TELKOM
NIKA
, Vol.14, No
.4, Dece
mbe
r
2016, pp. 14
24~143
1
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i4.4064
1424
Re
cei
v
ed Ma
y 30, 201
6; Revi
sed O
c
tob
e
r 24, 201
6; Acce
pted No
vem
ber 1
1
, 2016
Digital Image Watermarking Algorithm Using the
Intermediate Frequency
Hong
-an Li*
1
, Zhanli Li
2
, Zhuoming
Du
3
, Qi Wang
4
1,2
School of Mechan
ical En
gi
n
eeri
ng, Xi'
an U
n
iver
sit
y
of Sci
ence a
nd T
e
chnol
og
y, Xi'
an,
Chin
a
1,2
Colleg
e
of Computer Sci
e
n
c
e and T
e
chno
log
y
,
Xi'
an U
n
i
v
ersit
y
of Sci
e
n
c
e and T
e
chno
log
y
,
Xi'
an, Ch
i
n
a
3
School of Co
mputer Eng
i
n
e
e
rin
g
, Jiangs
u
Univ
ers
i
t
y
of
T
e
chn
o
lo
g
y
, Ch
angZ
h
ou, Ch
in
a
4
School of Infor
m
ation Sci
enc
e and T
e
chno
l
o
g
y
, North
w
e
s
t Universit
y
, Xi'
an, Chi
n
a
*Corres
p
o
n
d
e
n
author, emai
l: an6
86
0@1
26.c
o
m
A
b
st
r
a
ct
Digita
l
i
m
ag
e w
a
termarki
ng
is
one
of the
pro
p
o
sed
sol
u
tio
n
s for
copyri
ght pr
otection
of mu
ltimedi
a
data. T
h
is
tec
hni
que
is
bett
e
r tha
n
D
i
gita
l
Sig
natur
es
a
n
d
oth
e
r
meth
o
d
s b
e
caus
e
it
does
not
incr
e
a
se
overh
ead. W
a
t
e
rmarkin
g ad
d
s
the add
ition
a
l req
u
ire
m
ent
of robustness
.
T
o
im
prov
e the rob
u
stness
of
digit
a
l i
m
a
ge
w
a
termarki
ng
meth
od
base
d
on the i
m
age
freque
ncy, thi
s
pap
er ad
opt
s the inter
m
e
d
i
ate
freque
ncy to
e
m
b
e
d
the
w
a
termarkin
g
an
d
pro
pos
es a
n
digit
a
l
i
m
ag
e
w
a
termarki
ng
alg
o
rith
m
bas
e
d
o
n
robust pri
n
cip
a
l co
mp
one
nt
analys
is (RP
C
A) and d
i
scr
ete cosin
e
tra
n
sform (D
CT
). F
i
rstly, the high
freque
ncy part
and th
e low
freque
ncy part of
the i
m
ag
e ar
e
extracted by th
e RPCA a
l
gor
ithm. Bec
aus
e th
e
hig
h
frequ
ency
part has co
mplex statistica
l character
i
stics, this paper pr
o
c
esses the hi
g
h
freque
ncy p
a
rt
w
i
th "8×
8
" D
C
T
meth
od t
o
obta
i
n i
n
ter
m
e
d
i
a
te frequ
ency co
efficie
n
ts and th
en
the w
a
terma
rking
infor
m
ati
on is
embe
dde
d i
n
to
the obta
i
n
ed i
n
termed
i
ate fr
equ
ency co
effi
cients. T
he ex
peri
m
e
n
tal r
e
s
u
lts
show that the
pr
oposed algorithm
leads
to sati
sfactory robustness to the
attacks of im
puls
e noise
and
cropp
ing.
Ke
y
w
ords
:
digit
a
l w
a
ter
m
arkin
g
, rob
u
st princ
i
p
a
l c
o
mp
onent analysis, discrete cosine trans
f
orm
,
interm
ediate
fr
equ
ency p
a
rt
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Digital ima
ge
watermarkin
g
is a
pro
c
e
s
s
of em
be
dding
a pie
c
e of
di
gital co
ntent (image
)
into the
ori
g
in
al ima
ge
and
also
it p
r
ote
c
ts di
gital
co
ntent from
ille
g
a
l ma
nipulati
ons.
