I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
3
,
J
une
20
25
, pp.
2035
~
2043
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
3
.pp
2035
-
2043
2035
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
S
i
gn
l
a
n
gu
age
r
e
c
ogn
i
t
i
on
an
d
c
l
ass
i
f
i
c
at
i
on
u
si
n
g
b
l
e
n
d
e
d
e
n
se
m
b
l
e
m
ac
h
i
n
e
l
e
ar
n
i
n
g
A
k
as
h
R
aj
an
R
ai
, S
u
j
a
t
a R
aj
e
s
h
K
ad
u
D
e
pa
r
t
m
e
nt
of
I
nf
or
m
a
t
i
on
T
e
c
hnol
ogy,
T
e
r
na
C
ol
l
e
ge
of
E
ngi
ne
e
r
i
ng, U
ni
ve
r
s
i
t
y of
M
um
ba
i
, N
a
vi
M
um
ba
i
, I
ndi
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
M
a
r
3, 2024
R
e
vi
s
e
d
N
ov 25, 2024
A
c
c
e
pt
e
d
J
a
n 27, 2025
An
efficient
sign
language
recognition
system
(SLR)
is
the
most
sign
ificant
for
hearing
-
impaired
people
for
communication
.
The
body
movemen
ts
and
hand
gestures
are
utilized
to
characterize
the
vocabulary
in
dynami
c
sign
language.
The
SLR
is
a
challenging
problem
because
the
comput
ational
model
requires
simultaneous
spatial
-
temporal
modelling
for
a
num
ber
of
sources.
To
overcome
this
problem,
this
researc
h
proposes
the
b
lended
ensemble
machine
learning
(ML)
approaches
for
SLR.
Initially,
the
Indian
sign
language
(ISL)
dataset
is
collected
for
evaluating
the
effective
ness
of
the
model.
Then,
the
pre
-
processing
is
done
by
using
data
augmentation
and
normalization
techniques.
Then,
the
pre
-
processed
data
is
provided
to
the
segmentation
process
which
is
done
by
using
multi
-
threshold
entropy
function.
Then,
VGG
-
16
is
used
for
the
feature
extra
ction
process
to
extract
the
features
and
finally,
classification
is
carried
out
using
ensemble
ML.
An
effectivenes
s
of
the
proposed
method
is
validated
based
on
acc
uracy,
precision,
recall
,
and
F1
-
score,
wherein
it
achie
ves
better
results
of
9
9.57%,
0.92%,
0.95%
,
and
0.99%
as
compared
to
the
existing
works
like
s
upport
vector machine
(SVM) and
convolut
ional n
eural netwo
rk
(CNN).
K
e
y
w
o
r
d
s
:
B
le
nde
d e
ns
e
m
bl
e
ML
D
a
ta
a
ugm
e
nt
a
ti
on
I
S
L
r
e
c
ogni
ti
on
M
ul
ti
-
th
r
e
s
hol
di
ng
VGG
-
16
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
A
ka
s
h R
a
ja
n
R
a
i
D
e
pa
r
tm
e
nt
of
I
nf
or
m
a
ti
on T
e
c
hnol
ogy
,
T
e
r
na
E
ngi
ne
e
r
in
g C
ol
le
ge
,
U
ni
ve
r
s
it
y of
M
um
ba
i
N
a
vi
M
um
ba
i,
M
a
h
a
r
a
s
ht
r
a
, I
ndi
a
E
m
a
il
:
a
ka
s
hr
a
i9
32@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
C
om
m
uni
c
a
ti
on
is
s
ig
ni
f
ic
a
nt
f
or
pe
opl
e
to
di
s
c
ov
e
r
th
e
ir
ne
c
e
s
s
it
ie
s
a
s
w
e
ll
a
s
in
te
r
a
c
ti
ons
w
it
h
ot
he
r
pe
opl
e
.
T
he
r
e
a
r
e
num
be
r
of
de
a
f
a
nd
dum
b
p
e
opl
e
w
ho
m
a
jo
r
ly
de
pe
nd
on
s
ig
n
la
ngua
ge
to
c
om
m
uni
c
a
te
w
it
h
ot
he
r
s
[
1]
,
[
2]
.
T
he
s
ig
n
la
ngua
ge
s
a
r
e
m
e
a
s
ur
e
d
a
s
gr
a
phi
c
a
l
a
nd
non
-
ve
r
ba
l
f
or
m
o
f
c
om
m
uni
c
a
ti
ons
ut
il
iz
e
d
th
r
ough
di
f
f
e
r
e
nt
ly
-
a
bl
e
d
pe
opl
e
t
o
e
xpr
e
s
s
th
e
m
s
e
lv
e
s
or
in
te
r
a
c
t
w
it
h
th
e
ir
s
ur
r
oundings
.
G
lo
ba
ll
y,
th
e
s
ig
n
la
ngua
ge
is
th
e
m
os
t
e
m
e
r
gi
ng
a
s
w
e
ll
a
s
c
ha
ll
e
ngi
ng
ta
s
k.
T
h
e
s
ig
n
la
ngua
ge
e
f
f
ic
ie
nt
ly
he
lp
s
th
e
h
a
r
d
-
of
-
he
a
r
in
g
a
s
w
e
ll
a
s
s
p
e
e
d
-
im
pa
ir
e
d
s
oc
ie
ty
in
a
c
qui
r
in
g
a
c
a
de
m
ic
pr
of
ic
ie
nc
y,
pr
of
e
s
s
io
ns
a
s
w
e
ll
a
s
s
oc
ia
l
r
ig
ht
s
[
3]
,
[
4]
.
T
he
s
ig
n
la
ngua
ge
c
onve
r
ts
th
e
w
or
ds
,
s
e
nt
e
nc
e
s
,
num
be
r
s
a
nd
le
tt
e
r
s
of
na
tu
r
a
l
la
ngua
ge
to
e
na
bl
e
th
e
voc
a
ll
y
de
a
c
ti
va
te
d
pe
opl
e
to
in
te
r
a
c
t
w
it
h
th
e
ot
he
r
pe
opl
e
[
5]
.
I
n
s
ig
n
la
ngua
ge
,
th
e
m
e
a
ni
ng
a
nd
e
xt
r
a
c
ti
on
of
da
ta
is
e
xpr
e
s
s
e
d
th
r
ough
u
s
in
g
ha
nd
ge
s
tu
r
e
s
,
m
ove
m
e
nt
s
of
th
e
body,
f
a
c
ia
l
e
xpr
e
s
s
io
ns
a
s
w
e
ll
a
s
e
m
ot
io
ns
r
a
th
e
r
th
a
n
s
ound,
to
s
e
nd
th
e
m
e
s
s
a
ge
s
.
M
or
e
ove
r
,
s
ig
n
la
ngua
ge
m
in
im
iz
e
s
th
e
c
om
m
uni
c
a
ti
on
ga
p
a
m
ong
de
a
f
a
nd
dum
b
pe
opl
e
,
f
a
c
il
it
a
ti
ng
s
m
oot
h
c
om
m
uni
c
a
ti
on.
T
he
s
ig
n
la
ngua
g
e
s
va
r
y
f
r
om
r
e
gi
on
to
r
e
gi
on,
a
nd
na
ti
on
to
n
a
ti
on
[
6]
,
[
7]
.
T
he
num
be
r
of
r
e
s
e
a
r
c
he
r
s
de
te
r
m
in
e
a
n
e
xc
it
in
g
a
nd
e
xc
lu
s
iv
e
f
or
m
of
c
om
m
uni
c
a
ti
on
in
s
ig
n
la
ngua
ge
ove
r
va
r
io
us
na
ti
ons
.
T
he
m
a
c
hi
ne
le
a
r
ni
ng
(
M
L
)
a
nd
de
e
p
le
a
r
ni
ng
(
D
L
)
te
c
hni
que
s
ha
ve
obt
a
in
e
d
be
tt
e
r
e
nha
nc
e
m
e
nt
c
a
pa
bi
li
ti
e
s
in
s
ig
n
la
ngua
ge
r
e
c
ogni
ti
on
(
S
L
R
)
[
8]
,
[
9]
.
T
he
M
L
or
D
L
te
c
hni
que
s
a
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
3
,
J
une
2025
:
2035
-
2043
2036
im
pl
e
m
e
nt
e
d
f
or
th
e
a
ut
om
a
ti
c
r
e
c
ogni
ti
on
of
s
ig
n
la
ngua
g
e
g
e
s
tu
r
e
s
to
m
in
im
iz
e
c
om
m
uni
c
a
ti
on
ga
p
w
it
h
th
e
s
e
pe
opl
e
.
V
a
r
io
us
r
e
s
e
a
r
c
he
r
s
d
e
s
ig
n
th
e
ne
w
a
ppr
oa
c
he
s
f
or
S
L
R
f
r
om
t
he
a
dva
nt
a
ge
s
of
e
xi
s
ti
ng
a
ppr
oa
c
he
s
to
e
nha
nc
e
th
e
m
ode
l’
s
pe
r
f
or
m
a
nc
e
[
10]
,
[
11]
.
T
he
S
L
R
te
c
hni
que
s
a
r
e
pe
r
f
or
m
e
d
to
e
nha
nc
e
th
e
e
f
f
ic
ie
nc
y
of
th
e
m
ode
l
th
r
ough
m
in
im
iz
in
g
th
e
pr
oc
e
s
s
in
g
ti
m
e
,
de
ve
lo
pi
ng
r
e
li
a
bl
e
da
ta
ba
s
e
s
,
e
na
bl
in
g
qua
li
ty
e
nha
nc
e
m
e
nt
.
A
s
a
n
out
c
om
e
,
a
ut
om
a
ti
c
S
L
R
a
ppr
oa
c
he
s
a
r
e
r
e
qui
r
e
d
to
tr
a
ns
la
te
s
ig
ns
in
to
e
qui
va
le
nt
te
xt
or
s
ound
w
it
hout
a
n
a
s
s
is
ta
nc
e
of
tr
a
ns
la
to
r
s
[
12]
,
[
13]
.
M
o
s
tl
y
a
ppr
oa
c
he
s
a
r
e
di
vi
de
d
in
to
two
ty
pe
s
;
in
it
ia
ll
y,
th
e
a
ppr
oa
c
h
de
pe
nds
on
ha
nd
s
ha
pe
a
s
w
e
ll
a
s
ge
s
tu
r
e
m
ove
m
e
nt
.
T
he
n,
th
e
a
ppr
oa
c
he
s
de
pe
nd
on
th
e
s
e
que
n
c
e
of
im
a
ge
f
or
e
ve
r
y
S
L
R
.
T
he
r
e
s
e
a
r
c
h
e
r
s
ha
v
e
in
tr
oduc
e
d
th
e
S
L
R
te
c
hni
que
s
f
or
va
r
io
us
la
ngua
ge
s
s
u
c
h
a
s
I
ndi
a
n,
C
hi
ne
s
e
,
A
m
e
r
ic
a
n,
a
nd
G
e
r
m
a
n
s
ig
n
la
ngua
ge
.
T
hough
va
r
io
us
im
pr
ove
m
e
nt
s
ha
ve
be
e
n
de
ve
lo
p
e
d
f
or
S
L
R
,
s
ti
ll
s
om
e
m
ode
ls
ha
ve
f
a
il
e
d
to
m
a
na
ge
th
e
r
e
a
l
-
ti
m
e
da
ta
s
e
t
s
a
nd
la
c
k
in
s
om
e
c
a
s
e
s
[
14]
–
[
16]
.
N
ow
a
da
ys
,
not
e
w
or
th
y
pr
ogr
e
s
s
io
n
s
ha
ve
be
e
n
de
ve
lo
pe
d
in
th
e
a
r
e
a
of
S
L
R
,
e
s
pe
c
ia
ll
y
th
r
ough
M
L
a
ppr
oa
c
he
s
.
H
ow
e
ve
r
,
a
tt
a
in
in
g
th
e
be
t
te
r
a
c
c
ur
a
c
y
a
nd
r
obu
s
tn
e
s
s
ove
r
v
a
r
io
us
s
ig
n
la
ngua
ge
da
ta
s
e
t
is
c
h
a
ll
e
ngi
ng.
H
e
nc
e
,
th
e
m
ot
iv
a
ti
on
of
th
is
r
e
s
e
a
r
c
h
i
s
to
s
ol
ve
th
e
a
f
or
e
-
m
e
nt
io
ne
d
pr
obl
e
m
th
r
ough
pr
opos
in
g
th
e
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
ppr
oa
c
he
s
to
im
pr
ove
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
S
L
R
.
C
om
pa
r
e
d
to
pr
e
vi
ous
w
or
ks
li
ke
K
-
ne
a
r
e
s
t
ne
ig
hbor
(
K
N
N
)
,
n
a
ïv
e
B
a
ye
s
(
N
B
)
,
r
a
ndom
f
or
e
s
t
(
R
F
)
,
a
nd
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
(
S
V
M
)
,
th
e
pr
opos
e
d
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
ppr
oa
c
he
s
a
tt
a
in
be
tt
e
r
a
c
c
ur
a
c
y
a
nd
ge
ne
r
a
li
z
a
bi
li
ty
, r
e
pr
e
s
e
nt
in
g t
he
e
nha
nc
e
d e
f
f
e
c
ti
ve
ne
s
s
.
