TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3887 ~ 38
9
3
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.4446
3887
Re
cei
v
ed Se
ptem
ber 14, 2013; Revi
se
d No
vem
ber
13, 2013; Accepted Decem
ber 5, 201
3
The Contourlet Transform with Multiple Cycles
Spinning for Ca
tenary Image Denoising
Chan
gdong Wu*
1
,
2
,
Zhig
a
ng Liu
1
,
Hu
a Jiang
3
1
School of Elec
trical Eng
i
ne
eri
ng, South
w
e
s
t Jiaoto
ng U
n
ive
r
sit
y
, Che
n
g
du,
6100
31, Ch
ina
2
School of Elec
trical an
d Information En
gi
ne
erin
g, Xi
hu
a Universit
y
, Ch
en
gdu, 61
00
39, Chin
a
3
School of Co
mputer an
d Co
mmunicati
on E
ngi
neer
in
g, T
h
e e-mei C
a
mp
us of South
w
e
s
t Jiaotong
Univers
i
t
y
, E-m
e
i 61
42
02, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: superi
n_2
00
2@1
26.com
A
b
st
r
a
ct
In the catenary
ima
ges, no
ise
and artifacts a
r
e in
trod
uced
d
ue to the acq
u
i
s
ition tech
niq
u
e
s and
systems, w
h
ic
h may infl
uenc
e the ju
dg
e
m
e
n
t of catenar
y
qua
ilty an
d w
o
rking states. In this pa
per, th
e
contour
let tran
sform (
C
T
)
w
i
th p
e
rf
or
manc
e
s
of
multi-sc
al
e, multi-re
s
o
l
u
tion
an
d a
n
is
otropy is
pr
opos
ed
,
w
h
ich can
be
effectively a
p
p
l
ied to
i
m
a
ge
den
oisi
ng. H
o
w
e
ver, the CT
hasn
’
t t
he sp
i
nni
ng i
n
var
i
an
ce,
w
h
ich w
ill
lea
d
to th
e Gib
b
s
-like
ph
eno
mena. In
this
p
aper, th
e CT
w
i
th mu
ltipl
e
cycle sp
in
nin
g
is
introd
uced for
image
den
ois
i
ng, w
h
ich ca
n
effectively e
l
i
m
i
nate th
e vis
ual artifacts d
ue to the l
a
ck
of
translati
ona
l i
n
varia
n
ce. Mea
n
w
h
ile, the
diff
erent L
apl
aci
a
n pyra
mid (LP)
filters an
d
dir
e
ctiona
l filter b
a
nks
(DFB) are proposed to test noi
sy images
. Finally, test
the influ
ence
of cycle spinning numbers
for
denoising
effects by using
different cycle spinni
ng tim
e
s. T
he ex
periment
result
s show that
the propos
ed
meth
od h
a
s e
xcelle
nt de
nois
i
ng p
e
rfor
ma
n
c
e in ter
m
s of
the sign
al-to-
nois
e
ratio (S
NR) an
d the v
i
sual
effects, w
h
ich is also sup
e
rior
to some oth
e
r
existing
meth
ods in ov
erco
mi
ng the Gib
b
s
-like p
hen
o
m
ena,
backgr
oun
d s
m
o
o
thi
ng a
nd
preserv
a
tion of
edge sh
arp
n
e
ss and texture.
Ke
y
w
ords
: co
ntourl
e
t transfo
rm, spin
ni
ng in
varia
n
ce, caten
a
ry imag
e den
oisin
g
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
Caten
a
ry fau
l
t diagno
sis
unde
r ba
ckground n
o
ise i
s
a
challe
ngi
ng task. Duri
ng the
caten
a
ry fa
ult diagn
osi
s
b
a
se
d o
n
im
a
ge p
r
o
c
e
ssi
n
g
, imag
e q
u
a
lity is
cruci
a
l. Ho
wever,
the
caten
a
ry ima
ges
are ea
sy
to be influ
e
n
c
ed
by w
eat
h
e
r,
light
sou
r
ce,
sy
st
em a
nd
othe
r
fa
ct
ors,
whi
c
h
will a
p
pear noi
se. I
n
orde
r to a
c
curately an
al
yze the
fault
type, it is imp
o
rtant to
den
oise
the cate
nary i
m
age
s. Imag
e denoi
sin
g
i
s
to keep the
useful info
rm
ation and
re
d
u
ce th
e noi
se
of
the imag
e [1]
.