The
hid
d
en
watermark
should
be in
separable f
r
o
m
the ho
st image and robu
st
eno
ug
h
to
re
si
st any
manipul
ation
s
whil
e preserving the i
m
age q
ua
lity [1]. Watermarking a
d
d
s
the ad
ditional
requi
rem
ent of robu
stne
ss. An ideal
wate
rm
arki
ng system
woul
d embe
d an amou
n
t
o
f
informatio
n that could
no
t be remove
d or al
tered
without m
a
ki
ng the
cove
r obje
c
t enti
r
ely
unu
sabl
e [2]. So, waterm
arki
ng i
s
mai
n
ly to prev
en
t illegal co
py or cl
aim
s
th
e ownershi
p
of
digital me
dia.
It is
a m
e
a
s
ure
of imm
u
n
i
ty of wate
rm
ark a
gain
s
t a
ttempts to i
m
age
modifi
cat
i
on
and m
anip
u
la
tion like
comp
ressio
n, filteri
ng, rotati
o
n
, collision
atta
cks, resi
zing,
croppin
g
, etc [3
].
In ord
e
r to m
i
nimize
the d
a
mage
of the
origi
nal ima
ge an
d en
ha
nce th
e robu
stne
ss, a
ki
n
d
of
algorith
m
ha
s been p
r
op
ose
d
based
on the imag
e being laye
red. At present, the wavelet
decompo
sitio
n
method is t
he most famil
i
ar meth
o
d
ba
sed o
n
this ki
nd of algorith
m
[4-6].
C
a
o [7
] pr
o
p
o
s
ed
a wa
te
rma
r
k
i
n
g
a
l
go
rith
m b
a
s
e
d on
tr
an
s
f
or
m
do
ma
in
. Th
is alg
o
r
ith
m
has
som
e
ro
bustn
ess cha
r
acte
ri
stics, b
u
t it is
not efficient to re
sist the 1/4 a
nd 1/2 he
avily
cro
ppin
g
.
Dai
[8] propo
se
d a
digital
i
m
age
blind
watermarkin
g
algo
rithm
b
a
se
d o
n
wav
e
let
transfo
rm, wh
ich can effici
ently resi
st the croppin
g
at
tack, but the
robu
stne
ss to
resi
st the noi
se
attack is
not
stron
g
. Zha
n
g
[9] also p
r
o
posed a
digit
a
l imag
e waterma
r
ki
ng te
chnolo
g
y ba
se
d on
wavelet transform, whi
c
h i
s
oper
ated easily, but the ability to re
sist the
noi
se attack is
not
stron
g
. To im
prove the
rob
u
stne
ss of di
gita
l image
waterma
r
ki
ng
method, this
pape
r propo
ses a
new algo
rith
m
with stro
n
ger rob
u
stn
e
s
s
to
resi
st the atta
cks of
impul
se n
o
i
s
e a
nd
crop
ping
based on rob
u
st prin
cip
a
l compon
ent an
alysis
(RP
C
A) and di
screte
cosi
ne tran
sf
orm (DCT).
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Digital Im
age Wate
rm
arkin
g
Algorithm
Using the Interm
ediate Freq
uen
cy (Ho
n
g
-
an Li)
1425
2. Robus
t Principal Componen
t
Anal
y
s
is
The RP
CA method de
com
poses a
n
image into a lo
w ran
k
mat
r
i
x
image and
a spa
r
se
matrix image
[10-12]. If a
matrix could
be decomp
o
sed into a linear co
mbinatio
n of a low ra
nk
matrix and a spa
r
se matrix, as the equation M
=
A
+
E shows, then the two matrixes coul
d be
obtaine
d by usin
g so
me
math method
[13]. Where,
A is a low rank m
a
trix a
nd E is a
sp
arse
matrix. The l
o
w ran
k
mat
r
ix A could
be re
cove
red
from the
co
rru
pted mat
r
i
x
M
=
A
+
E, as
sho
w
n in Fig
u
re 1 [14]. The RPCA me
thod can ov
e
r
co
me the sh
ortco
m
ing of correl
ation d
a
ta
loss wh
en the
traditional P
C
A method i
s
used to redu
ce the hig
h
-di
m
ensi
on.
(a) Data_m
atrix
M
(b) L
o
w
_
ra
nk
_matrix
A
(c
) Spar
se m
a
trix
E
Figure 1. Co
mpone
nts De
comp
ositio
n by RPCA
In the RPCA
algorith
m
, the size of item
is arbitra
r
y, but the support set is sparse and
unkno
wn. Assumin
g
that
a given data matrix
D
∈
R
m×n
is a low rank o
r
approximate low ra
n
k
matrix, the optimal matrix A could b
e
obt
ained
by the
traditional P
C
A when the e
l
ements
of the
sep
a
rate
d sparse m
a
trix
E coul
d be
descri
bed
by an ind
epe
n
dent ide
n
tica
lly distribute
d
Gau
ssi
an di
stribution that i
s
to solve the
following o
p
timization p
r
o
b
l
em [14]:
,
‖
‖
.
