Chapter 7

Digital Watermarking

Advances in the development of digital data and the Internet have resulted in
changes in the modern way of communication. A digital multimedia content, as
opposed to an analog one, does not lose quality due to multiple copying processes.
However, this advantage of digital media is also their major disadvantage in terms
of copyright and the unauthorized use of data.
Cryptographic methods and digital watermarking techniques have been intro-
duced in order to protect the digital multimedia content. Cryptography is used to
protect the content during transmission from sender to recipient. On the other hand,
digital watermarking techniques embed permanent information into a multimedia
content. The digital signal embedded in the multimedia data is called digital
watermark. A watermarking procedure can be used for the following purposes:
ownership protection, protection and proof of copyrights, data authenticity protec-
tion, tracking of digital copies, copy and access controls.
The general scheme of watermarking is shown in Fig. 7.1. In general, a
watermarking procedure consists of watermark embedding and watermark detec-
tion. Although the low watermark strength is preferable in order to meet the
imperceptibility requirement, one must ensure that such a watermark is detectable,
as well. This can be achieved by using an appropriate watermark detector.
Watermark embedding can be based on additive or multiplicative procedures. In
multiplicative procedures, the watermark is multiplied by the original content.
Watermark detection can be blind (without using the original content) or
non-blind (in the presence of the original content).

7.1 Classification of Digital Watermarking Techniques

A number of different watermarking techniques have been developed. Most of them

can be classified into one of the categories given in Fig. 7.2. From the perceptual
aspect, the watermark can be classified as either perceptible or imperceptible.

© Springer International Publishing Switzerland 2016 349

S. Stanković et al., Multimedia Signals and Systems,
DOI 10.1007/978-3-319-23950-7_7
350 7 Digital Watermarking

Fig. 7.1 A block scheme of a watermarking procedure

Fig. 7.2 Classification of digital watermarking techniques

Noticeable watermark visibly changes the original content. It is sometimes used to

protect images and videos, but generally it is not very popular nowadays. This
technique involves embedding characters that uniquely identify the owners of the
content and appear as a background image, or as a visible sign. However, water-
mark of this type can be removed. Almost all currently used techniques fall into the
class of imperceptible techniques.
Imperceptible techniques are further divided into robust techniques, semi-fragile
and fragile watermarking techniques. Fragile watermarking assumes embedding of
certain watermark that will be significantly damaged or removed in an attempt to
modify the content. These techniques are useful in proving the data authenticity. In
semi-fragile watermarking, the watermark should be resistant to certain signal
7.2 Common Requirements Considered in Watermarking 351

processing techniques (e.g., compression), while it is fragile under any other attack.
However, the most commonly used techniques are based on the robust
watermarking and are considered in the next sections.
Robust techniques involve embedding a watermark in the original signal, such
that the watermark removal causes serious degradation of the signal quality.
Watermark should be designed to be robust to the standard signal processing
approaches (compression, filtering, etc.), as well as to intentional attempts to
remove the watermark.
Classification in terms of the embedding domain
Watermarking techniques are further divided by the domains in which the water-
mark is embedded. Namely, the watermark can be embedded directly in the signal
domain, or in one of the transform domains. The choice of the watermarking
domain depends on the type of multimedia data and the watermarking application.
The most frequently used transform domains are based on the DFT, DCT, and DWT
transforms. The transform domain watermarking is more convenient for modeling
the spectral characteristics of watermark according to the human perceptual model.
For highly nonstationary signals, the modeling can be achieved by using time-
frequency transforms.

7.2 Common Requirements Considered in Watermarking

Depending on the application and the type of data to be watermarked, the

watermarking procedure should fulfill a number of requirements. In the sequel,
we discuss some general and very common watermarking requirements.
1. The watermark should be accessible only to the authorized users. This issue is
referred as security of the watermarking procedure and it is generally achieved
by using cryptographic keys.
2. The watermark detectability should be assured regardless of the conventional
signal processing or malicious attacks that may be applied.
3. Generally, although one should provide an unremovable watermark, it should be
imperceptible within the host data.
4. The watermark should convey a sufficient amount of information.
As stated above, the first requirement is related to the security of the watermark
and watermarking procedure in general. In some applications, the specific security
keys (which can be encrypted) are used during the watermark embedding and
extraction. If the watermark is created as a pseudo-random sequence, then the
key used to generate a sequence can be considered as a watermarking key.
The next requirement is watermark robustness, which is one of the main
challenges when designing the watermarking procedure. The watermark should
be robust not only to the standard signal processing techniques, but also to the
malicious attacks aiming to remove the watermark. All algorithms that may lead to
352 7 Digital Watermarking

Table 7.1 Common attacks in audio and image watermarking procedures

Attacks
Audio watermarking Image watermarking
Resampling Requantization
Wow and flutter JPEG compression
Requantization Darkening
mp3 with constant bit rate Lightening
mp3 with variable bit rates Mean filter (of size 3
3, 5
5, 7
7)
Pitch scaling Median filter (of size 3
3, 5
5, 7
7)
Audio samples cropping Image cropping
Echo and time-scale modifications Image resize
Filtering Rotation
Amplitude normalization Adding noise Gaussian or impulse

the loss of the watermark information are simply called attacks. Some of the
common examples are compression algorithms, filtering, change of the data format,
noise, cropping signal samples, resampling, etc. The list of commonly present
attacks for audio signals and images is given in Table 7.1.
Perceptual transparency is one of the most important requirements. Watermark
should be adapted to the host content, and should not introduce any perceptible
artifacts or signal quality degradations. However, the imperceptibility is usually
confronted with the watermark robustness requirement. In order to be imperceptible,
the watermark strength should be low, which directly affects its robustness. Hence,
an efficient watermarking procedure should always provide a trade-off between the
imperceptibility and robustness. In order to perform the watermark embedding just
below the threshold of perception, various masking procedures can be employed.
In some applications it is desirable that the watermark convey a significant number
of bits, which will be extracted by detector. Hence, it is sometimes required that the
watermark data rate (payload) is high. The property that describes the ability to embed a
certain amount of information is known as a capacity of the watermarking algorithm.
Besides the general watermarking requirements discussed above, there could be
some specific requirements, as well, related to the following issues:
– Real-time implementation,
– Complete extraction/reconstruction of the watermark at the decoder,
– The absence of the original data during the watermark extraction (blind
extraction), etc.

