I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
15
,
No
.
5
,
Octo
b
er
2
0
2
6
,
p
p
.
4
7
0
0
~
4
7
1
0
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
5
.
p
p
4
7
0
0
-
4
7
1
0
4700
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Facia
l emo
tion re
co
g
nition throug
h
artificial
i
ntellige
nce
Ana
s
Q
ute
is
ha
t
1,
2
,
M
o
ha
m
ma
d H
a
s
s
a
n
1
,
Ash
ra
f
Al
Sh
a
ra
h
2
,
Ahm
e
d Q
uta
is
ha
t
3
1
D
e
p
a
r
t
me
n
t
o
f
C
o
mm
u
n
i
c
a
t
i
o
n
a
n
d
C
o
m
p
u
t
e
r
E
n
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
,
Al
-
A
h
l
i
y
y
a
A
mm
a
n
U
n
i
v
e
r
si
t
y
,
A
mm
a
n
,
J
o
r
d
a
n
2
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
i
c
a
l
En
g
i
n
e
e
r
i
n
g
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
Te
c
h
n
o
l
o
g
y
,
A
l
-
B
a
l
q
a
A
p
p
l
i
e
d
U
n
i
v
e
r
si
t
y
,
A
l
S
a
l
t
,
Jo
r
d
a
n
3
G
e
n
e
r
a
l
F
o
u
n
d
a
t
i
o
n
P
r
o
g
r
a
m,
S
o
h
a
r
U
n
i
v
e
r
si
t
y
,
S
o
h
a
r
,
O
m
a
n
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
2
9
,
2
0
2
5
R
ev
is
ed
J
u
l 2
9
,
2
0
2
6
Acc
ep
ted
Au
g
1
2
,
2
0
2
6
Th
is
p
a
p
e
r
p
re
se
n
ts
a
fu
z
z
y
m
in
-
m
a
x
d
e
e
p
n
e
u
ra
l
n
e
tw
o
rk
(F
M
M
DN
N)
fo
r
fa
c
ial
e
m
o
ti
o
n
re
c
o
g
n
it
io
n
(F
E
R).
Th
e
fu
z
z
y
m
in
-
m
a
x
n
e
u
ra
l
n
e
two
rk
(F
M
M
NN
)
will
re
p
lac
e
th
e
c
las
s
ifi
c
a
ti
o
n
lay
e
r
i
n
th
e
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
(CNN
).
Th
e
fu
z
z
y
m
in
-
m
a
x
(F
M
M
)
e
x
h
i
b
it
s
a
p
p
e
a
li
n
g
fe
a
tu
re
s
su
c
h
a
s o
n
li
n
e
a
d
a
p
tati
o
n
,
fa
st t
ra
in
in
g
,
i
n
terp
re
tab
il
it
y
,
a
n
d
n
o
ise
r
o
b
u
stn
e
ss
.
Th
e
fu
ll
y
c
o
n
n
e
c
ted
s
o
ftma
x
lay
e
r
in
t
h
e
CNN
is
re
p
lac
e
d
b
y
th
e
F
M
M
NN
,
wh
e
re
h
y
p
e
rb
o
x
e
s
will
b
e
c
re
a
te
d
a
n
d
u
se
d
f
o
r
c
las
sifica
ti
o
n
p
u
r
p
o
se
s
.
Th
e
F
ER2
0
1
3
d
a
tas
e
t
is
u
se
d
to
tes
t
th
e
p
e
rf
o
rm
a
n
c
e
o
f
t
h
e
p
r
o
p
o
se
d
m
e
th
o
d
.
Th
e
e
m
p
iri
c
a
l
re
s
u
lt
s
s
h
o
w
th
a
t
a
d
d
in
g
an
F
M
M
NN
lay
e
r
t
o
th
e
CNN
n
e
two
rk
imp
ro
v
e
s th
e
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
Dee
p
lear
n
in
g
Facial
em
o
tio
n
al
r
ec
o
g
n
itio
n
Fu
zz
y
m
in
-
m
a
x
n
eu
r
al
n
etwo
r
k
Haa
r
ca
s
ca
d
e
alg
o
r
ith
m
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
:
An
as Qu
teis
h
at
Dep
ar
tm
en
t o
f
C
o
m
m
u
n
icatio
n
an
d
C
o
m
p
u
ter
E
n
g
in
ee
r
i
n
g
,
Facu
lty
o
f
E
n
g
in
ee
r
in
g
Al
-
Ah
liy
y
a
Am
m
an
Un
i
v
er
s
ity
Am
m
an
,
J
o
r
d
a
n
E
m
ail:
a.
q
u
teis
h
at@
am
m
an
u
.
e
d
u
.
jo
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
p
u
r
s
u
it
o
f
lin
k
in
g
h
u
m
a
n
co
m
m
u
n
icatio
n
with
m
ac
h
i
n
e
s
y
s
tem
s
h
as
d
r
iv
en
ar
tific
ial
in
tellig
en
ce
d
ev
elo
p
m
e
n
t
an
d
im
p
r
o
v
em
e
n
ts
in
h
u
m
an
-
co
m
p
u
ter
in
ter
ac
tio
n
with
in
an
elec
tr
o
n
ic
wo
r
ld
[
1
]
–
[
5
]
.
T
h
e
ad
v
an
ce
m
e
n
t
o
f
f
ac
ial
em
o
tio
n
r
ec
o
g
n
itio
n
(
FER)
en
ab
les
t
h
is
r
esear
ch
to
i
n
ter
p
r
et
h
u
m
a
n
em
o
tio
n
al
s
ig
n
als
wh
ich
s
tan
d
as
a
co
r
e
c
o
m
p
o
n
en
t
o
f
th
is
wo
r
k
.
R
esear
ch
er
s
,
to
g
eth
er
with
tech
n
o
lo
g
is
ts
u
s
e
f
ac
ial
ex
p
r
ess
io
n
d
ata
to
ac
h
iev
e
m
ac
h
in
e
-
b
as
ed
h
u
m
a
n
em
o
tio
n
p
er
ce
p
tio
n
an
d
u
n
d
er
s
tan
d
i
n
g
s
o
m
ac
h
in
es
ca
n
r
esp
o
n
d
em
p
ath
etica
lly
in
h
u
m
an
-
m
ac
h
in
e
in
ter
ac
tio
n
s
[
6
]
.
M
u
ltip
le
p
r
o
f
ess
io
n
al
s
ec
to
r
s
alo
n
g
wi
th
o
p
e
r
atio
n
al
a
r
ea
s
r
ec
eiv
e
im
p
licatio
n
s
f
r
o
m
th
e
r
esear
ch
f
in
d
in
g
s
.
Ap
p
ly
in
g
m
o
d
er
n
tech
n
o
lo
g
y
to
h
u
m
a
n
-
c
o
m
p
u
ter
in
ter
ac
tio
n
wo
u
ld
r
en
d
er
s
y
s
tem
s
ca
p
a
b
le
o
f
in
ter
p
r
etin
g
h
u
m
a
n
e
m
o
tio
n
s
an
d
c
r
ea
tin
g
r
esp
o
n
s
es
with
in
cr
ea
s
ed
n
atu
r
al
u
n
d
er
s
tan
d
in
g
.
FER
f
u
n
ctio
n
s
as
a
v
ital
f
ield
f
o
r
h
u
m
an
-
c
o
m
p
u
te
r
in
ter
ac
ti
o
n
an
d
a
f
f
ec
tiv
e
co
m
p
u
tin
g
s
in
ce
it
h
elp
s
n
u
m
er
o
u
s
in
d
u
s
tr
ies
,
in
clu
d
i
n
g
h
ea
lth
ca
r
e,
ed
u
ca
tio
n
,
an
d
en
ter
tain
m
en
t
co
m
p
an
ies
,
r
u
n
th
eir
o
p
e
r
atio
n
s
.
T
ax
o
n
o
m
ic
r
esear
ch
i
n
FER
g
en
er
ates
au
to
m
atic
m
eth
o
d
s
th
at
an
aly
ze
f
ac
ial
e
x
p
r
e
s
s
io
n
s
to
d
is
co
v
er
em
o
tio
n
al
s
tates
in
h
u
m
an
s
e
f
f
ec
tiv
ely
[
7
]
.
FER
m
eth
o
d
s
d
u
r
in
g
th
eir
in
itial
s
tag
e
ex
tr
ac
ted
s
p
ec
if
ic
f
ac
ial
f
ea
tu
r
es
b
ased
o
n
m
an
u
al
f
a
c
e
lan
d
m
ar
k
s
s
u
ch
as
ey
e
-
to
-
e
y
e
an
d
e
y
e
-
to
-
n
o
s
e
-
to
-
m
o
u
th
d
is
tan
ce
s
.
T
h
e
f
ac
ial
f
ea
tu
r
e
e
x
tr
ac
tio
n
p
r
o
ce
s
s
d
ep
en
d
ed
o
n
p
r
i
n
cip
al
c
o
m
p
o
n
en
t
an
aly
s
is
(
PC
A
)
an
d
lo
ca
l
b
in
ar
y
p
atter
n
s
(
L
B
P)
,
am
o
n
g
o
t
h
er
tech
n
iq
u
es,
to
ex
tr
ac
t
d
ata
,
wh
ich
th
en
r
eq
u
ir
ed
class
if
icatio
n
[
8
]
.
T
h
ese
tech
n
iq
u
es
d
em
o
n
s
tr
ated
ex
citin
g
p
r
o
g
r
es
s
u
n
til
th
ey
f
ailed
to
d
eliv
er
e
x
p
ec
ted
r
esu
lts
wh
en
f
ac
e
d
wi
th
s
h
if
tin
g
li
g
h
tin
g
co
n
d
itio
n
s
,
d
i
f
f
er
en
t
h
ea
d
p
o
s
itio
n
in
g
,
a
n
d
p
a
r
tial o
b
s
cu
r
atio
n
o
f
s
p
ec
if
ic
f
ac
e
ar
ea
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
F
a
cia
l e
mo
tio
n
r
ec
o
g
n
itio
n
th
r
o
u
g
h
a
r
tifi
cia
l
in
tellig
en
ce
(
A
n
a
s
Qu
teis
h
a
t
)
4701
W
ith
th
e
in
tr
o
d
u
ctio
n
o
f
d
ee
p
lear
n
in
g
,
it
h
as
b
ec
o
m
e
th
e
b
ac
k
b
o
n
e
o
f
ar
tific
ial
in
tellig
en
ce
as
we
k
n
o
w
it
to
d
ay
.
I
t
h
as
f
o
u
n
d
its
way
to
FE
R
an
d
g
av
e
s
ig
n
if
ican
t
r
esu
lts
b
ec
au
s
e
th
e
f
ea
tu
r
es
ar
e
ex
tr
ac
ted
as
p
ar
t
o
f
t
h
e
d
ee
p
lear
n
in
g
s
tr
u
ctu
r
e
,
an
d
th
is
m
ad
e
th
e
n
etw
o
r
k
im
m
u
n
e
to
th
e
n
o
is
e
av
ailab
le
in
th
e
im
a
g
es.
On
e
o
f
th
e
m
o
s
t
co
m
m
o
n
ly
u
s
ed
d
ee
p
lear
n
in
g
p
ar
ad
i
g
m
s
is
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
[
9
]
–
[
1
2
]
.
Ho
wev
er
,
th
e
s
tan
d
ar
d
ar
c
h
itectu
r
e
o
f
a
C
NN
f
o
r
class
if
icatio
n
ty
p
ically
en
d
s
with
o
n
e
o
r
m
o
r
e
f
u
ll
y
co
n
n
ec
ted
lay
er
s
,
f
o
llo
wed
b
y
a
s
o
f
tm
ax
ac
tiv
atio
n
f
u
n
ct
io
n
.
W
h
ile
p
o
wer
f
u
l,
th
is
a
p
p
r
o
ac
h
h
as
k
n
o
wn
lim
itatio
n
s
s
u
ch
as o
v
er
co
n
f
id
en
ce
in
p
r
e
d
ictio
n
s
.
T
h
e
s
o
f
tm
ax
p
r
o
d
u
ce
s
a
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
th
at
alwa
y
s
s
u
m
s
to
1
,
o
f
ten
y
ield
in
g
h
i
g
h
co
n
f
id
en
ce
ev
en
f
o
r
am
b
ig
u
o
u
s
o
r
o
u
t
-
of
-
d
is
tr
ib
u
tio
n
s
am
p
les
[
1
3
]
.
T
h
e
s
ec
o
n
d
lim
itatio
n
is
th
e
f
ac
t th
at
th
e
f
u
lly
co
n
n
ec
ted
lay
e
r
s
ar
e
co
n
s
id
er
ed
as b
lack
b
o
x
es,
m
ak
in
g
an
y
attem
p
t
to
ex
p
lain
th
e
s
o
lu
tio
n
im
p
o
s
s
ib
le.
Fin
ally
,
C
NNs
ar
e
u
n
ab
le
to
h
an
d
le
am
b
ig
u
ity
an
d
u
n
ce
r
tain
ty
,
wh
ich
is
esp
ec
ially
co
n
ce
r
n
i
n
g
in
th
e
F
E
R
d
ataset
,
wh
er
e
an
e
x
p
r
ess
io
n
ca
n
lie
b
etwe
en
"
h
ap
p
y
"
a
n
d
"su
r
p
r
is
e,
"
o
r
b
e
a
b
len
d
o
f
m
u
ltip
le
em
o
tio
n
s
.
T
h
e
cr
is
p
,
d
eter
m
in
ed
n
atu
r
e
o
f
s
o
f
tm
a
x
d
o
es
n
o
t
m
o
d
el
th
i
s
u
n
ce
r
tain
ty
well
,
r
esu
ltin
g
in
lo
wer
-
t
h
an
-
ex
p
ec
t
ed
test
in
g
ac
cu
r
ac
y
.
