TELKOM
NIKA
, Vol.14, No
.1, March 2
0
1
6
, pp. 47~5
5
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i1.2634
47
Re
cei
v
ed Se
ptem
ber 3, 2015; Re
vi
sed
De
cem
ber 3,
2015; Accept
ed Ja
nua
ry 2,
2016
Demosaicing o
f
Color Images by Accurate Estimation
of Luminance
V.N.V. Sat
y
a
Prakash*
1
, K. Sat
y
a Pra
sad
2
, T. Ja
y
a
Chandr
a Prasad
3
1,3
Dept. of ECE, Rajeev Gan
d
h
i Memori
al co
l
l
eg
e of Engi
ne
erin
g and T
e
ch
nol
og
y (RGMC
E
T
)
,
Nandy
al, A.P.,
India
2
Dept. of ECE,
Ja
w
a
harl
a
l Ne
hru T
e
chnol
ogi
cal Univ
ersit
y
, Kakin
ada (JNT
UK),
Kakin
ada, A.P. India
*Corres
p
o
ndi
n
g
author, em
ail
:
prakashvnv
@
g
mail.c
o
m
1
,
p
rasad
_ko
dati@
ya
ho
o.co.in
2
,
jp.talar
i
@gm
a
il
.com
3
A
b
st
r
a
ct
Digita
l
c
a
mera
s acq
u
ire
col
o
r i
m
ag
es
usin
g a
sin
g
le
se
n
s
or w
i
th C
o
lor
filter Arrays.
A sin
g
l
e
color co
mpon
e
n
t per pixe
l is
acquir
ed us
in
g color
filter a
rrays and th
e remain
in
g tw
o compon
ents
are
obtai
ne
d usi
n
g de
mosa
icin
g tech
niq
ues.
T
he co
nv
e
n
ti
ona
l d
e
m
osa
i
cing tec
h
n
i
qu
es existe
nt i
n
duce
artifacts in res
u
ltant i
m
a
ges eff
e
cting r
e
constr
ucti
on
qua
lity. T
o
overco
me this dr
aw
back a
freque
ncy b
a
s
e
d
de
mos
a
ici
ng t
e
chn
iqu
e
is pr
opos
ed. T
h
e l
u
min
ance
a
n
d
chro
mi
na
nce compo
nents e
x
tracted
fro
m
th
e
freque
ncy d
o
m
ain
of the
ima
ge ar
e i
n
terp
ol
ated to
pro
duc
e inter
m
edi
ate
de
mos
a
ic
ed i
m
a
ges. A
nov
el
Neur
al Netw
ork Based Ima
g
e
Reco
nstructi
on Alg
o
rith
m is
appli
ed to the
intermedi
ate d
e
mosa
iced i
m
a
ge
to obtai
n res
u
ltant d
e
mosa
iced i
m
ages.
T
he result
s p
r
esente
d
in t
he pa
per pr
o
v
e the pro
p
o
s
ed
de
mos
a
ici
ng te
chni
que
exhi
bit
s
the best performanc
e
an
d is appl
icab
le to
a w
i
de variety
of ima
ges.
Ke
y
w
ords
: De
mos
a
ici
ng, freq
uency, col
o
r filt
er array, lu
mi
n
ance, chro
min
ance
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
A color im
ag
e is u
s
ually
comp
osed of
three colo
r plane
s an
d, accordingly,
three
sep
arate
sen
s
ors a
r
e re
qu
ired for a
ca
mera to mea
s
ure an ima
g
e. To redu
ce
the cost, ma
ny
came
ra
s u
s
e
a sin
g
le s
e
n
s
or
cov
e
re
d
wit
h
a
co
lo
r filter array (CFA). The mo
st com
mon
CFA
use
d
i
s
the
Bayer
CFA
[1, 2], whi
c
h i
s
sh
own
in Fig
u
re
1. In the
CFA-ba
sed
se
nso
r
config
uratio
n, only one col
o
r is m
e
a
s
ured at ea
ch pi
xel and the
missi
ng two colo
r value
s
are
estimated by
interpol
ation. The estim
a
tio
n
pr
o
c
e
ss i
s
known as
colo
r demo
s
ai
cin
g.
Figure 1. Bayer CFA Pattern alo
ng wit
h
its red, gree
n and blu
e
sa
mples
Demo
sai
c
in
g
probl
em i
s
a
spe
c
ial
case
of the im
age
recon
s
tru
c
tio
n. The
mo
st
straig
ht
forwa
r
d
app
roach for solving the
dem
o
s
ai
cing
pr
obl
em in
col
or i
m
age i
s
to
a
pply one
of t
he
stand
ard re
constructio
n method
s
for gray
sc
al
e i
m
age
s on
e
ach
ch
annel
sep
arately.
Many
method
s ha
ve been p
r
opo
sed fo
r singl
e ch
ann
el re
con
s
tru
c
tion such
as Interpolat
ion,
R
e
gu
la
r
i
z
a
tion
a
n
d
In
ve
rs
e F
ilte
r
ing
.
R
e
c
o
ns
tr
uc
ting
each
chan
nel
se
parately produ
ce
s a
r
tifacts.
Better re
con
s
tru
c
tion of t
he imag
e ca
n be o
b
taine
d
by takin
g
into co
nsi
deration the
cro
ss
cha
nnel
correlation. Seve
ral m
e
thod
s
have b
een
sugge
sted
to solve cross
chann
el correl
ation
demo
s
ai
cing
probl
em. Existing ap
proa
ches fo
r inte
r cha
nnel co
rrelation
s
improve
perfo
rma
nce
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 1, March 2
016 : 47 – 55
48
over in
dep
en
dent
cha
nnel
re
con
s
tructi
on. But
still there i
s
a
possibility for improvemen
t.
Template
mat
c
hin
g, filterin
g and
lumin
a
nce
ba
sed
ap
proa
ch
es are
basi
s
fo
r vari
ous te
ch
niqu
es,
but they are
limited in the ca
pability
of reco
ns
t
r
u
c
tion with dif
f
erent spatial
and chrom
a
tic
cha
r
a
c
teri
stics and results
in some a
r
tifa
cts in the reconstructe
d im
age
s.
Demo
sai
c
in
g
techniq
ue
s
can b
e bro
ad
l
y classified as freq
uen
cy
domain [4-7]
base
d
and
spatial
d
o
main
ba
sed
[8-10]. A d
e
t
ailed survey
of the vari
e
d
dem
osaici
n
g
techniqu
es is
pre
s
ente
d
in
[3] and [1
1].
