50 (IJCNS) International Journal of Computer and Network Security,

Vol. 2, No. 3, March 2010

Implementation of Watermarking for a Blind Image

using Wavelet Tree Quantization
S.M.Ramesh1, Dr.A.Shanmugam2, B.Gomathy3
1
Senior Lecturer, Dept.of ECE,
Bannari Amman Institute of Technology, Erode- 638401, India
smrameshme@yahoo.co.in
2
Professor, Dept.of ECE,
Bannari Amman Institute of Technology, Erode-638401, India
dras_bit@yahoo.com
3
Lecturer, Dept.of CSE,
Bannari Amman Institute of Technology, Erode- 638401, India
gomramesh@gmail.com

mid-frequency to ensure the transparency and robustness of

Abstract: This paper proposes implementation of
watermarking for a blind image using wavelet tree quantization
for copyright protection. In such a quantization scheme, there
exists a large significant difference while embedding a
watermark bit 1 and a watermark bit 0; it then does not require
any original image or watermark during the watermark
extraction. As a result, the watermarked images look lossless in
comparison with the original ones, and the proposed method can
effectively resist common image processing attacks; especially
for JPEG compression and low-pass filtering. Moreover, by
designing an adaptive threshold value in the extraction process,
our method is more robust for resisting common attacks such as
median filtering, average filtering, and Gaussian noise.
Experimental results show that the watermarked image looks
visually identical to the original, and the watermark can be
effectively extracted even after either an unintentional image
processing or intentional image attacks.
Keywords - Wavelet tree quantization, JPEG compression,
Low-pass filtering, Discrete Wavelet Transform.
1. Introduction
The Internet has popularized tremendously fast in our life in
the last decade. Due to the digitalization of documents,
images, music, videos, etc, people can access and propagate
them easily via the network. The watermarking technique
has been widely applied to digital contents for copyright
protection, image authentication, proof of ownership, etc.
This technique embeds information so that it is not easily
perceptible; that is, the viewer cannot see any information
embedded in the contents. The issue here is the detection of
the existence of watermarks in the digital contents to prove
ownership. The spatial and spectral domains are two
common methods for watermarking. In the spatial domain,
the watermark is embedded in the selected areas on the
texture of the host image. In the spectral domain, since the
spread spectrum communication is robust against many
types of interference and jamming, the host image is
transformed to the frequency domain using methods such as
Discrete Cosine Transform (DCT) or Discrete Wavelet
Transform (DWT); then the watermark is embedded in the
the watermarked image. For suggested inserting the
watermark into the perceptually significant portion of the
whole DCT transformed image, wherein a predetermined
range of low frequency components excludes the DC
component.
This watermarking scheme has been shown to be robust
against common attacks such as compression, filtering, and
cropping. The proposed a watermarking method, in
accordance with the multi-threshold wavelet coding
(MTWC), the successive subband quantization (SSQ) is
adopted to search for the significant coefficients. The
watermark is added by quantizing the significant coefficient
in the significant band by using different weights. The
proposed a watermarking method based on the qualified
significant wavelet tree (QSWT). The QSWT is derived
from the embedded zerotree wavelet algorithm (EZW). The
watermark is embedded in each of the two subbands of the
wavelet tree. Several watermarking methods used two sets of
coefficients, one to represent the watermark bit 0, and the
other to represent the watermark bit 1. According to the
embedded watermark bit, only a set of coefficients are
quantized each time. The proposed a blind watermarking
approach called differential energy watermarking. A set of
several 8×8 DCT blocks are composed and divided into two
parts to embed a watermark bit. The high frequency DCT
coefficients in the JPEG/MPEG stream are selectively
discarded to produce energy differences in the two parts of
the same set.
The wavelet coefficients of the host image are grouped
into wavelet trees, and each watermark bit is embedded
using two trees. One of the two trees is quantized with
respect to a quantization index, and both trees exhibit a
large statistical difference between the quantized tree and
the unquantized tree; the difference can later be used for
watermark extraction. EL [10] improved SHW [11] method
by adding the Human Visual System (HVS) to effectively
resist geometric attack. BKL [9] improved SHW [11]
method by using four trees to represent two watermark bits
 (IJCNS) International Journal of Computer and Network Security, 51
to improve visual quality. One of the four trees is quantized
according to the binary value of the two embedded
watermark bits. But these methods [9-11] cannot effectively
resist the attacks of low-pass filtering for JPEG
compression. In this paper, propose a blind watermarking
method based on wavelet tree quantization.
Figure 1. The watermark embedding procedure.
A watermarking technique is denoted as blind if the
original image is not needed for extraction. In previous
research, the watermark embedded in the significant
coefficients was found to be robust. The common issue is the
use of blind detection to find out whether the extracting
order is the same as the embedding order. Hence, propose a
watermarking method which embeds a watermark bit in the
maximum wavelet coefficient of a wavelet tree. The
proposed method is different from others, which use two
trees or two blocks to embed a watermark bit. We embed the
watermark by scaling the magnitude of the significant
difference between the two largest wavelet coefficients in a
wavelet tree to improve the robustness of the watermarking
Figure 2. The watermark extraction procedure.
The trees are so quantized that they exhibit a sufficiently
large enough energy difference between a watermark bit 1
and a watermark bit 0, which is then used for watermark
extraction. During extraction, an adaptive threshold value is
designed. A watermark bit 1 is extracted if the significant
Vol. 2, No. 3, March 2010
difference is greater than the adaptive threshold value;
otherwise, a watermark bit 0 is extracted. Experimental
results show that the proposed method is very efficient for
resisting various kinds of attacks.
2. Watermarking by Quantization of Wavelet
Trees
2.1 Wavelet Trees
A host image of size n by n is transformed into wavelet
coefficients using the 4-level discrete wavelet transform
(DWT).With 4-level decomposition and have 13 frequency
bands as shown in Fig. 1. The parent-child relationship can
be connected between these sub nodes to form a wavelet
tree. If the root consists of more than one node, then an
image will have many wavelet trees after the DWT. In this
case, we have 3 sub bands LH4, HL4, and HH4 as roots, and
the total wavelet trees are 3×(n/24×n/24) after an image of
size n by n is transformed by a 4-level wavelet transform. A
higher resolution level (such as level 3 in Fig.1) has more
