TELKOM
NIKA
, Vol.13, No
.1, March 2
0
1
5
, pp. 164~1
7
2
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v13i1.1319
164
Re
cei
v
ed Se
ptem
ber 25, 2014; Revi
se
d De
ce
m
ber
19, 2014; Accepted Janu
ary 4, 2015
A Detection Method for Transmission Line Insulators
Based on an Improved FCM Algorithm
Bo
Wen W
a
n
g
1
*, Quan Gu
2
1
Electrical Eng
i
ne
erin
g Col
l
e
g
e
, Northeast Di
anli U
n
iv
ersit
y
,
Jinli
n
, 132
01
2, Chin
a
2
School of Computing Science,
Univers
i
t
y
of Glasgo
w
,
Glas
go
w
,
G12
8QQ, United Kingdom
e-mail: w
a
ng
bo
w
e
n
_db
dl@
1
6
3
.com
1
, gu_q
u
an@
hotmai
l
.co
m
2
A
b
st
r
a
ct
An
im
p
r
o
v
ed
se
gm
en
ta
ti
on
Fu
z
z
y
C
-Me
an
s a
l
g
o
r
i
t
hm
(FC
M) i
s
p
r
op
o
s
ed
fo
r the im
a
g
e
recog
n
itio
n of t
r
ans
missi
on
lin
e ins
u
lators. In
this pa
per,
the
improve
d
W
i
e
ner filter
alg
o
rit
h
m
is firstly us
ed
to filtrate and
recover i
m
age
in pre-pr
oces
sing.
T
hen, th
e insu
lator i
m
age is se
g
m
e
n
ted bas
ed o
n
the
improve
d
al
go
rithm F
C
M. F
i
nally,
the c
ont
our of ins
u
lato
r is labe
ll
ed b
y
using c
onn
e
c
ted co
mp
one
n
t
lab
e
ll
ing a
l
g
o
ri
thm. Experi
m
e
n
tal resu
lts ha
ve show
n that the impr
ove
d
W
i
ener
filterin
g alg
o
rith
m may
effectively filter
and recov
e
r the i
m
a
ges; furthermo
re, the i
m
pr
ove
d
F
C
M ima
ge se
gme
n
tation a
l
g
o
rith
m
may acc
u
rate
ly
segment ins
u
l
a
tors from the i
m
a
ge.
Ke
y
w
ords
:
in
sulator, w
i
ener
filterin
g,
F
C
M,
conn
ected co
mpon
ent la
bel
lin
g
1. Introduc
tion
Since the po
wer in
du
stry has be
en pla
y
ing
a key role in the industri
a
lized
worl
d by
providin
g p
o
w
er
sou
r
ces,
it is very im
portant
to m
a
intain
po
wer tran
smi
ssi
on
line
s
withou
t a
failure, a
nd
a
n
in
sulato
r fai
l
ure
ha
s b
e
e
n
kno
w
n
as o
ne of th
e mai
n
cau
s
e
s
of t
he p
o
wer fail
ure
[1]. An insulat
o
r i
s
the
key
portion
of hig
h
volt
age tran
smissio
n
line,
who
s
e pe
rformance
di
re
ctly
influen
ce
s the ope
ration
safety of the entire tr
an
smissi
on line.
Most tran
smissi
on line
s
and
insul
a
tors are exposed in
the outdoo
r environm
ent
, and suffe
ri
ng from
sno
w
, rain, lig
htning
stri
ke for a l
ong time. [2]-[4]. Therefo
r
e, Regul
ar i
n
sp
ectio
n
of transmi
ssio
n lines an
d the
insul
a
tors is an impor
tant work
, [5]-[7].
Unli
ke the tra
d
itional man
u
a
l line insp
ection
method
s of transmi
ssi
on lines, a he
licopte
r
line in
spe
c
tio
n
mitigate
s
the da
mage
to groun
d ve
getation, an
d
lowe
rs both
the lab
o
r a
n
d
dang
er inh
e
rent in traditio
nal manu
al groun
d inspe
c
tions. Thi
s
me
thod improve
s
the insp
ecti
on
quality an
d ef
ficien
cy cha
r
acteri
ze
d
by
a flexible i
n
spectio
n
m
ode
. It take
s le
ss time to
gai
n
an
image, an
d is free from th
e
rest
rictio
ns i
m
pos
ed by th
e geo
gra
phi
cal enviro
n
me
nt [8]-[9]. Image
detectio
n
is
use
d
for t
r
an
smissio
n
lin
e
based
on th
e ae
rial p
hot
ogra
phy te
ch
nology [10]
-[11].
This m
e
thod
mitigates the
defect
s
that o
c
cur from
ma
nual d
e
tectio
n, improve
s
b
o
th the worki
n
g
efficien
cy an
d accu
ra
cy a
s
well a
s
gu
arante
e
s
i
n
sp
ection
quality
;
additionally,
it improve
s
the
safety of transmission line insp
ection and reduces t
he probability of accid
ents or emergencies
[12]. In addition, aeri
a
l ph
otogra
phy technolo
g
y
provi
des the
ba
sis for unma
nne
d aeri
a
l vehicle
(UAV) in
spe
c
tion technolo
g
y and
provides a the
o
re
tical meth
od
within th
e a
s
pe
cts of UAV
navigation a
n
d
flight conditi
on monito
ring
[13]-[15].
