TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 2574 ~ 2
5
8
1
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4825
2574
Re
cei
v
ed Se
ptem
ber 7, 2013; Re
vi
sed
Octob
e
r 17, 2
013; Accepte
d
No
vem
ber
19, 2013
Recognition Based on Metric-optimized Neighborhood
Preserving Embedding
Chen
Bo*, Z
h
ang Ye
Dep
a
rtment of Electrical E
ngi
neer
ing,
Xi
n Xi
ang U
n
ivers
i
t
y
,
East Jin Sui Street, Xi
n
X
ia
ng
cit
y
, HeN
an pr
ovinc
e
, Chin
a
*Corres
p
o
ndi
n
g
author, em
ail
:
chenbo
81
09
1
8
@so
hu.com
A
b
st
r
a
ct
F
a
ce reco
gn
iti
on is
a
bio
m
etric tech
nol
ogy
w
i
th gr
eat
dev
elo
pab
le
pote
n
t
ial. It has
a gr
eat d
eal
o
f
potenti
a
l a
p
p
l
i
c
ations
in p
ubl
ic security a
n
d
infor
m
atio
n s
e
curity. T
o
ov
erco
me th
e pr
obl
e
m
in th
e h
i
gh-
di
me
nsio
nal
fa
ce d
a
ta
proces
sing, th
e k-n
e
a
r
est ne
igh
bors
is ch
ose
by
Lin
ear D
i
scri
m
i
nat
e An
alysis
(LD
A
).
A Metric-opti
m
i
z
e
d
is
pro
pos
ed for N
e
ig
hb
orho
od
Pres
er
ving E
m
bed
di
n
g
(MONPE). MONPE alg
o
ri
thm,
w
i
th the dime
n
s
ions of data r
educ
ed
by LD
A, w
ill be reaso
nab
le in NPE
a
l
gorit
hm. On the other ha
nd, L
D
A
max
i
mi
z
e
s the
betw
een-cl
ass scatter and mi
ni
mi
z
e
s the w
i
thin-cl
a
ss scatter, so the neig
hbors of a sa
mpl
e
w
ill hav
e hi
gh
e
r
possi
bility to
be p
i
cked fro
m
the sa
me
c
l
as
s. With the ORL face d
a
tab
a
s
e
an
d the Y
a
le
datab
ase, the
recog
n
itio
n rat
e
an
d ru
n ti
me is
co
mpare
d
amon
g NPE,
MONPE and
CLMONPE.
T
he
simulati
on res
u
lts show
that CLMONPE
has
obvi
ous a
d
vant
age i
n
ap
plic
ati
on.
Ke
y
w
ords
: face recog
n
itio
n, ma
nifo
ld, sup
e
r
vised n
e
ig
hb
o
r
hoo
d pres
ervi
ng e
m
b
e
d
d
in
g
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
Manifold
re
aching
ha
s b
e
e
n
g
r
eat
used
in
th
e field
o
f
data
dinnin
g
, ma
chin
e
studying,
comp
uter visual re
ce
ntly.
The aim of m
anifold
stu
d
y is to gain the
topology st
ru
cture
and its l
a
w
among
the d
a
ta sp
ace, when the
colle
cted d
a
ta
p
e
rform the
man
i
fold structu
r
e
[1]. When
th
e
face and g
e
st
ure chan
ge, the gene
ral fa
ce re
co
gniti
o
n
method will
becom
e less effective. That’s
why manifold
study is cho
s
en in face recognition.
Whe
n
face an
d ge
sture
cha
nge
s, the colle
cte
d
data will
ch
a
nge in
nonli
n
ear
way. So t
he ge
neral d
a
ta dime
nsio
n de
crea
se
way su
ch a
s
P
C
A
[2, 3] and
L
D
A [4] ca
n not
perfo
rm th
e
true in
ne
r
structure of i
m
a
ge
spa
c
e. Bu
t the manifol
d
study
can
do
it well. T
he t
y
pical p
ape
rs abo
ut face
reco
gnition i
s
publi
s
hed
in
Scien
c
e i
n
th
e
same
i
ssu
e
b
y
Tene
nbau
m [5] an
d
Ro
wel
s
[6]. Th
e
y
sho
w
different ma
nifold
study m
e
thod
s a
s
Isometri
cal M
appin
g
(ISO
MAP) [7] and Locally Linea
r Embeddi
ng
(LLE) [8].
