lightweight robust watermarking for …repository.um.edu.my/505/1/final_pdf.pdf · lightweight...

i

LIGHTWEIGHT ROBUST WATERMARKING FOR DIGITAL

IMAGES USING MULTIPLE DOMAINS

THIBA A/P NARASIMHA BHARATHEYAR

DISSERTATION SUBMITTED IN FULFILMENT OF THE

REQUIREMENTS FOR THE DEGREE OF MASTER OF

COMPUTER SCIENCE

FACULTY OF COMPUTER SCIENCE AND INFORMATION

TECHNOLOGY

UNIVERSITY OF MALAYA

KUALA LUMPUR

JULY 2010

ii

UNIVERSITI MALAYA

ORIGINAL LITERARY WORK DECLARATION

Name of Candidate: (I.C/Passport No: ) Registration/Matric No: Name of Degree: Title of Project Paper/Research Report/Dissertation/Thesis ("this Work"): Field of Study:

I do solemnly and sincerely declare that:

(1) (2) (3) (4) (5) (6)

I am the sole author/writer of this Work; This Work is original; Any use of any work in which copyright exists was done by way of fair dealing and for permitted purposes and any excerpt or extract from, or reference to or reproduction of any copyright work has been disclosed expressly and sufficiently and the title of the Work and its authorship have been acknowledged in this Work; I do not have any actual knowledge nor do I ought reasonably to know that the making of this work constitutes an infringement of any copyright work; I hereby assign all and every rights in the copyright to this Work to the University of Malaya ("UM"), who henceforth shall be owner of the copyright in this Work and that any reproduction or use in any form or by any means whatsoever is prohibited without the written consent of UM having been first had and obtained; I am fully aware that if in the course of making this Work I have infringed any copyright whether intentionally or otherwise, I may be subject to legal action or any other action as may be determined by UM.

Candidate's Signature Date

Subscribed and solemnly declared before,

Witness's Signature Date

Name: Designation:

iii

Abstract

Due to rapid growth of the Internet, digital contents such as text, audio, video

and images are easily exchanged and disseminated by worldwide users. As a result,

unauthorized duplication arises and becomes a serious issue which needs to be combat

with. Therefore, several techniques has been researched and developed in guarding

these contents; there are cryptography, encryption, steganography and watermarking.

Out of these solutions, digital watermarking has emerged as a reliable technique in

safeguarding the digital contents; which works on imperceptibly altering an original

digital content to embed a message about the content, which can be later used for

authentication purpose.

Recently, many extensive researches have been carried out in the field of digital

image watermarking due to various types of attacks against the digital images.

Geometrical and image processing operations are two main types of attacks which can

easily distort digital images. In order to resist those attacks, robust and fragile

watermarks are required. A robust watermark is significantly essential in digital image

watermarking whereby the embedded watermark in an image can be easily recovered

even after it has been manipulated with geometrical or other attacks. Fragile watermark

has the ability to detect any unauthorized changes which has been made to the

embedded watermark.

In this dissertation, we introduce a new lightweight robust watermarking

approach based on edge features for spatial domain and wavelet domain. Edge features

are chosen as a key factor of our proposed watermarking technique, since the edge

information is considered to be robust feature which can withstand the scaling attack in

a watermarked image. Moreover, embedding strategy works by embedding watermark

iv

in robust edges of a cover image. Meanwhile, an inverse process of embedding step is

required in the watermark detection stage. In addition, the original cover image is not

needed in validating the detected watermark.

Experiments results demonstrated higher visual quality can be gained in spatial

domain, due to less distortion to the watermarked image. On the other hand, wavelet

domain revealed faster execution time with thresholding technique compared to spatial

domain, which is based on edge detection through convolution approach. By overall

experiment observations, we noticed that the proposed scheme is a simple algorithm

with low computational complexity and consists of moderate image distortions. These

are the three main factors contribute towards the robustness feature of our proposed

scheme. Therefore this scheme turns out to be a lightweight watermarking and it is

suitable to be implemented in many applications.

v

Acknowledgements

I would like to express the deepest appreciation to my dissertation supervisor,

Dr. Woo Chaw Seng, for being supportive throughout my dissertation. This dissertation

would not be completed without his invaluable guidance and comments. I would like to

thank the examiner’s committee member, Dr. S. Raviraja for his encouragement and

insightful comments.

I also like to express my gratitude to my fellow friends and colleagues, specially

to Hairul Aysa , Rajeswari, Lalitha, Noorliza and Farrieza who helped me a lot during

my hard times.

Sincere thanks to my husband Kumara Sastri, mother, sisters and brothers for

their moral and financial supports along the way. To them I dedicate this dissertation.

vi

Table of Contents

Original Literary Work Declaration …....….…..…..……...………………………….ii

Abstract...........................................................................................................................iii

Acknowledgements……..................................................................................................v

List of Figures…………...............................................................................................viii

List of Tables………….. ................................................................................................ix

List of Symbols and Abbreviations................................................................................x

Chapter 1: Introduction .................................................................................................1

1.1 Background ...........................................................................................................1

1.2 Motivation.............................................................................................................1

1.3 Overview of Digital Watermarking ......................................................................2

1.4 Goal and Objectives ..............................................................................................4

1.5 Project Scope.........................................................................................................4

1.6 Dissertation Organization......................................................................................5

Chapter 2: Literature Review........................................................................................7

2.1 Digital Image Watermarking.................................................................................7

2.2 Watermarking algorithm .......................................................................................8

2.3 Watermark properties............................................................................................9

2.4 Watermarking domain.........................................................................................10

2.5 Watermark embedding classification..................................................................11

2.6 Watermark detection techniques.........................................................................12

2.7 Watermarking Applications ................................................................................12

2.8 Robust Watermarking Techniques......................................................................13

2.8.1 Spatial domain..........................................................................13

2.8.2 Transform domain....................................................................14

2.8.3 Scale normalization and flowline curvature.............................17

2.8.4 Features point ...........................................................................18

2.8.5 Support Vector Machine ..........................................................18

2.9 Lightweight watermarking techniques................................................................19

2.9.1 Buyer-seller watermarking protocol ........................................19

2.9.2 Optimal Differential Energy Watermarking of DCT Encoded

Images and Video.....................................................................19

2.9.3 MMS Content Copyright Protection using Watermarking ......19

vii

2.10 Chapter Summary................................................................................................20

Chapter 3: Methodology...............................................................................................22

3.1 Introduction.........................................................................................................22

3.2 Generic Watermarking Process...........................................................................22

3.3 Proposed Watermarking Scheme ........................................................................23

3.3.1 Edge Detection.........................................................................24

3.3.2 Spatial Domain Watermarking.................................................26

3.3.3 Wavelet Domain Watermarking ..............................................30

3.3.3.1. Discrete Wavelet Transform ..............................................30

3.3.3.2. Level-One 2-Dimensional Discrete Wavelet Transform ...32

3.3.3.3. Level-Two 2-Dimensional Discrete Wavelet Transform...37

3.4 Chapter Summary................................................................................................41

Chapter 4: Experiments and Results...........................................................................42

4.1 Introduction.........................................................................................................42

4.2 Experiment Setup................................................................................................42

4.3 Interim Experiment Results.................................................................................43

4.3.1 LiREF watermarking in spatial domain ..................................44

4.3.2 LiREF watermarking in wavelet domain ................................46

4.4 Results Analysis ..................................................................................................47

4.4.1 LiREF watermarking in spatial domain ..................................47

4.4.2 LiREF watermarking in wavelet domain ................................49

4.5 Discussion ...........................................................................................................58

4.6 Overall Analysis..................................................................................................60

4.7 Chapter Summary................................................................................................64

Chapter 5: Conclusion.... ..............................................................................................66

5.1 Summary of the Dissertation...............................................................................66

5.2 Achievements......................................................................................................67

5.3 Future work .........................................................................................................68

References…………...... ................................................................................................69

viii

List of Figures

Figure Page

Figure 3.1 Generic watermarking process ......................................................................22

Figure 3.2 Watermark embedding steps in spatial domain .............................................27

Figure 3.3 Watermark detection steps in spatial domain ................................................29

Figure 3.4 Level-one 2-D DWT decomposed image......................................................31

Figure 3.5 Level-two 2-D DWT decomposed image......................................................31

Figure 3.6 Watermarks embedding steps in Level-one 2-D DWT .................................36

Figure 3.7 Watermarks detections steps in Level-one 2-D DWT...................................36

Figure 3.8 Watermarks embedding steps in Level-two 2-D DWT .................................38

Figure 3.9 Watermarks detection steps in Level-two 2-D DWT ....................................40

Figure 4.1 Cover images used for experiments...............................................................44

Figure 4.2 Edge detections for Lena image ....................................................................45

Figure 4.3 Cover image and its watermarked copy.........................................................45

Figure 4.4 Level-one 2-D DWT watermarking...............................................................46

Figure 4.5 Level-two 2-D DWT watermarking ..............................................................47

Figure 4.6 Comparison of p value among cover images in spatial domain. ..................49

Figure 4.7 p values for H1, V1, D1 sub bands of cover images; Threshold > 10...........51

Figure 4.8 p values for H1, V1, D1 sub bands of cover images; Threshold > 50..........51

Figure 4.9 p values for H1, V1, D1 sub bands of cover images; Threshold > 70..........52

Figure 4.10 p values for H1, V1, D1 sub bands of cover images; Threshold > 90........52

Figure 4.11 p values for H1, V1, D1 sub bands of cover images; Threshold > 130…...53





Figure 4.16 p values for H2, V2, D2 sub bands of cover images; Threshold > 130......57

Figure 4.17 Comparisons p values between Level-one and Level-two 2-D DWT........58

ix

List of Tables

Table Page

Table 4.1 LiREF watermarking experiment results in spatial domain............................48

Table 4.2 LiREF watermarking experiment results in Level-one 2-D DWT .................50

Table 4.3 LIREF watermarking experiment results in Level-two 2-D DWT.................54

Table 4.4 Comparison p values between Level-one and Level-two 2-D DWT.............57

Table 4.5 Execution time, PSNR and MSE in spatial domain........................................61

Table 4.6 Execution time, PSNR and MSE in level-one 2-D DWT...............................61

Table 4.7 Execution time, PSNR and MSE in level-two 2-D DWT...............................62

Table 4.8 PSNR values of recovered watermark from embedding schemes ..................64

x

List of Symbols and Abbreviations

1-D DWT One-Dimensional Discrete Wavelet Transformation

2-D DWT Two-Dimensional Discrete Wavelet Transformation

DCT Discrete Cosine Transform

DCT-SVD Discrete Cosine Transform Singular Value Decomposition

DDWT Distributed Discrete Wavelet Transformation

DFT Discrete Fourier Transform

DHT Discrete Hartley Transform

D-MR-DCT Distributed Multi-Resolution Discrete Cosine Transform

DRM Digital Rights Management

DT-CWT Dual Tree-Complex Wavelet Transform

DWT Discrete Wavelet Transform

DWT-SVD Discrete Wavelet Transform Singular Value Decomposition

ECC Error Correcting Code

Fuzzy-ART Fuzzy Adaptive Resonance Theory

HVS Human Visual System

ICA Independent Component Analysis

IDWT Inverse Discrete Wavelet Transform

JPEG Joint Photographic Experts Group

LiREF Lightweight Robust Watermarking Using Edge Features in Multiple

Domains

LSB Least Significant Bit

MMS Mobile Multimedia Service

MPEG Moving Picture Experts Group

MSE Mean Square Error

xi

NA Not Applicable

NVF Noise Visibility Function

PSNR Peak to Signal to Noise Ratio

RST Rotation, Scaling and Translation

SVD Singular Value Decomposition

SVM Support Vector Machine

1

Chapter 1: Introduction

1.1 Background

Rapid development of computer technologies and the Internet has made the

sharing and duplication of copyrighted material such as digital images, audio, music and

software becomes easier than before (Jun, Chi & Zhuang 2007; Reddy, Prasad & Rao

2009; Wang, Hou & Yang 2009). Moreover, there are many web sites which allows

user to upload or download unauthorized materials easily; which tends to promote

internet content piracy. One of the recent problems is the online leak of the unfinished

version of “X-Men Origins” movie before the film’s release, considered as serious

effect to the film makers as they invested huge amount of money on this blockbuster

movie especially during the economy downturn period (Lisa 2009). Therefore,

copyright protection has become a social issue.