Usually, the
r
e
are
two
m
a
in meth
od
s f
o
r m
u
lti-fo
cus image
de
noi
sing,
One
is the
spatial
dom
ai
n-ba
se
d met
hod
s, an
othe
r is t
he tran
sform
domai
n
based m
e
tho
d
s. Amo
ng th
ese
method
s, the
wavelet tra
n
s
form
(WT) h
a
s o
b
tained
good
re
sults
in handli
ng o
ne-di
men
s
ion
a
l
piecewi
s
e
sm
ooth fun
c
tion
s [2]. However, the
effi
c
i
en
c
y
an
d
s
p
ars
e
r
e
pr
es
en
ta
tio
n
o
f
W
T
a
r
e
limited by sp
atial isotropy
of t
he ba
sis f
unctio
n
s
built only along
with the finite dire
ction
s
, which
has low freq
u
ency
re
soluti
on an
d hi
gh t
i
me re
sol
u
tio
n
in the
high
freque
ncy
pa
rts [3]. Be
sid
e
s,
the efficient i
m
age
rep
r
e
s
entation h
a
s to accou
n
t for the
geom
etrical
struct
ure
perva
siv
e
in
natural
scene
s [4]. A
s
the
caten
a
ry im
a
ges cont
ain
much
texture
feature
s
, wav
e
lets may
n
o
t
be
the be
st spa
r
se represent
ation,
whi
c
h
has th
e goo
d
denoi
sing
re
sults,
while th
e pre
s
e
r
vatio
n
of
edge d
e
tails i
s
n’t ideal, the
artifacts m
a
rked by
softne
ss, ri
ngin
g
s,
halo
s
, and
co
lor ble
eding
can
be se
en alo
ng edg
es.
Rece
ntly, many mutli-sc
ale
transfo
rm
s are p
r
op
ose
d
. Duri
ng th
ese
transfo
rm
s, t
he
conto
u
rle
t
tran
sform
(CT) inh
e
r
i
ts th
e
mu
lti-
res
o
lu
tion
an
d time
-
f
re
q
uen
c
y
locali
zation
p
e
rform
a
n
c
e
s
of WT. It pro
v
ides mo
re
sparse
rep
r
e
s
entation at b
o
th sp
atial a
nd
dire
ctional re
solutio
n
s whi
c
h can efficie
n
tly repre
s
en
t
images cont
aining contou
rs an
d textures
[5].
Compa
r
e
d
with
WT, the CT
ca
n p
r
ovide m
o
re
useful perfo
rmances, whi
c
h can effect
ively
captu
r
e the i
m
porta
nt information of catenary ima
g
e
s
.
The CT ha
s the rem
a
rkabl
e virtues in d
enoi
sing two
-
dimen
s
ion
a
l image
s. Ho
we
ver, due
to do
wn-sa
m
pling a
nd
up
-sampli
ng, the
CT
ha
s the
perfo
rman
ce
s of shift-varia
n
t [2, 6], whi
c
h
will lead to G
i
bbs-like phe
nomen
a alo
n
g
with re
movi
ng imag
e noi
se. To ove
r
come the Gi
b
b
s-
like p
hen
om
ena, ma
ny method
s a
r
e
introd
uced
in
so
me p
a
pers. In [7], interval
wa
velet
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3887 – 38
93
3888
transfo
rm i
s
a
pplied, which can g
ene
rate
spa
r
ser
re
p
r
e
s
entatio
ns in
the vicinity of discontin
uities
than cl
assi
cal
WT. However, it is
still b
u
ilt on
the
WT. In [8], Wavelet based o
n
co
ntourl
e
t is a
non-re
dund
a
n
t spa
r
se
rep
r
e
s
entatio
n of image
, which ca
n effectively reflect visual
cha
r
a
c
teri
stics of the i
m
ag
e, while it
s m
a
in pu
rp
o
s
e
i
s
to redu
ce t
he redu
ndan
cy; the tran
sl
ation
varian
ce i
s
still exist. Recently, the nonsu
b
samp
le
d
contou
rlet transfo
rm an
d
the transl
a
tion
invariant
co
ntourlet tran
sfo
r
m a
r
e int
r
od
uce
d
,
whi
c
h
have de
si
rabl
e feature in
many imagi
n
g
appli
c
ation
s
such a
s
pattern recognitio
n
, edge dete
c
tion and ima
g
e
denoi
sing,
while all of them
are tra
n
slatio
n invariant at
the cost of high red
und
an
cy [2]. In
[9],
R Eslami an
d
H Radh
a ha
ve
achi
eved go
o
d
denoi
sin
g
result
s by em
ploying t
he
CT with cy
cle
spin
ning to d
enoi
sing. Whi
l
e
the different
Lapla
c
ian
pyramid (LP
)
filters
and di
re
ctional filter ba
nks (DFB) m
a
y have different
impact
s
on
d
enoi
sing effe
cts. Be
side
s, the diffe
rent
cycle
spi
nnin
g
numb
e
rs m
a
y influence the
image
den
oising effe
cts. In
this pa
per,
the
CT
wi
th
m
u
ltiple
cycle
spinnin
g
meth
od i
s
int
r
od
uce
d
to denoi
se.
