A Robust Blind Image Watermarking Method Using Local Maximum Amplitude Wavelet Coefficient Quantization

Abstract

In this paper, an innovative blind watermarking algorithm has been proposed for imagery applications. This algorithm has used the coefficients of the discrete wavelet transform of the host image in the form of super trees to embed the predefined binary watermark in the host image. In this scheme, a pseudo random sequence is generated to determine the exact wavelet super trees used for embedding procedure. In the next step, after choosing the maximum and second maximum amplitude coefficients of each super tree, the distance vector between two coefficients is computed. For embedding bit zero of the specified watermark, the values of the distance vector elements are decreased, while for embedding bit 1, those values will be increased based on the proposed formulas. The experimental results show that the proposed algorithm has significant robustness against image processing attacks, especially JPEG compression and also the PSNR value for the watermarked images generated by the proposed method is more than 42 dB.

Full text

Page 1

Advances in Electrical and Computer Engineering Volume 10, Number 3, 2010

A Robust Blind Image Watermarking Method
Using Local Maximum Amplitude Wavelet
Coefficient Quantization
Mehdi HAJIZADEH, Mohammad Sadegh HELFROUSH, Mohammad Javad DEHGHANI,
Ashkan TASHK
Department of Electrical Engineering
Shiraz University of Technology, Modarres Bolvd. P.O. Box 71555-313, Shiraz, Iran
Email: {m.hajizadeh, ms_helfroush, dehghani, a.tashk}@sutech.ac.ir

Abstract— In this paper, an innovative blind watermarking
algorithm has been proposed for imagery applications. This
algorithm has used the coefficients of the discrete wavelet
transform of the host image in the form of super trees to embed
the predefined binary watermark in the host image. In this
scheme, a pseudo random sequence is generated to determine
the exact wavelet super trees used for embedding procedure. In
the next step, after choosing the maximum and second
maximum amplitude coefficients of each super tree, the
distance vector between two coefficients is computed. For
embedding bit zero of the specified watermark, the values of
the distance vector elements are decreased, while for
embedding bit 1, those values will be increased based on the
proposed formulas. The experimental results show that the
proposed algorithm has significant robustness against image
processing attacks, especially JPEG compression and also the
PSNR value for the watermarked images generated by the
proposed method is more than 42 dB.

Index Terms— Blind, copyright
watermarking, wavelet trees
protection, image
I. INTRODUCTION
Nowadays, the burst of information due to the growth of
computer networks and multimedia technologies with no
exact supervision over them leads to the emersion of
copyright breaks and breaches. To defend against this
problem, some solutions have been proposed. One of such
proposed schemes is digital watermarking [1-3]. A
comprehensive and concise
watermarking is embedding a specific code or pattern
named watermark comprising copyright information into a
content based on a known algorithm [4]. The reverse
procedure is feasible using the algorithm and the embedded
watermark can be extractable completely.
The most important problem that digital watermarking used
for copyright applications is its robustness against
intentional and unintentional attacks. There have been many
watermarking strategies proposed for complying the
security, robustness, perceptual invisibility and protective
secrecy of watermarking systems [5]; however, all of them
suffer from some deficiencies like quality deterioration
along with accretion of payload. These problems are
aggravated when the original watermark is inaccessible or
the watermarking system has a blind scheme. If the
watermarking system is used for copyright protection of
image-type information, then the effect of this problem will
be more tangible.
definition for digital
The watermark bits may be inserted into the host image
either in spatial or in the transform domain. It is shown that
hiding information in the transform domain will lead to a
more robust watermarking system against most of the
attacks [6]. Among these methods, methods based on
discrete wavelet transform (DWT) are of great renown. This
is due to specific features of space-frequency localization,
multi-resolution display, linear computational complexity
and the excellent modeling related to Human Vision System
(HVS) [7].
Barni et al. [8], have introduced a blind watermarking
algorithm based on modifying the DWT coefficients of the
image. In this method, a binary pseudo random sequence is
embedded into the three largest components of the
frequency sub-bands of the host image using pixel-wise
masking. Distorting the largest frequency components can
easily eliminate the embedded watermark without seriously
destructing the watermarked image. In [9], the watermark is
inserted into the three mid-frequency sub-bands LH2, HL2
and HH2. From these three sub-bands, the minimum energy
sub-band is selected for embedding watermark. For
extraction of the watermark from watermarked image, the
wavelet transform filters can be used as the security keys. In
this method, exploiting a scaling parameter provides a trade-
off between the energy of the embedding watermark and the
image distortion. In the blind watermarking method
suggested by Huang et al. [10], the cover image is separated
into m blocks of the size n×n. Then, each block is
decomposed into wavelet domain and each bit of watermark
was embedded in the middle and low frequency sub-bands
of a wavelet coefficient of that block. For increasing the
robustness of this watermarking method against the image
processing attacks, multi-energy levels, Arnold transform
and multi-resolution wavelet transform are exerted.
Wang et al. [11], introduced a new blind watermarking
technique in which a tree quantization of the host image's
wavelet coefficients is used for embedding the watermark. A
special wavelet coefficient arrangement called super trees is
employed for quantizing. For embedding each bit of the
watermark, two super trees have been employed. In this
method, various frequency bands of wavelet transform have
been used for embedding the watermark bits so that
information energy is distributed in the whole spatial
domain. Accordingly, the watermarking scheme can be
robust and secure against intentional watermarking attacks.
Lien et al. [12] improved Wang et al. [11] method. In this
method, a blind wavelet-based watermarking method using
the HVS and the quantization of wavelet super trees is

