Digital Video Watermarking

Techniques for Secure
Multimedia Creation and
Delivery
CHAN Pik-Wah
A Thesis Submitted in Partial Fulfilment
of the Requirements for the Degree of
Master of Philosophy
in
Computer Science and Engineering
Supervised by
Prof. Michael R. Lyu
c _The Chinese University of Hong Kong
July 2004
The Chinese University of Hong Kong holds the copyright of this thesis. Any
person(s) intending to use a part or whole of the materials in the thesis in
a proposed publication must seek copyright release from the Dean of the
Graduate School.
Abstract of thesis entitled:
Digital Video Watermarking Techniques for Secure Multime-
dia Creation and Delivery
Submitted by CHAN Pik-Wah
for the degree of Master of Philosophy
at The Chinese University of Hong Kong in July 2004
There is an explosion of data exchange on the Internet and the
extensive use of digital media. Consequently, digital data own-
ers can quickly and massively transfer multimedia documents
through the Internet. It has aroused intense interest in mul-
timedia security and multimedia copyright protection. In this
paper, a comprehensive approach for protecting and managing
video copyrights with watermarking techniques is introduced.
We propose a novel hybrid digital video watermarking scheme
based on the scene change analysis, error correction code and
genetic algorithm. Our video watermarking algorithm is robust
against the attacks of frame dropping, averaging and statistical
analysis, which were not solved effectively in the past. It started
with a complete survey about current watermarking technolo-
gies. We have discovered that none of the existing schemes was
capable of resisting all attacks. Accordingly, we came up with
the idea of embedding a single watermark of different parts into
different scenes of a video. Then we analyze the strength of dif-
i
ferent watermarking schemes. A hybrid approach is applied to
form a super watermarking scheme which can resist most of the
attacks. In order to reinforce the robustness of the scheme, the
watermark is refined by an error correcting code, while the cor-
recting code is embedded as a watermark in the audio channel.
Furthermore, the fidelity of the scheme is enhanced by applying
genetic algorithm. It optimizes the quality of the watermarked
video. Also, our scheme allows blind retrieval of the embedded
watermark, which does not need the original video and the wa-
termark is perceptually invisible. The effectiveness of the scheme
is verified through a series of experiments, in which a number
of standard image processing attacks are conducted, and the ro-
bustness of our approach is demonstrated by using the criteria
of the latest StirMark test.
ii
iii
Acknowledgement
I would like to take this opportunity to express my gratitude to
my supervisor Prof. Michael R. Lyu, for his generous guidance
and patience given to me in the past two years. His numerous
support and encouragement, as well as his inspiring advice are
extremely essential and valuable in my research papers (con-
ference papers published in ICICS’2003 and WWW’2004 and
journal paper submitted to IEEE transaction TCSVT) and my
thesis.
I am also grateful for the time and valuable suggestion that
Prof. Irwin King, Prof. Tien-Tsin Wong and Prof. Roland Chin
have given in marking my term paper. Without their effort, I
will not be able to strengthen and improve my research project
and papers.
I would also like to show my gratitude to the Department of
Computer Science and Engineering, CUHK, for the provision of
the best equipment and pleasant office environment required for
high quality research.
Special thanks should be given to Mr. Edward Yau and Mr.
Sam Sze who have given me valuable suggestions, encourage-
ment and supports. And I would like to give my thanks to my
fellow colleagues, H. Y. Chan, C. H. Hoi, K. Z. Huang, K. Y.
iv
Lee, C. H. Chan, C. Y. IP, Y. Lam, C. H. Law, W. Hung, C. W.
Leung, C. W. Wong, J. Y. Zheng, Y. K. Yu, N. S. Lau and T.
H. Ng. They have given me support, and a joyful and wonderful
university life.
Finally, I am deeply indebted to my family for their uncon-
ditional love and support over the years.
v
This work is dedicated to my family for the support and
patience
vi
Contents
Abstract i
Acknowledgement iv
1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Objective . . . . . . . . . . . . . . . . . 3
1.3 Contributions . . . . . . . . . . . . . . . . . . . . 4
1.4 The Structure of this Thesis . . . . . . . . . . . . 6
2 Literature Review 7
2.1 Security in Multimedia Communications . . . . . 8
2.2 Cryptography . . . . . . . . . . . . . . . . . . . . 11
2.3 Digital Watermarking . . . . . . . . . . . . . . . . 14
2.4 Essential Ingredients for Video Watermarking . . 16
2.4.1 Fidelity . . . . . . . . . . . . . . . . . . . 16
2.4.2 Robustness . . . . . . . . . . . . . . . . . 17
2.4.3 Use of Keys . . . . . . . . . . . . . . . . . 19
2.4.4 Blind Detection . . . . . . . . . . . . . . . 20
2.4.5 Capacity and Speed . . . . . . . . . . . . . 20
2.4.6 Statistical Imperceptibility . . . . . . . . . 21
vii
2.4.7 Low Error Probability . . . . . . . . . . . 21
2.4.8 Real-time Detector Complexity . . . . . . 21
2.5 Review on Video Watermarking Techniques . . . 22
2.5.1 Video Watermarking . . . . . . . . . . . . 25
2.5.2 Spatial Domain Watermarks . . . . . . . . 26
2.5.3 Frequency Domain Watermarks . . . . . . 30
2.5.4 Watermarks Based on MPEG Coding Struc-
tures . . . . . . . . . . . . . . . . . . . . . 35
2.6 Comparison between Different Watermarking Schemes 38
3 Novel Watermarking Schemes 42
3.1 A Scene-based Video Watermarking Scheme . . . 42
3.1.1 Watermark Preprocess . . . . . . . . . . . 44
3.1.2 Video Preprocess . . . . . . . . . . . . . . 46
3.1.3 Watermark Embedding . . . . . . . . . . . 48
3.1.4 Watermark Detection . . . . . . . . . . . . 50
3.2 Theoretical Analysis . . . . . . . . . . . . . . . . 52
3.2.1 Performance . . . . . . . . . . . . . . . . . 52
3.2.2 Capacity . . . . . . . . . . . . . . . . . . . 56
3.3 A Hybrid Watermarking Scheme . . . . . . . . . . 60
3.3.1 Visual-audio Hybrid Watermarking . . . . 61
3.3.2 Hybrid Approach with Different Water-
marking Schemes . . . . . . . . . . . . . . 69
3.4 A Genetic Algorithm-based Video Watermarking
Scheme . . . . . . . . . . . . . . . . . . . . . . . . 73
3.4.1 Watermarking Scheme . . . . . . . . . . . 75
3.4.2 Problem Modelling . . . . . . . . . . . . . 76
3.4.3 Chromosome Encoding . . . . . . . . . . . 79
viii
3.4.4 Genetic Operators . . . . . . . . . . . . . 80
4 Experimental Results 85
4.1 Test on Robustness . . . . . . . . . . . . . . . . . 85
4.1.1 Experiment with Frame Dropping . . . . . 87
4.1.2 Experiment with Frame Averaging and Sta-
tistical Analysis . . . . . . . . . . . . . . . 89
4.1.3 Experiment with Lossy Compression . . . 90
4.1.4 Test of Robustness with StirMark 4.0 . . . 92
4.1.5 Overall Comparison . . . . . . . . . . . . . 98
4.2 Test on Fidelity . . . . . . . . . . . . . . . . . . . 100
4.2.1 Parameter(s) Setting . . . . . . . . . . . . 101
4.2.2 Evaluate with PSNR . . . . . . . . . . . . 101
4.2.3 Evaluate with MAD . . . . . . . . . . . . 102
4.3 Other Features of the Scheme . . . . . . . . . . . 105
4.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . 106
5 Conclusion 108
Bibliography 110
ix
List of Figures
2.1 Symmetric Cryptosystem . . . . . . . . . . . . . . 12
2.2 Asymmetric Cryptosystems . . . . . . . . . . . . 13
2.3 Watermarking Embedding and Detection Scenario 15
2.4 Classification map of existing digital video water-
mark techniques . . . . . . . . . . . . . . . . . . . 26
2.5 2 Scale 2-Dimensional Discrete Wavelet Transform 32
3.1 Overview of the watermarking process . . . . . . 43
3.2 Preprocessing the watermark . . . . . . . . . . . . 45
3.3 (a) Original watermark (b-i) Preprocessed water-
mark m
0
−m
7
(j) Encrypted watermark m
/
0
. . . 46
3.4 Scene change detection . . . . . . . . . . . . . . . 48
3.5 Embedding watermarks in a frame . . . . . . . . 49
3.6 (a) Original frame (b) Watermarked frame (c)
Extracted watermark corresponding to Figure 3.3(g)
(d) Recovered watermark. . . . . . . . . . . . . . 51
3.7 Possible improvement for scene based watermark-
ing scheme . . . . . . . . . . . . . . . . . . . . . 61
3.8 Overview of visual-audio hybrid watermarking scheme 62
3.9 (a) Original video watermark (b) Visualization of
averaging (c) Audio watermark (average of a) . . 64
x
3.10 Audio watermark embedding with MCLT . . . . . 66
3.11 One of the (a) original video frame and (b) wa-
termarked video frame . . . . . . . . . . . . . . . 67
3.12 Block of samples of the original wave content . . . 68
3.13 Block of samples of watermarked wave content . . 68
3.14 Overview of detection of the watermark . . . . . . 69
3.15 Hybrid approach with different scheme for differ-
ent scene . . . . . . . . . . . . . . . . . . . . . . . 71
3.16 Hybrid approach with different scheme for differ-
ent part of frame . . . . . . . . . . . . . . . . . . 72
3.17 The graph of three mutually orthogonal axes rep-
resenting the capacity, robustness and fidelity of
the watermarking scheme . . . . . . . . . . . . . . 73
3.18 The graph of two mutually orthogonal axes rep-
resenting the robustness and fidelity of the water-
marking scheme . . . . . . . . . . . . . . . . . . . 74
3.19 A illustrative diagram for GA-based optimization
process . . . . . . . . . . . . . . . . . . . . . . . . 77
3.20 The GA-based optimization process for part of
watermark . . . . . . . . . . . . . . . . . . . . . . 78
3.21 A 24-bit chromosome represents the sequence of
the scenes to embed . . . . . . . . . . . . . . . . . 80
3.22 The GA-based watermarking algorithm . . . . . . 82
3.23 Comparison between watermarked video with and
without GA optimization a) Original video frame
(b) Video frame watermarked with scene-based
scheme (c) Video frame watermarked with GA-
based scheme . . . . . . . . . . . . . . . . . . . . 84
xi
4.1 NC values under frame dropping . . . . . . . . . . 87
4.2 Scenario of statistical averaging attack . . . . . . 89
4.3 NC values under statistical averaging . . . . . . . 90
4.4 NC values under lossy compression . . . . . . . . 91
4.5 NC values under cropping . . . . . . . . . . . . . 94
4.6 NC values under PSNR . . . . . . . . . . . . . . . 95
4.7 NC values under different rescaling factor . . . . . 96
4.8 NC values under different noise added to the wa-
termarked video . . . . . . . . . . . . . . . . . . . 97
4.9 PSNR of the video under different GA generations 103
4.10 MAD of the video under different GA generations 104
4.11 A conceptual illustration on the performance of
the proposed scheme . . . . . . . . . . . . . . . . 106
xii
List of Tables
2.1 Basic Robustness Requirements . . . . . . . . . . 19
2.2 Classification of watermarking according to sev-
eral viewpoints . . . . . . . . . . . . . . . . . . . 24
2.3 Comparison between different watermarking schemes 39
4.1 Robustness comparison between different water-
marking schemes . . . . . . . . . . . . . . . . . . 98
4.2 Parameters Setting for GA-based experiment . . . 101
4.3 The computation time of the GA-based scheme . 102
4.4 PSNR comparison between different watermark-
ing schemes . . . . . . . . . . . . . . . . . . . . . 103
4.5 MAD comparison between different watermark-
ing schemes . . . . . . . . . . . . . . . . . . . . . 105
xiii
Chapter 1
Introduction
With the rapid growth of the Internet and multimedia systems
in distributed environments, it is easier for digital data owners
to transfer multimedia documents across the Internet. There-
fore, there is an increase in concern over copyright protection of
digital contents [1, 2, 3, 4]. Traditionally, encryption and control
access techniques were employed to protect the ownership of me-
dia. These techniques, however, do not protect against unautho-
rized copying after the media have been successfully transmitted
and decrypted. Recently, watermark techniques are utilized to
maintain the copyright [4, 5, 6, 7]. In this paper, we focus on
engaging the digital watermarking techniques to protect digital
multimedia intellectual copyright, and propose a new algorithm
particularly for video watermarking purpose.
1.1 Background
Multimedia and network security issues are classically handled
through cryptography, however, cryptography ensures confiden-
1
CHAPTER 1. INTRODUCTION 2
tiality, authenticity, and integrity only when a message is trans-
mitted through a public channel such as an open network. It
does not protect against unauthorized copying after the message
has been successfully transmitted. Digital watermarking is an
effective way to protect copyright of multimedia data even af-
ter its transmission. Watermarking is a concept of embedding a
special pattern, watermark, into a multimedia document so that
a given piece of copyright information is permanently tied to the
data. This information can later prove the ownership, identify a
misappropriating person, trace the marked document’s dissem-
ination through the network, or simply inform users about the
rights-holder or the permitted use of the data [6].
Digital watermarking remains a largely untested field. There
is only a very few number of industrial associations have pub-
lished the requirements for testing watermarking algorithms [8].
Numerous inventive watermarking approaches have been pro-
posed in these few years and most of them focus on digital im-
age watermarking. In recent years, image watermarking tech-
nique becomes mature, thus researcher starts to explore a more
challenging research topic – digital video watermarking. Most
of the proposed video watermarking schemes are based on the
techniques of image watermarking and directly applied to raw
video or compressed video. However, current image watermark-
ing schemes are not capable of adequately protecting video data
[9].
Video watermarking introduces some issues which is not present
in image watermarking. Due to large amounts of data and inher-
ent redundancy between frames, video signals are highly suscep-
CHAPTER 1. INTRODUCTION 3
tible to pirate attacks, including frame averaging, frame drop-
ping, frame swapping, statistical analysis, etc. Applying a fixed
image watermark to each frame in the video leads to problems
of maintaining statistical and perceptual invisibility. Further-
more, such an approach is necessarily video independent; as the
watermark is fixed while the frame changes. Applying indepen-
dent watermarks to each frame also presents a problem. Regions
in each video frame with little or no motion remain the same
frame after frame. Motionless regions may be statistically com-
pared or averaged to remove independent watermarks [10]. In
addition, video watermarking schemes must not use the original
video during watermark detection as the video usually is in very
large size and it is inconvenient to store it twice. We propose a
new video watermarking scheme to overcome these problems.
1.2 Research Objective
As video copyright protection is strongly concerned, a robust
video watermarking scheme is necessary. In order to design a
robust, invisible, blind and not removable video watermarking
scheme, a survey and investigation has been done on multime-
dia security issues and multimedia watermarking scheme. Vari-
ous watermarking scheme schemes are compared and evaluated.
Base on these, a new approach and procedures for multimedia
security based on watermarking are proposed [11, 12].
A Hybrid scene-based video watermarking scheme with er-
ror correcting code and genetic algorithm is proposed. In this
scheme, the watermark is decomposed into different parts and
CHAPTER 1. INTRODUCTION 4
embedded in the frames of different scenes in the video with
hybrid approaches. As identical watermark is used within each
motionless scene and independent watermarks are used for suc-
cessive different scenes, the proposed method is robust against
the attack of frame dropping, averaging, swapping, interpola-
tion and lossy compression. At the same time, error correcting
code is extracted from the video channel and embedded into the
audio channel, which provides extra information for recovery
of extracted watermark. The performance of the watermark-
ing scheme is enhanced by applying the genetic algorithm (GA)
optimization. Moreover, the scheme allows blind retrieval of
embedded watermark which does not need the original video.
Video watermarking is not a stand alone technology. It can be
associated with different applications to achieve a sophisticated
system. This research can be continuous by applying this new
developed scheme to specific environment or application and
examine its usefulness.
1.3 Contributions
Our research work has the following contributions:
• We have performed a complete survey on the current wa-
termarking technologies. It is noticed that none of the cur-
rent watermarking schemes can resist all attacks. With this
finding, we propose a hybrid watermarking scheme based
on scene change analyze [11], error correction codes and
genetic algorithm [12, 13].
CHAPTER 1. INTRODUCTION 5
• We proposed a new scheme which applies scene change de-
tections and scrambled watermarks in a video. The scheme
is robust against frame dropping, as the same part of the
watermark is embedded into the frames of a scene. For
different scenes, different parts of the watermark are used,
making the scheme robust against frame averaging and sta-
tistical analysis [11]. This scheme is innovative in attacking
the problems that are not solved effectively in the past.
• To increase the robustness, the watermark strength of the
scheme, we propose several hybrid approaches. The first
one is visual-audio hybrid watermarking scheme. As videos
consist of both video and audio channels, the robustness of
our scheme can be enhanced by including an audio water-
mark. Consequently, we embed error correcting codes of a
video watermark as an audio watermark, which can refine
the retrieved watermark during watermark detection [12].
• The second approach is another hybrid with different water-
marking schemes. As no existing scheme is resistant against
all attacks, we employ the hybrid scheme to embed different
parts of a watermark into different scenes. Thus, the pro-
posed scheme is capable of resisting most of the common
attacks [12].
• To increase the fidelity, the media quality, of the water-
marking scheme, we propose a GA-based watermarking
scheme. By employing GA, the quality of the watermarked
video is enhanced.
CHAPTER 1. INTRODUCTION 6
• Experiments have been done on these novel video water-
marking schemes to test and show its performance. The
robustness of our approach is demonstrated using the cri-
teria of the latest StirMark test [14].
• We compare the proposed scheme with the existing scheme
in different aspects and discuss the advantages and the dis-
advantages of our scheme.
Our approach cultivates an innovative idea in embedding dif-
ferent parts of a watermark according to scene changes, embed-
ding its error correcting codes as an audio watermark, applying
a hybrid approach to the proposed scheme and employing GA
to optimize the fidelity of the scheme. This approach is never
explored in the literature, and its advantages are clear and sig-
nificant. The effectiveness of this scheme is verified through a
number of experiments.
1.4 The Structure of this Thesis
This paper is organized as 5 chapters. The next chapter in-
troduces the issues related to multimedia security and different
multimedia watermarking techniques, and a survey on current
watermark techniques and video watermarking scheme are pro-
vided. Novel video watermarking scheme is described in chapter
3 and the experimental results in Chapter 4 are followed by. Fi-
nally, a conclusion would be given in chapter 5.
2 End of chapter.
Chapter 2
Literature Review
Nowadays the digital media is easily to be reproduced due to
the rapidly growth of internet and the multimedia technologies,
this drives to urgent need to resolve the security and copyright
protection issues. Therefore, the field of digital watermarking
grows extremely fast in these few years [15].
The purpose of a digital watermark is to embed auxiliary
information into a digital signal by making small changes that
are not perceptible to its intended recipient. For instance, in
the case of multimedia watermarking, the hidden signal should
not result in any visible or audible distortions. Because the
embedded signals enable invisible tags to be attached to digital
documents, watermarks are powerful tools that will play a role
in solving the growing digital property identification problem
[16].
This chapter overviews previous work in digital video water-
marking and related fields. We first have a look of two popular
security tools, digital signatures and cryptography. Then the
principle and the techniques of digital watermarking are dis-
7
CHAPTER 2. LITERATURE REVIEW 8
cussed. Besides, the essential differences and the advantages
that watermarking techniques provide over these existing tech-
nologies are explained. Moreover, the essential ingredients of
video watermarking are presented. The following section re-
views a number of techniques proposed in the literature. Then
different watermarking algorithms are implemented and evalu-
ated. Finally, a comparison among different video watermarking
is given.
2.1 Security in Multimedia Communications
In a decade years ago, multimedia documents are rarely available
to the mass consumer market. However, as the rapidly develop-
ment of the pervasive digital information technology, everyone’s
computer can have high quality video compression, increasing
network bandwidth and accessibility, dense portable storage me-
dia, and compounding processing power. Nevertheless, these
technological advances lead to another crisis. Multimedia users
had the ability to tamper with, produce copies of, and illegally
redistribute digital content. Without solving this security issue,
digital multimedia products and services cannot take-off in an
e-commerce setting [17].
Digital signature and cryptography are currently two stan-
dardized approaches to protect the digital contents. Digital sig-
nature is commonly used to authenticate digital transmissions.
It is based on public key cryptography and one-way hash func-
tions. By passing the document through a publicly available
one-way hash function, a unique identifier is generated, which is
CHAPTER 2. LITERATURE REVIEW 9
signed with the owner’s private key. Then, a string is produced
and it is referred to as the digital signature. In addition to
the signed document, the intended recipients obtain public keys
from certification authorities [16]. The document is authentic
only when it matches with the decrypted signature by applying
the hash function.
However, the document and signature are not bound in any
noteworthy manner. When transmitting the multimedia docu-
ments, they may become separated accidentally in transit or in-
tentionally by a malicious party. Thus, the receiver may not able
to verify the authentic document. In addition, this method of
tamper detection is too strict for multimedia objects. It dose not
allow the document to undergo compression and format changes
while still maintaining their authenticity. If just one bit differs
from the original, for instance due to lossy compression for effi-
cient network transfer, the hash identifier test will fail.
Use of cryptographically secure license keys is another method
for protecting digital intellectual property. The content of the
documents are protected form manipulation and stealing during
delivery as the assessment of the document is only permitted
to those who possess the appropriate key. However the criti-
cal flaw in this solution is that after transmitted and delivery
the document [17], the permitted recipient is able to access to
original proprietary data, which can then be reproduced per-
fectly and redistributed inexpensively. Thus, This technique is
not effective because it does not provide permanent protection
for the multimedia content after delivery. Moreover, with this
scheme, the intellectual property owner is not able to trace the
CHAPTER 2. LITERATURE REVIEW 10
responsibilities of pirating the properties.
According to the findings, we notice that an ideal solution
must somehow integrate security information directly into the
content of the multimedia document and the security informa-
tion should be inseparable from the document during its useful
lifespan. Moreover, the additional information should be per-
ceptually invisible as the multimedia documents are ultimately
processed by human viewers or listeners and the contents should
not be affected. Finally is the flexibility of the scheme. It should
be able to support identification of different copies of the docu-
ment.
Digital watermarking may be one of the suitable solutions. It
is an analogous techniques that have been used to protect valu-
able hardcopy documents, such as money, cheques and official
correspondence, for long time ago. Paper watermarks are faint
designs that are embedded by the manufacturer into the paper
used to produce such hardcopies. These marks are convincingly
hard to fake, and at the same time they do not obstruct the
normal processing, i.e. reading, and are impossible to be re-
moved without leaving any engram or causing severe damage
to the contents of the document. Digital watermarking tech-
nologies strive to achieve these goals in a digital environment
by inserting a retrievable watermark directly into the softcopy
data stream [17].
CHAPTER 2. LITERATURE REVIEW 11
2.2 Cryptography
Cryptography is the first technology that content owners would
turn to. It is probably the most common method of protecting
digital documents and certainly one of the best developed as
a science. Before delivery, the content is encrypted and the a
decryption key is provided only to those who have permission to
access the legitimate copies of the content. Then, the encrypted
file can be made available through the Internet, but would be
useless to a pirate without appropriate key. After encrypted,
the structure of the message is changed. It is meaningless and
unintelligible unless it is decrypted [18].
There are two kinds of cryptosystems: symmetric and asym-
metric [19]. Symmetric cryptosystems use the same key, known
as the secret key, to encrypt and decrypt a message, and asym-
metric cryptosystems use one key, named as public key, to en-
crypt a message and a different key, named as private key, to
decrypt it. Asymmetric cryptosystems are also called public key
cryptosystems .
Symmetric cryptosystems have a problem: ”how do you trans-
port the secret key from the sender to the recipient securely and
in a tamper proof fashion?” [19]. If you could send the secret
key securely, in theory, you then would simply use that secure
channel to send your message instead of encrypting your mes-
sage with symmetric cryptosystem. Commonly, trusted couriers
are used as a solution to this problem.
One example using symmetric cryptosystem is shown in Fig-
ure 2.1. Alice and Bob want to communicate in secret, while
CHAPTER 2. LITERATURE REVIEW 12
Message Encrypt
Key
Alice
Decrypt Message
Key
Bob
Eva
Figure 2.1: Symmetric Cryptosystem
Eve wants to eavesdrop. Alice and Bob could be military jets,
on-line businesses or just friends trying to have a private con-
versation. They cannot stop Eve listening to their radio signals,
so they can keep communication by using cryptography.
Alice and Bob exchange a digital key, so they both know it,
but it is otherwise secret [20]. Alice uses this key to encrypt
messages she sends, and Bob reconstructs the original messages
by decrypting with the same key. The encrypted messages are
useless to Eve, who does not know the key, and so cannot recon-
struct the original messages. With a good encryption algorithm,
this scheme can work well, but exchanging the key while keeping
it secret from Eve is a problem.
Asymmetric cryptosystem is another more efficient and reli-
able solution, such as RSA, which is the popular security tool
[20]. Asymmetric cryptosystems is different, because it splits
the key up into a public key for encryption and a secret key for
decryption. It’s not possible to determine the secret key from
the public key.
CHAPTER 2. LITERATURE REVIEW 13
Message Encrypt
Public Key
Alice
Decrypt Message
Secret Key
Bob
Key
Generator
Bob Public Key
Figure 2.2: Asymmetric Cryptosystems
In the Figure 2.2, Bob generates a pair of keys and tells ev-
erybody, including Eve, his public key, while only he knows his
secret key. Anyone can use Bob’s public key to send him an en-
crypted message, but only Bob knows the secret key to decrypt
it. This scheme allows Alice and Bob to communicate in secret
without having to meet.
CHAPTER 2. LITERATURE REVIEW 14
2.3 Digital Watermarking
Watermarking technique is a particular embodiment of multi-
media security. Digital Watermark is defined as a digital signal
or pattern inserted into a digital data, which can also be re-
ferred to copyright information. Watermarking is a key process
in the protecting copyright ownership of electronic data, includ-
ing image, videos, audio . . . etc. The term watermarking comes
from using the invisible ink to write secret messages [18]. The
additional requirement for watermarking is robustness. Even if
the existence of a watermark is known, such as the case in pub-
lic watermarking schemes, it should be ideally impossible for an
attacker to remove or destroy the embedded watermark with-
out rendering the cover object unusable. Generally, watermark
has three distinct properties imperceptible, inseparable from the
work, and undergoes the same transformation as the work [21].
A simple watermarking idea is shown in Figure 2.3. Water-
mark is a design of the watermark signal W to be added to the
host signal. The watermark signal, apart from depending on the
watermark information W’, may also depend on a key K and
the host data I into which it is embedded, shown in Equation
2.1
W = f
0
(I, K, W
/
) (2.1)
In watermarking algorithm, the host data I, such as stego-
image, is input to the watermarking algorithm and the algorithm
watermarks the image with a watermark W and output the
watermarked image I’ with the Equation 2.2:
CHAPTER 2. LITERATURE REVIEW 15
Watermarking
Watermark W
Stego-Image I
Secret / Public Key K
Watermarked
Image I
Detection
Algorithm
Watermark W
and/or Original
Image I
Test Image I
Secret / Public Key K
Watermark
or
Confidence
measure
Figure 2.3: Watermarking Embedding and Detection Scenario
I

