T
E
L
KO
M
NI
K
A
,
V
o
l.
1
4
,
N
o.
3
,
S
ept
em
ber
20
1
6
,
pp.
11
13
~
112
2
I
S
S
N
:
1
693
-
6
930
,
ac
c
r
edi
t
ed
A
b
y
D
IK
T
I,
D
e
c
r
e
e
N
o
:
58/
D
I
K
T
I
/
K
ep/
2013
D
O
I
:
10.
12928/
T
E
LK
O
M
N
I
K
A
.
v
1
4
i
3
.
3472
11
13
R
ec
ei
v
ed
J
anu
ar
y
6
,
2
01
6
;
R
ev
i
s
ed
May
2
2
,
2
01
6
;
A
c
c
ept
e
d
J
une
5
,
201
6
M
ul
ti
-
Cr
it
er
ia i
n
Discr
im
i
n
an
t
A
n
al
y
s
is t
o
F
i
n
d
t
h
e
D
omi
na
n
t Fe
a
t
u
r
e
s
A
r
i
f M
u
n
ta
s
a
*
,
I
n
d
a
h
A
g
u
s
ti
e
n
S
i
r
a
d
j
u
d
d
i
n
,
R
i
m
a
T
r
i
W
a
h
y
u
n
i
n
g
r
u
m
I
nf
or
m
at
i
c
s
E
ngi
neer
i
ng D
e
par
t
m
ent
,
E
ng
i
ner
i
ng
F
ac
ul
t
y
,
U
ni
v
er
s
i
t
y
of
T
r
unoj
oy
o
M
adur
a
,
R
y
T
el
ang P
o.
B
ox
2 K
am
al
,
B
an
gk
a
l
an
,
I
nd
one
s
i
a
*
C
or
r
es
po
ndi
ng a
ut
hor
,
e
-
ma
i
l
:
ar
i
f
m
unt
as
a@
i
f
.
t
r
u
noj
oy
o.
ac
.
i
d
A
b
st
r
act
A
c
r
uc
i
al
pr
obl
em
i
n
bi
om
et
r
i
c
s
i
s
e
nor
m
ou
s
di
m
en
s
i
on
al
i
t
y
.
I
t
w
i
l
l
ha
v
e a
n i
m
pac
t
on t
he
c
os
t
s
i
nv
o
l
v
ed.
T
he
r
ef
or
e,
t
he f
e
at
ur
e ex
t
r
ac
t
i
on p
l
ay
s
a s
i
gni
f
i
c
ant
r
o
l
e i
n b
i
om
et
r
i
c
s
c
om
p
ut
at
i
onal
.
I
n t
hi
s
r
es
ear
c
h,
a n
ov
e
l
appr
oa
c
h t
o
ex
t
r
ac
t
t
he f
e
at
ur
e
s
i
s
p
r
opo
s
ed f
or
f
ac
i
al
i
m
age r
ec
ogn
i
t
i
o
n
.
F
our
c
r
i
t
er
i
a o
f
t
he D
i
s
c
r
i
m
i
nan
t
A
na
l
y
s
i
s
h
a
v
e b
een
m
odel
ed
t
o f
i
nd
t
h
e
dom
i
na
nt
f
eat
ur
es
.
F
or
ea
c
h c
r
i
t
er
i
on
i
s
an
obj
e
c
t
i
v
e f
un
c
t
i
on,
i
t
w
as
d
er
i
v
ed t
o ob
t
ai
n t
h
e opt
i
m
um
v
a
l
ues
.
T
he opt
i
m
um
v
al
ue
s
c
a
n be s
ol
v
ed b
y
us
i
n
g ge
ner
a
l
i
z
ed t
he E
i
g
en
v
al
ue pr
o
bl
em
a
s
s
o
c
i
a
t
ed t
o t
h
e l
ar
ges
t
E
i
g
env
al
ue
.
T
h
e m
odel
i
ng r
es
ul
t
s
w
er
e
em
pl
oy
e
d t
o r
ec
o
gni
z
e t
he f
ac
i
al
i
m
age by
t
h
e m
ul
t
i
-
c
r
i
t
er
i
a p
r
oj
ec
t
i
on t
o t
he or
i
g
i
na
l
dat
a.
T
he t
r
ai
n
i
ng s
e
t
s
w
er
e al
s
o pr
o
c
es
s
ed
by
u
s
i
ng t
he E
i
gen
f
ac
e
pr
oj
e
c
t
i
on
t
o av
o
i
d t
h
e s
i
ngul
ar
i
t
y
pr
ob
l
em
c
as
e
s
.
T
he
s
i
m
i
l
ar
i
t
y
m
eas
ur
em
ent
s
w
er
e
per
f
or
m
ed
b
y
us
i
ng
f
our
di
f
f
er
ent
m
et
hod
s
,
i
.
e.
E
u
c
l
i
di
an
D
i
s
t
a
nc
e
,
M
anhat
t
an,
C
heb
y
s
h
ev
,
a
nd
C
anber
r
a.
F
e
at
ur
e e
x
t
r
a
c
t
i
on
and anal
y
s
i
s
r
e
s
ul
t
s
us
i
ng m
ul
t
i
-
c
r
i
t
er
i
a h
av
e
s
how
n
bet
t
er
r
e
s
ul
t
s
t
han
t
he
ot
her
a
ppe
ar
an
c
e
m
et
ho
d,
i
.
e.
E
i
genf
ac
e
(
P
C
A
)
,
F
i
s
her
f
a
c
e
(
L
i
ne
ar
D
i
s
c
r
i
m
i
nant
A
nal
y
s
i
s
or
LD
A
)
,
Lap
l
ac
i
anf
ac
e
s
(
L
oc
a
l
i
t
y
P
r
es
er
v
i
ng
P
r
oj
ec
t
i
on
or
LP
P
)
,
and
O
r
t
ho
gon
al
Lapl
a
c
i
anf
a
c
e
s
(
O
r
t
hog
ona
l
Lo
c
al
i
t
y
P
r
es
er
v
i
n
g P
r
o
j
ec
t
i
on
or
O
-
L
PP).
Ke
y
w
o
rd
s
:
mu
l
t
i
-
c
r
i
t
er
i
a
,
di
s
c
r
i
m
i
nant
ana
l
y
s
i
s
,
f
ea
t
ur
e
s
e
x
t
r
ac
t
i
on,
s
i
n
gul
a
r
i
t
y
pr
o
bl
em
,
f
a
c
i
al
r
e
c
og
ni
t
i
on
C
o
p
y
r
i
g
h
t
©
20
16 U
n
i
ver
si
t
a
s A
h
mad
D
ah
l
an
.
A
l
l
r
i
g
h
t
s r
eser
ved
.
1
.
I
n
tr
o
d
u
c
ti
o
n
H
um
ans
c
an r
ec
or
d hum
a
n f
ac
es
w
i
t
h s
t
or
i
ng i
m
por
t
ant
f
eat
ur
es
.
H
um
ans
hav
e al
s
o
been
ab
l
e
t
o r
ec
o
gni
z
e
a
per
s
on'
s
f
ac
e
w
i
t
h a
v
er
y
f
as
t
t
i
m
e.
T
he pr
oc
es
s
w
as
di
f
f
i
c
ul
t
t
o
i
m
pl
em
ent
on
a
c
o
m
put
er
.
Man
y
appr
o
ac
hes
ha
v
e
be
en
m
odel
ed
t
o
i
m
i
t
at
e
t
he
per
f
or
m
anc
e
of
t
he h
um
an br
ai
n,
f
or
bo
t
h obt
ai
ni
n
g t
h
e dom
i
na
nt
f
eat
ur
es
an
d t
he
dec
i
s
i
on m
ak
i
ng.
R
es
ear
c
her
s
h
av
e d
ev
el
op
ed t
he
bi
om
et
r
i
c
s
t
o m
i
m
i
c
t
he
h
um
an i
nt
e
l
l
i
g
enc
e.
T
he
bi
om
et
r
i
c
s
c
o
m
put
at
i
ona
l
pr
o
bl
em
i
s
t
he i
m
age di
m
ens
i
on
al
i
t
y
.
I
f
i
m
age di
m
ens
i
ona
l
i
t
y
us
e
d
i
s
hi
g
her
,
t
h
en
c
os
t
t
ak
en i
s
al
s
o m
or
e ex
pens
i
v
e.
Man
y
a
l
gor
i
t
hm
s
hav
e be
en
i
m
pr
ov
e
d t
o
r
e
duc
e
t
he
c
ur
s
e
di
m
ens
i
ona
l
i
t
y
an
d al
s
o o
bt
ai
n t
h
e hi
ghes
t
ac
c
ept
anc
e
r
at
e,
f
or
bot
h hol
i
s
t
i
c
m
et
hods
[
1
-
11]
a
nd
f
eat
ur
ed
-
bas
e
d ap
pr
oac
h
[
12]
,
and
ev
en c
om
bi
nat
i
on
o
f
t
hem
(
h
y
br
i
d
m
et
hod)
[
13
]
.
T
he
P
r
i
nc
i
pa
l
C
om
ponent
A
n
al
y
s
i
s
i
s
t
he
o
l
des
t
a
nd
t
h
e
s
i
m
pl
es
t
of
t
he
appear
a
nc
e
appr
o
ac
h.
