I
AE
S
I
nte
rna
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
9
,
No
.
3
,
Sep
tem
b
er
2020
,
p
p
.
424
~
428
I
SS
N:
2252
-
8938
,
DOI
: 1
0
.
1
1
5
9
1
/i
j
ai.
v
9
.i
3
.
p
p
424
-
4
2
8
424
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Ag
e
-
ba
sed facia
l
recog
nition us
ing
co
nv
o
luted
neura
l net
w
o
rk
deep learning
alg
o
rith
m
J
uli
us
Yo
ng
Wu J
ien
1
,
A
s
lin
a
B
a
ha
ru
m
2
,
Sh
a
liza
H
a
y
a
t
i
A.
Wa
ha
b
3
,
No
rdin Sa
a
d
4
,
M
uh
a
mm
a
d
O
m
a
r
5
,
No
o
rsid
i A
iz
ud
d
in
M
a
t
No
o
r
6
1
,
2,
3,
4
Us
e
r
Ex
p
e
rien
c
e
Re
se
a
rc
h
L
a
b
(UX
RL
),
F
a
c
u
lt
y
o
f
Co
m
p
u
ti
n
g
a
n
d
In
f
o
rm
a
ti
c
s,
Un
iv
e
rsiti
M
a
la
y
sia
S
a
b
a
h
,
8
8
4
0
0
Ko
ta Kin
a
b
a
l
u
,
S
a
b
a
h
,
M
a
lay
sia
5
F
a
c
u
lt
y
o
f
Bu
sin
e
ss
M
a
n
a
g
e
m
e
n
t,
Un
iv
e
rsiti
T
e
k
n
o
lo
g
i
M
A
RA
,
S
a
ra
w
a
k
Ca
m
p
u
s,
M
a
lay
si
a
6
UT
M
CRES
,
F
a
c
u
lt
y
o
f
Bu
il
t
En
v
iro
n
m
e
n
t
a
n
d
S
u
rv
e
y
in
g
,
Un
iv
e
rsiti
T
e
k
n
o
lo
g
i
M
a
lay
sia
,
Jo
h
o
r
Ba
h
ru
,
Jo
h
o
r,
M
a
lay
si
a
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Feb
5
,
2
0
2
0
R
ev
i
s
ed
A
p
r
1
2
,
2
0
2
0
A
cc
ep
ted
A
p
r
2
7
,
2
0
2
0
F
a
c
e
re
c
o
g
n
it
io
n
is
th
e
u
se
o
f
b
io
m
e
tri
c
in
n
o
v
a
ti
o
n
s
t
h
a
t
c
a
n
se
e
o
r
v
a
li
d
a
te
a
p
e
rso
n
b
y
se
e
in
g
a
n
d
in
v
e
stig
a
ti
n
g
d
e
sig
n
s
d
e
p
e
n
d
in
g
o
n
th
e
sh
a
p
e
o
f
th
e
in
d
iv
id
u
a
l.
F
a
c
e
re
c
o
g
n
it
io
n
is
u
se
d
larg
e
l
y
f
o
r
th
e
p
u
rp
o
se
o
f
w
e
ll
-
b
e
in
g
,
d
e
sp
it
e
th
e
f
a
c
t
th
a
t
p
a
ss
io
n
f
o
r
d
if
fe
re
n
t
a
re
a
s
o
f
u
se
is
g
ro
w
in
g
.
Ov
e
ra
ll
,
f
a
c
e
re
c
o
g
n
it
io
n
i
n
n
o
v
a
ti
o
n
s
a
re
w
o
rth
c
o
n
sid
e
rin
g
b
e
c
a
u
se
th
e
y
h
a
v
e
th
e
p
o
ten
ti
a
l
f
o
r
b
r
o
a
d
leg
a
l
ju
risd
ict
io
n
a
n
d
d
if
f
e
re
n
t
b
u
sin
e
ss
a
p
p
li
c
a
ti
o
n
s.
It
is
w
id
e
l
y
u
se
d
in
m
a
n
y
sp
a
c
e
s.
Ho
w
it
w
o
rk
s
is
a
p
ro
d
u
c
t
o
f
f
a
c
i
a
l
re
c
o
g
n
it
i
o
n
p
ro
c
e
ss
in
g
f
a
c
ial
g
e
o
m
e
tr
y
.
T
h
e
h
o
le
b
e
tw
e
e
n
th
e
e
a
r
a
n
d
th
e
g
o
o
d
w
a
y
f
ro
m
th
e
f
ro
n
t
to
th
e
jaw
a
re
th
e
m
a
in
v
a
riab
les
.
T
h
is
c
o
d
e
d
isti
n
g
u
ish
e
s
th
e
h
ig
h
li
g
h
t
o
f
th
e
f
a
c
e
th
a
t
is
im
p
o
rtan
t
f
o
r
y
o
u
r
fa
c
ial
se
p
a
ra
ti
o
n
a
n
d
c
re
a
tes
y
o
u
r
fa
c
ial
e
x
p
re
s
sio
n
.
T
h
e
re
f
o
re
,
th
is
stu
d
y
g
iv
e
s
a
n
o
v
e
rv
ie
w
o
f
a
g
e
d
e
tec
ti
o
n
u
sin
g
a
d
if
f
e
re
n
t
c
o
m
b
in
a
ti
o
n
o
f
m
a
c
h
in
e
lea
rn
in
g
a
n
d
im
a
g
e
p
ro
c
e
ss
in
g
m
e
th
o
d
s o
n
th
e
im
a
g
e
d
a
tas
e
t
.
K
ey
w
o
r
d
s
:
Ag
e
d
etec
tio
n
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
I
m
ag
e
p
r
o
ce
s
s
i
n
g
P
r
ed
ictio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Asl
i
n
a
B
ah
ar
u
m
,
UXR
L
,
Fac
u
lt
y
o
f
C
o
m
p
u
ti
n
g
an
d
I
n
f
o
r
m
a
tics
,
Un
i
v
er
s
iti Ma
la
y
s
ia
Sab
ah
,
8
8
4
0
0
Ko
ta
Kin
ab
alu
,
Sab
ah
,
Ma
la
y
s
ia.
