Amit Thesis

M. Tech (Research) Dissertation REPORT

Entitled

FPGA IMPLEMENTATION OF WATER MARKING

ALGORITHM FOR DIGITAL CAMERA

Submitted in partial fulfillment of the requirements

For the degree of M. Tech (Research)

SUBMITTED BY

Amit M. Joshi (Roll No.R07EC903)

GUIDED BY

Mr.Anand D. Darji

ELECTRONICS ENGINEERING DEPARTMENT SARDAR VALLABHBHAI NATIONAL INSTITUTE OF

TECHNOLOGY SURAT-395007

AUGUST-2009

Sardar Vallabhbhai National Institute of Technology Surat-395007, Gujarat, India.

Electronics Engineering Department

CERTIFICATECERTIFICATECERTIFICATECERTIFICATE

This is to certify that Mr. Amit M Joshi, Roll No.: R07EC903, has satisfactorily completed

a Dissertation on “FPGA Implementation of Watermarking Algorithm for Digital

Camera” as the partial fulfillment of the requirement for MASTER OF TECHNOLOGY

in ELECTRONICS ENGINEERING during the year 2008-2009.

Chairman Examiner Guide Head of the Department Prof. B. R Taunk Seal Of Department

ACKNOWLEDGEMENT

I take this opportunity to express my sincere gratitude to my guide Mr. Anand D. Darji,

Senior Lecturer, Electronics Engineering Department, for his esteemed guidance

throughout my dissertation work & his inspiration which motivated me to get fruitful

outcome through my work.

I gives me immense pleasure to thank Prof. B. R. Taunk, Head of the Department,

Electronics Department, SVNIT and Prof. N. Y. Desai, PG Incharge (Research),

Electronics Department, SVNIT for their co-operation at every stage

I would also like to thank my colleagues and family for providing me moral support,

constant inspiration, enthusiasm and co-operation for my dissertation work. Thanks to all

who have supported me directly or indirectly throughout my dissertation work.

Place: S.V.N.I.T., Surat Amit M Joshi

ACKNOWLEDGEMENT FOR SMDP-II

This project would have not been possible without support of high end EDA tools, which

were purchased under SMDP-II project by CEERI, pilani. I would like to thank MCIT,

New Delhi, without their support SMDP-II project would have never been started. I

would like to thank Incharge, Prof. N.Y.Desai and co-incharge, Mr. A.D. Darji for their

kind permission to work in VLSI Design Laboratory at our site. I would like to extend

my thanks to IIT, Bombay, which act as resource centre (RC) for SVNIT under this

project

Place: S.V.N.I.T., Surat Amit M Joshi

ABSTRACT

The rapid growth of techniques for storage and network distribution of digital media

information is demanding secure content for authentication.The image captured by digital

camera is major concern for the same as it. This image can easily altered by editig software.

Such problems have hindered the application of digital image for courtroom evidence, insurance

claim, copyright and journalistic photography. By using watermarking scheme, information such

as the author’s credentials and time stamps can be embedded into digital images for tamper-

proofing or authentication. Such an embedded watermark is transparent to general readers of

images, but it can be extracted and verified for detecting any tampering with images using a

relevant detection tool. For digital camera, we have to use invisible and robust watermarking

scheme for authenticity and ownership. The camera with inbuilt such capability is called as

“Secure Digital Camera”. Spatial domain watermarking is very widely used due to their less

computational complexity but suffers from less robustness. Frequency domain watermarking is

very popular due to transform domain property and higher robustness but it has larger

computational cost. We have adapted dual domain (spatial-frequency domain) scheme to take

advantage of both domain properties. Software implementation of this watermarking scheme is

well matured. Nowadays, hardware based watermarking has been drawn more attention as it

provides high performance and high speed. The proposed scheme provides higher MSE and

PSNR compared to other existing schemes. The main goal of research to adapt efficient FPGA

implementation of watermarking algorithm which provides ownership and authentication.

CONTENTS

List of Figures

List of Tables

List of Acronyms

Chapter 1. Introduction 1

1.1 Motivation 2

1.2 Data Hiding Techniques 5

1.3 Digital Camera Overview 5

1.4 Overview of Digital Watermarking Scheme 7

1.5 Objective of Research 8

1.6 Summary 9

Chapter 2. Digital Watermarking 10

2.1 Literature Survey of Watermarking 10

2.2 Basics of Watermarking 12

2.2.1 Characteristics of Watermarking 12

2.2.2 Types of watermarks 13

2.2.3 Classification of Watermarking Method 15

2.2.4 Conflicting Requirements of watermarking 15

2.3 Watermarking for Different Domain 16

2.3.1 Spatial domain Watermarking 17

2.3.2 Frequency Domain Watermarking 17

2.3.3 Spatial Domain Vs Frequency Domain Watermarking 18

3.2.4 Advantage of wavelet based Watermarking Techniques 19

2.4 Types of watermark attacks 19

2.5 Common Attack Techniques 20

2.6 Performance of Watermarked Image 21

2.7 Summary 22

Chapter 3. Wavelet Transform 23

3.1 History of Frequency Domain Transformation 23

3.2 Discrete Wavelet Transform (DWT) 23

3.2.1 Sub band for Wavelet Based Coding 24

3.2.2 Watermarking with Wavelet Decomposition Level 25

3.3 Implementation of Wavelet Transform 26

3.3.1 Relationship between Convolution and lifting scheme 26

3.3.2 Advantage of lifting scheme 27

3.3.3 Lifting based Scheme 28

3.4 Summary 33

Chapter 4. DWT Based Watermarking Algorithms 34

4.1 Algorithm 1 34

4.1.1 Watermark Embedding Method 34

4.1.2 Watermark Detection Method 35

4.2 Algorithm 2 35

4.2.1 Watermarking Embedding Algorithm 35

4.2.2 Watermark Extraction Process 37

4.2.3 Effect of Scaling factor α and Quantization factor q 37

4.3 Algorithm 3 39

4.3.1 The Embedding Method 40

4.3.2 Watermark Extraction Method 41

4.4 Proposed Watermarking Algorithm 42

4.4.1 Bit-plane slicing Method 43

4.5 Summary 46

Chapter 5. Implementation of Proposed Techniques 47

5.1 Hardware Implementation Lifting based DWT Scheme 50

5.1.1 Implementation of (lossless) Le gall 5/3 algorithm 50

5.1.2 Memory management 53

5.2 Watermark Embedding Hardware Implementation 56

5.2.1 Bit Plane Slicing Implementation 57

5.2.2 Random Number Generator 58

5.3 VLSI Implementation for Proposed Scheme 59

5.3.1 Architecture of Proposed scheme 59

5.3.2 Pin Diagram 60

5.4 Summary 61

Chapter 6. Results Analysis 62

6.1 MATLAB Results 62

6.2 FPGA Implementation Results 75

Chapter 7. Conclusion 81

References 82

Appendix A

Appendix B

Appendix C

LIST OF FIGURES

Figure 1.1 Watermarking Example

Figure 1.2 Watermarking Technique

Figure 1.3 Block Diagram of Digital Still Camera

Figure 1.4 Watermark Embedding Process

Figure 1.5 Watermark Extraction Process

Figure 2.1 Subband Coding

Figure 2.2 Level Of Decomposition for Wavelet Transform

Figure 2.3 Convolution based scheme

Figure 2.4 Lifting based filtering Method

Figure 3.1 Types of watermarks [34]

Figure 3.2 Conflicting Requirements

Figure 4.1 Watermark embedding and extraction step

Figure 4.2 Plot for different α versus PSNR for original and watermarked Images

Figure 4.3 Plot for α versus similarity factor for original and watermarked image

Figure 4.4 Plot for different q versus PSNR for original and watermarked Images

Figure 4.4 Plot for q versus similarity factor for original and watermarked image

Figure 4.6 Embedding Process

Figure 4.7 The pixel array graph of the block

Figure 4.8 Watermarking extracting process

Figure 4.9 Wavelet Decomposition

Figure 4.10 Bit Plane Slicing Scheme to LL Band

Figure 4.11 Explanation of Proposed Algorithm step by step

Figure 5.1 VLSI Flow for Proposed Scheme

Figure 5.2 Implementation Flow for Proposed Scheme

Figure 5.3 Watermark Detection Process

Figure 5.4 Lifting architecture for 5/3 wavelet

Figure 5.5 Single scale one dimensional wavelet lifting (5-3 lossless wavelet)

Figure 5.6 Predict phase

Figure 5.7 Update phase

Figure 5.8 1-D DWT unit

Figure 5.9 Memory mapping

Figure 5.10 Flowchart to calculate the 2D Forward wavelet Transform

Figure 5.11 Flowchart to calculate the 2D Inverse wavelet Transform

Figure 5.12 Bit Plane Slice Method

Figure 5.13 Random Number generation scheme

Figure 5.14 Architecture of proposed scheme

Figure 5.15 Pin diagram

Figure 6.1 Comparison of PSNR for all algorithms for gray and color images

Figure 6.2 Comparison of Similarity Factor for all algorithms for gray and color

images

Figure 6.3 Down sampling and up sampling process for Y component

Figure 6.4 Reconstructed Y plane after up sampling

Figure 6.5 Modified Down-Up sampling process with proper reconstruction of

I Image

Figure 6.6 Down sampling and up sampling process for Cb component

Figure 6.7 Down sampling and up sampling process for Cr component

Figure 6.8 Simulation of Legal 5/3 Scheme

LIST OF TABLES

Table 2.1 Complexity comparison of convolution and lifting based

implementation

Table 2.2 Daubenchies 9/7 analysis filter coefficient

Table 2.3 Daubenchies 9/7 synthesis filter coefficient

Table 2.4 Le Gall 5/3 analysis and synthesis filter coefficient

Table 6.1 Comparison of Original Image and Watermarked Image without any

attack

Table 6.2 Comparison of Original Image and Watermarked Image with Gaussian

noise attack

Table 6.3 Comparison of Original Image and Watermarked Image with 1 degree

anticlockwise rotation

Table 6.4 Comparison of Original Image and Watermarked Image with JPEG

Compression

Table 6.5 Comparison of Original Watermark and Extracted Watermarked without

any attack

Table 6.6 Comparison of Original Watermark and Extracted Watermark with 1

degree anticlockwise rotation

Table 6.7 Comparison of Original and Extracted Watermark with JPEG

Compression

Table 6.8 Comparison of Original Watermark and Extracted Watermark with

Gaussian noise

Table 6.9 Comparison of Original Image and Watermarked Image without any

attack

Table 6.10 Comparison of Original Image and Watermarked Image with Gaussian

attack

Table 6.11 Comparison of Original Image and Watermarked Image with 1 degree

anticlockwise rotation


Compression

Table 6.13 Comparison of Original Watermark and Extracted Watermarked without

any attack

Table 6.14 Comparison of Original Watermark and Extracted Watermark with

Gaussian noise

Table 6.15 Comparison of Original Watermark and Extracted Watermark with 1

degree anticlockwise rotation


Compression

Table 6.17 Synthesis Report for Proposed DWT, IDWT and Watermarking

Processor(SPARTEN 3E)

Table 6.18 Device Utilization Report for Proposed DWT, IDWT and Watermarking

Processor (SPARTEN 3E)

Table 6.19 Timing Reports for DWT,IDWT and Watermarking Processor

(SPARTEN 3E)

Table 6.20 Synthesis Report for Proposed DWT, IDWT and Watermarking

Processor (VIRTEX XCV50-6bg256)

Table 6.21 Device Utilization Report for Proposed DWT, IDWT and Watermarking

Processor (VIRTEX XCV50-6bg256)

Table 6.22 Timing Reports for DWT,IDWT and Watermarking Processor

(VIRTEX XCV50-6bg256)

Table 6.23 Comparison with Previous Work

Table 6.24 Device utilization comparison (VIRTEX XC4VLX25-10FF676)

Table 6.25 Timing analysis comparison (VIRTEX XC4VLX25-10FF676)

LIST OF ACRONYMS

DCT Discrete Cosine Transform

DRM Digital Right Management

CMOS Complementary Metal Oxide Semiconductor

LCD Liquid Crystal Display

ADC Analog to Digital Converter

DAC Digital to Analog Converter

JPEG Joint Photographic Experts Group

MPEG Motion Photographic Experts Group

DWT Discrete Wavelet Transform

IDWT Inverse Discrete Wavelet Transform

VLSI Very Large Scale Integrated

LSB Least Significant Bit

HVS Human Visual System

FPGA Field Programmable Gate Array

ASIC Application Specific Integrated Circuit

DFT Discrete Fourier Transform

FFT Fast Fourier Transform

RTL Register Transfer Level

PSNR Peak Signal to Noise Ratio

MSE Mean Square Error

SF Similarity Factor

I/O Input/Output

RAM Random Access Memory

ROM Read Only Memory

MUX Multiplexer

LUT Look Up Table

IOB Input/Output Blocks

GCLK Global Clock Buffer

FPGA Implementation of Watermarking Algorithm for Digital Camera

VLSI Design Lab., SVNIT, Surat 1

CHAPTER 1. INTRODUCTION

Digital photography is rapidly gaining popularity, and so the authentication of images taken by

digital cameras has become a great concern. Since the growth of computer networks and internet,

the information types of multimedia contents have changed. These kinds of digital contents are

easy to be stolen by way of transmitting on networks. For this reason, such technical mechanism

as copy protection, distribution tracing, and usage control are important issue for copyright

protection. Digital camera has been widely used in present days and images captured by it are

displayed on WWW and electronic commerce. But digital images are not easily acceptable in

courts, news reporting, and medical archiving because they can be altered or manipulated with

ease [1].

As the computers are more and more integrated via the network, the distribution of digital

media is becoming faster, easier, and requiring less effort to make exact copies. One of the major

impediments is the lack of effective intellectual property protection of digital media to

discourage unauthorized copying and distribution. Watermarking is a technique for inserting

information (watermark) into an image, which can be extracted or detected later for a variety of

purposes such as identification and/or authentication. Also referred to as simply watermarking, a

pattern of bits inserted into a digital image, audio or video file that identifies the file's copyright

information (author, rights, etc.). The name comes from the faintly visible watermarks imprinted

on stationery that identify the manufacturer of the stationery. The purpose of digital watermarks

is to provide copyright protection for intellectual property that's in digital format. Unlike printed

watermarks, which are intended to be somewhat visible, digital watermarks are designed to be

completely invisible, or in the case of audio clips, inaudible. Moreover, the actual bits

representing the watermark must be scattered throughout the file in such a way that they cannot

be identified and manipulated.

The digital watermark must be robust enough so that it can withstand normal changes to the

file, such as reductions from lossy compression algorithms. Satisfying all these requirements is

no easy feat, but there are a number of companies offering competing technologies. All of them

work by making the watermark appear as noise - that is, random data that exists in most digital

files anyway. To view a watermark, you need a special program that knows how to extract the

watermark data. Watermarking is also called data embedding and information hiding.



1.1 Motivation

Owing to the usage of internet, concerns about protecting and enforcing intellectual property

(IP) rights of the digital content are mounting. Unauthorized replication and manipulation of

digital content is relatively easy and can be achieved with inexpensive tools. Digital Rights

Management (DRM) systems ([2], [3]) address issues related to ownership rights of digital

content. Various aspects of content management – namely; content identification, storage,

representation, and distribution – and IP rights management are highlighted in DRM. Although

unauthorized access of digital content is being prevented by implementing encryption

technologies, these approaches do not prevent an authorized user from illegally replicating the

decrypted content. Digital watermarking is one of the key technologies that can be used in DRM

systems for establishing ownership rights, tracking usage, ensuring authorized access, preventing

illegal replication, and facilitating content authentication. Therefore, a two-layer protection

mechanism utilizing both watermarking and encryption is needed to build effective DRM

systems that can address IP rights and copyright issues [4].

Conventionally, in analog world, a painting is signed by the artist to attest the copyright, an

identity card is stamped by the steel seal to avoid forgery, and the paper money is identified by

the embossed portrait. Such kind of hand-written signatures, seals and watermarks have been

used from ancient times as a way to identify the source, creator of a document or a picture. For

example, a priceless painting of the 11th century in National Palace Museum named "Travelers

on a Mountain Path" had not been identified as the genuine work of Fan Kuan until Fan's

signature is found between the woods behind a group of travelers of the painting.

(a) Travelers on a Mountain (b) Fan’s signature

Fig 1.1 Watermarking Example



However, in the digital world, digital technology for manipulating images has made it difficult

to distinguish the visual truth. Besides, the characteristics of digitization bring significant hangs

in copyright issues, which create an urgent need to intellectual property protection on the

digitally recorded information. Digital watermarking has been proposed as a way to claim the

ownership of the source and owner. Unlike encryption, watermarking does not restrict access to

the data. Once the encrypted data is decrypted, the intellectual property rights are no longer

protected.

Over the past few years, the technology of the digital watermarking has gained prominence and

emerged as a leading candidate that could solve the fundamental problems of legal ownership

and content authentications for digital multimedia data. A great deal of research efforts has been

focused on digital image watermarking in recent years. The techniques proposed so far can be

divided into two groups according to the embedding domain. One group is spatial domain

approach.The other group is frequency domain approach.

The process of embedding a watermark in a multimedia object is termed as watermarking.

Watermark can be considered as a kind of a signature that reveals the owner of the multimedia

object. Content providers want to embed watermarks in their multimedia objects (digital content)

for several reasons like copyright protection, content authentication, tamper detection etc. A

watermarking algorithm embeds a visible or invisible watermark in a given multimedia object.

The embedding process is guided by use of a secret key which decided the locations within the

multimedia object (image) where the watermark would be embedded. Once the watermark is

embedded it can experience several attacks because the multimedia object can be digitally

processed. The attacks can be unintentional (in case of images, low pass filtering or gamma

correction or compression) or intentional (like cropping). Hence the watermark has to be very

robust against all these possible attacks. When the owner wants to check the watermarks in the

possibly attacked and distorted multimedia object, s/he relies on the secret key that was used to

embed the watermark. Using the secret key, the embedded watermark sequence can be extracted.

This extracted watermark may or may not resemble the original watermark because the object

might have been attacked. Hence to validate the existence of watermark, either the original

object is used to compare and find out the watermark signal (non-blind watermarking) or a

correlation measure is used to detect the strength of the watermark signal from the extracted

watermark (blind watermarking).In the correlation based detection the original watermark



sequence is compared with the extracted watermark sequence and a statistical correlation test is

used to determine the existence of the watermark.

Fig 1.2 Watermarking Technique

Each of the watermarking algorithms has its own applications; thus, all are equally important.

Over the past decade, numerous watermarking algorithms have been invented, and their software

implementations have been demonstrated ([5], [6]). However, only a few hardware

implementations are presented in the literature. A hardware based watermarking system can be

designed on a field programmable gate array (FPGA) board, Trimedia processor board [7], or

custom integrated circuit (IC) [8]. The choice between an FPGA and a cell-based IC is a trade-

off among cost, power consumption, and performance ([8], [9]). In this thesis, we present both

FPGA-based prototyping and a custom IC design to facilitate high-performance real-time

watermarking at the source end when the image is captured by an electronic component like a

digital camera.

We propose a new concept for digital cameras to solve some of the significant problems

associated with the use of digital images as evidence in a court of law. The integrity of digital

images as evidence rests on the accurate answering of a simple question: Who did what when?

Watermarking algorithms and Encryption schemes are used to identify from the digital image the



photographer, the camera, the time when the image was taken, and verify the image

authentication. We call a camera with this capability “Secure Digital Camera”.

1.2 Data Hiding Techniques

Cryptography: It scrambles a message into a code to obscure its meaning. Scrambling of

message is done with help of secret key. Scrambling message called as encrypted and it is again

decrypted with that secret key only. Cryptography provides security to message.

Steganography: With Steganography, the sender of a message would hide it in a host file. The

host file, or cover message, is the file that anyone can see. When people use steganography, they

often hide the true intent for communicating in a more common place communication scenario.

In steganography, usually the message itself is of value and must be protected through clever

hiding techniques and the “vessel” for hiding the message is not of value.

Watermarking: Watermarking is the direct embedding of additional information into the

original content or host signal. Ideally, there should be no perceptible difference between the

watermarked and original signal and the watermark should be difficult to remove or alter without

damaging the host signal. In watermarking, the effective coupling of message to the “vessel,”

which is the digital content, is of value and the protection of the content is crucial.

In case of steganography, where the method of hiding the message may be secret and the

message itself is secret, in watermarking, typically the watermark embedding process is known

and the message (except for the use of a secret key) does not have to be secret.

1.3 Digital Camera Overview

A block diagram of the proposed digital watermark is shown in Fig 1.1[10]. In the proposed

digital camera, the image is captured by an image sensor and converted to a digital signal by the

A/D converter. A Complementary Metal Oxide Semiconductor (CMOS) image sensor that has an

embedded A/D converter will be used. The captured image is stored temporarily in the scratch

memory, after which it is displayed on the LCD panel with the help of the controller. The

purpose of the Liquid Crystal Display (LCD) panel is to enable the user to see the image frame

before it is watermarked by the watermarking unit and stored in the camera, which can then be

further transmitted over the network, or transferred to flash memory, computer hard drive or

optical discs. The controller unit that is responsible will handle both color and monochrome



images. We will use robust and invisible watermarking scheme for implementation of

authenticity. The spatial domain co-efficient are converted to frequency domain co-efficient with

the help of frequency domain transformation methods like Discrete Cosine Transform (DCT),

Discrete Wavelet Transform (DWT) and Discrete Fourier Transform (DFT) [11]. The co-efficient

selection key provides the selected coefficients to embed the watermark. The watermark unit is

embedding the watermark to give “watermarked secure image”. User has been provided with a

flexibility to make a decision about insertion of watermark into an image. Based on requirement,

watermark may or may not be inserted into image. Images with and without watermark are

stored in two separate memories. Images without watermark are stored in temporary memory and

watermarked image are stored in flash memory. In case of non blind detection, images stored in

temporary memory can be used for detection.

Fig 1.3 Block Diagram of Digital Still Camera

A watermarking system can be implemented with either software or hardware. Most

watermarking algorithms have been developed for software implementations due to ease of use,

upgrading and flexibility. However, the software implementations have the limited speed

problem and are vulnerable to the off-line attacks. In software, implementation requires the

addition of a dedicated processor such as a DSP core that occupies considerably more area,

consumes significantly more power, and may still not perform adequately fast. The hardware

CMOSSensor

Sensor

ADCTemporaryMemory

LCD Controller

FrequencyTransformation

WatermarkingUnit

FlashMemory

InputImage

Co-efficientKey

Out

WatermarkedImage



implementation has advantage over the software implementation in terms of reliability and high

performance. The overall advantage is that hardware consumes less area, less power and faster

for implementing [12].If a chip is to be fitted in the digital devices, the output videos or images

can be marked right at the original although the same can be done using software after those

videos or images downloaded to the computer. But in this case embedding software will take

more time compared to hardware. The example of TV broadcast will highlight the significance

where digital media is to be marked in real time and hardware is the only solution.

1.4 Overview of Digital Watermarking scheme

The watermark extraction process is similar to the watermark embedding process. The whole

watermark embedding process is shown in Fig 1.4 [13] whereas the watermark extraction

process is shown in Fig 1.5 [13].

As with traditional color processing, we first convert an image from an RGB color space to the

YCbCr color space [14]. Then Y component of the image is down-sampled to form a grayscale

image of resolution of 1 M pixels (assuming the original is between 2 M and 8 M pixels true for

most digital cameras today). Afterwards, a watermark is embedded in the image of the

coefficients of the nth sub band level DWT of the image with secret key. Finally, the Y image

plane is converted back to spatial domain by IDWT and a watermarked image is formed by up-

sampling the image and adding it with the original Y, Cb and Cr color components. For

extraction process, the user has access of watermark (w), the coefficient selection key (c key)

Fig 1.4 Watermark Embedding Process

WC-key

Upsampling

YCbCr toRGB

IDWT

F

CbCr

DownSampling

RGB toYCbCr DWT

EmbeddingBlock

Y

Z



Fig 1.5 Watermark Extraction Process

([15],[16]) and the original image incase of non blind watermarking. Since only the user of the

digital camera knows the secret key for the watermarking, security against forgery is guaranteed.

