44

CHAPTER 3

HYBRID ROBUST WATERMARKING

FOR COLOR IMAGES

3.1 INTRODUCTION

A digital image consists of a set of pixels, which can be

conveniently captured using any electronic device, such as a camera, scanner
or camcorder. With the increased use of the Internet and proliferation of
image capturing devices, the access and distribution of images has become
convenient and tremendously attainable. Yet, most of the access and
distribution may be illicit or illegal. The protection of ownership and
prevention of unauthorized manipulation of digital images become an
important issue. The image content is used for content sensitive domains such
as photojournalism, courtroom evidence, medical applications, or commercial
transaction, where the originator of the content has to be verified by ensuring
the integrity of the content. Hence, a secure authentication system is needed to
guarantee the owner of the image in situations where the credibility of an
image may be questioned.

Among the various methods that have been introduced to protect

the intellectual property rights of digital images, the commonly used method
is digital watermarking. Some types of digital data such as logos, labels or
names representing the author’s ownership are embedded into the desired host
image as watermarks. The registration to the authentication center is
 45

necessary, which helps to resolve ownership disputes by identifying the

embedded watermark in the host image.

Color images are more important in our daily life which can
provide more information than grayscale images. The color images can hide a
greater amount of data and attain higher fidelity because the color perception
depends on not only the luminance but also the chrominance. Over the last
few decades, numerous watermarking methods have been developed and
tested with the aim of providing reliable ways of proving image ownership for
gray scale images. Some of the gray level images watermarking schemes are
also extended to the color images. However, their application to color images
might not be completely inadequate since they do not take implication of
HVS into consideration and in particular, its sensitivity to color brightness
and perception (Lusson et al 2013).

This chapter focuses on the problem of efficient color image

watermarking by combining the advantages of both spatial and frequency
domains. Watermark is generated for each channel of the color image by
extracting the spatial domain features using Gray Level Co-occurrence Matrix
as well as a unique owner identification number. This watermark is embedded
in PCA-based less correlated low and high frequency sub bands in such a way
that the perceptual quality of the image is preserved. Imperceptibility is
achieved by embedding the watermark in less correlated sub bands and
robustness is achieved by spreading the watermarks using Laplacian Pyramid
in contourlet transform. Simulation results show that the proposed scheme can
survive various image processing and signal processing attacks. The scheme
achieves high transparency, imperceptibility and robustness compared to
some of the existing schemes (Lie et al 2006, Li et al 2009, Haohao 2009,
Rahimi & Rabbani 2011 and Ranjbar et al 2013) in the literature.
 46

3.2 RELATED WORK

The image watermarking techniques proposed so far can be divided

into two broad categories based on watermark embedding: spatial domain and
transform domain. Moreover, hybrid domain techniques which combine the
advantages of the both domains are also available. The following sub-sections
brief each domain methodology and explore their advantages and limitations.

3.2.1 Spatial domain

Embedding watermarks into the spatial domain components of the

cover images is a simple method. It is one of the fundamental schemes used
since the beginning of digital watermarking in 1993 (Tirkel et al 1993). The
spatial domain-based methods select a number of pixels from the cover image
and modify the luminance values according to the watermark bits to be
embedded (Van Schyndel et al 1994, Pitas 1996, Chan & Cheng 2004, Ni
et al 2008, Deng et al 2010). The image containing these modified pixels
carries the watermark information. The same pixels used in the embedding
procedure should be selected from the watermarked image to extract the
watermark. The hidden watermark can be extracted by collecting all the
watermark bits embedded. Some important methods of spatial domain
embedding are explained below:

Least Significant Bit (LSB): Many spatial domain techniques

embed the watermark information in the Least Significant Bit (LSB) of every
pixel (Rajendra Acharya et al 2001, Awrangjeb & Kankanhalli 2003, De
Vleeschouwer et al 2003). Each LSB scheme embeds the watermark in the
image with different algorithms utilizing the insensitivity feature of the
Human Visual System (HVS) to perform minor changes in color and
brightness. LSB planes are suitable locations to be replaced by random bits
such that the watermark bits may be removed. The binary representation of
 47

the watermark is computed and LSB of every byte within the cover image is
overwritten. For example, Rajendra Acharya et al (2001) embeds encrypted
version of Electronic Patient Record (EPR) by replacing the LSB of gray
levels. The text file containing the patient’s data is encrypted by the logarithm
of ASCII codes of the text, the graphical signals are compressed and
subsequently interleaved with images. They pointed out the reduction of
storage and transmission overheads as a result of inserting patient information
in images. Moreover, direct embedding in medical images is not secure. LSB
manipulation is a fast and relatively inexpensive way of watermarking, but
tends to be vulnerable to spatial changes resulting from any image processing
or lossy compression methods such as JPEG.

Histogram: Histogram equalization is used in image processing

tasks to adjust contrasts. The aim of this technique is to distribute intensity
values on the histogram in a better way. This allows the image areas of lower
local contrast to gain a higher contrast. Histogram equalization accomplishes
this task by effectively spreading out the most frequent intensity values.
Histogram based schemes are commonly used for image watermarking
(Hwang et al 2006, Chrysochos et al 2007, Ni et al 2008, Deng et al 2010).
Pre-defined histogram values are used to embed the watermark. For example,

Chrysochos et al (2007) proposed a blind algorithm with an

asymmetric key to embed the watermark into histogram values. The
histogram shape is preserved after embedding watermark and the algorithm is
robust against geometrical attacks such as rotation, flipping, translation,
aspect ratio changes, resizing, warping, shifting, drawing and scattered tiles,
as well as their combinations. But, watermark embedding strategy could be
easily detected by comparing the histogram shape of the original image with
the watermarked image because of the existence of non-linear relationship
between the histogram representation and pixel representation. Moreover, this
 48

scheme has limited embedding capacity and less robustness to compression

and filtering attacks.

