Int
ern
at
i
onal
Journ
al of Ele
ctrica
l
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
,
No.
6
,
D
ece
m
ber
201
8,
pp. 502
1~50
31
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp5021
-
50
31
5021
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Pattern
Ap
proxi
mation B
ased
Gen
erali
zed Image
Noise
Reducti
on
U
sing Ad
aptive F
ee
df
orw
ar
d
Neur
al Netwo
rk
Nag
araj Bh
at
1
, U
.
Er
an
n
a
2
, Ma
noj
Kum
ar
Singh
3
1
Depa
rtment of
El
e
ct
roni
cs
&
C
om
m
unic
at
ion
E
ngine
er
ing, R
V
Coll
ege of
engi
n
ee
ring
,
Ind
ia
2
Hea
d
of
Depa
r
t
m
ent
,
Dep
art
m
e
nt
of El
ec
tron
ic
s
&
Com
m
unic
ati
on
Engi
n
ee
ring
,
BITM,
Indi
a
3
Dire
ct
or
,
Manu
ro
Tech
R
ese
ar
c
h
Pvt. Lt
d
.
,
India
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n
30
, 2
01
8
Re
vised
Ju
l
2
,
201
8
Accepte
d
J
ul
22
, 2
01
8
The
p
robl
em
of
noise
interfe
r
ence
with
th
e
image
al
wa
y
s
oc
cur
s
i
rre
spec
t
iv
e
of
whate
ve
r
pre
ca
ut
ion
is
ta
k
en.
Chal
l
engi
ng
iss
ues
with
noise
r
educ
t
ion
ar
e
dive
rsit
y
of
cha
r
ac
t
eri
sti
cs
involved
with
source
of
noise
and
in
result
;
it
is
diffi
cu
lt
to
dev
el
op
a
univ
ersa
l
soluti
on.
Thi
s
pape
r
h
as
prop
osed
neur
a
l
net
work
base
d
gene
ra
li
z
e
sol
uti
on
of
noise
red
uction
b
y
m
appi
ng
the
proble
m
as
pat
t
ern
appr
oximat
i
on.
Consideri
ng
the
statistical
r
el
a
ti
onship
among
loc
al
re
gion
pixels
in
the
noise
fre
e
image
as
norm
al
pa
tt
e
rns
,
fee
dforward
neu
ral
n
et
work
is
a
ppli
ed
to
ac
quir
e
th
e
knowledg
e
av
ai
l
able
withi
n
such
pat
t
ern
s.
Adapti
ven
e
ss
is
appl
ie
d
in
the
slope
of
tra
nsfer
func
ti
o
n
to
improve
the
le
arn
ing
proc
ess.
Acquir
ed
nor
m
al
pat
te
rns
kn
owledge
is
uti
lized
to
red
u
ce
th
e
l
eve
l
of
diffe
ren
t
t
y
p
e
o
f
noise
avail
able
withi
n
an
image
b
y
r
ec
orr
ec
t
ion
of
nois
y
pat
t
ern
s
through
pat
te
rn
appr
oxi
m
at
ion.
T
h
e
p
roposed
restor
a
ti
on
m
et
hod
does
not
nee
d
an
y
esti
m
at
ion
of
noise
m
odel
c
har
a
cteri
sti
cs
ava
ilable
in
the
i
m
age
not
onl
y
t
hat
it
ca
n
red
u
ce
the
m
ixe
r
of
diffe
ren
t
t
y
pes
of
noise
eff
ic
i
ent
l
y
.
The
p
roposed
m
et
hod
has
high
proc
essing
spee
d
al
ong
wit
h
sim
pli
ci
t
y
in
design.
Restor
ation
of
gra
y
s
ca
l
e
image
as
well
as
col
or
i
m
age
has
d
one,
which
has
suffere
d
from
diffe
r
ent
t
y
pes
of
noise
l
ike
,
Gauss
ia
n
no
ise, sal
t
&
pepe
r, spec
k
le n
oise
and
m
ixe
r
o
f
it.
Ke
yw
or
d:
Ad
a
ptive
slo
pe
Feed
forw
a
r
d
ar
chite
ct
ur
e
Neural
Netw
ork
No
ise
r
e
du
ct
io
n
Patt
ern
a
ppr
ox
i
m
at
ion
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Nag
a
ra
j
Bhat
,
Ma
noj
K
um
ar S
ingh
,
Ma
nuro Tec
h R
esearch
P
vt.
Ltd,
#20,2
nd
Cr
os
s
, Jy
oth
i
Nag
a
r,
Vidyara
nyap
ura, Ban
galo
re, I
nd
ia
.
Em
a
il
:
m
ks
ing
h@
m
anu
r
orese
arch.com
1.
INTROD
U
CTION
No
ise
a
rises
as
a
res
ult
of
m
od
el
le
d
or
unm
od
el
la
ble
pr
oces
ses
ha
ppeni
ng
within
t
he
pr
oductio
n
a
nd
captu
rin
g
of
a
real
sign
al
.
It
is
no
t
a
pa
rt
of
the
pe
rf
ect
sign
al
a
nd
is
c
ause
d
by
a
va
riet
y
of
sou
rce
s
li
ke
var
ia
ti
on
wit
hin
the
detect
or
sensiti
vity
,
en
vir
on
m
ental
va
riat
ion
s,
the
natu
re
of
rad
ia
ti
on
,
tra
ns
m
issi
on
or
qu
a
ntiza
ti
on
er
rors
et
c.
The
c
har
act
erist
ic
s
of
no
ise
rely
on
it
s
su
pply
.
S
ever
al
im
age
pr
oces
sin
g
pac
ka
ges
con
ta
in
op
e
rat
or
s
wh
ic
h
ad
d
the
no
ise
arti
f
ic
ia
ll
y
to
a
pic
ture.
Del
i
ber
at
el
y
cor
ruptin
g
an
i
m
age
with
no
ise
per
m
it
s
us
to
check
the
resist
ance
of
a
picture
a
nd
asse
ss
the
perform
ance
of
th
e
im
age
by
var
i
ous
no
ise
filt
ers.
N
oise
c
an
be
cl
assi
fied
into
t
wo
cat
e
gories
in
wh
ic
h
the
fi
rst
one
dep
e
nds
up
on
i
m
age
know
le
dg
e
a
nd
the
sec
ond o
ne
is in
dep
e
ndent
of
im
age k
no
wled
ge.
Ther
e
a
re
sev
e
ral
diff
e
re
nt
ty
pes
of
noise
c
an
ap
pear
unde
r
the
cat
eg
ory
of
im
age
data
dep
e
ndent
no
ise
li
ke
detect
or
noise
,
s
pe
ckle
noise
,
sal
t
&
peppe
r
noise
and
po
issi
on
no
ise
.
Each
ty
pe
of
noise
has
the
i
r
own
s
ource
of
ori
gin
an
d
t
he
ir
ef
fect
ove
r
the
im
age
are
al
so
uniq
ue.
Detect
or
noise
present
i
n
ne
arly
al
l
recorde
d
im
age
s
an
d
the
m
ain
reas
on
of
thi
s
no
ise
is
th
e
di
screte
natu
re
of
rad
ia
ti
on.
S
uch
kind
of
no
ise
can
m
od
el
ed
as
an
add
it
ive
m
odel
throu
gh
Ga
us
sia
n
distrib
ut
ion
.
S
pec
kle
no
ise
is
ca
us
e
d
by
an
er
r
or
in
the
process
of
tra
nsm
issi
on
an
d
corrupted
pix
el
s
ob
ta
ine
d
ei
the
r
with
the
m
axi
m
u
m
value
or
hav
i
ng
sin
gle
bits
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5021
-
5031
5022
flipp
e
d
over
.
S
al
t
&
pep
per
noise
is
caused
by
an
error
in
the
transm
issi
on
process
.
Corru
pted
pi
xels
by
thi
s
no
ise
are
ob
ta
i
ned
by
the
extrem
e
value
ei
t
her
at
m
axi
m
u
m
or
m
ini
m
u
m
.
Po
isso
n
noise
occurs
beca
us
e
of
nonlinea
rity
ava
il
able in the
det
ect
or
a
nd r
ec
order.
Re
storatio
n
of
i
m
age
is
an
i
m
po
rtant
a
nd
chall
eng
i
ng
pr
ob
le
m
in
i
m
ag
e
pro
cessi
ng.
Pr
act
ic
al
ly
,
add
it
ive
Ga
us
s
ia
n
no
ise
a
nd
Im
pu
lse
no
ise
m
od
el
can
represent
the
va
ri
ou
s
differe
nt
ty
pes
of
no
ise
s
that
app
ea
r
i
n
the i
m
age.
