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.
25
,
No
.
3
,
Ma
r
ch
2
0
2
2
,
p
p
.
1
4
8
9
~
1
5
0
0
I
SS
N:
2
5
0
2
-
4
7
5
2
,
DOI
: 1
0
.
1
1
5
9
1
/ijeecs.v
25
.i
3
.
p
p
1
4
8
9
-
1
5
0
0
1489
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ee
cs.ia
esco
r
e.
co
m
Linea
r
fusio
n app
ro
a
ch t
o
conv
o
lut
io
na
l neural net
w
o
rks for
facia
l emo
tion rec
o
g
nition
Usen Du
dek
ula
,
P
urna
cha
nd
N
S
c
h
o
o
l
o
f
E
l
e
c
t
r
o
n
i
c
s
En
g
i
n
e
e
r
i
n
g
,
V
I
T
-
A
P
U
n
i
v
e
r
s
i
t
y
,
A
m
a
r
a
v
a
t
i
,
A
n
d
h
r
a
P
r
a
d
e
s
h
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Au
g
12
,
2
0
2
1
R
ev
is
ed
Dec
31
,
2
0
2
1
Acc
ep
ted
J
an
21
,
2
0
2
2
F
a
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
i
o
n
is
a
c
h
a
ll
e
n
g
in
g
p
ro
b
lem
in
t
h
e
sc
ien
ti
fic
field
o
f
c
o
m
p
u
ter
v
isio
n
.
S
e
v
e
ra
l
fa
c
e
e
x
p
re
ss
io
n
re
c
o
g
n
i
ti
o
n
(F
ER)
a
lg
o
rit
h
m
s
a
re
p
ro
p
o
se
d
in
th
e
field
o
f
m
a
c
h
in
e
lea
rn
i
n
g
,
a
n
d
d
e
e
p
lea
rn
i
n
g
to
e
x
trac
t
e
x
p
re
ss
io
n
k
n
o
wle
d
g
e
fro
m
fa
c
ial
re
p
re
se
n
tatio
n
s.
Ev
e
n
th
o
u
g
h
n
u
m
e
ro
u
s
a
lg
o
rit
h
m
s
h
a
v
e
b
e
e
n
e
x
a
m
in
e
d
,
se
v
e
ra
l
issu
e
s
li
k
e
li
g
h
ti
n
g
c
h
a
n
g
e
s
,
ro
tatio
n
s
a
n
d
o
c
c
lu
si
o
n
s.
We
p
re
se
n
t
a
n
e
fficie
n
t
a
p
p
r
o
a
c
h
t
o
e
n
h
a
n
c
e
re
c
o
g
n
it
i
o
n
a
c
c
u
ra
c
y
i
n
th
is
stu
d
y
,
a
d
v
o
c
a
tes
tran
sfe
r
lea
rn
in
g
to
fin
e
-
tu
n
e
th
e
p
a
ra
m
e
ters
o
f
th
e
p
re
-
train
e
d
m
o
d
e
l
(VG
G
1
9
m
o
d
e
l)
a
n
d
n
o
n
-
p
re
-
train
e
d
m
o
d
e
l
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s
(CNN
s)
fo
r
t
h
e
tas
k
o
f
ima
g
e
c
las
sifica
ti
o
n
.
T
h
e
VG
G
1
9
n
e
two
rk
a
n
d
c
o
n
v
o
lu
t
io
n
a
l
n
e
two
rk
d
e
riv
e
tw
o
c
h
a
n
n
e
ls
o
f
e
x
p
re
ss
io
n
re
late
d
c
h
a
ra
c
terist
ics
fro
m
th
e
fa
c
ial
g
ra
y
sc
a
le
ima
g
e
s.
Th
e
li
n
e
a
r
fu
si
o
n
a
l
g
o
ri
th
m
c
a
lcu
late
s
th
e
c
las
s
b
y
tak
in
g
a
n
a
v
e
ra
g
e
o
f
e
a
c
h
c
las
sifica
ti
o
n
d
e
c
isio
n
o
n
train
i
n
g
sa
m
p
les
o
f
b
o
t
h
c
h
a
n
n
e
ls.
F
in
a
l
re
c
o
g
n
it
i
o
n
is
c
a
lcu
late
d
u
sin
g
c
o
n
v
o
l
u
ti
o
n
n
e
u
ra
l
n
e
two
r
k
a
rc
h
it
e
c
tu
re
fo
ll
o
we
d
b
y
a
so
f
tma
x
c
las
sifi
e
r.
S
e
v
e
n
b
a
sic
fa
c
ial
e
m
o
ti
o
n
s
(BEs):
h
a
p
p
i
n
e
ss
,
su
rp
r
ise
,
a
n
g
e
r,
sa
d
n
e
ss
,
fe
a
r,
d
is
g
u
st,
a
n
d
n
e
u
t
ra
l
fa
c
ial
e
x
p
re
ss
io
n
s
c
a
n
b
e
re
c
o
g
n
ize
d
b
y
t
h
e
p
ro
p
o
se
d
a
l
g
o
rit
h
m
,
T
h
e
a
v
e
ra
g
e
a
c
c
u
ra
c
ies
fo
r
sta
n
d
a
rd
d
a
ta
se
t’s
“
CK+
,
”
a
n
d
“
JA
F
F
E,
”
9
8
.
3
%
a
n
d
9
2
.
4
%
,
re
sp
e
c
ti
v
e
ly
.
Us
in
g
a
d
e
e
p
n
e
two
rk
wit
h
o
n
e
c
h
a
n
n
e
l,
th
e
p
ro
p
o
se
d
a
lg
o
rit
h
m
c
a
n
a
c
h
iev
e
we
ll
c
o
m
p
a
ra
b
le p
e
rfo
rm
a
n
c
e
.
K
ey
w
o
r
d
s
:
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
Dee
p
lear
n
in
g
Ma
ch
in
e
lear
n
in
g
T
r
an
s
f
er
lear
n
i
n
g
Vis
u
al
g
eo
m
etr
y
g
r
o
u
p
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
:
Pu
r
n
ac
h
an
d
N
Sch
o
o
l o
f
E
lectr
o
n
ics E
n
g
i
n
ee
r
in
g
,
VI
T
-
AP U
n
iv
e
r
s
ity
Am
ar
av
ati,
An
d
h
r
a
Pra
d
esh
,
I
n
d
ia
E
m
ail:
ch
an
d
u
in
ec
e@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
E
m
o
tio
n
al
ex
p
r
ess
io
n
s
ar
e
th
e
m
o
s
t
im
p
o
r
tan
t
as
it
en
c
o
d
es
n
o
n
-
v
er
b
al
wa
y
s
o
f
ex
p
r
ess
in
g
in
ter
io
r
em
o
tio
n
s
an
d
in
ten
tio
n
s
.
T
h
e
f
ac
ial
ac
tio
n
co
d
in
g
s
y
s
tem
(
F
AC
S)
is
a
s
u
itab
le
s
tr
u
ctu
r
e
th
at
u
s
es
ac
tio
n
u
n
its
(
AU)
to
id
en
tify
h
u
m
an
f
ac
ial
b
eh
a
v
io
r
o
n
th
e
f
ace
[
1
]
.
Mu
l
tifa
r
io
u
s
m
eth
o
d
s
co
n
f
id
e
in
t
h
e
ex
tr
ac
tio
n
o
f
th
e
f
ac
ial
r
eg
io
n
[
2
]
o
r
an
a
u
to
m
a
tic
s
o
lu
tio
n
to
id
e
n
tific
atio
n
[
3
]
.
Prim
ar
y
f
ac
ial
r
ec
o
g
n
itio
n
ex
tr
ac
tio
n
m
eth
o
d
s
in
clu
d
e
lo
ca
l
d
ir
ec
tio
n
al
p
att
er
n
(
L
DP)
,
lin
ea
r
d
is
cr
im
in
a
n
t
an
aly
s
is
(
L
DA)
,
lo
ca
l
b
in
ar
y
p
atter
n
s
(
L
B
P),
p
r
in
cip
al
c
o
m
p
o
n
en
t
a
n
aly
s
is
(
PC
A)
,
an
d
co
n
v
o
l
u
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
[
4
]
.
AUs
ar
e
co
r
e
in
d
icato
r
s
,
m
ak
in
g
a
d
ec
lar
atio
n
ab
o
u
t
th
e
c
o
r
r
e
s
p
o
n
d
en
t
em
o
tio
n
ac
tiv
atio
n
p
o
in
t
[
5
]
,
[
6
]
.
AUs
m
ay
n
o
t
o
n
ly
b
e
u
tili
ze
d
to
r
ev
ea
l
em
o
tio
n
s
an
d
tex
tu
r
es.
W
h
en
th
e
f
ac
e
r
ev
ea
ls
em
o
t
io
n
an
d
m
u
ltip
le
f
ilter
s
ca
n
b
e
ap
p
lied
f
o
r
t
h
e
ex
p
o
s
u
r
e
o
f
f
ac
ial
em
o
tio
n
s
C
o
n
v
o
lu
tio
n
al
n
e
u
r
al
n
etwo
r
k
s
ar
e
s
p
ec
if
ic
k
in
d
s
o
f
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
an
d
h
a
v
e
b
ee
n
wo
r
k
i
n
g
wi
th
r
ea
s
o
n
ab
le
p
er
f
o
r
m
an
ce
as
a
f
ea
tu
r
e
ex
tr
ac
t
o
r
[
7
]
.
Facial
ex
p
r
ess
io
n
id
en
tific
atio
n
ty
p
ically
r
eq
u
ir
es
th
r
ee
p
h
ases
o
f
p
r
ep
ar
atio
n
co
n
s
is
tin
g
o
f
th
e
ac
q
u
is
itio
n
o
f
th
e
f
ac
e,
th
e
r
etr
iev
al
o
f
th
e
f
ea
tu
r
es,
an
d
th
e
class
if
ier
[
8
]
.
Desp
ite
r
ec
en
t
r
ap
id
ad
v
an
ce
s
,
f
ac
ial
em
o
tio
n
r
ec
o
g
n
itio
n
(
FER)
r
em
ain
s
d
if
f
icu
lt
d
u
e
t
o
a
v
ar
iety
o
f
c
h
allen
g
es
s
u
c
h
as
im
p
r
o
v
e
m
en
ts
in
lig
h
tin
g
an
d
ac
ce
s
s
o
r
ies,
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.
25
,
No
.
3
,
Ma
r
ch
20
22
:
1
4
8
9
-
1
5
0
0
1490
p
ar
tial
o
cc
lu
s
io
n
s
,
ca
p
tu
r
in
g
h
ea
d
d
ef
lectio
n
o
f
th
e
f
ac
ial
a
r
ea
s
,
th
e
id
en
tific
atio
n
r
ate
r
em
ain
s
lo
w
d
u
e
to
th
e
in
f
lu
en
ce
o
f
d
if
f
e
r
en
t
a
m
b
ien
t a
s
well
as
d
if
f
er
e
n
ce
s
in
i
n
d
iv
i
d
u
al
p
e
o
p
le'
s
tr
aits
.
Ma
n
y
c
h
a
r
ac
ter
is
tics
m
ay
b
e
r
etr
iev
ed
a
n
d
lear
n
ed
f
o
r
a
g
o
o
d
f
ac
e
ex
p
r
ess
io
n
d
etec
ti
o
n
s
y
s
tem
u
s
in
g
d
ee
p
lear
n
i
n
g
a
n
d
s
p
ec
if
ically
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs).
No
tewo
r
th
y
is
th
at
i
n
th
e
ca
s
e
o
f
f
ac
ial
ex
p
r
ess
io
n
s
,
m
an
y
in
d
icatio
n
s
o
r
ig
in
ate
f
r
o
m
a
f
ew
r
eg
io
n
s
o
f
th
e
f
ac
e,
s
u
ch
as th
e
lip
s
an
d
ey
es,
wh
ile
o
th
er
p
o
r
tio
n
s
o
f
t
h
e
f
ac
e
p
lay
a
litt
le
r
o
le
in
th
eir
p
r
o
d
u
ctio
n
.
As
a
r
esu
lt,
th
e
m
ac
h
in
e
lear
n
in
g
f
r
am
ewo
r
k
s
h
o
u
ld
id
ea
lly
ju
s
t
f
o
cu
s
o
n
th
e
m
o
s
t
ess
en
tial
p
o
r
tio
n
s
o
f
th
e
f
ac
e,
wh
ile
b
ein
g
less
s
en
s
itiv
e
to
o
th
er
f
ac
ial
r
eg
io
n
s
.
Han
d
cr
a
f
ted
f
ea
tu
r
es
f
o
r
FER
ass
ig
n
m
en
ts
ar
e
n
o
lo
n
g
er
a
d
e
q
u
ate.
A
s
u
b
o
p
tim
al
s
o
lu
tio
n
t
o
th
ese
p
r
o
b
lem
s
ca
n
b
e
o
f
f
e
r
ed
b
y
d
ee
p
lear
n
in
g
m
eth
o
d
o
l
o
g
ies.
I
n
m
o
s
t
f
ield
s
o
f
m
ac
h
in
e
lear
n
in
g
an
d
co
m
p
u
ter
v
is
io
n
,
C
NN
h
as
b
ee
n
h
ig
h
ly
s
u
cc
ess
f
u
l,
th
e
r
e
s
e
a
r
c
h
e
r
s
u
s
e
C
N
N
f
o
r
o
b
j
e
ct
d
e
t
e
ct
i
o
n
i
n
a
w
i
d
e
r
a
n
g
e
o
f
a
p
p
l
i
c
a
t
i
o
n
s
[
9
]
,
[
1
0
]
.
