Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
DIGITAL WATERMARKING OF MULTIMEDIA DATA
THESIS SUBMITTED FOR THE AWARD OF THE DEGREE OF
Sottor of ^I)iIos(opf)p IN
ELECTRONICS ENGINEERING
BY
FAROOQ HUSAIN
Under the Supervision of
Dr. Omar Farooq Prof. Ekram Khan
DEPARTMENT OF ELECTRONICS ENGINEERING Z. H. COLLEGE OF ENGINEERING AND TECHNOLOGY
ALIGARH MUSLIM UNIVERSITY ALIGARH (INDIA)
2011
,' i 1
T8481
Phones Office: 2721148 Inter 2700920 Extn : 3125, 3126 Texlex: 564-230 AMU IN Fax 91-0571-2721148
DEPARTMENT OF ELECTRONICS ENGINEERING Z. H. COLLEGE OF ENGINEERING & TECHNOLOGY
ALIGARH MUSLIM UNIVERSITY, ALIGARH-202002, INDIA
Certificate
This is to certify that this thesis report entitled "DIGITAL WATERMARKING OF
MULTIMEDIA DATA" is submitted by Mr. Farooq Husain to the Department of
Electronics Engineering, Zakir Husain College of Engineering and Technology,
Aligarh Muslim University, Aligarh, India, for the award of the degree of Doctor of
Philosophy in Electronics Engineering. This is a record of the candidate's own work
carried out by him under our supervision and guidance. The matter embodied in this
thesis has not been submitted for the award of any other degree or diploma.
Dff Omar Farooq Prof. Ekram Khan Supervisor Co-Supervisor
ACKNOWLEDGEMENTS
Tirst and foremost, I wouCd [ik^ to tHan^^Dr. Omar Tarooq and^Prof.
'E^am Xfian, my supervisors, for their support, advice, and encouragement.
Tfieir exceptional scientific approach and experience have Seen an inspiration
to me in the years of my research wor^ ^Dr. Omar Tarooq and (prof 'El^ram
Kfian are not onfy academic mentors who advise me on research, 6ut a[so they
are great tutors with enthusiasm and patience to enlighten me on various
aspects. 'Without their generous support and assistance, this thesis wor^
would not have Seen what it is now.
I am highly indeSted to all the faculty memSers of (Department of
'Electronics Engineering for their academic and moral support during my
research at Jl.M.V. J4ligarh. I am very proud to have opportunities to wor/i^
with all these Sright minds during my stay for research atji. M. V. JAligarh. I
would also lih^ to thanh^allthe non-teaching staff of Elextronics Engineering
(Department for their nd and generous support during my (Ph.D.
I am extremely indeStedto my elder Brother (prof. Qayyum !Kusain, and
my sister-in-law (ShaShi)'Mumtaz (Begum for support and encouraging during
(ph.D. I am also thankful to my parents, my wife Shahnaz (Begum, my sons,
my nieces and my nephews for unfailing support.
I would li^ to than^to all the memSers of 94oradaSad Educational
'Trust (ME.f), !MoradaSad, and Director general, M I. T. MoradaSadfor
their support and encouragement during my research.
(Tarooq Jfusain)
TABLE OF CONTENTS
Page No.
Abstract i
Frequently used Abbreviations v
List of Figures viii
List of Tables xii
Chapter 1 Introduction 1
1.1 Overview 1
1.1.1 Applications of Watermarking 3
1.1.2 Requirements for Watermarking 4
1.2 Digital Watermarking of Multimedia Data 5
1.2.1 Text Watermarking 5
1.2.2 Audio Watermarking 6
1.2.3 Image Watermarking 6
1.2.4 Video Watermarking 6
1.3 Objectives of the Thesis 7
1.4 Outcome and Main Contributions of Thesis 7
1.5 Organization of Thesis 8
Chapter 2 Digital Watermarking Techniques 10
2.1 Introduction 10
2.2 Types of Watermarking 10
2.3 Attacks on Watermarks 11
2.3.1 Audio Watermark Attacks 12
2.3.2 Image Watermark Attacks 13
2.4 Overview of Transforms 14
2.4.1 Fourier Transform (FT) 14
2.4.2 Discrete Fourier Transform (DFT) 14
2.4.3 Short-Time Fourier Transform (STFT) 15
2.4.4 Discrete Cosine Transform (DCT) 16
2.4.5 Discrete Fractional Fourier Transform (DFRFT) 17
2.4.6 Discrete Wavelet Transform (DWT) 20
2.5 Digital Watermarking Techniques 21
2.5.1 Time-Domain/Spatial-Domain Watermarking 22
2.5.2 Transform-Domain Watermarking 25
2.6 DWT-based Kundur's Digital Image Watermarking Technique 28
Chapter 3 Digital Audio Watermarking 30
3.1 Introduction 30
3.2 Background 30
3.2.1 Chirp Signals 30
3.2.2 Related Works 33
3.3 Proposed Audio Watermarking Scheme 35
3.3.1 Watermark Embedding 35
3.3.2 Watermark Extraction 36
3.4 Single-Level and Multi-Level Watermarking 38
3.4.1 Three-Level Digital Audio Watermarking 38
3.5 Performance Measuring Parameters 40
3.5.1 Perceptual Transparency 40
3.5.2 The PEAQ Algorithm 41
3.5.3 Signal-to-Watermark Ratio (SWR) 42
3.5.4 Robustness against Audio Watermark Attacks 43
3.6 Simulation Results and Discussion 43
3.6.1 Robustness Evaluation for Single-Level and Multi-Level 47 Watermarking
3.7 Summary 67
Chapter 4 Digital Image Watermarking in DFRFT Domain 68
4.1 Introduction 68
4.2 Background 68
4.3 Digital Image watermarking using DFRFT 70
4.3.1 Watermark Embedding 71
4.3.2 Proposed Watermark Extraction 73
4.4 Simulation Results and Discussion 75
4.4.1 Performance Measuring Parameters 76
4.4.2 Imperceptibility Performance 77
4.4.3 Robustness of Watermarking Scheme 79
4.5 Summary 87
Chapter 5 Digital Image Watermarking using Combined Wavelet 88 and Fractional Fourier Transforms
5.1 Introduction 88
5.2 Background 88
5.3 Proposed Watermarking Scheme 89
5.3.1 Watermark Embedding 90
5.3.2 Watermark Extraction 94
5.4. Simulation Results and Discussion 95
5.5 Summary . 107
Chapter 6 Conclusions and Future Work 108
6.1 Conclusions 108
6.2 Future Work HO
References 111
Appendix-A: Test Images used for Watermarking 120
ABSTRACT
In the modern era of information technology, the use of multimedia to
exchange information is increasing day by day. The digital contents are transmitted
over wired/wireless networks and/or stored on storage media, for the purpose of
distribution and sharing of the information. However, these contents can be easily
replicated, modified and re-distributed by unauthorized third party. The ease with
which digital data can be copied and manipulated has generated a need for security
and authenticity of multimedia contents. Various techniques such as cryptography,
steganography and digital watermarking have been used in multimedia applications to
enhance the security. Cryptography deals with securing the multimedia contents by
encrypting them in a noise-like pattern using a secret key. Cryptography does not
completely solve the security problems because once the data decrypted, there is no
control on its dissemination. Steganography is the science of hiding the existence of
the message (information content) in another signal (referred to as a container or
dummy signal). The container signal is just a carrier of important information content.
Steganographic methods are not in general robust i.e., the hidden information can not
be recovered after data manipulation. Digital watermarking has been proposed as a
viable solution to improve multimedia security and to verify the authenticity of the
contents while offering the robustness against any attacks. In watermarking, the
additional information (watermark) is embedded into the digital contents
imperceptibly such that watermark may be detected/extracted later to make an
assertion about the multimedia data.
Watermarking has been proposed for various applications such as copyright
protection, copy control, broadcast monitoring, content authentication, etc. A
watermark scheme should have good imperceptibility; better robustness and higher
embedding capacity. A watermark is said to be imperceptible if the original host
signal and its watermarked version are perceptually indistinguishable to each other.
The property of resistance against distortion due to various signal processing
operations applied on watermarked multimedia signal is known as robustness and the
number of watermark bits embedded in a host signal without affecting its
imperceptibility is called the embedding capacity or payload.
Various algorithms for digital watermarking of multimedia data such as audio,
images and video have been proposed that involve embedding of watermark in the
spatial, transform and compressed domains. The properties of Human Auditory
System (HAS) and Human Visual System (HVS) have been explored in maximizing
embedding strength without causing significant perceptual distortion.
In this thesis, digital watermarking of multimedia data, mainly for audio and
images are presented. First, a digital audio watermarking scheme using chirp signal as
a watermark is presented, simulated and its performance is evaluated for single- and
multi-level embedding. In this thesis. Discrete Fractional Fourier Transform (DFRFT)
is used for watermarking of important information in digital images. First, an
extraction algorithm for DFRFT based watermarking scheme is proposed and its
performance is analyzed for various test images. Then, a novel watermarking
algorithm exploiting Discrete Wavelet Transform (DWT) together with DFRFT is
developed, in which watermark data has been embedded in high frequency sub-bands.
This thesis presents a digital audio watermarking scheme using chirp signal as
watermark. Three types of chirp signals (linear, quadratic and logarithmic) have been
studied. In this audio watermarking scheme, a property of HAS is considered in
selecting frequency ranges for chirp-based watermarks. The HAS shows low
sensitivity at frequencies less than 100 Hz. In this watermarking scheme, ranges of
frequencies for all three levels of chirp-based watermarks are below 100 Hz. This
watermarking scheme has been developed and simulated for single-level and multi
level watermark embedding. To obtain multi-level security in the audio data, multi
level chirp based audio watermarking is proposed. The performance of this scheme in
terms of Bit Error Rate (BER), Objective Difference Grade (ODG) using Perceptual
Evaluation of Audio Quality (PEAQ) model, and Signal-to-Watermark Ratio (SWR)
is evaluated under various audio watermark attacks for single-level and multi-level
watermark embedding schemes. This audio watermarking scheme is robust against
most of the audio watermark attacks such as low-pass filtering, interpolation,
decimation, re-sampling, amplitude scaling, addition of white Gaussian noise, MPS
compression. At the same time, it shows limited robustness against high-pass and
band-pass filtering for both single-level and multi-level watermarking schemes.
Multi-level chirp-based digital audio watermarking has different levels of robustness.
Hence, it can be used to embed information having different levels of security into a
host audio signal. By using multiple-level watermarking, the payload is increased.
The payload in three-level watemiarking is increased by 39.68 % of single-level
watermarking while maintaining imperceptibility of audio signal. ODG is a parameter
which is used to measure imperceptibility of watermarked audio signals. The average
ODG values of watermarked audio signals after one level, two levels and three levels
are -0.38, -0.46 and -0.48, respectively. These values of ODG show that
imperceptibility is maintained for single-level as well as for three-level audio
watermarking schemes.
For a DFRFT-based image watermarking scheme, a non-blind watermark
extraction scheme has been proposed. It has been assumed that the watermark is
embedded by either increasing (if watermark data is T ) the values of the pre-selected
set of contiguous DFRFT coefficients or leaving them unchanged (corresponding to
"C watermark data) of host images. The proposed algorithm is then used to extract the
watermark data by identifying the amount of changes in the same set of DFRFT
coefficients of watermarked image compared to the corresponding DFRFT
coefficients of host image. The performance of the proposed scheme has been
compared with a popular DWT-based Kundur's watermarking method. Performances
of both image watermarking techniques are measured in terms of Peak Signal-to-
Noise Ratio (PSNR) of watermarked images. It has been observed that for a set of five
test images, the average PSNR of watermarked image (for the same watermark data)
is about 51.24 dB where as that for DWT-based method is about 45.14 dB. Therefore,
the DFRFT based method is found to have better imperceptibility compared to DWT-
based Kundur's method. The performance of proposed extraction method is also
compared under various attacks and it has been observed that for attacks such as salt
& peppers noise, median filtering, Additive White Gaussian Noise (AWGN) and
JPEG compression, the proposed method has better robustness (measured in terms of
bit error rate of extracted data). However, for some attacks such as histogram
equalization and sharpening, the watermark extraction performance of DFRFT-based
method is found to be inferior in comparison to DWT-based Kundur's method.
In order to improve vratermark extraction performance under histogram
equalization and sharpening attacks, another image watermarking scheme has been
111
developed using a combination of DWT and DFRFT. The robustness of extraction
algorithm for histogram equaHzation and sharpening attacks can be enhanced by
exploiting the multi-resolution sub-band decomposition property of DWT and
applying DFRFT on a high frequency sub-band or a combination of high frequency
sub-bands of wavelet transformed host image. This scheme uses only one or two
levels of DWT 'decomposition for watermark embedding. Non-blind watermark
extraction scheme similar to the one described in previous paragraph is used to extract
embedded watermark. Average PSNR of watermarked image obtained using proposed
hybrid (combination of DWT and DFRFT) scheme is 51.35 dB, which shows
imperceptible watermarking of images. Simulation results show that for histogram
equalization attack, the BERs of extracted watermark for DWT and DFRFT-based
methods are obtained as 44.62-48.88% and 81.55-88.87%, respectively for the
considered set of images, whereas for combined DWT and DFRFT domain method,
the BER is 0-28.32%. Similarly, for sharpening attack, the BERs for DWT, DFRFT
and combined (DWT and DFRFT) methods are 32.56-47.57%, 70.32-77.93% and 0-
7.72% respectively. Therefore, the proposed hybrid domain watermarking method has
better robustness in comparison to watermark embedding in individual domain such
as DFRFT and DWT for histogram equalization and sharpening attacks.
IV
FREQUENTLY USED ABBREVIATIONS
ID-DFRFT : One-Dimensional Discrete Fractional Fourier Transform
2D-DFRFT : Two-Dimensional Discrete Fractional Fourier Transform
2D-DWT : Two-Dimensional Discrete Wavelet Transform
A/D : Analog-to-Digital
AWGN : Additive White Gaussian Noise
BER : Bit Error Rate
BPF : Band Pass Filter
BS : Broadcasting Services
CBR : Constant Bit Rate
CD : Compact Disc
D/A : Digital-to-Analog
DCT : Discrete Cosine Transform
DFRFT : Discrete Fractional Fourier Transform
DFT : Discrete Fourier Transform
DVD : Digital Versatile Disc
DWT : Discrete Wavelet Transform
EBU : European Broadcasting Union
FIR : Finite Impulse Response
FRFT ; Fractional Fourier Transform
FT : Fourier Transform
HAS : Human Auditory System
HF : High Frequency
HH : High-High frequency
HL : High-Low frequency
HPF : High Pass Filter
HRT : Hough Radon Transform
HVS : Human Visual System
IBM : International Business Machine
ITU ; International Telecommunication Union
ITU-R : ITU Radio communication
JND : Just Noticeable Distortion
JPEG : Joint Picture Experts Group
LH : Low-High frequency
LL : Low-Low frequency
LPF : Low Pass Filter
LSB : Least Significant Bit
MOS : Mean Opinion Score
MOV : Model Output Variable
MP3 : MPEG-1 Layer-3
MPEG : Motion Picture Experts Group
MSE : Mean Square Error
NMR : Noise-to-Mask Ratio
OCR : Optical Character Recognition
ODG : Objective Difference Grade
PDF : Portable Document Format
PEAQ : Perceptual Evaluation of Audio Quality
PSNR : Peak Signal-to-Noise Ratio
QF : Quality Factor
QIM : Quantization Index Modulation
RGB : Red-Green-Blue
RST : Rotation, Scaling, and Translation
SDG : Subjective Difference Grade
VI
SNR : Signal-to-Noise Ratio
SQAM : Speech Quality Assessment Material
STFT : Short-Time Fourier Transform
SWR : Signal-to-Watermark Ratio
SWRa : SWRof attacked watermarked audio signal with reference to host signal
SWRo : SWRof watermarked audio signal with reference to host audio
signal
TF : Time-Frequency
WWW : World Wide Web
XCF : Cross-Correlation Function
XOR : Exclusive OR
vu
LIST OF FIGURES
Figure No. Title Page No.
Illustration of (a) Cryptography (b) Watermarking „ Figure 1.1 ^ , . ,. j ' . ^ " ^ - ' ^ ° I
techniques applied on images
Block Diagram of Digital Watermarking (a) Watermark ^ ^ • Embedding (b) Watermark Detection
General overview of DWT-Based Kundur's Image ,Q ^ • Watermarking Technique
Spectrograms of chirp signals (a) Linear (0-100 Hz) (b) ^^ figure j . l Qy^^r^^j^ (Q.iQo Hz) (c) Logarithmic (10-100 Hz)
Figure 3.2 Waveformof a logarithmic chirp signal (10-100 Hz) 33
Figure 3.3 Block Diagram of Watermark Embedding Process 35
Figure 3.4 Block Diagram of Watermark Detection Process 37
„. ^ ^ Plot of cross-correlation function (XCF) vs. number of ~_ Figure J.5 ,. , 37
audio segments Waveforms of sections of (a) original audio signal (b)
„. - , First-Level watermarked audio signal (c) Second-Level .^ Figure 3 6 o \ / ^^
^ ' watermarked audio signal (d) Third-Level watermarked audio signal Waveforms of sections (a) original host audio signal (b)
Figure 3.7 chirp-based watermarked audio signal (c) difference 45 between two signals as shown in parts (a) and (b)
The variation of average BER in extracted watermark Figure 3.8 and 5'ffi?fl with increasing normalized cutoff frequency 48
«„/ after LPF attack for single-level watermarking
Watermark extraction performance in terms of BER Figure 3.9 with increasing normalized cutoff frequency c;„; for 48
low-pass filtered three-level watermarked audio signals
. ,^ Variation of iS' i fl with <»„, for low-pass filtered three-Figure 3.10 "'• 49
level watermarked audio signals.
Plots of average BER in extracted watermark and Figure 3.11 SWRa vs. .a*,, after HPF attack for single-level 50
watermarking
Vlll
Variation of average BER in extracted watermark with Figure 3.12 increasing(y^^^ after HPF attack for three-level 51
watermarked audio signals
Variation of average SWRa with increasing «„„ for
Figure 3.13 high-pass filtered three-level watermarked audio 51 signals
Variation of average BER in extracted watermark and Figure 3.14 SWRa with increasing co„^g after BPF attack for single- 52
level waterm£irking
Watermark detraction performance in terms of average Figure 3.15 BER with increasing &>„,;, after BPF attack for three- 53
level watermarking
^ Plot of average SWRa vs. (y„,„ after BPF attack for ^^ Figure 3.16 . 5 J
three-level watermarking
Variation of average BER in extracted watermark and Figure 3.17 SWRa with increasing interpolation factor A„ for up- 54
sampled single-level watermarked audio signals
Watermark extraction performance in terms of average Figure 3.18 BER with increasing interpolation factor A' ,, for up- 55
sampled tluee-level watermarked audio signals
Variation of average SWRa with increasing Figure 3.19 interpolation factor A„ after up-sampling of three-level 55
watermarked audio signals
Variation of BER in extracted watermark and SWRa Figure 3.20 with increasing decimation factor N^ for down- 56
sampled single-level watermarked audio signals
Watermark extraction performance in terms of average ^ ^, BER with increasing interpolation factor ^^ after Figure 3.21 c. -r 57
down-sampling of three-level watermarked audio signals
Variation of average SWRa with increasing decimation Figure 3.22 factor N^ for down-sampled three-level watermarked 57
audio signals
Variation of average BER in extracted watermark and Figure 3.23 5'ff7?a with increasing iV after re-sampling of single- 58
level watermarked audio signals
ix
Watermark extraction performance in terms of average Figure 3.24 BER with increasing N^ for re-sampled three-level 59
watermarked audio signals
^ ^^ Variation of average S'W i a with increasing iV, after re- ._ Figure 3.25 ,. r , , , , j ,• • > ^9
samplmg or three-level watermarked audio signals Variation of average BER and SWRa with increasing
Figure 3.26 SNR after adding white Gaussian noise of varying 60 power in single-level watermarked audio signals
Plot of the variation of average BER in extracted Figure 3.27 watermark with increasing SNR for attacked three-level 61
watermarked signals by AWGN
Variation of average SWRa with increasing SNR for Figure 3.28 attacked three-level watermarked audio signals by 61
AWGN
Variation of average BER and SWRa with increasing Figure 3.29 amplitude scaling factor A for amplitude-scaled single- 62
level watermarked audio signals
Watermark extraction performances in terms of average Figure 3.30 BER with increasing amplitude scaling factor A for 63
amplitude-scaled three-level watermarked audio signals
Variation of average SWRa with increasing amplitude Figure 3.31 scaling factor A for amplitude-scaled three-level 63
watermarked audio signal
Plots of variation of average BER and SWRa with Figure 3.32 increasing Constant Bit Rate (CBR) for MP3- 65
compressed single-level watermarked audio signals
Watermark extraction performances in terms of average Figure 3.33 BER with increasing CBR for MP3-compressed three- 65
level watermarked audio signals
Plot of variation of average SWRa with increasing CBR Figure 3.34 for MP3-compressed three-level watermarked audio 66
signals
Plot of variation of average BER in extracted . watermark and SWRa with increasing the number of
cropped samples of single-level watermarked audio signals
J.. , . Block Diagram of DFRFT-based Watermark Figure 4.1 f 72
Embedding Scheme
Figure 4.2 Block Diagram of Proposed Watermark Extraction 74
„. , _ Effects of Salt and Peppers Noise of density 0.01 on „„ Figure 4.3 . i . .• oO
^ watermark extraction
Figure 4.4 Effect of Median Filtering on watermark extraction 81
Effect of AWGN (with variance 0.0003) on watermark „-Figure 4.5 ^ . ^ ^ 82
° extraction
, , Effect of AV/GN (with varying variance) for 'Lena' „„ Figure 4.6 . \ J a ^ ^j.
° image
Figure 4.7 Effect of JPEG compression on watermark extraction 83
p. . g Effect of JPEG compression of varying QF on 'Lena' ^ . image
Figure 4.9 Effect of Low-Pass Filtering on watermark extraction 85
„. , .„ Effect of Histogram Equalization on watermark „ , Figure 4.10 . ° ^ 86
extraction
Figure 4.11 Effect of Sharpening attack on watermark extraction 87
^. , . Block diagram of watermark embedding for combined „„ Figure 5 1 90
DWT and DFRFT domain digital image watermarking
„. - „ Different sub-bands of an image for second-level DWT „^ Figure 5.2 , . . ^ 93
decomposition
^. ^ „ Block diagram of watermark extraction for combined „ . DWT and DFRFT domain digital Image watermarking
J.. . , Effect of Histogram Equalization on Watermark , „ , Figure 5.4 „ ^ . ^ ^ 106
Extraction
Figure 5.5 Effect of Sharpening on Watermark Extraction 106
,.. ^ , Effect of JPEG Compression (QF= 70%) on Watermark Figure 5.6 ^ , . 107
Extraction
XI
LIST OF TABLES
Table No.
Table 3.1
Table 3.2
Table 3.3
Table 4.1
Table 5.1
Table 5.2
Table 5.3
Table 5.4
Table 5,5
Table 5.6
Title
ODG values and its impairment description
Comparative average performance of chirp-based audio watermarking scheme using three types of chirp signals
Parameters and corresponding performance of Three-Level Watermarking
Comparison of DFRFT-based method with DWT-based method for fn'e host images
PSNR of watermarked 'Lena' image after embedding in various HF sub-bands
Performance Comparison of three watermark methods in terms of BER under different attacks for 'Lena' Image
Performance Comparison of three watermark methods in terms of BER under different attacks for 'Boats' Image
Performance Comparison of three watermark methods in terms of BER under different attacks for 'Goldhill' Image
Performance Comparison of three watermark methods in terms of BER under different attacks for 'Peppers' Image
Performance Comparison of three watermark methods in terms of BER under different attacks for 'Baboon' Image
Page No.
42
44
46
78
92
100
101
102
103
104
Xll
1
chapter 1 Introduction
CHapter 1
INTRODUCTION
The last two decades has seen remarkable advancement in digital
communication and computer technology. The advantages they offer are efficient
transmission of data, high computing power, availability of high bandwidth, easy
editing of digital contents, etc. Ii is now easy to copy digital content without any loss
in the quality of the digital media. The progress in this technology has also brought
some problems beside its advantages. The problem of ease to copy a digital content,
edit and transfer, it can violate copyright protection. Digital watermarking has been
proposed as a solution to prove ownership of digital data [1]. A watermark data (an
additional signal) is embedded into the original data (host signal) in such a way that it
remains present as long as the perceptible quality of the content of host signal is at an
acceptable level. The owner of the original data proves his/her ownership by
extracting the watermark data from the watermarked content (watermarked signal) in
case of multiple ownership claims.
1.1 Overview
The techniques used for data security are cryptography, steganography and
watermarking. Cryptography deals with securing contents of a message (host signal),
after encryption the message looks like noise and is useless. However, encryption
systems do not completely solve the security problem; because once decryption is
performed, there is no control on dissemination of the data [2], [3]. Hence, a
technique is required which always provide security to host signal. Steganography and
watermarking are regarded as the techniques which give protection to host signal all
the time. The basic idea of steganography is to embed a secret message in a media
which is interpreted by the intended user only. Steganography deals with concealing
the message, while watermarking has an additional requirement of securely
concealing the message, such that the message cannot be tampered. A good
steganographic system can embed large amount of information with no perceptual
degradation to the multimedia data, but a good watermarking system would embed
information that cannot be altered or removed without making the multimedia signal
unusable. A watermarking system involves trade off between imperceptibility,
robustness and embedding capacity [2]. It consists of two steps, viz. watermark
Cfiapter 1
embedding and watermark detection/extraction. Encryption and watermarking
techniques are shown in Figiire 1.1. Basic block diagram of watermark embedding
and detection is shown in Figure 1.2. Encryption key is a secret which is used in
cryptography to make hidden information more secure. The watermark data may be a
random number that is generated using a secret key which is known as seed of
watermark. Embedding is a process by which watermark data is fused with the
original signal (also known as host signal) to give watermarked signal. Detection
process checks the presence of watermark data in the watermarked signal but
extraction process finds the watermark data from the watermarked signal. User keys
are secret which are used for watermark embedding. These keys may be watermark
length, watermark location, watermark scaling factor etc.
Host Image
Encryption Key
Encryption
(a)
Encrypted Image
Watermark Data User key
Host Image (b)
Watermarked Image
Figure 1.1 Illustration of (a) Cryptography (b) Watermarking techniques applied on
images.
chapter 1
User key
Host signal ^
Watermark *
Watermark Embedding
Watermarked signal
User key
Watermarked signal
(a)
Watermark Detection
Detected watermark
(b)
Figure 1.2 Block Diagram of Digital Watermarking (a) Watermark E^mbedding (b)
Watermark Detection.
1.1.1 Applications of Watermarking
The main motivation behind the digital watermarking is copyright protection,
but it can also be used for content authentication, broadcast monitoring,
fingerprinting, and covert communication [4]-[7].
The application of watermarking for copyright protection is to evade other
parties from claiming the copyright of digital media. It is done by embedding
additional information (watermark data) that identifies the copyright owner of the
digital media. The content owner is recognized by his signature which can be
embedded as a watermark data.