96
1582-7445 © 2010 AECE
Digital Object Identifier 10.4316/AECE.2010.03016
[Downloaded from www.aece.ro on Tuesday, September 07, 2010 at 05:53:34 (UTC) by 85.185.144.179. Restrictions apply.]

Page 2

Advances in Electrical and Computer Engineering Volume 10, Number 3, 2010

presented. Each watermark bit is embedded using two
wavelet super trees. The watermark bits are embedded
repeatedly into bands in the three different orientations. The
method improves the performance of robustness against the
common image processing. Byun et al. [13] proposed a
watermarking method which employs quantization and
statistical characteristics of wavelet transform. In this
scheme, wavelet coefficients from the middle frequency
band are divided into 3×3 blocks. In case the standard
deviation of the coefficients of each 3×3 block is small, the
middle point of that block is quantized into the average
value of eight neighboring coefficients within a block. Then
it uses an average of coefficients in a block; otherwise if the
standard deviation of that block is larger than a threshold, it
is quantized using a quantization method. This method is
quite effective against JPEG compression with a quality
factor up to 30.
Yin et al. [14] proposed a wavelet-based blind watermarking
approach in which each bit of noisy watermark is used to
replace the third or fourth bit of the integer part of the
coefficient absolute values. The beneficiary of this method
is that it could overcome JPEG2000 compression attack, to a
great extent. Another watermarking algorithm based on
wavelet tree quantization has been proposed in [15]. The
wavelet trees have been structured from the coefficients of
the second and third frequency levels of wavelet transform.
Thus, each tree has five coefficients. A super tree consists of
four wavelet trees. Each super tree is divided into five 2×2
sub-blocks. Each sub-block has been separated into two
elements, namely “upper” and “lower”. The watermark
embedding algorithm is based on comparing the average
“upper” two elements and the average “lower” two
elements. The experimental results show that the method
performs well in JPEG compression, filtering and multiple
watermark attack.
A blind watermarking algorithm [16] has been proposed for
color images based on DWT and DFT. In this method, host
image is a color image. The color image is mapped from
RGB domain to YIQ domain. The luminance component Y
is decomposed into wavelet transform. The neighbor
coefficients are cleaved into 4×4 blocks and then, DFT is
applied to them. The security of this algorithm is improved
by using four keys.
Min-Jen Tsai et al. [17] proposed a novel wavelet tree
watermarking using double wavelet tree energy modulation.
The wavelet coefficients of host image are grouped into
super trees. In this scheme, four super trees and their
energies have been chosen for embedding one watermark
bit. The four super trees are converted into two groups and
the two groups within the energy difference are calculated.
The energy of each super tree is the sum of absolute values
of all the wavelet coefficients. Recently, a blind
watermarking method based on maximum wavelet
coefficients quantization was presented [18]. In this method,
a variable number of the wavelet transform coefficients are
formed into a block. The blocks are selected randomly from
different frequency sub-bands. For embedding watermark
bits, the local maximum coefficients are quantized. The
local maximum wavelet coefficient is always the greatest
one in a block. Therefore, this method offers high resistance
against attacks.
The methods proposed in [11, 12, 15] are based on
quantizing the significant coefficients in certain locations. In
other words, it is necessary that the same location to be
considered for significant coefficients in the decoding
process as those in the embedding one. Hence these methods
are not suitable for blind watermarking as the resulting
watermarked images are not robust against low pass filtering
and JPEG compression attacks.
In this paper, we quantize the significant coefficients such
that their locations should not necessarily be the same as in
the embedding phase with that of watermark extraction
phase. Quantizing of the largest local coefficients are based
the distance vector. The distance vector is obtained in such a
way that its value will be in the interval between the first
and second maximum coefficients of each super tree due to
their amplitude. Each bit of watermark is embedded
according to a predefined change in the distance vector
content. An adaptive threshold is employed for extracting
the watermark from watermarked image as the extracted
distance vector is compared with this threshold, and a bit
one or bit zero will be extracted due to a specific formula.
Experimental results show that the proposed algorithm has
higher PSNR than the other competitive methods and
watermarked images do not suffer from obvious visual
distortion.
The outline of the paper is as follows: Section II introduces
the proposed method including
preprocessing, blind embedding and extraction phases. The
detailed experimental results are brought in section III.
Finally, we will conclude the paper in Section IV.
the Super tree,
II. THE PROPOSED WATERMARKING METHOD
In this method, the host image named f has a size of M×M
pixels and the binary watermark named W is assumed to be
Nw bits. Every bit of W is whether 0 or 1. At first, for
embedding bits of watermark in host image, the wavelet
coefficients have been grouped into a sequence of super
trees using blocks of different sizes. These super trees have
been chosen in a pseudo random manner for embedding bits
of watermark in the host image. After choosing the
maximum coefficient in amplitude of each super tree, it’s
the time for computing the distance vector between this
coefficient and the second maximum coefficient, in
amplitude, by which the distance vector of each super tree
would be determined.
For embedding bit zero of watermark, a predefined value
is subtracted from the distance vector and for bit one, a
predefined value increased from the distance vector. For
extracting the watermark from watermarked image, an
adaptive threshold is exerted, so that if the extracted
distance vector is greater than this threshold, then it refers to
a bit one; otherwise, bit zero will be extracted. The
procedures of the proposed method described in the
following sections.
A. Super tree
Applying 3 level wavelet transformations to the host image
produces 10 different frequency sub-bands, which are
depicted in Figure 1. By using wavelet tree, it’s possible to
describe the similar spatial coefficients in different
frequency sub-bands. There would be many hierarchical
trees in the wavelet coefficients. In this paper, for modeling
wavelet coefficient in the form of tree structures, the second
and third wavelet level coefficients are used which are
named: (LH2, HL2, HH2) and (LH3, HL3, HH3),
respectively.
97
[Downloaded from www.aece.ro on Tuesday, September 07, 2010 at 05:53:34 (UTC) by 85.185.144.179. Restrictions apply.]

