Information Sciences 301 (2015) 44–60

Contents lists available at ScienceDirect

Information Sciences
journal homepage: www.elsevier.com/locate/ins

An image watermarking scheme in wavelet domain with

optimized compensation of singular value decomposition via
artificial bee colony
Musrrat Ali a, Chang Wook Ahn a,⇑, Millie Pant b, Patrick Siarry c
a
Department of Computer Engineering, Sungkyunkwan University, Suwon 440746, Republic of Korea
b
Department of Applied Science and Engineering, IIT Roorkee 247667, India
c
Laboratoire Images, Signaux et Systèmes Intelligents (LiSSi, EA 3956), Université Paris-Est Créteil Val de Marne, 61 avenue du Général de Gaulle,
94010 Créteil, France

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

Article history: Digital image watermarking is the process of authenticating a digital image by embedding
Received 4 July 2014 a watermark into it and thereby protecting the image from copyright infringement. This
Received in revised form 11 December 2014 paper proposes a novel robust image watermarking scheme developed in the wavelet
Accepted 27 December 2014
domain based on the singular value decomposition (SVD) and artificial bee colony (ABC)
Available online 3 January 2015
algorithm. The host image is transformed into an invariant wavelet domain by applying
redistributed invariant wavelet transform, subsequently the low frequency sub-band of
Keywords:
wavelet transformed image is segmented into non-overlapping blocks. The most suitable
Invariant wavelet transform
Artificial bee colony
embedding blocks are selected using the human visual system for the watermark embed-
Image watermarking ding. The watermark bits are embedded into the target blocks by modifying the first col-
Optimization umn coefficients of the left singular vector matrix of SVD decomposition with the help
Entropy of a threshold and the visible distortion caused by the embedding is compensated by mod-
ifying the coefficients of the right singular vector matrix employing compensation param-
eters. Furthermore, ABC is employed to obtain the optimized threshold and compensation
parameters. Experimental results, compared with the related existing schemes, demon-
strated that the proposed scheme not only possesses the strong robustness against image
manipulation attacks, but also, is comparable to other schemes in term of visual quality.
Ó 2015 Elsevier Inc. All rights reserved.

1. Introduction

The expansion of Internet and computer networks have made the digital media (e.g. audio, video, image, etc.) acquisition
and distribution very easy these days. The digital contents having the same quality as that of the original can be created eas-
ily with the help of advanced multimedia technologies. This sometimes leads to crucial critical issues of misuse of digital
content leading to severe consequences. To overcome this problem, digital watermarking has been introduced as a safeguard
for the copyright protection [13,39]. Numerous efforts have been devoted to the advancement of the image watermarking
schemes [3,7,8,11,19,54,59,63]. Watermarking schemes can be categorized as per their characteristic of implementation
and the requirements for the extraction of the watermark, a detailed review is provided in [25,29].

⇑ Corresponding author.
E-mail addresses: musrrat.iitr@gmail.com (M. Ali), cwan@skku.edu (C.W. Ahn).

http://dx.doi.org/10.1016/j.ins.2014.12.042
0020-0255/Ó 2015 Elsevier Inc. All rights reserved.
 M. Ali et al. / Information Sciences 301 (2015) 44–60 45

The prime objective of a robust watermarking scheme is to preserve the quality of the watermarked image (i.e. imper-
ceptibility) while protecting the watermark when the watermarked image undergoes to intentional or unintentional threats.
Therefore, a robust watermarking scheme is the one in which the watermark stays intact with the content of the host image
when it is subjected to different distortions or attacks like rotation, row or column flipping, scaling, and translation. Several
image watermarking schemes are proposed in the literature, improving these two features by taking the advantages of dif-
ferent transforms, such as the discrete Fourier transform (DFT), discrete fractional Fourier transform (DFrFT), discrete wave-
let transform (DWT), discrete cosine transform (DCT), singular value decomposition (SVD), and their combinations
[3,5,13,14,16,18,23,32,33,43,44,49,52,63,66]. In [60] a feature-based image watermarking technique by using multi-scale
Harris detector and wavelet moment theory is proposed. The digital watermark is embedded in the local feature regions
of the host image by modifying the selected wavelet moment invariants. An improved dual color image watermarking
scheme based on SVD employing the concept of the relationship between the second and third elements of the first column
of the eigenvectors has been introduced by Su et al. [55]. All the aforementioned DWT watermarking methods are not invari-
ant to rotation and flipping attacks applied to the watermarked image. Li et al. [35] proposed a redistributed invariant dis-
crete wavelet transform (RIDWT) domain image watermarking. The invariant wavelet domain is obtained by relocating the
pixels’ position to new one, then applied Haar wavelet transform and image normalization. This transform is invariant to the
rotation of multiple of 90°, row and column flipping. Excited by its performance in terms of robustness and imperceptibility
authors have recommended its application in other wavelet based watermarking schemes.
In the last two decades, artificial intelligence (AI) techniques such as evolutionary algorithms (EAs), support vector
machine, fuzzy logic and neural networks have played an important role in real life applications, including image water-
marking [1,3,10,12,24,41,42,56,58,59,62,64] for improving the performance. Under the category of evolutionary algorithms,
genetic algorithm (GA), particle swarm optimization (PSO), differential evolution (DE), and Firefly algorithm have made
numerous valuable contributions to the field of watermarking. An image watermarking technique which uses GA to find
the optimal scaling factors for watermark insertion is designed by Lai [31]. In [47], the introduced technique is making
use of a simple genetic algorithm in order to optimize the set of parameters for moments that significantly influences the
locality properties alongside with the overall performance of the watermarking procedure. In [51,61], PSO is used to find
the scaling factors and thresholds respectively for watermark embedding. Applications of DE algorithm for finding the opti-
mal parameters for image watermarking can be found in [3,5,6]. Recently, Agarwal et al. [1] implemented Firefly algorithm to
find the optimal values of multiple scaling factors (MSFs) for watermark embedding. A relatively new member of evolution-
ary algorithms, artificial bee colony (ABC), was introduced by Karaboga [27] in 2005. It simulates the intelligent foraging
behavior of a honey bee swarm. Numerical comparisons demonstrated that the performance of ABC algorithm is competitive
to other population-based algorithms with an advantage of employing fewer control parameters [28]. Due to its simplicity
and ease of implementation, ABC algorithm has captured much attention and has been applied to solve many practical opti-
mization problems [2,15,22,34,50]. It is worth pointing out that the ABC for the optimization problems in image processing
such as segmentation [22], and enhancement [15] has been implemented, where it has given a good performance. But to the
best of our knowledge, it has never been practiced in image watermarking.
The literature review of SVD based image watermarking reveals that, the watermarking schemes developed in spatial
domain or frequency domain, generally embed singular values of the watermark and the rest of the information is kept safe
for its extraction. Various researchers pointed out the false positive detection problem in most of the SVD-based algorithms
and proved that the verification watermark unreasonably can be extracted [4,20,21,36–38]. This creates an ambiguous sit-
uation, indicating the futility of such schemes for copyright protection purpose. To overcome this problem, numerous
researchers have proposed improved versions of SVD based image watermarking schemes. A robust image watermarking
scheme based on SVD that embeds the entire watermark is proposed by Mohammad et al. [45]. Run et al. [51] introduced
an image watermarking scheme employing SVD and embedding the principal component of the watermark. Particle swarm
optimization is applied to get the optimal scaling factors for embedding. It is based on the fact that SVD subspace (left and
right singular vectors) can preserve a significant amount of information about an image. Because different regions of an
image have different local features, so some visual models may be incorporated in finding the suitable embedding regions
to improve robustness while maintaining imperceptibility. Based on this concept, a blind SVD-based watermarking scheme
is presented in [9]. The host image is segmented into non-overlapping blocks of size 8  8, then the embedding blocks (most
textured) are selected depending upon the number of non-zero singular values. The watermark bits are embedded by mod-
ifying the coefficients in the first column of the left singular vector matrix of the target blocks. Lai [30] has introduced an
image watermarking scheme based on human visual system (HVS) and SVD. The embedding process of the scheme is same
as of [9], while the embedding blocks are selected based on the sum of visual and edge entropies. The scheme of Fan et al.
[17] is an advanced version of the scheme proposed by Chang et al. [9], that promoted the transparency of the scheme by
incorporating compensation operation. According to their scheme, the damage in the quality due to insertion of the water-
mark in the left singular vector matrix is compensated by modifying the right singular vector matrix. However, though these
SVD based watermarking schemes have solved the false positive detection problem, they are not robust against various
attacks and distortions. Moreover, in the watermarking scheme proposed by Fan et al. [17], the threshold and compensation
parameters are tuned and adjusted manually. The proper setting of these parameters for watermarking is more difficult than
expected, that generates significant optimization challenges.
In view of the above facts, this paper proposes a robust image watermarking scheme developed in an invariant wavelet
domain based on the singular value decomposition (SVD) and the artificial bee colony (ABC) algorithm. The RIDWT is applied
46 M. Ali et al. / Information Sciences 301 (2015) 44–60