2. Rese
arch
Metho
d
ISOMAP is a
kind of b
e
st
whol
e situatio
n
nonlin
ea
r di
mensi
on d
e
crease metho
d
. It can
ensure
data
conve
r
ge
nce
grad
ually. The ori
g
inal
i
dea is th
at: geod
esi
c
di
stance is a g
o
o
d
dissimila
rity measurement
, which ne
ed
to solve
the
p
r
oble
m
of ge
ode
sic
dista
n
c
e b
e
twe
en n
on-
neigh
borhoo
d
dots. T
he
problem
ca
n
b
e
solved by
computing
the
dista
n
ce ma
trix to co
nst
r
uct
sub
s
p
a
ce of the nonli
nea
r manifold di
stribution. The d
i
mensi
on can
be decrea
s
e
d
in this way.
Locally Linea
r Embeddi
ng
(LLE) ai
ms to
con
s
tr
u
c
t an
weig
ht vector betwee
n
the sampl
e
and its
neigh
borh
ood, a
n
d
it also
keep
s an
co
nst
a
n
t
weight valu
e for ea
ch
n
e
ighb
orh
ood
in
decrea
s
e
d
di
mensi
on
sp
a
c
e. T
h
is me
a
n
s th
at the
e
m
body m
ap i
s
u
nde
r l
o
cal
linea
r
co
nditi
on
and
th
e reco
nstru
c
t error seem
s
to
b
e
the
lea
s
t.
Thi
s
a
pproa
ch
can n
o
t only fi
nd the
no
nlin
ear
con
s
tru
c
t effi
ciently, but
also
ha
s the i
n
varian
ce
of transl
a
tion a
n
d
rotation.
Do
n
oho [9] i
n
vented
the HL
LE fro
m
LLE. The
HLLE
can fin
d
the latent
e
qual di
stan
ce
para
m
eter i
n
local m
anifol
d
.
Zhang
Ch
ang
sui p
r
e
s
ente
d
a metho
d
wh
ich
can
map f
r
om lo
w e
m
b
ody sp
ace to
high
spa
c
e
a
n
d
the metho
d
h
a
s b
een
proved in m
u
ti-ge
s
ture
face
im
age
re
con
s
tru
c
tion exp
e
rim
ent [10-11]. T
h
is
approa
ch imp
r
oved the no
n
linear d
e
crea
se metho
d
be
tter.
Re
cently different kin
d
s of
manifold stu
d
y
methods sp
ring up, e.g., Lapla
c
ian Eig
enmap,
local
tang
ent
sp
ace ali
g
n
m
ent (LTSA), and
so
o
n
. Lapl
aci
an E
i
genma
p
solv
e the
nonli
n
ear
function
throu
gh
L
apla
c
ia
n operator, by whi
c
h
th
e clo
s
ed
poi
nts
i
n
high dimen
s
i
on spa
c
e
can
be
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition
Based o
n
Me
tric-optim
ized
Neigh
borhoo
d Preserving
Em
bedding (Che
n Bo)
2575
mappe
d to low dimen
s
io
n spa
c
e
with little distan
ce
ch
ange
d. Local tangent spa
c
e alignme
n
t first
con
s
tru
c
t the
local lo
w di
mensi
on ma
n
i
fold by t
he tange
nt spa
c
e
of every sa
mple point, then it
get the whol
e
low embo
dy coo
r
din
a
te by
the arra
nge
ment of tangent spa
c
e.
Saul an
d Ro
wei
s
the
LLE
algo
rithm i
s
applie
d to t
he fa
ce p
o
se
image to
ob
tain its
distrib
u
tion. Z
hang
co
mbin
edthe L
L
E al
gorithm
with
t
he SVR
(Sup
port Ve
ctor
Regre
s
sion
) in
the
real fa
ce p
o
se estimatio
n
to get better
e
x
perime
n
tal result
s. But the manifold l
e
arnin
g
meth
o
d
s
have a
co
m
m
on d
r
a
w
ba
ck: they did
n
o
t give an
ex
plicit ma
ppin
g
fun
c
tion,
so can't extract the
cha
r
a
c
teri
stics of the new
sampl
e
. He X
i
aofei
gives a
Lapla
c
ian Ei
genma
p
linea
r app
roximati
on
method, the Locality Preserving Proj
ections, refe
rre
d
to as "LPP, obtained explicit proj
ecti
on
mappin
g
, an
d apply it to
face
re
co
gn
ition. Z
han
g
gives th
e lo
cal tange
nt space alig
nm
ent
algorith
m
of l
i
near ap
proximation m
e
th
od, line
a
r local tang
ent
sp
ace
alig
nmen
t (LLTSA
), a
n
d
applie
d to face re
cog
n
ition
to achi
eve effective re
sults.