.
(1)
The mat
r
ix
is solved
with
the sin
gula
r
value de
com
positio
n meth
od (SV
D
), th
en the
optimal sol
u
tion of the above optimizat
ion pro
b
lem
woul
d be obt
ained. If
is a spa
r
se noi
se
matrix, the traditional PCA
is not applicable.
Then, to recover the
low ran
k
matrix
become
s
a
doubl
e obje
c
t
i
ve optimizati
on pro
b
lem:
min
,
,
‖
‖
.
.
(2)
And then th
e
comp
romi
se f
a
ctor
λ
0
is intro
duced, the d
ouble
obje
c
ti
ve optimizati
on
probl
em is tra
n
sformed into
a single o
b
je
ctive optimiza
t
ion probl
em:
min
,
λ
‖
‖
.
.
(3)
This p
r
oble
m
is an NP problem, so th
e objec
tive functio
n
of the optimizatio
n probl
em
need
s to b
e
relaxed. Due t
o
the
nuclear
norm
of the matrix is a
n
e
n
velope of th
e matrix ra
nk,
the
0n
o
r
m
is equal to
1,1
-norm
unde
r ce
rtain conditio
n
s
. So the problem is relax
ed to the
followin
g
con
v
ex optimization pro
b
lem:
min
,
‖
‖
∗
λ
‖
‖
.
.
(4)
In the actua
l
calcul
ation,
the paper
[12] sugge
st
s
λ
1
,
⁄
. The abo
ve
optimizatio
n probl
em is al
so called
RPCA.
Whe
n
the image is in
the pro
c
e
s
s of
prepro
c
essing, the augme
n
ted
Lagrang
e
multipliers
(A
LM) [1
5] is
adopte
d
to
solve RP
CA
i
n
this pa
pe
r. The
ide
a
of
ALM i
s
a
s
the
following:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1424 – 143
1
1426
At first, the a
ugmente
d
La
gran
ge fun
c
tion is con
s
tru
c
ted:
L
,
,
,
‖
‖
∗
λ
‖
‖
〈
,
〉
‖
‖
2
⁄
(5)
Whe
n
YY
,
, the
alternatin
g m
e
thod is
ado
pted to solv
e the optimi
z
ation p
r
o
b
le
m
min
,
,
,
Y
,
. The matrix
and
are iterated altern
ately using the
exact ALM (EALM)
algorith
m
. If
, then:
min
,
,
,
min
‖
A
‖
∗
⁄
2
(6)
Update the matrix
accordi
ng to the obta
i
ned
:
min
,
,
,
min
‖
E
‖
⁄
2
(7)
Den
o
te
and
respe
c
tively to conve
r
ge to
∗
and
∗
, then the update ex
pre
ssi
on of
the matrix
is
:
∗
∗
(8)
At last update
the param
eter
:
,
‖
∗
∗
‖
‖
‖
⁄
,
otherwise
(9)
Whe
r
e,
1
is a consta
nt and
0
is a sm
all posi
t
ive number.
The RP
CA optimizatio
n
method i
s
a
spe
c
ial
ca
se in many
comm
on opti
m
ization
method
s. According
to thi
s
id
ea, this p
aper ap
plie
s
the RP
CA al
gorithm
into t
he digital
ima
ge
pro
c
e
ssi
ng fi
eld and i
n
the mea
n
time
a digital wa
termarkin
g
al
gorithm
base
d
on RP
CA
is
prop
osed. T
h
i
s
al
go
rithm n
o
t only a
c
hi
eves th
e
high
visual
qu
ality b
u
t also the
strong
ro
bust
n
e
s
s
to efficiently resi
st the attacks of impul
se noise and
croppi
ng.
