T
E
L
KO
M
N
I
KA
T
e
lec
om
m
u
n
icat
ion
,
Com
p
u
t
i
n
g,
E
lec
t
r
on
ics
an
d
Cont
r
ol
Vol.
1
8
,
No.
3
,
J
une
2020
,
pp.
1
2
9
2
~
1
30
1
I
S
S
N:
1693
-
6930,
a
c
c
r
e
dit
e
d
F
ir
s
t
G
r
a
de
by
Ke
me
nr
is
tekdikti
,
De
c
r
e
e
No:
21/E
/KP
T
/2018
DO
I
:
10.
12928/
T
E
L
KO
M
NI
KA
.
v1
8
i
3
.
15311
1292
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
A
u
t
om
a
t
e
d
B
an
gl
a
si
gn
l
an
g
u
age
t
r
a
n
sl
at
io
n
syst
e
m
f
o
r
al
p
h
ab
e
t
s
b
y m
e
a
n
s of
M
ob
il
e
N
e
t
T
az
k
ia
M
im
Angon
a
,
A
.
S
.
M
.
S
ia
m
u
z
z
am
an
S
h
aon
,
Kazi
T
ah
m
id
Ras
h
ad
Niloy
,
T
aj
b
ia
Karim
,
Z
ar
in
T
as
n
im
,
S
.
M
.
S
a
li
m
Re
z
a,
T
as
m
im
a
Nous
h
ib
a
M
ah
b
u
b
D
ep
ar
t
men
t
o
f
In
f
o
rmat
i
o
n
an
d
C
o
mmu
n
i
ca
t
i
o
n
T
ech
n
o
l
o
g
y
,
Ba
n
g
l
ad
e
s
h
U
n
i
v
ers
i
t
y
o
f
Pro
fe
s
s
i
o
n
al
s
,
Ban
g
l
ad
e
s
h
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
J
a
n
16
,
20
20
R
e
vis
e
d
F
e
b
7
,
2020
Ac
c
e
pted
F
e
b
28
,
2020
In
d
i
v
i
d
u
al
s
w
i
t
h
h
e
ari
n
g
a
n
d
s
p
e
ak
i
n
g
i
m
p
ai
rme
n
t
co
mmu
n
i
cat
e
u
s
i
n
g
s
i
g
n
l
an
g
u
a
g
e.
T
h
e
mo
v
eme
n
t
o
f
h
an
d
,
b
o
d
y
an
d
ex
p
r
es
s
i
o
n
s
o
f
face
are
t
h
e
mean
s
b
y
w
h
i
c
h
t
h
e
p
e
o
p
l
e,
w
h
o
are
u
n
ab
l
e
t
o
h
ear
an
d
s
p
eak
,
can
co
mmu
n
i
c
at
e.
Ban
g
l
a
s
i
g
n
a
l
p
h
ab
e
t
s
are
fo
rme
d
w
i
t
h
o
n
e
o
r
t
w
o
h
a
n
d
mo
v
eme
n
t
s
.
T
h
ere
ar
e
s
o
me
feat
u
res
w
h
i
ch
d
i
ffere
n
t
i
at
e
s
t
h
e
s
i
g
n
s
.
T
o
d
e
t
ec
t
an
d
rec
o
g
n
i
ze
t
h
e
s
i
g
n
s
,
a
n
al
y
zi
n
g
i
t
s
s
h
a
p
e
an
d
c
o
mp
ari
n
g
i
t
s
feat
u
res
i
s
n
eces
s
ary
.
T
h
i
s
p
a
p
er
ai
ms
t
o
p
r
o
p
o
s
e
a
mo
d
el
a
n
d
b
u
i
l
d
a
co
mp
u
t
er
s
y
s
t
emt
h
a
t
can
re
co
g
n
i
ze
Ban
g
l
a
Si
g
n
L
an
u
g
ag
e
al
p
h
ab
e
t
s
an
d
t
ran
s
l
at
e
t
h
em
t
o
co
rre
s
p
o
n
d
i
n
g
Ban
g
l
a
l
et
t
ers
b
y
mean
s
o
f
d
eep
co
n
v
o
l
u
t
i
o
n
al
n
eu
ra
l
n
et
w
o
r
k
(CN
N
).
CN
N
h
a
s
b
een
i
n
t
r
o
d
u
ced
i
n
t
h
i
s
mo
d
e
l
i
n
fo
rm
o
f
a
p
re
-
t
ra
i
n
e
d
mo
d
el
ca
l
l
ed
“M
o
b
i
l
e
N
et
”
w
h
i
ch
p
ro
d
u
ce
d
an
av
era
g
e
accu
racy
o
f
9
5
.
7
1
%
i
n
rec
o
g
n
i
z
i
n
g
3
6
Ba
n
g
l
a
Si
g
n
L
an
g
u
ag
e
al
p
h
ab
e
t
s
.
K
e
y
w
o
r
d
s
:
Ac
c
ur
a
c
y
B
a
ngla
s
ign
langua
ge
(
B
S
L
)
C
NN
C
onvolut
ion
M
obil
e
Ne
t
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
S
.
M
.
S
a
li
m
R
e
z
a
,
De
pa
r
tm
e
nt
of
I
nf
or
mat
ion
a
nd
C
omm
unica
ti
on
T
e
c
hnology
(
I
C
T
)
,
B
a
nglade
s
h
Unive
r
s
it
y
of
P
r
of
e
s
s
ionals
(
B
UP)
,
M
ir
pur
C
a
ntonm
e
nt,
Dha
ka
-
1216,
B
a
nglade
s
h
.
E
mail:
s
a
li
m4419@gm
a
il
.
c
om
1.
I
NT
RODU
C
T
I
ON
C
omm
unica
ti
on
is
one
of
human
s
oc
iety's
c
or
ne
r
s
tones
.
Huma
n
c
omm
unica
tes
by
the
mea
ns
of
s
poke
n
langua
ge
pr
im
a
r
i
ly.
T
he
pe
r
s
ons
who
a
r
e
una
ble
to
s
pe
a
k
a
nd
he
a
r
,
howe
ve
r
,
c
a
nnot
c
omm
unica
te
thi
s
wa
y.
T
he
r
e
f
or
e
,
many
c
ult
ur
e
s
ha
ve
de
ve
loped
their
own
langua
ge
f
or
c
omm
unica
ti
ng
with
the
s
pe
a
king
a
nd
he
a
r
ing
im
pa
ir
e
dpe
ople.
All
the
lette
r
s
a
nd
wor
ds
a
r
e
r
e
pr
e
s
e
nted
vis
ua
ll
y
to
the
im
pa
ir
e
d
pe
r
s
on.
T
his
langua
ge
is
c
a
ll
e
d
s
ign
langua
ge
.
S
ign
lang
ua
ge
mos
tl
y
c
ons
is
ts
of
s
pe
c
if
ic
ge
s
tur
e
s
made
with
one
or
two
ha
nds
whe
r
e
the
pos
it
ion
of
the
pa
lm
a
nd
plac
e
ment
of
f
inger
s
c
ombi
ne
d
make
up
l
e
tt
e
r
s
of
the
langua
ge
.
E
a
c
h
a
lphabe
t
ha
s
a
pr
e
de
ter
mi
ne
d
pa
tt
e
r
n
o
f
ha
nd
a
nd
f
inger
pos
it
ion.
T
he
va
r
iant
of
s
ing
langua
ge
f
or
B
a
ngla
langua
ge
is
r
e
f
e
r
r
e
d
to
a
s
th
e
B
a
ngla
S
ign
L
a
ngua
ge
(
B
S
L
)
.
F
igur
e
1
s
hows
the
ha
nd
s
ign
im
a
ge
s
f
or
B
S
L
vowe
ls
a
nd
F
igu
r
e
2
s
how
s
B
a
ngla
numer
ica
l
s
igns
.
L
e
a
r
ning
the
s
ign
langua
ge
c
a
n
be
quit
e
a
da
u
nti
ng
tas
k
to
a
n
a
ble
pe
r
s
on
be
c
a
us
e
s
pe
a
king
a
s
a
mea
ns
of
c
omm
unica
ti
on
be
c
omes
their
na
tur
e
.
As
a
r
e
s
ult
,
mos
t
pe
ople
of
thi
s
c
ountr
y
do
no
t
k
now
s
ign
langua
ge
of
B
a
ngla
langua
ge
.
As
s
uc
h
with
he
a
r
ing
a
nd
s
pe
e
c
h
im
pa
ir
e
d
pe
ople
ha
r
dly
be
c
ome
s
a
f
ull
f
unc
ti
oning
membe
r
of
the
s
oc
iety.
T
he
r
e
f
or
e
,
a
n
a
utom
a
ted
pr
oc
e
s
s
that
c
a
n
t
r
a
ns
late
the
s
ign
langua
ge
int
o
na
tur
a
l
s
poke
n
langua
ge
is
ne
c
e
s
s
a
r
y.
As
e
ve
r
y
l
a
ngua
ge
c
ons
is
ts
of
a
lphabe
ts
,
r
e
c
ognit
ion
a
nd
tr
a
ns
lation
of
a
lphabe
ts
is
the
f
ir
s
t
s
tep
towa
r
ds
a
c
ompl
e
te
r
e
a
l
-
ti
me
tr
a
ns
lation
s
ys
tem.
C
r
e
a
ti
ng
a
c
omput
e
r
s
ys
tem
that
c
a
n
r
e
c
ognize
B
S
L
a
lphabe
t
s
ign
f
r
om
a
n
im
a
ge
a
nd
tr
a
ns
late
it
to
B
a
ngla
let
ter
f
or
B
a
ngla
s
pe
a
king
pe
ople
ba
s
e
d
on
C
NN
[
1
]
,
is
the
obtec
ti
ve
of
thi
s
pa
pe
r
.
W
e
pr
e
s
e
nt
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
utomated
B
angla
s
ign
language
tr
ans
lat
ion
s
y
s
te
m
for
…
(
T
az
k
ia
M
im
A
ngona
)
1293
a
s
im
ple
a
nd
c
ompac
t
s
c
he
me
f
or
r
e
c
ognit
ion
o
f
B
S
L
in
thi
s
pa
pe
r
.
M
obil
e
Ne
t
V1
ha
s
be
e
n
us
e
d
ba
s
e
d
on
tr
a
ns
f
e
r
lea
r
ning
.
M
obil
e
Ne
t
is
a
c
las
s
of
li
ght
we
ight
de
e
p
c
onvolut
ional
ne
ur
a
l
ne
two
r
k
whic
h
a
r
e
va
s
tl
y
s
maller
in
s
ize
a
nd
f
a
s
ter
in
pe
r
f
or
manc
e
.
I
t
is
f
i
ne
-
tuned
to
a
da
pt
to
ou
r
pr
obl
e
m.
I
s
ha
r
a
-
L
ipi
da
tas
e
t
[
2
]
wa
s
us
e
d
to
tr
a
in
a
nd
e
va
luate
our
model
.
T
his
a
ppr
oa
c
h
a
ls
o
r
e
qui
r
e
s
les
s
s
tor
a
ge
than
other
e
xis
ti
ng
s
olut
ions
.
W
e
ha
ve
buil
t
a
model
f
or
de
s
ktop
c
omp
uter
,
but
due
to
low
memo
r
y
us
a
ge
,
it
c
a
n
a
ls
o
be
e
xtende
d
to
mobi
le
de
vice
s
a
s
we
ll
.
