I
nd
o
ne
s
ia
n J
o
urna
l o
f
E
lect
rica
l En
g
ineering
a
nd
Co
m
p
u
t
er
Science
Vo
l.
1
2
,
No
.
2
,
N
o
v
e
m
b
er
201
8
,
p
p
.
4
7
6
~
4
8
1
I
SS
N:
2502
-
4752
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ee
cs
.
v
1
2
.i
2
.
p
p
4
7
6
-
4
8
1
476
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
co
r
e.
co
m/jo
u
r
n
a
ls
/in
d
ex
.
p
h
p
/
ijeec
s
Celebri
ty Face
Re
co
g
nition us
ing
D
eep Lea
rning
Nur
At
eqa
h B
inti
M
a
t
K
a
s
i
m
1
,
N
ur
H
ida
y
a
h B
inti
Abd
Ra
h
m
a
n
2
,
Z
a
ida
h I
bra
hi
m
3
,
Nur
Na
bil
a
h Ab
u M
a
ng
s
ho
r
4
1
,
2,
3
F
a
c
u
l
ty
o
f
Co
m
p
u
ter an
d
M
a
th
e
m
a
ti
c
a
l
S
c
ien
c
e
s
,
Un
iv
e
rsiti
T
e
k
n
o
lo
g
i
M
A
RA
(Ui
T
M
),
S
h
a
h
A
lam
,
S
e
lan
g
o
r,
M
a
lay
sia
4
F
a
c
u
lt
y
o
f
Co
m
p
u
ter an
d
M
a
th
e
m
a
ti
c
a
l
S
c
ien
c
e
s
,
U
n
iv
e
rsiti
T
e
k
n
o
lo
g
i
M
A
RA
(Ui
T
M
),
Ca
m
p
u
s Ja
sin
,
M
e
lak
a
,
M
a
lay
sia
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
y
2
9
,
2
0
1
8
R
ev
i
s
ed
J
u
l 3
0
,
2
0
1
8
A
cc
ep
ted
A
u
g
3
,
2
0
1
8
F
a
c
e
re
c
o
g
n
it
io
n
is
o
n
e
o
f
th
e
we
ll
stu
d
ied
p
ro
b
lem
s
b
y
r
e
se
a
rc
h
e
rs
in
c
o
m
p
u
ter
v
isio
n
s.
Am
o
n
g
th
e
c
h
a
ll
e
n
g
e
s
o
f
th
is
tas
k
a
r
e
th
e
o
c
c
u
rre
n
c
e
o
f
d
if
fe
re
n
t
f
a
c
ial
e
x
p
re
s
sio
n
s
li
k
e
h
a
p
p
y
o
r
sa
d
,
a
n
d
d
if
f
e
re
n
t
v
i
e
ws
o
f
th
e
im
a
g
e
s
su
c
h
a
s
f
ro
n
t
a
n
d
si
d
e
v
ie
w
s.
T
h
is
p
a
p
e
r
e
x
p
e
ri
m
e
n
ts
a
p
u
b
li
c
ly
a
v
a
il
a
b
l
e
d
a
tas
e
t
th
a
t
c
o
n
sists
o
f
2
0
0
,
0
0
0
im
a
g
e
s
o
f
c
e
leb
rit
y
f
a
c
e
s.
De
e
p
L
e
a
rn
in
g
tec
h
n
iq
u
e
is
g
a
in
i
n
g
it
s
p
o
p
u
larity
in
c
o
m
p
u
ter v
isio
n
a
n
d
th
is
p
a
p
e
r
a
p
p
li
e
s
th
is
tec
h
n
iq
u
e
f
o
r
f
a
c
e
re
c
o
g
n
it
io
n
p
r
o
b
lem
.
On
e
o
f
th
e
tec
h
n
iq
u
e
s
u
n
d
e
r
d
e
e
p
lea
rn
in
g
is
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
(CNN
).
T
h
e
re
is
a
lso
p
re
-
train
e
d
CNN
m
o
d
e
ls
th
a
t
a
r
e
A
lex
Ne
t
a
n
d
G
o
o
g
L
e
Ne
t,
w
h
ich
p
ro
d
u
c
e
e
x
c
e
ll
e
n
t
a
c
c
u
ra
c
y
re
su
lt
s.
T
h
e
e
x
p
e
rime
n
tal
re
su
lt
s
in
d
ica
te
th
a
t
A
le
x
Ne
t
is
b
e
tt
e
r
th
a
n
b
a
sic
CNN
a
n
d
G
o
o
g
L
e
N
e
t
f
o
r
fa
c
e
re
c
o
g
n
it
io
n
.
K
ey
w
o
r
d
s
:
A
le
x
Net
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
Dee
p
lear
n
in
g
Face
r
ec
o
g
n
itio
n
Go
o
g
L
eNe
t
Co
p
y
rig
h
t
©
2
0
1
8
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
:
Nu
r
A
teq
ah
B
in
ti M
at
Kasi
m
,
Facu
lt
y
o
f
C
o
m
p
u
ter
an
d
Ma
t
h
e
m
a
tical
Scie
n
ce
s
,
U
n
i
v
er
s
itiT
ek
n
o
lo
g
i M
AR
A
(
UiT
M)
,
Sh
a
h
A
la
m
,
Selan
g
o
r
,
Ma
la
y
s
i
a.
E
m
ail:
ateq
ah
k
a
s
i
m
o
f
f
icial
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
g
r
ea
t
p
r
o
g
r
ess
o
f
au
to
m
atic
f
ac
e
r
ec
o
g
n
itio
n
i
n
r
ec
en
t
y
ea
r
s
h
a
s
m
ad
e
lar
g
e
-
s
ca
le
f
ac
e
id
en
ti
f
icatio
n
p
o
s
s
ib
le
f
o
r
m
a
n
y
p
r
ac
tical
ap
p
licatio
n
s
[
1
]
.
T
h
is
ap
p
licatio
n
i
s
w
id
el
y
u
s
e
d
w
h
e
n
t
h
e
i
m
a
g
e
s
f
o
r
th
e
p
er
s
o
n
s
to
b
e
r
ec
o
g
n
i
ze
d
ar
e
av
ailab
le
b
ef
o
r
eh
a
n
d
,
an
d
a
n
ac
cu
r
ate
r
ec
o
g
n
izer
i
s
n
ee
d
ed
f
o
r
a
lar
g
e
an
d
r
elativ
el
y
f
ix
ed
g
r
o
u
p
o
f
p
eo
p
le.
Fo
r
ex
am
p
le,
m
o
s
t
o
f
th
e
f
ac
e
r
ec
o
g
n
it
io
n
ap
p
licatio
n
is
u
s
ed
f
o
r
s
ea
r
ch
en
g
i
n
e
[
2
]
,
r
ec
o
g
n
itio
n
f
o
r
p
u
b
lic
f
ig
u
r
e
i
n
m
ed
ia
i
n
d
u
s
tr
y
,
an
d
v
id
eo
s
tr
ea
m
i
n
g
co
m
p
a
n
ie
s
f
o
r
m
o
v
ie
ch
ar
ac
ter
an
n
o
tatio
n
[
3
]
.