3. Discre
t
e Cosine Tran
sform
D
i
sc
re
te
Co
s
i
n
e
T
r
an
s
f
or
m (
D
C
T
)
ha
s
ad
va
n
t
age
s of
high comp
re
ssion
ratio, the
lowe
r
bit erro
r rate,
the information con
c
e
n
tra
t
ion and
sm
a
ll amount of cal
c
ulatio
n [16]. At the same
time, the digital waterm
arking b
a
sed o
n
DCT dom
a
i
n is very familiar. The b
a
si
c idea i
s
to
cho
o
se the l
o
w fre
que
ncy
coeffici
ents t
o
emb
ed the
watermarkin
g
into the im
a
ge [17]. Th
e l
o
w
freque
ncy pa
rt of image has mo
re ene
rgy than other parts, so the
waterm
arkin
g
embed
ded
in
the low freq
u
ency pa
rt co
uld obtain hi
gher
robu
stn
e
ss. The low frequen
cy p
a
rt is the sm
ooth
regio
n
of
an
origin
al ima
g
e
, so th
e m
o
dification
of this
part
will l
ead to
re
du
ce the q
uality
of the
image. T
h
e
h
i
gh frequ
ency
pa
rt is the
texture
of an
i
m
age
an
d th
e visu
al effe
ct is
goo
d if th
e
watermarkin
g
wa
s
emb
e
d
ded i
n
thi
s
p
a
rt, but it i
s
sen
s
itive to fi
lter a
nd
com
p
re
ssi
on. So,
we
adopt the
int
e
rme
d
iate fre
quen
cy coefficient
s
to em
bed the
waterma
r
ki
ng. T
w
o-dime
nsio
nal
DCT [1
8] can
be defined a
s
Equation
(1
0):
,
∑
∑
,
(10
)
The inverse o
f
two-dime
nsi
onal DCT can
be defined a
s
Equation
(1
1):
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Digital Im
age Wate
rm
arkin
g
Algorithm
Using the Interm
ediate Freq
uen
cy (Ho
n
g
-
an Li)
1427
,
∑
∑
,
(11
)
Whe
r
e,
can be define
d
as Equation (1
2
)
:
,
,
0
,
,
1
,
2
,⋯,
1
(12
)
The ide
a
of i
m
age p
r
o
c
e
s
sing m
e
thod
based on
DCT tran
sform i
s
that: Take the carrier
image
,
as a
two-dim
e
n
s
i
onal matrix a
nd use
DCT
transfo
rm di
rectly to pro
c
ess the
matrix and e
m
bed the
waterma
r
ki
ng i
n
to the ca
rri
er imag
e [19
]. This pape
r pro
c
e
s
ses t
h
e
texture layer
with
88
DCT m
e
thod, and th
en the wate
rmarking i
s
e
m
bedd
ed into
the obtained
intermediate frequenc
y
c
o
effic
i
ents
.
4. Watermar
king Embed and Extr
act
Algorithm
4.1. Watermarking Embe
d Algorithm
Embed a waterma
r
ki
ng im
age into a carrier ima
ge
with following
s
t
eps
:
Step 1.
At firs
t, us
e the RPCA algorithm to
extract t
he hig
h
freq
u
ency p
a
rt
and the
low freq
uen
cy part
from the origi
nal ca
rrie
r
imag
e
;
Step 2.
Then
, we apply
8
DCT tra
n
sfo
r
to
the high frequ
en
cy part
, and
then u
s
e
keys to ge
nerate
d
pseudo
ra
n
dom sequ
en
ces
and
, which a
r
e in
ord
e
r to record
the watermarking pixel val
ue (1 o
r
0);
Step 3.
Deno
te the matrix
as
,
, which is
the origi
nal
carri
er im
age
after the DCT
transfo
rm. Th
e interme
d
iat
e
frequ
en
cy coeffici
ents o
f
this block a
r
e nam
ed a
s
,
8
1
,
1
,
2
,⋯,
, then we hav
e formula
s
a
s
following:
,
,
σ
,
0
,
,
σ
,
1
(13
)
Whe
r
e,
σ
repre
s
ent
s embe
d
d
ing strength;
Step 4.
Com
b
ine the high
frequen
cy part
with the low freque
n
c
y part
and the
pro
c
e
ss of e
m
beddi
ng is
compl
e
te. Th
e image carrying the wate
rmarking i
s
na
med as
′
.
4.2. Watermarking Extr
a
c
t Algori
t
hm
Extract the waterma
r
ki
ng from
′
. The extractio
n
proce
ss i
s
outline
d
as follo
wing:
Step 1.
Apply DCT to the i
m
age
′
which
carrie
s the waterma
r
ki
ng;
Step 2.
Deno
te the matrix
as
′
,
, which i
s
after the DCT
transform of the image. The
interme
d
iate freque
ncy co
efficients a
r
e
′
,
8
1
1,
2,
⋯
,
. Then, cal
c
ulate the value
of each pixel
according to the followi
ng formul
a:
W
,
0
,
1
(14
)
Whe
r
e,
2
,
,
2
,
(15
)
Step 3.
Com
p
lete the wat
e
rma
r
ki
ng extractio
n
and th
e watermarki
ng is obtai
ne
d.
