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.
14
,
No
.
6
,
Dec
em
b
er
20
24
,
p
p
.
7
1
4
5
~
7
1
5
7
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
14
i
6
.
pp
7
1
4
5
-
7
1
5
7
7145
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Efficien
t
sm
a
rt
di
stribute
d f
a
ce
ide
ntif
ica
tion
using
t
he
M
ix
M
a
x
Sim
deci
sio
n f
unction
Sa
y
ed
M
o
ha
m
m
a
d Ahm
a
di,
Ro
uh
o
lla
h Dia
na
t
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
E
n
g
i
n
e
e
r
i
n
g
a
n
d
I
n
f
o
r
m
a
t
i
o
n
Te
c
h
n
o
l
o
g
y
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
si
t
y
o
f
Q
o
m
,
Q
o
m,
I
r
a
n
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
r
1
4
,
2
0
2
4
R
ev
is
ed
J
u
l 2
2
,
2
0
2
4
Acc
ep
ted
Au
g
6
,
2
0
2
4
Re
c
o
g
n
izin
g
a
larg
e
n
u
m
b
e
r
o
f
p
e
o
p
le
is
a
c
o
m
m
o
n
c
h
a
ll
e
n
g
e
in
fa
c
e
id
e
n
ti
fica
ti
o
n
a
p
p
li
c
a
ti
o
n
s,
i
n
v
o
l
v
in
g
d
e
c
re
a
se
d
a
c
c
u
ra
c
y
,
in
c
re
a
se
d
m
e
m
o
ry
a
n
d
ti
m
e
c
o
m
p
lex
it
ies
.
T
o
a
d
d
re
s
s
th
e
se
issu
e
s,
th
is
stu
d
y
in
tr
o
d
u
c
e
s
a
th
re
e
-
m
o
d
u
le
a
p
p
ro
a
c
h
:
“
to
il
e
rs
,
”
“
a
ffi
n
it
y
-
m
e
ter
,
”
a
n
d
“
d
e
c
isio
n
-
m
a
k
e
r
.
”
Un
li
k
e
th
e
ra
n
d
o
m
d
istri
b
u
t
io
n
m
e
th
o
d
s
u
se
d
in
p
re
v
io
u
s
so
lu
ti
o
n
s,
t
h
is
m
e
th
o
d
e
m
p
lo
y
s
c
lu
ste
ri
n
g
t
o
d
istri
b
u
t
e
th
e
p
ro
b
lem
in
to
s
u
b
n
e
two
r
k
s
c
a
ll
e
d
“
to
il
e
rs.
”
Th
e
t
o
il
e
r’s
m
o
d
u
le
c
a
lcu
late
s
th
e
li
k
e
li
h
o
o
d
o
f
tes
t
d
a
ta
b
e
lo
n
g
i
n
g
to
e
a
c
h
c
las
s
o
f
e
a
c
h
to
il
e
r,
u
si
n
g
th
e
las
t
lay
e
r
o
u
tp
u
ts
o
f
d
e
e
p
lea
rn
in
g
m
o
d
e
ls.
M
e
a
n
w
h
il
e
,
th
e
a
ffi
n
it
y
-
m
e
ter
m
o
d
u
le
d
e
term
in
e
s
t
h
e
sim
il
a
rit
y
b
e
twe
e
n
th
e
tes
t
d
a
ta
a
n
d
t
h
e
a
v
e
ra
g
e
o
f
e
a
c
h
c
las
s,
e
m
p
l
o
y
i
n
g
a
sim
il
a
rit
y
m
e
a
su
re
.
Th
e
d
e
c
isi
o
n
-
m
a
k
e
r
m
o
d
u
le
c
o
m
b
i
n
e
s
th
e
re
p
o
rts
fro
m
th
e
p
re
v
io
u
s
tw
o
m
o
d
u
les
a
n
d
se
lec
ts
th
e
fi
n
a
l
c
las
s,
u
t
il
izin
g
a
m
i
x
o
f
th
e
m
a
x
-
m
a
x
c
rit
e
rio
n
a
n
d
t
h
e
sim
il
a
rit
y
c
rit
e
rio
n
.
Th
e
p
r
o
p
o
se
d
m
e
th
o
d
o
u
t
p
e
rfo
rm
s
e
x
isti
n
g
s
o
lu
ti
o
n
s,
a
c
h
iev
i
n
g
im
p
ro
v
e
d
re
c
a
ll
,
p
re
c
isio
n
,
a
n
d
F
1
-
sc
o
re
.
It
e
ffe
c
ti
v
e
ly
a
d
d
re
ss
e
s
m
e
m
o
ry
,
sp
e
e
d
,
a
n
d
a
c
c
u
ra
c
y
issu
e
s
in
fa
c
e
id
e
n
ti
fica
ti
o
n
,
s
u
rp
a
ss
in
g
b
o
t
h
n
o
-
d
istri
b
u
ti
o
n
a
n
d
ra
n
d
o
m
m
e
th
o
d
s
o
n
G
li
n
t3
6
0
K,
VG
G
F
a
c
e
2
,
a
n
d
M
S
-
Ce
leb
-
1
M
d
a
tas
e
ts.
Ov
e
ra
ll
,
t
h
i
s
m
e
th
o
d
o
ffe
rs
a
m
o
re
e
fficie
n
t
a
n
d
a
c
c
u
ra
te
a
p
p
ro
a
c
h
b
y
d
istri
b
u
ti
n
g
t
h
e
p
ro
b
lem
in
to
s
u
b
n
e
two
rk
s
,
d
e
m
o
n
stra
ti
n
g
su
p
e
rio
r
p
e
rfo
rm
a
n
c
e
a
n
d
sc
a
lab
il
it
y
fo
r
larg
e
-
sc
a
le fac
e
re
c
o
g
n
it
i
o
n
a
p
p
l
i
c
a
ti
o
n
s
.
K
ey
w
o
r
d
s
:
C
lu
s
ter
in
g
Dee
p
lear
n
in
g
Dis
tr
ib
u
ted
lear
n
in
g
Face
id
en
tific
atio
n
Facial
r
ec
o
g
n
itio
n
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
:
Say
ed
Mo
h
am
m
a
d
Ah
m
a
d
i
D
e
p
a
r
t
m
e
n
t
o
f
C
o
m
p
u
t
e
r
E
n
g
i
n
e
e
r
i
n
g
a
n
d
I
n
f
o
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
s
i
t
y
o
f
Q
o
m
Qo
m
,
I
r
an
E
m
ail: sm
.
ah
m
ad
i@
s
tu
.
q
o
m
.
a
c.
ir
1.
I
NT
RO
D
UCT
I
O
N
Facial
r
ec
o
g
n
itio
n
h
as
b
ee
n
an
ac
tiv
e
r
esear
ch
ar
ea
in
m
ac
h
in
e
v
is
io
n
[
1
]
,
[
2
]
,
en
co
m
p
ass
in
g
v
ar
io
u
s
s
u
b
f
ield
s
s
u
ch
as
f
ac
e
d
etec
tio
n
[
3
]
,
[
4
]
,
alig
n
m
en
t
[
5
]
,
an
t
i
-
s
p
o
o
f
in
g
[
6
]
,
[
7
]
,
an
d
r
ec
o
g
n
itio
n
[
8
]
,
[
9
]
.
Face
r
ec
o
g
n
itio
n
in
v
o
lv
es
b
o
th
f
a
ce
id
en
tific
atio
n
an
d
v
e
r
if
ica
tio
n
,
wh
er
e
th
e
f
o
r
m
er
in
v
o
l
v
es
id
en
tify
in
g
th
e
p
er
s
o
n
in
a
n
in
p
u
t
im
ag
e
a
n
d
th
e
latter
in
v
o
lv
es
v
er
if
y
in
g
wh
eth
er
two
in
p
u
t
im
a
g
es
b
elo
n
g
t
o
th
e
s
am
e
p
er
s
o
n
.
Face
r
ec
o
g
n
itio
n
h
as
r
ec
en
tly
g
ain
ed
s
ig
n
if
ica
n
t
atten
tio
n
as
a
m
ea
n
s
o
f
b
io
m
et
r
ic
au
th
en
ticatio
n
d
u
e
to
th
e
c
o
r
o
n
av
ir
u
s
d
is
ea
s
e
2
0
1
9
(
C
OVI
D
-
19
)
p
an
d
em
ic
[
1
0
]
,
[
1
1
]
.
Dee
p
lear
n
in
g
tec
h
n
iq
u
es,
esp
ec
ially
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN
s
)
,
h
av
e
b
ee
n
at
th
e
f
o
r
ef
r
o
n
t
o
f
f
ac
e
r
ec
o
g
n
itio
n
m
eth
o
d
s
f
o
r
o
v
er
a
d
ec
ad
e
[
1
2
]
.
Ma
n
y
ap
p
licatio
n
s
o
f
m
ac
h
i
n
e
v
is
io
n
a
n
d
f
ac
e
p
r
o
ce
s
s
in
g
s
h
ar
e
c
o
m
m
o
n
ch
allen
g
es
with
f
ac
e
r
ec
o
g
n
itio
n
,
in
clu
d
i
n
g
v
a
r
y
in
g
lig
h
t
co
n
d
itio
n
s
s
u
ch
as
b
r
i
g
h
tn
ess
an
d
co
n
tr
ast,
d
if
f
er
e
n
t
p
o
s
es,
o
r
v
ar
io
u
s
ap
p
ea
r
an
ce
s
o
f
th
e
s
am
e
o
b
j
ec
t
in
m
u
ltip
le
im
ag
es.
Ad
d
i
tio
n
ally
,
f
ac
e
r
ec
o
g
n
itio
n
is
p
lag
u
ed
b
y
s
p
ec
if
ic
ch
allen
g
es,
s
u
ch
as
d
i
f
f
er
en
ce
s
in
f
ac
ial
ex
p
r
ess
io
n
s
,
m
ak
eu
p
,
an
d
a
g
e
i
n
v
a
r
io
u
s
im
a
g
es
o
f
a
s
in
g
le
p
er
s
o
n
.
Mo
r
eo
v
er
,
th
er
e
ar
e
tim
es
wh
en
in
tr
a
-
class
d
is
tan
ce
s
ex
ce
ed
in
ter
-
class
d
is
tan
ce
s
[
1
3
]
.
Un
lik
e
o
b
jects
in
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.
14
,
No
.
6
,
Dec
em
b
e
r
20
24
:
7
1
4
5
-
7
1
5
7
7146
o
b
ject
d
etec
tio
n
th
at
h
av
e
a
c
lear
d
is
tin
ctio
n
f
r
o
m
ea
ch
o
th
er
,
f
ac
es
in
f
ac
e
r
ec
o
g
n
itio
n
h
av
e
a
v
er
y
s
im
ilar
s
tr
u
ctu
r
e
to
ea
ch
o
th
e
r
.
T
h
e
m
o
r
e
in
d
iv
id
u
als
(
class
es)
i
n
v
o
lv
e
d
in
f
ac
e
r
ec
o
g
n
itio
n
,
th
e
m
o
r
e
g
en
e
r
al
p
r
o
b
lem
s
ar
is
e
d
u
e
t
o
th
ese
co
m
m
o
n
c
h
allen
g
es
an
d
s
p
ec
if
ic
f
ac
ial
f
ea
tu
r
es:
i
)
m
e
m
o
r
y
co
n
s
u
m
p
tio
n
p
r
o
b
lem
,
ii
)
in
cr
ea
s
ed
tem
p
o
r
a
l c
o
m
p
lex
ity
,
a
n
d
iii
)
in
c
r
ea
s
ed
ac
cu
r
ac
y
d
r
o
p
s
F
o
r
a
l
a
r
g
e
-
s
ca
l
e
f
a
c
e
i
d
e
n
t
i
f
ica
t
i
o
n
a
p
p
l
i
c
at
i
o
n
,
t
h
e
w
ei
g
h
t
m
a
t
r
i
x
r
e
q
u
i
r
e
s
m
o
r
e
m
e
m
o
r
y
t
h
a
n
m
o
s
t
c
u
r
r
e
n
t
g
r
a
p
h
i
c
s
p
r
o
c
es
s
i
n
g
u
n
i
t
s
(
G
PU
s
)
c
a
n
h
a
n
d
le
b
e
c
a
u
s
e
t
h
e
c
l
as
s
i
f
i
e
r
la
y
e
r
i
n
t
h
e
m
o
d
e
l
is
f
u
l
l
y
c
o
n
n
e
c
t
e
d
t
o
t
h
e
p
r
e
v
i
o
u
s
la
y
e
r
.
M
a
n
y
p
ar
a
m
e
t
e
r
s
i
n
t
h
i
s
p
a
r
t
a
l
o
n
e
i
m
p
o
s
e
a
h
e
a
v
y
c
o
m
p
u
t
a
ti
o
n
a
l
b
u
r
d
e
n
o
n
t
h
e
n
e
tw
o
r
k
t
o
c
o
m
p
u
t
e
t
h
e
l
o
s
s
f
u
n
c
t
i
o
n
an
d
u
p
d
a
t
e
t
h
e
w
e
i
g
h
t
s
.