As shown in above Fig 1.4, watermark is applied to frequency transform domain. There are

various schemes for spatial to frequency domain conversion like DFT,DCT,DWT etc. DWT is

recently used technique for frequency domain conversion. There are various methods available

for DWT based watermarking schemes. Arnold transform based watermarking scheme

([17],[18]) can be used where watermarked symbol is also transformed before embedding it to

the original image. Energy based watermarking scheme ([19],[20],[21],[22]) where, we embed

the watermark into the detail wavelet coefficients using a watermarking key, only in the first

level of the decomposed image. In middle band algorithm [23], one way to use one middle

frequency co-efficient as a scale to quantize the other co-efficent of the same middle frequency

band pair. Spread spectrum based watermarking [24] can be used for both spatial and spectral

domain watermarking.

1.5 Objective of Research

The objective of research is to develop a feasible and hardware implemented watermarking

scheme for digital cameras so as to provide a cost-effective and practical tool for authentication

and ownership verification of images taken by digital camera. The objectives of research for

thesis are as follows:

1) To study the different watermarking algorithms

2) To compare various watermarking algorithms to select best for FPGA/ASIC

implementation.

RGB toYCbCr

DownSampling

DWT MarkExtraction

Z

C-keyW

W’’



3) To propose the efficient algorithm which is useful for copyright protection and ownership

verification?

4) To achieve FPGA implementation of watermarking schemes and comparison based on

power, area and speed.

5) To synthesize the net list using synopsys Design Vision for UMC .18 µm standard Cell

library.

1.6 Organization of Thesis

The outline of rest of thesis is as follows: Chapter 2 includes literature survey of watermarking

and different characteristics of digital watermarking, types of watermarking and geometrical

attacks for watermarking. Chapter 3 describes the basics of Discrete Wavelet Transform (DWT)

and lifting based legall 5/3 scheme which has lesser computer complexity. Chapter 4 analyzes

the different algorithm for watermarking and also proposes scheme which has comparable PSNR

and MSE results.Chapter 5 emphasis implementation of wavelet based lifting scheme and

watermark hardware implementation. Chapter 6 provides results of different watermarking

algorithms, synthesis reports for wavelet based watermarking scheme and timing reports for

proposed scheme.



CHAPTER 2: DIGITAL WATERMARKING

Watermarking is the process that embeds data called a watermark, tag or label into a

multimedia object such that watermark can be detected or extracted later to make an assertion

about object. The object may be an image, audio, video or text. In general, any watermarking

scheme consists of three parts, such as the watermark, the encoder and the decoder.

2.1 Literature Survey of Watermarking

Previously, spatial domain watermarking techniques for image data was popular. Some of the

earliest techniques embed m sequences into the least significant bit (LSB) of the data to provide

effective transparent embedding technique ([25], [26]). M-sequences are chosen due to their

good correlation properties so that a correlation operation can be used for watermark detection.

Furthermore, these techniques are computationally inexpensive to implement. Such a scheme

was first proposed in and extended to two dimensional [27].Then, the authors reshape the m

sequence into two-dimensional watermark blocks which are added and detected on block by

block basis [25]. The block based method, known as variable two dimensional watermark

(VW2D) shown to be robust to JPEG compression. This technique has also been shown to be an

effective fragile watermarking scheme which can detect image alternation on a block basis [28].

Another spatial domain technique consists of embedding a texture based watermark into a

portion of the image with similar texture, it will be difficult to pensive the watermark. The

watermark is detected using a correlation detector [29]. Transformation domain watermarking is

useful for taking advantage of perceptual criteria in embedding process, for designing

watermarking techniques which are robust to common compression techniques, and for direct

watermark embedding of compressed bit stream. A common transform framework for images is

the block based DCT which is a fundamental building block of current image coding standards

such as JPEG and video coding standards such as the MPEG video coders [30] and the ITU H.26

x family of codecs.

First DCT based semi fragile watermarking algorithm for digital camera with FPGA

implementation was developed by Hyun Lim, Soon-Young Park and Seong jun Kang in early

2003 [31]. In 2004, Saraju P. Mohanty, N. Raganathan had also developed visible watermarking

scheme on DCT [32]. After that, algorithms for wavelet based approach were developed to adapt



JPEG2000 new millennium standard and to explore multiresolution property of wavelet ([33],

[34]). In 2005 later part, Sammy H. M. Hawk and Edumund Y. Lam proposed watermark

implementation technique in digital photography with DWT approach [35] for software based

implementation. In 2006, Lei Tian and Heng-Ming Tai had developed Secure Images captured

by Digital camera for DWT based approach [36]. Another spread spectrum watermarking

techniques provides better perceptual transparency and watermark robustness ([37], [38]). This

can also developed for secure digital camera application. A. Lumini and Dario Maio have

developed watermarking scheme with random binary sequence [39]. Yong-Gang Fu and Hui -

Rong Wang have done watermarking on threshold based scheme [40]. Hsien-Wen Tseng and

Chin-Chen Chang suggested watermarking with block processing method for with differential

expansion method [41]. Tung Shou Chen, Jeanne Chen and Jian-Guo Chen have developed

simple and efficient watermark technique for JPEG2000 Codec with scattered matrix watermark

[42].

The current literature is rich in watermarking algorithms developed for various types of media,

such as image, video, audio, and text data, and their software implementations. The algorithms

work in various domains like spatial, DCT, and wavelet and insert-extract different types of

watermarks including invisible robust, invisible fragile, and visible. These watermarking

algorithms primarily work off-line; i.e., the images are first acquired and then the watermarks are

inserted before the watermarked images are made available to the user. Thus, in these

approaches, there is a gap between image capture and image transmission. The objective of work

was to develop a hardware-based watermarking system to bridge this gap. The watermark chip

will be fitted in any electronic component that acquires the images, which are then watermarked

in real time while capturing.

A DCT domain invisible watermarking chip is presented by Tsai and Lu [43]. The

watermarking system embeds a pseudorandom sequence of real numbers with a selected set of

DCT coefficients and is extracted without using the original image. The chip is implemented

with TSMC 0.35 µm technology and has a die size of 3.064 x 3.064 mm2 and 46,374 gates. The

chip is estimated to consume 62.78 mW of power when operated at 50 MHz frequency with a

3.3V supply.

Garimella et al. [44] have proposed a watermarking VLSI architecture for invisible fragile

watermarking in the spatial domain. In this scheme, the differential error is encrypted and



interleaved along with the first sample. The watermark can be extracted by accumulating the

consecutive least significant bits (LSBs) of the pixels and then decrypting them. The extracted

watermark is then compared with the original watermark for image authentication. The

Application Specific Integrated Circuit (ASIC) is implemented using 0.13 µm technology. The

area of the chip is 3453 x 3453 µm2, and the chip consumes 37.6 µW of power when operated at

1.2 V. The critical path delay of the circuit is 5.89 ns.

Mohanty et al. [8] have proposed another watermarking hardware architecture that can insert

two visible watermarks in images in the spatial domain. This architecture can insert either of the

two watermarks depending on the requirements of the user. The chip is implemented with 0.35

µm2 technology and occupies an area of 3.34 x 2.89 mm2 and consumes 6.9286 mW when

operated at 3.3V and 292.27 MHz.

Fan et al. [45] have proposed a visible watermarking design based on an adaptive discrete

wavelet transform (DWT). They proposed efficiently reduced operational and resource-sharing

techniques using an existing algorithm. Host image and watermark are transformed into three-

level multi-resolution structures. The host image signal is divided into two sequences with the

same pattern length. Processing time is reduced by using a two-path parallel processing

architecture. The signal is sent to different processing elements by the demultiplexers. The

watermark image is embedded by modifying the coefficients of the image.

In this report, we describe a VLSI architecture that implements both invisible and robust

watermarking functionalities. We first decompose original image with first level DWT. Then, we

applied bit plane slicing method to obtain MSB to LSB plane. We embed the watermark in the

LSB plane. A binary watermark generated from pseudo random number generator is embedded

into selected co-efficient with the original image bit plane in the invisible robust watermarking

algorithm. The VLSI architecture is prototyped with a Xilinx FPGA and a custom IC design.

2.2 Basics of Watermarking

2.2.1 Characteristics of Watermarking

A watermark is designed to permanently reside in the host data. When the ownership of data is

in question, the information can be extracted to completely characterize the owner. To achieve

maximum protection of intellectual property with watermarked media, several requirements must

be satisfied. Digital watermarking is regarded as a possible solution to ownership problem



because it can embed the owner information into images to protect the copyright. A good

watermarking scheme should meet following requirements.

Imperceptible: The digital watermark should not be perceptible to the viewer. It means that

after the embedding process, the quality of the image should not be degraded. The watermark

should be imperceptible so as not to affect the viewing experience of the image or the quality of

the audio signal.

Undeletable: The watermark must be difficult or even impossible to remove by a malicious

cracker, at least without obviously degrading the host signal.

Statistically undetectable: A pirate should not be able to detect the watermark by comparing

several watermarked signals belonging to the same author.

Robustness: The watermark should be able to survive lossy compression techniques like JPEG,

which is commonly used for transmission and storage. The embedded mark should resist

malicious attacks such as compression, linear or nonlinear filtering, image enhancement,

resizing. The watermark should be retrievable even if common signal processing operations are

applied, such as signal enhancement, geometric image operations and noise filtering.

Unambiguous: Retrieval of the watermark should unambiguously identify the owner, and the

accuracy of identification should degrade gradually in the face of attacks.

Security: The embedded mark should only be accessible by authorized organization. Although

the watermarking algorithm is published to everyone, those without a secret key cannot access

the marks in the protected images.

Capacity: The amount of information that can be stored in a watermark depends on the

application. For application like digital images authentication, a payload of one bit is usually

sufficient.

2.2.2 Types of watermarks [46]

Visible: Visible watermarks are designed to be easily perceived by the viewer, and should

clearly identify the owner. The watermark must not detract from the image content itself.

However, most research currently focuses on invisible watermarks, which are imperceptible

under normal viewing conditions.

Fragile: Fragile watermarks are designed to be distorted, or to be broken, under the slightest

changes to the image. Semi-fragile watermarks are designed to break under all changes that

exceed a user-specified threshold.



Spatial: These are constructed in the image spatial domain, and embedded directly into an

image’s pixel data.

Spectral: These are constructed in the image transform domain, and watermarks are

incorporated into an image’s transform coefficients Discrete-Cosine Transform (DCT), Discrete-

Wavelet Transform (DWT).

Public & private: In some applications like copyright protection and data monitoring,

watermark extraction algorithm can use original un-watermarked data to find the watermark.

This is called as non oblivious or private watermarking. The original image is usually subtracted

from the suspected image before a detection algorithm is applied. Non-oblivious watermarking is

at present the only option for reliable copyright protection. In most other applications e.g. copy

protection and indexing, the watermark extraction algorithms do not have access to the original

watermark data. This renders the watermark extraction more difficult. Watermarking algorithm

of this kind is referred as public, blind or oblivious watermarking algorithm.

Fig 2.1 Types of watermarks [46].

Asymmetric & Symmetric: Asymmetric watermarking (also called asymmetric key

watermarking) is a technique where different keys are used for embedding and detecting the

watermark. In symmetric watermarking (or symmetric key watermarking), the same keys are

used for embedding and detecting watermarks.

The image captured by Digital Camera can not alter with malicious attacks so it should be

robust. We are using original image at the time of detection so it is private type of digital

watermarking.

Types of Watermark

Invisible Visible

Fragile Robust



2.2.3 Classification of Watermarking Method

There are two types of watermarking:

Irreversible watermarking

Reversible watermarking

Irreversible watermarking: Permanent distortion is one of the main drawbacks of all the

irreversible watermarking schemes. Attempts to recover the original signal after the signal

passing the authentication process is not guaranteed. Some common problems, such as salt-and

pepper artifacts and retrieval of original image can now be resolved using reversible

watermarking. Irreversible watermarking deals with copyright protection focusing more on

robustness.

Reversible watermarking: Reversible watermarking deals with fragile watermarking

applications like authentication and integrity of an image by embedding the assigned watermark

into the original image and it can also recover the original image from the suspected image. The

retrieved watermark can be used to determine the ownership by comparing the retrieved

watermark with the assigned one. Similar to conventional watermarking schemes, reversible

watermarking schemes have to be robust against the intentional or the unintentional attacks and

should be imperceptible to avoid the attraction of attacks and value lost. Therefore, the reversible

watermarking also has to satisfy all requirements of the conventional watermarking such as

robustness, imperceptibility, and readily embedding and retrieving. Reversible watermarking’s

application areas comprise of authenticity of the image where fragility is the main requirement.

For image authentication purpose for digital camera requires reversible watermarking approach

because original image should be retrieved at the time of extraction.

2.2.4 Conflicting Requirements of watermarking

The properties listed above are all related to each other. For instance, a very robust watermark

can be obtained by making many large modifications to the host data for each bit of the

watermark. Large modifications in the host data will be noticeable, however, and many

modifications per watermark bit will limit the maximum amount of watermark bits that can be

stored in a data object. Hence, a tradeoff should be considered between the different

requirements so that an optimal watermark for each application can be developed. The mutual

dependencies between the basic requirements are shown in Fig 2.2.



Fig 2.2 Conflicting Requirements

2.3 Watermarking In Different Domain

Generally, there are two broad watermarking schemes and categories: spatial domain

watermarking and frequency domain watermarking. In a spatial domain watermarking scheme,

the watermark is embedded by direct modifying the pixel value of an image. These spatial

domain watermarking schemes are simple and less robust to common signal processing

operations, since the watermark doesn’t spread all over the image and some common signal

processing easily erase the embedded watermark without affecting the quality of the

watermarked image. On the other hand, frequency domain watermarking schemes involve

embedding the watermark by modifying the transform co-efficient, after the image has been

transformed to the transform domain. Human eyes are more sensitive to noise in the lower

frequency range than in its higher frequency range counterpart. (The energy of most natural

images is concentrated on the lower frequency range.) Therefore, the quantization table applied

in lossy compression always reflects the human visual system that is less sensitive to

quantization noise at higher frequencies. In order to invisibly embed the watermark and survive

the lossy data compression, a reasonable trade-off is to embed the watermark into the middle-

frequency range of the image. To prevent an expert from extracting the hidden information

directly from the transformed domain, the watermarks are embedded by modifying the

relationship of the neighboring blocks of middle-frequency coefficients of the original image.

This is done instead of embedding by an adaptive operation. “Adaptive” watermarks may be

detected by using adaptive filters – as the watermark is based on the original image

characteristics only. The former is better as the coefficient governing relationship is known to the

Capacity

RobustnessImperceptibility



creator only. The oldest technique of embedding data into images is the Least Significant Bit

(LSB), and it is implicitly based on masking. What LSB does initially is to set the least significant

bit of each pixel to 0. This method satisfies the perceptual transparency property, since only the

least significant bit of an 8-bit value is altered. Data can be embedded into the image by choosing

the desired values of 0’s and 1’s for the LSB. This method was initially designed to work for gray

scale images. But it can be easily extended to color images by treating each plane as the single

plane in the former. This technique is useful for detecting size modifications or when some

editing may have been done to the image.

2.3.1 Spatial domain Watermarking

Spatial domain watermarking is performed by modifying values of pixel of a frame. Let us

denote a picture to be watermarked by P and values of its pixel samples by Pi, a watermarked

version of picture P by 'P and values of its pixel color samples by 'P i . Let us have as many

elements of watermark W with values Wi as number of pixels in picture P. Watermark W hereby

covers the whole picture P. Further, it is possible to increase the watermark strength by

multiplying watermark element values by weight factor α. Then the natural formula for

embedding watermark W into picture P is:

'i i iP P W (2.1)

The most common algorithm using spatial domain watermarking is LSB. The simplest

watermarking method in spatial domain is to just flip the least significant bit (LSB) of chosen

pixels in an image. Given the extraordinarily high channel capacity of using the entire cover for

transmission in this method, a smaller object may be embedded multiple times. Even if most of

these are lost due to attacks, a single surviving watermark would be considered a success.

2.3.2 Frequency Domain Watermarking

These methods are similar to spatial domain watermarking in that the values of selected

frequencies can be altered. Because high frequencies will be lost by compression or scaling, the

watermark signal is applied to lower frequencies, or better yet, applied adaptively to frequencies

containing important elements of the original picture. Upon inverse transformation, watermarks

applied to frequency domain will be dispersed over the entire spatial image, so these methods are

not as susceptible to defeat by cropping as the spatial technique.



a) Discrete Cosine Transform (DCT): DCT allows an image to be broken up into different

frequency bands, making it much easier to embed watermarking information into the middle

frequency bands of an image. The middle frequency bands are chosen such that they have

minimize to avoid the most visual important parts of the image (low frequencies) without over-

exposing themselves to remove through compression and noise attacks (high frequencies).In this

technique, the original image is divided into 8x8 blocks of pixels. Then, the 2-D DCT is applied

independently to each block. After that, coefficients of the middle-frequency range are picked

from the DCT coefficients.

b) Discrete Wavelet Transform (DWT) :In DWT based approach, each level produces four

level bands of data, one corresponding to the low pass band(LL), and three other corresponding to

horizontal (HL), vertical (LH) and diagonal (HH) high pass bands. The decomposed images show

a coarse approximation image in the lowest resolution low pass band, and three detail image in

higher bands. The low pass band can further be decomposed to obtain another level of

decomposition. Watermark data inserted into low frequencies is more robust to image distortions

that have low pass characteristics like filtering, lossy compression and geometric manipulation

and less robust to change of the histogram such as contrast/brightness adjustment, gamma

correction and cropping. On the other hand, watermark data inserted into middle and high

frequencies is typically less robust to low pass filtering, lossy compression and geometric

deformation of the image but extremely robust with respect to noise adding and nonlinear

deformation of the gray scale.

2.3.3 Spatial Domain Vs Frequency Domain Watermarking [47]

The watermark can be applied to frequency or in spatial domain. The frequency domain

methods are more robust than that of the spatial domain techniques. On the other hand, the

spatial domain watermarking schemes have less computational overhead compared to frequency

domain schemes. Nowadays, most researches focus on developing various watermarking

methods in the frequency domain since probably most currently popular information

transmission and compression techniques are in frequency domain e.g. like JPEG, MPEG, DCT,

etc. Several methods have been processed in frequency domain such like DFT, DCT and DWT.

Compared with other methods in frequency domain, DWT can better localize the feature whose

change is less sensitive to human eyes and localize the information in space and frequency.



Especially, DWT based watermarking methods have been researched intensively due to the fact

that the current image compression is based on wavelet domain, such like JPEG 2000.The spatial

domain watermarking scheme is generally fast and simple, but it doesn’t guarantee robustness

against common signal distortion, whereas frequency domain method based on a multi resolution

wavelet decomposition, which shows greater robustness against such common attacks. We have

focused on wavelet based frequency domain watermarking method for our digital camera.

2.3.4 Advantage of Wavelet based Watermarking Techniques [48]

1) It avoids blocking artifacts so perceptibility of water marked image is very good.

2) Wavelet-based watermarking schemes are more robust under transmission.

3) They are better matched to the HVS (Human Visual System) characteristics.

4) It exploits the multi resolution property of wavelet very well. Hence an image can be shown

at different levels of resolution and can be sequentially processed from low resolution to high

resolution.

5) Visual artifacts introduced by wavelet coded images are less evident compared to DCT

because wavelet transmission doesn’t decompose the image into blocks for processing. At

higher compression ratios blocking artifacts are noticeable in DCT; however, in wavelet

coded images it is much clearer.

6) DFT and DCT are full frame transform, and hence any change in the transform coefficients

affects the entire image except if DCT is implemented using a block based approach.

However DWT has spatial frequency locality, which means if signal is embedded it will

affect the image locally. Hence a wavelet transform provides both frequency and spatial

description for an image.

2.4 Types of watermark attacks

Simple attacks (other possible names include waveform attacks or noise attacks) are

conceptually simple. They attempt to impair the embedded watermark by manipulating the

whole watermarked data (host data plus watermark), without trying to identify and isolate the

watermark. Examples include linear and general non-linear filtering, waveform-based

compression (JPEG, MPEG), addition of noise, addition of a cropping, quantization in the pixel

domain, conversion to analog, and correction.



Detection-disabling attacks (other names include synchronization attacks) try to break the

correlation and make the recovery of the watermark infeasible for a watermark detector. This is

done mostly by geometric distortion like zooming, shift in spatial or temporal (for video)

direction, rotation, shear, cropping, pixel permutations, sub sampling, removal or insertion of

pixels or pixel clusters, or any other geometric transformation of the data. A typical property of

this type of attacks is that the watermark remains in the attacked data. Typically, it can be

recovered with increased intelligence (and thus, complexity) of the watermark decoder.

Ambiguity attacks (other possible names include confusion attacks, deadlock attacks, inversion

attacks, fake-watermark attacks and fake-original attacks) attempt to confuse by producing fake

data. In an ambiguity attack, the attacker tries to fake a watermark and an object such that the

watermark is embedded in the alleged “original” object.

Removal attacks attempt to: a) analyze the watermarked data, b) estimate the watermark or the

host data, c) separate the watermarked data into host data and watermark, and d) discard only the

watermark. Examples are collusion attacks, denoising, certain nonlinear filter operations and

compression attacks using synthetic modeling of the image (e.g. using texture models or 3-D

models).

2.5 Common Attack Techniques

Additive noise: This may stem in certain applications from the use of digital to analog (D/A) and

analog to digital (A/D) converters or from transmission errors. However, an attacker may

introduce perceptually shaped noise (thus, imperceptible) with the maximum unnoticeable

power. This action will typically increase the threshold at which the correlation detector works.

Cropping : This is a very common attack since in many cases the attacker is interested in a small

portion of the watermarked object, such as parts of a certain picture or frames of a video

sequence. With this in mind, in order to survive, the watermark needs to be spread over the

dimensions where this attack takes place.

Rotation and scaling: This has been the true battle horse of digital watermarking, especially

because of its success with still images. Correlation-based detection and extraction fail when

rotation or scaling is performed on the watermarked image because the embedded watermark and

the locally generated version do not share the same spatial pattern anymore. An exhaustive



search with different rotation angles and scaling factors does yield the correlation peak, but it is

prohibitively complex.

2.5 Performance of Watermarked Image

a) Similarity Factor (SF): To assess the change in quality of image, we require correlating each

coefficient of original image as well as watermarked image.

( , ) ' ( , )

( , ) ( , ) ' ( , ) ' ( , )

w m n w m nS F

w m n w m n w m n w m n

(2.2)

Where, w (m, n) = original Image, w (m, n) = watermarked Image

It gives correlation of original images and watermarked images. It should be ideally 1.

b) Mean Square Error (MSE): To measure the credibility of the modified image, the mean

square error (MSE) between the original watermark and the extracted watermark can be used.

MSE is defined as

2

1

1( , ') ( ( ) '( ) )

wN

iw

M S E w w w i w iN

(2.3)

Where w = Original watermark, w’= Extracted watermark & Nw = Length of watermark. A

small value of the MSE (For example, a value less than 0.02) indicates the extracted watermark is

still credible. When the MSE value is near 0.5, we can consider that the extracted watermark is

totally uncorrelated to the original one.

c) Peak Signal to Noise Ratio (PSNR): It is a ratio of maximum signal power to noise power.

225510logPSNR

MSE

(2.4)

The PSNR is term for the ratio between the maximum possible power of a signal and the power

of corrupting noise that affects the fidelity of watermarked image. Because many signals have a

very wide dynamic range, PSNR is usually expressed in terms of the logarithmic decibel scale.

