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.
18
,
No.
1
,
F
e
br
ua
r
y
2020
,
pp.
407
~
418
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
.
v18i1.
13958
407
Jou
r
n
al
h
omepage
:
ht
tp:
//
jour
nal.
uad
.
ac
.
id/
index
.
php/T
E
L
K
OM
N
I
K
A
C
on
t
in
u
ou
s
s
p
e
e
c
h
se
gm
e
n
t
at
io
n
u
si
n
g
l
oc
al
ad
a
p
t
iv
e
t
h
r
e
sh
ol
d
in
g
t
e
c
h
n
iq
u
e
i
n
t
h
e
b
lo
c
k
i
n
g b
lo
c
k
a
r
e
a
m
e
t
h
o
d
Roih
an
Aul
iya
Ulf
at
t
ah
,
S
u
k
m
awat
i
Nur
E
n
d
ah
,
Re
t
n
o
Ku
s
u
m
an
i
n
gr
u
m
,
S
a
t
r
iyo
Adh
y
In
fo
rma
t
i
c
s
D
ep
ar
t
men
t
,
Facu
l
t
y
o
f
Sci
en
ce
an
d
Ma
t
h
e
mat
i
c
s
,
D
i
p
o
n
eg
o
ro
U
n
i
v
ers
i
t
y
,
In
d
o
n
e
s
i
a
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
Aug
23
,
2019
R
e
vis
e
d
Nov
2
0
,
20
19
Ac
c
e
pted
De
c
20
,
20
19
Co
n
t
i
n
u
o
u
s
s
p
eech
i
s
a
fo
rm
o
f
n
at
u
ral
h
u
ma
n
s
p
eec
h
t
h
at
i
s
co
n
t
i
n
u
o
u
s
w
i
t
h
o
u
t
a
cl
ear
b
o
u
n
d
ar
y
b
et
w
een
w
o
rd
s
.
In
co
n
t
i
n
u
o
u
s
s
p
eec
h
reco
g
n
i
t
i
o
n
,
a
s
eg
me
n
t
a
t
i
o
n
p
ro
ce
s
s
i
s
n
e
ed
e
d
t
o
c
u
t
t
h
e
s
en
t
en
c
e
at
t
h
e
b
o
u
n
d
ary
o
f
each
w
o
r
d
.
Seg
men
t
at
i
o
n
b
ec
o
mes
a
n
i
m
p
o
r
t
an
t
s
t
ep
b
eca
u
s
e
a
s
p
eech
ca
n
b
e
reco
g
n
i
ze
d
fro
m
t
h
e
w
o
r
d
s
e
g
men
t
s
p
r
o
d
u
ced
b
y
t
h
i
s
p
ro
ce
s
s
.
T
h
e
s
e
g
men
t
at
i
o
n
p
r
o
ces
s
i
n
t
h
i
s
s
t
u
d
y
w
a
s
carri
e
d
o
u
t
u
s
i
n
g
l
o
ca
l
ad
a
p
t
i
v
e
t
h
re
s
h
o
l
d
i
n
g
t
ec
h
n
i
q
u
e
i
n
t
h
e
b
l
o
c
k
i
n
g
b
l
o
ck
area
me
t
h
o
d
.
T
h
i
s
s
t
u
d
y
ai
ms
t
o
co
n
d
u
c
t
p
erfo
rma
n
ce
co
mp
ari
s
o
n
s
fo
r
fi
v
e
l
o
ca
l
ad
ap
t
i
v
e
t
h
res
h
o
l
d
i
n
g
met
h
o
d
s
(N
i
b
l
ac
k
,
Sau
v
o
l
a,
Brad
l
e
y
,
G
u
an
g
l
e
i
X
i
o
n
g
an
d
Bern
s
e
n
)
i
n
co
n
t
i
n
u
o
u
s
s
p
eec
h
s
eg
me
n
t
a
t
i
o
n
t
o
o
b
t
ai
n
t
h
e
b
es
t
met
h
o
d
an
d
o
p
t
i
m
u
m
p
aramet
er
v
a
l
u
e
s
.
Bas
ed
o
n
t
h
e
res
u
l
t
s
o
f
t
h
e
s
t
u
d
y
,
N
i
b
l
ack
met
h
o
d
i
s
co
n
c
l
u
d
ed
as
t
h
e
b
e
s
t
met
h
o
d
fo
r
co
n
t
i
n
u
o
u
s
s
p
ee
ch
s
eg
men
t
at
i
o
n
i
n
In
d
o
n
e
s
i
a
n
l
an
g
u
a
g
e
w
i
t
h
t
h
e
accu
r
acy
v
a
l
u
e
o
f
9
5
%
,
an
d
t
h
e
o
p
t
i
m
u
m
p
aramet
er
v
a
l
u
e
s
fo
r
s
u
ch
me
t
h
o
d
are
w
i
n
d
o
w
=
7
5
an
d
k
=
0
.
2
.
K
e
y
w
o
r
d
s
:
B
locking
b
lock
a
r
e
a
C
onti
nuous
s
pe
e
c
h
L
oc
a
l
a
da
pti
ve
t
hr
e
s
holdi
ng
S
pe
e
c
h
s
e
gmenta
ti
on
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
ukmaw
a
ti
Nur
E
nda
h,
I
nf
or
matics
De
pa
r
tm
e
nt
,
F
a
c
ult
y
of
S
c
ienc
e
a
nd
M
a
thema
ti
c
s
,
Dipone
gor
o
Unive
r
s
it
y,
P
r
of
S
oe
da
r
to
S
t
.
,
S
.
H.
Ka
mpus
T
e
mbala
ng
UN
DI
P
,
S
e
mar
a
ng
,
J
a
wa
T
e
nga
h,
I
ndone
s
ia
.