A
d
d
i
t
io
n
a
l
l
y
,
t
h
e
l
a
r
g
e
n
u
m
b
e
r
o
f
n
e
u
r
o
n
s
i
n
t
h
e
o
u
t
p
u
t
l
a
y
e
r
r
e
s
u
l
ts
i
n
s
i
g
n
i
f
i
ca
n
t
o
v
e
r
f
i
t
t
i
n
g
o
n
t
h
e
n
e
t
w
o
r
k
.
D
u
e
t
o
t
h
e
s
p
e
c
i
f
i
c
n
a
t
u
r
e
o
f
t
h
e
f
a
c
e
s
t
r
u
c
t
u
r
e
,
t
h
e
m
o
r
e
c
l
a
s
s
es
t
h
e
r
e
a
r
e
,
t
h
e
m
o
r
e
p
r
o
n
o
u
n
c
e
d
t
h
e
d
r
o
p
i
n
a
c
c
u
r
a
c
y
.
W
h
e
n
al
l
d
at
a
a
r
e
t
r
a
i
n
e
d
a
n
d
i
d
e
n
t
i
f
i
e
d
t
o
g
e
t
h
e
r
,
t
h
e
r
e
i
s
a
h
i
g
h
d
e
g
r
e
e
o
f
d
a
t
a
d
i
s
p
e
r
s
i
o
n
,
m
a
k
i
n
g
i
t
m
o
r
e
c
h
a
l
l
e
n
g
i
n
g
t
o
d
i
s
t
i
n
g
u
is
h
d
i
f
f
i
c
u
l
t
,
i
.
e
.
,
s
i
m
i
l
a
r
d
a
t
a
.
A
d
o
p
t
i
n
g
a
d
i
s
t
r
i
b
u
t
e
d
a
p
p
r
o
ac
h
a
u
t
o
m
a
t
i
c
a
ll
y
s
o
l
v
es
t
h
e
p
r
o
b
l
e
m
s
o
f
m
e
m
o
r
y
a
n
d
t
i
m
e
.
T
h
i
s
s
t
u
d
y
a
i
m
s
t
o
p
r
o
p
o
s
e
a
d
i
s
t
r
i
b
u
te
d
a
p
p
r
o
a
c
h
t
h
a
t
d
o
e
s
n
o
t
s
a
c
r
i
f
ic
e
a
c
c
u
r
a
c
y
a
n
d
,
i
f
p
o
s
s
i
b
le
,
e
v
e
n
i
m
p
r
o
v
e
s
i
t
.
T
wo
g
en
er
al
ap
p
r
o
ac
h
es
h
av
e
b
ee
n
tak
en
to
ad
d
r
ess
th
e
ch
allen
g
es
p
o
s
ed
b
y
f
ac
e
r
ec
o
g
n
itio
n
with
lar
g
e
-
s
ca
le
d
atasets
.
T
h
e
f
ir
s
t
ap
p
r
o
ac
h
in
v
o
l
v
es
th
e
d
ev
el
o
p
m
en
t
o
f
n
ew
ar
ch
itectu
r
es
o
r
lo
s
s
f
u
n
ctio
n
s
,
wh
ich
h
av
e
led
to
th
e
cr
ea
tio
n
o
f
in
n
o
v
ativ
e
tech
n
iq
u
es
s
u
c
h
as
L
-
So
f
tm
ax
[
1
4
]
,
A
-
So
f
tm
ax
[
1
5
]
,
No
r
m
Face
[
1
3
]
,
C
o
s
Face
[
1
6
]
,
an
d
Ar
cFac
e
[
1
7
]
,
SF
ac
e
[
1
8
]
,
Gh
o
s
tFace
[
1
9
]
.
Alth
o
u
g
h
th
ese
e
f
f
o
r
ts
h
av
e
y
ield
ed
p
r
o
m
is
in
g
r
esu
lts
,
th
ey
ar
e
o
f
ten
n
o
t
p
r
ac
tical
f
o
r
v
er
y
h
ig
h
n
u
m
b
e
r
s
o
f
class
es
an
d
m
u
s
t
b
e
co
m
b
i
n
ed
wit
h
ad
d
itio
n
al
tech
n
iq
u
es.
T
h
e
s
ec
o
n
d
ap
p
r
o
ac
h
in
v
o
lv
es
tech
n
i
q
u
es
s
u
ch
as
s
o
f
tm
ax
d
is
s
ec
ti
o
n
[
2
0
]
,
ac
tiv
e
class
s
elec
tio
n
[
2
1
]
,
o
r
s
am
p
lin
g
[
2
2
]
,
w
h
ich
h
a
v
e
b
ee
n
em
p
lo
y
ed
to
m
ak
e
it
p
o
s
s
ib
le
to
i
m
p
lem
en
t
th
e
f
ir
s
t
ap
p
r
o
ac
h
f
o
r
v
er
y
h
ig
h
n
u
m
b
e
r
s
o
f
class
e
s
.
Alth
o
u
g
h
th
ese
t
ec
h
n
iq
u
es m
ay
r
esu
lt in
a
s
lig
h
t d
r
o
p
in
ac
cu
r
ac
y
,
th
ey
h
a
v
e
p
r
o
v
en
u
s
ef
u
l
in
m
ak
in
g
it
f
ea
s
ib
le
t
o
u
s
e
n
e
w
ar
ch
itectu
r
es
o
r
lo
s
s
f
u
n
ct
io
n
s
f
o
r
lar
g
e
-
s
ca
le
d
atasets
.
T
h
e
s
ec
o
n
d
ap
p
r
o
a
ch
f
o
cu
s
es
o
n
d
ec
o
m
p
o
s
in
g
th
e
p
r
o
b
lem
in
t
o
s
m
aller
s
u
b
p
r
o
b
lem
s
th
r
o
u
g
h
d
iv
id
e
-
an
d
-
co
n
q
u
er
tech
n
iq
u
e
s
.
E
r
r
o
r
-
co
r
r
ec
tin
g
o
u
tp
u
t m
o
d
el
(
E
C
OC
)
[
2
3
]
,
lab
el
m
ap
p
in
g
[
2
4
]
,
in
d
ep
e
n
d
en
t
s
o
f
tm
ax
m
o
d
el
(
I
SM)
[
2
5
]
,
m
u
lti
-
co
g
n
itio
n
s
o
f
tm
ax
m
o
d
el
(
MC
S
M)
[
2
6
]
ar
e
s
o
m
e
m
eth
o
d
s
o
f
th
is
ap
p
r
o
ac
h
.
T
h
e
E
C
OC
tech
n
iq
u
e
p
ar
titi
o
n
s
an
N
-
class
n
etwo
r
k
in
to
m
u
ltip
le
p
ar
allel
two
-
class
n
et
wo
r
k
s
.
E
ac
h
b
in
ar
y
n
etwo
r
k
ca
n
f
u
n
ctio
n
as
a
clu
s
ter
in
g
m
eth
o
d
to
ca
teg
o
r
iz
ed
d
ata
s
am
p
les.
T
h
e
o
r
ig
in
a
l
class
lab
el
ca
n
b
e
o
b
tain
ed
b
y
co
m
b
in
in
g
th
e
r
e
s
u
lts
o
f
b
in
ar
y
class
n
etwo
r
k
s
.
Ho
wev
er
,
th
is
ap
p
r
o
ac
h
r
eq
u
ir
es
estab
lis
h
in
g
a
co
r
r
esp
o
n
d
en
ce
b
etwe
en
b
in
a
r
y
an
d
o
r
ig
in
al
n
etwo
r
k
la
b
els.
Fu
r
th
er
m
o
r
e,
d
eter
m
in
in
g
t
h
e
n
u
m
b
e
r
o
f
b
in
ar
y
-
class
n
etwo
r
k
s
an
d
en
s
u
r
in
g
in
d
ep
en
d
en
ce
am
o
n
g
r
o
ws
an
d
co
lu
m
n
s
ar
e
cr
u
cial
to
m
in
im
izin
g
er
r
o
r
s
.
Alth
o
u
g
h
t
h
is
m
eth
o
d
h
as
n
o
t
b
ee
n
wid
ely
u
s
ed
f
o
r
f
ac
e
r
e
co
g
n
itio
n
,
it
co
u
ld
p
r
o
v
id
e
v
a
lu
ab
le
in
s
ig
h
ts
f
o
r
r
elate
d
s
tu
d
ies.
L
a
b
el
m
ap
p
in
g
is
a
v
ar
ian
t
o
f
th
e
E
C
OC
ap
p
r
o
ac
h
th
at
d
iv
id
es
th
e
N
-
clas
s
n
etwo
r
k
in
to
lar
g
er
s
u
b
n
etwo
r
k
s
in
s
tead
o
f
b
in
ar
y
-
class
s
u
b
n
etwo
r
k
s
.
T
h
e
s
u
b
n
e
two
r
k
s
ca
n
h
av
e
eith
er
e
q
u
al
o
r
m
ix
e
d
s
izes,
an
d
th
e
lab
el
m
ap
p
in
g
ap
p
r
o
ac
h
ca
n
b
e
ap
p
lied
ac
co
r
d
in
g
ly
.
W
h
ile
th
is
m
eth
o
d
h
as n
o
t
b
ee
n
d
ir
ec
tly
u
s
ed
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
,
it c
o
u
ld
in
s
p
ir
e
f
u
tu
r
e
r
esear
ch
in
th
e
ar
ea
.
T
h
e
I
SM
m
eth
o
d
[
2
5
]
,
wh
ich
in
v
o
lv
es
r
an
d
o
m
ly
d
is
tr
ib
u
ti
n
g
class
es
in
to
s
u
b
n
etwo
r
k
s
,
is
o
n
e
s
u
ch
ex
am
p
le
o
f
th
e
s
ec
o
n
d
ap
p
r
o
a
ch
.
Ho
wev
er
,
th
e
r
an
d
o
m
d
is
tr
ib
u
tio
n
o
f
class
es
ca
n
lead
to
er
r
o
r
s
,
d
etailed
in
th
e
p
r
o
p
o
s
ed
m
eth
o
d
s
ec
tio
n
.
o
n
e
s
o
lu
tio
n
to
clea
r
th
is
t
y
p
e
o
f
er
r
o
r
is
MCS
M.
MCS
M
p
r
o
v
id
es
a
d
escr
ip
tio
n
o
f
th
e
en
tire
ar
ch
itectu
r
e
u
s
ed
in
[
2
5
]
as
o
n
e
o
f
its
co
m
p
o
n
e
n
ts
,
ter
m
ed
a
co
g
n
itio
n
u
n
it.
I
t
th
en
c
o
n
cu
r
r
en
tly
tr
ain
s
s
ev
er
al
co
g
n
itio
n
u
n
its
.
All
co
g
n
itio
n
u
n
its
r
ec
ei
v
e
tes
t
d
ata.
Usi
n
g
a
v
o
tin
g
m
ec
h
an
is
m
,
th
e
p
r
ed
icted
class
is
d
eter
m
in
ed
b
y
th
e
class
with
th
e
m
o
s
t
v
o
tes
am
o
n
g
m
u
ltip
le
co
g
n
itiv
e
u
n
its
.
MCS
M
was
p
r
o
p
o
s
ed
to
ad
d
r
ess
th
e
p
r
o
b
lem
with
[
2
5
]
an
d
m
itig
ate
th
e
im
p
ac
t
o
f
e
r
r
o
r
s
ca
u
s
ed
b
y
a
n
im
p
r
o
p
e
r
r
a
n
d
o
m
d
is
tr
ib
u
tio
n
.
MCS
M
o
u
tp
er
f
o
r
m
s
I
SM
m
et
h
o
d
s
in
ter
m
s
o
f
ac
c
u
r
ac
y
.
H
o
wev
er
,
it
co
n
s
u
m
es
m
o
r
e
m
em
o
r
y
an
d
r
eq
u
ir
es
m
o
r
e
tim
e
f
o
r
tr
ain
in
g
an
d
p
r
e
d
ictio
n
.
T
h
is
s
tu
d
y
p
r
o
p
o
s
es
a
b
ette
r
s
o
lu
tio
n
f
o
r
th
e
m
en
tio
n
e
d
ty
p
e
o
f
er
r
o
r
.
I
t
u
tili
ze
s
in
tellig
en
t
d
is
tr
ib
u
tio
n
o
f
class
es
to
s
u
b
n
etwo
r
k
s
th
r
o
u
g
h
clu
s
ter
in
g
.
T
h
e
clu
s
ter
in
g
alg
o
r
ith
m
g
r
o
u
p
s
s
im
ilar
clas
s
es
to
g
eth
er
,
lead
i
n
g
to
a
r
e
d
u
ctio
n
in
d
ata
d
is
p
er
s
io
n
with
in
ea
ch
clu
s
ter
.
W
ith
in
cr
ea
s
ed
f
o
c
u
s
o
n
s
im
ilar
d
ata,
ea
ch
s
u
b
n
etwo
r
k
is
ex
p
ec
ted
to
im
p
r
o
v
e
ac
cu
r
ac
y
.