Later, Pan
g
Y
anWei et al.,
on the b
a
si
s
of
the locally lin
ear e
m
be
ddi
ng (L
LE) p
r
o
posed a
n
e
fficient
n
onlin
e
a
r sub
s
pa
ce learni
ng
m
e
thod,
kernel nei
ghb
orho
od keep
proje
c
tion, its main idea is
a linear tran
sformatio
n
matrix is introd
uced
to app
roxima
te the
classi
cal lo
cal
line
a
r e
m
bed
din
g
, and th
en
by the m
e
thod
of nu
cle
a
r
techni
que i
n
high di
men
s
ion
a
l spa
c
e
to solve. Z
hang, et
c. T
h
is p
ape
r p
r
opo
se
s a fa
ce
recognitio
n
m
e
thod ba
se
d on nea
re
st n
e
ighb
or ma
nifold. The met
hod ad
opts
manifold lea
r
ning
algorith
m
to cal
c
ulate th
e
mappin
g
of
the high di
mensi
on
spa
c
e to lo
w di
mensi
on lin
e
a
r
manifold, the
discrimina
nt analysis b
a
s
ed on n
earest neigh
bo
r rule of manifold. Wu, a
discrimi
nant
of manifol
d
l
earni
ng
algo
rithm is p
r
op
ose
d
, will
fa
ce
data
in
h
i
gh-di
men
s
io
nal
observation
space of ne
ste
d
in low di
me
nsio
nal
ma
nifold sp
ace. Un
like LLE, ISO
M
AP algorith
m
,
this method
USES the correl
a
tion
com
ponents anal
ysis (RCA
) to
establi
s
h a
nonlinear nested
discrimi
nant
rule in the d
a
ta. Yang, such
a
s
face
reco
gnition
method ba
se
d on extend
ed
ISOMAP is p
u
t forwa
r
d. T
he key is to e
s
timate the g
eode
si
c dista
n
ce, the m
e
th
od an
d geo
d
e
si
c
distan
ce a
s
a
feature vecto
r
, in pairs wit
h
FLD finally find the be
st proje
c
tion di
re
ction.
Re
cog
n
ize fa
ce
s mai
n
ly according
to the
feature
s
of
a
face, that i
s
t
o
say th
at the
r
e i
s
a
big differen
c
e
between
different
individu
als
with th
e
same
perso
n i
s
m
o
re
sta
b
le
mea
s
u
r
eme
n
t,
due to the
co
mplexity of face
cha
ngin
g
, so the
expression fe
ature
s
and featu
r
e
e
x
traction i
s
ve
ry
difficult. Du
e
to the dim
e
n
s
ion
s
of the
origin
al ima
g
e
is
quite
hig
h
, on the
ba
sis of the
o
r
igi
nal
image i
s
processed
dire
ctl
y
, will
increa
se the co
mple
xity of t
he algorithm, an
d the pe
rform
a
n
c
e
is al
so
a
ch
allenge
to the
comp
uter ha
rdwa
re,
so
th
e feature extraction
to b
e
come o
ne
of the
most
b
a
si
c
pr
oblem i
n
f
a
ce
re
cog
n
it
ion,
extraction
of
effective ide
n
t
ification feat
ure
s
i
s
the
ke
y to
solve the pro
b
lem.
It should be
pointed o
u
t that the cha
r
a
c
terist
i
cs
of the choi
ce of go
od or b
ad an
d adop
t
corre
s
p
ondin
g
cla
s
sifier h
a
s clo
s
e rela
tionshi
p.
For example, u
n
der
a cl
assifi
er
cla
ssifi
cati
on
effect goo
d chara
c
te
risti
c
, unde
r an
othe
r cl
assifier
m
a
y get wo
rse
cla
ssifi
cation
effect. Ho
w
to
maximize
the
extra
c
tion to
the characte
ri
stics of th
e
cl
assificatio
n
, i
s
the
ultimate
goal
of p
a
ttern
recognitio
n
, and it is also th
e core proble
m
of pattern reco
gnition
system.
