I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
pu
t
er
Science
Vo
l.
23
,
No
.
2
,
A
u
g
u
s
t
2
0
2
1
,
p
p
.
9
7
3
~
9
7
9
I
SS
N:
2502
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
23
.i
2
.
pp
973
-
9
7
9
973
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Aug
mented
binar
y
multi
-
la
bele
d C
NN
for
practical
f
a
cia
l
a
tt
ribut
e clas
sific
a
tion
M
o
ha
m
m
ed
B
er
ra
ha
l,
M
o
s
t
a
f
a
Aziz
i
M
ATS
I
Re
se
a
rc
h
Lab
,
E
S
TO,
M
o
h
a
m
m
e
d
F
irst
Un
iv
e
rsit
y
,
Ou
jd
a
,
M
o
r
o
c
c
o
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Mar
2
1
,
2
0
21
R
ev
is
ed
May
23
,
2
0
21
Acc
ep
ted
J
un
1
,
2
0
21
Bo
th
h
u
m
a
n
fa
c
e
re
c
o
g
n
i
ti
o
n
a
n
d
g
e
n
e
ra
ti
o
n
b
y
m
a
c
h
in
e
s
a
re
c
u
rre
n
tl
y
a
n
a
c
ti
v
e
a
re
a
o
f
c
o
m
p
u
ter
v
isio
n
,
d
ra
win
g
c
u
rio
si
ty
o
f
re
se
a
rc
h
e
rs,
c
a
p
a
b
le
o
f
p
e
rfo
rm
in
g
a
m
a
z
in
g
ima
g
e
a
n
a
l
y
sis,
a
n
d
p
ro
d
u
c
i
n
g
a
p
p
l
ica
ti
o
n
s
i
n
m
u
lt
i
p
le
d
o
m
a
in
s.
In
t
h
is
p
a
p
e
r,
we
p
ro
p
o
se
a
n
e
w
a
p
p
ro
a
c
h
fo
r
fa
c
e
a
tt
rib
u
tes
c
las
sifica
ti
o
n
(F
AC)
tak
in
g
a
d
v
a
n
tag
e
f
ro
m
b
o
t
h
b
i
n
a
ry
c
las
sifica
ti
o
n
a
n
d
d
a
ta
a
u
g
m
e
n
tatio
n
.
W
it
h
b
in
a
r
y
c
las
sifica
ti
o
n
we
c
a
n
re
a
c
h
h
ig
h
p
re
d
ictio
n
sc
o
re
s,
wh
il
e
a
u
g
m
e
n
ted
d
a
ta
p
r
e
v
e
n
t
o
v
e
rfit
ti
n
g
a
n
d
o
v
e
rc
o
m
e
th
e
lac
k
o
f
d
a
ta
fo
r
sk
e
tch
e
d
p
h
o
t
o
s.
O
u
r
a
p
p
ro
a
c
h
,
n
a
m
e
d
A
u
g
m
e
n
ted
b
in
a
ry
m
u
lt
il
a
b
e
l
CNN
(ABM
-
CNN
),
c
o
n
sists
o
f
t
h
re
e
ste
p
s:
i
)
sp
li
tt
in
g
d
a
ta
;
ii
)
tran
sfo
rm
e
d
-
it
to
sk
e
tch
(sim
p
li
f
ica
ti
o
n
p
ro
c
e
ss
)
;
iii
)
trai
n
se
p
a
ra
tely
e
a
c
h
a
tt
ri
b
u
t
e
with
two
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s;
t
h
e
wh
o
le
p
ro
c
e
ss
in
c
lu
d
e
s
two
n
e
t
wo
rk
s:
t
h
e
first
(re
sp
.
th
e
se
c
o
n
d
)
o
n
e
is
t
o
p
re
d
ict
a
tt
ri
b
u
tes
o
n
re
a
l
ima
g
e
s
(re
sp
.
sk
e
tch
e
s)
a
s
in
p
u
ts
.
T
h
ro
u
g
h
e
x
p
e
rime
n
ta
ti
o
n
,
we
fi
g
u
re
o
u
t
th
a
t
so
m
e
a
tt
rib
u
tes
g
iv
e
h
ig
h
p
re
d
icti
o
n
ra
tes
with
sk
e
tch
e
s
ra
th
e
r
t
h
a
n
with
re
a
l
ima
g
e
s.
On
th
e
o
t
h
e
r
h
a
n
d
,
we
b
u
il
d
a
n
e
w
fa
c
e
d
a
tas
e
t,
m
o
re
c
o
n
siste
n
t
a
n
d
c
o
m
p
lete
,
b
y
g
e
n
e
ra
ti
n
g
ima
g
e
s
u
sin
g
S
t
y
le
-
G
AN
m
o
d
e
l,
to
wh
ich
we
a
p
p
l
y
o
u
r
m
e
th
o
d
fo
r
e
x
trac
ti
n
g
fa
c
e
a
tt
rib
u
tes
.
As
re
su
lt
s,
o
u
r
p
ro
p
o
sa
l
d
e
m
o
n
stra
tes
m
o
re
p
e
rfo
rm
a
n
c
e
s
c
o
m
p
a
re
d
to
th
o
se
o
f
re
late
d
w
o
r
k
s.
K
ey
w
o
r
d
s
:
C
NN
Data
a
u
g
m
en
tatio
n
Dee
p
l
ea
r
n
in
g
Face
a
ttrib
u
tes
Face
s
k
etch
i
m
ag
e
I
m
ag
e
c
lass
if
icatio
n
M
u
lti
-
lab
el
lear
n
in
g
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
:
Mo
h
am
m
ed
B
er
r
a
h
al
Su
p
er
io
r
Sch
o
o
l o
f
T
ec
h
n
o
lo
g
y
Ou
jd
a
(
E
STO)
,
MA
T
SI
R
esear
ch
L
ab
Mo
h
am
m
ed
First Un
iv
er
s
ity
,
Ou
jd
a,
Mo
r
o
cc
o
E
m
ail:
m
.
b
er
r
ah
al@
u
m
p
.
ac
.
m
a
1.
I
NT
RO
D
UCT
I
O
N
I
n
ju
s
t
th
e
p
ast
f
ew
y
ea
r
s
,
ar
tific
ial
in
tellig
en
ce
h
as
tak
en
th
e
wo
r
ld
b
y
s
u
r
p
r
is
e,
d
iv
in
g
r
ap
id
p
r
o
g
r
ess
esp
ec
ially
in
t
h
e
f
iel
d
o
f
d
ee
p
lear
n
in
g
,
FAC
an
d
p
atter
n
r
ec
o
g
n
itio
n
p
r
o
b
lem
h
a
v
e
s
ee
n
tr
em
en
d
o
u
s
p
r
o
g
r
ess
.
T
h
is
co
n
tr
ib
u
tes
an
d
p
lay
s
an
im
p
o
r
tan
t
r
o
le
in
s
ec
u
r
ity
[
1
]
,
[
2
]
lik
e
ac
ce
s
s
co
n
tr
o
l
f
o
r
PC
s
o
r
s
m
ar
tp
h
o
n
e,
v
id
e
o
s
u
r
v
eillan
ce
,
cr
im
i
n
al
au
th
e
n
ticatio
n
,
f
ac
e
s
k
etch
[
3
]
a
n
d
f
ac
e
p
h
o
to
f
o
r
th
e
law
en
f
o
r
ce
m
e
n
t
ap
p
licatio
n
[
4
]
.
T
h
e
m
ain
p
r
o
p
er
ty
o
f
FAC
[
5
]
is
to
p
r
e
d
ict
m
u
ltip
le
f
ac
e
f
ea
tu
r
es,
s
tate
an
d
em
o
tio
n
[
6
]
o
n
th
e
g
iv
e
n
im
ag
e
o
r
f
ac
e
p
o
r
tr
ait.
Var
i
o
u
s
alg
o
r
ith
m
s
h
av
e
r
ea
ch
ed
a
n
ex
ce
llen
t
r
esu
lt
o
n
m
u
ltip
le
lev
els
f
o
r
FAC
,
eith
er
ap
p
ly
d
ir
ec
tly
C
NN
[
7
]
m
o
d
els
to
ex
tr
ac
t
f
ac
e
f
ea
tu
r
es,
o
r
u
s
in
g
m
eth
o
d
s
f
o
r
im
p
r
o
v
in
g
th
e
lear
n
i
n
g
b
y
d
is
tr
ib
u
tin
g
t
h
e
attr
ib
u
tes
in
to
t
wo
ca
teg
o
r
ies:
o
b
jecti
v
e
attr
i
b
u
tes
lik
e
we
ar
in
g
a
h
at,
ey
e
g
lass
es,
b
an
g
s
an
d
s
u
b
jectiv
e
o
n
es
lik
e
s
m
ilin
g
,
b
ig
lip
s
[
8
]
.
S
o
m
e
m
eth
o
d
s
f
o
cu
s
o
n
g
r
o
u
p
in
g
s
o
m
e
attr
ib
u
tes
o
n
th
e
b
asis
o
f
th
eir
in
ter
co
r
r
elatio
n
s
[
9
]
,
wh
ile
o
th
er
s
tar
g
et
d
etec
t
in
g
f
ac
e
la
n
d
m
ar
k
lo
ca
lizatio
n
[
1
0
]
to
r
e
d
u
ce
th
e
n
o
is
e.
Desp
ite
th
eir
wid
e
ap
p
licatio
n
,
FAC
r
em
ain
s
a
ch
allen
g
e
f
o
r
r
esear
ch
er
s
.
T
h
e
r
e
is
s
till
d
i
f
f
icu
lty
to
r
ec
o
g
n
ize
d
if
f
e
r
en
t
f
ac
e
attr
ib
u
tes
o
n
a
g
i
v
in
g
im
ag
e
an
d
it
m
ay
r
eq
u
ir
e
m
o
r
e
atten
tio
n
s
to
c
o
v
er
d
if
f
e
r
en
t
r
eg
io
n
s
o
f
th
e
f
ac
e,
an
d
a
l
ac
k
o
f
lar
g
e
d
atasets
with
h
eter
o
g
en
eo
u
s
f
ac
e
im
ag
es
an
d
s
k
etc
h
es,
o
r
an
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
23
,
No
.
2
,
Au
g
u
s
t 2
0
2
1
:
973
-
9
7
9
974
ad
d
itio
n
al
in
f
o
r
m
atio
n
t
h
at
ca
n
h
elp
to
i
n
cr
ea
s
e
o
u
r
ac
cu
r
a
cy
.
I
n
th
is
p
a
p
er
,
we
p
r
esen
t
a
n
ew
au
g
m
e
n
ted
b
in
ar
y
m
u
lti
-
lab
el
C
NN
-
b
ased
m
eth
o
d
(
AB
M
-
C
NN)
to
d
e
al
with
d
if
f
er
en
t
asp
ec
ts
o
f
FAC
f
o
r
r
ea
l
f
ac
ial
im
ag
es
an
d
s
k
etch
es.
