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.
22
,
No
.
3
,
J
u
n
e
2
0
2
1
,
p
p
.
1436
~
1
4
4
5
I
SS
N:
2
5
02
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
2
2
.i
3
.
pp
1
4
3
6
-
1
4
4
5
1436
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Custo
mer’s
spo
nt
a
neo
us facia
l ex
p
ress
io
n recog
nitio
n
G
o
la
m
M
o
rsh
ed,
H
a
m
i
m
a
h
Uj
ir,
I
rwa
nd
i H
ipi
ny
F
a
c
u
lt
y
o
f
C
o
m
p
u
ter
S
c
ien
c
e
a
n
d
In
fo
rm
a
ti
o
n
Tec
h
n
o
l
o
g
y
,
U
n
iv
e
rs
it
i
M
a
lay
sia
S
a
ra
wa
k
,
M
a
lay
sia
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l 1
6
,
2
0
2
0
R
ev
is
ed
J
u
n
4
,
2
0
2
1
Acc
ep
ted
J
u
n
1
3
,
2
0
2
1
In
th
e
field
o
f
c
o
n
su
m
e
r
sc
ien
c
e
,
c
u
sto
m
e
r
fa
c
ial
e
x
p
re
ss
io
n
is
o
ften
c
a
teg
o
rize
d
e
it
h
e
r
a
s
n
e
g
a
ti
v
e
o
r
p
o
siti
v
e
.
C
u
sto
m
e
r
wh
o
p
o
rtray
s
n
e
g
a
ti
v
e
e
m
o
ti
o
n
to
a
sp
e
c
ifi
c
p
r
o
d
u
c
t
m
o
stly
m
e
a
n
s
t
h
e
y
re
jec
t
t
h
e
p
r
o
d
u
c
t
wh
il
e
a
c
u
sto
m
e
r
with
p
o
siti
v
e
e
m
o
ti
o
n
is
m
o
re
li
k
e
ly
to
p
u
rc
h
a
se
th
e
p
ro
d
u
c
t.
To
o
b
se
rv
e
c
u
st
o
m
e
r
e
m
o
ti
o
n
,
m
a
n
y
re
se
a
rc
h
e
rs
h
a
v
e
stu
d
ied
d
iffere
n
t
p
e
rsp
e
c
ti
v
e
s
a
n
d
m
e
th
o
d
o
l
o
g
i
e
s
to
o
b
tain
h
i
g
h
a
c
c
u
ra
c
y
re
su
lt
s.
C
o
n
v
e
n
t
io
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
(CNN
)
is
u
se
d
to
re
c
o
g
n
ize
c
u
sto
m
e
r
sp
o
n
ta
n
e
o
u
s
fa
c
ial
e
x
p
re
ss
io
n
s.
Th
is
p
a
p
e
r
a
ims
to
re
c
o
g
n
ize
c
u
sto
m
e
r
sp
o
n
ta
n
e
o
u
s
e
x
p
re
ss
io
n
s
wh
il
e
t
h
e
c
u
sto
m
e
r
o
b
se
rv
e
d
c
e
rtain
p
r
o
d
u
c
ts.
We
h
a
v
e
d
e
v
e
l
o
p
e
d
a
c
u
sto
m
e
r
se
rv
ice
sy
ste
m
u
sin
g
a
CNN
t
h
a
t
is
train
e
d
t
o
d
e
tec
t
th
re
e
t
y
p
e
s
o
f
fa
c
ial
e
x
p
r
e
ss
io
n
,
i.
e
.
h
a
p
p
y
,
sa
d
,
a
n
d
n
e
u
t
ra
l.
F
a
c
ial
fe
a
tu
re
s
a
re
e
x
trac
ted
to
g
e
t
h
e
r
with
it
s
h
isto
g
ra
m
o
f
g
ra
d
ien
t
a
n
d
sli
d
in
g
win
d
o
w
.
T
h
e
re
su
lt
s
a
re
t
h
e
n
c
o
m
p
a
re
d
with
th
e
e
x
isti
n
g
w
o
rk
s
a
n
d
it
sh
o
ws
a
n
a
c
h
iev
e
m
e
n
t
o
f
8
2
.
9
%
su
c
c
e
ss
ra
te o
n
a
v
e
ra
g
e
.
K
ey
w
o
r
d
s
:
C
u
s
to
m
er
’
s
em
o
tio
n
Face
d
etec
tio
n
Facial
ex
p
r
ess
io
n
s
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
:
Go
lam
Mo
r
s
h
ed
Facu
lty
o
f
C
o
m
p
u
ter
Scien
ce
an
d
I
n
f
o
r
m
atio
n
T
ec
h
n
o
lo
g
y
Un
iv
er
s
iti Ma
lay
s
ia
Sar
awa
k
,
Ma
lay
s
ia
E
m
ail: g
m
r
iy
al@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Facial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
(
FER)
s
y
s
tem
s
ar
e
cu
r
r
en
tl
y
a
d
v
an
ce
d
in
te
r
m
s
o
f
h
u
m
an
ca
p
ab
ilit
y
in
d
etec
tin
g
an
d
r
esp
o
n
d
in
g
to
f
ac
ial
ex
p
r
ess
io
n
s
wh
ile
in
ter
ac
tin
g
with
ea
ch
o
th
e
r
.
Sev
e
r
al
ap
p
r
o
a
ch
es
ar
e
d
ev
elo
p
e
d
to
d
etec
t c
o
n
s
u
m
e
r
b
eh
av
io
r
le
v
els to
o
b
s
er
v
e
th
ei
r
s
atis
f
ac
tio
n
lev
el
[
1
]
,
[
2
]
.
T
h
ese
ap
p
r
o
ac
h
es a
r
e
u
s
ed
to
d
ev
elo
p
ap
p
licatio
n
s
m
ain
ly
u
s
ed
in
th
e
air
p
o
r
t,
m
ar
k
et,
s
h
o
p
p
in
g
m
all
et
ce
ter
a.
B
y
o
b
s
er
v
in
g
cu
s
to
m
er
’
s
f
ac
ial
ex
p
r
ess
io
n
s
an
d
an
aly
zin
g
th
em
,
th
e
p
r
o
v
id
er
ca
n
o
f
f
e
r
th
e
b
e
s
t
s
er
v
ice
s
u
itab
le
f
o
r
th
e
cu
s
to
m
er
s
.
T
h
is
r
esear
ch
is
m
o
tiv
ated
b
y
t
h
e
cu
s
to
m
er
s
er
v
ice
co
m
p
lain
ts
r
ec
eiv
ed
,
esp
ec
ially
at
th
e
g
o
v
er
n
m
en
t
o
f
f
ices,
an
d
n
o
s
tu
d
y
h
as
f
o
cu
s
ed
o
n
an
aly
s
in
g
th
e
s
eq
u
en
ce
o
f
f
ac
ial
ex
p
r
ess
io
n
s
d
u
r
in
g
t
h
e
in
ter
ac
tio
n
at
th
e
c
u
s
t
o
m
er
s
er
v
ice
d
esk
.
C
o
n
s
u
m
er
’
s
f
ee
lin
g
s
ca
n
b
e
o
b
s
er
v
e
d
f
r
o
m
t
h
eir
f
ac
i
al
ex
p
r
ess
io
n
s
.
T
h
r
o
u
g
h
f
a
ce
-
to
-
f
ac
e
co
m
m
u
n
icatio
n
,
a
s
er
v
ice
p
r
o
v
id
er
ca
n
f
in
d
a
n
af
f
ec
tiv
e
s
tate
b
y
u
s
in
g
f
ac
ial
ex
p
r
ess
io
n
s
.
Ma
n
y
r
esear
ch
er
s
h
av
e
illu
s
tr
ated
th
at
f
ac
ial
e
x
p
r
ess
io
n
s
ar
e
an
im
p
er
ativ
e
m
ar
k
er
o
f
f
ee
lin
g
[
3
]
-
[
5
]
.
R
esear
ch
er
s
h
av
e
d
is
co
v
er
ed
t
h
at
d
is
ap
p
o
in
ted
co
n
s
u
m
er
s
alwa
y
s
u
n
d
er
g
o
n
eg
ativ
e
em
o
tio
n
s
.
B
o
th
c
o
m
p
lain
t
a
n
d
n
o
n
-
co
m
p
lain
t
co
n
s
u
m
er
s
h
av
e
e
x
p
er
ien
ce
d
th
e
s
am
e
n
eg
ativ
e
e
m
o
tio
n
s
th
at
ca
u
s
e
s
er
v
ice
f
ail
u
r
e
[
6
]
.
I
n
a
d
d
itio
n
,
th
e
attem
p
t a
t ser
v
ice
r
ec
o
v
er
y
alwa
y
s
f
ails
if
th
e
s
er
v
ice
p
r
o
v
id
er
is
u
n
ab
le
to
d
is
co
v
er
th
e
n
eg
ativ
e
em
o
tio
n
s
o
f
th
e
co
m
p
lain
er
.
Sev
en
-
ex
p
r
ess
io
n
s
ch
em
a
[
7
]
,
[
8
]
r
ar
ely
o
cc
u
r
s
i
n
g
a
r
m
en
ts
s
h
o
p
p
in
g
an
d
th
er
ef
o
r
e
o
n
ly
th
r
ee
-
ex
p
r
ess
io
n
s
ch
em
a
is
u
s
ed
.
On
e
o
f
th
e
ex
a
m
p
les
in
th
i
s
f
ield
o
f
wo
r
k
is
Sen
tr
ib
u
te
[
9
]
,
w
h
ich
s
im
p
lifie
s
em
o
tio
n
al
an
aly
s
is
to
p
o
s
itiv
e,
n
eu
tr
al,
an
d
n
e
g
ativ
e.
T
o
o
b
s
er
v
e
cu
s
to
m
e
r
in
ter
est
th
r
o
u
g
h
a
s
p
ec
if
ic
p
r
o
d
u
ct
f
r
o
m
th
eir
ex
p
r
ess
io
n
lev
el,
r
esear
ch
er
s
d
ev
elo
p
r
o
b
u
s
t
FER
s
y
s
tem
s
u
s
in
g
d
ee
p
lear
n
i
n
g
ar
c
h
i
tectu
r
e
s
u
ch
as
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
[
1
0
]
,
[
1
1
]
.
C
NN
-
b
ased
FER
s
y
s
tem
s
ar
e
d
ev
elo
p
e
d
to
r
ec
o
g
n
ize
cu
s
to
m
er
s
'
s
e
v
en
k
in
d
s
o
f
f
ac
ial
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
C
u
s
to
mer’
s
s
p
o
n
ta
n
eo
u
s
fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
(
Go
la
m
Mo
r
s
h
ed
)
1437
ex
p
r
ess
io
n
s
f
r
o
m
a
d
if
f
er
e
n
t
p
er
s
p
ec
tiv
e.
I
n
co
n
s
u
m
er
s
cien
ce
,
cu
s
to
m
er
em
o
tio
n
al
f
ac
to
r
s
ar
e
clas
s
if
ied
in
to
two
m
ain
ca
teg
o
r
ies
as
p
o
s
it
iv
e
an
d
n
eg
ativ
e
[
1
2
]
.
W
h
ile
o
b
s
er
v
in
g
a
s
p
ec
if
ic
p
r
o
d
u
c
t
b
y
th
e
cu
s
to
m
er
,
p
o
s
itiv
e
em
o
tio
n
e
n
v
o
k
es
th
e
cu
s
to
m
er
to
ac
ce
p
t
o
r
p
u
r
ch
ase
it
[
1
3
]
.
T
h
is
p
o
s
itiv
e
ex
p
r
ess
io
n
is
o
b
s
er
v
e
d
f
r
o
m
th
eir
h
ap
p
y
ex
p
r
ess
io
n
.
B
esid
es,
n
eg
ativ
e
em
o
tio
n
s
b
y
th
e
cu
s
to
m
er
m
o
r
e
lik
ely
to
r
ejec
t
th
e
p
r
o
d
u
ct
.
T
h
is
n
eg
ativ
e
ex
p
r
ess
io
n
ca
n
b
e
o
b
s
er
v
ed
in
th
eir
s
ad
ex
p
r
e
s
s
io
n
.
T
o
an
aly
ze
th
e
cu
s
to
m
er
f
ac
ial
ex
p
r
ess
io
n
b
y
C
NN
n
etwo
r
k
we
ch
o
o
s
e
o
n
ly
th
r
ee
e
x
p
r
ess
io
n
s
as h
ap
p
y
,
s
ad
,
an
d
n
eu
tr
al.
T
h
e
ex
is
tin
g
m
eth
o
d
f
o
r
f
a
cial
ex
p
r
ess
io
n
s
h
as
s
o
m
e
l
im
itatio
n
s
.
T
h
e
two
-
d
im
e
n
s
io
n
al
(
2
D
)
ap
p
ea
r
an
ce
-
b
ased
lo
ca
l
ap
p
r
o
ac
h
f
o
r
th
e
ex
tr
ac
tio
n
o
f
in
tr
an
s
ien
t
f
ac
ial
f
ea
tu
r
es
an
d
r
e
co
g
n
itio
n
d
o
es
n
o
t
co
n
tain
en
o
u
g
h
f
ea
t
u
r
e
p
o
in
ts
wh
ich
r
esu
lts
in
an
in
co
r
r
ec
t
ac
cu
r
ac
y
r
ate.
C
u
s
to
m
er
s
ten
d
to
ac
t
f
aster
an
d
ch
an
g
e
th
eir
e
x
p
r
ess
io
n
s
s
p
o
n
tan
eo
u
s
ly
.
T
h
e
e
f
f
ec
tiv
e
way
to
u
n
d
er
s
tan
d
th
e
cu
s
to
m
er
'
s
n
ee
d
s
is
to
an
aly
ze
th
eir
s
p
o
n
tan
eo
u
s
ex
p
r
ess
io
n
s
.
