TELKOM
NIKA
, Vol.12, No
.4, Dece
mbe
r
2014, pp. 85
5~8
6
4
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i4.532
855
Re
cei
v
ed Se
ptem
ber 16, 2014; Revi
se
d Octob
e
r 29,
2014; Accept
ed No
vem
b
e
r
12, 2014
Image Deblurring Via an Adaptive Dictionary Learning
Strategy
Lei Li
1
, Ruiting Zhang
2
, Jiangmin Kan*
3
, Wenbin Li
4
1,3,
4
School of T
e
chn
o
lo
g
y
, Beij
ing F
o
restr
y
U
n
iversit
y
, 10
00
83, Beij
ing, C
h
i
na
1
Institute of Atmosph
e
ric Ph
ysics, Chin
ese
Academ
y of Sc
ienc
es, 100
029
, Beijin
g, Chin
a
2
Canvard C
o
ll
ege, Bei
jin
g T
e
chno
log
y
and
Busin
e
ss Univ
ersit
y
, 1
011
18,
Beiji
ng, Chi
n
a
*Corres
p
o
ndi
n
g
author: ka
nj
m@bjfu.e
du.cn
A
b
st
r
a
ct
Rece
ntly, spar
se repr
ese
n
tation
has
bee
n
app
lie
d to i
m
a
ge d
e
b
l
urri
ng.
T
he dicti
o
n
a
r
y
is the
funda
menta
l
p
a
rt of it
an
d
the
pro
per s
e
lecti
on
of d
i
ctionary
is v
e
ry i
m
porta
nt t
o
ac
hiev
e s
u
per
perfor
m
a
n
ce.
T
he glo
bal
l
earn
ed dictio
nary might achi
ev
e
inf
e
rior
perfor
m
a
n
ces si
nce
it c
oul
d n
o
t
min
e
t
h
e
specific inf
o
rmation suc
h
as the text
ure an
d edg
e w
h
ich is contai
ne
d in t
he blurre
d i
m
ag
e. How
e
ver, it i
s
a
computati
o
n
a
l
burd
en to
train
a n
e
w
dicti
o
n
a
ry for i
m
age
deb
lurri
ng w
h
i
c
h req
u
ir
es th
e w
hol
e i
m
age
(o
r
most parts)as i
nput; traini
ng the dicti
onary o
n
only a fe
w
patches w
ould r
e
sult in
over-fit
ting. T
o
addre
s
s
the prob
le
m, w
e
instea
d pro
p
o
se an
onl
ine
ada
ptio
n strategy to transfer t
he gl
ob
al le
arn
ed dicti
o
n
a
ry to a
specific i
m
age
. In our debl
u
rring al
gor
ith
m
, the sparse
coefficie
n
ts, latent imag
e, blu
r
kernel a
nd th
e
dictio
nary
ar
e upd
ated altern
atively.
An
d in
every step, the
glob
al l
ear
ne
d
diction
a
ry is u
pdate
d
in
an o
n
lin
e
form v
i
a s
a
mpl
i
ng
on
ly few
traini
ng
patc
hes
from
the tar
g
et no
isy i
m
age
. Since
our
ad
aptive
dicti
o
n
a
r
y
expl
oits the s
p
ecific i
n
for
m
ati
on,
o
u
r d
ebl
urr
i
ng
al
gorith
m
s
how
s sup
e
ri
or perfor
m
a
n
ce o
v
er
otherst
ate-
of-
the-art alg
o
rith
ms.
Ke
y
w
ords
: sp
arse repr
esent
ation, ad
aptive
di
ction
a
ry lear
nin
g
, ima
ge d
ebl
urrin
g
1. Introduc
tion
Image bl
uri
s
usually cau
s
ed
by relative motion
b
e
twee
n the
came
ra
andt
he
scene
durin
g the ex
posure time,
e.g., camera
sha
k
e. A
ndthi
s process ca
n be mod
e
led
as the follo
wi
ng
equatio
n:
(1)
Whe
r
e
i
s the
desi
r
ed
sh
arp
image;
is the
bl
urred im
ag
e,that is, the
degrade
d im
age
;
is the blu
r
ke
rnel that is a
s
sumedto
be
linear
ly shift-invariant;
is the additive noiseu
s
ually
assume
d a
s
following
Gau
ssi
an di
stributio
n;
d
enote
s
the
convolution
o
perato
r
.
Ima
ge
deblu
rri
ng
ca
n be
seen
a
s
aninverse
p
r
o
b
lem
of Eq.(1
)
. Given
only
the de
graded
imag
e a
s
in
p
u
t,
the goal
of bl
ind imag
e de
blurring i
s
to
inve
rse
the a
boveprocess and
to re
cov
e
r
b
o
th
and
,
whi
c
h a
r
e a
s
sume
d tobe u
n
kn
own in this ca
se.T
his i
s
as so-calle
d blind imag
e d
eblurrin
g
whi
c
h
is a fund
ame
n
tal and
chall
engin
g
probl
em in theima
ge and
sig
n
a
l
pro
c
e
ssi
ng l
i
terature [1]-[3].
Image
debl
urring
ha
s
ma
ny practi
cal
appli
c
ation
s
,such
a
s
rem
o
ving the
blu
r
from
con
s
u
m
er
photog
rap
h
s,
comp
utation
a
l photog
rap
h
y and ast
r
o
nomical imag
ing.Ho
weve
r,it is still an o
pen
que
stion to d
e
sig
n
an effici
ent deblu
rri
ng
algorith
m
.