7.3 Watermark Embedding

This section considers the additive and multiplicative watermark embedding

techniques.
7.3 Watermark Embedding 353

Additive embedding techniques can be defined as:

I w ¼ I þ αw; ð7:1Þ

where I represents the vector of signal samples or transform domain coefficients

used for watermarking, Iw is the vector of watermarked coefficients, w is the
watermark, while the parameter α controls the watermark strength. If the parameter
α should be adjusted to the signal coefficients then the watermark embedding can be
written as:

I w ¼ I þ αðI Þw: ð7:2Þ

Another frequently used approach is multiplicative embedding, given by the

relation:

I w ¼ I þ αwI: ð7:3Þ

In order to provide that the watermark does not depend on the sign of selected
watermarking coefficients, a modified version of Eq. (7.3) can be used:

I w ¼ I þ αwjI j: ð7:4Þ

Multiplicative watermarking is often used in the frequency domain, to ensure that

the watermark energy at a particular frequency is proportional to the image energy
at that frequency. An additional advantage of multiplicative watermark embedding
is that it is difficult to estimate and remove watermark by averaging a set of
watermarked signals, which is one of the common attacks.
Let us consider an example of robust image watermarking in the transform domain.
The two-dimensional DFT of an image is shown in Fig. 7.3a, while its
two-dimensional DCT is illustrated in Fig. 7.3b. Note that the DCT is real and has
only positive part of the spectrum, making it suitable for applications in watermarking.

Fig. 7.3 (a) The DFT of “Baboon” image, (b) The DCT of “Baboon” image
354 7 Digital Watermarking

Fig. 7.4 (a) Original “Lena” image, (b) watermarked “Lena” image

The region marked by the red line corresponds to low-frequency coefficients,

which contain most of the image energy. Consequently, the modification of
these coefficients can cause significant image quality degradation. Therefore, the
low-frequency coefficients are usually avoided in watermarking. Outside the
blue circle, we have high-frequency components. These components carry certain
image details and can be filtered out, without significant image degradation.
Therefore, the high-frequency components are also often omitted in watermarking.
It follows that the watermarking should be done in the middle frequency part
(between the blue and red circles in Fig. 7.3).
Consider the sorted DCT coefficients of an image. Given the nature of the DCT
transform, it is necessary to omit the first L coefficients (which are mostly the
low-frequency components) and choose the next M coefficients (mostly belonging
to the middle frequencies). Watermarking is then performed as:

I w ðiÞ ¼ I ðiÞ þ αjI ðiÞjwðiÞ for i ¼ L þ 1, L þ 2, . . . , L þ M; ð7:5Þ

where I(i) denotes a DCT coefficient of an image. The watermark w can be created
as a pseudo-random sequence. The inverse DCT is then applied to obtain the
watermarked image. The original and watermarked “Lena” images are shown in
Fig. 7.4 (peak signal to noise ratio PSNR ¼ 47 dB).

7.4 Watermark Detection

7.4.1 Hypothesis Testing Approach

The goal of each algorithm for watermark detection is to provide a reliable proof of
the watermark presence within the signal. Denote by Ix a set of coefficients on
which the watermark detection is performed (Ix can be either Iw or I depending
whether the watermark is present or not), and the watermark is w. A general
7.4 Watermark Detection 355

approach for watermark detection is based on a hypothesis testing problem. The

assumptions are:
H0: Ix does not contain watermark w,
H1: Ix contains watermark w.
The problem of watermark detection is based on a reliable threshold, used to
decide whether a watermark is present or not. The threshold is determined by defining
a criterion that ensures a minimum probability of detection error. Since the water-
mark detection can be viewed as a detection of signal in noise, the likelihood ratio is
used to minimize the error. Detection errors can occur in two cases: G10—when the
assumption of H0 is accepted as true, although the correct hypothesis is H1, G01—
when the assumption H1 is accepted as true, but the correct hypothesis is H0.
The criterion that determines the presence of the watermark is defined as
follows:


1, I x 2 R1 ,
ΦðI x Þ ¼ ð7:6Þ
0, I x 2 R0 ;

where R1 and R0 are regions in which the assumptions H1 and H0 are tested. In order
   ratio lis  defined
to minimize error during detection, a likelihood  by using the
conditional probability density functions p I x H 1 and p I x H0 :
  
p I x H 1
lðI x Þ ¼    : ð7:7Þ
p Ix H0

The minimum probability of error will be achieved when the region R1 is deter-
mined as:


p0 P01
R1 ¼ I x : lðI x Þ > ; ð7:8Þ
p1 P10

where p0 and p1 are a priori known probabilities of the assumptions H0 and H1

occurrence, while P01 and P10 are decision weights associated with G01 and G10,
respectively. The criterion for the detection can be written as:
8
< p P01
1, lðI x Þ > 0
ΦðI x Þ ¼ p1 P10 ð7:9Þ
: 0, otherwise:

Therefore, the detection is done by comparing the likelihood ratio with:

p0 P01
λ¼ : ð7:10Þ
p1 P10
356 7 Digital Watermarking

The threshold λ can be set to minimize the total probability of error that occurs
during detection:

Pe ¼ p0 P f þ p1 ð1  Pd Þ; ð7:11Þ

where Pf is the probability that the watermark is detected, when in fact it is not
present (false alarm), and (1  Pd) is the probability of watermark misdetection.
The error minimization procedure is commonly performed under the assumption
that P01 ¼ P10 and p0 ¼ p1, or in other words for λ ¼ 1. It means that the probabil-
ities of false alarm Pf and misdetection Pm ¼ (1  Pd) are the same.
In practice, we usually have a predefined maximum false alarm probability from
which the threshold λ is calculated as follows:

ð
1
 
p lH0 dl ¼ P f ; ð7:12Þ
λ

 
where p lH 0 is the pdf of l under H0. After the threshold λ is determined, the
probability of misdetection is calculated as:

ðλ
 
Pm ¼ p lH1 dl: ð7:13Þ
1

7.4.1.1 Additive White Gaussian Model

Let us consider the procedure to minimize the detection error in the case of additive
white Gaussian noise (AWGN) model, which is the simplest one encountered in
practice. This model assumes that the coefficients are uncorrelated and have a
Gaussian distribution. Note that the watermark is considered as a noisy signal:

I x ¼ I þ w þ n; ð7:14Þ

where Ix, I and w are the coefficients of the watermarked content, the original
content and the watermark, respectively. The watermarked content can be modified
in the presence of attack, which is modeled by noise n (white Gaussian noise).
Under the assumption that the original coefficients, as well as the noise coefficients,
are uncorrelated and follow the Gaussian distribution, Eq. (7.14) can be written as
follows:

I x ¼ I n þ w: ð7:15Þ

In also has the Gaussian distribution with the modified mean value and the variance
compared to the original content I. Now, the previously defined hypothesis can be
written as:
7.4 Watermark Detection 357

H0 : Ix ¼ In
H 1 : I x ¼ I n þ w:
  
p I x H 1
In order to minimize the similarity measure lðI x Þ ¼    ; it is necessary to know
p Ix H0
the conditional probability density function, which in the case of the Gaussian
distribution is defined as:

   Y
2
M ðI x ðiÞμx wðiÞÞ
1 
p I x H 1 ¼ pffiffiffiffiffiffiffiffiffiffie 2σ 2x
2πσ 2
i¼1 x
ð7:16Þ
   Y M
1 
ðI x ðiÞμx Þ2

p Ix H0 ¼ pffiffiffiffiffiffiffiffiffiffie 2σ 2x ;
2πσ x 2
i¼1

where μx is the mean value of signal coefficients used in watermark detection, while
M is the length of watermark. Now the measure of similarity is calculated as:

YM ðI ðiÞμx wðiÞÞ2
 x
   e 2σ 2
x
p I x H 1

lðI x Þ ¼    ¼ M
i¼1
: ð7:17Þ
p Ix H0 Y ðIx ðiÞμ2 x Þ2
e 2σ x

i¼1

Equation (7.17) can be written in a simplified form by applying the logarithmic

function:

X n
1 h 2 2
i
‘ðI x Þ ¼ ð I x ð i Þ  μ x Þ  ð I x ð iÞ  μ x  w ð iÞ Þ
2σ 2
i¼1 " x # ð7:18Þ
1 X n Xn X n
¼ 2 2I x ðiÞwðiÞ  2μx wðiÞ  w ði Þ ;
2
2σ x i¼1 i¼1 i¼1

where ‘(Ix) indicates the natural logarithm function of l(Ix). Note that the last two
terms within the brackets do not depend on Ix. Therefore, the term representing
linear correlation of Ix and w is used as a watermark detector:

X
M
D¼ I x ðiÞwðiÞ; ð7:19Þ
i¼1

which is optimal under the considered assumptions and is called the standard
correlator. In the case when the signal statistics is not distributed according to the
Gaussian distribution, other detector forms can be used.
358 7 Digital Watermarking

Fig. 7.5 Illustration of the errors that may occur in watermark detection

According to the procedure for determining the general detection threshold λ, we

can now determine the threshold for the standard correlator as:

ð
1
  
p DH0 dD ¼ P f ; ð7:20Þ
Tp

  
where p DH0 is the pdf of detector responses D under H0. The pdf of D under H0
and H1 are illustrated in Fig. 7.5. If the response of the detector is D < Tp, we
conclude that the watermark is not present, and vice versa. In the case of equal
probabilities Pf ¼ 1  Pd, the optimum threshold is A/2 (Fig. 7.5).
In order to determine the threshold and the probability of error, we need to know
how the watermark is embedded, the statistical characteristics of the image coeffi-
cients, as well as the characteristics of attacks.

7.4.2 A Class of Locally Optimal Detectors

According to the signal detection theory, it is difficult to define a general test that
maximizes the signal detection probability. Also, it is known that for detection of
weak signals a locally optimal detector can be created (in our case a watermark
signal is weak in comparison to the host signal). It is defined as follows:

D ¼ gLO ðI x Þ  w; ð7:21Þ

where gLO is the local optimum nonlinearity, defined by:

p0 ð I x Þ
gLO ðI x Þ ¼  ; ð7:22Þ
pðI x Þ

with p(Ix) and p0 (Ix) indicating the coefficients probability density function and its
derivative, respectively. Note that, the detector contains the nonlinear part gLO,
which is correlated with the watermark signal. If the coefficients have the Gaussian
distribution, the proposed detector corresponds to the standard correlator.
7.4 Watermark Detection 359

7.4.2.1 The Most Commonly Used Distribution Functions

and the Corresponding Detector Forms

The coefficients distribution for most images can be modeled by the Gaussian,
Laplace, generalized Gaussian or Cauchy distribution functions. For example,
recall that the generalized Gaussian function can be defined as:

αβ ðβjxμjÞα 1 Γð3=αÞ 1=2
GGF ¼ e , α > 0, β ¼ : ð7:23Þ
2Γð1=αÞ σ Γð1=αÞ

For α ¼ 1, this function is equal to the Laplace distribution, and for α ¼ 2 it is equal
to the Gaussian distribution. Figure 7.6 shows the coefficients distribution of an
image. The form of the detector, which corresponds to the generalized Gaussian
distribution, is given by:

X
M
D1 ¼ signðI x ðiÞÞjI x ðiÞjα1 wðiÞ; ð7:24Þ
i¼1

γ
while the detector form for Cauchy distribution, CF ¼ π , is equal to:
ðγ2 þðxδÞ2 Þ

X
M
2ðI x ðiÞ  δÞ
D2 ¼ wðiÞ: ð7:25Þ
i¼1 ðI x ði Þ  δ Þ2 þ γ 2

Note that x (in the pdf) corresponds to the watermarked coefficients Ix in the
detector form (M is the length of watermarked sequence and watermark). It is

Fig. 7.6 The distribution of

coefficients: Gaussian
(green line) and Laplace
distribution (purple line)
360 7 Digital Watermarking

important to emphasize that the locally optimum detector form can be quite
sensitive to the pdf variations.
A simple measure of detection quality can be defined as:

Dwr  Dww
R ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi; ð7:26Þ
σ wr 2 þ σ ww 2

where D and σ are mean values and standard deviations of detector responses, while
the indices wr and ww are used for right keys (watermarks) and wrong keys (wrong
trials), respectively. The wrong trial is any sequence which is not the watermark,
but is generated in the same way.

7.4.3 Correlation Coefficient and Similarity Measure

In order to determine the similarity between the original watermark w and the
watermark w* extracted from the protected data at the detection side, we can use the
similarity measure defined as follows:

w  w∗
Simðw, w∗ Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi : ð7:27Þ
w  w∗

The similarity measure is usually given in the form of the correlation coefficient,
which can be calculated as:

X
M
wðiÞw* ðiÞ
i¼1
ρðw, w*Þ ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð7:28Þ
X M X M  2
ðwðiÞÞ2 w* ðiÞ
i¼1 i¼1

where M is the length of watermark.

7.5 Examples of Watermarking Procedures

7.5.1 Audio Watermarking Techniques

Audio watermarking procedures are mainly based on the specific audio signal
characteristics and psychoacoustics. In the next subsections, a brief description of
audio watermarking approaches such as the spread-spectrum audio watermarking,
two-sets method and echo embedding, is provided.
7.5 Examples of Watermarking Procedures 361

7.5.1.1 Spread-Spectrum Watermarking

Spread-spectrum watermarking is an example of correlation based method that

assumes a pseudo-random sequence embedding, where the standard correlator is
used for detection. This is a commonly used watermarking approach. The pseudo-
random sequence, r(n), i.e., the wide-band noise sequence, can be embedded in the
time or in the transform domain. This sequence is used to modulate the binary
message υ ¼ f0; 1g, or equivalently b ¼ f1, 1g. The watermarked sequence wðnÞ
¼ br ðnÞ obtained in this way, is scaled according to the energy of the host signal s
(n), to provide a compromise between the watermark imperceptibility and robust-
ness. The watermark embedding can be done, for example, by an additive proce-
dure: sw ðnÞ ¼ sðnÞ þ αwðnÞ: A suitable pseudo-random sequence should have good
correlation properties, in the way that it should be orthogonal to the other pseudo-
random sequences. The commonly used sequence is called the m-sequence (max-
imum length sequence), and its autocorrelation is given by:
(
1MX1 1, for k ¼ 0,
wðiÞwði  kÞ ¼ 1 ð7:29Þ
M i¼0  , for k 6¼ 0:
M