Mo
r
eo
v
er
,
it
is
im
p
o
r
tan
t
to
an
aly
ze
th
e
s
eq
u
en
ce
o
f
f
ac
i
al
ex
p
r
ess
io
n
s
an
d
r
elate
th
em
to
th
e
ca
s
e
ev
e
n
ts
.
No
n
eth
eless
,
n
o
s
tu
d
y
h
as f
o
c
u
s
ed
o
n
th
is
ar
ea
.
T
h
e
C
NN
-
b
ased
FER
s
y
s
tem
s
o
lv
ed
a
lo
t
o
f
p
r
o
b
lem
s
s
u
ch
as
f
r
o
n
tal
f
ac
es,
n
o
is
e,
an
d
illu
m
in
atio
n
is
s
u
es.
B
y
im
p
lem
en
tin
g
d
r
o
p
o
u
t
an
d
c
o
n
tr
o
llin
g
o
v
er
f
itti
n
g
,
th
e
C
NN
n
etwo
r
k
b
ec
o
m
es
m
o
r
e
r
o
b
u
s
t
[
1
4
]
.
T
h
e
ac
cu
r
ac
y
o
f
th
e
C
NN
n
etwo
r
k
h
ig
h
ly
d
ep
e
n
d
ed
o
n
th
e
f
e
atu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
e
a
n
d
class
if
ier
.
I
n
m
o
s
t
o
f
th
e
C
NN
-
b
ased
s
y
s
tem
s
,
av
ailab
le
class
if
ier
s
s
u
ch
as
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
,
b
a
g
g
in
g
,
etc
ar
e
b
ein
g
u
s
ed
with
s
o
m
e
lim
itatio
n
s
[
1
5
]
.
T
o
tr
ain
th
e
C
NN
n
etwo
r
k
av
ai
lab
le
f
ac
ial
ex
p
r
ess
io
n
d
ataset
is
b
ein
g
u
s
ed
.
Du
e
to
th
e
lim
itatio
n
o
f
th
o
s
e
tech
n
iq
u
es
th
e
C
NN
n
etwo
r
k
ac
c
u
r
ac
y
r
ate
was
r
ed
u
ce
d
.
I
n
th
e
C
NN
n
etwo
r
k
,
th
e
d
ee
p
er
n
etwo
r
k
ca
n
h
elp
to
i
m
p
r
o
v
e
th
e
ac
cu
r
ac
y
r
ate
in
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
[
1
6
]
.
I
n
t
h
e
tr
ain
in
g
p
r
o
ce
s
s
,
ea
ch
lay
er
will
lear
n
f
r
o
m
th
e
p
o
s
s
ib
le
in
p
u
t
v
alu
e
f
r
o
m
th
e
p
r
ev
io
u
s
lay
er
s
.
I
f
th
e
f
ir
s
t
lay
er
ca
p
tu
r
es
lo
w
-
lev
el
f
ea
tu
r
es,
th
en
th
e
n
e
x
t
l
ev
el
will
also
ca
p
tu
r
e
lo
w
-
le
v
el
f
ea
tu
r
es.
As
a
r
esu
lt,
ea
c
h
lay
er
o
f
th
e
C
NN
n
etwo
r
k
is
tr
ain
e
d
with
l
o
w
-
lev
el
f
ea
tu
r
es
wh
ic
h
will
r
ed
u
ce
th
e
ac
cu
r
ac
y
le
v
el.
T
o
o
v
e
r
co
m
e
th
is
p
r
o
b
lem
,
m
o
r
e
lay
er
s
n
ee
d
t
o
b
e
ad
d
e
d
to
th
e
C
NN
n
etwo
r
k
th
at
will
h
elp
to
lear
n
f
r
o
m
v
a
r
io
u
s
lev
els
o
f
ab
s
tr
ac
tio
n
s
.
I
n
o
u
r
ex
p
e
r
im
en
t,
we
a
p
p
lied
m
u
ltip
le
lay
er
s
in
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
to
im
p
r
o
v
e
ac
cu
r
ac
y
.
T
h
i
s
p
a
p
e
r
p
r
es
e
n
ts
f
i
n
d
i
n
g
s
f
r
o
m
t
h
e
s
p
o
n
t
a
n
e
o
u
s
f
a
c
i
al
e
x
p
r
e
s
s
i
o
n
c
l
as
s
i
f
i
c
at
i
o
n
.
S
ec
t
i
o
n
2
d
e
s
c
r
i
b
es
t
h
e
r
e
l
a
te
d
w
o
r
k
s
i
n
t
h
is
f
i
e
l
d
.
W
e
p
r
es
e
n
t
d
at
a
c
o
ll
e
c
ti
o
n
,
d
a
ta
p
r
e
-
p
r
o
c
e
s
s
i
n
g
,
a
n
d
o
u
r
f
r
a
m
e
w
o
r
k
i
n
s
e
c
t
i
o
n
3
,
f
o
l
l
o
w
e
d
b
y
t
h
e
e
x
p
e
r
i
m
e
n
t
s
e
tt
i
n
g
,
r
e
s
u
l
ts
,
a
n
d
a
n
a
l
y
s
is
i
n
s
ec
t
i
o
n
4
.
F
i
n
a
ll
y
,
t
h
e
c
o
n
c
l
u
s
i
o
n
s
a
r
e
d
r
a
w
n
.
2.
B
ACK
G
RO
UND
C
o
n
s
u
m
er
s
alwa
y
s
s
h
o
w
th
eir
r
ea
ctio
n
s
wh
ile
b
u
y
in
g
a
p
r
o
d
u
ct
o
r
ex
p
er
ien
cin
g
a
s
er
v
ice
.
Usu
ally
,
th
eir
r
ea
ctio
n
r
ef
lects
s
atis
f
a
ctio
n
o
r
d
is
s
atis
f
ac
tio
n
.
I
n
f
a
ce
-
to
-
f
ac
e
cu
s
to
m
er
s
er
v
ice,
it
is
h
ar
d
to
d
etec
t
co
n
s
u
m
er
f
ee
lin
g
s
o
f
lik
e
o
r
d
is
lik
e
if
th
ey
co
m
e
f
r
o
m
d
if
f
er
en
t
co
u
n
tr
ies
an
d
cu
ltu
r
es
[
1
7
]
.
An
et
a
l
.
[
1
8
]
,
wh
en
s
h
o
win
g
d
is
s
atis
f
ac
tio
n
to
war
d
s
a
p
r
o
d
u
ct
o
r
s
er
v
ice,
th
e
co
n
s
u
m
er
wh
o
co
m
es
f
r
o
m
ea
s
ter
n
cu
ltu
r
es
s
h
o
ws
n
eg
ativ
e
e
m
o
ti
o
n
s
r
at
h
er
th
a
n
r
aise
t
h
eir
v
o
ice.
T
h
u
s
,
th
e
im
p
o
r
tan
ce
o
f
u
n
d
er
s
tan
d
in
g
co
n
s
u
m
er
ex
p
r
ess
io
n
s
h
as
r
ec
ei
v
ed
s
i
g
n
if
ican
t
atten
tio
n
f
r
o
m
r
es
ea
r
ch
er
s
in
ten
d
in
g
to
r
ed
u
c
e
th
e
f
ail
u
r
e
r
ate
o
f
th
e
co
m
p
an
y
.
Lu
et
a
l
.
[
1
9
]
p
r
o
p
o
s
es
a
r
ea
l
-
tim
e
au
to
m
ated
r
ec
o
m
m
en
d
e
r
s
y
s
tem
f
o
r
g
ar
m
en
ts
s
ales.
T
h
e
v
id
e
o
-
b
ased
r
ec
o
m
m
en
d
er
(
VAR)
s
y
s
tem
is
d
ev
elo
p
ed
to
o
b
s
er
v
e
t
h
e
cu
s
to
m
er
'
s
em
o
tio
n
s
wh
ile
tr
y
in
g
th
e
g
ar
m
en
t
p
r
o
d
u
ct
in
f
r
o
n
t
o
f
a
m
ir
r
o
r
.
W
ith
th
e
h
el
p
o
f
co
m
p
u
ter
v
is
io
n
tech
n
iq
u
es
an
d
cu
s
to
m
er
p
r
ef
er
en
ce
,
th
e
s
y
s
tem
ca
n
p
r
ed
ict
th
e
c
u
s
to
m
er
d
esire
d
p
r
o
d
u
ct
b
ased
o
n
th
eir
f
ac
ial
e
x
p
r
ess
io
n
s
also
r
ec
o
m
m
e
n
d
in
g
th
e
cu
s
to
m
er
p
r
ef
e
r
en
ce
.
T
h
e
f
ir
s
t
s
tep
o
f
th
is
s
y
s
tem
is
to
ca
p
tu
r
e
th
e
c
u
s
to
m
er
'
s
r
ea
cti
o
n
s
wh
ile
tr
y
in
g
o
r
o
b
s
er
v
in
g
a
g
ar
m
e
n
t
p
r
o
d
u
ct
in
f
r
o
n
t
o
f
th
e
m
i
r
r
o
r
,
u
s
in
g
a
lo
w
-
co
s
t
ca
m
er
a.
T
h
e
n
ex
t
s
tep
is
to
co
m
p
ar
e
cu
s
to
m
er
p
r
ef
e
r
en
ce
with
th
e
f
o
ca
l
cu
s
to
m
er
f
r
o
m
an
alr
ea
d
y
k
n
o
wn
d
atab
ase.
T
h
e
last
s
tep
is
to
r
ec
o
m
m
en
d
th
e
cu
s
to
m
er
v
a
r
io
u
s
p
r
o
d
u
cts
b
ased
o
n
th
eir
r
ev
iewin
g
.
T
h
e
VAR s
y
s
tem
u
s
es v
io
la
-
j
o
n
es [
2
0
]
f
ac
e
d
etec
to
r
s
to
d
etec
t
a
f
ac
e
in
ea
ch
f
r
am
e.
T
o
ac
h
iev
e
th
e
b
est
r
esu
lt,
o
n
ly
f
r
o
n
tal
f
ac
e
im
ag
e
f
r
am
es
ar
e
s
elec
ted
an
d
p
r
o
ce
s
s
ed
to
th
e
n
e
x
t
s
tep
s
.
Af
ter
d
etec
tin
g
th
e
f
ac
e
f
r
o
m
e
ac
h
f
r
am
e,
t
h
e
f
ac
e
is
n
o
r
m
alize
d
i
n
illu
m
in
atio
n
an
d
s
ize
wh
ich
will
b
e
s
e
n
t
f
o
r
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
h
e
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
[
2
1
]
class
if
ier
is
u
s
ed
to
class
if
y
f
ac
ial
e
x
p
r
ess
io
n
s
.
T
h
e
ex
p
r
ess
io
n
f
alls
in
to
t
h
r
e
e
s
ch
em
as
as
p
o
s
itiv
e,
n
e
g
ativ
e,
an
d
n
e
u
tr
al.
T
h
e
v
o
lu
n
teer
s
f
o
r
th
is
ex
p
er
im
en
t
wer
e
ask
ed
to
en
ac
t
ea
ch
ty
p
e
o
f
f
ac
ial
ex
p
r
ess
io
n
b
ein
g
h
ap
p
y
,
u
n
h
ap
p
y
,
a
n
d
n
eu
tr
al
th
r
ee
tim
es
wh
ile
lo
o
k
in
g
at
th
e
ca
m
er
a.
T
h
en
th
ey
wer
e
ask
ed
to
r
ea
d
t
h
e
g
i
v
en
s
en
ten
ce
to
ca
p
t
u
r
e
th
eir
f
ac
ial
r
ea
ctio
n
n
atu
r
ally
.
T
h
e
g
iv
en
s
en
ten
ce
was
p
u
r
p
o
s
ely
s
tr
u
ctu
r
ed
to
tr
ig
g
er
a
s
p
e
cif
ic
r
ea
ctio
n
f
r
o
m
th
e
v
o
lu
n
teer
s
.
I
n
th
e
last
s
tep
,
v
o
lu
n
teer
s
wer
e
ask
ed
to
m
a
k
e
a
p
ar
ticu
lar
ex
p
r
ess
io
n
ag
ai
n
.
All
th
e
s
tep
s
ar
e
r
ec
o
r
d
e
d
as
a
v
id
eo
clip
.
E
ac
h
v
id
eo
clip
was
th
en
ex
tr
ac
t
ed
f
r
am
e
b
y
f
r
am
e
an
d
th
e
b
est
i
m
ag
e
o
f
ea
ch
ex
p
r
ess
io
n
was
m
an
u
ally
p
ic
k
ed
.
T
h
e
ac
c
u
r
ac
y
r
ates
f
o
r
p
o
s
itiv
e,
n
eu
tr
al,
an
d
n
eg
ativ
e
ar
e
8
8
%,
8
1
%,
a
n
d
6
8
% r
esp
ec
tiv
ely
.
L
e
an
d
Vea
[
2
2
]
d
e
v
elo
p
e
d
a
cu
s
to
m
er
f
ac
ial
ex
p
r
ess
io
n
s
y
s
tem
b
y
u
s
in
g
Kin
ec
t
s
en
s
o
r
s
.
T
h
ey
p
r
o
p
o
s
ed
two
m
o
d
els
with
two
Kin
ec
t
s
en
s
o
r
s
V1
an
d
V2
.
T
h
o
s
e
s
en
s
o
r
s
ca
n
tr
ac
k
f
ac
ial
ap
p
ea
r
an
ce
th
r
o
u
g
h
a
t
h
r
ee
-
d
im
en
s
io
n
al
(
3D
)
m
a
s
k
.
