Content uploaded by Kevin Curran
Author content
All content in this area was uploaded by Kevin Curran on Jan 26, 2018
Content may be subject to copyright.
A novel approach to digital watermarking, exploiting
colour spaces
Fre
´de
´ric Lusson
a
, Karen Bailey
a
, Mark Leeney
a
, Kevin Curran
b,
n
a
Computing Department, Institute of Technology, Letterkenny, Co. Donegal, Ireland
b
Intelligent Systems Research Centre, University of Ulster, Derry, N. Ireland, UK
article info
Article history:
Received 16 December 2011
Received in revised form
8 October 2012
Accepted 14 October 2012
Available online 13 November 2012
Keywords:
Watermarking
Steganography
Image processing
Security
Information hiding
abstract
Watermarking is the process of embedding information in a carrier in order to protect
the ownership of text, music, video and images, while steganography is the art of hiding
information. Normally watermarks are embedded in images but remain visible in the
majority of commercial image databases, such as Getty (gettyimages.ie) or iStock Photo
(istockphoto.com). However this leaves traditional watermarking techniques vulnerable
to tampering. Thus the advantage of using steganographic techniques for watermarking
is that the watermark is resistant to detection and consequently to tampering. Robust-
ness is a characteristic of critical importance, in order that a watermark is to survive
image manipulation and enhancement processes, as well as intentional attacks, to ensure
piracy is prevented.
The aim of this work is to produce a novel hybrid digital watermarking technique,
based on the exploitation of both the RGB and the YCbCr colour spaces, using spatial
domain techniques. Results demonstrate that the proposed hybrid technique can with-
stand levels of geometric attacks and processing attacks up to a point where the
commercial value of the images tested would be lost. Results also demonstrate technical
and performance improvements over existing methods, in terms of security and
algorithm efficiency, while taking inspiration from steganography, to avoid drawing
attention to the fact that an image contains hidden information.
&2012 Elsevier B.V. All rights reserved.
1. Introduction
The unprecedented increase in piracy and digital
criminality over the past 10 years has stimulated interest
in the field of watermarking to enhance protection against
violations of copyrighted digital material, such as digital
images. According to a recent study carried out by TERA
Consultants for the International Chamber of Commerce
and made public in March 2010, the European creative
industries lost around 9.9 billion euros and over 186,000
jobs in 2008 because of piracy, mainly digital piracy [48].
Over the last 15 years, many watermarking methods
have been developed and tested with the aim of providing
reliable ways of proving image ownership. Surveys
detailing the most popular watermarking techniques
can be found in the literature [3,5,39,28]. This document
does not attempt to give a comprehensive review
of all the watermarking and steganographic techniques
developed over the past 15 years, as there is an
impressive amount of research in this area. Rather, this
paper discusses the most significant steps and techniques
developed in the context of watermark invisibility and
robustness in hiding information in digital images, in
order to propose a novel watermarking technique
approach. There are multiple data hiding methodologies
and algorithms, each solving a particular facet of the
watermarking problem, while no technique outperforms
the others from all points of view. Most of the water-
marking and steganography schemes are applied to
grey-scale images, or colour images first transformed into
Contents lists available at SciVerse ScienceDirect
journal homepage: www.elsevier.com/locate/sigpro
Signal Processing
0165-1684/$ - see front matter &2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.sigpro.2012.10.018
n
Corresponding author.
E-mail address: flusson@utvinternet.com (K. Curran).
Signal Processing 93 (2013) 1268–1294
grey-scale images before the embedding phase would
occur. However their application to colour images might
not be completely adequate since they do not take into
consideration the implication of the Human Visual Sys-
tem and in particular its sensitivity to colour brightness
and perception.
More recent watermarking studies [38,32] have turned
their attention to colour images rather than grey-scale
images. Effectively, colour may be more than just an
extension of grey scale. It is considered as a key element
for a number of image processing systems. In particular,
colour space transforms have played a central role in
coding, compression and transmission applications, in
television, video and image processing. While RGB chan-
nel format is a natural scheme for representing real-world
colour, each of the three channels is highly correlated
with the other two. YCbCr is a component colour space
used by digital video. Unlike the RGB model, YCbCr breaks
the visual information into black and white (luma) signal
and two colour components. It separates luminance from
chrominance (lightness from colour). With many more
rods than cones, the human eye is more attuned to
brightness and less to colour differences. Hence the YCbCr
colour system allows more attention to be paid to Y, and
less to Cb and Cr. As a result, using Cb and Cr values to
embed the watermark, rather than the Y channel, should
achieve watermark invisibility. Results show that water-
marks hidden in the YCbCr and XYZ colour spaces in
particular, are better recovered after JPEG compression
attacks. Similar results are noticed with Gaussian noise
attacks. These results demonstrate that the YCbCr and
XYZ colour spaces have large amount of perceptual
redundancy for colour pixels in this colour space. The
larger the extent of perceptual redundancy, the greater
the strength of the watermark signal that can be
embedded, and the higher the robustness of the
embedded watermark. Although the variety of attacks is
quite limited, the YCbCr colour space shows better overall
robustness to attacks while preserving the watermark
invisibility. Robustness tests done against geometric
attacks appear limited. Previous studies suggest that
DWT algorithms perform well against compression and
filtering.
2. Related work
Very early research focused on LSB insertion in the
spatial domain (pixel level) of images for its simplicity
and its potentially large capacity. Later scientific research
considered the frequency domain and the quantisation of
coefficients. Research conducted for the purpose of this
work would indicate that watermarking and steganogra-
phy techniques can be classified into Spatial Domain,
Frequency Domain and Adaptive methods. Adaptive
methods are treated as a special case here, because they
can either be applied to the spatial domain or to the
frequency domain. The following sections examine each
domain methodology and analyse their impact on achiev-
ing the optimum watermarking requirements.
2.1. Spatial domain methods
Histogram equalisation is used in image processing to
adjust contrasts [22]. The aim of this technique is to better
distribute intensity values on the histogram. This allows for
image areas of lower local contrast to gain a higher contrast.
Histogram equalisation accomplishes this by effectively
spreading out the most frequent intensity values.
Histogram-based data-hiding is a commonly used water-
marking scheme. In its simplest form, pre-defined histo-
gram values are used to embed the watermark. Chrysochos
et al. [8] chose a blind algorithm with an asymmetric key to
embed the watermark into histogram values. They show
that after embedding, the histogram shape is mainly pre-
served. They also demonstrate their algorithm to be robust
against geometrical attacks such as rotation, flipping, trans-
lation, aspect ratio changes and resizing, warping, shifting,
drawing and scattered tiles, as well as their combinations.
They did not test their algorithm against compression nor
against filtering attacks. In addition, the data hiding capacity
is very much restricted to 127 bits (for grey-scale images)
and 384 bits for colour images.
Such a scheme has the advantage of recovering the
original cover image from the combined image. In addition,
a modified histogram does not affect the visual perception
of the image. The main drawback of this technique is that
the embedding strategy can be detected more easily, just by
comparing the histogram shape of the original image versus
the watermarked image. Chrysochos et al. [8] and Bayley [4]
suggest that the main advantage of histogram based data
hiding is its robustness to rotations and other geometric
transformations. On the other hand, the main difficulty
associated with this technique is that there is a non-linear
relationship between its representation and the pixel repre-
sentation. Spatial domain methods concern the modification
of a pixel value directly on the spatial domain of an image
[36]. All studies referred to in this section are applied to
either JPG or BMP images. One of the simpler approaches to
hiding data within an image file is LSB insertion. Using this
method, the binary representation of the hidden data is
computed and LSB of each byte within the cover image is
overwritten. There is a trade-off between preserving the
image quality versus information hiding (watermark or
secret message) payload, although it is generally accepted
that modifying the LSB of each pixel does not visually alter
image quality. A reasonable capacity is a third the size of the
host image original size [47] (p. 34). This algorithm, pre-
sented by Shih and Wu [46] and Celik et al. [9], is easy to
break, by flipping the least significant bit of every pixel of
the image, or by embedding a new watermark on top of the
current one. On the other hand, it is easy to implement and
it requires less processing power. To alleviate this concern,
other algorithms [17] have been introduced whereby a
private key is used to define where the bit value should
be embedded (LSB, LSB2 or LSB3). Varying the bit position
used, makes it a lot more difficult to find which bit is used
to embed the watermark bit. One potential problem with
any of the LSB methods is that they can be discovered
visually by an adversary who is looking for unusual pat-
terns, or by using steganalysis tools. LSB manipulation is a
fast and relatively inexpensive way of hiding information
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1269
but it tends to be vulnerable to spatial changes resulting
from image processing or lossy compression. Such compres-
sion is a key advantage that JPEG images have over other
formats. High quality images can be stored in relatively
small files using JPEG compression method. LSB’s natural
shortcoming regarding weaknesses against image manipu-
lations such as JPEG compression, has led researchers to
look into the frequency domain. The next section discusses
these techniques.
