TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4200 ~ 4
2
0
5
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.142
2
4200
Re
cei
v
ed Se
ptem
ber 12, 2013; Revi
se
d De
ce
m
ber
8, 2013; Acce
pted Ja
nua
ry
6, 2014
Feature Extraction and Classification of Electric Power
Equipment Images Based on Corner Invariant Moments
Zhai Xueming*
1
, Zhang Dong
y
a
2
, De
w
e
n Wan
g
3
Dep
a
rtment of Contro
l and C
o
mputer Eng
i
n
e
e
rin
g
,North Ch
ina El
ectric Po
w
e
r Univ
ersit
y
,
No. 689 H
u
a
d
i
an Ro
ad, Bao
d
i
ng Cit
y, He
bei
Provinc
e
, Chin
a
*Corres
p
o
ndi
n
g
author, em
ail
:
zxm31
65@
12
6.com
1
, 7630
4
325
3@q
q
.com
2
, w
d
e
w
en@
gma
i
l
.
co
m
3
A
b
st
r
a
ct
F
eature extrac
tion a
nd acc
u
r
a
te classific
a
ti
on of
el
ectric
pow
er eq
uip
m
ent, hel
p to i
m
prove th
e
auto
m
ati
o
n
an
d i
n
tell
ige
n
t l
e
vel
of pow
er s
ystem
ma
n
age
me
nt. Ai
mi
ng
at the
pro
b
le
ms that a
pply
i
n
g
H
u
invari
ant
mo
ments to extract
imag
e
featur
e
co
mputes
lar
g
e an
d a
p
p
l
yin
g
corner v
e
ctor
to match h
a
s t
o
o
di
me
nsio
ns, this
pap
er
pre
s
ented
H
a
rris
corn
er
i
n
vari
ant mo
ments
al
gorith
m
.
T
h
is alg
o
rith
m onl
y
calcul
ates cor
n
er coord
i
n
a
tes other
tha
n
the
entire i
m
ag
e coord
i
nates, so
can cha
n
g
e
th
e poi
nt feature
into
feature v
e
ctors
,
and
re
duce
t
he c
o
rner
mat
c
hin
g
d
i
me
nsi
ons. C
o
mbi
n
e
d
w
i
th th
e SV
M (Sup
port V
e
ctor
Machi
ne) cl
ass
i
ficatio
n
meth
o
d
, w
e
cond
ucted a c
l
ass
i
ficat
i
on for
a l
a
rge
nu
mb
er of e
l
ec
trical e
qui
p
m
e
n
t
imag
es, and t
he resu
lt show
s that using H
a
rris corn
er
in
varia
n
t mo
men
t
s algorith
m
to
extract invari
an
t
mo
ments, and
classifyin
g by thes
e i
n
vari
ant
mo
ments can
achi
eve b
e
tter classificati
on a
ccuracy.
Ke
y
w
ords
: Hu
invari
ant mo
ments, Harris cor
ner
, feature ext
r
action, class
i
fi
cation
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
With the ra
pi
d develop
me
nt of intellige
n
t electri
c
g
r
i
d
con
s
tructio
n
and a
n
increa
sing
numbe
r of t
y
pes of
ele
c
tric p
o
wer
e
quipme
n
t, el
ectri
c
p
o
wer equip
m
ent
manag
eme
n
t is
developin
g
from the tra
d
itional text manag
em
ent t
o
multimedia
and intellig
ent inform
ation
manag
eme
n
t, thus leadi
ng
to greatly incre
a
sed workload an
d com
p
lexity of managem
ent. With
the popul
arit
y of digital camera, imag
e su
rveillan
c
e and oth
e
r
image
colle
ct
ing devices,
the
diso
rde
r
p
r
ob
lem of imag
e
informatio
n i
s
mo
re p
r
omi
nent. So it ha
s be
com
e
an
urge
nt probl
em
about
ho
w to
organi
ze
ord
e
rly tho
s
e
hu
ge n
u
mbe
r
a
nd va
riety of
electri
c
al
po
wer
equip
m
ent
in
orde
r to cla
ssify and browsi
ng qui
ckly [1].