Traditio
nal a
ppro
a
che
s
to rem
o
ve i
m
age
blur
are
alway
s
d
one by e
m
ploying
deconvolutio
n metho
d
. Ho
wever, th
e d
e
c
onvol
ution i
s
a
severely ill
-po
s
ed
p
r
obl
em, that is,th
ere
can
be expo
n
entially many
image
s
and kernel
s
that
satisfy Eq.(1).
To alleviate t
hese issue
s
,
many p
r
ior a
s
sumption
sa
nd regul
aritie
s a
r
e i
n
tro
d
u
c
ed
on
the
st
ructu
r
e
of
an
d
[3]-[9]. Th
e
statistical pri
o
rs of natural images are oft
en utilized. And a
commonly used prior on
is t
he
heavy-tailed
prio
r [4], that gradi
ents
of natural
ima
g
e
s
follow a hyp
e
r-L
apla
c
ian
di
stributio
n whi
c
h
wa
s ob
serve
d
byLevin et al..
Re
cently, usi
ng re
dun
dant
rep
r
e
s
entati
ons
and
spa
r
se p
r
op
erty i
n
nature ima
ges
ha
s
dra
w
n a lot of rese
arch
attention. Andsp
a
rse
represe
n
tation h
a
s al
so be
en
applied in i
m
age
deblurri
ngand achieved
sati
sfactory performance [7],[10]. Ca
i et
al.proposed
a blind motion
deblu
rri
ng
method by
exploitingth
e
sp
arsen
e
ss
of natu
r
al image
s
in over-com
plete
y=
x
k
+
n
x
y
k
n
xk
x
k
xk
x
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 855
– 864
856
pred
efined
dictionary (e.g. wavelet
s
, DCT
) to h
e
lp with kerne
l
estimation
a
nd sharp im
age
estimation [7
]. Jia et al. prop
osed a
newnon
-blin
d
image d
ebl
urri
ng meth
o
d
whi
c
h joi
n
tly
modele
d
the
sparse rep
r
e
s
entation of na
tural imag
e
p
a
tche
s an
d sparseg
r
a
d
ient
priors [11]. Li
et
al. prop
osed
a blind im
a
ge de
blurring
method
by
combi
n
ing th
e sh
arp
and
blur di
ction
a
ry
paira
nd th
e
spa
r
se g
r
a
d
i
ent pri
o
r,
an
d e
s
timating
the bl
ur
ke
rnel,sh
a
rp
im
age
and
sp
a
r
se
coeffici
ents resp
ectively d
u
ring
thed
ebl
urri
ng
pro
c
e
s
s [12]. T
he
redun
danta
n
d
over-comple
t
e
diction
a
ry ha
s been traine
d on image p
a
tche
s to hel
pexploit the sparse pri
o
r of
natural imag
es
[7],[11],[12].
Although de
b
l
urri
ng with a
prespe
cified
dictiona
ry is simplea
nd fast, it results in low
perfo
rman
ce
in mo
st
ca
se
s [13].
The
r
e
a
so
n i
s
that
the g
ene
ric b
a
se
s
(i.e., p
r
e-defin
ed
ba
sis
su
cha
s
curve
l
ets and fram
elets) d
o
not
exploit rich in
formation
c
o
n
tained in the b
l
urred imag
e or
the global lea
r
ned di
ction
a
rydoes not tra
n
sfer
well to
different types of image
s. In [14],Hu et al.
prop
osed a
deblu
rri
ng
method th
at exploited
s
parse
rep
r
e
s
e
n
tation u
s
ing
a blu
rry-sh
a
rp
diction
a
ry p
a
i
r
lea
r
ne
ddi
re
ctly from the
blurre
d
imag
e. And the
d
eblur kern
el
and
sha
r
p im
age
were estim
a
ted iteratively.
Ho
wever, the
adaptive strategy
[13] to
learn a n
e
w
diction
a
ryis d
e
sig
ned to work
wit
h
overlap
p
ing
p
a
tche
s
(on
e
per-pixel)of whole im
age. There
a
r
e ab
out
25
000
0
p
a
tche
s (8x8
) in
a512x5
12 im
age. Trainin
g
with such re
dund
ant pat
ches f
r
oma
sin
g
le imag
e i
s
time-con
sumi
ng.
Conve
r
sely, trainin
g
witho
n
ly a few pat
che
s
l
ead
s to
over-fitting. To bal
an
ce th
edilemm
a on
how
to quickly learn the di
ctionary an
d yetachi
e
ve higher a
c
cu
ra
cy, we pro
p
o
se a do
ma
in
adaptatio
nap
proa
ch
tra
n
sf
erri
ng th
e gl
o
bal le
arn
ed
di
ctiona
ry to th
esp
e
cifi
c ima
ge fo
r d
eblu
r
ring.
In our deblu
r
ring algo
rithm,thesp
a
rse co
efficients,
late
nt image, blur kern
el and th
edictio
nary are
update
d
alte
rnatively. And at every
ste
p
, thegl
ob
al l
earn
ed
dictio
nary i
s
u
pdat
ed in
an
onli
ne
form viasamp
ling only a
fe
w trai
ning
pat
che
s
from th
e blu
rre
d ima
ge,whi
c
h
ma
de the
updati
ng
pro
c
e
s
s mo
re efficie
n
t. T
he n
e
w lea
r
neddi
ctiona
ry
exploits the
sp
ecifi
c
inf
o
rmatio
n in
the
blurredim
age.
Therefore, it can
represent
patche
s
mu
ch better t
han the pre
s
pe
cifi
ed diction
a
ry’
s
.
Comp
ared t
o
the d
eblu
rri
ng meth
od
s [12],[15
], our m
e
thod
sho
w
s the
sup
e
rio
r
d
eblu
rrin
g
perfo
rman
ce.
