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.
1
5
,
No
.
2
,
A
p
r
il 2
0
2
6
,
p
p
.
1
5
6
6
~
1
5
7
7
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
2
.
p
p
1
5
6
6
-
1
5
7
7
1566
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
M
ultimo
da
l f
a
cia
l
ex
p
ress
io
n recog
nition us
ing
resi
d
ua
l
mo
g
rifier
lo
ng
sh
o
rt
-
ter
m memo
ry
M
a
m
a
t
ha
K
a
ri
y
a
pp
a
Ra
j
a
nn
a
1
,
T
hej
a
s
wini
Sh
a
nk
a
r
1
,
Ra
s
hm
i N
a
ra
s
im
ha
m
urt
hy
1
,
Na
nd
hin
i
Va
nn
iv
edu L
a
k
s
hm
a
na
n
2
,
H
a
ripra
s
a
d S
.
Ana
ntha
pa
dm
a
na
bh
a
ra
o
3
1
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
B
M
S
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
a
n
d
M
a
n
a
g
e
me
n
t
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
G
o
v
e
r
n
m
e
n
t
S
r
i
K
r
i
s
h
n
a
r
a
j
e
n
d
r
a
S
i
l
v
e
r
J
u
b
i
l
e
e
Te
c
h
n
o
l
o
g
i
c
a
l
I
n
st
i
t
u
t
e
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
s a
n
d
C
o
m
mu
n
i
c
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
,
J
a
i
n
U
n
i
v
e
r
s
i
t
y
,
B
e
n
g
a
l
u
r
u
,
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
J
u
l 1
2
,
2
0
2
5
R
ev
is
ed
J
an
2
0
,
2
0
2
6
Acc
ep
ted
Feb
6
,
2
0
2
6
M
u
lt
imo
d
a
l
fa
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
io
n
a
ims
t
o
imp
ro
v
e
e
m
o
ti
o
n
a
n
a
l
y
sis
b
y
in
teg
ra
ti
n
g
v
isu
a
l,
a
u
d
i
o
,
a
n
d
tex
tu
a
l
c
u
e
s
t
o
a
c
h
iev
e
a
c
c
u
ra
c
y
a
n
d
ro
b
u
stn
e
ss
.
Ho
we
v
e
r,
e
ffe
c
ti
v
e
l
y
re
c
o
g
n
izin
g
fa
c
ial
e
x
p
re
ss
io
n
s
a
c
ro
ss
v
id
e
o
,
te
x
t
,
a
n
d
a
u
d
io
p
re
se
n
ts
c
h
a
ll
e
n
g
e
s
d
u
e
t
o
in
c
o
n
siste
n
c
ies
in
h
o
w
e
m
o
ti
o
n
s
a
re
e
x
p
re
ss
e
d
a
m
o
n
g
t
h
e
se
m
o
d
a
li
ti
e
s.
To
o
v
e
rc
o
m
e
th
is
issu
e
,
th
is
re
se
a
rc
h
p
ro
p
o
se
s
a
re
sid
u
a
l
m
o
g
rifi
e
r
lo
n
g
sh
o
rt
-
term
m
e
m
o
ry
(RM
LS
TM
)
m
o
d
e
l
to
e
n
h
a
n
c
e
r
o
b
u
stn
e
ss
in
m
u
lt
imo
d
a
l
fa
c
ial
e
x
p
re
ss
io
n
re
c
o
g
n
it
io
n
.
B
y
in
teg
ra
ti
n
g
re
sid
u
a
l
c
o
n
n
e
c
ti
o
n
s
in
to
th
e
lo
n
g
sh
o
r
t
-
term
m
e
m
o
ry
(
LS
TM
)
,
th
e
m
o
d
e
l
imp
r
o
v
e
s
it
s
a
b
i
li
ty
to
c
a
p
tu
re
c
o
m
p
lex
d
e
p
e
n
d
e
n
c
i
e
s
a
m
o
n
g
v
a
rio
u
s
m
o
d
a
li
ti
e
s,
i
n
c
lu
d
in
g
v
id
e
o
,
tex
t
,
a
n
d
a
u
d
i
o
.
T
h
e
re
sid
u
a
l
c
o
n
n
e
c
ti
o
n
o
v
e
rc
o
m
e
s
th
e
v
a
n
is
h
in
g
g
ra
d
ie
n
t
p
r
o
b
lem
a
n
d
e
n
su
re
s
sta
b
le
trai
n
in
g
wit
h
b
e
tt
e
r
g
ra
d
ien
t
flo
w
in
d
e
e
p
e
r
n
e
two
rk
s.
T
h
e
m
o
g
rifi
e
r
m
e
c
h
a
n
i
sm
re
fin
e
s
th
e
in
p
u
t
fe
a
t
u
re
s
d
y
n
a
m
ica
ll
y
,
e
n
h
a
n
c
i
n
g
fe
a
tu
re
i
n
tera
c
ti
o
n
a
n
d
a
li
g
n
m
e
n
t
a
c
ro
ss
m
o
d
a
li
ti
e
s.
T
h
e
RM
L
S
TM
a
c
h
iev
e
s
9
9
.
5
7
%
a
n
d
9
7
.
8
3
%
a
c
c
u
ra
c
y
o
n
th
e
S
AV
EE
a
n
d
Y
o
u
T
u
b
e
d
a
ta
se
ts,
re
sp
e
c
ti
v
e
ly
,
o
u
tp
e
rfo
rm
i
n
g
b
o
t
h
th
e
mel
-
fre
q
u
e
n
c
y
c
e
p
stra
l
c
o
e
ffici
e
n
ts
t
ime
-
d
o
m
a
in
fe
a
tu
re
with
it
e
ra
ti
v
e
d
il
a
ted
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
tw
o
rk
(M
F
CC
T
-
1
DCN
N)
a
n
d
a
tt
e
n
t
io
n
-
b
a
se
d
m
u
lt
i
-
m
o
d
a
l
p
o
p
u
larity
p
re
d
icti
o
n
m
o
d
e
l
o
f
s
h
o
rt
-
fo
rm
v
id
e
o
s (AM
P
S
).
K
ey
w
o
r
d
s
:
Featu
r
e
in
ter
ac
tio
n
Gr
ad
ien
t f
lo
w
Hu
m
an
ex
p
r
ess
io
n
s
Mu
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
R
e
s
i
d
u
a
l
m
o
g
r
i
f
i
e
r
l
o
n
g
s
h
o
r
t
-
t
e
r
m
m
e
m
o
r
y
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
:
Ma
m
ath
a
Kar
iy
ap
p
a
R
ajan
n
a
Dep
ar
tm
e
n
t
o
f
E
l
ec
t
r
o
n
ics
a
n
d
C
o
m
m
u
n
i
ca
ti
o
n
E
n
g
i
n
ee
r
i
n
g
,
B
MS
I
n
s
t
it
u
te
o
f
T
ec
h
n
o
l
o
g
y
an
d
Ma
n
a
g
em
en
t
Yela
h
an
k
a,
B
en
g
alu
r
u
,
I
n
d
ia
E
m
ail: m
am
ath
ak
r
@
b
m
s
it.in
1.
I
NT
RO
D
UCT
I
O
N
E
m
o
tio
n
r
ec
o
g
n
itio
n
aim
s
to
id
en
tify
th
e
em
o
tio
n
o
f
e
v
e
r
y
u
tter
an
ce
with
in
a
d
ialo
g
u
e,
wh
er
e
h
u
m
an
em
o
tio
n
s
ar
e
ex
p
r
ess
e
d
th
r
o
u
g
h
d
i
f
f
er
en
t
m
o
d
alities
,
s
u
ch
as
v
is
u
al,
ac
o
u
s
tic
,
a
n
d
tex
t
u
al
[
1
]
.
T
h
e
f
ield
o
f
au
to
m
atic
em
o
tio
n
r
e
co
g
n
itio
n
h
as
m
ad
e
s
ig
n
if
ican
t
p
r
o
g
r
ess
u
s
in
g
in
d
i
v
id
u
al
m
o
d
alities
;
h
o
wev
er
,
co
n
s
id
er
ab
le
ch
allen
g
es
r
em
ai
n
b
e
f
o
r
e
th
is
p
r
o
b
lem
ca
n
b
e
f
u
lly
s
o
lv
e
d
[
2
]
.
Hu
m
a
n
em
o
t
io
n
al
ex
p
r
ess
io
n
is
in
h
er
en
tly
co
m
p
lex
,
as p
eo
p
le
co
m
b
in
e
m
u
ltip
le
em
o
tio
n
al
f
ac
to
r
s
[
3
]
.
T
h
e
s
am
e
em
o
tio
n
s
ca
n
b
e
r
ep
r
esen
ted
with
d
iv
er
s
e
s
ig
n
als,
s
u
ch
as
v
o
ca
l
ex
p
r
ess
io
n
s
an
d
n
o
n
v
er
b
al
cu
es,
in
clu
d
in
g
s
p
o
k
en
wo
r
d
s
,
f
ac
ia
l
ex
p
r
ess
io
n
s
,
an
d
b
o
d
y
m
o
v
e
m
en
ts
[
4
]
.
T
h
e
c
o
m
b
in
atio
n
o
f
v
a
r
io
u
s
in
d
icato
r
s
f
r
o
m
al
l
p
o
ten
tial
em
o
tio
n
s
o
u
r
ce
s
to
b
u
ild
a
m
u
ltimo
d
al
r
ec
o
g
n
itio
n
s
y
s
tem
h
o
l
d
s
g
r
ea
t
p
r
o
m
is
e
f
o
r
im
p
r
o
v
in
g
th
e
ac
cu
r
ac
y
o
f
s
in
g
le
-
s
o
u
r
ce
r
ec
o
g
n
itio
n
[
5
]
.
E
m
o
tio
n
r
ec
o
g
n
itio
n
s
y
s
tem
s
h
av
e
p
r
ac
tical
ap
p
licatio
n
s
ac
r
o
s
s
b
eh
av
io
r
al
an
aly
s
is
,
f
o
r
en
s
ic
wo
r
k
,
an
d
h
ea
lth
ca
r
e
s
ettin
g
s
[
6
]
.
R
esear
ch
in
p
s
y
ch
o
lo
g
y
i
n
d
ica
tes
th
at
a
d
r
i
v
er
’
s
em
o
tio
n
al
s
tate
d
ir
ec
tly
af
f
ec
t
s
r
o
ad
s
af
ety
,
with
n
e
g
ativ
e
e
m
o
tio
n
s
s
u
ch
as
a
n
g
er
an
d
s
ad
n
ess
in
cr
ea
s
in
g
th
e
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
Mu
ltimo
d
a
l fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
u
s
in
g
r
esid
u
a
l m
o
g
r
ifier
…
(
Ma
ma
th
a
K
a
r
iya
p
p
a
R
a
ja
n
n
a
)
1567
lik
elih
o
o
d
o
f
ca
r
ac
cid
e
n
ts
.
As
a
r
esu
lt,
tr
an
s
p
o
r
tatio
n
s
af
et
y
s
y
s
tem
s
r
eq
u
ir
e
r
ea
l
-
tim
e
em
o
tio
n
m
o
n
ito
r
in
g
d
u
e
to
th
e
cr
itical
d
r
iv
in
g
h
az
ar
d
s
p
o
s
ed
b
y
em
o
tio
n
al
in
s
ta
b
ilit
y
[
7
]
,
[
8
]
.
Dete
ctin
g
em
o
t
io
n
s
is
ch
allen
g
in
g
b
ec
au
s
e
in
d
iv
id
u
al
p
er
ce
p
tio
n
s
an
d
th
e
d
u
r
atio
n
o
f
em
o
tio
n
al
s
tates
im
p
ac
t
h
o
w
em
o
tio
n
s
ar
e
ex
p
r
ess
ed
[
9
]
.
T
h
e
in
itial
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
tech
n
i
q
u
es
r
elied
o
n
h
an
d
cr
a
f
ted
f
ea
tu
r
es,
wh
ich
r
esu
lted
in
lim
ited
ac
cu
r
ac
y
an
d
v
u
ln
er
ab
le
s
y
s
tem
d
esig
n
[
1
0
]
.
Dee
p
l
ea
r
n
in
g
(
DL
)
n
etwo
r
k
s
,
p
ar
ticu
lar
ly
c
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NN)
an
d
r
ec
u
r
r
e
n
t
n
eu
r
al
n
etwo
r
k
s
(
R
NN)
,
h
av
e
led
to
s
u
b
s
tan
tial
ad
v
a
n
ce
m
en
ts
in
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
class
if
icatio
n
,
en
h
an
cin
g
r
ec
o
g
n
itio
n
r
esu
lts
[
1
1
]
,
[
1
2
]
.
No
v
el
n
etwo
r
k
ar
ch
itectu
r
es
with
atten
tio
n
m
ec
h
a
n
is
m
s
,
as
well
as
d
ata
au
g
m
en
tatio
n
m
et
h
o
d
s
,
h
a
v
e
c
r
ea
ted
s
o
p
h
is
ticated
ap
p
r
o
ac
h
es
to
en
h
an
ce
f
ac
ial
e
x
p
r
ess
io
n
an
a
ly
s
is
an
d
g
en
er
aliza
tio
n
.
R
esear
ch
h
as
p
r
im
a
r
ily
f
o
cu
s
ed
o
n
d
etec
tin
g
em
o
tio
n
s
in
co
n
tr
o
lled
s
ettin
g
s
in
v
o
lv
i
n
g
in
d
iv
id
u
als
[
1
3
]
.
Mu
lti
-
f
a
cial
em
o
tio
n
r
ec
o
g
n
itio
n
p
r
e
s
en
ts
a
s
ig
n
if
ican
t
ch
allen
g
e
d
u
e
to
th
e
m
u
ltip
le
f
ac
e
d
etec
tio
n
m
o
d
els th
at
im
p
ac
t r
ea
l
-
tim
e
o
p
er
atio
n
[
1
4
]
.