I
n
the
ca
se of a
dd
it
ive
t
he
Gau
s
sia
n n
oi
se
m
od
el
,eac
h pixel
of the im
age gets t
he
c
hange
thr
ough
zer
o
-
m
ean
Ga
us
sia
n
distrib
ution
value.
S
uch
noise
ca
n
be
ef
fect
i
vely
rem
ov
ed
by
l
ocal
a
ver
a
ging
process
.
Linear
filt
er
li
ke
Gau
ssian
filt
er
can
re
m
ov
e
su
c
h
no
ise
in
an
e
ff
i
ci
ent
m
ann
er
but
blu
r
the
ed
ges.
I
n
i
m
pu
lse
no
ise
m
od
el
,
on
ly
a
certai
n
per
ce
ntage
of
the
tot
al
pix
el
s
pre
sent
in
the
i
m
age
get
a
cha
ng
e
by
rand
om
v
al
ue a
nd k
ee
ping t
he
r
em
ai
nin
g pixels as it
.
Affect
ed
pix
el
s
with
s
uc
h
noise
have
ve
ry
diff
e
re
nt
inten
sit
y
in
com
par
ed
to
their
nei
ghbors
.
It
i
s
diff
ic
ult
to
re
m
ov
e
the
nois
e
us
in
g
Ga
us
si
an
filt
ers
,
b
eca
us
e
s
harp
c
ha
nges
in
pi
xel
in
te
ns
it
y
are
trea
te
d
as
edg
e
.
To
rem
ov
e
su
c
h
ty
pe
of
no
ise
,
In
m
os
t
cases,
Me
dian
or
ra
nk
sta
ti
sti
cs
based
co
nc
epts
are
app
l
ie
d
to
dev
el
op
the
filt
ers
.
But
su
c
h
filt
ers
cannot
eff
ic
i
ently
rem
ov
e
the
Gaussi
an
noise
.
I
n
re
su
lt
,
in
m
os
t
o
f
th
e
conve
ntion
al
m
et
ho
ds,
to
re
m
ov
e
the
noi
s
e
in
any
recei
ve
d
im
age,
first
no
ise
c
harac
te
risti
cs
avail
able
in
th
e
i
m
age
ha
ve
to
est
i
m
a
te
and
then
filt
er
has
to
be
desig
ne
d
accor
dingly
.
Thi
s
is
tim
e
c
on
s
um
ing
an
d
costl
y
process
,
not
only
that
there
is
no
gu
a
ra
ntee
of
de
sired
s
uc
cess.
The
re
is
ver
y
le
ss
resea
rch
has
bee
n
done
i
n
the
de
sig
n
of
filt
er
w
hich
c
ou
l
d
ta
ke
care
of
im
age
no
i
se
irres
pecti
ve
of
their
m
od
el
char
act
e
ris
ti
cs.
T
he
s
it
utati
on
is
m
or
e
c
riti
cal
when
the
re
is
m
ix
ture
of
no
ise
a
vaila
ble
in
the
i
m
age,
for
ex
a
m
ple
,
if
no
isy
i
m
age
transm
i
t
te
d
thr
ough
the
f
a
ulty
tran
sm
issi
on
li
nes.
In
this
pa
pe
r,
know
le
dg
e
-
ba
sed
ap
proac
h
us
in
g
arti
fici
al
neural
netw
ork
(
A
NN)
has
app
li
ed
to
ov
e
rc
om
e
the
issues
avail
abl
e
with
conv
e
nt
ion
al
no
ise
re
storatio
n
ap
proach
e
s.
A
NN
has
the
capa
bi
li
ty
to
acqu
i
re
the
know
le
dg
e
from
inputs
thr
ough
le
arn
in
g
proce
dure
a
nd
pro
ve
n
to
be
unive
rs
al
approxim
at
o
r
[
1].
In
t
his
pa
per
per
ce
ptr
on
bas
ed
m
ulti
layer
feedfo
r
ward
ar
chite
ct
ur
e
has
app
li
ed
t
o
re
s
tore
t
he
im
ages.
Th
e
slop
e
of
act
iva
ti
on
f
un
ct
io
n
pl
ay
s
the
ver
y
im
po
rtant
ro
le
in
de
fining
the
conve
rg
e
nce
qual
it
y
of
le
arn
i
ng.
A
le
ss
slop
e
can
cause
sm
al
l
chan
ge
s
in
outp
ut
even
ther
e
is
a
high
c
ha
ng
e
in
in
pu
t
w
hile
high
value
of
s
lop
e
will
h
ave
the
re
v
erse
ef
fect.
Hen
ce
, r
at
he
r
than
le
ar
ning wi
th p
re
determ
ine f
ixed
sl
op
es
f
or
all
acti
vation
f
un
ct
io
ns
, o
pt
i
m
a
l value
of
sl
op
e
for
ea
ch
act
ivati
on
f
un
ct
io
n
has
ac
hieve
d
in
eac
h
it
erati
on
of
le
arn
i
ng
by
m
aki
ng
t
hem
sel
f
a
dap
ti
ve
.
To
un
der
sta
nd
the
ben
e
fit
of
a
dap
ti
ve
sl
op
e
s
over
fi
xe
d
one,
a
f
unct
ion
ap
pro
xim
at
ion
prob
le
m
has
consi
der
e
d
a
nd
ob
se
r
ved
t
hat
there
is
a
sig
nificant
im
pr
ov
e
m
ent.
Lat
er,
a
n
ada
ptive
a
rc
hitec
ture
has
a
pp
li
e
d
to
le
ar
n
the
c
or
relat
ion
patte
r
ns
avail
able
i
n
norm
al
i
m
age
pix
el
s
with
their
su
r
rou
nd
i
ng
pi
xels
usi
ng
gr
a
di
ent
m
et
ho
d.
Acqui
red
know
le
dge
has
util
iz
ed
to
re
-
co
rr
ect
the
relat
ion
s
hip
w
hich
ha
s
bee
n
distor
te
d
beca
us
e
of
the
prese
nce
of
noise
.
T
his
know
le
dg
e
base
d
recorrecti
on
does
n’
t
dep
e
nde
nt
up
on
t
he
m
od
el
of
no
ise
a
nd
ca
n
rem
ov
e
near
ly
al
l
t
ypes
of
noise
s
in
their
i
nd
i
vidual
pr
e
s
ence
or
in
a
m
ixed
f
or
m
.
Th
e
op
ti
m
al
siz
e
of
the
arch
it
ect
ure
ha
s
deci
ded
with
com
par
at
ive
pe
rfor
m
ances
over
va
rio
us
a
rc
hitec
tures
a
nd
ob
s
er
ved
that
s
m
al
le
r
siz
e
co
uld
be
t
he
bette
r
c
ho
ic
e.
T
he
propose
d
c
o
nce
pt
of
noise
re
storatio
n
has
al
s
o
a
pp
l
ie
d
ove
r
c
olo
r
i
m
ages
al
so
by providi
ng the i
nd
i
vidu
al
trainin
g ov
e
r
the a
vaila
ble
pri
m
ary color
in
form
ation
.
Ther
e
are
a
nu
m
ber
of
resear
cher
s
,
w
ho
has
giv
e
n
at
te
ntio
n
to
wards
noise
reducti
on.
In
[
2
]
,
im
age
denoisin
g
has
been
do
ne
th
r
ough
the
im
age
com
po
sit
ion.
It
is
a
pp
li
ed
to
fi
nd
the
c
om
po
nen
ts
of
im
age
achi
e
ve
d
in
m
ov
i
ng
fr
am
e
that
def
ine
s
it
s
local
sp
at
ia
l
inf
or
m
at
ion
.Co
nc
ept
of
global
filt
ering
has
pro
po
s
ed
in
[
3
]
by
est
im
at
ing
each
pi
xel
as
a
funct
ion
of
al
l
a
vai
la
ble
pix
el
s
in
the
i
m
age.
Ba
sed
on
nonloc
al
sel
f
si
m
il
arity
and
the
lo
w
ra
nk
a
ppr
ox
im
at
ion
a
i
m
age
denois
ing
m
et
ho
d
ha
s
bee
n
pr
opos
e
d
in
[
4
].
Whi
te
zero
-
m
ean
Gau
s
sia
n
noise
has
be
en
rem
ov
ed
t
hro
ugh
spa
rse
and
redu
nd
a
nt
rep
rese
ntati
on
in
[
5
]
.
A
filt
er
i
s
intende
d
in
[
6
]
to
enhance
th
e
Kuwa
har
a
fil
te
r
to
re
du
ci
ng
the
no
ise
.
