J. Vis. Commun. Image R.

31 (2015) 75–85

Contents lists available at ScienceDirect

J. Vis. Commun. Image R.

journal homepage: www.elsevier.com/locate/jvci

Digital watermark extraction using support vector machine with

principal component analysis based feature reduction q
Vivek Singh Verma a, Rajib Kumar Jha b,⇑, Aparajita Ojha a
a
PDPM, Indian Institute of Information Technology, Design and Manufacturing Jabalpur, 482005, India
b
Indian Institute of Technology Patna, Bihar 800013, India

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a new approach for watermark extraction using support vector machine (SVM) with
Received 13 August 2014 principal component analysis (PCA) based feature reduction. In this method, the original cover image is
Accepted 1 June 2015 decomposed up to three level using lifting wavelet transform (LWT), and lowpass subband is selected for
Available online 9 June 2015
data hiding purpose. The lowpass subband is divided into small blocks, and a binary watermark is
embedded into the original cover image by quantizing the two maximum coefficients of the block. In
Keywords: order to extract watermark bits with maximum correlation, SVM based binary classification approach
Lifting wavelet transform
is incorporated. The training and testing patterns are constructed by employing a reduced set of features
Support vector machine
Coefficient difference
along with block coefficients. Firstly, different features are obtained by evaluating the statistical param-
Feature reduction eters of each block coefficients, and then PCA is utilized to reduce this feature set. As far as security is
Watermarking concerned, randomization of coefficients, blocks, and watermark bits enhances the security of system.
Attacks Furthermore, energy compaction property of LWT increases the robustness in comparison to conven-
PCA tional wavelet transform. A comparison of the proposed method with some of the recent techniques
Digital image watermarking shows remarkable improvement in terms of robustness and security of the watermark.
Ó 2015 Elsevier Inc. All rights reserved.

1. Introduction greater part of these methods is to investigate watermark detect-

The ease of illegal reproduction of multimedia contents has
seriously affected the protection of multimedia intellectual prop-
erty rights. Currently, digital watermarking has enlisted consider-
able prominence within the research community due to its
remarkable benefits over the traditional data hiding techniques.
Watermarking technique should at least satisfy the following
requirements: (1) Imperceptible. (2) Without seriously affecting
the image quality it is difficult to remove/alter the watermark.
(3) Robust against various image processing operations, or attacks.
Broadly, watermarking techniques are developed either in the spa-
tial domain or in transform domain. A review of literature reveals
that transform domain techniques are typically more robust to var-
ious signal processing attacks as compared with spatial domain
techniques. Various spatial domain based techniques have been
proposed with focus on ensuring the integrity of digital media
and other related issues [1–10]. Numerous transform domain
based watermarking schemes have also been proposed and the
q
This paper has been recommended for acceptance by M.T. Sun.
⇑ Corresponding author.
E-mail addresses: viveksv10@gmail.com (V.S. Verma), jharajib@gmail.com
(R.K. Jha), aparajitaojha@gmail.com (A. Ojha).
able quality against different signal processing attacks (see for
example, [11–21]). Recently, Fu [21] proposed a novel discrete
cosine transform (DCT) based image watermarking scheme.
Using Bose–Chaudhuri–Hocquenghem (BCH) code, the watermark
bits are embedded into the host by modulating the relationships
between the selected DCT coefficients. Modulation is similar to
the JPEG compression and improves the robustness against JPEG
compression but not more effectively.
In order to achieve more robustness against geometrical
attacks, various content-based synchronization schemes have been
proposed. Recently, Deng et al. [22] presented a content-based
watermarking scheme that combines the invariant feature extrac-
tion with watermark embedding by using Tchebichef moments
(TM). In this approach, Harris Laplace detector is utilized to gener-
ate feature point based local disk which is invariant to the geomet-
ric distortions whereas, TMs are employed to describe the global
characteristics of the local invariant region. More recently, Gao
et al. [23] have proposed another variant of feature-based image
watermarking by incorporating the advantages of affine invariant
point detector, the image normalization, and orientation align-
ment. To improve the robustness, this scheme includes mainly
three components: feature selection procedure, geometrically
invariant region, and indirect inverse normalization. Although both