T
h
e
em
o
tio
n
lab
elin
g
is
co
n
d
u
cte
d
o
f
th
e
r
ec
o
r
d
in
g
o
f
cu
s
to
m
er
f
ac
ial
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.
22
,
No
.
3
,
J
u
n
e
2
0
2
1
:
1
4
3
6
-
1
4
4
5
1438
ex
p
r
ess
io
n
s
ev
er
y
5
s
ec
o
n
d
s
f
r
o
m
th
eir
r
ec
o
r
d
e
d
v
id
e
o
.
I
f
th
e
cu
s
to
m
er
ex
p
r
ess
io
n
s
ar
e
n
o
t
m
atch
ed
with
th
eir
ex
p
ec
ted
f
ac
ial
ex
p
r
ess
io
n
th
en
th
ey
lev
el
th
is
ex
p
r
ess
io
n
as
“Oth
er
s
”.
T
o
ev
alu
ate
th
e
s
en
s
o
r
p
er
f
o
r
m
an
ce
f
o
u
r
k
i
n
d
s
o
f
c
lass
if
ier
s
ar
e
b
ein
g
u
s
ed
s
u
c
h
as
B
ag
g
in
g
,
J
4
8
,
r
a
n
d
o
m
f
o
r
est,
an
d
r
a
n
d
o
m
c
o
m
m
ittee.
T
h
e
h
ig
h
est
s
u
cc
e
s
s
r
ate
f
o
r
b
o
th
Kin
ec
t
V1
an
d
Kin
ec
t
V2
is
ac
h
iev
e
d
u
s
in
g
r
an
d
o
m
c
o
m
m
ittee
with
9
3
.
0
9
% a
n
d
8
7
.
4
4
%.
Nak
an
o
an
d
Kato
[
2
3
]
d
ev
el
o
p
ed
a
s
y
s
tem
to
esti
m
ate
cu
s
to
m
er
e
x
p
ec
tatio
n
an
d
s
atis
f
ac
tio
n
lev
el
u
s
in
g
a
t
h
r
ee
-
d
im
e
n
s
io
n
al
C
NN.
Fro
m
th
e
cu
s
to
m
er
'
s
f
a
cial
ex
p
r
ess
io
n
s
an
d
b
o
d
y
m
o
tio
n
s
,
th
e
s
y
s
tem
ev
alu
ates
th
e
ex
p
ec
tatio
n
an
d
s
atis
f
ac
tio
n
lev
el
o
f
th
e
c
u
s
to
m
er
.
T
h
e
y
in
tr
o
d
u
ce
d
p
e
r
ce
p
tio
n
c
o
r
r
ec
tio
n
h
y
p
o
th
eses
(
PC
H)
to
ev
alu
at
e
th
e
m
ar
k
etin
g
f
r
o
m
c
o
n
s
u
m
er
s
atis
f
ac
tio
n
lev
el.
PC
H
co
n
d
u
ct
v
o
tin
g
f
o
r
cu
s
to
m
er
s
’
ex
p
ec
tatio
n
an
d
s
atis
f
ac
tio
n
lev
el
b
y
th
e
h
is
to
g
r
am
.
T
h
e
h
is
to
g
r
am
u
s
es
m
ar
k
etin
g
ev
o
l
u
tio
n
as
a
class
an
d
cu
s
to
m
er
s
n
u
m
b
e
r
a
s
a
d
eg
r
ee
.
T
h
e
ex
p
e
r
im
en
t
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
p
r
o
v
id
ed
7
8
.
8
%
ac
cu
r
ac
y
f
o
r
cu
s
to
m
e
r
ex
p
ec
tatio
n
an
d
6
9
.
7
% is
o
b
tain
e
d
f
r
o
m
th
e
s
at
is
f
ac
to
r
y
r
esu
lt.
Mo
u
lay
et
al
.
[
2
4
]
u
s
ed
a
p
atter
n
s
o
f
o
r
ien
ted
ed
g
e
m
a
g
n
itu
d
es
(
POEM
)
d
escr
ip
to
r
an
d
SVM
class
if
ier
to
r
ec
o
g
n
ize
a
cu
s
to
m
er
'
s
f
ac
ial
ex
p
r
ess
io
n
.
T
h
ey
f
o
cu
s
o
n
cu
s
to
m
er
f
ac
ial
ex
p
r
ess
io
n
s
o
n
a
s
p
ec
if
ic
p
r
o
d
u
ct.
T
h
e
ap
p
r
o
ac
h
s
y
s
tem
is
ev
alu
ated
u
s
in
g
J
ap
an
es
e
f
em
ale
f
ac
ial
ex
p
r
ess
io
n
s
(
J
AFFE
)
[
2
5
]
d
ata
s
et.
J
AFFE
d
ata
s
et
co
n
s
i
s
ts
o
f
2
1
3
im
ag
es
with
s
ev
en
b
asic
ex
p
r
ess
io
n
s
.
T
h
e
f
ac
es
ar
e
tak
e
n
f
r
o
m
1
0
J
ap
an
ese
m
o
d
els.
B
ased
o
n
th
e
d
if
f
er
en
t
f
ac
ial
ex
p
r
ess
io
n
s
,
th
e
d
atas
et
is
d
iv
id
ed
i
n
to
3
class
es:
n
eu
tr
al,
s
atis
f
ied
,
an
d
u
n
s
atis
f
ied
.
T
h
e
r
esu
lt
o
b
tain
ed
f
r
o
m
th
is
ex
p
e
r
im
en
t
is
9
8
.
3
9
%
f
o
r
n
e
u
tr
al,
9
8
.
5
%
f
o
r
n
o
t
s
atis
f
ied
,
an
d
9
7
.
0
5
% f
o
r
s
atis
f
ied
.
San
d
ee
p
d
ev
elo
p
ed
a
m
o
d
el
to
d
etec
t
cu
s
to
m
er
f
ac
ial
ex
p
r
e
s
s
io
n
b
y
u
s
in
g
n
eu
r
al
n
etwo
r
k
[
2
6
]
o
n
a
VGG
-
1
6
ar
ch
itectu
r
e.
T
h
e
V
GG
-
1
6
ar
ch
itectu
r
e
[
2
7
]
ca
n
p
r
o
ce
s
s
lar
g
e
-
s
ca
le
im
ag
e
r
ec
o
g
n
itio
n
a
n
d
ex
t
r
ac
t
f
ea
tu
r
es.
S.
K.
R
am
an
i
[
2
6
]
,
th
e
C
NN
n
etwo
r
k
is
co
n
s
id
er
ed
as
th
e
b
aselin
e
o
f
th
e
s
y
s
t
em
an
d
it
u
s
es
th
e
Go
o
g
le
clo
u
d
p
latf
o
r
m
v
is
io
n
ap
p
licatio
n
p
r
o
g
r
am
m
in
g
i
n
te
r
f
ac
e
(
API
)
f
o
r
f
ac
ial
ex
p
r
ess
i
o
n
d
etec
tio
n
wh
ic
h
co
u
ld
also
h
elp
to
m
in
im
ize
th
e
co
m
p
u
tatio
n
al
tim
e
f
o
r
p
r
o
ce
s
s
in
g
th
e
im
ag
es.
T
o
ca
p
tu
r
e
th
e
co
r
r
esp
o
n
d
i
n
g
ex
p
r
ess
io
n
a
ca
m
er
a
is
attac
h
ed
with
R
asb
er
r
y
Pi
wh
ich
r
e
co
r
d
s
th
e
cu
s
to
m
e
r
'
s
f
ac
ial
ex
p
r
ess
io
n
ev
er
y
f
iv
e
s
ec
o
n
d
s
.
T
h
e
in
p
u
t
im
ag
es
a
r
e
u
s
ed
to
tr
ain
th
e
v
is
io
n
AP
I
an
d
C
NN
m
o
d
el
to
class
if
y
im
ag
es
in
to
s
ev
en
ex
p
r
ess
io
n
ca
teg
o
r
ies.
T
h
e
s
o
f
tm
ax
f
u
n
ctio
n
is
wo
r
k
ed
as
a
n
o
u
tp
u
t
lay
er
in
t
h
is
n
etwo
r
k
.
Af
ter
tr
ain
in
g
th
e
n
etwo
r
k
,
r
esp
o
n
s
es
o
f
th
e
m
o
d
el
will
b
e
s
to
r
ed
in
an
ar
r
ay
.
Sin
ce
weig
h
t
is
co
n
s
id
er
ed
as
a
tim
e
-
d
ep
en
d
en
t
f
u
n
ctio
n
,
t
h
e
f
in
al
weig
h
t
will
b
e
u
s
ed
as
th
e
f
in
al
s
co
r
e
f
o
r
th
e
f
ee
d
b
ac
k
.
T
h
e
s
co
r
e
is
th
en
n
o
r
m
alize
d
in
a
r
an
g
e
o
f
-
1
to
1
.
I
f
th
e
v
alu
e
is
o
b
tain
ed
as
1
th
en
th
e
r
esu
lt
is
co
n
s
id
er
ed
as
h
ap
p
y
,
if
-
1
th
en
th
e
p
er
s
o
n
n
o
t
ha
p
p
y
with
t
h
e
s
er
v
ice
p
r
o
v
id
ed
.
On
th
e
o
th
er
h
an
d
,
if
0
is
t
h
en
co
n
s
id
er
ed
n
e
u
tr
al.
T
o
o
b
s
er
v
e
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el,
T
h
e
M&
M
in
itiativ
e
(
MMI
)
[
2
8
]
d
atab
ase
co
n
tain
in
g
a
s
er
ies
o
f
v
id
e
o
s
o
f
p
eo
p
le
s
h
o
win
g
em
o
tio
n
s
is
u
s
ed
.
2
0
im
ag
es
ar
e
s
ele
cted
f
r
o
m
ea
ch
v
i
d
eo
b
y
ch
an
g
in
g
th
e
tim
ef
r
am
e
to
a
s
in
g
le
im
ag
e
b
y
r
ep
ea
tin
g
o
v
er
th
e
v
ar
ian
ce
s
b
y
ap
p
ly
in
g
t
h
e
h
ig
h
est
f
ilter
f
r
o
m
r
ed
u
cin
g
th
e
f
ilter
s
ize
u
n
til
2
0
f
r
am
es
ar
e
g
en
er
ated
.
A
v
er
ag
e
ac
cu
r
ac
y
o
f
8
6
.
6
%
is
o
b
tain
ed
.
2
0
m
o
s
t
r
ep
r
esen
tativ
e
im
ag
e
s
ar
e
s
elec
ted
b
y
m
in
im
al
ch
an
g
es
in
t
h
e
tim
ef
r
am
e
as
a
s
in
g
le
im
a
g
e
b
y
iter
atin
g
o
v
er
th
e
d
if
f
er
e
n
ce
s
u
s
in
g
a
m
a
x
im
u
m
f
ilter
with
d
ec
r
ea
s
in
g
f
ilter
s
ize
u
n
til th
e
in
p
u
t im
a
g
e
f
r
a
m
es a
r
e
o
b
tain
ed
.
Lu
et
a
l
.
[
1
9
]
,
a
co
n
v
en
tio
n
a
l
co
m
p
u
ter
v
is
io
n
tech
n
o
lo
g
y
is
b
ein
g
u
s
ed
w
h
ich
is
n
o
t
ca
p
ab
le
o
f
d
ea
lin
g
with
er
r
o
n
e
o
u
s
in
p
u
t
s
an
d
p
ar
am
eter
s
.
I
n
th
e
s
y
s
tem
,
th
e
cu
s
to
m
er
p
r
e
f
er
en
ce
s
ar
e
ca
lcu
lated
b
y
m
atch
in
g
th
e
r
eg
io
n
o
f
in
ter
es
t
an
d
g
r
o
u
p
ed
th
em
in
to
s
im
ilar
r
eg
io
n
ex
p
r
ess
io
n
s
.
Du
e
to
th
e
lim
itatio
n
o
f
th
e
s
y
s
tem
,
it
o
n
ly
a
n
aly
ze
s
1
0
f
r
am
es
f
o
r
ea
ch
s
ec
o
n
d
o
f
3
0
f
r
am
es
-
p
er
-
s
ec
o
n
d
r
ec
o
r
d
in
g
wh
ich
p
r
o
d
u
ce
s
er
r
o
n
e
o
u
s
in
p
u
ts
in
th
is
s
tep
.
T
h
e
er
r
o
r
s
p
r
o
d
u
ce
d
in
th
is
s
tag
e
ca
n
n
o
t
b
e
r
eso
lv
ed
as
a
C
NN
n
etwo
r
k
.
L
e
an
d
Vea
[
2
2
]
,
Mo
u
la
y
et
a
l
.
[
2
4
]
,
th
e
s
y
s
tem
s
ar
e
d
e
v
elo
p
ed
u
s
in
g
a
m
u
lticlas
s
class
if
ier
.
Sin
ce
th
e
s
y
s
tem
s
ar
e
d
ev
elo
p
e
d
u
s
in
g
a
m
u
lticlas
s
class
if
ier
,
it
r
e
q
u
ir
es
m
o
r
e
co
m
p
u
tatio
n
al
tim
e
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
Nak
an
o
an
d
Kato
[
2
3
]
,
a
PC
H
clas
s
if
ier
is
in
tr
o
d
u
ce
d
in
th
e
tr
a
in
in
g
p
r
o
ce
s
s
with
a
lo
w
ac
cu
r
ac
y
r
ate.
T
h
e
m
o
d
el
in
[
2
6
]
ca
n
p
r
o
ce
s
s
a
lar
g
e
s
ca
le
o
f
d
ata
b
y
u
s
in
g
3
r
d
p
a
r
ty
API
wh
ich
is
n
o
t
cu
s
to
m
izab
le
an
d
n
ee
d
s
to
p
ay
f
o
r
ea
c
h
tim
e
tr
ain
in
g
.
