I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
1
6
,
No
.
4
,
A
u
g
u
s
t
20
2
6
,
p
p
.
2
1
3
4
~
2
1
5
7
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
1
6
i
4
.
pp
2
1
3
4
-
2
1
5
7
2134
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Lig
htweight
face
recog
nition ba
sed
on
PC
A coupled
with
PS
O
-
ba
sed
f
ea
ture
o
ptim
iza
ti
o
n in r
eso
urce
-
co
nstra
ined
env
iro
nments
Cha
im
a
a
K
ho
ud
da
,
Z
ineb G
o
t
t
i,
Sa
lma
Azz
o
uzi
,
M
o
ula
y
E
l H
a
s
s
a
n Cha
ra
f
La
b
o
r
a
t
o
r
y
o
f
R
e
se
a
r
c
h
i
n
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
,
F
a
c
u
l
t
y
o
f
S
c
i
e
n
c
e
s
,
I
b
n
To
f
a
i
l
U
n
i
v
e
r
si
t
y
,
K
e
n
i
t
r
a
,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
2
3
,
2
0
2
5
R
ev
is
ed
Ma
y
7
,
2
0
2
6
Acc
ep
ted
J
u
l 2
2
,
2
0
2
6
We p
re
se
n
t
a
stre
a
m
li
n
e
d
P
CA
-
P
S
O stru
c
tu
re
o
f
l
ig
h
twe
ig
h
t
fa
c
e
re
c
o
g
n
it
i
o
n
in
t
h
is
p
a
p
e
r,
wh
ich
c
o
m
b
in
e
s
p
ri
n
c
ip
a
l
c
o
m
p
o
n
e
n
t
a
n
a
ly
sis
(
P
CA)
to
re
d
u
c
e
d
ime
n
sio
n
s
with
p
a
rti
c
le
sw
a
rm
o
p
ti
m
iza
ti
o
n
(P
S
O)
to
a
d
a
p
t
iv
e
ly
se
lec
t
fe
a
tu
re
s.
In
c
o
n
tras
t
to
trad
i
ti
o
n
a
l
P
CA
-
b
a
se
d
m
o
d
e
ls,
o
u
r
m
o
d
e
l
d
y
n
a
m
ica
ll
y
c
h
o
o
se
s
th
e
m
o
st
i
n
fo
rm
a
ti
v
e
su
b
se
t
o
f
th
e
4
4
fe
a
tu
re
s
a
n
d
re
su
lt
s
in
a
sm
a
ll
b
u
t
i
n
fo
rm
a
ti
v
e
re
p
re
se
n
tatio
n
.
T
h
e
a
p
p
r
o
a
c
h
h
a
s
a
h
ig
h
re
c
o
g
n
it
i
o
n
a
c
c
u
ra
c
y
o
f
9
9
.
9
9
o
n
ORL
a
n
d
9
8
.
4
5
o
n
L
F
W,
wit
h
l
o
w
late
n
c
y
(a
p
p
ro
x
ima
tely
3
se
c
o
n
d
s
p
e
r
e
p
o
c
h
)
a
n
d
l
o
w
m
e
m
o
ry
fo
o
t
p
rin
t
,
a
n
d
wit
h
o
u
t
th
e
n
e
e
d
t
o
u
se
a
g
ra
p
h
ic ca
rd
to
r
u
n
i
t.
E
x
ten
si
v
e
c
o
m
p
u
tat
io
n
a
l
stu
d
ies
h
a
v
e
sh
o
wn
th
a
t
t
h
e
p
r
o
p
o
se
d
p
ip
e
li
n
e
is
m
u
c
h
fa
ste
r
a
n
d
c
o
n
su
m
e
s
les
s
m
e
m
o
ry
th
a
n
P
CA
-
ACO
a
n
d
li
g
h
twe
ig
h
t
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(
CNN
)
m
o
d
e
ls
li
k
e
M
o
b
i
leN
e
tV2
a
n
d
S
q
u
e
e
z
e
Ne
t
a
n
d
is
b
e
st
su
it
e
d
t
o
r
u
n
i
n
re
a
l
-
ti
m
e
in
CP
U
-
li
m
it
e
d
e
n
v
iro
n
m
e
n
t
s.
Th
e
sta
ti
stica
l
tes
ts
p
ro
v
e
t
h
e
b
e
tt
e
rm
e
n
t
o
f
o
u
r
m
e
th
o
d
c
o
m
p
a
re
d
to
th
e
p
a
st
m
e
th
o
d
s,
wh
e
re
th
e
c
h
a
n
g
e
s
a
re
sta
ti
stica
ll
y
sig
n
if
ica
n
t.
Th
is
re
se
a
rc
h
o
ffe
rs
a
n
e
fficie
n
t,
sim
p
le,
y
e
t
sc
a
lab
le
a
lt
e
rn
a
ti
v
e
t
o
d
e
e
p
lea
rn
in
g
-
b
a
se
d
re
c
o
g
n
it
i
o
n
s
y
ste
m
s,
e
sp
e
c
ially
w
h
e
n
e
m
b
e
d
d
e
d
i
n
teg
ra
ti
o
n
is
a
k
e
y
re
q
u
irem
e
n
t,
b
y
o
ffe
rin
g
ra
p
i
d
c
o
n
v
e
rg
e
n
c
e
,
a
d
a
p
ti
v
e
fe
a
tu
re
se
lec
ti
o
n
,
a
n
d
c
o
m
p
a
c
t
re
p
re
se
n
tatio
n
.
K
ey
w
o
r
d
s
:
Face
r
ec
o
g
n
itio
n
Featu
r
e
s
elec
tio
n
Op
tim
izatio
n
Par
ticle
s
war
m
o
p
tim
izatio
n
Prin
cip
al
co
m
p
o
n
en
t a
n
al
y
s
is
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
C
h
aim
aa
Kh
o
u
d
d
a
L
ab
o
r
ato
r
y
o
f
R
esear
ch
in
C
o
m
p
u
ter
Scien
ce
,
Facu
lty
o
f
Sc
ien
ce
s
,
I
b
n
T
o
f
ail
Un
iv
er
s
ity
Ken
itra
,
Mo
r
o
cc
o
E
m
ail: k
h
o
u
d
d
a.
c@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Au
to
m
atic
f
ac
e
r
ec
o
g
n
itio
n
h
as
b
ec
o
m
e
a
c
r
itical
co
m
p
o
n
en
t
i
n
a
wid
e
r
a
n
g
e
o
f
r
ea
l
-
wo
r
ld
ap
p
licatio
n
s
o
f
ten
im
p
er
ce
p
tib
le
to
u
s
er
s
s
u
ch
as
v
i
d
eo
s
u
r
v
eillan
ce
,
ac
ce
s
s
c
o
n
tr
o
l
(
en
tr
y
/ex
it)
,
au
th
en
ticatio
n
s
y
s
tem
s
,
an
d
h
u
m
an
–
co
m
p
u
ter
in
ter
ac
tio
n
.
Ho
wev
er
,
m
o
d
elin
g
th
e
h
u
m
an
f
ac
e
r
em
ain
s
a
ch
allen
g
in
g
task
d
u
e
to
v
ar
iat
io
n
s
in
lig
h
tin
g
co
n
d
itio
n
s
,
h
ea
d
p
o
s
es,
an
d
f
ac
ial
e
x
p
r
ess
io
n
s
.
T
h
ese
f
ac
t
o
r
s
m
ak
e
it d
if
f
icu
lt to
d
esig
n
f
ac
e
r
ec
o
g
n
itio
n
alg
o
r
ith
m
s
th
at
ar
e
b
o
th
ac
c
u
r
ate
an
d
co
m
p
u
tatio
n
ally
ef
f
icien
t.
Am
o
n
g
t
h
e
m
o
s
t
wid
ely
u
s
ed
m
eth
o
d
s
f
o
r
f
ac
ial
r
ep
r
esen
tatio
n
,
p
r
in
cip
al
co
m
p
o
n
en
t
an
aly
s
is
(
PC
A)
s
tan
d
s
o
u
t
f
o
r
its
ab
ilit
y
to
r
ed
u
ce
d
ata
d
im
en
s
io
n
ality
w
h
ile
p
r
eser
v
in
g
th
e
m
o
s
t
s
ig
n
if
ican
t
v
ar
iatio
n
s
i
n
f
ac
ial
im
ag
es.
Ap
p
ly
in
g
PC
A
to
th
e
r
ec
o
g
n
itio
n
task
tr
an
s
f
o
r
m
s
th
e
o
r
ig
in
al
h
ig
h
-
d
im
en
s
io
n
al
im
ag
e
s
p
ac
e
in
to
a
lo
wer
-
d
im
en
s
io
n
al
s
u
b
s
p
ac
e
co
m
m
o
n
ly
r
ef
er
r
ed
to
as
th
e
eig
en
f
ac
e
s
p
ac
e,
wh
ich
s
ig
n
if
ican
tly
r
ed
u
ce
s
th
e
co
m
p
u
tatio
n
al
lo
ad
w
h
ile
r
etain
in
g
d
is
cr
im
in
ativ
e
f
ac
ial
f
ea
tu
r
es.
Nev
er
th
eless
,
n
o
t
all
p
r
in
cip
al
co
m
p
o
n
en
ts
ex
tr
ac
ted
v
ia
PC
A
co
n
tr
ib
u
te
eq
u
ally
to
th
e
cla
s
s
if
icatio
n
task
.
Me
tah
eu
r
is
tic
alg
o
r
ith
m
s
s
u
ch
as
p
ar
ticle
s
war
m
o
p
t
im
izatio
n
(
PS
O)
ar
e
p
ar
ticu
la
r
ly
well
-
s
u
ited
f
o
r
h
ig
h
-
d
im
e
n
s
io
n
al
f
ea
tu
r
e
s
elec
tio
n
d
u
e
t
o
th
eir
a
b
ilit
y
to
ef
f
icien
tly
ex
p
l
o
r
e
lar
g
e
s
ea
r
ch
s
p
ac
es
with
o
u
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Lig
h
tw
eig
h
t fa
ce
r
ec
o
g
n
itio
n
b
a
s
ed
o
n
P
C
A
co
u
p
le
d
w
ith
P
S
O
-
b
a
s
ed
fea
tu
r
e
…
(
C
h
a
ima
a
K
h
o
u
d
d
a
)
2135
r
ely
in
g
o
n
g
r
ad
ie
n
t
in
f
o
r
m
atio
n
.
I
n
s
p
ir
ed
b
y
th
e
co
llectiv
e
b
eh
av
io
r
o
f
b
ir
d
s
f
lo
c
k
in
g
o
r
f
i
s
h
s
ch
o
o
lin
g
,
PS
O
is
a
p
o
p
u
latio
n
-
b
ased
o
p
tim
iz
atio
n
alg
o
r
ith
m
ca
p
ab
le
o
f
ef
f
i
cien
tly
s
ea
r
ch
in
g
th
e
f
ea
tu
r
e
s
p
ac
e
to
id
en
tif
y
th
e
o
p
tim
al
o
r
n
ea
r
-
o
p
tim
al
co
m
b
in
atio
n
o
f
co
m
p
o
n
e
n
ts
th
at
lead
to
ac
cu
r
ate
class
if
icatio
n
.
W
h
en
f
ea
tu
r
e
ex
tr
ac
tio
n
u
s
in
g
PC
A
is
f
o
llo
wed
b
y
f
ea
t
u
r
e
o
p
tim
izatio
n
u
s
in
g
PS
O,
th
e
o
v
er
all
f
ac
e
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
ca
n
b
e
s
ig
n
if
ican
tly
en
h
an
ce
d
in
te
r
m
s
o
f
b
o
th
ac
cu
r
ac
y
an
d
co
m
p
u
tatio
n
al
p
er
f
o
r
m
a
n
ce
.
I
n
th
is
p
ap
er
,
we
p
r
o
p
o
s
e
a
c
o
m
p
lete
f
ac
e
r
ec
o
g
n
iti
o
n
p
ip
elin
e
t
h
at
co
m
b
in
es
PC
A
f
o
r
d
im
en
s
io
n
ality
r
e
d
u
ctio
n
with
PS
O
f
o
r
o
p
tim
al
f
ea
tu
r
e
s
elec
tio
n
,
ev
al
u
ated
o
n
t
h
e
OR
L
f
ac
ial
d
ataset.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
aim
s
to
im
p
r
o
v
e
b
o
th
r
ec
o
g
n
itio
n
ac
c
u
r
ac
y
an
d
co
m
p
u
tatio
n
al
e
f
f
icien
c
y
b
y
r
etain
in
g
o
n
l
y
th
e
m
o
s
t
d
is
cr
im
in
ativ
e
p
r
in
ci
p
al
co
m
p
o
n
en
ts
.
T
o
v
alid
ate
o
u
r
ap
p
r
o
ac
h
,
we
p
r
esen
t
in
s
e
ctio
n
2
a
r
ev
iew
o
f
r
elate
d
wo
r
k
,
p
r
elim
in
ar
ies
in
s
ec
tio
n
3
,
f
o
llo
wed
b
y
o
u
r
m
eth
o
d
o
l
o
g
y
in
s
ec
tio
n
4
,
th
e
ex
p
er
im
en
tal
s
etu
p
an
d
r
esu
lts
in
s
ec
tio
n
5
,
a
co
m
p
ar
ativ
e
d
is
cu
s
s
io
n
in
s
ec
tio
n
6
,
an
d
co
n
clu
d
in
g
r
em
ar
k
s
in
s
ec
tio
n
7
.