E
mail:
s
ukma
wa
ti
020578
@
gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
C
onti
nuous
s
pe
e
c
h
r
e
c
ognit
ion
is
a
f
ur
ther
d
e
ve
lopm
e
nt
of
is
olate
d
wor
ds
r
e
c
ognit
ion
that
r
e
c
ognize
s
wor
ds
f
r
om
a
s
e
ntenc
e
us
ing
a
mac
hi
ne
lea
r
ning
a
lgor
it
hm
[
1
]
.
Huma
n
s
pe
e
c
h
is
a
c
onti
nuous
s
pe
e
c
h,
a
s
e
r
ies
of
wor
ds
c
ompos
e
d
c
onti
nuous
ly
without
a
c
lea
r
br
e
a
k
be
twe
e
n
wor
ds
.
C
onti
n
uous
s
pe
e
c
h
r
e
c
ognit
ion
tec
hnology
is
ne
e
de
d
s
o
that
the
mac
hine
c
a
n
unde
r
s
tand
human
s
pe
e
c
h
in
givi
ng
voice
c
omm
a
nds
[
2]
.
S
pe
e
c
h
r
e
c
ognit
ion
ha
s
be
e
n
wide
l
y
a
ppli
e
d
in
va
r
ious
f
ields
[3
-
6]
.
T
he
im
p
leme
ntation
o
f
c
onti
nuous
s
pe
e
c
h
r
e
c
ognit
ion
c
ons
is
ts
of
thr
e
e
major
s
tage
s
:
pr
e
-
pr
oc
e
s
s
ing,
f
e
a
tur
e
e
x
tr
a
c
ti
on,
a
nd
r
e
c
ognit
ion
[
7
]
.
P
r
e
-
pr
oc
e
s
s
ing
f
unc
ti
ons
to
pr
e
pa
r
e
s
pe
e
c
h
s
ignals
s
o
that
f
e
a
tur
e
e
xtr
a
c
ti
on
c
a
n
be
pe
r
f
o
r
med.
One
o
f
the
main
pr
oc
e
s
s
in
p
r
e
-
pr
oc
e
s
s
ing
is
the
s
e
gmenta
ti
on
pr
oc
e
s
s
.
S
e
gmenta
ti
on
p
r
oc
e
s
s
is
a
p
r
oc
e
s
s
of
divi
d
ing
c
ont
inuous
s
pe
e
c
h
int
o
ba
s
ic
unit
s
s
uc
h
a
s
wor
ds
,
phon
e
mes
or
r
e
c
ogniza
ble
s
yll
a
bles
[
8]
.
T
he
lac
k
o
f
mar
ke
r
s
that
indi
c
a
te
the
ini
ti
a
l
a
nd
f
inal
li
m
it
s
of
a
wo
r
d
whe
n
s
pe
a
king
incr
e
a
s
ingl
y
c
ompl
ica
tes
the
pr
oc
e
s
s
o
f
s
e
gmenta
ti
on,
e
s
pe
c
ially
whe
n
s
pe
a
king
c
onti
nuous
ly.
I
n
c
ontr
a
s
t
to
text
that
c
a
n
be
s
e
e
n
or
g
iven
it
s
s
e
gment
bounda
r
ies
by
r
e
c
ognizing
the
s
pa
c
e
be
twe
e
n
wor
ds
.
T
he
r
e
s
ult
s
of
thi
s
s
e
gmenta
ti
on
will
indi
r
e
c
tl
y
a
f
f
e
c
t
the
r
e
s
ult
s
of
r
e
c
ognit
ion
[
9]
.
T
he
r
e
s
e
a
r
c
h
r
e
late
d
to
c
onti
nuous
s
pe
e
c
h
s
e
gmen
tation
ha
s
be
e
n
c
onduc
ted
us
ing
s
e
ve
r
a
l
methods
,
a
udio
a
nd
vis
ua
l
f
us
ion
f
o
r
the
domain
of
T
ur
kis
h
langua
ge
[
10]
,
s
e
gmenta
ti
on
ba
s
e
d
on
ti
m
e
-
domain
f
e
a
tur
e
s
a
nd
f
r
e
que
nc
y
-
domain
f
e
a
tur
e
s
a
ppli
e
d
in
B
a
ngla
[
8
]
,
Hybr
id
of
t
im
e
-
domain
f
e
a
t
ur
e
s
a
nd
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.
18
,
No
.
1
,
F
e
br
ua
r
y
2020
:
407
-
418
408
f
r
e
que
nc
y
-
domain
f
e
a
tur
e
s
a
nd
media
n
f
il
ter
ing
in
T
a
mi
l
[
11]
,
s
e
gmenta
ti
on
with
dyna
mi
c
th
r
e
s
holdi
ng
a
nd
blocking
block
a
r
e
a
ins
ide
B
a
ngla
[
12]
.
B
a
s
e
d
on
s
ome
of
thes
e
s
tudi
e
s
,
c
onti
nuous
s
pe
e
c
h
s
e
gment
a
ti
on
c
a
n
be
done
by
c
onve
r
ti
ng
s
pe
e
c
h
s
ignal
r
e
pr
e
s
e
ntatio
ns
int
o
s
pe
c
tr
ogr
a
m
im
a
ge
s
.
T
he
s
pe
c
tr
ogr
a
m
i
mage
is
then
pr
oc
e
s
s
e
d
to
pr
oduc
e
wor
d
s
e
gments
,
one
method
t
ha
t
c
a
n
be
us
e
d
is
the
blocking
block
a
r
e
a
[
9]
.
B
locking
block
a
r
e
a
is
the
p
r
oc
e
s
s
of
making
wo
r
d
blocks
f
r
om
s
pe
c
tr
ogr
a
m
im
a
ge
s
in
the
f
o
r
m
o
f
binar
y
im
a
ge
s
thr
ough
s
e
ve
r
a
l
s
tage
s
,
na
mely
ge
n
e
r
a
ti
ng
s
pe
c
togr
a
ms
,
pe
r
f
or
mi
ng
dyna
mi
c
tr
e
s
hold
ing
with
c
lus
ter
ing
a
lgor
it
hms
on
s
pe
c
togr
a
m
im
a
ge
s
to
pr
oduc
e
binar
y
im
a
ge
s
a
nd
bou
nda
r
y
de
tec
ti
on.