2.2. Frequency domain
2.2.1. Discrete cosine transform
All DCT multiplications are on real numbers
[1,7,18,29,50]. DCT is widely used with image compres-
sion such as JPEG lossy compression, because it has a
strong ‘‘energy compaction’’ property [43]: most of the
signal information tends to be concentrated in a few low-
frequency components of the DCT. Frequency components
with minimal values are discarded, leaving only the
‘‘significant contributors’’ of an image. DCT based algo-
rithms are more robust to JPEG lossy compression which
is also based on the DCT. Unfortunately, these DCT based
schemes are not robust to basic transformations. Li and
Wang [30] proposed the modification of the Quantisation
Table (QT) part of the JPEG and used the middle frequency
coefficients to hide the message. The aim of quantisation
is to retain the valuable information while eliminating the
‘‘not so important’’ information. Sal Diaz and Grana
Romay [44] have proposed a multi-objective genetic
algorithm which searches the best localisation of the
DCT of an image to place the mark-image-DCT-
coefficients for minimal visual distortion and optimal
robustness. They measured the results of this algorithm
based on the Pareto-Front, which represents the trade-off
between image fidelity and robustness. Predicting attack
type and strength is however not a simple matter. It is
therefore very unlikely that this algorithm would fit all
conditions. Liu and Chou [32] have compared the efficacy
of a watermarking scheme between three different colour
spaces (YCbCr, XYZ, CIElab). The algorithm used the
frequency domain, extracting the wavelet coefficient with
highest perceptual redundancy in each colour band. To do
so, the DWT is performed independently on each colour
channel. Depending on the subband targeted, filtering
blocks of varying sizes are applied and the minimum
values of the Just Noticeable Coefficients (JND) are tar-
geted for embedding. Finally the inverse DWT is per-
formed to rebuild the combined image.
2.2.2. Discrete Fourier transform
The Fourier Transform is an important image proces-
sing tool which is used to decompose an image. The
output of the transformation represents the image in the
Fourier or frequency domain, while the input image is the
spatial domain equivalent. In the Fourier domain image,
each point represents a particular frequency contained in
the spatial domain image. The Fourier Transform is used
in a wide range of applications, such as image analysis,
image filtering, image reconstruction and image compres-
sion. The DFT is the sampled Fourier Transform and
therefore does not contain all frequencies forming an
image, but only a set of samples which is large enough
to fully describe the spatial domain image. The number of
frequencies corresponds to the number of pixels in the
spatial domain image, i.e. the images in the spatial and
Fourier domain are of the same size (Figs. 1–8).
The basic functions are sine and cosine waves with
increasing frequencies, i.e. F(0,0) represents the DC-
component of the image which corresponds to the aver-
age brightness and F(N1,N1) represents the highest
frequency. Low frequencies are responsible for the gen-
eral grey-level appearance of an image over smooth areas,
while high frequencies are responsible for detail (edges
and noise) [22]. The DFT, based on fast Fourier transform
methodology, uses phase modulation instead of magni-
tude components to hide messages, since phase modula-
tion has less visual effect. The output of the
transformation represents the image in the Frequency
Domain. This methodology has been used by Chi-Man
[12], Kim et al. [26], Kutamura et al. [27] and Qiang and
Huang [42], which demonstrate that DFT is preferable to
DCT when it comes to dealing with geometric manipula-
tions such as cropping and translation.
2.2.3. Discrete wavelet transform
The DWT provides a powerful insight into an image’s
spatial and frequency attributes. Hence, the DWT has
gained a lot of popularity (most of the recent research
on digital grey scaled image watermarking is based on
DWT [19,33,35], for the fast transformation approach that
translates an image from spatial domain to frequency
domain while still providing robustness. An example of
such a decomposition on ‘‘Lena’’ image can be seen in
Fig. 9. The watermark is inserted in the transform coeffi-
cients. The insertion process may be separated in 3
phases: computation of the DWT coefficients (using
various filters such as Haar, Daubechies [23]), addition
Fig. 1. First level DWT decomposition in RGB colour channels.
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941270
of the watermark to those coefficients (for example
modifying those that are above a given threshold in the
sub-bands other than the low pass sub-band) and com-
pute the inverse DWT to reconstruct the watermarked
image. Ghannam and Abou-Chadi [20] proposed a variant,
where embedding is performed in two bands representing
low and high frequency components in order to achieve
both imperceptibility and robustness.
2.3. Adaptive watermarking
All previously mentioned watermarking methods hide
the watermark in the spatial or the frequency domain, by
addition, multiplication or replacement. To do so, a fixed
raster (pixel grid or fixed size block image division for
example) is used for embedding and extracting the water-
mark. Further research in the HVS suggests that exploiting
certain image characteristics (corners, edges, luminance)
might protect the embedded data from deliberate attacks.
The hypothesis is as follow: if the watermark is hidden in
regions of an image less likely to be modified because of
their intrinsic value, the watermark’s survival probability
would be increased. Methodologies which use visually
significant regions in an image to hide the watermark
[2,38,41] are classified in this section. Cox [14] has
indicated that watermarks should be embedded into
regions with large magnitudes in the frequency coeffi-
cients, since geometric processing affects regions with
low coefficients.
Chen et al. [10] proposed a LSB-based solution to
embed the hidden message into pixels located in the
image edges. They combined the fuzzy edge detector with
the Canny edge detector to increase the embedding pay-
load. Miller et al. [37] has used the technique of informed
coding and embedding with a similar problem of low
embedding capacity. Schlauweg et al. [45] proposed an
algorithm where the watermarking position is deter-
mined by the image content, using textels (texture ele-
ments), particularly the grey-level blob, which has the
property of being scale invariant, therefore resistant to
geometric attacks. Lou et al. [34] proposed an adaptive
steganography scheme, capable of providing for a large
embedding capacity, while preserving the visual quality
of an image. They use the variation among the immediate
neighbouring pixel values to predict the embedding
capacity of each pixel. However, they did not measure
the robustness of their algorithm against attacks. A skin
tone detection steganography algorithm is proposed by
Cheddad et al. [6], which demonstrates robustness to
attacks, while keeping the stego data invisible, by embed-
ding in skin regions of an image. This is perfectly suited to
steganography where the cover image can be specifically
chosen with skin attributes in it. Unfortunately, this
technique is not generic enough to suit watermarking,
where one has no choice when it comes to choosing the
cover image. Genetic Algorithms are an important opti-
misation technique [24], in the area of evolutionary
computation. Khan and Mirza [25] have suggested the
idea of exploiting the HVS, combined with genetic algo-
rithms, to structure a watermark based on the cover
image and the intended attack, to optimise impercept-
ibility and robustness. One way to resist attacks is to
invert attack distortions at the decoding end. Gilani et al.
[21] concentrated on increasing the robustness of a
watermarking system by estimating a watermarked
image distortion, using distortion estimation functions
Fig. 3. Original and extracted watermark.
Fig. 4. Extracted watermark after JPEG compression.
Fig. 2. Original and watermarked ‘‘lena’’ image compared.
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1271
based on Genetic Programming. Results show superior
performance compared to the previously proposed tech-
nique by Piva et al. [40]. However, the technique uses
known attacks to generate the estimation functions, i.e.
the original image is tested against the attacked image to
generate the best estimation function. There are a lot of
undefined variables that can interfere with the process,
such as what attack or combination of attacks should the
host image be protected against, and for each attack, what
is the degree of the attack. To produce the results, Gilani
et al. had to pre-define the attacks, which unfortunately
do not reflect the reality.
More recently, Autrusseau and Le Callet [2] have
demonstrated the usefulness of combining the advances
in the understanding of the HVS with a Fourier space
watermarking technique, in providing good robustness
properties when subjected to many kinds of distortions.