A lot of researches have been done in the area of spatial domain

watermarking. Watermarks embedded in this domain are easy to get
destroyed and hence used to design fragile watermarks. Their natural
weaknesses against common image manipulations have led watermarking
researchers to look into the transform domain.

3.2.2 Transform domain

An image can be expressed in transform domain as frequency

coefficients by mapping the pixel values. The number of frequencies in the
transformed domain corresponds to the number of pixels in the spatial
domain. The idea behind the transform domain methods is to embed the
watermark into the transform domain coefficients of the host image and the
watermarked image is obtained by applying the inverse transform. Images in
transform domain have been comprehensively studied in the context of image
coding and digital watermarking (Bow 2002, Cox et al 2002).

The most important advantage of transform domain embedding is

the resultant imperceptibility. Low frequency coefficients of the transformed
image represent the overall shapes and outlines of features in the image, its
luminance, contrast characteristics and high frequencies represent sharp edges
and crispness in the image. Since some of these coefficients are not
significantly distorted after some attacks (JPEG compression, noise and
filtering), robust watermarking schemes are usually designed in this domain.
These schemes are more complex compared to spatial domain schemes, but at
the same time, they provide better robustness and imperceptibility.
Furthermore, the watermark is distributed irregularly over the image after the
inverse transform, which makes it difficult for the attacker to read or modify.
 49

The characteristics of the HVS are better captured by the transform domain
coefficients and hence watermarking in this domain gains more importance.
Important transforms used in image watermarking applications are briefly
explained below:

Discrete cosine transform (DCT): DCT is a kind of frequency domain

transformation method for real numbers, whose transform kernel is the cosine
function. Its energy compaction property makes it to use in image processing
applications, such as JPEG lossy compression. Frequency components of the
image with minimal values are discarded, leaving only the significant
contributors of the image. DCT has a wide range of applications in image
watermarking. But, unfortunately these schemes are not robust to basic image
transformations.

Lin & Chen (2000) embedded the watermark bits in low frequency
DCT coefficients by changing LSB. But, it suffers from low robustness
against JPEG compression because of the loss of bits after quantization of
DCT coefficients. Chu (2003) used image sub sampling to obtain subsampled
images and then each sub image is DCT transformed which resulted in higher
memory requirement for implementation. It is relatively weak under JPEG
compression and low pass filtering. Patra et al (2010) used a simple Chinese
Remainder Theorem based scheme with a good robustness against common
image processing operations. But, it is not a blind scheme and not robust to
severe noise attacks owing to the random spread of noise over the DCT
coefficients which adversely affect the extraction process.

Discrete fourier transform (DFT): DFT is an important image processing

tool for image decomposition. The basic functions of DFT are sine and cosine
waves with increasing frequencies. Each point in the Fourier domain of the
image represents a particular frequency contained in the spatial domain. DFT
is the sampled Fourier transform and hence does not contain all frequencies to
 50

form an image, but only having a set of samples which is large enough to
fully describe the spatial domain image. DFT uses phase modulation
components to embed watermarks because it causes fewer distortions in
images. Many researchers have used DFT to embed watermarks (Cheng &
Huang 2003, Pun 2006, Poljicak et al 2011). But these methods are not robust
to cropping and translation attacks.

For example, Poljicak et al (2011) proposed a blind watermarking

method to minimize the impact of watermark embedding on the image
quality. The watermark embedded in the optimum radius of the Fourier
transform magnitude maximizes the PSNR value to minimize quality
degradation of the image. The radius of watermark changes with different
image sizes and hence image-dependent. This method has a high level of
security, computationally simple, device independent, adaptable and robust.
However, the watermark is composed of circular dots of random noise and
results in low capacity.

Discrete wavelet transform (DWT): DWT is a multi-resolution description

of an image. The discrete wavelet transform is used extensively as a basis for
efficient digital image watermarking owing to its space-frequency
localization, multiresolution representation, superior HVS modeling, linear
complexity and adaptivity. Meerwald & Uhl (2001), Al-Otum & Al-Taba’a
(2009) proposed a family of modified wavelet-based watermarking techniques
with the selection of specific locations in the three detailed sub bands of the
first level of DWT decomposition of color images. The selective nature of the
modified pixel-wise method allows the scheme to be adaptive in terms of
imperceptibility and watermark size. However, this method can be only able
to detect the watermark existence with no extraction capability.

Liu (2010) estimated the noise detection threshold of each wavelet

coefficient in luminance and chrominance components of color images in
 51

order to satisfy transparency and robustness requirements of color image

watermarking. The thresholds are derived in a locally dynamic fashion based
on the wavelet decomposition, through which the perceptually significant
coefficients are selected for embedding watermark. Nevertheless, these
approaches are non-invariant to image rotation and translation attacks. DWT
have its own limitations in capturing the directional information such as
smooth contours and directional edges of the image. Hence, researchers move
forward to use transforms like curvelets and contourlets (Do & Vetterli 2003)
for better performance.

