I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
10
,
No
.
6
,
Dec
em
b
er
2
0
2
0
,
p
p
.
5
7
5
8
~
5
7
7
1
I
SS
N:
2
0
8
8
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v
1
0
i
6
.
pp
5
7
5
8
-
5
7
7
1
575
8
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m/in
d
ex
.
p
h
p
/I
JE
C
E
G
ende
r clas
sifica
t
io
n using
custo
m
co
nv
o
lutiona
l neural
netw
o
rk
s a
rchitec
ture
F
a
dh
la
n H
a
f
izhel
m
i K
a
m
a
r
u
Z
a
m
a
n
F
a
c
u
lt
y
o
f
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
,
Un
iv
e
rsiti
T
e
k
n
o
lo
g
i
M
A
R
A
,
M
a
la
y
sia
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
9
,
2
0
1
9
R
ev
i
s
ed
Ma
y
2
0
,
2020
A
cc
ep
ted
J
u
n
11
,
2
0
2
0
G
e
n
d
e
r
c
la
ss
i
f
ica
ti
o
n
d
e
m
o
n
stra
tes
h
ig
h
a
c
c
u
ra
c
y
in
m
a
n
y
p
re
v
i
o
u
s
w
o
rk
s.
Ho
w
e
v
e
r,
it
d
o
e
s
n
o
t
g
e
n
e
ra
li
z
e
v
e
r
y
w
e
ll
in
u
n
c
o
n
stra
i
n
e
d
s
e
tt
in
g
s
a
n
d
e
n
v
iro
n
m
e
n
ts.
F
u
rth
e
rm
o
re
,
m
a
n
y
p
ro
p
o
se
d
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
r
a
l
n
e
tw
o
rk
(CNN
)
b
a
se
d
so
lu
ti
o
n
s
v
a
ry
sig
n
if
ica
n
tl
y
in
th
e
ir
c
h
a
ra
c
ter
isti
c
s
a
n
d
a
rc
h
it
e
c
tu
re
s,
w
h
ich
c
a
ll
s
f
o
r
o
p
ti
m
a
l
CNN
a
r
c
h
it
e
c
tu
re
f
o
r
th
is
sp
e
c
if
i
c
tas
k
.
In
th
is
w
o
rk
,
a
h
a
n
d
-
c
ra
f
ted
,
c
u
sto
m
CN
N
a
r
c
h
it
e
c
tu
re
is
p
ro
p
o
se
d
to
d
isti
n
g
u
is
h
b
e
tw
e
e
n
m
a
le
a
n
d
f
e
m
a
le
f
a
c
ial
i
m
a
g
e
s.
T
h
is
c
u
s
to
m
CNN
re
q
u
ires
sm
a
ll
e
r
in
p
u
t
im
a
g
e
re
so
lu
ti
o
n
s
a
n
d
sig
n
if
ica
n
tl
y
f
e
w
e
r
train
a
b
le
p
a
ra
m
e
ters
th
a
n
so
m
e
p
o
p
u
lar
sta
te
-
of
-
th
e
-
a
rts
su
c
h
a
s
G
o
o
g
leN
e
t
a
n
d
A
le
x
Ne
t.
It
a
lso
e
m
p
lo
y
s
b
a
tch
n
o
rm
a
li
z
a
ti
o
n
lay
e
rs
w
h
ich
re
su
lt
s
in
b
e
tt
e
r
c
o
m
p
u
tatio
n
e
f
f
icie
n
c
y
.
Ba
se
d
o
n
e
x
p
e
rim
e
n
ts
u
s
in
g
p
u
b
li
c
ly
a
v
a
il
a
b
le
d
a
tas
e
ts
su
c
h
a
s
L
F
W
,
Ce
l
e
b
A
a
n
d
IM
DB
-
W
IKI
d
a
tas
e
ts,
th
e
p
ro
p
o
se
d
c
u
sto
m
CNN
d
e
li
v
e
r
e
d
th
e
f
a
ste
s
t
in
f
e
re
n
c
e
ti
m
e
in
a
ll
tes
ts,
w
h
e
r
e
it
n
e
e
d
s
o
n
ly
0
.
9
2
m
s
to
c
las
sify
1
2
0
0
im
a
g
e
s
o
n
G
P
U,
1
.
7
9
m
s
o
n
CP
U,
a
n
d
2
.
5
1
m
s
o
n
VP
U.
T
h
e
c
u
sto
m
C
NN
a
lso
d
e
li
v
e
rs
p
e
rf
o
r
m
a
n
c
e
o
n
-
p
a
r
w
it
h
sta
te
-
of
-
th
e
-
a
rts
a
n
d
e
v
e
n
su
rp
a
ss
e
d
th
e
s
e
m
e
th
o
d
s
in
Ce
leb
A
g
e
n
d
e
r
c
las
sif
ic
a
ti
o
n
w
h
e
re
it
d
e
li
v
e
re
d
th
e
b
e
st
re
su
lt
a
t
9
6
%
a
c
c
u
ra
c
y
.
M
o
re
o
v
e
r,
in
a
m
o
re
c
h
a
ll
e
n
g
in
g
c
ro
ss
-
d
a
tas
e
t
in
f
e
re
n
c
e
,
c
u
sto
m
CNN
train
e
d
u
sin
g
Ce
leb
A
d
a
tas
e
t
g
iv
e
s
th
e
b
e
st
g
e
n
d
e
r
c
la
ss
if
ic
a
ti
o
n
a
c
c
u
ra
c
y
f
o
r
tes
ts
o
n
IM
DB
a
n
d
W
IKI d
a
tas
e
ts at
9
7
%
a
n
d
9
6
%
a
c
c
u
ra
c
y
re
sp
e
c
ti
v
e
l
y
.
K
ey
w
o
r
d
s
:
C
NN
ar
ch
i
tectu
r
e
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
C
r
o
s
s
-
d
ataset
i
n
f
er
e
n
ce
Dee
p
lear
n
in
g
Gen
d
er
class
i
f
icatio
n
Co
p
y
rig
h
t
©
2
0
2
0
In
stit
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
F
ad
h
lan
Haf
iz
h
el
m
i K
a
m
ar
u
Z
a
m
an
,
Fac
u
lt
y
o
f
E
lectr
ical
E
n
g
in
ee
r
in
g
,
Un
i
v
er
s
iti T
ek
n
o
lo
g
i M
A
R
A
,
4
0
4
5
0
,
Sh
ah
A
la
m
,
Sela
n
g
o
r
,
Ma
la
y
s
ia
.
E
m
ail:
f
ad
h
la
n
@
u
i
t
m
.
ed
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
Hu
m
an
f
ac
ia
l
a
n
al
y
s
is
h
a
s
b
e
co
m
e
o
n
e
o
f
t
h
e
m
o
s
t
s
i
g
n
if
i
ca
n
t
ta
s
k
s
i
n
co
m
p
u
ter
v
is
io
n
,
s
in
ce
it
p
lay
s
a
v
ital
r
o
le
i
n
s
o
cial
in
t
er
ac
tio
n
s
.
L
i
k
e
o
t
h
er
tas
k
s
s
u
ch
a
s
t
h
e
c
h
ar
ac
ter
izatio
n
o
f
ag
e,
g
en
d
er
,
f
ac
ial
attr
ib
u
tes,
e
x
p
r
ess
io
n
s
,
a
n
d
p
er
s
o
n
alit
y
,
au
to
m
a
tic
g
e
n
d
er
class
i
f
icatio
n
h
as
v
ar
io
u
s
i
m
p
o
r
tan
t
ap
p
licatio
n
s
s
u
c
h
as
in
telli
g
en
t
u
s
er
i
n
te
r
f
ac
es,
u
s
er
id
en
ti
f
icatio
n
,
s
o
cial
in
ter
ac
tio
n
,
v
i
s
u
a
l
s
u
r
v
eillan
ce
,
co
llecti
n
g
d
em
o
g
r
ap
h
ic
s
tatis
tics
f
o
r
m
ar
k
eti
n
g
,
b
eh
a
v
io
u
r
r
ec
o
g
n
it
io
n
an
d
s
o
o
n
.
T
h
er
ef
o
r
e,
m
an
y
r
e
s
ea
r
ch
e
f
f
o
r
ts
h
a
v
e
b
ee
n
d
ev
o
ted
to
d
esig
n
au
to
m
ated
s
y
s
te
m
w
h
ich
ca
n
clas
s
if
y
g
e
n
d
er
s
[
1
-
4
]
.
A
lt
h
o
u
g
h
th
is
ta
s
k
h
as
b
ee
n
lar
g
el
y
ad
d
r
ess
ed
in
t
h
e
p
ast,
th
e
r
ep
o
r
ted
p
e
r
f
o
r
m
a
n
ce
s
ar
e
f
ar
f
r
o
m
o
p
ti
m
al
esp
ec
ial
l
y
u
n
d
er
u
n
co
n
s
tr
ai
n
ed
co
n
d
itio
n
s
[
5
,
6
]
.
M
o
r
eo
v
er
,
t
h
e
co
m
p
lex
it
y
o
f
th
is
tas
k
lar
g
el
y
d
ep
en
d
s
o
n
th
e
co
n
tex
t
o
f
th
e
ap
p
licatio
n
an
d
tr
ain
i
n
g
p
r
o
to
co
ls
.
Gen
d
er
cla
s
s
i
f
icatio
n
m
o
d
el
ca
n
b
e
tr
ai
n
ed
an
d
test
ed
f
r
o
m
t
h
e
s
a
m
e
d
ataset,
also
k
n
o
w
n
as
i
n
-
d
atase
t
i
n
f
er
e
n
ce
,
o
r
f
r
o
m
d
if
f
er
en
t
d
ataset,
a
ls
o
k
n
o
w
n
as
cr
o
s
s
-
d
ata
s
et
i
n
f
er
en
ce
.
B
esid
es,
f
ac
ia
l
i
m
a
g
es
u
s
ed
th
e
s
e
d
atasets
c
an
b
e
ca
p
tu
r
ed
u
n
d
er
co
n
tr
o
lled
o
r
u
n
co
n
tr
o
lled
/u
n
co
n
s
tr
ain
ed
en
v
ir
o
n
m
en
t
w
h
ic
h
w
ill
in
cr
ea
s
e
t
h
e
co
m
p
lex
it
y
o
f
th
e
tas
k
.
On
e
o
f
s
tate
-
of
-
t
h
e
-
ar
t
i
n
g
e
n
d
er
class
if
ica
tio
n
is
o
b
ta
in
ed
b
y
J
ia
an
d
C
r
is
tia
n
i
n
i
w
h
er
e
th
e
y
u
s
ed
4
m
il
lio
n
i
m
a
g
es
to
tr
ai
n
th
e
ir
m
et
h
o
d
ca
lled
C
-
P
eg
as
o
s
[
7
]
an
d
test
ed
it
u
s
i
n
g
cr
o
s
s
-
d
ata
s
et
in
f
er
e
n
ce
s
tr
ateg
y
o
n
u
n
co
n
s
tr
ain
ed
L
FW
d
ataset.
Ho
w
ev
er
,
m
o
r
e
r
ec
en
tl
y
,
Af
i
f
i
an
d
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:
2
0
8
8
-
8708
Gen
d
er c
la
s
s
ifica
tio
n
u
s
in
g
cu
s
to
m
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k…
(
F
a
d
h
la
n
Ha
fiz
h
elmi
K
a
ma
r
u
Za
ma
n
)
5759
A
b
d
elh
a
m
ed
[
8
]
p
er
f
o
r
m
ed
s
i
m
ilar
cr
o
s
s
-
d
ata
s
et
te
s
ts
an
d
b
ased
o
n
t
h
eir
r
e
s
u
l
ts
,
it
ca
n
b
e
o
b
s
er
v
ed
t
h
at
p
o
o
r
er
class
if
icatio
n
p
er
f
o
r
m
an
ce
m
a
y
b
e
o
b
tain
ed
a
s
a
r
e
s
u
lt,
i
n
w
h
ic
h
ac
co
r
d
in
g
to
th
e
m
i
s
d
u
e
to
d
i
f
f
er
en
t
c
o
n
d
i
t
i
o
n
s
o
f
c
o
l
l
e
c
t
i
n
g
i
m
a
g
e
s
i
n
d
i
f
f
e
r
e
n
t
d
a
t
a
s
e
t
s
,
s
u
c
h
a
s
o
c
c
l
u
s
i
o
n
s
,
i
l
l
u
m
i
n
a
t
i
o
n
c
h
a
n
g
e
s
,
b
a
c
k
g
r
o
u
n
d
s
,
e
t
c
.
R
ec
en
t
l
y
,
d
ee
p
n
e
u
r
al
n
et
w
o
r
k
s
,
m
o
r
e
s
p
ec
i
f
icall
y
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
(
C
N
N)
[
9
]
h
as
b
ec
o
m
e
t
h
e
g
o
ld
en
s
ta
n
d
ar
d
f
o
r
o
b
j
ec
t
r
ec
o
g
n
itio
n
.
C
NN
h
av
e
i
m
p
r
o
v
ed
n
ea
r
l
y
all
ar
ea
s
o
f
co
m
p
u
ter
v
i
s
io
n
in
cl
u
d
in
g
h
u
m
a
n
a
ctio
n
r
ec
o
g
n
itio
n
[
1
0
]
,
h
an
d
-
w
r
i
tten
d
ig
it
r
ec
o
g
n
itio
n
[
1
1
]
,
f
ac
e
v
er
if
icatio
n
a
n
d
class
i
f
icatio
n
[
1
2
-
1
4
]
an
d
f
ac
e
d
etec
tio
n
[
1
5
]
.
