Indonesi
an
Journa
l
of
El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
23
,
No.
1
,
Ju
ly
20
21
,
pp.
405
~
413
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v
23
.i
1
.
pp
405
-
413
405
Journ
al h
om
e
page
:
http:
//
ij
eecs.i
aesc
or
e.c
om
Bangla n
um
er
ical
sign langua
ge re
cognitio
n usin
g con
voluti
o
nal
neural n
etworks
F.
M.
Jave
d
Mehedi
Sh
am
rat
1
,
S
ovon C
ha
kr
abor
ty
2
,
Md. M
asum B
il
lah
3
, Moumi
ta K
ab
ir
4
, Naz
mus
Shakib
S
hadi
n
5
, S
il
vi
a
S
anj
ana
6
1,3
Depa
rtment
of
Software
Engi
n
ee
ring
,
Daffod
il
Inte
rna
ti
ona
l
Un
ive
rsit
y
,
Dhak
a,
Bangl
ad
esh
2,4
Depa
rtment
of
Com
pute
r
Sci
en
ce
and Engi
ne
ering,
Europ
ea
n
Univer
sit
y
of
B
anglade
sh,
Dh
aka,
B
angl
ad
esh
5,6
Depa
rtment
of
Com
pute
r
Sci
en
ce
and Engi
ne
ering,
Ahs
anul
l
ah University
of
Sc
ie
nc
e and
T
ec
hn
olog
y
,
Dhak
a,
Bangl
ad
esh
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
r 31,
2021
Re
vised
Ma
y
1
, 2021
Accepte
d
J
un
17, 202
1
The
amount
of
d
ea
f
and
m
ute
ind
ivi
dual
s
on
the
e
art
h
is
r
ising
a
t
a
n
al
arming
rat
e
.
Bangl
ad
esh
has
about
2.
6
mi
ll
ion
peop
le
who
are
unabl
e
to
i
nte
ra
ct
with
the
comm
unity
using
la
nguag
e.
Hea
ring
-
impair
e
d
ci
tizens
in
B
an
gla
desh
use
Bangl
ad
eshi
sign
la
nguage
(BS
L)
as
a
m
ea
ns
of
comm
unic
at
i
on.
In
thi
s
art
i
cl
e
,
we
prop
ose
a
new
m
ethod
for
Benga
l
i
sign
la
nguag
e
rec
ogni
ti
on
base
d
on
de
ep
convol
ut
iona
l
neur
al
n
et
works
.
Our
fra
m
ewo
rk
emplo
y
s
convol
uti
on
al
n
e
ura
l
n
et
works
(CNN
)
to
le
a
rn
fro
m
the
images
in
our
dataset
and
interpre
t
h
a
nd
signs
from
in
put
images.
Che
cki
ng
th
ei
r
co
l
lecti
ons
of
te
n
indi
c
at
ions
(we
used
te
n
se
ts
of
images
with
31
disti
nct
signs)
for
a
total
o
f
310
images.
The
proposed
s
y
st
em
ta
kes
snapshots
from
a
vide
o
b
y
using
a
webc
am
with
a
ppl
y
ing
a
computer
vision
-
base
d
appr
oac
h
.
After
that,
i
t
compare
s
those
pho
tos
to
a
pre
v
iousl
y
traine
d
d
at
ase
t
gen
erate
d
with
CNN
and
display
s
the
Benga
li
num
be
rs
(
০
-
৯
).
After
esti
m
at
ing
th
e
m
odel
on
our
dat
ase
t,
we
obtained
an
over
a
ll
ac
cur
acy
of
99.
8%.
W
e
want
t
o
strengt
he
n
thi
ngs
as
far
as
we
ca
n
to
m
ake
sile
nt
contac
t
with
th
e
m
aj
ority
o
f
socie
t
y
as
sim
ple
as
prob
ab
le
.
Ke
yw
or
d
s
:
Ba
nd
le
t t
ra
ns
m
issi
on
Be
ng
al
i si
gn la
ngua
ge
Deep C
N
N
Digit rec
ogniti
on
Im
age p
r
ocessi
ng
This
is an
open
acc
ess arti
cl
e
un
der
the
CC
BY
-
SA
l
ic
ense
.
Corres
pond
in
g
Aut
h
or
:
F.
M.
Jave
d
M
ehed
i
Sh
am
rat
Dep
a
rtm
ent o
f Sof
t
war
e
E
ng
i
neer
i
ng
Daffodil
Inter
na
ti
on
al
Uni
ver
s
it
y
102/1, S
ukra
ba
d,
Mi
r
pur
Roa
d,
D
hak
a
12
07,
Ban
glades
h
Em
a
il
: jav
edm
ehed
ic
om
@g
m
ai
l.com
1.
INTROD
U
CTION
Sign
la
ngua
ge
recog
niti
on
t
echnolo
gies
a
r
e
us
e
d
to
i
de
ntify
ind
ic
at
io
ns
of
num
ber
s,
al
phabets
,
phrases
,
or
s
om
e
oth
er
sig
ns
,
su
c
h
as
t
raffic
sign
al
ha
nd
m
ov
em
ents.
S
om
e
sci
entist
s
are
f
oc
us
ed
on
r
eal
-
tim
e
sy
m
bo
l
rec
ogni
ti
on
,
wh
il
e
ot
her
s
are
f
oc
us
ing
on
sta
ti
c
pi
ct
ur
es.
A
rtific
ia
l
neural
netw
ork
-
base
d
a
ppr
oach
e
s
for
real
-
tim
e
Ame
rican
sig
n
la
ngua
ge
(
AS
L
)
te
rm
[1
]
and
al
ph
a
bet
[
2]
re
cogniti
on
ha
ve
recently
bee
n
us
e
d.
Hand
si
gn
rec
ogniti
on w
it
h
th
e Mi
cro
s
of
t se
ns
or syst
em
[
3] is us
ed
in t
he rep
or
t t
o detec
t
sign
al
s
for
tw
o
real
-
world
ap
plica
ti
on
s:
arit
hm
et
ic
cal
culat
ion
an
d
the
r
oc
k
-
pa
pe
r
-
sci
ss
or
s
ga
m
e,
with
a
m
e
an
acc
ur
acy
of
over
90%
for
ASL.
Deep
le
arn
i
ng
has
been
us
e
d
to
identify
sign
s
in
I
nd
ia
n
sign
la
ng
uag
e
s
[
4],
[5
]
,
A
ra
bic
sign
la
nguag
e
[6], a
nd o
t
her la
ngua
ges.
In
t
he
m
od
er
n
era,
dee
p
le
a
rn
i
ng
a
nd
m
a
chine
le
a
rn
i
ng
[7
]
-
[
11
]
besides
Be
ng
al
i
si
gn
la
ngua
ge
(BSL)
rec
ognit
ion
st
ud
ie
s
ha
ve
gott
en
a
l
ot
of
at
te
ntio
n,
a
nd
var
i
ous
a
pproaches
for
im
pl
e
m
enting
a
Be
ng
al
i
sign
la
ng
uag
e
recog
niti
on
sc
hem
e
hav
e
be
en
sug
gested
.