to the host image to obtain an invariant wavelet domain, subsequently the low frequency sub-band of wavelet transformed
image is segmented into non-overlapping blocks. The most suitable blocks are selected using human visual system (sum of
visual and edge entropies) for the watermark embedding. The watermark is embedded into target blocks by modifying the
coefficients of singular vector matrices of SVD decomposition of the target blocks with the help of threshold and compen-
sation parameters which are optimized with the help of ABC algorithm. We have evaluated the performance of the proposed
scheme on eleven test images, and have compared it with the similar schemes available in the literature. Normalized cor-
relation and peak signal to noise ratio are the performance metrics employed for comparing the different schemes. Exper-
imental results demonstrated that the proposed scheme not only resist the attacks likely to distort the image but also, is
comparable in term of visual quality. The novelty of our proposed scheme may be highlighted as:

(a) The proposed scheme applies the redistributed invariant wavelet transform proposed by Li et al. [35] to obtain an
invariant wavelet domain.
(b) The embedding blocks are selected using HVS (sum of visual and edge entropies) for the watermark embedding
instead of random selection to obtain the better watermarked image with higher quality and robustness.
(c) It is clear from the survey of image watermarking literature that the performance of the schemes can be improved by
using SVD with the other transforms such as DWT, DCT, and DFT. Therefore, unlike the scheme [17], the proposed
scheme applies SVD in invariant wavelet transform domain.
(d) Instead of manually adjusting the threshold and compensation parameters, as in [17], we have obtained their opti-
mized value using ABC.
(e) It is worth pointing out that ABC that has been successfully implemented for solving complex optimization problems
in various fields is being applied for improving the image watermarking scheme for the first time in this paper to the
best of authors’ knowledge.

The rest of the paper is organized as follows. Section 2 gives the brief description of the concepts being used. The proposed
scheme is explained in detail in Section 3. Results are analyzed in Section 4. Finally, the concluding remarks and future
research directions are given in Section 5.

2. Review of algorithmic concepts being used

This section is dedicated to the brief description of the concepts being used to transform the host image into the fre-
quency domain and then the block selection for watermark embedding. Interested to know more in detail may refer to
the respective studies [27,30,35,39].

2.1. Entropy

Entropy is a basic concept of information theory [30]. It is described as a measure of the amount of disorder in a system. In
another way, it can be considered as the spread of states which a system can adopt. A low entropy system occupies a small
number of such states, while a high entropy system occupies a large number of states. For grayscale images, these states
correspond to the gray levels that which the individual pixels can adopt. It can be used to characterize the texture of the
image. According to Shannon’s definition, the entropy of a grayscale image is given by the following mathematical relation:

X
L1
E1 ¼  pi log pi ð1Þ
i¼0

where L signifies the number of gray levels and pi is the probability associated with gray level ‘‘i’’ such that 0 6 pi 6 1 and
PL1
i¼0 pi ¼ 1. It depends only on the probability distribution of the pixels intensities irrespective of their positions of occur-
rence. Therefore, the E1 is viewed as a global measure with respect to the image. It is pointed out in [40] that an exponential
form of entropy function can capture two dimensional (2D) spatial correlation of image in a better way in comparison to
conventional Shannon’s entropy. That is called edge entropy and uses the edge information of the image. The strength of
the edge of a pixel automatically considers the effect of neighboring pixel values, so the measure of edge entropy of the
image is dependent on the relative occurrence of these edge strengths irrespective of their position. The edge entropy is
defined as follows.

X
L1
E2 ¼ pi exp1pi ð2Þ
i¼0

where 1  pi denotes the ignorance or uncertainty of the pixel value.

 M. Ali et al. / Information Sciences 301 (2015) 44–60 47

2.2. Redistributed invariant discrete wavelet transform (RIDWT)

A successful watermarking scheme related to the copyright protection must be resistant to different forms of intentional
and unintentional attacks. Discrete wavelet transforms (DWT) [3], having numerous applications in different areas including
image processing and digital watermarking is one of the versatile mathematical transform [33,57,65]. Although most exist-
ing watermarking schemes based on DWT have survived common image processing attacks, but still there is a scope of
improvement for attacks like geometric distortions such as multiples of 90° rotation and image flipping. To overcome this
challenge, redistributed invariant wavelet transform is introduced by Li et al. [35]. It is based on the fact that a multiple
of 90° rotation and image flipping change only locations of the pixels in the image but their intensities are left unchanged.
According to it, pixels’ locations of the image are redistributed; then, the Haar wavelet transform and some normalized pro-
cedures are performed, and, finally, the invariant wavelet domain is obtained. According to its theory, the original image ðIÞ
of size M  N is divided into four (2  2) equal-sized sub-images and their respective mean ðlij Þ of intensities are calculated
 
and stored in a matrix form as: mean ¼
l11 l12 ; l P 0. With the help of these means a normalization matrix (NM) is
ij l21 l22
constructed as:
   
NM11 NM12 l11 þ l12 þ l21 þ l22 l11  l12 þ l21  l22
NM ¼ ¼
NM21 NM22 l11  l21 þ l12  l22 l11  l12 þ l22  l21
Then, the sign matrix (Sg) is constructed and is given by:
  
Sg 11 Sg 12 1 if NMij > 0
Sg ¼ ; Sg ij
Sg 21 Sg 22 1 if NMij < 0
Redistribute the original image ðIÞ, using the distribution relation given in (3) to obtain the redistributed image (RI).
8
> RIð2i  1; 2j  1Þ ¼ Iði; jÞ; 1 6 i 6 M=2; 1 6 j 6 N=2
>
>
>
>
< RIð2i  1; 2j  NÞ ¼ Iði; 3N=2  j þ 1Þ; 1 6 i 6 M=2; N=2 6 j 6 N
ð3Þ
>
> RIð2i  M; 2j  1Þ ¼ Ið3M=2  i þ 1; jÞ; M=2 6 i 6 M; 1 6 j 6 N=2
>
>
>
:
RIð2i  M; 2j  NÞ ¼ Ið3M=2  i þ 1; 3N=2  j þ 1Þ; M=2 6 i 6 M; N=2 6 j 6 N
Then, apply the one level Haar wavelet transform to the image RI that decompose it into four sub-bands, namely LL, HL, LH
 
DWT LL HL
and HH such that RI ! and multiply the sub-band of the transformed image by the Sign matrix, as given below
LH HH
and denote the resulted matrix as:
   
A11 A12 Sg 11  LL Sg 12  HL
A¼ ¼
A21 A22 Sg 21  LH Sg 22  HH
The matrix A is just the invariant wavelet domain if jNM 21 j < jNM 12 j, where the term j⁄j is the absolute value operator. Other-
!
AT11 AT21
wise, swap A12 with A21 and transpose each sub-band of A to get a refreshed matrix A, that is A ¼ where ATij
AT12 AT22
denotes the transpose matrix of Aij . The final matrix A is the invariant wavelet domain; that is invariant to multiples of
90° rotation and row or column flipping.

