TELKOM
NIKA
, Vol.13, No
.3, Septembe
r 2015, pp. 9
49~954
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i3.713
949
Re
cei
v
ed O
c
t
ober 7, 20
14;
Revi
se
d Apri
l 27, 2015; Accepte
d
May 1
3
, 2015
Implementation of K-Nearest Neighbors Face
Recognition on Low-power Processor
Eko Setia
w
a
n
1
, Adharul Muttaqin
2
Program of Informatio
n
T
e
chnol
og
y a
nd Co
mputer Scie
nc
e, Bra
w
ij
a
y
a U
n
iversit
y
,
Veteran R
o
a
d
No. 8 Mala
ng, Ja
w
a
T
i
mur,
Indon
esia
651
45
,
T
e
lp/F
ax 03
4
1
- 577 91
1
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: ekosetia
w
a
n
@
ub.ac.i
d
1
, adharu
l
@u
b.ac.id
2
A
b
st
r
a
ct
Face rec
ognition
is one
of
early
detection
in s
e
curit
y
system
. A
u
tom
a
tion enc
ourages
imple
m
entati
o
n
of fac
e
rec
o
gniti
on
in
s
m
a
ll
and
co
m
pac
t devic
es. Mo
st of face
rec
ogn
ition
res
e
a
r
ch
focused o
n
ly
on its accura
cy and perfor
m
e
d
on h
i
gh-
spee
d co
mput
er. F
a
ce recogniti
on that is
i
m
p
l
eme
n
t
ed
on
l
o
w-co
st p
r
oce
sso
r, such
as ARM proc
essor, ne
eds
prop
er al
gorith
m
. Our rese
ar
c
h
investi
gate
K-
Near
est Ne
ig
h
bor (K
NN)
alg
o
rith
m
in r
e
co
gni
z
i
n
g
face
o
n
ARM
proc
es
sor. T
h
is r
e
se
arch
soug
ht best k-
valu
e to cre
a
te
prop
er face r
e
cogn
iti
on w
i
th
l
o
w
-
pow
er proc
essor. T
he pr
o
pose
d
al
gor
ith
m
w
a
s tested on three d
a
tasets that w
e
re Olivet
ti Re
searc
h
La
boratory (ORL)
,
Yaleface a
nd
MUCT
. OpenC
V
w
a
s chosen
a
s
ma
in cor
e
i
m
a
ge pr
ocess
i
ng li
brary,
du
e to its hig
h
-s
pee
d. Propos
e
d
alg
o
rith
m w
a
s
imple
m
ente
d
o
n
ARM11 7
0
0
M
H
z
. 10-fo
ld
cross-vali
dati
o
n show
ed that
KNN face recogn
ition d
e
tect
ed
91.5%
face w
i
t
h
k=
1. Overal
l
ex
per
iment s
h
ow
ed that
prop
osed
al
gorith
m
detecte
d face
on 2.
66 s
on
A
R
M
process
o
r.
Ke
y
w
ords
: face recog
n
itio
n, K-Near
est Nei
ghb
ors, ARM processor
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Smart buildi
n
g be
come
s t
op issu
es in
last
y
ear
s.
S
o
me re
se
ar
che
s
kee
p
d
o
ing t
o
prop
ose
b
e
tter sma
r
t
buil
d
ing system. A
sma
r
t
build
ing shoul
d b
e
able
to mo
dify environm
ent
according to
the o
c
cupa
nt comm
and. Fi
rst, the sy
st
e
m
must id
enti
f
y who is it
s o
c
cupa
nt. Nat
u
ral
identificatio
n techni
que i
s
face
re
cog
n
ition that
gives better soluti
on. Applying
face recogniti
on,
system could identify
the
occupa
nt
witho
u
t disturbi
ng the other p
e
o
p
le.
For yea
r
s, fa
ce recognitio
n
be
come to
p issu
e
s
in
rese
arch p
a
p
e
r. Several t
e
ch
niqu
es
were appli
e
d
to propo
se
better re
co
gnition. Ch
a
nging in po
se, expressio
n
variation
s
and
differen
c
e
s
in
the position
of the light give difficulty to
resolve in face recognitio
n
. Basically, face
recognitio
n
can b
e
ap
pro
a
ch
ed
with f
a
cial
biom
etric featu
r
e
s
o
r
stati
s
tical
method. F
a
ci
al
biometri
c re
cognition
te
ch
nique offers high
a
c
curacy
with long
cal
c
ulatio
n. Hen
c
e,
statistical
approa
che
s
offer sp
eed
calcul
ation. Due to the
sp
eed, a stati
s
t
i
cal ap
proa
ch
is appli
ed m
o
re
widely than f
a
cial bi
ometri
c [1]. Princip
a
l Comp
one
n
t
Analysis (P
CA) an
d Line
ar Di
scrimin
a
n
t
Analysis
(L
DA) are two po
pular
statis
tical approa
che
s
in face
re
co
gnition.