Kar
m
ok
ar
et
al
.
[12]
introd
uc
ed
a
syst
e
m
of
BS
L
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
405
-
41
3
406
recog
niti
on
us
i
ng
a
ne
ur
al
net
work
in
2012,
with
a
93
pe
rc
ent
accuracy.
They
us
e
d
a
dataset
of
disti
nct
sk
in
ton
es
but
t
he
sam
e
con
te
xt,
al
lowing
for
f
ast
identific
at
ion.
Ra
ham
an
et
al.
sug
gested
a
com
pu
te
r
-
bas
e
d
Be
ng
al
i
sig
n
l
angua
ge
recog
niti
on
syst
em
in
20
14
[
13
]
.
Ra
him
et
al
.
[14]
ap
plied
t
he
backp
ropa
ga
ti
on
appr
oach
of
ANN
to
i
den
t
ify
ind
ic
at
ions
of
ce
rtai
n
popula
r
Be
ngal
i
te
r
m
s
in
2015.
Fin
ge
rtip
fin
der
al
gorithm
s
wer
e
us
ed
for
BSL
by
nu
m
ero
us
r
esearche
rs
in
2015
an
d
2016
[
15
]
-
[
17
]
.
F
or
t
he
ide
ntific
at
ion
of
Be
ng
al
i
Sig
n
Lan
gu
a
ge,
m
any
m
e
tho
ds
ba
sed
on
A
NN
[18]
ha
ve
be
en
s
uggested
.
The
pr
eci
sio
n
of
t
he
m
ajo
rity
of
t
he
te
chn
iq
ues
s
ee
m
s
to
be
prom
isi
ng
.
They
do,
howe
ve
r,
hav
e
dr
a
w
backs
du
e
t
o
the
usa
ge
of
lim
it
ed
dataset
s,
a
co
ntro
ll
e
d
backdro
p
or
scenari
o,
a
nd
in
certai
n
sit
ua
ti
on
s,
m
ajo
r
m
ist
akes
in
lig
htin
g
eff
ect
s
or
s
kin
ton
e,
w
hich
th
ey
m
on
it
or
to
pr
e
ve
nt
m
or
e
com
plica
ti
on
s.
The
a
ppro
ac
he
s
that
do
util
iz
e
neural
netw
orks
[
12
]
us
e
a
lot
of
pr
eprocessi
ng,
w
h
ic
h
is
n'
t
nece
ssaril
y
su
it
able
f
or
real
-
ti
m
e
app
li
cat
io
ns
.
U
sing
a
CNN,
the
pr
opos
e
d
syst
em
reco
gniz
es
only
the
Ba
ng
la
num
eric
di
gits
(C
N
N)
.
It
sens
ed
the
sym
bo
l
wit
h
on
ly
on
e
h
a
nd. It
is
sp
li
t i
nto
t
wo part
s: t
he
trai
ne
d
a
nd the
sig
ns detec
ti
on
par
t.
The
f
ollo
wing
sect
io
n
f
ollow
s
t
he
sam
e
structu
re.
This
porti
on
include
s
the
m
os
t
recent
dev
el
op
m
ents
in
Be
ng
al
i
Sig
ns
I
den
ti
ficat
ion.
Sect
ion
ii
descr
i
bes
the
researc
h
m
et
h
od
f
or
m
od
el
i
ng
th
e
whole
syst
em
.
The
ii
i
sect
ion
looks
at
the r
es
ults
of
the
syst
e
m
tha
t
has
be
en
est
ablis
hed.
Sect
ion
i
v
en
ds
with
a theory
, fl
aws
, and p
rop
os
al
s
for
pote
ntial
stud
y.
2.
METHO
DOL
OGY
This
pro
pose
d
syst
e
m
ta
kes
t
he
im
age
fr
om
a
ca
m
era
and
then
pr
e
proces
ses
the
im
ages.
To
ide
ntify
the
i
m
age
struc
ture
fir
stl
y
the
syst
e
m
app
li
es
b
an
dlet
tran
sform
at
ion
on
i
m
ages,
then
it
app
li
es
the
lo
gar
it
hm
rep
la
ce
te
ch
ni
qu
e
t
o
c
on
tr
ol
the
extra
li
gh
t
eff
ect
s
on
i
m
ages.
To
fix
ed
up
th
e
lo
w
-
res
olu
ti
on
im
age,
the
syst
e
m
al
so
use
d
t
he
D
-
LB
P
te
ch
nique.
The
n
it
m
eas
ur
es
the
pictu
re
dim
ension
[19].
As
d
im
e
ns
io
n
cal
culat
ion
is d
on
e
,
it
segm
ents
the
s
kin
c
ol
or
from
the
pictu
re
an
d
t
ran
s
f
orm
s
it
into
a
bi
na
ry
im
age,
rem
ov
e
s
the
broad
blob
s
f
ro
m
the
im
age,
an
d
us
es
deep
le
ar
ning
te
chn
iq
ues
to
const
ru
ct
a
qual
ifie
d
dataset
[
19
]
.
Cl
assifi
es
with
the
qu
al
ifie
d
dataset
after
the
recog
nizer
sect
ion
an
d
id
entifi
es
the
num
erical
dig
it
of
the
Ba
ng
la
.
I
n
this
pa
per
,
we
us
e
d
deep
le
a
rn
i
ng
m
et
ho
ds
for
Be
ng
al
i
num
erical
sign
detect
ion
.
In
Fig
ure
1
we
hav
e
d
is
play
ed
the e
ntire
pro
pose
d
m
od
el
d
ia
gr
am
.
Figure
1.
Pro
pos
e
d
m
od
el
d
ia
gr
am
2
.
1.
Data c
ollec
tion
We
c
ollec
t
i
m
a
ges
from
the
vi
deo
f
ram
e
us
in
g
the
we
bcam
.
To
ca
pture
the
i
m
age,
we
us
e
d
a X
ia
om
i
Vidlok W
77
w
ebcam
.
W
e
use
Ry
zen
39
00X
Core
i
3
3.90
(
ba
se
sp
ee
d)
G
H
z
PC
with
32 G
B
RAM f
or
runnin
g
the
syst
em
.
The
dataset
co
ns
i
sts
of
te
n
hand
signs
(a
t
otal
of
310
pictu
res
),
with
eac
h
si
gn
i
ng
cl
ass
co
ns
ist
in
g
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ba
ng
l
a numeri
cal
sig
n
l
angu
ag
e
rec
ogniti
on
us
in
g
c
o
n
v
o
l
u
t
i
o
n
a
l
…
(
F.
M. J
aved Mehe
di Sh
am
r
at)
407
of
31
im
ages
captu
red
at
va
rio
us
le
ng
t
hs
,
or
ie
ntati
on
s
,
a
nd
inte
ns
it
ie
s.
Be
ng
al
i
nu
m
erical
sign
la
ngua
ge
is
dep
ic
te
d
by
the
photo
s
.