The
plan
ned
Ga
bor
K
uw
a
ha
ra
filt
er
i
n
[
6
]
is
eco
no
m
i
cal
to
scal
e
back
the
noise
w
hile
no
t
losi
ng
the
data
on
the
edg
es
,
be
fore
the
pr
e
pa
rati
on
of
t
he
pictures
for se
gm
entat
ion
and alt
er
native i
m
age pro
ces
sing t
ech
niques.
The
p
e
rfor
m
ance
of
m
any
nonlinear
fil
te
rs
for
noise
reducti
on
in
i
m
age
sign
al
s
has
bee
n
inv
est
igate
d
in
[
7
].
Ac
quired
pictures
in
hy
per
s
pectral
im
aginati
on
s
quare
m
easur
e
disturbe
d
by
a
dd
it
ive
no
ise
that
ty
pi
cal
ly
assum
es
as
zer
o
-
m
ean
wh
it
e
on
e
.
In
fact,
the
re'
s
sti
ll
non
-
w
hite
noise
in
hy
per
s
pectral
pictures
(
HSIs
).
T
he
2
-
dim
e
ns
io
nal
filt
erin
g
ways
a
nd
m
ulti
di
m
ension
a
l
te
ns
or
dec
om
po
sit
ion
al
gorithm
s
cannot ta
ke
awa
y non
-
w
hite n
oise from
H
SIs
d
irect
ly
. T
her
e
fore,
a
pre
wh
it
enin
g denoisi
ng a
ns
we
r
s
upporte
d
te
ns
or
deco
m
po
sit
ion
is
pla
nned
in
[
8
]
.
T
her
e
a
re
se
ve
r
al
i
m
po
rtant
be
nef
it
s
of
de
noisi
ng
the
m
agn
et
ic
resona
nce
(MR)
im
ages
in
the
a
rea
of
hea
lt
h
care.Base
d
on
nonlo
cal
li
kelihoo
d
filt
er
,
[
9
]
has
pro
po
sed
a
m
od
ifie
d
n
on
l
ocal
filt
er
f
or
noise
re
duct
ion
in
MR
i
m
ages.A
f
us
io
n
base
d
con
c
ept
has
be
en
a
pp
li
ed
i
n
[
10
]
to
re
d
uce
t
he
im
pu
lse
noise
f
r
om
i
m
ages
w
hi
ch
ha
ve
bee
n
captu
re
d
th
rough
m
ulti
sens
or
s
.
Ba
se
d
on
an
a
nt
colo
ny
al
gorithm
and
gen
et
ic
al
go
rithm
i
m
age
restor
at
i
on
has
pr
e
sen
te
d
in
[11].
Ar
ti
fici
al
fish
-
swar
m
al
gorithm
has
app
li
ed
in
[
12]
fo
r
a
m
edical
DR
i
m
age
enhancem
ent.
[1
3]
Has
propose
d
an
adap
ti
ve
m
edian
filt
ering
al
gori
thm
,
to
rem
ov
e
the
i
m
pu
lse
no
ise
s
in
the
color
im
ages.
Ba
sed
on
the
deco
m
po
sit
io
n
an
d
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Patte
rn
Ap
pro
xi
ma
ti
on B
as
e
d
Gen
e
ra
li
zed
Image Noise
R
edu
ct
io
n usi
ng
Ad
ap
ti
ve
…
(
Nagar
aj B
ha
t
)
5023
interl
eavin
g, [
14]
h
as
pro
pose
d
a tra
ns
m
issi
on
tech
nique
w
hi
ch
hel
p
s
to
r
es
tore
th
e recei
ve
d
im
age w
hic
h was
c
o
m
pr
essed
by
vect
or
qua
ntiza
ti
on
.
I
n
[15]
,
wav
el
et
base
d
weig
hted
m
edi
an
filt
er
ha
s
be
en
pro
posed
to
do
denoise
d
the
MR
I
im
age.
The
rest
of
the
pa
per
is
orga
nized
as
fo
ll
ows:
Sect
ion
2
descr
i
bes
the
con
ce
pt
of
propose
d
patte
r
n
appr
ox
im
at
ion
us
ef
ul
f
or
nois
e
reducti
on.
I
n
sect
ion
3
the
detai
ls
of
ap
plied
ada
ptive
slo
pe
on
the
act
iv
at
ion
functi
on
pr
ese
nted.
T
he
sec
ti
on
4
pr
e
sen
ts
the
detai
l
exp
e
rim
ent
al
resu
lt
s
a
nd
a
naly
sis
f
or
f
unct
i
on
appr
ox
im
at
ion
an
d n
oise
resto
rati
on for g
ray
as w
el
l as c
olor
im
ages.
Co
nc
lusio
n
has pr
es
ented
i
n
sect
io
n 5.
2.
NOI
SE
R
E
D
UC
TI
ON IN
I
MAGE
AS
P
ATTE
RN LE
ARNI
NG &
A
PPRO
X
I
M
A
TION
Tech
nica
l
ly
,
im
age
can
be
c
on
si
der
e
d
a
s
va
riat
ion
of
li
gh
t
intensit
y
that
form
s
a
group
of
patte
r
ns
i
n
sp
at
ia
l
do
m
ai
n.
The
se
patte
r
ns
c
ou
l
d
al
so
be
sim
pler
in
natu
re
or
form
com
plex
p
at
t
ern
s
de
pe
nd
i
ng
upon
var
ying
in
inte
ns
it
y.
I
n
the
di
gital
do
m
ai
n
area
of
t
hese
patte
rn
s
in
her
e
n
t
in
t
he
f
or
m
of
pi
xel
val
ue
s
,
it
'
s
te
rr
ibly
to
ugh
to
ou
tl
ine
these
patte
r
ns
t
hro
ugh
pi
xels
glob
al
ly
,
howe
ver
it
'
s
do
able
to
w
at
ch
a
nd
pe
rcei
ve
t
he
char
act
e
risti
cs
of
th
os
e
patte
r
ns
in
the
local
reg
i
on
w
he
rever
the
pix
el
ar
ea
un
it
extrem
el
y
cor
relat
e
and
ca
n
be
t
hought
as
a
nei
ghbor
hood
patte
rn
outl
ine
d
by
pix
el
s
i
n
t
his
re
gi
on
as
s
how
n
in
Fi
g.1
&Fig.2
.
C
om
plexity
and
data
acce
ssible
inside
t
he
patte
r
ns
not
so
le
ly
dep
en
d
up
on
t
he
posit
ion
of
native
re
gion
,
howeve
r
add
it
io
nally
size o
f
the
native
reg
i
on.
A
ve
ry
sm
al
l
s
pati
al
reg
io
n
won'
t
carry
hi
gh
i
nfor
m
at
ive
patte
rn
w
her
e
ver
as
te
rr
i
bly
m
assive
siz
e
can
ca
rr
y
t
o
s
ever
al
le
ss
co
r
relat
ed
in
f
or
m
at
ion
.
Thes
e
na
ti
ve
patte
r
ns
can
be
c
onsid
ered
as
s
om
e
kind
of
functi
onal
f
orm
and
the
c
omplet
e
im
age
is
t
he
set
of
s
uch
f
un
ct
io
ns.
I
f
t
he
re
is
good
div
e
rsity
avail
able
in
th
e
i
m
age,in
the
resu
lt
s
there
ar
e
diff
ere
nt
ty
pes
of
local
fun
ct
ion
in
the
co
rr
es
pondin
g
f
unct
ion
set
of
i
m
age.
Con
si
der
i
ng
the
uni
ver
sal
a
ppr
ox
im
at
ion
capab
il
it
y
of
f
eedw
a
r
d
ne
ura
l
netw
ork,
it
is
possible
to
t
r
ai
n
the
neural
netw
ork
to
un
de
rstan
d
the
approxim
ate
ver
sio
n
of
local
patte
rn
s
in
a
i
m
age.
Pr
eseance
of
a
ny
typ
e
of
no
ise
,d
e
st
ro
y
t
he
local
f
un
ct
i
on
al
c
ha
racteri
sti
cs,w
hic
h
ca
n
be
c
orrecte
d
by
the
trai
ne
d
netw
ork
.
I
n
resu
lt
,
there
is
a
redu
ct
ion
of
noise
le
vel
with
ou
t
pr
e
know
le
dge
of
noise
char
a
ct
erist
ic
s
through
patte
r
n
rec
ogniti
on
con
ce
pt.
Local
patte
rn g
e
ne
rated in
a
blo
c
k
as
shown i
n
Fi
gu
re
1
a
nd Fig
ur
e
2
.