Discrete contourlet transform: Contourlet transform provides a flexible

image multiresolution presentation (Do & Vetterli 2003). The discrete
contourlet transform provides a sparse representation at both spatial and
directional resolutions. It is constructed by combining two distinct and
successive decomposition stages: a Laplacian Pyramid (LP) and a Directional
Filter Bank (DFB). The LP decomposes an image into a number of radial sub
bands plus an approximation image. Then a DFB is applied to each resulting
detailed sub bands where a maximum number of directions are used at the
finest sub band. In the first stage, image is transformed into a coarse version
to decompose each LP band pass image into many wedge-shaped sub bands.
The second stage applies appropriately 2-D quincunx filtering, critical sub
band decomposition and thus capturing directional information. Finally, the
image is represented as a set of directional sub bands at multiple scales. The
number of directional sub bands at each level is 2 . Contourlets are able to
capture directional edges superior to wavelets. Besides, contourlets performs
better in depicting the geometrical structure of images. Block diagram of CT
and the frequency decomposition is given in Figure 3.1. Contourlet transform
of the Baboon image which is decomposed into two pyramidal levels, again
decomposed into four and eight directional sub bands is shown in Figure 3.2
where, small coef cients are colored black, large coef cients colored white.
 52

Figure 3.1 (a) Block diagram of contourlet transform (b) resulting

frequency decomposition

Figure 3.2 Contourlet transform decomposition of Baboon image

Contourlet transform domain has attracted a considerable amount

of research attention in the field of image watermarking. Jayalakshmi et al
 53

(2006) proposed a non blind watermarking scheme by employing the pixels

selected from high frequency coefficients based on directional sub bands of
contourlet transform. The algorithm could not be able to combine the
watermark into the host image and hence the visual quality of the
watermarked image decreases with poor robustness. Zaboli & Moin (2007)
came up with a non-blind approach to watermark gray level images in
contourlet domain using HVS characteristics. Four-level contourlet
decomposition of the host image is done and a logo image is embedded using
scrambling with the help of a well-known Pseudo Noise (PN) sequence.

The scheme proposed by Li et al (2009) combined scale-space

feature based watermark synchronization with Non-Sampled Contourlet
Transform (NSCT) to embed the watermark. But, this scheme can only
watermark gray scale images and the PSNR values of the watermarked
images are comparably less. Haohao (2009) proposed YCbCr based
watermarking scheme in which the watermarks are embedded in the largest
detailed sub bands of contourlet coefficients. But it is less robust to image
processing and signal processing attacks. The level of watermark detection
after the attacks is very low since the watermarks are embedded in the largest
frequency sub bands. Akhaee et al (2010) introduced a robust blind scheme
along with a non-blind multiplicative watermarking scheme in which the
watermarks are embedded in the higher energy directional sub bands (edges)
of the image. But it is less robust to compression and rotation attacks.

Rahimi & Rabbini (2011) presented an adaptive dual watermarking

scheme in which watermark bits are embedded in the singular value vectors of
low pass contourlet sub bands. This scheme is developed for Digital Imaging
and Communications in Medicine (DICOM) images. But the method is less
robust to salt & pepper noise and motion blur attacks.
 54

Transform domain approaches embed watermarks using additive or

multiplicative methods. Embedding is performed either in the low frequency
coefficients or high frequency coefficients in order to resist various image
processing and signal processing attacks.

3.2.3 Hybrid domain

The analysis of the two different domains suggests an improved

robustness by combining both domains - a hybrid domain approach to image
watermarking. Jagadish et al (2004) embedded encrypted patient information
(watermark) interleaved in DCT, DFT and DWT domains for medical images.
But the resistance of this method against various image processing and signal
processing attacks are not analyzed. Lie et al (2006) embedded robust and
semi-fragile watermarks into the image. In this work, DCT domain is used for
watermarking gray scale JPEG images with the help of image-dependent
information. The binary logo modulated with image features is the robust
watermark. The image-dependent channel information is the semi-fragile
watermark. Both watermarks are embedded using quantization. The method is
well-balanced in terms of imperceptibility, robustness and computational cost.
However, the global authentication rate is around 60% against JPEG
compression, average filtering, Gaussian blurring and skewing.

Ranjbar et al (2013) proposed a blind and robust watermarking

method with two embedding stages. In the first stage, the odd description of
image is divided into non-overlapped fixed sized blocks and watermark is
embedded in the high frequency components of the contourlet transform
blocks. In the second stage, the watermark is embedded in the low frequency
components. But, this method is less resistant to median filter, Gaussian
noise, salt & pepper noise and JPEG compression attacks.
 55

Watermarks embedded in the high frequency sub bands of an

image is sensitive to many image processing attacks such as low pass
filtering, lossy compression, noise and geometrical distortions. On the other
hand, watermarks embedded in the low frequency sub bands are sensitive to
other image processing attacks such as histogram equalization and cropping
(Khalighi et al 2010). Hence, the traditional image watermarking methods are
not sufficient to withstand all types of attacks. So, there is an important
requirement to devise an efficient hybrid method to watermark color images
which can be resilient to a wide range of image processing and signal
processing attacks.

