Discrete Wavelet Transforms - Algorithms and Applications PDF

DISCRETE WAVELET
TRANSFORMS:
ALGORITHMS AND
APPLICATIONS
Edited by Hannu Olkkonen
Discrete Wavelet Transforms: Algorithms and Applications
Edited by Hannu Olkkonen
Published by InTech
Janeza Trdine 9, 51000 Rijeka, Croatia
Copyright © 2011 InTech

All chapters are Open Access articles distributed under the Creative Commons
Non Commercial Share Alike Attribution 3.0 license, which permits to copy,
distribute, transmit, and adapt the work in any medium, so long as the original
work is properly cited. After this work has been published by InTech, authors
have the right to republish it, in whole or part, in any publication of which they
are the author, and to make other personal use of the work. Any republication,
referencing or personal use of the work must explicitly identify the original source.
Statements and opinions expressed in the chapters are these of the individual contributors
and not necessarily those of the editors or publisher. No responsibility is accepted
for the accuracy of information contained in the published articles. The publisher
assumes no responsibility for any damage or injury to persons or property arising out
of the use of any materials, instructions, methods or ideas contained in the book.
Publishing Process Manager Ivana Lorkovic

Technical Editor Teodora Smiljanic
Cover Designer Jan Hyrat
Image Copyright Scott Bowlin, 2010. Used under license from Shutterstock.com
First published August, 2011

Printed in Croatia
A free online edition of this book is available at www.intechopen.com

Additional hard copies can be obtained from [email protected]
Discrete Wavelet Transforms: Algorithms and Applications, Edited by Hannu Olkkonen

p. cm.
ISBN 978-953-307-482-5
free online editions of InTech
Books and Journals can be found at
www.intechopen.com
Contents
Preface IX
Part 1 Discrete Wavelet Transform Based Hardware Algorithms 1
Chapter 1 Discrete Wavelet Multitone Modulation

for ADSL & Equalization Techniques 3
Sobia Baig, Fasih-ud-Din Farrukh and M. Junaid Mughal
Chapter 2 A Scalable Architecture for Discrete

Wavelet Transform on FPGA-Based System 25
Xun Zhang
Chapter 3 VLSI Architectures of Lifting-Based

Discrete Wavelet Transform 41
Sayed Ahmad Salehi and Rasoul Amirfattahi
Chapter 4 Simulation of Models and BER

Performances of DWT-OFDM versus FFT-OFDM 57
Khaizuran Abdullah and Zahir M. Hussain
Chapter 5 Several Kinds of Modified SPIHT Codec 67

Wenchao Zhang
Part 2 Image Processing Applications 79
Chapter 6 Multiresolution Approaches for Edge Detection and

Classification Based on Discrete Wavelet Transform 81
Guillermo Palacios, J. Ramón Beltrán and Raquel Lacuesta
Chapter 7 Low Bit Rate Video Compression Algorithm

Using 3-D Discrete Wavelet Decomposition 101
Awad Kh. Al-Asmari
Chapter 8 Low Complexity Implementation of Daubechies

Wavelets for Medical Imaging Applications 121
Khan Wahid
VI Contents
Chapter 9 Discrete Wavelets on Edges 135

Alexandre Chapiro, Tassio Knop De Castro,
Virginia Mota, Eder De Almeida Perez,
Marcelo Bernardes Vieira and Wilhelm Passarella Freire
Chapter 10 Discrete Wavelet Transform and Optimal Spectral

Transform Applied to Multicomponent Image Coding 151
Isidore Paul Akam Bita, Michel Barret, Florio Dalla Vedova,
Jean-Louis Gutzwiller and Dinh-Tuan Pham
Part 3 Discrete Wavelet Transforms for Watermaking 177
Chapter 11 Watermarking-Based Image Authentication

System in the Discrete Wavelet Transform Domain 179
Clara Cruz Ramos, Rogelio Reyes Reyes,
Mariko Nakano Miyatake and Héctor Manuel Pérez Meana
Chapter 12 Application of Discrete

Wavelet Transform in Watermarking 197
Corina Nafornita and Alexandru Isar
Part 4 Discrete Wavelet Transform Algorithms 219
Chapter 13 Shift Invariant Discrete Wavelet Transforms 221

Hannu Olkkonen and Juuso T. Olkkonen
Chapter 14 Condition on Word Length of Signals and

Coefficients for DC Lossless Property of DWT 231
Masahiro Iwahashi and Hitoshi Kiya
Chapter 15 Wavelet-Based Analysis and

Estimation of Colored Noise 255
Bart Goossens, Jan Aelterman, Hiêp Luong,
Aleksandra Pižurica and Wilfried Philips
Chapter 16 An Adaptive Energy Discretization

of the Neutron Transport Equation
Based on a Wavelet Galerkin Method 281
D. Fournier and R. Le Tellier
Preface
The discrete wavelet transform (DWT) algorithms have a firm position in processing
of signals in several areas of research and industry. As DWT provides both octave-
scale frequency and spatial timing of the analyzed signal, it is constantly used to solve
and treat more and more advanced problems. The DWT algorithms were initially
based on the compactly supported conjugate quadrature filters (CQFs). However, a
drawback in CQFs is due to the nonlinear phase effects such as spatial dislocations in
multi-scale analysis. This is avoided in biorthogonal discrete wavelet transform
(BDWT) algorithms, where the scaling and wavelet filters are symmetric and linear
phase. The BDWT algorithms are usually constructed by a ladder-type network called
lifting scheme. The procedure consists of sequential down and uplifting steps and the
reconstruction of the signal is made by running the lifting network in reverse order.
Efficient lifting BDWT structures have been developed for VLSI and microprocessor
applications. Only register shifts and summations are needed for integer arithmetic
implementation of the analysis and synthesis filters. In many systems BDWT-based
data and image processing tools have outperformed the conventional discrete cosine
transform (DCT) -based approaches. For example, in JPEG2000 Standard the DCT has
been replaced by the lifting BDWT.
A difficulty in multi-scale DWT analyses is the dependency of the total energy of the
wavelet coefficients in different scales on the fractional shifts of the analysed signal.
This has led to the development of the complex shift invariant DWT algorithms, the
real and imaginary parts of the complex wavelet coefficients are approximately a
Hilbert transform pair. The energy of the wavelet coefficients equals the envelope,
which provides shift-invariance. In two parallel CQF banks, which are constructed so
that the impulse responses of the scaling filters have half-sample delayed versions of
each other, the corresponding wavelet bases are a Hilbert transform pair. However,
the CQF wavelets do not have coefficient symmetry and the nonlinearity disturbs the
spatial timing in different scales and prevents accurate statistical analyses. Therefore
the current developments in theory and applications of shift invariant DWT
algorithms are concentrated on the dual-tree BDWT structures.
This book reviews the recent progress in discrete wavelet transform algorithms and
applications. The book covers a wide range of methods (e.g. lifting, shift invariance,
multi-scale analysis) for constructing DWTs. The book chapters are organized into
X Preface
four major parts. Part I describes the progress in hardware implementations of the
DWT algorithms. Applications include multitone modulation for ADSL and
equalization techniques, a scalable architecture for FPGA-implementation, lifting
based algorithm for VLSI implementation, comparison between DWT and FFT based
OFDM and modified SPIHT codec. Part II addresses image processing algorithms such
as multiresolution approach for edge detection, low bit rate image compression, low
complexity implementation of CQF wavelets and compression of multi-component
images. Part III focuses watermaking DWT algorithms. Finally, Part IV describes shift
invariant DWTs, DC lossless property, DWT based analysis and estimation of colored
noise and an application of the wavelet Galerkin method.
The chapters of the present book consist of both tutorial and highly advanced material.
Therefore, the book is intended to be a reference text for graduate students and
researchers to obtain state-of-the-art knowledge on specific applications. The editor is
greatly indebted to all co-authors for giving their valuable time and expertise in
constructing this book. The technical editors are also acknowledged for their tedious
support and help.
Hannu Olkkonen, Professor

University of Eastern Finland, Department of Applied Physics, Kuopio,
Finland
Part 1
Discrete Wavelet Transform Based

Hardware Algorithms
1
Discrete Wavelet Multitone Modulation for

ADSL & Equalization Techniques
Sobia Baig1, Fasih-ud-Din Farrukh2 and M. Junaid Mughal2
1Electrical
Engineering Department,
COMSATS Institute of Information Technology, Lahore
2Faculty of Electronic Engineering,
GIK Institute of Engineering Sciences and Technology, Topi

1,2Pakistan
1. Introduction
The reliable delivery of information over severe fading wireless or wired channels is a major
challenge in communication systems. At the heart of every communication system is the
physical layer, consisting of a transmitter, a channel and a receiver. A transmitter maps the
input digital information into a waveform suitable for transmission over the channel. The
communication channel distorts the transmitted waveform. One of the many sources of
signal distortion is the presence of multipath in the communication channel. Due to the
effect of the multipath signal propagation, inter-symbol interference (ISI) occurs in the
received waveform. Moreover, the transmitted signal gets distorted due to the effect of
various kinds of interference and noise, as it propagates through the channel. ISI and the
channel noise distort the amplitude and phase of the transmitted signal, which lead to
erroneous bit detection at the receiver. It is desirable for a good communication system that
its receiver is able to retrieve the digital information from the received waveform, even in
the presence of channel impairments such as, multipath effect and noise.
Orthogonal Frequency Division Multiplexing (OFDM) is a Multi-Carrier Modulation
(MCM) technique that enables high data rate transmission and is robust against ISI
(Saltzberg, 1967), (Weinstein and Ebert, 1971), (Hirosaki, 1981). It is a form of frequency
division multiplexing (FDM), where data is transmitted in several narrowband streams at
various carrier frequencies. The sub-carriers in an OFDM system are orthogonal under ideal
propagation conditions. By dividing the input bit-stream into multiple and parallel bit-
streams, the objective is to lower the data rate in each sub-channel as compared to the total
data rate and also to make sub-channel bandwidth lower than the coherence bandwidth of
the communication channel. Therefore, each sub-channel will experience flat-fading and will
have small ISI. Hence an OFDM system requires simplified equalization techniques, to
mitigate the inter-symbol interference. The ISI can be completely eliminated in OFDM
transceivers by utilizing the principle of cyclic prefixing (CP). Therefore, high data rate
communication systems prefer to apply multicarrier modulation techniques. OFDM has
been standardized for many digital communication systems, including ADSL, the 802.11a
and 802.11g Wireless LAN standards, Digital audio broadcasting including EUREKA 147
4 Discrete Wavelet Transforms: Algorithms and Applications
and Digital Radio Mondiale, Digital Video Broadcasting (DVB), some Ultra Wide Band
(UWB) systems, WiMax, and Power Line Communication (PLC) (Sari, et al., 1995)
(Frederiksen and Prasad, 2002), (Baig and Gohar, 2003).
Over the years, OFDM has evolved into variants, such as Discrete Multitone (DMT), and
hybrid modulation techniques, such as multi-carrier code division multiple access (MC-
CDMA), Wavelet OFDM and Discrete Wavelet Multitone (DWMT). Several factors are
responsible for the development of these variants, especially Wavelet based OFDM
techniques, which target several disadvantages associated with Multicarrier modulation
(MCM) techniques. Some of these drawbacks are:
 the spectral inefficiency associated with the guard interval insertion, which includes the
cyclic prefix
 the high degree of spectral leakage due to high magnitude side lobes of pulse shape of
sinusoidal carriers
 OFDM based communication system’s sensitivity to inter-carrier interference (ICI) and
narrowband interference (NBI)
Therefore, a Discrete Wavelet Transform (DWT) based MCM system was developed as an
alternative to DFT based MCM scheme (Lindsey, 1995). DWT based MCM techniques came
to be known as Wavelet-OFDM in wireless communications and as Discrete Wavelet
Multitone (DWMT) for harsh and noisy wireline communication channels such as Digital
Subscriber Line (DSL) or Power Line Communications (PLC) (Baig and Mughal, 2009).
This chapter describes the application of DWT in Discrete Multitone (DMT) transceivers and
its performance analysis in Digital Subscriber Line (DSL) channel, in the presence of
background noise, crosstalk etc. Time domain equalization techniques proposed for DWT
based multitone that is DWMT are discussed, along with the simulation results. The pros
and cons of adopting DWT instead of DFT in DMT transceivers will also be discussed,
highlighting the open areas of research.
2. Basics of wavelet filter banks & multirate signal processing systems

Wavelets and filter banks play an important role in signal decomposition into various
subbands, signal analysis, modeling and reconstruction. Some areas of DSP, such as audio
and video compression, signal denoising, digital audio processing and adaptive filtering are
based on wavelets and multirate DSP systems. Digital communication is a relatively new
area for multirate DSP applications. The wavelets are implemented by utilizing multirate
filter banks (Fliege, 1994). The discovery of Quadrature Mirror Filter banks (QMF) led to the
idea of Perfect Reconstruction (PR), and thus to subband decomposition. Mallat came up
with the idea of implementing wavelets by filter banks for subband coding and
multiresolution decomposition (Mallat, 1999). DWT gives time-scale representation of a
digital signal using digital filtering techniques. The DWT analyzes the signal at different
frequency bands with different resolutions by decomposing the signal into approximation
and detail coefficients. The decomposition of the signal into different frequency bands is
obtained simply by successive high pass and low pass filtering of the time domain signal.
2.1 Analysis and synthesis filter banks

Analysis filter banks decomposes input signal into frequency subbands. A two channel
analysis filter bank, as shown in Fig. 1, splits the input signal ( ) into a high frequency
Discrete Wavelet Multitone Modulation for ADSL & Equalization Techniques 5
component ( ) and a low frequency component ( ). The input signal ( ) is passed

through a low pass filter ( ) and a high pass filter ( ), yielding the ( ) and ( )
respectively.
Fig. 1. Two-channel Analysis filter bank.

Consequently, with the sampling frequency, = 2 , the available bandwidth from 0 to , is
divided into two halves, 0 ≤ ≤ 4 for the lower frequency signal ( )and 4≤ ≤
2 for the high frequency signal ( ). Therefore, the filtered signals ( )and ( )have
half the bandwidth of the input signal after being convolved with the low pass filter and
high pass filter respectively.The filtered and downsampled signal spectra are shown in Fig.
2. In matrix form the sub-band signals are represented as (Fliege, 1994),
( ) ( ) (− ) ( )
= (1)
( ), ( ) (− ) (− )
The two signal spectra overlap. The downsampling will produce aliased components of the
signals, that are functions of (− / ) in Eq. 1, since the filtered signals are not bandlimited
to . Two-channel synthesis filter bank is the dual of analysis filter bank, as shown in Fig. 3.
( ) and ( ) denote the lowpass and highpass filters, which recombine the upsampled
signals ( ) and ( ) into ( ), the reconstructed version of the input signal. The aliased
images are removed by the filter ( ) in the frequency range 4 ≤ ≤ 2, while the
filter ( ) eliminates the images in the upsampled signal ( ) in the frequency range
0 ≤ ≤ 4. Therefore, the signal ( ), output from the synthesis filter bank is (Fliege,
1994),
( )
( )= ( ) ( ) (2)
( )
/2 
Fig. 2. (Continued)
/2 

/2 
/2 
/2  2
 2
Fig. 2. Signal spectra in two-channel analysis filter bank. (a) Low pass & high pass filter
transfer functions. (b) low pass filtered signal spectrum ( ). (c) high pass filtered signal
spectrum ( ). (d) downsampled signal ( )spectrum. (e) downsampled signal ( )
spectrum (f) output signal spectra.
Fig. 3. Two-channel Synthesis filter bank.
2.2 Quadrature mirror filter bank

The analysis and the synthesis filter banks combine to form a structure commonly known as
the two-channel quadrature mirror filter (QMF) bank. QMF bank serves as the basic
building block in many multirate systems. A two-channel QMF bank is shown in Fig. 4. The
constituent analysis and synthesis filter banks have power complementary frequency
responses. The low pass and high pass filters in the analysis filter bank decompose the input
signal into sub-bands, and the decimation introduces a certain amount of aliasing, due to the
non-ideal frequency response of the analysis filters. However, the synthesis filters
characteristics are chosen with such frequency response, that the aliasing introduced by the
analysis filter bank is canceled out in the reconstruction process. The output signal ( ) is
the recovered version of the input signal ( ). Therefore, the output signal ( ) is
expressed as,
( ) (− ) ( )
( )= ( ) ( ) (3)
( ) (− ) (− )
( )= ( ) ( )+ ( ) (− ) (4)
The reconstructed signal ( ) consists of two terms, the first term that is the product of the
transfer function ( ) and ( ) is the desired QMF output, while the second term is the
product of the transfer function ( ) and (− ) is the aliasing term ( ) denotes the
aliasing components produced by the overlapping frequency responses of the analysis and
synthesis filter banks. For an alias-free filter bank, ( ) must be equal to zero. This
condition is mathematically expressed as (Vaidyanthan, 1993),
( )= ( ) (− ) + ( ) (− ) = 0 (5)
This condition may be satisfied by choosing ( ) = (− ) and ( )= (− ), then the

desired QMF output is represented as (Fliege, 1994),
( )= ( ) (− ) − ( ) (− ) (6)
Xˆ ( z )
Fig. 4. Two-channel QMF bank.