Figure
1.
Local
p
at
te
r
n gen
e
ra
te
d
in a
b
l
oc
k
Figure
2.
Local
p
at
te
r
n gen
e
ra
te
d
in a
b
l
o
ck
3.
NEED F
O
R
A
DA
PTI
VE
SLOPE
ACTI
V
ATIO
N
F
UNCTIO
N
I
N
FF
ARCHIT
EC
TURE
The
in
vestigat
ion
of
ne
ur
al
ne
twork
a
ppr
ox
i
m
at
ion
capab
il
it
ie
s
ha
s
been
done
in
past
by
a
nu
m
er
of
researc
her
s
.
H
ornik
[
1
]
has
pro
po
se
d
t
he
pro
of
of
un
i
ve
rsal
ap
pro
xim
a
ti
on
in
a
m
or
e
ge
ner
al
iz
ed
m
ann
er
,
with
s
uffici
ent
num
ber
of
hid
de
n
unit
s
in
a
sing
le
hi
dd
e
n
la
ye
r
a
nd
ar
bitrary
bounde
d
act
ivati
on
functi
on
,
sta
nd
a
rd
m
ultilay
er
feedfo
r
w
ard
netw
orks
are
unive
rsal
appr
ox
im
at
or
with
re
sp
ect
t
o
LP
(μ)
perform
ance
crit
eria,
f
or
a
r
bitrar
y
fi
nite
input
en
vir
onm
ent
m
easur
es
μ.
Howe
ver,
he
sai
d
ver
y
cl
early
that
that
al
l
act
ivati
on
func
ti
on
s
ψ wil
l n
ot
p
er
form
eq
ua
ll
y well
in
sp
ec
ific
learni
ng pr
ob
le
m
s.
In
the
bac
kpr
op
a
gatio
n
le
ar
ning
proce
ss,
weig
ht
update
d
rate
is
pro
p
or
ti
onal
to
the
der
iva
ti
ve
of
nonlinea
r
trans
fer
f
unct
ion
a
va
il
able
with
act
ive
un
it
s.
I
n
m
os
t
of
case
s,
si
gm
oid
curve
w
hich
has
bell
sh
ap
e
d
der
i
vatives
ha
s
app
li
ed
to
MLP
ne
ur
al
network.T
her
e
m
ay
be
a
chan
c
e
that
in
the
tr
ai
nin
g
phase,
outp
ut
of
li
near
accum
ulator
m
ay
app
e
ar
in
the
sat
urat
ion
re
gion
of
act
ivati
on
fun
ct
ion
.T
he
de
ri
vative
of
act
ivati
on
functi
on
is
very
s
m
al
l
in
sat
ur
at
ion
reg
i
on
i
n
res
ult
le
arn
i
ng
rate
bec
ome
s
ver
y
slo
w.
I
n
res
ult,i
t
m
ay
ta
ke
a
la
rg
e
nu
m
ber
of
it
erati
on
s
t
o
c
om
e
ou
t
from
t
his
sat
urat
ion
r
egio
n.
T
he
po
s
sible
so
luti
on
of
this
pro
ble
m
is
to
increase
the
r
egio
n
of
the
un
s
at
ur
at
e
d
re
gion
by
decr
e
sing
the
slo
pe
of
act
ivati
on
functi
on.
H
oweve
r
,
0
10
20
30
40
50
60
70
175
180
185
190
195
200
205
210
215
P
i
x
e
l
s
N
o
.
P
i
x
e
l
s
V
a
l
u
e
0
10
20
30
40
50
60
70
145
150
155
160
165
170
P
i
x
e
l
s
N
o
.
P
i
x
e
l
s
V
a
l
u
e
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5021
-
5031
5024
decr
easi
ng
the
slop
e
will
m
ake
the
syst
em
clo
ser
to
a
li
near
m
od
el
,
w
hich
in
ef
fect
dim
inishes
the
ad
va
ntage
of
ha
ving
t
he
m
ul
ti
la
ye
r
network.
hen
c
e
t
her
e
is
an
opt
i
m
u
m
value
of
slo
pe
nee
ded
at
each
it
erat
ion
as
accor
ding
t
o
th
e
la
nd
sc
a
pe
de
fine
d
by
t
he
e
r
ror
f
unct
io
n.
A
gain
t
he
value
of
t
he
sl
op
e
f
or
act
ivati
on
f
unct
ion
is n
ot sam
e fo
r
all
the neu
rons
.
The
c
om
ple
xity
inv
olv
e
d
wit
h
MLP
does
not
hav
e
al
l
the
slop
e
val
ues
befor
e
trai
ning
com
m
ence
hen
ce
t
her
e
is
need
t
o
pro
vi
de
the
ada
ptive
m
e
chan
ism
wh
ic
h
has
t
o
ta
ke
care
of
slop
es
of
act
i
vation
functi
on.
The
process
f
or
a
da
ption
of
slo
pe
s
can
be
der
i
ved
sim
ultaneou
s
ly
with
we
igh
ts
optim
iz
ation
in
te
rm
s
of
m
inim
iz
at
ion
of
e
r
ror
f
unct
ion.
Sp
eci
fical
ly
,
the
slo
pes
a
re
to
be
c
hose
n
so
as
t
o
m
ini
m
iz
e
the
perform
ance cri
te
rion
E
q
=
1
2
(
d
q
−
x
out
(
s
)
)
T
(
d
q
−
x
out
s
)
=
1
2
∑
(
d
q
h
−
x
out
,
h
s
)
2
n
h
=
1
(1)
Wh
e
re
‘s
’
denotes
the
nu
m
ber
of
la
ye
rs
i
n
the
netw
ork
a
nd
d
q
∈
ℜ
n
×
1
an
d
x
out
s
are
the
desire
d
a
nd
act
ual
ou
t
pu
ts
,
resp
e
ct
ively
of
the
network
du
e
to
qth
trai
ning
patte
rn
.
Co
nsi
der
an
act
ivat
ion
f
un
ct
i
on
of
the
sigm
oid
ty
pe
gi
ven
by
(
2
).
f
(
u
,
k
)
=
1
(
1
+
e
−
ku
)
(2)
Wh
e
re
u
is
the
input
to
the
nonlinea
rlit
y
and
k
is
t
he
slo
pe
par
am
et
er
wh
i
ch
ha
s
to
be
a
djust
ed
s
o
that
(
1
)
has
to
m
ini
m
iz
ed.
Con
si
der
i
ng
t
he
no
nlinear
li
ty
of
the
it
h
ne
uro
n
in
the
s
th
la
ye
r
of
the
netw
ork,
gr
a
dien
t
appr
oach can
be ap
plied
by
obta
ining
k
i
s
(
t
+
1
)
=
k
i
s
(
t
)
−
β
∂
E
q
∂
k
i
s
(3)
Using t
he
c
hain ru
le
,
the sec
ond t
erm
o
n
t
he rig
ht side
i
n
(
3
)
ca
n be r
e
w
ritt
en
as
∂
E
q
∂
k
i
s
=
∂
E
q
∂
u
i
s
∂
u
i
s
∂
x
out
,
i
s
∂
x
out
,
i
s
∂
k
i
s
=
−
δ
i
s
1
∂
x
ou
t
,
i
s
∂
u
i
s
⁄
∂
x
ou
t
,
i
s
∂
k
i
s
=
−
δ
i
s
f
k
(
u
,
k
)
f
u
(
u
,
k
)
(4)
Wh
e
re
δ
i
s
is
the
local
er
ror
f
or
the
it
h
ne
uro
n
of
the
sth
la
ye
r,
an
d
f
k
(
u
,
k
)
an
d
f
u
(
u
,
k
)
de
no
te
the
pa
rtia
l
der
i
vatives of the acti
vation
f
un
ct
io
n
with
k and
u
res
pecti
ve
ly
. H
ence the
slop
e
of
the act
ivati
on
fu
nctio
n
can
be defi
ned b
y
k
i
s
(
t
+
1
)
=
k
i
s
(
t
)
(
5
)
To
i
ncr
ease
the
stabil
it
y,
m
o
m
entum
ter
m
is also a
dd
e
d.