I
n
th
is
p
ap
e
r
,
th
e
f
u
zz
y
m
in
-
m
ax
n
eu
r
al
n
etwo
r
k
(
FMM
N
N)
[
1
4
]
–
[
1
7
]
is
in
tr
o
d
u
ce
d
to
r
ep
lace
th
e
s
o
f
tm
ax
lay
er
in
th
e
C
NN
to
s
o
lv
e
th
e
s
h
o
r
tc
o
m
in
g
s
o
f
th
e
s
o
f
tm
ax
f
u
lly
co
n
n
ec
ted
lay
e
r
.
T
h
e
f
u
zz
y
m
in
-
m
a
x
d
ee
p
n
eu
r
al
n
etwo
r
k
(
FMM
DNN)
will
u
s
e
th
e
C
NN
ab
ilit
y
to
ex
tr
ac
t
f
ea
t
u
r
es
an
d
will
b
e
tr
ain
ed
b
y
th
ese
f
ea
tu
r
es.
I
n
th
e
liter
atu
r
e
,
d
if
f
er
en
t
s
tu
d
ies
in
t
r
o
d
u
ce
d
f
u
zz
y
m
in
-
m
a
x
(
FMM
)
in
d
ee
p
lea
r
n
in
g
[
1
6
]
–
[
1
8
]
.
I
n
th
is
p
ap
er
,
th
e
lear
n
in
g
ab
ilit
y
of
FMM
NN
is
u
s
ed
as
th
e
f
u
lly
co
n
n
ec
ted
n
etwo
r
k
in
th
e
C
NN.
T
h
e
d
etails
o
f
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
ar
e
p
r
e
s
en
ted
in
th
e
n
ex
t
s
ec
tio
n
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
led
t
o
a
m
o
d
el
th
at
is
m
o
r
e
in
ter
p
r
etab
le,
b
etter
at
q
u
a
n
tif
y
in
g
u
n
ce
r
tain
ty
,
a
n
d
p
o
ten
tial
ly
m
o
r
e
r
o
b
u
s
t
to
th
e
n
o
is
y
an
d
am
b
ig
u
o
u
s
lab
els
f
o
u
n
d
in
d
atasets
lik
e
FER2
0
1
3
.
T
h
e
r
em
ain
d
er
o
f
t
h
e
p
a
p
er
is
o
r
g
an
ize
d
as
f
o
llo
ws
.
S
ec
tio
n
2
p
r
esen
ts
th
e
m
eth
o
d
o
lo
g
y
th
at
co
n
tain
s
th
e
C
NN
n
etwo
r
k
alo
n
g
with
th
e
FMM
n
etwo
r
k
i
n
o
r
d
er
to
c
r
ea
te
th
e
FMM
DNN.
T
h
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
ar
e
p
r
esen
ted
in
s
ec
tio
n
3
.
L
astl
y
,
s
ec
tio
n
4
p
r
esen
ts
th
e
co
n
clu
s
io
n
o
f
th
is
wo
r
k
.
2.
M
E
T
H
O
D
On
e
o
f
th
e
m
o
s
t
u
s
ed
d
ee
p
lea
r
n
in
g
n
e
u
r
al
n
etwo
r
k
s
is
th
e
C
NN,
th
is
n
etwo
r
k
au
to
m
atica
lly
ex
tr
ac
ts
th
e
f
ea
tu
r
es
with
o
u
t
an
y
h
u
m
a
n
s
u
p
er
v
is
io
n
[
1
9
]
.
I
n
C
NNs,
f
ea
tu
r
es
ar
e
lear
n
ed
d
ir
ec
tly
f
r
o
m
th
e
d
ata
d
u
r
in
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
,
r
ed
u
cin
g
th
e
n
ee
d
f
o
r
m
a
n
u
al
p
r
e
p
r
o
ce
s
s
in
g
o
f
th
e
in
p
u
t
d
ata.
I
n
f
ac
t,
th
is
f
ea
tu
r
e
is
f
o
r
d
ee
p
lear
n
in
g
tech
n
i
q
u
es
in
g
en
er
al.
C
NNs
lear
n
h
ier
ar
ch
i
ca
l
r
ep
r
esen
tatio
n
s
o
f
d
ata,
s
tar
tin
g
f
r
o
m
s
im
p
le
f
ea
tu
r
es
(
lik
e
d
o
ts
an
d
lin
es)
i
n
th
e
in
itial
lay
er
s
t
o
m
o
r
e
co
m
p
lex
f
ea
tu
r
es
(
lik
e
s
h
ap
es
a
n
d
o
b
jects)
in
d
ee
p
e
r
lay
er
s
.
T
h
is
allo
ws
C
NN
s
to
ef
f
ec
tiv
ely
ca
p
tu
r
e
co
m
p
lex
p
atter
n
s
in
d
ata
lik
e
im
ag
es
an
d
v
id
eo
s
.
Fig
u
r
e
1
p
r
esen
ts
th
e
d
if
f
er
e
n
t la
y
er
s
in
v
o
lv
ed
i
n
b
u
ild
i
n
g
a
C
NN
[
1
9
]
.
Fig
u
r
e
1
.
Hier
ar
c
h
ical
s
tr
u
ctu
r
e
o
f
C
NNs
I
n
ad
d
itio
n
,
C
NNs
r
eu
s
e
th
e
s
am
e
f
ilter
s
ac
r
o
s
s
th
e
e
n
tire
i
m
ag
e,
s
ig
n
if
ican
tly
r
e
d
u
cin
g
t
h
e
n
u
m
b
er
o
f
f
ea
tu
r
es
to
b
e
lear
n
ed
c
o
m
p
ar
ed
to
tr
ad
itio
n
al
n
eu
r
al
n
et
wo
r
k
s
.
T
h
r
o
u
g
h
th
eir
p
o
o
lin
g
lay
er
s
,
C
NNs
ca
n
r
ec
o
g
n
ize
p
atter
n
s
r
eg
ar
d
less
o
f
th
eir
p
o
s
itio
n
in
th
e
im
ag
e
.
T
h
is
is
p
ar
ticu
lar
ly
u
s
ef
u
l
f
o
r
task
s
lik
e
o
b
ject
d
etec
tio
n
an
d
im
ag
e
class
if
icatio
n
.
T
h
ey
also
r
ed
u
ce
th
e
s
ize
o
f
d
ata
r
ep
r
esen
tatio
n
s
,
d
ec
r
ea
s
in
g
co
m
p
u
tatio
n
al
co
m
p
lex
ity
an
d
in
cr
ea
s
in
g
ef
f
icien
cy
.
An
o
t
h
er
im
p
o
r
tan
t
f
ea
tu
r
e
is
p
a
r
al
lelizatio
n
,
as
C
NNs
ca
n
p
r
o
ce
s
s
m
u
ltip
le
p
ix
els
s
im
u
ltan
eo
u
s
ly
d
u
e
to
th
eir
ar
ch
itectu
r
e,
lead
in
g
to
f
a
s
ter
tr
ain
in
g
tim
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
2
0
2
6
:
4
7
0
0
-
4
7
1
0
4702
co
m
p
ar
ed
to
s
eq
u
e
n
tial
d
ata
p
r
o
ce
s
s
in
g
m
eth
o
d
s
.
T
h
is
f
ea
tu
r
e
is
a
s
ig
n
if
ican
t
ad
v
an
tag
e
wh
en
d
ea
lin
g
with
lar
g
e
d
atasets
[
1
9
]
.
I
n
a
C
N
N
,
t
h
e
i
n
p
u
t
l
a
y
e
r
f
i
r
s
t
r
e
c
e
i
v
es
t
h
e
r
a
w
d
a
t
a
(
t
h
e
m
a
i
n
t
y
p
e
o
f
i
n
p
u
t
d
a
t
a
)
p
r
o
v
i
d
e
d
t
o
t
h
e
m
o
d
e
l
.
T
h
i
s
l
a
y
e
r
s
i
m
p
l
y
p
as
s
es
t
h
e
i
m
a
g
e
t
o
t
h
e
n
e
x
t
l
a
y
e
r
w
it
h
o
u
t
a
n
y
c
o
m
p
u
t
a
t
i
o
n
s
,
b
u
t
i
t
m
u
s
t
b
e
a
d
j
u
s
t
e
d
t
o
a
c
c
o
m
m
o
d
a
t
e
t
h
e
i
m
a
g
e
p
r
o
p
e
r
t
i
e
s
(
s
u
c
h
as
h
e
i
g
h
t
,
wi
d
t
h
,
a
n
d
n
u
m
b
e
r
o
f
c
h
a
n
n
e
l
s
)
a
s
t
h
es
e
p
r
o
p
e
r
t
i
e
s
m
u
s
t
b
e
u
n
i
f
o
r
m
.
T
h
e
n
,
t
h
e
i
m
a
g
e
i
s
f
ed
i
n
t
o
t
h
e
c
o
n
v
o
l
u
ti
o
n
l
a
y
e
r
,
wh
i
c
h
p
e
r
f
o
r
m
s
a
c
o
n
v
o
l
u
t
i
o
n
o
p
e
r
a
t
i
o
n
b
e
t
w
ee
n
t
h
e
i
m
a
g
e
a
n
d
o
n
e
o
r
m
o
r
e
f
i
l
t
e
r
s
(
a
ls
o
c
a
l
l
e
d
k
e
r
n
el
s
)
t
o
r
es
u
l
t
i
n
a
f
e
a
t
u
r
e
m
a
p
c
o
n
ta
i
n
i
n
g
t
h
e
a
u
t
o
m
a
t
i
c
al
l
y
e
x
t
r
a
c
t
e
d
f
e
at
u
r
e
s
.
T
h
e
i
n
p
u
t
im
a
g
e
s
c
a
n
h
a
v
e
o
n
e
o
r
m
o
r
e
c
h
a
n
n
e
l
s
,
a
n
d
t
h
e
f
il
t
e
r
s
h
a
v
e
t
h
e
s
a
m
e
n
u
m
b
e
r
o
f
c
h
a
n
n
e
l
s
(
e
a
c
h
f
i
l
te
r
h
a
s
a
c
h
an
n
e
l
f
o
r
e
a
c
h
c
h
a
n
n
e
l
i
n
t
h
e
i
n
p
u
t
i
m
a
g
e
)
.
T
h
e
i
m
a
g
e
a
n
d
t
h
e
f
i
l
t
e
r
,
m
ea
n
i
n
g
e
a
c
h
c
h
a
n
n
e
l
i
n
t
h
e
i
m
a
g
e
w
i
t
h
i
ts
c
o
r
r
e
s
p
o
n
d
i
n
g
c
h
a
n
n
e
l
i
n
t
h
e
f
i
l
t
e
r
,
a
r
e
c
o
n
s
i
d
e
r
e
d
.
B
o
t
h
t
h
e
i
m
a
g
e
a
n
d
t
h
e
f
i
l
t
e
r
a
r
e
t
r
e
a
t
e
d
a
s
t
w
o
-
d
i
m
e
n
s
i
o
n
a
l
m
a
t
r
i
c
e
s
,
w
h
e
r
e
t
h
e
i
m
a
g
e
m
a
t
r
i
x
r
e
p
r
e
s
e
n
ts
p
i
x
e
l
v
a
l
u
e
s
,
an
d
t
h
e
f
i
l
t
e
r
m
a
t
r
i
x
c
o
n
t
a
i
n
s
w
ei
g
h
t
v
a
l
u
e
s
.
T
h
e
f
i
l
t
e
r
m
a
t
r
i
x
is
t
y
p
i
c
al
l
y
m
u
c
h
s
m
a
l
l
e
r
i
n
h
e
i
g
h
t
a
n
d
w
i
d
t
h
.
I
n
t
h
e
c
o
n
v
o
l
u
t
i
o
n
o
p
e
r
a
t
i
o
n
,
t
h
e
f
i
l
t
e
r
s
l
i
d
es
o
v
er
t
h
e
i
m
a
g
e
o
n
e
s
t
r
i
d
e
a
t
a
t
im
e
f
r
o
m
t
h
e
b
e
g
i
n
n
i
n
g
o
f
t
h
e
i
m
a
g
e
(
t
h
e
t
o
p
-
l
e
f
t
c
o
r
n
e
r
)
t
o
t
h
e
e
n
d
o
f
t
h
e
r
o
w
,
t
h
e
n
s
t
a
r
t
s
a
g
a
i
n
f
r
o
m
t
h
e
b
e
g
i
n
n
i
n
g
o
f
t
h
e
r
o
w
o
n
e
s
t
r
i
d
e
b
e
l
o
w
t
h
e
p
r
e
v
i
o
u
s
p
o
s
i
t
i
o
n
,
a
n
d
c
o
n
t
i
n
u
es
u
n
t
i
l
i
t
h
a
s
g
o
n
e
t
h
r
o
u
g
h
t
h
e
e
n
t
i
r
e
i
m
a
g
e
,
a
s
s
h
o
w
n
i
n
F
i
g
u
r
e
2
.
Fig
u
r
e
2
.
T
h
e
p
o
s
itio
n
s
o
f
a
3
×
3
f
ilter
o
n
a
5
×
5
im
ag
e
with
a
2
×
2
s
tr
id
e
T
h
is
r
eq
u
ir
es
p
r
ec
is
io
n
i
n
ca
lc
u
latin
g
th
e
d
im
e
n
s
io
n
s
o
f
b
o
t
h
th
e
f
ilter
an
d
th
e
s
tr
id
e
s
o
t
h
at
th
ey
f
it
th
e
in
p
u
t
im
ag
e
d
im
en
s
io
n
s
.
