TELKOM
NIKA
, Vol.13, No
.3, Septembe
r 2015, pp. 9
55~962
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i3.1803
955
Re
cei
v
ed Ma
rch 3
1
, 2015;
Re
vised June
18, 2015; Accepte
d
Jul
y
4
,
2015
Applications of Improved Ant Colony Optimization
Clustering Algorithm in Image Segmentation
J
unhui
Zh
ou
1
, Defa Hu
2*
1
School of Infor
m
ation, Hu
na
n Vocatio
nal C
o
l
l
ege for Nati
on
alities,
Yue
y
a
ng 4
140
00, Hun
an, Ch
i
n
a
2
School of Co
mputer an
d Informatio
n
Engi
n
eeri
ng, Hun
an
Univers
i
t
y
of Commerce,
Cha
ngsh
a
41
0
205, Hu
na
n, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: hdf666
@1
63.
com
A
b
st
r
a
ct
W
hen
express
i
ng th
e d
a
ta fe
ature
extractio
n
of th
e i
n
teres
t
ing
obj
ectives,
i
m
ag
e s
e
g
m
e
n
tation
i
s
to transform th
e data set of the f
eatures
of the orig
ina
l
i
m
age i
n
to
more
tight and g
e
n
e
r
al data set. T
h
is
pap
er expl
ores
the image se
g
m
e
n
tatio
n
tech
nol
ogy b
a
s
ed
on ant col
ony
opti
m
i
z
at
ion cl
usterin
g
alg
o
rit
h
m
and
pro
pos
es
an
i
m
prov
ed
ant co
lo
ny c
l
usteri
ng
alg
o
r
i
thm (ACCA).
It impr
oves
and
an
aly
z
e
s
the
computati
o
n
a
l formu
l
a of the simila
r
i
ty function an
d i
m
prov
es para
m
et
er selecti
on an
d setting by setti
n
g
ant clusteri
ng r
u
les. T
h
rou
gh t
h
is al
gor
ith
m
, it
can not on
ly a
cceler
a
te the
cl
usterin
g
spe
e
d
,
but it can als
o
have a
better clusteri
ng p
a
rtitioni
ng res
u
lt. T
he exp
e
ri
me
nt
al resu
lt show
s that the
meth
od of this pa
pe
r is
better than the
origi
n
a
l
OT
SU ima
ge se
g
m
ent
ation
meth
od i
n
accuracy, ra
pidity a
nd stab
i
lity.
Ke
y
w
ords
: ant
colony o
p
ti
mi
zation, cluster
i
n
g
alg
o
rith
m, i
m
age se
g
m
e
n
tation
Copy
right
©
2015 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The idea of i
m
age segm
e
n
tation is to map the
pixel
points to the
corre
s
po
ndin
g
feature
spa
c
e
a
c
cord
ing to the
fea
t
ures of diffe
rent inte
restin
g blo
c
ks
of a
r
ea in th
e ima
ge
su
ch a
s
gray
scale
and tex
t
ure a
nd th
en
cla
s
sify the
m
acco
rdi
ng to
ce
rtain rea
s
on
able simil
a
rity
criterio
n.
It
has
bee
n m
entione
d in t
he p
r
eviou
s
para
g
raph
th
at image
se
g
m
entation i
s
the key
step
in
image p
r
o
c
e
ssi
ng an
d image an
alysis;
therefore,
the image seg
m
entation effect will direct
ly
affect the accura
cy of the subsequ
ent image an
alysi
s
[1].
Previou
s
ly, Kapur ha
s
publ
ishe
d the
the
o
ry of
n
e
cessary optim
al m
a
tchin
g
d
egre
e
value
in imag
e
seg
m
entation
by usi
ng the
op
timal ent
ropy. The P
-
tile m
e
thod
pro
p
o
s
ed by
Doyle i
s
the early met
hod of autom
atic thre
shol
d
sele
cti
on ba
sed on gray histogram [2]. The computat
ion
of this theo
ry has l
o
w
req
u
irem
ents
on
hard
w
a
r
e
an
d it can
pro
c
ess differe
nt kind
s of ima
g
e
s,
however, its sh
ortcoming
s
a
r
e
also
very obv
io
us
and it
requi
res
huma
n
s to mea
s
u
r
e t
h
e
necessa
ry rat
i
o of the imag
e blocks to the sou
r
ce
ima
ge in adva
n
ce, resulting in
a low repe
ated
utilization fa
ct
or [3]. Du
nn h
a
s
rai
s
ed
an
applie
d theo
ry
. This theo
ry
extract
s
all p
i
xel informati
on
of the entire
image an
d then pe
rform
s
overall com
putation. Wit
h
the com
p
u
t
ational re
sult
, it
see
k
s the va
riance an
d the
n
the optim
al
matchin
g
val
ue [4]. Otsu
seeks the
opti
m
al sol
u
tion
b
y
usin
g the
ma
ximum varia
n
c
e
ratio
of th
e pixel info
rmation of th
e entire ima
ge, which i
s
the
famous O
T
SU method. T
h
is theo
ry has played a
fa