3. The NPE Algorithm
NPE algorith
m
and lo
cal
proje
c
tion
(L
LP) ther
e are
some
simila
r pro
pertie
s
,
but their
obje
c
tive function is comp
letely differen
t. NPE is
a li
near al
gorith
m
, so its fast and suitability for
pra
c
tical a
ppli
c
ation.
As stated ea
rlier, NPE method
s is the li
near a
pproximation of LLE method. And PCA
method, P
C
A method
is d
e
sig
ned to
ke
ep the
glob
al
stru
ctu
r
al
ch
ara
c
teri
stics,
and th
e meth
od
of NPE i
s
to
keep th
e lo
cal
feature
s
. Th
e
so
-calle
d
ke
ep lo
cal
featu
r
es i
s
to
use t
he
combi
nati
o
n
of the adjace
n
t point to one point on lin
ear said this.
In ha
s b
een
unde
r the
co
ndition
of dat
a set, we
ca
n choo
se
clo
s
e
neig
hbo
rs,
and
the
adja
c
en
cy graph. In so
me
ca
se
s, the d
a
ta point
di
stribution on
a nonlin
ear su
b
-
manifol
d
s,
b
u
t
only unde
r th
e co
ndition o
f
con
s
ide
r
ing
local n
earby, assumed to
be linea
r, this app
ro
ach i
s
likely to
be
reasona
ble. I
n
this
way,
we
ca
n
th
ro
ugh a point by
its neig
h
b
o
ring
the
wei
ght
coeffici
ent of
the line
a
r sai
d
lo
cal
ch
ara
c
teri
stics
of
a
d
jacent to
ref
l
ect thi
s
p
o
int
.
By minimizi
ng
the recon
s
tru
c
tion erro
r, the fixed weight
coefficie
n
t is as follo
ws:
2
(
)
||
||
ji
j
i
ji
Wx
w
x
(1)
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: 2574 – 2
581
2576
The re
co
nst
r
uction e
r
ror i
s
the sum of e
a
ch
p
o
int by its neigh
bori
n
g re
con
s
tru
c
ti
on erro
r
.
In ord
e
r to
re
alize th
e line
a
r rep
r
e
s
enta
t
ion in nei
gh
borin
g poi
nts,
we h
a
ve to
think a
bout t
he
initial point is mappe
d to a line. If you want
to get
a good m
a
p
,
should m
a
ke the followi
ng
obje
c
tive function to get the minimum.
2
(
)
||
||
ii
j
j
ij
Yy
w
y
(
2
)
j
x
refers to
val
ues for the j
t
h face
ima
g
e
an
d
ij
x
prefo
r
ms th
e n
e
ig
hborhoo
d of
j
x
in
k
distan
ce.
12
{,
,
,
}
n
Xx
x
x
(3)
12
{,
,
,
}
T
m
Yy
y
y
(4)
TT
YA
X
(
5
)
A
is
a
go
od ma
p
which can
fit
Equation (5) better.
A
s
sumin
g
i
x
is the
i
th vec
t
or
o
f
matrix
X
,
we can define
ii
i
j
j
j
Z
yw
y
a
nd the
extend matrix
()
Z
IW
Y
.Then th
e
followin
g
can
be simplify as:
2
()
(
)
ii
j
j
ij
Yy
w
y
2
()
i
i
Z
T
Z
Z
()
()
TT
T
YI
W
I
W
X
A
()
()
TT
T
A
XI
W
I
W
X
A
TT
A
XM
X
A
(
6
)
()
()
T
M
IW
IW
.Obviously
,
t
o
get
TT
A
XM
X
A
mean
s to solve the fo
llowing
equati
o
n
:
TT
XM
X
a
X
X
a
(
7
)
T
XU
S
V
(8)
T
XU
X
S
V
(
9
)
1,
2
{,
,
}
l
UU
U
U
is
th
e e
i
ge
nve
c
to
r o
f
ma
tr
ix
T
X
X
and
12
{,
,
,
}
l
VV
V
V
is t
he
eigenve
c
tor o
f
matrix
T
X
X
.
S
are
diago
nal matrix coming fro
m
the non-ze
ro eige
nvalue
s of
X
.