Ou
r
p
r
o
p
o
s
ed
m
eth
o
d
in
clu
d
es
th
r
ee
m
o
d
u
les:
to
iler
s
,
af
f
in
ity
-
m
eter
,
an
d
f
in
al
d
ec
i
s
io
n
m
ak
er
.
T
h
e
to
iler
’
s
m
o
d
u
le
u
s
es
s
u
p
er
v
is
ed
n
eu
r
al
n
etwo
r
k
s
to
class
if
y
f
ac
ial
im
ag
es
in
to
N
/m
clas
s
e
s
,
wh
er
e
N
is
th
e
to
tal
n
u
m
b
er
o
f
class
es
an
d
m
is
th
e
n
u
m
b
er
o
f
to
iler
s
.
T
h
e
af
f
in
ity
-
m
eter
ca
lcu
lates
th
e
s
im
ilar
ity
b
etwe
en
th
e
in
p
u
t
im
ag
e
an
d
th
e
av
er
a
g
e
f
ea
tu
r
e
s
o
f
ea
ch
class
f
o
r
ea
ch
to
iler
.
T
h
e
f
in
al
d
ec
is
io
n
m
ak
er
s
elec
ts
th
e
b
est
clas
s
f
r
o
m
ea
ch
to
iler
a
n
d
d
eter
m
i
n
es
th
e
f
in
al
o
u
tp
u
t.
Ou
r
ap
p
r
o
ac
h
o
f
f
er
s
s
ev
e
r
al
ad
v
a
n
tag
es
o
v
er
tr
ad
itio
n
al
f
ac
ial
r
ec
o
g
n
itio
n
m
eth
o
d
s
.
B
y
d
is
tr
ib
u
tin
g
th
e
wo
r
k
lo
ad
am
o
n
g
m
u
ltip
le
t
o
iler
s
,
we
ca
n
s
ig
n
i
f
ican
tly
r
ed
u
ce
th
e
tim
e
an
d
m
e
m
o
r
y
c
o
m
p
le
x
ities
.
Ad
d
itio
n
ally
,
th
e
u
s
e
o
f
d
ee
p
lear
n
in
g
allo
ws
f
o
r
a
m
o
r
e
ac
cu
r
ate
class
if
icatio
n
o
f
f
ac
ial
f
ea
tu
r
es,
lead
in
g
to
im
p
r
o
v
e
d
p
e
r
f
o
r
m
an
ce
.
Fin
al
ly
,
o
u
r
m
eth
o
d
ca
n
ea
s
ily
s
ca
le
to
ac
co
m
m
o
d
ate
lar
g
er
d
atasets
an
d
m
o
r
e
co
m
p
lex
class
if
icatio
n
task
s
.
Ov
er
all,
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
o
f
f
er
s
a
p
o
wer
f
u
l
an
d
ef
f
icien
t
ap
p
r
o
ac
h
to
f
ac
ial
r
ec
o
g
n
itio
n
.
W
e
b
eliev
e
th
at
th
is
m
eth
o
d
will
p
av
e
th
e
way
f
o
r
t
h
e
d
e
v
elo
p
m
e
n
t o
f
m
o
r
e
a
cc
u
r
ate
an
d
r
eliab
le
f
ac
ial
r
ec
o
g
n
itio
n
s
y
s
tem
s
th
at
ca
n
b
e
u
s
ed
in
a
wid
e
r
an
g
e
o
f
ap
p
licatio
n
s
.
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
E
fficien
t sma
r
t d
is
t
r
ib
u
ted
fa
ce
id
en
tifi
ca
tio
n
u
s
in
g
t
h
e
MixMa
xS
im
…
(
S
a
ye
d
Mo
h
a
mma
d
A
h
ma
d
i
)
7147
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
s
tu
d
y
ar
e:
i)
Pro
p
o
s
in
g
a
n
o
v
el
m
eth
o
d
to
en
h
an
ce
f
ac
e
id
e
n
tific
atio
n
ac
cu
r
ac
y
b
y
in
te
g
r
atin
g
clu
s
te
r
in
g
alg
o
r
ith
m
s
with
d
ee
p
lear
n
in
g
m
o
d
els
;
ii)
E
v
alu
atin
g
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
n
th
r
ee
wid
el
y
u
s
ed
f
ac
e
r
ec
o
g
n
itio
n
d
atasets
(
Glin
t3
6
0
K,
VGGFac
e2
,
MS
-
C
eleb
-
1
M)
an
d
d
em
o
n
s
tr
atin
g
s
ig
n
if
ican
t
im
p
r
o
v
e
m
en
ts
in
a
cc
u
r
ac
y
co
m
p
ar
ed
t
o
co
m
p
eti
tiv
e
m
eth
o
d
s
;
iii)
E
m
p
lo
y
i
n
g
s
tate
-
of
-
th
e
-
ar
t
an
d
r
o
b
u
s
t
b
ac
k
b
o
n
es,
s
u
ch
as
Ar
c
Face
,
SF
ac
e,
an
d
Gh
o
s
tFace,
f
o
r
f
ac
ial
f
ea
t
u
r
e
e
x
tr
ac
tio
n
;
a
n
d
iv
)
C
o
n
d
u
ctin
g
a
co
m
p
r
eh
e
n
s
iv
e
ex
p
er
im
en
tal
an
aly
s
is
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
,
in
clu
d
i
n
g
co
m
p
a
r
is
o
n
s
with
co
m
p
etitiv
e
m
eth
o
d
s
.
2.
P
RO
P
O
SE
D
M
E
T
H
O
D
T
h
e
lim
itatio
n
o
f
I
SM
[
2
5
]
.
T
h
is
lim
itatio
n
h
as
b
ee
n
g
e
n
er
a
lly
m
en
tio
n
e
d
in
[
2
6
]
a
n
d
is
th
o
r
o
u
g
h
ly
ex
am
in
ed
in
[
2
7
]
.
C
o
n
s
id
er
in
g
th
at
th
e
p
r
o
p
o
s
ed
m
eth
o
d
is
b
ased
o
n
th
e
I
SM
ap
p
r
o
ac
h
[
2
5
]
,
we
d
is
cu
s
s
th
e
lim
itatio
n
s
o
f
th
e
I
SM
m
eth
o
d
h
er
e.
I
SM
m
eth
o
d
h
as
a
we
ak
n
ess
th
at
ar
is
es
f
r
o
m
t
h
e
r
a
n
d
o
m
d
is
tr
ib
u
tio
n
o
f
class
es
am
o
n
g
s
u
b
n
ets.
T
o
h
e
lp
u
n
d
e
r
s
tan
d
th
is
wea
k
n
ess
,
co
n
s
id
er
two
s
ce
n
ar
i
o
s
.
I
n
t
h
e
f
ir
s
t
s
ce
n
ar
io
,
we
ca
n
m
ak
e
th
e
f
o
llo
win
g
ass
u
m
p
tio
n
s
:
a.
T
h
e
p
r
o
b
lem
co
n
s
is
ts
o
f
f
o
u
r
c
lass
es
: x
,
y
,
z,
a
n
d
w.
b.
T
h
e
test
s
am
p
le
t b
elo
n
g
s
to
cl
ass
x
.
c.
T
h
er
e
ar
e
two
s
u
b
n
ets:
A
an
d
B
.
d.
C
las
s
es x
an
d
w
ar
e
s
itu
ated
w
ith
in
s
u
b
n
et
A,
wh
ile
class
es y
an
d
z
r
esid
e
with
in
s
u
b
n
et
B
.
e.
C
las
s
es
x
,
y
,
an
d
z
b
ea
r
s
tr
ik
i
n
g
r
esem
b
lan
ce
s
to
o
n
e
an
o
th
er
,
y
et
th
ey
d
i
v
er
g
e
s
ig
n
if
ican
tly
f
r
o
m
class
w,
as v
is
u
ally
d
ep
icted
in
Fig
u
r
e
1
.
W
h
en
we
in
p
u
t
th
e
test
s
am
p
le
in
to
b
o
th
s
u
b
n
etwo
r
k
s
,
we
an
ticip
ate
th
at
s
u
b
n
etwo
r
k
A
will
y
ield
o
u
tp
u
t
v
alu
es
o
f
ap
p
r
o
x
im
a
tely
0
.
9
9
an
d
0
.
0
1
,
r
esp
ec
tiv
ely
.
T
h
is
ex
p
ec
tatio
n
ar
is
es
f
r
o
m
th
e
clo
s
e
r
esem
b
lan
ce
b
etwe
en
th
e
test
s
am
p
le
(
t)
an
d
class
x
,
wh
ile
it
s
ig
n
if
ican
tly
d
if
f
e
r
s
f
r
o
m
clas
s
w.
C
o
n
v
er
s
ely
,
in
s
u
b
n
etwo
r
k
B
,
we
f
o
r
esee
o
u
t
p
u
t
v
alu
es
a
r
o
u
n
d
0
.
4
8
an
d
0
.
5
2
,
as
t
ex
h
ib
its
s
im
ilar
ity
to
b
o
th
class
es
y
an
d
z.
As
a
r
esu
lt
o
f
th
e
m
a
x
-
m
ax
a
p
p
r
o
ac
h
,
s
u
b
n
et
A
is
ch
o
s
en
f
r
o
m
th
e
two
s
u
b
n
etwo
r
k
s
,
an
d
class
x
is
s
elec
ted
,
wh
ich
is
th
e
co
r
r
ec
t
ch
o
ice.
T
h
is
s
ce
n
ar
io
is
a
fo
r
tu
n
a
te
s
ta
tu
s
f
o
r
th
e
r
an
d
o
m
d
is
tr
ib
u
tio
n
o
f
class
es.
I
n
th
e
s
ec
o
n
d
s
ce
n
ar
io
,
we
ass
u
m
e
th
at
all
th
e
ass
u
m
p
tio
n
s
o
f
th
e
f
ir
s
t
s
ce
n
ar
io
ar
e
m
et,
ex
ce
p
t
f
o
r
ass
u
m
p
tio
n
d
,
wh
ich
ch
an
g
es a
s
f
o
llo
ws:
d
d
)
x
an
d
y
b
el
o
n
g
t
o
s
u
b
n
et
A,
a
n
d
w
an
d
z
b
elo
n
g
to
s
u
b
n
et
B
,
s
ee
Fig
u
r
e
2
.
Fig
u
r
e
1
.
Fo
r
t
u
n
ate
in
I
SM
[
2
7
]
.
T
h
e
test
im
ag
e
is
p
r
ed
icted
tr
u
e
in
r
a
n
d
o
m
d
is
tr
ib
u
ted
s
ce
n
ar
io
Fig
u
r
e
2
.
Un
f
o
r
tu
n
ate
in
I
SM
[
2
7
]
.
T
h
e
test
im
ag
e
is
n
o
t p
r
e
d
icted
in
r
a
n
d
o
m
d
is
tr
ib
u
ted
s
ce
n
ar
io
W
h
en
th
e
test
s
am
p
le
t
is
f
e
d
i
n
to
b
o
th
s
u
b
n
ets
in
th
e
s
ec
o
n
d
s
ce
n
ar
io
,
th
e
o
u
tp
u
t
v
alu
es
in
s
u
b
n
et
A
will
b
e
ap
p
r
o
x
im
ately
0
.
5
2
f
o
r
class
x
a
n
d
0
.
4
8
f
o
r
class
y
,
wh
ile
th
e
o
u
tp
u
t
v
alu
es
in
s
u
b
n
et
B
will
b
e
0
.
8
0
f
o
r
class
w
an
d
0
.
2
0
f
o
r
clas
s
z.
T
h
er
ef
o
r
e,
s
elec
tin
g
class
w
with
in
s
u
b
n
et
B
is
an
in
co
r
r
ec
t
ch
o
ice.
T
h
is
p
h
en
o
m
en
o
n
o
cc
u
r
s
b
ec
au
s
e
t
h
e
s
co
r
e
o
f
th
e
o
u
tp
u
t
v
alu
e
o
f
th
e
ac
tu
al
class
is
r
an
d
o
m
l
y
r
ed
u
ce
d
s
in
ce
it
is
p
lace
d
with
in
a
s
u
b
n
et
o
f
s
im
ilar
class
es.
Ho
wev
er
,
s
im
ilar
n
eg
ativ
e
class
es
with
in
o
th
er
clu
s
ter
s
g
et
h
ig
h
er
s
co
r
es
with
o
u
t
a
n
y
s
er
io
u
s
co
m
p
etito
r
s
d
u
e
to
th
e
m
ax
-
m
a
x
m
ec
h
an
is
m
,
ca
u
s
in
g
th
e
tr
u
e
class
to
u
ltima
tely
f
ail.
W
e
ca
ll
th
is
s
ce
n
ar
io
an
u
n
fo
r
tu
n
a
te
s
tatu
s
.
T
o
f
all
in
t
o
th
e
tr
ap
o
f
u
n
f
o
r
tu
n
ate,
two
co
n
d
itio
n
s
m
u
s
t
b
e
m
et
to
g
eth
er
:
C
o
n
d
itio
n
A
: T
h
e
ac
t
u
al
class
s
h
o
u
ld
b
e
i
n
th
e
s
am
e
s
u
b
n
etwo
r
k
as si
m
ilar
class
es.