Higher the value of PSNR gives better quality of image. For color images with three RGB values



per pixel, the definition of PSNR is the same except the MSE is the sum over all squared value

differences divided by image size and by three. Typical values for the PSNR in lossy image and

video compression are between 30 and 50 dB, where higher is better.

2.6 Summary

Digital watermarking is a technique for inserting ownership information (watermark) into an

image, which can be detected later on for identification and authentication. Digital watermarking

has been used to address issue through the embedding of a secret signal in a digital image that

enables ownership. Watermarking is defined as the practice of imperceptibility altering a

multimedia work to hide a message into the host image. In order for a watermark to be useful, a

watermarking scheme should at least meet basic requirements such as transparency, robustness

and security. Watermarking provides authentication and ownership verification for secure

images taken by digital camera. For digital camera application, invisible and robust

watermarking techniques is required.


23

CHAPTER 3. WAVELET TRANSFORM

Wavelets are functions defined over a finite interval and having an average value of zero. The

basic idea of the wavelet transform is to represent any arbitrary function as a superposition of a

set of such wavelets or basis functions. These basis functions or baby wavelets are obtained from

a single prototype wavelet called the mother wavelet, by dilations or contractions (scaling) and

translations (shifts).

3.1 History of Frequency Domain Transformation

The discovery of first of DCT in 1974 is an important achievement for the research community

working on frequency domain transformation. The DCT can be regarded as a discrete-time

version of the fourier-cosine series. It is a close relative of DFT, a technique for converting a

signal into elementary frequency components. In 1992, JPEG [49] established the first

international standard for still image compression where the encoders and decoders are DCT-

based. The DCT-based encoder can be thought of as essentially compression of a stream of 8x8

blocks of image samples. Each 8x8 block makes its way through each processing step, and yields

output in compressed form into the data stream. Because adjacent image pixels are highly

correlated, the Forward DCT (FDCT) processing step lays the foundation for achieving data

compression by concentrating most of the signal in the lower spatial frequencies. For a typical

8x8 sample block from a typical source image, most of the spatial frequencies have zero or near-

zero amplitude and need not be encoded. In principle, the DCT introduces no loss to the source

image samples; it merely transforms them to a domain in which they can be more efficiently

encoded. Since some of the current image compression techniques are based on the wavelet

domain, such as JPEG2000, DWT based watermarking methods have been researched

intensively.

3.2 Discrete Wavelet Transform (DWT)

Despite all the advantages of JPEG compression schemes based on DCT namely simplicity,

satisfactory performance, and availability of special purpose hardware for implementation; there

are some shortcomings also. Since, the input image needs to be “blocked”; correlation across the


24

block boundaries is not eliminated. This results in noticeable and annoying “blocking artifacts”

particularly at low bit rates. Over the past several years, the wavelet transform has gained

widespread acceptance in signal processing in general and in image compression research in

particular. In many applications, wavelet-based schemes (also referred as sub band coding)

outperform other coding schemes like the one based on DCT. Since there is no need to block the

input image and its basis functions have variable length, wavelet coding schemes at higher

compression avoid blocking artifacts. Wavelet-based coding is more robust under transmission

and decoding errors, and also facilitates progressive transmission of images. In addition, they are

better matched to the HVS characteristics. Because of their inherent multi resolution nature [50],

wavelet coding schemes are especially suitable for applications where scalability and tolerable

degradation are important.

3.2.1 Sub band for Wavelet Based Coding: Over the years, there have been many efforts

leading to improved and efficient design of filter banks and subband coding techniques. Since

1990, methods very similar and closely related to sub band coding [51] have been proposed by

various researchers under the name of Wavelet Coding (WC) using filters specifically designed

for image compression purpose. Such filters must meet additional and often conflicting

requirements ([52], [53]). These include short impulse response of the analysis filters to preserve

the localization of image features as well as to have fast computation, short impulse response of

the synthesis filters to prevent spreading of artifacts (ringing around edges) resulting from

quantization errors, and linear phase of both types of filters since nonlinear phase introduce

unpleasant waveform distortions around edges. Orthogonality is another useful requirement since

orthogonal filters ([54], [55]) in addition to preservation of energy; implement a unitary

transform between the input and the sub bands. The concept Sub band coding is given in Fig 3.1.

Unlike in the case of 1-D, in two-band Finite Impulse Response (FIR) systems linear phase and

orthogonally are mutually exclusive, and so orthogonality is sacrificed to achieve linear phase.

The input is passed to the low pass filter L and high pass filter H as shown in below Fig 3.1.After

it is passed to filter it is down sampled with the sampling factor 2. Then again low pass filter and

high pass filter both are passed through the high pass filter and low pass filter followed by down

sampling by two to obtain the LL, LH, HL and HH separated bands. The input is passed through


25

the 1st level of decomposition procedure to have one approximation co-efficient (LL) and three

detailed co-efficient LH, HL and HH.

Fig 3.1 Sub band Coding

3.2.2 Watermarking with Wavelet Decomposition Level: With different combination of a low

pass and high pass filter, an image is decomposed into low-low (LL), low-high (LH), high-low

(HL) and high-high (HH) bands. To obtain the next scaled wavelet co-efficient, the band LL is

further decomposed and sub sampled. This process is repeated several times, which is

determined by requirement of the user. Furthermore, from these DWT coefficients, the original

image can be reconstructed. This reconstruction of process is called the inverse DWT (IDWT).

There are two basics requirements to embedded watermark for Secure Digital Camera that

watermark should be invisible and robust for various attacks [56]. To have invisibility we can

insert it in the high frequency band (HH) but it provides less robustness. To have robustness we

can embed watermark into the low frequency band (LL) but then of course; the invisibility is the

issue, hence there is always trade off between invisibility and robustness [57]. So optimum way

is to choose middle frequency bands (LH) and (HL) to insert watermark symbol. So there is

always tradeoff between robustness and invisibility while choosing any watermarking algorithm.

With increase the level of decomposition, the robustness will increase. This requires these issues

should be considered at the cost of additional hardware while dealing with watermarking

embedding algorithm. Three level decomposition is given in Fig 3.2

H

L2

L

H

H

L

2

2

2

2

2

LL

LH

HL

HH

I/P


26

Fig 3.2 Level Of Decomposition for Wavelet Transform.

3.3 Implementation of Wavelet Transform

DWT can be implemented either using convolution technique or lifting techniques. Previously,

DWT implementations were based on the so-called convolution technique (filter bank scheme

[58]), which computes the DWT of a signal by iterating a sequence of high and low pass filtering

steps, followed by down sampling. In 1997, Sweldens proposed a new scheme, called Lifting

Scheme (LS), as an alternative way to compute the DWT [59]. The LS has immediately obtained

a noteworthy success, as it provides several advantages with respect to the convolution scheme

(filter bank scheme.)

3.3.1 Relationship between Convolution and lifting scheme

Convolution-based filtering consists in performing a series of dot products between the two

filter masks and the extended 1-D signal. This implementation is non-polyphase and suffers from

inefficient hardware utility, low throughput and perfect reconstruction is not guaranteed.

Convolution scheme is same as filter bank scheme. We define matrices called Quadrate Mirror

Filter (QMF) in terms of filter coefficients and multiply by our data vector to produce a step of

wavelet transform. The method for computing the DWT coefficients used is matrix

multiplication. It is found that a large fraction of entries of matrices used in this method are

zeros.

HL2

HH1

LH1

HL1

HH2

LH22

LH3

HL3 HH3

LL3


27

Fig 3.3 Convolution based scheme

Clearly, this method is inefficient for computational purpose – it would be preferable to take

the advantage of the fact that the matrices are sparse. The factorization of wavelet transform

enables this, replacing the process of multiplication with the sequence of trivial operations

involving simple multiplications, addition and shifting of data vectors. This has the advantages

of being faster and easily invertible, and allows less calculation. The Lifting Scheme (LS) takes

advantage of the fact that there is a simple relation between all multiresolution analyses that

share the same filter (h). It has been shown to be equivalent to the traditional method of

implementing the wavelet transform via the QMF matrices [60]. It turns out that, the procedure is

equivalent to the factorization of a matrix whose components are related to the z transform of the

filter. Lifting-based filtering consists of a sequence of very simple filtering operations for which

alternately odd sample values of the signal are updated with a weighted sum of even sample

values, and even sample values are updated with a weighted sum of odd sample values. To get

smooth coefficient, lifting scheme is used. For smooth coefficient division by two is needed.

Lifting is more efficient polyphase alternative to convolution implementation. It requires fewer

coefficients (less memory) and has twice throughput. It enables efficient lossy and lossless

compression to be achieved within a common framework, even within a single compressed data

stream.

3.3.2 Advantage of lifting scheme

1) Lifting scheme is fast: For long filters, Lifting scheme has a complexity of order n/2,

compared with a complexity of order n for classical wavelet implementation.

Ho(n)

H1(n)

Go(n)

G1(n)

X(n) Yo(K)

Y1(K)


28

2) All operations within a lifting step can be done entirely in parallel, while the only sequential

part is the order of lifting operations.

3) Lifting can be done in-place; therefore an auxiliary memory is not needed. At every

summation point the new stream replaces the old one.

4) Maps integer to integer transformation thus eliminating the use of floating point operator.

Thus more efficient in term of hardware implementation

5) Process is reversible and lossless.

Table 3.1 Complexity comparison of convolution and lifting based implementation [61]

Two pixel 1- level 1-D DWT

Convolution Lifting scheme

Filter Multiplication Addition Multiplication Addition

5/3 4 6 2 4

9/7 9 14 6 8

3.3.3 Lifting based Scheme

Lifting scheme is fast DWT implementation scheme [62], which can be used to construct both

first and second-generation wavelets. The idea behind the Lifting scheme is to use half of the

data samples to predict the other half and repeat this until all the samples are predicted. The

algorithm consists of three simple steps, applied repetitively on the samples: Split phase, Predict

phase and Update phase, as shown in Fig 3.4.

Split phase: Assume that the scheme starts at level 0. We denote the data set a k,0 where k

represents the data element and 0 signifies the iteration level 0. In the first stage, the data set is

split into two other sets:

The even samples k,1 and the odd samples k,1 (see Fig. 3.4). This is also referred to as the

Lazy Wavelet transform because it does not de-correlate the data, but just sub-samples the signal

into even and odd samples. We use negative indices according to the convention that smaller the

data set, the smaller the index

kk 2,0,1 (3.1)

12,0,1 kk (3.2)


29

Fig 3.4 Split, Predict and Update Phase for lifting scheme

Predict Phase (dual lifting): The next step is to use the even sub-set k,1 to predict the odd sub-

set k,1 using a prediction function )( ,1 kP . The more correlation presented in the original data,

the closer will the predicted value be to the original k,1 . For odd set k,1 will be replaced by the

difference between itself and its predicted value. Thus,

1, 1, 1,( )k k kP (3.3)

Different functions can be used for prediction of odd samples. The easiest choice is to predict

that an odd sample is just equal to its neighboring even sample. This prediction method is result

to the Haar wavelet. Obviously, this is an easy but not realistic choice, as there is no reason why

the odd samples should be equal to the even ones. Alternatively, second or higher degree

interpolation functions can be used for prediction. Depending on the degree of the interpolating

function N, we can measure failure to predict . N is referred to as the number of dual vanishing

moments and defines the degree of the polynomials that can be predicted by the dual wavelet.

Update Phase (primal lifting): In this stage the coefficients k,1 are lifted with the help of the

neighboring wavelet coefficients , so that a certain scalar quantity Q, e.g. the mean, is

preserved.

)()( ,0,1 kk QQ (3.4)

A new operator U is introduced that ensures the preservation of this quantity:

Split Predict Update

+

-+

+

kj ,1

kj ,

kj ,

odd samples γ

even samples λ


30

1, 1, 1,( )k k kU (3.5)

Operator U uses a wavelet coefficient of the current level ),( kj to update N’ even samples of the

same level ),( kj . N is also known as the number of real vanishing moments, not necessary equal

to N.

Inverse Transform: One of the advantages of the lifting scheme is that the inverse transform is

very trivial. The inversion rules are: revert the order of the operations, invert the signs in the

lifting steps, and replace the split

Step by a Merge step:

1-Update phase , , ( , )( )j k j k j kU (3.6)

2-Predict phase , , ,( )j k j k j kP (3.7)

3-Merge phase 1,2 , 1,2 1 ,j k j k j k j k (3.8)

These three phases are repeatedly applied on the even samples, transforming half of the samples

each pass (level), until all samples are transformed.

2-D transform: 2-D transform is performed by applying the 1-D transform algorithm

consecutively on the rows and columns of a 2-D signal. Starting from the first iteration level, the

1-D forward transform is first applied to all the rows, and then to all the columns. Subsequently

we move to the next iteration level and repeat the mentioned steps, and so on.

There are different form of Wavelet transform function have been proposed for still images.

From those function there has been growing interest on reversible integer to integer wavelet

transform. This comes from the fact that it maps integer to integer. Indeed, such transforms are

invertible in finite precision arithmetic, and approximate the linear wavelet transform from

which they are derived. Another wavelet transform to obtain lossy transformation is proposed by

Daubechies, with greater compression ratio. The interest in Daubechies 9/7 filter comes from the

fact that 9/7 algorithm satisfy the condition required to obtain wavelet transform like it utilize bi-

orthogonal pair of filters which gives linear phase, symmetric output. Since lossless compression

requires no data loss due to rounding, a reversible wavelet transform using lifting scheme with

Le Gall 5/3 filters are used. The equation below describes the 5/3 DWT algorithm

implementation.


31

(2 ) (2 2)(2 1) (2 1)

2

x n x ny n x n

(3.9)

(2 1) (2 1) 2(2 ) (2 )

4

y n y ny n x n

(3.10)

Where x the input signal and y is is is the output signal. The 5/3 filter allows repetitive

encoding and de-coding of an image without any loss. Of course, this is true when the

decompressed image values are not clipped when they fall outside the full dynamic range (i.e., 0-

255 for an 8 b/p image). Traditional wavelet transform implementations require the whole image

to be buffered and the filtering operation to be performed in vertical and horizontal directions.

While filtering in the horizontal direction is very simple, filtering in the vertical direction is more

cumbersome. Filtering along a row requires one row to be read; filtering along a column requires

the whole image to be read. The line-based wavelet transform overcomes this difficulty,

providing exactly the same transform coefficients as the traditional wavelet transform

implementation.

Lossy compression allows for some data to be lost during the compression, allowing non-

reversible wavelet transformations. The irreversible (lossy) transformation is implemented by

means of the biorthogonal Daubechies 9-tap/7-tap filter. The equations below shows how the 9/7

tap Daubechies algorithm can be implemented in lifting scheme.

Forward transformation:

2 1 2 1 2 2 2( )n n n nY X X X

2 2 2 1 2 1( )n n n nY X Y Y

2 1 2 1 2 2 2( )n n n nY Y Y Y

2 2 2 1 2 1( )n n n nY Y Y Y

2 1 2 1n nY K Y

2 2 /n nY Y k (3.11)

Inverse transformation:

2 2n nX k Y


32

2 1 2 1 /n nX X k

2 12 2 2 1( )nn n nX X X X

2 22 1 2 1 2( )nn n nX X X X

2 12 2 2 1( )nn n nX X X X

2 1 2 1 2 2 2( )n n n nX X X X (3.12)

Where the values of constants are:

1.586134342 , 0.052980118 , 1.230174105 , 0.882911075 and

0.443506852k

Table 3.2 Daubenchies 9/7 analysis filter coefficient

Analysis Filter coefficient

Index Low-pass filter High-pass filter

0 0.602949 1.115087

1 0.262864 -0.059127

2 -0.078223 -0.057543

3 -0.016864 0.091271

4 0.026748

Table 3.3 Daubenchies 9/7 synthesis filter coefficient

Synthesis Filter coefficients

Index Low-pass filter High-pass filter

0 1.115087 0.602949

1 0.591271 -0.266864

2 -0.057543 -0.07822

3 -0.091271 -0.016864

4 0.026748


33

Table 3.4 Le Gall 5/3 analysis and synthesis filter coefficient

Analysis filter coefficient Synthesis filter coefficient

Index Low-pass High-pass Low-pass High-pass

0 6/8 1 1 6/8

1 2/8 -1/2 -1/2 2/8

2 -1/8 - - -1/8

3.4 Summary

There are two Lifting based wavelet implementation one is Le Gall 5/3 and another is

Daubenchies 9/7. Le Gall 5/3 based watermarking scheme has coefficients which can be

implemented very easily without loss of its values as all coefficients have values of in terms of

negative powers of 2. But for Daubenchies 9/7 based lifting scheme has co-efficient which can

be complex for the implementation. To implement hardware efficient wavelet based

watermarking scheme, it is obvious choice to have Le Gall 5/3 based lifting scheme due to

simpler & lossless hardware implementation.

FPGA Implementation of Watermarking Algorithm For Digital Camera


CHAPTER 4. DWT BASED WATERMARKING ALGORITHMS

We have implemented total three different wavelet based watermarking algorithm.

Algorithm 1: Wavelet Based watermarking with Random binary sequence (A. Lumini et al.

[39]).

Algorithm 2: Wavelet Based watermarking with Threshold scheme (You Fu et al [40]).

Algorithm 3: Wavelet Based watermarking with Block processing method (Hsien-Wen et al.

[41]).

All three algorithms are studied and results are compared against robustness, security and

transparency.

4.1 Algorithm 1 [39]

4.1.1 Watermark Embedding Method

Step 1: Let X is the original gray-level image with size p q . First, we apply 1st level wavelet

transform on X to obtain four different frequency bands such as approximation co-efficient,

detailed co-efficient, horizontal and vertical co-efficient band.

Step 2: Apply 2nd level Decomposition on approximation co-efficient band X2. Arrange the

approximation co-efficient in the descending format such that Maximum co-efficient comes first

and minimum coefficient last.

Step 3: Generate the watermark key of Y which is of 64 x 64 using the random number generator

to embed into X2.

Step 4: Add the watermark to the 2nd level approximation co-efficient with the secret key applied

for selection of coefficient.

Z= X2+ a *Y (4.1)

Where Z= Watermarked Image, X2= 2nd level of decomposition level of X.

Y= Generated watermarked key,

a = scaling factor=0.1

Step 5: Save the embedded position, sub band label, then, we take the two-dimensional

IDWT of Z.

Step 6: To simulate an attack on watermark image, we have added Gaussian noise, 1 degree anti

clockwise rotation, compression and rescaling to check robustness of watermarking scheme.



The advantage of this scheme is simpler in implementation and disadvantage is availability of

same random number generator at transmitter and receiver side.

4.1.2 Watermark Detection Method

Step 1: Perform Wavelet transform to get recovered image.

Step 2: Remove the watermark from approximation co-efficient exactly inverse process of

embedding process.

Step 3: Perform 2 levels Inverse DWT transform to obtain the original image co- efficient.

Step 4: Compare the original Image with recovered Image.

4.2 Algorithm 2 [40]

The block diagram of the proposed image watermarking system is shown in Fig.4.1 below.

Suppose an 8-bit gray-scale image I of size M × N to be protected by insertion of a binary

watermark w, further assume w is of the size M/4 × N/4 (i.e. the size of lower frequency sub band

coefficient after two level decomposition), if not, a M/4 × N/4 binary image can be constructed

from w in several ways, a natural way is to periodically tile w.

4.2.1 Watermark Embedding Method

Step 1:We first decompose the original image into two levels by using a discrete bi-orthogonal

spline wavelet transform.

Step 2: A secret key K is used to generate one random permutation vectors Rp and Cp of the row

and column vectors Row = [1,2,…, M/4 ] and Col =[1,2,…, N/4 ] of the watermark image W,

and permute the pixel positions of the watermark W by using Rp and Cp as follows

wp (Rp (i),Cp ( j)) = w(i, j) (4.2)

where, wp represents the permuted watermark.

Step 3: For robust watermark embedding, the watermark is embedded into the lower frequency

sub band after two levels decomposition. Let CLL be the lower frequency coefficients at a

resolution level 2 discrete wavelet transform with orientation LL and C*LL be the corresponding

watermarked sub band coefficient.



Fig 4.1 watermark Embedding and extraction steps

Step 4: The embedding algorithm can be depicted as follows:

Let d (i, j)= C (i, j)/Q( i, j)LL

If wp (i, j) ≠ mod(d (i, j),2) ,

C*LL (i, j) = (d(i, j) +1) *Q(i, j) ;

Else, C*LL (i, j) = CLL (i, j).

Where, Q is constant matrix for controlling the embedding watermark strength, and

└ ┘ is the operation to get the integer part of the number.

Step 5: Then, after the inverse DWT, the watermarked image can be acquired. The decision of

the matrix Q is a key problem for the robustness of the watermark. The greater Q, the watermark

is more robust. We can have two methods to decide the element of Q, one simple way is to let all

the element of Q be the same number Const; the other is an adaptive way according to the

corresponding coefficients of the detail sub band.

Q1 (i, j) = Const. i =1,..., M /4 , j =1,2,.., N/4 (4.3)

2 2

2 2

| ( , ) | | ( , ) |( , ) log (2 )

2HL LHC i j C i j

Q i j

i = 1,...,M/4 , j = 1,2,.., N/4 (4.4)

where C2HL and C2

LH are the detail coefficients at the horizontal and vertical direction of the

level 2nd wavelet decomposition respectively, and α is a parameter for controlling the strength of

the watermark embedding.

For later formula, when the image has complex texture at the place (i, j), the coefficient Q (i, j)

is large, then we can have a strong watermark embedding. Whereas when the Q (i, j) is small, the



image is flat at (i, j), so we can have a light watermark, where α is a parameter for controlling the

strength of watermark embedding space. Hence in such a way, we can get robustness but less

degradation for original image.

4.2.2 Watermark Extraction Method

For watermark extraction, the inverse operations are carried out.

Step 1: The watermarked image is transformed by a DWT into two levels.

Step 2: For the lower frequency sub band coefficients CLL , the quantization matrix by (4.3) or

(4.4) is firstly computed.

Step 3: For each element of the coefficients we do as follows:

Let d (i, j) = C (i, j)/Q( i, j)LL

w* (i, j) = mod (d (i, j), 2);

Step 4: After inverse random permutation, leading to w* = w, i.e., the right watermark is

extracted. To evaluate the watermark detector response, we use the similarity factor to correlate

two values.

The advantage of scheme is that it is energy saving scheme but it is difficult to adjust the

threshold with accurate value of Q matrix.

4.2.3 Effect of Scaling factor α and Quantization factor q:

We have been implemented the different values of α varies from 0.1 to 0.5 for threshold based

scheme suggested by Yong Fu et al [17].The parameter α represents a compromise between

resistance and image fidelity. The goal of maximizing the watermark robustness and minimizing

Fig 4.2 Variation in PSNR with respect to α for original and watermarked Images



its perceptibility is competing objectives. The parameter α can be turned in order to improve the

performance of one to another: In fact, by increasing α the robustness is improved, while by

decreasing α more perceptual fidelity to the original is achieved. The algorithm has been

checked for different images of gray scale as well as color images. The graph plot for α against

PSNR is shown in below Fig. 4.2.As you observe in the plot, as value of α increases the PSNR

is decreases.

The Fig 4.3 shown below is plot for values of scaling factor α against PSNR. As value of

scaling factor α increases the correlation between original and watermarked image is also

decreases.

Fig 4.3 Variation in similarity factor with respect to α for original and watermarked image

The Fig 4.4 shows the effect of quantization factor on PSNR. The quantization factor varies

from value 0.1 to 0.5.It signifies the threshold of the images for watermark embedding process. A

Fig 4.4 Variation in PSNR with respect to q for original and watermarked Images



different threshold should be chosen for each image in order to preserve the contribution due to the

coherent portion among different images. As value of quantization factor increases the value of

PSNR decreases as shown in plot below.

As shown below in Fig. 4.5, Similarity factor values decreases as value of q increases. The color

images have more similarity compare to gray scale image.