B
a
s
e
d
on
th
e
na
tu
r
e
of
s
ig
n
la
ngu
a
ge
,
th
e
r
e
s
e
a
r
c
he
r
s
ut
il
iz
e
va
r
io
us
a
ppr
oa
c
he
s
f
or
S
L
R
.
V
a
r
io
us
ty
pe
s
of
S
L
R
a
r
e
a
na
ly
z
e
d
in
th
is
s
e
c
ti
on
to
e
li
m
in
a
te
th
e
c
om
m
uni
c
a
ti
on
ga
p.
B
e
c
a
us
e
of
th
e
va
r
yi
ng
na
tu
r
e
of
s
ig
ns
i
n e
ve
r
y s
ig
n
la
ngua
ge
, t
he
s
ig
n r
e
c
ogni
ti
on i
s
c
ha
ll
e
ng
in
g. A
th
ir
a
e
t
a
l
.
[
17
]
in
t
r
oduc
e
d t
he
S
V
M
f
o
r
s
e
lf
-
de
te
r
m
in
in
g
vi
s
io
n
-
ba
s
e
d
S
L
R
a
ppr
oa
c
h.
T
he
s
ugge
s
te
d
a
ppr
oa
c
h
ha
d
th
e
c
a
pa
bi
li
ty
to
r
e
c
ogni
z
e
th
e
s
in
gl
e
a
nd
double
ha
nde
d
of
s
ta
ti
c
a
s
w
e
ll
a
s
dyna
m
ic
ge
s
tu
r
e
s
us
in
g
r
e
a
l
-
ti
m
e
vi
de
o
of
I
ndi
a
n
s
ig
n
la
ngu
a
ge
(
I
S
L
)
.
I
n
th
a
t
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
,
s
ki
n
c
ol
or
s
e
gm
e
nt
a
ti
on
a
p
pr
oa
c
h
w
a
s
ut
il
iz
e
d
f
or
th
e
I
S
L
e
xt
r
a
c
ti
on.
T
he
s
ugge
s
te
d a
ppr
oa
c
h
e
f
f
ic
ie
nt
ly
m
in
im
iz
e
d
a
c
om
put
a
ti
ona
l
s
pe
e
d
to
a
n
e
xc
e
s
s
iv
e
a
m
ount
by
th
e
ut
il
iz
a
ti
on
of
Z
e
r
ni
ke
m
om
e
nt
s
f
r
a
m
e
e
xt
r
a
c
ti
on
a
ppr
oa
c
h.
H
ow
e
ve
r
,
th
e
i
nt
r
oduc
e
d
a
ppr
oa
c
h
w
a
s
pe
r
f
or
m
e
d
onl
y
w
it
h
f
iv
e
s
ta
ti
c
s
ym
bol
s
f
or
S
L
R
.
K
a
to
c
h
e
t
al
.
[
18]
pr
e
s
e
nt
e
d
th
e
S
V
M
a
nd
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
f
or
th
e
c
la
s
s
if
ic
a
ti
on
of
s
ig
n
la
ngua
ge
.
T
he
s
ugge
s
te
d
a
ppr
oa
c
h
ut
il
iz
e
d
th
e
ba
g
of
v
is
ua
l
w
or
ds
(
B
O
V
W
)
to
r
e
c
ogni
z
e
th
e
I
S
L
a
lp
ha
be
ts
(
A
-
Z
)
a
s
w
e
ll
a
s
di
gi
ts
(
0
-
9)
in
r
e
a
l
-
ti
m
e
vi
de
o
s
tr
e
a
m
.
S
pe
e
de
d
up
r
obus
t
f
e
a
tu
r
e
s
(
S
U
R
F
)
w
e
r
e
e
xt
r
a
c
te
d
f
r
om
th
e
hi
s
to
gr
a
m
s
,
a
nd
i
m
a
ge
s
w
e
r
e
pr
oduc
e
d
to
m
a
p
th
e
s
ig
n
w
it
h
c
ons
is
te
nt
la
b
e
ls
.
A
c
ol
la
bor
a
ti
ve
gr
a
phi
c
a
l
u
s
e
r
in
te
r
f
a
c
e
(
G
U
I
)
w
a
s
ge
ne
r
a
te
d
f
or
e
a
s
y
a
c
c
e
s
s
of
S
L
R
.
H
ow
e
ve
r
, t
he
s
ugge
s
te
d a
ppr
oa
c
h ut
il
iz
e
d t
he
l
a
r
ge
c
lu
s
te
r
da
ta
f
or
e
nha
nc
e
t
he
m
ode
l
pe
r
f
or
m
a
nc
e
.
S
ha
r
m
a
a
nd
S
in
gh
[
19]
de
ve
lo
pe
d
th
e
r
obus
t
c
om
put
e
r
-
vi
s
io
n
ba
s
e
d
C
N
N
by
d
e
pt
h
w
is
e
s
e
pa
r
a
bl
e
c
onvolut
io
n
(
D
S
N
)
f
o
r
th
e
r
e
c
ogni
ti
on
of
s
ig
n
la
ngua
ge
.
I
ni
ti
a
l
ly
,
th
e
I
S
L
da
ta
s
e
t
w
a
s
c
r
e
a
te
d
f
r
om
65
us
e
r
s
in
a
n
unc
ont
r
ol
le
d
e
nvi
r
onm
e
nt
.
T
he
n,
in
tr
a
-
c
la
s
s
va
r
ia
nc
e
in
th
e
da
ta
ba
s
e
w
a
s
pe
r
f
or
m
e
d
by
a
ugm
e
nt
a
ti
on
a
ppr
oa
c
h t
o e
nha
nc
e
t
he
ge
ne
r
a
li
z
a
ti
on c
a
pa
bi
li
ty
. T
he
C
N
N
w
a
s
ut
il
iz
e
d f
or
t
he
pr
oc
e
s
s
of
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
a
s
w
e
ll
a
s
c
la
s
s
if
ic
a
ti
on
of
I
S
L
s
ig
n
la
ngua
ge
.
T
he
s
ugge
s
te
d
a
ppr
oa
c
h
e
a
s
il
y
s
ol
ve
d
th
e
de
te
r
m
in
a
ti
on
of
two
-
ha
nd
I
S
L
ge
s
tu
r
e
pr
obl
e
m
.
B
ut
th
e
s
ugge
s
te
d
a
ppr
oa
c
h
ha
d
poor
pe
r
f
o
r
m
a
nc
e
in
c
a
s
e
of
de
te
r
m
in
e
th
e
s
im
il
a
r
ge
s
tu
r
e
s
.
N
a
ta
r
a
ja
n
e
t
al
.
[
20]
pr
e
s
e
nt
e
d
th
e
c
om
pl
e
te
D
L
a
ppr
oa
c
h
to
ha
ndl
e
th
e
S
L
R
,
pr
oduc
ti
on
ta
s
ks
,
a
nd
tr
a
ns
l
a
ti
on
in
r
e
a
l
-
ti
m
e
c
a
s
e
s
.
T
he
s
ugg
e
s
te
d
a
ppr
oa
c
h
ut
il
iz
e
d
a
M
e
di
a
P
ip
e
li
br
a
r
y
a
nd
hybr
id
C
N
N
w
it
h
B
i
-
di
r
e
c
ti
ona
l
lo
ng
s
hor
t
-
te
r
m
m
e
m
or
y
(
B
i
-
L
S
T
M
)
f
or
im
a
ge
a
nd
te
xt
e
xt
r
a
c
ti
on.
A
lt
e
r
na
ti
ve
ly
,
th
e
s
ig
n
ge
s
tu
r
e
vi
de
os
of
s
pe
e
c
h
s
e
nt
e
n
c
e
s
w
e
r
e
c
om
put
e
d
by
th
e
ut
il
iz
a
ti
on
of
hybr
id
ne
ur
a
l
m
a
c
hi
ne
tr
a
ns
la
ti
on
(
N
M
T
)
+
M
e
di
a
P
ip
e
+
dyna
m
ic
ge
ne
r
a
ti
ve
a
dve
r
s
a
r
ia
l
ne
twor
k
(
G
A
N
)
a
ppr
oa
c
h.
N
one
th
e
le
s
s
,
th
e
s
ugge
s
te
d
a
ppr
oa
c
h
in
s
uc
h
pr
ot
oc
ol
s
s
ig
ni
f
ic
a
nt
ly
a
f
f
e
c
te
d
th
e
pe
r
f
or
m
a
nc
e
.
N
a
ndi
e
t
al
.
[
21]
in
tr
oduc
e
d
th
e
C
N
N
in
te
gr
a
te
d
w
it
h
a
ugm
e
nt
a
ti
on,
ba
tc
h
nor
m
a
li
z
a
ti
on,
dr
opout,
s
to
c
ha
s
ti
c
pool
in
g
a
s
w
e
ll
a
s
di
f
G
r
a
d
opt
im
iz
e
r
f
or
f
in
ge
r
s
pe
ll
in
g
s
ta
ti
c
s
ig
n r
e
c
ogni
ti
on a
ppr
oa
c
h
f
or
I
S
L
a
lp
ha
be
t.
T
he
t
r
a
in
in
g,
t
e
s
ti
ng a
nd l
os
s
e
s
of
th
e
s
ugge
s
te
d
a
ppr
oa
c
h
w
e
r
e
a
tt
a
in
e
d
f
or
m
ul
ti
pl
e
s
e
pa
r
a
te
opt
im
iz
e
r
s
,
a
s
w
e
ll
a
s
th
r
e
e
ty
pe
s
of
pool
in
g
a
ppr
oa
c
he
s
.
N
one
th
e
le
s
s
,
th
e
s
ugge
s
te
d
a
ppr
oa
c
h
w
a
s
f
ut
il
e
i
n
de
te
c
ti
ng
th
e
dyna
m
ic
a
nd
r
e
a
l
-
ti
m
e
s
ig
ns
.
F
r
om
th
is
a
na
ly
s
is
,
s
om
e
li
m
it
a
ti
ons
ha
ve
be
e
n
id
e
nt
if
ie
d;
e
xe
c
ut
e
d
onl
y
w
it
h
th
e
m
in
i
m
um
s
ta
ti
c
s
ym
bol
s
,
poor
pe
r
f
or
m
a
nc
e
,
a
nd
f
ut
il
e
to
r
e
c
ogni
z
e
th
e
dyna
m
ic
a
nd
r
e
a
l
-
ti
m
e
s
ig
ns
.
B
a
s
e
d
upon
th
is
in
f
e
r
e
nc
e
,
th
is
r
e
s
e
a
r
c
h
pr
opos
e
d
e
m
pl
oys
I
S
L
r
e
c
ogni
ti
on,
a
nd
th
e
n
e
w
a
p
pr
oa
c
h
is
di
s
c
us
s
e
d
in
d
e
ta
il
in
th
e
f
ol
lo
w
in
g
s
e
c
ti
on. T
h
e
m
a
jo
r
c
ont
r
ib
ut
io
ns
of
t
hi
s
pa
pe
r
a
r
e
l
is
te
d a
s
f
ol
lo
w
s
:
−
T
he
pr
e
pr
oc
e
s
s
in
g
is
pe
r
f
or
m
e
d
by
us
in
g
da
ta
a
ugm
e
nt
a
ti
on
a
nd
da
ta
nor
m
a
li
z
a
ti
on
te
c
hni
que
s
.
T
h
e
m
ul
ti
t
hr
e
s
hol
d i
m
a
ge
e
nt
r
opy te
c
hni
que
i
s
us
e
d f
or
s
e
gm
e
nt
a
ti
on.
−
I
n
th
is
r
e
s
e
a
r
c
h,
th
e
nove
l
a
nd
r
obus
t
of
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
ppr
oa
c
h
is
pr
opos
e
d
f
or
S
L
R
w
hi
c
h
is
ut
il
iz
e
d i
n I
S
L
t
r
a
ns
la
ti
on s
ys
te
m
.
−
T
he
pr
opos
e
d
m
e
th
od
e
f
f
or
tl
e
s
s
ly
c
h
a
ll
e
nge
s
a
c
om
pl
e
x
pr
obl
e
m
f
or
de
te
r
m
in
in
g
two
-
ha
nd
ge
s
tu
r
e
s
of
I
S
L
w
it
h be
tt
e
r
r
e
c
ogni
ti
on outc
om
e
s
ove
r
ot
he
r
t
he
e
xi
s
ti
ng w
or
ks
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Si
gn l
anguage
r
e
c
ogni
ti
on and c
la
s
s
if
ic
at
io
n us
in
g bl
e
nde
d e
n
s
e
m
bl
e
m
ac
hi
ne
le
a
r
ni
ng
(
A
k
as
h R
aj
an R
ai
)
2037
T
he
s
tr
uc
tu
r
e
of
th
e
p
a
pe
r
is
a
r
r
a
nge
d
a
s
f
ol
lo
w
s
:
s
e
c
ti
on
2
pr
ovi
de
s
th
e
pr
opos
e
d
m
e
th
od.