Ho
w
ev
er
,
t
h
er
e
ar
e
t
w
o
p
r
o
b
lem
s
ass
o
ci
ated
w
it
h
C
NN
i
n
p
ar
ticu
lar
,
w
h
ic
h
is
(
1
)
th
e
e
n
o
r
m
o
u
s
s
ize
o
f
d
ata
r
eq
u
ir
e
d
to
tr
ain
t
h
e
n
et
w
o
r
k
s
u
ch
as
in
[
1
2
]
,
an
d
(
2
)
th
e
m
e
m
o
r
y
r
eq
u
ir
e
m
en
t
o
f
t
h
e
n
et
w
o
r
k
d
u
e
to
co
m
p
u
tat
io
n
o
f
m
ass
iv
e
p
ar
a
m
eter
s
o
f
te
n
li
m
it
s
th
e
ap
p
licatio
n
o
f
C
NN
o
n
em
b
ed
d
ed
p
lat
f
o
r
m
s
s
u
ch
as
i
n
m
o
b
ile
p
h
o
n
e
s
,
a
s
w
ell
as
o
n
clo
u
d
s
er
v
ice
s
.
Fo
r
ex
a
m
p
le,
t
w
o
s
tate
-
o
f
-
th
e
-
ar
ts
C
NN
ar
c
h
itect
u
r
e
ca
lled
Go
o
g
leNe
t
[
1
6
]
an
d
A
le
x
Net
[
1
7
]
b
o
th
co
n
tain
s
6
,
7
9
9
,
7
0
0
an
d
6
2
,
3
7
8
,
3
4
4
p
a
r
a
m
eter
s
r
esp
ec
tiv
el
y
.
An
o
t
h
er
ex
a
m
p
le,
a
1
6
-
la
y
er
C
NN
d
escr
ib
ed
in
[
1
8
]
h
a
s
a
w
ei
g
h
t
f
i
le
b
ig
g
er
t
h
a
n
5
0
0
MB
an
d
r
eq
u
ir
es
ab
o
u
t
3
.
1
×1
0
10
f
lo
atin
g
o
p
er
atio
n
s
p
er
i
m
ag
e.
T
h
u
s
,
C
N
N
ca
n
b
e
r
eg
ar
d
ed
as
a
h
i
g
h
-
ca
p
ac
i
t
y
cla
s
s
i
f
ier
h
a
v
i
n
g
v
er
y
lar
g
e
n
u
m
b
er
s
o
f
tr
ai
n
ab
le
p
ar
am
eter
s
th
a
t
r
eq
u
ir
e
s
C
NN
to
lear
n
f
r
o
m
lar
g
er
d
atasets
[
1
6
]
d
u
e
to
d
if
f
ic
u
lt
p
r
o
ce
s
s
o
f
t
u
n
i
n
g
a
n
d
esti
m
ati
n
g
ea
ch
p
ar
a
m
eter
f
r
o
m
s
m
al
l
n
u
m
b
er
o
f
s
a
m
p
le
s
.
T
o
r
ed
u
ce
th
e
ef
f
ec
t
o
f
t
h
ese
li
m
itatio
n
s
,
w
e
ca
n
o
p
ti
m
ize
t
h
e
C
NN
to
r
ed
u
ce
th
e
co
m
p
lex
i
t
y
o
f
th
e
la
y
er
s
b
y
e
m
p
lo
y
i
n
g
s
e
v
er
al
ap
p
r
o
ac
h
es
s
u
ch
as
eit
h
er
b
y
r
ed
u
c
in
g
t
h
e
n
u
m
b
er
o
f
co
n
v
o
lu
tio
n
al
la
y
er
s
,
a
n
d
/o
r
r
ed
u
ce
th
e
n
u
m
b
er
o
f
n
e
u
r
o
n
s
in
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
a
n
d
/o
r
r
ed
u
ce
th
e
r
eso
l
u
tio
n
o
f
th
e
i
n
p
u
t
i
m
ag
e
s
.
Ho
w
e
v
er
,
it
m
u
s
t
b
e
d
o
n
e
ca
r
ef
u
ll
y
a
s
to
e
n
s
u
r
e
th
at
th
e
r
es
u
lti
n
g
ar
ch
itect
u
r
e
ca
n
s
ti
ll
lear
n
t
h
e
task
at
h
an
d
,
e.
g
.
g
en
d
er
cl
ass
i
f
icatio
n
,
an
d
g
e
n
er
alize
well
o
n
u
n
s
ee
n
d
ata.
T
h
e
im
p
r
o
v
e
m
e
n
t in
co
m
p
u
tat
io
n
s
h
o
u
ld
n
o
t c
o
m
p
r
o
m
i
s
e
th
e
ac
cu
r
ac
y
o
f
t
h
e
clas
s
i
f
icatio
n
.
I
n
th
is
w
o
r
k
,
t
h
e
p
r
o
b
lem
o
f
g
e
n
d
er
class
i
f
icatio
n
a
n
d
h
ig
h
co
m
p
le
x
it
y
o
f
ex
iti
n
g
d
ee
p
n
eu
r
al
n
et
w
o
r
k
s
is
ad
d
r
ess
ed
,
b
y
f
o
cu
s
i
n
g
o
n
r
ed
u
cin
g
t
h
e
co
m
p
lex
it
y
o
f
t
h
e
C
NN
a
n
d
to
i
m
p
r
o
v
e
th
e
m
e
m
o
r
y
r
eq
u
ir
e
m
en
t
as
w
ell
a
s
t
h
e
ti
m
e
r
eq
u
ir
ed
f
o
r
n
et
w
o
r
k
in
f
e
r
en
ce
.
I
n
p
ar
ticu
lar
,
th
e
g
o
al
o
f
th
is
p
ap
er
is
to
p
r
o
p
o
s
e
a
co
m
p
lete
d
esig
n
o
f
a
lo
w
-
co
m
p
le
x
it
y
h
an
d
-
cr
af
ted
C
NN,
w
h
er
e
th
is
n
e
t
w
o
r
k
w
ill
b
e
test
ed
o
n
g
en
d
er
cl
ass
if
icatio
n
ta
s
k
u
n
d
er
a
v
er
y
ch
alle
n
g
i
n
g
u
n
co
n
s
tr
ain
ed
co
n
d
itio
n
s
as
well
as
u
n
d
er
g
o
in
g
ex
p
er
i
m
e
n
t
u
s
i
n
g
cr
o
s
s
-
d
atas
et
in
f
er
en
ce
i
m
p
le
m
e
n
tatio
n
s
.
T
h
is
r
elativ
el
y
s
i
m
p
ler
a
n
d
m
i
n
i
m
ized
m
o
d
el
ac
h
iev
ed
s
tate
-
of
-
th
e
-
ar
ts
p
er
f
o
r
m
a
n
ce
an
d
s
h
o
w
s
a
s
i
g
n
i
f
ica
n
t
b
o
o
s
t
in
in
f
er
e
n
ce
ti
m
e
w
h
e
n
co
m
p
ar
ed
ag
ain
s
t
s
e
v
er
al
ex
i
s
ti
n
g
C
N
N
ar
ch
itectu
r
e
s
.
Ho
w
e
v
er
,
A
h
an
d
-
cr
a
f
ted
ar
ch
itect
u
r
e
is
a
v
er
y
ch
alle
n
g
i
n
g
,
ti
m
e
-
co
n
s
u
m
i
n
g
an
d
r
eq
u
ir
e
ex
p
er
t
k
n
o
w
led
g
e
d
u
e
to
a
lar
g
e
n
u
m
b
er
o
f
ar
ch
itect
u
r
al
ch
o
ice
s
[
1
9
]
.
T
h
is
p
r
o
p
o
s
ed
si
m
p
li
f
ied
C
NN
ca
n
lear
n
f
r
o
m
r
elati
v
el
y
s
m
aller
d
atase
t
an
d
p
er
f
o
r
m
clas
s
i
f
icatio
n
o
n
a
lar
g
er
d
ataset.
On
e
o
f
t
h
e
i
m
p
o
r
tan
t
w
o
r
k
o
n
g
e
n
d
er
class
i
f
icat
io
n
ca
l
led
F
ac
e
T
r
ac
er
[
2
0
]
em
p
lo
y
s
co
m
b
in
atio
n
o
f
A
d
ab
o
o
s
t
an
d
S
u
p
p
o
r
t
Vec
to
r
Ma
ch
i
n
es
t
h
at
s
elec
t
a
n
d
tr
ai
n
o
n
t
h
e
o
p
ti
m
al
s
et
o
f
f
ea
t
u
r
es
f
o
r
ea
ch
attr
ib
u
te
b
ased
o
n
t
h
e
s
alie
n
t
s
tr
u
ct
u
r
e
o
f
f
ac
es.
Si
m
ilar
l
y
,
P
A
N
D
A
-
w
an
d
P
A
ND
A
-
1
[
2
]
co
m
b
i
n
e
s
d
ee
p
lear
n
i
n
g
a
n
d
p
ar
t
-
b
ased
m
o
d
els
b
y
tr
ai
n
i
n
g
p
o
s
e
-
n
o
r
m
a
lized
C
NN
s
to
cla
s
s
i
f
y
v
ar
io
u
s
h
u
m
a
n
attr
ib
u
te
s
s
u
c
h
as
h
ai
r
s
t
y
le,
g
en
d
er
,
ex
p
r
es
s
io
n
,
clo
t
h
es
s
t
y
le,
etc
f
r
o
m
t
h
at
w
o
r
k
s
well
f
o
r
i
m
ag
e
s
u
n
d
er
lar
g
e
v
ar
iab
ilit
y
o
f
p
o
s
e,
ap
p
ea
r
an
ce
,
v
ie
w
p
o
i
n
t,
o
cc
lu
s
i
o
n
,
an
d
ar
ticu
latio
n
.
L
i
u
et
al.
,
[
2
1
]
p
r
o
p
o
s
ed
ca
s
ca
d
e
s
o
f
d
u
al
C
NNs,
ca
lled
L
Net
an
d
A
Ne
t.
T
h
ese
C
NNs
ar
e
p
r
e
-
tr
ain
ed
i
n
d
i
f
f
e
r
en
t
s
e
s
s
io
n
s
b
u
t
j
o
in
tl
y
f
in
e
-
t
u
n
ed
w
i
th
attr
ib
u
te
tag
s
.
ANe
t
is
p
r
e
-
tr
ai
n
ed
f
o
r
attr
ib
u
te
p
r
ed
ictio
n
u
s
i
n
g
h
u
g
e
f
ac
e
id
en
titi
e
s
,
w
h
er
ea
s
L
N
et
is
p
r
e
-
tr
ai
n
ed
f
o
r
f
ac
e
lo
ca
lizatio
n
u
s
i
n
g
h
u
g
e
g
en
er
al
o
b
j
ec
t
ca
teg
o
r
ies.
T
h
is
ap
p
r
o
ac
h
s
u
r
p
as
s
ed
th
e
s
ta
te
-
of
-
th
e
-
ar
t
b
y
a
co
n
s
id
er
ab
le
m
ar
g
i
n
a
n
d
d
is
clo
s
es
i
m
p
o
r
tan
t
f
ac
ts
o
n
lear
n
in
g
f
ac
e
r
ep
r
esen
tatio
n
.
B
y
co
m
b
i
n
in
g
th
e
i
n
ter
m
ed
iate
la
y
er
s
o
f
a
d
ee
p
C
NN
u
s
i
n
g
a
s
ep
ar
ate
C
NN
f
o
llo
w
ed
b
y
a
m
u
lti
-
ta
s
k
lear
n
in
g
al
g
o
r
ith
m
,
H
y
p
er
f
ac
e
[
2
2
]
b
o
o
s
ts
u
p
th
eir
in
d
i
v
id
u
al
p
er
f
o
r
m
a
n
ce
s
b
y
ex
p
lo
iti
n
g
th
e
s
y
n
er
g
y
am
o
n
g
t
h
e
tas
k
s
.
T
h
e
au
th
o
r
s
s
h
o
w
ed
th
at
H
y
p
er
f
ac
e
is
ab
le
to
ex
tr
ac
t
b
o
th
h
o
lis
tic
a
n
d
lo
ca
l
in
f
o
r
m
a
tio
n
in
f
ac
es
a
n
d
th
u
s
o
u
tp
er
f
o
r
m
s
m
an
y
co
m
p
e
titi
v
e
al
g
o
r
ith
m
s
f
o
r
f
ac
e
d
ete
ctio
n
,
p
o
s
e
es
ti
m
atio
n
,
g
e
n
d
er
r
ec
o
g
n
itio
n
a
n
d
lan
d
m
ar
k
s
lo
ca
lizatio
n
.
J
ia
an
d
C
r
is
tian
in
i
[
7
]
p
r
esen
ted
a
s
i
m
p
le
y
et
e
f
f
ec
ti
v
e
class
if
ier
o
f
f
ac
e
i
m
a
g
es
ca
l
led
C
-
P
eg
a
s
o
s
w
h
ic
h
i
s
g
en
er
ate
d
b
y
tr
ain
i
n
g
a
lin
ea
r
c
lass
if
i
ca
tio
n
al
g
o
r
ith
m
o
n
a
m
as
s
i
v
e
d
ataset
w
h
ic
h
i
s
au
to
m
at
icall
y
a
s
s
e
m
b
led
an
d
lab
elled
.
T
h
ey
u
s
ed
f
o
u
r
m
il
lio
n
i
m
a
g
es
a
n
d
m
o
r
e
th
a
n
6
0
,
0
0
0
f
ea
tu
r
es
to
tr
ai
n
th
ese
clas
s
i
f
ier
s
.
