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
.
3
,
J
u
n
e
20
26
,
p
p
.
2494
~
2
5
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
3
.
p
p
2
4
9
4
-
2
5
1
0
2494
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
M
ultiling
ua
l
sig
ns reco
g
nition us
in
g
rec
urren
t
n
eura
l net
wo
rk
T
ho
us
ee
f
Ulla
K
ha
n
1
,
Dilee
p
M
a
richi
Ra
m
a
cha
nd
ra
2
1
D
e
p
a
r
t
me
n
t
o
f
M
a
st
e
r
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
s
,
V
i
d
y
a
V
i
k
a
s I
n
st
i
t
u
t
e
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
T
e
c
h
n
o
l
o
g
y
,
V
i
sv
e
sv
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
M
a
st
e
r
o
f
C
o
m
p
u
t
e
r
A
p
p
l
i
c
a
t
i
o
n
s
,
N
i
t
t
e
M
e
e
n
a
k
s
h
i
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
B
e
l
a
g
a
v
i
,
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
an
5
,
2
0
2
5
R
ev
is
ed
Ap
r
23
,
2
0
2
6
Acc
ep
ted
May
11
,
2
0
2
6
Re
c
o
g
n
it
i
o
n
o
f
sig
n
lan
g
u
a
g
e
is
a
c
ru
c
ial
ste
p
to
wa
rd
s
p
ro
v
id
in
g
in
d
iv
i
d
u
a
l
s
with
h
e
a
rin
g
a
n
d
sp
e
e
c
h
imp
a
irme
n
t
m
e
a
n
in
g
fu
l
c
o
m
m
u
n
ica
ti
o
n
,
b
u
t
th
e
fa
c
t
th
a
t
th
e
re
a
re
a
n
u
m
b
e
r
o
f
d
ist
in
c
t
sig
n
lan
g
u
a
g
e
s
a
n
d
g
e
stu
re
s
re
m
a
in
c
o
m
p
lex
m
a
k
e
s
it
a
c
h
a
ll
e
n
g
e
to
t
h
e
c
u
rre
n
t
a
u
to
m
a
ted
sy
ste
m
s.
T
h
e
p
re
se
n
t
p
a
p
e
r
d
e
sc
rib
e
s
a
re
a
l
-
ti
m
e
m
u
lt
il
in
g
u
a
l
si
g
n
la
n
g
u
a
g
e
re
c
o
g
n
it
i
o
n
s
y
ste
m
th
a
t
is
b
a
se
d
o
n
a
re
c
u
rre
n
t
n
e
u
ra
l
n
e
two
rk
wit
h
lo
n
g
sh
o
rt
-
ter
m
m
e
m
o
ry
(RNN
-
LS
TM
)
with
h
a
n
d
lan
d
m
a
rk
M
e
d
iaPi
p
e
-
b
a
se
d
h
a
n
d
lan
d
m
a
r
k
d
e
tec
ti
o
n
to
s
u
c
c
e
ss
fu
ll
y
re
c
e
iv
e
sp
a
ti
a
l
a
n
d
tem
p
o
ra
l
g
e
stu
re
fe
a
t
u
re
s.
Th
e
p
ro
p
o
se
d
sy
ste
m
wa
s
train
e
d
a
n
d
tes
ted
o
v
e
r
a
se
lf
-
c
o
ll
e
c
ted
se
t
o
f
a
lp
h
a
b
e
t
g
e
stu
re
s
o
f
th
e
Ch
in
e
se
,
Am
e
rica
n
,
a
n
d
I
n
d
ia
n
sig
n
lan
g
u
a
g
e
,
i
n
c
l
u
d
i
n
g
o
n
e
-
h
a
n
d
a
n
d
two
-
h
a
n
d
g
e
stu
re
s,
a
n
d
wa
s
ru
n
wit
h
Ke
ra
s
with
e
x
ten
siv
e
p
e
rfo
rm
a
n
c
e
e
v
a
lu
a
ti
o
n
m
e
tri
c
s.
Th
e
stre
n
g
th
a
n
d
g
e
n
e
ra
li
z
a
ti
o
n
a
b
il
it
ies
o
f
th
e
su
g
g
e
ste
d
a
p
p
r
o
a
c
h
a
s
p
a
rt
o
f
d
iffere
n
t
g
e
stu
re
p
a
tt
e
rn
s
a
n
d
v
a
riatio
n
s
in
u
se
rs
a
re
c
o
n
firme
d
b
y
e
x
p
e
rime
n
tal
o
u
tco
m
e
s
t
h
a
t
in
d
ica
te
h
i
g
h
r
e
c
o
g
n
it
i
o
n
ra
tes
o
f
9
9
.
5
8
%
,
9
9
.
6
2
%
,
a
n
d
9
9
.
6
3
%
o
f
th
e
Ch
i
n
e
se
,
Am
e
rica
n
,
a
n
d
In
d
ia
n
sig
n
lan
g
u
a
g
e
s
,
re
sp
e
c
ti
v
e
l
y
.
Th
e
se
re
su
lt
s
d
e
m
o
n
stra
te
th
e
p
ro
m
ise
o
f
th
e
g
iv
e
n
fra
m
e
wo
rk
a
s
a
d
e
d
ica
ted
a
ss
isti
v
e
sy
ste
m
o
f
c
o
m
m
u
n
ica
ti
o
n
a
n
d
g
iv
e
it
a
so
li
d
b
a
se
t
o
c
o
n
ti
n
u
e
it
s
d
e
v
e
lo
p
m
e
n
t
to
th
e
p
o
in
t
o
f
t
h
e
sy
st
e
m
o
f
t
h
e
c
o
n
ti
n
u
o
u
s si
g
n
la
n
g
u
a
g
e
re
c
o
g
n
i
ti
o
n
(CS
LR)
a
n
d
m
u
lt
imo
d
a
l
tran
sla
to
rs.
K
ey
w
o
r
d
s
:
Han
d
g
estu
r
es
Han
d
lan
d
m
a
r
k
d
etec
tio
n
I
m
ag
e
p
r
o
ce
s
s
in
g
L
o
n
g
s
h
o
r
t
-
ter
m
m
e
m
o
r
y
r
ec
o
g
n
itio
n
Mu
ltil
in
g
u
al
g
estu
r
e
r
ec
o
g
n
itio
n
Sig
n
alp
h
ab
ets
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
:
T
h
o
u
s
ee
f
Ulla
Kh
an
Dep
ar
tm
en
t o
f
Ma
s
ter
o
f
C
o
m
p
u
ter
Ap
p
licatio
n
s
,
Vid
y
a
Vik
as I
n
s
titu
te
o
f
E
n
g
in
ee
r
in
g
a
n
d
T
ec
h
n
o
lo
g
y
Vis
v
esv
ar
ay
a
T
ec
h
n
o
l
o
g
ical
Un
iv
er
s
ity
J
n
an
a
San
g
am
a,
Ma
c
h
h
e
,
B
elag
av
i,
Kar
an
atak
a
-
5
9
0
0
1
8
,
I
n
d
ia
E
m
ail
:
th
o
u
s
ee
f
5
8
8
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Ar
tific
ial
in
tellig
en
ce
(
AI
)
is
an
em
er
g
in
g
tech
n
o
l
o
g
y
i
n
th
e
p
r
esen
t
wo
r
ld
.
I
t
u
s
es
th
e
m
ac
h
in
e
lear
n
in
g
an
d
d
ee
p
lear
n
in
g
to
en
h
a
n
ce
tech
n
o
lo
g
ies.
AI
c
o
u
ld
also
b
e
u
s
ed
to
d
e
v
elo
p
a
d
v
an
ce
d
s
y
s
tem
s
s
o
th
at
h
ea
r
in
g
an
d
s
p
ee
ch
im
p
ai
r
ed
p
eo
p
le
ca
n
co
m
m
u
n
icate
with
o
th
er
m
em
b
er
s
o
f
s
o
ciety
.
T
h
er
e
ar
e
v
ar
io
u
s
s
ig
n
lan
g
u
a
g
es
th
at
a
r
e
e
x
er
ci
s
e
in
d
if
f
e
r
en
t
p
ar
ts
o
f
th
e
wo
r
ld
an
d
s
o
m
e
o
f
th
em
in
clu
d
e
Am
er
ican
,
I
n
d
ian
,
C
h
in
ese,
Ko
r
ea
n
s
ig
n
lan
g
u
a
g
e,
an
d
m
a
n
y
m
o
r
e
.
T
h
ese
s
ig
n
lan
g
u
ag
es
ca
n
n
o
t
b
e
id
en
t
if
ied
b
y
an
av
e
r
ag
e
p
er
s
o
n
an
d
th
is
m
ak
es
th
em
b
e
n
eg
lecte
d
in
th
e
s
o
ciety
.
T
h
e
s
y
s
tem
th
at
ac
k
n
o
wled
g
es
s
ig
n
lan
g
u
a
g
e
en
ab
les
th
e
n
o
r
m
al
p
eo
p
le
to
lear
n
th
e
lan
g
u
a
g
e
o
f
th
e
h
ea
r
in
g
a
n
d
s
p
ee
ch
im
p
air
ed
p
e
o
p
le
b
y
tr
an
s
latin
g
th
e
h
an
d
g
estu
r
es
to
tex
t
m
ess
ag
e
th
at
will
en
h
an
ce
th
e
co
m
m
u
n
icatio
n
b
etwe
en
th
e
m
.
Sig
n
lan
g
u
a
g
e
will
b
e
co
m
p
o
s
ed
o
f
letter
s
,
n
u
m
b
er
s
,
wo
r
d
s
,
an
d
s
en
ten
ce
s
.
Sig
n
s
ar
e
ca
p
tu
r
e
d
with
d
if
f
e
r
en
t
h
a
n
d
g
estu
r
es
th
r
o
u
g
h
ca
m
er
a,
p
r
e
-
p
r
o
ce
s
s
in
g
o
f
th
e
ca
p
tu
r
ed
im
ag
es
ar
e
p
er
f
o
r
m
ed
,
s
ig
n
s
ar
e
r
ec
o
g
n
ized
w
ith
r
ec
u
r
r
e
n
t
n
e
u
r
al
n
etwo
r
k
(
R
NN
)
alg
o
r
ith
m
an
d
d
is
p
lay
th
e
s
u
itab
le
ch
ar
a
cter
s
o
n
o
u
tp
u
t scr
ee
n
as sh
o
wn
in
Fig
u
r
e
1
.
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
ltil
in
g
u
a
l sig
n
s
r
ec
o
g
n
itio
n
u
s
in
g
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Th
o
u
s
ee
f U
lla
K
h
a
n
)
2495
I
n
o
u
r
s
y
s
tem
,
we
h
a
v
e
in
c
o
r
p
o
r
ate
d
o
n
e
-
h
a
n
d
ed
s
ig
n
lan
g
u
ag
es
C
h
in
ese
a
n
d
Am
er
ican
,
an
d
two
-
h
an
d
e
d
I
n
d
ian
s
ig
n
lan
g
u
ag
e
as
s
ee
n
in
Fig
u
r
e
2
.
