Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
6, N
o
. 2
,
A
p
r
il
201
6, p
p
.
61
1
~
62
0
I
S
SN
: 208
8-8
7
0
8
,
D
O
I
:
10.115
91
/ij
ece.v6
i
2.9
044
6
11
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
An Eff
e
ctive Nois
e Ad
apti
ve M
e
dian Filter for Rem
ovin
g
High
Density Impulse Noises
in
Col
o
r I
m
ages
S.
Abdul Saleem*, T.
Abdul
Raz
a
k **
* Bharathidas
a
n
University
, India
** Departmen
t
o
f
Computer Science, Ja
mal Moh
a
med Colleg
e
(A
utonomous), India
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Sep 21, 2015
Rev
i
sed
D
ec 29
, 20
15
Accepte
d
Ja
n 14, 2016
Im
ages
are nor
m
a
ll
y degrad
ed
b
y
s
o
m
e
form of impulse noises during th
e
acquisition, tran
smission and storage in
th
e
ph
y
s
i
cal m
e
d
i
a.
M
o
s
t
of the
rea
l
tim
e app
lic
ation
s
usuall
y r
e
quire
bright
and
cl
ear
im
ages
,
henc
e
dis
t
orted
o
r
degraded
im
age
s
need to
be
p
r
oces
s
e
d to
enh
a
nce
eas
y
iden
t
i
fic
a
tion
of
im
age deta
ils
an
d further works
on the im
age. In
this
paper we ha
ve anal
yz
ed
and tested th
e
number of exis
ting me
dian
filtering algor
ithms and their
limitations. As
a result we h
a
v
e
pr
oposed
a n
e
w effective no
ise ad
aptiv
e
m
e
dian filt
ering
algorithm
,
whi
c
h rem
oves the
im
pulse noises in the color
im
ages
while pres
erving the im
a
g
e deta
ils and enhancing th
e image quality
.
The proposed
method is a spatial dom
ain approach and us
es the 3×3
overlapp
i
ng win
dow to filter
th
e signa
l b
a
sed
on the
corr
ect
s
e
le
ction
of
neighborhood v
a
lues
to obtain
the ef
f
ective
median per
window. Th
e
performance o
f
the proposed
effectiv
e me
dian
filter has b
een
evaluated usin
g
MATLAB, simulations
on a both
gray
scal
e and
color
images that have
b
e
en
subjected to hig
h
density
o
f
cor
r
upti
on up to 90
% with impulse noises. The
results expose the effectiv
eness of our proposed algorithm when compared
with the
quant
it
ativ
e im
age
m
e
t
r
ics such
as PSNR, MSE, RMSE, I
E
F, T
i
m
e
and SSIM of ex
isting stand
a
rd
an
d adap
tive median filter
i
ng
algor
ithms.
Keyword:
Ed
ge prese
r
vat
i
on
Fre
que
ncy
do
m
a
i
n
Im
age param
e
ters
Im
age restorati
o
n
Im
pul
se n
o
i
s
e
Med
i
an
filters
Sp
atial do
m
a
in
Copyright ©
201
6 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
S. Abdul
Salee
m
,
Research Sc
holar, PG
& Rese
arch
De
part
m
e
nt
o
f
C
o
m
put
er
Sci
e
nce,
Jam
a
l
M
oham
e
d C
o
l
l
e
ge (
A
ut
on
om
ous)
,
No.
7
,
Race Cou
r
se Ro
ad
, Khaj
an
ag
ar,
Tiru
ch
irap
p
a
lli – 620
0
2
0
,
Tam
il Nad
u
,
Ind
i
a
Em
a
il: salee
m
n
t
s@g
m
ai
l.com
1.
INTRODUCTION
Im
age Process
i
ng is one
of t
h
e fast risi
ng
fields
in the area of c
o
m
puter science a
nd
engi
neeri
n
g.
The
gr
owt
h
o
f
t
h
i
s
fi
el
d
h
a
s bee
n
su
pe
r
i
or
by
t
h
e t
echn
o
l
o
gi
cal
a
dva
ncem
ent
s
i
n
di
gi
t
a
l
com
put
i
n
g,
pr
ocess
o
rs
, m
u
l
t
i
m
e
di
a dat
a
p
r
oces
si
n
g
a
n
d m
a
ss st
ora
g
e de
vi
ces.
Al
l
fi
el
ds
w
h
i
c
h
were
o
p
e
r
at
i
n
g
o
n
t
h
e
an
alog
si
g
n
a
ls are n
o
w in
creasin
g
l
y co
nv
ertin
g
in
t
o
th
e d
i
gital s
y
ste
m
s
fo
r th
eir ease o
f
u
s
e, reliab
ility an
d
flexibility.
Im
age processi
ng has bee
n
ex
te
nsively applied
in the a
r
ea
of
m
e
dical, photogra
phy, film
indust
r
y,
rem
o
te sen
s
in
g, traffic con
t
rol, astron
o
m
y, p
o
lice inv
e
s
tigatio
n
,
b
u
sin
e
ss, in
du
stry, t
r
ansp
ort traffic-con
t
ro
l,
m
i
li
t
a
ry
t
a
rget
anal
y
s
i
s
, an
d m
a
nufact
uri
ng
aut
o
m
a
t
i
on an
d co
nt
r
o
l
.
Im
age p
r
e-
pr
ocessi
ng t
e
c
hni
ques
suc
h
as
i
m
ag
e enh
a
n
c
emen
t, im
ag
e resto
r
ation
an
d
ob
j
ect
reco
gn
itio
n are u
s
ed
t
o
p
r
o
cess th
e imag
e
d
e
p
e
nd
ing
o
n
the
typ
e
of in
terferen
ce th
at
h
a
s cau
s
ed
th
e d
e
g
r
ad
atio
n [1
].
No
ises in
t
h
e
d
i
g
ital im
ag
es
are m
o
d
e
lled as thr
ee stand
a
rd
categ
ories, th
ey are add
itiv
e no
ises,
m
u
lt
i
p
l
i
cat
i
v
e
noi
ses a
n
d ra
n
dom
im
pul
se noi
ses
.
M
o
st
of t
h
e
di
gi
t
a
l
im
ages are n
o
r
m
a
ll
y
corru
pt
ed by
im
pul
se noi
se
du
ri
n
g
com
m
uni
cat
i
on [
2
]
.
T
h
e t
w
o c
o
m
m
o
n
im
pul
se n
o
i
s
e t
y
pes are ran
dom
-val
ue
d n
o
i
se and
sal
t
and pe
ppe
r
noi
se. I
n
t
h
e
r
a
nd
om
im
pul
se noi
se m
ode
l
,
im
age pi
xel
s
are ran
d
o
m
l
y corr
upt
e
d
by
t
w
o
fi
xed
ex
trem
e v
a
lu
es, 0
and
2
5
5
(for gray-scale
i
m
ag
e)
, g
e
n
e
rated
with
th
e sa
m
e
p
r
ob
ab
ility th
at is
P
i
s
noi
s
e
d
e
nsity, th
en
P1
is t
h
e
n
o
i
se
d
e
nsity o
f
salt
(
P/2
) a
n
d
P2
is
th
e no
ise d
e
n
s
ity
o
f
p
e
pp
er (
P/2
).
In
stead of
tw
o
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 2, A
p
ri
l
20
16
:
61
1 – 6
2
0
61
2
fixe
d
values, i
m
ages m
a
y be corrupte
d
by two fixe
d ranges
that
a
ppea
r
both e
n
ds
with
length of m
each
resp
ectiv
ely, t
h
at is [0
,m
]
d
e
no
tes salt an
d
[255
-m
,2
55
] d
e
no
tes p
e
p
p
e
r.
Here for no
ise d
e
n
s
ity
P,
P1=P2=P/2
.
Ano
t
h
e
r m
o
d
e
l
with
on
ly lo
w
in
ten
s
ity i
m
p
u
l
se no
ise an
d on
ly h
i
gh
i
n
ten
s
ity i
m
p
u
l
se n
o
i
se also
affect the
digit
a
l im
ages, that is
P1
≠
P2
. The salt an
d
pepper no
ises are satu
rated
v
a
lu
es th
at tak
e
m
a
x
i
m
u
m
an
d m
i
n
i
m
u
m allo
wed
v
a
lues of in
ten
s
ities. In
t
h
is p
a
pe
r we are m
e
r
e
ly co
n
s
i
d
ering
im
p
u
l
se no
i
s
e wit
h
P1= P
2
.
In a
n
y
si
gnal
p
r
oces
si
n
g
sy
st
em
, fil
t
e
ri
ng i
s
an esse
nt
i
a
l
part
whi
c
h i
n
vol
ves est
i
m
at
i
on of a si
g
n
al
deg
r
a
d
ed i
n
m
o
st
cases
by
i
m
pul
se noi
se.