Content authentication [4]-[5] is a detection process to ascertain whether the
contents of a digital media have changed. As a solution, a fragile watermark data [6]-
[7] embedded to the digital content (host signal) indicates whether the host signal has
been altered. If any tampering has occurred in the host signal, the change will also
occur in the recovered/detected watermark data. Content authentication can also
provide information about any part of the host signal that has been altered.
Cfiapterl
By embedding watermark, data into commercial advertisements, it can be
monitored whether the advertisements are broadcast at the correct instants by means
of an automated system [4]-[5]. The system receives the broadcast and searches these
watermark data, identifying where and when the advertisement has to be broadcast.
Fingerprinting is an approach to trace the source of illegal copies [4]-[5]. The
owner of the digital content (host signal) may embed different watermark data in the
copies of digital content customized for each recipient. In this manner, the owner can
identify the customer by extracting the watermark data in the case the host signal is
supplied to third parties.
Secret communication is another possible application of digital watemiarking
[4]-[5]. A watermark data can be embedded imperceptibly to a host signal to
communicate information from sender to an intended receiver while maintaining low
probability of intercept by other unimended receivers.
1.1.2 Requirements for Watermarking
The requirements for an effective watermarking scheme are imperceptibility,
robustness to intentional or unintentional signal processing operations and embedding
capacity. The efficiency of a digital watermarking process is evaluated according to
the properties of perceptual transparency (imperceptibility), robustness, embedding
capacity, computational cost, recovery of data with or without access to the original
hostsignaletc. [4], [5],[8]-[14][16].
Imperceptibility: A watermark is called imperceptible if the original host
signal and its watermarked version (watermarked signal) is perceptually
indistinguishable. The watermarks do not create visible artifacts in still images, alter
the bit rate of video or introduce audible artifacts in audio signals. Imperceptibility
for images and video signals is known as invisibility, but for audio and speech
signals, it is known as inaudibility. The imperceptibility of the watermark data is
tested by means of subjective experiments [15] such as Mean Opinion Score (MOS)
measurement through listening tests for audio and objective tests such as Signal-to-
Noise Ratio (SNR), and Objective Difference Grade (ODG) measurements. The
Cfiapter 1
imperceptibility of images can be measured by using objective tests such as Peak
SNR (PSNR) and Mean Square Error (MSE).
Robustness: Robustness to signal processing operations refers to the ability to
detect/extract the watermark data, after the watermarked signal has passed through
various signal processing operations. Robustness of a watermarking scheme can vary
from one signal processing operation to another. Depending on the application, the
digital watermarking technique can support different levels of robustness against
changes made to the watermarked content. If digital watermarking is used for
copyright owner identification, then the watermark data has to be robust against any
modifications.
Embedding capacity: The number of watermark bits embedded in a host
signal without affecting its imperceptibility is known as embedding capacity. The
embedding capacity is also known as payload.
1.2 Digital Watermarking of Multimedia Data
Digital watermarking is a technology that is used for the protection of any
multimedia data such as text, audio, images and video. In this section we briefly
discuss the watermarking methodology for each of these contents briefly.
1.2.1 Text Watermarking
Much of the early watermarking works were done in the area of text
watermarking [17]-[19]. These techniques were devised for watermarking electronic
versions of text documents such as portable document format (PDF) versions. Most of
this work is based on hiding the watermark data into the layout and formatting of the
document directly. Formatted text is probably the medium where watermarking can
be easily attacked. If the watermark data is in the format, then it can obviously be
removed by retyping the whole text using a new character font and a new format
where retyping can be either mginual or automated using optical character recognition
(OCR).
Cfiapterl
1.2.2 Audio Watermarking
The research on audio watermarking has been focused on either direct
watermarking of the audio signal or bit stream embedding where the audio is
represented in a compressed format. The use of perceptual models is an important
component in generating an effective and acceptable watermarking scheme for audio
[19], [21]-[23]. The requirements for audio watermarking are same as requirements of
watermarking for other media. The amount of information (watermark data) that can
be embedded robustly and imperceptibly is much lower than for visual media (images
and video) because audio signals are represented by much less samples per time
interval. The problem in audio watermarking is that the human auditory system
(HAS) is much more sensitive than the human visual system (HVS) and inaudibility
for audio signals is much more difficult to achieve than invisibility for images [21].
There are many watermarking techniques which are applicable to audio [21]-[23].
Three important techniques amongst them are echo coding [21], phase coding [21]
and spread spectrum coding [16], [22], and [23].
1.2.3 Image Watermarking
Most watermarking research and publications are focused on images [5] [8]
[20] [24] and [25]. The reason might be that there is a large demand for image
watermarking products due to the fact that there are many images available freely on
the World Wide Web (WWW) which need to be protected. Several watermarking
methods are in fact very similar and differ only in watermark signal design,
embedding method and detection/extraction method. Watermarking of images can
also be performed by using two dimensional (2D) versions of various transforms such
as Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT), Discrete
Wavelet Transform (DWT) and Discrete Fractional Fourier Transform (DFRFT) etc.
1.2.4 Video Watermarking
Video sequence consists of a series of consecutive and equally time-spaced
still images. Thus, the general- problem of watermarking seems very similar for
images and video sequences. The idea of image watermarking is directly applicable to
video sequences. However, additional requirements for video watermarking are low
CHapter 1
complexity, constant bit rate and compressed domain processing. Some video
watermarking approaches have already been earlier mentioned by several researchers
[19]-[20],and[24].
1.3 Objectives of the Thesis
The main objective of this thesis is to develop digital watermarking schemes
for multimedia data such as audio signals and images. It can be summarized as
follows:
1. To simulate and evaluate Ae performance of a digital audio watermarking
scheme using chirp signals as watermark.
2. To simulate and analyze the performance of a digital image watermarking
scheme in Discrete Fractional Fourier Transform (DFRFT) domain.
3. To propose and simulate a non-blind watermark extraction scheme in DFRFT
domain for image watermarking.
4. To simulate and demonstrate the performance of a digital image watermarking
scheme in combined DWT and DFRFT domains.
1.4 Outcome and Main Contributions of Thesis
In the present work, following are the main contributions:
• A chirp-based digital audio watermarking scheme is presented. Three types of
chirp signals (linear, quadratic and logarithmic) have been studied. This
scheme is simulated for single-level and multi-level watermark embedding.
The scheme is found to be robust against most of the audio watermark attacks
(except high-pass filtering, band-pass filtering and cropping), which are
generally applicable upon any audio watermarking technique. This scheme
shows limited robustness under high-pass and band-pass filtering attacks. The
multi-level watermarking scheme gives increased payload and multiple-level
robustness. The multiple- level robustness can be used to embed information
having different levels of security into a host audio signal.
• A digital image watermarking scheme in the DFRFT domain has been
proposed. In order to extract embedded watermark, a non-blind watermark
chapter 1
extraction scheme in DFRFT domain is proposed. Various simulation results
to evaluate the performance of this scheme under different attacks have been
studied and analyzed. This watermarking scheme shows imperceptibility
feature and is compared with existing DWT-based Kundur's watermarking
method. It is evident from the results that this method offers better robustness
than Kundur's method for some of the attacks such as salt and peppers noise,
median filtering, AWGN, and JPEG compression. This watermarking scheme
is more secure than other similar schemes because watermark can be
embedded using different DFRFT powers as secret keys, without its
knowledge watermark cannot be extracted.
• A digital image watermarking scheme in combined DWT and DFRFT
domains has been proposed for one-level and two-level DWT decomposition
with non-blind watermark extraction. Various simulation results of this
scheme under different watermark attacks have been reported. There is a
problem in DFRFT-based watermarking scheme that it offers higher BER in
watermark extraction for attacks such as histogram equalization and
sharpening. The proposed watermarking technique improves the watermark
extraction performance for these attacks.
1.5 Organization of Thesis
In Chapter 2, various digital watermarking techniques for audio signals and
images will be reviewed in brief
A digital audio watermarking scheme using chirp signal as a watermark has
been presented in Chapter 3 and its simulation results under various audio watermark
attacks has been discussed. In order to evaluate robustness of this scheme various
audio watermark attacks such as low-pass filtering, high-pass filtering, band-pass
filtering, interpolation, decimafion, re-sampling, amplitude scaling, addition of white
Gaussian noise, cropping and MP3 compression are applied on watermarked audio
signals. This chapter also presents simulation results for single-level and multi-level
watermark embedding and discusses various watermark attacks
In Chapter 4, digital image watermarking in DFRFT domain has been
developed, simulated and its robustness is evaluated under various image watermark
8
Cfiapter 1
attacks such as median filtering, salt and peppers noise, histogram equalization, JPEG
compression, low-pass filtering, addition of white Gaussian noise and sharpening.
Digital image watermarking scheme using combination of DWT and DFRFT
transforms has been given in Chapter 5. This scheme is simulated using one-level and
two-level DWT decomposition and DFRFT, and its watermark extraction
performance in terms of Bit Error Rate (BER) is measured under various image
watermark attacks.
Finally, Chapter 6 gives conclusion of the simulation results obtained in
Chapters 3 to 5 and provides some directions for further work.
p
CHapter 2 (Digitaf Watermar^ng
TecHniques
CfiapterZ
DIGITAL WATERMARKING TECHNIQUES
2.1 Introduction
Building up on the introduction of digital watermarking and its applications,
this chapter presents types of watermarking and attacks on watermarks. Overview of
various transforms which are used in transform domain watermarking will be given in
this chapter. At the end. various watermarking techniques for audio signals and
images will summarized in brief.
A digital watermark is a signal permanently embedded into digital multimedia
data i.e.. host signal (text, audio, images, and video) that can be detected or extracted
later b}' means of computing operations in order to make assertions about the data [2|
[26] [27]. The watermark is hidden in the host signal in such a way that it is
inseparable from the host signal and is resistant to any signal processing operation
which does not degrade the host signal perceptibly. Thus by means of watermarking,
the digital multimedia data (host signal) is still accessible but permanently marked.
Watermarks and attacks on watermarks are two sides of the same coin. The
goal of both is to preserve the value of the digital multimedia data. However, the goal
of a watermark is to be robust enough to resist attacks but not at the expense of
altering the value of the multimedia data being protected [26]. On the other hand, the
goal of the attack is to remove the watermark without destroying the value of the
protected data.
2.2 Types of Watermarking
Digital watermarking technology confirms the copyright of content owner by
inserting secret information (watermark) into the original host signal. It can integrate
closely with and hide in the original host signal such as text, audio, images and video
etc, with out destroying their commercial value. The term "digital watermarking' first
appeared in 1993. when Tirkel presented two watermarking techniques to hide
waterhiark data in images [28]. These methods were based on alteration of least
significant bits (LSBs) of the pixel values of an image. Various watermarking
methods are classified in the following categories:
10
Cfiapter2
• Robust, Fragile and Semi-fragile Watermarking: Robust watermarking is a
teclinique in which modification to the watermarked signal will not affect the
watermark. As opposed to this, fragile watermarking is a technique in which
watermark gets destroyed when watermarked signal is modified or tampered. A
semi-fragile watermark is robust to acceptable content preserving manipulations
such as lossy compression while fragile to malicious distortions such as content
modification.
• Visible and Transparent Watermarking: Visible watermarks are ones which are
embedded in visual content in such a way that they are visible when the content is
viewed. Transparent watermarks are imperceptible and they cannot be detected by
just viewing the digital content.
• Public and Private Watermarking: In public watermarking, users of the content
are authorized to detect the watermark while in private watermarking the users are
not authorized to detect the watermark.
• Asymmetric and Symmetric Watermarking: Asymmetric watermarking
(asymmetric key watermarking) is a technique where different keys are used for
embedding and detecting the watermark. In symmetric watermarking (symmetric
key watermarking) the same keys are used for embedding and detecting the
watermarks.
• Blind and Non-blind Watermarking: Watermarking in which original host signal
is not required for watermark detection/ extraction is known as blind
watermarking. If original host signal is required in watermark detection/
extraction then this watermarking is said non-blind (informed) type.
2.3 Attacks on Watermarks
In this section, various types of attacks applied to watermarked audio and
images will be discussed. Formally, any operation that affects the
detectability/extractability of a watermark can be considered as an attack on the
watermark signal. Attacks are of two types, standard attacks (unintentional attacks)
and malicious attacks (intentional attacks) Therefore, a distinction should be made
between the standard attacks and malicious attacks. The latter is defined as attacks
with the focus on removing the watermark, or renders it undetectable while
maintaining an acceptable signal quality level. A watermarked signal (audio, images
11
CdapterZ
or video) may be altered either purposely or intentionally. These attacks introduce
distortions in watermarked signal. Distortions introduced by these attacks should be
limited and do not produce excessive degradations; otherwise modified signal would
be unusable.
2.3.1 Audio Watermark Attacks
Most common attacks used against digital audio watermarking schemes are
discussed as under:
• Digital Filtering: It covers many attacks that can be described by simple
mathematical filters such as low-pass filtering, high-pass filtering, band-pass
filtering, time-scaling, etc.
• Sampling Rate Alteration: Converting digital audio to analog and then
con\erting it back to digital (re-sampling) is another form of attack. In this case,
frequenc}-, amplitude- and phase-dependent non-linearities can occur that destroy
watermark data. In this audio watermark attack, scaling or changing of sampling
frequency of watermarked audio signal is done. It can be done by three ways: up-
sampling, down-sampling and re-sampling. As the watermarked signal, up-
sampled, down-sampled or re-sampled, the sampling frequency is changed
according to the attack implemented and the characteristics of watermarked signal
are also changed. While playing the audio file as the sampling frequency is
changed the main information theme is lost so the audio signal must be re-scaled.
• Lossy Compression: The lossy encoding attack is a little more sophisticated and
harder to protect against. The idea behind lossy encoders, such as MPEG-1 layer-3
(MP3) compression, is to throw away any audio that is not important for human
hearing, based on psychoacoustic model [33]. However, if the encoder can
actually throw away the audio that is imperceptible to human listeners, then the
watermark is most likely be removed, since, by definition, it should not affect
human listeners significantly. In this attack, the watermarked audio signal is
compressed using MPS compression model with different Constant Bit Rates
(CBRs) from 56 kbps to 320 kbps. Then MPS file is de-compressed and converted
back to .wav audio file.
12
CfiapterZ
• Addition of White Gaussian Noise: This attack injects white Gaussian noise of
varying Signal-to-Noise Ratio (SNR) into watermarked audio signal.
• Amplitude Scaling: In this attack, amplitude of audio watermarked signal is
scaled by some integer factor.
• Cropping: This attack is used to remove some part of watermarked audio signal.
After cropping the watermarked audio signal, watermark detection/ extraction
performance is being determined.
2.3.2 Image Watermark Attacks
An image watermarking system could be judged against following attacks:
• Rotation: It is used to realign horizontal features of an image and it is certainly
the first modification applied to an image after it has been scanned.
• Cropping: Cropping refers to the removal of the some parts of an image. In some
cases, infringers are just interested in some specific part of the copyrighted
material. Cutting some part of the copyrighted material is called cropping.
• Scaling: Scaling can be divided into two groups, uniform and non-uniform
scaling. If the scaling is done by same factor in both horizontal and vertical
directions of an image, then it is called uniform scaling. Non-uniform scaling u.ses
different scaling factors in horizontal and vertical directions. In non-uniform
scaling, aspect ratio of an image changes.
• Compression: JPEG and .fPEG2000 are the most widely used compression
standards for images and any watermarking system should be resistant to some
degree of compression.
• Low pass filtering: This includes linear and non-linear filters. Frequently used
filters include median, Gaussian, and standard average filters.
• Sharpening; These filters can be used as an effective attack on some
watermarking schemes because they are very effective at detecting high frequency
noise introduced by some digital watermarking software.
• Histogram equalization: This usually improves the perceptual quality of an
image but may result into the removal of watermark data.
Cltaper2
• Noise addition: Any watermarking technique should be resistant after addition of
various types of noises such as salt and peppers noise with defined density and
additive white Gaussian noise (AWGN) with defined noise power.
2.4 Overview of Transforms
The watermark can be embedded into the signal directly acquired or the signal
may be transformed to embed watermark. In this section, an overview of various
types of transforms such Fourier Transform (FT). Discrete Fourier Transform (DFT).
Short Time Fourier Transform (STFT), Discrete Cosine Transform (DCT), Discrete
Fractional Fourier Transform (DFRFT) and Discrete Wavelet Transform (DWT) are
introduced which are commonly used for watermark embedding process.
2.4.1 Fourier Transform (FT)
The FT provides a means of transforming a continuous-time signal defined in
the time-domain into one defined in the frequency domain. It describes the continuous
spectrum of a non-periodic signal. The ¥1 X(f) of a continuous-time signal x(t) can be
expressed as:
.V(/)=Fnx(/) ]= [' x{t)e-''^'dt (2.1) J-CO
Inverse FT is expressed as:
.v(/) = InverseFTIXif)] = f X{f)e'^'^'df (2.2)
2.4.2 Discrete Fourier Transform (DFT)
Discrete Fourier Transform (DFT) is used in the case where both the time and
the frequency variables are discrete. Let x[n\ represents the discrete time signal and
X\k] represents the discrete frequency transform function. DFT of x[n\ can be
expressed as:
\\k] = DFT{x[n]]^Y^'\[n]e ^ , 0<k<N~l (2.3)
14
CHaperZ
Inverse DFT of A] ] is expressed as:
x\n\--Inverse DFT {X[k]\ = ^Y^\^\x[k]e ^ , 0 < «< A'- 1 (2.4)
where, n is tlie discrete-time index, k is the discrete frequenc}' index and .V is the time
period of both x[n] and X[k\.
The strongest components of the DFT are central components which contain
the low frequencies of the signal. DFT of a real signal/image is generally complex
valued, which results in phase and magnitude representation of a signal/image. Spatial
and circular shifts in the signal/image affect the phase representation of the
signal/image but not the magnitude representation. Scaling of an image is performed
b\' multiplying the pixel values of an image by a constant scaling factor. Scaling in the
spatial-domain causes inverse scaling in the frequency domain. Rotation of an image
results in cyclic shifting of pixel values of an image. Rotation in the spatial domain
causes the same rotation in the frequency domain.
2.4.3 Short-Time Fourier Transform (STFT)
STFT is used to determine the sinusoidal frequency and phase content of local
sections of a signal as it changes over time. In the continuous time case, the function
to be transformed is multiplied by a window function which is non-zero for only a
short period of time. The FT (one-dimensional function) of the resulting signal is
taken as the window slide along the time axis, resulting in a two-dimensional
representation of the signal. Mathematically, Continuous-Time STFT is written as:
.\(T.oj) = STFT{x{t)}= { x{t)w(t-r)e'""dt (2.5) J-CO
where, \i'(t) is the window function, x(t) is the signal to be transformed and X(T,CO) is
essentially the FT of x(t) vi'(t-x), a complex function representing the phase and the
magnitude of the signal over the time and frequency. The time index r is normally
considered to be "delav" time.
15
CfiapterZ
In the discrete-time case, the data to be transformed could be broken up into chunl'cs or
frames (which usually overlap each other to reduce artifacts at the boundary). Each
frame is Fourier transformed, and the complex result is added to a matrix, which
records magnitude and phase for each point in time and frequency. Discrete-time
STFT can be expressed as:
.\[m.co\ = STFT{x[n]] = y^ x[n\M[n ~m]e'""' (2.6)
where, xfn] is a discrete-time signal. In this case, time index m is discrete and angular
frequency ci)=2Kf\s continuous. The discrete-time index m is normally considered to
be "delay" time.
2.4.4 Discrete Cosine Transform (DCT)
A DCT expresses a sequence of finitely many data points in terms of sum of
cosine functions oscillating at different frequencies. DCTs are important to numerous
applications in science and engineering, from lossy compression of audio and images
to spectral methods for the numerical solution of partial differential equations. The
use of cosine rather than sine functions is critical in these applications. In particular, a
DCT is a Fourier-related transform similar to the DFT, but using only real numbers.
DCTs are equivalent to DFTs of roughly twice the length, operating on real data with
even symmetry (since the FT of a real and even function is real and even), where in
some variants the input/ or output data are shifted by half a sample. There are eight
standard DCT variants, of which four are common [34]. The most common variant of
DCT is the type-II DCT which is often referred as simply DCT. The inverse of type-II
DCT is known as type-ITI DCT and it correspondingly often named inverse DCT
(IDCT).
One dimensional DCT (ID-DCT) of signal x(n) is defined as:
'' ^ (2A/ + 1)LT (k) = DCT[x{n)]=-a{k)Y^x{n)cos^ '—, A =0,1,2,...,/V - 1 (2 .7 )
11=0
One dimensional inverse DCT (ID-IDCT) ofX(k) is expressed as:
CfiapterZ
.v(>?) = /DCT[A(A)] = j—ya(/c)A-(A)cos—'^"^'^^^-. /7 = 0,1.2 ,^'- 1 (2.8)
Uff, 2N
where, aiO) = ]/j2 and a{k) =1; ^ ;> 0
2.4.5 Discrete Fractional Fourier Transform (DFRFT)
Although the Fourier Transform (FT) is one of the most valuable and
frequently used tools in signal analysis and processing, but it is not suitable for the
analysis of non-stationary signals. Using FT, one can determine all the frequencies
present in a signal but it is not possible to determine locations (in time) of these
frequency components in the signal (i.e. it is impossible to know when they are
present). The FRFT was introduced in 1980 by Victor Namias [35] for the analysis of
non-stationary signals. McBride and Ken" later proposed the refinement and
mathematical defmhion of FRFT in 1987 [36]. In a very short span of time, FRFT has
established itself as a powerful tool for the analysis of time-varying non-stationary
signals [35]-[4I]. Furthermore, a general definition of FRFT for all classes of signals
(one-dimensional and multidimensional, continuous and discrete, and periodic and
non-periodic) was given by Cariolaro et al. [44]. Different approaches [39]-[41]
which have been developed to obtain Discrete FRFT (DFRFT) from Continuous
FRFT, but none of these approaches provide all the properties of continuous FRFT.
Santhanam and McClellan first reported the work on DFRFT in 1996 [45]. Thereafter
many more definitions of DFRFT came into existence and these definitions are
classified according to the methodology used for calculafions.
The FRFT has been found to have several applications in the areas of optics
and signal processing and it also leads to generalizafion of notion of space (or time)
and frequency domains which are central concepts of signal processing. FRFT has
also been found useful in applications such as in solving differential equations,
quantum mechanics, optical signal processing, study of space or time-frequency
distributions etc. [35], [37]-[38]. The FRFT may also have a significant impact in
other areas of science and engineering such as in physics, mathematics, pattern
recognition, image restoration and enhancement and matched filtering etc.
17
CHapterZ
Ff maps one-dimensional time signal/(?) into a one-dimensional frequency function
F(to). Although the FT provides the signal spectral content, it fails to indicate the time
location of the spectral components, which is important for non-stationary and time-
\arying signals. In order to describe these signals, time-frequency representations are
used. The Short-Time Fourier Transform (STFT) is an attempt to alleviate this
problem of FT [46]. It takes a non-stationary signal and divides it into windows of
signals for a specified short period of time. FT of each window is calculated
independently. It is a generalization of FT for localized time-frequency analysis.
Fractional Fourier Transform (FRFT) is the generalized form of the classical Fourier
Transform (FT). FRFT maps a one-dimensional time signal into a two-dimensional
function of time and frequency. The FT corresponds to a rotation in the time-
frequency plane over an angle equal to K/2. The FRFT corresponds to a rotation o\er
some arbitrary angle pn/l where 9 is a floating point number and is known as FRFT
power (also called the order of FRFT).
The p" order continuous FRFT of an analog signal/(/) is defined as; [42]
F''[f{t)]= ^K,,{t.u)f{t)dx, 0< |p [<2 (2.9) - X
Where, K p (t. ii) is the kernel function of the FRFT and it is given as:
f ^ i X ' - u ) '
1 - /COt« ( /" +11' til
exd 7 cota-j— , if a ^ lire In 'Y 1 " sin« J
(2.10) f)(u~t), if a = Inn-
Sill +1)- ifa = (ln + \)n
and p is the order of FRFT, a is the angle of rotation and n is an integer. The
parameters p and a are related as a = p7r/2.
The inverse of a ^ order FRFT is simply the FRFT with order -p. That is,
/ ( 0 = F-P\FP{f{t))] (2.11)
The concept of continuous FRFT can easily be extended to discrete FRFT (DFRFT).
Let/f'/j be a sampled periodic signal with a period Ts, the/?"' order DFRFT of f(t) can
CfiapterZ
be obtained by discretizing K^it.u) and replacing integral by summation in Eqn. (2.9)
[43]. giving as;
k
\ ( f i \ \ ?' W - / , 7 i4*^JS/l^^)'t't4J^-J
where, V is the number of samples in a period that is taken to be even.
The two-dimensional DFRFT (2D-DFRFT) of the image signal f(x. y) is determined
as
A/-1,V-1
The inverse 2D-DFRFTof the transformed image signal F„^„^_={m.n) is determined
bv
A/-I ,V-I
/(.Y,V-)= Y,YJ^"^"2^'"'"'^ A' „l_.,„(x,y,OT,H) (2.14) v=0 1=0
where, K,^^^^{x,y,m,n)^ K^^® K^^ is the 2D-DFRFT kernel K^^ and K^ are the 1D-
DFRFT kernels. Symbol ® denotes the convolution between K^^ and K^^^ • Flere,
a, = p^7T:'2and a, = p.iril are the fractional angles of 2D-DFRFT, and p^ and /?, are
the order or powers of 2D-DFRFT. An FRFT transformation domain is a combination
of the time and frequency domains. If /?, and pj ^re close to 1, the FRFT
transformation being dominantly in the frequency domain. On the hand, if p^ and /?,
are close to 0, FRFT is dominantly in the time domain. For the watermarking
proposed in this chapter, a frequency domain dominant case is used.
chapter 2
1A.6 Discrete Wavelet Transform (DWT)
The DWT is used to analyze a signal at different frequency bands with
different resolutions by decomposing the signal into a coarse approximation and detail
information [47]. DWT employs two sets of functions, called scaling functions and
wavelet functions, which are associated with low-pass and high-pass filters,
respectively. The decomposition of the signal into different frequency bands is simply
obtained by successive high-pass and low-pass filtering of the time-domain signal.
The original signal x[n\ is first passed through a half band high-pass filter }i[n] and a
low-pass filter h\n]. After the filtering, half of the samples can be eliminated
according to the Nyquist's rule, since the signal now has a highest frequency of n /2
radians instead of K. The signal can therefore be sub-sampled by 2, simply by
discarding every other sample [48J. This constitutes one level of decomposition and
can mathematically be expressed as follows:
" (2.15) yi„Ak]-2lx[>i]®h[2k->i]
where. )';„ ,;,| ]and yi,„,[k\ are the outputs of the high-pass and low-pass filters,
respectively, after sub-sampling by 2. Symbol ® denotes convolution operation
between two signals.