W
e
p
r
ese
n
t
a
d
e
e
p
l
e
a
r
n
i
n
g
-
b
a
s
e
d
s
y
s
t
e
m
f
o
r
f
a
c
e
e
x
p
r
e
s
s
io
n
i
d
e
n
t
i
f
i
c
a
ti
o
n
i
n
t
h
is
p
a
p
e
r
th
a
t
c
o
n
s
i
d
e
r
s
t
h
e
p
r
e
c
e
d
i
n
g
o
b
s
e
r
v
a
t
i
o
n
.
F
E
R
t
as
k
s
r
e
l
y
o
n
d
e
t
e
c
t
i
n
g
f
a
c
i
a
l
e
x
p
r
e
s
s
i
o
n
s
a
n
d
i
d
e
n
t
i
f
y
i
n
g
f
a
c
es
b
a
s
e
d
o
n
R
G
B
o
r
g
r
a
y
s
c
a
l
e
p
i
ct
u
r
e
s
.
T
r
a
d
i
t
i
o
n
al
F
E
R
t
a
s
k
s
d
e
p
e
n
d
o
n
h
a
n
d
-
c
r
a
f
t
e
d
f
e
a
t
u
r
e
s
.
F
e
at
u
r
e
s
m
a
y
b
e
d
iv
i
d
e
d
i
n
t
o
t
h
r
e
e
p
r
i
m
a
r
y
c
a
teg
o
r
i
e
s
:
a
p
p
e
a
r
a
n
ce
,
g
e
o
m
e
t
r
y
,
a
n
d
m
o
t
i
o
n
c
h
a
r
a
c
t
e
r
i
s
ti
c
s
,
r
es
p
e
c
ti
v
e
l
y
.
P
i
x
el
i
n
t
e
n
s
it
y
[
1
1
]
,
G
a
b
o
r
t
e
x
t
u
r
e
[
1
2
]
,
L
B
P
[
1
3
]
,
a
n
d
h
i
s
t
o
g
r
a
m
o
f
o
r
i
e
n
t
e
d
g
r
a
d
i
e
n
t
s
(
H
OG
)
[
1
4
]
,
a
r
e
s
o
m
e
o
f
t
h
e
m
o
s
t
o
f
t
e
n
u
t
i
l
i
z
e
d
a
p
p
e
a
r
a
n
c
e
c
h
a
r
a
c
t
e
r
is
t
ic
s
.
T
h
e
s
e
f
e
a
t
u
r
es
f
r
o
m
t
h
e
f
u
l
l
f
a
c
i
a
l
r
e
g
i
o
n
a
r
e
c
o
n
s
i
d
e
r
e
d
,
b
u
t
t
h
e
e
y
e
s
,
n
o
s
e
,
a
n
d
m
o
u
t
h
a
r
e
n
o
t
t
a
k
e
n
i
n
t
o
c
o
n
s
i
d
e
r
a
t
i
o
n
.
T
h
e
r
e
f
o
r
e
,
F
E
R
t
a
s
k
s
e
m
p
l
o
y
g
e
o
m
et
r
i
c
c
h
a
r
ac
t
e
r
i
s
t
i
cs
,
w
h
i
c
h
a
r
e
r
e
p
r
es
e
n
t
ed
b
y
t
h
e
g
e
o
m
e
t
r
ic
c
o
n
n
e
c
t
i
o
n
s
o
f
f
a
c
i
al
la
n
d
m
a
r
k
p
o
i
n
t
s
i
d
e
n
t
i
f
i
e
d
f
r
o
m
l
o
c
a
l
a
r
e
a
s
t
h
at
a
r
e
s
i
g
n
i
f
i
c
a
n
tl
y
lin
k
e
d
t
o
e
x
p
r
e
s
s
i
o
n
v
a
r
i
a
t
i
o
n
s
[
1
5
]
.
F
u
r
t
h
e
r
m
o
r
e
,
c
o
m
b
i
n
i
n
g
d
i
f
f
e
r
e
n
t
f
e
at
u
r
e
s
is
a
t
r
e
n
d
t
h
at
h
as
g
r
e
a
t
p
o
t
e
n
ti
a
l
[
1
6
]
.
T
w
o
-
s
t
a
g
e
m
u
l
t
i
-
t
as
k
f
r
a
m
e
w
o
r
k
t
o
e
x
p
l
o
r
e
F
E
R
.
W
i
t
h
t
h
e
u
s
e
o
f
li
n
e
a
r
-
c
h
a
i
n
o
r
t
i
c
o
t
r
o
p
i
n
-
r
e
l
e
as
i
n
g
f
ac
t
o
r
(
C
R
F
)
,
h
i
d
d
e
n
C
R
F
,
a
n
d
h
i
d
d
e
n
l
a
y
e
r
v
a
r
i
a
b
le
s
[
1
7
]
c
r
e
a
t
e
d
a
n
i
n
t
e
r
a
ct
i
v
e
,
m
u
l
t
i
-
d
i
m
e
n
s
i
o
n
a
l
m
o
d
e
l
o
f
t
h
e
h
i
d
d
e
n
l
a
y
e
r
.
A
s
a
r
e
s
u
l
t
o
f
t
h
is
m
o
d
e
,
a
s
i
m
il
a
r
it
y
a
n
a
l
y
s
is
is
u
s
e
d
t
o
d
e
t
e
r
m
in
e
h
o
w
a
n
e
x
p
r
e
s
s
i
o
n
c
h
a
n
g
e
s
.
A
l
r
e
a
d
y
e
x
is
t
i
n
g
m
e
t
h
o
d
s
f
o
r
f
a
c
i
a
l
r
ec
o
g
n
i
t
i
o
n
u
s
i
n
g
h
a
n
d
-
c
r
a
f
t
e
d
f
e
a
t
u
r
es
h
a
v
e
a
l
i
m
it
e
d
r
e
c
o
g
n
i
t
i
o
n
c
a
p
a
b
il
it
y
.
Nu
m
er
o
u
s
in
v
esti
g
atio
n
s
h
a
v
e
r
ec
en
tly
in
co
n
s
id
er
atio
n
o
f
d
ee
p
lear
n
in
g
,
s
tu
d
ied
FER
p
r
o
b
lem
s
in
p
atter
n
r
ec
o
g
n
itio
n
,
FER
h
as
h
ad
r
em
ar
k
a
b
le
s
u
cc
ess
[
1
8
]
.
Usi
n
g
d
ee
p
b
elief
n
etwo
r
k
s
(
DB
Ns),
tr
ain
ed
a
m
u
lti
-
lay
er
p
er
ce
p
tr
o
n
(
ML
P
)
to
d
etec
t
d
is
tin
ct
f
ac
ial
ex
p
r
ess
io
n
s
b
ased
o
n
th
e
lear
n
in
g
f
ea
tu
r
es.
ML
P
s
u
r
p
ass
es
b
o
th
SVM
an
d
R
F
class
if
ier
s
[
1
9
]
.
C
NN
f
o
r
FE
R
an
d
r
ep
o
r
ted
its
s
atis
f
ac
to
r
y
p
er
f
o
r
m
an
ce
in
th
e
`
`
C
K+
''
d
ata
s
et.
A
d
ata
au
g
m
en
tatio
n
s
tr
ateg
y
was
p
r
o
p
o
s
ed
to
ad
d
r
ess
th
e
lack
o
f
lab
eled
s
am
p
les
f
o
r
C
N
N
tr
ain
in
g
.
Sev
er
al
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
o
lo
g
ies
wer
e
also
u
s
ed
to
p
r
eser
v
e
e
x
p
r
ess
io
n
-
r
elat
ed
f
ea
tu
r
es
in
f
ac
ial
im
ag
es
[
2
0
]
.
C
o
m
b
i
n
ed
s
ev
er
a
l
C
NN
s
to
s
tu
d
y
FER
[
2
1
]
.
I
t
was
p
o
s
s
ib
le
to
co
m
b
in
e
th
ese
C
NNs
b
y
lear
n
in
g
th
e
s
et
weig
h
ts
o
f
t
h
e
n
etwo
r
k
r
esp
o
n
s
e
also
tr
ai
n
ed
s
ev
e
r
al
d
ee
p
C
NNs
f
o
r
r
o
b
u
s
t
FER
[
2
2
]
.
AU
-
in
s
p
ir
ed
d
ee
p
n
etwo
r
k
s
(
AUDN
s
)
[
2
3
]
.
As
a
r
esu
lt
o
f
AUDN
'
s
f
o
cu
s
o
n
a
s
in
g
le
f
ac
e
p
ictu
r
e
in
p
u
t
m
o
d
ality
,
its
r
ec
o
g
n
itio
n
ca
p
a
b
ilit
ies
ar
e
l
im
ited
.
I
n
a
h
ig
h
ly
d
ee
p
n
e
u
r
al
n
etwo
r
k
b
etter
ch
ar
ac
te
r
is
tics
s
p
ec
if
ic
f
o
r
ex
p
r
ess
io
n
r
ep
r
esen
tatio
n
.
Fo
u
r
in
ce
p
tio
n
lay
er
s
f
o
llo
we
d
a
m
ax
-
p
o
o
lin
g
lay
er
an
d
two
co
n
v
o
l
u
tio
n
al
n
etwo
r
k
s
.
Ho
wev
e
r
,
it
is
im
p
o
s
s
ib
le
to
tr
ain
th
is
n
etwo
r
k
with
o
u
t
th
e
u
s
e
o
f
c
o
m
p
u
tatio
n
a
l
p
o
wer
(
esp
ec
ially
GPUs
)
[
2
4
]
.
No
v
el
f
ac
ial
r
e
co
g
n
itio
n
ap
p
r
o
ac
h
u
s
in
g
a
g
u
id
ed
im
ag
e
f
ilter
an
d
a
c
o
n
v
o
lu
ti
o
n
al
n
e
u
r
al
n
etwo
r
k
[
2
5
]
.
Han
d
cr
a
f
ted
f
ea
tu
r
es
ar
e
th
e
f
o
u
n
d
atio
n
o
f
FE
R
ap
p
r
o
ac
h
es.
T
h
e
u
s
e
o
f
f
ac
i
al
d
ep
th
p
ictu
r
es
as
an
in
p
u
t
to
d
ee
p
n
etwo
r
k
s
is
,
u
n
f
o
r
t
u
n
ately
,
r
ar
e.
All
o
f
th
e
p
r
ev
io
u
s
s
tu
d
ies
h
av
e
m
ad
e
s
u
b
s
tan
tial
p
r
o
g
r
ess
in
th
e
f
ield
o
f
em
o
tio
n
id
en
tif
icatio
n
wh
en
co
m
p
ar
ed
to
p
r
e
v
io
u
s
ef
f
o
r
ts
,
b
u
t
th
ey
lack
a
clea
r
tech
n
iq
u
e
f
o
r
id
en
tify
in
g
k
e
y
f
ac
ial
ar
ea
s
f
o
r
em
o
tio
n
d
etec
tio
n
.
B
y
u
tili
zin
g
a
lin
ea
r
f
u
s
io
n
n
etwo
r
k
-
b
as
ed
ar
ch
itectu
r
e,
we
attem
p
t
to
s
o
lv
e
t
h
is
is
s
u
e
b
y
f
o
cu
s
in
g
o
n
t
h
e
em
o
tio
n
s
with
an
ac
c
u
r
ac
y
o
f
s
tan
d
ar
d
d
at
a
s
et’
s
“CK+,
”
an
d
“JAFF
E
,
”
9
8
.
3
%
an
d
9
2
.
4
%,
r
esp
ec
tiv
ely
.
T
h
is
p
ap
er
f
o
cu
s
es
o
n
th
e
p
r
o
b
lem
s
o
f
ch
a
r
a
cter
is
tics
ex
tr
ac
tio
n
an
d
f
ac
ial
ex
p
r
ess
io
n
d
etec
tio
n
.
First
o
f
all,
b
i
n
ar
y
f
ac
ial
im
ag
e
ch
an
n
els,
in
clu
d
in
g
g
r
a
y
im
ag
es
an
d
L
B
P
im
ag
es,
will
b
e
u
s
ed
b
y
FER
u
s
in
g
co
n
v
o
l
u
tio
n
n
eu
r
al
n
etwo
r
k
s
.
Seco
n
d
ly
,
a
m
eth
o
d
o
lo
g
y
f
o
r
f
in
e
-
t
u
n
in
g
is
u
s
ed
to
o
p
tim
ize
th
e
u
s
e
o
f
a
well
-
tr
ain
ed
p
r
e
-
T
r
ain
ed
n
e
t
wo
r
k
(
VGG1
9
m
o
d
el
o
n
I
m
a
g
eNe
t)
.
Pro
v
id
e
d
b
y
lin
ea
r
f
u
s
io
n
to
b
o
t
h
ch
an
n
e
l
o
u
tp
u
ts
.
Fin
al
r
ec
o
g
n
itio
n
is
ca
lcu
lated
u
s
in
g
co
n
v
o
lu
ti
o
n
n
eu
r
al
n
etwo
r
k
ar
ch
itectu
r
e
f
o
llo
wed
b
y
a
s
o
f
tm
ax
class
if
ier
.
