Information Sciences 269 (2014) 94–105

Contents lists available at ScienceDirect

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

Robust watermarking against geometric attacks using partial

calculation of radial moments and interval phase modulation
Sasan Golabi, Mohammad Sadegh Helfroush ⇑, Habibollah Danyali, Mehri Owjimehr
Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran

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

Article history: This paper presents a new image watermarking scheme which is blind and robust against
Received 25 December 2012 geometric attacks. The proposed algorithm is based on the features of radial moments. In
Received in revised form 13 July 2013 this algorithm the radial moments were first computed and then the watermark was
Accepted 19 November 2013
embedded into the differential phase of various blocks of the original image using a special
Available online 28 November 2013
type of PSK modulation called interval phase modulation (IPM). In order to embed the
watermark into the phase of moments, a particular method, namely the partial calculation
Keywords:
of moments was utilized for computation. The implementation results show the improve-
Blind watermarking
Robustness
ment in the robustness and quality of the watermarked images in comparison with other
Radial moment methods.
Interval phase modulation Ó 2013 Elsevier Inc. All rights reserved.
Partial calculation of moment

1. Introduction

The rapid growth of multimedia applications leads to an urgent need for adequate copyright protection techniques, espe-
cially for image data. Robust watermarking can be used to trace copies or to implement copyright protection methods. Some
watermarking algorithms are published in order to satisfy this robustness requirement [2,7,13–15,18].
For robust watermarking schemes, the embedded watermark must be robust against a variety of possible attacks. These
include robustness against signal processing attacks, such as filtering, additive noise, cryptographic, statistical and geometric
attacks. While many methods have a good performance against signal processing attacks, they lack robustness to geometric
transformations. Rotation and scaling attacks are considered more challenging than other attacks. This is due to the fact that
changing the image size or its orientation even slightly, could dramatically reduce the receiver ability to retrieve the water-
mark. It has been proven that even very small geometric distortions can prevent the detection of a watermark [12,16,19,22].
The existing methods that can resist geometric attacks are classified into the exhaustive search, invariant domain, embed-
ding template and feature-based methods. One concern in the exhaustive search is the computational cost in larger search
space. Invariant domain methods usually suffer from implementation issues and are vulnerable to cropping. The embedding
template-based techniques are vulnerable to template estimation attacks and cropping. By contrast, the feature-based
watermarking techniques use image-dependent features to represent invariant reference points for both embedding and
detection. They are resistant to various attacks, including cropping and random bending attacks (RBA) by binding the water-
mark synchronization with the image salient characteristics. These characteristics may be the whole image, some local re-
gions, or feature points. This group of the watermark synchronization techniques, also known as the second generation

⇑ Corresponding author.
E-mail addresses: s.golabi@sutech.ac.ir (S. Golabi), ms_helfroush@sutech.ac.ir (M. S. Helfroush), danyali@sutech.ac.ir (H. Danyali), m.owjimehr@
sutech.ac.ir (M. Owjimehr).

0020-0255/$ - see front matter Ó 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.ins.2013.11.020
 S. Golabi et al. / Information Sciences 269 (2014) 94–105 95