S
e
c
ti
on
3
pr
e
s
e
nt
s
th
e
c
l
a
s
s
if
ic
a
ti
on
u
s
in
g
bl
e
nde
d
e
n
s
e
m
bl
e
ML
te
c
hni
que
s
.
S
e
c
ti
on
4
pr
ovi
de
s
th
e
r
e
s
ul
ts
a
nd
di
s
c
us
s
io
n
,
a
nd
s
e
c
ti
on 5
c
ove
r
s
t
he
c
onc
lu
s
io
n of
t
hi
s
r
e
s
e
a
r
c
h.
2.
P
R
O
P
O
S
E
D
M
E
T
H
O
D
I
n
th
is
s
e
c
ti
on,
th
e
pr
opos
e
d
m
e
th
odol
ogy
i
s
pr
e
s
e
nt
e
d
to
de
s
c
r
ib
e
va
r
io
us
s
te
p
s
w
hi
c
h
a
r
e
c
a
r
r
ie
d
out
in
th
is
r
e
s
e
a
r
c
h.
T
h
e
da
ta
f
lo
w
s
in
th
is
pr
opos
e
d
m
e
th
o
d
in
vol
ve
s
c
ol
le
c
ti
on
of
da
ta
,
pr
e
pr
oc
e
s
s
in
g,
s
e
gm
e
nt
a
ti
on, f
e
a
tu
r
e
e
xt
r
a
c
ti
on a
nd c
la
s
s
if
ic
a
ti
on. F
ig
ur
e
1 s
ho
w
s
t
he
pr
oc
e
s
s
i
nt
r
ic
a
te
i
n S
L
R
.
F
ig
ur
e
1. W
or
kf
lo
w
of
t
he
pr
opos
e
d m
e
th
od
2.1.
D
at
as
e
t
D
a
ta
a
c
qui
s
it
io
n
is
s
ig
ni
f
ic
a
nt
f
or
r
e
s
e
a
r
c
h
w
or
k
s
a
nd
it
is
e
s
s
e
n
ti
a
l
f
or
M
L
de
ve
lo
pm
e
nt
.
T
he
one
of
th
e
m
os
t
s
ig
ni
f
ic
a
nt
c
ha
ll
e
nge
s
f
a
c
e
d
dur
in
g
th
e
r
e
s
e
a
r
c
h
is
th
a
t
th
e
r
e
a
r
e
no
be
nc
hm
a
r
k
da
ta
s
e
ts
a
va
il
a
bl
e
f
or
I
S
L
[
22]
.
H
e
nc
e
,
th
is
r
e
s
e
a
r
c
h
tr
ie
d
to
phy
s
ic
a
ll
y
de
ve
lo
p
th
e
da
ta
s
e
t
th
a
t
c
om
pr
is
e
s
of
a
lp
h
a
be
ts
(
A
-
Z
)
a
nd
num
e
r
ic
a
l
va
lu
e
s
(
0
-
9)
by
ut
il
iz
in
g
th
e
e
xt
e
r
na
l
w
e
b
c
a
m
e
r
a
.
S
ig
n
is
a
c
c
om
pl
is
he
d
by
65
va
r
io
us
pe
opl
e
w
hi
c
h c
ons
e
que
nc
e
s
i
n a
t
ot
a
l
of
1690 ge
s
tu
r
e
s
. F
r
om
t
hi
s
c
ol
le
c
te
d da
ta
s
e
t,
e
ve
r
y i
m
a
ge
i
s
s
c
a
le
d down f
r
om
a
m
a
xi
m
um
de
f
in
it
io
n
of
s
iz
e
256
×
256
.
F
ig
ur
e
2
s
how
s
th
e
s
a
m
pl
e
s
c
ol
le
c
te
d
da
ta
of
bot
h
a
lp
ha
be
ts
a
nd
num
e
r
ic
a
l
va
lu
e
s
.
F
ig
ur
e
2. S
a
m
pl
e
da
ta
s
e
t
of
ha
nd ge
s
tu
r
e
2.2.
P
r
e
-
p
r
oc
e
s
s
in
g
I
n
th
is
s
e
c
ti
on,
th
e
c
ol
le
c
te
d
da
ta
is
ut
il
iz
e
d
a
s
in
put
f
or
t
he
pr
e
-
pr
oc
e
s
s
in
g.
T
he
s
e
te
c
hni
que
s
in
vol
ve
da
ta
a
ugm
e
nt
a
ti
on,
a
nd
nor
m
a
li
z
a
ti
on
a
r
e
a
ppl
ie
d
in
th
e
c
ol
le
c
te
d
I
S
L
ge
s
tu
r
e
s
da
ta
.
I
n
da
t
a
a
ugm
e
nt
a
ti
on, t
he
c
ol
le
c
te
d i
nput
I
S
L
da
ta
s
e
t
ha
s
onl
y a
l
im
i
te
d numbe
r
of
s
a
m
pl
e
s
, he
nc
e
da
ta
a
ugm
e
nt
a
ti
on
is
pe
r
f
or
m
e
d
to
e
nha
nc
e
th
e
s
iz
e
of
th
e
im
a
ge
s
a
m
pl
e
s
[
23]
,
[
24]
.
T
he
da
ta
is
e
nha
nc
e
d
by
c
r
a
ti
ng
th
e
ne
w
s
im
il
a
r
s
a
m
pl
e
s
th
r
ough
tr
a
ns
f
or
m
in
g
th
e
a
c
tu
a
l
da
ta
.
T
he
m
ode
l
ge
ts
tr
a
in
e
d
ut
il
iz
in
g
49
,
920
im
a
ge
s
(
80%
)
a
nd
te
s
te
d
ut
il
iz
in
g
12
,
480
(
20%
)
of
th
e
to
ta
l
of
62
,
400
im
a
ge
s
.
T
he
tr
a
in
e
d
im
a
ge
s
a
r
e
a
ugm
e
nt
e
d
by
us
in
g
va
r
io
us
a
f
f
in
e
ope
r
a
ti
ons
s
uc
h
a
s
z
oom
in
g,
r
ot
a
ti
on,
s
ke
w
i
ng,
s
he
a
r
in
g,
he
ig
ht
a
nd
w
id
th
s
hi
f
t.
A
s
a
c
ons
e
que
nc
e
, e
ve
r
y i
m
a
ge
pr
oduc
e
s
90 ne
w
i
m
a
ge
s
, he
nc
e
ne
w
ly
pr
oduc
e
d i
m
a
ge
s
a
r
e
m
ul
ti
pl
ie
d i
nt
o 49
,
92
0
im
a
ge
s
,
w
hi
c
h
r
e
s
ul
te
d
in
4
,
542
,
720
im
a
ge
s
in
a
ugm
e
nt
e
d
tr
a
i
ni
ng
da
ta
.
T
he
n,
th
e
a
ugm
e
nt
e
d
d
a
ta
is
f
e
d
to
th
e
nor
m
a
li
z
a
ti
on t
e
c
hni
que
t
o s
ta
nda
r
di
z
e
t
he
i
nput
t
o a
l
a
r
ge
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
3
,
J
une
2025
:
2035
-
2043
2038
T
he
da
ta
nor
m
a
li
z
a
ti
on
is
ut
il
iz
e
d
to
e
s
ti
m
a
te
if
th
e
da
ta
di
s
tr
ib
ut
io
n
in
e
ve
r
y
in
put
pi
xe
l
is
unc
ha
ngi
ng
or
r
e
gul
a
r
,
a
lo
ngs
id
e
f
a
s
te
r
tr
a
in
in
g
c
onve
r
ge
n
c
e
.
T
h
e
nor
m
a
li
z
a
ti
on
ta
k
e
s
m
a
xi
m
um
a
nd
m
in
im
um
va
lu
e
s
to
s
e
t
th
e
da
t
a
in
b
e
twe
e
n
th
e
r
a
nge
of
0
a
nd
1
,
r
e
s
pe
c
ti
ve
ly
.
T
h
e
nor
m
a
li
z
a
ti
on
i
s
c
a
lc
ul
a
te
d
by us
in
g (
1)
:
=
X
−
X
m
i
n
X
m
a
x
−
X
m
i
n
(
1)
W
he
r
e
,
X
is
th
e
nor
m
a
li
z
e
d
va
lu
e
,
X
min
a
nd
X
ma
x
a
r
e
th
e
m
in
im
um
a
nd
m
a
xi
m
um
of
e
ve
r
y
f
e
a
tu
r
e
.
T
he
n,
th
e
nor
m
a
li
z
e
d da
ta
i
s
f
or
w
a
r
de
d f
or
t
he
s
e
gm
e
nt
a
ti
on pr
oc
e
s
s
.
2.3.
S
e
gm
e
n
t
at
io
n
I
n
th
is
s
e
c
ti
on,
th
e
pr
e
-
pr
oc
e
s
s
e
s
out
put
is
ut
il
iz
e
d
a
s
in
put
to
s
e
gm
e
nt
th
e
im
a
ge
d
a
ta
.
I
m
a
ge
s
e
gm
e
nt
a
ti
on
is
a
n
a
ppr
oa
c
h
f
or
e
xt
r
a
c
ti
ng
a
ta
r
ge
t
obj
e
c
t
f
r
om
th
e
ba
c
kgr
ound.
T
he
pr
e
-
pr
oc
e
s
s
e
d
da
ta
a
r
e
c
la
s
s
if
ie
d
a
c
c
or
di
ng
to
th
e
th
r
e
s
hol
d
va
lu
e
a
nd
de
pi
c
te
d
a
s
s
in
gl
e
a
nd
m
ul
ti
-
th
r
e
s
hol
d
s
e
gm
e
nt
a
ti
on.
H
e
nc
e
,
th
is
r
e
s
e
a
r
c
h
us
e
s
th
e
im
a
ge
s
e
gm
e
nt
a
ti
on
w
hi
c
h i
s
ba
s
e
d on ma
xi
m
um
e
nt
r
opy va
lu
e
s
. T
hi
s
m
ode
l
is
ut
i
li
z
e
d
to
va
li
da
te
th
e
opt
im
a
l
th
r
e
s
hol
d
va
lu
e
,
a
s
w
e
ll
a
s
is
u
s
e
d
to
o
pt
im
iz
e
th
e
da
ta
pr
e
s
e
nt
e
d
in
th
e
ba
c
kgr
ound.
H
e
r
e
a
f
te
r
,
a
n
id
e
nt
if
ic
a
ti
on
of
im
a
ge
gr
e
ys
c
a
le
is
a
c
c
om
pl
is
he
d
to
e
xt
r
a
c
t
th
e
ta
r
ge
te
d
a
r
e
a
s
a
nd
th
e
gr
e
ys
c
a
l
e
im
a
ge
r
a
nge
is
e
xpl
a
in
e
d
a
s
[
0
,
−
]
.
T
he
pi
xe
ls
w
hi
c
h
e
m
br
a
c
e
a
gr
e
ys
c
a
le
v
a
lu
e
a
r
e
c
ons
id
e
r
e
d
a
s
ta
r
ge
te
d
a
r
e
a
(
)
a
nd
ba
c
kgr
ound
a
r
e
a
(
)
.
T
he
s
e
two
a
r
e
a
s
in
vol
ve
r
a
ndomne
s
s
,
w
hi
c
h
is
th
e
n
c
om
bi
ne
d
w
it
h pos
s
ib
il
it
y c
onc
e
nt
r
a
ti
ons
r
e
pr
e
s
e
nt
e
d i
n (
2
)
to
(
4)
:
(
)
=
−
∑
=
0
(
2)
(
)
=
−
∑
1
−
−
1
=
+
1
1
−
(
3)
Ψ
(
)
=
(
)
+
(
)
(
4)
W
he
r
e
Ψ
(
)
is
ut
il
iz
e
d
to
a
tt
a
in
th
e
gr
e
a
te
r
va
lu
e
,
a
nd
is
th
e
opt
im
a
l
th
r
e
s
hol
d
va
lu
e
.
H
e
nc
e
,
a
l
a
r
ge
r
e
nt
r
opy
th
r
e
s
hol
di
ng
a
ppr
oa
c
h
is
ut
il
iz
e
d
f
or
m
ul
ti
-
th
r
e
s
hol
d
s
e
gm
e
nt
a
ti
on
a
nd
th
e
pr
obl
e
m
s
in
th
is
a
ppr
oa
c
h
a
r
e
e
xa
m
in
e
d a
s
di
m
e
ns
io
ns
. F
ur
th
e
r
m
or
e
, t
he
gr
e
ys
c
a
le
i
m
a
ge
va
lu
e
s
c
ons
id
e
r
e
d t
o s
e
gm
e
nt
t
he
i
nput
da
ta
a
nd t
he
n e
xpos
e
d t
o t
he
+
1
r
e
gi
on. An obje
c
ti
ve
f
unc
ti
on of
t
hi
s
a
ppr
oa
c
h i
s
pr
ovi
de
d i
n (
5)
.