Pro
ce
s
s
es
th
e
o
u
tco
m
es
an
d
d
o
es
f
ac
ial
e
x
p
r
ess
io
n
p
r
o
jectio
n
s
f
r
o
m
b
e
n
ch
m
ar
k
f
ac
ial
ex
p
r
ess
io
n
s
(
h
ap
p
in
ess
,
s
ad
n
ess
,
an
g
er
,
s
u
r
p
r
is
e,
d
is
g
u
s
t,
f
ea
r
,
an
d
n
at
u
r
al)
.
I
m
p
r
o
v
em
en
ts
o
n
VGG1
6
le
ad
in
g
t
o
VGG1
9
o
v
er
co
m
es
Alex
Net’
s
lim
itatio
n
s
an
d
i
m
p
r
o
v
es
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
.
VGG1
9
c
o
n
s
is
ts
o
f
s
o
m
e
u
n
n
ec
ess
ar
y
R
eL
U
u
n
its
i
n
th
e
ce
n
ter
o
f
th
e
n
et
wo
r
k
in
co
n
tr
ast
to
VGG1
6
[
1
0
]
.
O
u
r
an
aly
s
is
is
s
tr
u
ctu
r
ed
as
f
o
llo
ws.
Sectio
n
2
is
ab
o
u
t
th
e
in
tr
o
d
u
ctio
n
.
Se
ctio
n
2
is
ab
o
u
t
th
e
o
v
er
v
iew
o
f
th
e
d
ataset’
s
an
d
also
in
cl
u
d
es
co
m
p
r
eh
e
n
s
iv
e
C
NNs
an
d
s
o
m
e
p
r
e
-
tr
ai
n
ed
d
ee
p
n
etwo
r
k
m
o
d
els.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
is
d
ef
in
ed
in
s
ec
tio
n
3
.
T
h
e
test
r
esu
lts
an
d
in
ter
p
r
etatio
n
f
o
r
s
ec
tio
n
4
ar
e
an
n
o
u
n
ce
d
.
Sectio
n
5
p
r
o
v
id
es
th
e
co
n
clu
s
io
n
s
tatem
en
t.
2.
DATAS
E
T
’
S DE
SCR
I
P
T
I
O
N
E
m
o
tio
n
p
lay
s
a
m
ajo
r
r
o
le
i
n
in
ter
p
er
s
o
n
al
co
m
m
u
n
icatio
n
an
d
also
in
im
p
r
o
v
in
g
life
s
ty
le.
Face
em
o
tio
n
d
etec
tio
n
ca
n
also
ex
t
r
ac
t
a
s
et
o
f
f
ac
e
-
r
elate
d
attr
ib
u
tes,
s
u
ch
as h
ea
d
p
o
s
e,
ag
e,
e
m
o
tio
n
,
f
ac
ial
h
ai
r
,
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
Lin
ea
r
fu
s
io
n
a
p
p
r
o
a
ch
to
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks fo
r
fa
cia
l
emo
tio
n
r
ec
o
g
n
itio
n
(
Usen
Du
d
ek
u
la
)
1491
an
d
g
lass
es.
T
h
ese
attr
ib
u
te
s
ar
e
co
m
m
o
n
p
r
ed
ictio
n
s
,
n
o
t
au
th
e
n
tic
class
if
icatio
n
s
.
W
h
en
u
s
er
s
ad
d
th
em
s
elv
es
to
a
f
ac
e
s
er
v
ice
f
e
w
attr
ib
u
tes
ar
e
u
s
ef
u
l,
to
m
ak
e
s
u
r
e
th
at
th
e
ap
p
licatio
n
is
g
ettin
g
h
ig
h
-
q
u
ality
f
ac
e
d
ata
(
f
o
r
ex
am
p
le,
o
u
r
a
p
p
licatio
n
co
u
l
d
ad
v
is
e
u
s
er
s
to
tak
e
o
f
f
t
h
eir
s
u
n
g
lass
es
if
th
e
u
s
er
is
wea
r
in
g
s
u
n
g
lass
es).
T
h
e
tr
ial
wa
s
c
o
n
s
u
m
m
ate
o
n
two
p
u
b
licly
ac
ce
s
s
ib
le
f
ac
ial
ex
p
r
ess
io
n
d
atab
ases
:
J
AFFE
(
J
ap
an
es
e
f
em
ale
f
ac
ial
ex
p
r
ess
io
n
)
an
d
C
o
h
n
-
Kan
a
d
e
(
C
K+
)
.
W
e
test
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
ex
ten
s
iv
el
y
in
o
u
r
d
ataset
in
f
ac
ial
e
m
o
tio
n
al
ex
p
r
ess
io
n
class
if
icatio
n
.
T
h
e
ex
p
e
r
im
en
tal
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
u
tp
er
f
o
r
m
th
e
m
o
s
t
s
tate
-
of
-
th
e
-
ar
t
f
ac
ial
em
o
tio
n
al
ex
p
r
es
s
io
n
r
ec
o
g
n
itio
n
s
y
s
tem
s
.
Data
s
ets
ar
e
u
s
ed
f
o
r
tr
ain
in
g
an
d
1
0
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
test
in
g
p
u
r
p
o
s
es
b
y
s
p
litt
i
n
g
th
e
d
ata
i
n
to
8
0
%,
tr
ai
n
in
g
,
a
n
d
2
0
%,
f
o
r
test
in
g
.
T
h
e
d
ata
s
ets’
s
tatis
tic
s
ar
e
s
h
o
wn
in
Fig
u
r
e
1
.
Fig
u
r
e
1
.
Data
s
et’
s
s
tatis
tic
s
2
.
1
.
Co
nv
o
lutio
na
l neura
l net
wo
rk
s
C
o
m
p
ar
ed
t
o
h
a
n
d
-
c
r
af
ted
f
u
n
ctio
n
alities
,
C
NN
in
s
tan
tly
m
em
o
r
izes
f
u
n
ctio
n
s
f
o
r
lear
n
in
g
d
ee
p
v
is
u
al
v
ar
iatio
n
s
b
y
u
s
in
g
a
wid
e
v
ar
iety
o
f
tr
ain
in
g
d
ata
an
d
ca
n
ea
s
ily
eq
u
ate
its
test
p
r
o
ce
s
s
o
n
ac
ce
ler
atio
n
GPU
co
r
es.
T
h
e
f
o
llo
win
g
lay
er
s
ar
e
in
clu
d
ed
in
C
NN
ar
ch
itectu
r
e.
2
.
1
.
1.
Co
nv
o
lutio
na
l
la
y
er
Per
f
o
r
m
an
ce
o
v
er
t
he
in
p
u
t
is
tr
an
s
f
o
r
m
ed
in
to
g
r
o
u
n
d
b
r
ea
k
in
g
lay
er
s
.
E
n
ab
le
V
k
to
b
e
t
h
e
k
er
n
el
s
ize
n
,
m
f
ilter
ad
d
e
d
t
o
th
e
in
p
u
t
X
an
d
th
e
n
u
m
b
er
o
f
C
NN
n
eu
r
o
n
in
p
u
t
ties
is
n
an
d
m
.
T
h
e
r
esu
ltin
g
lay
e
r
o
u
tp
u
t is d
eter
m
i
n
ed
ac
co
r
d
in
g
ly
.
(
,
)
=
∑
∑
(
,
)
−
,
−
/
2
=
−
/
2
/
2
=
−
/
2
(
1
)
2
.
1
.
2
.
M
a
x
po
llin
g
Ma
x
Po
o
lin
g
is
a
co
n
v
o
lu
tio
n
m
eth
o
d
in
wh
ich
th
e
Ker
n
el
ta
k
es
th
e
h
ig
h
est
v
alu
e
f
r
o
m
th
e
r
eg
io
n
it
co
n
v
o
l
v
es.
Ma
x
Po
o
li
n
g
b
asically
tells
th
e
C
o
n
v
o
lu
tio
n
al
N
eu
r
al
Netwo
r
k
th
at
o
n
ly
th
at
i
n
f
o
r
m
atio
n
will
b
e
ca
r
r
ied
f
o
r
war
d
if
it
is
th
e
g
r
ea
test
in
f
o
r
m
atio
n
av
ailab
le
in
ter
m
s
o
f
am
p
litu
d
e.
B
y
u
s
in
g
th
e
m
ax
im
u
m
f
u
n
ctio
n
,
wh
er
e
m
is
th
e
f
ilter
s
ize
th
e
o
u
tp
u
t is ca
lcu
lated
a
f
o
llo
w
.
Ma
x
Po
o
lin
g
r
e
d
u
ce
s
X
i
in
p
u
t
.
(
)
=
{
+
1
,
+
1
|
|
≤
2
,
2
,
∈
(
2
)
2
.
1
.
3
.
Rec
t
if
ied linea
r
un
it
(
ReLU)
T
h
e
r
ec
tifie
d
lin
ea
r
ac
tiv
atio
n
u
n
it,
o
r
R
eL
U,
is
o
n
e
o
f
t
h
e
f
ew
lan
d
m
ar
k
s
in
th
e
d
ee
p
lear
n
in
g
r
ev
o
lu
tio
n
.
I
t'
s
s
im
p
le,
y
et
it
's
f
ar
s
u
p
er
io
r
to
r
ea
m
in
g
ac
tiv
atio
n
f
u
n
ctio
n
s
lik
e
s
ig
m
o
id
o
r
tan
h
.
R
eL
U
is
k
n
o
wn
to
b
e
u
s
ed
b
y
t
h
e
n
eu
r
a
l c
ell
to
ca
lcu
late
its
X,
o
u
tp
u
t
u
s
in
g
th
e
f
o
ll
o
win
g
ac
tiv
atio
n
f
u
n
ctio
n
.
R
(
X)
=
m
ax
(
0
,
x
)
(
3
)
2
.
1
.
4
.
F
ull
y
co
nn
ec
t
ed
la
y
er
(
F
C)
T
h
e
o
u
t
p
u
t
f
r
o
m
th
e
co
n
v
o
l
u
tio
n
al
lay
er
s
r
e
p
r
esen
ts
h
ig
h
-
lev
el
f
ea
tu
r
es
in
th
e
d
ata.
W
h
ile
th
at
o
u
tp
u
t
c
o
u
ld
b
e
f
latten
e
d
an
d
co
n
n
ec
ted
to
th
e
o
u
tp
u
t
lay
er
,
ad
d
in
g
a
f
u
lly
-
co
n
n
ec
ted
lay
er
is
a
b
est
way
o
f
lear
n
in
g
n
o
n
-
lin
ea
r
co
m
b
i
n
atio
n
s
o
f
th
es
e
f
ea
tu
r
es.
FC
lay
er
also
ter
m
ed
as
m
u
lti
-
lay
er
p
er
ce
p
tr
o
n
(
MLP
)
b
in
d
s
all
o
f
th
e
p
r
ec
e
d
in
g
l
ay
er
s
’
n
eu
r
o
n
s
to
ea
ch
n
e
u
r
o
n
o
f
its
lay
er
.
T
h
e
n
u
m
b
e
r
o
f
n
eu
r
o
n
s
in
th
e
co
m
p
letely
co
n
n
ec
ted
lay
e
r
w
as d
ef
in
ed
as X
with
a
s
ize
o
f
K
an
d
l
.
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.
25
,
No
.
3
,
Ma
r
ch
20
22
:
1
4
8
9
-
1
5
0
0
1492
f
(
x
)
=
σ
(
w
∗
x)
(
4
)
2
.
1
.
5
.
O
utput
la
y
er
T
h
er
e
ar
e
th
r
ee
s
o
r
ts
o
f
lay
er
s
in
a
s
tan
d
ar
d
n
eu
r
al
n
etwo
r
k
:
o
n
e
o
r
m
o
r
e
in
p
u
t
lay
er
s
,
o
n
e
o
r
m
o
r
e
h
id
d
en
la
y
er
s
,
an
d
o
n
e
o
r
m
o
r
e
o
u
tp
u
t
lay
er
s
.
Mo
r
e
ad
v
an
ce
d
,
n
o
v
el
n
eu
r
al
n
etwo
r
k
s
m
ay
in
clu
d
e
s
ev
er
al
lay
er
s
o
f
an
y
s
o
r
t,
an
d
ea
ch
la
y
er
m
a
y
b
e
d
esig
n
ed
d
i
f
f
er
e
n
t
ly
.
T
h
e
o
u
tp
u
t
lay
er
is
o
n
e
o
f
th
e
h
o
t
v
ec
to
r
s
th
at
r
ep
r
esen
t
th
e
in
p
u
t
im
ag
e
class
.
T
h
er
ef
o
r
e,
t
h
e
n
u
m
b
e
r
o
f
g
r
o
u
p
s
is
d
im
e
n
s
io
n
al.
T
h
e
o
u
tp
u
t
v
ec
to
r
class
X
is
d
er
iv
ed
.
c
(
x
)
=
{
i
|
ǁ
∃
A
j#
i
∶
x
j
≤
x
i
}
(
5
)
2
.
1
.
6
.
So
f
t
m
a
x
la
y
er
A
n
eu
r
al
n
etwo
r
k
'
s
ac
tiv
atio
n
f
u
n
ctio
n
is
an
ess
en
tial
co
m
p
o
n
en
t.