Face
r
ec
o
g
n
itio
n
is
a
d
if
f
ic
u
lt
task
b
ec
au
s
e
o
f
th
e
p
o
s
e,
lig
h
t,
f
ac
e
ex
p
r
ess
io
n
a
n
d
o
cc
lu
s
io
n
v
ar
iatio
n
s
,
wh
ich
f
r
e
q
u
en
tly
r
en
d
er
th
e
r
aw
p
ix
el
-
b
ased
r
ep
r
esen
tatio
n
s
in
ad
e
q
u
ate
to
p
r
o
v
id
e
s
tr
o
n
g
class
if
icatio
n
.
Alth
o
u
g
h
m
o
d
e
r
n
d
ee
p
lear
n
in
g
-
b
ased
a
p
p
r
o
a
ch
es
lik
e
co
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
(
C
NN
s
)
h
av
e
d
em
o
n
s
tr
ated
r
em
ar
k
a
b
le
r
esu
lts
,
th
ey
d
em
an
d
lar
g
e
-
s
c
ale,
an
n
o
tated
d
atasets
,
s
u
b
s
tan
tial c
o
m
p
u
tatio
n
al
r
eso
u
r
ce
s
,
an
d
im
p
lem
e
n
tatio
n
o
n
a
g
r
ap
h
ics
p
r
o
ce
s
s
in
g
u
n
it
(
GPU
)
,
m
ak
in
g
th
em
p
o
ten
t
ially
in
f
ea
s
ib
le
in
a
lim
ited
en
v
ir
o
n
m
en
t
lik
e
an
e
m
b
ed
d
e
d
s
y
s
tem
,
m
o
b
ile
d
ev
i
ce
,
o
r
e
d
g
e
c
o
m
p
u
tin
g
p
latf
o
r
m
.
T
h
is
d
r
aw
b
ac
k
is
d
r
iv
in
g
th
e
s
ea
r
ch
o
f
lig
h
tweig
h
t
an
d
ef
f
icien
t
s
o
lu
tio
n
s
th
at
ca
n
d
eliv
er
h
ig
h
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
with
o
u
t
th
e
b
u
r
d
en
o
f
d
ee
p
n
etwo
r
k
s
.
Ou
r
r
esear
ch
tak
es
a
d
v
an
tag
e
o
f
PC
A,
wh
ich
ca
n
b
e
ef
f
ec
ti
v
ely
u
s
ed
as
d
im
e
n
s
io
n
ality
r
e
d
u
ctio
n
to
ex
tr
ac
t
th
e
m
o
s
t
im
p
o
r
tan
t
f
ac
e
elem
en
ts
,
an
d
s
h
ed
th
e
u
n
n
e
ce
s
s
ar
y
o
n
es,
an
d
PS
O,
wh
ich
ca
n
b
e
ef
f
ec
tiv
ely
u
s
ed
as
f
ea
tu
r
e
s
elec
tio
n
to
i
d
en
tify
th
e
b
est
a
n
d
o
p
tim
al
s
u
b
s
et
o
f
co
m
p
o
n
en
ts
t
o
u
s
e
in
class
if
icatio
n
.
T
h
e
r
esu
lt
o
f
th
is
co
m
b
in
atio
n
is
a
co
m
p
u
tatio
n
ally
ef
f
icien
t
b
u
t
d
is
cr
im
in
ativ
e
f
r
am
ewo
r
k
,
w
h
ich
ca
n
r
ea
c
h
n
ea
r
-
p
er
f
ec
t
ac
c
u
r
ac
y
(
~9
9
.
9
9
%
o
n
OR
L
,
~9
8
.
4
5
%
o
n
L
FW
)
with
o
n
ly
a
s
m
all
am
o
u
n
t
o
f
tr
ain
in
g
tim
e
a
n
d
r
eso
u
r
ce
s
.
W
ith
in
ter
p
r
etab
ili
ty
,
lo
w
laten
cy
,
an
d
p
r
ac
tica
l
d
ep
lo
y
m
e
n
t,
PC
A+
PS
O
s
ta
n
d
s
as
a
p
o
ten
tial
s
o
lu
tio
n
wh
er
e
C
NNs c
an
b
e
e
x
ce
s
s
iv
e
o
r
in
f
ea
s
ib
le
an
d
p
r
o
v
id
e
a
tr
ad
e
o
f
f
b
etwe
en
ac
cu
r
a
cy
.
2.
RE
L
AT
E
D
WO
RK
S
Face
r
ec
o
g
n
itio
n
h
as
b
ee
n
s
t
u
d
ied
f
o
r
d
ec
ad
es,
s
tar
tin
g
w
ith
class
ical
ap
p
ea
r
an
ce
-
b
ase
d
m
eth
o
d
s
.
T
h
e
wo
r
k
in
[
1
]
in
tr
o
d
u
ce
d
th
e
E
ig
en
f
ac
es
ap
p
r
o
ac
h
b
ased
o
n
PC
A,
wh
ich
b
ec
am
e
o
n
e
o
f
th
e
f
ir
s
t
ef
f
ec
tiv
e
tech
n
iq
u
es
f
o
r
r
ep
r
esen
tin
g
f
a
cial
im
ag
es
in
a
r
e
d
u
ce
d
s
u
b
s
p
ac
e.
T
h
e
s
tu
d
y
in
[
2
]
later
p
r
o
p
o
s
ed
a
d
y
n
a
m
ic
tex
tu
r
e
m
o
d
el
u
s
in
g
lo
ca
l
b
in
a
r
y
p
atter
n
s
f
o
r
f
ac
ial
e
x
p
r
ess
io
n
r
ec
o
g
n
itio
n
.
W
ith
th
e
r
is
e
o
f
d
ee
p
lear
n
in
g
,
s
ev
er
al
m
o
d
els
h
av
e
ac
h
iev
ed
h
i
g
h
ac
cu
r
ac
y
i
n
co
m
p
lex
en
v
ir
o
n
m
en
ts
.
T
h
e
m
eth
o
d
in
[
3
]
d
esig
n
e
d
a
n
illu
m
in
atio
n
-
in
v
ar
ian
t
g
en
e
r
ativ
e
a
d
v
er
s
ar
ial
n
etwo
r
k
m
o
d
el
(IL
-
GAN)
th
at
im
p
r
o
v
es
p
e
r
f
o
r
m
an
ce
u
n
d
e
r
v
ar
y
in
g
lig
h
ti
n
g
co
n
d
itio
n
s
,
wh
ile
th
e
wo
r
k
in
[
4
]
p
r
esen
ted
Ar
cFac
e,
in
tr
o
d
u
cin
g
an
an
g
u
lar
m
ar
g
in
l
o
s
s
to
en
h
an
ce
f
ea
tu
r
e
d
is
cr
im
in
atio
n
i
n
d
ee
p
em
b
ed
d
in
g
s
.
T
h
e
ap
p
r
o
ac
h
in
[
5
]
co
m
b
in
ed
wa
v
elet
an
d
PC
A
f
ea
tu
r
es
t
h
r
o
u
g
h
a
d
ee
p
W
T
PC
A
-
L
1
m
o
d
e
l,
o
b
tain
in
g
r
o
b
u
s
t
r
ec
o
g
n
itio
n
u
n
d
er
n
o
is
e
an
d
o
cc
lu
s
io
n
.
I
n
p
ar
allel,
o
p
tim
izatio
n
alg
o
r
ith
m
s
h
av
e
g
ain
ed
atten
tio
n
f
o
r
im
p
r
o
v
in
g
f
ea
tu
r
e
s
elec
tio
n
an
d
class
if
icatio
n
.
T
h
e
s
tu
d
y
in
[
6
]
p
r
o
p
o
s
ed
a
PS
O
-
b
ased
b
lo
c
k
f
ea
tu
r
e
s
elec
tio
n
f
o
r
r
o
b
u
s
t
r
ec
o
g
n
itio
n
,
wh
ile
th
e
wo
r
k
in
[
7
]
ap
p
lied
PS
O
in
tr
an
s
f
er
lear
n
in
g
to
en
h
a
n
ce
ad
ap
tab
ilit
y
.
T
h
e
a
p
p
r
o
ac
h
i
n
[
8
]
in
tr
o
d
u
ce
d
an
en
h
an
ce
d
PS
O
v
ar
ia
n
t
f
o
r
b
et
ter
co
n
v
er
g
e
n
ce
i
n
class
if
icatio
n
task
s
.
T
h
e
r
esear
ch
in
[
9
]
u
tili
ze
d
f
r
ac
tio
n
al
-
o
r
d
er
m
u
lti
-
ch
an
n
el
m
o
m
e
n
ts
f
o
r
co
l
o
r
f
ac
e
r
ec
o
g
n
itio
n
,
im
p
r
o
v
in
g
ac
cu
r
ac
y
with
co
m
p
ac
t f
ea
tu
r
es.
Hy
b
r
id
s
y
s
tem
s
co
m
b
in
in
g
PS
O
an
d
d
ee
p
m
o
d
els
h
av
e
al
s
o
b
ee
n
in
v
esti
g
ated
.
T
h
e
s
tu
d
y
in
[
1
0
]
ap
p
lied
PS
O
with
Gab
o
r
a
n
d
d
ee
p
lear
n
in
g
to
s
elec
t
o
p
tim
a
l
f
ea
tu
r
es.
T
h
e
wo
r
k
i
n
[
1
1
]
in
teg
r
ated
PS
O
with
a
d
ee
p
b
elief
n
etwo
r
k
to
e
n
h
an
ce
f
ea
tu
r
e
o
p
tim
izatio
n
,
an
d
[
1
2
]
u
s
ed
PS
O
to
o
p
tim
ize
Ga
b
o
r
f
ilter
b
a
n
k
s
f
o
r
co
m
p
ac
t a
n
d
d
is
cr
im
in
ativ
e
r
e
p
r
esen
tatio
n
.
Oth
er
s
tu
d
ies
ex
p
lo
r
ed
a
d
ap
t
iv
e
o
r
im
p
r
o
v
ed
PS
O
v
ar
ian
ts
.
T
h
e
r
esear
ch
in
[
1
3
]
p
r
o
p
o
s
ed
an
im
p
r
o
v
e
d
PS
O
f
o
r
im
ag
e
f
ea
tu
r
e
co
m
p
en
s
atio
n
,
ac
h
ie
v
in
g
h
ig
h
p
e
r
f
o
r
m
an
ce
with
lo
w
co
m
p
lex
ity
.
T
h
e
w
o
r
k
in
[
1
4
]
in
tr
o
d
u
ce
d
a
g
u
i
d
ed
b
in
ar
y
PS
O
f
o
r
f
ea
tu
r
e
-
lev
el
f
u
s
io
n
,
an
d
[
1
5
]
d
ev
el
o
p
ed
a
n
etwo
r
k
-
s
tr
u
ctu
r
e
d
PS
O
f
o
r
s
tab
le
co
n
v
er
g
e
n
ce
.
T
h
e
s
tu
d
y
in
[
1
6
]
d
em
o
n
s
tr
ated
th
e
b
en
ef
its
o
f
o
p
tim
iza
tio
n
alg
o
r
ith
m
s
f
o
r
class
if
icatio
n
in
en
g
in
ee
r
in
g
a
p
p
licatio
n
s
.
C
o
m
p
r
eh
en
s
iv
e
s
u
r
v
e
y
s
,
s
u
ch
as
[
1
7
]
,
r
ev
iewe
d
th
e
e
v
o
lu
t
io
n
o
f
f
ea
tu
r
e
ex
tr
ac
tio
n
a
n
d
s
elec
tio
n
m
eth
o
d
s
in
f
ac
e
r
ec
o
g
n
itio
n
.
T
h
e
f
r
a
m
ewo
r
k
in
[
1
8
]
p
r
esen
ted
a
n
o
is
e
-
r
o
b
u
s
t
d
ictio
n
a
r
y
lear
n
in
g
ap
p
r
o
ac
h
,
wh
ile
[
1
9
]
a
n
d
[
2
0
]
s
h
o
we
d
th
e
ef
f
icien
cy
o
f
s
war
m
-
b
ased
lear
n
in
g
f
o
r
r
ea
l
-
tim
e
an
d
s
u
r
v
eillan
ce
ap
p
licatio
n
s
.
Ov
er
all,
th
ese
s
t
u
d
ies
co
n
f
ir
m
t
h
at
in
teg
r
atin
g
s
u
b
s
p
ac
e
lear
n
in
g
with
PS
O
-
b
ased
o
p
tim
izatio
n
im
p
r
o
v
es r
ec
o
g
n
itio
n
ac
cu
r
ac
y
an
d
c
o
m
p
u
tatio
n
al
ef
f
icien
c
y
.
T
h
e
r
esear
ch
in
[
2
1
]
d
ev
elo
p
ed
a
b
lu
r
-
in
v
ar
ian
t
b
in
a
r
y
d
e
s
cr
ip
to
r
to
im
p
r
o
v
e
r
o
b
u
s
tn
e
s
s
ag
ain
s
t
d
eg
r
ad
e
d
im
ag
es,
a
n
d
[
2
2
]
p
r
o
v
id
ed
a
co
m
p
r
eh
e
n
s
iv
e
o
v
er
v
iew
o
f
f
o
u
n
d
atio
n
al
f
ac
e
r
ec
o
g
n
itio
n
m
eth
o
d
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
2
1
3
4
-
2
1
5
7
2136
T
h
e
wo
r
k
i
n
[
2
3
]
co
m
b
in
ed
PS
O
with
Ad
aBo
o
s
t
to
h
an
d
le
ad
v
e
r
s
ar
ial
s
am
p
les
ef
f
ec
tiv
ely
,
wh
ile
[
1
2
]
o
p
tim
ized
Gab
o
r
f
ilter
b
a
n
k
s
u
s
in
g
PS
O
f
o
r
ef
f
icien
t a
n
d
d
is
cr
im
in
ativ
e
f
ea
tu
r
e
s
elec
tio
n
.