E
m
ail:
as
lin
a
@
u
m
s
.
ed
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
Facial
r
ec
o
g
n
i
tio
n
(
F
R
)
is
u
ti
lized
f
o
r
th
e
m
o
s
t
p
ar
t
f
o
r
wellb
ein
g
p
u
r
p
o
s
es,
d
esp
ite
th
e
f
ac
t
t
h
at
en
th
u
s
ias
m
f
o
r
d
if
f
er
e
n
t
ter
r
it
o
r
ies
o
f
u
til
izatio
n
i
s
d
ev
elo
p
i
n
g
[
1
]
.
I
n
all
ac
t
u
alit
y
,
F
R
in
n
o
v
atio
n
h
a
s
g
o
tte
n
n
o
te
w
o
r
t
h
y
co
n
s
id
er
atio
n
as
it
h
as
th
e
p
o
t
en
tial
f
o
r
a
w
id
e
s
co
p
e
o
f
la
w
au
t
h
o
r
izatio
n
an
d
d
if
f
er
en
t
b
u
s
in
e
s
s
ap
p
licatio
n
s
[
2
]
.
I
t
h
as
b
ee
n
g
en
er
all
y
u
ti
lized
i
n
n
u
m
er
o
u
s
s
p
ac
es,
f
o
r
e
x
a
m
p
le,
A
T
M,
s
o
cial
i
n
s
u
r
a
n
ce
f
r
a
m
e
w
o
r
k
,
d
r
iv
i
n
g
p
er
m
i
t
f
r
am
e
w
o
r
k
,
tr
ai
n
r
e
s
er
v
atio
n
f
r
a
m
e
w
o
r
k
,
o
b
s
er
v
in
g
a
s
s
i
s
ta
n
ce
,
an
d
id
en
ti
f
icatio
n
v
er
if
ica
tio
n
.
Ho
w
it
f
u
n
ctio
n
s
is
th
e
p
r
o
d
u
ct
f
o
r
f
ac
ial
ac
k
n
o
w
led
g
m
e
n
t
p
er
u
s
es
t
h
e
f
ac
e's
g
eo
m
etr
y
.
T
h
e
h
o
le
b
et
w
ee
n
th
e
ea
r
s
an
d
t
h
e
g
o
o
d
w
a
y
s
f
r
o
m
t
h
e
f
r
o
n
t
to
th
e
j
a
w
ar
e
k
e
y
v
ar
iab
les
[
3
]
.
T
h
e
co
d
e
d
is
tin
g
u
is
h
e
s
f
ac
ial
h
i
g
h
li
g
h
t
s
th
at
ar
e
cr
u
cial
to
y
o
u
r
f
ac
e
s
ep
ar
atio
n
an
d
p
r
o
d
u
ce
y
o
u
r
f
ac
ial
m
ar
k
.
B
ec
au
s
e
o
f
d
ee
p
lear
n
in
g
tech
n
iq
u
e
s
,
t
h
er
e
h
a
v
e
b
ee
n
cr
itical
ad
v
a
n
ce
s
i
n
FR
[
4
]
.
I
n
th
e
ea
r
l
y
s
ta
g
es,
e
x
p
lo
r
in
g
t
h
e
i
n
ter
est,
f
o
r
th
e
m
o
s
t p
ar
t,
f
o
cu
s
ed
o
n
F
R
w
it
h
a
d
ee
p
s
y
s
te
m
o
f
s
i
g
n
i
f
ic
an
t li
g
h
t o
r
p
ictu
r
e
f
ac
es.
Step
h
en
[
5
]
g
i
v
en
a
b
r
ief
r
ev
ie
w
o
f
th
e
tech
n
iq
u
e
s
o
f
d
ee
p
lear
n
in
g
a
n
d
f
ac
e
-
to
-
f
ac
e
lear
n
in
g
a
n
d
co
m
p
ar
es
s
o
m
e
o
f
th
e
b
asic
n
eu
r
a
l
f
o
r
m
u
las
b
ased
o
n
co
m
m
o
n
c
o
n
v
o
l
u
tio
n
n
eu
r
al
n
et
w
o
r
k
s
(
C
NNs).
T
h
e
d
ee
p
n
et
w
o
r
k
s
u
s
ed
in
F
R
,
s
u
ch
a
s
d
ee
p
b
elief
n
et
w
o
r
k
(
DB
N)
,
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
e
t
w
o
r
k
(
C
NN,
o
r
C
o
n
v
Net)
,
a
u
to
en
co
d
er
(
A
E
)
,
an
d
o
th
er
s
ar
e
an
al
y
ze
d
f
o
r
ar
ch
itect
u
r
e
[
6
]
.
Ma
n
d
al
[
7
]
ass
ess
ed
a
s
i
g
n
i
f
ica
n
t
m
ea
s
u
r
e
o
f
p
r
o
f
o
u
n
d
lear
n
in
g
s
tr
ate
g
ie
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
A
g
e
-
b
a
s
ed
f
a
cia
l rec
o
g
n
itio
n
u
s
in
g
co
n
v
o
lu
ted
n
e
u
r
a
l
... (
Ju
l
iu
s
Yo
n
g
W
u
Jien
)
425
f
o
r
FR
.
Sep
as
-
Mo
g
h
ad
d
a
m
[
6
]
s
tu
d
i
ed
o
f
FR
ar
r
an
g
e
m
e
n
ts
d
ep
en
d
en
t
o
n
a
n
o
th
er
,
all
th
e
m
o
r
e
in
co
r
p
o
r
atin
g
an
d
m
o
r
e
ex
tr
av
a
g
an
t
s
ta
g
g
er
ed
s
cien
ti
f
ic
class
i
f
icat
io
n
.
L
e
ar
n
ed
-
Miller
[
8
]
lo
o
k
ed
at
a
v
ar
iet
y
o
f
s
u
r
p
r
is
i
n
g
in
v
e
n
ti
v
e
s
tr
ateg
ie
s
in
t
h
e
L
ab
eled
Face
s
in
t
h
e
W
ild
(
L
FW
)
d
atab
ase.
2.