watermarking [22], has the worthwhile properties of invariance to noise, geometrical transformation and localization. This
group of techniques can be divided into three sub-categories: moment-based, histogram-based and feature point-based.
Histogram-based watermarking techniques [3] use histograms to solve the geometric invariance problem. The histogram
distribution of an image is approximately invariant under geometric attacks. Xiang et al. [19] propose an invariant image
watermarking in the low-frequency domain by using the histogram shape and mean in the Gaussian filtered low-frequency
component of images. The method proposed in [3] developed a geometrically robust image watermarking scheme by using a
histogram in a certain range to embed a watermark in circular regions centered on the Harris–Laplace feature points. The
major limitation of the histogram-based methods is their incapability to resist local transformation.
Feature point-based watermarking methods use the feature points to form local regions for the embedding and extracting
of the watermark. In [17], first, disks which represent the features of the image will be derived and 16 watermark bits will be
embedded in each of them by an indirect method. For example, the number of disks extracted for the image Lena is 8 and for
the pepper is 4. For watermark extraction, it is required that the exact location of the feature disks be identified. Due to com-
putational errors (the interpolation errors in the normalization process and the sampling errors in converting formulas from
continuous to discrete ones), the disks cannot be located accurately. Hence, this method is usually used only to detect the
presence of the watermark rather than extract it. In [4], two blind and non-blind watermarking methods are provided. In
the blind method, the desired watermark will be changed with respect to the normalized original image. Then an inverse
normalization operation is applied to the created watermark. Therefore, the watermark image and the host image will be
of the same size. Now, the watermark is added to the host image. Due to applying the normalization to the watermark image,
the normalization effects on the original image will be canceled. However, in these methods, watermark extraction involves
finding the exact position of regions and because of the interpolation and normalization errors, this could not be performed
accurately.
Due to their ability to represent the global properties of the image, moments have been used in many application fields of
image processing. Geometric moments are mainly used to capture global features of images. In [9], the absolute values of the
Zernike moments which are extracted from the normalized image are used for watermark embedding. These features are
invariant to geometric changes. The watermark will be changed and embedded into the host image using these features.
In [5], to achieve robustness against affine transformation, the watermark is embedded into a moment-based normalized
image. In [1,20], Zernike moments are used as geometrically robust image watermarks. Zhang et al. in [21] propose a geo-
metric invariant blind image watermarking by using invariant Tchebichef moments and independent component analysis
(ICA). In [10], a geometrically invariant watermarking method is proposed based on the orthogonal moments obtained by
polar harmonic transform (PHT). The magnitude of these moments is rotation invariant. However, they are not strictly trans-
lation and scale invariant. To achieve these properties, the image needs to be resized to a domain (x, y) e [1 1]  [1 1]
which increases the numerical and interpolation errors. Moreover, they have used direct modification of moments which
is less desirable in comparison with change indirectly the spatial pixels. A compensation image is proposed to solve the
reconstruction problem. Not only, it raises the complexity; but also, there is a limitation for the increasing of the PSNR.
Invariant moment-based methods have low computational costs and do not need the complete normalization of the im-
age. As the normalization process is one of the main sources of numerical errors, the features extracted by moments are more
accurate and robust against common geometric and signal processing attacks than other methods. However, moment-based
techniques are highly fragile against cropping attacks.
In this paper, a new robust watermarking scheme, which can effectively resist common geometric attacks, is proposed.
The phase of adaptive radial moments, which in this paper is computed in a unit circle, is used for watermark embedding.
According to the new phases after watermark embedding, pixels will be changed in the spatial domain. These phases are
robust against geometric attacks and therefore no normalization procedure is needed. Therefore, the computational cost
is low; moreover, the numerical accuracy will be high. Using a special type of phase modulation, called ‘‘interval phase mod-
ulation (IPM)’’, increases robustness against compression attacks. The results of implementation show the improved quality
of the watermarked image and more capacity, as compared with other common geometric invariant methods.
The rest of this paper is organized as follows: in Section 2, moments, particularly, radial moments and their invariant fea-
ture extraction will be reviewed. In Section 3, the watermark embedding and extraction procedures are explained. Imple-
mentation and results are described in Section 4 and finally Section 5 provides the conclusion for this work.

2. Moments

The most common moments, are geometric moments which can be calculated as:
ZZ
Mpq ¼ f ðx; yÞxp yq dxdy ð1Þ
R

where f(x, y) P 0 is a real bounded function with support on a finite region R; Mpq is the geometric moment of the order
(p + q). In the discrete case, for a digital image of size N  N we have:
X
N X
N
Mpq ¼ f ðx; yÞxp yq ð2Þ
x y
96 S. Golabi et al. / Information Sciences 269 (2014) 94–105

In the above formula, xpyq indicates the base of moment calculation, thus moment computation can be seen as the map-
ping of image onto these basis. In the pattern recognition applications, the most important moments are invariant moments.
These moments are actually those from which the features are extracted and these features are robust against some geomet-
ric changes. For example, the amplitude of Zernike and Legendre moments is rotation invariant. In general, rotation invariant
features can be achieved from any complex moments. To have translation and scale invariant features, the image normali-
zation can be used [8].
In this paper, the radial moments are used to achieve invariant features. For an image f(x, y), the radial moment Rpq of
order p and repetition q can be computed as:
Z 2pZ 1 pffiffiffiffiffiffiffi
Rpq ¼ r p eiqh f ðr cos h; r sin hÞrdrdh; i¼ 1 ð3Þ
0 0

where f(r, h) is the projection of the image on Polar coordinates over a unit disk, p = 1, 2, . . . , 1 and q can be any positive or
negative integer. According to the above equation, radial moments are defined in terms of polar coordinates (r, h) over a unit
disk. On the other hand, the intensity function of the image is defined in Cartesian coordinates (x, y). Therefore, an appro-
priate image mapping is necessary. There are two main approaches for coordinate mapping from Cartesian to Polar. In
the first approach, the square image is mapped onto a unit disk where the center of the image is assumed to be the origin
of coordinates. In this approach, all pixels outside the unit disk are removed, resulting in a loss of some image information. In
the second approach, the square image is mapped inside the unit disk where the center of the image is assumed to be the
coordinate origin. Since the second method shows no loss of information, in this paper the second square-to-circle mapping
approach is applied (Fig. 1).
In the discrete case, for a digital image of size N  N the integrals in the above formula are replaced by summations and
the image is normalized inside the unit disk using the second mapping transformation. Thus, the discrete radial moments are
defined as:
X
N1 X
N 1
Rpq ¼ kp r pjk eiqhjk f ðj; kÞ jrjk j 6 1 ð4Þ
j¼0 k¼0

where kp is the number of pixels which have |rjk| 6 1.