7.5.1.2 Two Sets Method

This blind audio watermarking procedure is based on the two sets A and B of audio
samples. A value d (watermark) is added to the samples within the set A, while it is
subtracted from the samples in B:

a*i ¼ ai þ d, b*i ¼ bi  d;

where ai and bi are samples from A and B, respectively.

h i When making decision
*
about watermark presence, the expected value E a  b is employed, where ā*
*

*
and b are mean values of samples ai and bi . This method is based on the
assumption that the mean values of the samples from different signal blocks are


the same, i.e., that E a  b ¼ 0 holds (which may not be always the case in the
practice). Only in this case, the watermark can be detected as:
h i
   
*
E a*  b ¼ E ða þ dÞ  b  d ¼ E a  b þ 2d ¼ 2d: ð7:30Þ

7.5.1.3 Echo Embedding

The echo embedding procedure can be realized according to:

xðnÞ ¼ sðnÞ þ αsðn  dÞ; ð7:31Þ

362 7 Digital Watermarking

where d represents a certain delay of the echo signal. The extraction of the
embedded echo requires the detection of delay d. The signal copy is usually delayed
for approximately 1 ms. The echo amplitude is significantly lower than the original
signal amplitude, and hence, the signal quality is not degraded. On the contrary, the
sound is enriched. There is also a variant of this procedure, where two delays are
considered: one is related to the logical value “1”, while the other is related to “0”.
The double echo embedding operation can be written as:

xðnÞ ¼ sðnÞ þ αsðn  dÞ  αsðn  d  ΔÞ; ð7:32Þ

where the difference between delays corresponding to “1” and “0” is denoted by Δ,
and its value does not exceed four samples. The delay detection is done by using the
cepstrum autocorrelation, which is the inverse Fourier transform of the -
log-magnitude spectrum. The complexity of cepstrum calculation is one of the
main disadvantages of this method.

7.5.1.4 Watermarking Based on the Time-Scale Modifications

Time-scale modifications are related to compressing and expanding of the time

axis. The basic idea of time-scale watermarking is to change the time scale between
two successive extremes (maximum and minimum). The interval between two
extremes is divided into N segments with equal amplitudes. The signal slope is
changed within a certain amplitudes range according to the bits that should be
embedded. Namely, the steep slope corresponds to bit “0”, while the mild slope
corresponds to bit “1”.

7.5.2 Image Watermarking Techniques

A simple watermarking algorithm for digital image protection is based on the

additive watermark embedding procedure in the 8
8 DCT domain. First, an
image is divided into 8
8 blocks of pixels as in the case of JPEG algorithm.
Then, the 2D DCT transform is applied to each block separately. The watermark is
embedded into the set of selected coefficients. In order to provide a good compro-
mise between the watermark imperceptibility and robustness, the coefficients are
selected from the middle frequency region, as illustrated in Fig. 7.7.
Watermark embedding is based on the standard additive procedure: I w ¼ I þ αw;
where I denotes the original middle-frequency DCT coefficients (from 8
8 block),
while Iw are the watermarked DCT coefficients. Next, we perform the inverse DCT
transform that results in watermarked 8
8 block. This is repeated for each block.
Watermark detection can be performed by using a standard correlation detector
(assuming that the distribution of selected coefficients can be modeled by the
Gaussian function). However, more accurate modeling can be obtained by using
7.5 Examples of Watermarking Procedures 363

Fig. 7.7 A region of

middle frequency DCT
coefficients within 8
8
block (shaded in gray)

generalized Gaussian and Cauchy function, where the corresponding detectors D1

and D2 (defined by Eqs. (7.24) and (7.25)) are used for detection.

7.5.3 The Procedure for Watermarking of Color Images

Unlike the previous procedure, where the block-based DCT is performed, here we
use the 2D DCT transform of the entire image. The procedure is described in the
sequel.
(a) The selection of coefficients for watermark embedding is done through the
following steps:
1. The color channels are separated (e.g., R, G, B), Fig. 7.8.
2. 2D DCT is computed for each color matrix
3. The matrices of DCT coefficients are transformed into vectors and sorting
operation is performed
4. The largest L coefficients are omitted and the next M coefficients are
selected for watermarking.
(b) Watermark embedding
Let us denote the sorted DCT coefficients by I, while w is the watermark
created as a pseudo-random sequence. The watermarked DCT coefficients are
calculated as:
 
I w ðiÞ ¼ I ðiÞ þ α  I ðiÞ  wðiÞ, i ¼ L þ 1, . . . , M þ L;

where i denotes the coefficient position in the sorted sequence.

(c) Reorder the sequence into matrix form.
(d) Calculate the 2D inverse DCT (with rounding to integer values).
364 7 Digital Watermarking

Fig. 7.8 Color image “Lena” and the separated color channels

7.5.4 An Overview of Some Time-Frequency Based

Watermarking Techniques

The time-frequency based watermarking can be used for different types of multime-
dia data: audio signals, images and video signals. The time-frequency domain can be
efficient regarding the watermark imperceptibility and robustness. Namely, the
watermark with specific time-frequency characteristics can be designed and adapted
to the host signal components, which enhances the efficiency of the watermarking
procedure. Note that the time-frequency representations defined for one-dimensional
signals can be extended to two-dimensional case in order to be applied to images. In
this case, they are usually referred as the space/spatial-frequency representations.
1. The watermark can be created with specific space/spatial-frequency character-
istics, while its embedding can be done even in the space domain. This approach
is among the first space/spatial-frequency based image watermarking proce-
dures. Namely, a two dimensional chirp signal is used as watermark:
  
W ðx; yÞ ¼ 2A cos ax2 þ by2 ¼ A e jðax þby Þ þ e jðax þby Þ :
2 2 2 2
ð7:33Þ
7.5 Examples of Watermarking Procedures 365

The watermark is embedded within the entire image:

I w ðx; yÞ ¼ I ðx; yÞ þ W ðx; yÞ: ð7:34Þ

It is interesting to observe that multiple different chirps with small amplitudes

can be used for watermarking. The parameters of the chirp signals and the
random sequence that defines the amplitudes of chirps, serve as the watermark
key. Since the watermark is embedded within the entire image in the spatial
domain, a proper masking that provides imperceptibility should be applied. Note
that the Wigner distribution provides an ideal representation for the chirp signal.
Hence, the watermark detection is performed by using a form of the Radon–
Wigner distribution:
 