2.3. Singular value decomposition (SVD)

The singular value decomposition (SVD) [39], a very powerful and useful matrix decomposition, is based on a theory of
linear algebra of diagonalization of a rectangular matrix. It can be used as a multimedia tool. It is very useful, particularly in
the context of data analysis, dimension reducing transformations of images, satellite data, etc. It is also the method of choice
for solving most linear least-squares problems. To better expose the various relationships among the original data it trans-
forms the correlated variables into a set of uncorrelated ones. Several watermarking schemes based on SVD have been pub-
lished in the literature [9,17,23,30,39,51]. According to the theory, a rectangular matrix B of order M  N can be partitioned
into the product of three matrices; an orthogonal matrix U, a diagonal matrix S and the transpose of an orthogonal matrix V.
It can be represented mathematically in its component matrices as:
2 3 2 3 2 3 2
b1;1 b1;2 b1;N u1;1 u1;2 u1;M s1;1 0  0 v 1;1 v 1;2 v 1;N 3T
6 b2;1 b2;2 7 6 u2;1 u2;2 7 6 0 s2;2  0 7 6 v 2;1 v 2;2 7
6 7 6 7 6 7 6 7
B¼6
6 .. ..
7 ¼ USV T ¼ 6
7 6 .. ..
76
7 6 .. .. .. ..
76
7 6 .. ..
7
7
4 . . 5 4 . . 5 4 . . . . 5 4 . . 5
bM;1 bM;N uM;1 uM;M 0 0 0 sM;N v N;1 v N;N
ð4Þ
48 M. Ali et al. / Information Sciences 301 (2015) 44–60

with UU T ¼ IM and VV T ¼ IN ; where IM and IN are the identity matrices of order M and N respectively. The elements of the
diagonal matrix S from first to last row are in descending order; i.e. satisfy the relation (5), where r is the rank of the matrix.
s1;1 P s2;2 P . . . P sr;r > srþ1;rþ1 ¼ srþ2;rþ2 . . . ¼ sM;N ¼ 0 ð5Þ

2.4. Artificial bee colony (ABC)

A simple and robust population based optimization scheme, artificial bee colony (ABC), was introduced by Karaboga [27]
in 2005. It simulates the intelligent foraging behavior of a honey bee swarm. Numerical comparisons demonstrated that the
performance of ABC algorithm is competitive to other population-based algorithms with an advantage of employing fewer
control parameters [28]. Due to its simplicity and ease of implementation, ABC algorithm has captured much attention and
has been applied to solve many practical optimization problems [2,15,22,34,50]. In ABC algorithm, the position of a food
source represents a possible solution to the optimization problem and the nectar amount of a food source corresponds to
the quality (fitness) of the associated solution. The number of the employed bees or the onlooker bees is equal to the number
of solutions in the population. Just like other evolutionary algorithms (EAs), it starts with an initial population of solutions of
size SN (food source positions), each having dimension D; X i ¼ ðxi;1 ; . . . ; xi;D Þ; i ¼ 1; . . . ; SN, where the index i denotes the ith
solution of the population. After initialization, the population of the positions (solutions) is subjected to a repeated cycle of
the search process of the employed bees, the onlooker bees and the scout bees. The cycle is repeated until a termination cri-
teria satisfied. In a robust search process, exploration and exploitation processes must be carried out together. In the ABC
algorithm, while onlookers and employed bees carry out the exploitation process in the search space, the scouts control
the exploration process. At the initialization stage, a set of SN food sources is randomly selected between lower
X min ¼ ðxmin;1 ; . . . ; xmin;D Þ and upper X max ¼ ðxmax;1 ; . . . ; xmax;D Þ bounds by the bees using the Eq. (6).
xi;j ¼ xmin;j þ randð0; 1Þ  ðxmax;j  xmin;j Þ ð6Þ
where rand(0, 1) is a uniformly distributed random number between zero and one.
In ABC algorithm, each cycle of searching of new food sources consists three steps that are discussed here one by one:
Employed bee: An employed bee produces a new solution depending on the local information and compare the fitness
value of the new solution to its parent solution. If the fitness value of the new solution is higher than that of its current solu-
tion, then replace the current solution with new one, otherwise continue with the current solution. In order to produce a new
solution Y i ¼ ðyi;1 ; yi;2 ; . . . ; yi;D Þ corresponding to the current solution X i ¼ ðxi;1 ; xi;2 ; . . . ; xi;D Þ, the ABC uses the following
expression (7):
yi;j ¼ xi;j þ /i;j  ðxi;j  xk;j Þ ð7Þ

where k 2 f1; 2; . . . ; SNg and j 2 f1; 2; . . . ; Dg are randomly chosen indexes such that k has to be distinct from i. /i;j is a ran-
dom number between 1 and 1 that controls the production of new solutions in the neighborhood of xi;j and represents the
comparison of two food positions visually by a bee.
Onlooker bee: The employed bees share the nectar information (fitness) of the food sources and their position information
with the onlooker bees waiting on the dance area of the hive. An onlooker bee evaluates the nectar information taken from
all employed bees and chooses a food source with a probability related to its nectar amount. As the nectar amount of a food
source increases, the probability of that food source chosen also increases. After arriving at the selected area, it chooses a
new food source in the neighborhood of the one in the memory depending on visual information as in the case of employed
bees. Evaluate the nectar amount of the candidate source, if the nectar is higher than that of the previous one, the bee mem-
orizes the new position and forgets the old one otherwise retain the old one. An artificial onlooker bee chooses a food source
depending on the probability value associated with that food source, Pi , calculated by the following expression:
fit
Pi ¼ PSN i ð8Þ
i¼1 fit i

where fiti is the fitness value of the ith solution and SN is the number of food sources which is equal to the number of
employed bees or onlooker bees. The fitness used in Eq. (9) is calculated using the following equation:
(
1
f ðX i Þþ1
if f ðX i Þ P 0
fiti ¼ ð9Þ
1 þ absðf ðX i ÞÞ elsewhere

where f ðX i Þ is the objective function value at the ith food position.

Scout bee: In ABC, if a position cannot be improved further through a predetermined number of cycles, then that food
source is assumed to be abandoned. The value of a predetermined number of cycles is an important control parameter of
the ABC algorithm, which is called ‘‘limit’’ for abandonment. In ABC, this is simulated by producing a position randomly
and replacing it with the abandoned one. Assume that the abandoned source is X i and j 2 f1; 2; . . . ; Dg, then the scout dis-
covers a new food source using (6) to be replaced with X i .
These three steps are repeated through a predetermined number of cycles called maximum cycle number (MCN) or until a
termination criterion is satisfied.
 M. Ali et al. / Information Sciences 301 (2015) 44–60 49

3. Proposed algorithm

This section is dedicated to the explanation of each step of the proposed scheme which are as follows:

3.1. Embedding region selection

The aim of this section is to highlight the usefulness of the human visual system (visual entropy and edge entropy) in the
selection of appropriate non-overlapping blocks for watermark embedding that achieve imperceptibility and robustness
requirements of watermarking. There are the varieties of image processing applications such as segmentation, coding and
compression where entropy has been used extensively. In steganography, the security of the hidden data is measured using
the relative entropy distance between the cover and the watermarked data [26]. Data hiding capacity is primarily deter-
mined by the entropy of the host media [46,48]. Some of the existing methods embed watermark in the image without
addressing human visual system (HVS), while considering HVS, the final watermarked image will have better quality. The
HVS model can be used not only to measure the perceptibility of watermark once it is embedded, but can also be used to
control that perceptibility during the embedding process. Maity and Kundu [40] proposed spatial domain and low lose trans-
form domain watermarking using entropy as image characteristics. In this study, we considered the visual entropy and edge
entropy, as also used in [30,40], in order to select watermark embedding regions that satisfy the requirements of the water-
marking scheme. It is incorporated by calculating visual entropy and edge entropy of each block with the help of (1) and (2),
respectively. Summed up these two values of entropy corresponding to each block and then sorted in ascending order based
on the magnitude. The lowest value blocks equal to the number of watermark bits are selected for watermark insertion.

3.2. Embedding

In order to embed a watermark W of size n  n into a host image I of size M  N, we utilize the concepts that are discussed
above in Section 2. A functional block diagram of the watermark embedding system is illustrated in Fig. 1. Referring to Fig. 1,
the watermark embedding process includes the following steps:

Step 1. Apply the redistributed invariant discrete wavelet transform (RIDWT) to the host image I, which transform it to the
matrix A with components A11 ; A12 ; A21 , and A22 . Select A11 that is in the size M=2  N=2 and partition it into 4  4
non-overlapping blocks Bi;j ; i ¼ 1; 2; . . . ; ðM=2Þ=4; j ¼ 1; 2; . . . ; ðN=2Þ=4. Obviously the numbers of non-overlapping
blocks should be greater than or equal to the numbers of the watermark bits.

Fig. 1. The illustration of watermark embedding procedure.