The
re
st of t
h
is
pap
er i
s
orga
nized
as follows. O
u
r ne
w im
age
deblu
rri
ng
al
gorithm
is
descri
bed i
n
detail in
se
ction 2.Experim
ental re
su
lt
s and comp
ari
s
on
with ot
her state-of-the
-art
approa
che
s
a
r
e presented i
n
se
ction 3. Ou
r
work a
r
e
summ
ari
z
ed i
n
se
ction 4.
2. Image De
blurring Via Diction
a
r
y
A
d
aption
2.1 Problem formulation
It is well-kn
o
w
n that n
a
tural image
pat
che
s
can be modele
d
via spa
r
se
repre
s
entatio
n
over a
n
over-compl
ete di
ctionary. L
e
t
is
an
imag
e pat
ch,
i
s the dim
ensi
on of the
feature,
is an over-co
m
plete
di
ctio
nary with
atom
s. Th
en th
e
rep
r
e
s
entatio
n can
be fo
rm
ulated
as
follow [13]:
(2)
Whe
r
e the re
pre
s
entatio
n coeffici
ent
is sparse.
In the
spa
r
se
re
pre
s
e
n
tation,
is the fu
n
damental
pa
rt anda
p
r
ope
r
sele
ction
of
is
very importa
nt. Popular
sele
ction
s
fo
r
are: Fou
r
i
e
r, wavel
e
t, wavelet pa
cket, co
sine
packet,an
d
curvelet et
c. [1
6]. Neve
rthel
ess, t
he
pred
efined
dictio
n
a
ryha
s it
s a
d
v
antage
s in
some
kind of p
r
obl
ems. The
r
efo
r
e,no
single
method
cu
rre
n
tly can re
prese
n
t an ima
ge de
sira
blya
nd
gene
rally. L
earni
ng the
diction
a
ry
adaptively
from
sam
p
leimag
es i
s
an
othe
r
option.
Neverth
e
le
ss,
the global
learn
eddi
ctio
nary doe
s
n
o
t exploit
rich info
rmatio
n contain
e
d
in
theblurre
d image, and the
r
efore, it may not
tran
sfer well to thespe
ci
fic image.
To overcome
this probl
em,
we co
uld ap
ply a si
milar
strategyli
ke K
SVD [13] to learn a
n
adaptive dicti
onary on the blurredimage. KSVD is
a kind of generali
z
at
ion of the K-means
algorithm. Given a
s
e
t
of image pat
ches
, KSVD learns
the dic
t
ionary
by
solving the fol
l
owin
g pro
b
le
m:
(3)
xR
n
n
DR
nk
k
x=D
R
k
DD
D
12
[
y
,y
,
.
.
.
,y
]
N
Y
D
2
20
D,
1
a
r
g
m
i
n
||y
-
D
||
.
.
||
||
,
.
i
N
ii
i
i
s
tT
i
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Im
age Deblu
r
ring Via an A
daptive Di
ctio
nary Lea
rni
n
g
Strategy (Lei
Li)
857
Whe
r
e
is the
numb
e
r of
p
a
tche
s,
is th
e maximum
o
f
nonzero ent
ries of
. Sin
c
e
probl
em (3)
is no
n-conve
x
, the dict
io
nary
and the
coeffici
ent
are learned
iteratively whi
l
e the other variabl
e is fixed.
Ho
wever, KS
VDis
de
signe
d to lea
r
n the
basi
s
with ov
erlap
p
ing
pat
che
s
(one
pe
r-pixel
)
of the
wh
ole i
m
age. T
hen
u
m
ber of im
ag
e pat
ch
e
s
is redun
dant
an
d ab
out
2500
00pat
che
s
(8
x8)
in
a 512x5
12 image.
T
r
aini
ng with su
ch manypatche
s
from
a
singl
e
image
i
s
tim
e
-con
sumi
ng,
as
the auth
o
rcla
imed that
ea
ch ite
r
ation
of their
algo
rithm take
s
a
pproxim
ately5 minute
s
o
n
a
wind
ows P
C
of 2.67 G
H
z
CPU
and
4
GBRAM [14].
Co
nversely, tr
ainin
g
with a few pat
ch
es
lead
sto over-f
itting. To balance the dile
mma on
ho
w to quicklylea
rn an ad
aptive dictiona
ry a
nd
yet achi
eve hi
gher a
c
curacy,we p
r
opo
se
ou
r ima
ge
de
blurring
alg
o
ri
thm via di
ctio
naryad
aptatio
n
to a sp
ecifi
c
domain. An
d
based o
n
th
e above
discussion,
we p
r
opo
se
ou
r i
m
age d
eblu
r
red
model an
d co
ncretethe reg
u
lari
zation te
rms:
(4)
Whe
r
e
. Eq.(4)
ha
s fou
r
te
rms, t
he first t
e
rmi
s
the
blu
r
red
imag
e
wh
ich
ca
n b
e
spa
r
sely rep
r
ese
n
ted und
er theblu
r
di
ctiona
ry
; the seco
nd term is the spa
r
se p
r
io
r ofthe
gradi
ent im
a
ge a
s
sume
d
to follo
w th
e hype
r-Lapl
acia
ndi
stributi
on [3], while
is the
g
r
adie
n
t
extraction filters a
nd
; the third term is
a
norm reg
u
l
a
rization to the blur kerne
l
estimation; a
nd the la
st term is th
e regula
r
iz
ation
on the targ
et dictiona
ry
with
, whi
c
h
controls the
compl
e
xity of the targ
et d
o
main
to
avoid the ove
r
-fitting proble
m
and m
a
ke
sit
possibl
e with
a few trainin
g
data for dicti
onary lea
r
nin
g
.