This equation has two sources of numerical and geometrical errors. The numerical error is due to approximating the inte-
gral by summations and the geometrical error is made by square to circular mapping transformation.
To avoid these errors, the method described in [8] is used. According to geometric moment calculations and with some
mathematical computations, radial moments can be expressed in terms of geometric moments:
q  pq  
pq
X2 X
q j
2 ð5Þ
Rpq ¼ ðiÞ M p2kj;2kþj
k¼0 j¼0
k j

It can be shown that the radial moments after a degree image rotation are changed as:

R0pq ¼ eiqa Rpq ð6Þ

where R0pq
is the radial moment of the rotated image. As can be seen from (6), the absolute value of radial moments is con-
stant under image rotation and therefore the amplitudes of radial moments can be defined as rotation invariant features.
jR0pq j ¼ jRpq j ð7Þ

Fig. 1. Mapping of square image inside the unit disk [8].

 S. Golabi et al. / Information Sciences 269 (2014) 94–105 97

However, as a large amount of image information is stored in the phase of moments by selecting only the amplitude, we
lose them. To provide radial coefficients with the rotation–invariant phase, a method that has been presented in [11] is used
to cancel the influence of the rotation on the phase of coefficients. This method has been performed by combining the phase
coefficients of different orders and repetitions to form the complex valued moments with rotation invariant phases:
n  lp m o
R00pq ¼ Rpq eiq\Rp0 ;1 p0 2 1; 3; 5; 7; . . . ; 2  max  1 ð8Þ
2
The phase of R00pq stays unchanged under the image rotation. Suppose that a is the rotation angle; R00pq and R00r
pq represent the
original radial feature and the rotated version, respectively, we have:

\R00pq ¼ \Rpq  q\Rp0 ;1 ð9Þ

After rotation, the phase of R00pq is computed using the following equations:
 
r r
\R00r 00
pq ¼ \Rpq  q \Rp0 ;1 ¼ ð\Rpq  qaÞ  qð\Rp0 ;1  aÞ ¼ \Rpq  q\Rp0 ;1 ¼ \Rpq ð10Þ

Therefore, the phase angle \R00r 00

pq of the rotated image is the same as the phase angle \Rpq of the non-rotated image. The value
of order p0 could be any odd number. However, as mentioned in [11], the coefficients with lower orders will be more robust
to noise and distortions and since R11 is zero for normalized images, by setting p0 = 3, (8) becomes:

R00pq ¼ Rpq eiq\R31 ð11Þ

As can be seen from Eq. (3), the scale changing of an image does not change the radial moment phase. Hence, by using
\R00pq , the features that are invariant against rotation and scaling are obtained. The translation invariance can usually be
achieved by using the central moments instead of the geometric ones:

X
N1 X
N1
lpq ¼ ðx  xÞp ðy  y
Þq f ðx; yÞ ð12Þ
x¼0 y¼0

where   ¼ M 01 =M 00 . Central moments are actually the moments which are invariant against image trans-
x ¼ M 10 =M 00 and y
lation. However, in the proposed method pixel distances from the image origin are used instead of the coordinate distances
and therefore, translation normalization can be avoided. As we know, normalization is a source of interpolation and numer-
ical errors. Therefore, the proposed method is invariant against rotation and scaling and translation changes and moreover,
the computational accuracy will be high.

3. The proposed method

Fig. 2 shows the diagram of the proposed watermark embedding scheme, which denotes embedding the watermark in
\R00pq . To set the phase of the invariant phase-radial moment to the desired values, the values of specified pixels are changed.
The approach to computing the phase of radial moments and changing the pixel values will be described in the following
subsections.

3.1. Computation of the phase of radial moments

In the proposed method, the phase of extracted features explained in Section 2 will be used for watermark embedding.
For this purpose, a method which is called here as ‘‘partial computation of moments (PCM)’’ will be used. This method
computes the radial moments by using geometric moments. Experimental results show that R42 results in the best invariant

Input Finding the phase of

Image Partial calculation of
image features
partitioning radial features

Watermark

Changing the phase using

Changing the pixels of
Watermarked Combining the interval modulation according
each partition to have
image partitions to the watermark bits
new phase

Fig. 2. Schematic diagram of watermark embedding.