A
n
eu
r
al
n
etwo
r
k
is
a
b
a
s
ic
lin
ea
r
r
eg
r
ess
io
n
m
o
d
el
with
o
u
t
an
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
ac
t
iv
atio
n
f
u
n
ctio
n
o
f
f
er
s
th
e
n
eu
r
al
n
etwo
r
k
n
o
n
-
lin
ea
r
ity
.
T
h
e
f
au
lt
is
d
is
tr
ib
u
ted
b
ac
k
o
v
er
th
e
So
ft
-
m
a
x
.
E
n
ab
le
N
to
b
e
th
e
in
p
u
t
v
ec
to
r
d
im
en
s
io
n
s
an
d
s
o
f
t
-
m
ax
will th
en
ca
lcu
late
m
ap
p
in
g
t
o
,
(
)
=
∑
=
1
(
6
)
2
.
2
.
P
re
-
t
ra
ined
m
o
dels
f
o
r
cla
s
s
if
ica
t
io
n
W
e
d
ef
in
e
p
r
e
-
tr
ain
ed
m
o
d
el
ty
p
es
f
o
r
class
if
icatio
n
.
T
r
ain
in
g
wh
en
we
wer
e
allo
tted
to
a
d
ee
p
er
n
etwo
r
k
.
Fu
r
th
e
r
,
tr
ain
in
g
a
d
ee
p
er
n
etwo
r
k
is
co
m
p
licate
d
b
y
o
v
er
f
itt
in
g
an
d
co
n
v
er
g
en
ce
ef
f
ec
ts
.
T
h
ese
ty
p
es
o
f
is
s
u
es
ar
e
o
p
tim
ally
r
eso
lv
ed
in
th
e
tr
a
n
s
f
er
lear
n
in
g
ar
ch
itectu
r
es.
T
r
an
s
f
er
l
ea
r
n
in
g
m
o
d
els
ar
e
d
is
u
n
ited
in
to
f
o
u
r
s
tates:
in
s
tan
c
e,
p
ar
am
ete
r
-
b
ased
,
f
e
atu
r
e
-
r
ep
r
esen
tatio
n
,
a
n
d
r
el
atio
n
al
-
k
n
o
wl
ed
g
e
.
T
r
an
s
f
er
lear
n
i
n
g
d
e
p
lo
y
s
a
p
r
e
-
ex
is
tin
g
ar
ch
itectu
r
e,
tr
ain
ed
o
n
a
lar
g
e
d
ataset,
an
d
f
u
r
t
h
e
r
p
r
o
ce
s
s
in
g
in
t
h
e
n
ex
t
s
tag
es.
Ou
r
wo
r
k
g
u
a
r
a
n
tees
co
n
s
is
ten
cy
b
y
r
ed
u
cin
g
th
e
c
o
s
ts
o
f
tr
a
in
i
n
g
with
n
ew
m
o
d
els
o
f
d
ee
p
lear
n
in
g
.
I
n
th
is
ar
ticle,
th
e
to
p
f
iv
e
p
r
e
-
tr
ai
n
ed
m
o
d
els
f
o
r
im
ag
e
class
if
icatio
n
th
at
ar
e
s
tate
-
of
-
th
e
-
ar
t
an
d
ar
e
wid
ely
u
s
ed
in
th
e
in
d
u
s
tr
y
as
well
as
r
esear
ch
.
T
h
e
in
d
i
v
id
u
al
m
o
d
els
ar
e
ex
p
lain
ed
i
n
m
u
ch
m
o
r
e
d
etail
b
el
o
w.
T
h
at
is
in
ce
p
tio
n
-
v
3
,
V
GG
-
1
6
,
VGG
-
1
9
,
R
esNet5
0
an
d
E
f
f
icien
tNet.
2
.
2
.
1
.
I
ncept
io
n
T
h
e
im
ag
e
d
ataset
was
u
s
e
d
in
th
e
2
0
1
2
lar
g
e
s
ca
le
v
is
u
al
r
ec
o
g
n
itio
n
ch
alle
n
g
e
(
I
L
SVR
C
)
.
I
n
ce
p
tio
n
-
v
3
s
u
p
p
lies
elev
ated
ac
cu
r
ac
y
o
u
tco
m
es
in
co
m
p
ar
is
o
n
with
I
n
ce
p
tio
n
-
v
1
an
d
I
n
c
ep
tio
n
-
v
2
.
T
ab
le
1
s
h
o
ws
th
e
ar
ch
itectu
r
e
o
f
in
ce
p
tio
n
v
3
.
C
NN
is
th
e
co
n
v
o
lu
ti
o
n
lay
er
,
P
is
th
e
m
ax
im
u
m
-
p
o
o
lin
g
lay
er
,
MX
is
th
e
m
ix
ed
lay
er
an
d
F
is
th
e
f
u
lly
co
n
n
ec
ted
lay
er
.
I
n
ce
p
tio
n
v
3
p
a
r
am
eter
s
ar
e
less
th
an
2
5
M,
it
e
n
ab
les
p
at
h
f
u
s
io
n
,
len
g
t
h
o
f
f
ea
tu
r
e
2
0
4
8
,
an
d
th
e
tr
ai
n
in
g
s
et
co
n
tain
s
I
m
ag
e
n
et
1
2
6
M
im
ag
es
o
f
1
k
s
u
b
jects.
T
ab
le
1
.
I
n
ce
p
tio
n
V3
ar
ch
itec
tu
r
e
i
n
p
u
t
i
ma
g
e
si
z
e
A
r
c
h
i
t
e
c
t
u
r
e
2
2
9
x
229
x
3
C
N
N
0
,
C
N
N
1
:
C
2
:
6
4
x
3
x
3
,
P
0
:
3
x
3
C
N
N
3
:
8
0
x
1
x
1
,
C
N
N
4
:
1
9
2
x
3
x
3
P
1
:
3
x
3
,
M
0
:
3
5
x
3
5
x
2
5
6
M
X
1
,
M
X
2
:
3
5
x
3
5
x
2
8
8
M
X
3
,
M
X
4
:
1
7
x
1
7
x
7
6
8
M
X
5
,
M
X
6
,
M
X
7
:
1
7
x
1
7
x
7
6
8
,
M
X
7
:
1
7
x
1
7
x
7
6
8
,
M
X
8
:
8
x
8
x
1
2
8
0
M
X
9
,
M
X
1
0
:
8
x
8
x
1
2
8
0
,
P
3
:
8
x
8
,
F
:
1
0
0
0
2
.
2
.
2
.
Vis
ua
l g
eo
m
et
r
y
g
r
o
up
(
VG
G
)
VGG
is
an
o
b
ject
r
ec
o
g
n
itio
n
m
o
d
el
p
r
e
-
tr
ai
n
ed
,
s
u
ch
as
I
n
c
ep
tio
n
-
v
3
.
I
n
th
is
ar
ticle,
we
d
is
cu
s
s
two
v
ar
ian
ts
o
f
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
m
o
d
els
n
am
ely
-
VGG1
6
an
d
VGG1
9
.
B
o
th
u
s
e
s
f
ea
tu
r
es
ex
tr
ac
ted
f
r
o
m
th
e
p
r
e
-
tr
ain
ed
VGG1
6
an
d
VGG
1
9
n
etwo
r
k
s
tr
ain
e
d
o
n
th
e
I
m
a
g
eNe
t
d
ataset
f
o
r
d
etec
tin
g
f
ac
ial
ex
p
r
ess
io
n
s
.
T
h
e
in
p
u
t
g
r
e
y
s
ca
le
im
ag
es
ar
e
p
r
e
-
p
r
o
ce
s
s
ed
d
u
r
in
g
th
e
tr
ain
in
g
s
tep
b
y
p
er
f
o
r
m
in
g
s
tr
en
g
th
n
o
r
m
aliza
tio
n
a
n
d
r
esizin
g
o
n
th
e
p
ix
el
v
alu
es.
Ho
wev
er
,
VGG
s
h
o
u
ld
b
e
d
esig
n
ed
wit
h
2
2
4
x
2
2
4
x3
i
n
p
u
ts
,
we
s
p
ec
if
y
pooling
=
'avg'
b
ef
o
r
e
class
if
icatio
n
lay
e
r
(
th
at
w
ay
VGG
ca
n
h
an
d
le
i
n
p
u
ts
o
f
an
y
s
ize)
.
I
n
an
o
th
er
way
,
VGG1
6
s
u
p
p
o
r
ts
d
o
wn
to
4
8
x
4
8
im
a
g
es
as
an
in
p
u
t.
W
e
in
s
tan
tia
te
o
u
r
m
o
d
el
with
K
er
as.a
p
p
licatio
n
s
.
v
g
g
1
6
.
VG
G
16
(
in
clu
d
e_
to
p
=
T
r
u
e,
weig
h
ts
=‘
im
ag
en
et’
,
in
p
u
t_
s
h
ap
e=
(
7
2
,
7
2
,
3
)
)
an
d
th
e
n
ad
d
o
wn
m
o
d
el
h
ea
d
f
u
lly
co
n
v
o
lu
tio
n
al
o
r
d
en
s
e.
T
h
ese
im
ag
es
ar
e
g
iv
en
as
in
p
u
t
to
th
e
VGG
n
etwo
r
k
.
T
h
e
VGG
co
n
tain
s
f
iv
e
p
o
o
lin
g
la
y
er
s
.
Fin
e
-
tu
n
in
g
en
a
b
les
o
n
e
to
u
p
d
ate
th
e
m
o
d
el
ar
ch
itect
u
r
e
b
y
elim
in
atin
g
th
e
lay
er
h
ea
d
s
th
at
wer
e
p
r
e
v
io
u
s
ly
en
tire
ly
co
n
n
ec
ted
,
o
f
f
er
in
g
n
ew,
n
ewly
in
itialized
lay
er
s
,
an
d
t
r
ain
in
g
th
e
n
ew
FC
lay
er
s
to
p
r
ed
ict
o
u
r
in
p
u
t g
r
o
u
p
.
T
r
ain
in
g
m
in
i
m
izes th
e
av
er
ag
e
p
r
ed
ictio
n
a
n
d
lo
g
-
lo
s
s
af
ter
th
e
s
o
f
tm
ax
lay
er
.
C
NN
is
th
e
co
n
v
o
lu
tio
n
lay
er
,
P
is
th
e
m
ax
i
m
u
m
p
o
o
lin
g
lay
er
,
MX
is
th
e
m
ix
ed
lay
e
r
an
d
F
is
th
e
f
u
lly
co
n
n
ec
ted
lay
e
r
.
V
GG1
6
p
ar
a
m
eter
s
1
3
8
M,
it d
is
ab
les
p
ath
f
u
s
io
n
,
t
h
e
len
g
th
o
f
f
ea
tu
r
e
4
0
9
6
,
an
d
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
Lin
ea
r
fu
s
io
n
a
p
p
r
o
a
ch
to
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks fo
r
fa
cia
l
emo
tio
n
r
ec
o
g
n
itio
n
(
Usen
Du
d
ek
u
la
)
1493
tr
ain
in
g
s
et
co
n
tain
in
g
I
m
ag
e
Net
1
2
6
M
im
ag
es
o
f
1
k
s
u
b
je
cts.
VGG1
9
h
as
1
4
4
M
p
ar
am
e
ter
s
,
it
en
ab
les
p
ath
f
u
s
io
n
,
th
e
len
g
th
o
f
f
ea
tu
r
e
4
0
9
6
,
an
d
th
e
tr
ain
in
g
s
et
co
n
ta
in
s
I
m
ag
eNe
t o
f
1
2
6
M
im
ag
es
o
f
1
k
s
u
b
jects.
T
h
e
a
r
ch
itectu
r
e
o
f
VGG1
6
an
d
V
GG1
9
ar
ch
itectu
r
e
s
h
o
wn
in
T
ab
le
2
an
d
T
ab
le
3
.
T
ab
le
2
.
Vg
g
1
6
a
r
ch
itectu
r
e
I
n
p
u
t
i
m
a
g
e
s
i
z
e
A
r
c
h
i
t
e
c
t
u
r
e
2
2
4
X
2
2
4
X
3
C
N
N
0
,
1
:
6
4
x
3
x
3
P
0
:
3
x
3
,
C
N
N
2
,
3
:
1
2
8
x
1
2
8
x
3
,
P
1
:
3
x
3
C
N
N
4
,
5
:
2
5
6
x
2
5
6
x
3
C
N
N
6
:
2
5
6
x
2
5
6
x
3
,
C
N
N
7
:
5
1
2
x
5
1
2
x
3
,
P
2
:
3
x
3
C
N
N
8
:
2
5
6
x
2
5
6
x
3
,
C
N
N
9
:
5
1
2
x
5
1
2
x
3
,
P
3
:
3
x
3
C
N
N
1
0
,
1
1
:
5
1
2
x
5
1
2
x
3
C
N
N
1
2
:
5
1
2
x
5
1
2
x
3
,
F
0
,
F
1
:
4
0
9
6
,
F
2
:
1
0
0
0
T
ab
le
3
.