B
y
e
m
p
lo
y
in
g
li
n
ea
r
class
i
f
ier
s
en
s
e
m
b
les,
w
h
e
n
test
ed
an
L
FW
d
ataset,
C
-
P
eg
a
s
o
s
ac
h
iev
ed
a
n
ac
cu
r
ac
y
o
f
9
6
.
8
6
%.
R
ec
en
t
l
y
,
Af
i
f
i
a
n
d
A
b
d
elh
a
m
e
d
[
8
]
p
r
o
p
o
s
ed
an
ap
p
r
o
ac
h
b
ased
o
n
t
h
e
b
eh
a
v
io
u
r
o
f
h
u
m
a
n
s
i
n
g
en
d
er
clas
s
i
f
icatio
n
.
T
h
e
y
r
el
y
o
n
f
o
g
g
y
f
ac
e
w
h
ich
co
m
b
i
n
ed
is
o
lated
f
ac
ial
f
ea
t
u
r
es
a
n
d
a
h
o
lis
tic
f
ea
t
u
r
e,
in
s
tead
o
f
d
ea
li
n
g
w
it
h
t
h
e
f
a
ce
i
m
a
g
e
as
a
s
o
le
f
ea
t
u
r
e.
T
h
en
,
th
e
y
u
s
e
f
o
g
g
y
f
ac
e
to
tr
ai
n
a
C
N
N
f
o
llo
w
ed
b
y
s
co
r
e
f
u
s
io
n
b
ased
o
n
A
d
a
B
o
o
s
t
to
class
if
y
t
h
e
g
e
n
d
er
class
.
An
tip
o
v
et
al
.,
[
2
3
]
s
u
g
g
ested
an
e
n
s
e
m
b
le
m
o
d
el
o
f
C
NN
to
e
n
h
a
n
ce
th
e
g
en
d
er
cla
s
s
i
f
icatio
n
a
cc
u
r
ac
y
f
r
o
m
f
ac
ial
i
m
ag
e
s
in
L
FW
d
atase
t.
T
h
eir
en
s
e
m
b
le
m
o
d
el
i
s
p
u
r
p
o
s
el
y
d
esi
g
n
ed
i
n
s
u
c
h
a
w
a
y
t
h
at
m
i
n
i
m
ized
t
h
e
m
e
m
o
r
y
r
eq
u
ir
e
m
en
ts
a
n
d
co
m
p
u
tatio
n
ti
m
e.
L
ik
e
w
is
e,
l
o
ca
l d
ee
p
n
eu
r
al
n
et
w
o
r
k
(
L
o
c
al
-
DNN)
[
3
]
is
p
r
o
p
o
s
ed
f
o
r
g
en
d
er
class
i
f
icat
io
n
w
h
er
e
it
is
r
elie
s
o
n
t
w
o
f
u
n
d
am
e
n
tal
id
ea
s
:
d
ee
p
ar
ch
i
tec
tu
r
es
a
n
d
lo
ca
l
f
e
at
u
r
es.
o
v
er
lap
p
in
g
r
e
g
io
n
s
i
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
6
,
Dec
em
b
er
2
0
2
0
:
5
7
5
8
-
5
7
7
1
5760
th
e
v
is
u
al
f
ie
ld
is
u
s
ed
to
tr
ain
th
e
m
o
d
el
b
y
d
is
cr
i
m
in
a
tiv
e
f
ee
d
-
f
o
r
w
ar
d
n
et
w
o
r
k
s
b
u
ilt
u
s
in
g
m
u
l
tip
le
la
y
er
s
.
T
h
e
au
th
o
r
s
s
h
o
w
ed
th
at
L
o
ca
l
-
DNN
o
u
tp
er
f
o
r
m
ed
o
th
er
d
ee
p
-
lear
n
i
n
g
-
b
ased
m
et
h
o
d
s
a
n
d
attai
n
s
s
tate
-
of
-
th
e
-
ar
t r
e
s
u
lt
s
i
n
m
u
ltip
le
b
en
c
h
m
ar
k
s
.
O
n
t
h
e
o
t
h
e
r
h
a
n
d
,
s
e
v
e
r
a
l
w
o
r
k
s
a
r
e
m
o
r
e
f
o
c
u
s
e
d
o
n
d
e
s
i
g
n
i
n
g
c
u
s
t
o
m
o
r
h
a
n
d
-
c
r
a
f
t
e
d
a
r
c
h
i
t
e
c
t
u
r
a
l
d
e
s
i
g
n
o
f
C
N
N
.
M
o
s
t
o
f
t
h
e
s
e
w
o
r
k
s
d
i
s
c
u
s
s
e
d
o
n
C
N
N
d
e
s
i
g
n
i
s
s
u
e
s
s
u
c
h
a
s
h
y
p
e
r
p
a
r
a
m
e
t
e
r
s
[
1
9
,
2
4
]
,
n
e
w
o
p
ti
m
izatio
n
o
f
o
b
j
ec
tiv
e
f
u
n
ctio
n
[
1
9
]
,
i
m
p
r
o
v
ed
tr
i
p
let
lo
s
s
f
u
n
ctio
n
[
2
5
]
,
s
tr
u
ct
u
r
e
co
m
p
r
es
s
io
n
o
f
C
NN
[
2
6
]
an
d
esti
m
ati
n
g
C
N
N
ar
ch
itect
u
r
es c
o
m
p
le
x
it
y
[
2
7
]
.
T
h
ese
m
et
h
o
d
s
s
h
ar
e
s
i
m
il
ar
o
b
j
ec
tiv
e
w
h
ic
h
i
s
t
o
f
i
n
d
t
h
e
m
o
s
t
o
p
t
i
m
i
z
e
d
h
a
n
d
-
c
r
a
f
t
e
d
C
N
N
a
r
c
h
i
t
e
c
t
u
r
e
t
h
a
t
p
r
o
d
u
c
e
s
a
h
i
g
h
r
e
c
o
g
n
i
t
i
o
n
p
e
r
f
o
r
m
a
n
c
e
w
h
i
l
e
m
a
i
n
t
a
i
n
i
n
g
i
t
s
c
o
m
p
l
e
x
i
t
y
t
o
b
e
a
s
l
o
w
a
s
p
o
s
s
i
b
l
e
t
o
a
l
l
o
w
f
a
s
t
e
r
c
o
m
p
u
t
a
t
i
o
n
s
a
n
d
l
e
s
s
m
em
o
r
y
r
e
q
u
i
r
e
m
e
n
t
.
A
d
d
itio
n
al
l
y
,
t
h
e
r
esear
ch
co
m
m
u
n
it
y
is
a
ls
o
ad
d
r
ess
in
g
a
m
u
c
h
m
o
r
e
r
ea
lis
tic
g
en
er
al
p
r
o
b
lem
f
o
r
g
en
d
er
c
la
s
s
i
f
icat
io
n
w
h
ic
h
is
g
en
d
er
cla
s
s
i
f
icat
io
n
f
r
o
m
f
ac
i
al
i
m
ag
e
s
i
n
th
e
w
ild
.
Ma
n
y
r
esear
ch
er
s
ar
e
n
o
w
f
o
cu
s
ed
o
n
ex
p
er
i
m
e
n
t
s
in
v
o
lv
i
n
g
r
ec
e
n
t
an
d
b
ig
g
er
d
ata
b
ases
th
at
e
n
co
m
p
a
s
s
m
o
r
e
v
ar
iatio
n
s
in
c
lu
d
i
n
g
id
en
tit
y
,
et
h
n
icit
y
,
ag
e,
illu
m
i
n
atio
n
s
,
i
m
a
g
e
r
eso
l
u
t
io
n
s
a
n
d
p
o
s
e
v
ar
iatio
n
s
.
T
h
is
h
as
ca
lled
f
o
r
s
o
lu
tio
n
s
o
n
h
o
w
to
ac
q
u
ir
e
a
m
o
d
el
th
a
t
ca
n
g
en
er
al
ize
w
e
ll
o
n
a
d
ataset
an
d
g
i
v
e
s
g
o
o
d
in
f
er
e
n
ce
p
er
f
o
r
m
a
n
ce
o
n
a
co
m
p
lete
l
y
n
e
w
u
n
s
ee
n
d
ata
s
et.
I
n
d
o
in
g
s
o
,
is
v
er
y
i
m
p
o
r
tan
t
to
en
s
u
r
e
t
h
e
m
o
d
el
d
id
n
o
t
r
e
q
u
ir
e
co
n
s
tan
t
r
etr
ain
in
g
an
d
ca
n
w
o
r
k
w
ell
in
ch
alle
n
g
i
n
g
,
u
n
co
n
s
tr
ai
n
ed
co
n
d
itio
n
s
.
T
h
u
s
,
a
cr
o
s
s
-
d
ataset
test
s
h
av
e
b
ee
n
ad
o
p
ted
p
r
ev
io
u
s
l
y
i
n
[
1
,
7
,
8
,
2
3
]
to
m
ea
s
u
r
e
t
h
e
p
er
f
o
r
m
an
ce
o
f
g
e
n
d
er
class
i
f
ier
o
n
n
e
w
d
ataset
s
t
h
at
p
r
esen
t
t
h
i
s
t
y
p
e
o
f
ch
a
lle
n
g
e.
T
h
e
r
est
o
f
t
h
i
s
p
ap
er
is
ar
r
an
g
ed
as
f
o
llo
w
s
.
I
n
Sectio
n
2
th
e
p
r
o
p
o
s
ed
cu
s
to
m
C
NN
is
ex
p
lai
n
ed
,
w
h
ic
h
ca
n
b
e
u
s
ed
to
au
to
m
atica
ll
y
c
la
s
s
i
f
y
g
en
d
er
s
.
Sectio
n
3
d
escr
ib
es
an
d
an
al
y
ze
s
th
e
r
es
u
lts
o
n
th
e
p
u
b
lic
l
y
a
v
ailab
le
d
atasets
s
u
c
h
as
L
F
W
,
C
eleb
A
a
n
d
W
I
KI
-
IM
DB
d
atasets
.
Fin
all
y
,
Sectio
n
4
d
r
a
w
s
s
o
m
e
co
n
cl
u
s
io
n
s
an
d
d
is
c
u
s
s
es
f
u
tu
r
e
w
o
r
k
.
2.
RE
S
E
ARCH
M
E
T
H
O
D
T
h
e
co
n
v
o
lu
t
io
n
al
n
eu
r
al
n
et
wo
r
k
ar
ch
itect
u
r
e
ad
o
p
ted
in
th
i
s
w
o
r
k
i
n
te
n
d
s
to
r
ed
u
ce
t
h
e
c
o
m
p
le
x
it
y
b
y
o
p
ti
m
izi
n
g
t
h
e
co
n
v
o
lu
ti
o
n
al
la
y
er
s
,
r
ed
u
ci
n
g
th
e
n
u
m
b
er
o
f
n
eu
r
o
n
s
i
n
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
an
d
r
ed
u
cin
g
th
e
r
eso
lu
tio
n
o
f
t
h
e
in
p
u
t
i
m
a
g
es.
I
n
d
esi
g
n
in
g
th
i
s
ar
ch
itect
u
r
e,
th
e
n
et
w
o
r
k
s
h
o
u
ld
s
till
b
e
ab
le
to
d
eliv
er
s
ta
te
-
of
-
th
e
-
ar
ts
p
er
f
o
r
m
an
ce
w
h
ile
m
ai
n
tai
n
i
n
g
a
v
e
r
y
o
p
ti
m
al
s
ize.
I
n
d
o
in
g
so
,
A
lex
Net
ar
ch
itect
u
r
e
is
u
s
ed
as
t
h
e
s
tar
ti
n
g
n
et
w
o
r
k
te
m
p
late.
A
le
x
Net
ar
ch
itec
t
u
r
e
h
ad
a
v
er
y
s
i
m
ilar
ar
ch
i
te
ctu
r
e
as
L
e
Net
b
y
L
eC
u
n
et
al
.,
[
2
8
]
an
d
it
co
n
tain
s
ei
g
h
t
m
ai
n
la
y
er
s
w
h
er
e
th
e
f
ir
s
t
f
i
v
e
ar
e
co
n
v
o
lu
tio
n
al
lay
er
s
.
So
m
e
o
f
th
e
co
n
v
o
l
u
tio
n
al
la
y
er
s
ar
e
f
o
llo
w
ed
b
y
Ma
x
P
o
o
lin
g
la
y
e
r
s
,
an
d
t
h
e
f
i
n
al
t
h
r
ee
la
y
er
s
ar
e
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
.
Firstl
y
,
th
e
i
n
p
u
t
is
r
ed
u
ce
d
in
to
s
m
al
ler
in
p
u
t
i
m
ag
e
r
eso
lu
tio
n
f
r
o
m
t
h
e
o
r
ig
in
al
2
2
7
×2
7
7
t
o
6
4
×
6
4
.
So
m
e
p
r
ev
io
u
s
w
o
r
k
s
o
n
f
ac
e
r
ec
o
g
n
it
io
n
s
h
o
w
ed
th
at
t
h
i
s
i
m
a
g
e
r
eso
l
u
tio
n
is
s
u
f
f
ic
ie
n
t
to
ac
h
ie
v
e
g
o
o
d
r
esu
lt
s
[
2
9
]
.
B
y
ch
a
n
g
in
g
t
h
e
s
ize
o
f
i
m
a
g
e
in
p
u
t
la
y
er
,
th
e
s
ize
o
f
s
u
b
s
eq
u
e
n
t
co
n
v
o
l
u
tio
n
la
y
er
s
n
ee
d
to
b
e
alter
ed
to
o
.