Several stu
d
i
e
s ha
d sh
o
w
n that PCA (k
no
wn a
s
Eigenface
)
and L
D
A (known as
Fishe
r
fa
ce)
method
had
high
accu
ra
cy in fa
ce
re
co
gnition [2]. B
o
th of th
ese
method
s
cal
c
ulated
eigenvalu
e
s
of face imag
e. Calculatio
n of ei
genval
ues requi
red
long step
s. Several meth
ods
were propo
se
d to improve
them in accu
racy and
sp
ee
d. Arif improv
ed the Fishe
r
face efficie
n
cy
by cal
c
ul
ate it
in two-dimen
s
ion.
The
imp
r
ovem
e
n
t gav
e go
od
re
sult
in 10
0% a
c
cu
racy
[3]. Two-
dimen
s
ion
ca
lculatio
n take
s more
step
s with time-co
s
t con
s
eq
uen
ce. Wijaya co
mbined h
o
list
i
c
feature an
d linear di
scrimi
nant analy
s
is
to redu
ce
the
training time
and keep the
accuracy [4].
Ho
wever, th
e previou
s
works we
re
applie
d on compute
r
equi
pped with hi
gh-p
o
wer
pro
c
e
s
sor. Si
nce
en
ergy
conservation
i
s
o
ne
of key i
s
sue
s
in
sma
r
t buildi
ng
re
search, lo
w po
wer
pro
c
e
s
sor be
come
s intere
sting to
pic to
be inve
stigat
ed a
nd
evalu
a
ted of
face reco
gnition. B
a
se
d
on Yong
et a
l
survey [2],
K Nea
r
e
s
t Neighb
or
(KNN) pl
aced thi
r
d ran
k
after
PCA and
LDA on
accuracy a
n
d
KNN pla
c
e
d
first ran
k
o
n
pro
c
e
ssi
ng speed. In this
study, we int
e
re
st to mea
s
ure
how suitable
KNN appli
ed in low power pr
oce
s
sor compa
r
e with PCA/Eigenface a
n
d
LDA/Fisherfa
ce.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
9
30
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 949 – 954
950
2. K-Near
est
Neighbo
r
K-Ne
are
s
t Neighb
or (K
NN) i
s
data cl
assifica
tio
n
method that
can b
e
used
as face
recognitio
n
method. Ea
ch pixel in fa
ce re
pre
s
e
n
ts uniqu
e info
rmation. Thi
s
pape
r recogn
ized
f
a
ce b
a
s
ed o
n
ea
ch pix
e
l
cla
ssif
i
cat
i
on.
Face
wa
s d
e
t
e
rmin
ed by
most
cl
as
s r
e
sult
e
d
in ea
ch
pixel cla
s
sification. In recognition, pixe
l matrix of fa
ce im
age
sh
ould b
e
resh
ape into
vect
or
before
cla
ssifi
cation. The p
r
opo
sed KNN
face re
co
gniti
on algo
rithm i
s
de
scribe
d a
s
follows:
1.
Modify dimen
s
ion
s
of M
-
ro
w an
d
N-colo
mn
face matrix (M x N) int
o
fac
e
trans
p
ose
vec
t
or
(1 x MN
)
2
.
A
r
r
a
g
e
e
a
c
h
f
a
c
e
v
e
c
t
o
r
i
n
t
o
m
a
t
r
i
x
f
o
r
m
(
K
x
M
N
)
w
i
t
h
K
i
s
n
u
m
b
e
r
o
f
t
r
a
i
n
i
n
g
f
a
c
e
image
s. Each ro
w rep
r
e
s
ents a si
ngle
image an
d each col
u
mn
would
rep
r
e
s
ent same pi
xels
positio
n in ea
ch face imag
e.
3.
Modify testing image matri
x
into face tra
n
sp
ose vecto
r
, as traini
ng i
m
age
s
(1 x M
N
).
4.