C
ontr
ibut
ion
of
thi
s
wor
k
c
a
n
be
wr
it
ten
a
s
f
oll
ows
:
-
B
S
L
a
lphabe
ts
r
e
c
ognit
ion
model
ha
s
be
e
n
pr
opos
e
d
a
nd
im
pleme
nted.
-
A
r
e
li
a
ble
de
s
ktop
a
ppli
c
a
ti
on
ha
s
be
e
n
buil
t
whic
h
c
a
n
tr
a
ns
late
36
B
S
L
a
lphabe
ts
.
F
igur
e
1.
B
S
L
vowe
ls
ha
nd
s
ign
[
3
]
F
igur
e
2.
B
S
L
numer
ica
ls
[
4
]
2.
RE
L
AT
E
D
WORKS
Va
r
ious
mac
hine
lea
r
ning
tec
hniques
with
im
a
ge
pr
oc
e
s
s
ing
ha
ve
a
lr
e
a
dy
be
e
n
u
s
e
d
by
va
r
ious
r
e
s
e
a
r
c
he
r
s
f
or
B
S
L
r
e
c
ognit
ion.
S
ome
of
thes
e
m
e
thods
a
r
e
dis
c
us
s
e
d
br
ief
ly
he
r
e
.
I
n
pa
pe
r
[
5
]
,
the
a
uthor
s
pr
opos
e
d
a
method
us
ing
AN
N
f
or
tr
a
ns
lating
th
e
B
S
L
a
lphabe
t
to
c
or
r
e
s
ponding
B
a
ngla
lette
r
.
F
or
im
a
ge
pr
oc
e
s
s
ing
ope
r
a
ti
ons
,
M
a
tl
a
b
tool
s
we
r
e
a
ppli
e
d.
A
tot
a
l
of
828
im
a
ge
s
we
r
e
c
oll
e
c
ted
f
o
r
tr
a
i
ning
a
nd
tes
ti
ng.
T
he
s
ys
tem
c
a
n
r
e
c
ognize
36
B
S
L
a
lphabe
ts
.
T
he
a
c
c
u
r
a
c
y
o
f
the
s
ys
tem
is
80.
902%
[
5
]
.
I
n
p
a
p
e
r
[
6
]
,
t
h
e
a
u
t
h
o
r
s
d
e
v
e
l
o
p
e
d
a
m
o
d
e
l
f
o
r
B
S
L
r
e
c
o
g
n
i
t
i
o
n
b
y
i
m
a
g
e
p
r
o
c
e
s
s
i
n
g
.
I
n
i
t
i
a
l
l
y
s
k
i
n
c
o
l
o
r
i
s
d
e
t
e
c
t
e
d
u
s
i
n
g
Y
C
b
C
r
s
k
i
n
d
e
t
e
c
t
i
o
n
a
l
g
o
r
i
t
h
m
.
E
x
t
r
a
c
t
e
d
f
e
a
t
u
r
e
s
f
r
o
m
i
m
a
g
e
s
w
e
r
e
f
e
d
i
n
t
o
S
V
M
c
l
a
s
s
i
f
i
e
r
f
o
r
t
r
a
i
n
i
n
g
a
n
d
t
e
s
t
i
n
g
.
T
h
e
y
c
o
l
l
e
c
t
e
d
i
m
a
g
e
s
o
f
1
5
B
a
n
g
l
a
c
o
n
s
o
n
a
n
t
h
a
n
d
s
i
g
n
s
.
E
a
c
h
c
l
a
s
s
o
f
s
i
g
n
s
h
a
s
5
2
o
r
5
6
i
m
a
g
e
s
o
f
2
9
d
i
f
f
e
r
e
n
t
b
a
r
e
h
a
n
d
s
.
T
h
e
9
5
%
a
v
e
r
a
g
e
a
c
c
u
r
a
c
y
h
a
s
b
e
e
n
r
e
c
e
i
v
e
d
f
o
r
t
r
a
i
n
i
n
g
s
e
t
a
n
d
8
6
%
h
a
s
b
e
e
n
r
e
c
e
i
v
e
d
f
o
r
e
v
a
l
u
a
t
i
o
n
s
e
t
.
F
o
r
t
h
e
l
e
t
t
e
r
s
‘
ঙ
’
a
n
d
‘
চ
’
i
t
g
o
t
v
e
r
y
l
o
w
a
c
c
u
r
a
c
y
[
6
].
C
NN
a
ppr
oa
c
h
wa
s
take
n
by
R
ony
e
t
a
l.
f
or
s
ing
le
-
ha
nd
B
S
L
c
ha
r
a
c
t
e
r
R
e
c
ognit
ion.
T
his
s
ys
tem
of
B
S
L
in
ter
pr
e
ter
ha
s
be
e
n
de
s
igned
with
the
pu
r
pos
e
of
pr
ovidi
ng
r
e
a
l
ti
me
c
omm
unica
ti
on
.
T
his
s
ys
tem
is
tr
a
ined
by
us
ing
T
e
ns
or
f
low
I
nc
e
pti
on
-
v3
M
ode
l
a
nd
tes
ted
by
a
pplyi
ng
C
NN
.
I
t
us
e
s
a
c
a
mer
a
to
de
tec
t
s
ign
in
r
e
a
l
t
im
e
.
T
he
y
a
c
hieve
d
a
n
a
ve
r
a
ge
r
e
c
ognit
ion
a
c
c
ur
a
c
y
of
92.
85%
[
7
].
Anothe
r
method
ha
s
be
e
n
pr
opos
e
d
by
S
ha
nta
e
t
a
l.
to
de
tec
t
B
S
L
us
ing
S
I
F
T
a
nd
C
NN
.
T
his
method
e
xtr
a
c
ts
f
e
a
tur
e
de
s
c
r
ipt
or
s
of
R
OI
(
R
e
gion
of
I
nter
e
s
t)
us
ing
S
I
F
T
,
then
c
lus
ter
s
the
f
e
a
tur
e
s
us
ing
K
-
mea
n
s
c
lus
ter
ing
to
ge
t
c
lus
ter
e
d
de
s
c
r
ipt
or
s
,
us
e
s
B
a
g
of
int
e
r
e
s
t
(
B
OI
)
to
r
e
pr
e
s
e
nt
c
lus
ter
e
d
f
e
a
tur
e
s
a
s
his
togr
a
ms
a
nd
then
input
the
h
is
togr
a
ms
in
C
NN
.
T
his
method
wa
s
done
both
wi
th
a
nd
without
S
I
F
T
whe
r
e
us
ing
S
I
F
T
yields
higher
a
c
c
ur
a
c
y
o
n
e
ve
r
y
s
ign.
Us
ing
S
I
F
T
thi
s
method
c
a
n
de
tec
t
38
ha
nd
s
ign
f
or
51
letter
s
wi
th
88.
52
%
a
c
c
ur
a
c
y
on
a
ve
r
a
ge
[
8
].
B
S
L
R
e
c
ognit
ion
Us
ing
De
e
p
C
NN
ha
s
be
e
n
in
tr
oduc
e
d
by
Hos
s
e
n
e
t
a
l.
T
h
is
s
ys
tem
is
made
to
de
tec
t
s
tatic
ha
nd
s
ign
f
r
om
i
mage
s
take
n
unde
r
va
r
ious
li
ghti
ng
c
ondit
ions
,
dif
f
e
r
e
nt
ba
c
kgr
ound
a
nd
pos
it
ions
of
the
ha
nd.
T
his
s
ys
tem
c
a
n
de
tec
t
37
B
S
L
a
lphabe
ts
with
va
li
da
ti
on
a
c
c
ur
a
c
y
of
84.
68%
[
9
].
I
s
lam
e
t
a
l.
pr
opos
e
d
a
mode
l
to
identif
y
B
S
L
d
igi
ts
us
ing
C
NN
.
T
his
model
us
e
s
mul
ti
-
laye
r
C
NN
with
2
s
ub
laye
r
s
a
nd
AD
AM
opti
mi
z
e
r
.
A
f
latte
n
a
nd
de
ns
e
laye
r
wa
s
us
e
d
a
f
ter
max
-
pooli
ng
laye
r
a
nd
on
the
f
inal
output
laye
r
s
of
tm
a
x
a
c
ti
va
ti
o
n
ha
s
be
e
n
us
e
d.
I
s
ha
r
a
-
L
ipi
da
tas
e
t
[
8]
w
a
s
us
e
d
a
nd
the
model
ha
s
s
hown
94.
88
%
a
c
c
ur
a
c
y
f
o
r
digi
t
0
t
o
9
[
10
].
R
e
c
ognit
ion
of
B
S
L
a
lphabe
ts
by
C
NN
wa
s
pr
o
pos
e
d
by
I
s
lam
e
t
a
l
.
T
he
y
a
ppli
e
d
their
model
on
I
s
ha
r
a
-
L
ipi
da
tas
e
t
[
2
]
whic
h
c
ontains
36
B
S
L
c
ha
r
a
c
ter
s
.
T
he
model
ha
d
mul
ti
laye
r
e
d
C
NN
a
nd
us
e
d
Ada
m
opti
mi
z
e
r
with
0.
001
lea
r
ning
late
.
On
thi
s
da
tas
e
t,
the
mod
e
l
a
c
hieve
d
92
%
a
c
c
ur
a
c
y
[
11
].
A
B
S
L
r
e
c
ognit
ion
s
ys
tem
of
f
e
r
e
d
by
Ha
s
a
n
e
t
a
l.
whic
h
a
ppli
e
d
AN
N
f
or
t
r
a
ini
ng
a
nd
c
las
s
if
ica
ti
on
a
nd
F
r
e
e
man
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
2
9
2
-
1
30
1
1294
C
ha
in
C
ode
wa
s
u
s
e
d
f
or
f
e
a
tur
e
e
xtr
a
c
ti
on.
YC
bC
r
wa
s
us
e
d
a
s
s
kin
de
tec
ti
on
a
lgor
it
hm.
T
he
s
ys
t
e
m
c
ould
r
e
c
ognize
20
a
lphabe
ts
a
nd
96.
5%
a
ve
r
a
ge
a
c
c
ur
a
c
y
r
a
te
wa
s
a
c
hieve
d.
B
ut
thi
s
s
ys
tem
c
a
n
de
tec
t
only
one
ha
nde
d
s
igns
[
12
].
Anothe
r
s
ys
tem
de
ve
loped
by
Ya
s
ir
e
t
a
l.
to
r
e
c
ognis
e
two
ha
nd
s
ign.
AN
N
wa
s
us
e
d
to
t
r
a
in
the
s
ys
tem
a
nd
Y
C
bC
r
a
lgor
it
hm
f
or
s
kin
de
tec
ti
on.
T
wo
dif
f
e
r
e
nt
tec
hniques
na
m
e
ly
P
C
A
a
nd
L
DA
we
r
e
us
e
d
a
nd
c
ompar
e
d
a
s
f
e
a
tur
e
e
xtr
a
c
to
r
whe
r
e
P
C
A
s
tands
f
or
pr
incipa
l
c
omponent
a
na
lys
is
a
nd
L
D
A
s
tands
f
or
L
inea
r
Dis
c
r
im
inant
Ana
lys
is
.
I
mage
s
of
15
B
a
ngla
a
lphabe
ts
we
r
e
c
oll
e
c
ted
f
r
om
22
pe
ople.