Learn
i
ng alg
ori
th
m
w
it
h
a
dapt
ive act
ivati
on
functi
on slo
pes
In
it
ia
li
ze the we
igh
ts i
n
the
n
e
twork
acco
rd
i
ng to
stan
dard i
niti
al
iz
at
ion
p
r
ocess
Fr
om
set the se
t of
trai
ning
d
a
ta
, d
e
rive
t
he n
et
work res
ponse
Local er
ror
is
obta
in
ed
with
re
sp
ect
to
t
he
de
s
ired net
w
ork re
sp
onse
an
d ob
t
ai
ned
act
ual ou
tpu
t acc
ordin
g
to foll
owin
g
e
quat
ions f
or
ouput l
ay
er a
nd h
i
dd
e
n
la
ye
r
unit
s
for
ou
t
pu
t l
ay
e
r:
δ
i
s
=
(
d
q
−
x
out
,
i
s
)
g
(
u
i
s
)
for
hidden
lay
er:
δ
i
s
=
∑
δ
h
s
+
1
w
h
i
s
+
1
n
2
h
=
1
g
(
u
i
s
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Patte
rn
Ap
pro
xi
ma
ti
on B
as
e
d
Gen
e
ra
li
zed
Image Noise
R
edu
ct
io
n usi
ng
Ad
ap
ti
ve
…
(
Nagar
aj B
ha
t
)
5025
The wei
ghts
of the
netw
ork
a
r
e upd
at
e
d
acc
ordi
ng to
w
ij
s
(
t
+
1
)
=
w
ij
s
(
t
)
+
μ
δ
i
s
x
out
,
j
s
The
sl
op
e
of t
he
acti
vation f
unct
ion are
up
da
te
d
acc
ordin
g t
o
k
i
s
(
t
+
1
)
=
k
i
s
(
t
)
+
β
δ
i
s
+
α
[
k
i
s
(
t
)
−
k
i
s
(
t
−
1
)
]
Stop the it
erati
on if
netw
ork
c
onve
rg
e
d
,
el
se
go b
ac
k
t
o
ste
p 2.
4.
E
X
PERI
MEN
TAL
R
USUL
TS &
ANALY
SIS
The
ne
ural
network
is
trai
ne
d
with
la
rg
e
num
ber
of
local
patte
rn
s
a
vaila
ble
in
the
im
ag
e
to
m
ap
th
e
ou
t
pu
t
sim
il
ar
to
the
i
nput
to
get
the
knowle
dg
e
of
po
s
sible
var
ia
ti
on
with
in
local
patte
rn.
I
n
the
noisy
im
age
,
local
patte
rns
lose
the
i
nh
e
rent
fu
nc
ti
on
al
rel
at
ion
an
d
this
can
be
c
orrecte
d
with
tr
ai
ne
d
netw
ork.
Ca
re
has
to
be
ta
ke
n
on
t
he
learni
ng
cha
ra
ct
erist
ic
s
it
sh
ould
be ge
ner
al
i
zat
ion
no
t t
ow
ard
s
m
e
m
or
iz
a
ti
on
.
4.1. Func
ti
on
Ap
pr
oxim
ati
on
A
s
a
t
e
s
t
c
a
se
t
o
un
de
r
s
t
a
nd
t
he
be
ne
f
i
t
of
t
he
a
da
pt
i
ve
s
l
op
e
i
n
f
e
e
df
or
w
a
r
d
a
r
c
hi
t
e
c
t
ur
e
,
a
m
a
t
he
m
a
ti
c
al
fun
c
t
i
on
as
de
f
i
ne
d
t
hr
o
ug
h
(
6
)
ha
s
c
on
s
i
de
r
e
d
f
or
a
p
pr
ox
i
m
a
t
i
on
pu
r
p
os
e
.
F
r
om
f
un
c
t
i
on
,
40
0
s
am
pl
e
s
i
n
t
he
r
a
ng
e
of
[
0
4
]
ha
ve
ge
ne
r
a
t
e
d
am
on
g
w
hi
c
h
od
d
s
a
m
pl
e
s
ha
ve
t
a
ke
n
f
o
r
t
r
a
i
ni
ng
p
ur
po
s
e
,
w
hi
l
e
e
ve
n
s
a
m
pl
e
s
ha
ve
a
pp
l
i
e
d
i
n
t
he
t
e
s
t
c
a
s
e
.
A
r
c
hi
t
e
c
t
ur
e
of
[
1
,
5,
1
]
ha
s
de
ve
l
op
e
d
t
o
a
pp
r
o
xi
m
a
te
t
he
f
un
c
t
i
on
w
i
t
h
a
nd
w
i
t
h
o
ut
a
da
pt
i
ve
s
l
op
e
of
s
i
gm
oi
d
f
un
c
t
i
on
a
nd
o
bt
a
i
ne
d
pe
r
f
o
r
m
a
nc
e
s
ha
ve
s
ho
w
n
i
n
F
i
g
ur
e
3.
T
he
ob
t
a
i
ne
d
va
l
ue
of
M
S
E
ha
s
a
l
s
o
s
h
ow
n
i
n
T
a
bl
e
1.
I
t
i
s
c
l
e
a
r
t
ha
t
t
he
r
e
i
s
a
m
uc
h
be
t
t
e
r
l
e
a
r
ni
ng
ha
s
ha
p
pe
ne
d
w
i
t
h
t
he
a
da
pt
i
ve
s
l
op
e
i
n
c
o
m
pa
ri
s
on
t
o
f
i
xe
d
s
l
o
pe
of
t
h
e
s
i
gm
oi
d
f
un
c
t
i
on
.
T
he
f
i
na
l
ob
t
a
i
ne
d
va
l
ue
of
t
he
s
l
op
e
f
o
r
hi
d
de
n
l
a
y
e
r
a
nd
o
ut
pu
t
l
a
y
e
r
no
de
s
ha
s
a
l
s
o
s
h
ow
n
i
n
T
a
bl
e
2.
I
t
c
a
n
a
l
s
o
ob
s
e
r
ve
t
ha
t
s
uc
h
ki
nd
of
va
l
ue
s
f
o
r
t
he
s
l
op
e
s
i
s
no
t
p
os
s
i
bl
e
t
o
de
f
i
ne
t
hr
o
ug
h
t
he
m
an
ua
l
a
pp
r
oa
c
h.
=
−
(
)
+
0
.
3
(6)
Table
1
.
E
rro
r i
n
le
ar
ning
of
f
un
ct
io
n
a
ppr
ox
i
m
ation
Mean sq
u
are
e
rr
o
r
Mean sq
u
are
e
rr
o
r
NN (
St
atic slo
p
e)
0
.00
2
4
NN(Adap
tiv
e slo
p
e)
0
.00
0
2
6
Table
2
.
T
he
f
i
nal slo
pe value
of
hidden
and
ou
t
pu
t
node
ac
ti
ve
f
un
ct
io
n
Lay
e
r
Slo
p
e valu
e
Hid
d
en
[
2
.15
1
4
1
.05
7
1
2.3
5
6
1
1
.81
6
7
2.4
4
3
2
];
Ou
tp
u
t
[
3
.4051
]
F
i
gu
r
e
3.
F
un
c
t
i
on
a
p
pr
o
xi
m
at
i
on
by
ne
ur
a
l
n
e
t
w
or
k
0
50
100
150
200
250
300
350
400
450
0
0
.
1
0
.
2
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
X
Y
FX
S
L
O
R
G
A
D
S
L
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5021
-
5031
5026
4.2
. Im
age
De
no
isi
ng
Qu
al
it
y
of
le
ar
ning
is
hea
vily
de
pe
nd
i
ng
up
on
the
siz
e
of
the
arc
hitec
tur
e
an
d
t
he
i
nput
data
.
Th
e
la
rg
er
siz
e
of
the
arc
hitec
ture
m
ay
cause
easi
er
le
arn
in
g
w
hi
le
loo
se
the
ge
ner
al
iz
at
ion
ca
pab
il
it
y.
Hen
c
e
it
is
ver
y
necessa
ry
to
hav
e
com
pact
siz
e
of
arch
it
ect
ur
e
as
m
uch
as
possi
ble.
Ther
e
is
no
direct
way
to
find
out
the
optim
al
size
of
a
rch
it
ect
ur
e,
ge
ne
rall
y
var
io
us
di
ff
e
rent
siz
e
arch
it
ect
ur
e
has
te
ste
d
directl
y
or
in
di
rectl
y
li
ke
pru
ning m
et
hod.
4.2.1
.
Op
tima
l
Siz
e
of
the
A
rchi
tectu
re
To
c
ollec
t
the
op
ti
m
al
t
ype
of
local
im
age
patte
rn,
it
is
ne
cessary
that
r
efere
nce
im
age
m
us
t
carry
div
e
rse
featu
re
s.
H
ence
a
gr
a
y
scal
e
of
“
Le
na
”
im
age,
ha
vi
ng
512*
512
pi
xels
has
ta
ke
n.
L
ocal
patte
r
ns
ha
ve
gen
e
rated
th
rough
the
local
r
egio
n
pix
el
s
.