Cal
c
ulate the
Eucluide
an Dista
n
ce (d)
of
each colu
mn (i) in test
ing image (x
) to
each col
u
mn
(i) in traini
ng i
m
age (y).
5.
Determine
cl
assificatio
n
b
a
se
d on
the
sh
o
r
t
e
st
dist
a
n
ce
of
w
hole
colu
mn in
e
a
c
h
row.
6.
Determine fa
ce re
co
gnitio
n
based on th
e k nea
r
e
s
t neighb
or an
d its dista
n
ce.
St
a
r
t
C
a
pt
ur
e f
r
o
m
c
a
m
e
r
a
L
B
P
-
cascad
e
da
n H
aar
-
c
ascad
e
fa
c
e
d
e
te
c
t
i
o
n
C
o
lle
c
t
r
e
f
e
r
e
n
c
e
fa
c
e
i
m
a
g
e
F
o
re
v
e
r l
o
o
p
D
i
m
e
n
s
i
o
n t
r
a
n
sf
or
m
a
t
i
on
o
f
r
e
fe
r
e
n
c
e
fa
c
e
i
m
a
g
e
(
1
x
M
N
)
V
e
ct
or
D
i
m
e
nsi
on
T
r
ansf
o
r
m
a
t
i
o
n
D
i
s
t
a
n
ce
ca
l
c
u
l
a
t
i
o
n
F
i
nd
k
-
ne
ar
est
n
e
i
g
h
b
o
r
Fin
i
s
h
Fa
c
e
re
c
o
g
n
i
t
i
o
n
re
s
u
l
t
Figure 1. Face Re
cog
n
ition
Flowcha
r
t
3. Implementation
The p
r
o
p
o
s
e
d
metho
d
was im
pleme
n
t
ed with
Op
e
n
CV lib
ra
ry [6] in
C lan
g
uage
to
ensure
fast i
m
age
pro
c
e
s
sing.
Ra
spb
e
rry Pi
wh
i
c
h
wa
s eq
uipe
d
with ARM11
700M
Hz core
, wa
s
sele
cted to
b
e
main
pro
c
essing
unit. Overall al
go
ri
thm that wa
s impleme
n
te
d in video fa
ce
recognitio
n
system, wa
s
sho
w
n
in Fi
g
u
re
1. Mai
n
pro
c
e
s
s reco
gnized fa
ce
contin
uou
sly
b
y
c
a
p
t
ur
in
g a pic
t
u
r
e fr
om a
t
ta
c
h
ed
w
e
bca
m
pe
r
o
d
i
ca
lly. C
a
p
t
ur
e
d
ima
g
e
wa
s pre
-
pr
oc
es
se
d
b
y
applying fa
ce
detection. Haar-cas
ca
de
wa
s appli
ed
due to its ro
b
u
stne
ss in face dete
c
tion.
Pre-
pro
c
e
s
sed im
age
woul
d be
pro
c
e
s
sed
b
y
propo
se
d f
a
ce
dete
c
tion
method. Th
e
whol
e p
r
oce
ss
woul
d be re
p
eated so as t
o
prod
uce a video-ba
s
ed fa
ce re
co
gnitio
n
.
4. Results a
nd Discu
ssi
on
Test
s
were
condu
cted
to e
v
aluate p
e
rfo
r
manc
e
of p
r
o
posed
algo
rithm. Te
sts carried
out
with thre
e face dataset whi
c
h a
r
e O
R
L [
7
], Yaleface [8] and MUCT
[9] dataset. Before ap
plying
as traini
ng d
a
taset, every
images
was
applie
d face
detectio
n
and
equalized in
same si
ze. The
result of p
r
e
p
r
ocessin
g
sh
owe
d
that
so
me faces co
uld n
o
t be
u
s
ed a
s
data
s
e
t. The 29
6
O
R
L
face
s, 15
6 Y
a
lefaces and
3644
MUCT
face
s
wo
uld b
e
u
s
ed
a
s
trai
ning
and te
sti
n
g d
a
ta. Figu
re
2 sho
w
s the overall ima
g
e
s
that are u
s
e
d
in training a
n
d testing.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
9
30
Im
plem
entation of K-Ne
arest Nei
ghb
ors Fa
ce
Reco
gnition on L
o
w
-p
ower… (E
ko Setiawan
)
951
Figure 2. ATT Face
Databa
se [7]
Figure 3. k-F
o
ld Cross Val
i
dation
Firs
t tes
t
ing
was
to find the k values
that
pro
d
u
c
ed
high
accu
ra
cy. Second
testing would
comp
are accura
cy and e
x
ecution tim
e
of
prop
osed metho
d
with co
mmo
n Eigenfa
c
e
and
Fishe
r
fa
se. Fi
nal testin
g would o
b
se
rve
time execut
io
n of co
ntinuo
us fa
ce recog
n
ition sy
stem
in
low-l
e
vel processor.
