TELKOM
NIKA
, Vol. 11, No. 6, June 20
13, pp. 3242
~
3
250
e-ISSN: 2087
-278X
3242
Re
cei
v
ed
Jan
uary 18, 201
3
;
Revi
sed Ap
ril 7, 2013; Accepte
d
April 2
2
, 2013
Monte-Carlo SURE for Choosing Regularization
Parameters in Image Deblurring
Yubing Han
*
,
Kelan Wan
g
, Mengna X
u
Schoo
l of Elect
r
onic a
nd Optic
a
l Eng
i
ne
eri
ng,
Nanj
i
ng U
n
ive
r
sit
y
of Scie
nce
and T
e
chnol
o
g
y
, N
anj
ing,
Chin
a, 21
009
4
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: han
yb
@mai
l.njust.ed
u
.cn
A
b
st
r
a
ct
Para
meter ch
oi
ce is crucia
l to regu
lari
z
a
t
i
on-
base
d
i
m
ag
e d
ebl
urrin
g
. In this paper, a Mo
n
t
e Carl
o
meth
od
is
use
d
to a
ppr
oxi
m
a
t
e the
opti
m
a
l
regu
lari
z
a
t
i
on
para
m
eter i
n
t
he s
ense
of St
ein
’
s
u
nbi
ase
d
risk
estimate (SUR
E) w
h
ich has
b
een
app
li
ed to
imag
e de
bl
urr
i
ng. T
he pr
opo
sed a
l
gor
ith
m
i
s
suitab
le for th
e
exact de
blurr
i
n
g
functio
n
s as
w
e
ll as
thos
e of not be
ing
e
x
presse
d an
al
yt
ically. We ju
stify our clai
ms b
y
prese
n
ting
ex
p
e
ri
ment
al r
e
sul
t
s for SURE-b
a
s
ed
opti
m
i
z
at
io
n w
i
th tw
o diffe
rent re
gul
ari
z
a
t
ion
al
gorith
m
s
of
T
i
khon
ov a
nd t
o
tal var
i
ati
on r
egu
lari
z
a
t
i
o
n
. Experi
m
ent
res
u
lts show
the
v
a
lid
ity of the
pr
opos
ed
alg
o
rit
h
m,
w
h
ich has si
mi
l
a
r perfor
m
a
n
ce
w
i
th the mini
mum MSE.
Ke
y
w
ords
: Mo
nte-Carl
o, Stei
n
’
s un
bias
ed ri
sk estimat
e
(SURE
), imag
e d
ebl
urrin, T
i
kho
nov reg
u
lar
i
z
a
ti
on,
total variation
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Image d
eblu
r
ring i
s
ve
ry
common
in im
age
pro
c
e
s
si
ng field.
Ho
wever, ima
ge
d
eblurrin
g
is a
ill-p
o
sed i
n
verse p
r
o
b
le
m. The
co
nce
p
t of ill-
p
o
sed
pro
b
lem
s
go
es
ba
ck to
Hadama
r
d
in t
h
e
begin
n
ing
of
this
centu
r
y [1]. Had
a
ma
rd e
s
sentially define
d
a
p
r
oble
m
to b
e
ill-po
se
d if t
h
e
solutio
n
is not
unique o
r
it is not contin
u
ous fun
c
tion
of the data, if
an arbi
t
r
arily small
pe
rturation
of the data can cau
s
e a
n
arbitra
r
ily large pertu
r
bati
on of the solution. A popular strategy for
solving inve
rse pro
b
lem
s
is to use regul
arizati
on te
ch
nique
s. Re
gu
larization met
hod is
a useful
strategy
whi
c
h ca
n
stabili
ze the p
r
obl
e
m
and to
obt
ain a
useful
and
stable
solution. However,
w
h
en
a
p
p
l
ying
th
is
me
th
od
, th
e
us
e
r
is fa
c
e
d w
i
th t
he difficult ta
sk of adj
ustin
g
re
gula
r
ization
para
m
eter to
obtain be
st p
e
rform
a
n
c
e.
Gene
rally, the effect of recon
s
tru
c
ted i
m
age is m
e
a
s
ured by mini
mizing m
ean
squ
a
re
d
error
(MSE),
as
we
all
know, th
e MS
E depe
nd
s o
n
the o
r
igin
al
sign
al
whi
c
h
is g
ene
rally
is
unavailabl
e o
r
un
kn
own a
prio
ri, a
pra
c
tical
app
ro
a
c
h i
s
to
repl
ace
the true
MSE by so
me
estimate
in t
he
sen
s
e
of
Stein’s
unbia
s
ed
ri
sk e
s
ti
mate (S
URE), whi
c
h
de
pe
nds on
the
gi
ven
data and pro
v
ides a mean
for unbia
s
ed
risk estim
a
te
of the true MSE [2, 3].