AN
N
c
las
s
if
ier
with
L
DA
pr
oduc
e
d
a
n
e
xc
e
ll
e
n
t
r
e
s
ult
.
T
he
r
e
s
ult
is
much
be
tt
e
r
than
AN
N
with
P
C
A
or
r
a
w
im
a
ge
[
13
]
.
A
ne
w
a
ppr
oa
c
h
wa
s
pr
opos
e
d
by
Ha
s
a
n
e
t
a
l.
H
ybr
id
c
olo
r
model
(
R
GB
-
YC
bC
r
-
HSV)
wa
s
us
e
d
f
or
s
kin
d
e
tec
ti
on.
A
c
ombi
na
ti
on
of
P
C
A
a
nd
F
L
DA
(
f
is
he
r
’
s
li
ne
a
r
dis
c
r
im
inant
a
na
l
ys
is
)
wa
s
int
r
oduc
e
d
to
e
xtr
a
c
t
f
e
a
tur
e
s
.
T
he
y
a
dopted
a
ne
w
c
las
s
if
ier
L
S
VM
.
T
he
s
ys
tem
c
a
n
identif
y
10
B
a
ngla
numb
e
r
s
[
14
]
.
Ya
s
ir
e
t
a
l.
int
r
oduc
e
d
a
S
I
F
T
(
S
ys
tem
I
nva
r
iant
F
e
a
tur
e
T
r
a
ns
f
or
m)
ba
s
e
d
a
ppr
oa
c
h
f
or
two
ha
n
de
d
s
ign
a
lphabe
ts
r
e
c
ognit
ion.
Af
te
r
no
r
maliza
ti
on
of
in
put
im
a
ge
s
ge
omet
r
ica
l
ba
s
e
d
S
I
F
T
method
is
us
e
d
f
or
e
xtr
a
c
ti
ng
f
e
a
tur
e
s
.
S
VM
wa
s
a
ppli
e
d
a
s
c
las
s
if
ier
.
C
ompar
ing
be
twe
e
n
c
las
s
if
ier
s
K
-
NN
(
e
xpe
r
im
e
ntal
r
e
s
ult
s
of
R
a
ha
man
e
t
a
l.
[
5
]
)
a
nd
S
VM
r
e
c
ognit
io
n
r
a
tes
,
k
-
NN
c
las
s
if
ies
be
tt
e
r
as
the
da
tas
e
t
wa
s
s
mall
[
15
].
Ka
r
mokar
e
t
a
l.
int
r
oduc
e
d
a
B
S
L
r
e
c
ognit
ion
s
ys
t
e
m
whic
h
r
e
c
ognize
s
47
a
lphabe
ts
of
B
a
ngla
S
ign
L
a
ngua
ge
by
NN
E
(
ne
ur
a
l
ne
twor
k
e
ns
e
mbl
e
)
.
A
method
c
ombi
ning
f
e
a
tur
e
e
xt
r
a
c
ti
on
a
nd
NN
E
wa
s
pr
opos
e
d.
A
tot
a
l
235
s
a
mpl
e
s
(
5
f
o
r
e
a
c
h
a
lphabe
t
)
we
r
e
us
e
d
f
or
tr
a
ini
ng
of
the
s
ys
tem.
NC
L
a
lgor
it
hm
wa
s
e
mpl
oye
d
f
or
f
a
s
ter
tr
a
ini
ng
.
T
he
e
xpe
r
i
menta
l
r
e
s
ult
s
hows
NC
L
with
f
e
a
tur
e
e
xtr
a
c
ti
on
gives
a
n
a
ppr
oxim
a
te
a
c
c
ur
a
c
y
of
93%
.
Va
r
ying
li
ghti
ng,
pr
e
s
e
nc
e
of
s
kin
c
olor
e
d
ob
jec
ts
,
qua
l
it
y
o
f
t
he
im
a
ge
c
a
us
e
s
it
to
f
a
c
e
pr
oblem
in
r
e
c
ognizing
the
a
lphab
e
ts
[
16
].
Az
iz
e
t
a
l.
a
ppr
oa
c
he
d
a
dif
f
e
r
e
nt
method
f
o
r
r
e
c
ognit
ion
of
B
S
L
.
T
he
pr
oc
e
s
s
of
r
e
c
ognit
ion
is
ba
s
e
d
on
dyna
mi
c
c
a
li
br
a
ti
on
of
s
kin
a
s
w
e
ll
a
s
ge
ometr
ic
ha
s
hing.
I
nput
im
a
ge
is
c
onve
r
ted
int
o
binar
y
im
a
ge
.
T
his
pr
oc
e
s
s
of
c
onve
r
s
ion
is
done
b
y
dyna
mi
c
c
olo
r
-
ba
s
e
d
s
kin
thr
e
s
hold
a
nd
me
a
n
s
hif
t
s
e
gmenta
ti
on
a
lgor
it
hm.
Ha
s
h
table
wa
s
tr
a
ine
d
with
1147
s
a
mpl
e
s
of
37
a
lphabe
ts
.
T
he
h
a
nd
s
ign
is
identif
ied
by
ha
s
hing
[
17
].
F
o
r
wor
k
in
g
with
i
mage
s
of
dif
f
e
r
e
nt
ba
c
kgr
ounds
,
a
model
wa
s
pr
o
pos
e
d
by
Ha
que
e
t
a
l
.
whic
h
identif
ies
two
ha
nde
d
s
igns
us
ing
P
C
A
a
lgo
r
it
hm
a
nd
AN
N.
T
he
mec
ha
nis
m
wa
s
to
pr
e
-
pr
oc
e
s
s
im
a
ge
s
a
nd
us
e
P
C
A
f
or
f
e
a
tur
e
e
xtr
a
c
ti
on.
K
-
Ne
a
r
e
s
t
Ne
ighbor
wa
s
c
hos
e
n
a
s
c
las
s
if
ier
.
I
t
c
a
n
r
e
c
ognize
26
B
a
ngla
a
lphabe
ts
with
a
n
a
ve
r
a
ge
a
c
c
ur
a
c
y
r
a
te
of
77.
846
%
.
T
he
mec
ha
nis
m
c
a
n
give
s
a
ti
s
f
a
c
tor
y
r
e
s
ult
s
f
or
input
im
a
ge
s
with
di
f
f
e
r
e
nt
ba
c
kgr
ounds
[
18
]
.
I
n
pa
pe
r
[
19
]
,
the
a
utho
r
s
de
ve
loped
a
B
S
L
r
e
c
o
gnit
ion
s
ys
tem.
T
he
s
ys
t
e
m
is
ba
s
ially
a
r
e
a
l
ti
me
c
omput
e
r
vis
ion
ba
s
e
d
s
ys
tem
whe
r
e
ha
nd
a
r
e
a
is
de
tec
ted
f
r
om
the
input
im
a
ge
us
ing
the
c
las
s
if
ier
na
mely
Ha
a
r
-
li
ke
f
e
a
t
ur
e
-
ba
s
e
d
c
a
s
c
a
de
d
c
las
s
if
ier
[
20
]
.
U
s
ing
K
-
Ne
a
r
e
s
t
Ne
ighbor
c
las
s
if
ier
,
ha
nd
s
ign
is
e
xtr
a
c
ted
by
the
pr
opos
e
d
s
ys
tem
f
r
om
Hue
a
nd
S
a
tur
a
ti
on
va
lues
of
HSV
c
olo
r
model
whe
r
e
HSV
s
tands
f
or
Hue
,
S
a
tur
a
ti
on,
a
nd
Va
lue.
HSV
c
olor
model
is
f
r
e
que
ntl
y
us
e
d
ins
tea
d
of
the
R
GB
c
olor
model
in
a
ppli
c
a
ti
on
pr
ogr
a
ms
s
uc
h
a
s
gr
a
phics
a
nd
pa
int
pr
ogr
a
ms
.
T
he
s
ys
tem
c
a
n
r
e
c
ogn
ize
30
c
ons
ona
nt
s
with
a
n
a
c
c
ur
a
c
y
of
94
.
75%
a
nd
6
vowe
ls
with
a
n
a
c
c
ur
a
c
y
of
98.
17%
.
T
he
s
ys
tem
c
a
n
give
output
in
r
e
a
l
ti
me.
T
h
e
s
ys
tem
ha
s
s
ome
pr
oblems
r
e
c
ognizing
the
s
im
il
a
r
ha
nd
s
igns
.
“
র
”
with
‘
ল
’
a
nd
‘
ফ
’
with
‘
ঝ
’
c
a
n
be
the
e
xa
mpl
e
s
of
s
im
il
a
r
ha
nd
s
igns
[
1
9
].
Hoque
e
t
a
l
pr
opos
e
d
a
de
tec
ti
on
s
ys
tem
na
mely
R
e
a
l
T
im
e
B
a
ngla
S
ign
L
a
ngua
ge
De
tec
ti
on.
F
a
s
ter
R
-
C
NN
ha
s
be
e
n
us
e
d
in
thi
s
pr
opos
e
d
s
ys
tem
f
or
de
tec
ti
ng
ha
nd
s
ign
a
nd
dif
f
e
r
e
nt
ba
c
kgr
ound
of
R
OI
whe
r
e
R
-
C
NN
s
tands
f
or
R
e
gion
-
C
onvolut
ional
N
e
ur
a
l
Ne
twor
k.
T
h
is
model
is
de
s
igned
to
de
tec
t
B
S
L
f
r
om
im
a
ge
s
with
di
f
f
e
r
e
nt
ba
c
kgr
ou
nd
a
nd
li
ghti
ng
in
r
e
a
l
ti
me
.
I
t
c
a
n
de
tec
t
10
letter
s
[
21
]
.
3.
B
AC
KROUND
An
ove
r
view
of
f
ine
-
tunni
ng
a
nd
M
obil
e
Ne
t
V1
is
given
in
thi
s
s
e
c
ti
on.
3.
1.
F
in
e
-
t
u
n
n
in
g
F
ine
-
tunni
ng
is
li
nke
d
with
tr
a
ns
f
e
r
l
e
a
r
n
i
n
g
.
W
h
e
n
w
e
u
s
e
k
n
o
w
l
e
d
g
e
t
h
a
t
w
a
s
g
a
i
n
e
d
f
r
o
m
o
n
e
t
a
s
k
a
n
d
a
p
p
l
y
i
t
t
o
a
n
e
w
t
a
s
k
,
t
h
a
t
’
s
t
r
a
n
s
f
e
r
l
e
a
r
n
i
n
g
.
F
i
n
e
-
t
u
n
n
i
g
i
s
a
p
r
o
c
e
s
s
w
h
e
r
e
i
t
t
a
k
e
s
a
n
a
l
r
e
a
d
y
t
r
a
i
n
e
d
m
o
d
e
l
a
n
d
t
h
e
n
t
u
n
e
s
o
r
t
w
e
a
k
s
t
h
a
t
m
o
d
e
l
t
o
p
e
r
f
o
r
m
a
s
e
c
o
n
d
t
a
s
k
.
I
t
a
l
l
o
w
s
t
o
t
a
k
e
t
h
e
a
d
v
a
n
t
a
g
e
o
f
t
h
e
g
a
i
n
e
d
k
n
o
w
l
e
d
g
e
w
i
t
h
o
u
t
h
a
v
i
n
g
t
o
d
e
v
e
l
o
p
i
t
f
r
o
m
s
c
r
a
t
c
h
a
s
b
u
i
l
d
i
n
g
a
m
o
d
e
l
f
r
o
m
s
c
r
a
t
c
h
i
s
q
u
i
t
e
d
i
f
f
i
c
u
l
t
.