5. Experimental Re
sults
and An
aly
s
is
In these experime
n
ts, the gray image
Lena with
512
512
pixel size is taken as the
origin
al ca
rri
e
r
image
whi
c
h is an o
pen
sou
r
ce imag
e
and the pixel
size of
watermarking ima
g
e
is
64
6
4
. The origi
n
al carrie
r ima
ge and the
waterma
r
ki
ng i
m
age a
r
e sho
w
n in Figu
re
2.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
9
30
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1424 – 143
1
1428
(a) T
he ori
g
in
al carrie
r ima
g
e
(b) T
he wate
rmarking ima
g
e
Figure 2. Orig
inal ca
rri
er im
age an
d wate
rmarkin
g
ima
g
e
After the wat
e
rma
r
ki
ng im
age i
s
em
be
ded into th
e
origin
al carrie
r imag
e, it is hard to
disting
u
ish Fi
gure
2(a) fro
m
Figu
re 3
(
a
)
visu
ally. Figure
3 sho
w
s
the emb
edde
d ca
rri
er im
a
g
e
and the extra
c
ted waterm
a
r
kin
g
imag
e u
s
ing o
u
r alg
o
rithm.
(a) T
he imag
e embed
ded t
he
w
a
te
rma
r
k
i
ng
(
b
)
Th
e
e
x
tr
ac
te
d
wa
te
r
m
ar
k
i
ng
Figure 3. The
embedd
ed carri
er ima
ge
and the extra
c
ted waterm
a
r
kin
g
imag
e
Use the rob
u
s
tne
ss a
nd in
visibility indexes to
evaluat
e the waterm
arki
ng techno
logy. As
we ca
n se
e from the Figu
re 2(a) a
nd th
e Figure 3
(
a
)
, they are very similar to each othe
r. Th
is
shows that the algorithm
invisib
ility is very strong. This is
a
subjective evaluation to the
algorith
m
. Except the subj
ective
evalua
tion, the obje
c
tive evaluati
on is mo
re i
m
porta
nt, such as
Pak
Sig
nal to
Noise
Ratio (PSNR) and
Normali
z
ed Coefficient (NC) [21]. The co
mpared re
sul
t
s
of extracting
watermarkin
g
image bet
ween ou
r al
go
rithm and th
e pape
r [9] who
s
e meth
o
d
is
based on
wa
velet transfo
rm are sho
w
n
in Table 1.
As we
can
see from Ta
bl
e 1, the PSNR va
lue of o
u
r algo
rithm i
s
31.48
90 an
d the NC
value is 0.99
70 while th
e
PSNR valu
e
of the pa
per
[9] is 3
0
.691
2
and th
e NC
value is 0.76
15.
So, in this aspect ou
r algo
rithm is better
than the pap
e
r
[9].
Table 1. Co
m
parin
g re
sult
s of PSNR and
NC
Algorithm PSNR
NC
Our
algo
rithm
31.4890
0.9970
Paper [4]
30.6912
0.7615
Table 2 sho
w
s that the extracte
d wate
rmarking
ima
g
e
of our algo
rithm is cl
ose
r
to the
origin
al watermarking im
ag
e than the
pa
per [9]. Ta
ble
1 and
Tabl
e
2 mea
n
that
our
algo
rithm
is
more
robu
st
than
the
p
aper [9]. Me
anwhile, in
orde
r to
me
asu
r
e
the
p
e
rform
a
n
c
e
of ou
r
algorith
m
we test
the algo
ri
thms with so
me
atta
cks,
such
a
s
croppi
ng atta
ck an
d
Salt & Pep
p
e
r
noise attack
[22].
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
9
30
Digital Im
age Wate
rm
arkin
g
Algorithm
Using the Interm
ediate Freq
uen
cy (Ho
n
g
-
an Li)
1429
Table 2. Co
m
parin
g re
sult
s of the extracted wate
rma
r
k
The original
w
ate
rmarking image
Our algo
rithm
Paper [4]
5.1. Salt and Pepper Nois
e Attac
k
The salt an
d
Pepper n
o
ise attack i
s
that bl
ack an
d white noi
s
es are add
e
d
into the
image to
ma
ke the im
age f
u
zzy. Wh
ethe
r it could
re
si
st the
salt a
n
d
pep
p
e
r
noi
se
attack o
r
not
that is a
very
importa
nt indi
cator to me
asure th
e p
e
rfo
r
mance of
an
algorith
m
. After the
salt an
d
pepp
er attack (the noi
se
den
sities respectively
are
0.01, 0.03, 0.05), t
he extract re
sult
s are
sho
w
n in Ta
b
l
e 3.