W −→I
/
(2.2)
Verification algorithm is a design of the corresponding ex-
traction method that recovers the watermark information from
the signal mixture, perhaps with the help of the key and the
original, shown in Equation 2.3 .
I
/
= g(I, I
/
, K) (2.3)
CHAPTER 2. LITERATURE REVIEW 16
2.4 Essential Ingredients for Video Watermark-
ing
Watermarking systems can be characterized by a number of
defining properties including embedding effectiveness, fidelity,
data payload, blind or informed detection, false positive rate,
robustness, security, cipher and watermark keys, modification
and multiple watermark, cost, tamper resistance, unobtrusive-
ness, ready detection, unambiguous, sensitivity, and scalability.
The relative importance of each property is dependent on the
requirement of the application and the role the watermark will
play. Some of them are common to most practical applications.
In this section, such general requirements are listed and briefly
discussed. The analysis focuses on image and video watermark-
ing.
2.4.1 Fidelity
What requirements should an ideal watermarking system have?
The first requirement would clearly be that of Fidelity [22]. A
watermarking system is of no use to anyone if it distorts the
cover image to the point of being useless, or even highly dis-
tracting. Ideally, the watermarked image should be perceptually
invisible even on the highest quality equipment.
Although visible watermarks tend to be more robust, for gen-
eral purpose applications it is desirable for the embedded mark
to be imperceptible to the human eye or ear. Invisibility is that
degree that an embedded watermark remains unnoticeable when
CHAPTER 2. LITERATURE REVIEW 17
a user views the watermarked contents. So far researchers have
tried to hide the watermark in such a way that it is impossible
to be noticed. However this requirement conflicts with other
requirements such as tamper resistance and robustness.
2.4.2 Robustness
The ideal watermark must also be highly robust, entirely resis-
tant to distortion introduced during either normal use, i.e. unin-
tentional attack, or a deliberate attempt to disable or remove the
watermark present, i.e. intentional, or malicious attack. Unin-
tentional attacks involve transforms that are commonly applied
to images during normal use, such as cropping, resizing, contrast
enhancement. . . etc.
Robustness is the resilience of an embedded watermark against
removal by signal processing. The use of music, images and
video signals in digital form, commonly involves many types of
distortions, such as lossy compression, or, in the image case, fil-
tering, resizing, contrast enhancement, cropping, rotation and
so on. For watermarking to be useful, the mark should be de-
tectable even after such distortions occurred. It is a common
opinion [18] that robustness against signal distortion is better
achieved if the watermark is placed in perceptually significant
parts of the signal. This depends on the behavior of lossy com-
pression algorithms, which operate by discarding perceptually
insignificant data not to affect the quality of the compressed
image, audio or video.
A particularly interesting form of unintentional attack is that
CHAPTER 2. LITERATURE REVIEW 18
of image compression. Meerwald [23] points out that lossy com-
pression and watermarking are inherently at odds; watermarking
seeks to encode information in extra bits that compression hopes
to remove. Thus, ideal watermarking and compression systems
are most likely inherently exclusive.
In malicious attacks, an attacker deliberately tries to disable
the watermark, often through a geometric distortion or the addi-
tion of noise. A final note is that robustness can include either
resilience to attack, or complete fragility. It may be the case
that some watermarking systems may require the watermark to
totally destroy the cover object if any tapering is present [24].
Consequently, watermarks hidden among perceptually insignif-
icant data are likely not to survive compression. In the image
watermarking case, the resistance to geometric manipulations,
such as translation, resizing, rotation and cropping is still an
open issue, yet such operations are very common and a solution
needs to be found before watermarking techniques are success-
fully applied to image copyright protection.
Most of Video watermarking scheme base on the techniques
of the image watermarking. But video watermarking introduces
some issues not present in image watermarking. Video signals
are highly susceptible to pirate attacks, including frame averag-
ing, frame dropping, frame swapping, statistical analysis, inter-
polation etc.
Petitcolas [25] provides us with a rough set of reliability or
robustness metrics, shown below in Table 2.1.
CHAPTER 2. LITERATURE REVIEW 19
Table 2.1: Basic Robustness Requirements
Level Zero Low Level Moderate
Standard JPEG Compression Quality 100-90 100-75 100-50
Color Correction (GIF) 256 256 16
Cropping 100-90 100-75 100-50
Gamma Correction 0.7-1.2 0.5-1.5
Scaling 1/2-3/2 1/3-2
Rotation ± 0-2 deg. ± 0-5 deg., 90 deg.
Horizontal Flip yes yes
Uniform Noise 1-5 1-15
Contrast ± 0-10 ± 0-25
Brightness ± 0-10 ± 0-25
Median Filter 33
2.4.3 Use of Keys
Another property of an ideal watermarking system is that it im-
plement the use of keys to ensure that the approach is not ren-
dered useless the moment that the algorithm becomes known
[22]. It may also be a goal that the system utilizes an asym-
metric key system such as in public/private key cryptographic
systems. Although private key systems are fairly easy to imple-
ment in watermarking, asymmetric key pairs are generally not.
The risk here is that embedded watermarking systems might
have their private key discovered, ruining security of the entire
system. This was exactly the case when a single DVD decoder
implementation left it’s secret key unencrypted, breaching the
entire DVD copy protection mechanism.
CHAPTER 2. LITERATURE REVIEW 20
2.4.4 Blind Detection
Blind detection refers to the ability to detect the watermark
without access to the original document. Because of the im-
mense size of uncompressed video files and the difficulty of in-
dexing them to search for a specific frame, it is an especially
important requirement in video watermarking.
2.4.5 Capacity and Speed
Slightly less important requirements of an ideal watermarking
system might be capacity, and speed. A watermarking system
must allow for a useful amount of information to be embedded
into the image. This can range from a single bit all the way
up to multiple paragraphs of text. Furthermore, in watermark-
ing systems destined for embedded applications, the watermark
detection (or embedding) may not be overly computationally
intensive as to preclude its use on low cost micro-controllers
Capacity is that amount of information that can be expressed
by an embedded watermark. Theoretical capacity of embedded
watermarks has been examined using information-theoretic con-
cepts. Depending on the application at hand, the watermarking
algorithm should allow a predefined number of bits to be hid-
den. General rules do no exist here, however, in the image case,
the possibility of embedding into the image at least 300-400 bits
should be granted. In any case, system designers should keep
well in mind that the number of bits could be hidden into data
is not unlimited; but very often is fairly small.
CHAPTER 2. LITERATURE REVIEW 21
2.4.6 Statistical Imperceptibility
The last possible requirement of an ideal watermarking system
is that of statistical imperceptibility [25]. The watermarking
algorithm must modify the bits of the cover in such a way that
the statistics of the image are not modified in any telltale fashion
that may betray the presence of a watermark. This requirement
is not quite as important here as it is in steganography, but some
applications may require it.
2.4.7 Low Error Probability
Even in the absence of attacks or signal distortions, the proba-
bility of failing to detect the watermark, i.e. false-negative, and
of detecting a watermark when, in fact, one does not exist, i.e.
false-positive, must be very small. Usually, statistically based
algorithms have no problem in satisfying this requirement; how-
ever such ability must be demonstrated, if watermarking is to
be legally credible.
2.4.8 Real-time Detector Complexity
For consumer-oriented watermarking applications, it is impor-
tant that the complexity of the detection and extraction algo-
rithms be low enough to execute within the specified real-time
deadlines.
CHAPTER 2. LITERATURE REVIEW 22
2.5 Review on Video Watermarking Techniques
As a method of intellectual property protection, digital water-
marks have recently stimulated significant interest and become a
very active area of research. Although watermarking is a recent
field of research, many techniques have already been proposed
both in the academic as well as in the industry. Various tech-
niques are applied in watermarking algorithms. They can be
classified into different types based on the offered functionalities.
In this section, a brief review of the current video watermarking
technologies is presented.
A digital document can be authenticated with what is known
as a digital watermark. A watermark is a secret code or image
incorporated into an original content, which acts to verify both
the owner and content of the document. The use of perceptually
invisible watermark is one of the copyright protection. A water-
marking algorithm consists of three parts: watermark, marking
algorithm and verification algorithm.
Each owner has a unique watermark. The marking algorithm
incorporates the watermark into the image or video. The veri-
fication algorithm authenticates the content, determining both
the owner and the integrity of the content. A variety of im-
perceptible watermarking schemes have been proposed over the
past few years. Numerous watermarking research tasks have
proposed many watermark techniques in terms of various appli-
cation areas. Moreover, different insertion and extraction meth-
ods can be found. We can classify the watermarking techniques
according to several points of view, as shown in Table 2.2. In this
CHAPTER 2. LITERATURE REVIEW 23
section, we focus on analyzing the video watermark processing
methods.
Digital watermarking can be applied to many different types
of documents, including text [26], audio [27, 28, 29, 30, 31, 32],
image [33, 34, 35, 36, 37, 38, 39] and video [40, 41, 42, 43, 44, 45].
Watermark techniques can be classified into visible [46, 47] and
invisible [29, 30, 40, 41] watermarks. In general, invisible water-
marks are mostly used . The oblivious meaningful video water-
marking remains a challenging problem since the original video
is often unavailable due to videos’ bulky volume. Watermarks,
on the other hand, need robustness to protect the ownership
from various attacks. They can be classified into three cate-
gories, robust [48, 49, 50, 51], semi-fragile [52] and fragile [53]
watermarks. Different applications would be chosen for different
levels of robustness according to the requirement. Applications
for copyright protection would require to use a robust water-
mark. Applications for proving integrity would employ a fragile
or semi-fragile watermark. Watermarks to be inserted can also
be classified into two types: noise type [54] and image type
[41, 42, 43, 44]. A noise type includes pseudo noise, Gaussian
random and chaotic sequence. A watermark can be a random
sequence with one information bit or multiple-bit meaningful in-
formation. The random sequence watermark is more robust in
general; however, embedding meaningful watermark is more im-
portant in some applications. For image types, there are binary
image, stamp, logo and label. Moreover, watermark process-
ing methods are classified into four categories: spatial domain,
frequency domain, compression domain and hybrid. Finally,
CHAPTER 2. LITERATURE REVIEW 24
watermark extraction methods can be classified as private [55],
semi-private [56] and public [57] watermarking, according to the
necessity of the original media.
Table 2.2: Classification of watermarking according to several viewpoints
Classification Contents
Inserted media category Text [26]
Image [27, 28, 29, 30, 31, 32]
Audio [33, 34, 35, 36, 37, 38, 39]
Video[40, 41, 42, 43, 44, 45]
Perceptivity of watermark Visible [46, 47]
Invisible[29, 30, 40, 41]
Robustness of watermark Robust [48, 49, 50, 51]
Semi-fragile [52]
Fragile [53]
Inserting watermark type Noise [54]
Information tagging
Image [41, 42, 43, 44]
Processing method Spatial domain LSB
Image checksum [58]
Patchwork [59]
Random function[60]
Frquency domain Look-up table
Spread spectrum DCT [61, 62]
Wavelet (DWT) [63, 64, 65, 66]
Fourier (DFT)[67, 68]
Compression domain MPEG1[42, 69, 70]
MPEG2[71, 72]
MPEG4[26, 73]
JPEG2000 [74]
Hybrid Visual-audio [75, 76]
Different watermarks [77]
Different watermarking scheme [78]
Necessary data for extraction Private [55]
Semi-private [56]
Public [57]
CHAPTER 2. LITERATURE REVIEW 25
2.5.1 Video Watermarking
Many digital watermarking schemes have been proposed in the
literature for still images and videos. Most of them operate
on uncompressed videos [10, 79, 80], while others embed water-
marks directly into compressed videos [72, 81].
Recently, researchers tend to investigate video watermarking
techniques that is robust and invisible. These schemes can be
distinguished in terms of the domain that the watermark being
embedded or detected, their capacity, real-time performance,
the degree to which all three axes are incorporated, and their
resistance to particular types of attacks [17]. A classification
map of the existing video watermarking techniques is presented
in Figure 2.4. They can be divided into 3 main groups based on
the domain that the watermark is embedded, they are spatial
domain, frequency domain and MPEG coding structure based.
Most of the proposed video watermarking scheme based on the
techniques of the image watermarking and applied to raw video
or the compressed video. As some issue in video watermarking
is not present in image watermarking, such as video object and
redundancy of the large amount video data, researchers have
make use of those characteristics to develop different schemes.
In the following sections, each class of algorithms is briefly de-
scribed. Besides, we present the important idea, strength and
limitation introduced by these schemes.
CHAPTER 2. LITERATURE REVIEW 26
Invisible Robust
Video Watermarking
Techniques
Spatial Domain
Method
Frequency Domain
Method
MPEG coding
Structure based
Method
Figure 2.4: Classification map of existing digital video watermark techniques
2.5.2 Spatial Domain Watermarks
We first review the video watermarking techniques in the spatial
domain. Algorithms in this class generally share the following
characteristics:
• The watermark is applied in the pixel or coordinate domain.
• No transforms are applied to the host signal during water-
mark embedding.
• The watermark is derived from the message data via spread
spectrum modulation.
• Combination with the host signal is based on simple oper-
ations, in the pixel domain.
• The watermark can be detected by correlating the expected
pattern with the received signal.
The main strengths of pixel domain methods are that they
are conceptually simple and have very low computational com-
plexities. As a result they have proven to be most attractive for
CHAPTER 2. LITERATURE REVIEW 27
video watermarking applications where real-time performance
is a primary concern. However, they also exhibit some major
limitations: The need for absolute spatial synchronization leads
to high susceptibility to de-synchronization attacks; lack of con-
sideration of the temporal axis results in vulnerability to video
processing and multiple frame collusion; and watermark opti-
mization is difficult using only spatial analysis techniques. The
three methods that fall into this class can be distinguished by
the dimensionality of the watermark pattern. Techniques based
on 1D and 2D spread spectrum modulation, and 3D Code Divi-
sion Multiple Access (CDMA) modulation have been proposed.
Several different methods enable watermarking in the spatial
domain. The simplest is to just flip the lowest-order bit of chosen
pixels in a grey scale or colour image. This will work well only
if the image is subjected to any human or noisy modification.
A more robust watermark can be embedded in an image in the
same way that a watermark is added to paper. Such techniques
may superimpose a watermark symbol over an area of the picture
and then add some fixed intensity value for the watermark to the
varied pixel values of the image. The resulting watermark may
be visible or invisible depending upon the value (large or small,
respectively) of the watermark intensity. One disadvantage of
spatial domain watermarks is that picture cropping, which is a
common operation of image editors, can be used to eliminate
the watermark.
Spatial watermarking can also be applied using colour sepa-
ration. In this way, the watermark appears in only one of the
colour bands. This renders the watermark visibly subtle so that
CHAPTER 2. LITERATURE REVIEW 28
it is difficult to detect under regular viewing. However, the wa-
termark appears immediately when the colours are separated
for printing or xerography. This renders the document useless
to the printer unless the watermark can be removed from the
colour band. This approach is used commercially for journal-
ists to inspect digital pictures from a photo-stockhouse before
buying non-watermarked versions.
Least Significant Bit Modification
The most straight-forward method of watermark embedding,
would be to embed the watermark into the least-significant-
bits of the cover object [82]. Given the extraordinarily high
channel capacity of using the entire cover for transmission in
this method, a smaller object may be embedded multiple times.
Even if most of these are lost due to attacks, a single surviving
watermark would be considered a success.
LSB substitution however despite its simplicity brings a host
of drawbacks. Although it may survive transformations such
as cropping, any addition of noise or lossy compression is likely
to defeat the watermark. An even better attack would be to
simply set the LSB bits of each pixel to one . . . fully defeat-
ing the watermark with negligible impact on the cover object.
Furthermore, once the algorithm is discovered, the embedded
watermark could be easily modified by an intermediate party.
An improvement on basic LSB substitution would be to use a
pseudo-random number generator to determine the pixels to be
used for embedding based on a given ”seed” or key [82]. Security
CHAPTER 2. LITERATURE REVIEW 29
of the watermark would be improved as the watermark could no
longer be easily viewed by intermediate parties. The algorithm
however would still be vulnerable to replacing the LSB’s with a
constant. Even in locations that were not used for watermarking
bits, the impact of the substitution on the cover image would
be negligible. LSB modification proves to be a simple and fairly
powerful tool for stenography, however lacks the basic robustness
that watermarking applications require.
Correlation-Based Techniques
Another technique for watermark embedding is to exploit the
correlation properties of additive pseudo-random noise patterns
as applied to an image [83]. A pseudo-random noise (PN) pat-
tern W(x,y) is added to the cover image I(x,y), according to the
equation shown below in Equation 2.4.
I
w
(x, y) = I(x, y) +k W(x, y) (2.4)
In Equation 2.4, k denotes a gain factor, and IW the resulting
watermarked image. Increasing k increases the robustness of
the watermark at the expense of the quality of the watermarked
image.
To retrieve the watermark, the same pseudo-random noise
generator algorithm is seeded with the same key, and the cor-
relation between the noise pattern and possibly watermarked
image computed. If the correlation exceeds a certain thresh-
old T, the watermark is detected, and a single bit is set. This
method can easily be extended to a multiple-bit watermark by
CHAPTER 2. LITERATURE REVIEW 30
dividing the image up into blocks, and performing the above
procedure independently on each block.
This basic algorithm can be improved in a number of ways.
First, the notion of a threshold being used for determining a log-
ical ”1” or ”0” can be eliminated by using two separate pseudo-
random noise patterns. One pattern is designated a logical ”1”
and the other a ”0”. The above procedure is then performed
once for each pattern, and the pattern with the higher resulting
correlation is used. This increases the probability of a correct
detection, even after the image has been subject to attack [83].
2.5.3 Frequency Domain Watermarks
Generally DCT, FFT and wavelet transform are used as the
methods of data transformation. In these methods, a watermark
that one wishes to embed distributively in overall domain of
an original data, and the watermark, is hardly to be deleted
once embedded. For transformed domain techniques, they have
hierarchical watermarking with Discrete Cosine Transform, sub-
band watermarking techniques, Discrete Wavelet Transform or
Discrete Fourier Transform.
The main strength offered by transform domain techniques
is that they can take advantage of special properties of alter-
nate domains to address the limitations of pixel-based methods
or to support additional features. For instance, designing a wa-
termarking scheme in the Discrete Cosine Transform (DCT) do-
main leads to better implementation compatibility with popular
video coding algorithms such as Moving Pictures Experts Group
CHAPTER 2. LITERATURE REVIEW 31
(MPEG)-2, and in the shift and rotation-invariant Fourier do-
mains facilitates the design of watermarks that inherit these
attractive properties. Besides, analysis of the host signal in a
frequency domain is a prerequisite for applying more advanced
masking properties of the HVS to enhance watermark robustness
and imperceptibility. Generally, the main drawback of transform
domain methods is their higher computational requirement. We
discuss the details of three methods here: Discrete Cosine Trans-
form, Discrete Wavelet Transform, and Discrete Fourier Trans-
form.
Discrete Cosine Transform
The classic and still most popular domain for image processing
is that of the Discrete Cosine Transform, or DCT. The DCT
allows an image to be broken up into different frequency bands,
making it much easier to embed watermarking information into
the middle frequency bands of an image. The middle frequency
bands are chosen such that they have minimize they avoid the
most visual important parts of the image (low frequencies) with-
out over-exposing themselves to removal through compression
and noise attacks (high frequencies) [83].
One such technique utilizes the comparison of middle-band
DCT coefficients to encode a single bit into a DCT block. To
begin, we define the middle-band frequencies (FM) of an 88
DCT block
CHAPTER 2. LITERATURE REVIEW 32
Discrete Wavelet Transform
Another possible domain for watermark embedding is that of
the wavelet domain. The DWT (Discrete Wavelet Transform)
separates an image into a lower resolution approximation im-
age (LL) as well as horizontal (HL), vertical (LH) and diagonal
(HH) detail components. The process can then be repeated to
computes multiple ”scale” wavelet decomposition, as in the 2
scale wavelet transform shown below in Figure 2.5.
One of the many advantages over the wavelet transform is
that that it is believed to more accurately model aspects of the
HVS as compared to the FFT or DCT. This allows us to use
higher energy watermarks in regions that the HVS is known
to be less sensitive to, such as the high resolution detail bands
LH,HL,HH). Embedding watermarks in these regions allow us
to increase the robustness of our watermark, at little to no ad-
ditional impact on image quality [83].
One of the most straightforward techniques is to use a similar
HH1
LL2
HH2
HL2
LH2
HL1
LH2
Figure 2.5: 2 Scale 2-Dimensional Discrete Wavelet Transform
CHAPTER 2. LITERATURE REVIEW 33
embedding technique to that used in the DCT. This can be
easily extended to multiple bit messages by embedding multiple
watermarks into the image. As in the spatial version, a separate
seed is used for each PN sequence, which are then added to the
detail coefficients. During detection, if the correlation exceeds
threshold for a particular sequence a ”1” is recovered; otherwise
a zero. The recovery process then iterates through the entire PN
sequence until all the bits of the watermark have been recovered.
Furthermore, as the embedding uses the values of the trans-
formed value in embedded, the embedding process should be
rather adaptive; storing the majority of the watermark in the
larger coefficients. The author [23] claims that the technique
should prove resistant to JPEG compression, cropping, and other
typical attacks.
Niu and Sun proposed a new wavelet-based digital water-
marking in [84]. This paper proposes a method of embedding a
digital watermark image in video. In the watermarking, the de-
composed watermark image with different resolution is embed-
ding in the corresponding resolution of the decomposed video
by means of multiresolution signal decomposing. The proposed
method is robust against the attack of frame dropping, averaging
and lossy compression. In [85], Serdean et al proposed another
DWT-base scheme. This scheme is a high capacity blind video
watermarking system invariant to geometrical attacks such as
shift, rotation, scaling and cropping. A spatial domain reference
watermark is used to obtain invariance to geometric attacks by
employing image registration techniques to determine and invert
the attacks. A second, high capacity watermark, which carries
CHAPTER 2. LITERATURE REVIEW 34
the data payload, is embedded in the wavelet domain according
to a human visual system model.
Mitchell et al proposed multiresolution video watermarking
using perceptual models and scene segmentation in [86]. The
watermark consists of static and dynamic temporal components
that are generated from a temporal wavelet transform of the
video scenes. To generate the watermark, the resulting wavelet
coefficient frames are modified by a perceptually shaped pseudo-
random sequence representing the author. The noise-like water-
mark is statically undetectable to thwart unauthorized removal.
Furthermore, the author representation resolves the deadlock
problem. The multiresolution watermark may be detected on
single frames without knowledge of the location of the frames in
the video scene.
Discrete Fourier Transform
M. Barni et al proposed a robust watermarking approach for
raw video in [87]. This approach first extracts the brightness
of the to-be-marked frame, computing its full-frame DFT and
then taking the magnitude of the coefficients. The watermark is
composed of two alphanumerical strings. The DFT coefficient
is altered, then IDFT. Only the first frame of each GOP is wa-
termarked, which was composed of twelve frames, leaving the
other ones uncorrupted. It is good robustness to the usual im-
age processing as linear/non-linear filtering, sharpening, JPEG
compression and resist to geometric transformations as scaling,
rotation and cropping. Decide to watermark one or more frames
CHAPTER 2. LITERATURE REVIEW 35
in GOP, a trade-off between time spent for marking and the de-
gree of robustness needed for the sequence can be achieve
2.5.4 Watermarks Based on MPEG Coding Structures
Video watermarking techniques that use MPEG-1, -2 and -4
coding structures as primitive components are primarily moti-
vated by the goal of integrating watermarking and compression
to reduce overall real-time video processing complexity. Com-
pression in block-based schemes like MPEG-2 is achieved by
using forward and bi-directional motion prediction to remove
temporal redundancy, and statistical methods to remove spatial
redundancy. One of the major drawbacks of schemes based on
MPEG coding structures is that they can be highly suscepti-
ble to re-compression with different parameters, as well as con-
version to formats other than MPEG. There are a number of
MPEG-2 and -4-based techniques that have been proposed, in-
cluding approaches based on GOP modification [69], high fre-
quency DCT coefficient manipulation, DCT block classification
[71], [88] and three more robust and general algorithms that will
be discussed in detail in this section. The two MPEG-2 water-
marking methods considered here embed hidden data by swap-
ping level-adjacent Variable Length Codeword (VLC) codeword
and manipulating mean luminance over regions of pixels.
Vassaux et al proposed a video object watermarking which is
based on the structure of MPEG-4 in [89]. This paper presents a
so-called scrambling technique which allows adapting any clas-
sical spread spectrum watermarking scheme operating in the
CHAPTER 2. LITERATURE REVIEW 36
spatial domain to the Mpeg-4 requirements concerning VO ma-
nipulation. This technique can be easily added to the embedding
and detection schemes without changing the watermarking al-
gorithm. It modifies some predefined pairs of quantized DCT
coefficient in the luminance block of pseudo-randomly selected
MBs. It is based on spread-spectrum techniques. Dividing the
image into blocks of equal size, them binary sequence is gen-
erated using secret key, and then adds to the image. Special
decomposition of mpeg-4 include the fact that VO have signifi-
cant value Watermark information has to present in each VO
In [90], Swanson, et al. presents a watermarking procedure to
embed copyright protection into video sequences which is object-
based transparent. To address issues associated with video mo-
tion and redundancy, individual watermarks are created for ob-
jects within the video. Each watermark is created by shaping an
author and video dependent pseudo-random sequence according
to the perceptual masking characteristics of the video. As a
result, the watermark adapts to each video to ensure invisibil-
ity and robustness. Furthermore, the noise like watermark is
statistically undetectable to prevent unauthorized removal.
Lu and Liao proposed another video object-based watermark-
ing in [91] which is resist to rotation and flipping. Video object
is very important concept in Mpeg4 standard. Video object
may be purposely cut and pasted for illegal use. In this pa-
per, a robust watermarking scheme foe video object protection
is proposed. For each segmented video object, a watermark is
embedded by a new technology designed based in the concept
of communication with side information. To solve the asyn-
CHAPTER 2. LITERATURE REVIEW 37
chronous problem caused by object placement, it proposes to
use eigenvectors of a video object for synchronization of rota-
tion and flipping.
In [92], Mobasseri proposed direct sequence watermarking
using m-frames This scheme applies a direct sequence spread
spectrum model to the watermarking of the digital video. First,
video signal is modeled as a sequence of bit planes arranged
along the time axis. Watermarking of this sequence is a two
layer operation. A controlling m-sequence first establishes a
pseudorandom order in the bit plane stream for later water-
marking. Watermark, defined as m-frames, supplant the tagged
bit planes. Moreover, attempts in corrupting the image to de-
stroy the watermark render the video useless before damaging
the seal itself. The watermarked video is also robust to video
editing attempts such as subsampling, frame reordering etc. The
watermark is also identifiable from very short segment of video.
Individual frames extracted from the video will also contain the
copyright information.
CHAPTER 2. LITERATURE REVIEW 38
2.6 Comparison between Different Watermark-
ing Schemes
In general, watermarking schemes can be roughly divided into
two categories: spatial domain watermark, and transformed do-
main watermark. We have chosen some representative water-
marking schemes in each category for implementation and per-
formed experiments to compare their robustness. They are:
Least Significant Bit (LSB) based watermarking scheme [93],
threshold-based correlation watermarking scheme [83], Direct
sequence watermark using m-frame [94, 95], DFT with template
matching [96], Discrete Wavelet Transform (DWT) based water-
marking scheme [97], Discrete Cosine Transform (DCT) based
watermarking scheme [98] and spread spectrum [99] watermark-
ing scheme. To evaluate the algorithms, the StirMark 4.0 bench-
mark program [100, 101] and 30 different images are used. The
tests are divided into the following sections: PSNR, compres-
sion, scaling, cropping, shearing, rotation, row/column removal,
and random geometric distortions. Each attack is considered by
itself and it is applicable after watermarking. For each image, we
assign a score of 1 if the watermark is correctly decoded in the
case. A value of zero is assigned if the watermark is incorrect.
The comparison is shown in Table 2.3.
From the result, we find that the watermarking schemes in
spatial domain are less robust than those in frequency domain.
LSB, threshold-based correlation and m-sequence watermarks
are perform worse than the other five implemented watermark-
ing algorithms. Therefore, these watermarking algorithms are
CHAPTER 2. LITERATURE REVIEW 39
Table 2.3: Comparison between different watermarking schemes
Attack Class LSB Threshold - based Correlation m-sequence / m-frame Spread Spectrum
JPEG Lossy Compression 0.20 0.75 0.7 0.85
PSNR 0.13 0.82 0.89 0.9
Add Noise 0.10 0.7 0.75 0.89
Median Filter 0.21 0.4 0.4 0.35
Row / Column Removal 0.4 0.63 0.7 0.69
Cropping 0.49 0.65 0.75 0.78
Rescale 0.22 0.5 0.62 0.83
Rotation 0.14 0.52 0.61 0.85
Affine 0.15 0.46 0.56 0.76
Geometrical Distortions 0.25 0.42 0.5 0.62
Shearing 0.27 0.3 0.54 0.85
Attack Class Mid-band DCT Mid-band DWT DFT template Matching Radon Transform
JPEG Lossy Compression 1 0.75 0.74 0.83
PSNR 0.98 1 0.81 0.78
Add Noise 0.95 0.73 0.86 0.75
Median Filter 0.4 0.3 0.25 0.3
Row / Column Removal 0.65 0.5 1 0.75
Cropping 0.62 0.76 0.89 0.85
Rescale 0.53 0.75 0.78 1
Rotation 0.5 0.52 1 0.98
Affine 0.35 0.45 0.98 0.83
Geometrical Distortions 0.64 0.75 0.37 0.75
Shearing 0.35 0.42 1 0.6
classified as fragile or semi-fragile watermarking. They can be
applied for the purpose of proving the integrity of a document.
The frequency domain watermarking schemes are relatively
more robust than the spatial domain watermarking schemes,
particularly in lossy compression, noise addition, pixel removal,
rescaling, rotation and shearing. DCT-based watermarking scheme
is the most robust to lossy compression. In this approach, an
image is broken up into different frequency bands by DCT, mak-
ing it much more easier to embed watermarking information into
the middle frequency bands of the image. The middle frequency
CHAPTER 2. LITERATURE REVIEW 40
bands are chosen to minimize the change of the most visually
important parts of the image (low frequencies) without over-
exposing themselves to the removal through compression and
noise attacks (high frequencies). Moreover, DWT-based water-
marking scheme is the most robust to noise addition.
DFT-based watermarking scheme with template matching
can resist a number of attacks, including pixel removal, rotation
and shearing. The purpose of the template is to enable resyn-
chronization of the watermark payload spreading sequence. It is
a key dependent pattern of peaks, which is also embedded into
DFT magnitude representation of the frame. The peaks are not
embedded by addition, but rather by modifying the value of the
target coefficient, such that it is at least two standard deviations
above the mean.
Radon transformation resists attacks by resealing and geo-
metric distortion. In the scheme, invariant watermarks use the
Radon transform and higher order spectra. A bispectrum fea-
ture vector of the image is used as the watermark. This approach
differs from the previous methods in that it embeds watermarks
into the phase of the higher order spectra. The Radon embed-
ding grid also outperforms the Fourier-Mellin based methods.
The weakness of the existing algorithms, however, includes:
i) The watermark is not robust to attacks which are specifically
targeted at to videos, such as frame dropping, averaging and sta-
tistical analysis. ii) The bit rate of the watermark is low. Some
algorithms embed only one bit information as the watermark.
iii) Existing techniques are not aware of the usefulness of the
audio channel in a video. iv) None of the existing watermarking
CHAPTER 2. LITERATURE REVIEW 41
schemes resists to all the attacks. v) A frequency domain water-
mark is more robust than a spatial domain watermark. To tackle
these problems, in this paper, we propose a novel watermarking
scheme based on scene changes with a hybrid approach.
2 End of chapter.
Chapter 3
Novel Watermarking Schemes
In this chapter, we present the proposed innovative digital video
watermarking scheme. A scene-based video watermarking scheme
is proposed, which is robust against the attacks of frame drop-
ping, averaging and statistical analysis, which were not solved
effectively in the past [11]. Moreover, a hybrid approach is pro-
posed, which can improve the robustness of the watermarking
scheme [12, 13]. To enhance the fidelity of the scheme, a GA-
based watermarking scheme is presented. In the following sec-
tions, the detail of each algorithm is described.
3.1 A Scene-based Video Watermarking Scheme
The new watermarking scheme we propose is based on scene
change analysis. Figure 3.1 shows an overview of our water-
marking process. In our scheme, a video and a watermark are
taken as the input, and the watermark is then decomposed into
different parts which are embedded in corresponding frames of
different scenes in the original video.
42
CHAPTER 3. NOVEL WATERMARKING SCHEMES 43
Video
Normalization
Scene Change Detection
DWT
DWT
Watermark
DWT Crop
Encryption
Key
Encryption
Encrypted
Watermark
Embedding Algorithm
if W[j] = 1,
Exchange C[i] with max(C[i], C[i+1], C[i+2], C[i+3], C[i+4])
else
Exchange C[i] with min(C[i], C[i+1], C[i+2], C[i+3], C[i+4])
Encrypted
Watermark
Watermarked
Video
8-bit planes
Bit Decomposition
Place the bit-plane
side by side
Figure 3.1: Overview of the watermarking process
CHAPTER 3. NOVEL WATERMARKING SCHEMES 44
As applying a fixed image watermark to each frame in the
video leads to the problem of maintaining statistical and per-
ceptual invisibility [102], our scheme employs independent wa-
termarks for successive but different scenes. However, applying
independent watermarks to each frame also presents a prob-
lem if regions in each video frame remain little or no motion
frame after frame. These motionless regions may be statistically
compared or averaged to remove the independent watermarks
[10, 103]. Consequencely, we use an identical watermark within
each motionless scene. With these mechanisms, the proposed
method is robust against the attacks of frame dropping, aver-
aging, swapping, and statistical analysis. This newly proposed
scheme consists of four parts, including: watermark preprocess,
video preprocess, watermark embedding, and watermark detec-
tion. Details are described in the following sections.
3.1.1 Watermark Preprocess
A watermark is scrambled into small parts in a preprocess, and
they are embedded into different scenes so that the scheme can
resist a number of attacks towards to the video. A 256-grey-
level image is used as the watermark, so 8 bits can represent
each pixel. The watermark is first scaled to a particular size as
follows:.
2
n
≤ m, n > 0 (3.1)
p +q = n, p, q > 0 (3.2)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 45
where m is the number of scene changes and n, p, q are positive
integers. The size of the watermark is represented as
64 2
p
64 2
q
(3.3)
Then the watermark is divided into 2
n
small images with
size 64 64. Figure 3.2 shows the procedure of the watermark
preprocess with m = 10, n = 3, p = 1, and q = 2.
In the next step, each small image is decomposed into 8 bit-
planes, and a large image m
n
can be obtained by placing the
bit-planes side by side only consisting of 0’s and 1’s. These
processed images are used as watermarks, and totally 2
n
inde-
pendent watermarks are obtained. To make the scheme more
robust, the processed watermarks m are transformed to the
wavelet domain and encrypted [104]. Sample preprocessed wa-
termarks are shown in Figure 3.3, where (a) is the original water-
mark, (b)-(i) represent the scrambled watermarks in the spatial
domain, and (j) shows the encrypted watermark of (b), i.e., m
/
0
.
Watermark
Crop
Encryption
Key
Encryption
Encrypted
Watermark
8-bit planes
Bit Decomposition
Place the bit-plane
side by side
Encrypted
Watermark
DWT
DWT
m0'
m0
m7
m7'
Figure 3.2: Preprocessing the watermark
CHAPTER 3. NOVEL WATERMARKING SCHEMES 46
3.1.2 Video Preprocess
Our watermark scheme is based on 4 levels Discrete Wavelet
Transform (DWT). All frames in the video are transformed to
the wavelet domain. The frames are decomposed in 4 level sub-
band frames by separable 2-D wavelet transform. It produces a
low-frequency sub-band LL
4
, and three series of high-frequency
subbands LH
j
, HL
j
, HH
j
, where j < 4. The low frequency sub-
band is a lowpass approximation of the original frame, and con-
Figure 3.3: (a) Original watermark (b-i) Preprocessed watermark m
0
− m
7
(j) Encrypted watermark m

0
CHAPTER 3. NOVEL WATERMARKING SCHEMES 47
tains most of energy of the frame. The other subbands include
edge components of horizontal, vertical and diagonal directions
at different scale and resolution, respectively. According to the
energy distribution, LL
4
is the most important then LH
j
, HL
j
,
and HH
j
. For different levels, the higher the level, the more
important the subbands. In our scheme, we only embed the
watermark in the middle frequency subbands. Our scheme is
based on 4 levels DWT, which is determined by experiments. If
less than 4 levels is applied, the capacity of the scheme would
be decreased; if larger than 4 levels is applied, the quality of the
watermark video is affected.
Moreover, scene changes are detected from the video by ap-
plying the histogram difference method on the video stream.
The histogram difference method is used for scene change de-
tection. Each frame is coded in 24-bit image, eight bits for
each color (red R, green G, blue B). Consequently, each pixel
is checked and classified into different classes. For efficiency
purpose, only the most significant two bits for each color are
considered. Then, the total difference of the whole histogram
(H) is calculated as:
H =
63

i,j=0
[P
a
(i) −P
b
(j)]
2
(3.4)
where P
a
and P
b
are the frequency distribution of the pixel level
of two images, a and b, i and j are two successive columns in
the color histogram. If H > threshold(T), we consider there is
a scene change. The threshold is again determined by experi-
ments.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 48
Figure 3.4: Scene change detection
Independent watermarks are embedded in frames of different
scenes. Within a motionless scene, an identical watermark is
used for each frame. As shown in Figure 3.4, watermark m
1
is
used for the first scene. When there is a scene change, another
watermark m
3
is used for the next scene. The watermark for
each scene can be chosen with a pseudo-random permutation
such that only a legitimate watermark detector can reassemble
the original watermark.
3.1.3 Watermark Embedding
Watermark is then embedded to the video frames by changing
position of some DWT coefficients with the following condition:
if W
j
= 1
Exchange C
i
with max(C
i
, C
i+1
, C
i+2
, C
i+3
, C
i+4
)
else
Exchange C
i
with min(C
i
, C
i+1
, C
i+2
, C
i+3
, C
i+4
)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 49
(3.5)
where C
i
is the i
th
DWT coefficient of a video frame, and W
j
is
the j
th
pixel of a corresponding watermark image [105]. When
the watermark W
j
= 1, we perform an exchange of the C
i
with
the maximum value among C
i
, C
i+1
, C
i+2
, C
i+3
, C
i+4
. When
W
j
= 0, we perform an exchange of the C
i
with the minimum
value among C
i
, C
i+1
, C
i+2
, C
i+3
, C
i+4
. With this algorithm, the
retrieval of the embedded watermark dose not need the original
image. The sequence of watermark coefficients used is depicted
in Figure 3.5, where higher frequency coefficients are embedded
to higher frequency parts of the video frame, and only the middle
frequency wavelet coefficient of the frame (middle frequency sub-
band) is watermarked [10].
HL1
LL4
HH1 LH1
LH2 HH2
HL2
LH3 HH3
HL3
HL4
LH4 HH4
Encrypted
Watermark mj
Figure 3.5: Embedding watermarks in a frame
CHAPTER 3. NOVEL WATERMARKING SCHEMES 50
3.1.4 Watermark Detection
The video is processed to detect the video watermark. In this
step, scene changes are detected from the tested video. Also,
each video frame is transformed to the wavelet domain with
4 levels. Then the watermark is extracted with the following
condition :
if WC
i
> median(WC
i
, WC
i+1
, WC
i+2
, WC
i+3
, WC
i+4
)
EW
j
= 1
else
EW
j
= 0
(3.6)
where WC
i
is the i
th
DWT coefficient of a watermarked video
frame, and EW
j
is the j
th
pixel of the extracted watermark [106].
When the watermark WC
j
is greater than median value among
WC
i
, WC
i+1
, WC
i+2
, WC
i+3
, WC
i+4
, the extracted watermark
is considered as one, i.e., EW
j
= 1; otherwise, it is considered
as zero, i.e., EW
j
= 0. With this algorithm, the retrieval of the
embedded watermark dose not need the original image. This is
an important property to video watermarking.
As an identical watermark is used for all frames within a
scene, multiple copies of each part of the watermark may be
obtained. The watermark is recovered by averaging the water-
marks extracted from different frames. This reduces the effect
if the attack is carried out at some designated frames. Thus we
can combine the 8 bit-planes and recover the 6464 size image,
CHAPTER 3. NOVEL WATERMARKING SCHEMES 51
i.e., 1/2
n
part of the original watermark.
If enough scenes are found and all parts of the watermark
are collected, the original large watermark image can be recon-
structed. This can be shown in Figure 3.6, where the original
frame, the watermarked frame, and the extracted watermark are
depicted.
Figure 3.6: (a) Original frame (b) Watermarked frame (c) Extracted water-
mark corresponding to Figure 3.3(g) (d) Recovered watermark.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 52
3.2 Theoretical Analysis
In this section, performance and the capacity of the proposed
watermarking scheme will be calculated.
3.2.1 Performance
In this section, the performance of the proposed algorithm is
calculated. Let T be the total number of frame in a video and
n
1
n
2
be the size of the video frame and m the total number
of scene change in the video.
Size of the video frame = n
1
n
2
Size of the watermark = m
1
m
2
Number of frames = T
Number of scene change = m
Prepare Watermark
To prepare the watermark for the scheme, a watermark is scram-
bled into small parts in preprocess, and they are embedded into
different scenes. The watermark is first scaled to a particular
size with the Equation 3.1.
2
n
≤ m, n > 0 (3.7)
p +q = n, p, q > 0 (3.8)
where m is the number of scene changes and n, p, q are positive
integers. The size of the watermark should be
64 2
p
64 2
q
(3.9)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 53
Then the watermark is divided into 2
n
small images with size
64 64.
Find N in Equation 3.7 = O(log m)
Find (q, p) in Equation 3.8= (
n
2
, n −
n
2
) = O(1)
Resize the watermark:
if size of watermark is smaller than video frame = O(m
1
m
2
)
if size of watermark is greater than video frame = O(64642
n
)
= O(m) = O(max[(m
1
, m
2
), m])
Generating different part of watermark = 2
n
64 64 = O(m)
Total running time = O(log m) + O(1)+ O(max[(m
1
, m
2
), m])
+ O(m) = O(max[(m
1
, m
2
), m])
Scene Change
Scene changes are detected from the video by applying the his-
togram difference method on the video stream. The histogram
difference method is used for scene change detection. Each frame
is coded in 24-bit image, eight bits for each color (red R, green
G, blue B). Consequently, each pixel is checked and classified
into different classes. For efficiency purpose, only the most sig-
nificant two bits for each color are considered. Then, the total
difference of the whole histogram (H) is calculated by Equation
3.10:
H =
63

i,j=0
[P
a
(i) −P
b
(j)]
2
(3.10)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 54
Scanning to generate the histogram for 1 frame = n
1
n
2
Create histogram = n
1
n
2
Compare the histogram = 64 64
Total running time = [2(n
1
n
2
) + 64] T = O(n
1
n
2
T)
Embedding watermark
Our watermark scheme is based on 4 levels Discrete Wavelet
Transform (DWT). All frames in the video are transformed to
the wavelet domain. The frame is decomposed in 4 level sub-
band frame by separable 2-D wavelet transform. It produces a
low frequency sub-band LL
4
, and three series of high frequency
subbands LH
j
, HL
j
, HH
j
, where j < 4.
Running time for DWT
= 2[n
1
n
2
+ n
1

n
2
2
2] + 2[
n
1
2

n
2
2
+
n
1
2

n
2
4
2] + 2[
n
1
4

n
2
4
+
n
1
4

n
2
8
2] + 2[
n
1
8

n
2
8
+
n
1
8

n
2
16
2]
= 4n
1
n
2
+ 2n
1
n
2
+n
1
n
2
+
n
1
n
2
2
=
15n
1
n
2
2
= O(n
1
n
2
)
When embedding the watermark, only the middle frequency
wavelet coefficient of the frame (middle frequency sub-band) is
watermarked, i.e., DWT coefficients of HL
1
, LH
1
, HL
2
, LH
2
,
HL
3
, LH
3
, HL
4
and LH
4
are watermarked [10].
Total number of pixel to watermark
CHAPTER 3. NOVEL WATERMARKING SCHEMES 55
=
n
1
n
2
2
+
n
1
2

n
2
2
2
+
n
1
4

n
2
4
2
+
n
1
8

n
2
8
2
=
n
1
n
2
2
+
n
1
n
2
8
+
n
1
n
2
32
+
n
1
n
2
128
=
85n
1
n
2
128
=O(n
1
n
2
)
Number of operation for watermark
=
85n
1
n
2
128
T
= O(n
1
n
2
T)
After the watermark is embedded, the video frame is inverse-
DWT. Running time for IDWT
= 2[n
1
n
2
+ n
1

n
2
2
2] + 2[
n
1
2

n
2
2
+
n
1
2

n
2
4
2] + 2[
n
1
4

n
2
4
+
n
1
4

n
2
8
2] + 2[
n
1
8

n
2
8
+
n
1
8

n
2
16
2]
= 4n
1
n
2
+ 2n
1
n
2
+n
1
n
2
+
n
1
n
2
2
=
15n
1
n
2
2
= O(n
1
n
2
)
Total running time for embedding watermark
= O(n
1
n
2
T) + 2O(n
1
n
2
T)
= O(n
1
n
2
T)
Finally, Running Time
=O(max[(m
1
, m − 2), m]) + O(n
1
n
2
T) + O(n
1
n
2
T) n
1
n
2

m
1
, m
2
=O(n
1
n
2
T)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 56
3.2.2 Capacity
Watermarking can be viewed as a communication problem with
side information available at the encoder and the decoder. The
problem is mathematically defined by distortion constraints, by
statistical models for the host signal, and by the information
available in the game between the information hider, the at-
tacker, and the decoder. Capacity of the watermark is defined
as how much information can be carried by the watermark when
it is embedded in an image. In particular, information theory
explains why the performance of watermark decoders that do
not have access to the host signal may surprisingly be as good
as the performance of decoders that know the host signal. Ca-
pacity expressions are derived under a parallel-Gaussian model
for the host-image source. [107]
In this section, we investigate the watermarking capacity
based domain-specified masking effects. We derive the capacity
when that power and noise constraints are not uniform across
sample, ie., the capacity issue in a variant state channel.
We consider an video as a channel with spatial-variant states,
which the power constraint of each state is determined by HSV
model or masking effect in some special domains. In this way,
each coefficient is considered as an independent random variable
with its own noise distribution. We will not consider a coefficient
as a communication channel [108, 109] because a channel usually
incites its reuse temporally, spatially, or in other domain.
Here we first define the symbols that will be used in this
section. Let X
1
, X
2
. . . X
N
be the changes of the coefficients in a
CHAPTER 3. NOVEL WATERMARKING SCHEMES 57
discrete video frame due to watermarking. The power constrain
of these values are the asking bounds determined by the source
coefficient values S
1
, S
2
. . . S
N
. We define a masking function
f s.t. E(XX
T
) ≤ f(S) where X = [X
1
, X
2
. . . X
N
]
T
and S =
[S
1
, S
2
. . . S
N
]
T
. In the receiver ed, consider Y = S
W
- S =
X - Z where Z are the noises added to the coefficients during
transmission.
Capacity = C
Host data = S = [S
1
, S
2
. . . S
N
]
T
Watermark = X = [X
1
, X
2
. . . X
N
]
T
Power constrain = E(XX
T
) ≤ f(S)
Noise = Z
Y = S
W
−S = X −Z (3.11)
Then, the maximum capacity of these multi-variant symbols in
Equation 3.12. We can assume X and Z are independent.
C = Max
p(x):E(XX
T
)≤f(s)
I(X; Y ) (3.12)
= Max
p(x)
[h(Y ) −h(Y [X)] (3.13)
= Max
p(x)
[h(Y ) −h(Z)] (3.14)
where p(.) represents any probability distribution, I(.;.) rep-
resents mutual information and h(.) represents the differential
entropy.
According to Theorem 9.6.5 in Elements of Information The-
ory [110], Y has zero mean and covariance K = E(XX
T
), the
differential entropy of Y, i.e., h(Y) satisfies the following
h(Y ) ≤
1
2
log(2Πexp)
n
[K[ (3.15)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 58
with equality iff Y ∼ N(0, K) and [.[ is the absolute value of the
determinant. Here, this theorem is valid no matter what the
range of K is.
Therefore, from 3.12 and [E(XX
T
) + E(ZZ
T
)[ = [f(s) +
E(ZZ
T
)[, we can see that
C =
1
2
log(2Πexp)
n
[f(S) +E(XX
T
)[ −h(Z) (3.16)
where we assume f(S) is diagonal and nonnegative s.t. [E(XX
T
)+
E(ZZ
T
)[ ≤ [f(s) +E(ZZ
T
)[. This assumption means that em-
bedded watermark values are mutually independent.
Equation 3.16 is the watermarking capacity in a variant-state
channel without specifying any type of noise. It is he capacity
given a noise distribution. If we look at Equation 3.16 and
Theorem 9.6.5 in [110] again, for all type of noise, we can find
that C will be at least
C
min
=
1
2
log(2Πexp)
n
[f(S) +E(ZZ
T
)[ (3.17)