Fig 4.5 Variation in similarity factor with respect to q for original and watermarked image

4.3 Algorithm 3[41]

It is a based on integer wavelet transform scheme with reversible watermarking scheme. It uses

the block processing scheme with differential expansion scheme.

,2

x yl h x y

(4.5)Symbol └ ┘ is the floor function meaning “the greatest integer less than or equal to”.

The inverse transformation of [4.5] is,

1,

2 2

hhx l y l

(4.6)

As grayscale values are bounded in [0,255], we have

10 255

2

hl

and 0 255

2

hl

(4.7)

Which is equivalent tomin(2(255 ),2 1)h l l (4.8)



Then, the process of embedding watermarking using difference expansion is described asfollows:

4.3.1 The Embedding Method

Fig 4.6 Embedding Process

Assume that the original image is a grayscale image with size of 256 X 256 pixels.

a) Confirm the location of Embedding:

Step 1: We first split the original image which is to be watermarked, into many non-overlapping

small blocks with 8 X 8 pixels in a scan-line order. For convenience, let us assume that the image

has 256 gray levels and 256 X 256 pixels. Thus, we will get 1024 small blocks.

Step 2: For every block, it is decomposed by 1-level IWT (Integer Wavelet Transform), and

produces a low frequency sub-band LL1, which is a 4x4 matrix, as shown in Fig. 4.7

Step 3: If the pair of values (b1, b6) and (b11, b16) all are expandable, then we use ‘1’ to stand

for the attribute, else ‘0’ is used to represent the attribute.

b1

b6

b11

b16

Fig 4.7 The pixel array graph of the block

OriginalImage

Embed watermarkdata into

expansible block

Watermark data

Divide into 8X8Blocks

IntegerWavelet

Transform

CheckExpansibility&generate Key

Watermarked Image

256X256



Definition of Expansibility: For a grayscale-valued pair, (x, y) we say its difference number is

expandable if it satisfies the following condition, such that

0 2 min(2(255 ),2 1)h b l l (4.9)

Where,

,2

x yl h x y

(4.10)

Symbol └ ┘ is the floor function meaning “the greatest integer less than or equal to”, For both

b=0 and b=1.

Step 4: According to expansibility of each block a 1024 bit secret key is generated composed by

‘0’ and ‘1’.

b) Embed Watermarking

Step 1: Watermark is embedded in the corresponding block where key is 1. Calculating the

average li and difference hi for the ith pair of pixels, denoted as (xi, yi) by using [4.4] for both the

pair of pixels of ith block

Step 2: Calculating, hi’ =2hi+ wi , where wi is the corresponding bit of watermark.

Step 3: Calculating new ith pair of pixels (xi’,yi’ ) after embedding watermark by the following

with equation [4.7] and [4.8].

Step 4: Repeat the former steps until both pixel pairs of each ith block.

Step 5: Construct new coefficients using new (xi’,yi’).

Step 6: Find inverse IWT of each block and rearrange 8x8 blocks to get watermark image.

4.3.2 Watermark Extraction Method

The extraction procedure can be gotten from the procedure of watermarking embedding. The

block diagram is shown below.

The steps are given below:

Step 1: The watermarked image is firstly split into 8 X 8 many non-overlapping small blocks

with pixels in a scan-line order.

Step 2: According to secret key, the expansible blocks is decomposed by 1-level wavelet lifting

scheme, for pair of pixels (b1,b6) and (b11,b16) of low frequency sub-band LL1, the following

procedures are adopted to extract watermarking extraction:



Fig 4.8 watermarking extracting process

1. Calculate the average and difference of the pair of pixels.

' '' , ' ' '

2

x yl h x y

(4.11)

2. Extract watermarking:

0

'( '),

2

hw LSB h h

(4.12)

3. Restore the original pixel values back using [4.8]

0 00 0

1' , '

2 2

h hx l y l

(4.13)After the restoring the original pixel finds the inverse IWT to the subtracted image, which is

similar to the original image. The result according to this algorithm is shown in Fig. 4.8.

4.4 Proposed Watermarking Algorithm

Spatial-domain digital watermarking methods are generally considered as having poor

performance after geometric distortion (such as cropping and scaling), common signal processing

(such as JPEG and filtering), or subterfuge attacks (such as changing the least-significant bits).It

has less computational cost due to their easy operation. So it is more efficient. On the other hand,

frequency-domain watermark techniques, have their high computation complexity, and provide

great robustness to different attack. However, we have adopted the combined approach with

spatial-frequency domain approach which has advantages of both domains. The frequency domain

Using DifferenceExpansion and

According to key

IWTSplit into 8X8block

WatermarkedImage

Watermarkeddata

SubtractedImage



transformation is done with lifting based wavelet scheme and spatial domain transformation done

with bit plane slicing.

(a) Original Image (b)1st level decomposed Image

Fig 4.9 Wavelet Decompostion

These operations correspond to the following filtering processes:

Top left: 2-D low pass filter (LL).

Top right: horizontal high pass and vertical low pass filter (HL).

Lower left: horizontal low pass and vertical high pass filter (LH).

Lower right: 2-D high pass filter (HH).

To apply this transform to a complete image, we group the pixels into different blocks and

apply DWT transformation to each block. To view the result, all the top left components in Fig

4.9 (a) of the different blocks were grouped together to form the top left sub image in Fig 4.9 (b)

and the same for the components in the other three positions. It is clear from Fig 4.9 (b) that the

most of the energy is contained in the top left (LL) sub image and the least energy is in the lower

right (HH) sub image. The top right (HL) and the lower left (LH) sub image contains the edges.

The energies of LL Band (88-89%), HL (4-5%), LH (6-7%) and HH (1-2%). So the LL band

contains most part of energy while HH band contains marginal part of the total energy of image.

4.4.1 Bit-plane slicing Method

Instead of highlighting gray level images, highlighting the contribution made to total image

appearance by specific bits might be desired. Suppose that each pixel in an image is represented

by 8 bits. Imagine the image is composed of 8, 1-bit planes ranging from bit plane 0 (LSB) to bit

plane 7 (MSB). In terms of 8-bit bytes, plane 0 contains all lowest order bits in the bytes

comprising the pixels in the image and plane 7 contains all high order bits. Separating a digital

image into its bit planes is useful for analyzing the relative importance played by each bit of the

image, implying, it determines the adequacy of numbers of bits used to quantize each pixel,



useful for image compression. In terms of bit-plane extraction for a 8-bit image, it is seen that

binary image for bit plane 7 is obtained by proceeding the input image with a threshold gray-

level transformation function that maps all levels between 0 and 127 to one level (e.g. 0)and

maps all levels from 129 to 253 to another (eg. 255).

a) Watermark Embedding Method:

Step 1: Read the original Image X.

Step 2: Compute the first level wavelet decomposition of original image X. We obtain one

approximation image (LL Band) and three detail images, corresponding to the horizontal,

vertical and diagonal details of the original host image.

Step 3: Convert the LL band of original image into different 8 planes start with MSB to LSB.

Step 4: Generate the required watermark of text, image or random number.

Step 5: Add the LSB plane of high frequencies with desired watermark plane.

Step 6: Reconstruct the image with adding all the 8 planes again with MSB to LSB.

Step 7: Apply 1st level inverse to LL band and three detailed co-efficient to obtain the

watermarked image.

Step 8: Apply different attacks to the watermarked image to check the robustness.

Fig 4.10 Bit Plane Slicing Scheme to LL Band

b) Watermark Extraction process

Step 1: Apply DWT to watermarked image to get approximation and details co-efficient.

Step 2: Apply bit plane slicing to the LL of the watermarked image to get 8 different planes

start from LSB to MSB.



Original Image 1st Level Decomposition

Watermarked Image Reconstruction from plane Watermarked LSB Plane

(a) Watermark Embedding Process

Watermarked Image Watermarked LSB Plane

Original Image LSB plane of original Image Watermark(b) Watermark Extraction process

Fig 4.11 Explanation of Proposed Algorithm step by step

Step 3: Apply DWT to the original image to have four sub bands of LL, LH, HL and HH.

+

IDWT

Watermark

-

DWT

DWT

Bit Plane Slicing

A

A



Step 4: Apply bit plane slicing to LL band of original image to get different planes from LSB

to MSB.

Step 5: Subtract the LSB plane of the original image from the LSB plane of watermarked

image to get watermark.

4.5 Summary

Spatial domain watermarking has advantage of less computational cost. Frequency domain

watermarking provides more robustness. The proposed algorithm has been developed to take

advantage of both spatial as well as frequency domain properties. The proposed algorithm

has been verified on MATLAB platforms for various attacks. It has comparable PSNR and

MSE results compare to other algorithms.



CHAPTER 5. IMPLEMANTATION OF PROPOSED SCHEME

We have implemented the proposed watermarking algorithm through HDL implementation.

We have targeted algorithm to implement on FPGA kit. The complete flow for proposed scheme

is shown below in Fig 5.1. The design of DWT,IDWT and watermarking is done through VHDL

based RTL coding. The functional simulation of RTL codes has been checked through Modelsim

simulator. The synthesis results are obtained in ISE 9.1 of Xilinx. The area, power and speed

performance has been verified and also generate .net list which is useful to download in FPGA

through design vision of Synopsys.

The implementation flow for proposed scheme is shown in Fig 5.2.The image is read through

DesignSpecification

Design Entry

Function simulation

Synthesis

Post Synthesis Simulation

Translate, Place & RouteMap

Target FPGA

I/p : -----------------0/p: .vhd file

I/p : .vhd and library0/p: .vsim or .sim file

I/p : .vhd and technology library andsymbol library0/p: net list .v file

I/p : net list .v file andVerilog Technology library0/p: .vsim or .sim file

HDL Code

Modelsim

ISE and Design Compiler

Modelsim and VCS

Area, Power and Timingreports

Fig 5.1 VLSI Flow for Proposed Scheme



Fig 5.2 Implementation Flow for Proposed Scheme

MATLAB and pixel is stored in datain.txt file. With help of text I/O package the datain.txt file has

read in VHDL and Legall 5/3 based Lifting wavelet applied to obtain transform domain co-

MATLABImage Read

Function

TextInput/Output

Functionto read

Pixel input

LiftingBasedDWT

TransformDomain

Co-efficientMatrix

Separate LL bandMemory to embed

watermark

Bit Plane SlicingMethod

WatermarkEmbedding

process in LSB

Random Number Generator forCo-efficient key and watermarkGeneration

Reconstruction ofPlane to obtain LL

Band ofwatermarked

image

Original Image

LiftingBasedIDWT

Watermarked Image

TextInput/output

Functionto write

out put Pixel

MATLABImage Write

Function

MATLABImplementation

MATLABImplementation

HDL Implementation



efficient matrix. LL band coefficient stored in separate memory to embed watermark. The RTL

code of bit Plane slicing has been developed to separate different planes from LSB to MSB. To

the selected co-efficient generated by random number generator, then watermark has embedded

to them. Then all planes are reconstructed with bit plane slicing RTL code to obtain LL band of

watermarked image. Then lifting based legall 5/3 IDWT has been applied to obtain pixel values.

The MATLAB function is used to construct watermarked image.

Watermarked Image Original Image

Fig 5.3 Watermark Detection Process

The detection process is very simple. We just use MATLAB and Text I/O function to read

original image and watermarked image. The DWT based Legall 5/3 lifting scheme is applied and

bit plane slicing RTL applied to obtain LSB of watermarked and original image. Then after,

Bit plane Slice toObtain LSB Plane

MATLAB & TextInput/Output

Read Function

DWT Based Lifting

MATLAB & TextInput/Output

Read Function

DWT Based Lifting

Bit plane Slice toObtain LSB Plane

-

Watermark

Co-efficient Selection Key



subtract these two LSB planes and watermark is extracted according to selection key generated

by random number generator.

5.1 Hardware Implementation Lifting based DWT Scheme

For implementation of hardware efficient DWT based scheme, lifting based scheme obviously

far better than traditional convolution scheme.

1) It requires less operations than the filter bank scheme, with a saving of up to one half for very

long filters;

2) It allows to compute an integer wavelet transform (IWT), that is, a wavelet transform that

maps integers to integers, thus enabling the design of embedded lossless and lossy image

encoders .

3) In terms of FPGA implementation, it uses fewer resources and can have pipeline stages

enabling higher operating speed and lower power consumption.

Lifting based wavelet scheme used in various approaches like Daubechies 9/7 and Le Gall 5/3.

But Le Gall 5/3 is proven more hardware efficient due to its simplicity and lossless

implementation. Since lossless compression requires no data loss due to rounding, a reversible

wavelet transform using lifting scheme with Le Gall 5/3 filters are used. The equation below

describes the 5/3 DWT algorithm implementation. The odd and even samples values can be

calculated by,

2

)22()2()12()12(

nxnxnxny extext

ext (5.1)

4

2)12()12()2()2(

nynynxny ext (5.2)

5.1.1 Implementation of (lossless) Le gall 5/3 algorithm

To implement this algorithm, the equations stated earlier are utilized. In lifting scheme this

algorithm is divided in two phases: predict phase and update phase. To find the value for predict

phase, at a time three inputs are required. Similarly in update phase one even input and two

values obtained form predict phase is needed. The structure in Fig 5.4 shows the 5/3

implementation.



Fig. 5.4 Lifting architecture for 5/3 wavelet

From the equations 5.1 and 5.2, it is found that odd and even coefficient can be easily calculated

using only shifting algorithm in VHDL. To save the memory during computation, once three

inputs are fetched from the memory, next time only two inputs are needed for calculation. So,

previous two inputs memory locations are utilized to store the values of predict phase (odd

coefficients). Hence, during computation total three memory locations are required. Fig. 5.5

shows how the coefficients are calculated with input pixels.

Fig 5.5 Single scale one dimensional wavelet lifting (5-3 lossless wavelet)

To perform 5/3 DWT 8 bit-gray scale image of LENA (256 X 256) is used. To implement the

predict phase and update phase necessary steps are shown below:

SplitPhase

PredictPhase

UpdatePhase

Xodd

Xeven

+

+

HP

LP

I/p value X



1. Split the input into odd and even positions.

2. perform predict step

3. perform update step

The architecture module of predict phase & update phase are shown below.

Fig 5.6 Predict phase

Fig 5.7 Update phase

The whole w-bit 1D-DWT structure, constructed according to lifting scheme is shown below in

Fig. 5.8. To process the image such a two one dimensional DWT unit for horizontal and vertical

transforms are required. To minimize the area, once the horizontal transform is completed same

Adder

ShiftReg

Adder

Y(2n + 1)

Y (2n ) Y(2n)

Y(2n - 1)

Adder

ShiftReg

subtractor

X(2n)

X(2n + 1)

X(2n + 2) Y(2n+1)



1D- DWT unit is utilized for vertical transformation. The coefficients computed in this way are

stored in internal memory.

Fig 5.8 1-D DWT unit

5.1.2 Memory management

As suggested, to obtain the forward wavelet transform, initially we need to read the three data.

From these three input data we get two coefficients detail (high) and approximate (low). We

need to manipulate these coefficients to display the image correctly at output. In VHDL, two

dimensional metric is not synthesizable. If we read the image having size n X n, then total n ^2

memory location are required to store each input pixels. For this thesis, 256 X 256 gray scale

image is utilized. So total memory location to store the input data are 65,536.Similarily total

65,536 memory locations are needed to store wavelet coefficient. Thus total memory locations

are 1, 31,072. Therefore the address size is calculated as 256 X 256 X 2 = 217 (17 bit address

size), and each data width is 8 bit as each pixel have size of 8 bits.

In 2-D wavelet transform, data is processed row wise and then after column wise. So during

the 1-D wavelet transform on rows, calculated detail coefficient should be placed at n/2 locations

form average coefficient. This mapping is shown in Fig 5.9. To store these coefficients also

single port RAM is designed in the FPGA. So before initialing the transformation, control offset

must be calculated from the size of image.

Thus to process the image, all rows are transferred to the FPGA from the SDRAM. The

coefficients from the corresponding to rows of the image itself are stored into single port RAM.



Fig 5.9 Memory mapping (a) input pixels (b) after forward DWT transformation

Here is short description of the VHDL design unit. The design was described in number of

entities as described below:

Legall_5_3_forward and Legall_5_3_inverse: These entity –architecture pairs implement

the whole 5/3 forward DWT transformation and inverse DWT transformation

respectively.

2 Dimensional DWT and IDWT: These entity –architecture pairs implement the 2-D

forward DWT transformation and inverse DWT transformation respectively. They are

utilized as a component in Legall_5_3_forward and Legall_5_3_inverse programs.

1 Dimensional DWT an IDWT: These entity-architecture pairs implement the 1-D DWT

transformation in forward and inverse direction respectively. They are utilized as a

component in 1 Dimensional DWT an IDWT programs.

The flow chart for 2 Dimensional forward wavelet transformations is shown in the Fig 5.10.

First we load the total number of decomposition level N to initiate the process. Afterwards,

counter is incremented until it reaches the specified value of row counter R and column counter

C. First forward transformation on row has been calculated and then same procedure is

applicable on coloumn. After complete process of transformation on row and coloumn the load

value of decomposition level is each time decremented and same steps are repeated as above

follows. The same way inverse wavelet transformation is performed as shown in the Fig 5.11.

Every time, Level of decomposition is decremented once row and coloumn inverse

transformation is completed.

0 256

65,53565,280

65,536

128

65,664 65,792

1, 31,0721,30 ,816

L H



2D Wavelet Transform

Load = N

X = 0

X = X + 1 Forward 1-D TransformOn rows

R < X ?

X = 0

X = X + 1

C < X ?

Decrement load by one

Load > 0

STOP

Forward 1-D TransformOn rows

YES

NO

START

Fig 5.10 Flowchart to calculate the 2D Forward wavelet Transform



Fig 5.11 Flowchart to calculate the 2D Inverse wavelet Transform

5.2 Watermark Embedding Hardware Implementation

There are two basic blocks required for watermark embedding process.

1) Bit plane Slicing scheme Implementation

2) Random Number Generator for key selection and watermark generation



5.2.1 Bit Plane Slicing Implementation

It is the spatial domain techniques. In Spatial domain scheme, watermark is directly embedding

into the pixel values. Algorithm splits the image into 8 planes from MSB to LSB. The whole

concept is explained in details as shown in below Fig 5.12.

Consider the case; following co-efficients are stored in memory after transformation is applied.

01111001 01100101 01001010 0010011010000100 10000110 10001001 10001101

The values read from memory are taken one by one in temp variable.

Fig 5.12 Bit Plane Slice Method

To separate out the values in different planes, we will xor the temp values with standard values

as follows. Here first value read from memory and stored in temp is 01111001.

LSB plane : 01111001 and 00000001 : The resultant values is 00000001. Seventh plane: 01111001 and 00000010 : The resultant values is 00000000. Sixth plane : 01111001 and 00000100 : The resultant values is 00000000. Fifth plane : 01111001 and 00001000 : The resultant values is 00001000. Fourth plane : 01111001 and 00010000 : The resultant values is 00010000. Third plane : 01111001 and 00100000 : The resultant values is 00100000. Second plane : 01111001 and 01000000 : The resultant values is 01000000. MSB plane : 01111001 and 10000000 : The resultant values is 00000000.

Next time another value read from memory and stored in temp as 01100101. and same procedure

of and operation with constants is follows as above. This way with all LSB plane of all the co-

efficient are obtained. The reconstruction of the planes are also very simple. We have to just add

all the resultant values plane s to obtain original value. With addition of all plane obtained

values:

Bit plane 7(MSB)

Bit plane 1(LSB)

One 8 bit byte



00000001 + 00000000 + 00000000 + 00001000 + 00010000 + 00100000 + 01000000 +

00000000 = 01111001

The LSB plane and plane next to LSB contains practically no visually significant information. LSB

manipulation gives an enormous amount of capacity in LSB plane for hiding information.

5.2.2 Watermark Generation and Selection Unit

We have used invisible and robust watermarking algorithm that insert the pseudorandom

numbers that insert pseudorandom numbers to host data. Therefore, we focused on structural

design aspect of watermarking generators using LFSRs. The binary watermark is generated by a

pseudorandom sequence generator. The watermark generation unit consists of an LFSR. The

LFSR has a multitude of uses in digital system design and is a very crucial unit in watermark

security and detection. It is a sequential and detection. It is a shift register with a combinational

feedback logic around it that causes it to cycle pseudo randomly through a sequence of binary

values. The LFSR consists of flip-flops (FFs) as sequential elements with feedback loops. The

feedback around an LFSR comes from a selected set of points called taps in FF chain; these taps

are fed back to the FFs after either xor, and, nor operation. Here an 8 bit LFSR is modeled so as

to use a one to many feedback structure and has been modified for a 2n looping sequence. It

calculates and holds the next values of the LFSR, which is then assigned to output signal.

Fig 5.13 Random Number generation scheme

It is one of the important blocks of watermarking process. It is used to generate the co-efficient

selection key and as well as watermark generation. As shown in Fig 5.13, 8 bit D-flip flop is



used so the maximum number of co-efficient selection key from random number generator is

255. The watermark is added according to key generated at the output. The whole process is

explained in below Fig 5.13. The random number generator is started with secret key provided to

its initial state. We have used as secret key 10101101 is the initial seed to start random number

generation. The same key has been used with same random number at the detection side.

5.3 VLSI Implementation for Proposed Scheme

5.3.1 Architecture of Proposed scheme

Main Memory

Fig 5.14 architecture for proposed scheme

AdderSubtractor

ShiftRegister

TemporaryStorage

Adder

ShiftRegister

Adder

RAM 1

ExternalRAM

LL BandStorage

LFSRFlip Flops

LogicalAND

WatermarkAdder

RAM 2



The architecture design for proposed scheme is defined as shown in Fig 5.14.The main memory

comprises of 2 times larger memory for size of original image. For size of 256 x256 image the

main memory requirement is 2*256*256=1,31,072. The memory is divided into two parts as

RAM1 as one for original image and second as RAM2 for watermarked image. At time of

detection in non blind scheme, RAM1 values are considered as original pixels values and RAM2

is watermarked values. At start three values are read from RAM for lifting based wavelet scheme

with legal 5/3 method as explained in section 5.1.

5.3.2 Pin Diagram

The pin diagram for wavelet based spatial domain watermarking chip is shown in the Fig 5.15.

The function of each pin is also explained in details.

Fig 5.15 Pin Diagram

The functional description of each pin is as follow:

Data In [7:0] : DATA Input bus. Original pixel value which was stored will be input on this bus

for operation.

CLK : Clock signal to chip.

Wmcontrol: Enables during embedding the watermark.

RESETZ: It is active low signal to reset the chip

START: It is an active low handshake signal to initiate data transfer operation on Data In

bus on every clock edge.

Data out [7:0]: DATA output bus. Watermarked pixel values are output on this bus.

Data In [7:0] Wavelet based

Spatial Domain

watermarking

chip

clk

wmcontrol

RESETZ

START

Data out [7:0]

READY

BUSY



READY: It is active high signal will be activated for one CLK cycle after the completion of

watermark embedding operation. It indicates that Data out bus has valid out on

bus.

BUSY: It is a active high signal. It indicates that watermarking process in progress. When

signal is active high, external access to RAM1’s are isolated. The data on data out

bus is not valid.

5.4 Summary

The proposed scheme uses wavelet based lifting scheme with legal 5/3 transformation to obtain

frequency domain co-efficient. The proposed scheme requires bit plane slicing hardware to

separate out 8 different Planes from LSB to MSB. We embed the watermark with random binary

watermark in LSB plane according co-efficient selection key.



CHAPTER 6. RESULTS ANALYSISWe have implemented different DWT based watermarking algorithms to check basic properties of

transparency, robustness and security of watermarking. We have used input image of 256 x 256

size of gray scale as well as color image. We have compared all algorithms for PSNR, MSE and

Similarity Factor results and also concluded the most suitable watermarking scheme for hardware

implementation. We have also measured the effect of quantization factor q and scaling factor α on

different watermarking schemes. We have used color images to prove watermarking scheme for

future generation secure digital camera.