RE
S
E
ARCH
M
E
T
H
O
D
Fig
u
r
e
1
s
h
o
w
s
t
h
e
f
lo
w
o
f
m
e
th
o
d
o
lo
g
y
.
T
h
is
f
lo
w
co
n
s
i
s
ts
o
f
p
h
ase
1
u
n
til p
h
a
s
e
3
as b
elo
w
:
P
h
ase
1
In
t
h
is
p
h
ase,
d
ata
is
ac
q
u
ir
ed
.
In
t
h
e
d
ata
p
r
ep
ar
atio
n
s
ta
g
e
,
s
a
m
p
le
s
of
f
ac
ial
i
m
a
g
e
s
ar
e
ac
q
u
ir
ed
an
d
th
en
u
n
d
er
g
o
p
r
e
-
p
r
o
ce
s
s
in
g
to
en
h
an
ce
t
h
e
q
u
al
it
y
of
i
m
a
g
es.
P
h
ase
2
In
th
is
p
h
ase,
s
e
g
m
e
n
ted
i
m
a
g
es
is
t
h
e
n
u
n
d
er
g
o
n
e
f
ea
tu
r
e
ex
tr
ac
tio
n
p
r
o
ce
s
s
h
er
e.
T
h
ese
ex
tr
ac
ted
f
ea
t
u
r
es
ar
e
th
e
n
u
s
ed
in
t
h
e
tr
ain
i
n
g
p
r
o
ce
s
s
.
P
h
ase
3
T
h
e
f
in
al
p
h
ase
is
th
e
p
r
ed
ictio
n
an
d
ev
al
u
atio
n
s
tag
e
w
h
er
e
each
b
u
ilt
m
o
d
el
is
u
s
ed
to
p
r
ed
ict
th
e
in
p
u
t
i
m
a
g
e.
T
h
e
ac
cu
r
ac
y
of
each
m
o
d
el
w
ill
be
ca
lcu
lated
an
d
ev
al
u
ated
.
Fig
u
r
e
1.
Flo
w
of
m
et
h
o
d
o
lo
g
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
9
,
No
.
3
,
Sep
te
m
b
er
20
20
:
4
2
4
–
428
426
3.
RE
SU
L
T
S
AND
D
I
SCU
SS
I
O
N
3
.
1
.
Da
t
a
a
cquis
it
io
n
T
h
r
ee
d
atab
ases
w
er
e
u
tili
ze
d
in
t
h
e
e
x
a
m
in
at
io
n
:
t
h
e
B
E
R
C
d
atab
ase,
th
e
P
AL
m
at
u
r
in
g
d
atab
ase,
an
d
th
e
FG
-
Net
m
at
u
r
i
n
g
d
atab
ase
[
9
]
.
T
h
e
B
E
R
C
d
atab
ase
co
n
tai
n
s
t
h
e
f
ac
e
p
ict
u
r
es
o
f
3
9
0
p
eo
p
le
th
e
ag
e
ex
ten
d
3
to
8
1
y
ea
r
s
o
f
a
g
e.
T
h
e
f
ac
ial
p
ic
tu
r
es
w
er
e
g
o
tten
u
tili
zi
n
g
a
co
m
p
u
ter
ized
ca
m
er
a
at
a
h
i
g
h
r
eso
lu
tio
n
o
f
1
6
0
0
×1
2
0
0
p
ix
els.
T
h
e
FG
-
NE
T
d
ev
elo
p
in
g
d
atab
ase
i
s
u
s
ed
in
t
h
i
s
i
n
v
e
s
ti
g
atio
n
,
w
h
ic
h
co
n
tain
s
1
0
0
2
f
ac
e
p
ict
u
r
es
f
r
o
m
8
2
o
n
e
o
f
a
k
in
d
s
u
b
j
ec
ts
,
an
d
ea
c
h
s
u
b
j
ec
t
h
a
s
6
–
1
8
f
a
ce
p
i
ctu
r
es
n
a
m
ed
w
it
h
g
r
o
u
n
d
tr
u
t
h
ag
es.
T
h
e
ag
es
ar
e
f
lo
w
ed
in
a
w
id
e
r
an
g
e
f
r
o
m
0
to
6
9
.
T
h
e
ag
e
tr
an
s
p
o
r
t
in
eith
er
th
e
n
u
m
b
er
o
f
p
ictu
r
es o
r
th
e
a
m
o
u
n
t o
f
s
u
b
j
ec
ts
is
esp
ec
iall
y
d
i
s
p
r
o
p
o
r
tio
n
ate
[
1
0
-
1
1
]
.
3
.
2
.
I
m
a
g
e
d
eno
is
ing
Nu
m
er
icall
y
,
p
ictu
r
e
co
m
m
o
tio
n
i
s
p
o
r
tr
ay
ed
a
s
a
m
u
lt
i
-
d
i
m
en
s
io
n
al
s
to
ch
a
s
tic
p
r
o
ce
s
s
[
1
2
]
.
T
h
er
ef
o
r
e,
p
ictu
r
e
co
m
m
o
ti
o
n
ca
n
b
e
n
u
m
er
icall
y
p
o
r
tr
ay
ed
b
y
s
cie
n
ti
f
ic
i
n
s
i
g
h
t
s
,
f
o
r
ex
a
m
p
le
th
r
o
u
g
h
li
k
eli
h
o
o
d
th
ic
k
n
e
s
s
cir
cu
latio
n
w
o
r
k
[
1
3
]
.
T
h
e
p
r
esen
tat
io
n
o
f
a
s
c
ien
t
if
ic
m
o
d
el
ca
n
ac
co
m
p
lis
h
b
etter
d
en
o
is
in
g
.
No
is
e
Mo
d
el
(
1
)
R
ay
le
ig
h
n
o
i
s
e
(
1
)
C
o
n
d
itio
n
1
is
th
e
li
k
eli
h
o
o
d
th
ick
n
e
s
s
ap
p
r
o
p
r
iatio
n
ca
p
ac
it
y
of
R
a
y
lei
g
h
co
m
m
o
tio
n
.