1.2 Motivation

Digital watermarking is a promising solution for protecting copyrighted digital

materials. It permits the owner of digital content to imperceptibly alter an original

digital content to embed a message about the content itself, which can be later used to

differentiate the original from the fake copy (Cox et al. 2008). Besides, it is useful for

assisting the authorities to diminish piracy and to ensure that the copyright holders to

receive their proper payment of royalties for their hard work (ScienceDaily 2008). For

example, record companies such as Sony Music and Universal Music Group embedded

anonymous watermarks into songs so that it can help them to trace the origins of the

illegally copied material (ScienceDaily 2008).

2

Watermarking works better than cryptography in copyright protection as it can

guard content even after it is being decrypted, it also survive from reencryption,

compression, digital-to-analog conversion and file format changes. Watermarking

within digital images is significantly stressed nowadays and many researches have been

performed in this area. Robustness and imperceptibility are two essential requirements

of digital image watermarking techniques (Dinghui, Haixia & Chao 2007). Digital

images can be easily altered from its original form using geometrical and image

processing operations, so robustness feature in image watermarking plays an important

role to authenticate the image. To improve imperceptibility of a watermark, perceptual

similarity between the original and watermarked image should be high.

The main contribution of this dissertation is in lightweight robust digital

watermarking for images. We investigated several robust watermarking and lightweight

watermarking methods for digital image. Next, we developed new robust lightweight

watermarking approaches based on edge features in spatial and wavelet domain.

1.3 Overview of Digital Watermarking

An overview of digital watermarking is briefly stated in this section including

essential terms in watermark, watermarking algorithm, watermark domain, watermark

embedding techniques, watermark properties, watermark detection techniques and its

applications. The detailed explanations on these topics is included in Chapter 2.

The following list explains several essential terms which are widely used in the

field of watermarking (Cox et al. 2008). Some of these terms are adopted in preparation

of this dissertation.

3

� Cover image or host image

The original unaltered image is referred as a cover image, which will hides or

“covers” the watermark or the secret message.

� Watermarked image

This is a cover image which has been embedded with watermark.

� Watermark

This is the secret message in the form of an image or pseudo random binary bits

to be embedded in a cover image.

� Watermark embedding

This is a process of altering and encoding a cover image with watermark.

Outcome of this process is a watermarked image.

� Watermark detection or extraction

This is a process of decoding the watermark from a watermarked image.

Generally, a watermarking algorithm consists of watermark embedding and

watermark detection steps. The embedding and detection processes can be implemented

in either spatial domain or transform domain of a cover image. Meanwhile, the method

of embedding watermark can be classified into visible or invisible watermarking. The

effectiveness of a watermarking algorithm is measured through fidelity, data payload,

robustness, and security properties (Cox et al. 2008; Wang, Hou & Yang 2009; Luo &

Tian 2008). There are two different techniques applied in validating the detected

watermark, such as blind and non-blind detection.

Digital watermark has been widely applied for the purpose of copyright

protection, fingerprinting, copy control, broadcast monitoring, and data authentication

(Wang, Yang & Cui 2008).

4

1.4 Goal and Objectives

The goal of this dissertation is to design and develop a lightweight robust

watermarking technique for digital images.

Following are the objectives in order to achieve the above goal:

� To study and review of existing literatures of robust and lightweight

watermarking for digital images.

� To analysis and design a new algorithm for robust lightweight

watermarking to improve weakness in current literatures.

� To implement, test and evaluate the proposed new algorithm developed.

1.5 Project Scope

The proposed lightweight robust digital watermarking is focused on gray scale

digital images. Watermark embedding and detection are employed through the edge

features of a cover image. The selected domains for our proposed scheme are based on

spatial and wavelet domains. Scaling attack is applied for testing the robustness against

geometrical distortions. A simple and a fast algorithm is employed in order to enhance

the robustness feature in this proposed watermarking approach. Further, this idea tends

to be blind watermarking detection since original image is not required for watermark

validation purpose.

5

1.6 Dissertation Organization

The dissertation is organized as follows. In Chapter 1, we defined watermarking

as a prominent solution for copyright protection, which works better than cryptography

(Cox et al. 2008). Next, a brief explanation on several essential terms used in the field

of watermarking is presented. This is followed by objective and scope definitions for

our proposed watermarking scheme.

Chapter 2 is describes of watermarking algorithm, properties, domain,

embedding classification, detection techniques and its applications. Final section

discussed about several existing robust watermarking methods and lightweight

watermarking methods for mobile platform. The corresponding method’s strengths and

weaknesses are highlighted. At the end, we define and describe a new lightweight

robust watermarking scheme using simple and fast algorithm.

The Chapter 3 explanation starts with the descriptions of generic watermarking

process and continued with discussion on our proposed watermarking scheme for spatial

and wavelet domains. A detailed step has been drawn out for watermark embedding and

detection process. In addition, edge features is employed in developing the scheme and

it is designed for gray scale digital images.

The experiment setting and experiment results of the proposed scheme are

included in Chapter 4. Five different cover images are evaluated during the experiments

and purposely attacked with scaling distortions. Next, the experiments results are

tabulated into tables and charts information. Detailed discussions, comparisons and

overall analysis are performed using this information.

6

Conclusion is made in Chapter 5. This includes the achievements of our

proposed lightweight robust watermarking scheme. Future works are also been

highlighted for further development of this scheme.

7

Chapter 2: Literature Review

2.1 Digital Image Watermarking

Cox et al. (2008) reveals that there is overlapping technical areas among

watermarking, steganography and information hiding. Generally watermarking refers to

the practice of imperceptibly altering a cover work to embed a message about the cover

work (Cox et al. 2008). Steganography is defined as the practice of imperceptibly

altering a cover work to embed a secret message. This secret message is unrelated to the

cover work. Meanwhile information hiding is defined as creating imperceptible

information as in watermarking or keeping the existence of the information secret. So it

can be concluded that watermarking and steganogarphy as the derived approaches from

information hiding.

Watermarking can be applied for both analog and digital media. One of the

analog approaches can be found on paper watermark such as currency notes (Cox et al.

2008), whereby an invisible portrait is embedded directly during the currency note

making process and only becoming visible as a result of a special viewing process.

Besides being invisible, the watermark signifies the authenticity of the note. Digital

watermarking means embedding secret messages within digital media such as text,

audio, video and image and can be extracted using specific algorithm. As commonly

known, digital media can be easily shared over the Internet using various

communication technologies and the watermark can be removed from the content for

the redistributions purpose. Due to this reason, a robust watermark is needed in

preventing unauthorized access to copyrighted digital media in wide ranges of

applications. This research work covers the robust watermark for digital images.

8

Robust watermarking for digital images can be defined as a method of

embedding watermarking signals or codes into digital images which can withstand

against the geometric distortions such as Rotation, Scaling And Translation (RST) and

non-geometric attacks such as Gaussian noise, Gaussian blur, contrast adjustment and

histogram equalization. Although many approaches has been introduced earlier in this

field, but it still shows some techniques severe to geometric attacks and need to be paid

higher attention in improving it.

In this chapter, we reviewed available literatures of robust watermarking

schemes with their related strength and weaknesses in Section 2.7 together with

available literatures in the area of lightweight watermarking for mobile platform in

Section 2.8. Meanwhile the following Section 2.1 to Section 2.6 includes a brief

explanation on digital watermarking from Section 1.2.

2.2 Watermarking algorithm

Watermarking algorithm consists of embedding and detection steps. A good

watermark need to be constructed (Lu, Lu & Chung 2006), before the embedding steps

takes place, since a well structured watermark can improve the watermark embedding

capacity and quality of watermarked image. Once the watermark is constructed, it is

embedded along with a chosen optional key within the cover image through selected

embedding algorithm. In addition, embedding step can works on spatial or frequency

domain. Finally, the end product of embedding step is the watermarked image which

can be classified as either visible or invisible watermark on it. Embedding domain and

its classification is further discussed in Section 2.3 and Section 2.4.

9

Detection is the reverse process of embedding. Further, it is identified as the

process of authenticating the watermarked image. As initial step, the watermarked

image is manipulated accordingly using detection algorithm whereby the embedded

watermark is located and extracted. The extracted watermark is then compared with the

original watermark and key which has been used in earlier in embedding step. In

practice, some algorithm may or may not require the presence of original image which

is referred as either blind or non-blind watermarking. These approaches are further

explained in Section 2.5. Further, a statistical computation is made based on this

comparison in order to verify the existence of watermark. Details of this watermark

algorithm is further discussed in Chapter 3.

2.3 Watermark properties

Watermark properties can be categorized as fidelity, payload, robustness,

security (Cox et al. 2008; Wang, Hou & Yang 2009; Luo & Tian 2008). The following

section describes the above requirements in brief.

a) Fidelity or Imperceptibility

Fidelity refers to the perceptual similarity between the cover image and

watermarked image. The embedded watermark should not degrade the cover

image perceptually; cover image and watermarked version should appear

similar for user’s view.

b) Data Payload

Data payload is defined as the number of watermark bits that can be encoded

in a cover image. The amount of data payload depends on the size of the

cover image. Higher data payload will cause more watermark bits to be

10

encoded in the cover image but in contrast it tends to reduce fidelity and

robustness rate. The amount of information that can be stored in the

watermark is based on application and the embedding-algorithm quality.

c) Robustness

A robustness feature is measured by the capability of extracting watermark

from a watermarked image even after it has gone through geometric or non-

geometric attacks. Higher robustness rate will increase the value of the

watermarked image but the fidelity rate will decrease. In addition, cost of

complexity tends to increase at the same time.

d) Security

A watermarked image is considered as secured if it is able to defeat hostile

attacks. A hostile attack refers to any process specifically designed to thwart

the watermark’s purpose such as unauthorized removal, embedding and

detection.

2.4 Watermarking domain

Two main domain used in digital image watermarking are spatial domain and

transform domain.

a) Spatial domain

In spatial domain, watermarking embedding is done by directly modifying

the pixel value of a cover image and the most common technique is by

altering least significant bit (LSB) of a cover image (Tirkel, Van Schyndel &

Osborne 1994). Overall, embedding watermarks in this particular domain

11

seems to be imperceptible and easy to implement since the computational

complexity is lower compared to other domain or techniques. However it is

not robust and vulnerable to compression, filtering and rotation attacks.

b) Transform or frequency domain

In this approach the selected cover image need to be transformed into

frequency domain using transformation approach (Woo, Du & Pham 2006;

Agarwal & Goyal 2007; Tang & Wang 2008; Liu et al. 2005); Discrete

Cosine Transform (DCT), Discrete Fourier Transform (DFT), Discrete

Wavelet Transform (DWT) or other transform approach. The transform

coefficient is modified with embedded watermarks, as a result, the inverse

transformed image produces the watermarked image. Although this approach

seems to be complicated and derive higher computational costs, it is more

robust against attacks compare to spatial domain.

2.5 Watermark embedding classification

Methods of embedding the watermark are classified into visible or invisible

watermarking. Visible watermarking is a method of embedding visible transparent

image such as company name, copyright, website address, logo or text which is overlaid

on the cover image (Winwatermark n.d.). The process allows the watermark to be

viewed, but still marks it clearly as the property of the owning organization with the

motive of copyrights authentication purpose. Visible watermarks discourage the

unauthorized copying but it still can be removed or altered (Hu, Kwong & Huang 2004)

by intruders. In contrast, invisible watermark refers to imperceptibly embedding

watermark information into a cover image. This is the preferred method among the

researches since it conceal the presence of watermark from naked eye.

12

2.6 Watermark detection techniques

Watermark detection mainly categorized into blind and non-blind techniques.

Blind techniques is employed with watermarked image for watermark detection and do

not require the original image (Jun, Chi & Zhuang 2007; Wang, Hou & Yang 2009;

Dorairangaswamy 2009), on the other hand, non-blind techniques require the original

image (Liu et al. 2005; Nasir, Weng & Jiang 2007; El-Taweel et al. 2005).