The typical
image cla
ssifi
cation retrieval
syst
em con
s
i
s
ts of data acqui
sition,
prep
ro
ce
ssin
g, feature
ex
traction,
cla
s
sificatio
n
de
cision
and
cl
a
ssifie
r
[2], in
whi
c
h fe
atu
r
e
extraction i
s
a key fa
ctor i
n
the cl
assification. F
eatu
r
es of the i
m
a
ge cont
ent in
clud
e imag
e
colo
r,
texture, shap
e and othe
r e
x
terior features [3].
For m
o
st ele
c
tric p
o
we
r equi
pm
ent image
s, those
colo
r is relatively simple, an
d there a
r
e n
o
obv
iou
s
texture ch
ang
es
like wood an
d
stone surfa
c
e,
so
color and texture feature al
way
s
are used as
auxiliary features [4]. A variety of electri
c
al
power eq
uip
m
ent is different from ea
ch othe
r in sha
p
e
s
, and
for the sam
e
electri
c
po
we
r
equipm
ent, the
sha
pe fea
t
ure h
a
s tra
n
s
lation,
rotati
on a
nd
scalin
g invari
ability. Therefore, t
h
is
pape
r focu
se
s mainly on shape featu
r
e
of electri
c
al
p
o
we
r equi
pm
ent as the obj
ect of re
sea
r
ch.
Dome
stic an
d foreign
ex
perts a
nd
scholars
have
done
a l
o
t of
re
se
arch
wo
rk on
the i
m
age
sha
pe fe
ature
extra
c
tion.
Hu.M.K con
d
u
c
ted seven
sh
ape i
n
varia
n
t
moment
s by
usin
g al
geb
ra
ic
invariant
mo
ment in
19
62
[5]. The
simpl
e
seven fe
atu
r
e ve
ctors
ca
n de
scrib
e
an
imag
e, but t
he
cal
c
ulatio
n is
relative la
rge.
Literatu
re [6]
applie
d Hu i
n
variant m
o
m
ents to
cla
ssi
fy and retri
e
val
electri
c
al p
o
wer equi
pme
n
t image
s, but not di
scusse
d the com
p
lexity of the algorithm.
Harri
s
co
rn
er
detectio
n
met
hod is
an effe
ctive mean
s t
o
gre
a
tly com
p
re
ss i
m
age f
eature
[7]. In an image,
corne
r
poi
nts a
r
e
the greate
s
t chan
ged
pi
xels in lo
cal
gray, ge
nerally
accou
n
ting fo
r only a
bout
0.05% of the
image pixel
s
.
The
co
rne
r
p
o
ints h
a
ve m
u
ch i
n
form
ation
and rotatio
n
invariant featu
r
es, al
so be a
b
le to adapt to ambient lig
ht chan
ge
s [8].
This p
ape
r p
r
ese
n
ts
Harri
s
corne
r
invari
ant
mome
nt algorith
m
. First, acco
rding
to Harri
s
corne
r
dete
c
ti
on op
erator,
detect im
age
co
rne
r
in
fo
rmation,
then record
the co
rne
r
coo
r
din
a
tes
and the
gray
value. Seco
n
d
, cal
c
ulate t
he six
Harr
i
s
corne
r
invari
a
n
t moment v
e
ctors
by Ha
rris
corne
r
inva
ri
ant mom
ent
algorith
m
. Th
is al
go
rith
m chang
ed point
features
into
feature ve
ctors,
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation of
Elec
tri
c
Powe
r Equipm
ent Im
ages… (Zh
a
i Xuem
ing)
4201
taking into
accou
n
t both th
e key po
sitio
n
of the
imag
e informatio
n, but also g
r
e
a
tly redu
ce
s the
amount of co
mputation. Th
e exper
im
ent sho
w
s that co
mbined
with
SVM classification algo
rith
m,
usin
g the six
Harris
corne
r
invaria
n
t moment ve
cto
r
s as featu
r
e
extracti
on v
e
ct
or
s t
o
cla
s
s
i
fy
image is effe
ctive.