Ass
u
m
in
g
th
at
th
e
d
im
e
n
s
io
n
s
o
f
th
e
in
p
u
t
im
ag
e
a
r
e
ℎ
×
,
th
e
d
im
en
s
io
n
s
o
f
th
e
f
ilter
ar
e
ℎ
×
,
an
d
th
e
d
im
e
n
s
io
n
s
o
f
th
e
s
tr
id
e
ar
e
ℎ
×
,
th
e
f
ea
tu
r
e
m
a
p
will
b
e
a
ℎ
×
m
atr
ix
as sh
o
wn
in
(
1
)
an
d
(
2
).
ℎ
=
ℎ
−
ℎ
ℎ
+
1
(
1
)
=
−
+
1
(
2
)
T
h
ese
two
am
o
u
n
ts
m
u
s
t
b
e
n
atu
r
al
n
u
m
b
er
s
.
Oth
er
wis
e,
p
a
d
d
in
g
th
e
in
p
u
t
im
a
g
e
m
ay
b
e
n
ee
d
ed
(
c
h
an
g
e
its
d
im
en
s
io
n
s
b
y
ad
d
i
n
g
ze
r
o
s
ar
o
u
n
d
it).
Ad
d
itio
n
ally
,
ea
c
h
s
tr
id
e
d
im
en
s
io
n
m
u
s
t
n
o
t
b
e
g
r
ea
ter
th
an
its
co
r
r
esp
o
n
d
in
g
f
ilter
d
im
e
n
s
io
n
to
en
s
u
r
e
th
at
all
p
ar
ts
o
f
th
e
in
p
u
t
im
a
g
e
ar
e
p
ass
ed
th
r
o
u
g
h
.
E
ac
h
e
n
tr
y
in
th
e
m
en
tio
n
e
d
m
atr
ix
co
r
r
esp
o
n
d
s
(
in
o
r
d
e
r
)
to
o
n
e
o
f
th
e
p
o
s
itio
n
s
th
r
o
u
g
h
wh
ic
h
th
e
f
i
lter
p
ass
es.
At
ea
ch
p
o
s
itio
n
th
e
f
ilter
p
ass
es th
r
o
u
g
h
d
u
r
in
g
its
s
tr
id
in
g
,
an
en
t
r
y
-
wis
e
p
r
o
d
u
ct
is
p
er
f
o
r
m
ed
b
e
twee
n
th
e
f
ilter
an
d
th
e
co
r
r
esp
o
n
d
in
g
p
atc
h
o
f
th
e
im
ag
e.
T
h
e
en
tr
y
-
wis
e
p
r
o
d
u
ct
o
p
er
atio
n
b
etwe
en
two
m
atr
ices
o
f
id
en
tical
d
im
en
s
io
n
s
is
p
er
f
o
r
m
ed
as
(
3
)
.
[
11
12
13
21
22
23
31
32
33
]
×
[
11
12
13
21
22
23
31
32
33
]
=
[
11
11
12
12
13
13
21
21
22
22
23
23
31
31
32
32
33
33
]
(
3
)
T
h
en
,
all
th
e
v
alu
es
r
esu
ltin
g
f
r
o
m
th
e
en
tr
y
-
wis
e
p
r
o
d
u
ct
o
p
er
atio
n
s
ar
e
s
u
m
m
e
d
with
a
d
if
f
er
en
t
v
alu
e
f
o
r
ea
ch
f
ilter
,
ca
lled
th
e
b
ias,
to
p
r
o
d
u
ce
th
e
co
r
r
esp
o
n
d
in
g
e
n
tr
y
o
f
th
e
f
ea
tu
r
e
m
ap
.
So
,
if
is
th
e
in
p
u
t
im
ag
e
,
W
is
th
e
f
ilter
,
an
d
is
th
e
b
ias,
th
en
ev
er
y
en
tr
y
o
f
th
e
f
ea
tu
r
e
m
ap
is
co
m
p
u
ted
as in
(
4
).
,
=
∑
(
(
,
,
)
(
+
,
+
,
)
)
,
,
+
(
4
)
T
h
e
o
u
tp
u
t
o
f
t
h
e
co
n
v
o
l
u
tio
n
lay
er
is
a
s
et
o
f
f
ea
tu
r
e
m
ap
s
,
o
n
e
f
o
r
ea
ch
f
ilter
.
T
h
ese
f
ea
tu
r
e
m
ap
s
h
ig
h
lig
h
t
th
e
d
if
f
er
e
n
t
f
ea
tu
r
e
s
lear
n
ed
f
r
o
m
th
e
in
p
u
t
im
a
g
e.
On
th
e
o
t
h
er
h
a
n
d
,
th
e
f
i
lter
s
ar
e
lear
n
ab
le
p
ar
am
eter
s
(
lik
e
weig
h
ts
in
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
)
.
T
h
e
m
u
ltip
le
f
ilter
s
lear
n
d
if
f
er
e
n
t
p
atter
n
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
F
a
cia
l e
mo
tio
n
r
ec
o
g
n
itio
n
th
r
o
u
g
h
a
r
tifi
cia
l
in
tellig
en
ce
(
A
n
a
s
Qu
teis
h
a
t
)
4703
(
d
etec
t
d
if
f
er
en
t
f
ea
tu
r
es).
An
ac
tiv
atio
n
f
u
n
ctio
n
is
ap
p
lied
af
ter
th
e
c
o
n
v
o
lu
tio
n
lay
er
,
a
d
d
in
g
n
o
n
-
lin
ea
r
ity
to
th
e
m
o
d
el.
On
e
c
o
m
m
o
n
ac
tiv
atio
n
f
u
n
ctio
n
u
s
ed
in
C
NNs
is
th
e
s
ig
m
o
id
(
)
,
wh
ich
is
u
s
ed
in
t
h
e
co
n
tex
t
o
f
b
in
ar
y
class
if
icatio
n
an
d
p
r
o
b
ab
ilis
tic
in
ter
p
r
etatio
n
task
s
,
wh
er
e
it
tr
an
s
f
o
r
m
s
its
in
p
u
t
in
to
an
o
u
tp
u
t
in
th
e
r
an
g
e
o
f
0
to
1
,
b
ased
o
n
(
5
).
(
)
=
1
1
+
−
(
5
)
Of
co
u
r
s
e,
th
e
in
p
u
ts
to
th
ese
f
u
n
ctio
n
s
ar
e
th
e
f
ea
tu
r
e
m
a
p
s
f
r
o
m
th
e
p
r
e
v
io
u
s
co
n
v
o
lu
ti
o
n
lay
er
s
,
wh
ich
ar
e
m
atr
ices.
So
,
th
ese
f
u
n
ctio
n
s
ar
e
a
p
p
lied
t
o
ea
c
h
en
tr
y
o
f
th
e
i
n
p
u
t
m
atr
ix
.
T
h
e
o
u
t
p
u
t
is
th
e
n
p
ass
ed
to
th
e
p
o
o
lin
g
lay
er
,
w
h
ich
r
e
d
u
ce
s
th
e
s
ize
o
f
th
e
d
ata
to
d
ec
r
ea
s
e
th
e
c
o
m
p
le
x
ity
o
f
th
e
m
o
d
el
an
d
in
cr
ea
s
e
its
ef
f
icien
cy
an
d
r
esis
tan
ce
to
s
m
all
v
ar
iatio
n
s
in
t
h
e
in
p
u
t
im
ag
es.
Ma
x
p
o
o
li
n
g
,
wh
ich
s
elec
ts
th
e
m
ax
im
u
m
v
alu
e
f
r
o
m
a
s
p
ec
i
f
ic
r
eg
io
n
(
ca
lled
a
win
d
o
w)
,
is
co
m
m
o
n
ly
u
s
ed
.
T
h
e
co
n
ce
p
t
o
f
s
tr
id
e
is
also
p
r
esen
t h
er
e
,
with
t
h
e
o
n
l
y
d
if
f
er
en
ce
b
ei
n
g
th
at
th
e
win
d
o
w
r
ep
lace
s
th
e
f
ilter
,
as sh
o
wn
i
n
Fig
u
r
e
3
.
Fig
u
r
e
3
.
Ma
x
p
o
o
lin
g
with
a
2
×
2
win
d
o
w
an
d
a
2
×
2
s
tr
id
e
A
s
eq
u
en
ce
o
f
c
o
n
v
o
lu
tio
n
la
y
er
,
ac
tiv
atio
n
f
u
n
ctio
n
,
an
d
p
o
o
lin
g
lay
er
is
u
s
u
ally
r
ep
e
ated
m
o
r
e
th
an
o
n
ce
in
C
NNs,
im
m
ed
iately
af
ter
th
e
in
p
u
t
lay
er
.
B
u
t
in
s
tead
o
f
th
e
r
aw
im
ag
es,
th
e
in
p
u
t
o
f
ea
ch
s
u
b
s
eq
u
en
t
c
o
n
v
o
lu
tio
n
lay
er
is
th
e
o
u
tp
u
t
o
f
its
p
r
ev
io
u
s
lay
er
,
wh
ich
will
b
e
o
n
e
o
r
m
o
r
e
f
ea
tu
r
e
m
ap
s
.
Ad
d
itio
n
ally
,
m
u
ltip
le
f
ea
t
u
r
e
m
ap
s
ar
e
tr
ea
ted
as
ch
a
n
n
els
o
f
a
s
in
g
le
i
n
p
u
t
im
ag
e
wh
en
f
ed
in
to
a
co
n
v
o
l
u
tio
n
lay
er
.
Af
ter
th
em
co
m
es
th
e
r
o
le
o
f
th
e
f
latten
in
g
lay
er
to
c
o
n
v
e
r
t
th
e
m
u
lti
-
d
im
en
s
io
n
al
in
p
u
t
m
atr
ices in
to
v
ec
to
r
s
,
as sh
o
w
n
in
Fig
u
r
e
4
.
Fig
u
r
e
4
.
T
h
e
f
u
lly
co
n
n
ec
ted
lay
er
I
n
th
is
p
ap
er
,
th
e
f
u
lly
co
n
n
e
cted
n
etwo
r
k
is
a
FMM
NN,
r
ep
lacin
g
th
e
o
r
i
g
in
al
s
o
f
tm
ax
n
etwo
r
k
.
T
h
e
FMM
NN
r
eq
u
ir
es
a
f
ix
ed
-
s
ize
f
ea
tu
r
e
v
ec
to
r
in
o
r
d
e
r
to
en
s
u
r
e
co
n
s
is
ten
t
r
esu
lts
w
ith
in
th
e
d
ataset.
Fo
r
th
is
to
b
e
im
p
lem
en
ted
,
a
f
ea
tu
r
e
v
ec
to
r
o
f
2
5
6
is
u
s
ed
.
T
h
is
is
d
o
n
e
b
y
en
s
u
r
in
g
th
at
th
e
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
(
GAP)
lay
er
h
as a
s
iz
e
o
f
2
5
6
,
as sh
o
wn
in
T
ab
le
1
.
T
h
e
f
ea
tu
r
e
v
ec
to
r
h
as
a
f
ix
e
d
len
g
th
o
f
2
5
6
,
a
n
d
t
h
is
is
f
ed
to
th
e
FMM
f
o
r
tr
ain
in
g
.
T
h
e
FMM
wa
s
in
tr
o
d
u
ce
d
in
1
9
9
2
[
1
4
]
,
[
1
5
]
.
FMM
cr
ea
tes
f
u
zz
y
h
y
p
er
b
o
x
es
as
d
ec
is
io
n
r
eg
io
n
s
.
E
ac
h
class
i
s
r
ep
r
esen
ted
b
y
o
n
e
o
r
m
o
r
e
h
y
p
e
r
b
o
x
es
i
n
th
e
f
ea
tu
r
e
s
p
ac
e
d
e
p
en
d
in
g
o
n
a
p
r
ed
ef
in
e
d
h
y
p
er
b
o
x
s
ize.
Hy
p
er
b
o
x
es
a
r
e
d
escr
ib
ed
b
y
m
in
im
u
m
an
d
m
ax
im
u
m
p
o
i
n
ts
.
A
m
em
b
er
s
h
ip
f
u
n
ctio
n
th
en
q
u
an
tifie
s
h
o
w
m
u
ch
a
g
iv
e
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
2
0
2
6
:
4
7
0
0
-
4
7
1
0
4704
s
am
p
le
b
elo
n
g
s
to
ea
ch
h
y
p
er
b
o
x
,
h
en
ce
to
ea
ch
class
.
T
h
e
m
em
b
er
s
h
ip
v
alu
e
f
o
r
a
n
in
p
u
t
p
atter
n
is
s
aid
t
o
b
e
o
n
e
if
it is
f
u
lly
co
n
tain
ed
with
in
th
e
h
y
p
er
b
o
x
.
T
h
e
d
ef
i
n
itio
n
o
f
ea
c
h
h
y
p
er
b
o
x
f
u
zz
y
s
et
is
as in
(
6
)
.
=
{
,
,
,
(
,
,
)
}
∀
∈
(
6
)
W
h
er
e
an
d
ar
e
m
ax
a
n
d
m
i
n
p
o
in
ts
,
r
esp
ec
tiv
ely
.
A
h
y
p
e
r
b
o
x
in
th
r
ee
d
im
en
s
io
n
s
is
s
h
o
wn
in
Fig
u
r
e
5
.
T
ab
le
1
.