D
ue
t
he m
os
t
s
t
r
ai
ght
f
or
w
ar
d ap
pr
oac
h,
t
h
e P
r
i
nc
i
pa
l
C
om
ponen
t
A
n
al
y
s
i
s
(
P
C
A
)
has
b
een i
m
pr
ov
ed
b
y
m
an
y
r
es
ear
c
her
s
,
i
.
e.
t
he L
i
ne
ar
D
i
s
c
r
i
m
i
nant
A
na
l
y
s
i
s
[
14
-
1
6]
,
t
h
e Loc
a
l
i
t
y
P
r
es
er
v
i
ng
P
r
oj
ec
t
i
o
n (
LP
P
or
La
pl
a
c
i
anf
ac
es
)
,
t
he D
i
s
c
r
i
m
i
nat
i
v
e C
om
m
on V
ec
t
or
[
17]
,
R
eg
ul
ar
i
z
e
d
D
i
s
c
r
i
m
i
nant
A
na
l
y
s
i
s
[
18]
,
LD
A
-
B
as
ed A
l
g
or
i
t
hm
s
[
19]
,
a
nd K
er
ne
l
P
r
i
nc
i
pa
l
C
om
ponen
t
A
na
l
y
s
i
s
[
20]
.
T
he
P
C
A
i
s
s
ubs
pac
e
m
et
hod
t
o
pr
oj
ec
t
t
he
or
i
gi
na
l
s
a
m
pl
e
s
et
s
t
o
t
he
E
i
gen
v
ec
t
or
of
t
he
c
ov
ar
i
anc
e
m
at
r
i
x
.
I
t
c
a
n
dem
ons
t
r
at
e
an
y
par
t
i
c
ul
ar
f
ac
i
al
i
m
age
i
n
t
he
c
oor
d
i
nat
e
s
pac
e
of
t
he
E
i
g
enf
ac
e.
T
he pr
oj
ec
t
i
on r
es
ul
t
s
c
an s
i
gni
f
i
c
an
t
l
y
r
educ
e t
h
e
i
m
age di
m
ens
i
o
nal
i
t
y
a
nd
al
s
o
pr
oduc
e
t
he
dom
i
nant
f
eat
ur
es
t
o
r
ec
og
ni
z
e
t
he
obj
ec
t
.
N
e
v
er
t
h
el
es
s
,
i
t
h
as
t
he
w
eak
nes
s
.
T
he
P
r
i
nc
i
pa
l
C
om
pone
nt
A
na
l
y
s
i
s
c
an
on
l
y
ef
f
i
c
i
ent
l
y
w
or
k
,
w
he
n
t
h
e
n
um
ber
of
t
r
ai
ni
ng
s
et
s
i
s
no
t
l
ar
ger
t
h
an t
he
i
m
age di
m
ens
i
onal
[
20
]
.
I
f
i
t
does
not
oc
c
ur
,
t
hen i
t
w
i
l
l
f
ai
l
t
o r
educ
e t
h
e
di
m
ens
i
ona
l
i
t
y
.
T
he LD
A
i
s
on
e of
ap
pe
a
r
anc
e m
et
hod as
t
he
de
v
el
o
pm
ent
r
es
ul
t
of
t
h
e
P
r
i
nc
i
pa
l
C
om
ponent
A
na
l
y
s
i
s
.
I
t
c
a
n
m
ap
and
r
ed
uc
e
t
he
di
m
ens
i
o
nal
i
t
y
b
ec
om
e
a
num
ber
of
c
l
as
s
es
.
T
he LD
A
i
s
not
a
l
s
o dep
e
nded o
n
t
he
n
um
ber
of
t
r
a
i
ni
ng s
et
s
but
d
epe
nds
on
t
he num
ber
of
c
l
as
s
es
.
T
he
LD
A
c
an
opt
i
m
i
z
e
t
he
pr
oj
ec
t
i
on
r
es
ul
t
s
b
y
opt
i
m
i
z
at
i
on
of
t
he
bet
w
een
an
d
w
i
t
h
i
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
13
–
1
122
1114
c
l
as
s
s
c
at
t
er
.
H
o
w
e
v
er
,
t
h
e LD
A
has
al
s
o
t
h
e
w
eak
nes
s
.
I
t
c
a
nn
ot
c
a
pt
ur
e
t
h
e m
ani
f
ol
d
non
-
l
i
n
ear
s
t
r
uc
t
ur
e.
I
t
c
a
nno
t
o
pt
i
m
al
l
y
per
f
or
m
di
m
ens
i
on
al
i
t
y
r
educ
t
i
on
w
he
n
t
he
t
r
ai
n
i
ng
c
l
as
s
es
ar
e m
or
e t
han t
h
e
i
m
age di
m
ens
i
onal
i
t
y
.
T
he
LP
P
p
er
f
or
m
s
t
he
dat
a
pr
oj
ec
t
i
o
n
of
t
he
or
i
gi
nal
s
a
m
pl
e
s
et
s
t
o
t
h
e
E
i
ge
nv
ec
t
or
of
t
he A
f
f
i
ni
t
y
M
at
r
i
x
.
I
t
w
as
g
ener
at
ed b
y
t
h
e
H
eat
K
er
n
el
m
at
r
i
x
.
T
he f
eat
ur
es
ex
t
r
ac
t
i
on
r
es
u
l
t
s
of
t
he
L
P
P
ar
e
t
h
e
l
oc
al
s
t
r
uc
t
ur
e,
w
h
er
e
t
h
e
y
c
a
n
r
epr
es
ent
and
r
et
a
i
n
t
h
e
l
oc
al
m
ani
f
ol
d.
H
o
w
e
v
er
,
t
he
LP
P
c
an
not
per
f
ec
t
l
y
r
es
t
or
e t
h
e i
nt
e
gr
at
i
o
n of
t
he
f
eat
ur
es
.
T
he
LP
P
ap
pr
oac
h
al
s
o h
as
t
he
s
i
m
i
l
ar
w
ea
k
nes
s
w
i
t
h
t
he
P
C
A
,
w
hi
c
h
i
s
t
he
per
f
or
m
anc
e
r
es
ul
t
s
d
epen
d
o
n
t
h
e
t
r
ai
n
i
ng s
et
s
us
ed
as
t
he
t
r
ai
n
i
ng
s
et
s
.
I
n
t
h
i
s
r
es
ear
c
h,
m
ul
t
i
-
c
r
i
t
er
i
a
i
n
t
h
e
D
i
s
c
r
i
m
i
nant
A
na
l
y
s
i
s
w
er
e
pr
opos
e
d.
M
ul
t
i
-
c
r
it
e
r
ia
i
s
a
w
a
y
t
o
c
apt
ur
e
t
h
e
m
ani
f
ol
d
s
t
r
uc
t
ur
e
f
r
om
f
our
di
r
ec
t
i
ons
.
F
or
eac
h
d
i
r
ec
t
i
o
n
i
s
des
c
r
i
be
d
b
y
us
i
ng
t
he
obj
ec
t
i
v
e
f
unc
t
i
on.
T
he
gen
er
at
i
ng
r
es
u
l
t
s
of
t
he
obj
ec
t
i
v
e
f
unc
t
i
o
n
w
er
e
c
al
c
u
l
at
ed
b
y
us
i
ng g
ener
a
l
i
z
e
d t
he
E
i
g
en
v
a
l
ue
pr
ob
l
em
c
or
r
es
pond
i
n
g t
o t
h
e l
ar
g
es
t
E
i
g
env
al
ue.
F
our
di
r
ec
t
i
ons
c
apt
ur
i
ng
h
as
p
r
ov
e
d
t
hat
t
he
dom
i
nan
t
f
eat
ur
es
pr
oduc
ed
c
a
n
pr
es
er
v
e
t
h
e
m
ani
f
ol
d
s
t
r
uc
t
ur
e s
o t
ha
t
t
h
e
y
c
a
n p
r
es
ent
t
h
e obj
ec
t
.
T
he ar
r
angem
ent
of
t
he paper
i
s
c
om
pos
ed as
f
ol
l
o
w
s
.
T
he s
ec
ond s
ec
t
i
on ex
pl
ai
ns
t
he
pr
opos
e
d a
ppr
oac
h
i
n d
et
ai
l
.
T
he s
i
m
i
l
ar
i
t
y
m
eas
ur
e
m
ent
s
ar
e w
r
i
t
t
e
n i
n t
he t
hi
r
d
par
t
.
T
he
f
our
t
h
s
ec
t
i
on r
epr
es
ent
s
t
he E
x
p
er
i
m
ent
al
r
es
u
l
t
s
.
T
he f
i
f
t
h par
t
di
s
c
us
s
es
and
c
o
m
par
es
t
he
pr
opos
e
d a
ppr
oac
h
r
es
ul
t
s
i
n
ot
h
er
m
et
hods
.
T
he
l
as
t
s
ec
t
i
on
r
es
um
es
t
he
r
es
ul
t
s
of
t
h
e
r
es
ear
c
h.
2.
R
e
sea
r
ch
M
et
h
o
d
T
he
Li
n
ear
D
i
s
c
r
i
m
i
nant
A
nal
y
s
i
s
i
s
t
he
m
et
hod
t
o
m
ax
i
m
i
z
e
t
he
v
a
l
u
es
bet
w
e
e
n
-
cl
a
ss
s
c
at
t
er
an
d t
o m
i
ni
m
i
z
e
w
i
t
h
i
n
-
c
l
as
s
s
c
at
t
er
.
I
t
i
s
t
he
enh
anc
em
ent
r
es
u
l
t
o
f
t
he
P
r
i
nc
i
pa
l
C
om
ponent
A
na
l
y
s
i
s
.
H
o
w
ev
er
,
t
he
L
i
ne
ar
D
i
s
c
r
i
m
i
na
nt
A
na
l
y
s
i
s
or
w
e
l
l
k
now
n a
s
LD
A
al
s
o
has
a l
i
m
i
t
at
i
on,
w
h
i
c
h i
s
m
ani
f
ol
d n
on
-
l
i
ne
ar
s
t
r
uc
t
ur
e
w
as
di
f
f
i
c
ul
t
t
o c
a
pt
ur
e.