Based
on
the
literatu
r
e
re
v
iewe
d
it i
s
evi
dent t
hat
t
he existing state
of
art dem
osaici
ng alg
o
rithm
s
indu
ce a
r
tifa
cts in
the
re
constructe
d im
age [3, 9], [1
1-12]. A d
e
tai
led
repo
rt o
n the
types of
artifa
cts
ob
se
rved
in de
mosa
ice
d
ima
ges i
s
p
r
esented
in
[1
3]. To mi
nimi
ze
occurre
nces
of artifact
s in
demo
s
ai
ced i
m
age
s,
the u
s
e
kno
w
le
dg
e de
rived fro
m
of local re
gio
n
or local patches [14
-
17] to interpol
ate/estima
te the
missi
ng color compo
nent
s is prop
osed.
An
iterative dem
osai
cin
g
tech
nique i
s
e
ssential to de
ri
ve local
kn
o
w
led
ge a
nd
achi
eve a
c
cu
rate
estimation [1
8, 19]. In recent times the
adoptio
n of
n
eural
networks in ima
ge p
r
oce
s
sing [2
0, 21]
to derive
kn
o
w
led
ge i
s
a
motivating fa
ctor fo
r t
he
a
u
thors of thi
s
pape
r. A neu
ral network b
a
s
ed
demo
s
ai
cing
techni
que p
r
o
posed in [22]
bears the
cl
ose
s
t simil
a
rit
y
to the work prop
osed he
re.
In [22] the
ro
tational inva
ri
ance i
s
u
s
ed
to tr
ain
the
neural n
e
two
r
ks a
nd
esti
mate the
missin
g
color pixel.
In this pap
er a freque
ncy
domain b
ased dem
osaici
ng tech
niqu
e
is pro
po
s
ed.
Based
on the freq
ue
ncy co
mpo
n
e
n
ts, lumina
nce and chromi
nan
ce inform
ation is u
s
ed
to represent the
image. Th
e input imag
e’s luminan
ce
and
chro
min
ance
compo
nents are se
greg
ated.
Usin
g
bilinea
r inte
rp
olation the l
u
minan
ce a
nd
three
ch
romi
nan
ce
cha
nn
els a
r
e
com
bi
ned to p
r
o
du
c
e
an intermedi
a
t
e image th
at exhibits e
r
rors o
r
artifa
cts.
To elimin
ate the e
rro
rs in t
he intermedi
a
t
e
image, a neu
ral network b
ase
d re
con
s
truction i
s
pro
p
ose
d. Re
con
s
tru
c
tion alg
o
r
ithm is refe
rred
to as Neural
Network Based Image
Re
con
s
tru
c
ti
on
Algorithm (NNIRA) in th
e prop
osed
system.
The
neu
ral
netwo
rks de
rive lo
cal
kn
owle
dge
fro
m
the i
m
ag
e pat
che
s
constructe
d.
The
kno
w
le
dge a
s
sist
s in estim
ating the missing
comp
on
ents.
The remai
nin
g pap
er i
s
o
r
gani
zed
as f
ollows
. Th
e
prop
osed
system is p
r
e
s
e
n
ted in
se
ction two. The exp
e
rim
ental
stu
d
y a
nd pe
rforman
c
e
com
pari
s
o
n
of ou
r p
r
op
ose
d
with
oth
e
r
state of a
r
t d
emosaici
ng
a
l
gorithm
s i
s
pre
s
ent
s in
section th
re
e. The
con
c
lu
si
ons
and
futu
re
work is di
scu
s
sed in the la
st se
ction of the pap
er.
2. Frequen
c
y
Domain N
N
IRA App
ro
ach (Prop
os
e
d Sy
stem)
NNI
RA ba
se
d techniqu
es have b
een
employed
to
recon
s
tru
c
t t
he ima
ge
but
to the
best of ou
r kn
owle
dge n
o
a
u
thor h
a
s
attempted to
co
nsid
er L
umin
ance co
mpo
n
ent (extra
cted
in
the Freq
uen
cy domain) to reco
nstruct th
e origin
al ima
ge (u
sing n
e
u
r
al networks).
The
step
s o
f
the p
r
opo
sed ima
ge
d
emosaici
ng t
ech
niqu
e
co
nsid
erin
g lu
minan
ce
informatio
n in
the frequen
cy domain are as follo
ws:
1)
Image Align
m
ent: Align the set of image
s pai
rwi
s
e usi
ng the l
o
w fre
que
ncy (lumina
n
ce
)
informatio
n of the CFA Fou
r
ier tra
n
sfo
r
m
images.
2) Lumina
n
ce/Chromi
nan
ce
Separ
ation: Extract the luminan
ce
an
d ch
romin
a
n
c
e info
rmatio
n
from ea
ch
of the input im
age
s. Bilinea
r inter
polatio
n is a
dopte
d
to combi
ne t
he lumin
an
c
e
and th
e three
ch
romi
nan
ce
ch
ann
els
re
sulting
in a
n
interme
diate i
m
age th
at is
not artefa
ct
free.
3)
Image Recon
s
tru
c
tion: Ne
ural Network
Base
d Imag
e Re
co
nstruction Algorith
m
(NNIRA
) i
s
use
d
for ima
ge recon
s
tru
c
tion
and
re
moval of a
r
tifacts ob
se
rve
d
in th
e inte
rmediate i
mag
e.
Duri
ng traini
n
g phase, the model is furni
s
he
d with bot
h the interme
diate image p
atche
s
P
for
i
1,
2,3,
…
.
.
N
and the
ori
gi
nal pat
che
s
Q
.After training, NNI
RA will
be abl
e to reconstruct
the corre
s
p
on
ding dem
osai
ced ima
ge fo
r any given error o
b
servatio
n.