This decomposition halves the time resolution since only half the number of
samples now characterizes the entire signal. However, this operation doubles the
frequency resolution, since the frequency band of the signal now spans only half the
previous frequency band, effectively reducing the uncertainty in the frequency by
half The above procedure, which is also known as the sub-band coding, can be
repeated for further decomposition. At every level, the filtering and sub-sampling will
resuh in half the number of samples (and hence half the time resolution) and half the
frequency band spanned (and hence doubles the frequency resolution) [49].
DWT decomposes the image into three spatial directions, i.e. horizontal,
vertical and diagonal. DWT of an image results into four sub-bands of different
frequencies, namely, LL (low-low frequencies), LH (low-high frequencies), HL
20
Cfiapter2
(high-low frequencies) and HH (high-high frequencies) sub-bands. In actual, LL and
MH sub-bands, represent low and high frequency components of an image,
respectively. Both LH and HL represent the medium frequency components of an
image. Hence, wavelets approximate properties of HVS more precisely. DWT is
computationally efficient and can be implemented by using simple filter convolution.
Magnitude of DWT coefficients is larger in the lowest bands (LL) at each level of
decomposition and is smaller for other bands (LH, HL and HH). The larger the
magnitude of DWT coefficients the more significant it is. High resolution sub-bands
help to easily locate edge and texture patterns in an image. DWT follows the HVS
more closely than DCT. DWT coded image is a multi-resolution description of an
image. Hence, an image can be shown at different levels of resolution and can be
sequentially processed from low resolution to high resolution. Visual artifacts
introduced by DWT coded images are less evident compared to DCT because DWT
does not decompose the image inl:o blocks for processing. At high compression ratios,
blocking artifacts are noticeable in DCT coded images. But blocking artifacts are not
found in DWT coded images. DFT and DCT are full frame transforms, hence any
change in transform coefficients affects the entire image, except if DCT is
implemented using a block-based approach. However, DWT has a spatial frequency
locality, which means if signal is embedded it will affect the image locality. Hence
DWT provides both frequency and spatial descripfion for an image. However,
computational complexity of DWT is significantly higher than DCT.
2.5 Digital Watermarking Techniques
Current watermarking techniques can be grouped into three main classes. The
first class includes the time-domain/ spatial-domain watermarking techniques. In
these techniques, the watermark signal is embedded by directly modifying the sample
values/ pixel values of the original audio signal / image. The second class of
watermarking includes the transform-domain watermarking techniques in which
watermark data is embedded in the transform-domain signal coefficients. The
transform-domain watermarking techniques have been found to have the greater
robustness against common signal processing operations. The third class of
watermarking is the compressed-domain watermarking in which watermark signal is
embedded in compressed format of original host signal using some standard
21
copter 2
compression algorithms such as JPEG & JPEG2000 for images and MPEG-1 layer-3
{MP3) for audio signals. Compressed-domain watermarking is more popular for video
data because most of the video material is available only in compressed format and
the computational cost of completely decoding and re-encoding the video for adding
the watermark in some other domain can be really prohibitive. Apart from this,
considerations similar to those drawn for still images are also, in general, valid for
video. In this chapter, a brief review of spatial-domain/ time-domain and transform-
domain watermarking techniques have been given. Now, time-domain/ spatial-domain
and transform domain watermarking techniques for audio signals and images will be
discussed in subsequent sub-sections.
2.5.1 Time-Domain/ Spatial-Domain Watermarking
The most straightforward way to hide a watermark signal within a host signal
is to directly embed a watermark in the original host signal. For audio signal, this
direct watermarking technique is called time-domain watermarking, whereas for still
images this corresponds to spatial-domain watermarking. Both types of watermarking
are also known as asset domain watermarking.
In man}- cases, the choice to embed the watermark in asset domain is only
possible due to low complexity, low delay, low cost or some other system
requirements. Another advantage of embedding in the asset domain is that in this way
temporal/spatial localization of the watermark is automatically achieved, thus
permitting a better characterization of the distortion introduced by the watermark and
its possible annoying effects.
Time-domain watermarking techniques are mostly used for audio signals. As
opposed to still images the sampling rate needs to be known for having correct
reproduction. The most common sampling rate for high quality audio is 44.1 kHz for
audio CD. Samples are usually quantized with 16 bits per channel. The number of
samples of audio track is also very different with respect to image. In the case of still
images, the asset domain corresponds to the spatial domain and the host feature set
coincides with pixel values. In still image, the number of host features is limited by
the image size, whereas in the case of an audio, the number of available host features
depends on signal duration.
22
CdapterZ
Several audio watermarking algorithms in time-domain have been proposed.
Tlie first and the most common one is to embed vvatermarlc in time domain. One of
the simplest techniques under this category is Least Significant Bit (LSB) alteration.
In this technique, LSB of each sample value of the host audio signal is made "0' or ' 1"
depending upon the watermark bit to be embedded [50]. A large amount of data can
be embedded into an audio signal using this method. However, the major
disadvantage of this method is that it has poor robustness to signal processing
operations. In this technique, watermark can be easily destroyed by any signal
processing attacks. Fxho hiding is another audio watermarking technique in time
domain which embeds the watermark by introducing an echo [51]-[53]. In the most
basic echo watermarking scheme, the watermark information is encoded in the signal
by modifying the delay between the host signal and the echo signal obtained from
host signal. It means that two different values of delays (offset values) i.e., Ati and At2
are used in order to encode either a '0' or a ' 1'. Both offset values have to be carefully
chosen in a way that makes the watermark both inaudible and extractable or
recoverable [51]. If only one echo was produced from the original audio signal, only
one bit of information could be encoded. Therefore, the original audio signal is
broken down into blocks before the encoding process begins. Once the encoding
process is completed, the blocks are concatenated back together to create the final
signal. Echo hiding can effectively embed imperceptible w'atermark data into an audio
signal but the watermark embedding process is signal dependent. Another category of
watermarking techniques in time-domain is Quantization Index Modulation (QIM)
watermarking methods [54]-[55]. QIM methods have shown a very good rate-
distortion-robustness trade-offs and are probably better than addifive spread spectrum
and generalized LSB methods, against bounded perturbations. QIM refers to
modulating an index or sequence of indices with the watermark information and
quantizing the host signal with the associated quantizer or sequence of quantizers
[55J-[56]. Due to its advantages of low computational complexity, large capacity,
great robustness and blind extracfion, it is widely used in recently developed digital
audio watermarking schemes [57].
The concept of spread spectrum used in communication systems is also
employed in digital watermarking. Generally, the message used as the watermark is a
naiTow-band signal compared to the host signal. The main idea of spread spectrum
23
CHapterZ
watermarking is to embed a narrow-band watermark signal into wide-band host audio
signal. The spread spectrum techniques offer the possibility of protecting the
watermark privacy by using a secret key to control the pseudo random sequence
generator that is needed in the process.
The basic idea of spread spectrum is to encode audio signal by spreading the
watermark information [16] across as much of the audible spectrum as possible.
Using this technique, a watermark can be embedded robustly into a host audio signal
v\ithout destroying its perceptual quality. Watermark can also be hided by taking
advantage of the masking property of the human ear. Audio masking is the effect by
which a faint but audible sound becomes inaudible in the presence of another louder
audible sound. In this technique, the masking regions are first computed and the
watermark is then embedded into these areas [21].
Spatial domain watermarking techniques for images include [28] [29]-[32]
[60]-[61]. Some of the earliest techniques [28] [61]-[62] embed M-sequence into the
least significant bit (LSB) of the host signal to provide an effective transparent
embedding technique. M-sequences are chosen due to their good coirelation
properties so that a correlation operation can be used for watermark detection.
Furthermore, these techniques are computationally inexpensive to implement. Such a
scheme was first proposed in [28] and extended to two dimensions in [62]. In [61] the
authors reshape the M-sequence into two-dimensional watermark blocks which are
added and detected on a block-b3'-block basis. This technique has also been shown to
be an effective fragile watermarking scheme which can detect image alterations on a
block basis [63].
Several spatial-domain watermarking techniques for images have already been
proposed [21]. One such technique consists of embedding a texture-based watermark
into a portion of the image with similar texture. The idea here is that due to the
similarity, it will be difficult to perceive the watermark. The watermark is detected
using a correlation detector. Another technique described as the patchwork method
divides the image into two subsets A and B where the brightness of one subset is
incremented by a small amount and the brightness of the other set is decremented by
the same amount. The incremental brightness level is chosen so that the change in
intensity remains imperceptible, The location of the subsets is secret and assuming
24
chapter 2
certain properties for the image data, the watermark is easily located by averaging the
difference between the values in two subsets. It is assumed that, on an average,
without the watermark, this value will go to zero for image data. In the example
where the pixels in subset "A" are incremented by one and the pixels in subset 'B' are
decremented by one, with N locations in the set. the expected value of the sum of
differences between the sets is given by 2N. For non-watermarked data, this value
should go to zero. A variation of this approach is described earlier [60], where more
information can be inserted in the host signal. Another spatial-domain technique has
already been proposed [65], where the blue component of an image in RGB format is
watermarked to ensure robustness while remaining fairly insensitive to human visual
system (HVS) factors.
2.5.2 Transform-Domain Watermarking
In transform-domain watermarking techniques, the watermark is inserted into
the coefficients of digital transform applied on the host asset or host signal. Most
commonl}' used transforms preferred for watermark embedding in the frequency
domain are DFT, DCT. DWT, DFRFT etc. Usually, transform-domain w-atermarking
techniques exhibit a higher robustness to the attacks. Also robustness against other
types of geometric transformations, e.g. scaling or shifting, is more easily achieved in
a transformed domain. Since such a domain can be expressly designed so it may be
invariant under a particular set of transformations. For instance, techniques operating
in the magnitude of DFT domain are intrinsically robust against shifting, since a shift
in the time/ space domain does not have any impact on DFT magnitude.
Perceptual constraints aiming at ensuring inx'isibility can also be readily
incorporated into frequency domain representations, e.g. by avoiding the modification
of low' spatial frequencies where alterations may produce very visible distortions. On
the other side, frequency-domain watermarking techniques do not allow to localize
precisely the watermarking disturbs in the asset space, thus making it difficult to tune
it to the HVS or HAS characteristics. Another drawback of transform-domain
techniques is computational complexity. As a matter of fact, many applications cannot
afford the extra time necessary to pass from the asset to the transformed domain and
backward.
25
CfiapterZ
For Iransfomi-domain audio watermarking, the one-dimensional (ID) versions
of various transforms that were used for 2D still images are the most suitable. Many
transform-domain image watermarking schemes have been developed since the early
days of watermarking research. At the beginning, DCT was preferred to DFT. mainly
for its assonance with JPEG coding standard. Transform-domain watermarking is
useful for taking advantage of perceptual criteria in the embedding process, for
designing watermarking techniques which are robust to common compression
techniques.
Various transform-domain audio watermarking techniques [57] [66-72J have
been proposed. The transforms can be DFT [57] [66]-[67]. DCT [68] or DWT [68]-
[72J. In these approaches, the amplitude or phase of the transformed coefficients has
been modified with some specified amount in order to caiTy out watermark data.
A common transform framework for images is the block-based DCT which is
a fundamental building block of coding standards such as the JPEG and \'ideo coding
standards such as the MPEG video coders [73]. One of the first block-based DCT
watermarking techniques has been earlier proposed [74]. The DCT was performed on
8.x8 blocks of data, a pseudorandom subset of the blocks was chosen and a triplet of
midrange frequencies were slight'y altered to encode a binary sequence. One of the
effective watermarking works [16] was first to describe how spread spectrum
principles borrowed from communicafion theory can be used in the context of
watermarking. The published results show that the technique is very effective both in
terms of image quality and robustness to the signal processing attempts to remove the
watermark. The technique was motivated by both perceptual transparency and
watermark robustness. One of the significant contribufions from this work was the
realization that the watermark should be inserted in the perceptually significant
portion of the image in order to achieve robustness against the attacks. A DCT was
performed on the whole image and the watermark was inserted in a predetermined
range of low frequency components excluding the DC component. The watermark
consists of a sequence of real numbers generated from a Gaussian distribution which
was added to the DCT coefficients. The watermark signal was scaled according to the
signal strength of the particular frequency component. This is a reasonable and simple
26
CfiapterZ
wax lo introduce some type of perceptual weighting into the watermarking scheme.
The watermark embedding algorithm is described as:
.V = .v(l + cr 11') (2 .16)
where, ,v is the original host signal, w is the watermark consisting of a random
Gaussian distributed sequence, a is a watermark scaling factor and x is the
watermarked signal. Parameter a is used to provide a good trade-off between
imperceptibility and robustness.
A \ariation of DCT-based watermarking is the variable lengths DCT-based
watermarking [75], where the DCT coefficients are sorted according to the magnitude
and only the n largest coefficients are marked that correspond to a user specified
percent of the total energy. This allows the user to trade off imperceptibility and
robustness to attack. Other DCT-based watermarking schemes use more elaborate
models of the HVS to incorporate an image adaptive watermark of maximum strength
subject to the imperceptibility criterion [20],[76],[77J. Two image adaptive
watermarking schemes are described earlier [77] and [81]. which are based on a
block-based DCT framework and wavelet framework. The perceptual models used
here can be described in terms of tluee different properties of HVS that have been
studied in the context of image coding: frequency sensitivity, luminance sensitivity,
and contrast masking [80]. A combination of the three components results in Just
Noticeable Distortion (.IND) thresholds for entire image. These models were first
de\'eloped to design more efficient image compression schemes than waveform
coding techniques alone could provide. This model was derived for the base line
mode of JPEG and showed a significant improvement in compression performance
when used to derive an image-adapfive quanfization table [78]. A similar model was
developed for wavelet-based compression using only frequency sensitivity to derive
perceptual weights for each of the sub-bands [79]. This model was used for a
wavelet-based watermarking scheme [77] and [80]. An image-dependent
watermarking technique is described in [75]. Two DCT-based approaches were
described in [81] and [82] where watermark detection does not require the original
image. The method in [83] is an expansion of the method proposed in [80] and [77] to
27
chapter 2
the case where the original host signal is not available for watermark detection. This
is an important feature for some applications such as authentication.
In [84]. a two-dimensional chirp signal was embedded in the low frequency
band ol~the host image in w avelet domain. The watermark was embedded in wavelet
domain and detected in the fractional Fourier transform (FRF'f) domain. This method
does not need the original image for watermark detection. It is more robust compared
to conventional spatial-domain algorithm. In [85], a blind digital image watermarking
algorithm using FRFT was presented. In this technique, multiple chirps were used as
watermark which was embedded in the spatial domain directly and w^atermark was
detected in FRFT domain. The watermark was detected conveniently according to the
propert}' that chirp signals can produce an impulse in the FRFT domain. This
algorithm has good security, imperceptibility and excellent resistant against the
attacks of JPEG compression, noise, cropping and filtering [85].
2.6 DWT-based Kundur's Digital Image Watermarking Technique
Kundur and Hatzinakos proposed a DWT-based robust watermarking
technique [86] known as digital image watemiarking using wa\elet-based fusion. This
watermarking method employs a multi-resolution wavelet decomposition of both the
host image and the watermark. Vi hen an image undergoes wavelet decomposition, its
components are separated into bands of approximately equal bandwidth on a
logarithmic scale much as the retina of the eye splits an image into several
components. It is, therefore, expected that use of the DWT will allow the independent
processing of the resulting components much like the human eye. Wavelet fusion
methods make use of information about the Iluman Visual System (HVS) to
determine what information, from each image, is important to retain in the composite.
Then somewhat complementary HVS rules can be used to robustly embed a
watermark imperceptibly inside the host image [86].
Firstly the host image and the watermark are transformed into the wavelet
domain. Then the detail images at each resolution level are segmented into non-
overlapping rectangles. The silence S is a numerical "measure of perceptual
importance. The silence S of each of these localized segments is computed using
information about the HVS.
28
CHapterZ
The watermark is embedded by simple scaled addition of the particular detail
component. The scaling of the watermark is a function of the silence of the region.
The greater the silence S, the stronger is the presence of the watermark. The
computation of silence is described in [86].
Then the inverse wavelet reconstruction of the fused image components is
performed to form the watermarked image. The general overview of this
watermarking method is shown in Figure 2.1.
To detect the received image to be credible, the cross-correlation between the
original watermark and the extracted watermark is evaluated. If the correlation
coefficient is reasonably high then the image is considered credible.
Most Image
Watermark
L Level DWT
Decomposition
First Level DWT
Decomposition
^ '
Fusion of Detail
Coefficients
i I
— •
Inverse L Level DWT
Reconstruction
Watermarkec Image
Figure 2.1 General overview of DWT-Based Kundur's Image Watermarking
Technique.
29
W Iffi
CHapter 3
^igitaCAucfio Watermar^ng
J
CHapterS
DIGITAL AUDIO WATERMARKING
3.1 Introduction
The previous chapter gave an overview of different digital watermarking
techniques. In this chapter, a chirp-based digital audio watermarking technique is
being presented. A watermarking scheme must satisfy the conflicting requirements
of robustness and high embedding capacity (payload). In digital watermarking,
robustness of watermark can be increased at the cost of decreasing embedding
capacity and vice-versa.
To enhance the payload and to introduce different levels of robustness in the
embedded watermark a multi-level chirp based watermarking is proposed.
Ad\'antages of using chirp signals as a watermark for digital audio are discussed in
this chapter (followed by simulation results). Finally, results of single-level and multi
level chirp-based digital audio watermarking under different attacks is summarized.
3.2 Background
In this section, overview of chirp signals and review of digital watermarking
using chirp signals as watermark are given.
3.2.1 Chirp Signals
A chirp signal is a cosine signal with frequency sweep. Chirps occur in natural
signals such as animal sounds (birds, frogs, whales, and bats), whistling sound, as
well as in man-made systems such as in radar and sonar [87]-[88]. Due to the
possibility of different rate of change in the chirp frequency it has also been used for
watermarking of audio signals and images [89]-[91]. Chirp signals can be classified as
linear, quadratic, and logarithmic depending upon the nature of sweep frequency with
the possibility of both an up sweep and a down sweep. These chirp signals are defined
as follows:
30
Cfiapter3
Linear chirp signal: It is specified by an instantaneous frequency sweep/(/)
which is a Unear function of time. This instantaneous frequency sweep is given
by:
/ , (0=./o+/^^ (3.1)
where, /? = (/,-/<,)//, (3.2)
/ ) is the initial frequency and /, is the final frequency at the target time (/,). The
time duration which is required to sweep the frequency from /„ to / j , is known as
target time (/,) of the chirp signal. Tlie parameter /^ is known as frequency
sweep rate and it ensures that the desired frequency breakpoint J\ at time /, is
maintained. If/, > /y then the corresponding chirp signal is known as up-sweep
chirp signal, else if /p > /, then it is known as down-sweep chirp signal
Spectrogram of a linear chirp signal of frequency sweep from 0 Hz to 100 Hz is
shown in Figure 3.1(a) for target time of 2 s.
Quadratic chirp signal: The instantaneous frequency sweep/(/•) is a quadratic
function of time and it is mathematically expressed as:
m = fo + /^r (3.3)
where./? = (./,-/o)/r,^ (3.4)
Spectrogram of a quadratic chirp signal of frequency sweep from 0 Hz to 100 Hz
and a target time of 2 s is shown in Figure 3.1(b).
Logarithmic chirp signal:; The instantaneous frequency sweep/(/) is a
logarithmic function of time and is defined as:
./i(0 = /o/?' • (3.5)
Putting / = /, for ./,(/) = ,/i in Eqn. 3.5,
,/i =fo/^' (3.6)
CfiapterS
Taking logioof both sides of Eqn. 3.6 and simplifying.
h = 'ogio(/i)-Jogio(/o)
log,o(>9) (3.7)
( i / ' i ) where, y9 = ( / , / /or" ' (3.8)
Spectrogram and waveform of a logarithmic chirp signal of frequency sweep (10
Hz -100 Hz) are shown in Fi^^es 3.1(c) and (3.2), respectively for /, equal to 2 s.
Initial frequency (/o) is kept (jqual to 10 Hz instead of 0 Hz because log,o(0) is
equal to -co.
(a)
(b)
(c)
0 . 5 Time
Figure 3.1 Spectrograms of chirp signals (a) Linear (0-100 Hz) (b) Quadratic (0-100 Hz) (c) Logarithmic (10-100 Hz).
32
CfiapterS
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
n
0.5 1 Time (sec.)
1.5
Figure 3.2 Waveform of a logarithmic chirp signal (10-100 Hz).
Chirp signal is selected as a watermark because it has a frequency sweep
instead of a fixed frequency. Due to this reason, chirp-based watermark is robust
against most of the audio attacks [92]. The chirp signals are used as a watermark in
digital watermarking of audio and images because they have a unique property that
they can be extracted from a high-noise background [92]. Chirp signals have been
used in watermarking because of its ability to give low side-lobes when correlated
with itself This gives an advantage of being detected/ extracted in the presence of
high background noise [92].
3.2.2 Related Works
In this sub-section, watennarking schemes using chirp signals as watermark
are discussed in brief Digital audio watermarking using chirp signals as a watermark
has been attempted earlier [89], [90] in which only linear chirps were used.
In [89], a spread spectrum audio watermarking algorithm was presented that
embeds linear chirps as a watermark messages. Different slopes (chirp rates) on the
time-frequency (TF) plane repn;sent watermark messages such that each slope
corresponds to a different message. Watermark messages were extracted using a line
detection algorithm based on the Hough-Radon transform (HRT). This method
33
CfiapterS
performs detection as well as extraction of embedded watermark bits. The
watermarking scheme detects the embedded watermark message even after common
signal-processing operations such as MPEG audio coding, re-sampling, low-pass
filtering and amplitude re-scaling.
In [901, linear chirps were embedded in audio signals as watermark. Different
chirp rates were used to represent different watermark messages such that each chirp
rate corresponds to a unique v/atermark message. The watermark messages were
embedded and extracted using line detection algorithm based on HRT. This scheme
correctly detects embedded watermark after common signal processing operations
such as MP3 compression, low-pass filtering, re-sampling, amplitude scaling and
additive noise [90J.
Linear chirp signals have also been used for watermarking digital images [91],
[84|-[85|. In [91], a robust image watermarking algorithm that embeds multiple
watermark bits in the host image was presented. Linear chirps were embedded as
watermark messages, where the slopes of the chirp on the TF plane represent
watermark messages, such that each slope corresponds to a unique message. Since
linear chirps are localized as a straight line in the TF plane, the HRT was used to
detect the watermark messages. The robustness of this scheme was evaluated using
the checkmark benchmark attacks [98]. This watermarking scheme detects the
watermark messages correctly after common image processing operations such as
.IPLG compression, filtering and re-sampling.
In [84], a two-dimensional (2D) chirp signal was embedded in the low
frequency band of the host image in wavelet domain. The watermark was embedded
in wavelet domain and detected in the fractional Fourier transform (FRFT) domain
blindly. It is more robust compared to conventional spatial-domain watermarking
algorithm. ' fc^
In [85], a blind digital image watermarking algorithm using FRFT was
presented. In this technique, multiple chirps were used as watermark which was
embedded in the spatial domain directly and watermark was detected in FRFT
domain. The watermark was detected by using the property that chirp signals can
produce an impulse in the FPJT domain. This algorithm has good security,
CHapterS
imperceptibility and excellent resistance against JPEG compression, noise, cropping
and filtering [85].
3.3 Proposed Audio Watermarking Scheme
In this section, a chirp-ba.sed digital audio watermarking scheme is presented.
Here the high correlation property of chirp signals is exploited and in phase and out of
phase chirps are embedded for T and '0', respectively. First, watermark embedding
then watermark extraction is discussed in subsequent sub-sections.
3.3.1 Watermark Embedding
In the proposed watermark embedding scheme, a chirp signal is generated
with its parameters such as its initial frequency ( /Q), final frequency ( /J) and target
lime (r,). The generated chirp signal is coded according to the watermark data to be
embedded. To embed a ' I' the chirp is generated using above parameters while for '0"
its phase is reversed. Finally these chirp signals corresponding to the watermark data
of N bits are concatenated to form a signal of duration exactly equal to audio signal.
Watermarked audio signal x„ is given by:
X,, + a \\\ (3.9)
where, x,, is the host audio signal, w^. is the concatenated chirp signal, and a is the
watermark scaling factor. The scheme of watermark embedding process is shown in
Figure 3.3. Watermark data l
Generation of chirp signal with
parameters
( / o - . / r O
.T Phase
Encoding
Host audio signal (J;,) / ^ \,«^ c
~Ky 1 '
Concatenation of chirp signals for watermark data
of N bits
Watermark amplitude scaling
(a)
w.
Watermarked audio signal (x,,,)
Figure 3.3 Block Diagram of Watermark Embedding Process.
35
CdapterS
3.3.2 Watermark Extraction
With the knowledge of the type of chirp (linear, quadratic or logarithmic) and
its parameters ( / Q , / and?,), the watermark can be extracted by measuring segment-
by-segment correlation of watermarked audio signal with the chirp signal. Watermark
extraction process is shown in Figure 3.4. A chirp signal x similar to the one used
during the process of embedding is generated. The watermarked audio signal x„ is
segmented into A' equal parts each of duration /, second and each /"' segment of i\, is
denoted by A-;, . The watermark bits from x„ are extracted by calculating the cross-
correlation coefficient between the chirp signal x andx , . Cross-correlation function
(XCF) of .Y and xl is determined by using the relationship given below:
ArF(/) = ^A-;,(A)x(/:), 1 < / > A (3.10)
where. L is the number of samples.
A sample plot of ACT of various segments of watermarked audio signal .v-;, with
specific chirp x that was used in watermark embedding process is shown in Figure
3.5.
The detected watermark bit Wj is obtained by:
[1 //• XCF{i)>0 vr, = ' (3.11)
[0 otherwise
The numbers of erroneously detected watermark bits are calculated by performing bit
wise XOR operation between original watermark bits (w,J and detected watermark
bits ( »;/), as follows:
1, M', IS in error ir/®vr„- '' (3.12)
0, M\, is not in error
Percentage Bit Error Rate (BER) is defined as
36
N
CfutpterS
(3,13)
where. A'\ is the total number of erroneously detected watermark bits and yV is the
total number of watermark bits. The performance of the proposed embedding method
is measured in terms of BER of the detected watermark bits, in comparison to the
original watermark bits. Higher the BER\ poorer is the performance of the watermark
algorithm.
Generation of chirp signal with parameters
(,/o.,/i-0
Watermarked audio signal (.v„ )
Division of x^^
in N equal and non-overlapping segments
Cross-correlation
measurement
Detection of watermark bits
and concatenation
Detected Watermark {wj \
Segments of x„
Figure 3.4 Block Diagram of Watermark Detection Process.
Cross-correlation function (XCF) is of high value at the places where the
amplitude of audio segments is of lower value. It is because the chirp signal as
compared to the audio signal has relatively higher amplitude at the places where the
audio amplitude is low, which results into a higher degree of similarity measure.
20 30 40 Segment number
Figure 3.5 Plot of cross-correlation function (XCF) vs. number of audio segments.