http://dx.doi.org/10.1016/j.jvcir.2015.06.001
1047-3203/Ó 2015 Elsevier Inc. All rights reserved.
76 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85
the methods have exhibited robustness against various important
attacks, their performance is not so good in case of geometrical dis-
tortions and multiple attacks.
In recent years, there has been a trend of using machine
learning and soft computing approaches to design more robust
and intelligent watermarking techniques. Yu et al. [24] proposed
a digital watermarking method in spatial domain based on artifi-
cial neural networks (ANNs), where employing multi-layer percep-
trons (MLPs) based neural networks to learn the characteristics of
the embedded watermark related to the watermarked image. Shieh
et al. [25] presented a genetic algorithm (GA) based optimization
scheme in watermarking, where GA was applied to find optimal
frequency bands for watermark embedding. Another adaptive
scheme based on fuzzy-ART was proposed by Chang et al. in [26]
using DCT domain. Adapting discrete wavelet transform (DWT)
and GA, Ramanjaneyulu et al. [27] have recently proposed a copy-
right protection scheme which outperforms the previous methods.
In this approach, the watermark embedding and extraction pro-
cesses are characterized with parameters and GA is utilized for
parameter optimization. Though, this method provides more
robustness against various types of image processing operations,
but, fails to achieve efficient robustness against JPEG compression
and sharpening operation.
A systematic review of literature indicates that the performance
of SVM based schemes are superior compared to that of traditional
learning methods. Particularly, the SVM based solution for the clas-
sification problem outperforms the solutions of traditional neural
networks, such as MLPs and radial basis functions [28]. Taking
the advantages of SVM, significant work has been done on water-
mark insertion and extraction. Tsai and Sun [28] proposed a color
image watermarking approach in spatial domain based on SVM,
where watermark extraction was considered as a binary classifica-
tion problem. Wang et al. [29] presented an image watermarking
approach using SVM to resist desynchronization attacks. In another
SVM based scheme proposed by Peng et al. [30], the special fre-
quency band and the property of image in multiwavelet domain
are utilized for watermarking. Though, the technique is quite
robust against several attacks, but, fails to resist JPEG compression,
median filtering, average filtering, and scaling attacks effectively.
Using regression of SVM, Wang et al. [31] and Shen et al. [32] pre-
sented robust watermarking schemes. Recently, Wang et al. [33]
proposed a robust image watermarking scheme based on the
SVM and Gaussian Hermite moments (GHMs). The original image
is transformed using non-subsampled contourlet transform
(NSCT) and corresponding lowpass subband is selected to embed
a binary watermark. During the watermark extraction, low-order
GHMs of training image are computed and watermark bits are
extracted from the watermarked image which is corrected by well
trained SVM model. The scheme is shown to be robust against
different geometric attacks, but, fails to resist JPEG compression
attacks effectively. More recently, Yang et al. [34] have also pro-
posed a robust image watermarking technique in undecimated dis-
crete wavelet transform (UDWT) domain. The scheme uses a fuzzy
support vector machine (FSVM) to learn geometric distortion
parameters by using low order Zernike moments. Watermark
extraction is carried out after watermarked image has been syn-
chronized without the original cover. One of the main drawback
of this scheme is excessive computational time in extraction. It is
also fragile to local geometrical distortions, such as random bend-
ing, and column or line removal, and fails to resist JPEG compres-
sion attacks more effectively.
Researchers have already done lots of noteworthy work in the
field of digital image watermarking. Even though, it is interesting
to point out that current methods designed for image integrity
may not be perfect. The proposed method provides more
robustness in all aspects of image integrity and outperforms espe-
cially in case of JPEG compression with high compression ratio.
1.1. Key contribution
Main aim of this paper is to design a system with enhanced fea-
tures of robustness and extraction of the watermark with maxi-
mum possible correlation. With this concern, the watermark
embedding and extraction procedures are designed in such a way
that maximum performance balance between capacity and robust-
ness could be achieved. In this scheme, the watermark embedding
is based on quantization of maximum coefficients of the block;
and, the watermark extraction is treated as binary classification
based. The novelty of the proposed work is that it uses computa-
tionally simple features set for training and testing the SVM, and
PCA is utilized to generate an optimally reduced set, which is not
experienced in existing schemes. In place of conventional wavelet
transform, lifting wavelet transforms have been used in the pro-
posed approach, as it is fast and requires less memory, and the
transform can be modified locally while preserving invertibility.
The motivation of using SVM is owing to the good learning capabil-
ity and generalization performance even when the watermarked
image is vigorously distorted [28,30]. Well established statistical
parameters such as Kurtosis, Skewness, and Entropy are used to
generate features set. Key features of the proposed scheme include
1. Using standard statistical parameters and PCA based feature
reduction makes the technique computationally more efficient.
2. Energy compaction property of LWT and good learning ability of
SVM can tolerate more image distortions.
3. Key based randomization of coefficients, blocks and watermark
bits enhances the security of the system.
4. In case of watermark extraction, neither original image nor
original watermark image is required.
Rest of the paper is organized as follows: In Section 2, we have
briefly sketched the concept of LWT and PCA based feature
reduction. Section 3 is devoted to the proposed watermarking
technique detailing embedding and extraction of watermark.
While in Section 4, experimental results and comparisons with dif-
ferent existing techniques are shown, conclusion and future scope
are discussed in Section 5.
2. Preliminaries
This section presents the background material for subsequent
sections. We begin with a brief introduction to LWT then illustrate
the general concept of PCA based dimensionality reduction (or fea-
ture reduction).
2.1. Lifting wavelet transform
Lifting wavelet transform was introduced by Sweldens [35]. The
lifting scheme conquers the limitations of the traditional wavelet
transform by reducing the computation time and memory require-
ments. By directly analyzing the problem in integer domain the
lifting wavelet simplifies the problem of reversibility which is
not experienced in conventional wavelet transform due to the
floating point coefficients, we get after transformation. Various
advantages of LWT over traditional wavelet transform helps on
developing watermarking schemes with increase of computational
efficiency.
In general, the lifting scheme includes three basic steps that are
splitting, prediction, and update [36,37].
 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85
Split: Splitting is also called as lazy wavelet transform. In this
step the input signal Sj is divided into even and odd samples:
Sj; 2k and Sj; 2kþ1 , and is expressed as:
SplitðSj Þ ¼ ½ev enðSj Þ; oddðSj Þ
Predict: This step can be viewed as a highpass filtering opera-
tion. Here, we predict odd samples Sj; 2kþ1 using even samples
Sj; 2k and the abstract difference dj1 is generated as:
dj1 ¼ Sj; 2kþ1  P½Sj; 2k 
where P[] is the prediction operator. The high frequency compo-
nent dj1 is nothing but an error between the original sample
and its predicted value.
Update: This step can be viewed as a lowpass filtering opera-
tion. The low-frequency component Sj1 represents a coarse
approximation to the original signal Sj which is obtained by
applying an update operator U[] as:
Sj1 ¼ Sj; 2k þ U½dj1 
The decomposition and reconstruction of given signal is shown in
Fig. 1.
2.2. PCA based feature reduction
The dimensionality reduction problem can be investigated as
follows: for given data set x ¼ ðx1 ; . . . ; xn ÞT with n-dimension, find
its lower dimensional representation, y ¼ ðy1 ; . . . ; yk ÞT with k 6 n,
that catches the content in the original data, according to some
method.
In the literature, PCA based solution for dimensionality reduc-
tion is among the most prominent scheme. It is also known as
the Karhunen–Loève transform, the singular value decomposition
(SVD), and the empirical orthogonal function (EOF) method
[38,39]. For given data set of n dimensions, PCA tries to observe a
linear subspace of dimension k 6 n in such a way that the data
points lie fundamentally on this linear subspace. Such a reduced
subspace endeavors to keep up most of the variability of the
information.
In essence, PCA tries to reduce the dimension of the data by
evaluating a couple of orthogonal linear combinations (LCs) of
the original variables with the largest variance. The first LC; y1 , is
the linear combination with the largest variance. We have
y1 ¼ xT v 1 , where the n-dimensional coefficient vector
v 1 ¼ ðv 11 ; . . . ; v 1n ÞT solves
v 1 ¼ arg max
kv ¼1k
VarfxT v g
where Var½u denotes variance of the vector u. Orthogonal to the
first linear combination LC, the second LC is the linear combination
Sj ,2 k
+ Sj - 1
Sj Split P U
– dj - 1
Sj ,2 k + 1
Decomposition
Fig. 1. Decomposition and reconstruction of lifting wavelet.
77
with the second largest variance, and so on. Number of LCs are same
as the number of variables in the given dataset. The first few LCs of
some datasets show most of the variance, so that the remaining can
be disregarded with negligible loss of information.
ð1Þ Since, the variance relies on the scale of the variables, it is
customary to first standardize each variable to have mean zero
and standard deviation one. After the standardization, the original
variables with perhaps distinctive units of estimation are all in
equivalent units. Accepting a standardized information with the
ð2Þ experimental covariance matrix
1
Rnn ¼ XX T ð5Þ
m
using spectral decomposition theorem, R can be defined as follows
R ¼ Z KZ T ð6Þ
where K ¼ diagðk1 ; . . . ; kn Þ is the diagonal matrix of the ordered
eigenvalues k1 6    6 kn , and Z is a n  n orthogonal matrix con-
ð3Þ taining the eigenvectors. It can be shown [40] that the LCs are given
by the n rows of the n  m matrix Y, where
Y ¼ ZT X ð7Þ
And, the subspace spanned by the first k eigenvectors has the smallest
mean square deviation from X among all subspaces of dimension k.
3. Proposed method
This section illustrates a new image watermarking technique
with good quality and reasonable resistance toward various image
processing operations. Firstly, the original image is decomposed
using LWT up to third level, and LH3 subband is obtained for
watermark insertion. Then, LH3 subband coefficients are randomly
shuffled using secret seed (key1), and consecutive non-overlapped
coefficients are grouped to form blocks. Then, all the blocks are
randomly shuffled using different seed (key2). In each block, the
difference of two maximum coefficients ðcðnÞ; cðn  1ÞÞ is calcu-
lated, called as coefficient difference Dmax . Watermark bits are also
randomly shuffled in same way by using a different seed (key3),
and embedded in each block by quantizing the two maximum
coefficients of the block. The detailed embedding procedure is
illustrated in watermark embedding section.
The SVM approach with PCA based feature reduction is applied
in case of watermark extraction. Firstly, the blocks of LH3 subband
coefficients of watermarked image are obtained in same way as in
embedding process. Then, different features of block coefficients
ð4Þ are evaluated with the help of standard statistical parameters such
as Kurtosis, Skewness, and Entropy. Then, obtained feature set is
processed using PCA to generate reduced feature set. The training
and testing data is prepared by concatenating the reduced feature
Sj ,2 k
–
U P Merge Sj
+
Sj ,2 k + 1
Reconstruction
78 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85