Im
pul
se n
o
i
s
es
are m
o
stly ca
use
d
duri
ng the process
of image
acq
u
i
sition
,
tran
sm
issio
n
th
rou
g
h
co
mm
u
n
i
catio
n
m
e
d
i
a an
d
sto
r
ag
e in
the p
h
y
sical d
e
v
i
ces. Sev
e
ral filtering
t
echni
q
u
es
ha
v
e
been
de
vel
o
ped
o
v
er t
h
e p
a
st
severa
l
dec
a
des f
o
r va
ri
o
u
s ap
pl
i
cat
i
ons
. The t
y
pe
of
noi
s
e
facto
r
and
in
ten
s
ity o
f
the noise th
at h
a
s
d
e
g
r
ad
ed
th
e im
a
g
e is also
tak
e
n
in
to
con
s
id
eratio
n
befo
re t
h
e filter
i
s
de
vel
o
ped
a
n
d
us
ed
.
Th
e techn
i
qu
es fo
r filtering
imag
e n
o
i
ses can
b
e
d
i
v
i
d
e
d
in
to
two
b
r
o
a
d categ
o
r
ies: spatial d
o
m
a
in
filterin
g
an
d
frequ
e
n
c
y d
o
m
ain
filterin
g
. Th
e sp
atial d
o
main
filterin
g
tech
n
i
qu
es
are b
a
sed
o
n
t
h
e d
i
rect
man
i
p
u
l
ation
o
f
th
e
im
ag
e p
i
x
e
ls wh
ere as
th
e frequ
ency d
o
m
ain
filtering
techn
i
ques h
a
v
e
t
o
do with
m
o
d
i
fyin
g
th
e
Fou
r
ier tran
sform
o
f
th
e in
terested
im
ag
e.
Th
e sp
atial do
main
filtering
is fu
rth
e
r
subd
ivid
ed
in
to
lin
ear filtering
an
d
n
o
n
lin
ear filtering
. In
lin
ear f
iltering
, a sin
g
l
e p
i
xel with
v
e
ry u
n
r
ep
resen
t
ativ
e v
a
lue
can
sign
ificantly affect th
e
m
ean
v
a
lu
e o
f
all th
e p
i
x
e
ls in
its n
e
ig
hb
ourhoo
d an
d
wh
en
th
e filter
n
e
igh
bou
rho
od stand
across an
edg
e
th
e filter
will in
terpo
l
ate n
e
w
v
a
lu
es fo
r p
i
x
e
ls
o
n
th
e edg
e
an
d so will
bl
u
r
t
h
at
edge
. Thi
s
m
a
y
be a
pr
o
b
l
e
m
i
f
sharp e
dges a
r
e re
qui
red i
n
t
h
e
o
u
t
p
ut
. These
pr
obl
em
s are rect
i
f
i
e
d
b
y
th
e
no
n
lin
ear filtering
. Ord
e
r statistic filters are
n
o
n
l
in
ear sp
atial fi
lters wh
o
s
e resp
on
se is b
a
sed
o
n
o
r
d
e
ri
n
g
th
e
p
i
x
e
ls co
n
t
ai
n
e
d
in
th
e i
m
ag
e area en
co
m
p
assed
b
y
th
e filter an
d
th
en
rep
l
acin
g
th
e v
a
l
u
e o
f
the
cen
ter
p
i
x
e
l
with
th
at v
a
lu
e
d
e
term
in
ed
b
y
th
e ran
k
i
ng
resu
lt [2
]. Th
e
b
e
st kno
wn
o
r
d
e
r-statistic n
o
n
lin
ear
filter is th
e m
e
d
i
an
filter.
A
num
ber
of
m
e
t
hods
ha
ve
bee
n
i
n
t
r
o
d
u
c
e
d t
o
rem
ove
im
pul
se n
o
i
s
e
fr
om
di
gi
t
a
l
i
m
ages. T
h
e
stan
d
a
rd
m
e
d
i
an
filter and
m
e
an
filter are u
s
ed
to
re
d
u
ce salt & p
e
p
p
e
r no
ise an
d
Gau
s
sian
no
ise resp
ectiv
ely.
Whe
n
these t
w
o noises exist in the
sam
e
image,
use of only one filteri
ng
m
e
thod ca
nnot
achieve the
de
sired
resu
lt.
Vector Med
i
an Filter (VMF)
is a sim
p
le rank
selec
tion
filter th
at i
d
en
tifies and
elim
in
ates th
e
fix
e
d
an
d rand
o
m
v
a
lu
ed im
p
u
l
se no
ises i
n
th
e
d
i
g
ital i
m
ag
es.
In
t
h
is filtering
alg
o
rith
m
,
th
e
v
ector
o
f
p
i
x
e
l
s
in
a
speci
fi
ed
wi
nd
ow
i
s
ra
nke
d
on
t
h
e
basi
s
o
f
s
u
m
of t
h
e
di
st
ances t
o
ot
her
vect
or
of
pi
xel
s
i
n
t
h
e a
not
he
r
window.
The center vector of
pi
xel
is
declared as
noisy i
f
its ra
nk is
bi
gge
r t
h
an a predefi
n
ed ra
nk
and its
di
st
ance f
r
om
a near
by
heal
t
h
y
vect
o
r
pi
xe
l
i
s
bi
gger
t
h
an
th
e pred
efined
th
resho
l
d. Th
e no
isy p
i
xel is
repl
ace
d wi
t
h
t
h
e vect
or m
e
di
an. T
h
e t
h
res
h
ol
d m
echani
s
m
fo
r det
ect
i
o
n
of
n
o
i
s
y
pi
xel
and
re
pl
aci
ng i
t
wi
t
h
vect
o
r
m
e
di
an i
s
sui
t
a
bl
e fo
r im
ages wi
t
h
noi
ses u
p
t
o
50
% of
noi
se
l
e
vel
.
Eve
n
t
h
ou
g
h
VM
F i
s
noi
se
ad
ap
tiv
e filter,
it is n
o
t
su
itab
l
e fo
r
h
i
gh
er
n
o
ise d
e
n
s
ities [3].
Stan
d
a
rd
Median
Filter
(SMF) is also
a sim
p
le ran
k
selectio
n
filter th
at attem
p
ts
to
elim
in
ate
i
m
p
u
l
se n
o
i
se
b
y
ch
ang
i
ng
the lu
m
i
n
a
n
ce v
a
lu
e o
f
th
e cen
t
er p
i
x
e
l of th
e
filterin
g
wind
ow with
th
e m
e
d
i
an
of
the lum
i
nance
values
of t
h
e
pixels
con
t
ain
e
d with
in th
e wi
nd
ow. Althou
gh th
e
SMF is si
m
p
le an
d
p
r
o
v
i
d
e
s a
reaso
n
a
b
l
e
noi
se rem
oval
perf
orm
a
nce, i
t
rem
oves t
h
i
n
l
i
n
es an
d bl
u
r
s
im
age det
a
i
l
s
even at
l
o
w
noi
s
e
densities. Furt
herm
ore, it has
no ada
p
ta
tion for va
rying
noi
se levels for a
reliable
m
e
dian signal. This
m
e
thod
affect
s t
h
e i
n
f
o
rm
ati
on
of t
h
e
unc
o
r
r
upt
e
d
t
r
ue
pi
xel
by
t
a
k
i
ng m
e
di
an i
t
s
e
l
f i
m
pul
se val
u
e [
4
]
.
Weigh
t
ed
Median
Filter (W
M
F
) and
Cen
t
er
Weigh
t
ed
Median
Filter (C
WMF) are m
o
d
i
fied
m
e
d
i
an
filters in
tro
d
u
c
ed
to
p
r
eserv
e
th
e im
ag
e d
e
tails o
f
all th
e spatial p
o
s
itio
n
s
b
y
g
i
v
i
n
g
m
o
re ex
tra
weigh
t
to
th
e
ap
pro
p
riate p
i
xels o
f
t
h
e filterin
g
wind
ow.
Th
ese filters
h
a
ve b
een propo
sed
to
avo
i
d
th
e
in
h
e
ren
t
d
r
awback
s
o
f
th
e stand
a
rd m
e
d
i
an
filter b
y
con
t
ro
llin
g
the tr
ad
e-off b
e
t
w
een
th
e no
ise
supp
ressi
on
an
d
d
e
tail
p
r
eserv
a
tio
n.
Bu
t th
eir d
e
tail
p
r
eserv
a
tio
n
on
im
ag
es is li
mited
as th
e ex
t
r
a weigh
t
g
i
ven to
a corru
p
t
ed
sig
n
a
l
can inc
r
ease noise of the
hi
gh
ly co
rrup
ted
d
i
g
ital i
m
ag
e an
d
t
h
ere is
n
o
ad
ap
tation
toward
s t
h
e v
a
rying
no
ise
r
a
tio
for
ch
oo
si
n
g
th
e
w
e
igh
t
an
d n
e
i
g
hbo
rhoo
d of
a p
a
r
ticu
l
ar
sign
al
[
5
].
