Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 5
,
O
c
tob
e
r
201
6, p
p
. 2
205
~221
0
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
5.1
010
5
2
205
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Speech Recognit
ion Usin
g Combi
n
ed Fuzzy and
Ant Col
o
ny Algorithm
Foo
a
d J
a
lili
1
,
Milad Jafari Barani
2
1
Depart
em
ent o
f
Com
puter
Engi
neering
S
c
ien
c
e
and Res
e
arch
Br
anch,
Is
lam
i
c
Az
ad Univers
i
t
y
Te
hran,
Iran
2
Young Researchers and
Elite Cl
ub, Urmia Br
anch, Islamic Az
ad
University
, Urmia, Iran
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Feb 7, 2016
Rev
i
sed
Au
g 4, 201
6
Accepted Aug 20, 2016
In recent
y
e
ars
various methods
has
been propo
sed for speech r
ecognition
and removing no
ise from the speech signal
be
cam
e an
im
portant
is
sue. In
this
paper a fu
zz
y
s
y
stem
has be
en
proposed for speech r
ecogni
ti
on that c
a
n
obtain
accu
rat
e
res
u
lts
us
ing
clas
s
i
fi
cat
ion of
s
p
eech s
i
gn
als
with “
A
nt
Colon
y
” a
l
gorit
hm
. F
i
rs
t, s
p
ee
ch s
a
m
p
les
are
given to th
e fuz
z
y
s
y
s
t
em
t
o
obt
a
i
n a
pat
t
e
r
n for e
v
e
r
y
se
t
of si
gna
l
s
t
h
at
c
a
n
be
he
l
p
ful
for di
m
e
nsi
ona
lity
reduction,
easier checking of outcome
and better recognitio
n of signals.
Then, th
e “
A
CO” algori
t
hm
is
us
ed to clus
ter th
es
e s
i
gnals
and
determ
ine
a
cluster for each
input signal.
Also, w
ith this m
e
thod we will be able t
o
recognize noise
and consid
er
it
in a se
p
a
rate clu
s
ter and
remove it from th
e
input signal
.
Re
sults show that t
h
e ac
cura
c
y
for
speech de
te
ction
and noise
removal is
desir
a
ble.
Keyword:
An
t co
lon
y
Clu
s
tering
Fuzzy logic
No
ise rem
o
v
a
l
Speec
h recognition
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Foo
a
d Jalili,
Depa
rt
em
ent
of C
o
m
put
er
En
gi
nee
r
i
n
g Sci
e
nce a
n
d
R
e
sear
ch B
r
a
n
c
h
,
Islamic Azad
Uni
v
ersity,
Teh
r
an
, Ir
an.
Em
a
il: fo
o
a
d
j
alili@g
m
ail.co
m
1.
INTRODUCTION
In last two
dec
a
des, speech
recogn
ition
was one of the most i
m
porta
nt
topics
of signal
processing
[1]
,
[2]
.
Va
ri
o
u
s
m
e
t
hods ha
v
e
been p
r
op
ose
d
i
n
rece
nt
y
ears t
h
at
have t
h
e
i
r ow
n st
re
ngt
h
and wea
k
ness
es e.g.
Particle Swarm Op
ti
m
i
zat
io
n
(PSO) co
m
b
in
ed
with
Feed
-f
or
w
a
rd
N
e
u
r
al-
n
etwo
rk
(FNN
)
th
at is called
PSO-
FNN [3], s
p
ee
ch rec
o
gnition with
fuzzy
T
-
S neural-networks [4]
a
n
d
s
p
eech
rec
o
gnit
i
on base
d on RBF
neu
r
al
net
w
or
k
s
[5]
.
Suc
h
m
e
t
h
o
d
s ha
d t
h
ei
r
speci
fi
c
probl
e
m
s for exam
ple they we
re not able to recognize
two sim
ilar signals or fuzzy neural
networks
had
high tim
e
com
p
lexity
problem
because
the m
a
ss of speec
h
si
gnal
s
.