I
n
itially
,
we
tr
an
s
f
o
r
m
o
u
r
m
u
lti
-
lab
e
l
p
r
o
b
lem
t
o
a
b
in
a
r
y
p
r
o
b
le
m
.
T
o
th
is
en
d
,
we
d
ev
elo
p
an
alg
o
r
ith
m
to
tr
an
s
f
o
r
m
o
u
r
p
r
o
b
lem
t
o
m
i
n
i
-
p
r
o
b
lem
s
o
f
b
in
ar
y
class
if
icatio
n
b
y
s
p
litt
in
g
th
e
d
ataset
o
f
ev
er
y
attr
ib
u
te
to
t
wo
class
es
(
0
o
r
1
)
if
th
e
attr
i
b
u
te
ex
is
ts
in
th
e
im
ag
e,
th
en
af
f
ec
ted
1
,
a
n
d
0
if
n
o
.
Af
ter
th
at
we
p
er
f
o
r
m
d
at
a
au
g
m
en
tatio
n
t
o
m
u
ltip
ly
d
ata
f
o
r
s
o
lid
lear
n
in
g
,
we
co
n
v
er
t
ev
er
y
im
a
g
e
in
th
e
d
ataset
in
to
9
im
a
g
es
b
y
ch
an
g
in
g
its
p
er
s
p
ec
tiv
es.
Du
e
to
th
e
lack
o
f
s
k
etch
es,
we
d
ec
id
e
to
tr
a
n
s
f
o
r
m
o
u
r
d
ata
t
o
s
k
etch
es.
W
e
r
u
n
o
u
r
C
NN
m
o
d
el
f
o
r
b
o
t
h
r
ea
l
im
ag
es
an
d
s
k
etch
es,
th
en
we
co
m
b
i
n
e
th
e
p
r
ed
ictio
n
r
esu
lts
f
o
r
ea
ch
attr
ib
u
te.
Ou
r
p
r
ed
ictio
n
m
o
d
u
le
i
n
AB
M
-
C
NN
o
u
tp
u
ts
4
0
f
ac
ial
attr
ib
u
tes
s
u
ch
as
h
air
co
lo
r
,
g
en
d
e
r
id
en
tific
ati
o
n
,
s
m
ilin
g
,
attr
ac
tio
n
,
an
d
h
a
t
o
r
g
lass
wea
r
in
g
.
On
th
e
o
t
h
er
h
a
n
d
,
we
cr
ea
te
o
u
r
o
wn
d
ataset
b
ased
o
n
g
en
er
ated
im
a
g
es
u
s
in
g
th
e
s
t
y
le
-
GAN
m
o
d
el
[
1
1
]
,
a
n
d
we
u
s
e
it
to
cr
ea
te
g
en
er
ated
f
ac
es.
Fin
ally
,
we
a
p
p
ly
o
u
r
AB
M
-
C
NN
m
o
d
el
t
o
ex
tr
ac
t
attr
ib
u
tes
an
d
s
av
e
it
in
a
C
SV
f
ile,
to
s
er
v
e
o
u
r
n
ex
t
r
esear
ch
in
c
o
m
p
u
ter
v
is
io
n
.
T
h
e
r
est
o
f
th
i
s
ar
ticle
is
f
o
r
m
u
lated
as
f
o
llo
ws:
I
n
th
e
s
ec
o
n
d
s
ec
tio
n
,
we
r
ec
all
b
ac
k
g
r
o
u
n
d
s
ab
o
u
t
c
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
(
C
NN)
,
d
ata
a
u
g
m
en
tatio
n
,
b
in
ar
y
class
if
icatio
n
an
d
m
u
lti
-
lab
el
lear
n
in
g
.
I
n
th
e
th
ir
d
s
ec
tio
n
,
we
d
is
cu
s
s
th
e
r
elat
ed
wo
r
k
s
.
W
e
p
r
esen
t
o
u
r
im
p
lem
en
tatio
n
in
t
h
e
f
o
u
r
th
s
ec
tio
n
.
B
ef
o
r
e
co
n
cl
u
d
in
g
,
th
e
f
if
th
s
ec
tio
n
s
u
m
m
a
r
izes
o
u
r
r
esu
lts
,
an
d
co
m
p
a
r
es th
em
with
p
r
ev
io
u
s
wo
r
k
s
.
2.
B
ACK
G
RO
UND
2
.
1
.
Co
nv
o
lutio
na
l neura
l net
wo
rk
(
CNN)
On
e
o
f
th
e
b
est
lear
n
in
g
alg
o
r
ith
m
s
th
e
u
n
d
er
s
tan
d
in
g
im
ag
e
co
n
ten
t,
th
e
C
NN
is
a
f
e
ed
f
o
r
war
d
m
u
ltil
ay
er
ed
h
ier
ar
ch
ical
n
et
wo
r
k
,
t
h
e
lay
e
r
s
ar
e
u
s
in
g
m
u
ltip
le
co
n
v
o
lu
tio
n
al
k
er
n
els
t
o
tr
an
s
f
o
r
m
th
e
g
iv
e
n
co
r
r
elate
d
d
ata,
f
o
r
e
x
tr
ac
tin
g
u
s
ef
u
l
f
ea
tu
r
es
f
r
o
m
th
e
m
,
i
n
th
e
o
th
e
r
h
an
d
th
e
o
u
tp
u
t
o
f
th
e
co
n
v
o
lu
tio
n
al
k
er
n
els
is
th
en
ass
ig
n
ed
to
ac
tiv
atio
n
f
u
n
ctio
n
h
o
is
th
e
n
o
n
lin
ea
r
p
r
o
ce
s
s
in
g
u
n
it,
to
h
elp
in
lear
n
in
g
ab
s
tr
ac
tio
n
s
an
d
em
b
ed
s
n
o
n
-
lin
ea
r
ity
in
th
e
f
ea
tu
r
e
s
p
ac
e.
T
h
is
way
it
g
en
e
r
ates
d
if
f
er
e
n
t
p
a
tter
n
s
h
elp
s
in
lear
n
in
g
o
f
s
em
an
tic
d
if
f
er
e
n
c
es
in
im
ag
es
[
7
]
.
Un
til
n
o
w
th
is
alg
o
r
ith
m
h
as
s
h
o
wn
an
am
az
in
g
p
er
f
o
r
m
an
ce
in
im
ag
es c
lass
if
icatio
n
,
d
etec
tio
n
,
s
eg
m
en
tatio
n
,
an
d
ex
tr
ac
t
in
g
f
ea
tu
r
es.
2
.
2
.
Da
t
a
a
ug
m
ent
a
t
io
n
Data
Au
g
m
en
tatio
n
ar
e
m
eth
o
d
s
th
at
u
s
ed
to
ex
p
an
d
t
h
e
s
ize
o
f
d
ataset
b
y
cr
ea
te
o
n
e
o
r
m
u
ltip
le
n
ew
d
ata
s
lig
h
tly
m
o
d
if
ied
o
f
ea
ch
ex
is
tin
g
d
ata,
th
e
p
u
r
p
o
s
e
o
f
th
ese
tech
n
iq
u
es
is
to
ass
is
t
m
ac
h
in
e
lear
n
in
g
alg
o
r
ith
m
s
to
r
e
d
u
cin
g
th
e
o
v
e
r
f
itti
n
g
.
I
n
o
u
r
ca
s
e
we
u
s
e
9
co
p
ies f
o
r
ea
c
h
f
ac
e
im
a
g
es
[
1
2
]
.
2
.
3
.
B
ina
ry
cla
s
s
if
ica
t
io
n
I
n
d
ee
p
lear
n
in
g
,
b
in
ar
y
class
if
icatio
n
is
th
e
c
r
ea
te
a
m
o
d
el
with
two
o
u
t
p
u
ts
ab
le
t
o
p
r
ed
i
ct
tr
u
e
o
r
f
alse
f
r
o
m
in
p
u
t
d
ata
th
at
ar
e
d
iv
id
ed
in
to
two
g
r
o
u
p
s
,
in
g
en
er
ally
it
ca
n
s
o
lv
e
ea
c
h
p
r
o
b
lem
with
av
er
a
g
e
d
o
u
b
le
class
if
ica
tio
n
is
s
u
es
,
s
in
ce
th
e
o
u
tp
u
t
is
s
im
p
le,
th
e
ac
cu
r
ac
y
o
f
th
ese
alg
o
r
ith
m
s
ar
e
h
ig
h
er
th
an
th
e
o
th
er
class
if
ier
s
an
d
it’s e
asy
f
o
r
th
e
m
o
d
el
to
p
r
ed
ict
r
esu
lts
[
1
3
]
.
2
.
4
.
M
ulti
-
la
bel
l
ea
rning
On
e
o
f
th
e
m
ajo
r
p
r
o
b
lem
s
o
f
class
if
icatio
n
th
at
Dee
p
lear
n
in
g
tr
y
t
o
m
o
d
el
is
to
ass
ig
n
m
o
r
e
th
a
n
two
class
es
to
m
u
ltip
le
o
u
tp
u
t
,
th
is
tech
n
ic
is
ca
lled
th
e
m
u
lt
i
-
lab
el
class
if
icatio
n
.
Mu
lti
-
lab
el
class
if
icatio
n
is
h
o
w
to
f
in
d
a
m
o
d
el
th
at
ca
n
m
ap
in
p
u
ts
to
b
i
n
ar
y
v
ec
to
r
s
,
th
is
tech
n
ic
is
a
g
en
er
aliz
atio
n
o
f
b
in
a
r
y
a
n
d
m
u
lticlas
s
clas
s
if
icatio
n
we
as
s
ig
n
a
v
alu
e
0
o
r
1
f
o
r
ea
ch
ele
m
en
t in
o
u
tp
u
t
[
1
4
]
.
3.
R
E
L
AT
E
D
WO
RK
Ma
o
et
a
l
.,
p
r
o
p
o
s
e
a
n
ew
al
g
o
r
ith
m
to
d
ea
l
with
th
e
f
ac
e
at
tr
ib
u
te
ex
tr
ac
tio
n
f
r
o
m
f
ac
ial
i
m
ag
e
s
,
th
e
alg
o
r
ith
m
ca
lled
d
ee
p
m
u
lti
-
task
m
u
lti
-
lab
el
C
NN
(
DM
M
-
C
NN)
,
b
y
d
iv
id
i
n
g
th
e
f
ac
ia
l
attr
ib
u
tes
in
two
ca
teg
o
r
ies
o
b
jectiv
e
an
d
s
u
b
jectiv
e
th
ey
m
an
ag
e
t
o
r
u
n
two
d
if
f
er
en
t
n
etwo
r
k
ar
c
h
itectu
r
e
s
tak
in
g
ad
v
an
tag
e
o
f
m
u
ltit
ask
lear
n
in
g
,
an
d
ad
o
p
tin
g
d
y
n
a
m
ic
weig
h
tin
g
s
ch
em
e
to
r
eso
lv
e
th
e
p
r
o
b
lem
o
f
d
iv
er
s
e
lear
n
in
g
co
m
p
lex
ities
[
8
]
.