W
e
us
e
d
the
data
f
or
the
trai
n
a
nd
te
st
ou
r
m
odel
to
reco
gnit
ion
Be
ng
al
i
n
um
erical
sign
s
. In Fi
gur
e 2
we sh
owed
the
dataset
im
ages s
am
ple.
Figure
2
.
Datas
et
s i
m
ages s
am
ples
2
.
2
.
Data pre
processin
g
2.2.1. B
an
dle
t
t
ra
nsform
at
i
on
A
3x
3
tra
ns
f
orm
at
ion
m
a
trix
is
need
e
d
for
pe
rsp
ect
ive
co
nversi
on.
W
el
l
after
the
tra
ns
it
ion
,
strai
ght
li
nes
would
re
m
ai
n
strai
gh
t.
We'
ll
need
fou
r
po
i
nts
from
t
he
input
i
m
age
and
f
our
poi
nts
from
the
ou
tp
ut
i
m
age
to
fin
d
t
his
tra
ns
it
ion
m
at
rix.
Th
ree
of
t
he
f
our
poi
nts
co
uld
not
be
in
sync
with
on
e
a
nothe
r.
T
hen
us
e
the
m
et
ho
d
cv
2
get
pe
rsp
ect
i
ve
tra
ns
f
orm
t
o
locat
e
the
tra
ns
it
ion
m
at
rix.
For
this
3x3
t
ran
sit
io
n
m
at
ri
x,
a
dd
cv2.w
a
r
p
Pe
rs
pecti
ve.
I
n
Fi
gure
3 we
exhib
it
ed
the
ou
t
put
of b
a
ndle
t t
ransform
at
ion
afte
r
im
ple
m
entati
on.
Figure
3
.
The
ou
t
pu
t
of
bandl
et
tran
s
form
at
i
on
2.2.2. L
ogarit
hm
r
epl
ace
Each
pix
el
'
s
va
lue
is
re
placed
with
it
s
lo
gar
i
thm
value
in
a
log
tra
nsfo
rm
a
ti
on
.
As
s
how
n
in
(
1)
c
an
be use
d
to
d
esc
ribe
l
og transf
orm
ation
s.
=
log
(
+
1
)
(1)
The
outp
ut
an
d
inp
ut
im
age
pix
el
values
a
re
s
and
r
,
res
pect
ively
,
and
c
is
a
con
sta
nt.
Sin
ce
the
input
i
m
age
has
a
pix
el
intensit
y
of
0,
log
(0)
eq
ua
ls
infin
it
y,
an
d
each
of
the
input
i
m
age'
s
pix
el
values
is
gi
ven
a
su
m
o
f on
e
. As
a r
es
ult,
1
is a
pp
li
ed
to
t
he
m
ini
m
u
m
a
m
ou
nt
to
re
nder it at
le
ast
1
.
Ex
pands
t
he
im
age'
s
dar
k
pi
xels
th
us
c
ompressi
ng
the
i
m
age'
s
li
gh
te
r
pix
el
s.
The
c
om
bin
at
ion
of
m
axi
m
u
m
and
m
ini
m
u
m
a
m
p
li
tud
e
val
ues
is
ref
e
rr
e
d
to
as
dynam
ic
ran
ge
.
Lo
wer
val
ues
are
rem
ov
e
d
wh
e
re
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
405
-
41
3
408
the
im
age'
s
dy
nam
ic
sp
ect
rum
exceeds
that
of
the
viewin
g
syst
em
.
W
e
us
e
l
og
tra
ns
f
orm
to
so
lve
thi
s
issue
.
The
lo
g
tra
ns
f
or
m
at
ion
com
pr
ess
es
the
dynam
ic
ran
ge
fi
rst,
the
n
upsca
le
s
the
picture
to
the
dis
play
dev
ic
e'
s
dynam
ic
ran
ge
.
L
ow
e
r
values
are
boos
te
d
i
n
this
way,
re
su
l
ti
ng
in
a
far
m
or
e
detai
le
d
pi
ct
ur
e.
I
n
Fi
gur
e
4
we
sh
owe
d
t
he
im
age a
fte
r
a
pply
ing
l
og
a
rithm
r
eplace
.
Figure
4
.
Afte
r
apply
ing
l
ogar
it
h
m
r
eplace
m
ent
2.2.3. D
-
LBP
Tw
o
com
ple
m
entary
m
easur
es:
local
sp
at
ia
l
patte
rn
s
a
nd
gray
scal
e
co
m
par
ison
,
acc
ordin
g
to
th
e
LBP
op
e
rato
r,
will
char
act
erize
two
-
dim
ension
al
surface
te
xtures.
T
he
ori
gin
al
LBP
op
e
rator
produces
la
bels
for
im
age
pix
e
ls
by
th
res
ho
l
di
ng
t
he
3x3
ne
ighbor
hood
of
each
pix
el
with
the
ce
nter
va
lue
an
d
t
reati
ng
th
e
resu
lt
as
a
bin
a
ry
integer.
T
he
te
xtu
re
desc
ri
ptor
is
then
the
histogram
of
t
hese
28
=
256
diff
e
re
nt
m
arks
.
This
op
e
rato
r,
w
he
n
c
ouple
d
with
a
sim
ple
local
co
ntrast
m
easur
e,
achi
eved
e
xcell
ent
uns
up
e
rv
ise
d
te
xture
segm
entat
ion
resu
lt
s.
The
L
BP
op
e
rator
is
denoted
by
the
fo
ll
owin
g
no
ta
ti
on:
Ru2
;
LBPP;
The
su
bsc
ript
ind
ic
at
es
that
the
operat
or
is
bein
g
us
e
d
in
t
he
(P,
R)
nei
ghbo
rho
od.
The
su
pe
rsc
ript
u2
denotes
us
i
ng
on
ly
sta
nd
a
rd
iz
e
d
pa
tt
ern
s
a
nd
la
be
li
ng
t
he
resid
ual
patte
rn
s
wi
th
a
sin
gle
dot
.
A
fter
obta
ini
ng
the
LBP
la
beled
picture
fl
(
x,
y
),
t
he
LB
P
histogram
can
be
i
den
ti
fie
d
as
(
1)
an
d
i
n
Fi
gure
5
we
s
howe
d
the
outp
ut
of
D
-
LB
P
after im
ple
m
entat
ion
.
=
∑
{
1
(
,
)
=
}
,
=
0
,
…
,
−
1
,
(2)
Figure
5
.
Afte
r
apply
ing
D
-
L
BP on t
he
im
a
ge
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ba
ng
l
a numeri
cal
sig
n
l
angu
ag
e
rec
ogniti
on
us
in
g
c
o
n
v
o
l
u
t
i
o
n
a
l
…
(
F.
M. J
aved Mehe
di Sh
am
r
at)
409
2
.
3
.
Pr
opose
d conv
olu
tion
neural
netw
or
k (CNN
)
m
odel
2.3.1.
Dimensi
on
c
alcula
tio
n
Jo
int
P
ho
t
ogra
ph
ic
E
xp
e
rts
G
rou
p
(JP
E
G
)
is
on
e
of
the
m
os
t
com
m
on
i
mage
com
pr
essi
on
m
et
ho
ds.