One
of the
difficult
i
es in speech
recognition is
incorrect
pronunciation of
words
or
having accent tha
t
p
u
t
th
e syste
m
in
un
certain
ty.
Fo
r so
lv
ing
this p
r
ob
le
m
,
a fu
zzy
m
o
d
e
l is u
s
ed
to h
e
lp
syste
m
in
u
n
certain
t
y
conditions. Fast and accurate speech rec
o
gnition is done
by reducing signal intervals that is fuzzy m
o
del’s
out
p
u
t
and t
h
e
di
st
i
n
ct
ness of si
gnal
s
i
s
done
by
t
h
e “AC
O
” al
gori
t
h
m
.
Al
so, wi
t
h
exi
s
t
e
nce of n
o
i
s
y
classes
for each sam
p
le, speech recognition can operate under noi
sy conditions and noise re
m
oval can be done by
in
iti
al sig
n
a
ls [6
]-[11
]
.
2.
ANT COL
O
N
Y
ALGO
RIT
H
M
Today
m
o
st of the resea
r
chers are tend to use
n
a
ture based
algo
rith
m
s
like “PSO”,
“ACO” a
nd
“Firefly” to solve variety of
probl
em
s. These algorithm
s
are popul
a
t
i
on-
base
d m
e
t
a
-he
u
ri
st
i
c
searc
h
alg
o
rith
m
s
. “ACO” algo
rith
m
as m
e
n
tio
n
e
d is a po
pu
latio
n-base
d algorit
h
m
that seeks to find t
h
e shortest
pat
h
bet
w
ee
n col
o
ny
and
fo
o
d
so
urce
by
fol
l
owi
n
g t
h
e
pat
h
s t
h
at
m
o
re ot
her a
n
t
s
secret
ed p
h
er
om
ones
from
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
220
5
–
22
10
2
206
th
em
selv
es. Th
is algorith
m
i
s
ab
le t
o
b
e
used
i
n
v
a
riety o
f
op
timizatio
n
p
r
ob
le
m
s
for e
x
am
ple: Traveling
sal
e
sm
an p
r
o
b
l
e
m
[1
2]
, dat
a
cl
ust
e
ri
n
g
[1
3]
,[
14]
a
n
d
t
e
xt
m
i
ni
ng
[1
5]
,[
1
6
]
.
T
h
i
s
al
g
o
ri
t
h
m
can be e
x
p
l
ai
ned
with
fou
r
ru
les:
Ants
are
seeki
n
g food and t
h
ey sear
ch
di
ffe
rent
pat
h
s
f
o
r
f
i
ndi
n
g
i
t
.
Each a
n
t sec
r
etes pherom
one
from
itself.
An
ts
are attract
ed
to ph
ero
m
o
n
e
sm
ell, so
p
a
th
with m
o
re ph
ero
m
o
n
e
is
p
a
th
with m
o
re traffic.
Phe
r
om
one e
v
apo
r
at
es
by
t
i
m
e, so
pat
h
wi
t
h
less
pherom
one m
eans pat
h
with less t
r
affic.
Th
is algor
ith
m h
a
s 4 ph
ases th
at
i
s
e
xpl
ai
ne
d as
f
o
l
l
o
w:
Initial pherom
one
am
ount for eac
h elem
ent of
phe
rom
one table is cal
culated
with (1)
where
Fitness
(
Sbest
)
in
d
i
cates fitn
ess of solu
tio
n
fo
r t
h
e
prob
lem
.
(1
)
Prob
ab
ility fu
nctio
n
(2
) is u
s
ed
for d
e
term
in
i
n
g
a p
a
t
h
.
Where
p
k
(r,
s
)
sh
ows th
e m
o
v
e
men
t
p
r
ob
ab
ility o
f
ant
k
from
poi
nt
r
to
s.
τ
(r,
s
)
indi
cat
es t
o
am
ou
nt
o
f
ph
er
o
m
one on t
h
e p
a
t
h
and
η
(r,
s
)
shows the fitne
s
s of
m
ovem
e
nt
.