I
n
th
e
s
am
e
f
ield
E
h
r
lich
a
n
d
Sh
ield
s
p
r
o
p
o
s
e
th
er
e
m
u
ltit
ask
lear
n
i
n
g
f
ac
ial
attr
ib
u
tes
ap
p
r
o
ac
h
[
1
5
]
,
wh
ile
Ma
n
d
el,
Pas
ca
n
u
p
r
o
p
o
s
e
a
m
eth
o
d
b
ased
o
n
s
h
ar
ed
f
ea
tu
r
es
b
etwe
e
n
th
ese
attr
ib
u
tes,
b
y
u
s
in
g
m
u
lti
-
task
r
estricte
d
b
o
ltzm
an
n
m
ac
h
in
e
(
MT
-
R
B
M)
[
1
6
]
,
th
ey
wer
e
a
b
le
to
lear
n
in
g
a
jo
i
n
t
f
ea
tu
r
e
r
ep
r
esen
tatio
n
f
r
o
m
f
ac
ial
la
n
d
m
ar
k
p
o
in
ts
f
o
r
all
attr
ib
u
t
es,
f
o
llo
win
g
b
y
ap
p
r
o
ac
h
s
u
b
s
is
t
o
f
a
b
o
tto
m
-
u
p
/to
p
-
d
o
wn
p
ass
f
o
r
lea
r
n
in
g
th
e
s
h
a
r
ed
r
e
p
r
esen
tatio
n
o
f
m
u
ltit
ask
m
o
d
els,
an
d
b
o
tto
m
-
u
p
p
ass
f
o
r
p
r
ed
ictio
n
o
f
task
s
.
T
h
is
a
p
p
r
o
ac
h
h
as r
ea
ch
ed
s
o
m
e
g
o
o
d
r
e
s
u
lt
b
y
a
v
er
ag
e
o
f
8
7
%
ac
cu
r
a
cy
f
o
r
all
attr
ib
u
tes
o
n
C
eleb
A
d
ataset
[
9]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
A
u
g
men
ted
b
in
a
r
y
mu
lti
-
la
b
el
ed
C
N
N
fo
r
p
r
a
ctica
l fa
cia
l a
ttr
ib
u
te
cla
s
s
ifica
tio
n
(
Mo
h
a
mme
d
B
err
a
h
a
l
)
975
Z
h
u
an
g
et
a
l
.
p
r
esen
t
a
n
o
v
e
l
f
o
r
m
u
lti
-
lab
el
lear
n
i
n
g
f
ac
i
al
attr
ib
u
tes
u
s
in
g
d
ee
p
tr
a
n
s
f
er
n
e
u
r
al
n
etwo
r
k
m
eth
o
d
n
am
ed
f
ac
e
m
u
lti
-
lab
el
tr
an
s
f
er
n
etwo
r
k
(
FMT
Net)
,
as
it
n
am
es
th
is
m
eth
o
d
r
esid
e
in
3
m
ajo
r
n
etwo
r
k
s
,
f
ac
e
d
etec
tio
n
,
m
u
lti
-
lab
el
lear
n
in
g
an
d
tr
an
s
f
er
lear
n
i
n
g
n
etwo
r
k
,
t
h
e
s
ec
o
n
d
n
etwo
r
k
co
n
s
is
t
o
f
p
r
ed
ictin
g
m
u
ltip
le
f
ac
ial
attr
ib
u
tes
s
im
u
ltan
eo
u
s
ly
,
to
in
cr
ea
s
e
p
er
f
o
r
m
a
n
ce
,
a
n
d
r
ed
u
ce
s
f
ea
tu
r
e
r
ed
u
n
d
an
cy
with
th
e
p
r
o
p
o
s
e
d
lo
s
s
weig
h
t
s
ch
em
e,
th
e
th
i
r
d
n
etwo
r
k
is
th
e
u
n
s
u
p
er
v
is
ed
lear
n
in
g
,
f
o
r
th
e
ad
ap
tatio
n
o
f
u
n
lab
eled
f
ac
ial
attr
ib
u
te
class
if
icatio
n
,
th
eir
m
eth
o
d
r
ea
ch
es
an
av
er
a
g
e
8
4
.
3
4
%
ac
c
u
r
ac
y
o
n
L
FW
A
d
ataset
[
9
]
.
H.
Din
g
et
a
l.
p
r
esen
t
n
o
v
el
to
im
p
r
o
v
e
attr
i
b
u
te
class
if
icatio
n
,
tr
u
e
ca
s
ca
d
e
n
etwo
r
k
lear
n
s
th
e
lo
ca
tin
g
o
f
th
e
f
ac
e
r
eg
i
o
n
an
d
p
er
f
o
r
m
s
attr
i
b
u
te
class
if
icatio
n
with
o
u
t a
lig
n
m
en
t
,
i
)
T
h
e
n
etwo
r
k
is
d
esig
n
ed
to
au
to
m
atica
lly
d
etec
t
s
p
ec
if
ic
r
eg
io
n
s
o
f
attr
ib
u
tes
;
ii
)
a
co
n
s
tr
u
ctio
n
an
d
co
m
b
i
n
atio
n
o
f
a
h
o
le
im
ag
e
-
b
ased
n
etwo
r
k
an
d
a
m
u
lt
ip
le
p
ar
t
-
b
ased
n
etwo
r
k
b
y
th
e
r
e
g
io
n
s
witch
lay
er
f
o
r
f
i
n
a
l
attr
ib
u
te
class
if
icatio
n
[
1
7
]
.
Kh
an
et
a
l
.
p
r
o
p
o
s
e
a
f
r
am
ewo
r
k
f
o
cu
s
i
n
g
o
n
th
r
ee
h
u
m
an
asp
ec
t
s
th
e
g
en
d
e
r
,
r
ac
e
an
d
ag
e,
th
e
f
r
am
ew
o
r
k
u
s
e
two
C
NN,
th
e
f
ir
s
t
as
s
eg
m
en
tatio
n
m
o
d
el
u
s
e
to
p
ar
s
es
a
f
ac
e
in
to
s
ev
en
d
an
ce
class
es
th
an
cr
ea
te
a
p
r
o
b
ab
i
lity
m
ap
s
f
o
r
ea
ch
f
ac
e
class
,
th
e
s
ec
o
n
d
m
o
d
el
is
to
ex
t
r
ac
t
f
ea
tu
r
es
f
r
o
m
p
r
o
b
a
b
ilit
y
m
ap
s
o
f
t
h
e
co
r
r
esp
o
n
d
in
g
class
f
o
r
ea
c
h
th
r
ee
a
s
p
ec
t m
en
tio
n
ed
b
ef
o
r
e
[
1
8
]
.
4.
I
M
P
L
E
M
E
NT
A
T
I
O
N
O
F
T
H
E
M
E
T
H
O
D
Ou
r
ap
p
r
o
ac
h
u
s
es
b
o
th
h
ig
h
p
r
ed
ictio
n
r
ate
f
o
r
b
in
ar
y
class
if
icatio
n
m
o
d
els
an
d
s
im
p
lifie
d
im
ag
es
in
to
s
k
etch
,
f
o
r
ex
tr
ac
t
in
g
s
o
m
e
f
ea
tu
r
es
with
m
o
r
e
v
is
ib
le
attr
ib
u
tes.
C
o
n
s
id
er
in
g
o
u
r
g
o
al,
we
d
ev
elo
p
an
alg
o
r
ith
m
th
at
co
n
s
is
ts
f
ir
s
t
t
o
d
etec
t
th
e
f
ac
e
o
n
tr
ain
in
g
d
ata,
th
en
it
tr
an
s
f
o
r
m
s
o
u
r
in
p
u
t
d
ata
to
s
k
etch
.
Nex
t,
it
s
p
lit
s
,
f
o
r
ea
ch
attr
ib
u
te,
th
e
p
r
ep
r
o
ce
s
s
ed
d
ata
i
n
to
1
(
f
o
ld
er
1
)
if
th
e
attr
ib
u
te
ex
i
s
ts
o
r
0
(
f
o
ld
er
0
)
if
n
o
t.
Af
ter
th
at,
we
au
g
m
en
ted
o
u
r
d
ata
f
o
r
m
o
r
e
ac
cu
r
ate
tr
a
in
in
g
,
we
f
ee
d
two
n
etwo
r
k
s
,
o
n
e
f
o
r
r
ea
l
im
a
g
es
an
d
an
o
th
er
f
o
r
s
k
etch
im
a
g
es
as
s
h
o
wn
in
F
ig
u
r
e
1
.
As
tr
ai
n
in
g
d
ata,
we
u
s
e
C
E
L
E
B
A
[
1
9
]
attr
ib
u
te
d
atasets
with
m
o
r
e
th
a
n
2
0
0
0
0
0
im
a
g
es,
ea
ch
with
4
0
attr
ib
u
tes.
A
t
last
,
we
g
en
er
ate
im
a
g
es
f
r
o
m
th
e
m
o
d
el
Sty
le
-
GAN
an
d
we
ap
p
ly
u
p
o
n
it o
u
r
m
eth
o
d
to
cr
ea
te
a
n
ew
d
ataset
ab
le
to
im
p
r
o
v
e
o
u
r
tr
ain
in
g
an
d
o
p
en
th
e
d
o
o
r
f
o
r
o
u
r
n
e
x
t r
esear
ch
.
F
a
c
e
D
e
te
c
to
r
T
r
a
n
s
f
o
r
m
to
s
k
e
tc
h
(
1
)
(
2
)
9
2
%
M
a
le
G
e
n
d
e
r
:
7
5
%
Y
o
u
n
g
7
3
%
N
o
_
B
e
a
r
d
4
1
%
Sm
ilin
g
(a
)
D
a
t
a
P
r
e
p
r
o
c
e
s
s
in
g
(b
)
A
BM
-
C
N
N
(c
)
P
r
e
d
ic
t
io
n
P
h
a
s
e
Fig
u
r
e
1
.
AB
M
-
C
NN
t
r
ain
in
g
m
o
d
el
4
.
1
.
CNN
a
rc
hite
ct
ure
W
e
in
itialize
th
e
p
ar
am
eter
s
f
o
r
o
u
r
C
NN
with
a
b
atch
s
ize
with
3
2
h
eig
h
t
an
d
wid
th
with
1
8
0
,
s
o
th
e
in
p
u
t
lay
er
h
as
1
8
0
*
1
8
0
*
3
n
eu
r
o
n
s
,
with
1
0
in
ter
m
ed
iat
e
h
id
d
en
lay
er
s
,
3
C
o
n
v
o
lu
tio
n
2
D
lay
er
s
(
1
6
,
3
2
,
6
4
)
,
3
Ma
x
_
Po
o
lin
g
2
D
lay
er
s
(
1
6
,
3
2
,
6
4
)
,
1
Dr
o
p
o
u
t
lay
er
(
6
4
)
,
1
Flatten
lay
er
(
3
0
9
7
6
)
,
2
Den
s
e
lay
er
s
(
1
2
8
,
2
)
,
an
d
2
o
u
tp
u
t
n
eu
r
o
n
s
f
o
r
th
e
b
i
n
ar
y
class
ificatio
n
as
s
h
o
wn
in
F
ig
u
r
e
2
.