significant coefficients than a lower resolution level (such as
level 2 in Fig.1). As we can see, when n = 512, there are 85
coefficients for a wavelet tree constructed from a node in
LH4 to LH1 following the parent-child relationship. For
avoiding attacks such as low pass filters, in our proposed
method we only need the largest two coefficients; these two
coefficients are selected from one coefficient of LH4 and
four coefficients of the same orientation in the same spatial
location in LH3 as shown in Fig. 2.
2.2 The Preprocess
With a four-level DWT and have 13 frequency bands as
shown in Fig.1. A higher level subband is more significant
than a lower subband. Using the LL4 subband as a root is
not suitable for embedding a watermark, since it is a low
frequency band that contains important information about
an image and easily causes image distortions. Embedding a
watermark in the HH4 subband is also not suitable, since the
subband can easily be eliminated, for example by a lossy
compression. The LH4 subband is more significant than the
HL4 subband, hence the LH4 subband has a higher priority
than the HL4 subband in the selection. A binary watermark
image W comprised of size Nw (≤ S = n/24×n/24) bits is
embedded. We represent each watermark bit as 1 or 0, and
use a 90 pseudorandom function with another seed to shuffle
Nw bits. According to the watermark bits embedded later, we
select Nw non-overlapping wavelet trees and compute the
global average significant difference of the total number of
the Nw wavelet trees using Eq.(1).
1 Nw
ε = { ------ ∑ (maxi – seci)} (1)
Nw i = 1
Where ε is the global average significant difference in all
Nw wavelet trees; {⋅} is the floor function; maxi is the local
maximum wavelet coefficient of the ith wavelet tree; seci is
52
the local second maximum wavelet coefficient of the ith
wavelet tree, 1≤ i ≤ Nw. If the embedded watermark bit is 1,
the local maximum coefficient is not quantized under
Maximum (ε ,T) Δi ≥ ε and is quantized under Δi <
Maximum (ε ,T) . If the embedded watermark bit is 0, the
local maximum coefficient is quantized by setting the local
maximum coefficient = local second maximum coefficient.
2.3 Watermark Embedding
Let maxi and seci be the local maximum wavelet
coefficient and the local second maximum wavelet
coefficient in a wavelet tree; the difference between both of
them is named as the local significant difference. The select
a threshold value β as an increment for quantization. The
larger the β is, the better robustness but the worse distortion
of the watermarked image will be. Each time at embed a
watermark bit and quantize the maximum coefficient in a
wavelet tree.
if (maxi < 0), then maxi = 0 (2)
Δi = maxi − seci, (3)
Where Δi denotes the significant difference between the
maximum coefficient and the second maximum coefficient
in the ith wavelet tree of Nw. Because some maximum
coefficients in a wavelet tree may be a negative value, and a
positive value has higher robustness than a negative value
under attacks, it will result in Δi more significant if we
modify the maximum coefficient from a negative value to a
positive value. When the Δi is more significant, it will be
more accurate at extracting watermark. To achieve the new
maximum coefficient be positive and to decrease the
distortion of watermarked image due to quantization, the
new maximum coefficient is set to a smallest positive value
zero here.
When embed a watermark bit 1, the maxi is quantized by
maxinew = maxi ,if (Δi ≥ Maximum (ε, T)) (4)
maxinew = maxi + β, if (Δi < Maximum (ε ,T)
and maxi is root) (5)
maxinew = maxi + β × γ , if (Δi < Maximum(ε ,T)
and maxi is root) (6)
Where maxinew denotes the new maximum coefficient in
the ith wavelet tree after embedding the watermark bit. The
maxi located at lower resolution level (child node) is less
robust than those located at higher resolution levels (root
node). The reason has been stated in Section II-B.
According to the band sensitivity, the coefficients quantized
at different resolution level are given different weights. The
quantized coefficient at the lower resolution level is given a
heavier weight than that at the higher resolution level.
Hence, if the maxi does not locate at the highest resolution
level, we will quantize the maxi by adding γ times of β
energy, here γ is a scale parameter and we set γ =1.5.
(IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 3, March 2010
Otherwise, the embed a watermark bit 0 according to Eq. (7)
as follows:
if (seci < 0 ), than maxinew = 0, secinew = 0;
otherwise maxinew = seci (7)
Where secinew denotes the new second maximum coefficient
in the ith wavelet tree after embedding the watermark bit 0.
When embedding a watermark bit 1, the maximum local
significant coefficient is quantized and added by an
incremental β if Δi < Maximum (ε, T); otherwise (i.e., Δi ≥
Maximum (ε, T)), it is kept the same as before. The reason
for not quantizing the maximum local significant coefficient
is that we won’t increase the distortion of the image. Some
images having small ε imply that their significant difference
is not obvious. We need an extra parameter T to improve the
robustness. The larger the T is, the higher probability the
maxi is quantized to a larger value; but, in the meanwhile,
the more distortion of the image will be as well. For
example, let ε =12 and Δi = 13. Suppose T is set to be less
than ε, such as T =11, maxi will not be quantized as Δi >ε
=12. On the other hand, if T is set to be larger than ε, such
as T =14, maxi will be quantized and increased by β as Δi
<T=14. On the contrary, when embedding a watermark bit
0, the value of maxi is quantized by decreasing to the local
second maximum seci and hence the new Δi will be equal to
0. Based on this strategy, there exists a large 91 energy
difference between embedding watermark bit 1 and
watermark bit 0.
3. Design of Watermark Decoder
3.1 The Decoder Design
In the proposed method, neither an original image nor an
original watermark image is required for the extraction
process. During the embedding process, embed a watermark
bit 1 by adding an energy β (or β × γ ) to the local maximum
wavelet coefficient in the wavelet tree, and embed a
watermark bit 0 by setting maxi = seci Hence, if the
wavelet tree was embedded a watermark bit 0, the local
significant difference between the largest two coefficients
will be close to zero; otherwise, if the wavelet tree was
embedded a watermark bit 1, the local significant difference
between the largest two coefficients will be greater than β.
In order to extract watermark bits correctly and the value
of y by Eq. (8).
1 Nw × α
y = --------- ∑ ϕ j ------ (8)
Nw × α j = 1
Where φ = {max1 - sec1 ,max2 - sec2 ,…..,maxi - seci }, for
i=1,2,…,N; the sorted φ is denoted as ϕ (φ) = { ϕ1,ϕ 2
...ϕ Nw }, ϕ1 < ϕ 2 < .... < ϕ Nw ; α is the scale
parameter, 0 < α ≤ 1; α is crucial to y and is used to
determine how many percentages of the significant
difference in φ can be used for the average. Hence, α marks
the minimal y value for extracting the watermark. The
larger α is, the larger the y will be. Suppose all embedded
watermark bits are 1 in the watermark. This means that the
 (IJCNS) International Journal of Computer and Network Security, 53
Vol. 2, No. 3, March 2010