Table 3. The
salt and p
epp
er noi
se attack re
sult
s of this pap
er
Noise density
After attack
Extr
acted watermarking
PSNR
NC
0.01
24.4008
0.9469
0.03
20.3230
0.8880
0.05
18.2125
0.8569
From
Tabl
e
3
we
can
see t
hat the
salt a
nd p
epp
er no
ise
den
sitie
s
resp
ectively a
r
e 0.01,
0.03, 0.05. The algo
rithm
of our pa
per
coul
d ex
tra
c
t the waterma
r
k, while the
salt and p
e
p
p
er
noise den
sity is bigge
r tha
n
0.01 the algorithm
of pa
per [9] coul
d
n
’t extract the watermarki
ng
image. The
r
e
f
ore, these show that
ou
r
algorith
m
ha
s better ability to resi
st the salt and p
e
p
p
er
noise attack than the pa
pe
r [9].
5.2. Croppin
g
Attack
Cro
ppin
g
attack is on
e o
f
the most common
atta
cks, whethe
r i
t
can re
si
st cro
ppin
g
attack effe
ctively or not is an impo
rta
n
t index
to measure rob
u
stne
ss of a waterm
arkin
g
algorith
m
. After crop
ping at
tack, the ex
tract re
sult
s are sho
w
n in T
a
ble 4.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
9
30
TELKOM
NIKA
Vol. 14, No. 4, Dece
mb
er 201
6 : 1424 – 143
1
1430
Table 4. The
results of cro
pping atta
ck
Cropping
Our algo
rithm
Paper [9]
The original imag
e after
cropping 1/4
The e
x
tracted
wa
termark
PSNR 12.1358
11.9392
NC
0.9802
0.6008
From T
able
4, whe
n
the
cro
ppin
g
rat
i
o is 1/4, the
PSNR valu
e of our
alg
o
rithm i
s
12.135
8, whil
e the pap
er [
9
] is 11.93
92
. Moreove
r
,
the visual
effect of ou
r al
gorithm i
s
be
tter
than the pa
per [9]. The NC value of the extr
acted waterm
a
r
kin
g
image
and the orig
inal
watermarkin
g
image
could
rea
c
h to
0.98
02, while th
e
pape
r [9] i
s
o
n
ly 0.600
8. So, ou
r alg
o
rit
h
m
has b
e
tter ro
bustn
ess to resi
st the crop
ping attack th
an the pap
er
[9].
The RP
CA al
gorithm
co
uld
most
ly avoid
the effect th
at ca
u
s
e
d
by
a few of d
e
g
enerate
points a
nd co
uld co
nverg
e
to the right result
s. In
the meantime, it would
keep
high calcul
ation
accuracy an
d
avoid the data loss
in the
process of dimensi
o
nalit
y
redu
ction. So, our algo
rithm
has hi
ghe
r pe
rforma
nce than the pap
er
[9] which i
s
b
a
se
d on wavelet transfo
rm
.
6. Conclusio
n
This p
ape
r e
x
tracts the lo
w freq
uen
cy part
and the spa
r
se pa
rt
of the original
carrie
r image
using the RPCA me
thod.
And then the spa
r
se part
is handled
with 8×8
DCT
method to
ge
t the interm
e
d
iate fre
qen
cy coefficie
n
ts and th
e waterma
r
ki
ng im
age i
s
em
bed
ded
into the obtai
ned intermedi
ate freque
ncy
coefficie
n
ts
of sparse p
a
rt
. The experi
m
ental re
sult
s
sho
w
that th
e algo
rithm
o
f
this pa
per
not only
coul
d obtain
high
visual
qualit
y but also co
uld
efficiently resi
st the attacks of cr
op
ping a
nd the salt an
d pepp
er noi
s
e.
In the future,
we will
do our best to find more effici
ent way
s
to improve the
ability to
resi
st attacks su
ch a
s
geo
metric di
storti
on, stro
nge
r noise and
so
on [23, 24].
Ackn
o
w
l
e
dg
ements
This
work i
s
partially supp
orted by the
Na
tion
al Nat
u
ral Sci
e
n
c
e
Found
ation o
f
China
(Grant No. 6
1402
206, 61
4721
66, U2
1
6114
), Sci
ent
ific Re
sea
r
ch
Program Fu
nded by Sha
a
nxi
Provinci
al Ed
ucatio
n Depa
rtment (P
rog
r
am No.1
6
J
K
1497
) a
nd th
e Xi'an
Unive
r
sity of S
c
ien
c
e
and Te
chn
o
l
ogy Cultivatio
n Found
ation
Project (201
4032
). We d
e
cla
r
e that th
ere i
s
no con
f
lict
of interest
s regarding
the
publi
c
atio
n
of this
arti
cle
and
would
l
i
ke to
than
k
the an
onymo
us
reviewers for
their valuabl
e
comme
nts a
nd su
gge
stio
ns.