1
2
log(2Πexp)
n
[E(ZZ
T
)[ (3.18)
=
1
2
[f(S) +E(ZZ
T
)
−1
+I[ (3.19)
When noise is Gaussian distribution. If we further assume that
noise are also independent in samples, then the watermarking
capacity will be:
C
min
=
n

i=1
1
2
log(1 +
P
i
N
i
) (3.20)
=
n

i=1
1
2
log(1 +
P
i
σ
2
n
) (3.21)
where P
i
and N
i
are the power constrains in the i
th
coefficient,
respectively. It is inserting that even though we use the multi-
CHAPTER 3. NOVEL WATERMARKING SCHEMES 59
variants to derive 3.20 instead of using Parallel Gaussian Chan-
nels, their results are the same in this special case.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 60
3.3 A Hybrid Watermarking Scheme
In the previous section, a novel scene-based watermarking scheme
is proposed, which is resistant to the attacks of the video proper-
ties, including frame averaging, frame dropping, and statistical
analysis. However, the scheme dose not improve the robustness
against the attacks by image processing on the video frames.
Therefore, we propose a hybrid approach to improve the per-
formance and the robustness of the watermarking scheme based
on the conclusion drawn from the survey and the properties of
a video.
The scene-based watermarking scheme can be improved with
two type of hybrid approaches: visual-audio hybrid watermark-
ing and hybrid with different watermarking schemes. Figure 3.7
shows the overall framework of the proposed scheme.
The visual-audio hybrid watermarking scheme applies both
video and audio watermarks in a video. Error correcting codes
are extracted from the video watermark and embedded as audio
watermark in the audio stream. This approach takes the ad-
vantage of watermarking the audio channel, because it provides
an independent means for embedding the error correcting codes,
which carry extra information for watermark extraction. There-
fore, the scheme is more robust than other schemes which only
use video channel alone. The hybrid approach with different wa-
termarking schemes can further be divided into to two classes:
independent scheme and dependent scheme. From the survey,
we find that no watermarking scheme can resist all watermark
attacks; hybrid with different watermarking schemes can be one
CHAPTER 3. NOVEL WATERMARKING SCHEMES 61
Scene based
Watermarking
Hybrid
Approach
Visual-audio
Hybrid
Watermark
Hybrid with
Different
Watermark
Independent
Watermark
Dependent
Watermark
Figure 3.7: Possible improvement for scene based watermarking scheme
of the solutions. It takes advantages of various watermarking
schemes by combining them in different ways.
3.3.1 Visual-audio Hybrid Watermarking
The visual-audio watermarking scheme combines a video wa-
termark and an audio watermark. We embed error correcting
codes of a video watermark as an audio watermark and refine
the retrieved video watermark during detection [11]. Figure 3.8
shows an overview of our visual-audio watermarking process. In
C
H
A
P
T
E
R
3
.
N
O
V
E
L
W
A
T
E
R
M
A
R
K
I
N
G
S
C
H
E
M
E
S
6
2
Video
Normalization
Scene Change Detection
DWT
DWT
Watermark
Crop
Encryption
Key
Encryption
Embedding Algorithm
if W[j] = 1,
Exchange C[i] with max(C[i], C[i+1], C[i+2], C[i+3], C[i+4])
else
Exchange C[i] with min(C[i], C[i+1], C[i+2], C[i+3], C[i+4])
Encrypted
Watermark
Watermarked
Video
8-bit planes
Bit Decomposition
Place the bit-plane
side by side
Encrypted
Watermark
Splitter
Video
Stream
Error Correcting
Code Extraction
Audio
Watermarking
IDWT
Merger
DWT
DWT
System
Stream
m0'
m0
m7
m7'
Audio
Stream
F
i
g
u
r
e
3
.
8
:
O
v
e
r
v
i
e
w
o
f
v
i
s
u
a
l
-
a
u
d
i
o
h
y
b
r
i
d
w
a
t
e
r
m
a
r
k
i
n
g
s
c
h
e
m
e
CHAPTER 3. NOVEL WATERMARKING SCHEMES 63
our scheme, an input video is split into audio and video streams,
which undergo separate watermarking procedures. On the one
hand, a video watermark is decomposed into various parts, em-
bedded in corresponding frames of different scenes in the original
video. On the other hand, error correcting codes are extracted
from the watermarks and embedded as an audio watermark in
the audio channel, which in turn makes it possible to correct and
detect the changes from the extracted video watermarks. This
additional protection mechanism enables our scheme to over-
come the corruption of a video watermark, thus the robustness
of the scheme is preserved under certain attacks.
Audio Watermark
The watermark embedded in the audio channel provides the er-
ror correction and detection capability for the video watermark.
In the detection phase, it would be extracted and used for re-
fining the video watermark. Disparate error correction coding
techniques can be applied, such as Reed-Solomon coding tech-
niques [103] and Turbo coding [111].
Error correcting codes play an important role in watermark-
ing, especially when the watermark is damaged significantly.
Error correcting codes overcome the corruption of a watermark,
and make the watermark survive through serious attacks. More-
over, our scheme benefits from audio watermarking as it pro-
vides an independent channel for embedding the error correct-
ing codes, which carry extra information for video watermark
extraction.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 64
The key to error correction is redundancy. The simplest error
correcting code is repeating everything several times. However,
in order to keep the audio watermark inaudible, we cannot em-
bed too much information into an audio channel. In our scheme,
we apply averaging to obtain the error correction code. Within
a small region of an image, the pixels are similar. Hence, an
average value of a small region can be fully utilized to estimate
the pixels within the particular region. The average value of the
pixels in each region is calculated as Equation 3.22
x

i=0
y

j=0
W
jz+rx+syz+i
(3.22)
where W is a pixel in the image, k is the k
th
block of the average
image, (r, s) is coordinate of region k, z is the width of the image,
(x, y) is the coordinate of the pixel in region k, and x y is the
size of the block. A sample is shown in Figure 3.9.
Figure 3.9: (a) Original video watermark (b) Visualization of averaging (c)
Audio watermark (average of a)
CHAPTER 3. NOVEL WATERMARKING SCHEMES 65
Audio Watermarking Embedding
In the audio channel, the audio watermarking is based on Mod-
ulated Complex Lapped Transform (MCLT) [112]. The MCLT
is a 2x over sampled DFT filter bank, used in conjunction with
analysis and synthesis windows that provide perfect reconstruc-
tion. The MCLT is well suited for noise suppression and echo
cancellation.
The MCLT is based on the oddly-stacked time-domain alias-
ing cancellation (TDAC) filter bank introduced by Princen, John-
son, and Bradley [113]. Its basis functions can be obtained by
cosine modulation of smooth windows, in the form [114].
After the wave is extracted from the audio channel, it is trans-
formed from original time domain to frequency (MCLT) domain.
The magnitude then is modulated according to the prepared wa-
termark in the previous section.
After addition of the watermark, we generate the time-domain
watermarked audio signal by combining the marked vector

y
with the original phase of

x , and passing those modified frames
through the inverse MCLT. Figure 3.10 shows the detail. For
the typical 44.1 kHz sampling, we use a length-2048 MCLT.
Only the MCLT coefficients within the 2-7 kHz subband are
modified and considered in the detection process, to minimize
carrier noise effects as well as sensitivity to downsampling and
compression.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 66
Block of Time Domain Sample of
the Original Content
MCLT
Magnitude
Block of Frequency (MCLT)
Domain Sample of the Original
Content
Phase
IMCLT
Block of Time Domain Sample of
the Watermarked Content
Magnitude
Block of Frequency (MCLT)
Domain Sample of the
Watermarked Content
Figure 3.10: Audio watermark embedding with MCLT
Watermarked Frame and Wave
After applying the algorithm described in the previous sections
on both video and audio channel, the watermarking process for
a video is completed. Figure 3.11 (a) shown the one of the orig-
inal frame, and Figure 3.11 (b) shown the watermarked frame.
Figure 3.12 shown the part of the original wave form, and Figure
3.13 shown the watermarked wave form.
Watermark detection
The watermark is detected through the process whose overview
is shown in Figure 3.14. A test video is split into video stream
and audio stream, and watermarks are extracted separately by
audio watermark extraction and video watermark extraction.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 67
(a) (b)
Figure 3.11: One of the (a) original video frame and (b) watermarked video
frame
Then the extracted watermarks undergo a refining process.
The video stream is processed to get the video watermark.
At the same time, error correcting codes are extracted from the
audio stream and the video watermark extracted is refined by
this information with the Equation 3.23:
RW
ij
=
EW
ij
f +Avg
k
g
f +g
(3.23)
where RW
ij
is the refined watermark, EW
ij
is the extracted
video watermark from Equation (7), Avg
k
is the extracted au-
dio watermark, k is the k
th
block of the average image, (i, j) is
coordinate of the video watermark, and f : g is a ratio of impor-
tance of the extracted video watermark to the audio watermark.
In all the subsequent experiments, we assume f = 0.5 and g =
0.5, f +g = 1.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 68
0 500 1000 1500 2000 2500 3000
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 3.12: Block of samples of the original wave content
0 500 1000 1500 2000 2500 3000
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Figure 3.13: Block of samples of watermarked wave content
CHAPTER 3. NOVEL WATERMARKING SCHEMES 69
Audio Watermark
Extraction
Audio
Stream
Video
Stream
Video Watermark
Extraction
Splitter
Watermark
Refining
Video
Watermarked
Video
Figure 3.14: Overview of detection of the watermark
3.3.2 Hybrid Approach with Different Watermarking
Schemes
No watermarking scheme is found in the current literature to be
capable of resisting all watermark attacks. The hybrid approach
can be a possible solution. As stated earlier, it can be classified
into independent schemes and dependent schemes. Independent
watermarking schemes include either different schemes for dif-
ferent scenes or different schemes for different parts of the frame.
Dependent watermarking schemes embed a watermark in each
frame with several different schemes.
We propose three approaches for the hybrid watermarking
schemes. They combine alien schemes in disparate ways. Four
watermarking schemes are chosen, each of which strives a dif-
ferent set of attacks. These four schemes are: Discrete Wavelet
Transform based watermarking (DWT), Discrete Cosine Trans-
form based watermarking, (DCT), Discrete Fourier Transform
CHAPTER 3. NOVEL WATERMARKING SCHEMES 70
based watermarking (DFT), and Radon Transform based water-
marking (RADON). As they embed the watermark in various
domains, their robustness properties are preserved. By combin-
ing the advantages of these watermarking schemes systemati-
cally, various kinds of attacks can be resisted altogether.
Independent Hybrid Watermarking
In the independent hybrid watermarking approach, the applied
watermarking schemes do not affect each other. We propose
two approaches to combine the employed watermarking schemes:
Different schemes for different scenes, and different schemes for
different parts of each frame.
Different schemes for different scenes In this approach, a water-
mark is still decomposed into different parts which are embed-
ded in the corresponding frames of different scenes in the original
video. Each part of the watermark, however, is embedded with a
different watermarking scheme. Figure 3.15 illustrates the idea.
Within a scene, all the video frames are watermarked with the
same part of a watermark by the same watermarking scheme.
When there is an attack on the watermarked video, different
watermarking schemes are resistant against it. Consequently,
some parts of the watermark still survive after the attack. This
approach thus enhances the chance of survival under several at-
tacks, and raises its robustness. The merit is that only part of
the watermark is damaged if the watermarked video is attacked,
provided that at least one of the watermarking schemes is resis-
tant against the attack. The disadvantage of this approach is
CHAPTER 3. NOVEL WATERMARKING SCHEMES 71
Figure 3.15: Hybrid approach with different scheme for different scene
that the accuracy of the extracted watermark is lower, compared
with the other schemes specified to a particular attack.
Different schemes for different parts of each frame This approach
is similar to the previous approach. However, four different wa-
termarking schemes are applied to each frame instead of dif-
ferent schemes for different scenes. The idea is illustrated in
Figure 3.16, where each video frame is divided into four parts,
and the watermark for that frame is also divided into four parts.
Then, each part of the watermark is embedded into the frame
in different domains.
When a watermarked video is attacked, part of the water-
mark in each frame may still survive. Therefore, information
for every part of the watermark can be retrieved, and the water-
mark can be approximately estimated. Although the accuracy of
the extracted watermark is reduced, it is more resistant against
attacks.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 72
Figure 3.16: Hybrid approach with different scheme for different part of frame
Dependent Hybrid Watermarking
In the dependent hybrid watermarking approach, each frame is
embedded with the same watermark serially by different water-
marking schemes in various domains. In our approach, again
four different watermarking schemes are applied. The frame is
first transformed to wavelet domain, and the watermark is em-
bedded. Then the frame is inversely transformed to frequency
domain by DCT to embed the same watermark, so on and so
forth for DFT and Radon transform, subsequently.
As the watermarks are embedded in different domains, they
can compensate each other in resisting against different attacks.
Nevertheless, different watermarking schemes may also affect
each other when under attacks.
Experiments are done to evaluate the effectiveness of different
approach and it is presented in Chapter 4.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 73
3.4 A Genetic Algorithm-based Video Wa-
termarking Scheme
The problem of designing a feasible watermarking scheme can be
viewed as an optimization problem with three conflicting goals:
higher fidelity (media quality index), better robustness (water-
mark strength) and larger data capacity. In the previous section,
we discussed applying a hybrid watermark approach to improve
the robustness of the scheme. In this section, genetic algorithm
(GA) is employed to enhance the fidelity [115].
Fidelity
Capacity
Robustness
Figure 3.17: The graph of three mutually orthogonal axes representing the
capacity, robustness and fidelity of the watermarking scheme
In general, higher fidelity, better robustness, and larger data
capacity is three goals that most watermarking scheme would
like to achieve. However, since these requirements are conflict-
ing with each other, designing an optimal watermarking scheme
CHAPTER 3. NOVEL WATERMARKING SCHEMES 74
Achievable
Non-achievable
region
Robustness
Fidelity
Figure 3.18: The graph of two mutually orthogonal axes representing the
robustness and fidelity of the watermarking scheme
has become an inherently difficult and interesting problem. For
instant, the fidelity requirement often limits the strength of em-
bedded signals, which consequently constraints the robustness
of a watermarking scheme against common or malicious manip-
ulations.
In the existing watermarking schemes, taking the application-
specific characteristics into consideration makes reasonable trade-
offs among the three requirements [115]. In other words, one can
view the embedding process as a selection of feasible embed-
ding points within an acceptable region in the space spanned
by three mutually orthogonal axes representing the prescribed
requirements, respectively as shown in Figure 3.17.
However, in most watermarking schemes, the process of de-
CHAPTER 3. NOVEL WATERMARKING SCHEMES 75
ciding the trade-off is always done in a heuristic manner, lacking
systematic explorations or even optimizations. In this section,
a GA-based watermark fidelity enhancing method is proposed,
which optimize the performance towards the non-achievable re-
gion as show in Figure 3.18
3.4.1 Watermarking Scheme
In the area of evolutionary computation, GA is an important
optimization technique [115]. Here we employ this techniques
to optimize the performance of our proposed scheme. To model
the problem as a GA problem, the fitness function, chromosome
and GA operators should be defined. In GA-based optimiza-
tions, the problem to be addressed is defined as an objective
function that indicates the fitness of any possible solution. Ac-
cording to the problem specific constraints, a population of can-
didate initialized, named as chromosome, which is a finite-length
strings.
During practical GA-based optimization processes, three GA
operators: reproduction, crossover, and mutation are applied to
the chromosomes repeatedly. Reproduction is a process in which
individual chromosomes are copied according to their objective
function values, that is, the fitness values. Chromosomes with
higher fitness values have higher probability to contribute more
offspring in the next generation. The objective function decides
the probability of chromosomes’ survival or removal during the
competition. Crossover is the procedure that pairs of chromo-
somes exchange portions of their genes to form new chromo-
CHAPTER 3. NOVEL WATERMARKING SCHEMES 76
somes, and consequently, new parameters other than the initial
ones can be produced for evaluation. Mutation is the occasional
random alternation of the value in some positions of chromo-
somes. Mutation can be regarded as a random walk through the
parameter space. By sparingly using mutation, chromosomes
with good performance but not obtainable by reproduction and
crossover may also have the chance to be selected. These op-
erators are used repeatedly to obtain successive generations of
chromosomes. [116]
Within a generation, only the chromosomes with the higher
fitness values can survive. This portion of those will be pass as
parent chromosomes to the next generation. After numbers of
generation, the chromosomes are optimized. We can obtain the
near-optimal solutions of modelled problem. [115].
3.4.2 Problem Modelling
By applying the GA to the scene-based video watermarking
scheme, the watermarked video quality is improved while still
keep the robustness of the watermark against image manipula-
tion.
The embedding position of the different parts of watermark
within the video are simulated as chromosomes. Then several
genetic-algorithmic operations are applied to optimize the video
frame quality after embedding. In our experiments, we use
image quality indicators, Mean Absolute Difference (MAD) to
measure the objective function values during optimization.
We choose the watermark embedding positions within a video
CHAPTER 3. NOVEL WATERMARKING SCHEMES 77
Figure 3.19: A illustrative diagram for GA-based optimization process
as our search space, i.e., search a scene that a particular part
of watermark is best to embed to. Then apply the genetic-
algorithmic operators to find the best combination of watermark
and video scene. A illustrative diagram is shown in the Figure
3.19. Assume a video is consists 8 scenes. The watermark is
scrambled into 8 parts and embedded into different scenes. The
optimal combination of watermark and scene are shown in the
chromosome.
By repeatedly applying the GA operations to each original
video frame and watermark image, we can get near-optimal em-
bedding positions for the original frame and the watermark im-
ages. Figure 3.20 shows the details of the GA-based optimiza-
tion process for each video scene.
In the watermark scheme, the input video frames are trans-
formed to frequency domain and the middle-frequency range
coefficients of the watermark are modified according to the wa-
termark. The basic idea is that the human eyes are sensitive to
the low frequency noise and the quantization step of lossy com-
pression may discard the high frequency components; therefore,
CHAPTER 3. NOVEL WATERMARKING SCHEMES 78
Figure 3.20: The GA-based optimization process for part of watermark
CHAPTER 3. NOVEL WATERMARKING SCHEMES 79
the reasonable trade-off is to embed the watermark into the
middle-frequency range of the video frames, i.e. watermarks are
embedded by modifying the middle-frequency coefficients within
each frame block of the original frame.
3.4.3 Chromosome Encoding
Assume we want to embed a part of the watermark image into
one of the scene in a video. The position of the scenes are encode
as a chromosome. The number of scene in the video is M and
we can user log
2
M bit to represent the position of the scene.
The scene of the i
th
watermark is best to be embedded to X
i
in
the video can be defined as:
¦(X
i
) [ 0 ≤ X
i
< M, X
i
,= X
j
if i ,= j and M > 1¦ (3.24)
The last two constraints of Equation 3.24 imply:
1. There are at least two scene-changes in the video, otherwise,
the search space of the GA is 1 and no optimization can be
done.
2. In a video, the video frames that have been embedded
should not be embedded again.
Figure 3.21 shows the sample chromosome which represents
the positions of the watermark that should be embedded to.
There are 8 scene changes in the video and 3-bit number need
to be used to represent the position of video scenes; therefore, a
24-bit chromosome represents the sequence of video scene to be
used.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 80
1 0 0 1 1 1 0 0 0 1 1 0 0 0 1 0 1 1 0 1 0 1 0 1
X0 X1 X2 X5 X4 X3 X6 X7
Figure 3.21: A 24-bit chromosome represents the sequence of the scenes to
embed
3.4.4 Genetic Operators
The initial chromosomes are generated with the above specifica-
tion. Then, the genetic-algorithmic operators are apply to the
chromosomes for optimization with the fitness function, they
are: reproduction, crossover, and mutation.
Fitness function
The fitness function f is a measure of profit we need during op-
timization. This application of GA aims to improve the fidelity,
i.e. the image quality, of the video frames, the reciprocal of
image similarity indicators (maximum absolute different MAD)
is chosen to be the fitness function. The definition of fitness
function f will be:
f =
1

7
x=0

7
y=0
[I
/
(x, y) −I(x, y)[
(3.25)
Where I
/
and I are the intensity values of the same pixel position
within a video frame after and before embedding, respectively.
After calculation of the fitness function value (defined in
Equation 3.25) for each parent chromosome, we employ the
biased-roulette-wheel method [refGA] in GA to generate N chil-
dren. The basic idea is that the higher a parent chromosome’s
CHAPTER 3. NOVEL WATERMARKING SCHEMES 81
fitness function value is, the higher probability it has to con-
tribute one or more offspring in the next generation.
Reproduction
Reproduction is a process that the children chromosomes are
generated according to the fitness function values of their par-
ents.
Crossover
After N legal parent chromosomes and M children chromosomes
are generated, crossover operator is applied to on these children
chromosomes. Firstly, the recently reproduced chromosomes are
randomly mated in a pair. Then an integer position p between
1 and the chromosome length (L) minus 1 is selected at random
for each pair where p ≤ L−1. Finally, each chromosome swaps
all the bits between the location of p+1 and chromosome length
with its mate.
Mutation
The last operator is mutation. Mutation is the random change
of bit values with small probability within a chromosome. Mu-
tation introduces some randomness into the optimization, thus,
new combination of watermark and video scene are generated.
This increases the chance of approaching the optimal, otherwise,
the optimization process is very slow. In our experiment, we use
0.05 as the mutation rate, i.e. a bit may change polarity (take
complement) at a probability of 0.05.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 82
Figure 3.22: The GA-based watermarking algorithm
After performing all the GA operations, more children chro-
mosomes may not satisfy the constraints of embedding positions.
After the checking and discarding process, we can get M legal
children chromosomes (M ≤ N). From the M + N chromo-
somes (M children and N parents), select N chromosomes with
the larger fitness function values to be the next parent genera-
tions.
Repeat all the operations mentioned above until the number
of generations specified has been done, and choose the best chro-
mosome, i.e. the one with the largest fitness function value, to
be the sequence of video scenes to be used. Figure 3.22 depicts
the block diagram of the proposed watermarking algorithm with
GA optimization.
The original video and the watermark are input to the sys-
tem. Firstly, the scene change analysis is applied to the video
and the watermark is preprocessed. Then, the GA-optimization
process will find out the almost-optimal combination of water-
CHAPTER 3. NOVEL WATERMARKING SCHEMES 83
mark and video scene. The watermarks are embedded in to the
video frames according to the result of the GA optimization. For
detection phase, the GA information should be passed to the
detector. The watermarked video is passed in to the detection
system; the watermark is extracted with the GA information
provided by the embedding phase. The original video frame,
the watermarked video frame with scene-based watermarking
scheme, and the the watermarked video frame with GA-based
watermarking scheme are shown in Figure 3.23. There is no sig-
nificant perceptual different between in Figure 3.23 (b) and (c).
However, the enhancement can be indicated by measuring with
the quality index and this experimental results will be shown in
the following section.
2 End of chapter.
CHAPTER 3. NOVEL WATERMARKING SCHEMES 84
(a)
(b) (c)
Figure 3.23: Comparison between watermarked video with and without GA
optimization a) Original video frame (b) Video frame watermarked with
scene-based scheme (c) Video frame watermarked with GA-based scheme
Chapter 4
Experimental Results
In this chapter, we present our experimental results on the scene-
based watermarking scheme, the hybrid watermarking scheme
and the GA-based watermarking scheme. The experiments are
basically divided into 2 types: test on robustness and test on
fidelity. In the following sections, we present the implementation
detail of the proposed schemes and the experimental results.
4.1 Test on Robustness
In this section, the robustness of the scene-based watermark-
ing scheme and the hybrid watermarking scheme is tested. To
implement the proposed watermarking scheme, the software Vir-
tualDub [117] is employed. The performance of the new video
watermarking scheme is evaluated through several experiments:
the experiment with various dropping ratio, the experiment with
various number of frame colluded, the experiment with various
quality factor of MPEG, and the test of Robustness with Stir-
Mark 4.0. Another DWT-based watermarking scheme, which
85
CHAPTER 4. EXPERIMENTAL RESULTS 86
embeds an identical watermark in all frames, is implemented
to compare with the proposed scheme. A video clip with 1526
frames of size 352 288 is used in our experiments. The video
consists of 10 scene changes. The experiments are done on a
desktop computer with Pentium 4 CPU 2.00GHz and 512MB
RAM.
Distinguishable attacks, including frame dropping, frame av-
eraging, lossy compression, and StirMark 4.0 [14], are carried out
to the watermarked video to test the robustness of our scheme.
The audio channel is also attacked by adding some noises into
it. After extracting and refining the watermarks, a quantitative
measurement is required to provide an objective judgment of
the extraction fidelity. Therefore, a similarity measurement of
the extracted and the referenced watermarks can be defined as
Equation 4.1, Normalized Correlation:
NC =

i

j
W
ij
RW
ij

i

j
W
2
ij
(4.1)
which is the cross-correlation normalized by the reference wa-
termark energy giving unity as the peak correlation [118], where
W
ij
is the original watermark and RW
ij
is the refined watermark
from Equation 3.23. We use this measurement to evaluate our
scheme in the experiments.
The NC values are retrieved when the watermarked video is
facing different attacks. The experimental results are described
in detail in the following sections.
CHAPTER 4. EXPERIMENTAL RESULTS 87
4.1.1 Experiment with Frame Dropping
As a video contains a large amount of redundancies between
frames, it may suffer attacks by frame dropping. This experi-
ment is aimed at examining the robustness of the scheme under
the frame dropping attack. Different percentages of frames are
dropped and the obtained results are shown in Figure 4.1.
0 10 20 30 40 50 60 70 80
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
% of Frame Dropped
NC Values under Frame Dropping
DWT−based watermarking scheme
Scene based watermarking scheme
Visual−audio hybrid watermarking scheme
Visual−audio scheme with audio attack
Hybrid approach with different scheme for different scene
Hybrid approach with different scheme for different part of frame
Hybrid approach with dependent watermark
Figure 4.1: NC values under frame dropping
From the experiment, we find that the scheme achieves better
performance than the DWT-based scheme without scene-based
watermarks [119, 120]. It is because in each scene, all frames
are embedded with the same watermark. It prevents the attack-
ers from removing the watermark by frame dropping. If they
try to remove one part of the watermark, they need to remove
CHAPTER 4. EXPERIMENTAL RESULTS 88
the whole trunk of frames (i.e., the whole scene), leading to a
significant damage to the video. In addition, when the frames
are dropped, the error is only introduced to a corresponding
small part of the watermark. For the DWT-based scheme (i.e.,
non-scene-based), however, the error is introduced to the whole
watermark, making the performance worse.
The performance of the scheme is significantly improved by
combining with an audio watermark, the visual-audio water-
marking scheme, especially when the dropping rate of video
frame is high. Due to the increased dropping rate, the error of
the extracted watermark is increased, which significantly dam-
ages the watermark. The error correcting codes from the audio
watermark provide information to correct the error and recover
the corruption of the video watermark. Consequently, the NC
values of the watermark are higher than these without the er-
ror correcting codes. Moreover, the error correcting codes are
embedded in the audio channel. As frame dropping would not
affect the audio channel much, our scheme benefits by allowing
uninterrupted error correcting codes to refine the watermark and
improve the NC values.
When the audio channel is also attacked, the NC values of
the extracted watermark are decreased. The error correcting
codes in the audio channel are altered by the attack. Although
the capability to recover the error in the video watermark is
dropped, the result is still better than the scheme without an
audio watermark, as the attacked audio watermark still contains
some information to recover the watermark in the video channel.
CHAPTER 4. EXPERIMENTAL RESULTS 89
4.1.2 Experiment with Frame Averaging and Statisti-
cal Analysis
Frame averaging and statistical analysis is another common at-
tack to the video watermark. When attackers collect a number
of watermarked frames, they can estimate the watermark by
statistical averaging and remove it from the watermarked video
[121, 122]. The scenario is shown in Figure 4.2.
Noise
Estimation
Noise
Estimation
Noise
Estimation
Video
Frames
+
Frames to be
attacked
Y1
Y2
Yn
W1
W2
Wn
+
W
-
Y
Figure 4.2: Scenario of statistical averaging attack
Experiments have been conducted to evaluate the proposed
scheme under this attack, and the results are shown in Figure
4.3. It is found that the proposed scheme can resist to statis-
tical averaging quite well. This is because our scheme crops a
watermark into pieces and embeds them into different frames,
making the watermarks resistant to attacks by frame averaging
for the watermark extraction. The identical watermark used
within a scene can prevent attackers from taking the advan-
tage of motionless regions in successive frames and removing
the watermark by comparing and averaging the frames statisti-
CHAPTER 4. EXPERIMENTAL RESULTS 90
cally [123]. On the other hand, independent watermarks used
for successive, yet different scenes can prevent the attackers from
colluding with frames from completely different scenes to extract
the watermark.
0 100 200 300 400 500 600 700 800 900 1000
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
NC values under statistical averaging
Number of Frame Colluded
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
Proposed Watermarkng Scheme
DWT−based Watermarking Scheme
Figure 4.3: NC values under statistical averaging
4.1.3 Experiment with Lossy Compression
This experiment is aimed at testing the robustness of the scheme
under attack by lossy compression. Figure 4.4 shows the NC
values of the extracted watermarks with different quality factors
of MPEG.
From the experiment, we note that the proposed scheme im-
proves the robustness for watermark protection. The perfor-
CHAPTER 4. EXPERIMENTAL RESULTS 91
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
NC Values under Lossy Compression
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
MPEG Quality Factor
DWT−based watermarking scheme
Scene based watermarking scheme
Visual−audio hybrid watermarking scheme
Visual−audio scheme with audio attack
Hybrid approach with different scheme for different scene
Hybrid approach with different scheme for different part of frame
Hybrid approach with dependent watermark
Figure 4.4: NC values under lossy compression
mance of the scheme is significantly improved by combining
with audio watermark again, especially when the quality fac-
tor of MPEG is low. This is because when the quality factor of
MPEG is low, the error of the extracted watermark is increased
and the watermark is damaged significantly. As the error cor-
recting codes are provided from the audio watermark, they are
not affected by the lossy compression attack applied to the video
channel. Consequently, the error correcting codes can overcome
the corruption of the video watermark, achiving higher NC val-
ues.
The proposed scheme without audio watermark has similar
performance with the other DWT-based scheme because both
CHAPTER 4. EXPERIMENTAL RESULTS 92
of them satisfy the following condition. Higher frequency DWT
coefficients of the watermark are embedded in the higher fre-
quency part of the video frame and high frequency sub-band
DWT coefficients (HH) of video frame are not watermarked.
This approach makes the watermark survive MPEG lossy com-
pression, as lossy compression removes the details (i.e. the high
frequency components) of the image [124].
The performance of the scheme is also improved by the hybrid
approach with different watermarking schemes. The NC values
of the hybrid approach under lossy compression are higher than
the scheme which only applying single watermarking scheme in a
video. From the survey, we find that the DCT-based watermark-
ing scheme is the most resistant one against lossy compression.
MPEG is based on DCT and lossy compression mostly only
modifies the highest frequency coefficients; therefore, modifying
the mid-band coefficients in this domain during watermarking a
video would reduce the effect of compression on the watermark.
When compression is applied to the watermarked video, the wa-
termark embedded in the video with DCT-based watermarking
scheme survives. Therefore, at least one forth of the watermark
can be retrieved from the video. This increases the robustness
of the scheme. However, the robustness of the scheme is not
improved by the hybrid approach with dependent watermark.
4.1.4 Test of Robustness with StirMark 4.0
StirMark 4.0 [100, 101] is a benchmark for testing robustness
of a watermarking scheme. In this experiment, we employ Stir-
CHAPTER 4. EXPERIMENTAL RESULTS 93
Mark 4.0 to test the robustness of the proposed schemes when
image processing is applied, including cropping, Peak Signal to
Noise Ratio (PSNR), adding noise, median filter, row/column
removal, rescaling, rotation and affine, and compare them with
the current techniques that exist in the literature. Figures 4.5,
4.8, 4.7 and Table 4.1 show the result of the watermarked video
under different attacks from StirMark.
Figure 4.5(d) shows the result of the NC values of the wa-
termark under different cropping ratios. The performance of
the watermarking schemes without audio watermark decreases
greatly with the increasing cropping ratio. When the cropping
ratio is greater than 60 present, the NC values of the retrieved
watermarks are less than 0.6. The visual-audio watermarking
scheme, on the other hand, gives better performance. The NC
values are greater than 0.7 even when the cropping ratio is 90
present. This shows that the audio watermark significantly im-
proves the robustness of the watermarking scheme. However, the
result shows that the performance of scheme is not improved by
the hybrid approach with different watermarking schemes. The
NC values of the hybrid approaches are similar to those of the
DWT-based watermarking scheme.
Figure 4.9 shows the result of the watermarking scheme under
different PSNR. The NC values of the watermarks are decreased
with the PSNR. The proposed scheme reveals improvement in
this experiment as well. The NC values of the DWT-based wa-
termarking scheme slightly higher than the hybrid approach. It
is because the watermark in the wavelet domain is not affected
by PSNR too much, but other watermarking schemes are af-
CHAPTER 4. EXPERIMENTAL RESULTS 94
10 20 30 40 50 60 70 80 90
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
Cropping Ratio
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
DWT−based watermarking scheme
Scene based watermarking scheme
Visual−audio hybrid watermarking scheme
Visual−audio scheme with audio attack
Hybrid approach with different scheme for different scene
Hybrid approach with different scheme for different part of frame
Hybrid approach with dependent watermark
Figure 4.5: NC values under cropping
fected. Thus, the overall performance the hybrid approach is
decreased. This is one of the disadvantages of the hybrid ap-
proaches.
When the watermarked video is rescaled, the proposed scheme
also portrays improvement. Figure 4.7 depicts the NC val-
ues when the watermarked video is rescaled with various fac-
tors. The performance of the scheme is significantly improved
again by visual-audio watermarking scheme, especially when the
rescaling factor is large. Furthermore, the improvement becomes
more evident with the increase of the rescaling factor. This is
because when the rescaling factor increases, the error of the ex-
tracted watermark is increased, which significantly damages the
watermark. The error correcting codes from the audio water-
CHAPTER 4. EXPERIMENTAL RESULTS 95
0 10 20 30 40 50 60 70 80 90 100
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
NC values under PSNR
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
PSNR
DWT−based watermarking scheme
Scene based watermarking scheme
Visual−audio hybrid watermarking scheme
Visual−audio scheme with audio attack
Hybrid approach with different scheme for different scene
Hybrid approach with different scheme for different part of frame
Hybrid approach with dependent watermark
Figure 4.6: NC values under PSNR
mark, on the other hand, provide information to correct the er-
ror and overcome part of the corruption in the video watermark
and produces higher NC values of the recovered watermark.
With the hybrid approaches, the robustness of the scheme is
increased. The NC values of the extracted watermark are higher
than the DWT-based and scene-based watermarking schemes.
In wavelet domain, the coefficients vary when the size of frame
or image is different. The coefficients in the Randon trans-
formed domain, however, do not vary too much when rescal-
ing. As shown in Figure 4.7, the hybrid approaches with differ-
ent schemes perform better than the scene-based watermarking
scheme. The hybrid approach with dependent watermark, nev-
ertheless, performs worse than other hybrid approaches. It is
CHAPTER 4. EXPERIMENTAL RESULTS 96
50 75 90 100 110 150 200
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
NC values under rescaling
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
Rescaling factor
DWT−based watermarking scheme
Scene based watermarking scheme
Visual−audio hybrid watermarking scheme
Visual−audio scheme with audio attack
Hybrid approach with different scheme for different scene
Hybrid approach with different scheme for different part of frame
Hybrid approach with dependent watermark
Figure 4.7: NC values under different rescaling factor
CHAPTER 4. EXPERIMENTAL RESULTS 97
0 10 20 30 40 50 60 70 80 90 100
0.5
0.55
0.6
0.65
0.7
0.75
0.8
0.85
0.9
0.95
1
NC values under different noise added to the watermarked video
N
o
r
m
a
l
i
z
e
d