B
y
r
e
p
l
a
c
i
n
g
t
h
e
t
o
p
l
a
y
e
r
w
i
t
h
a
c
l
a
s
s
i
f
e
r
,
a
p
r
e
-
t
r
a
i
n
e
d
m
o
d
e
l
c
a
n
b
e
e
a
s
i
l
y
f
i
n
e
-
t
u
n
e
d
.
T
h
e
p
r
e
-
t
r
a
i
n
e
d
w
e
i
g
h
t
s
c
a
n
b
e
u
s
e
d
t
o
s
p
e
e
d
u
p
c
o
n
v
e
r
g
e
n
c
e
r
a
t
h
e
r
t
h
a
n
u
s
i
n
g
r
a
n
d
o
m
w
e
i
g
h
t
s
[1
].
3.
2.
M
ob
il
e
Ne
t
V1
ar
c
h
it
e
c
t
u
r
e
M
obil
e
Ne
ts
a
r
e
a
c
las
s
of
low
late
nc
y,
low
powe
r
model
that
c
a
n
be
us
e
d
in
c
las
s
if
ica
ti
on,
de
tec
ti
on
a
nd
other
tas
ks
C
NN
a
r
e
good
f
or
.
As
they
a
r
e
s
mall
in
s
ize
,
thes
e
a
r
e
c
ons
ider
e
d
to
be
gr
e
a
t
d
e
e
p
lea
r
ning
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
utomated
B
angla
s
ign
language
tr
ans
lat
ion
s
y
s
te
m
for
…
(
T
az
k
ia
M
im
A
ngona
)
1295
models
to
be
us
e
d
in
c
omput
e
r
a
nd
mob
il
e
de
vi
c
e
s
.
I
t
is
buil
t
on
f
il
ter
s
tha
t
c
a
n
be
s
e
pa
r
a
ted
i
n
de
pth.
T
his
gives
the
models
a
n
oppor
tuni
ty
to
de
c
r
e
a
s
e
the
number
of
pa
r
a
mete
r
s
ne
e
de
d
f
or
c
onvolut
ional
ope
r
a
ti
ons
,
thus
r
e
duc
ing
the
model’
s
s
ize
.
T
he
y
a
r
e
pr
e
-
tr
a
ined
on
I
mage
Ne
t
[
22]
da
tas
e
t
[
23
].
T
he
a
r
c
hit
e
c
tur
e
f
oll
owe
d
a
pa
t
te
r
n
whic
h
wa
s
e
a
s
y
to
r
e
pli
c
a
te
.
-
T
he
f
il
te
r
s
ize
of
e
a
c
h
c
onvolut
ional
laye
r
that
is
n’
t
P
oint
wis
e
c
onvolut
ion
-
3x3
-
S
tr
ide
2
is
us
e
d
f
or
downs
a
mpl
ing
in
int
e
r
media
te
c
onvolut
ion.
-
T
he
output
is
f
latten
e
d
by
a
Globa
l
P
ooli
ng
laye
r
[
23
].
A
s
ubs
ti
tut
e
of
the
s
tanda
r
d
c
onvolut
ional
f
il
ter
s
in
M
obil
e
Ne
ti
s
de
pth
-
wis
e
s
e
pa
r
a
ble
f
il
ter
whic
h
c
ons
is
ts
of
two
laye
r
s
.
T
hos
e
a
r
e
d
e
p
t
h
-
w
i
s
e
a
n
d
p
o
i
n
t
-
w
i
s
e
c
o
n
v
o
l
u
t
i
o
n
.
F
i
g
u
r
e
3
i
l
l
u
s
t
r
a
t
e
s
t
h
e
c
o
m
p
a
r
i
s
o
n
b
e
t
w
e
e
n
s
t
a
n
d
a
r
d
a
n
d
d
e
p
t
h
-
w
i
s
e
(
d
w
)
s
e
p
a
r
a
b
l
e
f
i
l
t
e
r
.
M
o
b
i
l
e
N
e
t
h
a
s
1
3
c
o
n
v
o
l
u
t
i
o
n
b
l
o
c
k
s
.
F
i
g
u
r
e
4
d
e
s
c
r
i
b
e
s
t
h
e
M
o
b
i
l
e
N
e
t
a
r
c
h
i
t
e
c
t
u
r
e
.
A
b
a
t
c
h
n
o
r
m
a
l
i
z
a
t
i
o
n
a
n
d
R
e
l
u
l
a
y
e
r
i
s
a
f
t
e
r
e
a
c
h
c
o
n
v
o
l
u
t
i
o
n
l
a
y
e
r
.
L
a
s
t
l
y
t
h
e
r
e
i
s
a
g
l
o
b
a
l
a
v
e
r
a
g
e
p
o
o
l
i
n
g
l
a
y
e
r
a
n
d
a
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
w
h
i
c
h
f
e
e
d
s
i
n
t
o
S
o
f
t
m
a
x
f
u
n
c
t
i
o
n
f
o
r
c
l
a
s
s
i
f
i
c
a
t
i
o
n
[
23
].
F
igur
e
3.
S
tanda
r
d
a
nd
De
pth
-
wis
e
(
dw)
s
e
pa
r
a
ble
c
onvolut
ion
F
igur
e
4.
M
obil
e
Ne
t
a
r
c
hit
e
c
tur
e
[
23
]
4.
P
ROP
OS
E
D
S
Y
S
T
E
M
DE
S
CR
I
P
T
I
ON
T
he
r
e
s
e
a
r
c
h
wor
k
is
de
s
c
r
ibed
in
the
f
oll
owing
s
ub
-
s
e
c
ti
on
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
2
9
2
-
1
30
1
1296
4.
1.
Dat
as
e
t
p
r
e
p
ar
at
ion
W
e
us
e
d
B
S
L
’
s
f
ir
s
t
c
ompl
e
te
is
olate
d
c
ha
r
a
c
t
e
r
s
da
tas
e
t,
I
s
ha
r
a
-
L
ipi
da
tas
e
t
[
2
]
.
T
he
da
tas
e
t
c
ontains
30
c
ons
ona
nts
a
nd
6
vowe
ls
of
B
S
L
c
ha
r
a
c
ter
s
.
T
a
ble
1
s
hows
the
pr
o
pe
r
ti
e
s
of
the
da
tas
e
t.
T
he
da
tas
e
t
holds
36x50=
1800
im
a
ge
s
in
tot
a
l
.
All
im
a
ge
s
a
r
e
c
r
oppe
d
a
nd
r
e
s
ize
d
by
128x12
8
pixels
.
T
he
im
a
ge
s
we
r
e
take
n
with
s
a
me
ba
c
kgr
ound
a
nd
in
s
a
me
li
ghti
ng
c
ondit
ions
.
T
he
im
a
ge
s
a
r
e
i
n
.
J
P
E
G
f
or
mat
a
nd
they
a
r
e
c
onve
r
ted
int
o
bina
r
y
im
a
ge
be
f
or
e
f
e
e
ding
i
t
int
o
the
model.
F
igur
e
5
s
ho
ws
s
ome
s
a
mpl
e
s
of
the
da
tas
e
t.
T
he
36
B
S
L
a
lphabe
ts
(
অ
,
আ
,
ই
,
উ
,
এ
,
ও
,
ক
,
খ
,
গ
,
ঘ
,
চ
,
ছ
,
জ
,
ঝ
,
ট
,
ঠ
,
ড
,
ঢ
,
ত
,
থ
,
দ
,
ধ
,
ণ
,
প
,
ফ
,
ব
,
ভ
,
ম
,
য়
,
র
,
ল
,
স
,
হ
,
ড়,
ং
,
ং
)
a
r
e
a
s
s
igned
a
s
the
c
las
s
labe
ls
of
(
0,
1,
2
……,
36)
in
the
da
tas
e
t.
10
%
da
ta
we
r
e
s
e
pa
r
a
ted
f
r
om
the
d
a
tas
e
t
f
or
f
inal
tes
ti
ng
of
our
s
ys
tem.
T
a
ble
1.
Da
tas
e
t
p
r
ope
r
ti
e
s
D
a
ta
N
umbe
r
of
c
la
s
s
S
e
ts
pe
r
c
la
s
s
T
ot
a
l
numbe
r
of
da
ta
V
ow
e
ls
6
50
300
C
ons
ona
nt
s
30
50
1500
F
igur
e
5.
B
S
L
C
ha
r
a
c
ter
s
Da
tae
t
S
a
mpl
e
s
[
10]
Us
ing
s
tr
a
ti
f
ied
K
f
old
c
r
os
s
va
li
da
ti
on,
the
e
nt
ir
e
da
tas
e
t
wa
s
divi
de
d
int
o
tr
a
in
a
nd
tes
t
s
e
t.
F
or
the
pur
pos
e
of
tr
a
ini
ng
,
the
va
lue
o
f
K
wa
s
s
e
t
to
5.
T
he
da
ta
wa
s
s
pli
t
int
o
5
s
ubs
e
ts
a
nd
the
model
wa
s
tr
a
ined
on
5
-
1=
4
s
ubs
e
ts
.
S
o
the
las
t
s
ubs
e
t
w
a
s
hold
f
or
tes
ti
ng
a
nd
it
wa
s
done
f
or
e
a
c
h
of
the
s
ubs
e
ts
.
I
n
th
is
c
a
s
e
,
a
ll
the
s
a
mpl
e
s
ha
d
a
c
h
a
nc
e
to
b
e
in
the
tes
t
s
e
t
in
e
a
c
h
it
e
r
a
ti
on,
thus
buil
ding
a
mor
e
r
e
li
a
ble
s
ys
tem.
4.
2.
M
e
t
h
od
ology
T
he
pr
opos
e
d
method
us
e
s
Google
’
s
pr
e
tr
a
ined
model
M
obil
e
Ne
t
ve
r
s
ion
1
whic
h
ha
s
a
good
a
c
c
ur
a
c
y
on
I
mage
Ne
t
da
tas
e
t
[
23]
.
T
he
tr
a
ns
f
e
r
lea
r
ning
[
2
4
]
a
ppr
oa
c
h
wa
s
take
n
to
f
ine
tune
th
e
model
to
our
s
pe
c
if
ic
c
las
s
if
ica
ti
on
pr
ob
lem.
W
e
t
r
a
ined
M
obil
e
Ne
t
f
or
50
e
poc
hs
wi
th
pr
ior
we
ight
s
of
I
mage
Ne
t
[
22
]
da
tas
e
t
on
our
tr
a
ini
ng
s
e
t.
All
pr
e
-
tr
a
ined
we
ight
s
c
a
n
be
us
e
d
to
ini
ti
a
li
z
e
e
nti
r
e
ne
twor
k
we
ight
s
in
o
r
de
r
to
r
e
duc
e
t
r
a
ini
ng
ti
me
[
25
]
.
T
he
im
a
ge
s
we
r
e
pr
e
pr
oc
e
s
s
e
d
f
or
the
model.
T
he
las
t
6
laye
r
s
we
r
e
not
tr
a
ined
f
r
om
the
model
whic
h
ha
d
the
f
u
ll
y
c
onne
c
ted
or
de
ns
e
laye
r
.