Mohanty and Guturu [38] suggest extracting the most
perceptually important region of an image to embed the
watermark, using a combination of HVS metrics such as
intensity, contrast, location and edges. Cong [13] uses
Canny edge detection to isolate pixels of significance in
the representation of an image (feature points matching)
so that even after attacks, the watermark can be retrieved
Fig. 5. Watermark embedding in the RGB channels.
Fig. 6. ASCII watermark embedding in the CbCr channels.
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941272
by a blind process and identified. In the domain of
exploiting the HVS for steganography, Chen et al. [10]
have demonstrated an interesting hybrid edge detector
algorithm (a combination of fuzzy edge detector with
Canny edge detector) to improve payload capacity and
invisibility, in the spatial domain, using an LSB embedding
technique.
3. The proposed watermarking scheme
Watermarking algorithms studied by Cox et al. [15]
demonstrate that in most cases they only resolve some of
the potential attacks. A practical example of this is found
in Adobe Photoshop, in the form of a tool, designed by
Digimarc, to embed a numerical tag into photos, using
wavelet encoding techniques, which illustrates its resis-
tance to many basic distortions but not all: for example
rotating an image by 45 degrees before applying blurring
and sharpening filters will destroy the watermark.
To increase the chances of watermark survival after
multiple attacks, a combinative watermarking approach is
suggested by Shih [47] based on previous research done
in this area by Tsai et al. [49] and Shih and Wu [46],to
provide for high capacity watermarks, by splitting the
watermark into two parts; one part is embedded into the
Fig. 7. Watermark extraction from the CbCr channels.
Fig. 8. Watermark extraction from the RGB channels.
Fig. 9. MATLAB GUI comparing the original ‘‘lena’’ image and ‘‘lena’’
after the proposed hybrid watermarking method is performed.
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1273
spatial domain while the other part is embedded into the
frequency domain. The LSB technique is adopted in the
spatial domain, using a pseudo random number generator
to locate the pixels for the embedding process. Embed-
ding in the lower frequency components to avoid loss
after compression is the technique chosen in the fre-
quency domain. A random permutation of the watermark
is also used, showing a better resistance to cropping, in
particular. Chemak et al. [11], encode the watermark and
embed the result in both the frequency and the spatial
domain. They use the 5/3 wavelet decomposition adapted
for JPEG 2000 compression. Images transformed using
this algorithm is naturally resistant to this kind of attack.
It also has the advantage of not losing coefficient value
precision as it is an integer to integer conversion. They
propose to modify pixels luminance values using the LSB2
(second least significant bit) as their embedding techni-
que in the spatial domain.
Both studies suggest an improved robustness by com-
bining frequency and spatial domains in their algorithm, to
the detriment of performance, due to higher complexity.
However, it is suggested that the success of a watermarking
scheme should be founded not only on robustness and
invisibility, but also on the simplicity of its algorithm, to
enable better commercial viability. While processing power
(CPU and memory) increases all the time and becomes
cheaper with every new hardware release, image manipula-
tions require large amount of memory and fast CPU cycles
to be efficient. This is not an issue when serving individual
needs. However, for a commercial entity such as Getty
Images, which stores millions of images, algorithm effi-
ciency would certainly be an important decision factor. For
example, if it takes 2 s to embed a watermark into an image
using algorithm 1 and it takes 6 s using algorithm 2, it
would cost three times more in hardware investment to
choose the less efficient algorithm. If the most efficient
algorithm can perform as well as the less efficient one (in
terms of robustness and invisibility), the final decision
becomes easy to make.
Hence, this study proposes a completely different hybrid
approach, built upon the assumption that combining both
spatial features and colour space might improve robustness.
So rather than utilising a grey-scale image, advantage of the
colour redundancy that the YCbCr colour space offers will be
taken. A relatively simple algorithm for improved perfor-
mance has been focused on, inspired by two techniques
(additive pixel value and LSB substitution) widely used in
steganography. The following combined RGB and YCbCr
watermarking technique is proposed:
Non blind, additive pixel value in the RGB components,
using a JPEG black and white watermark (50 by 50
pixels).
Blind, LSB substitution in the Cb and Cr components of
the YCbCr colour space, using an ASCII text watermark.
This technique enables to hide more information,
without compromising invisibility.
Unlike the RGB model, YCbCr breaks the visual informa-
tion into black and white (luma) signal and two colour
components. It separates luminance from chrominance
(lightness from colour). With many more rods than cones,
the human eye is more attuned to brightness and less to
colour differences. Hence the YCbCr colour system allows
more attention to be paid to Y, and less to Cb and Cr. As a
result, using Cb and Cr values to embed the watermark,
rather than the Y channel, should achieve watermark invisi-
bility. The fact that YCbCr is applied in digital video (YUV and
YIQ are for analogue video), suggests that the YCbCr colour
space could be used to good effect with still digital images as
digital images are the building block of digital video and
there is a linear relationship between RGB and YCbCr.
Using the same line of thought, it may be possible to use
a blind watermarking technique in the YCbCr colour
domain. One would assume that the extra information
needed to extract the original watermark without prior
knowledge of the embedding process, will not affect the
invisibility quality of the proposed scheme. The second
important point to make about the YCbCr colour space is
that the translation between it and the RGB colour space
is linear and simple, therefore requiring little processing.
The RGB colour space is also exploited for the simple fact
that it is readily available to use on computer systems,
therefore requiring little computation overhead. Since it
is important to preserve the correlation between each
component (R, G, B), the same amount of watermark will
be embedded in each colour channel respectively, using
an additive LSB technique. Previous research in grey-
scale image watermarking has shown that it is more
robust to spread the watermark across the entire image
space rather than localising it to specific image regions.
For this reason, the watermark image will be rescaled to
the original image size before embedding.
It is proposed to use a blind watermarking technique in
the YCbCr colour domain. A small ASCII text used as the
watermark is the chosen option to embed in the YCbCr
because the amount of information to hide is small, there-
fore the chances of invisibility will remain high. The choice
of a black and white watermark in the RGB is motivated by
2factors:
Simplicity: the original image pixel is modified by a
small coefficient only when the watermark white pixel
occurs (value is 255).
Better chance of recovery as the colour black and white
is at the both end of the range (0 and 255).
It is hoped that using a hybrid watermarking system
will improve the general robustness against most digital
image attacks. The main motivation in adopting this
approach is to study:
How well any of the two watermarks survive a large
battery of attacks,
What impact embedding in the two different colour
space has on watermark invisibility,
How difficult it is to detect and potentially destroy the
watermarks,
How efficient this algorithm is in terms of processing
time.
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941274
3.1. Embedding phase
The process of watermark embedding is divided into
two phases.
Phase 1
First the image watermark ‘‘FL’’ is resized to the host
image size, in order to distribute the watermark evenly into
the host image, to increase resistance to attacks. The RGB
Image is decomposed into three matrices corresponding to
R, G and B channels, so is the resized watermark.
For each host image R, G, B matrix value, the equiva-
lent watermark R, G, B matrix weighted value (watermark
value k) is added so that:
Combined_image_R is equal to min
(original_Rþ(watermark_R k), 255)
Combined_image_G is equal to min
(original_Gþ(watermark_G k), 255)
Combined_image_B is equal to min
(original_Bþ(watermark_B k), 255)
where kis a constant that increases the strength of the
watermark. kequals to 0.01 is found to be the optimum in
this proposed algorithm. Below this value, the watermark
becomes imperceptible after extraction. Above this value,
the watermark becomes visible in the combined image
(Image 1).
Each new computed value is then recalibrated to an
integer, in the range 0 to 255, before recombining the 3
channels together to form Image 1.
For example:if original_R is equal to 254 and water-
mark_R is equal to 254, then the rounded value of
combined_image_R is equal to 256(254þ(254 0.01)),
which is clearly outside the upper range of 255. In this
case, combined_image_R is set to 255.
Phase 2
The combined RGB image (Image 1 in Fig. 16) is then
converted to YCbCr. The basic equation applied to convert
between RGB and YCbCr is shown in
Y¼0:299Rþ0:587Gþ0:114B ð1Þ
Cb ¼0:172R0:339G þ0:511B þ128
Cr ¼0:511R0:428G0:083Bþ128
The 3 component vectors (Y, Cb, Cr) are extracted. The
ASCII watermark is converted to a bit stream, using 8 bits
for each ASCII character. Although not implemented here,
the watermark can be encrypted before the binary con-
version to make it unreadable, if discovered.