Th
e
Pro
g
ressi
ve Switch
i
n
g
M
e
d
i
an
Filter (PSMF) is
ob
tain
ed b
y
co
m
b
in
in
g
t
h
e m
e
d
i
an
filter wit
h
an im
pulse det
ector and an im
pulse correct
or. The im
pulse detector aims to determ
ine
whethe
r the center
p
i
x
e
l of a g
i
v
e
n
filtering
wind
ow is co
rru
p
t
ed
or no
t.
If t
h
e cen
ter
p
i
x
e
l is id
en
tified
b
y
th
e d
e
tector as a
corrupted pi
xel, the
n
it is re
placed
with
the out
put of
the median
filter, othe
rwise, it is left uncha
n
ge
d.
In the
case wh
ere m
a
j
o
rity o
f
th
e ed
g
e
p
i
x
e
ls in
th
e i
m
ag
e are p
o
llu
ted
b
y
i
m
p
u
l
se
n
o
i
se,
filterin
g
is in
com
p
le
te
because the switching m
e
dian filter onl
y works on the cent
r
e va
lue of the
window and e
v
en for the s
m
allest
si
zed wi
n
d
o
w
,
3×3, i
t
i
s
not
pos
si
bl
e t
o
ha
ve an e
dge
pi
x
e
l
i
n
t
h
e cent
r
e of t
h
e sl
i
d
i
n
g wi
n
d
o
w
.
I
n
i
m
pul
se
correction
pha
s
e, an iterativ
e
correction process
foll
ows where only
th
e c
o
rrupted
pixels
are re
placed
by the
m
e
di
an of u
n
c
o
r
r
u
p
t
e
d
pi
xel
s
of a wi
nd
o
w
i
d
ent
i
f
i
e
d i
n
t
h
e latest
d
e
tectio
n
iteratio
n
.
The flag
is reset, mean
s,
th
e n
e
x
t
iteratio
n u
s
es th
e m
o
d
i
fied im
ag
e an
d
the m
odifie
d
flag im
age as
inputs
[6].
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Effective No
ise Ad
ap
tive
Med
ian
Filter f
o
r Remo
ving
Hig
h
Den
s
ity Imp
u
l
se No
ises
…
(S
.
Ab
du
l
S
a
leem
)
61
3
In th
is
p
a
p
e
r,
w
e
p
r
op
o
s
e a sp
atial do
m
a
in
meth
o
d
u
s
i
n
g th
e
o
v
e
rlapp
i
ng k
e
rn
el
w
i
ndow
to filter t
h
e
i
m
ag
e b
a
sed on
th
e
selectio
n o
f
n
e
i
g
hbo
r
i
ng
pixel val
u
es
and obtaining an e
fficient
med
i
an
p
e
r
w
i
ndow
p
o
s
ition
.
Fo
r each
windo
w
po
sitio
n
a m
e
d
i
an
is fou
n
d
fo
r selectiv
e p
i
x
e
ls in
th
e win
d
o
w
d
e
p
e
nd
ing
o
n
the
con
d
i
t
i
on
we p
u
rs
ue
d. T
h
e m
e
di
an i
s
t
e
st
ed,
and i
f
i
t
i
s
unaffect
ed
by
im
pul
se
noi
se, i
t
i
s
confi
r
m
e
d as t
h
e
effective
m
e
dian.
Th
e
org
a
n
i
zatio
n of t
h
is
p
a
p
e
r is as fo
llows:
Sect
i
o
n
2
di
sc
usses
rel
a
t
e
d
wo
rk
s
w
h
i
c
h
i
n
vol
ve
rem
o
v
a
l o
f
imp
u
l
se
no
ises usin
g
ad
ap
tiv
e
med
i
an
and
so
me o
f
its d
e
ri
vativ
e filters. Sectio
n
3
presen
ts th
e
p
r
op
o
s
ed
effectiv
e n
o
i
se ad
ap
tiv
e m
e
d
i
an
filterin
g
alg
o
rith
m
.
Th
e co
m
p
arison
o
f
p
r
opo
sed
filter with
o
t
h
e
r
n
on-lin
ear filters
b
y
using
q
u
a
n
titativ
e m
e
trics is g
i
v
e
n
u
n
d
e
r th
e h
e
ad
ing of resu
lts and
d
i
scu
ssion in
Sectio
n
4 a
n
d
fi
nal
l
y
t
h
e pa
per
i
s
co
nc
l
ude
d
wi
t
h
fut
u
re
di
rect
i
o
n i
n
Sect
i
on
5.
2.
RELATED WORKS
In
th
is section
,
we p
r
esen
t a b
r
ief
rev
i
ew of th
e ad
ap
tiv
e
med
i
an
filtering
alg
o
rith
m
s
. Th
e ad
ap
tiv
e
med
i
an
filters
are non
-lin
ear
o
r
d
e
red
statistic d
i
g
ital
filterin
g
tech
n
i
q
u
e
s
wh
ich
are
no
rmall
y
u
s
ed
to
redu
ce
high de
nsity noises ext
r
em
ely in
an im
age. It is one
of
the best
w
i
ndo
w
i
n
g
op
er
at
or
s
ou
t o
f
t
h
e man
y
wind
owing
operato
rs lik
e t
h
e m
ean
filter,
min
and
m
a
x
filter an
d th
e m
o
de filter.
Hwa
ng et.al.,
propose
d
a
n
Adaptive
Media
n
Filte
r
(AMF) to elim
inate th
e problem
s faced
by the
Stan
d
a
rd
Med
i
an
Filter an
d
Switch
i
ng
Med
i
an
Filters
. AMF ch
an
g
e
s its b
e
h
a
v
i
o
u
r
b
a
sed
o
n
t
h
e stat
istical
ch
aracteristics o
f
the im
ag
e in
sid
e
th
e filter
wind
ow. Th
e
perfo
r
m
a
n
ce o
f
Ad
ap
tiv
e
filter is u
s
u
a
lly sup
e
rio
r
t
o
n
on-ad
ap
tiv
e co
un
terp
arts. Th
e i
m
p
r
ov
ed
perfo
r
m
a
n
ce is
at th
e co
st o
f
ad
d
e
d
filter com
p
lex
i
t
y
. Mea
n
and
v
a
rian
ce are t
w
o
im
p
o
r
tan
t
statistical
m
e
a
s
u
r
es
b
a
sed
on wh
ich
ad
ap
tive filters can
b
e
d
e
sign
ed. In
practice
th
is filter im
p
o
s
es a li
m
it to
t
h
e wi
nd
ow size, S
xy
. W
h
en
t
h
is li
m
it is rea
c
h
e
d wh
ile th
e selected
m
e
d
i
an
is an
i
m
p
u
l
se, th
e i
m
p
u
l
siv
e
no
ise rem
a
in
s in
t
h
at windo
w of t
h
e im
ag
e. Th
e
ad
ap
tiv
e m
e
d
i
an
filter ach
iev
e
s goo
d
resu
lts in
m
o
st cases, bu
t ev
en
so
, co
m
p
u
t
atio
n
tim
e is
p
r
op
ortion
a
l to
the d
e
gree
o
f
co
rrup
tion
of th
e i
m
ag
e
b
e
ing
filtered
[4
].
Ran
k
Ord
e
red
Ad
ap
tiv
e Med
i
an
Filter (R
OAMF) also
p
r
op
o
s
ed
b
y
Hwan
g, H. an
d
Had
d
a
d
,
R.A.,
wh
ich
k
eeps t
h
e im
ag
e d
e
tails o
f
h
i
gh
ly co
rrup
ted d
i
g
ital i
m
ag
es b
y
switch
i
ng
th
e filterin
g
o
f
on
ly th
e
cor
r
u
p
t
e
d si
gn
al
s wi
t
h
a m
i
d-
ran
k
i
n
g
val
u
e chose
n
f
r
o
m
a nei
g
h
b
o
r
h
o
od t
h
at
va
ries
adaptively wi
th the
qua
nt
um
of im
pul
se n
o
i
s
e. A
M
F det
ect
s corr
upt
e
d
si
gnal
s
by
checki
n
g t
h
em
t
o
be bet
w
een m
i
nim
u
m
an
d
m
a
xim
u
m
of t
h
e m
e
di
an det
ect
ed nei
g
h
b
o
r
h
o
o
d
, i
t
fet
c
he
s a reliable median
from
an ada
p
tively va
rying
n
e
igh
bor
hoo
d
f
o
r
o
n
l
y th
e co
rr
up
ted sign
al
s and
w
o
rk
s
ver
y
w
e
ll fo
r all typ
e
s of
im
a
g
es
u
p
to
60
% no
ise
lev
e
ls. Th
e main
li
m
i
tatio
n
o
f
th
is filter is th
at th
e i
m
p
u
l
se rep
l
acing
m
e
d
i
an
is n
o
t
d
e
term
in
ed
fro
m
u
n
c
orru
p
t
ed
p
i
x
e
ls, im
p
u
l
se rep
l
acin
g
m
e
d
i
an
fro
m
a b
i
g
g
er wi
n
dow affects th
e im
ag
e
fid
e
lity, un
n
e
cessary
in
crease of
wi
n
dow-size
t
h
ou
gh
un
corrup
ted
p
i
x
e
ls
are
in
a sm
al
ler windo
w and
co
m
p
utatio
n
a
lly th
is filter is
co
stly [
1
], [4
].