98 S. Golabi et al. / Information Sciences 269 (2014) 94–105

feature and therefore, according to Eq. (8), we choose p = 4 and q = 2 as shown in Eq. (13). The radial moments are computed
as shown in Fig. 3.

R0042 ¼ R42 e2i\R31 ð13Þ

As can be seen in Fig. 3, the image matrix is divided into two matrices. In these matrices, gray blocks represent the zero
pixels and the white ones show the gray scale pixels. The goal of this conversion is computing, separately, the geometric
moments for four pixels which are symmetrical and have the same radial distance from the origin.
This figure shows that the geometric moment Mpq for the original matrix is the sum of Mpq of matrix 1 and Mpq of matrix 2.
Notice that Fig. 3 is only an example and there are various options for matrix partitioning; however, the pixels of one of them have
to be symmetric to the origin point. The desired watermark bits will be embedded into the phase of features. From (13) we have:

\R0042 ¼ \R42  2\R31 ð14Þ

And from Eq. (5) we can have:
R42 ¼ ðM 40  M04 Þ þ 2iðM31 þ M 13 Þ; R31 ¼ ðM 30 þ M12 Þ þ iðM 21 þ M 03 Þ ð15Þ
Now, for matrix 2, the phase computation for R0042 leads to:
   
2ðM 31 þ M13 Þ M 21 þ M03
\R0042 ¼ tg 1  2tg 1 ð16Þ
M 40  M04 M 30 þ M12
For more simplicity, the following notations are defined:

ðM40  M 04 Þ  A1; ðM31 þ M 13 Þ  A2

ð17Þ
ðM30 þ M 12 Þ  B1 ðM21 þ M 03 Þ  B2
To compute the geometric moments in matrix 1, the method presented by Hosny is used [8]. In this method as the pixels
have the same radial distance, for the computing of Mpq from Eq. (2) the numerical value of xpyq depends on the values of p
and q, whether they are even or odd.
Based on the results obtained in [8], the geometric moments can then be evaluated according to the following four cases:
Case 1: p and q are both even

Mpq ¼ bpq ðf ð1; 1Þ þ f ð1; 2Þ þ f ð2; 1Þ þ f ð2; 2ÞÞ ð18Þ

Case 2: p is even and q is odd

Mpq ¼ bpq ðf ð1; 1Þ  f ð1; 2Þ þ f ð2; 1Þ  f ð2; 2ÞÞ ð19Þ

Case 3: p is odd and q is even

Mpq ¼ bpq ðf ð1; 1Þ  f ð1; 2Þ þ f ð2; 1Þ þ f ð2; 2ÞÞ ð20Þ

Case 4: p and q are both odd

Mpq ¼ bpq ðf ð1; 1Þ þ f ð1; 2Þ þ f ð2; 1Þ  f ð2; 2ÞÞ ð21Þ

where f(i, j) is the intensity function of the pixel point (i, j), and bpq is a coefficient that depends on the distance of the pixels
from the origin. Since the pixels have the same radial distance from the origin, we have bpq = bqp. Before these calculations,
the original image defined in the square 512  512 is mapped to be inside the unit circle, where the coordinate origin is the
center of the circle. As a result, the value of bpq is too small.

f(1,1) f(1,2)
= +
f(2,1) f(2,2)

Initial matrix matrix1 matrix2

Fig. 3. Block partitioning for partial calculation of moments.

 S. Golabi et al. / Information Sciences 269 (2014) 94–105 99

Therefore, in matrix1 we have:

   
2ðf ð1; 1Þ þ f ð1; 2Þ þ f ð2; 1Þ  f ð2; 2ÞÞ f ð1; 1Þ  f ð1; 2Þ þ f ð2; 1Þ  f ð2; 2Þ
\R0042 ¼ tg 1  tg 1 ð22Þ
0 f ð1; 1Þ  f ð1; 2Þ þ f ð2; 1Þ þ f ð2; 2Þ
Eq. (22) can be expanded to m  m matrix shown in Fig. 4 as:
   
2ðf ð1; 1Þ þ f ð1; mÞ þ f ðm; 1Þ  f ðm; mÞÞ f ð1; 1Þ  f ð1; mÞ þ f ðm; 1Þ  f ðm; mÞ
\R0042 ¼ tg 1  2tg 1 ð23Þ
0 f ð1; 1Þ  f ð1; mÞ þ f ðm; 1Þ þ f ðm; mÞ
Finally, the phase of R0042 is computed as:

\R0042 ¼ \fA1 þ i2ðA2 þ b1 ½f ð1; 1Þ þ f ð1; mÞ þ f ðm; 1Þ  f ðm; mÞÞg  2\fB1 þ b2 ½f ð1; 1Þ  f ð1; mÞ þ f ðm; 1Þ
þ f ðm; mÞ þ iðB2 þ b2 ½f ð1; 1Þ  f ð1; mÞ þ f ðm; 1Þ  f ðm; mÞÞg ð24Þ
where b1 = b31 + b13 and b2 = b21 + b03.