Deta
iled
co
n
f
ig
u
r
atio
n
o
f
th
e
C
NN
f
ea
tu
r
e
ex
tr
ac
to
r
La
y
e
r
t
y
p
e
K
e
r
n
e
l
/
p
o
o
l
si
z
e
S
t
r
i
d
e
O
u
t
p
u
t
s
h
a
p
e
P
a
r
a
me
t
e
r
s
P
u
r
p
o
se
I
n
p
u
t
N
/
A
N
/
A
4
8
×
4
8
(
1
f
e
a
t
u
r
e
m
a
p
)
0
I
n
p
u
t
i
m
a
g
e
(
g
r
a
y
s
c
a
l
e
)
C
o
n
v
1
3
×
3
1
4
8
×
4
8
(
3
2
f
e
a
t
u
r
e
ma
p
)
3
2
0
I
n
i
t
i
a
l
f
e
a
t
u
r
e
d
e
t
e
c
t
i
o
n
C
o
n
v
2
3
×
3
1
4
8
×
4
8
(
3
2
f
e
a
t
u
r
e
ma
p
)
9
,
2
4
8
D
e
e
p
e
n
s
f
e
a
t
u
r
e
m
a
p
M
a
x
P
o
o
l
1
2
×
2
2
2
4
×
2
4
(
3
2
f
e
a
t
u
r
e
ma
p
)
0
D
o
w
n
sa
mp
l
i
n
g
a
n
d
t
r
a
n
s
l
a
t
i
o
n
i
n
v
a
r
i
a
n
c
e
C
o
n
v
3
3
×
3
1
2
4
×
2
4
(
6
4
f
e
a
t
u
r
e
ma
p
)
1
8
,
4
9
6
I
n
c
r
e
a
se
d
f
e
a
t
u
r
e
c
o
m
p
l
e
x
i
t
y
C
o
n
v
4
3
×
3
1
2
4
×
2
4
(
6
4
f
e
a
t
u
r
e
ma
p
)
3
6
,
9
2
8
D
e
e
p
e
n
s
f
e
a
t
u
r
e
m
a
p
M
a
x
P
o
o
l
2
2
×
2
2
1
2
×
1
2
(
6
4
f
e
a
t
u
r
e
ma
p
)
0
F
u
r
t
h
e
r
d
o
w
n
sa
mp
l
i
n
g
C
o
n
v
5
3
×
3
1
1
2
×
1
2
(
1
2
8
f
e
a
t
u
r
e
ma
p
)
7
3
,
8
5
6
H
i
g
h
-
l
e
v
e
l
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
C
o
n
v
6
3
×
3
1
1
2
×
1
2
(
1
2
8
f
e
a
t
u
r
e
ma
p
)
1
4
7
,
5
8
4
D
e
e
p
e
n
s
f
e
a
t
u
r
e
m
a
p
M
a
x
P
o
o
l
3
2
×
2
2
6
×
6
(
1
2
8
f
e
a
t
u
r
e
m
a
p
)
0
F
i
n
a
l
s
p
a
t
i
a
l
r
e
d
u
c
t
i
o
n
C
o
n
v
7
3
×
3
1
6
×
6
(
2
5
6
f
e
a
t
u
r
e
m
a
p
)
2
9
5
,
1
6
8
F
i
n
a
l
f
e
a
t
u
r
e
ma
p
g
e
n
e
r
a
t
i
o
n
GAP
6
×
6
N
/
A
2
5
6
f
e
a
t
u
r
e
ma
p
0
I
n
t
e
r
f
a
c
e
t
o
F
M
M
N
N
2
5
6
f
e
a
t
u
r
e
m
a
p
Fig
u
r
e
5
.
A
th
r
ee
-
d
im
e
n
s
io
n
al
h
y
p
er
b
o
x
T
h
e
co
m
b
i
n
ed
f
u
zz
y
s
et
class
if
y
in
g
th
e
K
th
p
atter
n
class
,
,
is
d
ef
in
ed
b
y
a
p
p
ly
in
g
th
e
d
ef
in
itio
n
o
f
a
h
y
p
er
b
o
x
f
u
zz
y
s
et
as in
(
7
)
.
=
∪
∈
(
7
)
W
h
er
e
is
th
e
n
u
m
b
er
o
f
s
et
h
y
p
er
b
o
x
es
lin
k
e
d
to
class
.
T
h
e
in
d
e
x
s
et
o
f
th
o
s
e
h
y
p
er
b
o
x
es
ass
o
ciate
d
with
class
.
A
k
ey
asp
ec
t
o
f
t
h
is
m
eth
o
d
in
v
o
lv
es
id
en
tify
i
n
g
an
d
a
d
ju
s
tin
g
class
b
o
u
n
d
ar
ies,
as
d
ep
icted
in
Fig
u
r
e
6
.
Fig
u
r
e
6
.
T
h
e
d
ec
is
io
n
b
o
u
n
d
a
r
y
cr
ea
ted
b
y
FMM
I
n
FMM
,
if
two
h
y
p
e
r
b
o
x
es
o
f
th
e
s
am
e
class
o
v
er
lap
,
n
o
th
in
g
will
h
ap
p
en
,
o
n
th
e
o
th
er
h
an
d
if
o
v
er
lap
p
i
n
g
o
cc
u
r
s
b
etwe
en
d
if
f
er
en
t
class
es
,
th
en
a
co
n
tr
a
ctio
n
p
r
o
ce
s
s
is
in
itiated
.
T
h
e
d
eg
r
ee
to
wh
ich
a
p
atter
n
b
elo
n
g
s
to
a
h
y
p
er
b
o
x
is
m
ea
s
u
r
ed
b
y
th
e
m
em
b
er
s
h
ip
f
u
n
ctio
n
(
ℎ
)
.
T
h
e
m
em
b
er
s
h
i
p
f
u
n
ctio
n
is
d
ef
in
ed
as
(
8
)
.
m
ax
p
o
i
n
t
mi
n
p
o
i
n
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
F
a
cia
l e
mo
tio
n
r
ec
o
g
n
itio
n
th
r
o
u
g
h
a
r
tifi
cia
l
in
tellig
en
ce
(
A
n
a
s
Qu
teis
h
a
t
)
4705
(
ℎ
)
=
1
2
∑
[
(
0
,
1
−
(
0
,
(
1
,
ℎ
−
)
)
)
=
1
+
(
0
,
1
−
(
0
,
(
1
,
−
ℎ
)
)
)
]
(
8
)
T
h
e
co
n
tr
ac
tio
n
an
d
e
x
p
an
s
io
n
p
r
o
ce
s
s
es
u
s
e
th
e
lear
n
in
g
a
lg
o
r
ith
m
in
FMM
.
T
h
e
lea
r
n
i
n
g
p
r
o
ce
s
s
in
FMM
is
a
s
u
p
er
v
is
o
r
p
r
o
ce
s
s
wh
er
e
a
p
air
o
f
in
p
u
t
v
ec
to
r
s
an
d
an
ass
o
ciate
d
class
is
g
iv
en
{
ℎ
,
ℎ
}.
T
h
e
lear
n
in
g
p
r
o
ce
s
s
b
eg
in
s
b
y
cr
ea
tin
g
n
ew
h
y
p
e
r
b
o
x
es
f
o
r
ea
ch
class
,
as
th
e
tr
ain
in
g
p
atter
n
s
ar
e
p
r
esen
ted
to
th
e
n
etwo
r
k
,
eith
er
n
ew
h
y
p
er
b
o
x
es
ar
e
cr
ea
ted
o
f
th
e
c
u
r
r
e
n
t
h
y
p
er
b
o
x
es
ar
e
ex
p
a
n
d
ed
a
n
d
co
n
tr
ac
ted
.
T
h
is
is
co
n
tr
o
lled
b
y
th
e
s
ize
o
f
t
h
e
h
y
p
er
b
o
x
an
d
t
h
e
n
u
m
b
er
o
f
class
es
th
e
d
ata
is
as
s
o
ciate
d
with
.
T
h
e
ex
p
a
n
s
io
n
cr
iter
ia
is
r
estricte
d
b
y
a
r
e
q
u
i
r
em
en
t th
at
as in
(
9
)
.
≥
∑
(
ma
x
(
,
ℎ
)
−
min
(
,
ℎ
)
)
=
1
(
9
)
W
h
er
e
0
≤
≤
1
is
th
e
h
y
p
er
b
o
x
s
ize.
Sh
o
u
ld
n
o
h
y
p
e
r
b
o
x
b
e
f
o
u
n
d
to
m
ee
t
t
h
e
ex
p
an
s
io
n
cr
iter
i
a,
a
n
ew
h
y
p
er
b
o
x
is
cr
ea
ted
i
n
th
e
n
et
wo
r
k
.
C
lass
es
ca
n
b
e
f
o
r
m
ed
th
at
ar
en
'
t
lin
ea
r
ly
s
ep
ar
ab
le
t
h
r
o
u
g
h
th
is
g
r
o
wth
p
r
o
ce
s
s
.
New
class
es
ca
n
b
e
ad
d
e
d
with
o
u
t
r
etr
ain
in
g
.
Ov
er
lap
p
in
g
is
a
p
o
s
s
ib
ilit
y
wh
en
ex
p
an
s
io
n
is
p
r
esen
ted
.
T
o
s
o
lv
e
th
is
is
s
u
e,
an
o
v
e
r
lap
test
is
co
n
d
u
cte
d
,
an
d
if
th
e
o
v
er
lap
o
cc
u
r
r
ed
b
etwe
en
d
if
f
er
e
n
t
class
es,
th
en
co
n
tr
ac
tio
n
is
d
o
n
e;
o
th
er
wis
e,
th
er
e
is
n
o
n
ee
d
to
p
er
f
o
r
m
a
n
y
o
p
er
atio
n
.
Alg
o
r
ith
m
1
s
h
o
ws th
e
p
s
eu
d
o
co
d
e
f
o
r
th
e
FMM
NN.
Alg
o
r
ith
m
1
.
Ps
eu
d
o
co
d
e
f
o
r
th
e
FMM
NN
Step 1: Initialization
Initialize an empty hyperbox set
ℋ
=
∅
Step 2: For each training sample
(
,
)
∈
Membership Evaluation: Compute membership value of
to all hyperboxes in
ℋ
that
belong to class
Hyperbox Selection: Select the hyperbox
ℎ
∗
with the highest membership value
Expansion Check: If expanding
ℎ
∗
with
satisfies the size constraint:
max
(
−
)
≤
∀
then expand the hyperbox, else, create a new hyperbox with
=
,
=
Overlap Test: Check for overlaps between the expanded hyperbox and hyperboxes
of other classes
Contraction: If overlap exists, apply contraction rules to eliminate overlap
while minimizing distortion
Step 3: Repeat
Continue until all samples are processed
Step 4: Return
Return the final set of hyperboxes
ℋ
W
h
en
d
esig
n
in
g
a
FMM
NN
th
er
e
ar
e
s
ev
er
al
h
y
p
er
p
ar
a
m
eter
s
th
at
n
ee
d
to
b
e
c
h
o
s
en
.
T
h
e
m
o
s
t
cr
itical
h
y
p
er
p
ar
a
m
eter
is
th
e
h
y
p
er
b
o
x
s
ize
θ.
As m
en
tio
n
e
d
ab
o
v
e
θ
ca
n
b
e
an
y
v
alu
e
f
r
o
m
0
to
1
.
W
h
en
θ
is
v
er
y
s
m
all
(
a
v
al
u
e
clo
s
e
t
o
0
)
,
th
e
n
etwo
r
k
will
ten
d
to
c
r
ea
te
a
h
y
p
er
b
o
x
f
o
r
ea
ch
in
p
u
t
s
am
p
le,
th
is
will
ca
u
s
e
o
v
er
f
itti
n
g
in
th
e
n
etwo
r
k
.
On
th
e
o
th
er
h
a
n
d
,
wh
en
θ
is
lar
g
e
,
th
e
n
etwo
r
k
will
b
e
s
m
aller
an
d
f
aster
to
lear
n
b
u
t
lo
s
es
th
e
ab
ilit
y
to
co
r
r
ec
tly
class
if
y
p
atter
n
s
in
o
th
er
wo
r
d
s
,
t
h
e
n
etwo
r
k
is
o
v
er
f
itted
.
T
h
e
m
o
s
t
co
m
m
o
n
l
y
u
s
ed
s
ize
f
o
r
th
e
h
y
p
er
b
o
x
es
is
0
.
0
5
,
th
is
s
ize
is
co
n
s
id
er
ed
m
o
d
er
ate
,
s
o
th
e
n
etwo
r
k
ca
n
lear
n
,
an
d
h
y
p
er
b
o
x
es
will
b
e
cr
ea
t
ed
,
ex
p
a
n
d
ed
,
an
d
co
n
tacte
d
ac
co
r
d
in
g
l
y
.
T
h
e
s
ec
o
n
d
h
y
p
er
p
ar
am
eter
is
th
e
s
en
s
itiv
ity
p
ar
am
eter
γ
.
T
h
e
p
ar
am
eter
is
u
s
ed
in
th
e
m
em
b
er
s
h
ip
f
u
n
ctio
n
;
it
c
o
n
tr
o
ls
h
o
w
f
ast
th
e
m
em
b
er
s
h
ip
v
alu
e
d
ec
r
ea
s
es
wh
en
th
e
in
p
u
t
lies
o
u
ts
id
e
a
h
y
p
er
b
o
x
b
o
u
n
d
a
r
y
.
T
h
e
v
alu
e
o
f
γ
v
ar
ies
b
etwe
en
0
.
5
an
d
5
.
I
n
th
is
p
a
p
er
γ
=
1
,
s
in
ce
th
is
is
th
e
ty
p
ical
v
alu
e
c
h
o
s
en
in
th
e
liter
atu
r
e.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
ex
p
er
im
en
ts
co
n
d
u
cted
wer
e
b
ased
o
n
th
e
FER2
0
1
3
d
ataset,
wh
ich
was
in
tr
o
d
u
ce
d
as
p
ar
t
o
f
an
in
ter
n
atio
n
al
co
n
f
er
en
ce
o
n
m
ac
h
i
n
e
lear
n
in
g
r
ep
r
esen
tatio
n
-
lear
n
in
g
ch
allen
g
e
an
d
m
a
d
e
p
u
b
lic
in
2013
[
2
0
]
.