I
n
t
h
is
p
ap
er
,
w
e
in
v
e
s
ti
g
ate
t
h
e
ap
p
licatio
n
o
f
d
ee
p
lear
n
i
n
g
m
et
h
o
d
s
n
a
m
el
y
C
o
n
v
o
lu
tio
n
Neu
r
al
Ne
t
w
o
r
k
(
C
NN)
an
d
t
w
o
p
r
e
-
tr
ain
ed
C
NN
m
o
d
els
t
h
at
ar
e
Alex
Ne
t
an
d
Go
o
g
L
eNe
t
f
o
r
f
ac
e
r
ec
o
g
n
i
tio
n
d
u
e
to
t
h
eir
ex
ce
lle
n
t a
cc
u
r
ac
y
p
e
r
f
o
r
m
a
n
ce
s
in
co
m
p
u
ter
v
i
s
io
n
.
2.
RE
L
AT
E
D
WO
RK
P
r
ev
io
u
s
l
y
,
r
esear
c
h
in
o
b
j
ec
t
r
ec
o
g
n
itio
n
u
s
es
h
an
d
cr
af
ted
f
ea
t
u
r
es
s
u
c
h
as
te
x
tu
r
e
f
ea
tu
r
es
f
o
r
f
all
ac
tiv
it
y
r
ec
o
g
n
itio
n
[
4
]
an
d
le
af
r
ec
o
g
n
itio
n
[
5
]
.
B
esid
es
t
h
at,
co
lo
r
f
ea
tu
r
es
h
a
v
e
also
b
e
en
ap
p
lied
f
o
r
f
r
u
it
r
ec
o
g
n
itio
n
[
6
]
w
h
er
e
it
i
n
v
o
l
v
es
id
en
tify
i
n
g
th
e
s
ig
n
i
f
ica
n
t
f
ea
tu
r
e
a
n
d
class
i
f
ier
to
o
b
tain
g
o
o
d
r
ec
o
g
n
itio
n
r
esu
lt
s
.
Ho
w
e
v
er
,
cu
r
r
en
tl
y
,
t
h
e
o
b
j
ec
t
r
ec
o
g
n
itio
n
r
esear
c
h
h
as
p
r
o
g
r
ess
ed
to
Dee
p
L
ea
r
n
in
g
(
D
L
)
w
h
er
e
n
o
h
an
d
cr
af
ted
f
ea
t
u
r
e
is
r
eq
u
ir
ed
an
d
y
et
th
e
r
es
u
lt
s
p
r
o
d
u
ce
d
ar
e
ex
ce
llen
t.
Dee
p
L
ea
r
n
i
n
g
(
D
L
)
w
a
s
ap
p
lied
to
s
o
lv
e
m
a
n
y
p
r
o
b
le
m
s
f
o
r
th
e
last
f
e
w
y
ea
r
s
.
T
h
e
p
r
o
b
l
e
m
s
r
an
g
e
f
r
o
m
co
m
p
u
ter
v
i
s
io
n
to
n
atu
r
al
la
n
g
u
a
g
e
p
r
o
ce
s
s
i
n
g
.
I
n
m
a
n
y
ca
s
es
D
L
o
u
tp
er
f
o
r
m
ed
o
th
er
e
x
i
s
ti
n
g
tech
n
iq
u
es
[
7
]
.
DL
m
e
th
o
d
s
s
tar
t
b
y
e
x
tr
ac
tin
g
a
r
ep
r
esen
tatio
n
o
f
th
e
f
ac
e
i
m
ag
e
u
s
i
n
g
lo
ca
l
i
m
a
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esia
n
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4752
C
eleb
r
ity
F
a
ce
R
ec
o
g
n
itio
n
u
s
in
g
Dee
p
Lea
r
n
in
g
(
N
u
r
A
teq
a
h
B
in
ti Ma
t Ka
s
im
)
477
d
escr
ip
to
r
s
.
T
h
en
th
e
y
a
g
g
r
eg
ate
s
u
c
h
lo
ca
l
d
escr
ip
to
r
s
u
s
i
n
g
p
o
o
lin
g
m
ec
h
a
n
i
s
m
i
n
to
an
o
v
er
al
l
f
ac
e
d
escr
ip
to
r
.
T
h
is
w
o
r
k
is
co
n
ce
r
n
ed
w
it
h
d
ee
p
ar
ch
itectu
r
es
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
.
T
h
e
d
ef
in
i
n
g
c
h
ar
ac
ter
is
tic
o
f
s
u
c
h
m
et
h
o
d
s
i
s
t
h
e
u
s
e
o
f
a
C
NN
f
ea
tu
r
e
ex
tr
ac
to
r
,
a
le
ar
n
ab
le
f
u
n
ctio
n
o
b
t
ain
ed
b
y
co
m
p
o
s
i
n
g
s
e
v
er
al
li
n
ea
r
an
d
n
o
n
-
li
n
ea
r
o
p
er
ato
r
s
.
Dee
p
Face
[
8
]
is
a
r
e
p
r
ese
n
tati
v
e
s
y
s
te
m
o
f
th
i
s
class
.
T
h
is
m
eth
o
d
u
s
e
s
a
d
ee
p
C
NN
to
class
if
y
f
ac
es
u
s
i
n
g
4
m
illi
o
n
e
x
a
m
p
le
s
o
f
tr
ain
i
n
g
i
m
a
g
es
s
p
an
n
i
n
g
4
0
0
0
u
n
iq
u
e
id
en
t
ities
.
T
h
e
s
a
m
e
C
NN
i
s
a
p
p
lied
to
p
air
f
ac
es
to
o
b
tain
d
escr
ip
to
r
s
an
d
th
e
m
atch
ed
s
i
m
ilar
it
y
i
m
ag
es
ar
e
p
er
f
o
r
m
ed
u
s
i
n
g
E
u
clid
ea
n
d
is
ta
n
ce
[
8
]
.
T
h
e
g
o
al
o
f
tr
ain
in
g
p
r
o
ce
s
s
i
s
to
m
a
x
i
m
is
e
th
e
d
is
ta
n
ce
b
e
t
w
ee
n
in
co
n
g
r
u
o
u
s
p
air
s
an
d
m
in
i
m
i
s
e
t
h
e
d
is
ta
n
ce
b
et
w
ee
n
co
n
g
r
u
o
u
s
p
air
s
o
f
f
ac
e
s
w
h
ic
h
is
p
o
r
tr
ay
i
n
g
th
e
s
a
m
e
id
en
tit
y
.
I
n
ad
d
itio
n
to
u
s
in
g
v
er
y
lar
g
e
a
m
o
u
n
t
o
f
tr
ai
n
in
g
d
ata,
Dee
p
Face
u
s
es
a
n
en
s
e
m
b
le
o
f
C
NN
as
w
ell
as
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
w
h
er
e
th
e
f
ac
e
i
m
a
g
es a
r
e
ali
g
n
ed
to
a
ca
n
o
n
ical
p
o
s
e
u
s
i
n
g
a
3
D
m
o
d
el.
An
o
th
er
ap
p
li
ca
tio
n
t
h
at
u
s
es
DL
is
a
u
to
m
atic
co
lo
r
izatio
n
o
f
b
lack
an
d
w
h
i
te
i
m
a
g
e
[
9
]
.
DL
ca
n
b
e
u
s
ed
to
co
lo
u
r
th
e
i
m
ag
e
b
y
u
s
i
n
g
th
e
o
b
j
ec
ts
an
d
th
eir
co
n
tex
t
w
ith
in
t
h
e
p
h
o
to
g
r
ap
h
.