2.7 Watermarking Applications

Watermarking has been widely used in various commercial fields; the following

list explains few specific areas where it is being applied (Wang, Yang & Cui 2008; Cox

et al. 2008; Woo 2007).

a) Copyright Protection

� Watermark helps the copyright owners to verify the illegal copies of their

works by embedding the watermark into their digital works. Later, the

successful detection of the watermark can be use to authenticate the

original owner. Besides, any unauthorized removal of the embedded

watermark will degrade the image imperceptibility.

b) Fingerprinting

� A hidden serial number is embedded within the digital material

purchased by a customer, which discourages them from redistributing the

content. It enables the intellectual property owner to identify which

customer broke his license agreement.

13

c) Copy Control

� Copyright owners can control the terms of use of their work with

watermarking, either as copy once, copy many or no copying at all.

d) Broadcast Monitoring

� Broadcast channels such as televisions and radios are monitored through

active monitoring techniques to check, when and whether the content is

transmitted, to verify advertising broadcasts and verify royalty payments,

and catching instances of piracy.

e) Data Authentication

� Watermark is used to detect any unauthorized modification applied on a

cover work. For example, checking for fraud passport photographs.

2.8 Robust Watermarking Techniques

Recently, there are many techniques for robust watermarking which have been

researched and developed to combat the content piracy issues. The following sections

discussed few methods which have been surveyed throughout the dissertation

preparation work. The approaches are mainly categorized under spatial and wavelet

domain.

2.8.1 Spatial domain

Tirkel, Van Schyndel and Osborne (1994), proposed a LSB approach, which

works by converting the corresponding pixels in a cover image into binary bits and

replacing the least significant bits with watermarking bits. The changes made onto least

significant bits do not degrade the visual effects and watermark detection tends to be

14

simple with this approach. However, as a drawback, watermark information can be

detected easily.

Nasir, Weng and Jiang (2007), proposed an algorithm based on embedding

binary image watermark which is permuted using secret code and Gray code. Later this

permuted watermark is encoded four times in different locations on blue components of

the colored cover image. The advantages of this scheme are; robust against various

image processing operations, and secure technique, since watermark extraction can be

done only with a correct key. The drawback is original image need to be presented

during watermark extraction.

Both multi-resolution and spatial domain used for embedding the watermark in

(Chemak, Bouhlel & Lapayre 2007). The embedded watermark is encoded using Error

Correcting Code (ECC) with turbo code. This algorithm is an efficient in embedding

large watermark bits into cover image and contributes good perceptual fidelity in image

after watermarking process.

2.8.2 Transform domain

An original image is transformed into wavelet domain and watermarks are

embedded in the difference value of the original image and referenced image (Liu et al.

2005), to overcome the watermark problem in spatial domain. This approach survived

various image processing attacks but it seems to be time consuming to do the

embedding process. In addition, original image need to be presented in the

watermarking extraction process. Meanwhile, Jun, Chi and Zhuang (2007), attempted a

randomly encrypt watermark before embedding it into middle sub-band of original

image in wavelet domain. Noise Visibility Function (NVF) has been adopted in

15

identifying the strength of watermark, and Independent Component Analysis (ICA)

algorithm used in watermark extraction whereby it eliminates the need of information

about original image or original watermarks. This approach is robust against Joint

Photographic Experts Group (JPEG) compression, additive noise, and filtering. In

contrast, this approach requires higher computational time, since an additional algorithm

is required in randomly encrypting watermark by scrambling for watermark embedding

process and furthermore, the watermark need to be descrambled and de extended for

watermark extraction purpose.

Another way of enhancing the robustness of a watermark is by decomposing

original image into low-frequency and mid-frequency bands in different resolution

levels by DWT. In (Luo & Tian 2008), the same watermark signals will be embedded

into both frequencies bands. Therefore, they are robust against image distortions and

noise adding, these techniques are vulnerable to some kind of attacks. Meanwhile,

Wang, Chang and Pan (2006), presented a DWT-based robust watermarking scheme

with Fuzzy Adaptive Resonance Theory (Fuzzy-ART). The sensitivity of Fuzzy-ART to

noise and outliers has been resolved by the robustness of the low-low frequency sub-

image. The strengths of this scheme are robust against common image processing and

geometric attacks, blind watermarking and lossless scheme. Bhatnagar and Raman

(2009) make use of gray scale logo image instead of randomly generated Gaussian noise

type watermark. This is done by transforming the original image into wavelet domain

and a reference sub-image is formed using directive contrast and wavelet coefficients.

Later the watermark embedded into reference image by modifying the singular values of

reference image using the singular values of the watermark. This scheme can resists the

ambiguity attack such as removal of watermark.

16

Agarwal and Goyal (2007), proposed an idea of embedding watermark codes

either in Discrete Hartley Transform (DHT) domain or in DCT domain based on the

number of edges in the blocks of the original image. This idea works by embedding the

watermark as block by block in different blocks of an original image depending on the

number of edges found in a given block in the image. Furthermore, the threshold

number of edges in the original image acts as a key in this algorithm and is used in the

watermark embedding and extraction process. This technique is robust against the

various attacks like JPEG compression, random and impulse noises, and cropping.

Robust watermarking based on invariant domain technique with blind

watermarking was introduced by Woo, Du and Pham (2006), eliminate the need of

image resynchronization. Enhanced with Human Visual System (HVS) masking

property of Dual Tree-Complex Wavelet Transform (DT-CWT) subbands, this scheme

defeat several image attacks includes RST operations, JPEG compression, and local

geometrical distortions.

Watermarking scheme based on Singular Value Decomposition (SVD) utilized

in (Lin et al. 2008; Bhatnagar & Raman 2007). In (Lin et al. 2008), SVD and

Distributed Discrete Wavelet Transformation (DDWT) are employed for watermark

embedding purpose. Therefore, this proposed scheme is robust and able to resists

geometric and signal processing attacks. Bhatnagar and Raman (2007), tried with

decomposing and transforming a cover image into four frequencies sub-bands using

Distributed Multi-Resolution Discrete Cosine Transform (D-MR-DCT) and applied

with SVD. Later the singular values of every sub-band are altered with singular values

of the watermark. Experiments proved in this method to be simple, lower complexity

and computational cost. In addition it is tend to be robust than Discrete Cosine

17

Transform Singular Value Decomposition (DCT-SVD) and Discrete Wavelet Transform

Singular Value Decomposition (DWT-SVD) methods.

Fu, Shen and Shen (2005), implemented a robust watermarking through

embedding watermarks in DFT domain in 4 subsampled images. Experiments shows

that, this technique is robust against some geometric and image processing operations.

A significant watermark embedding process is done into perceptually significant

wavelet coefficients using pixel wise masking (Mankar 2008). Robustness is achieved

by encoding the watermark for several times into the detail sub bands. It is

imperceptibly prominent to the HVS.

2.8.3 Scale normalization and flowline curvature

Woo, Du and Pham (2005), proposed a geometrically robust watermarking

based on scale normalization and flowline curvature. Watermark synchronization has

been adopted in an effective manner whereby the scale, translation invariance are

accomplished through scale normalization and meanwhile the rotation invariants

presented with selected feature points in flowline curvature calculations. It has been

experimented by selecting only two corners in recovering an image that underwent RST

transformation and has been verified to be robust against these RST operations.

Meanwhile the searching process becomes shorter since user needs to work with only

four robust corners in image recovery operations and it contributes a lower

computational complexity but tend to be weak against local transformations. Besides,

the original image needs to be presented in the process of watermark detection.

18

2.8.4 Features point

In (Tang & Wang 2008), Harris-Laplace, detector is used in detecting the feature

based corners and further enhanced. Annular region is extracted from this corners

followed by DCT and watermark embedding. An experiment shows this proposed

algorithm robust against several image operations and geometric distortions. A blind

image watermarking algorithm presented in (Wang, Hou & Yang 2009) by extracting

the feature point using a multi-scale Harris-Laplace detector. The resilience of the

watermarks against attacks enhanced by applying pseudo-Zernike moments. This

watermarking scheme able to resist geometric attacks and common image processing

operations. However, this scheme only allows lesser watermark bits to be encoded and

more time consuming in extracting feature point. Wang, Hou and Wu (2008), proposed

a scheme based on scale space theory, with the combination of image feature extraction

and image normalization. Feature points extraction is performed using Harris-Laplace

detector. This scheme tends to be invisible watermarking and able to resist common

signals processing and general geometric attacks.

2.8.5 Support Vector Machine

In (Wang, Yang & Cui 2008), a robust watermarking scheme is establish by

applying classification techniques based on Support Vector Machine (SVM). Besides,

the capacity of watermark embedding is improved greatly without adding extra

template. Proposed scheme proved to be robust against signals processing and

desynchronization attacks such as rotation, scaling, translation, row or column removal,

and local random bend.

19

2.9 Lightweight watermarking techniques

2.9.1 Buyer-seller watermarking protocol

Wu and Pang (2008), proposed a novel buyer-seller watermarking by generating

two independent watermarks W and W’, where W based on fingerprint of digital

contents by seller and W’ based on description by buyer. Buyer generates fingerprint

and matches with fingerprint supplied by seller for verification. This protocol is

efficient in terms of computation cost and communication overhead. Due to this, it is

feasible to be implemented for online application such as with mobile devices.

2.9.2 Optimal Differential Energy Watermarking of DCT Enc oded Images and

Video

In (Langelaar & Lagendijk 2001), a watermarking algorithm is proposed for

real-time watermarking of JPEG or Moving Picture Experts Group (MPEG) streams

based on DCT blocks. This technique avoids the need for decoding JPEG or MPEG

encoded information. Therefore, it is a lightweight watermarking process suitable for

consumer products.

2.9.3 MMS Content Copyright Protection using Watermarking

A Digital Rights Management (DRM) solution for Mobile Multimedia Service

(MMS) messages that uses a centralized approach and watermarking technology are

discussed in (Silva et al. 2003). The advantages of this scheme are able to detect and

avoid unauthorized dissemination of copyrighted content. Moreover this scheme is

robust to scaling attack, efficient in computational cost and provides full interoperability

between mobile phones.

20

2.10 Chapter Summary

A detail description of watermarking algorithm, properties, domain, embedding

classification, detection techniques and its applications is identified. In addition, several

existing literatures of robust and lightweight watermarking for digital images is studied

and reviewed carefully.

Digital image watermarking plays an important role in image authentication and

more prevalent compare to other techniques. Robustness is the crucial part in image

watermarking whereby the watermark resists the common image processing and

geometrical attacks. Therefore, many researches and ideas have been established

towards the field of robust watermarking.

We observed several weaknesses in some of the existing robust watermarking

schemes which have discussed in above sections. Those weaknesses are as the

following. In scheme (Woo, Du & Pham 2005; Liu et al. 2005; Nasir, Weng & Jiang

2007), original cover image need to be presented during watermarking extraction.

Watermark embedding consume much CPU time (Liu et al. 2005; Wang, Hou & Yang

2009). Lesser embedded watermark volume in cover image (Wang, Hou & Yang 2009)

and requirement of additional encryption technique for watermark (Jun, Chi & Zhuang

2007). Besides, we have also identified several strengths from the above discussions.

First, embedding watermark along edge is able to enhance the imperceptibility of

watermark (Agarwal & Goyal 2007). Second, a robust watermarking scheme is able to

defeat geometrical distortions such as scaling attack (Woo, Du & Pham 2006). Third, a

lightweight watermarking with efficient computational cost is suitable for mobile

platform.

21

Based on the above mentioned strengths and weaknesses, we decided to

develop a lightweight robust watermarking for digital image using robust edge features.

The robustness of our proposed scheme is improved by simple and fast watermarking

algorithm. This scheme is experimented separately in both spatial and wavelet domains.

Further explanation on our proposed scheme is included in Chapter 3.

22

Chapter 3: Methodology

3.1 Introduction

There are many robust watermarking schemes for digital images have been

researched and developed recently. Each of these approaches has their unique methods

and approaches to resist the geometrical and image processing attacks. In this chapter,

the general watermarking process is described briefly and followed with the detailed

discussion on the proposed scheme.

3.2 Generic Watermarking Process

Essentially, watermarking approach consists of embedding, and detection

phases. The generic idea has been illustrated in Figure 3.1 (Cox et al. 2008).

Embedder

Algo

Detector

Decision

(Yes/ No)

Watermark,

W

Start

Cover

image, I

Watermarked

image, I’

Detected

watermark,

W’

Stop

Figure 3.1 Generic watermarking process

23

As seen in Figure 3.1, an image is selected to be as the cover image, I, which

needs to be embedded with watermark. Watermark, W may consist of small image or

generated from pseudo random series consisting 1, and 0 bits. An embedder algorithm is

applied during embedding process; which will specifically select the regions to embed

the W in cover image. Watermarked image, I’ , is derived after the completion of

embedding process.