I
n
thi
s
r
e
s
e
a
r
c
h
,
on
wor
d
s
e
gmenta
ti
on
us
ing
dyna
mi
c
thr
e
s
holdi
ng
in
the
blocking
block
a
r
e
a
meth
od
with
the
a
ddit
ion
of
mor
phologi
c
a
l
ope
r
a
ti
ons
a
nd
ove
r
lapping
pr
oc
e
s
s
whic
h
is
then
c
a
ll
e
d
im
pr
ove
d
blocking
block
a
r
e
a.
T
his
is
done
be
c
a
us
e
it
wil
l
be
a
ppli
e
d
to
the
s
pe
e
c
h
s
e
gment
in
I
ndone
s
ian,
be
c
a
us
e
the
blocking
block
a
r
e
a
method
in
the
s
tudy
[
12
]
ha
s
a
ba
d
r
e
s
ult
if
a
ppli
e
d
in
the
I
ndone
s
ian
l
a
ngua
ge
domain.
T
his
may
be
due
to
I
ndone
s
ian
langua
ge
whic
h
ha
s
many
r
e
giona
l
diale
c
ts
,
s
o
that
a
wor
d
c
a
n
ha
ve
a
dif
f
e
r
e
nt
pa
tt
e
r
n.
Dyna
mi
c
th
r
e
s
holdi
ng
in
r
e
s
e
a
r
c
h
[
12]
us
e
s
a
s
ingl
e
thr
e
s
hold
f
or
the
e
nti
r
e
im
a
ge
or
g
lobal
thr
e
s
hold.
S
ing
le
thr
e
s
hold
in
s
uc
h
tec
hnique
will
be
di
f
f
icult
to
dis
ti
nguis
h
the
ba
c
kgr
o
und
a
nd
the
f
or
e
g
r
ound
f
ields
in
the
s
pe
c
tr
ogr
a
m
i
mage
s
with
mor
e
than
two
r
e
gions
due
to
va
r
ying
int
e
ns
it
ies
a
nd
nois
e
s
in
the
im
a
ge
s
[
13]
.
I
n
s
uc
h
c
ondit
ion
,
s
om
e
thr
e
s
hold
va
lues
a
r
e
ne
e
de
d
f
o
r
e
a
c
h
pix
e
l
in
a
pa
r
ti
c
ular
r
e
gion
us
ing
loca
l
a
da
pti
ve
th
r
e
s
holdi
ng
tec
hnique.
I
n
thi
s
s
tudy,
c
onti
nuous
s
pe
e
c
h
s
e
gmenta
ti
on
is
pe
r
f
or
med
us
ing
loca
l
a
da
pti
ve
thr
e
s
holdi
ng
tec
hnique
to
pr
oduc
e
binar
y
s
pe
c
tr
ogr
a
m
im
a
ge
s
.
T
his
thr
e
s
holdi
ng
tec
hnique
ha
s
be
e
n
a
ppli
e
d
to
binar
iza
ti
on
a
nd
im
a
ge
s
e
gmenta
ti
on
pr
oc
e
s
s
e
s
in
s
e
ve
r
a
l
pr
e
vious
s
tudi
e
s
[
14
-
17]
.
T
he
binar
y
im
a
g
e
,
r
e
s
ult
s
of
the
binar
iza
ti
on
is
then
pr
oc
e
s
s
e
d
us
ing
i
mpr
ov
e
d
blocking
block
a
r
e
a
method
s
o
that
ther
e
will
be
wor
d
blocks
ba
s
e
d
on
the
number
of
pixels
f
or
e
a
c
h
c
olum
n.
E
a
c
h
blo
c
k
is
a
wor
d
s
e
gment
that
r
e
s
ult
s
f
r
om
the
s
e
gmenta
ti
on
pr
oc
e
s
s
.
T
he
r
e
a
r
e
s
e
ve
r
a
l
loca
l
a
da
pti
ve
thr
e
s
holdi
ng
methods
including
Nibla
c
k
[
18]
,
S
a
uvola
[
19
]
,
B
r
a
dley
[
20
]
,
Gua
nglei
Xiong
[
21]
a
nd
B
e
r
ns
e
n
[
22
]
.
T
he
pe
r
f
or
manc
e
of
e
a
c
h
method
will
be
c
ompar
e
d
in
thi
s
s
tudy
f
or
c
onti
nuous
s
pe
e
c
h
s
e
gmenta
ti
on.
2.
RE
S
E
AR
CH
M
E
T
HO
D
T
his
r
e
s
e
a
r
c
h
is
divi
de
d
int
o
thr
e
e
s
t
e
ps
,
na
mely
da
ta
c
oll
e
c
ti
on,
s
e
gmenta
ti
on
a
nd
tes
ti
ng.
T
he
f
oll
owing
is
a
n
e
xplana
ti
on
of
e
a
c
h
s
tep.
2.
1.
Dat
a
c
oll
e
c
t
ion
T
he
da
ta
wa
s
take
n
by
r
e
c
or
ding
f
our
pe
ople
who
ha
ve
dif
f
e
r
e
nt
d
iale
c
ts
in
I
ndone
s
ia.