I
t
s
h
o
ws
th
e
o
n
e
-
h
an
d
e
d
a
n
d
two
-
h
a
n
d
ed
s
ig
n
lan
g
u
ag
es
in
clu
d
ed
i
n
th
is
s
tu
d
y
,
em
p
h
asis
in
g
th
e
v
ar
iatio
n
in
g
estu
r
e
p
atter
n
s
am
o
n
g
d
i
f
f
er
en
t
s
ig
n
lan
g
u
ag
e
s
y
s
tem
s
.
Fig
u
r
e
2
(
a)
s
h
o
ws
th
e
C
h
in
ese
s
ig
n
lan
g
u
a
g
e
al
p
h
a
b
ets
p
er
f
o
r
m
ed
u
s
in
g
s
in
g
le
h
an
d
g
estu
r
es
wh
ich
ar
e
co
m
p
ac
t
an
d
ef
f
icien
t
h
a
n
d
m
o
v
em
en
ts
.
T
h
e
Am
er
ica
n
s
ig
n
lan
g
u
a
g
e
alp
h
a
b
ets
ar
e
also
r
ep
r
esen
ted
u
s
in
g
o
n
e
-
h
a
n
d
g
estu
r
es
as
s
h
o
wn
in
Fig
u
r
e
2
(
b
)
.
T
h
e
I
n
d
ian
s
ig
n
lan
g
u
ag
e
alp
h
a
b
ets
m
ain
ly
in
v
o
lv
e
two
-
h
an
d
e
d
g
estu
r
es,
as sh
o
wn
in
Fig
u
r
e
2
(
c)
,
h
i
g
h
lig
h
tin
g
th
e
s
tr
u
ctu
r
al
co
m
p
lex
ity
h
a
n
d
l
ed
b
y
th
e
p
r
o
p
o
s
ed
r
ec
o
g
n
itio
n
m
o
d
el.
Fig
u
r
e
1
.
Sy
s
tem
ar
c
h
itectu
r
e
(
a)
(
b
)
(
c)
Fig
u
r
e
2
.
On
e
-
h
an
d
ed
an
d
two
-
h
an
d
e
d
s
ig
n
lan
g
u
ag
es c
o
n
s
id
er
ed
in
th
is
s
tu
d
y
o
f
(
a
)
C
h
in
e
s
e
s
ig
n
lan
g
u
a
g
e
al
p
h
ab
ets p
er
f
o
r
m
e
d
u
s
in
g
s
in
g
le
-
h
a
n
d
g
estu
r
es
,
(
b
)
Am
er
ican
s
ig
n
lan
g
u
ag
e
alp
h
ab
ets also
r
ep
r
esen
ted
with
o
n
e
-
h
a
n
d
g
estu
r
es
,
an
d
(
c
)
I
n
d
ian
s
ig
n
la
n
g
u
ag
e
alp
h
a
b
ets
,
wh
ich
p
r
im
ar
i
ly
in
v
o
lv
e
two
-
h
an
d
e
d
g
estu
r
es
W
e
g
o
o
v
e
r
r
e
le
v
a
n
t
r
ese
ar
ch
o
n
s
i
g
n
l
an
g
u
a
g
e
r
ec
o
g
n
iti
o
n
a
s
a
l
ite
r
a
tu
r
e
s
u
r
v
ey
.
K
asa
p
b
aş
i
et
a
l
.
[
1
]
,
p
r
o
p
o
s
e
d
co
n
v
o
l
u
ti
o
n
al
n
eu
r
al
n
et
wo
r
k
(
C
NN
)
m
o
d
e
l
c
o
n
s
is
ts
o
f
t
h
r
ee
c
o
n
v
o
l
u
t
io
n
al
l
ay
er
s
,
as
t
h
is
co
n
f
i
g
u
r
a
ti
o
n
y
iel
d
e
d
t
h
e
h
ig
h
est
ac
c
u
r
ac
y
d
u
r
i
n
g
t
h
e
ir
em
p
i
r
ic
al
t
ests
.
T
h
e
e
f
f
i
ci
en
c
y
o
f
9
9
.
3
8
%
wit
h
l
o
s
s
0
.
0
2
5
h
as
o
b
t
ai
n
e
d
o
n
A
m
e
r
ica
n
s
i
g
n
l
a
n
g
u
ag
e
.
G
h
o
tk
ar
et
a
l
.
[
2
]
,
m
o
d
el
r
e
co
g
n
i
ze
s
s
tati
c
I
n
d
ia
n
s
i
g
n
lan
g
u
a
g
e
d
i
g
its
,
ch
ar
ac
te
r
s
u
s
i
n
g
F
o
u
r
i
er
d
esc
r
i
p
t
o
r
t
o
e
x
t
r
ac
t
t
h
e
f
ea
t
u
r
es a
n
d
n
ea
r
es
t n
ei
g
h
b
o
u
r
h
o
o
d
c
lass
i
f
i
er
to
a
ch
ie
v
e
e
f
f
ic
ie
n
c
y
o
f
9
4
.
1
5
%
wit
h
E
u
cli
d
i
an
d
is
t
a
n
ce
.
Da
b
wa
n
e
t
a
l
.
[
3
]
,
a
p
p
r
o
ac
h
es
D
en
s
eNe
t
-
1
2
1
m
o
d
el
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
.
3
,
J
u
n
e
20
26
:
2
4
9
4
-
2
5
1
0
2496
wit
h
A
m
e
r
i
ca
n
s
i
g
n
la
n
g
u
a
g
e
ap
l
h
ab
et
d
atas
et
,
co
m
p
r
is
i
n
g
2
4
class
es
co
r
r
es
p
o
n
d
i
n
g
to
s
i
g
n
l
ett
e
r
s
,
ac
h
i
e
v
i
n
g
an
a
cc
u
r
ac
y
r
ate
o
f
9
7
%.
T
an
et
a
l
.
[
4
]
,
d
e
v
el
o
p
e
d
h
a
n
d
g
es
tu
r
e
r
e
co
g
n
it
io
n
wit
h
v
is
i
o
n
t
r
a
n
s
f
o
r
m
e
r
(
HGR
-
Vi
T
)
m
o
d
e
l
o
n
Na
ti
o
n
al
U
n
i
v
er
s
i
ty
o
f
S
in
g
a
p
o
r
e
(
NUS
)
h
a
n
d
g
est
u
r
es
d
atas
ets,
A
m
e
r
i
ca
n
s
ig
n
lan
g
u
a
g
e
n
u
m
b
e
r
s
a
n
d
A
m
e
r
i
c
an
s
i
g
n
l
an
g
u
a
g
e
c
h
ar
ac
t
e
r
s
,
a
n
d
o
b
tai
n
ed
a
n
e
f
f
ici
e
n
c
y
o
f
9
9
.
8
5
%,
9
9
.
3
6
%
,
a
n
d
9
9
.
9
8
%
r
es
p
ec
ti
v
el
y
.
L
ai
n
es
et
a
l
.
[
5
]
,
m
o
d
el
o
n
M
ex
ic
an
s
ig
n
l
an
g
u
a
g
e
u
s
in
g
s
ig
n
l
a
n
g
u
a
g
e
-
t
r
e
e
s
t
r
u
ct
u
r
e
s
k
el
et
o
n
i
m
a
g
e
(
SL
-
T
SS
I
)
D
e
n
s
eN
et
,
w
h
e
r
e
im
ag
es
a
r
e
p
r
o
ce
s
s
e
d
b
y
De
n
s
e
Net
-
1
2
1
,
w
h
i
ch
is
c
o
n
v
er
te
d
f
r
o
m
s
k
el
et
o
n
s
eq
u
e
n
ce
u
s
in
g
T
SS
I
w
it
h
ef
f
i
cie
n
cy
o
f
9
8
.
0
%.
Ku
m
a
r
i
a
n
d
A
n
a
n
d
[
6
]
,
h
y
b
r
i
d
a
r
c
h
ite
ct
u
r
e
o
f
C
NN
-
l
o
n
g
s
h
o
r
t
-
te
r
m
m
em
o
r
y
(
L
S
T
M
)
h
as
a
ch
ie
v
e
d
a
n
a
cc
u
r
ac
y
o
f
8
4
.
6
5
%
f
o
r
A
m
e
r
ica
n
s
i
g
n
la
n
g
u
a
g
e
test
e
d
o
n
1
0
0
cl
ass
es
o
f
w
o
r
d
-
le
v
el
A
m
e
r
i
ca
n
s
i
g
n
l
a
n
g
u
a
g
e
(
W
L
ASL
)
d
atas
et.
H
y
b
r
i
d
e
f
f
ici
en
t
c
o
n
v
o
lu
ti
o
n
m
o
d
el
[
7
]
,
wit
h
E
f
f
ic
ie
n
tN
et
-
B
3
a
n
d
a
f
ew
m
o
d
i
f
i
e
d
la
y
e
r
s
,
o
b
tai
n
s
ac
c
u
r
a
cy
o
f
9
3
.
1
7
%
t
o
r
ec
o
g
n
iz
e
h
u
n
d
r
ed
is
o
l
ate
d
d
y
n
a
m
i
c
B
a
n
g
l
a
s
ig
n
l
an
g
u
a
g
e
o
n
t
h
e
6
0
0
v
i
d
e
o
cli
p
s
d
atas
et
B
d
S
L
_
OPA
_
2
3
_
GE
S
T
UR
E
S
[
7
]
.
Gan
g
r
ad
e
a
n
d
B
h
ar
ti
[
8
]
,
u
s
ed
Mic
r
o
s
o
f
t
Kin
ec
t
s
en
s
o
r
t
o
id
en
tify
an
d
s
ep
ar
ate
h
a
n
d
p
o
r
ti
o
n
f
r
o
m
a
d
ep
th
p
ictu
r
e.
C
NN
u
s
ed
to
co
n
s
tr
u
ct
f
ea
tu
r
es
f
r
o
m
I
n
d
ian
s
ig
n
lan
g
u
ag
e
s
ig
n
als,
an
d
u
p
to
9
9
.
3
%
o
f
m
o
tio
n
s
ar
e
co
r
r
ec
tly
r
ec
o
g
n
ized
.
T
r
i
p
ath
i
et
a
l
.
[
9
]
,
m
o
d
el
co
n
v
e
r
ts
g
estu
r
es
to
tex
t
u
s
in
g
a
b
ag
o
f
v
is
u
al
wo
r
d
s
(
B
o
VW
)
co
n
ce
p
t.
A
co
llectio
n
o
f
4
2
,
0
0
0
p
h
o
to
s
in
cl
u
d
in
g
alp
h
ab
ets
an
d
d
ig
its
,
s
p
lit
in
to
3
5
class
es
with
1
,
2
0
0
im
ag
es
ea
ch
.