(
[
1
,
2
,
…
]
)
=
0
+
1
+
⋯
+
=
−
∑
0
(
1
)
0
(
1
)
1
−
1
=
0
+
∑
0
(
1
,
2
)
0
(
1
,
2
)
2
−
1
=
0
+
⋯
+
(
5)
∑
0
(
,
−
1
)
0
(
,
−
1
)
−
1
=
W
he
r
e
a
nd
is
th
e
pos
s
ib
il
it
y
of
gr
e
ys
c
a
le
va
lu
e
in
th
im
a
ge
a
nd
c
la
s
s
.
A
n
opt
im
a
l
th
r
e
s
hol
d
va
lu
e
i
s
de
te
r
m
in
e
d
a
t
th
e
va
lu
e
of
(
[
1
,
2
,
…
]
)
,
in
f
lu
e
nc
in
g
a
s
im
il
a
r
va
lu
e
of
1
∗
,
2
∗
,
…
∗
.
S
ti
ll
,
th
is
s
e
gm
e
nt
a
ti
on
ow
ns
t
he
e
f
f
ic
ie
nt
s
e
gm
e
nt
a
ti
on a
c
c
ur
a
c
y a
nd t
he
s
e
gm
e
nt
e
d v
a
lu
e
i
s
f
e
d t
o t
he
f
e
a
tu
r
e
e
xt
r
a
c
ti
on pr
oc
e
s
s
.
2.4.
F
e
at
u
r
e
e
xt
r
ac
t
io
n
T
he
s
e
gm
e
nt
e
d
da
ta
is
gi
ve
n
to
th
e
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
in
th
e
da
ta
s
e
t.
T
he
ge
ne
r
a
l
a
im
of
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
is
to
r
e
duc
e
th
e
di
m
e
ns
io
na
li
ty
a
nd
da
ta
c
om
pa
c
ti
on.
H
e
n
c
e
,
V
G
G
-
16
pr
e
-
t
r
a
in
e
d
a
r
c
hi
te
c
tu
r
e
is
us
e
d
f
or
th
e
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
pr
oc
e
s
s
.
T
he
V
G
G
-
16
[
25
]
,
[
26]
is
th
e
m
os
t
popula
r
im
a
ge
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
e
m
pl
oye
d
to
e
xt
r
a
c
t
a
la
r
ge
a
m
ount
of
da
ta
.
T
he
V
G
G
-
16
a
r
c
hi
te
c
tu
r
e
c
ons
is
ts
of
va
r
io
us
la
ye
r
s
s
uc
h
a
s
c
onvolut
io
n,
f
ul
ly
c
onne
c
te
d
(
F
C
)
a
s
w
e
ll
a
s
po
ol
in
g,
a
ls
o
a
ppl
ic
a
bl
e
in
A
le
xN
e
t
a
r
c
hi
te
c
tu
r
e
.
T
he
in
put
s
iz
e
of
th
is
a
r
c
hi
te
c
tu
r
e
is
f
ix
e
d
to
224
∗
224
pi
xe
ls
,
a
c
c
om
pa
ni
e
d
by
a
f
il
te
r
s
iz
e
of
3
∗
3
.
A
t
th
e
e
nd
of
th
is
a
r
c
hi
te
c
tu
r
e
,
it
in
vol
ve
s
th
e
a
c
ti
va
ti
on
f
unc
ti
on
w
hi
c
h
is
a
c
c
om
pl
is
he
d
f
or
di
s
tr
ib
ut
in
g
th
e
pr
oba
bi
li
ti
e
s
c
la
s
s
e
s
to
out
put
la
ye
r
s
.
I
n
th
is
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
f
or
th
e
pr
opos
e
d
m
e
th
od,
th
e
in
it
ia
l
la
y
e
r
of
VGG
-
16 a
r
c
hi
te
c
tu
r
e
i
s
f
or
w
a
r
de
d t
o
th
e
c
onvolut
io
na
l
la
ye
r
w
it
h
224
∗
224
pi
xe
l
im
a
ge
s
iz
e
. I
n t
he
r
e
c
ti
f
ie
d
li
ne
a
r
uni
t
(
R
e
L
U
)
a
c
ti
va
ti
on
f
unc
ti
on,
224
∗
224
im
a
ge
s
iz
e
is
f
or
w
a
r
de
d
by
pr
im
a
r
y
s
t
a
c
k
of
two
c
onvolut
io
ns
w
it
h
th
e
a
ppr
opr
ia
te
r
e
gi
on
of
3
∗
3
.
E
ve
r
y
la
ye
r
of
th
is
a
r
c
hi
te
c
tu
r
e
c
ons
is
t
s
of
64
f
il
te
r
s
a
nd
a
pa
ddi
ng
is
of
te
n
r
e
s
e
r
ve
d
w
it
h
a
pi
xe
l
s
iz
e
of
1
,
w
he
r
e
a
s
s
tr
id
e
va
lu
e
is
s
e
t
to
1
.
A
s
pa
ti
a
l
r
e
s
ol
ut
io
n
is
ta
ke
n
in
th
is
m
ode
l,
w
he
r
e
in
th
e
a
c
ti
va
ti
on
f
unc
ti
on
is
de
a
lt
w
it
h
a
pool
in
g
la
ye
r
w
it
h
th
e
s
tr
id
e
va
lu
e
of
2
×
2
pi
xe
ls
.
M
or
e
ove
r
,
in
th
e
f
ol
lo
w
in
g
s
te
p,
th
e
out
c
om
e
i
s
onc
e
m
or
e
f
or
w
a
r
de
d
to
th
e
c
onvolut
io
na
l
la
ye
r
s
a
nd
la
te
r
f
e
d
to
a
pool
in
g
la
ye
r
,
w
hi
c
h
is
r
e
s
ul
te
d
in
a
n
out
c
om
e
o
f
56
∗
56
∗
128
pi
xe
ls
.
T
hi
s
la
ye
r
e
ndur
e
s
th
e
f
in
a
l
pr
oc
e
s
s
of
c
onvolut
io
na
l
a
s
w
e
ll
a
s
pool
in
g
la
y
e
r
.
T
he
f
la
tt
e
n
la
ye
r
is
th
e
lo
c
a
t
e
d
a
m
ong
F
C
la
ye
r
s
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Si
gn l
anguage
r
e
c
ogni
ti
on and c
la
s
s
if
ic
at
io
n us
in
g bl
e
nde
d e
n
s
e
m
bl
e
m
ac
hi
ne
le
a
r
ni
ng
(
A
k
as
h R
aj
an R
ai
)
2039
w
hi
c
h
a
r
e
th
e
n
ut
il
iz
e
d
a
s
th
e
c
onvolut
io
na
l
s
ta
c
k
s
.
F
in
a
ll
y,
t
he
F
C
la
y
e
r
is
p
e
r
f
or
m
e
d
a
s
a
n
out
put
la
y
e
r
w
hi
c
h
in
vol
ve
s
1000
ne
ur
ons
,
w
he
r
e
a
s
,
th
e
in
it
ia
l
two
F
C
la
ye
r
s
r
e
pr
e
s
e
nt
a
n
out
c
om
e
of
4096
ne
ur
ons
.
T
he
n, t
he
e
xt
r
a
c
te
d f
e
a
tu
r
e
s
a
r
e
f
or
w
a
r
de
d t
o t
he
c
la
s
s
if
ic
a
ti
on
pr
oc
e
s
s
.
3.
C
L
A
S
S
I
F
I
C
A
T
I
O
N
U
S
I
N
G
B
L
E
N
D
E
D
E
N
S
E
M
B
L
E
M
A
C
H
I
N
E
L
E
A
R
N
I
N
G
T
E
C
H
N
I
Q
U
E
S
I
n
th
is
s
e
c
ti
on,
a
n
out
c
om
e
of
e
xt
r
a
c
t
e
d
f
e
a
tu
r
e
s
is
ut
il
iz
e
d
a
s
in
put
f
or
c
la
s
s
if
yi
ng
th
e
va
r
io
us
s
ig
n
la
ngua
ge
s
.
T
he
M
L
a
lg
or
it
hm
s
-
ba
s
e
d
r
e
c
ogni
ti
on
a
nd
c
la
s
s
i
f
ic
a
ti
on
of
s
ig
n
la
ngua
ge
s
pr
ovi
de
a
be
tt
e
r
pe
r
f
or
m
a
nc
e
.
I
n
th
is
c
la
s
s
if
ic
a
ti
on
pr
oc
e
s
s
,
th
e
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
ppr
oa
c
he
s
a
r
e
us
e
d
f
or
r
e
c
ogni
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
of
s
ig
n
la
ngua
ge
.
E
ns
e
m
bl
e
le
a
r
ni
ng
is
th
e
pr
oc
e
dur
e
of
in
te
gr
a
ti
ng
va
r
io
us
le
a
r
ni
ngs
to
e
nha
nc
e
th
e
pe
r
f
or
m
a
nc
e
,
w
he
r
e
e
ve
r
y
le
a
r
ni
ng
a
ppr
oa
c
h
tr
a
in
s
on
va
r
io
us
s
ubs
e
ts
of
f
e
a
tu
r
e
s
.
T
h
e
e
ns
e
m
bl
e
le
a
r
ni
ng
pe
r
f
or
m
a
nc
e
be
lo
ngi
ng
to
th
e
w
or
ki
ng s
ty
le
a
nd c
a
pa
b
il
it
y
of
th
e
s
e
le
c
te
d
ba
s
e
c
la
s
s
if
ie
r
s
.
E
ns
e
m
bl
e
le
a
r
ni
ng
m
ode
ls
e
nde
a
vor
to
e
nha
nc
e
pr
e
di
c
ta
bi
li
ty
th
r
ough
c
om
bi
ni
ng
th
e
va
r
io
us
a
ppr
oa
c
he
s
in
to
hi
ghl
y
de
pe
nda
bl
e
i
ndi
vi
dua
l
m
ode
l.
T
he
d
e
ta
il
e
d i
nf
or
m
a
ti
on a
bout
t
hi
s
bl
e
nde
d e
ns
e
m
bl
e
a
ppr
oa
c
h i
s
d
e
s
c
r
ib
e
d.
3.1. K
-
n
e
ar
e
s
t
n
e
ig
h
b
or
T
he
K
N
N
[
27]
a
lg
or
it
hm
is
unpr
e
te
nt
io
us
a
nd
e
a
s
y
-
to
-
im
pl
e
m
e
nt
w
he
n
c
om
pa
r
e
d
to
th
e
lo
gi
s
ti
c
r
e
gr
e
s
s
io
n
(
L
R
)
.
T
he
a
im
of
th
is
a
lg
or
it
hm
is
to
m
e
a
s
ur
e
th
e
di
s
ta
nc
e
a
m
ong
s
e
le
c
te
d
poi
nt
a
nd
ot
he
r
poi
nt
s
.
T
o
s
e
le
c
t
th
e
k
poi
nt
s
w
it
h
m
in
im
a
l
di
s
ta
n
c
e
,
th
is
m
ode
l
c
r
e
a
te
s
s
ta
ti
s
ti
c
s
on
c
l
a
s
s
if
ic
a
ti
on
ty
p
e
s
of
th
o
s
e
s
e
le
c
te
d
k
poi
nt
s
.
T
he
K
N
N
a
lg
or
it
hm
be
li
e
ve
s
th
a
t
th
e
obj
e
c
ts
a
r
e
s
im
il
a
r
to
e
a
c
h
ot
he
r
.
T
h
e
di
s
ta
n
c
e
i
s
m
a
in
ly
c
a
lc
ul
a
te
d
by
us
in
g
E
uc
li
de
a
n
a
nd
M
a
nha
tt
a
n
di
s
ta
nc
e
,
th
e
m
a
th
e
m
a
ti
c
a
l
f
or
m
ul
a
of
th
e
s
e
di
s
t
a
nc
e
is
f
or
m
ul
a
te
d i
n (
6)
a
nd (
7
)
:
:
(
,
)
=
√
∑
(
−
)
2
=
1
(
6)
ℎ
:
(
,
)
=
√
∑
|
−
|
=
1
(
7)
w
he
r
e
,
-
th
di
m
e
ns
io
na
l
va
lu
e
s
of
t
he
t
w
o point
s
w
hi
c
h
a
r
e
ut
il
iz
e
d i
n di
m
e
ns
io
na
l
s
p
a
c
e
.
3.2.
N
aï
ve
B
aye
s
T
he
N
B
[
28]
is
one
of
th
e
M
L
a
ppr
oa
c
he
s
,
w
hi
c
h
ut
il
iz
e
s
th
e
B
a
ye
s
th
e
or
e
m
to
s
ol
ve
th
e
pr
obl
e
m
pr
oba
bi
li
ty
f
unc
ti
on.