In rece
nt years, the
SURE criteri
o
n has
bee
n e
m
ployed in v
a
riety of
den
osin
g proble
m
s for
cho
o
si
ng re
gula
r
ization
para
m
eters, i
n
that
ca
se
of
den
osi
ng
algorith
m
s a
r
e not
bein
g
expre
s
sed
a
nalytically. It
has
been
dem
on
strated that M
onte
Carl
o m
e
thod i
s
p
r
a
c
t
i
cabl
e in
cal
c
ulation of S
U
RE [4]. Ho
we
ver,
its application
is not limited to denosin
g ca
se. In
this pape
r, we ex
tend the SURE method t
o
a
much b
r
o
ade
r cla
ss
of
pro
b
lems.
This p
ape
r i
s
organi
ze
d as follo
ws. I
n
se
ction 2,
we introdu
ce
the image
degrade
d
model an
d regula
r
ization
method. Section
3 de
scribes g
r
adi
ent
descent me
thod and im
age
recon
s
tru
c
tio
n
by a
n
it
erative al
go
rithm ba
sed
on Ti
kh
on
ov and
tota
l variation
(TV)
regul
ari
z
ation
.
In sectio
n
4, we
exten
d
Mo
nte-Ca
rlo SURE te
chniqu
e to i
m
age
debl
urri
ng
probl
em
s. In se
ction
5, we p
r
e
s
e
n
t experim
ent
al
re
sults
and
demo
n
strate
nume
r
ically that
SURE,
comp
uted u
s
in
g t
he Mo
nte-Ca
rlo
strate
gy
, faithfully app
roa
c
h th
e true MSE
cu
rve.
Finally the co
nclu
sio
n
is gi
ven in se
ction
6.
2. Problem
Formulation
It is
well
kn
o
w
n th
at
sign
al
s a
r
e
inevitab
ly
deg
rad
ed
d
u
ring
a
c
q
u
isiti
on, tra
n
smi
ssion a
n
d
stora
ge p
r
o
c
e
ss. In mo
st case
s, there a
r
e two
kin
d
s of
degra
ded fa
ctors, one i
s
the determi
nisti
c
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Monte-Ca
rlo SURE
for Ch
oosi
ng
Regul
arization Param
e
ters in Im
age Deblu
rrin
g
(Yubin
g
Ha
n)
3243
factors, such
as the
defe
c
ts of
cam
e
ra itse
lf, the
defocus
blur,
the motion
blur, an
d th
e
atmosp
he
ric
disturban
ce
s,
whi
c
h a
r
e m
a
inly ca
used
by image
acquisitio
n
sy
stem, the othe
r is
rand
om fa
cto
r
s,
su
ch a
s
p
hotoele
c
tri
c
n
o
ise,
cha
nnel
noise an
d
so on. In g
ene
ral, we a
s
su
me
the noise follows a certain probability dist
ribution, such as Ga
ussi
an distri
bution.
Let
2
1
L
L
R
X
be
the
i
deal
discrete
sig
nal
of a
co
ntinuo
us
scene.
2
1
M
M
R
Y
is
the
observed
deg
rade
d
signal.
2
2
1
1
M
,L
M
L
are th
e
sizes of the o
r
igin
al and
ob
se
rved si
gnal
,
respe
c
tively.
The deg
ra
de
d model of the gene
ral si
g
nal model i
s
given by:
E
HX
Y
(1)
whe
r
e
C)
N(0,
~
E
is zero-me
an
white
Gau
ssi
an n
o
ise
with a v
a
rian
ce
of C,
H is d
e
termi
n
istic
part of
de
gra
ded model,
whi
c
h
i
s
assumed
a
s
a
li
near o
perator, and
re
pre
s
ents any kin
d
s of
distortio
n
, blu
rrin
g
and d
o
wnsam
p
le in th
e pro
c
e
ss of i
m
age a
c
qui
si
tion.
In the variational fram
ewo
r
k [5, 6], the recon
s
tru
c
tio
n
sign
al is o
b
tained in
ge
neral
by
minimizi
ng a
co
st function
al of:
X
R
HX
Y
2
1
X
2
2
J
(2)
whe
r
e
2
denote
s
the Eu
clide
an no
rm,
2
2
HX
Y
2
1
is the data fid
e
li
ty term that measures th
e
con
s
i
s
ten
c
y of X to the given data Y, and
X
R
is a suitable re
gula
r
i
z
ation functio
n
that often
penali
z
ed
the
lack of smo
o
thne
ss i
n
X. The dete
r
mi
nation of reg
u
lari
zation
pa
ramete
r
λ
is an
importa
nt task and the mai
n
goal
in this
work is to opti
m
ize
λ
given Y [6-9].