r-rea
c
hin
g
influen
ce on ima
ge pro
c
e
s
sin
g
. It
is n
o
t only
used in
the
ima
ge bl
ock
se
g
m
entation
wit
h
threshold,
b
u
t also u
s
e
d
i
n
ma
ny rel
e
vant
fields. Th
e bi
gge
st advant
age
of this t
heory i
s
th
at it optimizes the
setting
of the thresh
old
correl
ation co
efficient whi
c
h hasn’t bee
n
pro
c
e
s
sed
p
e
rfectly an
d it is more accu
rate in see
k
in
g
this coefficie
n
t than the
previou
s
g
r
a
y
scal
e
di
fference hi
st method. Howe
ver, most of
the
histog
ram
s
o
b
tained fro
m
the images are not
sim
p
le mono
pol
ar-val
ue ima
ges a
nd OT
SU
crite
r
ion i
s
weak i
n
se
gm
enting the
s
e
image
s,
ma
ki
ng it difficult for se
gme
n
ta
tion to rea
c
h
a
s
a
tis
f
ac
tory objec
t
ive [5]. At pres
ent, wit
h
the in-d
e
p
th develo
p
me
nt of the rese
arch in thi
s
fi
eld,
it begin
s
to see the thresh
old segme
n
ta
tion in
the im
age blo
c
k se
gmentation
a
s
the trave
r
sa
l
algorith
m
sol
u
tion of ce
rtai
n pro
b
lem
s
a
nd it star
ts to
apply swarm intelligen
ce al
gorithm into t
he
pro
c
e
ssi
ng of
such pro
b
le
ms.
In orde
r to solve the above-me
ntione
d probl
em
s, this pape
r ha
s introdu
ce
d an
t colony
optimizatio
n clusteri
ng algo
rithm
into
ima
ge segm
ent
at
ion. It firs
tly introduc
e
s the c
h
arac
teris
t
ic
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 955 – 962
956
of
image se
gmentation a
nd
then
pro
poses an
im
proved ant col
ony optimi
z
ation cl
uste
ri
ng
a
l
g
o
r
ithm b
a
s
e
d
o
n
th
e
b
a
s
i
c
an
t c
o
lo
ny a
l
g
o
r
ith
m
, w
h
ic
h
imp
r
o
v
e
s
th
e
p
a
r
a
me
te
r
s
e
ttin
g
and
fitness fu
ncti
on of ACCA; accele
rate
s the segme
n
tation process a
nd find
s the o
p
timal thre
sh
ol
d
of the image
to be se
gmen
ted with ACCA as the th
re
shol
d se
arch
strategy
. Fin
a
lly, it compares
the perfo
rma
n
ce
s of ACCA and OTSU
in t
he segm
e
n
tation re
sult and converge
nce.
2. Image Segmenta
tion
Whe
n
expre
ssi
ng the d
a
t
a feature e
x
tracti
on of
the intere
stin
g obje
c
tives,
image
segm
entation
is to
tra
n
sfo
r
m the
data
set of the fe
ature
s
of
the
o
r
iginal
imag
e
into m
o
re
tig
h
t
and
gen
eral
data
set. If t
he va
rian
ce
gene
rated
in
the ima
ge
se
gmentation
i
s
tran
sferre
d t
o
a
highe
r link of
image proce
ssi
ng, the im
age segme
n
tation will hav
e a bigge
r varian
ce an
d it will
have a big
g
e
r
influen
ce o
n
the final image p
r
o
c
e
s
sing; the
r
efore, image seg
m
entation eff
e
ct
plays an obv
ious role in the entire ima
ge pr
o
c
e
s
sin
g
. Basically, image segm
entation can b
e
divided into the followi
ng four ste
p
s:
Figure 1. Pro
c
e
ss
cha
r
t of image segm
e
n
tation
What ha
s be
en used mo
re frequ
ently at pr
e
s
ent is an image
segmentatio
n definition
formed f
r
om t
he con
c
ept o
f
set in math
ematics. Express an
entire
image
with set
Q
; segm
ent
this image an
d form
N
blocks of area, na
mely divide set
Q
into
N
subsets. Every sub-set i
s
non-empty an
d every sub
s
et shoul
d me
et the followin
g
con
d
ition
s
:
(1) All
i
and
j
ij
should me
et that
ij
QQ
ф
=
;
(2)
()
1
N
i
i
QI
,
()
(
)
ij
QQ
ф
=
,
i
,
j
,
ij
;
(3) Whe
n
ij
,
P(
)
F
a
l
s
e
ij
QQ
;
(4) Whe
n
1
,
2
......
iN
=,
Tr
u
e
i
PQ
;
(5) Whe
n
1
,
2
......
iN
=,
i
Q
is the co
nne
cted
region.
Here, symbol
ф
is the con
c
ept of empty set in math
ematical
set
while
i
P
Q
is the
logical p
r
edi
cate
of the
element
s in
all sub
s
ets
i
Q
and it h
a
s
spe
c
ific mea
n
ing. Detaile
d
illustratio
n
wil
l
be made ab
out the five condition
s in the followi
ng p
a
ssag
e [6].