The filter banks, which are able to perfectly reconstruct the input signal are the perfect
reconstruction filter banks, that satisfy the perfect reconstruction condition. The desired
QMF output includes the function ( ) which gives perfect reconstruction of the input
signal if it is a mere delay, that is ( ) = (Fliege, 1994). Two-channel filter bank, shown
in Fig. 4, can be utilized to construct an octave-spaced wavelet filter bank with the help of a
tree type structure. Octave filter bank is constructed by the successive decomposition of the
low pass signal into constituent sub-bands, every time using the two-channel filter bank
(Qian, 2002). A three-level octave-spaced analysis filter bank is shown in Fig. 5 (a) and a
three-level octave-spaced synthesis filter bank is shown in Fig. 5 (b).
x̂(n)
Fig. 5. (a) Three-level analysis filter bank (b) Three-level synthesis filter bank.
2.3 Transmultiplexer
Transmultiplexers form an integral part of modems and transceivers based on filter banks
that work on the principle of perfect reconstruction. A simple two-channel filter bank can be
utilized to illustrate the perfect reconstruction condition. A transmultiplexer is the dual of
Sub-band coder (SBC) in structure. Fig. 6 shows a two-channel transmultiplexer filter bank,
which converts a time-interleaved signal at its input to a FDM signal, having separate bands
of spectrum multiplexed together and then converts it back into TDM signal at its output.
Transmultiplexers find application in modems and transceivers for digital communication
(Vaidyanthan, 1993).
X 0 (z) Xˆ 0 ( z )
X1 ( z) Xˆ 1 ( z )
Fig. 6. Two-channel Transmultiplexer.

3. Discrete multitone modulation technique

Discrete Multitone (DMT) modulation is a variant of OFDM associated with various loading
algorithms, so as to optimize a transceiver’s performance in wireline channels like
Asymmetrical Digital Subscriber Line (ADSL) and power line (Chow, et al., 1991). In
literature, several loading algorithms have been developed; allocating resources such as data
bits, or power in order to optimize high data rate, low average transmitting power, or low
bit error rate. Typically two of these parameters are kept constant and third is the goal of
optimization.
A conventional DFT based DMT transceiver block diagram is shown in Fig. 7. The channel
bandwidth is divided into N sub-channels. The input serial bit-stream is also split into
parallel sub-streams. The data bits assigned to sub-channels are according to a loading
algorithm. For water-filling bit-loading algorithm, greater number of bits is assigned to
higher SNR sub-channels. If the value of SNR of a sub-channel is below a pre-assigned
threshold, then no bits are allocated to that sub-channel. The assigned bits are mapped onto
Quadrature amplitude modulation (QAM) constellation forming a complex symbol. The
QAM symbols are then modulated onto orthogonal sub-carriers using Inverse Fast Fourier
Transform (IFFT). The QAM symbols are duplicated with their conjugate symmetric
counterparts and subjected to 2 point IFFT, in order to generate real samples for
transmission through the channel. A DMT symbol is thus formulated.
A guard band consisting of a few samples of the DMT symbol is pre-appended to the
symbol. This is the cyclic prefix, which consists of the last v samples of the DMT symbol,
circularly wrapped to the 2 DMT symbol. The length of cyclic prefix is chosen such that
it will be longer than the length of the channel response. The cyclic prefix added to a DMT
symbol lengthens the symbol period, making it longer than the worst possible delay spread,
which is caused by time delayed reflections of the original symbol arriving at the receiver.
Consequently, the cyclic prefix serves the purpose of absorbing any multipath interference.
Due to this cyclic extended symbol, the samples required for performing the FFT can be
taken anywhere over the length of the symbol, without degradation by the neighboring
symbols. However, the information sent in the cyclic prefix is redundant and reduces the
transceiver throughput by (2 + )/2 . Between the transmitter and receiver lies the
communication channel, which introduces both noise and distortion (mainly due to
multipath propagation) to the composite transmit signal. The channel can be modeled as a
finite impulse response (FIR) filter that possesses a frequency-selective fading characteristic.
The cyclic prefixed signal is transmitted through the channel, the output of which gives the
product of the channel impulse response and the transmitted symbols in frequency domain.
DMT receiver is basically the dual of the DMT transmitter, with the exception of the
equalization part. The equalization block consists of two parts, the time-domain equalizer
(TEQ) and the frequency-domain equalizer (FEQ). The purpose of TEQ is channel-
shortening and it immediately follows the channel, as shown in the Fig. 7. It serves to
shorten the channel impulse response, so that the equalized channel impulse response is less
than the length of the cyclic prefix . At the receiver the cyclic prefix samples are discarded
and remaining samples are subject to Fast Fourier transform (FFT). The frequency domain
equalizer divides the received sub-symbols by the FFT coefficients of the shortened channel
impulse response. The resulting signal is demodulated to recover the original data bits and
converted into a serial bit stream.
Fig. 7. Functional block diagram of a DMT transceiver.
3.1 Evolution of discrete wavelet multitone modulation

A major drawback of DFT-DMT is that the rectangular low-pass prototype filter results in
sinc shaped sub-band spectral response, the first side lobes of which are only 13dB down, as
pointed out by Sandzberg (Sandzberg, 1995). A dispersive channel will thus introduce Inter-
Carrier Interference (ICI) at significant levels. To mitigate this we can increase filter bank
genus, and design sub-channels with greater spectral isolation. We call this a lapped
transform, and much work has been done on the particular case of the Lapped Orthogonal
Transform (LOT) (Malvar, 1992). General Extended Lapped Transform (ELT) design is
computationally prohibitive, however Cosine Modulated Filter Banks (CMFB) can be
efficiently implemented utilizing Discrete Cosine Transform (DCT). In this way the design
procedure is simplified if we allow the transmultiplexer filters to be modulated versions of a
low pass, linear-phase prototype. Therefore, instead of designing filters, we now only
design one prototype filter. Modulated filter banks implementing lapped transforms with
applications to communications are generally referred to as Discrete Wavelet Multitone
(DWMT), to distinguish from DMT which uses a rectangular prototype.
Many contributions in literature have emphasized the need for DWMT in specific channel
conditions. Tzannes and Proakis have proposed DWMT in (Tzannes, et al., 1994), and
shown it to be superior to DFT-DMT. Authors suggest implementing DWMT in DSL
channel for improved performance (Doux et al., 2003). Studies have compared DMT and
DWMT performance in DSL channel (Akansu and Xueming 1998).
DWT exhibits better spectral shaping compared to the rectangular shaped subcarriers of
OFDM. Therefore, it offers much lower side lobes in transmitted signal, which reduces its
sensitivity to narrowband interference (NBI) and inter-carrier interference (ICI). However, it
cannot utilize CP to mitigate ISI created by the frequency-selective channel, as various DWT
symbols overlap in time domain (Vaidyanathan, 1993). Nevertheless, such MCM systems
based on DWT require an efficient equalization technique to counter the ISI created by the
channel.
4. Discrete Wavelet Multitone (DWMT) in Digital Subscriber Line (DSL)

A system based on Discrete Wavelet Multitone (DWMT) for modulating and
demodulating the required signal using Discrete Wavelet Transform as a basis function
has been suggested in wireless applications (Jamin andMähönen, 2005). The importance of
DWMT in wireless communication is a recognized area of research and on similar lines a
DWMT system can be implemented in wireline communication. It can be used as a
maximally decimated filter bank with its overlapping symbols in time-domain. Therefore,
this structure does not require the addition of CP which is an overhead in DMT and
DWMT based wireline systems (Vaidyanathan, 1993). On the other hand, the wavelet
filters also possess the advantages of having greater side-lobe attenuation and requires no
CP (Bingham, 1990). Therefore, the DWMT systems are bandwidth efficient by not using
the CP which creates the problem of bandwidth containment in DMT based systems.
However, application of the DWMT systems in a dispersive channel like ADSL
necessitates a robust channel equalization technique (Sandberg and Tzannes, 1995). In
literature some equalization techniques for DMT based multicarrier systems have been
suggested by many authors (Pollet and Peeters, 2000); (Acker et al., 2001); (Acker et al.,
2004); (Karp et al.,2003) and DWMT based multicarrier systems (Viholainen et al., 1999).
Equalization is a key factor in the design of modems based on DWMT modulation
technique and till date, it remains an open research area. When using the Discrete Wavelet
Packet Transform (DWPT) as a basis function in DWMT systems, it is difficult to equalize
the overlapped symbols in time domain. We emphasize on the design of equalizer for
DWPT based DWMT multicarrier systems. The proposed system is based on DWPT for
DWMT wireline systems and time-domain equalization is suggested for the equalization
process of overlapped symbols.
In this chapter, the time-domain equalization through a linear transversal filter is applied.
The equalization algorithms are based on Zero-Forcing (Z-F) and minimum mean squared
error (MMSE) criterion to a discrete wavelet-packet transform based DWMT transceiver for
a wireline ADSL channel. It is then compared with the system’s performance of a DMT
based ADSL system. For a fair comparison between the two systems, the DMT system also
utilizes the same time-domain equalization. The performance of the proposed wavelet-
packet based transceiver is also evaluated in the presence of near-end crosstalk (NEXT) and
far-end crosstalk (FEXT) for downstream ADSL. It is shown that the DWMT system
conserves precious bandwidth by not utilizing any CP, and gives improvement in bit error
rate (BER) performance over the DMT system with time-domain equalization (TEQ).
4.1 System model of DWMT

The DWMT system model’s block diagram is shown in Fig. 8. It divides the input data bit-
stream into multiple and parallel bit-streams. The proposed DWMT transceiver is based on
discrete wavelet packet transform (DWPT). The DWPT is implemented through a reverse
order perfect reconstruction filter bank transmultiplexer. Wavelet packets can be
implemented as a set of FIR filters, which leads to the filter bank realization of wavelet
transform, according to Mallat’s algorithm (Mallat, 1998). The blocked version of the input
signal ( ) is mapped to a variable QAM constellation according to the number of bits
loaded. This is interpolated and filtered by the branch synthesis filter ( ). The
combined signal is sent through the channel, and the received signal is filtered by an
equalizer filter. The equalized signal is passed through the corresponding analysis filter
( ) and decimated to retrieve the QAM encoded version of the transmitted signal. The
transmitted signal is recovered after QAM decoding.
Fig. 8. Functional Block diagram of DWMT system.
4.1.1 Water filling bit loading

Bit loading is usually applied to DMT modulated systems applied to wireline channels, by
first estimating the signal-to-noise ratio (SNR) of each sub-channel through channel
estimation techniques, which is followed by the distribution of bits to these sub-channels
according to their respective SNR. Water-Filling bit loading algorithm applied in the
proposed system is rate adaptive and it is suitable for achieving maximum bit rate and also
useful when considering the large number of sub-channels and variable QAM constellation
(Leke and Cioffi, 1997);(Yu and Cioffi, 2001). A discrete version of this algorithm is applied,
in which the bit-loading procedure initiates by determining the sub-channels that should be
turned off, due to very low SNR. The bits are assigned to channels according to their
capacity, expressed mathematically as (Thomas et al., 2002),
1  SNR 
b log 2 1   (7)
2  . m 
where SNR = εn.gn is the SNR of each sub-channel, εn is the sub-channel energy and gn is the
sub-channel SNR and it can be calculated as,
2
Hn
gn  (8)
2
where Hn is the ADSL channel impulse response and σ2 is the noise power, Γ is the SNR gap
and γm is the performance margin, which is the amount by which SNR can be reduced (Yu
and Cioffi,2001). The water filling bit-loading for the proposed system is shown in Fig. 9.
While considering the DWMT based communication system for the ADSL channel, it is
necessary to consider its frequency response and the effect of crosstalk, near-end crosstalk
(NEXT), and far-end crosstalk (FEXT) in system simulation. The ADSL channel impairments
and crosstalk is briefly discussed in the following section.
Fig. 9. ADSL channel frequency response & number of bits loaded according to discrete
water-filling algorithm.
4.2 ADSL channel

Digital Subscriber Line, commonly known as DSL is the most popular and ubiquitously
available wireline medium which provides high-speed Internet access over the twisted pair
telephone network. Fig. 10 shows a typical DSL network, which consists of copper lines
extending all the way from the central office (CO) to the customer’s premises. Current and
future applications such as Interactive Personalized TV, high definition TV (HDTV) and
video-on-demand through high-speed Internet access, will require more bandwidth.
Researchers are exploring cost-effective ways to exploit the existing copper infrastructure to
deliver greater bandwidth.
Fig. 10. A typical DSL network connecting subscribers to internet services through DSL to
the Central Office.
Although the DSL channel offers the advantage of utilizing the already in place telephone
lines to carry digital data, however there are different channel impairments that pose
difficulties in achieving the objective of high-speed and reliable communication (Cook, et

al.,1999). These channel impairments include different types of noise and interference. The
noise sources include crosstalk, impulse noise and narrow band noise (Thomas Starr, et al.,
2002). Also, interference in the communication signal may occur due to the electromagnetic
conduction (EMC) in the unshielded twisted pair (UTP) and DSL operating in the vicinity of
transmitters may pick up radio frequency interference (RFI) (Cook, et al.,1999). Moreover,
signal reflection may be induced due to bridge tabs, unterminated lines and load
mismatching in the telephone network. This leads to multipath signal propagation, due to
ISI occurs (Bingham, 2000). BER deterioration, due to ISI is a significant problem in the
communication systems utilizing the DSL channel. A typical telephone line frequency
response and its impulse response are shown in Fig. 11 and Fig. 12 respectively. Multicarrier
modulation is a possible solution to the ISI problem in DSL, which is already standardized
in Asymmetric digital subscriber line (ADSL), in the form of DMT modulation, as G.DMT
and G.lite ADSL.
Fig. 11. Frequency response of telephone line FIR channel.
Fig. 12. DSL channel impulse response.

4.2.1 Crosstalk
In a telephone network, each subscriber is connected to the CO through a twisted pair,
however, hundreds of such pairs are bound together in a cable. The twisting in the wires
keeps the electromagnetic coupling between them to a minimum, however, when the pairs
are numerous, all crosstalk between the pairs cannot be completely removed. Therefore, this
crosstalk constitutes a dominant impairment, where DSL channel is concerned. The DSL
crosstalk types, namely near end crosstalk (NEXT) and far-end crosstalk (FEXT) are
illustrated in Fig. 13 (Thomas Starr, et al., 2002). NEXT is the crosstalk due to the
neighboring transmitter on a different twisted pair line and its power increases with
increase in frequency. FEXT is the noise detected by the receiver located at the far end of the
cable from the transmitter. FEXT is typically less severe than NEXT, because FEXT is
attenuated as the cable length increases.
In this chapter, the performance of DWMT transceiver is evaluated for the downstream
ADSL channel. For this purpose, the NEXT and FEXT are modeled using the ADSL standard
G.992.1/G.992.2(ITU-T, 2003).
Fig. 13. NEXT and FEXT, the DSL crosstalks illustrated (Thomas Starr, et al., 2002).
The PSD of the ADSL transceiver disturbers for downstream is given by (ITU-T, 2003),
2
  f 
sin    
2   f o   1 1
PSDADSL , ds  Disturber  K ADSL , ds    2
 12
 16
,
fo  f   f   f HP 3 dB 
  1  1 
 fo   f HP 3 dB   f 
(0  f  ) (9)
where f is in Hz and the remaining parameters are defined in Table 1. The PSD of the ADSL
transceiver downstream NEXT is given by (ITU-T, 2003),
  NPSLn 
1.5
PSDADSL , ds  NEXT  PSDADSL , ds  Disturber  10 10  f NXT   f 1.5 ,(0  f   ) (10)
 
where f is in Hz and the remaining parameters are also given in Table 1. The PSD of the
ADSL transceiver downstream FEXT is given by (ITU-T, 2003),
2
  FPSLn 
1 2
PSDADSL , ds  FEXT  PSDADSL , ds  Disturber  H channel ( f )  10 10  dFXT  dFXT d f 2,
 
(0  f  ) (11)
where f is in Hz, and Hchannel(f) is the channel transfer function and the remaining
parameters are given in Table 1.
PSD of disturbers and NEXT is shown in Fig. 14(a) and Fig. 14 (b) displays the FEXT PSD for
downstream ADSL (ITU-T, 2003). The NEXT and FEXT for upstream can be computed in a
similar manner (ITU-T, 2003).
Parameter WPT-DWMT DFT-DMT

Number of disturbers 24 24
fLP3dB fs/2 fs/2
fHP3dB 138 kHz 138 kHz
KADSL 0.1104 watts 0.1104 watts
fNXT 160 kHz 160 kHz
NPSL 47.0 dB 47.0 dB
fFXT 160 kHz 160 kHz
dFXT 1.0 km 1.0 km
FPSL 45.0 dB 45.0 dB
Table 1. NEXT & FEXT Simulation Parameters.
Fig. 14. (a) PSD-disturber & PSD-NEXT for downstream ADSL in G.992.1/G.992.2 standard.
Fig. 14. (b) PSD-FEXT for downstream ADSL in G.992.1/G.992.2 standard.