Mean
while, t
o
test the i
n
fluen
ce of
d
e
noisi
ng effe
cts, the differe
nt cycle
spi
n
ning
times an
d L
P
filters an
d
DFB are e
m
ployed to
d
enoi
se. Expe
rimental
re
su
lts sh
ow th
at the
prop
osed met
hod can effectively apply in image den
oi
sing a
nd protect the image
texture details.
2. The Propo
sed Me
thod
2.1. The Con
t
ourle
t Trans
f
orm
The CT i
s
a
kind of multi-scale g
eomet
ri
c anal
y
s
is too
l
, which
can t
a
ke full adva
n
tage of
the geom
etri
c re
gula
r
ity of image int
r
insi
c
st
ru
ctu
r
es
and
obt
ain the a
s
ymptotic opti
m
al
rep
r
e
s
entatio
n. It ha
s the
p
e
rform
a
n
c
e
s
of multi-
re
sol
u
tion, lo
cal
a
nd m
u
lti-di
re
ctional, which
can
effectively re
pre
s
ent
smo
o
th cu
rvature details
typ
i
cal of the catenary ima
g
e
s. The
CT
is
con
s
tru
c
ted
by combi
n
ing
the LP with
DFB,
whi
c
h i
n
clu
d
e
s
the
sub
-
ba
nd d
e
c
omp
o
sitio
n
and
dire
ctional tra
n
sform. Firstly,
the LP transform ite
r
atively deco
m
po
se
s a two
-
di
mensi
onal im
ag
e
into low-pa
ss
and hi
gh-pa
ss sub-ban
ds,
and the
n
t
he same scale p
o
int
disco
n
tin
u
ities
a
r
e
lin
ked
into linea
r structure by DF
B [10], the DFB are
a
ppli
ed to the hig
h
-pa
s
s sub-b
and
s to furth
e
r
decompo
se t
he freq
uen
cy spe
c
tru
m
into
dire
ctional
sub-b
and
s [11]
, which i
s
ea
sily adjustabl
e
to
detect d
e
tails in any num
ber
(
2
n
) of dire
ction
s
[12]. Beca
use the n
u
mbe
r
of direction
a
l sub
-
band
s
can
vary bet
wee
n
scale
s
, CT
is a
true
two-dime
nsio
nal digital
i
m
age
rep
r
e
s
enting
method. Figu
re 1 sh
ows th
e
stru
cture of CT [13, 14].
Figure 1. A Flow G
r
aph of t
he Co
ntourl
e
t Tran
sform
The CT
con
s
i
s
ts of the followin
g
step
s:
Step 1. Deco
mpose the im
age with LP transfo
rm.
Step 2. Deco
mpose the hi
gh frequ
en
cy sign
al, synthe
size the sin
g
u
lar poi
nts di
stribute
d
in the s
a
me direc
t
ions
as
a c
oeffic
i
ent.
Step 3. Reco
nstru
c
t.
The CT ha
s more
directio
nal sub
-
ba
nd
s
a
nd di
re
cti
onal i
n
form
ation, the
partiti
oning
of
the freque
ncy
spe
c
trum di
a
g
ram i
s
sh
own in Figure 2.
B
a
nd
pa
ss
di
r
ecti
onal
subband
B
a
nd
pa
ss
di
r
ecti
onal
subband
I
nput
im
a
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Co
ntou
rl
et Tran
sform
with Multiple
Cycle
s
Spinni
ng for Cate
na
ry Im
age… (Cha
ngd
ong
Wu
)
3889
Figure 2. The
Diagram of the Directio
nal
Filt
er Banks
Freq
uen
cy Spectrum Parti
t
ioning
Figure 3 i
s
th
e CT
of the
Came
ram
an i
m
age
us
i
ng
2 LP level
s
a
nd 8
dire
ctio
ns at th
e
fines
t level.
Figure 3. The
Diagram of the Co
ntourl
e
t Tran
sform T
w
o Laye
r
s
De
comp
ositio
n
2.2. C
y
cle Spinning
The CT cont
ains
2(
0
,
1
.
.
.
)
n
n
dire
ction
basi
s
fun
c
tio
n
s, the a
s
p
e
c
t ratio of e
a
c
h b
a
si
s
function i
s
o
p
tional, whi
c
h can
be u
s
ed in im
ag
e denoi
sin
g
widely. Ho
wever, the CT is
con
s
tru
c
ted b
y
LP decomp
o
sition a
nd DFB, the tr
ansf
o
rm ha
s the d
o
wn
-samplin
g, which is ea
sy
to gene
rate
the Gibb
s-li
ke phen
omen
a due to the
lack
of spin
ning invari
an
ce. In ord
e
r
to
sup
p
re
ss the Gibb
s-li
ke ph
enome
na, the
CT with multiple cycle
spin
ning metho
d
is pro
p
o
s
ed b
y
Coifman a
nd
Don
oho.