Table 1. Cro
s
s Validation o
n
ORL d
a
taset
Num
b
er o
f
10-fo
ld
A
ccu
r
a
c
y
(%
)
K=1 K=3 K=5 K=7
1
93.3 83.3 73.3 70.0
2
96.7 86.7 90.0 86.7
3
100
90.0 90.0 86.7
4
93.3 86.7 86.7 83.3
5
96.7 93.3 86.7 86.7
6
93.3 86.7 80.0 76.7
7
79.3 72.4 65.5 69.0
8
93.1 89.7 75.9 79.3
9
86.2 79.3 82.8 82.8
10
82.8 82.8 79.3 72.4
A
v
e
r
age
91.5 85.1 81.0 79.3
Fold-1
Fold-2
…..
Fold-(N-1)
Fold-N
:
Train
i
n
g
:
T
estin
g
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 949 – 954
952
Tabel 2. KNN face accu
ra
cy
K Value
A
v
e
r
age A
c
cura
c
y
(
%
)
ORL da
taset
Y
a
leface dataset
MUCT dat
aset
K=1 91.5
78.8
70.0
K=3 85.1
75.0
63.6
K=5 81.0
71.8
62.1
K=7 79.3
64.8
60.2
First
te
sting wa
s cond
uct
ed
in 10
-fold
Cr
oss Vali
d
a
tion. Figu
re
3 sho
w
ed
how to
determi
ne
t
r
ai
ning and
testi
ng
im
age
s.
A
c
cura
cy on e
a
ch
fold wa
s cal
c
ulate
d
.
T
abel 1 wa
s
d
e
tail
accuracy
on
each fold
in
ORL
data
s
et.
The
total a
ccura
cy was
cal
c
ulate
d
o
n
av
erag
e of
10
-fold.
Table 2 sho
w
s the re
sult
s of accura
cy in different k v
a
lue on
several dataset.
Based o
n
Ta
ble 2, KNN d
one be
st accura
cy
on k e
qual to 1. It s
howed that KNN gave
91.5% on 29
5 ORL fa
ces,
78.8% on 156 Yalefa
ce
s
and 70% on
3644 M
U
CT face
s. Enorm
ous
numbe
r
of M
UCT
data
s
et
made th
e
system co
nf
use i
n
re
co
gnition.
K wa
s
equ
al
to 1
and
ORL
dataset woul
d be use
d
in next test due
to its
best re
sult. The nex
t phase te
st wa
s to comp
are
the pro
p
o
s
ed
method
with Eigenface an
d Fish
erfa
ce.
Tests
ca
rri
e
d
out to obtai
n informatio
n
on
how fea
s
ible
the propo
sed method
whe
n
impl
e
m
enting in l
o
w-po
we
r proce
s
sor. Te
sts
con
d
u
c
ted on
compute
r
wit
h
Intel Core i
7
, 2.8 GHz a
nd Ra
spb
e
rry Pi with Broadcom ARM1
1,
700M
Hz. Th
e
accuracy an
d executio
n time wo
uld
be
comp
are
d
. Accura
cy testi
ng wa
s h
e
ld
on
same
scena
ri
o of k-value
adju
s
tment. The final re
sults of testing were sh
own in Table 3 and
Table 4.
Tabel 3. Fa
ce
reco
gnition o
n
comp
uter
Me
t
h
od
A
ccu
r
a
c
y
(%
)
Learni
ng
require
ment
Learni
ng
time (s)
Recog
n
iti
on
time (s)
Total
ti
me
(s)
Eigenface 91.5
Y
e
s
3.935875
0.006887
3.94276
Fisherface 91.5
Y
e
s
2.997280
0.000540
2.99782
KNN 91.5
No
0
0.003689
0.00369
Tabel 4. Fa
ce
reco
gnition o
n
low-po
we
r pro
c
e
s
sor.