50 M. Ali et al. / Information Sciences 301 (2015) 44–60

Step 2. Select n  n appropriate blocks out of ðM=2Þ=4  ðN=2Þ=4 blocks for watermark embedding by utilizing the HVS
characteristics.
Step 3. Apply SVD to the selected block Bi;j to decompose it into three matrices U i;j ; Si;j , and V i;j . Then the relationship
between the entries in the first column of the U i;j matrix is examined. The watermark is embedded by changing
the relation between the second ðu2;1 Þ and the third ðu3;1 Þ entries in the first column. If the embedded binary water-
mark bit wi;j is 1, then the magnitude of ðu2;1 Þ must be greater than ðu3;1 Þ by a threshold value ðTÞ; i.e. the condition
ju2;1 j  ju3;1 j P T must hold. If the embedded binary watermark bit wi;j is 0, then the magnitude of ðu2;1 Þ must be
lesser than ðu3;1 Þ by a threshold value ðTÞ; i.e. the condition ju3;1 j  ju2;1 j P T must hold. When these two conditions
are violated, the entries of u2;1 and u3;1 should be modified as u02;1 and u03;1 , respectively, based on the following rules
(10)–(12). Fig. 2 shows the procedure of embedding one-bit of the watermark. Corresponding to each bit of the
watermark, there may be one of the four situations illustrated in Fig. 2. The bars in black color represent the actual
values of the coefficients, blue is the mean of the coefficients, the solid red bar is the portion of the threshold that is
added and dashed red bar is the subtracted portion of the threshold to the mean to achieve the new coefficients that
satisfy the embedding condition.
l ¼ ðju2;1 j þ ju2;1 jÞ=2 ð10Þ
(
u02;1 ¼ signðu2;1 Þ  ðl þ T=2Þ
if wi;j ¼ 1 & ðju2;1 j  ju3;1 j < TÞ; then ð11Þ
u03;1 ¼ signðu3;1 Þ  ðl  T=2Þ

(
u02;1 ¼ signðu2;1 Þ  ðl  T=2Þ
if wi;j ¼ 0 & ðju3;1 j  ju2;1 j < TÞ; then ð12Þ
u03;1 ¼ signðu3;1 Þ  ðl þ T=2Þ

Modifying the values of coefficients u2;1 and u3;1 of U i;j matrix affect the elements of second and third rows of the block
Bi;j . Thus, according to the second note of [17], the quality loss due to this change can be compensated by modifying
the elements of the first column of the matrix V i;j by the values D1 ; D2 ; D3 , and D4 respectively. Compensation oper-
ation was applied conditionally, because it may cause larger changes than without the compensatory operation. If the
difference between the original and compensated blocks is less than the difference between the original and modified
block without compensation, the compensatory operation is further used; otherwise no modification is performed on
matrix V i;j . Unlike the watermarking scheme proposed in [17], parameters T; D1 ; D2 ; D3 , and D4 are optimized using
ABC algorithm.
Step 4. Apply inverse SVD and inverse RIDWT to obtain watermarked image Iw .

3.3. Extraction

If there is a dispute regarding the rightful ownership of the image, then there is a need of verification process that is done
by extracting the embedded watermark. The detailed steps of the watermark extraction are described as follows.

Step 1. Apply steps 1–2 of the embedding process on the watermarked image Iw .
Step 2. Apply SVD to the selected block Bi;j to decompose it into three matrices U i;j ; Si;j , and V i;j . Then the relationship
between the entries in the first column of the U i;j matrix is examined. One-bit wi;j of the estimated watermark
W  is extracted depending upon the relationship between the second ðu2;1 Þ and the third ðu3;1 Þ entries in the first
column using Eq. (13).

1 if ju2;1 j P ju3;1 j
wi;j ¼ ð13Þ
0 otherwise
By repeating the step 2, all watermark bits are extracted and the estimated watermark is constructed.

3.4. Application of the artificial bee colony algorithm in finding optimal parameters

Numerous researchers have dealt with solving the problem of image watermarking as an optimization problem. Imper-
ceptibility and robustness are interrelated and mutually contradictory properties. Therefore a tradeoff will be often obtained
between these according to the needs of the application. In our previous work [3,5,6] we applied DE to find the optimal scal-
ing factors for embedding a grayscale watermark in different domains (i.e DWT, DCT, etc.). The proposed watermarking
scheme uses the binary watermark for embedding in an invariant wavelet domain using a threshold. Also, unlike our pre-
vious schemes, it does not directly embed the watermark into the host image. It quantizes the coefficients of the transformed
host image depending upon bits of the watermark. From an optimization point of view, all of these schemes are similar. In
the proposed algorithm the selection of the threshold, which plays an important part in balancing the imperceptibility and
the robustness of the watermarking algorithm, is very important. The objective function should include these two terms to
fulfil the requirements of robust watermarking scheme. The threshold in the watermarking scheme determines the
 M. Ali et al. / Information Sciences 301 (2015) 44–60 51

(a)

(b)

(c)

(d)
Fig. 2. (a–d) Illustration of different situations to embed watermark bits. The black bars represent the actual values of the coefficients, blue is the mean, and
solid and dashed red bars are the addition and subtraction to the mean to get the new coefficients. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

watermark strength that controls the imperceptibility and robustness. Use of small threshold favors the invisibility of the
watermark, but the watermarked image is less robust to several common attacks. On the other hand, high threshold favors
the robustness, but the quality of watermarked image is unacceptable. In most of the existing watermarking schemes this
parameter has been tuned and adjusted manually. Also, the compensation parameters are not obtained in an optimal
52 M. Ali et al. / Information Sciences 301 (2015) 44–60

manner. The proper setting of these parameters for watermarking is more difficult than expected. Actually, the values of
these parameters are image dependent. Therefore, an efficient and powerful algorithm is required that can find the suitable
values of these parameters automatically depending upon the image and watermark. Therefore, here we apply the artificial
bee colony algorithm to automatically determine these values to achieve a better performance. The procedure is started, ini-
tializing the population generated by random number generator between the specified ranges. The watermark bits are
embedded in the host image blocks by modifying the coefficients of the singular vector matrices with the help of these
parameters.
The watermarked image is examined for several attacks, and from these corrupted watermarked images the watermarks
are extracted using the extraction process. The objective function is computed for each solution and then the solution cor-
responding to higher fitness values are selected for the next generation. Now start the main loop of ABC. Perform reproduc-
tion process to generate new solutions by employed, onlooker and scout bees then evaluate the fitness. Continue the process
until the termination criterion is satisfied. At the end of execution, we will obtain the optimum parameters. The solution
corresponding to the best fitness of the final generation is used for watermark embedding. The objective function which
is used to evaluate the appropriateness of each candidate solution has to consider image quality and robustness against
attacks. For the assessment of image quality numerous of the quality metrics have been proposed in the literature. One of
the widely used metric for the quality assessment is the PSNR (peak-signal-to-noise-ratio) that we have been used in this
study to measure the quality of the distorted image. The similarity of the extracted watermark image (W  Þ with its original
watermark image (WÞ is evaluated by the normalized correlation (NC) metric which is the measure of robustness of the
scheme. The higher NC value is the indication that the extracted watermark is quite similar to the original watermark.
The mathematical expression of the objective function is given as follows [47]:
!
1X N
Minimize f ¼ 10  jPSNR  PSNRtarget j þ 1  NC i ð14Þ
N i¼1

where N is the number of attacks encountered in the procedure, NC i is the normalized correlation of extracted watermark
ðW  Þ corresponding to the ith attack, and PSNRtarget is a desired PSNR value. The incorporation of the target PSNR transforms
the optimization to a constrained procedure in order to ensure a minimum of image quality that must be acquired. As the
fitness function value decreases, PSNR approaches PSNRtarget and mean of NC i approaches 1.0. The mathematical formulae
used for PSNR of image quality assessment and normalized correlation (NC) are given below [63]:
!
ðIMAX Þ2
PSNRðI; I Þ ¼ 10log10 Pn Pn  2
ð15Þ
1
nn i¼1 j¼1 ðIi;j  I i;j Þ
Pn Pn
 i¼1 j¼1 W i;j  W i;j
NCðW; W Þ ¼ ð16Þ
nn
where I and I stand for the original and the processed images; IMAX is the maximum possible intensity value of the image I,
for an 8-bit per pixel representation IMAX is 255; subscripts i and j denote the location of the pixel value in the respective
images;  denotes the exclusive –OR (XOR) operation and n is the height or width of the square image.