T
h
e
d
ev
elo
p
m
en
t
o
f
en
d
-
to
-
e
n
d
s
o
lu
tio
n
s
is
s
till
in
p
r
o
g
r
es
s
,
as
r
esear
ch
er
s
h
av
e
p
er
f
o
r
m
ed
lim
ited
wo
r
k
to
ass
ess
co
llectiv
e
em
o
tio
n
s
in
ed
u
ca
tio
n
al
s
ettin
g
s
i
n
v
o
lv
in
g
s
tu
d
en
t
g
r
o
u
p
s
.
R
es
ea
r
ch
in
d
icate
s
th
at
r
ea
l
-
wo
r
ld
m
u
lti
-
s
u
b
ject
em
o
t
io
n
r
ec
o
g
n
itio
n
r
e
q
u
ir
es
f
u
r
th
er
d
ev
elo
p
m
en
t
b
e
y
o
n
d
cu
r
r
e
n
t
lim
itatio
n
s
[
1
5
]
.
Mu
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
u
tili
ze
s
v
i
d
eo
,
au
d
io
,
an
d
tex
t
d
ata
to
en
h
a
n
ce
a
cc
u
r
ac
y
.
Ho
wev
er
,
th
e
m
ain
ch
allen
g
es
in
clu
d
e
in
co
n
s
is
ten
cies
in
em
o
tio
n
p
r
esen
tatio
n
ac
r
o
s
s
d
if
f
er
en
t
m
o
d
alities
,
m
o
d
el
in
s
tab
ilit
y
,
an
d
in
e
f
f
ec
tiv
e
f
e
atu
r
e
f
u
s
io
n
.
T
o
ad
d
r
ess
th
ese
ch
allen
g
es,
th
is
r
esear
c
h
p
r
o
p
o
s
es
a
r
esid
u
al
m
o
g
r
if
ier
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
R
ML
STM
)
th
at
im
p
r
o
v
es
cr
o
s
s
-
m
o
d
al
alig
n
m
en
t
th
r
o
u
g
h
d
y
n
am
ic
f
ea
tu
r
e
tr
an
s
f
o
r
m
atio
n
,
wh
ile
r
esid
u
al
co
n
n
ec
tio
n
s
en
s
u
r
e
s
tab
le
tr
ain
in
g
.
Fu
r
th
er
m
o
r
e,
e
x
p
er
im
en
ts
o
n
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
d
em
o
n
s
tr
ate
th
at
R
ML
ST
M
o
u
tp
er
f
o
r
m
s
ex
is
tin
g
m
eth
o
d
s
.
T
h
is
r
esear
ch
ex
am
in
es
f
ac
ial
e
x
p
r
ess
io
n
r
ec
o
g
n
itio
n
,
u
tili
zin
g
v
ar
io
u
s
DL
tech
n
iq
u
es
al
o
n
g
with
th
eir
ad
v
a
n
tag
es
an
d
lim
itatio
n
s
.
T
h
is
an
aly
s
i
s
h
el
p
s
id
en
tify
r
esear
ch
g
ap
s
to
en
ab
le
th
e
d
ev
elo
p
m
en
t
o
f
i
m
p
r
o
v
e
d
r
ec
o
g
n
itio
n
tech
n
iq
u
es,
lead
i
n
g
t
o
b
etter
p
r
ec
is
io
n
.
S
i
n
g
h
e
t
a
l
.
[
1
6
]
s
u
g
g
e
s
t
e
d
a
3
D
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
(
3
D
C
NN
)
-
c
o
n
v
o
l
u
t
i
o
n
a
l
l
o
n
g
s
h
o
r
t
-
t
e
r
m
m
e
m
o
r
y
(
C
o
n
v
L
S
T
M
)
f
r
a
m
e
w
o
r
k
f
o
r
e
m
o
t
i
o
n
a
l
v
i
d
e
o
f
a
ci
a
l
r
e
c
o
g
n
i
t
i
o
n
.
B
y
c
o
m
b
i
n
i
n
g
C
N
N
a
n
d
l
o
n
g
s
h
o
r
t
-
t
e
r
m
m
e
m
o
r
y
(
L
S
T
M
)
,
t
h
e
m
o
d
e
l
e
f
f
e
c
tiv
e
l
y
c
a
p
t
u
r
e
d
s
p
a
ti
a
l
a
n
d
t
e
m
p
o
r
a
l
d
e
p
e
n
d
e
n
c
i
es
,
r
e
s
u
l
t
i
n
g
i
n
d
y
n
a
m
i
c
e
m
o
t
i
o
n
p
a
t
t
e
r
n
r
e
c
o
g
n
i
ti
o
n
.
H
o
w
ev
e
r
,
t
h
e
3
DC
NN
-
C
o
n
v
L
S
T
M
c
o
m
b
i
n
a
ti
o
n
f
a
c
e
d
o
v
e
r
f
i
t
t
i
n
g
p
r
o
b
l
e
m
s
w
h
e
n
d
e
a
l
i
n
g
w
i
t
h
h
i
g
h
-
d
i
m
e
n
s
i
o
n
al
s
p
a
t
i
o
te
m
p
o
r
a
l
d
a
t
a
.
S
i
n
g
h
e
t
a
l
.
[
1
7
]
d
ev
elo
p
ed
a
n
atten
tio
n
-
b
ased
2
DC
NN
wit
h
L
STM
to
p
er
f
o
r
m
s
p
ee
ch
e
m
o
tio
n
r
ec
o
g
n
itio
n
task
s
.
T
h
is
ar
ch
itectu
r
e
in
teg
r
ated
f
o
u
r
b
lo
ck
s
o
f
2
DC
NN
f
ea
tu
r
e
ex
tr
ac
to
r
s
with
2
DC
NN
-
L
STM
d
ep
en
d
e
n
cy
lear
n
er
s
an
d
in
co
r
p
o
r
ate
d
an
atten
tio
n
m
ec
h
an
is
m
to
f
ilter
L
STM
-
g
en
er
ated
s
ig
n
if
ican
t
d
ata,
f
o
llo
wed
b
y
a
d
r
o
p
o
u
t
o
p
er
atio
n
to
en
h
an
ce
em
o
tio
n
r
ec
o
g
n
itio
n
.
T
h
e
p
r
o
p
o
s
ed
m
et
h
o
d
e
x
h
ib
ited
s
en
s
itiv
ity
to
v
ar
io
u
s
s
p
ee
ch
p
atter
n
s
an
d
n
o
is
e
ex
p
o
s
u
r
e,
as th
e
m
o
d
el
f
ailed
to
p
r
o
p
er
ly
id
en
tify
ess
en
tial f
ea
tu
r
es.
Mid
d
y
a
et
a
l.
[
1
8
]
in
tr
o
d
u
ce
d
a
DL
-
b
ased
m
u
ltimo
d
a
l
em
o
tio
n
r
ec
o
g
n
itio
n
s
y
s
tem
u
s
in
g
au
d
io
-
v
is
u
al
m
o
d
alities
an
d
m
o
d
el
-
lev
el
f
u
s
io
n
.
Mo
d
el
-
le
v
el
f
u
s
io
n
was
p
er
f
o
r
m
e
d
to
estab
lis
h
th
e
b
es
t
m
u
ltimo
d
al
m
o
d
el
f
o
r
em
o
ti
o
n
r
ec
o
g
n
itio
n
b
y
c
o
m
b
in
i
n
g
au
d
io
an
d
v
id
eo
m
o
d
ality
d
ata.
T
h
e
m
o
d
el
d
ev
elo
p
e
d
an
o
p
tim
al
m
u
ltim
o
d
al
em
o
tio
n
r
ec
o
g
n
itio
n
s
y
s
tem
th
r
o
u
g
h
th
e
co
m
b
in
atio
n
o
f
au
d
io
an
d
v
id
eo
f
ea
tu
r
es
at
th
e
m
o
d
el
lev
el.
Ho
wev
er
,
wh
e
n
m
o
d
el
-
lev
el
f
u
s
io
n
c
o
m
b
in
e
d
s
ep
ar
ate
f
ea
tu
r
e
ex
tr
ac
t
o
r
s
,
th
e
ap
p
r
o
ac
h
b
ec
a
m
e
p
r
o
n
e
to
o
v
er
f
itti
n
g
wh
en
wo
r
k
in
g
with
s
m
all
d
atasets
with
lim
ited
em
o
tio
n
al
ex
p
r
ess
io
n
d
iv
er
s
ity
.
Allu
h
aid
a
n
et
a
l.
[
1
9
]
in
tr
o
d
u
ce
d
an
m
el
-
f
r
e
q
u
en
cy
ce
p
s
tr
al
co
ef
f
icien
ts
t
im
e
-
d
o
m
ain
f
ea
tu
r
e
with
iter
ativ
e
d
ilated
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
MFC
C
T
-
1
DC
NN)
to
r
ec
o
g
n
ize
s
p
ee
ch
ex
p
r
ess
io
n
s
.
T
h
e
C
NN
f
r
am
ewo
r
k
co
n
tain
ed
o
n
e
-
d
im
en
s
io
n
al
lay
er
s
al
o
n
g
with
ac
tiv
atio
n
lay
er
s
,
m
ax
-
p
o
o
lin
g
,
d
r
o
p
o
u
t
f
ea
t
u
r
es,
an
d
f
u
lly
co
n
n
ec
ted
(
FC
)
co
m
p
o
n
e
n
ts
f
o
r
s
p
ee
c
h
ex
p
r
ess
io
n
class
if
icatio
n
.
Ho
wev
er
,
th
er
e
was
s
p
ec
tr
al
in
f
o
r
m
atio
n
lo
s
s
in
MFC
C
T
-
1
DC
NN
b
ec
au
s
e
it
p
er
f
o
r
m
e
d
a
co
n
v
er
s
io
n
f
r
o
m
t
h
e
f
r
eq
u
en
cy
d
o
m
ain
to
t
h
e
tim
e
d
o
m
ain
.
C
h
o
et
a
l
.
[
2
0
]
p
r
esen
ted
an
AM
PS
f
o
r
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
iti
o
n
.
I
t
in
c
o
r
p
o
r
ated
b
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
B
iLST
M
)
with
s
elf
-
atten
tio
n
an
d
co
-
atten
tio
n
m
et
h
o
d
s
,
wh
ich
allo
wed
it
to
ac
h
iev
e
a
b
etter
u
n
d
er
s
ta
n
d
in
g
o
f
in
tr
a
-
an
d
in
te
r
-
m
o
d
al
r
elatio
n
s
h
ip
s
.
T
h
e
m
o
d
el
s
tr
u
g
g
led
to
ca
p
tu
r
e
f
in
e
d
etails
in
tem
p
o
r
al
co
n
n
ec
tio
n
s
b
etwe
en
au
d
io
an
d
v
i
s
u
al
co
m
p
o
n
en
ts
d
u
e
to
th
e
m
ix
ed
u
tili
za
tio
n
o
f
r
ec
u
r
r
en
t a
n
d
co
n
v
o
lu
tio
n
al
la
y
er
s
.
T
h
e
ex
is
tin
g
tech
n
iq
u
es su
f
f
er
ed
f
r
o
m
m
u
ltip
le
lim
itatio
n
s
,
in
clu
d
in
g
p
o
o
r
ca
p
ab
ilit
ies
to
d
etec
t
f
in
e
tem
p
o
r
al
co
n
n
ec
tio
n
s
b
etwe
en
au
d
io
an
d
v
is
u
al
m
o
d
alities
,
n
o
is
e
s
en
s
i
tiv
ity
an
d
d
iv
er
s
e
s
p
ee
ch
p
atter
n
s
b
ec
au
s
e
th
e
m
o
d
el
f
ailed
to
d
i
s
ce
r
n
p
r
o
p
e
r
f
ea
tu
r
es
wh
ile
h
an
d
lin
g
m
u
ltip
le
m
o
d
alities
,
th
er
eb
y
r
esu
ltin
g
i
n
co
m
p
licated
in
te
g
r
atio
n
ch
al
len
g
es.
T
o
o
v
er
co
m
e
th
ese
is
s
u
es,
th
e
R
ML
STM
is
p
r
o
p
o
s
ed
in
th
is
r
esear
ch
f
o
r
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
ar
e
as
f
o
llo
ws:
i)
T
h
e
R
ML
STM
is
p
r
o
p
o
s
ed
in
th
is
r
esear
ch
f
o
r
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
.
B
y
in
teg
r
atin
g
r
esid
u
al
co
n
n
ec
tio
n
s
in
to
L
ST
M,
th
e
m
o
d
el
ef
f
ec
tiv
ely
ca
p
tu
r
es
co
m
p
lex
d
e
p
en
d
e
n
cies
am
o
n
g
v
a
r
io
u
s
m
o
d
alities
s
u
ch
as v
id
eo
,
tex
t,
an
d
au
d
io
.
ii)
T
h
e
r
esid
u
al
co
n
n
ec
tio
n
p
r
o
v
id
es
s
tab
le
tr
ain
in
g
an
d
b
etter
g
r
ad
ien
t
f
lo
w
in
d
ee
p
e
r
n
e
two
r
k
s
wh
ile
ad
d
r
ess
in
g
th
e
v
an
is
h
in
g
g
r
ad
ien
t
is
s
u
e.
T
h
e
m
o
g
r
if
ie
r
m
ec
h
an
is
m
t
r
an
s
f
o
r
m
s
i
n
p
u
t
f
ea
tu
r
es
d
y
n
am
ically
,
t
h
er
eb
y
en
h
a
n
ci
n
g
f
ea
tu
r
e
in
ter
ac
tio
n
an
d
alig
n
m
en
t a
cr
o
s
s
m
o
d
alities
.
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.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
5
6
6
-
1
5
7
7
1568
iii)
E
f
f
icien
tNetB
0
is
u
s
ed
as a
n
ex
tr
ac
tiv
e
f
ea
tu
r
e
m
ec
h
an
is
m
f
o
r
im
ag
e/v
id
e
o
s
am
p
les th
r
o
u
g
h
its
ef
f
ec
tiv
e
v
is
u
al
p
atter
n
r
ec
o
g
n
itio
n
ca
p
ab
ilit
ies.
iv
)
T
ex
t
d
ata
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
ter
m
f
r
eq
u
en
cy
-
i
n
v
er
s
e
d
o
c
u
m
en
t
f
r
eq
u
en
cy
(
T
FID
F),
wh
ich
tr
an
s
f
o
r
m
s
v
er
b
al
in
f
o
r
m
atio
n
in
to
n
u
m
er
ical
v
ec
to
r
s
b
y
ass
ig
n
in
g
s
ig
n
if
ican
ce
to
im
p
o
r
tan
t
wo
r
d
s
with
o
u
t
r
ed
u
cin
g
th
e
s
tan
d
ar
d
ter
m
s
.
v)
MFC
C
ar
e
u
s
ed
to
ex
tr
ac
t
au
d
io
f
ea
tu
r
es
th
r
o
u
g
h
s
p
ec
tr
al
an
d
f
r
e
q
u
en
c
y
-
b
ased
ch
ar
ac
ter
is
tics
,
p
r
o
v
id
i
n
g
b
etter
au
d
io
s
ig
n
al
r
ep
r
esen
tatio
n
s
.