The cy
cle spi
nning may be
describ
ed a
s
follows:
1
,,
1,
1
1
(
{
[
(
(
))]
}
)
MN
ij
i
j
ij
IC
T
h
T
C
I
MN
(
1
)
W
h
er
e th
e
ima
g
e
s
ar
e of s
i
ze
(,
)
M
N
pixel
s
,
,
()
ij
CI
and
,
ij
C
are
the cy
cle
spi
nning
operator
and
inverse op
era
t
or, the sub
s
cript
i
and
i
are t
he line
dire
cti
on spinni
ng q
uantity to
the left or
right res
p
ec
tively,
j
and
j
are th
e column
dire
ction
spin
ning
quantity to th
e up
wa
rd
or d
o
wnwa
rd re
sp
ectivel
y
;
h
is th
e th
reshold
ope
rator;
T
and
1
T
are th
e con
t
ourlet
decompo
sitio
n
and re
co
nst
r
uctio
n
ope
rat
o
r [15].
The imag
e de
noisi
ng meth
od ba
sed o
n
CT with multi
p
le cycl
e spi
n
ning is a
s
foll
ows:
1. Spinning th
e origin
al ima
ge.
2. Multi-scale
decom
po
se the image.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3887 – 38
93
3890
3. Obtain the adaptive thre
shol
d and d
e
noise the ima
ge by CT.
4. Recon
s
tru
c
t image.
5. Inverse
spinnin
g
the i
m
age to o
b
tain the sam
e
se
quen
ce
image a
s
th
e origi
nal
image.
The m
a
in
ai
m of inte
nsity
image
d
enoi
sing
alg
o
rith
m is then
to
redu
ce
the
n
o
ise
level
while
p
r
eserv
i
ng the
ima
g
e
featu
r
e
s
[1
6]. Re
cently, den
oisi
ng
with ha
rd th
re
shol
d a
nd
soft
threshold
techniqu
e is
wi
dely used in
image
den
oisin
g
an
d h
a
s o
b
taine
d
some
succe
s
s.
Ho
wever, thi
s
metho
d
onl
y con
c
erns
o
ne point a
n
d
threshold
s
t
he imag
e term by term while
negle
c
ts thi
s
important fa
ct that the point
to be threshold
ed is depen
dent
on those in
its
neigh
borhoo
d
[17]. Usu
a
ll
y, the noise
of the ca
te
n
a
ry imag
es i
s
Ga
uss
whit
e noi
se, a
s
the
transfo
rme
d
Gau
ss
white
noise i
s
stil
l Gauss
whit
e noise, whi
l
e the sam
e
scal
e
and
all
dire
ctional
co
efficients gott
en by
CT
a
r
e
differe
nt. In t
h
is
pap
er, th
e
thre
sh
old
according
to
ea
ch
layer d
e
co
mp
osition
co
efficient
s an
d th
e noi
se inte
n
s
ity is
cho
s
e
n
, the select
ed threshold
has
self-a
dapta
b
il
ity.
3. Results a
nd Discu
ssi
on
The CT is
co
nstru
c
ted
by LP decomp
o
s
ition
a
nd
DF
B. The LP filters mainly h
a
v
e 5-3,
9.7, Pkva, Bu
rt, etc. the DFB mainly have Haa
r
, 5-
3, Pkva, Dmaxflat6, Cd, Sinc, etc. The choi
ce
of the LP filters and
DFB
has
ce
rtain i
n
fluen
ce
o
n
i
m
age d
enoi
si
ng qu
ality. In this pa
pe
r, the
different LP filters a
nd
DF
B are cho
s
en
to te
st the d
enoi
sing effe
cts. The te
sti
ng Ca
meram
an
Image co
ntai
ns the
35
Gaussi
an white n
o
ise.
Table 1. Ca
m
e
ram
an + G
a
ussian
White
Noi
s
e
(
35
)
Dfilter
Pfilter
Haar
5-3
Pkva
Dmaxflat6
Cd
Sinc
5-3
10.91
10.90
11.41
11.39
11.22
10.25
9-7
10.90
10.94
11.44
10.94
11.15
10.40
Pkva
10.50
10.67
11.01
11.08
10.81
10.35
Burt
10.84
10.89
11.41
11.43
11.12
10.60
Table 1
sho
w
s the S
N
R of choo
sing
the di
fferent
LP filters a
nd DFB. Experime
n
tal
results sho
w
t
hat it
could
g
e
t high
er SNR by
combi
n
g
the
9-7
filter
with the
Pkva
directio
nal filt
er.
That is be
ca
u
s
e th
e bio
r
th
ogon
al wavel
e
t ha
s t
he fe
ature
of linea
r ph
as
e, which can g
uaran
tee
the tran
sform coeffici
ent
s ar
e symm
etric a
nd sol
v
e the probl
em of edge
effect and
the
coeffici
ent ex
pan
sion.