Sk
in
m
as
k
in
g
is
a
s
tep
in
th
e
p
r
e
-
p
r
o
c
ess
in
g
s
tag
e
th
at
in
v
o
lv
es
co
n
v
er
tin
g
th
e
R
GB
to
HSV,
u
s
in
g
ca
n
n
y
ed
g
e
d
etec
t
io
n
to
f
i
n
d
k
ee
n
e
d
g
es.
T
h
e
m
i
n
i
-
b
atch
K
-
m
ea
n
s
clu
s
ter
in
g
a
lg
o
r
ith
m
is
u
s
ed
t
o
ag
g
r
eg
ate
t
h
e
s
p
ee
d
e
d
-
u
p
r
o
b
u
s
t
f
ea
tu
r
es
(
SUR
F
)
elem
en
ts
an
d
tr
an
s
late
th
em
in
to
a
v
is
u
al
lan
g
u
a
g
e
.
Od
ar
tey
et
a
l
.
[
1
0
]
,
Gh
an
ai
an
s
ig
n
lan
g
u
a
g
e
was
class
i
f
ied
an
d
r
ec
o
g
n
ize
d
u
s
in
g
t
h
e
s
u
g
g
ested
d
ee
p
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
DC
NN
)
,
wh
ich
ac
h
ie
v
ed
9
6
.
0
%
ac
cu
r
ac
y
.
T
o
len
tin
o
et
a
l
.
[
1
1
]
,
C
NN
m
o
d
el,
wh
er
e
im
ag
es
wer
e
tr
ain
ed
u
s
in
g
Ker
as.
Am
er
ican
s
ig
n
lan
g
u
ag
e
letter
r
ec
o
g
n
itio
n
ac
co
u
n
ted
f
o
r
9
0
.
0
4
%
o
f
th
e
s
y
s
tem
's
av
er
ag
e
test
in
g
ac
cu
r
ac
y
o
f
9
3
.
6
7
%.
L
o
m
as
et
a
l.
[
1
2
]
,
s
u
g
g
ested
C
NN
m
eth
o
d
to
u
tili
ze
a
co
m
p
o
u
n
d
co
e
f
f
icien
t
to
m
ea
s
u
r
e
all
th
e
d
ep
th
,
b
r
ea
d
th
,
an
d
r
eso
lu
tio
n
p
ar
am
eter
s
eq
u
ally
wh
en
u
tili
zin
g
th
e
m
eth
o
d
o
l
o
g
y
o
n
E
f
f
icien
tNet,
an
d
id
en
tifie
d
t
h
e
m
eth
o
d
b
y
s
elec
tin
g
B
0
m
o
d
el
as
ar
ch
itectu
r
e
f
o
r
test
in
g
with
a
d
ataset
o
f
Kag
g
le
.
R
in
ald
u
zz
i
et
a
l
.
[
1
3
]
,
m
o
d
el
u
s
in
g
a
m
ag
n
etic
p
o
s
itio
n
in
g
s
y
s
tem
,
wh
ich
h
as
b
etter
g
e
n
er
aliza
tio
n
f
ea
tu
r
es
an
d
ac
h
iev
es
a
class
if
icatio
n
ef
f
icien
c
y
o
f
r
o
u
g
h
ly
9
7
%
o
n
2
4
ch
a
r
ac
ter
s
.
Alaf
tek
in
et
a
l
.
[
1
4
]
,
T
u
r
k
is
h
s
ig
n
lan
g
u
ag
e
d
i
g
its
d
ataset
i
s
m
ak
in
g
u
s
e
to
tr
ain
in
g
th
e
y
o
u
o
n
ly
lo
o
k
o
n
c
e
v
er
s
io
n
4
-
cr
o
s
s
s
tag
e
p
ar
tial
(
YOL
Ov
4
-
C
SP
)
m
o
d
els,
an
d
th
eir
o
u
tco
m
es
ar
e
co
m
p
ar
e
d
in
ter
m
s
o
f
h
an
d
s
ig
n
al
id
en
tific
atio
n
,
wh
ich
y
ield
s
th
e
r
esu
lts
with
p
r
ec
is
io
n
9
8
.
9
5
%,
r
ec
all
9
8
.
1
5
%,
F
1
-
s
co
r
e
9
8
.
5
5
,
a
n
d
m
ea
n
av
er
a
g
e
p
r
ec
is
io
n
(
m
AP
)
9
9
.
4
9
% in
9
.
8
m
s
.
Usi
n
g
a
p
ath
a
g
g
r
e
g
atio
n
f
ea
tu
r
e
p
y
r
am
id
n
etwo
r
k
with
th
e
ODM
am
b
a
b
ac
k
b
o
n
e
n
etwo
r
k
,
YOL
O
-
Ma
m
b
a
m
o
d
el
p
r
o
d
u
c
ed
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
o
f
9
3
.
3
% o
n
u
n
iq
u
e
d
ataset
[
1
5
]
.
Sas
id
h
ar
an
et
a
l
.
[
1
6
]
,
u
s
ed
v
ar
iety
o
f
s
p
atio
tem
p
o
r
a
l
f
ea
tu
r
es
g
ath
e
r
ed
f
r
o
m
v
id
e
o
s
eq
u
en
ce
s
o
f
s
ig
n
la
n
g
u
a
g
e
m
o
v
em
e
n
ts
,
wh
ich
en
ab
le
to
id
en
tify
s
ig
n
als
b
y
ca
p
tu
r
in
g
b
o
th
t
h
e
tem
p
o
r
al
d
y
n
am
ics
o
f
th
e
m
o
v
em
e
n
ts
an
d
th
e
s
p
atial
ar
r
an
g
em
e
n
t.
T
h
e
ac
cu
r
ac
y
was
9
0
.
8
7
%
f
o
r
C
h
in
ese
s
ig
n
l
an
g
u
ag
e
an
d
8
9
.
4
6
%
f
o
r
Ar
a
b
ic
s
ig
n
lan
g
u
ag
e
ch
ar
ac
ter
s
ig
n
s
.
G
ated
r
ec
u
r
r
e
n
t
u
n
it
(
GR
U)
m
o
d
el
u
s
ed
f
o
r
ex
tr
ac
tio
n
o
f
tem
p
o
r
al
an
d
s
p
a
tial
ch
ar
ac
ter
is
tics
,
attain
ed
ac
cu
r
ac
y
o
f
9
2
.
1
3
%
an
d
9
3
.
9
8
%
o
n
th
e
d
atasets
DHG1
4
/2
8
an
d
SHR
E
C
'
1
7
,
r
esp
ec
tiv
ely
[
1
7
]
.
Sh
ar
m
a
an
d
Sin
g
h
[
1
8
]
,
C
NN
m
o
d
el
o
n
I
n
d
ian
s
ig
n
lan
g
u
a
g
e
g
estu
r
e
f
ea
tu
r
e
ex
tr
ac
tio
n
an
d
ca
teg
o
r
izatio
n
.
B
o
th
a
p
u
b
licly
ac
ce
s
s
ib
le
Am
er
ican
s
ig
n
lan
g
u
ag
e
d
ataset
an
d
a
s
elf
-
co
llected
I
n
d
ian
s
ig
n
lan
g
u
ag
e
d
ataset
ar
e
u
s
ed
to
ass
es
s
th
is
m
eth
o
d
'
s
p
er
f
o
r
m
an
ce
.
T
h
r
ee
d
atase
ts
wer
e
u
s
ed
in
to
tal,
an
d
th
e
ac
cu
r
ac
y
o
b
tain
e
d
was
9
2
.
4
3
%
,
8
8
.
0
1
%
,
an
d
9
9
.
5
2
%.
I
n
o
r
d
er
to
id
e
n
tify
g
estu
r
es,
T
ian
et
a
l.
[
1
9
]
d
ev
elo
p
ed
th
e
DAM
R
_
3
DNe
t
m
o
d
el,
wh
ich
y
ield
ed
ac
c
u
r
ac
y
o
f
5
5
.
3
%
an
d
9
2
.
9
3
% o
n
t
h
e
HM
DB
5
1
a
n
d
UC
F1
0
1
d
atasets
,
r
esp
ec
tiv
ely
.
Sre
em
ath
y
et
a
l
.
[
2
0
]
,
illu
s
tr
ates
h
o
w
to
id
en
tif
y
th
e
alp
h
ab
ets
in
I
n
d
ian
s
ig
n
lan
g
u
ag
e
b
y
im
a
g
e
p
r
o
ce
s
s
in
g
tech
n
iq
u
es a
n
d
m
a
ch
in
e
lear
n
in
g
.
Ad
d
itio
n
ally
,
Alex
Net,
Go
o
g
leNe
t,
VGG
-
1
6
,
an
d
VGG
-
1
9
wer
e
u
s
ed
in
m
ac
h
in
e
lear
n
in
g
t
ec
h
n
iq
u
e,
y
ield
in
g
ac
c
u
r
ac
ie
s
o
f
9
9
.
1
1
%,
9
5
.
8
4
%,
9
8
.
4
2
%,
an
d
9
9
.
1
1
%,
r
esp
ec
tiv
ely
.
C
h
o
n
g
a
n
d
L
ee
[
2
1
]
,
cr
ea
ted
a
p
r
o
t
o
ty
p
e
f
o
r
r
ec
o
g
n
izin
g
s
ig
n
lan
g
u
ag
e
u
s
in
g
th
e
leap
m
o
tio
n
co
n
tr
o
ller
with
g
o
al
o
f
r
ec
o
g
n
izin
g
Am
er
ican
s
ig
n
lan
g
u
a
g
e
.
T
h
e
r
esu
lts
o
f
th
e
e
x
p
er
i
m
en
t
s
h
o
wed
th
at
th
e
2
6
letter
s
co
u
ld
b
e
r
ec
o
g
n
ize
d
u
s
in
g
d
ee
p
n
eu
r
al
n
etwo
r
k
(
DNN
)
an
d
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM
)
at
8
0
.
3
0
% a
n
d
9
3
.
8
1
%,
r
esp
ec
tiv
ely
.
Per
u
m
al
et
a
l
.
[
2
2
]
,
s
u
g
g
est
L
STM
m
o
d
els
u
s
in
g
r
ec
tifie
d
lin
ea
r
u
n
it
(
Re
L
U)
ac
tiv
atio
n
f
u
n
ctio
n
s
,
wh
ich
h
av
e
b
ee
n
s
h
o
w
n
to
b
e
9
6
.
5
5
%
with
ap
p
r
o
p
r
ia
te
tr
ain
in
g
,
with
L
2
r
eg
u
lar
izatio
n
an
d
C
NN.
C
h
u
n
g
et
a
l
.
[
2
3
]
,
cr
ea
ted
an
d
ass
ess
ed
a
f
r
am
ewo
r
k
f
o
r
C
h
in
ese
s
ig
n
lan
g
u
ag
e
id
en
tific
at
io
n
th
at
m
ak
es
u
s
e
o
f
s
p
atio
tem
p
o
r
al
ch
ar
ac
te
r
is
tics
.