α
and
β
a
r
e c
o
ef
fi
ci
ent
s
u
s
ed
t
o
rega
rd
ei
t
h
er
t
a
bl
e
of
p
h
er
om
one o
r
i
n
f
o
rm
ati
on
gi
ve
n
b
y
p
r
ob
lem
.
At last
J
k
(r
)
is set o
f
trav
elled nod
es and
r
is last no
d
e
m
e
t.
,
,
.
,
∑
,
.
,
(2
)
To l
o
cal
u
pdat
e
of p
h
er
om
on
e t
a
bl
e t
h
at
causes p
r
eve
n
t
i
n
g ant
s
t
o
m
ove from
just
on
e pat
h
an
d m
a
ke
th
em
to
scan
new
p
a
th
s, (3)
will b
e
u
s
ed
.
,
1
.
,
.
(3
)
Whe
r
e
σ
i
s
l
o
cal
updat
e
coe
f
f
i
ci
ent
,
τ
0
i
s
i
n
i
t
i
al
am
ount
of p
h
er
om
one an
d
τ
(r,
s
)
is v
a
lu
e o
f
(r,
s
)
ele
m
ent in
phe
r
o
m
one t
a
b
l
e.
Fo
r g
l
ob
al update o
f
ph
ero
m
o
n
e
tab
l
e, each
ele
m
en
t o
f
tab
l
e is upd
ated
w
i
th
(4
).
,
1
.
,
.
∆
,
(4
)
Whe
r
e
ρ
i
s
p
h
e
r
om
one’
s
gl
o
b
a
l
up
dat
e
c
o
ef
f
i
ci
ent
(he
r
e i
s
t
a
ken
0
.
1
)
a
n
d
Δτ
(r,
s
)
is calcu
lated
with
(5
).
∆
,
,
(5
)
3.
PROP
OSE
D
METHO
D
Pro
p
o
se
d m
e
t
hod i
n
cl
ud
es t
h
ese st
eps:
fi
rst
a fuzzy
m
ode
l
i
s
desi
gne
d and t
h
e
n
AC
O
cl
ust
e
ri
n
g
alg
o
rith
m
will
b
e
ap
p
lied an
d
at last n
o
i
se
d
e
tectio
n
an
d rem
o
v
a
l will b
e
p
e
rform
e
d
.
3.
1.
D
e
s
i
g
n
in
g F
u
zzy
M
o
d
e
l
A
s
show
n
i
n
Fig
u
r
e
1,
x
n
i
s
i
nput
of f
u
zzy
m
odel
a
nd
y
R
is th
e ou
tpu
t
. Size o
f
th
e
o
u
t
p
u
t
s
is related
to
num
ber
of
r
u
l
e
s w
h
i
c
h a
r
e
rel
a
t
e
d t
o
t
h
e si
gn
al
si
ze. Thi
s
m
eans i
f
we
pe
rf
orm
m
o
re r
u
l
e
s
fo
r
pr
o
p
o
s
ed
m
odel
th
en
th
e
size of ou
tpu
t
will be in
creas
ed
where th
is in
cremen
t is un
d
e
r effect o
f
sign
al size. Ou
t
p
u
t
o
f
fu
zzy
m
odel, for eac
h input s
p
eech
signal i
n
c
ont
i
n
uous form
,
allocates
a value
as
a key.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
S
p
eech
Recognitio
n
Using
Comb
i
n
ed Fu
zzy
a
n
d
An
t C
o
lony Algo
rith
m (Fo
oad
Ja
lili
)
2
207
Fi
gu
re 1.
F
u
zz
y
m
odel
Fu
zzifica
tion
:
a Gaus
si
an m
e
m
b
ershi
p
f
u
nct
i
on
(6
) i
s
use
d
fo
r det
e
rm
i
n
i
ng a m
e
m
b
ershi
p
de
g
r
ee f
o
r ea
c
h
input signal to
each
fuzzy set
.