There a
r
e a
great
numb
e
r of insulator
detecti
o
n
stu
d
ies
of tran
smissi
on lin
e i
m
age
s al
l
over the wo
rl
d; howeve
r
, there
i
s
com
p
aratively few insulato
r det
ection
studie
s
aimed at the
aerial
tra
n
smissi
on li
ne i
n
spectio
n
im
ag
e. A kno
w
led
ge-b
a
sed
po
wer line
d
e
te
ction
metho
d
is
prop
osed for
a vision ba
se
d UAV su
rveil
l
ance and in
spectio
n
syste
m
[16]. The method u
s
e
s
an
improve
d
ca
n
n
y edge det
e
c
tor to dete
c
t
the edge of
a
tran
smissio
n
line. The d
e
tection effe
ct is
satisfa
c
to
ry d
ue to
the
sim
p
le im
age
ba
ckgro
und
in
the exp
e
rim
e
n
t, but the
ca
n
n
y edg
e d
e
te
ctor
doe
s not p
e
rf
orm
well
with
noise immu
nity when
used with
a co
mplex ba
ckground.
Refere
nce
[17] propo
se
s a ratio al
go
rithm used to
detect
the e
dge of the transmi
ssion li
ne and in
sul
a
tor
becau
se the
ratio algo
rith
m performs
well in noi
se
immunity and
produ
ce
s sa
tisfacto
ry effects.
The
ratio
alg
o
rithm i
s
th
e
method
u
s
ed
to dete
c
t line
a
r ta
rget
s, re
quirin
g
that
th
e di
re
ction
of
the
transmissio
n l
i
ne i
s
pa
rall
el
to the h
o
ri
zon
t
al edg
e of th
e imag
e, whil
e the i
n
sulato
r is a
nonli
n
e
a
r
target. Co
nse
quently, the edge d
e
tectio
n re
sults
are
incom
p
lete, g
i
ven som
e
e
dge
s had
be
en
los
t. The S
U
SAN edge scale invariant
feature
(SESI
F) algorithm,
pres
ented
in
referenc
e [18],
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Detection
Method for T
r
ansm
i
ssion Li
ne Insu
l
a
tors Based on an
Im
proved ....
(BoWen Wang)
165
use
s
an imp
r
oved gra
d
ient
edge dete
c
tion method to
extract insul
a
tors from a tran
smi
ssi
on line,
while
not
co
nformin
g
to t
he requi
rem
e
nts fo
r a
co
mplex ba
ckg
r
oun
d. Th
e h
i
gh voltag
e li
ne
feature extra
c
tion metho
d
base
d
on wavelet image
s, whi
c
h hav
e been p
r
e
s
e
n
ted in refe
re
nce
[19], utilizes the features of wa
velet theory, extracting a hi
gh vo
ltage transmi
s
si
on line to obt
ain
the po
sition
s
of a
high
voltage li
ne
and
i
n
sul
a
tors
accordin
g to
feat
ure
matching.
Given
that th
e
algorith
m
i
s
compl
e
x an
d
the
cal
c
ulati
on p
e
rio
d
i
s
long, th
e al
gorithm
may
meet l
o
w pi
xel
inspection image detection. References [20],[21]
provide an overview
for the national hi
gh
voltage in
sul
a
tor o
n
line
d
e
tection
meth
od a
nd
also
repre
s
e
n
t the
detectio
n
p
r
in
ciple
s
, d
e
tecti
o
n
equipm
ent, advantage
s a
nd disa
dvant
age
s of the
existing me
thods. The
referen
c
e
s
ha
ve
analyzed the
voltage dist
ribution m
e
th
od, the l
eakage current
detectio
n
me
thod, the pul
se
curre
n
t detection method
and the infra
r
ed temp
er
ature me
asure
m
ent method
in conta
c
tin
g
method
s a
ccordin
g to th
e
physi
cal
ch
a
r
acte
ri
stics of
the in
sulat
o
r defe
c
t lea
k
a
ge
curre
n
t. The
disa
dvantag
e
s
to
this met
hod
are hi
gh
labor
stre
n
g
th, low
safety, poo
r
efficien
cy, susceptibi
lity
to electroma
gnetic i
n
terfe
r
en
ce, po
ssi
b
ility of
false dete
c
tion
or lea
k
a
ge
detectio
n
, an
d a
susceptibility to environm
ental element
s (temperature
, humidity, etc.). Refe
rence [22] presents
an extra
c
tion
algorithm fo
r an Insul
a
tor
image u
s
in
g helicopter
ae
rial ph
otos. T
he algo
rithm
in
que
stion
seg
m
ents im
age
s u
s
ing th
e
maximum
entropy th
re
shol
d ba
se
d
on the g
e
n
e
tic
algorith
m
; it filters noi
se
in
se
gme
n
ted i
m
age
s
with
a
dual
-stru
c
tural filter
and
finally ide
n
tifies
the insulato
r
seri
al o
u
tline
usin
g an
ide
n
t
ific
ation met
hod in
the
co
nne
cted a
r
e
a
.
The al
gorith
m
may c
o
mpletely extrac
t ins
u
lato
r imag
e
s
from the a
e
rial ph
otos
on a sim
p
le
backg
rou
nd. The
robu
stne
ss of the algorithm
is to be improv
ed upo
n when de
aling
with a com
p
le
x backgro
und
.