C
o
r
r
e
l
a
t
i
o
n
Noise
DWT−based watermarking scheme
Scene based watermarking scheme
Visual−audio hybrid watermarking scheme
Visual−audio scheme with audio attack
Hybrid approach with different scheme for different scene
Hybrid approach with different scheme for different part of frame
Hybrid approach with dependent watermark
Figure 4.8: NC values under different noise added to the watermarked video
caused by the dependent properties of the watermarking scheme.
When one watermark is corrupted, it may affect the extraction
of other watermarks in the same frame. Therefore, the NC val-
ues of the extracted watermark are lower than those of other
hybrid approaches.
When additional noise is added to the watermarked video,
the proposed scheme also shows improvement. Figure 4.8 de-
picts the NC values when different noises are added to the wa-
termarked video. The performance of the scheme is significantly
improved by combining with an audio watermark, especially
when more noises are added.
There are several tests from the StirMark 4.0. The result
is summarized in Table 4.1. The proposed video watermarking
CHAPTER 4. EXPERIMENTAL RESULTS 98
scheme also shows improvement when the videos are under other
attacks, including: row removal, rotation, PSNR, affine. . . etc.
Attack Class DWT-based Scene-based Visual-audio hybrid Visual-audio hybrid
watermarking scheme watermarking scheme watermarking scheme watermarking scheme
with audio attack
Lossy Compression 0.61 0.62 0.82 0.69
PSNR 0.72 0.76 0.86 0.80
Add Noise 0.63 0.60 0.76 0.67
Median Filter 0.54 0.54 0.74 0.60
Row / Column Removal 0.69 0.71 0.85 0.75
Cropping 0.68 0.66 0.78 0.70
Rescale 0.63 0.62 0.75 0.69
Rotation 0.60 0.61 0.73 0.67
Affine 0.55 0.55 0.78 0.70
Overall 0.62 0.63 0.78 0.69
Attack Class Hybrid approach with different Hybrid approach with different Hybrid approach with
scheme for different scene scheme for different part of frame dependent watermark
Lossy Compression 0.71 0.72 0.68
PSNR 0.82 0.81 0.81
Add Noise 0.70 0.69 0.64
Median Filter 0.55 0.52 0.52
Row / Column Removal 0.77 0.78 0.74
Cropping 0.72 0.69 0.67
Rescale 0.71 0.68 0.63
Rotation 0.69 0.66 0.64
Affine 0.73 0.71 0.63
Overall 0.71 0.70 0.66
Table 4.1: Robustness comparison between different watermarking schemes
4.1.5 Overall Comparison
From the above results, the effectiveness of the scene-based hy-
brid schemes are demonstrated. The scene-based watermark-
ing scheme achieves higher NC values when attacks based on
video properties are launched. This indicates that the water-
CHAPTER 4. EXPERIMENTAL RESULTS 99
marking scheme work well by applying scene change detection
with scrambled watermarks. The performance of the scheme is
further improved by combining with an audio watermark, espe-
cially when the video watermark is corrupted, such as the attack
by lossy compression. When audio channel is also attacked, the
error correction information is altered. The overall performance,
however, still shows improvement. The robustness of the scheme
is also raised by engaging other hybrid approaches.
CHAPTER 4. EXPERIMENTAL RESULTS 100
4.2 Test on Fidelity
In this section, we focus on evaluating the performance of the
GA-based Watermarking Scheme. In the experiment, we prove
the fidelity enhancing effectiveness of the proposed optimization
process.
The GA-based watermarking scheme is implemented with
the GAlib [125]. To evaluate the fidelity of the watermarking
scheme, the peak signal-to-noise ratio (PSNR) and maximum
absolute difference (MAD) is used.
PSNR is given by
PSNR = 10 log
10
255
2
σ
2
q
[dB] (4.2)
where σ
2
q
is the mean square of the difference between the orig-
inal video frame and the watermarked one [126].
MAD is given by
MAD =
7

x=0
7

y=0
[I
/
(x, y) −I(x, y)[ (4.3)
Where I
/
and I are the intensity values of the same pixel position
within a video frame after and before embedding, respectively.
The performance of the GA-based video watermarking scheme
is evaluated through several experiments with different number
of generation in the GA-optimization process. Then, the quality
of the video is evaluated with PSNR and MAD. The quality of
the video watermarked by the scene-based watermarking scheme
and the hybrid watermarking scheme are also evaluated, and
compared with the GA-based scheme. In the experiment, two
CHAPTER 4. EXPERIMENTAL RESULTS 101
video clips are used. One of the video clips has 1526 frames of
size 352288 and it consists of 10 scene changes. Another video
clip has 4236 frames of size 352 288 and it consists of 22 scene
changes. The experiments are done on a desktop computer with
Pentium 4 CPU 2.00GHz and 512MB RAM.
4.2.1 Parameter(s) Setting
Table 4.2 shows the parameters setting for the GA-based video
watermarking experiments.
Table 4.2: Parameters Setting for GA-based experiment
Parameter Value
Population size 100
Mutation probability 0.05
Crossover probability 0.9
Score frequency 1
Flush frequency 25
In the experiment,we have used 0,
Number of generations 1, 5, 10, 20, 40, 100,
200, 400 and 600
Table 4.3 shows the computation time of the GA-based scheme.
4.2.2 Evaluate with PSNR
PSNR measures the signal to noise ratio of the watermarked
video, thus, we can evaluate the its fidelity. From the graph
4.9, it’s clear that the GA-based algorithm successfully reduces
the video frame distortion due to watermark embedding. As
CHAPTER 4. EXPERIMENTAL RESULTS 102
Table 4.3: The computation time of the GA-based scheme
Number of Computation time Computation time
Generation of Video 1 (s) of Video 2 (s)
0 50 141
5 62 179
10 80 234
20 103 289
40 124 354
100 156 423
200 185 534
400 226 620
600 259 702
the number of generations increases, the improvement of video
quality gradually approaches to a saturation value.
Table 4.4 depicts a comparison of PSNR with different wa-
termarking scheme. The PSNR of the video is reduced to 3/4
after GA is applied in optimizing the fidelity of the scheme. It
shows that the GA-based optimization effectively improves the
performance of the scheme.
4.2.3 Evaluate with MAD
MAD measures the different between the original video and the
watermarked video, thus, we can evaluate the quality of the
watermarked video. From the graph 4.10, it shows how the it-
erative generation numbers affect the optimizing performance.
The MAD of the watermarked video is decreased with the GA
CHAPTER 4. EXPERIMENTAL RESULTS 103
0 100 200 300 400 500 600
20
25
30
35
40
45
50
55
60
65
70
Generation number
P
S
N
R

(
d
B
)
PSNR of the video under different GA generation
Video 1
Video 2
Figure 4.9: PSNR of the video under different GA generations
Table 4.4: PSNR comparison between different watermarking schemes
Watermarking scheme PSNR of Video 1 PSNR of Video 2
Scene-based Watermarking scheme 40 33
Visual-audio hybrid watermarking 41 33
Hybrid approach with different scheme for different scene 43 34
Hybrid approach with different scheme for different part of frame 42 36
Hybrid approach with dependent watermark 35 27
GA-based watermarking scheme watermark scheme 52 42
CHAPTER 4. EXPERIMENTAL RESULTS 104
0 100 200 300 400 500 600
2.5
3
3.5
4
4.5
5
5.5
x 10
4
Generation Number
M
A
D
MAD of the video under different GA generation
Video 1
Video 2
Figure 4.10: MAD of the video under different GA generations
generation number. The optimization performance saturate af-
ter about 200 generations. Therefore, the performance of the
proposed GA-based watermarking scheme is converged to opti-
mal.
A comparison of MAD with different watermarking scheme
is shown in Table 4.5. It shows the enhancing effectiveness of
the GA-based watermarking scheme. The MAD of the video is
reduced to 3/5 after GA is applied to optimize the fidelity of the
scheme.
CHAPTER 4. EXPERIMENTAL RESULTS 105
Table 4.5: MAD comparison between different watermarking schemes
Watermarking scheme MAD of Video 1 MAD of Video 2
Scene-based Watermarking scheme 42168 50325
Visual-audio hybrid watermarking 42189 50489
Hybrid approach with different scheme for different scene 43984 52695
Hybrid approach with different scheme for different part of frame 43798 52786
Hybrid approach with dependent watermark 48652 55785
GA-based watermarking scheme watermark scheme 28346 31546
4.3 Other Features of the Scheme
Our proposed scheme is an invisible watermarking scheme. In
the scene-based watermarking scheme, as low frequency sub-
band DWT coefficients (LL) are not watermarked and image en-
ergy is concentrated on the lower frequency wavelet coefficients,
the watermark is perceptually invisible. If these coefficients are
altered, however, the perceptual quality will be affected [123].
Additionally, retrieval of the embedded watermarks does not
need the original video, i.e. a blind watermarking scheme. This
is an important performance feature of the scheme since it takes
a long time to transmit, store, and process the original video.
The experiments show that the proposed scheme is robust
to most of the existing attacks, however, there are still some
weaknesses in our scheme. The computation time of the GA-
based scheme is rather long if the number of GA generation
applied is large or the number of the scene change of the video
increases. Also, when the encoding method is applied again with
another watermark, the proposed scheme is not robust against
CHAPTER 4. EXPERIMENTAL RESULTS 106
it efficiently. The original watermark can only be extracted if
the second watermark is removed first.
4.4 Conclusion
From the experiment, we prove that our proposed scheme en-
hances two of the three prescribed watermarking requirements,
robustness and fidelity. The robustness-enhancement provided
by hybrid scene-based watermarking scheme and the fidelity-
enhancement provided by the GA-based watermarking scheme
are important steps toward a prefect watermarking scheme.
Non-achievable
region
Achievable
region
Robustness
Fidelity
A
A A
Figure 4.11: A conceptual illustration on the performance of the proposed
scheme
Figure 4.11 shows a conceptual illustration. When there are
two orthogonal axes, robustness and fidelity. The robustness is
indicated by the strength of the watermark, and the fidelity can
CHAPTER 4. EXPERIMENTAL RESULTS 107
be represented by quality index. Moreover, the curve represent
the best robustness under the fidelity performance constrain. To
optimize the performance of a watermarking scheme, we try to
move the point towards the curve.
The point A represents the performance of a scene-based wa-
termarking scheme. By applying the hybrid approaches, the
robustness of the scheme is improved and moves the point to
A’. When the watermarking scheme is further enhanced by GA,
i.e. increase the fidelity the scheme, the point moves along the
fidelity axes. Thus, it moves towards A”. Therefore, our pro-
posed scheme is approaching to the ”optimal” embedding con-
figuration.
2 End of chapter.
Chapter 5
Conclusion
This thesis investigates the knowledge of digital video water-
marking techniques for secure multimedia creation and delivery.
After noticing the importance of the multimedia security and
video watermarking in nowadays Internet world and reviewing
the state-of-the-arts technologies of the audio watermarking, im-
age watermarking and video watermarking, an innovative hybrid
digital video watermarking scheme with scene change analysis,
error correcting code and GA optimization is proposed. The pro-
cess of this comprehensive video watermarking scheme, includ-
ing watermark preprocessing, video preprocessing, watermark
embedding, and watermark detection, is described in detail. Ex-
periments are conducted to demonstrate that our scheme is ro-
bust against attacks by frame dropping, frame averaging, and
statistical analysis, and the robustness against the image pro-
cessing attacks is tested with StirMark benchmark. Moreover,
the fidelity of the scheme is evaluated.
Our approach cultivates an innovative idea in: (1) embedding
different parts of a watermark according to scene changes, (2)
108
CHAPTER 5. CONCLUSION 109
embedding its error correcting codes as an audio watermark,
(3) applying a hybrid approach to the proposed scheme, and
(4) employing the GA algorithm to enhance the fidelity. This
approach is never explored in the literature, and its advantages
are clear and significant. The effectiveness of this scheme is
verified through a number of experiments.
To conclude our work, we contribute on the followings:
• We have performed a complete survey on the current wa-
termarking technologies.
• We propose a scene-based watermarking scheme. The scheme
is robust against frame averaging, frame dropping, frame
swapping, statistical analysis, etc
• We propose a visual-audio hybrid watermarking scheme.
The robustness of our scheme can be enhanced by including
an audio watermark. We embed error correcting codes of
a video watermark as an audio watermark and refine the
retrieved watermark during watermark detection.
• We propose a hybrid approach with different watermarking
schemes. We employ the hybrid scheme to embed differ-
ent parts of a watermark into different scenes. There are
different ways to embed the watermarks.
• We propose a GA-based watermarking scheme increase the
fidelity, i.e. the media quality index, of the watermarking
scheme. By employing GA, we can optimize the combina-
tion of the watermark and the scenes in the video.
CHAPTER 5. CONCLUSION 110
• Experiment has been done on these novel video watermark-
ing schemes to test and show its performance. The robust-
ness of our approach is demonstrated using the criteria of
the latest StirMark test.
2 End of chapter.
Bibliography
[1] A. Piva, F. Bartolini, and M. Barni, ”Managing copyright
in open networks,” IEEE Transactions on Internet Com-
puting, Vol. 6, Issue. 3, pp. 18-26, May-June 2002.
[2] C. Lu, H. Yuan, and M. Liao, ”Multipurpose Watermarking
for Image Authentication and Protection,” IEEE Transac-
tions on Image Processing, Vol. 10, Issue. 10, pp. 1579-1592,
Oct. 2001.
[3] C. Lu, S. Huang, C. Sze, and H. Y. M. Liao, ”Cocktail
watermarking for digital image protection,” IEEE Transac-
tions on Multimedia, Vol. 2, pp. 209-224, Dec. 2000.
[4] J. Lee and S. Jung, ”A survey of watermarking techniques
applied to multimedia,” Proceedings 2001 IEEE Interna-
tional Symposium on Industrial Electronics (ISIE2001),
Vol. 1, pp. 272-277, 2001.
[5] M. Barni, F. Bartolini, R. Caldelli, A. De Rosa, and
A. Piva, ”A Robust Watermarking Approach for Raw
Video,” Proceedings 10th International Packet Video Work-
shop PV2000, Cagliari, Italy, May 1-2 2000.
111
BIBLIOGRAPHY 112
[6] F. Petitcolas (Eds)”Information hiding techniques for
steganography and digital watermarking Stefan Katzen-
beisser,” Artech House Books, Dec. 1999, ISBN 1-58053-
035-4.
[7] A. Eskicioglu and E. Delp, ”An overview of multimedia con-
tent protection in consumer electronics devices,” Proceed-
ings Signal Processing Image Communication 16 (2001),
pp. 681-699, 2001.
[8] N. Fates and F. Petitcolas, ”Watermarking schemes evalu-
ation tool,” Proceedings of IEEE Multimedia Software En-
gineering, pp. 328-331, Taiwan, Dec. 2000.
[9] A. Herrigel and J. Ruanaidh, ”Secure Copyright Protection
Techniques for Digital Images,” Processing in Workshop on
Information Hiding, LNCS, Springer Verlag.
[10] M. Swanson, B. Zhu, and A. Tewfik, ”Multiresolution Video
Watermarking using Perceptual Models and Scene Seg-
mentation,” Proceedings International Conference on Im-
age Processing (ICIP ’97), 3-Volume Set-Volume 2, Wash-
ington, DC, Oct. 26-29, 1997.
[11] P.W. Chan and M. Lyu, ”A DWT-based Digital Video
Watermarking Scheme with Error Correcting Code,” Pro-
ceedings Fifth International Conference on Information and
Communications Security (ICICS2003), Lecture Notes in
Computer Science, Springer, Vol. 2836, pp. 202-213, Huhe-
haote City, Inner-Mongolia, China, Oct. 10-13, 2003.
BIBLIOGRAPHY 113
[12] P.W. Chan, M.R. Lyu, R. Chin, ”Copyright Protection on
the Web: A Hybrid Digital Video Watermarking Scheme,”
Poster Proceedings 13th International World Wide Web
Conference (WWW’2004), pp. 354-355, New York, May 17-
22, 2004.
[13] P.W. Chan, M.R. Lyu and R.T. Chin ”A Novel Scheme for
Hybrid Digital Video Watermarking: Approach, Evaluation
and Experimentation,” submitted to IEEE Transactions on
Circuits and Systems for Video Technology.
[14] F. Petitcolas, M. Frederic Raynal, J. Dittmann, C.
Fontaine, N. Fates, ”A public automated web-based evalu-
ation service for watermarking schemes: StirMark Bench-
mark,” In Ping Wah Wong and Edward J. Delp, editors,
proceedings of electronic imaging, security and watermark-
ing of multimedia contents III, Vol. 4314, San Jose, Cali-
fornia, U.S.A., Jan. 20V26, 2001. The Society for imaging
science and technology (I.S. and T.) and the international
Society for optical engineering (SPIE). ISSN 0277-786X,
http://www.petitcolas.net/fabien/watermarking/stirmark/.
[15] K. Su, D. Kundur, and D. Hatzinakos, ”Statistical Invisi-
bility for Collusion-resistant Digital Video Watermarking,”
IEEE Transactions on Multimedia 2004
[16] K. Su, D. Kundur, and D. Hatzinakos, ”Spatially Localized
Image-dependent Watermarking for Statistical Invisibility
and Collusion Resistance,” IEEE Transactions on Multi-
media 2004
BIBLIOGRAPHY 114
[17] K. Su, ”Digital Video Watermarking Principles for Resis-
tance to Collusion and Interpolation Attacks,” Master of
Applied Science thesis, University of Toronto, Sept. 2001.
[18] I. Cox, M. Miller, and J. Bloom, ”Digital watermarking,”
Morgan Kaufmann Publishers, Oct. 2001, ISBN 1-55860-
714-5.
[19] F. Litterio, http://world.std.com/ franl/crypto/.
[20] P. Johnston, http://pajhome.org.uk/crypt/rsa/intro.html.
[21] Watermarking World, http://www.watermarkingworld.org/.
[22] M. Kutter and F. Hartung, ”Introduction to Watermark-
ing Techniques,” Proceedings Information Techniques for
Steganography and Digital Watermarking, S.C. Katzen-
beisser et al., Eds. Northwood, MA: Artec House, pp. 97-
119, Dec. 1999.
[23] H. Inoue, A. Miyazaki, and T. Katsura ”An Image Wa-
termarking Method Based on the Wavelet Transform”,
Kyushu Multimedia System Research Laboratory.
[24] I. Cox, M. Miller, J. Linnartz, and T. Kalker, ”A Review of
Watermarking Principles and Practices” Proceedings Digital
Signal Processing for Multimedia Systems, K.K. Parhi, T.
Nishitani, eds., New York, New York, Marcel Dekker, Inc.,
pp. 461-482, 1999.
[25] F. Petitcolas, ”Watermarking Schemes Evaluation”, IEEE
Signal Processing Magazine, Vol. 17, pp. 58-64, Sept. 2000.
BIBLIOGRAPHY 115
[26] Y. Kim, K. Moon, and I. Oh, ”A text watermarking al-
gorithm based on word classification and inter-word space
statistics,” Proceedings Seventh International Conference
on Document Analysis and Recognition 2003, pp. 775 -779,
Aug. 3-6, 2003 .
[27] R. Wolfgang, C. Podilchuk, and E. Delp, ”Perceptual Wa-
termarks for Digital Images and Video,” Proceedings of the
SPIE/IS and T International Conference on Security and
Watermarking of Multimedia Contents, Vol. 3657, pp. 40-
51, San Jose, CA, Jan. 25 - 27, 1999.
[28] R. Wolfgang and E. Delp, ”A Watermarking Technique for
Digital Imagery: Further Studies,” Proceedings of the In-
ternational Conference on Imaging Science, Systems, and
Technology, pp. 279-287, Las Vegas, Jun. 30 - Jul. 3, 1997.
[29] D. Sachs, R. Anand, and K. Ramchandran, ”Wireless im-
age transmission using multiple-description based concate-
nated codes,” Proceedings Data Compression Conference
DCC 2000, pp. 569, 2000.
[30] A. Doufexi, A. Nix, D. Bull, ”Robust wireless image trans-
mission using jointly-optimized modulation and source cod-
ing,” Proceedings IEEE 51st Vehicular Technology Confer-
ence 2000, VTC 2000-Spring Vol. 3, pp. 2039-2043, Tokyo,
2000.
[31] M. Buckley, M. Ramos, S. Hemami, and S. Wicker,
”Perceptually-based robust image transmission over wire-
BIBLIOGRAPHY 116
less channels,” Proceedings 2000 International Conference
on Image Processing, Vol. 2, pp. 128-131, 2000.
[32] H. Bassali, J. Chhugani, S. Agarwal, A. Aggarwal, and P.
Dubey, ”Compression tolerant watermarking for image veri-
fication,” Proceedings 2000 International Conference on Im-
age Processing, Vol. 1, pp. 430-433, 2000.
[33] L. Boney, A. Tewfik, and K. Hamdy, ”Digital watermarks
for audio signals,” Proceedings Third IEEE International
Conference on Multimedia Computing and Systems, pp.
473-480, Jun. 17-23, 1996.
[34] C. Lu, M. Liao, and L. Chen, ”Multipurpose audio wa-
termarking,” Proceedings 15th International Conference on
Pattern Recognition 2000, Vol. 3, pp. 282-285, 2000.
[35] S. Craver, M. Wu, and B. Liu, ”What can we reasonably
expect from watermarks?,” Proceedings IEEE Workshop on
the Applications of Signal Processing to Audio and Acous-
tics 2001, pp. 223-226, 2001.
[36] C. Xu, J. Wu and Q. Sun ”Digital audio watermarking and
its application in multimedia database,” Proceedings the
Fifth International Symposium on Signal Processing and Its
Applications ISSPA’99, Vol. 1, pp. 91-94, 1999.
[37] S. Foo, T. Yeo, D. Huang, ”An adaptive audio watermark-
ing system,” Proceedings IEEE Region 10 International
Conference on Electrical and Electronic Technology 2001,
Vol. 2, pp. 509-513, TENCON, 2001.
BIBLIOGRAPHY 117
[38] K. Kaabneh and A. Youssef, ”Muteness-based audio wa-
termarking technique,” Proceedings International Confer-
ence on Distributed Computing Systems Workshop, 2001,
pp. 379-383, Apr. 2001.
[39] C. Xu, J. Wu, and Q. Sun, ”A robust digital audio
watermarking technique” Proceedings the Fifth Interna-
tional Symposium on Signal Processing and Its Applications
ISSPA ’99, Vol. 1, pp. 95-98, 1999.
[40] X. Wang, Y. Liu, and S. Blostein, ”Video image transmis-
sion via mobile satellite channels,” Proceedings IEEE Inter-
national Conference on Communications, ICC 95, Vol. 1,
pp. 352-356, Seattle, 18-22 Jun. 1995.
[41] N. Checcacci, M. Barni, F. Bartolini, and S. Basagni, ”Ro-
bust video watermarking for wireless multimedia commu-
nications ,” Proceedings IEEEE Wireless Communications
and Networking Confernce 2000, WCNC. 2000, Vol. 3, pp.
1530-1535, 2000.
[42] C. Lin and S. Chang, ”Issues and Solutions for Authenticat-
ing MPEG Video,” Proceedings IEEE International Confer-
ence on Acoustics, Speach and Signal Processing, pp. 54-65,
Mar. 15-19, 1999.
[43] C. Lu and M. Liao, ”Video object-based watermarking: a
rotation and flipping resilient scheme,” Proceedings 2001
International Conference on Image Processing, Vol. 2, pp.
483-486, 2001.
BIBLIOGRAPHY 118
[44] E. Lin, C. Podilchuk, T. Kalker, and E. Delp, ”Stream-
ing Video and Rate Scalable Compression: What Are the
Challenges for Watermarking?,” Proceedings the SPIE In-
ternational Conference on Security and Watermarking of
Multimedia Contents III, Vol. 4314, San Jose, CA, Jan. 22-
25, 2001.
[45] F. Hartung and B. Girod, ”Watermarking of Uncompressed
and Compressed Video”, IEEE Transaction Signal Process-
ing, Vol. 66, no. 3 (Special issue on Watermarking), pp.
283-301, May 1998.
[46] R. Wolfgang and E. Delp, ”Overview of Image Security
Techniques with Applications in Multimedia Systems,” Pro-
ceedings the SPIE Conference on Multimedia Networks: Se-
curity, Displays, Terminals, and Gatweys, Vol. 3228, pp.
297-308, Dallas, Texa, Nov. 2-5, 1997.
[47] R. Wolfgang, C. Podilchuk, and E. Delp, ”Perceptual Wa-
termarks for Digital Images and Video,” Proceedings the
SPIE/IST International Conference on Security and Wa-
termarking of Multimedia Contents, Vol. 3657, pp. 40-51,
San Jose, CA, Jan. 25-27, 1999.
[48] T. Furon, and P. Duhamel, ”Robustness of asymmetric wa-
termarking technique”, Proceedings International Confer-
ence onImage Processing, 2000, Vol.3, pp. 21-24.
[49] R. Lancini, F. Mapelli, and S. Tubaro, ”A robust video
watermarking technique in the spatial domain,” Processing
BIBLIOGRAPHY 119
and Multimedia Communications 4th EURASIP-IEEE Re-
gion 8 International Symposium on Video/Image VIProm-
Com, pp. 251-256, 2002.
[50] M. Ramkumar and A. Akansu, ”Robust Protocols for Prov-
ing Ownership of Image,” IEEE Transactions on Multime-
dia.
[51] P. Lee and M. Chen, ”Robust error concealment algorithm
for video decoder,” IEEE Transactions on Consumer Elec-
tronics, Vol. 45, Issue. 3, pp. 851 -859, Aug. 1999.
[52] D. He, Q. Sun and Q. Tian, ”A semi-fragile object based
video authentication system,” Proceedings of the 2003 In-
ternational Symposium on Circuits and Systems ISCAS ’03,
Vol.3, pp. 814-817, May 25-28 2003.
[53] J. Fridrich, M. Goljan and A. Baldoza, ”New fragile authen-
tication watermark for images,” Proceedings. 2000 Interna-
tional Conference on Image Processing, Vol. 1, pp. 446-449,
Sept. 10-13, 2000.
[54] M. Swanson, B. Zhu, and A. Tewfik, ”Transparent robust
image watermarking,” Proceedings International Confer-
ence on Image Processing, 1996, Vol. 3, pp. 211-214, Sept.
16-19, 1996.
[55] Y. Steinberg, ”Watermarking identification for private and
public users: the broadcast channel approach,” Proceedings
the 2002 IEEE Information Theory Workshop, pp. 5-7, Oct.
20-25, 2002.
BIBLIOGRAPHY 120
[56] P. Wong and O. An, ”A novel semi-private watermarking
technique,” Proceedings IEEE International Symposium on
Circuits and Systems 2002 (ISCAS 2002), Vol. 2, pp. 448-
451, May 26-29, 2002.
[57] T. P.-C. Chen and T. Chen, ”A Framework for Optimal
Public Watermark Detection”, Carnegie Mellon University
Technical Report: AMP01-03.
[58] R. Wolfgang and E. Delp. ”A watermark for digital images,”
Proceedings International Conference on Images Process-
ing, pp. 219-222, Lausanne, Switzerland, Sept. 1996.
[59] H. Kii, J. Onishi, and S. Ozawa, ”The Digital Watermark-
ing Method by Using both Patchwork and DCT,” IEEE In-
ternational Conference on Multimedia Computing and Sys-
tems, Vol. 1, Florence, Italy, Jun. 07-11, 1999.
[60] N. Merhav, ”On random coding error exponents of water-
marking systems,” IEEE Transactions on Information The-
ory, Vol. 46 Issue. 2, pp. 420-430, Mar. 2000.
[61] M. Barni, F. Bartolini, A. Rosa and A. Piva, ”Capacity of
full frame DCT image watermarks,” IEEE Transactions on
Image Processing, Vol. 9 Issue. 8, pp. 1450-1455, Aug. 2000.
[62] A. Bors and I. Pitas, ”Image watermarking using DCT do-
main constraints,” Proceedings International Conference on
Image Processing 1996, Vol. 3, pp. 231-234, Sept. 16-19,
1996.
BIBLIOGRAPHY 121
[63] I. Hong, I. Kim and S. Han, ”A blind watermarking tech-
nique using wavelet transform,” Proceedings IEEE Interna-
tional Symposium on Industrial Electronics ISIE 2001, Vol.
3, pp. 1946-1950, 2001.
[64] X. Niu and S. Sun, ”A New Wavelet-Based Digital Water-
marking for Video,” Proceedings 9th IEEE Digital Signal
Processing Workshop, Texas, USA, Oct. 2000.
[65] K. Leung and B. Zeng, ”Wavelet-based digital watermark-
ing with halftoning technique,” Proceedings The 2001 IEEE
International Symposium on Circuits and Systems, 2001.
ISCAS 2001, Vol. 5, pp. 235-238, 2001.
[66] Y. Wang, J. Doherty, and R. Dyck, ”A wavelet-based wa-
termarking algorithm for ownership verification of digital
image,” IEEE Transactions on Image Processing, Vol. 11,
No. 2, Feb. 2002.
[67] P. Premaratne and C. Ko, ”A novel watermark embedding
and detection scheme for images in DFT domain,” Seventh
International Conference on Image Processing and Its Ap-
plications 1999, Vol. 2, pp. 780-783, Jul. 13-15, 1999.
[68] Q. Cheng and T. Huang, ”Optimum detection and decod-
ing of multiplicative watermarks in DFT domain,” Proceed-
ings IEEE International Conference on Acoustics, Speech,
and Signal Processing (ICASSP ’02) 2002, Vol. 4, pp. 3477-
3480, May 13-17, 2002.
BIBLIOGRAPHY 122
[69] J. Linnartz and J. Talstra, ”MPEG PTY-marks: Cheap de-
tection of embedded copyright data in DVD video,” Pro-
ceedings 5th European Symposium on Research in Computer
Security, pp. 221-240, 1998.
[70] L. Qiao and K. Nahrstedt, ”Watermarking method for
mpeg encoded video: Towards resolving rightful owner-
ship,” IEEE Multimedia Computing and Systems, Jun.
1998.
[71] T. Chung, M. Hong, Y oung-Nam Oh, Dong-Ho Shin, and
Sang-Hui Park, ”Digital watermarking for copyright pro-
tection of MPEG2 compressed video,” IEEE Transactions
on Consumer Electronics, Vol. 44, Issue. 3, pp. 895-901,
August 1998.
[72] S. Arena and M. Caramma, ”Digital watermarking applied
to MPEG2 coded video sequence exploiting space and fre-
quency masking,” Proceedings International Conference on
Image Processing (ICIP-2000), Vol. 3, pp. 438-441, Van-
couver, Canada, 2000
[73] A. Piva, R. Caldelli and A. De Rosa, ”A DWT-based object
watermarking system for MPEG-4 video streams,” Proceed-
ings 2000 International Conference on Image Processing,
Vol. 3, pp. 5-8, 2000.
[74] M. Suhail and M. Obaidat, ”On the digital watermarking in
JPEG 2000,” Proceedings The 8th IEEE International Con-
ference on Electronics, Circuits and Systems, 2001 (ICECS
2001), Vol. 2, pp. 871-874, Sept. 2-5, 2001.
BIBLIOGRAPHY 123
[75] J. Dittmann, A. Mukherje, and M. Steinebach, ”Media-
independent watermarking classification and the need for
combining digital video and audio watermarking for media
authentication,” Proceedings International Conference on
Information Technology: Coding and Computing 2000 pp.
62-67, 2000.
[76] J. Dittmann and M. Steinebach, ”Joint watermarking of
audio-visual data,” Processing 2001 IEEE Fourth Workshop
on Multimedia Signal, pp. 601-606, 2001.
[77] J. Fridrich, ”A hybrid watermark for tamper detection in
digital images,” Proceedings the Fifth International Sympo-
sium onSignal Processing and Its Applications ISSPA ’99,
1999, Vol. 1, pp. 301-304, 1999.
[78] X. Kang, J. Huang, Y. Shi, and Y. Lin, ”A DWT-DFT com-
posite watermarking scheme robust to both affine transform
and JPEG compression,” IEEE Transactions on Circuits
and Systems for Video Technology, Vol. 13, Issue: 8, pp.
776-786, Aug. 2003.
[79] F. Deguillaume, G. Csurka, J. Ruanaidh, and T. Pun, ”Ro-
bust 3D DFT video watermarking,” Proceedings Electronic
Imaging’ 99: Security and Watermarking of Multimedia
Contents, Vol. 3657, San Jose, CA, Jan. 1999.
[80] M. Ejima and A. Miyazaki, ”A wavelet-based watermark-
ing for digital images and video,” Proceedings International
Conference on Image Processing (ICIP-2000), Vol. 3, pp.
678-681, Vancouver, Canada, 2001.
BIBLIOGRAPHY 124
[81] F. Hartung and B. Girod, ”Watermarking of uncompressed
and compressed video,” Proceedings Signal Processing, Vol.
66, pp. 283-301, 1998.
[82] N. Johnson and S. Katezenbeisser, ”A Survey of Stegano-
graphic Techniques in Information Techniques for Steganog-
raphy and Digital Watermarking,” S.C. Katzenbeisser et
al., Eds. Northwood, MA: Artec House, pp. 43-75, Dec.
1999.
[83] G. Langelaar, I. Setyawan, R. Lagendijk, ”Watermarking
Digital Image and Video Data,” IEEE Signal Processing
Magazine, Vol 17, pp. 20-43, Sept. 2000.
[84] X. Niu and S. Sun, ”A New Wavelet-Based Digital Water-
marking for Video”, Proceedings 9th IEEE Digital Signal
Processing Workshop, Texas, USA, Oct. 2000.
[85] C. Serdean, M. Ambroze, M. Tomlinson, and G. Wade,
”Combating geometrical attacks in a dwt based blind video
watermarking system”, Proceedings Video/Image Process-
ing and Multimedia Communications 4th EURASIP-IEEE
Region 8 International Symposium on VIPromCom, pp.
263-266, 2002.
[86] D. Swanson, B. Zhu, and A. Tewfik, ”Multiresolution Video
Watermarking using Perceptual Models and Scene Seg-
mentation,” Proceedings International Conference on Im-
age Processing (ICIP ’97), Vol. 2, Washington, DC, Oct.
26-29, 1997.
BIBLIOGRAPHY 125
[87] M. Barni, F. Bartolini, R. Caldelli, A. De Rosa, and
A. Piva, ”A Robust Watermarking Approach for Raw
Video”, Proceedings 10th International Packet Video Work-
shop PV2000, Cagliari, Italy, May 1-2, 2000.
[88] M. Holliman, N. Memon, B. Yeo, and M. Yeung, ”Adaptive
public watermarking of DCT-based compressed images,”
Proceedings SPIE, Vol. 3312, pp. 284-295, 1997.
[89] B. Vassaux, P. Nguyen, S. Baudry, P. Bas, and J. Chas-
sery, ”Scrambling technique for video object watermarking
resisting to mpeg-4,” Proceedings Video/Image Processing
and Multimedia Communications 4th EURASIP-IEEE Re-
gion 8 International Symposium on VIPromCom, pp. 239-
244, 2002.
[90] M. Swanson, B. Zhu, B. Chau, and A. Tewfik, ”Object-
Based Transparent Video Watermarking,” Proceedings
IEEE Signal Processing Society 1997 Workshop on Multi-
media Signal Processing, Princeton, New Jersey, USA, Jun.
23-25, 1997.
[91] C. Lu and M. Liao, ”Video object-based watermarking: a
rotation and flipping resilient scheme,” Proceedings 2001
International Conference on Image Processing, Vol. 2, pp.
483-486, 2001.
[92] B. Mobasseri, ”Direct sequence watermarking of digital
video using m-frames,” Proceedings International Confer-
ence on Image Processing (ICIP-98), Vol. 3, pp. 399-403,
Chicago, Illinois, Oct. 4-7, 1998.
BIBLIOGRAPHY 126
[93] N. Memon, ”Analysis of LSB based image steganography
techniques Chandramouli,” Proceedings 2001 International
Conference on Image, Vol. 3. pp. 1019-1022, Oct. 7-10,
2001.
[94] B. Mobasseri, ”Direct sequence watermarking of digi-
tal video using m-frames,” Proceedings 1998 International
Conference on Image Processing ICIP 98, Vol. 2, pp. 399-
403, Oct. 4-7, 1998.
[95] B. Mobasseri, ”Direct sequence watermarking of digital
video using m-frames,” Proceedings International Confer-
ence on Image Processing (ICIP-98), Vol. 3, pp. 399-403,
Chicago, Illinois, Oct. 4-7, 1998.
[96] S. Pereira and T. Pun, ”Robust template matching for affine
resistant image watermarks, ” IEEE Transactions on Image
Processing, Vol.9 Issue. 6, pp. 1123-1129, Jun. 2000.
[97] I. Hong, I. Kim, and S. Han, ”A blind watermarking tech-
nique using wavelet transform” Proceedings IEEE Interna-
tional Symposium on Industrial Electronics ISIE 2001, Vol.
3, pp. 1946-1950, 2001.
[98] F. Duan, I. King, L. Xu, and L. Chan, ”Intra-block algo-
rithm for digital watermarking”, Proceedings IEEE 14th In-
ternational Conference on Pattern Recognition (ICPR’98),
Vol. 2, pp. 1589-1591, Aug. 17-20, 1998.
[99] M. George, J. Chouinard and N. Georganas, ”Digital water-
marking of images and video using direct sequence spread
BIBLIOGRAPHY 127
spectrum techniques,” Proceedings 1999 IEEE Canadian
Conference on Electrical and Computer Engineering, Vol.
1, pp. 116 121, May 9-12, 1999.
[100] F. Petitcolas and R. Anderson, ”Evaluation of Copyright
Marking Systems,” Proceedings IEEE multimedia Systems
(ICMCS’99), Florence, Italy, Jun. 7-11, 1999.
[101] M. Kutter and F. Petitcolas, ”A Fair Benchmark for Image
Watermarking Systems,” Proceedings of Electronic Imaging
’99, Security and Watermarking of Multimedia Contents,
Vol. 3657, pp. 226-239.
[102] N. Checcacci, M. Barni, F. Bartolini, and S. Basagni, ”Ro-
bust video watermarking for wireless multimedia commu-
nications,” Proceedings 2000 IEEE Wireless Communica-
tions and Networking Conference (WCNC 2000), Vol. 3,
pp. 1530-1535, 2000.
[103] L. Zhang, Z. Cao, and C. Gao, ”Application of RS-coded
MPSK modulation scenarios to compressed image com-
munication in mobile fading channel,” Proceedings 2000
52nd IEEE Vehicular Technology Conference, VTS-Fall
VTC.2000, Vol. 3, pp. 1198-1203, 2000.
[104] P. Dang and P. Chau, ”Image encryption for secure Inter-
net multimedia applications,” IEEE Transactions on Con-
sumer Electronics, Vol. 46, Issue. 3, pp. 395-403, Aug. 2000.
[105] F. Duan, I. King, L. Xu, and L. Chan, ”Intra-block max-
min algorithm for embedding robust digital watermark into
BIBLIOGRAPHY 128
images,” In Horace H.S. Ip and Arnold W.M. Smeulders,
editors, Proceedings of the IAPR International Workshop
on Multimedia Information Analysis and Retrieval, MI-
NAR’98, Lecture Notes in Computer Science, Vol. 1464,
pp. 255-264, Berlin Heidelberg, Germany, 1998. Springer-
Verlag.
[106] J. Cox, J. Kilian, F. Leighton, and T. Shamoon, ”Se-
cure spread spectrum watermarking for multimedia,” IEEE
Transactions on Image Processing, Vol. 612, pp. 1973-1987,
Dec. 1997.
[107] A. Moulin, ”The role of information theory in watermark-
ing and its application to image watermarking,” Signal Pro-
cessing 2001, Vol. 81, pp. 1121-1139, 2001.
[108] M. Ramkumar and A. Akansu, ”A Capacity for Data Hidig
in Internet Multimedia,” Symposium on Content Security
and Data Hiding in Digital Media, NJIT, Jersey City, May
1999.
[109] S. Servetto, C. Podilchu and K. Ramchadra, ”Capacity
Issues in Digital Image Waterarmking,” IEE International
Conference on Image Processing, Chicago, Oct. 1998.
[110] T. Cover and J. Thomas, ”Elements of Information Teory,”
John Wiley and Sones, Lnc., 1991.
[111] A. Ambroze, G. Wade, C. Serdean, M. Tomlinson, J.
Stander, and M. Borda, ”Turbo code protection of video
BIBLIOGRAPHY 129
watermark channel,” Proceedings IEEE Vision, Image and
Signal Processing, Vol. 148, Issue. 1, pp. 54-58, Feb. 2001.
[112] D. Kirovski and H. Malvar, ”Robust covert communication
over a public audio channel using spread spectrum,” Pro-
ceedings First Information Hiding Workshop, Pittsburgh,
PA, USA, 2001.
[113] J. Princen, A. Johnson, and A. Bradley, ”Subband/ trans-
form coding using filter bank designs based on time domain
aliasing cancellation,” Proceedings IEEE ICASSP, pp. 2161-
2164, Dallas, TX, Apr. 1987.
[114] H. Malvar, ”A Modulated Complex Lapped Transform and
its Applications to Audio Processing,” Proceedings IEEE
ICASSP, Mar. 1999.
[115] C. Huang and J. Wu, ”A watermark optimization tech-
nique based on genetic algorithms,” SPIE Electronic Imag-
ing 2000, San Jose, Jan. 2000.
[116] C. Huang and J. Wu, ”Using Genetic Algorithms as Water-
marking Performance Optimizers,” IEEE Signal Processing
Letters.
[117] VirtualDub is distributed under the GNU General Public
License, written by Avery Lee, virtualdub.org / phaeron
(at), http://www.virtualdub.org/.
[118] C. Hzu and J. Wu, ”Digital watermarking for video,” Pro-
ceedings 1997 13th International Conference on Digital Sig-
nal Processing, DSP 97, Vol. 1, pp. 217-220, Jul, 2-4, 1997.
BIBLIOGRAPHY 130
[119] X. Niu and S. Sun, ”A New Wavelet-Based Digital Wa-
termarking for Video,” 9th IEEE Digital Signal Processing
Workshop, Texas, USA, Oct. 2000.
[120] D. Kirovski, and H. Malvar, ”Robust spread-spectrum au-
dio watermarking,” Proceedings IEEE International Con-
ference on Acoustics, Speech, and Signal Processing, 2001,
Vol. 3, pp. 1345-1348, 2001.
[121] K. Su, D. Kundur and D. Hatzinakos, ”A Novel Ap-
proach to Collusion-Resistant Video Watermarking,” Pro-
ceedings Security and Watermarking of Multimedia Con-
tents IV SPIE, E. J. Delp and P. W. Wong, eds., Vol. 4675,
pp. 12, San Jose, California, Jan. 2002.
[122] K. Su, D. Kundur and D. Hatzinakos, ”A Content-
Dependent Spatially Localized Video Watermarked for Re-
sistance to Collusion and Interpolation Attacks,” Proceed-
ings IEEE International Conference on Image Processing,
Oct. 2001.
[123] Y. Wang, J. Doherty, and R. Dyck, ”A wavelet-based wa-
termarking algorithm for ownership verification of digital
image,” IEEE Transactions on Image Processing, Vol. 11,
No. 2, Feb. 2002.
[124] R. Wolfgang, C. Podilchuk, and E. Delp, ”The effect of
matching watermark and compression transforms in com-
pressed color images,” Proceedings International Confer-
ence on Image Processing (ICIP-98), Vol. 1, pp. 440-444,
Chicago (IL, USA), Oct. 1998.
BIBLIOGRAPHY 131
[125] The software for this work used the GAlib ge-
netic algorithm package, written by Matthew
Wall at the Massachusetts Institute of Technology,
http://lancet.mit.edu/ga/.
[126] E. Gilmore, M. Chouikha, ”Dictionary Approaches To Im-
age Compression and Reconstruction,” Proceedings of the
IASTED International Conference on Signal and Image,
Las Vegas, USA, Sept. 12-16, 1998.