,
,
,
(6
)
Fuzzy ope
rat
o
r
s
:
fu
zzy in
terface is calcu
lated
with
(7
)
wh
ere
N
adj
is adjust
ment coefficie
n
t and conside
r
ed
n/
4 t
h
at
n
i
s
nu
m
b
er of i
n
p
u
t
s
i
gnal
s
.
∏
∗
(7
)
The
out
put
of
a fuzzy
set
i
s
cal
cul
a
t
e
d wi
t
h
m
e
m
b
ershi
p
fu
nct
i
o
n s
o
o
u
t
put
of
n
ode
j
is calcu
lated
with
(8
).
α
α
/
∑
α
(8
)
Ag
greg
at
i
o
n:
s
i
nce decisi
on
making i
n
a fuzzy interface is
taken
re
garding to all
of t
h
e
rules, in this laye
r
they are all c
o
m
b
ined so
t
h
at
out
put
o
f
a n
o
d
e
k
is
calcu
lated
with
(9).
∑
,
1
,
2
,
,
(9
)
Whe
r
e
W
jk
is ru
le’s
weigh
t
.
After
o
b
t
ain
i
ng
t
h
e
o
u
t
p
u
t
sign
als, AC
O
will be u
s
ed
for cl
u
s
tering
th
em
.
3.
2.
Ant Colony Clustering Al
gorithm
C
l
ust
e
ri
ng i
s
an uns
uper
v
i
s
ed
assortm
e
nt
of sim
i
lar object
s or dat
a
i
n
t
o
som
e
cl
asses.
C
r
i
t
e
ri
on for
si
milar
ity o
f
o
b
j
ects is d
e
termin
ed
with
d
i
stan
ce
m
eas
urem
ent
.
M
a
ny
t
y
pes of di
st
ance can be
us
ed e.g.
Euclidean distance,
m
a
halan
obis distance and city
bl
ock di
st
ance. C
l
ust
e
r cent
e
rs are deci
si
on m
a
ki
ng
vari
abl
e
s t
h
at
can be
gai
n
ed
wi
t
h
m
i
nim
i
zi
ng t
h
e E
u
clidean distance over all
of
t
r
ai
ni
ng
sam
p
les i
n
n
-
dim
e
nsional space. Purpose
of clustering is
mi
nimizing the sum
of city block distance
over
N
sam
p
le
s and
allocating each of sam
p
les to one
of
K
cl
us
t
e
rs. Thi
s
i
s
d
one
by
(1
0)
w
h
ere
K
is number
of clusters
for
N
sa
m
p
les.
,
∑∑
(10
)
Centers of clus
ter are calculated with (11).
∑
∈
(11
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
220
5
–
22
10
2
208
Th
e ob
j
ectiv
e fu
n
c
tio
n
fo
r sam
p
le
i
is calcu
l
a
ted
with
(1
2).
∑
,
(12
)
Where S
train
i
s
num
ber of t
r
ai
ni
ng sam
p
l
e
s and m
i
i
s
defi
ned as a cl
ass t
h
at
sam
p
l
e
s are bel
ongs t
o
i
t
.
3.
3.
Deno
ising Speech Ssi
g
na
ls
In propose
d
method each signal belongs to a clus
ter but as we know signa
l contains
envi
ronm
ent
noi
se a
n
d t
h
i
s
pr
o
b
l
e
m
can affect
rec
o
gni
t
i
on
rat
e
. E
.
g
.
in
Figure 2
sp
ellin
g
of word
“fiv
e” in
six
different
typ
e
is sh
own
an
d no
ise is imp
r
essed th
e
sign
al an
d it is
clear
on the
im
ag
e. T
o
s
o
lv
e th
i
s
prob
lem
,
d
i
fferen
ce
of si
gnal
wi
t
h
noi
se a
n
d wi
t
h
out
noi
se i
s
c
o
nsi
d
e
r
ed
as a
n
o
i
s
e cl
ass an
d
gi
ve
n t
o
t
h
e f
u
zzy
ci
rcui
t
i
n
o
r
de
r t
o
allocate a nois
e class for that
signal.