T
h
e
h
an
d
lan
d
m
ar
k
s
f
ea
tu
r
es
ar
e
f
ed
in
to
a
b
id
ir
ec
tio
n
al
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
Bi
-
L
STM
)
s
y
s
tem
f
o
r
r
ec
o
g
n
izin
g
s
ig
n
.
A
co
n
ti
n
u
o
u
s
C
h
in
ese
s
ig
n
lan
g
u
ag
e
d
ataset
in
clu
d
i
n
g
1
,
2
0
0
s
am
p
le
f
ilm
s
en
co
m
p
ass
in
g
1
0
0
s
ig
n
als
,
an
d
8
p
ar
ticip
an
ts
,
th
e
r
ec
o
g
n
itio
n
r
ate
is
9
8
.
7
5
%.
L
i
et
a
l
.
[
2
4
]
,
tech
n
iq
u
e
f
o
r
v
o
ca
b
u
lar
y
s
eg
m
en
tatio
n
b
ased
o
n
th
e
ac
ce
ler
atio
n
s
ig
n
al
an
d
th
e
v
o
lu
m
e
s
tate
o
f
th
e
s
u
r
f
ac
e
elec
tr
o
m
y
o
g
r
ap
h
y
(s
EMG
)
s
ig
n
al.
T
h
e
av
er
ag
e
r
ec
o
g
n
itio
n
r
ate
o
f
s
ig
n
lan
g
u
ag
e
v
o
ca
b
u
lar
y
is
9
0
.
4
1
%
wh
e
n
th
e
m
u
lti
-
s
en
s
o
r
d
ec
is
io
n
f
u
s
io
n
ap
p
r
o
ac
h
o
f
h
id
d
e
n
Ma
r
k
o
v
m
o
d
el
is
em
p
lo
y
e
d
.
J
ian
g
et
a
l
.
[
2
5
]
,
g
iv
e
an
ex
te
n
s
iv
e
r
ev
iew
o
f
C
h
in
ese
s
ig
n
l
an
g
u
ag
e
r
ec
o
g
n
itio
n
ex
p
lain
in
g
h
o
w
it
d
e
v
elo
p
e
d
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
ltil
in
g
u
a
l sig
n
s
r
ec
o
g
n
itio
n
u
s
in
g
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Th
o
u
s
ee
f U
lla
K
h
a
n
)
2497
th
r
o
u
g
h
th
e
y
ea
r
s
b
eg
in
n
i
n
g
with
tr
ad
itio
n
al
f
ea
tu
r
e
-
b
ased
tech
n
iq
u
es
to
m
o
d
e
r
n
AI
an
d
d
ee
p
lear
n
i
n
g
-
b
ased
tech
n
o
lo
g
y
,
wh
ich
co
n
tex
t
u
alize
s
an
d
u
n
d
e
r
lin
es
th
e
m
eth
o
d
o
lo
g
y
u
s
ed
i
n
th
is
wo
r
k
.
L
iu
et
a
l
.
[
2
6
]
,
u
s
ed
C
NN
to
r
ec
o
g
n
ize
h
an
d
p
o
s
i
tio
n
s
an
d
m
o
v
em
e
n
t
tr
ajec
to
r
ies
b
y
co
m
b
in
in
g
s
tr
etch
y
s
tr
ain
s
en
s
o
r
s
with
body
-
m
o
u
n
ted
in
e
r
tial
m
ea
s
u
r
em
en
t
d
ev
ices,
4
8
C
h
in
ese
s
ig
n
lan
g
u
a
g
e
ter
m
s
u
s
ed
f
o
r
tr
ain
in
g
an
d
ac
h
iev
e
d
an
ef
f
icien
cy
o
f
9
5
.
8
5
%.
T
h
e
r
esear
ch
s
tu
d
ies
h
av
e
p
r
o
v
en
th
at
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
h
as
ad
v
an
ce
d
co
n
s
id
er
ab
ly
with
th
e
in
co
r
p
o
r
atio
n
o
f
d
ee
p
lear
n
in
g
,
k
ey
p
o
in
t
tr
ac
k
in
g
,
an
d
s
m
a
r
t
s
y
s
tem
d
esig
n
.
Als
h
ar
if
et
a
l
.
[
2
7
]
,
s
u
g
g
ested
a
r
ea
l
-
tim
e
s
y
s
tem
o
f
Am
er
ican
s
ig
n
lan
g
u
ag
e
in
ter
p
r
etatio
n
with
th
e
u
s
e
o
f
d
ee
p
lear
n
in
g
an
d
k
ey
p
o
in
t
-
b
ased
tr
ac
k
in
g
o
f
th
e
h
an
d
s
,
wh
ich
i
s
h
ig
h
ly
ac
cu
r
ate
an
d
r
ea
l
-
tim
e.
Ma
ash
i
et
a
l
.
[
2
8
]
p
r
o
p
o
s
ed
an
I
o
T
-
b
ased
s
m
ar
t
ass
is
tiv
e
co
m
m
u
n
icatio
n
s
y
s
te
m
b
u
ilt
o
n
h
y
b
r
id
d
ee
p
lear
n
in
g
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
m
o
d
els,
with
a
f
o
cu
s
o
n
s
ca
lab
ilit
y
a
n
d
r
ea
l
-
wo
r
l
d
ap
p
licatio
n
to
h
ea
r
in
g
-
im
p
air
ed
u
s
er
s
.
Al
-
L
atief
et
a
l.
[
2
9
]
p
r
esen
ted
a
d
escr
ip
tiv
e
co
m
p
ar
ativ
e
r
ev
i
ew
o
f
d
ee
p
lear
n
i
n
g
m
o
d
e
ls
in
th
e
r
ec
o
g
n
itio
n
o
f
s
ig
n
lan
g
u
a
g
e
an
d
s
y
s
tem
atica
lly
an
aly
ze
d
th
em
in
ter
m
s
o
f
ar
ch
itectu
r
e,
d
at
a
s
ets,
an
d
ch
allen
g
es,
wh
ich
will
co
n
tr
ib
u
te
to
im
p
o
r
tan
t
f
in
d
i
n
g
s
o
n
t
h
e
e
x
is
tin
g
tr
en
d
s
an
d
r
esear
ch
g
ap
s
.
Mo
r
eo
v
er
,
B
aih
an
et
a
l
.
[
3
0
]
h
av
e
s
u
g
g
ested
a
n
im
p
r
o
v
e
d
d
ee
p
lear
n
in
g
m
o
d
e
l
with
im
p
r
o
v
ed
h
y
b
r
id
o
p
tim
i
za
tio
n
alg
o
r
ith
m
,
wh
ic
h
s
h
o
w
s
b
etter
r
ec
o
g
n
itio
n
r
esu
lt
an
d
s
tab
ilit
y
,
f
u
r
t
h
er
c
o
n
f
ir
m
in
g
th
e
ef
f
icien
cy
o
f
a
d
v
an
ce
d
o
p
tim
izatio
n
al
g
o
r
ith
m
in
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
alg
o
r
ith
m
.
I
n
p
ar
allel,
th
e
r
esear
ch
co
m
m
u
n
ity
is
in
cr
ea
s
in
g
ly
ad
o
p
tin
g
tr
an
s
f
o
r
m
er
-
b
ased
an
d
m
u
lti
-
s
ca
le
tem
p
o
r
al
m
o
d
elin
g
ap
p
r
o
ac
h
es
f
o
r
co
n
tin
u
o
u
s
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
(
C
SLR)
,
ef
f
ec
tiv
ely
ad
d
r
ess
in
g
v
ar
iab
le
s
ig
n
d
u
r
atio
n
s
a
n
d
lo
n
g
-
r
an
g
e
tem
p
o
r
al
d
ep
en
d
en
cies
[
3
1
]
,
wh
ile
in
teg
r
ated
C
SLR
-
to
-
s
p
ee
ch
p
ip
elin
es
ar
e
b
ein
g
ex
p
lo
r
ed
to
m
o
v
e
b
ey
o
n
d
g
estu
r
e
class
if
icatio
n
to
war
d
e
n
d
-
to
-
en
d
ass
is
tiv
e
co
m
m
u
n
icatio
n
s
y
s
tem
s
[
3
2
]
.
R
ec
en
t
wo
r
k
s
a
r
e
also
f
o
cu
s
ed
o
n
th
e
r
o
b
u
s
tn
ess
,
s
ca
lab
ilit
y
,
an
d
p
r
ac
tical
d
ep
lo
y
m
e
n
t
o
f
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
(
SLR)
s
y
s
tem
s
.
Fo
r
C
SLR
,
atten
tiv
e
s
p
atio
-
tem
p
o
r
al
n
etwo
r
k
s
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
to
e
f
f
ec
tiv
ely
ca
p
tu
r
e
tem
p
o
r
al
d
ep
e
n
d
en
cie
s
o
v
er
lo
n
g
g
estu
r
e
s
eq
u
en
ce
s
[
3
3
]
.
I
n
ad
d
itio
n
,
p
o
s
e
-
esti
m
atio
n
-
d
r
iv
e
n
d
ee
p
l
ea
r
n
in
g
f
r
am
ewo
r
k
s
h
av
e
b
ee
n
in
v
esti
g
ated
to
e
n
h
an
ce
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
u
n
d
er
v
ar
iatio
n
s
in
s
ig
n
er
p
o
s
tu
r
e
an
d
v
iewp
o
i
n
t
[
3
4
]
.
T
h
ese
ad
v
an
ce
s
a
r
e
s
u
m
m
ar
is
ed
i
n
ex
ten
s
iv
e
r
ev
iew
s
tu
d
ies,
wh
ich
o
u
tlin
e
th
e
p
r
esen
t
ch
allen
g
es
an
d
r
esear
ch
av
en
u
es
f
o
r
SLR,
h
ig
h
lig
h
tin
g
th
e
n
ec
ess
ity
o
f
u
n
iv
er
s
al,
ef
f
ec
tiv
e
,
an
d
m
u
lti
m
o
d
al
SLR s
y
s
tem
s
[
3
5
]
.
T
h
e
o
r
g
an
izatio
n
o
f
th
is
d
o
c
u
m
en
t
is
as
f
o
llo
ws.
Sectio
n
1
o
u
tlin
es
r
elev
a
n
t
s
tu
d
ies
in
th
e
f
ield
o
f
s
ig
n
lan
g
u
a
g
e
r
ec
o
g
n
itio
n
.
S
ec
tio
n
2
in
clu
d
es
p
r
o
p
o
s
ed
m
eth
o
d
o
l
o
g
y
th
at
e
x
p
lain
s
d
a
taset,
p
r
ep
r
o
ce
s
s
in
g
,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
m
o
d
el
tr
ain
in
g
,
an
d
s
ig
n
r
ec
o
g
n
itio
n
.