To
unde
rstan
d
the
optim
al
arc
hitec
ture
in
noi
se
reducti
on
diff
e
rent
po
s
sibil
it
ie
s
hav
e
ex
plore
d
as
sh
ow
n
in
Ta
bl
e
3.
It
is
obser
ved
t
hat
arc
hitec
ture
siz
e
of
[
4
1
4]
has
delivere
d
the b
et
te
r
r
e
du
ct
ion
in
noise i
n
c
om
par
ison t
o othe
r
arc
hite
ct
ur
e.
Table
3
.
PS
NR (db)
values
in no
ise
r
e
du
ct
i
on as fu
nction o
f a
rch
it
ect
ure siz
e
(Noise
densi
ty
:
0
.
05, NI:
nois
y im
age)
Architectu
re
[
I
H
O]
Salt &Pep
p
er
[
NI:18
.45
]
Gau
ss
ian
[
NI:19
.08
]
Sp
eckl
e
[
NI:18
.85
]
[
6
4
1
6
6
4
]
2
3
.77
9
7
2
2
.11
7
9
2
4
.26
3
8
[
6
4
8
6
4
]
2
4
.07
9
8
2
2
.11
6
4
2
4
.43
2
8
[
1
6
8
16
]
2
4
.50
1
7
2
2
.53
2
2
2
5
.02
3
2
[
1
6
4
16
]
2
5
.44
7
5
2
2
.90
4
9
2
5
.94
8
6
[
1
6
2 16
]
2
6
.34
5
2
2
3
.01
6
9
2
6
.73
7
0
[
4
2
4
]
2
5
.41
0
6
2
2
.70
3
9
2
5
.96
5
4
[
4
1
4
]
2
6
.42
3
5
2
3
.02
8
6
2
6
.95
1
9
4.2.2
.
Op
timal
Trainin
g
Im
age
The
ne
xt
im
po
rtant
issue
is
to
determ
ine
w
hat
co
uld
be
na
ture
of
trai
ni
ng
im
age:
No
r
m
al
i
m
age
(w
it
ho
ut n
oise
)
o
r
N
oisy
i
m
age.
In
norm
al
i
m
age
le
arn
in
g
sam
e
o
utput ha
s co
ns
i
der
e
d
f
or targets,
wh
il
e
n
oisy
i
m
age
le
arn
ing
as
an
inp
ut
,
c
orres
pondin
g
norm
al
par
t
ha
s
con
si
der
e
d
f
or
ta
r
gets.
T
hr
ee
ver
sio
n
s
of
no
isy
i
m
age h
ave
de
velo
ped
: i
m
age co
r
rupted by s
al
t &
pepper
noise
, Gaussia
n no
ise
a
nd m
ixt
ur
e
of
both.
Wi
th the
arch
it
ect
ure
of
[4
1
4]
for
al
l
the
fo
ur
dif
fere
nt
po
s
sible
trai
nin
g
im
ages
,
le
arn
ing
has
app
li
ed
in
de
penden
tl
y
and
obta
ine
d
pe
rf
o
r
m
ance
in
no
ise
re
du
ct
io
n
ov
e
r
sam
e
no
i
se
de
ns
it
y
an
d
diff
e
ren
t
noise
de
ns
it
y
ha
ve
sho
w
n
in
Table
4
a
nd
in
Table
5.
It
is
obser
ve
d
that
norm
al
i
m
age
as
a
re
fer
e
nce
i
m
age
for
le
ani
ng
has
delive
re
d
the
bette
r
pe
rfor
m
ances
in
no
ise
reducti
on.
He
nce,
a
rch
it
ect
ure
[
4
1
4]
an
d
no
ise
fr
ee
im
age
of
“Le
na”
has
consi
der
e
d
f
or
final
le
arn
in
g
pu
r
pose.
T
he
convergenc
e
char
act
e
risti
c
of
le
ar
ning
for
“Lena”
im
a
ge
has
sh
ow
n
in
Fig
ure
4.
It
is
obse
rv
e
d
that
t
her
e
is
ver
y
fast
c
onve
rg
e
nce
ha
s
ta
ken
place.
Af
te
r
com
pletio
n
of
le
arn
in
g;
dif
fere
nt
ty
pes
of
im
ages
with
dif
fer
e
nt
ty
pes
of
no
ise
s
li
ke
G
auss
ia
n
no
ise
,
Salt
&
Pepp
e
r
no
ise
,
S
pec
kle noise
and m
ixtur
es of them
h
ave
ap
plied
for
te
st c
ases.
Table
4.
PS
NR (db
) wit
h diff
e
ren
t
form
s o
f
tr
ai
nin
g re
fer
e
nc
e i
m
age ‘
Le
na wit
h [4 1
4
]
F
F
arc
hitec
ture
si
ze
and em
bed
de
d no
ise
d
e
ns
it
y:
0
.
05 [WN:
im
a
ge wit
hout
no
i
se;
N
I
(
S
&P
): Im
age w
it
h sal
t & p
e
pper
noise
; NI
(G
a
us
s.
): im
age w
it
h Ga
us
sia
n no
ise
;
NI (M
ix.)
:
Im
age w
it
h
a
m
ixu
tu
re
of noises]
Tr
ain
g
I
m
ag
e
Salt &
Pep
p
er
Gau
ss
ian
Sp
eckale
Salt &
Pep
p
er
+
Gau
ss
ian
Salt &
Pep
p
er
+Sp
eckale
Salt &
Pep
p
er
+ Gaus
sian
+Sp
eckl
e
WN
2
6
.
4
9
4
3
2
3
.
0
4
8
0
2
6
.
9
5
6
8
2
2
.
3
9
0
1
2
5
.
3
1
4
5
2
2
.
2
0
2
3
NI(
S
&
P)
2
5
.
1
2
0
8
2
4
.
9
2
0
0
2
5
.
5
7
9
3
2
3
.
7
6
1
9
2
4
.
1
2
3
9
2
3
.
0
6
5
0
NI
(
Gau
ss
.)
2
3
.
2
1
7
1
2
6
.
5
7
5
5
2
3
.
4
4
0
0
2
4
.
9
1
5
6
2
2
.
3
4
0
8
2
3
.
6
1
7
0
NI
(
Mix.)
1
9
.
7
9
0
4
2
2
.
9
3
5
0
1
9
.
9
4
5
1
2
2
.
0
3
8
4
1
9
.
3
8
1
5
2
1
.
1
1
5
4
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Patte
rn
Ap
pro
xi
ma
ti
on B
as
e
d
Gen
e
ra
li
zed
Image Noise
R
edu
ct
io
n usi
ng
Ad
ap
ti
ve
…
(
Nagar
aj B
ha
t
)
5027
Figure
4
.
Lear
ning c
onve
rg
e
nces c
har
act
e
risti
c
Table
5.
PS
NR (db)
ov
e
r
im
a
ge wit
h diff
e
re
nt noise
de
ns
it
y of i
m
age ‘
Lena
with
[4 1 4]
FF
ar
chite
ct
ure
size
No
ise
d
en
sity
Salt &Pep
p
er
Gau
ss
ian
Sp
eckl
e
Salt &
Pep
p
er
+
Gau
ss
ian
Salt &
Pep
p
er
+Sp
eck
le
Salt &
Pep
p
er
+ Gaus
sian
+Sp
eckl
e
0
.1
0
.01
NI
:
1
5
.43
2
9
TNI
:
1
9
.28
7
0
TWN
:
2
4
.56
9
2
NI
:
2
5
.41
6
1
TNI
:
2
0
.16
5
0
TWN:
28
.49
0
8
1
7
.08
9
4
2
4
.56
9
2
1
9
.02
3
2
2
0
.03
7
4
2
0
.67
2
5
2
6
.82
3
2
1
6
.01
2
5
1
9
.54
9
8
2
5
.49
4
9
2
5
.65
6
6
2
0
.23
2
4
2
8
.61
8
8
1
3
.82
6
8
1
8
.57
9
7
2
3
.11
6
8
1
9
.02
9
2
2
0
.54
9
5
2
6
.47
9
0
1
3
.34
7
3
1
8
.47
0
9
2
3
.20
8
2
2
2
.75
1
0
2
0
.12
2
3
2
8
.08
0
0
1
2
.15
0
1
2
0
.49
5
1
1
8
.97
3
3
1
8
.29
9
5
2
0
.43
5
0
2
6
.20
2
6
4.2.3
.
Gra
y
Sc
ale Im
ag
e
D
e
n
oising
Five
dif
fer
e
nt
im
ages
hav
e
co
ns
ide
red
in
this
pap
er f
or
the
diff
e
re
nt
exp
e
r
i
m
ent
pu
r
po
se as
sh
ow
n
in
the
Fi
g
ure
5.