Only the Cb and Cr channels are selected to embed the
watermark. Each channel is converted into a 2Dmatrix,
which is segmented into blocks. The optimal block size,
which shows the best compromise between capacity,
imperceptibility and retrieval after attacks, is defined.
The reason for using blocks is based on the fact that there
is a better statistical chance of recovering the watermark
after geometric transformation. On the other hand, this
may increase the risk of detection due to the repeated
patterns, if the block size is guessed. Setting every LSB of a
block to the same value naturally leaves a ‘‘print’’, which
may be detected using steganalysis tools for example. One
way of guessing the block size would be to use iterations
starting with a block the size of the image. After each
iteration, the block size is reduced by one pixel, until a
pattern is discovered. A practical solution to this potential
problem (although not implemented for simplicity and
clarity) would be to alternate the use of the least sig-
nificant bit (LSB) and the second least significant bit
(LSB2) for embedding the watermark, based on a private
key randomly generated, containing a stream of ones and
zeros equal to the length of the watermark binary stream,
so that for each watermark bit:
If the corresponding private key bit position is zero,
the LSB is used
If the corresponding private key bit position is one, the
LSB2 is used
This would presume that the private key is output to
the user when it is randomly created and it has to be
associated with the original image used. A relational
database storing both information within the same record
would ensure that such a record is safely kept. Each block
position is computed based on the algorithm proposed by
Lin and Delp [31].
1. For every even pixel value of the combined image, if the
watermark bit is set to 1, then the LSB of Cb is set to 0 for
each Cb element in the block. The LSB of Cr is set to 1.
2. For every odd pixel value of the combined image, if the
watermark bit is set to 1, then the LSB of Cb is set to 1 for
each Cb element in the block. The LSB of Cr is set to 0.
This is repeated until the watermark binary stream is
fully embedded. The combined image is then converted
back to RGB and saved. The conversion from YCbCr back
to RGB is shown in
R¼Yþ1:371 Cr128
ðÞ ð2Þ
G¼Y0:698 Cr128ðÞ
B¼Yþ1:732 Cb128ðÞ
An important consideration regarding this method in
the YCbCr is the limited length of the text watermark,
depending on the image size and the block size.
3.2. Extraction phase
The watermark extraction is also done in two phases.
Phase 1
The watermarked RGB image is converted to YCbCr. The
3 component vectors (Y, Cb, Cr) are extracted. Only the Cb
and Cr channels are selected to extract the watermark.
Each channel is converted into a 2D matrix, which is
segmented into blocks the same block size used during
the embedding phase. For each block in Cb and Cr channels
in even positions (even rows and columns starting at
position [0, 0]):
In Cb the following condition is tested: the count of bits
equal to 0 is greater than the count of bits equal to 1.
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1275
In Cr the following condition is tested: the count of bits
equal to 1 is greater than the count of bits equal to 0.
If the two assumptions are true, then the watermark
bit is set to 1.
If the two assumptions are false, then the watermark
bit is set to 0.
The same process is reproduced for all odd block
positions (odd rows and columns starting at position
[1,1]), but this time the opposite values are taken in each
channel, such as:
In Cb the following condition is tested: the count of bits
equal to 1 is greater than the count of bits equal to 0.
In Cr the following condition is tested: the count of bits
equal to 0 is greater than the count of bits equal to 1.
If the two assumptions are true, then the watermark
bit is set to 1.
If the two assumptions are false, then the watermark
bit is set to 0.
The watermark bit stream obtained is converted back
to ASCII characters. The watermarked image is then
converted back to RGB, before starting Phase 2.
Phase 2
The combined image RGB (containing the watermark)
is decomposed into each colour channel. The original
image is also decomposed into each colour channel.
For each R, G, B matrix value from the original image, the
equivalent watermarked image R, G, B matrix weighted
value (combined image value k) is subtracted so that:
Watermark_R¼original_R (combined_image_R k)
Watermark_G¼original_G (combined_image_G k)
Watermark_B¼original_B (combined_image_B k)
where kis the same constant used in the embedding
process (k¼0.01). Each new computed watermark value is
then recalibrated to be in the range 0 to 255, before
recombining the 3 resultant channels together to form the
original watermark. The watermark obtained is finally
resized to its original size.
After Phases 1 and 2, the two watermarks are
extracted and clearly identified. Embedding into the
YCbCr allows one to extract the ASCII watermark without
needing the original image, nor any information regarding
the watermark. This is called a blind technique. However,
the original image was needed in order to extract the
original watermark from the RGB channels. It is called a
non blind (or informed) technique. The benefit of using a
blind technique in terms of image management is appre-
ciated, as there is no need to store the original image.
3.3. Image attacks
A series of attacks have been applied for each main
digital image attack category described earlier. To do so,
some attacks have been have automated (scaling, rotation,
compression, noise addition) using MATLAB. The proposed
algorithm has also been tested against JPEG 2000
compression attacks. Manual attacks have been applied
using Macromedia Fireworks, a popular image manipulation
software used to design and enhance digital images. These
attacks cover low pass filtering, gaussian blur, change of
brightness and contrast, cropping, sharpening and histo-
gram attacks. Finally, Stirmark was also used to test against
self similarity attacks, median filter attacks and to measure
the strength of the proposed algorithm.
3.4. Measure of invisibility
Both the PSNR and the SSIM metrics will be used to
measure the visual differences between the original
image and the combined image. The PSNR has been used
extensively in the past as a reference metric. The SSIM is
much more recent and is a candidate as a reasonable
alternative.
3.5. Measure of robustness
One of the most popular difference distortion measures
is the Normalised Mean Squared Error (NMSE) metric. A
correlation coefficient, computed by matching the pixel
similarity between the original watermark and the
extracted one, is also used. Regarding the extracted ASCII
watermark, the most objective method is that the text must
match exactly what was embedded in the YCbCr in order to
prove robustness to attacks. Although presence of ASCII
characters after extraction demonstrates that ASCII text was
embedded in the first place, failure will be considered if
more than a single character mismatch occurs, as the text
may not make sense any longer.
3.6. Steganalysis
To complete this experimentation, an evaluation of
how detectable the proposed algorithm is, is performed
using Stegdetect [19] and StegSecret [16]. Both are often
cited and used in steganography research studies. They
can detect a wide range of steganography algorithms.
Both tools provide potentially interesting feedback
regarding detection, given that it is known the proposed
technique is inspired by LSB steganography.
4. Results and analysis
Here, the results of the experiments are reported and
their interpretation is given. Experiments were completed
on a computer running Microsoft Windows Vista,
equipped with an Intel Core 2 Duo 2 GHz (Gigahertz)
CPU (Central Processing Unit) and 2 GB of RAM (Gigabytes
of Random Access Memory). The strengths and weak-
nesses of the proposed algorithm were analysed, as
regards in-visibility, robustness to attacks, potential
detection using steganalysis tools and potential destruc-
tion using Stirmark. Although seven test images have
been used during experimentation, only results based on
‘‘lena’’ image will be presented in this section.
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941276
4.1. Invisibility analysis
A visual test was performed to try to detect differences
between the original image and the combined image, for
each image. As seen in Fig. 9 for the particular case of
image ‘‘lena’’, the original image is ‘‘4.2.0.4-lena.tiff’’ and
the watermark image (after embedding ‘‘fl_small.jpg’’ and
the ASCII text ‘‘Copyright LYIT—2011’’) is watermarke-
d_img.bmp. Ten persons, chosen randomly amongst
family and friends, were given one minute to look
at the two images only, shown in Fig. 9. Any evidence
that could distinguish the original from the watermarked
Table 1
PSNR and SSIM results.
Colour space RGB RGB YCbCr YCbCr Hybrid Hybrid
Images PSNR SSIM PSNR SSIM PSNR SSIM
Crown 39.3 0.9909 48.29 0.9972 38.77 0.9882
Girl 40.36 0.9988 48.86 0.9981 39.81 0.9967
Lena 39.47 0.9991 48.55 0.9982 38.95 0.9972
Plane 38.83 0.9995 47.93 0.9976 38.31 0.9971
Boat 39.36 0.9989 48.34 0.9985 38.88 0.9973
Pepper 38.94 0.9916 48.01 0.9982 38.43 0.9899
Baboon 39.47 0.9995 48.56 0.9995 38.96 0.9989
Table 2
Watermark survival results after JPEG compression.