W
h
en
th
e
g
r
a
y
v
al
u
e
is
m
o
r
e
p
r
o
m
in
e
n
t
th
a
n
or
eq
u
iv
ale
n
t
to
t
h
e
lik
e
lih
o
o
d
t
h
ick
n
es
s
b
en
d
g
r
ad
es
to
o
n
e
s
id
e
a
n
d
t
h
e
f
u
n
d
a
m
e
n
ta
l
r
eg
io
n
on
t
h
e
co
r
r
ec
t
s
id
e
is
b
ig
g
er
,
t
h
at
is
,
t
h
e
d
i
m
est
i
m
at
i
o
n
of
co
m
m
o
t
io
n
f
o
c
u
s
e
s
is
m
o
r
e
d
is
tr
ib
u
ted
in
th
e
r
ig
h
t
s
id
e
of
t
h
e
ce
n
tr
al
ax
i
s
+
√
(
/
2
)
(
2
)
Gau
s
s
ia
n
n
o
is
e
(
2
)
C
o
n
d
itio
n
2
is
th
e
li
k
eli
h
o
o
d
th
ick
n
e
s
s
ap
p
r
o
p
r
iatio
n
ca
p
ac
it
y
of
t
h
e
Ga
u
s
s
ian
co
m
m
o
tio
n
.
T
h
e
g
r
a
y
v
al
u
e
of
th
e
cla
m
o
r
is
ar
o
u
n
d
th
e
f
o
ca
l
g
r
a
y
s
ca
le,
w
h
ic
h
is
g
e
n
er
all
y
d
ar
k
,
an
d
th
e
h
ig
h
c
o
n
tr
ast
co
m
m
o
tio
n
cir
cu
latio
n
is
les
s
[
1
4
]
.
Gau
s
s
ia
n
co
m
m
o
tio
n
is
li
k
e
w
i
s
e
ca
lled
o
r
d
i
n
ar
y
n
o
is
e.
I
t
s
li
k
eli
h
o
o
d
th
ic
k
n
e
s
s
co
m
p
lie
s
w
it
h
t
y
p
ical
cir
c
u
la
tio
n
.
It
is
a
g
en
er
all
y
u
tili
ze
d
co
m
m
o
tio
n
m
o
d
el.
Gau
s
s
ia
n
n
o
is
e
is
n
o
r
m
all
y
b
r
o
u
g
h
t
ab
o
u
t
by
a
w
f
u
l
li
g
h
tin
g
or
h
ig
h
te
m
p
er
at
u
r
e
d
u
r
in
g
p
r
o
cu
r
e
m
en
t
[
1
5
]
.
3
.
3
.
F
e
a
t
ure
e
x
t
ra
ct
io
n
Featu
r
e
ex
tr
ac
tio
n
is
a
s
tr
ateg
y
o
f
d
i
m
e
n
s
io
n
alit
y
d
ec
lin
e
b
y
w
h
ic
h
a
f
u
n
d
a
m
en
ta
l
g
a
m
e
p
lan
o
f
r
o
u
g
h
d
ata
is
r
ed
u
ce
d
to
lo
g
ic
all
y
s
en
s
ib
le
s
o
cial
a
f
f
a
ir
s
f
o
r
tak
i
n
g
ca
r
e
o
f
[
1
6
]
.
Natu
r
e
o
f
t
h
ese
tr
e
m
e
n
d
o
u
s
en
li
g
h
te
n
i
n
g
ass
o
r
t
m
en
ts
is
i
n
n
u
m
er
ab
le
ele
m
en
t
s
th
a
t
r
eq
u
i
r
e
a
h
u
g
e
a
m
o
u
n
t
o
f
en
r
o
lli
n
g
ad
v
an
ta
g
es
f
o
r
th
e
p
r
o
ce
s
s
[
1
7
]
.
Featu
r
e
ex
tr
ac
ti
o
n
is
t
h
e
n
a
m
e
f
o
r
m
et
h
o
d
s
th
at
s
elec
t
an
d
ad
d
itio
n
all
y
m
er
g
e
f
ac
to
r
s
i
n
to
f
ea
t
u
r
es,
e
f
f
ec
tiv
e
l
y
d
ec
r
ea
s
in
g
t
h
e
p
r
o
p
o
r
tio
n
o
f
d
ata
th
a
t
m
u
s
t
b
e
tak
e
n
ca
r
e
o
f
,
w
h
ile
s
till
ab
s
o
lu
t
el
y
a
n
d
th
o
r
o
u
g
h
l
y
d
ep
ictin
g
th
e
p
r
i
m
ar
y
i
n
s
tr
u
cti
v
e
as
s
o
r
t
m
e
n
t [
1
8
-
1
9
]
.
I
n
th
is
s
t
u
d
y
,
t
h
e
h
ig
h
li
g
h
ts
ca
n
b
e
s
ep
ar
ated
b
y
u
tili
z
in
g
C
NN.
Utilizi
n
g
lo
ca
le
-
b
ased
C
NN
f
in
d
i
n
g
k
e
y
p
o
s
i
tio
n
s
,
m
a
k
i
n
g
a
s
lid
i
n
g
w
i
n
d
o
w
o
n
t
h
e
p
ict
u
r
e,
an
d
m
o
v
i
n
g
t
h
e
s
lid
in
g
w
i
n
d
o
w
a
lo
n
g
t
h
e
p
ict
u
r
e
to
g
et
th
e
p
o
ten
tia
l
o
b
j
ec
tiv
e
zo
n
e,
C
NN
is
u
tili
ze
d
to
r
e
m
o
v
e
th
e
s
tan
d
ar
d
h
i
g
h
lig
h
t
s
o
f
t
h
e
o
b
j
ec
tiv
e
r
eg
io
n
,
th
at
is
,
to
g
et
t
h
e
y
ie
ld
o
f
f
i
x
e
d
m
ea
s
u
r
e
m
e
n
ts
as
p
er
co
n
v
o
lu
tio
n
,
p
o
o
lin
g
an
d
d
if
f
er
en
t
t
ask
s
.