S
ec
tio
n
3
in
clu
d
e
p
er
f
o
r
m
an
ce
m
etr
ics,
ac
cu
r
ac
y
,
lo
s
s
,
an
d
co
m
p
a
r
ativ
e
an
aly
s
is
.
Fin
ally
,
s
ec
tio
n
4
o
f
th
is
d
o
c
u
m
en
t c
o
n
clu
d
es with
a
s
u
m
m
ar
y
.
T
h
e
f
o
llo
win
g
ar
e
th
r
ee
s
tr
en
g
th
en
e
d
co
n
tr
ib
u
tio
n
p
o
in
ts
th
at
clea
r
ly
e
m
p
h
asize
th
e
n
o
v
elty
a
n
d
s
ig
n
if
ican
ce
o
f
o
u
r
p
r
o
p
o
s
ed
m
o
d
el:
i)
Un
if
ied
m
u
ltil
in
g
u
al
r
ec
o
g
n
itio
n
with
s
ig
n
er
-
i
n
d
ep
e
n
d
e
n
t
ev
alu
atio
n
:
th
e
p
r
o
p
o
s
ed
R
NN
-
L
STM
f
r
am
ewo
r
k
is
o
n
e
o
f
th
e
f
ew
s
y
s
tem
s
wh
ich
r
ec
o
g
n
is
e
C
h
in
ese,
Am
er
ican
,
a
n
d
I
n
d
ian
s
ig
n
lan
g
u
ag
es
in
clu
d
in
g
o
n
e
-
h
a
n
d
ed
an
d
tw
o
-
h
an
d
ed
g
estu
r
es
in
a
s
in
g
le
ar
ch
itectu
r
e.
W
e
ev
alu
ate
th
e
m
o
d
el
u
n
d
er
a
s
tr
ict
s
ig
n
er
-
in
d
ep
en
d
e
n
t
p
r
o
t
o
co
l,
s
h
o
win
g
s
tr
o
n
g
g
en
er
al
is
atio
n
to
u
n
s
ee
n
u
s
er
s
,
an
d
ad
d
r
ess
in
g
a
m
ajo
r
lim
itatio
n
in
e
x
is
tin
g
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
s
tu
d
i
es,
in
co
m
p
ar
is
o
n
to
m
an
y
p
r
e
v
io
u
s
wo
r
k
s
.
ii)
E
f
f
ec
tiv
e
tem
p
o
r
al
m
o
d
elin
g
u
s
in
g
Me
d
iaPip
e
lan
d
m
ar
k
s
a
n
d
L
STM
:
th
e
p
r
o
p
o
s
ed
m
o
d
e
l
co
m
b
in
es
th
e
h
an
d
lan
d
m
ar
k
ex
t
r
ac
tio
n
f
r
o
m
Me
d
iaPip
e
with
th
e
d
ee
p
tem
p
o
r
al
m
o
d
ellin
g
b
ased
o
n
s
tack
ed
L
STM
lay
er
s
.
T
h
is
ap
p
r
o
ac
h
ca
n
ca
p
tu
r
e
s
p
atial
h
an
d
co
n
f
ig
u
r
atio
n
s
an
d
tem
p
o
r
al
g
estu
r
es
d
y
n
am
ics
m
o
r
e
ef
f
ec
tiv
ely
th
a
n
th
e
f
r
am
e
-
b
ased
C
NN
m
o
d
els.
Su
ch
a
d
esig
n
r
esu
lts
in
co
n
s
is
ten
tly
b
etter
p
er
f
o
r
m
an
ce
o
n
all
th
r
ee
s
ig
n
lan
g
u
a
g
es.
No
v
elty
lies
in
th
e
co
m
b
in
at
io
n
o
f
lig
h
tweig
h
t
lan
d
m
ar
k
f
ea
tu
r
es
an
d
s
eq
u
en
ce
lear
n
in
g
f
o
r
r
ea
l
-
tim
e
m
u
ltil
in
g
u
al
r
ec
o
g
n
itio
n
.
iii)
Hig
h
ac
cu
r
ac
y
with
s
tatis
tic
al
r
o
b
u
s
tn
ess
an
d
o
v
er
f
itti
n
g
co
n
tr
o
l
:
th
e
m
o
d
el
attain
s
s
tate
-
of
-
th
e
-
ar
t
ac
cu
r
ac
y
a
b
o
v
e
9
9
.
5
%
o
n
C
h
in
ese,
Am
er
ican
,
a
n
d
I
n
d
ian
s
ig
n
lan
g
u
a
g
es
,
an
d
r
ep
o
r
ts
m
ea
n
ac
cu
r
ac
y
with
s
tan
d
ar
d
d
ev
iatio
n
s
o
v
er
m
u
ltip
le
ex
p
er
im
en
tal
r
u
n
s
.
T
h
e
co
n
tr
ib
u
tio
n
is
f
u
r
th
er
en
h
an
ce
d
b
y
th
e
in
clu
s
io
n
o
f
ea
r
ly
s
to
p
p
i
n
g
,
c
o
n
tr
o
lled
tr
ain
in
g
p
r
o
ce
d
u
r
es
an
d
s
tatis
tical
an
aly
s
i
s
,
wh
ich
s
h
o
w
n
o
t
o
n
ly
h
ig
h
p
e
r
f
o
r
m
an
ce
,
b
u
t
also
s
tab
ilit
y
,
r
ep
r
o
d
u
cib
ilit
y
an
d
r
esis
tan
ce
to
o
v
er
f
itti
n
g
,
o
f
ten
lack
in
g
in
co
m
p
ar
ab
le
s
tu
d
ies.
2.
M
E
T
H
O
D
T
h
e
r
ea
l
-
tim
e
m
u
ltil
in
g
u
al
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
s
y
s
tem
was
d
ev
el
o
p
ed
th
r
o
u
g
h
a
co
m
p
r
eh
e
n
s
iv
e
ap
p
r
o
ac
h
.
I
t
b
eg
in
s
with
d
ata
co
llectio
n
an
d
co
n
ti
n
u
es
with
p
r
e
-
p
r
o
c
ess
in
g
u
s
in
g
h
an
d
tr
ac
k
in
g
p
o
wer
ed
b
y
Me
d
ia
p
ip
e.
Fin
ally
,
R
NN
ar
e
em
p
l
o
y
ed
to
r
ec
o
g
n
ize
h
an
d
g
est
u
r
es,
as
illu
s
tr
ated
in
Fig
u
r
e
3
.
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
.
3
,
J
u
n
e
20
26
:
2
4
9
4
-
2
5
1
0
2498
Fig
u
r
e
3
.
Pro
p
o
s
ed
m
o
d
el
o
f
m
u
ltil
in
g
u
al
s
ig
n
lan
g
u
ag
e
r
ec
o
g
n
itio
n
2
.
1
.
Da
t
a
s
et
As
s
h
o
wn
in
T
a
b
le
1
,
th
e
d
ataset
u
s
ed
to
e
v
alu
ate
th
e
p
r
o
p
o
s
ed
m
u
ltil
in
g
u
al
s
ig
n
lan
g
u
a
g
e
r
ec
o
g
n
itio
n
s
y
s
tem
was
s
elf
-
co
llected
an
d
co
n
s
is
ts
o
f
is
o
lated
alp
h
ab
et
g
estu
r
es
f
r
o
m
C
h
in
ese
s
ig
n
lan
g
u
ag
e
,
Am
er
ican
s
ig
n
lan
g
u
ag
e,
a
n
d
I
n
d
ian
s
ig
n
la
n
g
u
a
g
e
.
A
to
tal
o
f
7
8
alp
h
ab
et
class
es
(
A
–
Z
f
o
r
ea
ch
lan
g
u
ag
e
)
wer
e
r
ec
o
r
d
ed
u
s
in
g
a
s
tan
d
ar
d
web
ca
m
er
a
u
n
d
er
v
ar
y
in
g
b
ac
k
g
r
o
u
n
d
c
o
n
d
itio
n
s
.
Data
wer
e
co
llected
f
r
o
m
m
u
ltip
le
p
ar
ticip
an
ts
s
p
an
n
in
g
d
if
f
er
en
t a
g
e
g
r
o
u
p
s
,
e
n
s
u
r
in
g
in
ter
-
u
s
er
v
a
r
iab
ilit
y
in
h
an
d
s
h
ap
e,
p
o
s
tu
r
e,
a
n
d
ex
ec
u
tio
n
s
ty
le.
Fo
r
ea
ch
alp
h
ab
et
class
,
3
0
g
estu
r
e
s
eq
u
e
n
ce
s
wer
e
r
ec
o
r
d
ed
p
er
p
ar
ti
cip
an
t,
wh
er
e
ea
ch
s
eq
u
en
ce
co
n
tain
s
3
0
f
r
am
es
an
d
r
esu
ltin
g
in
9
0
0
s
am
p
l
es
p
er
alp
h
ab
et.
Ov
er
all,
th
e
d
ataset
co
m
p
r
is
es
7
0
,
2
0
0
s
am
p
les p
er
la
n
g
u
a
g
e,
y
ield
in
g
a
to
tal
o
f
2
1
0
,
6
0
0
g
estu
r
e
s
am
p
les ac
r
o
s
s
all
th
r
ee
s
ig
n
lan
g
u
ag
es.
T
ab
le
1
.