S
and
V
are both
diago
nal matrix. So
X
is full rank mat
r
ix. The pro
b
lem
can be sim
p
lify as:
TT
X
MX
a
X
X
a
(10)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition
Based o
n
Me
tric-optim
ized
Neigh
borhoo
d Preserving
Em
bedding (Che
n Bo)
2577
4. The MONPE and CMO
N
PE Algorithm
NPE algo
rith
m ca
n be
eit
her
un
sup
e
rvi
s
ed
al
go
rithm
also
can b
e
sup
e
rvised
al
gorithm.
W
h
en
a kn
ow
n
c
l
ass
la
be
l, N
PE a
l
gor
ith
m
be
sup
e
rvise
d
al
gori
t
hm. Experim
ents
sh
ow th
at:
sup
e
rvised NPE algorithm
than unsu
p
e
r
vised
N
PE algorithm h
a
s
obviou
s
adva
n
tage
s: on the
one h
and, in
the actu
al face datab
ase, the same
fa
ce
data of
simil
a
r d
egree i
s
greate
r
tha
n
the
face d
a
ta of
different p
e
o
p
le. Data
of
simila
r de
gre
e
is
high
er, t
he smalle
r th
e re
co
nstruct
i
on
error of the li
near
rep
r
e
s
e
n
tation sh
oul
d. So
the su
pervisi
on of
a neigh
borho
od of embe
d
d
ing
algorith
m
tha
n
unsupe
rvised ke
ep n
e
ig
hborhoo
d em
beddi
ng alg
o
r
ithm re
co
nst
r
uctio
n
erro
r
is
smalle
r; Keep
neighb
orhoo
d, on t
he oth
e
r ha
nd, due
to the sup
e
rvi
s
ion
whe
n
sel
e
cting n
e
igh
b
o
r
embed
ding a
l
gorithm to d
i
stingui
sh the
classifi
catio
n
label, makes the ran
g
e
of paramet
er
values th
an
keep
sup
e
rvisi
on mu
ch
sm
aller n
e
igh
b
o
r
hoo
d em
bed
ding al
gorith
m
, which g
r
e
a
tly
redu
ce the
prog
ram
run
n
ing time. Ho
wever,
whether
NPE algorithm
of supe
rvise
d
or
unsupe
rvise
d
NPE al
gorith
m
, unde
r th
e
Euclide
an
me
tric, hig
h
di
m
ensi
onal
face
data m
a
y sh
ow
thin an
d
em
pty spa
c
e.
T
o
solve thi
s
probl
em
we
NPE alg
o
rith
m was put
forward. Fo
r
the
purp
o
ses of
writing,
the i
m
prove
d
al
g
o
rithm fo
r
M
O
NPE. Fo
r t
he p
u
rp
ose
of differe
nce, the
origin
al still consi
der the
supervi
sion
cla
ss
la
bel NPE
algorith
m
call
ed NPE algo
rithm.
On the thoug
hts of the above,
we prop
ose
d
MONP
E algorithm o
f
two situations: First,
sele
cting
nei
ghbo
r
con
s
id
er
cla
s
s lab
e
l
.
In ord
e
r to
facilitate the
discu
ssi
on in
this
articl
e, i
t
is
calle
d CLM
O
NPE.Secon
d
, sele
cting nei
ghbo
r cla
s
s label. It is still remem
b
e
r
ed
as MO
NPE.
MONPE can
be don
e in the followin
g
st
eps:
Step 1
:
Selec
t
the
k
neigh
b
o
rho
od of the
original d
a
ta.
Step 2: Reco
nstru
c
t the weight co
efficie
n
t under the
constraint of minimum erro
rs.
2
a
r
g
m
in
||
||
ji
j
i
ji
x
wx
(11)
j
x
is the neig
h
b
o
rho
od of
i
x
in the sam
e
cla
s
s
1,
2
,
3
,
,
j
k
.
Step 3
:
Take
PCA
algo
rithm to deal with
X
and get
X
. Compute the followi
n
g
equatio
n to get its minimu
m value.
2
arg
m
in
|
|
||
ii
j
j
ij
yw
y
(
1
2
)
Due to flip,
scale tran
sform, su
ch a
s
l
o
cal
geom
etri
c st
ru
cture i
s
relatively un
cha
nge
d,
so in the third
step cal
c
ulat
ion data dime
nsio
n red
u
cti
on is in the seco
nd ste
p
in the right und
er
the premi
s
e o
f
coefficient
matrix is co
nstant.