Referen
ces
[1]
Gong W
e
i, W
a
ng Qia
n
, Han
T
a
ilin. An Image W
a
termarki
ng Sch
e
me Ba
sed o
n
W
a
vel
e
t
T
r
ansform.
Journ
a
l of Ch
a
ngch
un U
n
iver
sity of Science
and T
e
ch
no
log
y
- Natural Sci
ence
. 20
14; 37
(3): 96-10
5.
[2]
Khan A, Sid
d
i
qa A, Munib
S, et al. A recent
surve
y
of
reversib
le w
a
t
e
rmarki
ng techni
qu
es.
Information sci
ences.
20
14; 2
79: 251-
27
2.
[3]
Anda
lib
i M, C
han
dler
DM. Digita
l Imag
e
W
a
termarkin
g
via Ad
aptiv
e
Log
o T
e
xturi
z
ation.
IEEE
T
r
ansactio
n
s o
n
Ima
ge Proce
ssing
. 20
15; 24
(12): 506
0-5
0
7
3
.
[4]
Prasad K
L
, Ra
o T
C
M, Kanna
n V. A H
y
brid
LW
T
-
D
W
T
Digital Imag
e W
a
termarkin
g
Sch
e
me usi
n
g
LSVR an
d
Q
R
-factorizatio
n
.
Internati
o
n
a
l J
o
urna
l of
App
l
i
ed E
n
g
i
ne
erin
g R
e
searc
h
.
2
016;
11(
6)
:
433
5-43
42.
[5]
Shukl
a
D, T
i
w
a
ri N, Dub
e
y
D. Surve
y
on
Digita
l W
a
term
arkin
g
T
e
chniq
ues.
Internati
o
nal Jo
urn
a
l of
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Digital Im
age Wate
rm
arkin
g
Algorithm
Using the Interm
ediate Freq
uen
cy (Ho
n
g
-
an Li)
1431
Sign
al Process
i
ng, Image Pro
c
essin
g
an
d Pattern Reco
gn
ition.
20
16; 9(1):
239-2
44.
[6]
Ne
yma
n
SN, Pradn
ya
n
a
INP, Sitohan
g B. A ne
w
c
o
p
y
ri
ght protectio
n
for vector map
using F
F
T
-
base
d
w
a
term
arkin
g
.
T
E
LKOMNIKA T
e
leco
mmu
n
icati
on C
o
mputi
ng Elect
r
onics a
nd Co
ntrol.
201
4;
12(2): 36
7-3
7
8
.
[7]
Cao F
u
d
e
. Researc
h
of di
gital
w
a
t
e
rmar
k
i
ng a
l
gor
ithm
based o
n
transform doma
i
n. Shand
on
g
Univers
i
t
y
. 2
0
0
9
.
[8]
Dai T
ao. Study on the d
i
git
a
l image
w
a
t
e
r
m
arki
ng a
l
g
o
rit
h
m base
d
on
w
a
vel
e
t transform. Z
hejian
g
Univers
i
t
y
. 2
0
0
7
.
[9]
Xim
i
ng Z
h
an
g. Digita
l
ima
ge
watermarki
ng te
chno
log
y
r
e
se
a
r
ch bas
ed o
n
w
a
vel
e
t transfo
rm
. Science
& T
e
chnol
ogy
Visio
n
. 201
4; 18: 112-1
14.
[10]
W
an T
,
Z
hu C,
Qin Z
.
Multi-focus imag
e fusi
o
n
base
d
on ro
b
u
st princi
pal co
mpon
ent an
al
ysis.
Pattern
Reco
gniti
on L
e
tters
. 2013; 34(
9): 1001-
10
08.
[11]
W
an T
,
Liao R,
Qin Z
.
A robu
st feature se
le
ction
a
ppr
oach
usin
g l
o
w
rank
matric
es for br
east tu
mor
s
in ultras
onic i
m
ages
. 18t
h IEEE Internatio
nal
Confer
ence
on
Image Proces
sing. 20
11: 16
45-1
648.
[12]
Can
des EJ, Ro
mberg J, T
ao T
.
Robust unc
ertaint
y
pri
n
ci
pl
es: exact sign
a
l
reconstructi
on
from high
l
y
incom
p
lete fre
que
nc
y
inform
ation.
IEEE Inform
Theory
. 20
06; 52: 48
9-50
9.