I
t
ac
ts
m
u
c
h
li
k
e
a
h
u
m
a
n
o
p
er
ato
r
.
T
h
ese
ca
p
ab
i
lit
y
le
v
er
a
g
es
o
f
t
h
e
h
ig
h
q
u
ali
t
y
a
n
d
v
er
y
lar
g
e
C
N
N
tr
ain
ed
f
o
r
I
m
a
g
eNe
t
a
n
d
co
-
o
p
ted
f
o
r
th
e
p
r
o
b
lem
o
f
i
m
a
g
e
co
lo
r
izatio
n
.
T
h
e
ap
p
r
o
ac
h
in
v
o
lv
e
s
th
e
u
s
e
o
f
v
er
y
lar
g
e
C
NN
a
n
d
s
u
p
er
v
i
s
ed
la
y
er
s
t
h
at
r
ec
r
ea
te
th
e
i
m
a
g
e
w
it
h
t
h
e
ad
d
itio
n
o
f
co
lo
r
f
ea
tu
r
es.
B
esid
es
th
at,
D
L
ca
n
b
e
u
til
iz
ed
to
ad
d
s
o
u
n
d
s
to
s
ilen
t
m
o
v
ies.
I
n
t
h
i
s
tas
k
it
w
ill
s
y
n
t
h
e
s
i
ze
s
o
u
n
d
s
to
m
atc
h
a
s
ile
n
t
v
id
eo
[
1
0
]
.
T
h
e
s
y
s
te
m
is
tr
ai
n
ed
u
s
in
g
1
0
0
0
ex
a
m
p
le
s
o
f
v
id
eo
s
w
it
h
s
o
u
n
d
o
f
a
d
r
u
m
s
tick
s
tr
ik
i
n
g
d
if
f
er
e
n
t
s
u
r
f
ac
es.
A
DL
m
o
d
el
as
s
o
ciate
s
w
it
h
t
h
e
v
id
eo
f
r
a
m
es
o
f
p
r
e
-
r
ec
o
r
d
ed
s
o
u
n
d
s
i
n
o
r
d
er
to
s
elec
t
s
o
u
n
d
s
to
p
la
y
th
a
t
b
est
m
atc
h
es
w
it
h
t
h
e
s
ce
n
e
[
1
0
]
.
T
h
e
s
y
s
te
m
w
as
t
h
en
e
v
al
u
ate
d
u
s
in
g
a
tu
r
i
n
g
-
test
lik
e
s
e
tu
p
w
h
er
e
h
u
m
an
s
h
ad
to
d
eter
m
i
n
e
w
h
ic
h
v
id
eo
h
a
d
th
e
r
ea
l
o
r
th
e
f
ak
e
(
s
y
n
t
h
e
s
ized
)
s
o
u
n
d
s
[
1
0
]
.
I
t u
s
e
d
b
o
th
C
NN
a
n
d
L
ST
M
r
ec
u
r
r
en
t n
e
u
r
al
n
et
w
o
r
k
s
[
1
0
]
.
DL
also
ca
n
b
e
u
s
ed
to
class
if
y
a
n
d
d
etec
t
tex
t
an
d
o
b
j
ec
ts
in
p
h
o
to
g
r
ap
h
s
[
1
1
]
.
State
-
of
-
t
h
e
-
ar
t
r
esu
lt
s
h
a
v
e
b
ee
n
ac
h
ie
v
ed
o
n
b
en
ch
m
ar
k
e
x
a
m
p
les o
f
th
i
s
p
r
o
b
lem
u
s
i
n
g
v
er
y
lar
g
e
C
NN.
A
b
r
ea
k
t
h
r
o
u
g
h
in
t
h
is
p
r
o
b
le
m
b
y
A
le
x
Kr
iz
h
ev
s
k
y
et
al.
r
esu
l
ts
o
n
t
h
e
I
m
a
g
e
Net
class
i
f
icat
io
n
p
r
o
b
lem
ca
l
l
ed
A
lex
Net
[
1
1
]
.
3.
RE
S
E
ARCH
M
E
T
H
O
D
3
.
1
.
T
he
Da
t
a
s
et
C
eleb
r
it
y
f
ac
e
d
atase
t
h
a
s
b
ee
n
u
s
ed
f
o
r
tr
ai
n
in
g
w
h
e
r
e
it
s
to
r
es
at
m
o
s
t
2
0
0
,
0
0
0
an
d
4
0
attr
ib
u
tes
[
1
2
]
.
Dif
f
er
e
n
t
f
ac
e
ex
p
r
ess
io
n
s
,
v
ie
w
s
an
d
b
ac
k
g
r
o
u
n
d
ar
e
th
e
s
a
m
p
le
o
f
4
0
attr
ib
u
tes
i
n
d
icate
d
in
th
is
d
ataset.
Fi
g
u
r
e
1
s
h
o
w
s
th
e
s
a
m
p
le
attr
ib
u
tes
in
c
lu
d
es
i
n
th
i
s
d
ata
s
et.
T
h
er
e
ar
e
d
i
f
f
er
e
n
t
at
tr
ib
u
tes
i
n
t
h
e
d
atasets
; g
e
n
d
er
is
o
n
e
o
f
t
h
e
e
x
a
m
p
le
s
o
f
t
h
e
attr
ib
u
te
s
[
1
2
]
.
Fig
u
r
e
1
.
C
eleb
r
it
y
f
ac
e
cla
s
s
i
f
ied
b
y
g
e
n
d
er
[
1
2
]
3
.
2
.
Co
nv
o
lutio
na
l N
eura
l N
et
w
o
rk
C
o
n
v
o
lu
tio
n
al
Neu
r
al
Net
w
o
r
k
s
(
C
N
N)
h
av
e
ta
k
e
n
th
e
co
m
p
u
ter
v
is
io
n
co
m
m
u
n
it
y
b
y
s
to
r
m
,
it
is
s
ig
n
i
f
ica
n
tl
y
i
m
p
r
o
v
i
n
g
th
e
s
tate
o
f
th
e
ar
t
i
n
m
an
y
w
a
y
s
in
co
m
p
u
ter
v
i
s
io
n
ap
p
licati
o
n
s
.
T
h
e
i
m
p
o
r
tan
t
in
g
r
ed
ie
n
ts
f
o
r
th
e
s
u
cc
es
s
o
f
s
u
c
h
m
et
h
o
d
s
is
th
e
av
ailab
ilit
y
o
f
lar
g
e
q
u
a
n
titi
e
s
o
f
tr
ain
in
g
d
ata
.
Ho
w
e
v
er
,
in
t
h
e
w
o
r
ld
o
f
f
ac
e
r
ec
o
g
n
itio
n
,
lar
g
e
s
ca
le
p
u
b
l
ic
d
atasets
h
av
e
b
ee
n
lac
k
i
n
g
,
an
d
lar
g
el
y
d
u
e
to
th
is
f
ac
to
r
,
m
o
s
t
o
f
t
h
e
r
ec
en
t
ad
v
a
n
ce
s
in
t
h
e
co
m
m
u
n
i
t
y
r
e
m
ai
n
r
estricte
d
to
I
n
ter
n
et
g
ia
n
t
s
s
u
c
h
a
s
Face
b
o
o
k
an
d
Go
o
g
le.