6.1 MATLAB RESULTS

The watermarking algorithms are simulated for following gray scale original images.

Lena (256X256) (greyscale, 8 bpp)

Baboon (256X256) (greyscale, 8 bpp)

Cameraman (256X256) (greyscale, 8 bpp)

Transparency:

Table 6.1 Comparison of Original Image and Watermarked Image without any attack

Parameter A. Luminiet al.[39]

Yong Fu et al[40] Hsien-WenTseng et al [41]

Proposedscheme

Scheme 1(α) Scheme 2(q)

LenaPSNR 50.7988 67.55 46.154 48.1513 60.000MSE 0.541 0.01143 1.5621 0.9953 0.065SF 0.9950 0.99928 0.99181 0.9925 0.996

BaboonPSNR 49.4857 66.063 46.193 47.7169 59.55MSE 0.7320 0.016097 1.5625 1.1000 0.072SF 0.9947 0.99914 0.9913 0.9917 0.995

CameramanPSNR 51.3804 68.149 46.193 48.2631 61.14MSE 0.4732 0.039832 1.5625 0.9700 0.050SF 0.9954 0.99942 0.99109 0.9932 0.998

As shown in Table 6.1, it has been concluded that Yong Fu et al. [17] algorithm for threshold

method with scheme 1 has provided better result compare to others. The proposed scheme has also



comparable results with higher PSNR values. Hsien Wen [18] algorithm which uses block

processing approach has less efficient because of artifacts.

Robustness:

To check the robustness of the algorithm, we have verified the algorithms with Gaussian noise

attack, JPEG compression and 1 degree anticlockwise rotation

Table 6.2 Comparison of original Image and Watermarked Image with Gaussian noise attack



Proposedscheme


LenaPSNR 20.068 20.079 20.029 20.014 20.071MSE 640.21 638.55 645.86 648.10 639.73SF 0.9910 0.99965 0.99348 0.9903 0.998



Table 6.3 Comparison of Original and Watermarked Image with 1 degree anticlockwise rotation



Proposedscheme


LenaPSNR 21.927 22.852 22.834 21.032 22.822MSE 417.20 337.19 338.63 512.70 339.54SF 0.9908 0.98581 0.99401 0.9901 0.9985





Table 6.4 Comparison of Original Image and Watermarked Image with JPEG Compression



ProposedschemeScheme 1(α) Scheme 2(q)

JPEG Compression 10%Lena

PSNR 41.943 42.879 41.758 41.794 42.097MSE 4.157 3.3512 4.3384 4.302 4.012SF 0.9997 0.99998 0.99109 0.9932 0.9998




PSNR 39.137 39.751 39.002 38.997 39.034MSE 7.932 7.1768 8.1822 8.23 8.123SF 0.9989 0.99979 0.99303 0.9947 0.9998




PSNR 37.752 37.883 36.17 36.364 36.225MSE 10.912 10.588 15.706 15.02 15.51SF 0.9989 0.99959 0.9920 0.9938 0.9993





From Table 6.2, 6.3 and 6.4, it has been concluded that proposed scheme has better robustness

compare to other schemes. The proposed scheme utilizes the advantages of wavelet domain and

spatial domain concept which will improve the robustness with lesser computational complexity.

Security:

Table 6.5 Comparison of Original and Extracted Watermarked without attack


Yong Fu et al[40] Hsien-WenTseng et al

[41]

ProposedSchemeScheme 1(α) Scheme 2(q)

LenaText Logo 0.9982 0.99931 0.99136 0.9920 0.9989Randomnumber

0.9985 0.99934 0.99138 0.9922 0.9991

BaboonText Logo 0.9972 0.99893 0.99113 0.9913 0.9985Randomnumber

0.9974 0.99895 0.99116 0.9916 0.9988

CameramanText Logo 0.9987 0.99947 0.99184 0.9925 0.9993Randomnumber

0.9989 0.99952 0.99188 0.9929 0.9996

Table 6.6 Comparison of Original and Extracted Watermark with 1 degree anti clock rotation



[41]

ProposedScheme



0.9978 0.99976 0.99140 0.9834 0.9989


0.9975 0.99910 0.99127 0.9025 0.9983


0.9979 0.99968 0.99137 0.9927 0.9991



Table 6.7 Comparison of Original and Extracted watermark with JPEG Compression



[41]

ProposedScheme



Text Logo 0.9970 0.99972 0.99542 0.9798 0.9979Randomnumber

0.9973 0.99974 0.99545 0.9801 0.9981


0.9955 0.99969 0.99538 0.9790 0.9965


0.9979 0.99982 0.999549 0.9805 0.9985



0.9965 0.99969 0.99457 0.9795 0.9975


0.9948 0.99959 0.99447 0.9782 0.9958


0.9973 0.99973 0.99464 0.9796 0.9977



0.9953 0.99957 0.99938 0.9783 0.9964


0.9937 0.99952 0.99635 0.9768 0.9945


0.9959 0.99963 0.99648 0.9783 0.9969



Table 6.8 Comparison of Original and Extracted Watermark with Gaussian noise



[41]

ProposedScheme



0.9972 0.99985 0.9966 0.9835 0.9990


0.9968 0.99975 0.9962 0.9522 0.9987


0.9976 0.99989 0.9975 0.9884 0.9992

We have used Text Logo and random number as watermark to all algorithms. From 6.6, 6.7 and 6.8,

it has been concluded that the proposed scheme has almost Similarity Factor 0.99 in all the cases.

Higher values of Similarity Factor show that extracted watermark is exactly similar to the original

watermark.

Color Images:

All the Tables 6.9-6.16 suggest that proposed scheme has better results for color images also.

Table 6.9 Comparison of Original Image and Watermarked Image without any attack




LenaPSNR 51.0551 68.067 46.154 49.5346 61.21MSE 0.510 0.010288 1.5621 0.7238 0.0492SF 0.9952 0.99929 0.99061 0.9935 0.997





Table 6.10 Comparison of Original Image and Watermarked Image with Gaussian noise attack




LenaPSNR 20.036 20.044 20.018 19.993 20.041MSE 644.90 643.72 647.52 651.32 644.10SF 0.9908 0.99871 0.99308 0.9900 0.997



Table 6.11 Comparison of Original and Watermarked Image with 1 degree anti clock rotation



Proposedscheme


LenaPSNR 22.331 23.514 0.9953 21.405 23.499MSE 380.15 289.51 290.97 470.5 290.5SF 0.9937 0.98661 0.9953 0.9915 0.9990



We have separate out RGB components from color images and convert it into YCbCr

component. Then luminance component of color image is processed for watermark embedding

process. This Y component has saved as gray image and all the attacks applied to luminance

component of color images. Table 6.9-6.12 shows the comparison of different images for all

algorithms. All the parameters like MSE, PSNR and Similarity Factor has calculated for color

images. The results shows that color images have even better results compare to gray level

images.






[41]

Proposedscheme



PSNR 45.768 46.228 44.062 44.474 46.004MSE 1.7230 1.5499 2.5519 2.321 1.632SF 0.9980 0.9997 0.9925 0.9935 0.9988




PSNR 42.425 42.9 41.755 42.087 42.789MSE 3.72 3.3347 4.34 4.021 3.421SF 0.9989 0.99984 0.99248 0.9935 0.9990




PSNR 40.868 41.084 40.229 40.407 41.123MSE 5.325 5.1347 6.169 5.921 5.021SF 0.9998 0.9999 0.9925 0.9940 0.99985





Table 6.13 Comparison of Original and Extracted Watermarked without attack



ProposedScheme


LenaText Logo 0.9985 0.99933 0.99139 0.9922 0.9992

Random number 0.9987 0.99937 0.99140 0.9925 0.9993Baboon

Text Logo 0.9974 0.99894 0.99115 0.9914 0.9987Random number 0.9975 0.99897 0.99117 0.9918 0.9989

CameramanText Logo 0.9988 0.99949 0.99187 0.9927 0.9996

Random number 0.9991 0.99955 0.99190 0.9932 0.9997

Table 6.14 Comparison of Original and Extracted Watermark with Gaussian noise



[41]

ProposedScheme


LenaText Logo 0.9971 0.99987 0.9969 0.9827 0.9986




Random number 0.9979 0.99990 0.9979 0.9891 0.9993

Table 6.15 Comparison of Original and Extracted Watermark with 1 degree anti clock rotation



[41]

ProposedScheme


LenaText Logo 0.9982 0.99934 0.99032 0.9832 0.9985




Random number 0.9987 0.99972 0.99086 0.9020 0.9992






[41]

ProposedScheme




BaboonText Logo 0.9956 0.99966 0.99536 0.9791 0.9967

Random number 0.9957 0.99972 0.99540 0.9793 0.9969Cameraman




BaboonText Logo 0.9946 0.99954 0.99448 0.9781 0.9958





BaboonText Logo 0.9942 0.99951 0.99444 0.9778 0.9954



The Fig 6.1 shows the comparison of PSNR against all algorithms for gray level and color

images. The algorithm 2 with scheme 1 of scaling factor α has much better results compare to

other algorithms. But it suffers from larger computational complexity. In comparison with that,

the proposed scheme has lesser computational cost and has comparable results for PSNR as

shown in Fig 6.1.



Fig 6.1 Comparison of PSNR for all algorithms for gray and color images

The Fig 6.2 also proves that proposed algorithm has better correlation for original and extracted

watermark.

Fig 6.2 Comparison of Similarity Factor for all algorithms for gray and color images

Down sampling of Color Image:

The complete process of down–up sampling of Lena 256 x 256 color image is shown below Fig

6.3. First color images which comprises of RGB components are converted to YCbCr where Y

contains luminance information and Cb and Cr contains the chrominance information of the image.

Then Y signal is down sampled at factor two to obtain down sample image. This image is used for

wavelet processing and to embed the watermark processing. This image is again up sampled with

factor two to obtain recover Y component. The up sampling process added zero to the image so that

reconstruction of the image is with is distorted as shown in below.



Fig 6.3 Down sampling and up sampling process for Y component

The problem with process that the up sampling adds the zero to the images so image after up

sample is distorted (recovered Y component) as shown in above Fig 6.3. To over come this

problem, we just simply add the original Y component value to the zero pedding values as shown

in below in Fig 6.4

+ =

Up sampled Y component Difference of Y Reconstructed Y

Fig 6.4 Reconstructed Y plane after up sampling

The complete modified process is shown in below Fig 6.5.

Fig 6.5 Modified Down-Up sampling process with proper reconstruction of Image



The chrominance signal can also be down sample as shown in Fig 6.6 for Cb component and Fig

6.7 for Cr component. We can see the difference between original and recovered up sampled

image in Fig 6.6 and Fig 6.7. As we can see, luminance component Y signal down/up sampling

process provides better results compare to chrominance signal Cb and Cr down/up sampling.

Fig 6.6 Down sampling and up sampling process for Cb component

Fig 6.7 Down sampling and up sampling process for Cr component



6.2 FPGA Implementation Results

Simulation results of Legall 5/3 as shown below.

Fig 6.8 Simulation of Legal 5/3 Scheme

a) SPARTEN Family:

Synthesis was performed by Xilinx project navigator ISE 10.1. During simulation as textio

library was utilized to read the gray scale image (lena 256 x 256) file. After processing results

are stored in text file. This text file is read through MATLAB to generate image.

Table 6.17 Synthesis Report for Proposed DWT, IDWT and Watermarking Processor

Resources DWT processor Watermarking

Processor

IDWT Processor

RAM/ROM 3 2 4

Adders/Subtractor 22 2 23

Latches 34 5 35

Muxs 4 3 4

Counters 3 2 3

Registers/Flip

Flops

382 62 381



The synthesis report and device utilization reports for proposed DWT, IDWT and Watermarking

Processor is shown in Table 6.17 and Table 6.18 respectively. The results are obtained for

Xc3s500e-4fg320-SPARTAN 3E FPGA using ISE 10.1 from Xilinx Tool.

Table 6.18 Device Utilization Report for Proposed DWT, IDWT and Watermarking Processor

Table 6.19 Timing Reports for DWT,IDWT and Watermarking Processor(SPARTEN)

Resources Path Time (ns) Frequency(MHz)

DWT Processor

Minimum Period 9.685 ns

103.254 MHz

Minimum input arrival time

before clock

6.084 ns

Maximum output required

time after clock

7.662 ns

IDWT Processor

Minimum Period 9.772ns

102.337 MHz


before clock

7.200 ns


time after clock

6.385 ns

Watermarking

Processor


433.088 MHz


before clock

4.134 ns


time after clock

4.394 ns

Resources DWT Processor IDWT Processor Watermarking

Processor

Slices 510/4656 (10%) 553/4656 (11%) 5/4656 (0%)

Slice Flip-Flops 361/9312 (3%) 358/9312 (3%) 9/9312 (0%)

4 inputs LUTs 976/9312 (10%) 1062/9312 (11%) 2/9312 (0%)

bounded IOBs 44/232 (18%) 44/232 (18%) 10/232 (4%)

GCLKs 1/24 (4%) 1/24 (4%) 1/24 (4%)



The timing analysis report is also generated using ISE Xilinx 10.1 and we obtained the maximum

operation frequency for each module .The detailed timing report is defined in Table 6.19

b) VIRTEX XCV50-6bg256 Family:

The synthesis report ,device utilization and timing reports for proposed DWT, IDWT and

Watermarking Processor are shown in Table 6.20,Table 6.21 and Table 6.22 respectively. The

results are obtained for XCV50-6bg256-VIRTEX Technology FPGA using ISE 10.1 from Xilinx

Tool.

Table 6.20 Synthesis Report for Proposed DWT, IDWT and Watermarking Processor

Resources DWT processor Watermarking

Processor

IDWT Processor

RAM/ROM 3 2 4

Adders/Subtractor 16 4 14

Latches 12 7 10

Muxs 6 4 5

Counters 2 2 3

Registers/Flip

Flops

469 68 542

Table 6.21 Device Utilization Report for Proposed DWT, IDWT and Watermarking Processor

Resources DWT Processor IDWT Processor Watermarking

Processor

Slices 528/768 (66%) 562/768 (73%) 5/768 (0%)

Slice Flip-Flops 363/1536 (23%) 363/1536 (23%) 9/1536 (0%)

4 inputs LUTs 997/1536 (64%) 1059/1536 (68%) 2/1536 (0%)

bounded IOBs 44/180 (24%) 44/180 (24%) 10/180 (5%)

GCLKs 1/4 (25%) 1/4 (25%) 1/4 (25%)



Table 6.22 Timing Reports for DWT,IDWT and Watermarking Processor(VIRTEX)

Resources Path Time (ns) Frequency(MHz)

DWT Processor


63.424 MHz


before clock

10.658 ns


time after clock

16.757 ns

IDWT Processor


64.508 MHz


before clock

10.582 ns


time after clock

12.767 ns

Watermarking

Processor


280.584 MHz


before clock

6.927 ns


time after clock

7.184 ns

Previously, S.P.Mohanty [63] has developed FPGA based implementation of an invisible

spatial domain watermarking encoder. The watermarking encoder consists of a watermark

generation unit and watermark insertion module. He has synthesized the prototype watermarking

encoder chip using Xilinx FPGA XCV50-6bg250-VIRTEX Technology. Mohanty has also used

LFSR module to insert binary watermark as watermark generation unit. He has design the all the

modules with VHDL. The comparison of the proposed scheme with Mohanty scheme is given in

Table 6.23

Table 6.23 Comparison with Previous Work

Units S.P.Mohanty et al[63] Proposed

Watermark Generation 4.916 ns 4.632 ns

Watermark Insertion 15.526 ns 3.590 ns



c) VIRTEX XC4VLX25-10FF676 Family:

The device utilization and timing reports for proposed DWT, IDWT and Watermarking

Processor are shown in Table 6.24 and Table 6.25 respectively. The results are obtained for

XC4VLX25-10FF676 VIRTEX Technology FPGA using ISE 10.1 from Xilinx Tool.

Table 6.24 Device utilization comparison

Logic Utilization S.P.Maity et al.[64] Proposed

Number of Slices 11536/10752 (107%) 1078/10752 (10%)

Number of Slice Flip-Flops 296/21504 (1%) 728/21504 (3%)

Number of 4 input LUTs 21692/21504 (100%) 2061/21504 (9%)

Number of bonded IOBs 649/448 (144%) 98/448 (21%)

Number of GCLKs 2/32 (6%) 3/32 (9%)

Table 6.25 Timing analysis comparison

Delay S.P.Maity et al.[64] Proposed

Minimum clock period

(Maximum Frequency)

4.555 ns

(212.542 MHz)

6.819 ns

(146.651 MHz)

Maximum output

required time after clock

102.195 ns 105.21 ns

Maximum combinational

path delay

103.346 ns 62.51 ns

S.P. Maity [64] has recently developed watermarking techniques based on spread spectrum

modulation techniques for authentication and copyright protection. He has synthesized the

watermark embedding and decoding logic with help of FPGA family of XC4VLX25-10FF676

VIRTEX Technology. He has embedded 64 x 64 watermark in 256 x256 original image size.

The synthesized results with proposed scheme has compared in Table 6.24. The timing report

comparison is also done with proposed scheme and S.P. Maity scheme in Table 6.25



Previously, Tsai and Lu also developed [43] DCT based invisible watermarking which has

synthesized on 0.35 um technology which operates on 50 MHz and consumes power of around

62.78 mw. Garimella [44] has developed watermarking in spatial domain with invisible fragile

property. The scheme proposed by him was synthesized in on 0.13 um technology and operate on

100 MHz clock frequency with 37.6 uW power consumption. Later on Mohanty [8] has developed

visible watermarking scheme with spatial domain on 0.35 um standard cell technology. The

watermarking algorithm consumes 6.93 mW power and operates on 292 MHz clock frequency.

Later on again Mohanty has developed DCT based Invisible watermarking scheme which has 0.3

mW power consumption and synthesized on 0.25 um technology. The proposed scheme has

synthesized on 0.18um technology with Design Vision of synopsys.



CHAPTER 7. CONCLUSION

The proposed algorithm has combined approaches of spatial and frequency domain. The results are

verified on MATLAB platform against various attacks. The proposed schemes have been

compared with different schemes. We have compared results of MSE, PSNR and Similarity Factor

for watermarked and original images. From Table 6.1-6.16, it has concluded that proposed scheme

has comparable results for transparency, security and robustness. It has overcome the problem of

block artifacts and advantage of both domain properties. From Table 6.17-6.25, it has proved that

it has also lesser computational complexity compare to other algorithms because we embed

watermark in wavelet domain. Overall, the proposed scheme has scheme provides comparable

results of MATLAB simulation which has effective hardware application suited for authentication

and ownership verification application.

82

REFRENCES

[1] P.Blythe and J. Fridrich, “Secure digital Camera”, Digital Forensic Research

Workshop,Aug,2004

[2] S. Emmanuel and M. S. Kankanhalli, “A Digital Rights Management Scheme for

Broadcast Video,” ACM-Springer Verlag Multimedia Systems Journal, vol. 8, no.

6, pp. 444–458, June 2003.

[3] D. Kundur and K. Karthik, “Digital Fingerprinting and Encryption Principles for

Digital Rights Management,” IEEE Signal Processing, vol. 52, 2004.

[4] A. M. Eskicioglu and E. J. Delp, “An Overview of Multimedia Content Protection

in Consumer Electronics Devices,” Elsevier Signal Processing: Image

Communication, vol. 16, pp. 681–699, 2001.

[5] F. A. P. Petitcolas, R. J. Anderson, and M. G. Kuhn, “Information Hiding - A

Survey,” Proceedings of the IEEE, vol. 87, no. 7, pp.1062–1078, July 1999.

[6] S. Voloshynovskiy, S. Pereira, T. Pun, J. J. Eggers, and J. K. Su, “Attacks on

Digital Watermarks: Classification, Estimation-based Attacks and Benchmarks,”

IEEE Communications Magazine, vol. 39, no. 9, pp. 118–126, August 2001.

[7] M. Maes, T. Kalker, J. P. M. G. Linnartz, J. Talstra, G. F. G. Depovere, and J.

Haitsma, “Digital Watermarking for DVD Video Copyright Protection,” IEEE

Signal Processing Magazine, vol. 17, no. 5, pp. 47–57, Sep 2000.

[8] S. P. Mohanty, N. Ranganathan, and R. K. Namballa, “A VLSI Architecture for

Visible Watermarking in a Secure Still Digital Camera (S2DC) Design,” IEEE

Transactions on Very Large Scale Integration Systems (TVLSI), vol. 13, no. 8, pp.

1002–1012, August 2005.

[9] N. J. Mathai, D. Kundur, and A. Sheikholeslami, “Hardware Implementation

Perspectives of Digital Video Watermarking Algortithms,”IEEE Transactions on

Signal Processing, vol. 51, no. 4, pp. 925–938, April 2003.

[10] O. B. Adamo, S. P. Mohanty, E. Kougianos, M. Varanasi, and W. Cai, “VLSI

Architecture and FPGA Prototyping of a Digital Camera for Image Security and

Authentication”, in Proceedings of IEEE Region 5 Technology and Science

Conference, pp. 154-158, 2006.

83

[11] Victor V. Hernandez Guzman, Mariko Nankano, “Analysis of a Wavelet based

watermarking Algorithm”, Proceeding of the 14th International Conference on

Electronics, Communication and Computers, 2004.

[12] Jianyog Huang and Changsheng Yang, “Image Digital Watermarking algorithm

Using Multiresolution wavelet Transform”, IEEE International conference on

systems, man and Cybernetics, 2004.

[13] Sammy H. M. Hawk and Edumund Y. Lam, “Watermark implementation in digital

photography”, Signal processing symposium,Dec-2005.

[14] Lei Tian and Heng-Ming Tai, “Secure Image Captured by Digital Camera”, IEEE,

2006

[15] L.J. Cox, J. Kilian, T. Leighton and T. Shanon, “Secure spread spectrum for

Multimedia”, NEC research institute, prinecton, NJ, technical report 95-10, 1995.

[16] I. J. Cox, J. Kilian, F.T. Leighton, and T. Shamoon, “Secure Spread Spectrum

Watermarking for Multimedia,” IEEE Transactions on Image Processing, vol. 6,

no. 12, pp. 1673-1687, Dec. 1997.

[17] A. Lumini and D. Maio, “A Wavelet-Based Image Watermarking Scheme”, The

International Conference on Information Technology: Coding and Computing

(ITCC’00), Las Vages, NV, pp. 122-127, March 27-29, 2000.

[18] Yong-Gang Fu and Hui-Rong Wang, “A Novel Discrete Wavelet Transform Based

Digital Watermarking Scheme”, 2nd International Conference on Anti-

counterfeiting, Security and Identification, 2008. ASID 2008. pp 55-58,Aug-2008

[19] Hsien-Wen Tseng , Chin-Chen Chang, “An extended difference expansion

algorithm for reversible watermarking”,Elsiver, Image and vision computing, pp

1148-13,2008

[20] Tung-Shou Chen, Jeanne Chen, Jian –Guo Chen, “A Simple and efficient

watermark technique based on JPEG2000 Codec”, Springer, Multi media systems,

pp 16-26, 2004.

[21] N. Memon and P. W. Wong, “Protecting Digital Media Content”, Communications

of the ACM, vol. 41, no. 7, pp. 34–43, July 1998.

[22] Nebu John Mathai, Deepa Kundar, Ali Sheikholeslani, “Hardware Implementation

Perspective of Digital Video Watermarking Algorithm”, IEEE Signal processing,

84

Vol. 51,No. 4, April 2003.

[23] Sammy H. M. Kwok, Edmund Y. Lam, “Watermarking Implementation in digital

photography”, Proceeding of 2005 International Symposium on Intelligent Signal

Processing and Communication System.