Fi
gu
re
2.
S
p
el
l
i
ng
o
f
wo
rd
“fi
v
e” i
n
si
x
di
f
f
e
r
ent
t
y
pes
The
n
, the
nois
e
class for a
particular signal
is
fetched a
n
d differe
n
ce
between it and the ori
g
inal
sig
n
a
l is calcu
lated
.
If t
h
ey are equ
a
l or sim
i
lar, th
en
no
ise
is rem
o
v
e
d
from
sig
n
a
l b
y
p
e
rfo
r
m
i
g
bu
t if
th
ey
are no
t, th
en
it sho
u
l
d
b
e
consid
ered
as a
n
e
w no
isy sign
al
an
d
pu
t it in
ano
t
h
e
r no
isy class. In
Fi
g
u
re 3a, th
e
ori
g
inal signal
with
noise is
s
h
own
a
n
d in Figure 3b t
h
e
noise ha
s
been
re
m
oved from
the signal a
n
d s
p
eech
sig
n
a
l
qu
ality i
s
h
i
g
h
.
(a)
(b
)
Fi
gu
re
3.
(a
)
O
r
i
g
i
n
al
s
p
eec
h
si
gnal
wi
t
h
e
n
v
i
ro
nm
ent
noi
se
(
b
)
Speec
h
si
g
n
al
aft
e
r
d
e
n
o
i
s
i
n
g
4.
R
E
SU
LTS AN
D ANA
LY
SIS
To m
a
ke a co
m
p
ari
s
on bet
w
een pr
o
p
o
s
ed
m
e
t
hod an
d ot
her m
e
t
hods a
dat
a
set
wi
t
h
45 per
s
o
n
s i
s
u
s
ed
th
at con
t
ain
s
G
a
u
ssian n
o
i
se
w
ith
15
db
,
20
db
,
25d
b, 30d
b
and a class w
itho
u
t
n
o
i
se.
In
sp
eech
reco
g
n
i
t
i
on o
f
part
i
c
ul
a
r
pers
on a pa
rt
i
c
ul
ar
cl
ass i
s
creat
ed fo
r t
h
at
per
s
on
. The
n
i
t
i
s
com
p
ared wi
t
h
ot
he
r
classes to
determ
ine the m
o
st sim
ilar class.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
S
p
eech
Recognitio
n
Using
Comb
i
n
ed Fu
zzy
a
n
d
An
t C
o
lony Algo
rith
m (Fo
oad
Ja
lili
)
2
209
Tabl
e
1. R
e
s
u
l
t
O
b
t
a
i
n
ed
f
r
om
Per
f
o
r
m
i
ng Pr
op
ose
d
M
e
t
h
o
d
on
Dat
a
set
Pr
oposed M
e
thod
PSO-
F
NN
15
96.
7
94.
8
20
97.
2
95.
7
25
96.
9
96.
7
30
97.
0
96.
2
Clean 97.
8
96.
7
Average Recognition
Rate
97.
12
96.
02
5.
CO
NCL
USI
O
N
In this pa
pe
r a com
b
ined fuz
z
y and ACO
based
m
e
thod had bee
n
proposed for s
p
eech recognition.
Fo
r
d
i
m
e
n
s
io
nality red
u
c
tio
n an
d
b
e
tter reco
gn
itio
n
rate fo
r si
g
n
a
ls
first
th
e p
r
op
osed
fu
zzy system
i
s
u
s
ed.