The wavelet packet transform (WPT) transmultiplexer in the proposed DWMT transceiver
gives perfect reconstruction of the transmitted signal, if ideal channel conditions are
assumed. However, an actual channel like ADSL is far from ideal, and therefore requires
some form of equalization to reliably retrieve the transmitted signal. Time domain
equalization is proposed here for DWMT based transceiver for ADSL. There are some
equalization techniques for ADSL proposed in literature (Acker et al., 2004);(SMÉKAL et al.,
2003);(Trautmann and Fliege, 2002); (Yap and McCanny, 2002).
4.3 Time domain equalization

In order to equalize the signal after it has been dispersed by the ADSL channel, time domain
equalization is proposed, and it is implemented through a linear transversal filter. The
equalizer filter is a linear function of the channel length L, and the filter coefficients are
optimized using the zero-forcing (ZF) and mean squared error (MSE) criterion (Farrukh et
al., 2007); (Farrukh et al., 2009).
4.3.1 ZF finite length equalizer

In ZF algorithm it cancels out the channel effect completely by multiplying the received
signal with the inverse of the channel impulse response, as shown in Fig. 15. With an infinite
length equalizer filter, it is possible to force the system impulse response to zero at all
sampling points (Proakis, 1995). However, since an infinite length filter is unrealizable.
Therefore, a finite length filter is considered that approximates the infinite length filter
(Proakis, 1995). The received signal y is the distorted version of the transmitted signal x after
convolution with the channel ch plus the channel noise r. The received signal can be
expressed in vector notation as,
= + (12)
The equalizer output vector z can be found by convolving a set of a training sequence input
samples h and equalizer tap weights c (Sklar, 2001),
= (13)
However, we continue with the assumption that channel state information is entirely known
at the receiver. Therefore, a square matrix h, consisting of channel coefficients is formulated
with the help of ZF criterion. The ZF algorithm defines that in order to minimize the peak
ISI distortion by selecting the equalizer filter weights c such that the equalizer output is
enforced to zero at sample points other than at the desired pulse. The weights are chosen
such that (Sklar, 2001)
1 for k  0
z( k )   (14)
0 for k  1, 2, , N
The equalizing filter has L=2N+1 taps. Equalizer filter coefficients are computed by (Sklar,
2001)
= (15)
The job of equalizing filter is to recover the transmitted signal x̂ from the received channel-
distorted signal y, as follows,
xˆ = yc
(16)
= xch c  rc
where x̂ is the distorted received signal which was transmitted through ADSL channel and
recovered after ZF equalization.
Fig. 15. A Linear transversal equalizer with coefficients optimized by Zero-Forcing criterion.
4.3.2 MMSE criterion

The MMSE criterion represents a more robust solution compared to the ZF since it considers
the effect of additive channel noise (Proakis, 1995);(Sklar, 2001). The MMSE criterion of
transversal equalizer filter coefficients optimizes the mean squared error of all the ISI terms
plus the noise at the equalizer output. A set of over determined equations is formed, in
order to derive a minimum MSE solution of the equalizer filter (Sklar, 2001). Therefore, for a
2N+1 tap filter, the matrix h will have dimensions of 4N+1 by 2N+1. Multiplying Eq. (13) by
hT (Sklar, 2001),
= (17)
= (18)
where is the cross correlation matrix and = is the autocorrelation matrix of the
input noisy signal, which are used to determine the equalizer coefficients c,
= (19)
For the MMSE solution of the equalizing filter, an over sampled non-square matrix h is
formed which is transformed to a square autocorrelation matrix Rhh, yielding the optimized
filter coefficients.
4.4 Simulation results

An ADSL system is investigated which is based on DWPT transmultiplexer. The system
utilizes M = 256 sub-channels and rate adaptive bit-loading algorithm is applied for bit
allocation to each sub-channel in channel environment which is based on ADSL along with
the crosstalk noise standards G.992.1/G.992.2 (ITU-T, 2003). For fair comparison, two
systems are simulated, which are based on DWMT and DMT transceiver using time-domain
equalization (TEQ) techniques for ADSL channel in the presence of AWGN and crosstalk
noise. The channel is considered to be stationary during symbol duration. MatLab is used
for all this simulation purpose and the parameters for simulation are specified in Table 2.
Parameter WPT-DWMT DFT-DMT

Data rate 1 Mbps 1 Mbps
Sampling Frequency 2.208 MHz 2.208 MHz
Modulation M-QAM (2, 4, 8, 16, 32, 64) M-QAM (2, 4, 8, 16, 32, 64)
Cyclic Prefix None 20%
FFT size (N) - 512
Wavelet-level 2 -
Number of bits/sub-channel 1 to 6 1 to 6
Table 2. DWMT & DMT System Simulation Parameters.
This corresponds to a system bandwidth of 2 MHz with data rate of 1 Mbps with discrete
wavelet packet filter which is used for transmitter and receiver end. The channel
equalization is performed by applying a linear equalizing filter in time-domain. The filter
coefficients for equalization are optimized by ZF algorithm and MMSE criterion. The ADSL
channel is simulated by an FIR filter of 100 taps.
The prototype filter for the synthesis and analysis part of the transmultiplexer is a discrete
wavelet filter using 2-level wavelet packet. The input symbols xk(n) are M-QAM modulated.
The equalizer frequency response of ZF equalizer FIR filter is shown in Fig. 16. Initially
DWMT transceiver and DMT systems are compared regarding the bit error rate (BER)
performance in AWGN channel, having identical time-domain zero-forcing channel
equalization. Although, the conventional DMT system equalization is a combination of time-
domain equalization (TEQ) and frequency-domain equalization (FEQ) techniques, in this
case DMT is equalized with a time-domain Zero-Forcing for fair comparison between the
two systems. The DWPT transform is applied utilizing Haar wavelet. Fig. 17 shows the
comparative performance of two systems in the presence of AWGN without crosstalk. The
BER curve, shown in Fig. 17, presents the fact that the two systems give almost identical
performance for lower SNR, and at higher SNR, the DWMT system exhibits an
improvement of 1 dB in / over the DMT system for an AWGN channel, at a of 1E-
6. It shows that both techniques using DMT and DWPT based ADSL without crosstalk
perform identically except at higher SNR. In the next step, the simulation is performed
according to the ADSL standard with crosstalk from G.992.1/G.992.2 (ITU-T, 2003).
Fig. 16. Equalizing Zero-Forcing filter frequency response.
Fig. 17. BER Comparison of DWMT & DMT systems in AWGN with ZF Equalization
techniques.
Fig. 18 shows the performance of DWMT and DMT systems in ADSL channel with AWGN,
NEXT and FEXT (crosstalk), utilizing time-domain equalization (TEQ) techniques. The
NEXT & FEXT represent the downstream crosstalk in ADSL channel according to the
G.992.1/G.992.2 standard (ITU-T, 2003), with the simulation parameters as described in
Table 1. DMT system is still equalized by ZF-TEQ, while the DWMT transceiver is equalized
by ZF-TEQ, time-domain MMSE (MMSE-TEQ). The BER curves shown in Fig. 18 validate
the fact that the wavelet packet transmultiplexer improves the performance of DWMT
transceiver, having ZF-TEQ by Eb/No margin of 1.0 db for BER of 1E-4, over a DMT
transceiver, having an identical equalizer. Moreover the MMSE-TEQ technique for DWMT
system shows an improvement of 2 dBs in Eb/No over ZF-TEQ technique for DWMT and a 3
dB gain over the ZF-TEQ equalized DMT system, at a BER of 1E-4.
Fig. 18. BER Comparison of DWMT & DMT systems for ADSL channel with AWGN, NEXT
& FEXT.
5. Pros & cons of applying DWT in multicarrier modulation techniques

DWMT modulation based transceiver, appears to be an interesting choice, when utilizing
multi-carrier modulation techniques in wireline systems. It not only recommends the unique
time-frequency localization advantage over the conventional frequency localized DMT
systems, but also preserves precious bandwidth, which is wasted in DMT based systems in
the form of cyclic prefix. However, when utilized in time dispersive channel like ADSL,
DWMT transceiver cannot do without an equalization technique because of the time
overlapped symbols. In this chapter DWMT based transceiver is discussed and its
performance analyzed for the ADSL channel, in comparison with a conventional DMT
modulation with ZF and MMSE algorithms using the time-domain equalization. DWMT
system based on WPT performs well in the presence of AWGN and crosstalk in comparison
with the DMT system for ADSL. ZF equalization algorithm does not consider noise, while
the MMSE criterion of optimizing the equalizer coefficients takes into account the effect of
channel noise. Therefore MMSE algorithm based DWMT transceiver gives better BER
performance in comparison with ZF criterion, since ZF is known to enhance channel noise.
The time-domain equalization is computationally complex in comparison to frequency
domain equalization, however it offers improved bit error rate.
6. Conclusion
The multirate digital signal processing techniques, including wavelets and filter banks are part
of new emerging technologies, which are finding applications in the field of digital
communications. DWT based Multicarrier modulation techniques have opened new avenues
for researchers, to avoid the spectral leakage and spectral inefficiency associated with Fourier
Transform based MCM techniques. Time domain equalizers based on ZF and MMSE
algorithms are utilized for DSL channel equalization in DWMT transceivers. MMSE based
equalizers outperform the ZF equalizers in terms of BER. The equalization techniques adopted
for DWMT transceiver is a topic of active research. Moreover, simulation results found in
literature have shown that DWT based MCM systems exhibit higher immunity to narrowband
interference (NBI). Therefore, WOFDM/DWMT can be considered as a viable alternative to
spectrally inefficient OFDM/DMT, however at the cost of higher computational complexity of
equalization.
7. References
Acker, K. V. , Leus, G. Moonen, M. van de Wiel, O. and Pollet, T. (2001). Per tone
equalization for DMT-based systems, IEEE Trans. on Communications, vol. 49, no. 1,
pp. 109-119.
Akansu, A.N.; Xueming Lin. (1998).A comparative performance evaluation of DMT (OFDM)
and DWMT (DSBMT) based DSL communications systems for single and multitone
interference. Proceedings of IEEE International Conference onAcoustics, Speech and
Signal Processing, 1998., vol.6, no., pp.3269-3272 vol.6, 12-15.
Alliance for Telecommunications Industry Solutions, (1995). American National Standard
for Telecommunications - Network and Customer Installation Interfaces -
Asymmetric Digital Subscriber Line (ADSL) Metallic Interface. ANSI T1.413 1995
ANSI. New York.
Jamin, A. Mähönen, P. (2005). Wavelet packet modulation for wireless communications,
Wireless Communications and Mobile Computing 5 (2): pp. 123-137.
Baig, S. and Gohar, N. D. (April 2003). Discrete Multi-Tone Transceiver atthe Heart of PHY
Layer of an In-Home Powerline Communication Local Area Network.IEEE
Communications Magazine,pp. 48-53.
Baig, S. and Mughal, M.J. (2009). Multirate signal processing techniques for high-speed
communication over power lines, IEEE Communications Magazine, vol.47, no.1,
pp.70-76, January 2009
Bingham, J. C. (2000). ADSL, VDSL, and Multicarrier Modulation, Wiley & Sons.
Chow, P. S., Tu, J. C. and Cio, J. M. (1991). Performance Evaluation ofa Multichannel
Transceiver System for ADSL and VHDSL Services.IEEE Journal on Select. Areas in
Commun., 9(6):909-919.
Cook, J.W.; Kirkby, R.H.; Booth, M.G.; Foster, K.T.; Clarke, D.E.A.; Young, G. (1999). The
noise and crosstalk environment for ADSL and VDSL systems.IEEE
Communications Magazine,vol.37, no.5, pp.73-78, May 1999.
Doux, C., V.; Lienard, J.; Conq, B.; Gallay, P.(2003). Efficient implementation of discrete
wavelet multitone in DSL communications.Video/Image Processing and Multimedia
Communications, 2003. 4th EURASIP Conference focused on , vol.1, no., pp. 393- 398.
Farrukh,F., Baig, S. and Mughal, M.J. (2007). Performance Comparison of DFT-OFDM and
Wavelet-OFDM with Zero-Forcing Equalizer for FIR Channel Equalization, Proc. of
IEEE International Conference on Electrical Engineering, LHR, Pakistan.
Farrukh,F., Baig, S. and Mughal, M.J. (2009). MMSE Equalization for Discrete Wavelet
Packet Based OFDM, Proc. of IEEEInternational Conference on Electrical Engineering,
ICEE’09, LHR, Pakistan.
Frederiksen, F.B. Prasad, R. (2002). An overview of OFDM and relatedtechniques towards
development of future wireless multimedia communications.Proceedingsof IEEE
Radio and Wireless Conference, RAWCON.pp. 19- 22.Aug. 11-14 Boston,
Massachusetts, USA.
Fliege, N.J.(1994). Multirate Digital Signal Processing-Multirate Systems-Filter Banks-Wavelets,

Wiley, New York.
H. S. D. (June 1988). The LOT: a link between block transform coding and multirate filter
banks.IEEE International Symposium on Circuits and Systems. pp. 835–838.Espoo,
Finland.
Hirosaki, B. (Jul 1981). An Orthogonally Multiplexed QAM System Using the Discrete
Fourier Transform.IEEE Transactions onCommunications, Volume 29, Issue 7, pp.
982-989.
Karp, T. , Wolf, M. Trautmann, S. and Fliege, N. J. (2003). Zero-Forcing Frequency Domain
Equalization for DMT Systems with Insufficient Guard Interval, Proceedings of IEEE
International Conference on Acoustics, Speech, and Signal Processing, pp. IV221- IV224,
Hong Kong, China.
Leke, A.and Cioffi, I. M. (1997). A Maximum Rate Loading Algorithm for Discrete Multitone
Modulation System, Proc. IEEE GLOBECOM '36.London. U.K., "01.3.DD. 1514-18.
Lindsey, A. R. and Dill, J. C. (1995). Wavelet packet modulation: a generalized method for
orthogonally multiplexed communications.Proceedings of IEEE 27th Southeastern
Symposium on System Theory, pp. 392–396.March 1995. Starkville,. Mississippi, USA.
Mallat, S. (1999) A Wavelet Tour of Signal Processing, 2nd ed. Academic Press New York.
Malvar, H. S. (1992). Extended lapped transforms: Properties, applications and fast
algorithms.IEEE Transactions on Signal Processing. Vol. 40, no. 11, pp. 2703–2714.
Pollet, Thierry and Peeters, Miguel. (2000). Equalization for DMT-Based Broadband
Modems, IEEE Comm. Mag.
Proakis, J. G. (1995). Digital Communications, 3rd edition, McGraw-Hill International
Editions.
Qian, Shie, (2002). Introduction to Time-Frequency and Wavelet Transforms,1st Ed, Prentice Hall
PTR.
Saltzberg, B. ( Dec 1967). Performance of an Efficient Parallel Data Transmission
System.IEEE Transactions on Communications, Volume 15, Issue 6, Page(s):805 — 81
Sandberg, S. D. and Tzannes, M. A. (1995).Overlapped discrete multitone modulation for
high speed copper wire communications, IEEE J. Select. Areas Commun., vol. 13, no.
9, pp. 1571--1585.
Sari, H. Karam, G. and Jeanclaude, I. (1995). Transmission techniquesfor digital terrestrial
TV broadcasting.IEEE Comms. Mag,Vol. 33, pp. 100—109.
Sklar, B. (2001). Digital Communications: Fundamentals and Applications, 2nd Ed. Ch. 3,
Englewood Cliffs, NJ: Prentice Hall.
Starr,T. Sorbara,M. Cioffi,J.M. Silverman,P.J. (2002). DSL Advances, Prentice Hall PTR.
Tzannes, M.A.; Tzannes, M.C.; Proakis, J.; Heller, P.N. (1994). DMT systems, DWMT
systems and digital filter banks.Proceedings of ICC '94, pp. 311 - 315 Vol.1.New
Orleans.
Van Acker K., Leus G., Moonen M., Pollet T., (2004). Improved initialization for time domain
equalization in ADSL, Signal Processing, vol. 84, pp. 1895-1908.
Vaidyanathan, P.P.. (1993).Multirate Systems and Filter Banks. Prentice-Hall, Inc.UpperSaddle
River, NJ.
Viholainen, A., J. Alhava, J. Helenius, J. Rinne, and M. Renfors, (1999).Equalization in filter
bank based multicarrier systems, in Proc. Int. Conf. on Electronics, Circuits and
Systems, Pafos, Cyprus, pp. 1467-1470.
Weinstein, S.; Ebert, P. (Oct 1971). Data Transmission by Frequency-Division Multiplexing

Using the Discrete Fourier Transform. IEEE Transactions on Communications,Volume
19, Issue 5,Part 1, pp:628 - 634.
Yap, K.S. and J.V. McCanny, (2002). A mixed cost-function adaptive algorithm for ADSL
time-domain equalization, Proc. of the IEEE International Conference on
Communication (ICC 2002), Vol. 3, New York, pp. 1798-1802.
Yu, W. and Cioffi, J. M. (2001). On Constant Power Water-Filling, IEEE ICC.
0
2
A Scalable Architecture for Discrete Wavelet