Fo
r
m
o
r
e
s
tab
le
an
d
i
m
p
r
o
v
ed
lear
n
in
g
,
2
lay
er
s
o
f
co
n
v
o
lu
tio
n
la
y
er
h
a
v
e
b
ee
n
ad
d
ed
to
th
is
n
et
w
o
r
k
.
Fi
n
all
y
,
to
r
ed
u
ce
th
e
tr
ain
ab
le
p
ar
am
eter
s
f
u
r
t
h
er
,
th
e
n
u
m
b
er
o
f
n
eu
r
o
n
s
ar
e
r
ed
u
ce
d
in
all
f
u
l
l
y
co
n
n
ec
ted
(
FC
)
la
y
er
s
.
T
h
e
n
u
m
b
er
o
f
n
e
u
r
o
n
s
i
s
later
r
ed
u
ce
d
in
th
i
s
cu
s
to
m
C
N
N
f
r
o
m
4
0
9
6
,
4
0
9
6
,
an
d
1
0
0
0
n
eu
r
o
n
s
to
1
0
0
,
5
0
an
d
2
n
e
u
r
o
n
s
at
ea
ch
FC
la
y
er
,
r
es
p
ec
tiv
el
y
.
I
t
is
i
m
p
o
r
tan
t
to
n
o
te
th
at
th
e
f
i
n
al
FC
la
y
er
ac
ts
as so
f
t
m
a
x
la
y
er
.
T
o
s
p
ee
d
u
p
th
e
tr
ain
in
g
,
th
e
n
o
r
m
aliza
t
io
n
la
y
er
s
i
n
th
e
n
e
t
w
o
r
k
w
h
ic
h
is
b
ased
o
n
lo
ca
l
r
esp
o
n
s
e
n
o
r
m
aliza
t
io
n
(
L
R
N)
ar
e
r
ep
lace
d
to
b
atch
n
o
r
m
aliza
tio
n
l
a
y
er
s
[
3
0
]
.
T
h
e
b
atch
n
o
r
m
a
li
za
tio
n
la
y
er
m
ak
e
s
n
o
r
m
aliza
t
io
n
a
p
ar
t
o
f
th
e
m
o
d
el
ar
ch
itectu
r
e
an
d
it
p
er
f
o
r
m
s
th
e
n
o
r
m
aliza
tio
n
f
o
r
ea
ch
tr
ain
i
n
g
m
i
n
i
b
atch
.
I
t
allo
w
s
th
e
n
et
w
o
r
k
to
b
e
tr
a
in
ed
a
h
i
g
h
er
lear
n
i
n
g
r
ate
s
th
at
is
o
th
er
w
i
s
e
d
if
f
ic
u
l
t
a
n
d
h
i
g
h
l
y
u
n
s
tab
le
w
it
h
n
o
r
m
al
L
R
N
la
y
er
s
.
B
atch
n
o
r
m
aliza
tio
n
h
a
s
b
ee
n
s
h
o
w
n
t
o
ac
h
iev
e
t
h
e
s
a
m
e
s
tate
-
of
-
t
h
e
-
ar
ts
p
er
f
o
r
m
an
ce
w
it
h
1
4
ti
m
e
s
f
e
w
er
tr
ain
in
g
it
er
atio
n
s
[
3
0
]
.
T
h
e
o
v
er
v
ie
w
o
f
th
e
ar
c
h
itect
u
r
e
o
f
th
e
p
r
o
p
o
s
ed
cu
s
to
m
C
NN
i
s
s
h
o
w
n
in
F
ig
u
r
e
1.
Su
b
s
eq
u
e
n
tl
y
,
t
h
e
la
y
er
s
,
s
iz
e
o
f
k
er
n
els,
n
u
m
b
er
o
f
k
er
n
els,
a
n
d
s
tr
id
es
f
o
r
ea
ch
la
y
er
u
s
ed
in
cu
s
to
m
C
N
N
ar
ch
itect
u
r
e
is
g
i
v
en
i
n
T
ab
le
1
.
A
cc
o
r
d
in
g
to
T
ab
le
1
,
in
to
tal,
th
e
c
u
s
to
m
C
NN
h
a
s
7
co
n
v
o
lu
tio
n
la
y
er
s
an
d
6
b
atch
n
o
r
m
al
izatio
n
la
y
er
s
i
n
b
etw
ee
n
o
f
ea
ch
co
n
v
o
l
u
tio
n
la
y
er
s
.
I
n
b
et
w
ee
n
o
f
ea
ch
co
n
v
o
l
u
tio
n
la
y
er
s
,
Ma
x
P
o
o
lin
g
la
y
er
is
ad
d
e
d
to
r
ed
u
ce
th
e
s
p
atial
d
i
m
e
n
s
io
n
o
f
t
h
e
in
p
u
t
v
o
lu
m
e
f
o
r
n
ex
t
la
y
er
s
.
T
h
e
ac
ti
v
atio
n
f
u
n
ct
io
n
u
s
ed
i
n
t
h
i
s
n
et
w
o
r
k
is
r
ec
ti
f
ied
li
n
ea
r
u
n
it
(
R
e
L
u
)
.
Sev
er
al
d
r
o
p
o
u
t
la
y
er
s
ar
e
al
s
o
ad
d
ed
w
ith
t
h
e
R
e
L
u
to
p
r
ev
e
n
t
o
v
er
f
itt
in
g
.
T
o
p
r
o
v
e
th
at
t
h
i
s
c
u
s
to
m
C
N
N
p
o
s
s
es
s
es
s
m
aller
n
u
m
b
er
o
f
p
ar
a
m
eter
s
as
co
m
p
ar
ed
to
s
e
v
er
al
e
x
is
ti
n
g
C
NN,
t
h
e
p
ar
a
m
eter
s
i
n
cu
s
t
o
m
C
NN
d
esi
g
n
i
s
co
m
p
u
ted
.
Her
e
it
is
s
h
o
w
n
i
n
d
etail
h
o
w
to
co
m
p
u
te
t
h
e
s
ize
o
f
o
u
tp
u
t
f
ea
tu
r
e
s
o
f
ea
c
h
co
n
v
o
lu
tio
n
la
y
er
an
d
h
o
w
to
co
m
p
u
te
t
h
e
n
u
m
b
er
o
f
p
a
r
a
m
eter
s
a
s
s
o
ciat
ed
w
i
th
co
n
v
o
lu
tio
n
la
y
er
s
a
n
d
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
o
f
a
C
NN.
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:
2
0
8
8
-
8708
Gen
d
er c
la
s
s
ifica
tio
n
u
s
in
g
cu
s
to
m
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k…
(
F
a
d
h
la
n
Ha
fiz
h
elmi
K
a
ma
r
u
Za
ma
n
)
5761
Fig
u
r
e
1
.
T
h
e
ar
ch
itectu
r
e
o
f
t
h
e
p
r
o
p
o
s
ed
cu
s
to
m
C
NN,
s
h
o
w
i
n
g
7
co
n
v
o
l
u
tio
n
la
y
er
s
,
2
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
an
d
a
s
o
f
t
m
ax
la
y
er
.
Fo
r
clar
ity
,
R
e
L
u
,
Ma
x
P
o
o
lin
g
,
d
r
o
p
o
u
t
an
d
b
atch
n
o
r
m
a
lizati
o
n
la
y
er
s
ar
e
n
o
t sh
o
w
n
.
Size
o
f
o
u
tp
u
t
f
ea
t
u
r
es a
n
d
ch
a
n
n
els ar
e
i
n
d
icate
d
in
th
e
f
i
g
u
r
e
u
s
i
n
g
th
is
f
o
r
m
at
:
o
u
tp
u
t wi
d
th
× o
u
tp
u
t h
eig
h
t ×
o
u
tp
u
t c
h
a
n
n
els
T
ab
le
1
.
T
h
e
lay
er
s
,
s
ize
o
f
k
e
r
n
els,
n
u
m
b
er
o
f
k
er
n
e
ls
,
s
tr
id
es
an
d
p
ad
d
i
ng
f
o
r
ea
ch
la
y
er
L
a
y
e
r
N
a
me
Ty
p
e
K
e
r
n
e
l
si
z
e
,
n
u
m
b
e
r
s o
f
k
e
r
n
e
l
s,
st
r
i
d
e
s a
n
d
p
a
d
d
i
n
g
1
'i
mag
e
i
n
p
u
t
'
I
mag
e
I
n
p
u
t
6
4
x
6
4
x
1
i
mag
e
s w
i
t
h
'z
e
r
o
c
e
n
t
r
e
' n
o
r
mal
i
z
a
t
i
o
n
2
'co
n
v
_
1
'
C
o
n
v
o
l
u
t
i
o
n
8
3
x
3
x
1
c
o
n
v
o
l
u
t
i
o
n
s
w
i
t
h
s
t
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
0
0
0
0
]
3
'b
a
t
c
h
n
o
r
m
_
1
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
8
c
h
a
n
n
e
l
s
4
'r
e
l
u
_
1
'
R
e
L
U
R
e
L
U
5
'max
p
o
o
l
_
1
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
0
0
0
0
]
6
'co
n
v
_
2
'
C
o
n
v
o
l
u
t
i
o
n
1
6
3
x
3
x
8
c
o
n
v
o
l
u
t
i
o
n
s w
i
t
h
s
t
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
0
0
0
0
]
7
'b
a
t
c
h
n
o
r
m
_
2
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
1
6
c
h
a
n
n
e
l
s
8
'r
e
l
u
_
2
'
R
e
L
U
R
e
L
U
9
'max
p
o
o
l
_
2
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
0
0
0
0
]
1
0
'co
n
v
_
3
'
C
o
n
v
o
l
u
t
i
o
n
3
2
3
x
3
x
1
6
c
o
n
v
o
l
u
t
i
o
n
s w
i
t
h
st
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
0
0
0
0
]
1
1
'b
a
t
c
h
n
o
r
m
_
3
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
3
2
c
h
a
n
n
e
l
s
12
'r
e
l
u
_
3
'
R
e
L
U
R
e
L
U
1
3
'max
p
o
o
l
_
3
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
0
0
0
0
]
1
4
'co
n
v
_
4
'
C
o
n
v
o
l
u
t
i
o
n
6
4
3
x
3
x
3
2
c
o
n
v
o
l
u
t
i
o
n
s w
i
t
h
st
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
15
'b
a
t
c
h
n
o
r
m
_
4
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
6
4
c
h
a
n
n
e
l
s
16
'r
e
l
u
_
4
'
R
e
L
U
R
e
L
U
1
7
'max
p
o
o
l
_
4
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
1
8
'co
n
v
_
5
'
C
o
n
v
o
l
u
t
i
o
n
1
2
8
3
x
3
x
6
4
c
o
n
v
o
l
u
t
i
o
n
s
w
i
t
h
s
t
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
1
9
'b
a
t
c
h
n
o
r
m
_
5
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
1
2
8
c
h
a
n
n
e
l
s
20
'r
e
l
u
_
5
'
R
e
L
U
R
e
L
U
2
1
'max
p
o
o
l
_
5
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
2
2
'co
n
v
_
6
'
C
o
n
v
o
l
u
t
i
o
n
2
5
6
3
x
3
x
1
2
8
c
o
n
v
o
l
u
t
i
o
n
s w
i
t
h
s
t
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
2
3
'b
a
t
c
h
n
o
r
m
_
6
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
2
5
6
c
h
a
n
n
e
l
s
24
'r
e
l
u
_
6
'
R
e
L
U
R
e
L
U
2
5
'max
p
o
o
l
_
6
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
2
6
'co
n
v
_
7
'
C
o
n
v
o
l
u
t
i
o
n
5
1
2
3
x
3
x
2
5
6
c
o
n
v
o
l
u
t
i
o
n
s w
i
t
h
s
t
r
i
d
e
[
1
1
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
2
7
'b
a
t
c
h
n
o
r
m
_
7
'
B
a
t
c
h
N
o
r
m
B
a
t
c
h
n
o
r
mal
i
z
a
t
i
o
n
w
i
t
h
5
1
2
c
h
a
n
n
e
l
s
28
'r
e
l
u
_
7
'
R
e
L
U
R
e
L
U
2
9
'max
p
o
o
l
_
7
'
M
a
x
P
o
o
l
i
n
g
2
x
2
m
a
x
p
o
o
l
i
n
g
w
i
t
h
s
t
r
i
d
e
[
2
2
]
a
n
d
p
a
d
d
i
n
g
[
1
1
1
1
]
3
0
'd
r
o
p
o
u
t
_
1
'
D
r
o
p
o
u
t
5
0
%
d
r
o
p
o
u
t
3
1
'f
c
_
1
'
F
C
1
0
0
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
32
'r
e
l
u
_
8
'
R
e
L
U
R
e
L
U
3
3
'd
r
o
p
o
u
t
_
2
'
D
r
o
p
o
u
t
5
0
%
d
r
o
p
o
u
t
3
4
'f
c
_
2
'
FC
5
0
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
35
'r
e
l
u
_
9
'
R
e
L
U
R
e
L
U
3
6
'd
r
o
p
o
u
t
_
3
'
D
r
o
p
o
u
t
5
0
%
d
r
o
p
o
u
t
3
7
'f
c
_
3
'
FC
2
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
3
8
'so
f
t
max
'
S
o
f
t
max
so
f
t
max
3
9
'cl
a
sso
u
t
p
u
t
'
C
l
a
ssi
f
i
c
a
t
i
o
n
c
r
o
sse
n
t
r
o
p
y
e
x
w
i
t
h
‘
mal
e
’
a
n
d
1
o
t
h
e
r
c
l
a
ss
2
.