Akko
u
l
et.al.,
p
r
op
o
s
ed
t
h
e
Ad
ap
tiv
e
Swit
ch
ing
Med
i
an
Filter (ASMF), wh
ich u
s
es decisio
n
and
co
rrectio
n wi
nd
ows t
h
at are
ad
ap
tiv
e to
effectiv
ely fi
n
d
i
m
p
u
l
se p
o
s
ition
s
an
d
si
g
n
al restorers. The
i
m
ag
e
fid
e
lity o
f
th
e resto
r
ed
o
u
t
pu
ts is b
e
tter at h
i
g
h
e
r
and
l
o
wer im
p
u
l
se n
o
i
se ratios. Th
is filter red
u
c
es
un
necessa
ry
i
n
crease i
n
wi
n
d
o
w
si
ze an
d t
h
e i
m
pul
se res
t
ori
n
g
val
u
e i
s
fr
om
am
ong
t
h
e nea
r
est
rel
i
abl
e
in
ten
s
ities wh
i
c
h
g
i
v
e
s b
e
st
po
ssib
l
e
restorat
io
n
ev
en
in h
i
gh
ly corru
p
t
ed
en
v
i
ron
m
en
t [9
]
.
Deci
si
o
n
B
a
se
d
Al
g
o
ri
t
h
m
s
(DB
A
s)
we
re
i
n
t
r
o
d
u
ced
by
bot
h S
r
i
n
i
v
a
s
a
n
et
.al
.
,
an
d
M
a
dh
u et
.al
.
,
wi
t
h
di
ffe
rent
app
r
oaches
,
w
h
i
c
h
det
ect
co
rr
upt
e
d
si
g
n
al
s
by
chec
ki
n
g
t
h
em
t
o
be bet
w
een m
i
nim
u
m
and
m
a
xim
u
m
of t
h
e m
e
di
an det
e
ct
ed nei
g
h
b
o
r
h
oo
d.
B
o
t
h
fet
c
h a
rel
i
a
bl
e m
e
di
an
fr
om
nei
g
hb
o
r
h
o
od
f
o
r
o
n
l
y
t
h
e
corrupted si
gnals. There
f
ore, their ap
p
r
o
aches work
efficien
tly well fo
r al
l typ
e
s o
f
im
a
g
es up
to
5
0
% n
o
i
se
l
e
vel
s
. The l
i
m
i
t
a
t
i
ons suc
h
a
s
im
pro
p
er a
n
a
l
y
ze of im
pul
s
e
det
ect
i
on a
n
d t
h
e ab
sence
of
val
i
d
m
e
di
an f
o
rce
th
eir algorith
ms to
rep
l
ace th
e sign
al with prev
iou
s
ly re
st
ored
value
.
T
h
es
e problem
s
m
a
ke the
horizontal and
d
i
ago
n
a
l streak
s in
rest
o
r
ed i
m
ag
es. Fu
rt
herm
o
r
e, th
ese
filters d
o
no
t co
nsid
er t
h
e preserv
a
tion
of i
m
ag
e
det
a
i
l
s
[1
0]
,
[1
1]
.
Aiswarya et.al., p
r
op
osed
t
h
e Decision
Based
Un
symmetri
c
Trimmed
Me
d
i
an
Filter (DBUTMF)
for
rem
ovi
ng
hi
gh
den
s
i
t
y
im
pul
se n
o
i
s
es i
n
i
m
ages a
n
d
vi
deo
s
,
w
h
i
c
h
o
v
erc
o
m
e
t
h
e pr
o
b
l
e
m
of st
reaki
ng
effect
s
in
DB
As. In
t
h
is algo
rith
m
th
e left an
d ri
g
h
t
ex
tream
e v
a
lu
es of th
e
sto
r
ed
array
ob
tain
ed fro
m
th
e
3
×
3
window a
r
e impulse
val
u
es a
n
d are
tr
imme
d. The
corrupted pixel
is re
placed
by the
median
of the
re
sultant
ar
r
a
y.
Ev
en
thou
gh
t
h
is ap
pr
oach
is
b
e
tter
t
h
an D
B
A
s
, it d
o
e
sn’t preserv
e
th
e im
ag
e d
e
tails at h
i
gher
n
o
i
se
d
e
nsities [12
]
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 2, A
p
ri
l
20
16
:
61
1 – 6
2
0
61
4
3.
PROPOSE
D
EFFECTIVE
NOISE
ADA
PTIVE ME
DIAN FILTER(E
NAMF)
In t
h
e p
r
op
ose
d
m
e
t
hod, t
h
e
si
ze of t
h
e
w
i
nd
ow i
s
fi
xe
d
,
h
o
we
ve
r, t
h
e
effect
i
v
e m
e
di
an m
a
y
be
d
i
fferen
t
fro
m
th
e v
a
lu
e at t
h
e mid
d
l
e o
f
t
h
e sorted
p
i
x
e
l v
a
lu
es. Th
e
p
r
op
o
s
ed
effectiv
e
m
e
d
i
an
filter is
designe
d
to
diminish the problem
f
aced by
the standard median
filter a
nd othe
r Ada
p
tive Median Filte
rs. The
pr
o
pose
d
al
g
o
r
i
t
h
m
i
s
t
h
e
m
odi
fi
cat
i
on
of
D
eci
si
on B
a
sed
Al
g
o
ri
t
h
m
(D
B
A
) o
f
Sri
n
i
v
asan et
.al
.
It
r
e
st
ores
th
e d
i
g
ital i
m
ag
es co
rrup
ted at h
i
gh
o
r
low im
p
u
l
se
no
ise ratio
s
b
y
switch
i
ng
o
n
l
y
th
e filtration
o
f
t
h
e
co
rrup
ted
im
ag
e sig
n
a
ls with
a
m
u
ch
reliab
l
e
m
i
d
-rank
ing
statistics v
a
lu
e to
k
eep
u
p
th
e sig
n
a
l con
t
en
t
o
f
th
e
restored im
age. Furthe
rm
ore the hori
zontal and
diagonal streaks in t
h
e DBA is rectifi
e
d
in
th
e
p
r
opo
sed
al
go
ri
t
h
m
by
r
e
st
ori
n
g t
h
e c
o
rrect
pi
xel
val
u
es f
r
om
t
h
e nei
g
h
b
o
ri
ng
pi
xel
s
i
n
t
h
e
ker
n
el
wi
n
d
o
w
.
Fi
gu
re
1.
B
l
oc
k
Di
ag
ram
of P
r
o
p
o
sed
m
e
t
hod
Th
e
b
l
o
c
k
d
i
ag
ram
o
f
pro
posed
filter is sho
w
n
in
Figu
re
1
and
ex
p
l
an
at
o
r
y steps of the p
r
op
o
s
ed
al
go
rith
m
for the
gray sc
ale and col
o
r images are a
s
follows.
3
.
1 A
l
go
rit
h
m-1
a
Input :
Gray
Scale Im
age Im
g
Out
put
:
De
noi
sed Im
age
b
Step
1
:
Set slidin
g
windo
w size
W
mi
n
=3×3, noisy im
age a an
d rest
ore
d
image
b
Step
2
:
Read the p
i
x
e
ls fro
m
th
e
wind
ow and sto
r
e it in
S
Step
3: Com
p
ute S
mi
n
, S
max
S
med
and Np
Step 4:
If S
mi
n
<a(i,j)<S
ma
x
, w
h
ere a
(
i
,
j
)
i
s
t
h
e pr
ocessi
ng c
e
nt
ral
pi
xel
,
then
it is co
n
s
i
d
er as un
co
rrup
ted
p
i
x
e
l
an
d retain
ed
. Oth
e
rwise go
t
o
step
5
.
Step 5: If S
mi
n
<
S
me
d
< S
ma
x
, w
h
e
r
e
S
me
d
is
t
h
e m
e
d
i
an
v
a
lu
e o
f
S, th
en
it is co
n
s
id
er as co
rrup
ted
p
i
xel and
replace b(i
,
j) by
S
me
d
. Oth
e
rwise go
to step
6.