Vg
g
1
9
a
r
ch
itectu
r
e
I
n
p
u
t
i
m
a
g
e
s
i
z
e
A
r
c
h
i
t
e
c
t
u
r
e
2
2
4
X
2
2
4
X
3
C
N
N
0
,
1
:
6
4
x
3
x
3
P
0
:
3
x
3
,
C
N
N
2
,
3
:
1
2
8
x
1
2
8
x
3
,
P
1
:
3
x
3
C
N
N
4
,
5
,
6
,
7
:
2
5
6
x
2
5
6
x
3
P
2
:
3
x
3
,
C
N
N
8
:
5
1
2
x
5
1
2
x
3
,
C
N
N
9
,
1
0
:
5
1
2
x
5
1
2
x
3
,
C
N
N
1
1
,
1
2
:
5
1
2
x
5
1
2
x
3
,
P
3
:
3
x
3
C
N
N
1
3
,
1
4
:
5
1
2
x
5
1
2
x
3
C
N
N
1
5
:
5
1
2
x
5
1
2
x
3
,
F
0
,
1
:
4
0
9
6
,
F
2
:
1
0
0
0
2
.
2
.
3
.
ResNet
-
50
R
es
N
e
t
-
5
0
i
s
a
n
i
m
a
g
e
r
e
c
o
g
n
i
t
i
o
n
m
o
d
e
l
w
h
i
c
h
h
a
s
b
e
e
n
p
r
e
-
t
r
a
i
n
e
d
d
e
e
p
l
e
a
r
n
i
n
g
o
f
t
h
e
C
N
N
,
a
s
u
b
s
e
t
o
f
d
e
e
p
n
e
u
r
a
l
n
e
t
w
o
r
k
s
.
R
e
s
Ne
t
-
5
0
i
s
5
0
l
a
y
e
r
s
d
e
e
p
a
n
d
i
s
t
r
a
i
n
e
d
o
n
1
m
i
l
l
i
o
n
i
m
a
g
e
s
i
n
1
,
0
0
0
g
r
o
u
p
s
i
n
t
h
e
I
m
a
g
e
N
e
t
d
at
a
b
a
s
e
.
I
n
ad
d
i
t
i
o
n
,
t
h
e
a
l
g
o
r
i
t
h
m
h
as
m
o
r
e
t
h
a
n
2
3
m
i
l
l
i
o
n
w
o
r
k
a
b
l
e
p
a
r
a
m
e
t
e
r
s
,
s
u
g
g
e
s
ti
n
g
a
d
e
e
p
a
r
c
h
i
te
c
t
u
r
e
t
h
a
t
al
l
o
ws
i
m
a
g
e
r
e
c
o
g
n
i
t
i
o
n
ea
s
ie
r
.
A
h
i
g
h
l
y
e
f
f
i
c
i
e
n
t
s
o
l
u
t
i
o
n
is
a
p
r
e
q
u
a
l
i
f
i
e
d
m
o
d
e
l
.
S
p
e
c
i
f
i
ca
l
l
y
,
t
h
e
f
i
v
e
s
t
a
g
e
s
o
f
r
e
s
i
d
u
al
b
l
o
c
k
s
c
o
n
s
i
s
t
o
f
t
h
e
R
es
N
e
t
-
5
0
m
o
d
e
l
.
E
v
e
r
y
r
es
i
d
u
a
l
b
l
o
c
k
h
as
t
h
r
e
e
l
a
y
e
r
s
o
f
1
*
1
a
n
d
3
*
3
c
o
n
v
o
l
u
t
i
o
n
b
l
o
c
k
s
.
E
v
e
r
y
l
a
y
e
r
is
f
e
d
i
n
t
o
t
h
e
n
e
x
t
l
a
y
e
r
i
n
t
r
a
d
it
i
o
n
a
l
n
e
u
r
a
l
n
e
tw
o
r
k
s
.
E
a
c
h
l
a
y
e
r
i
n
a
r
e
s
i
d
u
al
b
l
o
c
k
n
e
t
w
o
r
k
f
l
o
w
s
i
n
t
o
t
h
e
n
e
x
t
la
y
e
r
,
c
a
l
l
e
d
t
h
e
i
d
e
n
ti
t
y
c
o
n
n
ec
t
i
o
n
s
;
r
o
u
g
h
l
y
2
–
3
h
o
p
a
p
a
r
t
.
T
h
e
n
o
t
at
i
o
n
k
×
k
,
n
i
n
t
h
e
C
N
N
l
a
y
e
r
b
l
o
c
k
s
h
o
w
s
t
h
e
f
il
t
e
r
s
iz
e
k
a
n
d
n
c
h
a
n
n
e
l
s
.
T
h
e
c
o
m
p
l
e
te
l
y
l
i
n
k
e
d
l
a
y
e
r
o
f
1
0
0
0
n
e
u
r
o
n
s
i
s
d
e
n
o
t
e
d
b
y
f
u
l
l
y
c
o
n
n
e
c
t
e
d
(
F
C
)
1
0
0
0
.
D
j
a
n
g
o
r
e
s
t
f
r
a
m
e
w
o
r
k
(
D
R
F
)
is
a
l
i
b
r
a
r
y
t
h
a
t
c
r
e
a
t
es
a
v
e
r
s
a
ti
l
e
a
n
d
e
f
f
i
c
i
e
n
t
a
p
p
l
i
c
at
i
o
n
p
r
o
g
r
a
m
m
i
n
g
i
n
t
e
r
f
a
c
e
u
s
i
n
g
Dj
an
g
o
b
a
s
i
c
m
o
d
e
ls
.
R
es
N
e
t
-
5
0
a
r
c
h
it
e
c
t
u
r
e
i
s
s
h
o
w
n
i
n
F
i
g
u
r
e
2
.
F
r
o
m
t
h
e
e
x
p
e
r
i
m
e
n
t
a
l
o
u
t
c
o
m
es
,
we
o
b
s
e
r
v
e
d
t
h
a
t
t
h
e
VG
G
1
9
m
o
d
e
l
a
c
h
i
e
v
e
d
h
i
g
h
a
cc
u
r
a
c
y
o
n
b
o
t
h
d
a
t
a
s
et
s
c
o
m
p
a
r
e
d
t
o
th
e
R
es
N
e
t
-
5
0
s
i
n
c
e
i
t
i
s
a
v
e
r
y
d
e
e
p
n
e
t
w
o
r
k
w
it
h
1
4
4
m
i
l
l
i
o
n
p
a
r
a
m
e
t
e
r
s
(
t
r
a
i
n
ab
l
e
a
n
d
n
o
n
-
t
r
a
i
n
a
b
l
e
)
.
D
u
e
t
o
t
h
i
s
b
e
h
a
v
i
o
r
V
G
G
n
e
t
w
o
r
k
e
f
f
ic
i
e
n
t
l
y
c
a
p
t
u
r
e
s
t
h
e
f
e
a
t
u
r
e
s
(
e
d
g
es
)
b
y
m
i
n
i
m
i
z
i
n
g
t
h
e
o
v
e
r
f
i
t
ti
n
g
p
r
o
b
l
e
m
.
E
f
f
i
c
ie
n
t
N
e
t
-
B
0
t
o
B
7
,
p
a
r
a
m
et
e
r
s
ar
e
5
.
3
m
i
l
l
i
o
n
t
o
6
6
m
i
l
l
i
o
n
,
w
e
t
a
k
e
t
o
p
t
h
r
e
e
a
c
cu
r
a
c
y
v
a
l
u
e
s
o
f
e
a
c
h
E
f
f
i
c
i
e
n
tN
e
t
-
B
0
t
o
B
7
,
t
h
e
a
c
c
u
r
a
c
y
a
ls
o
i
n
c
r
e
a
s
e
s
e
v
e
r
a
l
p
a
r
a
m
e
t
e
r
s
i
n
c
r
e
as
e
m
o
d
e
l
,
t
h
e
p
a
r
a
m
e
t
e
r
s
l
i
k
e
t
r
ai
n
a
b
l
e
a
n
d
n
o
n
-
t
r
a
i
n
a
b
l
e
c
a
n
c
h
a
n
g
e
s
a
c
c
o
r
d
i
n
g
t
o
m
o
d
e
l
w
e
i
g
h
t
s
,
b
i
as
,
f
o
r
p
r
e
-
t
r
a
i
n
e
d
m
o
d
e
l
m
o
s
t
o
f
t
h
e
c
a
s
es
i
t
w
ill
u
s
e
l
is
t
e
d
p
a
r
a
m
e
t
e
r
s
i
n
T
a
b
le
s
4
a
n
d
5
.
T
h
e
b
i
a
s
a
n
d
w
e
i
g
h
ts
i
n
i
t
i
al
l
y
t
a
k
e
r
a
n
d
o
m
v
a
l
u
e
s
,
t
h
e
n
we
u
p
d
a
t
e
u
s
in
g
t
h
e
b
a
c
k
p
r
o
p
a
g
a
t
i
o
n
al
g
o
r
i
th
m
.
T
h
e
s
e
l
e
a
r
n
e
d
p
a
r
a
m
e
t
e
r
s
a
r
e
u
p
d
a
te
d
w
ei
g
h
ts
c
o
r
r
e
ct
l
y
,
t
h
e
m
o
d
e
l
w
il
l
p
r
e
d
i
c
t
t
h
e
a
cc
u
r
a
t
e
p
r
e
d
ic
t
i
o
n
.
T
h
e
l
e
a
r
n
e
d
p
a
r
a
m
e
t
e
r
s
a
r
e
n
o
t
u
p
d
a
t
e
d
p
r
o
p
e
r
l
y
,
t
h
e
r
e
c
o
g
n
i
t
i
o
n
r
a
t
e
i
s
d
o
w
n
.
T
h
e
i
n
t
u
i
t
i
o
n
b
e
h
i
n
d
s
e
l
ec
t
i
n
g
I
m
a
g
e
N
e
t
a
n
d
n
o
t
s
e
l
e
ct
i
n
g
o
t
h
e
r
a
r
c
h
it
e
c
t
u
r
es
l
ik
e
R
es
N
e
t
-
5
0
,
M
o
b
il
e
N
E
t
,
a
n
d
E
f
f
i
c
i
e
n
t
N
E
t
.
R
e
s
N
e
t
,
M
o
b
il
eN
E
t
’
s
a
r
e
f
a
s
t
e
r
a
n
d
s
m
a
l
le
r
t
h
a
n
o
t
h
e
r
m
a
j
o
r
n
e
t
wo
r
k
s
l
i
k
e
V
GG
-
1
9
f
o
r
t
h
e
r
e
i
s
a
s
m
a
ll
t
r
a
d
e
-
o
f
f
a
n
d
t
h
a
t
t
r
ad
e
-
o
f
f
i
s
a
c
c
u
r
a
cy
c
o
m
p
a
r
e
d
t
o
I
m
a
g
e
N
e
t
,
M
o
b
i
le
N
E
t'
s
a
r
e
t
y
p
i
c
al
l
y
a
r
e
n
’
t
as
ac
c
u
r
a
t
e
b
u
t
r
e
l
at
i
v
e
l
y
s
m
a
l
l
ac
cu
r
a
t
e
r
e
d
u
c
t
i
o
n
,
t
o
p
S
t
a
t
e
-
of
-
t
h
e
-
A
r
t
p
r
e
-
t
r
a
i
n
e
d
m
o
d
e
l
s
f
o
r
i
m
a
g
e
cl
a
s
s
i
f
i
c
a
t
i
o
n.
Fig
u
r
e
2
.
R
esNet
-
5
0
ar
c
h
itectu
r
e
[
2
6
]
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.
25
,
No
.
3
,
Ma
r
ch
20
22
:
1
4
8
9
-
1
5
0
0
1494
T
ab
le
4
.
E
f
f
icien
t N
et’
s
f
am
il
y
ef
f
icien
cy
o
n
I
m
ag
eNe
t
M
o
d
e
l
To
p
a
c
c
u
r
a
c
y
P
a
r
a
me
t
e
r
s
Ef
f
i
c
i
e
n
t
N
e
t
-
B0
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B1
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B2
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B3
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B4
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B5
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B6
Ef
f
i
c
i
e
n
t
N
e
t
s
-
B7
7
7
.
3
%
7
9
.
2
%
8
0
.
3
%
8
1
.
7
%
8
3
.
0
%
8
3
.
7
%
8
4
.
2
%
8
4
.
2
%
5
.
3
M
7
.
8
M
9
.
2
M
1
2
M
1
9
M
3
0
M
4
3
M
6
6
M
T
ab
le
5
.
Per
tain
ed
m
o
d
els f
o
r
to
p
ac
cu
r
ac
y
M
o
d
e
l
To
p
a
c
c
u
r
a
c
y
P
a
r
a
me
t
e
r
s
R
e
sN
e
t
-
50
I
n
c
e
p
t
i
o
n
_
v
3
VGG
-
16
VGG
-
19
7
7
.
1
5
%
7
8
.
8
%
7
8
.
2
%
8
4
.
2
%
2
5
M
2
4
M
1
3
8
M
1
4
4
M
2
.
2
.
4
.
E
f
f
icient
Net
s
T
h
e
E
f
f
icien
tNet
ca
teg
o
r
y
co
n
s
is
ts
o
f
eig
h
t
m
o
d
els
f
r
o
m
B
0
to
B
7
,
ea
ch
with
a
co
r
r
esp
o
n
d
in
g
n
u
m
b
er
o
f
m
o
d
els
r
ef
er
r
in
g
to
v
ar
ian
ts
with
h
ig
h
er
p
ar
a
m
eter
s
an
d
g
r
ea
ter
ac
cu
r
ac
y
.
E
f
f
icien
tNet
is
a
co
m
m
u
n
ity
o
f
n
eu
r
al
n
etwo
r
k
m
o
d
els.
Sin
ce
2
0
1
2
,
th
e
I
m
a
g
eNe
t
d
ataset
h
as
ex
p
a
n
d
ed
a
s
it
h
as
g
r
o
wn
m
o
r
e
co
m
p
lex
.