T
o
de
te
r
m
in
e
th
e
pos
te
r
io
r
di
s
tr
ib
ut
io
n,
ba
c
kgr
ound
knowle
dge
is
e
xpl
a
in
e
d
a
s
pr
io
r
di
s
tr
ib
ut
io
n,
a
s
w
e
ll
a
s
a
s
s
oc
i
a
te
d
w
it
h
th
e
pe
r
c
e
pt
io
na
l
da
t
a
in
th
e
pr
oc
e
dur
e
of
pr
oba
bi
li
ty
f
unc
ti
on.
T
he
B
a
ye
s
ia
n
s
ta
ti
s
ti
c
s
i
s
a
n
a
ppr
oa
c
h
f
or
de
te
r
m
in
in
g
th
e
da
t
a
a
n
d
e
va
lu
a
ti
ng
th
e
p
a
r
a
m
e
te
r
s
a
c
c
or
di
ng
to
th
e
B
a
ye
s
’
th
e
or
e
m
,
w
hi
c
h
is
e
xpr
e
s
s
e
d
in
(
8)
a
nd
th
e
(
)
a
r
e
e
s
ti
m
a
te
d
by
us
in
g
(
9)
ba
s
e
d
on
th
e
to
ta
l
pr
oba
bi
li
ty
r
ul
e
.
(
|
)
=
(
|
)
(
)
(
)
(
8)
(
)
=
∑
(
)
(
|
)
=
1
(
9)
E
ve
nt
ua
ll
y,
th
e
pr
oba
bi
li
ty
de
pe
nds
on
a
p
a
r
ti
c
ul
a
r
gr
oup
a
nd
th
e
gr
e
a
te
r
va
lu
e
is
ta
ke
n
a
s
c
la
s
s
if
ic
a
ti
on
out
c
om
e
.
T
h
e
r
e
a
r
e
va
r
io
us
ty
pe
s
of
N
B
s
uc
h
a
s
B
e
r
noul
li
,
m
ul
ti
nom
in
a
l
a
nd
G
a
us
s
ia
n
N
B
,
w
hi
c
h
di
f
f
e
r
e
nt
ia
te
th
e
f
e
a
tu
r
e
ve
c
to
r
a
s
di
s
c
r
e
te
or
c
ont
in
uou
s
.
T
he
B
e
r
noul
li
N
B
c
la
s
s
if
ie
r
is
s
ui
ta
bl
e
in
a
s
it
ua
ti
on
w
he
r
e
th
e
f
e
a
tu
r
e
ve
c
to
r
s
im
it
a
te
to
th
e
B
e
r
noul
li
di
s
tr
ib
ut
io
n
s
uc
h
a
s
bi
na
r
y
di
s
tr
ib
ut
io
n.
T
he
m
ul
ti
nom
ia
l
N
B
c
la
s
s
if
ie
r
is
a
ppl
ic
a
bl
e
f
or
th
e
di
s
c
r
e
te
f
e
a
tu
r
e
ve
c
to
r
a
nd
im
it
a
te
s
th
e
m
ul
ti
va
r
ia
te
di
s
tr
ib
ut
io
ns
.
F
in
a
ll
y,
th
e
G
a
us
s
ia
n
N
B
is
e
m
pl
oye
d
onl
y
onc
e
th
e
f
e
a
tu
r
e
ve
c
to
r
s
ha
ve
c
ont
in
uous
va
r
ia
bl
e
s
a
nd i
m
it
a
te
or
e
s
ti
m
a
te
t
o t
he
a
c
tu
a
l
di
s
tr
ib
ut
io
n.
3.3. Ran
d
om
f
or
e
s
t
R
F
[
29]
is
a
n
il
lu
s
tr
a
ti
ve
e
ns
e
m
bl
e
le
a
r
ni
ng
a
ppr
oa
c
h
th
a
t
de
pe
nds
on
s
tr
a
te
gi
e
s
of
th
e
ba
ggi
ng
f
a
m
il
y. T
he
R
F
ut
il
iz
e
s
de
c
is
io
n t
r
e
e
(
D
T
)
a
s
ba
s
e
l
e
a
r
ne
r
s
a
nd de
ve
lo
ps
va
r
io
us
D
T
by uti
li
z
in
g t
he
s
a
m
pl
in
g
a
ppr
oa
c
h
w
it
hout
pl
a
c
e
m
e
nt
.
T
he
R
F
is
a
n
a
ppr
oa
c
h
f
or
a
ggr
e
g
a
ti
ng
or
e
ve
n
ba
ggi
ng
da
ta
w
hi
c
h
is
de
pl
oye
d
to
m
in
im
iz
e
a
s
ig
ni
f
ic
a
nt
pa
r
a
m
e
te
r
of
t
he
va
r
ia
nc
e
i
n t
he
out
put
. I
n c
la
s
s
if
ic
a
ti
on pr
oc
e
s
s
,
e
a
c
h t
r
e
e
c
on
s
id
e
r
s
th
e
out
put
a
s
t
he
c
la
s
s
e
s
a
nd t
he
c
la
s
s
e
s
w
it
h gr
e
a
te
r
numbe
r
of
out
c
om
e
s
a
r
e
s
e
le
c
te
d a
s
f
in
a
l
out
put
. T
he
t
e
s
t
s
a
m
pl
e
s
a
c
qui
r
e
out
c
om
e
s
on
e
v
e
r
y
de
ve
lo
pe
d
D
T
a
s
w
e
ll
a
s
e
s
ti
m
a
te
th
e
te
s
t
s
a
m
pl
e
c
a
te
gor
y
by
th
e
vot
in
g
s
tr
a
te
gy.
A
t
th
e
w
or
ki
ng
of
R
F
,
va
r
io
us
s
a
m
pl
e
s
a
r
e
a
r
bi
tr
a
r
il
y
c
hos
e
n
to
de
v
e
lo
p
th
e
num
be
r
of
DT
,
w
hi
le
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
3
,
J
une
2025
:
2035
-
2043
2040
th
e
pr
oc
e
dur
e
of
D
T
de
ve
lo
pm
e
nt
,
th
e
s
e
l
e
c
te
d
f
e
a
tu
r
e
s
a
nd
it
s
s
pl
it
node
s
a
r
e
a
r
bi
tr
a
r
y.
F
or
th
e
c
la
s
s
if
ic
a
ti
on
ta
s
k, t
he
out
put
of
R
F
i
s
t
he
c
la
s
s
s
e
le
c
te
d by the
m
a
jo
r
it
y of
t
he
c
la
s
s
.
3.4. S
u
p
p
or
t
ve
c
t
or
m
ac
h
in
e
S
V
M
[
30]
is
w
id
e
ly
us
e
d
f
or
th
e
pr
obl
e
m
s
of
c
la
s
s
if
ic
a
ti
on,
de
te
c
ti
on
a
s
w
e
ll
a
s
r
e
gr
e
s
s
io
n.
S
V
M
is
th
e
m
os
t
e
ne
r
gy
e
f
f
ic
ie
nt
a
s
it
ut
il
iz
e
s
th
e
s
ubs
e
t
of
tr
a
in
in
g
poi
nt
s
in
de
c
is
io
n
f
unc
ti
on.
T
he
S
V
M
e
m
pl
oys
e
f
f
ic
ie
nt
ly
in
hi
gh
-
di
m
e
ns
io
na
l
f
e
a
tu
r
e
ve
c
to
r
s
,
he
nc
e
hype
r
pl
a
ne
di
m
e
ns
io
ns
a
r
e
of
te
n
le
s
s
th
a
n
1
in
f
e
a
tu
r
e
ve
c
to
r
.
T
he
pr
oc
e
s
s
of
id
e
nt
if
yi
ng
th
e
de
s
ir
e
d
hype
r
pl
a
ne
i
s
de
te
r
m
in
in
g
a
m
a
xi
m
um
m
a
r
gi
n.
T
he
pur
po
s
e
is
to
e
nha
nc
e
t
he
m
a
r
gi
n f
or
i
m
pr
ovi
ng t
he
r
obus
tn
e
s
s
, a
s
w
e
ll
a
s
f
or
r
e
duc
in
g t
he
c
la
s
s
if
ic
a
ti
on’
s
e
r
r
or
r
a
te
. T
he
s
a
m
pl
e
poi
nt
s
w
hi
c
h
de
te
r
m
in
e
a
gr
e
a
te
r
m
a
r
g
in
a
r
e
known
a
s
s
uppor
t
ve
c
to
r
s
.
T
he
S
V
M
is
us
e
d
f
or
m
a
ki
n
g
pr
e
di
c
ti
ons
a
nd
is
e
s
s
e
nt
ia
l
f
or
f
it
on
s
pa
r
s
e
da
t
a
.
T
he
goa
l
of
S
V
M
is
to
c
la
s
s
if
y
th
e
s
a
m
pl
e
in
to
one
or
m
or
e
c
la
s
s
e
s
th
r
ough
th
e
e
xt
e
ns
io
n
th
a
t
s
pe
c
if
ie
s
to
th
e
m
ul
ti
-
c
la
s
s
pr
obl
e
m
.
T
he
r
e
a
r
e
a
num
be
r
of
hype
r
pl
a
ne
s
th
a
t
di
s
c
r
im
in
a
te
th
e
two
c
l
a
s
s
e
s
;
how
e
ve
r
,
th
e
a
im
i
s
a
tt
a
in
e
d
by
id
e
nt
if
yi
ng
a
n
opt
im
a
l
s
e
pa
r
a
ti
ng
hype
r
pl
a
ne
w
hi
c
h pl
a
c
e
s
out
e
r
m
os
t
f
r
om
bot
h c
la
s
s
e
s
.
4.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
T
hi
s
s
e
c
ti
on
il
lu
s
tr
a
te
s
th
e
r
e
s
ul
ts
a
nd
di
s
c
u
s
s
io
n
of
th
e
pr
op
os
e
d
m
e
th
od
f
or
S
L
R
.
T
he
pr
opos
e
d
m
e
th
od
f
or
S
L
R
is
e
xe
c
ut
e
d
by
ut
il
iz
in
g
th
e
pl
a
tf
or
m
of
pyt
ho
n
3.8
w
it
h
W
in
dow
s
10
O
S
,
16
G
B
R
A
M
w
it
h
in
te
l
-
i7
pr
oc
e
s
s
or
.
T
h
e
pr
opos
e
d
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
ppr
o
a
c
h
is
a
na
ly
z
e
d
ba
s
e
d
on
v
a
r
io
us
a
s
s
e
s
s
m
e
nt
m
e
tr
ic
s
na
m
e
d
a
c
c
ur
a
c
y,
pr
e
c
i
s
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
.
T
he
m
a
th
e
m
a
ti
c
a
l
e
xpr
e
s
s
io
ns
of
th
e
s
e
m
e
tr
ic
e
s
a
r
e
e
xpr
e
s
s
e
d i
n (
10
)
a
nd (
11
).
=
+
+
+
+
(
10)
=
+
(1
0
)
=
+
(
12)
1
−
=
2
2
+
+
(1
1
)
W
he
r
e
T
P
is
tr
ue
pos
it
iv
e
,
T
N
is
tr
ue
ne
g
a
ti
ve
, F
P
is
f
a
ls
e
pos
it
i
ve
, a
nd F
N
is
f
a
ls
e
ne
g
a
ti
ve
.
4.1. P
e
r
f
or
m
an
c
e
an
al
ys
is
I
n
th
is
s
e
c
ti
on,
th
e
e
f
f
e
c
ti
ve
ne
s
s
of
th
e
pr
opos
e
d
m
e
th
od
f
or
S
L
R
is
a
na
ly
z
e
d
by
us
in
g
th
e
s
ta
nda
r
d
da
ta
s
e
t.
T
a
bl
e
s
1
to
3
s
how
th
e
out
c
om
e
s
dur
in
g
s
e
gm
e
nt
a
ti
on,
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
f
or
S
L
R
on
I
S
L
da
ta
s
e
t.
T
he
pe
r
f
or
m
a
nc
e
m
e
tr
ic
e
s
na
m
e
d
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
a
r
e
ut
il
iz
e
d
to
e
s
ti
m
a
te
a
n e
f
f
e
c
ti
ve
ne
s
s
of
t
he
pr
opos
e
d m
e
th
od.
T
a
bl
e
1
s
how
s
th
e
m
ul
ti
-
th
r
e
s
hol
di
ng
out
c
om
e
s
w
it
h
va
r
io
us
s
e
gm
e
nt
a
ti
on
te
c
hni
que
s
f
or
S
L
R
on
I
S
L
da
ta
s
e
t.
T
h
e
m
ul
ti
-
th
r
e
s
hol
di
ng
a
tt
a
in
s
be
tt
e
r
p
e
r
f
or
m
a
nc
e
a
s
c
om
pa
r
e
d
to
th
e
e
xi
s
ti
ng
m
e
th
ods
w
it
h
va
r
io
us
a
s
s
e
s
s
m
e
nt
m
e
tr
ic
e
s
.