2.2 Optimiza
tion Algorith
m
Obviou
sly, the obje
c
t funct
i
on in pro
b
le
m (4
) i
s
non
-convex. It can be ea
sily pro
v
ed that
probl
em (4) i
s
jointlyco
n
ve
x with re
spe
c
t to all of its variabl
es. And
we propo
se
analg
o
rithm t
hat
alternately o
p
timize
s one
variable whi
l
e the ot
hers are fixed. The metho
d
is descri
b
e
d
in
Algorithm 1.
1)
Sparse repr
esen
tatio
n
sub-pro
b
lem:
At
the begin
n
ing ofea
ch iteration, we fix
and
to
estimate the
coeffici
ent
of each image p
a
tc
h, the prob
lem is then transfo
rme
d
to:
(5)
Whe
r
e
. Ob
viously, we
coul
d de
com
pose theEq.
(5) fo
r ea
ch image
pat
ch
sep
a
rately, a
nd solve e
a
ch probl
em a
s
follow:
(6)
There is only
one unkno
wn variable in
the abov
e op
timizationp
ro
blem, whi
c
h can be
solve
d
efficiently via the ba
si
s pursuitalg
o
rithm [
17] or the accel
e
rated p
r
oxim
al gra
d
ient
(APG)m
ethod
s [18],[19].
2)
Late
nt imag
e estima
t
ion
sub-pro
b
le
m:
Similar to
the ideas p
r
opo
sed in [2
0],we assu
me
thesp
a
rse
co
efficients
of the late
nt ima
ge p
a
tch
with
are th
e
sa
me a
s
th
e
o
ne of
blurredim
age
relative to
.Thus,
wh
en
the
curren
t estimation
is
kno
w
n,
we
can
r
e
c
on
stru
ct latent imag
e
v
ia the following p
r
obl
em
:
N
0
T
i
D
12
[
,
,
...,
]
N
2
1
,
x
,k
,D
1
22
20
{
x
,
k
,
D
}
a
r
g
m
i
n
|
|
y
-
x
k
|
|
(
||
R
y
-
D
k
|
|+
λ
||
||
)
||
G
x
||
+
|
|
k
|
|
+
γ
||
D
-
D
|
|
i
i
Fi
i
i
i
F
,
DD
k
b
D
b
G
[0
.
5
0
.
8
]
2
L
D
0
D
kD
a
2
21
1
{
}
a
r
g
m
in
(
||R
y
-
D
|
|
+
λ
||
|
|
)
i
k
ii
b
i
i
i
Dk
kk
bb
D
2
21
{}
a
r
g
m
i
n
|
|
R
y
-
D
|
|
+
λ
||
|
|
)
i
k
ii
b
i
i
i
Rx
i
D
Ry
i
D
b
i
Rx
i
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 855
– 864
858
(7)
The solution to
problem
(7) can
be
obtained qui
ckly
accordingto [10].In practi
ce, we set the
as fi
rst-order
gradi
ent
filters
,
and
set
the
same
a
s
confi
guratio
ns in
[10].
3)
Blur kern
e
l estimation su
b-problem
:
As the other
sub
-
p
r
obl
ems, all other variable
s
excep
t
k are fixed. T
he minimi
zati
on of model (4) re
du
ce
s to the followin
g
probl
em:
(8)
This is a l
e
a
s
t squa
re
p
r
oblem
with
T
i
kho
nov
regu
larization,whi
c
h i
s
ea
sily
solved
in th
e
freque
ncy
-
do
main.
4)
Diction
a
r
y
adaption
sub
-
problem
:
In
this
su
b-pro
b
lem,we
up
d
a
te the
dicti
onary
of the
latent image
whi
c
h ha
s be
en estimate
d
inthe
latent image e
s
timat
i
on sub
-
p
r
obl
em. We set
other vari
able
s
exce
pt
fixed, and dictio
nary updat
ing de
grad
es to the
followin
g
pro
b
lem:
(9)
Whe
r
e
i
s th
e
identity mat
r
ix,
,
,
is th
e nu
mbe
r
o
f
patch
es
used
toupdate
the
diction
a
ry, a
n
d
and
.Probl
em (9
)
can
b
e
solved in the
followin
g
clo
s
ed formul
atio
n:
(10)
Ho
wever, th
e
inverse
of a
matrix is time
-c
o
n
sumin
g
.
Weutili
zed
th
e sto
c
h
a
sti
c
gradi
ent
desce
nt [21] to update the
d
iction
ary effectivel
y and efficiently. This on-li
ne me
thod iseffe
ctive
and ea
sy to impleme
n
t. The algorith
m
for updatin
gt
he
diction
a
ry is p
r
esented in Al
gorithm 2.
Algorithm 1
Blind image d
eblurrin
g
via diction
a
ry ad
aption.
-
Input: blurred
image
, sharp di
ctiona
ry
, kernel si
ze,i
teration n
u
m
ber
, the number of
patch
es .
-
Output: Estimated latent image
and blur
kernel
.
-
Initialization:
1. For:
;
2.
Sparse re
pre
s
entatio
n: up
date sp
arse coefficient
s
v
ia
minimizi
ng th
e sub
-
p
r
obl
e
m
(5);
3.
Latent image
estimation: u
pdate the late
nt
image
via minimizi
ngthe
sub-problem
(7);
4.
Blur ke
rnel e
s
timation: upd
ate the blur
kernel
viaminimizing the
su
b-p
r
obl
em (8
);
5.
Dictio
nary Up
dating: upd
ate the sha
r
p di
ct
iona
ry
via minimizi
ng th
e sub
-
p
r
obl
e
m
(9);
6. end
Fo
r.
Algorithm 2
Optimiz
a
tion for Problem (9)
-
Input:
, latent
image
, sparse c
oeffici
ents
, thenumbe
r of patche
s
.