T
h
is
r
esear
ch
p
ap
er
is
f
u
r
t
h
er
o
r
g
a
n
ized
as
f
o
llo
ws
.
S
ec
tio
n
2
d
etails
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
.
S
ec
tio
n
3
p
r
esen
ts
th
e
r
esu
lts
an
d
d
is
cu
s
s
io
n
. T
h
e
c
o
n
clu
s
io
n
o
f
th
is
s
tu
d
y
is
p
r
o
v
i
d
ed
in
s
ec
tio
n
4
.
2.
P
RO
P
O
SE
D
M
E
T
H
O
DO
L
O
G
Y
I
n
th
is
r
esear
ch
,
th
e
R
ML
STM
is
p
r
o
p
o
s
ed
f
o
r
m
u
ltimo
d
a
l
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
u
tili
zin
g
d
ata
f
r
o
m
t
h
e
SAVE
E
a
n
d
Yo
u
T
u
b
e
s
o
u
r
ce
s
,
wh
ich
in
c
lu
d
e
v
i
d
eo
/im
ag
es,
tex
t,
a
n
d
au
d
io
.
T
o
p
r
o
v
i
d
e
h
ig
h
-
q
u
ality
in
p
u
ts
,
p
r
ep
r
o
ce
s
s
in
g
tech
n
iq
u
es
ar
e
ap
p
lied
,
s
u
ch
as
im
ag
e
n
o
r
m
aliza
tio
n
an
d
p
u
n
ctu
atio
n
r
em
o
v
al.
I
m
ag
e
n
o
r
m
aliza
tio
n
is
ap
p
lied
to
s
tan
d
a
r
d
ize
p
ix
e
l
in
ten
s
ities
,
p
u
n
ctu
atio
n
r
em
o
v
al
is
u
s
ed
f
o
r
tex
t
clea
n
in
g
,
an
d
MFC
C
f
ea
tu
r
es
ar
e
d
ir
ec
tly
ex
tr
ac
te
d
f
r
o
m
au
d
io
d
ata
to
ca
p
tu
r
e
k
e
y
f
r
e
q
u
e
n
cy
ch
ar
ac
ter
is
tics
.
T
h
e
f
ea
tu
r
e
ex
t
r
ac
tio
n
s
tep
in
clu
d
es
E
f
f
icie
n
tNetB
0
f
o
r
v
is
u
al
d
ata,
T
FID
F
f
o
r
tex
t
u
a
l
in
f
o
r
m
atio
n
,
an
d
MFC
C
f
o
r
s
p
ee
ch
s
ig
n
als.
T
h
e
ex
tr
ac
ted
f
ea
tu
r
es
f
r
o
m
all
m
o
d
alities
ar
e
f
u
s
ed
to
cr
ea
te
a
u
n
if
ied
m
u
ltim
o
d
al
r
ep
r
esen
tatio
n
.
T
h
en
,
th
is
f
u
s
ed
d
ata
is
f
e
d
in
t
o
th
e
R
ML
STM
m
o
d
el,
wh
ic
h
u
s
es
r
esi
d
u
al
co
n
n
ec
tio
n
s
to
o
v
er
co
m
e
th
e
v
an
is
h
in
g
g
r
ad
i
en
t
is
s
u
e
an
d
p
r
o
v
id
e
m
o
d
el
s
tab
ilit
y
,
wh
ile
th
e
m
o
g
r
if
ier
a
p
p
r
o
ac
h
tr
a
n
s
f
o
r
m
s
in
p
u
t
f
ea
tu
r
es
d
y
n
am
ically
,
th
er
eb
y
en
h
a
n
cin
g
f
ea
t
u
r
e
alig
n
m
en
t
an
d
in
ter
ac
tio
n
a
m
o
n
g
v
ar
io
u
s
m
o
d
alities
.
L
astl
y
,
th
e
m
o
d
el
class
if
ies
f
ac
ial
ex
p
r
ess
io
n
s
b
ased
o
n
m
u
ltimo
d
al
in
p
u
ts
.
Fig
u
r
e
1
s
h
o
w
s
th
e
f
lo
w
d
iag
r
am
o
f
th
is
r
esear
ch
.
Fig
u
r
e
1
.
Flo
w
d
ia
g
r
am
o
f
th
e
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
2
.
1
.
Da
t
a
s
et
T
h
e
S
A
V
E
E
[
2
1
]
,
Y
o
u
T
u
b
e
[
7
]
,
C
K
+
[
1
6
]
,
a
n
d
R
A
V
D
E
E
S
[
1
8
]
d
a
t
a
s
e
t
s
a
r
e
u
s
e
d
t
o
c
o
l
l
e
c
t
d
a
t
a
f
o
r
t
h
i
s
r
e
s
e
a
r
c
h
.
T
h
e
s
e
d
a
t
a
s
e
t
s
a
r
e
a
p
p
l
i
e
d
t
o
c
r
e
a
t
e
a
n
a
u
t
o
m
a
t
e
d
e
m
o
t
i
o
n
i
d
e
n
t
i
f
i
c
a
t
i
o
n
s
y
s
t
e
m
t
h
a
t
s
u
p
p
o
r
t
s
v
i
d
e
o
,
i
m
a
g
e
,
t
e
x
t
,
a
n
d
a
u
d
i
o
.
T
h
e
S
A
V
E
E
d
a
t
a
s
e
t
i
n
c
l
u
d
e
s
4
8
0
w
o
r
d
s
,
1
0
g
e
n
e
r
a
l
s
e
n
t
e
n
c
e
s
,
3
s
t
a
n
d
a
r
d
s
e
n
t
e
n
c
e
s
,
a
n
d
2
e
m
o
t
i
o
n
-
s
p
e
c
i
f
i
c
s
e
n
t
e
n
c
e
s
.
T
h
e
Y
o
u
T
u
b
e
d
a
t
a
s
e
t
i
n
c
l
u
d
e
s
s
e
v
e
n
d
i
s
t
i
n
c
t
e
m
o
t
i
o
n
s
:
a
n
g
e
r
,
d
i
s
g
u
s
t
,
f
e
a
r
,
h
a
p
p
i
n
e
s
s
,
n
e
u
t
r
a
l
,
s
a
d
n
e
s
s
,
a
n
d
s
u
r
p
r
i
s
e
,
a
l
l
o
f
w
h
i
c
h
a
r
e
r
e
c
o
r
d
e
d
a
n
d
e
v
a
l
u
a
t
e
d
b
y
1
0
i
n
d
i
v
i
d
u
a
l
s
.
T
h
e
C
K
+
d
a
t
a
s
e
t
c
o
n
t
a
i
n
s
5
9
3
s
e
q
u
e
n
c
e
s
f
r
o
m
1
2
3
s
u
b
j
e
c
t
s
,
o
f
w
h
i
c
h
3
2
7
s
e
q
u
e
n
c
e
s
a
r
e
l
a
b
e
l
e
d
w
i
t
h
e
m
o
t
i
o
n
s
.
S
e
v
e
n
e
m
o
t
i
o
n
s
a
r
e
l
a
b
e
l
e
d
,
i
n
c
l
u
d
i
n
g
s
u
r
p
r
i
s
e
,
f
e
a
r
,
d
i
s
g
u
s
t
,
c
o
n
t
e
m
p
t
,
s
a
d
n
e
s
s
,
h
a
p
p
i
n
e
s
s
,
a
n
d
a
n
g
e
r
.
T
h
e
R
A
V
D
E
E
S
d
a
t
a
s
e
t
c
o
n
t
a
i
n
s
1
,
4
4
0
a
u
d
i
o
f
i
l
e
s
r
e
c
o
r
d
e
d
b
y
1
2
f
e
m
a
l
e
a
n
d
1
2
m
a
l
e
a
c
t
o
r
s
,
c
o
v
e
r
i
n
g
e
i
g
h
t
d
i
s
t
i
n
c
t
e
m
o
t
i
o
n
s
:
a
n
g
e
r
,
c
a
l
m
n
e
s
s
,
d
i
s
g
u
s
t
,
f
e
a
r
,
n
e
u
t
r
a
l
i
t
y
,
h
a
p
p
i
n
e
s
s
,
s
a
d
n
e
s
s
,
a
n
d
s
u
r
p
r
i
s
e
.
A
l
l
r
e
c
o
r
d
e
d
a
u
d
i
o
f
i
l
e
s
h
a
v
e
a
4
8
k
H
z
s
a
m
p
l
e
r
a
t
e
a
n
d
16
-
b
i
t
r
e
s
o
l
u
t
i
o
n
.
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
Mu
ltimo
d
a
l fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
u
s
in
g
r
esid
u
a
l m
o
g
r
ifier
…
(
Ma
ma
th
a
K
a
r
iya
p
p
a
R
a
ja
n
n
a
)
1569
2
.
2
.
P
re
pro
ce
s
s
ing
Af
ter
d
ata
co
llectio
n
,
n
o
r
m
aliza
tio
n
,
an
d
p
u
n
ctu
atio
n
r
em
o
v
al
ar
e
ap
p
lied
to
p
r
e
p
r
o
ce
s
s
th
e
d
ata.
No
r
m
aliza
tio
n
is
u
s
ed
f
o
r
im
a
g
e/v
id
eo
d
ata,
wh
ile
p
u
n
ctu
ati
o
n
r
em
o
v
al
is
ap
p
lied
to
tex
t
d
ata.
No
r
m
aliza
tio
n
is
a
p
r
e
p
r
o
ce
s
s
in
g
m
eth
o
d
ap
p
lied
to
r
ed
u
ce
t
h
e
d
i
f
f
er
en
ce
s
in
f
ac
e
im
ag
es,
s
u
ch
as
v
ar
i
atio
n
s
in
lig
h
tin
g
,
t
o
ac
h
iev
e
b
etter
im
ag
e
q
u
ality
.
I
t e
n
h
an
ce
s
im
ag
e
in
ten
s
ity
,
r
esu
ltin
g
in
im
p
r
o
v
e
d
clar
ity
an
d
h
ig
h
er
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
.
T
h
e
m
ath
em
atic
al
ex
p
r
ess
io
n
f
o
r
th
is
n
o
r
m
aliza
tio
n
is
p
r
esen
ted
in
(
1
)
.
W
h
er
e
d
en
o
tes
th
e
n
o
r
m
alize
d
im
a
g
e,
an
d
d
en
o
t
e
m
in
im
u
m
a
n
d
m
a
x
im
u
m
im
ag
e
in
ten
s
ities
.
=
−
−
(
1
)
Pu
n
ctu
atio
n
r
em
o
v
al
h
elp
s
t
o
s
tan
d
ar
d
ize
th
e
tex
tu
al
d
ata
b
y
elim
in
atin
g
u
n
n
ec
ess
ar
y
s
y
m
b
o
ls
th
at
d
o
n
o
t
co
n
tr
ib
u
te
to
s
em
an
tic
m
ea
n
in
g
,
th
u
s
en
h
an
cin
g
m
o
d
el
ef
f
icien
cy
.
B
y
r
em
o
v
in
g
p
u
n
ctu
atio
n
,
n
o
is
e
is
r
ed
u
ce
d
,
lead
in
g
to
b
etter
t
o
k
en
izatio
n
a
n
d
f
ea
tu
r
e
ex
tr
ac
tio
n
.
T
h
e
f
o
r
m
u
la
f
o
r
r
em
o
v
in
g
p
u
n
ctu
atio
n
is
p
r
esen
ted
in
(
2
)
.
W
h
er
e
d
e
n
o
tes
th
e
s
et
o
f
p
u
n
ctu
atio
n
m
a
r
k
s
an
d
d
en
o
tes
th
e
to
k
en
ized
tex
t
af
ter
p
u
n
ct
u
atio
n
r
e
m
o
v
al.
Af
ter
p
er
f
o
r
m
in
g
t
h
e
p
r
e
p
r
o
ce
s
s
in
g
,
f
ea
tu
r
e
ex
tr
ac
tio
n
is
ap
p
lied
to
im
ag
e/v
id
e
o
,
tex
t
,
an
d
a
u
d
io
d
ata
to
ex
tr
ac
t
m
ea
n
in
g
f
u
l a
n
d
r
elev
an
t in
f
o
r
m
atio
n
f
r
o
m
ea
ch
o
f
th
ese
m
o
d
alities
.
=
{
∈
:
∉
}
(
2
)
2
.
3
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
E
f
f
icien
tNetB
0
is
u
s
ed
as
an
ex
tr
ac
tiv
e
f
ea
tu
r
e
m
ec
h
a
n
is
m
f
o
r
im
a
g
e/v
id
eo
s
am
p
les
th
r
o
u
g
h
its
ef
f
icien
t
co
m
p
u
tatio
n
o
f
v
is
u
al
p
atter
n
s
.
T
ex
t
d
ata
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
T
FID
F,
wh
ich
tr
an
s
f
o
r
m
s
v
er
b
al
in
f
o
r
m
atio
n
in
to
n
u
m
er
ical
v
ec
to
r
s
b
y
em
p
h
asizin
g
s
ig
n
if
i
ca
n
t
wo
r
d
s
with
o
u
t
d
im
in
is
h
i
n
g
s
tan
d
ar
d
ter
m
s
.
MFC
C
an
aly
ze
s
au
d
io
s
ig
n
als
th
r
o
u
g
h
s
p
ec
tr
al
a
n
d
f
r
eq
u
en
cy
-
b
ased
ch
a
r
ac
ter
is
tics
to
cr
e
ate
ef
f
ec
tiv
e
au
d
io
s
ig
n
al
r
ep
r
esen
tatio
n
s
.