difference between the maximum wavelet coefficients and
the second maximum wavelet coefficients for any embedded
wavelet tree is greater than β. The value of α should be set
as small as possible to avoid extraction errors (see Eq.
(9),(10)); the reason for this is that it can exclude those big
significant differences of the embedded wavelet trees in Eq.
(8). On the other hand, if all embedded watermark bits are 0
in the watermark, the value of α must be set as large as
possible. Therefore, α is sensitive to the content of the
watermark.
3.2 Watermark Extraction
Following Eqs. (8),(9) and (10), it would be easy to extract
the watermark. If the local significant difference is greater
than or equal to y, where 0 < y ≤ β, then the embedded
watermark bit could be 1; otherwise, the embedded
watermark bit could be 0. The watermark bit can be
extracted based on Eq. (9) and (10) as follows:
watermark bit = 1, if (maxi - seci ) > y (9)
watermark bit = 0, otherwise (10)
The peak signal-to-noise ratio (PSNR) to evaluate the
quality between the attacked image and the original image.
After extracting the watermark, the normalized correlation
coefficient (NC) is computed using the original watermark
and extracted watermark to judge the existence of the
watermark. The value of the NC coefficient is defined as
follows:
1 w h -1 w w -1
NC = ----------- ∑ ∑ w (i, j) × w ' (i, j) (11)
wh ×ww i = 0 j = 0
Where wh and ww are the height and width of the
watermark. w (i, j) and w' (i, j) are the values of the
coordinate (i, j) in the original watermark and the extracted
watermark, respectively. Here w (i, j) is set to 1 if it is a
watermark bit 1; otherwise, it is set to -1. w' (i, j) is set in
the same way. So the value of w(i, j) × w' (i, j) is either -1 or
1.
Table 1: Normalized Correlation Coefficients (NC) after
Attacks by JPEG Compression with the Quality Factors
(QF) 10, 20, 30, 40, 50, 60, 70, 80, 90, AND 100.