4. Experimental results

This section is dedicated to the performance evaluation of the proposed watermarking scheme and comparison with its
parent schemes proposed by Lai [30], Fan et al. [17], that have been explained briefly in the introduction. For the sake of
simplicity, we will call these HVS + SVD and SVD respectively. To make our scheme more convincing, it is also compared with
the schemes proposed in [16,32,52] that we will call FWT, FrFT, and DWTQ respectively. A summary of the comparison of the
schemes is given in Table 1. SVD embeds the watermark in spatial domain while all the remaining schemes and the proposed
scheme embed the watermark in the frequency domain. The spatial domain (SVD) scheme has higher capacity than other
schemes. The capacity of the proposed scheme is equal to HVS + SVD scheme, but it is lesser than SVD. In Table 1, ‘N/A’
in the row corresponding to the capacity represents that the capacities for these schemes are not calculated, because these
depend on watermark size. Eleven standard test images of size 512  512 are taken for the experiments that are given in
Fig. 3(a–k). We have collected them from some open image databases that are available through the Internet. A binary logo of

Table 1
Summarization illustration of the schemes.

Attribute SVD HVS + SVD FWT DWTQ FrFT Ours

Host image 512  512 512  512 512  512 512  512 512  512 512  512
Watermark 32  32 32  32 20  50 32  16 64  64 32  32
Visual logo Yes Yes Yes Yes Yes Yes
Spatial/frequency Spatial Frequency Frequency Frequency Frequency Frequency
Parameter value Manually Manually Manually Automatic Manually Automatic
Capacity (b/p) 0.0625 0.0156 N/A 0.0104 N/A 0.0156
 M. Ali et al. / Information Sciences 301 (2015) 44–60 53

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

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

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

Fig. 3. (a–k) Test image: Baboon, Clown, Couple, Houses, Kiel, Lena, Lighthouse, Man, Pepper, Sailboat, and Zelda respectively, (l) watermark image.

size 32  32 that is given in Fig. 3 (lÞ, is taken as the watermark to verify image copyright. All the schemes have considered
the same size of the host and watermark images for the experimental analysis. In order to validate the robustness of the pro-
posed scheme, several image manipulation attacks are applied to degrade the quality of the watermarked image (Iw Þ that are
briefly described in Table 2 with their indexes. All the schemes are coded in MATLAB and executed on a personal computer
(PC) with Core 2 Duo processor, 4 GB RAM and Windows 7. The parameter settings of ABC are, population (number of food
locations) size SN = 20, limit = 50, and maximum cycle number (MCN) 100 [2]. In order to justify the proposed approach,

Table 2
Description of the attacks used for the robustness analysis.

Index Attack’s description

0 No attack applied
1 Median filtering with window size 2  2
2 Rescaling 512 ? 256 ? 512
3 Gamma correction with gamma value 0.2
4 Anticlockwise rotation by 90°
5 Circular cropping with 256 radius and centered at the center of the image
6 Histogram equalization
7 Gaussian noise addition with mean zero and standard deviation 0.001
8 JPEG compression with quality factor 75
9 Salt and pepper noise with noise density 0.05
10 Deleted 20 rows and 20 columns from random locations
11 Gaussian low pass filter with window size 9  9
12 Sharpening
13 Contrast adjustment (enhanced 40%)
14 Flipping of rows
15 Flipping of columns
54 M. Ali et al. / Information Sciences 301 (2015) 44–60

results are compared to the schemes [16,17,30,32,52]. To run the schemes for comparison, we have taken the recommended
parameter settings from their respective studies. The best results among the comparison in Tables 3–5 are highlighted in
bold and the tie among the results is shown in italic.

4.1. Imperceptibility

For an invisible watermarking, the embedded watermark must be imperceptible. Imperceptibility or transparency relates
to the properties of the human visual system [8]. In general, if the original and the watermarked images are mutually

Table 3
PSNR between watermarked and original images (dB).

Image PSNRtarget SVD HVS + SVD FWT DWTQ FrFT Ours

Baboon 40 42.2943 37.7290 22.8836 44.6238 41.2227 40.0256
Clown 49 48.9664 47.8601 22.8102 41.1843 41.2973 49.0194
Couple 42 46.2274 41.3311 22.8192 43.1897 41.2221 42.0148
Houses 40 41.8111 35.3789 22.4779 38.9901 41.2623 40.1107
Kiel 42 42.9307 40.5057 22.8338 43.0604 41.2284 42.0380
Lena 44 48.3212 43.2054 22.7807 42.6139 41.2173 44.0207
Lighthouse 43 44.8060 41.7695 22.6162 42.8834 41.2171 42.9883
Man 45 45.4251 43.1866 21.9677 41.2490 41.3289 45.0227
Pepper 43 45.0615 42.0315 22.7679 43.7749 41.2274 43.0222
Sailboat 40 45.3672 33.9873 22.5829 41.2905 41.2165 40.0038
Zelda 45 50.2175 44.4398 22.8682 49.2493 41.2186 45.0180
Average 45.5844 41.0386 22.6735 42.919 41.2417 43.0258

Table 4
Comparison of the schemes in terms of NC value of extracted watermark obtained from the watermarked images without any distortion attacks.

Image SVD HVS + SVD FWT DWTQ FrFT Ours

Baboon 1.0000 1.0000 0.9082 0.8936 1.0000 1.0000
Clown 0.9268 0.8066 0.9268 0.7734 1.0000 0.9941
Couple 1.0000 1.0000 0.9219 0.8262 1.0000 1.0000
Houses 1.0000 0.9971 0.9248 0.8965 0.9688 1.0000
Kiel 1.0000 1.0000 0.9111 0.8047 1.0000 1.0000
Lena 1.0000 1.0000 0.9326 0.7920 1.0000 1.0000
Lighthouse 1.0000 1.0000 0.9209 0.8594 1.0000 1.0000
Man 0.9854 0.9521 0.9355 0.8848 1.0000 0.9922
Pepper 0.9951 1.0000 0.9336 0.8184 1.0000 1.0000
Sailboat 1.0000 1.0000 0.9258 0.8545 1.0000 1.0000
Zelda 0.9990 0.9980 0.9375 0.8105 1.0000 1.0000
Average 0.9915 0.9776 0.9253 0.8376 0.9972 0.9988

Table 5
Comparison of the schemes in terms of average NC value obtained by taking over the eleven host images corresponding to the different distortion attacks.

Image SVD HVS + SVD FWT DWTQ FrFT Ours

0 0.9915 0.9776 0.9253 0.8376 0.9972 0.9988
1 0.7435 0.9423 0.8167 0.5832 0.6236 0.9076
2 0.7312 0.9518 0.6963 0.6978 0.5531 0.9134
3 0.9835 0.9727 0.8494 0.6808 0.5247 0.9973
4 0.4983 0.4987 0.4975 0.4656 0.4775 0.9988
5 0.8690 0.8311 0.8916 0.7893 0.7781 0.8490
6 0.9782 0.9731 0.9510 0.7108 0.9454 0.9982
7 0.9031 0.9384 0.9279 0.6832 0.9912 0.9830
8 0.8389 0.9579 0.9039 0.8065 0.7476 0.9574
9 0.8398 0.8017 0.8881 0.5365 0.8249 0.8104
10 0.9250 0.8714 0.8936 0.6408 0.8579 0.9347
11 0.9466 0.9705 0.9197 0.7699 0.9481 0.9964
12 0.9610 0.9579 0.9761 0.6813 0.9966 0.9972
13 0.8754 0.7303 0.9305 0.7006 0.8631 0.9443
14 0.4807 0.4822 0.5028 0.5286 0.4719 0.9988
15 0.5150 0.5566 0.4982 0.4758 0.4755 0.9988

Average 0.8175 0.8384 0.8168 0.6618 0.7548 0.9552

 M. Ali et al. / Information Sciences 301 (2015) 44–60 55

identical, the watermarking technique is said to be imperceptible. In other words, a good watermarking scheme causes no
artifacts or quality loss in the images. Common trend for analyzing imperceptibility is through objective and subjective
methods, is utilized in the present study as well. As reported in the literature, imperceptibility through objective process
is carried out by considering peak signal to noise ratio (PSNR). The PSNR values obtained by the watermarking schemes
are given in Table 3. In this paper, the threshold is selected, to ensure that the PSNR value is higher than the target PSNR
(PSNRtarget). From this Table it is clear that the proposed scheme provides the satisfactory results, in every case the PSNR
value is higher than 40 dB. SVD scheme provides the best results in most of the cases. While the performance of FWT is very
poor in the comparison. On the other hand, the subjective process takes advantage of the tonal distribution comparison
between original and watermarked images. The best way to obtain the tonal distribution of an image is the histogram. By
looking at the histogram for a specific image, the viewer can judge the entire tonal distribution at a glance. It tells how much
detail is at the light end of the spectrum and how much at the dark end. Therefore, here we have compared the histogram of
the original and watermarked images to show the imperceptibility, for the brevity of space we have given only one, corre-
sponding to the image ‘Man’ for the illustration in Fig. 4. From this figure, it is clear that the shape of the histograms of the
original and watermarked images is similar which further proves the imperceptibility of the proposed scheme.