3.
RE
S
E
ARCH
M
E
T
H
O
D
S
3
.
1
.
Da
t
a
c
o
llect
io
n a
nd
p
re
-
p
ro
ce
s
s
ing
I
m
ag
es
f
o
r
th
r
ee
ty
p
es
o
f
f
ac
i
al
ex
p
r
ess
io
n
s
(
n
eu
tr
al,
h
ap
p
y
an
d
s
ad
)
ar
e
co
llected
in
r
ea
l
-
tim
e.
T
h
e
m
ain
ca
m
er
a
is
p
lace
d
i
n
f
r
o
n
t
o
f
th
e
s
u
b
jects
to
e
n
ab
le
t
h
e
r
ea
l
-
tim
e
r
ec
o
r
d
in
g
o
f
th
e
f
a
cial
ex
p
r
ess
io
n
s
o
f
th
e
h
u
m
an
f
ac
e
wh
ile
th
ey
ar
e
o
b
s
er
v
in
g
a
s
p
ec
if
ic
p
r
o
d
u
ct.
A
d
ig
ital c
am
er
a
is
aim
ed
d
ir
ec
tly
at
th
e
s
u
b
ject’
s
f
ac
e
to
o
b
s
er
v
e
th
eir
ex
p
r
ess
io
n
.
T
h
e
ca
m
er
a
is
p
air
ed
with
a
l
in
u
x
-
b
ased
o
p
er
atin
g
s
y
s
tem
with
8
GB
R
A
M
an
d
I
n
tel
i5
Pro
ce
s
s
o
r
.
T
h
e
im
ag
e
s
eq
u
e
n
ce
s
ar
e
r
ec
o
r
d
ed
3
0
f
r
am
es
p
er
s
ec
o
n
d
.
A
t
o
tal
n
u
m
b
er
o
f
5
3
s
u
b
jects
p
ar
ticip
ated
i
n
th
e
ex
p
er
im
en
t
an
d
r
ec
o
r
d
ed
th
ei
r
th
r
ee
ty
p
es
o
f
f
ac
ial
e
x
p
r
ess
io
n
s
.
E
ac
h
v
id
eo
is
r
ec
o
r
d
e
d
b
etwe
en
1
0
-
1
2
s
ec
.
T
h
en
,
a
v
id
eo
c
o
n
tain
in
g
a
s
p
ec
if
ic
p
r
o
d
u
ct
s
lid
esh
o
w
is
s
h
o
wn
to
th
e
s
u
b
jects
as sh
o
wn
in
Fig
u
r
e
1
.
T
h
e
s
u
b
ject’
s
s
p
o
n
tan
eo
u
s
ex
p
r
ess
io
n
is
ca
p
tu
r
ed
wh
ile
th
ey
o
b
s
er
v
e
d
th
e
p
r
o
d
u
cts.
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
C
u
s
to
mer’
s
s
p
o
n
ta
n
eo
u
s
fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
(
Go
la
m
Mo
r
s
h
ed
)
1439
Fig
u
r
e
1
.
Data
g
at
h
er
in
g
s
etu
p
T
h
e
wh
o
le
ex
p
er
im
en
t
was
co
n
d
u
cted
with
m
e
d
iu
m
illu
m
in
atio
n
an
d
p
o
s
ed
o
f
th
e
f
r
o
n
tal
p
r
o
f
ile
s
o
th
at
th
e
s
y
s
tem
ca
n
b
e
ev
alu
at
ed
in
d
if
f
e
r
e
n
t
co
n
d
itio
n
s
.
I
n
t
h
is
ex
p
er
im
en
t,
o
n
ly
s
p
o
n
tan
e
o
u
s
ex
p
r
ess
io
n
was
r
ec
o
r
d
e
d
wh
ich
r
e
f
er
s
to
lo
w
-
in
ten
s
ity
f
ac
ial
ex
p
r
ess
io
n
s
.
Ou
r
p
r
o
p
o
s
ed
m
o
d
el
is
d
ev
elo
p
ed
u
s
in
g
two
d
if
f
er
en
t
class
if
ier
s
wh
ich
ar
e
SVM
[
2
9
]
an
d
So
f
t
m
ax
[
3
0
]
.
T
h
ese
m
o
d
els
ar
e
tr
ain
ed
u
s
in
g
th
e
FER2
0
1
3
d
ataset
an
d
co
llected
r
ea
l
-
tim
e
f
ac
ial
ex
p
r
ess
io
n
im
a
g
es
in
h
ap
p
y
,
s
ad
,
a
n
d
n
e
u
r
al
ca
te
g
o
r
ies
as
s
h
o
wn
in
Fig
u
r
e
s
2
an
d
3
.
T
h
e
FER2
0
1
3
d
ataset
co
n
s
is
ts
o
f
b
lu
r
r
y
im
ag
es
an
d
s
o
m
e
o
f
th
e
im
a
g
es
ar
e
n
o
t
alig
n
ed
p
r
o
p
er
l
y
[
3
1
]
.
Fig
u
r
e
2
.
Data
s
et
ex
am
p
les
–
s
ad
f
ac
e
Fig
u
r
e
3
.
Data
s
et
ex
am
p
les
–
h
ap
p
y
f
ace
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.
22
,
No
.
3
,
J
u
n
e
2
0
2
1
:
1
4
3
6
-
1
4
4
5
1440
C
NN
is
o
n
e
o
f
th
e
class
o
f
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
wh
ich
h
av
e
p
r
o
v
e
n
an
e
f
f
ec
tiv
e
an
d
m
o
r
e
ef
f
icien
t
ter
m
in
th
e
ar
ea
o
f
i
m
ag
e
r
ec
o
g
n
itio
n
a
n
d
class
if
i
ca
tio
n
.
I
n
o
u
r
ex
p
er
im
en
t,
we
p
r
o
p
o
s
e
two
m
o
d
els
as
C
NN
-
SV
M
(
m
o
d
el
A)
an
d
C
NN
-
So
f
tm
ax
(
m
o
d
el
B
)
as
s
h
o
wn
in
Fig
u
r
e
s
4
an
d
5
.
B
o
th
m
o
d
els
co
n
s
is
t
o
f
a
co
n
v
o
lu
tio
n
al
lay
er
,
p
o
o
lin
g
lay
er
,
r
ec
tifie
d
-
lin
er
f
u
lly
(
also
k
n
o
wn
as
R
eL
u
lay
er
)
,
c
o
n
n
ec
ted
la
y
er
,
an
d
class
if
icatio
n
lay
er
[
3
2
]
.
T
h
e
o
n
ly
d
if
f
er
e
n
ce
in
th
e
class
if
icatio
n
lay
er
s
wh
er
e
m
o
d
el
A
u
s
es
SVM
an
d
m
o
d
el
B
u
s
es So
f
tMa
x
as th
e
class
if
ier
.
Fig
u
r
e
4.
C
NN
-
SVM
(
m
o
d
el
A)
Fig
u
r
e
5
.
C
NN
-
So
f
tMa
x
(
m
o
d
el
B
)
C
o
n
v
o
lu
tio
n
al
lay
e
r
s
ar
e
r
esp
o
n
s
ib
le
to
ca
lcu
late
th
e
co
n
v
o
lu
tio
n
s
p
atially
o
f
a
g
r
o
u
p
o
f
n
eu
r
o
n
s
alo
n
g
with
f
ilter
s
th
at
ca
n
b
e
lear
n
ab
le
[
3
3
]
.
T
h
e
f
ea
tu
r
e
m
ap
s
ar
e
cr
ea
ted
b
y
ca
lcu
latin
g
th
e
d
o
t
p
r
o
d
u
ct
o
f
th
e
f
ilter
weig
h
ts
alo
n
g
with
th
e
in
p
u
t
n
eu
r
o
n
s
th
at
ar
e
r
esp
o
n
s
ib
le
to
p
ass
to
t
h
e
n
ex
t
c
o
r
r
esp
o
n
d
in
g
lay
e
r
f
o
r
co
m
p
u
tatio
n
s
[
3
4
]
.
E
ac
h
la
y
e
r
is
f
o
r
m
ed
b
ased
o
n
k
er
n
el
s
ize,
d
ep
th
,
s
tr
id
e,
a
n
d
p
ad
d
in
g
.
T
h
e
k
er
n
el
s
ize
r
ef
er
s
to
th
e
s
p
atial
s
ize
o
f
th
e
co
n
v
o
lu
tio
n
al
f
ilter
.
Dep
th
r
e
p
r
esen
ts
th
e
d
ep
th
o
f
th
e
f
ilter
;
s
tr
id
e
r
ef
er
s
to
th
e
s
ize
o
f
th
e
s
lid
es
ac
r
o
s
s
a
lo
c
al
r
eg
io
n
o
f
t
h
e
in
p
u
t
d
ata
a
n
d
p
ad
d
in
g
r
eg
u
lates
th
e
e
d
g
ed
p
ad
d
e
d
v
alu
e
.
T
h
e
p
o
o
lin
g
lay
e
r
is
r
esp
o
n
s
ib
le
to
d
ec
r
ea
s
e
th
e
s
p
atial
d
im
en
s
io
n
o
f
th
e
in
p
u
t
[
3
5
]
.
T
h
e
p
r
o
ce
s
s
i
s
o
b
tain
ed
b
y
p
er
f
o
r
m
in
g
a
s
lid
in
g
win
d
o
w
o
n
a
s
in
g
le
d
ep
th
s
lice.
T
h
is
tech
n
iq
u
e
is
u
s
ed
to
co
m
p
u
te
th
e
av
er
ag
e
o
r
m
ax
im
u
m
v
al
u
e
o
f
th
e
d
im
e
n
s
io
n
al
lo
ca
tio
n
.
T
h
e
Re
L
u
l
ay
er
is
r
esp
o
n
s
ib
le
to
r
em
o
v
e
th
e
n
eg
ativ
e
v
alu
e
f
r
o
m
th
e
i
n
p
u
t
d
ata
an
d
r
ep
lac
e
b
y
th
e
v
alu
e
‘
0
’
[
3
6
]
.
I
n
t
h
e
f
u
lly
c
o
n
n
ec
ted
lay
e
r
,
th
e
n
e
u
r
o
n
s
'
s
p
atial
s
ize
is
d
ec
r
ea
s
ed
to
1
*
1
m
atr
ix
.
T
h
e
class
if
icatio
n
lay
er
is
r
esp
o
n
s
ib
le
to
m
ea
s
u
r
e
th
e
p
r
e
d
icted
p
r
o
b
ab
ilit
ies
an
d
lab
el
th
em
in
to
r
esp
ec
tiv
e
class
es.
Ho
wev
er
,
o
u
r
m
o
d
el
A
u
s
es
a
m
u
lticla
s
s
cla
s
s
if
ier
S
VM
in
s
tead
o
f
th
e
s
o
f
tm
ax
f
u
n
ctio
n
d
ev
el
o
p
ed
b
y
Vap
n
ik
[
3
7
]
.
T
h
e
o
u
tp
u
t
o
f
th
e
r
esu
lt
will
b
e
tr
an
s
lated
in
to
ca
s
e
y
∈
{−
1
,
+
1
)
as in
[
3
8
]
.
min
1
‖
‖
2
2
+
∁
∑
ma
x
(
0
,
1
−
′
(
+
)
)
2
=
1
(
1
)
(
)
=
∑
=
1
(
2
)
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
C
u
s
to
mer’
s
s
p
o
n
ta
n
eo
u
s
fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
(
Go
la
m
Mo
r
s
h
ed
)
1441
I
n
m
o
d
el
B
,
th
e
s
o
f
tm
ax
f
u
n
ct
io
n
is
im
p
lem
en
ted
to
class
if
y
th
e
m
u
ltip
le
class
es.
I
n
th
is
e
x
p
er
im
en
t,
th
e
s
o
f
tm
ax
f
u
n
ctio
n
is
u
s
ed
b
ec
au
s
e
th
e
s
o
f
tm
ax
is
wel
l
k
n
o
wn
n
o
r
m
alize
d
f
u
n
ctio
n
to
r
e
ac
t
with
th
e
b
lu
r
r
y
im
ag
es
r
ath
er
th
an
o
th
e
r
s
tan
d
ar
d
n
o
r
m
ali
za
tio
n
.
I
n
(
2
)
r
e
f
er
s
th
e
So
f
tMa
x
f
u
n
ctio
n
[
3
9
]
.
On
ce
th
e
C
NN
n
etwo
r
k
is
tr
ain
e
d
,
o
u
r
p
r
o
p
o
s
ed
s
y
s
tem
is
test
ed
with
th
e
tr
ain
ed
C
NN
n
etwo
r
k
an
d
th
e
r
esu
lts
ar
e
as
s
h
o
wn
in
Fig
u
r
e
s
6
an
d
7
.
Fig
u
r
e
6
.
R
ea
l
-
tim
e
f
ac
ial
ex
p
r
ess
io
n
ex
p
er
im
en
t
-
s
ad
f
ac
e
Fig
u
r
e
7
.
R
ea
l
-
tim
e
f
ac
ial
ex
p
r
ess
io
n
ex
p
er
im
en
t
-
h
ap
p
y
f
ac
e
3.
2
.
F
r
a
m
ewo
r
k
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
is
d
ev
e
lo
p
ed
u
s
in
g
th
e
C
NN
n
etwo
r
k
an
d
tr
ain
ed
u
s
in
g
o
n
l
y
h
ap
p
y
,
s
ad
,
an
d
n
eu
tr
al
im
ag
es.