The
hea
der
s
(
f
irst
fe
w
byte
s)
in
t
he
m
ajo
rit
y
of
file
form
at
s
pro
vid
e
va
luable
detai
ls
about
t
he
file
.
JPE
G
head
e
rs,
f
or
e
xam
ple,
pro
vid
e
detai
ls
s
uch
a
s
hei
gh
t,
w
ei
ght,
c
olor
de
pth
(
gr
ay
scal
e
or
RGB),
an
d
so
on.
We
fin
d
the
res
olut
ion
of
a
J
PEG
i
m
age
with
out
util
iz
ing
a
ny
e
xter
nal
li
br
a
ries
in
t
his
s
of
t
w
are
by
rea
ding
thes
e
head
e
rs.
2.3.2.
Segm
en
t
skin
co
l
or
This
m
e
tho
d
is
on
ly
us
ed
to
de
te
ct
the
color
of
a
hum
an
being,
su
c
h
as
a
han
d
or
a
face
[
20
]
.
It
rea
ds
the
RGB
pictu
re
an
d
the
n
m
easur
e
s
the
im
age'
s
propor
ti
ons.
It
senses
t
he
sk
in
c
olor
af
te
r
trans
form
ing
the
RGB pict
ure t
o YCbC
r [20]
, [
21
]
.
In F
i
gure
6,
t
he
s
kin col
or is se
gm
ented fr
om
a d
iffe
rent
b
ack
gro
und.
Figure
6
.
S
kin
colo
r
se
gm
entat
ion
2.3.3.
E
xt
r
ac
t N lar
ges
t
bl
obs
This
r
ole
assist
s
in
the
ide
ntif
ic
at
ion
of
broa
d
bl
ob
s
of
an
i
tem
in
a
bin
ar
y
picture.
It
ta
kes
the
blob
reg
i
on
after
get
ti
ng
al
l
of
t
he
bl
ob
p
r
operti
es from
the
bi
nar
y
pictu
re.
It
al
so
detect
s
the
la
r
gest
e
ntit
y
fr
om
the
chosen
boun
da
ries
an
d
est
abli
sh
es
bounda
rie
s
in
the
or
igi
na
l
picture.
T
his
m
et
ho
d
em
ploy
s
a
custom
featur
e
to r
em
ov
e t
he N lar
gest
blobs
f
r
om
the b
in
ar
y pict
ur
e
seen
i
n
Fi
gure
7.
Figure
7
.
Ext
ra
ct
the lar
g
est
blo
bs f
ro
m
the
bin
ary im
age
2.3.4.
Gener
ate tr
ain n
e
twor
k
This
m
et
ho
d
e
ven
us
e
d
a
w
ebcam
to
bu
il
d
the
qu
al
ifie
d
net
wor
k
but
pr
e
processe
d
the
we
bcam
i
m
ages
util
iz
ing
the
sam
e
process
a
s
m
entio
ne
d
a
bove
.
T
his
im
plies
that
the
trai
ne
d
da
ta
set
was
generate
d
us
in
g
t
he
la
r
ge
st
bin
a
ry
bl
ob
of
an
im
age.
To
beg
i
n,
m
ulti
ple
photos
of
t
he
sam
e
sign
w
her
e
nee
de
d.
F
or
each
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
405
-
41
3
410
sy
m
bo
l,
ap
pro
xim
a
te
ly
31
phot
os
wer
e
ta
ken
[
22
]
.
To
create
the
div
i
sion
s
,
the
sug
gested
m
et
hod
stores
var
i
ou
s
sig
n
phot
os
int
o
se
pa
rate
f
old
e
rs.
T
h
is
m
achine
w
as
giv
e
n
31
ph
oto
s
t
o
cre
at
e
a
qual
ifie
d
network
t
o
i
m
pr
ove
acc
uracy
.
Af
te
r
ta
ki
ng
pictu
res,
t
hi
s
m
et
ho
d
ge
ne
rates
a
qual
ifi
ed
dataset
util
iz
ing
a
c
onvol
ut
ion
al
neural
net
work
[23]
m
e
tho
dol
og
y.
For
the
e
du
cat
e
d
net
work,
the
propose
d
f
ram
ewo
r
k
u
sed
the
Alex
ne
t
of
the
dee
p
le
ar
ni
ng
process
.
T
his
fram
ewo
rk,
in
pa
rtic
ular
,
e
m
plo
ys
the
conv
olu
ti
onal
netw
ork'
s
Tra
ns
fe
r
Learn
i
ng
proc
ess.
F
or
the
c
onvoluti
onal
la
y
er
data,
it
resized
t
he
bin
a
ry
pictu
re
blob
i
nto
a
[
28
x28].
The
n,
us
in
g
a
ra
ndom
sel
ec
ti
on
process,
it
ge
ne
r
at
ed
a
qual
ifie
d
dataset
a
nd
a
resea
rch
dataset
.
The
te
rm
"q
ualifi
ed
dataset
"
re
fer
s
to
a
cl
assifi
ca
ti
on
database
,
wh
e
reas
"t
est
dataset
"
re
fer
s
to
a
cl
assifi
ca
ti
on
database
t
hat
is
us
e
d
in
co
njun
ct
ion
with
the
trai
ned
dataset
.
Fo
r
the
qu
al
if
ie
d
dataset
,
the
m
achine
us
ed
750
im
ages
a
nd
for
the
evaluati
on
dataset
,
it
us
ed
250
i
m
ages.
In
it
ia
l
le
arn
rate
=
0.
0001,
Ma
x
Ep
ochs
=
20
,
and
Mi
ni
batch
siz
e
= 64
wer
e
u
se
d.
2
.
4
.
Clas
si
fy
i
ng
tr
ained
net
w
ork
This
syst
em
lo
ads
the
t
rainin
g
dataset
at
fi
r
st,
then
t
he
sig
n
ha
s
bee
n
ta
ke
n
w
he
n
the
fi
gure
window
is
opene
d.
T
his
syst
em
can
detect
Ba
ngla
nu
m
erical
signs
only
usi
ng
one
ha
nd.
T
he
pro
po
se
d
syst
em
has
us
e
d
the
we
bc
a
m
i
m
age
for
recog
niti
on
of
the
sign.
Wh
e
n
com
plete
d
th
e
la
rg
est
bi
nary
blo
b
im
ag
e
then
it
do
e
s
cl
assify
with
the
c
us
to
m
trai
nin
g
dat
aset
that
wa
s
first
cre
at
ed.
The
pro
posed
m
od
el
us
ed
A
le
xN
et
cl
assifi
er to
det
ect
the signs.
F
inall
y, it
g
ives
the outcom
e th
at
is shown i
n Fi
gure
8.
Wh
e
n
t
he
fig
ure
portal
is
rai
sed,
this
de
vice
load
s
th
e
trai
ning
dataset
fi
rst,
the
n
the
sy
m
bo
l.
On
ly
on
e
ha
nd
is
inten
ded
to
ide
ntify
a
Ba
ngla
nu
m
erical
sym
bo
l.