T
her
ef
or
e,
i
t
i
s
nec
es
s
ar
y
t
o
be
i
m
pr
ov
ed.
I
n
t
h
i
s
r
es
ear
c
h,
f
our
d
i
f
f
er
ent
d
i
r
ec
t
i
ons
ar
e
pr
op
os
e
d
t
o
obt
a
i
n
t
h
e
pr
oj
ec
t
i
on s
p
ac
e b
y
m
ul
t
i
-
c
r
i
t
er
i
a i
n D
i
s
c
r
i
m
i
nant
A
na
l
y
s
i
s
.
S
up
pos
e
C
r
e
pr
es
ent
e
d c
l
as
s
es
,
an
d
i
s
t
at
ed
c
l
as
s
i
nd
ex
.
T
he
m
em
ber
s
o
f
C
c
l
as
s
ar
e
X
1
,
X
2
,
X
3
,
.
.
,
X
C
.
T
he
pr
opos
ed
m
et
hod
has
opt
i
m
i
z
e
d t
h
e
B
et
w
e
en
-
C
l
as
s
S
c
at
t
er
(
S
b
)
and T
ot
al
-
C
l
as
s
S
c
at
t
er
(
S
b
+
S
W
)
t
o
obt
ai
n
t
h
e
dom
i
nant
f
eat
ur
es
.
I
t
c
an
be
p
er
f
or
m
ed
b
y
der
i
v
at
i
on
of
t
he
m
ul
t
i
-
c
r
it
e
r
i
a
in
D
is
c
r
im
in
a
n
t
A
na
l
y
s
i
s
,
w
hi
c
h ar
e
*
4
*
3
*
2
*
1
an
d
,
,
A
A
A
A
.
T
he
y
c
an be s
t
at
ed as
m
ax
i
m
um
ar
gum
ent
o
f
t
he
B
et
w
een
-
C
l
as
s
S
c
at
t
er
an
d
T
ot
al
-
C
l
as
s
S
c
at
t
er
as
f
ol
l
o
w
s
:
+
=
A
S
A
A
S
S
A
A
w
T
w
b
T
a
)
(
m
ax
ar
g
*
1
(
1)
T
w
T
w
b
T
a
A
S
A
A
S
S
A
A
+
=
)
(
m
ax
ar
g
*
2
(
2)
(
)
+
=
A
S
S
A
A
S
A
A
w
b
T
b
T
a
m
ax
ar
g
*
3
(
3)
(
)
T
w
b
T
b
T
a
A
S
S
A
A
S
A
A
+
=
m
ax
ar
g
*
4
(
4)
T
he v
al
ue of
*
4
*
3
*
2
*
1
an
d
,
,
A
A
A
A
c
an c
apt
ur
e t
he
obj
ec
t
f
eat
ur
es
f
r
om
t
he di
f
f
er
ent
di
r
ec
t
i
on.
T
ot
a
l
c
r
i
t
er
i
a
c
a
n
be
obt
a
i
n
ed
b
y
ad
di
ng
E
q
uat
i
on
(
1)
,
(
2)
,
(
3)
a
nd
(
4)
as
s
een
i
n
t
he
f
ol
l
o
w
i
ng equ
at
i
on
:
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
Mul
t
i
-
C
r
i
t
er
i
a
i
n
D
i
s
c
r
i
m
i
na
n
t
A
n
al
y
s
i
s
t
o F
i
n
d t
h
e D
o
mi
nant
F
eat
ur
es
(
A
r
i
f
Mun
t
as
a
)
1115
(
)
(
)
(
)
(
)
+
+
+
+
+
+
+
=
+
+
+
+
+
+
+
=
T
w
b
T
b
T
w
b
T
b
T
T
w
T
w
b
T
w
T
w
b
T
a
T
w
b
T
b
T
w
b
T
b
T
T
w
T
w
b
T
w
T
w
b
T
a
A
S
S
A
A
S
A
A
S
S
A
A
S
A
A
S
A
A
S
S
A
A
S
A
A
S
S
A
A
S
S
A
A
S
A
A
S
S
A
A
S
A
A
S
A
A
S
S
A
A
S
A
A
S
S
A
A
)
(
)
(
m
ax
ar
g
m
ax
ar
g
m
ax
ar
g
)
(
m
ax
ar
g
)
(
m
ax
ar
g
*
(
5)
I
f
t
he
v
al
ue
of
A
i
s
s
et
of
c
r
i
t
er
i
a
{
A
1
,
A
2
,
A
3
,
…
,
A
l
}
,
t
h
en
t
he
obj
ec
t
i
v
e
f
unc
t
i
o
n
of
t
he
pr
op
os
ed
m
et
hod c
an be
w
r
i
t
t
en as
f
ol
l
o
w
s
:
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
+
+
+
+
+
+
+
=
T
w
b
T
T
b
T
w
b
T
b
T
T
w
T
T
w
b
T
w
T
w
b
T
A
A
S
S
A
t
ra
ce
A
S
A
t
ra
ce
A
S
S
A
t
ra
ce
A
S
A
t
ra
ce
A
S
A
t
ra
ce
A
S
S
A
t
ra
ce
A
S
A
t
ra
ce
A
S
S
A
t
ra
ce
A
)
(
)
(
m
ax
ar
g
*
(
6)
T
he
E
quat
i
o
n
(
6)
c
an
be
s
ol
v
ed
b
y
s
p
l
i
t
t
i
n
g
f
or
eac
h
obj
ec
t
i
v
e
f
unc
t
i
o
n
an
d
f
ol
l
o
w
ed
b
y
t
he m
at
r
i
x
t
r
ac
e.
T
he m
at
r
i
x
t
r
ac
e c
an
be
c
al
c
u
l
at
e
d b
y
u
s
i
ng
ge
ner
a
l
i
z
ed
t
h
e E
i
ge
nv
al
ue
pr
obl
em
as
s
how
n i
n t
he
f
ol
l
o
w
i
ng equ
at
i
on:
(
)
(
)
+
=
A
S
A
t
ra
ce
A
S
S
A
t
ra
ce
A
w
T
w
b
T
A
)
(
m
ax
ar
g
*
1
(
7)
(
)
(
)
+
=
T
w
T
T
w
b
T
A
A
S
A
t
ra
ce
A
S
S
A
t
ra
ce
A
)
(
m
ax
ar
g
*
2
(
8)
(
)
(
)
(
)
+
=
A
S
S
A
t
ra
ce
A
S
A
t
ra
ce
A
w
b
T
b
T
A
m
ax
ar
g
*
3
(
9)
(
)
(
)
(
)
+
=
T
w
b
T
T
b
T
A
A
S
S
A
t
ra
ce
A
S
A
t
ra
ce
A
m
ax
ar
g
*
4
(
10)
T
he
obj
ec
t
i
v
e
f
unc
t
i
o
n
of
t
he
E
qu
at
i
on
(
7)
,
(
8)
,
(
9)
,
and
(
10)
c
an
be
s
t
at
ed
r
es
pe
c
t
i
v
e
l
y
a
s
f
o
llo
w
s
:
(
)
(
)
A
S
A
A
S
S
A
A
J
w
T
w
b
T
+
=
)
(
(
11)
(
)
(
)
T
w
T
T
w
b
T
A
S
A
A
S
S
A
A
J
)
(
)
(
+
=
(
12)
(
)
A
S
S
A
A
S
A
A
J
w
b
T
b
T
+
=
)
(
(
13)
(
)
(
)
(
)
T
w
b
T
T
b
T
A
S
S
A
A
S
A
A
J
+
=
)
(
(
14)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
13
–
1
122
1116
Mi
n
i
m
i
z
i
n
g of
t
he
v
a
l
u
e of
J
(
A
)
as
s
ho
w
n
i
n
E
quat
i
o
n (
11)
c
an
be
o
bt
a
i
ne
d b
y
der
i
v
at
i
on
of
t
he f
unc
t
i
o
n of
J
t
o
A
an
d
i
t
i
s
s
et
t
o
0 v
al
ue
as
f
ol
l
o
w
s
:
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
*
*
)
(
*
)
(
0
*
0
*
*
0
*
*
0
*
*
0
*
2
*
2
0
*
2
*
2
0
*
*
)
(
2
2
A
S
S
A
S
A
S
S
A
S
A
J
A
S
A
J
A
S
S
A
S
A
S
A
A
S
S
A
A
S
S
A
S
A
A
S
A
S
S
A
A
S
A
A
S
S
A
S
A
A
S
A
S
S
A
A
S
S
A
S
A
A
S
S
A
A
S
A
S
A
A
S
S
A
S
S
A
A
S
A
S
A
A
S
S
A
S
A
A
S
S
A
A
S
A
S
A
A
S
S
A
S
A
A
S
S
A
A
S
A
dv
d
A
S
A
A
S
S
A
dv
d
A
J
dA
d
w
b
w
w
b
w
w
w
b
w
w
T
w
b
T
w
b
w
T
w
w
b
T
w
T
w
b
w
T
w
w
b
T
w
b
w
T
w
b
T
w
w
T
w
b
w
b
T
w
w
T
w
b
w
T
w
b
T
w
w
T
w
b
w
T
w
b
T
w
T
w
T
w
b
T
+
=
+
=
−
+
=
+
−
+
=
+
−
+
=
+
−
+
=
+
−
+
=
+
−
+
=
+
−
+
=
+
−
+
=
λ
(
15)
T
he
m
ax
i
m
u
m
v
al
ue
of
E
q
uat
i
on
(
1
5)
c
an
b
e
ob
t
ai
ne
d
b
y
m
i
ni
m
i
z
at
i
o
n
t
h
e
v
al
u
e
of
S
w
and
m
ax
i
m
i
z
at
i
on
t
h
e
v
a
l
u
e
of
S
b
+
S
w
.