2. Harris Co
r
n
er Inv
a
riant Moment Fea
t
ures
2.1. Hu In
v
a
riant Momen
t
s
Hu inva
riant
moment
s a
r
e
seven
sh
ap
e inva
ria
n
t m
o
ments that
Hu.M.K co
nd
ucted
by
usin
g alge
bra
i
c invari
ant m
o
ment in 1
9
6
2
. In x-y plan
e, an MxN im
age
f
(
x,
y
), the (p
+q
)-th o
r
der
moment
s and
central m
o
m
ents are defin
ed re
spe
c
tive
ly as:
D
q
p
pq
dxdy
y
x
f
y
x
m
)
,
(
(1)
D
q
p
pq
dxdy
y
x
f
y
y
x
x
)
,
(
)
(
)
(
(2)
Whe
r
e
00
10
/
m
m
x
，
00
01
/
m
m
y
, the
s
e
are the
ce
ntroid
of the
i
m
age.
Wh
en
f
(
x,
y
) i
s
the obj
ect
de
nsity, the
ze
ro-o
rde
r
m
o
m
ents
m
00
are t
he
sum
of th
e
den
sity, that
is al
so
the
m
a
ss
of the object.
The cent
ral moment
s
pq
are the metri
cs that mea
s
ure
the ba
rycenter di
strib
u
tion
in regio
n
R
. T
he normali
ze
d central mo
ments, den
oted by
pq
, are
defined a
s
:
2
/
)
2
(
00
q
p
pq
pq
(3)
Hu
co
ndu
cte
d
seven i
n
variant m
o
ment
s’ fu
n
c
tion
s
by usi
ng th
e
normali
zed
se
con
d
-
orde
r a
nd thi
r
d-ord
e
r
ce
ntral mo
ment
s. The
s
e
se
ve
n functio
n
s h
a
ve tran
slatio
n, rotation
a
n
d
scaling invari
ability.
02
20
1
2
11
2
02
20
2
4
)
(
2
21
03
2
12
30
3
)
3
(
)
3
(
2
03
21
2
12
30
4
)
(
)
(
2
03
21
2
12
30
03
21
03
21
2
03
21
2
12
30
12
30
12
30
5
)
(
)
(
3
)
)(
3
(
)
(
3
)
(
)
)(
3
(
(4)
)
)(
(
4
)
(
)
(
)
(
21
03
12
30
11
2
21
03
2
12
30
02
20
6
2
21
03
2
12
30
03
21
30
12
2
21
03
2
12
30
12
30
03
21
7
)
(
)
(
3
)
)(
3
(
)
(
3
)
(
)
)(
3
(
2.2. Harris Corner Inv
a
ria
n
t Momen
t
Algorithm
It’s a relative
effective poin
t
feature
extracti
on
algo
rit
h
m by u
s
ing
Harri
s
op
erator
whi
c
h
only use
s
th
e first-ord
e
r
gray differe
n
c
e an
d Ga
us
sian filteri
ng.
Thus it ha
s
spe
ed calcul
ating
and
strong
timeline
ss.
In
addition,
Ha
rris corn
er
po
ints a
r
e
not
sen
s
itive to
image
rotatio
n
,
transl
a
tion a
nd gray tran
sform
a
tion, so the Harr
is
corne
r
poi
nts also
have th
e advanta
g
e
s
of
high stability, simple
operation
a
nd anti-i
n
terference ability.