I
n
a
m
o
r
e
r
ec
en
t
c
o
n
tr
ib
u
tio
n
,
[
2
4
]
p
r
o
p
o
s
ed
a
h
y
b
r
id
o
p
tim
izatio
n
m
eth
o
d
c
o
m
b
in
i
n
g
q
u
an
tu
m
-
in
s
p
ir
ed
f
ir
e
f
ly
a
n
d
ar
tific
ial
b
ee
c
o
lo
n
y
alg
o
r
ith
m
s
f
o
r
m
u
lti
-
p
o
s
e
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
,
f
u
r
th
er
co
n
f
ir
m
in
g
th
e
ef
f
ec
tiv
en
ess
o
f
s
war
m
-
b
ased
o
p
tim
izatio
n
in
f
ac
ial
an
aly
s
is
task
s
.
I
n
s
u
m
m
a
r
y
,
p
r
e
v
io
u
s
s
tu
d
ies
h
ig
h
lig
h
t
th
at
in
teg
r
atin
g
f
ea
tu
r
e
ex
tr
ac
tio
n
with
s
war
m
-
b
ased
o
p
tim
izatio
n
tech
n
iq
u
es
s
ig
n
if
ican
tly
im
p
r
o
v
es
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
wh
ile
m
ai
n
tain
in
g
co
m
p
u
tatio
n
al
ef
f
icien
cy
.
B
u
ild
in
g
o
n
th
ese
in
s
ig
h
ts
,
th
e
p
r
esen
t
wo
r
k
i
n
tr
o
d
u
ce
s
an
e
n
er
g
y
-
ef
f
icien
t
f
ac
e
r
ec
o
g
n
itio
n
f
r
am
ewo
r
k
th
at
co
m
b
in
es
PC
A
f
o
r
d
im
en
s
io
n
ality
r
ed
u
cti
o
n
with
PS
O
f
o
r
o
p
tim
al
f
ea
tu
r
e
s
elec
tio
n
.
T
h
is
in
teg
r
atio
n
aim
s
to
ac
h
iev
e
a
b
alan
ce
d
tr
ad
e
-
o
f
f
b
etwe
en
ac
cu
r
ac
y
,
s
tab
ilit
y
,
an
d
co
m
p
u
ta
tio
n
al
co
s
t,
m
ak
in
g
it su
itab
le
f
o
r
r
ea
l
-
wo
r
ld
b
io
m
etr
ic
ap
p
licatio
n
s
.
3.
P
RE
L
I
M
I
NAR
I
E
S
3
.
1
.
P
rincipa
l
co
m
po
nent
a
na
ly
s
is
PC
A
is
a
wid
ely
u
s
ed
d
im
en
s
i
o
n
ality
r
e
d
u
ctio
n
tech
n
iq
u
e
in
m
ac
h
in
e
lear
n
i
n
g
a
n
d
d
ata
an
aly
s
is
.
I
ts
o
b
jectiv
e
is
to
r
etain
th
e
m
o
s
t
in
f
o
r
m
ativ
e
p
atter
n
s
in
th
e
d
at
a
wh
ile
r
ed
u
cin
g
th
e
n
u
m
b
e
r
o
f
f
ea
tu
r
es,
t
h
er
eb
y
s
im
p
lify
in
g
th
e
d
ataset
with
o
u
t sig
n
if
ican
t lo
s
s
o
f
v
ar
ian
ce
.
PC
A
tr
an
s
f
o
r
m
s
th
e
o
r
ig
in
al
co
r
r
elate
d
v
a
r
iab
les
in
to
a
n
e
w
s
et
o
f
u
n
c
o
r
r
elate
d
v
ar
iab
l
es
ca
lled
p
r
in
cip
al
co
m
p
o
n
en
ts
.
T
h
ese
co
m
p
o
n
en
ts
ar
e
o
b
tain
ed
b
y
p
er
f
o
r
m
in
g
a
n
eig
e
n
v
alu
e
d
e
co
m
p
o
s
itio
n
o
f
th
e
d
ata'
s
co
v
ar
ian
ce
m
atr
ix
.
T
h
e
eig
en
v
ec
to
r
s
r
e
p
r
esen
t
th
e
d
ir
ec
tio
n
s
o
f
m
ax
im
u
m
v
ar
ian
ce
,
an
d
th
ei
r
ass
o
ciate
d
eig
en
v
alu
es
in
d
ica
te
th
e
am
o
u
n
t
o
f
v
ar
ia
n
ce
ca
p
tu
r
ed
b
y
ea
ch
d
ir
ec
tio
n
.
B
y
s
elec
tin
g
th
e
to
p
co
m
p
o
n
en
ts
with
th
e
h
ig
h
est
eig
en
v
alu
es,
PC
A
p
r
o
jects
t
h
e
d
ata
in
to
a
lo
wer
-
d
im
en
s
i
o
n
al
s
u
b
s
p
ac
e
th
at
p
r
eser
v
es
th
e
ess
en
tial
s
tr
u
ctu
r
e
o
f
th
e
o
r
ig
in
al
d
ataset.
PC
A
i
s
p
ar
ticu
lar
ly
ef
f
ec
tiv
e
f
o
r
v
is
u
alizin
g
an
d
an
aly
zin
g
h
ig
h
-
d
im
en
s
io
n
al
d
ata,
r
ed
u
cin
g
r
ed
u
n
d
a
n
cy
,
an
d
im
p
r
o
v
i
n
g
t
h
e
co
m
p
u
tat
io
n
al
ef
f
icien
c
y
o
f
d
o
wn
s
tr
ea
m
m
ac
h
i
n
e
lear
n
in
g
alg
o
r
ith
m
s
.
T
h
e
b
lack
a
x
es
in
Fig
u
r
e
1
r
e
p
r
esen
t
th
e
o
r
i
g
in
al
f
ea
tu
r
es
"
R
ad
iu
s
"
an
d
"Ar
ea
"
o
f
th
e
d
at
aset.
PC
A
id
en
tifie
s
two
n
ew
o
r
th
o
g
o
n
al
ax
es,
PC
₁
an
d
PC
₂,
th
at
r
ep
r
esen
t
th
e
d
ir
ec
tio
n
s
o
f
m
ax
im
u
m
v
ar
ian
ce
in
th
e
d
ata.
T
h
ese
n
ew
ax
es r
esu
lt f
r
o
m
a
r
o
tatio
n
o
f
th
e
o
r
i
g
in
al
f
e
atu
r
e
s
p
ac
e.
Fig
u
r
e
1
.
PC
A
alg
o
r
ith
m
f
ea
tu
r
es e
x
tr
ac
tio
n
wo
r
k
f
lo
w
PC
₁
ca
p
tu
r
es
th
e
g
r
ea
test
am
o
u
n
t
o
f
v
ar
iatio
n
,
wh
ile
PC
₂,
b
ein
g
o
r
th
o
g
o
n
al
to
PC
₁,
ca
p
tu
r
es
th
e
r
em
ain
in
g
v
ar
ia
n
ce
.
Sin
ce
th
e
d
ata
p
o
in
ts
ar
e
m
o
r
e
s
p
r
ea
d
o
u
t
alo
n
g
PC
₁
th
an
PC
₂,
p
r
o
je
ctin
g
th
e
d
ata
o
n
to
PC
₁
allo
ws
a
r
ed
u
ctio
n
f
r
o
m
two
d
im
en
s
io
n
s
to
o
n
e,
wh
i
le
p
r
eser
v
in
g
m
o
s
t
o
f
th
e
r
el
ev
an
t
s
tr
u
ctu
r
e
an
d
in
f
o
r
m
atio
n
.
3
.
2
.
P
CA
f
o
r
f
a
cia
l f
ea
t
ures e
x
t
ra
ct
io
n
A
k
ey
s
tep
in
t
h
e
p
r
o
p
o
s
ed
p
ip
elin
e
is
th
e
u
s
e
o
f
PC
A
as
a
d
im
en
s
io
n
ality
r
ed
u
ctio
n
te
ch
n
iq
u
e
t
o
ex
tr
ac
t
th
e
m
o
s
t
r
elev
an
t
f
ac
ial
f
ea
tu
r
es
f
r
o
m
th
e
tr
ain
in
g
i
m
ag
es.
E
ac
h
f
ac
e
im
ag
e
is
f
ir
s
t
r
esh
ap
ed
in
to
a
v
ec
to
r
an
d
s
tack
ed
co
l
u
m
n
-
wi
s
e
to
f
o
r
m
a
d
ata
m
atr
ix
,
wh
er
e
ea
ch
co
lu
m
n
co
r
r
esp
o
n
d
s
to
o
n
e
tr
ain
in
g
im
ag
e.
T
h
e
m
ea
n
f
ac
e
v
ec
to
r
is
th
en
co
m
p
u
ted
an
d
s
u
b
tr
ac
ted
f
r
o
m
all
im
ag
e
v
ec
to
r
s
,
ce
n
ter
in
g
th
e
d
ata
ar
o
u
n
d
th
e
o
r
ig
in
.
T
h
is
s
tep
en
s
u
r
es th
at
th
e
p
r
in
cip
al
co
m
p
o
n
en
ts
r
ef
lect
th
e
d
ir
ec
tio
n
s
o
f
m
ax
im
u
m
v
ar
ia
n
ce
in
th
e
d
ataset,
co
r
r
esp
o
n
d
i
n
g
t
o
th
e
m
o
s
t in
f
o
r
m
ativ
e
d
if
f
er
e
n
ce
s
in
f
ac
ial
s
tr
u
ctu
r
e
an
d
ap
p
ea
r
an
ce
.
T
h
e
co
v
ar
ian
ce
m
at
r
ix
o
f
th
e
ce
n
ter
ed
d
ata
is
ca
lcu
lated
to
ca
p
tu
r
e
h
o
w
p
ix
el
in
ten
s
ities
v
ar
y
to
g
eth
er
ac
r
o
s
s
th
e
tr
ain
in
g
s
et
.
Du
e
to
th
e
h
ig
h
d
im
en
s
io
n
ali
ty
o
f
f
ac
ial
im
a
g
es,
a
co
m
p
u
ta
tio
n
al
tr
ick
is
u
s
ed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Lig
h
tw
eig
h
t fa
ce
r
ec
o
g
n
itio
n
b
a
s
ed
o
n
P
C
A
co
u
p
le
d
w
ith
P
S
O
-
b
a
s
ed
fea
tu
r
e
…
(
C
h
a
ima
a
K
h
o
u
d
d
a
)
2137
b
y
f
o
r
m
in
g
a
s
m
aller
co
v
a
r
ia
n
ce
m
atr
ix
=
,
wh
er
e
is
th
e
m
ea
n
-
ce
n
ter
ed
d
ata
m
atr
ix
.
E
ig
en
v
alu
e
d
ec
o
m
p
o
s
itio
n
is
th
en
a
p
p
lie
d
to
o
b
tain
th
e
eig
en
v
alu
es
an
d
eig
e
n
v
ec
to
r
s
,
w
h
ich
r
e
p
r
esen
t
th
e
p
r
in
ci
p
al
d
ir
ec
tio
n
s
o
f
v
ar
iatio
n
.
T
h
ese
eig
en
v
ec
to
r
s
,
o
f
ten
r
ef
er
r
ed
to
as
eig
e
n
f
ac
es,
d
ef
i
n
e
a
n
ew
o
r
th
o
n
o
r
m
al
b
asis
s
p
an
n
in
g
a
s
u
b
s
p
ac
e
th
at
ca
p
t
u
r
es
th
e
m
o
s
t
s
alien
t
f
ac
ial
ch
ar
ac
ter
is
tics
.
T
h
e
ei
g
en
v
al
u
es
ar
e
s
o
r
te
d
i
n
d
escen
d
i
n
g
o
r
d
er
,
an
d
th
e
to
p
co
m
p
o
n
e
n
ts
ar
e
r
etain
ed
.
E
ac
h
f
ac
e
im
ag
e
is
th
en
p
r
o
jecte
d
in
to
th
is
r
ed
u
ce
d
-
d
im
en
s
io
n
al
s
p
ac
e
b
y
co
m
p
u
tin
g
its
d
o
t
p
r
o
d
u
ct
with
th
e
s
elec
ted
eig
en
f
ac
es,
p
r
o
d
u
cin
g
a
co
m
p
ac
t
f
ea
tu
r
e
v
ec
to
r
th
at
p
r
eser
v
es
th
e
m
o
s
t d
is
cr
im
in
ativ
e
in
f
o
r
m
atio
n
,
as illu
s
tr
ated
in
Fig
u
r
e
2
.
Fig
u
r
e
2
.
PC
A
wo
r
k
f
lo
w
f
o
r
e
ig
en
f
ac
e
g
e
n
er
atio
n
3
.
3
.
P
a
rt
icle
s
wa
rm
o
ptim
iz
a
t
io
n
PSO
i
s
a
p
o
p
u
latio
n
-
b
ased
m
e
tah
eu
r
is
tic
in
s
p
ir
ed
b
y
th
e
co
ll
ec
tiv
e
b
eh
av
io
r
o
f
s
war
m
s
in
n
atu
r
e.