Transform on FPGA-Based System
Xun Zhang
Institut Superieur d’Electronique de Paris-ISEP
France
1. Introduction
In recent year, the Forward and Inverse Discrete Wavelet Transform (FDWT/IDWT) (S.Mallet,
1999) has been widely used as an alternative to the existing time-frequency representations
such as DFT and DCT. It has become a powerful tool in many areas, such as image
compression and analysis, texture discrimination, fractal analysis, pattern recognition and so
on. The recent and future developments of high definition digital video and the diversity of
the terminals had led to consider a multi-resolution codec. In this context, the FDWT/IDWT
as well as the others computational functions such as Motion Estimation (ME) are required to
be scalable and flexible to support rich multimedia applications and adapt to the fast changing
of standards requirement. In this background, a universal, extremely scalable and flexible
computational architecture which can adapt to variable workload would be more and more
important and suitable for the multimedia application in the future.
In the literature, there have been several proposals devoted to the hardware implementation of
FDWT/IDWT. Some proposals(M.A.Trenas et al., 2002) (et al, 2002) (Lee & Lim, 2006)(Ravasi,
2002)(P.Jamkhandi et al., 2000)(Tseng et al., 2003)addressed the importance of flexibility and
proposed programmable DWT architectures based on two types: VLSI or FPGA architecture.
The VLSI architectures have large limitations in terms of flexibility and scalability compared
to the FPGA architectures. Even though some recent solutions proposed programmable
and scalable for either variable wavelet filters(Olkkonen & T.Olkkonen, 2010) (Lee & Lim,
2006) or the structure of FDWT, they remind, in addition to their cost, dedicated to specifics
algorithms and cannot be adapted to future solutions. In another hand, the existing FPGA
architectural solutions are mainly ASIC like architectures and use external off-the-shelf
memory components which represent a bottleneck for data access. The possibility of
parallelizing the processing elements offered by FPGAs associated to a sequential access to
data and bandwidth limitations do not enhance the overall computing throughput. The
very powerful commercial VLIW digital signal processor obtains its performance thanks to
a double data-path with a set of arithmetic and logic operators with a possibility of parallel
executions and a wide execution pipeline. However, these performances are due to a high
frequency working clock. Even though these DSP has a parallel but limited access to a set
of instructions, the data memory access remains sequential. The performance requirement is
paid by high circuit complex and power consumption. Most of work focuses on the reuse of
devices likes FPGAs for different applications or different partitions of one applications.
In order to square up these needs, we propose a novel DWT architecture and implementation
method. The proposed architecture can support multi-standard by reconfiguring the
26
Will-be-set-by-IN-TECH
T1 T1 T1
T2 T7 T2
T3 T3 T30
T31
T4 T8 T4 T32
T5
T9 T5
T6 T6 T6
(A1) (A2) (A3)
Fig. 1. Application adaptive configuration
interconnection between date memories and processing elements. Moreover, the number
of processing element and its working frequency could be reconfigured dynamically. A
controller plays a key role as a reconfigurable interface allowing multiple accesses to local
memory, external memory through a DMA and feeding the processing element in an
optimal fashion. An implementation method is developed to identify parallelism level of
processing element and working frequency as well as to find out the tradeoff between power
consumption and performance. In comparing with others VLSI and ASIC architecture, double
size of memory can be economic in using our novel architecture.
In the following paper, we start, in Section2 by presenting a definition of adaptation in two
manners: the application adaptive and task adaptive, within the system complexe context. We
then give a brief overview of DWT algorithm in Section 3 where we detail a reconfigurable
DWT hardware processor architecture. In order to experimentally explicit our proposed
system, Section 4 focus on the detail of our proposed reconfigurable architecture which
supports our DWT algorithm implementation. Section 5 focus on the implmentation, analysis
and validation of system. Finally, Section 6 summarizes and concludes this work.
2. Levels of adaptation
In the multimedia computing environment, adaptation can be seen in two manners: the
application adaptive and task adaptive. Following the adaptation of computing environement
the different applications or different standard of one application can be switched in run-time.
For example, the multimedia terminal switches it use from playing a movie to answering
a video call. The task adaptive consists of the switching different versions of a task of an
application, this situation can occur for instance in down scaling or up scaling situations.
2.1 Application adaptive

For a given domain, applications can be described by a set of processing tasks and sub tasks.
The difference between the applications could be represented with common processing tasks
and specific processing tasks. Figure 1-A2 shows an example of two applications A1 and
A2 featuring common tasks (continuous lines) and specific tasks (dash lines). Switching from
application A1 to application A2requires replacement of specific tasks and the communication
between newly loaded tasks and common tasks. In some cases, the simultaneous execution
A Scalable
A Scalable Architecture
Architecture forTransform
for Discrete Wavelet Discrete Wavelet
on FPGA-Based Transform on FPGA-Based System
System 273
of two applications is required. To achieve this, different versions of specific tasks must be
available.
2.2 Task adaptive

Each task of an application commonly consists of a set of sub-tasks or a set of operators
depending on the complexity of task as shown in figure1-A3. To enable task adaptivity,
different versions of a task for a given algorithm must be defined and characterised in terms
of power, area, throughput, efficiency and other objectives. For the same task, it must be
also possible to change the type of algorithm in order to adapt the application to the future
standards.
In this background, the adaptability of application helps us to configure partially one part
of application for adapting to a new application. The task adaptive level permits us mainly
to make a small change in the task to make the application adapt to different sceneries. In
this paper, we focus on the task adaptive so as to realize muti DWT processing algorithms by
using partial reconfiguration technique.
3. 2-D DWT processing algorithm

A survery of 2-D DWT architecture can be referenced in the paper Olkkonen & T.Olkkonen
(2010). The two dimensional (2D) forward discrete wavelet transform (FDWT) is a rapid
decomposition in the multimedia application domain. The FDWT is computed by successive
low-pass and high-pass filtering. The output of each filter is decimated that is every second
value is removed halving de length of the output S.Mallet (1999). The output of each filter
stage is made of transform coefficients and each filter stage represents a level of transform.
The low pass result is then transformed by the same process and this is repeated until the
desired level is reached. In the Inverse discrete wavelet transform (IDWT), the approximation
and detail coefficients at every level are up-sampled by two, passed through the low pass and
high pass filters and then added. This process is continued through the same number of levels
as in the decomposition process to obtain the original signal. In this paper we will focus on
the implementation of IDWT, the same approach will be applied to FDWT.
3.1 Classical processing approach

The classical approach to 2D decoding is to process each layer in the tree decomposition
separately and to process the vertical and horizontal layers successively one after the other.
The performance of this approach is strongly limited by the management of temporary data
required between two successive layers and between horizontal and vertical filtering. For a
2D image with N rows and N columns and L levels, the amount of data to be filtered on each
layer increase ( for IDWT) by a factor of four from one layer to the next, and the total amount of
processed data along the whole tree reconstruction process is given by the following equation:
N×N 4L − 1
L
D=
i −1
= ∑
3 × 4 L −1
×N×N (1)
i =1 4
To process a N × N image, a temporal memory of size
4L − 1
D−N×N = ( )×N×N (2)
3 × 4 L −1
is required. As an example, for 2 level resolution a temporal memory of 0.25 N × N size is
required. For a given layer, the filtering process is achieved horizontally and vertically; thus
28
two read accesses and two writes accesses are necessary and the total amount of data read and
written is expressed as Dw = Dr = 2 × D. The memory bandwidth B, in bidirectional access
case, can be considered as the production of the total amount of data processed for a frame
per second ( f ps) Td f = ( Dr + Dw ) × f ps and the number of bits Nb of a coefficient:
B = Td f × Nb (3)
As an example, for a gray level image of 512 × 512 pixels with 25 frame per second, 8 bits
per pixel and 2 levels of reconstruction, a bandwidth of 260 Mb/s is required. These results
illustrate the memory management problem as the main bottleneck of the classical approach.
3.2 Proposed processing approach

In order to reduce the memory size and to optimize the overall system performance, the
wavelet algorithm is redesigned to exploit efficiently the inherent processing parallelism. This
processing parallelism is possible if the required data is accessible in parallel, accordingly a
data partitioning is used. The degree of parallelism and thus of the data partitioning will
depend on the level of transformation, the number of levels and data dependency.
The proposed organization is shown in figure 2 depicting the memory fragmentation (2-a)
and tasks allocation (2-b) on processing elements for two level IDWT. It is a compromise
and intermediate solution between a massive parallelism and a sequential execution. The
processing tasks are mainly filtering operations witch are organized and allocated to a
processing element so that the among of data processed is the same. Indeed, if we consider a
W × W bloc, an IDWT will be processed in three phases as shown in figure (3). In phase Φ1
W ×W
The processing element PE1 requires 2 × W 2 × 2 =
W
2 data accesses to reconstruct the LL
bloc meanwhile the processing element PE2 can process vertically the W × 2
W
remaining data
(HL and HH). In phase Φ2 , when the two processing elements terminate their executions, the
LL bloc is reconstructed and the pressing element PE2 can resume its vertical executions on
the W ×
2
W
available data. In phase Φ3 , after the termination of PE2, data is available to process
the horizontal pass on a bloc of W × Wdata. Using PE1 and PE2 in parallel, the data processed
by each PE is of W × W
2 . This architecture is scalable and can be extended to different levels of
resolutions by an adequate choice of processing elements.
4. System overall architecture

With the down scaling technology, the modern chips can integrate a huge quatity of mixed
grain hardware resources ranging from several hard microprocessors, hard arithmetic
operators to hundred of thousand of simple gates allowing the integration of various soft
cores. The problem of resources management becomes then very acute especially in
reconfigurable systems. In these systems, the management of reconfigurations is a very
important part in the design phase due to the complexity of hardware reconfigurations and
the reconfigurability needs of an aplication.
In the different proposed solutions, the two parts of reconfiguration that are reconfigurable
capabilities of the hardware and the different reconfigurations possibilities of an application
are not taken into account. A layered reconfiguration management approach through a hierar-
chical decomposition of a system will allow us to solve this problem.
The proposed adaptable architectue shown in figure 1- c, allowing the adaptation of differents
applications and an application in different conditions, is organised as a set of clusters. Each
cluster is designed to execute a sub-set of tasks. This clusters are parallelisable, so that the
A Scalable
System 295
Fig. 2. Processing approach in 2D IDWT onto two-level(a); task allocation(b)
PE1 PE2 PE2
HL L PE1
L L
LH HH HL H H H PE2
Φ1 Φ2 Φ3
Fig. 3. 2-D IDWT Processing phases
30
same set of processing are performed on multiple data blocs. Each cluster is composed of
an heterogeneous multiprocessor cores that allow software reuse, one or several Reconfig-
urable Processing Modules (RPU), a Reconfigurable Communication Module (RCM), and on
chip memory. The RPM allows hardware acceleration and can be configured in a way that
supports different versions of a task. The reconfigurable interface (RIF) is used to build the
interconnection between differents modules. Each RPM can be reconfigured at runtime.
Each cluster has a local configuration manager implemented in an on chip processor that
controls the sequences of reconfigurations of the cluster. In this local configuration level, all
clusters are configurable in parallel and independently. The reconfiguration process allocates
dynamicaly to differents tasks of an application the adequate hardware ressources and
optimal operation frequency and voltage. The presence of local configuration managers
allows the acceleration of the adaptation process. To control the overall system, a global
reconfiguration level is necessary. In this level, the necessary informations are managed in
order to modify the global organisation of the system by configuring the communication
between clusters and the elements of a cluster, allowing for instance to switch from an
application to another.
The overall architectureM.Guarisco et al. (2007) is depicted in figure 4. Three memory blocks
are present, while the first one and the last one repectively store original data image and
deliver computed data, the second block feeds the processing elements. In addition to these
three blocks, the system is composed of a reconfigurable processing unit two data organization
units and control unit. This last one unit allows to connect the right memory to the right Unit
at the right time. Once the memory bloc 1 is full (and as a consequence memory bloc 3 is
empty, or at least, all his bytes are read or store in external memory), each memory datapath
is switch allowing new picture datas to be treated. A new cycle begins, memory 3 is this time
filled and datas in memory 1 are transforming.
Fig. 4. Porposed DWT processing system architecture
4.1 RPU instance

The reconfigurable Processing Unit (RPU) allows the implementation of different types of
wavelet filter. A filter (task) is a set of arithmetic and logic operators. A configuration of RPU
consists of a type of filter or a version of a filter. For a given filter, corresponding operators can
be connected by different ways in order to carry out different filter versions. These different
A Scalable
System 317
Fig. 5. General architecture of a RPU
versions can be parallel, in pipeline, sequential or an association of both methods. A possible

architecture of the RPU and conction with reconfigurable interface is shown in the figure5
Chart1 lists number of computation operators needed (number of additioner, shifter,
multiplier by filtering operation). We have choose two filters in order to illustrate adaptation
at task level.
Filters Additions Shifts Multiplications
5/3 5 2 0
2/6 5 2 0
SPB 7 4 1
9/7-M 8 2 1
2/10 7 2 2
5/11-C 10 3 0
5/11-A 10 3 0
6/14 10 3 1
SPC 8 4 2
13/7-T 10 2 2
13/7-C 10 2 2
9/7-F 12 4 4
Table 1. Different filter types of wavelet transform

32
Table 1 lists the number of main computational requirements (the number of additions,
shifts, and multiplications per filtering operation). We choose two filters to illustrate the task
adaptive level.
4.1.0.1 The 5/3 lifting based wavelet transform
The IDWT 5/3 lifting based wavelet transform has short filter length for both low-pass and
high-pass filter. They are computed through following equations :
D [n] = S0 [n] − [1/4( D [n] + D [n − 1]) + 1/2] (4)
S[n] = D0 [n] + [1/2(S0 (n + 1) + S0 [n])] (5)
The equations for FDWT 5/3 are given bellow:
D [n] = D0 [n] − [1/2(S0 (n + 1) + S0 [n])]] (6)

S[n] = s0 [n] + [1/4( D [n] + D [n − 1]) + 1/2] (7)
D [n] is the even term and S[n] is the odd term. The corresponding data flow graph(DFG)
is shown in figure 6. It is composed of two partitions: odd and even. Each partition is
implemented in the corresponding data path of the RPU. The register file is used to hold
intermediate computation results.
4.1.0.2 The 9/7 − F based FDWT
The 9/7-F FDWT is an efficient approach which is computed through following equations:
203
D1 [n] = D0 [n] + [ (−S0 [n + 1] − S0 [n]) + 0.5] (8)
128
217
S1 [ n ] = S0 [ n ] + [ (− D1 [n] − D1 [n − 1]) + 0.5] (9)
4096
113
D [n] = D1 [n] + [ ( D [n + 1] + D1 [n]) + 0.5] (10)
128 1
1817
S [ n ] = S1 [ n ] + [ ( D [n] + D1 [n − 1]) + 0.5] (11)
4096 1
There is similarities between equations of 5/3 filter and those of 9/7 − F filter which implies
same similarities between the data flow graph of the two filters. It is clear that by duplicating
the dataflow graph of filter 5/3 and inserting four multipliers we obtain the data flow graph of
the 9/7 filter. Moreover, if we consider the table 1, we can see that by partially reconfiguring
the 9/7 filter we can implement all the list of the table. The reconfiguration of 9/7 filter
consists of suppressing or disconnecting unused operators and generation of an adequate
control and an efficient data management.
4.2 Reconfigurable interface

The reconfigurable interface is the key element of Reconfigurable Prcessing Unit (RPU). One of
its functionality is to connect together the RPU and control communication protocol between
the RPU and internal memory. Th controler has to generate adresses for writing and reading
operations in memory. A reconfigurable sequencer is used in order to manage the operation
and communication sequence. The reconfigurable interface is composed of a three levels
pipelined structure for calcul units apart from the one of the first level. Steps of pipeline
are : reading (R), execution (E), and writing (W). In our bench test, two versions of interfaces
holding different filters implementation are defined. The pipeline stages are :
A Scalable
System 339
Fig. 6. IDWT 5/3 (a) and 9/7 DFG (b)
• Read (R): The source operands from the on chip memory are sent to the register file. The
contro module gives an order to the reading file address generator integrated into the
control module for reading the row or column resource from the memory module (SRAM)
to the RPU at the address pointed to be by a read counter. Two data are read in one clock
cycle.
• Execution (E): The data available in the regiter file is used bythe data-path to process in
parallel the two parts of the filter. As the high pass filter part requires the previous result
of low pass filter part, the execution is delayed by one clock cylce for high pass filter results.
This operation is executed in one clock cycle.
• Writeback (W): The results of computation are written back to on chip memory at the
address pointed to by a write counter.
The figure7 illustrates the operating mode of the three stages pipeline. Because of sequential
acces to one memory bloc, the computations of the first level are performed as shown in (a)
allowing the exection of three operations in one clock cycle. For the remaining porcessing,
thanks to the parallel read, execute and write, six operations are executed in one clock cycle
(b).
34
Fig. 7. Pipeline organization: special case (a) and normal case(b)
4.3 Memory access

Seemless memory is make up of several fixed size blocks. Each block is a dual port memory
with simultaneous read-write access. Size of memory block correspond to image size in the
first level of transformation in the case of the iDWT (inverse DWT). In our experimentation we
choose an image of 32x32 pixels or bytes (we work on grey level pictures, that’s why a pixel is
constitued by one byte only). Because of this organisation, when the first level is proceded, the
two data paths of the processing elements are sequentially feeded, that require two memory
access cycles. However, for others, datas are read from (or ordered in) two different parallel
memory blocks for one processing element in parallel.
4.4 Detailed operations

To explain the operating details of the system, consider an original 8x8 image as shown in
figure 8-a. One of the task of this architecture is to rearrange the the pixels in the memory
bloc. In order to benefit from the parallelism, Data organizing mudule arranges the pixels as
shown in figure 8-b. So, due to the utilisation of memory bloc divided in four independent
dual port memories, the processing controller can reach, for a given i, Si and Di which are
normally two consecutive pixel in the image and those which we need to calculate at the same
time the two coefficient of the DWT. If we want to calculate two new samples at each clock
cycle, we have to reach two consecutive elements (Si and Di) at the same cycle.
So, in a first time, each processing element can calculate 1D-DWT in line. As we have two
element, the system can compute two 1D-DWT in the same time. In a second time, the system
computes the 1D-DWT, but now, in columns. Thus, we save a precious time and we can
theoretically achieve an infinite number of levels. Let be Tload , the needed time to fill a memory
bloc at the frequency of the data (it corresponds to the time of a complete reading or writing,
pixel after pixel, of the whole memory bloc), we can say that the execution time of the first
A Scalable
System 35
11
Fig. 8. Original Image(a); Reorganized Image(b)
level is Tload
2 (we reach two pixels at one clock cycle, so, it divides by two Tload , we have two
PE that divides again by two Tload but we have to achieve two time the 1D-DWT. Finally the
execution time is of Tload
2 ). Moreover, we know that for the next level, we need only the low
frequency coefficients which represent only a quarter of the total result of the previous level.
The execution time is then the result of an arithmetic suite which is represented in equation 4.
1 Tload
n
Texec = ∑ i = 1 n 4i − 1 2
(12)
If the number of level tends towards infinite, the execution time is then of 2∗T3load . A data-out
unit allows getting back the DWT coefficients in an ordered way. This controller can be easily
modified to adapt the structure of the data flow to the system.
4.5 Target platform