A
t
t
h
at
p
o
in
t,
th
e
y
ie
ld
v
e
cto
r
s
o
f
th
e
s
u
b
s
eq
u
en
t
s
ta
g
e
ar
e
g
r
o
u
p
ed
(
class
i
f
ier
s
s
h
o
u
ld
b
e
p
r
ep
ar
ed
b
y
t
h
eir
h
i
g
h
li
g
h
ts
)
;
an
d
th
e
f
ac
ial
ag
e
i
s
an
t
icip
ate
d
u
tili
zi
n
g
t
h
e
C
NN
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J
A
r
ti
f
I
n
tell
I
SS
N:
2252
-
8938
A
g
e
-
b
a
s
ed
f
a
cia
l rec
o
g
n
itio
n
u
s
in
g
co
n
v
o
lu
ted
n
e
u
r
a
l
... (
Ju
l
iu
s
Yo
n
g
W
u
Jien
)
427
3
.
4
.
T
ra
ini
ng
s
et
3
.
4
.
1
.
Co
nv
o
lutio
na
l
neura
l
net
w
o
rk
s
t
ruct
ure
T
h
e
class
ical
n
et
w
o
r
k
s
tr
u
ct
u
r
e
o
f
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
s
y
s
t
e
m
ap
p
ea
r
s
i
n
Fig
u
r
e
2
.
T
h
e
c
o
n
v
o
l
u
tio
n
n
eu
r
al
s
y
s
te
m
co
n
tai
n
s
a
f
e
w
"
co
n
v
o
lu
tio
n
la
y
er
s
"
an
d
"
ex
am
i
n
in
g
la
y
er
s
"
to
p
r
o
ce
s
s
th
e
in
f
o
r
m
atio
n
s
i
g
n
al.
A
t
t
h
at
p
o
in
t,
t
h
e
m
ap
p
i
n
g
b
et
w
ee
n
th
e
i
n
f
o
s
ig
n
al
a
n
d
th
e
y
ield
r
es
u
lt
is
ac
k
n
o
w
l
ed
g
ed
in
th
e
f
u
ll
ass
o
ciatio
n
la
y
er
.
E
v
er
y
co
n
v
o
lu
tio
n
ex
tr
icate
s
t
h
e
h
i
g
h
lig
h
ts
o
f
th
e
i
n
f
o
r
m
atio
n
s
i
g
n
al
th
r
o
u
g
h
a
co
n
v
o
lu
t
io
n
ac
tiv
it
y
o
f
a
co
n
v
o
lu
t
io
n
f
il
ter
[
2
0
-
2
2
]
.
E
x
a
m
i
n
i
n
g
la
y
er
is
l
ik
e
w
i
s
e
ca
lled
th
e
"
ass
e
m
b
l
y
"
la
y
er
.
I
t
s
ca
p
ac
it
y
is
to
u
ti
lize
th
e
s
ta
n
d
ar
d
o
f
n
e
ig
h
b
o
r
h
o
o
d
r
elatio
n
s
h
ip
to
d
o
w
n
e
x
a
m
p
le,
w
h
ic
h
less
e
n
s
i
n
f
o
r
m
atio
n
(
d
ec
r
ea
s
es c
alcu
latio
n
)
a
n
d
h
o
ld
s
v
al
u
ab
le
d
ata
in
th
e
s
y
s
te
m
[
2
3
]
.
Af
ter
th
e
p
ict
u
r
e
g
o
es
th
r
o
u
g
h
all
co
n
v
o
l
u
tio
n
la
y
er
s
an
d
test
i
n
g
la
y
er
s
,
th
e
ele
m
en
t
m
ap
p
in
g
is
n
o
r
m
all
y
c
h
an
g
ed
o
v
er
i
n
to
h
i
g
h
l
ig
h
t
v
ec
to
r
y
ield
b
y
f
u
ll
a
s
s
o
ciatio
n
ac
ti
v
it
y
,
a
s
it
w
er
e,
t
h
e
f
u
l
l
as
s
o
ciatio
n
la
y
er
is
f
r
a
m
ed
.
T
h
e
ele
m
en
t
m
ap
p
in
g
ca
n
b
e
ass
o
ciate
d
p
er
s
is
te
n
tl
y
f
o
r
c
o
m
m
o
n
l
y
,
an
d
t
h
e
la
s
t
f
u
ll
a
s
s
o
ciatio
n
la
y
er
is
th
e
y
ield
la
y
er
.
B
y
an
d
b
y
,
th
e
y
ield
la
y
er
is
f
r
eq
u
e
n
tl
y
u
ti
lized
as
th
e
ele
m
en
t
o
f
p
ictu
r
e
ex
tr
ac
tio
n
,
an
d
th
e
s
e
h
ig
h
li
g
h
ts
ar
e
g
e
n
er
all
y
u
tili
ze
d
f
o
r
r
elap
s
e
ch
ar
ac
ter
izatio
n
p
r
ep
ar
in
g
(
f
o
r
p
r
ep
ar
in
g
SVM
o
r
So
f
t M
a
x
)
[
2
4
-
2
5
]
.
Fig
u
r
e
2.
T
h
e
class
ical
s
tr
u
ct
u
r
e
of
C
NN
4.
CO
NCLU
SI
O
N
I
t
ten
d
s
to
b
e
an
ticip
ated
t
h
at
th
e
p
r
o
g
r
a
m
m
ed
f
ac
ial
a
g
e
e
v
alu
atio
n
s
tr
ate
g
y
d
ep
en
d
en
t
o
n
C
NN
h
a
s
p
r
ef
er
r
ed
ex
ec
u
tio
n
o
v
er
t
h
at
d
ep
en
d
en
t
o
n
ar
tific
ial
f
ea
t
u
r
es
o
f
S
VM
.
T
h
e
ex
p
la
n
atio
n
m
i
g
h
t
b
e
t
h
at
t
h
e
C
NN
ca
n
p
r
o
cu
r
e
p
r
o
g
r
ess
i
v
el
y
i
n
e
x
h
a
u
s
tib
le
an
d
s
i
g
n
i
f
ica
n
t
h
i
g
h
li
g
h
ts
i
n
f
ac
ial
p
i
ctu
r
es
b
y
lear
n
in
g
,
w
h
ile
t
h
e
f
al
s
el
y
p
la
n
n
ed
h
ig
h
lig
h
t
s
ca
n
j
u
s
t
co
v
er
s
o
m
e
f
i
x
ed
an
d
s
i
n
g
le
h
i
g
h
lig
h
ts
i
n
th
e
p
ict
u
r
e,
an
d
th
e
d
e
m
ea
n
o
r
o
f
h
i
g
h
l
ig
h
t
s
is
n
o
t
r
ic
h
.