Step 6:
If
Np>
=
5 and
b(i,j-1)=0, the
n
it is cons
i
d
er a
s
corrupte
d
pi
xel
and re
place b(i,j) by S
mi
n
. If N
p
>=
5
and
b
(
i
,
j
-
1
)
=
2
55
, t
h
e
n
re
pl
a
ce t
h
e co
rr
upt
ed pi
xel
b
(
i
,
j
)
by
S
ma
x
. Otherwise re
place the b(i,j)
by the
m
ean
val
u
e o
f
pre
v
i
o
usl
y
p
r
oce
ssed
pi
xel
s
b(i
-
1
,
j
)
a
n
d
b
(
i
,
j-
1)
.
Step 7:
If Np<
5
the
n
re
place
the b(i,j)
by S
me
d
.
Step
8
:
Rep
eat
th
e abov
e step
s fo
r all th
e p
i
x
e
ls in
th
e im
ag
e
3.
2
Al
g
o
ri
thm
-
1b
In
p
u
t: RGB Im
age
Im
g
Ou
t
p
u
t: No
ise Filtered
im
ag
e
MLFI
Step1:
Input the RGB im
age
Im
g = im
read(<
RGB Im
age>)
Step
2
:
Sp
lit the i
m
ag
e in
to three layers
n
a
mely Red
Ch
an
nel, Green Ch
ann
e
l and
Bl
u
e
C
h
ann
e
l.
M
L
I
(
0
)
=
R
e
d(
Im
g)
M
L
I
(
1
)
=
G
r
e
e
n(
Im
g)
MLI(
2)
= Blue(
I
m
g
)
MLI
=
∫
Re
d(
I
m
g)+Green
(Im
g)+Blue
(Im
g)
Step3:
Take
ea
ch layer a
n
d c
h
eck
for im
pulse
noise in each pixels
usi
n
g 3×3
kernel
window.
Step
4
:
Ap
p
l
y t
h
e
p
r
op
o
s
ed
effectiv
e
n
o
i
se ad
ap
tiv
e m
e
d
i
an
filter w
ith
an
app
r
op
riate
v
a
lu
e
fro
m
an
accep
ted
nei
g
hb
o
r
h
o
o
d
t
o
R
e
d C
h
a
n
nel
an
d
ot
he
r set
o
f
val
u
es t
o
Gre
e
n a
n
d
B
l
ue C
h
an
nel
s
.
M
L
F(I
)
=
EN
AM
F(M
L
I(
I)
).
U
n
c
o
r
r
upt
e
d P
i
x
e
l
N
o
is
y P
i
xe
l
Mod
i
f
i
ed
P
i
xel
Input Noisy
Imag
e
Impuls
e
Detection
Res
t
ored
image
Original
Pix
e
l
Effect
ive
Est
i
m
a
tion of
Med
i
an
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Effective No
ise Ad
ap
tive
Med
ian
Filter f
o
r Remo
ving
Hig
h
Den
s
ity Imp
u
l
se No
ises
…
(S
.
Ab
du
l
S
a
leem
)
61
5
Step
5
:
C
o
n
cat
en
ate th
e filtered
three ch
an
nels MLFI = cat(MLF(i
))
Step
6
:
Disp
lay th
e no
ise
filtered co
l
o
r im
ag
e
Th
e t
w
o d
e
d
i
cated
step
s of t
h
e propo
sed filter are:
St
ep
1:
Aa
da
pt
i
v
e det
ect
i
o
n
o
f
i
m
pul
si
ve l
o
c
a
t
i
ons i
n
t
h
ree
chan
nel
s
.
Step 2: Corre
ction of the
detected im
pul
sive pixels
with an a
p
propriate value
from
an
acceptable
n
e
igh
bor
hoo
d fr
o
m
th
e w
i
n
dow
o
n
thr
ee ch
an
n
e
ls.
Th
e
p
r
o
p
o
s
ed
filter h
a
s ad
ap
tiv
e d
e
tection
o
f
i
m
p
u
l
se no
ises th
at lead
s to b
eco
m
e
m
a
x
i
m
u
m sig
n
a
l
extraction and im
pulse rest
ori
ng
value is
from
am
ong
the nea
r
est rel
i
able intensities give
best
possible
r
e
stor
atio
n ev
en
in
h
i
gh
ly corr
up
ted env
i
ronmen
t up
to
9
0
%
no
ise
lev
e
l. Th
e ho
riz
ontal
and diagonal streaks
are v
e
ry less wh
en
co
m
p
ared
with
o
t
h
e
r ad
ap
tiv
e no
n-
lin
ear filters. Th
e perfo
r
m
a
n
ce o
f
th
e filterin
g
pro
cess
is qu
an
tified
by u
s
ing
m
e
trics su
ch
as Peak Sign
al to
N
o
is
e Ratio (P
SNR
)
, M
e
a
n
S
q
uar
e
Er
ro
r
(M
SE)
,
Ro
ot
M
ean Sq
uare
Err
o
r
(RM
S
E)
, Tim
e
f
actor, Im
age Enhanc
e
m
ent Factor (IEF
) and the Struct
ural Similarity
Ind
e
x
(SSIM
)
t
h
at clearly sho
w
t
h
e
b
e
ttermen
t of
o
u
r
p
r
o
p
o
s
ed
effectiv
e
n
o
n
lin
ear filter fro
m
o
t
h
e
r adap
tiv
e
filters. Th
e abov
e sai
d
m
e
trics
are
repr
esen
ted
in equ
a
tion
(1
) th
rou
g
h
(5
).
)
255
(
log
10
2
10
MSE
PSNR
(1
)
M
i
N
j
n
m
R
n
m
O
MN
MSE
11
2
)]
,
(
]
,
(
[
)
1
(
(2
)
2
/
1
]
11
2
)]
,
(
]
,
(
[
)
1
[(
M
i
N
j
n
m
R
n
m
O
MN
RMSE
(3
)
M
i
N
j
M
i
N
j
n
m
O
n
m
D
n
m
O
n
m
N
IEF
11
2
11
2
))
,
(
)
,
(
(
))
,
(
)
,
(
(
(4
)
)
)(
(
)
2
)(
2
(
2
2
2
1
2
2
1
2
C
C
C
C
SSIM
R
O
R
O
OR
R
O
(5
)
Wh
ere,
O is th
e orig
in
al imag
e; R, th
e
re
st
ore
d
i
m
age;
D, de
-
noi
se
d i
m
age;
µ
O
and µ
R
are the
avera
g
es of O and
R respecti
v
ely;
σ
O
2
and
σ
R
2
are varianc
e
s of
O and R
respectively;
σ
OR
is the correlation
coefficient
between O a
n
d R;
C
1
and C
2
are sm
a
ll co
n
s
tan
t
s fo
r stab
ilize th
e co
m
p
u
t
atio
n
;
C
1
=(k
1
+L)
2
C
2
=
(k
2
+L)
2
k
1
=0
.01
an
d k
2
=0
.03
b
y
d
e
fau
lt; L=2
5
5
.
4.
RESULTS
A
N
D
DI
SC
US
S
I
ON
The
per
f
o
r
m
a
nce of t
h
e p
r
o
p
o
se
d EN
AM
F
al
go
ri
t
h
m
for
vari
ous i
m
ages at
di
ffe
rent
noi
se l
e
vel
s
vary
i
n
g fr
om
1
0
%
t
o
90
%
i
s
t
e
st
ed by
usi
n
g M
A
TLAB
.
R
e
sults of
both gray scale
and c
o
lor sta
n
dard images
are sh
ow
n i
n
t
h
e Fi
g
u
res 2
,
1
1
an
d 12 re
spe
c
t
i
v
el
y
.
Fi
gure
2 (a) i
s
t
h
e
Lena.jpg
im
age corrupted
with 30% of
salt an
d
p
e
p
p
e
r n
o
i
se. Th
e same i
m
ag
e is re
sto
r
ed
w
ith
VMF, SMF, ROAMF, DB
A and
p
r
op
osed
filter are
sh
own
in
Figur
e 2
(
b
)
, 2
(
c)
, 2(
d)
, 2(
e)
and 2(f) respectively
.
The sam
e
im
age corrupte
d
with
90% of s
a
lt and
p
e
pp
er no
ise an
d
restored
with
VM
F, SMF, ROAMF, DBA and
p
r
op
osed
filter are sh
own
in
Figu
re 2
(
g
)
,
2(
h)
, 2
(
i
)
,
2
(
j
)
,
2(
k)
an
d
2(l
)
resp
ectively.