Con
d
ition
(1) indi
cate
s tha
t
all the bl
ocks of
are
a
afte
r ima
ge
seg
m
entation
ca
n’t hav
e
the same pix
e
l point; in other word
s, a
pixel point can only belo
ng to a point and a cla
ss.
In
mathemati
cs,
it means that
the segm
ent
ed su
bset
s can’t have inte
rse
ction.
Con
d
ition (2
) sh
o
w
s
that after the image
segm
e
n
tation, all pi
xel points
sh
all be allo
cat
ed to their o
w
n corre
s
po
n
d
ing
sub
s
et
s, nam
ely that every
pixel p
o
int
shall h
a
ve
its
own
cl
ass
an
d that the
r
e
i
s
n
o
u
n
cl
assi
fied
pixel point.
Con
d
ition
(3) demo
n
st
rate
s that
after i
m
age
se
gme
n
tation i
s
fini
she
d
, the pi
xel
points in diffe
rent
sub
-
blo
c
k of a
r
e
a
defi
n
itely have di
fferent cha
r
a
c
teri
stics, tha
t
is to
say th
at
different sub-blocks of are
a
have no
sa
me pixel
poin
t
s. Conditio
n
(4) m
anife
sts
that after ima
ge
segm
entation
is compl
e
te
d, every seg
m
ented sub
-
block of are
a
has its o
w
n
characte
ri
stics,
whi
c
h are different from ot
her sub
-
blo
cks of area,
na
mely that all
the pixel poin
t
s of the sam
e
sub
-
blo
c
k of
area
have
ce
rtain
sam
e
o
r
simila
r
ch
aracteri
stics. Condition
(5
)
states that
after
image segm
e
n
tation is co
mpleted, all p
i
xel points of
the sam
e
su
b-block of are
a
are conn
ecte
d.
By analyzin
g
the ab
ove five con
d
ition
s
, it
can b
e
known that th
e found
ation
of image
segm
entation
is the discon
tinuity and similarity of
the image pixels. Discontinuit
y
means that the
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Applicatio
ns
of Im
proved
Ant Colon
y
O
p
tim
i
zation Cl
usteri
ng Algo
rithm
in… (Ju
nhui Zho
u
)
957
cha
r
a
c
teri
stics of the pixels in different
sub
-
bl
o
c
ks of
area are different to a ce
rtain extent and
they
form in
coh
e
re
nt pix
e
l ch
ar
acte
ri
stic
s in
different
sub
-
blo
c
ks of
are
a
before it forms
discontin
uity.
Similarity refe
rs to that the
pixels
of the
same
su
b-bl
o
ck
of are
a
ha
ve certai
n pixel
c
h
ar
ac
ter
i
s
t
ics
with high s
i
milar
i
ty s
u
c
h
as
the
grayscale value
and c
o
lor
featur
e of the pixel.
In
pra
c
tice, th
e
pra
c
tical ap
plicatio
ns
of both ima
g
e
pro
c
e
s
sing
and im
age
analysi
s
a
r
e
the
appli
c
ation
s
for certain
specifi
c
o
c
casions. T
herefore, mo
st of
the imag
e
segm
entation
in
pra
c
tical
ap
pl
ication
s
d
o
e
s
n’t need
to m
eet all the
ab
ove condition
s; in
stead, th
ey only ne
ed
to
satisfy one o
r
several
con
d
i
tions a
c
cordi
ng to spe
c
ific
appli
c
ation re
quire
ment
s [7].
OTSU im
age
seg
m
entatio
n method
ex
tract
s
gr
ay hi
stogram of a
ll pixel point
s in the
sou
r
ce imag
e
to perform t
he first-step
pro
c
e
ssi
ng. Its se
co
nd
ste
p
is to co
mp
are a
nd get t
h
e
maximum val
ue from th
e
histog
ram i
n
formatio
n calculated in th
e
last ste
p
an
d
then dete
r
mi
ne
the threshold
of the image segm
entati
on acco
rdin
g
to the maxi
mum varian
ce. This is the
method
of self-
adaptive thr
e
s
hold
s
e
lec
t
i
on. The
s
egm
entation pr
inc
i
ple of OTSU image
segm
entation
method
is to see
k
0
-
a
n
d
1-or
der m
o
ments defin
ed o
n
the
b
a
si
s of the
g
r
ay
histog
ram
of
the ima
ge
an
d its advanta
ge i
s
that
it
d
oesn’t requi
re
any p
r
io
r inf
o
rmatio
n, saving
some
man
p
o
w
er
su
pe
rvisi
on, therefo
r
e,
it can b
e
ap
plied in m
a
n
y
practi
cal fi
elds
and it h
a
s
become an i
m
porta
nt theory in image
segm
entatio
n.
At the same time, the sho
r
tcomi
n
g
s
of
OTSU se
gm
entation
met
hod
are al
so
very o
b
viou
s. Its
optimal
sol
u
tion, n
a
m
ely to
see
k
the
maximum m
a
tchin
g
d
egree valu
e
req
u
ire
s
to
tr
av
erse all
info
rmation.