O
v
er
all
,
t
h
is
s
t
u
d
y
ai
m
s
to
g
i
v
e
a
n
o
v
er
v
ie
w
o
f
th
e
i
m
p
r
o
v
e
m
e
n
t
o
f
th
e
ac
cu
r
ac
y
i
n
d
etec
ti
n
g
t
h
e
ag
e
o
f
f
ac
ial
r
ec
o
g
n
it
io
n
.
B
esid
es,
f
ac
ial
i
m
a
g
e
s
d
ataset
s
th
at
w
er
e
p
r
o
p
o
s
ed
to
u
s
e
in
t
h
is
p
ap
er
f
o
r
f
u
r
t
h
e
r
s
tu
d
y
.
Facial
i
m
a
g
es
f
r
o
m
a
d
if
f
er
en
t
r
eg
io
n
i
n
th
e
w
o
r
ld
s
u
c
h
as
C
X
R
f
r
o
m
Af
r
ica
o
r
f
r
o
m
th
e
w
ester
n
c
o
u
n
tr
y
ca
n
b
e
ap
p
lied
to
o
.
Fu
r
th
er
m
o
r
e,
t
h
e
to
tal
n
u
m
b
er
o
f
d
atasets
u
s
e
as
tr
ain
i
n
g
a
n
d
v
alid
atio
n
s
u
g
g
es
ted
to
in
cr
ea
s
e
as it
m
i
g
h
t i
m
p
r
o
v
e
o
v
er
all
p
er
f
o
r
m
an
ce
.
RE
F
E
R
E
NC
E
S
[1
]
Io
a
n
n
is
A
.
Ka
k
a
d
iaris,
G
e
o
rg
e
T
o
d
e
rici,
G
e
o
rg
io
s
Ev
a
n
g
e
lo
p
o
u
lo
s,
G
e
o
rg
io
s
P
a
ss
a
li
s,
T
h
e
o
h
a
ri
s.
“
3
D
-
2
D
f
a
c
e
re
c
o
g
n
it
io
n
w
it
h
p
o
se
a
n
d
il
l
u
m
i
n
a
ti
o
n
n
o
rm
a
li
z
a
ti
o
n
.
”
Co
mp
u
ter
Vi
sio
n
a
n
d
Im
a
g
e
Un
d
e
rs
ta
n
d
in
g
.
1
5
4
:
1
3
7
-
1
5
1
,
2
0
1
7
.
[2
]
Jia
n
g
Jin
g
Lv
,
X
iao
Hu
S
h
a
o
,
Jia
S
h
u
i
H
u
a
n
g
,
X
ian
g
Do
n
g
Zh
o
u
,
X
i
Z
h
o
u
.
“
Da
ta
a
u
g
m
e
n
tatio
n
f
o
r
f
a
c
e
re
c
o
g
n
it
io
n
.
”
Ne
u
ro
c
o
m
p
u
ti
n
g
.
2
3
0
:
1
8
4
-
1
9
6
,
2
0
1
7
.
[3
]
V
a
ib
h
a
v
Ja
in
,
Din
e
sh
P
a
tel.
“
A
GP
U
b
a
se
d
im
p
lem
e
n
tatio
n
o
f
Ro
b
u
st
F
a
c
e
d
e
tec
ti
o
n
S
y
ste
m
.
”
Pro
c
e
d
ia
C
o
mp
u
ter
S
c
ien
c
e
.
87
:
1
5
6
-
1
6
3
,
2
0
1
6
.
[4
]
A
n
d
re
w
J.L
o
g
a
n
,
Ga
e
l
E.
G
o
rd
o
n
,
G
u
n
ter L
o
ff
ler.
“
Co
n
tri
b
u
ti
o
n
s o
f
in
d
iv
id
u
a
l
f
a
c
e
f
e
a
tu
re
s to
f
a
c
e
d
isc
rim
in
a
ti
o
n
”
,
Vi
sio
n
Res
e
a
rc
h
.
1
3
7
:
2
9
-
39
,
2
0
1
7
.
[5
]
L
a
n
e
,
T
.
M
.
,
Niru
,
M
.
,
A
ra
n
,
N.,
S
u
ry
a
,
G
.
,
S
tep
h
e
n
,
A
.
B.
e
t
a
l.
“
De
e
p
L
e
a
rn
in
g
M
o
d
e
ls
o
f
th
e
R
e
t
in
a
l
Re
sp
o
n
se
t
o
Na
tu
ra
l
S
c
e
n
e
s,”
in
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
3
0
th
C
o
n
fer
e
n
c
e
o
n
Ne
u
ra
l
In
f
o
rm
a
ti
o
n
Pr
o
c
e
ss
in
g
S
y
ste
ms
(
NIPS
2
0
1
6
)
,
Ba
rc
e
lo
n
a
,
S
p
a
i
n
.
[6
]
A
.
S
e
p
a
s
-
M
o
g
h
a
d
d
a
m
,
F
.
M
.
P
e
re
ira,
P
.
L
.
Co
rre
ia.
“
F
a
c
e
re
c
o
g
n
it
io
n
:
a
n
o
v
e
l
m
u
lt
i
-
lev
e
l
tax
o
n
o
m
y
b
a
se
d
su
rv
e
y
,
”
ET
Bi
o
me
trics
.
9
(2
)
:
5
8
-
67
,
2
0
2
0
.
[7
]
S
.
S
o
n
g
,
V
.
Ch
a
n
d
ra
se
k
h
a
r,
B.
M
a
n
d
a
l
e
t
a
l.