Data
s
et
s
u
m
m
ar
y
ta
b
le
D
a
t
a
s
e
t
a
t
t
r
i
b
u
t
e
C
h
i
n
e
se
si
g
n
l
a
n
g
u
a
g
e
A
meric
a
n
si
g
n
l
a
n
g
u
a
g
e
I
n
d
i
a
n
s
i
g
n
l
a
n
g
u
a
g
e
N
u
mb
e
r
o
f
p
a
r
t
i
c
i
p
a
n
t
s
10
10
10
A
l
p
h
a
b
e
t
c
l
a
ss
e
s
2
6
(
A
-
Z)
2
6
(
A
-
Z)
2
6
(
A
-
Z)
S
a
mp
l
e
s
p
e
r
c
l
a
ss
p
e
r
p
a
r
t
i
c
i
p
a
n
t
3
0
se
q
u
e
n
c
e
s
×
30
f
r
a
mes
3
0
se
q
u
e
n
c
e
s
×
3
0
f
r
a
mes
3
0
se
q
u
e
n
c
e
s
×
3
0
f
r
a
mes
S
a
mp
l
e
s
p
e
r
c
l
a
ss (
t
o
t
a
l
)
9
0
0
9
0
0
9
0
0
To
t
a
l
sam
p
l
e
s
p
e
r
l
a
n
g
u
a
g
e
7
0
,
2
0
0
7
0
,
2
0
0
7
0
,
2
0
0
D
a
t
a
r
e
p
r
e
se
n
t
a
t
i
o
n
M
e
d
i
a
p
i
p
e
h
a
n
d
l
a
n
d
m
a
r
k
s
(
N
u
mP
y
)
M
e
d
i
a
p
i
p
e
h
a
n
d
l
a
n
d
m
a
r
k
s
(
N
u
mP
y
)
M
e
d
i
a
p
i
p
e
h
a
n
d
l
a
n
d
m
a
r
k
s
(
N
u
mP
y
)
Tr
a
i
n
/
v
a
l
i
d
a
t
i
o
n
/
t
e
st
sp
l
i
t
7
0
%/
1
5
%
/
1
5
%
7
0
%/
1
5
%
/
1
5
%
7
0
%/
1
5
%
/
1
5
%
S
i
g
n
e
r
o
v
e
r
l
a
p
b
e
t
w
e
e
n
t
r
a
i
n
a
n
d
t
e
s
t
N
o
(
si
g
n
e
r
-
i
n
d
e
p
e
n
d
e
n
t
s
p
l
i
t
)
N
o
(
si
g
n
e
r
-
i
n
d
e
p
e
n
d
e
n
t
s
p
l
i
t
)
N
o
(
si
g
n
e
r
-
i
n
d
e
p
e
n
d
e
n
t
s
p
l
i
t
)
T
o
m
itig
ate
id
en
tity
b
ias,
th
e
d
ataset
was
s
p
lit
s
u
ch
th
at
s
ig
n
er
s
in
th
e
tr
ain
in
g
s
et
d
o
n
o
t
ap
p
ea
r
in
th
e
test
s
et,
allo
win
g
f
o
r
s
ig
n
e
r
-
in
d
ep
e
n
d
en
t
a
n
d
f
air
ev
alu
at
io
n
o
f
g
en
er
alis
atio
n
.
T
h
e
ex
tr
ac
ted
s
eq
u
en
ce
s
o
f
h
an
d
lan
d
m
a
r
k
s
wer
e
d
iv
i
d
ed
in
to
tr
ain
in
g
,
test
in
g
,
a
n
d
v
ali
d
atio
n
s
u
b
s
ets
with
a
f
ix
ed
s
p
lit
r
atio
.
T
h
is
s
tr
ict
d
esig
n
o
f
th
e
d
ata
s
et
h
elp
s
p
r
ev
en
t
u
s
er
-
s
p
ec
if
ic
o
v
e
r
f
itti
n
g
an
d
allo
ws
a
f
air
co
m
p
ar
is
o
n
o
f
th
e
p
r
o
p
o
s
ed
R
NN
-
L
STM
m
o
d
el
ac
r
o
s
s
m
u
ltip
le
s
ig
n
lan
g
u
ag
es.
I
t
u
s
es
Op
en
C
V
with
f
r
am
es
to
ca
p
tu
r
e
th
e
h
an
d
g
estu
r
es
th
r
o
u
g
h
ca
m
er
a
an
d
s
to
r
ed
in
th
e
s
h
ap
es
ar
r
ay
in
PNG
f
o
r
m
at
with
cv
2
.
r
ec
tan
g
le(
f
r
am
e,
(
0
,
4
0
)
,
(
3
0
0
,
4
0
0
)
,
(
2
5
5
,
2
5
5
,
2
5
5
)
,
2
)
h
av
in
g
a
n
y
c
o
lo
r
b
ac
k
g
r
o
u
n
d
u
s
in
g
“c
v
2
.
r
ec
tan
g
le(
im
ag
e,
s
tar
t_
p
o
i
n
t,
en
d
_
p
o
in
t,
co
lo
r
,
th
ic
k
n
ess
)
”.
Min
im
u
m
9
0
im
a
g
es
o
f
s
ize
3
0
0
×
3
6
0
p
i
x
els
with
24
-
b
it
d
ep
th
with
d
if
f
er
en
t
p
o
s
tu
r
e
s
f
r
o
m
1
0
d
if
f
e
r
en
t
p
eo
p
le
was
ca
p
tu
r
ed
f
o
r
ea
ch
s
ig
n
alp
h
ab
et
f
r
o
m
A
–
Z
in
th
r
ee
d
if
f
er
e
n
t
s
ig
n
lan
g
u
ag
es.
T
h
e
s
am
p
les
o
f
ca
p
tu
r
ed
h
an
d
g
estu
r
es
im
ag
es
f
o
r
d
ataset
cr
ea
tio
n
ar
e
s
h
o
wn
in
Fig
u
r
e
4.
Fig
u
r
e
4
(
a
)
s
h
o
ws
r
ep
r
esen
tativ
e
s
am
p
les
o
f
C
h
in
ese
s
ig
n
lan
g
u
ag
e
al
p
h
ab
ets
ca
p
t
u
r
ed
f
r
o
m
m
u
ltip
le
u
s
er
s
,
d
em
o
n
s
tr
atin
g
co
m
m
o
n
v
ar
iatio
n
s
ac
r
o
s
s
in
d
iv
id
u
als.
Fig
u
r
e
4
(
b
)
illu
s
tr
ates
s
am
p
le
Am
er
ican
s
ig
n
lan
g
u
a
g
e
alp
h
ab
et
g
estu
r
es,
in
clu
d
in
g
d
if
f
er
en
ce
s
in
f
in
g
er
c
o
n
f
ig
u
r
atio
n
an
d
m
o
v
em
en
t.
Fig
u
r
e
4
(
c
)
p
r
esen
ts
s
am
p
le
I
n
d
ian
s
ig
n
l
an
g
u
ag
e
al
p
h
ab
ets,
h
ig
h
lig
h
ti
n
g
v
ar
iatio
n
s
in
h
an
d
p
o
s
tu
r
e
an
d
o
r
ie
n
tatio
n
th
at
af
f
ec
t v
is
u
al
ap
p
ea
r
an
ce
.
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
ltil
in
g
u
a
l sig
n
s
r
ec
o
g
n
itio
n
u
s
in
g
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Th
o
u
s
ee
f U
lla
K
h
a
n
)
2499
A
B
C
…
Z
A
B
C
…
Z
(
a)
(
b
)
A
B
C
…
Z
(
c)
Fig
u
r
e
4
.
C
ap
tu
r
e
d
im
ag
es o
f
h
an
d
g
estu
r
es u
s
ed
f
o
r
d
ataset
cr
ea
tio
n
o
f
(
a)
C
h
in
ese
s
ig
n
la
n
g
u
ag
e
alp
h
ab
ets
g
estu
r
es
,
(
b
)
Am
er
ica
n
s
ig
n
la
n
g
u
ag
e
alp
h
ab
ets
g
estu
r
es
,
a
n
d
(
c)
I
n
d
ian
s
ig
n
lan
g
u
ag
e
alp
h
ab
ets
g
estu
r
es
2
.
2
.
P
re
-
pro
ce
s
s
ing
us
ing
M
edia
P
ipe
f
o
r
ha
nd
la
nd
m
a
rks
det
ec
t
io
n
Me
d
iaPip
e
is
u
s
ed
f
o
r
h
a
n
d
s
k
ey
p
o
i
n
ts
tr
ac
k
in
g
as
s
h
o
wn
i
n
Fig
u
r
e
5
,
with
m
o
d
el
co
m
p
l
ex
ity
o
f
0
an
d
a
m
in
im
u
m
d
etec
tio
n
a
n
d
tr
ac
k
in
g
co
n
f
id
en
ce
o
f
0
.
5
.
Ac
tio
n
will
b
e
tak
en
o
n
th
e
r
an
g
e
o
f
s
eq
u
en
ce
with
f
r
am
e
n
u
m
b
er
in
g
o
f
s
eq
u
en
c
e
len
g
th
o
f
9
0
.
Op
e
n
C
V
in
ter
p
r
etes
an
im
ag
e
as
B
G
R
f
o
r
m
at
b
y
d
ef
a
u
lt;
b
u
t
wh
ile
p
r
o
ce
s
s
in
g
an
im
ag
e,
it
will
b
e
co
n
v
er
ted
to
R
GB
f
o
r
m
at
as
cv
tC
o
lo
r
(
im
ag
e,
cv
2
.
C
OL
OR
_
B
G
R
2
R
G
B
)
.
T
h
e
d
if
f
er
en
t
c
o
lo
u
r
f
o
r
m
ats
o
f
s
ig
n
lan
g
u
ag
e
im
ag
e
is
u
s
ed
in
r
ec
o
g
n
itio
n
o
f
s
ig
n
lan
g
u
ag
e
as seen
in
Fig
u
r
e
6
.
I
n
th
is
m
o
d
el,
we
h
av
e
u
s
ed
R
GB
an
d
B
GR
f
o
r
m
ats f
o
r
th
e
i
m
ag
e
p
r
o
ce
s
s
in
g
.
Fig
u
r
e
5
.
Me
d
iaPip
e
'
s
p
o
r
tr
ay
al
o
f
h
an
d
lan
d
m
ar
k
s
Fig
u
r
e
6
.
Dif
f
e
r
en
t f
o
r
m
ats f
o
r
im
ag
e
p
r
o
ce
s
s
in
g
Fra
m
es
ar
e
cr
ea
ted
with
Op
en
C
V
as
s
ee
n
in
Fig
u
r
e
7
,
b
y
r
ea
d
in
g
PNG
f
o
r
m
atted
i
m
ag
es,
an
d
Me
d
iaPip
e
d
etec
tio
n
will
b
e
d
o
n
e
o
n
f
r
a
m
es
with
h
an
d
s
g
estu
r
es.
Key
p
o
in
ts
a
r
e
e
x
tr
ac
t
ed
an
d
s
to
r
e
d
in
th
e
n
u
m
p
y
ar
r
ay
.
Fig
u
r
e
7
s
h
o
ws
th
e
ex
tr
ac
tio
n
o
f
k
ey
p
o
in
ts
th
r
o
u
g
h
h
a
n
d
lan
d
m
a
r
k
d
etec
tio
n
f
r
o
m
th
e
ca
p
tu
r
ed
im
ag
es u
s
in
g
Me
d
iaPip
e
f
o
r
t
h
e
s
ig
n
alp
h
ab
ets in
C
h
in
ese
s
ig
n
lan
g
u
a
g
e
,
Am
er
ican
s
ig
n
l
an
g
u
ag
e
,
an
d
I
n
d
ian
s
ig
n
lan
g
u
ag
e
r
esp
ec
tiv
ely
.
Fig
u
r
e
7
(
a)
s
h
o
ws
th
e
d
etec
ted
h
an
d
la
n
d
m
ar
k
s
f
o
r
C
h
i
n
ese
s
ig
n
lan
g
u
ag
e
g
estu
r
es,
Fig
u
r
e
7
(
b
)
p
r
esen
t
s
lan
d
m
ar
k
ex
tr
ac
tio
n
f
o
r
Am
er
ican
s
ig
n
lan
g
u
ag
e
g
estu
r
es,
an
d
Fig
u
r
e
7
(
c)
illu
s
tr
ates
th
e
d
etec
ted
lan
d
m
ar
k
s
f
o
r
I
n
d
ian
s
ig
n
lan
g
u
ag
e
g
estu
r
es
,
clea
r
ly
i
n
d
icatin
g
k
ey
j
o
in
t
p
o
s
itio
n
s
u
s
ed
f
o
r
f
ea
tu
r
e
e
x
tr
ac
tio
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
.