Normalized
Quality Factors Correlation Extracted
(QF) Coefficients Watermark
(NC)

10 0.33

20 0.65 E

30 0.78 ES

40 0.82 E A

ESHA 50 0.89 ES
(a) (b)
60 0.96 ESH
Figure 3. (a) The original image of Lena of size 512×512.
(b) The original binary watermark of size 32×16. 70 1 ESHA

80 1 ESHA

90 1 ESHA

100 1 ESHA

The Lena (512×512 pixels, 8 bits/pixel) obtained from

attacking the Stir mark benchmark and Photo Impact 11
software tools to simulate common image attacks. While
there is no attack, for the sake of brevity only the Lena
image and the binary watermark are shown in Fig. 3. Fig. 4
shows the watermarked image and the extracted result. Pre-
ESHA determine the scale parameter T at 10, β =7, γ =1.5, and α
(a) (b) =0.7. In the following, consider both geometric and
nongeometric attacks. Nongeometric attacks include JPEG
Figure 4. (a) Watermarked Lena with PSNR = 41.88dB compression, low pass filtering, histogram equalization, and
(b) The extracted watermark with NC = 1 sharpening. From Table 1, the proposed method can
correctly extract the watermark while the quality factors are
greater than 70 and it becomes worse if the quality factor is
4. Experimental Results
54 (IJCNS) International Journal of Computer and Network Security,
Vol. 2, No. 3, March 2010