W
e
int
r
oduc
e
d
a
de
n
s
e
laye
r
with
36
c
las
s
l
a
be
l
s
whic
h
is
the
number
of
our
c
las
s
e
s
a
nd
‘
s
of
tm
a
x’
a
s
a
c
ti
va
ti
on
f
unc
ti
on
ins
tea
d
of
thes
e
6
laye
r
s
.
T
he
pr
opos
e
d
model
us
e
d
AD
AM
opti
mi
z
e
r
whe
r
e
lea
r
ning
r
a
te
wa
s
0.
001
a
nd
c
os
t
f
unc
t
ion
wa
s
c
a
tegor
ica
l
c
r
os
s
e
ntr
opy.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
utomated
B
angla
s
ign
language
tr
ans
lat
ion
s
y
s
te
m
for
…
(
T
az
k
ia
M
im
A
ngona
)
1297
T
he
c
onvolut
ional
laye
r
s
e
xtr
a
c
t
the
f
e
a
tur
e
s
o
f
the
im
a
ge
s
.
F
r
om
the
las
t
c
onvolut
ion
laye
r
“
c
onv_pw_13_r
e
lu”
50176
f
e
a
tur
e
s
we
r
e
e
xtr
a
c
ted
f
or
e
a
c
h
im
a
ge
.
Dur
ing
the
t
r
a
ini
ng
pr
oc
e
s
s
,
the
e
xtr
a
c
ted
f
e
a
tur
e
s
f
r
om
tr
a
ins
e
t
a
r
e
us
e
d
to
tr
a
in
the
de
ns
e
laye
r
.
De
ns
e
laye
r
a
c
ts
a
s
a
c
la
s
s
if
ier
whic
h
c
las
s
if
ies
the
im
a
ge
s
int
o
t
he
c
or
r
e
s
ponding
c
las
s
labe
ls
.
Af
t
e
r
tr
a
ini
ng
,
the
model
wa
s
e
va
luate
d
by
tes
t
s
e
t.
T
he
model
pr
e
dicte
d
the
c
las
s
labe
ls
of
the
im
a
ge
s
be
longi
n
g
to
tes
t
s
e
t.
F
igur
e
6
il
lus
tr
a
tes
the
block
diagr
a
m
of
our
pr
opos
e
d
a
ppr
oa
c
h.
T
he
pr
opos
e
d
model
is
tr
a
ined
us
ing
s
tr
a
ti
f
ied
k
f
old
c
r
os
s
va
li
da
ti
on
s
c
he
me.
T
he
da
tas
e
t
of
1800
im
a
ge
s
is
s
huf
f
led
a
nd
s
pli
t
uni
f
or
ml
y
in
to
5
f
olds
.
F
or
tes
ti
ng,
one
-
f
old
is
s
e
t
a
s
ide
a
nd
t
he
other
f
our
f
olde
r
s
a
r
e
us
e
d
f
or
t
r
a
ini
ng.
T
his
is
r
e
pe
a
ted
f
ive
ti
mes
in
or
de
r
to
us
e
e
a
c
h
f
old
a
s
t
e
s
t
da
ta.
T
he
mea
s
ur
e
s
f
or
e
va
luation
we
r
e
P
r
e
c
is
ion,
R
e
c
a
ll
,
F
1
-
s
c
or
e
a
nd
s
uppor
t.
P
r
e
c
is
ion
is
the
r
a
ti
o
of
pos
it
ive
obs
e
r
va
ti
ons
t
ha
t
we
r
e
p
r
e
dicte
d
a
c
c
ur
a
tely
to
to
tal
pr
e
dicte
d
o
bs
e
r
va
ti
ons
that
we
r
e
pos
it
ive.
F1
-
s
c
or
e
is
the
a
ve
r
a
ge
of
pr
e
c
is
ion
a
nd
r
e
c
a
ll
.
R
e
c
a
ll
is
the
r
a
ti
o
of
pos
it
ive
e
ve
nts
that
we
r
e
a
c
c
ur
a
tely
pr
e
dict
e
d
to
a
ll
the
obs
e
r
va
ti
ons
in
the
c
las
s
.
S
uppo
r
t
s
hows
t
he
nu
mber
of
s
a
mpl
e
s
in
the
c
las
s
e
s
.
F
igur
e
6.
Our
pr
opos
e
d
model
5.
RE
S
UL
T
AN
D
DI
S
CU
S
S
I
ON
T
he
tes
ti
ng
s
e
t
wa
s
20%
o
f
the
w
h
o
l
e
d
a
t
a
s
e
t
.
T
h
e
m
o
d
e
l
h
a
s
b
e
e
n
t
r
a
i
n
e
d
f
o
r
5
0
e
p
o
c
h
s
.
T
h
e
a
v
e
r
a
g
e
a
c
c
u
r
a
c
y
f
o
r
3
6
a
l
p
h
a
b
e
t
s
o
f
t
h
e
p
r
o
p
o
s
e
d
m
o
d
e
l
i
s
9
5
.
7
1
%
.
T
h
e
e
v
a
l
u
a
t
i
o
n
o
f
t
h
e
m
o
d
e
l
f
o
r
e
a
c
h
i
t
e
r
a
t
i
o
n
i
s
g
i
v
e
n
i
n
T
a
b
l
e
2
.
A
m
o
r
e
d
e
t
a
i
l
e
d
r
e
p
o
r
t
i
s
s
h
o
w
n
i
n
T
a
b
l
e
3
.
I
n
e
a
c
h
i
t
e
r
a
t
i
o
n
,
t
h
e
r
e
p
o
r
t
s
h
o
w
s
t
h
e
f
o
l
l
o
w
i
n
g
m
e
a
s
u
r
e
s
:
p
r
e
c
i
s
i
o
n
,
r
e
c
a
l
l
a
n
d
F
1
-
s
c
o
r
e
.
A
v
e
r
a
g
e
p
r
e
c
i
s
i
o
n
a
n
d
r
e
c
a
l
l
t
h
e
m
o
d
e
l
r
e
t
u
r
n
e
d
w
a
s
9
6
%
.
T
h
e
c
l
a
s
s
e
s
w
i
t
h
h
i
g
h
e
s
t
p
r
e
c
i
s
i
o
n
w
e
r
e
‘
ক
’
a
n
d
‘
দ
’
.
T
h
e
c
l
a
s
s
e
s
w
i
t
h
t
h
e
l
o
w
e
s
t
p
r
e
c
i
s
i
o
n
w
e
r
e
‘
স
’
a
n
d
‘
ড
’
.
A
s
s
o
m
e
i
m
a
g
e
s
w
e
r
e
b
l
u
r
a
n
d
c
o
m
p
l
e
x
i
n
s
h
a
p
e
t
h
e
m
o
d
e
l
c
o
u
l
d
n
’
t
r
e
c
o
g
n
i
z
e
t
h
e
m
c
o
r
r
e
c
t
l
y
.
T
h
e
m
o
d
e
l
g
o
t
c
o
n
f
u
s
e
d
b
e
t
w
e
e
n
s
o
m
e
h
a
n
d
s
i
g
ns
s
u
c
h
a
s
‘
স
’
a
n
d
‘
গ
’
a
s
t
h
e
y
a
r
e
s
i
m
i
l
a
r
i
n
s
h
a
p
e
,
s
o
t
h
e
s
e
c
l
a
s
s
e
s
h
a
d
a
l
o
w
a
c
c
u
r
a
t
e
p
r
e
d
i
c
t
i
o
n
r
a
t
e
.
T
a
ble
4
pr
e
s
e
nts
a
c
ompar
is
on
a
mong
e
xis
ti
ng
models
a
nd
ou
r
pr
opos
e
d
model
.
Our
pr
opos
e
d
a
r
c
hit
e
c
tur
e
ga
ve
be
tt
e
r
pe
r
f
or
manc
e
than
mos
t
of
thes
e
e
xis
ti
n
g
models
.
Us
ing
c
onve
nti
ona
l
C
NN
,
I
s
lam
e
t
a
l.
us
e
d
I
s
ha
r
a
-
L
ipi
da
tas
e
t
[
1]
whic
h
is
the
s
a
me
da
tas
e
t
that
we
us
e
d.
T
he
y
buil
t
their
o
wn
C
NN
model
whic
h
ga
ve
a
n
a
ve
r
a
ge
a
c
c
ur
a
c
y
of
92.
00
%
.
Our
pr
opos
e
d
model
M
obil
e
Ne
t
ga
ve
a
be
tt
e
r
a
c
c
ur
a
c
y
r
a
te
than
that
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
2
9
2
-
1
30
1
1298
T
a
ble
2.
M
ode
l
e
va
luation
r
e
s
ult
I
te
r
a
ti
ons
(
F
or
5
-
f
ol
d
)
A
c
c
ur
a
c
y (
%
)
1
st
95.71
2
nd
96.39
3
rd
92.00
4
th
96.94
5t
h
97.54
A
ve
r
a
ge
95.71
T
a
ble
3.
Ave
r
a
ge
r
e
s
ult
f
or
e
a
c
h
i
ter
a
ti
on
I
te
r
a
ti
ons
(
F
or
5 f
ol
d
)
P
r
e
c
is
io
n
R
e
c
a
ll
F1
-
s
c
or
e
S
uppor
t
1
st
0.96
0.96
0.96
840
2
nd
0.97
0.96
0.96
831
3
rd
0.93
0.92
0.92
825
4
th
0.97
0.97
0.97
817
5
th
0.98
0.98
0.98
812
A
ve
r
a
ge
0.96
0.96
0.96
4125
T
a
ble
4.
C
ompar
is
on
a
mong
e
xis
ti
ng
a
nd
our
p
r
op
os
e
d
model
M
ode
ls
M
e
th
odol
ogy
D
a
ta
s
e
t
C
la
s
s
A
ve
r
a
ge
a
c
c
ur
a
c
y
R
a
hma
n e
t
a
l.
[
5
]
M
ul
i
-
L
a
ye
r
N
e
ur
a
l
N
e
twor
k
w
it
h i
ma
ge
P
r
oc
e
s
s
in
g
23x36=
828 s
a
mpl
e
s
36 a
lp
ha
be
ts
80.90%
U
ddi
n
e
t
a
l.
[
6
]
Y
C
bC
r
a
lg
or
it
hm a
nd S
V
M
C
la
s
s
if
ie
r
800 s
a
mpl
e
s
15 c
ons
ona
nt
s
86%
R
ony e
t
a
l
[
7
]
C
N
N
1900 s
a
mpl
e
s
38 a
lp
ha
be
ts
,
3
s
pe
c
ia
l
c
ha
r
a
c
te
r
92.85%
I
s
la
m e
t
a
l.
[
11
]
C
N
N
1800 s
a
mpl
e
s
(
I
s
ha
r
a
-
L
ip
i
da
ta
s
e
t
[
1
])
36 a
lp
ha
be
ts
92%
P
r
opos
e
d mode
l
M
obi
le
N
e
t
1800 s
a
mpl
e
s
(
I
s
ha
r
a
-
L
ip
i
da
ta
s
e
t
[
1
])
36 a
lp
ha
be
ts
95.71%
6.
DE
VL
OP
I
NG
T
HE
DE
S
KT
OP
AP
P
L
I
CA
T
I
O
N
T
he
pr
opos
e
d
a
r
c
hit
e
c
tur
e
M
obil
e
Ne
t
wa
s
int
e
gr
a
ted
in
a
c
omput
e
r
vis
ion
ba
s
e
d
a
ppli
c
a
ti
on.