3. Image Rec
onstr
uction
based on
Re
gulariza
t
ion
Whe
n
the lin
ear op
erator
H is a blu
r
or conv
olutio
n operator, re
constructio
n
the origin
al
sign
al X from
the ob
servat
ion is
calle
d
deconvol
utio
n or d
eblu
rri
ng. Re
gula
r
ization metho
d
is
cru
c
ial to ima
ge deblu
rri
ng
processin
g
. There are tw
o issu
es to b
e
con
s
ide
r
e
d
in regul
ari
z
ati
on,
the type of
re
gulari
z
atio
n t
e
rm
and
the
sele
ction
of
regula
r
ization
para
m
eter,
th
ey all h
a
ve
cl
ose
con
n
e
c
tion
with the effe
ct of the
re
store
d
ima
g
e
.
In this pa
per,
we mai
n
ly discu
ss
two
regul
ari
z
ation
methods: Ti
khonov re
gul
a
r
ization a
nd T
V
regula
r
ization
3.1. Tikhono
v
Regulariza
t
ion
Gene
rally we can
choo
se
regula
r
ization function
a
s
dxdy
R
X
X
, and let
2
s
s
, where
X
is the gra
d
ient of X, then we h
a
ve
dxdy
R
2
Ω
X
X
, which
is Tikhon
ov
regul
ari
z
ation
[10, 11].
dxdy
λ
Y
J
2
X
HX
2
1
X
2
2
(3)
The Euler-L
a
g
ran
ge eq
uati
on of (3) i
s
:
Y
H
X
HX
H
T
T
2
(4)
Whe
r
e
is the Lapla
c
ian o
p
e
rato
r.
3.2. Total Variation (TV)
Regula
r
iza
t
ion
Whe
n
s
s
, then
regula
r
i
z
atio
n function b
e
com
e
s
dxdy
R
Ω
X
X
, this is
total
variation (TV
)
regula
r
i
z
atio
n [12].
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3242 – 3
250
3244
dxdy
λ
Y
J
X
HX
2
1
X
2
2
(5)
The Euler-L
a
g
ran
ge eq
uati
on of (5) i
s
:
Y
H
X
X
X
HX
H
T
T
div
(6)
If
0
X
, the diffus
i
on c
oeffic
i
ent
2
2
1
1
y
x
X
X
X
. That means,
in the flat region
of image, a
d
d
ing a
great
smo
o
thing
effect will le
ad to a
bad
stairca
s
e
effect. In ord
e
r
to
overcome thi
s
phe
nome
n
o
n
, we introd
u
c
e a pa
ram
e
ter
, s
o
:
dxdy
J
X
HX
Y
X
X
X
X
2
2
2
1
min
arg
min
arg
ˆ
(7)
And the co
rre
spo
ndin
g
Euler-Lag
ran
ge
equatio
n is:
Y
H
X
X
X
HX
H
T
T
div
(8)
Whe
r
e
2
2
2
y
x
u
u
u
.
3.3. Gradient Desc
ent Me
thod
Large-scale
equation
problem
i
s
ine
v
it
able in image resto
r
ati
on processin
g
. The
dimen
s
ion
of some
matrix i
s
too la
rge, fo
r
exampl
e, if the si
ze
of a g
i
ven image i
s
256
256, a
n
d
then the
si
ze
of ope
rato
r
H is 25
6
2
25
6
2
. The
r
efore, som
e
di
re
ct metho
d
s such
as Ga
uss
elimination
m
e
thod a
nd L
U
de
co
mpo
s
i
t
ion ca
n n
o
t be ap
plied i
n
pra
c
tice be
cause la
rge m
a
trix
can
not b
e
stored i
n
o
u
r
comp
uter. T
h
e iterat
ive
al
gorithm
s
can
avoid the
d
e
com
p
o
s
ition
of
matrix and th
e amou
nt of
stora
ge i
s
le
ss than th
at
of dire
ct metho
d
s. In this
pa
per,
we resto
r
e
ima
g
e
b
y
gr
ad
ie
n
t
de
sc
en
t me
th
od
w
h
ich
is a ty
pe of iterative
meth
od.
G
r
adi
ent desce
nt
is al
so
kno
w
n a
s
ste
epe
st desce
n
t
method. In this iterat
ive algorithm, the i
n
verse op
erat
ion of H can
be
avoided and some
pri
o
r knowl
edge
of the
sol
u
tion c
an be
effecti
v
ely combi
n
e
d
in the ite
r
at
ive
p
r
oc
es
s
.
A simple form
of the gradie
n
t descent m
e
thod is give
n by:
k
k
k
J
τ
X
X
X
1
(9)
Whe
r
e k i
s
the numbe
r of iteration.
is th
e iterative ste
p
, which i
s
a small en
oug
h
to ensure th
e
conve
r
ge
nce
of iterative
algorith
m
.
k
X
is
the
es
timate of
X
after
k ti
mes iterative
cal
c
ul
ation
and
k
J
X
is the negative gradi
e
n
t of
X
J
at
k
X
.