T
h
e
f
ir
s
t step
is
im
ag
e
p
r
e
-
p
r
o
ce
s
s
in
g
wh
ich
i
n
v
o
lv
es d
ata
au
g
m
en
tatio
n
an
d
n
o
r
m
a
lizatio
n
a
s
s
h
o
wn
in
Fig
u
r
e
8
.
T
h
e
d
ata
au
g
m
en
tatio
n
in
v
o
lv
es
s
ca
lin
g
an
d
r
o
tatio
n
a
n
d
t
h
e
n
o
r
m
al
izatio
n
p
r
o
ce
s
s
is
to
r
ed
u
ce
th
e
illu
m
in
atio
n
an
d
p
o
s
e
o
f
th
e
in
p
u
t
im
ag
e.
T
h
e
f
ac
e
is
d
etec
ted
u
s
in
g
Vio
la
-
J
o
n
es
[
2
0
]
al
g
o
r
ith
m
d
etec
tio
n
th
en
cr
o
p
p
ed
an
d
n
o
r
m
alize
d
as
64
x
6
4
p
ix
els.
T
h
en
th
e
f
ac
ial
p
ar
ts
ar
e
ex
tr
a
cted
an
d
r
esized
as
3
2
x
6
4
p
i
x
els
.
Af
ter
th
e
p
r
ep
r
o
ce
s
s
in
g
s
tep
,
th
e
im
ag
es
ar
e
f
e
d
to
th
e
C
NN
n
etwo
r
k
t
o
class
if
y
th
e
im
a
g
es
in
to
th
r
ee
ca
teg
o
r
ies
as
h
ap
p
y
,
s
ad
,
an
d
n
e
u
tr
al
.
T
h
e
m
ai
n
co
n
ce
p
t
o
f
C
NN
is
th
at
it
ex
tr
ac
t
s
f
ac
ial
in
f
o
r
m
atio
n
th
r
o
u
g
h
its
weig
h
t
an
d
f
ield
s
h
ar
in
g
p
r
o
ce
s
s
.
T
h
e
n
u
m
b
er
o
f
tr
ain
ed
p
ar
am
eter
s
is
r
ed
u
ce
d
in
th
e
s
h
a
r
in
g
p
r
o
ce
s
s
an
d
th
e
p
r
o
p
a
g
atio
n
o
f
im
ag
e
in
f
o
r
m
atio
n
th
r
o
u
g
h
t
h
e
lay
er
s
ar
e
b
ein
g
ca
lcu
lated
b
y
co
n
v
o
l
u
tio
n
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
I
n
th
e
l
ast
s
tag
e,
th
e
b
ac
k
-
p
r
o
p
ag
ati
o
n
alg
o
r
ith
m
is
ap
p
lied
to
t
h
o
s
e
im
ag
es
w
h
ich
p
r
o
v
id
e
a
m
ax
im
u
m
v
alu
e
o
f
1
to
d
etec
t
co
r
r
ec
t
f
ac
ial
ex
p
r
ess
io
n
.
W
e
to
o
k
th
e
m
ax
im
u
m
n
u
m
b
e
r
1
f
o
r
th
e
ex
p
ec
ted
ex
p
r
ess
io
n
an
d
0
f
o
r
o
th
er
ex
p
r
ess
io
n
s
.
T
h
ese
v
alu
es
ar
e
ca
lcu
lated
in
th
e
f
in
al
lay
er
o
f
th
e
s
o
f
tm
a
x
f
u
n
ctio
n
.
I
n
th
e
test
in
g
p
h
as
e,
ea
ch
in
p
u
t
f
r
am
e
will
b
e
m
atch
ed
with
th
e
g
iv
e
n
tr
ain
ed
C
NN
m
o
d
el
to
r
ec
o
g
n
ize
th
e
f
ac
ial
ex
p
r
ess
io
n
o
f
th
e
cu
s
to
m
er
wh
ile
o
b
s
er
v
in
g
th
e
p
r
o
d
u
ct.
Fig
u
r
e
8
.
Ov
e
r
all
f
r
am
ewo
r
k
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.
22
,
No
.
3
,
J
u
n
e
2
0
2
1
:
1
4
3
6
-
1
4
4
5
1442
4.
RE
SU
L
T
S AN
D
AN
AL
Y
SI
S
T
h
e
f
ac
e
lan
d
m
a
r
k
s
an
d
h
is
to
g
r
am
o
f
o
r
ien
ted
g
r
ad
ien
ts
(
H
OG
)
f
ea
tu
r
es
ar
e
ex
tr
ac
ted
an
d
f
ed
in
to
a
m
u
lti
-
class
SVM
an
d
So
f
tm
ax
class
if
ier
.
I
n
o
u
r
ex
p
e
r
im
en
t,
we
im
p
lem
en
t
f
ac
ial
lan
d
m
a
r
k
an
d
HOG
f
ea
tu
r
e
lay
er
s
in
b
o
th
m
o
d
els.
T
ab
le
1
s
h
o
ws
th
e
clas
s
if
icatio
n
r
es
u
lts
f
o
r
m
o
d
el
A
an
d
m
o
d
el
B
.
B
ased
o
n
T
ab
le
1
,
m
o
d
el
B
g
iv
es
a
h
ig
h
e
r
ac
cu
r
ac
y
r
esu
lt
i
n
co
m
p
ar
is
o
n
to
m
o
d
el
A
i
n
all
e
x
p
er
im
e
n
ts
.
T
h
e
tr
ain
in
g
tim
e
f
o
r
th
e
o
b
tain
ed
r
esu
lt
is
5
0
2
1
s
ec
o
n
d
s
o
n
c
o
r
e
i5
a
n
d
L
in
u
x
-
b
ased
o
p
er
atin
g
s
y
s
tem
.
Fig
u
r
e
9
s
h
o
ws
th
e
ac
cu
r
ac
y
r
esu
lt
o
f
m
o
d
el
A
a
n
d
m
o
d
el
B
.
T
h
e
x
-
ax
is
r
ep
r
esen
ts
th
e
ex
p
er
im
en
t
t
y
p
e
an
d
th
e
y
-
ax
is
r
ep
r
esen
ts
th
e
ac
cu
r
ac
y
r
ate
o
f
m
o
d
els
A
an
d
B
.
B
ased
o
n
T
ab
le
1
,
m
o
d
el
A'
s
ac
cu
r
ac
y
is
s
lig
h
tl
y
lo
wer
co
m
p
a
r
ed
to
m
o
d
el
B
b
y
4
.
5
%.
T
h
is
r
esu
l
t
d
if
f
er
s
b
ec
a
u
s
e
in
C
NN
o
v
er
f
itti
n
g
is
s
u
e
an
d
f
u
lly
c
o
n
n
ec
t
lay
er
u
n
ab
le
to
ex
tr
ac
t
th
e
in
f
o
r
m
atio
n
f
r
o
m
p
r
ev
io
u
s
lay
e
r
s
[
3
0
]
.
B
esid
es,
a
d
d
in
g
m
o
r
e
f
ea
tu
r
es
f
o
r
ea
ch
m
o
d
el
ca
n
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
r
ate.
I
n
o
u
r
m
o
d
el
B
,
o
n
e
ex
tr
a
f
u
lly
co
n
n
ec
te
d
lay
er
was
ad
d
ed
to
im
p
r
o
v
e
th
e
ac
cu
r
ac
y
lev
el.
B
ased
o
n
Fig
u
r
e
9
,
w
e
ch
o
o
s
e
m
o
d
el
B
in
t
h
e
n
e
x
t
ex
p
e
r
im
e
n
t
s
in
ce
it
p
r
o
v
id
es
m
o
r
e
ac
cu
r
ac
y
th
an
m
o
d
el
A.
T
h
e
ac
cu
r
ac
y
p
er
ce
n
ta
g
e
is
ca
lcu
lated
b
y
u
s
in
g
(
3
)
.
Fr
o
m
T
ab
le
2
,
it
ca
n
b
e
co
n
clu
d
ed
th
at
th
e
h
ap
p
y
ex
p
r
ess
io
n
h
as
t
h
e
h
i
g
h
est
s
u
cc
ess
r
ate
wh
ich
is
8
9
%
an
d
t
h
e
n
eu
t
r
al
ex
p
r
ess
io
n
wh
ich
s
u
cc
ess
r
ate
is
8
4
%
an
d
th
e
lo
west su
cc
ess
r
ate
is
th
e
s
ad
ex
p
r
es
s
io
n
wh
ich
is
7
5
%.
Fig
u
r
e
9
.
C
o
m
p
a
r
is
o
n
b
etwe
e
n
m
o
d
el
A
a
n
d
m
o
d
el
B
T
ab
le
1
.
C
lass
if
icatio
n
r
esu
lt f
o
r
m
o
d
el
A
an
d
m
o
d
el
B
Ex
p
e
r
i
m
e
n
t
s
M
o
d
e
l
A
M
o
d
e
l
B
C
N
N
7
2
.
4
%
7
3
.
5
%
C
N
N
+
F
a
c
i
a
l
La
n
d
mar
k
s
7
3
.
5
%
7
4
.
4
%
C
N
N
+
F
a
c
i
a
l
La
n
d
mar
k
s
+
H
O
G
6
8
.
7
%
7
3
.
2
%
C
N
N
+
F
a
c
i
a
l
La
n
d
mar
k
s
+
H
O
G
+
S
l
i
d
i
n
g
W
i
n
d
o
w
s
7
1
.
5
%
7
5
.
5
%
=
+
+
+
+
(
3
)
T
ab
le
2
.
C
o
n
f
u
s
io
n
m
atr
ix
f
o
r
h
ap
p
y
,
s
ad
an
d
n
eu
tr
al
e
x
p
r
ess
io
n
F
a
c
i
a
l
Ex
p
r
e
ss
i
o
n
P
r
e
d
i
c
t
e
d
H
a
p
p
y
A
c
t
u
a
l
Tr
u
e
N
e
g
a
t
i
v
e
-
6
F
a
l
se
P
o
s
i
t
i
v
e
-
4
F
a
l
se
N
e
g
a
t
i
v
e
-
2
Tr
u
e
P
o
s
i
t
i
v
e
-
41
S
a
d
A
c
t
u
a
l
Tr
u
e
N
e
g
a
t
i
v
e
-
12
F
a
l
se
P
o
s
i
t
i
v
e
-
3
F
a
l
se
N
e
g
a
t
i
v
e
-
10
Tr
u
e
P
o
s
i
t
i
v
e
-
28
N
e
u
t
r
a
l
A
c
t
u
a
l
Tr
u
e
N
e
g
a
t
i
v
e
-
8
F
a
l
se
P
o
s
i
t
i
v
e
-
5
F
a
l
se
N
e
g
a
t
i
v
e
-
3
Tr
u
e
P
o
s
i
t
i
v
e
-
37
T
ab
le
3
s
h
o
ws
th
e
ac
cu
r
ac
y
lev
el
o
f
th
e
cu
s
to
m
er
'
s
th
r
ee
k
in
d
s
o
f
f
ac
ial
ex
p
r
ess
io
n
s
f
o
r
d
if
f
er
en
t
p
r
o
p
o
s
ed
m
o
d
els.
B
ased
o
n
t
h
e
tab
le
t
h
e
a
v
ailab
le
class
if
ier
is
b
ein
g
u
s
ed
to
class
if
y
th
e
cu
s
to
m
er
f
ac
ial
ex
p
r
ess
io
n
in
to
d
if
f
er
e
n
t
ca
teg
o
r
ies.
L
u
et
a
l.
[
1
9
]
,
cu
s
t
o
m
er
r
ea
ctio
n
is
b
ein
g
o
b
s
er
v
ed
to
ca
p
tu
r
e
th
e
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
C
u
s
to
mer’
s
s
p
o
n
ta
n
eo
u
s
fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
(
Go
la
m
Mo
r
s
h
ed
)
1443
cu
s
to
m
er
p
r
ef
e
r
en
ce
s
an
d
ac
h
iev
ed
s
u
cc
ess
r
ates
f
o
r
p
o
s
itiv
e,
n
eu
tr
al,
an
d
n
eg
ativ
e
a
r
e
8
8
%,
8
1
%,
an
d
6
8
%
r
esp
ec
tiv
ely
.
Nak
an
o
a
n
d
Kat
o
[
2
3
]
,
th
e
m
o
d
el
is
d
ev
elo
p
e
d
u
s
in
g
th
e
C
NN
n
etwo
r
k
an
d
a
ch
iev
ed
7
8
.
8
%
ac
cu
r
ac
y
f
o
r
c
u
s
to
m
er
e
x
p
ec
t
atio
n
an
d
6
9
.
7
%
f
o
r
s
atis
f
ac
tio
n
.
I
n
th
eir
m
o
d
el,
th
e
PC
H
class
if
ier
was
ap
p
lied
in
th
eir
tr
ain
in
g
p
r
o
ce
s
s
to
o
b
tain
th
e
r
esu
lt.
L
e
an
d
Vea
[
2
2
]
,
Kin
ec
t
b
ased
c
u
s
to
m
ized
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
e
is
u
s
ed
i
n
th
eir
m
o
d
el.
VGG
-
1
6
ar
c
h
itectu
r
e
is
b
ein
g
u
s
ed
with
a
C
NN
-
b
ased
f
ea
tu
r
e
ex
tr
ac
tio
n
tech
n
iq
u
e
an
d
ac
h
iev
e
d
an
av
er
ag
e
s
u
cc
ess
r
ate
o
f
8
6
.
6
%
[
2
6
]
.