It’s
co
nsidered corner poi
nts that t
h
e
pixels
cha
n
g
ed g
r
eate
s
t i
n
local g
r
ay. Ha
rri
s
corn
er d
e
tectio
n
operator
alg
o
rithm
wo
rks as
follows: this
method u
s
e
s
a recta
ngul
a
r
win
d
o
w
or
a Gau
ssi
an
wind
ow to m
o
ve on the i
m
age,
and then we can g
e
t the derived pa
rtial 2
2 st
ru
cture
M from the original templat
e
windo
w. Ne
xt,
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4200 – 4
205
4202
cal
c
ulate th
e eigenvalu
e
s
1
and
2
of
M
. Acco
rdin
g
to the defined
corne
r
re
sp
o
n
se fu
nctio
n
R
in
formula
(5
), calcul
ate the
R
value
of ea
ch pixel. Fi
na
lly, we sele
ct a se
rie
s
of
co
rne
r
coo
r
dina
te
s
by using lo
cal
non-m
a
xima sup
p
re
ssion t
o
get the app
rop
r
iate thre
shold.
B
C
C
A
I
w
I
I
w
I
I
w
I
w
M
y
v
u
y
x
v
u
y
x
v
u
x
v
u
2
,
,
,
2
,
2
2
2
)
(
)
(
)
det(
B
A
k
C
AB
C
ktr
M
R
(5)
2
1
)
(
C
tr
Whe
r
e
I
x
and
I
y
are ze
ro order g
r
ay gra
d
i
ent.
Assu
me we get
n co
rne
r
coo
r
din
a
tes by
Ha
rri
s corner dete
c
ting
ope
rator. Th
ese
are
recorded a
s
seri
es (
x
1
, y
1
)
,
(
x
2
, y
2
)…(
x
n
, y
n
), and co
rn
er co
ordi
nate
gray values
are de
noted
by
f
(
i
，
j
).
Then a
c
cordi
ng to the formul
a (1) and fo
rmu
l
a (2), we de
fine the discrete co
rne
r
order
moment
s and
central m
o
m
ents a
s
follows:
D
q
p
pq
dxdy
y
x
f
y
x
m
)
,
(
(6)
D
q
p
pq
dxdy
y
x
f
y
y
x
x
)
,
(
)
(
)
(
(7)
Literatu
re [9]
did furth
e
r re
sea
r
ch of
Hu
invariant
mo
ments metho
d
, and
p
r
ove
d
that
Hu
invariant m
o
ments
only h
a
ve tran
slatio
n and
rotati
o
n
invariability.
Next, we
discuss the influ
e
nce
of scal
e
and
contrast o
n
the co
rne
r
inv
a
riant
mo
me
nts. The scali
ng facto
r
is d
enoted by
, and
the cont
ra
st cha
nge fa
cto
r
is
k
. And th
e positio
n is
cha
nge fro
m
(
i
,
j
) to(
i’
,
j’
) , accordi
ngly, the
grad
ation
f
(
i
,
j
) is converted into
f
(
i’
,
j’
) from the
same lo
catio
n
. There are
the followin
g
relat
i
on
shi
p
s:
0
j
i
'
j
'
i
(8)
0
k
)
j
,
i
(
kf
)
'
j
,
'
i
(
'
f
The origi
nal cente
r
m
o
me
nt
is
pq
, afte
r the trans
f
ormation is
’
pq
. And the
coo
r
dinate
s
of the ce
nter of grav
ity are re
spe
c
tivel
y
(
i
,
j
)an
d
(
i
,
j
). Th
en
the relatio
n
ship bet
wee
n
the
transfo
rme
d
center mom
e
n
t
and the origi
nal
ce
nter mo
ment is shown as form
ula (9):
D
pq
q
p
q
p
q
p
D
q
p
pq
k
j
i
f
j
j
i
i
k
j
i
f
j
j
i
i
)
,
(
)
(
)
(
)
,
(
)
(
)
(
(9)
Particul
arly
’
00
=
k
’
00
，
and the normali
zed ce
ntral mo
ments a
r
e def
ined a
s
:
pq
q
p
q
p
r
pq
q
p
r
pq
pq
k
k
k
2
/
)
(
00
00
)
(
)
(
(10)
Combi
ned
wi
th formul
a
(3
) a
nd
(4
), we can
ea
sily get th
e rela
tionshi
p of i
n
variant
moment
s bet
wee
n
the orig
inal and tra
n
sformed a
s
formula (1
1).