Page 3

Advances in Electrical and Computer Engineering Volume 10, Number 3, 2010

To rearrange trees into a super tree form, the m×n blocks of
third level of wavelet transformed image, an in companion
with 2m×2n blocks in the second level of the same wavelet
transformed image are employed. The watermarking
capacity or in fact, the number of super trees can be
increased by changing the dimensions of blocks. In the
proposed algorithm, only one bit of watermark can be
embedded in each super tree. For instance, if the size of the
blocks in the third level of is 2×2, then the maximum
number of embedded watermark bits in the host image with
size of M×M pixels will be
[15]. However, if the size of blocks in the same level
becomes 4×4, then the number of embedded watermark bits
will be .
)2/2/(3
MM
××
, similar to
)2/2/(3
44
MM
××
55

Figure 1. Tree structure of wavelet coefficients that correspond to the
same spatial area

The coefficients of higher levels in wavelet transformation
have more influence in comparison to those of the lower
levels. Accordingly, the usage of coefficients of LL3 band
for embedding the watermark is not suitable, as the low
frequency bands include higher significant information.
Therefore, a very small change in their content will lead to
large damage in the host image quality. Also, embedding
watermark bits in the coefficients of HH3 bands is not
suitable; because these coefficients will be also damaged
easily; for example, using lossy compression algorithm. The
LH3 frequency band has more priority over HL3 band;
however the same concept is right for the same frequency
bands in the second level of wavelet transformed images.
Consequently, the important and desirable target is that the
bits of watermark, first must be embedded at the frequency
sub-bands LH3 and LH2 and then HL3 and HL2 and finally
can be embedded in HH3 and HH2 bands [18].