During watermark detection stage, a detector algorithm will work on image I’ , to

detect the watermark which has been embedded earlier. The detected watermark is

identified as W’. W and W’ is compared to verify the similarity and existence of

watermark in I’ .

As mentioned earlier in Section 2.9, we have designed and developed a new

robust watermarking algorithm which is simple and fast. Section 3.3 explains the

proposed scheme in detail.

3.3 Proposed Watermarking Scheme

An essential principle of watermarking is to embed the watermark into

significant regions of a cover image whereby it can withstand the image processing

attacks. Taking this into consideration, we employed a similar approach to conventional

LSB (Tirkel, Van Schyndel & Osborne 1994) image watermarking, as known as

lightweight robust watermarking using edge features in multiple domains (LiREF).

LiREF algorithm is built with edge detection technique in developing the watermarking

system. This idea is fine tuned for gray scale digital images, it can be extended for

colour image watermarking.

24

LiREF watermarking embed watermark into robust edge features of the cover

image, which is less imperceptible to human eyes. On the other hand, an inverse process

of embedding step is required in the watermark detection stage. We implemented this

approach both in spatial and wavelet domains of the cover image. Simple and fast

algorithm is developed using MatLab software to enhance the robustness of LiREF

scheme. In addition, to test the robustness of this scheme against image attacks, the

watermarked image is purposely distorted with scaling attack.

In the next section, we briefly describe the edge detection and scaling attack

which is implemented in the proposed scheme. In addition, we also provide some brief

explanation about MatLab software. Further, we include the detailed discussions about

proposed watermarking implementation steps.

3.3.1 Edge Detection

Edges of an image mainly consist of meaningful and significant information.

Edge detection is a process of locating an edge of an image (Vincent & Folorunso 2009)

which is more meaningful and easier to analyze (Wikipedia, The Free Encyclopedia

2010a). In the proposed scheme, robust edges of a cover image are essential for

watermark embedding and detection process. Moreover, edge information is considered

to be robust reference points which can withstand the scaling attack in a watermarked

image. The robust edges are identified using edge detection operators such as Canny,

Roberts, Prewitt and Sobel and thresholding technique (Neoh & Hazanchuk 2004;

Senthil & Bhaskaran 2008). These two techniques are discussed as the following.

25

a) Edge Detection Operators

Roberts, Prewitt and Sobel are classified as gradient-based edge detectors which

represent edges by determining the level of variance between different pixels

(Neoh & Hazanchuk 2004). On the other hand, Canny edge detector works by

determining the optimum edges by minimizing error rate, well localized and

only one response to a single edge (Senthil & Bhaskaran 2008). Hence it is

known as optimal detector. In the proposed scheme, we tested all four edge

detectors separately in spatial domain in order to determine the robust edges in

cover image and also to identify the best edge detection operator for a particular

cover image.

b) Thresholding

Threshold techniques are used to detect edges in an image by making decisions

based on local pixel information (Teller 1996), it is classified as simple and an

efficient methods (Huang & Bai 2008). In this proposed scheme, a minimum

threshold value based on transform coefficients is chosen for watermark

embedding at level-one and level-two in the wavelet domain. We manually set a

group of threshold consisting 10, 50, 70, 90, and 130 coefficient or pixel values.

All these threshold values are tested with different cover image at level-one and

level-two separately. The image coefficients which are lower than the threshold

is being filtered out and watermark embedded in remaining regions. The best

threshold value is identified by looking at the watermark embedding and

detection performance. More robust edges are revealed and more watermark bits

are able to be embedded through this technique.

26

c) Scaling Attack

Scaling attack is a kind of geometrical attack which works by scaling up or

down the watermarked image. Scaling up enlarge image size, meanwhile in

scaling down, image is being shrunk. In this scheme, we scaled down cover

image into 256 x 256 pixels dimensions for watermark embedding. Further, it is

restored as it original dimensions. For watermark detection, the scaled up image

is resize back into 256 x 256 pixels dimensions. We employed imresize function

in MatLab to perform scaling process.

d) MatLab Software

MatLab is a fourth-generation programming language integrated with interactive

and numerical computing environment (Wikipedia, The Free Encyclopedia

2010b; The MathWorks 2009b). It is developed by The MathWorks, Inc.

MatLab allows matrix-based manipulation for scientific and engineering use,

plotting of functions and data, implementation of algorithms, creation of user

interfaces, and interfacing with programs in other languages (Wikipedia, The

Free Encyclopedia 2010b).

3.3.2 Spatial Domain Watermarking

At first, our proposed LiREF watermarking is tested for cover image in the

spatial domain as this domain is far less complex as no transform is required (Hameed,

Mumtaz & Gilani 2006). Embedding and detection strategies are clearly described as

the following.

27

a) Watermark Embedding

As seen in Figure 3.2, the first step in embedding process is to read in the cover

image which needs to be watermarked and it is identified as, I. Image I is being

normalized, which means it is fixed into standard size by transforming into 256 x 256

pixels size.

Start

Cover

image, I

Normalisation

of image size

into 256 x 256

pixels

Edge detection

Read pixel along

the edge

Embed bit-8

watermark at

bit-8 location of

pixel

Restore original

image

dimension

Watermarked

image, I’

Stop

Count number

of embedded

bits

Embedded

bits, M

Watermark,

W

End of edge

Yes

No

Got to next edge

Figure 3.2 Watermark embedding steps in spatial domain

28

Edge detection operator such as Canny, Roberts, Prewitt and Sobel are applied

separately on I, in detecting the robust edges of the image. Once the robust edges are

detected, the pixel values along these edges being accessed and converted to binary

values. In addition, the watermark is also converted to binary values. Later, the bit-8 of

watermark is embedded into bit-8 location of the pixel value. The design approach for

generating watermark bits is using pseudo random sequence as series 1, and 0 bits. It is

fixed as 128 bits of length. The embedding process starts from top left edge of the most

robust edge detected in earlier step and move on to right direction of the robust edge.

On the other hand, number of embedded bits is being total up and identified as M.

Embedding process is terminated when all the robust edges in I has been watermarked.

Finally, image I will be restored into original dimensions. The scaling attack

takes place here. The output of the embedding process is the watermarked image, I’ .

b) Watermark Detection

The first step in this detection stage works similar as the normalization step in

embedding stage, refer Figure 3.3. Watermarked image, I’ is being normalized into

standard size by transforming into 256 x 256 pixels size. The same edge detection

operator which is used in embedding stage earlier, is used in detecting the robust edges

of the normalized image, I’ . Then, the pixel values along these robust edges are

extracted and converted to binary values. In addition, the watermark is also converted to

binary values. The bit-8 of each extracted pixel value is compared with the bit-8

watermark, W. Number of match is counted and defined as detected watermark bits, E.

This process works for all the robust edges in I’ .

29

Percentage match, p between embedded watermark and detected watermark is

calculated as given by the equation (1):

100*/ EMp = ………………… (1)

where M is total number embedded watermark bits and E is total number of

detected watermark bits.

Start

Watermarked

image, I’

Normalisation

of image size

into 256 x 256

pixels

Edge detection

Read pixel along

the edge

Compare bit-8

pixel value with

bit-8 watermark

Watermark,

W

Detcted

watermark

bits, E

Calculate

percentage

match, p

Stop

Embedded

bits, M

Detected

watermark

bits, E

Percentage

match, p

End of edge

Go to next edge

Yes

No

Count number

of detected bits

Figure 3.3 Watermark detection steps in spatial domain

30

3.3.3 Wavelet Domain Watermarking

Watermarking in wavelet domain has higher robustness compare to spatial

domain (Jun, Chi & Zhuang 2007; Hameed, Mumtaz & Gilani 2006). This is achieved

through the processes called as wavelet transform and inverse wavelet transform.

Wavelet transform decompose a cover image from its spatial domain into different sub

band frequencies of transform domain. Embedding of watermark into sub band

coefficients reveals higher imperceptibility to HVS. On the other hand, inverse wavelet

transform constructs the embedded watermark to be randomly disseminated throughout

the image. Hence, the watermark information becomes inaccessible and more robust

against any unauthorized alterations.

In this proposed approach, we employed LiREF watermarking into a cover

image at level-one and level-two of its wavelet domain using Two-Dimensional

Discrete Wavelet Transformation (2-D DWT) and Inverse Discrete Wavelet Transform

(IDWT). Brief introduction to 2-D DWT and IDWT are included in following sections.

3.3.3.1. Discrete Wavelet Transform

One-Dimensional Discrete Wavelet Transformation (1-D DWT) decomposes all

the rows signals in a cover image into high frequency and low frequency sub bands.

Meanwhile, by repeating 1-D DWT into column signals of same cover image, four

frequencies coefficients sets are created. The four frequencies coefficients denoted as

one low pass sub band, approximation (A1), and three high pass sub bands, which are

horizontal (H1), vertical (V1) and diagonal (D1) as shown in Figure 3.4. Hence, this is

known as level-one 2-D DWT.

31

Figure 3.4 Level-one 2-D DWT decomposed image

A1 coefficients consist of the most information of the cover image, while the

other three coefficients contain the edge and textures components (Jun, Chi & Zhuang

2007; Jin & Peng 2006). Therefore, embedding watermark into A1 coefficients are very

sensitive to the HVS (Reddy, Prasad & Rao 2009). Due to this, we proposed the

watermark to be embedded into three high pass sub band coefficients, to improve the

robustness of watermarking. All four coefficients are reconstructed back into original

image using IDWT; this is inverse process of DWT.

In order to perform level-two 2-D DWT, sub band A1 in level-one 2-D DWT is

further decomposed into next level, consisting of one low pass sub band coefficients,

and three high pass sub bands coefficients, denoted as approximation (A2), horizontal

(H2), vertical (V2) and diagonal (D2). Figure 3.5 illustrates the level-two 2-D DWT

decomposition. As watermark embedding in A2 sub band may degrade the image,

alternatively the watermark is embedded in H2, V2 and D2 sub bands.

Figure 3.5 Level-two 2-D DWT decomposed image

32

In proposed scheme, the DWT function in MatLab can perform level-one and

level-two 2-D DWT with a Daubechies wavelet decomposition filter (The MathWorks

2009a). The following sections discussed LiREF watermarking algorithm for wavelet

domain in further detail.

3.3.3.2. Level-One 2-Dimensional Discrete Wavelet Transform


This approach works by embedding watermark bits in cover image in its wavelet

domain at level-one. During the embedding process as seen in Figure 3.6, the cover

image is being normalized its size into 256 x 256 pixels. Normalized image is further

decomposed into level-one DWT using the Daubechies filter. This process will

constructs four sub bands coefficients denoted as A1, H1, V1 and D1. The watermark

embedding is done at H1, V1 and D1 sub band, excluding the A1, since embedding at

this region will degrade the image quality (Jun, Chi & Zhuang 2007). Besides, HVS is

sensitive to the changes in low frequency regions compare to high sub bands (Jin &

Peng 2006).

Coefficients at H1, V1, D1 sub bands are converted from real number values

into 8-bits unsigned integer format. Next, robust edges in H1, V1 and D1 sub bands are

determined using thresholding technique. This task is done as described in Section

3.3.1(b). Coefficients which are above threshold value, T are selected and converted to

binary values. In addition, the watermark is also converted to binary values. The bit-8 of

selected coefficients in H1, V1, and D1 sub bands are substituted with the bit-8 of

watermark. The design approach for generating watermark bits is using pseudo random

sequence as series 1, and 0 bits and it is fixed as 128 bits of length. The substitution

33

process repeats until all the selected coefficients are substituted with watermark. Once

the substitution is completed, all coefficients at H1, V1, D1 sub bands are converted

back into real number format. All four sub bands (A1, H1, V1 and D1) are inversed

transform using IDWT function to retrieve the spatial domain of the image. The image

is normalized into its original size. Finally the watermarked image, I’ has been created.