3.2. Watermark embedding

In order to partition the original image for the calculating of the phase of radial moments, we generate a mask image that
has the same size as the original image with a 508  508 matrix of 0’s inside and the pixel values of the original image on
four sides in two rows and two columns corresponding to original image.
At first we consider a 4  4 matrix using the corresponding pixels of original image around the origin of the mask image
and embed a watermark bit by changing the radial feature of this matrix. The size of the inner matrix is enlarged by adding
two columns and two rows on both sides and construct a 8  8 matrix to embed the next bit. At each step, we add two
columns and two rows on each side of the previous inner matrix and construct a new one to embed the next watermark
bit into it. This process is continued to the final matrix with size 508  508. Steps of partitioning the host image are shown
in Fig. 5.
Using the phase for watermark embedding is the same as phase modulation. For some reasons, such as using summation
instead of integration and the created interpolation after image normalization, the values and features mentioned above will
accompany some errors. Consequently, in this paper we use a particular type of phase modulation called interval phase mod-
ulation (IPM) for more robustness against numerical errors. The phase partitioning diagram is illustrated in Fig. 6. In this fig-
ure, it can be observed that the phase of R0042 is first computed. Then the 0–360° range is divided into 30° intervals.
Corresponding to the embedding bit, the angle of R0042 , will change at its own interval as below:

 Embedding bit = 0
(
\R0042old 30i þ 5 6 \R0042old 6 30i þ 10 i ¼ f0; 1; . . . ; 11g
\R0042new ¼ ð25Þ
30i þ 7:5 otherwise

 Embedding bit
( = 1 00
\R42old 30i þ 20 6 \R0042old 6 30i þ 25 i ¼ f0; 1; . . . ; 11g
\R0042new ¼ ð26Þ
30i þ 22:5 otherwise

In the embedding process, according to Fig. 5 and Eq. (24), the value of each symmetric pixel can be modified to change
the phase of \R0042 . This can be considered as a key to embedding and extracting the watermark bits. In this paper, the value of
f(1, 1) and f(m, 1) are changed. In Eq. (24), it can be seen that k=2 changes in f(1, 1) and f(m, 1) only changes the last part of the
above equation by b2 k. As mentioned before, the value of b2 is too small and therefore a great change needs to be made in
pixel values to have a specified change in feature angle. This leads to great damage in image quality. In order to avoid this, we
add b2 k=2 instead of k=2, which has a small value and its effect on image quality becomes much weaker. For a desired phase
of R0042 , b2 k is given by:

f(1,1) f(1,m)

= +

f(m,1) f(m,m)

matrix1 matrix2

Fig. 4. Block partitioning for the partial calculation of moments in m  m matrix.

100 S. Golabi et al. / Information Sciences 269 (2014) 94–105

Fig. 5. Partitioning the original image. Gray color shows zero pixels and white ones represent original gray scale pixels.

Fig. 6. Interval phase modulation diagram.

   
\R42  \R0042 0
b2 k ¼ tg B 1  B0 2 ð27Þ
2
with
B0 1 ¼ B1 þ b2 ½f ð1; 1Þ  f ð1; mÞ þ f ðm; 1Þ þ f ðm; mÞ
B0 2 ¼ B2 þ b2 ½f ð1; 1Þ  f ð1; mÞ þ f ðm; 1Þ  f ðm; mÞ
The watermark embedding procedure:

(1) Partition the original image by using the generated mask image as mentioned in this section.
(2) At the first step of partitioning (4  4 matrix), calculate \R0042 according to Eq. (24).
(3) Examine the computed phase at the 30° intervals according to the embedding bit as mentioned in Eqs. (25) and (26)
and change this phase.
(4) Find b2 k according to the changes in phase and Eq. (27).
(5) Change the value of pixels f(1, 1) and f(4, 1) by b2 k=2.
(6) Perform steps (1)–(5) for partitioning at the next level (8  8 matrix) to the last level (508  508 matrix) to embed 127
bits.

3.3. Extraction process

The extraction process is formulated as follows:

(1) Construct a 512  512 binary matrix in which the first and the last two rows and columns are 1 and put a matrix of 1s
based on the steps of the embedding process (Fig. 5) at its center. This matrix is shown in Fig. 7a.
(2) Compute the rotation angle h based on the method which is shown in Fig. 7c and estimate the scaling factor s by com-
paring the size of the watermarked image and rotated image.
 S. Golabi et al. / Information Sciences 269 (2014) 94–105 101

y
2.5 1
1 1 x Rotated image
y
2.5 1 (rotation ) = tg −1
x

a b c
Fig. 7. Matrices of extraction process. White color represents 1 and gray color shows 0.