3
,
J
u
n
e
20
26
:
2
4
9
4
-
2
5
1
0
2500
A
B
C
…
Z
(
a)
A
B
C
…
Z
(
b
)
A
B
C
…
Z
(
c)
Fig
u
r
e
7
.
Han
d
lan
d
m
ar
k
d
ete
ctio
n
p
r
o
ce
s
s
u
s
in
g
Me
d
iaPip
e
o
n
d
if
f
er
en
t sig
n
lan
g
u
a
g
es
o
f
(
a)
C
h
in
ese
s
ig
n
lan
g
u
ag
e
g
estu
r
es,
(
b
)
Am
er
ic
an
s
ig
n
lan
g
u
ag
e
g
estu
r
es,
a
n
d
(
c)
I
n
d
ian
s
ig
n
la
n
g
u
a
g
e
g
estu
r
es
2
.
3
.
F
e
a
t
ure
ex
t
r
a
ct
io
n
Ker
as
is
u
s
ed
f
o
r
m
o
d
el
an
d
lay
er
s
.
I
n
itially
we
lo
ad
n
u
m
p
y
ar
r
ay
in
th
e
d
ata
p
ath
an
d
a
p
p
en
d
th
at
with
win
d
o
w,
s
eq
u
e
n
ce
,
an
d
lab
els,
wh
o
s
e
ar
ch
itectu
r
e
s
ee
n
in
Fig
u
r
e
8.
X
-
c
o
o
r
d
in
at
e
is
r
ep
r
esen
ted
as
n
u
m
p
y
a
r
r
ay
an
d
Y
-
co
o
r
d
in
a
te
is
r
ep
r
esen
ted
as
ca
teg
o
r
izin
g
th
e
lab
els.
T
h
en
we
ca
r
r
y
o
u
t
s
p
litt
in
g
o
f
x
_
T
r
ain
,
y
_
T
r
ain
,
an
d
x
_
T
est,
y
_
T
est
with
r
esp
ec
t
t
o
x
a
n
d
y
co
o
r
d
i
n
ates.
No
w
Ker
a
s
m
o
d
el
is
u
s
ed
to
im
p
lem
en
t th
r
ee
la
y
er
s
o
f
L
ST
M
alg
o
r
ith
m
with
R
e
LU
ac
tiv
atio
n
an
d
6
4
,
1
2
8
,
an
d
6
4
r
esp
ec
tiv
ely
.
Fig
u
r
e
8
.
Ar
c
h
itectu
r
e
o
f
Ker
a
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
Mu
ltil
in
g
u
a
l sig
n
s
r
ec
o
g
n
itio
n
u
s
in
g
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Th
o
u
s
ee
f U
lla
K
h
a
n
)
2501
Nex
t
Ker
as
m
o
d
el
is
u
s
ed
to
im
p
lem
en
t
two
lay
er
s
o
f
d
e
n
s
e
alg
o
r
ith
m
with
R
e
LU
an
d
o
n
e
la
y
er
with
s
o
f
tm
ax
ac
tiv
atio
n
.
So
f
t
m
ax
co
n
v
e
r
ts
v
ec
to
r
v
alu
es
to
p
r
o
b
a
b
ilit
ies.
I
ts
d
er
iv
ativ
e
is
m
o
n
o
to
n
ic,
as
is
th
e
R
eL
U
f
u
n
ctio
n
its
elf
,
Fig
u
r
e
9
illu
s
tr
ates
th
is
.
T
h
e
f
u
n
ctio
n
r
etu
r
n
s
0
if
it
g
ets
an
y
n
eg
ativ
e
in
p
u
t,
b
u
t
r
etu
r
n
s
a
p
o
s
itiv
e
n
u
m
b
er
if
it
r
ec
ei
v
e
s
an
y
p
o
s
itiv
e
v
alu
e.
C
o
n
s
eq
u
en
tly
,
th
e
o
u
tp
u
t'
s
r
an
g
e
is
0
t
o
in
f
in
ite.
T
h
er
e
is
a
m
ath
em
atica
l
f
o
r
m
u
la
as
i
n
(
1
)
th
at
r
ep
r
esen
ts
R
eL
U.
Nex
t
Ker
as
m
o
d
el
f
it
with
x
_
T
r
ain
,
y
_
T
r
ai
n
.
Fin
ally
,
m
o
d
el
will
b
e
s
av
ed
as
J
SO
N
f
ile
with
th
e
n
am
e
m
o
d
el.
h
5
,
wh
ich
s
to
r
e
r
ec
o
r
d
s
o
f
m
o
d
el,
lay
e
r
s
,
weig
h
ts
an
d
tr
ac
k
a
b
les.
(
)
=
(
0
,
)
(
1
)
Fig
u
r
e
9
.
R
eL
U
f
u
n
ctio
n
ality
2
.
4
.
M
o
del t
ra
ini
ng
T
r
ain
in
g
o
f
m
o
d
el
co
n
tain
s
K
er
as,
T
en
s
o
r
B
o
ar
d
,
la
b
ellin
g
,
t
r
ain
-
test
s
p
litt
in
g
,
an
d
L
STM
an
d
Den
s
e
alg
o
r
ith
m
s
with
2
0
0
e
p
o
ch
s
.
B
y
u
s
in
g
an
o
p
tim
izatio
n
tec
h
n
iq
u
e
lik
e
g
r
ad
ien
t
d
escen
t
i
n
co
m
b
in
atio
n
with
b
ac
k
p
r
o
p
ag
atio
n
o
v
e
r
tim
e
to
d
eter
m
in
e
th
e
g
r
ad
ie
n
ts
n
ee
d
ed
f
o
r
th
e
o
p
tim
izatio
n
p
r
o
ce
s
s
,
an
R
NN
em
p
lo
y
in
g
L
STM
u
n
its
,
as
illu
s
tr
ated
in
Fig
u
r
e
1
0
,
ca
n
b
e
tr
ain
ed
u
n
d
er
s
u
p
er
v
is
io
n
u
s
in
g
a
s
et
o
f
t
r
ain
in
g
s
eq
u
en
ce
s
.
T
h
is
will
al
ter
ea
c
h
weig
h
t
o
f
th
e
L
STM
n
etwo
r
k
p
r
o
p
o
r
tio
n
ately
to
th
e
d
e
r
iv
ativ
e
o
f
th
e
er
r
o
r
(
at
th
e
L
STM
n
etwo
r
k
'
s
o
u
tp
u
t la
y
er
)
with
r
esp
ec
t to
t
h
e
co
r
r
esp
o
n
d
in
g
weig
h
t.
Fig
u
r
e
1
0
.
Ar
ch
itectu
r
e
o
f
L
S
T
M
W
e
em
p
lo
y
L
STM
i
n
th
e
s
u
g
g
ested
s
y
s
tem
,
wh
ich
h
as
th
r
e
e
g
ates:
in
p
u
t,
o
u
t
p
u
t,
a
n
d
f
o
r
g
et
g
ates.
L
STM
g
ates
ar
e
s
ig
m
o
id
ac
tiv
atio
n
f
u
n
ctio
n
s
th
at
p
r
o
d
u
ce
a
v
alu
e
b
etwe
en
0
an
d
1
,
wh
i
ch
is
o
f
ten
eith
er
0
o
r
1
.
Fig
u
r
e
1
1
illu
s
tr
ates h
o
w
th
e
s
ig
m
o
id
f
u
n
ctio
n
r
ep
r
esen
ts
L
STM
.
T
h
e
L
STM
g
ates a
r
e
illu
s
tr
ated
b
y
(
2
)
.
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
.
3
,
J
u
n
e
20
26
:
2
4
9
4
-
2
5
1
0
2502
=
(
+
ℎ
−
1
+
)
=
(
+
ℎ
−
1
+
)
=
(
+
ℎ
−
1
+
)
˜
=
(
+
ℎ
−
1
+
)
=
⊙
−
1
+
⊙
˜
ℎ
=
⊙
ℎ
(
)
(
2
)
W
h
er
e
s
tan
d
s
f
o
r
in
p
u
t
g
ate
,
s
tan
d
s
f
o
r
f
o
r
g
et
g
ate
,
s
tan
d
s
f
o
r
o
u
tp
u
t
g
ate
,
d
escr
ib
es
s
ig
m
o
id
f
u
n
ctio
n
,
weig
h
t
f
o
r
th
e
r
esp
ec
tiv
e
g
ate
(
)
n
eu
r
o
n
s
,
ℎ
−
1
o
u
tp
u
t
o
f
th
e
p
r
ev
io
u
s
L
ST
M
b
lo
ck
(
at
tim
estam
p
−
1
)
,
in
p
u
t
at
cu
r
r
en
t
tim
estam
p
,
b
iases
f
o
r
th
e
r
esp
ec
tiv
e
g
ates
(
)
,
ce
ll
s
tat
e
(
m
em
o
r
y
)
at
tim
estam
p
(
)
,
an
d
˜
ca
n
d
id
ate
f
o
r
ce
ll st
ate
at
tim
estam
p
(
)
.
E
r
r
o
r
g
r
a
d
ien
ts
d
is
ap
p
ea
r
ex
p
o
n
en
tially
with
th
e
le
n
g
th
o
f
tim
e
b
etwe
en
s
ig
n
i
f
ican
t
e
v
e
n
ts
,
wh
ich
ar
e
a
d
r
awb
ac
k
o
f
g
r
ad
ien
t
d
e
s
ce
n
t
f
o
r
co
n
v
en
tio
n
al
R
NNs.
B
u
t
with
L
STM
u
n
its
,
th
e
e
r
r
o
r
s
tay
s
in
th
e
ce
ll
o
f
th
e
L
STM
u
n
it
ev
en
af
te
r
er
r
o
r
v
alu
es
ar
e
b
ac
k
-
p
r
o
p
ag
a
ted
f
r
o
m
t
h
e
o
u
t
p
u
t
lay
er
.
Un
til
th
e
L
STM
u
n
it
lear
n
s
to
s
h
u
t o
f
f
th
e
v
al
u
e,
th
i
s
“
er
r
o
r
ca
r
o
u
s
el
”
k
ee
p
s
f
ee
d
in
g
er
r
o
r
s
b
ac
k
t
o
ea
ch
o
f
its
g
ates.