T
he
e
xi
s
ti
ng
w
or
ks
s
uc
h
a
s
li
ke
O
ts
u
th
r
e
s
hol
di
ng
a
r
e
c
om
p
a
r
e
d
w
it
h
th
e
pr
opos
e
d
m
ul
ti
-
th
r
e
s
hol
di
ng
te
c
hni
que
.
T
h
e
s
ugge
s
te
d
m
ul
ti
-
th
r
e
s
hol
di
ng
te
c
hni
que
a
tt
a
in
s
th
e
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
of
99.57%
,
0.92,
0.95
,
a
nd
0.99,
c
or
r
e
s
pondingl
y
th
e
s
e
r
e
s
ul
ts
s
how
e
f
f
e
c
ti
ve
out
c
om
e
s
.
T
a
bl
e
1. P
e
r
f
or
m
a
nc
e
a
na
ly
s
i
s
f
or
s
e
gm
e
nt
a
ti
on
M
e
t
hods
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
s
c
or
e
E
dge
-
ba
s
e
d
91.23
0.82
0.84
0.88
R
e
gi
on
-
ba
s
e
d
93.06
0.83
0.88
0.90
A
da
pt
i
ve
t
hr
e
s
hol
di
ng
95.28
0.86
0.91
0.93
O
t
s
u t
hr
e
s
hol
di
ng
97.38
0.90
0.93
0.96
M
ul
t
i
-
t
hr
e
s
hol
di
ng
99.57
0.92
0.95
0.99
T
a
bl
e
2
pr
e
s
e
nt
s
th
e
V
G
G
-
16
a
r
c
hi
te
c
tu
r
e
w
it
h
va
r
io
us
pr
e
-
tr
a
in
e
d
m
ode
ls
f
or
S
L
R
on
I
S
L
da
ta
s
e
t.
T
he
V
G
G
-
16
a
tt
a
in
s
s
upe
r
io
r
pe
r
f
or
m
a
nc
e
a
s
c
om
pa
r
e
d
to
t
he
e
xi
s
ti
ng
m
e
th
ods
w
it
h
va
r
io
us
a
s
s
e
s
s
m
e
nt
m
e
tr
ic
e
s
.
T
he
e
xi
s
ti
ng
w
or
ks
s
u
c
h
a
s
li
ke
R
e
s
N
e
t5
0,
I
m
a
ge
N
e
t,
I
nc
e
pt
io
nN
e
t
a
nd
A
L
e
xN
e
t
a
r
e
c
om
pa
r
e
d
w
it
h
th
e
pr
opos
e
d
te
c
hni
que
.
T
he
pr
opos
e
d
V
G
G
-
16
te
c
h
ni
que
s
im
ul
ta
ne
ous
ly
a
tt
a
in
s
th
e
a
c
c
ur
a
c
y,
pr
e
c
is
io
n, r
e
c
a
ll
,
a
nd F
1
-
s
c
or
e
of
99.57%
, 0.92, 0.95
,
a
nd 0.99 a
nd t
he
s
e
r
e
s
ul
t
s
s
how
s
e
f
f
e
c
ti
ve
out
c
om
e
s
.
T
a
bl
e
3
e
xhi
bi
t
s
th
e
r
e
s
ul
t
of
th
e
pr
opos
e
d
B
le
nde
d
E
ns
e
m
bl
e
M
L
m
e
th
od
w
it
h
va
r
io
us
c
la
s
s
if
ic
a
ti
on
m
ode
ls
f
or
S
L
R
on
I
S
L
da
ta
s
e
t.
T
he
pr
opos
e
d
a
ppr
oa
c
h
a
tt
a
in
s
s
ig
ni
f
ic
a
nt
pe
r
f
or
m
a
nc
e
a
s
c
om
pa
r
e
d
to
th
e
e
xi
s
ti
ng
m
e
th
ods
w
it
h
va
r
io
us
a
s
s
e
s
s
m
e
nt
m
e
tr
ic
e
s
.
T
he
e
xi
s
ti
ng
w
or
ks
s
uc
h
a
s
li
ke
K
N
N
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Si
gn l
anguage
r
e
c
ogni
ti
on and c
la
s
s
if
ic
at
io
n us
in
g bl
e
nde
d e
n
s
e
m
bl
e
m
ac
hi
ne
le
a
r
ni
ng
(
A
k
as
h R
aj
an R
ai
)
2041
N
B
,
R
F
,
a
nd
S
V
M
a
r
e
c
om
pa
r
e
d
w
it
h
th
e
pr
opos
e
d
bl
e
nde
d
e
ns
e
m
bl
e
M
L
te
c
hni
que
.
T
he
pr
opos
e
d
m
e
th
od
a
tt
a
in
s
th
e
a
c
c
ur
a
c
y,
pr
e
c
is
io
n,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
of
99.57%
,
0.92,
0.95
,
a
nd
0.99,
r
e
s
pe
c
ti
ve
ly
.
T
he
bl
e
nde
d e
ns
e
m
bl
e
M
L
m
e
th
od a
c
c
om
pl
is
he
s
e
f
f
e
c
ti
ve
out
c
om
e
s
w
he
n c
om
pa
r
e
d t
o i
ndi
vi
dua
l
a
ppr
oa
c
he
s
.
T
a
bl
e
2. P
e
r
f
or
m
a
nc
e
a
na
ly
s
i
s
f
or
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
M
e
t
hods
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
s
c
or
e
R
e
s
N
e
t
50
94.29
0.85
0.89
0.93
I
m
a
ge
N
e
t
95.03
0.87
0.91
0.94
I
nc
e
pt
i
onN
e
t
96.29
0.88
0.93
0.96
A
l
e
xN
e
t
98.47
0.91
0.94
0.97
VGG
-
16
99.57
0.92
0.95
0.99
T
a
bl
e
3. P
e
r
f
or
m
a
nc
e
a
na
ly
s
i
s
f
or
c
la
s
s
if
ic
a
ti
on r
e
s
ul
t
s
on I
S
L
da
ta
s
e
t
M
e
t
hods
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
s
c
or
e
KNN
95.74
0.86
0.90
0.94
NB
96.38
0.87
0.92
0.95
RF
97.34
0.88
0.93
0.97
S
V
M
98.98
0.91
0.94
0.98
B
l
e
nde
d E
ns
e
m
bl
e
M
L
99.57
0.92
0.95
0.99
4.2.
C
om
p
ar
at
iv
e
an
al
ys
is
T
a
bl
e
4
pr
e
s
e
nt
s
th
e
c
om
pa
r
a
ti
ve
r
e
s
ul
ts
of
th
e
pr
opos
e
d
w
or
k
w
it
h
e
xi
s
ti
ng
w
or
ks
.
T
h
e
pr
opos
e
d
bl
e
nde
d
e
ns
e
m
bl
e
M
L
is
va
li
da
te
d w
it
h
m
e
th
od,
da
ta
s
e
ts
,
a
c
c
ur
a
c
y,
pr
e
c
is
io
n
,
a
nd
r
e
c
a
ll
f
or
I
S
L
da
ta
s
e
t.
T
he
e
xi
s
ti
ng w
or
ks
s
uc
h a
s
[
17]
, [
19]
‒
[
21]
a
r
e
us
e
d t
o a
na
ly
z
e
i
n c
o
nt
r
a
s
t
to
t
he
pr
opos
e
d m
e
th
od.
T
a
bl
e
4. C
om
pa
r
a
ti
ve
r
e
s
ul
ts
of
pr
opos
e
d w
or
k w
it
h e
xi
s
ti
ng w
or
k on I
S
L
da
ta
s
e
t
M
e
t
hods
A
c
c
ur
a
c
y (
%
)
P
r
e
c
i
s
i
on
R
e
c
a
l
l
F1
-
s
c
or
e
S
V
M
[
17]
90.1
N
/
A
N
/
A
N
/
A
C
N
N
-
D
S
C
[
19]
92.43
0.89
0.94
0.98
VGG
-
16+B
i
-
L
S
T
M
[
20]
98.56
N
/
A
0.9843
0.9758
C
N
N
[
21]
99.36
N
/
A
N
/
A
N
/
A
P
r
opos
e
d
bl
e
nde
d e
ns
e
m
bl
e
ML
99.57
0.92
0.95
0.99
4.3.
D
is
c
u
s
s
io
n
T
hi
s
s
e
c
ti
on
il
lu
s
tr
a
te
s
th
e
li
m
it
a
ti
ons
of
e
xi
s
ti
ng
w
or
ks
a
nd
e
xpl
a
in
s
how
th
e
pr
opos
e
d
m
e
th
od
ove
r
c
om
e
s
s
u
c
h
li
m
it
a
ti
ons
. T
he
li
m
it
a
ti
ons
of
th
e
e
xi
s
ti
ng
w
or
ks
s
uc
h
a
s
S
V
M
[
17]
onl
y
pe
r
f
or
m
e
d
w
it
h
th
e
5
s
ta
ti
c
s
ym
bol
s
f
or
S
L
R
.
S
V
M
-
C
N
N
[
18]
ut
il
iz
e
d
th
e
la
r
ge
c
lu
s
te
r
da
ta
f
or
e
nha
nc
in
g
th
e
m
ode
l’
s
pe
r
f
or
m
a
nc
e
. T
he
C
N
N
-
de
pt
h s
e
pa
r
a
bl
e
c
onvolut
io
n (
D
S
C
)
[
19
]
ha
d poor
pe
r
f
or
m
a
nc
e
i
n c
a
s
e
of
r
e
c
ogni
ti
on
of
th
e
s
im
il
a
r
ge
s
tu
r
e
s
.
C
N
N
[
21]
a
ppr
oa
c
h
r
e
c
ogni
z
e
d
onl
y
th
e
s
ta
ti
c
s
ig
ns
.
H
ow
e
ve
r
,
th
e
pr
opos
e
d
bl
e
nde
d
e
ns
e
m
bl
e
M
L
m
e
th
od
ta
c
kl
e
s
th
e
s
e
li
m
it
a
ti
ons
.
T
he
pr
opos
e
d
a
tt
a
in
s
be
tt
e
r
r
e
s
ul
ts
a
nd
a
c
c
om
pl
is
he
s
a
n
a
c
c
ur
a
c
y
of
99.57%
,
pr
e
c
is
io
n
of
0.92,
r
e
c
a
ll
of
0.95
,
a
nd
F
1
-
s
c
or
e
of
0.99.
O
n
th
e
ot
he
r
ha
nd,
th
e
e
xi
s
ti
ng
w
or
k
C
N
N
-
D
S
C
[
19]
a
tt
a
in
s
th
e
a
c
c
ur
a
c
y
of
92.43%
.
T
he
r
e
by
,
th
e
e
xi
s
ti
ng
w
or
ks
e
xhi
bi
t
poor
pe
r
f
or
m
a
nc
e
in
c
ont
r
a
s
t
to
t
he
pr
opos
e
d
bl
e
nde
d e
n
s
e
m
bl
e
M
L
m
e
th
od.
5.
C
O
N
C
L
U
S
I
O
N
T
he
a
im
of
th
is
r
e
s
e
a
r
c
h
i
s
to
a
c
c
om
pl
is
h
th
e
r
e
a
l
-
ti
m
e
r
e
c
ogni
ti
on
of
a
lp
ha
be
ts
a
nd
num
e
r
ic
a
l
va
lu
e
s
in
I
S
L
by
pr
ovi
di
ng
th
e
i
m
a
ge
pr
oc
e
s
s
in
g
-
ba
s
e
d
r
e
c
ogni
ti
on
m
e
th
od.
H
e
nc
e
in
th
is
r
e
s
e
a
r
c
h,
th
e
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
ppr
oa
c
h
is
in
tr
oduc
e
d
f
or
I
S
L
r
e
c
ogni
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on.
I
n
th
is
in
tr
oduc
e
d
m
e
th
od, t
he
bl
e
nde
d e
ns
e
m
bl
e
l
e
a
r
ni
ng i
nvol
ve
s
f
our
m
a
in
M
L
a
ppr
oa
c
he
s
s
uc
h a
s
K
N
N
, N
B
, R
F
, a
nd S
V
M
.
A
s
a
r
e
s
ul
t,
th
is
r
e
s
e
a
r
c
h
e
nha
nc
e
s
th
e
S
L
R
a
c
c
ur
a
c
y,
a
s
w
e
ll
a
s
m
in
im
iz
e
s
th
e
m
ode
l’
s
c
om
put
a
ti
ona
l
c
om
pl
e
xi
ty
by
ut
il
iz
in
g
th
e
e
f
f
e
c
ti
ve
bl
e
nde
d
e
ns
e
m
bl
e
M
L
a
p
pr
oa
c
h.