Ho
wev
e
r
,
m
o
s
t
o
f
t
h
em
ar
e
n
o
t
ef
f
icien
t
in
ter
m
s
o
f
lo
ad
h
an
d
lin
g
.
Mo
r
e
ef
f
i
cien
t
m
eth
o
d
s
f
o
r
s
m
aller
m
o
d
els
h
a
v
e
b
ee
n
i
n
tr
o
d
u
ce
d
in
r
ec
en
t
y
ea
r
s
.
So
m
u
ch
s
o
th
at
wh
en
s
ca
lin
g
d
o
wn
th
e
m
o
d
el,
s
ca
lin
g
is
p
er
f
o
r
m
e
d
o
n
d
ep
th
,
d
is
tan
ce
,
an
d
r
eso
lu
tio
n
f
o
cu
s
in
g
o
n
all
th
r
ee
in
c
o
m
b
in
atio
n
h
as
y
ield
e
d
m
o
r
e
ef
f
icien
t r
esu
lts
.
3.
P
RO
P
O
SE
D
M
O
D
E
L
T
h
is
wo
r
k
p
r
o
p
o
s
es
a
n
ew
a
p
p
r
o
ac
h
f
o
r
f
ac
ial
em
o
tio
n
r
ec
o
g
n
itio
n
u
s
in
g
a
lin
ea
r
f
u
s
io
n
tech
n
iq
u
e.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
co
n
s
is
ts
o
f
th
e
f
o
llo
win
g
s
tep
s
:
f
ea
t
u
r
e
ex
t
r
ac
tio
n
f
r
o
m
C
NN
an
d
f
ea
tu
r
e
ex
t
r
ac
tio
n
f
r
o
m
v
g
g
1
9
,
a
n
d
f
in
ally
f
u
s
io
n
o
f
d
if
f
er
en
t
o
u
tp
u
ts
.
R
ec
o
g
n
izin
g
f
ac
ial
em
o
tio
n
s
with
th
e
h
el
p
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el.
W
h
en
it
co
m
es
to
f
ac
e
d
etec
tio
n
,
th
e
p
r
e
s
en
t
in
v
esti
g
atio
n
u
s
es
th
e
wi
d
ely
u
til
i
z
ed
Vio
la
-
J
o
n
es
f
r
am
ewo
r
k
[
2
7
]
.
T
h
e
e
n
tire
m
et
ho
d
o
lo
g
y
is
d
ep
icted
in
Fig
u
r
e
3
.
T
h
e
C
K+
[
2
8
]
an
d
th
e
J
AFFE
[
2
9
]
f
ac
ial
p
ictu
r
es
ar
e
u
s
ed
i
n
b
en
ch
m
ar
k
in
g
d
atasets
,
an
d
r
ea
l
en
v
ir
o
n
m
en
ts
v
ar
y
in
r
o
tatio
n
,
ev
en
f
o
r
th
e
s
am
e
s
u
b
ject.
T
h
is
v
ar
iatio
n
is
in
d
ep
en
d
en
t
o
f
f
ac
ial
em
o
tio
n
s
an
d
m
ay
im
p
a
ct
th
e
ac
cu
r
ac
y
o
f
FER
's
r
ec
o
g
n
itio
n
.
T
h
e
f
ac
e
ar
ea
is
alig
n
e
d
u
s
in
g
r
o
tatio
n
r
ec
tific
atio
n
to
s
o
lv
e
t
h
is
p
r
o
b
lem
.
[R
X,
R
Y,
1
]
=[
R
x
'
,
R
y
’
,
1
]
=
[
c
os
θ
s
in
θ
0
−
s
in
θ
c
os
θ
0
0
0
1
]
(
7
)
Fig
u
r
e
3
.
Pro
p
o
s
ed
m
o
d
el
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
Lin
ea
r
fu
s
io
n
a
p
p
r
o
a
ch
to
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks fo
r
fa
cia
l
emo
tio
n
r
ec
o
g
n
itio
n
(
Usen
Du
d
ek
u
la
)
1495
I
n
wh
ich
(
R
x
,
R
y
)
r
ep
r
esen
t
s
th
e
in
itial
p
o
in
t
in
th
e
f
ac
e
p
ictu
r
e,
an
d
(
R
x
'
,
R
y
'
)
r
ep
r
esen
ts
th
e
co
o
r
d
in
ate
(
x
,
y
)
af
ter
th
e
r
o
ta
tio
n
tr
an
s
f
o
r
m
atio
n
is
p
e
r
f
o
r
m
ed
.
θ
d
en
o
tes
th
e
an
g
le
o
f
r
o
ta
tio
n
f
r
o
m
o
n
e
ey
e
ce
n
t
er
to
th
e
o
th
er
,
as
m
ea
s
u
r
ed
f
r
o
m
th
e
ce
n
t
e
r
o
f
th
e
ey
e
.
O
n
th
e
h
o
r
izo
n
tal
ax
is
,
th
e
a
x
is
s
tar
tin
g
v
alu
e
is
ze
r
o
,
a
ll
id
en
tifie
d
f
ac
e
a
r
ea
s
ar
e
r
ed
u
ce
d
t
o
7
2
×
7
2
p
ix
els
af
ter
r
o
tatio
n
r
ec
tific
atio
n
.
W
h
ile
a
s
m
aller
f
ac
e
ar
ea
m
ig
h
t
in
cr
ea
s
e
t
h
e
s
p
ee
d
o
f
FER,
it
ca
n
also
lea
d
to
lo
s
in
g
f
ac
ial
f
ea
tu
r
es,
esp
ec
ially
f
o
r
t
h
e
in
f
o
r
m
atio
n
ac
q
u
ir
e
d
f
r
o
m
f
ac
ial
L
B
P
im
ag
es,
r
elate
d
f
ac
ial
r
eg
io
n
s
,
s
u
ch
as
m
o
u
t
h
s
,
ey
es,
an
d
e
y
eb
r
o
ws,
ar
e
m
o
r
e
r
em
ar
k
ab
le
i
n
L
B
P im
ag
es th
an
in
g
r
a
y
s
ca
le
im
ag
es.
Ou
r
d
esig
n
is
to
am
en
d
th
e
ex
ac
tn
ess
o
f
em
o
tio
n
ex
p
r
ess
io
n
cla
s
s
if
ica
tio
n
b
y
m
o
d
er
n
co
n
v
o
lu
tio
n
al
n
etwo
r
k
ar
c
h
itectu
r
e.
I
n
clu
d
in
g
g
r
ay
im
ag
es
an
d
L
B
P
im
ag
es,
will
b
e
u
s
ed
b
y
FER
u
s
in
g
co
n
v
o
l
u
tio
n
n
e
u
r
a
l
n
etwo
r
k
s
.
Activ
atio
n
s
o
f
r
ec
t
if
ied
lin
ea
r
u
n
it
(
R
eL
u
)
af
ter
ea
ch
co
n
v
o
l
u
tio
n
lay
e
r
ar
e
a
d
d
ed
a
n
d
s
o
f
t
-
m
ax
class
if
ie
r
s
ar
e
u
s
ed
f
o
r
an
ac
tiv
atio
n
f
u
n
ctio
n
in
th
e
f
latten
in
g
lay
er
.
T
h
e
f
ir
s
t
s
tep
o
f
f
ea
t
u
r
e
ex
tr
ac
tio
n
f
r
o
m
co
n
v
o
l
u
tio
n
n
e
u
r
al
n
etwo
r
k
a
r
ch
itectu
r
e
is
s
h
o
wn
i
n
T
ab
le
6
.
T
h
e
n
,
VGG
1
9
is
ch
o
s
en
f
o
r
Facial
f
ea
tu
r
es
ex
tr
ac
tio
n
,
a
p
p
ly
in
g
tr
an
s
f
er
l
ea
r
n
in
g
o
u
r
p
ar
tial
VGG1
9
n
etwo
r
k
is
tr
ain
ed
o
n
th
e
Flatten
m
o
d
el
I
m
ag
eNe
t
d
ataset
with
a
d
r
o
p
o
u
t
r
atio
o
f
0
.
5
,
6
4
d
en
s
e
la
y
er
s
o
f
R
eL
U
ar
ch
itectu
r
e,
an
d
s
ev
en
ac
tiv
atio
n
p
o
in
ts
.
So
f
tm
ax
class
if
ier
is
d
ep
lo
y
ed
in
th
e
s
am
e.
T
h
e
in
p
u
t
g
r
ey
s
ca
le
im
ag
es
ar
e
p
r
e
-
p
r
o
ce
s
s
ed
d
u
r
in
g
th
e
tr
ain
in
g
s
tep
b
y
p
er
f
o
r
m
in
g
s
tr
en
g
th
n
o
r
m
aliza
tio
n
a
n
d
r
esizin
g
o
n
t
h
e
p
ix
el
v
alu
es.
T
h
ese
im
ag
es
ar
e
g
iv
en
as
in
p
u
t
to
th
e
VGG
n
etwo
r
k
.
T
h
e
VG
G
co
n
tain
s
f
iv
e
p
o
o
lin
g
la
y
er
s
.
R
ath
er
th
an
u
s
in
g
VGG1
6
,
it
is
o
p
tim
al
to
u
s
e
VGG1
9
(
VGG1
9
h
as m
o
r
e
m
e
m
o
r
y
)
,
an
d
t
h
e
b
est p
er
f
o
r
m
a
n
ce
.
T
h
e
in
p
u
t
d
ata
h
as
a
s
ize
o
f
1
x
7
2
x
7
2
.
Af
ter
th
at,
we'
ll
wo
r
k
o
n
th
e
f
ir
s
t
f
o
u
r
b
lo
c
k
s
.
T
h
e
f
if
th
b
lo
ck
C
o
n
v
5
_
1
f
t,
Su
m
m
ar
izes
th
e
p
ar
am
eter
s
f
o
r
th
e
la
y
er
s
in
th
is
b
lo
ck
,
wh
ich
ar
e
m
e
n
tio
n
ed
in
T
ab
le
7
.
T
h
e
lear
n
in
g
r
ates
o
f
th
e
f
if
th
b
lo
c
k
'
s
lay
er
s
ar
e
(
0
.
0
0
1
u
s
ed
f
o
r
lay
er
s
o
f
th
e
f
if
th
b
lo
ck
)
th
a
n
th
eir
o
r
ig
in
al
v
alu
es
(
0
.
0
1
u
s
ed
f
o
r
lay
e
r
s
o
f
p
r
ev
i
o
u
s
b
lo
ck
s
)
to
en
s
u
r
e
th
at
th
e
y
ca
n
lear
n
m
o
r
e
ef
f
ec
tiv
e
in
f
o
r
m
atio
n
.
B
ec
au
s
e
we
d
ec
r
ea
s
in
g
1
0
tim
es
th
e
o
r
ig
in
al
v
alu
e
.
W
eig
h
t
d
ec
ay
,
wh
ich
r
e
d
u
ce
s
y
o
u
r
co
ef
f
icien
ts
to
ze
r
o
,
g
u
a
r
an
tees
th
at
y
o
u
r
ea
ch
a
lo
ca
l
o
p
tim
u
m
with
s
m
all
-
m
ag
n
itu
d
e
p
a
r
a
m
eter
s
.
T
h
is
is
cr
itical
to
p
r
e
v
en
t
an
o
v
e
r
-
f
itti
n
g
s
itu
atio
n
.
I
n
ad
d
itio
n
,
b
y
in
cr
ea
s
in
g
th
e
co
n
v
ex
ity
o
f
th
e
o
b
jectiv
e
f
u
n
ctio
n
,
th
e
m
o
d
el
b
ec
o
m
es
ea
s
ier
to
o
p
tim
i
z
e.
W
h
en
it
co
m
es
to
o
p
tim
al
weig
h
t
d
ec
ay
,
it
d
ep
e
n
d
s
o
n
th
e
t
o
tal
n
u
m
b
er
o
f
b
at
ch
r
u
n
s
a
n
d
weig
h
t
u
p
d
ates.
T
ab
le
6
.
C
NN
ar
ch
itectu
r
e
f
o
r
im
ag
e
class
if
icatio
n
La
y
e
r
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
S
t
r
i
d
e
A
c
t
i
v
a
t
i
o
n
F
u
n
c
t
i
o
n
C
o
n
v
o
l
u
t
i
o
n
(
C
1
)
64
7
x
7
4
x
4
Re
LU
P
o
o
l
i
n
g
3
x
3
2
x
2
C
o
n
v
o
l
u
t
i
o
n
(
C
2
)
32
5
x
5
1
x
1
Re
LU
C
o
n
v
o
l
u
t
i
o
n
(
C
3
)
64
5
x
5
1
x
1
Re
LU
C
o
n
v
o
l
u
t
i
o
n
(
C
4
)
32
5
x
5
1
x
1
Re
LU
P
o
o
l
i
n
g
3
x
3
2
x
2
F
l
a
t
t
e
n
so
f
t
m
a
x
T
ab
le
7
. P
ar
am
eter
s
f
o
r
th
e
p
r
o
p
o
s
ed
m
o
d
el
P
a
r
a
me
t
e
r
s
V
a
l
u
e
Le
a
r
n
i
n
g
r
a
t
e
0.
00
1
W
e
i
g
h
t
D
e
c
a
y
0
.
0
1
M
o
me
n
t
u
m
0.