In this pape
r, A novel detection met
hod is
prop
ose
d
for aerial transmi
ssi
on line
insp
ectio
n
image
s, the first step is to re
pro
c
e
ss th
e spatial
swit
ch
ing of aeri
a
l photo color
a
n
d
image
s; the seco
nd is to
segment a
e
ria
l
phot
os
ba
se
d on an im
proved FCM
al
gorithm; finall
y
,
the u
s
e
of id
entification
in
the
con
n
e
c
ted a
r
ea
to i
dentify the o
u
tline of
an i
n
sul
a
tor i
n
t
he
image
s.
2. Image reproces
sing
The expe
rim
ent prove
s
th
at the effect is
poo
r when
dire
ctly swit
ching the
colo
ur imag
e
to a grey-scal
e
map
and th
en segm
entin
g the in
su
lato
r imag
es. T
h
e process
switche
s
the
RGB
spa
c
e
of the
colo
ur im
ag
e to the HSI
spa
c
e,
H d
enote
s
Hue,
S denote
s
S
a
turation, a
n
d
I
denote
s
Inten
s
ity. The swit
chin
g pro
c
e
s
s is sho
w
n a
s
follows:
2
BG
H
BG
(1)
2
()
(
)
arccos
2(
)
(
)
(
)
RG
R
B
RG
R
B
G
B
3
1m
i
n
(
,
,
)
SR
G
B
RG
B
,
[0
,1
]
S
(2)
3
RG
B
I
,
0,
25
5
[]
I
(3)
The HSI sp
ace m
odel i
s
cl
ose to
p
eople'
s visua
l
perceptio
n
about colo
ur; thre
e
comp
one
nts l
a
ck co
rrelatio
n, whe
r
e the
H co
mp
o
nent
is insen
s
itive to ray and
shado
w an
d th
e
S com
pon
ent
ha
s an
effect
on th
e ima
g
i
ng o
b
ject
al
o
ng
with chan
ges in lig
ht in
tensity. As
su
ch,
the model is
able to distin
guish the obj
ects in diffe
re
nt colou
r
s.
A vast numb
e
r
of expe
rime
nts
have
proven that the va
lue of
the S
compon
ent is
smalle
r
whe
n
the ray
of light is stronge
r co
ncurrent with
cha
nge
s in li
ght intensity; con
s
eq
uently, the S
comp
one
nt is sele
cted a
s
a grey
-scale
map
of the i
m
age
segm
e
n
tation and t
r
ansfe
rred to the
pixel spa
c
e [0
, 255].
Aerial ph
otos may degra
d
e
image
s du
ring the
imagi
ng pro
c
e
s
s, so it is ne
ce
ssary to
resto
r
e tho
s
e
images. Ima
ge re
storatio
n pro
c
e
s
ses
the degrade
d
images for t
he purpo
se of
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 164 – 1
7
2
166
resto
r
in
g the
origin
al im
ag
e. Thi
s
p
ape
r uses an
imp
r
oved
Wie
n
e
r
filtering
algo
rithm to
re
sto
r
e
image
s.
Wien
er filteri
ng requi
re
s a
minimum m
e
an squa
re
error b
e
twe
en t
he inp
u
t ima
g
e
(,
)
I
ij
and the
re
st
oring i
m
ag
e
(,
)
Ri
j
. Assuming t
he imag
e
si
gnal a
pproximates
stabl
e
ran
dom
pro
c
e
ss,
it
sh
all meet
:
2
mi
n
(
(
[
(
,
)
(
,
)
]
)
)
WE
I
x
y
R
x
y
(4)
Whe
r
e
()
E
is a
mathemati
c
a
l
expectatio
n
. The exp
r
e
ssion of the
sp
atial dom
ain
for
Wien
er filterin
g is sh
own as follow:
'
ˆ
(,
)
[
(,
)
]
DD
f
xy
E
f
xy
E
D
(5)
Whe
r
e E is the mathemati
c
al expe
ctation in
the nei
ghbo
urh
ood
of point (x, y)
; D is the
varian
ce yield
in the neighb
ourh
ood of p
o
int (x, y); and
'
D
is the varia
n
ce yield of n
o
ise.
E
,
D
and
(,
)
f
xy
are known in expression (5). It
is
ne
ce
ssa
ry to estimate the noi
se
var
i
anc
e
'
D
for the input ima
ge. Achievin
g
this and
det
ermini
ng the
size of the te
mplate in the
aerial
photos
is a difficult
process. T
he
complex
area
of the im
ages will
be fuzzy
if the templ
a
te
is too large;
however, if the template is t
oo sm
all the effect of noise reduction will be
unde
sirable.
Therefore, a
type
of improved
Wien
er filtering a
l
gorithm i
s
pre
s
ente
d
. The
prop
osed alg
o
rithm is
sho
w
n a
s
follows:
1) Use the S
obel Op
erato
r
to filter in
p
u
t image
s, a
nd the ed
ge
and no
n-edg
e will be
obtaine
d.
2) Buil
d a
nei
ghbo
urh
ood
estimate
of n
o
ise
varia
n
ce
'
D
in
a no
n-ed
ge. Supp
osi
n
g
NE
D
is a local vari
ance of point (x, y), build a
local vari
an
ce
in a
55
neigh
bo
urho
od area
of point (x,
y).
Wh
en all points
in
th
e neigh
bou
rho
o
d
a
r
e within
t
he a
r
ea,
it is
not ne
ce
ssary to cali
brate; if
some p
o
ints i
n
the neigh
bo
urho
od a
r
e n
o
t in the
area,
the point is calibrate
d as n
o
ise.