Ou
r
m
o
d
el
was
tr
ain
ed
b
et
wee
n
1
5
an
d
2
0
ep
o
ch
s
,
all
th
e
d
ata
av
ailab
le
is
p
ass
ed
th
r
o
u
g
h
th
e
n
eu
r
al
n
etwo
r
k
1
5
with
s
o
m
e
attr
ib
u
te
an
d
2
0
tim
es
with
o
th
er
s
.
W
e
u
s
e
R
elu
as
ac
t
iv
atio
n
f
u
n
ctio
n
,
wh
ile
th
e
o
p
tim
izer
f
u
n
ctio
n
was
‘
Ad
am
’
,
an
d
we
u
s
ed
Acc
u
r
ac
y
as m
etr
ics.
W
e
tr
ain
o
v
er
4
m
illi
o
n
p
a
r
am
eter
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
23
,
No
.
2
,
Au
g
u
s
t 2
0
2
1
:
973
-
9
7
9
976
C
o
n
v
2
D
(
1
6
)
M
a
x
P
o
o
l
i
n
g
2
D
C
o
n
v
2
D
(
3
2
)
C
o
n
v
2
D
(
6
4
)
M
a
x
P
o
o
l
i
n
g
2
D
M
a
x
P
o
o
l
i
n
g
2
D
D
r
o
p
o
u
t
F
l
a
t
t
e
n
D
e
n
s
e
(
1
2
8
)
D
e
n
s
e
(
2
)
Fig
u
r
e
2
.
C
NN
m
o
d
el
lay
e
r
s
4
.
2
.
Ste
p o
f
o
ur
wo
r
k
−
F
i
r
s
t
p
h
a
s
e
:
P
r
e
p
a
r
e
o
u
r
d
a
t
a
;
S
p
l
i
t
o
u
r
d
a
t
a
s
e
t
i
n
t
o
4
0
f
o
l
d
e
r
s
.
E
v
e
r
y
f
o
l
d
e
r
c
o
n
t
a
i
n
s
s
u
b
f
o
l
d
e
r
a
t
t
r
i
b
u
t
e
s
a
n
d
o
p
p
o
s
i
t
e
a
tt
r
i
b
u
t
es
,
a
n
d
e
v
e
n
e
v
e
r
y
s
u
b
f
o
l
d
e
r
c
o
n
t
a
i
n
s
t
h
r
ee
s
u
b
f
o
l
d
e
r
s
:
t
r
a
i
n
s
,
v
a
li
d
a
t
i
o
n
a
n
d
t
e
s
ts
.
W
e
d
i
s
t
r
i
b
u
t
e
o
u
r
d
a
ta
i
n
t
h
e
f
o
l
l
o
w
i
n
g
f
o
r
m
:
7
0
%
f
o
r
t
h
e
t
r
a
i
n
f
o
l
d
e
r
,
a
n
d
1
5
%
f
o
r
e
a
c
h
v
a
l
i
d
a
ti
o
n
a
n
d
t
e
s
t
.
−
Seco
n
d
p
h
ase:
T
r
an
s
f
o
r
m
d
at
a;
f
o
r
r
u
n
n
in
g
alg
o
r
ith
m
s
o
n
b
o
th
r
ea
l
h
u
m
a
n
f
ac
e
p
h
o
to
s
an
d
s
k
etch
o
n
es,
u
n
f
o
r
t
u
n
ately
th
e
d
ataset
av
ailab
le
o
f
s
k
etch
p
h
o
to
s
is
lim
ited
.
Fo
r
th
at
r
ea
s
o
n
,
we
ar
e
g
o
i
n
g
to
tr
an
s
f
o
r
m
C
E
L
E
B
A
to
s
k
etch
.
−
T
h
ir
d
p
h
ase
:
Data
s
et
p
r
e
p
r
o
ce
s
s
in
g
;
First
o
f
all,
a
p
p
ly
th
e
a
u
g
m
en
ted
f
u
n
ctio
n
o
n
d
atasets
,
r
ea
d
all
im
a
g
es
in
o
u
r
t
r
ain
f
o
l
d
er
an
d
tr
an
s
f
o
r
m
-
it
to
m
atr
ices
an
d
v
ec
to
r
s
t
o
tr
ain
o
u
r
m
o
d
el
to
r
ea
ch
th
e
g
o
al
o
f
t
h
e
b
est
p
er
f
o
r
m
an
ce
.
−
Fo
u
r
th
Ph
ase:
B
u
ild
our
m
o
d
e
l
s
;
o
n
ce
we
f
in
is
h
tr
ain
in
g
o
u
r
m
o
d
els,
we
s
av
e
th
em
.
E
v
er
y
m
o
d
el
is
f
it
o
n
a
tr
ain
in
g
da
taset,
th
a
n
th
e
test
d
ataset
is
u
s
ed
to
v
alid
ate
th
e
ac
cu
r
ac
y
o
f
o
u
r
f
in
al
m
o
d
el.
−
Fifth
Ph
ase:
T
h
e
E
v
alu
ati
on
o
f
o
u
r
f
i
n
al
m
o
d
el
q
u
ality
b
y
p
r
e
d
icti
n
g
o
u
ts
id
e
d
ata
a
n
d
c
o
m
p
ar
e
th
e
r
esu
lts
.
−
Six
th
Ph
ase:
W
e
co
m
b
in
e
th
e
p
r
ed
ict
r
esu
lt
o
n
b
o
th
r
ea
l
im
ag
e
an
d
s
k
etch
im
ag
e
b
y
ca
lcu
latin
g
b
est
ac
cu
r
ac
y
f
o
r
ea
ch
attr
ib
u
te.
−
Sev
en
th
Ph
ase:
C
r
ea
te
n
ew
d
ataset
th
at
co
n
tain
s
g
en
er
ated
f
ac
ial
im
ag
es
b
y
Sty
le
-
GAN.
W
e
also
g
en
er
ate
a
C
SV f
ile
f
o
r
attr
ib
u
te
class
if
icatio
n
r
esu
lts
.
T
h
is
will b
e
d
o
n
e
in
f
u
t
u
r
e
wo
r
k
.
4
.
2
.
H
a
rdwa
re
cha
ra
ct
er
is
t
i
cs
T
o
test
o
u
r
DL
-
m
o
d
el
,
we
u
s
e
th
e
r
em
o
tely
ac
ce
s
s
ib
le
h
ig
h
-
p
er
f
o
r
m
an
ce
c
o
m
p
u
tin
g
(
HPC
)
in
f
r
astru
ctu
r
e
C
lu
s
ter
HPC
-
MA
R
W
AN
[
2
0
]
:
−
C
o
m
p
u
te
No
d
es: 2
*
I
n
tel
Xeo
n
Go
ld
6
1
4
8
(
2
.
4
GHz
/2
0
-
co
r
e
)
/ 1
9
2
GB
R
AM
−
GPU
No
d
e
: 2
*
NVI
DI
A
T
esla P1
0
0
/ 1
9
2
GB
R
AM
−
Sto
r
ag
e
No
d
e:
2
*
I
n
tel
Xeo
n
Sil
v
er
4
1
1
4
(
2
.
2
GHz
/2
0
-
co
r
e
)
/
1
8
*
SATA
6
T
B
5.
RE
SU
L
T
AND
DI
SCUS
SI
O
N
W
e
s
tar
t
o
u
r
ex
p
er
ien
ce
with
tr
ain
in
g
an
d
co
m
p
a
r
in
g
r
esid
u
al
n
etwo
r
k
s
u
s
in
g
o
n
C
eleb
A
d
ataset
.
Af
ter
th
e
im
p
lem
en
tatio
n
o
f
A
B
M
-
C
NN
alg
o
r
ith
m
,
we
ev
alu
ate
ea
ch
n
eu
r
al
n
etwo
r
k
o
n
ev
er
y
f
ac
ial
attr
ib
u
te
,
b
y
th
eir
ef
f
icien
cy
u
s
in
g
ac
c
u
r
ac
y
,
lo
s
s
,
v
alid
atio
n
ac
cu
r
a
cy
,
v
alid
atio
n
t
r
ain
in
g
,
ep
o
c
h
s
an
d
tr
ain
in
g
tim
e.
W
e
co
n
d
u
cted
th
e
e
x
p
er
im
e
n
t
with
7
0
%
o
f
o
u
r
d
ataset
f
o
r
tr
ain
in
g
,
1
5
%
f
o
r
test
in
g
an
d
th
e
s
am
e
r
ate
f
o
r
v
alid
atio
n
,
a
b
atch
s
ize
o
f
3
2
,
an
d
a
n
u
m
b
er
o
f
e
p
o
ch
s
b
etw
ee
n
1
5
an
d
2
0
.
Fo
r
all
attr
ib
u
t
es,
we
h
av
e
alm
o
s
t
4
m
illi
o
n
tr
ain
ab
le
p
ar
am
ete
r
s
.
As
s
h
o
wn
in
T
ab
le
1
,
th
e
r
esu
lts
o
f
o
u
r
ap
p
r
o
ac
h
ar
e
p
r
o
m
is
in
g
.
W
e
n
o
tice
a
b
etter
p
er
f
o
r
m
an
ce
,
a
n
av
er
a
g
e
o
f
all
attr
ib
u
tes
r
ea
ch
in
g
9
0
.
0
5
%
ac
cu
r
ac
y
with
r
ea
l
i
m
ag
es
an
d
8
8
.
9
3
%
ac
cu
r
ac
y
with
s
k
etch
,
wh
ic
h
g
iv
es
u
s
a
v
er
y
g
o
o
d
p
r
e
d
ictio
n
r
esu
lt.
W
e
s
ee
s
o
m
e
attr
ib
u
tes
ex
ce
ed
th
e
9
6
%
ac
cu
r
ac
y
m
a
r
k
s
s
u
ch
as
(
g
en
d
er
,
b
ald
n
ess
,
ey
eg
lass
,
p
ale
s
k
in
,
h
air
c
o
lo
r
)
,
wh
ile
th
e
o
th
er
s
d
o
n
o
t
g
o
d
o
w
n
to
7
4
.
6
1
%
ac
cu
r
ac
y
;
th
is
p
er
ce
n
t
ag
e
r
etu
r
n
s
to
th
e
o
v
al
f
ac
e
att
r
ib
u
te
in
s
k
etch
,
wh
ile
th
e
av
e
r
ag
e
d
u
r
atio
n
o
f
an
ep
o
ch
is
1
0
m
in
,
o
r
alm
o
s
t
4
h
f
o
r
2
0
ep
o
ch
s
with
1
4
0
m
s
/
s
t
ep
.
T
h
e
r
esu
lt
o
f
tr
ain
in
g
r
ea
l
i
m
ag
e
is
b
etter
th
a
n
s
k
etch
f
o
r
m
o
s
t
attr
ib
u
tes.