C
o
n
d
itio
n
B
: Sim
ilar
class
es s
h
o
u
ld
b
e
d
is
tr
ib
u
ted
in
d
if
f
er
e
n
t su
b
n
etwo
r
k
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.
14
,
No
.
6
,
Dec
em
b
e
r
20
24
:
7
1
4
5
-
7
1
5
7
7148
On
e
ap
p
r
o
ac
h
to
o
v
er
co
m
in
g
th
e
wea
k
n
ess
o
f
I
SM
is
to
u
s
e
clu
s
ter
in
g
to
p
r
ev
en
t
th
e
o
cc
u
r
r
en
ce
o
f
C
o
n
d
itio
n
B
.
T
h
is
m
ea
n
s
th
at
ef
f
o
r
ts
a
r
e
m
a
d
e
to
p
r
ev
e
n
t
s
im
ilar
class
es
f
r
o
m
b
ein
g
d
is
tr
ib
u
ted
ac
r
o
s
s
d
if
f
er
en
t
clu
s
ter
s
an
d
,
in
s
tead
,
to
p
lace
th
em
in
a
u
n
if
ie
d
clu
s
ter
.
C
lu
s
ter
in
g
alg
o
r
ith
m
s
,
s
u
ch
as
k
-
m
ea
n
s
,
ca
n
b
e
u
s
ed
f
o
r
t
h
is
p
u
r
p
o
s
e.
T
h
i
s
ap
p
r
o
ac
h
m
ay
s
ee
m
to
r
esu
lt
in
s
ev
er
e
co
m
p
etitio
n
with
i
n
clu
s
ter
s
an
d
l
o
w
o
u
tp
u
t
v
alu
es
f
o
r
win
n
in
g
cla
s
s
es.
Ho
wev
er
,
th
is
p
er
ce
p
tio
n
is
n
o
t
ac
cu
r
ate
as
clu
s
ter
in
g
ac
tu
ally
d
ec
r
ea
s
es
th
e
d
is
p
er
s
io
n
o
f
d
ata
with
in
ea
ch
clu
s
ter
an
d
allo
ws
th
e
n
etwo
r
k
to
f
o
cu
s
m
o
r
e
o
n
d
is
tin
g
u
is
h
in
g
b
etwe
en
s
im
ilar
class
e
s
.
T
h
u
s
,
th
e
n
eg
a
tiv
e
ef
f
ec
t is n
eu
tr
alize
d
b
y
th
i
s
p
o
s
itiv
e
p
r
o
p
er
t
y
.
I
n
co
n
tr
ast to
I
SM,
th
is
m
eth
o
d
ca
n
d
eter
m
in
e
t
h
e
clu
s
ter
to
wh
ich
th
e
in
p
u
t im
ag
e
b
elo
n
g
s
.
Fo
r
ea
ch
clu
s
ter
,
a
r
ep
r
esen
tativ
e
ca
n
b
e
d
eter
m
in
ed
an
d
c
o
m
p
ar
ed
w
ith
th
e
f
ea
tu
r
e
v
ec
to
r
o
f
th
e
in
p
u
t
im
ag
e
to
ass
ig
n
it
to
th
e
clu
s
ter
with
th
e
h
ig
h
est
s
im
ilar
ity
.
T
h
is
is
n
o
t
p
o
s
s
ib
le
with
I
SM,
wh
ich
r
elies
o
n
th
e
m
ax
-
m
ax
m
ec
h
an
is
m
.
T
h
er
ef
o
r
e,
an
alter
n
ativ
e
cr
iter
io
n
n
ee
d
s
to
b
e
u
s
ed
.
W
e
u
s
e
n
ea
r
est
n
eig
h
b
o
r
with
s
im
ilar
ity
m
ea
s
u
r
e
in
th
e
a
ffin
ity
-
mete
r
m
o
d
u
le,
in
s
tead
o
f
clu
s
ter
ce
n
ter
s
.
Fig
u
r
e
3
s
h
o
ws
th
e
m
ain
id
ea
o
f
th
is
s
tu
d
y
:
Fig
u
r
e
3
(
a)
s
h
o
w
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
in
v
o
l
v
es
d
is
tr
ib
u
tin
g
all
clas
s
es
in
to
m
u
ltip
le
to
ile
r
s
u
s
in
g
clu
s
ter
in
g
.
E
ac
h
to
iler
c
o
n
tain
s
s
im
ilar
cl
ass
es.
I
n
Fig
u
r
e
3
(
b
)
test
p
h
as
e,
two
ca
n
d
id
ate
ca
teg
o
r
ies
ar
e
co
n
s
id
er
ed
:
(
b
1
)
p
r
ed
ictin
g
th
e
test
s
am
p
le
u
s
i
n
g
ea
c
h
s
u
b
n
et
(
t
o
iler
'
s
m
o
d
u
l
e)
an
d
s
elec
tin
g
t
h
e
class
with
th
e
h
ig
h
est
o
u
tp
u
t
v
alu
e
(
m
ax
-
m
a
x
)
,
an
d
(
b
2
)
d
eter
m
in
in
g
th
e
cl
o
s
est
class
to
th
e
s
am
p
le
b
ased
o
n
th
e
af
f
in
ities
o
f
th
e
to
iler
(
n
ea
r
est
class
)
.
Fin
ally
,
(
b
3
)
t
h
e
p
r
ed
icted
class
is
ch
o
s
en
to
m
ax
im
ize
th
e
p
r
o
d
u
ct
o
f
th
e
o
u
tp
u
t
v
alu
e
a
n
d
th
e
s
im
ilar
ity
.
(
a)
(
b
)
Fig
u
r
e
3
.
T
h
e
m
ain
c
o
n
ce
p
t i
n
th
e
s
tu
d
y
(
a
)
p
r
e
p
r
o
ce
s
s
in
g
p
h
ase,
an
d
(
b
)
p
o
s
tp
r
o
ce
s
s
in
g
p
h
ase
2
.
1
.
F
ra
m
ewo
r
k
o
v
er
v
iew
W
e
b
eg
in
b
y
u
s
in
g
a
p
r
e
-
tr
ai
n
ed
C
NN
[
2
8
]
,
[
2
9
]
to
ex
tr
ac
t
f
ea
tu
r
es
f
o
r
all
im
ag
es
in
ea
ch
d
ataset,
r
esu
ltin
g
in
a
co
llectio
n
o
f
f
ea
tu
r
es
d
en
o
te
d
as
=
{
}
=
1
,
wh
er
e
ℝ
,
N
r
ep
r
esen
ts
th
e
n
u
m
b
er
o
f
im
ag
es,
D
is
a
d
ataset,
an
d
d
r
ep
r
esen
ts
th
e
f
ea
tu
r
e
v
ec
t
o
r
d
im
en
s
io
n
.
T
o
im
p
lem
en
t
o
u
r
f
ac
e
id
en
tific
atio
n
ar
ch
itectu
r
e,
we
b
r
ea
k
th
e
id
e
n
tific
atio
n
p
r
o
b
lem
in
to
th
r
ee
m
o
d
u
les.
T
h
e
f
i
r
s
t
m
o
d
u
le
co
m
p
r
is
es
a
g
r
o
u
p
o
f
to
iler
s
,
with
ea
ch
b
ein
g
tr
ain
ed
to
id
en
tify
a
s
p
ec
if
ic
s
u
b
s
et
o
f
class
es.
C
lu
s
ter
in
g
is
u
s
ed
to
d
is
tr
ib
u
te
all
class
es
am
o
n
g
th
e
to
iler
s
,
en
s
u
r
in
g
th
at
ea
c
h
to
iler
is
r
es
p
o
n
s
ib
le
f
o
r
id
en
tif
y
in
g
s
im
il
ar
class
es.
Du
r
in
g
test
in
g
,
th
e
ex
tr
ac
ted
f
ea
tu
r
e
v
ec
to
r
f
r
o
m
th
e
test
im
ag
e
is
g
iv
en
as
in
p
u
t
to
all
to
iler
s
.
E
ac
h
to
iler
(
wh
ich
is
a
n
eu
r
al
n
etwo
r
k
)
ca
lcu
lates
an
d
r
etu
r
n
s
th
e
o
u
tp
u
t
v
alu
es
a
s
s
o
ciate
d
with
ea
ch
class
f
o
r
th
e
test
d
ata.
T
h
e
s
ec
o
n
d
m
o
d
u
le,
th
e
a
ffin
ity
-
mete
r
m
o
d
u
le,
ca
lcu
lates
th
e
s
im
ilar
ity
o
f
th
e
in
p
u
t
to
th
e
av
er
ag
e
f
ea
tu
r
es
o
f
ea
ch
class
o
f
ea
ch
to
iler
s
ep
ar
ately
.
Fin
ally
,
th
e
th
ir
d
m
o
d
u
l
e,
th
e
fin
a
l
d
ec
is
io
n
ma
ke
r
m
o
d
u
le,
s
u
m
m
ar
izes
th
e
r
ep
o
r
ts
r
ec
eiv
ed
f
r
o
m
th
e
f
ir
s
t
an
d
s
ec
o
n
d
m
o
d
u
les
a
n
d
s
elec
ts
a
class
as
th
e
f
in
al
ca
teg
o
r
y
th
at
h
as
th
e
h
ig
h
est
o
v
er
all
s
co
r
e
o
b
tain
e
d
f
r
o
m
th
e
f
ir
s
t
m
o
d
u
le'
s
o
u
t
p
u
t
v
alu
e
an
d
th
e
af
f
in
ity
s
co
r
e
f
r
o
m
th
e
s
ec
o
n
d
m
o
d
u
le.
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
E
fficien
t sma
r
t d
is
t
r
ib
u
ted
fa
ce
id
en
tifi
ca
tio
n
u
s
in
g
t
h
e
MixMa
xS
im
…
(
S
a
ye
d
Mo
h
a
mma
d
A
h
ma
d
i
)
7149
2
.
2
.
F
ra
m
ewo
r
k
m
o
du
les
Ou
r
p
r
o
p
o
s
ed
m
eth
o
d
c
o
n
s
is
ts
o
f
th
r
ee
m
o
d
u
les
:
i
)
to
iler
s
,
ii
)
af
f
in
ity
m
eter
,
a
n
d
iii
)
f
in
a
l
d
ec
is
io
n
m
ak
er
.
E
ac
h
m
o
d
u
le
p
lay
s
a
cr
u
cial
r
o
le
in
en
s
u
r
in
g
a
cc
u
r
ate
an
d
ef
f
icien
t
o
u
tco
m
es.
T
h
is
m
o
d
u
lar
ar
ch
itectu
r
e
allo
ws
f
o
r
s
p
ec
i
alize
d
p
r
o
ce
s
s
in
g
wh
ile
m
ai
n
tain
in
g
a
clea
r
f
lo
w
o
f
in
f
o
r
m
atio
n
b
etwe
en
co
m
p
o
n
en
ts
.
2
.
2
.
1
.
T
o
ilers
E
ac
h
to
iler
is
a
s
u
p
er
v
is
ed
n
e
u
r
al
n
etwo
r
k
th
at
w
o
r
k
s
in
p
ar
allel
with
o
th
er
to
iler
s
a
n
d
is
r
esp
o
n
s
ib
le
f
o
r
id
en
tify
i
n
g
ap
p
r
o
x
im
ately
/
class
e
s
,
wh
er
e
N
is
th
e
to
tal
n
u
m
b
er
o
f
class
es
an
d
m
i
s
t
h
e
n
u
m
b
e
r
o
f
to
iler
s
.
T
h
e
d
is
tr
ib
u
tio
n
o
f
cl
ass
es
am
o
n
g
th
ese
to
iler
s
is
p
er
f
o
r
m
ed
u
s
in
g
clu
s
ter
in
g
a
lg
o
r
ith
m
s
s
u
ch
as
k
-
m
ea
n
s
,
wh
ich
g
r
o
u
p
s
s
im
ilar
class
es
in
th
e
s
am
e
clu
s
ter
an
d
d
is
s
im
ilar
class
e
s
in
d
if
f
er
en
t
clu
s
ter
s
.
E
ac
h
to
iler
is
tr
ain
ed
u
s
in
g
d
ata
co
r
r
esp
o
n
d
in
g
to
its
class
e
s
,
wh
i
ch
ar
e
th
e
f
ea
tu
r
es
o
f
im
ag
es
ex
tr
ac
ted
b
y
a
p
r
e
-
tr
ain
ed
LRe
s
N
et1
0
0
-
IR
m
o
d
el
.
Du
r
in
g
th
e
test
p
h
ase,
th
e
to
iler
’
s
m
o
d
u
le
d
eter
m
in
es
th
e
s
im
ilar
ity
o
f
th
e
in
p
u
t d
ata
to
ea
ch
o
f
its
ass
ig
n
ed
class
es a
n
d
s
en
d
s
th
e
r
esu
lts
to
th
e
n
ex
t m
o
d
u
le
f
o
r
f
u
r
th
e
r
p
r
o
ce
s
s
in
g
.
a.