(3) Rotate and scale the constructed binary matrix by h and s respectively and multiplied the watermarked image by it, to
have the masked watermarked image.
(4) Construct a new 512  512 zero matrix but with different pixel values as shown in Fig. 7b and compute b2 by com-
puting M21 + M03 for this matrix. Rotate and scale it by the angle h and the factor s.
(5) Compare this rotated and scaled matrix with the result of Step (3) to find the location of f(1, 1) and f(m, 1) at each step.
The floating part of f(1, 1) and f(m, 1) is b2 k=2.
(6) Compute the value of k based on the result of steps (4) and (5).
(7) Add k to the pixel values f(1, 1) and f(m, 1) of the masked watermarked image and compute \R0042 of this matrix.
(8) Extract the watermark bit as:

(
0 30i < \R0042 6 30i þ 15 i ¼ f0; 1; . . . ; 11g
watermark bit ¼ ð28Þ
1 30i þ 15 < \R0042 6 30i þ 30 i ¼ f0; 1; . . . ; 11g

A schematic diagram of this process is illustrated in Fig. 8.

As can be seen from (24), \R42 remained unchanged by the proposed method and therefore, it was used directly to com-
pute b2 k: It is clear that various signal attacks change the pixel values and thus the angle and amplitude of moments will be
changed. However, these attacks affect the phase of moments much lower than their amplitudes. Also, these changes can be
reduced more using the proposed embedding method.

4. Implementation results

In this section, the performance of the proposed method against geometric transformation has been illustrated and eval-
uated. The proposed algorithm has been tested on different types of images of size 512  512, namely Lena, Baboon and Pep-
per as shown in Fig. 9(a1, b1, c1). A pseudorandom sequence of size 127-bits is used as the watermark pattern. The
performance of the proposed watermarking scheme is evaluated in two aspects: imperceptibility and robustness. The results
show that the proposed scheme can satisfy the requirements of both imperceptibility and robustness.
The original images and the watermarked images (Lena, Baboon and Peppers) obtained by the proposed algorithm are
shown in Fig. 9(a1, b1, c1) and (a2, b2, c2), respectively. The differences between the original images and the watermarked
versions are magnified by a factor 10,000 and are shown in Fig. 9(a3, b3, c3). Based on (27), it is clear that the proposed

Watermarked Image Partial calculation of Finding the phase of

image partitioning radial features features

No If the phase is greater Finding the interval

Watermark bit is 0 than interval means? which include the phase

Yes

Watermark bit is 1

Fig. 8. Schematic diagram of watermark extraction.

102 S. Golabi et al. / Information Sciences 269 (2014) 94–105

a1 a2 a3

b1 b2 b3

c1 c2 c3
Fig. 9. Originl and watermarked images, (a1) original image of Lena, (a2) watermarked image of Lena with watermark of size 127 bits, (a3) difference image
between (a1) and (a2). (b1) original image of baboon, (b2) watermarked image of baboon with watermark of size 127 bits, (b3) difference image
between (b1) and (b2). (c1) original image of pepper, (c2) watermarked image of pepper with watermark of size 127 bits, (c3) difference image between (c1)
and (c2).

method changes 127 pixel values of original image, less than 1. If intensities of 127 pixels of the original image are changed
by 1, PSNR will be greater than 80 dB. So the PSNR of watermarked image will be greater than this.
After extracting the watermark, the bit error rate (BER) is computed using the original watermark and the extracted
watermark to evaluate the correctness of an extracted watermark. Bit Error Rate for the embedded and extracted watermark
sequences of length N bits is listed as:

1X N
BERðW; W 0 Þ ¼ jWðiÞ  W 0 ðiÞj ð29Þ
N i¼1

where W and W 0 are the original watermark and the extracted watermark sequences.
 S. Golabi et al. / Information Sciences 269 (2014) 94–105 103

In order to demonstrate, the robustness of the proposed watermarking algorithm, the watermarked image is attacked by
some geometric attacks. After these attacks on the watermarked image the extracted watermark sequence is compared with
the original one. In the following sub-sections, two attacks are examined: rotation and scaling. Our method shows good per-
formance against these geometric attacks.

4.1. Rotation

Table 1 illustrates the robustness of the proposed method against the rotation attack for various degree rotations in terms
of BER. As can be seen, this method has a good performance by various rotations especially in small angels.

4.2. Scaling

There are two types of image scaling, increasing and decreasing the size of image. For both types the scaling factor is
examined and the watermark is extracted as described in the extraction sub-section. Table 2 shows the results for various
scaling factors from 60% scaling down to 160% scaling up in terms of BER. As the results illustrate, the watermarked images
show strong resistance to the scaling attack.