Ho
wev
er
,
in
F
ig
u
r
e
3
we
n
o
tice
i
n
s
o
m
e
attr
ib
u
tes
tr
ain
i
n
g
n
eu
r
al
with
s
k
etc
h
th
at
we
ca
n
p
r
ed
ict
b
etter
r
esu
lts
f
o
r
th
e
f
o
llo
win
g
attr
ib
u
te
(
b
al
d
,
b
ig
lip
s
,
d
o
u
b
le
ch
i
n
,
ey
eg
l
ass
es,
n
o
b
ea
r
d
)
.
W
e
co
n
clu
d
e
th
at
we
ca
n
s
k
etch
m
o
d
els
to
p
r
e
d
ict
th
ese
attr
ib
u
tes
in
s
tead
o
f
r
ea
l
im
ag
es
m
o
d
els.
Ou
r
ac
cu
r
ac
y
ju
m
p
s
f
r
o
m
9
0
.
4
% to
9
0
.
5
4
%,
a
s
u
ch
r
ate
g
i
v
es m
o
r
e
p
r
ec
is
ely
p
r
ed
ictio
n
th
an
th
e
n
o
r
m
al
m
o
d
els.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
A
u
g
men
ted
b
in
a
r
y
mu
lti
-
la
b
el
ed
C
N
N
fo
r
p
r
a
ctica
l fa
cia
l a
ttr
ib
u
te
cla
s
s
ifica
tio
n
(
Mo
h
a
mme
d
B
err
a
h
a
l
)
977
T
ab
le
1
.
R
esu
lts
o
f
tr
ain
in
g
AB
M
-
C
NN
alg
o
r
ith
m
o
n
ev
er
y
f
ac
ial
attr
ib
u
te,
b
o
t
h
with
c
ele
b
A
d
ataset
D
a
t
a
s
e
t
C
e
l
e
b
A
R
e
a
l
I
mag
e
s
C
e
l
e
b
a
A
S
k
e
t
c
h
I
mag
e
s
A
t
t
r
i
b
u
t
e
s
A
c
c
u
r
a
c
y
%
Lo
ss
%
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
%
V
a
l
i
d
a
t
i
o
n
Lo
ss
%
Ep
o
c
h
s
A
c
c
u
r
a
c
y
%
Lo
ss
%
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
%
V
a
l
i
d
a
t
i
o
n
Lo
ss
%
Ep
o
c
h
s
1
_
N
o
_
B
e
a
r
d
_
9
4
.
7
1
3
.
1
95
1
2
.
1
20
9
5
.
4
1
1
.
9
9
5
.
8
1
0
.
9
20
2
_
W
a
v
y
H
a
i
r
_
8
6
.
4
37
8
3
.
6
3
6
.
8
20
8
1
.
1
4
0
.
8
82
3
9
.
7
20
3
_
5
o
’
c
l
o
c
k
S
h
a
d
o
w
_
94
17
9
4
.
1
1
7
.
1
15
9
1
.
8
1
8
.
7
9
2
.
1
1
8
.
2
20
4
_
A
t
t
r
a
c
t
i
v
e
_
81
41
80
42
15
8
0
.
3
4
1
.
1
8
0
.
5
4
1
.
2
20
5
_
B
r
o
w
n
_
H
a
i
r
_
8
6
.
5
30
87
3
0
.
5
15
8
1
.
3
3
9
.
8
81
4
0
.
8
20
6
_
B
a
g
s
_
U
d
e
r
_
E
y
e
_
8
3
.
5
35
83
38
15
8
3
.
2
3
6
.
2
8
3
.
1
3
6
.
4
5
20
7
_
S
i
d
e
b
u
r
n
s_
9
8
.
6
7
.
8
9
6
.
5
7
.
8
20
9
6
.
6
8
.
3
9
6
.
7
7
.
8
20
8
_
B
i
g
_
Li
p
s
_
7
7
.
4
4
8
.
2
7
7
.
7
4
7
.
5
15
8
1
.
9
3
8
.
9
8
1
.
7
0
.
3
9
20
9
_
B
i
g
_
N
o
s
e
_
83
3
8
.
5
8
3
.
1
38
15
7
7
.
2
4
8
.
9
7
7
.
2
4
9
.
2
20
1
0
_
A
r
c
h
e
d
_
Ey
e
b
r
o
w
_
83
36
8
3
.
8
3
5
.
2
15
8
2
.
9
3
5
.
9
8
3
.
4
35
20
1
1
_
B
l
o
n
d
_
H
a
i
r
_
96
1
2
.
8
9
6
.
1
1
2
.
5
15
9
3
.
6
1
5
.
5
9
3
.
8
1
5
.
5
20
1
2
_
B
l
u
r
r
y
_
9
5
.
8
11
9
6
.
5
1
0
.
2
15
9
5
.
7
1
1
.
7
9
5
.
3
1
8
.
2
20
1
3
_
B
u
s
h
y
_
Ey
e
b
r
o
w
_
92
2
1
.
8
9
0
.
8
2
2
.
1
15
9
0
.
5
23
9
0
.
9
22
20
1
4
_
C
h
u
b
b
y
_
95
1
3
.
5
9
5
.
1
1
2
.
8
15
94
1
3
.
9
95
1
3
.
3
20
1
5
_
E
y
e
g
l
a
sses
_
9
9
.
5
2
.
2
9
9
.
5
2
15
9
9
.
6
2
.
7
9
9
.
5
2
.
7
20
1
6
_
M
o
u
t
h
_
O
p
e
n
_
93
1
9
.
5
9
3
.
5
1
8
.
5
15
9
0
.
2
2
3
.
7
9
0
.
4
2
3
.
9
20
1
7
_
M
u
s
t
a
c
h
e
_
9
7
.
2
9
9
8
.
1
8
.
7
15
9
6
.
1
9
.
7
9
6
.
1
9
.
3
20
1
8
_
N
a
r
r
o
w
E
y
e
s
_
8
9
.
8
26
90
2
5
.
4
15
8
9
.
8
2
6
.
3
8
9
.
8
2
6
.
4
20
1
9
_
O
v
a
l
_
F
a
c
e
_
7
6
.
8
5
2
.
5
7
5
.
6
5
1
.
6
20
7
4
.
6
5
2
.
7
75
5
1
.
9
20
2
0
_
P
a
l
e
_
S
k
i
n
_
9
8
.
3
8
.
7
9
6
.
7
8
.
8
20
9
5
.
8
1
5
.
5
9
5
.
7
1
6
.
4
20
2
1
_
D
o
u
b
l
e
_
C
h
i
n
_
9
5
.
9
11
96
1
0
.
5
15
96
11
9
6
.
1
1
0
.
5
20
2
2
_
P
o
i
n
t
y
_
N
o
se
_
77
4
8
.
9
7
5
.
8
0
.
4
9
20
7
5
.
5
4
9
.
7
7
5
.
7
4
9
.
7
20
2
3
_
R
e
c
e
d
e
H
a
i
r
_
9
3
.
9
1
5
.
7
9
3
.
5
1
5
.
3
20
9
2
.
8
1
6
.
9
93
1
6
.
2
20
2
4
_
R
o
s
y
C
h
e
e
k
s
_
9
4
.
9
1
2
.
1
95
1
2
.
2
20
9
4
.
2
1
4
.
2
9
3
.
8
1
4
.
7
9
20
2
5
_
S
m
i
l
i
n
g
_
9
3
.
8
1
9
.
7
9
1
.
9
1
9
.
5
20
9
0
.
8
2
1
.
6
9
1
.
4
2
0
.
5
20
2
6
_
B
a
l
d
_
99
3
.
5
9
8
.
6
3
.
2
15
9
9
.
4
4
9
9
.
3
4
.
1
20
2
7
_
S
t
r
a
i
g
h
t
H
a
i
r
_
85
3
9
.
3
8
1
.
8
3
9
.
5
20
8
0
.
9
41
8
1
.
2
4
0
.
7
20
A
v
e
r
a
g
e
9
0
.
4
2
3
.
3
6
8
9
.
9
4
2
1
.
4
2
8
8
.
9
3
2
4
.
9
5
8
9
.
0
9
2
3
.
5
5
Fig
u
r
e
3
.
C
o
m
p
a
r
is
o
n
o
f
L
o
s
s
an
d
Acc
u
r
ac
y
f
o
r
ev
er
y
attr
ib
u
te
5
.
1
.
Co
m
pa
riso
n ABM
-
CN
N
wit
h F
AC
ex
is
t
ing
m
et
ho
ds
I
n
th
is
s
ec
tio
n
,
we
co
m
p
ar
e
th
e
p
er
f
o
r
m
an
ce
o
f
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
AB
M
-
C
NN
with
v
ar
io
u
s
FAC
m
eth
o
d
s
,
lik
e
Pan
d
a
[
2
1
]
,
th
is
m
eth
o
d
u
s
e
p
ar
t
b
ased
m
o
d
el
s
to
ex
tr
ac
t
f
ea
tu
r
es
an
d
SVM
[
2
2
]
as
class
if
ier
s
,
lik
e
L
Nets+A
Net
wh
ich
u
s
e
it
to
f
o
r
ea
ch
attr
ib
u
te,
b
esid
es
two
lo
ca
lizatio
n
n
etwo
r
k
s
an
d
o
n
e
attr
ib
u
te
n
etwo
r
k
.
MO
ON
[
2
3
]
a
n
a
p
p
r
o
ac
h
wh
ic
h
u
s
es
a
co
m
b
in
atio
n
o
f
o
p
tim
izatio
n
m
eth
o
d
s
r
eso
lv
in
g
th
e
is
s
u
es
o
f
im
b
alan
ce
p
r
o
b
lem
f
o
r
m
u
lti
-
lab
el,
as
NSA
[
2
4
]
an
ap
p
r
o
ac
h
u
s
es
th
e
s
eg
m
en
t
-
b
ased
m
eth
o
d
s
f
o
r
FAC
,
MCNN
-
AUX
[
2
5
]
in
an
o
th
er
h
an
d
b
ased
o
n
attr
ib
u
te
lo
ca
ti
o
n
wh
ich
d
iv
i
d
es
th
em
in
to
n
i
n
e
g
r
o
u
p
s
,
last
o
n
e
th
e
MCF
A
[
2
6
]
m
eth
o
d
wh
ic
h
ex
p
lo
its
th
e
in
h
er
en
t
d
ep
e
n
d
en
cies
b
etwe
en
FAC
an
d
a
u
x
iliar
y
task
s
(
f
ac
e
d
etec
tio
n
an
d
FLD)
.
As
s
h
o
w
n
in
T
ab
le
2
AB
M
-
C
NN
s
u
r
m
o
u
n
t
th
e
FAC
m
eth
o
d
s
we
ju
s
t
m
en
tio
n
ed
with
a
n
av
er
ag
e
ac
c
u
r
ac
y
o
f
9
0
.