T
h
e
d
ataset
co
n
ta
in
s
3
5
,
8
8
7
g
r
ay
s
ca
le
f
ac
e
im
ag
es
s
ized
4
8
×
4
8
p
ix
els,
l
ab
elled
with
s
ev
en
em
o
tio
n
class
es:
an
g
er
,
d
is
g
u
s
t,
f
ea
r
,
h
ap
p
in
ess
,
s
ad
n
ess
,
s
u
r
p
r
is
e,
an
d
n
eu
t
r
al.
I
m
a
g
es
wer
e
g
ath
er
ed
f
r
o
m
th
e
web
an
d
au
to
m
atica
lly
r
e
g
is
ter
ed
s
o
t
h
at
f
ac
es
ar
e
a
p
p
r
o
x
im
ately
ce
n
ter
ed
.
T
h
e
o
r
ig
i
n
al
d
escr
ip
tio
n
an
d
d
ataset
s
tati
s
tics
(
p
er
-
class
co
u
n
ts
)
ar
e
r
ep
o
r
ted
in
t
h
e
ch
al
len
g
e
p
ap
e
r
/an
n
o
u
n
ce
m
e
n
t.
FER2
0
1
3
b
ec
am
e
a
d
e
f
ac
to
b
en
ch
m
a
r
k
f
o
r
d
ee
p
lear
n
in
g
FER
wo
r
k
b
ec
au
s
e
it
is
lar
g
e,
r
ea
lis
tic,
an
d
ea
s
ily
ac
ce
s
s
ib
le.
T
h
es
e
q
u
alities
m
ad
e
it
a
co
m
m
o
n
te
s
tb
ed
f
o
r
c
o
n
v
o
lu
tio
n
al
a
r
ch
it
ec
tu
r
es,
tr
an
s
f
er
lear
n
in
g
,
an
d
d
ata
-
au
g
m
en
tatio
n
s
tr
ateg
ies
[
2
1
]
.
T
o
f
u
r
th
e
r
im
p
r
o
v
e
th
e
d
ataset,
av
o
id
th
e
r
i
s
k
o
f
o
v
er
f
itti
n
g
,
wh
ic
h
is
a
c
o
m
m
o
n
is
s
u
e
with
d
atasets
tr
ain
ed
o
n
lim
ited
d
atasets
,
an
d
to
en
h
an
ce
th
e
m
o
d
el’
s
r
o
b
u
s
tn
ess
to
v
ar
iatio
n
s
in
f
ac
ial
p
o
s
es
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
2
0
2
6
:
4
7
0
0
-
4
7
1
0
4706
lig
h
tin
g
,
a
s
et
o
f
d
ata
au
g
m
e
n
tatio
n
tech
n
iq
u
es is
ap
p
lied
to
th
e
tr
ain
in
g
s
et
[
2
2
]
.
T
a
b
le
2
s
h
o
ws th
e
g
eo
m
etr
ic
an
d
p
h
o
to
m
etr
ic
tr
a
n
s
f
o
r
m
atio
n
s
ap
p
lied
r
a
n
d
o
m
l
y
to
ea
ch
tr
ain
in
g
im
ag
e
a
f
ter
ea
ch
e
p
o
ch
.
T
ab
le
2
.
Au
g
m
en
tatio
n
tec
h
n
i
q
u
es a
p
p
lied
A
u
g
m
e
n
t
a
t
i
o
n
t
e
c
h
n
i
q
u
e
P
a
r
a
me
t
e
r
/
r
a
n
g
e
P
u
r
p
o
se
R
a
n
d
o
m
h
o
r
i
z
o
n
t
a
l
f
l
i
p
P
=
0
.
5
S
i
mu
l
a
t
e
s f
a
c
e
s
l
o
o
k
i
n
g
i
n
t
h
e
o
p
p
o
s
i
t
e
d
i
r
e
c
t
i
o
n
,
i
n
c
r
e
a
si
n
g
p
o
s
e
i
n
v
a
r
i
a
n
c
e
R
a
n
d
o
m r
o
t
a
t
i
o
n
±
1
0
d
e
g
r
e
e
s
A
c
c
o
u
n
t
s
f
o
r
s
l
i
g
h
t
h
e
a
d
t
i
l
t
s a
n
d
m
i
n
o
r
p
o
s
e
v
a
r
i
a
t
i
o
n
s
R
a
n
d
o
m w
i
d
t
h
/
h
e
i
g
h
t
sh
i
f
t
±
0
.
1
(
1
0
%
o
f
t
o
t
a
l
)
S
i
mu
l
a
t
e
s m
i
n
o
r
mi
s
a
l
i
g
n
me
n
t
s i
n
f
a
c
e
d
e
t
e
c
t
i
o
n
/
c
r
o
p
p
i
n
g
R
a
n
d
o
m
z
o
o
m
0
.
9
t
o
1
.
1
S
i
mu
l
a
t
e
s
v
a
r
i
a
t
i
o
n
s
i
n
d
i
s
t
a
n
c
e
f
r
o
m
t
h
e
c
a
mera
B
r
i
g
h
t
n
e
ss
a
d
j
u
s
t
me
n
t
±
0
.
2
I
n
c
r
e
a
ses
r
o
b
u
s
t
n
e
ss
t
o
v
a
r
y
i
n
g
l
i
g
h
t
i
n
g
c
o
n
d
i
t
i
o
n
s
T
h
e
au
g
m
en
tatio
n
will
n
o
t
af
f
ec
t
th
e
t
o
tal
n
u
m
b
er
o
f
im
ag
es
in
th
e
d
ataset,
in
s
tead
,
ev
e
r
y
tim
e
an
im
ag
e
is
p
r
esen
ted
in
a
n
ew
ep
o
ch
,
a
r
an
d
o
m
tr
a
n
s
f
o
r
m
ati
o
n
is
ap
p
lied
t
o
th
is
im
ag
e
.
T
h
is
g
u
ar
an
tees
th
e
C
NN
n
ev
er
s
ee
s
th
e
s
am
e
im
a
g
e
twice.
T
h
e
d
ataset
was
d
iv
id
ed
in
to
two
s
ets,
8
0
%
o
f
th
e
d
ata
was
u
s
ed
f
o
r
tr
ain
in
g
(
2
8
,
7
0
9
s
am
p
les,
d
iv
i
d
ed
in
ev
e
n
p
r
o
p
o
r
tio
n
s
b
etwe
en
s
ev
en
class
es)
an
d
2
0
%
o
f
th
e
d
ata
f
o
r
test
in
g
.
Sev
er
al
to
o
ls
a
n
d
tech
n
iq
u
es
wer
e
u
s
ed
in
th
e
m
o
d
el
tr
ai
n
in
g
p
r
o
ce
s
s
;
a
n
ac
o
n
d
a
in
teg
r
ated
d
e
v
elo
p
m
e
n
t
en
v
ir
o
n
m
en
t
(
I
DE
)
was
u
s
ed
to
im
p
lem
e
n
t
th
e
FMM
DNN.
Op
en
C
V
lib
r
a
r
y
was
u
s
ed
t
o
in
clu
d
e
th
e
Haa
r
ca
s
ca
d
e
alg
o
r
ith
m
,
an
d
th
e
f
u
l
ly
co
n
n
ec
ted
n
etwo
r
k
was
r
ep
l
ac
ed
b
y
th
e
FMM
NN.
T
h
e
F
MM
NN
was
s
et
to
a
h
y
p
er
b
o
x
s
ize
o
f
0
.
0
5
.
T
h
e
m
o
d
el
was
tr
ain
ed
an
d
test
ed
1
0
tim
es,
th
e
co
n
f
u
s
io
n
m
at
r
ix
f
o
r
t
h
e
av
er
a
g
e
r
esu
lts
is
s
h
o
wn
in
T
ab
le
3
.
T
ab
le
3
.
C
o
n
f
u
s
io
n
m
atr
ix
f
o
r
th
e
FER2
0
1
3
d
ataset
Tr
u
e
/
P
r
e
d
i
c
t
e
d
A
n
g
r
y
D
i
sg
u
st
F
e
a
r
H
a
p
p
y
S
a
d
S
u
r
p
r
i
s
e
N
e
u
t
r
a
l
R
e
c
a
l
l
(
%)
A
n
g
r
y
5
7
4
.
8
4
7
.
9
1
2
4
.
5
4
2
8
.
7
4
5
7
.
4
8
2
8
.
7
4
9
5
.
8
60
D
i
sg
u
st
5
.
5
5
6
7
.
7
1
6
.
6
6
2
.
2
2
1
5
.
5
4
3
.
3
3
9
.
9
9
61
F
e
a
r
1
1
2
.
6
4
3
0
.
7
2
6
1
4
.
4
3
0
.
7
2
1
1
2
.
6
4
6
1
.
4
4
6
1
.
4
4
60
H
a
p
p
y
5
3
.
2
2
1
7
.
7
4
3
5
.
4
8
1
5
9
6
.
6
3
5
.
4
8
1
7
.
7
4
1
7
.
7
4
90
S
a
d
1
3
7
.
1
7
2
4
.
9
4
8
7
.
2
9
3
7
.
4
1
7
7
3
.
1
4
3
7
.
4
1
1
4
9
.
6
4
62
S
u
r
p
r
i
s
e
3
3
.
2
8
1
6
.
6
4
9
1
.
5
2
2
4
.
9
6
1
6
.
6
4
6
3
2
.
3
2
1
6
.
6
4
76
N
e
u
t
r
a
l
1
3
5
.
5
2
3
6
.
9
6
2
4
.
6
4
3
6
.
9
6
7
3
.
9
2
3
6
.
9
6
8
8
7
.
0
4
72
P
r
e
c
i
s
i
o
n
(
%)
42
67
48
90
48
78
53
T
h
e
co
n
f
u
s
io
n
m
atr
ix
r
ev
ea
ls
d
is
tin
ct
p
atter
n
s
o
f
m
is
class
if
icatio
n
th
at
a
r
e
c
o
m
m
o
n
in
th
e
FER
s
y
s
tem
s
,
th
is
p
r
o
v
id
es
a
d
ee
p
er
u
n
d
er
s
tan
d
in
g
o
f
th
e
m
o
d
e
l’
s
d
ec
is
io
n
b
o
u
n
d
ar
y
.
T
h
e
m
o
d
el
s
h
o
ws
th
at
th
e
h
ap
p
y
em
o
tio
n
h
as
th
e
h
ig
h
es
t
ac
cu
r
ac
y
s
in
ce
a
r
o
u
n
d
2
5
%
o
f
th
e
im
a
g
es
ar
e
o
f
h
a
p
p
y
e
m
o
tio
n
th
is
g
a
v
e
th
e
m
o
d
el
lar
g
e
n
u
m
b
e
r
o
f
t
r
ain
in
g
s
am
p
les.
On
th
e
o
th
er
h
a
n
d
,
d
is
g
u
s
t
h
a
d
th
e
lo
west
n
u
m
b
e
r
o
f
s
am
p
les
in
th
e
d
ataset
with
a
p
er
ce
n
tag
e
o
f
1
.
5
%
o
f
t
h
e
3
5
8
8
7
im
ag
es.
T
h
e
m
o
d
el
class
if
ied
d
is
g
u
s
t
w
ith
an
ac
cu
r
ac
y
o
f
6
1
%
wh
ich
is
s
im
ilar
to
th
e
o
th
er
em
o
tio
n
s
.
T
h
is
s
h
o
ws
th
at
th
e
m
o
d
el
is
n
o
t
af
f
ec
ted
b
y
th
e
lo
w
n
u
m
b
er
o
f
s
am
p
les,
th
is
is
d
u
e
to
th
e
u
s
e
o
f
au
g
m
en
tatio
n
in
th
e
d
ataset
.
T
o
f
u
r
th
er
an
aly
ze
th
e
p
er
f
o
r
m
an
ce
o
f
t
h
e
m
o
d
el
r
ec
all
,
p
r
ec
is
io
n
,
an
d
F1
-
s
co
r
e
a
r
e
u
s
ed
.
T
h
ese
q
u
an
titi
es a
r
e
d
ef
in
e
d
as f
o
llo
ws.
i)
R
ec
all:
th
e
r
atio
o
f
co
r
r
ec
t
p
r
e
d
ictio
n
s
f
o
r
a
s
p
ec
if
ic
em
o
tio
n
to
th
e
to
tal
n
u
m
b
er
o
f
im
ag
es
th
at
ac
tu
ally
b
elo
n
g
t
o
th
at
em
o
tio
n
.
R
ec
all
is
d
ef
in
ed
as (
1
0
)
.
=
+
(
1
0
)
ii)
Pre
cisi
o
n
:
th
e
r
atio
o
f
c
o
r
r
ec
t
p
r
ed
ictio
n
s
f
o
r
a
s
p
ec
if
ic
em
o
tio
n
to
t
h
e
to
tal
n
u
m
b
e
r
o
f
p
r
ed
ictio
n
s
th
at
wer
e
class
if
ied
as b
elo
n
g
in
g
t
o
th
at
em
o
tio
n
[
2
3
]
.
Pre
cisi
o
n
is
d
ef
in
ed
as (
1
1
)
.
=
+
(
1
1
)
iii)
F1
-
s
co
r
e:
r
ep
r
esen
ts
th
e
h
ar
m
o
n
ic
m
ea
n
o
f
p
r
ec
is
io
n
an
d
r
e
ca
ll,
an
d
is
u
s
ed
as
a
p
er
f
o
r
m
an
ce
m
etr
ic
t
o
ev
alu
ate
class
if
icatio
n
m
o
d
els with
im
b
alan
ce
d
d
ataset.