3) Cal
c
ul
ate the noi
se vari
ance
'
NE
D
, the formula is sho
w
n as follo
w:
(,
)
'
(,
)
(,
)
R
NE
xy
NE
D
xy
Dx
y
S
(6)
Whe
r
e R i
s
th
e point set of the curre
n
t non-e
dge; S is the numbe
r of pixels.
3. Image seg
menta
t
ion b
ased on an i
m
prov
ed fuzz
y
c-means algorithm
A fuzzy
c-me
ans alg
o
rith
m (F
CM)i
s
a
wid
e
ly-u
sed
image
segm
entation m
e
thod. Th
e
method
obtai
ns th
e m
e
mb
ership
of e
a
ch sampl
e
p
o
i
n
t for
all
cla
s
s
cent
re
s th
ro
ugh th
e
obje
c
tive
function and determines
the affiliation of sample
points to aut
omatica
lly cl
assify the data
sampl
e
. FCM
may prevent
the proble
m
s that occur
i
n
threshold
settings. The chara
c
te
risti
c
s of
FCM a
r
e
suit
able fo
r the u
n
ce
rtainty an
d fuzzine
s
s e
x
isting in im
a
ges. T
he F
C
M algo
rithm
also
belon
gs to
a
type of
un
supervi
sed
cla
ssifi
cation
m
e
thod; a
clu
s
ter p
r
o
c
e
s
s n
eed
s n
o
m
a
n
ual
intervention a
nd is ap
plicab
le to multiple appli
c
ation fie
l
ds for auto
m
atic image
se
gmentation.
Princi
ple of a traditional F
C
M algo
rithm
:
suppo
se
sa
mple set
12
,,
,
n
X
xx
x
belong
s
to the
p-di
m
ensi
on E
u
cli
dean
spa
c
e,
p
i
x
R
,
1,
2
,
,
in
. Firs
t, c
l
assify s
a
mple
set X into
dif
f
e
rent
cl
as
se
s:
c mut
u
al
ly disjoints
su
bset
s
12
,,
,
c
Vv
v
v
; c is the numb
e
r of
cla
s
ses. Ea
ch
cla
ss
ha
s st
rong
coh
e
sive
ness reflecte
d in the f
eat
ure
s
of ima
g
e
s a
nd la
rge
differen
c
e
s
exist
betwe
en the
cla
s
ses. The
obje
c
tive function is:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Detection
Method for T
r
ansm
i
ssion Li
ne Insu
l
a
tors Based on an
Im
proved ....
(BoWen Wang)
167
2
11
(,
)
(
)
(
)
cn
m
mi
j
i
j
ji
J
UV
u
d
(7)
whe
r
e m
i
s
t
he weighte
d
index,
meetin
g
(1
,
)
m
; n
is
th
e nu
mb
er
o
f
p
i
xe
ls
in
th
e
image
s; c is the p
r
e
s
et n
u
m
ber
of
cla
s
ses, me
eting
1<c
≤
n;
V
is th
e p
×
c cl
uste
r ce
nter
matri
x
;
and
ij
u
expresses
wh
ether sample
j
x
is in t
he
cla
ss i
of t
he
clu
s
ter, n
a
m
ely, the me
mbershi
p
of
sampl
e
j
x
to
i
v
. The valu
e of
ij
u
is 0
o
r
1,
so
U i
s
a
0
-
1
ma
trix of c×n
an
d me
ets th
e
constraint
conditions:
1
1
c
ij
i
u
1,
2
,
j
n
(8)
It also meets:
1
1
n
ij
j
u
1,
2
,
ic
(9)
It may calcula
t
e the Euclide
an dista
n
ce
ij
d
betwe
en the
sampl
e
j
x
and clu
s
ter cente
r
i
v
:
2
()
()
T
i
j
ji
ji
ji
dx
v
x
v
x
v
(10)
whe
r
e
p
j
x
R
, and
p
i
vR
.Und
er the co
nstrai
nt con
d
i
t
ions, upd
ate U to obtain:
21
1
1
c
m
ij
i
j
k
j
k
ud
d
0
kj
d
(11)
If
ik
,
ij
u
=1;
whe
n
ik
,
ij
u
=0.
i
v
is calculat
ed ba
sed o
n
U
:
11
()
()
nn
mm
ii
j
j
i
j
jj
vu
x
u
(12)
(,
)
J
UV
expre
s
ses th
e quad
rati
c sum of the we
ighting di
stan
ce fro
m
the sample to the
clu
s
ter
cente
r
. The value
reflect
s
the con
s
i
s
te
n
c
y degree of the cla
s
ses in
the definition of
spe
c
ific differences. The
cl
uster i
s
more comp
act if
(,
)
J
UV
is smalle
r. The algorith
m
upd
ates
U
and V throu
gh iteration
and altern
ation. Wh
e
n
two perio
d
s
of nearby iterations
meet
()
(
1
)
tt
VV
, the iteration will stop,
where t is
the number of iteration times
and
is the
preset convergence thre
shold value. Then, the objective
function will reach a mi
nimum.
The FCM al
g
o
rithm attra
c
ts ea
ch samp
le in the cluster cente
r
bu
t at a slower rate of
conve
r
ge
nce. To solve the
s
e pr
oble
m
s,
the pape
r pre
s
ent
s a type of FCM algo
ri
thm corre
c
tio
n
ij
u
to improve the rate of co
nv
erge
nce of the clu
s
ter
cent
er.