No
matter
what
LP filters a
r
e
use
d
, a
s
lo
ng
as the
DFB
i
s
Haa
r
wavel
e
t,
the den
oisi
n
g
effects are
n
’t ideal, tha
t
is be
ca
u
s
e
the analytic pro
pertie
s
o
f
haar filter
are
relatively poo
r. Beside
s, a
s
the op
erati
on time is
lon
g
by combi
n
i
ng the Df
ilter filter, Dmaxflat6,
Sinc
with oth
e
r Pfilters for denoi
sin
g
, which
w
ill
affect the im
age
denoi
sing
efficien
cy, the
9
-
7
filter and P
k
v
a
filter a
r
e
ch
ose
n
for th
e
LP filter
an
d
DFB of
CT. T
o
text the effectivene
ss of
our
prop
osed alg
o
rithm, we de
noisi
ng the n
o
isy Ca
m
e
ra
man and L
e
n
a
image
s by wavelet tran
sform
(WT
), contou
rlet tran
sfo
r
m
(CT
)
a
nd
co
ntourlet tr
an
sform with
mul
t
iple cycl
e
spi
nning
(M
CSCT).
Figure 4 and
Table 2 sho
w
the experime
n
tal results.
In the
Figure 4
and
Table 2, MCSCT1 exp
r
e
s
se
s the CT u
s
ing
cycle
spi
nning o
ne tim
e
,
MCSCT
2
and
MCSCT
3
on
the analo
g
y of this.
In
the
Table
2,
is t
he n
o
ise
vari
ance. SNR i
s
the
signal
-to
-
noi
se ratio. The
hi
ghe
r
of
SNR, the bett
e
r of image d
enoi
sing effe
cts.
From the Fig
u
re 4, the re
sults of the
s
e
figur
es in
dicate that the CT metho
d
works well
and b
r
ing
s
so
me advanta
g
e
s in
kee
p
ing
the conto
u
r t
e
xtures th
an
WT, this imp
r
ovement of the
denoi
sing p
e
r
forma
n
ce is reasona
bly explained by
the fact that the CT
can
represent more
dire
ctional d
e
tails than
WT. Ho
weve
r, all
of them still have the artifact phenom
ena
. The
prop
osed me
thod is bette
r at detail information
reservation an
d
flicker rest
rain than oth
e
r
method
s. Th
e re
sults of T
able 2 indi
cat
e
that
the MCSCT m
e
tho
d
can
outpe
rform the
WT
and
CT in SNR m
easure, the m
o
re cy
cle spin
ning times, th
e highe
r SNR values.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Co
ntou
rl
et Tran
sform
with Multiple
Cycle
s
Spinni
ng for Cate
na
ry Im
age… (Cha
ngd
ong
Wu
)
3891
Figure 4. The
Denoi
sin
g
Result
s of Different De
noi
sin
g
Method
s wi
th Came
rama
n and Le
na
(
25
)
Table 2. SNR Contra
st of Different Den
o
isin
g Metho
d
s
Image
SNR
WT
CT
MCSCT1
MCSCT2
MCSCT3
Camerama
n
15 12.21
12.33
12.73
12.75
12.80
20 11.47
12.14
12.47
12.49
12.53
25 10.89
11.78
12.18
12.22
12.27
30 10.41
11.37
11.80
11.88
11.91
Lena
15 11.42
12.01
12.39
12.44
12.48
20 10.82
11.66
12.09
12.14
12.20
25 10.38
11.13
11.63
11.68
11.69
30 9.94
10.15
10.88
10.89
10.96
Figure 5. The
Catena
ry De
noisi
ng Resu
l
t
s of Different
Denoi
sin
g
Method
s
N
oisy
im
a
g
e
Me
thod
of
W
T
Me
thod
of
CT
Me
thod
of
M
C
S
C
T
1
Me
thod
of
MCS
C
T
2
Me
thod
of
MCS
C
T
3
N
ois
y
im
a
g
e
Me
thod
of
WT
Me
thod
of
CT
Me
thod
of
MCS
C
T
1
Me
thod
of
MCS
C
T
2
Me
thod
of
M
C
S
C
T
3
N
oi
sy
i
m
age
1
Me
th
od
of
W
T
Me
t
hod o
f
C
T
Me
th
od
of
M
C
S
C
T
1
Me
th
od
of
M
C
S
C
T
2
Me
th
od
of
M
C
S
C
T
3
N
oi
sy
i
m
a
g
e
2
Me
th
od
of
W
T
Met
h
od
of
C
T
Me
th
od
of
MCS
C
T
1
Me
th
od
of
M
C
S
C
T
2
Me
th
od
of
M
C
S
C
T
3
N
oi
sy
i
m
age
3
Me
th
od
of
W
T
Me
t
hod o
f
C
T
Me
th
od
of
MCS
C
T
1
Me
th
od
of
M
C
S
C
T
2
M
e
t
h
o
d
o
f
MC
SC
T
3
N
ois
y
im
a
g
e
4
Me
th
od
of
W
T
Me
t
h
od o
f
C
T
Me
th
od
of
MCS
C
T
1
Me
th
od
of
M
C
S
C
T
2
Me
th
od
of
M
C
S
C
T
3
N
ois
y
im
a
g
e
5
Me
th
od
of
W
T
Me
t
h
od o
f
C
T
Me
th
od
of
MCS
C
T
1
Me
th
od
of
M
C
S
C
T
2
Me
th
od
of
M
C
S
C
T
3
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3887 – 38
93
3892
Table 3. SNR Contra
st of Different Den
o
isin
g Metho
d
s
Image
SNR
WT
CT
MCSCT1
MCSCT2
MCSCT3
Image1
9.09
10.06
10.34
10.38
10.41
Image2
9.63
10.73
11.04
11.07
11.10
Image3
9.68
10.96
11.32
11.35
11.39
Image4
7.91
9.69
10.13
10.17
10.21
Image5
13.62
13.84
14.02
14.07
14.11
The
cate
nary
imag
e q
ualit
y is vital i
n
t
he fault
diag
nosi
s
. In
this pap
er, th
e p
r
opo
se
d
method i
s
ap
plied to de
no
ise the
actu
a
l
caten
a
ry im
age
s, the de
noisi
ng results are sho
w
n
in
Figure 5 and
Table 3.