Compression (%) (M) NMSE RExtracted watermarks
50.025421 0.84
45 0.24498 0.64
85 0.24067 0.30
95 0.22021 0.09
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1277
image was previously removed, such as the file name
which appears below each image, in Fig. 9.Each
person was then asked if they could find any differences
between the two images, and if yes to point them out. Four
out of the ten persons thought that the two images were
the same. The remaining six suspected the two images
were not the same, but when asked to point out the
differences, no one was able to point them out with 100%
conviction. 100% conviction means without any hesitation
nor change of opinion. Having completed this simple
subjective test, the PSNR and SSIM values were also
computed, after individual watermarks were embedded,
and then with the hybrid watermark. We measured the
PSNR and SSIM of the combined image after embedding
the image watermark in the RGB only, the PSNR and SSIM
of the combined image after embedding the ASCII water-
mark in the YCbCr only and the PSNR and SSIM of the
combined image after embedding both watermarks.
A SSIM value of 1 means the two images compared are
identical. A PSNR value greater than 35 dB means the two
images compared are not visibly different. Table 1 sum-
marises the results. Looking at the results, the first observa-
tion to make is that there is very little difference between the
PSNR values, from one image to another, within each colour
space, although each image tested has different character-
istics to the others. Embedding in the YCbCr only, also shows
a higher PSNR value compared to the RGB only. We believe
this is because LSB embedding in the YCbCr has less visual
impact than the additive method in the RGB and the image
watermark binary payload is significantly greater than the
ASCII watermark binary payload.
The SSIM values are all around 0.99, showing no
appreciable difference between the original image and
the watermarked image, regardless of the embedding
technique used and the colour space chosen. These results
tally with the visual inspection, in demonstrating the
invisibility of the proposed watermarking scheme. The
comparable results, despite using very different images
(from the set of seven images), also suggests that the
proposed hybrid watermarking algorithm should remain
invisible regardless of the kind of image used.
4.2. Robustness analysis
This section analyses each watermark survival after
attacks against the water-marked image ‘‘lena’’. For each
attack, the NMSE and the Correlation coefficient (R) are
computed. Both metrics are used to evaluate the distor-
tion of the watermark image (FL logo) after attacks. As
well as these two metrics, a visual inspection on the
extracted watermarks is performed to judge their survi-
val. Attacks are performed using tools such as MATLAB:
annotated with (M), Macromedia Fireworks: annotated
with (F) and Stirmark: annotated with (S).
4.2.1. JPEG compression attack
JPEG is one of the most widely used compression algo-
rithms and any watermarking system should be resilient to
some degree of compression. JPEG compression with differ-
ent quality factors are applied to the watermarked image
‘‘lena’’. Table 2 summarises the results. It can be clearly seen
that the ASCII text watermark has not survived any level of
JPEG compression. However the watermark image ‘‘FL’’ is still
perceivable up to 85% of compression (15% JPEG quality
factor). At this level, the image has lost its commercial value.
A plotted NMSE and correlation coefficient values are illu-
strated in Fig. 10. The correlation coefficient values computed
and graphed in Fig. 10, reflect very precisely what can be
observed of the watermark image after extraction. In light of
these results, the proposed algorithm is robust against JPEG
compression up to 85% compression.
4.2.2. JPEG 2000 compression attack
JPEG 2000 is another kind of compression algorithm
which uses wavelet instead of the DCT. JPEG 2000
Compressor 1.0 [1], was used for the purpose of this test.
Different quality factors were applied to the watermarked
image ‘‘lena’’. Table 3 summarises the results. It can be
clearly seen that the ASCII text watermark has not
survived any level of JPEG 2000 compression. However
the watermark image ‘‘FL’’ is still perceivable at 90% of
data lost in the image (or 10% JPEG 2000 quality factor).
Fig. 10. NMSE and correlation values of watermark at various JPEG quality factors.
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941278
A plotted NMSE and correlation coefficient values are
illustrated in Figs. 11–13. The correlation graph in Fig. 11
confirms that even after applying a compression ratio of
90%, the watermark image is still clearly identifiable. This
demonstrates the proposed algorithm is resistant to JPEG
2000 compression.
Fig. 11. NMSE and correlation values of watermark at various JPEG 2000 quality factors.
Fig. 12. NMSE and correlation values of watermark at various noise addition levels.
Fig. 13. NMSE and correlation values of watermark at various image resizing levels.
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1279
4.2.3. Noise addition attack
Noise has been added to the watermarked image lena at
varying degrees, 100% meaning all pixels were modified,
50% meaning half of the number of pixels were modified
and so on. The results are illustrated in Table 4. The ASCII
text watermark does not survive the noise addition at 100%
(all pixels modified), but the watermark image extracted is
clearly identifiable at this level of noise addition. This is
confirmed by a correlation coefficient of 0.86. In all other
cases, both watermarks survive noise addition attacks.
4.2.4. Resizing attacks
To perform this attack, the watermarked image lena is
first down-sized by a percentage and then up-sized to the
original image size, losing information in the process. Table 5
shows the results. Attacks are measured by the percentage
of original image size reduction. For example a value of
90 (90% of the original image) means the attack reduces the
image size by 10%. Table 5 clearly shows that the ASCII text
watermark does not survive any resizing levels. However, the
watermark image ‘‘FL’’ is still recognisable after resizing at
Table 3
Watermark survival results after JPEG compression.
Compression (%) (M) NMSE RExtracted watermarks
10 0.015541 0.94
50 0.24503 0.79
80 0.24283 0.63
90 0.23991 0.53
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941280
20% of the original image size. At this level however, the
image has clearly lost its commercial value.
4.2.5. Rotation attacks
Rotation clockwise or anti-clockwise, by a very small
amount (0.1 degree), is usually enough to disturb the entire
bit map, to such an extent that the embedded information
might be lost, but the image commercial value remains.
Rotation attacks (between þ1.1 and 1 degree) have been
performed and the results are summarised in Table 6.
Between 1and1degreerotation,theASCIItextwater-
mark survives the attack. Outside of this range of attacks, it
becomes unreadable. The extracted watermark image is
however not very distinct after any attack. Therefore it can
be concluded that the proposed watermarking scheme is
moderately resistant to rotation attacks. Clockwise rotations
and anti-clockwise rotations of more than one degree do not
allow recovery of either of the two watermarks.
Table 4
Watermark survival results after noise addition.
Attack level (%) (M) NMSE RExtracted watermarks
100 0.00032 0.86
50 0.0087 0.94
33 0.0117 0.95
25 0.01278 0.96
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1281
4.2.6. Filtering and histogram attacks
All these attacks were performed using Macromedia
Fireworks. Brightness and contrast were increased slightly
(by a value of 1) so that the commercial quality of the
image remained. From the results in Table 7, it can be
seen that both watermarks have survived the attack. The
combined Hue Saturation and Lightness (HSL) attack
(increase by a factor of 1 for each) shows that only the
watermark image survives. The Gaussian blur filter
attack also demonstrates that only the watermark image
survives. The Histogram attack destroys the ASCII text
watermark and the watermark image is barely identifi-
able. The Sharpening attack also destroys the ASCII water-
mark. However the watermark image is well preserved
(Table 8).
4.2.7. Median filter and self similarity attacks
These two attacks are grouped here because they
are filter attacks both performed using Stirmark. The
self similarity test is performed on RGB, YUV, HSV
or LAB colour space. A mask is defined to select which
channel to attack. In this case, ‘‘s for spatial’’ was chosen
Table 5
Watermark survival results after resizing attacks.
%of original image (M) NMSE RExtracted watemarks
90 0.024304 0.88
50 0.024494 0.79
20 0.024514 0.68
10 0.24526 0.46
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941282
Table 6
Watermark survival results after rotation.
Angle (degree) (M) NMSE RExtracted watermarks
00.024512 0.63
10.24095 0.08
1.1 0.23955 0.03
0.5 0.24327 0.27
10.23932 0.12
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1283
when defining the test. As can be seen in Table 9, the
watermark image is completely destroyed after
both attacks. However the ASCII text watermark is pre-
served 100% in the case of the self similarity attack.