P ωx ; ω y ; W v ¼ jFT 2D ðI w ðx; yÞW v ðx; yÞÞj2 ¼
ð 1 ð 1 2
  jðxωx þyω y Þ
 ð7:35Þ
 I ð x; y ÞW ð x; yÞe dxdy  ;
 w v 
1 1

where FT2D denotes the 2D Fourier transform, while:

W v ðx; yÞ ¼ e jðav x þbv y2 þcv xyÞ

2
: ð7:36Þ

Different values of parameters av, bv, and cv define a set of projection planes.
The additional term cvxy is used to detect some geometrical transformations, as
well. In order to make a decision about the watermark presence within the
image, the maxima of the Radon–Wigner distribution are calculated:
 
Mðav ; bv ; cv Þ ¼ max P ωx ; ω y ; W v ; ð7:37Þ
ωx , ω y

and compared with a reference threshold. This procedure provides robustness to

various attacks, some being a median filter, geometrical transformations (trans-
lation, rotation and cropping simultaneously applied), a high-pass filter, local
notch filter and Gaussian noise.
2. Digital audio watermarking can be done by using time-frequency expansion and
compression. The audio signal is firstly divided into frames of size 1024 sam-
ples, where the successive frames have 512 samples overlapping. If the original
frame is lengthened or shortened, the logical value 1 is assigned, otherwise the
“normal frames” corresponds to the logical value 0. The watermark is a sequence
obtained as a binary code of the alphabet letters, converted to the ASCII code.
The frames with signal energy level above a certain threshold are selected. The
signal is transformed to frequency domain and a psychoacoustic model is used to
determine the masking threshold for each selected frame. The frames length is
changed in frequency domain by adding or removing four samples with ampli-
tudes that do not exceed the masking threshold. It prevents a perceptual distor-
tion. In order to preserve the total signal length, the same number of expanded
366 7 Digital Watermarking

and compressed frames is used (usually an expanded frame is followed by a

compressed frame). The detection procedure is non-blind, i.e., the original signal
is required. The difference between the original and watermarked samples in
time domain will have diamond shape for the pair expanded-compressed frame
(Diamond frames), while the difference is flat and close to zero for unaltered
frames. The pair of Diamond frames is used to represent the binary 1, while the
logical values 0 are assigned to the unaltered frames. Hence, it is possible to
detect binary values, and consequently the corresponding alphabetical letters.
3. A spread spectrum based watermarking in the time-frequency domain
The watermark is created as:

wi ðnÞ ¼ aðnÞmi ðnÞ pi ðnÞ cos ðω0 ðnÞnÞ; ð7:38Þ

where mi(n) is the watermark before spreading, pi(n) is the spreading code or the
pseudo-noise sequence (bipolar sequence taking the values þ1 and 1 with
equal probabilities), while ω0 is the time-varying carrier frequency which
represents the instantaneous mean frequency of the signal. The parameter a(n)
controls the watermark strength. The masking properties of the human auditory
system are used to shape an imperceptible watermark. The pseudo-noise
sequence is low-pass filtered according to the signal characteristics. Two differ-
ent scenarios of masking have been considered. The tone- or noise-like charac-
teristics are determined by using the entropy:

X
ωmax
H¼ Pðxi Þlog2 Pðxi Þ: ð7:39Þ
i¼1

The probability of energy for each frequency (within a window used for the
spectrogram calculation) is denoted by P(xi), while ωmax is the maximum
frequency. A half of the maximum entropy Hmax ¼ log2ωmax is taken as a
threshold between noise-like and tone-like characteristics. If the entropy is
lower than Hmax, it is considered as a tone-like, otherwise it is a noise-like
characteristic.
The time-varying carrier frequency is obtained as the instantaneous mean
frequency of the host signal, calculated by:

X
ωmax
ωTFDðn; ωÞ
ω¼0
ωi ð n Þ ¼ : ð7:40Þ
X
ωmax
TFDðn; ωÞ
ω¼0

The instantaneous mean frequency is computed over each time window of the
STFT, and the TFD(n,ω) is the energy of the signal at a given time and
frequency.
7.5 Examples of Watermarking Procedures 367

Finally, after the watermark is modulated and shaped, it is embedded in the

time domain as: swi ðnÞ ¼ si ðnÞ þ wi ðnÞ. During the detection, the demodulation
is done by using the time-varying carrier and then the watermark is detected by
using the standard correlation procedure.
4. Watermarking approach based on the time-frequency shaped watermark
In order to ensure imperceptibility constraints, the watermark can be modeled
according to the time-frequency characteristics of the signal components. For
this purpose the concept of nonstationary filtering is adapted and used to create a
watermark with specific time-frequency characteristics. The algorithm includes
the following steps:
1. Selection of signal regions suitable for watermark embedding;
2. Watermark modeling according to the time-frequency characteristics of the
host signal;
3. Watermark embedding and watermark detection in the time-frequency
domain.
Due to the multicomponent nature of multimedia signals (e.g., speech sig-
nals), the cross-terms free time-frequency distributions (TFD) should be used,
such as the spectrogram and the S-method. If a region selected from the TFD is:

D ¼ fðt; ωÞ : t 2 ðt1 ; t2 Þ, ω 2 ðω1 ; ω2 Þg; ð7:41Þ

we can define a time-frequency mask as follows:

1 for ðt; ωÞ 2 D and jTFDðt; ωÞj > ξ,
LM ðt; ωÞ ¼ ð7:42Þ
0 for ðt; ωÞ 2
= D or jTFDðt; ωÞj < ξ:

The parameter ξ is a threshold which can be calculated as a portion of the TFD

maximum: ξ ¼ λ10λlog10 ðmaxðjTFDðt;ωÞjÞÞ (λ is a constant). The mask LM contains the
information about the significant components within the region D. Hence, if we
start with an arbitrary random sequence p, the modeled watermark is obtained at
the output of the nonstationary (time-varying) filter:
X
wðtÞ ¼ LM ðt; ωÞSTFT p ðt; ωÞ; ð7:43Þ
ω

where STFTp stands for the short-time Fourier transform of p. The watermark
embedding is done according to:

STFT Iw ðt; ωÞ ¼ STFT I ðt; ωÞ þ STFT wkey ðt; ωÞ; ð7:44Þ

where Iw, I and w are related to the watermarked signal, original signal and
watermark, respectively.
368 7 Digital Watermarking

The watermark detector can be made by using the correlation in the time-
frequency domain:

X
M
D¼ i
STFT wkey STFT Iiw : ð7:45Þ
i¼1

Note that the time-frequency domain provides a larger number of coefficients

for correlation (compared to time or frequency domains), which enhances the
detection performance.