T
im
e
co
m
p
lex
ity
W
ith
a
5
1
2
-
d
im
en
s
io
n
al
in
p
u
t
f
ea
tu
r
e
v
ec
to
r
(
=
5
1
2
)
a
n
d
ea
ch
t
o
iler
h
an
d
lin
g
ap
p
r
o
x
im
ately
/
class
es,
th
e
weig
h
t
m
atr
ix
in
t
h
e
p
r
e
-
f
in
al
lay
er
h
as
d
im
en
s
i
o
n
s
o
f
×
/
.
Giv
en
th
at
ea
ch
to
iler
is
tr
ain
ed
s
o
lely
o
n
d
ata
r
elev
a
n
t
to
its
ass
ig
n
ed
class
es
an
d
th
e
en
tire
d
ataset
s
ize
is
,
ea
ch
to
iler
is
t
r
ain
ed
with
a
b
o
u
t
/
d
ata.
T
h
is
lead
s
to
a
to
tal
tim
e
co
m
p
lex
ity
f
o
r
ea
c
h
ep
o
ch
(
E
p
o
ch
T
im
e
)
:
=
×
×
=
×
×
2
(
1
)
I
f
th
e
to
iler
s
a
r
e
tr
ain
e
d
in
p
a
r
allel
o
n
m
u
ltip
le
m
ac
h
in
es,
t
h
e
To
ta
l
Time
r
eq
u
ir
ed
in
P
a
r
a
llel
Mo
d
e
ca
n
b
e
ca
lcu
lated
as
(
2
)
:
=
×
×
×
=
×
×
×
2
(
2
)
Ho
wev
er
,
if
th
e
y
ar
e
tr
ain
e
d
s
eq
u
en
tially
,
th
e
To
ta
l
Time
r
eq
u
ir
ed
in
S
eria
l
Mo
d
e
n
ee
d
to
b
e
ca
lcu
lated
as
(
3
)
:
=
×
×
×
(
3
)
I
f
o
n
ly
a
s
in
g
le
to
iler
is
em
p
lo
y
ed
(
m=1
)
,
th
e
n
ec
ess
ar
y
n
u
m
b
er
o
f
tim
e
u
n
its
will b
e:
(
=
1
)
=
(
=
1
)
=
×
×
×
,
(
4
)
wh
ich
is
tim
es sl
o
wer
th
an
th
e
d
is
tr
ib
u
ted
m
o
d
el
in
s
er
ial
r
u
n
n
in
g
an
d
2
tim
es
s
lo
wer
th
an
th
e
d
is
tr
ib
u
ted
m
o
d
el
in
p
ar
allel
r
u
n
n
in
g
.
Hen
ce
,
th
e
g
r
ea
ter
th
e
n
u
m
b
e
r
o
f
to
iler
s
,
th
e
f
ewe
r
to
tal
o
p
er
ati
o
n
s
ar
e
n
ee
d
e
d
.
B
u
t
o
n
th
e
o
th
er
h
an
d
,
th
e
r
is
k
o
f
o
v
er
f
itti
n
g
will
in
cr
ea
s
e.
T
h
er
ef
o
r
e,
a
b
alan
ce
p
o
in
t
s
h
o
u
ld
b
e
c
o
n
s
id
er
ed
b
etwe
en
th
e
co
m
p
lex
ity
o
f
tim
e
(
th
e
n
u
m
b
er
o
f
r
e
q
u
ir
ed
o
p
e
r
atio
n
s
)
an
d
th
e
ac
cu
r
ac
y
o
f
t
h
e
s
y
s
tem
.
b.
Me
m
o
r
y
co
m
p
lex
ity
E
ac
h
to
iler
n
ee
d
s
to
b
e
tr
ain
ed
f
o
r
e
ep
o
ch
s
,
with
ea
ch
s
tep
o
f
ea
ch
ep
o
ch
r
e
q
u
ir
in
g
b
b
atch
es
o
f
d
ata.
T
h
ese
b
atch
es
a
r
e
m
u
ltip
lied
b
y
th
e
weig
h
t
m
at
r
ix
with
d
im
en
s
io
n
s
5
1
2
×
C
/m
,
r
esu
ltin
g
in
a
C
/m
v
ec
to
r
.
T
h
er
ef
o
r
e,
in
ea
c
h
s
tep
,
we
r
eq
u
ir
e
b
×
512
×
4
b
y
tes
to
s
to
r
e
i
n
p
u
t
d
ata
a
n
d
512
×
C
/m
×
4
b
y
tes
f
o
r
s
to
r
i
n
g
weig
h
t m
atr
ix
.
2
.
2
.
2
.
Af
f
ini
t
y
m
et
er
T
h
is
m
o
d
u
le
ca
lcu
lates
th
e
s
im
ilar
ities
b
etwe
en
th
e
in
p
u
t
d
ata
an
d
th
e
av
er
a
g
e
f
ea
tu
r
es
o
f
ea
ch
class
f
o
r
ea
ch
to
iler
.
First,
we
ca
lcu
late
th
e
av
er
ag
e
f
ea
tu
r
e
v
ec
to
r
s
o
f
tr
ain
in
g
d
ata
f
o
r
ea
c
h
class
:
̄
=
∑
=
1
,
=
1
,
2
,
3
,
.
.
.
,
,
(
5
)
wh
er
e
d
en
o
tes
th
e
s
p
ec
if
ic
cl
ass
in
th
e
d
ata
s
et,
̅
is
th
e
m
ea
n
o
f
f
ea
tu
r
e
v
ec
to
r
s
o
f
class
,
C
is
th
e
to
tal
n
u
m
b
er
o
f
class
es
in
th
e
d
ata
s
et,
is
th
e
j
-
th
f
ea
tu
r
e
v
ec
to
r
o
f
th
e
class
f
r
o
m
th
e
tr
ain
in
g
d
ata,
is
th
e
n
u
m
b
er
o
f
tr
ain
i
n
g
s
am
p
les f
o
r
class
.
Nex
t,
we
ca
lcu
late
th
e
af
f
in
ity
m
atr
ix
b
ased
o
n
th
e
s
im
ilar
ity
b
etwe
en
th
e
in
p
u
t f
ea
t
u
r
e
v
ec
t
o
r
an
d
th
e
m
ea
n
f
ea
tu
r
e
v
ec
to
r
s
o
f
all
c
lass
es f
o
r
ea
ch
to
iler
.
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.
14
,
No
.
6
,
Dec
em
b
e
r
20
24
:
7
1
4
5
-
7
1
5
7
7150
(
1
,
2
,
.
.
.
,
)
in
(
1
,
2
,
.
.
.
,
)
(
,
)
,
i
t
tm
tt
i
n
c
iC
A
c
o
s
f
F
=
=
=
,
(
6
)
wh
er
e
r
ep
r
esen
ts
th
e
f
ea
tu
r
es
o
f
th
e
in
p
u
t
s
am
p
le,
d
en
o
tes
th
e
n
u
m
b
e
r
o
f
to
iler
s
,
is
th
e
n
u
m
b
er
o
f
class
es
in
t
-
th
to
iler
.
r
ep
r
esen
ts
th
e
af
f
in
ity
b
etwe
en
th
e
in
p
u
t a
n
d
all
clas
s
es
in
to
iler
t
,
an
d
̅
d
en
o
tes th
e
av
er
ag
e
f
ea
t
u
r
e
v
ec
t
o
r
o
f
class
f
o
r
to
iler
.
a.
T
im
e
co
m
p
lex
ity
I
f
we
ass
u
m
e
th
at
we
h
av
e
n
im
ag
es f
o
r
ea
ch
id
e
n
tity
,
ca
lcu
l
atin
g
th
e
m
ea
n
o
f
t
h
ese
n
f
ea
tu
r
e
v
ec
to
r
s
will
r
eq
u
ir
e
×
512
o
p
er
atio
n
s
.
Ad
d
itio
n
ally
,
ca
lcu
latin
g
th
e
a
f
f
in
ity
b
etwe
en
an
y
two
f
ea
tu
r
e
v
ec
to
r
s
r
eq
u
ir
es 5
1
2
o
p
er
atio
n
s
.
W
ith
class
es,
th
e
to
tal
n
u
m
b
er
o
f
o
p
er
atio
n
s
n
ee
d
e
d
is
512
2
×
×
o
p
er
atio
n
s
.
2
.
3
.
3
.
F
ina
l dec
is
io
n m
a
k
er
T
h
is
m
o
d
u
le
is
r
esp
o
n
s
ib
le
f
o
r
s
elec
tin
g
th
e
f
in
al
class
lab
el
f
o
r
th
e
in
p
u
t
im
ag
e
b
ased
o
n
t
h
e
r
esu
lts
o
b
tain
ed
f
r
o
m
ea
c
h
to
iler
.
T
h
e
f
o
llo
win
g
s
tep
s
ar
e
p
er
f
o
r
m
ed
f
o
r
ev
er
y
to
iler
:
-
T
h
e
m
o
d
u
le
f
in
d
s
th
e
class
with
th
e
h
ig
h
est
o
u
tp
u
t
s
co
r
e
f
r
o
m
t
h
e
to
iler
’
s
m
o
d
u
le.
Su
b
s
eq
u
en
tly
,
it
r
etr
iev
es th
e
af
f
in
ity
v
alu
e
o
f
t
h
at
class
f
r
o
m
th
e
af
f
in
ity
-
m
et
er
m
o
d
u
le
an
d
m
u
ltip
lies
th
e
two
v
alu
es.
-
T
h
e
m
o
d
u
le
also
f
in
d
s
th
e
class
wi
th
th
e
h
ig
h
est
af
f
in
ity
v
alu
e
f
r
o
m
t
h
e
af
f
in
ity
-
m
eter
m
o
d
u
le
an
d
r
etr
iev
es its
to
iler
r
esp
o
n
s
e
v
al
u
e
f
r
o
m
th
e
to
iler
’
s
m
o
d
u
le.
I
t
m
u
ltip
lies
th
ese
two
v
alu
es a
s
well.
-
T
h
e
m
o
d
u
le
s
elec
ts
th
e
m
a
x
im
u
m
m
u
ltip
lied
v
al
u
e
o
b
tain
ed
f
r
o
m
s
tep
s
a
a
n
d
b
,
an
d
t
h
e
class
co
r
r
esp
o
n
d
in
g
t
o
th
at
v
alu
e
is
co
n
s
id
er
ed
as th
e
c
h
o
s
en
class
o
f
th
at
to
iler
.
Alth
o
u
g
h
th
e
th
r
e
e
m
o
d
u
les
ar
e
lo
g
ically
s
ep
ar
ate,
in
p
r
ac
tic
e,
th
e
f
ir
s
t
m
o
d
u
le
p
er
f
o
r
m
s
it
s
task
an
d
s
en
d
s
th
e
r
esu
lts
to
th
e
th
ir
d
m
o
d
u
le.
B
y
ex
a
m
in
in
g
th
e
r
e
s
u
lts
o
f
th
e
f
ir
s
t
m
o
d
u
le,
th
e
th
ir
d
m
o
d
u
le
m
ay
d
ec
id
e
n
o
t
to
u
s
e
th
e
s
ec
o
n
d
m
o
d
u
le
to
s
p
ee
d
u
p
p
er
f
o
r
m
a
n
ce
.
T
h
is
is
b
ec
au
s
e
if
th
e
s
el
ec
ted
class
ac
co
r
d
in
g
to
th
e
r
esu
lts
o
f
th
e
f
ir
s
t
m
o
d
u
le
h
as
a
v
e
r
y
h
ig
h
s
co
r
e,
its
d
ec
is
io
n
is
ac
cu
r
ate
an
d
tr
u
e
an
d
alm
o
s
t
alwa
y
s
v
er
if
ied
b
y
th
e
s
ec
o
n
d
m
o
d
u
l
e.
T
h
e
s
ec
o
n
d
m
o
d
u
le
is
ac
tu
ally
u
s
ed
as
an
au
x
iliar
y
to
o
l
f
o
r
ca
s
es
wh
er
e
t
h
e
f
ir
s
t
m
o
d
u
le
is
u
n
ce
r
tain
an
d
h
as
a
lo
w
s
co
r
e.
Fu
r
th
er
d
et
ails
r
eg
ar
d
in
g
th
is
p
o
in
t
will
b
e
p
r
o
v
id
ed
in
th
e
ex
p
er
im
en
ts
s
ec
tio
n
.
Alg
o
r
it
h
m
1
p
r
o
v
id
es
a
m
o
r
e
d
etailed
d
escr
ip
tio
n
o
f
t
h
e
p
o
s
t
-
p
r
o
ce
s
s
in
g
s
tep
(
test
p
h
ase)
.
Alg
o
r
ith
m
1
.
Po
s
t
-
p
r
o
ce
s
s
in
g
p
h
ase
I
N
P
U
T
S
:
face:
112
×
112
×
3
RGB face image, c: number of subnets, thr: threshold
O
U
T
P
U
T
:
: predicted class label
#
Feature
extraction
1
F = extract_features(
face
)
#
Best classes from
toilers
module
2
for m in
range(0, c):
3
=
(
(
,
)
)
4
=
(
(
,
)
)
5
=
0
.