B. Preprocessing
One of the most important features of watermarking
algorithms is their security. Accordingly, for increasing the
security of a typical watermarking algorithm, the size of
blocks in the third level of wavelet transformed image is
chosen 2×2. Suppose that the target is embedding 512 bits
in a host image with size 512×512 and only the coefficient
of mid-frequency bands LH3 and LH2 are to be used. Thus,
it’s possible to embed up to 1024 watermark bits in the host
image. By using a sequence of pseudo random numbers, 512
number of super tree among the whole number of super tree;
i.e. 1024, are selected in which watermark bits are going to
be embedded. By treating in this way, the total number of
coefficients in a super tree will be equal to 20 which
comprise 4 coefficients belong to frequency sub-band LH3
and 16 ones belong to frequency sub-band LH2. Before
embedding watermark bits in the host image, the value of E
and di can be computed according to the equations (1) and
(2) as following:

N
Nid
N
1
=

isabsmabsd
,...,2 , 1
=−=


Where E is mean distance vector for all super trees; di is the
amplitude of distance vector for ith super tree; mabsi and
sabsi are the first and second maximum coefficients due to
the amplitudes of ith wavelet super tree, respectively.
Moreover, Nw is the number of watermark bits. Coefficient
mabsi has two situations: the maximum positive coefficient
or the minimum negative one. The coefficients Δ11, Δ12, Δ21
and Δ22 have positive constant values which has a
significant influence on the PSNR of the proposed
algorithm. For instance, by increasing these coefficients, it is
possible to embed watermark bits in the host image with
higher energy; however, this will lead to PSNR reduction.
Before embedding watermark bits, if coefficients Δ11, Δ12,
Δ21 and Δ22 are greater than E then the new Δ21 and Δ22 are
calculated as in following formula:

(1)
W
i
i
W
E
W
,...,2 , 1
=
1
=
∑

(2)
Wiii
N
HL1
LH1
HH1
HL2
LH2
HH2
HH3
LH3
HL3
LL3
(3)
⎩
⎨
⎧
>Δ=Δ−Δ+Δ=Δ
>Δ=Δ−Δ+Δ=Δ
Eif Eand E
Eif Eand E
new
22
new
12
new
21
new
11
222212
2121 11


C. The proposed algorithm for watermark embedding
In this stage, for embedding 0 and 1 bit in each super tree, it
is essential that at the first step the maximum coefficient
according to its amplitude be determined. For embedding bit
1 in a wavelet super tree, mabsi is quantized due to the
following proposed formula:

⎧
×+
mabs mabs
ii

On the other hand, for embedding bit 0 in a wavelet super
tree, the following proposed formula is used:

⎧
×−
mabsmabs
ii

Where sgn(.) function operates as in:
⎧
−
1
mabsif

Where
amplitude coefficient of super tree. In this algorithm, the
value of d
until the suitable value for di is achieved. The suitable value
of di for bit 1 equals with E+Δ12 and its appropriate value for
bit 0 is equal with E-Δ22; according to these notes, it is
essential that di which is the amplitude of distance vector for
ith super tree, be calculate
For embedding bit 1, if di is smaller than E+Δ11, then di has
()
) 4 (
)()()sgn(
)(
11 12
11
⎩
⎨
Δ+≤−Δ+
Δ+>
=
EddE
Edmabs
mabs
ii
ii
new
i
()
) 5 (
)()()sgn(
)(
2122
21
⎩
⎨
Δ−≥−Δ+
Δ−<
=
EdEd
Edmabs
mabs
ii
ii
new
i

(6)
⎩
⎨
<
≥
=
0
01
)sgn(
i
i
i
mabsif
mabs

w is the new value for the maximum
ne
i
mabs
d.
i is increased or decreased by quantizing mabsi

98
[Downloaded from www.aece.ro on Tuesday, September 07, 2010 at 05:53:34 (UTC) by 85.185.144.179. Restrictions apply.]

Page 4

Advances in Electrical and Computer Engineering Volume 10, Number 3, 2010

to be increased due to equation (4) till its value approaches
the E+Δ12 value. For embedding bit 0, if di is greater than E-
Δ21, then it must be decreased according to equation (5) until
its value approaches to E-Δ22. Consequently, it is obvious
that Δ11≤Δ12 and Δ21≤Δ22 and they must be observed in the
preprocessing step to be objected.
D. Proposed algorithm for watermark extraction
In the proposed method, there is no need of original image
for extracting the watermark bits. In the embedding
algorithm for one bits embedding, the value of di has been
increased and for the zero bits embedding value of di has
been decreased. Now, in the extraction algorithm, it's
necessary to extract and choose a threshold (T), so that for
the di larger than T, a bit 1 is detected and for the di smaller
than T, a bit 0 is extracted. Accordingly, the value of T can
be calculated as in the following formulas:

1
j
nj
W
nN
−
+=

⎥
⎢⎣

Where
{
211
,
sabsmabssabsmabs
−−=φ
and
{
...,,...,,
2121
W
N
ϕϕϕϕϕφ<<=
⎣⎦ . is the symbol of floor function and α determines the
number of distance vectors essential for finding threshold T.
If α equals zero, then threshold T is to be computed due to
all of distance vectors.
Note that as the value of α increases, the number of distance
vectors used for computing T parameter decreases. The
reason for such process is because of this fact that those
distance vectors which have the largest and smallest
amplitudes at both ends treat as noise; so, it's essential to get
omitted.
After computing the threshold T, it is the time to extract the
watermark image as in equation (9),
(
⎩
otherwise
0
(7)
''
1
;
2
jj
n
ϕ
N
j
sabsmabsT
W
−==
∑
−
ϕ