Fo
r
t
he
identific
at
io
n
of
t
he
sym
bo
l,
the
pro
po
se
d
syst
em
us
ed
a
we
bc
a
m
picture.
It
beg
i
ns
by
pre
proces
sin
g
the
pi
ct
ur
e
with
the
te
chn
iq
ue
m
entioned
in
the
2.1,
2.2
,
and
2.3
sect
io
ns
ab
ove.
When
the
la
r
gest
bina
ry
bl
ob
pictu
r
e
is
over
,
it
ca
n
be
cl
assifi
e
d
us
in
g
the
c
us
tom
trai
ning
dataset
th
at
was
ge
ner
at
ed
at
t
he
be
ginnin
g.
Finall
y,
i
t
pro
du
ce
s
the
resu
lt
(
০
,
১
,
২
,
৩
,
৪
,
৫
,
৬
,
৭
,
৮
,
৯
)
ind
ic
at
e
d
in
Fi
gure
8.
Figure
8
.
Be
ng
al
i nu
m
erical
sign
la
ngua
ge r
ecognit
ion
3.
RESU
LT
S
A
ND
D
IS
C
USS
ION
To
dev
el
op
th
e
area
of
si
gn
la
ngua
ge
inte
rpretat
ion
,
the
pro
po
s
ed
m
od
el
has
us
e
d
a
var
ie
ty
of
appr
oach
es
.
We
dev
el
op
e
d
the
propose
d
m
od
el
to
im
pro
ve
the
te
c
hn
i
qu
e
on
Be
ng
al
i
nu
m
erical
sign
detect
ion.
The
accuracy
of
ea
ch
sig
n'
s
identific
at
ion
is
cal
culat
ed
by
this m
et
ho
d
f
or
ex
pe
rim
ental
resu
lt
s.
W
e
m
easur
ed
the
perform
ance
of
tw
o
m
od
el
s
us
in
g
pr
eci
sio
n,
recall
,
true
negat
ive
rate,
and
accu
racy
afte
r
com
pl
et
ing
the
trainin
g
a
nd te
sti
ng
phase.
Th
e f
or
m
ulas that we
us
e
d
a
re as
show
n
i
n (3)
-
(
4)
:
=
+
(3)
=
+
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ba
ng
l
a numeri
cal
sig
n
l
angu
ag
e
rec
ogniti
on
us
in
g
c
o
n
v
o
l
u
t
i
o
n
a
l
…
(
F.
M. J
aved Mehe
di Sh
am
r
at)
411
=
+
+
+
+
(5)
Ra
te
=
+
(6)
Table
1
s
hows
how
the
syst
em
us
es
25
te
sts
on
ha
nd
sig
ns
for
eac
h
di
git,
with
9
sig
ns
c
orrectl
y
recog
nized
at
the
m
axi
m
u
m
pea
k
a
nd
1
sign
rec
ogniti
on
le
vel
i
s
c
om
par
at
ively
low.
T
he
preci
sion
is
achieve
d
100%
f
or
eac
h
si
gn
(
০
-
৯
)
,
be
sid
es
Re
cal
l
al
so
gain
100%
f
or
9
sig
n
s
,
exce
pt
one
Be
ngal
i
sign
(
৩
)
.
Fo
r
al
l
sign
s
,
Tru
e
ne
gative
gain
0%,
it
car
ried
a
huge
im
pact
on
ov
e
rall
accuracy.
Fin
al
ly
,
ou
r
m
od
el
fo
r
th
e
Be
ng
al
i
num
er
ic
al
sign
’s
la
nguag
e
detect
ion
achieve
d
overa
ll
99
.8
%
acc
uracy
.
Fig
ure
9
de
picte
d
the
val
ue
of
ever
y
te
st
re
sul
t
(P
recisi
on,
Re
cal
l,
Tru
e
N
egati
ve,
Acc
uracy
)
f
or
eac
h
sign.
Fig
ure
10
s
hows
t
he
a
ccur
ac
y
evaluati
on
of ten se
par
at
e
nu
m
erical
Beng
al
i si
gn
s
.
Table
1
.
O
utco
m
es f
or eac
h n
um
erical
Bengal
i si
gn
Ban
g
la Sign
Precisio
n
Recall
Tr
u
e Negativ
e
Accurac
y
০
100%
100%
0%
100%
১
100%
100%
0%
100%
২
100%
100%
0%
100%
৩
100%
98%
0%
98%
৪
100%
100%
0%
100%
৫
100%
100%
0%
100%
৬
100%
100%
0%
100%
৭
100%
100%
0%
100%
৮
100%
100%
0%
100%
৯
100%
100%
0%
100%
Figure
9
.
Pr
eci
sion, recall
, t
rue
ne
gative,
acc
ur
acy
gr
a
ph
for
eac
h
sig
n
Figure
10
.
C
om
par
ing
the
ac
cur
acy
of ten
di
ff
ere
nt Ben
gal
i nu
m
erical
sig
ns
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
23
, N
o.
1
,
Ju
ly
20
21
:
405
-
41
3
412
In
T
a
ble
2
we
com
par
ed
s
ome
pr
e
vious
wor
ks
,
wh
ic
h
are
on
Be
ngal
i
an
d
dif
fere
nt
la
ngua
ge
si
gn
s
detect
ion.
T
he
m
et
ho
dolo
gi
es
they
ha
ve
b
ee
n
us
e
d,
arti
cl
e
publish
ing
ye
ar
,
a
nd
accuracy
of
their
m
et
ho
dolo
gies,
we
de
picte
d
in
Table
2.
T
hey
ha
ve
m
any
si
m
il
ari
ti
es
with
our
pro
pose
d
syst
em
,
a
few
m
et
ho
dolo
gies
they
ha
ve
be
en
us
ed
inclu
ding
CN
N,
ar
ti
fici
al
neu
ral
netw
ork,
sup
port
vecto
r
m
a
chine,
Conve
x
hu
ll
m
et
ho
d,
3D
co
nvolu
ti
onal
ne
ur
al
netw
ork
(
CNN
),
key
m
axim
u
m
cur
vat
ur
e
points
,
s
roke
s
ub
-
segm
ent
vecto
rs,
e
quiv
olu
m
e
tric
par
ti
ti
on,
Gabo
r
filt
er,
ke
rn
el
PC
A,
m
r
phologica
l
ope
rati
on,
S
VM.
Most
of
the
m
od
el
s
cha
ng
e
d
t
he
im
ages
into
gray
sca
le
and
di
rectl
y
app
ly
s
om
e
sk
in
detect
ion
al
gorithm
s
to
de
te
ct
the
sh
a
pe
of
the
ha
nd.
O
n
the
othe
r
ha
nd,
our
pr
opos
e
d
m
od
el
pr
e
processe
s
the
i
m
ages
as
well
to
get
the
highest
accuracy
on
te
st
an
d
trai
ning
phases,
the
n
we
a
pp
ly
c
orr
esp
ondingly
di
m
ension
cal
cu
la
ti
on
,
se
gm
ent
sk
in
colo
r,
ext
ract
N
la
rg
e
st
blob
s,
an
d
trai
ne
d
our
net
work
at
the
processi
ng
phase.