I
n
th
is
m
eth
o
d
,
ea
ch
i
n
d
iv
id
u
al
s
o
lu
tio
n
ca
lled
a
p
ar
ticle
r
e
p
r
esen
ts
a
p
o
te
n
tial
s
o
lu
tio
n
to
an
o
p
tim
izatio
n
p
r
o
b
lem
.
T
h
ese
p
ar
ticles
"m
o
v
e"
th
r
o
u
g
h
th
e
s
ea
r
c
h
s
p
ac
e
b
y
u
p
d
atin
g
th
eir
p
o
s
itio
n
s
an
d
v
elo
cities
at
ea
ch
iter
atio
n
b
ased
o
n
b
o
th
th
ei
r
o
wn
b
est
-
k
n
o
wn
p
o
s
itio
n
s
(
p
er
s
o
n
al
b
est,
o
r
p
b
est
)
an
d
th
e
g
lo
b
al
b
est
-
k
n
o
wn
p
o
s
itio
n
f
o
u
n
d
b
y
th
e
s
war
m
(
g
lo
b
al
b
est,
o
r
g
b
est).
At
ea
ch
iter
atio
n
,
t
h
e
f
itn
ess
o
f
ev
e
r
y
p
ar
ticle
is
ev
alu
ated
ac
co
r
d
in
g
to
a
p
r
e
d
ef
in
ed
o
b
jectiv
e
f
u
n
ctio
n
.
B
ased
o
n
th
e
c
o
m
p
ar
is
o
n
b
etwe
en
its
cu
r
r
en
t
s
tate
an
d
th
e
b
est
-
k
n
o
w
n
s
o
lu
tio
n
s
,
ea
ch
p
ar
ticle
ad
ju
s
ts
its
v
elo
city
ac
co
r
d
in
g
l
y
.
T
h
e
u
p
d
ated
v
elo
city
is
t
h
e
n
u
s
ed
t
o
co
m
p
u
te
a
n
ew
p
o
s
it
io
n
,
an
d
th
e
p
r
o
ce
s
s
co
n
tin
u
es u
n
til co
n
v
er
g
en
ce
is
ac
h
iev
ed
o
r
a
s
to
p
p
in
g
cr
iter
i
o
n
is
m
et,
as d
ep
icted
i
n
Fig
u
r
e
3
.
Fig
u
r
e
3
.
PS
O
wo
r
k
f
l
o
w
to
f
in
d
b
est p
ath
PSO
i
s
in
s
p
ir
ed
b
y
th
e
co
llectiv
e
b
eh
av
io
r
o
f
n
atu
r
al
s
war
m
s
s
u
ch
as
f
lo
ck
s
o
f
b
ir
d
s
o
r
s
ch
o
o
ls
o
f
f
is
h
.
E
ac
h
in
d
iv
i
d
u
al
in
th
e
s
war
m
,
k
n
o
w
n
as
a
p
ar
ticle,
ex
p
lo
r
es
th
e
s
o
lu
tio
n
s
p
ac
e
b
y
a
d
ju
s
tin
g
its
p
o
s
itio
n
b
ased
o
n
b
o
th
its
p
er
s
o
n
al
ex
p
er
ien
ce
an
d
t
h
at
o
f
t
h
e
g
r
o
u
p
.
At
iter
atio
n
n
,
th
e
v
elo
city
=
(
+
1
)
o
f
a
p
ar
ticle
is
u
p
d
ated
ac
c
o
r
d
i
n
g
to
(
1
)
:
(
+
1
)
=
·
(
)
+
1
·
1
·
(
−
(
)
)
+
2
·
2
·
(
−
(
)
)
(
1
)
w
h
er
e
is
th
e
in
er
tia
weig
h
t c
o
n
tr
o
llin
g
m
o
m
en
tu
m
,
1
,
2
ar
e
c
o
g
n
itiv
e
an
d
s
o
cial
ac
ce
ler
atio
n
co
ef
f
icien
ts
,
1
,
2
ar
e
r
an
d
o
m
v
alu
es
in
[
0
,
1
]
,
is
th
e
b
est
p
o
s
itio
n
f
o
u
n
d
b
y
th
e
p
ar
ticle,
is
th
e
b
est
p
o
s
itio
n
f
o
u
n
d
b
y
th
e
s
war
m
.
T
h
e
n
ew
p
o
s
itio
n
is
th
en
co
m
p
u
ted
as (
2
)
.
(
+
1
)
=
(
)
+
(
+
1
)
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
2
1
3
4
-
2
1
5
7
2138
E
ac
h
p
ar
ticle
m
ain
tain
s
a
r
ec
o
r
d
o
f
its
o
wn
b
est
-
k
n
o
wn
s
o
lu
tio
n
(
p
ers
o
n
a
l
b
est
)
an
d
i
s
in
f
lu
en
ce
d
b
y
t
h
e
g
lo
b
ally
b
est
-
k
n
o
wn
p
o
s
itio
n
am
o
n
g
all
p
a
r
ticles
(
g
lo
b
a
l
b
e
s
t
)
.
T
h
e
s
war
m
co
n
v
er
g
es
wh
e
n
p
ar
ticles
s
tab
ilize
ar
o
u
n
d
th
e
o
p
tim
u
m
.
3
.
4
.
P
SO
f
o
r
f
e
a
t
ure
s
elec
t
io
n
I
n
th
is
s
tu
d
y
,
PSO
was
em
p
lo
y
ed
as
a
f
ea
tu
r
e
s
elec
tio
n
tech
n
iq
u
e
t
o
im
p
r
o
v
e
f
ac
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
af
ter
d
im
en
s
io
n
a
lity
r
ed
u
ctio
n
b
y
PC
A
.
W
h
ile
PC
A
ef
f
ec
tiv
ely
r
ed
u
ce
s
th
e
n
u
m
b
er
o
f
f
ea
tu
r
es,
n
o
t
all
r
etain
ed
co
m
p
o
n
en
ts
co
n
tr
ib
u
te
eq
u
ally
to
class
if
icatio
n
ac
cu
r
ac
y
.
T
h
er
ef
o
r
e,
PS
O
wa
s
u
s
ed
as
a
g
lo
b
al
o
p
tim
izatio
n
alg
o
r
ith
m
to
id
en
tify
th
e
m
o
s
t
in
f
o
r
m
ativ
e
s
u
b
s
et
o
f
PC
A
co
m
p
o
n
en
ts
.
PS
O
is
we
ll
-
s
u
ited
f
o
r
b
in
ar
y
f
ea
tu
r
e
s
elec
tio
n
p
r
o
b
lem
s
d
u
e
to
its
ab
ilit
y
to
ef
f
icien
tly
ex
p
lo
r
e
h
ig
h
-
d
im
en
s
io
n
al
d
is
cr
ete
s
p
ac
es with
o
u
t r
eq
u
ir
in
g
g
r
a
d
i
en
t in
f
o
r
m
atio
n
.
T
ec
h
n
ically
,
ea
ch
p
ar
ticle
in
t
h
e
PS
O
s
war
m
r
ep
r
esen
ts
a
b
in
ar
y
v
ec
to
r
o
f
le
n
g
th
DDD,
w
h
er
e
DDD
is
th
e
n
u
m
b
er
o
f
p
r
i
n
cip
al
co
m
p
o
n
e
n
ts
.
A
b
it
v
alu
e
o
f
1
i
n
d
icate
s
th
e
in
clu
s
io
n
o
f
a
c
o
m
p
o
n
en
t,
wh
ile
0
in
d
icate
s
ex
clu
s
io
n
.
T
h
e
p
a
r
ticles
ar
e
in
itialized
r
an
d
o
m
ly
an
d
ev
alu
ated
u
s
in
g
a
f
itn
ess
f
u
n
ctio
n
b
ased
o
n
class
if
icatio
n
ac
cu
r
ac
y
.
Fo
r
ea
ch
p
ar
ticle
:
i)
th
e
s
elec
ted
co
m
p
o
n
e
n
ts
ar
e
u
s
ed
to
p
r
o
ject
tr
ain
in
g
an
d
test
f
ac
e
im
ag
es,
ii)
a
n
ea
r
est
-
n
eig
h
b
o
r
class
if
ier
with
E
u
clid
ea
n
d
is
tan
ce
is
ap
p
lied
,
an
d
iii)
th
e
r
e
co
g
n
itio
n
ac
c
u
r
ac
y
is
m
ea
s
u
r
ed
an
d
ass
ig
n
ed
as th
e
p
ar
ticle'
s
f
itn
es
s
s
co
r
e.
Du
r
in
g
iter
atio
n
s
,
p
a
r
ticles
u
p
d
ate
th
eir
v
elo
city
an
d
p
o
s
itio
n
b
ased
o
n
th
eir
p
er
s
o
n
al
b
es
t
s
o
lu
tio
n
(
)
an
d
th
e
g
l
o
b
al
b
est
f
o
u
n
d
b
y
th
e
s
war
m
(
)
.
T
h
e
v
elo
cit
y
u
p
d
ate
in
cl
u
d
es
th
r
ee
ter
m
s
:
in
er
tia
(
m
o
m
en
tu
m
)
,
co
g
n
itiv
e
attr
ac
tio
n
(
to
war
d
)
,
a
n
d
s
o
cial
attr
ac
tio
n
(
to
war
d
)
.
A
s
ig
m
o
id
f
u
n
ctio
n
is
ap
p
lied
to
m
a
p
th
e
u
p
d
ated
v
e
lo
city
to
a
p
r
o
b
ab
ilit
y
f
o
r
s
ele
ctin
g
ea
ch
f
ea
tu
r
e.
T
h
is
iter
ativ
e
p
r
o
ce
s
s
co
n
tin
u
es
u
n
til
co
n
v
er
g
en
ce
,
lead
in
g
to
t
h
e
s
elec
tio
n
o
f
an
o
p
tim
al
o
r
n
ea
r
-
o
p
tim
al
f
ea
tu
r
e
s
u
b
s
et.
T
h
e
r
esu
lts
s
h
o
w
th
at
PS
O
ef
f
ec
ti
v
ely
r
ed
u
ce
s
th
e
d
im
e
n
s
io
n
ality
o
f
PC
A
v
ec
to
r
s
,
r
etain
s
o
n
ly
th
e
m
o
s
t
d
is
cr
im
in
ativ
e
f
ea
tu
r
es,
an
d
s
ig
n
if
ican
tly
im
p
r
o
v
es
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
ac
h
iev
in
g
r
ates
o
v
er
9
9
.
9
9
%.
Ad
d
itio
n
ally
,
t
h
e
r
ed
u
ce
d
f
ea
tu
r
e
s
et
en
h
an
ce
s
co
m
p
u
tatio
n
al
ef
f
icien
c
y
,
m
ak
in
g
th
e
s
y
s
tem
f
aster
an
d
m
o
r
e
s
u
itab
le
f
o
r
d
e
p
lo
y
m
en
t.
3
.
5
.
O
RL
da
t
a
ba
s
e
T
h
e
OR
L
d
atab
ase
o
f
f
ac
es
c
o
n
s
is
ts
o
f
4
0
0
g
r
ay
s
ca
le
im
ag
es
f
r
o
m
4
0
d
is
tin
ct
in
d
iv
id
u
al
s
,
with
1
0
im
ag
es
p
er
s
u
b
ject.
T
h
e
p
h
o
to
g
r
ap
h
s
wer
e
tak
e
n
u
n
d
er
v
ar
y
in
g
co
n
d
itio
n
s
,
in
clu
d
i
n
g
d
if
f
er
en
ce
s
in
lig
h
tin
g
,
f
ac
ial
ex
p
r
ess
io
n
s
(
e.
g
.
,
s
m
ilin
g
o
r
n
eu
tr
al,
o
p
e
n
o
r
clo
s
ed
ey
es),
an
d
f
ac
ial
d
etails
(
s
u
ch
as
p
r
esen
ce
o
r
ab
s
en
ce
o
f
g
lass
es).
All
s
u
b
jects
wer
e
p
o
s
itio
n
ed
f
r
o
n
tally
,
with
s
lig
h
t
v
ar
iatio
n
s
in
h
ea
d
o
r
ien
tatio
n
,
an
d
ca
p
tu
r
ed
a
g
ain
s
t a
u
n
if
o
r
m
d
ar
k
b
ac
k
g
r
o
u
n
d
,
as illu
s
tr
ate
d
in
Fig
u
r
e
4
.
Fig
u
r
e
4
.
OR
L
d
atab
ase
s
am
p
l
e
im
ag
es
E
ac
h
im
ag
e
h
as
a
r
eso
lu
tio
n
o
f
9
2
×1
1
2
p
ix
els
an
d
is
en
co
d
e
d
in
8
-
b
it
g
r
ay
s
ca
le
(
2
5
6
g
r
e
y
lev
els
p
er
p
ix
el)
.
T
h
e
d
ataset
is
o
r
g
an
iz
ed
in
to
4
0
f
o
ld
er
s
,
n
am
e
d
s
1
to
s
4
0
,
ea
ch
c
o
r
r
esp
o
n
d
in
g
to
a
s
u
b
ject.
W
ith
in
ea
ch
f
o
ld
e
r
,
1
0
im
ag
es a
r
e
n
a
m
ed
f
r
o
m
1
.
p
g
m
t
o
1
0
.
p
g
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Lig
h
tw
eig
h
t fa
ce
r
ec
o
g
n
itio
n
b
a
s
ed
o
n
P
C
A
co
u
p
le
d
w
ith
P
S
O
-
b
a
s
ed
fea
tu
r
e
…
(
C
h
a
ima
a
K
h
o
u
d
d
a
)
2139
W
h
ile
OR
L
i
s
a
well
-
e
s
tab
lis
h
ed
b
en
ch
m
ar
k
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
r
esear
ch
,
it
h
as
ce
r
tain
lim
itatio
n
s
.