4.3. Rotation with scaling

Fig. 10(a1 and a2), both are the examples of the rotation of the watermarked image in Fig. 9(a2) by 34°, one with cropping
and the other without cropping. Experimental results show that the similarity measure for the cropping case is low. There-
fore, as expected, for Fig. 10(a1) the proposed method does not have a good performance, but for Fig. 10(a2), it is more effi-
cient. Fig. 10(a2) is rotation of the watermarked image by 34°, but to show all pixels, we apply image resizing by factor 1.39.
Thus, rotation without cropping will use scaling itself. However, in general, another scaling can be applied to the rotated
watermarked image. In this case, based on what has been mentioned in the extraction sub-section, the embedding locations

Table 1
BER of the proposed method for various rotation angles for test images.

Rotation angle 0° 1° 2° 5° 8° 15° 30° 45° 80° 140° 170° 180°

Lena 0 0 0 0 0 0.05 0 0.37 0.007 0.16 0.32 0.31
Baboon 0 0 0 0 0.007 0.039 0 0.38 0.33 0.025 0 0
Pepper 0 0 0 0 0.007 0.039 0.078 0.36 0.047 0.071 0.023 0

Table 2
BER of the proposed method for various scaling factors for test images.

Scaling factor 0.6 0.7 0.9 1.1 1.6

Lena 0.41 0.37 0.06 0.39 0.37
Baboon 0.44 0.38 0.46 0.39 0.38
Pepper 0.37 0.40 0.15 0.31 0.18

Fig. 10. (a1) Image of Lena after 34 degree rotation and with cropping, (a2) 34 degree rotation with scaling of Lena image without cropping.
104 S. Golabi et al. / Information Sciences 269 (2014) 94–105

Table 3
BER of the proposed method for various geometric attacks on various images.

Image Attack
1° rotation 5° rotation 1° rotation and 2 scaling 1° rotation and 5 scaling and shift the origin to (2, 3)
Lena 0 0 0.11 0.13
Baboon 0 0 0.09 0.12
Pepper 0 0 0.07 0.1

Table 4
Comparison between robustness of the proposed method and the method of [17] (BER).

Attack The proposed method Method of [16]

1° rotation with cropping
Lena 0.45 0.62
Baboon 0.44 0.72
Pepper 0.42 0.5
5° rotation with cropping
Lena 0.48 1
Baboon 0.49 1
Pepper 0.47 1
1.1 scaling
Lena 0.39 0.5
Baboon 0.39 0.54
Pepper 0.31 0.75

Table 5
Comparison between robustness of the proposed method and the method of [6] (BER).

Attack The proposed method Method of [6]

30° rotation
Lena 0 0.60
Baboon 0 0.73
Pepper 0.078 0.69
5° rotation
Lena 0.18 0.60
Baboon 0.35 0.66
Pepper 0.16 0.61
1.5 scaling
Lena 0.39 0.40
Baboon 0.39 0.53
Pepper 0.31 0.44
Lena 0.06 0.60
[0.7 0.9] scaling
Baboon 0.46 0.86
Pepper 0.15 0.73

are found and the watermark sequence is extracted. In Table 3, the robustness of the proposed method against various geo-
metric attacks is evaluated. Note that, all of the practical implementations are performed by the nearest interpolation.
In each case, the bit error rate criterion (BER) has been used to evaluate the correction of the extracted bits. The results
indicate the high performance of the proposed method.
Tables 4 and 5 show the robustness of the proposed method for various geometric attacks, as compared to other pub-
lished results in [17,6], respectively. As papers [17,6] have used the attacks mentioned in Tables 4 and 5, these algorithms
have not been simulated and the comparison has been made based on these attacks. As can be seen, the proposed method
provides better performance. Also, the comparison with the method described in [4] shows that the robustness results of the
proposed method for some geometric attacks are close to the results of the method of [4], but the quality and the capacity of
the proposed algorithm are higher than [4].

5. Conclusion

A blind and robust watermarking scheme was presented. This scheme is robust against geometric attacks and uses the
features of radial moments. At first, the invariant phase radial moments were computed, and the watermark was embedded
 S. Golabi et al. / Information Sciences 269 (2014) 94–105 105

into the differential phase of various blocks of the original image by use of a special type of phase modulation called interval
phase modulation. In order to embed the watermark into the phase of moments, a particular method called partial calcula-
tion of moments has been used. The implementation results show the improvement in the robustness and quality of water-
marked images in comparison with other methods. This method has some advantages in comparison with other common
methods: (1) This scheme does not use any kind of normalization, leading to the elimination of interpolation errors which
occur during RST (Rotation, Scale, Transition) normalization. (2) The proposed method has the ability to detect and extract
the watermark and since the interval modulation is used, the watermarking scheme has high robustness. (3) This algorithm
can be adopted according to the user necessity. In order to achieve more robustness or quality larger or smaller intervals are
used. (4) This method has the simple implementation and the ability to apply on various image formats. (5) The extraction
process does not need any preprocessing and the watermark bits are extracted directly from the attacked image. This prop-
erty increases the accuracy of the extraction process and decreases the computational errors.