0
E
is the set of all non-edg
e points in a
n
image;
1
E
is the
set of all edg
e points.
First, divide t
he no
n-e
dge
area i
n
to m
u
tua
lly disj
ointed blo
c
k. Next, divide
the edg
e
points into
nearby bl
ock, forming
mutually
di
sj
ointed and
contin
uou
s sub
c
la
sse
s
and
prelimi
narily segm
enting
i
m
age
s,
thus obtai
nin
g
c m
u
tually disjoi
n
t
ed sub
c
la
sses.
(0)
i
X
, where,
{1
,
2
,
,
}
ic
; when
j
x
is l
o
cated in the
i
th
(0)
i
X
,
(0)
1
ij
u
, or
(0
)
0
ij
u
. Up
on ite
r
ati
on, the
membe
r
ship
update e
quati
on (11
)
is mo
dified as follo
w:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 164 – 1
7
2
168
21
1
1
0
1,
0
1,
0
0
c
m
ij
kj
j
k
j
k
ij
j
k
j
dd
x
E
d
ux
E
d
ot
h
e
r
w
i
s
e
(13)
The clu
s
te
r center up
date
equatio
n (12
)
is modified a
s
follow:
10
1
0
((
)
)
((
)
1
)
jj
j
j
mm
ii
j
j
j
i
j
xE
xE
xE
xE
vu
x
x
u
(14)
Whe
n
1
j
x
E
(ed
ge
point), the
cal
c
ulatio
n eq
ua
tion
is co
nsi
s
t
ent with
the g
eneral F
C
M;
whe
n
0
j
x
E
(non
-e
dge poi
nt), the membe
r
shi
p
will not ch
ange, no mat
t
er how it is i
t
erated,
without the n
eed for recalculation. The
r
e
f
ore, it
shall recal
c
ul
ate an
d determi
ne the edg
e blo
c
k.
FCM only
a
p
p
lies
the gre
y
feature
s
of an
im
age
wh
en
seg
m
entin
g the
imag
e,
without
considering the spatial features
of a
pixel. This
paper utilizes
the
Markov Random Field (M
RF)
to segm
ent
a se
co
nda
ry image. Imp
o
rting t
he M
R
F an
d Gib
b
s rand
om field dist
ributi
o
n
improve
s
the
prio
r p
r
oba
bili
ty pixel distri
buti
on; be
ca
u
s
e the
MRF
and Gi
bb
s ra
ndom fiel
d ha
ve
parity, the MRF may be expre
s
sed b
y
one Gibbs
distributio
n. The probabili
ty of
the pixel i
belon
ging to the cla
s
s i may be sho
w
n a
s
follow:
(
)
exp[
(
)
]
[
(
)
]
i
iN
i
i
lL
pX
j
X
n
j
n
l
(15)
Whe
r
e
()
i
nj
is the numbe
r of node
s wh
en n
e
ighb
ourhoo
d
i
N
is labelled
as j; L is the
cla
ss set
.
Next, utilize t
he pri
o
r probability
ij
p
provided by the Gibbs mo
del, the probability value
of the pixel i
belon
ging to the cla
s
s j. Membe
r
ship
ij
u
may be upda
ted as
(1
)
ij
ij
pu
, where
co
ntrol
s
the
weig
ht facto
r
; the value
will
increa
se
alo
ng
with the
n
o
ise
in the
im
age
s, an
d the
scope i
s
01
. Formula (1
3) ma
y be updated
as:
21
1
1
0
(1
)
,
0
1,
0
0
c
m
ij
i
j
k
j
j
k
j
k
jk
j
ij
ot
h
e
r
w
i
s
e
pd
d
x
E
d
xE
d
u
(16)
The procedu
res of an imp
r
oved FCM al
gorithm a
r
e li
sted a
s
follows:
1) Cal
c
ul
ate the set of marginal point
s a
nd non
-ed
g
e
points.
2) In
put
()
t
U
into the follo
win
g
equatio
n to
calcul
ate the
C-clu
s
ter cen
t
er mat
r
ix
()
t
V
, t=0;
initialize othe
r param
eters, inclu
d
i
ng the
value of iteration paramete
r
and Ma
rkov
factor
.
3) Utilize equ
ation
(11)
to update
()
t
U
an
d
(1
)
t
U
; cal
c
ulate
(1
)
t
V
ba
sed
on
(1
)
t
U
; the first
time of iteration uses F
C
M
to classify prelimina
r
ily.
4) Utilize equation (13) to calc
ul
ate the prior local probability
ij
P
.
5) Input the prio
r local probability
ij
P
into equation (14
)
; use eq
uati
on (14
)
and
(12) to
cal
c
ulate
the membe
r
ship matrix
()
t
U
and cl
uster
cent
re
()
t
V
.
6)
Cho
o
se
a
pro
per
matrix norm to
compa
r
e
()
t
V
and
(1
)
t
V
; if
()
(
1
)
tt
VV
, s
t
op
iteration; or t
=
t+1, an
d ret
u
rn to ste
p
(4
).
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Detection
Method for T
r
ansm
i
ssion Li
ne Insu
l
a
tors Based on an
Im
proved ....