-
Output:Dic
tionary
.
22
2
x
a
r
g
mi
n
|
|
y
-x
k
|
|
(
|
|
R
k
-D
|
|
)
|
|
G
x
|
|
Fi
i
i
x
G
1
G[
1
,
1
]
21
GG
T
0.
5
22
k
a
r
g
m
i
n
|
|
y
-
x
k
|
|
||k||
F
F
k
x
D
22
20
D
1
0
D
D
D
a
r
g
m
i
n
(
||
R
x
-
D
||
+
|
|
D
-
D
||
)
arg
m
i
n
(
t
r(D
D
(
C
C
+
I
))-
2D
(C
X
+
D
)
)
a
r
g
m
in
t
r
(
D
D
A
)
2
tr
(
D
B
)
M
k
ii
F
i
TT
T
T
TT
I
12
X=[
R
x
,
R
x
,
,
R
x
]
M
12
C=[
,
,
,
]
M
M
A=
C
C
I
T
0
B=
C
X
D
T
1
0
D
(
CC
I)
(C
X
D
)
TT
y
0
D
T
M
xk
k
1,
2
,
,
tT
i
x
k
D
0
D
x
i
M
D
t
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Im
age Deblu
r
ring Via an A
daptive Di
ctio
nary Lea
rni
n
g
Strategy (Lei
Li)
859
1.
Ran
domly draw
patche
s
from
latent ima
g
e
to form
and their
corre
s
p
ondin
g
sp
arse coefficient
s
.
2. Cal
c
ulate
and
;
3. For:
to
do
4.
Upd
a
te the
column via opt
imizing Probl
em (9):
,
5. end
Fo
r.
3. Experiments
The propo
se
d algorith
m
which i
s
imple
m
ented an
d corre
s
p
ondin
gexperi
m
ent
s to verify
it
sef
f
e
ct
iv
ene
ss
a
r
e
ca
rrie
d
out
i
n
MA
T
L
A
B
.
In all o
u
r
experim
en
ts, we
set
,
,
,
,
,
e
mpiri
c
ally.
T
he initial sharp
di
ctionary
is learned from th
e
s
h
arpimages
us
ing
KSVD [13].
The
initial kernel
is
set to be
theGauss
i
an
k
e
rnel with
.
Each
iteratio
n of
our alg
o
r
ithmtakes a
pproxim
ately
2 min
u
tes on
a
wind
ows
PC of
2.50G
Hz
CPU an
d 4 G
B
RAM. Blurred image
s for experime
n
ta
re synthe
sized
with three dif
f
erent ki
nd
s o
f
blur kern
el,in
c
ludi
ng motio
n
blur ke
rn
el(dire
c
tion 45
degree an
d length5 pixel
s
), Gaussia
n
blur
kernel
(sta
nda
rd deviation
5 pixels)an
d
averag
e blur ke
rnel (5
pixels), an
d then addit
i
ve
Gau
ssi
ann
oise followin
g
st
anda
rd devia
tion of
0.01 are ad
ded to
theblurred i
m
age. Fo
r color
image
s, we d
i
rectly
split th
em into th
ree
ch
ann
els,
RGB and
de
bl
ur e
a
ch
cha
n
nel respe
c
tively.
We compa
r
e
d
our re
sult
with the state
-
of-the
-a
rt
algorithm [12],[15],[23] which
is a blind ima
ge
deblu
rri
ng m
e
thodb
ased
on de
co
nvolu
t
ion metho
d
.
The final
re
sults a
r
e eval
uatedin te
rm
s of
PSNR
(pe
a
k sig
nal to
n
o
ise
ratio
)
a
nd SSIM(st
ru
ctural
si
milari
ty index) [22
]. Becau
s
e t
h
e
authors
did n
o
t publi
s
h th
e
i
r code,
we
directly q
uote t
he de
blu
rrin
g
re
sults fo
r th
e gray image
s in
[12]. For the colo
r imag
es,
we only com
pare o
u
r al
go
rithm with the
ones in [15].
Table 1. qua
n
t
itative comparative evalu
a
ti
on. PSNR i
s
ch
osen a
s
perfo
rman
ce
measure
File Blurt
y
pe
PSNR
Dilips'
s
method
Li's method
Dong’s
methed
Our
method
house.tif
motion 27.96
29.01
26.16
31.07
gaussian 27.06
29.26
30.60
30.05
average
28.21
29.
29
31.04
31.09
boats.tif
motion 27.13
28.55
25.52
29.95
gaussian 26.73
27.11
29.13
29.27
average
26.43
27.98
29.01
29.07
cameraman.tif
motion 25.83
26.22
25.86
28.71
gaussian 21.67
25.59
25.56
26.50
average
23.77
26.18
24.32
26.15
barbar
a.tif
motion 25.92
25.99
26.36
28.71
gaussian 21.76
26.02
25.29
26.50
average
24.97
24.67
24.82
26.15
baboon.tif
motion 23.54
24.44
25.43
25.71
gaussian 22.89
24.99
24.98
24.12
average
22.51
23.58
23.63
24.02
*Best re
sults
are note
d
in b
o
ld.