A
us
e
r
int
e
r
f
a
c
e
wa
s
c
r
e
a
ted
with
the
he
lp
of
P
ython
T
kint
e
r
.
T
he
homes
c
r
e
e
n
a
nd
t
r
a
ns
it
ion
f
r
om
homes
c
r
e
e
n
of
the
a
ppli
c
a
ti
on
is
s
hown
in
F
igur
e
7.
T
he
a
ppli
c
a
ti
on
ta
ke
s
a
ha
nd
s
ign
im
a
ge
a
s
in
put
a
nd
the
im
a
ge
is
f
e
d
int
o
the
t
r
a
ined
model.
I
t
c
las
s
if
ies
the
im
a
ge
a
nd
r
e
tur
ns
the
c
or
r
e
s
ponding
B
a
ngla
letter
in
r
e
a
l
ti
me.
T
he
B
a
ngla
letter
is
dis
playe
d
in
the
s
c
r
e
e
n
in
text
f
or
mat
.
I
t
a
ls
o
s
hows
the
top
5
pr
o
ba
bil
s
it
c
c
las
s
of
the
input
im
a
ge
.
He
r
e
,
the
c
las
s
ha
ving
the
highes
t
pr
oba
bil
it
y
is
s
e
lec
ted
a
s
the
c
las
s
of
that
pa
r
ti
c
ular
ha
nd
s
ign.
F
igur
e
7.
Us
e
r
int
e
r
f
a
c
e
of
B
S
L
de
s
ktop
tr
a
ns
lation
s
ys
tem
f
or
a
lphabe
ts
7.
CONC
L
USI
ON
AN
D
F
UT
UR
E
WORK
W
it
hout
the
c
r
e
a
ti
on
of
s
ign
langua
ge
he
a
r
ing
a
nd
s
pe
e
c
h
im
pa
ir
e
d
pe
ople
would
f
or
e
ve
r
be
una
ble
to
c
omm
unica
te
with
other
pe
ople
a
nd
e
ve
n
e
a
c
h
other
.
W
hil
s
t
s
ign
langua
ge
c
a
n
e
na
ble
them
t
o
e
xpr
e
s
s
thems
e
lves
,
lac
k
of
knowle
dge
of
s
ign
langua
ge
a
mong
the
ge
ne
r
a
l
pe
op
le
c
r
e
a
tes
a
ga
p
of
c
omm
u
nica
ti
on.
T
he
pa
pe
r
pr
e
s
e
nts
a
quick
a
nd
c
ompac
t
mehod
f
or
tr
a
ns
lation
of
B
S
L
a
lphabe
ts
in
to
B
a
ngla
L
a
ngua
ge
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
utomated
B
angla
s
ign
language
tr
ans
lat
ion
s
y
s
te
m
for
…
(
T
az
k
ia
M
im
A
ngona
)
1299
T
he
model
c
ould
r
e
c
ognize
36
a
lphabe
ts
wit
h
a
n
a
ve
r
a
ge
a
c
c
ur
a
c
y
of
95.
71
%
.
I
t
a
ls
o
ta
ke
s
les
s
c
omput
a
ti
ona
l
ti
me,
r
e
s
ult
ing
in
les
s
powe
r
c
ons
umpt
ion
whic
h
is
ne
c
e
s
s
a
r
y
nowa
da
ys
.
I
t
wa
s
us
e
d
to
buil
d
a
n
a
utom
a
ted
tr
a
ns
lation
s
ys
tem
f
or
c
omput
e
r
.
T
his
r
e
s
e
a
r
c
h
is
the
f
ir
s
t
s
tep
towa
r
ds
making
a
s
y
s
tem
that
c
a
n
be
us
e
d
in
da
y
-
to
-
da
y
s
it
ua
ti
ons
to
c
omm
uni
c
a
ti
on
a
s
a
s
ys
tem
ne
e
ds
to
r
e
c
ognize
the
a
lphab
e
ts
f
ir
s
t
be
f
or
e
it
c
a
n
be
tr
a
ined
to
s
tar
t
r
e
c
ognizing
wor
ds
.
B
y
us
ing
im
a
ge
s
of
higher
r
e
s
olut
ion,
the
model
c
ould
be
e
nha
nc
e
d.
T
his
model
c
a
nnot
de
tec
t
B
S
L
c
ha
r
a
c
ter
s
in
dif
f
e
r
e
nt
ba
c
kgr
ounds
a
nd
li
ghti
ng
c
ondit
ions
.
T
his
a
ppr
o
a
c
h
c
a
n
be
im
pr
ove
d
us
ing
a
mor
e
ve
r
s
a
ti
le
da
tas
e
t.
I
n
f
utur
e
,
it
c
a
n
be
us
e
d
to
r
e
c
ognize
a
ll
the
B
S
L
c
ha
r
a
c
ter
s
(
51
a
lphabe
ts
)
.
T
his
a
ppr
oa
c
h
is
the
ba
s
ic
s
tep
in
buil
ding
a
mo
r
e
a
c
c
ur
a
te
B
S
L
r
e
c
ognit
ion
s
ys
tem
in
r
e
a
l
ti
me.
W
e
ha
ve
buil
t
a
de
s
ktop
tr
a
n
s
lation
s
ys
tem
a
nd
due
to
low
s
tor
a
ge
it
c
a
n
be
e
xtende
d
to
mobi
le
de
vice
s
.
RE
F
E
RE
NC
E
S
[1
]
I.
G
o
o
d
fel
l
o
w
,
Y
.
Ben
g
i
o
an
d
A
.
Co
u
rv
i
l
l
e,
“Co
n
v
o
l
u
t
i
o
n
al
N
e
t
w
o
rk
s
,
”
D
eep
Le
a
r
n
i
n
g
,
M
IT
P
r
es
s
,
2
0
1
6
.
[2
]
M.
San
zi
d
u
l
Is
l
am,
S.
Su
l
t
an
a
Sh
arm
i
n
M
o
u
s
u
m
i
,
N
.
A
.
J
es
s
a
n
,
A
.
Sh
ah
a
ri
ar
A
za
d
Rab
b
y
,
an
d
S.
A
k
h
t
er
H
o
s
s
ai
n
,
“Is
h
ara
-
L
i
p
i
:
T
h
e
Fi
r
s
t
Co
m
p
l
e
t
e
Mu
l
t
i
p
u
r
p
o
s
eO
p
en
A
cces
s
D
at
a
s
et
o
f
Is
o
l
a
t
ed
Ch
arac
t
ers
f
o
r
Ban
g
l
a
Si
g
n
L
an
g
u
a
g
e,
”
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
B
a
n
g
l
a
S
p
eec
h
a
n
d
L
a
n
g
u
a
g
e
P
r
o
ces
s
i
n
g
(ICB
S
LP
)
,
p
p
.
1
-
4
,
2
0
1
8
.
[3
]
T
.
H
o
q
u
e
a
n
d
F.
K
ab
i
r,
“A
u
t
o
mat
e
d
Ban
g
l
a
S
i
g
n
L
an
g
u
a
g
e
T
ra
n
s
l
at
i
o
n
Sy
s
t
em
:
Pro
s
p
ec
t
s
,
L
i
m
i
t
a
t
i
o
n
s
an
d
A
p
p
l
i
ca
t
i
o
n
s
,
”
Co
n
fere
n
ce
:
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
In
f
o
r
m
a
t
i
cs
,
E
l
ect
r
o
n
i
c
s
a
n
d
V
i
s
i
o
n
(IC
I
E
V
)
,
p
p
.
8
5
6
-
8
6
2
,
2
0
1
6
.
[4
]
A
.
M.
J
arman
,
S.
A
rs
h
ad
,
N
.
A
l
am,
a
n
d
M.
J
.
I
s
l
am,
“
A
n
A
u
t
o
ma
t
ed
Be
n
g
a
l
i
S
i
g
n
L
an
g
u
ag
e
Rec
o
g
n
i
t
i
o
n
S
y
s
t
em
Bas
ed
o
n
Fi
n
g
er
t
i
p
Fi
n
d
er
A
l
g
o
ri
t
h
m
ISSN
:
2
1
8
6
-
0
1
1
4
,
”
In
t
.
J.
E
l
ect
r
o
n
.
In
f
o
r
m
a
t
i
c
s
,
v
o
l
.
4
,
n
o
.
1
,
2
0
1
5
.
[5
]
M.
A
.
Rah
man
,
“Reco
g
n
i
t
i
o
n
o
f
St
a
t
i
c
H
an
d
G
es
t
u
re
s
o
f
A
l
p
h
a
b
et
i
n
Ban
g
l
a
Si
g
n
L
an
g
u
a
g
e,
”
IO
S
R
J.
Co
m
p
u
t
.
E
n
g
.
,
v
o
l
.
8
,
n
o
.
1
,
p
p
.
0
7
-
1
3
,
2
0
1
2
.
[6
]
J
.
U
d
d
i
n
,
F.
N
.
A
rk
o
,
N
.
T
ab
as
s
u
m,
T
.
R.
T
ri
s
h
a,
an
d
F.
A
h
med
,
“Ban
g
l
a
s
i
g
n
l
a
n
g
u
ag
e
i
n
t
er
p
ret
a
t
i
o
n
u
s
i
n
g
b
a
g
o
f
feat
u
re
s
an
d
S
u
p
p
o
r
t
V
ect
o
r
Mach
i
n
e,
”
3
rd
I
n
t
.
Co
n
f
.
E
l
e
ct
r
.
In
f
.
Co
m
m
u
n
.
Tech
n
o
l
.
E
ICT,
p
p
.
1
-
4
,
2
0
1
7
.
[7
]
A
.
J
.
Ro
n
y
,
K
.
H
.
Sa
i
k
a
t
,
M.
T
an
zeem,
a
n
d
F.
M.
Rah
at
H
as
a
n
Ro
b
i
,
“A
n
effect
i
v
e
a
p
p
r
o
ach
t
o
c
o
mmu
n
i
c
at
e
w
i
t
h
t
h
e
d
eaf
an
d
mu
t
e
p
e
o
p
l
e
b
y
reco
g
n
i
zi
n
g
c
h
aract
er
s
o
f
o
n
e
-
h
a
n
d
b
an
g
l
a
s
i
g
n
l
an
g
u
a
g
e
u
s
i
n
g
c
o
n
v
o
l
u
t
i
o
n
al
n
eu
ra
l
-
n
e
t
w
o
rk
,
”
4
th
In
t
.
Co
n
f
.
E
l
ec
t
r
.
E
n
g
.
In
f
.
Co
m
m
u
n
.
Tech
n
o
l
.
i
CE
E
i
CT
,
p
p
.
7
4
-
7
9
,
2
0
1
8
.
[8
]
Su
l
t
a
n
a
S
h
an
t
a,
Sh
i
ri
n
&T
a
i
fu
r
A
n
w
ar,
Sa
i
f&Ray
h
an
u
l
K
ab
i
r,
Md
.
(
2
0
1
8
),
“Ba
n
g
l
a
Si
g
n
L
a
n
g
u
ag
e
D
et
ec
t
i
o
n
U
s
i
n
g
SIF
T
an
d
CN
N
,
”
Co
n
f
er
e
n
ce:
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
Co
m
p
u
t
i
n
g
,
Co
m
m
u
n
i
c
a
t
i
o
n
a
n
d
Ne
t
wo
r
k
i
n
g
Tech
n
o
l
o
g
i
es
(ICCCNT)
,
p
p
.