Fo
r
e
x
a
m
p
le,
t
h
e
m
o
s
t
r
ec
en
t
f
ac
e
r
ec
o
g
n
itio
n
m
et
h
o
d
b
y
Go
o
g
le
w
as
tr
ain
ed
u
s
in
g
2
0
0
m
illi
o
n
i
m
a
g
e
s
an
d
eig
h
t
m
illi
o
n
u
n
iq
u
e
id
e
n
titi
e
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4752
I
n
d
o
n
esia
n
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l
.
1
2
,
No
.
2
,
No
v
e
m
b
er
201
8
:
4
7
6
–
4
8
1
478
T
h
e
cu
r
r
en
t
C
NN
m
o
d
els
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
ten
d
to
b
e
d
ee
p
er
an
d
lar
g
er
to
f
it
lar
g
e
a
m
o
u
n
t
o
f
t
h
e
d
ata
f
r
o
m
t
h
e
p
u
b
lic
r
eso
u
r
ce
s
s
u
c
h
as
th
e
I
n
ter
n
e
t.
A
cc
o
r
d
in
g
to
Xian
g
W
u
[
1
3
]
th
e
p
er
f
o
r
m
an
ce
o
f
C
NN
h
as
g
r
ea
tl
y
i
m
p
r
o
v
ed
,
f
o
r
ex
a
m
p
le,
th
e
ac
c
u
r
ac
y
o
n
t
h
e
ch
alle
n
g
i
n
g
L
FW
b
en
ch
m
ar
k
h
as
b
ee
n
i
m
p
r
o
v
ed
f
r
o
m
9
7
%
to
9
9
%
[
1
3
]
.
T
h
is
i
m
p
r
o
v
e
m
e
n
t
i
s
m
ain
l
y
d
u
e
to
th
e
f
ac
t
th
at
C
N
N
ca
n
lea
r
n
a
co
m
p
le
x
d
ata
d
is
tr
ib
u
tio
n
f
r
o
m
t
h
e
lar
g
e
-
s
ca
le
tr
ain
i
n
g
d
ataset.
Sev
er
al
r
ec
e
n
t
p
ap
er
s
h
a
v
e
a
ls
o
h
y
p
o
th
es
ized
th
at
C
N
N
d
ev
elo
p
an
u
n
d
er
s
tan
d
i
n
g
ab
o
u
t
o
b
j
ec
ts
b
ased
o
n
th
e
tr
ain
i
n
g
d
ata,
as
s
u
c
h
th
at
t
h
e
y
ar
e
ev
e
n
ab
le
to
g
en
er
ate
n
e
w
i
m
a
g
e
s
[
1
4
]
.
Ho
w
e
v
er
h
u
m
a
n
i
s
v
er
y
ca
p
ab
le
to
r
ec
o
g
n
ize
u
n
f
a
m
iliar
o
b
j
ec
ts
,
b
y
id
e
n
ti
f
y
i
n
g
th
eir
i
m
p
o
r
tan
t
f
ea
t
u
r
es,
m
ai
n
l
y
th
eir
s
h
ap
es.
T
h
ey
ca
n
also
id
e
n
ti
f
y
o
b
j
ec
ts
in
v
ar
io
u
s
f
o
r
m
s
s
u
c
h
as
d
i
f
f
er
en
t
s
ca
les,
o
r
ien
ta
tio
n
s
,
co
lo
u
r
s
o
r
b
r
ig
h
t
n
ess
.
T
h
er
ef
o
r
e,
it
r
em
ai
n
s
to
b
e
s
ee
n
h
o
w
C
NN
co
m
p
ar
e
to
h
u
m
a
n
s
i
n
ter
m
s
o
f
“se
m
a
n
t
ic
g
en
er
aliza
tio
n
”
.
Fig
u
r
e
2
ill
u
s
tr
a
tes
th
e
ar
ch
itectu
r
e
o
f
a
C
NN.
T
h
e
in
p
u
t
is
an
i
m
ag
e
u
s
ed
f
o
r
r
ec
o
g
n
itio
n
,
d
u
r
i
n
g
co
n
v
o
lu
tio
n
al
p
r
o
ce
s
s
,
th
e
o
u
tp
u
t
o
f
th
e
i
m
ag
e
b
ec
a
m
e
ac
t
iv
atio
n
m
ap
.
C
o
n
v
o
lu
tio
n
al
l
a
y
er
ac
ts
as
a
f
i
lter
to
w
ar
d
s
t
h
e
i
n
p
u
t
i
n
ter
m
s
o
f
s
izes,
p
ad
d
in
g
,
f
ea
t
u
r
es
a
n
d
etc.
P
o
o
lin
g
la
y
er
is
o
p
er
atin
g
as
a
r
ed
u
ce
r
f
o
r
n
u
m
b
er
o
f
p
ar
am
e
ter
s
.
B
o
th
la
y
er
ac
ted
as
f
ea
tu
r
es
e
x
tr
ac
ti
o
n
to
p
r
o
d
u
ce
a
g
en
er
ic
f
ea
tu
r
es.
A
t
th
e
e
n
d
,
th
e
o
u
tp
u
t
la
y
er
ac
t
as
f
u
ll
y
co
n
n
ec
ted
la
y
er
.
T
h
er
e
ar
e
a
f
e
w
l
a
y
er
s
t
h
at
lie
o
n
t
h
e
o
u
tp
u
t
la
y
er
s
u
c
h
as
o
u
tp
u
t
g
en
er
ato
r
la
y
er
f
o
r
g
en
er
ati
n
g
th
e
lo
s
s
w
h
ile
tr
ai
n
i
n
g
t
h
e
i
m
a
g
e
[
1
5
]
.
Fig
u
r
e
2
.
T
h
e
im
a
g
e
o
f
C
NN
a
r
ch
itect
u
r
e
[
1
5
]
3
.
3
.
Alex
Nex
A
le
x
Net
ac
h
ie
v
ed
th
e
to
p
5
er
r
o
r
s
f
r
o
m
2
6
%
to
1
5
.
3
%
in
I
m
ag
e
Net
L
ar
g
e
Scale
Vi
s
u
al
R
ec
o
g
n
itio
n
C
h
al
len
g
e
(
I
L
SVR
C
)
[
1
0
]
.
T
h
e
n
et
w
o
r
k
h
ad
s
i
m
ilar
ar
ch
i
tectu
r
e
as
L
eNe
t
b
y
Ya
n
n
L
e
C
u
n
et
al
b
u
t
w
a
s
d
ee
p
er
[
8
]
.
I
t
also
h
as
m
o
r
e
f
i
lter
s
p
er
la
y
er
w
i
th
s
tac
k
ed
co
n
v
o
l
u
tio
n
a
l
la
y
er
s
co
n
s
is
tin
g
o
f
1
1
x
1
1
,
5
x
5
,
3
x
3
,
co
n
v
o
lu
tio
n
s
,
m
a
x
p
o
o
lin
g
,
d
r
o
p
o
u
t,
d
ata
au
g
m
e
n
tat
io
n
,
R
eL
U
ac
tiv
at
io
n
s
,
SG
D
w
it
h
m
o
m
en
t
u
m
[
1
0
]
.