(8)
10,
2
<≤
⎥⎦
⎢ ×
α
=
α
W
N
n

}
'
N
'
N
'
2
'''
,...,
WW
sabsmabs
−

}
;
W
N
ϕ<

(9)
)
⎨
⎧
≥−
=
Tsabsmabs
bitwatermark
ii
1
''

III. EXPERIMENTAL RESULTS
The one most common and standard filter for DWT over
images is Le Gall 5/3 filter. Le Gall’s 5/3 filter has rational
coefficients and has been used in image processing as a
standard for image compression [20]. Therefore, in this
paper is used the discrete wavelet transform with Le Gall
5/3 filter. To evaluate the efficiency of the proposed
watermarking method, a significant number of natural gray
scale images with size 512×512 (M=512) and a 32×16 (Nw
=512) watermark have been employed. For simulation, the
value of α and the size of blocks in the third level of wavelet
transform are considered 0.5 and 2×2, respectively. There
are also some other supposed values for Δij as the following:
Δ11 = 47, Δ12 = 53, Δ21 = 18 and Δ22 = 25.
Moreover, the frequency sub-bands LH3 and LH2 are used
for embedding the watermark bits of watermark image. The
peak signal-to-noise ratio (PSNR) criterion is used
comprehensively for comparing the quality of original
image f with size of M×M pixels and the watermarked one
(fw). For PSNR computation of f and fw, the following
formulas are used:

()()()
11
M

255
(log20
10
MSE

Another criterion which is used in this paper is normalized
correlation coefficient (NC). The NC is used as a measure
between embedded watermark (W) and the extracted
watermark (We) which is calculated as:
ww
jiWejiW
NC
×
Where wh and ww are length and width of watermark,
respectively. In addition, W(i,j) and We(i,j) indicates the bits
of the original watermark and the extracted watermark,
respectively. In this paper, if the value of watermark equals
1, then a value 1 is placed in W(i,j), otherwise a -1 value is
placed in W(i,j). We(i,j) takes -1 or 1 values in the same way
as W(i,j) does. Therefore, the value of W(i,j)×We(i,j) always
equals either 1 or -1.

(10)
2
2
,,
yxfyxf
MSE
M
x
M
y
w
∑∑
==
−
=

(11)
)
PSNR=

(12)
wh
ij
ww
hw
×
=∑∑
==
11
), (), (







Figure 2. The host images of size 512×512

For distinguishing the existence of the watermark in a
watermarked image, the value of the computed NC is
compared with a thresholdρ; if
watermark bits; otherwise, it does not. The probability of
false positive error can be computed by [21],
ρ≥
NC
, the image contains
P fP
N
(13)
A
E
AN
E
N
NA
W
f
PP
A
P
W
W
W
P
) 1 (
)
2
1
(
−
⎟⎟
⎠
⎞
⎜⎜
⎝
⎛
=
−
×
+
=∑
ρ

Where
reasonable to assume
is the probability of W(i,j)≠We(i,j) and it is
5 . 0
=
, with choosing properρ ,
would be calculated, accordingly. For example, let N
ρ = 0.23 and
5 . 0
=
; the value of
from equation (13) and leads to value 1.03×10

E P
E P

P fP
w=512,
E P
can be computed
P fP
-7.
99
[Downloaded from www.aece.ro on Tuesday, September 07, 2010 at 05:53:34 (UTC) by 85.185.144.179. Restrictions apply.]

Page 5

Advances in Electrical and Computer Engineering Volume 10, Number 3, 2010


(a)

(b)

(c)

(d)
Figure 3. (a) The host image of Lena of size 512×512, (b) The original
binary watermark of size 32×16, (c) Watermarked Lena image with
PSNR=43.17 dB, (d) The extracted watermark with NC=1