W
it
h
eac
h
im
age
,
dif
fer
e
nt
ty
pe
s
of
no
ise
:
s
al
t
&
pe
pper
noise
,
Ga
us
sia
n
noise
,
Sp
e
ckl
e
noise
,
m
ixtur
e
of
sal
t
&
pepper
al
ong
with
Ga
ussi
an
a
nd
s
pec
kale
no
ise
ha
ve
ap
plied.
As
sh
ow
n
i
n
Fi
g
ure
6,
diff
e
re
nt
ve
rsi
on
of
a
noisy
i
m
age
in
1
st
r
ow
a
nd
de
noise
d
ver
si
on
of
t
he
co
rr
es
p
onding
im
age
by
propose
d
m
et
ho
d
ha
ve
s
how
n
in
2
nd
r
ow.
The
c
orres
pondin
g
P
SN
R
va
lues
for
no
isy
im
age
and
de
no
ise
im
age
ha
s
sho
wn
in
T
able
6
a
nd
Table
7
f
or
dif
fer
e
nt
noise
de
ns
it
y.
Perfor
m
ance
com
par
is
on
is
al
so
pr
es
e
nted
a
gainst
the
W
i
ner
filt
er
.
It
is
ob
s
er
ve
d
that
perform
ance
obta
ined
with
ne
ur
al
net
w
ork
is
m
uch
bette
r
in
c
om
par
is
on
to
the
Wine
r
filt
e
r
perform
ance.
I
t
is
interest
in
g
to
note
t
hat
wit
h
high
noise
le
vel
the
re
is
m
or
e
re
duct
io
n
in
noise
as
obse
r
ve
i
n
Table
7
in
c
om
par
ison
to
T
able
6.
It
is
al
so
obser
ved
th
at
m
ixtur
e
of
no
ise
has
bee
n
ha
nd
le
d
by
pr
opos
e
d
so
luti
on
ve
ry
eff
ic
ie
ntly
.
I
n
F
i
g
ur
e
7,
t
hr
e
e
d
i
f
f
e
r
e
nt
pa
t
t
e
r
ns
f
or
m
e
d
by
s
am
e
l
oc
a
ti
on
l
oc
a
l
pi
xe
l
s
f
r
om
no
i
s
e
f
r
e
e
im
a
ge
(
T
R)
,
no
i
s
y
im
a
ge
(
WN
)
a
nd
r
e
s
t
or
e
d
i
m
a
ge
(
A
P
)
ha
s
s
h
ow
n.
I
t
c
a
n
ob
s
e
r
ve
t
ha
t
c
or
r
e
c
t
e
d
pa
t
t
e
r
n
i
s
ve
r
y
c
l
os
e
t
o
t
he
t
r
ue
pa
t
t
e
r
n.
F
i
g
ur
e
5.
I
m
a
ge
s
c
on
s
i
de
r
e
d
f
or
e
x
pe
r
i
m
e
nt
s
(
f
r
om
le
f
t
t
o
r
i
gh
t
)
:
L
e
na
,
E
l
i
a
ne
,
B
oa
t
,
V
e
g
e
t
a
bl
e
,
a
nd
S
i
y
a
s
ha
0
2
4
6
8
10
12
14
16
18
20
0
0
.
5
1
1
.
5
2
2
.
5
3
x
1
0
-3
e
r
r
o
r
p
l
o
t
i
n
l
e
a
r
n
i
n
g
i
t
e
r
a
t
i
o
n
n
o
.
M
S
E
e
r
r
o
r
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5021
-
5031
5028
F
i
g
ur
e
6.
D
e
no
i
s
e
d
im
a
ge
f
rom
di
f
f
e
r
e
nt
t
y
pe
s
of
no
i
s
y
im
ag
e
s
(
f
r
om
l
e
f
t
to
r
i
g
ht
:
s
a
l
t
&
pe
pp
e
r
,
G
a
us
s
i
a
n
no
i
s
e
,
s
pe
c
ka
l
e
no
i
s
e
,
s
a
l
t
&
p
e
pp
e
r
+
G
a
us
s
i
a
n
n
oi
s
e
,
s
a
l
t
&
pe
p
pe
r
+
s
pe
c
kl
e
,
s
a
l
t
&
pe
p
pe
r
+
G
a
us
s
i
a
n
+
s
pe
c
k
l
e
no
i
s
e
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
& C
om
p
Eng
IS
S
N: 20
88
-
8708
Patte
rn
Ap
pro
xi
ma
ti
on B
as
e
d
Gen
e
ra
li
zed
Image Noise
R
edu
ct
io
n usi
ng
Ad
ap
ti
ve
…
(
Nagar
aj B
ha
t
)
5029
Table
6
.
N
oise
reducti
on in
i
m
ages
with
[4 1
4] FF
a
rc
hitec
ture
size wit
h n
oise
den
sit
y:
0
.
05 and
value
s
are
PSN
R i
n ‘db’ (
NI
:
noisy
i
m
age;
W
D:
Win
er
denoise
d
im
age; ND
:
Ne
ur
al
denoise
d
im
age)
I
m
ag
e
Salt &
Pep
p
er
Gau
ss
ian
Sp
eckl
e
Salt & Pep
p
er
+
Gau
ss
ian
Salt & Pep
p
er
+Sp
eck
le
Salt &
Pep
p
er
+ Gaus
sian
+Sp
eckl
e
Lena
NI
:
1
8
.45
5
9
W
D
:
2
1
.12
1
5
ND
:
2
6
.49
4
3
1
9
.11
6
7
2
3
.58
3
0
2
3
.04
8
0
1
8
.81
9
7
2
4
.93
1
7
2
6
.95
6
8
1
6
.21
7
1
2
0
.22
4
7
2
2
.39
0
1
1
6
.02
9
6
2
1
.10
0
3
2
5
.31
4
5
1
4
.52
7
7
2
0
.08
9
0
2
2
.20
2
3
E
li
a
n
e
NI
:
1
8
.41
2
9
W
D
:
2
1
.04
2
4
ND :
2
6
.67
2
1
1
9
.19
7
6
2
3
.54
5
2
2
3
.30
1
3
1
8
.31
8
4
2
4
.48
7
2
2
6
.74
5
5
1
6
.26
8
2
2
0
.29
7
5
2
2
.77
7
8
1
5
.74
3
7
2
0
.88
4
7
2
5
.14
0
2
1
4
.40
4
0
2
0
.05
1
0
2
2
.33
1
7
Bo
at
NI
:
1
8
.46
8
1
W
D
:
2
1
.14
1
2
ND
:
2
4
.42
0
9
1
9
.09
7
8
2
3
.31
9
8
2
1
.66
4
0
1
8
.47
4
6
2
4
.89
2
6
2
4
.59
2
7
1
6
.17
2
5
2
0
.10
7
5
2
1
.26
3
0
1
5
.89
2
6
2
1
.10
0
7
2
3
.63
6
9
1
4
.31
4
8
1
9
.90
3
3
2
1
.17
1
1
Veg
etab
le
NI
:
1
8
.36
6
6
W
D
:
2
0
.87
9
1
ND
:
2
5
.36
8
8
1
9
.11
7
9
2
3
.55
3
0
2
2
.63
5
0
1
8
.89
2
6
2
4
.72
4
5
2
6
.00
2
4
1
6
.04
6
1
1
9
.91
8
6
2
1
.85
2
1
1
6
.01
2
8
2
0
.92
3
3
2
4
.60
9
1
1
4
.55
1
5
1
9
.95
3
5
2
1
.80
1
7
Siy
ash
a
NI
:
1
7
.63
2
0
W
D
:
1
9
.84
4
4
ND
:
2
2
.26
9
7
1
9
.33
4
7
2
3
.20
5
4
2
0
.55
7
9
2
1
.26
7
0
2
5
.99
5
3
2
3
.40
8
8
1
5
.70
1
9
1
8
.84
1
3
1
9
.53
2
6
1
6
.67
7
2
2
0
.40
0
6
2
2
.08
9
2
1
4
.78
7
7
1
8
.97
8
9
1
9
.48
2
2
Table
7
.
N
oise
reducti
on in
im
ages
with
[4 1
4] FF
a
rc
hitec
ture
size wit
h n
oise
den
sit
y:
0
.