1
2
1
k
，
2
2
4
2
k
，
3
3
6
3
k
，
4
3
6
4
k
，
5
6
12
5
k
，
6
4
8
6
k
，
7
6
12
7
k
(11)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation of
Elec
tri
c
Powe
r Equipm
ent Im
ages… (Zh
a
i Xuem
ing)
4203
From th
e fo
rmula
(11
)
, we can
see
th
at afte
r the
transfo
rmatio
n
of scale
an
d
co
ntra
st,
the invariant
moments a
r
e the (
2
/k
) integer po
wer
com
pare
d
with the origin
al invariant
moment
s, no
longer remai
n
ing invari
ant
features. We reg
r
oup th
e above form
ulas in o
r
de
r to
eliminate
th
e influen
ce of
scalin
g
fa
ctor and co
ntra
st f
a
ctor.
So
we
get the
six i
n
variant
mome
n
t
s
vec
t
ors
as
follows
:
2
1
2
1
，
2
1
3
2
3
4
3
，
4
5
4
，
4
1
6
5
，
5
7
6
(12)
Specific al
go
rithm step
s:
(A)
Fi
rst, we sho
u
ld
m
a
ke the
colle
cted image
s
size
norm
a
lizi
ng a
nd g
r
aying.
Next, filter
the prep
ro
ce
ssed ima
g
e
s
b
y
using difference ope
rato
rs, an
d cal
c
ul
ate
I
x
,
I
y
and
I
xy
. Then use t
h
e
5x5 Gau
ssi
an
templates to
smooth the i
m
age, after removing noi
se, we ca
n get
M
.
(B) Cal
c
ul
ate
the corne
r
resp
on
se function
R
of the
corre
s
p
ondin
g
pixels acco
rding to
M
, where
R=AB-C
2
-K(A
+B
)
2
. Then sel
e
ct a se
rie
s
o
f
corne
r
coordinate
s
by using lo
cal n
o
n
-
maxima
sup
p
r
essio
n
to
get
the a
p
p
r
op
ri
ate thre
shold.
Re
co
rd
the corne
r
co
ordi
n
a
tes as
(
x
1
, y
1
),
(
x
2
, y
2
)…(
x
n
, y
n
), and the
correspon
ding
gray value
s
, whe
r
e corner
numbe
r is
n
.
(C) Cal
c
ulate
the corn
er p
o
ints orde
r moments
m
pq
and ce
nter mo
ments
pq
accor
d
ing
to the formula
(6) an
d (7
), whe
r
e
i
=1,2,…,
n
.
(D) No
rmali
z
e the above
corne
r
point
s cente
r
mom
ents a
c
cordi
n
g to the formula (3
),
and get
pq
. Then calculate
the seven co
rne
r
point
s in
variant mome
nt vectors
1
-
7
.
(E) In actual
process, sin
c
e the seven
in
variant mo
ment vectors vary wide, we u
s
e
*
i
=|log|
i
||. Then we cal
c
ulate the six
corner p
o
in
t
invariant m
o
ments ve
ct
ors
acco
rdin
g to
formula (12), whi
c
h are all have translat
i
on, rotation and scaling invariability.
