
uate
d on
ORL face Data
base. Two im
ages per eac
h
cl
ass are used
fo
r t
r
ai
ni
n
g
an
d rem
a
i
n
i
ng 8
are use
d
fo
r t
e
st
i
ng. R
e
s
u
l
t
s
of t
h
e p
r
op
ose
d
m
e
t
hod
wi
t
h
OR
L
face
database
a
r
e s
h
own in Ta
ble 2.
Table
2.
Expe
rim
e
ntal results for
whole
images using OR
L
Face Databa
se
I
m
age Type
Recognition m
e
thods
PCA
LDA
M2
D-PCA
Raw
62.
75%
71.
6%
74.
2%
Histogr
am
85.
2%
88.
74%
98.
4%
Reco
gn
itio
n accu
racy
o
f
th
e
p
r
op
o
s
ed
m
e
th
o
d
is also ev
al
u
a
ted
on
AR
Datab
a
se. Si
x
sam
p
les p
e
r
each class a
r
e
use
d
for trai
ning a
n
d rem
a
ining
20 sam
p
les are
use
d
for testing. T
h
e
perform
a
nce of the
pr
o
pose
d
m
e
t
hod
i
s
s
h
o
w
n i
n
Tabl
e
3.
Tabl
e
3. E
x
per
i
m
e
nt
al
resul
t
s
fo
r
wh
ol
e i
m
ages usi
n
g
AR
Fa
ce Dat
a
ba
se
I
m
age Type
Recognition m
e
thods
PCA
LDA
M2
D-PCA
Raw
51.
74%
74.
3%
87.
3%
Histogr
am
69.
5%
86.
25%
94.
6%
3.
2.
2.
Experiment 2
In t
h
e sec
o
nd e
xpe
ri
m
e
nt
, pr
o
pos
ed m
e
t
h
o
d
i
s
ap
pl
i
e
d
on
se
gm
ent
e
d i
n
p
u
t
im
ages suc
h
as
l
e
ft
u
ppe
r
,
left lo
wer,
righ
t upp
er and
rig
h
t
lower p
a
rtitio
n
s
. Recogn
itio
n
rates
o
f
PCA and
M2D-PC
A m
e
th
od
s are
co
m
p
u
t
ed
on
Yale d
a
tab
a
se B. It
is clearly ev
id
en
t th
at t
h
e
p
r
op
o
s
ed
m
e
th
o
d
is ev
alu
a
ted on
seg
m
en
ted
ORL
face database a
nd
perform
a
nce of the propos
e
d m
e
thod is
99.
2
4% with his
t
ogram
e
qualised sam
p
les where as
on
raw sam
p
les the recognition
rate is 69.
34%. M2
D-PCAm
ethod is also evalua
te
d on segm
ented AR face
dat
a
base
wi
t
h
5 t
r
ai
ni
ng sa
m
p
l
e
s and per
f
o
r
m
a
nce of t
h
e p
r
o
p
o
se
d
m
e
t
hod i
s
96
.
3
2
%
wi
t
h
hi
st
og
ram
eq
u
a
lised
sam
p
les bu
t wh
ich is 52
.2
6%
o
n
raw seg
m
en
ted
imag
es.
Tab
l
e
4
.
Recog
n
ition
Rates
of Propo
sed m
e
t
h
od
ov
er b
e
n
c
h
m
a
rk
d
a
tab
a
ses
Input I
m
ages
ORL
Face Datab
a
se
AR Face
Da
tabase
Yale database B
PCA
M
2
D-PCA
PCA
M
2
D-PCA
PCA
M
2
D-PCA
Raw
L
e
ft
upper
38.
14%
62.
15%
22.
54%
42.
51%
32.
16%
38.
12%
Raw L
e
ft lower
34.
18%
64.
27%
23.
65%
43.
56%
26.
38
36.
31%
Raw
Right
upper
32.
48%
58.
67%
21.
97%
41.
29%
28.
12%
42.
43%
Raw Right L
o
wer
33.
68%
54.
19%
21.
49%
40.
5%
27.
53%
41.
21%
Raw
Fusion
55.
28%
69.
34%
32.
54%
52.
26%
36.
42%
48.
34%
Histogr
am
L
e
ft
Up
per
69.
34%
74.
29%
45.
21%
54.
12%
65.
21%
87.
21%
Histogr
am
L
e
ft
L
o
wer
72.
35%
75.
37%
43.
52%
53.
68%
59.
31%
79.
14%
Histogr
am
Right
Upper
72.
52%
76.
21%
42.
26%
48.
69%
63.
13%
89.
55%
Histogr
am
Right
Lower
74.
91%
73.
24%
44.
63%
49.
57%
60.
16%
83.
57%
Histogr
am
Fusion
82.
15%
99.
24%
65.
89%
96.
32%
74.
34%
96.
12%
4.