set along with corresponding block coefficients. The prepared Step 3. Consecutive non-overlapped coefficients of LH3 sub-
training pattern and the reference watermark are used as input band are grouped to form blocks, and all these blocks are ran-
to train the SVM. Finally, the watermark is extracted by the trained domly shuffled using secret seed key2.
SVM. The complete extraction procedure is illustrated in water- Step 4. Using (10), the average coefficient difference c is
mark extraction section. computed.
Step 5. Generate a one dimensional vector of binary watermark
3.1. Watermark embedding procedure W of length N w by concatenating the reference information and
signature information.
In this paper, watermark W of length N w is divided into two Step 6. For all N w bits of watermark do following:
components, the reference information Rf of length N r and the 6.1. In each block find the two largest coefficients cðnÞ
signature information S of length N s . The reference information bits and cðn  1Þ.
are generated according to secret seed key3. One dimensional 6.2. If the embedded watermark bit is 1, modify cðnÞ
vector of signature information bits (i.e. logo watermark image) using (8).
are randomly shuffled using the same seed key3. Here, we 6.3. If the embedded watermark bit is 0, modify cðnÞ
have concatenated them into a single and represented as using (9).
W ¼ Rf þ S ¼ w1 ; . . . ; wNr ; wNr þ1 ; . . . ; wNr þNs ¼ w1 ; . . . ; wNw . The ref- Step 7. All the blocks and updated coefficients of LH3 subband
erence information is utilized to generate a training pattern, while, are inversely shuffled using the same keys as used in Step 3 and
the signature information is utilized to generate a testing pattern. Step 2.
The mathematical formulation to embed watermark bits is as Step 8. Watermarked image is obtained by doing inverse LWT
follows. For each watermark bit, quantize the maximum operation.
coefficients of the corresponding block in LH3 subband using (8)
and (9).
If the watermark bit is 1, 3.2. Watermark extraction procedure

ci ðnÞ þ T; if Dmax
i < maxðc; TÞ
ci ðnÞ ¼ ð8Þ Watermark extraction problem can be treated as a classification
ci ðnÞ; otherwise
problem involving binary classes, and the SVM is used to realize
and, if the watermark bit is 0, watermark extraction. In the proposed watermark embedding pro-
cedure, we observed that there is a nonlinear function relationship
ci ðnÞ ¼ ci ðnÞ  Dmax
i ð9Þ between watermark bit wi 2 f1; 0g and two maximum coefficients
where T is a threshold value used in quantization, and is the Dmax in ith block, i.e., wi ¼ f ðci ðn  1Þ; ci ðnÞÞ. The SVM is employed here,
i
difference between two maximum coefficients ðci ðnÞ; ci ðn  1ÞÞ of as it has powerful nonlinear mapping ability in comparison to
the corresponding ith block. The average coefficient difference value other machine learning methods. It is assumed that the water-
c of all Nw blocks is expressed as: marked image may get distorted from different image processing
$P % operations or attacks. This results in the change of LWT coefficients
Nw max
i¼1 Di of an image after attacking. So, the designed detector should have
c¼ ð10Þ
Nw high ability to detect the watermark under this contaminated envi-
ronment. In machine learning point of view, the watermark detec-
here, N w is the total number of blocks in LH3 subband where the tion can be realized as an ability of generalization. In proposed
watermark bits are embedded. technique, the SVM is used as watermark detector. Because, the
The block schematic of the proposed embedding process is SVM has good generalization ability, and it can help on improving
illustrated in Fig. 2. The steps associated in embedding of water- the robustness of watermarking system.
mark bits are listed as follows: Here, our scheme is oblivious because the original image is not
necessary in case of watermark detection. As discussed above in
Step 1. The LH3 subband is obtained by three level decomposi- embedding section, that two classes of watermark, i.e. signature
tion of original image using LWT. information and reference information are embedded into the
Step 2. LH3 subband coefficients are then randomly shuffled cover image. Firstly, we extract the LH3 subband coefficient blocks
using secret seed key1. of watermarked image where the reference information is

Fig. 2. Watermark embedding process.