I
t
is
r
elativ
ely
s
m
all
in
s
ize,
co
n
tain
s
well
-
alig
n
ed
im
a
g
es
with
u
n
if
o
r
m
b
ac
k
g
r
o
u
n
d
s
,
a
n
d
ex
h
ib
its
lim
ited
v
ar
iab
ilit
y
in
p
o
s
e,
illu
m
in
atio
n
,
an
d
e
x
p
r
ess
io
n
c
o
m
p
a
r
ed
t
o
r
ea
l
-
wo
r
ld
s
ce
n
a
r
io
s
.
T
h
ese
co
n
s
tr
ain
ts
m
ak
e
it
ea
s
ier
to
ac
h
iev
e
h
ig
h
ac
c
u
r
a
cy
b
u
t
lim
it
th
e
g
en
e
r
aliza
b
ilit
y
o
f
r
esu
lts
to
m
o
r
e
ch
alle
n
g
in
g
,
u
n
c
o
n
s
tr
ain
ed
en
v
ir
o
n
m
en
ts
.
T
h
is
d
ataset
s
er
v
es
a
s
th
e
b
en
ch
m
ar
k
f
o
r
o
u
r
e
x
p
er
im
e
n
ts
.
W
e
ap
p
ly
PC
A
to
ex
tr
ac
t
f
ac
ial
f
ea
tu
r
es
an
d
s
u
b
s
eq
u
e
n
tly
u
s
e
an
o
p
tim
izatio
n
alg
o
r
ith
m
(
e.
g
.
,
PS
O
o
r
AC
O)
to
en
h
an
ce
f
ea
tu
r
e
s
elec
tio
n
an
d
im
p
r
o
v
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
.
B
ased
o
n
th
ese
p
r
i
n
cip
les,
we
d
e
s
cr
ib
e
in
th
e
n
ex
t
s
ec
tio
n
h
o
w
PC
A
an
d
PS
O
ar
e
in
teg
r
ated
in
to
o
u
r
co
m
p
lete
f
ac
e
r
ec
o
g
n
itio
n
p
ip
elin
e
.
3
.
6
.
Co
m
pu
t
a
t
io
na
l
c
o
m
plex
it
y
T
h
e
d
ec
is
io
n
to
u
s
e
PC
A
an
d
PS
O
in
o
u
r
p
ap
er
is
b
ased
o
n
th
e
co
m
p
u
tatio
n
al
ef
f
icien
c
y
,
esp
ec
ially
in
co
m
p
ar
is
o
n
with
th
e
cu
r
r
e
n
t
d
ee
p
lear
n
in
g
m
eth
o
d
s
.
PC
A
is
a
lin
ea
r
s
u
b
s
p
ac
e
lear
n
in
g
m
eth
o
d
,
with
th
e
b
u
lk
o
f
t
h
e
co
m
p
u
tatio
n
b
e
in
g
th
e
co
m
p
u
tatio
n
o
f
th
e
co
v
ar
ia
n
ce
m
atr
i
x
o
f
th
e
d
ata
an
d
th
e
eig
en
d
ec
o
m
p
o
s
itio
n
.
T
o
c
o
m
p
u
te
th
e
eig
en
d
ec
o
m
p
o
s
itio
n
o
f
a
d
ataset
o
f
s
ize
N
im
ag
es,
o
f
d
i
m
en
s
io
n
ality
D,
th
e
co
m
p
lex
ity
o
f
th
e
co
m
p
u
tatio
n
is
o
f
o
r
d
er
O(
D
3
)
in
t
h
e
e
x
tr
em
e
ca
s
e,
wh
ich
is
co
m
p
u
ted
d
ir
ec
tly
,
b
u
t
in
p
r
ac
tice,
th
e
s
o
-
ca
lled
s
m
all
co
v
ar
ia
n
ce
tr
ick
ca
n
r
ed
u
ce
th
e
co
m
p
u
tatio
n
to
O(
N
3
)
wh
en
N
is
s
m
all
co
m
p
ar
ed
to
D,
wh
ich
h
ap
p
e
n
s
in
o
u
r
OR
L
d
ataset.
Af
ter
c
o
m
p
u
tin
g
th
e
ei
g
en
f
ac
es,
it
o
n
ly
tak
es
O(
D
k
)
to
p
r
o
ject
ea
ch
n
ew
im
ag
e
in
t
o
th
e
r
ed
u
ce
d
s
u
b
s
p
ac
e,
wh
ich
is
m
u
ch
less
th
an
p
er
-
s
am
p
le
p
r
o
ce
s
s
in
g
tim
e,
o
n
c
e
th
e
eig
en
f
ac
es h
a
v
e
b
ee
n
ca
lcu
lated
.
B
y
co
n
tr
ast,
d
ee
p
lear
n
in
g
m
o
d
els
lik
e
C
NNs
h
av
e
L
-
H
-
W
-
C
-
K
-
F
co
m
p
u
tatio
n
ally
co
m
p
lex
ity
o
f
O(
L
H
W
C
K
2
F),
wh
er
e
L
is
th
e
n
u
m
b
er
o
f
lay
er
s
,
H
W
d
im
en
s
io
n
s
o
f
f
ea
tu
r
e
m
ap
s
,
C
is
th
e
n
u
m
b
er
o
f
in
p
u
t
ch
an
n
els,
K
is
th
e
k
er
n
e
l
s
ize,
an
d
F
is
th
e
n
u
m
b
er
o
f
f
ilter
s
p
er
lay
er
.
T
h
is
co
m
p
le
x
ity
s
ca
les
g
r
ea
tly
with
n
etwo
r
k
d
e
p
th
an
d
im
ag
e
r
eso
lu
tio
n
,
an
d
m
u
s
t c
o
n
s
u
m
e
lar
g
e
am
o
u
n
ts
o
f
GPU
ac
ce
ler
atio
n
an
d
m
em
o
r
y
b
an
d
wid
th
,
w
h
ich
ar
e
u
s
u
ally
n
o
t a
v
ailab
le
in
e
d
g
e
o
r
em
b
e
d
d
ed
d
ev
ices.
Usi
n
g
PC
A
t
o
r
ed
u
ce
d
i
m
en
s
io
n
s
an
d
PS
O
to
s
e
lec
t
f
e
at
u
r
es
ad
a
p
ti
v
e
ly
,
o
u
r
f
r
am
ew
o
r
k
h
as
a
co
m
p
u
tati
o
n
a
ll
y
li
g
h
twe
i
g
h
t
f
o
o
t
p
r
in
t,
an
d
h
i
g
h
r
e
c
o
g
n
i
ti
o
n
a
cc
u
r
ac
y
.
T
h
e
PC
A
+
PS
O
te
c
h
n
iq
u
e
h
as
b
e
en
u
s
e
d
to
d
o
wn
s
am
p
le
t
h
e
o
r
i
g
i
n
al
1
0
,
3
0
4
-
d
i
m
e
n
s
i
o
n
al
im
ag
e
v
e
ct
o
r
s
to
a
s
m
all
er
s
iz
e
o
f
4
4
f
ea
t
u
r
es
h
e
n
c
e
s
las
h
i
n
g
th
e
s
to
r
a
g
e
as
we
ll
as
p
e
r
-
s
am
p
le
class
if
ica
ti
o
n
co
m
p
l
ex
it
y
t
o
O
(
k
N
t
r
ai
n
)
n
e
ar
est
n
ei
g
h
b
o
r
cl
ass
i
f
ic
ati
o
n
.
T
h
is
s
h
o
ws
a
d
ef
i
n
i
te
c
o
m
p
u
t
ati
o
n
al
b
e
n
ef
it
o
v
e
r
d
ee
p
n
etw
o
r
k
s
,
o
f
f
er
in
g
a
h
i
g
h
-
s
p
e
ed
,
r
es
o
u
r
ce
-
e
f
f
ici
e
n
t
a
n
d
s
ca
l
ab
le
s
o
lu
ti
o
n
t
h
a
t
ca
n
b
e
d
ep
l
o
y
e
d
t
o
lo
w
-
p
o
we
r
a
n
d
co
n
s
tr
a
in
e
d
s
y
s
t
em
s
i
n
r
e
al
-
ti
m
e
.
T
h
e
PS
O
im
p
lem
en
tatio
n
in
o
u
r
s
tu
d
y
is
m
o
d
if
ied
to
b
in
ar
y
f
ea
tu
r
e
s
elec
tio
n
,
u
n
lik
e
th
e
t
r
ad
itio
n
al
co
n
tin
u
o
u
s
im
p
lem
en
tatio
n
o
f
PS
O.
T
h
e
co
n
v
e
n
tio
n
al
PS
O
u
p
d
ate
eq
u
atio
n
is
u
s
ed
t
o
f
ir
s
t
u
p
d
ate
th
e
p
ar
ticle
v
elo
cities:
+
1
=
⋅
+
1
⋅
1
⋅
(
−
)
+
2
⋅
2
⋅
(
−
)
(
3
)
w
h
er
e
w=
0
.
9
is
th
e
in
er
tia
we
ig
h
t,
1
=
2
.
0
an
d
2
=
2
.
0
ar
e
th
e
co
g
n
itiv
e
a
n
d
s
o
cial
p
ar
am
eter
s
an
d
1
,
2
U(
0
,
1
)
a
r
e
r
an
d
o
m
f
ac
to
r
s
.
I
n
o
r
d
er
t
o
co
n
v
er
t
co
n
tin
u
o
u
s
v
elo
cities
to
b
in
ar
y
s
elec
tio
n
s
o
f
f
ea
tu
r
es,
we
u
s
e
a
s
ig
m
o
id
tr
an
s
f
er
f
u
n
ctio
n
:
(
+
1
)
=
1
1
+
−
+
1
(
4
)
T
h
e
n
ew
lo
ca
tio
n
o
f
ea
ch
p
a
r
t
icle
is
th
en
d
eter
m
in
ed
p
r
o
b
ab
ilis
tically
:
th
e
f
ea
tu
r
e
is
ch
o
s
en
(
(
+
1
)
=
1
)
wh
en
a
r
a
n
d
o
m
n
u
m
b
er
i
n
[
0
-
1
]
is
s
m
aller
th
an
(
(
+
1
)
:
o
th
er
wis
e,
th
is
f
ea
tu
r
e
is
s
et
to
0
.
Al
s
o
,
to
m
ak
e
s
u
r
e
th
at
th
e
p
ar
ticle
v
el
o
cities d
o
n
o
t g
et
to
o
lar
g
e
an
d
d
o
m
in
ate
th
e
s
ig
m
o
id
,
a
cla
m
p
in
g
r
a
n
g
e
o
f
-
6
,
6
]
is
u
s
ed
s
o
th
at
th
e
o
u
tp
u
t o
f
th
e
s
ig
m
o
id
is
in
a
u
s
ef
u
l
r
an
g
e
o
f
p
r
o
b
ab
ilit
ies.
T
h
is
is
an
ef
f
ec
tiv
e
m
eth
o
d
t
o
co
n
v
er
t
th
e
co
n
tin
u
o
u
s
PS
O
in
to
a
b
in
ar
y
PS
O
th
at
ca
n
b
e
u
s
ed
in
f
ea
tu
r
e
s
elec
tio
n
an
d
en
s
u
r
e
t
h
e
co
n
v
er
g
en
ce
s
tab
ilit
y
,
as
w
ell
as,
ad
ap
tiv
e
s
elec
tio
n
o
f
th
e
m
o
s
t
in
f
o
r
m
ativ
e
p
r
in
cip
al
co
m
p
o
n
en
ts
.
Ou
r
ap
p
r
o
ac
h
o
f
d
ef
in
in
g
th
e
clam
p
i
n
g
an
d
b
in
a
r
y
m
ap
p
in
g
b
y
ex
p
licitly
d
ef
in
in
g
th
e
clam
p
in
g
,
an
d
o
f
f
er
in
g
ac
cu
r
a
te
co
n
tr
o
l
o
v
er
f
ea
tu
r
e
s
u
b
s
et
ex
p
lo
r
atio
n
wh
ich
is
v
ital
to
a
ttain
in
g
s
u
ch
h
ig
h
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
(
~9
9
.
9
9
%
)
as seen
in
o
u
r
e
x
p
er
im
e
n
ts
.
4.
M
E
T
H
O
DO
L
O
G
Y
I
n
th
is
s
tu
d
y
,
we
p
r
o
p
o
s
e
a
f
ac
e
r
ec
o
g
n
itio
n
a
n
d
f
ea
tu
r
e
o
p
t
im
izatio
n
s
y
s
tem
th
at
co
m
b
in
es
PC
A
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
a
n
d
PSO
f
o
r
o
p
tim
al
f
ea
tu
r
e
s
elec
tio
n
.
T
h
e
s
y
s
tem
is
ev
alu
ated
o
n
th
e
b
en
ch
m
a
r
k
OR
L
f
ac
e
d
atab
ase,
wh
ich
co
n
tain
s
g
r
ay
s
ca
le
im
ag
es
o
f
4
0
in
d
iv
i
d
u
als
ca
p
tu
r
ed
u
n
d
er
v
ar
y
in
g
l
ig
h
tin
g
co
n
d
itio
n
s
,
f
ac
ial
ex
p
r
ess
io
n
s
,
an
d
p
o
s
e
o
r
ien
tatio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
2
1
3
4
-
2
1
5
7
2140
T
h
e
m
eth
o
d
o
lo
g
y
f
o
llo
ws
a
s
tr
u
ctu
r
ed
s
eq
u
en
ce
o
f
s
ix
k
ey
s
tag
es,
as
illu
s
tr
ated
in
Fi
g
u
r
e
5
.