I
t is
d
ef
in
ed
as (
1
2
)
.
1
−
s
c
or
e
=
2
×
×
+
(
1
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
F
a
cia
l e
mo
tio
n
r
ec
o
g
n
itio
n
th
r
o
u
g
h
a
r
tifi
cia
l
in
tellig
en
ce
(
A
n
a
s
Qu
teis
h
a
t
)
4707
T
ab
le
4
s
h
o
ws
th
e
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
f
o
r
th
e
FER
d
ataset.
T
h
e
F1
-
s
co
r
e
is
u
s
ed
to
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
class
if
icatio
n
m
o
d
els.
Sin
ce
t
h
e
FER2
0
1
3
d
ataset
is
an
im
b
alan
ce
d
d
ataset,
th
e
r
esu
lts
in
t
h
e
co
n
f
u
s
io
n
m
atr
ix
will
n
o
t
r
ef
lect
th
e
f
u
ll
s
to
r
y
b
e
h
in
d
th
e
r
esu
lts
.
T
h
e
p
r
o
p
o
s
ed
m
o
d
el
ac
h
iev
es
ex
ce
p
tio
n
al
p
er
f
o
r
m
an
ce
o
n
h
ap
p
y
(
9
0
%
r
ec
all,
9
0
%
p
r
ec
is
io
n
,
an
d
9
0
%
F1
-
s
co
r
e)
an
d
s
tr
o
n
g
p
er
f
o
r
m
an
ce
o
n
s
u
r
p
r
is
e
(
7
6
%
r
ec
all,
7
8
%
p
r
ec
is
io
n
,
an
d
7
7
%
F1
-
s
co
r
e)
.
T
h
ese
em
o
tio
n
s
ar
e
ch
ar
ac
ter
ized
b
y
h
ig
h
l
y
d
is
tin
ct
an
d
ea
s
ily
r
ec
o
g
n
izab
le
f
ac
ial
ac
tio
n
u
n
its
(
AUs):
f
o
r
in
s
tan
ce
,
h
ap
p
y
h
as
ch
ee
k
r
aiser
(
AU6
)
an
d
lip
co
r
n
e
r
p
u
ller
(
AU1
2
)
,
wh
ile
s
u
r
p
r
is
e
h
as
i
n
n
er
b
r
o
w
r
aiser
(
AU1
)
,
o
u
ter
b
r
o
w
r
aiser
(
AU2
)
,
an
d
u
p
p
er
lid
r
aiser
(
AU5
)
.
T
h
e
C
NN
f
ea
tu
r
e
ex
tr
ac
to
r
s
u
cc
ess
f
u
lly
is
o
lates
th
ese
u
n
iq
u
e
f
ea
tu
r
es,
r
esu
ltin
g
in
a
well
-
s
ep
ar
ated
h
y
p
er
b
o
x
i
n
th
e
FMM
NN
f
ea
tu
r
e
s
p
ac
e.
T
ab
le
4
.
C
lass
if
icatio
n
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
Emo
t
i
o
n
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
sc
o
r
e
(
%)
A
n
g
r
y
42
60
49
D
i
sg
u
st
67
61
64
F
e
a
r
48
60
53
H
a
p
p
y
90
90
90
S
a
d
48
62
54
S
u
r
p
r
i
s
e
78
76
77
N
e
u
t
r
a
l
53
72
61
On
e
ca
n
also
n
o
tice
f
r
o
m
T
ab
l
es
1
an
d
2
th
at
th
e
f
ea
r
an
d
s
u
r
p
r
is
e
em
o
tio
n
s
h
av
e
s
ig
n
if
ica
n
t o
v
er
lap
,
th
is
is
o
b
v
io
u
s
s
in
ce
th
e
m
o
d
e
l
m
is
class
if
ied
9
1
.
5
2
f
ea
r
s
am
p
les
as
s
u
r
p
r
is
e
s
am
p
les.
T
h
is
ca
n
b
e
ju
s
tifie
d
b
y
lo
o
k
in
g
at
th
e
b
io
lo
g
ical
asp
ec
ts
o
f
b
o
t
h
f
ea
r
an
d
s
u
r
p
r
is
e
em
o
tio
n
s
.
B
o
th
em
o
tio
n
s
s
h
ar
e
k
ey
AUs,
n
o
tab
l
y
AU1
an
d
AU2
,
wh
ic
h
ca
u
s
e
t
h
e
ey
es
to
wid
en
.
T
h
e
p
r
im
a
r
y
d
if
f
er
en
ce
lies
in
th
e
m
o
u
th
s
h
ap
e
(
s
tr
etch
e
d
in
f
ea
r
,
d
r
o
p
p
ed
jaw
in
s
u
r
p
r
is
e
)
.
T
h
e
lo
w
-
r
eso
lu
tio
n
4
8
×
4
8
im
ag
es,
c
o
m
b
in
e
d
with
th
e
i
n
h
er
en
t
s
u
b
tlety
o
f
th
ese
d
if
f
er
en
ce
s
,
lead
to
o
v
er
l
ap
p
in
g
f
ea
t
u
r
e
v
ec
to
r
s
in
th
e
2
5
6
s
p
ac
e.
T
h
is
in
tu
r
n
s
u
g
g
ests
th
at
th
e
FMM
NN
h
y
p
er
b
o
x
es
f
o
r
f
ea
r
an
d
s
u
r
p
r
is
e
ar
e
ad
jace
n
t
o
r
s
lig
h
tl
y
in
ter
s
ec
t
in
th
e
f
ea
tu
r
e
s
p
ac
e,
r
ef
lectin
g
th
e
am
b
ig
u
ity
o
f
th
e
in
p
u
t d
ata
its
elf
.
T
h
e
class
es
an
g
r
y
(
4
2
%
p
r
ec
is
io
n
)
,
f
ea
r
(
4
8
%
p
r
ec
is
io
n
)
,
a
n
d
s
ad
(
4
8
%
p
r
e
cisi
o
n
)
ex
h
ib
it
th
e
lo
west
p
r
ec
is
io
n
s
co
r
es.
T
h
is
in
d
icate
s
th
at
th
e
m
o
d
el
f
r
eq
u
en
tly
m
i
s
class
if
ies o
th
er
em
o
tio
n
s
as th
ese
th
r
ee
.
A
clo
s
er
lo
o
k
at
th
e
an
g
r
y
e
m
o
tio
n
o
n
e
ca
n
n
o
tice
t
h
at
th
e
m
o
d
el
m
is
class
if
ies
1
3
7
.
1
7
s
ad
s
am
p
les
an
d
1
1
2
.
6
4
f
ea
r
s
am
p
les
as
an
g
r
y
.
T
h
is
is
lik
ely
d
u
e
to
th
e
s
h
ar
ed
p
r
esen
c
e
o
f
b
r
o
w
lo
wer
e
r
(
AU4
)
i
n
b
o
th
a
n
g
r
y
a
n
d
s
ad
ex
p
r
ess
io
n
s
,
an
d
th
e
g
en
e
r
al
in
ten
s
ity
o
f
n
e
g
ativ
e
em
o
ti
o
n
s
.
Similar
ly
,
in
th
e
s
ad
e
m
o
tio
n
th
e
m
o
d
e
l
m
is
class
if
ies
1
4
9
.
6
4
n
eu
tr
al
s
am
p
les
as
s
ad
.
T
h
is
is
a
co
m
m
o
n
c
h
allen
g
e,
as
a
"
n
eu
tr
al
"
f
ac
e
with
s
lig
h
t
f
atig
u
e
o
r
r
elax
atio
n
ca
n
b
e
e
asil
y
co
n
f
u
s
ed
with
a
lo
w
-
in
ten
s
ity
"
s
ad
"
ex
p
r
ess
io
n
,
esp
ec
ially
in
th
e
ab
s
en
ce
o
f
d
y
n
am
ic
in
f
o
r
m
atio
n
.
T
o
f
u
r
th
er
th
e
o
v
er
all
ac
cu
r
a
cy
o
f
th
e
s
y
s
tem
co
m
p
ar
ed
with
ac
cu
r
ac
ies
f
r
o
m
th
e
liter
atu
r
e
ar
e
s
h
o
wn
in
T
ab
le
5
.
T
h
e
r
esu
lts
s
h
o
w
th
at
th
e
p
r
o
p
o
s
ed
d
ee
p
FMM
clas
s
if
icatio
n
s
y
s
tem
g
av
e
r
esu
lts
th
at
ar
e
v
er
y
n
ea
r
to
th
e
r
esu
lts
p
u
b
lis
h
ed
in
th
e
liter
atu
r
e.
T
h
is
s
h
o
ws
th
at
wh
en
o
th
er
v
e
r
s
io
n
s
o
f
th
e
FMM
NN
ar
e
u
s
ed
th
e
r
esu
lts
will
im
p
r
o
v
e,
as
th
e
n
ewe
r
v
er
s
io
n
s
o
f
th
e
n
etwo
r
k
h
a
v
e
h
i
g
h
er
ac
cu
r
ac
y
an
d
h
a
v
e
th
e
ab
ilit
y
to
ex
p
lain
th
e
o
b
tain
e
d
r
esu
lts
.
T
ab
le
5
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
b
etwe
en
o
b
tain
ed
r
esu
lts
a
n
d
p
u
b
lis
h
ed
r
esu
lts
P
a
p
e
r
/
M
o
d
e
l
Te
st
a
c
c
u
r
a
c
y
(
%)
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F1
-
sc
o
r
e
(
%)
R
e
sN
e
t
1
5
2
[
2
0
]
39
69
68
68
VGG
-
19
[
2
0
]
44
71
72
70
Tr
a
n
sf
e
r
l
e
a
r
n
i
n
g
(
A
l
e
x
N
e
t
)
[
2
4
]
66
-
-
-
D
e
e
p
e
r
C
N
N
[
2
5
]
75
-
-
-
R
e
si
d
u
a
l
a
t
t
e
n
t
i
o
n
(
R
A
N
)
[
2
6
]
72
-
-
-
S
e
l
f
-
c
u
r
e
n
e
t
w
o
r
k
(
S
C
N
)
[
2
7
]
89
-
-
-
F
M
M
D
N
N
71
61
69
64
As
ex
p
lain
ed
ea
r
lier
,
t
h
e
p
a
p
er
in
tr
o
d
u
ce
s
th
e
FMM
NN
in
s
tead
o
f
th
e
s
o
f
tm
a
x
lay
er
,
th
e
m
ain
ad
v
an
tag
e
o
f
u
s
in
g
th
e
FMM
NN
is
th
e
g
e
o
m
etr
ic
in
te
r
p
r
et
ab
ilit
y
an
d
a
r
o
b
u
s
t
m
ec
h
an
is
m
f
o
r
u
n
ce
r
tain
ty
q
u
an
tific
atio
n
in
th
e
class
if
icatio
n
p
r
o
ce
s
s
.
T
h
e
C
NN
is
co
n
s
id
er
ed
a
b
lack
b
o
x
s
in
ce
weig
h
t
m
atr
ices
ar
e
o
p
aq
u
e
.
T
h
e
FMM
NN
o
n
th
e
o
th
er
h
a
n
d
,
in
tr
o
d
u
c
es
g
eo
m
et
r
ical
in
ter
p
r
etab
ilit
y
to
h
an
d
le
th
e
u
n
ce
r
tain
t
y
o
f
th
e
class
if
icatio
n
p
r
o
ce
s
s
.
E
ac
h
class
is
p
r
esen
ted
at
least
o
n
e
h
y
p
e
r
b
o
x
,
an
d
a
d
ec
is
io
n
b
o
u
n
d
ar
y
is
cr
ea
ted
b
etwe
en
th
ese
h
y
p
er
b
o
x
es.
Fu
r
th
er
m
o
r
e,
s
in
ce
th
e
h
y
p
er
b
o
x
es
ar
e
b
o
u
n
d
ed
b
y
m
i
n
im
u
m
an
d
m
ax
im
u
m
p
o
in
ts
f
u
zz
y
if
-
th
en
r
u
les ca
n
b
e
ex
tr
ac
ted
to
ju
s
tify
th
e
r
ese
ar
ch
d
ec
is
io
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
15
,
No
.
5
,
Octo
b
er
2
0
2
6
:
4
7
0
0
-
4
7
1
0
4708
I
n
th
e
s
o
f
tm
ax
class
if
ier
u
n
ce
r
tain
ty
is
in
f
er
r
ed
f
r
o
m
th
e
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
,
h
o
wev
er
,
in
FMM
NN,
u
n
ce
r
tain
ty
is
ex
p
licitly
q
u
an
tifie
d
b
y
th
e
f
u
zz
y
m
em
b
er
s
h
ip
f
u
n
ctio
n
,
w
h
ich
q
u
an
tifie
s
th
e
d
eg
r
ee
to
wh
ich
a
n
in
p
u
t
f
ea
tu
r
e
v
e
cto
r
b
elo
n
g
s
to
a
h
y
p
er
b
o
x
.
T
o
f
u
r
th
er
illu
s
tr
ate
th
is
p
o
in
t,
two
s
am
p
les
ar
e
s
tu
d
ied
f
r
o
m
th
e
d
ataset
s
h
o
wn
in
T
ab
le
6
.
T
h
e
f
ir
s
t
s
am
p
le
is
a
c
lear
s
am
p
le
with
a
tr
u
e
class
“
h
ap
p
y
”,
th
is
s
am
p
le
th
e
s
o
f
tm
ax
p
r
o
b
ab
i
lity
was
0
.
9
8
an
d
th
e
FM
MN
N
m
em
b
er
s
h
ip
was
0
.
9
9
f
o
r
th
e
m
a
x
im
u
m
m
em
b
er
s
h
ip
.