In this paper, the proposed method has been compared with
the methods proposed by Wang et al. [11], Li et al. [12] and
Lin et al. [18] with the predefined value ρ = 0.23.
The proposed algorithm is implemented on sixteen diverse
gray scale images such as Lena, Pepper and Goldhill which
belong to the database introduced in [22] and are shown in
Figure 2. The original Lena image and its watermarked one
including the binary watermark "ITCI" are depicted. The
used binary watermark has 256 zero bits and 256 one bits.
The results related to the achieved PSNR and the average of
distance vector E for each image, are depicted in Table I.
Please note that the computed PSNR for all examined
images is greater than 42dB. Moreover, the related NC of
extracted watermark image equals to 1. The watermarked
versions of Lena, Pepper and Goldhill images had been
exposed to many different attacks such as JPEG
compression, median filtering, histogram equalization,
average filtering, Gaussian filtering, scaling, cropping and
rotation. JPEG is one of the most popular formats of images
found in internet or taken by digital cameras. The quality
factor (QF) is a number in the interval [0–100] which states
the compression percentage. By QF ratio reduction, the
compression ratio increases yet in turn the quality of image
is reduced. In this paper, the corresponding watermarked
images had been compressed with various QF values.
Even under the high compression ratios, the quality of
watermarked images is still satisfactory such that the
extracted watermarks and the original watermarks are of
high correlation. The proposed method is able to detect
watermarks for quality factors greater than 10. The results of
this attack are shown in Table II.
Median filtering, as a nonlinear operation, simultaneously
reduces noise and preserves edges. Also, Average filtering
provides some smoothing over the image details. To
simulate filtering-type attacks, the watermarked image is
filtered by a median and an average filter. The histogram
equalization is used to improve the contrast of images by
altering the values in an intensity image so that the output
image histogram almost seems uniform. Table III, comprises
the resulting NC values for applying median and average
filters of sizes like 3×3, 5×5 and 7×7 and also histogram
equalization. In Table IV, the NC results of Gaussian noise
attacks with zero mean and variances like 0.001, 0.002 and
0.003 are shown. For applying scaling attacks to the
watermarked images, at first the watermarked images with
size 512×512 have gotten rescaled to the size 256×256.
Then, they got rescaled again to their original size. For
applying the cropping attack, the quarter part of the
watermarked image would be substituted with zeros or 255
contents. In rotation attack, the watermarked image has been
rotated with various angles. Rotating with small angles, the
rotated watermarked image does not have visual distortion
very much but the extracted watermark will severely be
damaged.

Table I. The PSNRs of watermarked images and mean of distance
vectors (E)
Crown Pepper GoldhillLena ImagesTank Truck Elaine House
44.19 43.89 42.49 42.52 43.12 42.94 43.82 43.17
PSNR
18.21 17.21 25.81 30.13 16.70 27.50 21.74 28.05
E
Trucks2Trucks1
Living Room
Airplane Bridge Pirate Tiffany Baboon Images
44.09 44.17 42.22 43.00 42.28 42.38 43.19 43.26
PSNR
22.92 18.23 27.25 15.28 41.68 43.39 16.91 35.01
E

Table II. Normalized correlation coefficients (NC) upon attacks of
JPEG compression with the quality factors (QF) 10, 15, 20 25, 30, 35,
40, 50, 60, 70, 80, 90. (a) Lena, (b) Goldhill, (c) Peppers
15 10 QF 35 30 25 20
0.96 0.91 0.89 0.77 0.66 0.49
(a) Lena/NC
0.94 0.92 0.84 0.79 0.69 0.46
(b) Goldhill/NC
0.94 0.92 0.87 0.76 0.64 0.45
(c) Peppers/NC
90 80 70 60 50 40 QF
1 1 1 1 0.99 0.99
(a) Lena/NC
1 1 1 0.99 0.99 0.97
(b) Goldhill/NC
1 1 1 1 0.98 0.95
(c) Peppers/NC

Table III. Normalized correlation coefficients (NC) upon attacks of
Median filter, Histogram equalization and Average filter. (a) Lena, (b)
Goldhill, (c) Peppers
Median filter
Attacks
Average filter
Histogram
(7×7)
(5×5)
(3×3)
equalization
(7×7)
(5×5)
(3×3)
Images
0.44 0.60 0.88 0.89
0.53 0.67 0.86
(a) Lena/NC
0.39 0.56 0.83 0.83
0.39 0.55 0.80
(b) Goldhill/NC
0.41 0.61 0.86 0.88
0.42 0.64 0.86
(c) Peppers/NC

Table IV. Normalized correlation coefficients (NC) upon attacks of
Gaussian noise, Scaling 256×256, Cropping 1/4 and Rotation. (a) Lena,
(b) Goldhill, (c) Peppers
Scaling
Gaussian noise
Attacks
Rotation Cropping
0.25° 0.3° -0.25° -0.3° 1/4 256×256 0.003
0.69
0.002
0.80
0.001 Images
0.74 0.66 0.71 0.63 0.58 0.73 0.95
(a) Lena/NC
0.72 0.63 0.66 0.59 0.40 0.74 0.70 0.83 0.95
(b) Goldhill/NC
0.72 0.66 0.74 0.69 0.55 0.73 0.74 0.86 0.94
(c) Peppers/NC