E
a
c
h
pe
r
s
on
s
a
ys
20
s
e
ntenc
e
s
,
that
is
:
S1
abang
be
r
c
e
r
it
a
s
e
s
uatu
y
ang
bagus
S2
bapak
ibu
pe
r
gi
be
r
s
ama
adik
S3
bibi
mulai
ter
k
e
nal
s
or
e
ini
S4
c
incin
k
aw
in
dar
i
bahan
pe
r
mata
S5
dia
puny
a
dua
mobi
l
hit
am
S6
hidup
it
u
s
e
pe
r
ti
s
e
k
otak
c
ok
lat
S7
k
amu
jangan
jadi
judes
juga
S8
k
apan
k
it
a
main
bola
pa
ntai
S9
k
ar
e
na
k
e
ju
adalah
s
us
u
s
api
S
10
k
ompor
k
r
e
dit
be
r
w
a
r
na
me
r
ah
muda
S
11
maaf
atas
k
e
jadi
an
s
e
nin
lal
u
S
12
mak
an
k
uning
telur
s
e
tengah
matang
S
13
mas
ini
s
k
e
r
e
ta
be
r
baju
bir
u
tua
S
14
nanti
s
iang
s
aja
k
ata
be
r
bahay
a
S
15
pabr
ik
gula
pas
ir
ada
li
ma
S
16
paman
me
ninggal
s
aat
dulu
s
e
k
ali
S
17
pantun
tentang
pis
ang
dan
s
ay
ur
S
18
s
iapa
s
uk
a
anak
k
e
c
il
lucu
S
19
s
ulap
tepung
ter
igu
r
as
a
r
oti
S
20
tukang
ti
pu
s
udah
ter
tangkap
juga
2.
2.
S
e
gm
e
n
t
at
ion
T
he
r
e
a
r
e
f
ive
p
r
oc
e
s
s
e
s
in
c
onti
nuous
s
pe
e
c
h
s
e
gmenta
ti
on,
na
mely
G
e
ne
r
a
t
e
S
pe
c
tr
ogr
a
ms
,
B
inar
iza
ti
on,
M
or
phologi
c
a
l
Ope
r
a
ti
ons
,
I
mpr
ove
d
B
locking
B
lock
Ar
e
a
s
a
nd
B
ounda
r
y
D
e
tec
ti
on
a
s
s
hown
in
F
igur
e
1.
2.
2.
1.
Gener
at
e
s
p
e
c
t
r
ogr
am
Ge
ne
r
a
te
S
pe
c
tr
ogr
a
m
c
onve
r
ts
s
ound
s
ignal
s
i
nto
im
a
ge
s
of
s
ound
s
ignal
int
e
ns
it
y
that
ha
ve
dif
f
e
r
e
nt
de
ns
it
ies
.
T
he
s
pe
c
togr
a
m
f
unc
ti
ons
to
identif
y
a
nd
g
r
oup
the
s
ound
inpu
t
in
phone
mi
c
wa
y
.
T
he
im
a
ge
o
f
the
s
pe
c
tr
ogr
a
m
is
then
c
onve
r
ted
i
nto
a
gr
a
ys
c
a
le
im
a
ge
t
o
be
a
ble
to
do
b
inar
iza
ti
on
us
ing
l
oc
a
l
a
da
pti
ve
t
hr
e
s
holdi
ng.
T
he
im
a
ge
of
the
s
pe
c
tr
ogr
a
m
f
r
om
the
s
pe
e
c
h
s
ignal
of
"
dia
puny
a
d
ua
mobil
hit
am
"
is
s
hown
in
F
igu
r
e
2.
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
C
onti
nuous
s
pe
e
c
h
s
e
gm
e
ntat
ion
us
ing
local
a
dapt
ive
thr
e
s
holdi
ng
tec
hniqu
...
(
R
oihan
A
uli
y
a
Ulfat
ta
h)
409
F
igur
e
1.
P
r
oc
e
s
s
b
lock
of
s
e
gmenta
ti
on
F
igur
e
2.
S
pe
e
c
h
s
ignal
a
nd
s
pe
c
tr
ogr
a
m
i
mage
o
f
“
dia
puny
a
dua
mobil
hit
am
”
2.
2.
2.
B
in
ar
izat
ion
u
s
in
g
local
ad
ap
t
ive
t
h
r
e
s
h
old
in
g
T
he
binar
y
i
mage
of
s
pe
c
tr
ogr
a
m
is
obtaine
d
thr
ough
a
B
inar
iza
ti
on
p
r
oc
e
s
s
us
ing
loca
l
a
da
pti
ve
thr
e
s
holdi
ng
tec
hnique.
T
h
is
tec
hnique
will
pr
o
duc
e
a
th
r
e
s
hold
va
lue
us
e
d
to
gr
oup
the
int
e
ns
it
y
of
the
input
im
a
ge
int
o
two
va
lues
(
ba
c
kgr
ound
o
r
f
o
r
e
gr
ound)
.
a.
Niblac
k
Niblac
k
de
ter
mi
ne
s
the
thr
e
s
hold
va
lue
ba
s
e
d
on
t
he
loca
l
mea
n
a
nd
loca
l
s
tanda
r
d
de
viation.
B
oth
a
r
e
c
a
lcula
ted
in
a
window
with
the
s
ize
of
m
x
n
b
a
s
e
d
on
the
ne
ighbor
hood
va
lue
of
the
pixels
,
s
o
that
e
a
c
h
pixel
ha
s
a
dif
f
e
r
e
nt
th
r
e
s
hold
va
lue.
T
he
f
o
r
mul
a
t
o
c
a
lcula
te
the
thr
e
s
hold
va
lue
is
[
18]
:
(
,
)
=
(
,
)
+
·
σ
(
,
)
(
1)
whe
r
e
:
k
:
a
c
ons
tant
that
ha
s
a
va
lue
be
twe
e
n
0
a
nd
1
(
,
)
:
the
loca
l
mea
n
of
the
pixel
in
the
loca
l
window
(
,
)
:
the
loca
l
s
tanda
r
d
de
viation
of
the
pixel
in
the
loc
a
l
window
T
he
pr
oc
e
s
s
of
obtaining
a
binar
y
im
a
ge
us
ing
Nib
lac
k
is
s
hown
in
F
igu
r
e
3.
b.
S
a
uvola
S
a
uvola
de
ter
mi
ne
s
the
thr
e
s
hold
va
lue
ba
s
e
d
on
the
loca
l
mea
n
a
nd
the
loca
l
s
tanda
r
d
de
viation,
the
s
a
me
a
s
Niblac
k,
be
c
a
us
e
it
is
a
de
ve
lopm
e
n
t
of
Niblac
k
method.