[24] R. Gonzalez and R. Woods, Digital Image Processing, Second edition, Prentice

Hall, 2002

[25] R.G. Schynded, a. Tirkel and C. Osborne, “A digital watermark”, Proc. IEEE Int.

Conf. Image Processing ( ICIP), 1994, pp 86-90.

[26] R.B. Wolfgang and E.J. Delp, “A watermark for digital images”, in Proc. Int.

Conf. Image Processing, Lausanne Switzerland, Sept. 1996, vol 3, pp 215-218.

[27] R.G. van Schyndel and C. Osborne, “A two-dimensional watermark”, in proc.

DICTA, 1993, pp. 378-393.

[28] R.B.Wolfgang and E.J. Delp, “fragile watermarking using vw2d watermark”, in

Proc. SPIE, security and watermarking of multimedia contents, an Jose,CA,Jan

1999,vol. 3657,pp 204-213.

[29] W. bender, d. Gruhi and N. Morimoto, “Techniques for data hiding”, IBM syst. J.,

vol. 35,nos. 3-4 pp,313-336,1996.

[30] B.G. Haskell, A. puri and A.N. Netravali, “digital video: An Introduction to

MPEG 2”, New York: Chapman & Hall, 1997.

[31] Hyun Lim, Soon-Young Park and Seong jun Kang, “FPGA Implementation of

Image Watermarking Algorithm for a Digital camera”, Signal processing & VLSI,

IEEE, 2003.

[32] Saraju P. Mohanty, N. Raganathan, “VLSI Implementation of Visible

Watermarking for a secure Digital Camera Design”, Proceeding of 17th

International Conference on VLSI design,2004.

[33] Victor V. Hernandez Guzman, Mariko Nankano, “Analysis of a Wavelet based

watermarking Algorithm”, Proceeding of the 14th International Conference on

Electronics, Communication and Computers, 2004.

[34] Jianyog Huang and Changsheng Yang, “Image Digital Watermarking algorithm

Using Multiresolution wavelet Transform”, IEEE International conference on

systems, man and Cybernetics, 2004.

85

[35] Sammy H. M. Hawk and Edumund Y. Lam, “Watermark implementation in digital

photography”, Signal processing symposium,Dec-2005.

[36] Lei Tian and Heng-Ming Tai, “Secure Image Captured by Digital Camera”, IEEE,

2006

[37] L.J. Cox, J. Kilian, T. Leighton and T. Shanon, “Secure spread spectrum for

Multimedia”, NEC research institute, prinecton, NJ, technical report 95-10, 1995.

[38] I. J. Cox, J. Kilian, F.T. Leighton, and T. Shamoon, “Secure Spread Spectrum

Watermarking for Multimedia,” IEEE Transactions on Image Processing, vol. 6,

no. 12, pp. 1673-1687, Dec. 1997.

[39] A. Lumini and D. Maio, “A Wavelet-Based Image Watermarking Scheme”, The

International Conference on Information Technology: Coding and Computing

(ITCC’00), Las Vages, NV, pp. 122-127, March 27-29, 2000.

[40] Yong-Gang Fu and Hui-Rong Wang, “A Novel Discrete Wavelet Transform Based

Digital Watermarking Scheme”, 2nd International Conference on Anti-

counterfeiting, Security and Identification, 2008. ASID 2008. pp 55-58,Aug-2008

[41] Hsien-Wen Tseng , Chin-Chen Chang, “An extended difference expansion

algorithm for reversible watermarking”,Elsiver, Image and vision computing, pp

1148-13,2008

[42] Tung-Shou Chen, Jeanne Chen, Jian –Guo Chen, “A Simple and efficient

watermark technique based on JPEG2000 Codec”, Springer, Multi media systems,

pp 16-26, 2004.

[43] T. H. Tsai and C. Y Lu, “A Systems Level Design for Embedded Watermark

Technique using DSC Systems,” in Proceedings of the IEEE International

Workshop on Intelligent Signal Processing and Communication Systems, 2001.

[44] A. Garimella, M. V. V. Satyanarayan, R. S. Kumar, P. S. Murugesh, and U. C.

Niranjan, “VLSI Implementation of Online Digital Watermarking Techniques with

Difference Encoding for the 8-bit Gray Scale Images,” in Proceedings of the

International Conference on VLSI Design, 2003, pp. 283–288.

[45] Y. C. Fan, L. D. Van, C. M. Huang, and H. W. Tsao, “Hardware-Efficient

Architecture Design of Wavelet-based Adaptive Visible Watermarking,” in

Proceedings of 9th IEEE International Symposium on Consumer Electronics, 2005,

86

pp. 399–403.

[46] Christine I. Podilchuk and Edward J. Delp, “Digital watermarking: Algorithms and

applications,” IEEE Signal Processing Magazine, Vol. 18, Issue 4, pp 33-46.,July-

2001

[47] Tefas, A., Pitas, I., “Robust Spatial Image Watermarking Using Progressive

Detection”, In Proceedings of the IEEE International Conference on Acoustics,

Speech, and Signal Processing ,Salt Lake City,USA,Vol. 3,pp. 1973–1976,2001

[48] Vidhyasagar M. Potdar, Song Han, Elizabeth Chang, “A survey of Digital Image

Watermarking Techniques”, 3rd International Conference on Industrial Informatics

(INDICON 2005), pp 709-716, 10-12 Aug-2005.

[49] Froment , J. and Mallat, S. “Second Generation Compact Image Coding with

Wavelets”, Academic Press Professional, Inc .,San Diego, CA,USA, pp. 655-678,

1993

[50] Mallat S G. “A theory of multi resolution signal decomposition: The wavelet

representation”, IEEE Trans. Pattern Analysis and Machine Intelligence, 1989, vol.

11, no. 7, pp. 674-693, July 1989.

[51] Woods, J. W. and S. D. O’Neil, “Sub band coding of images,” IEEE Trans.

Acoust., Speech, Signal Processing, vol. ASSP-34, pp. 1278-1288, Oct. 1986.

[52] Strang, G. and Nguyen, T.Wellesley, “Watermark for copyright protection”,

Cambridge Press, Wellesley, MA, 1996,

[53] Vetterli, M. and Kovacevic, “DCT-DWT Watermarking”, Englewood Cliffs, NJ,

Prentice Hall, 1995, http://cm.bell-labs.com/who/jelena/Book/home.html

[54] Donovan Artz, “Digital Steganography: Hiding Data within Data”, IEEE Internet

Computing, Vol 5, pp. 75-80,May-June 2001

[55] E. Lam, “Image Restoration in Digital Photography”, IEEE Transactions on

Consumer Electronics, vol. 49, no. 2, pp. 269-274, May 2003.

[56] Tefas, A., Pitas, I., “Robust Spatial Image Watermarking Using Progressive

Detection”, In Proceedings of the IEEE International Conference on Acoustics,

Speech, and Signal Processing ,Salt Lake City,USA,Vol. 3,pp. 1973–1976,2001

[57] Tsai, T.H., Lu, C.Y., “A Systems Level Design for Embedded Watermark

Technique using DSC Systems.” presented at the IEEE Int. Workshop Intelligent

http://cm.bell-labs.com/who/jelena/Book/home.html

87

Signal Processing Communication System, Nashville, TN, pp 20-23,Nov 2001.

[58] Amara Grapes, “An introduction to wavelets”, IEEE Computational Science and

Engineering, vol. 2, num. 2, published by the IEEE Computer Society, Summer

1995,

[59] Mohanty, S.P., Ranganathan, N., Namballa, R.K.“VLSI Implementation of

Invisible Digital Watermarking Algorithms Towards the Development of a Secure

JPEG Encoder.” In: Proceedings of the IEEE Workshop on Signal Processing

Systems, pp 183–188, Aug 27-29,2003

[60] Majid Rabbani and Rajan Joshi. “An overview of the jpeg2000 still image

compression standard”, Signal Processing: Image Communication Journal, Vol. 17,

pp 3-48, 2002.

[61] W. Sweldens, “The lifting scheme: A new philosophy in Biorthogonal wavelet

constructions,” in Proc. SPIE, vol. 2569, 1995, pp. 68–79.

[62] C.C. Liu,Y.H. Shiau, and J.M. Jou, “Design and Implementation of a Progressive

Image Coding Chip Based on the Lifted Wavelet Transform,” in Proc. of the 11th

VLSI Design/CAD Symposium, Taiwan, 2000

[63] S.P.Mohanty, Renuka Kumara C.,Sridhara Nayak, “FPGA Based Implementation

of an Invisible-Robust Image Watermarking Encoder,” In: Lecture notes in

computer science,vol. 3356,pp 344-353,2004

[64] Santi P Maity, Hafizur Rahman, Sudip Ghosh,Pranab Ray, “Spread Spectrum

Image Watermarking with Digital Design,” IEEE International Advance

Computing Conference (IACC 2009),Patiala, India, pp 868-873,2009

[65] Saraju P. Mohanty, N. Raganathan and K. Balakrishna, “ A Dual Voltage-

Frequency VLSI chip for Image Watermarking in DCT Domain”, IEEE

Transaction on circuits and systems II(TCAS-II), vol. 53,pp 394-398,May 2006.

PUBLICATION

[1] Amit M Joshi, Anand Darji, “Secure Digital Camera with DWT

Based Watermarking”, International Journal of Emerging Technologies and

Application in Engineering Technology and Sciences, vol 2,issue 2,pp

813-818,July-2009 (Published).

[2] Amit M Joshi, Anand Darji, “Efficient Dual Domain Watermarking Scheme

for Secure Images ”, IEEE International Conference on Advanced in

recent Technologies in Communication & Computing,ARTCOM-2009

(Selected).

[3] Amit M Joshi,Anand Darji,Taral Chhatbar, Abhishek Agrahari, “Clock

Skew Analysis & Optimization of Sequential Circuit”, National Conference on

Inteligent Systems NCIS-07 at MJCET, Hyderabad, August-2007 (Selected and

Presented)

APPENDIX A

INTERNATIONAL JOURNAL OF EMERGING TECHNOLOGIES AND APPLICATIONS INENGINEERING, TECHNOLOGY AND SCIENCES (IJ-ETA-ETS)

ISSN: 0974-3588 | July ’09 – December ’09 | Volume 2 : Issue 2 | Page: 813

SECURE DIGITAL CAMERA WITH DWT BASEDWATERMARKING

1 Amit M. Joshi, 2 Anand D. Darji

1, 2 ECED, SVNIT, Surat

1 [email protected], 2 [email protected]

ABSTRACT:The rapid growth of digital media information is demanding secure content for authentication. The imagecaptured by digital camera can easily altered by editing software. Such problems have hindered the applicationof digital image for courtroom evidence, insurance claim, copyright and journalistic photography. By usingwatermarking scheme, information such as the author’s own identity embedded into digital images forauthentication. Such an embedded watermark can be extracted later on as a proof for ownership. For digitalcamera, we have to use invisible and robust watermarking scheme for authenticity and ownership. The camerawith inbuilt such capability is called as “Secure Digital Camera”. Previously, spatial domain watermarkingschemes are very widely use. But nowadays, frequency domain schemes are very popular due to their transformdomain property. DWT based watermarking is upcoming technology of frequency domain watermarking fordigital camera due to its multiresolution property. The main goal of research is to adapt DWT basedwatermarking algorithm which provides ownership verification.

Keywords- Digital Watermarking, Discrete Wavelet Transform, Secure Digital Camera, WaveletDecomposition

1. INTRODUCTIONDigital camera has been widely used in present daysand images captured by it are displayed on WWWand electronic commerce. But digital images are noteasily acceptable in courts, as a proof and medicalimage archiving because they can be altered ormanipulated with ease. Watermarking is a techniquefor inserting ownership information (watermark) intoan image, which can be detected later on foridentification and authentication. In 1998, USCongress passed The Digital Millennium CopyrightAct (DMCA) which makes it illegal to circumventany technological measure that protects an owner’sintellectual property rights of digital content. Digitalwatermarking provides security of owner againstcopyright violation. Digital watermarks aretransparent signatures. They are integrated withindigital files as noise, or random information thatalready exists in the file. Thus, detection and removalof watermark becomes more difficult. Typically,watermarks are dispersed throughout the entiredigital file such that the manipulation of one portionof the file does not alter the underlying watermark.There are several data hiding techniques available forsecurity apart from watermarking such ascryptography and steganography. Cryptography dealswith scrambling of message to form encryptedversion with help of secret key. With steganography,the sender of a message would hide it in a host file.The host file, or cover message, is the file thatanyone can see in steganography, usually the

message itself is of value and must be protectedthrough clever hiding techniques and the “vessel” forhiding the message is not of value. Watermarking isthe direct embedding of additional information intothe original content or host signal. Ideally, thereshould be no perceptible difference between thewatermarked and original signal and the watermarkshould be difficult to remove or alter withoutdamaging the host signal. In watermarking, theeffective coupling of message to the “vessel,” whichis the digital content, is of value and the protection ofthe content is crucial. Previously, spatial domainwatermarking techniques for image data was popular.Some of the earliest techniques embed m sequencesinto the Least Significant Bit (LSB) of the data toprovide effective transparent embedding technique[1]. M-sequences were extended to 2 dimensional [2].The block based method proposed in [3] to be robustfor JPEG Compression. First DCT based semi fragilewatermarking algorithm for digital camera withFPGA implementation was developed by Hyun Lim,Soon-Young Park and Seong jun Kang in early2003[4]. In 2004, Saraju P. Mohanty, N. Raganathanhad also developed visible watermarking scheme onDCT [5]. After that some algorithms for waveletbased approach were starting to explore to adaptJPEG2000 new millennium standard and exploremultiresolution property of Wavelet [6]. In 2005 laterpart, Sammy H. M. Hawk and Edumund Y. Lamproposed watermark implementation technique indigital photography with DWT approach [7] for

mailto:[email protected]

mailto:[email protected]



software based implementation. Afterwards, there areseveral attempts to develop secure images capturedby digital camera for DWT based approach. The flowof paper is as following manner. Section 1 includesthe literature review of digital watermark and scopeof watermarking for digital camera. Section 2describes the overview of digital camera and theblock diagram of digital camera. Section 3 emphasison different properties of digital watermarking andwatermark embedding and detection process. Section4 describes three different algorithmsimplementation. Section 5 and 6 provides results andconclusion for DWT based watermarking for digitalcamera.2. DIGITAL CAMERA OVERVIEWA block diagram of the proposed camera withwatermark embedded unit is shown in Fig. 1. In thedigital camera, the image is captured by a sensor andconverted to a digital signal by ADC. The capturedimage is stored temporarily in the Temp. memory,after which it is displayed on the LCD with the helpof the control unit. The display is used to enable theuser to see the image frame before it is watermarkedby the watermark unit and stored in the camera,which can be further transmitted over the network, ortransferred to Flash memory, computer hard drive oroptical discs. The control unit is responsible forcontrolling the entire sequence of events. There arethree main blocks of our interest compression unit,watermark unit and Co-efficient Key Generator. Userhas been provided with a flexibility to make adecision about insertion of watermark into an image.Based on requirement, watermark may or may not beinserted into image. Images with and withoutwatermark are stored in two separate memories.Images without watermark are stored in Temp.memory and watermarked image are stored in Flashmemory. In case of non blind detection, imagesstored in Temp. memory can be used for detection.The objective of research is to develop a feasible andeffective DWT based digital watermarking scheme[8] to provide a cost-effective and practical tool forauthentication and ownership verification of digitalimages taken by Secure Digital Camera.

Fig 1 Block Diagram of Secure Digital Camera

2. PROPERTIES OF DIGITALWATERMARKING

Watermarking provides user its ownership forcopyright protection. A good watermarking schemeshould meet following requirements.Security: The embedded mark should only beaccessible by authorized organization. Although thewatermarking algorithm is published to everyone,those without a secret key cannot access the marks inthe protected images.Robustness: It should resist malicious attacks such ascompression, rotation, image enhancement, resizing.Imperceptibility: The digital watermark should notbe perceptible to the viewer.Capacity: It shows total number of bits can beembedded per unit of cover image.3.1 Spatial Domain Vs Frequency DomainWatermarkingThe watermark can be applied to frequency or inspatial domain. The frequency domain methods aremore robust than that of the spatial domaintechniques. On the other hand, the spatial domainwatermarking schemes have less computationalcomplexity compared to frequency domain schemes.Nowadays, most researches focus on developingvarious watermarking methods in the frequencydomain since most currently popular informationtransmission and compression techniques are infrequency domain e.g. like JPEG, MPEG, DCT, etc.Compared with other transforms in frequencydomain, DWT can better localize the feature whosechange is less sensitive to human eyes and localizethe information in space and frequency [9].Especially, DWT based watermarking methods havebeen researched intensively due to the fact that thecurrent image compression is based on waveletdomain, such like JPEG 2000.The spatial domainwatermarking scheme is generally fast and simple,but it doesn’t guarantee robustness against commonsignal distortion. The frequency domain methodbased on a multi resolution wavelet decomposition,which shows greater robustness against suchcommon attacks. We have focused on wavelet basedfrequency domain watermarking method for ourSecure Digital Camera.3.2 Watermarking with Wavelet DecompositionLevelFor Secure Digital Camera, watermarking schemeshould be invisible and robust for various attacks[10]. To have invisibility we insert watermark in highfrequency band (HH) but it provides less robustnessto the system. To have robustness we can embedwatermark into the Low frequency band (LL) but ofcourse the invisibility then the issue, so there isalways trade off between invisibility and robustness[11].With increase the level of decomposition, therobustness will increase but computationalcomplexity will also be increased. So these issuesmust be considered while dealing with watermarking



embedding schemes. The area of research to developwatermarking DWT based algorithm for SecureDigital Camera.3.3 Implementation of Digital WatermarkFor Secure CameraThe whole watermark embedding process is shown inFig 2.Whereas the watermark extraction process isshown in Fig 3.

Fig. 2 Proposed Watermark Embedding Process

Fig. 3 Watermark Extraction Process

As with traditional color processing, we first convertan image from an RGB color space to the YCbCrgray space. Then the Y component of the image isdown-sampled to form a grayscale image ofresolution of 1 M pixels (assuming that the original isbetween 2M and 8M pixels, true for most digitalcameras today). Afterwards, a watermark isembedded in the image by Key of co-efficientselection for DWT of the image. Finally, the image isconverted back to spatial domain by IDWT and awatermarked image is formed by up-sampling theimage and adding it with the original Cb and Cr colorcomponents. The Key provides co-efficient selectionplaces at which watermark to be embedded. This Keyis only known to user and same key must be use atthe time of extraction. Since only the user of theSecure Digital Camera knows the key for thewatermarking, so security against forgery isguaranteed. First image is converted to transformdomain by DWT and then watermarking is applied tothe frequency domain co-efficient. There are variousmethods available for DWT based watermarkingschemes.3.4 Performance of Watermarked Image3.4.1 Similarity Factor (SF): To assess the changein quality of image, it is required to correlate eachcoefficient of original image as well as watermarkedimage. It is a normalized correlation.

( , ) '( , )

( , ) ( , ) '( , ) '( , )

w m n w m nSF

w m n w m n w m n w m n

Where, w (m, n) = original Image,

w (m, n) = watermarked Image3.4.2 Mean Square Error (MSE): To measure thecredibility of the modified image, the Mean SquareError (MSE) between the original watermark and theextracted watermark can be used. MSE is defined as

2

1

1( , ') ( ( ) '( ))

wN

iw

MSE w w w i w iN

Where, w = Original watermark,

w = Extracted watermark,Nw = Length of watermark.

4. DWT BASED WATERMARKINGALGORITHMSIn the DWT based watermarking method, a twodimensional image is divided into four frequencybands namely, low frequency band (LL), Highfrequency band (HH), Low-high frequency band(LH) and High-low frequency band (HL). LL bandcontains important information of the image so itsvalue is likely to remain constant, because it is highenergy band and its coefficient values do not getchanged with attack. This band is used to calculatesignature of an image. We have explained threedifferent algorithms for DWT based watermarkingschemes.Algorithm 1:It performs two level wavelet decomposition of thehost image and obtains the sum of the ten highestapproximations co-efficient as their values are likelyto remain constant. This is passed to the randomnumber generator as a seed. In implementing theabove technique we have assumed that the samerandom number generator is available at bothreceiver and the transmitter end. One of thelimitations of this technique is that detection of therecovered image depends on the availability oforiginal image.Following are steps for watermarking scheme.Embedding Process:1) Read the host image and perform two levelwavelet decomposition of the host image2) Sum the 10 highest approximation coefficients anddivide it by 512 and store it in ‘M’ variable.3) Pass the value of ‘M’ to random number generatoras initial state which will generate the watermark4) Add this watermark to the detailed diagonal co-efficient of the transformed image.

( , ) ( , ) ( , )x xd m n d m n w m n



Where,

( , )xd m n = watermarked image,

( , )xd m n = original image

w (m, n) = watermark to be embedded = scaling factor for watermark =0.01

5) Reconstruct the image by doing inverse wavelettransformation which will give the watermarkedimage6) To simulate an attack on watermark image, wehave added salt and pepper noise, 180 degreeclockwise rotation and rescaling to check robustnessof watermarking scheme.Detection Process:7) Perform two level wavelet decomposition of awatermarked image8) Sum the 10 highest approximation coefficients anddivide it by 512 and store it in ‘M1’ variable9) Pass the value of ‘M1’ to random numbergenerator as initial state which will give watermark.10) Remove this watermark from the detaileddiagonal co-efficient of the transformed watermarkedimage to get the recovered image

' ( , ) ( , )

( , )x xd m n d m n

w m n

Where = scaling factor for watermark11) Reconstruct the recovered image by two levelinverse wavelet transformations.Algorithm 2:This watermark scheme is based on wave lift integertransform. It uses simple random number to performwatermarking.The following steps are performed:Embedding Process:1) Read an input image2) Perform 2 level integer wave lift scheme of legall5/3 to obtain approximation image and three detailedcoefficient.3) Generate the watermark with pseudo randomnumber generator4) Add the watermark image with secret key toneighboring approximation co-efficient5) Perform wavelet transform to generatewatermarked image6) To simulate an attack on watermark image, wehave added salt and pepper noise, 180 degreeclockwise rotation, compression and rescaling tocheck robustness of watermarking scheme.Detection Process:7) Perform Inverse wavelet transform to getrecovered image.8) Remove the watermark from approximation co-efficient exactly inverse process of embeddingprocess.9) Perform 2 levels Inverse integer wave lift schemeof legall 5/3 to obtain transform domain co-efficient

10) Compare the original Image with recoveredImage.Algorithm 3:This watermark scheme is based on the fixedthreshold of wavelet co-efficient. We assume that thebinary watermark is of length Nw, and consists ofelements from the set {−1, 1}. Let OI be the originalimage. We embed the watermark into the detailwavelet coefficients using a watermarking key.Embedding Process:1) Read the host image OI.2) Compute the first level wavelet decomposition ofthe original image. We obtain one approximationimage and three detail images, corresponding to thehorizontal, vertical and diagonal details of the image.We denote the Xth detail image of the original by d x

(m n) , where x Є {h, v, d} (h, v, d stands for“horizontal”, “vertical” and “diagonal” co-efficient,respectively). The approximation image is ax (m, n).3) For each detail image, we consider the coefficientlocation (m, n) and we compute a watermarking key kx (m n), where x Є {h, v, d}

kx (m, n) = 1 if dx (m, n) > qMx

= 0 otherwiseWhere q is a user-defined variable and Mx= maximumof the value (m, n).4) For the given detail image d x (m n), if theassociated value of the key is zero, we do not embeda mark.5) To embed a watermark bit w (m, n); we increase ordecrease the detail image coefficient with a givenquantity:

( , ) ( , ) ( , )x xd m n d m n w m n Where x Є {h, v, d}, α =0.01 = user-defined positive

variable.6) The watermarked image is computed from theapproximation image a(m, n) and from the detailimages d’ x (m, n), where x Є {h, v, d}. Theparameters q and α are user-defined. It is easy tonotice that a larger value for α and a smaller value forq will increase the robustness of the method, but willmake the marking process more visible.Detection Process:7) To extract the watermark, compute the first leveldiscrete wavelet decomposition of the received imageRI and of original image OI. We obtain twodecomposition images r (m, n) and o (m, n). Wedenote the xth detail image of received image by rdx,where where x Є {h, v, d} (h, v, d stands forhorizontal, vertical and diagonal respectively.)Theapproximation image is ra (m, n).8) Compute the watermarking key kx (m, n). Wemake use of the key to locate the coefficients wherethe mark was embedded. We extract the watermarkas follows:



[ ( , ) ( , )]1, 0

[ ( , ) ( , )]( , ) 1, 0

( 1,1),

x x

x x

rd m n d m nif

rd m n d m nrw m n if

rand otherwise

Where rw (m, n) is the recovered watermark bit.9) If the watermark had been embedded in differentlocations several times, then the most common bitvalue is assigned for the recovered bit. Otherwiserandom guess is made for this value.