1
-
6
,
2
0
1
8
.
[9
]
M.
A
.
H
o
s
s
en
,
A
.
G
o
v
i
n
d
ai
a
h
,
S.
Su
l
t
an
a,
an
d
A
.
Bh
u
i
y
a
n
,
“Ben
g
a
l
i
s
i
g
n
l
a
n
g
u
ag
e
rec
o
g
n
i
t
i
o
n
u
s
i
n
g
d
eep
co
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
,
”
Jo
i
n
t
7
th
I
n
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
In
f
o
r
m
a
t
i
cs
,
E
l
ec
t
r
o
n
i
cs
&
V
i
s
i
o
n
(ICI
E
V
)
a
n
d
2
0
1
8
2
n
d
I
n
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
Im
a
g
i
n
g
,
V
i
s
i
o
n
&
P
a
t
t
e
r
n
R
ec
o
g
n
i
t
i
o
n
(
i
cIV
P
R
),
p
p
.
3
6
9
-
3
7
3
,
2
0
1
8
.
[1
0
]
M.
S.
Is
l
am,
S.
S.
S.
Mo
u
s
u
m
i
,
A
.
K
.
M.
S.
A
za
d
R
ab
b
y
,
S.
A
.
H
o
s
s
ai
n
,
an
d
S.
A
b
u
j
ar,
“A
p
o
t
en
t
m
o
d
e
l
t
o
reco
g
n
i
ze
b
a
n
g
l
a
s
i
g
n
l
a
n
g
u
ag
e
d
i
g
i
t
s
u
s
i
n
g
co
n
v
o
l
u
t
i
o
n
al
n
e
u
ral
n
e
t
w
o
rk
,
”
P
r
o
ce
d
i
a
Co
m
p
u
t
.
S
ci
.
,
v
o
l
.
1
4
3
,
p
p
.
6
1
1
-
6
1
8
,
2
0
1
8
.
[1
1
]
M.
S.
Is
l
am,
S.
S.
S.
Mo
u
s
u
m
i
,
A
.
S.
A
.
Rab
b
y
,
an
d
S.
A
.
H
o
s
s
a
i
n
,
“A
Si
mp
l
e
an
d
Mi
g
h
t
y
A
rr
o
w
h
ead
D
e
t
ect
i
o
n
T
ech
n
i
q
u
e
o
f
Ban
g
l
a
Si
g
n
L
an
g
u
ag
e
Ch
arac
t
ers
w
i
t
h
C
N
N
,
”
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
R
ecen
t
Tr
en
d
s
i
n
Im
a
g
e
P
r
o
ces
s
i
n
g
a
n
d
P
a
t
t
e
r
n
R
ec
o
g
n
i
t
i
o
n
v
o
l
.
1
0
3
5
,
p
p
.
4
2
9
-
4
3
7
,
2
0
1
9
.
[1
2
]
M.
M.
H
as
an
,
M.
K
h
a
l
i
l
u
zz
aman
,
S.
A
.
H
i
me
l
,
an
d
R.
T
.
Ch
o
w
d
h
u
r
y
,
“H
an
d
s
i
g
n
l
a
n
g
u
a
g
e
reco
g
n
i
t
i
o
n
f
o
r
Ba
n
g
l
a
al
p
h
ab
e
t
b
as
e
d
o
n
Freeman
Ch
ai
n
Co
d
e
an
d
A
N
N
,
”
4
th
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
A
d
v
a
n
ce
s
i
n
E
l
ect
r
i
ca
l
E
n
g
i
n
eer
i
n
g
(ICA
E
E
)
,
p
p
.
7
4
9
-
7
5
3
,
2
0
1
7
.
[1
3
]
R.
Y
as
i
r
a
n
d
R.
A
.
K
h
an
,
“
T
w
o
-
h
a
n
d
e
d
h
a
n
d
g
es
t
u
re
re
co
g
n
i
t
i
o
n
fo
r
Ba
n
g
l
a
s
i
g
n
l
a
n
g
u
ag
e
u
s
i
n
g
L
D
A
a
n
d
A
N
N
,
”
Th
e
8
th
In
t
er
n
a
t
i
o
n
a
l
Co
n
f
e
r
en
ce
o
n
S
o
f
t
w
a
r
e,
Kn
o
wl
ed
g
e,
In
f
o
r
m
a
t
i
o
n
M
a
n
a
g
e
m
en
t
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
(S
KIM
A
)
,
p
p
.
7
-
1
1
,
2
0
1
4
.
[1
4
]
S.
M.
K
.
H
as
a
n
a
n
d
M.
A
h
ma
d
,
“A
n
e
w
ap
p
ro
ac
h
o
f
s
i
g
n
l
a
n
g
u
ag
e
reco
g
n
i
t
i
o
n
s
y
s
t
em
f
o
r
b
i
l
i
n
g
u
a
l
u
s
e
rs
,
”
In
t
e
r
n
a
t
i
o
n
a
l
Co
n
f
er
e
n
ce
o
n
E
l
ec
t
r
i
ca
l
&
E
l
ec
t
r
o
n
i
c
E
n
g
i
n
eer
i
n
g
(ICE
E
E
),
p
p
.
3
3
-
3
6
,
2
0
1
5
.
[1
5
]
F.
Y
as
i
r,
P.
W
.
C.
Pras
ad
,
A
.
A
l
s
ad
o
o
n
,
an
d
A
.
E
l
c
h
o
u
emi
,
“SIFT
b
as
e
d
ap
p
ro
ac
h
o
n
Ban
g
l
a
s
i
g
n
l
an
g
u
ag
e
reco
g
n
i
t
i
o
n
I
E
E
E
8
t
h
In
t
er
n
at
i
o
n
al
W
o
rk
s
h
o
p
o
n
Co
mp
u
t
a
t
i
o
n
a
l
In
t
el
l
i
g
en
ce
an
d
A
p
p
l
i
cat
i
o
n
s
(IW
CI
A
),
p
p
.
3
5
-
3
9
,
2
0
1
6
.
[1
6
]
B.
Ch
an
d
raK
arm
o
k
ar,
K
.
Md
.
Ro
k
i
b
u
l
A
l
am,
an
d
M.
K
i
b
r
i
a
Si
d
d
i
q
u
ee,
“Ban
g
l
a
d
es
h
i
S
i
g
n
L
an
g
u
ag
e
Rec
o
g
n
i
t
i
o
n
E
mp
l
o
y
i
n
g
N
e
u
ral
N
et
w
o
r
k
E
n
s
emb
l
e,
”
In
t
.
J.
Co
m
p
u
t
.
A
p
p
l
.
,
v
o
l
.
5
8
,
n
o
.
1
6
,
p
p
.
4
3
-
4
6
,
2
0
1
2
.
[1
7
]
K
.
E
.
A
zi
z,
A
.
W
ad
u
d
,
S.
Su
l
t
an
a,
M.
A
.
H
u
s
s
ai
n
,
an
d
A
.
Bh
u
i
y
a
n
,
“Ben
g
a
l
i
Si
g
n
L
an
g
u
a
g
e
Reco
g
n
i
t
i
o
n
u
s
i
n
g
d
y
n
ami
c
s
k
i
n
cal
i
b
rat
i
o
n
an
d
g
e
o
met
r
i
c
h
as
h
i
n
g
,
”
6
th
In
t
.
Co
n
f
.
In
f
o
r
m
a
t
i
c
s
,
E
l
ec
t
r
o
n
.
V
i
s
.
7
th
In
t
.
S
ym
p
.
Co
m
p
u
t
.
M
ed
.
H
e
a
l
.
Tech
n
o
l
.
ICIE
V
-
I
S
CM
H
T
,
p
p
.
1
-
5
,
2
0
1
7
.
[1
8
]
P.
H
aq
u
e,
B.
D
a
s
,
an
d
N
.
N
.
K
as
p
y
,
“
T
w
o
-
H
a
n
d
e
d
Ban
g
l
a
S
i
g
n
L
an
g
u
ag
e
Rec
o
g
n
i
t
i
o
n
U
s
i
n
g
Pri
n
ci
p
al
C
o
mp
o
n
en
t
A
n
a
l
y
s
i
s
(PC
A
)
an
d
K
N
N
A
l
g
o
ri
t
h
m,
”
2
nd
In
t
.
Co
n
f
.
E
l
e
ct
r
.
Co
m
p
u
t
.
Co
m
m
u
n
.
E
n
g
.
E
CCE
,
p
p
.
1
-
4
,
2
0
1
9
.
[1
9
]
M.
A
.
Rah
aman
,
M.
J
as
i
m,
M.
H
.
A
l
i
,
an
d
M.
H
as
an
u
z
zaman
,
“Real
-
t
i
me
co
m
p
u
t
er
v
i
s
i
o
n
-
b
as
e
d
Ben
g
al
i
s
i
g
n
l
an
g
u
a
g
e
reco
g
n
i
t
i
o
n
,
”
1
7
t
h
In
t
.
Co
n
f
.
Co
m
p
u
t
.
In
f
.
Tech
n
o
l
.
ICCIT
,
p
p
.
1
9
2
-
1
9
7
,
2
0
0
3
.
[2
0
]
J
as
i
m,
Ma
h
mo
o
d
&
Z
h
an
g
,
T
ao
&H
a
s
an
u
zzaman
,
M
d
.
,
“A
Real
-
T
i
me
C
o
mp
u
t
er
V
i
s
i
o
n
-
Ba
s
ed
St
a
t
i
c
a
n
d
D
y
n
a
mi
c
H
an
d
G
es
t
u
re
R
ec
o
g
n
i
t
i
o
n
Sy
s
t
em,
”
In
t
er
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
Im
a
g
e
a
n
d
G
r
a
p
h
i
c
s
,
v
o
l
.
4
,
2
0
1
4
.
[2
1
]
O
.
B.
H
o
q
u
e,
M.
I.
J
u
b
ai
r,
M.
S.
Is
l
am,
A
.
F
.
A
k
as
h
,
an
d
A
.
S.
Pau
l
s
o
n
,
“Real
T
i
me
Ban
g
l
ad
e
s
h
i
Si
g
n
L
an
g
u
ag
e
D
et
ec
t
i
o
n
u
s
i
n
g
Fa
s
t
er
R
-
CN
N
,
”
In
t
.
Co
n
f
.
In
n
o
v.
E
n
g
.
Tech
n
o
l
.
ICIE
T
,
2
0
1
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
1693
-
6930
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
,
Vol.
1
8
,
No
.
3
,
J
une
2020:
1
2
9
2
-
1
30
1
1300
[2
2
]
L
.
Fei
-
Fei
,
J
.
D
e
n
g
,
a
n
d
K
.
L
i
,
“Ima
g
eN
e
t
:
C
o
n
s
t
r
u
ct
i
n
g
a
l
a
rg
e
-
s
cal
e
i
mag
e
d
a
t
ab
a
s
e,
”
J.
V
i
s
.
,
v
o
l
.
9
,
n
o
.
8
,
p
p
.
1
0
3
7
-
1
0
3
7
,
2
0
1
0
.
[2
3
]
A
.
G
.
H
o
w
ard
et
a
l
.