I
n
s
u
m
m
ar
y
,
all
im
ag
es
ar
e
r
esized
to
a
f
ix
ed
r
eso
l
u
tio
n
,
th
e
ir
co
n
tr
ast
is
en
h
an
ce
d
,
an
d
p
ix
el
in
ten
s
ities
ar
e
n
o
r
m
alize
d
b
ef
o
r
e
b
ein
g
v
ec
to
r
ized
in
to
co
l
u
m
n
f
o
r
m
f
o
r
s
u
b
s
eq
u
en
t p
r
o
ce
s
s
in
g
.
Fig
u
r
e
5
.
Sch
em
atic
d
iag
r
am
o
f
th
e
PC
A
–
PS
O
f
ac
e
r
ec
o
g
n
it
io
n
ar
ch
itectu
r
e
4
.
1
.
P
re
pro
ce
s
s
ing
a
nd
da
t
a
s
et
prepa
ra
t
io
n
T
h
e
OR
L
f
ac
e
d
atab
ase
co
n
tai
n
s
4
0
0
g
r
ay
s
ca
le
f
ac
ial
im
ag
es
d
is
tr
ib
u
ted
ac
r
o
s
s
4
0
s
u
b
d
ir
ec
to
r
ies
(
s
1
to
s
4
0
)
,
ea
ch
co
r
r
esp
o
n
d
in
g
to
a
s
in
g
le
s
u
b
ject
an
d
co
n
t
ain
in
g
1
0
im
a
g
es
in
(
.
p
g
m
)
f
o
r
m
at.
T
o
e
n
s
u
r
e
u
n
if
o
r
m
ity
in
th
e
i
n
p
u
t
d
ata,
e
ac
h
im
ag
e
u
n
d
er
g
o
es th
e
f
o
llo
win
g
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
:
a.
R
esizin
g
:
All
im
ag
es
ar
e
r
esized
to
a
co
m
m
o
n
r
eso
lu
tio
n
o
f
1
1
2
×
9
2
p
ix
els
to
s
ta
n
d
ar
d
ize
i
n
p
u
t
d
im
en
s
io
n
s
ac
r
o
s
s
th
e
d
ataset.
b.
His
to
g
r
am
eq
u
aliza
tio
n
:
Glo
b
al
h
is
to
g
r
am
eq
u
aliza
tio
n
i
s
ap
p
lied
to
en
h
an
ce
co
n
tr
a
s
t
an
d
m
itig
ate
v
ar
iatio
n
s
ca
u
s
ed
b
y
lig
h
tin
g
o
r
b
ac
k
g
r
o
u
n
d
in
c
o
n
s
is
ten
cies.
c.
No
r
m
aliza
tio
n
:
Pix
el
in
ten
s
iti
es
ar
e
s
ca
led
to
th
e
[
0
,
1
]
r
an
g
e
to
en
s
u
r
e
co
n
s
is
ten
t
b
r
ig
h
tn
e
s
s
an
d
f
ac
ilit
ate
n
u
m
er
ical
s
tab
ilit
y
in
s
u
b
s
eq
u
en
t step
s
.
d.
Flatten
in
g
:
E
ac
h
p
r
ep
r
o
ce
s
s
ed
im
ag
e
is
r
esh
ap
ed
in
to
a
c
o
lu
m
n
v
ec
to
r
an
d
s
to
r
ed
in
a
d
ata
m
atr
ix
o
f
d
im
en
s
io
n
s
(
Nu
m
p
ix
els ×
Nu
m
im
ag
es),
wh
ich
s
er
v
es a
s
in
p
u
t to
th
e
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
ip
elin
e.
E
ac
h
im
ag
e
v
ec
t
o
r
is
ass
o
ciate
d
with
a
co
r
r
esp
o
n
d
i
n
g
lab
el
i
n
d
icatin
g
th
e
s
u
b
ject
id
en
tity
(
f
r
o
m
1
to
4
0
)
.
4
.
2
.
T
ra
ini
ng
–
v
a
lid
a
t
io
n
-
t
est
pa
rt
it
io
nin
g
T
h
e
d
ataset
is
d
iv
id
ed
in
to
t
r
ain
in
g
,
v
alid
atio
n
,
a
n
d
test
in
g
s
u
b
s
ets
u
s
in
g
s
tr
atif
ied
s
a
m
p
lin
g
p
er
s
u
b
ject
to
en
s
u
r
e
p
r
o
p
o
r
tio
n
al
r
ep
r
esen
tatio
n
.
T
h
e
v
alid
atio
n
s
u
b
s
et,
r
ep
r
esen
tin
g
a
f
i
x
ed
p
r
o
p
o
r
tio
n
o
f
th
e
tr
ain
in
g
d
ata,
is
s
p
ec
if
ically
r
eser
v
ed
f
o
r
PS
O
f
itn
ess
ev
alu
atio
n
.
T
h
is
c
o
n
f
ig
u
r
atio
n
p
r
e
v
en
ts
o
v
er
f
itti
n
g
b
y
en
s
u
r
in
g
t
h
at
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
d
o
es
n
o
t
d
ir
ec
tly
ass
ess
p
er
f
o
r
m
an
ce
o
n
th
e
d
ata
u
s
ed
f
o
r
PC
A
m
o
d
el
co
n
s
tr
u
ctio
n
.
Stra
tific
atio
n
h
e
lp
s
p
r
eser
v
e
class
d
iv
e
r
s
ity
i
n
b
o
t
h
p
a
r
titi
o
n
s
an
d
p
r
e
v
en
t
s
class
im
b
alan
ce
,
wh
ich
co
u
ld
o
th
er
wis
e
d
is
to
r
t
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
an
d
b
ias th
e
ev
alu
atio
n
r
esu
lts
.
4
.
3
.
P
CA
-
ba
s
ed
f
ea
t
ure
e
x
t
r
a
ct
io
n
PC
A
i
s
ap
p
lied
to
r
ed
u
ce
t
h
e
d
im
en
s
io
n
ality
o
f
h
ig
h
-
r
eso
lu
tio
n
f
ac
ial
im
ag
es
wh
ile
p
r
e
s
er
v
in
g
th
e
m
o
s
t
s
ig
n
if
ican
t
v
a
r
ian
ce
.
T
h
e
s
elec
tio
n
o
f
p
r
in
ci
p
al
co
m
p
o
n
en
ts
ca
p
tu
r
in
g
9
5
%
o
f
t
h
e
to
tal
v
ar
ian
c
e
is
m
o
tiv
ated
b
y
th
e
n
ee
d
to
b
alan
ce
d
im
en
s
io
n
ality
r
e
d
u
ctio
n
with
th
e
p
r
eser
v
atio
n
o
f
d
is
cr
im
in
ativ
e
in
f
o
r
m
atio
n
—
a
cr
iter
io
n
f
r
eq
u
en
tly
ad
o
p
ted
in
PC
A
-
b
ased
f
ac
e
r
ec
o
g
n
itio
n
s
tu
d
ies.
T
h
e
p
r
o
ce
s
s
in
v
o
lv
es th
e
f
o
llo
win
g
s
tep
s
:
a.
Me
an
ce
n
ter
in
g
:
T
h
e
m
ea
n
f
a
ce
v
ec
to
r
is
co
m
p
u
ted
ac
r
o
s
s
all
tr
ain
in
g
s
am
p
les
an
d
s
u
b
tr
ac
ted
f
r
o
m
ea
c
h
im
ag
e
v
ec
to
r
.
T
h
is
en
s
u
r
es th
a
t th
e
d
ata
is
ce
n
ter
ed
a
r
o
u
n
d
th
e
o
r
ig
in
in
f
ea
tu
r
e
s
p
ac
e
.
b.
C
o
v
ar
ian
ce
m
atr
ix
ca
lcu
latio
n
:
R
ath
er
th
an
c
o
m
p
u
tin
g
t
h
e
f
u
ll
co
v
ar
ia
n
ce
m
atr
ix
o
f
(
s
ize
Nu
m
p
i
x
els
×
Nu
m
p
ix
els),
a
m
o
r
e
ef
f
icie
n
t
ap
p
r
o
a
ch
is
u
s
ed
b
y
c
o
m
p
u
tin
g
a
r
e
d
u
ce
d
co
v
ar
ia
n
ce
m
atr
ix
o
f
(
s
ize
n
u
m
_
s
am
p
les
×
n
u
m
_
s
am
p
les
)
,
lev
er
ag
i
n
g
th
e
tr
a
n
s
p
o
s
e
tr
ic
k
.
c.
E
ig
en
f
ac
e
c
o
m
p
u
tatio
n
:
E
ig
e
n
v
alu
e
d
ec
o
m
p
o
s
itio
n
is
a
p
p
lied
to
th
e
r
e
d
u
ce
d
co
v
ar
ian
ce
m
atr
ix
.
T
h
e
r
esu
ltin
g
eig
en
v
ec
to
r
s
ar
e
p
r
o
jecte
d
b
ac
k
in
to
th
e
o
r
ig
in
al
im
ag
e
s
p
ac
e
to
f
o
r
m
eig
en
f
ac
es,
wh
ich
s
er
v
e
as
p
r
in
cip
al
co
m
p
o
n
e
n
ts
ca
p
tu
r
in
g
th
e
m
o
s
t d
is
cr
im
in
ativ
e
v
ar
i
atio
n
s
in
th
e
d
ata.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Lig
h
tw
eig
h
t fa
ce
r
ec
o
g
n
itio
n
b
a
s
ed
o
n
P
C
A
co
u
p
le
d
w
ith
P
S
O
-
b
a
s
ed
fea
tu
r
e
…
(
C
h
a
ima
a
K
h
o
u
d
d
a
)
2141
d.
Dim
en
s
io
n
ality
r
ed
u
ctio
n
:
T
h
e
eig
en
f
ac
es
a
r
e
r
an
k
e
d
ac
co
r
d
in
g
to
t
h
eir
co
r
r
esp
o
n
d
in
g
eig
en
v
alu
es
in
d
escen
d
in
g
o
r
d
er
.
A
cu
m
u
lativ
e
en
er
g
y
p
lo
t
is
u
s
ed
to
d
eter
m
in
e
th
e
n
u
m
b
er
o
f
co
m
p
o
n
en
ts
r
eq
u
ir
ed
to
r
etain
9
5
%
o
f
th
e
to
tal
v
a
r
ian
ce
.
B
o
th
tr
ain
in
g
an
d
test
im
a
g
es
ar
e
th
en
p
r
o
jecte
d
o
n
to
t
h
is
r
ed
u
ce
d
PC
A
s
u
b
s
p
ac
e
to
g
en
e
r
ate
co
m
p
ac
t
f
ea
tu
r
e
v
ec
to
r
s
.
4
.
4
.
F
e
a
t
ure
s
elec
t
io
n us
ing
P
SO
W
h
ile
PC
A
ef
f
ec
tiv
ely
r
ed
u
c
es
th
e
d
im
en
s
io
n
ality
o
f
th
e
f
ea
tu
r
e
s
p
ac
e,
n
o
t
all
r
etain
e
d
p
r
in
ci
p
al
co
m
p
o
n
en
ts
co
n
tr
ib
u
te
eq
u
all
y
to
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
.
T
h
e
ch
o
ice
o
f
PS
O
is
m
o
tiv
ated
b
y
its
ab
ilit
y
to
ef
f
icien
tly
ex
p
lo
r
e
h
ig
h
-
d
im
en
s
io
n
al
d
is
cr
ete
s
p
ac
es
with
o
u
t
r
eq
u
ir
in
g
g
r
ad
ien
t
in
f
o
r
m
atio
n
,
m
ak
in
g
it
p
ar
ticu
lar
ly
s
u
itab
le
f
o
r
b
in
a
r
y
f
ea
tu
r
e
s
elec
tio
n
p
r
o
b
lem
s
s
u
ch
as
o
u
r
s
.
T
o
f
u
r
th
er
r
ef
i
n
e
th
e
s
elec
tio
n
o
f
r
elev
an
t f
ea
tu
r
es,
PSO
is
em
p
l
o
y
ed
as a
g
l
o
b
al
s
ea
r
ch
alg
o
r
it
h
m
.
T
o
im
p
lem
en
t PSO in
o
u
r
ca
s
e,
we
f
o
llo
wed
th
ese
m
ai
n
s
tep
s
:
a.
Par
ticle
en
co
d
in
g
:
E
ac
h
p
ar
tic
le
r
ep
r
esen
ts
a
b
i
n
ar
y
v
ec
to
r
,
wh
er
e
a
v
al
u
e
o
f
1
in
d
icate
s
t
h
e
in
clu
s
io
n
o
f
a
s
p
ec
if
ic
PC
A
co
m
p
o
n
en
t a
n
d
0
d
en
o
tes ex
cl
u
s
io
n
.
b.
Swar
m
in
itializatio
n
:
A
p
o
p
u
l
atio
n
o
f
p
a
r
ticles
is
r
an
d
o
m
ly
in
itialized
.
T
h
e
s
ea
r
c
h
d
y
n
am
i
cs
ar
e
g
o
v
er
n
e
d
b
y
th
e
in
e
r
tia
weig
h
t
w,
th
e
co
g
n
itiv
e
f
ac
to
r
c1
,
an
d
th
e
s
o
cial
f
ac
to
r
c2
,
wh
ich
co
n
t
r
o
l
th
e
b
ala
n
ce
b
etwe
en
ex
p
lo
r
atio
n
a
n
d
ex
p
lo
itatio
n
.