CfiapterS
3.4 Single-Level and Multi-Level Watermarking
In single-level watermarking, only one watermark data is embedded in the
host signal. Multi-level watermarking can be defined as a process of embedding
multiple watermarks to the same host signal, where each watermark can be detected
or extracted securely with corresponding keys without the knowledge of other
watermarks in the host signal [101]. One of the main purposes of multi-level
watermarking is to enable secure, user-specific watermarking for several users such as
media creators, distributors and end users. Different applications use watermarks for
different purposes. Each individual application has its own set of mutually conflicting
requirements such as embedding capacity, robustness, and perceptual quality. Two or
more watermarks can be embedded into a host signal to simultaneously satisfy the
requirements of multiple applications.
Multiple watermarks can be embedded in non-overlapping areas of the host
signal. This can be done by dividing the information carrying coefficients of the host
signal into multiple groups and embedding different watermarks into different groups
of coefficients [100]. This division can be done in time, space or frequency. Muhiple
watermarks are usually embedded into non-overlapping segments to avoid
interference among different levels. It can also be embedded into a host signal in an
overlapping fashion, where different watermarks are embedded successively, and they
all share the same watermark embedding space. Overlapped embedding is susceptible
to interference and conflicts among the different watermark levels [100].
An example of multi-level audio watermarking where three different
watermarks have been embedded at three levels and segments of watermarked audio
signals for all the three levels are shown in Figure 3.6. There is no visual difference
among segments of original audio signal and corresponding three levels watermarked
audio signals. It can be said from visual inspection of wavefonns shown in Figure 3.6
that this audio watermarking technique is imperceptible.
3.4.1 Three-Level Digital Audio Watermarking
Multi-level watermarking is implemented in audio signal on three different
levels. The three levels of different chirp-based watermarks have been embedded to
38
CfiapterS
increase embedding capacity (payload) as well as to offer different levels of
robustness. Hence, it can be used to embed information having different levels of
security into a host audio signal. Three levels of watermarks are embedded by
following the steps given below:
Step-I: hiitialK'. the parameters for three levels of watermarks i.e. chirp signals are
selected. Chirp parameters are initial frequency ( /(,), final frequency {J\) and target
time (/,). Different chirp parameters are used for three levels of watermark.
Step-U: All the three levels of watermarks are embedded one over another in host
audio signal.
Step-lII: First level of watermark with chirp parameters /„ = 10 Hz. /', = 60 Hz. and
/, = 0.25 sec. is embedded in original host audio signal and it produces first-level
watermarked signal (.\\,,).
Step-IV: The second level of watermark with chirp parameters /g = 10 Hz. /, = 30
Hz. and /,= 1 sec. is embedded in .Y,,, which produces second-level watermarked
signal (A,,,,).
Step-V: Finally, third level of w'atermark with chirp parameters /i, = 10 Hz, /,= 15
Hz. and [^= 1.5 sec. is embedded in x^^.j vvhich gives third-level watermarked signal
(.T,,,). In three-level watermarking,x^, is also known as final watermarked audio
signal and is denoted by x„,.
Step-VI: All the three levels of watermarks that were embedded at three levels in
audio signal can be extracted by determining cross-correlation function of each
segment of x„ and specific chirp signal (x) that was used for a particular level of
watermark.
39
C/iapter3
(a) (b)
ai us Tlnw (sec)
(c) (d)
Figure 3.6 Waveforms of sections of (a) original audio signal (b) First-Level
watermarked audio signal (c) Second-Level watermarked audio signal (d) Third-Level
watermarked audio signal.
3.5 Performance Measuring Parameters
Important parameters to evaluate the performance of a digital watermarking
scheme are its perceptual transparency and robustness.
3.5.1 Perceptual Transparency
Perceptual transparency or imperceptibility of a watermarked audio signal
(also known as inaudibility) can be accessed either by subjective or objective
evaluation. Imperceptibility of a M'atermarked audio signal can be checked either by
40
CHapterS
performing listening tests given by Mean Opinion Score (MOS) (subjective) or by
measuring Objective Difference Grade (ODG) using Perceptual Evaluation of Audio
Quality (PFAQ) [101] model (objective). MOS gives a numerical indication of the
perceived quality of audio received after being transmitted and eventually compressed
using codecs. MOS is expressed as a number on a scale of 1 to 5, 1 being the worst
and 5 the best. MOS can also be used to evaluate perceptual quality of watermarked
audio signals because MOS gives a perceived quality measurement of difference
between watermarked and original host audio signal. Trained subjects are asked to
listen to the audio and rate them on a scale of 1 to 5. Then an average rating is
calculated which is known as MOS. The embedding process should not introduce any
perceptual artifacts in the original host audio signal. The perceptual audio quality of
chirp-based watermarked audio signals is evaluated using PEAQ basic model based
on the International Telecommunication Union (ITU) recommendation (ITU-R
BS.1387) [101]-[102]. PEAQ model measures ODG of watermarked audio signal.
3.5.2 The PEAQ Algorithm
PEAQ algorithm is the ITU-R recommendation (ITU-R BS.1387) [101]-[102]
for perceptual evaluation of wide-band audio coders. Two versions of PEAQ model
are available and these are basic model and advanced model. The PEAQ algorithm
models the fundamental properties of the Human Auditory System (HAS) with
physiological and psychoacoustic effects. This algorithm uses both original and
watermarked audio signals to find differences between them. An ODG is evaluated
using a total of eleven Model Output Variables (MOV) of basic version of PEAQ
model. ODG values mimic the listening test ratings and have values -4.0 for very
annoying quality to zero for imperceptible (inaudible) difference quality. The ODG is
calculated by PEAQ algoritlim specified in ITU-R BS-1387-l and it corresponds to
the Subjective Difference Grade (SDG) used in human based audio tests. SDG is the
difference grade between the listener's ratings of the reference and coded signals. The
SDG has a negative value when the listener successfully distinguishes the reference
from the coded signal. The SDG has a positive value when the listener erroneously
identifies the coded signal as the reference. The SDG equal to zero shows the
transparency region and any impairment is imperceptible while an SDG of -4
indicates a very annoying impairment. The ODG scores are equivalent to SDG
41
_________________^________^^^^^^^^ CfiapterS
scores determined through subjective listening tests. The subjective quality
assessment of audio is performed using SDG which is evaluated with respect to
original reference audio signal. In case of objective quality assessment, a set of audio
parameters like Signal-to-Noise Ratio (SNR) or Noise-to-Mask Ratio (NMR) is
computed. These parameters should be chosen so that their values are strongly
correlated to the perceived quality loss due to watermark embedding. ITU provides a
recommendation to evaluate the perceived audio quality with two methods. Basic
PEAQ and Advanced PEAQ models. Both are full reference quality indexes. They are
computed with respect to the original reference audio signal. The resulting indexes are
named ODG. The ODG values and its corresponding impairment description are
given in Table 3.1. The ODG for audio v/atermarking techniques is defined as:
ODG = Grade of a watermarked audio signal - Grade of a host audio signal
Table 3.1 ODG values and its impairment description
Impairment Description
Imperceptible
Perceptible, but not annoying
Slightly annoying
Annoying
Verj annoying
ITU-R Grade
5.0
4.0
3.0
2.0
1.0
ODG Score
0
-1.0
-2.0
-3.0
-4.0
3.5.3 Signal-to-Watermark Ratio (SWR)
Objective measurement of audio quality is also performed on the watermarked
audio signal ( . r j and attacked watermarked audio signal (.r,„,). It is determined by
measuring SWR of the watermarked audio signal (xj with reference to the host
signal (.V,,) and denoted by SWRo and SWR of the attacked watermarked audio signal
(A-,„,) with reference to x^, and denoted by SWRa.
SWRo: It is the ratio of power of original host audio signal ( J;, ) to the power of error
signal obtained by subtracting watermarked audio signal (A-,J from x,,. SWRo is
defined by:
42
chapters
SWRo = 10 log,
J^^lii) 1=0
A \ - i
Y.i^„i')-xji)f 1=0
(3.14)
where. .\\ is the number of samples in x,, and x„,
SWRa: It is the ratio of power of original host audio signal (.Y,, ) to the power of error
signal obtained by subtracting attacked watermarked audio signal (,v„„) from .r„.
SWRu is defined bv:
.v,~i
SWRa = 10 log X /' '
101 ;=0
. 'V, - l
^[.v,(/)-x„„,(/)]^ /=0
(3.15)
where. A'\ is the number of samples in j / , and A-,„.
3.5.4 Robustness against Audio Watermark Attacks
The watermark should not get degraded or destroyed as a result of
unintentional or intentional (malicious) signal distortions. The robustness
performance of the proposed watermark embedding scheme is measured in terms of
BER as defined in Eqn. 3.13.
3.6 Simulation Results and Discussion
In this secfion, various simulation results for single-level and multi-level
chirp-based digital audio watermarking schemes are presented and results are
discussed. For chirp-based watermark embedding, 14 audio files selected from Speech
Quality Assessment Material (SQAM) [103] are used. These audio files are sampled
at 44.1 kHz and have a resolution of 16 bits per sample. In actual, these are wide-band
stereo audio files (6 speech files and 8 music files) chosen from a set of audio files
provided by European Broadcasting Union (EBU) for non-commercial research and
testing purposes.
43
Cfiapter3
In this work, all the three types of chirp signals (linear chirp, quadratic chirp,
and logarithmic chirp) have been used for digital audio watermarking A comparative
axerage performance of chirp-based audio watermarking using three types of chirp
signals is given in Table 3.2. In the subsequent discussion about audio watermarking
scheme, a logarithmic chirp signal is selected to embed watermark because it offers
lower BER i.e. zero in extracted watermark at approximately same level of ODG.
which is an important criterion of any audio watermarking schemes.
Table 3.2 Comparative average performance of chirp-based audio watermarking
scheme using three types of chirp signals
1 Type of Chirp Signal
! Linear
( Quadratic
Logarithmic
Frequency Range
(,/o',/,)'Hz
(0-60)
(0-60)
(10-60)
Target Time (/,), second
0.25
0.25
0.25
BER
1.02
3.18
0
ODG
-0.3854
-0.3738
-0.3851
SWRo (dB)
30
30
30
Waveforms of sections of an original audio signal and chirp-based watennarked audio
signal are shown in Figure 3.7. It can be observed that the difference between the
original and watermarked audio is very small as shown in Figure 3.7 (c).
44
CfiapUrS
(a)
(b)
o.sts
(C)
Figure 3.7 Waveforms of sections (a) original host audio signal (b) chirp-based
watermarked audio signal (c) difference between two signals as shown in parts (a) and
(b).
45
CfiapterS
The parameters selected for single-level chirp-based watermark embedding are
given in Table 3.2 and number of bits to be embedded per audio file is 63. A low
frequency sweep range is chosen for chirp-based digital audio watermarking because
the human ear has poor sensitivity at low frequencies [104]. This will cause the
watermarked signal to have imperceptible difference in comparison to original signal
even at lower SWRo.
Various parameters and corresponding performance of three-level audio
watermarking on SQAM audio files are given in Table 3.3. For three-level audio
watermarking scheme, logarithmic chirp signal of frequenc}' sweep of 10-60 Hz for
first leveL 10-30 Hz for second level and 10-15 Hz for third level have been used to
generate the chirp streams. These frequency ranges have been chosen in order to keep
the watermarks imperceptible and to have different frequency sweep rates. The
lengths of the chirps are kept different for all the three watermark levels for
minimizing the BER in watermark extraction.
Objective assessments of the audio quality have also been carried out by
calculating the SWRo at each v/0.termark level. The average SWRo was evaluated for
all the audio files under test at three levels and results obtained are being given in
Table 3.3. This SWRo can be improved if the length of the chirp is increased, but for
long chirp, lesser bits can be embedded in the audio signal. It is important to note in
Table 3.3 that the target time of the chirp is increased with the increase in the level of
watermark embedding. This is done because with the increase in the watermark level
the SWRo decreases; thus to have zero BER. longer durafion chirp helps to have better
correlation.
Table 3.3 Parameters and corresponding performance of Three-Level Watermarking
Watermark Level (WL)
1 II III
Frequency Range
(./o^A) Hz
10-60
10-30 10-15
Target Time (/,) (second)
0.25
1 1.5
ODG
-0.38
-0.46 -0.48
SWRo (dB)
30 27.04 25.09
Number of Bits
embedded per audio
file
63 15 10
46
Cflatter 3
3.6.1 Robustness Evaluation for Single-Level and Multi-Level
Watermarking
To evaluate robustness of the single-level and multi-level chirp-based digital
audio watermarking schemes, various audio attacks such as low-pass filtering, high-
pass filtering, band-pass filtering, interpolation, decimation, re-sampling, addition of
white Gaussian noise, amplitude scaling, MPEG-3 Layer-1 (MPS) compression and
cropping are applied. Robustness of the chirp-based audio watermarking schemes can
be evaluated by measuring correlation between the original watermark data that was
embedded and the recovered watermark data.
Low-Pass Filtering
A Finite Impulse Response (FIR) low-pass filter (LPF) of order 50 with
different normalized upper cutoff frequencies («„, ) was designed to attack the single-
level and multi-level watermarked audio signals. Plots of average BER in extracted
watermark vs. co^^i and SWRa vs. &»,,, for single-level audio watermarking are shown in
Figure .3.8. Figure 3.9 shows watermark extraction performance with increasing cutoff
frequency <y,,, for all the three levels of watermarks. Variation of SWRa with
increasing co^^i for three-level watermarked audio signal is shown in Figure 3.10. It is
observed from the Figures 3.8 and 3.9 that single-level and multi-level chirp-based
audio watermarking schemes are robust against low-pass filtering attack (i.e. after
low-pass filtering, watermarks are extracted without any error) due to the watermark
occupying the low frequency (below 100 Hz). BER in extracted watermarks for all the
three levels is also zero because all the three watermarks occupy low frequency
ranges as shown in Table 3.3. 'e^
The (o„i^ of LPF varies between 0.1 (4410 Hz) and 0.9 (39690 Hz). The
embedded watermark frequency is below 100 Hz for both single-level and multi-level
audio watermarking. The embedded watermark is extracted from the low-pass filtered
watermarked audio signal without error because value of &>,,/ is much higher than the
embedded watermark. This implies that even if 90% of the bandwidth (BW) is lost in
filtering the watermark can still be recovered although the signal becomes useless.
47
C/iapter3
The SWRu of low-pass filtered watermarked audio signal increases with increasing
(0^,1 from 0.1 to 0.9 because the BW of the watermarked audio signal increases for
increasing the \'alue of r;';,,, .
1 35
0.9 :
0.8
0.7
30
25
0.6 ' ^ - - - ^ , - m _ - ^ ^ 20 T3 a: ^^ ~
UJ 0.5 / -"^ "I m m^ a:
0.3 B" , 10
0.2 5
0.1
0 - • * • • • • • • • • 0
0 0.1 0.2 0.2 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Cutoff Frequency ( U)„L )
Figure 3.8: The variation of average BER in extracted watermark and ^WRa with
increasing normalized cutoff frequency 6;„/ after LPF attack for single-level
watermarking.
1 —•—Level-I
0.9 : Level-ll
0.8 I ••^-- Level-Ill i
0.7 '•
0.6 :
^ 0.5 i CO
0.4
0.3
0.2 !
i 0.1 i
0 I A • A • • k: • • -k'
0 0.2 0.4 0.6 0.8 1
Normalized Cutoff Frequency ( Wn^)
Figure 3.9 Watermark extraction performances in terms of BER with increasing
normalized cutoff frequency r ),,, for low-pass filtered three-level watermarked audio
signals.
48
Cfutpter3
30
25
20
-a J 15
10
0 ^
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Normalized Cutoff Frequency ( WnL)
Figure 3.10 Variation of SWRci with w^^i for low-pass filtered three-level
watermarked audio signals.
High-Pass Filtering
A high-pass FIR digital filter of order 50 is applied to attack the single-le\el
and multi-level watermarked audio signals. Average BER in extracted watermark and
SWRu for different values of normalized lower cutoff frequency ((o, ^ ) of high-pass
filter (HPF) was evaluated for single-level audio watermarking and results are shown
in Figure 3.11. A plot of average BER in extracted watermark for all the three levels
of watermarks (in multi-level watermarking) vs. (y„„ is shown in Figui-e 3.12.
Variation of SWRa of high-pass filtered multi-level watermarked audio signal with
increasingco^^^ is shown in Figure 3.13.
It is evident from the Figure 3.11 that forr/; ^ up to 0.04 (1764 Hz), the
watermark from the single-level watermarked audio signal is fully recovered without
any error because the chirp-based watermark embedded is of low frequency (below
100 Hz) and the HPF has smooth transition from stop-band to pass-band. The
embedded chirp-based watermark is not removed but it is attenuated at lower values
of cutoff frequencies of HPF. However, for higher cutoff frequencies, errors start to
occur due to embedded watermark gets severely attenuated. The chirp signal can be
49
Cliapter3
extracted from a high noise background. As the cutoff frequency (a)„„) of HPF
increases, the attenuation in the watermarked frequency band also increases, causing
more bit errors and lowering the SWRa of watermarked audio signal.
Due to the same reason as above, SWRa of high-pass filtered multi-level
watermarked audio signal also decreases. In multi-level (three-level) audio
watermarking, first level of watermark can be extracted without any error up to w,,,/ =
0.02 (882 Hz) due to the watermark being at low frequency band (10-60 Hz) and of
small chirp duration (0.25 second). The second and third levels of watermarks are
extracted without any error up to fi;„„ = 0.06 (2446 Hz) as they are at very low
frequency bands (10-30 Hz for second and 10-15 Hz for third levels watermarks) and
of larger chirp durations (Isec ibr second and 1.5 sec for third level of watermarks).
100
80
a: [11
60
40
20
0 —
0 0.01 0,02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Normalized Cutoff Frequency ( WnH)
Figure 3.11 Plots of average BER in extracted watermark and SWRa vs. «„„ after
HPF attack for single-level watermarking.
50
CHapter3
100
80
UJ
m
60
40 !
20
0 I
0
-•—Level-I
Level-ll
A - Level-lll
0.01 0.02 0.03 0.04 0.05 0.06 0.07
Normalized Cutoff Frequency ( W^H )
0.1
Figure 3.12 Variation of average BER in extracted watermar]< witli increasing &i,,/,
after HPF attack for three-level watermarked audio signals.
0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
Normalized Cutoff Frequency (
0.09 0.1
Figure 3.13 Variation of average SWRa with increasing a*,,/ for high-pass filtered
three-level watermarked audio signals.
Band-Pass Filtering
A band-pass FIR digital filter of order 50 with varying normalized lower
cutoff frequency (fc' j/,) and a fixed normalized upper cutoff frequency (ft*,,,/,) of 0.9
51
Cluipter3
is designed to attack the watermcirked audio signals. The band-pass fiUer (BPF) used
is similar to HPF that was used previously except upper cutoff frequency (&>„,;,) is
kept 0,9 (39690 Hz). In this BPF, lower cutoff frequency («„,;j) is varied from 0.01
(441 Hz) to 0.09 (3969 Hz). Plots of average BER vs. co,^,, and SWRa vs.r ,,, for
single-level audio watermarking are shown in Figure 3.14. A plot of variation of
average BER for all the three levels of watermarks (multi-level audio watermarking)
for increasing w,,,, is shown in Figure 3.15. Variation oi SWRu of band-pass filtered
multi-level watermarked audio signals with increasing (J»„,/, is shown in Figure 3.16.
It is observed from simulation result that the embedded watermark can sustain the
limited BPF attack up to the &>„,„= 0.05 (2205 Hz) for single-level watermarked audio
signal.
The same reasoning (as high-pass filtering discussed earlier) is applicable
here. Due to long chirp durations for second and third levels of watermarks, the lower
BER in watermark extraction occurs. This also shows that the multi-level
watermarking has different levels of robustness. Hence, this can be used to embed
information having different levels of security information into a host audio signal.
100 „ „ 25
80 • / I 20
60 \ / 15 m IT \ / •°
DO \ \ a. 40 \ / 10 w
20
0 • • -0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Normalized Lower Cutoff Frequency ( u)„,e)
Figure 3.14 Variation of average BER in extracted watermark and SWRa with
increasing «„!„ after BPF attack for single-level watermarking.
52
chapter 3
a. Ml m
100
90
80
70
60
50
40
30 •
20
10
0
0
A-
0.01
-•—Level-I Level-ll
A •• Level-lll
0.02 0.03 0.04 0.05 0.06 0.07 0.08
Normalized Lower Cutoff Frequency ( Wnie )
0.09 0.1
Figure 3.15 Watermark extraction performances in terms of average BER with
increasing w,,,,; after BPF attack for three-level watermarking.
0.03 0.04 0.05 0.06 0.07
Normalized Lower Cutoff Frequency (
0.09
Figure 3.16 Plot of average SWKa vs. ft;,,,/, after BPF attack for three-level
watermarking.
Up-Sampling by Interpolation
l.'p-sampling by interpolation factor (.V„) is used as an attack to the
watermarked audio signals and is performed by inserting zero samples in the original
signal and then low-pass filtering it. The LPF is used to remove the unwanted images
generated in the spectrum of zei-o-samples inserted signal. Therefore, this filter is also
53
CfiapterS
known as anti-imaging LPF. The plots of Average BHR vs. A',, and SWRa vs. A'„ for
single-level watermarking are shown in Figure 3.17. A variation of average BER of
three levels of watermarks with increasing interpolation factor A',, for three-level
audio watermarked signal is shown in Figure 3.18. Variation of SlVRa of up-sampled
three-level watermarked audio signal with increasing interpolation factor A',, is shown
in Figure 3.19. It is observed from the Figures 3.17, 3.18 and 3.19 that this scheme is
robust to up-sampling by interpolation for both single-level and multi-level
watermarking. By up-sampling using interpolation, no spectral distortion is
introduced. Therefore, the watermark can be extracted without any error and SlVRa
remains unaltered for interpolated watermarked audio signals and is equal to SWR for
watermarked audio signal. Due to interpolation by some integer factor, the sampling
frequency is multiplied by that interpolation factor (up-sampling), and then this up-
sampled signal is passed through a LPF to reject unwanted images produced by
interpolation process. Therefore, there is negligible spectral distortion in interpolated
signal. The same reasoning is true for three-level audio watermarking scheme.
0.9
0.8
0.7
1 - • - B E R , 35 - • — SWRa
30
25
: 20 I LU 0.5 ra CD Q : 1 5 $
CO
10
5
0.4
0.3
0.2
0.1
0 • • • • • • • • • " ' i 0 1 2 3 4 5 6 7 8 9 10 11
Interpolation Factor ( Nu)
Figure 3.17 Variation of average BER in extracted watermark and SWRa with
increasing interpolation factor A ,, for up-sampled single-level watermarked audio
signals.
54
1
0.9
0.8
0.7
0.6 o: UJ 0.5 m
0.4
0.3
0.2
0.1
0 -A-
2
-A-
3
-A-
4
- • - Level-I
Level-ll
-A-Level- I l l
-A-
5
-A-
6
-A-
7
Interpolation Factor ( N ,)
-A-
8
CdapterS
-A-
9 10 11
Figure 3.18 Watermark extraction performances in terms of average BER with
increasing interpolation factor A',, for up-sampled three-level watermarked audio
smnals.
30
25
- 20 03
TO 1 5 on
w 10
5
0
4 5 6 7 8
Interpolation Factor ( Nu)
10 11
Figure 3.19 Variation of average SWRa with increasing interpolation factor iV„ after
up-sampling of tiiree-level watermarked audio signals.
55
CdapterS
Down-Sampling by Decimation
In order to evaluate the robustness to down-sampling attack, a decimator with
decimation factor A;, was designed. Down-sampling by decimation is performed by
pre-alias low-pass filtermg followed by down-sampling. This pre-alias low-pass filter
eliminates the spectrum of X(w) in the range nlNj< ro< TI. Average BER in extracted
watermark and SWRa vs. A' , plots for single-level watermarking are shown in Figure
3.20. Watermark extraction performance in terms of average BER of three levels of
watermarks with increasing decimation factor .V,y is shown in Figure 3.21. Variation
Q^ SWRa of the three-level watermarked audio signals with increasing A' , is shown in
Figure 3.22. It is evident from the Figures 3.20. 3.21 and 3.22 that single-level and
multi-level (three-level) audio watermarking schemes are robust to down-sampling by
decimation, but SWRa decreases with increasing A',, due to low-pass filtering which is
performed before down-sampling. Although the spectrum is spread during decimation
process, but the watermark is preserved because it lies in a very low frequency band
(below 100 Hz). In down sampling by decimation, SWRa decreases on increasing .-V,
due to increase in removal of frequency components in this process. This is true for
both single-level and multi-level (three-level) audio watermarking schemes.
1 30
0.9 •BER 0.8 I ^ _»_SWRa
0.7 i
25
i 20 0.6 ^ ~ ^ ^ ^ S
a:
0.4 ^ 5 10
5
0.3
0.2
0.1
0 -• • • • • • • • •• 0 1 2 3 4 5 6 7 8 9 10 11
Decimation Factor ( Nd)
Figure 3.20 Variation of BEFL in extracted watermark and SWRa with increasing
decimation factor Nj for down-sampled single-level watermarked audio signals.
56
Cfiapter 3
1
0.9
0.8
0.7
0.6 a: UJ 0.5 CQ
0.4
0.3
0.2
0.1
0
• Level-I Level-ll
•Level-Ill
-A-
2
-A-
3
-Ai-
4
-A-
5
-A-
6
-A-
7
-A-
8
-A-
9 10 11
Decimation Factor ( N j)
Figure 3.21 Watermark extraction performances in terms of average BER with
increasing interpolation factor N^ after down-sampling of three-level watermarked
audio signals.
25
20
CQ 1 5
•D
« § 10
5 6 7 8
Decimation Factor ( N^ )
10 11
Figure 3,22 Variation of average SWRa with increasing decimation factor Nj for
down-sampled three-level watermarked audio signals.
57
Cfiapter3
Re-Sampling
Re-sampling a signal by some arbitrary rational factor (A',. = P/Q) is
equivalent to up-sampling (interpolation) by integer factor (P) followed by down-
sampling (decimation) by another integer factor (Q). Re-sampling by varying
arbitrary rational factor (N,.) from 0.1 to 2 in step size of 0.1 is applied to attack the
single-level and multi-level (three-level) watermarked audio signals. Plots of Average
BER vs. A', and SWRa vs. PI,, for single-level watermarking are shown in Figure
3.23. A variation of average BER of three levels of watermarks vs. N,. for multi-
le\el watermarking scheme is shown in Fig. 3.24. A plot of SWRa of the re-sampled
multi-level watermarked audio signals vs. iV,. is shown in Figure 3.25. It is evident
from the Figures 3.23 and 3.24 that the single-level and multi-level audio
watermarking schemes are robust to re-sampling. If P < Q, in this case, on increasing
the value of -V,., the SWRa increases up to SWR of watermarked audio signal. This
case has the performance same as that of purely down-sampling by decimation. If P >
Q. in this case, on increasing the value of A',.. the SWRa remains constant. This case
has the performance same as that of purely up-sampling by interpolation. In both
cases, watermark can be extracted without any error due to watermark is of very low
frequency (below 100 Hz). The results of multi-level audio watermarking scheme can
be explained by using above given reasoning that were used for single-level audio
watermarking scheme.