The exp
e
rim
ental results
of deblu
r
red i
m
age i
n
term
s ofPSNR a
n
d
SSIM are
p
r
esented
in Table 1 an
d Table 2 re
spe
c
tively. From the two table
s
, we ca
n con
c
lu
de that our alg
o
ri
thm
sho
w
s su
peri
o
r pe
rform
a
n
c
e than
Dilips’s onb
oth
PSNR a
nd SSIM. From the de
blurred ima
g
e
in
Figure 1 a
nd
Figure 2,the
r
e exists
noi
se
amplificat
io
n
in sm
ooth reg
i
ons i
n
Dili
ps’
s
meth
od, su
ch
as fa
ce
re
gi
on in B
a
rb
ara, ba
ckgro
u
n
d
re
gi
on i
n
House an
d
Cameraman. I
n
contra
st, o
u
r
algorith
m
is able toget m
o
re a
c
ceptab
le resu
lts on
the deblu
rre
d
images in th
ese
s
mo
oth a
nd
M
x
12
X=[
R
x
,
R
x
,
,
R
x
]
M
12
C=[
,
,
,
]
M
A=
C
C
I
T
0
B=
C
X
D
T
1
j
k
-t
h
j
1
u(
b
D
)
d
A[
,
]
jj
j
j
jj
2
1
du
max
(
u
,
1
)
j
j
j
0.01
15
00
0.
01
5
T
0.
1
2
000
M
0
D
k
1
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 855
– 864
860
textured re
gi
ons
with fewer rin
g
ing
effects. O
nep
ossible expl
ana
tion is that th
e assum
ed p
r
ior
distribution of
edges used i
n
[12],[15],[2
3] does
not hold well whereas ou
r m
e
thodexpl
oits the
informatio
n from the blurre
d image.
Table 2.qu
ant
itative compa
r
at
ive evaluat
ion. SSIM is chosen a
s
perf
o
rma
n
ce mea
s
ure
File Blurt
y
pe
PSNR
Dilips'
s
method
Li's method
Dong’s
methed
Our m
e
thod
house.tif
motion 0.816
0.832
0.7144
0.864
gaussian 0.768
0.787
0.8383
0.829
average
0.807
0.821
0.7984
0.854
boats.tif
motion 0.845
0.836
0.6736
0.895
gaussian 0.821
0.842
0.7840
0.865
average
0.804
0.801
0.7575
0.857
cameraman.tif
motion 0.847
0.848
0.7608
0.849
gaussian 0.695
0.728
0.7169
0.809
average
0.708
0.797
0.6519
0.772
barbar
a.tif
motion 0.837
0.841
0.7282
0.855
gaussian 0.723
0.732
0.6634
0.808
average
0.754
0.801
0.6273
0.789
baboon.tif
motion 0.787
0.772
0.7946
0.855
gaussian 0.747
0.825
0.8060
0.858
average
0.718
0.816
0.7961
0.837
*Best re
sults
are note
d
in b
o
ld.
(a)
(b)
(c
)
(
d
)
(e)
(f)
(g)
(h)
(i)
(j)
(k
)
(l)
(m)
(n)
(o)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Im
age Deblu
r
ring Via an A
daptive Di
ctio
nary Lea
rni
n
g
Strategy (Lei
Li)
861
(p)
(q)
(r)
(s)
(t)
(u)
(v)
(w)
(x)
(y)
Figure 1. Deb
l
ur re
sult
s
(a)
sharp (b) bluraverag (c) Dilip
s's
(d) Dong’
s (e) our m
e
thod
image
method
method
(a)
sha
r
p (b) bl
urga
ussia
n
(c)
Dilips's
(d) Dong’
s (e) ou
r metho
d
image
method
method
(a)
sharp (b) blurmotion
(c)
Dilips's
(d) Dong’
s (e) our method
image
method
method
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 855
– 864
862
(a)
sharp (b) bluraverag (c) D
ilips's
(d) Dong’
s (e) our method
image
method
method
(a)
sharp (b) blurgaussian (c) Dilips'
s
(d) Dong’
s (e) our method
image
method
method
(a)
sharp (b) blurmotion (c) Dilip
s's
(d) Dong’
s (e) our m
e
thod
image
method
method
Figure 2. Col
o
r De
blu
r
re
sults
From Fig
u
re
2 and Ta
ble 3, we
can
see th
at our met
hodi
s amu
c
h more
effectivedebl
urri
ng meth
o
d
. The d
eblu
rrin
g
re
sult
s
with competi
ng meth
ods
are
com
p
a
r
ed in
Figure 2. We
can
se
e that
there
are
m
any noi
se
re
sidu
als
and
artifact
s arou
nd ed
ge
s in
the
deblu
r
red im
age
s in Dilip
s's metho
d
[15
].Our method
leads to the
best visu
al q
uality. It not onl
y
can remove the blurring ef
fects an
d noi
se, but
also
can recon
s
truct more and
sharpe
r ima
ge
edge
s than
Dilips'
s metho
d
s
.
Table 3. PSNR and SSIM result
s of debl
urri
ng imag
es.
File Blurt
y
pe
PSNR SSIM
Dilips'
s
method
Our
method
Dilips'
s
method
Our
method
Parrot.j
pg
motion 22.56
25.76
0.738
0.835
gaussian 23.76
25.89
0.771
0.834
average
22.44
25.34
0.705
0.809
Lena.jpg
motion 25.98
28.61
0.782
0.851
gaussian 25.23
26.93
0.722
0.783
average
24.41
26.74
0.651
0.757
*Best re
sults
are note
d
in b
o
ld.
We have al
so done the
deblu
rri
ng ex
perim
ents o
n
Barba
r
awith
the global l
earn
ed
diction
a
ry a
n
d
the
adaptiv
e di
ctiona
ry,as
sho
w
n
in Fi
g.3. The P
N
S
R
a
nd SSIM
are
25.76
dBa
nd
0.7884
re
spe
c
tively withou
t dictiona
ry a
daption
whil
e
t
he co
unte
r
p
a
rts
are
26.5
0dB and
0.8
08.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Im
age Deblu
r
ring Via an A
daptive Di
ctio
nary Lea
rni
n
g
Strategy (Lei
Li)
863
Debl
urring
wi
thadaptive
di
ctiona
ry
i
s
sh
owe
d
the
sup
e
rio
r
p
e
rfo
r
m
ance b
e
cau
s
eit exploited
the
spe
c
ific info
rmation in the image.