k
k
J
J
X
X
X
X
X
(10)
For the regul
arization fun
c
tion (2),
the n
egative gra
d
i
ent is given b
y
:
X
X
HX
H
Y
H
X
X
L
J
T
T
(11)
Whe
r
e
X
L
is the differential op
erato
r
of
X
R
, iterative algorith
m
is as follo
ws:
1
1
1
1
1
k
k
k
T
T
k
k
k
L
J
τ
k
X
X
HX
H
Y
H
X
X
X
X
X
X
X
(12)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Monte-Ca
rlo SURE
for Ch
oosi
ng
Regul
arization Param
e
ters in Im
age Deblu
rrin
g
(Yubin
g
Ha
n)
3245
4. Regulari
z
ation Param
e
ter
Determination
4.1. Choosin
g Regur
ariza
t
ion Parame
ter Bas
e
d On
The Minimum SURE
For the d
e
g
r
a
ded mo
del, th
e pro
bability den
sity
of the observed Y
can
be exp
r
e
s
sed a
s
the expone
ntial distrib
u
tion
[2].
X
Y
X
Y
X
Y
g
exp
b
f
T
(13)
Whe
r
e
Y
C
Y
C
Y
1
T
n
exp
)
det(
2
π
1
)
b(
2
1
)
(
,
Y
C
H
(Y)
1
T
φ
,
HX
C
H
X
2
1
(X)
1
T
T
g
. Ap
pare
n
tly, a
sufficie
n
t statistics for esti
mating X is given by
)
φ
(
Y
u
. Therefore, any re
aso
nabl
e of Y will be
a fun
c
tion
onl
y of u. Mo
re
spe
c
ifically, from the
Rao-Blac
kwell theorem [14], it follows
that if
λ
X
ˆ
is an estimat
e
of X which
is not a
function only of u, then the estimate
)
u
X
E(
λ
ˆ
has
lesser o
r
equal MSE than that of
u
X
λ
λ
h
ˆ
, therefo
r
e, in the seq
uel, we only co
n
s
ide
r
metho
d
s
that
depe
nd on th
e data via u.
Whe
r
e
u
λ
h
is a function of u that depend
s o
n
the obse
r
va
tions Y and
t
he sub
s
cri
p
t
denote
s
t
hat the esti
mation is
rel
a
ted to reg
u
l
a
rization pa
rameters. For the
es
timate
u
X
λ
λ
h
ˆ
, MSE is defined a
s
:
2
λ
2
2
λ
2
(
h
E
N
1
E
N
1
u)
X
X
X
ˆ
(14)
2
λ
2
λ
2
λ
2
λ
h
N
1
E
N
1
E
λ
u
X
X
X
argmin
ˆ
argmin
(15)
The sufficient
statistics u li
es in the
ran
ge space
)
(H
T
, s
o
u
X
λ
λ
h
ˆ
a
l
s
o
be
lo
ngs
to this spa
c
e
.
Denote by
H
HH
H
P
T
T
the orthog
on
al proje
c
tion
onto
, where
H is ra
nk-
deficie
nt. Then,
2
2
2
2
λ
2
E
E
E
E
E
P)X
(I
X
P
PX
X
P)
(I
P)X
(I
X
P
PX
X
X
λ
ˆ
ˆ
ˆ
ˆ
(16)
If
λ
X
ˆ
lies in the
space
, then
0
ˆ
λ
X
P)
(I
.and
P)X
(I
is con
s
ta
nt and in
dep
ende
nt
of
λ
X
ˆ
. T
h
er
e
f
or
e, in
th
is
c
a
s
e
, it is
suffic
i
ent
to obtain the es
timate of t
he firs
t term f
o
r
optimiz
ing
λ
X
ˆ
.
Next we co
n
s
ide
r
the sp
e
c
ific Ga
ussia
n
distrib
u
tion
expressio
n
. Suppo
se H
has the
sing
ular valu
e decompo
sit
i
on
T
Q
U
H
for some
unitary matri
c
es U a
nd Q.
Let H ha
s ra
nk r
so that
is a
m
n
diago
nal mat
r
ix wh
ose
the
first r
diag
on
al eleme
n
ts
are
equal
to
0
2
i
.
Proje
c
tion ma
trix is
T
VV
P
, whe
r
e
V equal
s to the first
r colu
mns of Q, a
n
d
let
X
V
X'
T
. The
s
u
ffic
i
ent s
t
atis
tic
for es
timating
'
X
is
u
V
Y
C
H
V
u'
T
1
T
T
, and
'
u
is a Ga
ussian rand
o
m
vector with a
mean
'
u
and a covari
an
ce
HV
C
H
V
C'
1
T
T
.