34

Figure 3.6 Watermarks embedding steps in Level-one 2-D DWT

Start

Cover image, I

Normalisation of image size into 256 x 256

pixels

Level-one 2-D DWT

Normalisation coefficients

into unsigned 8-bit integer





Pick a threshold value, T



Embed bit-8 watermark into

bit-8 of coefficient Watermark,

W


into real number

IDWT

Restore original image

dimension

Watermarked image, I’

Stop

Low pass sub band,

A1

High pass sub band,

H1

High pass sub band,

V1

High pass sub band,

D1

Count number of embedded

bits

Embedded bits in H1,

M

Embedded bits in V1,

M

Embedded bits D1, M

coefficient > T

Yes

Go to next coefficient

No

Yes

Embed bit-8 watermark into bit-

8 of coefficient


into real number


bits

coefficient > T

Yes


No

Yes

Embed bit-8 watermark into bit-

8 of coefficient


into real number


bits

coefficient > T

Yes


No

Yes

End of coefficient

End of coefficient

End of coefficient

35


The watermarked image, I’ is accessed for the detection stage, refer Figure 3.7.

Besides, the watermark, W and threshold value, T which has been used in embedding

stage is needed for detection purpose. Image I’ is normalized from its original size into

256 x 256 pixels. Next, it is decomposed into level-one 2-D DWT constructing four

frequencies sub bands; A1, H1, V1, and D1.

All coefficients at the high pass sub bands are normalized from real number

values into 8-bits unsigned integer format and those coefficients which are same value

as or above threshold, T are selected and converted to binary values. In addition,

watermark is also converted to binary values. During the detection process, the bit-8 of

each selected coefficient in H1, V1, and D1 sub bands are being read and compared

with the bit-8 watermark, W. Then, the number of match is counted and defined as

detected watermark bits, E. This process works for all the selected coefficients at high

pass sub bands. Finally, a percentage match, p for H1, V1 and D1 sub bands is

calculated based on equation (1), in Section 3.3.2(b).

36

Figure 3.7 Watermarks detections steps in Level-one 2-D DWT

37

3.3.3.3. Level-Two 2-Dimensional Discrete Wavelet Transform


Cover image, I is accessed and normalized into 256 x 256 pixels size, refer to

Figure 3.8. Normalized image is further decomposed into level-one 2-D DWT. This is

followed by level-two decomposition at A1 sub band coefficients, which will constructs

the A2, H2, V2, and D2 sub bands coefficients.

Coefficients at H2, V2, D2 sub bands are converted from real number values

into 8-bits unsigned integer format. Next, robust edges in H2, V2 and D2 sub bands are

determined using thresholding technique. This task is done as described in Section

3.3.1(b). Coefficients which are above threshold value, T are selected and converted to

binary values. In addition, the watermark is also converted to binary values. The bit-8 of

selected coefficients in H2, V2, and D2 sub bands will be substituted with the bit-8 of

watermark. The design approach for generating watermark bits is using pseudo random

sequence as series 1, and 0 bits. It is fixed as 128 bits of length. The substitution process

repeats until all the selected coefficients are substituted with watermark.

Once the substitution is completed, all coefficients at H2, V2, D2 sub bands are

converted back into real number format. Then, A2, H2, V2 and D2 sub bands are

inversed transformed into level-two wavelet domain by employing the IDWT function.

Output of this process is the A1 sub band. A1 sub band together with H1, V1, D1 sub

bands are inversed transform again, in order to retrieve the spatial domain of the image.

Further, the image is normalized into its original size. Finally the watermarked image

has been created denoted as I’ .

38

Start

Normalisation

of image size

into 256 x256

pixels

Level-one 2-D

DWT

Normalisation

coefficients

into unsigned

8-bit integer

Restore

original image

dimension

Watermarked

image, I’

Stop

Low pass

sub band,

A1

High pass

sub band,

H1

High pass

sub band,

V1

High pass

sub band,

D1

Low pass

sub band,

A2

High pass

sub band,

H2

High pass

sub band,

V2

High pass

sub band,

D2

Normalisation

coefficients

into unsigned

8-bit integer

Normalisation

coefficients

into unsigned

8-bit integer

Cover

image, I

Level-two 2-D

DWT

Level-one 2-D

IDWT

Low pass

sub band,

A1

Pick a

threshold

value, T

Pick a

threshold

value, T

Pick a

threshold

value, T

Embed bit-8

watermark into

bit-8 of coefficient

Watermark,

W

Normalisation

coefficients

into real

number

Level-two 2-D

IDWT

Count number

of embedded

bits

Embedded

bits in H2,

M

Embedded

bits in V2,

M

Embedded

bits D2, M

coefficient > T

Yes


No

Yes

Embed bit-8

watermark into bit-

8 of coefficient

Normalisation

coefficients

into real

number

Count number

of embedded

bits

coefficient > T

Yes


No

Yes

Embed bit-8

watermark into bit-

8 of coefficient

Normalisation

coefficients

into real

number

Count number

of embedded

bits

coefficient > T

Yes


No

Yes

End of

coefficient

End of

coefficient

End of

coefficient

Figure 3.8 Watermarks embedding steps in Level-two 2-D DWT

39


I’ is accessed and normalized into 256 x 256 pixels, refer Figure 3.9. Normalized

image is decomposed into level-one 2-D DWT and followed by level-two

decomposition at A1 sub band coefficients, which creates the A2, H2, V2, D2 sub bands

coefficients.

All coefficients at H2, V2 and D2 sub bands are normalized from real number

values into 8-bits unsigned integer format and those coefficients which are same value

as or above the threshold, T are selected and converted to binary values. In addition,

watermark is also converted to binary values. During the detection process, the bit-8 of

each selected coefficient in H2, V2, and D2 sub bands are being read and compared

with the bit-8 watermark, W. Next, the number of match is counted and defined as

detected watermark bits, E. This process works for all the selected coefficients at high

pass sub bands (H2, V2 and D2). Finally, a percentage match, p for H2, V2 and D2 sub

bands is calculated based on equation (1), in Section 3.3.2(b).

40

Start

Normalisation

of image size

into 256 x 256

pixels

Level-one 2-D

DWT

Normalisation

coefficients

into unsigned

8-bit integer

Low pass

sub band,

A1

High pass

sub band,

H1

High pass

sub band,

V1

High pass

sub band,

D1

Low pass

sub band,

A2

High pass

sub band,

H2

High pass

sub band,

V2

High pass

sub band,

D2

Normalisation

coefficients

into unsigned

8-bit integer

Normalisation

coefficients

into unsigned

8-bit integer

Watermarked

image, I’

Level-two 2-D

DWT

Compare bit-8

coefficient with

bit-8 watermark

Watermark,

W

Stop

Threshold

value, T

Detected

watermark

bits, E

Detected

watermark

bits, E

Yes

Calculate

percentage match,

p

Calculate

percentage match,

p

Calculate

percentage match,

p

Percentage

match, p


Percentage

match, p

Percentage

match, p

Embedded

bits in H1,

M

Embedded

bits in V1,

M

Embedded

bits in D1,

M

coefficients >= T

End of

coefficient

Yes

No

Count number of

detected bits

Detected

watermark

bits, E

Compare bit-8

coefficient with

bit-8 watermark

Yes


coefficients >= T

End of

coefficient

Yes

No

Count number of

detected bits

Compare bit-8

coefficient with

bit-8 watermark

Yes


coefficients >= T

End of

coefficient

Yes

No

Count number of

detected bits

Figure 3.9 Watermarks detection steps in Level-two 2-D DWT

41

3.4 Chapter Summary

A detailed description of watermarking generic process, and proposed LiREF

watermarking for spatial and wavelet domain is explained.

A lightweight robust watermarking based on robust edge features to withstand

the geometric attacks is designed for both spatial and wavelet domains of a cover image.

The proposed scheme is basically derived from the LSB approach with the combination

of edge detection technique. Robust edges become the essential features in identifying

the best location for watermark embedding and more robust to image attacks. Moreover,

this proposed algorithm appears to be simple and has low computational complexity,

which means it is reliable to be implemented on the platform of mobile device.

42

Chapter 4: Experiments and Results

4.1 Introduction

The proposed LiREF watermarking algorithm is experimented and results are

presented through a careful analysis and observations. In order to accomplish this, the

proposed algorithm is tested based on the strategies presented in this dissertation.

Primarily, the algorithm is tested separately in spatial and wavelet domain of TIFF kind

of gray scale images.

4.2 Experiment Setup

a) Hardware Specifications

The following are minimum hardware requirements for the experiment:

i) Processor : Intel Pentium-IV compatible PC

ii) Memory : 1 GB of RAM

iii) Hard Drive : 20 GB hard disk or higher

iv) Display : LCD Monitor (1024 x 768) or higher resolution monitor

b) Software List

The following software is used for the implementation of the proposed

idea:

i) Operating System : Windows XP

ii) Development tool : MatLab R2008a

iii) Runtime environment: Windows XP and above or Windows

Server 2003.

43

c) Image data configuration

The proposed idea is developed and evaluated for five different 8-bit

gray scales cover images in TIFF (Tagged Image File Format) image format.

TIFF images do not use compression and do not degrade in quality each time the

image is edited compared to JPEG images. Due to this exceptional feature of

TIFF format, it becomes prominent image type for the proposed scheme. Each

cover image used in this experiments need to be normalized into 256 x 256

pixels for watermark embedding purpose.

d) Watermark configuration

A watermark is generated using pseudo random sequences consisting 1

or 0 binary bits with the length of 128-bits.

4.3 Interim Experiment Results

LiREF watermarking is experimented with spatial domain of a cover image

according to algorithms described in Section 3.3.2. On the other hand, watermark

algorithms described in Section 3.3.3.2 and 3.3.3.3 have been experimented with level-

one and level-two wavelet domains separately. Embedding process works specifically at

robust edges of cover image, which are identified through edge detection technique in

spatial domain, and thresholding technique for wavelet domain. Inverse embedding

process is accomplished in both domains for watermark detection.

Five gray scale images have been used for the experiment as illustrated in Figure

4.1.

44

Figure 4.1 Cover images used for experiments

To investigate the robustness of the proposed algorithm against attacks, we

purposely attacked the watermarked image by scaling attack. This attack is caused by

MatLab program.

4.3.1 LiREF watermarking in spatial domain

This approach requires edge detection to be employed on spatial domain of

cover image, so that watermark is embedded on these robust edges. Figure 4.2 shows

the sample of edge detections for “Lena” image using Canny, Roberts, Prewitt and

Sobel operators. As seen in Figure 4.2, more edges can be spotted using Canny

followed by Sobel, Prewitt and Roberts edge detector.

(a) Baboon.tiff (b) Barbara.tiff

(e) Lena.tiff

(d) Peppers.tiff (c) Cameraman.tiff

45

Figure 4.2 Edge detections for Lena image

Figure 4.3 shows cover image of “Lena”, its watermarked copy, and difference

image. Roberts edge detection operator is used in edge detection process. Very little

differences can be notice along edges in the image.

Figure 4.3 Cover image and its watermarked copy

(a) Canny (b) Roberts

(c) Sobel (d) Prewitt

(a) Cover image (b) Watermarked Image (c) Difference image

46

4.3.2 LiREF watermarking in wavelet domain

LiREF watermarking scheme is experimented on both level-one and level-two

DWT domains separately. The Daubechies wavelet is used to produce the wavelet

coefficients. The motive is to test the robust coefficients levels to embed and extract the

watermark. Watermark is only embed in the three high pass sub bands denoted as

horizontal, diagonal, and vertical coefficients; when its value above thresholding value.

Figure 4.4 illustrate level-one 2-D DWT watermarking for “Lena” image, which

shows cover image of “Lena”, its watermarked copy, and difference image. The

selected threshold value is above 70.

Figure 4.4 Level-one 2-D DWT watermarking

Figure 4.5 shows level-two 2-D DWT watermarking for “Lena” image,

consisting of cover image of “Lena”, watermarked copy, and difference image. Selected

threshold value is above 70.


47

Figure 4.5 Level-two 2-D DWT watermarking

4.4 Results Analysis

All the cover images are tested with LiREF watermarking scheme. Watermarked

image might be attacked either on purpose or accidentally, so the watermarking system

should able to detect and extract the watermark. To test the robustness of the proposed

algorithm against attacks, watermarked image is attacked by scaling distortion.

Finally, a percentage match, p between embedded watermark and detected

watermark is calculated based on equation (1), in Chapter 3, Section 3.3.2(b), for spatial

and wavelet domains.