T
he
di
f
f
e
r
e
nc
e
is
that
ther
e
is
a
n
R
va
lue
in
the
S
a
uvola
f
o
r
mul
a
.
R
is
the
dyna
mi
c
r
a
nge
of
the
s
tanda
r
d
de
viation
or
the
maximum
va
lue
of
the
s
tanda
r
d
de
viation
obtaine
d.
T
he
f
or
mul
a
to
c
a
l
c
ulate
S
a
uvola
thr
e
s
hold
va
lue
is
[
19]
:
(
,
)
=
(
,
)
∗
[
1
+
(
(
,
)
−
1
)
]
(
2)
whe
r
e
:
=
the
mea
n
f
o
r
a
ll
windows
=
the
s
tanda
r
d
de
viation
f
or
a
ll
windows
k
=
a
c
ons
tant
(0
-
1)
R
=
the
dyna
mi
c
r
a
nge
of
the
s
tanda
r
d
de
viation
T
he
pr
oc
e
s
s
of
obtaining
a
binar
y
im
a
ge
us
ing
S
a
uvola
is
s
hown
in
F
igu
r
e
4
.
c.
B
r
a
dley
B
r
a
dley
de
ter
mi
ne
s
the
th
r
e
s
hold
va
lue
ba
s
e
d
on
t
he
loca
l
mea
n
a
nd
the
a
ve
r
a
ge
va
lue
o
f
b
r
ight
ne
s
s
.
T
he
loca
l
mea
n
is
a
ls
o
c
a
lcula
ted
in
a
window
w
it
h
the
s
ize
o
f
m
x
n
ba
s
e
d
on
the
ne
ighbo
r
hood
va
lue
of
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.
18
,
No
.
1
,
F
e
br
ua
r
y
2020
:
407
-
418
410
the
pixels
,
s
o
that
e
a
c
h
pixel
ha
s
a
dif
f
e
r
e
nt
thr
e
s
hold
va
lue.
M
e
a
nwhile,
the
a
ve
r
a
ge
va
lue
of
br
ight
ne
s
s
de
pe
nds
on
the
c
ons
tant
va
lue
T
(
in
the
r
a
nge
o
f
1
-
100)
.
T
he
th
r
e
s
hold
va
lue
of
B
r
a
dley
metho
d
c
a
n
be
c
a
lcula
ted
a
s
f
oll
ows
[
20]
:
thr
e
s
hold
=
∗
(
1
−
100
)
(
3)
t
he
pr
oc
e
s
s
of
obtaining
a
binar
y
im
a
ge
us
ing
B
r
a
d
ley
is
s
hown
in
F
igu
r
e
5
.
d.
Gua
nglei
Xiong
Gua
nglei
Xiong
de
ter
mi
ne
s
the
th
r
e
s
hold
va
lue
b
a
s
e
d
on
the
loca
l
mea
n
or
the
loca
l
media
n
,
a
nd
de
pe
nds
on
a
c
e
r
tain
c
ons
tant
va
lue
in
the
r
a
nge
of
0
-
255.
I
n
thi
s
s
tudy,
the
thr
e
s
hold
va
lue
is
de
t
e
r
mi
ne
d
us
ing
loca
l
mea
n,
be
c
a
us
e
the
a
ve
r
a
ge
va
lue
of
ne
ighbor
hood
f
or
a
pixel
with
the
s
ize
of
m
x
n
is
m
or
e
a
ble
to
r
e
pr
e
s
e
nt
the
va
lue
o
f
the
pixel
than
us
ing
the
lo
c
a
l
media
n.
T
he
t
hr
e
s
hold
va
lue
of
Gua
nglei
Xion
g
c
a
n
be
c
a
lcula
ted
a
s
f
oll
ows
[
21]
:
T
=
m
e
a
n
−
C
a
tau
T
=
m
e
d
ia
n
−
C
(
4)
T
he
pr
oc
e
s
s
of
obtaining
a
binar
y
im
a
ge
us
ing
Gua
nglei
Xiong
is
s
hown
in
F
igur
e
6
.
e.
B
e
r
ns
e
n
B
e
r
ns
e
n
de
ter
mi
ne
s
the
thr
e
s
hold
va
lue
ba
s
e
d
on
the
loca
l
mea
n,
the
loca
l
c
ont
r
a
s
t
,
a
nd
the
th
r
e
s
hold
va
lue
of
c
ontr
a
s
t
.
L
oc
a
l
c
ontr
a
s
t
i
s
the
ini
ti
a
l
de
ter
mi
na
nt
o
f
the
th
r
e
s
hold
f
o
r
two
c
ondit
ions
,
th
a
t
is
,
if
the
loca
l
c
ontr
a
s
t
va
lue
is
les
s
than
the
thr
e
s
hold
v
a
lue
of
c
ontr
a
s
t
k,
the
pixel
wil
l
be
s
e
t
to
the
ba
c
kgr
ound
or
the
f
or
e
gr
ound
de
pe
nd
ing
on
the
global
m
idgr
e
y
va
lue.
W
he
r
e
a
s
if
the
loca
l
c
ontr
a
s
t
va
lue
>
=
the
t
hr
e
s
hold
va
lue
of
c
ontr
a
s
t
,
it
will
be
s
e
t
to
the
ba
c
kgr
oun
d
or
the
f
o
r
e
gr
ound
de
pe
nding
on
the
loca
l
mea
n
va
lue
.
T
he
t
hr
e
s
hold
va
lue
o
f
B
e
r
ns
e
n
method
is
c
a
lcula
ted
a
s
f
oll
ows
[
22]
:
(
,
)
=
0
.
5
(
m
a
x
(
,
)
+
m
i
n
(
,
)
)
(
5)
a
nd
the
pr
ovided
c
ont
r
a
s
t
is
c
a
lcula
ted
a
s
f
oll
ows
:
(
,
)
=
(
,
)
−
m
i
n
(
,
)
≥
k
(
6)
whe
r
e
:
m
a
x
(
,
)
=
T
he
m
a
xim
um
gr
a
y
va
lue
in
the
loca
l
window
m
i
n
(
,
)
=
T
he
mi
ni
mum
gr
a
y
va
lue
in
the
loca
l
window
=
T
he
th
r
e
s
hold
va
lue
of
c
ontr
a
s
t
T
he
pr
oc
e
s
s
of
obtaining
a
binar
y
im
a
ge
us
ing
B
e
r
ns
e
n
is
s
hown
in
F
igur
e
7
.