1
.
Size
o
f
t
he
o
utput
f
ea
t
ures o
f
a
co
nv
o
lutio
n la
y
er
a
n
d
m
a
x
po
o
lin
g
la
y
er
T
o
co
m
p
u
te
th
e
n
u
m
b
er
o
f
o
u
tp
u
t
f
ea
t
u
r
es
o
f
a
C
NN,
let’
s
d
en
o
te
th
e
w
id
t
h
o
f
o
u
tp
u
t
i
m
ag
e
as
,
th
e
w
id
th
o
f
in
p
u
t
i
m
ag
e
as
,
t
h
e
w
id
t
h
o
f
co
n
v
o
lu
tio
n
k
er
n
e
ls
la
y
er
as
,
th
e
n
u
m
b
er
o
f
k
er
n
els
u
s
ed
as
,
th
e
s
tr
id
e
o
f
th
e
co
n
v
o
lu
tio
n
as
,
th
e
p
ad
d
in
g
as
an
d
th
e
p
o
o
l
s
ize
as
.
T
h
e
s
ize
o
f
th
e
o
u
tp
u
t
o
f
a
co
n
v
o
lu
t
io
n
la
y
er
an
d
Ma
x
P
o
o
lin
g
la
y
er
ar
e
g
iv
en
b
y
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
6
,
Dec
em
b
er
2
0
2
0
:
5
7
5
8
-
5
7
7
1
5762
=
−
+
2
+
1
(
1
)
=
−
+
1
(
2
)
2
.
2
.
P
a
ra
m
et
er
s
o
f
a
co
nv
o
lutio
n la
y
er
I
n
a
C
NN,
t
h
er
e
ar
e
t
w
o
p
ar
a
m
eter
s
f
o
r
ea
c
h
la
y
er
,
n
a
m
el
y
th
e
w
ei
g
h
ts
an
d
b
ia
s
es.
T
h
e
to
tal
n
u
m
b
er
o
f
p
ar
a
m
eter
s
i
s
t
h
e
s
u
m
o
f
al
l
w
ei
g
h
t
s
a
n
d
b
iase
s
.
L
et
t
h
e
n
u
m
b
er
o
f
w
ei
g
h
ts
an
d
b
iases
o
f
th
e
co
n
v
o
lu
t
io
n
la
y
er
b
e
=
2
an
d
=
r
esp
ec
tiv
el
y
,
w
h
er
e
is
th
e
n
u
m
b
er
o
f
ch
an
n
el
s
o
f
th
e
i
n
p
u
t
i
m
ag
e.
T
h
u
s
,
w
e
ca
n
co
m
p
u
te
th
e
n
u
m
b
er
o
f
p
ar
a
m
eter
s
o
f
t
h
e
co
n
v
o
lu
tio
n
la
y
er
:
=
+
(
3
)
2
.
3
.
P
a
ra
m
et
er
s
o
f
a
n F
C
la
y
er
co
nn
ec
t
ed
t
o
a
c
o
nv
o
luti
o
n la
y
er
L
et
t
h
e
n
u
m
b
er
o
f
w
ei
g
h
t
s
a
n
d
b
ias o
f
a
FC
L
a
y
er
w
h
ic
h
is
c
o
n
n
ec
ted
to
a
co
n
v
o
l
u
tio
n
la
y
er
d
en
o
ted
as
=
2
an
d
=
r
esp
ec
tiv
el
y
,
w
h
er
e
is
th
e
n
u
m
b
er
o
f
n
e
u
r
o
n
s
i
n
th
e
FC
L
a
y
er
an
d
is
th
e
s
ize
o
f
th
e
o
u
tp
u
t
i
m
a
g
e
o
f
th
e
p
r
ev
io
u
s
co
n
v
o
lu
ti
o
n
la
y
er
.
T
h
u
s
,
t
h
e
n
u
m
b
er
o
f
p
ar
am
e
ter
s
o
f
th
e
co
n
v
o
lu
tio
n
la
y
er
ca
n
b
e
co
m
p
u
ted
f
r
o
m
:
=
+
(
4
)
2
.
4
.
P
a
ra
m
et
er
s
o
f
a
n F
C
la
y
er
co
nn
ec
t
ed
t
o
a
no
t
her
F
C
la
y
er
L
et
t
h
e
n
u
m
b
er
o
f
w
ei
g
h
ts
an
d
b
iases
o
f
a
FC
L
a
y
er
w
h
ic
h
is
co
n
n
ec
ted
to
an
F
C
L
a
y
er
b
e
=
−
1
an
d
=
r
esp
ec
tiv
el
y
,
w
h
er
e
−
1
is
th
e
n
u
m
b
er
o
f
n
e
u
r
o
n
s
in
t
h
e
p
r
ev
io
u
s
F
C
L
a
y
er
.
T
h
u
s
,
w
e
ca
n
co
m
p
u
te
th
e
n
u
m
b
er
o
f
p
ar
a
m
eter
s
o
f
t
h
e
co
n
v
o
lu
tio
n
la
y
er
:
=
+
(
5
)
F
in
a
l
ly
,
t
o
t
al
n
u
m
b
e
r
o
f
p
a
r
am
e
t
e
r
s
=
+
+
c
an
b
e
d
et
e
r
m
in
e
d
.
I
t
is
im
p
o
r
t
an
t
t
o
n
o
t
e
t
h
at
th
e
r
e
a
r
e
n
o
p
a
r
am
et
e
r
s
as
s
o
ci
a
t
e
d
w
i
th
a
p
o
o
l
in
g
,
d
r
o
p
o
u
t
,
an
d
R
eL
u
l
ay
e
r
s
.
T
h
e
p
o
o
l
s
iz
e
,
s
t
r
i
d
e
,
an
d
p
a
d
d
i
n
g
a
r
e
a
l
l
c
o
n
s
i
d
e
r
e
d
a
s
h
y
p
e
r
p
a
r
am
e
t
e
r
s
w
h
o
s
e
v
a
lu
e
i
s
s
et
b
e
f
o
r
e
th
e
l
e
a
r
n
in
g
p
r
o
c
e
s
s
b
eg
in
s
.
A
l
th
o
u
g
h
b
e
t
t
e
r
r
e
s
u
l
t
s
c
an
b
e
a
ch
i
ev
e
d
w
h
en
h
y
p
e
r
p
a
r
am
e
te
r
s
a
r
e
p
r
o
p
e
r
ly
a
d
ju
s
t
e
d
[
2
4
]
,
th
e
s
e
h
y
p
e
r
p
a
r
am
e
t
e
r
s
a
r
e
c
o
n
s
i
d
e
r
e
d
n
o
n
-
t
r
a
in
a
b
l
e
a
n
d
ar
e
e
x
t
e
r
n
a
l
t
o
th
e
m
o
d
el
.
T
h
r
o
u
g
h
o
u
t
th
i
s
w
o
r
k
,
t
h
e
c
o
m
p
u
t
er
s
y
s
t
em
th
a
t
is
u
s
e
d
i
n
th
e
ex
p
e
r
im
en
t
s
r
u
n
s
o
n
I
n
t
e
l
i
7
-
6
7
0
0
C
PU
@
3
.
4
0
G
Hz
,
w
i
th
1
6
G
B
o
f
R
A
M
an
d
u
s
e
s
G
T
X
1
0
8
0
T
i
a
s
t
h
e
m
a
in
G
PU
.
T
h
e
h
y
p
e
r
p
a
r
am
e
te
r
s
u
s
e
d
i
n
t
r
a
in
i
n
g
th
e
cu
s
t
o
m
C
N
N
a
r
e
as
f
o
l
l
o
w
s
:
m
o
m
en
tu
m
=
0
.
9
,
m
in
i
b
a
t
c
h
s
i
ze
=
5
0
0
,
L
2
r
e
g
u
l
a
r
i
z
at
i
o
n
=
0
.
0
0
1
,
i
n
i
t
i
al
l
e
a
r
n
in
g
r
ate
=
0
.
1
,
l
e
a
r
n
r
a
t
e
d
r
o
p
f
a
c
t
o
r
=
0
.
9
,
a
n
d
l
e
a
r
n
r
at
e
d
r
o
p
p
e
r
i
o
d
=
1
0
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
I
n
th
i
s
s
ec
tio
n
,
th
e
r
es
u
lt
s
o
f
g
en
d
er
clas
s
if
icatio
n
e
x
p
er
im
en
ts
o
n
s
ev
er
al
d
ataset
s
ar
e
p
r
esen
ted
.
C
r
itical
d
is
c
u
s
s
io
n
s
o
n
th
e
p
e
r
f
o
r
m
an
ce
o
f
p
r
o
p
o
s
ed
cu
s
to
m
C
N
N
is
p
r
esen
ted
i
n
ter
m
s
o
f
its
ac
cu
r
ac
y
a
n
d
in
f
er
en
ce
s
p
ee
d
o
n
GP
U,
C
P
U
an
d
an
e
m
b
ed
d
ed
s
y
s
te
m
.
T
h
e
p
er
f
o
r
m
an
ce
in
g
en
d
er
r
ec
o
g
n
itio
n
is
al
s
o
co
m
p
ar
ed
ag
ai
n
s
t
s
tate
-
of
-
th
e
-
ar
ts
m
eth
o
d
s
s
u
c
h
as
A
lex
N
et
an
d
Go
o
g
leNe
t
to
h
ig
h
li
g
h
t
th
e
s
u
p
er
io
r
it
y
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
B
esid
es,
th
e
cu
s
to
m
C
N
N
is
also
t
ested
u
n
d
er
a
cr
o
s
s
-
d
ataset
i
n
f
er
e
n
ce
co
n
s
tr
ai
n
t,
w
h
er
e
cu
s
to
m
C
NN
tr
ai
n
ed
u
s
in
g
C
eleb
A
d
ataset
i
s
test
ed
o
n
I
MD
B
an
d
W
I
KI
d
atasets
.
3
.
1
.
Da
t
a
s
et
s
des
cr
iptio
n
I
n
th
e
e
x
p
er
i
m
e
n
ts
s
ev
er
al
p
u
b
licl
y
av
a
ilab
le
d
atasets
n
a
m
e
l
y
t
h
e
lab
elled
f
ac
es
i
n
t
h
e
W
ild
(
L
FW
)
d
ataset,
C
eleb
Face
s
Attr
ib
u
t
es
Data
s
et
(
C
eleb
A
)
d
ataset
an
d
I
MD
B
-
W
I
KI
d
ataset
a
r
e
u
s
ed
.
T
h
e
L
FW
d
ataset
[
3
1
]
co
n
tain
s
1
3
,
3
2
3
p
h
o
to
s
o
f
5
,
7
4
9
ce
leb
r
ities
ta
k
en
u
n
d
er
u
n
co
n
s
tr
ain
ed
e
n
v
i
r
o
n
m
e
n
ts
w
h
ic
h
ar
e
th
en
d
i
v
id
ed
in
to
1
0
-
f
o
ld
cr
o
s
s
v
alid
atio
n
.
E
ac
h
f
o
ld
co
n
tain
s
b
o
th
m
ale
a
n
d
f
e
m
ale
i
m
a
g
es,
as
s
u
g
g
e
s
ted
b
y
th
e
r
es
tr
icted
p
r
o
to
co
l.
P
er
f
o
r
m
an
ce
is
m
ea
s
u
r
ed
u
s
in
g
th
e
r
estricte
d
p
r
o
to
co
l,
in
w
h
ich
o
n
l
y
g
e
n
d
er
lab
els ar
e
av
ailab
le
in
tr
ai
n
i
n
g
.
L
FW
g
en
d
er
lab
els
u
s
ed
in
t
h
is
w
o
r
k
ar
e
d
eter
m
i
n
ed
b
y
Af
if
i
an
d
A
b
d
elh
a
m
ed
[
8
]
,
w
h
er
e
v
alu
e
s
o
f
attr
ib
u
tes
s
u
g
g
e
s
ted
b
y
Ku
m
ar
et
al.
,
[
3
2
]
ar
e
u
s
ed
to
lab
el
th
e
i
m
ag
e
s
b
ased
o
n
g
e
n
d
er
.
Su
b
s
eq
u
e
n
tl
y
,
t
h
e
y
r
e
m
o
v
e
in
co
r
r
ec
t
lab
els
b
y
m
a
n
u
al
l
y
r
ev
ie
w
i
n
g
ea
c
h
ca
te
g
o
r
y
o
f
m
ale
an
d
f
e
m
ale
i
m
ag
e
s
th
r
ee
ti
m
e
s
.
I
n
th
i
s
p
ap
er
,
a
v
ar
ian
t
o
f
L
FW
d
ataset
ca
lle
d
L
FW
A
[
3
3
]
w
h
ich
co
n
tai
n
s
th
e
s
a
m
e
i
m
a
g
e
s
av
ailab
le
i
n
t
h
e
o
r
i
g
in
a
l
L
F
W
d
ataset
is
u
s
ed
,
h
o
w
e
v
er
,
i
m
a
g
es
in
L
FW
A
d
atase
t
ar
e
alig
n
ed
u
s
i
n
g
a
co
m
m
er
cial
f
ac
e
ali
g
n
m
e
n
t s
o
f
t
w
ar
e.