Abstract of thesis entitled: Digital Video Watermarking Techniques for Secure Multimedia Creation and Delivery Submitted by CHAN Pik-Wah for the degree of Master of Philosophy at The Chinese University of Hong Kong in July 2004

There is an explosion of data exchange on the Internet and the extensive use of digital media. Consequently, digital data owners can quickly and massively transfer multimedia documents through the Internet. It has aroused intense interest in multimedia security and multimedia copyright protection. In this paper, a comprehensive approach for protecting and managing video copyrights with watermarking techniques is introduced. We propose a novel hybrid digital video watermarking scheme based on the scene change analysis, error correction code and genetic algorithm. Our video watermarking algorithm is robust against the attacks of frame dropping, averaging and statistical analysis, which were not solved effectively in the past. It started with a complete survey about current watermarking technologies. We have discovered that none of the existing schemes was capable of resisting all attacks. Accordingly, we came up with the idea of embedding a single watermark of different parts into different scenes of a video. Then we analyze the strength of difi

ferent watermarking schemes. A hybrid approach is applied to form a super watermarking scheme which can resist most of the attacks. In order to reinforce the robustness of the scheme, the watermark is refined by an error correcting code, while the correcting code is embedded as a watermark in the audio channel. Furthermore, the fidelity of the scheme is enhanced by applying genetic algorithm. It optimizes the quality of the watermarked video. Also, our scheme allows blind retrieval of the embedded watermark, which does not need the original video and the watermark is perceptually invisible. The effectiveness of the scheme is verified through a series of experiments, in which a number of standard image processing attacks are conducted, and the robustness of our approach is demonstrated by using the criteria of the latest StirMark test.

ii

iii .

CUHK. as well as his inspiring advice are extremely essential and valuable in my research papers (conference papers published in ICICS’2003 and WWW’2004 and journal paper submitted to IEEE transaction TCSVT) and my thesis. Chan. I will not be able to strengthen and improve my research project and papers. Sam Sze who have given me valuable suggestions. C. Y. K. I would also like to show my gratitude to the Department of Computer Science and Engineering. Y. His numerous support and encouragement. Tien-Tsin Wong and Prof. Hoi. encouragement and supports. Special thanks should be given to Mr. H. for his generous guidance and patience given to me in the past two years. Prof. Huang. H. Lyu. iv . for the provision of the best equipment and pleasant office environment required for high quality research. K. Without their effort. And I would like to give my thanks to my fellow colleagues.Acknowledgement I would like to take this opportunity to express my gratitude to my supervisor Prof. Edward Yau and Mr. Irwin King. I am also grateful for the time and valuable suggestion that Prof. Z. Michael R. Roland Chin have given in marking my term paper.

Finally. Ng. Yu. K. Leung. Lam. S. C. C. IP. Y. H. J. v .Lee. Y. W. C. Zheng. N. and a joyful and wonderful university life. Y. Y. C. W. C. Hung. Lau and T. H. H. They have given me support. W. Chan. I am deeply indebted to my family for their unconditional love and support over the years. Law. Wong.

This work is dedicated to my family for the support and patience vi .

. .1 Background . . . . i iv 1 1 3 4 6 7 . . . . . . Essential Ingredients for Video Watermarking 2.3 Use of Keys . Digital Watermarking . . . . . . . . . . . . . . . . . . . . .Contents Abstract Acknowledgement 1 Introduction 1. .4. . . 2. . . .4. . . .4 Research Objective . . . . . . . . .4. . . . . . . . .6 Statistical Imperceptibility .2 2. . .3 1. . . . . . . . . . 8 11 14 16 16 17 19 20 20 21 2 Literature Review 2. . . . . 2. . vii . . . . . . .2 1. . . . . .1 2. 2. Contributions .2 Robustness . 2. . . . . . .4 Blind Detection . . . . . . . . . . . . . The Structure of this Thesis .4. . . . . . . . . . . . . . . . . 1. Cryptography . . .5 Capacity and Speed . . . .1 Fidelity . . . . . . . .3 2. . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . .4. . . . . . . . . . . . . . .4. . . . . . . . . . .4 Security in Multimedia Communications . . . . . . . .

3. .4 Watermark Detection . . .1 Watermarking Scheme . . . . . . . . . .6 Video Watermarking . . . . . . . .5 Review on Video Watermarking Techniques . . . . . . . . . 21 2. . .1 A Scene-based Video Watermarking Scheme . . . . . .1 2. . . . . . . .5. . . . . . . . . . 21 2. . . . .1 Watermark Preprocess . . . .5. . . . . . . . .4. .1. . . . . . . . . . .1. . 3. . . . .5. .4.3 3. . . . 3. . 3. viii . 30 Watermarks Based on MPEG Coding Structures . . .2. . . . . . .4 2. . 44 3. .2 2. . .1. .4. . . . . . . . . . . . . 35 Comparison between Different Watermarking Schemes 38 42 3 Novel Watermarking Schemes 3. . . . . . 46 48 50 52 52 56 60 61 3.7 Low Error Probability . A Genetic Algorithm-based Video Watermarking Scheme . 42 3.3 Chromosome Encoding . .8 Real-time Detector Complexity . . . . .2 Hybrid Approach with Different Watermarking Schemes . . . . . . 22 2. .2 Capacity . . . . .1 Performance . .3 Watermark Embedding 3. . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . .4. . .2. . 3. . . . . . . . . . . . 25 Spatial Domain Watermarks . . . . . . . . . . . . 26 Frequency Domain Watermarks .2. 3.2 3. . .4 . .1 Visual-audio Hybrid Watermarking . .2 Video Preprocess . . .4. . . . . . . . . . . . . . . . . . .3. . Theoretical Analysis . . . .3 2. . . .3.1. A Hybrid Watermarking Scheme . . . . 69 . . . . . . . . . . . . . 73 75 76 79 3.5. .2 Problem Modelling .

. . .2. . . . . . . 102 Other Features of the Scheme . . .1.4 Genetic Operators . . . . . . . . .0 . . . . .2 Parameter(s) Setting .2 4. . . . . 87 Experiment with Frame Averaging and Statistical Analysis . . . . .1. 92 4. . .3 4. 90 Test of Robustness with StirMark 4. . .5 Overall Comparison . . . . . . 89 Experiment with Lossy Compression . . . .4 4. . . . . . . . . .2. . .2 Test on Robustness . . . . . .1. . . . . . . . . .4. . . . 80 85 4 Experimental Results 4. . . 85 Experiment with Frame Dropping . . . .1. . . . .1 4. 101 Evaluate with PSNR . .4 4.3 4. . 101 4. . . . . . . . .1. . . . . . . .3.3 Evaluate with MAD . .1 4. . . . . .1 4. . 98 Test on Fidelity . . . . . . . . . . . . . . . . . . . .2. 100 4. . . 105 Conclusion . . . . . . . . . . . . . . . . . . . 106 108 110 5 Conclusion Bibliography ix . .

. . . . .7 3. 46 Scene change detection . .3 3. 49 (a) Original frame (b) Watermarked frame (c) Extracted watermark corresponding to Figure 3. . .3 2. 64 3. 61 Overview of visual-audio hybrid watermarking scheme 62 (a) Original video watermark (b) Visualization of averaging (c) Audio watermark (average of a) . . . . . . . . . . . . . . . . . . . . . .8 3. . . . . . . . . . . . . . . . . . . 48 Embedding watermarks in a frame . .List of Figures 2. . . . . .9 x . . . . 45 (a) Original watermark (b-i) Preprocessed watermark m0 − m7 (j) Encrypted watermark m0 .3(g) (d) Recovered watermark. . . . . . . . . . .5 3. . . . . . . . .2 2.2 3.6 Overview of the watermarking process . . . . . . . . .1 3.5 2 Scale 2-Dimensional Discrete Wavelet Transform 32 3. . . . .4 Symmetric Cryptosystem . . . . . 26 2. . . 13 Watermarking Embedding and Detection Scenario 15 Classification map of existing digital video watermark techniques . . 12 Asymmetric Cryptosystems . 51 Possible improvement for scene based watermarking scheme . . . . .4 3. . . . . 43 Preprocessing the watermark . . . .1 2. . . . . . . . . . . .

. . . . . . . robustness and fidelity of the watermarking scheme . . . . . . . . . . . . . A 24-bit chromosome represents the sequence of the scenes to embed . . . . . . . . . 66 3. . . . . . . . . . . . . . 72 3.14 Overview of detection of the watermark . . . . . . . Comparison between watermarked video with and without GA optimization a) Original video frame (b) Video frame watermarked with scene-based scheme (c) Video frame watermarked with GAbased scheme . . . . . . . . . . . . . . . . . . . 84 . . . . .21 3. . . . . . . . . . .23 process . .11 One of the (a) original video frame and (b) watermarked video frame . . 69 3.22 3. . 73 3. . . . . . .10 Audio watermark embedding with MCLT . . . . . . 74 3. . . . .16 Hybrid approach with different scheme for different part of frame . .15 Hybrid approach with different scheme for different scene . . . . 82 . . . . .3. . 71 3. . . .20 3. . . . . . 80 . . . . . . . . .17 The graph of three mutually orthogonal axes representing the capacity. . . .19 A illustrative diagram for GA-based optimization 3. . . . . . . . . . . . . . . .13 Block of samples of watermarked wave content . . . . . . . . . . . . . The GA-based watermarking algorithm . . . . . . . . . 78 .12 Block of samples of the original wave content . . . . xi . 67 3. 77 . . . . . . The GA-based optimization process for part of watermark . .18 The graph of two mutually orthogonal axes representing the robustness and fidelity of the watermarking scheme . . . 68 3. . . . . . 68 3.

4. . . . . . 96 NC values under different noise added to the watermarked video . . . . . . . .7 4. . . . . . . .8 4. . . . . 91 NC values under cropping . . .3 4. . . .6 4. 95 NC values under different rescaling factor . . . . .2 4. .11 A conceptual illustration on the performance of the proposed scheme . .5 4. . . . . . . . 89 NC values under statistical averaging . . . . 97 PSNR of the video under different GA generations 103 4. . . . . . . .10 MAD of the video under different GA generations 104 4. . . . . . . . . 87 Scenario of statistical averaging attack . . . . . . . 94 NC values under PSNR . . . .4 4. . . . . . . . . .1 4. . . . . . 106 xii . . . .9 NC values under frame dropping . . . 90 NC values under lossy compression . . . .

. . .4 Basic Robustness Requirements .3 4. . . . . 105 Parameters Setting for GA-based experiment . . . . . . .5 MAD comparison between different watermarking schemes . . . . . . . .2 4. 19 Classification of watermarking according to several viewpoints . . . . . .2 2. . . . . . . 4. . 102 .List of Tables 2. . xiii . . . . . . . . . . . . . 103 . . .1 2. . . . . . . . . . .1 4. . . . . . . . 24 Comparison between different watermarking schemes 39 Robustness comparison between different watermarking schemes . . . . . . . . 101 . . . . . The computation time of the GA-based scheme PSNR comparison between different watermarking schemes . 98 . . . . . . . .3 4. .

there is an increase in concern over copyright protection of digital contents [1. we focus on engaging the digital watermarking techniques to protect digital multimedia intellectual copyright. do not protect against unauthorized copying after the media have been successfully transmitted and decrypted. 5. encryption and control access techniques were employed to protect the ownership of media. 1.Chapter 1 Introduction With the rapid growth of the Internet and multimedia systems in distributed environments. watermark techniques are utilized to maintain the copyright [4. These techniques. Recently. however. Therefore. however. 7]. 2. cryptography ensures confiden1 . 3. In this paper. it is easier for digital data owners to transfer multimedia documents across the Internet. Traditionally. 4]. 6.1 Background Multimedia and network security issues are classically handled through cryptography. and propose a new algorithm particularly for video watermarking purpose.

into a multimedia document so that a given piece of copyright information is permanently tied to the data. Numerous inventive watermarking approaches have been proposed in these few years and most of them focus on digital image watermarking. Due to large amounts of data and inherent redundancy between frames. In recent years. There is only a very few number of industrial associations have published the requirements for testing watermarking algorithms [8]. However. image watermarking technique becomes mature. It does not protect against unauthorized copying after the message has been successfully transmitted. Video watermarking introduces some issues which is not present in image watermarking. thus researcher starts to explore a more challenging research topic – digital video watermarking. watermark. authenticity. Digital watermarking remains a largely untested field. or simply inform users about the rights-holder or the permitted use of the data [6]. Watermarking is a concept of embedding a special pattern. Most of the proposed video watermarking schemes are based on the techniques of image watermarking and directly applied to raw video or compressed video.CHAPTER 1. INTRODUCTION 2 tiality. video signals are highly suscep- . and integrity only when a message is transmitted through a public channel such as an open network. identify a misappropriating person. Digital watermarking is an effective way to protect copyright of multimedia data even after its transmission. This information can later prove the ownership. current image watermarking schemes are not capable of adequately protecting video data [9]. trace the marked document’s dissemination through the network.

1. 12]. Various watermarking scheme schemes are compared and evaluated. Applying a fixed image watermark to each frame in the video leads to problems of maintaining statistical and perceptual invisibility. as the watermark is fixed while the frame changes. Applying independent watermarks to each frame also presents a problem. etc. a survey and investigation has been done on multimedia security issues and multimedia watermarking scheme. a new approach and procedures for multimedia security based on watermarking are proposed [11. video watermarking schemes must not use the original video during watermark detection as the video usually is in very large size and it is inconvenient to store it twice. statistical analysis. blind and not removable video watermarking scheme. In this scheme. We propose a new video watermarking scheme to overcome these problems.2 Research Objective As video copyright protection is strongly concerned. Base on these. such an approach is necessarily video independent.CHAPTER 1. A Hybrid scene-based video watermarking scheme with error correcting code and genetic algorithm is proposed. In addition. Furthermore. including frame averaging. frame dropping. frame swapping. invisible. Motionless regions may be statistically compared or averaged to remove independent watermarks [10]. INTRODUCTION 3 tible to pirate attacks. In order to design a robust. Regions in each video frame with little or no motion remain the same frame after frame. a robust video watermarking scheme is necessary. the watermark is decomposed into different parts and .

we propose a hybrid watermarking scheme based on scene change analyze [11]. 13]. With this finding. error correcting code is extracted from the video channel and embedded into the audio channel. interpolation and lossy compression. averaging. INTRODUCTION 4 embedded in the frames of different scenes in the video with hybrid approaches. At the same time. . which provides extra information for recovery of extracted watermark. The performance of the watermarking scheme is enhanced by applying the genetic algorithm (GA) optimization.3 Contributions Our research work has the following contributions: • We have performed a complete survey on the current watermarking technologies. This research can be continuous by applying this new developed scheme to specific environment or application and examine its usefulness. 1. swapping. error correction codes and genetic algorithm [12. As identical watermark is used within each motionless scene and independent watermarks are used for successive different scenes. Moreover. It is noticed that none of the current watermarking schemes can resist all attacks. It can be associated with different applications to achieve a sophisticated system. Video watermarking is not a stand alone technology.CHAPTER 1. the proposed method is robust against the attack of frame dropping. the scheme allows blind retrieval of embedded watermark which does not need the original video.

the watermark strength of the scheme. Consequently. we propose several hybrid approaches. By employing GA. we employ the hybrid scheme to embed different parts of a watermark into different scenes. • To increase the fidelity. As videos consist of both video and audio channels. of the watermarking scheme. the proposed scheme is capable of resisting most of the common attacks [12]. making the scheme robust against frame averaging and statistical analysis [11]. as the same part of the watermark is embedded into the frames of a scene. the robustness of our scheme can be enhanced by including an audio watermark. This scheme is innovative in attacking the problems that are not solved effectively in the past. The first one is visual-audio hybrid watermarking scheme. • The second approach is another hybrid with different watermarking schemes.CHAPTER 1. the media quality. The scheme is robust against frame dropping. we propose a GA-based watermarking scheme. Thus. the quality of the watermarked video is enhanced. As no existing scheme is resistant against all attacks. which can refine the retrieved watermark during watermark detection [12]. we embed error correcting codes of a video watermark as an audio watermark. • To increase the robustness. different parts of the watermark are used. INTRODUCTION 5 • We proposed a new scheme which applies scene change detections and scrambled watermarks in a video. . For different scenes.

The robustness of our approach is demonstrated using the criteria of the latest StirMark test [14]. Our approach cultivates an innovative idea in embedding different parts of a watermark according to scene changes. • We compare the proposed scheme with the existing scheme in different aspects and discuss the advantages and the disadvantages of our scheme. The next chapter introduces the issues related to multimedia security and different multimedia watermarking techniques. .4 The Structure of this Thesis This paper is organized as 5 chapters. applying a hybrid approach to the proposed scheme and employing GA to optimize the fidelity of the scheme. and a survey on current watermark techniques and video watermarking scheme are provided. This approach is never explored in the literature. Novel video watermarking scheme is described in chapter 3 and the experimental results in Chapter 4 are followed by. INTRODUCTION 6 • Experiments have been done on these novel video watermarking schemes to test and show its performance. and its advantages are clear and significant.CHAPTER 1. embedding its error correcting codes as an audio watermark. Finally. 1. 2 End of chapter. a conclusion would be given in chapter 5. The effectiveness of this scheme is verified through a number of experiments.

The purpose of a digital watermark is to embed auxiliary information into a digital signal by making small changes that are not perceptible to its intended recipient. the field of digital watermarking grows extremely fast in these few years [15]. watermarks are powerful tools that will play a role in solving the growing digital property identification problem [16]. this drives to urgent need to resolve the security and copyright protection issues. Then the principle and the techniques of digital watermarking are dis7 . For instance. We first have a look of two popular security tools. Because the embedded signals enable invisible tags to be attached to digital documents. digital signatures and cryptography.Chapter 2 Literature Review Nowadays the digital media is easily to be reproduced due to the rapidly growth of internet and the multimedia technologies. This chapter overviews previous work in digital video watermarking and related fields. in the case of multimedia watermarking. Therefore. the hidden signal should not result in any visible or audible distortions.

everyone’s computer can have high quality video compression. and compounding processing power. Without solving this security issue.CHAPTER 2. It is based on public key cryptography and one-way hash functions. a comparison among different video watermarking is given. Digital signature is commonly used to authenticate digital transmissions. Multimedia users had the ability to tamper with. and illegally redistribute digital content. digital multimedia products and services cannot take-off in an e-commerce setting [17]. produce copies of. Moreover. Besides. which is . 2. dense portable storage media. LITERATURE REVIEW 8 cussed. the essential ingredients of video watermarking are presented. these technological advances lead to another crisis. By passing the document through a publicly available one-way hash function. Digital signature and cryptography are currently two standardized approaches to protect the digital contents. multimedia documents are rarely available to the mass consumer market. as the rapidly development of the pervasive digital information technology. increasing network bandwidth and accessibility. a unique identifier is generated. Nevertheless. Then different watermarking algorithms are implemented and evaluated.1 Security in Multimedia Communications In a decade years ago. The following section reviews a number of techniques proposed in the literature. the essential differences and the advantages that watermarking techniques provide over these existing technologies are explained. However. Finally.

a string is produced and it is referred to as the digital signature. the receiver may not able to verify the authentic document. Thus. If just one bit differs from the original. However. When transmitting the multimedia documents. LITERATURE REVIEW 9 signed with the owner’s private key. with this scheme. the hash identifier test will fail. This technique is not effective because it does not provide permanent protection for the multimedia content after delivery. the intellectual property owner is not able to trace the . the permitted recipient is able to access to original proprietary data. Thus. which can then be reproduced perfectly and redistributed inexpensively.CHAPTER 2. for instance due to lossy compression for efficient network transfer. In addition. The content of the documents are protected form manipulation and stealing during delivery as the assessment of the document is only permitted to those who possess the appropriate key. this method of tamper detection is too strict for multimedia objects. Use of cryptographically secure license keys is another method for protecting digital intellectual property. the document and signature are not bound in any noteworthy manner. In addition to the signed document. the intended recipients obtain public keys from certification authorities [16]. Then. they may become separated accidentally in transit or intentionally by a malicious party. It dose not allow the document to undergo compression and format changes while still maintaining their authenticity. Moreover. The document is authentic only when it matches with the decrypted signature by applying the hash function. However the critical flaw in this solution is that after transmitted and delivery the document [17].

and are impossible to be removed without leaving any engram or causing severe damage to the contents of the document. cheques and official correspondence. for long time ago. Digital watermarking may be one of the suitable solutions. i.e. These marks are convincingly hard to fake. Moreover. Paper watermarks are faint designs that are embedded by the manufacturer into the paper used to produce such hardcopies. such as money. Finally is the flexibility of the scheme. According to the findings. and at the same time they do not obstruct the normal processing. reading. the additional information should be perceptually invisible as the multimedia documents are ultimately processed by human viewers or listeners and the contents should not be affected. .CHAPTER 2. It is an analogous techniques that have been used to protect valuable hardcopy documents. Digital watermarking technologies strive to achieve these goals in a digital environment by inserting a retrievable watermark directly into the softcopy data stream [17]. we notice that an ideal solution must somehow integrate security information directly into the content of the multimedia document and the security information should be inseparable from the document during its useful lifespan. It should be able to support identification of different copies of the document. LITERATURE REVIEW 10 responsibilities of pirating the properties.

the encrypted file can be made available through the Internet.1. Alice and Bob want to communicate in secret. trusted couriers are used as a solution to this problem. Before delivery. but would be useless to a pirate without appropriate key. One example using symmetric cryptosystem is shown in Figure 2. It is probably the most common method of protecting digital documents and certainly one of the best developed as a science. while . you then would simply use that secure channel to send your message instead of encrypting your message with symmetric cryptosystem. the structure of the message is changed. It is meaningless and unintelligible unless it is decrypted [18]. There are two kinds of cryptosystems: symmetric and asymmetric [19]. in theory. named as private key. Symmetric cryptosystems have a problem: ”how do you transport the secret key from the sender to the recipient securely and in a tamper proof fashion?” [19]. Then. Commonly. to encrypt and decrypt a message.CHAPTER 2. and asymmetric cryptosystems use one key. Asymmetric cryptosystems are also called public key cryptosystems . Symmetric cryptosystems use the same key. named as public key. the content is encrypted and the a decryption key is provided only to those who have permission to access the legitimate copies of the content. LITERATURE REVIEW 11 2.2 Cryptography Cryptography is the first technology that content owners would turn to. to encrypt a message and a different key. to decrypt it. known as the secret key. After encrypted. If you could send the secret key securely.

and so cannot reconstruct the original messages. Asymmetric cryptosystems is different. so they can keep communication by using cryptography. but it is otherwise secret [20].CHAPTER 2. . It’s not possible to determine the secret key from the public key. such as RSA. Asymmetric cryptosystem is another more efficient and reliable solution. so they both know it. With a good encryption algorithm. on-line businesses or just friends trying to have a private conversation. They cannot stop Eve listening to their radio signals. Alice and Bob could be military jets. Alice and Bob exchange a digital key. The encrypted messages are useless to Eve. who does not know the key. this scheme can work well. but exchanging the key while keeping it secret from Eve is a problem. which is the popular security tool [20]. because it splits the key up into a public key for encryption and a secret key for decryption. Alice uses this key to encrypt messages she sends.1: Symmetric Cryptosystem Eve wants to eavesdrop. LITERATURE REVIEW 12 Alice Bob Key Key Message Encrypt Eva Decrypt Message Figure 2. and Bob reconstructs the original messages by decrypting with the same key.

his public key. while only he knows his secret key. Bob generates a pair of keys and tells everybody. but only Bob knows the secret key to decrypt it.2: Asymmetric Cryptosystems In the Figure 2. . Anyone can use Bob’s public key to send him an encrypted message. including Eve.2. LITERATURE REVIEW Bob Public Key Alice Bob 13 Public Key Secret Key Key Generator Message Encrypt Decrypt Message Figure 2. This scheme allows Alice and Bob to communicate in secret without having to meet.CHAPTER 2.

inseparable from the work. apart from depending on the watermark information W’. and undergoes the same transformation as the work [21].2: . W ) (2. etc. the host data I. Watermarking is a key process in the protecting copyright ownership of electronic data. shown in Equation 2. is input to the watermarking algorithm and the algorithm watermarks the image with a watermark W and output the watermarked image I’ with the Equation 2. A simple watermarking idea is shown in Figure 2. K. Even if the existence of a watermark is known. LITERATURE REVIEW 14 2. may also depend on a key K and the host data I into which it is embedded.3 Digital Watermarking Watermarking technique is a particular embodiment of multimedia security. . . such as stegoimage. it should be ideally impossible for an attacker to remove or destroy the embedded watermark without rendering the cover object unusable. Watermark is a design of the watermark signal W to be added to the host signal.1) In watermarking algorithm.CHAPTER 2. The term watermarking comes from using the invisible ink to write secret messages [18]. videos. The watermark signal. watermark has three distinct properties imperceptible.1 W = f0 (I. Generally. audio . Digital Watermark is defined as a digital signal or pattern inserted into a digital data. such as the case in public watermarking schemes. which can also be referred to copyright information. including image.3. The additional requirement for watermarking is robustness.

K) (2. shown in Equation 2. perhaps with the help of the key and the original.CHAPTER 2.3: Watermarking Embedding and Detection Scenario I W −→ I (2.3) .2) Verification algorithm is a design of the corresponding extraction method that recovers the watermark information from the signal mixture. LITERATURE REVIEW 15 Watermark W Stego-Image I Watermarking Watermarked Image I Secret / Public Key K Watermark W and/or Original Image I Test Image I Detection Algorithm Watermark or Confidence measure Secret / Public Key K Figure 2. I = g(I. I .3 .

1 Fidelity What requirements should an ideal watermarking system have? The first requirement would clearly be that of Fidelity [22]. robustness. cipher and watermark keys. blind or informed detection.CHAPTER 2. data payload. unambiguous. Ideally. LITERATURE REVIEW 16 2. ready detection. The analysis focuses on image and video watermarking. sensitivity. security. fidelity. tamper resistance.4 Essential Ingredients for Video Watermarking Watermarking systems can be characterized by a number of defining properties including embedding effectiveness. for general purpose applications it is desirable for the embedded mark to be imperceptible to the human eye or ear. cost. Although visible watermarks tend to be more robust. unobtrusiveness. 2. Invisibility is that degree that an embedded watermark remains unnoticeable when . such general requirements are listed and briefly discussed. The relative importance of each property is dependent on the requirement of the application and the role the watermark will play. In this section. or even highly distracting. A watermarking system is of no use to anyone if it distorts the cover image to the point of being useless. modification and multiple watermark. and scalability.4. the watermarked image should be perceptually invisible even on the highest quality equipment. false positive rate. Some of them are common to most practical applications.

the mark should be detectable even after such distortions occurred. such as cropping. intentional. commonly involves many types of distortions. filtering. 2. contrast enhancement. So far researchers have tried to hide the watermark in such a way that it is impossible to be noticed. . such as lossy compression. It is a common opinion [18] that robustness against signal distortion is better achieved if the watermark is placed in perceptually significant parts of the signal. LITERATURE REVIEW 17 a user views the watermarked contents.2 Robustness The ideal watermark must also be highly robust. or malicious attack. The use of music. or.e. or a deliberate attempt to disable or remove the watermark present. A particularly interesting form of unintentional attack is that . etc. images and video signals in digital form. unintentional attack. For watermarking to be useful. i. cropping. rotation and so on. which operate by discarding perceptually insignificant data not to affect the quality of the compressed image. Robustness is the resilience of an embedded watermark against removal by signal processing. .CHAPTER 2. entirely resistant to distortion introduced during either normal use. Unintentional attacks involve transforms that are commonly applied to images during normal use. resizing.4. This depends on the behavior of lossy compression algorithms. resizing. However this requirement conflicts with other requirements such as tamper resistance and robustness.e. in the image case. i. audio or video. contrast enhancement.

resizing. watermarking seeks to encode information in extra bits that compression hopes to remove.1. In the image watermarking case. Meerwald [23] points out that lossy compression and watermarking are inherently at odds. often through a geometric distortion or the addition of noise. Petitcolas [25] provides us with a rough set of reliability or robustness metrics. interpolation etc. ideal watermarking and compression systems are most likely inherently exclusive. Thus. . But video watermarking introduces some issues not present in image watermarking. frame dropping. watermarks hidden among perceptually insignificant data are likely not to survive compression. Video signals are highly susceptible to pirate attacks. an attacker deliberately tries to disable the watermark.CHAPTER 2. Most of Video watermarking scheme base on the techniques of the image watermarking. rotation and cropping is still an open issue. such as translation. including frame averaging. frame swapping. or complete fragility. statistical analysis. Consequently. the resistance to geometric manipulations. In malicious attacks. It may be the case that some watermarking systems may require the watermark to totally destroy the cover object if any tapering is present [24]. A final note is that robustness can include either resilience to attack. shown below in Table 2. LITERATURE REVIEW 18 of image compression. yet such operations are very common and a solution needs to be found before watermarking techniques are successfully applied to image copyright protection.

LITERATURE REVIEW 19 Table 2. yes 1-5 ± 0-10 ± 0-10 Moderate 100-50 16 100-50 0. breaching the entire DVD copy protection mechanism. yes 1-15 ± 0-25 ± 0-25 3×3 2.1: Basic Robustness Requirements Level Zero Standard JPEG Compression Quality Color Correction (GIF) Cropping Gamma Correction Scaling Rotation Horizontal Flip Uniform Noise Contrast Brightness Median Filter 100-90 256 100-90 Low Level 100-75 256 100-75 0. asymmetric key pairs are generally not.3 Use of Keys Another property of an ideal watermarking system is that it implement the use of keys to ensure that the approach is not rendered useless the moment that the algorithm becomes known [22]. . 90 deg. The risk here is that embedded watermarking systems might have their private key discovered. ruining security of the entire system.5 1/3-2 ± 0-5 deg. This was exactly the case when a single DVD decoder implementation left it’s secret key unencrypted. Although private key systems are fairly easy to implement in watermarking.7-1.5-1.4.2 1/2-3/2 ± 0-2 deg. It may also be a goal that the system utilizes an asymmetric key system such as in public/private key cryptographic systems.CHAPTER 2..

2. This can range from a single bit all the way up to multiple paragraphs of text. the possibility of embedding into the image at least 300-400 bits should be granted. system designers should keep well in mind that the number of bits could be hidden into data is not unlimited. Theoretical capacity of embedded watermarks has been examined using information-theoretic concepts.4. LITERATURE REVIEW 20 2.5 Capacity and Speed Slightly less important requirements of an ideal watermarking system might be capacity. Because of the immense size of uncompressed video files and the difficulty of indexing them to search for a specific frame. but very often is fairly small. however. in watermarking systems destined for embedded applications. Furthermore. the watermark detection (or embedding) may not be overly computationally intensive as to preclude its use on low cost micro-controllers Capacity is that amount of information that can be expressed by an embedded watermark. it is an especially important requirement in video watermarking. In any case. A watermarking system must allow for a useful amount of information to be embedded into the image. and speed.CHAPTER 2. General rules do no exist here.4 Blind Detection Blind detection refers to the ability to detect the watermark without access to the original document.4. in the image case. . the watermarking algorithm should allow a predefined number of bits to be hidden. Depending on the application at hand.

4. must be very small. it is important that the complexity of the detection and extraction algorithms be low enough to execute within the specified real-time deadlines. This requirement is not quite as important here as it is in steganography. one does not exist.4. 2. The watermarking algorithm must modify the bits of the cover in such a way that the statistics of the image are not modified in any telltale fashion that may betray the presence of a watermark.4. false-negative. false-positive. i.7 Low Error Probability Even in the absence of attacks or signal distortions. statistically based algorithms have no problem in satisfying this requirement. however such ability must be demonstrated. i. 2.e. .e. and of detecting a watermark when. the probability of failing to detect the watermark. but some applications may require it. in fact. Usually. if watermarking is to be legally credible.8 Real-time Detector Complexity For consumer-oriented watermarking applications.6 Statistical Imperceptibility The last possible requirement of an ideal watermarking system is that of statistical imperceptibility [25]. LITERATURE REVIEW 21 2.CHAPTER 2.

determining both the owner and the integrity of the content.2. We can classify the watermarking techniques according to several points of view. Each owner has a unique watermark. In this section. Although watermarking is a recent field of research. A digital document can be authenticated with what is known as a digital watermark. a brief review of the current video watermarking technologies is presented. A watermark is a secret code or image incorporated into an original content. marking algorithm and verification algorithm.5 Review on Video Watermarking Techniques As a method of intellectual property protection. LITERATURE REVIEW 22 2. Moreover. The marking algorithm incorporates the watermark into the image or video. Various techniques are applied in watermarking algorithms. The verification algorithm authenticates the content. many techniques have already been proposed both in the academic as well as in the industry. In this . A watermarking algorithm consists of three parts: watermark. digital watermarks have recently stimulated significant interest and become a very active area of research. They can be classified into different types based on the offered functionalities. which acts to verify both the owner and content of the document. The use of perceptually invisible watermark is one of the copyright protection. different insertion and extraction methods can be found.CHAPTER 2. as shown in Table 2. A variety of imperceptible watermarking schemes have been proposed over the past few years. Numerous watermarking research tasks have proposed many watermark techniques in terms of various application areas.