4.2. Robustness against attacks

This section evaluates the robustness of the proposed scheme against the attacks that are given in Table 2. The quality of
the extracted watermark is determined by calculating the NC value using (16) and the corresponding results are tabulated in
Tables 4 and 5. A higher NC value indicates that the extracted watermark is more similar to the original watermark. Table 4
shows the results of the watermark extracted from the images that have not been subjected to any distortion attacks. From
the Table, it is clear that the SVD, HSV + SVD, FrFT, and the proposed schemes have extracted the watermark exactly same as
embedded in seven (63%), seven (63%), ten (90%), and nine (81%) cases respectively out of eleven cases. In the rest of the case
it is not exactly same, slightly different from the embedded watermark that is supported from the NC values given in Table 4.
On an average of NC values of extracted watermark proposed scheme outperformed all the other schemes. The average of NC
values corresponding to each attack over the eleven test images is given in Table 5 for the robustness comparison of the
schemes. The HVS + SVD, FWT, FrFT and the proposed schemes obtained the results better the other schemes in three
(18%), two (12%), one (6%) and ten (62%) cases respectively out of sixteen cases. For the qualitative comparison of the
schemes, sample of the attacks applied to the watermarked image ‘Man’ is shown in Fig. 5 and extracted watermark images
are given in Fig. 6. From Fig. 6 it is clear that the proposed scheme extract the watermark very well in comparison to the
other schemes even though the watermarked image seriously distorted. In visual quality of retrieved watermark all the
schemes have given more or less similar results in all the cases except rotation, Gaussian noise and flipping, where the pro-
posed scheme has given a better performance in comparison to the others. Hence, the performance of the proposed scheme
is better than or comparable to the other scheme considered for the comparison.

4.3. Non parametric statistical analysis

In order to compare the performance of the schemes in a more rigorous and meaningful way, the comparisons have been
realized using the hypothesis testing through a non-parametric test named Wilcoxon’s signed rank test [53]. It is imple-
mented through a well-known statistical software package SPSS. As a null hypothesis, it is assumed that the median of dif-
ferences of two sample values is zero. Whereas, the alternative hypothesis is that there is a significant difference in the
median values of the two samples. The statistical significance test is conducted at the 5% level of significance. In total there
are ninety-six data samples, sixteen corresponding to the each scheme (i.e. proposed scheme and five other schemes). Two

6000
Host image
Watermarked image
5000

4000
Pixel's count

3000

2000

1000

0
0 50 100 150 200 250
Gray level

Fig. 4. Imperceptibility illustration through histograms taking image ‘Man’.

56 M. Ali et al. / Information Sciences 301 (2015) 44–60

Index: 0 Index: 1 Index: 2 Index: 3

Index: 4 Index: 5 Index: 6 Index: 7

Index: 8 Index: 9 Index: 10 Index: 11

Index: 12 Index: 13 Index: 14 Index: 15

Fig. 5. Visual representation of the attacks applied to the watermarked images.

samples are taken for the comparison at a time, one corresponding to proposed scheme and the other corresponding to some
other scheme being compared, so in total there are eighty pairs. To establish that the results of the proposed scheme are
statistically significant, based on the NC values, the outcomes of the test are presented in Table 6, where ‘R+’ is the number
of positive ranks, ‘R’ is the number of negative ranks, ‘S+’ and ‘S’ are the sum of ranks of absolute value of the difference
between the two test variables greater than zero and less than zero respectively. That is called the sum of positive and neg-
ative ranks respectively. We assign one of the two signs (+, and Þ based on the test outcome, where the ‘+’ sign indicates
that the proposed scheme is significantly better than the other scheme, while the ‘’ sign shows no significant difference
between the two schemes. Table 6 reports the p-values produced by the Wilcoxon’s signed rank test, the acceptance or rejec-
tion of the null hypothesis depends on it. From Table 6, the p-value less than or equal to 0.05 (5% significance level) is strong
evidence against the null hypothesis. It indicates that the median value of the normalized correlation values produced by
proposed scheme is statistically significant and has not occurred by chance. From the statistical results, we observe that
there is a significant difference between the values obtained by the proposed scheme and the other scheme. The proposed
scheme provides results better than the SVD, HVS + SVD, FWT, DWTQ, and FrFT schemes in 81%, 56%, 87%, 100%, and 75%
cases, respectively, in the rest of the cases it performed equally.
 M. Ali et al. / Information Sciences 301 (2015) 44–60 57

Index SVD HSV+SVD FWT DWTQ FrFT Ours

0

10

11

12

13

14

15

Fig. 6. Visual representation of extracted watermark images from image ‘Man’ by proposed scheme and the other related schemes.

4.4. False positive detection and security

As pointed out by many researchers in watermarking literature, the false positive detection problem arises in most of
the SVD-based algorithms. It happens due to only the singular value insertion of the watermark into the host image in
the embedding process. In the extraction process, singular values are extracted from the watermarked image and the
rest of the information is provided by the owner. That creates an ambiguous situation and the ownership problem can-
not be resolved in such a condition. One of the solutions to this problem is to embed and extract the entire watermark
instead of singular values of the watermark. It means, anyone who claims ownership has to extract the entire watermark
from the watermarked image. Since our scheme follows the same principle, insertion of entire watermark, which means
the final extracted watermark will be decided by the all watermark information bits, not by the partial watermark infor-
mation. The false-positive detection problem like in SVD-based watermarking scheme will not occur in case of our pro-
posed scheme.
The watermark security of the proposed scheme lies on the positions of embedding blocks 4  4. An attacker with mali-
cious intentions can attempt to know the positions of embedding blocks so that it can destroy the watermark. The positions
of these blocks do not directly correspond to the host image. These are the positions in a cell array which stores these blocks.
Depending upon the way of scanning, along diagonals, rows and columns, these blocks can be picked up from the image by at
least sixteen ways. Whereas, in a cell array these can be arranged by thirteen ways. The probability (say p) by which an
attacker can correctly guess the positions of embedding blocks in the watermarked image can be given by the following
expression.

p ¼ 1=ðR ways of pickingÞ  1=ðR ways of arrangingÞ ¼ ð1=16Þ  ð1=13Þ ¼ 0:48%

Therefore, it shows that the success rate is less than a half percent. The attacker will have to put in a lot of effort and time
before he is able to remove the watermark bits by knowing the exact location of the embedding blocks. So, it is very difficult
for the attacker to correlate these cell array positions of blocks with their actual positions in the watermarked image as it
does not know by which way the blocks have been picked from the image and how the blocks have been arranged in a cell
array.
58 M. Ali et al. / Information Sciences 301 (2015) 44–60

Table 6
Statistical comparison of the proposed scheme versus other schemes based on Wilcoxon’s test.

Index R+ R Tie S+ S p-value Sig. R+ R Tie S+ S p-value Sig.

SVD Vs Ours HVS + SVD Vs Ours
0 4 0 7 10.0 0.0 0.068  4 0 7 10.0 0.0 0.068 
1 11 0 0 66.0 0.0 0.003 + 2 9 0 14.0 52.0 0.091 
2 11 0 0 66.0 0.0 0.003 + 3 8 0 14.0 52.0 0.091 
3 11 0 0 66.0 0.0 0.003 + 6 2 3 33.0 3.0 0.035 +
4 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
5 3 8 0 20.0 46.0 0.248  5 6 0 38.0 28.0 0.657 
6 11 0 0 66.0 0.0 0.003 + 8 1 2 41.0 4.0 0.028 +
7 11 0 0 66.0 0.0 0.003 + 8 1 2 44.0 1.0 0.011 +
8 11 0 0 66.0 0.0 0.003 + 2 9 0 20.0 46.0 0.248 
9 3 8 0 11.0 55.0 0.050 + 5 6 0 31.0 35.0 0.859 
10 6 5 0 41.0 25.0 0.477  10 1 0 65.0 1.0 0.004 +
11 11 0 0 66.0 0.0 0.003 + 6 1 4 25.0 3.0 0.063 
12 11 0 0 66.0 0.0 0.003 + 10 1 0 65.0 1.0 0.004 +
13 10 1 0 55.0 11.0 0.050 + 11 0 0 66.0 0.0 0.003 +
14 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
15 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +

FWT Vs Ours DWTQ Vs Ours

0 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
1 9 2 0 62.0 4.0 0.010 + 11 0 0 66.0 0.0 0.003 +
2 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
3 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
4 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
5 3 8 0 11.0 55.0 0.051  10 1 0 62.0 4.0 0.010 +
6 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
7 10 1 0 64.0 2.0 0.006 + 11 0 0 66.0 0.0 0.003 +
8 10 1 0 63.0 3.0 0.008 + 11 0 0 66.0 0.0 0.003 +
9 1 10 0 2.0 64.0 0.006 + 11 0 0 66.0 0.0 0.003 +
10 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
11 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
12 11 0 0 66.0 10.0 0.003 + 11 0 0 66.0 0.0 0.003 +
13 10 1 0 55.0 11.0 0.051  10 1 0 65.0 1.0 0.004 +
14 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
15 11 0 0 66.0 0.0 0.003 + 11 0 0 66.0 0.0 0.003 +
Index R+ R Tie S+ S p-value Sig.
FrFT Vs Ours
0 1 2 8 3.0 3.0 1.000 
1 11 0 0 66.0 0.0 0.003 +
2 11 0 0 66.0 0.0 0.003 +
3 11 0 0 66.0 0.0 0.003 +
4 11 0 0 66.0 0.0 0.003 +
5 9 2 0 55.0 11.0 0.050 +
6 10 0 1 55.0 0.0 0.005 +
7 7 4 0 38.0 28.0 0.657 
8 10 1 0 64.0 2.0 0.006 +
9 5 6 0 22.0 44.0 0.328 
10 11 0 0 66.0 0.0 0.003 +
11 10 0 1 55.0 0.0 0.005 +
12 3 3 5 10.0 11.0 0.917 
13 10 1 0 55.0 11.0 0.050 +
14 11 0 0 66.0 0.0 0.003 +
15 11 0 0 66.0 0.0 0.003 +

5. Conclusions

The proposed work is motivated by the well-known theory that the quality damage due to insertion of watermark’s bit in
one singular vector matrix of a selected block may be compensated by modifying the entries of the other singular vector
matrix of the same block. The parameters modifying the singular vector matrices for insertion and compensation are opti-
mized through ABC. This paper provides another scope of improving image watermarking scheme by using ABC. The blocks
for insertion of watermark bits are selected using the human visual system. The whole process of watermarking takes place
in a wavelet domain that is invariant to rotation and flipping. The strength of the scheme against image manipulation attacks
is tested on a set of eleven test images and fifteen image manipulation attacks. The results are compared with the other sim-
ilar schemes pictorially as well as numerically. It is observed that the proposed scheme performs competitively in terms of
imperceptibility and robustness. Integration of transforms has improved the robustness of the proposed scheme in
 M. Ali et al. / Information Sciences 301 (2015) 44–60 59

comparison to the schemes considered for the comparison in most of the cases. However, we noticed that the resistance of
the proposed scheme towards the median filtering, rescaling, and salt and pepper noise attacks is still very limited. Future
course of action could be to make it more resistant to these attacks. This scheme is free from a false positive detection prob-
lem. Thus, the proposed scheme can be seen as a better alternative over the existing robust watermarking schemes in resolv-
ing the rightful ownership. Its performance on colored images and video will also make an interesting area of research which
is the authors plan to do in future. Also, the study of its mathematical properties and theoretical analysis may be considered
as a future research direction. Additionally, the proposed research can be considered as a potential step toward exploring the
scope of other naturally behaved optimization schemes in watermarking.

Acknowledgments

The authors would like to thank the anonymous referees for their critical and constructive comments that have greatly
improved the quality and readability of the paper. This work was supported by the National Research Foundation of Korea
(NRF) grant funded by the Korea government (MSIP) (NRF- 2012R1A2A2A01013735).

References

[1] C. Agarwal, A. Mishra, A. Sharma, P. Bedi, Optimized gray-scale image watermarking using DWT–SVD and Firefly algorithm, Expert Syst. Appl. 41
(2014) 7858–7867.
[2] B. Akay, A study on particle swarm optimization and artificial bee colony algorithms for multilevel thresholding, Appl. Soft Comput. 13 (2013) 3066–
3091.
[3] M. Ali, C.W. Ahn, An optimized watermarking technique based on self-adaptive DE in DWT–SVD transform domain, Signal Process. 94 (2014) 545–556.
[4] M. Ali, C.W. Ahn, Comments on ‘‘Optimized gray-scale image watermarking using DWT–SVD and Firefly algorithm’’, Expert Syst. Appl. 42 (2015) 2392–
2394.
[5] M. Ali, C.W. Ahn, M. Pant, A robust image watermarking technique using SVD and differential evolution in DCT domain, Opt. – Int. J. Light Electron Opt.
125 (2014) 428–434.
[6] M. Ali, C.W. Ahn, P. Siarry, Differential evolution algorithm for the selection of optimal scaling factors in image watermarking, Eng. Appl. Artif. Intell. 31
(2014) 15–26.
[7] S.H. Amiri, M. Jamzad, Robust watermarking against print and scan attack through efficient modeling algorithm, Signal Process. Image Commun. 29
(2014) 1181–1196.
[8] G. Bhatnagar, Q.M. Jonathan Wu, A new logo watermarking based on redundant fractional wavelet transform, Math. Comput. Model. 58 (2013) 204–
218.
[9] C.-C. Chang, P. Tsai, C.-C. Lin, SVD-based digital image watermarking scheme, Pattern Recognit. Lett. 26 (2005) 1577–1586.
[10] K.W. Chau, Application of a PSO-based neural network in analysis of outcomes of construction claims, Autom. Constr. 16 (2007) 642–646.
[11] B. Chen, G. Coatrieux, G. Chen, X. Sun, J.L. Coatrieux, H. Shu, Full 4-D quaternion discrete Fourier transform based watermarking for color images, Digit.
Signal Process. 28 (2014) 106–119.
[12] C. Cheng, K. Chau, Y. Sun, J. Lin, Long-term prediction of discharges in manwan reservoir using artificial neural network models, in: J. Wang, X.-F. Liao,
Z. Yi (Eds.), Adv. Neural Networks, Springer, Berlin, Heidelberg, 2005, pp. 1040–1045.
[13] I.J. Cox, J. Kilian, F.T. Leighton, T. Shamoon, Secure spread spectrum watermarking for multimedia, IEEE Trans. Image Process. 6 (1997) 1673–1687.
[14] S. Dadkhah, A. Abd Manaf, Y. Hori, A. Ella Hassanien, S. Sadeghi, An effective SVD-based image tampering detection and self-recovery using active
watermarking, Signal Process. Image Commun. 29 (2014) 1197–1210.
[15] A. Draa, A. Bouaziz, An artificial bee colony algorithm for image contrast enhancement, Swarm Evol. Comput. 16 (2014) 69–84.
[16] E.H. Elshazly, O.S. Faragallah, A.M. Abbas, M.A. Ashour, E.-S.M. El-Rabaie, H. Kazemian, et al, Robust and secure fractional wavelet image watermarking,
Signal, Image Video Process. (2014), http://dxdoi.org/10.1007/s11760-014-0684-x.
[17] M.-Q. Fan, H.-X. Wang, S.-K. Li, Restudy on SVD-based watermarking scheme, Appl. Math. Comput. 203 (2008) 926–930.
[18] L. Gao, L. Qi, Y. Wang, E. Chen, S. Yang, L. Guan, Rotation invariance in 2D-FRFT with application to digital image watermarking, J. Signal Process. Syst.
72 (2012) 133–148.
[19] Q. Gu, T. Gao, A novel reversible robust watermarking algorithm based on chaotic system, Digit. Signal Process. 23 (2013) 213–217.
[20] J.-M. Guo, H. Prasetyo, Security analyses of the watermarking scheme based on redundant discrete wavelet transform and singular value
decomposition, AEU – Int. J. Electron. Commun. 68 (2014) 816–834.
[21] J.-M. Guo, H. Prasetyo, False-positive-free SVD-based image watermarking, J. Vis. Commun. Image Represent. 25 (2014) 1149–1163.
[22] K. Hanbay, M.F. Talu, Segmentation of SAR images using improved artificial bee colony algorithm and neutrosophic set, Appl. Soft Comput. 21 (2014)
433–443.
[23] H.-T. Hu, L.-Y. Hsu, Exploring DWT–SVD–DCT feature parameters for robust multiple watermarking against JPEG and JPEG2000 compression, Comput.
Electr. Eng. (2014), http://dxdoi.org/10.1016/j.compeleceng.2014.08.001.
[24] Z.-K. Huang, K.-W. Chau, A new image thresholding method based on Gaussian mixture model, Appl. Math. Comput. 205 (2008) 899–907.
[25] E. Hussein, M.A. Belal, Digital watermarking techniques, applications and attacks applied to digital media: a survey, Int. J. Eng. Res. Technol. 1 (2012) 1–
8.
[26] B.G. Joachim J. Eggers, R. Bäuml, A communications approach to image steganography, in: Secur. Watermarking Multimed. Contents, vol. IV, 2002, pp.
1–12.
[27] D. Karaboga, An Idea Based on Honey Bee Swarm for Numerical Optimization, Tech. Rep. TR06, Erciyes Univ. Eng. Fac. Comput. Eng. Dep., 2005.
[28] D. Karaboga, B. Akay, A comparative study of Artificial Bee Colony algorithm, Appl. Math. Comput. 214 (2009) 108–132.
[29] A. Khan, A. Siddiqa, S. Munib, S.A. Malik, A recent survey of reversible watermarking techniques, Inf. Sci. (Ny) 279 (2014) 251–272.
[30] C.-C. Lai, An improved SVD-based watermarking scheme using human visual characteristics, Opt. Commun. 284 (2011) 938–944.
[31] C.-C. Lai, A digital watermarking scheme based on singular value decomposition and tiny genetic algorithm, Digit. Signal Process 21 (2011) 522–527.
[32] J. Lang, Z. Zhang, Blind digital watermarking method in the fractional Fourier transform domain, Opt. Lasers Eng. 53 (2014) 112–121.
[33] S.-H. Lee, DWT based coding DNA watermarking for DNA copyright protection, Inf. Sci. (Ny) 273 (2014) 263–286.
[34] B. Li, Y. Li, L. Gong, Protein secondary structure optimization using an improved artificial bee colony algorithm based on AB off-lattice model, Eng. Appl.
Artif. Intell. 27 (2014) 70–79.
[35] L. Li, H.-H. Xu, C.-C. Chang, Y.-Y. Ma, A novel image watermarking in redistributed invariant wavelet domain, J. Syst. Softw. 84 (2011) 923–929.
[36] H.-C. Ling, R.C.-W. Phan, S.-H. Heng, On an optimal robust digital image watermarking based on SVD using differential evolution algorithm, Opt.
Commun. 284 (2011) 4458–4459.
[37] H.-C. Ling, R.C.-W. Phan, S.-H. Heng, On the security of a hybrid watermarking algorithm based on singular value decomposition and Radon transform,
AEU – Int. J. Electron. Commun. 65 (2011) 958–960.
60 M. Ali et al. / Information Sciences 301 (2015) 44–60