T
h
e
PS
O
h
y
p
er
p
ar
a
m
eter
s
(
p
o
p
u
latio
n
s
ize,
in
er
t
ia
weig
h
t,
an
d
ac
ce
ler
atio
n
co
ef
f
icien
ts
)
ar
e
s
et
ac
co
r
d
in
g
to
em
p
ir
ical
g
u
id
elin
es
f
r
o
m
th
e
liter
atu
r
e
to
en
s
u
r
e
an
ef
f
ec
tiv
e
tr
ad
e
-
o
f
f
b
etwe
en
s
ea
r
ch
s
p
ac
e
ex
p
lo
r
atio
n
a
n
d
c
o
n
v
er
g
e
n
ce
s
p
ee
d
.
c.
Fit
n
ess
ev
alu
atio
n
:
T
h
e
f
itn
ess
o
f
ea
ch
p
ar
ticle
is
ev
alu
ate
d
u
s
in
g
r
ec
o
g
n
itio
n
ac
c
u
r
ac
y
o
n
a
v
alid
atio
n
s
u
b
s
et,
r
ep
r
esen
tin
g
a
f
ix
ed
p
r
o
p
o
r
tio
n
o
f
th
e
tr
ain
i
n
g
d
ata,
t
o
av
o
id
o
v
e
r
f
itti
n
g
.
T
h
is
is
co
m
p
u
ted
u
s
in
g
a
k
-
n
ea
r
est
n
eig
h
b
o
r
(
k
-
NN)
class
if
ier
with
E
u
clid
ea
n
d
is
tan
ce
,
ch
o
s
en
f
o
r
its
s
im
p
licity
,
in
ter
p
r
etab
ilit
y
,
an
d
s
tr
o
n
g
p
e
r
f
o
r
m
an
ce
in
lo
w
-
d
im
en
s
io
n
al
PC
A
s
u
b
s
p
ac
e
s
,
an
d
r
estricte
d
to
th
e
co
m
p
o
n
en
ts
s
elec
ted
b
y
th
e
p
ar
ticle'
s
b
in
ar
y
m
ask
.
d.
Up
d
ate
r
u
les
an
d
co
n
v
e
r
g
en
ce
:
Du
r
in
g
ea
ch
iter
atio
n
,
p
ar
ticl
es
u
p
d
ate
th
eir
v
elo
cities
an
d
p
o
s
itio
n
s
b
ased
o
n
th
eir
p
e
r
s
o
n
al
b
est
an
d
th
e
g
lo
b
al
b
est
s
o
lu
tio
n
s
.
A
co
n
s
tr
ain
t
is
en
f
o
r
ce
d
to
av
o
id
o
v
er
ly
m
in
im
al
s
u
b
s
ets.
T
h
e
alg
o
r
ith
m
co
n
tin
u
es
u
n
til
co
n
v
er
g
en
ce
is
ac
h
i
ev
ed
,
at
wh
ich
p
o
in
t
th
e
g
lo
b
al
b
est
p
ar
ticle
m
ask
is
s
elec
ted
as th
e
o
p
tim
al
f
ea
tu
r
e
s
u
b
s
et
f
o
r
class
if
icati
o
n
.
4
.
5
.
Cla
s
s
if
ica
t
io
n
us
ing
E
uc
lid
ea
n
dis
t
a
nce
On
ce
th
e
o
p
tim
al
s
u
b
s
et
o
f
p
r
i
n
cip
al
co
m
p
o
n
en
ts
h
as b
ee
n
s
elec
ted
,
b
o
th
tr
ain
i
n
g
an
d
test
im
ag
es a
r
e
p
r
o
jecte
d
o
n
to
th
e
r
esu
ltin
g
P
C
A
s
u
b
s
p
ac
e.
A
n
ea
r
est
-
n
eig
h
b
o
r
class
if
ier
b
ased
o
n
E
u
clid
ea
n
d
is
tan
ce
is
th
en
u
s
ed
f
o
r
f
in
al
r
ec
o
g
n
itio
n
.
F
o
r
ea
ch
test
im
ag
e,
its
f
ea
tu
r
e
v
ec
to
r
is
co
m
p
ar
ed
to
all
tr
ain
in
g
v
ec
to
r
s
b
y
co
m
p
u
tin
g
th
e
E
u
clid
ea
n
d
is
tan
ce
.
T
h
e
test
im
ag
e
is
ass
ig
n
ed
th
e
lab
el
o
f
th
e
tr
ain
in
g
s
am
p
le
with
th
e
s
m
allest d
is
tan
ce
.
4
.
6
.
E
v
a
lua
t
i
o
n
m
et
rics a
nd
perf
o
rm
a
nce
a
na
ly
s
is
T
h
e
p
r
ed
ictiv
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
ass
es
s
ed
u
s
in
g
th
e
f
o
llo
win
g
ev
alu
atio
n
m
etr
ics:
a.
Acc
u
r
ac
y
:
T
h
e
p
r
o
p
o
r
tio
n
o
f
c
o
r
r
ec
tly
class
if
ied
test
im
ag
es o
u
t o
f
t
h
e
to
tal.
b.
C
o
n
f
u
s
io
n
m
atr
ix
:
A
m
atr
ix
(
o
f
ten
v
is
u
alize
d
as
a
h
ea
tm
ap
)
th
at
illu
s
tr
ates
th
e
d
is
tr
ib
u
ti
o
n
o
f
p
r
e
d
icted
lab
els v
er
s
u
s
tr
u
e
lab
els ac
r
o
s
s
all
class
es.
c.
Pre
cisi
o
n
,
r
ec
all
,
an
d
F1
-
s
co
r
e:
C
las
s
-
s
p
ec
if
ic
p
er
f
o
r
m
an
ce
m
ea
s
u
r
es
th
at
h
ig
h
lig
h
t
h
o
w
well
th
e
m
o
d
el
d
is
tin
g
u
is
h
es
b
etwe
en
d
if
f
er
en
t
id
en
titi
es.
T
h
ese
ar
e
es
p
ec
ially
r
elev
an
t
in
m
u
lticlas
s
cla
s
s
if
icatio
n
p
r
o
b
lem
s
s
u
ch
as f
ac
e
r
ec
o
g
n
i
tio
n
.
I
n
ad
d
itio
n
,
a
PS
O
co
n
v
er
g
en
ce
cu
r
v
e
is
p
lo
tted
to
d
e
m
o
n
s
tr
ate
th
e
s
tab
ilit
y
an
d
ef
f
icien
cy
o
f
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
o
v
er
s
u
c
ce
s
s
iv
e
iter
atio
n
s
.
Ov
er
all,
th
e
p
r
o
p
o
s
ed
p
i
p
elin
e
in
teg
r
ates
d
im
en
s
io
n
ality
r
ed
u
ctio
n
(
PC
A)
,
ev
o
lu
tio
n
ar
y
o
p
tim
izatio
n
(
PS
O)
,
an
d
a
s
im
p
le
y
et
ef
f
ec
tiv
e
d
is
tan
ce
-
b
ased
class
if
ier
(
E
u
clid
ea
n
n
ea
r
est
n
eig
h
b
o
r
)
.
PC
A
en
s
u
r
es
ef
f
icien
t
ex
tr
ac
tio
n
o
f
k
ey
f
ac
ial
f
ea
tu
r
es,
wh
ile
PS
O
o
p
tim
ally
s
elec
ts
th
e
m
o
s
t
d
is
cr
im
i
n
ativ
e
s
u
b
s
et.
T
h
e
n
ea
r
est
-
n
eig
h
b
o
r
r
u
le
p
r
o
v
id
es
a
lig
h
tweig
h
t a
n
d
in
ter
p
r
etab
l
e
class
if
icatio
n
m
eth
o
d
.
E
x
p
er
im
en
tal
r
esu
lts
o
n
th
e
OR
L
f
ac
e
d
ataset
s
h
o
w
h
ig
h
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
,
ac
h
iev
in
g
o
v
er
9
9
.
9
9
%
ac
c
u
r
ac
y
with
r
ap
id
PS
O
co
n
v
er
g
en
ce
.
T
h
ese
o
u
t
co
m
es
v
alid
ate
th
e
e
f
f
ec
tiv
en
ess
,
ef
f
icien
cy
,
an
d
p
r
ac
tical
ap
p
licab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
a
p
p
r
o
ac
h
f
o
r
r
ea
l
-
w
o
r
ld
f
ac
e
r
ec
o
g
n
itio
n
s
ce
n
a
r
io
s
.
5.
E
XP
E
R
I
M
E
N
T
R
E
SU
L
T
S
I
n
th
is
s
tu
d
y
,
a
s
er
ies
o
f
ex
p
e
r
im
en
ts
wer
e
co
n
d
u
cted
u
s
in
g
th
e
OR
L
f
ac
e
d
atab
ase
to
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
f
ac
e
r
ec
o
g
n
itio
n
p
i
p
elin
e,
wh
ic
h
in
teg
r
ates
PC
A
f
o
r
f
e
atu
r
e
e
x
tr
ac
tio
n
an
d
PSO
f
o
r
f
ea
tu
r
e
s
elec
tio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
1
6
,
No
.
3
,
J
u
n
e
20
2
6
:
2
1
3
4
-
2
1
5
7
2142
T
h
e
OR
L
d
ataset
c
o
n
s
is
ts
o
f
4
0
0
g
r
ay
s
ca
le
f
ac
ial
im
ag
es
f
r
o
m
4
0
in
d
iv
id
u
als,
with
e
ac
h
s
u
b
ject
r
ep
r
esen
ted
b
y
1
0
v
ar
iatio
n
s
u
n
d
er
d
if
f
er
e
n
t
lig
h
tin
g
c
o
n
d
it
io
n
s
,
f
ac
ial
ex
p
r
ess
io
n
s
,
an
d
s
lig
h
t
p
o
s
e
ch
an
g
es.
Fig
u
r
e
6
p
r
o
v
id
es
ex
am
p
les
o
f
s
u
b
jects
f
r
o
m
th
e
d
ataset,
illu
s
tr
atin
g
th
e
v
is
u
al
d
iv
er
s
ity
ac
r
o
s
s
in
d
iv
id
u
als.
All
im
ag
es
wer
e
p
r
ep
r
o
ce
s
s
ed
an
d
s
tan
d
ar
d
ized
to
a
r
eso
l
u
tio
n
o
f
1
1
2
×
9
2
p
ix
els,
r
esu
l
tin
g
in
1
0
,
3
0
4
r
aw
p
ix
el
f
ea
tu
r
es p
e
r
im
ag
e.
Fig
u
r
e
6
.
Dif
f
e
r
en
t su
b
jects f
r
o
m
OR
L
d
atab
ase
Prio
r
to
f
ea
tu
r
e
ex
tr
ac
tio
n
,
ea
ch
f
ac
ial
im
a
g
e
u
n
d
er
g
o
es
a
s
tan
d
ar
d
ized
p
r
e
p
r
o
ce
s
s
in
g
p
r
o
ce
d
u
r
e
t
o
en
h
an
ce
q
u
ality
an
d
en
s
u
r
e
c
o
n
s
is
ten
cy
ac
r
o
s
s
th
e
d
ataset.
First,
all
im
ag
es a
r
e
r
esized
to
a
u
n
if
o
r
m
r
eso
lu
tio
n
o
f
1
1
2
×
9
2
p
ix
els.
His
to
g
r
am
eq
u
aliza
tio
n
is
th
en
a
p
p
lied
to
m
itig
ate
lig
h
tin
g
in
c
o
n
s
is
ten
cies
an
d
im
p
r
o
v
e
co
n
tr
ast.
Fin
ally
,
p
ix
el
in
te
n
s
ities
ar
e
n
o
r
m
alize
d
to
th
e
[
0
,
1
]
r
an
g
e
to
estab
lis
h
a
co
n
s
is
ten
t
in
ten
s
ity
s
ca
le
ac
r
o
s
s
all
im
ag
es.
E
ac
h
p
r
ep
r
o
ce
s
s
ed
im
ag
e
is
th
en
f
latten
e
d
in
to
a
co
lu
m
n
v
ec
t
o
r
wh
ile
p
r
eser
v
in
g
its
p
ix
el
-
wis
e
s
tr
u
ctu
r
e
f
o
r
s
u
b
s
eq
u
e
n
t f
ea
tu
r
e
ex
tr
ac
tio
n
s
tag
es.
T
o
ass
ess
th
e
im
p
ac
t
o
f
tr
ain
i
n
g
d
ata
av
ailab
ilit
y
o
n
r
ec
o
g
n
itio
n
p
e
r
f
o
r
m
an
ce
,
th
r
ee
tr
ai
n
-
test
s
p
lit
co
n
f
ig
u
r
atio
n
s
wer
e
ev
al
u
ated
:
9
0
%
tr
ain
in
g
with
1
0
%
test
in
g
,
8
0
%/2
0
%,
an
d
7
0
%/3
0
%.
Stra
tifie
d
s
am
p
lin
g
was
em
p
lo
y
ed
to
en
s
u
r
e
th
at
ea
ch
s
u
b
ject
was
p
r
o
p
o
r
tio
n
ally
r
ep
r
esen
ted
ac
r
o
s
s
all
s
p
lits
.
T
h
is
ap
p
r
o
ac
h
p
r
eser
v
ed
class
d
is
tr
ib
u
tio
n
s
a
n
d
m
ain
tain
ed
a
b
alan
ce
d
d
at
aset
f
o
r
b
o
th
tr
ain
in
g
an
d
test
in
g
p
h
ases
.