7.6 Examples

7.1. Consider a vector with a few image DFT coefficients chosen for watermarking.
DFT ¼ [117 120 112 145 136 115].
The watermarking procedure should be done in the following way:
(a) Sort the vector of DFT coefficients.
(b) Add a watermark given by w ¼ [3.5 2 4 5 9 7].
(c) Assume that the sequence wrong ¼ [3 2 5 7 2 4] provides the highest
response of the correlation based detector among large number of wrong
trials (wrong keys) used for testing.
(d) Prove that the watermark can be successfully detected by using the
standard correlator.
Solution:

DFT sort ¼ ½ 112 115 117 120 136 145 ;

DFT w ¼ DFT sort þ w ¼ ½ 108:5 113 121 125 145 138 :

In order to ensure a reliable watermark detection using the standard correlator,

the detector response for the watermark should be higher than the maximal
detector response when using wrong trials:

ΣDFT w  w > ΣDFT w  wrong,

ΣDFT w  w ¼ 842:25,
ΣDFT w  wrong ¼ 86:5:

Having in mind the results, we may conclude that the watermark detection is
successful.
7.2. Write a program in Matlab which perform the image watermarking as follows:
(a) Calculate and sort the DCT coefficients of the considered image;
7.6 Examples 369

Fig. 7.9 Results of

watermark detection

(b) Create a watermark as a pseudo-random sequence, e.g.,

watermark ¼ 0.5rand(1500,1);
(c) After omitting the strongest 1500 coefficients, embed the watermark into
the next 1500 coefficients by using the multiplicative procedure with
α ¼ 0.8;
(d) Check if the watermark is imperceptible within the protected image;
(e) Perform the watermark detection in the DCT domain by using the stan-
dard correlator. It is necessary to demonstrate that the detector response
for the watermark is higher than the detector response for any of the
100 wrong trials (Fig. 7.9).
Solution:
alfa¼0.8;
Det¼zeros(2,100);
image¼imread(’lena512.bmp’);
image¼image(1:2:512,1:2:512);
N¼256;
DCT1¼dct2(image);
Vector¼DCT1(:);
[g,v]¼sort(abs(Vector));
watermark¼0.5*rand(1500,1);
Vectorwat¼Vector;
Vectorwat(v(N*N-1500-1500þ1:N*N-1500))¼Vector(v(N*N-1500-
1500þ1:N*N-1500))þalfa*abs(Vector(v(N*N-1500-1500þ1:N*N-
1500))).*watermark;
DCTwat¼DCT1;
DCTwat(:)¼Vectorwat;
370 7 Digital Watermarking

imagewat¼idct2(DCTwat);
figure,imshow(uint8(imagewat))
DCTwat1¼dct2(imagewat);
DCTwat1¼DCTwat1(:);
x¼DCTwat1(v(N*N-1500-1500þ1:N*N-1500));
for k¼1:100
wrong¼0.5*rand(1500,1);
Det(1,k)¼sum(x.*watermark);
Det(2,k)¼sum(x.*wrong);
end
figure,
plot(1:100,Det(2,1:100),’r’,1:100,Det(1,1:100),’g’)

7.3. Consider the watermarking procedure described in the sequel. A block of the
8
8 DCT coefficients is selected. The watermark is added to the block
coefficients: I w ¼ I þ w. The watermarked image is exposed to the quantiza-
tion attack defined by the quantization matrix Q. Determine which watermark
samples will contribute to the difference between the watermarked and the
original coefficient after quantization attack.

45 20 54 81 0 0 0 0 5 4 1 3 0 0 0 0
15 77 0 11 0 0 0 0 3:5 5 0 5 0 0 0 0
21 0 0 39 0 0 0 0 3 3 2 5 0 0 0 0
27 44 52 75 0 0 0 0 0 2 0 6:5 0 0 0 0
DCT ¼ w¼
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
3 5 7 9 11 13 15 17
5 7 9 11 13 15 17 19
7 9 11 13 15 17 19 21
9 11 13 15 17 19 21 23
Q¼
11 13 15 17 19 21 23 25
13 15 17 19 21 23 25 27
15 17 19 21 23 25 27 29
17 19 21 23 25 27 29 31

Solution:
Approach I: It is possible to perform the quantization of the original and the
watermarked coefficients, to compare them and to select the coefficients that
are different after quantization.
7.6 Examples 371

15 4 8 9 0 0 0 0 17 5 8 9 0 0 0 0
3 11 0 1 0 0 0 0 4 12 0 1 0 0 0 0
3 0 0 3 0 0 0 0 3 0 0 3 0 0 0 0
3 4 4 5 0 0 0 0
DCT q ¼ DCT wq ¼ 3 4 4 5 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Hence, the positions of the selected coefficients are: (1,1), (1,2), (2,1), (2,2).
Approach II: Select the watermark samples higher than Q/2.

1:5 2:5 3:5 4:5 5:5 6:5 7:5 8:5 5 4 1 3 0 0 0 0

2:5 3:5 4:5 5:5 6:5 7:5 8:5 9:5 3:5 5 0 5 0 0 0 0
3:5 4:5 5:5 6:5 7:5 8:5 9:5 10:5 3 3 2 5 0 0 0 0
Q 4:5 5:5 6:5 7:5 8:5 9:5 10:5 11:5 2
¼ w¼ 0 0 6:5 0 0 0 0
2 5:5 6:5 7:5 8:5 9:5 10:5 11:5 12:5 0 0 0 0 0 0 0 0
6:5 7:5 8:5 9:5 10:5 11:5 12:5 13:5 0 0 0 0 0 0 0 0
7:5 8:5 9:5 10:5 11:5 12:5 13:5 14:5 0 0 0 0 0 0 0 0
8:5 9:5 10:5 11:5 12:5 13:5 14:5 15:5 0 0 0 0 0 0 0 0

The selected watermark samples will produce a difference between the

quantized original and watermarked image coefficients.
7.4. Based on the principles introduced in the previous example, the image
watermarking procedure is implemented as follows:
(a) DCT is calculated for the 8
8 image blocks
(b) The watermark w is added to the quantized coefficients:
Iq(i,j) ¼ K(i,j)Q(i,j), (K(i,j) are integers), but the coefficients quantized
to zero value are immediately omitted
(c) The selection of coefficients suitable for watermarking is done according
to the constraints:
– The watermarked DCT coefficients after quantization with Q should
have non-zero values;
– The watermarked coefficients should not be rounded to the same value
as the original coefficients.
Analyze and define the values of K(i,j) and watermark w, that satisfy the
above constraints.
Solution and explanation:
To ensure that the watermarked DCT coefficients after quantization with Q,
have non-zero values, the following relation should hold:
372 7 Digital Watermarking

Qði; jÞ
jK ði; jÞQði; jÞj  jwj  ; ð7:46Þ
2

or equivalently, the watermark should satisfy the condition:

Condition 1

Qði; jÞ
jwj  jK ði; jÞQði; jÞj  : ð7:47Þ
2

The watermarked coefficients will not be quantized to the same value as the
original coefficients if the following condition is satisfied:
Condition 2

Kði, jÞQði, jÞ þ w < Kði, jÞQði, jÞ  Q=2

or : ð7:48Þ
Kði, jÞQði, jÞ þ w  Kði, jÞQði, jÞ þ Q=2

From Eq. (7.48) we have: jwj > Qði; jÞ=2. Combining Condition 1 and
Condition 2, we get:


Qði; jÞ   Qði; jÞ
w ½jK ði; jÞj  1=2Qði; jÞ,  [ , ½jK ði; jÞj  1=2Qði; jÞ ;
2 2

Note that jK ði; jÞj  2 should hold.