In order to demonstrate the feasibility of the proposed FDWT/IDWT architecture, we have
implemented a reconfigurable architecture IDWT targeting an FPGA Xilinx from the Virtex
4 family. The virtex-4 circuit hold partial reconfiguration. Partial reconfiguration of Xilinx
FPGA’s is achieved using partial configuration datas Inc. (2004). The target architecture,
as shown as in the figure 9, is make up of static modules (PowerPC, ICAP, BRAM, PLB
Bus) and reconfigurable units(the scalable RPU and hierarchic on-chip memory). ICAP is
used to achieve the partial reconfiguration through the embeded processor PowerPC. The
reconfiguration datas are stocked in BRAM memory of FPGA and are loaded via ICAP.
5. Implementation results
We have modeled the architecture in HDL in the sofware suite ISE from Xilinx. The simulation
results agree with our theoretical waiting. Indeed, we can perform with this architecture
a very high number of levels. According to the simulation results, we can run a working
frequency of 67MHz. But as we use the internal memory of an FPGA, we are limited and we
can reach an image size of only 128x128 pixels. The solution consists of a small modification
of the data organizing units to allow the architecture to treat macro-bloc instead of a whole
picture.
Figure10 illustrates the placement and routing of one RPU on Xilinx Virtex-4 FPGA. Three
mains parts of system like: reconfigurable interface(middle bloc), four registre blocs and two
datapaths of IDWT algorithm. The configuration file of each is independant, which is named
36
Virtex−4 FX12 Platform
DWT fontion Unit
datapath
odd
register file
On chip memory
datapath
Reconfigurable Interface
even
(1)
RPU
On chip memory
datapath
odd
register file
(2)
On chip memory
datapath
even
(3)
LL LH
I/O
HL HH
Organisating
PowerPC Unit
PLB
accessing memory
ICAP Controller
Controller BRAM
DDR
Fig. 9. Target architecture.
Partial bitstreams RPU bitstream size(Kbyte) Reconfiguration time

Ko ms
static 582 Ko 21 sec (JTAG)
part
Partial bitstream 1 R_com_1 33Ko 0.57 ms
Partial bitstream 2 R_com_2 63Ko 0.67 ms
Partial bitstream 3 R_f_53 28Ko 0.26 ms
Partial bitstream 3 R_d_97 11Ko 0.16 ms
Table 2. Measured reconfiguration time of different bitstream files for 2-D IDWT.
A Scalable
System 37
13
Fig. 10. placement and routing schema of DWT processing unit with Xilinx PlanAhead tool
as partial bitstreams. The different partial bitstreams are stored in the on chip BRAM. The
static bitstream is loaded using cable. To measure the execution time of each partial bitstream,
a free running hardware timer is used. The measurement results are in table 2. In this table,
the mains modules are : the different part between filter 5/3 and 9/7 ( R_d_97 ), 5/3 filter
functional module (R_f_53), interface for 5/3 filter (R_com_1) and: communication interface
for 9/7 filter (R_com_2).
The on chip PowerPC processor is used for autoconfiguration through HWICAP. As the
PowerPC is an element of the system, it is used to detect external or internal events and
accordingly loads automatically the adequate configuration to adapte the system to the given
situation and then making the auto-adaptive. The HWICAP makes auto-configuration easier,
in fact a C program on PowerPC allows the transfer of 512x32 bit blocks of the partial bitstream
from the configuration memory to a fixed size buffer of the HWICAP peripheral, which the
transfer from the buffer to the ICAP. The total reconfiguration time can approximated by the
following equation:
Tcon f ig = TICAP + TBRAM (13)

Where TICAP is the time required to transfer configuration from the buffer to the ICAP, and
TBRAM is the time required to transfer data from configuration memory to the HWICAP buffer.
Table 2 shows different parts of the system, the size of corresponding bitstream file and their
configuration time. The system consists of a static part and reconfigurable parts ( Part1 and
Part2 are the two versions of reconfigurable communication allowing the switching between
38
two filters, Part3 corresponds to 5/3 filter, and Part4 is the difference between 5/3 filter
and 9/7 filter ). The configuration time is measured using a free running counter (timer)
incremented every system clock cycle, and capturing the start time and the end time. We see
that the configuration time as expected depends linearly on the size of bitstream.
Type of Resolution Area(mm2 Max frequency of Memory

architecture for VLSI and operation(MHz) Requirement
ASIC)(CLB for (KB)
FPGA)
Proposed 32x32 153 CLBs 50 1.024
architecture
64x64 538 CLBs 50 4.960
ASICs one image 8.796 mm2 50 2 memory frames
based(Tseng frame
et al., 2003)
Zero-padding 32x32 4.26mm2 50 6.99
scheme (et al,
2002)
Table 3. Implementation results

To compare the measured configuration time with the minimum possible value, the value for
the reconfiguration of Virtex-4 FPGA could be obtained with this equation: Tcon f ig = L/r,
where L is the length of the configuration and r is the transfer rate. As an example, for a file of
63KB size, and a clock frequency of 100 MHz as used in our experimentation, the minimum
theoretical time is 0.63 ms, which is much less than 90 ms that as given in table 2. This is due to
PowerPC that acts as the configuration manager in our system. Large part of time is spent to
copy reconfiguration data from on chip or external memory to HWICAP buffer. The difference
between the measured configuration time (0.97 ms) and the computed time (0.63 ms) is due
to the imprecision of the measurement method. In fact, the capture of start and stop time is
achieved using software, which tacks additional clock cycles. In table 2 we can see also that
the main part of reconfiguration time is wasted for the transmission of reconfiguration files.
The reconfiguration time includes two part times: Tbram , the total load time for transferring
the reconfiguration bitstreams from memory on chip to buffer of ICAP with package
512byte. Ticap , the total configuration time through the ICAP port is grouped by the sum
of configuration time for one package. Hence, the reconfiguration time is decided largely
by the size of reconfiguration bit files and the number of reconfiguration bit files. The
reconfiguration manager makes possible to reduce the reconfiguration time through hiding
partial reconfiguration process in the execution process. It is obvious that the configuration
time can be improved. A solution we are studying is based on a specific hardware
reconfiguration manager capable to transfers the configuration data from on chip memory
to ICAP.
Moreover, the chart 3 compare our approach with the other architecture. We observe primary
two parameters based on different resolution of image. At the same working condition, the
area of DWT computing module is variety according to the size of image(153CLBs for 32x32
and 538CLBs for 64x64) where including the adding of memory requirement(1,024KB for
32x32 and 4,960KB for 64x64). Thus the area of circuit can be used efficiently according to
A Scalable
System 39
15
the defnite size of image. The size of of memory requirment is scalable and thus the correcte
size of memory can be configured dynamically to adapt to the requirement of bandwidth
of memory. The other work shown in this table are based on ASIC(Tseng et al., 2003)and
VLSI (et al, 2002). the area of circuit and the size of memory are fixed and thus the maximum
size of memory must be previewered, which may lead to the urgent or surplus of memory
access.
This proposed architecture features small area and low momory requirments. Processing time
for a 32x32 image blocks is 43s which is lower than others traditional design. Using a 64X64
image blocks gives a good performance throughput which takes 86s for the transformation,
for two-level 2D IDWT, which is capable to perform the image CCIR(720X576) format image
signal at 50 f rame/s.
6. Conclusion
In this book chapter, , we have described auto-adaptive and reconfigurable hybrid architecture
for F/IDWT signal processing application. Two levels of auto adaptation are defined in
order to minimize the reconfiguration overhead. The application adaptive level in which
different applications of a domain are classified and characterized by a set of tasks. The task
adaptive level in which for a given task, a set of versions are defined and characterized for
use in a situation to adapt the application to different constraints like energy, and bandwidth
requirement.
The proposed architecture is a universal, scalable and flexible featuring two levels of
reconfiguration in order to enable the application adaptivity and task adaptivity. We
demonstrated through the case study that it can be used for any types of filters, any size
of image and any level of transformation. The memory is organized as a set of independent
memory blocks. Each memory block is a reconfigurable module. The high scalability of the
architecture is achieved through the flexibility and ease of choosing the number of memory
blocks and processing elements to match the desired resolution. The on-chip memory is used
not only to hold the source image, but also to store the temporary and final result. Hence, there
is no need of temporal memory. The processor has no instructions and then no decoder, in fact,
the hardware reconfigurable controller plays the role of a specific set of instructions and their
sequencing. For a given set of tasks, a set of configurations are generated at compile time and
loaded in run time by the configuration manager via configuration memory. The prototype
has been tested on FPGA developpment cart of Xilinx with 65nm CMOS technology. The
prototyping chip can be reconfigured to adapte 5/3 filter or 9/7 filter. In comparing with
others ASIC architecture at the same working frequency, our proposed architecture requires
less memory bloc and fewer hardware resource than the others.
7. References
et al, S. (2002). Vlsi implmentation of 2-d dwt/idwt cores using 9/7-tap filter banks based on
the non-expansive symmetric extension scheme, in IEEE (ed.), Proceedings of the 15th
International Conference on VLSI Design, number 435 in ASP-DAC ’02, IEEE Computer
Society, Washington, DC, USA.
Inc., X. (ed.) (2004). Two flows for partial reconfiguration: module-based or different based, Xilinx.
Lee, S.-W. & Lim, S.-C. (2006). Vlsi design of a wavelet processing core, in IEEE (ed.), IEEE
Transactions on Circuits and Systems for Video Technology, Vol. 16.
40
M.A.Trenas, J.Lopez & E.L.Zapata (2002). A configurable arechitecture for the wavelet packet
transform, The journal of VLSI Signal Processing, Vol. 32, pp. 151–163.
M.Guarisco, X.Zhang, H.Rabah & S.Weber (2007). An efficient implementation of scalable
architecture for discrete wavelet transform on fpga, in IEEE (ed.), 6th IEEE Dallas
Circuits and Systems workshop, pp. 1–3.
Olkkonen, H. & T.Olkkonen, J. (2010). Discrete Wavelet Transform Structures for VLSI
Architecture Design, intech, Hannu Olkkonen and Juuso T. Olkkonen (2010).
Discrete Wavelet Transform Structures for VLSI Architecture Design, VLSI,
Zhongfeng Wang (Ed.), ISBN: 978-953-307-049-0, InTech, Available from:
http://www.intechopen.com/articles/show/title/discrete-wavelet-transform-
structures-for-vlsi-architecture-design.
P.Jamkhandi, A.Mukherjee, K.Mukherjee & Franceschini, R. (2000). Parallel
hardware/software architecture for computation of discret wavelet tranform
using the recursive merge filtering algorithm, Proceeding International Parallel Distrib.
workshop, pp. 250–256.
Ravasi, M. e. a. (2002). A scalable and programmable architecture for 2-d dwt decoding, IEEE
Transactions on Circuits and Systems for Video Technology, Vol. 12, pp. 671–677.
S.Mallet (1999). A theory for multi-resolution signal decomposition: The wavelet
representation, number 7, IEEE, pp. 674–693.
Tseng, P.-C., Huang, C.-T. & Chen, L.-G. (2003). Reconfigurable discrete wavelet transform
architecture for advancedmultimedia, in IEEE (ed.), Signal Processing Systems,
number 137-141.
0
3
VLSI Architectures of Lifting-Based Discrete

Wavelet Transform
Sayed Ahmad Salehi and Rasoul Amirfattahi
Isfahan University of Technology, Department of Electrical and Computer Engineering,
Digital Signal Processing Research Lab., Isfahan
Iran
1. Introduction
The advantages of the wavelet transform over conventional transforms, such as the Fourier
transform, are now well recognized. Because of its excellent locality in time-frequency
domain, wavelet transform is remarkable and extensively used for signal analysis,
compressing and denoising. Defining DWT by Mallat [1] provided possibility of its digitally
hardware or software implementation. The discrete wavelet transform (DWT) performs a
multiresolution signal analysis which has adjustable locality in both the space (time) and
frequency domains [1]. Unlike the Fourier transform, the wavelet transform has many
possible sets of basis functions. A trade-off can be made between the choice of basis functions
and the complexity of the corresponding hardware implementations. Using finite impulse
response (FIR) filters and then subsampling is the classical method for implementing the
DWT. Due to the large amount of computations required, there have been many research
efforts to develop new rapid algorithms [2]. In 1996, Sweldens presented a lifting scheme
for a fast DWT, which can be easily implemented by hardware due to significantly reduced
computations [3]. This method is entirely based on a spatial interpretation of the wavelet
transform. Moreover, it provides the capability of producing new mother wavelets for
the wavelet transform, based on space domain features. Due to recent advances in the
technology, implementation of the DWT on field programmable gate array (FPGA) and digital
signal processing (DSP) chips has been widely developed. As described in Sect. 3, in the
lifting scheme the structural processing elements, including multipliers, are arranged serially;
hence, the number of multipliers in each pipeline stage determines the clock speed of the
structure. Based on [4], the main challenges in the hardware architectures for 1-D DWT are
the processing speed and the number of multipliers, while for 2-D DWT it is the memory
issue that dominates the hardware cost and the architectural complexity. The reason is the
limitation of the on-chip memory and the power consumption [4,5].
2. DWT structures
The wavelet transform provides a time-frequency domain representation for the analysis
of signals. Therefore, there are two main methods to produce and implement wavelet
transforms. These methods are based on time domain or frequency domain features. The
frequency based method is Filter Banks (FB) and the time based one is called Lifting Scheme
(LS). We will describe them in following sections.
42
2.1 Filter banks structure

In the FB method, for one level of wavelet decomposition, the input signal is divided into two
separate frequency parts by passing it simultaneously through a pair of low pass, H (z), and
high pass, G (z), filters, as shown in Fig. 1. Then, subsampling the filter’s output to produce
the low pass and high pass outputs (s,d). Therefore, the FB method performs the DWT based
on convolving filter taps and samples of the input signal. H (z) and G (z) can be written in this
form:
H (z) = h0 + h1 z−1 + h2 z−2 + ... + h N z− N
G (z) = g0 + g1 z−1 + g2 z−2 + ... + g M z− M .
d
G(z) 2
x
s
H(z) 2
Fig. 1. Filter Banks Block diagram

As an example consider the CDF(2,2) wavelet. H (z) and G (z) for this transform are
−1 1 3 1 −1
H ( z ) = √ z2 + √ z + √ + √ z −1 + √ z −2 .
4 2 2 2 2 2 2 2 4 2
−1 1 −1
G ( z ) = √ z2 + √ z + √
2 2 2 2 2
The low pass filter has 5 taps and the high pass has 3 taps, so we call it 5/3 wavelet. Although
FB structure is the prior one but it is only capable of providing wavelet transforms in the
frequency domain and not in the time domain. Moreover, in general, the FB filter coefficients
are not integer numbers; hence, they are not appropriate for hardware implementation. In
addition, the number of arithmetic computations in the FB method is very large.
2.2 Lifting structure

The LS method is a new method for constructing and performing wavelets based on the time
(space) domain [3].
As shown in Fig. 2, at first the LS structure splits the input signal samples into even and odd
samples. Then P function is applied on even samples as a prediction function. The word
prediction is used here because P function predicts odd samples using even samples.The
difference between this prediction and the actual value of odd sample, creates the high
frequency part of the signal which is called "detail" coefficients (d). Then applying the U
function on detail signal and combining the result with even samples update them so that the
output coefficients (s) have the desired properties. Usually the desired properties of s is the
same as the properties of input signal (x) but with half size. So the s signal is an approximation
for x and is called approximation coefficient.
Note that the details and approximation coefficients (d,s) in lifting scheme, respectively, are
the same as high pass and low pass outputs in FB.
Based on the above description we have
d = xodd − P ( xeven ),
VLSI Architectures
VLSI Architectures ofDiscrete
of Lifting-Based Lifting-Based Discrete Wavelet Transform
Wavelet Transform 433
for prediction block and

s = xeven + U (d)
for update block.
Equations for P and U functions are determined based on the implemented wavelet, also the
number and arrangement of P and U blocks in the lifting structure are different for various
types of wavelets.
x d
-
x
split P U
s
+
x
Fig. 2. Block diagram of a lifting stage
We can write matrix equations for P and U blocks respectively as following:

xeven (z) 1 0 xeven (z)
=
d(z) t(z) 1 xodd (z)

P

s(z) 1 s(z) xeven (z)
= .
d(z) 0 1 d(z)

U
Generally speaking, if we have more than one lifting step, the matrix equation is(3):

s(z) k 0 m 1 si (z) 1 0 xeven (z)
0 1/k i∏
= (1)
d(z) =1
0 1 ti (z) 1 xodd (z)
In (1), k and 1/k are normalization factors. The last matrix is used only for normalization
and may be omitted in many applications such as compression. The relation between FB
coefficients and LS equations is (3):

h (z) ho (z) k 0 m 1 si (z) 1 0
E (z) = e
ge ( z ) go ( z )
= ∏
0 1/k i=1 0 1 ti (z) 1
Matrix E (z) is called a polyphase matrix, where according to the FB structure, he and ho are
even and odd taps of the low pass filter and ge and go are even and odd taps of the high
pass filter, respectively. si (z) and ti (z) are related to filter coefficients in FB structure. In other
words si (z) and ti (z) can be obtained from FB by factorization algorithm presented in [5].
Example: Let consider the previous example, 5/3 wavelet, in LS. This wavelet consist of one
lifting step (one P unit and one U unit together is a lifting step). For this wavelet the prediction
of each odd sample in signal is the average of two adjacent even samples. Then P block
calculates the difference between the real value of signal sample and its prediction:
1
d(n ) = x (2n + 1) − [ x (2n ) + x (2n + 2)].
2
U block updates even samples to have the same property as the original signal. It uses two
most recently computed differences for update procedure:
1
s(n ) = x (2n ) + (d(n − 1) + d(n )).
4
44
So the matrix equation for 5/3 wavelet is

√
s(z) 2 0 1 14 (1 + z−1 ) 1 0 xeven (z)
= √1 −1 (1 + z )
d(z) 0
2
0 1 2 1 xodd (z)