Usi
ng the SVD de
comp
osi
t
ion of H,
we have:
r
1
T
r
1
T
U]
C
Λ
[U
C'
X'
U]
C
Λ
[U
μ
'
(17)
Whe
r
e
Λ
is a
r
r
diagon
al mat
r
ix with
diago
na
l eleme
n
ts
0
σ
2
i
a
nd
r
[A]
is the
r
r
top-left
prin
ciple bl
ock of size r of the matrix A. Since
0
C
,
C'
is invertible. Therefore:
'
T
X
g
u'
X'
u'
X'
|
u'
exp
q
f
)
(
)
(
(18)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3242 – 3
250
3246
Whe
r
e
'
u
is the sufficient statistic of
'
X
,
)
|
(
X'
u'
f
is exponential
distributio
n,
X'
U]
C
Λ
[U
X'
X'
r
1
T
T
2
1
)
(
g
,
}
)
2
/
1
(
exp{
)
'
det(
)
2
(
1
)
(
u'
C'
u'
C
u'
1
T
n
q
.
In order to estimate
2
X
P
PX
E
λ
ˆ
, we must co
mpute:
}
VX'
X
E{
X}
VV
X
E{
PX}
X
E{
PX}
P
X
E{
T
λ
T
T
λ
T
λ
T
T
λ
ˆ
ˆ
ˆ
ˆ
(19)
Sinc
e
)
(Vu
u)
X
'
h
(
h
λ
λ
λ
ˆ
an
d
)
X'
|
(u'
f
is ex
pone
ntial fa
milies
dist
ri
bution. L
e
t
)
'
(V
V
)
'
(
T
u
h
u
k
is the estima
tion of
λ
X
ˆ
, then:
'
q
h
E
h
Tr
E
'
'
q
h
E
h
Tr
E
q
'
h
E
h
Tr
E
q
k
E
k
Tr
E
k
E
h
E
λ
λ
T
λ
λ
λ
λ
u
)
(u'
(u)V
u
(u)
P
u
)
(u
(u)V
u
(u)
VV
u'
)
(u'
)V
(Vu
u'
)
(Vu'
V
u'
)
(u'
)
(u'
u'
)
u'
)X'
(u'
)VX'
(Vu'
T
λ
T
T
T
T
T
T
T
λ
T
λ
ln
ln
ln
ln
(
(20)
So the unbia
s
ed estimate o
f
the MSE
X
X
E
ˆ
P
P
ca
n be obtain
e
d
by [2]:
ML
λ
λ
h
h
Tr
h
h
S
X
u
u
(u)
P
(u)
P
PX
(u)
T
T
λ
ˆ
2
-
2
)
(
2
2
(21)
Whe
r
e
Y
C
H
H)
C
(H
X
1
T
1
T
ML
ˆ
is the
maximum li
kelih
ood e
s
t
i
mation, and
the
)
(
den
otes t
h
e
Moore-Pe
nro
s
e p
s
eu
do in
verse.
4.2. Monte
-
Carlo Realization Of
The Unbiased
Ris
k
Estimation
The cru
c
ial st
ep for evalu
a
t
ing the SURE is the com
putation of
u
h
Tr
T
)
(
u
P
; h
o
wever,
in most
re
con
s
tru
c
tion al
go
rithms
(e
spe
c
ially iterative algorith
m
),
(u)
h
is not available
explicitly,
so
we u
s
e
Monte-Ca
rlo
techni
que to
app
roximate
the tra
c
e. T
o
co
mpute
u
h
Tr
T
)
(
u
P
, we
introduc
e
the
following theorem at firs
t.
Theorem
1 [3]: let
N
b
R
be a
ze
ro-m
ean
i.i.d. ran
dom ve
ct
or. Assum
e
(u)
h
has th
e
se
con
d
-o
rd
er Taylor expan
sion, so we h
a
ve
)
(
)
(
lim
)
(
0
u
P
b
u
P
b
u
u
P
b
h
ε
h
E
h
Tr
T
T
(22)
Proof: The se
con
d
-o
rd
er T
a
ylor expan
si
on of
)
(
b
u
ε
h
can b
e
written a
s
:
h
h
e
J
h
h
2
)
(
)
(
)
(
b
u
u
b
u
(23)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Monte-Ca
rlo SURE
for Ch
oosi
ng
Regul
arization Param
e
ters in Im
age Deblu
rrin
g
(Yubin
g
Ha
n)
3247
Whe
r
e
)
(
u
h
J
is the Ja
cobi
an
matrix of
(u)
h
evaluated at Y
and
h
e
represe
n
ts the e
rro
r
vector. In this case, the co
mpone
nts a
r
e
bound
ed in the expe
ctatio
n sen
s
e, then
:
2
2
)}
(
{
)
(
C
u
P
P
b
b
u
P
u
ε
b
u
P
b
b
h
h
T
h
T
λ
λ
T
b
J
tr
e
E
J
b
E
h
h
b
E
(24)
Whe
r
e
C
e
b
E
h
T
b
P
be
cau
s
e b
ha
s finit
e
high
er-orde
r
mom
ents
a
nd the
com
p
onent
s of
h
e
are bou
nde
d
in the expect
a
tion se
nse. P is proje
c
tio
n
matrix. Wh
en
0
, we have:
u
u
P
u
P
u
b
u
P
b
b
)
(
h
Tr
J
tr
(
h
)
ε
(
h
E
T
λ
h
λ
λ
T
)}
(
{
)
1
lim
0
(25)
So
)
(
)
(
lim
)
(
0
u
P
b
u
P
b
u
u
P
b
λ
T
T
h
ε
h
E
h
Tr
. Acco
rdi
ng to this t
heorem, we
prop
ose the followin
g
algo
rithm to appro
x
imate
u
(u)
P
λ
h
Tr
N
2
1
.