T
h
e
s
ec
o
n
d
s
am
p
le
was
an
am
b
ig
u
o
u
s
s
am
p
le,
in
th
is
s
am
p
le
th
e
s
o
f
tm
ax
g
av
e
a
m
ax
im
u
m
p
r
o
b
a
b
ilit
y
o
f
0
.
5
5
,
wh
ile
th
e
FMM
NN
m
em
b
er
s
h
ip
was
0
.
4
8
f
o
r
t
h
e
co
r
r
ec
t
class
(
f
ea
r
)
an
d
0
.
4
5
f
o
r
wr
o
n
g
f
ea
tu
r
e.
T
h
is
h
i
g
h
co
m
p
etitio
n
b
etwe
en
th
e
class
es p
r
o
v
id
es
a
m
o
r
e
tr
a
n
s
p
ar
en
t m
ea
s
u
r
e
o
f
u
n
ce
r
tain
ty
.
T
ab
le
6
.
C
o
m
p
a
r
is
o
n
b
etwe
en
s
o
f
tm
ax
an
d
FMM
NN
u
n
ce
r
ta
in
ty
h
an
d
lin
g
S
a
mp
l
e
Tr
u
e
c
l
a
ss
S
o
f
t
M
a
x
m
a
x
p
r
o
b
a
b
i
l
i
t
y
F
M
M
N
N
ma
x
me
mb
e
r
sh
i
p
F
M
M
N
N
s
e
c
o
n
d
ma
x
m
e
mb
e
r
sh
i
p
C
l
e
a
r
sa
mp
l
e
H
a
p
p
y
0
.
9
8
(
h
a
p
p
y
)
0
.
9
9
(
h
a
p
p
y
)
0
.
0
1
(
n
e
u
t
r
a
l
)
A
mb
i
g
u
o
u
s s
a
mp
l
e
F
e
a
r
0
.
5
5
(
f
e
a
r
)
0
.
4
8
(
f
e
a
r
)
0
.
4
5
(
su
r
p
r
i
se)
4.
CO
NCLU
SI
O
N
T
h
is
p
ap
er
p
r
esen
te
d
a
d
ee
p
le
ar
n
in
g
n
e
u
r
al
n
etwo
r
k
b
ased
o
n
C
NN
an
d
FMM
.
T
h
e
s
o
f
tm
a
x
lay
er
in
th
e
C
NN
was
r
ep
lace
d
b
y
an
FMM
NN
.
T
h
e
p
r
o
p
o
s
ed
n
etw
o
r
k
will
u
s
e
th
e
ex
tr
ac
ted
f
ea
t
u
r
es
f
r
o
m
t
h
e
C
NN
an
d
cr
ea
te
a
n
d
r
ef
in
e
h
y
p
e
r
b
o
x
es.
T
h
e
c
r
ea
ted
h
y
p
e
r
b
o
x
es
ar
e
th
en
ex
p
a
n
d
ed
an
d
co
n
tr
ac
ted
ac
co
r
d
i
n
g
to
th
eir
ass
o
ciate
d
class
e
s
.
T
h
e
p
r
o
p
o
s
ed
n
etwo
r
k
is
test
ed
o
n
th
e
FER
d
ataset.
T
h
e
r
esu
lts
o
b
tain
ed
s
h
o
w
th
a
t
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
ca
n
d
et
ec
t
em
o
tio
n
s
in
p
ictu
r
es
with
g
o
o
d
ac
cu
r
ac
y
.
Fu
r
th
er
m
o
r
e,
o
th
er
ty
p
es
o
f
FMM
n
eu
r
al
n
etwo
r
k
s
ca
n
b
e
u
s
ed
t
o
im
p
r
o
v
e
th
e
r
esu
lts
in
f
u
tu
r
e
wo
r
k
,
an
d
o
th
er
d
atasets
ca
n
b
e
u
s
ed
to
tr
ain
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
An
as Qu
teis
h
at
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Mo
h
am
m
ad
Hass
an
✓
✓
✓
✓
✓
✓
Ash
r
af
Al
Sh
ar
ah
✓
✓
✓
✓
✓
✓
Ah
m
e
d
Qu
tais
h
at
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t
th
e
f
in
d
in
g
s
o
f
th
is
s
tu
d
y
ar
e
o
p
en
ly
av
ailab
le
in
Kag
g
le
at
h
ttp
s
://www.
k
ag
g
le.
co
m
/d
atasets
/m
s
am
b
ar
e/f
er
2
0
1
3
.
RE
F
E
R
E
NC
E
S
[
1
]
J.
P
u
st
e
j
o
v
sk
y
a
n
d
N
.
K
r
i
s
h
n
a
sw
a
m
y
,
“
Emb
o
d
i
e
d
h
u
ma
n
c
o
m
p
u
t
e
r
i
n
t
e
r
a
c
t
i
o
n
,
”
K
I
-
K
ü
n
st
l
i
c
h
e
I
n
t
e
l
l
i
g
e
n
z
,
v
o
l
.
3
5
,
n
o
.
3
–
4
,
p
p
.
3
0
7
–
3
2
7
,
N
o
v
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
0
7
/
s1
3
2
1
8
-
021
-
0
0
7
2
7
-
5.
[
2
]
Z.
L
y
u
,
“
S
t
a
t
e
-
of
-
t
h
e
-
a
r
t
h
u
ma
n
-
c
o
mp
u
t
e
r
-
i
n
t
e
r
a
c
t
i
o
n
i
n
m
e
t
a
v
e
r
se,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
H
u
m
a
n
–
C
o
m
p
u
t
e
r
I
n
t
e
r
a
c
t
i
o
n
,
v
o
l
.
4
0
,
n
o
.
2
1
,
p
p
.
6
6
9
0
–
6
7
0
8
,
N
o
v
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
8
0
/
1
0
4
4
7
3
1
8
.
2
0
2
3
.
2
2
4
8
8
3
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
F
a
cia
l e
mo
tio
n
r
ec
o
g
n
itio
n
th
r
o
u
g
h
a
r
tifi
cia
l
in
tellig
en
ce
(
A
n
a
s
Qu
teis
h
a
t
)
4709
[
3
]
Z.
Lv
,
F
.
P
o
i
e
si
,
Q
.
D
o
n
g
,
J.
L
l
o
r
e
t
,
a
n
d
H
.
S
o
n
g
,
“
D
e
e
p
l
e
a
r
n
i
n
g
f
o
r
i
n
t
e
l
l
i
g
e
n
t
h
u
m
a
n
–
c
o
m
p
u
t
e
r
i
n
t
e
r
a
c
t
i
o
n
,
”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
2
,
n
o
.
2
2
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
2
2
2
1
1
4
5
7
.
[
4
]
J.
L
i
u
,
“
C
h
a
t
G
P
T
:
p
e
r
s
p
e
c
t
i
v
e
s
f
r
o
m
h
u
ma
n
–
c
o
mp
u
t
e
r
i
n
t
e
r
a
c
t
i
o
n
a
n
d
p
s
y
c
h
o
l
o
g
y
,
”
Fr
o
n
t
i
e
rs
i
n
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
7
,
Ju
n
.
2
0
2
4
,
d
o
i
:
1
0
.
3
3
8
9
/
f
r
a
i
.
2
0
2
4
.
1
4
1
8
8
6
9
.
[
5
]
I
.
N
a
f
i
sah
e
t
a
l
.
,
“
D
e
e
p
l
e
a
r
n
i
n
g
-
b
a
sed
f
e
a
t
u
r
e
s
e
l
e
c
t
i
o
n
f
o
r
d
e
t
e
c
t
i
o
n
o
f
a
u
t
i
sm
s
p
e
c
t
r
u
m
d
i
so
r
d
e
r
,
”
Fr
o
n
t
i
e
rs
i
n
Art
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
8
,
J
u
n
.
2
0
2
5
,
d
o
i
:
1
0
.
3
3
8
9
/
f
r
a
i
.
2
0
2
5
.
1
5
9
4
3
7
2
.
[
6
]
I
.
G
o
o
d
f
e
l
l
o
w
,
Y
.
B
e
n
g
i
o
,
a
n
d
A
.
C
o
u
r
v
i
l
l
e
,
D
e
e
p
l
e
a
r
n
i
n
g
.
C
a
m
b
r
i
d
g
e
,
U
n
i
t
e
d
K
i
n
g
d
o
m:
M
I
T
P
r
e
ss,
2
0
1
6
.
[
7
]
S
.
A
l
Zu
’
b
i
e
t
a
l
.
,
“
A
n
o
v
e
l
d
e
e
p
l
e
a
r
n
i
n
g
t
e
c
h
n
i
q
u
e
f
o
r
d
e
t
e
c
t
i
n
g
e
mo
t
i
o
n
a
l
i
m
p
a
c
t
i
n
o
n
l
i
n
e
e
d
u
c
a
t
i
o
n
,
”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
1
,
n
o
.
1
8
,
S
e
p
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s
1
1
1
8
2
9
6
4
.
[
8
]
M
.
K
h
a
n
e
t
a
l
.
,
“
A
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
f
o
r
f
a
c
i
a
l
e
m
o
t
i
o
n
s
r
e
c
o
g
n
i
t
i
o
n
u
si
n
g
p
r
i
n
c
i
p
a
l
c
o
m
p
o
n
e
n
t
a
n
a
l
y
si
s a
n
d
n
e
u
r
a
l
n
e
t
w
o
r
k
t
e
c
h
n
i
q
u
e
s,
”
T
h
e
Ph
o
t
o
g
ra
m
m
e
t
r
i
c
Re
c
o
r
d
,
v
o
l
.
3
7
,
n
o
.
1
8
0
,
p
p
.
4
3
5
–
4
5
2
,
D
e
c
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
1
1
/
p
h
o
r
.
1
2
4
2
6
.
[
9
]
H
.
S
.
G
i
l
l
a
n
d
B
.
S
.
K
h
e
h
r
a
,
“
A
n
i
n
t
e
g
r
a
t
e
d
a
p
p
r
o
a
c
h
u
si
n
g
C
N
N
-
R
N
N
-
LSTM
f
o
r
c
l
a
ssi
f
i
c
a
t
i
o
n
o
f
f
r
u
i
t
i
mag
e
s,”
M
a
t
e
ri
a
l
s
T
o
d
a
y
:
Pro
c
e
e
d
i
n
g
s
,
v
o
l
.
5
1
,
p
p
.
5
9
1
–
5
9
5
,
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
ma
t
p
r
.
2
0
2
1
.
0
6
.
0
1
6
.
[
1
0
]
S
.
A
l
g
h
y
a
l
i
n
e
,
“
O
p
t
i
m
i
se
d
C
N
N
a
r
c
h
i
t
e
c
t
u
r
e
s
f
o
r
h
a
n
d
w
r
i
t
t
e
n
A
r
a
b
i
c
c
h
a
r
a
c
t
e
r
r
e
c
o
g
n
i
t
i
o
n
,
”
C
o
m
p
u
t
e
rs,
M
a
t
e
r
i
a
l
s
&
C
o
n
t
i
n
u
a
,
v
o
l
.
7
9
,
n
o
.
3
,
p
p
.
4
9
0
5
–
4
9
2
4
,
2
0
2
4
,
d
o
i
:
1
0
.
3
2
6
0
4
/
c
mc
.
2
0
2
4
.
0
5
2
0
1
6
.
[
1
1
]
Y
.
Zh
a
n
g
e
t
a
l
.
,
“
C
a
r
r
y
i
n
g
o
u
t
C
N
N
c
h
a
n
n
e
l
p
r
u
n
i
n
g
i
n
a
w
h
i
t
e
b
o
x
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
N
e
u
r
a
l
N
e
t
w
o
r
k
s
a
n
d
L
e
a
rn
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
4
,
n
o
.
1
0
,
p
p
.
7
9
4
6
–
7
9
5
5
,
O
c
t
.
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS
.
2
0
2
2
.
3
1
4
7
2
6
9
.
[
1
2
]
K
.
S
h
a
h
e
e
d
e
t
a
l
.
,
“
D
S
-
C
N
N
:
A
p
r
e
-
t
r
a
i
n
e
d
X
c
e
p
t
i
o
n
m
o
d
e
l
b
a
se
d
o
n
d
e
p
t
h
-
w
i
s
e
se
p
a
r
a
b
l
e
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
f
o
r
f
i
n
g
e
r
v
e
i
n
r
e
c
o
g
n
i
t
i
o
n
,
”
E
x
p
e
rt
S
y
st
e
m
s w
i
t
h
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
1
9
1
,
A
p
r
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
sw
a
.
2
0
2
1
.
1
1
6
2
8
8
.
[
1
3
]
T.
P
e
a
r
c
e
,
A
.
B
r
i
n
t
r
u
p
,
a
n
d
J
.
Z
h
u
,
“
U
n
d
e
r
st
a
n
d
i
n
g
s
o
f
t
ma
x
c
o
n
f
i
d
e
n
c
e
a
n
d
u
n
c
e
r
t
a
i
n
t
y
,
”
2
0
2
1
,
a
rX
i
v
:
2
1
0
6
.
0
4
9
7
2
.
[
1
4
]
P
.
K
.
S
i
m
p
s
o
n
,
“
F
u
z
z
y
m
i
n
-
m
a
x
n
e
u
r
a
l
n
e
t
w
o
r
k
s
-
p
a
r
t
1
:
c
l
a
ss
i
f
i
c
a
t
i
o
n
,”
I
EEE
T
ra
n
s
a
c
t
i
o
n
s
o
n
N
e
u
ra
l
N
e
t
w
o
rks
,
v
o
l
.
3
,
n
o
.
5
,
p
p
.
7
7
6
–
7
8
6
,
1
9
9
2
,
d
o
i
:
1
0
.
1
1
0
9
/
7
2
.
1
5
9
0
6
6
.