3. Rese
arch
Metho
d
Based o
n
the idea prese
n
ted
by Alleysso
n et al [23].luminan
ce and ch
rom
i
nan
ce
informatio
n are en
cod
ed separately from the F
ourie
r spe
c
trum o
f
a Bayer CFA image. They
sho
w
e
d
that a
Bayer CFA image
x,
y
ca
n
be written a
s
a su
m of the
red, g
r
ee
n, a
nd blu
e
colo
r ch
ann
el
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Dem
o
sai
c
in
g of Color Im
ag
es b
y
Accurate Es
tim
a
tion of Lum
inance
(V.N.V. Satya Prakash)
49
∑
x,
y
T
x,
y
with
T
x,
y
1
cos
Π
x
1
cos
Π
y
4
⁄
re
d
T
x,
y
1
cos
П
x
cos
Π
y/2
gr
een
T
x,
y
1
cos
Π
x
1
cos
Π
y
4
⁄
blue
(
1
)
The imag
e
x,
y
is the i
th
color
cha
nnel ima
g
e, and
T
x,
y
is a modulatio
n matrix,
whi
c
h is ‘1’ a
t
the measu
r
ed po
sition
s of the image,
and ‘0’ else
whe
r
e. As th
ese m
odulati
on
function
s are
a com
binatio
n of co
si
ne
s, their Fou
r
ie
r tran
sform i
s
a
combi
nation
of Dira
cs. Usi
ng
the fact that
a prod
uct in spatial do
mai
n
corr
e
s
po
nd
s to a convol
ution in frequ
ency dom
ain,
we
obtain:
u,
v
2
u,
v
∗
δ
u,
v
δ
u
,v
δ
u
,v
δ
u,
v
2
∗
δ
u
,v
(2)
δ
u
1
2
,v
1
2
δ
u
1
2
,v
1
2
δ
u
1
2
,v
1
2
Whe
r
e Fo
urie
r tran
sform
s
are indi
cate
d in bold. Any color ima
ge
u,
v
can be represented a
s
a
sum of
a scal
ar representi
ng its lumi
na
nce
α
u,
v
and
a le
ngth thre
e ve
ctor
u,
v
that is
called
chromin
a
n
c
e
and re
presen
ts oppo
nent colors as b
e
lo
w in Equation
(3)
u,
v
u,
v
u,
v
u,
v
α
u,
v
u,
v
u,
v
u,
v
(
3
)
Lumina
n
ce is defined a
s
α
2
/2
, we obtai
n:
u,
v
u,
v
u,
v
u,
v
u,
v
2
u,
v
u,
v
/2
2
u,
v
u,
v
u,
v
u,
v
u,
v
u,
v
2
u,
v
(4)
Usi
ng this
de
finition, we can see that the fi
rst term i
n
Equation
(2
) co
rrespon
d
s
to the
luminan
ce
sig
nal
α
u,
v
and the
two oth
er te
rm
s represent th
e ch
romi
nan
ce
u,
v
.Because of
the modulatio
n function
s, the lumina
nce
part appe
ars in low frequ
ency re
gion
of the spe
c
trum,
and a the chrominan
ce p
a
rt appears in high freq
uen
cy region.
Usi
ng
a lo
w
pass filte
r
, th
e lumi
nan
ce
i
n
formatio
n from the
ima
g
e
s i
s
extra
c
te
d In thi
s
pape
r, a f
r
eq
uen
cy dom
ai
n ap
pro
ach i
s
u
s
e
d, whi
c
h u
s
e
s
only t
he lo
w frequ
enci
es for im
age
alignme
n
t, which
are le
ss
pron
e to ali
a
sing. As thi
s
i
s
also
the p
art
of the CFA F
ourie
r tran
sfo
r
m
that contain
s
the luminan
ce informatio
n, we appl
y ou
r algorithm di
rectly on the raw CFA imag
es.
Next, we
se
parate th
e i
m
age
s into l
uminan
ce
an
d ch
romi
nan
ce u
s
in
g a l
ow p
ass filte
r
, and
interpol
ate the two se
parately.
The fre
que
ncy domain a
p
p
roa
c
h
pre
s
e
n
ted by Vand
ewall
e
et al [24]
is
used to alig
n
Bayer
CFA i
m
age
s
and
this
algo
rithm
sele
cts only t
he lo
w f
r
eq
ue
ncy info
rmati
on
sin
c
e thi
s
part
of the
spe
c
trum is le
ss corr
upted
by
aliasi
ng a
nd
to get lu
mi
nan
ce info
rm
ation a
pa
rt of
Spectrum en
codi
ng is d
on
e as
sho
w
n i
n Equation
(2
), therefo
r
e we can
dire
ctly apply alignm
ent
algorith
m
to the ra
w CFA i
m
age
s. First we pe
rfor
m pl
anar
rotation
estimation, fo
llowe
d by pla
nar
shift estimatio
n
. The rotation angle is est
i
mated by co
mputing the frequ
en
cy con
t
ent
δ
of the
image a
s
a function of the
angle for e
a
ch of the input image
s.
δ
|
ρ
,
θ
|
dr
d
θ
,
∞
δ
∆
δ
/
δ
∆
δ
/
(
5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 1, March 2
016 : 47 – 55
50
Whe
r
e
ρ
,
θ
is th
e Fou
r
ier transfo
rm of
the CFA im
age
, convert
ed in pol
ar
coo
r
din
ates, t
he rotatio
n a
ngle bet
wee
n
two imag
es
can th
en be f
ound at the
maximum of
the
correl
ation b
e
twee
n two
su
ch fun
c
tio
n
s. Next,
the rotatio
n
is can
c
el
ed, a
nd the
shifts are
estimated
by
com
puting
the le
ast
squ
a
re
s fit of a
plane
thro
ug
h the
(line
a
r) pha
se
difference
betwe
en the
image
s. As
we o
nly use
the low fr
equ
ency info
rmat
ion of the im
age
s, we d
o
not
need to sepa
rate luminan
ce and chromi
nan
ce for thi
s
phase. The
use of the ra
w se
nsor d
a
ta for
the image ali
gnment allo
ws a highe
r preci
s
ion of
th
e alignme
n
ts,
as no additi
onal filtering
or
interpol
ation
errors are introdu
ced.
3.1. Luminance / Chromin
a
nce Sepa
ra
tion
Here we
sep
arate the lum
i
nan
ce an
d chromi
nan
ce i
nformatio
n in
each of the i
m
age
s
in order to int
e
rpol
ate the
m
se
parately. As in
di
cate
d
in Equatio
n
(2),
we
extra
c
t the lu
mina
nce
sign
al from the CFA image
s usin
g a low-pa
ss filter
sp
ecified by Alleysson et al [23]. The three
chromin
an
c
e
parts
(for re
d, gree
n and
blue) a
r
e th
en obtain
ed
by subtractin
g this lumin
a
n
ce
informatio
n from the
re
d,
gree
n a
nd
bl
ue
ch
a
nnel
s of
the CFA
i
m
age and
d
emodul
ating the
result. This
result
s in a l
u
minance im
age
α
and three
ch
romin
an
c
e
image
s
,
and
all
at
the origin
al image si
ze.