(BoWen Wang)
169
In order to i
dentify the insulato
r outlin
e, it
is nece
s
sary to use
identification
s
in the
connected area. The
specific steps are shown
as follows: the insulato
r area will be obtained
after imag
e segmentatio
n; use th
e ide
n
tification
s in th
e co
nne
cted
area to
identi
f
y the insulat
o
r
outline, sp
ecif
ically labellin
g the insul
a
to
r edge, u
s
in
g the 8-nei
ghb
o
u
rho
od meth
od. Assume the
target i
s
whit
e, the value i
s
25
5, the ba
ckgro
und i
s
black a
nd th
e value i
s
0.
First, compl
e
tely
scan the bin
a
ry image a
n
d
label all target pixe
ls, an
d then label
each pixel in
an 8-conn
ected
area;
c
o
mpare eight
neighborhood pixels
, inc
l
uding
t
he upper surf
ac
e, lo
wer
s
u
rfac
e, left, right
,
upper left, upper right, lower
right
and l
o
wer left. Finally, label the
comparative results.
Speci
a
lly
pro
c
e
ss the p
i
xels witho
u
t the eight ne
arby pixels.
Finally, throu
gh the
pa
ral
l
elogram
rul
e
[22], the
positio
n of t
he in
sulato
r ca
n be
accurately de
tected a
c
cord
ing to its feature
s
.
4. Experimental resul
t
s
For thi
s
expe
riment, two g
r
oup
s
of ima
ges in th
e in
sulator i
m
age
l
i
brato
r
y in th
e matlab
7.4 platform
were rand
oml
y
cho
s
en, in
cluding
30
ima
ges i
n
ea
ch
g
r
oup
(size: 1
28×160
). So
me
image
s po
ssessed a
certa
i
n amount of noise.
Experiment
1
sho
w
s the
result of im
ag
e
processing.
We
ran
doml
y
sele
cted a
n
image
from two g
r
ou
ps of the ima
ge library. Th
e experim
ent
al re
sults a
r
e
sho
w
n in fig
u
r
e 1. Figu
re1
(a)
is the
o
r
igin
al
imag
e; Figu
re 1
(b
) i
s
th
e
result
of
extra
c
ting th
e S
co
mpone
nt afte
r
swit
chin
g from
the RGB
sp
a
c
e to the
HSI
spa
c
e. T
he
S comp
one
nt
is withi
n
the
scope
of 0-2
55. Figu
re1
(b)
sho
w
s a p
r
o
m
inent in
sula
tor area withi
n
the S
com
p
onent ima
ge
and a hi
ghe
r
saturability in
the
image
s, so t
he obj
ect
s
can be
distin
g
u
ish
ed by dif
f
erent
colou
r
s; Figu
re 1
(
c) is th
e re
sul
t
of
Wien
er filtrati
on of the S comp
one
nt, restori
ng the
image; and
Figure 1 (d
) is the re
sult
of
improve
d
Wi
ener filtration
. The result
s show that th
e differen
c
e
s
within a sm
all sco
pe will
be
smooth
ene
d, utilizing the al
gorit
hm
s presented in the p
aper.
(a) O
r
igin
al image
(b) S compo
nent
of HIS
(c) Wiener filt
ering
(d) The propos
ed algorithm
Figure 1. The
result
s of the
image pre-p
r
oce
s
sing
Experiment
2
sh
ows th
e in
sulato
r
re
cog
n
ition
re
sults
obtaine
d utili
zing
the
algo
rithms i
n
the pape
r. Th
e experim
ent
al effect is sh
own in Fi
gure
2. Figure
2
(a
) is the p
r
e
-
proce
s
sed im
a
ge,
Figure 2 (b
) i
s
the re
sult o
f
image se
g
m
entation
util
izing the F
C
M algorith
m
, and Fig
u
re
2 (c) i
s
the re
sult of i
m
age
seg
m
e
n
tation utilizi
ng the im
p
r
o
v
ed FCM
alg
o
rithm. In the
experim
ent, the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 164 – 1
7
2
170
numbe
r of
c
classe
s is thre
e, t
he cl
uste
r
centre i
s
(40,
120, 20
0), th
e co
ndition
st
oppin
g
iterati
on
is 0.000
1 and
the maximum numbe
r of iteration time
s is 100.
As shown in
Figure 2(b),
a tradition
al
FC
M al
gorit
hm wa
s utili
zed to
se
gm
ent the
insul
a
tor
are
a
.
Some noi
se
points or the
small
area i
n
clu
d
ing fe
wer p
o
ints
we
re not rest
rain
ed.
As shown in
Figure 2(c), the
FCM algorithm present
ed in this
paper
was utilized to segment
image
s. Th
e
Markov
ra
n
dom fiel
d
wa
s utili
zed
to
se
con
darily
segment
imag
es, a
n
d
had
the
potential to restrai
n
noi
se
and a sm
aller area. Figu
re
2(d
)
is the re
cog
n
ition re
sult. The red line
has b
een u
s
e
d
to label the insul
a
tor outli
ne.