Step 1: generate a ze
ro-m
ean
i.i.d rand
om vector
N
R
b
with unit varianc
e
.
Step 2: evaluate
(u)
h
.
Step 3: let
b
u
ε
z
, and evaluate
(z)
λ
h
.
Step 4: comp
ute
(u)
(z)
P
λ
λ
T
h
h
b
N
2
1
and
u
(u)
P
h
Tr
N
2
1
.
5. Experimental Re
sults
In this
se
ctio
n, the
stand
a
r
d
cam
e
ra
ma
n an
d p
eppe
rs ima
g
e
s
with si
ze
of 2
5
6
×
25
6 a
r
e
adopte
d
as te
st image
s. In degradatio
n, 3×3
Gau
s
sia
n
ke
rnel
with
a varian
ce of
1 is u
s
ed to b
l
ur
the ori
g
inal i
m
age
s, an
d then the
white
Gau
ssi
an
noi
se
with a
stan
dard
deviatio
n
of 0.1 i
s
ad
de
d
to the blurred
image.
Tikho
nov reg
u
lari
zation
a
nd TV re
gul
ariz
ation a
r
e
adopte
d
in
deblu
rri
ng al
gorithm.
Figure 1
(
a
)
and Fi
gu
re
3(a
)
a
r
e th
e
origi
nal im
a
ges;
Figu
re
1(b
)
a
nd Fi
g
u
re
3(b)
are
the
degrade
d im
age
s; Figu
re
1(c)
and F
i
gure
3(c)
a
r
e the
re
sto
r
ed im
age
s
usin
g Ti
khon
ov
regul
ari
z
ation
with
optimi
z
ed p
a
ra
meters; Fig
u
re
1
(
d
)
a
nd Fi
gu
re
3(d
)
a
r
e
the
restored
imag
es
usin
g TV re
gulari
z
atio
n
with optimi
z
e
d
paramet
e
r
s. From the v
i
sual
persp
ective, the image
resto
r
e
d
by
T
i
kho
nov
regul
arization
can
not
pre
s
e
r
ve
the
detail
s
o
f
the e
dge
s
o
f
image,
whil
e
the image re
store
d
by TV regula
r
izatio
n is mu
ch b
e
tter in keepi
ng details of
image, and the
edge is m
o
re visible co
mpared to Tikho
nov re
g
u
l
arization. Th
e pea
k sign
al-to-noi
se ratio
(PSNR) val
u
es
are li
sted
in Ta
ble 1
where
the
outp
u
t PSNR ba
sed o
n
true M
SE and
Mont
e-
Carl
o SURE resp
ectively a
r
e given a
nd
the PS
NR b
a
s
ed o
n
the
s
e
method a
r
e
si
milar. Figu
re 2
and
Figu
re
4
are
the S
U
RE and
MSE curves of
cam
e
ram
an and
pepp
ers re
sp
ectively.
We
use
Monte-
Ca
rlo
SURE to sele
ct reg
u
la
rizatio
n
para
m
eter in
stea
d of MSE
whi
c
h ha
s b
een
descri
bed i
n
se
ction 4. It can
be
see
n
that t
he cu
rves of SURE
and MSE o
b
tained
by two
recon
s
tru
c
tio
n
algo
rithm
s
are ve
ry cl
ose.
No
w we co
mpa
r
e t
he regula
r
i
z
a
t
ion pa
ramet
e
r
λ
sele
cted by SURE an
d MSE. For came
raman imag
e in Tikho
nov re
gulari
z
atio
n, MSE and SURE
rea
c
h the
minimum al
most at the
same p
o
int
26
.
0
SURE
MSE
. In TV re
gulari
z
atio
n,
036
.
0
SURE
MSE
. While
for
pe
ppers i
m
ag
e,
the pa
ram
e
te
r
λ
sele
cted
b
y
MSE and S
URE
ha
s
a small
gap,
whe
n
the Ti
khonov regul
a
r
izatio
n is
used, the optim
um re
gula
r
ization pa
ram
e
ters
are
40
.
0
MSE
and
35
.
0
SURE
. For the
T
V
reg
u
lari
zat
i
on, the
opti
m
um
reg
u
larization
para
m
eters a
r
e
046
.
0
MSE
and
041
.
0
SURE
resp
ectively. Fro
m
these
resul
t
s, we can se
e th
e
effectivene
ss of the Mo
nte-
Carl
o SURE alg
o
rith
m and it is a good way to select
the
regul
ari
z
ation
param
eter in
image debl
urring.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3242 – 3
250
3248
6. Conclusio
n
In this
pap
er,
we
develo
p
e
d
the
unbia
s
ed e
s
timate
of the MSE i
n
imag
e d
ebl
urri
ng
by
extending
th
e SURE met
hod,
whi
c
h i
s
used to
cho
o
se
the
opti
m
al regul
ari
z
ation p
a
ra
me
ter
.