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:
2
0
8
8
-
8708
Gen
d
er c
la
s
s
ifica
tio
n
u
s
in
g
cu
s
to
m
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k…
(
F
a
d
h
la
n
Ha
fiz
h
elmi
K
a
ma
r
u
Za
ma
n
)
5763
C
eleb
A
[
3
4
]
is
a
lar
g
e
-
s
ca
le
d
ataset
w
it
h
lar
g
e
f
ac
ial
d
i
v
er
s
i
ties
,
h
u
g
e
q
u
a
n
tit
ies,
a
n
d
co
m
p
r
eh
en
s
i
v
e
an
n
o
tatio
n
s
t
h
at
h
as
m
o
r
e
t
h
an
2
0
2
,
5
9
9
im
a
g
es
f
r
o
m
1
0
,
1
7
7
id
en
titi
es,
h
a
v
i
n
g
4
0
b
in
ar
y
attr
ib
u
tes
an
n
o
tatio
n
s
a
n
d
5
la
n
d
m
ar
k
l
o
ca
tio
n
s
f
o
r
ea
ch
i
m
a
g
e.
T
h
e
i
m
a
g
es
in
th
is
d
atase
t
also
c
o
n
tain
s
b
ac
k
g
r
o
u
n
d
clu
tter
an
d
p
o
s
e
v
ar
iatio
n
s
.
I
MD
B
-
W
I
KI
d
ataset
[
3
5
,
3
6
]
co
n
tain
s
5
2
4
,
2
3
0
im
a
g
es
w
h
i
ch
m
ad
e
it
o
n
e
o
f
th
e
lar
g
e
s
t
p
u
b
lic
f
ac
e
d
ataset
av
ailab
le.
T
h
ese
f
ac
e
i
m
a
g
es
ar
e
cr
a
w
led
f
r
o
m
I
MD
B
an
d
W
ik
ip
ed
ia
w
eb
s
i
tes.
T
h
is
d
ataset
in
to
tal
co
n
tai
n
s
4
6
0
,
7
2
3
f
ac
e
i
m
ag
e
s
f
r
o
m
2
0
,
2
8
4
ce
leb
r
ities
f
r
o
m
I
MD
b
an
d
6
2
,
3
2
8
f
r
o
m
W
ik
ip
ed
ia
w
it
h
g
e
n
d
er
an
d
ag
e
lab
els
ar
e
s
u
p
p
lied
f
o
r
tr
ain
in
g
.
Fo
r
I
MD
B
-
W
I
KI
d
ataset,
o
n
l
y
p
h
o
to
s
th
at
h
av
e
t
h
e
s
ec
o
n
d
s
tr
o
n
g
es
t
f
a
ce
d
etec
tio
n
b
elo
w
a
ce
r
tain
th
r
es
h
o
ld
ar
e
c
h
o
s
en
,
th
u
s
t
h
e
to
tal
n
u
m
b
er
o
f
i
m
a
g
es
u
s
ed
i
n
th
is
w
o
r
k
f
r
o
m
th
i
s
d
ataset
ar
e
3
3
,
1
8
1
i
m
ag
e
s
a
n
d
3
,
2
1
0
im
ag
e
s
f
o
r
I
MD
B
an
d
W
I
KI
r
esp
ec
tiv
el
y
.
I
n
all
ex
p
er
i
m
en
t
s
,
all
f
ac
e
i
m
ag
e
s
ar
e
ali
g
n
ed
,
cr
o
p
p
e
d
an
d
r
esized
to
6
4
×
6
4
.
So
m
e
ex
a
m
p
les o
f
th
e
i
m
a
g
es
f
r
o
m
L
FW
,
C
eleb
A
,
I
MD
B
-
W
I
KI
d
atasets
ar
e
s
h
o
w
n
i
n
Fig
u
r
e
2
.
Su
b
s
eq
u
e
n
tl
y
,
t
h
e
n
u
m
b
er
o
f
m
ale
an
d
f
e
m
ale
i
m
a
g
e
s
f
o
r
ea
ch
d
ataset
ar
e
s
u
m
m
ar
ized
in
T
ab
le
2
.
Fig
u
r
e
2
.
Sa
m
p
le
o
f
m
ale
(
o
n
t
h
e
lef
t)
an
d
f
e
m
ale
i
m
ag
e
s
(
o
n
th
e
r
ig
h
t)
f
r
o
m
L
FW
an
d
C
eleb
A
d
ataset
r
esp
ec
tiv
e
l
y
T
ab
le
2
.
Nu
m
b
er
o
f
f
e
m
ale
an
d
m
ale
i
m
a
g
es c
o
n
s
i
s
t in
t
h
e
d
atasets
u
s
ed
in
t
h
i
s
w
o
r
k
D
a
t
a
se
t
s
F
e
mal
e
i
mag
e
s #
M
a
l
e
i
mag
e
s #
L
F
W
2
,
9
6
6
1
0
,
2
6
8
C
e
l
e
b
A
1
1
8
,
1
6
5
8
4
,
4
3
4
I
M
D
B
1
8
,
8
0
3
1
4
,
3
7
8
W
I
K
I
1
,
3
7
9
1
,
8
3
1
3
.
2
.
P
a
ra
m
et
er
s
o
f
cus
t
o
m
CNN
a
rc
hite
ct
ure
Firstl
y
,
th
e
co
m
p
le
x
it
y
o
f
t
h
e
cu
s
to
m
C
NN
ar
c
h
itect
u
r
e
ar
e
co
m
p
ar
ed
ag
ai
n
s
t
Go
o
g
l
eNe
t
an
d
A
le
x
Net
ar
ch
itectu
r
e
b
y
co
m
p
u
tin
g
th
e
n
u
m
b
er
o
f
p
ar
a
m
e
ter
s
.
T
h
e
f
ea
tu
r
e
s
ize
(
o
u
tp
u
t
)
o
f
a
co
n
v
o
lu
tio
n
la
y
er
an
d
Ma
x
P
o
o
lin
g
la
y
er
ar
e
co
m
p
u
ted
u
s
i
n
g
(
1
)
an
d
(
2
)
.
A
f
ter
w
a
r
d
,
t
h
e
n
u
m
b
er
o
f
p
ar
am
eter
s
o
f
a
co
n
v
o
lu
tio
n
la
y
er
,
n
u
m
b
er
o
f
p
ar
am
eter
s
o
f
a
n
FC
la
y
er
co
n
n
ec
ted
to
a
c
o
n
v
o
lu
t
io
n
la
y
er
,
an
d
n
u
m
b
er
o
f
p
ar
a
m
eter
s
o
f
a
n
F
C
la
y
er
co
n
n
ec
ted
to
a
n
o
th
er
FC
la
y
er
ar
e
ca
lcu
lated
u
s
i
n
g
(
3
)
,
(
4
)
an
d
(
5
)
r
esp
ec
tiv
el
y
.
T
h
ese
p
ar
a
m
eter
s
ar
e
g
iv
e
n
i
n
T
ab
le
3
.
B
ased
o
n
T
ab
le
3
,
th
e
to
tal
n
u
m
b
er
o
f
p
ar
a
m
e
ter
s
o
f
c
u
s
to
m
C
NN
is
2
,
0
4
1
,
7
9
6
p
ar
a
m
eter
s
.
I
t is
in
ter
e
s
ti
n
g
to
n
o
te
t
h
at
9
5
% o
f
t
h
e
p
ar
a
m
eter
s
co
m
e
s
f
r
o
m
t
h
e
6
t
h
a
n
d
7
th
co
n
v
o
lu
t
io
n
la
y
er
,
a
n
d
t
h
e
f
ir
s
t
FC
la
y
er
.
T
h
is
n
u
m
b
er
o
f
p
ar
am
eter
s
is
s
u
b
s
eq
u
e
n
tl
y
co
m
p
ar
e
d
ag
a
in
s
t
t
h
e
n
u
m
b
er
o
f
p
ar
a
m
eter
s
o
f
Go
o
g
leNe
t a
n
d
A
le
x
Net
ar
ch
i
t
ec
tu
r
e,
w
h
ic
h
is
g
i
v
en
i
n
T
ab
le
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
6
,
Dec
em
b
er
2
0
2
0
:
5
7
5
8
-
5
7
7
1
5764
T
ab
le
3
.
Nu
m
b
er
o
f
p
ar
am
eter
s
f
o
r
ea
ch
la
y
er
in
t
h
e
p
r
o
p
o
s
e
d
cu
s
to
m
C
NN
L
a
y
e
r
F
e
a
t
u
r
e
S
i
z
e
W
e
i
g
h
t
s
B
i
a
se
s
P
a
r
a
me
t
e
r
s
I
n
p
u
t
I
mag
e
6
4
×
6
4
×
3
0
0
0
C
o
n
v
-
1
6
2
×
6
2
×
8
2
1
6
8
2
2
4
B
a
t
c
h
N
o
r
m
6
2
×
6
2
×
8
0
16
16
C
o
n
v
-
2
2
9
×
2
9
×
1
6
1
1
5
2
16
1
1
6
8
B
a
t
c
h
N
o
r
m
2
9
×
2
9
×
1
6
0
32
32
C
o
n
v
-
3
1
3
×
1
3
×
3
2
4
6
0
8
32
4
6
4
0
B
a
t
c
h
N
o
r
m
1
3
×
1
3
×
3
2
0
64
64
C
o
n
v
-
4
7
×
7
×
6
4
1
8
4
3
2
64
1
8
4
9
6
B
a
t
c
h
N
o
r
m
7
×
7
×
6
4
0
1
2
8
1
2
8
C
o
n
v
-
5
5
×
5
×
1
2
8
7
3
7
2
8
1
2
8
7
3
8
5
6
B
a
t
c
h
N
o
r
m
5
×
5
×
1
2
8
0
2
5
6
2
5
6
C
o
n
v
-
6
4
×
4
×
2
5
6
2
9
4
9
1
2
2
5
6
2
9
5
1
6
8
B
a
t
c
h
N
o
r
m
4
×
4
×
2
5
6
0
5
1
2
5
1
2
C
o
n
v
-
7
3
×
3
×
5
1
2
1
1
7
9
6
4
8
5
1
2
1
1
8
0
1
6
0
B
a
t
c
h
N
o
r
m
3
×
3
×
5
1
2
0
0
1
0
2
4
FC
-
1
1
0
0
×
1
4
6
0
8
0
0
1
0
0
4
6
0
9
0
0
FC
-
2
5
0
×
1
5
0
0
0
50
5
0
5
0
S
o
f
t
max
2
×
1
1
0
0
2
1
0
2
T
o
t
a
1
P
a
r
a
me
t
e
r
s
#
2
0
4
1
7
9
6
T
ab
le
4
.
C
o
m
p
ar
is
o
n
o
n
n
u
m
b
er
o
f
p
ar
am
eter
s
o
f
cu
s
to
m
C
NN
ag
ai
n
s
t G
o
o
g
le
Net
an
d
A
l
ex
Net
C
N
N
M
o
d
e
l
s
I
mag
e
I
n
p
u
t
S
i
z
e
M
i
n
i
B
a
t
c
h
S
i
z
e
M
a
i
n
L
a
y
e
r
s
P
a
r
a
me
t
e
r
s
c
u
s
t
o
m C
N
N
6
4
×
64
5
0
0
7
c
o
n
v
,
3
F
C
l
a
y
e
r
s
2
,
0
4
1
,
7
9
6
G
o
o
g
l
e
N
e
t
2
2
4
×
224
30
2
2
c
o
n
v
w
i
t
h
i
n
c
e
p
t
i
o
n
l
a
y
e
r
s
6
,
7
9
9
,
7
0
0
A
l
e
x
N
e
t
2
2
7
×
227
30
5
c
o
n
v
,
3
F
C
l
a
y
e
r
s
6
2
,
3
7
8
,
3
4
4
A
cc
o
r
d
in
g
to
T
a
b
le
4
,
A
lex
Net
h
as
th
e
lar
g
e
s
t
n
u
m
b
er
o
f
p
ar
am
eter
s
w
it
h
m
o
r
e
th
an
6
2
M
p
ar
am
eter
s
w
h
ile
Go
o
g
leNe
t
h
as
6
.
8
M
p
ar
am
eter
s
.
T
h
is
r
en
d
er
s
A
le
x
Net
to
r
eq
u
ir
e
lar
g
er
m
e
m
o
r
y
s
p
ac
e
d
u
r
in
g
tr
ai
n
in
g
a
s
co
m
p
ar
ed
to
o
th
er
m
eth
o
d
s
.
T
h
e
p
r
o
p
o
s
ed
cu
s
to
m
C
NN
p
o
s
s
e
s
s
e
s
lo
w
e
s
t
n
u
m
b
er
o
f
p
ar
am
eter
s
w
h
e
n
co
m
p
ar
ed
ag
ain
s
t
th
e
Go
o
g
leNe
t
an
d
A
le
x
Net
p
ar
a
m
eter
s
.
I
n
f
ac
t,
cu
s
to
m
C
NN
h
a
s
3
ti
m
es
s
m
al
ler
n
u
m
b
er
o
f
p
ar
a
m
eter
s
th
a
n
Go
o
g
leNe
t
an
d
3
0
ti
m
es
s
m
a
ller
th
a
n
Alex
Net.
T
h
is
h
as
a
p
o
s
iti
v
e
i
m
p
ac
t
o
n
t
h
e
m
e
m
o
r
y
r
eq
u
ir
e
m
en
t
f
o
r
th
i
s
c
u
s
to
m
C
NN
d
u
r
i
n
g
tr
ain
i
n
g
,
w
h
ich
en
ab
le
th
e
c
u
s
to
m
C
NN
to
b
e
tr
ain
ed
w
ith
m
i
n
i
b
atch
s
ize
o
f
5
0
0
,
co
m
p
ar
ed
to
Go
o
g
leNe
t
an
d
A
le
x
Net
w
h
ic
h
is
s
et
to
o
n
l
y
3
0
th
r
o
u
g
h
o
u
t
th
e
ex
p
er
i
m
e
n
ts
.