7.5. Define the form of a locally optimal watermark detector, that corresponds to
the watermarked coefficients pdf, assuming that they are selected by using the
criterion jK ði; jÞj  2. The coefficients pdf can be modeled by a function
illustrated in Fig. 7.10.
Solution:
In the considered case, the coefficients pdf from Fig. 7.10 can be approxi-
mately described by using the following function:
I 2n
x

I 2n e j a j ;
2γ
 Ix
pðI x Þ ¼ a
ð7:49Þ
1þ a
x

where parameter a defines the positions of the pdf maxima, while n controls
the pdf decay between the maximum and the origin. The parameter γ is usually
equal to 1/2, 1 or 2.
A locally optimal detector can be defined as:

p0 ð I x Þ
Dopt ¼   w; ð7:50Þ
pð I x Þ
7.6 Examples 373

Fig. 7.10 The distribution

of coefficients after
omitting low-frequency
components

which in the case of the specified function p becomes:

0 1
X  2γ
M B γ 2γ1 I xi n C
Dopt ¼ wi B I
@a2γ xi sgn    2n C
A; ð7:51Þ
a I xi
i¼1 I xi 1þ a

where M is the number of watermarked coefficients.

7.6. By using the results obtained in the Example 7.4, derive the condition for
watermarked coefficients selection which would provide the robustness to a
certain JPEG quantization degree defined by the matrix Q0 . By robustness we
assume that the coefficients pdf is preserved even after quantization with
matrix Q0 , in order to provide successful detection by using locally optimal
detector. Assume that the watermark embedding is done according to:
 
I ði; jÞ
I w ði; jÞ ¼ round Qði; jÞ þ Qði; jÞw; ð7:52Þ
Qði; jÞ

where Q is the quantization matrix with high quality factor QF (i.e., a low
compression ratio).
Solution:
In order to provide robustness to the quantization defined by the matrix Q0
(quality factor QF0 ), the criterion for coefficients selection should be modified.
The watermarked coefficients will be robust after applying Q0 if they are not
rounded to zero, i.e., if the following condition is satisfied:
0
Q ði; jÞ
jK ði; jÞQði; jÞj  jQði; jÞwj > : ð7:53Þ
2
374 7 Digital Watermarking

Note that the worst case is assumed in Eq. (7.53): the coefficient and the
watermark have opposite signs. Hence, we may observe that for efficient
watermark detection, the coefficients should be selected for watermarking if:
0
Q ði; jÞ
jK ði; jÞj  w þ : ð7:54Þ
2Qði; jÞ

Therefore, if Q with an arbitrary high QF is used for watermark embedding,

the robustness is satisfied even for matrix Q0 as long as the criterion Eq. (7.54)
is satisfied. In this way the procedure provides the full control over the
robustness to any JPEG quantization degree.
Note that if the criterion is satisfied for QF0 < QF, then the watermark
detection will certainly be successful for any quantization Qx defined by QFx
for which QFx > QF0 holds.
7.7. A speech watermarking procedure in the time-frequency domain can be
designed according to the following instructions:
1. Voiced speech regions are used for watermarking.
2. Watermark is modeled to follow the time-frequency characteristics of
speech components in the selected region.
3. Watermark embedding and detection is done in the time-frequency domain
by using the S-method and time-varying filtering procedure.
(a) Design a time-frequency mask for watermark modeling and define the
modeled watermark form.
(b) Define a watermark detection procedure in the time-frequency domain
which includes the cross-terms.
Solution:
(a) By using the S-method (with L ¼ 3), a voiced speech region is selected:

D ¼ fðt; ωÞ : t 2 ðt1 ; t2 Þ, ω 2 ðω1 ; ω2 Þg;

where t1 and t2 are the start and end points in the time domain, while
frequency range is ω 2 ðω1 ; ω2 Þ. According to Eq. (7.42), the time-
frequency mask can be defined as:


1 for ðt; ωÞ2D and jSMðt; ωÞj > ξ
LM ðt; ωÞ ¼ ;
0 for ðt; ωÞ=
2D or jSMðt; ωÞj < ξ

where parameter λ within the energy floor ξ can be set to 0.7. The
illustration of speech region is given in Fig. 7.11a, the corresponding
mask is shown in Fig. 7.11b, while the time-frequency representation of
the modeled watermark is shown in Fig. 7.11c. The modeled version of
the watermark is obtained by using Eq. (7.43).
7.6 Examples 375

Fig. 7.11 (a) Speech region selected for watermarking, (b) mask function, (c) time-frequency
representation of modeled watermark

Note that the time-frequency characteristics of watermark correspond to

the speech components. Hence, it would be difficult to remove the
watermark without introducing serious signal quality degradation.
(b) The watermark detection can be performed by using the S-method with
L ¼ 32 to intentionally produce the cross-terms:

X
N X
N
,j
D¼ i
SMwkey SMxiw þ SMiwkey SMix,wj ; ð7:55Þ
i¼1 i; j¼1
i6¼ j

where the index w is related to watermark and xw to watermarked

coefficients. Although the cross-terms are usually undesirable in the
time-frequency analysis, they may increase performance of watermark
detector.
7.8. In analogy with one-dimensional case described in the previous example,
design a space/spatial-frequency based image watermarking procedure.
Note: Space/spatial-frequency representation is calculated for each pixel and it
reflects the two-dimensional local frequency content around the pixel. The 2D
form of the STFT for the window of size N
N is extended from the 1D version
as:

X
N=21 X
N=21
I ðn1 þ i1 , n2 þ i2 Þwði1 ; i2 Þe j N ðk1 i1 þk2 i2 Þ :
2π
STFT ðn1 ; n2 ; k1 ; k2 Þ ¼
i1 ¼N=2 i2 ¼N=2

Solution:
Space/spatial-frequency representation can be used for classification between
the flat and busy image regions. Namely, busy image regions are preferred
in watermarking, because it is easier to provide watermark imperceptibility.
376 7 Digital Watermarking