3.3 PROPOSED HYBRID WATERMARKING SCHEME

We propose a blind scheme to efficiently watermark color images

which combines the advantages of spatial domain characteristics such as data
hiding capacity and transform domain properties such as imperceptibility and
robustness. The watermark is embedded in less correlated low and high
frequency sub bands of the color image in an imperceptible way. The
verification of the watermark is done in a blind manner. The functional block
diagram of the color image watermarking scheme is given in Figure 3.3. In
this scheme, GLCM is constructed for each channel of the color image.

GLCM is a statistical method to examine the spatial relationship of

pixels. It is also known as the gray-level spatial dependence matrix. GLCM
calculates how often a pixel with gray level (grayscale) intensity value
occurs horizontally adjacent to a pixel with the value .
 56

Watermark Embedding Process Watermark Extraction Process

Original Image Watermarked Image

Contourlet Attack
Transform

Sub bands
GLCM Construction
Inverse
a11 ... ... a1n Contourlet Contourlet
Transform Transform
a21 ... ... a2 n
... ... ... ...
an1 ... ... ann n n

PCA Inverse PCA

Key Generation PCA

Secret key

Watermark XOR
generation

XOR
Concatenate

Resultant
sub bands

Ownership-id

Ownership-id
XOR

Extracted watermark

Figure 3.3 Functional block diagram of the proposed color image

watermarking scheme
 57

In this work, each element, ( , ) in GLCM specifies the number of

times that pixel with value occurs diagonally adjacent to a pixel with the
value . All possible ways of relationship between adjacent pixels can be
efficiently found out using diagonally adjacent GLCM. The secret key
computed from the GLCM is used as a seed for watermark generation. Less
correlated contourlet sub bands are selected for watermark embedding with
the help of Principal Component Analysis (PCA).

PCA (Cheng 2006) is an efficient tool for reducing dimensionality

which is a mathematical procedure that uses an orthogonal transformation to
convert a set of observations of correlated variables into a set of values of
uncorrelated variables called principal components. PCA plots the data into a
new coordinate system where the data with maximum covariance is plotted
together as the first principal component. Similarly, there are the second and
third principal components and so on. The first principal component has the
maximum energy concentration. PCA is employed in this method to
efficiently choose suitable significant sub bands for imperceptibly embedding
the watermark.

3.3.1 Watermark generation process

The watermark generation process consists of two phases: secret

key generation and watermark generation. The secret key generation process
computes the eigen value of the moment = × where is the GLCM
of the cover image . Here we use mean eigen value as the secret key,
which is used for watermark generation. The watermark generation algorithm
takes and the image as input, applies contourlet transform to , finds the
size of the sub bands generated and outputs a watermark of length bits. In
this process, a random number is used to uniquely identify the host image
 58

owner. Figure 3.4 and Figure 3.5 show the secret key and watermark
generation algorithms respectively.

Algorithm - secret key generation ( )

1. Compute the GLCM [ ] × of the host image where the image is

represented as n bits per pixel.

2. Compute the transpose of the matrix as .

3. Calculate the moment = × .

4. Compute the eigen values , ,…… of the matrix .

=[ , ,…… ]

5. Convert these eigen values into integers by multiplying each eigen

value by 10 where is a positive integer, defined by the user.

6. Compute the mean eigen value K.

=( + + )/

7. Output the secret key K.

Figure 3.4 Secret key generation

In the algorithm, the value of is selected only by the user to

convert the eigen values into integers. If it produces long integers, watermark
generation process takes care of that value such that watermark is embedded
imperceptibly. The secret key is used as a seed for watermark generation. A
 59

random number is used for uniquely identifying the owner. The following
steps are contained in the watermark generation process:

Algorithm - watermark generation ( , )

1. Apply contourlet transform to the image .

2. = size of sub bands generated.

3. = binary( ).

4. Select an owner-identification number .

5. = binary ( ).

6. = , where ‘ ’ denotes the concatenation of strings.

7. = length( ) .

8. If ( = ) , go to Step 10.

else if ( < ), then pad ( ) zeros to the string and go to step 10.

9. If ( > ),

= ceil ( / ).

= ( × ) .

Pad zeros to .

10. Output the watermark .

Figure 3.5 Watermark generation

3.3.2 Watermark embedding

Numerous sub bands are generated after contourlet decomposition

of the image and selection of specific independent bands for embedding
watermark is done efficiently with the help of PCA. Instead of embedding the
watermark either completely in higher frequency bands or lower frequency
 60

bands as in the existing algorithms, the proposed hybrid method uses PCA for
the selection of less correlated sub bands with the aim of imperceptible
watermark embedding. PCA is applied to every low frequency and high
frequency bands of the contourlet transform, from which a less correlated low
frequency sub band and a less correlated high frequency sub band are
selected for embedding the watermark. Split the image into three color
channels, RGB. For each channel, the watermark embedding algorithm given
in Figure 3.6 is applied.

Algorithm -watermark embedding ( )

1. Apply contourlet transform on .

2. From the contourlet sub-bands, and are selected using PCA to

embed the watermark.

3. = binary ( ) ; = binary ( ).

4. = , = // , represents XOR operation//

5. Perform inverse PCA to spread the watermark throughout the image.

6. Perform inverse contourlet transform to get the watermarked image.

Figure 3.6 Watermark embedding process

3.3.3 Watermark extraction

This hybrid method requires only the secret key and the binary
user identification number for watermark extraction and hence it is a blind
scheme. The watermark is computed using and as per the watermark
generation procedure. At the extraction stage, the watermarked image is
 61

decomposed into three color channels (RGB) and the least correlated sub
bands and containing the watermarks are obtained after applying the
contourlet transform and PCA. The values are XOR-ed with and the
resultant sub bands are again XOR-ed with and respectively to extract
the watermark . The extracted watermark is compared with the computed
watermark for verification.