T
h
e
in
p
u
t
i
m
ag
e
r
e
s
o
lu
tio
n
f
o
r
cu
s
to
m
C
NN
is
al
s
o
s
ig
n
i
f
i
ca
n
tl
y
s
m
all
er
at
6
4
×6
4
p
ix
els,
as
co
m
p
ar
ed
to
2
2
4
×
2
2
4
p
i
x
els
a
n
d
2
2
7
×2
2
7
p
ix
els
r
eq
u
ir
ed
b
y
Go
o
g
leNe
t
a
n
d
A
l
ex
Net
r
esp
ec
ti
v
el
y
.
I
n
to
tal,
th
e
s
ize
o
f
in
p
u
t
f
o
r
cu
s
to
m
C
NN
is
9
2
%
s
m
aller
th
an
i
n
p
u
t
f
o
r
Go
o
g
leNe
t
an
d
A
le
x
Net.
T
h
is
w
i
ll
allo
w
th
e
cu
s
to
m
C
N
N
to
b
e
tr
ain
ed
m
u
c
h
f
aster
a
n
d
ca
n
b
e
tr
ain
ed
o
n
r
elativ
e
l
y
ch
ea
p
e
r
co
m
p
u
ter
s
y
s
te
m
w
it
h
le
s
s
er
s
p
ec
i
f
icatio
n
s
.
Ne
v
er
th
ele
s
s
,
b
ased
o
n
r
es
u
lts
p
r
esen
ted
i
n
s
u
b
s
eq
u
e
n
ce
,
th
e
c
ap
ab
ilit
y
o
f
cu
s
to
m
C
NN
to
ex
tr
ac
t
an
d
lear
n
h
i
g
h
l
y
co
m
p
le
x
p
ar
a
m
eter
s
f
r
o
m
s
ig
n
i
f
ican
t
l
y
s
m
aller
i
m
ag
e
s
in
m
u
c
h
f
aster
ti
m
e
w
it
h
o
u
t
s
ac
r
i
f
ici
n
g
th
e
ac
c
u
r
a
c
y
is
d
e
m
o
n
s
tr
ated
.
Sin
ce
tr
ai
n
in
g
lar
g
e
d
ataset
s
o
n
C
P
U
o
r
Vis
u
a
l
P
r
o
ce
s
s
in
g
Un
it
(
VP
U)
is
p
ain
f
u
ll
y
s
lo
w
,
it
is
m
o
r
e
ap
p
r
o
p
r
iate
to
s
h
o
w
th
e
co
m
p
u
tat
io
n
al
ad
v
an
ta
g
e
o
f
cu
s
to
m
C
N
N
b
y
m
ea
s
u
r
in
g
t
h
e
av
er
ag
e
i
n
f
er
e
n
ce
ti
m
e
f
o
r
1
2
0
0
im
a
g
es
f
r
o
m
1
0
f
o
ld
s
o
f
L
FW
d
ataset
u
s
in
g
cu
s
to
m
C
NN,
Go
o
g
leNe
t
a
n
d
A
le
x
Net.
T
h
e
in
f
er
en
ce
s
ar
e
r
u
n
o
n
GT
X1
0
8
0
T
i
(
GP
U)
,
I
n
tel
i7
-
6
7
0
0
@
3
.
4
0
GHz
(
C
P
U)
an
d
Mo
v
id
iu
s
Ne
u
r
al
C
o
m
p
u
t
e
Sti
ck
(
Mo
v
id
iu
s
NC
S)
(
VPU)
.
T
h
e
av
er
ag
e
in
f
er
en
ce
ti
m
e
m
ea
s
u
r
ed
in
t
h
i
s
ex
p
er
i
m
e
n
t is
g
i
v
en
i
n
Fi
g
u
r
e
3.
A
cc
o
r
d
in
g
to
Fi
g
u
r
e
3
,
c
u
s
to
m
C
NN
r
eq
u
ir
es
o
n
l
y
0
.
9
2
m
s
to
clas
s
if
y
1
2
0
0
i
m
ag
e
s
o
n
GP
U,
w
h
ile
Go
o
g
leNe
t
a
n
d
A
le
x
Net
r
eq
u
ir
es
s
i
g
n
i
f
ica
n
tl
y
lo
n
g
er
in
f
er
en
ce
ti
m
e
at
4
.
4
2
m
s
a
n
d
1
.
9
5
m
s
r
esp
ec
ti
v
el
y
.
On
C
P
U,
th
e
a
v
er
ag
e
i
n
f
er
en
c
e
ti
m
e
f
o
r
cu
s
to
m
C
N
N
is
j
u
s
t
1
.
7
9
m
s
,
w
h
ile
Go
o
g
leNe
t
a
n
d
A
le
x
Net
r
eq
u
ir
e
s
6
2
.
3
7
m
s
a
n
d
2
0
.
4
1
m
s
r
esp
ec
t
iv
el
y
.
Fo
r
VP
U,
cu
s
to
m
C
NN
ag
ai
n
r
eq
u
ir
e
s
leas
t
a
m
o
u
n
t
o
f
i
n
f
er
en
ce
ti
m
e
o
f
j
u
s
t
2
.
5
1
m
s
,
w
h
ile
Go
o
g
leN
et
an
d
A
lex
Net
r
eq
u
ir
es
9
5
.
7
8
m
s
an
d
5
5
.
6
0
m
s
r
esp
ec
tiv
el
y
.
T
h
is
p
u
t
i
n
to
p
er
s
p
ec
tiv
e
th
at
c
u
s
to
m
C
N
N
ca
p
italize
o
n
its
les
s
er
r
eq
u
ir
e
m
e
n
ts
o
f
i
m
a
g
e
s
ize
a
n
d
p
ar
am
eter
s
w
h
ich
en
ab
les
c
u
s
to
m
C
NN
to
b
e
i
n
f
er
r
ed
at
s
i
g
n
if
ican
t
l
y
h
i
g
h
e
r
s
p
ee
d
o
n
GP
U,
C
P
U
an
d
VP
U
en
v
ir
o
n
m
en
t
.
T
h
e
m
o
s
t
i
m
p
o
r
tan
t
tak
ea
w
a
y
is
t
h
i
s
w
ill
allo
w
c
u
s
to
m
C
N
N
to
b
e
d
ep
lo
y
ed
i
n
r
ea
l
-
ti
m
e
tas
k
s
w
h
et
h
er
it
is
r
u
n
o
n
GP
U/C
P
U
-
b
a
s
ed
s
y
s
te
m
o
r
o
n
co
m
p
letel
y
e
m
b
e
d
d
ed
s
y
s
te
m
s
u
c
h
as
th
o
s
e
r
ely
i
n
g
o
n
m
o
r
e
en
er
g
y
-
e
f
f
icie
n
t
VP
U
s
u
c
h
as
Mo
v
id
iu
s
N
C
S.
I
n
t
h
e
f
o
llo
win
g
s
ec
tio
n
,
it
i
s
s
h
o
w
n
t
h
at
ev
en
t
h
o
u
g
h
cu
s
to
m
C
NN
w
o
r
k
s
o
n
s
m
aller
n
u
m
b
er
o
f
p
ar
am
eter
s
,
it
s
p
er
f
o
r
m
a
n
ce
in
g
en
d
er
clas
s
i
f
icatio
n
i
s
in
f
ac
t
o
n
-
p
ar
an
d
o
cc
asio
n
all
y
b
etter
t
h
an
s
tate
-
of
-
t
h
e
-
ar
ts
.
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:
2
0
8
8
-
8708
Gen
d
er c
la
s
s
ifica
tio
n
u
s
in
g
cu
s
to
m
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k…
(
F
a
d
h
la
n
Ha
fiz
h
elmi
K
a
ma
r
u
Za
ma
n
)
5765
Fig
u
r
e
3
.
C
o
m
p
ar
is
o
n
b
et
w
ee
n
th
e
a
v
er
ag
e
i
n
f
er
en
c
e
ti
m
e
o
f
1
2
0
0
im
a
g
es
f
r
o
m
L
FW
d
ataset
f
o
r
cu
s
to
m
C
NN,
Go
o
g
leNe
t a
n
d
Alex
Ne
t o
n
d
if
f
er
e
n
t i
n
f
er
e
n
ce
e
n
v
ir
o
n
m
e
n
t
s
–
GP
U
(
G
T
X
1
0
8
0
T
i)
,
C
P
U
(
i7
-
6
7
0
0
C
P
U
@
3
.
4
0
G
Hz,
an
d
VP
U
(
Mo
v
id
iu
s
Ne
u
r
al
C
o
m
p
u
te
Sti
c
k
)
3
.
3
.
P
er
f
o
rm
a
nce
i
n g
ender
cla
s
s
if
ica
t
io
n
T
h
e
cu
s
to
m
C
N
N
is
tr
ain
e
d
to
class
if
y
g
en
d
er
f
r
o
m
f
ac
e
i
m
ag
e
s
in
L
FW
A
d
atas
et
an
d
its
p
er
f
o
r
m
a
n
ce
is
e
v
al
u
ated
in
ter
m
s
o
f
ac
c
u
r
ac
y
,
t
r
u
e
p
o
s
itiv
e
r
ate
(
T
P
R
)
,
f
alse
p
o
s
itiv
e
r
ate
(
FP
R
)
an
d
P
r
ec
is
io
n
.
T
h
ese
p
er
f
o
r
m
a
n
ce
s
ar
e
av
er
ag
ed
o
v
er
1
0
r
u
n
s
f
r
o
m
1
0
f
o
ld
s
a
s
cr
o
s
s
v
alid
a
tio
n
as
m
en
tio
n
ed
ea
r
lier
.
T
w
o
v
ar
ia
n
ts
o
f
i
m
a
g
e
s
ar
e
u
s
ed
,
n
a
m
el
y
th
e
g
r
a
y
s
ca
le
an
d
R
GB
i
m
ag
e
s
.
T
h
e
p
er
f
o
r
m
an
ce
f
o
r
g
en
d
er
class
i
f
icatio
n
u
s
i
n
g
c
u
s
to
m
C
NN
is
co
m
p
ar
ed
ag
ai
n
s
t
t
h
e
p
er
f
o
r
m
an
ce
o
f
Go
o
g
leNe
t
a
n
d
A
le
x
Net
cla
s
s
i
f
ier
s
d
o
in
g
th
e
s
a
m
e
tas
k
an
d
is
s
h
o
w
n
i
n
Fi
g
u
r
e
s
4
an
d
5.
Fig
u
r
e
4
.
C
o
m
p
ar
is
o
n
b
et
w
ee
n
th
e
ac
c
u
r
ac
y
o
f
c
u
s
to
m
C
N
N,
Go
o
g
leNe
t a
n
d
A
le
x
Net
w
i
th
r
esp
ec
t
to
class
i
f
icatio
n
u
s
i
n
g
g
r
a
y
s
ca
le
an
d
R
GB
i
m
a
g
es
A
cc
o
r
d
in
g
to
Fi
g
u
r
e
4
,
at
leas
t
s
li
g
h
t
l
y
b
etter
p
er
f
o
r
m
an
ce
i
s
ac
h
ie
v
ed
u
s
i
n
g
R
GB
i
m
a
g
es
co
m
p
ar
ed
w
it
h
g
r
a
y
s
ca
le
i
m
a
g
es
f
o
r
all
test
ed
class
i
f
ier
s
.
0
.
0
1
ac
cu
r
ac
y
i
m
p
r
o
v
e
m
en
t
is
o
b
tain
ed
b
y
cu
s
to
m
C
NN,
Go
o
g
leNe
t
an
d
A
le
x
Net
w
h
en
u
s
in
g
R
GB
i
m
a
g
es
a
s
o
p
p
o
s
ed
to
u
s
in
g
g
r
a
y
s
ca
le
i
m
ag
es.
A
cc
o
r
d
in
g
to
Fig
u
r
es
4
an
d
5
,
Go
o
g
leNe
t
g
i
v
es
th
e
b
est
ac
c
u
r
ac
y
f
o
r
R
G
B
im
a
g
es
w
it
h
an
ac
cu
r
ac
y
o
f
0
.
9
6
,
w
h
ile
cu
s
to
m
C
NN
d
eli
v
er
s
0
.
9
5
ac
cu
r
ac
y
,
b
etter
th
a
n
A
le
x
Net
w
h
ic
h
i
s
at
0
.
9
0
ac
cu
r
ac
y
.
I
n
ter
m
s
o
f
T
P
R
,
Go
o
g
leNe
t
ag
ain
d
eliv
er
s
b
e
s
t
p
er
f
o
r
m
a
n
ce
at
0
.
9
8
T
P
R
,
w
h
ile
c
u
s
to
m
C
NN
a
n
d
A
le
x
Net
d
e
li
v
er
s
0
.
9
7
an
d
0
.
9
3
T
P
R
r
esp
ec
tiv
el
y
.
Si
m
ilar
l
y
,
Go
o
g
leNe
t
d
eliv
er
s
b
est
p
er
f
o
r
m
a
n
ce
f
o
r
FP
R
an
d
P
r
ec
is
io
n
,
w
h
er
e
cu
s
to
m
C
N
N
d
eliv
er
s
s
ec
o
n
d
-
b
est
p
er
f
o
r
m
a
n
ce
,
f
o
llo
w
ed
b
y
A
le
x
Net
w
it
h
th
e
w
o
r
s
t
p
er
f
o
r
m
an
ce
o
f
al
l
th
r
ee
class
i
f
ier
s
.