 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85 79

embedded. Then, different statistical parameters such as: Kurtosis where f i ð1Þ; f i ð2Þ; . . . ; f i ðmÞ are reduced features, ci ð1Þ;
(p1 ), Skewness (p2 ), Entropy (p3 ), Standard deviation (p4 ), Mean ci ð2Þ; . . . ; ci ðnÞ are coefficients of the corresponding ith block,
(p5 ), Variance (p6 ), Mode (p7 ), Median (p8 ), Covariance (p9 ), and wi is desired output, i ¼ 1; 2; . . . ; N r .
Poisson probability density function (p10 ), Moment (p11 ), and 5.2. The ‘‘RBF’’ kernel of SVMs is described as follows:
Quantiles (p12 ) of all these blocks are evaluated. This set of param-
eters (p1 ; p2 ; . . . ; p12 ) is called as features set and given as input to Kðxi ; xÞ ¼ expðkxi  xk2 =r2 Þ ð12Þ
the PCA for feature reduction. The output of PCA is called as here, r is the width parameter of ‘‘RBF’’ kernel.
reduced features set, and it is represented as (f 1 ; . . . ; f m ), where 5.3. The optimal model can be defined as follows:
m is the total number of reduced features. Considering the block X
Nr
1X Nr XNr
size of 2  2, the total number of LH3 subband coefficient blocks Maximize ai  ai aj yi yj Kðxi ; xj Þ
are 1024, in which, 512 blocks are used to prepare training pattern i¼1
2 i¼1 j¼1
ð13Þ
and remaining 512 blocks are utilized to generate the testing pat- X
Nr

tern. So, the size of features set fpi ðkÞg is 512  12, and PCA is uti- s:t: ai yi ¼ 0; 0 6 ai 6 C; i ¼ 1; . . . ; N r
lized to reduce this set. i¼1

A set of training patterns is constructed by employing the where ai ði ¼ 1; 2; . . . ; N r Þ are training parameter, and C is the
reduced features set along with the coefficients of corresponding penalty parameter. If the optimal solution is a ¼
blocks. Similarly, the set of testing patterns (where the signature ða1 ; a2 ; . . . ; aNr Þ, then the decision function y can be expressed
information is embedded) is generated. The SVM can be trained as
by utilizing the set of training patterns. Finally, the trained SVM !
X
Nr
is applied to classify a set of testing patterns. Following the results y ¼ f ðxÞ ¼ sign ai yi Kðxi ; xÞ þ b ð14Þ
produced by the classifier (the trained SVM), the digital watermark i¼1

can be extracted.
here, b 2 R is a bias.
The block schematic of the proposed extraction process is illus-
Step 6. To extract the signature information S0 , construct the
trated in Fig. 3. Extraction of the binary watermark is formulated as
testing pattern w0 for well trained SVM. w0 ¼ fx0u ¼
follows: 0 0 0
ðf u ð1Þ; f u ð2Þ; . . . ; f u ðmÞ; c0u ð1Þ; c0u ð2Þ; . . . ; c0u ðnÞÞg. Then, by using
Step 1. The LH3 subband coefficients of watermarked image are well trained SVM in Eq. (14), we can obtain their corresponding
obtained in same way as in embedding process. outputs fy0u ju ¼ 1; . . . ; N s g, i.e.
Step 2. Using secret keys key1 and key2, the coefficients and 
y0u ¼ f ðx0u Þ; u ¼ 1; . . . ; Ns
corresponding blocks of LH3 subband are randomly shuffled. 0 0 0
x0u ¼ ðf u ð1Þ; f u ð2Þ; . . . ; f u ðmÞ; c0u ð1Þ; c0u ð2Þ; . . . ; c0u ðnÞÞ 2 w0
Step 3. Generate features set (fpi ðkÞjk ¼ 1; 2; . . . ; lg) of such
blocks where the reference information (Rf ) bits are embedded. ð15Þ
Step 4. Apply PCA onto this features set for feature reduction, Thus, the signature information bits ðS0 Þ can be obtained by
and obtain the reduced features set ff i ðkÞjk ¼ 1; 2; . . . ; mg, 
1; if y0u ¼ 1;
where m 6 l. w0u ¼ ð16Þ
Step 5. Training the SVM 0; if y0u ¼ 1:
5.1. Construct a training pattern w by employing the Step 7. The extracted signature information bits are stored in
reduced set of features along with coefficients of blocks one dimensional vector S0 of length N s , which are inversely shuf-
where the reference information ðw1 ; w2 ; . . . ; wNr Þ has
fled using the same seed key3 as used in embedding process.
been embedded:
Step 8. Reshape the extracted S0 same as the size of original logo
n o watermark image.
w ¼ ðxi ; yi Þ 2 RN  Rji ¼ 1; 2; . . . ; N r
¼ fðf i ð1Þ; f i ð2Þ; . . . ; f i ðmÞ; ci ð1Þ; ci ð2Þ; . . . ; ci ðnÞÞ; wi ji 4. Experimental results and discussions
¼ 1; 2; . . . ; Nr g
This section is devoted to the experimental results using the
ð11Þ proposed scheme and its comparison with some of the recent

Fig. 3. Watermark extraction process.

80 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n) (o)

Fig. 4. Standard benchmark images (a) Lena, (b) Goldhill, (c) Peppers, (d) Man, (e) Airport, (f) Tank, (g) Truck, (h) Elaine, (i) Boat, (j) Barbara, (k) Mandril, (l) House, (m)
Jetplane, (n) Livingroom and (o) Woman.

Fig. 6. (a) Watermarked image ‘‘Lena’’ with PSNR = 41.5107 dB. (b) Extracted
Fig. 5. (a) Original image ‘‘Lena’’ of size 512  512. (b) Original binary watermark of watermark with NC = 1, and BER = 0.
size 32  16.

watermarking schemes. Large dataset of images i.e. more than 50
real life images and standard benchmark images including
Lena; Goldhill; Peppers; etc. (512  512), 8 bits/pixel are chosen for
the experiment [41], shown in Fig. 4. For the sake of brevity, only
‘‘Lena’’ and the watermark image (32  16) are shown in Fig. 5
along with the watermarked image (Fig. 6(a)) and the extracted
watermark (Fig. 6(b)), to demonstrate the visual quality of the
extracted watermark and the cover image. Some necessary param-
eters used in proposed watermarking technique are as follows: the
embedding threshold T = 11 and block size of 2  2. According to
testing results on different test images, it is observed that the
 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85 81

Table 1
Results of test image ‘‘Lena’’ under different attacks.