6
if
>
:
7
=
8
=
9
=
10
=
_
(
,
)
11
if
>
ℎ
:
12
return
#
Best classes from
affinity
-
meter
module
13
for m in range(0, c):
14
=
(
(
,
)
)
15
=
_
(
,
)
16
if
>
:
17
=
18
=
19
=
20
=
(
,
)
[
]
#
Final decision
21
If
×
>
×
:
22
= label of
.
23
otherwise:
24
= label of
.
25
return
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
E
fficien
t sma
r
t d
is
t
r
ib
u
ted
fa
ce
id
en
tifi
ca
tio
n
u
s
in
g
t
h
e
MixMa
xS
im
…
(
S
a
ye
d
Mo
h
a
mma
d
A
h
ma
d
i
)
7151
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
E
x
perim
ent
a
l
s
et
t
ing
s
a.
Data
s
ets
T
h
is
s
tu
d
y
u
s
ed
th
e
f
o
llo
win
g
d
atasets
:
i)
T
h
e
Glin
t3
6
0
k
d
a
tab
ase
[
2
2
]
is
r
en
o
wn
ed
as
th
e
clea
n
est
an
d
m
o
s
t
ex
ten
s
iv
e
r
eso
u
r
ce
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
.
W
ith
in
th
is
d
ata
b
ase,
a
r
em
ar
k
ab
l
e
3
6
0
,
2
3
2
u
n
iq
u
e
in
d
iv
id
u
als
ar
e
r
ep
r
esen
ted
,
ac
co
m
p
an
ied
b
y
an
im
p
r
ess
iv
e
c
o
llectio
n
o
f
1
7
m
illi
o
n
i
m
ag
es
;
ii)
MS
-
C
eleb
-
1
M
[
3
0
]
c
o
n
tain
s
1
0
0
K
id
en
titi
es,
ea
ch
id
en
tity
h
as
ab
o
u
t
1
0
0
f
ac
ial
im
ag
es
;
an
d
iii)
VGGF
ac
e2
[
3
1
]
co
n
tain
s
m
o
r
e
th
an
3
.
3
M
i
m
ag
es o
f
o
v
e
r
9
K
id
en
titi
es.
Fo
r
ea
ch
d
ataset,
th
e
class
e
s
wer
e
s
elec
ted
r
an
d
o
m
ly
.
T
h
e
y
wer
e
d
iv
id
ed
in
to
th
r
ee
p
ar
t
s
:
6
0
%
f
o
r
tr
ain
in
g
,
2
0
%
f
o
r
v
alid
atio
n
,
an
d
th
e
r
est
f
o
r
test
in
g
.
I
n
co
n
s
id
er
atio
n
o
f
th
e
v
ar
y
in
g
n
u
m
b
er
o
f
s
am
p
les
in
ea
ch
class
,
we,
th
er
e
f
o
r
e,
co
n
s
id
er
th
e
ev
al
u
atio
n
s
am
p
le
s
ize
eq
u
iv
ale
n
t
to
t
h
e
n
u
m
b
e
r
o
f
test
s
am
p
les,
i.e
.
,
f
iv
e
f
o
r
ea
ch
.
I
n
th
e
ab
s
en
ce
o
f
s
u
f
f
icien
t
s
am
p
les
in
s
o
m
e
class
es,
n
ew
d
ata
wer
e
g
en
er
ated
u
s
in
g
d
ata
au
g
m
en
tatio
n
,
th
r
o
u
g
h
r
an
d
o
m
r
o
tatio
n
,
f
lip
,
b
r
ig
h
tn
ess
,
an
d
co
n
tr
ast tr
an
s
f
o
r
m
atio
n
s
.
b.
Featu
r
e
ex
tr
ac
tio
n
T
h
e
p
r
o
p
o
s
ed
alg
o
r
ith
m
em
p
lo
y
s
s
o
m
e
o
f
th
e
latest
an
d
m
o
s
t
p
o
wer
f
u
l
f
ea
tu
r
e
ex
tr
ac
to
r
s
in
s
tead
o
f
th
e
r
aw
im
ag
es:
-
Ar
cFac
e
[
1
7
]
:
T
h
e
b
ac
k
b
o
n
e
ar
ch
itectu
r
e
o
f
t
h
is
ex
tr
ac
to
r
i
s
L
R
esNet1
0
0
E
-
I
R
[
1
7
]
,
wh
i
ch
is
a
r
esid
u
al
n
etwo
r
k
[
2
8
]
.
I
t w
as tr
ain
ed
u
s
in
g
th
e
Ar
cFac
e
lo
s
s
f
u
n
ctio
n
an
d
th
e
MS
-
C
eleb
-
1
M
[
3
0
]
d
a
taset.
-
S
f
ace
[
1
8
]
:
T
h
is
s
tu
d
y
em
p
l
o
y
s
th
e
I
R
esNet5
0
v
er
s
io
n
o
f
th
e
f
ea
tu
r
e
ex
tr
ac
to
r
as
th
e
b
ac
k
b
o
n
e
an
d
u
tili
ze
s
C
o
s
Face
[
1
6
]
as
th
e
lo
s
s
f
u
n
ctio
n
.
T
h
e
au
th
o
r
s
tr
ain
ed
th
is
m
o
d
el
f
r
o
m
s
c
r
atch
u
s
in
g
th
e
VGGFac
e2
[
3
1
]
d
ataset.
-
Gh
o
s
tFaceN
et
[
1
9
]
:
T
h
is
b
ac
k
b
o
n
e
in
co
r
p
o
r
atin
g
Gh
o
s
tMo
d
u
les.
T
h
e
u
tili
ze
d
v
er
s
io
n
in
th
is
s
tu
d
y
is
th
e
f
ir
s
t
an
d
co
n
f
ig
u
r
e
d
with
a
wi
d
th
o
f
1
an
d
a
s
tr
id
e
o
f
3
.
All
b
ac
k
b
o
n
es
tak
e
1
1
2
×1
1
2
R
GB
im
ag
es
an
d
co
n
v
er
t
t
h
em
in
to
5
1
2
-
d
im
en
s
io
n
al
f
e
atu
r
e
v
ec
t
o
r
s
.
W
e
u
tili
ze
d
th
e
p
r
e
-
t
r
ain
ed
weig
h
ts
p
r
o
v
id
ed
in
.
c.
Selectio
n
o
f
to
iler
s
’
r
ep
r
esen
tativ
es
T
h
er
e
a
r
e
s
o
m
e
ca
n
d
i
d
ates
f
o
r
c
h
o
o
s
in
g
a
r
ep
r
esen
tativ
e
f
o
r
ea
ch
clu
s
ter
:
i)
T
h
e
m
ea
n
ca
n
p
o
ten
tially
s
er
v
e
as
th
e
r
ep
r
esen
tativ
e
f
o
r
ea
ch
clu
s
ter
.
Ho
wev
er
,
in
p
r
ac
tice,
it
m
ay
n
o
t
a
cc
u
r
ately
r
ep
r
esen
t
lar
g
e
-
s
ca
le
f
ac
e
f
ea
tu
r
es
d
u
e
to
th
e
n
atu
r
e
o
f
d
ata
d
is
tr
ib
u
t
io
n
,
r
ef
er
to
T
ab
le
1
;
ii)
Alter
n
ativ
ely
,
a
s
tr
o
n
g
ca
n
d
id
ate
f
o
r
b
ein
g
th
e
clu
s
te
r
r
ep
r
esen
tativ
e
is
th
e
clo
s
est
d
ata
p
o
in
t
to
th
e
s
am
p
le
f
ea
tu
r
es.
I
n
o
th
e
r
wo
r
d
s
,
th
e
test
s
am
p
le
b
elo
n
g
s
to
th
e
clu
s
ter
co
n
tain
in
g
th
e
clo
s
est
d
ata
p
o
in
t
,
r
ef
er
to
T
ab
le
1
;
an
d
iii)
T
h
e
m
ax
im
u
m
v
alu
e
o
f
all
to
iler
o
u
tp
u
ts
ca
n
b
e
u
s
ed
to
d
eter
m
in
e
th
e
clu
s
ter
to
wh
ich
ea
c
h
d
ata
b
elo
n
g
s
(
I
SM
m
eth
o
d
)
.
T
ab
le
1
.
C
lu
s
ter
in
g
ac
cu
r
ac
y
R
e
p
r
e
se
n
t
a
t
i
v
e
Tr
a
i
n
A
c
c
u
r
a
c
y
%
Te
st
A
c
c
u
r
a
c
y
%
M
e
a
n
6
7
.
0
8
5
9
.
7
2
N
e
a
r
e
st
9
8
.
0
8
9
8
.
1
7
*
F
o
r
2
0
,
0
0
0
c
l
a
ss
e
s,
u
si
n
g
k
-
mea
n
s
w
i
t
h
1
0
c
l
u
s
t
e
r
s
d.
Un
ce
r
tain
ty
ar
ea
Fo
r
a
p
r
o
b
lem
in
v
o
lv
in
g
2
0
,
0
0
0
class
es,
with
in
th
e
Glin
t3
6
0
k
d
ataset,
m
o
s
t
o
f
th
e
test
s
am
p
les
y
ield
o
u
tp
u
t
v
alu
es
r
an
g
in
g
f
r
o
m
0
.
6
an
d
0
.
8
.
Ap
p
r
o
x
im
ately
8
0
%
o
f
th
e
o
u
tp
u
t
v
alu
es
o
f
t
h
e
s
e
test
s
am
p
les
ar
e
ex
ce
ed
a
s
p
ec
if
ic
th
r
esh
o
ld
(
a
p
p
r
o
x
im
ately
0
.
4
f
o
r
2
0
,
0
0
0
class
es)
in
th
e
m
ax
-
m
ax
a
p
p
r
o
ac
h
,
s
ee
Fig
u
r
e
4
.
Nea
r
ly
1
0
0
%
o
f
t
h
ese
v
alu
es
r
ep
r
esen
t
co
r
r
ec
t
p
r
e
d
ictio
n
s
as
s
h
o
wn
in
Fig
u
r
e
5
.
I
n
o
t
h
er
w
o
r
d
s
,
alm
o
s
t
all
o
f
th
e
n
etwo
r
k
'
s
p
r
ed
ictio
n
s
h
av
e
b
ee
n
ac
cu
r
ate
f
o
r
o
u
tp
u
t
v
alu
es
g
r
ea
ter
th
an
0
.
4
.
W
h
en
th
e
n
etwo
r
k
'
s
co
n
f
id
en
ce
lev
el
d
r
o
p
s
to
a
r
o
u
n
d
0
.
4
,
th
e
er
r
o
r
p
r
e
d
ictio
n
r
ate
in
cr
ea
s
es.
C
o
n
s
eq
u
en
tly
,
an
y
p
r
o
p
o
s
ed
s
o
lu
tio
n
s
h
o
u
ld
p
r
im
ar
ily
co
n
ce
n
tr
ate
o
n
th
is
p
ar
ticu
lar
r
an
g
e.
I
t
s
h
o
u
ld
b
e
n
o
te
d
th
at
th
is
th
r
esh
o
ld
v
alu
e
m
ay
v
ar
y
d
ep
en
d
i
n
g
o
n
th
e
ty
p
e
o
f
d
ata
s
et,
b
ac
k
b
o
n
e
ty
p
e,
lo
s
s
f
u
n
cti
o
n
,
a
n
d
ar
ch
itectu
r
al
f
ea
tu
r
es
.
Fo
r
th
is
r
ea
s
o
n
,
we
em
p
lo
y
ed
a
d
y
n
a
m
ic
ap
p
r
o
ac
h
to
d
eter
m
in
e
it.
I
n
t
h
is
m
an
n
er
,
f
o
r
v
alid
atio
n
d
atasets
,
we
co
n
s
id
er
th
e
lar
g
est
v
alu
e
as
th
e
th
r
esh
o
ld
,
e
n
s
u
r
i
n
g
th
at
at
least
9
9
.
5
%
o
f
th
e
d
ata
with
o
u
tp
u
t
v
al
u
es
h
ig
h
e
r
th
an
th
is
th
r
esh
o
ld
ar
e
co
r
r
ec
tly
p
r
ed
icted
.
I
n
o
th
er
wo
r
d
s
,
we
f
in
d
a
p
o
in
t
wh
er
e
th
e
m
ajo
r
ity
o
f
er
r
o
r
-
p
r
o
n
e
d
ata
h
av
e
o
u
t
p
u
t
v
alu
es
lo
wer
th
an
it,
an
d
we
r
ef
er
to
th
is
r
eg
io
n
as
th
e
u
n
ce
r
tain
ty
r
eg
io
n
.
Ou
r
m
eth
o
d
aim
s
to
co
n
ce
n
tr
ate
its
f
o
cu
s
o
n
th
is
r
e
g
io
n
.