3. Experiment
3.1. Image Choosing Pre
p
roces
sing
In this experi
m
ent, we ch
o
o
se five types of
electri
c
p
o
we
r equi
pm
ent shot in the factory
environ
ment
as
sho
w
n i
n
Figure 1. Fro
m
left to
righ
t the five image
s are tra
n
sformer,
circuit
brea
ke
r, en
ergy meter, swi
t
ch an
d curre
n
t transfo
rme
r
. We u
s
e
18
0 image
s a
s
t
o
tal sam
p
le
s, in
whi
c
h
90 i
m
a
ges a
r
e
used
for SVM
trai
ning
and
th
e
othe
rs are
u
s
ed
for testin
g cl
assificatio
n
accuracy. Th
e expe
riment
al environme
n
t is ma
tla
b
R2008,
com
b
in
ed with
lib
svm-3-17
software
packa
ge for
SVM training
and testin
g.
Figure 1. Images u
s
e
d
in the Experime
n
t
Harri
s
co
rne
r
invariant mo
ment algo
rith
m and SVM classificatio
n
a
r
e sh
own in F
i
gure 2.
Figure 2. Experime
n
tal Pro
c
ed
ure
Digital images
Preprocessin
g
SVM
model
training
SVM
cla
ssif
i
cat
i
o
n
Output
Har
r
i
s
cor
n
e
r
invariant
moment
s
algorith
m
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4200 – 4
205
4204
Image
pre
p
ro
ce
ssi
ng
ope
rations in
clud
e
no
rmali
z
ing
i
m
age
si
ze,
graying a
nd
en
han
cing
image contra
st with histog
ram equ
alization method.
First, dete
c
t co
rne
r
coo
r
di
nates by Ha
rris
detectin
g
ope
rator. Th
en, calcul
ate Ha
rri
s co
rn
er inva
riant mome
nt
vectors
1
-
6
, and put these
six vectors as extracted feature
vectors. In this paper, we choose LIBSVM classifier for traini
ng
and
cla
ssifi
cation. Acco
rd
ing to the
training
sa
mpl
e
s,
we train
these
sampl
e
s, u
s
in
g cross-
validation m
e
thod to o
b
tai
n
the o
p
timal
paramete
r
s
g
and
c
, thus
, getting the trained mode
l
.
Next, input th
e sa
mple
s to
be teste
d
to t
he trai
ned SV
M model, a
n
d
re
cord the m
i
scl
assificatio
n
numbe
r an
d classificatio
n
a
c
cura
cy.
3.2. Experimental Data
Acquisition
After prep
ro
cessing to the
colle
cted im
age
s and det
ecting
Harri
s
corn
er
coo
r
dinate
s
,
image
s
sho
w
n in
Figu
re
3.
In the
p
r
oce
s
s of
sel
e
ctin
g Harris corn
er
co
ordi
nate
s
, pa
ram
e
ter
k
is
0.04. Then we cal
c
ulate
d
Harri
s
co
rne
r
invariant mo
ment vectors
1
-
6
, a part
of the training
sampl
e
s d
a
ta
sho
w
n in Ta
ble 1.
Figure 3. Images afte
r the Corne
r
Dete
ction
Table 1. Part
of the Trainin
g
Data
Image class
1
2
3
4
5
6
Transform
er
0.0940
0.0993
1.0166
0.1727
0.0670
1.0050
Breaker
0.0902
0.0975
1.0008
0.1812
0.0672
0.9942
Energ
y
Meter
0.0931
0.0848
1.0000
0.1907
0.0715
1.0445
Sw
itch
0.0927
0.0987
1.0147
0.1772
0.0670
0.9999
CT
0.0987
0.0949
1.0186
0.1806
0.0716
1.0252
Acco
rdi
ng to
SVM algorithm Vpani
k prop
os
ed [1
0], we sele
cted RBF(radi
al basi
s
function
) a
s
kernel fu
nctio
n
, and
dete
r
mined th
e
op
timal pa
rame
ters (
g,
c
)=(2,0.0625
) by
u
s
ing
cro
s
s
-validat
ion al
gorith
m
in training
[11], wh
ere
g
is th
e
width
paramete
r
o
f
RBF,
c
is t
he
penalty factor.