Be
si
des, after de
eply
investigatin
g
OTSU m
e
tho
d
and
promo
t
ing it to two
-
dime
nsi
on, i
t
s pri
n
ci
ple i
s
to d
e
fine t
he
comp
ari
s
o
n
b
e
twee
n the d
a
ta to be extracted in
th
e source ima
ge
and othe
r im
pure
data a
s
the
trace
of
a ma
trix. The
cha
nge
of this computation
a
l
equatio
n i
s
d
e
termin
ed
by
0
,
s
t
,
1
,
s
t
and
,
j
s
t
. Additionally:
0
(,
)
st
ij
io
j
o
ws
t
p
(
1
)
1
(,
)
st
i
io
j
o
s
ti
p
(
2
)
(,
)
st
j
j
io
j
o
s
tj
p
(
3
)
Here, perform accumulative summation on
0
,
ws
t
,
1
,
s
t
and
j
,
s
t
and their re
sul
t
s
need to
be i
n
c
orpo
rated i
n
all
,
s
t
. The
com
p
lexity of this com
putation
has
re
ached
(4
)
OL
and
it can’t meet
the real
-time
requi
rem
ent i
n
pra
c
tica
l ap
plicatio
ns
sin
c
e it is ve
ry time-con
sumi
n
g
.
OTSU i
m
ag
e
se
gmentatio
n metho
d
red
u
ce
s it
s time
compl
e
xity by ado
pting
sp
e
c
ial
cal
c
ul
atio
n
prin
ciple,
ho
wever,
this m
e
thod
re
quire
s ex
ce
ssive
i
n
vestment
an
d it i
s
n
o
t ap
plica
b
le
as well
[8].
Ant col
ony o
p
t
imization
ha
s the
cha
r
a
c
teristi
c
s of
self
-organi
zation,
pa
rallel
an
d
positive
feedba
ck an
d
it ha
s st
ron
g
rob
u
stn
e
ss.
In the
sea
r
ch
pro
c
e
s
s, ev
ery ant i
s
i
n
d
epen
dent from
each other a
n
d
they comm
unicate only throu
gh ph
ero
m
one. The
r
ef
ore, ant col
o
n
y
algorithm can
be seen
as
a dist
ributed
multi-ag
ent system an
d it perfo
rm
s in
depe
ndent
solution
sea
r
ch in
many point
s
of the proble
m
spa
c
e
at the same
tim
e
. It not only increa
se
s th
e relia
bility of the
algorith
m
, bu
t is also ma
ke
s the algo
rithm have stronge
r glo
b
a
l sea
r
ching
ability. At the
begin
n
ing of
the algorith
m
, a single a
r
tificial ant se
a
rch
es th
e sol
u
tion disord
e
r
ly. After som
e
-
time evolution, the artificial ants sp
ont
aneo
usly
se
a
rch
som
e
sol
u
tions
close to the optimal
solutio
n
th
rou
gh the
ph
ero
m
one. T
h
is is a
pro
c
e
s
s fro
m
di
sorde
r
to
order.
The
o
p
timal mat
c
hi
ng
degree value
is the optimal
solution [9].
3.
Impro
v
ed Ant
Colon
y
Clustering
Algorithm
3.1. Principle of An
t Col
o
n
y
Sy
stem
Optimiza
tion
Every ant in the ant
colo
ny here
is a
n
a
u
t
onomo
u
s
e
n
tity. It contin
uou
sly intera
cts
with
the enviro
n
m
ent; it has intelligen
ce,
dynamics
a
n
d
movability and it ca
n p
r
ovide te
chni
cal
sup
port fo
r d
a
ta cl
uste
ring
. In this al
go
rithm, every
ant sta
n
d
s
fo
r a
data
obje
c
t an
d its
ne
xt
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 955 – 962
958
moving p
o
sit
i
on i
s
dete
r
mined
acco
rding to th
e
simila
rity fun
c
tion
and
probability tra
n
s
fer
function b
e
tween itself an
d
the ant in the neigh
borho
od sp
ace. In the mea
n
whil
e, it dynamically
update
s
its
cl
ass nu
mbe
r
a
c
cordi
ng to
cl
usteri
ng
rule
sets. T
he m
o
vement of ant
makes it affe
ct
itself and it
s
obje
c
t in the
neigh
borhoo
d. It can fo
rm a faste
r
a
nd bette
r cl
u
s
ter
after
certain
iteration
s
th
rough
little lo
cal
neigh
bo
rhood
inform
a
t
ion. In the f
o
llowin
g
p
a
ragra
ph, the
an
t
routing
sketch of Figure 2
vividly indicates the p
r
inci
p
l
e of ant colo
ny system op
timization [10
]
.