Attacks PSNR (dB) NC BER CR (%) Extracted watermark

Tr(0, 10) 19.06 0.7148 0.1426 85.74
Tr(10, 0) 17.93 0.7617 0.1191 88.08
Tr(2, 15) 17.15 0.7891 0.1055 89.45
Scl(0.5) 33.57 0.9453 0.0273 97.26
Scl(0.9) 37.47 0.9844 0.0078 99.21
Scl(1.5) 40.01 0.9961 0.0020 99.80
Rt(50 ) 11.45 0.7305 0.1348 86.52
Rt(100 ) 9.67 0.6758 0.1621 83.78
Rt(150 ) 8.87 0.6680 0.1660 83.39
Cr(10%) 23.80 0.9922 0.0039 99.60
Cr(20%) 19.66 0.9609 0.0195 98.04
Cr(50%) 11.94 0.8789 0.0605 93.94
Md(33) 34.71 0.9258 0.0371 96.28
Md(55) 31.05 0.8203 0.0898 91.01
Avg(33) 31.65 0.8633 0.0684 93.16
Avg(55) 28.21 0.7813 0.1094 89.06
GF(33) 38.53 0.9570 0.0215 97.85
Gs(0.05) 13.63 0.8320 0.0840 91.60
Slp(0.01) 25.36 0.9141 0.0430 95.70
Spn(0.01) 25.54 0.9102 0.0449 95.50
Shp 23.38 0.9648 0.0176 98.24
HE 19.07 0.9336 0.0332 96.67

‘‘RBF kernel’’ (see Eq. (12)) shows better performance over other where Iði; jÞ and I0 ði; jÞ are the gray level intensity value at location
kernel functions, such as ‘‘Linear kernel’’ and ‘‘Polynomial kernel’’ ði; jÞ of the original image and attacked watermarked image,
under different attacks. Finally, it is observed that on considering respectively; and m1 ; m2 are image dimensions.
the size of reference information as 512 bits, the possibility of The NC and BER can be computed as follows:
watermark extraction with maximum correlation is high. PP
i W ij W 0ij
j
Performance of proposed watermarking technique is investi- NC ¼ ð19Þ
gated by measuring its imperceptibility and robustness. For imper- hw
ceptibility, peak-signal-to-noise rate (PSNR) is employed to where W ij and W 0ij are the values located at ði; jÞ of the embedded
evaluate the difference between the original image I and water- and extracted watermark and set to 1 if it is a watermark bit 1
marked image I0 . For robustness, normalized correlation coefficient otherwise it is set to 1; and h; w are height and width of the water-
(NC) and bit error rate (BER) are used to measure difference mark, respectively.
between the original watermark W and the extracted watermark
B
W 0 . It needless to note that the larger PSNR, the better impercepti- BER ¼ ð20Þ
bility and a method has higher NC value or lower BER, it is more hw
robust. where, B is the number of erroneously detected bits, h  w is the
The PSNR is given by watermark image dimensions.
As discussed in Section 3.2, the watermark extraction problem
2552
PSNR ¼ 10log10 ð17Þ is realized as a classification problem. Therefore, the efficiency of
MSE watermark extraction can be analyzed on the basis of classification
where MSE is the mean-square error between the original image ability of the learning machine, and it is realized as classification
and the attacked image, and it is defined as follows: rate (CR), which is expressed as:
Pm1 1 Pm2 1 2 T1
i¼0 j¼0 ðIði; jÞ  I0 ði; jÞÞ CR ¼  100 ð21Þ
MSE ¼ ð18Þ T2
m1  m2

Fig. 7. The relationship between the PSNR, threshold value (T) and the number of
embedded watermark bits. Fig. 8. PSNR of 15 watermarked images.
82 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85

Table 2
NC values of extracted watermark for JPEG attack with different QF over fifteen images.

Images JPEG attacks (QF)

20 30 40 50 60 70 80 90 100
Lena 0.9063 0.9609 0.9688 0.9805 1 1 1 1 1
Goldhill 0.9297 0.9531 0.9609 0.9727 1 1 1 1 1
Peppers 0.9375 0.9453 0.9688 0.9844 1 1 1 1 1
Man 0.9102 0.9492 0.9766 0.9922 1 1 1 1 1
Airport 0.9258 0.9414 0.9727 0.9805 1 1 1 1 1
Tank 0.9648 0.9922 0.9922 0.9961 1 1 1 1 1
Truck 0.9414 0.9570 0.9648 0.9727 1 1 1 1 1
Elaine 0.9297 0.9492 0.9766 0.9922 1 1 1 1 1
Boat 0.9102 0.9648 0.9805 1 1 1 1 1 1
Barbara 0.9570 0.9766 0.9688 0.9805 1 1 1 1 1
Mandril 0.9297 0.9531 0.9609 0.9766 1 1 1 1 1
House 0.9102 0.9375 0.9531 0.9749 1 1 1 1 1
Jetplane 0.8906 0.9375 0.9692 0.9961 1 1 1 1 1
Livingroom 0.9297 0.9492 0.9688 0.9805 1 1 1 1 1
Woman 0.9180 0.9609 0.9766 0.9961 1 1 1 1 1

Table 3
NC values of extracted watermark for different attacks over fifteen images.

Images Attacks
Md(33) Md(55) Avg(33) Avg(55) GF(33) Gs(0.05) Slp(0.01) Spn(0.01) Shp HE
Lena 0.9258 0.8203 0.8633 0.7813 0.9570 0.8320 0.9141 0.9102 0.9648 0.9336
Goldhill 0.9375 0.8477 0.8867 0.8398 0.9609 0.8320 0.9180 0.9297 0.9414 0.8984
Peppers 0.9180 0.7773 0.8594 0.7852 0.9609 0.8008 0.9102 0.8906 0.9766 0.9375
Man 0.9141 0.8281 0.8906 0.8164 0.9727 0.8281 0.8945 0.9219 0.9336 0.9297
Airport 0.8828 0.7695 0.8906 0.8125 0.9727 0.7969 0.9023 0.9258 0.9414 0.8789
Tank 0.9688 0.8555 0.9141 0.8633 0.9922 0.8438 0.9297 0.9336 0.9805 0.8750
Truck 0.9453 0.8281 0.9063 0.8594 0.9609 0.8516 0.9102 0.8945 0.9531 0.9844
Elaine 0.9336 0.8203 0.8750 0.8125 0.9688 0.8438 0.9141 0.8789 0.9453 0.8984
Boat 0.9258 0.7891 0.8555 0.7930 0.9570 0.8242 0.9102 0.8789 0.9297 0.8594
Barbara 0.9023 0.8125 0.8750 0.8125 0.9727 0.8164 0.8789 0.9023 0.9102 0.9766
Mandril 0.8984 0.8125 0.9219 0.8555 0.9336 0.8125 0.9297 0.9141 0.9336 0.8945
House 0.8711 0.7617 0.8086 0.7461 0.9258 0.9180 0.9375 0.8789 0.9688 0.8203
Jetplane 0.8867 0.8086 0.8516 0.7813 0.9102 0.8125 0.9492 0.8984 0.9336 0.8086
Livingroom 0.9258 0.8398 0.8945 0.8359 0.9727 0.8164 0.9648 0.9375 0.9297 0.8398
Woman 0.9531 0.8633 0.8828 0.8438 0.9766 0.8203 0.9883 0.8789 0.9688 0.8555

Table 4
NC values of extracted watermark for different attacks over fifteen images.