References

[1] M. Alghoniemy, A.H. Tewfik, Geometric invariance in image watermarking, IEEE Trans. Image Process. 13 (2) (2004) 145–153.
[2] I. Cox, J. Kilian, T. Leighton, T. Shamoon, Secure spread spectrum watermarking for multimedia, IEEE Trans. Image Process. 6 (12) (1997) 1673–1687.
[3] C. Deng, X. Gao, X. Li, D. Tao, Local histogram based geometric invariant image watermarking, Signal Process. 90 (12) (2010) 3256–3264.
[4] P. Dong, J.G. Brankov, N.P. Galatsanos, Digital watermarking robust to geometric distortions, IEEE Trans. Image Process. 14 (12) (2005) 2140–2150.
[5] P. Dong, N. Galatsanos, Affine transformation resistant watermarking based on image normalization, IEEE Conf. Image Process. 3 (2002) 489–492.
[6] X. Gao, C. Deng, X. Li, D. Tao, Geometric distortion insensitive image watermarking in affine covariant regions, IEEE Trans. Syst., Man, Cybernet. 40 (3)
(2010) 278–286.
[7] F. Hartung, M. Kutter, Multimedia watermarking techniques, Proc. IEEE 87 (7) (1999) 1079–1107.
[8] K.M. Hosny, Fast and accurate method for radial moment’s computation, Pattern Recogn. Lett. 31 (2) (2010) 143–150.
[9] H.S. Kim, H.K. Lee, Invariant image watermark using Zernike moments, IEEE Trans. Circ. Syst. Video Technol. 13 (8) (2003) 766–775.
[10] L. Li, Sh. Li, A. Abraham, J.Sh. Pan, Geometrically invariant image watermarking sing polar harmonic transforms, Inform. Sci. 199 (2012) 1–19.
[11] S. Li, M.C. Lee, C.M. Pun, Complex Zernike moments features for shape-based image retrieval, IEEE Trans. Syst., Man, Cybernet. 39 (1) (2009) 227–237.
[12] C. Lin, M. Wu, J. Bloom, I. Cox, M. Miller, Y. Lui, Rotation, scale and translation resilient watermarking for images, IEEE Trans. Image Process. 10 (5)
(2001) 767–782.
[13] W. Lu, W. Sun, H. Lu, Robust watermarking based on DWT and nonnegative matrix factorization, Comput. Electr. Eng. 35 (1) (2009) 183–188.
[14] S. Mabtoul, E.I. Elhaj, D. Aboutajdine, Robust semi-blind digital image watermarking technique in DT-CWT domain, Int. J. Comput. Sci. (2009) 8–12.
[15] H. Munawer, A. Otum, N.A. Samara, A robust blind color image watermarking based on wavelet-tree bit host difference selection, J. Signal Process. 90
(8) (2010) 2498–2512.
[16] J. O’Ruanaidh, T. Pun, Rotation, scale, and translation invariant digital image watermarking, Signal Process. 66 (3) (1998) 303–317.
[17] C.W. Tang, H.M. Hang, A Feature-based robust digital image watermarking scheme, IEEE Trans. Signal Process. 51 (4) (2003) 950–959.
[18] W. Tzenghuang, S. Yentan, Y. Jenchang, A robust watermarking technique for copyright protection using discrete wavelet transform, WAEAS Trans.
Comput. 9 (5) (2010) 485–495.
[19] S. Xiang, H.J. Kim, J. Huang, Invariant image watermarking based on statistical features in the low-frequency domain, IEEE Trans. Circ. Syst. Video
Technol. 18 (6) (2008) 777–790.
[20] Y.Q. Xin, S. Liao, M. Pawlak, Geometrically robust image watermark via pseudo-Zernike moments, IEEE Conf. Electr. Comput. Eng. 2 (2004) 939–942.
[21] L. Zhang, G. Qian, W. Xiao, Z. Li, Geometric invariant blind image watermarking by invariant Tchebichef moments, Opt. Express 15 (5) (2007) 2251–
2261.
[22] D. Zheng, J. Zhao, LPM-Based RST invariant digital image watermarking, IEEE Trans. Circ. Syst. Video Technol. 3 (2003) 1951–1954.