(a)
(b)
(
c)
(d)
Figure 2. Ant colo
ny clu
s
tering sketch
In
Figu
re 2, 5000
obj
ect
s
are rand
oml
y
distri
buted
in
a ci
rcular regio
n
with
t
he sam
e
200-unit radi
us an
d 10 art
i
ficial ants
clu
s
ter the
s
e obj
ects. (a
), (b
), (c) and (d) in
Figure 2
sho
w
the re
sult of
ant clu
s
teri
ng
in different i
t
erat
ion
s
. Th
e ran
domly
distrib
u
ted o
b
ject
s are firstly
clu
s
tere
d into
several
sma
ll classe
s wit
h
in a sho
r
t time and then
gradu
ally cl
ustered into
big
cla
s
ses repe
ating the
a
b
o
v
e process; t
he p
h
e
r
om
o
n
e
inten
s
ity in
the path
p
r
e
s
ents
a
bigg
er
and
bigge
r differe
nce. Finally, all the ants choo
se the pat
h with high p
hero
m
on
e intensity [11].
3.2. Algorith
m
Descrip
tio
n
3.2.1. Ant Ex
pression
The alg
o
rith
m simul
a
tes t
he be
havior
of the age
nt ant to “sea
rch pro
p
e
r
swa
r
m” in
th
e
movement
s. Its basi
c
id
e
a
is to rand
omly proj
ect
the obje
c
ts to be cl
ust
e
red i
n
a t
w
o-
dimen
s
ion
a
l grid an
d every subject ha
s a randomly
initial position
.
Every ant c
an move in the
grid
s. Mea
s
u
r
e the
swa
r
m
simila
rity of the curr
ent o
b
j
ect in th
e lo
cal enviro
n
me
nt. Tran
sfer t
he
sw
arm
simila
rity to the pro
bability of the moving obje
c
t throu
gh p
r
obability tran
sfer fun
c
tion
and
pick u
p
o
r
p
u
t do
wn the
su
bject
s
at thi
s
prob
ability. T
he joint
actio
n
of multi
-
age
nt ant ma
ke
s
the
obje
c
ts of the
same
class i
n
the same
spatial domai
n
cluste
r toget
her [12].
An ant agent refers to a da
ta object; in this
way, it can redu
ce the
comp
utation time and
the storage
space. Its beh
aviors
ar
e si
mple: wh
en it
doe
sn’t find
a pro
p
e
r
col
o
n
y, it keep
s on
sea
r
ching;
when it
ha
s fo
u
nd the
relatively prope
r
col
ony, it sto
p
s
moving. It rep
eats thi
s
process
until all the
agent a
n
ts h
a
ve found th
e prope
r po
sitions. W
e
u
s
e the foll
owing qui
ntuple
to
descri
be the
state of every
ant:
n
i
v
s
c
y
x
i
i
i
i
i
1
),
,
,
,
,
(
Here,
n
is the
numbe
r of
d
a
ta,
i
x
and
i
y
are t
he
coo
r
din
a
tes
of wh
ere
the
th
i
ant is
locate
d,
i
c
is the
cla
ss
numb
e
r
of the ant,
i
s
re
flects
wheth
e
r
the ant i
s
in
the moveme
nt state or
the stability state and
i
v
is the spee
d of the ant in its movement [13].
3.2.2. Calcul
ation of Similarit
y
Assu
ming the
r
e is an a
n
t
i
a
in the place
r
at the mome
nt
t
, define the av
erag
e simila
ri
ty
of
i
a
and its nei
ghbo
ur obj
ect
j
a
as:
()
ma
x
(,
)
1
()
1
(1
(
1
)
/
)
ji
ij
i
aN
a
i
da
a
fa
av
v
(
4
)
In this formula,
i
Na
is the neighbou
rho
od of
i
a
. In
the grid, every agent a
n
t will have 8
blan
ks in its
neigh
borhoo
d
if this ant is
locate
d
anywhere b
u
t the boun
dary; if the ant is in the
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Applicatio
ns
of Im
proved
Ant Colon
y
O
p
tim
i
zation Cl
usteri
ng Algo
rithm
in… (Ju
nhui Zho
u
)
959
boun
dary of the gri
d
, there
are only 3 g
r
i
d
s in its n
e
igh
bourhoo
d [14
].
max
v
is the maximum speed
and
i
v
is the
m
o
vement
spe
ed. Ta
ke
a d
e
crea
sing
ra
n
dom n
u
mbe
r
here
to m
a
ke
the a
n
t mov
e
fast at the
b
eginni
ng in
o
r
de
r to fo
rm
a cl
uste
r fa
st and
ro
ughly
.
Then th
e
speed
re
du
ce
s
rand
omly to refine the clu
s
tering result.
,
ij
da
a
is the dista
n
ce b
e
twee
n
ant
i
a
and
j
a
, whi
c
h ca
n b
e
cal
c
ulated
in the
prep
ro
ce
ssin
g pha
se of the clu
s
terin
g
a
nd its definitio
n is as follo
ws:
m
k
jk
ik
j
i
a
a
a
a
d
1
2
)
(
)
,
(
(
5
)
In the fo
rmul
a,
m
is the
num
ber of p
r
op
ert
i
es,
is the coefficient to
adjust the
similarity
betwe
en the
data obj
ect
s
and it dete
r
mines th
e cl
usteri
ng n
u
m
ber a
nd
con
v
ergen
ce
sp
eed,
makin
g
the value of
gra
d
u
a
lly decrea
s
e
in the cluste
ring pro
c
e
s
s.