5
4
%
o
f
all
attr
ib
u
tes
o
n
C
eleb
A
s
p
ec
ially
in
ce
r
tain
attr
ib
u
tes
(
B
ald
,
B
ig
_
L
ip
s
,
B
lo
n
d
_
Hair
,
Mu
s
tach
e,
Nar
r
o
w_
E
y
es,
Pale_
Sk
in
,
Sid
eb
u
r
n
s
,
Sm
ilin
g
,
W
av
y
_
Hair
)
wh
ic
h
o
v
er
p
ass
all
th
e
FAC
m
eth
o
d
s
,
h
o
wev
er
,
in
s
o
m
e
attr
ib
u
te
AB
M
-
C
N
N
f
ails
to
o
b
tain
a
h
ig
h
r
ate
as
h
is
co
m
p
etito
r
(
B
ag
s
_
Un
d
er
_
E
y
e
,
B
ig
_
No
s
e,
No
_
B
ea
r
d
,
R
o
s
y
_
C
h
ee
k
s
)
.
On
th
e
wh
o
le,
th
e
p
er
f
o
r
m
an
ce
co
m
p
ar
is
o
n
b
etwe
en
all
th
e
co
m
p
etin
g
m
eth
o
d
s
s
h
o
ws
th
e
ef
f
ec
tiv
en
ess
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
.
T
h
e
b
in
ar
y
class
if
icatio
n
m
eth
o
d
h
as
p
r
o
v
en
h
er
h
i
g
h
s
co
r
e
p
r
e
d
ictio
n
o
v
er
m
u
lti
-
class
o
r
m
u
lti
-
lab
le
u
s
ed
b
y
r
elate
d
wo
r
k
,
in
FAC
m
eth
o
d
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4
7
5
2
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l.
23
,
No
.
2
,
Au
g
u
s
t 2
0
2
1
:
973
-
9
7
9
978
T
ab
le
2
.
C
o
m
p
a
r
is
o
n
with
s
tate
-
of
-
th
e
-
ar
t FAC
m
eth
o
d
s
D
a
t
a
set
C
E
L
E
B
A
M
o
d
e
l
s
P
a
n
d
a
LN
e
t
s+A
N
e
t
M
O
O
N
N
S
A
M
C
N
N
-
AUX
M
C
F
A
A
B
M
-
C
N
N
A
t
t
r
i
b
u
t
e
s
1
_
Ey
e
g
l
a
ss
e
s
9
8
.
0
0
9
9
.
0
0
9
9
.
4
7
9
9
.
5
1
9
9
.
6
3
1
0
0
.
0
9
9
.
6
0
2
_
A
r
c
h
e
d
-
Ey
e
b
r
o
w
s
7
8
.
0
0
7
9
.
0
0
8
2
.
2
6
8
2
.
5
6
8
3
.
4
2
8
3
.
0
0
8
3
.
0
0
3
_
M
u
st
a
c
h
e
9
3
.
0
9
5
.
0
0
9
6
.
8
2
9
5
.
8
6
9
6
.
8
8
9
7
.
0
0
9
7
.
2
0
4
_
5
o
’
c
l
o
c
k
S
h
a
d
o
w
8
8
.
0
0
9
1
.
0
0
9
4
.
0
3
9
3
.
1
3
9
4
.
5
1
9
4
.
0
0
9
4
.
0
0
5
_
B
a
g
s
_
U
n
d
e
r
_
E
y
e
7
9
.
0
0
7
9
.
0
0
8
4
.
9
2
8
4
.
8
6
8
4
.
9
2
8
5
.
0
0
8
3
.
5
0
6
_
P
o
i
n
t
y
_
N
o
s
e
7
1
.
0
0
7
2
.
0
0
7
6
.
4
6
7
6
.
4
7
7
7
.
4
7
7
7
.
0
0
7
7
.
0
0
7
_
B
a
l
d
9
6
.
0
0
9
8
.
0
0
9
8
.
7
7
9
8
.
0
3
9
8
.
9
0
9
9
.
0
0
9
9
.
4
0
8
_
B
i
g
_
Li
p
s
6
7
.
0
0
6
8
.
0
0
7
1
.
4
8
6
9
.
2
8
7
1
.
4
7
7
2
.
0
0
8
1
.
9
0
9
_
B
i
g
_
N
o
s
e
7
5
.
0
0
7
8
.
0
0
8
4
.
0
0
8
3
.
8
1
8
4
.
5
3
8
4
.
0
0
8
3
.
0
0
1
0
_
B
l
o
n
d
_
H
a
i
r
9
3
.
0
0
9
5
.
0
0
9
5
.
8
6
9
5
.
7
6
9
6
.
0
1
9
6
.
0
0
9
6
.
0
0
1
1
_
N
a
r
r
o
w
_
E
y
e
s
8
4
.
0
0
8
1
.
0
0
8
6
.
5
2
8
6
.
8
8
8
7
.
2
3
8
7
.
0
0
8
9
.
7
5
1
2
_
B
u
s
h
y
_
Ey
e
b
r
o
w
8
6
.
0
0
9
0
.
0
0
9
2
.
6
2
9
2
.
6
6
9
2
.
8
4
9
2
.
0
0
9
2
.
0
0
1
3
_
A
t
t
r
a
c
t
i
v
e
8
1
.
0
0
8
1
.
0
0
8
1
.
6
7
8
2
.
7
6
8
3
.
0
6
8
3
.
0
0
8
1
.
0
0
1
4
_
C
h
u
b
b
y
8
6
.
0
0
9
1
.
0
0
9
5
.
4
4
9
4
.
9
4
9
5
.
6
7
9
6
.
0
0
9
5
.
0
0
1
5
_
D
o
u
b
l
e
_
C
h
i
n
8
8
.
0
0
9
2
.
0
0
9
6
.
3
2
9
5
.
8
0
9
6
.
3
2
9
6
.
0
0
9
6
.
0
0
1
6
_
M
o
u
t
h
-
O
p
e
n
9
3
.
0
0
9
2
.
0
0
9
3
.
5
4
9
3
.
7
8
9
3
.
7
4
9
3
.
0
0
9
3
.
0
0
1
7
_
S
i
d
e
b
u
r
n
s
9
3
.
0
0
9
6
.
0
0
9
7
.
5
9
9
7
.
1
7
9
7
.
8
5
9
8
.
0
0
9
8
.
5
0
1
8
_
B
l
u
r
r
y
8
6
.
0
0
8
4
.
0
0
9
5
.
6
7
9
5
.
9
6
9
6
.
1
7
9
6
.
0
0
9
5
.
8
0
1
9
_
O
v
a
l
_
F
a
c
e
6
5
.
0
0
6
6
.
0
0
7
5
.
7
3
7
4
.
9
3
7
5
.
8
4
7
5
.
0
0
7
5
.
6
1
2
0
_
P
a
l
e
_
S
k
i
n
9
1
.
0
0
9
1
.
0
0
9
7
.
0
0
9
7
.
0
0
9
7
.
0
5
9
7
.
0
0
9
8
.
3
0
2
2
_
N
o
_
B
e
a
r
d
9
3
.
0
0
9
5
.
0
0
9
5
.
5
8
9
6
.
1
7
9
6
.
0
5
9
6
.
0
0
9
5
.
4
0
2
3
_
B
r
o
w
n
_
H
a
i
r
7
7
.
0
0
8
0
.
0
0
8
9
.
3
8
8
8
.
2
5
8
9
.
1
5
8
8
.
0
0
8
6
.
5
0
2
4
_
R
e
c
e
d
e
_
H
a
i
r
8
5
.
0
0
8
9
.
0
0
9
3
.
5
6
9
2
.
2
5
9
3
.
8
1
9
4
.
0
0
9
3
.
9
0
2
5
_
R
o
s
y
_
C
h
e
e
k
s
8
7
.
0
0
9
0
.
0
0
9
4
.
8
2
9
4
.
7
9
9
5
.
1
6
9
5
.
0
0
9
4
.
4
0
2
6
_
S
t
r
a
i
g
h
t
_
H
a
i
r
6
9
.
0
0
7
3
.
0
0
8
2
.
2
6
8
0
.
4
1
8
3
.
5
8
8
5
.
0
0
8
5
.
0
0
2
7
_
S
m
i
l
i
n
g
9
2
.
0
0
9
2
.
0
0
9
2
.
6
0
9
2
.
7
0
9
2
.
7
3
9
3
.
0
0
9
3
.
8
0
2
8
_
W
a
v
y
_
H
a
i
r
7
7
.
0
0
8
0
.
0
0
8
2
.
4
7
8
1
.
7
0
8
3
.
9
1
8
5
.
0
0
8
6
.
0
0
A
v
e
r
a
g
e
84
.
0
0
8
5
.
6
4
8
9
.
5
9
8
9
.
2
9
9
0
.
0
3
8
9
.
9
6
9
0
.
5
4
5
.
2
.
G
ener
a
t
e
news
im
a
g
es a
nd
predict
a
t
t
ribute
f
o
r
new
da
t
a
s
et
I
n
t
h
i
s
p
a
r
t
,
w
e
t
r
y
t
o
c
r
e
a
t
e
a
n
e
w
d
a
t
a
s
e
t
b
y
g
e
n
e
r
a
t
i
n
g
i
m
ag
e
s
f
r
o
m
S
t
y
l
e
-
G
A
N
a
n
d
a
p
p
l
y
o
n
t
h
e
m
A
B
M
-
C
NN
f
o
r
t
h
e
e
x
t
r
a
c
ti
o
n
o
f
p
e
r
t
i
n
e
n
t
at
t
r
i
b
u
t
es
o
v
e
r
5
0
%
p
r
e
d
i
c
t
e
d
r
a
t
e
a
n
d
s
a
v
e
d
-
i
t
i
n
C
SV
f
i
le
l
i
k
e
is
s
h
o
w
n
i
n
F
i
g
u
r
e
4
.
W
e
ci
t
e
f
ew
r
e
a
s
o
n
s
w
it
h
a
d
a
p
t
e
d
AB
M
-
C
N
N
t
o
e
x
t
r
a
c
t
f
e
a
t
u
r
es
f
o
r
m
G
A
N
m
o
d
e
ls
[
2
7
]
,
i
)
c
r
e
a
t
e
d
a
t
a
s
e
t
o
f
h
i
g
h
q
u
a
l
i
t
y
c
o
m
p
a
r
e
d
t
o
e
x
i
s
t
i
n
g
d
a
t
a
s
e
t
s
a
n
d
u
n
l
i
m
i
t
e
d
d
a
t
a
;
ii
)
c
r
e
a
t
e
d
a
t
a
s
e
t
s
w
i
t
h
d
i
f
f
e
r
e
n
t
f
a
c
e
p
o
s
e
s
a
n
d
f
a
c
e
t
r
a
n
s
f
o
r
m
a
t
i
o
n
;
i
i
i
)
t
h
e
l
a
c
k
o
f
l
a
r
g
e
d
a
t
a
b
a
s
e
s
f
o
r
p
o
r
t
r
a
i
t
s
;
iv
)
t
h
e
p
r
o
b
l
e
m
o
f
p
o
l
i
c
y
i
n
c
e
r
t
a
i
n
p
r
i
v
a
t
e
d
a
t
a
s
e
t
;
v
)
R
e
s
e
a
r
c
h
i
s
m
o
v
i
n
g
t
o
w
a
r
d
s
t
h
e
p
a
t
h
o
f
d
a
t
a
g
e
n
e
r
a
t
i
o
n
a
n
d
u
n
s
u
p
e
r
v
i
s
e
d
l
e
a
r
n
i
n
g
.