Fi
na
ll
y,
we
app
li
ed
the
Alex
Net cla
ssier to classi
fy the im
ages.
Fo
r
tun
at
el
y, our
m
od
el
ac
hieve
d
the h
i
gh
est
acc
ur
acy
(9
9.8%)
a
m
on
g
the ex
ist
in
g rel
at
ed
w
orks
.
Table
2
.
C
om
par
at
ive a
naly
sis o
n pr
e
vious
re
la
te
d
w
orks
Year
Metho
d
s
Ov
erall
Accurac
y
G.
A
.
Rao
et al
.
[
2
4
]
2018
CNN
9
2
.88
%
Ch
o
wd
-
h
u
ry
et al
.
[
2
5
]
2017
Artif
icial neu
ral
n
e
two
rk, Sup
p
o
rt
v
ecto
r
m
achi
n
e,
Co
n
v
ex
hu
ll
m
e
th
o
d
8
4
.11
%
El
-
b
ad
aw
y
et al
.
[
2
6
]
2017
3
D Co
n
v
o
lu
ti
o
n
al Neural
N
etwo
rk (
CNN)
98%
Kai
m
al
[
2
7
]
2017
Key
Maxi
m
u
m
Cu
rvatu
re
Po
in
ts, St
r
o
k
e Sub
-
seg
m
en
t
Vectors,
Equ
iv
o
lu
m
etric
Pa
rtition
8
5
.4%
Ud
d
in
et al
.
[
2
8
]
2016
Gab
o
r
f
ilter,
Kerne
l PCA,
Morp
h
o
lo
g
ical op
eration
,
SV
M
9
9
.5%
Ou
r
Prop
o
sed
Mod
el
2021
CNN
9
9
.8%
4.
CONCL
US
I
O
N
Ma
chine
le
ar
ni
ng
an
d
deep
le
arn
in
g
doin
g
a
gr
eat
ro
le
in
rece
nt
tim
es
in
i
m
age
pr
oc
essing,
the
m
edical
sect
or
and
var
i
ous
pu
rposes.
T
o
ide
ntify
Be
ng
al
i
num
erical
sign
s
from
the
inp
ut
i
m
ages,
our
cu
rr
e
nt
m
od
el
ta
kes
aro
un
d
a
m
inu
te
.
Ou
r
ai
m
is
to
r
end
e
r
this
proc
ess
as
s
m
oo
th,
fast,
an
d
real
-
ti
m
e
as
po
ssible.
The
dataset
that
we
hav
e
us
e
d
f
or
trai
ning,
te
sti
ng,
a
nd
validat
ion
c
onta
ins
th
e
Be
ng
al
i
sig
n
(
০
-
৯
)
.
W
e
i
ntend
t
o
com
plete
ou
r
da
ta
set
by
incl
udin
g
al
l
Be
ng
a
li
a
lph
a
bets.
We
ha
ven
'
t
inc
orp
or
at
ed
gestur
e
rec
ogniti
on
i
n
our
researc
h,
nor
ha
ve
we
incl
uded
the
rec
ogni
ti
on
of
te
rm
s
a
nd
f
ull
sente
nc
es,
s
o
we'
re
e
xc
it
ed
to
em
bed
this
into
ou
r
m
od
el
.
W
e
i
nten
d
to
integrate
our
f
ram
ewo
rk
i
nto
a
portable
c
om
pu
te
r
and
I
oT
m
e
dia,
sp
eci
fical
ly
sm
artph
on
e
s,
and
operate
it
as
an
a
ppli
cat
ion
afte
r
im
plem
enting
bo
th
of
these
m
et
hods
.
It
w
ould
be
conve
niently
avail
able
to
al
l
people
in
this
m
ann
er
,
assist
ing
us
in
the
in
corp
or
at
io
n
of
su
c
h
te
chnolo
gi
es
into
our
c
ultur
e
.
REFERE
NCE
S
[1]
N
Sara
wate,
M
Chan
LE
U,
and
Cemil
OZ,
“
A
r
ea
l
-
ti
m
e
Am
eri
c
an
Sign
La
ngu
a
ge
word
rec
ogn
i
ti
on
s
y
st
em
base
d
on
neur
a
l
ne
tworks
and
a
proba
bil
isti
c
m
odel,”
Tur
ki
sh
Journal
of
El
e
ct
rica
l
E
ngine
ering
&
C
omputer
Scienc
e
s
,
vol.
23
,
no
.
1
,
pp
2107
-
2123,
201
3,
doi
:
10
.
3906/elk
-
1303
-
167
.
[2]
B
Garc
ia
and
S
Alar
con
Viesc
a
,
“
Rea
l
-
t
ime
Am
eri
ca
n
Sign
L
angua
ge
R
ec
og
nit
ion
with
Con
volut
ional
Neur
a
l
Networks
,”
Con
vol
uti
ona
l
N
eural
Ne
tworks
for
V
isual
R
ec
ogni
ti
o
n
,
vol
.
2
,
pp
.
225
-
232,
2
012
.
[3]
Z
Z.
Ren,
J.
Yua
n,
J.
Meng
and
Z.
Zha
ng
,
"Robus
t
Part
-
Based
Hand
Gesture
Rec
o
gnit
ion
Us
ing
Kinec
t
Sensor,"
i
n
IEE
E
Tr
ansacti
o
ns on
Multimedi
a
,
vol
.
15
,
no
.
5
,
pp.
1110
-
1120
,
Aug.
2013,
doi:
10.
1109/T
MM
.
2
013
.
2246148.
[4]
S.
Padm
ava
thi
,
M.
S.
Saipreet
h
y
,
an
d
V.
Va
ll
i
a
m
m
ai
,
“
India
n
Sign
La
ngu
age
C
har
acte
r
R
ec
ogn
it
ion
using
Neur
al
Networks,”
I
JC
A
Spe
ci
al
Iss
ue
on
Rece
nt
Tr
ends
in
Pattern
Re
c
ognit
ion
and
Image
Ana
ly
sis
RT
PR
IA
,
vol
.
1
,
pp
:
40
-
45,
2013
.
[5]
N.
Pri
y
adha
rsin
i
and
N.
Ra
je
s
wari
,
“
Sign
Language
Re
cogn
it
io
n
Us
ing
Co
nvolut
ional
Ne
ura
l
Networks,
”
Inte
rnational
Jo
urnal
on
Rece
nt
and
Innov
ati
on
Tr
ends
in
Computing
and
Com
m
unic
at
ion
,
vol
.
5,
no.
6,
pp
.
23
21
-
8169,
2017
.
[6]
O
Al
-
Jarra
h
an
d
Alaa
Ha
la
wa
ni,
“
Rec
ogn
it
io
n
of
gesture
s
i
n
Arabi
c
Sign
La
nguag
e
using
neur
al
n
et
work
s,”
Arti
ficial
Intelli
g
enc
e
and
Soft
C
omputing
,
vol
.
1
33,
no.
1
-
2
,
pp
.
28
-
30,
2006,
do
i
:
10.