Fig. 7.12 (a) Busy region

and its spectrogram, (b) flat
region and its spectrogram

The examples of busy and flat image regions are shown in Fig. 7.12a, b,
respectively. Note that, unlike the busy regions, the flat regions contain
small number of significant components in the space/spatial-frequency
domain, which can be used as a criterion for regions classification.
Following analogy with the procedure for speech signals, watermark can
be modeled according to the local frequency characteristics defined by the
mask L:
XX  
wkey ðn1 ; n2 Þ ¼ STFT p ðn1 ; n2 ; ω1 ; ω2 ÞL n1 , n2 , ω1 , ω2 ; ð7:56Þ
ω1 ω2

where STFTp is a short-time Fourier transform of the two-dimensional pseudo-

random sequence. The mask is obtained as:


1 for ðω1 , ω2 Þ : jSTFTðn1 , n2 , ω1 , ω2 Þj2 > ξ,
Lðn1 , n2 , ω1 , ω2 Þ ¼
0 for ðω1 , ω2 Þ : jSTFTðn1 , n2 , ω1 , ω2 Þj2  ξ:

Watermark embedding and detection can be done in the space/spatial-

frequency domain in the same way as in the case of speech signals.
References 377

References

1. Al-khassaweneh M, Aviyente S (2005) A time-frequency based perceptual and robust

watermarking scheme. Proc. of EUSIPCO 2005, Antalya
2. Barni M, Bartolini F (2004) Watermarking systems engineering. Marcel Dekker, Inc.,
New York, NY
3. Barkat B, Sattar F (2010) Time-frequency and time-scale-based fragile watermarking methods
for image authentication. EURASIP J Adv Signal Process 2010, 408109. doi:10.1155/2010/
408109
4. Battiato S, Emanuel S, Ulges A, Worring M (2012) Multimedia in forensics, security, and
intelligence. IEEE Multimedia Mag 19(1):17–19
5. Briassouli A, Strintzis MG (2004) Locally optimum nonlinearities for DCT watermark detec-
tion. IEEE Trans Image Process 13(12):1604–1618
6. Cox IJ, Miller ML, Bloom JA (2002) Digital watermarking. Academic, London
7. Dittmann J, Megı́as D, Lang A, Herrera-Joancomartı́ J (2006) Theoretical framework for a
practical evaluation and comparison of audio watermarking schemes in the triangle of robust-
ness, transparency and capacity. In: Transactions on Data Hiding and Multimedia Security I,
vol 4300, Lecture Notes in Computer Science. Springer, Berlin, pp 1–40
8. Djurović I, Stanković S, Pitas I (2001) Digital watermarking in the fractional Fourier trans-
formation domain. J Netw Comput Appl 24(2):167–173
9. Esmaili S, Krishnan S, Raahemifar K (2003) Audio watermarking time-frequency character-
istics. Can J Elect Comput Eng 28(2):57–61
10. Foo SW, Ho SM, Ng LM (2004) Audio watermarking using time-frequency compression
expansion. Proc. of the Int. Symp. on Circuits and Systems, ISCAS 04, 3: III - 201-4
11. Hannigan BT, Reed A, Bradley B (2001) Digital watermarking using improved human visual
system model. Proc SPIE 4314:468–474
12. Heeger D (1997) Signal detection theory, Teaching Handout. Department of Psychology,
Stanford University, Stanford, CA
13. Hernandez JR, Amado M, Perez Gonzales F (2000) DCT-domain watermarking techniques for
still images: detector performance analysis and a new structure. IEEE Trans Image Process
9:55–68
14. Kang X, Huang J, Zeng W (2008) Improving robustness of quantization-based image
watermarking via adaptive receiver. IEEE Trans Multimedia 10(6):953–959
15. Katzenbeisser S, Petitcolas F (2000) Information hiding: techniques for steganography and
digital watermarking. Artech House, Norwood, MA
16. Kirovski D, Malvar HS (2003) Spread-spectrum watermarking of audio signals. IEEE Trans
Signal Process 51(4):1020–1033
17. Mobasseri BG, Zhang Y, Amin MG, Dogahe BM (2005) Designing robust watermarks using
polynomial phase exponentials. Proc. of Acoustics, Speech, and Signal Processing
(ICASSP’05), Philadelphia, PA, vol 2, pp ii/833–ii/836
18. Muharemagić E, Furht B (2006) Survey of watermarking techniques and applications,
Chapter 3. In: Furht B, Kirovski D (eds) Multimedia watermarking techniques and applica-
tions. Auerbach Publication, Boca Raton, FL, pp 91–130
19. Nikolaidis A, Pitas I (2003) Asymptotically optimal detection for additive watermarking in the
DCT and DWT domains. IEEE Trans Image Process 12(5):563–571
20. Proceedings of the IEEE: Special Issue on Identification and Protection of Multimedia
Information, vol 87, July 1999
21. Podilchuk CI, Zheng W (1998) Image adaptive watermarking using visual models. IEEE J
Select Areas Commun 16:525–539
22. Stankovic L, Stankovic S, Djurovic I (2000) Space/spatial-frequency based filtering. IEEE
Trans Signal Process 48(8):2343–2352
23. Stanković S, Djurović I, Herpers R, Stanković LJ (2003) An approach to the optimal water-
mark detection. AEUE Int J Elect Commun 57(5):355–357
378 7 Digital Watermarking

24. Stanković S, Djurović I, Pitas I (2001) Watermarking in the space/spatial-frequency domain

using two-dimensional Radon-Wigner distribution. IEEE Trans Image Process 10:650–658
25. Stankovic S, Orovic I, Chabert M, Mobasseri B (2013) Image watermarking based on the
space/spatial-frequency analysis and Hermite functions expansion. J Elect Imaging 22
(1):013014
26. Stankovic S, Orovic I, Mobasseri B, Chabert M (2012) A Robust procedure for image
watermarking based on the Hermite projection method. Automatika 53(4):335
27. Stanković S, Orović I, Žarić N (2008) Robust watermarking procedure based on JPEG-DCT
image compression. J Elect Imaging 17(4):043001
28. Stanković S, Orović I, Žarić N (2010) An application of multidimensional time-frequency
analysis as a base for the unified watermarking approach. IEEE Trans Image Process 19
(2):736–745
29. Stanković S (2000) About time-variant filtering of speech signals with time-frequency distri-
butions for hands-free telephone systems. Signal Process 80(9):1777–1785
30. Stanković S, Orović I, Žarić N (2008) Robust speech watermarking in the time-frequency
domain. EURASIP J Adv Signal Process 2008:519206
31. Steinebach M, Dittmann J (2003) Watermarking-based digital audio data authentication.
EURASIP J Appl Signal Process 2003(10):1001–1015
32. Wang FH, Pan JS, Jain LC (2009) Innovations in digital watermarking techniques. Studies in
computational intelligence. Springer, New York, NY
33. Wickens TD (2002) Elementary signal detection theory. Oxford University Press, Oxford