Comp
utation
and
appli
c
ati
on of
SURE
need
to eval
u
a
te the
tra
c
e,
ho
weve
r, the
co
mputatio
n
of
the tra
c
e m
a
y turn out to
be no
ntrivial,
esp
e
ci
ally
wh
en
the deblu
r
ring re
con
s
truction algo
rithm
doe
s not hav
e explicit ana
lytical form. In this
pap
er,
we u
s
e the
Monte-Ca
rlo
method to so
lve
this p
r
o
b
lem.
The
co
ntributi
on of
ou
r
wo
rk i
s
th
at the
Monte-Ca
rlo
SURE
metho
d
is exten
ded
to
the appli
c
atio
n of image d
ebluui
ng. Th
e advantag
e of
this metho
d
for sel
e
ctin
g parameters is
that the MSE can be e
s
ti
mated pu
rely
from t
he me
asu
r
ed d
a
ta
without ne
ed
the kno
w
le
dg
e
o
f
original image. Experiment result
s sho
w
that th
e o
p
timal pa
ram
e
ter
obtaine
d
by Mo
nte-Carlo
SURE is p
e
rf
ectly agre
ed
with the true
minimum val
ue of MSE.
Figure 1. Visual Com
p
a
r
ison of SURE-optimiz
e
d
De
blurred
Re
sul
t
s for Cam
e
ra
man. (a
)
Origin
al imag
e. (b) Degrad
ed image. (c) Rest
o
r
ed im
age by usi
ng
Tikho
bov re
g
u
lari
zation.
(d)
Re
store
d
image by usi
n
g TV regula
r
i
z
ation.
0
0.
01
0.
02
0.
03
0.
04
0.
05
0.
06
0.
07
0.
08
0.
09
0.
1
0
0.
005
0.
01
0.
015
0.
02
0.
025
0.
03
MS
E
Tr
u
e
M
S
E
SU
R
E
0
0.
05
0.
1
0.
15
0.
2
0.
25
0.
3
0.
35
0.
004
0.
006
0.
008
0.
0
1
0.
012
0.
014
0.
016
0.
018
0.
0
2
MS
E
Tr
u
e
M
S
E
SU
R
E
Figure 2. MSE(
λ
) and SURE(
λ
) for
Ca
meram
an. (a
)
MSE(
λ
) a
nd
SURE(
λ
) ba
sed on
Tikho
nov Re
gulari
z
atio
n. (b) MSE(
λ
) an
d SURE(
λ
) b
a
se
d on TV Regula
r
ization.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Monte-Ca
rlo SURE
for Ch
oosi
ng
Regul
arization Param
e
ters in Im
age Deblu
rrin
g
(Yubin
g
Ha
n)
3249
Figure 3. Visual Com
p
a
r
ison of SURE-opt
imize
d
De
blurred
Re
sul
t
s for Peppe
rs.
(a)Origi
nal im
age. (b
) De
graded ima
ge. (c
)
Rest
ore
d
image by usi
n
g Tikho
nov
regul
ari
z
ation
.
(d) Re
sto
r
ed
image by usi
ng TV regul
arization.
0
0.
01
0.
02
0.
03
0.
04
0.
05
0.
0
6
0.
0
7
0.
08
0.
09
0.
1
0
0
.
005
0.
01
0
.
015
0.
02
0
.
025
0.
03
MS
E
SU
R
E
Tr
u
e
M
S
E
0
0.
0
5
0.
1
0.
15
0.
2
0.
2
5
0.
3
0.
3
5
0.
4
0.
45
0.
5
0.
0
0
2
0.
0
0
4
0.
0
0
6
0.
0
0
8
0.
0
1
0.
0
1
2
0.
0
1
4
0.
0
1
6
0.
0
1
8
MS
E
T
r
ueM
S
E
S
URE
Figure 4. MSE(
λ
) and SURE(
λ
) for
ca
meram
an for
pepp
ers. (a)
MSE(
λ
) and
SURE(
λ
) ba
sed
on Tikh
onov regula
r
ization.
(b) MSE(
λ
) a
nd SURE
(
λ
)
based on TV
regul
ari
z
ation
.
Table 1. Co
m
pari
s
on of M
SE and SURE in Terms of
Output PSNR (dB)
Regularization
method
Image
Input
PSNR
Output PSN
R ba
sed
on MSE
Output PSN
R ba
sed
on SURE
Tikhonov
Camerama
n
19.7608
23.7006
23.7006
Peppers
19.9057
26.0399
26.0122
TV
Camerama
n
19.7608
24.9044
24.9044
Peppers
19.9057
27.8045
27.7497
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3242 – 3
250
3250
Ackn
o
w
l
e
dg
ement
This wo
rk
i
s
partially
supp
orted by
the Na
tion
al Natural S
c
ien
c
e
Found
ation o
f
Chin
a
(112
730
17).