R
eL
U
ac
tiv
at
io
n
is
attac
h
ed
af
ter
ev
er
y
co
n
v
o
l
u
tio
n
al
a
n
d
f
u
ll
y
-
co
n
n
ec
ted
la
y
er
.
A
le
x
Net
w
as
tr
ai
n
ed
f
o
r
6
d
ay
s
s
i
m
u
lta
n
eo
u
s
l
y
o
n
t
w
o
Nv
id
ia
Gef
o
r
ce
GT
X
5
8
0
GPUs
w
h
ic
h
is
t
h
e
r
ea
s
o
n
f
o
r
wh
y
t
h
eir
n
et
w
o
r
k
is
s
p
lit
i
n
to
t
w
o
p
ip
elin
e
s
[
1
6
].
A
le
x
Net
m
e
th
o
d
is
d
esig
n
ed
b
y
th
e
S
u
p
er
Vis
io
n
g
r
o
u
p
,
wh
ich
is
co
n
s
i
s
ti
n
g
o
f
Geo
f
f
r
e
y
Hi
n
to
n
,
A
le
x
Kr
izh
e
v
s
k
y
,
a
n
d
I
l
y
aS
u
ts
k
e
v
er
[
1
0
].
Fig
u
r
e
3
illu
s
tr
ates t
h
e
ar
ch
ite
ctu
r
e
o
f
A
le
x
Net.
Fig
u
r
e
3
.
T
h
e
im
a
g
e
o
f
A
le
x
N
et
ar
ch
itectu
r
e
[
1
0
]
3
.
4
.
G
o
o
g
L
eNe
t
Go
o
g
L
eNe
t
ac
h
ie
v
ed
th
e
to
p
-
5
er
r
o
r
r
ate
o
f
6
.
6
7
%
ac
co
r
d
i
n
g
to
“
Dee
p
s
p
ar
s
e
r
ec
tif
ier
n
et
w
o
r
k
s
I
n
P
r
o
ce
e
d
in
g
s
o
f
th
e
1
4
th
I
n
ter
n
atio
n
al
C
o
n
f
er
en
ce
o
n
A
r
ti
f
icial
I
n
te
lli
g
en
ce
an
d
Stati
s
ti
cs
”
[
1
7
]
.
T
h
is
w
as
m
u
c
h
c
lo
s
er
an
d
al
m
o
s
t
s
i
m
il
ar
to
h
u
m
a
n
le
v
el
p
er
f
o
r
m
an
c
e.
As
it
tu
r
n
s
o
u
t,
t
h
is
w
as
ac
t
u
all
y
r
ath
er
h
ar
d
to
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esia
n
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4752
C
eleb
r
ity
F
a
ce
R
ec
o
g
n
itio
n
u
s
in
g
Dee
p
Lea
r
n
in
g
(
N
u
r
A
teq
a
h
B
in
ti Ma
t Ka
s
im
)
479
d
o
an
d
r
eq
u
ir
ed
s
o
m
e
h
u
m
a
n
tr
ain
in
g
i
n
o
r
d
er
to
b
ea
t
Go
o
g
L
eNe
ts
ac
c
u
r
ac
y
.
Af
ter
a
f
e
w
d
a
y
s
o
f
tr
ai
n
i
n
g
,
th
e
h
u
m
a
n
ex
p
er
t (
An
d
r
ej
K
ar
p
ath
y
)
w
a
s
ab
le
to
ac
h
ie
v
e
a
t
o
p
-
5
er
r
o
r
r
ate
o
f
5
.
1
% f
o
r
s
in
g
le
m
o
d
el
a
n
d
3
.
6
%
f
o
r
en
s
e
m
b
le.
I
t
u
s
ed
b
atch
n
o
r
m
al
izatio
n
,
i
m
a
g
e
d
is
to
r
tio
n
s
an
d
R
M
Sp
r
o
p
.
T
h
is
m
o
d
u
le
is
b
ased
o
n
s
ev
er
al
v
er
y
s
m
all
co
n
v
o
l
u
tio
n
s
in
o
r
d
er
to
d
r
asti
ca
ll
y
r
ed
u
ce
t
h
e
n
u
m
b
er
o
f
p
ar
a
m
eter
s
.
Go
o
g
L
eNe
t
’
s
ar
ch
i
tectu
r
e
co
n
s
is
t
o
f
2
2
la
y
er
b
u
t
r
ed
u
ce
d
th
e
n
u
m
b
er
o
f
p
ar
a
m
e
ter
s
f
r
o
m
6
0
m
illi
o
n
(
A
le
x
Net)
to
4
m
illi
o
n
.
Fig
u
r
e
4
ill
u
s
tr
ates
th
e
ar
c
h
i
tectu
r
e
o
f
Go
o
g
L
e
Net
[
1
7
]
.
I
n
Go
o
g
L
eNe
t,
t
h
er
e
ar
e
9
in
ce
p
tio
n
m
o
d
u
le
s
o
cc
u
r
ed
f
o
r
co
n
s
id
e
r
in
g
th
e
clu
s
ter
i
n
g
an
d
n
et
w
o
r
k
w
it
h
i
n
th
e
n
e
t
w
o
r
k
.
Du
r
i
n
g
t
h
e
i
n
ce
p
tio
n
m
o
d
u
les,
th
e
m
o
d
u
le
r
an
g
e
b
ei
n
g
ca
lc
u
lated
an
d
r
e
m
o
v
in
g
th
e
f
u
ll
y
co
n
n
ec
ted
la
y
er
s
.
Me
an
w
h
ile,
p
o
o
lin
g
in
t
h
e
in
ce
p
tio
n
m
o
d
u
le
s
r
ed
u
ce
s
t
h
e
n
u
m
b
er
s
o
f
p
ar
a
m
eter
s
i
n
v
o
lv
ed
.
B
esid
es
t
h
at,
s
h
ad
o
w
n
e
t
w
o
r
k
an
d
a
u
x
iliar
y
class
i
f
ier
ar
e
ad
d
ed
to
p
r
o
v
id
e
b
etter
o
u
tp
u
ts
.
F
u
r
th
er
m
o
r
e,
Go
o
g
L
eNe
t
h
as
m
o
r
e
la
y
er
s
d
u
e
to
th
e
9
in
ce
p
tio
n
m
o
d
u
les t
h
at
r
ep
ea
tin
g
t
h
e
co
n
v
o
lu
tio
n
al,
p
o
o
lin
g
,
s
o
f
t
m
ax
a
n
d
co
n
ca
t p
r
o
ce
s
s
es [
1
7
]
,
[
1
8
]
.
Fig
u
r
e
4
.
I
m
a
g
e
o
f
Go
o
g
L
eNe
t a
r
ch
itect
u
r
e
[
1
7
]
4.
RE
SU
L
T
S AN
D
AN
AL
Y
SI
S
Ma
tlab
R
2
0
1
8
b
is
u
s
e
as
th
e
to
o
l
to
tr
ain
an
d
test
th
e
d
ataset
f
o
r
th
is
p
ap
er
.
T
h
e
s
ize
o
f
th
e
i
m
ag
e
i
s
d
ef
in
e
as 1
0
0
x
1
0
0
x
3
w
h
ic
h
m
ea
n
s
th
at
t
h
e
s
ize
o
f
t
h
e
i
m
a
g
e
is
1
0
0
x
1
0
0
p
ix
els
an
d
th
e
v
a
lu
e
3
i
n
d
icate
s
t
h
at
th
e
tr
ain
in
g
i
m
ag
e
i
s
a
co
lo
r
i
m
a
g
e.