4.4.1 LiREF watermarking in spatial domain

Table 4.1 lists results of LiREF watermarking experiments in spatial domain for

cover images after scaling attack, together with edge detection techniques, embedded,

detected watermark volume and p values.


48

Table 4.1 LiREF watermarking experiment results in spatial domain

Based on the figures in Table 4.1, a line chart has been plotted to show the

comparison of p value among different cover images with related edge detection

operators. The line chart is shown in Figure 4.6.

Attack : Scaling (Original dimension > 256 x 256 > Original dimension)

Cover image Edge

detection operator

Embedded watermark bits /

volume

Extracted watermark bits /

volume Percentage match,

p

Baboon.tiff Canny 10645 5225 49

Sobel 2808 1310 47

Prewitt 2704 1233 46

Roberts 1292 662 51

Barbara.tiff Canny 5913 2927 50

Sobel 2954 1421 48

Prewitt 2909 1483 51

Roberts 2339 1181 50

Cameraman.tiff Canny 5747 3133 55

Sobel 2503 1346 54

Prewitt 2509 1544 62

Roberts 2341 1693 72

Lena.tiff Canny 5113 2475 48

Sobel 2573 1225 48

Prewitt 2523 1221 48

Roberts 2393 1179 49

Peppers.tiff Canny 4962 2454 49

Sobel 2581 1232 48

Prewitt 2567 1241 48

Roberts 2587 1324 51

49

0

10

20

30

40

50

60

70

80

90

100

Baboon.tiff Barbara.tiff Cameraman.tiff Lena.tiff Peppers.tiff

Cover Image

Per

cen

tage

mat

chCanny

Sobel

Prewitt

Roberts

Figure 4.6 Comparison of p value among cover images in spatial domain. As in the above Figure 4.6, among the cover images, “Cameraman” has the highest p value with all edge detection operators. The rest of images have highest p with Roberts edge detector, except for “Barbara” which shows highest p value with Prewitt edge detector.

4.4.2 LiREF watermarking in wavelet domain

a) Level-one 2-D DWT

Table 4.2 lists the embedded and detected watermark volume from

watermarked image in level-one 2-D DWT and their p values. Line charts

have been plotted based on the derived p value for H1, V1, D1 sub bands in

cover images together with selected threshold value, refer to Figure 4.7 to

Figure 4.11.

50

Table 4.2 LiREF watermarking experiment results in Level-one 2-D DWT

NA = Not Applicable. Detected watermark bits and percentage match, p becomes as NA, when there is no or zero embedded bits found in H1, D1, or V1 sub bands for a particular cover image at certain threshold value.


Level-one 2-D DWT


volume Detected watermark

bits / volume Percentage match,

p

Cover image Threshold

value H1 V1 D1 H1 V1 D1 H1 V1 D1

>10 3485 3215 2018 1440 1318 341 41 41 17

>50 321 62 14 64 8 0 20 13 0

>70 110 8 0 30 1 NA 27 13 NA

>90 41 4 0 0 0 NA 0 0 NA

Baboon.tiff

>130 0 0 0 NA NA NA NA NA NA

>10 1130 2188 984 510 960 62 45 44 6

>50 56 64 0 6 9 NA 10 14 NA

>70 14 8 0 0 0 NA 0 0 NA


Barbara.tiff


>10 1509 1404 856 848 757 484 56 54 57

>50 186 305 38 94 195 38 51 64 100

>70 92 182 8 55 159 3 60 87 38

>90 46 102 2 32 52 2 70 51 100

Cameraman.tiff

>130 6 26 0 4 21 NA 67 81 NA

>10 990 1719 517 454 852 125 46 50 24

>50 47 155 0 18 43 NA 38 28 NA

>70 7 62 0 3 26 NA 43 42 NA

>90 1 23 0 0 7 NA 0 30 NA

Lena.tiff

>130 0 3 0 NA 0 NA NA 0 NA

>10 1108 1373 353 598 684 72 54 50 20

>50 66 180 1 28 90 0 42 50 0

>70 21 89 0 6 23 NA 29 26 NA

>90 8 33 0 1 8 NA 13 24 NA

Peppers.tiff

>130 0 1 0 NA 0 NA NA 0 NA

51

0

10

20

30

40

50

60

70

80

90

100

Baboon.tiff Barbara.tiff Cameraman.tiff Lena.tiff Peppers.tiffCover Image

Pe

rce

nta

ge

ma

tch

H1

V1

D1

Figure 4.7 p values for H1, V1, D1 sub bands of cover images; Threshold > 10. As in the above Figure 4.7, D1 sub band reveals lower p value for all cover images, except for “Cameraman”. H1 sub band shows highest p value for all cover images, except for “Lena”.

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

ch

H1

V1

D1

NA(D1)

Figure 4.8 p values for H1, V1, D1 sub bands of cover images; Threshold > 50. As in the above Figure 4.8, “Cameraman” has the highest p value at D1 sub band. In contrast, “Baboon” and “Peppers” has zero p value followed by “Barbara” and “Lena” which reveal NA results at D1 sub band.

52

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

chH1

V1

D1

NA(D1)

Figure 4.9 p values for H1, V1, D1 sub bands of cover images; Threshold > 70. As in the above Figure 4.9, “Barbara” shows zero p value for H1, V1 and NA for D1 sub bands. In further, “Baboon”, “Lena” and “Peppers” also reveal NA results for D1 sub band.

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

ch

H1

V1

D1

NA(H1)

NA(V1)

NA(D1)

Figure 4.10 p values for H1, V1, D1 sub bands of cover images; Threshold > 90. As in the above Figure 4.10, “Barbara” shows NA results for H1, V1 and D1 sub bands. Detectable watermark is found in “Cameraman” at all sub bands, “Lena” at V1 sub band, and “Peppers” at H1 and V1.

53

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

chH1

V1

D1

NA(H1)

NA(V1)

NA(D1)

Figure 4.11 p values for H1, V1, D1 sub bands of cover images; Threshold > 130. As in the above Figure 4.11, only “Cameraman” has revealed detectable watermark at H1 and V1 sub bands. “Lena” and “Peppers” have zero p value at V1 sub bands. The rest of the results are NA.

b) Level-two 2-D DWT

Table 4.3 lists the embedded and detected watermark volume from

watermarked image in level-two 2-D DWT and their p values. Line charts

have been plotted based on the derived p value for H2, V2, D2 sub bands in

cover images together with selected threshold value, refer to Figure 4.12 to

Figure 4.16.

54

Table 4.3 LIREF watermarking experiment results in Level-two 2-D DWT


Level-two 2-D DWT


volume Detected watermark

bits / volume Percentage match,

p

Image Threshold H2 V2 D2 H2 V2 D2 H2 V2 D2

>10 1319 1398 1033 684 717 493 52 51 48

>50 288 193 86 122 91 33 42 47 38

>70 117 81 27 53 37 8 45 46 30

>90 26 34 9 7 12 0 27 35 0

Baboon

>130 4 4 0 2 2 NA 50 50 NA

>10 912 1044 480 481 539 222 53 52 46

>50 116 212 11 53 106 4 46 50 36

>70 72 104 2 37 43 1 51 41 50

>90 37 51 1 15 24 0 41 47 0

Barbara

>130 10 12 0 5 2 NA 50 17 NA

>10 687 565 479 385 299 249 56 53 52

>50 192 208 83 102 120 43 53 58 52

>70 116 167 45 56 94 20 48 56 44

>90 79 138 27 41 74 20 52 54 74

Cameraman.tiff

>130 47 97 5 24 43 2 51 44 40

>10 635 811 473 348 398 240 55 49 51

>50 124 302 60 60 141 16 48 47 27

>70 72 162 20 43 74 8 60 46 40

>90 35 108 5 21 45 2 60 42 40

Lena.tiff

>130 11 52 0 4 26 NA 36 50 NA

>10 854 789 405 433 389 214 51 49 53

>50 183 273 46 93 125 22 51 46 48

>70 104 196 15 52 96 5 50 49 33

>90 61 140 3 25 67 0 41 48 0

Peppers.tiff

>130 24 74 0 10 32 NA 42 43 NA

NA = Not Applicable. Detected watermark bits and percentage match, p becomes as NA, when there is no or zero embedded bits found in H2, D2, or V2 sub bands for a particular cover image at certain threshold value.

55

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

ch

H2

V2

D2

Figure 4.12 p values for H2, V2, D2 sub bands of cover images; Threshold > 10. As in the above Figure 4.12, all cover images reveal consistent p values at varies sub bands. Meanwhile, “Barbara” has lowest p value at D2 sub band.

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

ch

H2

V2

D2

Figure 4.13 p values for H2, V2, D2 sub bands of cover images; Threshold > 50. As in the above Figure 4.13, all cover images have very much consistent p value with the range 40%-60% at H2 and V2 sub bands. There are drastic differences of p value found at D2 sub band for all cover images except for “Cameraman”.

56

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

ch

H2

V2

D2

Figure 4.14 p values for H2, V2, D2 sub bands of cover images; Threshold > 70. As in the above Figure 4.14, inconsistent p values are found for all cover images at all sub bands.

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

ch

H2

V2

D2

Figure 4.15 p values for H2, V2, D2 sub bands of cover images; Threshold > 90. As in the above Figure 4.15, “Cameraman” has shown detectable watermark at this stage with the highest p value at D2 sub band and in contrast, zero p value are obtained for “Baboon”, “Barbara” and “Peppers” at t his sub band. Inconsistent p values are found at H2 and V2 sub bands.

57

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cen

tage

mat

chH2

V2

D2

NA(D2)

Figure 4.16 p values for H2, V2, D2 sub bands of cover images; Threshold > 130 As in the above Figure 4.16, most of the cover images reveal NA results at D2 sub band except for “Cameraman”. H2 and V2 sub bands have varies performance of p value for all kind of cover images.

We compare our experiment results obtained in level-one and level-two 2-D

DWT as listed in Table 4.4 for threshold value above 70. A related column chart has

been tabulated based on figures in Table 4.4, refer to Figure 4.17.

Table 4.4 Comparison p values between Level-one and Level-two 2-D DWT Percentage Match, p

Threshold > 70

Level-one 2-D

DWT

Level-two 2-D

DWT

Image H1 V1 D1 H2 V2 D2

Baboon.tiff 27 13 NA 45 46 30

Barbara.tiff 0 0 NA 51 41 50

Cameraman.tiff 60 87 38 48 56 44

Lena.tiff 43 42 NA 60 46 40

Peppers.tiff 29 26 NA 50 49 33

NA = Not Applicable

58

Figure 4.17 Comparisons p values between Level-one and Level-two 2-D DWT. As in the above Figure 4.17, “Cameraman” reveals the highest p values in both level of 2-D DWT. This is followed by “Lena”, “Peppers”, “Baboon” and “Barbara” respectively.

4.5 Discussion

Based on the derived results from above, several ideas has been discussed as the

following. In spatial domain approach, more watermark volume are embedded in cover

images using Canny edge detection, since it able to reveal more robust edges compare

to other edge detectors (Senthil & Bhaskaran 2008), refer Table 4.1. The highest

watermark volume embedded in “Baboon” followed by “Barbara”, “Cameraman”,

“Lena”, and “Peppers”. Highest p value for overall cover images is shown with Roberts

edge detection, except for “Barbara”, where the highest p shown with Prewitt edge

detection. “Cameraman”, shows highest p since it was not affected by the scaling attack

as the rest of the images, refer to Figure 4.6. Hence, “Cameraman” is used as control

item for this experiment.

As referring to Table 4.2, level-one 2-D DWT watermarking reveals more

watermarks able to embed in lower threshold value. Seen from Figure 4.7 to 4.11, if the

H1

H1

H1

H1

H1

V1

V1

V1

V1

V1

D1 D1

D1

D1 D1

H2H2

H2

H2

H2V2

V2

V2

V2V2

D2

D2D2

D2D2

0

10

20

30

40

50

60

70

80

90

100


Cover Image

Per

cent

age

mat

ch

H1

V1D1

H2

V2D2

* NA

* * * *

59

thresholding value is increased from 10 to 130, p values will drop from the range of

40-60% to 0%. This can be seen by analyzing the H1 sub band in “Pepper”, whereby

when the threshold changes from 70 to 50, the p values changed from 29% to 42%.