[
1
5
]
P
.
K
.
S
i
m
p
s
o
n
,
“
F
u
z
z
y
mi
n
-
ma
x
n
e
u
r
a
l
n
e
t
w
o
r
k
s
-
p
a
r
t
2
:
c
l
u
s
t
e
r
i
n
g
,
”
I
EEE
T
r
a
n
sa
c
t
i
o
n
s
o
n
F
u
zzy
S
y
s
t
e
m
s
,
v
o
l
.
1
,
n
o
.
1
,
F
e
b
.
1
9
9
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TFU
ZZ.
1
9
9
3
.
3
9
0
2
8
2
.
[
1
6
]
T.
R
.
C
h
a
v
a
n
a
n
d
A
.
V
.
N
a
n
d
e
d
k
a
r
,
“
A
c
o
n
v
o
l
u
t
i
o
n
a
l
f
u
z
z
y
mi
n
-
ma
x
n
e
u
r
a
l
n
e
t
w
o
r
k
,
”
N
e
u
ro
c
o
m
p
u
t
i
n
g
,
v
o
l
.
4
0
5
,
p
p
.
6
2
–
7
1
,
S
e
p
.
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
e
u
c
o
m.
2
0
2
0
.
0
4
.
0
0
3
.
[
1
7
]
W
.
H
u
a
n
g
,
M
.
S
u
n
,
L
.
Z
h
u
,
S
.
-
K
.
O
h
,
a
n
d
W
.
P
e
d
r
y
c
z
,
“
D
e
e
p
f
u
z
z
y
mi
n
–
ma
x
n
e
u
r
a
l
n
e
t
w
o
r
k
:
a
n
a
l
y
s
i
s
a
n
d
d
e
si
g
n
,
”
I
EEE
T
r
a
n
s
a
c
t
i
o
n
s
o
n
N
e
u
ra
l
N
e
t
w
o
rks
a
n
d
L
e
a
r
n
i
n
g
S
y
s
t
e
m
s
,
v
o
l
.
3
5
,
n
o
.
6
,
p
p
.
8
2
2
9
–
8
2
4
0
,
J
u
n
.
2
0
2
4
,
d
o
i
:
1
0
.
1
1
0
9
/
TN
N
LS.
2
0
2
2
.
3
2
2
6
0
4
0
.
[
1
8
]
M
.
K
a
l
b
h
o
r
,
S
.
S
h
i
n
d
e
,
D
.
E
.
P
o
p
e
s
c
u
,
a
n
d
D
.
J
.
H
e
m
a
n
t
h
,
“
H
y
b
r
i
d
i
z
a
t
i
o
n
o
f
d
e
e
p
l
e
a
r
n
i
n
g
p
r
e
-
t
r
a
i
n
e
d
mo
d
e
l
s
w
i
t
h
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
c
l
a
ss
i
f
i
e
r
s
a
n
d
f
u
z
z
y
mi
n
–
m
a
x
n
e
u
r
a
l
n
e
t
w
o
r
k
f
o
r
c
e
r
v
i
c
a
l
c
a
n
c
e
r
d
i
a
g
n
o
si
s
,
”
D
i
a
g
n
o
st
i
c
s
,
v
o
l
.
1
3
,
n
o
.
7
,
A
p
r
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
d
i
a
g
n
o
st
i
c
s
1
3
0
7
1
3
6
3
.
[
1
9
]
T.
K
a
t
t
e
n
b
o
r
n
,
J.
Le
i
t
l
o
f
f
,
F
.
S
c
h
i
e
f
e
r
,
a
n
d
S
.
H
i
n
z
,
“
R
e
v
i
e
w
o
n
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
(
C
N
N
)
i
n
v
e
g
e
t
a
t
i
o
n
r
e
mo
t
e
se
n
si
n
g
,
”
I
S
PR
S
J
o
u
rn
a
l
o
f
P
h
o
t
o
g
r
a
m
m
e
t
ry
a
n
d
Re
m
o
t
e
S
e
n
s
i
n
g
,
v
o
l
.
1
7
3
,
p
p
.
2
4
–
4
9
,
M
a
r
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
sp
r
s
j
p
r
s
.
2
0
2
0
.
1
2
.
0
1
0
.
[
2
0
]
P
.
G
i
a
n
n
o
p
o
u
l
o
s,
I
.
P
e
r
i
k
o
s,
a
n
d
I
.
H
a
t
z
i
l
y
g
e
r
o
u
d
i
s,
“
D
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
e
s
f
o
r
f
a
c
i
a
l
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
:
a
c
a
se
st
u
d
y
o
n
F
ER
-
2
0
1
3
,
”
i
n
Ad
v
a
n
c
e
s
i
n
H
y
b
r
i
d
i
z
a
t
i
o
n
o
f
I
n
t
e
l
l
i
g
e
n
t
Me
t
h
o
d
s
,
C
h
a
m,
S
w
i
t
z
e
r
l
a
n
d
:
S
p
r
i
n
g
e
r
,
2
0
1
8
,
p
p
.
1
–
16
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
319
-
6
6
7
9
0
-
4
_
1
.
[
2
1
]
T.
K
a
n
a
d
e
,
J
.
F
.
C
o
h
n
,
a
n
d
Y
.
T
i
a
n
,
“
C
o
mp
r
e
h
e
n
s
i
v
e
d
a
t
a
b
a
s
e
f
o
r
f
a
c
i
a
l
e
x
p
r
e
ssi
o
n
a
n
a
l
y
si
s
,
”
i
n
Pro
c
e
e
d
i
n
g
s
F
o
u
rt
h
I
EE
E
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
A
u
t
o
m
a
t
i
c
F
a
c
e
a
n
d
G
e
s
t
u
re
Re
c
o
g
n
i
t
i
o
n
,
2
0
0
0
,
p
p
.
4
6
–
53
,
d
o
i
:
1
0
.
1
1
0
9
/
A
F
G
R
.
2
0
0
0
.
8
4
0
6
1
1
.
[
2
2
]
M
.
R
.
H
a
ssa
n
e
t
a
l
.
,
“
I
n
t
e
g
r
a
t
i
n
g
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
s
a
n
d
d
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
t
e
c
h
n
i
q
u
e
s
f
o
r
i
m
p
r
o
v
e
d
b
r
e
a
st
c
a
n
c
e
r
d
e
t
e
c
t
i
o
n
,
”
J
o
u
rn
a
l
o
f
W
i
re
l
e
ss
M
o
b
i
l
e
N
e
t
w
o
r
k
s,
U
b
i
q
u
i
t
o
u
s
C
o
m
p
u
t
i
n
g
,
a
n
d
D
e
p
e
n
d
a
b
l
e
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
1
6
,
n
o
.
2
,
p
p
.
5
1
7
–
5
3
4
,
Ju
n
.
2
0
2
5
,
d
o
i
:
1
0
.
5
8
3
4
6
/
JO
W
U
A
.
2
0
2
5
.
I
2
.
0
3
1
.
[
2
3
]
S
.
D
u
,
Y
.
Ta
o
,
a
n
d
A
.
M
.
M
a
r
t
i
n
e
z
,
“
C
o
m
p
o
u
n
d
f
a
c
i
a
l
e
x
p
r
e
ssi
o
n
s
o
f
e
m
o
t
i
o
n
,
”
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
N
a
t
i
o
n
a
l
A
c
a
d
e
m
y
o
f
S
c
i
e
n
c
e
s
,
v
o
l
.
1
1
1
,
n
o
.
1
5
,
p
p
.
E1
4
5
4
–
E1
4
6
2
,
A
p
r
.
2
0
1
4
,
d
o
i
:
1
0
.
1
0
7
3
/
p
n
a
s.
1
3
2
2
3
5
5
1
1
1
.
[
2
4
]
S
.
R
a
m
a
l
i
n
g
a
m
a
n
d
F
.
G
a
r
z
i
a
,
“
F
a
c
i
a
l
e
x
p
r
e
ss
i
o
n
r
e
c
o
g
n
i
t
i
o
n
u
si
n
g
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
,
”
i
n
2
0
1
8
I
n
t
e
r
n
a
t
i
o
n
a
l
C
a
rn
a
h
a
n
C
o
n
f
e
re
n
c
e
o
n
S
e
c
u
ri
t
y
T
e
c
h
n
o
l
o
g
y
(
I
C
C
S
T
)
,
O
c
t
.
2
0
1
8
,
p
p
.
1
–
5
,
d
o
i
:
1
0
.
1
1
0
9
/
C
C
S
T.
2
0
1
8
.
8
5
8
5
5
0
4
.
[
2
5
]
O
.
A
r
r
i
a
g
a
,
M
.
V
.
-
T
o
r
o
,
a
n
d
P
.
G
.
P
l
ö
g
e
r
,
“
R
e
a
l
-
t
i
me
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
f
o
r
e
m
o
t
i
o
n
a
n
d
g
e
n
d
e
r
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
i
n
ES
A
N
N
2
0
1
9
-
Pr
o
c
e
e
d
i
n
g
s
,
2
7
t
h
E
u
ro
p
e
a
n
S
y
m
p
o
s
i
u
m
o
n
Art
i
f
i
c
i
a
l
N
e
u
r
a
l
N
e
t
w
o
rks
,
C
o
m
p
u
t
a
t
i
o
n
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
Ma
c
h
i
n
e
L
e
a
r
n
i
n
g
,
2
0
1
9
,
p
p
.
2
2
1
–
2
2
6
.
[
2
6
]
T.
S
h
e
n
a
n
d
H
.
X
u
,
“
F
a
c
i
a
l
e
x
p
r
e
ssi
o
n
r
e
c
o
g
n
i
t
i
o
n
b
a
s
e
d
o
n
m
u
l
t
i
-
c
h
a
n
n
e
l
a
t
t
e
n
t
i
o
n
r
e
s
i
d
u
a
l
n
e
t
w
o
r
k
,
”
C
o
m
p
u
t
e
r
M
o
d
e
l
i
n
g
i
n
En
g
i
n
e
e
ri
n
g
&
S
c
i
e
n
c
e
s
,
v
o
l
.
1
3
5
,
n
o
.
1
,
p
p
.
5
3
9
–
5
6
0
,
2
0
2
3
,
d
o
i
:
1
0
.
3
2
6
0
4
/
c
mes.
2
0
2
2
.
0
2
2
3
1
2
.
[
2
7
]
K
.
W
a
n
g
,
X
.
P
e
n
g
,
J
.
Y
a
n
g
,
S
.
L
u
,
a
n
d
Y
.
Q
i
a
o
,
“
S
u
p
p
r
e
ss
i
n
g
u
n
c
e
r
t
a
i
n
t
i
e
s
f
o
r
l
a
r
g
e
-
sc
a
l
e
f
a
c
i
a
l
e
x
p
r
e
ss
i
o
n
r
e
c
o
g
n
i
t
i
o
n
,
”
i
n
2
0
2
0
I
EEE/
C
VF
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
a
n
d
P
a
t
t
e
r
n
R
e
c
o
g
n
i
t
i
o
n
(
C
V
PR)
,
J
u
n
.
2
0
2
0
,
p
p
.
6
8
9
6
–
6
9
0
5
,
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
4
2
6
0
0
.
2
0
2
0
.
0
0
6
9
3
.
B
I
O
G
RAP
H
I
E
S O
F
AUTH
O
RS
Ana
s
Q
u
te
ish
a
t
re
c
e
iv
e
d
h
i
s
B.
S
c
.
i
n
El
e
c
tro
n
ic
E
n
g
i
n
e
e
rin
g
fro
m
P
rin
c
e
ss
S
u
m
a
y
a
Un
iv
e
rsity
f
o
r
Tec
h
n
o
l
o
g
y
,
J
o
rd
a
n
i
n
2
0
0
3
.
In
2
0
0
5
h
e
fin
ish
e
d
h
is
M
.
S
c
.
in
El
e
c
tro
n
ic
S
y
ste
m
De
sig
n
fro
m
th
e
Un
iv
e
rsiti
S
a
in
s
M
a
lay
sia
.
H
e
g
o
t
h
is
P
h
.
D.
in
2
0
0
9
i
n
Co
m
p
u
tati
o
n
a
l
In
tell
ig
e
n
c
e
fr
o
m
th
e
Un
i
v
e
rsiti
S
a
in
s
M
a
lay
sia
.
He
is
h
e
a
d
o
f
t
h
e
D
e
p
a
rtme
n
t
o
f
Co
m
m
u
n
ica
ti
o
n
a
n
d
C
o
m
p
u
ter
En
g
in
e
e
rin
g
,
Al
-
Ah
li
y
y
a
Am
m
a
n
Un
iv
e
rsit
y
.
C
u
rre
n
tl
y
,
h
e
is
a
n
a
ss
o
c
iate
p
ro
fe
ss
o
r
a
t
Al
-
Ba
lq
a
Ap
p
l
ied
Un
i
v
e
rsity
.
His
re
se
a
rc
h
in
tere
sts
in
c
lu
d
e
a
rti
ficia
l
n
e
u
ra
l
n
e
two
r
k
s,
p
a
tt
e
rn
re
c
o
g
n
it
io
n
,
in
tel
li
g
e
n
t
sy
ste
m
s,
a
n
d
e
m
b
e
d
d
e
d
s
y
ste
m
s.
He
h
a
s
p
u
b
li
s
h
e
d
m
a
n
y
re
se
a
rc
h
p
a
p
e
rs
in
i
n
tern
a
ti
o
n
a
l
jo
u
rn
a
ls.
He
c
a
n
b
e
c
o
n
tac
ted
a
t
e
m
a
il
:
a
n
a
s.q
u
teish
a
t@
b
a
u
.
e
d
u
.
j
o
.
Evaluation Warning : The document was created with Spire.PDF for Python.