Images Attacks
Tr(0, 10) Tr(10, 0) Tr(2, 15) Scl(0.5) Scl(0.9) Scl(1.5) Rt(50 ) Rt(100 ) Rt(150 ) Cr(10%) Cr(20%) Cr(50%)

Lena 0.7148 0.7617 0.7891 0.9453 0.9844 0.9961 0.7305 0.6758 0.6680 0.9922 0.9609 0.8789
Goldhill 0.7188 0.7930 0.7852 0.9570 0.9766 0.9844 0.7148 0.7188 0.6680 0.9805 0.9570 0.8398
Peppers 0.7539 0.7695 0.7773 0.9414 0.9648 0.9766 0.7227 0.6836 0.6445 0.9727 0.9336 0.8477
Man 0.7930 0.7813 0.7461 0.9453 0.9648 0.9766 0.7383 0.7031 0.6992 0.9727 0.9688 0.8164
Airport 0.7891 0.7188 0.8086 0.9531 0.9688 0.9805 0.7695 0.7617 0.7031 0.9844 0.9492 0.8555
Tank 0.6875 0.7695 0.7500 0.9883 0.9922 0.9961 0.6797 0.6289 0.6680 0.9883 0.9648 0.8984
Truck 0.8086 0.8203 0.7930 0.9531 0.9531 0.9648 0.7500 0.7461 0.7031 0.9570 0.9453 0.8633
Elaine 0.7422 0.7539 0.7891 0.9531 0.9922 0.9961 0.7266 0.6602 0.6680 0.9844 0.9688 0.8789
Boat 0.7617 0.8203 0.7539 0.9219 0.9688 0.9766 0.7813 0.7070 0.6914 0.9844 0.9688 0.9102
Barbara 0.7422 0.7773 0.8008 0.9453 0.9805 0.9883 0.7344 0.7109 0.6797 0.9805 0.9492 0.8398
Mandril 0.8125 0.8164 0.8086 0.9531 0.9531 0.9492 0.7422 0.7383 0.6602 0.9609 0.9414 0.8555
House 0.6367 0.6758 0.7148 0.8828 0.9180 0.9297 0.6289 0.6484 0.6484 0.9258 0.9414 0.7695
Jetplane 0.7227 0.7500 0.7539 0.9297 0.9453 0.9727 0.6836 0.6953 0.6758 0.9570 0.9394 0.8477
Livingroom 0.7852 0.7930 0.7539 0.9453 0.9570 0.9609 0.7422 0.7539 0.6758 0.9727 0.9492 0.8672
Woman 0.7148 0.7422 0.7461 0.9688 0.9805 0.9961 0.6367 0.6602 0.6367 0.9883 0.9844 0.8594

X
N  
where T 1 is the total number of classes correctly classified by the N Nn
machine, and T 2 is the total number of classes. In case of different Pfp ¼ P ð1  PE Þn ð22Þ
n¼dððþ1Þ=2ÞNe
n E
attack conditions, Table 1 shows the performance of machine in
terms of CR.
The watermark detection can be analyzed on the basis of com- where PE is the probability that Wði; jÞ – W 0 ði; jÞ, and it is reasonable
parison of NC value with the detection threshold value . If NC P , to assume P E ¼ 0:5. For example, N ¼ 512;  ¼ 0:23, and PE ¼ 0:5,
we say that the extracted watermark is detected successfully; the estimated value of Pfp is 1:03  107 . In this paper, we set
otherwise, it is not. The probability of false positive error Pfp is  ¼ 0:23. Now, we say that the watermark is detected successfully,
incorporated to decide  [22,23], if NC P 0:23.
 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85 83

4.1. Imperceptibility
To analyze the performance of the proposed technique in terms
of imperceptibility, we have executed the watermark
embedding algorithm on different test images. In case of no attack,
Fig. 6 shows watermarked version of ‘‘Lena’’ image with
PSNR = 41.5107 dB, NC = 1, and BER = 0. Here, one can observe that
watermarked image is not distinguishable from their original ones,
which demonstrate good imperceptibility of the proposed water-
marking technique.
We have also analyzed the performance balance between
imperceptibility and robustness. A larger value of embedding
threshold (T) will increase the watermark strength, but decrease
the PSNR, and vice versa. On the other hand, given a fixed T or
watermark strength, the more watermark bits are embedded, the
lower PSNR value is, and vice versa. The relationship between
the PSNR and these two factors are shown in Fig. 7. In case of dif-
ferent attack conditions, the performance of imperceptibility vs
robustness is shown in Table 1.
In proposed method, the total embedding capacity is 1024 bits,
in which 512 bits are reserved for signature information and
remaining 512 bits for reference information. In our experiments,
on selecting the embedding threshold T ¼ 11, maximum robust-
ness is observed. At this value of T, the PSNR values of fifteen
watermarked images are graphically presented in Fig. 8, and the
average PSNR turns out to be 39.27 dB, which is quite reasonable.
4.2. Robustness
To investigate the robustness of the proposed method, effect of
various attacks on watermarked images is studied. The attacks
include both geometric and non-geometric types are listed as fol-
lows: JPEG compression with given quality factor (JPEG(QF)),
median filtering (Md), average filter (Avg), Gaussian filter (GF),
sharpening (Shp), histogram equalization (HE), Gaussian noise
(Gs), salt and pepper noise (Slp), speckle noise (Spn), translation
(Tr), scaling (Scl), cropping (Cr), and rotation (Rt). Table 1 shows
experimental results obtained on standard ‘‘Lena’’ image. Here,
one can observe that the proposed method outperforms in various
attack conditions and extracted watermarks are identifiable in
every case.
This robustness test is also performed over different images and
results are shown in Tables 2–4. Table 2 shows the performance
results in terms of NC values for different JPEG attacks over fifteen
images. Here, it is observed that for all images with JPEG ðQF > 50Þ
attack, the obtained NC ¼ 1, that means the extracted watermark is
exactly same as the original one. While, in case of various geomet-
ric and non-geometric attack conditions, the performance results
in terms of NC values are given in Tables 3 and 4.
4.3. Security analysis
The security requirements for a watermarking system vary
greatly from application to application. Some watermarking appli-
cations must have the ability to embed, detect, and/or remove
watermarks, whereas others must be restricted from performing
some, or all of these actions. Therefore, secure watermarks are
required to enforce these restrictions. To make the system more
secure, two level security approach is incorporated. Firstly, it is
provided by two different keys ðkey1; key2Þ to randomize the coef-
ficients of LH3 subband and blocks. Secondly, it is controlled by the
third key ðkey3Þ to randomize watermark bits.
Comparisons of the proposed technique with existing tech-
niques [21,27,30–34] using NC and BER as parameters of robust-
ness are shown in Figs. 9–11, Tables 5, and 6. The NC values and
the BER values in every comparison show that the proposed tech-
nique surmounted the existing techniques in different attack con-
ditions. By analyzing these comparisons, it is observed that the
proposed technique can resist maximum number of signal

Fig. 9. The comparison results with [21], for common image processing operations
(BER).