If the user inputs an invalid secret key, then the XOR operation
creates a noise in the image, so that the embedded watermark cannot be
extracted properly. The major benefit of this scheme is that only the person,
who has a valid secret key, can claim as the owner of the image. The
computational complexity of the proposed scheme is O ( ) where is the
number of pixels in the image.

3.4 EXPERIMENTAL RESULTS AND DISCUSSION

Several experiments have been performed in order to test the

efficiency of the proposed scheme, using seven color images of size
512 × 512 (Lena, Baboon, Barbara, Goldhill, Pepper, Fruits, Bridge and
Flowers) from the database of test images website. CT and MRI images are
also used from the database of DICOM images website for sake of
comparison. 9-7 orthogonal filters with two levels of multilevel
decomposition and PKVA ladder filters are used for directional
decomposition of contourlet transform. The performance evaluation is done
in terms of transparency, payload and robustness.

3.4.1 Transparency

The transparency of the embedded watermark is evaluated by

checking whether there exists any noticeable difference between the host
image and its watermarked counterpart. Peak signal-to-Noise Ratio (PSNR) is
computed between the host and watermarked images to evaluate the quality of
 62

the watermarked image and to assess the transparency of the embedded

watermark. Higher PSNR leads to good quality watermarked images.

255
= 10 (dB) (3.1)

where MSE is the mean square error between host and watermarked images.
Figure 3.7 shows the host Lena, Baboon and Pepper images and their
corresponding watermarked counterparts. It is highly difficult to find the
difference between the watermarked and the original images in vision. The
PSNR values of the watermarked images are 54.68, 53.55 and 58.32 dB
respectively. It is obvious from Table 3.1 that watermarked images are of
higher PSNR values greater than 53.55 dB with very high quality.

Table 3.1 PSNR and payload values of watermarked images

Images PSNR (dB) Payload (bits)

Baboon 53.55 200
Lena 54.68 256
Pepper 58.32 256
Fruits 54.97 288
Barbara 53.92 200
Flowers 54.94 256
Bridge 55.20 256
Goldhill 56.21 256

3.4.2 Capacity

The number of bits embedded in the watermarked image (payload)

is computed to analyze the capacity of the proposed scheme. Table 3.1
illustrates the payload embedded in the watermarked images. From the table,
one can see that the capacity is greater than or equal to 200 bits. Moreover,
 63

among the watermarked images, Baboon has minimum payload and Fruits has
maximum payload and this is due to the fact that capacity is image dependent.

(a) Lena (b) Watermarked Lena (PSNR= 54.68 dB)

(c) Baboon (d) Watermarked Baboon (PSNR= 53.55 dB)

(e) Pepper (f) Watermarked Pepper (PSNR= 58.32 dB)

Figure 3.7 Host and watermarked test images

 64

3.4.3 Robustness

In image watermarking schemes, robustness of the watermark

represents the resistance to various types of traditional image processing and
signal processing attacks which may lead to the distortion or removal of
watermark. We have tested the watermarked images with some common
attacks available from Stirmark website. The robustness of the embedded
watermark to various geometrical and signal processing attacks is measured in
terms of Normalized Correlation (NC), Bit Error Rate (BER), Normalized
Hamming Similarity (NHS) and Global Authentication Rate.

Normalized Correlation (NC): Normalized correlation is adopted as a

measure for evaluating the robustness of a watermarking scheme. NC value is
1 when no attacks are performed. The NC value between the host and
watermarked image is computed as follows:

(, ) (, )
( , )= (3.2)
2
(, ) 2
(, )

where is the host image and is the watermarked image.

Bit Error Rate (BER): Bit error rate is also used to measure the robustness
of the proposed scheme against common image processing and signal
processing attacks. BER is the ratio of wrongly extracted watermarked bits to
the total number of watermark bits embedded. Lesser the BER, more robust
the watermark is towards attacks. The BER value will be zero if the image is
not subjected to any attack.

Normalized Hamming Similarity (NHS): Normalized Hamming Similarity

(Li et al 2009) between the embedded watermark and extracted watermark
is defined as
 65

= ( , )/ (3.3)

where (. , . ) denotes the number of bits different in the two binary

strings. It is computed to evaluate the effectiveness of the proposed algorithm

Global authentication rate: In Lie et al (2006), global authentication rate is

used to make an objective decision about the accuracy of the extracted
watermarks. The global authentication rate is a quantitative measure of
correlation between the extracted and the computed watermarks. The global
G
authentication rate between the embedded watermark w and extracted
watermark w' is expressed as:

( , ) =1

1 ( )( ( )
) ( ) (3.4)
.

where denotes the number of watermarked channels and denotes the

G
watermark length. The value = 1 represents the high credibility of the
image content and no manipulation or processing happened in the
watermarked image.

Various types of attacks are applied to the watermarked images and

the robustness of the proposed scheme is calculated using BER and NC
values.

JPEG compression attack: JPEG compression is used commonly for image

compression applications. In these experiments, the watermarked images are
compressed with different quality factors ranging from 5 to 90. Table 3.2
gives the quantitative results of robustness in terms of BER for Lena, Baboon,
Pepper and Goldhill images. The results illustrate that the embedded
watermark is still robust against the JPEG compression even with a quality
 66

factor 5. Figure 3.8 depicts the high NC values even after JPEG compression
of quality factor 70 for different images.