This might be explained by the fact that this test is
performed on RGB, YUV, HSV or LAB but not on YCbCr.
The ASCII watermark is only preserved at 72%, in the case of
the median filter attack. Therefore, it can be conclusively
deduced that the proposed algorithm is resistant to Stirmark
self similarity test. It is resistant to Stirmark median filter
attack but to a lesser degree (Table 10).
4.2.8. Cropping attacks
Macromedia Fireworks (F) was used to conduct these
attacks. For each cropping phase, the image centre is
preserved and only a percentage of the outer part of the
image is removed. In doing so, the visually significant part
of the image is assumed to be located towards the centre
Table 7
Watermark survival results after attacks using Fireworks.
Attacks (F) NMSE RExtracted watermarks
Brightness contrast 0.0049176 0.88
HSL 0.019349 0.75
Gaussian blur 1 0.24438 0.58
Gaussian blur 2 0.24519 0.37
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941284
of the image. Cropping at 10% means 90% of the water-
marked image remains. It is worth mentioning that the
pixels of both the watermarked image and the image
were realigned (synchronised) after attack. There is a
strong linear correlation between the level of cropping
and the amount of watermark extracted. This is explained
by the fact that the watermark ‘‘FL’’ is spread evenly
across the image, increasing its chance of survival.
Table 8
Watermark survival results after attacks using Fireworks (continued).
Attacks (F) NMSE RExtracted watermarks
Histogram 0.012551 0.37
Sharpening 0.015023 0.80
Table 9
Watermark survival results after Stirmark attacks.
Attacks (S) NMSE RExtracted watermarks
Median filter 0.0 0
Self similarity 00
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1285
4.2.9. Collage attacks
The Letterkenny Institute of Technology logo was
borrowed from their web-site (www.lyit.ie), with their
agreement. It was applied on different areas of the water-
marked ‘‘lena’’ image. The LYIT logo suited this experi-
ment well because of its strong colour contrast with the
‘‘lena’’ image. Also because of its width, nearly equivalent
to the original image width, the impact of the collage on
the original watermark (if any) should not go un-noticed.
Results in Table 11 clearly show that the LYIT logo
overwrites the RGB pixel values that it covers. It also
partially replaces 3 characters (‘‘ght’’ with ‘‘uUT’’) of the
ASCII watermark, when applied in the area of the image
where embedding occurred (collage top). Based on these
results, it is reasonable to assume that the proposed
watermarking scheme will resist collage attacks unless
most of the original image is covered, in which case, its
original commercial value would be lost.
4.2.10. Clipping attacks
Results in Table 12 demonstrate that clipping any area of
the image will remove any watermark bits present.
However, because the watermark image ‘‘FL’’ is spread
across the entire image, it is very likely to survive clipping
attacks, unless a very large portion of the original image is
clipped, in which case, its original commercial value would
be lost.
4.2.11. Remarks
In most cases, either the ASCII text watermark or the
watermark image ‘‘FL’’ survives after attack. In the case of
Stirmark tests and rotation attacks, it is mainly the ASCII
Table 10
Watermark survival results after cropping attacks.
Attack level (%) (F) NMSE RExtracted watermarks
10 0.00913 0.94
25 0.00813 0.74
501 0.0068 0.65
Table 11
StegDetect detection results.
Image name Original Watermarked
Lena jphide(
nn
) jphide(
nn
)
Crown jphide(
nnn
) jphide(
nn
)
Girl jphide(
nn
) negative
Plane jphide(
nn
) jphide(
nnn
)
Boat skipped (false positive) skipped (false positive)
Peppers jphide(
n
) skipped (false positive)
Baboon skipped (false positive) skipped (false positive)
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941286
watermark that is recovered, while only the watermark
image FL survives the JPEG and JPEG 2000 compression
attacks and most filter attacks. In the case of noise addition,
both watermarks are recovered, with the exception of the
100% noise addition which destroys the ASCII watermark.
These results emphasise the need for such an hybrid algo-
rithm. Test results conducted after geometrical attacks in
particular demonstrate the importance to synchronise pixels
of original images with pixels of images obtained after
attacks. A slight mis-alignment would destroy the watermark
image ‘‘FL’’. Finally, looking at all graph results, it can be
observed that the coefficient Ris a much more useful marker
of watermark image fidelity, before and after attack, than the
NMSE. Although the NMSE is widely used in the literature,
significant variations of its values have been noticed, depend-
ing on the type of attack performed. This demonstrates, in
this particular study, that the NMSE may not be such a useful
metric.
4.3. Security analysis
During this testing, after watermark extraction, a visible
pattern on the watermark image itself has been noticed, as
seen in Figs. 14 and 15. This pattern is characterised by the
presence of light blue squares and light yellow squares in the
upper half of the image. This pattern is a direct result of the
algorithm applied when embedding the ASCII watermark in
the YCbCr. Although it is only visible when extracting the
watermark image ‘‘FL’’, it was necessary to find out if the
tested embedded technique was easily detectable using
steganography techniques. To do so, StegDetect and StegSe-
cret have been used. With each steganalysis tool, each
original images was tested first to set a proper benchmark.
Each image was then tested after the watermarks were
embedded. A sensitivity of 5 for StegDetect was chosen,
which is the mid-range sensitivity level. The results are listed
in Tables 13 and 14. It is interesting to note that at a
sensitivity level of 5, most of the original images, which are
not supposed to contain any hidden information, report a
positive detection to the ‘‘jphide’’ method. This raises ques-
tions regarding the accuracy of StegDetect. However, apart
from the image plane, which shows an increased detection
probability that the ‘‘jphide’’ algorithm is used, results
Table 12
StegSecret detection results.
Image name Original Watermarked
Lena No detection No detection
Crown No detection No detection
Girl No detection No detection
Plane No detection No detection
Boat Hiderman program detected! Hiderman program detected!
Peppers No detection No detection
Baboon Hiderman program detected! Hiderman program detected!
Fig. 14. NMSE and correlation values of watermark at various image rotations.
Fig. 15. Visible pattern on the extracted watermark image.
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1287
Fig. 16. Embedding and extraction time.
Table 13
Watermark survival results after collage attacks.
Region (F) NMSE RExtracted watermarks
Bottom 0.01482 0.93
Centre 0.01486 0.89
Top 0.01478 0.94
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941288
between the original image and the watermarked image are
similar for the other images. These results demonstrate that
the proposed algorithm is not detected accurately by
StegDetect.
Similarly to the previous experiment, the original images
boat and baboon show a positive detection of the Hiderman
program. Unfortunately, no useful information could be
found on Hiderman’s algorithm, apart from the fact that it
is used to protect privacy by hiding information in many
different file formats. The result when running the boat
original and watermarked image displays a message saying
Steganography found at marker position 15,547. The result
when running the baboon original and watermarked image
displays a message saying Steganography found at marker
position 178,438. This also raises questions regarding the
detection accuracy of StegSecret. Overall, StegSecret has not
managed to detect the proposed watermarking scheme.
Finally, the Stirmark PSNR test was used to measure the
strength of the proposed watermarking method. If after
performing this test, the two watermarks can be extracted
and they are positively identifiable, the strength of this
water-marking technique will be proven. Using Stirmark,
starting from 0, each step is incremented by 10 until 100 are
reached. Table 15 shows only the significant steps. Using
Stirmark as a benchmark to measure the strength of the
proposed algorithm, it can be concluded that only the ASCII
text watermark is identifiable throughout each step, showing
the strength of the proposed embedding technique in the
YCbCr. However the technique used to embed in the RGB is
weak, judging by Stirmark standards, as the watermark
image is not recoverable from step 20 upwards.
4.4. Capacity analysis
Capacity is more a concern of steganographers rather
than those implementing watermarking techniques. As
such, it is not essential to embed a lengthy watermark in
order to uniquely identify an image to its author. For
example, a unique social security number associated to a
country identifier and a time stamp or an image name is
all that is required to provide uniqueness.
The proposed embedding technique uses two
watermarks:
1. An ASCII text which is short: 20 characters in total,
each character is 8 bit, therefore 160 bits long.
2. An watermark image: 50 by 50 pixels, each pixel is
converted to 8 bit before embedding, therefore 20,000
bits long.
Due to the algorithm used to embed in the YCbCr,
limits are placed on the ASCII watermark length that can
be used. If applied to an image of size 512 by 512 pixels
and a block of size 19 by 19 pixels, the maximum number
of watermark text characters to embed would be 84.