35
30
25 S"
20 E n
15 i
10
5
0 0.1 0.3 0.5 07 0.9 1.1 1.3 1.5 1.7 1.9 2.1
Re-Sampling Factor ( Nr)
Figure 3.23 Variation of average BER in extracted watermark and SWRa with
increasing A'',, after re-sampling of single-level watermarked audio signals.
58
CfiapterS
1
0 9 — • — Level-I Level-ll
• A - Level-Ill 0.8
0.7
0 . 6 DC ^0.5
0.4
0.3
0.2
0.1 ; i
0 ' A - -A- -A- -A- -A- -A- - A - - A - -A- -A- -A- -A- -A- -A- -A- - A- - A- - A- - A- - A 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Re-Sampling Factor (N^)
Figure 3.24 Watermark extraction performance in terms of average BER with
increasing A',, for re-sampled tliree-level watermarked audio signals.
0
0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9
Re-Sampling Factor (N,.)
Figure 3.25 Variation of average SWRa with increasing N,. after re-sampling of three-
level watermarked audio signals.
Addition of White Gaussian Noise
Additive white Gaussian noise (AWGN) of varying noise power is injected
into the watermarked audio signals to give Signal-to-Noise Ratio (SNR) in the range -
59
Cfiapter3
5 dB to 50 dB. Two plots of average BER vs. SNR and SWRa vs. SNR for single-
level watermarking scheme are shown in Figure 3.26. Watermark extraction
performance in terms of average BER with increasing SNR for three levels of
watermarks is shown in Figure 3.27. Another plot of SWRa of AWGN-injected multi
level watermarked audio signal vs. SNR is shown in Figure 3.28.
In single-level audio watermarking, as the SNR increases (i.e. noise decreases)
SWRa reaches its maximum value of SWR (i.e. 30 dB), in this case, noise power is
negligible. Therefore, BER in extracted watermark is also zero for higher SNR. At
SNR= 10 dB, SWRa is also approximately equal to 10 dB. As the SNR decreases (i.e.
noise increases) the noise starts to dominate and SWRa becomes equal to or less than
SNR.
In multi-level audio watermarking as the results given in Figures 3.27 and
3.28, for the first, second and third levels of watermarks, BER is zero at SNR = 20
dB. SNR= 0 dB and SNR= -5 dB, respectively. The second and third levels of
watermarks are more robust against AWGN since these watermarks are of very lower
frequency band and have longer duration as compared to first level of watermark.
a: hi CD
10 15 20 25
SNR (dB)
30 35 40 45 50
Figure 3,26 Variation of average BER and SWRa with increasing SNR after adding
white Gaussian noise of varying power in single-level watermarked audio signals.
60
CAapterS
a: LU m
^^^=B=c
«—Level-I «—Level-l l
Level-
10 15 20 25
SNR (dB)
30 35 40 45 50
Figure 3.27 Plot of the variation of average BER in extracted watermark with
increasing SNR for attaclced three-level watermarked signals by AWGN.
m
TO
a. CO
30
20
10
- • • •
-10 10 20 30 40 50
-10
SNR(dB)
Figure 3.28 Variation of average SWRa with increasing SNR for attacked three-level
watermarked audio signals by AWGN.
Amplitude Scaling
Amplitude scaling of single-level and multi-level (three-level) watermarked
audio signals is performed for different scaling factor (A). Average BER and SWRa
61
Cfiapter3
vs. amplitude scaling factor [A) plots for single-level watermarking are shown in
Figure 3.29. A plot of average BER vs. amplitude scaling factor (A) for multi-level
watermarking is shown in Figure 3.30. A variation of SWRa of amplitude-scaled
multi-level watermarked audio signal with increasing amplitude scaling factor (A) is
shov\n in Figure 3.31. It is observed from the Figures 3.29, 3.30 and 3.31 that average
BER in extracted watermark is zero for single-level and three-le\'el watermarking
schemes. Therefore, chirp-based audio watermarking scheme for single-level and
three-level is robust to amplitude scaling. Amplitude scaling of the watermarked
audio signal does not result into spectral modification. Therefore, there is zero BER in
the e.xtracted watermark and SWRa remains unaltered, i.e. fixed at 30 dB for single
level watermarking. The same reasoning is applicable to multi-level watermarking as
above.
1 35
0.9 -4 30
0.8
0.7 25 --•—BER ^
0-6 _,^SWRa I 20 ^
^ 0.5 ; -15 °^ m
0.4 ; '^ § en
0.3 10 0.2
: 5
0.1
0 Itti ««>««« « * • • • • * • • 0
0 1 2 3 4 5 6 7 8 9 10 Amplitude Scaling Factor ( A)
Figure 3.29 Variation of average BER and SWRa with increasing amplitude scaling
factor (A) for amplitude-scaled single-level watermarked audio signals.
62
chapters
1
0.9
0.8 - • - Level-I
Level-ll
0.7 -Ar-Level-
0.6 a: uj 0.5 m
0.4
0.3
0.2
0.1
0 iitiitll l i A A • • • • A A-
2 3 4 5 6 7 8 9 10
Amplitude Scaling Factor ( A )
Figure 3.30 Watermark extraction performances in terms of average BER with
increasing amplitude scaling factor (A) for amplitude-scaled three-level watermarked
audio signals.
30
25 MIMM»M • • • • • • • • •
_ 20
(0 10 :
0 1 2 3 4 5 6 7 8 9 1 0
Amplitude Scaling Factor ( A)
Figure 3.31 Variation of average SWRa with increasing amplitude scaling factor (A)
for amplitude-scaled three-level watermarked audio signal.
63
Cluipter3
MP3 Compression
This is one of the simplest and most common attacks applied to the audio
watermarked signals. Since the proposed digital audio watermarking is carried out on
uncompressed audio, therefore, it is necessary to evaluate the robustness towards
compression attacks. Simulation experiments were carried out for a wide range of
MPEG-1 Layer-3 (MP3) compression attacks having Constant Bit Rate {CBR)
ranging from 56 kbps to 320 kbps. Plots of average BER in extracted watermark and
SWRa vs. Constant Bit Rate [CBR) of MP3 compression for single-level audio
watermarking are shown in Figure 3.32. A variation of average BER in extracted
v\atermark with increasing CBR of MP3 compression for multi-level watermarking is
shown in Figure 3.33, A plot of SWRa of MP3-compressed three-level watermarked
audio signal vs. CBR of MP3 compression is shown in Figure 3.34. It is evident from
the Figures 3.32 and 3.33 that single-level and multi-level watermarking schemes are
robust to MP3 compression (because the BER=0). As the bit rate of MP3 compression
algorithm is reduced, higher compression ratio is obtained. By reducing the bit rate of
MP3 compression, more redundant information from the audio signal is eliminated.
The SWRa decreases with reduction in the bit rate (increasing compression ratio) of
MP3. This is because of the fact that MP3 removes less significant information from
the audio to achie\e higher compression (i.e. lower bit rate). The watermark is
correct!}' extracted from the iMP3-compressed watermarked audio signal for MP3 bit
rate from 56 kbps to 320 kbps. For multi-level watermarking, BER in extracted
watermark is zero for all the three levels of watermarks for MP3-compressed audio
signals. SWRa of MP3-compressed multi-level watermarked audio signals varies in
the same fashion as for MP3-compressed single watermarked audio signals.
64
Cfiapter3
1 30
0.9
0.8 25
•BER 0-7 r " ^ -B-SWRa 20
0.6 S a: ^ S 0.5 ; 15 ^
CO
10 0.4
0.3
0-2 5
0.1
0 - • - • — • — • — • — « • • • • • -• 0
50 100 150 200 250 300 350
WIP3 Bit Rate (kbps)
Figure 3.32 Plots of variation of average BER and SWRa with increasing Constant
Bit Rate (CBR) for MP3-compressed single-level watermarked audio signals.
1
0.9
0.8
0.7
0.6
0.4
- •— Level-I Level-ll
•A - Level-Ill
0.3 ;
0.2 !
0.1
0 A - A • • A- • A • - -A - • A A A A A A
50 100 150 200 250 300 350
MP3 Bit Rate (kbps)
Figure 3.33 Watermark extraction performances in terms of average BER with
increasing CBR for MP3-compressed three-level watermarked audio signals.
65
CdapterS
0 L
50 100 150 200 250 300 350
MP3 Bit Rate (kbps)
Figure 3.34 Plot of variation of average SWRa with increasing CBR for MP3-
compressed three-level watermarked audio signals.
Cropping Attack
Cropping attack performance for single-level watermarking is shown in Figure
3.35. In cropping attack, some of the samples are removed from the end of the
watermarked audio signal; BER increases on increasing the percentage crop (ratio of
number of samples cropped to total number of samples in watermarked audio signal)
of the audio signal but SWRa decreases on increasing the percentage crop. If the
signal is cropped from the beginning or middle of the audio file, then the watermark is
not detectable from the point beyond which the signal has been cropped, this is due to
the offset problem of samples. Since the correlation of the chirp and the watermarked
audio is carried out on frame by frame basis, therefore, cropping will cause an
alignment problem of the chiip with the frame. This will result into errors during the
watermark extraction. If cropping of samples from an audio signal is done from the
beginning or middle of the signal then delaying of remaining samples take place
which is known as offset problem. Due to offset problem in cropping, watermark
extraction is not possible.
66
chapter 3
30
25
20 _ m T3
1 5 ^
10
5
0
CO
100
Figure 3.35 Plot of variation of average BER in extracted watermark and SWRa with
increasing the number of cropped samples of single-level watermarked audio signals.
3.7 Summary
The chapter presented single-level and multi-level chirp-based digital audio
watermark embedding and extraction. Owing to the correlation property of the chirp
signal, the chirp-based digital audio watermarking can be extracted from a high noise
background. Both single-level and multi-level audio watermarking schemes are
found to be robust to attacks such as low-pass filtering, interpolation, decimation, re
sampling, addition of white Gaussian noise (AWGN), MP3 compression, and
amplitude scaling. The chirp-based digital audio watermarking schemes (single-level
and three-level) show limited robustness to high-pass filtering and band-pass filtering
attacks. It can be said that the payload in three-level watermarking has been increased
by 39.68 % of single-level watermarking while maintaining imperceptibility of audio
signal. Multi-level chirp-based digital audio watermarking has various degree of
robustness at different levels. Hence, it can be used to embed information having
different levels of security into a host audio signal.
67
CHapter 4
(Digit aC Image Watertnar^ng in (D^c^T (Domain
m
Cfiapter4
DIGITAL IMAGE WATERMARKING IN DFRFT
DOMAIN
4.1 Introduction
In the previous chapter the development of chirp-based multi-level
watermarking techniques for audio signals was presented. The present and next
chapters deal with embedding of the watermark data in images. Due to vast amount of
data inherent in an image, a large amount of information can be embedded into them.
Though in images, embedding can be performed in any one of the following three
domains, spatial (image) domain, frequency (transform) domain and in compressed
bit-stream domain. But the work presented in this thesis is focused on the
development of watermark embedding technique in images transformed using
Discrete Fractional Fourier Transform (DFRFT) (in this chapter) and using the
combination of DFRFT and discrete wavelet transforms (DWT) (in the next chapter).
The chapter is organized as follows. It starts with background and review of
existing works in related area. Then, algorithms for watermark embedding in DFRFT
domain and extraction of watermark data are presented. The next section briefly
describes various parameters used for performance measurement, followed by the
simulation results and discussions. Finally the chapter is summarized at the end.
4.2 Background
Digital watermarking is a technology that is used in copyright protection,
source authentication and integrity of network [21] [60] [105]-[109]. The available
image watermarking techniques can be broadly categorized in three groups: spatial
(pixel)-domain watermarking techniques [28] [107], frequency (transform) domain
techniques [84]-[85], [108]-[11(3] and compression-domain techniques [117]-[120].
The spatial-domain watermarking algorithms directly embed watermark information
into a set of pre-selected pixels in images and/or selected frames of videos [28].
Although these techniques are simple, but have shortcoming that they are not robust
enough to various image processing operations. In transform- (frequency) domain
image watermarking, the original image is first transformed into the frequency-
68
CHapter 4
domain, and then the transformed coefficients are modified by the watermart:
information. The commonly used transforms for this purpose are Discrete Fourier
Transform (DFT) [108]-[110], Discrete Cosine Transform (DCT) [111]-[112],
Discrete Wavelet Transform (DWT) [113]-[114], and Discrete FRFT (DFRFT) [84]-
[851, [115]-[116] or any other that can be used for transforming digital media for the
purpose of digital watermarking in transform-domain. In recent years it has been
recognized that watermark embedding in transform domain is much more robust than
spatial-domain/ time-domain vv'atermarking. A major difficulty in watermarking in a
transform domain is that there is no accurate control on the allowable distortion at any
pixel in the spatial domain. Transform-domain watermarking methods are more
complicated and require more computational power. In compression-domain
watermarking schemes, the watermark bits are either directly inserted into the
encoded bit-stream, or in a partially decoded bit-stream. Benefits of these schemes are
lesser burden of decompressing and re-compressing the content and avoiding further
perceptual degradation of the host content [117]-[120]. Out of three, the transform-
domain watermarking algorithms attract more attention of recent study for their
excellent stability and robustness [113], [121].
The works presented in this chapter are mainly deal with watermarking in
transform domain in general and Discrete Fractional Fourier Transform (DFRFT) in
particular. Despite its many rich features, DFRFT has not fully explored for the
watermarking purpose. Most of the existing works on DFRFT-based watermarking
schemes were concentrated on the embedding and error detection of watermark data
in DFRFT-domain [84]-[85], [115]-[116], but to the best of the author's knowledge,
no work has been reported on the extraction of watermark data embedded in DFRFT-
domain. This chapter introduces a novel scheme for non-blind extraction of the data
embedded in the DFRFT-domain.
A digital image watermarking in DFRFT-domain was first studied by Djurovic et al.
[115], in which, original image was transformed by DFRFT, and transformed
coefficients were modified by the watermark information. This method can be used
only for watermark detection and is robust against lossy compression attacks up to
compression ratio of only 50%. In [116], multiple chirps were used as watermark
which were embedded in the spatial-domain directly, and detected in the DFRFT-
69
Cliapter4
domain. This watermarking technique can be used for watermark detection of chirp
signals only. In [85], a chirp signal was used as a watermark and was embedded in
DFRFT-domain of the image. Due to impulsive nature of chirp signals in DFRFT-
domain, the presence of watermark can be detected conveniently. In [84] spatial chirp
signal as a watermark was embedded into low frequency band of wavelet transform of
host image. The watermark detection was carried out in DFRFT-domain, without the
need of the original image. A common problem of all these previous works [84]-[85],
[115]-[116] is that they were developed only for watermark detection. These schemes
are useful only for authentication purpose, where presence of watermark can be
verified. However, in many applications such as copyright verifications, it is
necessary to extract the embedded data. The vvork presented in this chapter is
different from previous works in the sense that it proposes a watermark extraction
scheme in DFRFT-domain, which is the main contribution of this chapter.
In this chapter, an algorithm to extract the watermark data embedded in
DFRFT-domain is proposed. One of the advantage of using DFRFT as a
transformation tool for watermark embedding is that by using varying powers of 2D-
DFRFT, multiple watermarks (i.e. higher payload) can be embedded in comparison to
other image watermarking techniques such as those based on DFT and DCT. The
DFRFT powers and watermark location can be used as part of secret keys for such
type of watermarking techniques. By changing watermark location, it is possible to
embed different watermarks of varying watermark lengths. Therefore, higher
embedding capacity and higher degree of security are two attractive features of
DFRFT-domain image watermarking.
4.3 Digital Image Watermarking using DFRFT
In this section, a digital image watermarking technique using 2D-DFRFT is
presented. As mentioned, most of the earlier works on image watermarking using
DFRFT were concentrated on embedding and watermark detection (rather than
extraction) [84]-[85], [115]-[116]. The novelty of the present work is the development
of an algorithm for extraction of watermark data embedded in DFRFT transformed
images. Both embedding and watermark extraction algorithms are presented in
following subsections.
70
Cfiapter4
4.3.1 Watermark Embedding
The watermark embedding algorithm using 2D-DFRFT presented here which
is similar to that demonstrated in [115]. The technique has been developed to embed a
binary image (as watermark data) into a host image. The block diagram of the
watermark embedding scheme is shown in Figure 4.1. In this scheme, the host image
is first transformed using 2D-DFRFT with a fixed fractional transformation power
pair (pi, P2). A binary image (watermark image) of size 32*32 pixels (having total of
1024 bits of data) is considered as watermark data. It should be ensured that the size
of watermark data is less than the number of DFRFT coefficients of host image. In
order to embed the watermark image, a set of DFRFT coefficients of host image are
selected, which are altered depending upon corresponding watermark data. To select
the set of DFRFT coefficients, the complex-valued DFRFT coefficients are sorted in
the ascending order of their magnitude, then required number of coefficients (equal to
length of watermark data) starting from an arbitrary position are selected. The starting
position of watermark embedding may be considered as a variable and is the part of
watermark key. The watermark should not be embedded in the coefficients with
smaller magnitudes because that would make it very sensitive to noise removal or
compression operation, but it should neither be embedded in the coefficients with
very large magnitudes because that would disturb the image too much [115]. The
watermark data (after scaling) is then added to the selected DFRFT coefficients (in
order). The modified array of DFRFT coefficients are then re-arranged at their
original positions and the inverse DFRFT of which produces the resulting
watermarked image.
The selection of fractional transformation power 'p' is an important parameter
in the embedding process. Since the DFRFT for 0<p<2 has the ability to describe the
image or signal information in the time (or spatial) as well as in frequency domains. If
p is chosen close to 0, the DFRFT can be considered as the transformation being
dominantly in the time-domain. On the other hand, if p is chosen close to 1, the
DFRFT is dominantly in the frequency-domain. Since the present work deals with
watermark embedding in the frequency domain of the DFRFT, therefore p is
considered as close to unity. Also, the starting position of embedding is chosen
towards higher magnitude coefficients of the host image. Another factor of interest is
71
Cliapter4
scaling factor, which is simply the multiplication of amplitude of watermark data.
Watermark scaling is done to achieve imperceptible watermarking and to extract
watermark data correctly. Watermark
Data Watermark Scaling
2D-DFRFT z Re-ordering of DFRFT Coefficients in
ascending order
z. Watermark Embedding
/"'
z;
inverse 2D-DFRFT
Ph =[-Pr-P2]
Figure 4.1 Block Diagram of DFRFT-based Watermark Embedding Scheme.
In order to explain the embedding process mathematically, let the host image
is 7 ' and 'w' is the watermark binary image. Z denotes the complex-valued 2D-
DFRFT of image /, i.e. Z=DFRFT (I) = X+JY, where X and Tare real and imaginary
parts of Z respectively. These DFRPT coefficients are sorted according to their
magnitude in ascending order and the i"' element of sorted array Z is denoted
as Z [/] = X^ [i] + j Y^ [i]. The watermark w is embedded into the selected host DFRFT
coefficients ZJ/], i = L + l,...,L + M, where L is the index in sorted array Z, from
where the watermark embedding will be started and M is the watermark length. After
embedding, the modified DFRFT coefficients are stored in array Z" . whose elements
are obtained as
z :'['] = ' Z J / ] + w [ ; - i , + 1], // L + l<i<L +
ZAi] otherwise (4.1)
72
CHapter4
Then the modified array Z" is rearranged with original index in the 2D-array
Z" and the watermarked image is obtained by computing the inverse DFRFT ofZ".
This watermarked image is denoted by/" . The embedding process described above is
also depicted in Figure 4.1. Parameters used to control the watermark embedding are:
DFRFT powers (p), watermark location (/,) and watermark length (M). These
parameters are the part of keys generated in the embedding process and will be used
for watermark extraction.
4.3.2 Proposed Watermark Extraction
As mentioned earlier, most of the existing works on DFRFT-based image
watermarking schemes [84]-[85], [115]-[116], were concentrated mainly for detection
of watermark data only. These schemes are useful only for the authentication purpose
where presence of watermark need to be verified. However, in many applications such
as signature or copyright verifications, it is necessary to extract the embedded data.
To the best of our knowledge, no attempt is made to extract the watermark data
embedded in DFRFT domain. This research proposes an algorithm for extraction of
watermark data from watermarked image, which was embedded in DFRFT domain.
The block diagram of the proposed watermark extraction technique is shown
in Figure 4.2. It is basically a non-blind approach in which the knowledge of the
original host image is required. The watermarked and host images are first
transformed by using 2D-DFRFr with the same fractional power (/ ^ = [Pi^Pj]) as
used during the embedding process. The DFRFT coefficients of the watermarked and
host images are then sorted in ascending order, similar to the ordering performed
during embedding. The ordered DFRFT coefficients of host image are then subtracted
from the ordered DFRFT coefficients of watermarked image. The element-wise
subtraction from index L+1 to L+M results the extracted watermark data w /-
In order to explain the extraction procedure further, let 7 ' and ' /" ' ' represent
the host image and the received v\/atermarked image, respectively. Also let,
• Z denotes the complex-valued 2D-DFRFT of image /, i.e. Z=DFRFT (I) =
X+jY, whereXand Fare real and imaginary parts of Z, respectively;
73
Cliapter4
• Z'" denotes the complex-valued 2D-DFRFT of /'" . i.e.
Z'"=DF/?F7'(7™)=.\'"•+/}"'"•, where X''"and Y'''' are real and imaginary
parts of Z'" , respectively;
• Z, is an array containing elements of Z arranged in the ascending order of
their magnitudes;
• Zy" is an array with elements of Z"" arranged in ascending order of their
magnitudes;
• ZJ/] and Z™[/] are the i' elements of sorted arrays Z^ and Z[^\
respectively.
Then the extracted watermark from the received watermarked image is given by
u-, .,[/] = Z™[/]-Z,.[/] jor L + l<i<L + M (4.2)
/ ^
2D-DFRFT
Ph =[P,-P2] .
z Sorting of
DFRFT Coefficients of Host Image
Zs
2D-DFRFT
Ph ^[P^^Pi]
Z"'' Sorting of
DFRFT Coefficients of Watermarked
Image
z';V]-ZM for
L + \<i<L + M
^"exl
Figure 4,2 Block Diagram of Proposed Watermark Extraction.
Due to the different attacks on the watermarked image, it may happen that the
extracted watermarks are different than the original embedded watermarks. Also, due
to attacks and arithmetic subtraction, the extracted watermarks Wgxt may be multi
valued, in contrast to' two-valued original watermarks. In order to compare the
robustness of the proposed technique, it is necessary to convert extracted watermark
into bi-level (to reconstruct the binary images of watermark) samples through hard
74
Cfiapter4
thresholding. That is, the extracted watermark samples are compared to a threshold
'Th' (which is half of the scaling factor used in embedding). Let ' w'' is the binarized
extracted watermarks, which is obtained as follows;
1, ;/• H;,„ >Th ,^ ^.
0, otherwise
where, 'Th' is the threshold. The numbers of erroneously extracted watermark bits are
calculated by performing bit-wise XOR operation between original watermark bits
( w ) and extracted watermark bits (w ), as follows:
w ®w =\ , (4.4) 0, VI' is not in error
Bit Error Rate (BER) is defined as
BER = ^xl()0 % (4.5)
where, A'\ is the total number of erroneously extracted watermark bits and N is the
total number of watermark bits.
4.4 Simulation Results and Discussion
In order to evaluate the performance of the DFRFT-based watermarking
technique described previously, five standard images {Lena Boats, Goldhill, Peppers
and Baboon) are considered as host image in which watermark data is to be
embedded. All these are 8-bits gray-scale images of size 512*512 pixels. The pixels
of a binarised (1 bit/pixel) signature image of size 32*32, 'fh.bmp' are used as
watermark data. The parameters used for watermark embedding and extractions are as
follows:
• Watermark location (L) = 259000,
• Watermark length (M) = 32*32 = 1024,
• Scaling factor for watermark data = 44,
• DFRFT powers used p,, --= [p^,p^]= [0.99, 0.99] and
75
•
Cfiapter4
Threshold value for binEirization of extracted watermarks, Th = scaling
factor/2=22.
All results of DFRFT-based embedding and extraction method are compared
with a DWT-based image watermarking method proposed by Kundur and Hatzinakos
[861.
4.4.1 Performance Measuring Parameters
The performance of the proposed image watermarking algorithm is measured
in terms of Peak Signal-to-Noise Ratio (PSNR) of watermarked image and Bit Error
Rate (BER) in watermark extraction. These parameters are defined as follows:
Imperceptibility
One of the important features of digital watermarking is that the perceptual
quality of the watermarked image should not degrade much as compared to original
one. A better watermark algorithm should be such that the watermarks do not create
visible artifacts in the host image. Imperceptibility of watermarked images is
evaluated by measuring its PSNR. Imperceptibility may also be checked by visual
inspection of watermarked image and comparing with original host image.
Peak Signal to Noise Ratio (PSNR)
PSNR is used to measure the objective quality of watermarked images. The
host image is considered as a reference. The value of PSNR is the measure of
similarity between watermarked and host images. It is measure of how much invisible
the embedded watermark is. This property is known as imperceptibility property. The
watermark image with PSNR value of more than 50 dB is as good as the original
image. It is defined as:
^ 5 ^ = 10 log. MSE
.dB (4.6)
Where, MSE is the mean square error between watermarked and host images. MSE is
defined as
76
Cfiapter4
MSE 1
A/ ,V
M X N , , .v=l >•=!
XZ[p( '>)-/''( '-V)f ^ .K ' I' []Lj a / v i V (4.7)
where, p(x, y) and p'(x, y) are pixels of host and watermarked images at (x, y)
coordinates, respectively.
Bit Error Rate (BER)
The performance of extraction algorithm is measured in terms of percentage
bit error rate (BER) of the extracted watermark data. The BER in watermark
extraction is defined as follows (see Eqn. 4.8).
BER = --^yJOO % (4.8)
Where, ,V is the number of erroneously extracted watermark bits and N is the total
number of watermark bits.
4.4.2 Imperceptibility Performance
In order to measure the imperceptibility feature of our DFRFT-based image
watermarking scheme, the binary image 'fli.bmp' is embedded into each of five
standard host images (Lena, Boats, Goldhill, Peppers and Baboon). The five
watermarked images are obtained corresponding to the original five host images. The
DFRFT-based watermarking scheme is compared with existing DWT-based Kundur's
image watermarking scheme in terms of perceptual quality. Original host images and
watermarked images for DFRFT-based method and DWT-based Kundur's method are
shown in Table 4.1 along with corresponding PSNR values. It can be observed that
PSNR of watermarked images obtained by applying DFRFT-based image
watermarking scheme, is more than 50 dB for all the five test images. Whereas the
PSNR values for DWT-based Kundur's watermarking method is approximately 45 dB
for all the five images. Therefore, it can be said that our DFRFT-based watermarking
scheme gives better fidelity than DWT-based Kundur's watermarking method.