Sim
i
larly, The c
o
lor im
age
Rose.jpg
wi
t
h
b
o
t
h
2
0
% a
n
d
90
% of
noi
se
de
nsi
t
y
,
Pepp
er.jpg
imag
e
with
bo
th 30
% and
90% no
ise
d
e
n
s
ities are sh
own in
Figu
re 11
an
d 12
resp
ectiv
ely. Th
ese im
ag
es were restored
with
VSM,
SMF, ROAMF, DB
A and
Pro
p
o
s
ed
Filter are sh
own
in
Fig
u
r
e
11
(
a
)
t
h
rou
g
h
11
(
l
) an
d Figur
e
1
2
(
a
)
thr
oug
h 12(
l)
r
e
sp
ectiv
ely.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 2, A
p
ri
l
20
16
:
61
1 – 6
2
0
61
6
(a)
(b
)
(c)
(d
)
(e)
(
f)
(g
)
(h
)
(i)
(j)
(k
)
(l)
Figure
2. (a
)
Gray-Scale Le
na
Im
age with
30% Noise
De
nsity and sam
e
im
age restore
d
with,
(b)
VM
F, (c)
SM
F, (d
)
R
O
A
M
F,
(e
) DBA
,
(f
)
P
r
op
ose
d
ENAM
F Alg
o
r
ith
m
,
(g
) Gray
-
S
cale
Lena
Im
age with 90
% Noise
D
e
n
s
ity an
d same i
m
ag
e r
e
stor
ed w
ith,
(h)
VMF,
(
i
) SMF,
(j
) ROA
M
F,
(k
) D
B
A,
(
l
)
Pro
p
o
s
ed
EN
A
M
F
Algo
rith
m
Tab
l
es
1
thro
ug
h 8 show th
e
q
u
a
n
titativ
e measu
r
es
an
d
t
h
eir correspo
n
d
i
n
g
graph
s
are
sh
own
in th
e
Fi
gu
res
3 t
h
r
o
ug
h
10
. The
v
a
ri
at
i
ons
of P
S
NR
an
d S
S
I
M
m
e
t
r
i
c
s of t
h
e pr
o
p
o
s
ed
ENAM
F al
g
o
r
i
t
h
m
i
n
gra
p
hs
gi
ve
n i
n
Fi
gu
res
3,
5
,
6,
8 a
n
d
10
cl
earl
y
sh
ow
t
h
e e
ffect
i
v
e
n
ess
o
f
ou
r
pr
o
pos
ed
a
l
go
ri
t
h
m
.
Tabl
e
1. C
o
m
p
ari
s
o
n
of
PS
N
R
val
u
es
o
f
di
f
f
ere
n
t
al
g
o
ri
t
h
m
s
fo
r
Lena
.jpg
i
m
ag
e at d
i
fferen
t no
ise
d
e
n
s
ities (%)
Noise
Density
VMF
SMF
ROAM
F
DBA
ENAMF
10
40.
898
0
39.
796
8
45.
326
8
45.
333
8
45.
379
8
20
40.
927
3
39.
792
4
44.
466
5
44.
473
7
44.
505
0
30
40.
873
4
39.
815
6
43.
652
4
43.
698
4
43.
789
9
40
40.
535
6
39.
847
4
42.
928
0
42.
946
9
42.
996
1
50
40.
022
0
39.
957
1
42.
279
5
42.
370
2
42.
449
5
60
39.
307
7
40.
155
3
41.
746
0
41.
758
1
41.
935
9
70
38.
701
5
40.
539
4
41.
206
8
41.
383
7
41.
476
6
80
38.
277
0
41.
164
8
40.
609
9
40.
663
6
40.
684
0
90
37.
921
42.
407
39.
834
7
39.
926
0
39.
996
Tabl
e 2.
C
o
m
p
ari
s
o
n
of
M
S
E val
u
es
o
f
di
ffe
r
e
nt
al
g
o
ri
t
h
m
s
fo
r
Lena
.jpg
i
m
ag
e at d
i
fferen
t no
ise
d
e
n
s
ities (%)
Noise
Density
VMF
SMF
ROAM
F
DBA
ENAMF
10
5.
2878
6.
814
1.
9072
1.
9086
1.
9028
20
5.
2523
6.
8209
2.
3357
2.
3343
2.
3332
30
5.
3179
6.
7846
2.
8044
2.
8027
2.
8004
40
5.
748
6.
7351
3.
3134
3.
2991
3.
2879
50
6.
4696
6.
5671
3.
8470
3.
8353
3.
8238
60
7.
6262
6.
2741
4.
3499
4.
3468
4.
3406
70
8.
7685
5.
7431
4.
9549
4.
9412
4.
9293
80
9.
112
4.
9728
5.
6506
5.
5811
5.
5265
90
10.
494
3
3.
7355
6.
9012
6.
7996
6.
5360
Tabl
e 3.
C
o
m
p
ari
s
o
n
of
SS
IM
val
u
e
s
of di
f
f
e
r
ent
al
go
ri
t
h
m
s
fo
r
Lena
.jpg
i
m
ag
e at d
i
fferen
t no
ise
d
e
n
s
ities (%)
Noise
Density
VMF
SMF
ROAM
F
DBA
ENAMF
10
0.
7417
3
0.
3982
7
0.
9784
5
0.
9781
1
0.
9780
3
20
0.
6669
4
0.
4228
1
0.
9682
1
0.
9682
8
0.
9683
9
30
0.
4170
3
0.
3967
4
0.
9481
5
0.
9482
0.
9482
5
40
0.
2167
7
0.
2972
2
0.
9271
1
0.
9258
2
0.
9251
7
50
0.
1095
7
0.
1783
9
0.
8957
2
0.
8917
3
0.
8996
1
60
0.
0591
9
0.
1119
2
0.
8549
7
0.
8450
0.
8629
4
70
0.
0391
4
0.
0726
2
0.
7957
2
0.
7742
8
0.
8134
5
80
0.
0275
4
0.
0457
6
0.
7108
9
0.
6587
0
0.
7467
0
90
0.
0108
0
0.
0242
0
0.
5478
4
0.
4314
8
0.
5771
4
Tabl
e
4. C
o
m
p
ari
s
o
n
of
PS
N
R
val
u
es
o
f
di
f
f
ere
n
t
al
g
o
ri
t
h
m
s
fo
r
Ro
s
e
.
j
p
g
C
o
lor im
ag
e at differen
t
n
o
i
se
den
s
ities (%)
Noise
Densit
y
VMF
SMF
ROAM
F
DBA
ENAMF
10
39.
311
7
37.
841
9
44.
761
3
43.
742
5
44.
023
9
20
39.
337
2
37.
842
6
43.
406
3
42.
824
8
42.
939
5
30
39.
280
9
37.
859
5
42.
351
3
41.
934
6
41.
974
7
40
38.
978
9
37.
903
2
41.
433
9
41.
112
6
41.
160
3
50
38.
438
0
37.
988
7
40.
758
9
40.
517
4
40.
507
7
60
37.
690
8
38.
201
4
39.
846
4
39.
800
9
39.
924
6
70
36.
974
8
38.
535
5
39.
208
2
39.
218
3
39.
297
4
80
36.
306
7
39.
207
1
38.
619
1
38.
657
6
38.
690
3
90
35.
764
8
40.
542
8
37.
658
0
37.
681
9
37.
970
5
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Effective No
ise Ad
ap
tive
Med
ian
Filter f
o
r Remo
ving
Hig
h
Den
s
ity Imp
u
l
se No
ises
…
(S
.
Ab
du
l
S
a
leem
)
61
7
Tabl
e 5.
C
o
m
p
ari
s
o
n
of
IE
F v
a
l
u
es of di
f
f
ere
n
t
al
g
o
ri
t
h
m
s
fo
r
Ro
s
e
.
j
p
g
colo
r im
ag
e at d
i
fferen
t no
ise
d
e
n
s
ities (%)
Noise
Density
VMF
SMF
ROAM
F
DBA
ENAMF
10
64.
941
6
8.
6023
1847.
8
3
0
1483.
3
0
3
1373.
0
3
1
20
37.
710
6
9.
0624
728.
25
37
613.
40
74
514.
80
36
30
14.
204
8
4.
3171
408.
08
00
373.
12
23
245.
84
48
40
5.
8017
7.
6858
207.
57
04
205.
05
94
144.
30
70
50
2.
7089
5.
8298
136.
88
41
137.
12
17
91.
608
10
60
1.
5172
3.
9963
91.
144
50
75.
896
70
53.
552
80
70
0.
9815
9
2.
7281
56.
456
50
40.
384
40
34.
050
00
80
0.
7061
2
1.
8728
33.
928
40
23.
035
40
23.
123
00
90
0.
5583
7
1.
3293
12.
176
40
8.
3254
0
9
.
3831
Tabl
e 6.
C
o
m
p
ari
s
o
n
of
SS
IM
val
u
e
s
of di
f
f
e
r
ent
al
go
ri
t
h
m
s
fo
r
Ro
s
e
.
j
p
g
colo
r im
ag
e at d
i
fferen
t no
ise
d
e
n
s
ities (%)
Noise
Density
VMF
SMF
ROAM
F
DBA
ENAMF
10
0.
8879
4
0.
6534
0.
9878
1
0.
9883
5
0.
9901
7
20
0.
8286
9
0.
6640
6
0.
9820
5
0.
9827
3
0.
9843
4
30
0.
6587
2
0.
6515
1
0.
9722
4
0.
9737
5
0.
9741
5
40
0.
4199
7
0.
6097
1
0.