“
M
u
lt
im
o
d
a
l
M
u
lt
i
-
S
trea
m
D
e
e
p
L
e
a
rn
in
g
f
o
r
Eg
o
c
e
n
tri
c
A
c
ti
v
it
y
Re
c
o
g
n
it
io
n
,
”
in
Pro
c
e
e
d
in
g
s
o
f
th
e
2
9
t
h
IEE
E
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
Vi
sio
n
a
n
d
Pa
t
ter
n
Rec
o
g
n
it
io
n
W
o
rk
sh
o
p
s,
CVP
RW
2
0
1
6
,
p
p
.
3
7
8
–
3
8
5
,
2
0
1
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8938
I
n
t J
A
r
ti
f
I
n
tell
,
Vo
l.
9
,
No
.
3
,
Sep
te
m
b
er
20
20
:
4
2
4
–
428
428
[8
]
G
.
B.
Hu
a
n
g
,
M
.
Ra
m
e
sh
,
T
.
B
e
rg
,
a
n
d
E.
L
e
a
rn
e
d
-
M
il
ler
e
t
a
l.
“
L
a
b
e
led
F
a
c
e
s
in
th
e
W
il
d
:
A
D
a
tab
a
se
f
o
r
S
tu
d
y
in
g
F
a
c
e
Re
c
o
g
n
it
io
n
in
Un
c
o
n
stra
i
n
e
d
E
n
v
iro
n
m
e
n
ts,"
in
"
T
e
c
h
n
ica
l
Re
p
o
rt
07
-
4
9
,
"
Un
iv
e
rsity
of
M
a
ss
a
c
h
u
se
tt
s,
Am
h
e
rst.
2
0
0
7
.
[9
]
G
.
C
y
b
e
n
k
o
,
A
p
p
ro
x
ima
ti
o
n
b
y
su
p
e
rp
o
sit
io
n
s
o
f
a
si
g
m
o
id
a
l
f
u
n
c
ti
o
n
,
M
a
t
h
e
m
a
ti
c
s
o
f
c
o
n
tr
o
l,
sig
n
a
ls
a
n
d
s
y
ste
m
s
.
2
(4
)
:
3
0
3
-
3
1
4
,
1
9
8
9
.
[1
0
]
O.
De
lallea
u
,
Y.
Be
n
g
i
o
,
“
S
h
a
ll
o
w
v
s.
d
e
e
p
su
m
-
p
ro
d
u
c
t
n
e
t
w
o
rk
s,
”
in
:
Ad
v
a
n
c
e
s
in
Ne
u
r
a
l
I
n
fo
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
,
p
p
.
6
6
6
–
6
7
4
,
2
0
1
1
.
[1
1
]
Z.
L
u
,
H.
P
u
,
F
.
W
a
n
g
,
Z.
Hu
,
L.
W
a
n
g
,
“
T
h
e
e
x
p
re
s
siv
e
p
o
w
e
r
o
f
n
e
u
ra
l
n
e
tw
o
rk
s:
A
v
ie
w
f
ro
m
th
e
w
id
th
,
”
in
:
Ad
v
a
n
c
e
s i
n
Ne
u
ra
l
In
f
o
rm
a
ti
o
n
Pro
c
e
ss
in
g
S
y
ste
ms
,
p
p
.
6
2
3
1
-
6
2
3
9
,
2
0
1
7
.
[1
2
]
M
.
Us
m
a
n
(Ak
ra
m
),
A
a
sia
Kh
a
n
u
m
,
“
Re
ti
n
a
l
im
a
g
e
s:
b
lo
o
d
v
e
ss
e
l
se
g
m
e
n
tatio
n
b
y
th
re
sh
o
ld
p
r
o
b
in
g
,
”
i
n
:
I
EE
E
S
y
mp
o
si
u
m
o
n
In
d
u
s
tri
a
l
El
e
c
tro
n
ics
&
Ap
p
li
c
a
ti
o
n
s
(
IS
IEA
)
,
3
–
6
th
,
Oc
to
b
e
r,
2
0
1
0
,
P
e
n
a
n
g
,
M
a
l
a
y
sia
,
p
p
.
4
9
3
-
4
9
7
,
2
0
1
0
.
[1
3
]
H.
N.
M
h
a
sk
a
r,
T
.
P
o
g
g
io
,
“
De
e
p
v
s.
sh
a
ll
o
w
n
e
tw
o
rk
s:
A
n
a
p
p
ro
x
im
a
ti
o
n
th
e
o
ry
p
e
rsp
e
c
ti
v
e
,
”
An
a
lys
is
a
n
d
Ap
p
li
c
a
ti
o
n
s
,
1
4
(
0
6
)
:
8
2
9
-
8
4
8
2
0
1
6
.
[1
4
]
H.
M
h
a
sk
a
r,
Q.
L
iao
,
T
.
P
o
g
g
io
,
“
L
e
a
rn
in
g
f
u
n
c
ti
o
n
s:
w
h
e
n
is
d
e
e
p
b
e
tt
e
r
t
h
a
n
sh
a
l
lo
w
,
”
a
rX
iv
p
re
p
r
in
t
a
rX
iv
:1
6
0
3
.
0
0
9
8
8
.
[1
5
]
G
.
F
.
M
o
n
t
u
f
a
r,
R.
P
a
sc
a
n
u
,
K
.
Ch
o
,
Y.
Be
n
g
io
,
“
On
t
h
e
n
u
m
b
e
r
o
f
li
n
e
a
r
re
g
io
n
s
o
f
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
s,
”
in
:
Ad
v
a
n
c
e
s i
n
n
e
u
r
a
l
i
n
f
o
rm
a
ti
o
n
p
ro
c
e
ss
in
g
sy
ste
ms
,
p
p
.
2
9
2
4
–
2
9
3
2
,
2
0
1
4
.
[1
6
]
G
a
sa
h
,
M
.
,
Ba
h
a
ru
m
,
A
.
,
Zain
,
N
.
H.M
.
,
Ha
lam
y
,
S
.
,
Ha
n
a
p
i,
R.
,
No
o
r,
N.A
.
M
.