15 and
value
s
are
PSN
R i
n ‘db’ (
NI
:
noisy
i
m
age;
W
D:
Win
er
denoise
d
im
age; ND
:
Ne
ur
al
denoise
d
im
age)
I
m
ag
e
(0.1
5
)
Salt &
Pep
p
er
Gau
ss
ian
Sp
eckl
e
Salt & Pep
p
er
+
Gau
ss
ian
Salt & Pep
p
er
+Sp
eck
le
Salt & Pep
p
er
+ Gaus
sian
+Sp
eckl
e
Lena
NI
:
13
.65
6
4
W
D
:
1
8
.61
4
7
ND
:
2
3
.05
1
5
1
5
.06
1
9
1
6
.27
3
1
1
6
.21
6
3
1
4
.36
9
2
2
0
.65
9
4
2
4
.25
8
7
1
2
.19
1
4
1
5
.14
4
2
1
6
.03
9
2
1
1
.77
9
5
1
8
.14
0
6
2
1
.58
4
8
1
0
.70
3
5
1
5
.79
4
1
1
6
.86
5
3
E
li
a
n
e
NI
:
13
.76
2
3
W
D
:
1
8
.72
4
4
ND :
2
3
.46
4
7
1
5
.20
5
5
1
6
.37
2
8
1
6
.40
7
4
1
3
.88
7
9
2
0
.35
8
3
2
3
.67
0
2
1
2
.34
9
4
1
5
.39
9
9
1
6
.49
3
3
1
1
.46
3
8
1
7
.74
1
5
2
0
.92
9
7
1
0
.64
4
8
1
6
.04
2
9
1
7
.24
0
3
Bo
at
NI
:
13
.73
0
2
W
D
:
1
8
.69
3
7
ND
:
2
2
.13
3
0
1
5
.00
8
6
1
6
.18
6
1
1
5
.84
8
7
1
3
.88
5
4
2
0
.58
9
8
2
2
.78
8
4
1
2
.21
4
3
1
5
.15
9
5
1
5
.94
4
4
1
1
.52
0
0
1
8
.10
2
8
2
0
.76
8
1
1
0
.64
6
1
1
5
.96
3
0
1
7
.05
1
8
Veg
etab
le
NI
:
13
.50
6
0
W
D
:
18
.39
4
8
ND
:
2
2
.34
5
8
1
5
.03
3
3
1
6
.24
4
2
1
6
.22
2
2
1
4
.52
3
2
2
0
.49
5
5
2
3
.91
5
7
1
2
.06
6
2
1
4
.95
1
2
1
5
.87
9
4
1
1
.74
9
9
1
7
.88
4
6
2
1
.01
7
2
1
0
.72
8
2
1
5
.64
8
9
1
6
.62
5
9
Siy
ash
a
NI
:
1
2
.86
6
1
W
D
:
1
7
.20
6
3
ND
:
1
9
.67
6
1
1
4
.97
7
5
1
6
.19
4
0
1
5
.53
3
5
1
6
.65
3
2
2
1
.29
8
6
2
2
.47
6
2
1
1
.30
1
7
1
3
.76
5
1
1
4
.29
8
0
1
2
.30
6
1
1
7
.69
6
4
1
9
.63
9
8
1
0
.58
7
5
1
4
.35
0
3
1
4
.85
8
2
Fig
ure
7
.
Re
co
rr
ect
io
n of
im
a
ge bloc
k
c
o
r
rupted by s
pec
kle noise
4.2.4
.
C
olo
r
I
ma
ge Den
oise
W
it
h
the
sam
e
pr
inciple
as
f
or
gray
scal
e
im
age,co
lo
r
im
age
nois
e
resto
rati
on
is
al
so
app
li
ed
.
C
olor
i
m
age
of
Le
na
is
ta
ken
f
or
trai
ning
pur
pose
an
d
in
div
i
du
al
c
olor
m
a
trix
(Re
d,
G
re
en
&
B
lue
)
proces
s
ind
e
pende
ntly
.D
if
fer
e
nt
ty
pe
s
of
noise
ha
ve
app
li
e
d
f
or
t
est
case
over
Lena
a
nd
Ve
ge
ta
ble
Im
age
as
show
n
in
Fig
ur
e
8
.
T
he
obta
ine
d
pe
rfor
m
ance
in
te
rm
s
of
PSN
R
has
sho
wn
i
n
Table
8.
It
is
obser
ve
d
that
th
ere
is
sign
ific
a
nt
im
pr
ovem
ent in noi
s
e reducti
on
w
it
h
al
l var
ie
ti
es
of
no
ise
.
0
10
20
30
40
50
60
70
140
160
180
200
220
240
260
P
i
x
e
l
s
n
u
m
b
e
r
P
i
x
e
l
s
V
a
l
u
e
WN
AP
TR
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
87
08
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
5021
-
5031
5030
Fig
ure
8.
D
e
n
o
i
s
e
d
c
ol
or
i
m
a
ge
f
rom
di
f
f
e
r
e
n
t
t
y
pe
s
of
no
i
s
y
im
a
ge
s
(
f
r
om
le
f
t
t
o
r
i
gh
t
:
s
a
l
t
&
pe
p
pe
r
,
G
a
us
s
i
a
n
n
oi
s
e
,
s
pe
c
k
l
e
n
oi
s
e
,
s
a
l
t
&
pe
pp
e
r
+
G
a
us
s
i
a
n
n
oi
s
e
,
s
a
l
t
&
pe
pp
e
r
+
s
pe
c
kl
e
,
s
a
l
t
&
pe
pp
e
r
+
G
a
us
s
i
a
n
+
s
pe
c
kl
e
no
i
s
e
)
Table
8
.
N
oise
reducti
on in
col
or
im
ages w
it
h [4 1
4
]
F
F a
r
chite
ct
ur
e
siz
e
with
no
ise
d
e
nsi
ty
: 0.
2
a
nd
va
lues
are P
SN
R i
n ‘db’ (
NI
:
noisy
im
age; ND
:
Ne
ur
al
de
no
ise
d
i
m
age
)
I
m
ag
e
(0.0
5
)
Salt &
Pep
p
er
Gau
ss
ian
Sp
eckale
Salt & Pep
p
er
+
Gau
ss
ian
Salt & Pep
p
er
+Sp
eck
le
Salt & Pep
p
er
+ Gaus
sian
+Sp
eckl
e
Lena
NI
:
1
8
.18
7
7
ND :
23
.97
9
8
1
9
.34
8
5
2
2
.50
5
6
1
9
.11
9
1
2
3
.96
1
2
1
6
.15
1
5
2
0
.49
4
7
1
5
.97
3
7
2
1
.20
7
7
1
4
.58
1
2
1
9
.42
0
1
Veg
etab
le
NI
:
1
7
.97
1
5
1
9
.20
7
7
1
9
.81
9
3
1
5
.90
7
2
1
6
.08
2
4
1
4
.65
4
2
ND
:
2
1
.99
1
8
2
1
.57
7
8
2
3
.12
1
2
1
9
.73
5
7
2
0
.85
8
8
1
9
.12
4
0
5.
CONCL
US
I
O
N
U
ni
ver
sal
nois
e
resto
rati
on
ha
s
pro
pose
d
in
this
pa
per
wit
h
intel
li
gen
t
m
a
nn
e
r
by
the
use
of
arti
fici
al
neural
netw
ork
.
The
p
r
opos
e
d
m
et
ho
d
ha
s
ta
ken
a
ver
y
ne
w
co
nce
pt
of
l
ocal
patte
r
n
form
by
the
pix
e
ls
with
neig
hbo
rin
g
pi
xels
to
reduce
the
dif
fer
e
nt
ty
pes
of
noses
.
I
f
this
patte
rn
is
hav
i
ng
so
m
e
kind
of
ir
regul
ar
i
ti
es
because
of
noise
,
ap
pro
xim
a
t
ion
intel
li
gen
c
e
of
neura
l
net
work
s
uppo
rt
enou
gh
t
o
co
rrec
t
this.
The
p
r
opose
d
m
et
ho
d
pe
rform
ances
ha
ve
s
how
n
c
om
petitive
pe
rfo
rm
ance
with
a
da
ptive
W
ei
nner
filt
er.
D
evel
op
m
ent
of
so
luti
on
nee
de
d
only
on
e
i
m
age
at
the
time
of
trai
n
in
g
in
res
ult
there
is
a
ver
y
fast
i
m
ple
m
entat
ion
of
the
le
arn
in
g
pr
oce
ss
.
P
rese
nted
m
et
ho
d
has
va
li
dated
with
m
os
t
com
m
on
ty
pes
of
noise
s
in
pr
act
ic
e
an
d
the
resu
lt
s
wer
e
v
e
ry
sat
isfact
ory
.
Evaluation Warning : The document was created with Spire.PDF for Python.