Fig. 11. The comparison results with [30], for common image processing operations
(BER).

Table 5
The comparison results with [31,32], for common image processing operations (BER).

Attacks [31] [32] Proposed

JPEG(50) 0.1992 0.3625 0
JPEG(90) 0.0490 0.2712 0
Filter(33) 0.4043 0.0675 0.0371
Slp 0.0693 0.0350 0.0430
Gs 0.0381  0.0840
Scl(50%)  0.0605 0.0273
Scl(150%) 0.0410 0 0.0020
Fig. 10. The comparison results with [27], for common image processing operations Rt(50 )  0.0963 0.1348
(NC).
84 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85

Table 6
The comparison results with [33,34], for common image processing operations (BER).

Attacks Images
Lena Mandril Barbara
[33] [34] Proposed [33] [34] Proposed [33] [34] Proposed
Md(33) 0.0276 0.0469 0.0371 0.3140 0.0723 0.0508 0.0422 0.0491 0.0488
GF(33) 0.0117 0.0388 0.0215 0.1802 0.0435 0.0332 0.0188 0.0352 0.0137
Avg(33) 0.0413 0.0740 0.0684 0.3706 0.1079 0.0391 0.0669 0.0713 0.0625
Gs(0.05) 0.0061 0.0540 0.0840 0.0471 0.0474 0.0938 0.0068 0.0542 0.0918
Slp(0.01) 0.0081 0.0369 0.0430 0.0684 0.0491 0.0352 0.0166 0.0400 0.0605
HE 0.3354 0.5435 0.0332 0.4946 0.6226 0.0527 0.5044 0.7170 0.0117
Scl(0.5) 0.0906 0.1372 0.0273 0.4783 0.1479 0.0234 0.0911 0.1016 0.0273
JPEG(30) 0.0903 0.1328 0.0195 0.3806 0.0601 0.0234 0.1011 0.1113 0.0117
JPEG(40) 0.0681 0.0925 0.0156 0.3467 0.0571 0.0195 0.0796 0.0867 0.0156
JPEG(50) 0.0466 0.0769 0.0098 0.2917 0.0479 0.0117 0.0588 0.0681 0.0098
JPEG(70) 0.0242 0.0459 0 0.1714 0.0403 0 0.0256 0.0466 0
JPEG(90) 0.0063 0.0342 0 0.0432 0.0393 0 0.0071 0.0369 0
JPEG(70)+Gs(0.05) 0.1267 0.0864 0.0938 0.1858 0.0437 0.0703 0.0310 0.0815 0.0781
JPEG(50)+Avg(33) 0.0847 0.1125 0.0645 0.4055 0.1135 0.0332 0.1008 0.1055 0.0586
JPEG(70)+Md(33) 0.0435 0.0657 0.0332 0.3350 0.0859 0.0449 0.0603 0.0618 0.0605
Rt(100 )+Scl(0.9) 0.0747 0.0498 0.1543 0.3293 0.1287 0.1504 0.0977 0.0869 0.1660
Rt(50 )+JPEG(50) 0.0746 0.0798 0.1328 0.3201 0.0750 0.1230 0.0833 0.0781 0.1270
Scl(0.9)+JPEG(70) 0.0835 0.0630 0.0117 0.3591 0.0667 0.0273 0.0925 0.0601 0.0156
Rt(150 )+Tr(2,15) 0.0186 0.0374 0.1680 0.2151 0.1267 0.1543 0.0605 0.0793 0.1426