1016/S0004
-
3702(01)00141
-
2
.
[7]
F.
M.
Jave
d
M
ehe
di
Sham
rat,
Z.
T
asni
m
,
P.
Ghos
h,
A.
Maj
um
der
and
M.
Z.
Hasan
,
"P
ersona
lization
of
J
ob
Circ
ul
ar
Announce
m
ent
to
Applicants
Us
ing
D
ec
ision
Tr
ee
Cl
assific
a
ti
on
Alg
orit
hm
,
"
2020
IEE
E
Int
ernati
on
al
Confe
renc
e
fo
r
Innov
ati
on
in
Technology
(
INOCON
)
,
Bangl
uru,
India
,
2020
,
pp.
1
-
5,
d
oi:
10.
1109/INOCO
N50539.2020.
92
98253.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ba
ng
l
a numeri
cal
sig
n
l
angu
ag
e
rec
ogniti
on
us
in
g
c
o
n
v
o
l
u
t
i
o
n
a
l
…
(
F.
M. J
aved Mehe
di Sh
am
r
at)
413
[8]
F.
M.
Jave
d
Me
hedi
Sham
rat
,
P.
Ghos
h,
M.
H.
Sadek,
M.
A.
Kaz
i
and
S.
Shult
ana
,
"Im
ple
m
entati
on
of
Mac
hin
e
Le
arn
ing
Algori
thms
to
Dete
ct
t
he
Prognos
is
Rat
e
of
Kidne
y
Di
sea
se,
"
2020
IE
EE
Inte
rnat
iona
l
Confe
renc
e
for
Inno
vat
ion
in
Te
chnol
ogy
(
INOC
ON)
,
Bangl
uru,
I
ndia
,
2020,
pp.
1
-
7,
doi
:
10
.
1109/
INO
CON
50539.
2020.
9298026.
[9]
P.
Ghos
h,
F.
M.
Jave
d
Mehe
di
Sham
rat
,
S.
Shulta
n
a,
S.
Afrin,
A.
A.
Anjum
and
A.
A.
Khan,
"O
pti
m
iz
at
io
n
of
Predic
ti
on
Me
th
od
of
Chronic
Kidne
y
D
isea
se
Us
in
g
Mac
hine
Learni
ng
Algorit
h
m
,
"
2020
15th
I
nte
rnational
Joint
Symposium
on
Arti
fi
c
ial
Int
el
lige
nce
and
Natural
Language
Proce
ss
ing
(
iSA
I
-
NL
P)
,
Bangkok,
Th
ai
l
and,
2020,
pp.
1
-
6,
doi
:
10
.
110
9/i
SA
I
-
NLP51646.2020.
937678
7.
[10]
P.
Ghos
h
et
al
.
,
"Effi
cient
Prediction
of
Cardi
ov
a
scula
r
Dise
ase
Us
ing
Mac
hine
Le
arn
ing
Algori
t
hm
s
W
it
h
Rel
ief
and
LASS
O
Feat
ure
Sel
ecti
on
Te
chni
qu
es,
"
in
IEE
E
A
cc
ess
,
vol
.
9,
pp.
19304
-
19326,
2021,
do
i:
10.
1109/ACCESS
.
2021.
3053759.
[11]
F.M.
Jave
d
Meh
edi
Sham
rat,
As
aduz
z
aman,
A
.
K.
M.
Sa
zzadur
Rahm
an,
R
.
T
.
H.
Tusher
,
and
Za
rrin
T
asnim,
“
A
Com
par
at
ive
A
naly
s
is
of
Parkinson
Disea
se
P
red
iction
Us
ing
Mac
hine
Lear
ning
Approac
he
s,”
Inte
rnationa
l
Journal
of
Scien
ti
fic
&
Technol
ogy
R
ese
arch
,
vol
.
8
,
no
.
11
,
pp:
2
576
-
2580,
2019
.
[12]
B.
C.
Karm
okar
,
K.
M.
R.
Al
am,
and
M.
K.
Sidd
ique
e
,
"Bangl
ad
eshi
sign
la
ngua
ge
rec
ogni
ti
on
e
m
plo
y
ing
neur
al
net
work
ense
m
b
le
,
"
In
te
rnat
iona
l
Journal
o
f
Co
mputer
Appl
i
cati
ons
(IJCA
),
vol.
58,
no
.
16
,
pp
.
43
-
46,
Novem
be
r
2012,
doi
:
10
.
51
20/9370
-
3846
.
[13]
M.
A.
Raha
m
an,
M.
Jasim
,
M.
H.
Ali
and
M.
Hasanuz
za
m
an
,
"Rea
l
-
t
ime
computer
vision
-
bas
ed
Benga
li
Sign
La
nguag
e
r
ec
og
nit
ion,
"
2014
17
th
Int
ernati
onal
Confe
renc
e
on
Computer
and
I
nformation
Tech
nology
(
ICCIT)
,
2014,
pp
.
192
-
1
97,
doi
:
10
.
1109
/ICCIT
e
chn.
201
4.
7
073150.
[14]
Md.
Abdur
Rahi
m
,
Ta
nzi
l
la
h
W
ahi
d,
and
Md.
Khale
d
Ben
Islam
,
“
Visual
Rec
ognit
ion
of
Benga
l
i
Sign
La
nguag
e
using
Artifi
c
ia
l
Neura
l
Ne
twork,”
Inte
rnat
ional
J
ournal
of
Computer
Application
s
,
vol.
94
,
no
.
1
7,
pp.
1
-
5
,
2014
,
doi:
10
.
5120/16
448
-
0572.
[15]
A.
M.
Jarm
an,
S.
Ars
had,
N.
Alam,
and
M.
J.
Islam,
“
An
aut
om
at
ed
bengali
sign
la
nguage
r
ec
ogni
ti
on
s
y
st
e
m
base
don
finge
r
tip
finde
r
a
lgori
th
m
,
”
Inte
rnation
al
Journal
of
Elec
troni
cs
&
Inf
orm
ati
cs
,
vol.
4
,
no.
1,
pp
.
1
-
10
,
2015.
[16]
M.
Hasan,
T
.
H.
Saji
b
and
M.
De
y
,
"A
m
ac
hi
ne
learni
ng
bas
ed
appr
oa
ch
for
the
de
te
c
ti
on
a
nd
rec
ogni
ti
on
of
Bangl
a
sign
language
,
"
2016
Inte
rnational
Confe
renc
e
on
Me
dic
al
Eng
i
nee
ring,
Heal
th
Informatic
s
and
Technol
ogy
(
Me
diTec
)
,
2016,
pp.
1
-
5,
doi: 10.
110
9/MEDITE
C.
20
16.
7835387.
[17]
S.
T.
Ahm
ed
a
nd
M.
A.
H.
Akhand,
"Bangl
a
deshi
Sign
La
n
guage
Re
cogni
t
i
on
using
finge
rt
ip
positi
on
,
"
2016
Inte
rnational
Co
nfe
renc
e
on
Me
dic
al
Eng
ine
erin
g,
Healt
h
Infor
matic
s
and
Technol
ogy
(
Me
diTec
)
,
2016,
pp.