0
0
1
O
p
t
i
mi
z
e
r
A
d
a
m
T
h
is
is
cr
itical
to
p
r
e
v
en
t
a
n
o
v
er
-
f
itti
n
g
s
itu
atio
n
.
I
n
ad
d
itio
n
,
b
y
in
cr
ea
s
in
g
th
e
co
n
v
ex
ity
o
f
th
e
o
b
jectiv
e
f
u
n
ctio
n
,
th
e
m
o
d
el
b
ec
o
m
es
ea
s
ier
to
o
p
tim
i
z
e.
W
h
en
it
co
m
es
to
o
p
tim
al
wei
g
h
t
d
ec
a
y
,
it d
ep
en
d
s
o
n
th
e
to
tal
n
u
m
b
e
r
o
f
b
atch
e
s
r
u
n
s
an
d
weig
h
t
u
p
d
ates.
Acc
o
r
d
in
g
to
o
u
r
em
p
ir
ical
r
esear
ch
o
f
Ad
am
,
th
e
lo
wer
th
e
o
p
tim
u
m
weig
h
t
d
e
ca
y
is
,
th
e
lo
n
g
e
r
th
e
r
u
n
tim
e/
n
u
m
b
er
o
f
b
atc
h
r
u
n
s
ar
e.
Mo
m
en
tu
m
is
u
tili
z
ed
to
r
ed
u
ce
wei
g
h
t
c
h
an
g
e
v
ar
i
atio
n
s
ac
r
o
s
s
s
u
cc
ess
iv
e
iter
atio
n
s
,
an
d
it
s
tar
ted
at
a
ze
r
o
in
itial
v
alu
e
.
T
h
e
p
ar
tial
VGG1
9
n
etwo
r
k
i
s
u
s
ed
to
ex
tr
ac
t
an
ex
p
r
ess
io
n
r
e
lated
f
ea
tu
r
e
v
ec
to
r
f
r
o
m
f
ac
e
g
r
ay
s
ca
le
im
ag
es.
T
h
e
C
NN
is
u
s
ed
t
o
ex
t
r
ac
t
a
f
ea
tu
r
e
v
ec
to
r
f
r
o
m
L
B
P
f
a
ce
p
ictu
r
es.
T
h
er
e
ar
e
two
ca
s
ca
d
ed
f
u
ll
-
c
o
n
n
ec
t
lay
er
s
o
n
ea
ch
f
ea
tu
r
e
v
ec
to
r
to
r
ed
u
ce
th
e
s
ize.
B
o
th
tr
ain
i
n
g
m
o
d
els
o
b
tain
co
r
r
esp
o
n
d
i
n
g
m
o
d
el
wei
g
h
ts
.
C
o
r
r
esp
o
n
d
in
g
weig
h
ts
to
lin
ea
r
f
u
s
io
n
ap
p
r
o
ac
h
to
n
ew
m
o
d
el
C
NN
ar
ch
itectu
r
e.
C
NN
i
s
th
e
m
o
s
t
co
m
m
o
n
n
etwo
r
k
m
o
d
el
am
o
n
g
t
h
e
m
an
y
p
o
s
s
ib
le
d
ee
p
lear
n
in
g
m
o
d
els.
T
h
e
f
in
al
C
NN
ar
ch
itectu
r
e
is
s
h
o
wn
in
T
ab
le
8
.
T
h
e
VGG
1
9
s
y
s
tem
u
s
es
a
tu
n
in
g
s
y
s
te
m
t
o
ex
tr
ac
t
em
o
tio
n
s
f
r
o
m
g
r
ay
-
lev
el
im
ag
es.
T
h
e
v
ec
to
r
is
d
er
iv
ed
b
y
C
NN
f
r
o
m
L
B
P
f
ac
e
im
ag
es.
T
h
e
two
ca
s
ca
d
e
d
f
u
lly
co
n
n
ec
ted
lay
e
r
s
ar
e
co
m
m
u
n
icate
d
b
y
ea
c
h
v
ec
to
r
.
VGG1
9
an
d
C
NN
b
o
th
co
n
s
tr
u
ct
a
f
u
s
ed
v
ec
to
r
=
{
1
,
2
7
}
.
T
h
e
f
ir
s
t
elem
en
t
is
f
o
llo
w
e
d
by
(
8
)
,
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.
25
,
No
.
3
,
Ma
r
ch
20
22
:
1
4
8
9
-
1
5
0
0
1496
F
u
= α
.
S
i
(
1
−
α
)
l
i
(
8
)
w
h
er
e
α
weig
h
ts
o
f
th
e
g
r
a
y
-
s
ca
le
im
ag
es.
E
v
alu
ated
b
y
cr
o
s
s
-
v
alid
atio
n
.
T
h
e
ca
te
g
o
r
ical
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
o
f
s
o
f
t
-
m
ax
,
th
e
in
p
u
t is a
s
et
o
f
m
u
lti
-
class
.
=
∑
=
1
(
9
)
T
h
e
co
s
t f
u
n
ctio
n
,
wh
ich
is
d
e
f
in
ed
b
y
f
(
y
=
k
/
x
)
.
(
,
)
=
−
∑
.
l
og
(
)
=
1
(
10
)
T
r
u
e
lab
el
Z
i
in
d
icate
s
an
d
Y
i
is
th
e
o
u
tp
u
t
o
f
a
s
o
f
t
-
m
ax
f
u
n
ctio
n
.
B
ac
k
p
r
o
p
ag
atio
n
is
b
ased
o
n
th
e
o
p
tim
izatio
n
alg
o
r
ith
m
o
f
g
r
a
d
ien
t d
escen
t.
T
ab
le
8
.
Pro
p
o
s
ed
C
NN
ar
ch
it
ec
tu
r
e
La
y
e
r
F
i
l
t
e
r
s
K
e
r
n
e
l
S
i
z
e
D
r
o
p
o
u
t
A
c
t
i
v
a
t
i
o
n
F
u
n
c
t
i
o
n
C
o
n
v
o
l
u
t
i
o
n
(
C
1
)
6
5
x
5
Re
LU
P
o
o
l
i
n
g
2
x
2
C
o
n
v
o
l
u
t
i
o
n
(
C
2
)
16
5
x
5
Re
LU
P
o
o
l
i
n
g
2
x
2
C
o
n
v
o
l
u
t
i
o
n
(
C
3
)
1
2
0
5
x
5
0
.
2
5
Re
LU
F
l
a
t
t
e
n
D
e
n
se
(
8
4
)
0
.
5
R
e
LU
D
e
n
se
l
a
y
e
r
S
o
f
t
ma
x
T
h
e
n
o
v
elty
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
is
lin
ea
r
w
eig
h
ted
f
u
s
io
n
is
a
p
r
o
ce
d
u
r
e
wh
er
e
th
e
r
esu
ltin
g
f
u
s
ed
p
ictu
r
e
is
m
o
r
e
in
f
o
r
m
ativ
e
an
d
c
o
m
p
lete
t
h
an
an
y
o
f
th
e
in
p
u
t
o
f
th
e
p
i
ctu
r
e
b
y
co
m
b
in
in
g
r
elev
an
t
in
f
o
r
m
atio
n
f
r
o
m
a
s
et
o
f
im
ag
es
i
n
to
th
e
in
d
iv
id
u
al
p
ictu
r
e.
I
m
ag
e
f
u
s
io
n
tec
h
n
iq
u
es
ca
n
im
p
r
o
v
e
th
e
q
u
ality
o
f
a
p
p
licatio
n
d
ata.
T
h
e
m
er
g
er
tec
h
n
iq
u
e
was
u
s
ed
to
in
teg
r
ate
VGG
-
1
9
a
n
d
C
NN
d
ec
is
io
n
s
.
T
h
e
f
u
s
io
n
alg
o
r
ith
m
ca
lcu
lates
t
h
e
class
b
y
ta
k
in
g
an
av
er
ag
e
o
f
ea
ch
class
if
icatio
n
d
ec
is
io
n
o
n
tr
ai
n
in
g
s
am
p
les.
Mo
r
e
r
eliab
le
(
h
ig
h
er
p
r
ec
is
io
n
)
class
if
icatio
n
is
weig
h
ed
an
d
s
ig
n
if
ican
tly
co
n
tr
ib
u
tes
t
o
d
ec
is
io
n
m
ak
in
g
.
W
e
h
av
e
u
s
ed
th
e
f
u
s
io
n
r
u
le
o
f
th
e
weig
h
ted
s
u
m
to
e
v
al
u
ate
th
e
b
est
o
p
p
o
r
tu
n
ity
o
f
a
f
u
s
io
n
r
esu
lt
o
f
a
p
ar
ticu
lar
em
o
tio
n
,
co
m
p
ar
is
o
n
m
u
ltimo
d
al
with
u
n
im
o
d
al
m
o
d
els.
Du
e
to
th
e
h
ig
h
ca
p
a
city
o
f
d
ee
p
C
NN,
th
e
av
er
ag
e
f
ea
tu
r
e
f
u
s
io
n
m
o
d
el
im
p
r
o
v
es
th
e
p
er
f
o
r
m
an
ce
s
u
b
s
tan
tiv
ely
.
T
ak
i
n
g
an
a
v
er
ag
e
o
f
m
u
ltip
le
m
o
d
els
wi
ll
r
ed
u
ce
th
e
v
ar
ia
n
ce
.
T
h
is
s
h
o
ws
th
at
m
u
ltimo
d
al
o
f
f
er
s
a
v
er
y
ef
f
ec
tiv
e
way
o
f
im
p
r
o
v
in
g
ac
cu
r
ac
y
r
ates,
as o
r
i
g
in
ally
p
r
o
p
o
s
ed
.
4.
E
XP
E
R
I
M
E
N
T
A
L
RE
SUL
T
S AN
D
A
NALY
SI
S
B
u
ilt
o
n
th
e
T
en
s
o
r
f
lo
w
s
y
s
t
em
,
we
test
th
e
ac
h
iev
em
en
t
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
o
n
th
e
wi
n
d
o
ws
in
ter
n
et
Go
o
g
le
C
o
lab
p
latf
o
r
m
.
T
wo
p
u
b
lically
av
ailab
le
d
atasets
ar
e
u
s
ed
f
o
r
f
ac
ial
em
o
tio
n
im
ag
es.
T
h
e
r
esu
lts
ar
e
illu
s
tr
ated
in
Fig
u
r
es
4
(
a)
an
d
(
b
)
.
Dis
p
lay
in
g
p
r
ec
is
io
n
cu
r
v
e
b
y
r
ed
an
d
l
o
s
s
b
y
g
r
ee
n
an
d
s
tab
ilizes
lo
s
s
f
o
llo
win
g
2
5
to
5
0
ep
o
c
h
s
,
an
d
th
eir
test
r
esu
lts
ar
e
also
s
h
o
wn
.
T
h
e
av
er
ag
e
ac
cu
r
ac
y
o
f
r
ec
o
g
n
itio
n
f
o
r
'
C
K+
an
d
J
AF
FE'
d
ata
s
ets
i
s
9
8
.
3
%
an
d
9
2
.
4
%,
r
esp
ec
tiv
ely
.
W
e
also
m
ea
s
u
r
e
o
u
r
p
r
o
ce
s
s
b
y
ass
ess
in
g
ac
cu
r
ac
ies
b
ased
o
n
s
in
g
le
-
ch
a
n
n
el
f
a
cial
im
a
g
es.
W
e
test
th
e
f
ea
s
ib
ilit
y
o
f
o
u
r
m
eth
o
d
o
lo
g
y
.
C
o
m
p
ar
ed
to
o
th
er
C
NN
r
elate
d
ap
p
r
o
ac
h
es,
o
u
r
a
p
p
r
o
ac
h
p
r
o
d
u
ce
s
im
p
r
o
v
ed
e
f
f
icien
cy
with
f
ea
tu
r
es
lik
e
h
is
to
g
r
am
o
f
o
r
ien
ted
g
r
ad
ien
ts
(
HOG)
,
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
,
an
d
K
-
NN
Ou
r
m
eth
o
d
'
s
b
en
ef
it
is
ac
co
m
p
lis
h
ed
b
y
m
ak
in
g
g
o
o
d
u
s
e
o
f
th
e
c
o
m
p
lem
e
n
tar
y
o
f
v
ar
io
u
s
f
ac
ial
im
ag
e
s
o
u
r
c
es,
wh
ile
th
e
o
th
er
s
tr
ateg
y
o
n
l
y
u
s
es
FER
ap
p
r
o
ac
h
es.
C
o
m
p
ar
is
o
n
s
b
etwe
en
o
u
r
a
p
p
r
o
ac
h
an
d
th
e
o
th
er
s
tate
-
of
-
th
e
-
a
r
t
FER
ap
p
r
o
ac
h
es
ar
e
s
h
o
wn
in
T
ab
l
e
9
.
4
.
1
.
So
m
e
o
f
t
he
cla
s
s
if
ica
t
io
n f
a
ults a
re
dis
cus
s
ed
T
h
e
SVM
alg
o
r
ith
m
is
n
o
t
s
u
i
tab
le
f
o
r
la
r
g
e
s
ets
o
f
d
ata.
SVM
d
o
es
n
o
t
d
o
t
h
at
well
as
th
e
r
e
is
m
o
r
e
n
o
is
e
in
th
e
co
llectio
n
,
i.e
.
o
v
er
lap
o
f
tar
g
et
g
r
o
u
p
s
.