processing attacks effectively, especially JPEG compression. Thus,
the simulation results show that proposed technique provides bet-
ter solution for copyright protection under different conditions.
5. Conclusions
Combining lifting wavelet transforms and support vector
machine with principal component analysis, we present a new
watermark extraction scheme. It successfully incorporates the
classification technique based on SVM with PCA based features
reduction to extract the watermark with maximum possible corre-
lation. The energy compaction property of LWT and good learning
ability of SVM can tolerate more image distortions. Well estab-
lished statistical parameters and PCA based dimensionality reduc-
tion makes the algorithm more simpler. Using different keys to
randomize coefficients, blocks, and watermark bits enhances the
security of system. The broad exploratory works have demon-
strated that the proposed scheme has conquered those challenging
attacks such as JPEG compression, translation, scaling, cropping,
rotation, and some common image processing operations. The
experimental results show noteworthy work of the proposed tech-
nique contrasted with the comparable existing techniques under
diverse signal processing attacks. However, further improvements
are required in case of rotation and translation operations.
Acknowledgments
The authors would like to thank the supportive remarks and
valuable suggestions from the anonymous reviewers.
References
[1] N. Nikolaidis, I. Pitas, Robust image watermarking in the spatial domain, Signal
Process. 66 (3) (1998) 385–403.
[2] X. Gao, L. An, X. Li, D. Tao, Reversibility improved lossless data hiding, Signal
Process. 89 (10) (2009) 2053–2065.
[3] L. Luo, Z. Chen, M. Chen, X. Zeng, Z. Xiong, Reversible image watermarking
using interpolation technique, IEEE Trans. Inform. Forensics Sec. 5 (1) (2010)
187–193.
[4] H. Luo, F.-X. Yu, H. Chen, Z.-L. Huang, H. Li, P.-H. Wang, Reversible data hiding
based on block median preservation, Inform. Sci. 181 (2) (2011) 308–328.
[5] X. Gao, L. An, Y. Yuan, D. Tao, X. Li, Lossless data embedding using generalized
statistical quantity histogram, IEEE Trans. Circ. Syst. Video Technol. 21 (8)
(2011) 1061–1070.
[6] L. An, X. Gao, Y. Yuan, D. Tao, Robust lossless data hiding using clustering and
statistical quantity histogram, Neurocomputing 77 (1) (2012) 1–11.
[7] L. An, X. Gao, Y. Yuan, D. Tao, C. Deng, F. Ji, Content-adaptive reliable robust
lossless data embedding, Neurocomputing 79 (0) (2012) 1–11.
[8] L. An, X. Gao, X. Li, D. Tao, C. Deng, J. Li, Robust reversible watermarking via
clustering and enhanced pixel-wise masking, IEEE Trans. Image Process. 21 (8)
(2012) 3598–3611.
[9] X. Li, J. Li, B. Li, B. Yang, High-fidelity reversible data hiding scheme based on
pixel-value-ordering and prediction-error expansion, Signal Process. 93 (1)
(2013) 198–205.
[10] X.-T. Wang, C.-C. Chang, T.-S. Nguyen, M.-C. Li, Reversible data hiding for high
quality images exploiting interpolation and direction order mechanism, Digital
Signal Process. 23 (2) (2013) 569–577.
[11] I.J. Cox, J. Kilian, F. Leighton, T. Shamoon, Secure spread spectrum
watermarking for multimedia, IEEE Trans. Image Process. 6 (12) (1997)
1673–1687.
[12] A. Bohra, O. Farooq, Izharuddin, Blind self-authentication of images for robust
watermarking using integer wavelet transform, {AEU} – Int. J. Electron.
Commun. 63 (8) (2009) 703–707.
[13] D.Y. Chen, O. Ming, J.L. Wu, A shift-resisting public watermark system for
protecting image processing software, IEEE Trans. Consumer Electron. 46 (3)
(2000) 404–414.
[14] M.-S. Hsieh, D.-C. Tseng, Y.-H. Huang, Hiding digital watermarks using
multiresolution wavelet transform, IEEE Trans. Ind. Electron. 48 (5) (2001)
875–882.
[15] S.-H. Wang, Y.-P. Lin, Wavelet tree quantization for copyright protection
watermarking, IEEE Trans. Image Process. 13 (2) (2004) 154–165.
[16] W.-H. Lin, S.-J. Horng, T.-W. Kao, P. Fan, C.-L. Lee, Y. Pan, An efficient
watermarking method based on significant difference of wavelet coefficient
quantization, IEEE Trans. Multimedia 10 (5) (2008) 746–757.
[17] X. Wang, X. Li, B. Yang, Z. Guo, Efficient generalized integer transform
for reversible watermarking, IEEE Signal Process. Lett. 17 (6) (2010) 567–
570.
[18] B. Lei, I.Y. Soon, F. Zhou, Z. Li, H. Lei, A robust audio watermarking scheme
based on lifting wavelet transform and singular value decomposition, Signal
Process. 92 (9) (2012) 1985–2001.
[19] Q. Gu, T. Gao, A novel reversible robust watermarking algorithm based on
chaotic system, Digital Signal Process. 23 (1) (2013) 213–217.
[20] V.S. Verma, R.K. Jha, Improved watermarking technique based on significant
difference of lifting wavelet coefficients, Signal, Image Video Process. (2014)
1–8.
[21] Y. Fu, Robust oblivious image watermarking scheme based on coefficient
relation, Optik – Int. J. Light Electr. Opt. 124 (6) (2013) 517–521.
[22] C. Deng, X. Gao, X. Li, D. Tao, A local Tchebichef moments-based robust image
watermarking, Signal Process. 89 (8) (2009) 1531–1539.
[23] X. Gao, C. Deng, X. Li, D. Tao, Geometric distortion insensitive image
watermarking in affine covariant regions, IEEE Trans. Syst., Man, Cybernet.,
Part C: Appl. Rev. 40 (3) (2010) 278–286.
[24] P.-T. Yu, H.-H. Tsai, J.-S. Lin, Digital watermarking based on neural networks
for color images, Signal Process. 81 (3) (2001) 663–671.
[25] C.-S. Shieh, H.-C. Huang, F.-H. Wang, J.-S. Pan, Genetic watermarking based
on transform-domain techniques, Pattern Recognit. 37 (3) (2004) 555–
565.
[26] C.-H. Chang, Z. Ye, M. Zhang, Fuzzy-art based adaptive digital watermarking
scheme, IEEE Trans. Circ. Syst. Video Technol. 15 (1) (2005) 65–81.
 V.S. Verma et al. / J. Vis. Commun. Image R. 31 (2015) 75–85
[27] K. Ramanjaneyulu, K. Rajarajeswari, Wavelet-based oblivious image
watermarking scheme using genetic algorithm, IET Image Process. 6 (4)
(2012) 364–- 373.
[28] H.-H. Tsai, D.-W. Sun, Color image watermark extraction based on support
vector machines, Inform. Sci. 177 (2) (2007) 550–569.
[29] X.-Y. Wang, H.-Y. Yang, C.-Y. Cui, An SVM-based robust digital image
watermarking against desynchronization attacks, Signal Process. 88 (9)
(2008) 2193–2205.
[30] H. Peng, J. Wang, W. Wang, Image watermarking method in multiwavelet
domain based on support vector machines, J. Syst. Softw. 83 (8) (2010) 1470–
1477.
[31] F. Wang, W. Song, J. Hou, Regression of SVM based robust watermarking
algorithm, in: Int. Conf. Signal Process., 2008, pp. 2197–2200.
[32] R. Shen, Y. gang Fu, H. tao Lu, A novel image watermarking scheme based on
support vector regression, J. Syst. Softw. 78 (1) (2005) 1–8.
[33] X.-Y. Wang, E.-N. Miao, H.-Y. Yang, A new SVM-based image watermarking
using Gaussian–Hermite moments, Appl. Soft Comput. 12 (2) (2012) 887–903.
85
[34] H.-Y. Yang, X.-Y. Wang, C. peng Wang, A robust digital watermarking
algorithm in undecimated discrete wavelet transform domain, Comput.
Electr. Eng. 39 (3) (2013) 893–906.
[35] W. Sweldens, The lifting scheme a construction of second generation wavelets,
SIAM J. Math. Anal. 29 (2) (1998) 511–546.
[36] W. Sweldens, The lifting scheme: a custom-design construction of
biorthogonal wavelets, Appl. Comput. Harmon. Anal. 3 (2) (1996) 186–200.
[37] R. Claypoole, R. Baraniuk, R. Nowak, Adaptive wavelet transforms via lifting,
in: IEEE Int. Conf. Acoustics, Speech and Signal Process., vol. 3, 1998, pp. 1513–
1516.
[38] J. Jackson, A User’s Guide to Principal Components, John Wiley and Sons, New
York, 1991.
[39] I. Jolliffe, Principal Component Analysis, Springer-Verlag, 1986.
[40] K. Mardia, J. Kent, J. Bibby, Multivariate Analysis, Academic Press, New York,
1979.
[41] F.A.P. Petitcolas, Weakness of Existing Watermarking Schemes, 1997. <http://
www.petitcolas.net/fabien/watermarking/stirmark/index.html>.