T
h
e
f
ir
s
t
co
n
v
o
lu
t
io
n
la
y
er
o
f
C
N
N
ex
tr
ac
t
s
th
e
e
d
g
es
o
f
t
h
e
i
m
ag
e
p
r
esen
ted
.
Fig
u
r
e
5
s
h
o
w
s
th
e
r
esu
lts
o
f
t
h
e
ce
leb
r
it
y
f
ac
e
r
ec
o
g
n
itio
n
w
it
h
v
alid
atio
n
ac
cu
r
ac
y
o
f
9
9
.
7
2
%.
T
h
e
elap
s
ed
ti
m
e
i
s
4
8
s
ec
o
n
d
s
to
co
m
p
lete
t
h
e
p
r
o
ce
s
s
w
it
h
m
ax
i
m
u
m
2
3
2
iter
atio
n
s
a
n
d
4
ep
o
ch
s
.
Fu
r
t
h
er
m
o
r
e,
th
e
av
er
a
g
e
f
o
r
iter
atio
n
p
er
ep
o
ch
is
5
8
.
Me
an
w
h
ile,
th
e
iter
atio
n
f
r
eq
u
e
n
c
y
is
3
0
iter
atio
n
s
an
d
p
atien
ce
is
5
an
d
th
e
lear
n
i
n
g
r
ate
0
.
0
1
w
it
h
th
e
s
ch
ed
u
le
o
f
l
ea
r
n
in
g
r
ate
is
co
n
s
tan
t.
Fig
u
r
e
5
.
A
cc
u
r
ac
y
o
f
tr
ai
n
i
n
g
4
.
1
.
Alex
Net
A
le
x
Net
is
o
n
e
o
f
t
h
e
p
r
e
-
tr
ain
ed
C
NN
m
o
d
els.
T
h
e
s
ca
le
o
f
th
e
s
ize
an
d
c
o
lo
r
is
d
ef
in
ed
as
2
2
7
x
2
2
7
x
3
.
T
h
e
ac
cu
r
ac
y
h
a
s
ac
h
i
ev
ed
1
0
0
%.
Fig
u
r
e
5
s
h
o
w
s
t
h
e
r
esu
lt
f
o
r
A
le
x
Net.
T
h
e
v
a
lid
atio
n
ac
cu
r
ac
y
is
1
0
0
%
w
h
ich
m
ea
n
s
t
h
at
i
t
h
a
s
ac
h
ie
v
ed
is
v
er
y
ac
c
u
r
ate.
T
h
e
elap
s
ed
ti
m
e
i
s
9
m
in
8
s
ec
o
n
d
s
f
o
r
th
e
p
r
o
ce
s
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
5
0
2
-
4752
I
n
d
o
n
esia
n
J
E
lec
E
n
g
&
C
o
m
p
Sci,
Vo
l
.
1
2
,
No
.
2
,
No
v
e
m
b
er
201
8
:
4
7
6
–
4
8
1
480
to
co
m
p
lete
an
d
ac
h
ie
v
e
s
u
c
h
ac
cu
r
ac
y
w
it
h
6
ep
o
ch
s
an
d
1
0
2
iter
atio
n
s
.
Hen
ce
,
th
e
iter
atio
n
s
p
er
ep
o
ch
ar
e
1
7
.
T
h
e
lear
n
in
g
r
ate
s
c
h
ed
u
l
e
is
co
n
s
ta
n
t
w
i
th
0
.
0
0
0
1
.
T
h
e
tr
ain
i
n
g
p
r
o
ce
s
s
ac
h
ie
v
es
1
0
0
%
ac
cu
r
ac
y
s
i
n
ce
f
ir
s
t e
p
o
ch
.
Fig
u
r
e
6
.
Th
e
r
esu
lt p
er
f
o
r
m
a
n
ce
o
f
A
le
x
Net
4
.
2
.
G
o
o
g
L
eNe
t
Go
o
g
L
eNe
t
i
s
an
o
th
er
p
o
p
u
lar
p
r
e
-
tr
ain
ed
C
NN
m
o
d
el
th
at
h
as
b
ee
n
r
ep
o
r
ted
to
p
r
o
d
u
ce
v
er
y
h
i
g
h
ac
cu
r
ac
y
[
8
]
.
Size
i
m
a
g
e
is
d
e
f
i
n
e
at
2
2
7
x
2
2
7
x
3
.
Fo
r
th
is
ex
p
er
im
e
n
t,
Go
o
g
L
eNe
t r
ec
ei
v
e
s
th
e
s
a
m
e
ac
c
u
r
ac
y
as
A
le
x
Net
w
h
ic
h
is
1
0
0
%.
Fig
u
r
e
6
s
h
o
w
s
t
h
e
r
esu
lt
s
o
f
Go
o
g
L
eNe
t.
T
h
e
v
alid
atio
n
ac
cu
r
ac
y
is
1
0
0
%
eq
u
iv
ale
n
t
to
1
af
ter
it
r
ea
c
h
ed
th
e
f
i
n
al
i
ter
at
io
n
.
T
h
e
co
m
p
letio
n
ti
m
e
i
s
1
4
m
in
u
te
s
4
7
s
ec
o
r
d
s
w
i
th
6
ep
o
ch
s
an
d
m
a
x
i
m
u
m
iter
atio
n
i
s
1
0
2
.
Hen
ce
,
t
h
e
iter
atio
n
s
p
er
ep
o
ch
ar
e
1
7
.
A
s
s
h
o
w
n
o
n
t
h
e
g
r
ap
h
,
t
h
e
lear
n
in
g
r
ate
s
c
h
ed
u
le
i
s
co
n
s
t
an
t a
t 0
.
0
0
0
1
.
Fig
u
r
e
7
.
R
esu
lt p
er
f
o
r
m
a
n
ce
p
r
o
d
u
ce
d
b
y
Go
o
g
L
eNe
t
5.
CO
NCLU
SI
O
N
T
ab
le
1
s
h
o
w
s
th
e
p
r
o
m
i
s
i
n
g
r
esu
lts
p
r
o
d
u
ce
d
b
y
C
NN,
A
le
x
Net
a
n
d
Go
o
g
L
eNe
t
d
esp
ite
th
e
d
if
f
er
e
n
ce
s
in
g
e
n
d
er
,
f
ac
e
e
x
p
r
ess
io
n
s
,
h
air
s
t
y
le,
f
ea
t
u
r
e
s
,
an
d
b
ac
k
g
r
o
u
n
d
o
f
th
e
i
m
ag
es
i
n
th
e
d
ata
s
et.
B
ased
o
n
th
e
r
esu
lts
d
is
p
la
y
e
d
in
T
ab
le
1
,
w
e
ca
n
s
ee
th
a
t
A
le
x
Net
an
d
Go
o
g
L
e
Net
h
av
e
b
etter
ac
cu
r
ac
y
w
h
ic
h
is
1
0
0
%
co
m
p
ar
ed
to
C
NN
w
h
ic
h
is
9
9
.
7
2
%.