However when the thresholding increase to 130, it seems only single watermark bit

embedded in V1 sub band, and no watermark embedded in H1 and D1 sub bands.

Percentage match shown as 0% for V1 sub band and Not Applicable (NA) results for

H1 and D1 sub bands respectively.

As seen in Table 4.2, highest watermark volumes are embedded in V1 sub band

for “Barbara”, “Lena” and “Peppers”. Meanwhile for “Baboon” and “Cameraman”

highest watermark embedded in H1 sub band. Theoretically, watermark embedding

volume tend to be higher for H1 and V1 sub bands, since more edge information can be

seen at these sub bands (Reddy, Prasad, & Rao 2009). Furthermore, no watermark

embedded in D1 sub band when thresholding set above 70 for all images except for

“Cameraman”, which is embedded with 8 bits of watermark.

Table 4.3 lists the highest watermark embedded volume for all cover images

with thresholding above 10 in level-two 2-D DWT. But these numbers tend to decrease

when the thresholding is increased, refer Figure 4.13 to Figure 4.16. By comparing

results in Table 4.2 and Table 4.3, it is found that less watermark volume being

embedded in level-two compare to level-one. This is caused by the less number of

coefficients located in level-two sub bands. In contrast, higher p values found in level-

two, for instance, when thresholding is set above 70, p values for “Lena” are 60%, 46%

and 40% for H2, V2, D2 sub bands, meanwhile in level-one, p values are 43%, 42%

and NA for H1, V1 and D1 respectively, refer Figure 4.17. From Table 4.3, we

observed that p still can be found in H2 and V2 sub band even with thresholding above

60

130. On the other hand, lower watermark volume being embedded and detected in D2

sub band.

Through the observation of individual features of cover images, we conclude

some facts as the following. “Baboon” consists of more edge information on its fur,

which makes more watermarks embedded along these edges. “Barbara” represents the

second high textured regions found on its veil and pants. “Cameraman” which has high

contrast region between the shirt and background is not attacked by scaling distortion as

its dimension remains same throughout the watermarking process. Due to this factor,

“Cameraman” reveals highest p value among the rest images. “Lena” has smooth

background and textures on its hat, represents second lowest watermark embedded

volume. Lower watermarks have been embedded in “Peppers” since it has very smooth

content with less edge information.

4.6 Overall Analysis

In order to further evaluate the above experiments, the experiments’ results are

assessed based on Peak To Signal To Noise Ratio (PSNR) (Jun, Chi & Zhuang 2007;

Reddy, Prasad & Rao 2009) and execution time (Jun, Chi & Zhuang 2007). PSNR is

related to imperceptibility, which means it is used to indicate the changes in the

watermarked image and cover image. The PSNR is computed according to equation (2):

[ ]∑∑−

=

−

=

−•

=

•=1

0

21

010 ),('),(

1,

255log20

N

j

M

i

jiIjiINM

MSEdBMSE

PSNR …(2)

where MSE is the Mean Square Error of cover image and watermarked image.

M is the length (pixel) of the image and N is the width (pixel). 255 is gray level range of

image, I(i,j) and I’(i,j) are gray level values at pixel(i,j) of cover image and

watermarked image respectively. Execution time for a watermark algorithm refers to

61

actual CPU cycles time used in embedding watermark in cover image and watermark

detection.

Computed execution time, PSNR and MSE values are shown in Table 4.5, Table

4.6 and Table 4.7 for spatial domain with Roberts edge detection, level-one 2-D DWT

and level-two 2-D DWT respectively.

Table 4.5 Execution time, PSNR and MSE in spatial domain

Table 4.6 Execution time, PSNR and MSE in level-one 2-D DWT

Roberts edge detection

Embedding time

Detection time

PSNR MSE

Baboon.tiff 1.19 1.05 +31.05 dB 7.1461

Barbara.tiff 1.31 1.09 +33.27 dB 5.5349

Cameraman.tiff 1.31 0.98 +68.68 dB 0.0939

Lena.tiff 1.38 1.11 +38.03 dB 3.1983

Peppers.tiff 1.34 1.17 +37.49 dB 3.4031

Level-one 2-D DWT, Threshold > 70

Embedding time

Detection time

PSNR MSE

Baboon.tiff 1.19 0.52 +30.58 dB 7.5397

Barbara.tiff 1.19 0.50 +32.56 dB 6.0085


Lena.tiff 1.27 0.56 +36.06 dB 4.0155

Peppers.tiff 1.20 0.55 +35.66 dB 4.2045

62

Table 4.7 Execution time, PSNR and MSE in level-two 2-D DWT

Higher PSNR in spatial domain is identified compare to wavelet domain, due to

differences in image distortions in both domains, as seen in Table 4.5 to 4.7.

Watermarked image in spatial domain only being distort by scaling attack but in

wavelet domain it has undergone three types of distortions; there are back and forth

image transformation between spatial and wavelet domain, back and forth

normalization of coefficients from real number into 8-bit integer format and scaling

attacks. These distortions have degraded the visual quality of watermarked image in

wavelet domain and weakens its robustness. The experiment result also shows PSNR in

level-two is lower compare to level-one due to the corruption of the cover image with

level-two image decomposition and scaling attack, refer to Figure 4.5. In contrast, level-

two is more reliable and robust compare to level-one, since level-two located at higher

level in wavelet domain and it is more generic in revealing the robust edges of an

image. This can be seen by referring to comparisons of p values among level-one and

level-two in Table 4.4. By observing results in Table 4.5 to 4.7 again, we found that

PSNR in all domains are above 30dB for scaling attacks, which means our proposed

watermarking scheme is still robust and could detect the watermark successfully with

good visual quality of watermarked image (Chemak 2008; Al-Khassaweneh 2007). In

Level-two 2-D DWT, Threshold > 70

Embedding time

Detection time

PSNR MSE

Baboon.tiff 1.31 0.45 +30.48 dB 7.6298

Barbara.tiff 1.23 0.56 +32.04 dB 6.3795


Lena.tiff 1.30 0.52 +34.66 dB 4.7135

Peppers.tiff 1.34 0.56 +34.21 dB 4.9659

63

addition, the proposed scheme is robust to scaling, even when the watermarked image is

zoomed in into its original size.

It is also observed that wavelet domain reveals faster execution time compare to

spatial, with thresholding technique. Thresholding is a fast and simple implementation

for identifying robust edges in cover image compared to slower convolution approach

used by edge detection technique in spatial domain. Level-two 2-D DWT utilized more

CPU cycles times for embedding, since the embedder need to checks for the level-one

decomposition and level-two decomposition. In term of similarity, both levels in

wavelet domain have similar detection time. Overall embedding time of the embedder is

around 1 to 2 seconds and the detection time of the detector is around 0 to 2 seconds, as

seen in Table 4.5 to 4.7. This proves that our proposed scheme has low computational

complexity.

The proposed LiREF watermarking for level-two wavelet domain is compared

with previous research works (Lin et al. 2008; Wang, Chang & Pan 2006) for

robustness to scaling attack. Table 4.8 represents the PSNR values for recovered

watermark for several embedding schemes that have been scaling attack. Cover image

used is “Lena”. It is observed that our proposed scheme outperforms the other two

schemes against scaling attack. In addition, our scheme appears to be robust with simple

algorithm without the use of Fuzzy ART or SVD. Comparison for spatial domain is

omitted, since by theoretically it is not robust to geometrical attacks (Liu et al. 2005;

Jun, Chi & Zhuang 2007).

64

Table 4.8 PSNR values of recovered watermark from embedding schemes

The robust performance of our proposed scheme lies on the fact of simple and

fast algorithm. This is because watermark data are placed on the robust detected edges

using edge detection and thresholding and do not require any additional variables and

their alteration to obtain optimum watermarked results. Moreover, the proposed scheme

reveals fast execution time in all domains. Due to simple and fast algorithm and

moderate PSNR values we conclude our proposed scheme to be robust lightweight

watermarking, whereby it is suitable to be implemented in real time environment

specifically for mobile platform.

4.7 Chapter Summary

Digital images can be easily distort with geometrical attacks and effects the

robustness features of image watermarking. In this dissertation, we developed a

watermarking scheme based on the LiREF for spatial and wavelet domain which is

robust to scaling attack. Watermark embedding strategy in spatial domain works by

extraction of robust edges in cover image using edge detection techniques and

watermark embedded along these edges. In wavelet domain, the watermark is

embedded into three detail high pass sub band of a cover image using thresholding,

which can guarantee the visual quality of the watermarked image. An inverse process of

watermark embedding for each domain is used separately for the purpose of watermark

Watermarking scheme PSNR [dB] of recovered watermark

A Robust Watermark Scheme for Copyright Protection (Lin et al. 2008)

32.632dB

A DWT-based Robust Watermarking Scheme with Fuzzy ART ( Wang, Chang & Pan 2006)

29.80 dB

The proposed approach +34.66 dB

65

detection. Both approaches have been experimented and proven to be robust to scaling

attack. This algorithm is very efficient in term of processing time.

66

Chapter 5: Conclusion

5.1 Summary of the Dissertation

Digital watermarking is a method of imperceptibly altering an original digital

content to embed a message about the content itself, which later can be used for the

purpose of copyright protection, fingerprinting, copy control, broadcast monitoring and

data authentication. Watermarking has better performance compared to cryptography,

since it is robust against varies image processing attacks. Moreover, watermarking

within digital image becomes significantly important nowadays.

In this dissertation research work, we have successfully implemented a new

lightweight robust watermarking using robust edge features in multiple domains

(LiREF). The proposed algorithm mainly focused on gray scale digital images. For

simplicity, a watermark is generated through pseudo random sequences and embedded

separately in spatial and wavelet domains along the robust edges of a cover image.

Those robust edges are distinguished using edge detection operator and thresholding

techniques for spatial and wavelet domain respectively. The embedded watermark is

successfully detected without the information of original cover image using an inverse

embedding process. Hence, this is known as blind watermark detection scheme. The

robustness of our proposed scheme is tested with resistance of watermarked image

against scaling attack.

67

5.2 Achievements

The objectives stated earlier in Chapter 1, have been achieved in order to

accomplish the dissertation goal which is to design and develop a lightweight robust

watermarking using robust edge features in multiple domains.

� A careful study and review of several literatures has been carried out

successfully in the area of robust watermarking and lightweight

watermarking for digital image. The findings are written in Chapter 2.

� A new algorithm has been analyzed and designed for lightweight robust

watermarking to improve weakness in current literatures. For watermark

embedding, we used edge detection operation with spatial and thresholding

technique with wavelet domain to discover robust edges in cover image. An

inverse process of watermark embedding for each domain is used separately

for the purpose of watermark detection. In order to test the robustness of

watermarked image against geometrical attack, we purposely distort the

watermarked image with scaling attack. The complete explanation is

included in Chapter 3.

� A new algorithm has been implemented, tested and evaluated in Chapter 4.

Through several experiments and analysis, it is proven that this algorithm is

a robust to scaling attack and very efficient in term of processing time.

Experimental results have demonstrated higher PSNR is gained in spatial

domain, due to less distortion to watermarked image. Meanwhile, wavelet domain

reveals lower computational complexity with thresholding method compared to slower

convolution approach used by edge detection technique in spatial domain. Even though,

level-two wavelet domain has lower PSNR, it shows more reliable p value compare to

68

level-one. Moreover, this proposed domain approach still shows better PSNR values of

distort watermarked image compared to (Lin et al. 2008; Wang, Chang, & Pan 2006).

Overall experiment observations, we noticed that the proposed scheme is simple

algorithm with low computational complexity and consists of moderate PSNR. These

are the three main factors contributed towards the robustness feature of our proposed

scheme. Therefore, this scheme turns out to be as robust lightweight watermarking and

it is suitable to be implemented in real time environment specifically for mobile

platform.

5.3 Future work

Our future work will be based on as the following:

� To apply spread spectrum approach in order to improve the

watermark detection rate.

� To experiment other geometrical attacks such as rotation and

translation and several image processing attacks against the

proposed idea.

� This proposed scheme is anticipated to be tested on mobile

platform and extended algorithm scheme for color images.

69

References

Agarwal, N. & Goyal, A.K., 2007, ‘Robust Watermarking in Transform Domain using

Edge Detection Technique’, Conference on Computational Intelligence and Multimedia Applications, vol. 4, pp.59-63. Retrieved January 28, 2009, from IEEEXplore database.