The results of scaling, cropping and rotation attacks and
their related NCs are shown in Table IV. The proposed
algorithm has been compared with the algorithm proposed
by Wang et al. [11], Li et al. [12] and Lin et al. [18] based
on Lena image. The results of comparisons are depicted in
Table V. In addition, the results of applying JPEG
compression attacks to Lena image for both the proposed
method in this paper and the one proposed by Byun et al.
[13] have been compared with each other and the results of
this comparison have been taken in Table VI. The results
show that the proposed method has higher PSNR than the

100
[Downloaded from www.aece.ro on Tuesday, September 07, 2010 at 05:53:34 (UTC) by 85.185.144.179. Restrictions apply.]

Page 6

Advances in Electrical and Computer Engineering Volume 10, Number 3, 2010

other competitive ones. Besides, the robustness of
watermarked image against the applied attacks in the
proposed method is higher than the algorithm proposed by
Wang et al. [11] and Li et al. [12] and can compete with the
one proposed by Lin et al. [18]. Finally, the robustness of
the watermark against JPEG compression in the proposed
algorithm is the highest compared with competing methods
in this paper.

Table V. Comparing the proposed method with Wang[11], Li[12] and
Lin[18]
Proposed
Method
(PSNR=43.17
dB)
Lin et al. [18]
(PSNR =
42.02dB)
Li et al. [12]
(PSNR=40.6dB)
Wang et al.[11]
(PSNR =
38.2dB)
Attacks/NC
0.86 0.90 0.35 0.51
Median filter (3×3)
0.52 0.53 NA NA
Median filter (5×5)
0.49 0.34 0.15 NA
JPEG (QF = 10)
0.77 0.67 0.34 NA
JPEG (QF = 20)
0.91 0.82 0.52 0.15
JPEG (QF = 30)
0.99 0.96 0.52 0.26
JPEG (QF = 50)
1 0.97 0.63 0.57
JPEG (QF = 70)
1 0.99 0.78 1
JPEG (QF = 90)
0.74 0.59 0.46 0.37
Rotation (0.25°)
0.71 0.60 0.50 0.32
Rotation (-0.25°)
0.73 0.66 0.61 NA
Cropping (1/4)
0.58 0.88 0.35 NA
Scaling (256×256)
0.88 0.95 NA NA
Average filter (3×3)
0.44 0.47 NA NA
Average filter (7×7)
0.89 0.79 NA NA
Histogram
equalization