CO
NCL
USI
O
N
In this pa
pe
r,
the M2D-PC
A w
ith
m
a
p reduce approac
h
is propose
d for face rec
o
gni
tion unde
r
v
a
rying
illu
m
i
n
a
tio
n env
i
ro
nmen
t. It is clearly o
b
s
erv
a
b
l
e th
at th
e recog
n
ition
resu
lts
o
f
M
2
D-PC
A
is b
e
tter
in
co
m
p
ariso
n
with
app
r
o
aches lik
e PC
A, LDA fo
r th
e
whole im
age and
segm
ented im
a
g
es as
give
n
a
b
ove.
R
e
sul
t
s
of t
h
e
pr
o
pose
d
m
e
t
h
od e
v
al
uat
e
d
u
s
i
ng f
o
ur (
2
x
2)
, Si
x (
3
x 2
)
and Ei
g
h
t
(4
x 2) e
qual
p
a
rt
i
t
i
ons,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 4
,
Au
gu
st 2
016
:
16
10
–
1
616
1
616
b
u
t
t
h
e p
e
rforman
ce is sup
e
rio
r
with
4 p
a
rtitio
n
s
. Th
e
p
e
rfo
r
m
a
n
ce is ev
alu
a
ted on
stan
d
a
rd
d
a
tab
a
ses su
ch
as ex
tend
ed
Yale d
a
tab
a
se B, ORL and
AR. Fu
tu
re scop
e
o
f
th
e
work
is to
redu
ce th
e time co
m
p
lex
ity
with
furth
e
r
o
p
tim
iz
atio
n
in th
e featu
r
e selection
.
REFERE
NC
ES
[1]
J.
F.
Pe
re
ira
,
et al.
, “A robust featur
e extr
actio
n algorithm based
on class-Modular Image Principal Component
Anal
y
s
is
for f
a
ce ver
i
fi
cat
ion,
”
Acoustics, Speech and Sig
nal Processing (
I
CASSP)
,
2011 IEEE Internationa
l
Conference on
, pp.
1469-1472
, 2011.
[2]
L. Zhao,
et al.
,
“
R
esearch on K
P
CA and NS-LDA Com
b
ined Face
Recogn
ition
,
”
Computationa
l Intelligen
ce an
d
Design (
I
SCID)
,
2012 Fifth In
ter
national S
y
mposium on
, vol. 1, p
p
. 140-143
, 201
2.
[3]
J. Yang,
et al.
, “
T
wo-dimensional PCA: a new ap
proach to
appe
ar
ance-b
a
sed face
representation
and recognition
,
”
IEEE Transactio
ns on Pattern
An
alysis and Ma
ch
ine In
telligen
ce
, vol/issue: 26(1), pp.
131
,137,
200
4.
[4]
M.
D.
Ma
rsic
o,
et al.
, “
R
obust Face R
ecogni
tion for Uncont
rolled Pose and
Illum
i
nat
i
on C
h
anges,”
IEEE
Transactions on
Systems, Man
,
a
nd Cybernetics:
Systems
, vol/issue: 43(1)
, pp
. 149
,163, 2013.
[5]
G. Guodong,
et al.
, “
F
ace re
c
ognition b
y
sup
port vector m
a
c
h
ines,”
Proceed
ings of the IEEE Internationa
l
Conferenc
e
on
A
u
tomatic
Fac
e
a
nd Gesture
Reco
gnition
, pp
. 196
–201, 2000
.
[6]
Raju A
.,
et al.
,
“A Co
mparative Analy
s
is of Histogram Equali
zation based Tech
niques fo
r Contrast Enhancement
and Brightn
e
ss Preserving,”
I
n
ternational Jo
urnal of Signa
l Pro
cessing, I
m
age Processing and Patter
n
Recogn
ition
, vo
l/issue: 6(5)
, pp
.
353-366, 2013
.
[7]
D. Zhang
and Z. H. Zhou
, “2DPCA:
Two-directional two-dimensional PCA
for effic
i
en
t fa
ce r
e
pres
ent
a
tion
an
d
recognition,”
N
e
urocomputing
, v
o
l. 69
, pp
. 224-2
31, 2005
.
[8]
J. Yang and C
.
Liu, “Horizontal and
Vertical 2DPCA-Based
Di
scriminant A
n
aly
s
is fo
r Face Verification on
a
Large-S
c
al
e Dat
a
bas
e
,
”
Information Forensics a
nd S
ecurity, IEEE Transactions on
, vol/issue: 2(
4), pp. 781,79
2
,
2007.
[9]
F
.
S
a
m
a
ria and A. Harter, “
P
ara
m
eteris
at
ion of a S
t
ochas
tic M
odel for Hum
a
n F
a
ce Ident
i
fic
a
t
i
on,
”
Pr
oceed
ings
of
2nd IEEE Works
hop on App
lica
t
ions
of Computer
Vision
, Sarasota FL
, D
ecem
ber
1994.
[10]
A. M. Mar
tin
ez
and R. Ben
a
vent
e
,
“T
he
AR fa
ce
da
t
a
ba
se
,”
CVC
Tech. Report
#2
4
, 1998
.
Evaluation Warning : The document was created with Spire.PDF for Python.