The ch
ang
e o
f
paramete
r
d
i
rectly affects
the value ran
ge of the simi
larity function
and
the value of
is the numb
e
r
of ants in the neigh
bou
rho
o
d
[15].
3.2.3. Calcul
ation of Mov
e
ment Probabilit
y
Proba
bility transfe
r fun
c
ti
on i
s
the fu
nction
of
i
f
a
and it tran
sfers the
avera
g
e
simila
rity of the data obj
e
c
ts a
s
the m
o
vement pro
bability
m
P
of the agent ants. I
f
i
f
a
is v
e
ry
small; the
n
m
P
is very hi
gh;
otherwi
se,
m
P
is very lo
w.
We
sele
ct Si
gmoid fu
ncti
on a
s
the
prob
ability transfer fun
c
tion
becau
se Sig
m
oid f
unctio
n
has non
-line
a
rity. Only one param
eter i
s
need
ed to be
adju
s
ted in th
e movement
and it has fa
ster conve
r
g
e
n
c
e spee
d.
Define the
movement p
r
obability
m
P
of
i
a
whi
c
h ha
sn’t
found prop
er colony in
the
rand
om move
ment as:
1
S
igm
o
id
mi
Pf
a
(
6
)
In this form
ula,
1
()
1
cx
cx
e
Sigmo
i
d
x
e
is the n
a
tu
ral expo
nenti
a
l form. If the para
m
eter
c
is
big, the curve
saturat
e
s fa
st and the con
v
ergen
ce
spe
ed is al
so fast
[16].
3.2.4. Descri
p
tion of
Clus
tering Rules
C
l
us
te
r
i
ng
Ru
l
e
:
th
e
c
l
ass
i
n
for
m
a
t
i
on
i
c
of
i
a
is up
dat
ed a
c
cording
to the foll
o
w
ing
rule
s:
(a) If
i
a
finds a
pr
oper c
l
ass
i
fic
a
tion and
s
t
ops
moving, its
c
l
ass
num
ber will
be
c
h
anged
by the class n
u
mbe
r
of the agent ant wit
h
t
he minimu
m similarity in
its neighb
orh
ood;
(b) If
i
a
ch
ang
e
s
fro
m
h
a
lt
state to m
o
vement
state,
its
cla
s
s nu
mber will
also be
cha
nge
d by its mark nu
mb
er du
ring its
movement;
(c
) If
i
a
keep
s
its moveme
nt state u
n
cha
nged, its cl
a
s
s num
ber w
ill also
re
mai
n
the
same.
The
clu
s
terin
g
rul
e
s are ve
ry sim
p
le a
n
d
t
he ag
ent a
n
t
s can u
pdate
their
state
wi
th local
informatio
n a
nd fo
rm
clu
s
tering
dyna
mi
cally. Th
e
clu
s
terin
g
re
sult
will
be
given
directly by t
he
cla
ss n
u
mbe
r
of the agent ant [17].
3.2.5. Descri
p
tion of the
Algorithm Pr
ocess
Thro
ugh the
above anal
ysis, the pro
c
e
ss of multi
-
age
nt ant clusteri
ng alg
o
r
ithm is
descri
bed a
s
Figure 3.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 955 – 962
960
Figure 3. The
process of a
n
t cluste
ring
algorith
m
4. Analy
s
is o
f
Experimen
t
al Results
This p
ape
r comp
ares th
e seg
m
entat
ion efficien
cy by using
OTSU an
d
ACCA
segm
entation
method
s o
n
the sta
nda
rd test ima
g
e
Coin
s. In th
is expe
rime
n
t, the hard
w
are
memory i
s
2.5G and the
run-time e
n
vironment
is M
a
tlab20
12. The co
mpa
r
ison re
sult of the
matchin
g
val
ue an
d the
p
r
ocessin
g
tim
e
by the
s
e t
w
o
seg
m
enta
t
ion metho
d
s is in
dicated
in
Table 1. Th
e
segm
ented
e
x
perime
n
tal result
s are sh
own in
Figu
re
4. Figure
4(a
)
is the
origi
n
a
l
image; Fig
u
re 4(b) i
s
the
i
m
age
se
gme
n
ted by
O
T
SU segm
entati
on meth
od a
nd Fig
u
re
4(c) is
the image
se
gmented
by ACCA segm
e
n
tation metho
d
. We can ge
t the standa
rd of the pra
c
t
i
cal
appli
c
ation
s
from the
s
e ob
vious compa
r
ison
s.
Table 1. Co
m
pari
s
on of Ma
tching
Deg
r
e
e
Value and
Processin
g
Time after Cal
c
ulation
Experiment No.
OTSU imag
e seg
m
entation metho
d
ACCA image seg
m
entation metho
d
Image threshold
Time (ms)
Image threshold
Time (ms)
1
103 24.97
121
18.37
2
103 25.05
128
17.93
3
103 24.38
136
17.34
4
103 24.72
133
17.61
5
103 25.02
123
18.02
It can be
se
en from T
abl
e 1 that ACCA meet
s th
e pra
c
tical a
pplication req
u
irem
ents
and it greatly redu
ce
s the p
r
ocessin
g
time and
the co
mputational ti
me of the thresh
old.