α
∗
α
ʘ
T
∗
α
ʘ
T
∗
α
ʘ
T
∗
with
121
242
121
/4,
010
141
010
/
4
(
6
)
The matri
c
e
s
and
are two demod
ulation
(or interpol
ation) filters, an
d the symbol
ʘ
is u
s
e
d for
a point
wi
se
multiplicatio
n of
two mat
r
ice
s
.F
rom separated and
demo
dulate
d
luminan
ce
a
nd
chromin
a
nce
si
gnal
s we
compu
t
e their
hig
h re
sol
ution
versi
on
s
u
s
ing
Normali
z
ed
Convolution
(NC) ap
pro
a
ch
prop
osed
in [
27]
on e
ach
o
f
the fou
r
ch
a
nnel
se
parate
l
y.
A Gaussia
n
weig
hting fun
c
tion (appli
c
a
b
ility functi
on
) is u
s
ed to h
ave the high
est co
ntrib
utions
from sa
mple
s close to the con
s
id
ere
d
pi
xel and used
a varian
ce
σ
2
.
A pixel of the
high re
soluti
on image i
s
comp
uted fro
m
the pixels in a neighb
o
u
rho
o
d
arou
nd it as:
P
′
(
7
)
Whe
r
e
is an
1
vector conta
i
ning the neighbo
rho
od pi
xels,
is an
matrix
of
basi
s
fun
c
tion
s sa
mpled at
the local
coo
r
dinate
s
of the pixels
,and
is an
weightin
g
matrix contai
ning the Gau
ssi
an wei
ghts sample
d at
the pixel coo
r
dinate
s
. The first eleme
n
t of
the
1
vec
t
or
P
′
gives the
inte
rpolate
d
pixel
value. Fo
r n
e
ighb
ourhoo
d
sele
ction, a
circula
r
regio
n
with ra
dius four time
s the
pixel di
stan
ce
of the
high
re
sol
ution ima
ge i
s
u
s
ed.
Du
e to t
he
nonu
niform g
r
id, the numb
e
r of pixels
in this regi
on
may vary dep
endin
g
on the
position.
We p
e
rfo
r
m
bilinea
r interp
olation for th
e lumina
nce cha
nnel
, as
well as
for eac
h
of
the ch
romi
na
nce
ch
ann
els
,
and
then
we a
dd lumin
an
c
e
and
ch
romin
ance togeth
e
r
,
whi
c
h re
sult
s in an interm
ediate high
resol
ution col
our ima
ge
P
. If
we c
o
ns
ider
pixels of th
e
high re
sol
u
tio
n
image to be
compute
d
then P can be
given as:
P
′
(
8
)
Her
e
P
induce
s
erro
r / artifacts, a
s
the final high
resolution imag
e
are compute
d
by
fitting a polyn
omial
surfa
c
e
to co
mpute
high
re
solutio
n
imag
e an
d
also
scale
of
the appli
c
a
b
ili
ty
function
play
s a
de
cisive
role in th
e qu
ality of interp
olation
su
ch
as
usin
g
Lo
w-o
r
d
er
NC
with a
large a
ppli
c
a
b
ility window
can
not re
con
s
tru
c
t small d
e
tails in the i
m
age.
3.2. Image Recons
truc
tio
n
A
ssu
ming
P
is
the intermedi
ate image ob
tained from Equation (8) a
nd
Q
is the perfect
demo
s
ai
ced i
m
age; we fo
rmulate the im
age corruptio
n pro
c
e
ss a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Dem
o
sai
c
in
g of Color Im
ag
es b
y
Accurate Es
tim
a
tion of Lum
inance
(V.N.V. Satya Prakash)
51
P
ῼ
Q
(
9
)
Whe
r
e
ῼ
R
→R
is a
n arbit
r
ary
stoch
asti
c
erro
r
process ind
uce
d due to
the polynomi
a
l
surfa
c
e fitting
. Then, demo
s
ai
cing le
arni
ng obje
c
tive become
s
:
g
ar
g
m
i
n
g
A
|
|
g
P
Q
|
|
(
1
0
)
In Equation (10) we
see that the task is to find a fun
c
tion
′
g
′
that best
approximate
s
ῼ
. We can no
w treat the ima
ge dem
osaici
ng problem i
n
a unified fra
m
ewo
r
k by
cho
osi
ng ap
p
r
op
riate
ῼ
in different situati
ons.
3.3. Image Recons
truc
tio
n De
m
o
saic Aut
o
-e
ncod
e
r
[IRD
A]
Let
Q
be the origin
al data
for
i
1,2,
3,
…
.
.
N
and
P
be the corru
pted versio
n o
f
corre
s
p
ondin
g
Q
.
hP
φ
W
b
(
1
1
)
Q
P
φ
W
′
hP
b
′
(
1
2
)
Whe
r
e
φ
P
1
e
x
p
P
is th
e sigmoi
d act
i
vation function whi
c
h is
applie
d comp
onent-
w
ise to vec
t
or
s
,
h
is the
hidde
n layer activation,
Q
P
is an approximation of
Q
and
ω
W,
b,
W
′
,b
′
represents t
he wei
ghts
a
nd bia
s
e
s
. IRDA
ca
n be trained
with va
riou
s optimi
z
ation
methods
to minimiz
e
the
rec
o
ns
truc
tion loss
.
ω
ar
gmi
n
ω
∑
‖Q
Q
P
‖
(
1
3
)
After training
IRDA, we
m
o
ve on to tra
i
ning the n
e
xt layer by usi
ng the hi
dde
n laye
r
activation of the first layer
as the inp
u
t of the next layer.
3.4. Neural Net
w
o
r
k Ba
se
d Image Rec
onstr
uction
Algorithm
In this
section, we will
descri
be the st
ructure and optimi
z
ati
on obj
ective of the
prop
osed m
o
del Neural Network Ba
se
d Image
Re
c
onstructio
n
A
l
gorithm
(NNI
RA). Due to
the
fact that dire
ctly proce
ssi
ng
t
he entire im
age is intract
able, we
in
ste
ad dra
w
ove
r
l
appin
g patch
es
from the imag
e as ou
r data
obje
c
ts.