W
ith
in
ea
ch
co
n
f
ig
u
r
atio
n
,
th
e
tr
ain
i
n
g
an
d
test
in
g
im
a
g
es
wer
e
r
an
d
o
m
ly
s
elec
ted
to
in
t
r
o
d
u
c
e
v
ar
iab
ilit
y
a
n
d
to
s
im
u
late
r
ea
lis
tic
an
d
d
iv
er
s
e
r
ec
o
g
n
itio
n
s
ce
n
a
r
io
s
.
I
n
th
e
tr
ain
in
g
p
h
ase,
PC
A
i
s
ap
p
lied
t
o
c
o
m
p
u
te
a
s
et
o
f
o
r
t
h
o
n
o
r
m
al
b
asis
v
ec
to
r
s
co
m
m
o
n
l
y
r
ef
er
r
ed
to
as
eig
en
f
ac
es
wh
ich
r
ep
r
esen
t
th
e
d
ir
ec
tio
n
s
o
f
g
r
ea
test
v
ar
ian
ce
with
in
th
e
f
ac
ial
im
ag
e
d
ata.
T
h
e
m
ea
n
f
ac
e
is
f
ir
s
t
co
m
p
u
ted
b
y
av
er
ag
in
g
th
e
p
ix
el
in
ten
s
iti
es
ac
r
o
s
s
all
tr
ain
in
g
im
ag
es.
E
ac
h
tr
ain
in
g
im
ag
e
is
th
en
ce
n
ter
ed
b
y
s
u
b
tr
ac
tin
g
th
is
m
ea
n
f
ac
e,
r
esu
ltin
g
in
a
ze
r
o
-
m
ea
n
d
ata
m
atr
ix
.
T
o
r
ed
u
ce
co
m
p
u
tatio
n
al
co
m
p
lex
ity
,
th
e
s
o
-
ca
lled
"c
o
v
ar
ia
n
ce
m
atr
ix
t
r
ick
"
is
em
p
lo
y
ed
:
in
s
tead
o
f
ca
lcu
latin
g
th
e
co
v
a
r
ian
ce
m
atr
ix
in
th
e
h
ig
h
-
d
im
en
s
io
n
al
p
ix
el
s
p
ac
e,
it
is
c
o
m
p
u
ted
in
th
e
l
o
wer
-
d
im
en
s
io
n
al
im
ag
e
s
p
ac
e.
T
h
e
eig
en
v
ec
to
r
s
o
f
th
is
s
m
aller
co
v
ar
ian
ce
m
atr
i
x
ar
e
th
en
b
ac
k
-
p
r
o
jecte
d
in
to
th
e
o
r
ig
in
al
p
i
x
el
s
p
ac
e
to
f
o
r
m
t
h
e
ac
tu
al
eig
e
n
f
ac
es
u
s
ed
in
s
u
b
s
eq
u
en
t
f
ea
tu
r
e
e
x
tr
ac
tio
n
.
Fig
u
r
e
7
illu
s
tr
ates
th
e
m
ea
n
f
ac
e
a
n
d
th
e
to
p
eig
en
f
ac
es
o
b
tain
e
d
th
r
o
u
g
h
th
is
p
r
o
ce
s
s
,
h
ig
h
lig
h
tin
g
t
h
e
m
o
s
t
s
ig
n
if
ican
t
v
ar
iatio
n
s
in
th
e
d
ataset.
E
ac
h
eig
en
v
ec
to
r
is
ass
o
ciate
d
with
an
eig
en
v
al
u
e
th
at
i
n
d
icate
s
th
e
am
o
u
n
t
o
f
v
ar
ian
ce
ex
p
lain
ed
b
y
th
e
c
o
r
r
esp
o
n
d
in
g
eig
e
n
f
a
ce
.
T
o
d
eter
m
in
e
t
h
e
d
im
e
n
s
io
n
ality
r
eq
u
ir
ed
to
ca
p
t
u
r
e
m
o
s
t
o
f
th
e
v
ar
iab
ilit
y
in
th
e
tr
ain
in
g
s
et,
th
e
cu
m
u
l
ativ
e
s
u
m
o
f
eig
en
v
alu
es
is
co
m
p
u
ted
a
n
d
p
lo
tted
.
B
ased
o
n
th
is
cu
m
u
lativ
e
v
ar
ian
ce
cu
r
v
e,
it
was
o
b
s
er
v
ed
th
at
ap
p
r
o
x
im
ately
th
e
to
p
1
0
0
p
r
in
cip
al
co
m
p
o
n
en
ts
ac
co
u
n
t
f
o
r
o
v
er
9
5
%
o
f
th
e
to
tal
v
ar
ian
ce
.
C
o
n
s
eq
u
en
tly
,
b
o
th
tr
ain
in
g
an
d
t
esti
n
g
im
ag
es
ar
e
p
r
o
jecte
d
o
n
to
th
is
r
ed
u
ce
d
s
u
b
s
p
ac
e,
r
esu
ltin
g
in
c
o
m
p
ac
t
f
ea
tu
r
e
v
ec
t
o
r
s
th
at
s
er
v
e
as i
n
p
u
t to
s
u
b
s
eq
u
en
t p
r
o
ce
s
s
in
g
s
tag
es.
Fig
u
r
e
8
illu
s
tr
ates
th
e
d
is
tr
ib
u
tio
n
o
f
eig
en
v
alu
es
f
o
r
th
e
f
i
r
s
t
1
0
0
p
r
i
n
cip
al
co
m
p
o
n
e
n
ts
.
Sin
ce
PC
A
eig
en
v
alu
es
r
e
p
r
esen
t
t
h
e
am
o
u
n
t
o
f
v
ar
ian
ce
ca
p
tu
r
ed
b
y
e
ac
h
co
m
p
o
n
e
n
t,
t
h
ey
ar
e
a
r
r
a
n
g
ed
in
d
escen
d
in
g
o
r
d
er
.
As
s
h
o
wn
in
th
e
g
r
ap
h
,
th
e
f
ir
s
t
f
ew
co
m
p
o
n
en
ts
,
p
ar
ticu
lar
ly
th
e
to
p
1
0
,
ac
c
o
u
n
t
f
o
r
th
e
lar
g
est
p
r
o
p
o
r
tio
n
o
f
v
ar
ian
ce
,
wh
ile
th
e
eig
en
v
alu
es
o
f
th
e
r
em
ain
i
n
g
co
m
p
o
n
en
ts
ap
p
r
o
ac
h
ze
r
o
.
T
h
is
v
is
u
aliza
tio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
Lig
h
tw
eig
h
t fa
ce
r
ec
o
g
n
itio
n
b
a
s
ed
o
n
P
C
A
co
u
p
le
d
w
ith
P
S
O
-
b
a
s
ed
fea
tu
r
e
…
(
C
h
a
ima
a
K
h
o
u
d
d
a
)
2143
clea
r
ly
h
ig
h
lig
h
ts
h
o
w
th
e
v
ar
ian
ce
is
co
n
ce
n
tr
ated
i
n
th
e
i
n
itial
co
m
p
o
n
e
n
ts
an
d
b
ec
o
m
es
n
eg
lig
ib
le
in
th
e
later
o
n
es,
h
elp
in
g
u
s
d
eter
m
in
e
an
o
p
tim
al
n
u
m
b
er
o
f
co
m
p
o
n
en
ts
to
r
etain
.
Fig
u
r
e
7
.
Me
an
f
ac
es a
n
d
t
o
p
9
eig
en
f
ac
es
Fig
u
r
e
8
.
PC
A
eig
en
v
alu
e
d
is
tr
ib
u
tio
n
Fo
llo
win
g
th
e
eig
en
v
alu
e
d
is
tr
ib
u
tio
n
,
it
b
ec
o
m
es
ess
en
tial
to
ass
es
s
h
o
w
well
th
e
r
ed
u
ce
d
f
ea
tu
r
es
r
ep
r
esen
t
b
o
th
tr
ain
in
g
an
d
te
s
tin
g
s
ets.
T
h
i
s
is
ac
h
iev
ed
th
r
o
u
g
h
a
v
is
u
al
in
s
p
ec
tio
n
o
f
th
eir
d
is
tr
ib
u
tio
n
in
th
e
PC
A
-
tr
an
s
f
o
r
m
ed
s
p
ac
e.
Fig
u
r
e
9
is
p
ar
ticu
lar
l
y
s
ig
n
i
f
ican
t
as
it
v
is
u
alize
s
th
e
d
is
tr
ib
u
tio
n
o
f
f
ea
tu
r
e
s
p
ac
e
b
et
wee
n
th
e
tr
ain
in
g
a
n
d
test
in
g
d
ata
wh
e
n
p
r
o
jecte
d
o
n
to
th
e
f
ir
s
t
two
p
r
in
cip
al
c
o
m
p
o
n
en
ts
.
I
f
t
h
ese
d
is
tr
ib
u
tio
n
s
wer
e
f
ar
ap
ar
t,
it
c
o
u
ld
in
d
icate
p
o
o
r
g
en
e
r
aliza
tio
n
an
d
r
ed
u
ce
d
r
o
b
u
s
tn
ess
o
f
th
e
m
o
d
el.
H
o
wev
er
,
th
e
o
v
er
lap
o
b
s
er
v
ed
b
etwe
en
t
h
e
two
s
e
ts
s
u
g
g
ests
th
at
th
e
ex
tr
ac
ted
f
ea
tu
r
es
ar
e
s
im
ilar
ly
d
is
tr
ib
u
ted
.
T
h
is
s
p
atial
p
r
o
x
im
ity
im
p
lies
th
at
th
e
m
o
d
el
g
en
er
alize
s
well
an
d
is
c
ap
ab
le
o
f
ac
h
iev
in
g
h
i
g
h
p
r
e
d
ictio
n
ac
cu
r
ac
y
o
n
b
o
th
tr
ain
in
g
an
d
u
n
s
ee
n
test
d
ata,
as su
p
p
o
r
ted
b
y
th
e
q
u
a
n
titativ
e
r
esu
lts
.
T
h
is
n
u
m
b
er
is
u
s
ed
to
g
r
ap
h
ically
d
ep
ict
th
e
f
ea
tu
r
e
s
p
ac
e,
h
av
in
g
u
n
d
e
r
g
o
n
e
th
e
P
C
A
o
f
th
e
o
r
ig
in
al
h
ig
h
d
im
en
s
io
n
al
f
ac
i
al
d
ata
to
th
e
f
ir
s
t
two
p
r
in
cip
al
co
m
p
o
n
e
n
ts
.
T
h
e
tr
ain
in
g
a
n
d
test
s
am
p
le
ar
e
p
lo
tted
to
d
e
m
o
n
s
tr
ate
h
o
w
t
h
ey
ar
e
d
is
tr
ib
u
ted
h
i
g
h
lig
h
tin
g
th
e
d
is
tr
ib
u
tio
n
an
d
co
n
ce
n
tr
a
tio
n
o
f
d
ata
p
o
in
ts
.
I
t
is
o
b
v
io
u
s
th
at
s
o
m
e
o
v
er
la
p
is
p
r
esen
t,
b
u
t
m
o
s
t
clu
s
ter
s
o
f
p
o
in
ts
th
at
r
ep
r
esen
t
v
ar
io
u
s
s
u
b
jects
ar
e
f
ar
ap
ar
t,
wh
ich
m
ea
n
s
th
at
th
e
P
C
A
tr
an
s
f
o
r
m
atio
n
r
ep
r
esen
ts
a
s
ig
n
if
ican
t
p
ar
t
o
f
th
e
v
a
r
ia
n
ce
in
th
e
d
ata
an
d
r
etain
s
d
is
cr
im
in
ativ
e
d
ata.
T
h
is
v
is
u
aliza
tio
n
is
u
s
ef
u
l
to
esti
m
ate
q
u
alitativ
ely
th
e
s
ep
a
r
ab
ilit
y
o
f
class
es
an
d
s
u
itab
ilit
y
o
f
th
e
r
ed
u
ce
d
f
ea
tu
r
e
s
p
ac
e
to
d
o
wn
s
tr
ea
m
PS
O
-
b
ased
s
elec
tio
n
an
d
class
if
icatio
n
.
Mo
r
eo
v
er
,
th
e
n
u
m
b
er
also
g
iv
es
s
o
m
e
im
p
licit
in
f
o
r
m
atio
n
o
n
th
e
in
tr
icac
y
o
f
th
e
f
ea
tu
r
e
d
is
tr
ib
u
tio
n
.
T
h
e
d
im
en
s
io
n
ali
ty
r
ed
u
ctio
n
i
n
b
o
th
m
aj
o
r
co
m
p
o
n
e
n
ts
in
d
icate
s
th
at
PC
A
d
o
es
n
o
t
d
ec
r
ea
s
e
d
im
en
s
io
n
ality
b
u
t
r
at
h
er
t
h
e
f
ea
tu
r
e
s
p
ac
e
is
d
elica
te
an
d
v
a
r
ied
an
d
th
is
is
wh
y
a
s
elec
tio
n
m
ec
h
a
n
is
m
s
u
c
h
as
PS
O
is
n
ee
d
ed
to
s
elec
t
t
h
e
m
o
s
t
in
f
o
r
m
ativ
e
s
u
b
s
et
o
f
f
ea
tu
r
es.
T
h
e
r
elativ
ely
s
p
r
ea
d
o
u
t
test
d
ata
Evaluation Warning : The document was created with Spire.PDF for Python.