9603
2
0.
9630
8
0.
9673
9
50
0.
2197
4
0.
5329
5
0.
9414
6
0.
9425
9
0.
9496
2
60
0.
1023
6
0.
4617
6
0.
9218
9
0.
9147
7
0.
9249
8
70
0.
0459
6
0.
4014
4
0.
8812
6
0.
8698
8
0.
8986
5
80
0.
0245
1
0.
3503
9
0.
8179
9
0.
7910
3
0.
8484
2
90
0.
0133
9
0.
3001
0.
6617
2
0.
6149
0.
7257
6
Tabl
e 7.
C
o
m
p
ari
s
on
o
f
IEF
val
u
es
o
f
di
ffe
r
e
nt
al
g
o
ri
t
h
m
s
fo
r
Pepp
er.jpg
co
lor im
ag
e at d
i
fferen
t
no
ise
d
e
nsities
Noise
Densit
VMF
SMF
ROAM
F
DBA
ENAMF
10
48.
058
4
12.
980
5
853.
22
9
976.
04
07
861.
69
7
20
32.
359
6
13.
379
2
456.
12
95
470.
88
66
405.
91
1
2
30
14.
380
3
12.
726
7
264.
84
90
264.
64
91
218.
67
6
0
40
6.
6947
10.
064
8
183.
27
9
180.
89
45
139.
48
8
4
50
3.
4134
6.
9723
125.
88
87
116.
10
01
93.
064
7
60
2.
0618
4.
5126
87.
707
0
77.
860
6
63.
650
3
70
1.
3939
2.
9145
60.
875
7
50.
461
9
44.
075
8
80
1.
0705
1.
9490
37.
209
5
27.
922
2
28.
560
7
90
0.
9069
1.
3489
19.
687
8
12.
748
0
14.
420
0
Tabl
e 8.
C
o
m
p
ari
s
o
n
of
SS
IM
val
u
e
s
of di
f
f
e
r
ent
al
go
ri
t
h
m
s
fo
r
Pepp
er.jpg
co
lor im
ag
e at d
i
fferen
t
no
ise
d
e
nsities
Noise
Densi
VMF
SMF
ROAM
F
DBA
ENAMF
10
0.
8360
4
0.
5646
1
0.
9902
1
0.
9883
1
0.
9888
8
20
0.
7580
7
0.
5779
7
0.
9721
8
0.
9694
0.
9698
3
30
0.
5425
4
0.
5505
3
0.
9704
1
0.
9681
3
0.
9699
8
40
0.
3137
5
0.
4530
4
0.
9507
4
0.
9510
3
0.
9515
8
50
0.
1676
3
0.
3337
8
0.
9329
2
0.
9278
9
0.
9306
8
60
0.
0942
3
0.
2357
8
0.
8982
5
0.
8955
8
0.
9082
1
70
0.
0537
3
0.
1742
6
0.
8592
9
0.
8408
7
0.
8787
7
80
0.
0307
4
0.
1293
6
0.
7851
2
0.
7549
3
0.
8178
4
90
0.
0195
4
0.
0390
6
0.
6432
2
0.
5727
3
0.
6898
2
Fi
gu
re
3.
N
o
i
s
e De
nsi
t
y
ver
s
us P
S
NR
f
o
r
Gray
-
S
cal
e
Len
a
Im
ag
eat d
i
fferen
t
no
ise
d
e
nsities
Fi
gu
re
4.
N
o
i
s
e De
nsi
t
y
ver
s
us M
S
E
fo
r
G
r
ay
-
Scale Len
a
Im
ag
e at
d
i
fferen
t no
ise
d
e
n
s
ities
Fi
gu
re
5.
N
o
i
s
e De
nsi
t
y
ver
s
us S
S
IM
f
o
r
G
r
ay
-Scal
e
Len
a
Im
ag
e at d
i
fferen
t
no
ise
d
e
nsities
Fi
gu
re
6.
N
o
i
s
e De
nsi
t
y
ver
s
us P
S
NR
f
o
r
R
o
se
Im
ag
e at d
i
fferen
t
no
ise
d
e
n
s
ities
34
36
38
40
42
44
46
10
20
30
40
50
60
70
80
90
PSNR
(dB
)
Noise
Density
(%)
VMF
SMF
ROAM
F
DB
A
Proposed
0
2
4
6
8
10
12
10
20
30
40
50
60
70
80
90
MSE
Noise
Density
(%)
VMF
SMF
ROAM
F
DB
A
Proposed
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
10
20
30
40
50
60
70
80
90
SSIM
Noise
Density
(%)
VMF
SMF
ROAM
F
DB
A
Proposed
0
10
20
30
40
50
10
20
30
40
50
60
70
80
90
PSNR
(dB
)
Noise
Density
(%)
VMF
SMF
ROAM
F
DB
A
Proposed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 2, A
p
ri
l
20
16
:
61
1 – 6
2
0
61
8
Fi
gu
re 7.
N
o
i
s
e
De
nsi
t
y
ver
s
us IEF f
o
r col
o
r
R
o
s
e
Im
ag
e at d
i
fferen
t
no
ise
d
e
n
s
ities
Fi
gu
re 9.
N
o
i
s
e
De
nsi
t
y
ver
s
us IEF f
o
r Pe
p
p
er
Im
age
at d
i
fferen
t
no
i
s
e d
e
n
s
ities
Fi
gu
re
8.
N
o
i
s
e De
nsi
t
y
ver
s
us S
S
IM
f
o
r c
o
l
o
r
Ro
se
Im
ag
e at d
i
fferen
t
no
ise
d
e
nsities
Fi
gu
re 1
0
. N
o
i
s
e
De
nsi
t
y
ve
rs
us
S
S
IM
f
o
r Pe
ppe
r
Im
ag
e at d
i
fferen
t
no
ise
d
e
n
s
ities
(a)
(b
)
(c)
(d
)
(e)
(
f)
(g
)
(h
)
(i)
(j)
(k
)
(l)
Fig
u
re 11
. (a) Ro
se Im
ag
e
with
20
%
No
ise Den
s
ity
an
d same i
m
ag
e r
e
stor
ed w
ith (b
) VMF,
(
c
)
SM
F,
(d
)
ROAM
F
,
(e
)
D
B
A,
(f
)
Pr
op
o
s
ed
EN
AM
F
Alg
o
rithm
(g
)
Rose
Im
age with 9
0
%
N
o
ise
Density
a
n
d
sa
m
e
im
age
resto
r
ed
with (
h
) VM
F,
(i) SM
F, (
j
)
ROAM
F
,
(k
) D
B
A, (l)
P
r
op
os
ed
EN
AM
F A
l
go
rithm
0
500
1000
1500
2000
10
20
30
40
50
60
70
80
90
IEF
Noise
Density
(%)
VMF
SMF
ROAM
F
DB
A
Proposed
0
200
400
600
800
1000
1200
10
20
30
40
50
60
70
80
90
IEF
Noise
Density
(%)
VMF
AMF
ROAM
F
DB
A
Proposed
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
10
20
30
40
50
60
70
80
90
SSIM
Noise
Density
(%)
VMF
SMF
ROAM
F
EASMF
Proposed
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
10
20
30
40
50
60
70
80
90
SSIM
Noise
D
en
sity
(%)
VMF
SMF
ROAM
F
DB
A
Proposed
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
An
Effective No
ise Ad
ap
tive
Med
ian
Filter f
o
r Remo
ving
Hig
h
Den
s
ity Imp
u
l
se No
ises
…
(S
.