E(z)
and the polyphase matrix is
−1 −1 1 z −1 + √

√ z + √
3 − √
1 z √ 1
E (z) = 4 2
−
2 2 4 2 2 2 2 2 .
√1 − √1 z √1
2 2 2 2 2
We propose the following lemma for using in hardware implementation of LS as will be
describe in section 3.1.
Lemma1: Factorization can be done so that si (z) and ti (z) are first-order or lower-order
polynomials.
Proof sketch: After a polyphase matrix representing a wavelet transform with finite filters
is factored into lifting steps, each step becomes a Laurent polynomial. Since the difference
between the degrees of the even and odd parts of a polynomial is never greater than 2, it is
always possible to find the common divisor of the first-order or lower-order polynomials.
Also, the lifting factorization process is non-unique and so there is freedom in the form
of factorization. Hence, a classical wavelet filter can always be factored into first-order or
lower-order Laurent polynomials (i.e., si (z) or ti (z)).
Compared to the FB method, the LS method has many advantages [6,7]. The most important
one is the number of arithmetic computations. In the LS method, the number of arithmetic
operations, additions and multiplications, is nearly one-half of that of the FB, which is why
the LS structure is more efficient. Even the amount of computations in some types of DWT
can be reduced to a quarter of that needed for FB [8]. Furthermore the LS structure has
the advantage of implementing the Integer Wavelet Transform (IWT) efficiently. IWT is a
wavelet-like transform in which all of the decomposition coefficients are integer [9]. The IWT
is appropriate for hardware implementation of the DWT [10]. The practical advantages of
using the lifting-based IWT have been described in [11]. Moreover, by using the LS, it is easy
to implement the DWT in a fully in-place method, which is memory efficient [2,12].
Regarding the above explanation, the LS structure is used to implement 5/3 and 9/7 wavelets,
which are used, respectively, for lossless and lossy compression in the JPEG 2000 standard.
3. 1-DDWT
In this section, some types of lifting-based DWT processing elements and 1-D structures are
explained.
3.1 Basic functional units for lifting scheme

As pointed out in Lema1, factorization can be done so that si(z) and ti(z) are first order or
lower-order polynomials. Figure 3 shows three possible categories of the basic processing
unit and the related polynomials in such a factorization. Different kinds of lifting-based DWT
architectures can be constructed by combining the three basic lifting elements. Most of the
applicable DWTs like 9/7 and 5/3 wavelets consist of processing units, as shown in Fig. 3(a),
which is simplified as Fig. 4. This unit is called the processing element (PE). In Fig. 4, A, B
and C are input samples which arrive successively. To implement the P unit, A and C receive
VLSI Architectures
D D D
B + B + B +
a × +
× ×
a ×
+ a b
C Z A C Z
A A
-1 -1
a(1+z ) a+bz a
Fig. 3. Basic functional units for LS
Fig. 4. Most convenient basic PE for LS
even samples while B receives odd samples. On the other hand, for the U unit, A and C
are odd samples and B receives even samples. Now, the structure of Fig. 4 can be used to
implement 5/3 and 9/7 wavelets. For instance, Fig. 5 and Fig. 6 shows the architecture of
the 5/3 and 9/7 wavelets respectively, where each white circle represents a PE. In Fig. 6, the
R R R
P
U
Fig. 5. Lifting Structure for 5/3 wavelet
R R R
Fig. 6. Lifting Structure for 9/7 wavelet

input and output layers are essential (basic) layers and are fixed for each wavelet type, while
by changing the number of extended layers, the type of wavelet can be changed accordingly.
For example, omission of a single extended (added) layer in Fig. 6 will change the related
architecture from 9/7 type to 5/3 type. The black circles in Fig. 6 represent needed stored data
for computing outputs (s,d). R0 , R1 and R2 , are registers that get their values from new input
samples and are called data memory. The other three black circles which store the results
of previous computations are known as temporary memory. The number of data memory
46
registers is constant and is equal to 3, while the number of temporary memory registers is
(2e + 1), where e is the number of extended layers [13]. This structure can be implemented
by using combinatorial circuits so that, when the input samples are fed to the architecture,
outputs are ready to be used after a delay time. Also, the implementation of the structure
can be performed via a pipelined structure by adding some registers. The number of pipeline
stages depends on the added registers. Increasing the pipeline stages results in increases in
the clock frequency, system latency and number of required registers [4]. Note that 2-D DWT
architectures are constructed from 1-D DWT units as row-wise and column-wise DWT units.
The data of a complete row is saved for each memory in a column-wise unit. So, the sum of
the data and temporary memories in the column-wise DWT unit determines the amount of
needed internal memory [14,15,16]. The pipeline registers do not affect the required internal
memory [17].
3.2 1-D DWT structures

By combining the functional units described in previous section we can construct 1
dimensional DWT. The architectures presented in Fig. 7 can be applied to implement
the lifting-based 1-D DWT. The structure shown in Fig. 7(a) processes all input samples
concurrently, in parallel form. In Fig. 7(b), input samples arrive in pairs at consecutive clock
pulses and the results for each pair are ready after five cycles. However, due to the pipelined
structure, the clock frequency of Fig. 7(b) is higher than that of Fig. 7(a). There is a trade-off
between the clock speed and the number of pipeline stages.
4. Hardware architectures
4.1 Simple hardware implementation
Figure 7 shows an architecture proposed in [18] for the 9/7 wavelet. Indeed, it is the hardware
Implementation of Fig. 6. Accordingly, the architecture presented in Fig. 9 can be used for the
5/3 wavelet.
4.2 Minimizing hardware architectures

In the structure of Fig. 8, there are two similar cascaded blocks which are different only in the
multiplier’s coefficients. According to [19], one of the similar blocks may be omitted as shown
in Fig. 10. Note that in Figs. 7 and 8, the delay unit represented by z−1 is implemented by
one register, while in Fig. 10 each delay unit contains two consecutive registers. Investigation
on the architecture depicted in Fig. 10 shows that this hardware contains one P and one U
unit. Note that, as mentioned in Sect. 2, both the P and U units can be implemented with the
functional unit depicted in Fig. 4. Therefore, the structure shown in Fig. 10 is implemented
by two similar sections which can be reduced to one section. The resulting architecture for the
9/7 wavelet is shown in Fig. 11. In this structure, U1(0) represents the current output of the
U1 unit and P1(-1) represents the previous output of the P1 unit, and so on. The control signal,
"S", which has four states, selects the inputs of the multiplexers sequentially. In the first state,
two consecutive input samples arrive and the P1 function with α coefficient is performed on
them. In the second state, the U1 function with β coefficient will be imposed on the result of
the previous state (first state’s output). Similarly, in the third and fourth states, computations
for P2 and U2 units will be performed on the results of the previous states. Thus, P2 and
U2 produce final outputs for the structure. The data flow for achieving a pair of wavelet
coefficients using the proposed structure is shown in Table 1.
VLSI Architectures
x x x x
(a)
x x
x x T=0
T=1
T=2
T=3
(b)
Fig. 7. 1-D DWT structures based on lifting a)parallel architecture b) sequential-pipelined
architecture
U1 U2
i0
Z + Z + o0
+ β × + δ ×
α × + γ × +
i1 Z + Z Z + Z
o1
P1 P2
Fig. 8. Lifting-based hardware architecture for 9/7 wavelet

S in1 in2 out F(factor)
0 i1 i0 P1 α
1 i0(-1) P1 U1 β
2 P1 (−1) U2 P2 γ
3 U1 (−1) P2 U2 ζ
Table 1. Time sequence for structure of Fig. 11
The calculation of consecutive wavelet coefficients is periodic and continuous; therefore, the
sequence of control signal "S" for data flow can be easily generated by a simple logic circuit.
Figure 11 shows the hardware architecture for Fig. 11. The 5/3 wavelet implementation of
the proposed architecture is depicted in Fig. 13. It is clear that only the number of coefficients
48
R0
o0
i0 +
+ 1/4 ×
-1/2 × +
i1 + o1
R1 R2
Fig. 9. Lifting-based hardware architecture for 5/3 wavelet
Z + o0
i0 β
+ ×
δ
α
γ
× +
i1 o1
Z + Z
Fig. 10. Propose architecture in [19]
in1 out
+
s F
×
in2
+ s
Z
s
Fig. 11. Minimized structure
and delay block registers, that is, the z−1 blocks, have been modified from four to two. So,
changing the wavelet type changes these two quantities, coefficients and registers, only. Both
P and U units in LS can be implemented by means of the PE shown in Fig. 4. We explored
this feature in the previous section and implemented a 1-D DWT structure containing only
one PE. We call this method the "folded method". The folded structure is an alternative for the
proposed method in [12] by which the lifting-based structures can be designed systematically.
As shown in Fig. 14 for 9/7 wavelet, the method in [12] produces systolic architecture, but
folded method produces folded architecture. In folded structure, the output of the PE unit is
fed back through the delay registers to the PE’s input. By incorporating different numbers of
delay registers and coefficients with PE, the structure for different wavelets can be designed.
For example the folded structure for 5/3 and 9/7 wavelets has two and four delay registers,
respectively. Also the coefficients for 5/3 wavelet are −21 and 14 while for 9/7 they are α, β, γ, δ.
VLSI Architectures
out
+
s
×
s
+ s
Fig. 12. Hardware implementation for Fig.10
out
+
s ×
s
+ s
Fig. 13. Minimized architecture for 5/3 wavelet
(a) Systolic method
out
PE
(b) Folded method

Fig. 14. 1-D Lifting-Based DWT for 9/7 wavelet by Systematic Design Method
In order to show the efficiency of our architecture, several architectures are chosen for
comparison. Ignoring the pipeline registers, the results of comparison for the 9/7 wavelet are
given in Table 2. It is obvious that compared to other architectures, the number of processing
units is reduced in the folded architecture, thus requiring less area to implement the DWT.
Having smaller 1-D DWT units is very effective in multidimensional architectures or in 2-D
DWT, where it is needed to increase the number of 1-D DWT units to achieve a higher
performance [20]. The cost is that, in the proposed architecture, the clock pulses required
to compute outputs are more than those in the previous architectures. This requirement is
due to the sequential states required to complete the computation of each output.
50
Architecture Multiplier Adder Register

Lifting[18] 4 8 6
Proposed in [19] 2 4 6
Systolic[12] 4 8 4
Folded [30] 1 2 4
Table 2. Comparison of some different 1-D DWT architectures for 9/7 wavelet (J −→ ∞)
5. 2-D DWT structures

In this section, we review four convenient structures for 2-D DWT. It is assumed that the 2-D
wavelets reviewed in the following structures are separable, so that the 2-D wavelet transform
can be reduced to a 1-D wavelet transform performed on rows and columns, respectively.
In the direct method (step-by-step method), shown in Fig. 15, the input frame, stored in
external memory, arrives at the 1-D DWT, row by row. The primary outputs are wavelet
coefficients in the row direction and are stored in the external memory. After scanning all
the rows of the frame, again the coefficients are transferred from the external memory to the
1-D DWT block, but this time in the column-wise direction. The secondary outputs of the
1-D DWT block are 2-D DWT coefficients of the input frame, which are stored in the external
memory again. If computation of coefficients for one more decomposition level is needed, this
procedure must be repeated for the LL part of the previous level, whose size is a quarter of the
input frame size. This routine will be repeated for higher levels. The direct method hardware
is simple, but its latency and the number of external memory accesses are large. The number
of external memory accesses for computing a J-level 2-D DWT of an N × N input image can
be calculated by the expression below, where half of the sum is related to the external memory
reads and the other half is related to the external memory writes
1 1 1
4 × (1 + + + · · · + ( ) J −1 ) × N 2 (2)
4 16 4
External memory 1-D

(N*N) DWT
Fig. 15. Direct method

The line-based method can be implemented in two forms: single level and multilevel. In the
line-based single level method, which is shown in Fig. 16, each level of DWT is performed by
a 2-D DWT block. In this method, only internal memory is used to compute one level DWT
for both the row and column directions, hence, there is no external memory access during
the computation of one level 2-D DWT (except for reading rudimentary inputs and writing
final results for that level). The required internal memory is the sum of the data memory and
the temporal memory (black circles shown in Fig. 6) for each line. So the amount of needed
internal memory is 6N for 9/7 wavelet and 4N for 5/3 wavelet [13,21]. But the results of the
DWT in the row direction for even rows can be used for the computation of the DWT in the
column direction without storing them. Hence, the required internal memory for 9/7 and 5/3
2-D wavelets is reduced to 5N and 3N, respectively. Recently, some new modifications have
been made for the 2-D DWT block. For example, in [20] the number of data entrances has been
increased by using more 1-D DWT units in the 2-D DWT block. Although this modification
VLSI Architectures
11
increases the speed, it requires more internal memory and the size of the circuit is increased.
In [22], by replacing registers with line buffers and controlling data flow in a structure like Fig.
13, with 5 more registers, a 2-D DWT block for 5/3 wavelet has been proposed.
Obviously, for a higher-level 2-D DWT, only LL coefficients of the previous level are used, so
the total number of external memory access for a J-level 2-D DWT on an N × N image is
1 1 1
+ + · · · + ( ) J −1 ) × N 2
2 × (1 + (3)
4 16 4
The structure of Fig. 17 performs all levels of 2-D DWT, using only internal memory. So,
External memory 1-level

(N*N) 2-D DWT
Fig. 16. Single level line-based method

the total number of external memory accesses for a J-level 2-D DWT is limited to 2N 2 , which
corresponds to reading the input image for the first level and writing the final DWT results.
The line-based multilevel structure, shown in Fig. 17, is much faster than the previous
structures, but it needs a larger amount of hardware and so its hardware utilization (i.e., the
average value of the area of working parts versus the whole area of the hardware) is low [23],
but the 2-D recursive architecture proposed in [24] improves the hardware utilization for the
J-level 2-D DWT. In Fig. 17, the required internal memory for the 9/7 wavelet is obtained from
equation (4).
1 1 1
5N × (1 + + + · · · + ( ) J −1 ) (4)
2 4 2
External memory J-level

(N*N) 2-D DWT
Fig. 17. Multilevel line-based method

There is a trade-off between the size of the internal memory and the number of external
memory accesses in the 2-D DWT structures mentioned previously. Now, a block-based
structure that parameterizes the aforementioned trade-off is introduced. The block-based
structure is similar to the line-based method, but instead of considering the total length of
a row for DWT in the row direction, only a part of it with length M pixels is considered (Fig.
18). It means that the first M columns of the main frame (the gray area in Fig. 18) are used
as the input frame and 2-D DWT coefficients are computed for them. So the required internal
memory, which is determined by the length of the rows, is decreased. As an example, for the
9/7 wavelet the internal memory size will be decreased from 5N to 5M (where M is a fraction
of N). It is possible to consider a block of image by partitioning the image in both the row and
column directions. In this method, the block (or window) slides across the image and both
the row- and column-wise 1-D DWT will be performed on them [25]. The size of tile windows
may be reduced to 2 × 2 pixels [26].
52
Fig. 18. Scan method in block-based structure
6. Scan methods for block-based structure

However, in the above-mentioned method, there is a problem in the boundary region between
two M-pixel sections. To compute the DWT for the beginning pixel of the nextM-pixel section,
values of K previous pixels are needed. These K pixels produce values of temporary memory
(black circles shown in Fig. 19). K is equal to n t − 2, where nt is the number of filter taps
corresponding to the desired DWT. For the 9/7 wavelet, as shown in Fig. 19, K is equal to
7 (shaded circles). To solve the boundary problem, the overlapped scan method has been
Fig. 19. Boundary region for 9/7 wavelet

proposed in [27]. A new M-pixel section begins from the last K pixels in the previous section.
So two sections are overlapped in K pixels, and this causes the number of external memory
N 2 M instead of N 2 . The number of external memory writes is limited to writing
reads to be ( M −K )
the output results and is equal to N 2 .
We can use an alternate scan method for the overlapped region at the boundary between two
M-pixel sections. The relationships for the new scan method are different from the previous
method. The temporary memory at the boundary of two M-pixel sections (black circles in Fig.
19) can be saved for the computation of the next section. These saved values will be used
for the computation of the next M pixels. Hence, a new M-pixel section begins without any
overlap with the previous section. The required memory to save temporary data is L × N,
where L is the value of the temporary memory in the related DWT core. For example, in Fig.
VLSI Architectures
13
Direct Single level line-based Block-Based

Internal Memory Size 0 5N 5M 5M
External Memory Reads 2N 2 N2 N2 M N 2 (1 + M
4 ) − 4N
M−K
External Memory Write 2N 2 N2 N2 N 2 (1 + M
4 ) − 4N
Control complexity Low Medium Medium Medium