To text the protecting a
b
ility of the image text
ure deta
ils, the partial
picture
s
of catenary
denoi
sing
re
sults ba
sed o
n
different den
oisin
g
metho
d
s are sh
own
in Figure 6.
Figure 6. The
Partial Pictures of Cate
nary
Denoi
sing
Re
sults Ba
se
d on Differe
nt Denoi
sin
g
Method
s
From
the
results
of
Figu
re
5,
Figure 6
and
Table
3, it c
an be
seen t
hat the
caten
a
ry
image
s d
enoi
sed
by
WT
show some
bl
ur
artifact
s,
which
have
so
me ri
ngin
g
a
r
tifacts a
r
o
und
the
edge
s in ima
ges.
The imag
es
denoi
se
d by CT have mo
re detail information than
WT, the SNR is highe
r
about 1dB th
an WT, that is be
cau
s
e th
e CT ha
s ri
ch
er direction
a
l informatio
n. The MCS
C
T
can
effectively de
noise the
ima
ges an
d eli
m
i
nate the
Gi
bb
s-li
ke
ph
eno
mena,
whi
c
h
is o
u
tpe
r
formi
ng
in den
oisin
g
pro
c
ed
ures
without lo
sin
g
the u
s
ef
ul
informatio
n
su
ch a
s
texture d
e
tails,
su
ch
observation
is con
s
iste
nt
with all the
d
enoi
sed im
ag
es, ju
st a
s
is the case wit
h
the o
b
jecti
v
e
SNR v
a
lue
s
.
Then our con
c
lu
sions a
r
e a
s
follows:
1.
T
he
CT
de
noi
sing
metho
d
works
well
a
nd
b
r
ing
s
so
me
adva
n
tag
e
s
in
keepi
ng
the
conto
u
r textures than
WT, althoug
h t
hey have the Gib
b
s-li
ke p
hen
o
m
ena.
2.
Use
the
MCS
C
T
method
for
denoi
sing, th
e protectin
g
a
b
ility of the image
edge
s
and
textures i
s
m
o
re
abu
ndant
than the
WT
and
CT,
whi
c
h
can
effecti
v
ely eliminat
e the Gi
bb
s-li
ke
phen
omen
a.
3.
With
the
increasi
ng
of
cycle spi
nnin
g
numbe
rs, the image deno
ising effect
s are
improve
d
, wh
ile the d
enoi
sing time i
s
i
n
crea
sed.
Gen
e
r
ally spea
kin
g
,
the suitable
cycle
spinnin
g
numbe
rs are
about thre
e times.
4.
T
he
n
o
ise
i
s
st
rong
er,
th
e
v
a
lue
s
of
S
NR
are
lo
we
r,
th
e
de
noi
sing
e
ffects
are
mo
re
d
e
c
r
e
as
ed
.
5.
Use
th
e
sam
e
method
to
d
e
noise
the
im
age
s
with
the
sa
me
noi
se,
the
cha
nge
s
of
the image det
ails are less, the den
oisin
g
effects a
r
e be
tter.