From the a
n
a
l
ysis of the co
mpari
s
o
n
of the
above figu
res, it ca
n be
see
n
that ACCA ca
n
sea
r
ch the
optimal thre
shol
d fast
a
nd qui
ckly. The meth
od
of this p
a
per
ha
s certain
effectivene
ss in segme
n
ta
tion effect, segment
atio
n spe
ed
a
nd converg
e
n
c
e and
it
h
a
s b
e
tter
segm
entation
effects and
clear texture
compa
r
ed
with
OTSU imag
e
segme
n
tatio
n
method.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Applicatio
ns
of Im
proved
Ant Colon
y
O
p
tim
i
zation Cl
usteri
ng Algo
rithm
in… (Ju
nhui Zho
u
)
961
(a) O
r
igin
al image
(b) O
T
SU se
gmentation
(c
) A
CCA
seg
m
entation
Figure 4. Co
mpari
s
o
n
of image segm
e
n
tation effect
s by different
method
s
5. Conclusio
n
This p
ape
r has
reali
z
ed
ant colony
optimizatio
n
cluste
rin
g
algorith
m
in
image
segm
entation
,
simulate
d the be
havior
of “searchi
n
g
the ap
pro
p
ri
ate swa
r
m” i
n
the move
m
ent;
dynamically update
s
its
class-m
ar
k a
c
cording to
certain
clu
s
tering rul
e
s a
n
d
guide
s the
ant
individual
s to
find the opti
m
al se
gme
n
tation thre
sh
ol
d qui
ckly. By com
pari
ng t
he pe
rform
a
n
c
e
s
of the sta
nda
rd ima
ge te
st
and th
e trad
itional
imag
e
segm
entation
method
s, th
e expe
riment
al
result ha
s d
e
mon
s
trate
d
that for the
same im
age
to be
segm
e
n
ted, the
col
ony intellig
en
ce
optimizatio
n clusteri
ng alg
o
r
ithm ha
s a fast opt
imizatio
n spe
ed an
d high optimi
z
a
t
ion quality.
Ackn
o
w
l
e
dg
ements
This
wo
rk
wa
s supp
orted
by the Natio
n
a
l Natu
ral S
c
ience Fou
n
d
a
tion of Chin
a (G
rant
No: 612
024
6
4
).
Referen
ces
[1]
Naos
hen
g Qia
o
, Ping Su
n. Stud
y
of Improv
ed Ots
u Algor
ithm and Its Rat
i
on Eva
l
u
a
tion
Anal
ys
is for
PCB Photo
e
le
ctric Image S
egme
n
tatio
n
.
Optik-Internati
ona
l Jour
nal f
o
r Lig
h
t an
d Electron Optic
s
.
201
4; 125(
17): 478
4-47
87.
[2]
Xi
nso
ng W
a
n
g
, Guofen
g Q
i
n. Pav
e
ment
Im
age S
egme
n
tation
Bas
e
d
on F
C
M
Alg
o
rithm Us
in
g
Neig
hb
orh
ood Information.
T
E
LKOMNIKA Indo
nesi
an J
o
u
r
nal
of Electric
al En
gin
eer
ing
.
201
2; 10(
7):
161
0-16
14.
[3]
Dirk H
o
lz, Sv
e
n
Be
hnk
e. Ap
p
r
oximate
T
r
iangul
ation
a
n
d
R
egi
on Gro
w
i
n
g
for Efficie
n
t S
egme
n
tatio
n
and Smo
o
thi
n
g
of Range Ima
ges.
Rob
o
tics and Auto
no
mo
us Systems
. 2
014; 62(
9): 128
2-12
93.
[4]
Shib
ai Yi
n, et
al. Efficient M
u
ltilev
e
l Ima
g
e
Se
g
m
en
ta
ti
on th
ro
u
g
h
Fu
zzy
En
tro
p
y
Maxi
mi
za
ti
on
an
d
Graph Cut Opti
mizatio
n
.
Pattern Recognition
.
2014; 4
7
(9): 2
894-
290
7.
[5]
Sebasti
an T
Gollmer, et al. Using im
ag
e Segme
n
ta
tio
n
for Evaluati
ng 3
D
Statistical Shap
e Mod
e
l
s
Built
w
i
th Gro
u
p
w
is
e C
o
rre
s
pon
de
nce Opti
mizatio
n
.
Co
mputer Visi
on
a
nd Imag
e Un
derstan
din
g
.
201
4; 125: 28
3
-
303.
[6]
Sijie
Ni
u, et a
l
. Automate
d R
e
tinal
La
ye
rs Se
gmentati
on in SD-OCT
Images
Usi
ng Du
al-
G
radie
n
t
an
d
Spatia
l C
o
rrel
a
tion Sm
oothn
e
ss Constra
i
nt.
Co
mp
uters
i
n
Biol
ogy and
M
edici
ne
. 20
14;
54(1): 11
6
-
128.
[7]
Sung
H
a
Ka
ng
, Riccar
do M
a
r
c
h. Multi
p
h
a
se
Image S
egm
e
n
tation
vi
a E
q
u
a
ll
y
Dista
nce
d
Multipl
e
W
e
ll
Potentia
l.