A
d
etai
led
d
escr
ip
tio
n
o
f
th
ese
m
eth
o
d
s
is
p
r
o
v
id
ed
in
th
e
s
u
b
s
eq
u
e
n
t sectio
n
s
.
2
.
3
.
1
.
E
f
f
icient
Net
B
0
f
o
r
ima
g
e/v
ideo
da
t
a
E
f
f
icien
tNet
is
a
s
ca
led
-
u
p
m
o
d
el
th
at
o
p
tim
izes
b
o
th
ac
c
u
r
ac
y
an
d
ef
f
icien
c
y
.
E
f
f
icien
t
NetB0
is
a
k
ey
lay
er
in
t
h
e
m
o
b
ile
in
v
e
r
ted
b
o
ttlen
ec
k
MBC
o
n
v
,
with
co
m
p
o
u
n
d
s
ca
lin
g
ap
p
lie
d
to
th
r
ee
co
m
p
o
n
e
n
ts
:
d
ep
th
,
wid
th
,
an
d
r
eso
lu
tio
n
.
T
h
e
co
n
s
tan
ts
,
,
an
d
d
en
o
te
d
ep
th
,
wid
th
an
d
r
eso
lu
tio
n
,
r
esp
ec
tiv
ely
,
an
d
ar
e
d
er
iv
e
d
f
o
r
b
etter
r
esu
lts
[
2
2
]
.
E
f
f
icien
tNetB
0
s
ea
r
ch
es
f
o
r
th
ese
s
ca
lin
g
co
e
f
f
icien
t
s
with
less
s
y
s
te
m
ca
p
ac
ity
in
s
m
aller
m
o
d
els.
2
.
3
.
2
.
T
er
m
f
re
qu
ency
-
inv
er
s
e
do
cum
ent
f
re
qu
ency
f
o
r
t
ex
t
da
t
a
T
FID
F is
ap
p
lied
to
ex
tr
ac
t f
ea
tu
r
es f
r
o
m
r
aw
tex
t d
ata,
wh
e
r
e
weig
h
ts
ar
e
ass
ig
n
ed
to
ea
ch
ter
m
in
a
d
o
cu
m
e
n
t
b
ased
o
n
ter
m
f
r
e
q
u
en
cy
(
T
F)
an
d
in
v
e
r
s
e
d
o
c
u
m
en
t
f
r
e
q
u
en
c
y
(
I
DF)
.
Hig
h
e
r
-
weig
h
t
ter
m
s
h
av
e
g
r
ea
ter
s
ig
n
if
ican
ce
c
o
m
p
a
r
ed
to
lo
wer
-
weig
h
t te
r
m
s
[
2
3
]
.
T
h
e
weig
h
t f
o
r
ea
ch
ter
m
is
ca
lcu
lated
u
s
in
g
(
3
)
.
,
=
,
(
)
(
3
)
W
h
er
e
,
is
th
e
n
u
m
b
e
r
o
f
o
cc
u
r
r
en
ce
s
o
f
ter
m
in
d
o
c
u
m
en
t
,
is
th
e
to
tal
n
u
m
b
er
o
f
d
o
c
u
m
en
ts
an
d
d
en
o
tes
th
e
d
o
cu
m
en
ts
with
ter
m
.
T
FID
F
is
a
ty
p
e
o
f
s
co
r
in
g
m
ea
s
u
r
e
m
en
t
ap
p
r
o
ac
h
wid
ely
u
s
ed
in
s
u
m
m
ar
izatio
n
an
d
d
ata
r
etr
i
ev
al.
T
F
esti
m
ates
th
e
f
r
eq
u
en
cy
o
f
to
k
e
n
an
d
g
iv
es
m
o
r
e
s
ig
n
if
ican
ce
to
co
m
m
o
n
t
o
k
en
s
in
a
g
iv
en
d
o
cu
m
en
t.
Ho
we
v
er
,
I
DF
esti
m
ates
th
e
r
ar
ity
o
f
to
k
en
s
in
th
e
co
r
p
u
s
.
I
n
th
is
m
an
n
er
,
if
u
n
co
m
m
o
n
w
o
r
d
s
ap
p
ea
r
in
m
o
r
e
th
an
o
n
e
d
o
c
u
m
en
t,
th
ey
ar
e
c
o
n
s
id
er
ed
s
ig
n
if
ican
t.
I
n
a
g
r
o
u
p
o
f
d
o
c
u
m
en
ts
,
th
e
I
DF
weig
h
ts
a
to
k
en
u
s
in
g
(
4
)
.
W
h
er
e,
(
)
is
th
e
f
r
eq
u
e
n
cy
o
f
in
an
d
(
)
is
th
e
in
v
er
s
e
f
r
eq
u
en
cy
.
T
h
e
T
F
-
I
DF is estima
ted
b
y
co
m
b
in
in
g
T
F a
n
d
I
DF,
as r
ep
r
esen
ted
in
(
5
)
.
(
)
=
(
)
(
4
)
−
=
×
(
5
)
T
FID
F
is
ap
p
lied
to
esti
m
ate
th
e
s
ig
n
if
ican
t
ter
m
weig
h
ts
,
an
d
th
e
f
in
al
o
u
t
p
u
t
o
f
T
FID
F
is
in
th
e
f
o
r
m
o
f
a
weig
h
t
m
atr
ix
.
T
h
e
s
co
r
es
in
cr
ea
s
e
g
r
ad
u
ally
with
th
e
T
FID
F
co
u
n
t,
b
u
t
a
r
e
b
alan
ce
d
b
y
t
h
e
f
r
eq
u
e
n
cy
o
f
th
e
w
o
r
d
.
I
t
tr
an
s
f
o
r
m
s
v
ar
iab
le
-
le
n
g
th
tex
t
in
to
f
ea
tu
r
e
v
ec
to
r
o
f
f
ix
e
d
-
len
g
t
h
th
er
eb
y
p
r
o
v
id
i
n
g
s
im
p
ler
in
teg
r
atio
n
with
r
ec
o
g
n
itio
n
m
o
d
els.
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.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
5
6
6
-
1
5
7
7
1570
2
.
3
.
3
.
M
F
CC
f
o
r
a
ud
io
da
t
a
T
h
e
ac
o
u
s
tic
f
ea
tu
r
es
o
f
an
au
d
io
s
ig
n
al
r
ep
r
esen
t
th
e
p
h
y
s
ical
p
r
o
p
er
ties
o
f
s
p
ee
ch
in
ter
m
s
o
f
am
p
litu
d
e,
f
r
eq
u
en
cy
,
an
d
lo
u
d
n
ess
.
T
h
e
ac
o
u
s
tic
f
ea
tu
r
e
s
e
t
in
clu
d
es
d
is
tin
ct
s
p
ec
tr
al
f
ea
tu
r
es,
v
o
ice
q
u
ality
f
ea
tu
r
es,
an
d
tim
e
-
d
o
m
ain
f
e
atu
r
es
to
d
escr
ib
e
f
ac
ial
ex
p
r
ess
io
n
s
.
MFC
C
p
r
o
v
id
es
s
p
e
ctr
al
in
f
o
r
m
atio
n
o
f
s
p
ee
ch
an
d
m
o
d
els
h
u
m
a
n
au
d
ito
r
y
p
er
ce
p
tio
n
.
I
n
MFC
C
,
th
e
ce
p
s
tr
u
m
is
o
b
tain
ed
b
y
ap
p
ly
in
g
a
d
is
cr
ete
co
s
in
e
tr
an
s
f
o
r
m
to
th
e
lo
g
a
r
ith
m
o
f
th
e
s
h
o
r
t
-
tim
e
p
o
wer
s
p
ec
tr
u
m
o
f
t
h
e
s
ig
n
al.
T
h
e
co
ef
f
icien
t
o
f
th
e
m
el
-
s
ca
le
s
p
ec
tr
u
m
is
an
im
p
r
o
v
ed
tec
h
n
iq
u
e
ad
a
p
ted
f
r
o
m
t
h
e
ce
p
s
tr
u
m
.
T
h
e
m
el
-
s
ca
le
co
n
s
is
ts
o
f
a
u
n
i
f
o
r
m
s
p
ac
e
o
f
tr
ian
g
u
lar
f
ilter
b
an
k
s
[
2
4
]
.
Hen
ce
,
th
e
b
an
d
wid
th
o
f
in
d
iv
id
u
al
f
ilter
s
en
h
a
n
ce
s
th
e
lo
g
ar
ith
m
ic
s
ca
le
an
d
n
o
r
m
alize
s
th
e
f
r
eq
u
e
n
cy
.
C
o
n
s
id
er
th
e
m
ag
n
itu
d
e
r
esp
o
n
s
e
o
f
a
s
eq
u
en
c
e
o
f
th
e
tr
ian
g
u
lar
f
ilter
,
g
iv
e
n
as
m
el
f
r
eq
u
en
cy
.
T
h
e
m
el
-
s
ca
le
f
o
r
m
is
ex
p
r
es
s
ed
in
(
6
)
.
W
h
er
e,
(
)
d
en
o
tes
th
e
m
el
f
r
eq
u
e
n
cy
s
ca
le
an
d
d
en
o
tes
th
e
ac
tu
al
f
r
eq
u
en
c
y
.
T
h
e
m
el
f
r
eq
u
en
cy
co
n
v
er
ts
th
e
ac
tu
al
f
r
eq
u
en
c
y
o
f
p
o
wer
s
p
ec
tr
u
m
m
ag
n
itu
d
es
o
f
th
e
in
p
u
t sp
ee
ch
s
ig
n
al
t
h
r
o
u
g
h
a
m
el
-
s
ca
le
f
ilter
b
an
k
,
as g
i
v
en
in
(
7
)
.
(
)
=
2595
10
(
1
+
700
)
(
6
)
(
,
)
=
(
(
,
)
)
(
7
)
W
h
er
e
(
,
)
r
ep
r
es
en
ts
th
e
m
e
l
f
r
e
q
u
en
cy
co
ef
f
ic
ien
ts
an
d
i
s
a
s
eq
u
en
c
e
o
f
co
ef
f
ic
ie
n
t
s
o
b
t
ai
n
ed
th
r
o
u
g
h
th
e
m
el
-
s
ca
le
f
il
te
r
b
an
k
.
M
F
C
C
r
ep
r
e
s
en
t
s
th
e
t
o
n
e
a
s
p
ec
t
s
o
f
s
p
ee
ch
th
at
ar
e
s
ig
n
i
f
ic
an
t
f
o
r
f
e
te
ct
io
n
f
ac
ia
l
em
o
t
io
n
s
b
y
p
i
tch
,
s
tr
es
s
p
at
te
r
n
s
,
an
d
v
o
c
al
in
to
n
at
io
n
.
Fu
r
th
er
m
o
r
e
,
i
t
r
ed
u
c
e
s
h
ig
h
-
d
im
en
s
io
n
a
l
au
d
io
in
to
lo
w
-
d
im
en
s
io
n
al
f
e
atu
r
e
s
et
wh
i
le
p
r
e
s
er
v
i
n
g
d
is
cr
im
in
at
iv
e
f
ea
tu
r
e
s
th
er
eb
y
en
h
an
c
in
g
cla
s
s
if
i
ca
tio
n
p
er
f
o
r
m
an
ce
.
2
.
4
.
F
e
a
t
ure
f
us
io
n
Featu
r
e
f
u
s
io
n
is
p
e
r
f
o
r
m
ed
af
ter
ex
tr
ac
tin
g
f
ea
tu
r
es
f
r
o
m
E
f
f
icien
tNetB
0
f
o
r
v
id
eo
/i
m
ag
e
d
ata,
T
FID
F
f
o
r
tex
t,
an
d
MFC
C
co
ef
f
icien
ts
f
o
r
au
d
io
.
Fu
s
io
n
is
ca
r
r
ied
o
u
t
b
y
co
n
ca
ten
atin
g
th
e
ex
tr
ac
ted
f
ea
tu
r
e
v
ec
to
r
s
in
to
a
s
in
g
l
e
v
ec
to
r
f
o
r
m
u
ltimo
d
al
r
e
p
r
esen
tatio
n
.
T
h
is
f
u
s
io
n
es
tab
lis
h
es
es
s
en
tial
co
n
n
ec
tio
n
s
b
etwe
en
d
if
f
er
e
n
t
m
o
d
alities
,
allo
win
g
th
e
m
to
lev
er
ag
e
ea
ch
o
t
h
er
'
s
d
ata
to
m
ax
im
iz
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
.
T
h
e
in
teg
r
ated
tech
n
iq
u
e
h
el
p
s
ad
d
r
ess
v
ar
iatio
n
s
in
ex
p
r
ess
io
n
r
ep
r
esen
tatio
n
s
ac
r
o
s
s
d
if
f
er
en
t
m
o
d
alities
.
T
h
e
in
ter
co
n
n
ec
ted
f
ea
tu
r
e
v
ec
to
r
p
r
o
v
id
es a
b
et
ter
u
n
d
er
s
tan
d
i
n
g
o
f
f
ac
ial
ex
p
r
ess
io
n
s
,
lead
in
g
to
im
p
r
o
v
e
d
r
ec
o
g
n
itio
n
r
esu
lts
.
R
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
d
e
p
en
d
s
o
n
th
e
co
n
ca
te
n
ated
f
e
atu
r
e
v
ec
to
r
,
as
i
t
s
er
v
es
as
in
p
u
t
to
th
e
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
,
wh
ich
b
en
ef
it
s
f
r
o
m
th
e
ad
d
itio
n
al
in
f
o
r
m
atio
n
p
r
o
v
id
ed
b
y
co
m
b
in
in
g
m
u
ltimo
d
al
d
ata.
2
.
5
.
Rec
o
g
nitio
n
T
h
e
f
u
s
ed
f
ea
tu
r
es
ar
e
p
r
o
v
id
ed
as
in
p
u
t
to
th
e
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
.