Fig
u
r
e
4
.
Pre
d
ict
f
ac
ial
attr
ib
u
tes f
r
o
m
g
en
e
r
ated
im
ag
es u
s
i
n
g
AB
M
-
C
NN
6.
CO
NCLU
SI
O
N
I
n
th
is
p
a
p
er
,
we
p
r
o
p
o
s
e
a
n
o
v
el
f
r
am
ewo
r
k
ca
lled
A
u
g
m
en
ted
B
in
ar
y
m
u
lti
-
lab
el
C
NN
m
eth
o
d
(
AB
M
-
C
NN)
to
im
p
r
o
v
es
th
e
p
er
f
o
r
m
an
ce
o
f
FAC
.
T
h
e
m
ain
wo
r
k
o
f
o
u
r
m
eth
o
d
is
h
an
d
lin
g
ea
ch
attr
ib
u
t
e
as
a
b
in
ar
y
p
r
o
b
lem
“e
x
is
ts
o
r
n
o
t
e
x
is
tin
g
”,
th
en
t
r
an
s
f
o
r
m
in
g
th
e
d
ata
g
iv
e
n
in
s
k
et
ch
an
d
r
u
n
n
in
g
two
n
eu
r
al
n
etwo
r
k
s
.
T
h
e
f
i
r
s
t
o
n
e
is
f
o
r
p
r
ed
ictin
g
r
ea
l
im
ag
es
a
n
d
th
e
s
ec
o
n
d
f
o
r
s
k
etch
im
ag
es.
E
x
p
er
im
e
n
ts
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esian
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4
7
5
2
A
u
g
men
ted
b
in
a
r
y
mu
lti
-
la
b
el
ed
C
N
N
fo
r
p
r
a
ctica
l fa
cia
l a
ttr
ib
u
te
cla
s
s
ifica
tio
n
(
Mo
h
a
mme
d
B
err
a
h
a
l
)
979
th
e
p
u
b
lic
C
eleb
A
d
ataset
h
a
v
e
d
em
o
n
s
tr
ated
th
at
AB
M
-
C
NN
ac
h
iev
es
h
ig
h
e
r
p
e
r
f
o
r
m
a
n
ce
co
m
p
ar
ed
with
s
ev
er
al
r
elate
d
wo
r
k
s
o
n
FAC
m
eth
o
d
s
.
Fin
ally
,
we
g
en
er
ate
im
ag
es
v
ia
Sty
le
-
GAN
a
n
d
ap
p
ly
o
n
it
o
u
r
ap
p
r
o
ac
h
to
r
eso
lv
e
s
o
m
e
ex
ci
tin
g
p
r
o
b
lem
s
th
at
r
esear
ch
er
s
u
s
u
ally
f
ac
e
in
th
e
f
ield
.
RE
F
E
R
E
NC
E
S
[1
]
I.
Id
rissi,
M
.
B
o
u
k
a
b
o
u
s,
M
.
A
z
izi,
O.
M
o
u
ss
a
o
u
i,
a
n
d
H.
El
F
a
d
il
i,
“
To
wa
r
d
a
d
e
e
p
lea
rn
in
g
-
b
a
se
d
in
tru
si
o
n
d
e
tec
ti
o
n
s
y
ste
m
fo
r
Io
T
a
g
a
in
st b
o
t
n
e
t
a
tt
a
c
k
s,”
IAE
S
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
Arti
fi
c
ia
l
In
tell
ig
e
n
c
e
IJ
-
AI
,
v
o
l
.
1
0
,
n
o
.
1
,
p
p
.
1
1
0
-
1
2
0
,
2
0
2
1
,
d
o
i:
1
0
.
1
1
5
9
1
/i
jai.
v
1
0
.
i1
.
p
p
1
1
0
-
1
2
0
.
[2
]
I.
Id
rissi,
M
.
Az
izi,
a
n
d
O.
M
o
u
ss
a
o
u
i,
“
Io
T
se
c
u
rit
y
with
De
e
p
Lea
rn
in
g
-
b
a
se
d
In
tru
si
o
n
De
tec
ti
o
n
S
y
ste
m
s:
A
sy
ste
m
a
ti
c
li
tera
tu
re
re
v
iew
,
”
Fo
u
rth
I
n
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
In
telli
g
e
n
t
Co
m
p
u
t
in
g
in
D
a
ta
S
c
ien
c
e
s
ICDS
,
2
0
2
0
,
p
p
.
1
-
1
0
,
d
o
i:
1
0
.
1
1
0
9
/ICD
S
5
0
5
6
8
.
2
0
2
0
.
9
2
6
8
7
1
3
.
[3
]
H.
Ha
n
,
B.
F
.
Kla
re
,
K.
Bo
n
n
e
n
,
a
n
d
A.
K.
Ja
in
,
“
M
a
tch
i
n
g
C
o
m
p
o
site
S
k
e
tch
e
s
to
F
a
c
e
P
h
o
to
s:
A
Co
m
p
o
n
e
n
t
-
Ba
se
d
Ap
p
r
o
a
c
h
,
”
IEE
E
T
ra
n
s
.
o
n
I
n
f
o
rm
a
ti
o
n
Fo
re
n
sic
s
a
n
d
S
e
c
u
rity
,
v
o
l.
8
,
n
o
.
1
,
p
p
.
1
9
1
-
2
0
4
,
Ja
n
.
2
0
1
3
,
d
o
i:
1
0
.
1
1
0
9
/T
IF
S
.
2
0
1
2
.
2
2
2
8
8
5
6
.
[4
]
C.
D.
F
r
o
wd
,
e
t
a
l
.
,
“
Ca
tch
in
g
E
v
e
n
M
o
re
Offe
n
d
e
rs
wit
h
E
v
o
F
I
T
F
a
c
ial
Co
m
p
o
sites
,
”
T
h
ir
d
In
t
.
Co
n
f
e
re
n
c
e
on
Eme
rg
in
g
S
e
c
.
T
e
c
h
.
,
2
0
1
2
,
p
p
.
2
0
-
2
6
,
d
o
i
:
1
0
.
1
1
0
9
/E
S
T.
2
0
1
2
.
2
6
.
[5
]
X.
Zh
e
n
g
,
Y.
G
u
o
,
H.
Hu
a
n
g
,
Y.
Li
,
a
n
d
R.
He
,
“
A
S
u
rv
e
y
o
f
De
e
p
F
a
c
ial
Attri
b
u
te
An
a
ly
sis,”
In
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Co
m
p
u
ter
Vi
si
o
n
,
v
o
l.
1
2
8
,
n
o
.
8
-
9
,
p
p
.
2
0
0
2
-
2
0
3
4
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
1
2
6
3
-
0
2
0
-
0
1
3
0
8
-
z.
[6
]
M
.
Bo
u
k
a
b
o
u
s
a
n
d
M
.
Az
izi
,
“
Re
v
iew
o
f
Lea
rn
in
g
-
Ba
se
d
Tec
h
n
iq
u
e
s
o
f
S
e
n
ti
m
e
n
t
An
a
l
y
si
s
fo
r
S
e
c
u
rit
y
P
u
rp
o
se
s,”
In
n
o
v
.
i
n
S
m
a
rt Ci
ti
e
s
Ap
p
.,
v
o
l.
4
,
p
p
.
9
6
-
1
0
9
,
2
0
2
1
,
d
o
i:
d
o
i.
o
r
g
/
1
0
.
1
0
0
7
/9
7
8
-
3
-
0
3
0
-
6
6
8
4
0
-
2
_
8
.
[7
]
A.
Kh
a
n
,
A.
S
o
h
a
il
,
U.
Za
h
o
o
ra
,
a
n
d
A.
S
.
Qu
re
sh
i
,
“
A
su
rv
e
y
o
f
th
e
re
c
e
n
t
a
rc
h
it
e
c
tu
re
s
o
f
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
s,”
Arti
fi
c
ia
l
In
telli
g
e
n
c
e
Rev
iew V
o
lu
me
,
v
o
l.
5
3
,
n
o
.
8
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
0
4
6
2
-
0
2
0
-
0
9
8
2
5
-
6.
[8
]
L.
M
a
o
,
Y.
Ya
n
,
J.
X
u
e
,
a
n
d
H.
Wan
g
,
“
De
e
p
M
u
lt
i
-
tas
k
M
u
lt
i
-
la
b
e
l
CNN
fo
r
Eff
e
c
ti
v
e
F
a
c
ial
Attri
b
u
te
Clas
sifica
ti
o
n
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s o
n
Af
fec
ti
v
e
Co
m
p
u
ti
n
g
,
d
o
i:
1
0
.
1
1
0
9
/T
AFF
C
.
2
0
2
0
.
2
9
6
9
1
8
9
.
[
9
]
N
.
Z
h
u
a
n
g
,
Y
.
Y
a
n
,
S
.
C
h
e
n
,
H
.
W
a
n
g
,
a
n
d
C
.
S
h
e
n
,
“
M
u
l
t
i
-
l
a
b
e
l
L
e
a
r
n
i
n
g
B
a
se
d
D
e
e
p
T
r
a
n
s
f
e
r
N
e
u
r
a
l
N
e
t
w
o
r
k
f
o
r
F
a
c
i
a
l
A
t
t
r
i
b
u
t
e
C
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
P
a
t
t
e
r
n
R
e
c
o
g
n
i
t
,
v
o
l
.
8
0
,
p
p
.
2
2
5
-
2
4
0
,
2
0
1
8
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
p
a
t
c
o
g
.
2
0
1
8
.
0
3
.
0
1
8
.
[1
0
]
Y.
Wu
,
T.
Ha
ss
n
e
r,
K.
Kim
,
G
.
M
e
d
io
n
i,
a
n
d
P
.
Na
tara
jan
,
“
F
a
c
ial
Lan
d
m
a
rk
De
tec
ti
o
n
with
Twe
a
k
e
d
Co
n
v
o
l
u
ti
o
n
a
l
Ne
u
ra
l
Ne
two
rk
s,
”
IEE
E
T
ra
n
s
.
o
n
Pa
tt
e
rn
An
a
l
y
sis
a
n
d
M
a
c
h
in
e
I
n
telli
g
e
n
c
e
,
v
o
l.
4
0
,
n
o
.
1
2
,
p
p
.
3
0
6
7
-
3
0
7
4
,
2
0
1
8
,
d
o
i:
1
0
.
1
1
0
9
/T
P
AMI
.
2
0
1
7
.
2
7
8
7
1
3
0
.
[1
1
]
T.
Ka
rra
s,
S
.
Lain
e
,
M
.
Aitt
a
la,
J.
He
ll
ste
n
,
J.
Leh
ti
n
e
n
,
a
n
d
T
.