T
he
der
i
v
at
i
o
n
r
es
ul
t
of
t
he
E
qu
at
i
on
(
15)
c
an
be
s
ol
v
ed
b
y
us
i
ng t
h
e G
ener
a
l
i
z
ed
E
i
g
en
v
al
ue pr
o
bl
em
c
or
r
es
pond
i
n
g t
o t
h
e l
ar
g
es
t
E
i
g
en
v
a
l
ue
as
f
o
llo
w
s
:
(
)
(
)
A
S
S
S
w
w
b
1
*
A
−
+
=
λ
(
16)
T
he E
quat
i
on (
1
2)
,
(
13)
,
a
nd
(
14)
c
an b
e
der
i
v
e
d
w
i
t
h t
he
s
am
e pr
oc
es
s
as
s
h
o
w
n i
n
E
qu
at
i
on (
1
5)
an
d f
ol
l
o
w
ed
b
y
gen
er
al
i
z
ed
E
i
gen
v
a
l
u
e
pr
obl
em
as
f
ol
l
ow
s
:
(
)
(
)
(
)
T
T
w
T
w
b
T
A
S
S
S
1
*
A
−
+
=
λ
(
17)
(
)
A
S
S
S
A
w
b
b
1
*
−
+
=
λ
(
18)
(
)
(
)
(
)
T
T
w
b
T
b
T
A
S
S
S
A
1
*
−
+
=
λ
(
19)
T
he
c
al
c
ul
at
i
o
n
r
es
u
l
t
s
of
E
quat
i
ons
(
1
6)
,
(
17)
,
(
18)
a
n
d
(
19)
ar
e
s
um
m
ed
t
o
obt
a
i
n
t
he
dom
i
nant
f
eat
ur
es
as
f
ol
l
o
w
s
∑
=
=
4
1
*
*
s
S
A
A
(
20)
T
he v
al
ue of
S
w
an
d
S
b
c
an
be r
epr
es
ent
e
d as
f
ol
l
o
w
s
:
(
)
(
)
∑
∑
=
∈
−
−
−
=
C
i
x
T
c
c
i
w
i
X
X
N
S
1
)
1
(
1
ω
µ
µ
(
21)
F
our
s
i
m
i
l
ar
i
t
y
m
eas
ur
em
e
nt
s
hav
e bee
n us
ed t
o obt
ai
n t
h
e m
at
c
hi
ng r
es
ul
t
s
,
i
.
e.
t
he
E
uc
l
i
d
i
an
D
i
s
t
a
nc
e (
d
1)
,
Manh
at
t
a
n (
d
2)
,
C
h
eb
y
s
he
v
(
d
3)
an
d C
anb
er
r
a (
d4)
.
T
he s
i
m
i
l
ar
i
t
y
m
eas
ur
e
m
ent
m
et
hods
c
an
be r
epr
es
ent
e
d i
n t
h
e f
ol
l
o
w
i
ng
eq
uat
i
on
:
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
Mul
t
i
-
C
r
i
t
er
i
a
i
n
D
i
s
c
r
i
m
i
na
n
t
A
n
al
y
s
i
s
t
o F
i
n
d t
h
e D
o
mi
nant
F
eat
ur
es
(
A
r
i
f
Mun
t
as
a
)
1117
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
T
j
i
C
i
j
j
i
C
i
m
j
T
j
i
i
m
j
j
i
i
C
i
T
m
j
i
j
i
m
j
i
j
i
i
C
i
T
i
i
i
b
X
W
X
X
m
X
m
X
m
X
m
m
m
S
i
i
i
i
*
*
1
*
1
)
(
1
*
)
(
1
*
1
)
(
1
1
1
1
1
1
1
∑
∑
∑
∑
∑
∑
∑
∑
=
=
=
=
=
=
=
=
=
=
−
−
=
−
−
=
µ
µ
µ
µ
µ
µ
(
22)
∑
∑
=
=
−
=
m
i
n
j
j
i
j
i
Y
X
d
1
1
,
,
1
(
23)
∑
∑
=
=
−
=
m
i
n
j
j
i
j
i
Y
X
d
1
1
,
,
2
(
24)
(
)
∑
∑
∑
∑
=
=
=
=
+
−
=
m
i
n
j
j
i
j
i
m
i
n
j
j
i
j
i
Y
X
Y
X
d
1
1
,
,
1
1
,
,
3
(
25)
(
)
∑
∑
=
=
−
=
m
i
n
j
j
i
j
i
Y
X
d
1
1
,
,
4
m
ax
(
26)
3.
R
e
su
l
t
s
a
n
d
A
n
a
l
y
s
i
s
I
n t
hi
s
s
ec
t
i
on,
t
h
e pr
op
os
ed
m
et
hod w
i
l
l
b
e
t
es
t
ed
b
y
us
i
ng t
hr
ee
f
ac
i
a
l
i
m
age
dat
a
bas
es
.
T
he
y
ar
e
us
u
al
l
y
us
ed
t
o
m
eas
ur
e
t
he
per
f
or
m
anc
e
of
t
he
pr
opos
e
d
m
et
hod.
T
hr
ee
dat
a
bas
es
ar
e
ar
e
t
h
e U
ni
v
er
s
i
t
y
of
B
er
n,
t
he
Y
A
LE
,
an
d t
h
e
A
T
&
T
or
O
l
i
v
et
t
i
R
es
ear
c
h
Labor
at
or
y
(
O
R
L)
f
ac
i
a
l
i
m
age
d
at
a
bas
es
.
F
or
eac
h
d
at
ab
as
e
w
i
l
l
be
r
an
dom
l
y
t
es
t
ed
b
y
us
i
ng
t
hr
ee,
f
our
,
f
i
v
e
an
d
s
i
x
t
r
ai
ni
ng
s
et
s
.
F
or
eac
h
F
ac
i
al
i
m
age
dat
ab
as
e
ha
s
t
he
di
f
f
er
ent
di
m
ens
i
ons
as
s
ho
w
n i
n T
abl
e
1
.
T
abl
e 1
.
F
ac
i
a
l
I
m
age D
at
a
bas
e A
t
t
r
i
b
ut
es
f
or
E
x
per
i
m
ent
s
No
F
ac
i
al
I
m
age
D
at
abas
e
N
um
ber
O
f
c
l
a
s
s
es
,
i
m
age
s
a
m
pl
e
s
f
or
al
l
c
l
as
s
es
T
r
ai
ni
ng
S
e
ts
H
i
gh,
W
i
dt
h
,
and
D
i
m
ens
i
on
s
1
T
he U
ni
v
er
s
i
t
y
o
f
B
e
rn
30 C
l
as
s
es
,
10
I
m
age
s
300
140,
120,
and
16.
800
P
i
x
el
s
2
T
he
YA
L
E
15 C
l
as
s
es
,
11
I
m
age
s
165
136,
104,
and
14.
144
P
i
x
el
s
3
T
he O
R
L
40 C
l
as
s
es
,
10
I
m
age
s
400
112,
92,
and 10.
304
P
i
x
el
s
3.
1.
E
v
a
l
u
a
ti
o
n
th
e
P
r
o
p
o
s
e
d
M
e
th
o
d
o
n
th
e
U
n
i
v
e
r
s
i
t
y
o
f B
e
r
n
F
a
c
i
a
l
I
m
a
g
e
D
a
ta
b
a
s
e
T
he
U
ni
v
er
s
i
t
y
of
B
er
n
F
ac
i
al
I
m
age
(
U
o
B
)
D
at
ab
as
e
i
n
v
ol
v
e
d
t
hi
r
t
y
per
s
ons
,
f
or
eac
h
per
s
on
has
t
en
i
m
ages
w
i
t
h
di
f
f
er
ent
pos
es
.
T
he
or
i
g
i
nal
s
i
z
e
of
t
he
U
o
B
f
ac
i
al
i
m
age
dat
abas
e
i
s
512 p
i
x
el
s
f
or
hei
g
ht
and
342 pi
x
e
l
s
f
or
w
i
d
t
h [
2
2]
.
I
n t
hi
s
r
es
ear
c
h,
a
l
l
of
i
m
ages
w
er
e r
es
i
z
e
d
i
nt
o 1
40 p
i
x
el
s
f
or
hei
ght
a
nd 120 p
i
x
el
s
f
or
w
i
dt
h.
T
he i
m
age s
a
m
pl
e of
t
he U
oB
c
an be s
een i
n
F
i
gur
e 1
.
I
t
has
t
h
e di
f
f
er
ent
pos
es
f
or
eac
h c
l
as
s
,
but
i
t
h
as
t
he s
am
e ex
pr
es
s
i
ons
,
w
h
i
c
h i
s
nor
m
al
ex
pr
es
s
i
on.
F
our
s
c
enar
i
os
ha
v
e
bee
n
per
f
or
m
ed t
o ev
a
l
u
at
e t
h
e
pr
opos
e
d m
et
hod,
t
hes
e ar
e us
i
ng
t
w
o,
t
hr
ee,
f
our
,
a
nd f
i
v
e f
ac
i
al
i
m
ages
,
t
he
y
ha
v
e b
een c
h
os
en r
a
ndom
l
y
.
T
h
e ex
per
i
m
ent
a
l
r
es
ul
t
s
ar
e de
pen
di
n
g on t
h
e t
r
ai
n
i
ng s
et
s
us
e
d,
f
or
bot
h t
he n
um
ber
of
t
r
ai
ni
n
g s
e
t
s
and pos
es
.