decreased. The less the quality factor is, the more vague the In this paper, propose a wavelet-tree-based blind
extracted watermark is. watermarking method by quantizing the maximum wavelet
coefficient in a wavelet tree. Embed a watermark bit by
quantizing the maximum wavelet coefficient in a wavelet
tree. The trees are so quantized that they exhibit a
4.1 Experiment Analysis sufficiently large enough energy difference between a
Compare the proposed method with BKL [9], EL [10], watermark bit 0 and a watermark bit 1, which difference,
and SHW [11] using the Lena image. The watermark denoted as significant difference, is then useful for later
consists of 256 bit 1and 256 bit 0. The results are shown in watermark extraction. During extraction, an adaptive
Table 2. From Table 2, the PSNR of the proposed method is threshold value is designed. The magnitude of the
better than those of [11]. In this method, it is not so good for significant difference in a wavelet tree is compared to the
the rotation attacks with degree greater than ± 0.70; but it is adaptive threshold value. Furthermore, regarding each
far better than the listed methods; especially for low pass wavelet tree embedded with a watermark bit, we not only
filtering attacks for JPEG compression. can embed more bits in an image but can extract the
watermark without any need of the original image and
Table 2: Comparison of the Proposed Method and the watermark. Moreover by designing an adaptive threshold
Methods in [9], 10] and [11]. value in the extraction process. In addition to the copyright
protection, the proposed method can also be applied to data
hiding and image authentication.
Attacks / SHW EL BKL Proposed
NC PSNR = PSNR= PSNR= Method
38.2dB 40.6dB 41.54dB PSNR =
References
41.88dB
[1] D. P. Mukherjee, S. Maitra, and S. T. Acton, "Spatial
JEPG NA 0.15 0.17 0.33 domain digital watermarking of multimedia objects for
(QF = buyer authentication," IEEE Trans. Multimedia, vol. 6,
10) pp. 1-15, Feb. 2004.
[2] N. Nikolaidis and I. Pitas, "Robust image watermarking
JEPG NA 0.34 0.61 0.65 in the spatial domain," Signal Processing, vol. 66, pp.
(QF = 385-403, 1998.
20)
[3] A. K. Mahmood and A. Selin,"Spatially adaptive
JEPG 0.15 0.52 0.79 0.78
wavelet thresholding for image
(QF = watermarking,"presented at Proc.IEEE ICME, Toronto,
30) 2006 of Conference.
[4] J. R. Hernandez, M. Amado, and F. Perez- Gonzalez,
JEPG 0.23 0.52 0.83 0.82 "DCT- omain watermarking techniques for still images:
(QF = detector performance analysis and a new structure,"
40) IEEE Trans.Image Processing, vol. 9, pp.55- 68, Jan.
2000.
JEPG 0.28 0.52 0.89 0.89 [5] L. Sin-Joo and J. Sung-Hwan, "A survey of
(QF =
watermarking techniques applied to multimedia," in
50)
Proc. IEEE ISIE 2001, pp. 272-277
JEPG 0.46 0.59 0.94 0.96 [6] V. M. Potdar, S. Han, and E. Chang,"A survey of digital
(QF = image watermarking techniques,"presented at Proc.
60) IEEE INDIN 2005 of Conference.
[7] H. J. Wang, P. C. Su, and C. C. J. Kuo, "Wavelet-based
JEPG 0.57 0.63 0.97 1 digital image watermarking," Optics Express,vol.3, pp.
(QF = 491- 496, Dec. 1998.
70) [8] M. Hsieh, D. Tseng, and Y. Huang, "Hiding digital
watermarks using multiresolution wavelet transform,"
JEPG 0.89 0.71 0.99 1
IEEE Trans. Ind. Electron., vol. 48, pp. 875-882, Oct.
(QF =
80)
2001.
[9] B. K. Lien and W. H. Lin,"A watermarking method
JEPG 1 0.78 1 1 based on maximum distance wavelet tree
(QF = quantization," presented at 19th Conf. Computer
90) Vision, Graphics and Image Processing, 2006 of
Conference.
JEPG 1 0.88 1 1 [10] E. Li, H. Liang, and X. Niu,"An integer wavelet based
(QF = multiple logo-watermarking scheme," presented at
100) Proc. IEEE WCICA, 2006 of Conference.
[11] S. H. Wang and Y. P. Lin, "Wavelet tree quantization
5. Conclusion for copyright protection watermarking," IEEE Trans.
 (IJCNS) International Journal of Computer and Network Security, 55
Vol. 2, No. 3, March 2010