Note: due t
o
the need
to use L
D
A algorithm o
f
data proce
ssi
ng, so n
o
matter
CLMO
NPE al
gorithm
or M
O
NPE alg
o
rit
h
m, are
ca
rri
ed out u
nde
r con
s
id
erin
g
the cla
s
s lab
e
l.
The differe
nce is: wheth
e
r
to con
s
ide
r
when sele
cting
neighb
or cl
a
ss la
bel.
5. Experiment and Simulation
5.1. Experiment Designa
tion
(a) T
he pu
rpo
s
e of the exp
e
rime
nt
The algo
rithm
is pre
s
ented
in this pape
r, on t
he basi
s
of con
s
ide
r
in
g the classification of
face data i
n
formation, a
nd co
mbini
n
g with
the
algorith
m
M
O
NPE algo
ri
thm is prop
ose
d
,
therefo
r
e, th
e obje
c
tive
of this pa
pe
r is
to co
mp
are and eva
l
uate
the su
pervisi
on
of the
algorith
m
, MONPE alg
o
rit
h
m and
CL
M
O
NPE adva
n
t
ages and di
sadva
n
tage
s of
the
algo
rithm.
There are two indicators:
(1) the
same
con
d
ition, different cl
assification accu
ra
cy of the algorithm.
(2) the
same
con
d
ition, the
length of the different algo
rithm runni
ng time.
Acco
rdi
ng to the above two indicators, in ord
e
r to a
c
curately refle
c
t the ch
ara
c
teristics
of the sa
me
algorith
m
an
d
the co
mpa
r
ison bet
wee
n
the alg
o
rithm,
the obje
c
tive
of this pa
pe
r is
as
follows
:
(1) In ORL
face datab
ase, the different
alg
o
rit
h
ms, und
er different training for
comp
ari
s
o
n
, the ch
ang
e of testing its cl
a
ssi
fi
cation a
c
curacy an
d progra
m
run
n
in
g time.
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: 2574 – 2
5
81
2578
(2) In the Y
a
l
e
face d
a
tab
a
se, for
different algo
rithm
s
, com
p
a
r
e u
nder
different
training
numbe
r, test the ch
ang
e of its cla
ssifi
cati
on accu
ra
cy and prog
ram
runni
ng time.
(b) T
he ste
p
s of the experiment
Both NPE al
gorithm
an
d
MONPE al
go
rithm, in
ord
e
r
to g
e
t the
result
s of th
e
optimal
algorith
m
of the execution
of various p
a
rameters
t
o
m
a
ke t
he f
i
nal
cla
ssif
i
cat
i
on
ac
cur
a
cy
of
t
h
e
highe
st value
.
In detail in this pa
per, the
spe
c
ific test
belo
w
step
s:
Step 1: Prep
are te
st data
;
In orde
r to
co
mp
are the
re
sults
und
e
r
differe
nt di
mensi
o
n
face data, thi
s
pap
er expe
riments on the
ORL an
d Yale face data
b
a
s
e data.
Step 2: Implementation alg
o
rithm of the optim
al cla
s
si
f
i
cat
i
on ac
cu
r
a
cy
;
I
n
det
er
mining
the nu
mbe
r
o
f
trainin
g
a
n
d
traini
ng
mod
e
, sel
e
ct
the
optimal valu
e
of the
hig
h
e
s
t
classification
accuracy. According to th
e determin
a
tion of a
large
number of e
x
perime
n
ts, the optimal when
testing the ne
ighbo
r to 1.
Step 3:
Dete
rmine
the
re
sults; In
O
R
L
and
Ya
le
fa
ce
datab
ase, all ta
ke
40
different
training m
e
th
ods. Th
en, their average.
Step 4: Di
scussthe
abov
e calculated
re
sults,
an
alyze
th
e ch
ara
c
teri
stics of
ea
ch
algorith
m
and
compa
r
in
g the advantag
es
and disadva
n
tage
s of the algorith
m.