I
n
t
h
i
s
c
as
e,
t
he
num
ber
of
f
eat
ur
es
us
ed
i
s
t
w
ent
y
-
n
i
ne.
A
s
k
no
w
n,
t
he
pr
opos
e
d
m
et
hod
has
been
pr
od
uc
ed C
-
1 c
l
as
s
es
,
C
r
epr
es
ent
s
n
um
ber
of
c
l
as
s
es
of
t
he t
r
ai
ni
ng s
et
s
,
w
hi
c
h
ar
e
t
hi
r
t
y
c
l
as
s
es
.
F
or
eac
h s
c
e
nar
i
o,
t
w
e
nt
y
ex
p
er
i
m
ent
s
hav
e
be
en c
o
nduc
t
e
d,
w
hi
c
h
ar
e us
i
ng
t
e
n
unt
i
l
t
w
en
t
y
-
n
i
n
e f
eat
ur
es
as
t
he
s
i
m
i
l
ar
i
t
y
m
eas
ur
e
m
ent
s
.
T
he m
ax
i
m
u
m
r
ec
ogn
i
t
i
on r
a
t
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
13
–
1
122
1118
pr
oduc
e
d
b
y
t
h
e
pr
o
pos
ed
m
et
hod
us
i
ng
t
w
o,
t
hr
ee
,
f
our
an
d
f
i
v
e
t
he
t
r
ai
ni
ng
s
e
t
s
ar
e
75.
42%
,
87.
6
2%
,
96.
11%
,
a
nd 98
.
6
7%
,
r
es
pec
t
i
v
e
l
y
.
T
he hi
gh
es
t
r
ec
ogni
t
i
o
n r
at
e has
o
c
c
ur
r
ed w
he
n
us
i
ng
f
i
v
e t
r
a
i
n
i
ng s
et
s
.
T
he
ex
per
i
m
ent
al
r
es
u
l
t
s
of
t
he
pr
opos
e
d m
et
hod dem
ons
t
r
at
ed t
h
at
t
h
e
hi
g
hes
t
r
ec
og
ni
t
i
o
n
r
at
e
ha
s
oc
c
ur
r
ed
w
hen
us
i
ng
f
i
v
e
t
r
ai
n
i
ng
s
et
s
,
w
h
er
eas
t
h
e
s
i
m
i
l
ar
i
t
y
us
ed
i
s
t
he E
uc
l
i
d
i
a
n D
i
s
t
anc
e
.
T
he l
ow
es
t
r
ec
ogni
t
i
o
n r
at
e oc
c
ur
r
ed w
h
en
us
i
ng t
w
o t
r
ai
n
i
ng s
et
s
.
F
i
gur
e
2a,
2b,
2c
,
and
2d
d
epi
c
t
t
he ex
p
er
i
m
ent
al
r
es
u
l
t
s
t
w
o t
r
ai
ni
ng s
et
s
us
i
ng
t
w
o,
t
hr
e
e,
f
our
,
and
f
i
v
e
t
r
a
i
ni
ng
s
et
s
(
s
ee
F
i
gur
e
2)
.
T
he
us
age
of
t
he
f
eat
ur
es
i
nf
l
u
enc
ed
t
he
r
ec
ogn
i
t
i
on
r
at
e
r
es
ul
t
s
.
T
he m
ax
i
m
u
m
f
eat
ur
es
us
ed
ar
e num
ber
of
c
l
as
s
es
m
i
nus
one
(
30
–
1 =
29)
f
eat
ur
es
.
F
i
gur
e
1.
T
he U
n
i
v
er
s
i
t
y
of
B
er
n F
ac
i
a
l
I
m
age [
22]
F
i
gur
e
2.
E
x
per
i
m
ent
al
R
es
ul
t
s
U
s
i
ng t
he
P
r
opos
e
d M
et
ho
d on
t
he
U
n
i
v
er
s
i
t
y
of
B
er
n F
ac
i
a
l
I
m
age D
at
ab
as
e
F
i
gur
e 3.
I
m
age T
es
t
s
w
er
e
er
r
or
r
ec
ogn
i
z
ed [
22]
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
Mul
t
i
-
C
r
i
t
er
i
a
i
n
D
i
s
c
r
i
m
i
na
n
t
A
n
al
y
s
i
s
t
o F
i
n
d t
h
e D
o
mi
nant
F
eat
ur
es
(
A
r
i
f
Mun
t
as
a
)
1119
T
he s
m
al
l
es
t
ac
c
ept
anc
e r
at
e
w
as
pr
o
duc
ed
b
y
us
i
n
g t
w
o t
r
a
i
n
i
ng s
et
s
.
T
he s
i
m
i
l
ar
f
eat
ur
es
of
t
he
di
f
f
er
ent
c
l
a
s
s
es
i
nduc
e
d t
he r
ec
o
gn
i
z
i
ng er
r
or
s
.
T
he s
i
m
i
l
ar
f
eat
u
r
es
ar
e c
aus
e
d
b
y
t
he s
i
m
i
l
ar
i
m
age of
t
he t
es
t
i
ng
and
t
he
t
r
ai
ni
n
g
s
et
f
ound as
s
e
en
i
n F
i
gu
r
e 3,
t
h
e f
i
r
s
t
c
ol
um
n
i
s
t
he
t
es
t
i
ng
s
et
,
w
her
eas
t
he
s
ec
on
d
c
ol
um
n
i
s
t
he
i
m
age
m
at
c
hi
ng
f
ound.
I
n
F
i
g
ur
e
3
s
ho
w
s
t
hat
t
he i
m
age m
at
c
hi
n
g f
ound ar
e f
al
s
e.
H
o
w
e
v
er
,
t
h
e s
i
m
i
l
ar
i
t
y
bet
w
ee
n i
m
ages
t
o eac
h
ot
her
w
i
l
l
a
l
s
o pr
o
duc
e
t
he
s
i
m
i
l
ar
f
eat
ur
es
.
3.
2.
E
v
a
l
u
a
ti
o
n
th
e
P
r
o
p
o
s
e
d
M
e
th
o
d
o
n
th
e
Y
A
L
E
F
a
c
i
a
l
I
m
a
g
e
D
a
ta
b
a
s
e
T
he s
ec
ond ev
a
l
uat
i
o
n i
s
ex
per
i
m
ent
al
us
i
n
g t
he
Y
A
LE
f
ac
i
al
i
m
age dat
ab
as
e.
T
o
ev
a
l
u
at
e
t
he
pr
o
pos
ed
m
et
hod,
s
c
enar
i
o
us
ed
i
s
t
he
s
a
m
e
as
t
he
f
i
r
s
t
a
s
s
es
s
m
ent
.
T
he
Y
A
L
E
f
ac
i
al
i
m
age dat
ab
as
e has
t
he s
m
al
l
er
c
l
as
s
es
t
han t
he
U
ni
v
er
s
i
t
y
of
B
er
n f
ac
i
al
i
m
age dat
ab
as
e
[
23]
.
T
he
Y
A
L
E
f
ac
i
a
l
i
m
age d
at
a
bas
e
has
f
i
f
t
een
c
l
as
s
es
,
f
or
eac
h c
l
as
s
has
el
ev
en
i
m
age
pos
es
.
T
he Y
A
L
E
i
s
f
ac
i
al
i
m
age dat
a
bas
e t
h
at
has
t
he d
i
f
f
er
ent
pos
es
,
l
i
gh
t
i
n
gs
and
ex
pr
es
s
i
ons
as
s
ee
n
i
n
F
i
g
ur
e
4.
I
n
t
h
i
s
r
es
ear
c
h,
t
he
s
a
m
e
s
c
enar
i
o
w
as
us
ed
t
o
e
v
al
uat
e
t
h
e
pr
opos
e
d
m
et
hod,
w
hi
c
h
i
s
us
i
ng
f
our
s
c
enar
i
os
.
F
or
eac
h
s
c
enar
i
o,
t
he
pr
op
os
e
d
m
et
hod
w
as
t
es
t
ed us
i
n
g t
w
o unt
i
l
f
our
t
een f
eat
ur
es
(
t
h
e
num
ber
of
c
l
as
s
es
m
i
nus
one)
.
T
he
ex
p
er
i
m
ent
al
r
es
ul
t
s
us
i
n
g t
h
e pr
op
os
e
d m
et
hod has
pr
o
duc
ed r
ec
ogn
i
t
i
on r
at
e 81
.
48%
(
F
i
gur
e 5
a)
,
90%
(
F
i
gur
e
5b)
,
90
.
47%
(
F
i
g
ur
e
5c
)
,
and 91.
1
1%
(
F
i
gur
e
5d)
f
or
t
w
o,
t
hr
e
e,
f
our
,
and
f
i
v
e
t
r
ai
ni
ng
se
t
s r
e
sp
e
ct
i
ve
l
y.
F
i
gur
e
4.
T
he
Y
A
L
E
F
ac
i
al
I
m
age [
23]
F
i
gur
e
5.
E
x
per
i
m
ent
al
R
es
ul
t
s
U
s
i
ng t
he
P
r
opos
e
d M
et
ho
d on
t
he
Y
A
L
E
F
ac
i
al
I
m
age
D
at
ab
as
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
13
–
1
122
1120
T
he r
ec
ogni
t
i
on r
at
e h
as
i
nc
r
eas
ed pr
op
or
t
i
o
na
l
t
o t
he t
r
ai
n
i
n
g s
et
s
us
ed,
t
ho
ugh t
h
e
r
ec
ogni
t
i
o
n r
at
e
i
nc
r
e
as
i
n
g
i
s
not
s
i
gn
i
f
i
c
ant
.