A
clo
s
er
lo
o
k
o
n
t
h
e
a
cc
u
r
ac
y
o
f
test
ed
cla
s
s
i
f
ier
s
f
o
r
ea
ch
t
est
f
o
ld
i
n
Fi
g
u
r
e
6
s
h
o
w
s
t
h
a
t
cu
s
to
m
C
NN
a
n
d
Go
o
g
leNe
t
h
as
co
m
p
ar
ab
le
p
er
f
o
r
m
an
ce
a
n
d
m
o
r
e
s
tab
le
w
it
h
j
u
s
t
s
li
g
h
t
f
l
u
ct
u
atio
n
s
in
ac
cu
r
ac
y
w
h
e
n
co
m
p
ar
ed
to
A
lex
Net.
T
h
ese
ex
ce
lle
n
t
p
er
f
o
r
m
a
n
ce
s
o
f
c
u
s
to
m
C
NN
ar
e
q
u
ite
i
m
p
r
es
s
iv
e
co
n
s
id
er
in
g
it
s
s
ig
n
i
f
ica
n
tl
y
le
s
s
co
m
p
le
x
ar
c
h
itect
u
r
e
w
h
ic
h
co
n
tain
s
f
e
w
e
r
p
ar
am
e
ter
s
a
n
d
i
n
f
er
r
ed
at
s
ig
n
if
ican
t
l
y
h
i
g
h
er
s
p
ee
d
th
an
s
ta
te
-
of
-
th
e
-
ar
ts
s
u
ch
as Go
o
g
leNe
t a
n
d
A
lex
Net
class
i
f
ier
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
10
,
No
.
6
,
Dec
em
b
er
2
0
2
0
:
5
7
5
8
-
5
7
7
1
5766
(
a)
(
b
)
Fig
u
r
e
5
.
P
er
f
o
r
m
a
n
ce
m
ea
s
u
r
ed
in
ter
m
s
o
f
av
er
ag
e
te
s
t a
cc
u
r
ac
y
,
tr
u
e
p
o
s
iti
v
e
r
ate,
f
al
s
e
p
o
s
itiv
e
r
ate,
an
d
p
r
ec
is
io
n
f
o
r
cu
s
to
m
C
NN,
G
o
o
g
leNe
t a
n
d
A
lex
Net
f
o
r
(
a)
g
r
a
y
s
ca
le
i
m
a
g
e
s
,
an
d
(
b
)
R
GB
im
a
g
es
Fig
u
r
e
6
.
T
h
e
p
er
f
o
r
m
an
ce
m
e
asu
r
ed
as a
cc
u
r
ac
y
f
o
r
cu
s
to
m
C
NN,
Go
o
g
le
Net
an
d
Alex
Net
f
o
r
ea
ch
tes
t f
o
ld
in
L
FW
d
ataset
Af
ter
w
ar
d
s
,
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
g
e
n
d
er
clas
s
i
f
icatio
n
f
o
r
L
FW
A
a
n
d
C
eleb
A
d
ataset
is
e
v
alu
a
ted
i
n
th
e
f
o
r
m
o
f
R
ec
ei
v
er
Op
er
atin
g
C
h
ar
ac
ter
is
tic
s
(
R
O
C
)
cu
r
v
e.
I
t
is
o
n
e
o
f
t
h
e
m
o
s
t
i
m
p
o
r
tan
t
ev
al
u
atio
n
m
etr
ics
f
o
r
ch
ec
k
i
n
g
a
n
y
cla
s
s
i
f
icatio
n
m
o
d
el’
s
p
er
f
o
r
m
a
n
ce
.
A
r
ea
Un
d
er
th
e
C
u
r
v
e
(
A
U
C
)
is
u
s
ed
to
m
ea
s
u
r
e
th
e
p
er
f
o
r
m
a
n
ce
at
v
ar
io
u
s
th
r
es
h
o
ld
s
s
et
tin
g
s
u
s
ed
b
y
t
h
e
f
i
n
al
s
o
f
t
m
ax
la
y
er
.
T
h
e
R
OC
c
u
r
v
e
f
r
o
m
th
is
e
x
p
er
i
m
e
n
t
is
s
h
o
w
n
Fi
g
u
r
e
7
.
B
ased
o
n
Fig
u
r
e
7
,
f
o
r
L
FW
A
d
atasets
,
Go
o
g
leNe
t
h
as
th
e
b
est
A
U
C
,
f
o
llo
w
ed
cu
s
to
m
C
NN
a
n
d
Al
ex
Net.
I
n
ter
e
s
ti
n
g
l
y
,
t
h
e
A
U
C
f
o
r
cu
s
to
m
C
NN
a
n
d
Go
o
g
le
Net
is
n
o
t
to
o
m
u
c
h
d
if
f
er
e
n
t
f
r
o
m
o
n
e
an
o
t
h
er
t
h
u
s
p
r
o
v
in
g
t
h
at
c
u
s
to
m
C
N
N
ag
ai
n
ca
n
d
eliv
er
p
er
f
o
r
m
an
ce
o
n
p
ar
w
it
h
Go
o
g
le
Net
clas
s
i
f
ier
d
esp
ite
i
ts
m
u
c
h
s
i
m
p
ler
ar
c
h
itect
u
r
e.
Mo
r
eo
v
er
,
th
e
A
U
C
o
f
c
u
s
to
m
C
NN
o
n
C
eleb
A
d
ataset
ex
ce
ed
th
e
A
UC
o
f
Go
o
g
leNe
t
an
d
A
le
x
Net,
m
ak
in
g
cu
s
to
m
C
NN
as
th
e
b
est
p
e
r
f
o
r
m
i
n
g
clas
s
i
f
ier
f
o
r
C
eleb
A
d
ata
s
et.
T
h
is
e
x
c
ep
tio
n
al
p
er
f
o
r
m
a
n
ce
o
n
C
e
le
b
A
d
atase
t
also
p
r
o
v
e
s
t
h
at
e
v
en
th
o
u
g
h
c
u
s
to
m
C
NN
co
n
tai
n
s
f
e
w
er
p
ar
a
m
et
er
s
th
at
t
h
e
o
th
er
test
ed
clas
s
i
f
ier
s
,
it
ca
n
lear
n
to
class
i
f
y
t
h
e
g
e
n
d
er
o
f
m
o
r
e
th
an
2
0
0
K
i
m
a
g
es c
o
r
r
ec
tl
y
m
o
s
t o
f
th
e
ti
m
e
.
T
h
en
,
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
cu
s
to
m
C
NN
f
o
r
g
en
d
er
cla
s
s
i
f
icatio
n
is
co
m
p
ar
ed
ag
ai
n
s
t
s
ev
er
al
o
th
er
s
tate
-
of
-
t
h
e
-
ar
ts
.
Si
m
ilar
p
r
o
to
co
ls
as
u
s
ed
b
y
t
h
e
o
r
ig
in
al
p
u
b
licatio
n
s
o
f
th
e
s
e
m
et
h
o
d
s
ar
e
u
s
ed
,
an
d
th
e
r
es
u
lts
ar
e
g
i
v
en
in
T
ab
le
5
.
Fo
r
L
FW
A
d
ata
s
et,
L
Ne
ts
+
A
Net
[
2
1
]
d
eliv
er
s
t
h
e
b
est
r
esu
lt
w
it
h
9
8
%
ac
cu
r
ac
y
w
h
ile
c
u
s
to
m
C
N
N
d
eliv
er
s
9
5
%
ac
cu
r
ac
y
.
Ho
w
e
v
er
,
cu
s
to
m
C
NN
d
eliv
er
s
b
etter
ac
cu
r
ac
y
co
m
p
ar
ed
to
s
ev
er
al
m
et
h
o
d
s
s
u
c
h
as
Alex
Ne
t,
Face
T
r
ac
e
r
[
2
0
]
,
P
A
NDA
-
w
[
2
]
,
a
n
d
R
C
NN
Gen
d
er
[
2
2
]
.
I
n
f
ac
t,
cu
s
to
m
C
NN
d
eli
v
er
s
th
e
s
a
m
e
ac
c
u
r
ac
y
a
s
[
3
7
]
+A
Ne
t.
Fo
r
C
e
leb
A
d
ataset,
c
u
s
to
m
C
NN
d
eliv
er
s
th
e
b
est
p
er
f
o
r
m
an
ce
,
s
u
r
p
as
s
i
n
g
a
ll
s
tate
-
of
-
th
e
-
ar
ts
w
it
h
9
6
%
g
e
n
d
er
class
i
f
icatio
n
ac
c
u
r
ac
y
.
C
u
s
to
m
C
NN
is
ev
e
n
3
%
b
etter
th
a
n
Go
o
g
l
eNe
t
an
d
2
%
b
etter
th
a
n
t
h
e
r
ec
en
t
H
y
p
er
Face
[
2
2
]
.
T
h
e
f
o
llo
w
i
n
g
Fi
g
u
r
e
8
s
h
o
w
t
h
e
ex
a
m
p
l
e
o
f
f
e
m
a
le
an
d
m
ale
s
a
m
p
les
w
i
th
i
n
co
r
r
ec
t c
lass
i
f
icatio
n
f
o
r
L
FW
an
d
C
eleb
A
d
ataset.
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:
2
0
8
8
-
8708
Gen
d
er c
la
s
s
ifica
tio
n
u
s
in
g
cu
s
to
m
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k…
(
F
a
d
h
la
n
Ha
fiz
h
elmi
K
a
ma
r
u
Za
ma
n
)
5767
(
a)
(
b
)
Fig
u
r
e
7
.
R
OC
f
o
r
(
a)
L
FW
an
d
(
b
)
C
eleb
A
d
atasets
T
ab
le
5
.
A
cc
u
r
ac
y
o
f
g
e
n
d
er
class
i
f
icatio
n
f
o
r
co
m
p
ar
ed
ag
a
in
s
t
s
ev
er
al
s
tate
-
of
-
t
h
e
-
ar
t
s
M
e
t
h
o
d
C
l
a
ssi
f
i
c
a
t
i
o
n
A
c
c
u
r
a
c
y
(
%)
L
F
W
A
C
e
l
e
b
A
F
a
c
e
T
r
a
c
e
r
[
2
0
]
91
84
P
A
N
D
A
-
w
[
2
]
93
86
P
A
N
D
A
-
1
[
2
]
97
92
[
3
7
]
+
A
N
e
t
95
91
L
N
e
t
s+
A
N
e
t
[
2
1
]
98
94
C
a
st
r
i
l
l
o
n
-
S
a
n
t
a
n
a
e
t
a
l
[
1
]
98
NA
R
C
N
N
G
e
n
d
e
r
[
2
2
]
95
91
M
u
l
t
i
t
a
sk
F
a
c
e
[
2
2
]
97
93
H
y
p
e
r
F
a
c
e
[
2
2
]
97
94
C
-
P
e
g
a
so
s
[
7
]
97
NA
A
F
I
F
[
8
]
96
NA
A
l
e
x
N
e
t
90
91
G
o
o
g
l
e
N
e
t
96
93
c
u
st
o
m
C
N
N
95
96
(
a)
(
b
)
Fig
u
r
e
8
.
E
x
a
m
p
les o
f
f
e
m
ale
(
to
p
)
an
d
m
ale
(
b
o
tto
m
)
s
a
m
p
l
es
w
i
th
i
n
co
r
r
ec
t c
lass
i
f
icat
io
n
f
o
r
,
(
a)
L
FW
d
ataset
an
d
(
b
)
C
eleb
A
d
ataset
A
cc
o
r
d
in
g
to
Fig
u
r
e
8
,
s
o
m
e
o
f
s
u
p
p
o
s
ed
l
y
f
e
m
ale
s
u
b
j
ec
t
m
a
y
h
av
e
m
asc
u
li
n
e
f
ea
t
u
r
es
s
u
c
h
as
m
o
r
e
a
n
g
u
lar
f
ac
e
s
h
ap
e
w
it
h
s
tr
o
n
g
j
a
w
li
n
e.
L
ik
e
w
i
s
e,
t
h
e
s
u
p
p
o
s
ed
ly
m
ale
s
u
b
j
ec
ts
m
a
y
h
a
v
e
f
e
m
i
n
i
n
e
f
ea
t
u
r
es
s
u
ch
a
s
r
o
u
n
d
er
f
ac
e
an
d
ch
ee
k
s
.
Ot
h
er
f
ac
to
r
s
th
at
m
a
y
co
n
tr
ib
u
te
to
m
is
cl
ass
i
f
icatio
n
is
p
o
s
e
v
ar
iatio
n
s
an
d
o
b
j
ec
ts
o
r
o
cc
l
u
s
io
n
s
p
r
esen
t
in
f
ac
ial
i
m
a
g
es,
e.
g
.
:
s
u
n
g
la
s
s
es
a
n
d
m
icr
o
p
h
o
n
es.
Ho
w
ev
er
,
th
ese
m
is
cla
s
s
i
f
icatio
n
s
ca
n
b
e
r
ed
u
ce
d
b
y
h
a
v
in
g
m
o
r
e
s
a
m
p
les
o
f
m
ale
an
d
f
e
m
ale
p
o
s
s
ess
in
g
t
h
e
p
o
s
e
v
ar
iatio
n
s
an
d
o
b
j
ec
ts
o
r
o
cc
lu
s
io
n
s
i
n
tr
ai
n
in
g
d
ataset.
Evaluation Warning : The document was created with Spire.PDF for Python.