5.2. Face Da
tabase Sele
ction for Expe
riment
It Is by the
AT&T Camb
ridge that ORL fa
ce data
b
a
se la
boratory with simult
aneo
us
spe
e
ch
at Ca
mbrid
ge univ
e
rsity,
visi
on and rob
o
ti
cs grou
p,
ma
de up
of 40 different peopl
e
e
a
ch
of 10 different image
s
is con
s
tru
c
te
d. The
s
e
im
age
s a
r
e in
different p
e
riod
s, different
illumination, u
nder the
con
d
i
tion of different fa
cial expression a
nd fa
cial detail
s. Seen a
s
:
Figure 1. Face Image in the Datab
a
se o
f
ORL
Figure 2. Face Image in the Datab
a
se o
f
Yale
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition
Based o
n
Me
tric-optim
ized
Neigh
borhoo
d Preserving
Em
bedding (Che
n Bo)
2579
Yale face
dat
aba
se
contai
ns 1
5
16
5 im
age
s of the i
ndividual, ea
ch p
e
rson, 1
1
image
s
contai
ning
un
der diffe
rent l
i
ght co
ndition
s (e.
g
.,
on th
e left side
of the light, the right lig
htin
g,
central lig
ht),
different exp
r
ession
s,
su
ch
as ha
ppy, sa
d, sle
ep,
su
rp
rise
d,
blin
k of
an eye) of
th
e
face imag
e. Seen a
s
Figu
re
2
:
5.3. The Sim
u
lation Result
In ord
e
r to
re
duce the influ
ence of rand
om
ne
ss an
d
some
di
sturb
ance, on th
e
premi
s
e
of all param
eter optimal,
to con
s
ide
r
each tr
ainin
g
seve
ral ra
ndomly selected 40 different
training
meth
ods,
and th
en
comp
are
the ave
r
ag
e. In ad
dition, be
cau
s
e
of the imp
r
o
v
e
d
algorith
m
in time co
mplexit
y
with little ch
ange,
so
the
test re
sults
o
n
ly
comp
are the a
c
cura
cy
of
cla
ssif
i
cat
i
on.
(a) T
he re
sult
s of ORL d
a
taba
se
Figure 3. The
Result for 2 Traini
ng Time
s
Figure 4. The
Result for 3 Traini
ng Time
s
Figure 5. The
Result for 4 Traini
ng Time
s
Figure 6. The
Result for 5 Traini
ng Time
s
Table.
Cla
s
sif
i
cat
i
on A
c
cu
r
a
cy
f
o
r
O
R
L Face
Ima
ge Datab
a
se
Training time 2
Training time 3
Training time 4
Training time 5
NPE
79.45%
87.69%
91.73%
92.83%
CLMONPE
79.70%
87.41%
91.90%
93.30%
MONPE
51.56%
87.18%
91.70%
91.73%
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: 2574 – 2
581
2580
(b) T
he re
sult
s of Yale data
base
Figure 7. The
Result for 2 Traini
ng Time
s
Figure 8. The
Result for 3 Traini
ng Time
s
Figure 9. The
Result for 4 Traini
ng Time
s
Figure 10. Th
e Re
sult for 5
Trainin
g
Tim
e
s
Table 2. Class
i
fic
a
tion Ac
curac
y
for Yale Fac
e
Image Databas
e
Training time 2
Training time 3
Training time 4
Training time 5
NPE
55.87%
67.58%
73.83%
77.36%
CLMONPE
55.07%
68.33%
73.48%
77.89%
MONPE
28.35%
65.69%
72.00%
76.64%
6. Conclusio
n
It
can
be
in
cl
uded
that:
(1
) CLMO
NPE algorith
m
i
s
compa
r
ed with
the NPE algorithm
and M
O
NPE
algo
rithm, in
took the
opt
imal pa
ra
m
e
ters,
CL
MO
NPE algorith
m
are
supe
rio
r
to
NPE algorithm, less
MONPE al
g
o
rith
m in trainin
g
for po
or
perfo
rman
ce.