R
ec
og
ni
t
i
on
er
r
or
s
w
er
e c
aus
e
d
b
y
t
he
i
m
age
v
ar
i
ant
s
s
uc
h
as
ex
pr
es
s
i
o
ns
,
bu
t
t
he
l
i
ght
i
ng
ef
f
ec
t
c
an
be
s
t
i
l
l
r
ec
o
gn
i
z
ed.
T
he ex
per
i
m
ent
a
l
r
es
ul
t
s
i
n de
t
ai
l
c
an
be
s
ee
n i
n
F
i
g
ur
e 5.
I
n F
i
gur
e 5 c
an a
l
s
o b
e s
ho
w
n t
h
at
t
he s
m
al
l
er
f
eat
ur
es
us
ed ha
v
e pr
od
uc
ed,
t
h
e l
o
w
er
r
ec
ogn
i
t
i
on r
at
e,
t
h
e m
or
e
f
eat
ur
es
us
ed has
al
s
o c
ont
r
i
but
ed
t
he l
ar
ger
ac
c
ept
anc
e
r
at
e as
s
een i
n
F
i
gur
e 5.
3.
3
.
E
v
a
l
u
a
ti
o
n
th
e
P
r
o
p
o
s
e
d
M
e
th
o
d
o
n
th
e
O
R
L
F
a
c
i
a
l
I
m
a
g
e
D
a
ta
b
a
s
e
T
he l
as
t
ev
a
l
uat
i
o
n i
s
ex
e
c
ut
ed b
y
us
i
n
g t
he O
R
L f
ac
i
al
i
m
age dat
ab
as
e.
T
he O
R
L
f
ac
i
al
i
m
age
dat
a
bas
e
h
as
f
our
hundr
ed
i
m
ages
.
T
he
y
ar
e
gai
ned
f
r
om
f
or
t
y
per
s
ons
,
f
our
eac
h
per
s
on
h
as
t
en
d
i
f
f
er
ent
po
s
es
.
T
he
v
ar
i
ant
of
pos
es
c
ons
i
s
t
s
of
ex
pr
es
s
i
ons
,
ac
c
es
s
or
i
es
,
an
d
pos
i
t
i
on of
t
he f
ac
i
al
p
os
e
.
T
he ex
pr
es
s
i
ons
of
t
he O
R
L f
ac
e i
m
age dat
a
bas
e ar
e s
m
i
l
i
n
g,
neut
r
a
l
,
ope
n
and
c
l
os
e
e
y
es
.
T
he ac
c
es
s
or
i
es
of
t
h
e
O
R
L i
s
us
i
n
g g
l
as
s
es
or
no
t
,
w
h
er
eas
t
he
pos
i
t
o
n
of
t
he
f
ac
i
al
pos
e
i
s
des
c
r
i
bed
b
y
pos
e
o
w
ne
d f
or
eac
h per
s
on
s
uc
h as
r
i
ght
,
l
ef
t
,
up
and
do
w
n
[
2
4]
.
F
i
gur
e
6
d
em
ons
t
r
at
ed
t
en
per
s
ons
o
f
t
he
O
R
L
f
ac
e
i
m
age
dat
a
bas
e
w
i
t
h
t
he
di
f
f
er
ent
pos
es
,
ex
pr
es
s
i
on
s
and ac
c
es
s
or
i
es
.
I
n t
h
i
s
r
es
e
ar
c
h,
t
he
dom
i
nant
f
eat
ur
es
of
t
he pr
op
os
ed m
et
hod r
es
ul
t
s
ha
v
e
been
r
educ
ed
t
o n
um
ber
c
l
as
s
m
i
nus
o
ne (
4
0
-
1 =
39)
.
T
he pr
opos
e
d
m
et
hod w
as
ev
al
uat
e
d
b
y
us
i
n
g
t
w
o,
t
hr
ee,
f
our
an
d f
i
v
e t
r
ai
ni
ng
s
et
s
as
dem
ons
t
r
at
ed
i
n F
i
g
ur
e 7
,
i
.
e.
7a,
7
b,
7c
,
and
7
d
r
es
pec
t
i
v
el
y
.
F
i
gur
e
6.
T
he O
R
L F
ac
i
a
l
I
m
age [
24]
F
i
gur
e
7.
E
x
per
i
m
ent
al
R
es
ul
t
s
U
s
i
ng t
he
P
r
opos
e
d M
et
ho
d on
t
he
O
R
L F
ac
i
al
I
m
age D
at
a
bas
e
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
IS
S
N
:
1
693
-
6
930
Mul
t
i
-
C
r
i
t
er
i
a
i
n
D
i
s
c
r
i
m
i
na
n
t
A
n
al
y
s
i
s
t
o F
i
n
d t
h
e D
o
mi
nant
F
eat
ur
es
(
A
r
i
f
Mun
t
as
a
)
1121
T
w
ent
y
-
one
unt
i
l
t
h
i
r
t
y
-
n
i
ne
f
eat
ur
es
has
b
een
ut
i
l
i
z
e
d
f
or
s
i
m
i
l
ar
i
t
y
m
eas
ur
em
ent
s
.
T
he
ev
a
l
u
at
i
on
r
es
u
l
t
s
ha
v
e
pr
o
duc
ed
t
he
r
ec
o
gni
t
i
o
n
8
5.
6
2%
,
91
.
07%
,
97.
08%
,
and
97.
5%
f
or
t
w
o
,
t
hr
ee,
f
our
an
d f
i
v
e t
r
a
i
n
i
ng
s
et
s
r
es
pec
t
i
v
el
y
.
T
he us
ag
e of
f
eat
ur
es
has
af
f
ec
t
ed t
he r
ec
ogn
i
t
i
o
n
r
at
e o
bt
a
i
ne
d.
T
he
hi
gher
r
ec
ogni
t
i
o
n r
at
e c
a
n b
e
on
l
y
ac
h
i
e
v
ed
b
y
us
i
ng
t
h
e
m
or
e d
o
m
i
nant
f
eat
ur
es
.
T
he
m
or
e
dom
i
nant
f
eat
ur
es
r
educ
e
d,
t
h
e
l
o
w
es
t
r
ec
og
ni
t
i
on
r
at
e
obt
a
i
ne
d.
T
he
num
ber
of
t
r
ai
n
i
ng
s
et
s
has
a
l
s
o
i
nf
l
ue
nc
ed
t
he
r
ec
ogni
t
i
o
n r
at
e o
bt
a
i
n
ed.
T
he hi
ghes
t
r
ec
ogni
t
i
o
n
r
at
e
w
as
obt
ai
n
ed
w
h
en
ex
p
er
i
m
ent
al
r
es
ul
t
s
us
i
ng
f
i
v
e
t
r
ai
ni
n
g
s
et
s
and
t
h
i
r
t
y
-
f
i
ve
unt
i
l
t
hi
r
t
y
-
ni
ne
d
om
i
nant
f
eat
ur
es
,
w
hi
c
h
i
s
9
7.
5%
.
I
t
m
eans
,
onl
y
f
i
v
e
of
t
w
o
h
undr
e
d i
m
ages
w
er
e f
al
s
e r
ec
og
ni
t
i
o
n.
3.
4
.
D
i
s
c
u
s
s
i
o
n
a
n
d
C
o
m
p
a
r
i
s
o
n
to
O
th
e
r
M
e
th
o
d
s
T
he pr
opos
ed m
et
hod h
as
pr
ov
e
d t
h
at
m
ul
t
i
-
c
r
it
e
r
ia
o
f
D
is
c
r
i
m
in
a
nt
A
na
l
y
s
i
s
c
an b
e
i
m
pl
em
ent
ed
t
o r
ec
ogn
i
z
e
t
he f
ac
i
al
i
m
age.
T
he
obj
e
c
t
i
v
e
f
unc
t
i
on of
t
he
pr
op
o
s
ed
m
et
hod as
w
r
i
t
t
en
i
n
E
qu
at
i
on (
11)
,
(
1
2)
,
(
13)
and (
14)
c
an b
e de
r
i
v
e
d and s
o
l
v
ed b
y
us
i
ng g
ener
a
l
i
z
e
d t
h
e
E
i
g
en
v
a
l
ue pr
o
bl
em
.
F
our
c
r
i
t
er
i
a of
t
he pr
opos
e
d
m
et
hod m
odel
ed ha
v
e
pr
oduc
e
d t
he
dom
i
nant
f
eat
ur
es
t
h
at
c
a
n
be
i
m
pl
em
ent
ed
t
o
r
ec
ogn
i
z
e
t
he
f
ac
i
a
l
i
m
age.
T
he
pr
opos
ed
m
et
hod w
as
as
s
es
s
ed b
y
u
s
i
ng t
hr
e
e f
ac
i
al
i
m
age da
t
a
bas
es
,
w
hi
c
h
ar
e t
he U
oB
t
he
Y
A
L
E
,
an
d
T
he
O
R
L
f
ac
e
i
m
age
dat
a
bas
es
.
T
he
r
ec
ogni
t
i
o
n
r
at
e
has
di
s
p
l
a
y
e
d
t
ha
t
t
he
pr
opos
ed
m
et
ho
d
has
pr
o
duc
ed
t
h
e
l
ar
ger
a
c
c
ept
anc
e r
a
t
e
on
t
he
O
R
L f
ac
e i
m
age
dat
a
bas
e
t
h
an o
t
her
f
ac
i
al
i
m
age
dat
ab
as
es
,
w
hi
c
h
ar
e
t
he
U
n
i
v
er
s
i
t
y
of
B
er
n
a
n
d
t
he
Y
A
L
E
f
ac
e
i
m
age
dat
abas
es
.
T
he
pr
opos
e
d
m
et
hod
has
pr
od
uc
ed
t
h
e
l
o
w
es
t
r
ec
ogn
i
t
i
o
n
r
at
e
t
ha
n
t
h
e
U
ni
v
er
s
i
t
y
o
f
B
er
n
and
t
h
e
O
R
L f
ac
e i
m
age dat
a
bas
e
s
.