Table 3. Comparison of different 2-D DWT structures for one level 9/7 wavelet
19, L is 4 for the 9/7 wavelet without pipelining. The storage of temporary memory may be
fulfilled in internal or external memory. If internal memory is used to save temporary data, the
number of external memory accesses is equal to N 2 read and N 2 write operations. However,
if external memory is used to save temporary memory, both of the external memory reads and
writes are increased by an amount of ( M N − 1) N × L. Hence, the number of external memory
reads as well as the number of external memory writes is equal to N 2 + ( M N − 1) N × L. In
this expression, the ( M − 1) coefficient is the number of M-pixel sections. Due to hardware
N
limitations (the limit size of internal memory on FPGA ICs), we select the second case for
implementation. Comparisons of the aforementioned methods for one level 2-D DWT are
given in Table 3. The table shows the values of the internal memory size and external memory
accesses for the three algorithms. It is shown that in our proposed algorithm the internal
memory size is between those of two other algorithms (the direct method and the line-based
method). The same conclusion is true for the external memory size. The table also shows
the order of complexity for the control circuits of these methods based on [28]. Similar
comparisons for J-level (J −→ ∞) 2-D DWT are given in Table 4. Note that M has been
considered to be fixed for all levels of 2-D DWT. It means that the width of the M-pixel section
for the current level is the same as the one in the previous level. The J-level structure can
be implemented either in the form of a single level (Fig. 15) or multilevel structure (Fig.
16), and the relevant values are listed in Table 4. We observe that the parameter M can be
determined according to hardware limitations, such as internal memory. The conclusion from
the two tables is that the block-based structure with the new scan method, in comparison
with the direct method, needs more internal memory, but needs only about one-half of the
external memory accesses. Due to the shorter access time for internal memory, the clock
pulse frequency will increase, and based on the energy model in [5] the power consumption
will decrease. In comparison with other methods, the new method remarkably decreases
the needed internal memory at the cost of a soft increase in the number of external memory
accesses.
7. Experimental results
The folded 1-D DWT architecture was described in VHDL code and simulated by
Active-HDL6.3 software. Then the relevant VHDL code was synthesized by the Synplify7.5.1
software tool to be implemented on IC XC2V40 (from the VirtexII family of Xilinx FPGAs).
The maximum estimated frequency for implementing Fig. 12 on this IC is 122.4 MHz, which
is practical for real-time implementation of the 9/7 wavelet for large images. The maximum
frequency to implement the 5/3 wavelet on the IC is 163.1 MHz. Also, the block-based
architecture with the new scan method was modeled and simulated for the 5/3 wavelet with
N = 1024 and 8-bit pixels. The code was synthesized by Synplify7.5.1 for implementation on
VirtexII. After post place and route simulation, the clock pulse frequency achieved was 97
MHz. The structure receives one pixel as input per each clock pulse. So, according to the
54
Block-Based
Direct Line-Based Single Level Line-Based Multilevel Line-Based
Single Multi-
Level level (27) (30) (27) (30)
Internal
Memory 0 5N 10N 5M 5M (5M ) J (5M ) J
Size
External
4 N2 M 4 N2 K 16N 2
Memory 8 2
3N
4 2
3N N2 3 ( M−K ) 3
N (1 + M
4 2 4
) − 8N N2 + 3 M−K N2 + M − 8N
Reads
External
16N 2
3 N (1 + M ) − 8N N2 + − 8N
8 N2 4 2 4 2 4 2 4
Memory 3 3N N2 3N N2 M
Write
Table 4. Comparison of different 2-D DWT structures for J level 9/7 wavelet (J −→ ∞)
calculations below, the folded structure can be used to perform 3 levels of 2-D DWT for 70
frames (1024×1024 pixels) per second, and it is suitable for use in real-time hardware video
codec.
1024 × 1024 × (1 + 14 + 16
1
)
t= = 14.2ms,
97MHz
1 1
nf = = 70( f rame/s).
t 14.2ms
8. Conclusion
Lifting Scheme and some different lifting based architectures for DWT presented in this
chapter. Then we focused on the size (area) of the architecture. An architecture to minimize
the number of multipliers and adders has been investigated for implementation of 1-D DWTs.
All types of 1-D DWTs can be implemented by modifying only the number of registers and
coefficients of the architecture. Thus, the folded architecture, which has fixed form units for
all DWT types, presents a new folded method for systematic implementation of DWT. It is
possible to design a software program to produce the folded architecture for different types
of wavelets. What is needed is to define coefficients (α, β, ...) required for each step of the
desired wavelet in the lifting scheme. The folded method can be extended for large and
complex structures such as multilevel discrete wavelet packet transforms [29] to reduce the
area. Also, we have reviewed the 2-D DWT block-based structure and shown its power
to trade off between the internal memory size and the number of external accesses by a
controlling parameter.
9. References
[1] S. Mallat, A theory for multiresolution signal decomposition: the wavelet representation.
IEEE Trans. Pattern Anal. Mach. Intell. 11, 674–693 (1989)
[2] T. Acharya, C. Chakrabarti, A survey on lifting-based discrete wavelet transform
architectures. J. VLSI Signal Process. 42, 321–339 (2006)
[3] I. Daubechies, W. Sweldens, Factoring wavelet transform into lifting steps. J. Fourier
Anal. Appl. 4, 247–269 (1998)
VLSI Architectures
15
[4] C.-T. Huang, P.-C. Tseng, L.-G. Chen, Flipping structure: an efficient VLSI architecture
for liftingbased discrete wavelet transform, IEEE Trans. Signal Process. 52 (2004), pp.
1080–1089
[5] N.D. Zervas et al., Evaluation of design alternatives for the 2-D-discrete wavelet
transform. IEEE Trans. Circuits Syst. Video Technol. 11(12), 1246–1262 (2001)
[6] K. A. Kotteri, S. Barua, A. E. Bell, and J. E. Carletta, "A comparison of hardware
implementations of the biorthogonal 9/7 DWT: convolution versus lifting," IEEE Trans.
Circuits Syst. II, Expr. Br., vol. 52, no. 5, pp. 256-260, 2005.
[7] M. Maurizio, M. Guido, P. Gianluca, and Z. Maurizio, "Novel JPEG 2000 compliant DWT
and IWT VLSI implementations," J. VLSI Signal Process., vol. 35, no. 2, pp. 137-153, Sep.
2003.
[8] J. Reichel, On the arithmetic and bandwidth complexity of the lifting scheme, in Proc. of
International Conference on Image Processing (2001), pp. 198–201
[9] R. Calderbank, I. Daubechies, W. Sweldens, B.-L. Yeo, Wavelet transforms that map
integers to integers. Appl. Comput. Harmon. Anal. 5(3), 332–369 (1998)
[10] A. Jensen, A. La Cour-Harbo, Ripples in Mathematics: The Discrete Wavelet Transform
(Springer, Berlin, 2001)
[11] K.G. Oweiss et al., A scalable wavelet transform VLSI architecture for real-time signal
processing in high-density intra-cortical implants. IEEE Trans. Circuits Syst. 54(6),
1266–1278 (2007)
[12] C.-T. Huang, P.-C. Tseng, L.-G. Chen, Efficient VLSI architectures of lifting-based discrete
wavelet transform by systematic design method, in IEEE International Symposium on
Circuits and Systems, vol. 5 (2002), pp. 565–568
[13] W.-H. Chang, Y.-S. Lee, W.S. Peng, C.-Y. Lee, A line-based, memory efficient and
programmable architecture for 2D DWT using lifting scheme, in IEEE International
Symposium on Circuits and Systems, vol. 4 (2001), pp. 330–333
[14] K. Andra, C. Chakrabarti, T. Acharya, A VLSI architecture for lifting-based forward and
inverse wavelet transform. IEEE Trans. Signal Process. 50(4), 966–977 (2002)
[15] O. Fatemi, S. Bolouki, Pipeline memory-efficient and programmable architecture for 2D
discrete wavelet transform using lifting scheme, in Proceedings of IEE Circuits, Devices
and Systems, December 2005, pp. 703–708
[16] Lan, N. Zheng, Y. Liu, Low power and high-speed VLSI architecture for lifting-based
forward and inverse wavelet transform. IEEE Trans. Consumer Electron. 51(2), 379–385
(2005)
[17] C.Y. Xiong, J.W. Tian, J. Liu, A note on Şflipping structure: an efficient VLSI architecture
for liftingbased discrete wavelet transformŤ. IEEE Trans. Signal Process. 54(4), 1910–1916
(2006)
[18] J.M. Jou, Y.H. Shiau, C.C. Lio, Efficient VLSI architectures for the biorthogonal wavelet
transform by filter bank and lifting scheme, in Proceedings of IEEE ISCAS 2001, pp.
529–533
[19] C.J. Lian, K.F. Chen, H.H. Chen, L.G. Chen, Lifting based discrete wavelet transform
architecture for JPEG2000, in IEEE International Symposium on Circuits and Systems
(ISCAS 2001) Sydney, May 2001
[20] C.Y. Xiong, J.W. Tian, J. Liu, Efficient architectures for two-dimensional discrete wavelet
transform using lifting scheme. IEEE Trans. Image Process. 16(3), 607–614 (2007)
56
[21] P.-C. Tseng, C.-T. Huang, L.-G. Chen, Generic RAM-based architecture for
two-dimensional discrete wavelet transform with line-based method, in Asia-Pacific
Conference on Circuits and Systems (2002), pp. 363–366
[22] H. Varshney, M. Hasan, S. Jain, Energy efficient novel architectures for the lifting-based
discrete wavelet transform. IET Image Process. 1(3), 305–310 (2007)
[23] C.-T. Huang, P.-C. Tseng, L.-G. Chen, Hardware implementation of shape-adaptive
discrete wavelet transform with the JPEG2000 defaulted 9/7 filter bank, in IEEE
International Conference on Image Processing, vol. 3 (2003), pp. 571–574
[24] L.Hongyu, M.K. Mandal, B.F. Cockburn, Efficient architectures for 1-D and 2-D
lifting-based wavelet transforms. IEEE Trans. Signal Process. 52(5), 1315–1326 (2004)
[25] C.H. Yang et al., A block-based architecture for lifting scheme discrete wavelet transform.
IEICE Trans. Fundam. 90(5), 1062–1071 (2007)
[26] J.W. Kim et al., Tiled interleaving for multi-level 2-D discrete wavelet transform, in IEEE
Int. Symp. Circuits Syst., May 2007, pp. 3984–3987
[27] C.-T. Huang, P.-C. Tseng, L.-G. Chen, Memory analysis and architecture for
two-dimensional discrete wavelet transform, in IEEE International Symposium on
Circuits and Systems (ISCAS 2004)
[28] C.-T. Huang, P.-C. Tseng, L.-G. Chen, Analysis and VLSI architecture for 1-D and 2-D
discrete wavelet transform. IEEE Trans. Signal Process. 53(4), 1575–1586 (2005)
[29] C. Wang, W.S. Gan, Efficient VLSI architecture for lifting-based discrete wavelet packet
transform. IEEE Trans. Circuits Syst. 54(5), 422–426 (2007)
[30] S.A. Salehi, S. Sadri, "Investigation of Lifting-Based Hardware Architectures for Discrete
Wavelet Transform," Journal of Circuits Systems and Signal Processing, vol. 28, N.1,
pp1-16, 2009.
4
Simulation of Models and BER Performances of

DWT-OFDM versus FFT-OFDM
Khaizuran Abdullah1 and Zahir M. Hussain2,3
1Electricaland Computer Engineering Department,
International Islamic University Malaysia,
2Department of Electrical Engineering, University of Kufa,
3School of Electrical and Computer Engineering, RMIT University,
1Malaysia
2Iraq
3Australia
1. Introduction
Orthogonal Frequency Division Multiplexing (OFDM) is a multicarrier modulation system.
The transmission channel is divided into a number of subchannel in which each subchannel is
assigned a subcarrier. Conventional OFDM systems use IFFT and FFT algorithms at the
transmitter and receiver respectively to multiplex the signals and transmit them
simultaneously over a number of subcarriers. The system employs guard intervals or cyclic
prefixes (CP) so that the delay spread of the channel becomes longer than the channel impulse
response (Peled & Ruiz, 1980; Bahai & Saltzberg, 1999; Kalet, 1994; Beek et al.,1999; Bingham,
1990; Nee and Prasad, 2000). The system must make sure that the cyclic prefix is a small
fraction of the per carrier symbol duration (Beek et al.,1999; Steendam & Moeneclaey, 1999).
The purpose of employing the CP is to minimize inter-symbol interference (ISI). However a CP
reduces the power efficiency and data throughput. The CP also has the disadvantage of
reducing the spectral containment of the channels (Ahmed, 2000; Dilmirghani & Ghavami,
2007, 2008). Due to these issues, an alternative method is to use the wavelet transform to
replace the IFFT and FFT blocks (Ahmed, 2000; Dilmirghani & Ghavami, 2007, 2008; Akansu &
Xueming, 1998; Sandberg & Tzannes, 1995). The wavelet transform is referred as Discrete
Wavelet Transform OFDM (DWT-OFDM). By using the transform, the spectral containment of
the channels is better since they are not using CP (Ahmed, 2000; Dilmirghani & Ghavami,
2007, 2008). The illustration of the superior subchannel containment attributes in wavelet has
been described in detailed by (Sandberg & Tzannes, 1995) as compared to Fourier. The wavelet
transform also employs Low Pass Filter (LPF) and High Pass Filter (HPF) operating as
Quadrature Mirror Filters satisfying perfect reconstruction and orthonormal bases properties.
It uses filter coefficients as approximate and detail in LPF and HPF respectively. The
approximated coefficients is sometimes referred to as scaling coefficients, whereas, the detailed
is referred to wavelet coefficients (Abdullah et al., 2009; Weeks, 2007). In some literatures, these
two filters are also called subband coding since the signals are divided into sub-signals of low
and high frequencies respectively. The purpose of this chapter is to show the simulation study
of using the Matrices Laboratory (MATLAB) on the wavelet based OFDM particularly DWT-
OFDM as alternative substitutions for Fourier based OFDM. MATLAB is preferred for this
approach because it offers very powerful matrices calculation with wide range of enriched
toolboxes and simulation tools. To the best of the authors’ knowledge, there is no study on the
descriptive procedures of simulations using MATLAB with regards of flexible transformed
models in an OFDM system, especially when dealing with wavelet transform. Therefore, this
chapter is divided into three main sections: section 2 will explain conventional FFT-OFDM,
section 3 will describe in detail the models for DWT-OFDM, and section 4 will discuss the Bit
Error rate (BER) result regarding those two transformed platforms, DWT-OFDM versus FFT-
OFDM.
2. Fourier-based OFDM
A typical block diagram of an OFDM system is shown in Figure 1. The inverse and forward
blocks can be FFT-based or DWT-based OFDM.
Fig. 1. A Typical model of an OFDM transceiver with inverse and forward transformed
blocks which can be substituted as FFT-OFDM or DWT-OFDM.
The system model for FFT-based OFDM will not be discussed in detail as it is well known in
the literature. Thus, we merely present a brief description about it. The data dk is first being
processed by a constellation mapping. M-ary QAM modulator is used for this work to map
the raw binary data to appropriate QAM symbols. These symbols are then input into the
IFFT block. This involves taking N parallel streams of QAM symbols (N being the number of
sub-carriers used in the transmission of the data) and performing an IFFT operation on this
parallel stream. The output in discrete time domain is as follows:
N 1 n
1 j 2 i
X k (n) 
N
 Xm (i )e N (1)
i 0
Where xk(n) | 0 ≤ n ≤ N −1, is a sequence in the discrete time domain and Xm(i) | 0 ≤ i ≤ N −
1 are complex numbers in the discrete frequency domain. The cyclic prefix (CP) is lastly
added before transmission to minimize the inter-symbol interference (ISI). At the receiver,
the process is reversed to obtain the decoded data. The CP is removed to obtain the data in
the discrete time domain and then processed to FFT for data recovery. The output of the FFT
in the frequency domain is as follows:
N 1 n
j 2 i
U m (i )   U k ( n )e N (2)
i 0
Simulation of Models and BER Performances of DWT-OFDM versus FFT-OFDM 59
3. Wavelet-based OFDM
As mentioned in the previous section, the inverse and forward block transforms are flexible
and can be substituted with FFT or DWT-OFDM. We have discussed briefly about FFT-
OFDM. Thus, this section will describe wavelet based OFDM particularly about DWT-
OFDM transceiver. This section is divided into three parts: a description of the DWT-OFDM
transmitter and receiver models as well as the Perfect Reconstruction properties’ discussion.
3.1 Discrete Wavelet Transform (DWT) transmitter

From Figure 1, it is obvious that the transmitter first uses a 16 QAM digital modulator which
maps the serial bits d into the OFDM symbols Xm, within N parallel data stream Xm(i) where
Xm(i) |0 ≤ i ≤ N − 1. The main task of the transmitter is to perform the discrete wavelet
modulation by constructing orthonormal wavelets. Each Xm(i) is first converted to serial
representation having a vector xx which will next be transposed into CA as shown in details
as in Figure 2. This means that CA not only its imaginary part has inverting signs but also its
form is changed to a parallel matrix. Then, the signal is up-sampled and filtered by the LPF
coefficients or namely as approximated coefficients. This coefficients are also called scaling
coefficients. Since our aim is to have low frequency signals, the modulated signals xx
perform circular convolution with LPF filter whereas the HPF filter also perform the
convolution with zeroes padding signals CD respectively. Note that the HPF filter contains
detailed coefficients or wavelet coefficients. Different wavelet families have different filter
length and values of approximated and detailed coefficients. Both of these filters have to
satisfy orthonormal bases in order to operate as wavelet transform. The number of CA and
CD depends on the OFDM subcarriers N. Samples of this processing signals CA and CD that
pass through this block model is shown in Figure 4. The above mentioned signals are
simulated using MATLAB command [Xk] = idwt(CA;CD;wv) where wv is the type of wavelet
family.
Fig. 2. Discrete Wavelet Transform (DWT)-OFDM transmitter model.