In the trainin
g
pha
se, the
model is su
pplied with b
o
th the corrupted erro
r image
patch
es
P
, for
i
1,
2,3,
…
.
.
N
and the ori
ginal pat
che
s
Q
. After training, IRDA will be able to
recon
s
tru
c
t the corre
s
p
ondi
ng cle
an ima
ge given any
error ob
se
rva
t
ion.
To co
mbine
spa
r
se codin
g and n
eural
netwo
rk
s a
n
d
avoid over-fitting, we train
IRDA
to minimize t
he re
con
s
truction loss re
g
u
larized by a sparsity-ind
uci
ng term.
M
P,
Q,
ω
∑
‖
Q
Q
P
‖
δ
KM
Γ
‖
Γ
λ
‖
W
‖
W
′
(14
)
KM
Γ
Γ
∑
Γ
|
Γ
|
log
Γ
Γ
1
Γ
log
Γ
Γ
,
Γ
∑
hP
(15
)
Whe
r
e
h.
and
Q
.
are d
e
fined i
n
Equation
(11
)
and E
quatio
n (12
)
respe
c
tively. Here
Γ
is the
averag
e a
c
tivation of the
hidden l
a
yer. We re
gula
r
i
z
e the
hidde
n layer
rep
r
e
s
entatio
n to
be
spa
r
se by ch
oosi
ng small
Γ
so that the
KM
divergen
ce term will encou
rage the mea
n
activation
of hidde
n uni
ts to be
smal
l. Hen
c
e the
hidde
n uni
t
s
will be
ze
ro
most of the ti
me and
achi
eve
spa
r
sit
y
.
After training of the firs
t IRDA, we us
e
hQ
and
hP
as th
e cle
an an
d
error in
put
respec
tively for the
s
e
c
ond IRDA. Since
hQ
) lie
s in
a di
fferent spa
c
e
from
Q
, the me
aning
of
applying
ῼ
.
to
hQ
is not cle
a
r. We discarded
P
and used
ῼ
hQ
)) a
s
the erro
r input.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 1, March 2
016 : 47 – 55
52
We the
n initi
alize
a de
ep
netwo
rk with
the wei
ghts
obtaine
d fro
m
K
lay
e
re
d IRDA
s.
The network has on
e inpu
t layer, one output and
2K
1
hidde
n layers. The entire netwo
rk i
s
then traine
d u
s
ing the
stan
dard b
ack-propag
ation al
g
o
rithm to mini
mize the follo
wing o
b
je
ctive:
M
P,
Q,
ω
∑
‖
Q
Q
P
‖
λ
∑
‖W
‖
(
1
6
)
Here we removed
the sparsity
regul
a
riza
tion because
the pre-
trained weights will
serve a
s
re
gu
larization to the netwo
rk [25], acco
rdin
g to [26], during pre-t
r
aini
ng
and fine tuning
stage
s the l
o
ss fu
nctio
ns
are o
ptimized
with L-B
F
GS
algorith
m
to
achi
eve faste
s
t co
nverg
en
c
e
in our settings
.
3.5. Hidden
La
y
e
r Featur
e Analy
s
is
Duri
ng the training of im
age recon
s
tructi
on d
e
mo
sai
c
ing a
u
to-encode
rs [IRDA], the
error trai
ning
data is typically generate
d
with i
ndiscriminately
sel
e
cted simple error
di
stribut
ion
rega
rdl
e
ss
of the
ch
ara
c
te
ristics of the
particula
r
trai
ning data. Ho
wever,
it
is p
r
opo
sed
that t
h
is
pro
c
e
s
s de
se
rves a lot
of
attention i
n
real
wo
rl
d
problem
s; the
clea
n trai
ning
data i
s
i
n fa
ct
usu
a
lly subj
e
c
t to error. Hence, if we tend to es
timat
e the distri
but
ion of error
a
nd more u
s
in
g it
to gene
rate e
rro
r trai
ning
data, the en
suing IRDA wi
ll learn to
be
more
rob
ust
to errors in t
he
input data an
d pro
d
u
c
e be
tter feature
s
. It can be sug
geste
d that training ima
ge
recon
s
tru
c
tio
n
auto-e
n
code
rs [IRDA] with totally different color int
e
n
s
ity images th
at fit
to specifi
c
situatio
ns
can
also imp
r
ove
the perfo
rma
n
ce.
The re
sultin
g
image obtai
ned po
st neu
ral net
work
reco
nstructio
n
is co
nsi
dere
d as the
demo
s
ai
ced i
m
age. In the
next se
ction t
he expe
ri
me
ntal study to
evaluate the
perfo
rman
ce
of
the prop
osed
demo
s
ai
cing
techni
que i
s
pre
s
ente
d.
4. Experimental Stud
y
a
nd Performa
nce Comp
ari
s
ons
:
To
validate our pro
po
s
ed
system
the most
comm
o
n
ly
use
d
Ko
dak data
s
et [28]
is
con
s
id
ere
d i
n the expe
ri
mental stu
dy and p
e
rfo
r
mance comp
arison
s. The
Kodak
data
s
et
con
s
i
s
ts
of 2
4 imag
es ea
ch
having
a
high
re
so
lutio
n
of 7
68x51
2
and
24
bit
color
depth.
T
he
Kodak d
a
taset con
s
ide
r
ed
in the experi
m
ental
study
and pe
rform
a
nce
comp
ari
s
ons i
s
sh
own
in
Figure 2. Th
e pro
po
s
ed d
emosaici
ng t
ech
niqu
e
is d
e
velope
d usi
ng MATLAB. To evaluate
the
perfo
rman
ce
of demo
s
ai
ced ima
ge
s
o
btained
in
th
e
propo
se
d system, colo
r-pea
k sig
nal
-to-
noise ratio (CPSNR) d
e
fine
d in [23] is
used. The CPS
NR i
s
com
put
ed usi
ng.
10
log
(
1
7
)
Whe
r
e CMSE represents th
e Colo
r Mea
n
Square Erro
r and is defin
e
d
as:
∑∑
∑
,
,
,
,
(
1
8
)
Whe
r
e
1
and
2
are the inte
nsity levels
of original i
m
age a
nd d
e
mosaiced i
m
age
s of
height
and wi
dth
. ‘
’ value varie
s
from 1 t
o 3 for the three col
or pla
n
es.