Rotation attack: The robustness of the proposed hybrid watermarking

scheme is analyzed against rotation attack with varying rotation angles from
-50 to 50 . Table 3.2 shows the higher robustness of the proposed approach
against rotation with reduced BER and high NC values. Figure 3.8 show that
the proposed scheme has high NC values of 0.99 against rotation.

0.95

0.9

JPEG 70
NC values

Gaussian noise(0.01)
0.85

Salt & Pepper noise(0.01)

Rotation(0.5 degrees)
0.8
Scaling(0.2)

0.75

0.7

Figure 3.8 NC values against JPEG, noise, rotation and scaling attacks

Noise attacks: The Gaussian and salt and pepper noise attacks of noise
density 0.01 are tested. From Table 3.2, one can see that the corresponding
BER values are less than 5% which indicates the higher robustness of the
scheme against Gaussian noise and salt & pepper noise attacks. Also Figure
 67

3.8 indicates that the resulting higher NC values are really high and 0.98 for
all test images under noise attacks.

Image scaling: The watermarked images are scaled down by 2% and 5%.
The reduced images are re-scaled back to its original dimensions and the
watermarks are extracted. Table 3.2 shows the BER values of the extracted
watermarks under scaling attacks. It is observed that BER is around 7%.
Figure 3.8 shows that the NC values are between 0.95 and 0.97 which
indicates that the embedded watermark is more affected by scaling attack than
any other attacks.

Filtering attacks: Different types of ltering attacks have been tested, such
as Gaussian, sharpening, wiener and median filters of different window sizes
to the watermarked images. The corresponding results are shown in Table 3.2
and Figure 3.9. Obviously, the proposed scheme resists the specified attacks
and the watermark survives with higher NC values between 0.93 to 0.99. NC
values are exactly 1 for Gaussian and median filters. It is interesting to see
that the BER values after Gaussian filtering is zero for all test images. The
experimental results show high robustness of the scheme against Gaussian,
wiener and median filter attacks.
 68

Table 3.2 BER (%) of the extracted watermarks under common image
& signal processing attacks

Images
Attacks Amount of distortion
Lena Baboon Pepper Goldhill
5 3.71 3.77 3.98 3.96
10 3.59 3.79 3.36 3.80
20 3.69 3.81 3.96 3.83
JPEG 30 3.81 3.85 3.79 3.81
50 2.56 3.59 4.17 3.82
80 3.03 4.40 4.58 2.93
90 3.45 3.44 4.41 3.02
3×3 4.21 3.34 2.61 3.78
Median filter
5×5 4.35 4.01 3.72 4.14
3×3 0.00 0.00 0.00 0.00
Gaussian filter 5×5 0.00 0.00 0.00 0.00
7×7 0.00 0.00 0.00 0.00
Wiener filter 3×3 3.68 3.64 4.56 4.82
Sharpening - 2.41 4.12 4.64 4.00
0.01 3.83 3.68 4.80 3.45
0.02 4.94 3.78 4.73 3.72
Gaussian noise
0.03 4.84 4.50 4.94 3.36
0.05 4.85 4.07 4.87 3.58
0.01 3.82 3.63 4.37 2.38
0.02 2.73 3.53 4.67 3.54
0.05 3.58 3.83 3.68 3.49
Salt & pepper noise
0.06 3.45 3.73 3.98 3.58
0.07 3.58 2.96 4.11 4.15
0.08 3.93 3.20 3.55 3.92
-5 4.34 4.36 4.44 2.05
-2 3.04 2.03 3.33 1.60
-1 3.50 2.61 2.77 2.56
-0.5 3.61 2.18 2.75 2.98
Rotation
1 3.66 1.74 2.25 3.34
0.5 3.36 2.18 3.21 2.33
2 2.92 2.03 3.26 4.00
5 4.44 2.74 4.47 3.35
0.2 4.63 5.34 4.78 5.80
Scaling
0.5 5.23 6.23 7.01 6.90
JPEG 50 + Med.filter (3×3) - 4.29 4.43 4.71 4.71
JPEG 50 +Gaussian noise (0.01) - 4.48 2.61 3.22 2.93
JPEG 50 +scaling (0.2) - 6.25 4.36 6.66 5.41
 69

0.95

0.9
NC values

Median filter (3x3)

0.85 Sharpening
Gaussain filter (3x3)
Gaussian (3 3)
0.8
Wiener filter
0.75

0.7

Figure 3.9 NC values against different types of filtering attacks

Combined attacks: Some additional attacks are also performed to the

watermarked test images which combines two or more attacks in order to
evaluate the robustness of the proposed scheme. JPEG compression attack
combined with median filter, Gaussian noise and scaling are applied to the
watermarked images. The results are shown in Table 3.2. When JPEG
compression of quality factor 50 is combined with median filter and Gaussian
noise, very good NC values of 0.99 and very less BER values (< 4.75%) are
attained. But when scaling is combined with JPEG, the watermarks are
extracted with comparatively high BER values around 6.25% and less NC
values around 0.96. Hence, our scheme is less resilient to scaling attack.