Al-Khassaweneh, M.A. 2007, ‘Image watermarking in the time-frequency domain’,

Doctor of Philosophy Dissertation, Michigan State University. Retrieved March 27, 2009, from Digital Dissertations (UMI) @ProQuest database.

Bhatnagar G., & Raman, B. 2007, ‘Robust Watermarking Using Distributed MR-DCT

and SVD’, Seventh International Conference on Advances in Pattern Recognition, pp. 21-24. Retrieved May 11, 2009, from IEEEXplore database.

Bhatnagar, G., & Raman, B. 2009, ‘A new robust reference watermarking scheme based

on DWT-SVD’, Computer Standards and Interfaces, vol. 31, no. 5, pp. 1002-1013. Retrieved March 27, 2009, from ScienceDirect database.

Chemak, C., Elmarzouqi, N., Bouhlel, M.S., & Lapayre, J.C. 2008, “A robust

watermarking scheme against contrasts and scaling for PocktNeuro Project”, ICGST International Journal on Graphics, Vision and Image Processing, GVIP, vol. 8, no. 1, pp. 33-40. Retrieved March 20, 2009, http://www.icgst.com/gvip/Volume8/Issue1/P1150729007.pdf

Chemak, C., Bouhlel, M.S., & Lapayre, J.C. 2007, ‘A new scheme of robust image

watermarking: the double watermarking algorithm’, Proceedings of the 2007 Summer Computer Simulation Conference, pp. 1201-1208. Retrieved March 27, 2009, from ACM (Association for Computing Machinery) Digital Library database.

Cox I.J., Miller M.L., Bloom J.A., Fridrich J. & Kalker T. 2008, Digital Watermarking

and Steganography, 2nd edn, Morgan Kaufmann Publishers. Dinghui, Z., Haixia, D., & Chao, Z. 2007, ‘Researches on Digital Image

Watermarking’, The Eighth International Conference on Electronic Measurement and Instruments. Retrieved October 8, 2009, from IEEEXplore database.

Dorairangaswamy, M.A. 2009, ‘A Novel Invisible and Blind Watermarking Scheme For

Copyright Protection of Digital Images’, International Journal of Computer

70

Science and Network Security, vol. 9, no. 4. Retrieved January 3, 2010, from http://paper.ijcsns.org/07_book/200904/20090411.pdf

El-Taweel, G.S., Onsi, H.M., Samy, M., & Darwish M.G. 2005, ‘Secure and Non-Blind Watermarking Scheme for Color Images’, ICGST International Journal on Graphics, Vision and Image Processing, vol. SI1. Retrieved January 3, 2010, from http://www.icgst.com/gvip/v4/P1150442004.pdf

Fu, Y., Shen, R. & Shen, L. 2005, ‘Robust Image Scheme Based on Subsampling’,

Third International Conference on Information Technology and Applications, vol. 2, pp. 361-365. Retrieved February 4, 2009, from IEEEXplore database.

Hameed, K., Mumtaz, A., & Gilani, S.A.M. 2006, ‘Digital Image Watermarking in the

Wavelet Transform Domain’, World Academy of Science, Engineering and Technology. Retrieved January 1, 2010, from http://www.waset.org/journals/waset/v13/v13-16.pdf

Hu, Y., Kwong, S., Huang, J. 2004, ‘Using invisible watermarks to protect visibly

watermarked images’, Proceedings of International Symposium Circuits and Systems, vol. 5, pp. V-584 - V-587. Retrieved January 3, 2010, from IEEEXplore database.

Huang, T.L., & Bai, X. 2008, ‘An Improved Algorithm for Medical Image

Segmentation’, Second International Conference on Genetic and Evolutionary Computing, pp.289-292. Retrieved Feb 23, 2010, from IEEEXplore database.

Jin, C., & Peng, J. 2006, ‘A robust wavelet-blind digital watermarking algorithm’,

Information Technology Journal, vol. 5, no. 2, pp. 358-363. Retrieved October 1, 2009, from Science Alert database.

Jun, Y., Chi, J.R. & Zhuang, X.D. 2007, ‘A New Wavelet-based Robust Watermarking

for Digital Image’, IEEE International Conference on Networking, Sensing and Control, pp.1391-1394. Retrieved February 4, 2009, from IEEEXplore database.

Langelaar, G. C., & Lagendijk, R. 2001, ‘Optimal Differential Energy Watermarking of

DCT Encoded Images and Video’, IEEE Transactions on Image Processing, vol.10, pp. 148-158. Retrieved Feb 21, 2010, from IEEEXplore database.

Lin, C.H., Liu, J.C., Shih, C.H., & Lee, Y.W. 2008, ‘A Robust Watermark Scheme for

Copyright Protection”, International Conference on Multimedia and Ubiquitous Engineering, pp.132-137. Retrieved March 27, 2009, from IEEEXplore database.

71

Lisa, R.F. 2009, In digital age, can movie piracy be stopped?, CNN.com/Technology, Retrieved May 9, 2009, from http://www.cnn.com/2009/TECH/05/01/wolverine.movie.piracy/

Liu, J.L., Lou, D.C., Chang, M.C. & Tso, H.K. 2005, ‘A robust watermarking scheme

using self-reference image’, Computer Standards & Interfaces, vol. 28, no. 3, pp. 356-367. Retrieved February 4, 2009, from ScienceDirect database.

Lu, W., Lu, H., & Chung, F. 2006, ‘Feature based watermarking using watermark

template match’, Applied Mathematics and Computations, vol. 177, no. 1, pp. 377-386. Retrieved March 27, 2009, from ScienceDirect database.

Luo, K. & Tian, X. 2008, ‘A New Robust Watermarking Scheme based on Wavelet

Transform’, Congress on Image and Signal Processing, vol. 1, pp. 312-316. Retrieved February 4, 2009, from IEEEXplore database.

Mankar, V.H, Das, T.S, Saha, S., & Sarkar, S.K. 2008, ‘Robust Image Watermarking

Under Pixel Wise Masking Framework’, First International Conference on Emerging Trends in Engineering and Technology”, pp.68-72. Retrieved March 27, 2009, from IEEEXplore database.

Nasir, I., Weng, Y. & Jiang, J. 2007, ‘A New Robust Watermarking Scheme for Color

Image in Spatial Domain’, Third International IEEE Conference on Signal-Image Technologies and Internet-Based System, pp.942-947. Retrieved May 10, 2009, from IEEEXplore database.

Neoh, H.S., & Hazanchuk, A. 2004, ‘Adaptive Edge Detection for Real-Time Video

Processing using FPGAs’, GSPX Conference, vol. 7, no. 3, pp. 2-3. Retrieved Feb 23, 2010, from http://www.altera.com/literature/cp/gspx/edge-detection.pdf

Reddy, P.R., Prasad, M.V.N.K., Rao, D.S. 2009, ‘Robust Digital Watermarking of

Images using Wavelets’, International Journal of Computer and Electrical Engineering, vol. 1, no. 2, pp. 1793-8163. Retrieved October 8, 2009, from IJCEE database.

ScienceDaily 2008, Robust Watermarking Offers Hope Against Digital Piracy.

Retrieved May 9, 2009, from http://www.sciencedaily.com/releases/2008/12/081208092114.htm

Senthil, V., & Bhaskaran, R. 2008, ‘Robustness Analysis of Blind and Non-Blind

Multiple Watermarking using Edge Detection and Wavelet Transforms’, International Symposium on Information Technology, vol. 4, pp.1-7. Retrieved at 16 January 2010 from IEEEXplore database.

72

Silva, L. M. , Carvalho, N., Germano, P., Reis, P., Santos, A. , & Carvalho, P. 2003,

‘MMS Content Copyright Protection using Watermarking’, Proceedings Of The First International Mobile IPR Workshop. Retrieved Feb 20, 2010, from http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.105.6772&rep=rep1&type=pdf#page=48

Tang, Y.L. & Wang, C.P. 2008, ‘A Robust Watermarking Algorithm Based on Salient

Image Features’, International Conference on Intelligent Information Hiding and Multimedia Signal Processing, pp.1137-1140. Retrieved January 27, 2009, from IEEEXplore database.

Teller, S. 1996, Application Challenges to Computational Geometry, The

Computational Geometry Impact Task Force's Report. Retrieved Feb 23, 2010, from http://people.csail.mit.edu/seth/pubs/taskforce/techrep.html

The MathWorks 2009a, Wavelet Toolbox Documentation. Retrieved October 9, 2009,

from http://www.mathworks.com/access/helpdesk/help/toolbox/wavelet The MathWorks 2009b, MATLAB - The Language of Technical Computing. Retrieved

March 6, 2009, from http://www.mathworks.com/products/matlab/ Tirkel, A.Z., Van Schyndel, R.G., & Osborne, C.F. 1994, ‘A digital watermark’, Image

Processing Proceedings, vol. 1, pp. 86-90. Retrieved January 2, 2010, from IEEEXplore database.

Vincent, O.R., & Folorunso, O. 2009, ‘A Descriptive Algorithm for Sobel Image Edge

Detection’, Proceedings of Informing Science & IT Education Conference (InSITE), pp. 97-107. Retrieved Feb 21, 2010, from http://proceedings.informingscience.org/InSITE2009/InSITE09p097-107Vincent613.pdf

Wang, H.J., Chang, C.Y., & Pan, S.W. 2006, ‘A DWT-based Robust Watermarking

Scheme with Fuzzy ART’, International Joint Conference on Neural Networks. pp. 750-1757. Retrieved at Feb 15, 2010, from IEEEXplore database.

Wang, X.Y., Hou, L.M. & Wu, J. 2008, ‘A feature-based robust digital image

watermarking against geometric attacks’, Image and Vision Computing, vol. 26, no.7, pp. 980-989. Retrieved March 27, 2009, from ScienceDirect database.

73

Wang, X.Y., Hou, L.M. & Yang, H.Y. 2009, ‘A feature-based image watermarking

scheme robust to local geometrical distortions’, Journal of Optics : Pure and Applied Optics. Retrieved March 27, 2009, from Institute of Physics (IOP) Journals database.

Wang, X.Y., Yang, H.Y. & Cui, C.Y. 2008, ‘An SVM-based robust digital image watermarking against de synchronization attacks’, Signal Processing, vol. 88, no. 9, pp. 2193-2205. Retrieved March 26, 2009, from ScienceDirect database.

Wikipedia, The Free Encyclopedia 2010a, Segmentation (image processing). Retrieved

February 23, 2010, from http://en.wikipedia.org/w/index.php?title=Segmentation_(image_processing)&oldid=345424480

Wikipedia, The Free Encyclopedia 2010b, MATLAB. Retrieved March 6, 2010, from

http://en.wikipedia.org/wiki/MATLAB. Winwatermark n.d., Digital Photography Related - Digital Watermarking Processing.

Retrieved October 18, 2009, from http://www.winwatermark.com/resource/visible-watermarking-and-invisible-watermarking

Woo, C.S. 2007, ‘Digital image watermarking methods for copyright protection and

authentication’, Doctor of Philosophy Thesis, Queensland University of Technology. Retrieved April 5, 2009, from http://eprints.qut.edu.au/16457/1/Chaw-Seng_Woo_Thesis.pdf

Woo, C.S., Du, J. & Pham, B.L. 2005, ‘Geometrically Robust Digital Image

Watermarking Using Scale Normalisation and Flowline Curvature’, Proceedings Security, Steganography and Watermarking of Multimedia Contents (SPIE2005) 5681, pp. 528-538. Retrieved January 28, 2009, from http://eprints.qut.edu.au/2756/1/2756_1.pdf

Woo, C.S., Du, J. & Pham, B.L. 2006, ‘Geometric Invariant Domain for Image Watermarking’, International Workshop on Digital Watermarking. Retrieved January 28, 2009, from http://eprints.qut.edu.au/5620/1/5620.pdf

Wu, Y., & Pang, H. 2008, ‘A Lightweight Buyer-SellerWatermarking Protocol’,

Advances in Multimedia, vol. 2008, no. 2, pp. 1-7. Retrieved Feb 15, 2010, from ACM (Association for Computing Machinery) Digital Library database.

lightweight robust watermarking for …repository.um.edu.my/505/1/final_pdf.pdf · lightweight...

Documents