,
“Mo
b
i
l
e
N
et
s
:
E
ff
i
ci
en
t
C
o
n
v
o
l
u
t
i
o
n
a
l
N
eu
ra
l
N
e
t
w
o
rk
s
fo
r
Mo
b
i
l
e
V
i
s
i
o
n
A
p
p
l
i
cat
i
o
n
s
,
”
2
0
1
7
,
[
O
n
l
i
n
e],
A
v
ai
l
ab
l
e:
h
t
t
p
s
:
/
/
arx
i
v
.
o
rg
/
ab
s
/
1
7
0
4
.
0
4
8
6
1
[2
4
]
D
.
Rat
h
i
,
“O
p
t
i
m
i
zat
i
o
n
o
f
T
ra
n
s
fer
L
earn
i
n
g
fo
r
S
i
g
n
L
an
g
u
a
g
e
Reco
g
n
i
t
i
o
n
T
ar
g
et
i
n
g
Mo
b
i
l
e
P
l
at
f
o
r
m,
”
In
t
e
r
n
a
t
i
o
n
a
l
Jo
u
r
n
a
l
o
n
R
ece
n
t
a
n
d
In
n
o
va
t
i
o
n
T
r
en
d
s
i
n
Co
m
p
u
t
i
n
g
a
n
d
Co
m
m
u
n
i
ca
t
i
o
n
,
p
p
.
1
9
8
-
2
0
3
,
2
0
1
8
.
[2
5
]
I.
K
emen
d
i
k
b
u
d
e
t
al
.
,
“W
E
M
O
T
E
–
W
o
rd
E
mb
e
d
d
i
n
g
b
as
ed
Mi
n
o
r
i
t
y
O
v
er
s
amp
l
i
n
g
T
ech
n
i
q
u
e
fo
r
Imb
a
l
an
ced
E
mo
t
i
o
n
an
d
Sen
t
i
me
n
t
Cl
a
s
s
i
fi
ca
t
i
o
n
,
”
P
l
o
S
O
n
e
,
v
o
l
.
8
,
n
o
.
3
,
p
p
.
1
-
1
6
,
2
0
1
8
.
GRAP
H
I
E
S
OF
AU
T
HO
RS
Ta
zki
a
M
i
m
A
ng
o
n
a
recei
v
ed
t
h
e
B.
Sc
i
n
In
fo
rmat
i
o
n
an
d
Co
mmu
n
i
ca
t
i
o
n
T
ec
h
n
o
l
o
g
y
fro
m
Ban
g
l
ad
e
s
h
U
n
i
v
er
s
i
t
y
o
f
Pro
fes
s
i
o
n
a
l
s
.
Sh
e
i
s
w
o
rk
i
n
g
o
n
s
e
v
eral
re
s
earch
p
r
o
j
e
ct
s
cu
rren
t
l
y
.
H
er
re
s
earch
i
n
t
eres
t
s
i
n
c
l
u
d
e
Co
m
p
u
t
er
V
i
s
i
o
n
,
Pat
t
ern
Rec
o
g
n
i
t
i
o
n
,
A
rt
i
fi
c
i
a
l
In
t
e
l
l
i
g
e
n
ce,
Mach
i
n
e
L
earn
i
n
g
an
d
d
a
t
a
mi
n
i
n
g
.
A
.
S
.
M
.
Si
a
m
uzza
m
a
n
Sh
a
o
n
recei
v
ed
t
h
e
B.
Scan
d
i
n
In
fo
rma
t
i
o
n
a
n
d
Co
mm
u
n
i
cat
i
o
n
T
ech
n
o
l
o
g
y
fro
m
Ban
g
l
a
d
es
h
U
n
i
v
er
s
i
t
y
o
f
Pro
fes
s
i
o
n
al
s
.
H
i
s
res
earc
h
i
n
t
eres
t
s
i
n
cl
u
d
e
Co
mp
u
t
er
V
i
s
i
o
n
,
Pat
t
er
n
Reco
g
n
i
t
i
o
n
an
d
A
rt
i
fi
c
i
al
In
t
el
l
i
g
e
n
ce.
Ka
zi
Ta
h
m
i
d
R
a
s
ha
d
N
i
l
o
y
rece
i
v
e
d
t
h
e
B.
Sc
i
n
In
fo
rma
t
i
o
n
an
d
Co
mmu
n
i
ca
t
i
o
n
T
ech
n
o
l
o
g
y
fro
m
Ban
g
l
a
d
es
h
U
n
i
v
er
s
i
t
y
o
f
Pro
fes
s
i
o
n
al
s
.
H
i
s
res
earc
h
i
n
t
eres
t
s
i
n
cl
u
d
e
Co
mp
u
t
er
V
i
s
i
o
n
an
d
Mach
i
n
e
L
earn
i
n
g
.
Ta
jbi
a
K
a
ri
m
i
s
s
erv
i
n
g
as
a
l
ect
u
rer
i
n
In
fo
rma
t
i
o
n
an
d
Co
mmu
n
i
ca
t
i
o
n
T
ech
n
o
l
o
g
y
(ICT
)
d
ep
ar
t
men
t
at
Ban
g
l
a
d
es
h
U
n
i
v
er
s
i
t
y
o
f
Pro
f
es
s
i
o
n
al
s
(BU
P).
Prev
i
o
u
s
l
y
s
h
e
w
o
r
k
ed
i
n
E
l
ec
t
ri
ca
l
an
d
E
l
ec
t
ro
n
i
c
E
n
g
i
n
eeri
n
g
d
e
p
a
rt
men
t
o
f
A
meri
ca
n
In
t
ern
a
t
i
o
n
a
l
U
n
i
v
er
s
i
t
y
-
Ba
n
g
l
ad
e
s
h
.
Be
s
i
d
es
b
ei
n
g
an
aca
d
emi
c
i
a
n
,
s
h
e
i
s
v
o
l
u
n
t
ari
l
y
w
o
rk
i
n
g
a
s
a
s
t
u
d
en
t
ad
v
i
s
o
r
o
f
t
h
e
ICT
d
e
p
art
me
n
t
s
t
u
d
en
t
s
e
n
ro
l
l
ed
i
n
2
0
1
9
a
t
BU
P.
T
i
l
l
n
o
w
s
h
e
h
as
p
u
b
l
i
s
h
e
d
t
w
o
j
o
u
rn
a
l
an
d
t
h
ree
co
n
feren
ce
p
ap
er
s
.
Sh
e
h
as
a
d
i
v
ers
i
fi
e
d
res
earc
h
i
n
t
eres
t
i
n
c
l
u
d
i
n
g
E
l
ec
t
ro
n
i
c
s
,
Si
g
n
al
p
r
o
ces
s
i
n
g
,
D
a
t
a
s
ci
en
ce
et
c.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
KA
T
e
lec
omm
un
C
omput
E
l
C
ontr
o
l
A
utomated
B
angla
s
ign
language
tr
ans
lat
ion
s
y
s
te
m
for
…
(
T
az
k
ia
M
im
A
ngona
)
1301
Za
ri
n
Ta
s
n
i
m
i
s
w
o
rk
i
n
g
as
a
L
ect
u
rer
i
n
t
h
e
D
ep
ar
t
men
t
o
f
In
fo
rmat
i
o
n
an
d
Co
mmu
n
i
ca
t
i
o
n
T
ech
n
o
l
o
g
y
(ICT
)
u
n
d
er
t
h
e
Fac
u
l
t
y
o
f
Sci
en
ce
an
d
T
ec
h
n
o
l
o
g
y
(FS
T
),
Ban
g
l
ad
e
s
h
U
n
i
v
er
s
i
t
y
o
f
Pro
fes
s
i
o
n
a
l
s
(BU
P).
Sh
e
ach
i
e
v
ed
B.
Sc.
(E
n
g
i
n
eeri
n
g
)
d
eg
re
e
i
n
E
l
ect
r
i
cal
,
E
l
ec
t
r
o
n
i
c
an
d
Co
mm
u
n
i
cat
i
o
n
(E
E
C
E
)
E
n
g
i
n
eer
i
n
g
fro
m
Mi
l
i
t
ar
y
In
s
t
i
t
u
t
e
o
f
Sci
e
n
ce
an
d
T
ech
n
o
l
o
g
y
(MIS
T
),
D
h
ak
a,
Ba
n
g
l
ad
e
s
h
i
n
2
0
1
4
a
n
d
M.
Sc.
(E
n
g
i
n
eeri
n
g
)
d
eg
ree
i
n
T
el
ec
o
mmu
n
i
ca
t
i
o
n
s
an
d
W
i
rel
e
s
s
S
y
s
t
ems
f
ro
m
t
h
e
U
n
i
v
ers
i
t
y
o
f
L
i
v
erp
o
o
l
,
U
K
i
n
2
0
1
9
.
Sh
e
h
as
o
n
e
n
a
t
i
o
n
a
l
j
o
u
r
n
al
an
d
o
n
e
i
n
t
ern
at
i
o
n
a
l
co
n
fere
n
ce
p
u
b
l
i
ca
t
i
o
n
s
.
S
.
M
.
Sa
l
i
m
R
eza
i
s
an
A
s
s
i
s
t
a
n
t
Pro
fe
s
s
o
r
o
f
t
h
e
D
ep
ar
t
men
t
o
f
ICT
,
Ban
g
l
ad
e
s
h
U
n
i
v
er
s
i
t
y
o
f
Pro
fes
s
i
o
n
a
l
s
(BU
P),
D
h
a
k
a,
Ban
g
l
a
d
es
h
.
H
e
i
s
a
Pro
fes
s
i
o
n
al
Memb
er
o
f
IE
E
E
an
d
man
y
o
t
h
er
p
r
o
fes
s
i
o
n
a
l
o
rg
a
n
i
z
at
i
o
n
s
.
H
e
i
s
a
k
een
p
l
an
n
er
&
i
mp
l
eme
n
t
e
r
w
i
t
h
p
r
o
fi
c
i
en
c
y
i
n
man
a
g
i
n
g
acad
em
i
c
an
d
ad
m
i
n
i
s
t
rat
i
v
e
act
i
v
i
t
i
es
en
t
ai
l
i
n
g
o
p
erat
i
o
n
a
l
p
o
l
i
c
i
es
/
n
o
rms
,
facu
l
t
y
ap
p
rai
s
al
/
co
m
p
u
t
er
s
ci
e
n
ce
t
rai
n
i
n
g
.
H
e
h
as
mo
re
t
h
an
5
0
res
earc
h
p
ap
er
s
i
n
ISI
/
Sc
o
p
u
s
In
d
e
x
ed
j
o
u
rn
a
l
an
d
ma
n
y
rep
u
t
ed
In
t
er
n
at
i
o
n
al
co
n
feren
c
es
.
Ta
s
m
i
m
a
N
o
us
h
i
ba
M
a
hbub
,
cu
rre
n
t
l
y
s
t
u
d
y
i
n
g
i
n
BSc
i
n
In
fo
rm
at
i
o
n
an
d
Co
mmu
n
i
ca
t
i
o
n
T
ec
h
n
o
l
o
g
y
at
Ban
g
l
a
d
es
h
U
n
i
v
ers
i
t
y
o
f
Pro
fes
s
i
o
n
al
s
,
i
s
w
o
r
k
i
n
g
o
n
s
ev
era
l
res
earc
h
p
r
o
j
ec
t
s
.
H
er
re
s
earc
h
i
n
t
ere
s
t
s
i
n
cl
u
d
e
I
o
T
,
Cy
b
er
Secu
r
i
t
y
,
Si
g
n
a
l
Pro
ces
s
i
n
g
an
d
Mach
i
n
e
L
earn
i
n
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.