2.
2.
3.
M
or
p
h
ologi
c
al
o
p
e
r
at
ion
s
M
or
phologi
c
a
l
ope
r
a
ti
ons
a
r
e
pe
r
f
or
med
to
r
e
c
ons
tr
uc
t
a
nd
e
li
mi
na
te
im
pe
r
f
e
c
ti
ons
in
the
im
a
ge
s
tr
uc
tur
e
the
b
inar
y
i
mage
[
23
,
24
]
that
i
mpr
ove
r
e
s
ult
s
f
r
om
the
s
e
gmenta
ti
on
pr
oc
e
s
s
to
make
it
mor
e
sm
ooth.
a.
E
r
os
ion
E
r
os
ion
a
im
s
to
r
e
duc
e
the
e
dge
of
the
objec
t.
T
his
pr
oc
e
s
s
matc
he
s
whe
the
r
ther
e
a
r
e
a
ny
objec
ts
(
im
a
ge
pixels
)
of
the
f
or
e
gr
ound
that
c
ome
int
o
c
ontac
t
with
the
ba
c
kgr
ound
,
if
a
ny,
the
f
or
e
gr
ound
va
lue
that
make
s
a
c
ontac
t
is
c
ha
nge
d
a
c
c
or
d
ing
to
the
ba
c
kgr
ound
va
lue
[
25
]
.
b.
Dilatio
n
Dilation
is
the
oppos
it
e
of
e
r
os
ion
with
the
s
a
me
c
onc
e
pt
[
25]
.
T
his
pr
oc
e
s
s
matc
he
s
whe
ther
ther
e
is
a
pa
r
t
o
f
the
e
leme
nt
s
tr
uc
tu
r
e
that
c
omes
i
nto
c
ontac
t
with
the
ba
c
kgr
ound
whe
n
the
c
e
nter
of
the
e
leme
nt
is
f
o
r
e
gr
ound.
I
f
ther
e
is
,
the
ba
c
kgr
ound
va
lue
matc
he
s
the
f
o
r
e
gr
ound
va
lue.
2.
2.
4.
I
m
ag
e
b
locki
n
g
u
s
in
g
im
p
r
ove
d
b
locki
n
g
b
lock
ar
e
a
T
he
im
p
r
ove
d
b
locking
block
a
r
e
a
method
a
im
s
t
o
c
ha
nge
the
b
inar
y
im
a
ge
of
a
s
pe
c
tr
ogr
a
m
that
ha
s
gone
th
r
ough
mor
phologi
c
a
l
ope
r
a
ti
ons
int
o
a
block
im
a
ge
by
a
pplyi
ng
the
c
onc
e
pt
of
ove
r
lapping
c
olum
ns
.
T
he
method
wor
ks
by
b
r
e
a
king
the
im
a
g
e
int
o
s
e
ve
r
a
l
f
r
a
mes
then
c
a
lcula
ti
ng
the
lum
inan
c
e
va
lue
0
a
nd
the
lum
inanc
e
va
lue
1
f
or
e
a
c
h
f
r
a
me
.
I
f
the
lum
inanc
e
va
lue
0
r
e
a
c
he
s
45%
or
mor
e
than
the
number
of
pixels
in
the
f
r
a
me,
the
f
r
a
me
will
be
c
olor
e
d
i
n
blac
k.
W
he
r
e
a
s
if
the
lum
inanc
e
va
lue
1
r
e
a
c
he
s
55%
or
mor
e
,
the
f
r
a
me
will
be
c
olor
e
d
in
whi
te.
F
u
r
ther
mor
e
,
a
ll
of
the
pixel
va
lues
in
the
f
r
a
me
will
be
c
ha
nge
d
to
white
if
the
f
r
a
me
is
white
a
nd
the
ne
xt
f
r
a
me
is
white.
As
ide
f
r
om
that,
a
ll
o
f
the
pixel
va
lue
s
in
t
he
f
r
a
me
a
r
e
c
ha
nge
d
to
blac
k.
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
C
onti
nuous
s
pe
e
c
h
s
e
gm
e
ntat
ion
us
ing
local
a
dapt
ive
thr
e
s
holdi
ng
tec
hniqu
...
(
R
oihan
A
uli
y
a
Ulfat
ta
h)
411
F
igur
e
3.
F
lowc
ha
r
t
to
o
btain
b
inar
y
im
a
ge
u
s
ing
n
ibl
a
c
k
F
igur
e
4.
F
lowc
ha
r
t
to
o
btain
b
inar
y
i
mage
u
s
ing
s
a
uvola
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.
18
,
No
.
1
,
F
e
br
ua
r
y
2020
:
407
-
418
412
F
igur
e
5.
F
lowc
ha
r
t
to
o
btain
b
inar
y
i
mage
u
s
ing
b
r
a
dley
F
igur
e
6.
F
lowc
ha
r
t
to
o
btain
b
inar
y
i
mage
u
s
ing
g
ua
nglei
x
iong
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
C
onti
nuous
s
pe
e
c
h
s
e
gm
e
ntat
ion
us
ing
local
a
dapt
ive
thr
e
s
holdi
ng
tec
hniqu
...
(
R
oihan
A
uli
y
a
Ulfat
ta
h)
413
F
igur
e
7.
F
lowc
ha
r
t
to
o
btain
b
inar
y
i
mage
u
s
ing
b
e
r
ns
e
n
2.
2.
5.