The
n
AC
O cl
u
s
t
e
ri
ng al
go
ri
t
h
m
i
s
perfo
rm
ed on t
h
e si
g
n
al
s
to
allo
cate th
em to
th
eir
appropriate clusters
with
m
i
nim
i
zi
ng t
h
e
ci
t
y
-bl
o
ck di
st
ance bet
w
e
e
n s
i
gnal
s
an
d cl
us
t
e
r cent
e
rs. T
h
i
s
m
e
t
hod as sh
ow
n i
n
res
u
l
t
s
l
eads
to
b
e
tter qu
ality in
co
m
p
arison
with
o
t
h
e
r pro
p
o
s
ed
al
g
o
rith
m
s
an
d
h
a
s l
o
wer tim
e co
m
p
lex
ity reg
a
rd
t
o
fu
zzy
syste
m
and city-bloc
k
dista
n
ce
m
easur
e that is easier to
com
pute rathe
r
than E
u
clide
a
n distance a
n
d other
distance m
easures.
REFERE
NC
ES
[1]
R. M. Hegde,
et al.
, “Significan
ce of
the modified group
delay
f
eatur
e in sp
eech
recognition
,
”
Audio, Speech
, an
d
Language Processing, IEEE Transactions on
, vol/issue: 15(1)
, pp
. 190-202, 2007.
[2]
B. Ghaza
le
,
et al.
, “
A
com
p
arat
iv
e stud
y
of trad
iti
onal and n
e
wl
y
proposed featur
e
s
for recognit
i
on
of speech und
er
stre
ss,
”
Sp
eech
a
nd Audio
Pr
oc
es
s
i
ng, IE
EE
T
r
ans
actions
on,
vol/issue: 8(4), pp. 4
29-442, 2000
.
[3]
Zhang J. R.,
et al.
, “
A
h
y
brid
partic
le swarm
optim
izat
ion-ba
ckpropagation algorithm for feedforward neur
al
network tr
aining
,”
Applied Ma
th
ematics and
Co
mputation
, vol.
185, pp
. 1026-1
037, 2007
.
[4]
X. Y. Zhang an
d P. W
a
ng, “
I
mproved T-S Fuzz
y
Neur
al
Ne
twork in Applicat
io
n of Speech Rec
ognition S
y
st
em
,”
Computer Engin
eering and
App
l
ications
, vol. 45
,
pp. 246-248
, 20
09.
[5]
G. Li, “Th
e
Stu
d
y
of th
e Speech
Recognition
Based on the O
p
timized RBF
Neural Network
,
” Master D
e
gree
Thesis of
Taiy
uan University
of
Techno
log
y
, Taiy
uan
,
2009
.
[6]
D. Ververid
is a
nd C. Kotropou
los, “
E
m
o
tional
speech
re
cogni
tion: Resour
ces, featur
es, and
m
e
thods,”
Speech
communication
, vol/issue: 48(9), pp.
1162-1181
, 2006.
[7]
H. Jiang, “
C
onfidence m
easures for speech reco
gnition: A survey,
”
Speech com
m
unication
, vol/issue: 45(4), pp.
455-470, 2005
.
[8]
K. Palecek
and J. Chaloupk
a, “A
udio-visual speech recogn
ition
in
nois
y
audio
env
i
ronm
ents,”
Telecommunications
and Signal Processing (
T
SP)
,
2013 36th In
ternational Conf
erence on. IEEE
, 2013.
[9]
P. Dai and Y
.
Soon, “
R
obust speech re
cogni
tion
b
y
usi
ng spe
c
tr
al
subtract
ion with
noise pe
ak shift
i
ng,”
I
E
T Signa
l
Proc
e
ssing
, vo
l/issue: 7(8), pp. 6
84-692, 2013
.
[10]
M. A. Anusuy
a
and S. K. Katti, “
S
peech reco
gnition b
y
m
ach
ine, a rev
i
ew,” arXiv preprint a
r
Xiv: 1001.226
7
,
2010.
[11]
O.
Kalinli,
et a
l
.
, “
N
oise adapt
i
v
e
tra
i
ning for
ro
bust autom
a
ti
c speech r
ecogn
itio
n,”
Aud
i
o, Sp
eech, and Language
Processing, IEEE
Transactions on
, vol/issue: 18
(8), pp
. 1889-19
01, 2010
.
[12]
C. A. Coello,
et al.
,
“
E
volut
ionar
y
algor
ithm
s
for
solving m
u
lti-ob
jec
tive
probl
em
s,” Spring
er,
200
7.