I
t
s
h
o
w
s
t
h
at
t
h
e
m
ac
h
i
n
e
h
as
p
er
f
ec
tl
y
r
ec
o
g
n
ized
all
ce
leb
r
it
y
i
m
a
g
es i
n
t
h
e
d
atase
t
u
s
i
n
g
Alex
Ne
t a
n
d
Go
o
g
L
eN
et.
T
h
is
is
d
u
e
to
th
e
tr
ain
in
g
o
f
m
illi
o
n
s
o
f
d
ata.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
d
o
n
esia
n
J
E
lec
E
n
g
&
C
o
m
p
Sci
I
SS
N:
2502
-
4752
C
eleb
r
ity
F
a
ce
R
ec
o
g
n
itio
n
u
s
in
g
Dee
p
Lea
r
n
in
g
(
N
u
r
A
teq
a
h
B
in
ti Ma
t Ka
s
im
)
481
Me
an
w
h
ile,
t
h
e
s
p
ee
d
o
f
p
r
o
ce
s
s
i
n
g
C
NN
r
ec
o
r
d
ed
th
e
f
a
s
test
to
co
m
p
lete
tr
ai
n
in
g
co
m
p
ar
ed
to
A
le
x
Net
an
d
Go
o
g
L
eNe
t.
T
h
i
s
is
d
u
e
to
t
h
e
n
u
m
b
er
o
f
la
y
er
s
in
t
h
e
s
e
m
o
d
els.
T
h
e
m
o
r
e
n
u
m
b
er
o
f
la
y
er
s
,
th
e
m
o
r
e
ti
m
e
it
ta
k
es
to
p
r
o
d
u
ce
th
e
r
e
s
u
lt
s
.
C
NN
co
m
p
l
etes
th
e
ex
ec
u
tio
n
o
r
co
n
v
er
g
es
af
ter
4
8
s
ec
o
n
d
s
w
h
ile
A
le
x
Net
ac
h
iev
ed
1
0
0
%
ac
cu
r
ac
y
i
n
9
m
i
n
u
tes
a
n
d
8
s
ec
o
n
d
s
.
Go
o
g
L
eNe
t
r
eq
u
ir
es
1
4
m
i
n
u
te
s
an
d
4
7
s
ec
o
n
d
s
to
co
n
v
e
r
g
e
o
r
co
m
p
l
ete
th
e
ex
ec
u
tio
n
.
I
n
s
elec
ti
n
g
w
h
ic
h
m
o
d
u
le
to
u
s
e,
v
ar
io
u
s
f
ac
to
r
s
n
ee
d
to
b
e
co
n
s
id
er
ed
s
u
ch
as
t
h
e
av
ai
l
ab
ilit
y
o
f
lar
g
e
a
m
o
u
n
t
o
f
tr
ain
i
n
g
d
ata,
th
e
a
m
o
u
n
t
o
f
ti
m
e
t
h
at
ca
n
b
e
s
p
ar
ed
f
o
r
th
e
tr
ain
in
g
p
r
o
ce
s
s
an
d
th
e
n
u
m
b
e
r
o
f
er
r
o
r
s
th
at
ca
n
b
e
ac
ce
p
ted
.
Fu
tu
r
e
r
esear
c
h
i
n
cl
u
d
es
th
e
i
m
p
r
o
v
e
m
en
t
o
f
t
h
e
C
NN
ar
ch
itect
u
r
e
a
n
d
ex
p
er
i
m
e
n
t o
n
o
th
er
p
r
e
-
tr
ai
n
e
d
C
NN
m
o
d
els.
T
ab
le
1
.
C
o
m
p
ar
is
o
n
s
b
et
w
ee
n
C
NN,
A
le
x
Net
Go
o
g
L
eNe
t
Pa
r
a
m
e
t
e
r
s
C
N
N
S
c
r
a
t
c
h
A
l
e
x
N
e
t
Go
o
g
L
e
N
e
t
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
9
9
.
7
2
%
1
0
0
.
0
0
%
1
0
0
.
0
0
%
El
a
p
se
d
T
i
me
4
8
se
c
o
r
d
s
9
m
i
n
u
t
e
s
8
se
c
o
n
d
s
1
4
mi
n
u
t
e
s
4
7
se
c
o
n
d
s
N
u
mb
e
r
o
f
E
p
o
c
h
4
6
6
N
u
mb
e
r
o
f
I
t
e
r
a
t
i
o
n
2
3
2
1
0
2
1
0
2
V
a
l
i
d
a
t
i
o
n
F
r
e
q
u
e
n
c
y
30
3
3
ACK
NO
WL
E
D
G
E
M
E
NT
S
T
h
e
au
th
o
r
s
w
o
u
ld
lik
e
to
th
an
k
Facu
l
t
y
o
f
C
o
m
p
u
ter
an
d
Ma
t
h
e
m
at
ical
Scien
ce
s
,
Un
i
v
er
s
iti T
ek
n
o
lo
g
i M
A
R
A
,
Sh
a
h
A
la
m
,
Selan
g
o
r
,
f
o
r
s
p
o
n
s
o
r
in
g
t
h
is
r
esear
c
h
.
RE
F
E
R
E
NC
E
S
[
1
]
J.
Ha
sh
e
m
i,
Q.
Qiu
a
n
d
G
.
S
a
p
iro
,
“
In
telli
g
e
n
t
S
y
n
th
e
sis
Driv
e
n
M
o
d
e
l
Ca
li
b
ra
ti
o
n
:
F
ra
m
e
w
o
rk
a
n
d
F
a
c
e
Re
c
o
g
n
it
io
n
A
p
p
li
c
a
ti
o
n
”
,
In
ter
n
a
ti
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
ter Visio
n
(ICC
V
)
2
0
1
7
.
[
2
]
R.
De
sa
i
a
n
d
B.
S
o
n
a
w
a
n
e
,
“
Gist,
HO
G
,
a
n
d
DW
T
-
b
a
se
d
Co
n
ten
t
-
b
a
se
d
Im
a
g
e
Re
tri
e
v
a
l
f
o
r
F
a
c
ial
Im
a
g
e
s”
,
In
tern
a
ti
o
n
a
l
.
[
3
]
X
.
Qin
,
Y.
Zh
o
u
,
Z.
He
,
Y.
W
a
n
g
a
n
d
Z.
T
a
n
g
,
“
A
F
a
ste
r
R
-
CNN
b
a
se
d
M
e
th
o
d
f
o
r
Co
m
ic
Ch
a
ra
c
ters
F
a
c
e
De
tec
ti
o
n
”
,
1
4
th
IA
P
R
In
tern
a
ti
o
n
a
l
Co
n
f
e
e
n
c
e
o
n
Do
c
u
m
e
n
t
A
n
a
l
y
sis a
n
d
Re
c
o
g
n
it
i
o
n
(ICDA
R)
2
0
1
7
.
[
4
]
W
.
On
g
V
u
i
Jiu
n
n
,
N.
S
a
b
ri
a
n
d
Z.
Ib
ra
h
im
,
“
I
m
a
g
e
-
b
a
s
e
d
Hu
m
a
n
F
a
ll
Re
c
o
g
n
it
io
n
u
sin
g
G
a
u
ss
ian
M
ix
tu
re
M
o
d
e
l
a
n
d
S
u
p
p
o
rt
V
e
c
t
o
r
M
a
c
h
i
n
e
”
,
In
t
e
rn
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
Co
n
tro
l
T
h
e
o
ry
a
n
d
A
p
p
li
c
a
ti
o
n
s,
v
o
l.