Ab
du
l
S
a
leem
)
61
9
(a)
(b
)
(c)
(d
)
(
e
) (
f)
(g
)
(h
)
(i)
(j)
(k
)
(l)
Figure 12. (a) Pepper
Im
age
with 30% Nois
e
De
nsity
an
d
sa
m
e
i
m
ag
e r
e
sto
r
ed
w
ith
(b
)
V
M
F,
(
c
) SMF,
(d
)
ROAM
F
,
(e
)
D
B
A,
(f
)
Pr
op
o
s
ed
EN
AM
F
Alg
o
rithm
,
(g
)
Pep
p
er
Im
age with
90
%
Nois
e De
nsity
an
d s
a
m
e
im
age resto
r
ed
with
(
h
)
VM
F,
(i)
SM
F,
(j) R
OAM
F
,
(k
) D
B
A, (l)
P
r
op
os
ed
E
N
AM
F Algo
rithm
Th
e Tab
l
e
1
an
d
Tab
l
e 4
clearly sh
o
w
the PSNR v
a
l
u
es o
f
th
e filtered
i
m
ag
es from d
i
fferen
t
alg
o
rith
m
s
, wh
ich
realize the p
r
eserv
a
tion o
f
im
ag
e q
u
a
lity o
f
ou
r prop
o
s
ed
ENAM
F. Tab
l
e 2 sho
w
s th
e
MSE of th
e filtered
im
ag
es fro
m
d
i
fferen
t
al
g
o
rith
m
s
,
wh
ich
realize
ou
r
p
r
o
p
o
s
ed
al
g
o
rith
m
h
a
s th
e m
i
n
i
m
u
m
erro
r
rate wh
en
co
m
p
ared
with
o
t
h
e
r
filterin
g
resu
lts
. The SSIM
v
a
lu
es o
f
th
e tested imag
es are shown in
Tabl
es 3,
6 an
d 8. F
r
om
t
h
ese t
a
bl
es and t
h
ei
r cor
r
esp
o
ndi
ng
gra
p
h
s
, Fi
g
u
re 5
,
8 an
d 1
0
sho
w
s t
h
e bet
t
erm
e
nt
o
f
o
u
r propo
sed
filter wh
en
co
m
p
ared
with o
t
h
e
r
n
o
n
-
li
n
e
ar filters. Th
e
IMF v
a
lu
es o
f
th
e p
r
o
p
o
s
ed
filter
resem
b
l
e
s wi
t
h
DB
A,
whi
c
h
are gi
ve
n i
n
T
a
bl
e 7 an
d
9
re
spectively. The streaking
effect such as horizontal
an
d d
i
ag
on
al str
eak
s t
h
at
n
o
rmall
y
o
ccur
in D
B
A
s
ar
e r
e
ctif
ied
b
y
cor
r
ect selectio
n
of
th
e n
e
i
g
hbo
rhood
p
i
x
e
ls in
ou
r
propo
sed
filterin
g
al
g
o
rith
m
wh
ich
in
t
u
rn
g
i
v
e
s a
b
e
tter
v
i
su
al
p
e
rcep
tio
n
as shown
i
n
figu
res
2(l
)
, 11
(l
)
a
n
d 12
(l
).
5.
CO
NCL
USI
O
N
AN
D F
U
T
U
RE D
I
RE
CTI
O
N
In
th
is p
a
p
e
r, a n
e
w effectiv
e n
o
i
se adap
tiv
e m
e
d
i
an
filter is p
r
op
o
s
ed
wh
ich
g
i
v
e
s b
e
tter
perform
a
nce in com
p
arison
with VM
F, SM
F
,
ROAM
F a
n
d DBA in term
s
of P
S
NR, MS
E, RMSE, SS
I
M
and
IEF m
e
trics. Th
e
p
r
o
p
o
s
ed
al
g
o
rith
m
is faster th
an
ROAMF since it us
es a sm
all and fixe
d
window of size
3
×
3.
In add
ition
,
it affects a sm
o
o
t
h
tran
siti
o
n
b
e
t
w
een
the p
i
x
e
l
v
a
l
u
es
b
y
u
tilizin
g the correlation
between
nei
g
hb
o
r
i
n
g p
r
ocesse
d pi
xel
s
whi
l
e
p
r
ese
r
v
i
ng e
dge
det
a
i
l
s t
hus l
eadi
n
g
t
o
bet
t
e
r e
d
g
e
prese
r
vat
i
o
n
.
The
p
r
op
o
s
ed
filter is tes
t
ed
fro
m
lo
w to
h
i
g
h
noise d
e
n
s
itie
s o
n
d
i
fferen
t g
r
ayscale i
m
ag
es an
d
co
lo
r im
ag
es th
at
yield recogniz
a
ble and patches free re
storation. Th
e signi
ficant di
ffe
rence i
n
PSNR, SSIM a
n
d visua
l
p
e
rcep
tion
with
o
t
h
e
r co
m
p
etitiv
e filters q
u
a
n
tifies a d
o
m
i
n
an
ce
o
f
th
e pro
p
o
s
ed
filter. In
fu
t
u
re, fu
zzy
lo
g
i
c
b
a
sed
ad
ap
tiv
e switch
i
n
g
m
e
d
i
an
filter
will play th
e do
m
i
n
a
n
t
ro
le in d
i
g
ital i
m
ag
e restoratio
n
.
REFERE
NC
ES
[1]
Rafael C
.
Gonzalez and
Rich
ard
E. Wood.
Digita
l Image
Processing
. 3
rd
Edition, Prentice-Hall, 200
9.
[2]
Abdul Saleem S, and Abdul Razak T. Survey
o
n
Color
Image Enhancement
Techniqu
es using Spatial Filter
i
n
g
.
International Jo
urnal of Co
mputer Applic
ations (
I
JCA)
. 2014; 94(
9): 39-45.
[3]
Manglem Sigh
Kh., Bora Prabin K.,
and B
i
ren
d
ra Singh. S.
V
ector
Median
Filter
for Removal of Impulse Noise
from Color Imag
es.
IU-Journal o
f
Elec
trical and
Electroni
cs Engineering
. 2004; 4
(
1): 1063-1072
.
[4]
Hwang H. and
Hadded R.A. Adaptiv
e Median
Filter: New al
gorithm
s
and results.
IEEE T
r
ansactions. Image
Processing.
199
5; 4(4): 499–502
.
[5]
Browning D.R.K. The Weigh
t
ed Median
Filter.
Communications of
the ACM.
19
84; 27(8): 807-8
18.
[6]
Zhou
Wang and
David Zhang
.
Progressive
Switching
Median
Filter
for Rem
o
val of
Im
pulse
noise from
High
l
y
Corrupted Im
ag
es
.
IEEE Transaction s
on Circuit and S
y
stems-
II, Ano
l
og and
Digital S
i
gnal
Processing.
1999;
46(1).
[7]
Z. W
a
ng,
A.C.
Bovik, H.R
.
S
h
eikh and
E.P
.
S
i
m
oncelli
. Im
age
Qualit
y As
s
e
s
m
ent: F
r
om
error m
eas
urem
ent t
o
struc
t
ural simila
rity
.
IEEE Transactions on
Image
Processing
. 200
4; 13(1).
[8]
Kwa
me
Ose
i
Boa
t
i
n
g,
Be
nj
a
min Wey
ori Asuba
m a
nd Da
vid
Sanka Laar. Improving the Eff
ective n
e
ss of the
M
e
dian F
i
l
t
er
.
I
n
ternational Jou
r
nal of Electr
onic and Communication
Engin
eering
. 2012; ISSN 0974-2166; 5(1)
:
85-97.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
6, No
. 2, A
p
ri
l
20
16
:
61
1 – 6
2
0
62
0
[9]
Akkoul S., Led
ee. Roger
,
Leco
nge. R, Harba
R. A New
Adaptive Switching
Median Filter
.
Signal Processing
Letters.
IE
EE
. 2
010; 17(6).
[10]
S
r
inivas
an K.S
.
and Ebene
zer D
.
A new fas
t
and effici
ent decision based algorithm
for removal of high density
impulse noise.
I
EEE signal processing
. 2007; 14
(3):189-192.
[11]
Nair Madhu S.,
Revath
y
K.
and
Tatava
rti Rao
.
“Removal of Salt and Peppe
r Noise in Images: A new Decision-
Based Algorith
m
”,
Proceeding
of
Int
e
rnational
Multi
Con
f
eren
ce
of
Engin
eers a
nd Computer Scientists
. 2008; 1:
19-21: IMECS.
[12]
Ais
w
ar
y
a
K
., J
a
yar
a
j V
.,
Eben
e
zer D.
A New
and Efficient Algorithm for th
e R
e
moval of High
density
Salt and
Pepper Noise in
Images and Vid
e
os.
IEEE
. 2010;
978-0-7695-3941.
BIOGRAP
HI
ES OF
AUTH
ORS
Abdul Saleem has done B.Sc. Ph
y
s
ics from St. John’s College,
Pa
lay
a
mkottai, M.Sc.
Ph
y
s
ics
from Jamal Mohamed Colleg
e
(Aut
onomous), Tiruchir
appalli, Af
f
iliated
to
Bharath
i
dasan
University
, MC
A from Manon
maniam
Sundaranar University
, Tir
unelveli. C
u
rrently
h
e
is
working as an Assistant Professor of Computer
Science, Jamal
Mohame
d College and th
e Part
Tim
e
Res
e
arch
S
c
holar
in th
e s
a
m
e
coll
ege
.
His
area
of
inte
res
t
i
s
Im
age P
r
oces
s
i
ng
Abdul Razak is
working as an Associate Professor of Com
p
u
t
er Sci
e
nce
,
Ja
m
a
l Moham
e
d
College, Tiru
chirappalli. He has
a teaching exp
e
r
i
ence of 27
y
e
ar
s. He has been guiding M.Phil.
S
c
holars
for pas
t
18
years
and P
h
.D S
c
holars
for
pas
t
thre
e
yea
r
s
.
His
areas
of r
e
s
earch in
ter
e
s
t
includ
e Network Security
, Image Processing
and Data Mining. He has published around 20
research
ar
tic
les
in int
e
rna
tiona
l
j
ournals.
Evaluation Warning : The document was created with Spire.PDF for Python.