T
he
e
f
f
ic
ie
nc
y
of
th
e
m
ode
l
w
hi
c
h
i
s
tr
a
in
e
d
a
nd
te
s
te
d
on
th
e
I
S
L
c
ol
le
c
te
d
da
ta
s
e
t
is
s
tu
di
e
d.
T
h
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
r
e
ve
a
l
th
a
t
th
e
pr
opos
e
d
m
e
th
od
a
c
c
om
pl
is
he
s
th
e
a
c
c
ur
a
c
y
of
99.57%
,
pr
e
c
is
io
n
of
0.92,
r
e
c
a
ll
of
0.95
,
a
nd
F
1
-
s
c
or
e
of
0.99
r
e
s
pe
c
ti
ve
ly
.
I
n
f
ut
ur
e
w
o
r
k,
th
e
pe
r
f
or
m
a
nc
e
of
th
e
pr
opos
e
d
m
e
th
od
w
il
l
be
e
nha
nc
e
d
f
or
a
num
be
r
o
f
r
e
a
l
-
ti
m
e
a
ppl
ic
a
ti
ons
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
3
,
J
une
2025
:
2035
-
2043
2042
F
U
N
D
I
N
G
I
N
F
O
R
M
A
T
I
O
N
A
ut
hor
s
s
ta
te
no f
undi
ng i
nvol
ve
d.
A
U
T
H
O
R
C
O
N
T
R
I
B
U
T
I
O
N
S
S
T
A
T
E
M
E
N
T
T
hi
s
jo
ur
na
l
us
e
s
th
e
C
ont
r
ib
ut
or
R
ol
e
s
T
a
xonomy
(
C
R
e
di
T
)
to
r
e
c
ogni
z
e
in
di
vi
dua
l
a
ut
hor
c
ont
r
ib
ut
io
ns
, r
e
duc
e
a
ut
hor
s
hi
p di
s
put
e
s
,
a
nd f
a
c
il
it
a
te
c
ol
la
bo
r
a
ti
on.
N
am
e
o
f
A
u
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
A
ka
s
h R
a
ja
n
R
a
i
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
S
uj
a
ta
R
a
je
s
h K
a
du
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
l
i
z
a
t
i
on
M
:
M
e
t
hodol
ogy
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
da
t
i
on
Fo
:
Fo
r
m
a
l
a
na
l
ys
i
s
I
:
I
nve
s
t
i
ga
t
i
on
R
:
R
e
s
our
c
e
s
D
:
D
a
t
a
C
ur
a
t
i
on
O
:
W
r
i
t
i
ng
-
O
r
i
gi
na
l
D
r
a
f
t
E
:
W
r
i
t
i
ng
-
R
e
vi
e
w
&
E
di
t
i
ng
Vi
:
Vi
s
ua
l
i
z
a
t
i
on
Su
:
Su
pe
r
vi
s
i
on
P
:
P
r
oj
e
c
t
a
dm
i
ni
s
t
r
a
t
i
on
Fu
:
Fu
ndi
ng a
c
qui
s
i
t
i
on
C
O
N
F
L
I
C
T
O
F
I
N
T
E
R
E
S
T
S
T
A
T
E
M
E
N
T
A
ut
hor
s
s
ta
te
no c
onf
li
c
t
of
i
nt
e
r
e
s
t.
D
A
T
A
A
V
A
I
L
A
B
I
L
I
T
Y
T
he
da
ta
th
a
t
s
uppor
t
th
e
f
in
di
ngs
of
th
is
s
tu
dy
a
r
e
ope
nl
y
a
va
il
a
bl
e
in
I
ndi
a
n
S
ig
n
L
a
ngua
ge
a
t
ht
tp
s
:/
/ww
w
.ka
ggl
e
.c
om
/d
a
ta
s
e
ts
/p
r
a
th
um
a
r
ik
e
r
i/
in
di
a
n
-
s
ig
n
-
la
ngua
ge
-
is
l,
r
e
f
e
r
e
nc
e
num
be
r
[
22]
.
R
E
F
E
R
E
N
C
E
S
[
1]
M
.
I
.
S
a
l
e
e
m
,
A
.
S
i
ddi
qui
,
S
.
N
oo
r
,
M
.
A
.
L
uque
-
N
i
e
t
o,
a
nd
E
.
N
a
va
-
B
a
r
o,
“
A
m
a
c
hi
ne
l
e
a
r
ni
ng
ba
s
e
d
f
ul
l
dupl
e
x
s
ys
t
e
m
s
uppor
t
i
ng m
ul
t
i
pl
e
s
i
gn l
a
ngua
ge
s
f
or
t
he
de
a
f
a
nd m
ut
e
,”
A
ppl
i
e
d Sc
i
e
nc
e
s
, v
ol
. 13, no. 5, 2023, doi
:
10.3390/
a
pp13053114.
[
2]
G
.
H
.
S
a
m
a
a
n
e
t
al
.
,
“
M
e
di
a
P
i
pe
’
s
l
a
ndm
a
r
ks
w
i
t
h
R
N
N
f
or
dyna
m
i
c
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on,”
E
l
e
c
t
r
oni
c
s
,
vol
.
11,
no.
19
,
2022, doi
:
10.3390/
e
l
e
c
t
r
oni
c
s
11193228.
[
3]
M
.
S
.
A
m
i
n
a
nd
S
.
T
.
H
.
R
i
z
vi
,
“
S
i
gn
ge
s
t
ur
e
c
l
a
s
s
i
f
i
c
a
t
i
on
a
nd
r
e
c
ogni
t
i
on
u
s
i
ng
m
a
c
hi
ne
l
e
a
r
ni
ng,”
C
y
be
r
n
e
t
i
c
s
and
Sy
s
t
e
m
s
,
vol
. 54, no. 5, pp. 604
–
618, 2023, doi
:
10.1080/
01969722.2022.2067634.
[
4]
D
.
K
ot
ha
di
ya
,
C
.
B
ha
t
t
,
K
.
S
a
pa
r
i
ya
,
K
.
P
a
t
e
l
,
A
.
B
.
G
i
l
-
G
onz
á
l
e
z
,
a
nd
J
.
M
.
C
or
c
ha
do,
“
D
e
e
ps
i
gn:
s
i
gn
l
a
ngua
g
e
de
t
e
c
t
i
on
a
nd
r
e
c
ogni
t
i
on us
i
ng de
e
p l
e
a
r
ni
ng,”
E
l
e
c
t
r
oni
c
s
, vol
. 11, no. 11, 2022, doi
:
10.339
0/
e
l
e
c
t
r
oni
c
s
11111780.
[
5]
C
.
L
u,
M
.
K
oz
a
k
a
i
,
a
nd
L
.
J
i
ng,
“
S
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on
w
i
t
h
m
ul
t
i
m
oda
l
s
e
ns
or
s
a
nd
de
e
p
l
e
a
r
ni
ng
m
e
t
hods
,”
E
l
e
c
t
r
oni
c
s
,
vol
. 12, no. 23, 2023, doi
:
10.3390/
e
l
e
c
t
r
oni
c
s
12234827.
[
6]
Q
.
M
.
A
r
e
e
b,
M
a
r
ya
m
,
M
.
N
a
de
e
m
,
R
.
A
l
r
ooba
e
a
,
a
nd
F
.
A
nw
e
r
,
“
H
e
l
pi
ng
he
a
r
i
ng
-
i
m
pa
i
r
e
d
i
n
e
m
e
r
ge
nc
y
s
i
t
ua
t
i
ons
:
a
de
e
p
l
e
a
r
ni
ng
-
ba
s
e
d a
ppr
oa
c
h,”
I
E
E
E
A
c
c
e
s
s
, vol
. 10, pp. 8502
–
8517, 2022, doi
:
10.1109/
A
C
C
E
S
S
.2022.3142918.
[
7]
O
.
M
.
S
i
nc
a
n
a
nd
H
.
Y
.
K
e
l
e
s
,
“
U
s
i
ng
m
ot
i
on
hi
s
t
or
y
i
m
a
ge
s
w
i
t
h
3D
c
onvol
ut
i
ona
l
ne
t
w
or
ks
i
n
i
s
ol
a
t
e
d
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on,”
I
E
E
E
A
c
c
e
s
s
, vol
. 10, pp. 18608
–
18618, 2022, doi
:
10.1109/
A
C
C
E
S
S
.2022.3151362.
[
8]
S
.
D
a
s
,
M
.
S
.
I
m
t
i
a
z
,
N
.
H
.
N
e
om
,
N
.
S
i
ddi
que
,
a
nd
H
.
W
a
ng,
“
A
hybr
i
d
a
ppr
oa
c
h
f
or
B
a
ngl
a
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on
us
i
ng
de
e
p
t
r
a
ns
f
e
r
l
e
a
r
ni
ng
m
ode
l
w
i
t
h
r
a
ndom
f
o
r
e
s
t
c
l
a
s
s
i
f
i
e
r
,”
E
x
pe
r
t
Sy
s
t
e
m
s
w
i
t
h
A
ppl
i
c
at
i
ons
,
vol
.
213,
2023,
doi
:
10.1016/
j
.e
s
w
a
.2022.118914.
[
9]
J
. B
or
a
, S
. D
e
hi
ngi
a
, A
. B
or
ua
h, A
.
A
. C
he
t
i
a
,
a
nd D
. G
ogoi
, “
R
e
a
l
-
t
i
m
e
a
s
s
a
m
e
s
e
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on u
s
i
ng m
e
di
a
pi
pe
a
n
d
de
e
p l
e
a
r
ni
ng,”
P
r
oc
e
di
a C
om
put
e
r
Sc
i
e
nc
e
, vol
. 218, pp. 1384
–
1393, 2022, do
i
:
10.1016/
j
.pr
oc
s
.2023.01.117.
[
10]
M
.
H
r
úz
,
I
.
G
r
ube
r
,
J
.
K
a
ni
s
,
M
.
B
ohá
č
e
k,
M
.
H
l
a
vá
č
,
a
nd
Z
.
K
r
ňoul
,
“
E
ns
e
m
bl
e
i
s
w
ha
t
w
e
ne
e
d:
i
s
ol
a
t
e
d
s
i
gn
r
e
c
ogni
t
i
on
e
di
t
i
on,”
Se
ns
or
s
, vol
. 22, no. 13, 2022, doi
:
10.3390/
s
22135043.
[
11]
R
.
R
a
s
t
goo,
K
.
K
i
a
ni
,
a
nd
S
.
E
s
c
a
l
e
r
a
,
“
R
e
a
l
-
t
i
m
e
i
s
ol
a
t
e
d
ha
nd
s
i
gn
l
a
ngu
a
ge
r
e
c
ogni
t
i
on
us
i
ng
de
e
p
ne
t
w
or
ks
a
nd
S
V
D
,”
J
our
nal
of
A
m
bi
e
nt
I
nt
e
l
l
i
ge
nc
e
and H
um
ani
z
e
d C
om
put
i
ng
, vol
. 13, no. 1, pp.
591
–
611, 2022, doi
:
10.1007/
s
12652
-
021
-
02920
-
8.
[
12]
D
.
K
ot
ha
di
ya
,
C
.
B
ha
t
t
,
K
.
S
a
pa
r
i
ya
,
K
.
P
a
t
e
l
,
A
.
-
B
.
G
i
l
-
G
onz
á
l
e
z
,
a
nd
J
.
M
.
C
or
c
ha
do
,
“
D
e
e
ps
i
gn:
S
i
gn
l
a
ngua
ge
d
e
t
e
c
t
i
on
a
n
d
r
e
c
ogni
t
i
on us
i
ng de
e
p l
e
a
r
ni
ng
,”
E
l
e
c
t
r
oni
c
s
, vol
. 1
1
, no.
11
, 2022, doi
:
10.115
91/
i
j
e
c
e
.v12i
3.pp2996
-
3004.
[
13]
S
.
G
üne
y
a
nd
M
.
E
r
kuş
,
“
A
r
e
a
l
-
t
i
m
e
a
ppr
oa
c
h
t
o
r
e
c
ogni
t
i
on
of
T
ur
ki
s
h
s
i
gn
l
a
ngua
ge
by
us
i
ng
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
,
”
N
e
ur
al
C
om
put
i
ng and A
ppl
i
c
at
i
ons
, vol
. 34, no. 5, pp. 4069
–
4079, 2022, doi
:
10.3390/
e
l
e
c
t
r
oni
c
s
11111780
.
[
14]
A
. S
i
ngh, A
. W
a
dha
w
a
n,
M
. R
a
khr
a
, U
. M
i
t
t
a
l
, A
.
A
l
A
hda
l
, a
nd S
.
K
. J
ha
, “
I
ndi
a
n s
i
gn l
a
ngua
ge
r
e
c
ogni
t
i
on
s
ys
t
e
m
f
or
dyna
m
i
c
s
i
gns
,”
2022
10t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
n
c
e
on
R
e
l
i
abi
l
i
t
y
,
I
nf
oc
om
T
e
c
h
nol
ogi
e
s
and
O
pt
i
m
i
z
at
i
on
(
T
r
e
nds
and
F
ut
ur
e
D
i
r
e
c
t
i
ons
)
(
I
C
R
I
T
O
)
,
N
oi
da
, I
ndi
a
, 2022, pp. 1
-
6, doi
:
10.1109
/
I
C
R
I
T
O
56286.