Fo
r
ex
a
m
p
le,
in
Fig
u
r
e
6
,
w
h
ich
illu
s
tr
ates
th
e
d
is
tr
ib
u
tio
n
o
f
to
iler
u
n
it
ca
n
d
id
ates,
th
e
th
r
esh
o
ld
v
alu
e
is
0
.
4
9
,
an
d
alm
o
s
t
all
o
u
tp
u
t
v
alu
es
ab
o
v
e
th
is
th
r
esh
o
ld
h
av
e
b
e
en
ac
cu
r
ately
p
r
e
d
icted
.
Fig
u
r
e
7
d
ep
icts
th
e
d
is
tr
ib
u
t
io
n
o
f
af
f
in
ity
-
m
eter
u
n
it
ca
n
d
id
ate
v
alu
es,
wh
ic
h
is
v
er
y
s
im
ilar
in
s
h
ap
e
to
Fig
u
r
e
6
.
T
h
is
im
p
lies
th
at
in
th
e
u
n
ce
r
tain
ty
r
eg
io
n
,
with
t
h
e
co
llab
o
r
atio
n
o
f
t
h
e
o
p
i
n
io
n
s
o
f
b
o
th
u
n
its
,
a
h
ig
h
er
ac
c
u
r
ac
y
ca
n
b
e
ac
h
ie
v
ed
.
I
n
f
ac
t,
w
h
en
ev
e
r
th
e
o
u
tp
u
t
v
alu
e
o
f
a
test
d
ata
in
p
u
t
f
r
o
m
to
iler
u
n
it
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.
14
,
No
.
6
,
Dec
em
b
e
r
20
24
:
7
1
4
5
-
7
1
5
7
7152
ex
ce
ed
s
th
e
th
r
esh
o
ld
,
th
e
f
r
a
m
ewo
r
k
n
o
lo
n
g
e
r
awa
its
th
e
o
p
in
io
n
o
f
th
e
af
f
i
n
ity
-
m
eter
u
n
it
an
d
r
elies
s
o
lely
o
n
th
e
to
iler
u
n
it'
s
r
ep
o
r
t
f
o
r
d
ec
is
io
n
-
m
ak
in
g
.
I
n
a
d
d
itio
n
,
wh
en
co
n
s
id
er
in
g
th
e
s
im
ilar
ity
b
etwe
en
th
e
f
ea
tu
r
es
o
f
th
o
s
e
class
es
an
d
t
h
e
test
d
ata,
it
is
s
ig
n
if
ican
tly
lo
w.
Nev
er
th
eless
,
m
ix
in
g
th
ese
two
cr
iter
ia,
i.e
.
,
s
im
ilar
ity
an
d
th
e
m
ax
-
m
ax
c
r
iter
io
n
,
ca
n
lead
to
m
o
r
e
r
o
b
u
s
t r
esu
lts
.
Fig
u
r
e
4
.
C
u
m
u
lativ
e
f
r
eq
u
en
c
y
o
f
test
s
am
p
les
b
ased
o
n
th
e
o
u
tp
u
t v
alu
e,
with
er
r
o
r
s
o
cc
u
r
r
i
n
g
b
ef
o
r
e
t
h
e
r
ed
lin
e
(
u
n
ce
r
tain
t
y
ar
ea
)
Fig
u
r
e
5
.
R
atio
o
f
co
r
r
ec
t a
n
d
in
co
r
r
ec
t p
r
ed
ictio
n
s
f
o
r
m
a
x
-
m
ax
o
u
tp
u
t v
alu
es in
th
e
test
d
ataset,
with
in
co
r
r
ec
t p
r
ed
ictio
n
s
o
cc
u
r
r
in
g
b
ef
o
r
e
th
e
r
e
d
lin
e
(
u
n
ce
r
tain
ty
ar
ea
)
Fig
u
r
e
6
.
Fre
q
u
en
cy
o
f
ea
c
h
m
ax
-
m
ax
v
alu
e
o
f
to
iler
’
s
ca
n
d
id
ates f
o
r
v
alid
ati
o
n
d
ata
(
G
h
o
s
tFace,
MS
-
C
eleb
-
1
M,
2
0
,
0
0
0
class
es,
6
s
u
b
n
ets)
Fig
u
r
e
7
.
Fre
q
u
en
cy
o
f
ea
c
h
s
im
ilar
ity
v
alu
e
o
f
af
f
in
ity
-
m
eter
ca
n
d
id
ates f
o
r
v
alid
atio
n
d
ata
(
Gh
o
s
tFace,
MS
-
C
eleb
-
1
M,
2
0
,
0
0
0
class
es,
6
s
u
b
n
ets)
e.
C
lu
s
ter
in
g
o
f
class
es
T
h
is
s
tu
d
y
d
em
o
n
s
tr
ates
th
at
th
e
u
s
e
o
f
clu
s
ter
in
g
ca
n
allev
iate
th
e
er
r
o
r
s
ca
u
s
ed
b
y
t
h
e
r
an
d
o
m
d
is
tr
ib
u
tio
n
o
f
d
ata
in
to
iler
’
s
m
o
d
u
le.
T
h
e
m
o
r
e
ac
c
u
r
ate
th
e
clu
s
ter
in
g
is
p
er
f
o
r
m
ed
,
t
h
e
h
ig
h
er
t
h
e
f
in
al
p
er
f
o
r
m
an
ce
will
b
e.
Fo
r
s
im
p
licity
an
d
to
p
r
o
v
e
th
e
c
o
r
r
ec
tn
ess
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
,
we
u
s
e
a
v
e
r
y
s
im
p
le
clu
s
ter
in
g
al
g
o
r
ith
m
,
k
-
m
ea
n
s
.
Acc
o
r
d
in
g
to
th
is
ap
p
r
o
ac
h
,
clu
s
ter
in
g
was
ap
p
lied
o
n
all
f
ea
tu
r
e
v
ec
to
r
s
o
f
th
e
tr
ain
in
g
d
ata.
Ho
wev
er
,
th
e
p
r
o
b
lem
is
th
at
n
o
t
all
s
am
p
les
r
elate
d
to
ea
ch
class
ar
e
g
r
o
u
p
ed
in
to
a
s
in
g
le
clu
s
ter
.
Nev
er
th
e
less
,
to
ad
d
r
ess
th
is
i
s
s
u
e,
ea
c
h
class
is
a
s
s
ig
n
ed
to
a
clu
s
ter
th
at
h
as
th
e
h
ig
h
est
n
u
m
b
er
o
f
s
am
p
les
in
it.
I
n
t
h
e
test
p
h
ase,
we
ca
n
esti
m
ate
th
e
to
iler
(
clu
s
ter
)
o
f
a
test
s
am
p
le
u
s
in
g
th
e
af
f
in
ity
-
m
eter
m
o
d
u
le.
T
h
is
is
b
ec
au
s
e
th
e
n
ea
r
est
n
eig
h
b
o
r
s
o
f
ea
ch
to
iler
p
r
o
v
i
d
e
a
b
etter
m
ea
s
u
r
e
f
o
r
esti
m
atin
g
th
e
clu
s
ter
co
m
p
ar
ed
to
th
e
ce
n
ter
s
o
f
th
e
clu
s
ter
s
.
Fro
m
th
e
m
en
tio
n
ed
d
es
cr
ip
tio
n
s
,
it
ca
n
b
e
in
f
er
r
ed
th
at
th
e
p
r
o
v
id
e
d
f
r
a
m
ewo
r
k
is
in
d
ep
en
d
en
t
o
f
t
h
e
d
ataset
an
d
also
th
e
b
ac
k
b
o
n
e.
Ou
r
ex
p
er
im
en
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
E
fficien
t sma
r
t d
is
t
r
ib
u
ted
fa
ce
id
en
tifi
ca
tio
n
u
s
in
g
t
h
e
MixMa
xS
im
…
(
S
a
ye
d
Mo
h
a
mma
d
A
h
ma
d
i
)
7153
r
esu
lts
in
T
ab
le
2
also
co
n
f
ir
m
th
is
claim
.
T
h
ese
ex
p
er
im
en
ts
wer
e
co
n
d
u
cted
o
n
th
e
d
atasets
M
S
-
C
eleb
-
1
M
[
3
0
]
,
VGGFac
e2
[
3
1
]
,
with
v
ar
io
u
s
n
ew
b
ac
k
b
o
n
es
in
clu
d
i
n
g
SF
ac
e
[
1
8
]
,
Gh
o
s
tFaceN
et
[
1
9
]
.
I
n
all
o
f
th
ese
ca
s
es,
an
im
p
r
o
v
em
e
n
t in
ac
cu
r
ac
y
is
o
b
s
er
v
ed
in
th
e
p
r
o
p
o
s
ed
s
ce
n
ar
io
co
m
p
ar
ed
to
co
m
p
etin
g
m
eth
o
d
s
.
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
r
esu
lts
o
n
Gh
o
s
tFaceN
et
an
d
SF
ac
e
D
a
t
a
s
e
t
B
a
c
k
b
o
n
e
C
l
a
s
ses
D
i
st
r
i
b
u
t
i
o
n
-
F
r
e
e
To
i
l
e
r
s
I
S
M
[
2
5
]
o
u
r
p
r
e
+
I
S
M
p
o
st
o
u
r
p
r
e
+
o
u
r
p
o
st
Pre
*
Rec
F1
Pre
Rec
F1
Pre
Rec
F1
Pre
Rec
F1
MS
-
C
e
l
e
b
-
1
M
[
3
0
]
G
h
o
s
t
F
a
c
e
N
e
t
[
1
9
]
5
,
0
0
0
9
7
.
1
9
9
6
.
6
8
9
6
.
5
9
2
9
7
.
2
1
9
6
.
7
4
9
6
.
6
4
9
7
.
1
4
9
6
.
6
6
9
6
.
5
6
9
7
.
2
7
9
6
.
8
0
9
6
.
7
1
1
0
,
0
0
0
9
5
.
2
9
9
4
.
4
6
9
4
.
2
7
3
9
5
.
9
3
9
5
.
2
2
9
5
.
0
8
9
5
.
8
5
9
5
.
1
0
9
4
.
9
7
9
6
.
0
0
9
5
.
2
9
9
5
.
1
7
2
0
,
0
0
0
9
4
.
2
5
9
3
.
1
6
9
2
.
9
2
6
9
5
.
7
9
9
5
.
0
5
9
4
.
9
1
9
5
.
6
4
9
4
.
8
8
9
4
.
7
3
9
5
.
8
7
9
5
.
1
3
9
5
.
0
0
5
0
,
0
0
0
9
1
.
1
9
8
9
.
2
1
8
8
.
7
0
16
9
5
.
1
5
9
4
.
2
3
9
4
.
0
8
9
4
.
8
7
9
3
.
9
0
9
3
.
7
3
9
5
.
2
3
9
4
.
3
2
9
4
.
1
8
VGG
-
F
a
c
e
2
[
3
1
]
S
F
a
c
e
[
1
8
]
5
,
0
0
0
9
1
.
3
1
9
0
.
0
3
8
9
.
7
3
2
9
1
.
6
0
9
0
.
3
5
9
0
.
0
7
9
1
.
4
7
9
0
.
1
2
8
9
.
8
4
9
1
.
9
9
9
0
.
6
8
9
0
.
4
6
6
,
0
0
0
9
0
.
9
1
8
9
.
5
0
8
9
.
2
0
3
9
1
.
3
3
8
9
.
9
9
8
9
.
7
3
9
1
.
0
2
8
9
.
5
8
8
9
.
3
1
9
1
.
7
3
9
0
.
3
7
9
0
.
1
5
7
,
0
0
0
9
0
.
5
1
8
9
.
0
3
8
8
.
7
0
3
9
0
.
9
4
8
9
.
5
5
8
9
.
2
6
9
0
.
8
0
8
9
.
3
4
8
9
.
0
4
9
1
.
4
7
9
0
.
0
7
8
9
.
8
4
8
,
0
0
0
9
0
.
1
2
8
8
.
5
4
8
8
.
2
3
3
9
0
.
7
2
8
9
.
2
7
8
8
.
9
7
9
0
.
6
0
8
9
.
0
8
8
8
.
7
7
9
1
.
2
1
8
9
.
7
8
8
9
.
5
3
8
,
9
0
0
8
9
.
6
5
8
8
.
0
5
8
7
.
6
9
4
9
0
.
4
2
8
8
.
9
5
8
8
.
6
2
9
0
.
1
0
8
8
.
4
9
8
8
.
1
6
9
0
.
8
7
8
9
.
4
4
8
9
.
1
7
*
Ev
a
l
u
a
t
i
o
n
me
t
r
i
c
s
a
r
e
P
r
e
c
i
si
o
n
(
P
r
e
)
,
R
e
c
a
l
l
(
R
e
c
)
,
a
n
d
F
1
-
sc
o
r
e
f.
I
m
p
lem
en
tatio
n
d
etails
A
s
in
g
le
ar
ch
itectu
r
e
was
em
p
lo
y
ed
i
n
all
t
h
e
ex
p
er
im
en
ts
,
f
ea
tu
r
in
g
a
5
1
2
-
d
im
en
s
io
n
al
v
e
cto
r
in
th
e
in
p
u
t
lay
er
,
f
o
llo
wed
b
y
L
2
n
o
r
m
aliza
tio
n
.