Jour
nal of Visu
al C
o
mmunic
a
tio
n
and I
m
ag
e Re
prese
n
tatio
n
.
2014; 25(
6): 144
6-14
59.
[8]
Ozge Oztimur
Karad
ag, F
a
to
s T
Yarman V
u
ral. Ima
ge S
egme
n
tatio
n
b
y
F
u
s
i
o
n
of
Lo
w
L
e
vel
an
d
Domai
n
Spec
ifi
c
Information v
i
a Markov R
a
n
dom F
i
el
ds.
Pattern Recognition Letters
. 20
1
4
; 46(1): 75-
82.
[9]
Nan Ge
ng, D
o
ngji
an
He, Ya
n
s
hua
ng S
o
n
g
. Camera
Image
Mosaic
ing
Ba
sed o
n
a
n
Opti
mized S
URF
Algorit
hm.
T
E
LKOMNIKA Indones
ian J
ourn
a
l of Electrica
l
Engi
neer
in
g
. 2012; 10(
8): 218
3-21
93.
[10]
Sina T
abakhi,
Parham Mor
adi, F
a
rdi
n
Akhla
ghi
an. An Unsu
pervis
ed
F
eature Sel
e
ction Alg
o
rithm
Ba
se
d on
An
t
C
o
l
ony
Op
ti
miza
ti
o
n
.
E
n
g
i
ne
erin
g Ap
plic
ati
ons
of Artificia
l
Intell
ige
n
ce
. 2
014; 32: 12-
123.
[11]
Aritra Ch
o
w
d
h
u
r
y
, S
w
a
g
a
ta
m Das. Auto
mati
c Sha
pe
Indep
en
dent
Clusteri
ng Ins
p
ire
d
b
y
Ant
Dy
namics.
Sw
arm a
nd Evo
l
ut
ion
a
ry Co
mp
utatio.
201
2; 3: 33-45.
[12]
Sala
bat K
han,
Abdu
l R
auf B
a
ig, W
a
se
em S
hahz
ad. A
Nov
e
l A
n
t C
o
lo
n
y
Optimizatio
n
B
a
sed
Si
ngl
e
Path Hi
erarch
i
c
al Cl
assific
a
ti
on Al
gorithm f
o
r Pred
icting
Gene Ontol
o
g
y
.
Ap
pli
ed S
o
ft Computi
n
g
.
201
4; 16: 34-4
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 3, September 20
15 : 955 – 962
962
[13]
Abou
l Ella H
a
ssani
en, et al. MRI Breast
Canc
er Dia
gn
osis H
y
bri
d
Appro
a
ch Usi
n
g
Adaptive Ant-
Based Se
gme
n
tation a
nd Mu
ltila
yer P
e
rcept
ron Ne
ural N
e
tw
o
r
ks Class
ifie
r.
Applie
d Soft Computi
n
g
.
201
4; 14(A): 6
2
-71.
[14]
Arash Ghan
ba
ri, et al. A C
oop
erative Ant
Colon
y
Opti
mizatio
n
-Genet
ic Algorithm
Appro
a
ch for
Constructi
on o
f
Energ
y
Dem
and
F
o
recast
i
n
g Kn
o
w
l
e
d
ge-
Based
E
x
p
e
rt
S
y
stems.
K
n
o
w
ledge-B
a
se
d
System
s
. 20
13
; 39: 194-2
06.
[15]
Md Mo
niru
l Ka
bir, Md
Sha
h
j
a
han, K
a
zu
yu
ki
Murase. A
Ne
w
H
y
br
id
Ant C
o
lo
n
y
Optimiz
a
tion A
l
gor
ith
m
for F
eature Sel
e
ction.
Expert
Systems w
i
th Appl
icatio
ns.
2
012; 39(
3): 374
7-37
63.
[16]
Amio
y Kum
a
r, Madas
u Ha
nm
and
lu
, HM Gup
t
a. Ant Colo
n
y
Optimizatio
n
B
a
sed F
u
zz
y B
i
nar
y Dec
i
si
on
T
r
ee for Bimo
dal
Ha
nd K
nuc
kle Ver
i
ficati
on
S
y
stem.
Ex
pe
rt Systems w
i
th Ap
plic
ations
.
201
3; 4
0
(2):
439-
449.
[17]
RJ Kuo, et al.
Automatic Ke
rnel C
l
uste
ri
ng
w
i
t
h
Bee
Col
o
n
y
Optimizati
on Alg
o
rithm.
Information
Scienc
es
. 201
4; 283(1): 1
07-
122.
[18]
Pabl
o L
o
y
ol
a,
et al. Pre
d
icti
n
g
W
eb
User B
ehav
ior usin
g Lear
nin
g
-Bas
e
d
Ant
C
o
lo
n
y
Optimizatio
n
.
Engi
neer
in
g Applic
atio
ns of Artificial Intel
lig
e
n
ce
. 201
2; 25(
5): 889-8
97.
Evaluation Warning : The document was created with Spire.PDF for Python.