Using smaller block size would increase watermark capa-
city but to the detriment of robustness, as it was observed
during experimentation. Overall the proposed algorithm
provides for a reasonable size watermark, which can be a
simple logo combined with a short ASCII text of up to 20
characters long, or two pieces of text of 2500 characters
and 20 characters respectively.
Table 14
Watermark survival results after clipping attacks.
Region (F) NMSE RExtracted watermarks
Top 0.011551 0.92
Bottom 0.011874 0.94
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1289
4.5. Complexity analysis
Algorithm complexity refers to processing power
required to embed and extract the watermark. This is an
important factor that any commercial entity would take
into consideration, particularly if the volume of images to
protect is significant. In order to evaluate the cpu time in
seconds (computer resource) that the proposed algorithm
is using, the time it takes to embed and extract the
watermarks for each image has been computed, on a
computer running a dual core 2 GHz processor and 4
gigabytes of memory on a 32 bits Operating System
running Windows 7. Results are available in Table 16
and graphed in Fig. 16.
Regardless of the image used, the time it takes to
embed is approximately equal to the time it takes to
extract the two watermarks, which is around 0.3 to 0.4 s.
This is significantly faster (four times faster) than the 1.2 s
it took to embed in the frequency domain using the afore
mentioned DWT method. As observed during the tests,
measuring the robustness of a watermark is a difficult
task to achieve: the range of distortion is almost infinite
Table 15
Watermarking strength.
PSNR test NMSE RExtracted watermarks
10 0 0.64
20 0 0.10
50 00
90 00
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941290
and difficult to model or define [21]. Having said that, a
wide range of attacks have been measured, giving a very
comprehensive idea on how well the proposed algorithm
can withstand these attacks. Due to the hybrid technique
proposed, the proposed watermarking algorithm can
resist a wide range of attacks, with the exception of a
useful but limited resistance to rotation and histogram
attacks. In general, it was observed that when one water-
mark was destroyed, the other remained intact, so in a
sense they are complementary. However it is very difficult
to predict what an attacker will do and how well this
algorithm can resist a combination of attacks, as the
variety of such possible attacks is large. In light of the
experimentation results, this hybrid algorithm does not
impair the image quality, is fairly secure and is efficient in
terms of processing power required to implement. During
the tests, no improvement (nor deterioration) in robust-
ness was observed, by embedding first in the RGB rather
than the YCbCr colour space. This suggests that embed-
ding in the RGB and the YCbCr does not significantly affect
one over the other.
4.6. Watermarking algorithm comparison
Table 17 summarises the comparison of watermark
recovery after attacks, between the proposed algorithm
and a DWT technique [32] applied in 3 different colour
spaces (YCbCr, XYZ and Cielab). Table 18 provides a
comparison of the PSNR values and SSIM values.
Table 19 shows metrics for other tests performed on the
hybrid technique. It is worth mentioning that attacks
studied in the aforementioned study are in areas where
frequency based algorithms perform better traditionally
than spatial methods. As expected from previous studies,
the watermark integrity after compression, filtering and
Table 17
Image attack result comparison between the proposed technique and [32].
Attacks Liu and Chou [32] Proposed scheme
YCbCr XYZ Cielab RGB YCbCr NMSE R
Original watermark
JPEG compression (CR¼32) 0.024 0.79
Low pass filtering 00
Gaussian noise 0.024 0.37
Scaled down (x4) 0.024 0.68
Scaled up (x4) 0.024 0.74
Table 16
CPU time used to embed and extract the watermarks.
Images Embedding time (s) Extraction time (s)
Lena 0.33 0.42
Crown 0.37 0.36
Girl 0.37 0.41
Plane 0.31 0.36
Boat 0.33 0.34
Peppers 0.34 0.34
Baboon 0.36 0.36
Table 18
PSNR/SSIM comparison between the proposed technique and [32].
Image: Lena Liu and Chou [32] Proposed scheme
YCbCr XYZ Cielab RGB YCbCr Hybrid
PSNR 40.38 40.99 41.44 39.47 48.55 38.95
SSIM Not provided 0.9991 0.9982 0.9972
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1291
scaling attacks is good in particular when embedded in
the YCbCr. So are the results of the proposed technique, in
particular where the watermark is embedded in the RGB
and to a lesser extend when the image is subjected to a
low pass filter. This may be explained by the fact that the
black and white watermark ‘‘FL’’ is resized to the original
image before embedding. In addition, each RGB compo-
nent of the original image pixel is modified by the same
small coefficient. The combination of these two factors
may contribute to a better survival even after JPEG
compression, where traditionally spacial domain methods
are the worst performer. The strong performance of the
proposed method lies in the fact that it has shown good
probability of recovery of the watermark across a much
wider range of attacks. No other technique studied could
show resistance to such an array of attacks. Furthermore,
no other watermarking scheme studied provides informa-
tion regarding potential detection to steganalysis tools.
5. Conclusion
The aim was to propose a watermarking algorithm,
inspired by steganography techniques, to hide the watermark
so that it is undetectable and thus harder to remove or
destroy. Results show that the proposed hybrid watermark-
ing technique is undetectable to visual inspection and
steganalysis tools failed to detect it. This demonstrates that
the hybrid algorithm is unnoticed, even when two separate
watermarks are used (with one of them containing a
significant number of bytes to hide). The hybrid watermark-
ing method can withstand levels of geometric and processing
attacks, up to a point where the commercial value of the
images tested would be lost. In fact, no other studies have
demonstrated robustness to such a large array of attacks.
It is also interesting to note that despite the current trend,
which favours frequency domain embedding (DWT in parti-
cular) over spatial domain embedding, it has been
Table 19
Various image attack results for the proposed technique.
Attacks Proposed scheme
RGB YCbCr NMSE R
JPEG 2000 (CR¼50) 0.024 0.79
Rotation 1deg. 0.024 0.08
Brightness/Contrast 0.004 0.88
Sharpening 0.015 0.80
Stirmark self similarity test 00
Cropping 10% 0.009 0.94
Collage 0.014 0.93
Clipping 0.011 0.92
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941292
demonstrated that the use of spatial domain techniques can
perform very well in terms of robustness, while being more
efficient in terms of processing. On the security aspect, the
proposed hybrid algorithm was benchmarked against Stir-
mark, showing good resilience of the ASCII watermark to the
PSNR test, at all levels of attacks. The stringent Stirmark test
removed the watermark image ‘‘FL’’, but the ASCII watermark
remained. The proposed method possessed invisibility,
robustness, security, capacity and complexity characteristics
as opposed to the majority of methods which tend to focus
only on invisibility and robustness. It also requires a low
processing power which increases the commercial viability of
any watermarking scheme.
References
[1] F. Alturki, R. Mersereau, Robust oblivious digital watermarking
using image transform phase modulation, Image Processing 2
(2000) 84–87.
[2] F. Autrusseau, P. Le Callet, Signal Processing 6 (87) (2007)
1363–1383.
[3] K. Bailey, K. Curran, An evaluation of image based steganography
methods, Multimedia Tools and Applications 30 (1) (2006) 55–88.
[4] H. Bayley, The Lost Language of Symbolism, Citadel Press, 1966,
London, UK.
[5] A. Cheddad, Digital image steganography: survey and analysis of
current methods, Signal Process 90 (3) (2010) 727–752. /http://dx.
doi.org/10.1016/j.sigpro.2009.08.010S, ISSN: 0165-1684.
[6] A. Cheddad, J. Condell, K. Curran, P. Mc Kevitt, A skin tone detection
algorithm for an adaptive approach to steganography, Signal
Processing 89 (2009) 2465–2478.
[7] J.S. Cho, et al., Enhancement of robustness of image watermarks
embedding into coloured image, based on WT and DCT, Information
Technology: Coding and Computing (2000) 483–488.
[8] E. Chrysochos, V. Fotopoulos, A. Skodras, M. Xenos, Reversible
image watermarking based on histogram modification, in: Proceed-
ings of the 11th Panhellenic Conference on Informatics with
International Participation , vol. B, 2007, pp. 93–104.
[9] M.U. Celik, et al., Lossless generalised-LSB data embedding, IEEE
Transactions on Image Processing (2005) 253–265.