I
n
o
u
r
p
r
o
p
o
s
ed
m
o
d
el,
a
f
u
lly
m
u
lti
-
co
n
n
ec
t
lay
er
is
ap
p
lied
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
with
s
o
f
tm
ax
f
u
n
c
t
io
n
wh
ich
h
el
p
s
th
e
m
o
d
el
t
o
o
b
tain
its
h
ig
h
est
ac
cu
r
ac
y
r
esu
lt
f
r
o
m
th
e
g
i
v
e
n
in
p
u
t
an
d
p
r
o
v
id
e
a
s
u
cc
ess
r
ate
o
f
8
9
%
as
h
ap
p
y
,
7
5
%
as
s
ad
,
an
d
8
4
%
as
n
eu
tr
al.
B
ased
o
n
th
ese
r
esu
lts
,
it is
p
r
o
v
en
t
h
at
o
u
r
p
r
o
p
o
s
ed
m
o
d
el
h
as a
b
etter
p
er
f
o
r
m
an
ce
co
m
p
ar
e
d
to
th
e
ex
is
tin
g
s
tu
d
ies m
en
tio
n
ed
in
T
ab
le
3
.
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
b
etwe
en
o
th
er
m
o
d
els
S
y
st
e
m
C
l
a
s
si
f
i
e
r
F
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
A
c
c
u
r
a
c
y
(
%)
A
v
i
d
e
o
-
b
a
se
d
a
u
t
o
ma
t
e
d
r
e
c
o
m
me
n
d
e
r
(
V
A
R
)
sy
st
e
m f
o
r
g
a
r
me
n
t
s
[
1
9
]
.
S
V
M
Lo
c
a
l
b
i
n
a
r
y
p
a
t
t
e
r
n
(
LB
P
)
H
a
p
p
y
-
8
8
%,
n
e
u
t
r
a
l
-
8
1
%
a
n
d
s
a
d
6
8
%
P
o
t
e
n
t
i
a
l
i
t
y
o
f
3
D
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s t
o
e
st
i
ma
t
e
c
u
st
o
mer
e
x
p
e
c
t
a
t
i
o
n
a
n
d
s
a
t
i
sf
a
c
t
i
o
n
[
2
3
]
.
P
e
r
c
e
p
t
i
o
n
c
o
r
r
e
c
t
i
o
n
h
y
p
o
t
h
e
s
e
s
(
P
C
H
)
C
N
N
B
a
se
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
A
v
e
r
a
g
e
a
c
c
u
r
a
c
y
l
e
v
e
l
6
9
.
7
%
A
c
u
s
t
o
mer
e
mo
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
t
h
r
o
u
g
h
f
a
c
i
a
l
e
x
p
r
e
ss
i
o
n
u
s
i
n
g
k
i
n
e
c
t
se
n
s
o
r
s
v
1
a
n
d
v
2
[
2
2
]
B
a
g
g
i
n
g
,
J4
8
,
R
a
n
d
o
m
F
o
r
e
st
,
a
n
d
R
a
n
d
o
m
C
o
mm
i
t
t
e
e
K
i
n
e
c
t
b
a
se
d
c
u
s
t
o
mi
z
e
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
.
F
o
r
se
n
s
o
r
k
i
n
e
c
t
v
1
u
si
n
g
C
#
a
n
d
k
i
n
e
c
t
v
2
u
s
i
n
g
c
+
+
A
v
e
r
a
g
e
a
c
c
u
r
a
c
y
f
o
r
sen
s
o
r
V
1
B
a
g
g
i
n
g
-
9
0
.
4
5
%
J4
8
-
8
8
.
5
4
%
R
a
n
d
o
m f
o
r
e
s
t
-
9
1
.
8
7
%
R
a
n
d
o
m
c
o
mm
i
t
t
e
e
-
9
3
.
0
8
9
%
A
v
e
r
a
g
e
a
c
c
u
r
a
c
y
f
o
r
sen
s
o
r
V
2
B
a
g
g
i
n
g
-
8
3
.
0
5
%
J4
8
-
7
9
.
3
8
%
R
a
n
d
o
m f
o
r
e
s
t
-
8
5
.
3
2
%
R
a
n
d
o
m
c
o
mm
i
t
t
e
e
-
8
7
.
4
4
%
F
a
c
i
a
l
e
x
p
r
e
ss
i
o
n
d
e
t
e
c
t
i
o
n
u
s
i
n
g
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
c
u
st
o
mer
-
b
a
se
d
s
e
r
v
i
c
e
[
2
6
]
S
o
f
t
ma
x
C
N
N
b
a
se
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
.
A
v
e
r
a
g
e
a
c
c
u
r
a
c
y
r
a
t
e
8
6
.
6
%
P
r
o
p
o
se
d
m
e
t
h
o
d
S
o
f
t
ma
x
F
a
c
e
l
a
n
d
mar
k
,
H
O
G
a
n
d
sl
i
d
i
n
g
w
i
n
d
o
w
H
a
p
p
y
-
8
9
%,
n
e
u
t
r
a
l
-
8
4
%,
s
a
d
-
7
5
%
5.
CO
NCLU
SI
O
N
T
h
e
m
ain
p
u
r
p
o
s
e
o
f
th
is
r
e
s
ea
r
ch
is
to
r
ec
o
g
n
ize
h
u
m
a
n
ex
p
r
ess
io
n
in
co
n
s
u
m
e
r
s
c
ien
ce
an
d
co
m
p
ar
e
it
to
th
e
ex
is
tin
g
m
eth
o
d
s
in
th
e
f
ac
ial
ex
p
r
ess
io
n
f
ield
o
f
s
tu
d
y
.
C
u
r
r
en
tly
,
th
e
s
y
s
tem
is
lim
i
ted
to
r
ec
o
g
n
ize
o
n
ly
th
r
ee
k
in
d
s
o
f
f
ac
ial
ex
p
r
ess
io
n
s
as
h
ap
p
y
,
s
ad
,
an
d
n
eu
tr
al.
F
o
r
f
u
t
u
r
e
wo
r
k
,
o
u
r
s
tu
d
ies
will
b
e
ex
ten
d
ed
to
m
o
r
e
ex
p
r
ess
io
n
s
.
I
n
o
u
r
ex
p
er
im
en
t,
we
a
ch
iev
ed
8
9
%,
7
5
%,
an
d
8
4
%
ac
cu
r
ac
y
r
esu
lts
in
h
ap
p
y
,
s
ad
,
a
n
d
n
eu
tr
al
ex
p
r
ess
io
n
s
wh
ich
ca
n
b
e
im
p
r
o
v
ed
b
y
u
s
in
g
a
m
u
lti
-
class
class
if
ier
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
A
CK
NO
WL
E
DG
E
M
E
NT
S
T
h
e
au
t
h
o
r
s
f
u
lly
ac
k
n
o
wled
g
ed
Un
i
v
er
s
iti
Ma
lay
s
ia
Sar
a
wak
f
o
r
th
e
s
u
p
p
o
r
t
wh
ich
m
ak
es
th
is
im
p
o
r
tan
t stu
d
y
v
iab
le
a
n
d
ef
f
ec
tiv
e
.
RE
F
E
R
E
NC
E
S
[1
]
R.
Zh
i
,
X.
Hu
,
C.
Wan
g
,
a
n
d
S
.
Li
u
,
“
De
v
e
lo
p
m
e
n
t
o
f
a
d
irec
t
m
a
p
p
in
g
m
o
d
e
l
b
e
twe
e
n
h
e
d
o
n
ic
ra
ti
n
g
a
n
d
fa
c
ial
re
sp
o
n
se
s
b
y
d
y
n
a
m
ic
fa
c
ial
e
x
p
re
ss
io
n
re
p
re
se
n
tatio
n
,
”
Fo
o
d
Re
se
a
rc
h
In
ter
n
a
t
io
n
a
l
,
v
o
l.
1
3
7
,
No
v
.
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j.
fo
o
d
re
s.
2
0
2
0
.
1
0
9
4
1
1
.
[2
]
S
.
S
.
S
a
m
a
n
t
,
M
.
J
.
C
h
a
p
k
o
a
n
d
H.
S
e
o
,
“
P
re
d
ictin
g
c
o
n
su
m
e
r
li
k
in
g
a
n
d
p
re
fe
re
n
c
e
b
a
se
d
o
n
e
m
o
ti
o
n
a
l
re
sp
o
n
se
s
a
n
d
se
n
so
r
y
p
e
rc
e
p
ti
o
n
:
A
st
u
d
y
with
b
a
sic
tas
te
so
lu
ti
o
n
s
,”
Fo
o
d
Res
e
a
rc
h
In
ter
n
a
ti
o
n
a
l,
v
o
l.
1
0
0
,
p
p
.
3
2
5
-
3
3
4
,
Ju
l.
2
0
1
7
,
d
o
i:
1
0
.
1
0
1
6
/
j.
fo
o
d
re
s.
2
0
1
7
.
0
7
.
0
2
1
.
[3
]
S
.
G
.
M
o
o
re
,
“
Atti
tu
d
e
p
re
d
ict
a
b
il
it
y
a
n
d
h
e
l
p
fu
l
n
e
ss
in
o
n
l
in
e
re
v
iew
s
:
Th
e
ro
le
o
f
e
x
p
lai
n
e
d
a
c
ti
o
n
s
a
n
d
re
a
c
ti
o
n
s
,
”
J
o
u
rn
a
l
o
f
C
o
n
s
u
me
r R
e
se
a
rc
h
,
v
o
l.
4
2
,
n
o
.
1
,
p
p
.
3
0
-
44
,
2
0
1
5
,
d
o
i:
h
tt
p
s:/
/d
o
i.
o
rg
/
1
0
.
1
0
9
3
/j
c
r/
u
c
v
0
0
3
.
[4
]
M
.
D
.
R
o
c
k
lag
e
.
,
a
n
d
R.
H
.
F
a
z
io
,
“
T
h
e
e
n
h
a
n
c
in
g
v
e
rsu
s
b
a
c
k
firi
n
g
e
ffe
c
ts
o
f
p
o
siti
v
e
e
m
o
ti
o
n
in
c
o
n
su
m
e
r
re
v
ie
ws
,
”
J
o
u
rn
a
l
o
f
M
a
rk
e
ti
n
g
R
e
se
a
rc
h
,
v
o
l
.
57
,
n
o
.
2
,
p
p
.
3
3
2
-
3
5
2
,
F
e
b
.
2
0
2
0
,
d
o
i:
1
0
.
1
1
7
7
/0
0
2
2
2
4
3
7
1
9
8
9
2
5
9
4
.
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.
22
,
No
.
3
,
J
u
n
e
2
0
2
1
:
1
4
3
6
-
1
4
4
5
1444
[5
]
N.
Hu
ss
a
in
,
H.
Ujir,
I
.
Hip
i
n
y
a
n
d
J.
L.
M
in
o
i.
,
“
3
D
fa
c
ial
a
c
ti
o
n
u
n
i
ts
re
c
o
g
n
it
i
o
n
f
o
r
e
m
o
ti
o
n
a
l
e
x
p
re
ss
io
n
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
o
f
Rec
e
n
t
T
e
c
h
n
o
l
o
g
y
a
n
d
En
g
i
n
e
e
rin
g
,
v
o
l.
8
,
n
o
.
2
8
,
p
p
.
1
3
1
7
-
1
3
2
3
,
2
0
1
9
.
[6
]
L.
J.
Ha
rriso
n
-
Walk
e
r
,
“
T
h
e
e
ffe
c
t
o
f
c
o
n
su
m
e
r
e
m
o
ti
o
n
s
o
n
o
u
tc
o
m
e
b
e
h
a
v
i
o
rs
f
o
ll
o
wi
n
g
se
rv
ice
f
a
il
u
re
,”
J
o
u
rn
a
l
o
f
S
e
rv
ice
s M
a
rk
e
ti
n
g
,
v
o
l
.
33
,
n
o
.
3
,
p
p
.
2
8
5
–
3
0
2
,
2
0
1
9
,
d
o
i:
1
0
.
1
1
0
8
/JS
M
-
04
-
2
0
1
8
-
0
1
2
4
.
[7
]
H.
Ujir
a
n
d
M
.
S
p
a
n
n
,
“
S
u
rfa
c
e
n
o
rm
a
ls
with
m
o
d
u
lar
a
p
p
r
o
a
c
h
a
n
d
we
i
g
h
ted
v
o
t
in
g
sc
h
e
m
e
in
3
D
fa
c
ial
e
x
p
re
ss
io
n
c
las
sifica
ti
o
n
,
”
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
C
o
mp
u
ter
a
n
d
In
fo
rm
a
ti
o
n
T
e
c
h
n
o
lo
g
y
,
v
ol
.
3
,
n
o
.
5
,
p
p
.
9
0
9
-
9
1
8
,
S
e
p
.
2
0
1
4
.
[8
]
P
.
Tarn
o
ws
k
i
.,
M.
Ko
l
o
d
z
iej,
A
.
M
a
jk
o
ws
k
i
a
n
d
R
.
J.
Ra
k
.,
“
E
m
o
ti
o
n
re
c
o
g
n
it
io
n
u
si
n
g
fa
c
ial
e
x
p
re
ss
io
n
s
,”
i
n
ICCS
,
p
p
.
1
1
7
5
-
1
1
8
4
,
Ju
n
.
2
0
1
7
,
d
o
i:
1
0
.
1
0
1
6
/j
.
p
ro
c
s.
2
0
1
7
.
0
5
.
0
2
5
.
[9
]
J.
Yu
a
n
.,
S.
M
c
d
o
n
o
u
g
h
.
,
Q.
Y
o
u
.
,
a
n
d
J.
Lu
o
,
“
S
e
n
tr
ib
u
te:
ima
g
e
se
n
ti
m
e
n
t
a
n
a
ly
s
is f
r
o
m
a
m
id
-
lev
e
l
p
e
rsp
e
c
ti
v
e
,
”
i
n
Pr
o
c
e
e
d
in
g
s
o
f
t
h
e
S
e
c
o
n
d
I
n
te
rn
a
ti
o
n
a
l
W
o
rk
sh
o
p
o
n
Iss
u
e
s
o
f
S
e
n
ti
me
n
t
Disc
o
v
e
ry
a
n
d
Op
i
n
io
n
M
i
n
in
g
,
2
0
1
3
,
p
p
.