[13]
Hua
ng
Xi
aos
h
eng, H
uan
g P
i
ng,
Ca
o Yi
qi
n. A block-sp
arse RPCA
al
gorithm for m
o
vin
g
ob
ject
detectio
n
bas
e
d
on PCP.
Jour
nal of East Chi
na Jia
o
ton
g
Un
iversity
. 201
3; 30(5): 30-
31.
[14]
Shi J, Z
h
eng
X, W
e
i
Z
.
Sur
v
e
y
o
n
a
l
g
o
rith
ms of lo
w
-
ra
nk
matri
x
recov
e
r
y
.
Ap
pl
icatio
n Rese
arch of
Co
mp
uters
. 20
13; 30(6): 1
601
-160
5.
[15]
Lan
g Shi
n
a
n
, Z
hao Bo. A n
e
w
i
m
ag
e pr
oce
ssing m
e
t
hod f
o
r discrim
inati
n
g inter
nal
la
yer
s
from radi
o
echo so
un
din
g
data of ice s
heets via
a co
mbin
ed
ro
bust
princi
pal c
o
m
pon
ent an
al
ysi
s
and total
variati
on ap
pro
a
ch. 201
4: 839
-840.
[1
6
]
H
q
W, R
e
nh
ou L
.
An
en
cry
p
te
d
bi
na
ry
i
m
a
g
e
di
gital
w
a
ter
m
arking
al
gorit
hm bas
ed
on
DCT
.
Sm
a
ll
Micro Co
mp
ute
r
System
. 200
3
;
24(1): 103-1
0
7
.
[17]
Bo Ji
nag,
F
e
n
g
L
i
. Co
lor
ima
ge
dig
i
tal
w
a
te
rmarking
a
l
gor
i
t
hm bas
ed
on
discrete
w
a
v
e
l
e
t transfor
m
and HVS.
Micr
oco
m
p
u
ter & Its Applic
ations
. 200
4:
55-5
7
.
[18]
Li
Liu,
Yaj
i
an
Z
hou, Bi
n Z
h
a
ng. D
i
gita
l W
a
t
e
rmarkin
g M
e
thod
for QR
Co
de Ima
ges
Ba
sed
on
DCT
and SVD.
Infra
r
ed an
d Las
er Engi
neer
in
g
. 2013; (12): 3
05-
312.
[19]
Ben
w
e
i
Z
h
u, W
unan
W
an. Rese
arch on
D
i
gita
l
W
a
termar
k
ing
Alg
o
rithm
Based
o
n
L
SB
for QR C
o
d
e
.
Journ
a
l of Ch
e
ngd
u Univ
ersit
y
of Informatio
n T
e
chno
lo
gy
. 201
2; 27(6): 54
2-54
6.
[20]
S
y
e
d
Al
i Kh
a
y
a
m
.
T
he Discret
e Cos
i
ne
T
r
ansform (DCT
):
T
h
eor
y a
nd A
p
p
l
i
c
ation.
Information Theory
and C
odi
ng
. 20
03; 10(5): 4-7.
[21]
Jie W
u
. A Dig
ital W
a
termarki
n
g Alg
o
rithm Ba
se on D
C
T
and Blocke
d Imag
e.
Artificial Inte
ll
i
g
en
ce
and
Reco
gniti
on T
e
chno
logy
. 2
007
: 1424-1
4
2
7
.
[22]
S Rungr
au
ngsi
l
p, M Ketcham, Virutt Kosolvij
ak.
Data hid
i
n
g
meth
od for QR code b
a
se
d on w
a
termar
k
by compar
e DCT
w
i
th DFT
do
ma
in
. Intern
ation
a
l Co
nfer
ence o
n
Com
puter an
d Co
mmunicati
o
n
T
e
chnolog
ies. 201
2:
26-2
7
.
[23]
Che
n
CH, T
ang YL, Hs
ieh
W
S
, et al. Image T
a
mper
Detectio
n an
d Rec
o
ver
y
b
y
Intersecti
n
g
Sign
atures.
T
E
LKOMNIKA T
e
leco
mmunic
a
tion C
o
mputi
ng Electro
n
ics
and Co
ntrol.
2014; 12(
4):
112
3-11
31.
[24]
Xu
Y
X
Y, Zha
ng Q, Zho
u
C. A Nov
e
l
DWT-
Based
W
a
termarkin
g
for Imag
e
w
i
th T
he SIFT
.
T
E
LKOMNIKA T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol.
20
13; 11(1): 1
91-
198.
Evaluation Warning : The document was created with Spire.PDF for Python.