3.5 COMPARISON WITH THE EXISTING SCHEMES

In this section, the performance of the proposed method is

compared with other well-known robust image watermarking schemes given
in Lie et al (2006), Li et al (2009), Haohao (2009), Rahimi & Rabbani
 70

(2011), and Ranjbar et al (2013) in terms of PSNR, NHS, global

authentication rate, NC and BER .

Table 3.3 gives the comparison of PSNR values of the proposed

scheme with the traditional image watermarking schemes such as Li et al
(2008) and Ranjbar et al (2013) for Lena, Pepper and Baboon images. It can
be seen that the proposed scheme achieves a very good PSNR value greater
than 53.55 dB which ensures high quality of watermarked images.

Table 3.3 Comparison of PSNR (dB) of watermarked images

Image Li et al (2009) Ranjbar et al (2013) Proposed method

Lena 44.03 38.72 54.68
Pepper 45.28 38.95 58.32
Baboon 45.57 38.16 53.55

Table 3.4 gives the comparison of NHS mean values of Li et al

(2009) with that of the proposed method for the images Lena, Pepper and
Baboon. The performance of the hybrid method is better in terms of NHS
mean values for the different image processing and signal processing attacks.
Li et al’s scheme is less robust when the rotation is more than 30 degrees for
added Gaussian noise and for some more attacks.
 71

Table 3.4 Comparison of NHS values under various attacks

Lena Pepper Baboon

Attacks Proposed Li et al Proposed Li et al Proposed Li et al
method (2009) method (2009) method (2009)
Added Gaussian noise 0.965 0.906 0.944 0.891 0.950 0.819
Median filter ( 3 × 3) 0.977 0.969 0.988 0.906 0.945 0.860
JPEG 30 0.958 0.900 0.990 0.867 0.962 0.813
JPEG 50 0.962 0.969 0.935 0.932 0.965 0.860
JPEG 70 0.971 0.969 0.990 0.958 0.963 0.980
Med. Filter (3×3) + JPEG 0.965 0.988 0.987 0.906 0.960 0.833
90
Rotation (10 degrees) 0.963 0.963 0.959 0.975 0.939 0.866
Rotation (30 degrees) 0.937 0.922 0.959 0.919 0.938 0.906
Rotation (45 degrees) 0.934 0.891 0.949 0.888 0.923 0.922
Rotation (60 degrees) 0.934 0.850 0.954 0.896 0.924 0.875
Rotation (90 degrees) 0.967 0.844 0.936 0.869 0.970 0.813
Rot. 10 degrees + 0.971 0.969 0.972 0.956 0.941 0.917
cropping
Rot. 15 degrees + 0.979 0.969 0.995 0.984 0.944 0.938
cropping
Rot. 30 degrees + 0.949 0.922 0.955 0.980 0.937 0.927
cropping
Rot. 45 degrees + 0.987 0.969 0.966 0.953 0.923 0.863
cropping
Rot. 60 degrees + 0.949 0.938 0.954 0.885 0.931 0.828
cropping
Scaling 0.6 × 0.830 0.844 0.823 0.898 0.843 0.750
Scaling 0.8 × 0.841 0.888 0.852 0.969 0.865 0.906
Scaling 1.2 × 0.860 0.960 0.886 0.980 0.855 0.860
Scaling 1.4 × 0.883 0.969 0.866 0.906 0.838 0.875
Translation x-40 and y-40 0.950 1.000 0.962 1.000 0.944 0.969
Rot. 10 deg. + scaling 0.848 0.938 0.864 0.938 0.854 0.896
0.9 ×
Rot. 30 deg. + scaling 0.864 0.875 0.866 0.919 0.852 0.875
0.9 ×
 72

Table 3.5 compares the performance of Haohao (2009) with our

hybrid method. represents the number of color channels from which the
watermarks are extracted. is the threshold value and is the average
measure of the correlation values between all contourlet sub bands of attacked
image and that of the corresponding contourlet sub bands of the host image.
values should be much higher than the threshold for a very good
correlation. In the proposed method, values are four times greater than the
threshold values, even in the case of median filtering, JPEG, Gaussian noise
and rotations. But in Haohao (2009) method, values are comparably very
less than the threshold values and the watermark can be extracted only
from the two color channels i.e, = 2 for rotation attack. But the proposed
method is capable of extracting watermark from all the three channels. The
improved and values of our method ensure the robustness against median
filtering, JPEG, Gaussian noise and rotation attacks.

Table 3.5 Comparison of , and values under various attacks with

Haohao (2009) scheme

T R D

Proposed Proposed Proposed

Haohao Haohao Haohao
Attacks method method method
Y/Cb/Cr Y/Cb/Cr Y/Cb/Cr
R/G/B R/G/B R/G/B

No attack 0.24/0.24/0.22 2.54/0.23/0.35 1.00/0.99/0.99 8.59/8.39/6.94 3 3

Median
filtering 0.24/0.24/0.23 0.54/0.28/0.43 1.01/0.99/0.99 0.80/0.59/0.85 3 3

JPEG 0.20/0.21/0.23 2.19/0.07/0.21 1.01/1.01/0.99 3.28/0.11/0.23 3 3

Gaussian
noise 0.19/0.20/0.19 3.26/4.43/2.83 0.99/1/0.99 7.83/8.87/6.55 3 3

Rotation 0.16/0.23/0.17 1.22/0.16/0.40 1.20/1.05/1.00 0.93/0.55/0.82 3 2

 73

In Table 3.6, the comparison of global authentication rate of Lie

et al (2006) with the proposed method is shown. It clearly depicts that the
proposed method is robust in terms of very good global authentication rate
with more than 92 % against JPEG compression, average filtering and
skewing compared to Lie et al (2006) for Lena image.