3.3. Results and An
aly
s
is
Table 2. Cla
s
sificatio
n
Re
sults by Harri
s
Corne
r
Invari
ant Moment
Vectors
Image class
Training samples
Test samples
Miscla
ssif
i
cation
number
Accuracy(%
)
Transform
er
19
17
3
82.35
Breaker
18
18
2
88.89
Energ
y
Meter
20
16
2
87.50
Sw
itch
17
19
2
89.47
CT
16
20
2
90.00
Figure 4. Co
mpari
s
o
n
of Cla
ssifi
cation
Accu
ra
cy
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion an
d Cla
ssifi
cation of
Elec
tri
c
Powe
r Equipm
ent Im
ages… (Zh
a
i Xuem
ing)
4205
Use Ha
rri
s
co
rne
r
invari
ant
moment vect
ors
1
-
6
to train and te
st. The re
sult di
splayed
of wrong n
u
m
bers a
nd
cl
assificatio
n
a
c
cura
cy, a
s
s
hown in Ta
bl
e 2. Similarly
,
we al
so
cla
ssify
these five t
y
pes ima
g
e
s
based on
seven
Hu i
n
variant m
o
ments. In o
r
de
r to facili
tate
comp
ari
s
o
n
, the cla
s
sificati
on re
sult and
comp
ari
s
o
n
a
r
e sh
own in F
i
gure 4.
From the a
bove experi
m
ental re
sult
s, t
he classi
fication accu
racy ha
s re
ach
ed a
relatively
satisfied status. But
there also
ex
ists som
e
misj
udgm
e
n
t, the rea
s
o
n
s le
d to the
s
e
justice inclu
d
e
that these image
s are m
o
stly shot fro
m
fact
ory environm
ent. The equipm
ent has
different deg
rees of wear,
con
s
um
ption,
and other
di
ssi
pation in t
he enviro
n
m
ents, whi
c
h a
l
so
affect the ima
ge feature ext
r
actio
n
to so
me extent.
First, we get the co
rne
r
poi
nts dete
c
ted by Ha
rri
s corner op
erato
r
, then acco
rdin
g to the
Harri
s
corner invaria
n
t m
o
ment al
gorit
hm, ch
ang
e
corne
r
poi
nts to the six f
eature
vecto
r
s.
Formul
a evid
ence that the six f
eature vectors have t
r
an
slation,
rotation and
sca
ling invaria
b
ili
ty.
Next, c
o
mbined
with SVM c
l
ass
i
fic
a
tion, us
e the
ex
trac
ted s
i
x feature vec
t
ors to c
l
ass
i
fy thos
e
electri
c
po
we
r equip
m
ent
image
s. The
experime
n
t
result sho
w
s that Harri
s
corne
r
invari
ant
moment algo
rithm can
be
use
d
a
s
sh
ap
e
featu
r
es
to
descri
be
an i
m
age fe
ature
.
Furthe
r,
Harris
corne
r
inva
ri
ant mom
ents are
calculate
d
only o
n
the
co
rne
r
p
o
int
s
of the
targe
t
so
can
grea
tly
redu
ce extraction spe
ed an
d data pro
c
e
s
sing.
4. Conclusio
n
In the shooti
ng pro
c
e
s
s, electri
c
al p
o
w
er
e
quip
m
e
n
t images a
r
e easily affe
cted by
sho
o
ting angl
es, distan
ce
and othe
r factors, whil
e the Harris
corn
er and Hu in
variant mome
nts
both have translation, rot
a
tion a
nd
scali
ng invariabilit
y. Therefore,
we
unified the corner feat
ure
and i
n
varia
n
t
moment
s, ch
angin
g
the
p
o
int featu
r
e
i
n
to featu
r
e v
e
ctors. T
h
is a
l
gorithm
sele
cted
six Harris
corner invari
ant
moment ve
ct
ors a
s
ex
tra
c
ted featu
r
e
for the
next ima
ge
cla
ssifi
cati
on.