W
h
ile
L
STM
n
etwo
r
k
s
ar
e
wid
ely
u
s
ed
f
o
r
s
eq
u
en
tial
d
ata
p
r
o
ce
s
s
in
g
,
th
ey
s
tr
u
g
g
le
to
ca
p
tu
r
e
co
m
p
lex
d
ep
en
d
e
n
cies
am
o
n
g
m
u
ltip
le
m
o
d
alities
in
f
ac
ial
e
x
p
r
ess
io
n
r
ec
o
g
n
itio
n
.
T
r
ad
itio
n
al
L
S
T
M
[
2
5
]
p
r
o
ce
s
s
es
in
p
u
t
s
eq
u
en
ce
s
in
d
ep
en
d
en
tl
y
,
lim
itin
g
its
ab
ilit
y
to
ca
p
t
u
r
e
in
tr
icate
r
elatio
n
s
h
ip
s
am
o
n
g
v
is
u
al,
au
d
io
an
d
tex
t
d
ata.
L
STM
h
as
s
ev
er
al
l
im
itatio
n
s
,
s
u
ch
as
in
ef
f
ec
tiv
e
h
an
d
lin
g
o
f
m
u
ltimo
d
al
d
e
p
e
n
d
en
cies,
d
if
f
icu
lty
in
ca
p
tu
r
in
g
f
in
e
-
g
r
ain
e
d
in
t
er
ac
tio
n
s
am
o
n
g
v
ar
io
u
s
in
p
u
t
m
o
d
alities
,
an
d
s
lo
wer
c
o
n
v
er
g
en
ce
d
u
e
to
u
n
id
ir
ec
tio
n
al
p
r
o
ce
s
s
in
g
.
T
h
ese
is
s
u
es
r
ed
u
ce
its
ef
f
icien
cy
in
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
,
wh
er
e
b
o
th
s
p
atial
an
d
tem
p
o
r
al
d
ep
en
d
e
n
cies m
u
s
t b
e
m
o
d
e
led
ef
f
ec
tiv
ely
.
Stan
d
ar
d
L
STM
s
tr
u
g
g
les
wit
h
h
an
d
lin
g
c
o
m
p
lex
in
ter
ac
tio
n
s
,
p
ar
ticu
lar
ly
in
m
u
ltimo
d
al
s
ce
n
ar
io
s
.
T
h
e
m
o
g
r
if
ier
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
ML
STM
)
a
d
d
r
es
s
es
th
is
b
y
in
tr
o
d
u
cin
g
t
h
e
m
o
g
r
if
ier
m
ec
h
an
is
m
,
wh
ich
d
y
n
am
ically
tr
an
s
f
o
r
m
s
in
p
u
t
f
ea
tu
r
es
m
u
ltip
le
tim
e
s
b
ef
o
r
e
f
ee
d
in
g
th
em
in
to
th
e
L
STM
ce
ll.
T
h
is
en
h
an
ce
s
f
e
atu
r
e
in
ter
ac
tio
n
a
n
d
d
ata
alig
n
m
en
t
ac
r
o
s
s
m
o
d
alities
,
im
p
r
o
v
i
n
g
m
o
d
el
f
le
x
ib
ilit
y
an
d
f
ea
t
u
r
e
r
ep
r
esen
tatio
n
in
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
.
T
h
e
ML
STM
im
p
r
o
v
es
m
u
lti
m
o
d
al
in
f
o
r
m
atio
n
f
u
s
io
n
b
y
d
y
n
am
ically
ad
a
p
tin
g
f
ea
tu
r
e
im
p
o
r
tan
ce
,
en
h
an
cin
g
ex
p
r
ess
iv
e
p
o
wer
a
n
d
g
en
e
r
aliza
tio
n
.
I
ts
alter
n
ati
v
e
u
p
d
ate
m
ec
h
an
is
m
allo
ws
d
ee
p
er
cr
o
s
s
-
m
o
d
al
f
ea
tu
r
e
r
ef
i
n
em
en
t,
lea
d
in
g
to
m
o
r
e
ac
cu
r
ate
e
m
o
ti
o
n
r
ec
o
g
n
itio
n
.
B
y
im
p
r
o
v
in
g
m
u
ltimo
d
al
f
ea
tu
r
e
in
teg
r
atio
n
,
m
in
im
izin
g
in
f
o
r
m
atio
n
lo
s
s
an
d
en
ab
li
n
g
q
u
ick
e
r
co
n
v
er
g
e
n
ce
,
th
e
ML
STM
b
ec
o
m
es
h
ig
h
ly
s
u
itab
le
f
o
r
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
.
I
t
ex
ten
d
s
th
e
s
tan
d
ar
d
L
STM
b
y
ad
d
i
n
g
two
g
atin
g
u
n
its
o
n
to
p
o
f
t
h
e
L
STM
,
en
h
an
cin
g
th
e
in
ter
ac
tio
n
s
p
ac
e
b
etwe
en
n
etwo
r
k
in
p
u
ts
an
d
o
u
tp
u
ts
.
Fig
u
r
e
2
s
h
o
ws
th
e
ML
STM
s
tr
u
ctu
r
e.
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
Mu
ltimo
d
a
l fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
u
s
in
g
r
esid
u
a
l m
o
g
r
ifier
…
(
Ma
ma
th
a
K
a
r
iya
p
p
a
R
a
ja
n
n
a
)
1571
Fig
u
r
e
2
.
C
ell
s
tr
u
ctu
r
e
o
f
ML
STM
I
n
ML
STM
,
th
e
i
n
p
u
t
(
)
an
d
p
r
ev
io
u
s
tim
estam
p
o
u
tp
u
t
ℎ
(
−
1
)
ar
e
alter
n
ativ
ely
s
cr
ee
n
e
d
b
ef
o
r
e
e
n
ter
in
g
th
e
L
STM
.
T
h
e
in
p
u
t
(
)
is
p
ass
ed
th
r
o
u
g
h
a
s
ig
m
o
id
th
r
esh
o
ld
u
n
it
to
att
ain
th
e
co
n
tr
o
l
s
tate
(
)
.
T
h
e
s
ig
m
o
id
f
u
n
ctio
n
en
s
u
r
es
th
at
th
e
v
alu
es
o
f
(
)
lie
b
etwe
en
[
0
,
1
]
.
T
h
e
(
)
an
d
ℎ
(
−
1
)
tr
an
s
f
o
r
m
ev
er
y
elem
en
t
in
ℎ
(
−
1
)
to
v
ar
io
u
s
d
eg
r
ee
s
.
I
f
(
)
h
as
a
v
al
u
e
o
f
1
f
o
r
a
g
iv
en
elem
en
t,
a
n
d
th
e
co
r
r
esp
o
n
d
in
g
elem
en
t
o
f
ℎ
(
−
1
)
is
p
ass
ed
in
to
th
e
n
etwo
r
k
with
its
ac
tu
al
v
al
u
e.
I
f
(
)
m
atch
es
an
elem
en
t
v
alu
e
o
f
0
.
5
in
ℎ
(
−
1
)
,
th
e
r
esp
ec
tiv
e
elem
en
t
in
ℎ
(
−
1
)
f
lo
ws
in
to
th
e
n
eu
r
al
n
etwo
r
k
af
te
r
s
p
litt
in
g
ev
er
y
elem
e
n
t
v
alu
e
in
to
(
)
.
T
h
e
v
alu
e
o
f
ev
e
r
y
el
em
en
t
in
(
)
is
a
d
er
iv
ativ
e
f
r
o
m
th
e
(
)
weig
h
t
at
ea
ch
lay
er
,
u
p
d
ated
co
n
tin
u
o
u
s
ly
d
u
r
in
g
tr
ain
in
g
to
r
ed
u
ce
n
etwo
r
k
lo
s
s
.
T
h
e
u
p
d
ate
p
r
o
ce
s
s
f
o
r
(
)
an
d
ℎ
(
−
1
)
′
ar
e
m
ath
em
atica
lly
f
o
r
m
u
lated
in
(
8
)
an
d
(
9
)
.
(
)
=
(
∙
(
)
+
)
(
8
)
ℎ
(
−
1
)
′
=
2ℎ
(
−
1
)
∙
(
−
1
)
+
ℎ
(
−
1
)
(
9
)
W
h
er
e
is
th
e
we
ig
h
t
o
f
th
e
in
p
u
t
(
)
th
at
co
n
tr
o
ls
ℎ
(
−
1
)
,
an
d
is
a
b
i
as.
T
h
e
ℎ
(
−
1
)
is
co
n
v
er
ted
in
to
(
)
by
ℎ
(
−
1
)
′
.
T
h
e
f
in
al
o
u
tp
u
t
o
f
th
e
R
NN
is
o
b
tain
ed
f
r
o
m
(
)
.
I
n
th
e
tr
ain
i
n
g
p
h
ase,
th
e
n
eu
r
al
n
etwo
r
k
is
u
p
d
ated
u
s
in
g
th
e
t
im
e
s
er
ies
(
)
.
I
f
a
p
ar
ticu
lar
elem
en
t
with
in
th
e
tim
e
s
er
ies
o
f
in
p
u
t
s
h
o
ws
a
d
r
am
atic
ch
a
n
g
e,
th
en
th
e
n
e
two
r
k
n
ee
d
s
to
b
e
b
etter
tu
n
ed
.
T
h
e
ML
STM
ad
d
r
ess
es
th
is
b
y
i
n
tr
o
d
u
cin
g
ℎ
(
−
1
)
′
as
a
th
r
esh
o
ld
u
n
it
o
f
(
)
co
n
tr
o
l.
T
h
e
ℎ
(
−
1
)
′
is
p
r
o
ce
s
s
ed
t
h
r
o
u
g
h
a
s
ig
m
o
i
d
th
r
esh
o
ld
s
tr
u
ctu
r
e
to
g
en
er
ate
th
e
c
o
n
tr
o
l
s
tate
(
)
.
B
y
u
s
in
g
(
)
to
tr
an
s
f
o
r
m
ea
ch
elem
e
n
t
(
)
,
th
e
n
etwo
r
k
u
s
es
th
e
o
b
tain
ed
lo
s
s
v
alu
e
to
f
in
e
-
tu
n
e
t
h
e
weig
h
ts
o
f
th
e
lay
er
f
o
r
g
r
ad
ie
n
t
u
p
d
ates.
T
h
e
u
p
d
ate
p
r
o
ce
s
s
f
o
r
(
)
an
d
(
−
1
)
′
ar
e
g
iv
en
in
(
1
0
)
a
n
d
(
1
1
)
.
(
)
=
(
∙
ℎ
(
−
1
)
′
+
)
(
1
0
)
(
)
′
=
2
(
)
∙
(
)
+
(
)
(
1
1
)
W
h
er
e
is
th
e
in
p
u
t
weig
h
t
ℎ
(
−
1
)
′
o
n
(
)
,
p
r
o
v
id
in
g
a
d
eq
u
ate
co
n
tr
o
l,
an
d
is
th
e
b
ias
m
atr
ix
,
(
)
is
co
n
v
er
ted
in
to
(
)
u
s
in
g
(
)
′
.
Utilizin
g
r
esid
u
al
co
n
n
ec
tio
n
s
with
in
ML
STM
f
o
r
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
en
h
a
n
c
es
th
e
s
y
s
tem
’
s
ab
ilit
y
to
p
r
o
c
ess
d
iv
er
s
e
in
p
u
t
m
o
d
alities
in
clu
d
in
g
a
u
d
io
,
tex
t
an
d
im
ag
e/v
id
eo
.
R
esid
u
al
c
o
n
n
ec
tio
n
s
en
s
u
r
e
th
at
th
e
o
r
ig
in
al
in
p
u
t
f
ea
tu
r
es
r
em
ain
in
tact
as
th
ey
p
ass
th
r
o
u
g
h
u
n
to
u
ch
ed
tr
a
n
s
f
o
r
m
atio
n
s
,
p
r
eser
v
in
g
ess
en
tial
d
ata
f
r
o
m
all
m
o
d
alities
ac
r
o
s
s
v
ar
io
u
s
p
r
o
ce
s
s
in
g
lay
er
s
.
T
h
is
s
ig
n
if
ican
tly
r
ed
u
ce
s
th
e
ch
an
ce
o
f
g
r
ad
ien
t
v
a
n
is
h
in
g
d
u
e
to
im
p
r
o
v
ed
g
r
ad
i
en
t
f
lo
w
with
in
th
e
n
etwo
r
k
,
allo
win
g
d
ee
p
er
ar
c
h
itectu
r
es
to
p
er
f
o
r
m
m
o
r
e
ef
f
ec
tiv
ely
.
T
h
e
ML
STM
lev
er
ag
es
th
ese
r
esid
u
al
co
n
n
ec
tio
n
s
to
m
ain
tain
b
o
t
h
tr
an
s
f
o
r
m
ed
a
n
d
r
aw
in
p
u
t
d
at
a,
en
ab
lin
g
th
e
s
y
s
tem
to
d
etec
t
s
u
b
tle
em
o
tio
n
al
cu
es
ac
r
o
s
s
d
if
f
er
en
t
m
o
d
alities
.
R
esid
u
al
co
n
n
ec
tio
n
s
in
ML
STM
im
p
r
o
v
e
p
r
ec
is
io
n
in
d
etec
tin
g
co
m
p
le
x
f
ac
ial
ex
p
r
ess
io
n
s
b
y
p
r
o
m
o
tin
g
r
ap
i
d
m
o
d
el
lea
r
n
in
g
,
wh
ile
m
ain
tain
in
g
s
tab
le
f
ea
tu
r
e
ex
tr
ac
tio
n
a
n
d
co
m
b
in
atio
n
.
T
h
ey
h
elp
p
r
e
s
er
v
e
f
ea
tu
r
es
f
r
o
m
ex
ce
s
s
iv
e
d
is
to
r
tio
n
,
f
ac
ilit
atin
g
ef
f
e
ctiv
e
m
o
d
elin
g
o
f
m
icr
o
-
ex
p
r
ess
io
n
s
an
d
v
o
ca
l
t
o
n
e
v
ar
iatio
n
s
.