The demo
s
ai
cing
pe
rform
ance re
sults based
on
CP
SNR obtain
e
d a
r
e ta
bulat
ed in
Ta
ble
1. In addition
to the CPSNR the PSNR re
sults
co
nsid
erin
g the
red, gre
en
and blu
e ch
a
nnel
informatio
n is also pre
s
ented. Base
d on the
re
sults it is o
bse
rved that
the propo
sed
demo
s
ai
cing
techni
que i
s
robu
st and
effectively wo
rks on
the
wide
variety of im
age
s p
r
e
s
ent
in
the Kodak d
a
t
aset.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
9
30
Dem
o
sai
c
in
g of Color Im
ag
es b
y
Accurate Es
tim
a
tion of Lum
inance
(V.N.V. Satya Prakash)
53
Figure 2. Kodak Image d
a
taset of 24 im
age
s co
nsi
d
e
r
ed for te
sting
Table 1. Average PSNR (o
f red, gree
n a
nd blue
cha
n
n
els) and
CP
SNR results i
n
dB obtaine
d
usin
g our p
r
o
posed dem
osaicin
g tech
niq
ue
Image No.
Red
c
h
an
nel
Green cha
nnel
Blue
c
h
an
nel
CPSNR
1(“Stone building”)
43.41
45.4
43.45
43.93
2(“Re
d door”)
44.69
47.94
45.92
45.84
3(‘Hats”)
46.59
48.74
46.86
47.22
4(“Girl in red”)
45.78
48.09
46.2
46.5
5(“moto
r cross bikes”)
44.15
46.12
44.48
44.77
6(“sail boat”)
44.03
46.27
44.12
44.61
7(“Wido
w
”) 46.74
48.91
47.12
47.42
8(“Ma
r
ket Place”)
42.54
44.99
42.65
43.16
9(“Sail boat”)
46.43
48.45
46.39
46.92
10(“sail boat rac
e”)
46.48
48.53
46.4
46.96
11(“sail boat at pi
er”)
44.48
46.83
44.9
45.21
12(“couple on
be
ach”)
46.24
48.54
46.32
46.82
13(“Mou
ntain stream”)
42.58
44.21
42.5
42.98
14(“Wate
r raflets
”)
43.38
46.1
44.41
44.38
15(“Girl
w
i
th Painted face”)
44.9 47.69
45.61
45.81
16(“Island”) 45.12
47.41
45.24
45.72
17(“R
oman statu
e”)
46.3
47.94
46.01
46.61
18(“Mod
el girl”)
44.34
46
44.08
44.67
19(“Light h
ouse in maine”)
44.54
46.74
44.7
45.14
20(“Propeller pla
ne”)
45.61
47.78
45.85
46.24
21(“H
ead Light”)
44.52
46.47
44.51
45.01
22(“Ba
r
n and Po
nd”)
44.78
46.71
44.81
45.28
23(“T
w
o
maca
w
s
”)
46.66
49.33
47.61
47.63
24(“Mou
ntain chalet”)
44.01
45.71
43.33
44.16
Avg
.
44.93
47.12
45.14
45.54
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 1, March 2
016 : 47 – 55
54
To com
pa
r
e
the perfo
rma
nce of the p
r
opo
sed d
em
osai
c
in
g tech
nique
with the other
state of a
r
t d
e
mosaici
ng t
e
ch
niqu
es th
e avera
ge PS
NR
of the red
,
green,
blue
cha
nnel
s a
nd
the
CPSNR re
sults obtain
ed
is co
nsi
de
r
ed. T
he p
e
rforman
c
e
of the pro
p
o
s
ed dem
osaicing
techni
que
is comp
ared wit
h
lea
r
ne
d sim
ultaneo
us
sp
arse codi
ng (LSSC)
[2
9],
Iterative Re
sid
ual
Interpolatio
n (IRI) [19], multi-directio
nal weight
ed inte
rp
olation and g
uided filter (M
DWI
-
GF
) [30],
MLRI [31], similarity and colo
r differen
c
e (SACD) base
d
demo
s
aicin
g
[7], adaptive resi
du
al
interpol
ation
(ARI) [32] a
n
d
multilaye
r
neural n
e
two
r
k (NN) b
a
se
d dem
osaici
n
g
techniq
ue [
22].
The
NN te
ch
nique
propo
sed in [2
2] be
ars the
clo
s
e
s
t simil
a
rity to the
work
propo
sed
he
re.
The
averag
e PSNR of the
re
d,
gree
n, blu
e
chann
el
an
d CPSNR i
s
con
s
ide
r
ed
for
compa
r
ison. T
h
e
results
obtain
ed a
r
e
sh
own in T
able
2
of the pa
pe
r. Com
pared
with multilaye
r ne
ural
net
work
(NN) ba
se
d
demo
s
ai
cing
tech
niqu
e [
22], the
p
r
o
posed demo
s
ai
cing
techn
i
que exhibits
an
improvem
ent
of 8.82 dB in
aver
ag
e. An improvem
ent
of 4.1 dB
and
4.23dB is
re
ported
ba
sed
on
our p
r
op
osed
techniq
ue a
gain
s
t the LSSC [29]
and
ARI [32]. Based on the results it is evid
ent
that our p
r
o
p
ose
d dem
osaicin
g techni
que o
utper
fo
rms th
e exist
i
ng state
of art dem
osaicing
techni
que
s.
Table 2. Perf
orma
nce com
pari
s
on
re
sult
A
l
g
o
ri
th
m
PSNR Red
cha
n
nel
PSNR Gree
n ch
annel
PSNR Blue
cha
nnel
CPSNR
LSSC [29]
40.53
44.31
40.65
41.44
IRI [19]
38.82
42.38
39.15
39.77
MDWI-GF [30]
–
–
–
37.3
MLRI [31]
40.59
42.97
39.86
40.94
SACD [7]
40.88
40.93
41.02
–
ARI [32]
40.81
43.66
40.21
41.31
NN [22]
–
–
–
36.72
Propose
d
44.93
47.12
45.14
45.54
5. Conclusio
n
In this pa
per a frequ
en
cy domain
ba
sed
dem
osaici
ng techniqu
e
is propo
se
d
.
The
luminan
ce a
n
d ch
romin
an
c
e info
rmatio
n obtaine
d from the frequ
ency dom
ain
is extracte
d. A
bilinea
r interp
olation tech
ni
que implem
e
nted on t
he luminan
ce a
n
d chromina
nce information
is
use
d
to
pa
rtially derive
t
he mi
ssing
compon
ents a
nd p
r
od
uce
an inte
rme
d
i
a
te imag
e. T
he
interme
diate i
m
age is
split
into patch
es.