5. RESULTSWe have been implemented three different DWTbased watermarking algorithms to check basicproperties of transparency, robustness and security ofwatermarking. We have used input image as 256x256 jpg image.

Transparency:

Table 1 Similarity Factor and MSE comparison oforiginal and watermarked Image

Robustness:Attacks Adding

noise asSalt &pepper

Rotateclockwise

180 degree

Rescaling

Algorithm-1MSE-1.84 x

10-3

5.77x10-4 0.0081 50.40 x10-4


102

1.81x101 6.56 x 102 33.01 x10-6


10 -4

6.2 x 10-4 6.93 x 10-4 44.93 x10-4

Table 2 MSE Comparison of Watermarked andOriginal Image for different algorithm after attacks

Attacks Addingnoise asSalt &pepper

Rotateclockwise 180degree

Rescaling

Algorithm-1SF-1.0000

0.9990 0.9812 0.9963


0.9617 0.9950 1.0000


0.9754 0.9826 0.9855

Table 3 Similarity Factor Comparison ofWatermarked and Original Image for different

algorithm after attacks

Security:Attacks Adding

noise asSalt &pepper

Rotateclockwis

e 180degree

Rescaling


10-8

3.0213 x10-8

2.106 x10-4

1.64 x 10-7


10-4

1.6852 x10-4

1.353 x10-4

1.31 x 10-4


10-4

5.892 x10-4

6.251 x10-4

4.93 x 10-4

Table 4 Similarity Factor Comparison of Recoveredand Original Image for different algorithm after

attacks

6. CONCLUSIONSecure Digital Camera requires robust and invisiblewatermarking scheme for captured image. The resultsare verified on MATLAB platform against variousattacks. Three algorithms have been developed forrobustness, invisibility and security measurementwith standard parameter. Table 1 suggests thatalgorithm 2 provides better transparency compare toothers schemes because of better Similarity Factorand less Mean Square Error(MSE). Table 2 and 3reveal that algorithm 1 provides greater robustnessand correlation compare to other schemes forwatermarked and original image. Table 4 and 5emphasis that algorithm 2 provides better security inorder to correlate Similarity Factor and Mean SquareError for original and recover watermark. Overallalgorithm 2 provides better transparency, security androbustness among all three algorithms. Further workrequired for blind watermarking technique to developdigital camera with their watermarking capability.

Similarity Factor MSEAlgorithm-

10.9970 4.5632 x 10-4

Algorithm-2

1.0000 3.0029 x 10-6

Algorithm-3

0.9873 5.1027



Attacks Addingnoise asSalt &pepper

Rotateclockwis

e 180degree

Rescaling


0.9996 0.3194 0.9985

Algorithm-2(SF-0.9617

0.9994 0.9617 1.0000

Algorithm-3(SF-0.9477

0.9257 0.9400 0.9439

Table 5 MSE Comparison of Recovered and Originalwatermark for different algorithm after attacks

7. REFERENCES[1]. R.G. Van Shindig, A.Z.Tirkel and C. F. Osborne,

“A Two-Dimensional Digital Watermark,” inproc. DICTA, 1993, pp. 378-393.

[2]. R.G. Van Schyndel, A.Z.Tirkel, N.R.A.Mee,C.F.Osborne., “A Digital Watermark,” FirstIEEE Image Processing Conference, HoustonTX, November 15-17, 1994, vol II, p.86-90.

[3]. A.N. Netravali and B.G.Haskell, “DigitalPictures-Representation, compression andstandards,” 2nd Edition, New York, PlennumPress, 1995.

[4]. Hyun Lim, Soon-Young Park and Seong JunKang, “FPGA Implementation of ImageWatermarking Algorithm for a Digital camera,”Signal processing & VLSI, IEEE, 2003.

[5]. Saraju P. Mohanty, N. Raganathan, Ravi K.Namballa, “VLSI Implementation of VisibleWatermarking for a Secure Digital CameraDesign,” Proceeding of 17th InternationalConference on VLSI design, 2004.

[6]. Victor V. Hernandez Guzman, Mariko NankanoMiyatake, Hector M. Perez Meana “Analysis of aWavelet based watermarking Algorithm,”Proceeding of the 14th International Conferenceon Electronics, Communication and Computers,2004.

[7]. Sammy H. M. Kwok and Edmund Y. Lam,“Watermark implementation in digitalphotography,” Processing Of 2005 InternationalSymposium on Intelligent Signal Processing andCommunication Systems, Dec-2005.

[8]. Nebu John Mathai, Deepa Kundur, AliSheikholeslani, “Hardware ImplementationPerspective of Digital Video WatermarkingAlgorithm,” IEEE Signal processing, Vol. 51,No. 4, April 2003.

[9]. Jianyog Huang and Changsheng Yang, “ImageDigital Watermarking algorithm UsingMultiresolution wavelet Transform,” IEEEInternational conference on systems, man andCybernetics, 2004.

[10]. Pennebaker, W. B., Mitchell, J. L. ,“Arithmetic coding articles,” IBM J. Res.Dev., vol. 32, no 6 (Nov1988), pp 717-774.

[11]. Froment, J. and Mallat S, “SecondGeneration Compact Image Coding withWavelets,” in C.K. Chui, editor, WatermarkEmbedding Techniques, vol. 2, Academic Press,NY, 1992..

1

Efficient Dual Domain Watermarking Scheme for Secure Images

Amit M Joshi PG Student, Electronics Engg. Department,

Sardar Vallabhbhai National Institute of Technology, Surat, India.

[email protected]

Anand Darji Lecturer, Electronics Engg. Department,

Sardar Vallabhbhai National Institute of Technology, Surat, India.

[email protected]

Abstract— Day to day growth of multimedia technology draws great attention for security. Digital images can be easily altered with software. This problems demand copyright protection and ownership verification. Digital watermarking is proposed as one of the ways to accomplish this, in which Digital watermarks are generally embedded into digital images in a manner that make the watermark invisible to a human observer as such watermarks do not cause degradation in the visual quality, or in the usefulness of the images. Digital watermarking is a technique to insert owner’s identity into images for authentication. Digital watermarking has emerged as a new area of research for the Intellectual Property (IP) protection of images. Spatial domain watermarking has advantage of less computational cost. Frequency domain watermarking provides more robustness. The proposed algorithm has been developed to take advantage of both spatial as well as frequency domain properties.

Keywords— Digital watermark, Discrete Cosine Transform, Discrete Wavelet Transform, Least Significeny Bit, Ownership

I. INTRODUCTION Due to the rapid and extensive growth of electronic

publishing industry, data can now be distributed much faster and easier. The digital multimedia information such as images, texts, music and videos are exchanged through open channels of internet. This information can be easily copied, modified or tampered without any authorization during multimedia transmission. So there is a need of security for information. One possible solution to this proprietary problem can be achieved with digital watermarking. A new emerging technology, digital watermarking protects digital media by embedding a robust signal directly into the media, thus providing a promising way to protect the digital media from illegal copyright and manipulation. Digital watermarking is a process of embedding data called as watermark (may be text, image or random number) into multimedia object in order to protect the owner’s right to the object. The concept of digital watermarking arose while trying to solve problems related to the copyright of intellectual property in digital media. It is used as a means to identify the owner or distributor of digital data. There is a wide range of application of digital watermarking, including copy right protection, authentication, finger printing, copy control and broadcast. For different kind of applications, digital watermark should show different properties. Here, we mainly discuss the ownership verification and authentication

problem. For this kind of application, digital watermarking should have properties like invisibility, robustness, high detection reliability, etc. The embedded data may be either perceptible or imperceptible. In general, Digital watermarking system consists of three main stages: embedding stage, manipulation stage and extraction stage as shown in Figure 1. The embedding stage involves hiding of owner’s identity for authentication, watermark, inside the image without affecting the visual quality of original image. The watermark can be of any form in terms of pseudo random number, text and image.

Embedding Stage Manipulation Stage Extraction Stage

Figure 1. Block Diagram of Watermark process

The key can be used to enforce security in which it prevents unauthorized parties from recovering and manipulating stage, also referred to some attacks like adding noise, signal distortion and different geometric operations. The extraction stage is an important stage to make assertions about data ownership verification. Digital watermarking is a major sub discipline of information hiding. A watermarking system can be either imperceptible (invisible) or visible. Imperceptible watermarking systems may be more robust towards non destructive attacks where as visible watermarking may be fragile. This kind of watermarking is ideal for image authentication applications. Different approaches are adopted to provide efficient imperceptible watermarking algorithms. Previously, spatial domain watermarking techniques were very popular due to their less computational complexity. Some of the earliest techniques used to embed m sequences into Least Significant Bit (LSB) of the data to provide effective transparent embedding technique [1].Another Spatial Domain technique consist of embedding a texture based watermark into a portion of the image with similar texture, it will be difficult to pensive the watermark. The watermark is detected using a correlation detector [2].DCT based transform domain watermarking started in later part of 90s. In 2004, Saraju P. Mohanty, N. Raga Nathan had also developed visible watermarking scheme on DCT [3]. After some period, wavelet based approach were starting due to their multiresolution property for wavelet [4]. In 2005 later part, Sammy H., M.

Original Image

Watermarked Image

Recovered Image

2

Hawk and Edumund Y. Lam proposed watermark implementation technique in digital photography with DWT approach [5]. In 2007, Cong JIN, Liang-Gang PAN and Ting SU proposed the scheme for Image watermark based of DWT with Visual based concept [6]. In 2008, M.A.EI-Iskandarani, Saad M. Saad and Ahmed M. Abubahia has proposed novel concept of DWT based watermarking with Lifting scheme which uses filter parameterization concept for efficient image watermarking[7]. The flow of the paper is as follow: Section II provides different properties of watermarking. In Section III explains three different algorithms for watermarking. Section IV gives results for different algorithm and Finally, Section V describes conclusion

II. PROPERTIES OF WATERMARKING

A. Types of Watermark 1) Visible: Visible watermarks are designed to be easily

perceived by the viewer, and clearly identify the owner. The watermark must not detract from the image content itself, however. Most research currently focuses on invisible watermarks, which are imperceptible under normal viewing conditions.

2) Fragile: Fragile watermarks are designed to be distorted, or to be broken, under the slightest changes to the image. Semi-fragile watermarks are designed to break under all changes that exceed a user-specified threshold. 3) public and private: In some applications like copyright protection and data monitoring, watermark extraction algorithm can use original un-watermarked data to find the watermark. This is called as non oblivious or private watermarking. The original image is usually subtracted from the suspected image before a detection algorithm is applied. Non-oblivious watermarking is at present the only option for reliable copyright protection. 4) Asymmetric and symmetric: Asymmetric watermarking (also called asymmetric key watermarking) is a technique where different keys are used for embedding and detecting the watermark. In symmetric watermarking (or symmetric key watermarking), the same keys are used for embedding and detecting watermarks.

B. Requirements for watermarking A watermark is designed to permanently reside in the host

data. When the ownership of data is in question, the information can be extracted to completely characterize the owner. To achieve maximum protection of intellectual property with watermarked media, several requirements must be satisfied:

1) Perceptual Transperency: In many applications, such as copyright and usage tracking, embedding data or additional information, the algorithms must embed data without affecting the perceptual quality of the underlying host signal. A data-embedding procedure is truly imperceptible if humans cannot differentiate between the original host signal and a host signal with inserted data.

2) Robustness: The watermark should be retrievable even if common signal processing operations are applied, such as signal enhancement, geometric image operations and noise filtering. This modification include additive Gaussian or non-Gaussian noise,, linear filtering JPEG, wavelet, rotation, spatial or temporal scaling, nonlinear filtering e.g., median filtering. In general the robustness of watermark against the processing techniques depends on the embedding depth and amount of information bits of watermark [8].

3) Security:A secure data-embedding procedure cannot be broken unless the unauthorized user has access to a secret key that controls the insertion of the data in the host signal .

4) Capacity:The capable size of embedding information is defined as the embedding capacity. The embedding capacity should not be extremely low to affect the accuracy of the retrieved watermark and the recovered image.

5) Robustnes:The watermark must be able to be easily and securely embedded and retrieved by the owner. Therefore, the overheads of embedding process and retrieving process should be limited in a reasonable range.

C. Watermarking for Different Domain Generally, there are two broad watermarking schemes and

categories: spatial domain watermarking and frequency domain watermarking. In a spatial domain watermarking scheme, the watermark is embedded by direct modifying the pixel value of an image These spatial domain watermarking schemes are simple and less robust to common signal processing operations, since the watermark doesn’t spread all over the image and some common signal processing easily erase the embedded watermark without affecting the quality of the watermarked image. On the other hand, frequency domain watermarking schemes involve embedding the watermark by modifying the transform co-efficient, after the image has been transformed to the transform domain.

1) Spatial domain Watermarking: Spatial domain watermarking is performed by modifying values of pixel of a frame. Let us denote a picture to be watermarked by P and values of its pixel samples by iP , a watermarked version of picture P by 'P and values of its pixel color samples by 'iP . Let us have as many elements of watermark W with values Wi as number of pixels in picture P . Watermark W hereby covers the whole picture P . Further, it is possible to increase the watermark strength by multiplying watermark element values by weight factor α. Then the natural formula for embedding watermark W into picture P is 'i iP P Wα= +

The most common algorithm using spatial domain watermarking is LSB. The simplest watermarking method in spatial domain is to just flip the least significant bit (LSB) of chosen pixels in an image. Given the extraordinarily high channel capacity of using the entire cover for transmission in this method, a smaller object may be embedded multiple times. Even if most of these are lost due to attacks, a single surviving watermark would be considered a success [9].

3

2) Frequency Domain Watermarking :These methods are similar to spatial domain watermarking in that the values of selected frequencies can be altered. Because high frequencies will be lost by compression or scaling, the watermark signal is applied to lower frequencies, or better yet, applied adaptively to frequencies containing important elements of the original picture. Upon inverse transformation, watermarks applied to frequency domain will be dispersed over the entire spatial image, so these methods are not as susceptible to defeat by cropping as the spatial technique.

a) Discrete Cosine Transform (DCT) :The DCT allows an image to be broken up into different frequency bands, making it much easier to embed watermarking information into the middle frequency bands of an image. The middle frequency bands are chosen such that they have minimize they avoid the most visual important parts of the image (low frequencies) without over- exposing themselves to removal through compression and noise attacks (high frequencies).

b) Discrete Wavelet Transform (DWT):In DWT based approach, each level produces four level bands of data, one corresponding to the low pass band(LL), and three other corresponding to horizontal (HL), vertical (LH) and diagonal (HH) high pass bands. The decomposed images shows a coarse approximation image in the lowest resolution low pass band, and three detail image in higher bands. The low pass band can further be decomposed to obtain another level of decomposition. Watermark data inserted into low frequencies is more robust to image distortions that have low pass characteristics like filtering, lossy compression and geometric manipulation and less robust to change of the histogram such as contrast/brightness adjustment, gamma correction and cropping. On the other hand, watermark data inserted into middle and high frequencies is typically less robust to low pass filtering, lossy compression and l geometric deformation of the image but extremely robust with respect to noise adding and nonlinear deformation of the gray scale.

III. PROPOSED WATERMARKING ALGORITHM Spatial-domain digital watermarking methods are generally

considered as having poor performance after geometric distortion (such as cropping and scaling), common signal processing (such as JPEG and filtering), or subterfuge attacks (such as changing the least-significant bits).It has less computational cost due to their easy operation. So it is more efficient. On the other hand, frequency-domain watermark techniques, have their high computation complexity, and provide great robustness to different attack. However, we have adopted the combined approach with spatial-frequency domain approach which has advantages of both domains. The frequency domain transformation is done with simple haar wavelet and spatial domain transformation done with bit plane slicing. This paper describes the technique of wavelet transform for decomposition of image into different bands. For the image decomposition and reconstruction the Haar transform has been applied as a basic tool used in the wavelet transform. Here some mathematical analysis of Haar transforms extraction using decomposition and reconstruction

matrices. is described for image compression and image pixels features

A. Haar Wavelet: Wavelets are functions generated from a single function by

its dilations and translations. The Haar transform forms the simplest compression process of this kind. In 1-dimension, transforms a 2-element vector [x (1), x (2)]T into [y (1), y (2)]T by relation:

(1) (1)(2) (2)

y xT

y x⎡ ⎤ ⎡ ⎤

=⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦ Where,

1 111 12

T⎡ ⎤

= ⎢ ⎥−⎣ ⎦ (1)

is an orthonormal matrix as its rows are orthogonal to each other (their dot products are zero).Therefore T−1 = TT and it is possible to recover x from y by relation

(1) (1)(2) (2)

Tx yT

x y⎡ ⎤ ⎡ ⎤

=⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦ (2)

In 2-dimensions x and y become 2 × 2 matrices. We can transform at first the columns of x, by pre-multiplying by T, and then the rows of the result by post-multiplying by TT to find

y = T x TT and in the next step x = TT yT (3)

To show more clearly what is happening we can use a specific matrix x of the form

12

a b a b c d a b c dx imlying that y

c d a b c d a b c d+ + + − + −⎡ ⎤ ⎡ ⎤

= =⎢ ⎥ ⎢ ⎥+ − − − − +⎣ ⎦ ⎣ ⎦ (4)

(a) Original Image (b)1st level decomposed Image

Figure 2. Wavelet Decompostion with Haar Wavelet

These operations correspond to the following filtering processes: Top left: 2-D low pass filter (LL). Top right: horizontal high pass and vertical low pass filter (HL). Lower left: horizontal low pass and vertical high pass filter (LH). Lower right: 2-D high pass filter (HH). To apply this transform to a complete image, we group the pixels into different blocks and apply (3) to each block. To view the result, all the top left components in Figure 2(a) of the different blocks in y were grouped together to form the top left sub image in Figure 2(b) and the same for the components in the other three positions. It is clear from Figure 2(b) that the most of the energy is contained in the top left (LL) sub image and the least energy is in the lower right (HH) sub image. The top right (HL) and the lower left (LH) sub image contains the

4

edges. The energies of LL Band (88-89%), HL(4-5%),LH(6-7%) and HH(1-2%). So the LL band contains most part of energy while HH band contains marginal part of the total energy of image.

B. Bit-plane slicing Instead of highlighting gray level images, highlighting the contribution made to total image appearance by specific bits might be desired. Suppose that each pixel in an image is represented by 8 bits. Imagine the image is composed of 8, 1-bit planes ranging from bit plane1-0 (LSB) to bit plane 7 (MSB). In terms of 8-bits bytes, plane 0 contains all lowest order bits in the bytes comprising the pixels in the image and plane 7 contains all high order bits. separating a digital image into its bit planes is useful for analyzing the relative importance played by each bit of the image, implying, it determines the adequacy of numbers of bits used to quantize each pixel, useful for image compression. In terms of bit-plane extraction for a 8-bit image, it is seen that binary image for bit plane 7 is obtained by proceeding the input image with a threshold gray-level transformation function that maps all levels between 0 and 127 to one level (e.g 0)and maps all levels from 129 to 253 to another (eg. 255). Watermark Embedding Method: Step 1: Read the original Image X. Step 2: Compute the first level wavelet decomposition of original image X. We obtain one approximation image (LL Band) and three detail images, corresponding to the horizontal, vertical and diagonal details of the original host image . Step 3: Convert the LL band of original image into different 8 planes start with MSB to LSB.

Figure 3. Bit Plane Slicing to LL Band

Step 4: Generate the required watermark of text, image or random number. Step 5: Add the LSB plane of high frequencies with desired

watermark plane. Step 6: Reconstruct the image with adding all the 8 planes again with MSB to LSB. Step 7: Apply 1st level inverse to LL band and three detailed co-efficient to obtain the watermarked image. Step 8: Apply different attacks to the watermarked image to check the robustness. Watermark Extraction process: Step 1: Apply DWT to the watermarked image to get approximation and details co-efficient. Step 2: Apply bit plane slicing to the LL of the watermarked image to get 8 different planes start from LSB to MSB. Step 3: Apply DWT to the original image to have four sub bands of LL, LH, HL and HH.

Step 4: Apply bit plane slicing to LL band of original image to get different planes from LSB to MSB. Step 5: Subtract the LSB plane of the original image from the LSB plane of watermarked image to get watermark

IV. EXPERIMENTEL RESULTS The proposed algorithm is implemented and results are

obtained on MATLAB platform. The proposed algorithm has been compared with some another algorithm is proposed by Jun Tian [10] based on differential expansion. In this approach, whole image is divided into the blocks and then 1st level Integer wavelet transform is obtained to check expansibility of each block and generate secret key. If block is expansible then watermark is embedded otherwise it is not embed. At extraction part with secret key, the expansible blocks is decomposed by 1-level wavelet lifting scheme. This algorithm is also used for image authentication and ownership verification application. But due to division of image in blocks it gives less PSNR and high MSE. We have removed the concept of blocks and convert in different level of decomposition and then apply bit plane slicing to embedded the watermark. With proposed our approach, we overcome block artifacts issues. So we get good PSNR and MSE compare to existing approaches. We compared our results with differential expansion method results.

A. Performance of an Algorithm Before presenting the results, it is necessary that the factor, which determines the performance of an algorithm, is clearly stated. Essentially, the most important factor is Similarity factor which gives difference between the original image and the reconstructed image to achieve complete reversibility and others are Mean square error and Peak signal to noise ratio (db).

5

Original Image 1st level decomposition

Watermarked Image Reconstruction from plane Watermarked LSB Plane (a) Watermark Embedding Process

Watermarked Image Watermarked LSB Plane

Original Image LSB Plane of original Image Watermark

(b) Watermark Extraction process

Figure 4. Explanation of Praposed algorithm with Step by Step

B. Performance of an Algorithm Before presenting the results, it is necessary that the factor, which determines the performance of an algorithm, is clearly stated. Essentially, the most important factor is Similarity factor which gives difference between the original image and the reconstructed image to achieve complete reversibility and others are Mean square error and Peak signal to noise ratio (db).

C. Simulation Parameters The simulated models are set to the following parameters: Original Image: Lena (512X512) (greyscale,8 bpp) Baboon (512X512) (greyscale,8 bpp) Cameraman (512X512) (greyscale,8 bpp)

Bit Plane Slicing

DWT

IDWT

Watermark

DWT

DWT

A

A

6

D. Mean Square Error(MSE)

It is measured as the average of square of error introduce. It is defined as,

{ }21 1

0 0( , ) 1 ( , )M N

r cI r c I r c

M S EM N

− −

= =⎛ ⎞−⎜ ⎟=⎜ ⎟⎝ ⎠

∑ ∑

M x N is size of image. I(r,c) and I1(r,c) denotes row element “r” and column element “c” of Original and Watermarked image.