1
-
8
,
d
o
i:
1
0
.
1
1
4
5
/2
5
0
2
0
6
9
.
2
5
0
2
0
7
9
.
[1
0
]
G
.
Yo
lcu
.
,
I.
Oz
tel,
S.
K
a
z
a
n
,
C
.
Oz
a
n
d
F
.
Bu
n
y
a
k
,
“
De
e
p
lea
rn
in
g
-
b
a
se
d
fa
c
e
a
n
a
ly
sis
sy
ste
m
f
o
r
m
o
n
it
o
ri
n
g
c
u
sto
m
e
r
in
tere
st,”
J
o
u
rn
a
l
o
f
A
mb
ien
t
In
telli
g
e
n
c
e
a
n
d
H
u
ma
n
ize
d
Co
mp
u
ti
n
g
,
v
o
l.
11
,
n
o
.
1
,
p
p
.
237
-
2
4
8
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
2
6
5
2
-
0
1
9
-
0
1
3
1
0
-
5
.
[1
1
]
H.
Zh
a
n
g
.
,
A.
Jo
lfae
i
,
a
n
d
M
.
Ala
z
a
b
,
“
A
fa
c
e
e
m
o
ti
o
n
re
c
o
g
n
it
i
o
n
m
e
th
o
d
u
sin
g
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
r
a
l
n
e
two
r
k
a
n
d
ima
g
e
e
d
g
e
c
o
m
p
u
t
in
g
,
”
IEE
E
Acc
e
ss
,
v
o
l
.
7
,
p
p
.
1
5
9
0
8
1
-
1
5
9
0
8
9
,
Oc
t.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/ACCE
S
S
.
2
0
1
9
.
2
9
4
9
7
4
1
.
[1
2
]
P
.
M
.
A.
De
sm
e
t
,
“
F
a
c
e
s
o
f
p
ro
d
u
c
t
p
lea
su
re
:
2
5
p
o
siti
v
e
e
m
o
ti
o
n
s
in
h
u
m
a
n
-
p
ro
d
u
c
t
in
tera
c
ti
o
n
s
,
”
In
ter
n
a
t
io
n
a
l
J
o
u
rn
a
l
of
De
sig
n
,
v
o
l
.
6
,
n
o
.
2
,
De
c
.
2
0
1
2
.
[1
3
]
M
.
S
.
Ba
laji
a
n
d
A.
Qu
a
z
i,
“
Cu
sto
m
e
rs’
e
m
o
ti
o
n
re
g
u
lat
io
n
stra
teg
ie
s
in
se
rv
ice
fa
il
u
re
e
n
c
o
u
n
t
e
rs,”
Eu
ro
p
e
a
n
J
o
u
rn
a
l
o
f
M
a
rk
e
ti
n
g
,
v
o
l.
5
1
,
n
o
.
5
/6
,
p
p
.
9
6
0
-
9
8
2
,
2
0
1
7
,
d
o
i:
1
0
.
1
1
0
8
/E
JM
-
03
-
2
0
1
5
-
0
1
6
9
.
[1
4
]
A.
D.
G
a
v
ril
o
v
,
A.
J
o
rd
a
c
h
e
,
M.
Va
sd
a
n
i
a
n
d
J.
De
n
g
,
“
P
re
v
e
n
ti
n
g
m
o
d
e
l
o
v
e
rfit
ti
n
g
a
n
d
u
n
d
e
rfi
tt
in
g
i
n
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
s,”
In
ter
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
S
o
ft
wa
re
S
c
ien
c
e
a
n
d
C
o
mp
u
ta
ti
o
n
a
l
In
telli
g
e
n
c
e
(IJ
S
S
CI)
,
v
o
l
.
10
,
n
o
.
4
,
p
p
.
19
-
28
,
De
c
.
2
0
1
8
,
d
o
i
:
1
0
.
4
0
1
8
/IJS
S
CI.
2
0
1
8
1
0
0
1
0
2
.
[1
5
]
K.
M
o
h
a
n
,
A.
S
e
a
l,
O.
Kre
jca
r
a
n
d
Ya
z
id
i,
“
A.
F
ER
-
n
e
t:
F
a
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
io
n
u
si
n
g
d
e
e
p
n
e
u
ra
l
n
e
t,
”
Ne
u
ra
l
Co
m
p
u
t
in
g
a
n
d
Ap
p
li
c
a
t
io
n
s
,
p
p
.
1
-
12
,
2
0
2
1
,
d
o
i:
1
0
.
1
0
0
7
/s
0
0
5
2
1
-
0
2
0
-
0
5
6
7
6
-
y
.
[1
6
]
T.
Vijay
a
k
u
m
a
r.
,
“
Co
m
p
a
ra
ti
v
e
stu
d
y
o
f
c
a
p
su
le
n
e
u
ra
l
n
e
two
rk
in
v
a
rio
u
s
a
p
p
li
c
a
ti
o
n
s,”
J
o
u
rn
a
l
o
f
Arti
fi
c
ia
l
In
telli
g
e
n
c
e
,
v
o
l.
1
,
n
o
.
1,
p
p
.
19
-
27
,
2
0
1
9
,
d
o
i:
1
0
.
3
6
5
4
8
/
jaic
n
.
2
0
1
9
.
1
.
0
0
3
.
[1
7
]
C.
H.
S
a
m
p
a
io
,
W.
J.
Lad
e
i
ra
,
a
n
d
F
.
D.
O.
S
a
n
ti
n
i
,
“
Ap
p
s
fo
r
m
o
b
il
e
b
a
n
k
in
g
a
n
d
c
u
st
o
m
e
r
sa
ti
sfa
c
ti
o
n
:
A
c
ro
ss
-
c
u
lt
u
ra
l
st
u
d
y
,”
I
n
ter
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Ba
n
k
M
a
rk
e
ti
n
g
,
v
o
l
.
35
,
n
o
.
7
,
p
p
.
1
1
3
3
–
1
1
5
3
,
2
0
1
7
,
d
o
i:
1
0
.
1
1
0
8
/IJBM
-
09
-
2
0
1
5
-
0
1
4
6
[1
8
]
K.
An
,
M
.
K.
H
u
i
a
n
d
K.
Leu
n
g
,
“
Wh
o
s
h
o
u
ld
b
e
re
sp
o
n
sib
le
?
Eff
e
c
ts
of
v
o
ice
a
n
d
c
o
m
p
e
n
sa
ti
o
n
o
n
re
sp
o
n
si
b
il
it
y
a
tt
rib
u
t
io
n
,
p
e
rc
e
iv
e
d
j
u
stice
,
a
n
d
p
o
st‐c
o
m
p
lai
n
t
b
e
h
a
v
i
o
rs
a
c
ro
ss
c
u
lt
u
re
s
,”
I
n
ter
n
a
t
io
n
a
l
J
o
u
r
n
a
l
o
f
Co
n
fl
ic
t
M
a
n
a
g
e
me
n
t
,
v
o
l.
12
,
n
o
.
4
,
p
p
.
3
5
0
–
3
6
4
,
2
0
0
1
,
d
o
i:
1
0
.
1
1
0
8
/eb
0
2
2
8
6
3
.
[1
9
]
S.
Lu
,
L.
Xia
o
a
n
d
M
.
Din
g
,
“
A
v
id
e
o
-
b
a
se
d
a
u
t
o
m
a
ted
re
c
o
m
m
e
n
d
e
r
(VA
R)
sy
ste
m
fo
r
g
a
rm
e
n
ts
,”
M
a
rk
e
ti
n
g
S
c
ien
c
e
,
v
o
l.
35
,
n
o
.
3
,
p
p
.
4
8
4
-
5
1
0
,
2
0
1
6
,
d
o
i:
1
0
.
1
2
8
7
/mk
sc
.
2
0
1
6
.
0
9
8
4
.
[2
0
]
N.
T.
De
sh
p
a
n
d
e
a
n
d
S
.
Ra
v
ish
a
n
k
a
r,
“
F
a
c
e
d
e
tec
ti
o
n
a
n
d
re
c
o
g
n
it
io
n
u
sin
g
Vi
o
la
-
Jo
n
e
s
a
l
g
o
r
it
h
m
a
n
d
f
u
si
o
n
o
f
P
CA an
d
AN
N
,”
A
d
v
a
n
c
e
s i
n
Co
mp
u
ta
ti
o
n
a
l
S
c
ien
c
e
s a
n
d
T
e
c
h
n
o
lo
g
y
,
v
o
l.
10
,
n
o
.
5
,
p
p
.
1
1
7
3
-
1
1
8
9
,
2
0
1
7
.
[2
1
]
S.
Ku
m
a
r,
S.
S
in
g
h
a
n
d
J.
Ku
m
a
r,
“
M
u
lt
i
p
le
fa
c
e
d
e
tec
ti
o
n
u
sin
g
h
y
b
ri
d
fe
a
tu
re
s
wit
h
S
VM
c
las
s
ifi
e
r
,”
i
n
Da
t
a
a
n
d
Co
mm
u
n
ica
ti
o
n
Ne
two
rk
s
,
p
p
.
2
5
3
-
2
6
5
,
2
0
1
9
,
d
o
i:
1
0
.
1
0
0
7
/9
7
8
-
9
8
1
-
13
-
2
2
5
4
-
9
_
2
3
.
[2
2
]
H.
T.
Le
a
n
d
L.
A
.
Ve
a
,
“
A
c
u
sto
m
e
r
e
m
o
ti
o
n
re
c
o
g
n
it
i
o
n
th
r
o
u
g
h
fa
c
ial
e
x
p
re
ss
io
n
u
sin
g
k
i
n
e
c
t
se
n
so
r
v
1
a
n
d
v
2
:
A
c
o
m
p
e
ra
ti
v
e
a
n
a
ly
sis
,”
i
n
P
ro
c
e
e
d
in
g
s
o
f
t
h
e
1
0
th
I
n
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
Ub
i
q
u
it
o
u
s
In
fo
rm
a
ti
o
n
M
a
n
a
g
e
me
n
t
a
n
d
C
o
mm
u
n
ica
ti
o
n
,
2
0
1
9
,
p
p
.
1
-
7
,
d
o
i:
1
0
.
1
1
4
5
/2
8
5
7
5
4
6
.
2
8
5
7
6
2
8
.
[2
3
]
T.
Na
k
a
n
o
a
n
d
S.
Ka
to
,
“
P
o
ten
ti
a
li
ty
o
f
3
D
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
two
rk
s
to
e
stim
a
te
c
u
sto
m
e
r
e
x
p
e
c
tatio
n
a
n
d
sa
ti
sfa
c
ti
o
n
,
”
2
0
1
7
I
EE
E
6
th
Glo
b
a
l
Co
n
fer
e
n
c
e
o
n
Co
n
su
m
e
r
El
e
c
tro
n
ics
(GCCE),
2
0
1
7
,
p
p
.
1
-
4,
d
o
i
:
1
0
.
1
1
0
9
/G
CCE.
2
0
1
7
.
8
2
2
9
3
9
1
.
[2
4
]
M
.
S
.
Bo
u
z
a
k
ra
o
u
i
,
A.
S
a
d
iq
,
a
n
d
N.
E
n
n
e
y
a
,
“
A
c
u
sto
m
e
r
e
m
o
t
io
n
re
c
o
g
n
it
io
n
t
h
ro
u
g
h
fa
c
ial
e
x
p
re
ss
io
n
u
sin
g
P
OE
M
d
e
sc
rip
t
o
r
a
n
d
S
V
M
c
las
si
fier,”
i
n
Pro
c
e
e
d
in
g
s o
f
t
h
e
2
n
d
i
n
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
Bi
g
Da
ta
,
Cl
o
u
d
a
n
d
Ap
p
li
c
a
ti
o
n
s
,
2
0
1
7
,
p
p
.
1
-
6
,
d
o
i:
1
0
.
1
1
4
5
/
3
0
9
0
3
5
4
.
3
0
9
0
4
3
6
.
[2
5
]
M
.
J.
Ly
o
n
s,
M
.
Ka
m
a
c
h
i
,
a
n
d
J.
G
y
o
b
a
,
“
Ja
p
a
n
e
se
fe
m
a
le
fa
c
ial
e
x
p
re
ss
io
n
s
(JA
F
F
E),
”
Da
t
a
b
a
se
o
f
d
i
g
it
a
l
ima
g
e
s
,
v
o
l.
3
,
1
9
9
7
.
[2
6
]
S
.
K.
Ra
m
a
n
i
,
“
F
a
c
ial
e
x
p
re
ss
io
n
d
e
tec
ti
o
n
u
sin
g
n
e
u
ra
l
n
e
two
rk
f
o
r
c
u
sto
m
e
r
b
a
se
d
se
rv
ice
,”
In
2
0
1
8
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
,
C
o
mm
u
n
ica
ti
o
n
,
a
n
d
S
ig
n
a
l
Pr
o
c
e
ss
in
g
(ICCCS
P)
,
p
p
.
1
-
4
,
F
e
b
.
2
0
1
8
,
d
o
i:
1
0
.
1
1
0
9
/ICCC
S
P
.
2
0
1
8
.
8
4
5
2
8
6
4
.
[2
7
]
A.
Ca
ro
p
p
o
,
A.
Leo
n
e
,
a
n
d
P.