The novelty o
f
the propo
se
d demo
s
ai
cin
g techni
que i
s
the ado
ption
of neu
ral n
etwork to l
e
arn from the
local
pat
ch
es a
nd
esti
mate the mi
ssi
ng
comp
one
nts.
A novel Ne
ural
Network Based Im
a
g
e
Re
con
s
truction Algorith
m
is present
ed in
the prop
osed
demosaici
ng
techniqu
e. The ex
perim
e
n
tal results a
nd perfo
rma
n
c
e compa
r
iso
n
results p
r
e
s
e
nted prove t
he rob
ustn
ess and de
mo
sa
icin
g efficie
n
cy over
exi
s
tent state of
ar
t
demo
s
ai
cing
techni
que
s.Evaluation of t
he p
r
op
osed demo
s
ai
cing
techni
que
on addition
al
im
age
datasets is
co
nsid
ere
d as t
he future of the wo
rk p
r
e
s
ented he
re.
Referen
ces
[1]
BE Ba
yer. Col
o
r imagi
ng arr
a
y. US Patent 3,
971,0
65. 19
76.
[2]
Ousman Bo
uk
ara, La
urent B
i
tjoka
a
, Gamraïkr
éo Dja
o
w
é
a
.
Nond
estructi
ve Determi
nati
on of Be
ans
W
a
ter Absor
p
ti
on C
a
p
a
cit
y
U
s
ing
CF
A Ima
ges A
nal
ys
is F
o
r Har
d
-T
o-Cook Eva
l
uati
on.
Internatio
na
l
Journ
a
l of Elec
trical an
d Co
mputer Eng
i
n
eeri
ng (IJECE)
. 20
13; 3(3): 31
7-3
28.
[3]
X Li, B Gunturk
,
L Z
hang. Image dem
osaic
i
n
g
: A sy
st
ematic
surve
y
.
SPIE
. 200
8; 682
2: 68
221.
[4]
Dubois
E.
F
r
e
que
ncy-do
mai
n
metho
d
s for
de
mos
a
icki
ng
of Bayer-sa
m
ple
d
col
o
r i
m
a
ges
. In Signal
Processi
ng Let
ters, IEEE. 2005; 12(12): 8
47-
850.
[5]
Nai-
Xi
an
g Li
an
, Lanla
n
Ch
an
g, Yap-Pe
ng T
an, Z
agor
odn
o
v
V.
Adaptive
F
iltering for C
o
lor F
ilter Array
De
mos
a
icki
ng.
In Image Processin
g
, IEEE
Transacti
ons o
n
. 2007; 1
6
(10):
251
5-25
25.
[6]
Lu F
a
ng, A
u
OC, Yan
Ch
e
n
, Katsag
ge
lo
s AK, Ha
nli
W
ang,
Xi
ng
W
en.
Joi
n
t D
e
mosa
icin
g
an
d
Subp
ixel-B
ase
d
Dow
n
-Sa
m
p
l
ing for Bayer I
m
a
ges: A F
a
st F
r
equency-
Do
ma
in An
alysis
Appro
a
ch.
In
Multimedia, IEEE
T
r
ansactions
on. 2012; 14(4): 1359-1369.
[7]
R Nir
ub
an, T
Sree
Re
ng
a
Raj
a
, R
De
e
pa. Sim
ilar
i
t
y
and
Vari
anc
e
of C
o
lor
Diff
e
renc
e Bas
ed
Demos
a
ici
ng.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2015; 1
3
(2): 2
38-2
46.
[8]
Den
g
w
e
n Z, Xiaol
iu S, Weim
ing D.
Co
lo
ur de
mos
a
icki
ng
w
ith direct
ion
a
l
filtering a
nd
w
e
ightin
g.
In
Image Process
i
ng, IET
.
2012; 6(8): 108
4-1
0
9
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Dem
o
sai
c
in
g of Color Im
ag
es b
y
Accurate Es
tim
a
tion of Lum
inance
(V.N.V. Satya Prakash)
55
[9]
Xi
ao Z
h
ou, F
a
nfan Ya
ng, C
h
un
xi
ao Z
h
ang,
Che
ng
y
o
u W
a
ng.
Improve
d
a
daptiv
e d
e
m
os
aickin
g
us
in
g
directi
ona
l w
e
i
ghtin
g.
In C
o
mputer Sci
enc
e & Ed
ucati
o
n
(ICC
SE), 9th
Internatio
na
l C
onfere
n
ce
on
.
201
4: 615-
618.
[10]
Z
hang D, Xia
o
lin W
u
.
Col
o
r demosaick
in
g via directi
o
nal li
ne
ar mi
n
imu
m mean
squar
e-error
estimation.
In Image Proc
essi
ng, IEEE
T
r
ansacti
ons o
n
. 200
5; 14(12): 2
167
-217
8.
[11]
D Meno
n, G Calva
g
n
o
. Color im
age
d
e
mosa
ickin
g
: An overvi
e
w
.
Signa
l Proc
essin
g
: Imag
e
Co
mmun
icati
o
n (Elsevi
e
r)
. 20
11; 26.
[12]
Kim Y, Jeo
ng
J.
F
our Directi
on R
e
sid
u
a
l
In
terpol
ation
for
De
mos
a
icki
ng.
In Circuits and S
y
stems for
Vide
o T
e
chnol
og
y
,
IEEE
T
r
ansactions o
n
. 99
: 1.
[13]
Mascha
l
RA, Youn
g SS, Re
yno
l
ds
JP, Krape
ls K, Fann
ing J, Cor
b
in T
.
N
e
w I
m
ag
e
Qu
al
ity
Assessment Algorithm
s for CFA Dem
o
s
a
ic
ing.
In Sensors Journ
a
l, IEEE. 2013; 13(
1): 371
-378.
[14]
Xi
an
gdo
ng C
h
en, Li
w
e
n H
e
, G
w
a
n
g
g
il J
eon
, Jechan
g Jeo
ng.
Multid
irecti
ona
l W
e
ig
hted
Interpol
ati
o
n
and R
e
fine
me
nt Method for Bayer Pattern CF
A Demosai
cking.
In Circu
its and S
y
ste
m
s for Vide
o
T
e
chnolog
y, IE
EE
T
r
ansactio
n
s on. 20
15; 2
5
(8): 127
1-1
2
8
2
.