[38] H.-C. Ling, R.C.-W. Phan, S.-H. Heng, Comment on ‘‘Robust blind image watermarking scheme based on redundant discrete wavelet transform and
singular value decomposition’’, AEU – Int. J. Electron. Commun. 67 (2013) 894–897.
[39] R. Liu, T. Tan, An SVD-based watermarking scheme for protecting rightful ownership, IEEE Trans. Multimed. 4 (2002) 121–128.
[40] S.P. Maity, M.K. Kundu, DHT domain digital watermarking with low loss in image informations, AEU – Int. J. Electron. Commun. 64 (2010) 243–257.
[41] S.P. Maity, S. Maity, J. Sil, C. Delpha, Collusion resilient spread spectrum watermarking in M-band wavelets using GA-fuzzy hybridization, J. Syst. Softw.
86 (2013) 47–59.
[42] S.P. Maity, S. Maity, J. Sil, C. Delpha, Perceptually adaptive MC-SS image watermarking using GA-NN hybridization in fading gain, Eng. Appl. Artif. Intell.
31 (2014) 3–14.
[43] N.M. Makbol, B.E. Khoo, Robust blind image watermarking scheme based on redundant discrete wavelet transform and singular value decomposition,
AEU – Int. J. Electron. Commun. 67 (2013) 102–112.
[44] N.M. Makbol, B.E. Khoo, A new robust and secure digital image watermarking scheme based on the integer wavelet transform and singular value
decomposition, Digit. Signal Process 33 (2014) 134–147.
[45] A.A. Mohammad, A. Alhaj, S. Shaltaf, An improved SVD-based watermarking scheme for protecting rightful ownership, Signal Process. 88 (2008) 2158–
2180.
[46] P. Moulin, J.A. O’Sullivan, Information-theoretic analysis of information hiding, IEEE Trans. Inf. Theory. 49 (2003) 563–593.
[47] G.A. Papakostas, E.D. Tsougenis, D.E. Koulouriotis, Moment-based local image watermarking via genetic optimization, Appl. Math. Comput. 227 (2014)
222–236.
[48] M. Ramkumar, A.N. Akansu, Capacity estimates for data hiding in compressed images, IEEE Trans. Image Process 10 (2001) 1252–1263.
[49] S. Rawat, B. Raman, A blind watermarking algorithm based on fractional Fourier transform and visual cryptography, Signal Process. 92 (2012) 1480–
1491.
[50] F.J. Rodriguez, M. Lozano, C. García-Martínez, J.D. González-Barrera, An artificial bee colony algorithm for the maximally diverse grouping problem, Inf.
Sci. (Ny) 230 (2013) 183–196.
[51] R.-S. Run, S.-J. Horng, J.-L. Lai, T.-W. Kao, R.-J. Chen, An improved SVD-based watermarking technique for copyright protection, Expert Syst. Appl. 39
(2012) 673–689.
[52] M.J. Sahraee, S. Ghofrani, A robust blind watermarking method using quantization of distance between wavelet coefficients, Signal, Image Video
Process 7 (2011) 799–807.
[53] D.J. Sheskin, Handbook of Parametric and Nonparametric Statistical Procedures, third Ed., 2003.
[54] Q. Su, Y. Niu, G. Wang, S. Jia, J. Yue, Color image blind watermarking scheme based on QR decomposition, Signal Process. 94 (2014) 219–235.
[55] Q. Su, Y. Niu, H. Zou, X. Liu, A blind dual color images watermarking based on singular value decomposition, Appl. Math. Comput. 219 (2013) 8455–
8466.
[56] R. Taormina, K. Chau, R. Sethi, Artificial neural network simulation of hourly groundwater levels in a coastal aquifer system of the Venice lagoon, Eng.
Appl. Artif. Intell. 25 (2012) 1670–1676.
[57] S. Tedmori, N. Al-Najdawi, Image cryptographic algorithm based on the Haar wavelet transform, Inf. Sci. (Ny) 269 (2014) 21–34.
[58] H.-H. Tsai, Y.-S. Lai, S.-C. Lo, A zero-watermark scheme with geometrical invariants using SVM and PSO against geometrical attacks for image
protection, J. Syst. Softw. 86 (2013) 335–348.
[59] H.-H. Tsai, H.-C. Tseng, Y.-S. Lai, Robust lossless image watermarking based on a-trimmed mean algorithm and support vector machine, J. Syst. Softw.
83 (2010) 1015–1028.
[60] X.-Y. Wang, Y.-P. Yang, H.-Y. Yang, Invariant image watermarking using multi-scale Harris detector and wavelet moments, Comput. Electr. Eng. 36
(2010) 31–44.
[61] Y.-R. Wang, W.-H. Lin, L. Yang, An intelligent watermarking method based on particle swarm optimization, Expert Syst. Appl. 38 (2011) 8024–8029.
[62] C.L. Wu, K.W. Chau, Y.S. Li, Predicting monthly streamflow using data-driven models coupled with data-preprocessing techniques, Water Resour. Res.
45 (2009) 1–23.
[63] X. Wu, W. Sun, Robust copyright protection scheme for digital images using overlapping DCT and SVD, Appl. Soft Comput. 13 (2013) 1170–1182.
[64] J. Zhang, K.-W. Chau, Multilayer Ensemble Pruning via Novel Multi-sub-swarm Particle Swarm Optimization, J. Univ. Comput. Sci. 15 (2009) 840–858.
[65] L. Zhao, Y. Jia, Transcale control for a class of discrete stochastic systems based on wavelet packet decomposition, Inf. Sci. (Ny) 296 (2015) 25–41.
[66] P.-P. Zheng, J. Feng, Z. Li, M. Zhou, A novel SVD and LS-SVM combination algorithm for blind watermarking, Neurocomputing 142 (2014) 520–528.