E. Peak Signal to Noise Ratio(PSNR)

It is a ratio of maximum signal power to noise power. PSNR = 10 log (2552 / MSE)

F. Similarity Factor(SF)

SF = ( )∑ ∑ ∑ ∑−

=

−

=

−

=

−

=′1

0

1

0

1

0

1

02 ),(/),(*),(M

r

N

c

M

r

N

ccrfcrfcrf

M x N is size of image. f(r,c) and f’(r,c) denotes row element “r” and column element “c” of original and reconstructed image.

TABLE I. COMPARISION ORIGINAL MSE,PSNR AND SIMILARITY FACTOR OF WATERMARKED AND ORIGINAL IMAGE

Image MSE PSNR SF Proposed Scheme

Lena 0.050 61.14 0.998 Cameraman 0.065 60.00 0.996 Baboon 0.072 59.55 0.995

Jun Tian[10] Lena 0.692 49.7274 0.990 Cameraman 1.643 45.9727 0.984 Baboon 1.260 47.1253 0.982

TABLE II. COMPARISION ORIGINAL MSE,PSNR AND SIMILARITY FACTOR OF WATERMARKED AND ORIGINAL IMAGE AFTER SALT AND PEEPER

NOISE ADDED TO WATERMARKED IMAGE Image MSE PSNR SF Proposed Scheme Lena 0.062 60.21 0.997

Cameraman 0.067 59.74 0.996 Baboon 0.078 59.20 0.996


TABLE III. COMPARISION ORIGINAL MSE,PSNR AND SIMILARITY FACTOR OF WATERMARKED AND ORIGINAL IMAGE AFTER 180 DEGREE

CLOCKWISE ROTATION TO WATERMARKED IMAGE

Image MSE PSNR SF Proposed Scheme Lena 10.60 37.87 0.973



TABLE IV. COMPARISION ORIGINAL MSE,PSNR AND SIMILARITY FACTOR OF WATERMARKED AND ORIGINAL IMAGE AFTER RESIZING OF

WATERMARKED IMAGE

Image MSE PSNR SF Proposed Scheme Lena 75.0 29.38 0.989



V. CONCLUSION The proposed algorithm has combined approaches of spatial and frequency domain. The results are verified on MATLAB platform against various attacks. The proposed scheme has been compared with differential expansion schemes. We have compared results of MSE and PSNR for watermarked and original images. From Table I-IV, it is conclude that proposed scheme has better robustness to withstand various geometrical attacks. It has overcome the problem of block artifacts and advantage of both domain properties. We embed the watermark in just adding in the LSB plane and achieved the higher PSNR results. It has also lesser computational complexity compare to other algorithms because we embed the watermark in 1st decomposition level.

REFERENCES [1] R.G. Van Shindig, A.Z.Tirkel and C. F. Osborne, “A Two-Dimensional

Digital Watermark,” in proc. DICTA, pp. 378-393,1993. [2] S. Arivazhagan and L. Ganesan, “Texture Segmentation Using Wavelet

Transform,” Pattern recognition Letters, Vol. 24,pp 3197-3203,December-2003

[3] Saraju P. Mohanty, N. Raganathan, Ravi K. Namballa, “VLSI Implementation of Visible Watermarking for a Secure Digital Camera Design,” Proceeding of 17th International Conference on VLSI design, 2004.

[4] A. Gavlasova,A. Prochazka, and M. Mudrova, “Wavelet Use for Image Classification,” 15th International Conference on Process Control,2005

[5] Sammy H. M. Kwok and Edmund Y. Lam, “Watermark implementation in digital photography,” Processing Of 2005 International Symposium on Intelligent Signal Processing and Communication Systems, Dec-2005.

[6] Cong Jin, Liang-Gang Pan, Ting Su, “A Blind Watermarking Scheme Based on Visual Model for Copyright Security,” MCAM,pp 454-463,2007.

[7] M.A.EI-Iskandarani,Saad M. Saad, Ahmed M. Abubahiam, “An Efficient Digital Image Watermarking,” ICCST,2008

[8] W. bender, d. Gruhi and N. Morimoto,“Techniques for data hiding,” IBM syst. J., vol. 35,nos. 3-4 pp. 313-336,1996.

[9] C. Fei, D. Kundur, and R.H. Kwong, “The choice of watermark domain in the presence of compression”, in Proc. IEEE Int. Conf. on Infermation Technology: Coding and Computing,pp79-84,April 2001.

[10] Jun Tain, “ Reversible Watermarking by Difference expansion” Multimedia and Security workshop at ACM-02 France,Dec-06,2002.

CLOCK SKEW ANALYSIS AND OPTIMIZATION OF SEQUENTIAL

CIRCUIT

Amit M Joshi *, Anand D. Darji , Taral D Chhatbar , Abhishek Aghrahari SARDAR VALLABHAI NATIONAL INSTITUTE OF ENGINEERING & TECHNOLOGY

SURAT-395007; e-mail: [email protected]

Abstract. A well-designed clock system is a

fundamental requirement in high-speed computers.

As the demand for high-speed computers grows, the

design of their clock systems have become a major

concern not only in achieving high performance, but

also in reducing assembly and maintenance costs. This

paper gives an idea about important issues of clock

skew which gives more focus in today research area.

As technology advances and timing cycle reduces it

becomes an important parameter of sequential design.

In this paper analysis and optimization techniques for

clock skew is discussed. We have also focused on

double clocking & zero clocking constraints for

optimization. To overcome these constraints clock

period minimization is required. H tree is one of the

standard technique which is used in clock distributed

network to reduce the clock skew. Approaches like

retiming and transistor sizing techniques are

discussed for skew optimization.

Key words: Gufe skew; Double clocking; Zero

clocking; H- Tree.

1. INTRODUCTION

We specify system timing such that every system

memory element has an expected arrival time for

the active edge of its clock signal. Clock-edge

inaccuracy is the difference between the actual and

expected arrival time of this clock edge. We define

path clock skew as the sum of the clock-edge

inaccuracies of the pair’s source and destination.

System clock skew is the largest path clock skew in

the system. It is the value of the worst-case timing

inaccuracy among all paths. We can break it into

inter board skew, on-board inter chip skew, and so

on to the smallest timed component. The challenge

to designers of clock-distribution networks is how

to control system clock skew so that it becomes an

acceptably small fraction of the system clock

period. As a rule, most systems cannot tolerate a

clock skew of more than 10% of the system clock

period. If system clock skew goes beyond the

design limit, system behavior is affected in terms of

setup and hold times, which results in long and

short paths. No scheme is immune from these

problems-even flip-flop machines can malfunction

when clock skew is present. Clock-powering trees

are a source of clock. We can minimize the clock

skew by placing all the gates at a given tree level,

or delay which is the difference between its hest-

case and worst-case delays. This difference is called

gufe skew. In addition,we can realize elements at

each tree level by using electrically matched

devices and careful wiring.

2. IS CLOCK SKEW OPTIMIZATION SAFE?

A common misconception about changing clock

skews is believed to be an “unsafe” optimization. A

small change in either gate or interconnect delays

may cause a circuit with hazardously small

tolerance to malfunction. But one can build in

safety margins, to ensure skewing errors do not

disrupt circuit functionality. These safety margins

ensure that the circuit will operate in the presence

of unintentional process-dependent skew variations.

Introducing deliberate delays within the clocking

network has been a tactic that has long been used

by designers [1]. From the design point of view,

zero skew clock tree construction techniques may

be adapted to build fixed skew clock networks. A

practical proof of this concept was demonstrated in

a pipelined data buffer chip using the concept of

skewed clocks, which was previously designed and

fabricated [2].

Register A Register B

Fig. 1 The concept of zero clocks skew

In fact it is a misconception to believe that zero

skew is entirely safe. To see this, consider a shift

register consisting of register A whose output is

connected to register B with no combinational logic

between them as shown in the Fig. 1 [3,4]. Even

for a circuit designed for a skew, a small

unintentional positive skew at register B will cause

double clocking i.e. a short path constraint

violation. Such problem may be avoided by the use

of safety margins with introducing the delibrate

skew to the circuit. A small amout of delibrate

positive skew at register A provides an effective

safety margin aginst short path violation.

D Q

D Q

clk

3. CLOCK SKEW OPTIMIZATION

Use of intentional non-zero clock skews can speed

up the clock period for a sequential circuit since it

facilitates cycle-borrowing. Now its time to find the

optimal skews at each memory element in an edge-

triggered circuit, where the optimality is defined in

terms of minimizing the clock period [5]. The

problem is posed as a linear program, and this

formulation is described in this section.

Fig. 2 Ccombinational block in sequential circuit

Consider a combinational block in a sequential

circuit, as shown in Fig. 2. Let FFi and FFj be a pair

of flip flops at the input and output of the

combinational block respectively. Where Xi and Xj

are clocks for FFi and FFj respectively. Denote the

delay of the combinational block between them as

),( jid . d(i ,j) being the minimum delay and

),( jid being the maximum delay. If Thold is the

flip-flop hold time, Tset-up the flip-flop set-up time,

and P the clock period, then in the presence of

skews the double-clocking and zero-clocking

constraints are discussed as below..

Double-clocking/short path constraint:

To avoid double-clocking, the data from the next

clock cycle should arrive at FFj no earlier than Thold

after the current clock. The data leaves FFi at time

Xi and reaches FFj at Xi + d(i,j). Thus, the double

clocking constraint illustrated in Fig. 3.

Fig. 3 Double clocking

Zero-clocking/ long path constraint:

To avoid zero-clocking, the data from the current

clock cycle should arrive at FFj no later than a time

Tsetup before the next clock. Since the data leaves

FFi at time Xi, the latest time by which it will reach

FFj is Xi+ ),( jid . The zero-clocking constraint

shown in Fig. 4 may be expressed as

setupTPXjjidXi ),(

Fig 4. Zero clocking

4. CLOCK PERIOD MINIMIZATION

The problem of minimizing the clock period P by

controlling the clock skew at each flip-flop, subject

to correct timing, can now be formulated as the

following linear program. Minimize P Such that ,

holdTXjjidXi ),(

setupTPXjjidXi ),(

The linear program assumes that the delays of the

combinational segments are constant.The above

formulation may be altered to model uncertainties

in the clock skews. If it can be guaranteed that in

the manufactured circuit, the skew at flip-flop k will

be within the range [Xk- /2 , Xk + /2]. Where Xk is

the designed value of the skew. Then the difference

between any skews (Xi – Xj) in the manufactured

circuit, must be within of the designed value of

(Xi – Xj). A linear program that is guaranteed to

produce a functional manufactured circuit is

formulated below.

Minimize P Such that,

holdTXjjidXi ),(

setupTPXjjidXi ),(

5.EFFICIENT SOLUTION OF SKEW

OPTIMIZATION PROBLEM

The constraint of the linear program are rewritten as

),( jidTXjXi hold

PjidTXiXj setup ),(

Note that the skews at the primary inputs and the

primary outputs may be set to zero under the

assumption that they cannot be controlled or

adjusted. For a constant value of P, the constraint

matrix for this problem reduces to a system of

difference constraints [6]. For such a system, it is

possible to construct a constraint graph. The

Xi Xj

FFi

CK

Combinational

FFj

CK

Xj

Xi

holdTXjjidXi ),(

Thold d (i,j)

Xi

Xj

Input Latch

Clock

Output Latch

Clock

Output Data

Input Latch

Clock

Output

Latch

Clock

Output

Data

d (i,j) Tsetup

vertices of the constraint graph correspond to the Xi

variables, and each constraint Xa-Xb > C

corresponds to a directed edge from node b to node

a with a weight of c. If P is achievable, then a set of

skews that satisfies P can be found by solving the

longest path problem on this directed graph. The

clock period P is feasible, provided the

corresponding constraint graph contains no positive

cycles. This observation was utilized in for efficient

solution of the skew optimization problem [7]. The

optimal clock period is obtained by performing a

binary search. The procedure continues until the

smallest feasible clock period is found. The

following skeletal pseudo code is used to perform

the binary search for the optimal clockperiod.

Algorithm:

Construct the Constraint graph;

Pmin= Plow ;

Pmax=Phigh;

While (Pmax – Pmin) >

{

P= ( Pmax + Pmin)/2;

If constraint graph has a positive cycle

Pmin = P;

Else

Pmax = P;

}

6. CLOCK DISTRIBUTED NETWORK

Clock distribution network is required to propagate

the clock signal to every register. It must be

designed carefully to achieve performance

requirements on clock signal at sink nodes.

Objective of the clock tree distibution methods are

to achieve either zero skew or fixed skew. H tree

stucture is classical method for clock distibution

network design as shown in Fig 5.

Fig.5 H Tree Structure

The H tree ensures that all sinks of the clock tree

are equidistant from the clock source. The H-tree

works well for regular structure such as systolic

arrays to reduce skew. In practice, asymmetric

distribution of clock pins are very much common

hence H-tree may not be effective for clock routing.

To address this issue, the method of recursive

geometric matching was proposed [8]. It is an

extension of the H-tree idea to structures with a

smaller degree of regularity. Large high

performance circuits have internal clock that is the

function of external clock. The internal clock cycle

time comprises of several components, such as

delay associated with storage devices on the chip,

clock skew & signal transmission etc. The internal

clock signal is distributed to the circuits through

distributed network. The clock skew within a chip

is the variation in the internal clock signal of the

circuit. The main contributor to the skew are

RC transmission delay variations, device variations,

and localized loading varations. In the past, clock

skew is the only small portion of clock time. But as

cycle time decreases, the propagation of clock skew

to the cycle time has dramatically increased. Clock

skew introduces propogation delay between internal

and external clock, which makes interfacing

difficult. Current techniques to control clock skew

focus on controlling the RC variation by equalizing

the routing loading and/or distance. Loading

variation is controlled by adding the dummy loads

to equalize load capacitance of the distribution

network. Effectivness of these techniques depend

on precision by which they equalize load and

distance.

7. CLOCK SKEW OPTIMIZATION AND

RETIMING

In a sequential circuit, due to difference in

interconnect delays on the clock distribution

network, clock signal do not arrive at all of the FF’s

at the same time.There is a skew between the clock

arrival times at different FF’s.Several methods for

achiving zero clock skew have been developed.

Viewing clock skew as a managable resours rather

then a liability allows a designer to introduce

intentional skews to improve the performance of the

circuit. To illustrate this, consider the example in

Fig 6 where each inverter has a unit delay. The

combinational circuit blocks CC1 and CC2 have

delays of 3.0 and 1.0 units, respectively, therefore

the fastest allowable clock has a period of 3.0 units.

However if a skew of +1.0 unit is applied to the

clock line to FF B, the circuit can run with a clock

period of 2.0 units.

Fig. 6 Nonzero clock skew

A second approach to timing optimization is the

procedure of retiming [1]. Retiming involves the

relocation of FF’s across logic gates to allow the

circuit to be operated under a faster clock without

changing its functionality. For the circuit in Fig 6.

the period of the original circuit can be minimized

to 2.0 units by moving FF B to the left across the

FF

A

ck

FF

B

ck

FF

C

ck

CC1 CC2

I1 I2 I3 I4

inverter I3. This results in the combinational circuit

blocks CC1 & CC2 having delays of 2.0 units.

Fig.7 Retiming for clock period minimization

8. OPTIMAZATION WITH TRANSISTOR SIZING

It has been assumed that the delay of each

combinational path is fixed .However, the drive

strength of gates can be altered by transistor sizing

[9]. Thereby changing the combinational path

delays. To illustrate this, consider a pair of

combinational stages separated by latches, as shown

in Fig 8. We will consider area as the cost function

here, but the same idea is valid for other correlated

cost functions such as power. If only the sizing

transformation were to be applied, the nature of the

cost-delay tradeoff curve for each stage would be as

shown in Fig 9.If the timing specification on Stg1

were stringent, then the optimal sizing solution

would lie beyond the knee of the curve, at a point

such as point A. Similarly, if the specifications on

Stg2 were loose, then the optimal solution might lie

at a point such as C. The cost of sizing solution

would then be [Area(A) + Area(C)].

Fig 8 Non zero clock skew

Now consider a situation where a small skew S, is

applied to the registers between the two stages,

allowing Stg1 to borrow time from Stg2.This would

lead to a looser specification on Stg1, bringing the

solution point from A to B, and a tighter

specification for stage2, moving its solution from C

to D. The cost of the overall sizing solution, [Area

(B) + Area(D)] can be seen to be much smaller than

the cost with zero skew.For a small amount of cycle

borrowing, the cost at the solution point A was well

above the knee of the curve. Therefore, the

judicious use of sizing in conjunction with

deliberate skew can deliver solutions of

significantly better quality than sizing alone. It is

important to perform the two optimizations

concurrently, rather than sequentially, So

to fully incorporate the mutual interactions between

skew and sizing.

Fig 9 Sizing in conjunction with clock skew

9. CONCLUSION

The clock system is an integral part of synchronous

computers, yet it is not a widely studied aspect of

their design. Early attention to system timing issues

can provide benefits in system performance as well

as product development time. The goal of the clock

system designer is to control system clock skew at

the system operating frequency as well as to

minimize electrical hazards that may add

undesirable components to the clock signals.

10. REFERENCE

(1) K.D. Wagner. Clock system design.IEEE Design

and test of Computers, pages 9-27, October 1988.

(2) M. Heshami and B.A. Wooley. A 250 MHz skewed

–clock pipelined data buffer. IEEE Journal of

Solid-State Circuits, 31(3):376-383,March 1996.

(3) H.Sathyamurthy, S.S. Spanker, and J. P. Fish

burn. Speeding up piplined circuits through a

combinational of gate sizing and clock skew

optimization. In proceeding of the IEEE/ACM

International Conference on Computer-Aided

design, pages 467-470.1995

(4) H.Sathyamurthy, S.S. Spanker, and J. P. Fish

burn. Speeding up piplined circuits through a

combinational of gate sizing and clock skew

optimization. IEEE Transactions on Computer-

Aided Design of Integrated Circuits and systems,

17(2):173-182, February 1998.

(5) J. P. Fish burn. Clock skew optimization.IEEE

Transaction on Computers, 39 (7) :945-951, July

1990

(6) T.H.Coremen, C.E. Leiserson and R.L.Rivest.

Introduction to Algorithms. McGraw-Hill, New

York,NY, 1990

(7) R.B. Deokar and S.S. Spanker. A graph-theoretic

approach to clock skews optimization. In

Proceeding of the IEEE International Symposium

on Circuits and systems, pages 1.407-1.410,1994

(8) A.Kahang,J.Cong, and G.Robins.High-

performance clock routing based on recursive

geometric matching.In Proceeding of the

ACM/IEEE Design Automation Conference,pages

322-327,1991

(9) S.S. Sapatnekar and S.M. Kang. Design

automation for timing driven Layout synthesis.

Kluwer Academic Publishers, Boston, MA, 1993.

F

F

1

Stage 1

Stg1

Stage 2

Stg2

F

F

2

F

F

3

FF

A

CK

FF

B

CK

FF

C

CK

CC1 CC2

I1 I2 I3 I4

APPENDIX B

Lifting Scheme In traditional convolution (filtering) based approach for computation of the forward

DWT, the input signal (x) is filtered separately by a low-pass filter (~h ) and a high-pass

filter (~g ). The two output streams are then sub-sampled by simply dropping the alternate

output samples in each stream to produce the low-pass (L) and high-pass (H) sub band

outputs as shown in Fig 1. The two filters (~h ,

~g ) form the analysis filter bank. The

original signal can be reconstructed by a synthesis filter bank (h, g) starting from L and H

as shown in Fig 2.

Fig 1 Analysis and Synthesis Filter Bank

The condition for the perfect reconstruction of a signal are given by equations [1]-[2].

~ ~1 1( ) ( ) ( ) ( ) 2h z h z g z g z− −+ = [1]

~ ~1 1( ) ( ) ( ) ( ) 0h z h z g z g z− −− + − = [2]

Where ( )h z is the z-transform of the FIR (Finite Impulse Response) filter h . h can be

expressed as a Laurent polynomial of degree p as

0( )

pi

ii

h z h z−

=

=∑ [3]

By splitting the filter coefficients into odd and even parts, the filters can be rewritten as

follows.

2 1 2( ) ( ) ( )e oh z h z z h z−= + [4]

2 1 2( ) ( ) ( )e og z g z z g z−= + [5]

~ ~ ~2 1 2( ) ( ) ( )e oh z h z z h z−= + [6]

~ ~ ~2 1 2( ) ( ) ( )e og z g z z g z−= + [7]

The corresponding polyphase representation is [34]:

( ) ( )( )

( ) ( )e e

o o

h z g zP z

h z g z⎡ ⎤

= ⎢ ⎥⎣ ⎦

and ~ ~

~

~ ~

( ) ( )( )

( ) ( )

e o

e o

h z h zP z

g z g z

⎡ ⎤⎢ ⎥=⎢ ⎥⎣ ⎦

Often ( )P z is unity is called dual of ~

( )P z and for perfect reconstruction, they are related

as ~

1( ) ( )TP z P z I− = ,

where I is a 2x2 identity matrix. Now wavelet transform in terms of polyphase matrix

can be expressed as equations (3.8)-(3.9) for forward DWT and inverse DWT

respectively [34].

~

1

( ) ( )( )

( ) ( )e

o

L z x zP z

H z z x z−

⎡ ⎤ ⎡ ⎤=⎢ ⎥ ⎢ ⎥

⎣ ⎦ ⎣ ⎦ [8]

1

( ) ( )( )

( ) ( )e

o

x z L zP z

z x z H z−

⎡ ⎤ ⎡ ⎤=⎢ ⎥ ⎢ ⎥

⎣ ⎦ ⎣ ⎦ [9]

If the determinant of ( )P z is unity, it can be shown by applying Cramer’s rule [34] that

~1 1( ) ( )h z z g z−= − −

~1 1( ) ( )g z z h z− −=

~1 1( ) ( )h z z g z− −= − −

~1 1( ) ( )g z z h z− −= − .

Daubechies and Sweldens in [34] have shown that the polyphase representation can

always be factored into lifting steps by Euclidean algorithm to find the greatest common

divisor. Thus the polyphase representation becomes:

1

1 ( ) 1 0( )

0 1/ 0 1 ( ) 1

n i

i i

K O a zP z

K b z=

⎡ ⎤ ⎡ ⎤ ⎡ ⎤= Π⎢ ⎥ ⎢ ⎥ ⎢ ⎥⎣ ⎦ ⎣ ⎦ ⎣ ⎦

[10]

Where K is a constant and acts as a scaling factor. ( )ia z and ( )ib z are the Laurent

polynomials which correspond to the updaters and the predictors of the lifting steps. By

factoring the wavelet filters into lifting steps, it is expected that the computation

performed on each stage will be much less complex.

We can say, the main feature of the lifting based DWT scheme is to break up the

high pass and low pass filters into a sequence of upper and lower triangular matrices and

convert the filter implementation into banded matrix multiplications. Lifting based

approach has several advantages over conventional convolution based approach [34].

Lifting Architecture

Lifting scheme can be mainly divided into three parts as shown in the Fig 2 splitting,

lifting and scaling.

Split Step: The input signal x[n] is split into odd and even subsequences, [ ]d n and [ ]s n

respectively (also known as Lazy Wavelet Transform).

2i is x∈

2 1i id x +∈

Predict Step: This step predicts data in the subsequence [ ]d n using the samples in the

[ ]s n and replaces the samples in [ ]d n using prediction error.

Fig 2 Basic Lifting Scheme

[ ] [ ] ( [ ])d n d n P s n= −

()P = Prediction Operator

Update Step: This step updates the data in [ ]s n using the data in [ ]d n .

[ ] [ ] ( [ ])s n s n U d n= +

()U =Update Operator

APPENDIX C

TEST IMAGES

Gray Scale Images:

Lena (256x256) Baboon (256x256) Cameraman(256x256)

Color Images:

Lena (256x256) Baboon (256x256) Cameraman(256x256)

Watermarked Images

Text Logo (64 x 64) Text Logo (64 x 64) Random Sequence (64x64)

Documents

Amit Thesis