(a)
Initial
(b) Adaptive
(c
) Deblur with
(d) Deblur with
dictiona
ry
dictiona
ry
Figure 3. Co
mpari
s
o
n
of deblurrin
g
re
su
lts on Ba
rb
ara with the glo
bal
learned di
ctiona
ry and the
adaptive di
ctionary. The P
N
SR a
r
e 25.7
6dB and
26.5
0dBre
sp
ectiv
e
ly and the SSI are 0.7884
and 0.80
8
4. Conclusio
n
In this
pap
er,
we
propo
se
d an
onlin
e a
daptiv
e di
ctio
naryst
r
ategy
for ima
ge
de
blurring
algorith
m
ba
sed on the
sp
arserepr
esen
tation. Our st
rategy tran
sf
ers
a glob
al learn
ed di
ctio
nary
to a spe
c
ific i
m
age to
ma
ke the trade
-o
ff betwee
n
computation
a
l
co
st an
d a
c
cura
cy. The
n
e
w
learn
ed
dictio
nary
exploite
d the
specifi
c
inform
at
ion
such
a
s
the
te
xture a
nd
ed
ge info
rmatio
n in
the blu
r
red i
m
age.The
r
efo
r
e, it ca
n
rep
r
e
s
ent
patche
s
better th
an th
e p
r
e
s
pe
cifie
d
di
ctiona
ry a
nd
the glob
al learn
ed di
cti
onary.Th
e a
ppro
a
ch
taken is
ba
se
d on
spa
r
se
and redu
n
dant
rep
r
e
s
entatio
ns i
n
ove
r
co
mplete di
ctio
narie
s l
earne
d. With
only
a fe
w trai
ni
ng
sampl
e
s,
our
method
can
speed u
p
the
learni
ng p
r
o
c
e
ss via
d
o
m
ain ad
aptat
ion. Furthe
rmore, the d
o
m
ain
adaptatio
n al
lows us to
circumve
nt theoverfi
tting
probl
em effe
ctively. Also, the adaptive
diction
a
ryexpl
oited the sp
e
c
ific informati
on in
the ima
ge, whi
c
h ma
ke
s our d
ebl
urri
ng alg
o
rit
h
m
more
effectiv
e. Experime
n
ts h
a
vede
monst
r
ated
that ou
r im
a
ge d
eblu
rrin
g
metho
d
yi
elds
s
u
pe
r
i
or
pe
r
f
or
ma
nc
es
o
v
e
r
o
t
h
e
r
me
th
od
s
.
Ackn
o
w
l
e
dg
ements
This work
was
supported Beijing Higher Ed
ucation Young Elite Teacher P
r
oject(NO.
YETP1949),F
undam
ental
Re
sea
r
ch F
und
s for th
e
Ce
ntral
Universitie
s
u
nder Grant
No.
TD20
13
-4, an
d Nation
al Na
tural Scie
nce Found
ation of
China u
nde
r Grant
No. 30
9011
64.
Referen
ces
[1]
R. F
e
rgus, B. Sing
h, A. Hertzmann, ST
. R
o
w
e
is, W
T
.
F
r
eema
n
. Remo
ving cam
e
ra s
hake from
a
singl
e ph
otogr
aph.
ACM T
r
an
sactions o
n
Graph
ics.
200
6; 25: 787–
79
4.
[2]
A. Levin. Bl
in
d
motion
de
blur
ring
usin
g ima
ge statistics.
A
d
vanc
esin
Ne
u
r
al Infor
m
ati
o
n
Processi
ng
System
s (NIPS
)
. 2006; 2: 4.
[3]
Q. Shan, J. Jia
,
A. Agar
w
a
l
a
. High-
qu
alit
y m
o
tion
de
blurri
n
g
froma si
ngl
e i
m
age.
ACM T
r
ansacti
ons o
n
Graphics.
20
08
; 27(73):10.
[4]
A. Levin,
R. F
e
rgus, F
.
Dur
a
nd, W
T
.
F
r
eeman. Im
ag
e a
nd d
epthfrom
a conv
enti
o
n
a
l
camera
w
i
th
a
code
d ap
erture
.
ACM T
r
ansactions on Grap
hi
cs
. 2007; 26: 7
0
-71.
[5]
J. Jia.
Singl
e
ima
ge
moti
o
n
deb
lurri
ng u
s
ing trans
par
e
n
cy
. Proceedings ofIEEE Conference on
Comp
uterVisi
o
n
and Patter
n
Reco
gniti
on. 2
007: 1–
8.
[6]
N. Joshi,
R. S
z
eliski,
D. Krie
gman.
PSF
estimatio
n
us
ing
s
harp
ed
ge
pred
i
c
tion
. Proc
eed
i
ngs
of
IEEE
Confer
ence
on
Computer Vis
i
on an
d
Pattern
Recog
n
itio
n. 2
008: 1–
8.
[7]
J.-F
. Cai, H. Ji, C. Li
u,
Z
.
Shen.
Bli
n
d
motio
n
d
e
b
l
urrin
g
fro
m
a
singl
e i
m
ag
e
usin
g sp
arse
appr
oxi
m
ati
o
n
.
Proceedings
of IEEE
Conference on Com
puter V
i
sion
andPattern Rec
ognition. 2009:
104
–1
11.
[8]
A. Levin, Y. W
e
iss, F
.
Durand, W
T
. F
r
e
e
man.