40. Moreover. including text [26]. They can be classified into three categories. 50. 34. 32]. 42. 47] and invisible [29. Watermarks. logo and label. compression domain and hybrid. 30. 51]. 44]. need robustness to protect the ownership from various attacks. For image types. frequency domain. Gaussian random and chaotic sequence.CHAPTER 2. Watermark techniques can be classified into visible [46. Digital watermarking can be applied to many different types of documents. embedding meaningful watermark is more important in some applications. invisible watermarks are mostly used . stamp. 36. 45]. The oblivious meaningful video watermarking remains a challenging problem since the original video is often unavailable due to videos’ bulky volume. The random sequence watermark is more robust in general. 35. we focus on analyzing the video watermark processing methods. Applications for copyright protection would require to use a robust watermark. audio [27. 30. robust [48. image [33. In general. A watermark can be a random sequence with one information bit or multiple-bit meaningful information. 41. 43. . 28. on the other hand. LITERATURE REVIEW 23 section. there are binary image. Watermarks to be inserted can also be classified into two types: noise type [54] and image type [41. A noise type includes pseudo noise. 42. semi-fragile [52] and fragile [53] watermarks. however. Applications for proving integrity would employ a fragile or semi-fragile watermark. watermark processing methods are classified into four categories: spatial domain. 39] and video [40. 31. 43. 49. 38. 37. 41] watermarks. 29. 44. Finally. Different applications would be chosen for different levels of robustness according to the requirement.

39] Video[40. 66] Fourier (DFT)[67. 31. 70] MPEG2[71. Table 2. 51] Semi-fragile [52] Fragile [53] Inserting watermark type Noise [54] Information tagging Image [41. 62] Wavelet (DWT) [63. 42. 41.CHAPTER 2. LITERATURE REVIEW 24 watermark extraction methods can be classified as private [55]. 43. 64. 38. 44. according to the necessity of the original media. 34. 32] Audio [33. 41] Robustness of watermark Robust [48. 73] JPEG2000 [74] Hybrid Visual-audio [75. semi-private [56] and public [57] watermarking. 44] Processing method Spatial domain LSB Image checksum [58] Patchwork [59] Random function[60] Frquency domain Spread spectrum Look-up table DCT [61. 47] Invisible[29. 45] Perceptivity of watermark Visible [46. 40. 37. 76] Different watermarks [77] Different watermarking scheme [78] Necessary data for extraction Private [55] Semi-private [56] Public [57] . 28. 49. 50. 35. 43. 72] MPEG4[26. 30. 68] Compression domain MPEG1[42. 36. 29.2: Classification of watermarking according to several viewpoints Classification Inserted media category Contents Text [26] Image [27. 30. 65. 42. 69.

CHAPTER 2. Most of the proposed video watermarking scheme based on the techniques of the image watermarking and applied to raw video or the compressed video. In the following sections.1 Video Watermarking Many digital watermarking schemes have been proposed in the literature for still images and videos. real-time performance. They can be divided into 3 main groups based on the domain that the watermark is embedded. These schemes can be distinguished in terms of the domain that the watermark being embedded or detected. such as video object and redundancy of the large amount video data. LITERATURE REVIEW 25 2. their capacity. 80]. Besides. each class of algorithms is briefly described. 81]. 79. Most of them operate on uncompressed videos [10. the degree to which all three axes are incorporated. while others embed watermarks directly into compressed videos [72. and their resistance to particular types of attacks [17]. strength and limitation introduced by these schemes.5. . researchers have make use of those characteristics to develop different schemes. we present the important idea.4. As some issue in video watermarking is not present in image watermarking. A classification map of the existing video watermarking techniques is presented in Figure 2. researchers tend to investigate video watermarking techniques that is robust and invisible. they are spatial domain. Recently. frequency domain and MPEG coding structure based.

in the pixel domain.4: Classification map of existing digital video watermark techniques 2.CHAPTER 2. • The watermark can be detected by correlating the expected pattern with the received signal. Algorithms in this class generally share the following characteristics: • The watermark is applied in the pixel or coordinate domain. As a result they have proven to be most attractive for . LITERATURE REVIEW 26 Invisible Robust Video Watermarking Techniques Spatial Domain Method Frequency Domain Method MPEG coding Structure based Method Figure 2.2 Spatial Domain Watermarks We first review the video watermarking techniques in the spatial domain.5. • The watermark is derived from the message data via spread spectrum modulation. The main strengths of pixel domain methods are that they are conceptually simple and have very low computational complexities. • No transforms are applied to the host signal during watermark embedding. • Combination with the host signal is based on simple operations.

CHAPTER 2. LITERATURE REVIEW

27

video watermarking applications where real-time performance is a primary concern. However, they also exhibit some major limitations: The need for absolute spatial synchronization leads to high susceptibility to de-synchronization attacks; lack of consideration of the temporal axis results in vulnerability to video processing and multiple frame collusion; and watermark optimization is difficult using only spatial analysis techniques. The three methods that fall into this class can be distinguished by the dimensionality of the watermark pattern. Techniques based on 1D and 2D spread spectrum modulation, and 3D Code Division Multiple Access (CDMA) modulation have been proposed. Several different methods enable watermarking in the spatial domain. The simplest is to just flip the lowest-order bit of chosen pixels in a grey scale or colour image. This will work well only if the image is subjected to any human or noisy modification. A more robust watermark can be embedded in an image in the same way that a watermark is added to paper. Such techniques may superimpose a watermark symbol over an area of the picture and then add some fixed intensity value for the watermark to the varied pixel values of the image. The resulting watermark may be visible or invisible depending upon the value (large or small, respectively) of the watermark intensity. One disadvantage of spatial domain watermarks is that picture cropping, which is a common operation of image editors, can be used to eliminate the watermark. Spatial watermarking can also be applied using colour separation. In this way, the watermark appears in only one of the colour bands. This renders the watermark visibly subtle so that

CHAPTER 2. LITERATURE REVIEW

28

it is difficult to detect under regular viewing. However, the watermark appears immediately when the colours are separated for printing or xerography. This renders the document useless to the printer unless the watermark can be removed from the colour band. This approach is used commercially for journalists to inspect digital pictures from a photo-stockhouse before buying non-watermarked versions.
Least Significant Bit Modification

The most straight-forward method of watermark embedding, would be to embed the watermark into the least-significantbits of the cover object [82]. Given the extraordinarily high channel capacity of using the entire cover for transmission in this method, a smaller object may be embedded multiple times. Even if most of these are lost due to attacks, a single surviving watermark would be considered a success. LSB substitution however despite its simplicity brings a host of drawbacks. Although it may survive transformations such as cropping, any addition of noise or lossy compression is likely to defeat the watermark. An even better attack would be to simply set the LSB bits of each pixel to one . . . fully defeating the watermark with negligible impact on the cover object. Furthermore, once the algorithm is discovered, the embedded watermark could be easily modified by an intermediate party. An improvement on basic LSB substitution would be to use a pseudo-random number generator to determine the pixels to be used for embedding based on a given ”seed” or key [82]. Security

CHAPTER 2. LITERATURE REVIEW

29

of the watermark would be improved as the watermark could no longer be easily viewed by intermediate parties. The algorithm however would still be vulnerable to replacing the LSB’s with a constant. Even in locations that were not used for watermarking bits, the impact of the substitution on the cover image would be negligible. LSB modification proves to be a simple and fairly powerful tool for stenography, however lacks the basic robustness that watermarking applications require.
Correlation-Based Techniques

Another technique for watermark embedding is to exploit the correlation properties of additive pseudo-random noise patterns as applied to an image [83]. A pseudo-random noise (PN) pattern W(x,y) is added to the cover image I(x,y), according to the equation shown below in Equation 2.4. Iw (x, y) = I(x, y) + k × W (x, y) (2.4)

In Equation 2.4, k denotes a gain factor, and IW the resulting watermarked image. Increasing k increases the robustness of the watermark at the expense of the quality of the watermarked image. To retrieve the watermark, the same pseudo-random noise generator algorithm is seeded with the same key, and the correlation between the noise pattern and possibly watermarked image computed. If the correlation exceeds a certain threshold T, the watermark is detected, and a single bit is set. This method can easily be extended to a multiple-bit watermark by

the notion of a threshold being used for determining a logical ”1” or ”0” can be eliminated by using two separate pseudorandom noise patterns.5. and performing the above procedure independently on each block. even after the image has been subject to attack [83]. One pattern is designated a logical ”1” and the other a ”0”. For transformed domain techniques. FFT and wavelet transform are used as the methods of data transformation. designing a watermarking scheme in the Discrete Cosine Transform (DCT) domain leads to better implementation compatibility with popular video coding algorithms such as Moving Pictures Experts Group . and the watermark. and the pattern with the higher resulting correlation is used. In these methods. a watermark that one wishes to embed distributively in overall domain of an original data. This basic algorithm can be improved in a number of ways. First. they have hierarchical watermarking with Discrete Cosine Transform. The main strength offered by transform domain techniques is that they can take advantage of special properties of alternate domains to address the limitations of pixel-based methods or to support additional features. LITERATURE REVIEW 30 dividing the image up into blocks. is hardly to be deleted once embedded.3 Frequency Domain Watermarks Generally DCT. Discrete Wavelet Transform or Discrete Fourier Transform. For instance. This increases the probability of a correct detection. The above procedure is then performed once for each pattern.CHAPTER 2. subband watermarking techniques. 2.

To begin. We discuss the details of three methods here: Discrete Cosine Transform. Discrete Cosine Transform The classic and still most popular domain for image processing is that of the Discrete Cosine Transform. or DCT. analysis of the host signal in a frequency domain is a prerequisite for applying more advanced masking properties of the HVS to enhance watermark robustness and imperceptibility.CHAPTER 2. the main drawback of transform domain methods is their higher computational requirement. we define the middle-band frequencies (FM) of an 8×8 DCT block . and Discrete Fourier Transform. LITERATURE REVIEW 31 (MPEG)-2. The DCT allows an image to be broken up into different frequency bands. and in the shift and rotation-invariant Fourier domains facilitates the design of watermarks that inherit these attractive properties. making it much easier to embed watermarking information into the middle frequency bands of an image. One such technique utilizes the comparison of middle-band DCT coefficients to encode a single bit into a DCT block. The middle frequency bands are chosen such that they have minimize they avoid the most visual important parts of the image (low frequencies) without over-exposing themselves to removal through compression and noise attacks (high frequencies) [83]. Discrete Wavelet Transform. Besides. Generally.

vertical (LH) and diagonal (HH) detail components. such as the high resolution detail bands LH. One of the most straightforward techniques is to use a similar LL2 HL2 HL1 LH2 HH2 LH2 HH1 Figure 2. as in the 2 scale wavelet transform shown below in Figure 2. The DWT (Discrete Wavelet Transform) separates an image into a lower resolution approximation image (LL) as well as horizontal (HL). LITERATURE REVIEW Discrete Wavelet Transform 32 Another possible domain for watermark embedding is that of the wavelet domain. Embedding watermarks in these regions allow us to increase the robustness of our watermark. This allows us to use higher energy watermarks in regions that the HVS is known to be less sensitive to.5: 2 Scale 2-Dimensional Discrete Wavelet Transform .CHAPTER 2. at little to no additional impact on image quality [83].HH). One of the many advantages over the wavelet transform is that that it is believed to more accurately model aspects of the HVS as compared to the FFT or DCT.5.HL. The process can then be repeated to computes multiple ”scale” wavelet decomposition.

A spatial domain reference watermark is used to obtain invariance to geometric attacks by employing image registration techniques to determine and invert the attacks. as the embedding uses the values of the transformed value in embedded. rotation. In [85]. The proposed method is robust against the attack of frame dropping. if the correlation exceeds threshold for a particular sequence a ”1” is recovered. Furthermore. cropping. A second. The author [23] claims that the technique should prove resistant to JPEG compression. This scheme is a high capacity blind video watermarking system invariant to geometrical attacks such as shift. and other typical attacks. which are then added to the detail coefficients. scaling and cropping. the decomposed watermark image with different resolution is embedding in the corresponding resolution of the decomposed video by means of multiresolution signal decomposing. storing the majority of the watermark in the larger coefficients. Niu and Sun proposed a new wavelet-based digital watermarking in [84]. This paper proposes a method of embedding a digital watermark image in video. LITERATURE REVIEW 33 embedding technique to that used in the DCT. In the watermarking. During detection. averaging and lossy compression. Serdean et al proposed another DWT-base scheme. high capacity watermark. As in the spatial version. This can be easily extended to multiple bit messages by embedding multiple watermarks into the image. otherwise a zero. The recovery process then iterates through the entire PN sequence until all the bits of the watermark have been recovered.CHAPTER 2. which carries . the embedding process should be rather adaptive. a separate seed is used for each PN sequence.

Discrete Fourier Transform M. JPEG compression and resist to geometric transformations as scaling. then IDFT. leaving the other ones uncorrupted. To generate the watermark. Barni et al proposed a robust watermarking approach for raw video in [87]. The DFT coefficient is altered. is embedded in the wavelet domain according to a human visual system model. The noise-like watermark is statically undetectable to thwart unauthorized removal.CHAPTER 2. Furthermore. the resulting wavelet coefficient frames are modified by a perceptually shaped pseudorandom sequence representing the author. rotation and cropping. The watermark consists of static and dynamic temporal components that are generated from a temporal wavelet transform of the video scenes. The watermark is composed of two alphanumerical strings. computing its full-frame DFT and then taking the magnitude of the coefficients. LITERATURE REVIEW 34 the data payload. This approach first extracts the brightness of the to-be-marked frame. Decide to watermark one or more frames . The multiresolution watermark may be detected on single frames without knowledge of the location of the frames in the video scene. the author representation resolves the deadlock problem. Mitchell et al proposed multiresolution video watermarking using perceptual models and scene segmentation in [86]. which was composed of twelve frames. Only the first frame of each GOP is watermarked. It is good robustness to the usual image processing as linear/non-linear filtering. sharpening.

Compression in block-based schemes like MPEG-2 is achieved by using forward and bi-directional motion prediction to remove temporal redundancy. including approaches based on GOP modification [69]. a trade-off between time spent for marking and the degree of robustness needed for the sequence can be achieve 2. [88] and three more robust and general algorithms that will be discussed in detail in this section. Vassaux et al proposed a video object watermarking which is based on the structure of MPEG-4 in [89].5. The two MPEG-2 watermarking methods considered here embed hidden data by swapping level-adjacent Variable Length Codeword (VLC) codeword and manipulating mean luminance over regions of pixels. DCT block classification [71]. This paper presents a so-called scrambling technique which allows adapting any classical spread spectrum watermarking scheme operating in the . high frequency DCT coefficient manipulation. -2 and -4 coding structures as primitive components are primarily motivated by the goal of integrating watermarking and compression to reduce overall real-time video processing complexity. There are a number of MPEG-2 and -4-based techniques that have been proposed. as well as conversion to formats other than MPEG.4 Watermarks Based on MPEG Coding Structures Video watermarking techniques that use MPEG-1. LITERATURE REVIEW 35 in GOP. and statistical methods to remove spatial redundancy.CHAPTER 2. One of the major drawbacks of schemes based on MPEG coding structures is that they can be highly susceptible to re-compression with different parameters.

To address issues associated with video motion and redundancy. Special decomposition of mpeg-4 include the fact that VO have significant value Watermark information has to present in each VO In [90]. a robust watermarking scheme foe video object protection is proposed. Video object may be purposely cut and pasted for illegal use.CHAPTER 2. As a result. et al. Swanson. them binary sequence is generated using secret key. Each watermark is created by shaping an author and video dependent pseudo-random sequence according to the perceptual masking characteristics of the video. Dividing the image into blocks of equal size. To solve the asyn- . In this paper. LITERATURE REVIEW 36 spatial domain to the Mpeg-4 requirements concerning VO manipulation. Furthermore. This technique can be easily added to the embedding and detection schemes without changing the watermarking algorithm. and then adds to the image. the noise like watermark is statistically undetectable to prevent unauthorized removal. individual watermarks are created for objects within the video. Lu and Liao proposed another video object-based watermarking in [91] which is resist to rotation and flipping. the watermark adapts to each video to ensure invisibility and robustness. presents a watermarking procedure to embed copyright protection into video sequences which is objectbased transparent. It modifies some predefined pairs of quantized DCT coefficient in the luminance block of pseudo-randomly selected MBs. a watermark is embedded by a new technology designed based in the concept of communication with side information. For each segmented video object. It is based on spread-spectrum techniques. Video object is very important concept in Mpeg4 standard.

frame reordering etc. LITERATURE REVIEW 37 chronous problem caused by object placement. Watermark. First. video signal is modeled as a sequence of bit planes arranged along the time axis. The watermark is also identifiable from very short segment of video. it proposes to use eigenvectors of a video object for synchronization of rotation and flipping.CHAPTER 2. Mobasseri proposed direct sequence watermarking using m-frames This scheme applies a direct sequence spread spectrum model to the watermarking of the digital video. . A controlling m-sequence first establishes a pseudorandom order in the bit plane stream for later watermarking. Moreover. supplant the tagged bit planes. attempts in corrupting the image to destroy the watermark render the video useless before damaging the seal itself. In [92]. The watermarked video is also robust to video editing attempts such as subsampling. defined as m-frames. Watermarking of this sequence is a two layer operation. Individual frames extracted from the video will also contain the copyright information.

We have chosen some representative watermarking schemes in each category for implementation and performed experiments to compare their robustness.0 benchmark program [100. Therefore. 95]. LITERATURE REVIEW 38 2. 101] and 30 different images are used. the StirMark 4. we find that the watermarking schemes in spatial domain are less robust than those in frequency domain. From the result.6 Comparison between Different Watermarking Schemes In general. compression. Discrete Wavelet Transform (DWT) based watermarking scheme [97]. threshold-based correlation and m-sequence watermarks are perform worse than the other five implemented watermarking algorithms. The tests are divided into the following sections: PSNR. scaling. The comparison is shown in Table 2. these watermarking algorithms are . They are: Least Significant Bit (LSB) based watermarking scheme [93].3. row/column removal. DFT with template matching [96]. and transformed domain watermark. Discrete Cosine Transform (DCT) based watermarking scheme [98] and spread spectrum [99] watermarking scheme. To evaluate the algorithms.CHAPTER 2. Direct sequence watermark using m-frame [94. rotation. For each image. watermarking schemes can be roughly divided into two categories: spatial domain watermark. Each attack is considered by itself and it is applicable after watermarking. A value of zero is assigned if the watermark is incorrect. we assign a score of 1 if the watermark is correctly decoded in the case. LSB. threshold-based correlation watermarking scheme [83]. shearing. and random geometric distortions. cropping.

83 0.56 0.3 0.63 0. rescaling.based Correlation 0.21 0.78 0.3 m-sequence / m-frame 0.98 0.7 0. DCT-based watermarking scheme is the most robust to lossy compression.65 0.83 0.10 0.75 0.14 0.78 1 0.76 0.75 0.3: Comparison between different watermarking schemes Attack Class JPEG Lossy Compression PSNR Add Noise Median Filter Row / Column Removal Cropping Rescale Rotation Affine Geometrical Distortions Shearing LSB 0.5 0.65 0. The middle frequency .85 0.85 Attack Class JPEG Lossy Compression PSNR Add Noise Median Filter Row / Column Removal Cropping Rescale Rotation Affine Geometrical Distortions Shearing Mid-band DCT 1 0. rotation and shearing.62 0.6 classified as fragile or semi-fragile watermarking.7 0.85 0.25 0.62 0.35 0.54 Spread Spectrum 0.4 0.61 0.82 0.35 0.20 0.76 0.4 0. In this approach.13 0.5 0.45 0.27 Threshold .78 0.75 0. pixel removal.64 0.75 1 0.5 0.89 0.98 0.46 0.73 0.4 0.52 0.83 0.86 0.98 0.69 0.35 Mid-band DWT 0.42 DFT template Matching 0. making it much more easier to embed watermarking information into the middle frequency bands of the image.53 0.52 0.75 0.5 0.85 1 0.4 0.89 0.75 0.49 0.37 1 Radon Transform 0. LITERATURE REVIEW 39 Table 2.CHAPTER 2.89 0.15 0.3 0.74 0. particularly in lossy compression.9 0. noise addition.62 0.22 0.75 0.25 1 0. They can be applied for the purpose of proving the integrity of a document.42 0. The frequency domain watermarking schemes are relatively more robust than the spatial domain watermarking schemes.75 0.95 0.81 0. an image is broken up into different frequency bands by DCT.7 0.75 0.

The purpose of the template is to enable resynchronization of the watermark payload spreading sequence.CHAPTER 2. DFT-based watermarking scheme with template matching can resist a number of attacks. Some algorithms embed only one bit information as the watermark. however. LITERATURE REVIEW 40 bands are chosen to minimize the change of the most visually important parts of the image (low frequencies) without overexposing themselves to the removal through compression and noise attacks (high frequencies). A bispectrum feature vector of the image is used as the watermark. invariant watermarks use the Radon transform and higher order spectra. including pixel removal. includes: i) The watermark is not robust to attacks which are specifically targeted at to videos. The weakness of the existing algorithms. but rather by modifying the value of the target coefficient. such that it is at least two standard deviations above the mean. In the scheme. This approach differs from the previous methods in that it embeds watermarks into the phase of the higher order spectra. DWT-based watermarking scheme is the most robust to noise addition. Moreover. averaging and statistical analysis. which is also embedded into DFT magnitude representation of the frame. iii) Existing techniques are not aware of the usefulness of the audio channel in a video. iv) None of the existing watermarking . ii) The bit rate of the watermark is low. such as frame dropping. rotation and shearing. The Radon embedding grid also outperforms the Fourier-Mellin based methods. It is a key dependent pattern of peaks. Radon transformation resists attacks by resealing and geometric distortion. The peaks are not embedded by addition.

LITERATURE REVIEW 41 schemes resists to all the attacks.CHAPTER 2. 2 End of chapter. To tackle these problems. in this paper. v) A frequency domain watermark is more robust than a spatial domain watermark. . we propose a novel watermarking scheme based on scene changes with a hybrid approach.

which is robust against the attacks of frame dropping. which can improve the robustness of the watermarking scheme [12. 13]. and the watermark is then decomposed into different parts which are embedded in corresponding frames of different scenes in the original video.Chapter 3 Novel Watermarking Schemes In this chapter. In the following sections. we present the proposed innovative digital video watermarking scheme. A scene-based video watermarking scheme is proposed. averaging and statistical analysis. Moreover. a GAbased watermarking scheme is presented. a video and a watermark are taken as the input. In our scheme. Figure 3. 42 . 3.1 A Scene-based Video Watermarking Scheme The new watermarking scheme we propose is based on scene change analysis. the detail of each algorithm is described.1 shows an overview of our watermarking process. a hybrid approach is proposed. To enhance the fidelity of the scheme. which were not solved effectively in the past [11].

C[i+2]. C[i+1]. NOVEL WATERMARKING SCHEMES 43 DWT Video Normalization Scene Change Detection DWT Watermarked Video Embedding Algorithm if W[j] = 1. Exchange C[i] with max(C[i]. C[i+3].CHAPTER 3. C[i+3]. C[i+4]) else Exchange C[i] with min(C[i]. C[i+2]. C[i+4]) Key 8-bit planes Encryption DWT Crop Bit Decomposition Encrypted Watermark Watermark Encryption Encrypted Watermark Place the bit-plane side by side Figure 3.1: Overview of the watermarking process . C[i+1].

The watermark is first scaled to a particular size as follows:. However.1. we use an identical watermark within each motionless scene. Consequencely.1 Watermark Preprocess A watermark is scrambled into small parts in a preprocess. q > 0 (3. the proposed method is robust against the attacks of frame dropping. and statistical analysis. This newly proposed scheme consists of four parts. watermark embedding.CHAPTER 3. A 256-greylevel image is used as the watermark. averaging. With these mechanisms. and watermark detection.2) . video preprocess. NOVEL WATERMARKING SCHEMES 44 As applying a fixed image watermark to each frame in the video leads to the problem of maintaining statistical and perceptual invisibility [102]. our scheme employs independent watermarks for successive but different scenes. These motionless regions may be statistically compared or averaged to remove the independent watermarks [10. 103]. including: watermark preprocess. 2n ≤ m.1) (3. n > 0 p + q = n. so 8 bits can represent each pixel. p. and they are embedded into different scenes so that the scheme can resist a number of attacks towards to the video. applying independent watermarks to each frame also presents a problem if regions in each video frame remain little or no motion frame after frame. 3. swapping. Details are described in the following sections.

. Sample preprocessed watermarks are shown in Figure 3.CHAPTER 3.3) Then the watermark is divided into 2n small images with size 64 × 64. m0 . Key 8-bit planes m0 DWT Encryption Encrypted Watermark m0 ' Crop Watermark Bit Decomposition m7 ' DWT Encryption Encrypted Watermark m7 Place the bit-plane side by side Figure 3. the processed watermarks m are transformed to the wavelet domain and encrypted [104].2: Preprocessing the watermark .2 shows the procedure of the watermark preprocess with m = 10. n = 3. The size of the watermark is represented as 64 · 2p × 64 · 2q (3. where (a) is the original watermark. i. and q = 2. Figure 3. q are positive integers. p. In the next step. and a large image mn can be obtained by placing the bit-planes side by side only consisting of 0’s and 1’s. p = 1. each small image is decomposed into 8 bitplanes.e. and totally 2n independent watermarks are obtained.3. (b)-(i) represent the scrambled watermarks in the spatial domain. and (j) shows the encrypted watermark of (b). To make the scheme more robust. These processed images are used as watermarks. NOVEL WATERMARKING SCHEMES 45 where m is the number of scene changes and n.

HHj . All frames in the video are transformed to the wavelet domain. and con- Figure 3. The low frequency subband is a lowpass approximation of the original frame. and three series of high-frequency subbands LHj . HLj .3: (a) Original watermark (b-i) Preprocessed watermark m0 − m7 (j) Encrypted watermark m0 .2 Video Preprocess Our watermark scheme is based on 4 levels Discrete Wavelet Transform (DWT). The frames are decomposed in 4 level subband frames by separable 2-D wavelet transform.CHAPTER 3. NOVEL WATERMARKING SCHEMES 46 3. where j < 4.1. It produces a low-frequency sub-band LL4 .

each pixel is checked and classified into different classes. the capacity of the scheme would be decreased. we only embed the watermark in the middle frequency subbands. the total difference of the whole histogram (H) is calculated as: 63 H= i. blue B). The threshold is again determined by experiments. we consider there is a scene change. NOVEL WATERMARKING SCHEMES 47 tains most of energy of the frame. only the most significant two bits for each color are considered. If less than 4 levels is applied. Moreover. The other subbands include edge components of horizontal. the more important the subbands. a and b. . LL4 is the most important then LHj . If H > threshold(T ). the quality of the watermark video is affected. scene changes are detected from the video by applying the histogram difference method on the video stream. which is determined by experiments. According to the energy distribution. Our scheme is based on 4 levels DWT. For different levels. respectively. and HHj .CHAPTER 3. Then. eight bits for each color (red R. HLj . the higher the level. The histogram difference method is used for scene change detection. In our scheme. vertical and diagonal directions at different scale and resolution.4) where Pa and Pb are the frequency distribution of the pixel level of two images. Each frame is coded in 24-bit image. For efficiency purpose. Consequently. if larger than 4 levels is applied. i and j are two successive columns in the color histogram.j=0 [Pa (i) − Pb (j)]2 (3. green G.

Ci+1 . Ci+2 .3 Watermark Embedding Watermark is then embedded to the video frames by changing position of some DWT coefficients with the following condition: if Wj = 1 Exchange Ci with max(Ci .1.4. watermark m1 is used for the first scene. The watermark for each scene can be chosen with a pseudo-random permutation such that only a legitimate watermark detector can reassemble the original watermark.CHAPTER 3.4: Scene change detection Independent watermarks are embedded in frames of different scenes. an identical watermark is used for each frame. Within a motionless scene. When there is a scene change. Ci+4 ) else Exchange Ci with min(Ci . As shown in Figure 3. Ci+3 . Ci+3 . 3. Ci+2 . Ci+4 ) . another watermark m3 is used for the next scene. Ci+1 . NOVEL WATERMARKING SCHEMES 48 Figure 3.

The sequence of watermark coefficients used is depicted in Figure 3. Ci+2 . When the watermark Wj = 1. Ci+3 . Ci+1 . where higher frequency coefficients are embedded to higher frequency parts of the video frame. Ci+4 . we perform an exchange of the Ci with the maximum value among Ci . NOVEL WATERMARKING SCHEMES 49 (3.5: Embedding watermarks in a frame . and Wj is the j th pixel of a corresponding watermark image [105].5. Ci+2 .CHAPTER 3. When Wj = 0. LL4 HL 4 LH 4 HH4 HL3 HL2 HH3 HL1 LH3 LH2 HH2 LH1 HH1 Encrypted Watermark mj Figure 3. Ci+1 . With this algorithm. and only the middle frequency wavelet coefficient of the frame (middle frequency subband) is watermarked [10]. we perform an exchange of the Ci with the minimum value among Ci . Ci+4 . the retrieval of the embedded watermark dose not need the original image.5) where Ci is the ith DWT coefficient of a video frame. Ci+3 .

In this step. Also.. i.4 Watermark Detection The video is processed to detect the video watermark. W Ci+4 .6) where W Ci is the ith DWT coefficient of a watermarked video frame. and EWj is the j th pixel of the extracted watermark [106]. multiple copies of each part of the watermark may be obtained. W Ci+3 . The watermark is recovered by averaging the watermarks extracted from different frames. i.1. W Ci+1 . When the watermark W Cj is greater than median value among W Ci .e. scene changes are detected from the tested video. EWj = 1. W Ci+4 ) else EWj = 0 (3.e. NOVEL WATERMARKING SCHEMES 50 3. . This is an important property to video watermarking. With this algorithm. W Ci+3 . W Ci+1 . This reduces the effect if the attack is carried out at some designated frames. W Ci+2 .. each video frame is transformed to the wavelet domain with 4 levels.CHAPTER 3. Thus we can combine the 8 bit-planes and recover the 64 × 64 size image. As an identical watermark is used for all frames within a scene. the extracted watermark is considered as one. W Ci+2 . otherwise. Then the watermark is extracted with the following condition : if W Ci > EWj = 1 median(W Ci . the retrieval of the embedded watermark dose not need the original image. EWj = 0. it is considered as zero.

Figure 3.3(g) (d) Recovered watermark. This can be shown in Figure 3.. the watermarked frame.6. If enough scenes are found and all parts of the watermark are collected. and the extracted watermark are depicted. where the original frame.6: (a) Original frame (b) Watermarked frame (c) Extracted watermark corresponding to Figure 3.CHAPTER 3. the original large watermark image can be reconstructed. NOVEL WATERMARKING SCHEMES 51 i. . 1/2n part of the original watermark.e.

1 Performance In this section. 3. p. the performance of the proposed algorithm is calculated. q are positive integers.9) .8) where m is the number of scene changes and n.2 Theoretical Analysis In this section. and they are embedded into different scenes.7) (3. n > 0 p + q = n.CHAPTER 3. NOVEL WATERMARKING SCHEMES 52 3. Size of the video frame = n1 × n2 Size of the watermark = m1 × m2 Number of frames = T Number of scene change = m Prepare Watermark To prepare the watermark for the scheme.1. a watermark is scrambled into small parts in preprocess. q > 0 (3. Let T be the total number of frame in a video and n1 × n2 be the size of the video frame and m the total number of scene change in the video. The watermark is first scaled to a particular size with the Equation 3. The size of the watermark should be 64 · 2p × 64 · 2q (3. 2n ≤ m. p.2. performance and the capacity of the proposed watermarking scheme will be calculated.

m]) Scene Change Scene changes are detected from the video by applying the histogram difference method on the video stream. Then. only the most significant two bits for each color are considered. For efficiency purpose. blue B). each pixel is checked and classified into different classes. p) in Equation 3. eight bits for each color (red R. m2 ).j=0 . the total difference of the whole histogram (H) is calculated by Equation 3.CHAPTER 3.7 = O(log m) Find (q. m2 ). m]) Generating different part of watermark = 2n × 64 × 64 = O(m) Total running time = O(log m) + O(1)+ O(max[(m1 . m]) + O(m) = O(max[(m1 .10: 63 [Pa (i) − Pb (j)]2 (3. m2 ).8= ( n . green G. The histogram difference method is used for scene change detection. NOVEL WATERMARKING SCHEMES 53 Then the watermark is divided into 2n small images with size 64 × 64.10) H= i. Each frame is coded in 24-bit image. Find N in Equation 3. Consequently. n − n ) = O(1) 2 2 Resize the watermark: if size of watermark is smaller than video frame = O(m1 m2 ) if size of watermark is greater than video frame = O(64×64×2n ) = O(m) = O(max[(m1 .

HL4 and LH4 are watermarked [10]. HL3 ..CHAPTER 3. i. only the middle frequency wavelet coefficient of the frame (middle frequency sub-band) is watermarked. and three series of high frequency subbands LHj . LH2 .e. HHj . Total number of pixel to watermark . All frames in the video are transformed to the wavelet domain. LH1 . The frame is decomposed in 4 level subband frame by separable 2-D wavelet transform. It produces a low frequency sub-band LL4 . DWT coefficients of HL1 . HLj . Running time for DWT = 2[n1 × n2 + n1 × n2 × 2] + 2[ n1 × n2 + n1 × 2 2 2 2 n2 n1 n2 n1 n2 n1 n2 4 + 4 × 8 × 2] + 2[ 8 × 8 + 8 × 16 × 2] = 4n1 n2 + 2n1 n2 + n1 n2 + n12n2 = 15n1 n2 2 = O(n1 n2 ) n2 4 × 2] + 2[ n1 × 4 When embedding the watermark. LH3 . HL2 . NOVEL WATERMARKING SCHEMES 54 Scanning to generate the histogram for 1 frame = n1 × n2 Create histogram = n1 × n2 Compare the histogram = 64 × 64 Total running time = [2(n1 × n2 ) + 64] × T = O(n1 n2 T ) Embedding watermark Our watermark scheme is based on 4 levels Discrete Wavelet Transform (DWT). where j < 4.

the video frame is inverseDWT.CHAPTER 3. NOVEL WATERMARKING SCHEMES 55 = + = + = =O(n1 n2 ) n1 ×n2 2 n1 ×n2 2 85n1 n2 128 2 n1 ×n2 8 n1 n2 2 × 2 + 2 + 2 ×n + n132 2 + n1 ×n2 128 n1 n2 4 × 4 n1 n2 8 × 8 Number of operation for watermark = 85n1 n2 × T 128 = O(n1 n2 T ) After the watermark is embedded. Running Time =O(max[(m1 . m2 =O(n1 n2 T ) n2 4 × 2] + 2[ n1 × 4 n1 n2 ≥ . m − 2). Running time for IDWT = 2[n1 × n2 + n1 × n2 × 2] + 2[ n1 × n2 + n1 × 2 2 2 2 n2 n1 n2 n1 n2 n1 n2 4 + 4 × 8 × 2] + 2[ 8 × 8 + 8 × 16 × 2] = 4n1 n2 + 2n1 n2 + n1 n2 + n12n2 = 15n1 n2 2 = O(n1 n2 ) Total running time for embedding watermark = O(n1 n2 T ) + 2O(n1 n2 T ) = O(n1 n2 T ) Finally. m]) + O(n1 n2 T ) + O(n1 n2 T ) m1 .

. We will not consider a coefficient as a communication channel [108. or in other domain. we investigate the watermarking capacity based domain-specified masking effects. X2 . and the decoder. . In this way. by statistical models for the host signal. which the power constraint of each state is determined by HSV model or masking effect in some special domains. and by the information available in the game between the information hider.. ie.2 Capacity Watermarking can be viewed as a communication problem with side information available at the encoder and the decoder. XN be the changes of the coefficients in a . The problem is mathematically defined by distortion constraints. 109] because a channel usually incites its reuse temporally. spatially. information theory explains why the performance of watermark decoders that do not have access to the host signal may surprisingly be as good as the performance of decoders that know the host signal. the attacker. NOVEL WATERMARKING SCHEMES 56 3.2. each coefficient is considered as an independent random variable with its own noise distribution. We consider an video as a channel with spatial-variant states. the capacity issue in a variant state channel.CHAPTER 3. We derive the capacity when that power and noise constraints are not uniform across sample. Here we first define the symbols that will be used in this section. In particular. Let X1 . Capacity expressions are derived under a parallel-Gaussian model for the host-image source. Capacity of the watermark is defined as how much information can be carried by the watermark when it is embedded in an image. [107] In this section.

5 in Elements of Information Theory [110].t. . The power constrain of these values are the asking bounds determined by the source coefficient values S1 . XN ]T Power constrain = E(XX T ) ≤ f (S) Noise = Z Y = SW − S = X − Z (3.. NOVEL WATERMARKING SCHEMES 57 discrete video frame due to watermarking.S = X . i. C = M axp(x):E(XX T )≤f (s) I(X. .12) (3.12. Y has zero mean and covariance K = E(XX T ). . I(. XN ]T and S = [S1 .14) where p(. .) represents the differential entropy. the maximum capacity of these multi-variant symbols in Equation 3.) represents any probability distribution.) represents mutual information and h(. . .e. the differential entropy of Y. In the receiver ed. h(Y) satisfies the following h(Y ) ≤ 1 log(2Π exp)n |K| 2 (3. . E(XXT ) ≤ f (S) where X = [X1 .CHAPTER 3. . X2 . consider Y = SW .. Capacity = C Host data = S = [S1 . We define a masking function f s. According to Theorem 9. S2 . We can assume X and Z are independent.6. S2 . SN ]T Watermark = X = [X1 . X2 . SN ]T . Y ) = M axp(x) [h(Y ) − h(Y |X)] = M axp(x) [h(Y ) − h(Z)] (3. SN .11) Then.15) . . .Z where Z are the noises added to the coefficients during transmission.. S2 .13) (3.