Farther, comparing the subjective quality of v/atermarked images with corresponding
original host images (also given in Table 4.1), it can be observed that watermarked
77
Cliapter4
images and original host images are almost similar and no artifacts are visible in the
watermark image. Therefore the proposed watermark algorithm maintains the
imperceptibility feature.
Table 4.1 Comparison of DFRFT-based method with DWT-based method for five
host images
78
Cfuipter4
4.4.3 Robustness of Watermarking Scheme
An image watermarking scheme can not be considered robust until it is able to
sustain the image processing attacks. In order to measure the robustness of the
proposed watermark scheme, some common image processing attacks such as
corrupted with salt & peppers noise, median filtering, addition of Gaussian noise,
JPEG compression, low-pass filtering, histogram equalization, and sharpening (as
discussed in Chapter 2) are applied to the watermarked image. The robustness
performance of our DFRPT-based watermarking method is compared with a popular
DWT-based Kundur's watermarking method [86]. The effects of various attacks are
discussed in subsequent texts.
Salt and Peppers Noise
This noise creates isolated black pixels on white areas of image and white
pixels in black areas of image. The salt and peppers noise of density d^O.Ol has been
added to each of the watermarked image and performance of watermark extraction
technique is evaluated in terms of BER and compared with DWT-based Kundur's
watermarking method. The corresponding results for each of the five test images are
shown in Figure 4.3. From this figure, it can be observed that in DWT-based
Kundur's watermarking method, the BER is always higher than 30%, whereas for the
proposed method, the corresponding BER is consistently less than 15%.
Since, in DFRFT-based technique, watermark was embedded in lower
frequency content of host image and salt & peppers noise can be roughly thought as a
signal with plenty of high frequency components. Therefore, salt and peppers type of
attack will affect most of the high frequency content in the signal, without having any
severe effect on low frequency information of host image. Due to this reason, salt and
peppers noise will not affect the watermark that was embedded in low frequency. At
the same time, in DWT-based Kundur's watermarking method, watermark was
embedded in high frequency DWT coefficients. In this case, salt and peppers noise
affects the embedded watermark and therefore increasing the BER. This is the reason
why, our technique offers smaller BER in comparison to Kundur's method for salt
and peppers noise attack. Therefore, it can be said that our watermarking technique is
more robust than Kundur's technique for salt and pepper noise.
79
Cfiapter4
• DWT
HDFRFT
Lena Boats Goldhill Peppers Baboon
Images
Figure 4.3 Effects of Salt and Peppers Noise of density 0.01 on watermark extraction.
Median Filtering
The median filtering teclmique is widely used image processing technique for
smoothing finer details with the major edges preserved. A median filter may increase
or decrease the intensity of pixel value, since median value depends on the intensity
of the neighborhood around of pixel in the image. Median filtering is effective in
removing salt and pepper noise (isolated high or low values). A 3x3 median filter is
applied to the watermarked images and watermark extraction performance in terms of
BER is determined for our method and Kundur's watermarking method. Watermark
extraction performance for 3x3 median filtering is shown in Figure 4.4. It can be
observed from the Figure 4.4 that our watermarking technique is more robust than
Kundur's watermarking method for all the five images used in watermarking as it
offers smaller BER in watermark extraction in comparison to later method.
80
chapter 4
• DWT HDFRFT
35 30 i 25 j
^ 20 m 1 5 I
1 0 I 5 I 0 I fc fc fc fc I
Lena Boats Goldhill Peppers Baboon
Images
Figure 4.4 Effect of Median Filtering on watermark extraction.
Addition of Gaussian Noise
A wliite Gaussian noise with zero mean and variance of 0.0003 is added to all
five watermarked images and wsitermark extraction performance in terms of BER has
been measured for DFRFT-based watermarking method as well as for Kundur's
watermarking method. The BER performances of two methods are shown in Figure
4.5. Watermark extraction perfarmance for AWGN with zero mean and varying
variance (0.0001 to 0.001) for 'Lena' watermarked image is shown in Figure 4.6. It
can be seen from Figures 4.5 and 4.6 that our watermarking technique is more robust
than Kundur's watermarking method for Gaussian noise attack as it offers smaller
BER in watermark extraction. One of the possible reasons for better robustness of
DFRFT-based watermarking scheme over DWT-based Kundur's scheme is that in our
scheme, watermarks are added into those DFRFT coefficients that are of high
magnitudes (or of high energy). The added noise shows little effect on these
coefficients, therefore making the proposed technique more robust to additive
Gaussian noise attacks. On the other hand, in DWT-based Kundur's method, the
watermark was embedded by randomly selecting the set of coefficients in a sub-band,
without any consideration of their magnitude (or energy).
81
chapter 4
a: m
• DWT
SDFRFT
Lena Boats Goldhill Peppers
Images
Baboon
Figure 4.5 Effect of AWGN (with variance 0.0003) on watermark extraction.
-•-DWT -»-DFRFT
0.0001 0.0003 0.0005 0.0007 0.0009
Variance of AWGN
Figure 4.6 Effect of AWGN (with varying variance) for 'Lena' image.
JPEG Compression
JPEG is currently one of the most widely used compression algorithms for
images and any watermarking system should be resilient to some degree of
compression. JPEG compression algorithm with quality factor of 70% is applied to all
the five watermarked images and watermark extraction performance in terms of BER
is measured for each of compressed images. The corresponding results are compared
in Figure 4.7. It is evident from the Figure 4.7 that as JPEG compression at a fixed
quality factor (QF) of 70 % is applied on all the five watermarked images obtained
from our DFRFT-based and DWT-based Kundur's watermarking methods, the BER
82
Cliapter4
in watermark extraction in our scheme is smaller than Kundur's method. This is due
to the fact that in our scheme, watermark was embedded in low frequency contents
but in Kundur's method, watermark was embedded in high frequency DWT
coefficients and JPEG compression quantizes high frequency contents more
coercively than the low-frequency contents.
JPEG compression is also applied for varying QF from 10% to 100% with the
step of 10% on watermarked images obtained from our method and Kundur's method.
Watermark extraction performance for JPEG compression for varying QF is shown in
Figure 4.8. As the QF increases, the compression ratio decreases, therefore, BER in
watermark extraction decreases due to the large quantization errors of high frequency
contents. Also, it can be seen that BER of our method is almost consistently lower
than the BER obtained in Kundur's method.
In our watermarking method, the BER in watermark extraction at QF = 10% is
higher than the BER at QF=20% as more high frequency contents which are likely to
be rounded to zero value. But, when QF increases from 20% to 100%. BER becomes
approximately constant due to watermark has been embedded in low frequency
contents. On the other hand, in DWT-based method, as QF increases from 10% to
100%, the high frequency contents increase. Therefore, BER in watermark extraction
decreases from 27% to 0.5% due to watermark was embedded in high frequency
contents of the image. As QF increases, high frequency contents of the image
increases, therefore, the watermark is extracted correctly.
• DWT
S DFRFT
Lena Boats Goldhill Peppers Baboon
Images
Figure 4,7 Effect of JPEG compression on watermark extraction.
83
Cliapter4
en UJ QQ
- • -DWT
-•-DFRFT
10 20 30 40 50 60 70 80
JPEG Quality Factor (QF)
90 100
Figure 4.8 Effect of JPEG compression of varying QF on 'Lena' image.
Low-Pass Filtering
Low-pass filtering is employed to remove high spatial frequency noise from a
digital image. A LPF tends to retain the low frequency information within an image
while reducing the high frequency information. Low-pass filtering operation is
performed on five watermarked images and watermark extraction performance in
terms of BER for these images is shown in Figure 4.9. In our DFRFT-based image
watermarking, since watermark is embedded in low frequency contents, low-pass
filtering of watermarked images offers smaller BER in comparison to DWT-based
Kundur's scheme. In Kundur's watermarking method, watermark was embedded in
high frequency DWT coefficients. However, since Goldhill and Baboon images have
more high frequency contents, therefore BER in watermark extraction for these
images are higher or approximately equal to that of DWT-based Kundur's method.
84
Cfiapter4
• PWT
SDFRFT
Lena Boats Goldhill Peppers
Images
Baboon
Figure 4.9 Effect of Low-Pass Filtering on watermark extraction.
Histogram Equalization
The process of adjusting intensity values of images is known as histogram
equalization. Histogram equalization transforms the intensity values so that the
histogram of the output image approximately matches a specified histogram. It
produces an output image having values evenly distributed throughout the range. It is
applied to all five watermarked images and watermark extraction performance in
terms of BER for our watermarking technique and Kundur's watermarking method
has been measured and is shown in Figure 4.10. Modification of pixel intensities of
images due to histogram equalization may result into change in watermark data.
Therefore, BER in watermark extraction increases. As observed from Fig. 4.10, it is
apparent that our watermarkmg method offers higher BER (poor robustness) in
watermark extraction in comparison to Kundur's watermarking method.
85
Cfiapter4
• DWT
SDFRFT
Lena Boats Goldhill Peppers Baboon
Images
Figure 4.10 Effect of Histogram Equalization on watermark extraction.
Sharpening
Higli-pass filtering (HPF) is the basic step in most of the sharpening methods.
An image is sharpened when contrast is enhanced between adjoining areas with little
variation in brightness or darkness. The kernel of the HPF is designed to increase the
brightness of the center pixel relative to neighboring pixels. Sharpening operation is
performed on the five watermarked images and watermark extraction performance in
terms of BER for these sharpened images is shown in Figure 4.11. It is observed from
the Figure 4.11 that our watermarking method offers higher BER in watermark
extraction for all five images in comparison to Kundur's watermarking method. BER
in watermark extraction for our DFRFT-based image watermarking scheme is higher
than DWT-based Kundur's scheme for all the five watermarked images due to
watermark was embedded in low frequency contents (in low frequency DFRFT
coefficients), which are severely affected (rejected) due to high pass filtering. In
DWT-based Kundur's image watermarking scheme, watermark was embedded in
high frequency DWT coefficients, and sharpening shows little effect on these
watermark data.
86
chapter 4
• DWT
HDFRFT
Lena Boats Goldhill Peppers Baboon
Images
Figure 4.11 Effect of Sharpening attack on watermark extraction.
4.5 Summary
In this chapter, a DFRFT-based watermarking scheme for gray scale images
with novel non-blind watermark extraction algorithm has been presented. It has been
observed that this watermarking technique maintains imperceptibility feature of
watermarked images. The results of our technique have been compared with existing
DWT-based Kundur's watermarking method. Our technique is better than Kundur's
method for some of the attacks such as salt and peppers noise, median filtering,
AWGN, and JPEG compression. Our watermarking technique has to be more secure
than other similar watermarking techniques because watermark can be embedded
using a secret DFRFT powers which are essential information required for watermark
extraction. For our DFRFT-based image watermarking scheme, a non-blind
watermark extraction scheme has been proposed.
87
CHapter 5
(Digitaf Image Watermar^ng using
ComSinecfWavefet and Tractiona[ Courier
Transforms
^ n
CfiapterS
DIGITAL IMAGE V^ATERMARKING USING COMBINED
WAVELET AND FRACTIONAL FOURIER TRANSFORMS
5.1 Introduction
In previous chapter a digital image watermarking scheme based on DFRFT
was discussed. It was observed that DFRFT-based watermarking scheme has
relatively poor robustness compared to DWT-based Kundur's watermarking scheme
for attacks such as sharpening. This necessitates the need to develop a watermarking
technique which can improve the robustness of DFRFT-based watermarking
technique for such attack. One of the reasons for the relatively poor robustness of
DFRFT-based watermarking scheme is that in this scheme the watermark data was
embedded in low frequency regions, which was filtered out by sharpening, thereby
increasing the BER (hence reducing the robustness). One of the possible solutions to
improve the robustness against such attack may be to embed the watermark data in
high frequency regions of the image. The high frequency components of the image
can be selected by selecting the high frequency sub-bands of wavelet transformed
image. That is if DFRFT-based watermarking scheme is applied in high frequency
sub-bands of DWT transformed image, then the robustness may be improved. With
this motivation, a watermarking scheme combining DWT with DFRFT has been
presented in this chapter. The watermark embedding and extraction for digital image
watermarking using combination of DWT and DFRFT will be described. Various
simulation results for this scheme and comparisons with other watermarking schemes
will be presented and discussed in this chapter.
5.2 Background
Among the many available transforms, DCT and DWT are probably the most
widely used and explored transforms for watermarking applications [111]-[114].
DWT is a hierarchical transform (unlike FFT and DCT), which offers the possibility
of analyzing a signal at different resolutions. DWT has been used in digital image
watermarking more frequently due to its excellent spatial localization property" and
multi-resolution characteristics [122]-[123]. A major advantage of the DWT lies in
the fact that it performs an analysis similar to that of the Human Visual System
88
CfiapterS
(HVS), as it also split the image into several frequency bands, similar to HVS. Some
other advantages of DWT are as follows:
(i) It has better energy compaction as compared to FFT,
(ii) Unlike DCT, DWT is not a block-based transform, and so it does not suffer with
annoying blocking artifacts, which is generally associated with the DCT.
(iii) The multi-resolution property of DWT offers more degrees of freedom
compared with the DCT.
However, as it operates over the entire image, the DWT requires larger
memory and has higher computational complexity than the block-based DCT.
Due to these advantages of DWT, it has been explored for various applications
including the digital watermarking. A number of watermarking schemes are
developed using either DWT alone [121-[122] or DWT in conjunction with other
transforms, such as DCT and DFT [123]-[126]. In this chapter an attempt is made to
develop watermark embedding and extraction algorithms using combination of DWT
and DFRFT.
Digital image watermarking based on DWT and DFRFT was also attempted
earlier [84], in which a two-dimensional spatial-domain chirp signal (as watermark)
was embedded in low frequency sub-band of the host image in wavelet-domain. The
two-dimensional spatial-domain chirp watermark was detected in DFRFT domain
blindly.
However, our watermarking approach differs from that of Hong et al. [84] in
the sense that the watermark is embedded in selected DFRFT coefficients of single or
combination of high frequency (HF) sub-bands of wavelet-transformed host image. A
non-blind watermark extraction method is also suggested.
5.3 Proposed Watermarking Scheme
As mentioned earlier, one of the problems with DFRFT-based watermarking
scheme is that it offers higher BER for the attacks such as histogram equalization and
sharpening. The proposed watermarking technique is aimed to circumvent this
89
CfiapterS
problem by exploiting the multi-resolution sub-band decomposition property of DWT
and applying DFRFT of selected sub-bands in transformed image only, instead of
applying DFRFT of the whole image. In following sub-sections, proposed watermark
embedding scheme is introduced, which is followed by the corresponding extraction
method.
5.3.1 Watermark Embedding
The block diagram of watermark embedding process is shown in Figures 5.1.
As shown in figure, the host image (7) is first wavelet transformed into approximation
and detailed sub-bands. Then the DFRFT transform is applied to some selected high-
frequency sub-bands.. The resulting DFRFT coefficients are then arranged in
ascending order and watermark data (w) is embedded in some of these coefficients.
After embedding, inverse DFRFT and inverse DWT are applied to obtain the
watermarked image (/').
In this scheme, DWT decomposition of a gray scale host image is performed
using Haar filters. A singIe-le 'el DWT decomposes an image into multi-resolution
sub-bands, consisting of approximation coefficients (LLl) as well as horizontal
(LHl), vertical (HLl) and diagonal (HHl) details coefficients, where T indicate the
sub-band obtained after F ' level of decomposition. This process is repeated on
approximafion coefficients sub-band (LLl) to compute 2" level decomposition,
which further breaks LLl sub-band in to LL2, LH2, HL2 and HH2 sub-bands. The
process may be repeated on i" -level approximation sub-band to get (i+l)"^-level sub-
bands.
/
DWT DFRFT of Selected
Sub-Bands - H
~i
w
+
Inverse DFRFT of Modified
Sub-Bands — •
Inverse DWT
/"' , '
Figure 5.1 Block diagram of watermark embedding for combined DWT and DFRFT
domain digital image watermarking.
90
CfiapterS
In this algorithm, the watermark data is embedded in selected HF sub-bands,
i.e. in the one or two sub-bands selected among LHl, HLl, HHl, LH2, HL2 and
HH2. The criterion to select the sub-bands is the perceptibility of watermarked
image (to be discussed later in this section). Each of the selected HF sub-bands is
further transformed using DFRFT and then the transformed coefficients are
concatenated in a single one-dimensional array (using raster scan). The DFRFT
coefficients of concatenated array are sorted in ascending order of their magnitudes.
Some of these coefficients (equal to the watermark length) in contiguous order are
selected for watermark embedding and the watermark is added in these coefficients.
These watermarked and remaining DFRFT coefficients of sorted array are then placed
at their original locations which are finally placed back in HF sub-bands of which
they belong. The inverse DFRFT is applied to the selected HF sub-bands, which were
earlier transformed using DFRFT. Then inverse DWT is performed to obtain the
watermarked image (/').
• Criterion for Selection of High Frequency Sub-bands for Embedding
As mentioned earlier, since the watermark is embedded in the selected HF
sub-bands of DWT transformed image, it becomes important to discuss criterion for
the selection of suitable HF sub-bands. In this work, the selection criterion is based on
imperceptibility of the watermarked image. For the investigation purpose, an
experiment was performed in which the watermark data of fixed length (1024) is
embedded in either one of the HF sub-bands or various combinations of two different
HF sub-bands, of a two levels DWT transformed 'LENA' image. The PSNR value of
the watermarked image is measured for each case. The results are listed in Table 5.1.
It can be observed from the table that if watermark is embedded in any one of the HF
sub-band of the first level of decomposition (i.e. LHl, HLl, and HHl), the resulting
watermarked image has poor imperceptibility (PSNR of the range of 35-40 dB) as
evident from first three rows of Table 5.1. Therefore, LHl, HLl and HHl sub-bands
are not suitable for watermark embedding. However, if watermark is embedded in
other high-frequency sub-bands (LH2, HL2 and HH2) as well as various
combinations of HF sub-bands listed in Table 5.1, the watermarked image has good
imperceptibility, as evident from PSNR values (more than 51 dB) of watermarked
image in the table.
91
Cfiapter5
Table 5.1; PSNR of watermarked 'Lena' image after embedding in various HF sub-
bands
S.No.
1
2 -1
4
5
6
7
8
9
10
11
12
13
14
15
Sub-bands in DWT
LHI
HLl
HHl
LHl+HLl
HLl+HHl
LHl+HHl
LH2
HL2
HH2
LH1+LH2
HL1+HL2
HH1+HH2
LH2+HL2
HL2+HH2
LH2+HH2
PSNR (dB)
35.05
31.85
40.00
51.35
51.33
51.35
51.43
51.36
51.37
51.35
51.34
51.35
51.38
51.29
51.35
The watermark embedding process can be summarized in the following steps:
Step-1: The first step in the watermark embedding process is that the n-ievels DWT
decomposition of the host image (I), which results into (3n+l) sub-bands, one
approximation sub-band (LL„) and 3n detailed sub-bands J^{LH,+HL,+HH,) .That
is;
X"„. = DWT (/)= LL„ + LH„ + HL„ + HH„ + LH„ , + HL„_, + //W,,^, + + Lf/, + //L, + //// , (5.1)
where, n is the number of decomposition levels (here n = 1 or 2) in DWT.
sb' represents a sub-band with index /. Different sub-bands along with their indexes
for two-level decomposition (n=2) of an image are shown in Figure 5.2.
92
CfiapterS
Step-2: Select a set of sub-bands from the detailed sub-bands (HF sub-bands), say this
set is ,v'. where i may have any value from 2 to 7 (refer Fig. 5.2). Perform DFRFT of
each HF sub-bands of set s'. That is,
R =V\s'] (5.2)
where, F denotes DFRFT transformation operator and R is an array containing the
DFRFT of sets'.
LL2 1
LH2 3
HL2 2
HH2 4
LHl 6
HLl 5
HHl 7
Figure 5.2 Different sub-bands of an image for second-level DWT decomposition.
Step-3: Sort the elements of set R (set containing DFRFT of s") in ascending order of
their magnitudes and store the sorted data in an array Z,.
Step-4: Select first 'M' DFRFT coefficients of Z,, where M is the watermark length
for watermark embedding. Let these coefficients of Z,are indexed asZ^i], where
k=l,2,...,M.
Step-5: Add watermark data \<,{k] to Z^k] to get watermarked DFRFT
coefficients z;.'[/I].
Z';[k] = Z^[k]+w[k], \<k<M (5.3)
93
CfiapterS
Step-6: These watermarked and remaining DFRFT coefficients of sorted array are
then placed at their original locations which are finally placed back in HF sub-bands
of which they belong.
Step-7: Apply inverse DFRFT on each modified sub-bands obtained in Step-6
Step-8: Perform inverse DWT to obtain a watermarked image (/").
5.3.2 Watermark Extraction
It is basically a non-blind approach in which the knowledge of the original
host image is required. It is assumed that host image is also available at the receiver.
Additionally, the extraction process also requires information such as number of
decomposition levels, powers of DFRFT transform and index of sub-bands in which
watermark was embedded etc. This additional information may be send as headers in
the watermarked image. The block diagram of the proposed watermark extraction
scheme is shown in Figures 5.3, The watermarked and host images are transformed
by using 'Haar' wavelet transform. Then the detailed sub-bands (where watermark
was embedded) of both host and watermark images are transformed using DFRFT
(using the same respective DFRFT powers as were used during watermark embedding
process). These transformed coefficients of both images are concatenated in two one-
dimensional arrays (scanned in raster order) correspondingly. The resulting arrays are
sorted in ascending order of their magnitudes, similar to the ordering done during
watermark embedding process. Then the first M coefficients of array containing
ordered DFRFT coefficients of host image are subtracted (element-wise) from the
array containing ordered DFRFT coefficients of watermarked image, which results
into extracted watermark. In this manner, extracted watermark has been obtained.
In order to elaborate the step-by-step watermark extraction procedure, let
/ and /"represent the host and the received watermarked images respectively. The
steps involved in extraction are as follows:
Step-1: The first step in the watermark extraction process is to apply one or two
levels of 2D-DWT using 'Haar' filters on both / and /".
94
CfiapterS
Step-2: Using the index of HP' sub-band from the header, perform DFRFT of HF sub-
bands (in which watermark was embedded) for both host and watermarked images.
/ " •
DWT
DWT
DFRFT of Selected
Sub-Bands
DFRFT of Selected
Sub-Bands
^'
if
+
Figure 5.3 Block diagram of watermark extraction for combined DWT and DFRFT
domain digital Image watermarking.
Step-3: If R is an array containing the DFRFT of set of HF sub-bands for original
host image (/) and R" is an array containing the DFRFT of set of HF sub-bands for
watermarked image (/'"), then sort the elements of sets R and R ' in ascending order
of their magnitudes and store the sorted data in an arrays Z, and Z,", respectively.
Step-4: The watermark is then extracted by performing element-wise subtraction of
first M elements of array Z, frora Z"' according to Eqn. (5.4) given below;
M' = Z:[k]~-Z^[k]for\<k<M (5.4)
5.4 Simulation Result and Discussion
In this section, simulation results of proposed digital image watermarking
scheme are presented and compared with DFRFT-based scheme discussed in previous
chapter (Chapter-4) and DWT-based watermarking method proposed by Kundur and
Hatzinakos [86]. Implementation parameters of all the three watermarking methods
(DWT-based method, DFRFT-based method and proposed method) are summarized
as follows.
95
chapters
In order to evaluate the performance of proposed watermarking method and
for comparison with other methods, five standard 8-bit gray scale images namely
Lena. Boats. Goldhill, Peppers and Baboon, each of size 512*512 pixels are
considered. Pixels of a binarized image of size 32*32 pixels {\fh.hmp') are taken as
watermark data. Therefore, the length of watermark data is 32*32=1024 (i.e. M=
1024). The watermark data are scaled using a scaling factor of 44. The scaling factor
determines the threshold {Th) for the binarization of extracted watermark data, i.e.
7'/7=scaling factor/2=22.
The remaining parameters for different methods are as follows;
> DWl-based Kundur's method [86]:
• Number of decoraposition levels= four;
• Filter: 'Haar' filter;
• Sub-band in which watermark is embedded: high-frequency sub-bands
of4"Mevel.
>• DFRFT-based method:
• The watermark locafion (L) =259000;
• DFRFT powers p, =[p,.Pj]= [0.99, 0.99].
"y Proposed (combined DWT and DFRFT-based) method:
• Number of decomposifion levels= two;
• Wavelet filter: 'Haar' filter;
• Watermark location (L) = 0;
• DFRFT powers P, =[p„p,\= [0.99, 0.99];
• Sub-band in which watermark is embedded: combination of high-
frequency sub-bands of 1 ' and 2" * levels.
96
CfiapterS
Using these parameters, the three methods are simulated and their
performances under various attacks, measured in terms of BER of extracted
watermark data are summarized in Tables 5.2-5.6 for 'Lena', 'Boats', 'Goldhill',
'Peppers' and 'Baboon' images, respectively. In each table, the first column includes
various channel attacks, second and third columns list BER under various attacks for
DWT and DFRFT-based methods, respectively. The remaining columns in each table
give BER performance of proposed method with different combinations of high
frequency sub-bands considered for watermark embedding, and under various attacks.
Watermark extraction performance for salt and peppers noise attack for
uatermark embedded in each ol' the host image is presented in the second row of the
corresponding tables. It has been observed from these results that the proposed
method consistently outperforms (offers smaller BER in watermark extraction) the
DWT-based scheme and has almost similar performance to that of DFRFT-based
method for almost all the host images. A detailed analysis of these results reveal that
for almost all images, the DWT-based method results BER of the order of 32-33%,
DFRFT-based scheme gives BER of 12-13% whereas the combined method has BER
performance of approx. 10-14%o. Further, for each host image, a slight variation in
BER (approx. 3.5 - 5%) of the proposed method is observed by varying the sub-bands
in which watermark was embedded.
Third row of the corresponding tables gives the watermark extraction
performance in terms of BER for median filtering attack for above mentioned three
methods. It is evident from the results given in tables that the DWT-based method
gives BER of the order 16.5-31.74%, DFRFT-based method resuhs BER of 4.5-
20.61% whereas the proposed (combined DWT and DFRFT domains) method has
BER performance of 10.16-13.28%. It can be observed from these results that the
proposed (combined) watermarking technique offers smaller BER in watermark
extraction than DWT-based method and higher BER than DFRFT-based method for
four images. For Baboon image, BER in extracted watermark is smallest than other
two methods.
Results of watermark extraction performance for additive white Gaussian
noise (AWGN) for three methods are given in fourth row of corresponding tables. It is
97
CfiapterS
observed from the detailed result analysis that for almost all images, the DWT-based
method results BER of the order of 10.84-13-74 %, DFRFT-based method gives BER
of 3.8-4.18% whereas the proposed method has BER performance of approx. 3.02-
5.66%.These results show that proposed (combined) method offers smaller BER than
DWT-based method and approximately equal to DFRFT-based method for all the five
images. The possible reason for this is that in DFRFT-based method, watermarks are
added into those DFRFT coefficients that are of high magnitudes (or of high energy).
