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