“
Ev
a
lu
a
ti
o
n
o
f
P
o
sit
iv
e
E
m
o
ti
o
n
i
n
Ch
il
d
re
n
M
o
b
il
e
L
e
a
rn
in
g
A
p
p
li
c
a
ti
o
n
,
”
B
u
ll
e
ti
n
o
f
El
e
c
trica
l
En
g
i
n
e
e
rin
g
a
n
d
In
fo
rm
a
t
ics
9
(2
)
:
8
1
8
-
8
2
6
,
2
0
2
0
.
[1
7
]
Yin
Bi,
M
in
g
so
n
g
L
v
,
Ya
n
g
ji
e
W
e
i,
Na
n
G
u
a
n
,
W
a
n
g
Yi,
“
M
u
lt
i
-
f
e
a
tu
re
f
u
sio
n
f
o
r
th
e
rm
a
l
fa
c
e
re
c
o
g
n
it
io
n
,
”
V
o
l
u
m
e
7
7
:
3
6
6
-
3
7
4
,
2
0
1
6
.
[1
8
]
Ay
a
n
S
e
a
l,
De
b
o
to
s
h
Bh
a
tt
a
c
h
a
rj
e
e
,
M
it
a
Na
sip
u
ri.
“
Hu
m
a
n
fa
c
e
re
c
o
g
n
it
io
n
u
si
n
g
ra
n
d
o
m
f
o
re
st
b
a
se
d
f
u
sio
n
o
f
à
-
tro
u
s
w
a
v
e
let
t
ra
n
s
f
o
rm
c
o
e
ff
ic
ien
ts
f
ro
m
th
e
r
m
a
l
a
n
d
v
isib
le
i
m
a
g
e
s
,
”
AEU
-
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
El
e
c
tro
n
ic
s
a
n
d
Co
mm
u
n
ica
t
io
n
s
,
7
0
(8
)
:
1
0
4
1
-
1
0
4
9
,
2
0
1
6
.
[1
9
]
Li
-
F
a
n
g
Zh
o
u
,
Y
u
e
-
W
e
iDu
,
W
e
i
-
S
h
e
n
g
L
i,
Jia
n
-
Xu
n
M
i,
X
iao
L
u
a
n
.
“
P
o
se
-
ro
b
u
st
f
a
c
e
re
c
o
g
n
it
io
n
w
it
h
Hu
ffm
a
n
-
L
BP
e
n
h
a
n
c
e
d
b
y
Div
id
e
-
a
n
d
-
Ru
l
e
stra
teg
y
”
,
Pa
tt
e
rn
Rec
o
g
n
it
io
n
,
78
:
4
3
-
55
,
2
0
1
8
.
[2
0
]
T
h
ib
a
u
lt
Na
p
o
leo
n
,
A
y
m
a
n
A
l
fa
lo
u
.
“
P
o
se
i
n
v
a
rian
t
f
a
c
e
re
c
o
g
n
it
io
n
:
3
D
m
o
d
e
l
f
ro
m
sin
g
le
p
h
o
t
o
,
”
Op
ti
c
s
a
n
d
L
a
se
rs
in
E
n
g
i
n
e
e
rin
g
,
89
:
1
5
0
-
1
6
1
,
2
0
1
7
.
[2
1
]
M
ich
e
le
Na
p
p
i
,
S
tef
a
n
o
Ricc
iar
d
i,
M
a
ss
im
o
T
istar
e
ll
i.
“
De
c
e
iv
i
n
g
f
a
c
e
s:
W
h
e
n
p
las
ti
c
su
rg
e
ry
c
h
a
ll
e
n
g
e
s
f
a
c
e
re
c
o
g
n
it
io
n
,
”
Ima
g
e
a
n
d
V
isio
n
C
o
mp
u
t
in
g
,
54
:
7
1
-
82
,
2
0
1
6
.
[2
2
]
Ya
n
L
ian
g
,
Yu
n
Zh
a
n
g
,
X
ian
-
X
i
a
n
Zen
g
.
“
P
o
se
-
i
n
v
a
rian
t
3
D
f
a
c
e
re
c
o
g
n
it
io
n
u
si
n
g
h
a
lf
f
a
c
e
”
S
ig
n
a
l
Pro
c
e
ss
in
g
:
Ima
g
e
Co
mm
u
n
i
c
a
t
io
n
,
57:
84
-
90
,
2
0
1
7
.
[2
3
]
Ch
e
n
f
e
i
X
u
,
Qih
e
L
iu
,
M
a
o
Ye
.
“
Ag
e
in
v
a
rian
t
fa
c
e
re
c
o
g
n
it
io
n
a
n
d
re
tri
e
v
a
l
b
y
c
o
u
p
led
a
u
to
-
e
n
c
o
d
e
r
Ne
tw
o
rk
s
,
”
Ne
u
ro
c
o
mp
u
ti
n
g
,
Vo
lu
m
e
2
2
2
:
6
2
-
71
,
2
0
1
7
.
[2
4
]
S
u
n
g
Eu
n
Ch
o
i,
Ja
e
ik
Jo
,
S
a
n
g
h
a
k
L
e
e
,
He
e
s
e
u
n
g
Ch
o
i,
Ig
-
Ja
e
Ki
m
,
Ja
ih
ie
Ki
m
.
“
A
g
e
f
a
c
e
si
m
u
latio
n
u
si
n
g
a
g
in
g
f
u
n
c
ti
o
n
s
o
n
g
lo
b
a
l
a
n
d
l
o
c
a
l
f
e
a
tu
re
s
w
it
h
re
sid
u
a
l
im
a
g
e
s
,
”
Exp
e
rt
S
y
ste
ms
wit
h
Ap
p
li
c
a
ti
o
n
,
V
o
lu
m
e
8
0
:
1
0
7
-
125
,
2
0
1
7
.
[2
5
]
Q.
X
u
,
M
.
Z
h
a
n
g
,
Z.
G
u
,
G
.
P
a
n
,
“
Ov
e
r
f
it
ti
n
g
re
m
e
d
y
b
y
sp
a
rsify
in
g
re
g
u
lariz
a
ti
o
n
o
n
f
u
ll
y
-
c
o
n
n
e
c
ted
l
a
y
e
rs
o
f
CNN
S
,”
Ne
u
ro
c
o
mp
u
ti
n
g
,
3
2
8
:
69
-
74
,
2
0
1
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.