T
h
e
s
o
f
tm
ax
f
u
n
ctio
n
is
u
s
ed
in
th
e
f
in
al
lay
er
o
f
a
n
eu
r
al
n
etwo
r
k
m
o
d
el
f
o
r
class
if
icatio
n
task
s
,
wh
ich
tak
es th
e
ex
p
o
n
en
tial o
f
ea
ch
o
u
tp
u
t a
n
d
n
o
r
m
alize
s
th
em
b
y
d
i
v
id
in
g
with
t
o
tal
o
f
al
l th
e
ex
p
o
n
en
tials
.
T
h
is
p
r
o
ce
s
s
tu
r
n
s
r
aw
o
u
tp
u
t
s
co
r
es
in
to
p
r
o
b
ab
ilit
ies.
T
h
is
p
r
o
ce
d
u
r
e
g
u
ar
a
n
tees
th
at
th
e
o
u
tp
u
t
v
alu
es
m
ay
b
e
in
ter
p
r
ete
d
as
p
r
o
b
ab
ilit
ies
as
th
ey
f
all
b
etwe
en
0
a
n
d
1
a
n
d
ad
d
u
p
t
o
1
.
T
h
e
n
etwo
r
k
'
s
p
r
ed
ictio
n
s
m
ay
b
e
in
ter
p
r
eted
p
r
o
b
a
b
ilis
tically
u
s
in
g
th
e
s
o
f
tm
a
x
f
u
n
ctio
n
(
3
)
.
(
⃗
)
=
∑
=
1
(
3
)
W
h
er
e
is
s
o
f
tm
ax
,
⃗
is
in
p
u
t
v
ec
to
r
,
is
s
tan
d
ar
d
e
x
p
o
n
en
tial
f
u
n
ctio
n
f
o
r
in
p
u
t
v
ec
to
r
,
is
n
u
m
b
er
o
f
class
es
in
th
e
m
u
lti
-
clas
s
clas
s
if
ier
,
is
s
tan
d
ar
d
ex
p
o
n
en
tial
f
u
n
ctio
n
f
o
r
o
u
tp
u
t
v
ec
to
r
,
an
d
is
s
tan
d
ar
d
ex
p
o
n
e
n
tial f
u
n
ctio
n
f
o
r
o
u
t
p
u
t v
ec
to
r
.
Fig
u
r
e
1
1
.
L
STM
s
ig
m
o
id
f
u
n
ctio
n
2
.
5
.
Sig
n
re
co
g
nitio
n
Me
d
iaPip
e
is
u
s
ed
to
p
r
e
d
ict
h
an
d
lan
d
m
ar
k
s
f
r
o
m
a
n
u
m
p
y
ar
r
ay
s
u
p
p
lied
i
n
a
J
SON
f
ile
with
a
th
r
esh
o
ld
o
f
0
.
8
.
T
h
e
p
r
ed
ict
(
)
f
u
n
ctio
n
,
wh
ic
h
y
ield
s
p
r
e
d
ictio
n
s
in
th
e
f
o
r
m
at
s
u
p
p
lied
b
y
t
h
e
n
etwo
r
k
'
s
o
u
tp
u
t
lay
e
r
,
will
b
e
u
s
ed
b
y
th
e
m
o
d
el
t
o
m
ak
e
p
r
e
d
ictio
n
s
b
ased
o
n
th
ese
h
an
d
k
ey
p
o
i
n
ts
.
T
h
e
ar
g
m
ax
(
)
Nu
m
Py
f
u
n
ctio
n
m
u
s
t
b
e
u
s
ed
to
tr
an
s
late
th
e
m
u
lticlas
s
clas
s
if
icatio
n
r
esu
lts
,
wh
ic
h
ar
e
an
ar
r
a
y
o
f
p
r
o
b
a
b
ilit
ies,
in
to
a
s
in
g
le
class
o
u
tp
u
t p
r
e
d
ictio
n
.
Fin
ally
,
ac
cu
r
ac
y
will b
e
ca
lcu
alted
an
d
d
is
p
lay
ed
.
T
h
e
ac
cu
r
ac
y
r
ep
r
esen
tatio
n
o
f
th
e
p
r
o
p
o
s
ed
s
ig
n
lan
g
u
a
g
e
r
ec
o
g
n
itio
n
s
y
s
tem
f
o
r
m
u
l
tip
le
s
ig
n
lan
g
u
ag
es
is
s
h
o
wn
in
Fig
u
r
e
1
2
.
Par
ticu
lar
ly
,
th
e
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
o
b
tain
e
d
f
o
r
C
h
in
e
s
e
s
ig
n
lan
g
u
ag
e
is
s
h
o
wn
in
Fig
u
r
e
1
2
(
a)
,
wh
ic
h
s
h
o
ws
th
e
ef
f
icien
cy
o
f
th
e
m
o
d
el
in
r
ec
o
g
n
is
in
g
o
n
e
-
h
an
d
ed
g
estu
r
e
p
atter
n
s
.
T
h
e
r
esu
lts
o
f
ac
cu
r
ac
y
f
o
r
A
m
er
ican
s
ig
n
lan
g
u
ag
e
a
r
e
g
iv
en
in
Fig
u
r
e
1
2
(
b
)
.
T
h
e
ac
cu
r
ac
y
p
er
f
o
r
m
an
ce
o
f
I
n
d
ian
s
ig
n
lan
g
u
ag
e
is
g
iv
e
n
in
Fig
u
r
e
1
2
(
c)
.
T
h
e
r
esu
lts
s
h
o
w
th
e
r
o
b
u
s
tn
ess
o
f
t
h
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
f
o
r
b
o
th
o
n
e
-
h
a
n
d
an
d
two
-
HGR
s
ce
n
ar
io
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
Mu
ltil
in
g
u
a
l sig
n
s
r
ec
o
g
n
itio
n
u
s
in
g
r
ec
u
r
r
en
t n
eu
r
a
l n
etw
o
r
k
(
Th
o
u
s
ee
f U
lla
K
h
a
n
)
2503
A
B
C
…
Z
(
a)
A
B
C
…
Z
(
b
)
A
B
C
…
Z
(
c
)
Fig
u
r
e
1
2
.
R
ec
o
g
n
itio
n
ac
c
u
r
a
cy
ac
h
iev
ed
b
y
t
h
e
p
r
o
p
o
s
ed
s
y
s
tem
ac
r
o
s
s
d
if
f
er
en
t sig
n
lan
g
u
ag
es
o
f
(
a)
C
h
in
ese
s
ig
n
lan
g
u
a
g
e
alp
h
ab
ets,
(
b
)
Am
e
r
ican
s
ig
n
lan
g
u
ag
e
alp
h
ab
ets
,
an
d
(
c)
I
n
d
ia
n
s
ig
n
lan
g
u
ag
e
al
p
h
ab
ets
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
ab
le
2
s
h
o
ws
th
e
e
x
p
er
im
e
n
ts
wer
e
r
ep
ea
ted
f
i
v
e
tim
es
with
d
if
f
er
en
t
r
a
n
d
o
m
in
itialis
ati
o
n
s
wh
ile
k
ee
p
in
g
th
e
s
am
e
s
ig
n
er
-
in
d
ep
en
d
en
t
tr
ain
/
v
alid
atio
n
/tes
t
s
p
lit
to
ev
alu
ate
th
e
s
tati
s
ti
ca
l
r
o
b
u
s
tn
ess
an
d
s
tab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
m
u
lti
lin
g
u
al
s
ig
n
lan
g
u
a
g
e
r
ec
o
g
n
itio
n
s
y
s
tem
.
T
h
e
m
ea
n
r
ec
o
g
n
i
tio
n
ac
cu
r
ac
y
a
n
d
its
s
tan
d
ar
d
d
ev
iatio
n
f
o
r
ea
ch
s
ig
n
lan
g
u
ag
e
ar
e
s
h
o
wn
in
T
ab
le
2
.
T
h
e
s
m
all
s
tan
d
ar
d
d
ev
iatio
n
v
alu
es
in
d
icate
th
at
th
e
p
r
o
p
o
s
ed
R
NN
-
L
STM
m
o
d
el
h
as
co
n
s
is
ten
t
an
d
r
eliab
le
p
e
r
f
o
r
m
an
ce
ac
r
o
s
s
m
u
ltip
le
r
u
n
s
,
wh
ich
im
p
lies
ef
f
ec
tiv
e
o
v
er
f
i
ttin
g
co
n
tr
o
l a
n
d
r
o
b
u
s
t g
en
er
a
lis
atio
n
ca
p
ab
ilit
y
am
o
n
g
d
if
f
e
r
en
t sig
n
er
s
.
T
ab
le
2
.
Acc
u
r
ac
y
s
tab
ilit
y
ac
r
o
s
s
m
u
ltip
le
ex
p
er
im
e
n
tal
r
u
n
s
S
i
g
n
l
a
n
g
u
a
g
e
M
e
a
n
a
c
c
u
r
a
c
y
(
%)
S
t
a
n
d
a
r
d
d
e
v
i
a
t
i
o
n
(
%)
C
h
i
n
e
se
si
g
n
l
a
n
g
u
a
g
e
9
9
.
5
8
±
0
.
2
1
A
meric
a
n
si
g
n
l
a
n
g
u
a
g
e
9
9
.
6
2
±
0
.
1
8
I
n
d
i
a
n
s
i
g
n
l
a
n
g
u
a
g
e
9
9
.
6
3
±
0
.
1
7
N
o
t
e
:
r
e
s
u
l
t
s
a
r
e
a
v
e
r
a
g
e
d
o
v
e
r
f
i
v
e
i
n
d
e
p
e
n
d
e
n
t
r
u
n
s
u
si
n
g
t
h
e
sam
e
si
g
n
e
r
-
i
n
d
e
p
e
n
d
e
n
t
d
a
t
a
s
e
t
sp
l
i
t
w
i
t
h
d
i
f
f
e
r
e
n
t
r
a
n
d
o
m
i
n
i
t
i
a
l
i
z
a
t
i
o
n
s.
T
h
e
s
tan
d
ar
d
F1
-
s
co
r
e,
wh
i
ch
ca
lcu
lates
th
e
n
u
m
b
er
o
f
tim
es
th
e
m
o
d
el
p
r
o
d
u
ce
d
ac
cu
r
ate
p
r
ed
ictio
n
s
,
h
as
b
ee
n
u
s
ed
to
ass
ess
th
e
m
o
d
el
ac
cu
r
ac
y
p
er
f
o
r
m
an
ce
.
Fals
e
p
o
s
itiv
e
(
FP
)
,
f
alse
n
eg
ativ
e
(
FN)
,
tr
u
e
p
o
s
itiv
e
(
T
P),
an
d
tr
u
e
n
eg
ativ
e
(
T
N)
ar
e
th
e
m
ea
s
u
r
es
th
at
ar
e
em
p
lo
y
ed
.
I
t
u
s
es
(
4
)
to
(
6
)
to
d
eter
m
in
e
th
e
m
o
d
el'
s
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
.
=
+
(
4
)
=
+
(
5
)
1
−
=
+
1
2
(
+
)
(
6
)
Evaluation Warning : The document was created with Spire.PDF for Python.