Experiment
3
is the
se
gm
entation effe
ct us
ing
the i
m
age
se
gme
n
tation alg
o
rit
h
m. The
experim
ental
results
are
sh
own i
n
Fig
u
re
3. For t
he ex
perim
ent, two
gro
u
p
s
of 3
0
image
s
we
re
cho
s
e
n
fro
m
the in
sulato
r
image li
bra
r
y. We
ch
ose t
he
widely u
s
ed K-M
ean
s
algorith
m
, FCM
algorith
m
and
algorith
m
s
p
r
esented i
n
the pa
per
to
compa
r
e. Com
pare
d
to the
FCM al
gorith
m
,
the K-Me
an
s
algorith
m
is faster in
cal
c
u
l
ating s
epa
rat
i
on time
s. Th
e ho
rizontal a
x
is indi
cate
s t
h
e
numbe
r of
pi
xels, and
the
vertical
axis indicates th
e num
ber of
wrong
pixel
s
for the im
a
g
e
segm
entation
algo
rithm. T
he K-M
ean
s
algorith
m
inv
o
lves m
o
re e
rro
r
clu
s
ter p
i
xels, while t
h
e
FCM al
go
rith
m involves l
e
ss. T
he al
go
rithms p
r
e
s
ent
ed in thi
s
p
a
p
e
r h
a
ve a g
o
od segm
entat
ion
effect, effectively reduci
n
g
the numbe
r of erro
r clu
s
te
r pixels.
(a) T
he pre-p
r
ocesse
d ima
ge
(b) The F
C
M algo
rithm
(c) The p
r
op
o
s
ed al
gorith
m
(d) The re
gi
on label
Re
su
lt
Figure 2. The
result
s of the
insulato
r
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
A Detection
Method for T
r
ansm
i
ssion Li
ne Insu
l
a
tors Based on an
Im
proved ....
(BoWen Wang)
171
Figure 3. The
compa
r
i
s
on result
s of thre
e different alg
o
rithm
s
5. Conclusio
n
s
This pa
per p
r
ese
n
ts an im
age re
co
gniti
on met
hod of
transmi
ssion
line insulato
rs ba
sed
on an Improved FCM Alg
o
r
ithm. Du
ring
the pre-
processing
stage,
an impr
ove
d
Wiene
r filteri
ng
algorith
m
wa
s p
r
e
s
e
n
ted t
o
filter
and
re
store
th
e
ima
ges.
Du
ring
t
he im
age
seg
m
entation
sta
ge,
an improved
FCM imag
e
segm
entatio
n algo
rithm
wa
s also pre
s
ente
d
. Final
ly, identifications
were utilize
d
in conne
cte
d
area
s to id
entify t
he insulator’
s outlin
e. The experimental re
sul
t
s
sho
w
e
d
that the improve
d
Wiene
r filteri
ng algo
ri
thm
may effectively filter and restore imag
e
s
,
and the im
proved FCM im
age
segm
ent
ation algo
rith
m
can
se
gme
n
t insul
a
tors from ima
g
e
s
a
nd
effectively reduce numb
e
r
of erro
r clu
s
te
r pixels. T
he
next step of this re
se
arch i
s
to improve t
he
segm
entation
time of the F
C
M algo
rithm
.
Ackn
o
w
l
e
dg
ement
This work wa
s su
ppo
rted in part by the Scientific an
d
Techn
o
logi
cal Planning P
r
oje
c
t of
Jilin Provin
ce (201
005
6
5
), the Scientific Re
se
arch Fund
of Jilin Pro
v
incial Edu
c
ation
(201
203
07), t
he Sci
entific
and T
e
chnolo
g
ical Pl
anni
n
g
Proje
c
t of
Jilin Provin
ce
(2012
0332
), t
he
Scientific
Re
sea
r
ch Fu
nd
of Jilin Provinci
al Ed
u
c
ation
(20
1
3
434)
and th
e Jilin
provi
n
ce
developm
ent
and
Reform Commi
ssion proje
c
ts (20
1
3
C0
48).
Referen
ces
[1]
Zhu H, Li WG, Lin Y. Present
and
future de
velo
pment of
d
e
tectio
n methods for com
posite insulator
.
Insulators a
nd
Surge Arrester
s
. 2006; 8(1): 1
33-1
37.
[2]
Qaddo
umi NN
, El-Hag AH,
Saker Y. Outdoor Insu
l
a
tors
T
e
sting Using
Artificial Ne
ur
al Net
w
o
r
k-
Based
Ne
ar-F
i
e
ld M
i
cro
w
a
v
e
T
e
chniqu
e.
IEEE Transactions on Instru
mentation and
Measur
em
ent
.
201
3; 63(2): 26
0-26
6.
[3]
Buxt
on B, Be
mard F
,
Abdal
l
ahi A, Ho
uari
F
R
,
Delmiro J.
Devel
opme
n
t of
an E
x
tensi
o
n of the Otsu
Algorit
hm for Multidime
n
si
ona
l Image
Segme
n
tatio
n
of
T
h
in-F
ilm Blood Sl
ide
s
.
International
Confer
ence on
Computi
ng:
T
heory a
nd Ap
pl
icatio
ns
. 200
7; 1(1): 552-
56
2.
[4]
Chen SY, Song SF, Li LX
.
Su
rve
y
on sm
art grid tec
hno
lo
g
y
.
Pow
e
r System T
e
ch
no
logy
.
200
9; 33(
8):
1-7.
[5]
Hassa
npo
ur R
,
Shah
ba
hram
i A, W
a
n
g
S.
A d
aptive
g
aussi
an
mixtu
r
e
mod
e
l
for
skin co
lor
.
Procee
din
g
s of
W
o
rld Acade
m
y
of Scie
nce,
Engin
eeri
ng a
nd T
e
chnol
og
y. 2008; 31: 1-6.
[6]
Lu GQ,
X
u
HG
, Li YB.
L
i
n
e
de
te
cti
o
n ba
se
d o
n
ch
ai
n co
de d
e
t
e
c
tio
n. IEE
E
Internati
o
n
a
l
Co
nferenc
e
on Veh
i
cul
a
r El
ectronics a
nd
Safet
y
. 20
05; 1
:
98-103.