Image Processing, vol. 13, pp. 154-165, Feb. 2004.
[12] G. C. Langelaar and R. L. Lagendijk, "Optimal
differential energy watermarking of DCT encoded
images and video,"IEEE Trans. Image Processing,
vol. 10, pp. 148-158, Jan.2001.
[13] F. A. P. Petitcolas, Weakness of existing watermark
Scheme 1997[online] Available:
http://www.petitcolas.net/fabien/watermarking/stirma
rk/index.html.
[14] PhotoImpact 11 software, http://www.ulead.com,
Ulead system, Inc.
Guide Award” five times from Tamil Nadu state Government. He
is also the recipient of “Best Outstanding Fellow Corporate
Member Award” by Institution of Engineers (IE),India -2004 and
“Jewel of India” Award by International Institute of Education and
Management, New Delhi–2004 and “Bharatiya Vidya Bhavan
National Award for Best Engineering College Principal 2005” by
Indian Society for Technical Education (ISTE). “Education
Excellence Award” by All India Business& Community
Foundation, New Delhi.

Authors Profile

S.M.Ramesh received the B.E degree

in Electronics and Communication
Engineering from National Institute of
Technology
(Formerly Regional Engineering
College), Trichy, Bharathidhasan
University, India in the year 2001 and
the M.E, degree in Applied Electronics
from RVS College of Engineering and
Technology, Dindugal, Anna
University, India in the year 2004. From 2004 to 2005, he served
as a Lecturer in the Department of Electronics and Communication
Engineering, Maharaja Engineering College, Coimbatore, India.
From 2005 to 2006, he served as a Lecturer in the Department of
Electronics and Communication Engineering, Nandha Engineering
College, Erode, India. Since June 2006, he served as Senior
Lecturer, in the Department of Electronics and Communication
Engineering, Bannari Amman Institute of Technology,
Sathyamangalam, and Erode, India. He is currently pursuing the
Ph.D. degree in Anna University, Chennai-India, working closely
with Prof. Dr.A.Shanmugam and Prof Dr.R.Harikumar.

Dr.A.Shanmugam received the B.E,

degree in Electronics and
Communication Engineering from PSG
College of Technology., Coimbatore,
Madras University, India in the year
1972 and the M.E, degree in Applied
Electronics from College of
Engineering, Guindy, Chennai, Madras
University, India in the year 1978 and
received the Ph.D. in Computer
Networks from PSG College of
Technology., Coimbatore, Bharathiyar University, India in the year
1994.From 1972 to 1976, he served as a Testing Engineer at Test
and Development Center, Chennai, India. From 1978 to 1979, he
served as a Lecturer in the Department of Electrical Engineering,
Annamalai University, India. From 1979 to 2002, he served
different level as a Lecturer, Asst.Professor, Professor and Head in
the Department of Electronics and Communication Engineering of
PSG College of Technology, Coimbatore, India. Since April
2004.He assumed charge as the Principal, Bannari Amman
Institute of Technology, Sathyamangalam, Erode, India.

He works in field of Optical Networks, broad band computer

networks and wireless networks, Signal processing specializing
particularly in inverse problems, sparse representations, and over-
complete transforms.Dr.A.Shanmugam received “Best Project