I
n
th
e
s
ce
n
ar
io
s
wh
er
e
th
e
n
u
m
b
er
o
f
f
ea
tu
r
es
f
o
r
ea
c
h
d
ata
p
o
in
t
ex
ce
ed
s
th
e
n
u
m
b
e
r
o
f
tr
ain
in
g
d
ata
s
am
p
les,
t
h
e
SVM
wo
u
ld
b
e
u
n
d
er
p
e
r
f
o
r
m
in
g
.
An
d
th
e
SVM
Mo
d
el
is
a
r
ep
r
esen
tatio
n
o
f
ex
am
p
les
as
d
o
ts
in
s
p
ac
e,
m
ap
p
ed
in
s
u
c
h
a
way
th
at
th
e
ex
am
p
les
o
f
th
e
Sep
ar
ate
ca
teg
o
r
ies
o
r
class
es
ar
e
class
if
ied
in
to
two
ca
teg
o
r
ies.
Div
is
io
n
o
f
th
e
p
lan
e
t
h
at
m
ax
im
izes
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
Lin
ea
r
fu
s
io
n
a
p
p
r
o
a
ch
to
co
n
vo
lu
tio
n
a
l n
eu
r
a
l n
etw
o
r
ks fo
r
fa
cia
l
emo
tio
n
r
ec
o
g
n
itio
n
(
Usen
Du
d
ek
u
la
)
1497
m
ar
g
in
b
etwe
en
th
e
two
th
er
e
is
v
ar
io
u
s
g
r
o
u
p
s
.
T
h
is
is
b
ec
au
s
e
th
e
s
ep
ar
atio
n
p
lan
e
h
as
th
e
wid
est
d
is
tan
ce
to
th
e
L
o
wer
s
,
th
e
cl
o
s
est tr
ain
in
g
d
ata
p
o
in
ts
in
an
y
class
.
E
r
r
o
r
in
a
g
en
er
aliza
tio
n
o
f
t
h
e
t
o
tal
class
if
ier
.
HOG
h
as
a
d
ec
en
t
s
co
r
e
f
o
r
h
u
m
an
id
e
n
tific
atio
n
.
Ho
w
ev
er
,
it
h
as
a
d
r
awb
ac
k
th
a
t
is
v
er
y
s
u
s
ce
p
tib
le
to
th
e
r
o
tatio
n
o
f
th
e
im
a
g
e.
KNN
-
Pre
cisi
o
n
r
elies
o
n
d
ata
ac
c
u
r
ac
y
.
W
ith
lar
g
e
d
ata,
t
h
e
p
r
ed
ictiv
e
s
tag
e
ca
n
b
elo
n
g
.
R
esp
o
n
s
iv
e
to
th
e
d
ata
s
ize
a
n
d
in
s
ig
n
if
ican
t
f
u
n
ctio
n
s
.
T
h
e
KNN
alg
o
r
ith
m
'
s
d
r
awb
ac
k
is
th
at
it
u
s
es
b
o
t
h
E
q
u
al
ch
a
r
ac
ter
is
tics
with
co
r
r
e
latio
n
s
in
co
m
p
u
tatio
n
.
T
h
is
c
an
lead
t
o
m
is
tak
es
in
class
if
icatio
n
,
p
ar
ticu
lar
ly
if
th
er
e
is
ju
s
t a
s
m
all
s
u
b
s
et
o
f
tr
aits
th
at
a
r
e
h
elp
f
u
l
f
o
r
class
if
icatio
n
.
(a
)
(
b
)
Fig
u
r
e
4
.
J
AFFE
tr
ain
in
g
ac
cu
r
ac
y
an
d
v
alid
atio
n
l
o
s
s
,
(
a)
an
d
th
e
ir
test
an
aly
s
is
test
r
esu
lts
,
C
K+
tr
ain
in
g
ac
cu
r
ac
y
an
d
v
alid
atio
n
lo
s
s
an
d
(
b
)
th
e
ir
test
an
aly
s
is
test
r
esu
lts
.
Nu
m
b
er
o
f
e
p
o
ch
s
in
x
a
x
is
,
tr
ain
an
d
test
lo
s
s
,
th
eir
ac
cu
r
ac
y
(
%)
i
n
y
a
x
is
T
ab
le
9
.
C
o
m
p
a
r
is
o
n
o
f
r
ec
o
g
n
itio
n
r
ate
with
s
o
m
e
o
f
th
e
e
x
is
tin
g
tech
n
iq
u
es
M
e
t
h
o
d
JA
F
F
E
D
a
t
a
b
a
se
C
K
+
D
a
t
a
b
a
s
e
LB
P
+
O
R
B
f
e
a
t
u
r
e
s
[
3
0
]
8
8
.
5
5
%
F
i
sh
e
r
f
a
c
e
[
3
1
]
8
9
.
2
%
D
e
e
p
F
e
a
t
u
r
e
s
+
H
O
G
[
3
2
]
9
0
.
5
8
%
S
u
n
e
t
a
l
.
[
3
3
]
9
2
.
0
0
%
C
N
N
[
3
4
]
7
6
.
5
%
Pr
o
p
o
sed
M
o
d
e
l
9
2
.
4
%
LB
P
+
O
R
B
f
e
a
t
u
r
e
s
[
3
0
]
D
e
e
p
f
e
a
t
u
r
e
s
+
H
O
G
[
3
2
]
I
n
c
e
p
t
i
o
n
[
3
5
]
D
y
n
a
mi
c
c
a
sc
a
d
e
c
l
a
ss
i
f
i
e
r
[
3
6
]
A
t
t
e
n
t
i
o
n
a
l
c
o
n
v
o
l
u
t
i
o
n
a
l
[
3
7
]
Pr
o
p
o
sed
M
o
d
e
l
9
3
.
2
%
9
4
.
1
7
%
9
3
.
2
%
9
7
.
8
%
9
8
.
0
%
9
8
.
3
%
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.
25
,
No
.
3
,
Ma
r
ch
20
22
:
1
4
8
9
-
1
5
0
0
1498
4
.
2
.
Q
ua
lit
a
t
iv
e
a
na
ly
s
es o
f
t
he
pro
po
s
ed
m
o
del
W
eb
ca
m
f
ac
ial
im
ag
e
s
am
p
les
ar
e
o
b
tain
ed
f
o
r
test
in
g
.
T
h
e
r
o
b
u
s
tn
ess
o
f
o
u
r
p
r
o
ce
d
u
r
es
is
m
ea
s
u
r
ed
.
T
h
e
f
ac
ial
ar
ea
o
b
s
er
v
ed
in
ea
c
h
s
u
b
-
f
ig
u
r
e
is
a
g
r
ee
n
r
ec
tan
g
u
lar
.
Fig
u
r
e
s
5
(
a
)
an
d
(
b
)
.
Dem
o
n
s
tr
ates
s
o
m
e
g
o
o
d
i
d
e
n
tific
atio
n
ca
s
es:
E
x
tr
em
ely
p
r
ec
is
e
f
ac
ial
g
estu
r
es.
T
h
e
a
cc
u
r
ac
y
o
f
m
u
ltip
le
f
ac
ial
ex
p
r
ess
io
n
s
is
o
v
er
0
.
9
5
lik
e
an
g
er
an
d
h
ap
p
y
,
i
n
clu
d
in
g
w
h
en
a
n
o
teb
o
o
k
o
cc
lu
d
es
th
e
to
p
ics
p
ar
tly
Fig
u
r
e
5
(
c)
.
I
llu
s
tr
ates so
m
e
in
s
tan
ce
s
o
f
f
ailed
f
ac
ial
r
ec
o
g
n
itio
n
,
as "U
n
k
n
o
wn
"
o
r
m
is
lab
el.
(
a)
(
b
)
(
c)
Fig
u
r
e
5
.
Usi
n
g
f
ac
ial
web
ca
m
o
f
s
u
cc
ess
f
u
l r
ec
o
g
n
itio
n
o
f
f
ac
ial
ex
p
r
ess
io
n
s
with
b
etter
ac
cu
r
ac
y
f
o
r
th
e:
(
a)
h
ap
p
y
e
x
p
r
ess
io
n
in
a
n
g
r
y
p
ar
tial o
cc
lu
s
io
n
s
,
(
b
)
h
ea
d
d
e
f
lectio
n
an
d
a
n
o
teb
o
o
k
o
cc
lu
d
es p
ar
tially
,
an
d
(
c)
f
ailed
to
r
ec
o
g
n
itio
n
f
ac
ial
ex
p
r
ess
io
n
5.
CO
NCLU
SI
O
N
In
th
is
s
tu
d
y
,
an
im
p
r
o
v
ed
FER tec
h
n
iq
u
e
ca
n
p
r
o
ce
s
s
b
o
th
g
r
ay
s
ca
le
an
d
L
B
P p
ictu
r
es o
f
th
e
f
ac
e
at
th
e
s
am
e
tim
e.
W
e
co
n
ten
d
t
h
at
th
e
two
p
ict
u
r
e
ch
a
n
n
els
n
o
w
b
ein
g
u
tili
z
ed
co
m
p
lem
en
t
o
n
e
an
o
th
er
,
co
llect
b
o
th
l
o
ca
l
an
d
g
lo
b
al
in
f
o
r
m
atio
n
f
r
o
m
f
a
cial
im
ag
es,
an
d
en
h
an
ce
r
ec
o
g
n
itio
n
.
T
o
m
ak
e
f
u
ll
u
s
e
o
f
th
e
ch
ar
ac
ter
is
tics
th
at
h
av
e
b
ee
n
r
etr
iev
e
d
f
r
o
m
t
h
e
v
a
r
io
u
s
p
ictu
r
e
c
h
an
n
els,
a
weig
h
te
d
f
u
s
io
n
m
eth
o
d
is
p
r
esen
ted
.
T
o
a
u
to
m
atica
lly
ex
tr
ac
t
th
e
ch
ar
ac
ter
is
tics
o
f
f
ac
e
em
o
tio
n
s
f
r
o
m
f
ac
ial
g
r
ay
s
ca
le
p
ictu
r
es,
a
p
ar
tial
VGG1
9
n
etwo
r
k
is
b
u
ilt.
T
o
tr
ain
th
e
n
etwo
r
k
u
s
in
g
I
m
ag
eNe
t'
s
b
asic
p
ar
am
eter
s
,
f
in
e
-
tu
n
in
g
is
p
er
f
o
r
m
ed
.
C
NN
is
b
u
ilt
to
a
u
to
m
atica
lly
ex
tr
ac
t
f
ac
e
ex
p
r
ess
io
n
ch
ar
ac
ter
is
tics
f
r
o
m
L
B
P
p
ictu
r
es.
A
f
ter
th
at,
a
weig
h
ted
f
u
s
io
n
ap
p
r
o
a
ch
is
p
r
esen
te
d
to
co
m
b
in
e
th
e
two
ch
a
r
ac
ter
is
tics
to
m
ak
e
m
ax
im
u
m
u
s
e
o
f
th
e
co
m
p
lem
en
tar
y
f
ac
e
in
f
o
r
m
at
io
n
.
T
h
e
r
esu
lts
o
f
th
e
r
ec
o
g
n
itio
n
ar
e
b
ased
o
n
th
e
lin
ea
r
f
u
s
io
n
m
eth
o
d
to
a
n
o
v
el
m
o
d
el
o
f
C
NN
ar
ch
ite
ctu
r
e
u
s
in
g
co
r
r
esp
o
n
d
i
n
g
w
eig
h
ts
f
r
o
m
a
s
o
f
tm
a
x
cla
s
s
if
ier
.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
l
o
g
y
was
co
m
p
ar
ed
to
s
o
m
e
o
f
th
e
p
r
io
r
m
et
h
o
d
s
.
B
ased
o
n
th
e
co
m
p
ar
is
o
n
r
es
u
lts
,
it
ap
p
ea
r
s
th
at
th
e
p
r
o
p
o
s
ed
m
eth
o
d
o
l
o
g
y
o
u
tp
er
f
o
r
m
s
s
o
m
e
o
f
t
h
e
ex
is
tin
g
m
eth
o
d
s
.
B
ased
o
n
th
e
o
u
tco
m
es
o
f
th
e
ex
p
er
im
en
ts
,
we
m
ay
in
f
er
th
at
o
u
r
s
u
g
g
ested
ap
p
r
o
ac
h
ca
n
b
e
u
s
ed
f
o
r
f
ac
ial
em
o
tio
n
s
r
ec
o
g
n
itio
n
.
T
h
er
ef
o
r
e
a
m
o
d
er
n
m
o
d
el
m
u
s
t
b
e
d
e
v
elo
p
e
d
th
at
ca
n
lo
wer
tr
ain
i
n
g
tim
e
an
d
d
e
liv
er
y
in
r
ea
l
-
tim
e
ap
p
licatio
n
s
.
Ou
r
wo
r
k
f
o
r
th
e
f
u
tu
r
e
will
b
e
t
o
s
im
p
lify
th
e
n
etwo
r
k
f
u
r
th
er
a
n
d
ac
ce
ler
ate
th
e
alg
o
r
ith
m
.
W
e
p
lan
to
co
n
ce
n
tr
ate
o
n
o
th
er
f
a
cial
im
ag
es c
h
an
n
els to
ex
p
an
d
th
e
f
u
s
io
n
n
etwo
r
k
f
u
r
th
er
.
Evaluation Warning : The document was created with Spire.PDF for Python.