[7]
Jou F
,
F
an KC
, Chan
g YL. Ef
ficient match
i
n
g
of l
a
rge-s
i
ze
histogr
ams.
Pattern Recogn
i
t
ion L
e
tters
.
200
4; 25(3): 27
7-28
6.
[8]
Roll
an
d JP, Vo V, Bloss B. Fast algor
ithms
for
histogram matchin
g
:
Appl
icat
io
n to textu
r
e s
y
nthesi
s
.
Journ
a
l of Elec
tronic Ima
g
i
n
g
.
2009; 9(1): 3
9
-
45.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 13, No. 1, March 2
015 : 164 – 1
7
2
172
[9]
Morovic J, Sh
a
w
J, Su
n PL.
A fast, non-it
erative
an
d e
x
act histogr
am
matchin
g
al
gor
ithm.
Pattern
Reco
gniti
on L
e
tters
. 2002; 23(
3): 127-1
35.
[10]
Jahn
e
B.
D
i
git
a
l I
m
ag
e Pr
oc
essin
g
C
once
p
t
s.
Algorith
m
s and Scie
nt
ific Appl
icatio
ns
. S
p
rin
ger, 1
997;
1: 236-2
42.
[11]
Liu
XH, Guo
C, An H. A modi
fi
ed
w
i
en
er filterin
g for restoration
of rin
g
-
code
d ap
ertur
e
imag
es in
inerti
al confi
n
e
m
ent fusion.
Acta Optica Sinica
. 2004; 2
4
(8)
:
1045-1
0
5
0
.
[12]
T
apas K, David M, Mount NS, Netan
y
a
h
u
CD, Piatko
RS, Angela
YW
. An Efficient k-Means
Clusteri
ng A
l
g
o
rithm: Ana
l
ys
is an
d Implem
entatio
n.
IEEE Transactions
on Patter
n
A
nalysis
and
Machi
ne Intell
i
genc
e
. 200
2; 24(7): 881-
89
2.
[13]
Bharat P,
Dur
ga T
.
Improv
e
d
K-Me
doi
ds
Clusteri
n
g
Bas
ed
on
Cluster
Vali
dity In
dex
an
d Obj
e
ct
Density
. IEEE 2nd Intern
atio
n
a
l Adva
nce C
o
mput
in
g Conf
e
r
ence. 20
10; 1:
379-3
84.
[14]
Dame
ng
D, D
e
ju
n M. A
F
a
st Appr
oac
h to
K-m
eans
Cl
usterin
g
for
T
i
me Series
Bas
ed
on S
y
m
bol
i
c
Repr
esentati
o
n
.
Internation
a
l
Journ
a
l of Adv
anc
e
m
ents in
Co
mp
uting T
e
chno
logy
. 20
12
; 4
(
5
)
: 2
33-
239.
[1
5
]
C
h
en
JJ, Song
A, Zh
an
g W. Hy
b
r
i
d
Cl
u
s
te
rin
g
Meth
ods
Base
d o
n
A
d
aptive
K-harm
onic
Mea
n
s.
Internatio
na
l Journ
a
l of Adva
nce
m
e
n
ts in C
o
mputi
ng T
e
ch
nol
ogy
. 20
12; 4(6): 10-2
3
.
[16]
Li Z
R
, Li
u Y, H
a
yw
a
r
d R.
Kn
o
w
ledge bas
ed pow
er
li
ne det
ection
for UAV surveil
l
a
n
ce an
d
ins
pecti
on
system
s
. Proc
eed
ings
of Internatio
nal C
onfe
r
ence o
n
Imag
e and Vis
i
o
n
C
o
mputi
ng, 20
0
8
; 1: 1-6.
[17]
Katrasnik J, Pernus F
,
Likar B. A surve
y
of
mobil
e
rob
o
ts for distributi
on
po
w
e
r lin
e insp
ection.
IEEE
Tra
n
s
a
c
ti
on
s on
Po
we
r D
e
l
i
v
ery
, 2010; 25(
1): 485-4
93.
[18]
Basu M. An improve
d
SUS
A
N edg
e dete
c
tion
scal
e
inv
a
ria
b
le featur
e
s
algorit
hm-a surve
y
.
IEEE
T
r
ansactio
n
s o
n
System
, Man
and C
y
b
e
rn
ati
cs. 2002; 32(
3)
: 252 -260.
[19]
Sarab
a
n
d
i M, Park M. Extrac
tion of po
w
e
r li
ne maps from
millim
eter-
w
av
e pol
arimetric
SAR imag
es.
IEEE Trans. A
n
tennas Propag
. 2000; 4
8
(2): 180
2–
180
9.
[20]
W
ang
X. An in
sulator o
n
li
ne d
e
tecting Meth
o
d
summar
y
.
P
o
rcelai
n Arrester
. 2002; 14(
6): 34-39.
[21]
Hua
ng
XN, Z
h
ang
Z
L
. An
e
x
tractio
n
alg
o
ri
thm for ins
u
l
a
tors Patro
l
h
e
li
copter
aeri
a
l
i
m
age.
Grid
techno
lo
gy
. 20
10; 34(1):1
94-
197.
[22]
Lin J
C
. A gl
as
s insu
lators
De
fect
Diag
nos
is
method
Base
d
on c
o
lor
ima
g
e
s.
Grid technology
. 20
11;
35(1): 12
7-1
3
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.