C
o
n
s
eq
u
en
tly
,
th
e
s
y
s
tem
b
ec
o
m
es
m
o
r
e
ad
ep
t
at
r
ec
o
g
n
izin
g
em
o
tio
n
s
in
co
n
tex
t,
a
d
ap
ti
n
g
to
ch
a
n
g
es
i
n
s
p
ea
k
e
r
id
en
tity
an
d
e
n
v
ir
o
n
m
en
tal
co
n
d
itio
n
s
,
wh
ile
s
u
cc
ess
f
u
lly
p
r
o
ce
s
s
in
g
v
is
u
al
in
f
o
r
m
atio
n
lo
s
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.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
5
6
6
-
1
5
7
7
1572
T
h
e
R
ML
STM
in
tr
o
d
u
ce
s
k
ey
in
n
o
v
atio
n
s
th
at
e
n
h
an
ce
m
u
l
tim
o
d
al
f
ac
ial
e
x
p
r
ess
io
n
r
ec
o
g
n
itio
n
b
y
im
p
r
o
v
in
g
f
ea
tu
r
e
in
teg
r
atio
n
an
d
tem
p
o
r
al
m
o
d
elin
g
.
T
h
e
m
o
g
r
if
ier
m
ec
h
a
n
is
m
d
y
n
am
i
ca
lly
tr
an
s
f
o
r
m
s
an
d
r
ef
in
es
in
p
u
t
in
ter
ac
tio
n
s
am
o
n
g
m
o
d
alities
,
en
s
u
r
in
g
ef
f
ec
ti
v
e
co
n
tex
tu
al
u
n
d
er
s
tan
d
i
n
g
a
n
d
d
ata
f
u
s
io
n
.
T
h
e
R
ML
STM
ad
d
r
ess
e
s
v
an
is
h
in
g
g
r
ad
ien
t
is
s
u
es,
en
ab
lin
g
d
e
ep
er
n
etwo
r
k
tr
ain
in
g
wh
ile
r
etain
in
g
lo
n
g
-
ter
m
d
ep
en
d
e
n
cies.
Ad
d
itio
n
ally
,
its
iter
ativ
e
f
ea
tu
r
e
r
ef
in
em
en
t
o
p
tim
izes
h
id
d
en
s
tate
r
ep
r
esen
tatio
n
s
,
lead
in
g
to
ac
cu
r
ate,
d
is
cr
im
in
ativ
e
em
b
e
d
d
in
g
s
f
o
r
em
o
tio
n
r
ec
o
g
n
itio
n
.
T
h
r
o
u
g
h
th
ese
ad
v
a
n
ce
m
e
n
ts
,
th
e
R
ML
STM
s
ig
n
if
ican
tly
im
p
r
o
v
es r
ec
o
g
n
i
tio
n
ac
cu
r
ac
y
in
m
u
ltim
o
d
al
f
a
cial
ex
p
r
ess
io
n
an
aly
s
is
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
R
ML
STM
is
s
im
u
lated
in
MA
T
L
AB
(
R
2
0
2
0
b
)
with
s
y
s
tem
r
eq
u
ir
em
e
n
ts
o
f
W
in
d
o
ws
1
0
OS,
16
GB
R
AM
,
an
d
a
n
i7
p
r
o
ce
s
s
o
r
.
T
h
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
is
e
v
alu
ated
u
s
in
g
m
etr
ics
o
f
ac
c
u
r
ac
y
,
p
o
s
itiv
e
p
r
ed
ictiv
e
v
alu
e
(
PP
V)
,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
an
d
Ma
tth
ews
co
r
r
elatio
n
co
ef
f
icien
t
(
MCC
)
.
T
h
e
m
ath
em
atica
l
eq
u
atio
n
s
f
o
r
th
ese
m
etr
ics
ar
e
g
iv
en
in
(
1
2
)
to
(
1
6
)
.
W
h
er
e,
,
,
,
an
d
a
r
e
tr
u
e
p
o
s
itiv
es,
f
alse p
o
s
itiv
es,
f
alse n
eg
ativ
es
,
an
d
f
alse p
o
s
itiv
es,
r
esp
ec
tiv
ely
.
=
+
+
+
+
(
1
2
)
=
+
(
1
3
)
=
+
(
1
4
)
=
+
(
1
5
)
=
×
−
×
√
(
+
)
(
+
)
(
+
)
(
+
)
(
1
6
)
3
.
1
.
P
er
f
o
r
m
a
nce
o
f
m
ultimo
da
l f
ea
t
ures f
o
r
SAVE
E
a
nd
Yo
uT
u
be
da
t
a
s
et
s
I
n
th
is
s
u
b
s
ec
tio
n
,
th
e
p
er
f
o
r
m
an
ce
o
f
m
u
ltimo
d
al
f
ea
tu
r
es
f
o
r
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
i
s
an
aly
ze
d
u
s
in
g
v
ar
io
u
s
m
etr
ics.
T
ab
le
1
s
h
o
ws
th
e
p
er
f
o
r
m
a
n
ce
r
esu
lts
f
o
r
th
e
SAVE
E
d
at
aset,
wh
ile
T
ab
le
2
p
r
esen
ts
th
e
p
er
f
o
r
m
an
ce
r
esu
lts
f
o
r
th
e
Yo
u
T
u
b
e
d
ataset.
T
h
e
h
y
b
r
id
f
ea
tu
r
es
lev
er
a
g
e
co
m
p
lete
m
o
d
ality
co
m
p
lem
en
tar
ity
b
etwe
en
au
d
io
,
v
id
e
o
/im
ag
e
,
an
d
tex
t
to
a
ch
iev
e
o
p
tim
al
p
er
f
o
r
m
an
ce
.
T
h
e
h
y
b
r
id
m
o
d
el
in
teg
r
ates
au
d
io
s
ig
n
als,
v
is
u
a
l
s
ig
n
als
an
d
lin
g
u
is
tic
in
f
o
r
m
atio
n
,
r
esu
ltin
g
in
s
tr
o
n
g
er
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
an
d
im
p
r
o
v
e
d
d
ec
is
io
n
-
m
ak
i
n
g
.
B
y
c
o
m
b
in
i
n
g
t
h
e
p
a
r
ticip
atin
g
m
o
d
alities
,
th
e
p
e
r
f
o
r
m
an
ce
m
etr
ics
s
h
o
w
s
u
p
er
io
r
ac
cu
r
ac
y
an
d
p
r
ec
is
io
n
,
alo
n
g
s
id
e
en
h
an
ce
d
s
p
ec
if
icity
an
d
s
en
s
itiv
ity
,
as
well
as
a
h
ig
h
er
MCC
s
co
r
e
co
m
p
a
r
ed
to
in
d
iv
id
u
al
o
r
d
u
al
m
o
d
alities
.
T
h
e
h
y
b
r
id
f
ea
tu
r
es
ac
h
iev
e
9
9
.
5
7
%,
9
9
.
3
1
%,
9
9
.
1
6
%,
9
8
.
6
5
%,
an
d
9
8
.
8
3
%
f
o
r
ac
c
u
r
ac
y
,
PP
V,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
an
d
MCC
,
r
esp
ec
tiv
el
y
,
o
n
th
e
SAVE
E
d
ataset.
Similar
ly
,
h
y
b
r
id
f
ea
tu
r
es
ac
h
iev
e
9
7
.
8
3
%,
9
7
.
5
6
%,
9
7
.
2
4
%,
9
6
.
2
4
%,
an
d
9
6
.
7
5
%
f
o
r
ac
cu
r
ac
y
,
PP
V,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
a
n
d
MCC
,
r
esp
ec
tiv
ely
,
o
n
t
h
e
Yo
u
T
u
b
e
d
ataset.
T
ab
le
1
.
Per
f
o
r
m
an
ce
o
f
d
if
f
er
en
t f
ea
tu
r
e
e
x
tr
ac
tio
n
o
n
SAVE
E
d
ataset
M
u
l
t
i
m
o
d
a
l
f
e
a
t
u
r
e
s
A
c
c
u
r
a
c
y
(
%)
P
P
V
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
M
C
C
(
%)
A
u
d
i
o
9
7
.
5
2
9
7
.
2
8
9
7
.
1
4
9
6
.
6
7
9
6
.
8
1
V
i
d
e
o
/
i
m
a
g
e
9
5
.
4
3
9
5
.
1
9
9
5
.
0
7
9
4
.
5
8
9
4
.
7
2
Te
x
t
9
3
.
6
1
9
3
.
3
7
9
3
.
2
2
9
2
.
7
9
9
2
.
9
3
A
u
d
i
o
+
v
i
d
e
o
9
9
.
2
1
9
8
.
8
9
9
8
.
7
6
9
8
.
3
4
9
8
.
4
9
V
i
d
e
o
+
t
e
x
t
9
7
.
3
3
9
7
.
1
1
9
7
.
0
5
9
6
.
5
9
9
6
.
7
4
Te
x
t
+
a
u
d
i
o
9
5
.
4
8
9
5
.
2
6
9
5
.
1
3
9
4
.
6
2
9
4
.
7
9
H
y
b
r
i
d
f
e
a
t
u
r
e
s
(
a
u
d
i
o
,
v
i
d
e
o
,
t
e
x
t
)
9
9
.
5
7
9
9
.
3
1
9
9
.
1
6
9
8
.
6
5
9
8
.
8
3
T
ab
le
2
.
Per
f
o
r
m
an
ce
o
f
d
if
f
er
en
t f
ea
tu
r
e
e
x
tr
ac
tio
n
o
n
Yo
u
T
u
b
e
d
ataset
M
u
l
t
i
m
o
d
a
l
f
e
a
t
u
r
e
s
A
c
c
u
r
a
c
y
(
%)
P
P
V
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
M
C
C
(
%)
A
u
d
i
o
9
5
.
7
2
9
5
.
4
8
9
5
.
3
1
9
4
.
1
2
9
4
.
3
6
V
i
d
e
o
9
3
.
6
1
9
3
.
4
2
9
3
.
2
5
9
2
.
0
7
9
2
.
2
9
Te
x
t
9
1
.
5
5
9
1
.
3
3
9
1
.
1
7
9
0
.
0
5
9
0
.
2
8
A
u
d
i
o
+
v
i
d
e
o
9
7
.
2
1
9
6
.
8
9
9
6
.
6
6
9
5
.
4
1
9
5
.
6
3
V
i
d
e
o
+
t
e
x
t
9
5
.
3
3
9
5
.
1
2
9
4
.
9
8
9
3
.
8
1
9
4
.
0
5
Te
x
t
+
a
u
d
i
o
9
3
.
4
7
9
3
.
2
6
9
3
.
1
1
9
1
.
8
9
9
2
.
1
2
H
y
b
r
i
d
f
e
a
t
u
r
e
s
(
a
u
d
i
o
,
v
i
d
e
o
,
t
e
x
t
)
9
7
.
8
3
9
7
.
5
6
9
7
.
2
4
9
6
.
4
9
9
6
.
7
5
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
Mu
ltimo
d
a
l fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
u
s
in
g
r
esid
u
a
l m
o
g
r
ifier
…
(
Ma
ma
th
a
K
a
r
iya
p
p
a
R
a
ja
n
n
a
)
1573
3
.
2
.
P
er
f
o
r
m
a
nce
o
f
re
co
g
ni
t
io
n f
o
r
SAVE
E
a
nd
Yo
u
T
u
be
da
t
a
s
et
s
I
n
th
is
s
u
b
s
ec
tio
n
,
th
e
p
er
f
o
r
m
an
ce
o
f
r
ec
o
g
n
itio
n
f
o
r
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
is
an
aly
ze
d
u
s
in
g
v
ar
io
u
s
m
etr
ics.
T
ab
le
3
p
r
esen
ts
th
e
p
er
f
o
r
m
an
ce
f
o
r
th
e
SAVE
E
d
ataset,
wh
ile
T
ab
le
4
s
h
o
ws
th
e
p
er
f
o
r
m
an
ce
f
o
r
th
e
Yo
u
T
u
b
e
d
ataset.
T
h
e
R
ML
STM
o
u
tp
er
f
o
r
m
s
v
is
io
n
tr
a
n
s
f
o
r
m
er
(
ViT
)
,
s
win
tr
an
s
f
o
r
m
er
s
,
b
id
i
r
ec
tio
n
al
en
co
d
er
r
ep
r
esen
tatio
n
s
f
r
o
m
tr
a
n
s
f
o
r
m
er
(
B
E
R
T
)
,
m
u
lti
-
lay
e
r
p
er
ce
p
tr
o
n
(
ML
P),
R
NN,
L
STM
,
an
d
ML
STM
b
y
in
c
o
r
p
o
r
atin
g
a
d
v
an
ce
d
m
ec
h
an
is
m
s
f
o
r
d
etec
tin
g
tem
p
o
r
al
p
atter
n
s
an
d
tr
an
s
f
o
r
m
in
g
d
y
n
am
ic
f
ea
tu
r
e
ass
o
ciatio
n
s
.
T
h
e
m
em
o
r
y
a
n
d
g
atin
g
m
ec
h
an
is
m
s
in
R
ML
STM
en
ab
le
it
to
h
an
d
le
s
eq
u
e
n
tial
d
ata
m
o
r
e
ef
f
ec
tiv
ely
c
o
m
p
ar
e
d
to
M
L
P
an
d
R
NN.
Ad
d
itio
n
al
f
e
atu
r
es
in
R
ML
STM
im
p
r
o
v
e
u
p
o
n
s
tan
d
a
r
d
L
STM
an
d
ML
STM
b
y
u
tili
zin
g
atten
tio
n
-
lik
e
m
ec
h
an
is
m
s
a
n
d
g
r
ad
ien
t
u
p
d
ate
m
eth
o
d
s
to
ad
d
r
ess
n
o
n
-
lin
ea
r
d
ata
p
atter
n
s
an
d
lo
n
g
-
ter
m
d
ep
en
d
en
cies.
T
h
e
R
ML
STM
ac
h
iev
es
9
9
.
5
7
%
,
9
9
.
3
1
%,
9
9
.
1
6
%,
9
8
.
6
5
%,
an
d
9
8
.
8
3
%
f
o
r
ac
cu
r
ac
y
,
PP
V,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
an
d
MCC
,
r
esp
ec
tiv
ely
,
o
n
th
e
SAVE
E
d
ataset.
Similar
l
y
,
R
ML
STM
ac
h
iev
es
9
7
.
8
3
%,
9
7
.
5
6
%,
9
7
.