S
icili
a
n
o
,
“
Co
m
p
a
riso
n
b
e
twe
e
n
d
e
e
p
lea
rn
in
g
m
o
d
e
ls
a
n
d
trad
i
ti
o
n
a
l
m
a
c
h
in
e
lea
rn
in
g
a
p
p
ro
a
c
h
e
s
f
o
r
fa
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
i
o
n
i
n
a
g
e
in
g
a
d
u
lt
s
,”
J
o
u
r
n
a
l
o
f
Co
mp
u
te
r
S
c
ien
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
v
o
l.
35
,
n
o
.
5
,
p
p
.
1
1
2
7
-
1
1
4
6
,
2
0
2
0
,
d
o
i:
1
0
.
1
0
0
7
/s1
1
3
9
0
-
0
2
0
-
9
6
6
5
-
4
.
[2
8
]
M
.
F
.
Va
lstar
a
n
d
M
.
P
a
n
ti
c
.
,
“
In
d
u
c
e
d
d
is
g
u
st,
h
a
p
p
in
e
ss
a
n
d
su
rp
rise
:
a
n
a
d
d
it
i
o
n
t
o
t
h
e
M
M
I
f
a
c
ial
e
x
p
re
ss
io
n
d
a
ta
b
a
se
,
”
i
n
Pro
c
e
e
d
i
n
g
s
o
f
I
n
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
L
a
n
g
u
a
g
e
Res
o
u
rc
e
s
a
n
d
Eva
lu
a
ti
o
n
,
W
o
rk
sh
o
p
o
n
EM
OTION
,
2
0
1
0
,
p
p
.
65
-
70
.
[2
9
]
X.
Qi,
T.
Wan
g
a
n
d
J.
Li
u
,
“
Co
m
p
a
riso
n
o
f
su
p
p
o
rt
v
e
c
to
r
m
a
c
h
i
n
e
a
n
d
S
o
ftma
x
c
las
sifiers
in
c
o
m
p
u
ter
v
isio
n
,”
i
n
2
0
1
7
S
e
c
o
n
d
I
n
ter
n
a
ti
o
n
a
l
Co
n
f
e
re
n
c
e
o
n
M
e
c
h
a
n
ica
l,
Co
n
tro
l
a
n
d
Co
m
p
u
ter
En
g
in
e
e
rin
g
(IC
M
CCE)
,
2
0
1
7
,
p
p
.
1
5
1
-
1
5
5
,
d
o
i:
1
0
.
1
1
0
9
/ICM
C
CE.
2
0
1
7
.
4
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
C
u
s
to
mer’
s
s
p
o
n
ta
n
eo
u
s
fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
(
Go
la
m
Mo
r
s
h
ed
)
1445
[3
0
]
J.
S
h
a
o
a
n
d
Y.
Qia
n
,
“
Th
re
e
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
m
o
d
e
l
s
fo
r
fa
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
io
n
in
t
h
e
wild
,
”
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l
.
3
5
5
,
p
p
.
82
-
92
,
Au
g
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
1
6
/j
.
n
e
u
c
o
m
.
2
0
1
9
.
0
5
.
0
0
5
.
[3
1
]
G
ian
n
o
p
o
u
lo
s,
P
a
n
a
g
io
ti
s,
I
.
P
e
rik
o
s
a
n
d
I
.
Ha
tzily
g
e
ro
u
d
is,
“
De
e
p
lea
rn
in
g
a
p
p
ro
a
c
h
e
s
fo
r
fa
c
ial
e
m
o
ti
o
n
re
c
o
g
n
i
ti
o
n
:
A
c
a
se
stu
d
y
o
n
F
E
R
-
2
0
1
3
,
”
Ad
v
a
n
c
e
s
i
n
Hy
b
ri
d
iza
t
io
n
o
f
I
n
telli
g
e
n
t
M
e
th
o
d
s
,
p
p
.
1
-
1
6
,
2
0
1
8
,
d
o
i:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
3
1
9
-
6
6
7
9
0
-
4
_
1
.
[3
2
]
J.
Va
n
Kle
e
f,
“
T
o
wa
rd
s
h
u
m
a
n
-
li
k
e
p
e
rf
o
rm
a
n
c
e
fa
c
e
d
e
tec
ti
o
n
:
A
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
a
p
p
r
o
a
c
h
,
”
Un
ive
rs
it
y
o
f
T
we
n
te
,
2
0
1
6
.
[3
3
]
G
.
Lin
a
n
d
W.
S
h
e
n
,
“
Re
se
a
rc
h
o
n
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
two
rk
b
a
se
d
o
n
imp
r
o
v
e
d
Re
L
u
p
iec
e
wise
a
c
ti
v
a
ti
o
n
fu
n
c
ti
o
n
,
”
Pr
o
c
e
d
ia
c
o
mp
u
ter
sc
ien
c
e
,
v
o
l
.
1
3
1
,
p
p
.
9
7
7
-
9
8
4
,
2
0
1
8
,
d
o
i:
1
0
.
1
0
1
6
/
j.
p
r
o
c
s.2
0
1
8
.
0
4
.
2
3
9
.
[3
4
]
N.
M
a
,
X.
Zh
a
n
g
,
H.
T
.
Zh
e
n
g
a
n
d
J.
S
u
n
,
“
S
h
u
ff
len
e
t
v
2
:
P
ra
c
ti
c
a
l
g
u
i
d
e
li
n
e
s
fo
r
e
fficie
n
t
CN
N
a
rc
h
it
e
c
tu
re
d
e
sig
n
,”
i
n
Pr
o
c
e
e
d
in
g
s
o
f
t
h
e
Eu
ro
p
e
a
n
c
o
n
fer
e
n
c
e
o
n
c
o
mp
u
ter
v
isio
n
(ECCV
)
,
2
0
1
8
,
p
p
.
1
1
6
-
1
3
1
,
d
o
i:
1
0
.
1
0
0
7
/9
7
8
-
3
-
0
3
0
-
0
1
2
6
4
-
9
_
8
.
[3
5
]
S.
Re
n
,
K.
He
,
R.
G
irs
h
ick
,
X.
Z
h
a
n
g
a
n
d
J.
S
u
n
,
“
Ob
jec
t
d
e
tec
ti
o
n
n
e
two
r
k
s
o
n
c
o
n
v
o
l
u
ti
o
n
a
l
fe
a
t
u
r
e
m
a
p
s
,
”
IEE
E
Tr
a
n
sa
c
ti
o
n
s
o
n
Pa
tt
e
rn
An
a
l
y
sis
and
M
a
c
h
in
e
In
telli
g
e
n
c
e
,
”
v
o
l.
39
,
n
o
.
7
,
1
4
7
6
-
1
4
8
1
,
2
0
1
6
,
d
o
i:
1
0
.
1
1
0
9
/
TP
AMI.
2
0
1
6
.
2
6
0
1
0
9
9
.
[3
6
]
A.
F
.
Ag
a
ra
p
,
“
De
e
p
lea
rn
in
g
u
si
n
g
re
c
ti
fied
li
n
e
a
r
u
n
it
s
(Re
L
u
),
”
a
rXiv p
re
p
ri
n
t
a
rX
iv:
1
8
0
3
.
0
8
3
7
5
,
2
0
1
8
.
[3
7
]
C.
Co
r
tes
a
n
d
V.
Va
p
n
i
k
,
“
S
u
p
p
o
rt
-
v
e
c
to
r
n
et
wo
r
k
s,”
M
a
c
h
i
n
e
L
e
a
rn
in
g
,
v
o
l.
2
0
,
n
o
.
3
,
p
p
.
273
–
2
9
7
,
1
9
9
5
,
d
o
i:
1
0
.
1
0
0
7
/BF
0
0
9
9
4
0
1
8
.
[3
8
]
T.
Zh
a
n
g
,
W.
Zh
e
n
g
,
Z.
C
u
i
,
Y.
Zo
n
g
,
J.
Ya
n
a
n
d
K.
Ya
n
,
“
A
d
e
e
p
n
e
u
ra
l
n
e
two
r
k
-
d
r
iv
e
n
fe
a
tu
re
l
e
a
rn
in
g
m
e
th
o
d
fo
r
m
u
lt
i
-
v
iew
fa
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
i
o
n
,
”
IEE
E
T
ra
n
sa
c
ti
o
n
s
o
n
M
u
lt
ime
d
i
a
,
v
o
l.
18
,
n
o
.
1
2
,
p
p
.
2
5
2
8
-
2
5
3
6
,
2
0
1
6
,
d
o
i:
1
0
.
1
1
0
9
/
TM
M
.
2
0
1
6
.
2
5
9
8
0
9
2
.
[3
9
]
L.
Ch
e
n
,
M
.
Zh
o
u
,
W.
S
u
,
M
.
Wu
,
J
.
S
h
e
a
n
d
K.
Hir
o
ta,
“
S
o
f
tma
x
re
g
re
ss
io
n
b
a
se
d
d
e
e
p
sp
a
r
se
a
u
to
e
n
c
o
d
e
r
n
e
two
rk
fo
r
fa
c
ial
e
m
o
ti
o
n
re
c
o
g
n
it
i
o
n
i
n
h
u
m
a
n
-
ro
b
o
t
i
n
tera
c
ti
o
n
,
”
I
n
f
o
rm
a
ti
o
n
S
c
ien
c
e
s
,
v
o
l.
428
,
p
p
.
49
-
61
,
2
0
1
8
,
d
o
i:
1
0
.
1
0
1
6
/
j.
i
n
s.2
0
1
7
.
1
0
.
0
4
4
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
G
o
la
m
M
o
r
shed
is
c
u
rre
n
tl
y
wo
rk
i
n
g
a
s
a
C
h
ief
Tec
h
n
o
lo
g
y
Offic
e
r
a
t
H
M
N
g
ro
u
p
o
f
c
o
m
p
a
n
ies
a
n
d
S
e
n
io
r
S
o
ftwa
re
En
g
i
n
e
e
r
a
t
Ac
c
o
u
n
t
S
o
ft
E
n
terp
r
ise
S
d
n
B
h
d
.
He
g
ra
d
u
a
te
d
fro
m
Un
i
v
e
rsiti
M
a
lay
sia
S
a
ra
wa
k
i
n
2
0
1
8
wit
h
a
b
a
c
h
e
l
o
r’s
d
e
g
re
e
in
c
o
m
p
u
ter
sc
ien
c
e
(S
o
ftwa
re
En
g
in
e
e
ri
n
g
).
H
a
m
im
a
h
Ujir
is
c
u
rre
n
tl
y
a
S
e
n
io
r
Lec
tu
re
r
a
t
F
C
S
IT,
UN
IM
AS.
S
h
e
re
c
e
iv
e
d
h
e
r
P
h
D
fr
o
m
Un
iv
e
rsity
o
f
Birmi
n
g
h
a
m
,
Un
it
e
d
Ki
n
g
d
o
m
in
2
0
1
3
.
Ha
m
ima
h
’s
re
se
a
rc
h
in
tere
sts
lay
in
th
e
in
terd
isc
ip
li
n
a
ry
field
o
f
c
o
m
p
u
t
e
r
v
isio
n
,
with
re
late
d
in
tere
sts
b
e
in
g
in
c
o
m
p
u
ter
g
ra
p
h
ics
,
ima
g
e
p
ro
c
e
ss
in
g
,
a
n
d
m
a
th
e
m
a
ti
c
a
l
m
e
th
o
d
s.
He
r
p
re
v
io
u
s
a
n
d
c
u
rre
n
t
wo
rk
s
in
c
l
u
d
e
3
D
p
h
y
sic
a
l
sim
u
latio
n
a
n
d
3
D
sta
ti
c
a
n
d
d
y
n
a
m
ic
fa
c
ial
e
x
p
re
ss
io
n
a
n
a
ly
sis.
S
h
e
a
lso
c
o
n
d
u
c
ted
re
se
a
rc
h
wo
rk
s o
n
a
c
a
d
e
m
ic q
u
a
li
ty
i
n
h
i
g
h
e
r
e
d
u
c
a
ti
o
n
.
Ir
wa
n
d
i
g
ra
d
u
a
ted
with
a
P
h
D
i
n
Co
m
p
u
ter
Visio
n
fro
m
Un
i
v
e
rsity
o
f
Brist
o
l
b
a
c
k
in
2
0
1
4
.
He
a
lso
h
o
ld
s
a
n
M
S
c
(2
0
0
7
)
a
n
d
a
BS
c
(
2
0
0
3
)
fr
o
m
Un
iv
e
rsiti
Tek
n
o
lo
g
i
M
a
lay
sia
.
He
is
c
u
rre
n
tl
y
se
rv
in
g
a
s
th
e
De
p
u
ty
D
e
a
n
o
f
Un
d
e
rg
ra
d
u
a
tes
a
t
th
e
F
a
c
u
lt
y
o
f
C
o
m
p
u
te
r
S
c
ien
c
e
&
In
fo
rm
a
ti
o
n
Tec
h
n
o
lo
g
y
,
Un
i
v
e
rsiti
M
a
lay
sia
S
a
ra
wa
k
.
His
re
se
a
rc
h
in
tere
sts
a
re
,
b
u
t
n
o
t
li
m
it
e
d
to
,
C
o
m
p
u
ter
Visio
n
,
P
a
tt
e
rn
M
a
tch
in
g
a
n
d
A
u
g
m
e
n
te
d
Re
a
li
ty
.
M
o
re
sp
e
c
ifi
c
a
ll
y
,
h
e
is
a
c
ti
v
e
ly
p
u
rsu
i
n
g
re
se
a
rc
h
g
ra
n
ts/p
o
stg
ra
d
u
a
tes
o
n
th
e
to
p
ics
o
f
Visu
a
l
An
ima
l
Bio
m
e
tri
c
a
n
d
t
h
e
u
se
o
f
Au
g
m
e
n
te
d
Re
a
li
ty
(AR) fo
r
As
siste
d
L
iv
i
n
g
/
Ed
u
c
a
ti
o
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.