The added noise exhibits a little effect on these coefficients, therefore making the
proposed technique more robust to additive Gaussian noise attacks.
Fifth row of corresponding table gives the watermark extraction performance
for JPEG with quality factor of 70%. BER in extracted watermark for DWT-based
method and DFRFT-based method are 7.16-9.34% and 3.61-4.88%, respectively.
Watermark extraction performance for combined method is 6.45-10.45%. It is
observed from these results that proposed (combined) method offers highest BER in
comparison to other two methods. JPEG compression algorithm works in the principle
that as compression ratio increases (quality factor decreases); more and more high
frequency components are quantized to zero values. Since in proposed method,
watermark data are embedded in a set of HF sub-bands and vv'hen watermarked image
is subjected to JPEG compression, high frequency components are distorted, therefore
increasing the BER. At the same time, watermark was embedded in low frequency
components in DFRFT-based method as well as in comparatively low range of high
frequencies in DWT-based Kundur's method compared to proposed method
(combined DWT and DFRFT Method).
Low-pass filtering operation is applied on all the watermarked images and
watermark extraction performance in terms of BER is determined for three methods
mentioned above and results are given in sixth row of corresponding table.
Watermark extraction performance in terms of BER for DWT-based method and
DFRFT-based method are 20.83-29.84 % and 11.72-28.32 %, respectively. For
combined method, BER in extracted watermark is 10.45-11.91%). It is observed from
these results that proposed method offers smallest BER in watermark extraction in
comparison to DFRFT-based method and DWT-based method for five watermarked
images because watermark embedding in proposed (combined) method is performed
98
CfuipterS
in comparatively high frequencj' components than other two methods. Therefore, our
method offers smallest error in watermark extraction than other two methods after
low-pass filtering attack.
Results of watermark extraction performance for histogram equalization are
given in seventh row of corresponding table. The BERs in extracted watermark for
DWT-based method and DFRFT-based method are 44.62-48.88% and 81.55-88.87%,
respectively. For proposed (combined) method, BER is 0-28.32%. For this attack,
proposed method offers best watermark extraction performance in comparison to
other two methods due to modification of pixel intensities of images by histogram
equalization which results into change in watermark data.
Watermark extraction performance in terms of BER for sharpening attack is
gi\'en in eighth row of corresponding table. The BERs for DWT-based method and
DFRFT-based method are 32.56-47.57% and 70.32-77.93%, respectively. The BER
for combined method is 0-7.72% after sharpening attack. It is observed from these
results that proposed (combined) method offers smallest BER in watermark extraction
for all the images used in watermarking in comparison to DFRFT-based method and
DWT-based method. Reason of best watermark extraction performance for
sharpening attack in comparison two other methods is same as given for low pass
filtering attack.
99
01
S
c
c 01 u
.4>
01 •o
cc
o
E O
•o O
0)
E
i-
E u 01
01 01
c o
E o U O)
u c E
01
s H
•a o JS
-4-»
U
S H
• o c «
H
Q T3
0 1
C • • •
. 2
E o U • • T5 O
"O 0)
o o
d.
H -o
§ 5 to «i
"U o
H J=
^ CI
<
H ^ Q c
• " •
1/3
73 B C«
J2 1
3 m
x =
- H ' ^
- ] d
= 5
f S
X ^ - T
5 X +
f^ !?
— "+
M 2J
• ^ "
^ •" "-• >-> " T -
_ — M •*"
" T
^ — X ^ • J -
f S
E
( N _J
( N
!J
en • *
_: ~~
•* o ^ '—
• ^
r- o
_ — ( N
—
—• ( N
( N
—
r^ l O
f-i ^-
m r^ ^ -'
f O • —
^ —
,— q ( N
m - ^
^^
^ ON
•<t' ""
ON ON
r i ^"
i n
'—'
to
tN n
»}
u a a
P H
«3 CU
u- i
^ d ""
i T i
^ d '
l o T T
d «
" D
-* d —
l O T f
d —
i n
^, d
i n
^ d
i n • ^ ^
d
i n
^ d
i n
^ d "*
>n -* d ""*
i n
^ d ""*
O N
^'
NO
" •
a +-
c .2 "55 < ^
's-
T i -
r-i i n
o-i
ON
^ ^
-• NO
rn
oo q i n
<Ti
• ^
0^ NO
^"
oo
<n
^ T)-'
^ ^'
r~-r^ i n
0 0
rn
ON
""
z 0 ^ <
i n
^ d ""
i n T f
d ' '
i n T f
d ~ - i
i n • *
d ^
i n
^ d ^
i n f )
d r—
i n
^ d
i n
_ d
i n
^ d .—
i n
xl-d ~
NO
q d • " •
NO
^-d "~"
— ^
NO
0 0
a OH
B
E "
i n • ^
d '
i n • *
d '
^ oo d •^
m NO
d r ^
i n i n
d "—
• *
oo d
i n NO
d '—'
i n
' t d • — '
i n i n
d r—
i n • *
d "~
i n NO
d *""
r^ m ^ "^
O N
' '
1 — 1
( N
PH-C
o .•::
0 0 O N
q d
0 0 ON
o d
0 0 O N
O
d
O N
( N
d
o-i n
d
oo 0 ^
d
o
o
o
0 0 NO
d
m i
r i ( N
r^ ( N
^
0 0
0 0 0 0
ON
i n
'^
B _ O
s -v L, N
O 03
CO O N
q d
oo ON
o d
CO O N
O
d
r^ q •^
0 0
oo
d
• *
i n
r>i
o
o
o
r- r-l ' • " '
— CT;
cS
< — 1
NO
rn
oo
r^
i n r^
l » _B
"a 9J a «
s
o «
a Vi
u C3 +^
C 01 X.
.4>
01
S 3
w CQ
E
o
c
u
s
CI
u JZ
c o
« c E o U V y c « E u
a a, IT)
z H
•o o x:
-^rf
o
s a:
j - O B 55
H ^ O T3 1" C
• » M
' ^
:w ["O i O ^
IS 1 (U
y ; O 0, O
P-1
1
H U. »:
Q
1
i H
!
i 1
•0 o
« s
O x: •b^
u S
H ^ Q e • • ^
• 0 e « X : 1
3 (/5
— ^ _ j " — '
— " X •*•
^ +
•J?
<N S _; I
+
I 3 •J +
^ — X 3
->% X ^
3 IJ ~ X "• +
fS X X
fS - i
r l I - 1
r^ r~ ro '""
ON
^ • —
0-1 l O
_ l —
( N O
-—'
vq ( N
NO
r4 .—
— M
<N
'—'
r~; (N
.—'
<} oo c '
i n m o ~
IT)
q d
( N
(N
' '
_ O ( N
~
r^ O
a a 0)
4 j !Z1
i/n
^ d ~"
'/-I • ^
d
m .
d 1 —
i n
d ^
wo ^, d
i n
d
i n
^ d
i r i 'a; d
•n ^ d t —
m ^ d '
i n TT
d
« • *
d • " "
i n (N
o
o '—
6JD
^3 ro
O N
en
^ ^
^ ^'
NO
(
m " r^
^
ON i n
• * '
T f
r^ ^ i n
r—
00
r^
ON NO
ON
- rf
ON OS
(
^ r-~ rn
z 0 ^ ^
i n • *
d *""
i n -:r d '
m • *
d ^
m
d —
m ^ d
d - — •
i n T T
d
m
d —
i n • *
d —•
i n
^, d
'
NO
ON
NO
'—' d '—
0 0 oo
'
\r ON
0 PH
c _o '3?
s o U
o w
N — '
i n
^ d • " "
m, rr d ""
i n
^ d ^
i n
d 1 — 1
m • *
d '—'
d —"
in i
^ d *~~
i n
d —
ITi I t
d —
m t
d • " •
i n NO
d
<n • *
d • " •
t-~
o\ r- *~
ON 00
I N
Hi
22 CD
a. •-\ ^
^ d
— ^
NO NO
•—
m
d ™
r- r^ o^
i n
d n
ON r i
d
00 00
•* d
NO NO
—
ON i n
oo
( N CT)
od
t ^
d CNI
r-1 00
NO oo
m 00
^ ^
s u en o 1
*^ 3
o
i n ON
—; d
i n Q\
' ' d
ON
d
GN r'^
d
^ r '- j
rJ
o
o
00 ON O
d
ON
^ d
M ^ rn
i n ( N
(N
r-i rn d r-
NO i n
'5 a
(/3
E ^ —" •— Xi
2 o
O -s. o
!«. J y n
• ^ ^
M
- -c 01 L.
.aj £ •5
01 "0 s 3
r^
o !/l
s i. 0>
C
CA
o J: 01
E ^ u w E u o» -1^
s ^ 4J 01 u
JS -!-» !« o s o
.— 'u
a-E o U 01
c S5
E >-
c2
0)
a.
1/)
T3 O
J=
ii
s
" 3 c o:
H
c IS E o U . * •a o JS
^ 01
(/! O a o u ^
H -O t i l ®
xs
^1 Q S
<
H ^ 0
e
,o 3
VI
f S
— £ = =F
+
r^i
-J -^ = ?
rs '^ X
^ +
" T
f s : ^
2^ "T +
5 = J 7 +
_ — J 3 ~ ?
r t
r j _J X
rN X J
r-ly-i
C<
r^ r'
, '. _
00 (N
r'i
i n
r l
r j
in
t^ i n
__ en ( N
t ^
—'
m
d '
r i n
en
—
r~ r-~ r^
-
0\ TT
r4
r-r-r^
!/l
0.
(/5 Z
i n TT
d
i n
^ d
m i TT
d
i n
d
i n m d
i n
d
i n
d
i n
•*_ d
i n
d "—
i n
d " " •
^0
— d •—'
i n r' i
d ^
i n
yo
\ D i n
d (N
6JD _C
0) ••-<
.— r-; en
r-i o r i
t —
00
r^
oo
<> rn
00
-*'
^ ' t
00
r-i
00
r-i
0^
^ •*
M t '
OS
o
^ ^ Z
0
>n -* d
m -* d
i n
^ d
so
d
i n (N
d
OS
Os'
i n
d
i n r f d
m
d
<n
d '—
oo ( N
ON
OO
o Cs'
— SO
rn
so m 00
0 A N >-5
e
mi • *
d ^
i n • ^
d
-* r-d —
i n ITi
d
i n r^
d '—
d
i n i n
d
i n
^ d -—
i n in
d • —
i n
d •""
i n
^ d '—'
i n so d —
(N r<-i
00
^
2 Ml
OH • -1 »•
5:=
o
OS en
d
oo o^ o
d
OS i n
d
00 r-
d
so so
o
o
00 OS
o d
o
so r- r—
C-l
^ en
00 < r
00
so r--sd
S
E-2 t , N
o «
o
oo OS O
d
r l
d
0^
d
OS - T
d
rn
o
o
0 0 OS
q d
r<-i
d
^ IS;
r-i
• * so
oi
__ r~-d
^ r--
^
OiJ
'c 0)
a «
o
1/1 i-<u a. a. a.
a u
.4*
c
P3
o
s
o
E
u
E 0)
u
s o
'H S5 a E o
u u B
E u a u
(X
T3 O
f • ^ ^
1 Pi t^ Q "O s «5
H
Q
• 41 C
1 < m
: £i
1 E 1 o
! ^
'"S 1 ®
J=
T5 0)
O C O
O H
i j
H tb aj
H
' <
•a o
01
S
-a o
^ -w 0)
s
H ^ Q s
• • •
c C5 ^
1
s in
= =
r^
-J d
^ : ; « • J ^
r s : : : ; T - 'Jm,
" ^ X
+
(N 5 J 5 * +
<s [! J =
^ — • « .M
~ T
r I 1 =
^ — IC £ • " "+
r s 7 "
"Z.
<N
«
f S
[ j
( N
~
r i r j
« '
V O
o •a^
ao rn ro
1/-,
r-i • —
n r^ ^ -
o o f ^ .
'—
^ oi -
r^ vq f ,
—
oo c^, rn
'"'
OS
oo ( N
'~~
( N
( N
~
^ 0^ (N
r *-; ( N ( T l
!« L,
a a 0,
= «
in Z
i n
d • —
i r i
'^ d '—
v r j
^ d
i n T T
d
i n
^ d
' n
d
i n • *
d
m
d
i n
'a-d
i n • *
d ' '
i n
-* d ""
i n ' I t
d ""
' O
q d
3 0 •sf
o ~~
OX)
-E
B
.2 '5 fo"
"~ • * •
m rf '
r*^
' t
•— OO
n " ,
— ^
^ '*'
O VO
i n
r^
^
' '_ Tt -
oo OS
• * '
— T f
•*
t ^
O
ON 0 0
d "~
0 ^ <
d ""*
i n • ^
d "
i n
^ d
i n
^ d
i n
•* d
i n
d .—
m
'*. o
i n
d
i n
_ d .—<
i n
, d ' '
i n
ro d *""
NO
o d '~
ON i n
'— o 0 0
0
B
'a?
m i
d ""
m -* d ~"
• n •* d —"
i n T f
d —
i n T T
d '—
IT] m d 1 —
i n I * -
d
i n
d
i n i t
d —
i n T T
d • ^
i n
'^ d "-
^ r-d "
( N
r-—
r<~, 0 0
d C^l
in 2 6X1
o .•::
o
0 0
o o d
o
n d
r~ i
d
c-• *
d
o
o
o
ON f i
d
0 0
q • ^
t ^
~ *~"
i n i n
oo
( N NO
Tt-'
^
E-2 U N BX):;= O S5
o
oo o-o d
o
0 0 0 0
d
ON r* ,
d
r r^
—
O
O
o
oo NO
d
m 0 0
f N
r ' l 0 0
oi
r-oo
r
( N ' t
i r i r - i
W) S
'a
rn o
c o o
Si cs
03
175
c
"73 C 3
ca
E
o
0)
E
c o _j2
C. E o U 0) (J c « E
a
•o o J3
4«*
•o s s: H
Q
0 1
s t H t
X i
E o U . . •o o
"O 7 5
O
o
a.
H "O u. o S -c DS -«-•
• 3 , O H J=
^ 0)
<
H ^ Q c *
B
« J2 1
3 in
_ r j
•r I
- . ( N
J ri ^? — f ^
— X = 4^
N a s ft
(s : i J =: = ?
— H -J- X
^? _
J 5 - X +
_ •— 3 S +
f S • "
—
f N
^
( N
!J
. — 1
r i "-
^ r~; d • ~
l A
r i
r"-)
—
M ^ O
^ • —
r i —
— o r i •"
"/ (^1
d
__
0 0
m O^
ON ON
( N
"
^ O
_" '—'
'/-) -* —1
' '
• *
^ ^ m
^ b
a a
OH
«3 «
'y~.
^ C
^ yr. VC
d '
< N 0 0
_; ~
r^
i; —
> r i v O
d '—
so
d
-* o _J
'
>/-) ' d —
r",
( —
m ^ d ""*
>o o d " " •
0 0 ( N
f i • ~
— ^ d (N
^ r-— m
fa c .5
^ IK
< N 1 / ^
rn
O v
r 'S
( N
T f '
ON
r^
OV T l -
^
t N
•
0 0 oo
^
^
ON
^
MD
rn
O N
NO
•*
0 0 0 0
'
0 0
'— ^
• *
0 0
d • ^
z 0
u-^ ( N
d • ~
NO
o d '—
t ^
' <>
0 0
r i r^ r~
^ r^ NO
Ti
^ d '~~
u- i
r-j d ^
1 / ^
d
1/-1
0 0
- r-NO
U-,
^ NO
^
NO
—; r-'
S
Is fa S fa
' / - I T t ;
d •
• ^
r--d ~~
n r i
^ -"
• *
d
» r i i r i
d ^
—
VT) NO
d • ~ "
m l ^
d —
rr o
—
i n
^ d • " "
r-) • — 1
_; "~
O N
__ —'
ro 0 0
r-' ( N
^ 0 0
ON f N |
Vi
2 bxi
OH -2
hJ fa
o
( N
d
0 0 NO
d
i n
( N
r--( N
—
r'i
O
O
0 0 ON
o d
0 0 0 0
d
— NO
f ^
( N
^ r-'
r--CO
0 0 0 0
0 0 0 0
od •*
S O
E -s. U N
M fa
o
O -i n
d
0 0 t ^
d
i n r-i < N
n oo ( N
oo
i n
o
oo ON
o d
i n ON
d
O N
-* d
t-~-
^ i n
r-i r t~^
n O N
r- r-~
r~-m r- '
eic
'c a s .
t/2
CHapterS
Now, some important results of proposed watermarking scheme are shown by
plots in Figures 5.4-5.6. The watermark extraction performance of proposed
(combined) method for histogram equalization and sharpening attacks improves a lot.
But at the same time, BER in extracted watermark increases for JPEG compression
attack.
So far, the performance of proposed method under various attacks and
compared with other similar methods (DWT-based method and DFRFT-based
method). However, in order to emphasize the advantages and shortcomings of the
method, the results for attacks such as histogram equalization, sharpening and JPEG
compression are graphically shown in Figures 5.4-5.6, respectively.
In all these figures, only the results of embedding watermark data in sub-band
HH2 (representative of single HF sub-band). HLl+HFIl (representative of
combination of two HF sub-bands of first level decomposition), HL2+HH2
(representative of combination of two HF sub-bands of second level decomposition)
and HL1+HH2 (representative of combination of two HF sub-bands of first level and
second level decompositions) have been shown.
It is observed from the Figure 5.4 that proposed image watermarking scheme
(using a set of high frequency sub-bands) offers drastic reduction of BER in extracted
watermark in comparison to other two methods because of modification of pixel
intensities of images by histogram equalization which results into change in
watermark data.
It is evident from the results shown in Figure 5.5 that proposed image
watermarking scheme using a set of high frequency sub-bands offers smallest BER in
watermark extraction in comparison to other two methods for all the images used in
watermarking. Reason of this improvement in watermark extraction performance for
sharpening attack is that in our proposed watermarking scheme watermark is
embedded in comparatively high frequency components of an image in comparison to
other methods of image watermarking.
Plots of watermark extraction performance for JPEG compression are shown
in Figure 5.6. It is evident from this figure that proposed (combined) method using a
105
Cfiapter5
set of HF sub-bands offers highest BER in watermark extraction in comparison to
both DFRFT-based method and DWT-hased method. JPEG compression algorithm
works in the principle that as compression ratio increases (quality factor decreases);
more and more high frequency components are quantized to zero values. Since in
proposed method, watermark data are embedded in a set of HF sub-bands and when
watermarked image is subjected to JPEG compression, high frequency components
are distorted severely in comparison to two other methods of image watermarking,
therefore increasing the BER in watermark extraction.
m CQ
100
90
80
70
60
50
40
30
20
10
0
Lena Boats Goldhill Peppers Baboon
Images
• DFRFT
E Kundur (DWT)
E3DWT-DFRFT(HH2)
nDWT-DFRFT(HL1+HH1)
0DWT-DFRFT (HL2+HH2)
nDVVT-DFRFT(HH1+HH2)
Figure 5.4 Effect of Histogram Equalization on Watermark Extraction.
Lena Boats (Boldhill Peppers Baboon
Images
Figure 5.5 Effect of Sharpening on Watermark Extraction.
• DFRFT
m Kundur (DWT)
nD\An"-DFRFT(HH2)
nD\/\/T-DFRFT(HL1+HH1)
B DWT-DFRFT (HL2+HH2)
DD\/VT-DFRFT(HH1+HH2)
106
CfidpterS
• DFRFT
m Kundur (DWT)
SDWT-DFRFT(HH2)
E3DWT-DFRFT(HL1+HH1)
@ DWT-DFRFT (HL2+HH2)
0DWT-DFRFT(HH1+HH2)
Lena Boats Goldhill Peppers Baboon
Images
Figure 5.6 Effect of JPEG Compression (QF = 70%) on Watermark Extraction.
5.5 Summary
It was observed in Chapter 4 ttiat DFRFT-based image watermarking scheme
shows relatively poor robustness compared to DWT-based Kundur's watermarking
scheme for sharpening and histogram equalization attacks. In order to reduce BER in
extracted watermark, a watermarking scheme in combined DWT and DFRFT
domains has been developed and its performance has been evaluated. This scheme
improves the watermark extraction performance for sharpening and histogram
equalization attacks. But at the same time watermark extraction performance for
JPEG compression deteriorates.
107
chapter 6 Condusions andTuture
m ^
Cfiapter 6
6.1 Conclusions
In this thesis, digital watermarking schemes for audio and images have been
proposed and simulated. Performance of these schemes has been evaluated in terms of
BER in watermark extraction. Three watermarking schemes are presented and
simulated in this thesis. The first scheme proposed in this thesis is for audio signals,
which uses chirp signal to embed watermark. This scheme is used for single-level as
well as multiple-level watermark embedding. The other two watermarking schemes
have been developed and simulated for gray-scale images. These methods are digital
image watermarking in DFRFT-domain and in combined DWT and DFRFT domains.
Based on the simulation results of these watermarking techniques, following
conclusions can be drawn which are given in subsequent paragraphs:
Chirp-based digital audio watermarking scheme for single-level embedding is
robust for low-pass filtering, up-sampling by interpolation, down-sampling by
decimation, re-sampling, AWGN, amplitude scaling and MP3 compression. However,
this watermarking scheme shows limited robustness under high-pass and band-pass
filtering attacks. To measure impercepfibility of watermarked audio signals. ODG
measurement of watermarked audio signals is performed using PEAQ model. ODG
value obtained for single-level embedding is -0.38 for SWRo= 30 dB which shows
imperceptibility of watermarking.
MulU-level chirp-based audio watermarking is also robust under various
attacks such as low-pass filtering, interpolation, decimation, re-sampling, addition of
white Gaussian noise, amplitude scaling, and MP3 compression. This scheme also
shows limited robustness for different levels of watermarks. ODG values for second-
level and third-level watermarked signals are -0.46 for SWRo= 27.04 dB and -0.48 for
SWRo = 25.09 dB, respectively. For imperceptible audio watermarking, the ODG
values of watermarked audio signals should be in between 0 and -1. The ODG values
of watermarked audio signals obtained after first level, second level and third level
watermark embedding are -0.38, -0.46 and -0.48, respectively These ODG values are
in between 0 and -I. These ODG values indicate that imperceptibility of watermarked
audio signals exhibit for all the ttiree levels of watermarks. The payload in three-level
108
Cfiapter 6
watermarking is increased by 39.68% of single-level watermarking while
imperceptibility (inaudibility) of audio signal is maintained.
Second watermarking scheme is the DFRFT-based digital image
watermarking scheme and it has been compared with existing DWT-based Kundur's
watermarking method. This technique is more secure than other techniques because
watermark can be embedded using a secret DFRFT powers and this information is
essential for extracting the watermark information. A novel non-blind extraction
method for DFRFT-domain watermarking scheme is suggested. Various simulation
results of this scheme under different attacks have been studied and analyzed. It has
been observed that in general, the watermarking scheme maintains imperceptibility of
watermarked images. Further, the average PSNR (averaged over five test images) for
DFRFT-domain image watermarking scheme is 51.29 dB while the average PSNR for
DWT-based method is 45.14 dB. DFRFT-domain image watermarking scheme has
been found superior for som:e of the attacks such as salt and peppers noise, median
filtering, AWGN, and JPEG compression than DWT-based Kundur's method. BER
performance of DFRFT-based method (BER<15%) under salt and peppers noise has
superior performance than DWT-based Kundur's method (BER>30%). In DFRFT-
based method, BER in extracted watermark for AWGN attack is less than DWT-
based Kundur's method. For DFRFT-based method, BER in extracted watermark is
approximately constant at 4% with increasing variance of AWGN. BER in extracted
watermark for DFRFT-based watermarking method for varying JPEG quality factor
(30%- 90%) is less than 5%.
Third watermarking scheme for digital images is based on combined DWT
and DFRFT domains with non-blind watermark extraction algorithm. Average PSNR
of watermarked image obtained by using hybrid (combined DWT and DFRFT
domains) method is 51.35 dB. This PSNR value indicates that this proposed scheme
satisfies imperceptibility feature of watermarking. BERs in extracted watermark for
histogram equalization attack for DWT and DFRFT-based methods are 44.62-48.88%
and 81.55-88.87%, respectively for the considered set of images. For hybrid domain
method, BER in extracted watermark is 0-28.32% for histogram equalization attack
for the same set of images. The BER performances for sharpening attack for DWT,
DFRFT and hybrid domain methods are 32.56-47.57%, 70.32-77.93% and 0-7.72%,
109
Cfiapter 6
respectively. Therefore, the proposed hybrid domain method exhibits better
robustness in comparison to watermarking performed in individual domain for
histogram equalization and sharpening attacks. At the same time, watermark
extraction performance for hybrid domain method for JPEG compression deteriorates
in comparison to individual domain watermarking schemes (DWT-based and DFRFT-
based methods).
6.2 Future Work
In this thesis, chirp-based digital audio watermarking scheme has been
presented and its performance in terms of BER of extracted watermark under different
audio attacks has been evaluated. This watermarking scheme can also be extended to
other multimedia signals such as ECG and images.
Digital image watermarking schemes in DFRFT-domain and combined DWT
and DFRFT domains can easily be extended to video signals. By using varying
DFRFT powers, multi-level watermarking can be developed for images and video
signals with higher embedding capacity.
10
^fe ererences
^ference
REFERENCES [1]. M. Natarajan, and M. Makhdumi, "Safeguarding the Digital Contents: Digital
Watermarking". DESIDOC Journal of Library & Information Technology, Vol. 29, No. 3, pp. 29-35, May 2009.
[2]. M. Barni, and F. Bartolini, "Watermarking Systems Engineering Enabling Digital Assets Security and Other Applications", Marcel Dekker, 2004.
[3]. C. S. Lu, "Multimedia Security: Steganography and Digital Watermarking Techniques for Protection of Intellectual Property", Idea Group Publishing, 2005.
[4]. 1. J. Cox, M. L. Miller and J. A. Bloom, "Watermarking Applications and Their Properties", Proceedings of International Conference on Information Technology: Coding and Computing (ITCC'2000), pp. 6-10, Las Vegas, Nevada, March 2000.
[5j. G. C. Langelaar, 1. Setyavvan, and R. L. Lagendijk, "Watermarking Digital Image and Video Data", IEEE Signal Processing Magazine, Vol. 17. No. 5, pp. 20-46, September 2000.
[6]. R. S. Alomari, and A. Al-Jabar, "A Fragile Watermarking Algorithm for Content Authentication", International Journal of Computing and Information Sciences, Vol. 2, No. 1. pp. 27-37, April 2004.
[7]. P. Meenakshi Devi, M. Venkatsan, and K. Duraiswamy, "A Fragile Watermarking Scheme for Image Authentication with Tamper Localization using Integer Wavelet Transform", Journal of Computer Science, Vol. 5. No. II, pp. 831 -837, 2009.
[8]. M. A. Dorairangaswamy, "A Novel Invisible and Blind Watermarking Scheme for Copyright Protection of Digital Images", IJCSNS International Journal of Computer Science and Network Security, Vol. 9. No. 4, pp. 71-78. April 2009.