18) − log(2Π exp)n |E(ZZ T )| 2 1 (3. |E(XX T )+ E(ZZ T )| ≤ |f (s) + E(ZZ T )|.16) 2 where we assume f(S) is diagonal and nonnegative s. If we look at Equation 3. NOVEL WATERMARKING SCHEMES 58 with equality iff Y ∼ N (0.16 is the watermarking capacity in a variant-state channel without specifying any type of noise.21) σn i=1 2 where Pi and Ni are the power constrains in the ith coefficient.16 and Theorem 9. This assumption means that embedded watermark values are mutually independent. we can find that C will be at least 1 (3. this theorem is valid no matter what the range of K is. Equation 3. If we further assume that noise are also independent in samples. Therefore.12 and |E(XX T ) + E(ZZ T )| = |f (s) + E(ZZ T )|.| is the absolute value of the determinant. Here.17) log(2Π exp)n |f (S) + E(ZZ T )| Cmin = 2 1 (3. then the watermarking capacity will be: n 1 Pi log(1 + ) (3.CHAPTER 3. for all type of noise.6. K) and |. It is he capacity given a noise distribution.19) = |f (S) + E(ZZ T )−1 + I| 2 When noise is Gaussian distribution.t. It is inserting that even though we use the multi- .5 in [110] again. respectively.20) Cmin = Ni i=1 2 n 1 Pi = log(1 + 2 ) (3. we can see that 1 C = log(2Π exp)n |f (S) + E(XX T )| − h(Z) (3. from 3.

NOVEL WATERMARKING SCHEMES 59 variants to derive 3.CHAPTER 3.20 instead of using Parallel Gaussian Channels. their results are the same in this special case. .

a novel scene-based watermarking scheme is proposed. Error correcting codes are extracted from the video watermark and embedded as audio watermark in the audio stream. and statistical analysis. hybrid with different watermarking schemes can be one .3 A Hybrid Watermarking Scheme In the previous section. which carry extra information for watermark extraction. The scene-based watermarking scheme can be improved with two type of hybrid approaches: visual-audio hybrid watermarking and hybrid with different watermarking schemes. NOVEL WATERMARKING SCHEMES 60 3. The hybrid approach with different watermarking schemes can further be divided into to two classes: independent scheme and dependent scheme. we propose a hybrid approach to improve the performance and the robustness of the watermarking scheme based on the conclusion drawn from the survey and the properties of a video. including frame averaging. This approach takes the advantage of watermarking the audio channel. Figure 3. the scheme is more robust than other schemes which only use video channel alone.7 shows the overall framework of the proposed scheme. which is resistant to the attacks of the video properties. However. The visual-audio hybrid watermarking scheme applies both video and audio watermarks in a video. Therefore. we find that no watermarking scheme can resist all watermark attacks. From the survey.CHAPTER 3. because it provides an independent means for embedding the error correcting codes. the scheme dose not improve the robustness against the attacks by image processing on the video frames. frame dropping. Therefore.

CHAPTER 3.7: Possible improvement for scene based watermarking scheme of the solutions. In . NOVEL WATERMARKING SCHEMES 61 Scene based Watermarking Hybrid Approach Visual-audio Hybrid Watermark Hybrid with Different Watermark Independent Watermark Dependent Watermark Figure 3.1 Visual-audio Hybrid Watermarking The visual-audio watermarking scheme combines a video watermark and an audio watermark.3. It takes advantages of various watermarking schemes by combining them in different ways. 3. We embed error correcting codes of a video watermark as an audio watermark and refine the retrieved video watermark during detection [11].8 shows an overview of our visual-audio watermarking process. Figure 3.

C[i+4]) Watermarked Video IDWT DWT Normalization Scene Change Detection Video Stream DWT 62 . C[i+3]. C[i+2]. C[i+3]. C[i+2]. C[i+1].CHAPTER 3. NOVEL WATERMARKING SCHEMES Figure 3.8: Overview of visual-audio hybrid watermarking scheme Audio Stream Audio Watermarking Merger Key Error Correcting Code Extraction 8-bit planes m0 DWT Encryption Crop Watermark Bit Decomposition m0 ' Encrypted Watermark m7 ' DWT Encryption Encrypted Watermark m7 Video System Stream Splitter Place the bit-plane side by side Embedding Algorithm if W[j] = 1. C[i+4]) else Exchange C[i] with min(C[i]. C[i+1]. Exchange C[i] with max(C[i].

This additional protection mechanism enables our scheme to overcome the corruption of a video watermark. On the one hand. thus the robustness of the scheme is preserved under certain attacks. Disparate error correction coding techniques can be applied. our scheme benefits from audio watermarking as it provides an independent channel for embedding the error correcting codes. a video watermark is decomposed into various parts. an input video is split into audio and video streams. Moreover. Error correcting codes play an important role in watermarking. NOVEL WATERMARKING SCHEMES 63 our scheme. Error correcting codes overcome the corruption of a watermark. error correcting codes are extracted from the watermarks and embedded as an audio watermark in the audio channel. In the detection phase. which undergo separate watermarking procedures. which in turn makes it possible to correct and detect the changes from the extracted video watermarks. Audio Watermark The watermark embedded in the audio channel provides the error correction and detection capability for the video watermark. which carry extra information for video watermark extraction. and make the watermark survive through serious attacks. it would be extracted and used for refining the video watermark.CHAPTER 3. On the other hand. . especially when the watermark is damaged significantly. embedded in corresponding frames of different scenes in the original video. such as Reed-Solomon coding techniques [103] and Turbo coding [111].

k is the k th block of the average image. Hence. s) is coordinate of region k. Figure 3. (x. The simplest error correcting code is repeating everything several times. However. (r.CHAPTER 3.9: (a) Original video watermark (b) Visualization of averaging (c) Audio watermark (average of a) .22 x y i=0 j=0 Wj×z+r×x+s×y×z+i (3. A sample is shown in Figure 3. in order to keep the audio watermark inaudible. we cannot embed too much information into an audio channel. and x × y is the size of the block. Within a small region of an image. z is the width of the image.9. an average value of a small region can be fully utilized to estimate the pixels within the particular region.22) where W is a pixel in the image. In our scheme. the pixels are similar. we apply averaging to obtain the error correction code. NOVEL WATERMARKING SCHEMES 64 The key to error correction is redundancy. y) is the coordinate of the pixel in region k. The average value of the pixels in each region is calculated as Equation 3.

The MCLT is a 2x over sampled DFT filter bank. The MCLT is well suited for noise suppression and echo cancellation. For the typical 44. and passing those modified frames through the inverse MCLT.10 shows the detail. Figure 3. Only the MCLT coefficients within the 2-7 kHz subband are modified and considered in the detection process. The MCLT is based on the oddly-stacked time-domain aliasing cancellation (TDAC) filter bank introduced by Princen. . After addition of the watermark. Johnson. The magnitude then is modulated according to the prepared watermark in the previous section. to minimize carrier noise effects as well as sensitivity to downsampling and compression. After the wave is extracted from the audio channel. in the form [114]. we use a length-2048 MCLT. we generate the time-domain watermarked audio signal by combining the marked vector y with the original phase of x . used in conjunction with analysis and synthesis windows that provide perfect reconstruction.CHAPTER 3. the audio watermarking is based on Modulated Complex Lapped Transform (MCLT) [112]. Its basis functions can be obtained by cosine modulation of smooth windows. NOVEL WATERMARKING SCHEMES Audio Watermarking Embedding 65 In the audio channel.1 kHz sampling. and Bradley [113]. it is transformed from original time domain to frequency (MCLT) domain.

Figure 3. A test video is split into video stream and audio stream.CHAPTER 3. Figure 3. . Watermark detection The watermark is detected through the process whose overview is shown in Figure 3.12 shown the part of the original wave form.10: Audio watermark embedding with MCLT Watermarked Frame and Wave After applying the algorithm described in the previous sections on both video and audio channel. and Figure 3. the watermarking process for a video is completed. NOVEL WATERMARKING SCHEMES 66 Block of Time Domain Sample of the Original Content Block of Time Domain Sample of the Watermarked Content MCLT Phase IMCLT Magnitude Magnitude Block of Frequency (MCLT) Domain Sample of the Original Content Block of Frequency (MCLT) Domain Sample of the Watermarked Content Figure 3. and Figure 3. and watermarks are extracted separately by audio watermark extraction and video watermark extraction.13 shown the watermarked wave form.11 (b) shown the watermarked frame.11 (a) shown the one of the original frame.14.

In all the subsequent experiments.CHAPTER 3. Avgk is the extracted audio watermark. NOVEL WATERMARKING SCHEMES 67 (a) (b) Figure 3. At the same time.11: One of the (a) original video frame and (b) watermarked video frame Then the extracted watermarks undergo a refining process. we assume f = 0. The video stream is processed to get the video watermark.23) f +g where RWij is the refined watermark.5 and g = 0. EWij is the extracted video watermark from Equation (7). RWij = .23: EWij f + Avgk × g (3. (i. f + g = 1. error correcting codes are extracted from the audio stream and the video watermark extracted is refined by this information with the Equation 3.5. k is the k th block of the average image. and f : g is a ratio of importance of the extracted video watermark to the audio watermark. j) is coordinate of the video watermark.

2 −0.6 −0.CHAPTER 3.2 −0.4 −0.8 −1 0 500 1000 1500 2000 2500 3000 Figure 3.6 0.8 −1 0 500 1000 1500 2000 2500 3000 Figure 3.13: Block of samples of watermarked wave content .12: Block of samples of the original wave content 1 0.6 −0.8 0.2 0 −0.4 −0.6 0.2 0 −0. NOVEL WATERMARKING SCHEMES 68 1 0.8 0.4 0.4 0.

As stated earlier. Independent watermarking schemes include either different schemes for different scenes or different schemes for different parts of the frame. The hybrid approach can be a possible solution.14: Overview of detection of the watermark 3. (DCT). These four schemes are: Discrete Wavelet Transform based watermarking (DWT). Discrete Fourier Transform . Four watermarking schemes are chosen. We propose three approaches for the hybrid watermarking schemes. it can be classified into independent schemes and dependent schemes. Dependent watermarking schemes embed a watermark in each frame with several different schemes.3. Discrete Cosine Transform based watermarking.CHAPTER 3. They combine alien schemes in disparate ways. NOVEL WATERMARKING SCHEMES 69 Audio Stream Audio Watermark Extraction Video Splitter Watermark Refining Watermarked Video Video Stream Video Watermark Extraction Figure 3. each of which strives a different set of attacks.2 Hybrid Approach with Different Watermarking Schemes No watermarking scheme is found in the current literature to be capable of resisting all watermark attacks.

a watermark is still decomposed into different parts which are embedded in the corresponding frames of different scenes in the original video. The disadvantage of this approach is . As they embed the watermark in various domains. This approach thus enhances the chance of survival under several attacks. however. Within a scene. various kinds of attacks can be resisted altogether. different watermarking schemes are resistant against it. Figure 3. provided that at least one of the watermarking schemes is resistant against the attack. Independent Hybrid Watermarking In the independent hybrid watermarking approach. By combining the advantages of these watermarking schemes systematically. some parts of the watermark still survive after the attack. The merit is that only part of the watermark is damaged if the watermarked video is attacked. We propose two approaches to combine the employed watermarking schemes: Different schemes for different scenes. is embedded with a different watermarking scheme. all the video frames are watermarked with the same part of a watermark by the same watermarking scheme. Consequently. Different schemes for different scenes In this approach. NOVEL WATERMARKING SCHEMES 70 based watermarking (DFT). and different schemes for different parts of each frame. and raises its robustness. the applied watermarking schemes do not affect each other. and Radon Transform based watermarking (RADON).15 illustrates the idea. Each part of the watermark.CHAPTER 3. their robustness properties are preserved. When there is an attack on the watermarked video.

and the watermark can be approximately estimated. Therefore. NOVEL WATERMARKING SCHEMES 71 Figure 3. it is more resistant against attacks. each part of the watermark is embedded into the frame in different domains. Although the accuracy of the extracted watermark is reduced. The idea is illustrated in Figure 3. compared with the other schemes specified to a particular attack. where each video frame is divided into four parts. and the watermark for that frame is also divided into four parts.CHAPTER 3. four different watermarking schemes are applied to each frame instead of different schemes for different scenes. Different schemes for different parts of each frame This approach is similar to the previous approach.16. When a watermarked video is attacked.15: Hybrid approach with different scheme for different scene that the accuracy of the extracted watermark is lower. However. Then. part of the watermark in each frame may still survive. information for every part of the watermark can be retrieved. .

each frame is embedded with the same watermark serially by different watermarking schemes in various domains. . Experiments are done to evaluate the effectiveness of different approach and it is presented in Chapter 4.16: Hybrid approach with different scheme for different part of frame Dependent Hybrid Watermarking In the dependent hybrid watermarking approach. Nevertheless. The frame is first transformed to wavelet domain. As the watermarks are embedded in different domains. In our approach. again four different watermarking schemes are applied.CHAPTER 3. so on and so forth for DFT and Radon transform. and the watermark is embedded. NOVEL WATERMARKING SCHEMES 72 Figure 3. subsequently. different watermarking schemes may also affect each other when under attacks. they can compensate each other in resisting against different attacks. Then the frame is inversely transformed to frequency domain by DCT to embed the same watermark.

NOVEL WATERMARKING SCHEMES 73 3. In the previous section. However.17: The graph of three mutually orthogonal axes representing the capacity. since these requirements are conflicting with each other. higher fidelity.CHAPTER 3. and larger data capacity is three goals that most watermarking scheme would like to achieve. Fidelity Robustness Capacity Figure 3. better robustness (watermark strength) and larger data capacity. robustness and fidelity of the watermarking scheme In general.4 A Genetic Algorithm-based Video Watermarking Scheme The problem of designing a feasible watermarking scheme can be viewed as an optimization problem with three conflicting goals: higher fidelity (media quality index). In this section. we discussed applying a hybrid watermark approach to improve the robustness of the scheme. designing an optimal watermarking scheme . better robustness. genetic algorithm (GA) is employed to enhance the fidelity [115].

17.18: The graph of two mutually orthogonal axes representing the robustness and fidelity of the watermarking scheme has become an inherently difficult and interesting problem. which consequently constraints the robustness of a watermarking scheme against common or malicious manipulations. the fidelity requirement often limits the strength of embedded signals. the process of de- . in most watermarking schemes. However. In other words. taking the applicationspecific characteristics into consideration makes reasonable tradeoffs among the three requirements [115]. NOVEL WATERMARKING SCHEMES Robustness Non-achievable region 74 Achievable Fidelity Figure 3. respectively as shown in Figure 3.CHAPTER 3. In the existing watermarking schemes. For instant. one can view the embedding process as a selection of feasible embedding points within an acceptable region in the space spanned by three mutually orthogonal axes representing the prescribed requirements.

According to the problem specific constraints. Chromosomes with higher fitness values have higher probability to contribute more offspring in the next generation.1 Watermarking Scheme In the area of evolutionary computation.4. a population of candidate initialized.18 3. The objective function decides the probability of chromosomes’ survival or removal during the competition. NOVEL WATERMARKING SCHEMES 75 ciding the trade-off is always done in a heuristic manner. which is a finite-length strings. crossover. named as chromosome. the fitness function. Reproduction is a process in which individual chromosomes are copied according to their objective function values. that is. GA is an important optimization technique [115]. lacking systematic explorations or even optimizations. To model the problem as a GA problem. and mutation are applied to the chromosomes repeatedly. which optimize the performance towards the non-achievable region as show in Figure 3. In this section. In GA-based optimizations. During practical GA-based optimization processes. chromosome and GA operators should be defined. three GA operators: reproduction. a GA-based watermark fidelity enhancing method is proposed. the fitness values. Crossover is the procedure that pairs of chromosomes exchange portions of their genes to form new chromo- .CHAPTER 3. the problem to be addressed is defined as an objective function that indicates the fitness of any possible solution. Here we employ this techniques to optimize the performance of our proposed scheme.

NOVEL WATERMARKING SCHEMES 76 somes. Mutation is the occasional random alternation of the value in some positions of chromosomes. the chromosomes are optimized.4. only the chromosomes with the higher fitness values can survive. This portion of those will be pass as parent chromosomes to the next generation. After numbers of generation. chromosomes with good performance but not obtainable by reproduction and crossover may also have the chance to be selected. the watermarked video quality is improved while still keep the robustness of the watermark against image manipulation.2 Problem Modelling By applying the GA to the scene-based video watermarking scheme. The embedding position of the different parts of watermark within the video are simulated as chromosomes. These operators are used repeatedly to obtain successive generations of chromosomes. We choose the watermark embedding positions within a video . By sparingly using mutation. 3. and consequently. we use image quality indicators. Then several genetic-algorithmic operations are applied to optimize the video frame quality after embedding. Mutation can be regarded as a random walk through the parameter space. [115]. Mean Absolute Difference (MAD) to measure the objective function values during optimization. In our experiments. [116] Within a generation. new parameters other than the initial ones can be produced for evaluation.CHAPTER 3. We can obtain the near-optimal solutions of modelled problem.

20 shows the details of the GA-based optimization process for each video scene. A illustrative diagram is shown in the Figure 3.e. we can get near-optimal embedding positions for the original frame and the watermark images. therefore. . Assume a video is consists 8 scenes. The optimal combination of watermark and scene are shown in the chromosome. search a scene that a particular part of watermark is best to embed to. The basic idea is that the human eyes are sensitive to the low frequency noise and the quantization step of lossy compression may discard the high frequency components.19: A illustrative diagram for GA-based optimization process as our search space. NOVEL WATERMARKING SCHEMES 77 Figure 3.19. Then apply the geneticalgorithmic operators to find the best combination of watermark and video scene. The watermark is scrambled into 8 parts and embedded into different scenes. Figure 3. the input video frames are transformed to frequency domain and the middle-frequency range coefficients of the watermark are modified according to the watermark. In the watermark scheme. i. By repeatedly applying the GA operations to each original video frame and watermark image.CHAPTER 3..

20: The GA-based optimization process for part of watermark . NOVEL WATERMARKING SCHEMES 78 Figure 3.CHAPTER 3.

the search space of the GA is 1 and no optimization can be done. otherwise. 2. a 24-bit chromosome represents the sequence of video scene to be used. i. In a video. 3.e. the video frames that have been embedded should not be embedded again.CHAPTER 3.4. The scene of the ith watermark is best to be embedded to Xi in the video can be defined as: {(Xi ) | 0 ≤ Xi < M.21 shows the sample chromosome which represents the positions of the watermark that should be embedded to. Figure 3. (3. There are at least two scene-changes in the video. NOVEL WATERMARKING SCHEMES 79 the reasonable trade-off is to embed the watermark into the middle-frequency range of the video frames.24) . Xi = Xj if i = j and M > 1} The last two constraints of Equation 3. therefore. There are 8 scene changes in the video and 3-bit number need to be used to represent the position of video scenes.24 imply: 1. watermarks are embedded by modifying the middle-frequency coefficients within each frame block of the original frame. The number of scene in the video is M and we can user log2 M bit to represent the position of the scene.3 Chromosome Encoding Assume we want to embed a part of the watermark image into one of the scene in a video. The position of the scenes are encode as a chromosome.

crossover.21: A 24-bit chromosome represents the sequence of the scenes to embed 3.e.4 Genetic Operators The initial chromosomes are generated with the above specification. After calculation of the fitness function value (defined in Equation 3. and mutation. Then. y) − I(x. i. respectively. The basic idea is that the higher a parent chromosome’s . The definition of fitness function f will be: f= 1 (x. the genetic-algorithmic operators are apply to the chromosomes for optimization with the fitness function.4.CHAPTER 3. y)| (3. NOVEL WATERMARKING SCHEMES 80 100111000110001011010101 X0 X1 X2 X3 X4 X5 X6 X7 Figure 3. This application of GA aims to improve the fidelity. the reciprocal of image similarity indicators (maximum absolute different MAD) is chosen to be the fitness function. we employ the biased-roulette-wheel method [refGA] in GA to generate N children. the image quality.25) 7 x=0 7 y=0 |I Where I and I are the intensity values of the same pixel position within a video frame after and before embedding.25) for each parent chromosome. of the video frames. Fitness function The fitness function f is a measure of profit we need during optimization. they are: reproduction.

i. otherwise. Mutation The last operator is mutation. Reproduction Reproduction is a process that the children chromosomes are generated according to the fitness function values of their parents.e. . the higher probability it has to contribute one or more offspring in the next generation. Then an integer position p between 1 and the chromosome length (L) minus 1 is selected at random for each pair where p ≤ L − 1. Mutation introduces some randomness into the optimization. Finally. NOVEL WATERMARKING SCHEMES 81 fitness function value is. we use 0. Mutation is the random change of bit values with small probability within a chromosome. new combination of watermark and video scene are generated. In our experiment. crossover operator is applied to on these children chromosomes.CHAPTER 3. thus. the recently reproduced chromosomes are randomly mated in a pair. the optimization process is very slow. Firstly. Crossover After N legal parent chromosomes and M children chromosomes are generated. This increases the chance of approaching the optimal.05. a bit may change polarity (take complement) at a probability of 0. each chromosome swaps all the bits between the location of p+1 and chromosome length with its mate.05 as the mutation rate.

Figure 3. Firstly. we can get M legal children chromosomes (M ≤ N ).22 depicts the block diagram of the proposed watermarking algorithm with GA optimization. the scene change analysis is applied to the video and the watermark is preprocessed. NOVEL WATERMARKING SCHEMES 82 Figure 3. i. select N chromosomes with the larger fitness function values to be the next parent generations. After the checking and discarding process. more children chromosomes may not satisfy the constraints of embedding positions. Repeat all the operations mentioned above until the number of generations specified has been done. The original video and the watermark are input to the system. to be the sequence of video scenes to be used. Then.e.22: The GA-based watermarking algorithm After performing all the GA operations.CHAPTER 3. the one with the largest fitness function value. the GA-optimization process will find out the almost-optimal combination of water- . From the M + N chromosomes (M children and N parents). and choose the best chromosome.

and the the watermarked video frame with GA-based watermarking scheme are shown in Figure 3. . the enhancement can be indicated by measuring with the quality index and this experimental results will be shown in the following section. the watermarked video frame with scene-based watermarking scheme. 2 End of chapter. NOVEL WATERMARKING SCHEMES 83 mark and video scene. the watermark is extracted with the GA information provided by the embedding phase.CHAPTER 3. The original video frame. For detection phase. the GA information should be passed to the detector. The watermarks are embedded in to the video frames according to the result of the GA optimization. The watermarked video is passed in to the detection system. There is no significant perceptual different between in Figure 3.23 (b) and (c). However.23.

23: Comparison between watermarked video with and without GA optimization a) Original video frame (b) Video frame watermarked with scene-based scheme (c) Video frame watermarked with GA-based scheme . NOVEL WATERMARKING SCHEMES 84 (a) (b) (c) Figure 3.CHAPTER 3.

4. the hybrid watermarking scheme and the GA-based watermarking scheme. the experiment with various quality factor of MPEG. the experiment with various number of frame colluded. we present the implementation detail of the proposed schemes and the experimental results. To implement the proposed watermarking scheme. the software VirtualDub [117] is employed. which 85 .0.Chapter 4 Experimental Results In this chapter. the robustness of the scene-based watermarking scheme and the hybrid watermarking scheme is tested. Another DWT-based watermarking scheme. The performance of the new video watermarking scheme is evaluated through several experiments: the experiment with various dropping ratio. and the test of Robustness with StirMark 4. we present our experimental results on the scenebased watermarking scheme. In the following sections. The experiments are basically divided into 2 types: test on robustness and test on fidelity.1 Test on Robustness In this section.

The NC values are retrieved when the watermarked video is facing different attacks. The video consists of 10 scene changes. EXPERIMENTAL RESULTS 86 embeds an identical watermark in all frames. is implemented to compare with the proposed scheme. After extracting and refining the watermarks. where Wij is the original watermark and RWij is the refined watermark from Equation 3. and StirMark 4.1) which is the cross-correlation normalized by the reference watermark energy giving unity as the peak correlation [118]. frame averaging. are carried out to the watermarked video to test the robustness of our scheme. The audio channel is also attacked by adding some noises into it. Normalized Correlation: NC = i j Wij × RWij 2 i j Wij (4.23. . We use this measurement to evaluate our scheme in the experiments.00GHz and 512MB RAM. Distinguishable attacks. The experimental results are described in detail in the following sections.0 [14]. a similarity measurement of the extracted and the referenced watermarks can be defined as Equation 4. a quantitative measurement is required to provide an objective judgment of the extraction fidelity. including frame dropping. The experiments are done on a desktop computer with Pentium 4 CPU 2.CHAPTER 4.1. A video clip with 1526 frames of size 352 × 288 is used in our experiments. Therefore. lossy compression.

If they try to remove one part of the watermark. 120].1 Experiment with Frame Dropping As a video contains a large amount of redundancies between frames.8 0. Different percentages of frames are dropped and the obtained results are shown in Figure 4.1: NC values under frame dropping From the experiment. we find that the scheme achieves better performance than the DWT-based scheme without scene-based watermarks [119. it may suffer attacks by frame dropping.55 0. all frames are embedded with the same watermark. they need to remove . EXPERIMENTAL RESULTS 87 4.85 Normalized Correlation 0. This experiment is aimed at examining the robustness of the scheme under the frame dropping attack. It prevents the attackers from removing the watermark by frame dropping.CHAPTER 4.1.1.7 0. It is because in each scene.65 DWT−based watermarking scheme Scene based watermarking scheme Visual−audio hybrid watermarking scheme Visual−audio scheme with audio attack Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark 0 10 20 30 40 50 % of Frame Dropped 60 70 80 0.5 Figure 4.6 0.95 0.9 0. NC Values under Frame Dropping 1 0.75 0.

Consequently. the result is still better than the scheme without an audio watermark. leading to a significant damage to the video. . as the attacked audio watermark still contains some information to recover the watermark in the video channel. non-scene-based). EXPERIMENTAL RESULTS 88 the whole trunk of frames (i. our scheme benefits by allowing uninterrupted error correcting codes to refine the watermark and improve the NC values. the error correcting codes are embedded in the audio channel. the visual-audio watermarking scheme. the error is introduced to the whole watermark. which significantly damages the watermark. when the frames are dropped. Due to the increased dropping rate..CHAPTER 4. In addition. the NC values of the extracted watermark are decreased. the error is only introduced to a corresponding small part of the watermark. especially when the dropping rate of video frame is high. the error of the extracted watermark is increased. The error correcting codes from the audio watermark provide information to correct the error and recover the corruption of the video watermark. Although the capability to recover the error in the video watermark is dropped. The performance of the scheme is significantly improved by combining with an audio watermark. As frame dropping would not affect the audio channel much. making the performance worse. When the audio channel is also attacked.e. however. the NC values of the watermark are higher than these without the error correcting codes. Moreover.. The error correcting codes in the audio channel are altered by the attack.e. the whole scene). For the DWT-based scheme (i.

The identical watermark used within a scene can prevent attackers from taking the advantage of motionless regions in successive frames and removing the watermark by comparing and averaging the frames statisti- . making the watermarks resistant to attacks by frame averaging for the watermark extraction. 122]. When attackers collect a number of watermarked frames.CHAPTER 4. Y1 Noise Estimation W1 Video Frames Y2 Noise Estimation W2 + W - + Y Yn Noise Estimation Wn Frames to be attacked Figure 4. and the results are shown in Figure 4. This is because our scheme crops a watermark into pieces and embeds them into different frames. EXPERIMENTAL RESULTS 89 4. It is found that the proposed scheme can resist to statistical averaging quite well. The scenario is shown in Figure 4.2: Scenario of statistical averaging attack Experiments have been conducted to evaluate the proposed scheme under this attack.3.2 Experiment with Frame Averaging and Statistical Analysis Frame averaging and statistical analysis is another common attack to the video watermark.2.1. they can estimate the watermark by statistical averaging and remove it from the watermarked video [121.

65 0.1.75 0.6 0. NC values under statistical averaging 1 Proposed Watermarkng Scheme DWT−based Watermarking Scheme 0.4 shows the NC values of the extracted watermarks with different quality factors of MPEG. independent watermarks used for successive.7 0. From the experiment.55 0. Figure 4. yet different scenes can prevent the attackers from colluding with frames from completely different scenes to extract the watermark. On the other hand.3 Experiment with Lossy Compression This experiment is aimed at testing the robustness of the scheme under attack by lossy compression.95 0. The perfor- .5 0 100 200 300 400 500 600 Number of Frame Colluded 700 800 900 1000 Figure 4.8 0.9 0. EXPERIMENTAL RESULTS 90 cally [123].85 Normalized Correlation 0.CHAPTER 4. we note that the proposed scheme improves the robustness for watermark protection.3: NC values under statistical averaging 4.

5 0.65 0.95 DWT−based watermarking scheme Scene based watermarking scheme Visual−audio hybrid watermarking scheme Visual−audio scheme with audio attack Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark 91 0. As the error correcting codes are provided from the audio watermark. they are not affected by the lossy compression attack applied to the video channel.4: NC values under lossy compression mance of the scheme is significantly improved by combining with audio watermark again.9 0. achiving higher NC values. the error correcting codes can overcome the corruption of the video watermark.2 0.CHAPTER 4. Consequently.1 0.7 0. The proposed scheme without audio watermark has similar performance with the other DWT-based scheme because both .6 0.9 1 Figure 4.8 Normalized Correlation 0.4 0.5 0 0.3 0. This is because when the quality factor of MPEG is low.7 0.75 0.8 0.6 MPEG Quality Factor 0. EXPERIMENTAL RESULTS NC Values under Lossy Compression 0. especially when the quality factor of MPEG is low.55 0.85 0. the error of the extracted watermark is increased and the watermark is damaged significantly.

we employ Stir- .4 Test of Robustness with StirMark 4.1.0 StirMark 4. EXPERIMENTAL RESULTS 92 of them satisfy the following condition. therefore.CHAPTER 4. Therefore. the high frequency components) of the image [124]. we find that the DCT-based watermarking scheme is the most resistant one against lossy compression.0 [100. the watermark embedded in the video with DCT-based watermarking scheme survives. This increases the robustness of the scheme. From the survey. The NC values of the hybrid approach under lossy compression are higher than the scheme which only applying single watermarking scheme in a video. MPEG is based on DCT and lossy compression mostly only modifies the highest frequency coefficients. The performance of the scheme is also improved by the hybrid approach with different watermarking schemes. 101] is a benchmark for testing robustness of a watermarking scheme. However. In this experiment. When compression is applied to the watermarked video. as lossy compression removes the details (i. at least one forth of the watermark can be retrieved from the video. 4. modifying the mid-band coefficients in this domain during watermarking a video would reduce the effect of compression on the watermark. the robustness of the scheme is not improved by the hybrid approach with dependent watermark. Higher frequency DWT coefficients of the watermark are embedded in the higher frequency part of the video frame and high frequency sub-band DWT coefficients (HH) of video frame are not watermarked.e. This approach makes the watermark survive MPEG lossy compression.

However. including cropping.0 to test the robustness of the proposed schemes when image processing is applied.5. adding noise. The NC values of the watermarks are decreased with the PSNR.6. and compare them with the current techniques that exist in the literature.1 show the result of the watermarked video under different attacks from StirMark. 4. EXPERIMENTAL RESULTS 93 Mark 4. Peak Signal to Noise Ratio (PSNR). the NC values of the retrieved watermarks are less than 0. row/column removal. The proposed scheme reveals improvement in this experiment as well. gives better performance. The NC values of the DWT-based watermarking scheme slightly higher than the hybrid approach. but other watermarking schemes are af- . rotation and affine. on the other hand. median filter. The NC values of the hybrid approaches are similar to those of the DWT-based watermarking scheme. rescaling.CHAPTER 4. When the cropping ratio is greater than 60 present. The visual-audio watermarking scheme. Figure 4.9 shows the result of the watermarking scheme under different PSNR. The performance of the watermarking schemes without audio watermark decreases greatly with the increasing cropping ratio. Figure 4. Figures 4. 4. The NC values are greater than 0.7 even when the cropping ratio is 90 present.5(d) shows the result of the NC values of the watermark under different cropping ratios. the result shows that the performance of scheme is not improved by the hybrid approach with different watermarking schemes. This shows that the audio watermark significantly improves the robustness of the watermarking scheme.8. It is because the watermark in the wavelet domain is not affected by PSNR too much.7 and Table 4.

This is because when the rescaling factor increases. the improvement becomes more evident with the increase of the rescaling factor. The performance of the scheme is significantly improved again by visual-audio watermarking scheme.7 0. EXPERIMENTAL RESULTS 1 DWT−based watermarking scheme Scene based watermarking scheme Visual−audio hybrid watermarking scheme Visual−audio scheme with audio attack Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark 94 0. the error of the extracted watermark is increased.55 0. the overall performance the hybrid approach is decreased. Figure 4. This is one of the disadvantages of the hybrid approaches.5 10 20 30 40 50 Cropping Ratio 60 70 80 90 Figure 4. especially when the rescaling factor is large.75 0. the proposed scheme also portrays improvement. The error correcting codes from the audio water- . which significantly damages the watermark.8 0.CHAPTER 4.65 0.95 0.85 Normalized Correlation 0.9 0. Thus.5: NC values under cropping fected. When the watermarked video is rescaled. Furthermore.6 0.7 depicts the NC values when the watermarked video is rescaled with various factors.

In wavelet domain.55 0.6 0. on the other hand. The hybrid approach with dependent watermark.85 Normalized Correlation 0.CHAPTER 4. the robustness of the scheme is increased. EXPERIMENTAL RESULTS NC values under PSNR 1 95 0.6: NC values under PSNR mark. do not vary too much when rescaling. provide information to correct the error and overcome part of the corruption in the video watermark and produces higher NC values of the recovered watermark. It is . As shown in Figure 4.95 0.8 0. The coefficients in the Randon transformed domain.65 0.9 0.5 0 10 20 30 40 50 PSNR 60 70 80 90 100 Figure 4. nevertheless. performs worse than other hybrid approaches. With the hybrid approaches. the coefficients vary when the size of frame or image is different. The NC values of the extracted watermark are higher than the DWT-based and scene-based watermarking schemes.7.7 DWT−based watermarking scheme Scene based watermarking scheme Visual−audio hybrid watermarking scheme Visual−audio scheme with audio attack Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark 0. the hybrid approaches with different schemes perform better than the scene-based watermarking scheme. however.75 0.

CHAPTER 4. EXPERIMENTAL RESULTS

96

NC values under rescaling 1

0.95

0.9

0.85 Normalized Correlation

0.8

0.75

0.7

0.65

0.6

0.55

DWT−based watermarking scheme Scene based watermarking scheme Visual−audio hybrid watermarking scheme Visual−audio scheme with audio attack Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark

0.5 50

75

90

100

110 Rescaling factor

150

200

Figure 4.7: NC values under different rescaling factor

CHAPTER 4. EXPERIMENTAL RESULTS
NC values under different noise added to the watermarked video 1 DWT−based watermarking scheme Scene based watermarking scheme Visual−audio hybrid watermarking scheme Visual−audio scheme with audio attack Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark

97

0.95

0.9

0.85 Normalized Correlation

0.8

0.75

0.7

0.65

0.6

0.55

0.5

0

10

20

30

40

50 Noise

60

70

80

90

100

Figure 4.8: NC values under different noise added to the watermarked video

caused by the dependent properties of the watermarking scheme. When one watermark is corrupted, it may affect the extraction of other watermarks in the same frame. Therefore, the NC values of the extracted watermark are lower than those of other hybrid approaches. When additional noise is added to the watermarked video, the proposed scheme also shows improvement. Figure 4.8 depicts the NC values when different noises are added to the watermarked video. The performance of the scheme is significantly improved by combining with an audio watermark, especially when more noises are added. There are several tests from the StirMark 4.0. The result is summarized in Table 4.1. The proposed video watermarking

CHAPTER 4. EXPERIMENTAL RESULTS

98

scheme also shows improvement when the videos are under other attacks, including: row removal, rotation, PSNR, affine. . . etc.
Attack Class DWT-based watermarking scheme Scene-based watermarking scheme Visual-audio hybrid watermarking scheme Visual-audio hybrid watermarking scheme with audio attack Lossy Compression PSNR Add Noise Median Filter Row / Column Removal Cropping Rescale Rotation Affine Overall 0.61 0.72 0.63 0.54 0.69 0.68 0.63 0.60 0.55 0.62 0.62 0.76 0.60 0.54 0.71 0.66 0.62 0.61 0.55 0.63 0.82 0.86 0.76 0.74 0.85 0.78 0.75 0.73 0.78 0.78 0.69 0.80 0.67 0.60 0.75 0.70 0.69 0.67 0.70 0.69

Attack Class

Hybrid approach with different scheme for different scene

Hybrid approach with different scheme for different part of frame 0.72 0.81 0.69 0.52 0.78 0.69 0.68 0.66 0.71 0.70

Hybrid approach with dependent watermark 0.68 0.81 0.64 0.52 0.74 0.67 0.63 0.64 0.63 0.66

Lossy Compression PSNR Add Noise Median Filter Row / Column Removal Cropping Rescale Rotation Affine Overall

0.71 0.82 0.70 0.55 0.77 0.72 0.71 0.69 0.73 0.71

Table 4.1: Robustness comparison between different watermarking schemes

4.1.5

Overall Comparison

From the above results, the effectiveness of the scene-based hybrid schemes are demonstrated. The scene-based watermarking scheme achieves higher NC values when attacks based on video properties are launched. This indicates that the water-

such as the attack by lossy compression. however. still shows improvement. The performance of the scheme is further improved by combining with an audio watermark. EXPERIMENTAL RESULTS 99 marking scheme work well by applying scene change detection with scrambled watermarks. especially when the video watermark is corrupted. When audio channel is also attacked.CHAPTER 4. The robustness of the scheme is also raised by engaging other hybrid approaches. . The overall performance. the error correction information is altered.

y)| (4. The GA-based watermarking scheme is implemented with the GAlib [125]. the peak signal-to-noise ratio (PSNR) and maximum absolute difference (MAD) is used. The quality of the video watermarked by the scene-based watermarking scheme and the hybrid watermarking scheme are also evaluated. the quality of the video is evaluated with PSNR and MAD. The performance of the GA-based video watermarking scheme is evaluated through several experiments with different number of generation in the GA-optimization process.CHAPTER 4. MAD is given by 7 7 M AD = x=0 y=0 |I (x. In the experiment. two . we focus on evaluating the performance of the GA-based Watermarking Scheme. EXPERIMENTAL RESULTS 100 4. Then. respectively. we prove the fidelity enhancing effectiveness of the proposed optimization process. y) − I(x. PSNR is given by 2552 P SN R = 10 log10 2 [dB] σq (4.2 Test on Fidelity In this section. In the experiment. and compared with the GA-based scheme.2) 2 where σq is the mean square of the difference between the original video frame and the watermarked one [126]. To evaluate the fidelity of the watermarking scheme.3) Where I and I are the intensity values of the same pixel position within a video frame after and before embedding.