[10] W.J. Chen, C.C. Chang, T.H. Ngan Le, High payload steganography
mechanism using hybrid edge detector’, Expert Systems with
Applications 37 (2010) 3292–3301.
[11] C. Chemak, M.S. Bouhlel, J.C. Lapayre, A new scheme of robust
image watermarking: the double watermarking algorithm, in:
Proceedings of the 2007 Summer Computer Simulation Conference,
San Diego, California, 2007, pp. 1201–1208.
[12] P. Chi-Man, A novel DFT-baseddigital watermarking system forimages,
in: Proceedings of the Signal Processing, 2006 8th International
Conference, vol. 4, Beijing, China, 2006.
[13] J. Cong, A novel watermarking algorithm for resistant geometric
attacks using feature points matching, Information Management &
Computer Security 14 (1) (2006) 75–98.
[14] I.J. Cox, Secure spread spectrum watermarking for multimedia, IEEE
Transactions Image Processing (1997) 1673–1687.
[15] I.J. Cox, M.L. Miller, J.A. Bloom, Digital Watermarking, Morgan
Kaufmann, San Fransisco, CA, 2001.
[16] I.J. Cox, M.L. Miller, J.A. Bloom, Digital Watermarking and Funda-
mentals, Morgan Kaufmann, San Fransisco, CA, 2002.
[17] N. Cvejic, T. Seppanen, Increasing robustness of LSB audio stegano-
graphy by reduced distortion LSB coding, Journal of Universal
Computer Science (2005) 56–65.
[18] S. Daly, A.B. Watson, Digital Images and Human Vision, The MIT
Press, Boston, USA 1993.
[19] S. Emek, M. Pazarci, Additive vs. image dependent DWT-DCT based
watermarking, in: B. Gunsel, A. Jain, A. Murat Tekalp, B. Sankur (eds.),
Multimedia Content Representation, Classification and Security, 2006,
pp. 98–105.
[20] S. Ghannam, F.E.Z. Abou-Chadi, WPT versus WT for a Robust
Watermarking Technique, IJCSNS International Journal of Compu-
ter Science and Network Security 9 (1) (2009).
[21] L. Gilani, A. Khan, A. Mirza, Distortion Estimation in Digital Image
Watermarking using Genetic Programming, 2006.
[22] R.C. Gonzalez, R.E. Woods, Digital Image Processing, Third edition,
Prentice Hall, 2008, New York, USA.
[23] R.C. Gonzalez, R.E. Woods, S.L. Eddins, Digital Image Processing
using MATLAB, Second edition, Gatesmark Publishing, USA, 2009.
[24] C. Huang, J. Wu, A watermark optimization technique based on
genetic algorithms, in: Proceedings of the SPIE Electronic Imaging,
San Jose, CA, 2000.
[25] A. Khan, A.M. Mirza, Genetic Perceptual Shaping: Utilizing cover
image and Conceivable Attack Information During Watermarking
Embedding, Elsevier, 2005, Amsterdam, Holland.
[26] B.S. Kim, et al., Robust digital watermarking method against
geometrical attacks, Real-Time Imaging (2003) 139–149.
[27] I. Kitamura, S. Kanai, T. Kishinami, Copyright protection of vector
map using digital watermarking method based on discrete Fourier
transform, Geoscience and Remote Sensing Symposium 3 (2001)
1191–1193.
[28] G.C. Langelaar, I. Setyawan, R.L. Lagendijk, Watermarking digital
image and video data. A state of the art overview, IEEE Signal
Processing Magazine 17 (5) (2000) 20–46.
[29] G.C. Langelaar, R.L. Lagendijk, Optimal differential energy water-
marking of DCT encoded images and video, IEEE Transactions on
Image Processing (2001) 148–158.
[30] X. Li, J. Wang, A steganographic method based upon JPEG and
particle swarm optimization algorithm, Information Sciences 177
(15) (2007) 3099–31091.
[31] E. Lin, E. Delp, Spatial synchronisation using watermark key
structure, in: Proceedings of the SPIE Conference Security, Stegano-
graphy and Watermarking of Multimedia Contents, San Jose, CA,
2004.
[32] K.C. Liu, C.H. Chou, Robustness comparison of color image water-
marking schemes in uniform and non-uniform color spaces, Inter-
national Journal of Computer Science and Network Security 7 (7)
(2007) 239–247.
[33] D.C. Lou, J.M. Shieh, H.X. Tso, A robust buyer–seller watermarking
scheme based on DWT, International Journal of Pattern Recognition
and Artificial Intelligence (2006) 79–90.
[34] D.C. Lou, N.I. Wu, C.M. Wang, Z.H. Lin, C.S. Tsai, A novel adaptive
steganography based on local complexity and human vision sensi-
tivity, Journal of Systems and Software (2010) 1236–1248.
[35] J.Q. Ni, et al., A RST-invariant robust DWT-HMM watermarking
algorithm incorporating Zernike moments and template, in: Pro-
ceedings of the Knowledge-Based Intelligent Information and
Engineering Systems, Pt 1, 2005, p. 1233–1239.
[36] N. Nikolaidisa, I. Pitas, A method for watermark casting on digital
images, IEEE Transactions on Circuits and Systems for Video
Technology (1998) 775.
[37] M.L. Miller, G.J. Doerr, I.J. Cox, Applying informed coding and
embedding to design a robust high-capacity watermark, IEEE
Transactions on Image Processing (2004) 792.
[38] S.P. Mohanty, P. Guturu, A novel invisible color image water-
marking scheme using image adaptive watermark creation and
robust insertion-extraction, ACM Transactions on Multimedia
Computing Communications and Applications 5 (2) (2008) 1–24.
(Art. 12).
[39] F.A.P. Petitcolas, R.J. Anderson, Evaluation of copyright marking
systems, in: Proceedings of the IEEE International Conference on
Multimedia Computing and Systems, ICMS ‘99, I, Florence, Italy,
June 1999, pp. 574–579.
[40] A. Piva, et al., DCT-based watermark recovering without resorting
to the uncorrupted original image, in: Proceedings of the Interna-
tional Conference Image Processing, vol. 1, 1997, pp. 520–523.
[41] F. Qi, J. Wu, G. Shi, Extracting regions of attention by imitating the
human visual system, in: Proceedings of the IEEE International
Conference on Acoustics, Speech and Signal Processing, icassp,
2009, pp.1905–1908.
[42] C. Qiang, T.S. Huang, Optimum detection and decoding of multi-
plicative watermarks, IEE Transactions on Image Processing (2000)
1123–1129.
[43] K.R. Rao, P. Yip, Discrete Cosine Transform: Algorithms, Advantages
Applications, Academic Press, Boston, 1990.
[44] D. Sal Diaz, M. Grana Romay, Introducing a watermarking with
multi-objective genetic algorithm, in: Proceedings of the 2005
Conference on Genetic and Evolutionary Computation, GECCO’05,
2005.
[45] M. Schlauweg, D. Profrock, B. Zeibich, E. Muller, Self-synchronizing
robust textel watermarking in Gaussian scale-space, in: Proceed-
ings of the 10th ACM workshop on Multimedia and security,
MM&Sec 2008, Oxford, UK, ACM 2008, 2008.
F. Lusson et al. / Signal Processing 93 (2013) 1268–1294 1293
[46] F.Y. Shih, S.Y. Wu, Combinational image watermarking in the spatial
and frequency domains, Pattern Recognition 36 (2003) 969.
[47] F.Y. Shih, Digital Watermarking and Steganography—Fundamentals
and Techniques, CRC Press, 2008, Florida, USA.
[48] TERA Consultants, ‘Building a Digital Economy: The importance of
Saving Jobs in the EU’s Creative Industries, March 2010. Referencing,
/http://www.teraconsultants.fr/assets/publications/PDF/2010-Mar
s-Etude_Piratage_TERA_full_report-En.pdfS(last visited 17/03/
2011).
[49] M.J. Tsai, K. Yu, Y.Z. Chen, Joint wavelet and spatial transformation
for digital watermarking, IEEE Transactions on Consumer Electro-
nics (2000) 241.
[50] S. Voloshynovskiy, S. Pereira, V. Iquise, T. Pun, Attack modelling:
towards a second generation watermarking benchmark, Signal
Processing 81 (6) (2001) 1177–1214.
F. Lusson et al. / Signal Processing 93 (2013) 1268–12941294