9
,
n
u
m
b
e
r
4
4
,
2
0
1
6
.
[
5
]
Z.
Ib
ra
h
im
,
N.
S
a
b
ri
a
n
d
N.
N
.
M
o
h
d
M
a
n
g
h
o
r,
“
L
e
a
f
R
e
c
o
g
n
it
io
n
Us
in
g
T
e
x
tu
re
F
e
a
tu
re
s
fo
r
He
rb
a
l
P
lan
t
Id
e
n
ti
f
ica
ti
o
n
’,
I
n
tern
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
El
e
c
tri
c
a
l
En
g
in
e
e
rin
g
a
n
d
Co
m
p
u
ter
S
c
ien
c
e
(IJEECS
),
V
o
l
.
9
,
N
o
.
1
2
0
1
8
,
p
p
.
1
5
2
-
1
5
6
.
[
6
]
N.
S
a
b
ri
a
n
d
Z.
Ib
ra
h
im
,
“
P
a
lm
Oil
F
re
sh
F
ru
i
t
Bu
n
c
h
Rip
e
n
e
s
s
G
ra
d
in
g
Id
e
n
ti
f
ica
ti
o
n
u
sin
g
Co
lo
r
F
e
a
tu
re
s”
,
Jo
u
rn
a
l
o
f
F
u
n
d
a
m
e
n
tal
a
n
d
A
p
p
l
ied
S
c
ien
c
e
,
2
0
1
7
,
9
(4
S
),
p
p
.
5
6
3
-
5
7
9
.
[
7
]
Ha
d
a
d
Y (2
0
1
8
)
Am
a
z
in
g
A
p
p
li
c
a
ti
o
n
o
f
De
e
p
L
e
a
rn
in
g
[
8
]
A
.
Ko
rt
y
le
ws
k
i,
B.
E
g
g
e
r
a
n
d
A.
S
c
h
n
e
id
e
r,
“
Em
p
iri
c
a
ll
y
A
n
a
l
y
z
in
g
th
e
Eff
e
c
t
o
f
Da
tas
e
t
Bias
e
s
o
n
De
e
p
F
a
c
e
Re
c
o
g
n
it
io
n
S
y
ste
m
s
”
,
Co
m
p
u
ter V
isi
o
n
a
n
d
P
a
tt
e
rn
Re
c
o
g
n
it
io
n
(
CVP
R)
2
0
1
8
.
[
9
]
D.
V
a
rg
a
,
C.
A
.
S
z
a
b
o
a
n
d
T
.
S
z
i
ra
n
y
i,
“
A
u
to
m
a
ti
c
Ca
rto
o
n
Co
l
o
ri
z
a
ti
o
n
b
a
se
d
o
n
Co
n
v
o
lu
ti
o
n
a
l
N
e
u
ra
l
Ne
tw
o
rk
”
,
15
th
In
tern
a
ti
o
n
a
l
W
o
rk
h
o
p
o
n
C
o
n
ten
t
-
Ba
se
d
M
u
lt
im
e
d
ia In
d
e
x
in
g
,
Ju
n
e
2
0
1
7
,
p
p
.
1
9
-
21.
[
1
0
]
Bro
w
n
lee
,
J.
(2
0
1
6
,
Ju
ly
2
9
).
8
I
n
sp
iratio
n
a
l
A
p
p
li
c
a
ti
o
n
s o
f
De
e
p
L
e
a
rn
in
g
.
[
1
1
]
A
.
Kriz
h
e
v
sk
y
,
I.
S
u
tsk
e
v
e
r
a
n
d
G
.
E.
Hin
to
n
,
“
Im
a
g
e
N
e
t
Cla
ss
if
ic
a
ti
o
n
w
it
h
De
e
p
Co
n
v
o
lu
ti
o
n
a
l
Ne
tw
o
rk
s”
,
A
d
v
a
n
c
e
s in
Ne
u
ra
l
In
f
o
rm
a
ti
o
n
P
r
o
c
e
ss
in
g
S
y
ste
m
s 2
5
,
2
0
1
2
.
[
1
2
]
S
.
Ya
n
g
,
P
.
L
u
o
,
C.
C
.
L
o
y
,
a
n
d
X
.
T
a
n
g
,
"
F
ro
m
F
a
c
ial
P
a
rts
Re
sp
o
n
se
s
to
F
a
c
e
De
tec
ti
o
n
:
A
De
e
p
L
e
a
rn
in
g
A
p
p
ro
a
c
h
"
,
in
IEE
E
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
V
isio
n
(
ICCV)
,
2
0
1
5
.
[
1
3
]
W
u
X
,
He
R,
S
u
n
Z,
T
a
n
T
,
“
A
Li
g
h
t
CNN
f
o
r
De
e
p
F
a
c
e
Re
p
re
se
n
tatio
n
w
it
h
No
isy
L
a
b
e
ls”
,
IEE
E
T
ra
n
sa
c
ti
o
n
s o
n
In
f
o
rm
a
ti
o
n
F
o
re
n
sic
s an
d
S
e
c
u
rit
y
(2
0
1
8
).
[
1
4
]
Ho
ss
e
in
i
H,
X
iao
B,
Ja
isw
a
l
M
,
P
o
o
v
e
n
d
ra
n
R
On
t
h
e
L
i
m
it
a
ti
o
n
o
f
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
s
in
Re
c
o
g
n
izin
g
Ne
g
a
ti
v
e
I
m
a
g
e
.
[
1
5
]
G
u
p
ta,
D.,
Ja
in
,
K.,
Ja
in
,
A
.
,
&
A
n
a
l
y
ti
c
s
V
id
h
y
a
Co
n
ten
t
T
e
a
m
.
(2
0
1
7
,
Ju
n
e
2
9
).
A
rc
h
it
e
c
tu
re
o
f
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
s (CN
Ns
)
d
e
m
y
st
if
ied
.
[
1
6
]
G
a
o
H (2
0
1
7
)
A
W
a
l
k
-
th
ro
u
g
h
A
lex
Ne
t
.
[
1
7
]
G
lo
ro
t,
X.,
Bo
r
d
e
s,
A
.
,
&
Be
n
g
io
,
Y.
(2
0
1
1
).
De
e
p
s
p
a
rse
re
c
ti
f
i
e
r
n
e
tw
o
rk
s
In
P
ro
c
e
e
d
i
n
g
s
o
f
th
e
1
4
th
I
n
tern
a
ti
o
n
a
l
Co
n
f
e
re
n
c
e
o
n
A
rti
f
icia
l
In
telli
g
e
n
c
e
a
n
d
S
tatisti
c
s.
JML
R
W
&
CP
V
o
l
u
m
e
(V
o
l.
1
5
,
p
p
.
3
1
5
-
3
2
3
).
[
1
8
]
A
ro
ra
,
S
.
,
Bh
a
sk
a
ra
,
A
.
,
G
e
,
R.
,
&
M
a
,
T
.
P
r
o
v
a
b
le b
o
u
n
d
s f
o
r
lea
rn
in
g
so
m
e
d
e
e
p
re
p
re
se
n
tatio
n
s
.
ICM
L
2
0
1
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.