The expe
rim
ent sh
ows it
can
achi
eve
high
cla
ssifi
cation a
c
cura
cy. So it’s fast and fea
s
ibl
e
that
usin
g Ha
rri
s corne
r
invaria
n
t moments a
l
gorithm to
ex
tract imag
e feature, with th
e advantag
e of
s
h
ort time and c
o
mplexity.
Ackn
o
w
l
e
dg
ements
This
wo
rk i
s
supp
orte
d
by Natio
n
a
l
Na
tu
ral S
c
ience Fo
und
ation of
Chi
na (No.
6107
4078
).
Referen
ces
[1]
D Lv, P Yu. Information
Mana
geme
n
t S
y
stem of S
ubstatio
n
Eq
u
i
pme
n
t base
d
on ima
g
e
.
T
e
leco
mmunic
a
tions for Elect
r
ic Pow
e
r System
. 2
007; 2
8
(1
74):61-
65.
[2]
P.Bi. A Comp
ariso
n
of Ap
pr
oach
e
s for Image
Class
ificat
ion.
Mo
dern El
ectronic
T
e
c
h
n
o
lo
gy
. 20
09;
18,59-
64.
[3]
P.Y.Liu. Stud
y on the Algor
it
hms for Con
t
ent-base
d
Image F
eatur
e Extractio
n
/Dis
e
r
tation. M.S.
T
hesis. Chang
chun:Ji
lin U
n
iv
ersit
y
;2
00
5.
[4]
G.Q. Yang, R.Y. Cui. Method for Natur
a
l
Im
age Discrim
ination
based
on T
e
x
t
ure
Featur
e.
Appl
icati
o
n
Rese
arch of C
o
mputers
. 20
1
0
; 27(7): 27
83-
278
5.
[5]
HU
MK
．．
Visu
a
l
patter
n
rec
o
gniti
on
b
y
mo
ment i
n
vari
a
IEEE Transacti
ons
on Infor
m
ation
Theory
.
196
2; 8(2): 179
-187.
[6]
D Lv.
Res
e
a
r
ch a
n
d
Impl
ementati
o
n
of
Electric
Po
w
e
r E
qui
pme
n
t Pictures
Mana
geme
n
t
S
y
stem/Dis
erta
tion MS.
T
hesi
s
. Baodin
g
:Nor
th Chin
a Electri
c
Po
w
e
r Un
iver
sit
y
; 20
06.
[7]
Harris CG, Stephens MJ.
A combi
ned c
o
rn
e
r
and
ed
ge d
e
t
e
ctor
. Procee
di
ngs of th
e 4th
Alve
y Vis
i
o
n
Confer
ence. 1
988; 14
7-1
51.
[8]
QW
Jing, W
R
Li. A N
e
w
W
a
y
of Parts Shap
e R
e
co
gniti
on B
a
sed
on H
a
rris C
o
rner D
e
tectio
n
.
Mi
cro
c
om
pu
te
r
Inform
ation
. 2
010; 26(
6): 182
-184.
[9]
MR Din
g, JL C
han
g, JX P
eng
. Re
search on invari
ant
mom
ents
al
gorithm.
Journ
a
l of Dat
a
Acqu
isitio
n
and Proc
essin
g
. 1992; 7(
1): 1-9.
[10]
Vapn
ik VN. T
h
e nature of stat
istical
le
arn
i
ng
theor
y.Ne
w
Y
o
rk:Spring
e
r-Ver
lag; 19
95.
[11]
GH F
eng.
Par
a
meter Optim
i
zing
fo
r S
upp
o
r
t Vector Mac
h
in
e C
l
assific
a
tion.
C
o
mput
er
Eng
i
n
eeri
n
g
and Ap
plic
atio
n
. 2011; 4
7
(3): 123-
125.
Evaluation Warning : The document was created with Spire.PDF for Python.