Un
ders
t
andi
ng
and
ev
alu
a
ting bli
nd deco
n
vol
u
tio
n
alg
o
rith
ms
. Pr
oceedings
of IEEE Conferenc
e on Computer Vision andP
attern Recognition. 2009:
0
D
D
0
D
D
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 855
– 864
864
196
4–
197
1.
[9]
N. Joshi, C. Zitnick, R.
Szeli
ski, D. Kriegm
an.
Ima
ge de
blurri
ng
and de
noisi
ng usin
g color
pr
iors
.
Proceedings of
IEEE Conference on Computer Visi
on and Pattern Recognit
ion. 2009:
1550–1557.
[10]
D. Krishna
n, R. F
e
rgus. F
a
st image dec
onv
oluti
on usi
ng h
y
p
e
rl
apl
acia
n priors.
Adva
nc
es in Neur
a
l
Information Pr
ocessi
ng Syste
m
s (NIPS)
. 20
09; 22: 1–
9.
[11]
C. Jia, B. Evans.
Patch-bas
ed
imag
e d
e
conv
oluti
on vi
a j
o
i
n
t mode
lin
g of s
p
arse pr
iors
.
Pr
ocee
din
g
s o
f
the 18th IEEE Internati
o
n
a
l
Co
nferenc
e on im
age Proc
essi
n
g
(ICIP). 2011: 681
–6
84.
[12]
H. Li, Y. Z
h
ang
, H. Z
han
g, Y.
Z
hu, J. Sun.
B
l
ind i
m
a
ge de
bl
urring
b
a
sed
o
n
sp
arse prior
of
dicti
onary
pair
.
Proce
edi
ngs of the 21t
h Inter
natio
na
l
Confere
n
ce o
n
Pattern
Rec
ogn
ition (ICPR
)
. 2012: 305
4–
305
7.
[13]
M. Aharon, M. Elad, A. Bruck
s
tein. K-SVD: An al
gor
ithm f
o
r desi
g
n
i
ng
o
v
er compl
e
te d
i
ction
a
ries fo
r
sparse re
prese
n
tation.
IEEE Transactions on Signal Process
i
ng
. 20
06; 54:
431
1–
432
2.
[14]
Z
.
Hu, J.B. Hu
ang, M.H. Yan
g
.
Singl
e i
m
a
g
e
de
blurri
ng w
i
t
h ada
ptive d
i
ction
a
ry le
arni
ng
. Proceedi
ng
s
of the 17th IEEE Internatio
nal
Confer
ence
on
Image Proces
sing
(ICIP)
. 20
10: 116
9–
11
72
.
[15]
K. Dilip, F.
T
a
y,
T
.
and Rob.
Bl
ind d
e
co
nvol
uti
on usi
ng a n
o
r
m
a
l
i
z
e
d
spars
i
ty me
asure
. Pr
ocee
din
g
s of
IEEE Confere
n
c
e on Com
pute
r
Vision a
nd Pa
tte
rn Recog
n
iti
on(CVP
R
). 201
1: 233–
24
0.
[16]
JL. Starck, MK. Nguy
en, F.
Murtag
h. W
a
velets
an
d cur
v
elets for
i
mag
e
d
e
co
nvol
utio
n: a c
o
mbi
n
e
d
appr
oach.
Si
gn
al Process
i
ng
.
200
3; 83: 227
9
– 2283.
[17]
SS. Che
n
, DL
. Don
oho, MA
. Saun
ders. A
t
omic d
e
comp
ositio
nb
y
basi
s
purs
u
it.
SIAM jour
na
l
o
n
scientific co
mp
uting
. 19
98; 20
: 33–61.
[18]
P.
T
s
eng. On accel
e
rated
pr
oximal
gra
d
ie
n
t
methods for c
onve
x
-c
oncav
e
optimizati
on.
SIAM Journa
l
on Optim
i
z
a
tion
. 2008.
[19]
S. Ji, J. Ye.
An
accel
e
rate
d gr
adi
ent
meth
od
for trace n
o
r
m
mi
ni
mi
z
a
t
i
o
n
. Procee
din
g
s of I
n
ternati
o
n
a
l
Confer
ence
on
Machin
e Le
arnin
g
. 200
9: 45
7-46
4.
[20]
Y. Lou, A. L.
Bertozzi, S. S
oatto. Direct sparse d
e
b
l
urri
ng.
Journ
a
l of
Mathe
m
atica
l
Imagi
ng a
n
d
Vision
. 201
1; 39: 1–12.
[21]
J. Mairal, F
.
B
a
ch, J. Ponc
e,
G. Sapiro.
On
line
dicti
o
n
a
ry l
earn
i
ng
for sp
arse co
di
ng
. P
r
ocee
din
g
s of
the 26th An
nu
a
l
Internatio
na
lC
onfere
n
ce
o
n
Machi
ne Le
arn
i
ng. 20
09: 68
9
–69
6.
[22]
Z. W
ang, AC.
Bovik, HR.
Sh
eikh, EP. S
i
m
once
lli. Im
a
g
e
qua
lit
y ass
e
ss
ment: from err
o
r visi
bil
i
t
y
to
structural similarit
y
.
IEEE Transactions on I
m
age Proc
essing
. 200
4; 14: 6
00–
61
2.
[23]
W
e
ishe
ng D
o
n
g
, Lei Z
h
a
ng,
Guangm
ing S
h
i, Xi
n Li. No
nl
o
c
all
y
ce
ntraliz
e
d
sparse r
epre
s
entatio
n for
imag
e restorati
on.
IEEE Trans. on Im
age Processing
. 201
3; 22(4): 16
20-
16
30.
Evaluation Warning : The document was created with Spire.PDF for Python.