2
4
%,
9
6
.
2
4
%,
an
d
9
6
.
7
5
%
f
o
r
ac
cu
r
ac
y
,
PP
V,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
an
d
MC
C
,
r
esp
ec
tiv
ely
,
o
n
th
e
Yo
u
T
u
b
e
d
ataset.
R
ML
STM
o
u
tp
er
f
o
r
m
s
tr
an
s
f
o
r
m
er
-
b
ase
d
s
tr
u
ctu
r
es
d
u
e
to
its
s
u
p
er
i
o
r
ab
ilit
y
to
ca
p
tu
r
e
l
o
n
g
-
te
r
m
d
ep
en
d
e
n
cies.
I
n
co
n
tr
ast
to
tr
an
s
f
o
r
m
er
s
,
R
ML
STM
p
er
f
o
r
m
s
b
etter
o
n
m
o
d
e
r
ate
d
atasets
b
y
e
f
f
e
ctiv
ely
m
ain
tain
in
g
s
eq
u
en
tial
d
ep
en
d
e
n
cies.
Fu
r
th
er
m
o
r
e
,
as
a
tim
e
-
s
er
ies
an
d
s
eq
u
en
tial
m
o
d
el,
R
ML
STM
d
em
o
n
s
tr
ates
a
clea
r
ad
v
an
tag
e
o
v
er
t
r
an
s
f
o
r
m
e
r
m
o
d
els in
th
is
co
n
tex
t.
T
ab
le
3
.
Per
f
o
r
m
an
ce
o
f
d
if
f
er
en
t c
lass
if
ier
o
n
SAVE
E
d
ataset
C
l
a
s
si
f
i
e
r
A
c
c
u
r
a
c
y
(
%)
P
P
V
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
M
C
C
(
%)
V
i
T
8
7
.
6
4
8
7
.
4
6
8
7
.
2
8
8
6
.
5
2
8
6
.
4
1
S
w
i
n
t
r
a
n
sf
o
r
mers
8
9
.
2
7
8
9
.
1
4
8
9
.
0
4
8
8
.
7
6
8
8
.
5
9
B
ER
T
9
0
.
9
3
9
0
.
5
7
9
0
.
3
6
8
9
.
6
4
8
9
.
4
4
M
LP
9
1
.
5
2
9
1
.
3
8
9
1
.
1
9
9
0
.
0
5
9
0
.
2
4
R
N
N
9
3
.
6
7
9
3
.
4
9
9
3
.
3
1
9
2
.
1
2
9
2
.
3
4
LSTM
9
5
.
7
3
9
5
.
5
8
9
5
.
4
2
9
4
.
1
9
9
4
.
4
5
M
LST
M
9
7
.
4
9
9
7
.
2
9
9
7
.
0
8
9
6
.
0
7
9
6
.
2
5
R
M
LSTM
9
9
.
5
7
9
9
.
3
1
9
9
.
1
6
9
8
.
6
5
9
8
.
8
3
T
ab
le
4
.
Per
f
o
r
m
an
ce
o
f
d
if
f
er
en
t c
lass
if
ier
o
n
Yo
u
T
u
b
e
d
ata
s
et
C
l
a
s
si
f
i
e
r
A
c
c
u
r
a
c
y
(
%)
P
P
V
(
%)
S
p
e
c
i
f
i
c
i
t
y
(
%)
S
e
n
s
i
t
i
v
i
t
y
(
%)
M
C
C
(
%)
V
i
T
8
5
.
7
3
8
5
.
3
5
8
4
.
5
9
8
4
.
3
1
8
4
.
0
9
S
w
i
n
t
r
a
n
sf
o
r
mers
8
7
.
2
7
8
7
.
0
8
8
6
.
7
3
8
6
.
4
7
8
6
.
2
3
B
ER
T
8
9
.
6
2
8
9
.
2
7
8
8
.
8
5
8
8
.
5
2
8
8
.
1
8
M
LP
9
0
.
8
3
9
0
.
5
5
9
0
.
2
7
8
9
.
4
6
8
9
.
7
1
R
N
N
9
1
.
1
5
9
1
.
7
1
9
1
.
4
6
9
0
.
6
2
9
0
.
0
6
LSTM
9
3
.
4
1
9
3
.
1
4
9
3
.
6
3
9
2
.
0
4
9
2
.
3
3
M
LST
M
9
5
.
0
6
9
5
.
3
8
9
5
.
0
9
9
4
.
6
0
9
4
.
1
7
R
M
LSTM
9
7
.
8
3
9
7
.
5
6
9
7
.
2
4
9
6
.
4
9
9
6
.
7
5
T
h
e
in
teg
r
atio
n
o
f
r
esid
u
al
co
n
n
ec
tio
n
s
in
R
ML
STM
im
p
r
o
v
es
g
r
a
d
ien
t
f
lo
w
an
d
m
itig
ates
v
an
is
h
in
g
g
r
ad
ien
t iss
u
es in
d
e
ep
er
n
etwo
r
k
s
.
B
y
in
co
r
p
o
r
ati
n
g
th
ese
r
esid
u
al
co
n
n
ec
tio
n
s
,
R
ML
STM
en
s
u
r
es
b
etter
p
r
o
p
a
g
atio
n
ac
r
o
s
s
lay
er
s
,
lead
in
g
to
en
h
an
ce
d
le
ar
n
in
g
s
tab
ilit
y
an
d
ef
f
icien
cy
d
u
r
in
g
tr
ain
in
g
.
Sp
ec
if
ically
,
R
ML
STM
ex
ce
l
s
d
u
e
to
its
ab
ilit
y
to
ca
p
tu
r
e
lo
n
g
-
ter
m
d
ep
en
d
en
cies
in
s
eq
u
en
tial
d
ata
wh
ile
h
an
d
lin
g
less
co
m
p
u
tatio
n
al
ti
m
e
co
m
p
ar
e
d
to
tr
a
n
s
f
o
r
m
e
r
m
o
d
els.
T
ab
le
5
s
h
o
ws
th
e
co
m
p
lex
ity
an
aly
s
is
o
f
r
ec
o
g
n
itio
n
m
o
d
el
s
f
o
r
th
e
SAVE
E
a
n
d
Yo
u
T
u
b
e
d
atasets
.
T
h
e
ViT
,
Swin
T
r
a
n
s
f
o
r
m
e
r
,
B
E
R
T
,
ML
P,
R
NN,
L
STM
,
an
d
ML
STM
ar
e
co
n
s
id
er
e
d
as
ex
is
tin
g
m
eth
o
d
s
,
an
d
th
e
p
r
o
p
o
s
ed
R
ML
STM
is
co
m
p
ar
ed
with
th
ese
m
o
d
els.
T
h
e
co
m
p
lex
ity
is
an
aly
ze
d
i
n
ter
m
s
o
f
tr
ain
i
n
g
tim
e
an
d
test
in
g
tim
e
f
o
r
b
o
t
h
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
a
tasets
.
T
h
e
R
ML
STM
ac
h
iev
e
s
a
tr
ain
in
g
tim
e
o
f
1
7
4
m
i
n
u
tes
an
d
a
test
in
g
tim
e
o
f
0
.
0
3
s
ec
o
n
d
s
f
o
r
th
e
SA
VE
E
d
ataset.
Similar
ly
,
th
e
R
ML
STM
ac
h
iev
es
a
tr
ain
in
g
tim
e
o
f
1
7
6
m
in
u
tes
an
d
a
test
in
g
tim
e
o
f
0
.
0
4
s
ec
o
n
d
s
f
o
r
th
e
Yo
u
T
u
b
e
d
ataset.
T
h
is
is
d
u
e
to
its
s
u
p
er
io
r
r
esid
u
al
co
n
n
ec
tio
n
s
an
d
g
atin
g
m
ec
h
an
is
m
s
co
m
p
ar
ed
to
ViT
,
s
win
tr
an
s
f
o
r
m
er
,
B
E
R
T
,
ML
P,
R
NN,
L
STM
,
an
d
ML
STM
.
I
n
co
n
tr
ast
to
tr
an
s
f
o
r
m
er
m
o
d
els
s
u
ch
as
ViT
,
s
win
tr
an
s
f
o
r
m
er
,
an
d
B
E
R
T
,
th
e
R
ML
STM
m
ain
tain
s
a
s
eq
u
en
tial
p
r
o
ce
s
s
in
g
s
tr
u
ctu
r
e
with
lin
ea
r
tim
e
co
m
p
lex
ity
.
C
o
m
p
ar
ed
to
L
STM
an
d
ML
STM
,
th
e
r
esid
u
al
co
n
n
ec
tio
n
in
R
ML
STM
p
r
o
v
id
es
b
etter
co
n
v
er
g
en
ce
th
r
o
u
g
h
e
n
h
a
n
ce
d
g
r
ad
ie
n
t
f
lo
w,
th
er
eb
y
r
e
d
u
cin
g
t
h
e
v
an
is
h
in
g
g
r
ad
ien
t iss
u
e.
Fu
r
th
er
m
o
r
e,
it r
eq
u
ir
es f
ewe
r
p
ar
am
eter
s
t
h
an
ML
P a
n
d
R
NN,
th
u
s
r
ed
u
ci
n
g
c
o
m
p
u
tatio
n
ti
m
e.
T
h
e
p
r
o
p
o
s
ed
R
ML
STM
ac
h
iev
es
p
-
v
alu
es
o
f
0
.
0
0
8
an
d
0
.
0
0
1
2
f
o
r
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
,
r
esp
ec
tiv
ely
.
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.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
5
6
6
-
1
5
7
7
1574
T
ab
le
5
.
C
o
m
p
le
x
ity
an
aly
s
is
o
f
r
ec
o
g
n
itio
n
m
o
d
els f
o
r
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
C
l
a
s
si
f
i
e
r
S
A
V
EE
Y
o
u
T
u
b
e
Tr
a
i
n
i
n
g
t
i
me
(
mi
n
)
Te
st
i
n
g
t
i
m
e
(
s)
Tr
a
i
n
i
n
g
t
i
me
(
mi
n
)
Te
st
i
n
g
t
i
m
e
(
s)
V
i
T
1
9
8
0
.
2
1
9
5
0
.
2
1
S
w
i
n
t
r
a
n
sf
o
r
mers
1
9
4
0
.
1
6
1
9
2
0
.
1
9
B
ER
T
1
8
9
0
.
1
3
1
8
9
0
.
1
7
M
LP
1
8
6
0
.
1
1
1
8
7
0
.
1
5
R
N
N
1
8
4
0
.
9
1
8
4
0
.
1
2
LSTM
1
8
0
0
.
0
7
1
8
1
0
.
0
9
M
LST
M
1
7
7
0
.
0
5
1
7
8
0
.
0
7
R
M
LSTM
1
7
4
0
.
0
3
1
7
6
0
.
0
4
3
.
3
.
Co
m
pa
ra
t
iv
e
a
na
ly
s
is
T
h
is
s
ec
tio
n
p
r
esen
ts
a
co
m
p
ar
ativ
e
an
aly
s
is
o
f
th
e
p
r
o
p
o
s
ed
R
ML
STM
with
ex
is
tin
g
m
eth
o
d
s
is
p
r
o
v
id
e
d
f
o
r
th
e
SAVE
E
,
Yo
u
T
u
b
e,
C
K+
,
an
d
R
AVDE
E
S
d
atasets
.
T
h
e
3
D
-
C
NN
-
C
o
n
v
L
STM
[
1
6
]
,
L
STM
+
a
tten
tio
n
+CNN
-
2
D
[
1
7
]
,
DL
-
b
ased
f
ea
tu
r
e
ex
tr
ac
to
r
n
etwo
r
k
s
[
1
8
]
,
MFC
C
T
-
C
N
N
[
1
9
]
,
a
n
d
AM
PS
[
2
0
]
ar
e
c
o
n
s
id
er
ed
as
ex
is
tin
g
m
eth
o
d
s
to
d
em
o
n
s
tr
ate
t
h
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
p
r
o
p
o
s
ed
R
ML
STM
.
T
h
e
R
ML
STM
ac
h
iev
es
ac
cu
r
ac
y
r
ates
o
f
9
9
.
5
7
%,
9
7
.
8
3
%,
9
7
.
2
8
%,
an
d
9
4
.
6
3
%
f
o
r
th
e
SAVE
E
,
Yo
u
T
u
b
e
,
C
K+
,
an
d
R
AVDE
E
S d
ataset
s
,
r
esp
ec
tiv
ely
.
T
ab
le
6
s
h
o
ws t
h
e
co
m
p
a
r
ativ
e
an
aly
s
is
.
T
ab
le
6
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
o
f
p
r
o
p
o
s
ed
m
o
d
el
o
n
f
o
u
r
d
if
f
er
en
t d
atasets
D
a
t
a
s
e
t
M
e
t
h
o
d
A
c
c
u
r
a
c
y
(
%)
S
A
V
EE
3D
-
C
N
N
-
C
o
n
v
LST
M
[
1
6
]
9
8
.
8
3
LSTM
+
a
t
t
e
n
t
i
o
n
+
C
N
N
-
2
D
[
1
7
]
5
7
.
5
0
DL
-
b
a
se
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
o
r
n
e
t
w
o
r
k
s
[
1
8
]
99
M
F
C
C
T
-
C
N
N
[
1
9
]
93
R
M
LSTM
9
9
.
5
7
Y
o
u
T
u
b
e
A
M
P
S
[
2
0
]
7
9
.
7
R
M
LSTM
9
7
.
8
3
C
K
+
3D
-
C
N
N
-
C
o
n
v
LST
M
[
1
6
]
9
5
.
1
0
R
M
LSTM
9
7
.
2
8
R
A
V
D
EES
LSTM
+
a
t
t
e
n
t
i
o
n
+
C
N
N
-
2
D
[
1
7
]
7
4
.
4
4
DL
-
b
a
se
d
f
e
a
t
u
r
e
e
x
t
r
a
c
t
o
r
n
e
t
w
o
r
k
s
[
1
8
]
86
M
F
C
C
T
-
C
N
N
[
1
9
]
92
R
M
LSTM
9
4
.
6
3
3
.
4
.