[9]. A. A. Haj, and A. Odeh, "Robust and Blind Watermarking of Relational Database Systems", Journal of Computer Science, Vol. 4, No. 12. pp. 1024-1029, 2008.
[10]. N. V. Dharwadkar, and B. B. Amberkar, "An Efficient Non-Blind Watermarking Scheme for Color Images", International Journal of Computer Applications, Vol. 2, No. 3,pp. 60-66, May 2010.
[II]. C. Karabat, and M. Keskinoz, "Robust Non-Blind Detection for Spread Spectrum Watermarking System", Proceedings of International Conference on Intelligent Information Hiding and Multimedia Signal Processing, pp. 1221-1224, August 2008.
[12]. S. W. Foo, "Non-Blind Audio Watermarking using Compression-Expansion of Signals", Proceedings of the IEEE Asia Pacific Conference on Circuits and Systems, Nanyang University, Singapore, pp. 1288-1291, 2008.
[13]. M. Maes, T. Kalker, J.P. Linnartz, J. Talstra, G. Depovere, and J. Haitsma, "Digital Watermarking for DVD Video Copy Protection", IEEE Signal Processing Magazine. Vol. 17, No. 5, pp. 47-57, September 2000.
[14]. F. A. P. Petitcolas, "Watermarking Schemes Evaluation", IEEE Signal Processing Magazine, Vol. 17, No. 5, pp. 58-64, September 2000.
[15]. S. H. Kwok, C.C. Yang, and K.Y. Tarn, "Intellectual Property Protection for Electronic Commerce Applications", Journal of Electronic Commerce Research Vol 5 No I pp 1-13,2004.
Ill
I^fference
[16], I. J. Cox, J. Kilian, T. Leighton, and T. Shamoon, "Secure Spread Spectrum Watermarking for Multimedia", IEEE Transaction on Image Processing, Vol. 6, pp. 1673-1687, Dec.1997,
[17]. M. Abdullah, and F. Wahab, "Key Based Text Watermarking of e-Text Documents in an Object Based Environment using z-axis for Watermark Embedding", World Academy of Science, No. 46, pp. 199-202, 2008.
[18]. M. Y. Kim, 'Text Watermarking by Syntactic Analysis", Proceedings of 12"' WSEAS International Conference on Computers, Heraklion, Greece, pp. 904-909, July 2008.
[19]. M. D. Swanson, M. Kobayashi, and H. Tewfik, "Multimedia Data Embedding and Watermarking Technologies", Proceedings of the IEEE, Vol. 86, No. 6, pp. 1064-1087, June 1998.
[20]. R. B. Wolfgang, C. I. Podilchuk, and E. J. Delp, "Perceptual Watermarks for Image and Video", Proceedings of the IEEE, Vol. 87, No. 7, pp. 1108-1126, July 1998.
[21]. W. Bender, D. Gruhl, N. Moromoto, and A. Lu. "Techniques for Data Hiding", IBM Systems Journal, Vol. 35, Nos. 3&4, pp. 313-336, 1996.
[22]. D. Kirovsk, C. D. Jones, and H. S. Malvar, "Spread Spectrum Watermarking of Audio Signals", IEEE Transactions on Signal Processing. Vol. 51, No. 4, pp. 1020-1033, April 2003.
[23]. H. Malik, S. Khokhar. and A. Rashid, "Robust Audio Watermarking using Frequency Selective Spread Spectrum Theory", Proceedings of IEEE International Conference on .Acoustics, Speech and Signal Processing, Vol. 5, pp. 385-388. May 2004.
[24]. M. Ejima, and A. Miyazaki, "A Wavelet Based Watermarking for Digital Images and Video", lElCE Transactior Fundamentals, Vol. E83-A, No. 3, pp. 532-540, March 2000.
[25]. M. Kutter and F. A. P. Petitcolas, "Fair Evaluation Methods for Image Watermarking Systems", Journal of Electronic Imaging, Vol. 9, No. 4, pp. 445-455. October 2000.
[26]. M. Potdar, S. Han, and E. Chang, "A Survey of Digital Image Watermarking Techniques", Proceedings of IEEE International Conference on Industrial Infonnatics, pp. 709-716,2005.
[27]. N. Provos and P. Honeyman, "Hide and Seek: An Introduction to Steganography", IEEE Journal of Security and Privacy, Vol. l,No. 3, pp. 32-44, 2003.
[28]. R. G. Schyndel, A. Tirkel, and C. F. Osborne, "A Digital Watermark", Proceedings of IEEE International Conference on Image Processing, [CIP-1994, pp. 86-90, 1994.
[29]. L. Li, and B. L. Guo, "Localised Image Watermarking in Spatial-Domain Resistant to Geometric Attacks", International Journal of Electronics and Communications, Vol. 63, No. 2, pp. 123-131, February 2009.
[30]. B. Surekha, and G. N. Swamy, "A Spatial Domain Public Image Watermarking", International Journal of Security and its Applications, Vol. 5, No. 1, pp. 1-11, January 2011.
[31]. P. S. Huang, C. S. Chiang, C. P. Chang, and T. M. Tu, "Robust Spatial Watermarking Technique for Colour Images via Direct Saturation Adjustment", Proceedings of lEE Conference on Vision, Image and Signal Processing, Vol. 152. pp. 561-574, 2005.
12
'l^ffference
[32]. A. Basil, T. S. Das, S. Maiti, N. Islam, and S. K. Sarkar, "FPGA based Implementation of Robust Spatial-Domain Image Watermarking Algorithm", Proceedings of 4" International Conference on Computers and Devices for Communication, CODEC-2009. pp. 1-4, Kolkata, 2009.
[33]. D. Naveen, and A. Jhansi Rani, "Implementation of Psychoacoustic Model in Audio Compression using Munich and Gamma chirp Wavelets", International Journal of Engineering and Technology, Vol. 2, No. 5, pp. 1066-1072, 2010.
[34]. R. B. Durai C, and V. Duraisamy, "A Genetic Approach to Content-based Image Retrieval using DCT and Classification Technique", International Journal of Engineering and Technology, Vol. 2, No. 6, pp. 2020-2024, 2010.
[35]. Victor Namias, "The Fractional Order Fourier Transform and its Applications to Quantum Mechanics", IMA Journal of Applied Mathematics, Vol. 25, No. 3, pp 241-265, 1980.
[36]. A. C. McBride and F.H. Kerr, "On Namias's Fractional Fourier Transform", IMA Journalof Applied Mathematics, Vol. 39, No. 2, pp 159-175, 1987.
[37]. D. Mendlovic and H.M, Ozaktas, "Fractional Fourier Transforms and Their Optical Implementation-l", Journal of Optical Society of America, Vol. 10, No. 9, pp 1875-1881. September 1993.
[38]. H. M. Ozaktas and D. Mendlovic, "Fractional Fourier Transforms and Their Optical Implementation-II", Journal of Optical Society of America, Vol. 10, No. 12, pp 2522-2531,1993.
[39]. S. C. Pei, M.H. Yeh and C.C. Tseng, "Discrete Fractional Fourier Transform based on Orthogonal Projections", IEEE Transactions on Signal Processing, Vol. 47, No. 2. pp 1335-1348, Feb. 1999.
[40]. S. C. Pei and M.H. Yeh, "A Novel Method for Discrete Fractional Fourier Transform Computation", Proceedings of IEEE International Symposium on Circuits and Systems, ISCAS 2001, Vol. 2, pp. 585-588, 6-9 May 2001.
[41]. C. Candon, M.A. Kutty, and H.M. Ozaktas, "The Discrete Fractional Fourier Transform", IEEE Transactions on Signal Processing, Vol. 48, No. 5, pp. 1329-1337, May 2000.
[42]. H. M. Ozaktas, O. Arikan, 0. Kutty, and M.A Bazdagt, "Digital Computation of the Fractional Fourier Transform", IEEE Transactions on Signal Processing, Vol. 44, No. 9 pp. 2141-2150, September 1996.
[43]. S. C. Pei and M.H. Yeh, "Two Dimensional Discrete Fractional Fourier Transform", Signal Processing, Elsevier, Vol. 67,1^0. 1, pp. 99-108, May 1998.
[44]. G. Cariolaro, T. Erseghe, P. Kraniauskas, and N. Laurenti, "Multiplicity of Fractional Fourier Transforms and Their Relationships", IEEE Transactions on Signal Processing, Vol. 48, No.l, pp. 227-241, Jan. 2000.
[45]. B. Santhanam and J. H. McClellan, "The Discrete Rotational Fourier Transform", IEEE Transactions on Signal Processing, Vol. 44, No. 4, pp. 994-998, April 1996.
[46]. Gabor, "Theory of Communication", Journal of the lEE, Vol. 93, pp. 429-457, 1946.
ii:
^ference
[47]. S. G. Mallat, "A Theory for Multiresolution Signal Decomposition: The Wavelet Representation", IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. l i ,No. 7, pp. 674-693, July 1989.
[48]. P. Addison, "The Little Wave with the Big Future", Physics World, Vol. 17. pp. 35-39, March 2004.
|49]. K. P. Soman, K. I. Ramachandran, "Insight into Wavelets; From Theory to Practice", 2nd Edition, Prentice Hall of India, 2004.
[50]. Y. Xiong and Z. X. Ming, "Covert Communication Audio Watermarking Algorithm Based on LSB", International Conference on Communication Technology, lCCT-2006, pp. 1-4,2006.
[51]. N. F. Johnson, and S.C. Katzenbeisser, "A Survey of Steganographic Techniques", First Edition, Artech House Bosten, M.A., 2000.
[52]. B. S. Ko, R. Nishimura, and Y. Suzuki, "Time-Spread Echo Method for Digital Audio Watermarking", IEEE Transactions on Multimedia, Vol. 7, No. 2, pp. 212-221, April 2005.
[53]. H. O. Oh, J. W. Seok, J. W. Hong, and D. H. Youn, "New Echo Embedding Technique for Robust and Imperceptible Audio Watermarking", Proceedings of ICASSP 2001, pp. 1341-1344,2001.
[54]. Y. W. Liu, and J. 0. Smith, "Watermarking Sinusoidal Audio Representations by Quantization Index Modulation in Multiple frequencies", Proceedings of ICASSP 2004, Vol. 5, pp. 373-376,2004.
[55]. B. Chen and G. W. Worneli, "Quantization Index Modulation: A Class of Provably Good Methods for Digital Watennarking and Information Embedding", IEEE Transactions on Information Theory, Vol. 47, No. 4, pp. 1423-1443, May 2001.
[56]. B. Chen, and G. W. Worneli, "Digital Watermarking and Information Embedding using Dither Modulation", Proceedings of IEEE Second Workshop on Multimedia Signal Processing, M.I.T., U.S.A., pp. 273-278, 1998.
[57]. S. Yang, W. Tan, Y. Chen, and W. Ma, "Quantization-Based Digital Audio Watermarking in Discrete Fourier Transform Domain", Journal of Multimedia, Vol. 5, No. 2, pp. 151-158, April 2010.
[58], W. Zeng, and B. Liu, "Statistical Watermark Detection Technique without using Original Images for Resolving Rightful Ownerships of Digital Images", IEEE Transactions on Image Processing, Vol. 8, No. 11, pp. 1534-1548, November 1999.
[59]. K. Tanaka, Y. Nakamura, and K. Matsui, "Embedding Secret Information into a Dithered Multi-Level Image", Proceedings of Conference on Military Communications, pp. 216-220, 1990.
[60]. 1. Pitas, "A Method of Signature Casting on Digital Images" Proceedings of IEEE International Conference on Image Processing, ICIP-1996, Vol.3, pp. 215-218, 1996.
[61]. R. B. Wolfgang and E. J. Delp, "A Watermark for Digital Images" Proceedings of IEEE International Conference on Image Processing, ICIP-1996, Vol.3, pp. 219-222, 1996.
[62]. R. G. Schyndel, and C. Osborne, "A Two-Dimensional Watermark" Proceedings of DICTA-1993, pp. 378-383.
14
^ference
I63j. R. B. Wolfgang and E. J. Delp, "'Fragile Watermarking using the vw2d watermark". Proceedings of SPIE Conference on Security and Watermarking of Multimedia Contents, San Jose, Vol. 3657, pp. 204-213, 1999.
[64]. S. Walton, "Information Authentication for a Slippery New Age", Dr Dobbs Journal, Vol. 20, No. 4, pp. 18-26, April 1995.
[65]. M. Kutter, F. Jordan, and F. Bossen, "Digital Watermarking of Color Images using Amplitude Modulation", Journal of Electronic Imaging, Vol. 7, No. 2, pp. 326-332, April 1998.
[66]. L. Xie, J. Zhang and H. He, "Robust Audio Watermarking Scheme Based on Nonuniform Discrete Fourier Transform", Proceedings of IEEE International Conference on Engineering of Intelligent Systems, pp. 1-5, 2006.
[67]. J. Singh, P. Garg, and A. N. De, "Audio Watermarking using Spectral Modifications", International Journal of Signal Processing, Vol. 5, pp. 297-301, 2009.
[68]. X. Y. Wang and H. Zhao, "A Novel Synchronization Invariant Audio Watermarking Scheme Based on DWT and DCT", IEEE Transactions on Signal Processing, Vol. 54, No. 12, pp. 4835-4840, Dec. 2006.
[69]. S. Vongpraphip, and, M. Ketcham "An Intelligence Audio Watermarking Based on DWT-SVD using ATS", WRI Global Congress on Intelligent Systems, GCIS-2009, Vol. 3, pp. 150-154, May 2009.
[70]. Y. Wu and S. Shimamoto, "A Study on DWT-Based Digital Audio Watermarking for Mobile Ad Hoc Network". Proceedings of IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing, Vol. 2, pp. 247-251, 2006.
[71]. R. Vieru, R. Tahboub, C. Constantinescu and V. Lazarescu, "New Results using the Audio Watermarking Based on Wavelet Transform", Proceedings of International Symposium on Signals, Circuits and Systems, Vol. 2, pp. 441-444, 2005.
[72], X. Quan and H. Zhang, "Audio Watermarking Based on Psychoacoustic Model and Adaptive Wavelet Packets". Proceedings of 7th International Conference on Signal Processing, Vol. 3, pp. 2518-2521, 2004.
[73]. B. G. Haskell, A. Puri, and A. N. Netravali, "Digital Video: An Introduction to MPEG-2", New York: Chapman & Hall, 1997.
[74]. E. Koch, and J. Zhao, "Towards Robust and Hidden Image Copyright Labeling", Presented at Nonlinear Signal Processing Workshop, Thessaloniki, Greece, 1995.
[75]. F. M. Boland, J.J.K. O'Ruanaidh, and C. Dautzenberg, "Watermarking Digital Images for Copyright Protection", Proceedings of IEEE International Conference on Image Processing and Its Applications, pp. 321-326, 1995.
[76]. M. D. Swanson, B. Zhu, and A.H. Tewfik, "Transparent Robust Image Watermarking", Proceedings of International Conference on Image Processing, Lausanne, Switzerland, pp.211-214, 1996.
[77]. C. Podilchuck, and W. Zeng, "Image Adaptive Watermarking using Visual Models", IEEE Journal on Selected Areas of Communications, Vol. 16, pp. 525-539, May 1998.
[78]. A. B. Watson, "DCT Quantization Matrices Visually Optimized for Individual Images", Proceedings of SPIE Conference on Human Vision, Visual Processing and Digital Display-IV, pp. 202-216. 1992.
'^ference
[79]. A. B. Watson, G. Y. Yang, J. A. Solomon, and J. Villasenor, "Visibility of Wavelet Quantization Noise", IEEE Transactions on Image Processing, Vol, 6, No.8, pp. 1164-1175, Aug. 1997.
[80]. C. I. Podilchuck, and W. Zeng, "Digital Image Watermarking using Visual Models". Journal of SPIE Human Vision and Electronic Imaging-II. pp. 100-111. Feb. 1997.
[81]. A. G. Bors, and I. Pitas, "Image Watermarking using DCT Domain Constraints", Proceedings of IEEE International Conference on Image Processing, Lausanne, Switzerland, Vol. 3, pp. 231-234, 1996.
[82]. A. Piva, M. Barni, F. Bartolini, and Cappellini, "DCT-based Watermarking Recovering Without Resorting to the Uncorrupted Original Image", Proceedings of IEEE International Conference on Image Processing, Santa Barbara, CA, Vol. 1, pp. 520-523, 1997.
o [83]. W. Zeng, and B. Liu, "A Statistical Watermark Detection Technique Without using Original Images for Resolving Rightful Ownerships of Digital Images", IEEE Transactions on Image Processing, Vol. 8, No. 11, pp. 1534-1548, Nov. 1999.
[84]. W. D. Hong, L. D. Ming, Y. Jun and C. F. Xiong , "An Improved Chirp Typed Blind Watermarking Algorithm Based on Wavelet and Fractional Fourier Transform". Proceedings of IEEE International Conference on Images and Graphics, ICIG 2007, pp. 291-296.22-24 Aug. 2007.
[85]. Z. Feng, M.U. Xiaomin and Y. Shouyi. "Multiple Chirp Typed Blind Watermarking Algorithm Based on Fractional Fourier Transform", Proceedings of International Symposium on Intelligent Signal Processing and Communication Systems, ISPACS 2005, pp. 141-144, 13-16 Dec. 2005.
[86]. D. Kundur and D. Hatzinakos, " A Robust Digital Image Watermarking Method using Wavelet Based Fusion", Proceedings of IEEE International Conference on Image Processing, ICIP 1997, Vol. 1, pp. 544-547, 1997.
[87]. G. Wang, X.G. Xia, and V.C. Chen, "Adaptive Filtering Approach to Chirp Estimation and Inverse Synthetic Aperture Radar Imaging of Maneuvering Targets", Journal of Optical Engineering (SPIE Digital Library), Vol. 42, Issue 1, pp. 190-199, 2003.
[88]. M. Barbu, E. J. Kaminsky, and R. E. Trahan, "Fractional Fourier Transform for Sonar Signal Processing", Proceedings of MTS/IEEE, OCEANS-2005, Vol. 2, pp. 1630-1635, 2005.
[89]. S. Erkiiciik, S. Krishnan. and M. Zeylinglu, "Robust Audio Watermarking using Chirp Based Technique", Proceedings of IEEE International Conference on Multimedia, Vol. 2. pp. 513-516, July 2003,
[90]. S. Erkuciik, S. Krishnan, and M. Zeylinglu, "A Robust Audio Watermark Representation Based on Linear Chirps," IEEE Transactions on Multimedia, Vol. 8, No. 5, pp. 925-936, October 2006.
[91]. A. Ramalingam and S. Krishnan, "A Robust image Watermarking using a Chirp Detection-based Technique", Proceedings of lEE Conference on Vision, Image and Signal Processing, Vol, 152, Issue No. 6, pp. 771-778, 2005.
[92]. J. M. Blackledge and E. Coyle, "Self-Authentication of Audio Signals by Chirp Coding", Proceedings of the 12th International Conference on Digital Audio Effects (DAFx-09), Como, Italy, pp. 1-8, September 1-4, 2009.
l^gference
[93]. I. J. Cox, J. P. Linnartz, and T. Shamoon, "Public Watermarks and Resistance to Tampering", Proceedings of International Conference on Image Processing, ICIP-1997, pp. 26-29, 1997.
[94]. .1. C. Davis, "Protecting Intellectual Property in Cyberspace", IEEE Technology and Society Magazine", Vol. 17, No. 2, pp. 12-25, 1998.
[95J. F. Hartung, and M. Kuttar, "Multimedia Watermarking Techniques", Proceedings of the IEEE Conference, Special Issue on Identification and Protection of Multimedia Information, Vol. 87, No. 7, pp. 1079-1107, July 1999.
[96]. S. Craver, N, Memon, B.L. Yeo, and M. Yeung, "Can Invisible Watermarks Resolve Rightful Ownerships", Proceedings of SPIE Storage and Retrieval for Still Image and Video Databases, Vol. 3022, pp. 310-321, Feb. 1997.
[97]. W. Zeng, "Visual Optimization in Digital Image Watermarking", Proceedings of Workshop on Multimedia and Security at ACM Multimedia'99, pp.81-88, October 1999.
[98]. P. Meerwald, and S. Pereira, "Attacks, Applications and Evaluation of Known Watermarking Algorithms with Checkmark", Proceedings of SPIE symposium on Electronic Imaging, Vol. 4675, Jan. 2002.
[99]. A. Z. Tirkel, G. A. Rankin, R. G. V. Schyndel, W. J. Ho, N. R. A. Mee, and C. F. Osborne, "Electronic Water Mark", DICTA-1993, Macquarie University, Sydney, pp. 666-672. Dec. 1993.
[100].N. P. Sheppard, R. Safavi-Naini and P. Ogunbona, "On Multiple Watermarking", Proceedings of ACM Multimedia Conference, Ottawa, pp. 3-6, Canada, 2001.
[101]. International Telecommunication Union (ITU), "Method for Objective Measurements of Perceived Audio Quality (PEAQ)", ITU-R BS.1387, 1998.
[1021.P. Kabal, "An Examination and Interpretation of ITU-R BS.1387: Perceptual Evaluation of Audio Quality (PEAQ)", Technical Report, McGill University, Version 2, 2002.
[ 103]. http://sound.media.mit.edu/mpeg4/audio/sqam/Dec.2008.
[104]. J. Benesty, M. M. Sondhi, and Y. Huang, "Springer Handbook of Speech Processing", Springer-Verlag Berlin Heidelberg, 2008.
[105]. P. W. Wong, "A Public Key Watermark for Image Verification and Authentication", Proceedings of IEEE International Conference on Image Processing, Chicago, Illinois, USA, Vol. 1, pp. 425-429. October 1998.
[106].l. Pitas, and T.H. Kaskalis, "Apply Signatures on Digital Images", Proceedings of IEEE workshop on Non-linear Signal and Image Processing, Neos Marmaros, Greece, Vol. 1 , pp. 460-463, June 1995.
[107].R. B. Wolfgang, and E. J. Delp, "A Watermark for Digital Images", Proceedings of International Conference on Image Processing, Vol. 3, pp. 219-222, Sept. 1996.
[108].Kitamytra, S. Kanai, T. Kanai. and T. Kishinami. "Copyright Protection of Vector Map using Digital Watermarking Method based on Discrete Fourier Transform", Proceedings of IEEE International Symposium on Geosciences and Remote Sensing, pp.1191-1193,2001.
117
^ference
[109J.P. Premaratne, and C.C. Ko, "A Novel Watermarking Embedding and Detection Scheme for Images in DFT domain", Proceedings of 7th International Conference on hnage Processing and Its Applications, pp. 780-783, 1999.
[110]. V. Solachidis, and I. Pitas, " Circularly Symmetric Watermark Embedding in 2D-DFT domain", IEEE Transactions on Image Processing, Vol. 10, No. 11, pp. 1741-1753, Nov. 2001.
[111].J. R. Hernndez, M. Amado, and P.P. Gonzlez, "DCT-Domain Watermarking Techniques for Still Images Detector Performance Analysis and a New Structure", IEEE Transactions on Image Processing, Vol. 9, No. 1, pp. 55-68, Jan. 2000.
[112].M. Barni, F. Bartolini, V. Cappellini and A. Piva, "A OCT Domain System for Robust Image Watermarking", Signal Processing (Special Issue on Watermarking ), Vol. 66, No. 3, pp. 357-372, May 1998.
[I 13].X. Huang, Y. Luo, M. Tan, and D. Lin, "An Image Digital Watermarking based on DWT in Invariant Wavelet Domain", Proceedings of IEEE International Conference on Images and Graphics, ICIG 2007, pp. 329-336, 22-24 Aug.2007.
[il4].C. Yaw Low, A.B. Jin Teoh and C. Tee, "Fusion of LSB and DWT Biometric Watermarking for Offline Handwritten Signature", Proceedings of Congress on Image and Signal Processing, CISP'08. Vol. 5, pp. 702-708, May 2008.
[115]. I. Djurovic, S. Stankovic, and I, Pitas, "Digital Watermarking in the Fractional Fourier Transformation Domain", Journal of Network and Computer Applications, Vol. 24, No. 4, pp. 167-173, July 2001.
[1161.F.Q. Yu, Z.K. Zhang, and M.H. Xu, "A Digital Watermarking Algorithm for Image Based on Fractional Fourier Transform", Proceedings of 1st IEEE International Conference on Industrial Electronics and Applications, ICIEA 2006, pp. 1-5, 2006.
[117].R. Petrovic, and D, Yang, "Efficient and Secure Forensic Marking in Compressed domain Descriptions/ Claims", US Patent Application Number 20090326961, 16 April 2009.
[1 ]8].B. G. Mobasseri, and R. J. Berger, "A Foundation for Watermarking in Compressed Domain", IEEE Signal Processing Letters, Vol. 12, No. 5, pp. 399-402, May 2005.
[119].S. Biswas, S. R. Das, and E. M. Petriu, "An Adaptive Compressed MPEG-2 Video Watermarking Scheme", IEEE Transactions on Instrumentation and Measurement, Vol. 54, No. 5, pp. 1853-1861, October 2005.
[120], A. M. Alattar, E.T. Lin, and M. U. Celik, "Digital Watermarking of Low Bit-Rate Advanced Simple Profile MPEG-4 Compressed Video", IEEE Transactions on Circuits and Systems for Video Technology, Vol. 13, No. 8, pp. 787-800, August 2003.
[121].A. A. Haj, and A. Abu-Errub, "Performance Optimization of Discrete Wavelet Transform Based Image Watermarking using Genetic Algorithms", Journal of Computer Science, Vol. 4, No. 10, 2008.
[122].R. Safabakhsh, S. Zaboli, and A. Tabibiazar, "Digital Watermarking on Still Images using Wavelet Transform" Proc. of International Conference on Information
. Technology: Coding and Competing- ITCC-2004.
[123]. A. M. Kothari, A. C. Suthar, and R. S. Gajre, "Performance Analysis of Digital Image Watermarking Technique- Combined DWT- DCT over individual DWT", International Journal of Advanced Engineering & Applications, Jan. 2010.
118
(Reference
[124J. A. A. Haj, "'Combined DWT-DCT Digital Image Watermarking'", Journal of Computer Science, Vol, 3, No, 9, pp. 740-746, Sept. 2007.
[125]. A. A. Abdulfetah, X. Sun, H. Yang, and N. Mohammad, "Robust Adaptive Image Watermarking using Visual Models in DWT and DCT domain". Information Technology Journal, Vol. 9, No. 3, pp. 460-466, 2010.
[126], X. Rang, J. Huang, Y. Q. Shi, and Y. Lin, "A DWT-DFT Composite Watermarking Scheme Robust to Both Affme Transform and JPEG Compression", IEEE Transactions on Circuits and Systems for Video Technology, Vol. 13, No. 8, pp. 776-786, August, 2003.
119
Jippencfi^
Jlppendh(^
TEST IMAGES USED FOR WATERMARKING
(a) 'LENA' Image
(b) 'BOATS' Image
120
Appendhc
(c) 'GOLDHILL' Image
(d) 'PEPPERS' Image
121
Jippendv(,
(e) 'BABOON' Image
122