[13]
E. D.
Lumer an
d B. Faieta, “Diversity
and
ad
ap
tation
in populations of
cluster
i
n
g
ants,”
Proceed
i
ngs of the
third
internationa
l co
nference on S
i
mulation o
f
adaptive b
e
havior:
from animals to an
imats 3: from an
imals to animats
3
. MIT Press, 19
94.
[14]
C. F. Tsai,
et a
l
.
, “ACODF: a
novel data clustering approach
for data mining in large datab
a
ses,”
Journal of
Systems and Software
, vol/issue:
73(1), pp
. 133-1
45, 2004
.
[15]
J. Handl
and B
.
Mey
e
r, “Improved an
t-based
cluste
ring
and sor
ting in
a docum
ent r
e
tri
e
va
l in
t
e
rfac
e
,
”
Paralle
l
Problem Solving
from Nature—PPSN
VII
.
Spring
e
r Berl
in He
idel
berg
, pp
. 913-92
3, 2002
.
[16]
K.
M.
Hoe,
et al.
, “Homogeneous ants for web doc
ument similarity
modeling and
cat
egoriz
ation
,
” Ant
Algorithm
s
,
Springer Ber
lin
Heidelb
e
rg, pp.
256-261, 2002
.
Train
i
ng
Al
g
orith
m
SNR
(
db)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE
Vo
l. 6
,
N
o
. 5
,
O
c
tob
e
r
20
16
:
220
5
–
22
10
2
210
BIOGRAP
HI
ES OF
AUTH
ORS
Fooad jalili was
born in
urmia,
Iran in
1987
and
graduated
B.Sc in
computer
en
gineer
ing from
Is
lam
i
c Azad U
n
ivers
i
t
y
Urm
i
a
branch
and M
.
S
c
from
Is
lam
i
c
Azad Univers
i
t
y
Qazvin br
anch
in art
i
fi
cia
l
in
te
lligen
ce
. Re
cen
t
l
e he
is with
Islam
i
c Az
ad Uni
v
ersit
y
Sci
e
nc
e
and rese
arch
Tehran b
r
anch
a
s
a Phd student
in ar
tifi
c
ia
l in
t
e
llig
enc
e
. At
th
e sam
e
tim
e h
e
is professor at
Is
lam
i
c Azad
U
n
ivers
i
t
y
of Ur
m
i
a and h
i
s
res
e
arch
inter
e
s
t
s
in
clude
: Im
age
pr
oces
s
i
ng, S
i
gn
al
processing, Image Authentication, Cr
y
p
tograph
y
, Speech Reco
gnition, Neur
al
networks and
F
u
z
z
y
sy
st
e
m
s
.
M
ilad jafar
i
bar
a
ni
was born i
n
naqadeh, Ir
an in 1985. He
received B.E. degr
ee in computer
engine
ering
fro
m
Is
lam
i
c Azad
Univers
i
t
y
,
Ur
m
i
a, Ir
an,
res
p
e
c
tiv
el
y M
.
S
c
de
gree
in
artif
ici
a
l
intelligen
ce in 2
014 from Qazvin branch
, Islamic
Azad Univ
ersity
, Qazvin, Iran
.
He is professor
in Is
lam
i
c Az
ad
Univers
i
t
y
of
Urm
i
a and Uni
v
ersity
of Applied Scien
ce
and
Techno
log
y
of
Urmia and h
e
is
member of Yo
ung Resear
chers
and
Elite Club
Urmia Bran
ch, Islamic A
zad
Univers
i
t
y
, Urm
i
a, Ir
an. His
res
earch
inter
e
s
t
s
i
n
clude im
ag
e pr
oces
s
i
ng, wat
e
r
m
arking, Im
age
Authenti
cat
ion,
optim
iza
tion
algori
t
hm
s,
Cr
y
p
togr
aph
y
, neural networks and image
authen
tic
ation
.
Evaluation Warning : The document was created with Spire.PDF for Python.