T
he r
ec
ogni
t
i
on
er
r
or
s
on t
h
e
Y
A
L
E
f
ac
e i
m
age
dat
a
bas
e
w
er
e
c
aus
ed
b
y
l
i
gh
t
i
ng
ef
f
ec
t
.
I
t
m
eans
t
he
pr
opos
ed
m
et
hod
c
annot
n
or
m
al
i
z
e
l
i
gh
t
i
n
g
ef
f
ec
t
.
T
he
ac
c
ept
anc
e r
a
t
e r
es
u
l
t
s
of
t
he
pr
opos
e
d m
et
hod
w
e
r
e al
s
o c
om
par
ed t
o t
h
e o
t
her
m
et
hods
,
w
hi
c
h ar
e t
he
P
r
i
nc
i
pa
l
C
om
ponent
A
n
al
y
s
i
s
o
r
w
e
l
l
-
k
now
n E
i
ge
nf
ac
es
,
F
i
s
her
f
ac
e,
Lap
l
ac
i
a
nf
ac
es
,
and
O
r
t
ho
gona
l
Lap
l
ac
i
anf
ac
es
(
O
-
Lapl
ac
i
anf
ac
es
)
as
s
ho
w
n
i
n T
abl
e
2,
3,
and 4.
T
abl
e 2.
C
om
par
i
s
on R
es
u
l
t
s
on t
h
e U
n
i
v
er
s
i
t
y
of
B
er
n
F
ac
i
al
I
m
age D
at
a
bas
e
T
r
ai
ni
ng
S
et
U
s
ed
R
ec
ogni
t
i
on R
a
t
e
(%
)
E
i
genf
a
c
e
F
i
s
her
f
a
c
e
Lapl
ac
i
anf
a
c
es
O
-
Lapl
ac
i
anf
a
c
es
T
he
P
r
opos
ed
M
et
hod
2
66.
25
66.
67
67.
5
63.
33
75.
42
3
71.
43
71.
43
71.
90
66.
19
87.
62
4
83.
33
81.
67
86.
67
77.
22
96.
11
5
92.
00
91.
33
90.
00
89.
33
98.
67
T
abl
e 3.
C
om
par
i
s
on R
es
u
l
t
s
on t
h
e
Y
A
L
E
F
ac
i
al
I
m
age D
at
abas
e
T
r
ai
ni
ng
S
et
U
s
ed
R
ec
ogni
t
i
on R
a
t
e
(
%
)
E
i
genf
a
c
e
F
i
s
her
f
a
c
e
Lapl
ac
i
anf
a
c
es
O
-
Lapl
ac
i
anf
a
c
es
T
he P
r
opos
ed
M
et
hod
2
43.
5
45.
7
56.
5
55.
7
80.
74
3
48.
9
64.
5
68.
5
70.
1
90.
00
4
52.
2
72.
7
74.
6
77.
3
90.
47
5
57.
8
77.
5
78.
3
82.
1
91.
11
T
abl
e 4.
C
om
par
i
s
on R
es
u
l
t
s
on t
h
e O
R
L F
ac
i
a
l
I
m
age D
at
a
bas
e
T
r
ai
ni
ng
S
et
Us
e
d
R
ec
ogni
t
i
on R
a
t
e
(
%
)
E
i
genf
a
c
e
F
i
s
her
f
a
c
e
Lapl
ac
i
anf
a
c
es
O
-
Lapl
ac
i
anf
a
c
es
T
he P
r
opos
ed
M
et
hod
2
66.
3
71.
1
76.
1
79.
6
85.
62
3
75.
4
84.
2
86.
6
88.
6
91.
07
4
82
89.
5
90.
42
94.
08
97.
08
5
85.
9
92.
25
93.
15
96.
35
97.
50
T
he pr
opos
ed m
et
hod
has
pr
ov
ed t
hat
t
he
ex
per
i
m
ent
al
r
es
ul
t
s
out
p
er
f
or
m
ed
t
o ot
h
er
m
et
hods
f
or
al
l
f
ac
i
al
i
m
age dat
a
bas
es
,
th
e
U
ni
v
er
s
i
t
y
of
B
er
n,
t
h
e
Y
A
L
E
,
an
d
t
he O
R
L f
ac
i
al
i
m
age dat
a
bas
es
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
13
–
1
122
1122
4
.
C
o
n
c
l
u
s
i
o
n
T
hr
ee dat
abas
es
w
er
e e
m
pl
o
y
ed t
o e
v
a
l
u
at
e
t
h
e
pr
opos
e
d m
et
hod.
T
he ev
al
u
at
i
on
r
es
ul
t
s
s
ho
w
ed
t
hat
t
he
pr
opos
ed
m
et
hod
c
an
r
ec
og
ni
z
e
t
he
f
ac
i
a
l
i
m
age.
T
he
r
ec
ogni
t
i
o
n
r
at
e
r
es
ul
t
s
ar
e l
ar
ger
t
h
an o
t
her
m
et
hods
s
uc
h as
E
i
g
enf
ac
e,
F
i
s
her
f
ac
e,
La
pl
ac
i
anf
ac
es
,
an
d
O
r
t
hogon
al
L
ap
l
ac
i
anf
ac
es
.
T
he
hi
ghes
t
ac
c
ept
anc
e
r
at
es
w
er
e
ac
h
i
e
v
e
d
w
h
en
us
i
n
g
f
i
v
e
t
r
ai
n
i
ng s
et
s
f
or
t
he U
ni
v
er
s
i
t
y
of
B
er
n,
t
h
e
Y
A
LE
,
a
n
d t
he O
R
L f
ac
i
al
i
m
age da
t
abas
es
.
T
he
r
ec
ogni
t
i
o
n
r
es
ul
t
h
as
pr
o
duc
ed
t
h
e
l
o
w
es
t
ac
c
ept
anc
e
r
at
e
w
h
en
us
i
ng
t
h
e
Y
A
LE
f
ac
e.
R
ec
ogn
i
t
i
on
er
r
or
s
on t
h
e
Y
A
L
E
f
ac
i
al
i
m
age dat
abas
e
w
er
e c
aus
e
d b
y
l
i
ght
i
n
g ef
f
ec
t
.
F
ur
t
her
m
or
e,
t
he
l
i
m
i
t
at
i
on
of
t
he
pr
op
os
ed m
et
hod
c
an be
r
ec
o
v
er
ed
b
y
us
i
n
g t
h
e R
e
t
i
n
ex
m
et
hods
t
o r
em
ov
e t
he
l
i
gh
t
i
ng
ef
f
ec
t
.
R
ef
er
en
ces
[1
]
M
unt
as
a
A
.
N
ew
m
odel
l
i
ng
of
m
odi
f
i
ed t
w
o di
m
en
s
i
o
n
al
f
i
s
h
er
f
a
c
e ba
s
ed f
e
at
ur
e
e
x
t
r
ac
t
i
on
.
T
EL
KO
M
N
I
KA
T
el
ec
om
m
uni
c
at
i
on
C
om
put
i
ng E
l
ec
t
r
oni
c
s
a
nd
C
ont
r
ol
.
2014
;
12(
1
):
11
5
-
12
2
.
[2
]
AM
M
ar
t
i
nez
,
AC
Ka
k
.
PC
A v
e
rs
u
s
L
D
A.
I
E
E
E
T
r
an
s
.
P
at
t
er
n A
nal
.
M
ac
h.
I
nt
el
l
.
20
01;
23(
2)
:
22
8
-
233.
[3
]
P
N
B
el
hum
eur
,
J
P
H
es
panh
a
,
DJ
K
r
i
eg
m
an
.
E
i
g
enf
a
c
e
s
v
s
.
F
i
s
her
f
ac
es
:
R
e
c
og
ni
t
i
on
u
s
i
ng
c
l
as
s
s
pe
c
i
fi
c
l
i
ne
ar
pr
o
j
ec
t
i
o
n.
I
E
E
E
T
r
ans
.
P
at
t
er
n A
n
al
.
M
ac
h.
I
nt
el
l
.
1997
;
19(
7)
:
71
1
-
72
0.
[4
]
C
hen
LF
,
Li
a
o
H
Y
M
,
K
o
M
T
,
Li
n
J
C
,
Y
u
G
J
.
A
ne
w
LD
A
bas
ed
f
a
c
e
r
e
c
ogn
i
t
i
on
s
y
s
t
e
m
w
hi
c
h
c
an
s
ol
v
e
t
he
s
m
a
l
l
s
am
pl
e s
i
z
e pr
obl
e
m
.
P
at
t
er
n R
e
c
o
g
ni
t
i
on
.
2
000;
3
3:
1
713
-
1726
.
[5
]
Ki
m
T
,
K
i
t
t
l
er
J
.
Lo
c
a
l
l
y
l
i
nea
r
di
s
c
r
i
m
i
n
ant
anal
y
s
i
s
f
or
m
ul
t
i
-
m
od
al
l
y
di
s
t
r
i
but
ed
c
l
a
s
s
e
s
f
or
f
ac
e
r
ec
og
ni
t
i
on w
i
t
h a
s
i
ngl
e
m
o
de
l
i
m
age.
I
E
E
E
T
r
a
ns
.
P
at
t
er
n A
nal
.
M
ac
h.
I
nt
el
l
.
2
005;
27
(
3)
:
318
-
32
7.
[6
]
M
unt
as
a A
.