N
oisy
im
a
g
e
1
Me
t
h
o
d
o
f
W
T
Me
t
h
od
of
CT
Me
t
h
o
d
o
f
MC
S
C
T
1
M
e
t
h
od
of
M
C
S
C
T
2
M
e
t
h
od
of
M
C
S
C
T
3
No
i
s
y
i
m
age
2
Me
t
h
od
o
f
W
T
Me
t
h
od
of
CT
M
e
th
od
of
MCSCT
1
M
e
th
od
of
M
C
SCT
2
M
e
th
od
of
M
C
SCT
3
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Co
ntou
rl
et Tran
sform
with Multiple
Cycle
s
Spinni
ng for Cate
na
ry Im
age… (Cha
ngd
ong
Wu
)
3893
4. Conclusio
n
Acco
rdi
ng to
the multi-scal
e, multi-re
sol
u
tion a
nd
ani
sotro
p
y virtue
s of
CT,
re
asonabl
e
sele
cting th
e
LP filter an
d
DFB, an im
a
ge de
noi
sing
method
ba
se
d on
CT i
s
a
p
p
lied. In a
ddit
i
on,
aim to the
sh
ortco
m
ing
of CT
without th
e cycl
e in
va
riance, the con
t
ourlet tra
n
sfo
r
m with
multi
p
le
cycle
spin
nin
g
is propo
se
d to overcom
e
the Gi
bb
s-li
ke ph
enom
e
na eme
r
ge
d in the pro
c
e
s
s of
image
den
oising. In thi
s
pa
per,
seve
ral
grou
ps
of
sta
ndard im
age
s and
a
c
tual
catenary
imag
e
s
are
chosen
to text the v
a
lidity, the e
x
perim
e
n
t re
sults sho
w
t
hat the
MCS
C
T m
e
thod
can
effectively su
ppre
s
s
noise
and
eliminate
the
Gibb
s-li
ke ph
enom
en
a than
WT a
nd
CT, the
m
o
re
cycle
spin
nin
g
times, the better de
noisi
ng effects,
g
enerally spe
a
k
ing, a
c
cordi
ng to the co
st of
denoi
sing tim
e
, the suitabl
e cycle
spin
ni
ng times
a
r
e
about three. In sho
r
t, the p
r
opo
se
d meth
od
is an
effectiv
e imag
e de
n
o
isin
g meth
o
d
owi
ng to
th
e value
s
of S
N
R and
effectively preserv
i
ng
details a
nd te
xture informat
ion of origin
al
images.
Ackn
o
w
l
e
dg
ements
This stu
d
y is sup
p
o
r
ted
by National
Natural Sci
ence Foun
d
a
tion of Chi
na (NO.
5137
7136;
NO. 513
0714
4
)
, Te
chn
o
logi
cal
Re
sea
r
ch
and
Devel
o
p
m
ent Pro
g
ra
m of the Mi
ni
stry
of Railway
s (2011
J0
16-B) and Fun
damental
Re
se
arch Fu
n
d
s for Cent
ral Unive
r
siti
es
(SWJT
U
11
CX141) in
Chin
a.
Referen
ces
[1]
Yu
yin
g
Sh
i, Yo
ngg
ui Z
h
u, Ji
n
g
jin
g
Liu.
Semi
-imp
lic
it Imag
e
Den
o
isi
ng
Alg
o
r
ithm for D
i
ffer
ent Bo
un
dar
y
Con
d
itio
ns.
T
E
LKOMNIKA Indon
esia
n Jour
nal of
Electric
al
Engin
eeri
ng.
2
013; 11(
4): 205
8-20
63.
[2]
QiuShe
ng
Lia
n
, ShuZ
h
en
Che
n
. T
he T
r
ansl
a
tion
Inva
riant C
onto
u
rl
et-like T
r
ansfo
rm for Imag
e
Den
o
isi
ng.
Acta Autom
a
tica Sinica
. 2
0
0
9
; 35
(5): 505-5
08.
[3]
Yang T
i
an
xu,
W
angji
a
n
w
e
i
,
Hu Xia
ogu
an
g. Res
earc
h
on the Mech
anic
a
l
State Parame
ter Extractio
n
Method of Hi
g
h
Volt
ag
e Circ
u
it Breakers.
T
E
LKOMNIKA Indo
nesi
an Jo
u
r
nal of Electric
al Eng
i
ne
erin
g.
201
3; 11(5): 27
71-2
779.
[4]
Arthur L da Cu
nha, Jia
npi
ng
Z
hou. T
he Nonsubs
ampl
ed Conto
u
rlet
T
r
ansfo
rm:
T
heor
y, Design, an
d
Appl
icatio
ns.
IEEE Transactions on
I
m
age Processing
. 20
06; 15(1
0
): 308
9-31
01.
[5]
Huai
bi
n W
A
NG
, Yuanc
ha
o LI
U, Ch
und
on
g
W
A
NG. Resea
r
ch o
n
Bl
ind
W
a
termarki
ng A
l
gorithm
Base
d
on C
h
a
o
tic En
cr
y
p
ti
on
an
d C
ontour
let T
r
ansform.
Journ
a
l
of Infor
m
ati
o
n
& Co
mputati
o
nal Sc
ie
nce
.
201
1; 8(15): 37
03-3
711.