B
ou
n
d
ar
y
d
e
t
e
c
t
ion
an
d
wor
d
s
e
gm
e
n
t
at
io
n
T
he
block
i
mage
that
ha
s
be
e
n
obtaine
d
is
then
pr
oc
e
s
s
e
d
to
de
ter
mi
ne
the
ini
ti
a
l
a
nd
the
f
inal
bounda
r
ies
f
or
e
a
c
h
block
.
T
he
c
oo
r
dinate
s
of
the
i
nit
ial
a
nd
the
f
inal
bounda
r
ies
of
e
a
c
h
block
a
r
e
c
a
lcula
ted
a
c
c
or
ding
to
the
ove
r
a
ll
va
lue
of
the
block
im
a
g
e
c
olum
n.
F
u
r
ther
mor
e
,
the
r
e
s
ult
s
of
thes
e
pe
r
c
e
nt
a
ge
s
a
r
e
us
e
d
a
s
a
guidanc
e
in
the
pr
oc
e
s
s
of
c
utt
ing
the
voi
c
e
[
12]
.
2.
3.
T
e
s
t
in
g
S
e
gmenta
ti
on
tes
ti
ng
is
c
onduc
ted
by
us
ing
f
ive
loca
l
a
da
pti
ve
thr
e
s
holdi
ng
methods
a
nd
im
pr
ove
d
blocking
block
a
r
e
a
method
wi
th
a
c
ombi
na
ti
on
of
s
e
ve
r
a
l
pa
r
a
mete
r
s
.
I
t
ha
s
s
ix
s
c
e
na
r
ios
a
s
f
oll
ows
.
−
S
c
e
na
r
io
1,
the
s
e
gmenta
ti
on
us
e
s
Niblac
k
method
with
incr
e
ment
p
r
oc
e
s
s
whe
r
e
the
window
v
a
lue
incr
e
a
s
e
by
15
s
tar
ti
ng
f
r
om
window
s
ize
=
15
to
window
s
ize
=
165.
−
S
c
e
na
r
io
2
,
the
s
e
gmenta
ti
on
us
e
s
S
a
uvola
metho
d
with
a
incr
e
ment
pr
oc
e
s
s
whe
r
e
the
window
v
a
lue
incr
e
a
s
e
by
15
s
tar
ti
ng
f
r
om
window
s
ize
=
15
to
window
s
ize
=
165
.
−
S
c
e
na
r
io
3
,
the
s
e
gmenta
ti
on
us
e
s
B
r
a
dley
metho
d
with
a
incr
e
ment
pr
oc
e
s
s
whe
r
e
the
window
v
a
lue
incr
e
a
s
e
by
15
s
tar
ti
ng
f
r
om
window
s
ize
=
15
to
window
s
ize
=
165.
−
S
c
e
na
r
io
4
,
the
s
e
gmenta
ti
on
us
e
s
Gua
nglei
Xion
g
method
with
incr
e
ment
pr
oc
e
s
s
whe
r
e
the
win
dow
va
lue
incr
e
a
s
e
by
15
s
tar
ti
ng
f
r
om
window
s
ize
=
15
to
window
s
ize
=
165
.
−
S
c
e
na
r
io
5
,
the
s
e
gmenta
ti
on
us
e
s
B
e
r
ns
e
n
meth
od
with
incr
e
ment
pr
oc
e
s
s
whe
r
e
the
window
v
a
lue
incr
e
a
s
e
by
15
s
tar
ti
ng
f
r
om
window
s
ize
=
15
to
wind
ow
s
ize
=
165
.
−
S
c
e
na
r
io
6
,
c
ompar
ing
the
f
ive
s
c
e
na
r
ios
a
bove
.
S
e
gmenta
ti
on
tes
ti
ng
us
e
s
80
s
e
ntenc
e
s
in
the
f
o
r
m
o
f
c
onti
nuous
s
pe
e
c
h
whic
h
c
ons
is
ts
of
20
s
e
ntenc
e
s
s
poke
n
by
4
di
f
f
e
r
e
nt
pe
ople
(
O1
,
O2
,
O3
,
O4
)
.
T
he
n
c
a
lcula
te
the
a
c
c
ur
a
c
y
in
s
e
gmenting
wor
d
c
or
r
e
c
tl
y
with
e
qua
ti
on
a
s
be
low.
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.
18
,
No
.
1
,
F
e
br
ua
r
y
2020
:
407
-
418
414
=
∑
ℎ
∑
ℎ
∗
100%
(
7)
=
∑
ℎ
ℎ
ℎ
∑
ℎ
∗
100%
(
8)
3.
RE
S
UL
T
S
A
ND
AN
AL
YSI
S
3.
1.
T
e
s
t
in
g
r
e
s
u
lt
F
igur
e
8
s
hows
the
a
c
c
ur
a
c
y
of
s
e
gmenta
ti
on
r
e
s
ul
ts
us
ing
Niblac
k
with
the
highes
t
a
c
c
ur
a
c
y
of
99
%
f
or
wor
d
a
c
c
ur
a
c
y
a
nd
95%
f
or
s
e
ntenc
e
s
e
gmenta
ti
on
a
c
hieve
d
whe
n
the
window
s
ize
is
75x75.
F
igur
e
9
s
hows
the
a
c
c
ur
a
c
y
of
s
e
gmenta
ti
on
r
e
s
ult
s
us
i
ng
S
a
uvola
with
the
highes
t
a
c
c
ur
a
c
y
of
97%
f
or
wor
d
a
c
c
ur
a
c
y
a
nd
86%
f
or
s
e
ntenc
e
s
e
gmenta
ti
on
a
c
hieve
d
whe
n
the
window
s
ize
is
90x90
.
F
ig
ur
e
10
s
hows
the
a
c
c
ur
a
c
y
of
s
e
gmenta
ti
on
r
e
s
ult
s
us
ing
B
r
a
dle
y
with
the
highes
t
a
c
c
ur
a
c
y
of
88%
f
o
r
wor
d
a
c
c
u
r
a
c
y
a
nd
58%
f
or
s
e
ntenc
e
s
e
gmenta
ti
on
a
c
hieve
d
whe
n
th
e
window
s
ize
is
90
x
90.