[15]
W
a
chira Ki
n
y
u
a
, M
w
a
n
g
i Eli
j
ah, Jeo
n
G
w
a
ngg
il.
A pe
nto
m
i
no-b
a
se
d p
a
th ins
p
ire
d
d
e
mosa
ickin
g
techni
qu
e for the bay
er col
o
r filter array.
In AF
RICON. 2015
: 1-5.
[16]
Micha
e
l G, Kir
y
ati N.
Exa
m
p
le b
a
sed
de
mos
a
ici
ng.
In Image Pro
c
essin
g
(ICIP), 2014 IEEE
Internatio
na
l C
onfere
n
ce o
n
. 201
4: 183
2-18
36.
[17]
Mori
ya S, Mak
i
ta J,
Kuno T
,
Matoba N, Su
giur
a H.
Adva
nced d
e
m
os
ai
cing
meth
ods
base
d
on the
chan
ges
of col
o
rs in
a l
o
ca
l
regi
on.
In
Consumer E
l
ectr
onics, IEEE
T
r
ansactions
on.
2006; 52(1):
206-
214.
[18]
Lu YM, K
a
rza
nd M, V
e
tterli
M.
De
mos
a
icki
ng
by Alt
e
rnati
ng Pr
oj
ecti
ons:
T
heory
an
d F
a
st One-Ste
p
Im
plem
entation
. In Image Processi
ng, IEEE
T
r
ansactions
on. 201
0; 19(8)
: 2085-2
0
9
8
.
[19]
W
e
i Ye, Kai-K
uan
g Ma.
C
o
l
o
r Ima
ge
De
mosaici
ng
Usin
g
Iterative Res
i
dua
l Interp
olati
on.
In Image
Processing, IEEE
T
r
ansactions
on. 2015; 24(
12): 5879-
5891.
[20]
Burger H
C
, Schul
er CJ, Ha
rmelin
g S.
Image d
e
n
o
isi
ng:
Can pl
ain
ne
ural n
e
tw
orks compete w
i
th
BM3D?.
In Co
mputer Visi
on
and Pattern R
e
cogn
it
ion (CVP
R), 2012 IEEE Confer
ence o
n
. 2012: 239
2-
239
9.
[21]
Budi
Ra
hma
n
i
,
Supri
y
a
d
i.
Earl
y M
ode
l
of
T
r
affic Sign R
e
min
der
Based
o
n
N
e
ural
Net
w
o
r
k.
TELKOMNIKA.
2012; 1
0
(4): 7
49-7
58.
[22]
Yi-Qing
W
a
n
g
.
A multi
l
ayer n
e
u
ral n
e
tw
ork for ima
ge de
mos
a
ickin
g
.
In Image Process
i
ng
(ICIP), 2014
IEEE International Conferen
ce on. 2014:
1852-1856.
[23]
Alle
yss
on
D,
Su
sstrunk S,
Herault J.
Linear demosaic
ing ins
p
ired by
the hum
an vis
ual system
.
In
Image Process
i
ng, IEEE T
r
an
sacti
ons o
n
. 20
05; 14(4): 4
39-
449.
[24]
P Vand
e
w
a
l
l
e
, S S¨usstrunk,
M Vetterli. A
F
r
eque
nc
y Do
main A
ppro
a
c
h
to Re
gistrati
on of Al
iase
d
Images
w
i
t
h
A
pplic
atio
n to S
uper-R
eso
l
utio
n.
EURASIP J
ourn
a
l o
n
Ap
pli
ed Si
gna
l Proc
essin
g
, Speci
a
l
Issue on Su
per
-Resol
utio
n Imagi
ng
. 20
06.
[25]
D Robinson, P
Milanfar. Statistical performanc
e analy
sis of
super-res
o
lution.
IEEE Transactions
on
Imag
e Process
ing
. 20
06; 15:
141
3-14
28.
[26]
QV Le, A
Co
ates, B Proc
hn
o
w
,
AY
N
g
. On
optimiz
ation
m
e
thods
for
dee
p l
earn
i
ng.
L
earn
i
ng
.
20
11:
265-
272.
[27]
T
Q
Pham, LJ van Vliet, K Schutte. Rob
u
st F
u
sion of Irregul
arl
y
Sam
p
le
d Data Usi
ng Ada
p
tive
Normal
i
ze
d Co
nvol
ution.
EUR
ASIP Journal
o
n
Appl
ied S
i
gn
al Process
i
ng
. 200
6.
[28]
T
he Kodak color image datas
et, Image Avail
able [Online]. Avai
lable: http://r0k.us/graphics/kodak/.
[29]
J Maira
l
, F
Bac
h
, J Po
nce, G
Sapir
o
, A Z
i
ss
erman.
N
on-
lo
cal sp
arse
mo
dels f
o
r i
m
a
g
e
restoratio
n.
In
Proc. IEEE 12th Int. Conf. Comput. Vis. 200
9: 2272
–2
27
9.
[30]
Xi
an
gdo
ng C
h
en, Li
w
e
n H
e
, G
w
a
n
g
g
il J
eon
, Jechan
g Jeo
ng.
Multid
irecti
ona
l W
e
ig
hted
Interpol
ati
o
n
and R
e
fine
me
nt Method for Bayer Pattern CF
A Demosai
cking.
In Circu
its and S
y
ste
m
s for Vide
o
T
e
chnolog
y, IE
EE
T
r
ansactio
n
s on. 20
15; 2
5
(8): 127
1-1
2
8
2
.
[31]
D Kiku, Y Monno, M
T
anaka,
M Okutomi.
Mini
mi
z
e
d-l
a
p
l
acia
n resid
ual
interp
olati
on fo
r color i
m
ag
e
de
mos
a
icki
ng.
In Proc. IST
/
SPIE Electronic
Imagin
g
. 201
4; 9023: 1-8.
[32]
Yusuke
Mon
n
o
,
Daisuk
e Kik
u
,
Masa
yuk
i
T
anaka, Mas
a
toshi
Okutomi.
Ad
a
p
tive R
e
si
dua
l
Interpol
atio
n
for Color Imag
e De
mosa
icki
n
g
.
In IEEE
Internatio
nal C
onf
erenc
e on Ima
ge Process
i
ng
(ICIP). 2015:
1-5.
Evaluation Warning : The document was created with Spire.PDF for Python.