Me
t
h
od
A
ccu
r
a
c
y
(%
)
Learni
ng
require
ment
Learni
ng
time (s)
Recog
n
iti
on
time (s)
Total
ti
me
(s)
Eigenface 91.5
Y
e
s
299.221
0.459
299.680
Fisherface 91.5
Y
e
s
234.252
0.027
234.279
KNN 91.5
No
0
0.152
0.152
Table
3
sho
w
ed that fa
ce
reco
gnition
ex
ecut
io
n time
on
comp
uter
wa
s n
o
t mo
re
than
4
se
con
d
s. T
a
ble 4 info
rmed that e
x
ecution ti
m
e
of fishe
r
f
a
ce
and
eig
enface in
cre
a
se
d
signifi
cantly into 299.7
se
cond
s and
234
.3 se
cond
s re
spe
c
tively in low-po
we
r pro
c
e
s
sor. Both
of
these
pop
ula
r
metho
d
s
were n
o
t app
ropriate l
o
w-p
o
we
r p
r
o
c
e
s
sor
due to it
s lon
g
exe
c
u
t
ion
time. The pro
posed metho
d
took sh
orte
r time t
han the eigen or fisherfa
ce. This fact confirm
e
d
that KNN wa
s p
r
op
er fa
ce
re
cog
n
ition
method i
n
lo
w-p
o
wer processor. On ne
xt
sce
nari
o
, KNN
woul
d be sele
cted a
s
the re
cog
n
ition met
hod.
Final test ai
med to observe perform
an
ce of
co
ntinu
ous fa
ce re
cognition on l
o
w-po
we
r
pro
c
e
s
sor. E
x
perime
n
t ap
plied fa
ce
d
e
tection
by
Local Bina
ry Pattern a
n
d
Ha
ar-Ca
s
ca
de.
Dete
ction
re
sult wa
s
re
cog
n
ize
d
by K
N
N. Exper
im
en
t wa
s do
ne
b
y
sho
w
ing
p
r
i
n
ted te
st ima
g
e
on onlin
e ca
mera. Th
e camera ca
ptu
r
ed ima
ge.
Figure 4 sh
ows the face detectio
n
and
recognitio
n
proce
s
s on lo
w-power p
r
o
c
e
s
sor.
Tabel 4. CP
U Load an
d me
mory usage.
Camera Pixe
l
CPU Lo
ad
(%
)
Memor
y
Usage
(MB)
1024 x
768
84
67.6
960 x 6
4
0
79
62.8
640 x 4
8
0
82
59.1
320 x 2
4
0
85
53.2
160 x 1
2
0
75
51.5
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Im
plem
entation of K-Ne
arest Nei
ghb
ors Fa
ce
Reco
gnition on L
o
w
-p
ower… (E
ko Setiawan
)
953
Tabel 5. Execution time
of
face reco
gnition
Trial
Executi
on time (s)
Trial
Executi
on time (s)
1 2.32
16
2.73
2 2.76
17
2.70
3 2.67
18
2.66
4 2.65
19
2.70
5 2.81
20
2.61
6 2.91
21
2.61
7 2.92
22
2.67
8 2.29
23
2.42
9 2.56
24
2.28
10 2.52 25
2.85
11 2.52 26
2.79
12 2.45 27
2.71
13 2.45 28
2.95
14 2.66 29
2.85
15 2.68 30
3.00
A
v
e
r
age time (s
)
2.66
Experiment e
v
aluated CP
U load a
nd
memory u
s
a
ge of contin
u
ous fa
ce recognition o
n
several ca
m
e
ra pixel si
ze. The syst
em sho
w
dif
f
erent CP
U Load a
nd m
e
mory on e
a
ch
resolution. Detail CPU L
o
ad and m
e
m
o
ry usage
wa
s
sh
own cle
a
r
ly on Table
4. Smaller ca
mera
resolution
de
livered
sm
all
e
r m
e
mo
ry
usa
ge.
CPU load
sho
w
n
anom
aly b
e
twee
n vari
o
u
s
resolution. T
he experi
m
e
n
t sho
w
n tha
t
640 x 480 resolution g
a
v
e optimum in CPU loa
d
and
memory
u
s
a
ge. Th
e
re
sol
u
tion p
r
od
uced n
o
t hi
g
h
CPU load
a
n
d mem
o
ry
usage. Expe
rim
ent
also
perfo
rme
d
face
re
co
gn
ition 30 time
s seq
uentially
and
re
corded
each exe
c
uti
on time. Based
on Tabl
e 5, the average ti
me of the face re
cog
n
ition
pro
c
e
ss
wa
s 2.66 se
co
nd
s on e
m
be
dd
ed
sy
st
em
s.