40.9 1 25 In the experiment. thus.we have used 0.CHAPTER 4.2.05 0.2: Parameters Setting for GA-based experiment Parameter Population size Mutation probability Crossover probability Score frequency Flush frequency Value 100 0. EXPERIMENTAL RESULTS 101 video clips are used. we can evaluate the its fidelity. From the graph 4. Number of generations 1. 4. 100. 10. 400 and 600 Table 4.2 Evaluate with PSNR PSNR measures the signal to noise ratio of the watermarked video. Table 4. 200. 4.2. it’s clear that the GA-based algorithm successfully reduces the video frame distortion due to watermark embedding. The experiments are done on a desktop computer with Pentium 4 CPU 2.3 shows the computation time of the GA-based scheme.2 shows the parameters setting for the GA-based video watermarking experiments. As .00GHz and 512MB RAM.1 Parameter(s) Setting Table 4. 5. 20. Another video clip has 4236 frames of size 352 × 288 and it consists of 22 scene changes. One of the video clips has 1526 frames of size 352×288 and it consists of 10 scene changes.9.

we can evaluate the quality of the watermarked video. it shows how the iterative generation numbers affect the optimizing performance. The PSNR of the video is reduced to 3/4 after GA is applied in optimizing the fidelity of the scheme. From the graph 4. It shows that the GA-based optimization effectively improves the performance of the scheme. thus.2.CHAPTER 4.3: The computation time of the GA-based scheme Number of Generation 0 5 10 20 40 100 200 400 600 Computation time Computation time of Video 1 (s) 50 62 80 103 124 156 185 226 259 of Video 2 (s) 141 179 234 289 354 423 534 620 702 the number of generations increases.10. Table 4.4 depicts a comparison of PSNR with different watermarking scheme.3 Evaluate with MAD MAD measures the different between the original video and the watermarked video. 4. EXPERIMENTAL RESULTS 102 Table 4. the improvement of video quality gradually approaches to a saturation value. The MAD of the watermarked video is decreased with the GA .

CHAPTER 4.9: PSNR of the video under different GA generations Table 4.4: PSNR comparison between different watermarking schemes Watermarking scheme Scene-based Watermarking scheme Visual-audio hybrid watermarking Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark GA-based watermarking scheme watermark scheme PSNR of Video 1 40 41 43 42 35 52 PSNR of Video 2 33 33 34 36 27 42 . EXPERIMENTAL RESULTS 103 PSNR of the video under different GA generation 70 Video 1 Video 2 65 60 55 50 PSNR (dB) 45 40 35 30 25 20 0 100 200 300 Generation number 400 500 600 Figure 4.

the performance of the proposed GA-based watermarking scheme is converged to optimal.5 3 2.CHAPTER 4. EXPERIMENTAL RESULTS x 10 4 104 MAD of the video under different GA generation 5. The optimization performance saturate after about 200 generations.5 MAD 4 3.10: MAD of the video under different GA generations generation number. Therefore. A comparison of MAD with different watermarking scheme is shown in Table 4.5. It shows the enhancing effectiveness of the GA-based watermarking scheme. The MAD of the video is reduced to 3/5 after GA is applied to optimize the fidelity of the scheme.5 5 Video 1 Video 2 4.5 0 100 200 300 Generation Number 400 500 600 Figure 4. .

Also. The computation time of the GAbased scheme is rather long if the number of GA generation applied is large or the number of the scene change of the video increases.3 Other Features of the Scheme Our proposed scheme is an invisible watermarking scheme. however. when the encoding method is applied again with another watermark. retrieval of the embedded watermarks does not need the original video. the watermark is perceptually invisible. Additionally.5: MAD comparison between different watermarking schemes Watermarking scheme Scene-based Watermarking scheme Visual-audio hybrid watermarking Hybrid approach with different scheme for different scene Hybrid approach with different scheme for different part of frame Hybrid approach with dependent watermark GA-based watermarking scheme watermark scheme MAD of Video 1 42168 42189 43984 43798 48652 28346 MAD of Video 2 50325 50489 52695 52786 55785 31546 4. This is an important performance feature of the scheme since it takes a long time to transmit. however. there are still some weaknesses in our scheme. the proposed scheme is not robust against . In the scene-based watermarking scheme. the perceptual quality will be affected [123]. a blind watermarking scheme. EXPERIMENTAL RESULTS 105 Table 4. and process the original video. If these coefficients are altered. store. The experiments show that the proposed scheme is robust to most of the existing attacks.CHAPTER 4. as low frequency subband DWT coefficients (LL) are not watermarked and image energy is concentrated on the lower frequency wavelet coefficients.e. i.

When there are two orthogonal axes. robustness and fidelity. EXPERIMENTAL RESULTS 106 it efficiently. The original watermark can only be extracted if the second watermark is removed first.CHAPTER 4. robustness and fidelity. we prove that our proposed scheme enhances two of the three prescribed watermarking requirements.11: A conceptual illustration on the performance of the proposed scheme Figure 4. and the fidelity can . The robustness is indicated by the strength of the watermark. 4. Robustness Achievable region Non-achievable region A A A Fidelity Figure 4.11 shows a conceptual illustration.4 Conclusion From the experiment. The robustness-enhancement provided by hybrid scene-based watermarking scheme and the fidelityenhancement provided by the GA-based watermarking scheme are important steps toward a prefect watermarking scheme.

our proposed scheme is approaching to the ”optimal” embedding configuration. Thus. we try to move the point towards the curve. the robustness of the scheme is improved and moves the point to A’. When the watermarking scheme is further enhanced by GA. 2 End of chapter. it moves towards A”.CHAPTER 4. i.e. By applying the hybrid approaches. Therefore. The point A represents the performance of a scene-based watermarking scheme. . increase the fidelity the scheme. EXPERIMENTAL RESULTS 107 be represented by quality index. the curve represent the best robustness under the fidelity performance constrain. To optimize the performance of a watermarking scheme. the point moves along the fidelity axes. Moreover.

Experiments are conducted to demonstrate that our scheme is robust against attacks by frame dropping. watermark embedding. (2) 108 . image watermarking and video watermarking. error correcting code and GA optimization is proposed. frame averaging. Our approach cultivates an innovative idea in: (1) embedding different parts of a watermark according to scene changes. and the robustness against the image processing attacks is tested with StirMark benchmark. After noticing the importance of the multimedia security and video watermarking in nowadays Internet world and reviewing the state-of-the-arts technologies of the audio watermarking.Chapter 5 Conclusion This thesis investigates the knowledge of digital video watermarking techniques for secure multimedia creation and delivery. the fidelity of the scheme is evaluated. The process of this comprehensive video watermarking scheme. and watermark detection. an innovative hybrid digital video watermarking scheme with scene change analysis. and statistical analysis. including watermark preprocessing. Moreover. video preprocessing. is described in detail.

etc • We propose a visual-audio hybrid watermarking scheme. we contribute on the followings: • We have performed a complete survey on the current watermarking technologies. . • We propose a hybrid approach with different watermarking schemes. of the watermarking scheme. The effectiveness of this scheme is verified through a number of experiments. By employing GA. We employ the hybrid scheme to embed different parts of a watermark into different scenes. • We propose a scene-based watermarking scheme. • We propose a GA-based watermarking scheme increase the fidelity. and its advantages are clear and significant. This approach is never explored in the literature. the media quality index.CHAPTER 5. CONCLUSION 109 embedding its error correcting codes as an audio watermark. (3) applying a hybrid approach to the proposed scheme. There are different ways to embed the watermarks. we can optimize the combination of the watermark and the scenes in the video. frame swapping. The robustness of our scheme can be enhanced by including an audio watermark.e. i. We embed error correcting codes of a video watermark as an audio watermark and refine the retrieved watermark during watermark detection. To conclude our work. The scheme is robust against frame averaging. frame dropping. statistical analysis. and (4) employing the GA algorithm to enhance the fidelity.

CHAPTER 5. The robustness of our approach is demonstrated using the criteria of the latest StirMark test. . 2 End of chapter. CONCLUSION 110 • Experiment has been done on these novel video watermarking schemes to test and show its performance.

and H. 2. 3. 2001. Vol. Barni. Liao. pp. Liao. 209-224. Vol. Dec. Piva. 111 . ”A survey of watermarking techniques applied to multimedia. ”Cocktail watermarking for digital image protection. ”Managing copyright in open networks. Vol. F. Y. [3] C. ”A Robust Watermarking Approach for Raw Video. pp. A. Cagliari. 1. Italy. 1579-1592. Caldelli. S.” IEEE Transactions on Internet Computing. Huang. R. Issue. 10. pp.” IEEE Transactions on Multimedia.Bibliography [1] A. and M. May 1-2 2000.” Proceedings 2001 IEEE International Symposium on Industrial Electronics (ISIE2001). Vol. Sze. ”Multipurpose Watermarking for Image Authentication and Protection. May-June 2002. 2001. Lu. Barni.” Proceedings 10th International Packet Video Workshop PV2000. Lu.” IEEE Transactions on Image Processing. Yuan. 18-26. Oct. H. and M. De Rosa. C. Bartolini. Issue. Bartolini. pp. Piva. and A. Jung. F. Lee and S. 2000. 272-277. M. [5] M. [2] C. 6. 10. [4] J.

Oct. pp. Petitcolas (Eds)”Information hiding techniques for steganography and digital watermarking Stefan Katzenbeisser. pp. ”A DWT-based Digital Video Watermarking Scheme with Error Correcting Code. B.W.” Proceedings Fifth International Conference on Information and Communications Security (ICICS2003). ”An overview of multimedia content protection in consumer electronics devices. Petitcolas. Chan and M. and A. Zhu. Springer. Herrigel and J. Inner-Mongolia. Dec. 1997. Tewfik. [10] M. 10-13. 328-331. . Dec. 3-Volume Set-Volume 2. [9] A. ”Multiresolution Video Watermarking using Perceptual Models and Scene Segmentation. China.” Proceedings Signal Processing Image Communication 16 (2001). [8] N. ISBN 1-58053035-4. DC.” Processing in Workshop on Information Hiding. ”Watermarking schemes evaluation tool. Vol. [7] A. Ruanaidh. Washington. 1999. LNCS. Delp. Oct. 2836.BIBLIOGRAPHY 112 [6] F.” Proceedings of IEEE Multimedia Software Engineering. 681-699. 26-29. 2003. Eskicioglu and E. Taiwan. Springer Verlag. Swanson. Lyu. 2001. [11] P. ”Secure Copyright Protection Techniques for Digital Images. 202-213. 2000.” Proceedings International Conference on Image Processing (ICIP ’97).” Artech House Books. Huhehaote City. pp. Lecture Notes in Computer Science. Fates and F.

BIBLIOGRAPHY

113

[12] P.W. Chan, M.R. Lyu, R. Chin, ”Copyright Protection on the Web: A Hybrid Digital Video Watermarking Scheme,” Poster Proceedings 13th International World Wide Web Conference (WWW’2004), pp. 354-355, New York, May 1722, 2004. [13] P.W. Chan, M.R. Lyu and R.T. Chin ”A Novel Scheme for Hybrid Digital Video Watermarking: Approach, Evaluation and Experimentation,” submitted to IEEE Transactions on Circuits and Systems for Video Technology. [14] F. Petitcolas, M. Frederic Raynal, J. Dittmann, C. Fontaine, N. Fates, ”A public automated web-based evaluation service for watermarking schemes: StirMark Benchmark,” In Ping Wah Wong and Edward J. Delp, editors, proceedings of electronic imaging, security and watermarking of multimedia contents III, Vol. 4314, San Jose, California, U.S.A., Jan. 20V26, 2001. The Society for imaging science and technology (I.S. and T.) and the international Society for optical engineering (SPIE). ISSN 0277-786X, http://www.petitcolas.net/fabien/watermarking/stirmark/. [15] K. Su, D. Kundur, and D. Hatzinakos, ”Statistical Invisibility for Collusion-resistant Digital Video Watermarking,” IEEE Transactions on Multimedia 2004 [16] K. Su, D. Kundur, and D. Hatzinakos, ”Spatially Localized Image-dependent Watermarking for Statistical Invisibility and Collusion Resistance,” IEEE Transactions on Multimedia 2004

BIBLIOGRAPHY

114

[17] K. Su, ”Digital Video Watermarking Principles for Resistance to Collusion and Interpolation Attacks,” Master of Applied Science thesis, University of Toronto, Sept. 2001. [18] I. Cox, M. Miller, and J. Bloom, ”Digital watermarking,” Morgan Kaufmann Publishers, Oct. 2001, ISBN 1-55860714-5. [19] F. Litterio, http://world.std.com/ franl/crypto/. [20] P. Johnston, http://pajhome.org.uk/crypt/rsa/intro.html. [21] Watermarking World, http://www.watermarkingworld.org/. [22] M. Kutter and F. Hartung, ”Introduction to Watermarking Techniques,” Proceedings Information Techniques for Steganography and Digital Watermarking, S.C. Katzenbeisser et al., Eds. Northwood, MA: Artec House, pp. 97119, Dec. 1999. [23] H. Inoue, A. Miyazaki, and T. Katsura ”An Image Watermarking Method Based on the Wavelet Transform”, Kyushu Multimedia System Research Laboratory. [24] I. Cox, M. Miller, J. Linnartz, and T. Kalker, ”A Review of Watermarking Principles and Practices” Proceedings Digital Signal Processing for Multimedia Systems, K.K. Parhi, T. Nishitani, eds., New York, New York, Marcel Dekker, Inc., pp. 461-482, 1999. [25] F. Petitcolas, ”Watermarking Schemes Evaluation”, IEEE Signal Processing Magazine, Vol. 17, pp. 58-64, Sept. 2000.

BIBLIOGRAPHY

115

[26] Y. Kim, K. Moon, and I. Oh, ”A text watermarking algorithm based on word classification and inter-word space statistics,” Proceedings Seventh International Conference on Document Analysis and Recognition 2003, pp. 775 -779, Aug. 3-6, 2003 . [27] R. Wolfgang, C. Podilchuk, and E. Delp, ”Perceptual Watermarks for Digital Images and Video,” Proceedings of the SPIE/IS and T International Conference on Security and Watermarking of Multimedia Contents, Vol. 3657, pp. 4051, San Jose, CA, Jan. 25 - 27, 1999. [28] R. Wolfgang and E. Delp, ”A Watermarking Technique for Digital Imagery: Further Studies,” Proceedings of the International Conference on Imaging Science, Systems, and Technology, pp. 279-287, Las Vegas, Jun. 30 - Jul. 3, 1997. [29] D. Sachs, R. Anand, and K. Ramchandran, ”Wireless image transmission using multiple-description based concatenated codes,” Proceedings Data Compression Conference DCC 2000, pp. 569, 2000. [30] A. Doufexi, A. Nix, D. Bull, ”Robust wireless image transmission using jointly-optimized modulation and source coding,” Proceedings IEEE 51st Vehicular Technology Conference 2000, VTC 2000-Spring Vol. 3, pp. 2039-2043, Tokyo, 2000. [31] M. Buckley, M. Ramos, S. Hemami, and S. Wicker, ”Perceptually-based robust image transmission over wire-

Dubey. Chhugani. J. pp. Foo. Aggarwal. [33] L. Bassali. 2000. 91-94. Vol. Lu. Tewfik.” Proceedings 15th International Conference on Pattern Recognition 2000. 128-131. pp. 473-480. T. [35] S. pp. M. D. ”Compression tolerant watermarking for image verification. 1996. ”Multipurpose audio watermarking. Craver. 2001. 223-226. ”What can we reasonably expect from watermarks?.” Proceedings 2000 International Conference on Image Processing. S. and B. pp. 3. M. 1.” Proceedings IEEE Workshop on the Applications of Signal Processing to Audio and Acoustics 2001. and P. Sun ”Digital audio watermarking and its application in multimedia database. Huang. 509-513. Chen. 2. 1999. A. J. Wu.” Proceedings IEEE Region 10 International Conference on Electrical and Electronic Technology 2001. and L. Yeo. Vol. Xu. 17-23. Vol. Jun. pp. [36] C. Liao. ”Digital watermarks for audio signals. Boney. [32] H. A. . pp. 2001. ”An adaptive audio watermarking system.” Proceedings the Fifth International Symposium on Signal Processing and Its Applications ISSPA’99. pp. Liu. 282-285.” Proceedings 2000 International Conference on Image Processing. Vol. and K. Vol. [34] C. [37] S. TENCON. 1.BIBLIOGRAPHY 116 less channels. 2. 2000. Wu and Q.” Proceedings Third IEEE International Conference on Multimedia Computing and Systems. 2000. Agarwal. Hamdy. 430-433.

2000. and S. 18-22 Jun.” Proceedings 2001 International Conference on Image Processing. ICC 95. [43] C. J. Mar. M. 95-98. ”Issues and Solutions for Authenticating MPEG Video. 2001. Wu. Lu and M. [42] C. ”Robust video watermarking for wireless multimedia communications . 1999. Seattle. Kaabneh and A. 1999. ”Video image transmission via mobile satellite channels. Basagni. 1995.” Proceedings IEEE International Conference on Communications. WCNC.BIBLIOGRAPHY 117 [38] K. and Q. 1530-1535. 2. 3.” Proceedings International Conference on Distributed Computing Systems Workshop. [39] C. Vol. 15-19. 483-486. . 2001. pp.” Proceedings IEEE International Conference on Acoustics. Vol.” Proceedings IEEEE Wireless Communications and Networking Confernce 2000. 352-356. Sun. Blostein. Xu. 1. ”Video object-based watermarking: a rotation and flipping resilient scheme. 2000. 1. pp. ”Muteness-based audio watermarking technique. Apr. Vol. Vol. and S. Chang. Liu. pp. [40] X. 379-383. Bartolini. Speach and Signal Processing. 2001. Youssef. F. pp. ”A robust digital audio watermarking technique” Proceedings the Fifth International Symposium on Signal Processing and Its Applications ISSPA ’99. 54-65. Liao. Lin and S. Wang. Barni. pp. Checcacci. pp. [41] N. Y.

1999. Girod.3. Delp. pp. 2-5. pp. San Jose. and Gatweys. Texa. Jan. C. May 1998. Kalker. 3657. [46] R. Vol. Nov. Jan. ”Overview of Image Security Techniques with Applications in Multimedia Systems. Mapelli. Wolfgang. Proceedings International Conference onImage Processing. Furon. pp. and E. 40-51. Vol. 25-27. Tubaro. IEEE Transaction Signal Processing. Displays. and P. ”Perceptual Watermarks for Digital Images and Video.” Processing . 2000. and S. Delp. [47] R. Delp. Wolfgang and E. Vol. Vol. 3228. Lin. C. [48] T. San Jose. Podilchuk. Lancini. 21-24. [49] R. ”A robust video watermarking technique in the spatial domain. ”Watermarking of Uncompressed and Compressed Video”. 297-308. Duhamel. no.BIBLIOGRAPHY 118 [44] E. 1997. Vol. Podilchuk. Terminals. T.” Proceedings the SPIE/IST International Conference on Security and Watermarking of Multimedia Contents. CA. 4314. 3 (Special issue on Watermarking). ”Robustness of asymmetric watermarking technique”. ”Streaming Video and Rate Scalable Compression: What Are the Challenges for Watermarking?. 283-301. [45] F. 66. CA. 2225. and E. pp.” Proceedings the SPIE Conference on Multimedia Networks: Security. F. 2001. Hartung and B.” Proceedings the SPIE International Conference on Security and Watermarking of Multimedia Contents III. Dallas.

” Proceedings. Steinberg. pp. ”Transparent robust image watermarking. 45. Akansu. 1999. 10-13.” IEEE Transactions on Multimedia. Baldoza. Lee and M. Q. pp. M. Chen. Issue. 1. Ramkumar and A. Vol. 16-19. pp. [52] D. Tian. [54] M. [51] P. .” IEEE Transactions on Consumer Electronics. ”Robust Protocols for Proving Ownership of Image. 814-817. Tewfik. 851 -859. [50] M.” Proceedings International Conference on Image Processing. ”New fragile authentication watermark for images. 211-214. [53] J. 3. 20-25. ”A semi-fragile object based video authentication system. Sept. Sept. Vol. 2002. pp. Aug. Fridrich. 1996. 2000. Zhu. Swanson. 1996. May 25-28 2003. 2000 International Conference on Image Processing. ”Robust error concealment algorithm for video decoder. pp. 5-7. B. Vol. and A.BIBLIOGRAPHY 119 and Multimedia Communications 4th EURASIP-IEEE Region 8 International Symposium on Video/Image VIPromCom. Goljan and A.” Proceedings the 2002 IEEE Information Theory Workshop. 446-449. Vol. [55] Y.3. 3. pp. 2002. Sun and Q. 251-256.” Proceedings of the 2003 International Symposium on Circuits and Systems ISCAS ’03. He. Oct. ”Watermarking identification for private and public users: the broadcast channel approach.

2000. Rosa and A. Jun.” IEEE International Conference on Multimedia Computing and Systems. Mar.-C. Sept. 2002. pp. J. Vol. pp. Vol. 448451. P. Lausanne. ”Image watermarking using DCT domain constraints. 46 Issue. ”On random coding error exponents of watermarking systems. Ozawa. Chen and T.” IEEE Transactions on Image Processing. 2. and S. Delp. ”A watermark for digital images. [57] T. 219-222. An. Aug. 07-11.” IEEE Transactions on Information Theory. [58] R. pp. ”A novel semi-private watermarking technique. Chen. Carnegie Mellon University Technical Report: AMP01-03. pp. Onishi. Vol. Wolfgang and E. 420-430. Bors and I. 8.” Proceedings International Conference on Image Processing 1996. A. Italy.” Proceedings International Conference on Images Processing. 1996. Pitas. [60] N. Sept.BIBLIOGRAPHY 120 [56] P. ”The Digital Watermarking Method by Using both Patchwork and DCT. 1999. ”Capacity of full frame DCT image watermarks. Piva. [59] H. pp.” Proceedings IEEE International Symposium on Circuits and Systems 2002 (ISCAS 2002). May 26-29. 16-19. Switzerland. Bartolini. F. ”A Framework for Optimal Public Watermark Detection”. Vol. 2000. 3. Barni. 231-234. Wong and O. . 2. Kii. [61] M. [62] A. 1996. Vol. 9 Issue. Merhav. 1450-1455. Florence. 1.

” Proceedings The 2001 IEEE International Symposium on Circuits and Systems.BIBLIOGRAPHY 121 [63] I. No. 1946-1950.” Proceedings 9th IEEE Digital Signal Processing Workshop. Kim and S.” Proceedings IEEE International Conference on Acoustics. Sun. 2001. ISCAS 2001. Cheng and T. 2002. pp. Vol. 2. 4. I. and R. USA. [68] Q. 5. ”A blind watermarking technique using wavelet transform. Zeng. Ko. Wang. ”Optimum detection and decoding of multiplicative watermarks in DFT domain. 11. [66] Y. ”A wavelet-based watermarking algorithm for ownership verification of digital image. [64] X. Speech. Hong. Feb. 2001. Leung and B. 2. [67] P. 2002.” Seventh International Conference on Image Processing and Its Applications 1999. Premaratne and C. 34773480. 2001. pp. 2000. Vol. and Signal Processing (ICASSP ’02) 2002. May 13-17. [65] K. J. 235-238. 3. Han.” Proceedings IEEE International Symposium on Industrial Electronics ISIE 2001. ”A New Wavelet-Based Digital Watermarking for Video. Huang. Doherty. 13-15. 1999. Niu and S. Vol.” IEEE Transactions on Image Processing. Dyck. pp. Oct. Jul. Vol. Texas. ”A novel watermark embedding and detection scheme for images in DFT domain. . pp. ”Wavelet-based digital watermarking with halftoning technique. Vol. 780-783.

Suhail and M. [74] M.” Proceedings The 8th IEEE International Conference on Electronics. pp.” IEEE Transactions on Consumer Electronics. 2000 [73] A.BIBLIOGRAPHY 122 [69] J.” Proceedings 2000 International Conference on Image Processing. 221-240. Arena and M. Qiao and K. 2001 (ICECS 2001). pp. pp. Dong-Ho Shin. Nahrstedt. ”Digital watermarking for copyright protection of MPEG2 compressed video. 2001. Vancouver. Vol. 2-5. [72] S. [71] T. Y oung-Nam Oh. Piva. Vol. Obaidat.” Proceedings International Conference on Image Processing (ICIP-2000). ”MPEG PTY-marks: Cheap detection of embedded copyright data in DVD video. 871-874. ”Watermarking method for mpeg encoded video: Towards resolving rightful ownership. pp. Canada. 3. pp. 1998. 438-441. Vol. 3. R.” Proceedings 5th European Symposium on Research in Computer Security.” IEEE Multimedia Computing and Systems. Circuits and Systems. De Rosa. ”Digital watermarking applied to MPEG2 coded video sequence exploiting space and frequency masking. [70] L. Linnartz and J. Caldelli and A. 895-901. 3. Caramma. ”On the digital watermarking in JPEG 2000. Issue. 44. . and Sang-Hui Park. 5-8. Sept. 2. Vol. 1998. ”A DWT-based object watermarking system for MPEG-4 video streams. 2000. Chung. August 1998. Talstra. M. Hong. Jun.

Kang. [77] J. ”A wavelet-based watermarking for digital images and video. Csurka. 1999. pp. ”Mediaindependent watermarking classification and the need for combining digital video and audio watermarking for media authentication. Steinebach. pp. 1. [79] F. Vol. Issue: 8. J.” Proceedings International Conference on Image Processing (ICIP-2000). Deguillaume. Steinebach. Huang. 301-304. Vol. 3657.” IEEE Transactions on Circuits and Systems for Video Technology. J. Miyazaki.” Processing 2001 IEEE Fourth Workshop on Multimedia Signal. pp. Fridrich. 776-786.” Proceedings the Fifth International Symposium onSignal Processing and Its Applications ISSPA ’99. ”A hybrid watermark for tamper detection in digital images. San Jose. Mukherje. 1999. and Y. and T. 678-681. Lin. 2001. Aug. ”Joint watermarking of audio-visual data. Dittmann and M. Vol. CA. Ruanaidh. ”Robust 3D DFT video watermarking. Dittmann. Canada. 601-606. 1999. [80] M. A. 62-67. 2000.” Proceedings International Conference on Information Technology: Coding and Computing 2000 pp.BIBLIOGRAPHY 123 [75] J. G. Pun. [78] X. Shi. and M. pp. ”A DWT-DFT composite watermarking scheme robust to both affine transform and JPEG compression. Ejima and A. 3. Jan. . 2003. Vancouver. [76] J. Vol. 2001. 13.” Proceedings Electronic Imaging’ 99: Security and Watermarking of Multimedia Contents. Y.

and G. Zhu. Johnson and S. Langelaar. [85] C. 283-301. ”Combating geometrical attacks in a dwt based blind video watermarking system”. pp. 2000. Girod. Proceedings Video/Image Processing and Multimedia Communications 4th EURASIP-IEEE Region 8 International Symposium on VIPromCom. and A. 43-75. Proceedings 9th IEEE Digital Signal Processing Workshop. ”Watermarking Digital Image and Video Data. I. [84] X. pp. Setyawan..” S. 20-43.” IEEE Signal Processing Magazine. Sun.” Proceedings Signal Processing. 66. pp. 1998. M. Vol 17. Lagendijk. Eds. 263-266. ”A New Wavelet-Based Digital Watermarking for Video”. pp. ”A Survey of Steganographic Techniques in Information Techniques for Steganography and Digital Watermarking. Vol. 26-29. R. 1999.” Proceedings International Conference on Image Processing (ICIP ’97). USA. MA: Artec House. Katezenbeisser. Hartung and B. Tomlinson. [82] N. Northwood. 2002. Katzenbeisser et al. Tewfik. [83] G. 2000. 2. M. Niu and S.C. 1997. ”Watermarking of uncompressed and compressed video. B. Serdean. . Vol. ”Multiresolution Video Watermarking using Perceptual Models and Scene Segmentation. Texas. Wade. Washington. Swanson. Ambroze.BIBLIOGRAPHY 124 [81] F. Sept. [86] D. Dec. Oct. Oct. DC.

pp. Yeo. A. Caldelli. 4-7. Cagliari. 284-295. 2000. B. [91] C. ”Direct sequence watermarking of digital video using m-frames. Italy. 1998. . Zhu. 483-486. [90] M. Chau. 3. and A. Tewfik. Holliman. pp. 1997. [89] B. Vol. and J. 399-403. B. 2. Chicago. Mobasseri.BIBLIOGRAPHY 125 [87] M. R.” Proceedings 2001 International Conference on Image Processing. and A. New Jersey. Piva. pp. 1997. Vol. USA. Bas. Yeung. ”A Robust Watermarking Approach for Raw Video”. Baudry. May 1-2.” Proceedings IEEE Signal Processing Society 1997 Workshop on Multimedia Signal Processing. Jun. Illinois. F. De Rosa. Bartolini. Vassaux. ”Video object-based watermarking: a rotation and flipping resilient scheme. Nguyen. Oct. Vol. Proceedings 10th International Packet Video Workshop PV2000. Swanson. Chassery. P. Liao. 2002. 3312. Princeton.” Proceedings SPIE. and M. [92] B. N. S. P. ”Scrambling technique for video object watermarking resisting to mpeg-4. pp. 239244. ”Adaptive public watermarking of DCT-based compressed images. Barni. [88] M. 2001. ”ObjectBased Transparent Video Watermarking. B. Memon.” Proceedings Video/Image Processing and Multimedia Communications 4th EURASIP-IEEE Region 8 International Symposium on VIPromCom. 23-25. Lu and M.” Proceedings International Conference on Image Processing (ICIP-98).

” Proceedings International Conference on Image Processing (ICIP-98). 1998. Vol. pp. [96] S.9 Issue. 17-20. 4-7. Oct. Aug. L. and S. Vol. George. 1019-1022. Vol. [97] I. I. 3. Kim. Memon. 4-7. Oct. 1589-1591. ” IEEE Transactions on Image Processing. 2001. pp. Mobasseri. Vol. 1123-1129. 1998.” Proceedings 2001 International Conference on Image. and L. Oct. Mobasseri. Chan. ”A blind watermarking technique using wavelet transform” Proceedings IEEE International Symposium on Industrial Electronics ISIE 2001. [94] B. Xu. Vol. pp. 399-403. Vol.BIBLIOGRAPHY 126 [93] N. 1946-1950. Chouinard and N. Chicago. 1998. Jun. Hong. Georganas. ”Digital watermarking of images and video using direct sequence spread . ”Analysis of LSB based image steganography techniques Chandramouli. 2. [98] F. ”Direct sequence watermarking of digital video using m-frames. Han. Illinois. I. pp. 2. 2001. 3. ”Intra-block algorithm for digital watermarking”. pp. Duan. 3. J.” Proceedings 1998 International Conference on Image Processing ICIP 98. Pun. 6. 7-10. [99] M. 2000. Proceedings IEEE 14th International Conference on Pattern Recognition (ICPR’98). King. ”Robust template matching for affine resistant image watermarks. [95] B. Pereira and T. ”Direct sequence watermarking of digital video using m-frames. pp. 399403.

3. pp. ”Robust video watermarking for wireless multimedia communications. Security and Watermarking of Multimedia Contents. [105] F. 1999. Vol. and S. Chan. King. 3657.” Proceedings 2000 IEEE Wireless Communications and Networking Conference (WCNC 2000). [101] M. 7-11. 2000. Chau.2000. [100] F. Vol. VTS-Fall VTC. 1198-1203. 1999. ”Image encryption for secure Internet multimedia applications.” IEEE Transactions on Consumer Electronics. Petitcolas. pp. Kutter and F. 2000. Checcacci. Cao. Italy. May 9-12. Basagni. 395-403. ”Intra-block maxmin algorithm for embedding robust digital watermark into . Barni. pp. Vol. Dang and P. 46. Issue. Jun. Bartolini. ”Evaluation of Copyright Marking Systems. Vol. Z.” Proceedings 1999 IEEE Canadian Conference on Electrical and Computer Engineering. pp. 1530-1535. [103] L. and L. Anderson. Aug. and C.BIBLIOGRAPHY 127 spectrum techniques. ”A Fair Benchmark for Image Watermarking Systems.” Proceedings 2000 52nd IEEE Vehicular Technology Conference. Gao. 2000. Duan. M. Vol. F.” Proceedings of Electronic Imaging ’99.” Proceedings IEEE multimedia Systems (ICMCS’99). pp. Zhang. [104] P. L. Petitcolas and R. Florence. 1. 3. 226-239. 3. [102] N. Xu. I. ”Application of RS-coded MPSK modulation scenarios to compressed image communication in mobile fading channel. 116 121.

Lecture Notes in Computer Science. 255-264. Dec. Ramkumar and A. pp. Lnc. 1973-1987. Ip and Arnold W. Germany. Berlin Heidelberg. J. Proceedings of the IAPR International Workshop on Multimedia Information Analysis and Retrieval. Jersey City. [106] J. Oct. Shamoon. Vol. [110] T. 2001. [111] A. pp. [107] A. Leighton. Cox.” John Wiley and Sones. ”The role of information theory in watermarking and its application to image watermarking. Smeulders.” Signal Processing 2001. NJIT. pp. ”Secure spread spectrum watermarking for multimedia. 1998. Vol. 1121-1139. Akansu. 1998. [108] M. MINAR’98. 1991. ”A Capacity for Data Hidig in Internet Multimedia. J. Servetto. F. Chicago. M. Stander. G.” IEEE Transactions on Image Processing.S. Kilian. Podilchu and K. [109] S.. C. Ambroze. 612. ”Capacity Issues in Digital Image Waterarmking.” Symposium on Content Security and Data Hiding in Digital Media.M. Borda.” IEE International Conference on Image Processing. Cover and J. Ramchadra. May 1999. Wade. and M. editors. Moulin. and T.” In Horace H.BIBLIOGRAPHY 128 images. Tomlinson. Vol. Thomas. SpringerVerlag. 1997. 81. 1464. ”Turbo code protection of video . ”Elements of Information Teory. Serdean. C.

Apr. 2000. Malvar. Princen. Huang and J.” Proceedings 1997 13th International Conference on Digital Signal Processing.virtualdub. San Jose.” Proceedings First Information Hiding Workshop. [113] J. Wu. [115] C. pp. Vol. Mar. Malvar. [112] D. Wu. [114] H. 217-220.” SPIE Electronic Imaging 2000. Hzu and J. ”Robust covert communication over a public audio channel using spread spectrum. 1. TX. 1997. Issue. [116] C. 1987. 1999. PA. A. http://www.” Proceedings IEEE ICASSP. 2001. [118] C. 21612164. 2-4. DSP 97. pp. Image and Signal Processing. pp. ”Using Genetic Algorithms as Watermarking Performance Optimizers.” Proceedings IEEE Vision. Wu. ”A watermark optimization technique based on genetic algorithms. Bradley. USA. written by Avery Lee. virtualdub. Johnson. Huang and J. Jul. and A.” Proceedings IEEE ICASSP.” IEEE Signal Processing Letters. Feb. Dallas. [117] VirtualDub is distributed under the GNU General Public License. Pittsburgh. Jan. ”Subband/ transform coding using filter bank designs based on time domain aliasing cancellation. Kirovski and H. 148. ”Digital watermarking for video.org/. . ”A Modulated Complex Lapped Transform and its Applications to Audio Processing.org / phaeron (at). 2001. 54-58.BIBLIOGRAPHY 129 watermark channel. Vol. 1.

” Proceedings Security and Watermarking of Multimedia Contents IV SPIE. Dyck. Su. ”A Novel Approach to Collusion-Resistant Video Watermarking. USA.BIBLIOGRAPHY 130 [119] X. W. [120] D. Delp and P. C. and R. ”A New Wavelet-Based Digital Watermarking for Video. Hatzinakos. 2002. Wolfgang. 440-444. California. and Signal Processing. ”A ContentDependent Spatially Localized Video Watermarked for Resistance to Collusion and Interpolation Attacks. Oct. Doherty.” Proceedings IEEE International Conference on Acoustics.” Proceedings IEEE International Conference on Image Processing. J. [123] Y. Su. Jan. 2001. Vol. . Chicago (IL. USA). Wang. Kirovski. Niu and S. No. Speech. ”Robust spread-spectrum audio watermarking.” Proceedings International Conference on Image Processing (ICIP-98). Malvar. Kundur and D. Vol. pp. Oct.” IEEE Transactions on Image Processing. Podilchuk. 2001. 4675. Texas. 1998. [121] K. Vol. Delp. 3. pp. J. 1. 1345-1348. D.. and E. ”The effect of matching watermark and compression transforms in compressed color images. E. and H. Sun. D. Kundur and D. 2002. 2.” 9th IEEE Digital Signal Processing Workshop. [122] K. Wong. ”A wavelet-based watermarking algorithm for ownership verification of digital image. pp. 12. Vol. 11. 2001. Hatzinakos. [124] R. 2000. San Jose. Oct. Feb. eds.

written by Matthew Wall at the Massachusetts Institute of Technology. . Gilmore. http://lancet.mit.BIBLIOGRAPHY 131 [125] The software for this work used the GAlib genetic algorithm package. Chouikha. USA.” Proceedings of the IASTED International Conference on Signal and Image. ”Dictionary Approaches To Image Compression and Reconstruction. [126] E.edu/ga/. Sept. M. 1998. 12-16. Las Vegas.

Sign up to vote on this title
UsefulNot useful