1
-
5,
doi:
10
.
1109/M
EDIT
EC.
2016
.
7
835364.
[18]
Golam
Ka
y
as,
S
akha
wa
t
Hos
sai
n,
and
Him
el
and
Mahra
jHasa
n,
“
Autom
at
ic
Re
co
gnit
ion
of
Bang
l
a
Sign
La
nguag
e
Us
ing
Artifi
cial
Neura
l
Ne
tworks
(AN
N
S)
For
D
ea
f
And
Dum
b
t
o
Bridge
The
Co
m
m
unic
at
ion
Ga
p,
”
In
te
rnationa
l
Journal
of
R
ec
e
nt
Ad
vanc
es
in Mult
idi
sc
ipl
inar
y
R
ese
arch
,
vol
.
3,
no
.
7
,
pp
.
1649
-
1654,
2016
.
[19]
M.
T.
Hoque,
M.
Rifa
t
-
Ut
-
Ta
uw
ab,
M.
F.
Kabir
,
F.
Sarke
r,
M.
N.
Huda
and
K.
Abdulla
h
-
Al
-
Mam
un,
"A
utomate
d
Bangl
a
sign
language
tra
nsl
at
i
on
sy
st
em:
Pro
spec
ts,
li
m
itati
o
ns
and
appl
ic
ations,"
2016
5th
Inte
rnationa
l
Confe
renc
e
on
I
nfo
rm
ati
cs,
El
e
c
tronic
s and Visi
on
(
ICIEV
)
,
2016,
pp
.
856
-
862
,
doi:
10
.
1109/ICI
EV.
2016.
77601
23.
[20]
K.
B.
Shaik,
P.
Gane
san
,
V.
Kali
st
,
B.
S.
Sath
ish,
and
J.
M
.
M.
Jenit
h
a,
“
Com
par
at
ive
Stud
y
of
Skin
Co
lor
Dete
c
ti
on
and
Segm
ent
ation
in
HS
V
and
YCbC
r
Color
Spac
e,”
Proce
dia
Comp
ute
r
Scienc
e
,
vo
l.
57
,
pp.
41
-
4
8
,
2015,
doi
:
10
.
10
16/j
.
pro
cs.
2015.
07.
362.
[21]
V.
Bhame,
R
.
S
ree
m
at
h
y
and
H
.
Dhum
al
,
"V
isi
on
base
d
hand
gesture
r
ec
ogni
t
ion
using
e
cc
en
t
ric
app
roa
ch
fo
r
hum
an
computer
int
er
ac
t
ion,
"
2
014
Inte
rnation
al
Confe
renc
e
o
n
Adv
anc
es
in
Computing,
Co
mm
unic
ati
ons
and
Informatic
s (
ICACCI)
,
2014,
pp
.
949
-
953,
doi
:
10
.
1109/ICACCI.
2
014.
6968545.
[22]
G.
A.
Rao,
K.
Sy
ama
la
,
P.
V.
V
.
Kishore
and
A.
S.
C.
S.
Sastr
y
,
"D
ee
p
convol
u
ti
onal
n
eur
al
n
etw
orks
for
sign
la
nguag
e
re
cog
nit
ion,
"
2018
C
onfe
renc
e
on
Signal
Proce
s
sing
And
Com
municat
ion
Eng
ine
ering
S
yste
ms
(
SPA
CES)
,
2018,
pp
.
194
-
197
,
d
oi:
10
.
1109/SP
ACES.2018.
8316
344.
[23]
Jonatha
n
Long,
Eva
n
Shelh
amer
,
Tr
evor
Darr
el
l
,
“
Full
y
Convo
lu
ti
onal
Networks
for
Sem
ant
i
c
S
egmenta
t
ion,”
in
IEE
E
Tr
ansacti
o
n
s on
Pattern
An
aly
sis and
Ma
ch
ine
In
te
l
li
gen
ce
,
2015,
pp
.
3431
-
3440.
[24]
M.
J.
Hos
sein
and
M.
Sabbir
Ejaz
,
"Rec
ogni
ti
on
o
f
Benga
l
i
Sign Langua
ge
using N
ovel
Dee
p
Conv
olut
ional
Neura
l
Network,
"
2020
2nd
Inte
rnation
al
Confe
renc
e
o
n
Sustainabl
e
Technol
og
ie
s
for
Industry
4.
0
(
ST
I)
,
2020,
pp.
1
-
5,
doi:
10
.
1109/ST
I50764.
2020.
93
50418.
[25]
A.
R.
Chowdhu
r
y
,
A.
B
iswas,
S.
M.
F.
Hasan
,
T
.
M.
R
ahma
n
and
J.
Uddin
,
"Benga
l
i
Sign
la
nguag
e
to
te
x
t
conve
rsion
usin
g
art
ifici
al
neur
al
net
work
and
support
vec
tor
m
ac
hine
,
"
2017
3rd
Inte
rnational
Confe
renc
e
o
n
El
e
ct
rica
l
Infor
mation
and
Com
municat
ion
Tech
nology
(
EICT)
,
2017,
pp
.
1
-
4
,
doi
:
10.
1109
/E
ICT
.
2017.
8275248.
[26]
M.
El
Bada
w
y
,
A.
S.
El
ons,
H.
A.
Shedee
d
a
nd
M.
F.
Tol
ba,
"A
rab
ic
sign
la
nguage
re
co
gnit
ion
with
3
D
convol
uti
on
al
ne
ura
l
n
et
works
,
"
2017
Ei
ghth
Int
ernati
onal
Conf
ere
nce
on
In
te
l
ligent
Computing
and
Information
Syste
ms
(
ICICIS)
,
2017,
pp
.
66
-
7
1,
doi
:
10
.
1109/I
NTEL
CIS.2017.8260028.
[27]
M.
M.
Hasan,
A.
Y.
Srizon
,
A
.
Sa
y
e
ed
and
M.
A.
M.
Hasan,
"Classific
a
ti
on
of
Sign
La
ngu
a
ge
Char
ac
t
ers
b
y
Appl
y
ing
a
Dee
p
Convolut
ional
Neura
l
Networ
k,
"
2020
2nd
Inte
rnational
Conf
ere
nce
on
Ad
va
nce
d
Informatio
n
and
Comm
unic
a
ti
on
Te
chnol
ogy
(
ICAICT
)
,
2020,
pp.
434
-
438,
do
i:
10
.
1109/ICAI
CT51780.
2020.
9
333456.
[28]
M.
A.
Uddin
and
S.
A.
Chowdh
ur
y
,
"H
and
sign
la
nguage
re
cog
nit
ion
for
Bangla
al
phab
et
using
Support
Vec
tor
Mac
hine,"
2016
Inte
rnational
Co
nfe
renc
e
on
Inn
ovat
ions
in
Scie
nce
,
Engi
n
ee
rin
g
and
Technol
og
y
(
ICISET)
,
2016,
pp.
1
-
4
,
doi
:
10
.
1109/ICISE
T.
20
16.
7856479
.
Evaluation Warning : The document was created with Spire.PDF for Python.