Table VI. Comparing the proposed method with Byun [13]
40 30 20 10 QF
Byun [13]
(PSNR=41.95)
Proposed Method
(PSNR=43.17)
100 90 80 70 60 50
1 0.99 0.970.940.90 0.830.81 0.77 0.600.49
1 1 1 1 1 0.990.990.910.770.49
IV. CONCLUSIONS
In this paper, we have proposed a novel blind watermarking
algorithm which has used the wavelet coefficients in the
form of super trees for embedding watermark in the host
image. The watermark bits have been embedded into the
second and third level coefficients of wavelet transform of
the host image based on distance vector. The extraction of
watermark is done based on an adaptive threshold, so that if
the distance vector extracted from the watermarked image is
greater than that threshold, bit 1 will be extracted; otherwise,
bit 0 will be detected. Proposed algorithm is tested with
various types of image processing attacks. The experimental
results have proved that the proposed algorithm has
significant robustness against common attacks such as
median filtering, JPEG compression, average filtering,
Gaussian noise, rotation,
equalization. In general, the proposed method is a good
useful tool for ownership identification and copyright
protection.
cropping and histogram
REFERENCES
[1] F. A. P. Petitcolas, R. J. Anderson and M. G. Kuhn, “Attacks on
Copyright marking system,” Proc. Of 2nd Int. Workshop on
Information Hiding, Portland, Oregon, USA, pp. 218-238, April
1998.
[2] C. C. Chang, P. Y. Lin, J. Sh. Yeh, “Preserving robustness and
removability for digital watermarks using subsampling and
difference correlation,” Information Sciences, vol. 179, pp. 2283–
2293, 2009.
[3] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, “Secure spread
spectrum watermarking for multimedia,” IEEE Trans. Image
Process., vol. 6, no. 12, pp. 1673–1687, Dec. 1997.
[4] I.J. Cox, M.L. Miller, and J.A. Bloom, “Digital Watermarking,”
Morgan Kaufmann Publishers, 2001.
[5] I. J. Cox and M. L. Miller, “The first 50 years of electronic
watermarking,” J. Appl. Signal Process., vol. 2, pp. 126–132, 2002.
[6] I. J. Cox, J. Kilian, F. T. Leighton, and T. Shamoon, “Secure spread
spectrum watermarking for multimedia,” IEEE Trans. Image
Process., vol. 6, no. 12, pp. 1673–1687, Dec. 1997.
[7] P. Meerwald and A. Uhl, “A survey of wavelet- domain
watermarking algorithms,” Proc. Of SPIE, Electronic Imaging,
Security and Watermarking of Multimedia Contents ІІІ, vol. 4314,
Jan. 2001.
[8] M. Barni, F. Bartolini, and A. Piva, “Improved wavelet-based
watermarking through pixel-wise masking,” IEEE Trans. Image Pro.
, vol. 10, no. 5, pp. 783-791, May 2001.
[9] P. Tay and J. P. Havlicek, “Image watermarking using wavelets,”
Proc. The 45th Midwest Symposium on Circuits and Systems, vol. 3,
pp. 258-261, 2002.
[10] J. Huang & C. Yang, “Image digital watermarking algorithm using
multiresolution wavelet transform,” In Proc. Of the IEEE Int. Conf.
on Systems, Man and Cybernetics, vol. 3, pp. 2977–2982, Oct. 2004.
[11] Sh. H. Wang and Y. P. Lin, “Wavelet Tree Quantization for
Copyright Protection Watermarking”, IEEE Trans. On image
Processing, vol. 13, no. 2, pp. 154-165, Feb. 2004.
[12] E. Li, H. Liang and X. Niu, “Blind Image Watermarking Scheme
Based on Wavelet Tree Quantization Robust to Geometric Attacks,”
In Proceedings of the IEEE WCICA, vol. 2, pp. 10256–10260, 2006.
[13] K. Byun, S. Lee, and H. Kim, “A watermarking method using
quantization and statistical characteristics of wavelet transform,”
Proc. IEEE PDCAT, pp. 689–693, 2005.
[14] Jui-Cheng Yen, Hun-Chen Chen, and Jui-Hsiang Juan, “Blind
Watermarking Based on The Wavelet Transform,” The 7th Int. Conf.
on Parallel and Distributed Computing, Applications and
Technologies (PDCAT’06), pp. 474-478, Dec. 2006.
[15] G. D. Wu, P. H. Huang, “Image watermarking using structure based
wavelet tree quantization”, In ICIS, Melbourne, Australia, pp. 315–
319, July 2007.
[16] A. Hu, N. Chen, “A Blind Watermarking Algorithm for Color Image
Based on Wavelet Transform and Fourier Transform”, the 9th
International Conference for Young Computer Scientists, Nov.
2008.
[17] M. J. Tsai, C. Ting Lin and J. Liu , “A Wavelet-based Watermarking
Scheme Using Double Wavelet Tree Energy Modulation”, the 15th
IEEE International Conference on Image Processing (ICIP 2008) ,
pp.417-420, Oct. 2008.
[18] W. H. Lin, Y. R. Wang, S. Horng, T. Kao, Yi Pan, “A blind
watermarking method using maximum wavelet coefficient
quantization,” Expert Systems with Applications , vol. 36, no. 9, pp.
11509-11516, Nov. 2009.
[19] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet
coefficients,” IEEE Trans. Signal Process., vol. 41, no. 12, pp.
3445–3462, Dec. 1993.
[20] C. Christopoulos, A. Skodras, and T. Ebrahimi, “The JPEG2000 still
image coding system: an overview,” IEEE Transactions on
Consumer Electronics, vol. 46, no. 4, pp. 1103–1127, Nov. 2000.
[21] D. Kundur and D. Hatzinakos, “Digital watermarking using
multiresolution wavelet decomposition,” in Proc. IEEE ICASSP,
Seattle, vol. 5, pp. 2969–2972, May 1998.
[22] F. A. P. Petitcolas, Weakness of Existing Watermark Scheme 1997
[Online]
http://www.petitcolas.net/fabien/watermarking/stirmark/index.html.
[23] W. H. Lin, S. J. Horng, T. W. Kao, P. Fan, C. L. Lee, Yi Pan, “An
Efficient Watermarking Method Based on Significant Difference of
Wavelet Coefficient Quantization,” IEEE Transactions on
Multimedia, Vol. 10, no. 5, pp. 746-757, Aug. 2008.
Available:

101
[Downloaded from www.aece.ro on Tuesday, September 07, 2010 at 05:53:34 (UTC) by 85.185.144.179. Restrictions apply.]