Referen
ces
[1]
Had
a
mard J. L
e
ctures o
n
on
Cauc
h
y
'
s
Pro
b
l
e
m in
i
n
Li
near
Partial Differ
e
ntial Eq
uati
ons
. Ne
w
Hav
e
n
:
Yale U
n
ivers
i
t
y
Press; 1923.
[2]
Eldar YC. Gen
e
raliz
ed SUR
E
for Expon
enti
a
l
Families: App
l
icatio
ns to Reg
u
lariz
a
tio
n
.
IEEE Trans. on
Sign
al Process
i
ng
. 20
09; 57(
2
)
: 471-48
1.
[3]
Gir
y
es
R, E
l
a
d
M, El
dar
Y
C
. T
he proj
ec
ted
GSURE
for a
u
tomatic
param
eter tu
ni
ng
in
iterativ
e
shrink
age met
hods.
App
l
i
ed
and C
o
mput
ati
ona
l Har
m
o
n
ic
Analys
is
. 201
1; 30(3): 407- 4
2
2
.
[4]
Rama
ni S, Bl
u
T
,
Unser M. Monte-C
a
rlo S
u
r
e
: A Bl
ack-Box Optimization of R
egulariz
a
tion Param
e
ter
s
for General D
e
noisi
ng Al
gorit
hms.
IEEE Tra
n
sactions
on I
m
age Proc
essing
. 200
8; 17(9)
: 1540-1
5
5
4
.
[5] Karl
WC.
Reg
u
lari
z
a
tio
n
in i
m
a
ge restor
ati
on an
d recons
truction
. in Ha
ndb
ook of Image an
d Vide
o
process
i
ng, A. Bovik, Ed., 2nd
ed. Ne
w
York:
Elsevier. 20
05
: 183–2
02.
[6]
Park SC, Park
MK, Kang M
G. Super-reso
l
u
tion
imag
e re
constructio
n
: a
technic
a
l
ove
r
vie
w
.
IEEE
Sign
al Process
i
ng Ma
ga
z
i
n
e
.
200
3; 20(3): 21
–36.
[7]
Girard DA. T
he fast Monte-Ca
rlo
Cross-V
a
lid
atio
n a
n
d
CL pr
oce
dures
: comments, n
e
w
res
u
lts
an
d
app
licati
on to i
m
age rec
o
ver
y
probl
ems.
Com
p
ut. Statist.
,
199
5; 10: 205
–
231.
[8]
Golub
GH, He
ath M, W
a
h
b
a
G. General
ize
d
cross-val
i
d
a
tio
n
as
a m
e
tho
d
for ch
oosi
n
g
a
go
od r
i
d
g
e
param
eter.
T
e
chno
metrics
. 19
79; 21: 21
5-22
3.
[9]
Rice J. Choic
e
of smoothing
param
eter in d
e
conv
oluti
on p
r
obl
ems.
Cont
empor
ary Math.
, 1986; 59:
137
–1
51.
[10]
T
i
khonov
AN, Goncharsk
y A
V
,
St
epan
ov V
V
et al. N
u
mer
i
cal M
e
tho
d
s for the S
o
luti
on
of Ill-Pos
e
d
Probl
ems. Klu
w
e
r
Acad
emic
Publ
ishers. 19
95.
[11]
T
i
khonov AN.
Reg
u
lar
i
zatio
n
of incorr
ectl
y pose
d
pr
obl
e
m
s.
Soviet athematical Dok
l
ady
. 196
3; 4:
162
4-16
27.
[12] Youn
g
M.
T
he T
e
chnic
a
l W
r
iter
’
s
Ha
ndb
ook
.
Mill Vall
e
y
, C
A
: Universit
y
Sc
i
ence; 19
89.
[13]
Rudi
n LI, Osher S, F
a
temi E. Nonli
near
tot
a
l variati
on b
a
s
ed no
ise rem
o
val a
l
gor
ithm
s.
Physica D:
N
o
n
l
i
n
e
a
r
Phen
om
en
a
. 19
92;
60(1): 259-
26
8.
[14]
Sny
m
an JA. Practical Mathematical Optimizati
on: An Intr
oduction to Basi
c Optimization T
heor
y
and
Classic
a
l a
nd
Ne
w
Grad
ient-
B
ased Al
gorit
h
m
s. Springer P
ublis
hi
ng; 20
05
.
[15]
Ka
y S M.
F
und
amenta
l
s
of Statistical
Sig
n
a
l
Proce
ssi
ng: Es
timation
T
heory, U
p
p
e
r S
add
l
e
R
i
ver, N
J
:
Prentice H
a
ll, Inc.; 1993.
Evaluation Warning : The document was created with Spire.PDF for Python.