2022.9964940
.
[
15]
S
.
K
.
S
i
ngh
a
n
d
A
.
C
h
a
t
u
r
ve
d
i
,
“
A
r
e
l
i
a
bl
e
a
n
d
e
f
f
i
c
i
e
n
t
m
a
c
h
i
ne
l
e
a
r
ni
ng
pi
pe
l
i
n
e
f
o
r
a
m
e
r
i
c
a
n
s
i
gn
l
a
ng
ua
ge
ge
s
t
u
r
e
r
e
c
og
ni
t
i
o
n
us
i
n
g
E
M
G
s
e
ns
o
r
s
,”
M
u
l
t
i
m
e
di
a
T
oo
l
s
and
A
p
pl
i
c
a
t
i
ons
,
v
ol
.
8
2,
no
. 1
5,
pp
.
238
33
–
238
71
, 2
02
3,
do
i
:
10
.1
007
/
s
1
104
2
-
02
2
-
14
11
7
-
y.
[
16]
A
.
S
.
N
a
ndhi
ni
,
D
.
S
.
R
oopa
n,
S
.
S
hi
ya
a
m
,
a
nd
S
.
Y
oge
s
h,
“
S
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on
us
i
ng
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
J
our
nal
of
P
hy
s
i
c
s
:
C
onf
e
r
e
nc
e
S
e
r
i
e
s
, vol
. 1916, no. 1, 2021, doi
:
10.1088/
1742
-
6596/
1916/
1/
012091.
[
17]
P
.
K
.
A
t
hi
r
a
,
C
. J
. S
r
u
t
hi
,
a
nd
A
.
L
i
j
i
ya
,
“
A
s
i
gne
r
i
nde
pe
nde
n
t
s
i
g
n
l
a
n
gua
ge
r
e
c
ogni
t
i
on
w
i
t
h
c
o
-
a
r
t
i
c
ul
a
t
i
on
e
l
i
m
i
na
t
i
o
n f
r
o
m
l
i
ve
vi
de
os
:
a
n
I
ndi
a
n
s
c
e
na
r
i
o,”
J
our
nal
o
f
K
i
ng
Sau
d
U
ni
v
e
r
s
i
t
y
-
C
om
pu
t
e
r
and
I
nf
or
m
at
i
on
Sc
i
e
nc
e
s
,
v
ol
.
34
,
no.
3,
p
p.
771
–
781,
2022, d
oi
:
10.10
16/
j
.j
ks
uc
i
.2019
.05.0
02.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
Si
gn l
anguage
r
e
c
ogni
ti
on and c
la
s
s
if
ic
at
io
n us
in
g bl
e
nde
d e
n
s
e
m
bl
e
m
ac
hi
ne
le
a
r
ni
ng
(
A
k
as
h R
aj
an R
ai
)
2043
[
18]
S
.
K
a
t
oc
h,
V
.
S
i
ngh,
a
nd
U
.
S
.
T
i
w
a
r
y,
“
I
ndi
a
n
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on
s
ys
t
e
m
us
i
ng
S
U
R
F
w
i
t
h
S
V
M
a
nd
C
N
N
,”
A
r
r
ay
,
vol
. 14, 2022, doi
:
10.1016/
j
.a
r
r
a
y.2022.100141.
[
19]
S
.
S
ha
r
m
a
a
nd
S
.
S
i
ngh,
“
R
e
c
ogni
t
i
on
of
i
ndi
a
n
s
i
gn
l
a
ngua
ge
(
I
S
L
)
us
i
ng
de
e
p
l
e
a
r
ni
ng
m
ode
l
,”
W
i
r
e
l
e
s
s
P
e
r
s
ona
l
C
om
m
uni
c
at
i
ons
, vol
. 123, no. 1, pp. 671
–
692, 2022, doi
:
10.1007/
s
11277
-
021
-
09152
-
1.
[
20]
B
.
N
a
t
a
r
a
j
a
n
e
t
al
.
,
“
D
e
ve
l
opm
e
nt
of
a
n
e
nd
-
to
-
e
nd
de
e
p
l
e
a
r
ni
ng
f
r
a
m
e
w
or
k
f
or
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on,
t
r
a
ns
l
a
t
i
on,
a
nd
vi
de
o
ge
ne
r
a
t
i
on,”
I
E
E
E
A
c
c
e
s
s
, vol
. 10, pp. 104358
–
104374, 2022, doi
:
10.1109/
A
C
C
E
S
S
.2022.3210543.
[
21]
U
.
N
a
ndi
,
A
.
G
hor
a
i
,
M
.
M
.
S
i
ngh,
C
.
C
h
a
ngda
r
,
S
.
B
ha
kt
a
,
a
nd
R
.
K
.
P
a
l
,
“
I
ndi
a
n
s
i
gn
l
a
ngua
ge
a
l
pha
be
t
r
e
c
ogni
t
i
on
s
ys
t
e
m
us
i
ng
C
N
N
w
i
t
h
di
f
f
G
r
a
d
opt
i
m
i
z
e
r
a
nd
s
t
oc
ha
s
t
i
c
pool
i
ng,”
M
ul
t
i
m
e
di
a
T
ool
s
and
A
ppl
i
c
at
i
ons
,
vol
.
82,
no.
7,
pp.
9627
–
9648
,
2023, doi
:
10.1007/
s
11042
-
021
-
11595
-
4.
[
22]
P
.
A
r
i
ke
r
i
,
“
I
ndi
a
n
s
i
gn
l
a
ngua
ge
da
t
a
s
e
t
,”
K
aggl
e
.
2020.
[
O
nl
i
ne
]
.
A
va
i
l
a
bl
e
:
ht
t
ps
:
/
/
w
w
w
.ka
ggl
e
.c
om
/
da
t
a
s
e
t
s
/
pr
a
t
hum
a
r
i
ke
r
i
/
i
ndi
a
n
-
s
i
gn
-
l
a
ngua
ge
-
i
s
l
/
da
t
a
[
23]
C
.
C
he
n
e
t
al
.
,
“
E
nha
nc
i
ng
M
R
i
m
a
ge
s
e
gm
e
nt
a
t
i
on
w
i
t
h
r
e
a
l
i
s
t
i
c
a
dve
r
s
a
r
i
a
l
da
t
a
a
ugm
e
nt
a
t
i
on,”
M
e
di
c
al
I
m
age
A
nal
y
s
i
s
,
vol
. 82, 2022, doi
:
10.1016/
j
.m
e
di
a
.2022.102597.
[
24]
W
.
M
.
S
a
l
a
m
a
a
nd
M
.
H
.
A
l
y,
“
D
e
e
p
l
e
a
r
ni
ng
i
n
m
a
m
m
ogr
a
phy
i
m
a
ge
s
s
e
gm
e
nt
a
t
i
on
a
nd
c
l
a
s
s
i
f
i
c
a
t
i
on:
a
ut
om
a
t
e
d
C
N
N
a
ppr
oa
c
h,”
A
l
e
x
andr
i
a E
ngi
ne
e
r
i
ng J
ou
r
nal
, vol
. 60, no. 5, pp. 4701
–
4709, 2021, doi
:
10.1016/
j
.a
e
j
.2021.03.048.
[
25]
S
. S
ha
r
m
a
, K
. G
ul
e
r
i
a
, S
. T
i
w
a
r
i
, a
nd S
. K
um
a
r
, “
A
de
e
p l
e
a
r
ni
ng ba
s
e
d c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k m
ode
l
w
i
t
h V
G
G
16 f
e
a
t
ur
e
e
xt
r
a
c
t
or
f
or
t
he
de
t
e
c
t
i
on
of
A
l
z
he
i
m
e
r
di
s
e
a
s
e
us
i
ng
M
R
I
s
c
a
n
s
,”
M
e
as
ur
e
m
e
nt
:
Se
ns
or
s
,
vol
.
24,
2022,
doi
:
10.1016/
j
.m
e
a
s
e
n.2022.100506.
[
26]
G
.
S
.
N
i
j
a
guna
,
J
.
A
.
B
a
bu,
B
.
D
.
P
a
r
a
m
e
s
ha
c
ha
r
i
,
R
.
P
.
de
P
r
a
do,
a
nd
J
.
F
r
nda
,
“
Q
ua
nt
um
f
r
ui
t
f
l
y
a
l
gor
i
t
hm
a
nd
R
e
s
N
e
t
50
-
V
G
G
16 f
or
m
e
di
c
a
l
di
a
gnos
i
s
,”
A
ppl
i
e
d Sof
t
C
om
put
i
ng
, vol
. 136, 2023, doi
:
10.1016/
j
.a
s
oc
.2023.110055.
[
27]
F
.
U
t
a
m
i
ni
ngr
um
,
I
K
.
S
om
a
w
i
r
a
t
a
,
a
nd
G
.
D
.
N
a
vi
r
i
,
“
A
l
pha
be
t
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on
us
i
ng
k
-
ne
a
r
e
s
t
ne
i
ghbor
opt
i
m
i
z
a
t
i
on
,”
J
our
nal
of
C
om
put
e
r
s
, vol
.
14
, no.
1
, pp.
63
–
70
, 20
19
, doi
:
10.1
7706/
j
c
p.14.1 63
-
70
.
[
28]
J
. K
. A
l
w
a
n, D
. S
. J
a
a
f
a
r
, a
nd I
. R
. A
l
i
, “
D
i
a
b
e
t
e
s
di
a
gno
s
i
s
s
ys
t
e
m
u
s
i
ng m
odi
f
i
e
d N
a
i
ve
B
a
ye
s
c
l
a
s
s
i
f
i
e
r
,”
I
ndone
s
i
an J
ou
r
nal
of
E
l
e
c
t
r
i
c
al
E
ngi
ne
e
r
i
ng and C
om
put
e
r
S
c
i
e
nc
e
, vol
. 28, no. 3, pp. 1766
–
1774, 2
022, doi
:
10.11591/
i
j
e
e
c
s
.v28.i
3.pp1766
-
1774.
[
29]
S
.
R
houa
s
,
A
.
E
l
A
t
t
a
oui
,
a
nd
N
.
E
l
H
a
m
i
,
“
E
nha
nc
i
ng
c
ur
r
e
nc
y
pr
e
di
c
t
i
on
i
n
i
nt
e
r
na
t
i
ona
l
e
-
c
om
m
e
r
c
e
:
B
a
ye
s
i
a
n
-
opt
i
m
i
z
e
d
r
a
ndom
f
or
e
s
t
a
ppr
oa
c
h
us
i
ng
t
he
K
l
a
r
na
da
t
a
s
e
t
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
E
l
e
c
t
r
i
c
al
and
C
om
put
e
r
E
ngi
ne
e
r
i
ng
,
vol
.
14,
no.
3,
pp. 3177
–
3186, 2024, doi
:
10.11591/
i
j
e
c
e
.v14i
3.pp3177
-
3186.
[
30]
M
.
B
a
n
s
a
l
a
nd
S
.
G
upt
a
,
“
D
e
t
e
c
t
i
on
a
nd
r
e
c
ogni
t
i
on
of
ha
nd
ge
s
t
ur
e
s
f
or
i
ndi
a
n
s
i
gn
l
a
ngua
ge
r
e
c
ogni
t
i
on
s
ys
t
e
m
,”
2021
6t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
Si
gnal
P
r
oc
e
s
s
i
ng,
C
om
put
i
ng
and
C
ont
r
ol
(
I
SP
C
C
)
,
S
ol
a
n,
I
ndi
a
,
2021,
pp.
136
-
140,
doi
:
10.1109/
I
S
P
C
C
53510.2021.9609448
.
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Akash
Rajan
Rai
received
a
bachelor
degree
in
information
&
tec
hnology
from
the Univer
sity of Mumbai
in 2020. He
is curren
tly pursuing
his
master
’s
d
egree
in
information
technology
fro
m
Terna
Engineering
College.
His
areas
of
research
in
clude
machine
learning,
data
analysis,
and
web
development
frameworks
.
He
can
b
e
contacted
at
email:
akashrai93
2@
gmail.co
m.
Sujata
Rajesh
Kadu
received
a
Ph.D.
degree
in
electroni
cs
and
telecommunic
atio
n
from
the
University
of
Mumbai
in
2022.
She
is
currently
an
Assistant
Profes
sor
at
the
University
of
Mumbai,
Terna
Engineering
College
.
Her
areas
of
research
include
signal processing, ima
ge segmentation
and classifica
tion,
obje
ct
-
based image analysis,
and mul
tiresol
ution
segmentat
ion
. She can be contacted a
t email: sujatakadu@
gmail.com.
Evaluation Warning : The document was created with Spire.PDF for Python.