(2
)
CLM
O
NPE
algorith
m
is
comp
ared
with the
NPE al
gorithm, the
i
n
crea
se
of d
e
clin
e in
dim
ensi
on, in th
eir
respe
c
tive re
ach
e
s
a ste
ady state,
steady
CL
MO
NPE algo
rith
m is
sup
e
ri
or to the
NPE
algorith
m
. MONPE alg
o
rit
h
m and
CL
M
O
NPE alg
o
rit
h
m. (3
) The
comp
ari
s
o
n
result
s sho
w
that
the LDA alg
o
r
ithm is relati
vely close to
the di
stan
ce
betwe
en poi
n
t
s in cla
ss, a
nd the discre
t
e
points of the
dista
n
ce bet
wee
n
the
cla
s
s this
pap
er puts forward mea
s
u
r
e
s
to optimize t
h
e
neigh
borhoo
d
embeddi
ng algorith
m
.Thi
s algo
rith
m adopt
s linea
r discrimina
nt analysis d
a
t
a
dimen
s
ion
re
ductio
n
algo
rithm usin
g th
e sel
e
cted
k neighb
or, to
some
exten
t, improves t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Re
cog
n
ition
Based o
n
Me
tric-optim
ized
Neigh
borhoo
d Preserving
Em
bedding (Che
n Bo)
2581
reliability of t
he Euclidean metr
ic. F
u
rther
work i
s
that high
di
mensional face
data
compari
s
on
CLMONPE al
gorithm and
NPE algo
rithmwill promot
e the optimizati
on of the
measure to other
dimen
s
ion re
ductio
n
meth
od.
Referen
ces
[1]
W
Bleds
oe. M
an-mac
h
in
e fa
cial
Rec
ogn
itio
n. Pa
noramic
Research Inc.
Pa
lo Alto,
CA. 1966; Rep
PRI:22
[2]
Berto R, P
o
g
g
i
o
T
.
F
a
ce reco
gniti
on: F
e
atur
e vers
us tem
p
l
a
tes.
lEEE
Trans. on PAMI
. 1
993;
15(
10):
104
2-10
52.
[3]
Z
i
qua
n Hon
g
. Alge
braic fe
atu
r
e extr
action o
f
image forrec
ogn
ition.
Patte
rnRec
ogn
itio
n.
1991; 2
4
(3)
:
211-
219.
[4]
Nakam
u
ra O,
Mathur S, Minami T
.
Identific
ation of hum
an
faces based o
n
isod
ensit
y m
aps.
Pattern
Reco
gniti
on.
1
991; 24(
3): 263
-272.
[5]
Lad
es M, Vorb
ugg
en J, Bu
h
m
ann J, et a
l
. Distort
io
n inv
a
r
i
anto
b
ject rec
o
gniti
on i
n
the
d
y
nam
ic Li
nk
architectur
e
.
IEEE Trans, onComputers
. 19
9
1
; 42(3): 30
0-3
1
1
[6]
Samari
a F
,
Yo
ung
S. HMM-
base
d
arc
h
itec
ture
for fac
e
id
entificati
on Im
age
an
d Vis
i
o
n
Com
puti
n
g
.
199
4; 12(8).
[7]
T
eenba
um JB, de Silva V, La
ngford JC. A Glob
al
Geometri
c F
r
ame
w
o
r
k for Nonl
in
ear Di
mensi
ona
li
t
y
Red
u
ction.
Se
i
ence.
20
00; 29
0(55
00): 23
19-
232
3.
[8]
Ro
w
e
is
SL, S
aul
L. No
nli
n
e
a
r dim
ensi
o
n
a
l
i
t
y
re
ducti
on
b
y
l
o
ca
ll
y l
i
ne
ar
embe
ddi
ng.
S
e
ie
nce.
2
000;
290(
550
0): 232
3-23
26.
[9]
Donoho DL, Grimes
C.
He
SSian
ei
gen
maps: New
l
o
ca
lly li
ne
ar e
m
b
edd
ing
tech
niq
ues for
hig
h
-
di
me
nsio
nal da
ta.
Proc. of the
Natio
nal Aca
d
e
m
y
of Scie
nces
. 2003; 10
0(10)
: 5591-5
5
9
6
.
[10]
Khair
udi
n. RB
F
NN Co
ntrol o
f
A T
w
o-
Link
F
l
e
x
ib
le M
ani
pu
l
a
tor Incorp
orati
ng Pa
yl
o
ad.
TE
L
K
OM
N
I
KA
T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol.
20
10; 08(2): 1
57-
164.
[11]
P Srikanth, Ash
w
a
n
i Kumar
Cha
nde
l. Inver
s
e S-T
r
ansform Based Dec
i
sion T
r
ee for Po
w
e
r S
y
stem
Faults Identification.
T
E
LKO
M
NIKA T
e
leco
mmu
n
icati
o
n
Co
mp
uting
Ele
c
tronics a
nd
Contro
l
. 20
11
;
09(1): 99-
10
6.
Evaluation Warning : The document was created with Spire.PDF for Python.