F
igur
e
11
s
hows
the
a
c
c
ur
a
c
y
of
s
e
gmenta
ti
on
r
e
s
ult
s
us
ing
Gua
ng
lei
Xiong
with
th
e
highes
t
a
c
c
ur
a
c
y
of
89%
f
or
wor
d
a
c
c
ur
a
c
y
a
nd
60%
f
or
s
e
ntenc
e
s
e
gm
e
ntation
a
c
hieve
d
whe
n
the
window
s
ize
is
90x90.
F
igur
e
12
s
hows
the
a
c
c
ur
a
c
y
of
s
e
gmenta
ti
on
r
e
s
ult
s
us
ing
B
e
r
ns
e
n
with
the
hi
ghe
s
t
a
c
c
ur
a
c
y
of
97%
f
o
r
wor
d
a
c
c
ur
a
c
y
a
nd
86%
f
or
s
e
ntenc
e
s
e
gmenta
ti
on
a
c
hieve
d
whe
n
the
window
s
ize
is
75x75
a
nd
90
x
90.
F
igur
e
8.
Gr
a
ph
of
s
e
gmenta
ti
on
r
e
s
ult
s
f
or
s
c
e
na
r
i
o
1
F
igur
e
9.
Gr
a
ph
of
s
e
gmenta
ti
on
r
e
s
ult
s
f
or
s
c
e
na
r
i
o
2
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
C
onti
nuous
s
pe
e
c
h
s
e
gm
e
ntat
ion
us
ing
local
a
dapt
ive
thr
e
s
holdi
ng
tec
hniqu
...
(
R
oihan
A
uli
y
a
Ulfat
ta
h)
415
F
igur
e
10
.
Gr
a
ph
of
s
e
gmenta
ti
on
r
e
s
ult
s
f
or
s
c
e
na
r
io
3
F
igur
e
11
.
Gr
a
ph
of
s
e
gmenta
ti
on
r
e
s
ult
s
f
or
s
c
e
na
r
io
4
F
igur
e
12
.
Gr
a
ph
of
s
e
gmenta
ti
on
r
e
s
ult
s
f
or
s
c
e
na
r
io
5
A
c
ompar
is
on
of
the
highes
t
r
e
s
ult
s
of
thes
e
f
iv
e
s
c
e
na
r
ios
c
a
n
be
s
e
e
n
in
F
igu
r
e
13
.
F
r
om
thi
s
f
igur
e
,
it
c
a
n
be
s
e
e
n
that
the
highes
t
a
c
c
ur
a
c
y
of
e
a
c
h
method
is
a
c
hieve
d
whe
n
the
window
s
ize
is
7
5x75
or
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.
18
,
No
.
1
,
F
e
br
ua
r
y
2020
:
407
-
418
416
90x90.
Above
a
nd
be
low
thi
s
window
s
ize
,
a
c
c
ur
a
c
y
d
e
c
r
e
a
s
e
s
.
T
his
c
a
n
s
how
that
the
opti
mal
a
c
c
ur
a
c
y
is
in
the
r
a
nge
60
-
90
or
90
-
105
.
F
igur
e
13
.
Gr
a
ph
of
s
e
gmenta
ti
on
r
e
s
ult
s
f
or
s
c
e
na
r
io
6
3.
2.
Anal
ys
is
of
s
e
gm
e
n
t
at
ion
r
e
s
u
lt
T
he
be
s
t
r
e
s
ult
is
in
f
luenc
e
d
by
the
pa
r
a
mete
r
va
lues
f
o
r
Niblac
k
method
a
s
c
a
n
be
s
e
e
n
in
F
igur
e
14
a
nd
F
igur
e
15.
F
igur
e
14.
R
e
s
ult
of
b
inar
iza
ti
on
(
l
e
f
t)
a
nd
r
e
s
ult
o
f
im
pr
ove
d
b
locking
b
lock
a
r
e
a
(
r
ight
)
u
s
ing
nibl
a
c
k
m
e
thod
with
p
a
r
a
mete
r
v
a
lues
(
window
=
15
a
nd
k
=
0.
2
)
T
he
two
im
a
ge
s
a
bove
s
how
a
c
ompar
is
on
of
the
s
e
gmenta
ti
on
r
e
s
ult
s
with
dif
f
e
r
e
nt
window
pa
r
a
mete
r
s
that
a
f
f
e
c
t
the
s
e
gmenta
ti
on
r
e
s
ult
s
.
T
he
s
e
gmenta
ti
on
r
e
s
ult
s
with
window
=
75
c
a
n
pr
oduc
e
be
tt
e
r
s
pe
c
tr
ogr
a
m
binar
y
im
a
ge
s
than
us
ing
w
indow
=
15,
the
qua
li
ty
o
f
the
b
locking
r
e
s
ult
s
of
the
s
pe
c
tr
ogr
a
m
im
a
ge
is
a
ls
o
indi
r
e
c
tl
y
inf
luenc
e
d
by
the
qua
li
ty
of
the
bina
r
y
im
a
ge
.
F
igur
e
15.
R
e
s
ult
of
b
inar
iza
ti
on
(
l
e
f
t)
a
nd
r
e
s
ult
o
f
im
pr
ove
d
blocking
b
lock
a
r
e
a
(
r
igh
t)
u
s
ing
nibl
a
c
k
method
with
pa
r
a
mete
r
v
a
lues
(
window
=
7
5
a
nd
k
=
0.
2
)
B
a
s
e
d
on
F
igur
e
8
unti
l
F
igu
r
e
12
,
it
is
s
hown
that
the
window
s
ize
ha
s
a
n
e
f
f
e
c
t
on
the
a
c
c
ur
a
c
y
o
f
the
s
e
gmenta
ti
on
a
s
ide
f
r
om
ha
ving
other
pa
r
a
mete
r
s
.
T
he
a
c
c
ur
a
c
y
of
the
s
e
gmenta
ti
on
will
r
e
a
c
h
Evaluation Warning : The document was created with Spire.PDF for Python.