Ail
a
,
“
An
a
ly
z
in
g
a
n
d
Im
p
ro
v
in
g
th
e
Im
a
g
e
Qu
a
li
ty
o
f
S
t
y
leG
AN
,
”
Pro
c
.
o
f
t
h
e
IEE
E/
CVF
Co
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
Vi
sio
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
i
ti
o
n
CV
PR
,
2
0
2
0
.
[1
2
]
C.
S
h
o
rte
n
a
n
d
T.
M
.
K
h
o
sh
g
o
f
t
a
a
r,
“
A
su
r
v
e
y
o
n
Im
a
g
e
Da
ta
Au
g
m
e
n
tati
o
n
fo
r
De
e
p
Lea
rn
in
g
,
”
J
o
u
r
n
a
l
o
f
Bi
g
Da
ta
,
v
o
l.
6
,
n
o
.
1
,
p
p
.
1
-
4
8
,
2
0
1
9
,
d
o
i:
1
0
.
1
1
8
6
/s
4
0
5
3
7
-
0
1
9
-
0
1
9
7
-
0.
[1
3
]
Bin
a
ry
c
las
sifica
ti
o
n
-
Wi
k
i
p
e
d
ia.
[
On
li
n
e
].
A
v
a
il
a
b
le:
h
tt
p
s://
e
n
.
wi
k
ip
e
d
ia.o
rg
/wi
k
i/
Bi
n
a
ry
_
c
las
sifica
ti
o
n
.
[1
4
]
A.
Bu
y
u
k
c
a
k
ir,
H.
Bo
n
a
b
,
a
n
d
F
.
Ca
n
,
“
A
n
o
v
e
l
o
n
l
in
e
sta
c
k
e
d
e
n
se
m
b
le
fo
r
m
u
lt
i
-
la
b
e
l
stre
a
m
c
las
sifica
ti
o
n
,
”
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
In
fo
rm
a
ti
o
n
a
n
d
Kn
o
wled
g
e
M
a
n
a
g
e
me
n
t
,
2
0
1
8
,
p
p
.
1
0
6
3
-
1
0
7
2
,
d
o
i:
1
0
.
1
1
4
5
/3
2
6
9
2
0
6
.
3
2
7
1
7
7
4
.
[1
5
]
M
.
E
h
rli
c
h
,
T
.
J.
S
h
ield
s,
T.
Alm
a
e
v
,
a
n
d
M
.
R.
Am
e
r,
“
F
a
c
ial
Attri
b
u
tes
Clas
sifica
ti
o
n
u
sin
g
M
u
lt
i
-
Tas
k
Re
p
re
s
e
n
tatio
n
Lea
rn
in
g
,
”
Pro
c
.
I
EE
E
Co
n
f
.
C
o
mp
.
Vi
si
o
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
it
io
n
W
o
rk
sh
o
p
s
,
2
0
1
6
,
p
p
.
4
7
-
55.
[1
6
]
H.
L.
Ca
,
M
.
M
a
n
d
e
l,
R
.
P
a
sc
a
n
u
,
Y.
Be
n
g
i
o
,
a
n
d
B.
U.
Ca
,
“
Lea
r
n
in
g
Alg
o
rit
h
m
s
fo
r
th
e
Clas
sific
a
ti
o
n
Re
stricte
d
Bo
lt
z
m
a
n
n
M
a
c
h
i
n
e
,
”
J
o
u
rn
a
l
o
f
M
a
c
h
i
n
e
L
e
a
r
n
in
g
Res
e
a
rc
h
,
v
o
l.
1
3
,
p
p
.
6
4
3
-
6
6
9
,
2
0
1
2
.
[1
7
]
H.
Din
g
,
H.
Z
h
o
u
,
S
.
K.
Zh
o
u
,
a
n
d
R.
Ch
e
ll
a
p
p
a
,
“
A
De
e
p
Ca
sc
a
d
e
Ne
two
rk
fo
r
Un
a
li
g
n
e
d
F
a
c
e
Attri
b
u
te
Clas
sifica
ti
o
n
,
”
3
2
n
d
AA
AI
Co
n
fe
re
n
c
e
Arti
fi
c
ia
l
I
n
telli
g
e
n
c
e
,
2
0
1
8
,
p
p
.
6
7
8
9
-
6
7
9
6
.
[1
8
]
K.
Kh
a
n
,
M
.
Atti
q
u
e
,
R
.
U.
Kh
a
n
,
I.
S
y
e
d
,
a
n
d
T.
S
.
Ch
u
n
g
,
“
A
M
u
lt
i
-
Tas
k
F
ra
m
e
wo
rk
f
o
r
F
a
c
ial
Attri
b
u
tes
Clas
sifica
ti
o
n
th
r
o
u
g
h
E
n
d
-
to
-
En
d
F
a
c
e
P
a
rsin
g
a
n
d
De
e
p
Co
n
v
o
l
u
ti
o
n
a
l
Ne
u
ra
l
Ne
two
rk
s,”
S
e
n
so
r
s
,
v
o
l.
2
0
,
n
o
.
2
,
d
o
i:
1
0
.
3
3
9
0
/s2
0
0
2
0
3
2
8
.
[1
9
]
Z.
Li
u
,
P
.
Lu
o
,
X.
Wan
g
,
a
n
d
X.
Tan
g
,
“L
a
rg
e
-
sc
a
le
Ce
leb
F
a
c
e
s
A
tt
rib
u
tes
(Ce
leb
A)
Da
tas
e
t.
”
[On
li
n
e
].
Av
a
il
a
b
le:
h
tt
p
:
//
m
m
lab
.
ie.cu
h
k
.
e
d
u
.
h
k
/
p
ro
jec
ts/Cele
b
A.h
tml
[2
0
]
Hig
h
P
e
rfo
rm
a
n
c
e
Co
m
p
u
ti
n
g
(H
P
C)
&
AMD.
[On
l
in
e
].
A
v
a
il
a
b
le
:
h
tt
p
s://
ww
w.m
a
rwa
n
.
m
a
/i
n
d
e
x
.
p
h
p
[2
1
]
N.
Zh
a
n
g
,
M
.
P
a
lu
r
i,
M
.
Ra
n
z
a
to
,
T.
Da
rre
ll
,
a
n
d
L
.
Bo
u
rd
e
v
,
“
P
AN
DA
:
P
o
se
Alig
n
e
d
Ne
t
wo
rk
s
fo
r
De
e
p
Attri
b
u
te
M
o
d
e
li
n
g
,
”
Pro
c
.
s
o
f
th
e
IEE
E
Co
n
f.
o
n
Co
mp
u
ter
Vi
si
o
n
a
n
d
P
a
tt
e
rn
Rec
o
g
n
it
io
n
,
p
p
.
1
6
3
7
-
1
6
4
4
,
2
0
1
3
.
[2
2
]
B.
S
c
h
o
l
k
o
p
f
a
n
d
A.
J.
S
a
m
o
la,
“
S
m
o
la.
A:
Lea
rn
in
g
with
Ke
r
n
e
ls
-
S
u
p
p
o
r
t
Ve
c
to
r
M
a
c
h
i
n
e
s,
Re
g
u
lariz
a
ti
o
n
,
Op
ti
m
iza
ti
o
n
a
n
d
Be
y
o
n
d
,
”
M
IT
Pre
ss
,
Ca
mb
rid
g
e
,
M
A,
2
0
0
1
.
[2
3
]
E.
M
.
Ru
d
d
,
M
.
G
ü
n
t
h
e
r,
a
n
d
T
.
E.
Bo
u
lt
,
“
M
OO
N:
A
m
ix
e
d
o
b
jec
ti
v
e
o
p
ti
m
iza
ti
o
n
n
e
two
r
k
f
o
r
t
h
e
re
c
o
g
n
it
i
o
n
o
f
fa
c
ial
a
tt
rib
u
tes
,
”
Eu
r
o
p
e
a
n
C
o
n
f
.
Co
mp
u
ter
Vi
sio
n
,
2
0
1
6
,
p
p
.
19
-
3
5
,
d
o
i:
1
0
.
1
0
0
7
/9
7
8
-
3
-
3
1
9
-
4
6
4
5
4
-
1
_
2
.
[2
4
]
U.
M
a
h
b
u
b
,
S
.
S
a
rk
a
r,
a
n
d
R.
Ch
e
ll
a
p
p
a
,
“
S
e
g
m
e
n
t
-
Ba
se
d
M
e
t
h
o
d
s
fo
r
F
a
c
ial
Attri
b
u
te
De
tec
ti
o
n
fro
m
P
a
rt
ial
F
a
c
e
s,”
IEE
E
T
ra
n
s
.
Af
fec
.
C
o
mp
u
ti
n
g
,
v
o
l
.
1
1
,
n
o
.
4
,
p
p
.
6
0
1
-
6
1
3
,
2
0
2
0
,
d
o
i:
1
0
.
1
1
0
9
/T
AFF
C.
2
0
1
8
.
2
8
2
0
0
4
8
.
[2
5
]
E
.
M
.
Ha
n
d
a
n
d
R.
Ch
e
ll
a
p
p
a
,
“
Attri
b
u
tes
fo
r
Im
p
r
o
v
e
d
Att
rib
u
tes
:
A
M
u
lt
i
-
Tas
k
Ne
two
r
k
fo
r
Attri
b
u
te
Clas
sifica
ti
o
n
,
”
a
rXiv C
o
rn
e
ll
U
n
ive
rs
it
y
,
2
0
1
6
.
[2
6
]
N.
Zh
u
a
n
g
,
Y.
Ya
n
,
S
.
C
h
e
n
,
a
n
d
H.
Wa
n
g
,
“
M
u
l
ti
-
tas
k
Lea
rn
in
g
o
f
Ca
sc
a
d
e
d
CNN
f
o
r
F
a
c
ial
Attri
b
u
te
Clas
sifi
c
a
ti
o
n
,
”
2
4
th
In
t
.
Co
n
f
.
o
n
Pa
tt
e
rn
Rec
o
g
n
it
i
o
n
IC
PR
,
2
0
1
8
,
d
o
i
:
1
0
.
1
1
0
9
/ICP
R
.
2
0
1
8
.
8
5
4
5
2
7
1
.
[2
7
]
M
.
Be
rra
h
a
l
a
n
d
M
.
Az
izi,
“
Re
v
iew
o
f
DL
-
Ba
se
d
G
e
n
e
ra
ti
o
n
Tec
h
n
iq
u
e
s
o
f
Au
g
m
e
n
te
d
Im
a
g
e
s
u
sin
g
P
o
rtraits
S
p
e
c
ifi
c
a
ti
o
n
,
”
Fo
u
rth
In
t
.
C
o
n
f
.
on
I
n
telli
g
e
n
t
Co
m
p
u
ti
n
g
in
Da
t
a
S
c
ien
c
e
s
ICDS
,
2
0
2
0
,
p
p
.
1
-
8
,
d
o
i:
1
0
.
1
1
0
9
/ICDS
5
0
5
6
8
.
2
0
2
0
.
9
2
6
8
7
1
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.