T
he H
um
an F
a
c
i
al
E
x
pr
es
s
i
o
n C
l
as
s
i
f
i
c
at
i
o
n U
s
i
ng t
he C
en
t
er
K
er
n
el
S
ub
s
p
ac
e b
as
e
d
t
he R
i
d
ge R
eg
r
es
s
i
o
n.
J
our
n
al
of
C
om
put
er
S
c
i
enc
e
.
2
015;
1
1(
11)
:
1054
-
105
9.
[7
]
I
G
ede
P
as
ek
S
u
t
a
W
i
j
ay
a,
K
e
i
i
c
h
i
U
c
h
i
m
ur
a,
G
ou
K
o
ut
a
k
i
.
F
ac
e
R
e
c
og
ni
t
i
on
U
s
i
ng
H
o
l
i
s
t
i
c
F
eat
ur
e
s
and
L
i
ne
ar
D
i
s
c
r
i
m
i
nant
A
n
al
y
s
i
s
S
i
m
pl
i
f
i
c
at
i
on.
T
EL
KO
M
N
I
KA
T
el
ec
om
m
uni
c
a
t
i
on
C
om
put
i
ng
El
e
c
t
r
oni
c
s
and
C
o
nt
r
o
l
.
20
12;
10(
4)
:
775
-
787.
[8
]
M
unt
as
a
A
.
A
new
appr
o
ac
h:
T
he l
oc
a
l
f
eat
ur
e ex
t
r
ac
t
i
o
n b
as
ed
on
t
he
new
r
egu
l
at
i
on o
f
t
he
l
oc
al
l
y
pr
es
er
v
i
ng pr
oj
e
c
t
i
o
n.
A
p
pl
i
ed M
at
h.
S
c
i
.
20
15;
9(
102)
:
50
65
-
5078.
[9
]
Q
i
an T
i
an.
F
ac
e R
e
c
og
ni
t
i
on
U
s
i
ng
I
nv
ar
i
a
nc
e w
i
t
h a
S
i
ng
l
e T
r
ai
ni
n
g S
a
m
pl
e.
T
E
L
KO
M
N
I
KA
.
201
4;
12(
4)
:
921
-
93
2
.
[
10]
M
unt
as
a
A
.
F
ac
i
a
l
r
ec
o
gni
t
i
on us
i
ng s
q
uar
e di
ago
nal
m
at
r
i
x
bas
ed on t
w
o
-
di
m
ens
i
ona
l
l
i
ne
ar
di
s
c
r
i
m
i
na
nt
a
nal
y
s
i
s
.
In
t. R
e
v
.
C
o
m
p
u
t. S
o
ftw
a
r
e
(
I.R
E
.C
O
.S
.)
.
201
5;
1
0(
7)
:
7
18
-
7
25.
[
11]
D
ai
D
Q
,
Y
uen
P
C
.
F
ac
e
r
ec
o
gni
t
i
on
by
r
egul
ar
i
z
ed
di
s
c
r
i
m
i
nant
an
al
y
s
i
s
.
I
EEE
T
ra
n
s
.
Sy
s
t
.
,
M
a
n
,
C
y
ber
n.
B
,
C
y
b
er
n
.
2007
;
37(
4
)
:
1080
-
108
5.
[
12]
M
unt
as
a
A
,
M
oc
ham
m
ad
K
a
u
t
s
ar
S
ho
pan,
M
aur
i
dhi
H
er
y
P
ur
nom
o
,
K
ond
o
K
uni
o
.
E
nha
n
c
em
ent
of
t
he A
dap
t
i
v
e S
hap
e V
ar
i
a
nt
s
A
v
er
age V
al
u
es
by
U
s
i
ng E
i
g
h
t
M
o
v
em
ent
D
i
r
e
c
t
i
o
ns
f
o
r
M
ul
t
i
-
F
eat
ur
e
s
D
et
ec
t
i
on
of
F
a
c
i
a
l
S
k
et
c
h.
IT
B
J
. IC
T
.
201
2;
6
c
(
1)
:
1
-
20.
[
13]
S
angee
t
a N
Ka
k
a
rw
a
l
,
R
at
n
adeep
R
D
es
hm
u
k
h.
H
y
br
i
d
F
eat
ur
e E
x
t
r
ac
t
i
on T
ec
hni
qu
e f
or
Fa
c
e
R
ec
ogn
i
t
i
o.
I
n
t
er
na
t
i
on
al
J
our
n
al
of
A
dv
a
nc
e
d C
om
put
er
S
c
i
e
nc
e
and A
ppl
i
c
at
i
on
s
.
2
012;
3(
2)
:
60
-
62.
[
14]
P
F
Hs
i
e
h
,
DS
W
ang
,
CW
H
s
u.
A
l
i
near
f
ea
t
ur
e
ex
t
r
ac
t
i
o
n
f
or
m
ul
t
i
c
l
as
s
c
l
as
s
i
f
i
c
at
i
on
pr
obl
e
m
s
ba
s
ed
on
c
l
as
s
m
e
an
an
d
c
ov
ar
i
anc
e
di
s
c
r
i
m
i
n
ant
i
nf
or
m
at
i
on.
IE
E
E
Tr
a
n
s
.
P
a
tt.
A
n
a
l
.
M
a
c
h
.
In
te
l
l
.
2
0
06;
28(
2)
:
223
-
2
35.
[
15]
H
Yu
, J
Y
ang.
A
di
r
ec
t
LD
A
al
gor
i
t
h
m
f
or
hi
gh
-
d
i
m
e
ns
i
ona
l
dat
a w
i
t
h appl
i
c
at
i
on t
o f
ac
e r
ec
og
ni
t
i
on
.
P
at
t
er
n R
e
c
o
gni
t
i
on
.
200
1;
3
4:
2067
-
2070
.
[
16]
J
i
ep
i
ng Y
e,
Q
i
Li
.
A
T
w
o
-
S
t
age Li
near
D
i
s
c
r
i
m
i
nan
t
A
nal
y
s
i
s
v
i
a Q
R
-
D
ec
om
po
s
i
t
i
on.
I
E
EE
T
r
ans
ac
t
i
on
s
on
P
at
t
er
n A
na
l
y
s
i
s
and
M
ac
hi
ne I
nt
el
l
i
g
enc
e
.
2005;
27(
6)
:
92
9
-
941
.
[
17]
H
ak
an C
ev
i
k
al
p
,
M
ar
i
an N
ea
m
t
u,
M
i
t
c
h
W
i
l
k
e
s
,
A
t
a
l
ay
B
ar
k
ana
.
D
i
s
c
r
i
m
i
na
t
i
v
e C
om
m
on
V
ec
t
or
s
f
or
F
ac
e R
ec
o
gni
t
i
on.
I
EEE T
ra
n
s
.
o
n
P
at
t
er
n
A
nal
y
s
i
s
and
M
ac
hi
ne I
nt
e
l
l
i
gen
c
e
.
2
005
;
27(
1
):
4
-
13.
[
18]
DQ
Da
i
,
PC
Y
uen.
R
e
gul
ar
i
z
ed
D
i
s
c
r
i
m
i
na
nt
A
n
al
y
s
i
s
a
nd
I
t
s
A
p
pl
i
c
at
i
on
t
o
F
ac
e
R
ec
ogn
i
t
i
on.
P
at
t
er
n R
e
c
o
gni
t
i
on
.
200
3;
3
6(
3)
:
845
-
847
.
[
19]
J
Lu,
K
N
P
l
at
a
ni
ot
i
s
,
AN
V
en
et
s
a
nopo
ul
o
s
.
F
a
c
e R
ec
ogn
i
t
i
on
Us
i
n
g
L
DA
-
B
as
ed A
l
gor
i
t
h
m
s
.
I
E
EE
T
r
ans
.
N
eur
a
l
N
et
w
or
k
s
.
200
3;
14:
1
95
-
2
00.
[
20]
G
B
audat
, F
A
nouar
.
G
en
er
al
i
z
ed D
i
s
c
r
i
m
i
nant
A
na
l
y
s
i
s
U
s
i
n
g a K
er
nel
A
ppr
o
ac
h.
N
eur
al
C
om
put
at
i
o
n
.
2
000;
12(
1
0)
:
2
3
85
-
240
4.
[
21]
R
H
uang,
Q
Li
u
,
H
Lu
,
S
Ma
.
S
ol
v
i
ng t
he S
m
al
l
S
i
z
e P
r
obl
e
m
of
LD
A
.
P
r
o
c
.
1
6
th
In
t
’
l
C
o
n
f. P
a
tt
e
r
n
R
ec
ogn
i
t
i
on
.
2
002;
3:
29
-
3
2.
[
22]
T
he U
ni
v
er
s
i
t
y
of
B
er
n,
ht
t
p:
/
/
w
w
w
.
i
am
.
uni
be
.
c
h
/
f
k
i
/
dat
aba
s
es
/
i
am
-
fa
c
e
s
-
d
at
ab
as
e
.
[
23]
T
he
Y
A
LE
C
ent
er
f
o
r
C
om
put
at
i
o
nal
V
i
s
i
on and
C
ont
r
ol
.
Y
al
e
F
a
c
e
D
at
aba
s
e
,
h
ttp
:
/
/
c
v
c
.
y
al
e
.
edu
/
pr
o
j
ec
t
s
/
y
al
ef
ac
es
/
y
al
ef
a
c
es
.
ht
m
l
.
[
24]
T
he
O
R
L,
R
es
ear
c
h
C
ent
er
o
f
A
t
t
,
U
K
.
O
l
i
v
e
tti
-
A
tt
O
R
L
F
a
c
e
D
at
aba
s
e,
h
ttp
://w
w
w
.
uk
.
r
es
ear
c
h.
at
t
.
c
o
m
/
f
a
c
eda
ba
s
e.
Evaluation Warning : The document was created with Spire.PDF for Python.