[6]
Qian Z
hao, B
o
Ye, Jiali
n
Cao. Image
Den
o
isi
ng Ba
sed on Impr
o
v
ed No
n-loc
a
l
Means a
n
d
Nons
ubsam
pl
e
d
Co
ntour
let
T
r
ansform W
i
ener
F
ilteri
ng.
Journ
a
l
of
Co
mp
utation
a
l
Informatio
n
System
s
. 20
10
; 6(2): 601-60
9
.
[7]
W
e
i Sio
n
g
Le
e, Ashraf A. K
a
s
s
im. Sign
al
an
d
Imag
e Ap
pro
x
im
atio
n Usi
n
g
Interval W
a
v
e
l
e
t T
r
ansform.
IEEE Transactions on I
m
age
Processing.
20
07; (16): 46-5
6
.
[8]
Ramin Es
lami,
Ha
yder R
a
d
h
a
. W
a
velet-ba
sed cont
ourl
e
t transform an
d its app
licati
on to ima
g
e
codi
ng.
Proces
sings of IEEE ICIP
. 2004; (5): 318
9-31
95.
[9]
Ramin Esl
a
mi, Ha
yd
er Rad
ha.
T
he Contour
le
t
T
r
ansform for Image De-n
oi
sing Usi
ng C
y
c
l
e Spi
nni
ng
.
A silo
mar C
onf
erenc
e on Si
gn
als, Systems,
a
nd Co
mputers.
Pacific Grove, USA
. 2003; 1
9
82-1
986.
[10]
Peng F
e
ng, Yi
ng ju
n Pa
n, Biao W
e
i, et.al.
E
nha
ncin
g re
tinal im
ag
e b
y
the Cont
ourl
e
t transform.
Pattern Recognition Letters
. 200
7; 28: 516-
522.
[11]
Peng F
e
n
g
,
Jin
g
W
ang,
Biao
W
e
i, et al. A
Fusi
o
n
Algor
ith
m
for GF
P Image an
d Phase
Contrast Imag
e
of
Arab
ido
p
sis Cell
Bas
ed on
SF
L-Conto
u
rlet
T
r
ansform.
C
o
m
p
u
t
a
t
i
o
n
a
l
and
Ma
them
a
t
i
c
al Me
th
od
s in
Medici
ne
. 20
13
; (2013): 1-10.
[12]
Zhilin
g Lo
ng, Nicol
as H. You
nan. Multisc
a
le
texture se
gme
n
tation
vi
a a contour
let conte
x
tu
al hi
dde
n
Markov model.
Digital S
i
gn
al
Processi
ng.
20
13; (23): 85
9-8
69.
[13]
NG Chit
ali
y
a,
AI T
r
ivedi. An
Efficient Meth
o
d
for F
a
c
e
F
e
a
t
ure E
x
tractio
n
an
d R
e
co
gniti
on
bas
ed
on
Conto
u
rlet T
r
ansforms a
nd P
r
incip
a
l
Com
p
o
nent A
nal
ys
is.
Proced
ia Co
mputer
Sc
ienc
e.
201
0; (2): 5
2
-
61.
[14]
M DO, M Vetterli. T
he contourlet transform an effi
cie
n
t direc
t
iona
l multir
eso
l
utio
n ima
ge re
prese
n
tatio
n
IEEE Trans. Image Proc
ess
. 2
005; (14): 2
091
-210
6.
[15]
Don
g
L
i
a
ng, M
i
n S
h
e
n
, Qin
g
-w
e
i
Gao,
et
a
l
. A Meth
od
for
Image
De
nois
i
ng B
a
se
d
on t
he
Cont
ourl
e
t
T
r
ansform Using Rec
u
rsive
C
y
c
l
e Sp
inn
i
n
g
.
ACT
A
Electronica si
nica
. 20
05; 33(1
1
): 204
4-20
46.
[16]
Xi
ao
fen
g
L
i
, J
un
Xu, Jij
u
n
Lu
o, et.al. Inte
nsi
t
y
im
ag
e d
e
n
o
i
s
ing
for l
a
ser
a
c
tive ima
g
i
ng s
y
stem
usi
n
g
nons
ubs
ampl
e
d
contour
let tra
n
sform and SU
RE appr
oac
h.
Optik
. 2012; (1
23): 808-
81
3.
[17]
Jingl
on
g C
h
e
n
,
Yan
y
a
ng Z
i
,
Z
hen
gji
a
H
e
, et al. A
d
a
p
ti
ve red
u
n
d
a
n
t multi
w
av
el
et d
eno
isin
g
w
i
t
h
improve
d
ne
ig
hbor
ing c
oeffi
cients for ge
a
r
bo
x fault d
e
tection.
Mech
a
n
ical Syste
m
s
and Si
gna
l
Processi
ng
. 20
13; (38): 54
9-5
68.
Evaluation Warning : The document was created with Spire.PDF for Python.