Dis
cus
s
io
n
T
h
e
ad
v
an
tag
es
o
f
R
ML
STM
an
d
th
e
lim
itatio
n
s
o
f
ex
is
tin
g
m
eth
o
d
s
ar
e
d
is
cu
s
s
ed
in
th
is
s
ec
tio
n
.
T
h
e
ex
is
tin
g
tech
n
iq
u
es
s
u
f
f
er
f
r
o
m
m
u
ltip
le
lim
itatio
n
s
,
in
clu
d
in
g
p
o
o
r
ca
p
a
b
ilit
ies
in
d
etec
tin
g
f
in
e
tem
p
o
r
al
c
o
n
n
ec
tio
n
s
b
etwe
e
n
au
d
i
o
an
d
v
is
u
al
m
o
d
alitie
s
,
n
o
is
e
s
en
s
itiv
ity
an
d
d
iv
er
s
e
s
p
ee
ch
p
atter
n
s
.
T
h
ese
m
o
d
els
f
ail
to
d
is
ce
r
n
th
e
p
r
o
p
er
f
ea
tu
r
es
wh
ile
h
an
d
lin
g
m
u
ltip
le
m
o
d
alities
,
lead
in
g
to
co
m
p
lex
in
teg
r
atio
n
ch
allen
g
es.
T
o
o
v
e
r
co
m
e
t
h
ese
is
s
u
es,
th
e
R
ML
STM
is
p
r
o
p
o
s
ed
f
o
r
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
b
ec
au
s
e
it
im
p
r
o
v
es
b
o
th
f
ea
tu
r
e
r
ep
r
esen
tatio
n
an
d
s
eq
u
e
n
tial
lear
n
in
g
m
eth
o
d
s
.
E
ac
h
r
esid
u
al
elem
en
t
h
elp
s
r
ed
u
ce
g
r
a
d
ien
t
v
an
is
h
in
g
d
u
r
in
g
b
ac
k
p
r
o
p
ag
atio
n
,
en
ab
lin
g
ef
f
ec
tiv
e
tr
ain
in
g
in
DL
m
o
d
els.
I
n
ML
STM
,
th
e
in
p
u
t
r
e
p
r
esen
tatio
n
s
u
n
d
e
r
g
o
c
o
n
tin
u
o
u
s
tr
an
s
f
o
r
m
atio
n
s
th
at
im
p
r
o
v
e
th
e
r
elatio
n
s
h
ip
b
etwe
en
tem
p
o
r
al
an
d
v
is
u
al
f
ea
tu
r
es.
T
h
is
m
ec
h
a
n
is
m
im
p
r
o
v
es
th
e
m
o
d
el'
s
ef
f
icien
cy
in
d
etec
tin
g
s
u
b
tle
f
ac
ial
m
o
v
em
en
ts
an
d
tem
p
o
r
al
r
elatio
n
s
h
ip
s
.
T
h
e
u
s
e
o
f
R
ML
STM
s
tr
en
g
th
en
s
th
e
s
y
s
tem
's
p
er
f
o
r
m
an
ce
ag
ain
s
t
f
ac
ial
ex
p
r
ess
io
n
n
o
is
e,
r
esu
ltin
g
in
h
ig
h
er
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
.
T
h
e
r
esu
lts
ar
e
a
n
aly
ze
d
in
ter
m
s
o
f
q
u
an
titativ
e
an
d
c
o
m
p
ar
ativ
e
an
aly
s
is
f
o
r
b
o
th
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
.
T
h
e
R
ML
STM
ac
h
iev
es
9
9
.
5
7
%,
9
9
.
3
1
%,
9
9
.
1
6
%,
9
8
.
6
5
%,
an
d
9
8
.
8
3
%
f
o
r
ac
cu
r
ac
y
,
PP
V,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
an
d
MCC
f
o
r
th
e
SAVE
E
d
ataset.
Similar
ly
,
R
ML
STM
ac
h
iev
es
9
7
.
8
3
%,
9
7
.
5
6
%,
9
7
.
2
4
%,
9
6
.
2
4
%,
an
d
9
6
.
7
5
%
f
o
r
ac
cu
r
ac
y
,
PP
V,
s
p
ec
if
icity
,
s
en
s
itiv
ity
,
a
n
d
MCC
f
o
r
th
e
Yo
u
T
u
b
e
d
at
aset.
4.
CO
NCLU
SI
O
N
T
h
e
R
ML
STM
is
p
r
o
p
o
s
ed
in
th
is
r
esear
ch
f
o
r
m
u
ltimo
d
al
f
ac
ial
e
x
p
r
ess
io
n
r
ec
o
g
n
i
tio
n
.
T
h
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atas
ets
ar
e
u
s
ed
,
wh
ich
in
cl
u
d
e
v
id
eo
/im
ag
es,
tex
t,
a
n
d
a
u
d
i
o
m
o
d
alities
.
T
h
r
ee
p
r
elim
in
ar
y
o
p
er
atio
n
s
ar
e
p
er
f
o
r
m
ed
:
n
o
r
m
alizin
g
im
ag
es,
r
em
o
v
in
g
p
u
n
ctu
atio
n
f
r
o
m
tex
t
d
ata,
an
d
d
ir
ec
tly
ex
tr
ac
tin
g
MFC
C
f
ea
tu
r
es
f
r
o
m
au
d
i
o
in
p
u
ts
.
I
n
t
h
e
f
ea
t
u
r
e
ex
tr
ac
tio
n
s
tep
,
E
f
f
icien
t
NetB0
is
u
s
ed
f
o
r
p
r
o
ce
s
s
in
g
v
is
u
al
d
ata,
wh
il
e
T
FID
F
an
d
MFC
C
ar
e
e
m
p
lo
y
ed
f
o
r
tex
t
u
al
d
ata
a
n
d
s
p
ee
ch
s
ig
n
als,
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
Mu
ltimo
d
a
l fa
cia
l e
xp
r
ess
io
n
r
ec
o
g
n
itio
n
u
s
in
g
r
esid
u
a
l m
o
g
r
ifier
…
(
Ma
ma
th
a
K
a
r
iya
p
p
a
R
a
ja
n
n
a
)
1575
r
esp
ec
tiv
ely
.
E
f
f
icien
tNetB
0
s
er
v
es
as
an
ex
tr
ac
tiv
e
f
ea
tu
r
e
m
ec
h
an
is
m
f
o
r
im
ag
e/v
i
d
eo
s
am
p
les
d
u
e
to
its
ef
f
icien
t
co
m
p
u
tatio
n
o
f
v
is
u
al
p
atter
n
s
.
T
h
e
T
FID
F
p
r
o
c
ess
es
tex
t
d
ata
b
y
co
n
v
er
tin
g
v
er
b
al
in
f
o
r
m
atio
n
in
to
n
u
m
er
ical
v
ec
to
r
s
,
en
h
an
cin
g
th
e
im
p
o
r
tan
ce
o
f
m
ea
n
in
g
f
u
l
wo
r
d
s
with
o
u
t
d
im
in
is
h
in
g
s
t
an
d
ar
d
ter
m
s
.
T
h
e
MFC
C
an
aly
ze
s
au
d
io
s
ig
n
als
th
r
o
u
g
h
s
p
ec
tr
al
a
n
d
f
r
eq
u
en
cy
-
b
ased
ch
a
r
ac
ter
is
tics
to
cr
e
ate
ef
f
ec
tiv
e
au
d
io
s
ig
n
al
r
ep
r
esen
tatio
n
s
.
Su
b
s
eq
u
en
tly
,
th
e
f
ea
tu
r
es
f
r
o
m
ea
c
h
m
o
d
ality
ar
e
f
u
s
ed
,
an
d
th
e
co
m
b
in
e
d
f
ea
tu
r
es
ar
e
in
p
u
t
in
to
th
e
R
ML
STM
.
T
h
e
g
atin
g
an
d
m
em
o
r
y
m
ec
h
an
is
m
s
in
R
ML
STM
en
h
an
ce
its
ca
p
ab
ilit
ies
f
o
r
h
an
d
lin
g
s
eq
u
e
n
tial
d
ata,
th
e
r
eb
y
en
h
a
n
cin
g
m
u
ltimo
d
al
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
.
T
h
e
R
ML
STM
ac
h
iev
es
9
9
.
5
7
%
an
d
9
7
.
8
3
%
ac
cu
r
ac
y
f
o
r
th
e
SAVE
E
an
d
Yo
u
T
u
b
e
d
atasets
,
r
esp
ec
tiv
ely
.
I
n
th
e
f
u
tu
r
e,
v
ar
io
u
s
DL
tech
n
iq
u
es
ca
n
b
e
ap
p
lied
to
f
u
r
th
er
en
h
an
ce
r
ec
o
g
n
itio
n
p
er
f
o
r
m
an
ce
.
Ad
d
itio
n
ally
,
th
e
d
ev
elo
p
e
d
m
o
d
el
co
u
ld
b
e
i
m
p
r
o
v
e
d
to
en
h
a
n
ce
r
o
b
u
s
tn
ess
ag
ain
s
t
n
o
is
y
en
v
ir
o
n
m
e
n
ts
an
d
ad
v
er
s
ar
ial
attac
k
s
,
ex
p
lo
r
in
g
p
r
e
v
en
tiv
e
m
ea
s
u
r
es.
ACK
NO
WL
E
DG
M
E
N
T
S
T
h
e
au
th
o
r
s
g
r
atef
u
lly
ac
k
n
o
wled
g
e
th
e
s
u
p
p
o
r
t
an
d
f
a
cilities
p
r
o
v
id
ed
b
y
B
MS
I
n
s
titu
te
o
f
T
ec
h
n
o
lo
g
y
a
n
d
Ma
n
ag
em
en
t,
Yela
h
an
k
a,
Go
v
er
n
m
en
t
S
r
i
Kr
is
h
n
ar
ajen
d
r
a
Sil
v
er
J
u
b
ilee
T
ec
h
n
o
lo
g
ical
I
n
s
titu
te,
B
an
g
alo
r
e,
an
d
J
ain
Un
iv
er
s
ity
,
Kan
ak
ap
u
r
C
am
p
u
s
,
B
en
g
alu
r
u
,
f
o
r
ca
r
r
y
in
g
o
u
t
th
is
r
esear
ch
wo
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
Ma
m
ath
a
Kar
iy
ap
p
a
R
ajan
n
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
T
h
ejaswin
i Sh
an
k
ar
✓
✓
✓
✓
✓
✓
R
a
s
h
m
i
N
a
r
a
s
i
m
h
a
m
u
r
t
h
y
✓
✓
✓
✓
✓
✓
✓
Nan
d
h
in
i V
an
n
i
v
ed
u
L
ak
s
h
m
an
an
✓
✓
✓
✓
✓
✓
✓
✓
Har
ip
r
asad
S
.
An
an
th
ap
ad
m
an
ab
h
ar
ao
✓
✓
✓
✓
✓
✓
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
ataset
g
e
n
er
ated
d
u
r
in
g
th
e
cu
r
r
en
t
s
tu
d
y
ar
e
av
ailab
l
e
in
th
e
SAVE
E
,
C
K+
,
an
d
R
AVDE
E
S
r
ep
o
s
ito
r
y
at
h
ttp
://k
ah
la
n
.
ep
s
.
s
u
r
r
ey
.
ac
.
u
k
/s
av
ee
/,
h
ttp
s
://g
ts
.
ai/d
ata
s
et
-
d
o
wn
lo
ad
/
ck
-
d
ataset
-
ai
-
d
ata
-
co
llectio
n
/
,
an
d
h
ttp
s
://www.
i
n
n
o
v
atian
a
.
co
m
/en
/d
atasets
/r
av
d
ess
.
RE
F
E
R
E
NC
E
S
[
1
]
Y
.
D
a
i
,
J.
L
i
,
Y
.
Li
,
a
n
d
G
.
L
u
,
“
M
u
l
t
i
-
m
o
d
a
l
g
r
a
p
h
c
o
n
t
e
x
t
e
x
t
r
a
c
t
i
o
n
a
n
d
c
o
n
se
n
s
u
s
-
a
w
a
r
e
l
e
a
r
n
i
n
g
f
o
r
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
i
n
c
o
n
v
e
r
sa
t
i
o
n
,
”
K
n
o
w
l
e
d
g
e
-
Ba
s
e
d
S
y
st
e
m
s
,
v
o
l
.
2
9
8
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
k
n
o
sy
s
.
2
0
2
4
.
1
1
1
9
5
4
.
[
2
]
Y
.
Le
i
a
n
d
H
.
C
a
o
,
“
A
u
d
i
o
-
v
i
su
a
l
e
mo
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
w
i
t
h
p
r
e
f
e
r
e
n
c
e
l
e
a
r
n
i
n
g
b
a
se
d
o
n
i
n
t
e
n
d
e
d
a
n
d
mu
l
t
i
-
mo
d
a
l
p
e
r
c
e
i
v
e
d
l
a
b
e
l
s
,
”
I
EEE
T
ra
n
s
a
c
t
i
o
n
s
o
n
Af
f
e
c
t
i
v
e
C
o
m
p
u
t
i
n
g
,
v
o
l
.
1
4
,
n
o
.
4
,
p
p
.
2
9
5
4
–
2
9
6
9
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
TA
F
F
C
.
2
0
2
3
.
3
2
3
4
7
7
7
.
[
3
]
H
.
M
.
S
h
a
h
z
a
d
,
S
.
M
.
B
h
a
t
t
i
,
A
.
J
a
f
f
a
r
,
a
n
d
M
.
R
a
s
h
i
d
,
“
A
m
u
l
t
i
-
m
o
d
a
l
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
f
o
r
e
m
o
t
i
o
n
r
e
c
o
g
n
i
t
i
o
n
,
”
I
n
t
e
l
l
i
g
e
n
t
Au
t
o
m
a
t
i
o
n
a
n
d
S
o
f
t
C
o
m
p
u
t
i
n
g
,
v
o
l
.
3
6
,
n
o
.
2
,
p
p
.
1
5
6
1
–
1
5
7
0
,
2
0
2
3
,
d
o
i
:
1
0
.
3
2
6
0
4
/
i
a
sc.
2
0
2
3
.
0
3
2
5
2
5
.
Evaluation Warning : The document was created with Spire.PDF for Python.