T
h
e
f
in
al
lay
er
,
r
ef
er
r
ed
to
as
th
e
class
if
ier
lay
er
,
co
n
s
is
ts
o
f
n
eu
r
o
n
s
,
wh
ich
r
ep
r
esen
ts
th
e
s
u
b
n
etwo
r
k
'
s
in
d
ex
.
W
e
u
t
ilized
th
e
Ar
cFac
e
lo
s
s
f
u
n
ctio
n
f
o
r
tr
ain
in
g
,
em
p
lo
y
in
g
th
e
Ad
am
o
p
tim
ize
r
with
a
lear
n
in
g
r
ate
o
f
0
.
0
0
1
,
a
b
eta1
o
f
0
.
9
,
an
d
a
b
eta2
o
f
0
.
9
9
.
All scen
ar
io
s
wer
e
tr
ain
ed
f
o
r
5
0
ep
o
ch
s
,
an
d
th
e
ex
p
er
im
en
ts
wer
e
co
n
d
u
cted
with
v
ar
y
in
g
class
s
izes.
g.
Similar
ity
v
alu
e
o
r
s
u
b
n
etwo
r
k
s
o
u
tp
u
t
v
alu
e?
I
t
h
as
b
ee
n
n
o
te
d
th
at
t
h
e
d
ec
i
s
io
n
-
m
ak
in
g
u
n
it
m
u
ltip
lies
th
e
v
alu
es
en
ter
e
d
f
r
o
m
t
h
e
p
r
ev
io
u
s
u
n
its
f
o
r
ea
ch
o
f
th
ese
two
ca
n
d
id
at
es
an
d
s
elec
ts
th
e
clas
s
with
th
e
h
ig
h
er
p
r
o
d
u
ct.
Ho
wev
er
,
e
x
p
er
im
en
ts
in
d
icate
th
at
th
e
im
p
o
r
ta
n
ce
o
f
ca
n
d
id
ates
f
r
o
m
to
iler
an
d
af
f
in
ity
-
m
eter
u
n
its
m
ay
n
o
t
b
e
e
q
u
al
,
d
ep
en
d
in
g
o
n
th
e
ty
p
e
o
f
d
atasets
,
tr
ain
in
g
m
eth
o
d
o
lo
g
ies,
l
o
s
s
f
u
n
ctio
n
s
u
s
ed
in
th
e
t
r
ain
in
g
o
f
s
u
b
n
etwo
r
k
s
,
an
d
o
t
h
er
f
ac
to
r
s
.
As
th
e
ex
tr
ac
ted
f
ea
tu
r
es
b
ec
o
m
e
m
o
r
e
ac
cu
r
ate,
t
h
e
af
f
in
ity
-
m
eter
u
n
it
ca
n
d
id
ate
g
ain
s
h
i
g
h
er
im
p
o
r
ta
n
ce
in
d
ec
is
io
n
-
m
ak
in
g
.
T
h
er
ef
o
r
e,
an
ad
d
itio
n
al
co
ef
f
icien
t
ter
m
,
d
en
o
ted
as
w
,
is
in
tr
o
d
u
ce
d
to
th
e
f
o
r
m
u
la.
I
n
o
th
er
w
o
r
d
s
,
ass
u
m
in
g
:
i)
T
h
e
af
f
in
ity
-
m
eter
u
n
it
ca
n
d
id
at
e
class
is
d
en
o
ted
as
x
an
d
i
i)
T
h
e
to
iler
u
n
it
ca
n
d
id
ate
class
is
d
en
o
ted
as
y
,
T
h
e
s
elec
ted
class
will b
e
th
e
o
n
e
th
at
m
a
x
im
izes th
e
ex
p
r
es
s
io
n
(
7
)
am
o
n
g
all
s
u
b
n
etwo
r
k
s
:
(
×
×
,
×
)
,
(
7
)
wh
er
e
a
r
ep
r
esen
ts
th
e
o
u
tp
u
t
v
alu
e
o
f
x
,
b
r
ep
r
esen
ts
th
e
s
im
ilar
ity
v
alu
e
o
f
x
,
c
r
ep
r
esen
t
s
th
e
o
u
tp
u
t v
alu
e
y
,
an
d
d
r
ep
r
esen
ts
th
e
s
im
ilar
ity
v
alu
e
o
f
y
.
As
m
en
tio
n
e
d
,
w
r
ep
r
esen
ts
th
e
im
p
o
r
tan
ce
o
f
th
e
a
f
f
in
ity
-
m
ete
r
u
n
it
ca
n
d
id
ate
,
r
an
g
in
g
b
etwe
en
ze
r
o
an
d
two
.
I
f
its
v
alu
e
is
less
th
an
o
n
e,
it
s
ig
n
if
ies
a
h
ig
h
er
im
p
o
r
tan
ce
g
iv
en
to
th
e
af
f
in
ity
-
m
eter
u
n
it
ca
n
d
id
ate.
I
f
it
is
g
r
ea
ter
th
an
o
n
e,
it
in
d
icate
s
a
h
ig
h
er
i
m
p
o
r
tan
ce
g
iv
en
to
th
e
to
iler
u
n
it c
a
n
d
id
ate.
T
h
e
q
u
esti
o
n
a
r
is
es
: H
o
w
to
d
eter
m
in
e
th
e
v
alu
e
o
f
w
?
W
e
co
n
s
id
er
2
0
0
p
o
te
n
tial v
alu
es f
o
r
w
in
th
e
r
an
g
e
o
f
ze
r
o
to
two
with
in
ter
v
als
o
f
0
.
0
1
.
Su
b
s
eq
u
en
tl
y
,
we
ca
lcu
late
th
e
ac
cu
r
ac
y
f
o
r
ea
ch
o
f
th
ese
w
v
alu
es
o
n
t
h
e
v
alid
atio
n
d
ata
s
et.
T
h
e
p
o
in
t
wh
er
e
t
h
e
h
ig
h
est
ac
cu
r
ac
y
is
ac
h
iev
e
d
is
th
en
r
eg
ar
d
ed
as
t
h
e
f
in
al
v
alu
e
f
o
r
w
.
As
d
ep
icte
d
in
Fig
u
r
e
8
,
f
o
r
th
e
MS
-
C
eleb
-
1
M
d
atasets
with
2
0
,
0
0
0
ca
teg
o
r
ies
an
d
6
s
u
b
n
etwo
r
k
s
,
an
d
f
o
r
v
alid
ati
o
n
d
atasets
,
th
e
o
p
tima
l
w
is
1
.
4
,
y
ield
in
g
th
e
h
ig
h
est
ac
cu
r
ac
y
o
f
9
5
.
1
3
1
.
T
h
is
s
am
e
w
is
ap
p
lied
to
test
d
ataset
in
Fig
u
r
e
9
,
an
d
it
is
o
b
s
er
v
ed
th
at
n
ea
r
ly
o
p
tim
al
r
esu
lt
s
ar
e
ac
h
iev
ed
.
T
h
e
b
est
w
f
o
r
tes
t
d
ataset
s
i
s
1
.
3
6
(
with
an
ac
cu
r
ac
y
o
f
9
5
.
1
6
3
)
,
v
e
r
y
clo
s
e
to
o
u
r
ch
o
s
en
v
alu
e
o
f
w=
1
.
4
(
ac
cu
r
ac
y
o
f
9
5
.
1
5
6
)
.
Ad
d
iti
o
n
ally
,
th
e
f
ig
u
r
e
illu
s
tr
a
tes
t
h
at
s
o
lely
u
tili
zin
g
th
e
t
o
iler
u
n
it
cr
iter
io
n
(
i.e
.
,
w=
0
)
wo
u
ld
r
esu
lt
in
an
ac
cu
r
ac
y
o
f
9
4
.
8
3
7
,
wh
ile
u
s
in
g
o
n
ly
th
e
af
f
in
ity
-
m
eter
u
n
it
cr
iter
io
n
(
i.e
.
,
w=
2
)
wo
u
ld
y
ield
an
ac
cu
r
ac
y
o
f
9
5
.
1
1
4
.
T
h
e
c
o
m
b
in
atio
n
o
f
b
o
th
c
r
iter
ia
s
u
r
p
ass
es
th
e
p
e
r
f
o
r
m
a
n
ce
o
f
ea
ch
cr
iter
io
n
in
d
i
v
id
u
ally
.
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.
14
,
No
.
6
,
Dec
em
b
e
r
20
24
:
7
1
4
5
-
7
1
5
7
7154
Fig
u
r
e
8
.
Fin
d
i
n
g
b
est th
r
esh
o
ld
p
o
in
t f
r
o
m
v
alid
atio
n
d
ataset
with
2
0
0
0
0
class
es a
n
d
6
s
u
b
n
etwo
r
k
s
(
Gh
o
s
tFaceN
et+
MS
-
C
eleb
-
1
M)
Fig
u
r
e
9
.
Acc
u
r
ac
y
(
r
ec
all)
in
d
if
f
er
en
t
p
o
ten
tial th
r
esh
o
l
d
v
alu
es a
n
d
s
elec
ted
th
r
esh
o
ld
v
alu
e
f
o
r
test
d
ataset
(
Gh
o
s
tFaceN
et+
MS
-
C
eleb
-
1
M)
3
.
2
.
M
et
ho
d
co
m
pa
riso
n
T
h
e
f
o
llo
win
g
f
ac
e
id
e
n
tific
atio
n
s
ce
n
ar
io
s
wer
e
ev
alu
ated
in
ea
ch
ca
s
e:
a.
Dis
tr
ib
u
tio
n
-
Fre
e:
A
s
in
g
le
n
e
two
r
k
.
b.
I
SM
: M
u
ltip
le
s
u
b
n
ets in
p
ar
allel,
r
an
d
o
m
class
d
is
tr
ib
u
tio
n
b
etwe
en
s
u
b
n
ets,
an
d
th
e
m
ax
-
m
ax
p
r
ed
ictio
n
m
ec
h
an
is
m
.
c.
Ou
r
+
−
Po
s
t:
M
u
ltip
le
s
u
b
n
ets
i
n
p
ar
allel,
d
is
tr
ib
u
tio
n
class
ed
b
y
clu
s
ter
in
g
,
m
a
x
-
m
ax
d
ec
is
io
n
m
eth
o
d
.
d.
Ou
r
+
Po
s
t:
Mu
lt
ip
le
s
u
b
n
ets
in
p
ar
allel,
clu
s
ter
in
g
-
b
ased
class
d
i
s
tr
ib
u
tio
n
,
an
d
a
m
o
d
if
ied
m
ax
-
m
ax
p
r
ed
ictio
n
m
ec
h
a
n
is
m
(
p
r
o
p
o
s
ed
m
et
h
o
d
)
.
T
ab
le
3
s
u
m
m
ar
izes
th
e
r
esu
lt
s
o
f
f
ac
e
id
e
n
tific
atio
n
f
o
r
all
s
ce
n
ar
io
s
.
Dif
f
er
en
t
r
esu
lts
f
o
r
d
if
f
e
r
en
t
d
atasets
ca
n
b
e
ca
u
s
ed
b
y
d
if
f
er
en
ce
s
in
th
e
lev
el
o
f
clea
n
lin
ess
o
r
d
if
f
icu
lties
o
f
th
e
d
ata
s
et
im
ag
es.
An
o
th
e
r
r
ea
s
o
n
is
th
at
th
e
p
r
e
-
tr
ai
n
ed
m
o
d
el
was
tr
ain
ed
u
s
in
g
MS
-
C
eleb
-
1
M
d
ataset,
wh
ich
is
wh
y
it
h
as
b
etter
r
esu
lts
th
an
o
t
h
er
d
atasets
.
Of
co
u
r
s
e,
in
th
is
ca
s
e,
th
e
p
r
o
p
o
s
ed
m
eth
o
d
w
o
r
k
e
d
b
etter
th
an
o
th
e
r
m
eth
o
d
s
.
W
h
en
th
e
n
u
m
b
e
r
o
f
class
es
is
s
m
all,
th
er
e
is
a
lo
wer
r
is
k
o
f
o
v
er
f
itti
n
g
,
an
d
co
n
s
eq
u
e
n
tly
,
th
e
p
r
o
p
o
s
ed
m
o
d
el
an
d
o
th
er
d
is
tr
ib
u
tio
n
m
eth
o
d
s
m
ay
n
o
t
n
ec
ess
ar
ily
lead
to
p
er
f
o
r
m
an
ce
im
p
r
o
v
e
m
en
t.
Ho
wev
er
,
as
th
e
n
u
m
b
e
r
o
f
class
es
in
cr
ea
s
es,
th
e
p
r
o
p
o
s
ed
m
eth
o
d
co
n
s
i
s
ten
tly
o
u
tp
er
f
o
r
m
s
o
th
er
m
et
h
o
d
s
.
Acc
o
r
d
i
n
g
to
Evaluation Warning : The document was created with Spire.PDF for Python.