Table 3.6 Comparison of global authentication rate for Lena image with
Lie et al (2006)

Global Authentication Rate (%)

Malicious Processing
Proposed method Lie et al (2006)
JPEG quality factor=10% 97.20 53.52
JPEG quality factor=20% 97.48 62.11
Average filter (5 × 5) 96.22 54.98
Gaussian blurring (radius=2) 96.66 53.91
Skewing (2 degrees) 92.55 61.62

The proposed hybrid method is also compared with a contourlet

domain watermarking scheme for DICOM images proposed by Rahimi and
Rabbani (2011). The corresponding BER values for filtering, noise, resizing
and motion blur attacks are given in Table 3.7. The proposed method is highly
robust against these attacks with very less BER for the watermarked DICOM
image. Table 3.7 shows the improved performance of the proposed method
under various attacks.
 74

Table 3.7 Comparison of BER (%) of on the watermarked DICOM

images

Rahimi & Rabbani

Attacks Proposed method
(2011)
No attack 0 0
Average filter (3 × 3) 4.53 37.08
Median filter (3 × 3) 4.83 25.15
Wiener filter (3 × 3) 4.07 36.39
Salt & pepper noise 2.66 32.73
(0.003)
Salt & pepper noise 3.42 44.03
(0.005)
Resize (25%) 6.45 43.29
Resize (50%) 7.25 25.71
Motion (10,45) 4.07 45.77

Table 3.8 Comparison of BER (%) of the extracted watermark of our

method with Ranjbar et al (2013) for Lena image

Ranjbar et al
Attacks Proposed method
(2013)
JPEG compression 2.56 7.49
(QF = 50)
Median filter (3 × 3) 4.21 13.37
Salt & pepper noise (20%) 2.73 11.71
Averaging filter (3 × 3) 3.02 3.41
Gaussian noise (10%) 2.81 11.71
Rotation (0.50) 3.36 -
Scaling(50%) 4.63 0
Cropping(25%) 4.01 0
Sharpening 2.41 0
 75

Also a comparison of BER values of the proposed method with a

recent contourlet domain watermarking scheme proposed by Ranjbar et al
(2013) has been done and the results are shown in Table 3.8 and Figure 3.10
for watermarked Lena image. Simulation results listed in this table indicate
that against all the attacks except scaling, cropping and sharpening, our
method outperforms Ranjbar et al (2013). Moreover, their method is not
robust against rotation and median filter.

The proposed method attains high NC values for all the attacks
compared as shown in Figure 3.10. The method in Ranjbar et al (2013) is not
robust against rotation and median filter.

Simulation results given in Tables 3.3-3.8 and Figure 3.10 indicate

that the proposed method is better than recently proposed robust contourlet-
based schemes for a wide variety of attacks in terms of BER and NC values.

1
0.95
0.9
0.85
NC values

0.8
Ranjbar et al (2013)
0.75
0.7 Proposed scheme
0.65
0.6
0.55
0.5

Figure 3.10 Comparison of NC values with Ranjbar et al (2013) for

Lena image
 76

3.6 SUMMARY

This chapter deals with a robust, blind, color image watermarking

scheme with GLCM-based watermark generation and PCA-based watermark
embedding in contourlet domain. A unique watermark is generated for every
image. No additional memory space (logo) is required to store the watermark
as compared to the existing schemes. Authenticity is achieved by including
the owner identification number as well as an image-dependent user-id in the
watermark. The proposed method possesses imperceptibility, authenticity,
security and robustness characteristics as opposed to the majority of image
watermarking methods which focus only on imperceptibility and robustness.
It is impossible for an attacker to compute the secret key from the
watermarked image because the watermark is securely embedded in the
frequency domain even though the secret key generation process is an open
algorithm. The hybrid Contourlet based color image watermarking scheme is
benchmarked against Stirmark, showing good resilience against different
types of attacks.

The robustness of the proposed scheme against various image

processing and signal processing attacks owes to the special structure of the
Laplacian pyramid (Song et al 2008). Although the watermark is embedded in
the less correlated low frequency and high frequency sub bands, it is likely to
spread out into all the sub bands when the watermarked image is
reconstructed back. The low frequency sub bands of the watermarked image
contain the watermark and hence the proposed scheme is very robust against
high frequency attacks such as low pass filtering, compression, noise and
geometrical distortions which will destroy the high frequencies of the image.
Some watermarks are preserved in the high frequency sub bands too. So, the
proposed scheme is robust against low frequency attacks such as histogram
equalization and cropping. But, since the Laplacian pyramid and directional
 77

filter banks do not do down samplings, it causes the contourlet transform not
to have translational invariance (Shunqing et al 2012). This is why our
scheme is less robust to scaling attack. The comparison of the proposed
method with other existing schemes shows improved performance of the
scheme in terms of transparency and robustness. The person who owns a valid
secret key and a user-id only can claim as the rightful owner of the image.
This method can be efficiently applied for copyright protection and ownership
verification of digital images.

Although there is a lot of work focusing towards different

watermarking schemes for color images in the literature, not much attention
has been given to 3-D anaglyph images. In the next chapter, we propose an
efficient watermarking scheme for 3-D anaglyph images.