Te
s
t
res
u
lt
s
sho
w
ed t
h
at
K
N
N f
a
c
e
re
co
g
n
ition was fe
asibl
e
to be i
m
pleme
n
ted
on
embed
ded system
s.
Figure 4. Con
t
inuou
s face reco
gnition
5. Conclusio
n
The
re
sea
r
ch con
c
lude
d
som
e
info
rmations ab
o
u
t face
re
co
gnition o
n
l
o
w-po
we
r
pro
c
e
s
sor. K-Nea
r
e
s
t Neig
hbor face recognition
deliv
ered
be
st a
c
curacy
91.5
%
on k=1. K
N
N
sho
w
e
d
the faster exe
c
uti
on time com
pare
d
with P
C
A and L
D
A
.
Time execu
t
ion of KNN to
recogni
ze fa
ce wa
s 0.152
se
con
d
s o
n
h
i
gh-p
r
o
c
e
s
sor. Face dete
c
ti
on and recog
n
ition only ne
ed
2.66 se
co
nd to re
cogni
ze o
n
low-po
we
r ARM11
b
a
se
d system . O
v
eral propo
se
d method wo
rk
w
e
ll on
low
-
po
w
e
r
pr
oc
es
so
r
.
Ap
pr
o
p
r
i
ate
fa
c
e
dete
c
tion
on
l
o
w-p
o
we
r system have
pote
n
tia
l
y
to boost reco
gnition.
Referen
ces
[1]
Rabi
a J, Hami
d RA. A Survey of F
a
ce Rec
ogn
ition T
e
chn
i
qu
es.
Journa
l of Informati
on Processi
ng
.
200
9; 5(2).
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 949 – 954
954
[2]
Peter NB, Joao
PH, David JK.
Eigenfac
es vs. F
i
sherface
s: Recogn
it
io
n
Using C
l
ass
Specific L
i
ne
ar
Projecti
on
. 2014.
[3]
Arif M. Ne
w
Modell
i
n
g
of Modifie
d
T
w
o Dim
ens
io
na
l F
i
sherface
Based F
eatur
e Extracti
on
.
T
E
LKOMNIKA T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol
. 20
14; 12(1): 1
15-
122.
[4]
IGPS W
ija
ya, K Uchim
u
ra,
G Koitaki.
F
a
ce Rec
o
g
n
i
t
ion Us
ing
H
o
listic F
e
ature
s
and
Li
near
Discrimi
nant A
nal
ysis
Simp
lifi
c
ation.
T
E
LKO
M
NIKA T
e
leco
mmu
n
icati
on
Co
mp
uting
Ele
c
tronics a
n
d
Contro
l
. 201
2; 10(4): 77
1-7
8
2
.
[5]
Yong
X,
Xia
o
zhao F
,
Xu
el
o
ng L, Jia
n
Y, Jane
Y, Ho
ng
L, Shaoh
ua T
.
Data Uncerta
i
nt
y
i
n
F
a
c
e
Reco
gniti
on
.
IEEE Transaction on Cybernetics
. 2014.
[6]
OpenCV
Dev
T
eam. 2014. T
he Op
en
CV R
e
ferenc
e M
anu
al
Rel
eas
e 2.4.
9.0. access
ed
on S
eptem
be
r
10th 20
14.
[7]
F Samaria, A
Harter.
Para
meterisati
on of a
Stochastic Mo
del for H
u
ma
n
F
a
ce Identific
ation
. 2nd IEEE
W
o
rkshop o
n
Appl
icatio
ns of Comp
uter
Visi
on.Saras
o
ta (F
lorid
a
). 199
4.
[8]
P Belh
umeur,
J Hesp
anh
a, D
Kriegm
an. Eig
enfaces
vs. F
i
s
herfaces: R
e
co
gniti
on Us
in
g C
l
ass Sp
ecifi
c
Lin
ear Proj
ecti
on.
IEEE Transactions on Pattern An
alysis and Machine Int
e
lligence
. 19
97
: 711-72
0.
[9]
S Milborro
w
,
J Morkel, F
Nicol
ls. T
he
MUCT
Landmarked F
a
ce D
a
tabas
e.
Pattern Recog
n
iti
o
n
Associati
on of South Africa
. 2
010.
Evaluation Warning : The document was created with Spire.PDF for Python.