The detailed and approximated coefficients must be orthogonal and normal to each other.
By assigning g as LPF filter coefficients and h as HPF filter coefficients, the orthonormal
bases can be satisfied via four possible ways (Weeks, 2007): <g, g*>= 1, <h, h*>= 1, <g, h*>= 0
and <h, g*>= 0. The symbol * indicates its conjugate, and the symbol < , > is referring to the
dot product. The result which yields to 1 is related to the normal property whereas the
result yielding to 0 is for orthogonal property accordingly.
Fig. 3. The processed signals of one symbol DWT-OFDM system using bior5.5 in DWT
transmitter. Top: data CA, Middle: data CD, Bottom: data Xk, corresponding to Figure 2.
Both filters are also assumed to have perfect reconstruction property. The input and output
of the two filters are expected to be the same. A further discussion can be found in section
3.3.
3.2 Discrete Wavelet Transform (DWT) receiver

The DWT receiver is the reverse process which is simulated using the MATLAB command
[ca; cd] = dwt(Uk;wv). The receiver system model that processes the data ca, cd and Uk is
shown in Figure 4. The parameter wv is to indicate the wavelet family that is used in this
simulation. Uk is the front-end receiver data. This data is decomposed into two filters, high
and low pass filters corresponding to detailed and approximated coefficients accordingly.
The ca signal which is the output of the approximated coefficients or low pass filter will
finally be processed to the QAM demodulator for data recovery. To perform that operation,
data is first transposed before converting into parallel representation. The output Um(i) is
passed to QAM demodulator. The index i depends on the number of OFDM subcarriers.
The data cd is explained next. Due to the effect of CD data generated in the transmitter, Uk
has some zeroes elements which is decomposed as the detailed coefficients. The signal
output of these coefficients is cd. Comparing to ca, the cd signal is discarded because it does
not contain any useful information instead. Samples of this processing signals that pass
through the DWT-OFDM receiver model is shown in Figure 5.
Fig. 4. Discrete Wavelet Transform (DWT)-OFDM receiver model.
Fig. 5. The processed signals of one symbol DWT-OFDM system using bior5.5 in DWT
receiver. Top: data ca. Middle: data cd. Bottom: data Uk, corresponding to Figure 4.
3.3 Perfect reconstruction

A block diagram of perfect reconstruction (PR) system operation is illustrated in Figure 6.
The PR property is performed by a two-channel filter bank which is represented by the LPF
and HPF. The first level of analysis filter in the receiver part can be folded and the decimator
and the expander are cancelled out by each other.
Fig. 6. A simple and modified model of two-channel filter bank illustrating a perfect
reconstruction property with the superscript number is referring to the steps.
To satisfy a perfect reconstruction operation, the output Yk(i) is expected to be the same as
Xk(i). With the exception of a time delay, the input can be considered as Yk(i) = Xk(i-n) where
n can be substituted as 1 to describe this simple task. The steps to perform the mathematical
operation of PR can be summarized as follows (Weeks, 2007):
1. Selecting the filter coefficients for ga, i.e., a and b. Thus, ga = {a; b}.
2. ha is a reversed version of ga with every other value negated. Thus, ha = {b;−a}. If the
system has 4 filter coefficients with ga = {a; b; c; d}, then ha = {d;−c; b;−a}.
3. hs is the reversed version of ga, thus hs = {b; a}.
4. gs is also a reversed version of ha, therefore gs = {−a; b}.
The above steps can be rewritten as follows:
ga ={a,b}, ha={b,-a}, hs={b,a}, gs={-a,b} (3)

Considering that the input with delay are applied to ha and ga in Figure 4, then the output of
these filters are
Zk(i) = b(Xk(i) − a(Xk(i-1)) (4)
Wk(i) = a(Xk(i) + b(Xk(i-1)) (5)

Considering also that Zk(i) and Wk(i) are delayed by 1, then i can be replaced by (i-1) as
follows
Zk(i-1) = a(Xk(i-1) + b(Xk(i-2)) (6)
Wk(i-1) = b(Xk(i-1) − a(Xk(i-2)) (7)

The output Yk(i) can be written as
Yk(i) = gsZk(i) + hsWk(i) (8)

or,
Yk(i) = −aZk(i) + bZk(i-1) + bWk(i) + aWk(i-1) (9)

Substituting equations (5), (6), (7) and (8) into (9) yields to
Yk(i) = 2(a2 + b2)Xk(i-1) (10)

The output Yk(i) is the same as the input Xk(i) except that it is delayed by 1 if we substitute
the coefficient factor 2(a2 +b2) by 1. The PR condition is satisfied.
4. Simulation results
Simulation variables and their matrix values are shown in Table I. The number of samples
for the subcarriers N is 64, and the number of samples for the symbols ns is 1000. Data is
similar between FFT and DWT OFDM in all parameters except the multiplexed one. For
DWT-OFDM, it is required the transmitted signal to have double the data of FFT-OFDM.
This is due to the fact that the DWT transmitter has zeroes padding component. An element
value in the table that has a multiplier is referred to its matrix representation of row and
column. If the element has 64 x 1000, it means that it has 64 numbers of rows and 1000
numbers of columns.
Variables and Parameters FFT-OFDM DWT-OFDM
Minimum Subcarriers 64 64
requirement OFDM symbols 1000 1000
input binary generated 64 x 1000 64 x 1000
parallel transmitted data 64 x 1000 64 x 1000
Transmitter
serial transmitted data 1 x 64000 1 x 64000
multiplexed data transmitted 64000 x 1 128000 x 1
multiplexed data received 64000 x 1 128000 x 1
serial received data 1 x 64000 1 x 64000
Receiver
parallel received data 64 x 1000 64 x 1000
output binary recovered 64 x 1000 64 x 1000
Table 1. Simulation variables and their matrix values.

The curves in Figure 8 could have been better if we used more number of samples for the
symbols. However, this yields longer time of running the simulations. Other variables are
listed according to their use as in Figures 1, 2 and 3. Figure 7 shows the OFDM symbols in
time domain for the two transformed platforms FFT and DWT. Some of the simulation
parameters related to this figure are: the OFDM symbol period To = 9 ms, the total
simulation time t = 10 × To = 90 ms, the sampling frequency fs = 71.11 kHz, the carriers
spacing ΔN = 1.11 kHz and the bandwidth B = ΔN × 64 = 71.11 kHz. Thus, the simulation
satisfied the Nyquist criterion where fs < 2B. Both platforms used the same parameters. It is
interesting to observe that the DWT-OFDM symbol is less in term of the mean of amplitude
vectors as compared to FFT-OFDM. The mean of FFT is 1.4270, whereas, the mean of DWT
is -9.667E-04. This is due to the fact that zero - padding was performed in the DWT
Fig. 7. An OFDM symbol in time domain: FFT-OFDM (Top), DWT-OFDM (Bottom).
Fig. 8. BER performance for DWT-OFDM.

(transmitter) system model. As a result, most samples in the middle of DWT-OFDM symbol
is almost zeroes. The DWT-OFDM performance can be observed from Figure 8. The wavelet
families Biorthogonal and Daubechies are compared with FFT-OFDM. It is shown that
bior5.5 is superior among all others. It outperforms FFT and Daubechies by about 2 dB and
bior3.3 by 8 dB at 0.001 BER.
5. Conclusions
Simulation approaches using MATLAB for wavelet based OFDM, particularly in DWT-
OFDM as alternative substitutions for Fourier based OFDM are presented. Conventional
OFDM systems use IFFT and FFT algorithms at the transmitter and receiver respectively to
multiplex the signals and transmit them simultaneously over a number of subcarriers. The
system employs guard intervals or cyclic prefixes so that the delay spread of the channel
becomes longer than the channel impulse response. The system must make sure that the
cyclic prefix is a small fraction of the per carrier symbol duration. The purpose of employing
the CP is to minimize inter-symbol interference (ISI). However a CP reduces the power
efficiency and data throughput. The CP also has the disadvantage of reducing the spectral
containment of the channels. Due to these issues, an alternative method is to use the wavelet
transform to replace the IFFT and FFT blocks. The wavelet transform is referred as Discrete
Wavelet Transform OFDM (DWT-OFDM). By using the transform, the spectral containment
of the channels is better since they are not using CP. The wavelet based OFDM (DWT-
OFDM) is assumed to have ortho-normal bases properties and satisfy the perfect
reconstruction property. We use different wavelet families particularly, Biorthogonal and
Daubechies and compare with conventional FFT-OFDM system. BER performances of both
OFDM systems are also obtained. It is found that the DWT-OFDM platform is superior as
compared to others as it has less error rate, especially using bior5.5 wavelet family.
6. References
Abdullah, K.; Mahmoud, S. & Hussain, Z.M. (2009). Performance Analysis of an Optimal
Circular 16-QAM for Wavelet Based OFDM Systems. International Journal of
Communications, Network and System Sciences (IJCNS), Vol. 2, No. 9, (December
2009), pp 836-844, ISSN 1913-3715.
Ahmed, N. (2000). Joint Detection Strategies for Orthogonal Frequency Division
Multiplexing. Dissertation for Master of Science, Rice University, Houston, Texas.
pp. 1-51, April.
Akansu, A. N. & Xueming, L. (1998). A Comparative Performance Evaluation of DMT
(OFDM) and DWMT (DSBMT) Based DSL Communications Systems for Single and
Multitone Interference, Proceedings of the IEEE International Conference on
Acoustics, Speech and Signal Processing, vol. 6, pp. 3269 - 3272, May.
Bahai, A. R. S. & Saltzberg, B. R. (1999). Multi-Carrier Digital Communications - Theory and
Applications of OFDM. Kluwer Academic. ISBN: 0-306-46974-X 0-306-46296-6. New
York.
Baig, S. R.; Rehman, F. U. & Mughal, M. J. (2005). Performance Comparison of DFT, Discrete
Wavelet Packet and Wavelet Transforms in an OFDM Transceiver for Multipath
Fading Channel. 9th IEEE International Multitopic Conference, pp. 1-6, December.
Bingham, J. A. C. (1990). Multicarrier Modulation for Data Transmission: An Idea Whose

Time Has Come. IEEE Communications Magazine, Vol. 28, no 5, pp. (5-14).
Dilmirghani, R. & Ghavami, M. (2007). Wavelet Vs Fourier Based UWB Systems, 18th IEEE
International Symposium on Personal, Indoor and Mobile Radio Communications,
pp.1-5, September.
Kalet, I. (1989). The multitone Channel. IEEE Transactions on Communications, Vol. 37, No. 2,
(Feb 1989), pp. (119-124).
Mirghani, R. & Ghavami, M. (2008). Comparison between Wavelet-based and Fourier-based
Multicarrier UWB Systems. IET Communications, Vol. 2, Issue 2, pp. (353-358).
Nee, R. V. & Prasad, R. (2000). OFDM for Wireless Multimedia Communications, Boston: Artech
House. ISBN 0-89006-530-6.
Peled, A. & Ruiz, A. (1980). Frequency Domain Data Transmission Using Reduced
Computational Complexity Algorithms, Proceedings IEEE International
Conference on Acoustics, speech and Signal Processing (ICASSP 1980), Denver, pp.
964-967.
Sandberg, S. D. & Tzannes, M. A. (1995). Overlapped Discrete Multitone Modulation for
High Speed Copper Wire Communications. IEEE Journal on Selected Areas in
Communications, vol. 13, no. 9, pp. (1571-1585).
Steendam, H. & Moeneclaey, M. (1999). Analysis and Optimization of the Performance of
OFDM on Frequency-Selective Time-Selective Fading Channels. IEEE Transactions
on Communications, Vol. 47, No. 12, (Dec 99), pp. (1811- 1819).
Van de Beek, J. J.; Dling, P.; Wilson, S. K. & Brjesson, P. O. (1999). Orthogonal Frequency
Division Multiplexing (OFDM), URSI Review of Radio Science 1996-1999, Oxford
Publishers.
Weeks, M. (2007). Digital Signal Processing Using Matlab and Wavelets, Infinity Science Press
LLC. ISBN-10: 0-7637-8422-2.
5
Several Kinds of Modified SPIHT Codec

Wenchao Zhang
Institute of Electronics, Chinese academy of science
People’s Republic of China
1. Introduction
SPIHT (Set Partitioning In hierachical trees), being an efficient coding method for wavelet
coefficients, has acquired more and more widely application, especially in image/video
compression fields. But, conventional SPHIT have some obvious limitions. For example,
when for the color image compression, polarmetric SAR image compression, or multi-
spectrum image compression and other multi-channel image compression, there are only
very limited image planes(R,G,B for color image, HH, HV,VVfor polarmetric SAR images)
but there exist large amount of information redundancy among the image planes. So,
considering the support length of discrete wavelet transform, we can’t use the set 8-partition
methods such as 3D-SPIHT which has been used for video compression. Another example,
when the input image is unsymmetrical such as 16x512 image block even only one line
image content, which is very common in hardware design, because it means the larger final
chip die size for the many line buffers. But, the traditional SPIHT codec can acquire the best
compression performances only for the image is symmetrical in horizontal and vertical
dimension. To the above specific applications, I will discuss several kinds of modified
SPIHT in this chapter, most of them are the author’s newly research result.
In the following section, We will give a simple description about the traditional SPIHT
codec, then, we will take the polarimetric SAR intensity image compression as example and
give some specific compression method for multi-channel image compression. Certainly,
before encoding for the wavelet coefficients, we need a 3D matrix transform to remove the
information redundancy among the image plane and in each image plane. For the
unsymmetrical image compression used in hardware design, an unsymmetrical SPIHT
codec is detailed addressed, at the time, its specific case for only 1 line image compression,
1D SPIHT codec is also given.
2. Traditional SPIHT codec[1]

SPIHT can be used to compress the traditional image, but it need 2D discrete wavelet
transform (DWT) before encode the wavelet coefficients. The correlation of DWT coefficients
not only exist in each subband inside but also among different subbands. Conventional
SPIHT codec is to encode coefficients by SOT (Spatial Orientation Tree), which utilize the
correlation of the same subband and different subbands. Conventional SPIHT codec, having
many advantages such as simpleness, embedded code stream when compared with other
encoders, is a very efficient compression method.
According the SOT structure, the set partitioning can be defined as:
T (i , j )  c(i , j )  D(i , j )

D(i , j )  O( i , j )  L(i , j ) (1)
L(i , j )  D( k , l ) ( k , l )  O( i , j )
 
Here, T (i , j ) is spatial orientation tree and c(i , j ) is a root node of the tree; D(i , j ) , O(i , j )
and L(i , j ) are node c(i , j ) ’s descendant node set, direct descendant node set and indirect
descendant node set. Direct node set O(i , j ) can be further partitioned into 4 nodes just as
the following:
O(i , j ) = {c(2 i , 2 j ), c(2i , 2 j  1), c (2 i  1, 2 j ), c(2 i  1, 2 j  1)} (2)
Conventional SPIHT encoding process can be divided into sorting pass and refinement pass.
During the encoding process, 3 lists are used to record the corresponding encoding
information, which include List of insignificant set (LIS), list of insignificant pixel (LIP) and
list of significant pixel (LSP). The basic operation is significance test and the significance test
function is just as the following:
1,

Sn T   
 
max c i , j  2
n
(3)
 i , j T

 0, otherwise ,
The encoding process can be simply described as:

Output the initialized threshold n   log 2 (max c(i , j )  ; set LSP as NULL; add the root nodes
having and no having descendant nodes to LIS and LIP respectively.
For a spatial orientation tree T (i , j ) , its nodes are partitioned into nodes c(i , j ) and
descendant nodes D(i , j ) according to formula 1. If node c(i , j ) is significant, move it from
LIP into LSP. If D(i , j ) is significant, then partition D(i , j ) into O(i , j ) and L(i , j ) . The
significant nodes of O(i , j ) will also be moved to LSP according to the significance test. If
L(i , j ) is significant, then L(i , j ) will also be partitioned 4 set and test the significance of
every set. This process will be continued on.
Sorting pass is for every element c(i , j ) of LSP; output the nth efficient bit of c(i , j ) for the
current threshold n.
Decrease the n and continue the sorting pass and refinement pass until meet the required bit
number or the end of encoding.
The decoding process is same to the encoding process, which, compared with EZW, make it
un-necessary to transmit position information because the position information is
embedded in the encoding process.
3. Multi-channel image compression

SPIHT, utilizing the information redundancy of 2D wavelet coefficients and adopting set 4
division, can be used to compress 2D image data. Naturally, it can be enlongated to 3D
video compression, that is to say, if we adopt 3D DWT to remove the information
redundancy of video data and choose set 8-partition instead of set 4-partition to encode 3D
Several Kinds of Modified SPIHT Codec 69
wavelet transform coefficients, this method is called 3D-SPIHT. But at some cases, we only
need to compress multi-channel image, such as color images (R,G, B), polarimetric SAR
image (HH,HV,VV), multi-spectrum image, ect. Because the third dimension usually have
only limited image planes and can’t be processed by supported discrete wavelet transform,
however there exist much information redundancy among each image channel. In this case,
consider it’s simplity, 2D DWT for each image plane and 1D DCT can be used to remove the
information redundancy. In the following section 3.1-3, taking polarimetric SAR image as
example, 3D matrix transform and the related compression method including bit allocation
based encoding method and 3D SPIHT Embedded method will be addressed.
3.1 3D matrix transform for multi-channel image

At first, 3D-matrix [2-3] is adopted to represent the multi-polarimetric SAR intensity images.
The 1st, 2nd and 3rd planes represent the HH, HV and VV polarimetric SAR intensity image
respectively. Assuming the image dimensions are: M  N , the multi-polarimetric SAR
images can be written as a M  N  3 matrix f ( x , y , z) ( x  0,1, 2, M  1;
y  0,1, 2 N  1; z  0,1, 2) . In order to remove the redundancy of the matrix, DWT, KLT,
DCT and other linear transforms can be used.
Because there are only 3 components in polarimetric channel and KLT is dependent on the
statistic properties of data, DCT other than DWT or KLT or other linear transforms is chosen
to remove the redundancy of polarimetric channels. According to DCT transform definition,
the DCT transform can be represented as:
1 2
F( x , y ,0)   f ( x , y , z)
3 z0
(4)
2 2 (2 z  1)Z
F( x , y , Z )   f ( x , y , z)cos 6
3 z0
(Z  1, 2)
The 3D-matrix can be composed in other orders, such as HH, VV, HV acted as the 1st, 2nd
and 3rd planes respectively. The components of like-polarimetric(HH and VV) have strong
correlation, while the components of cross-polarimetric (HH/VV and HV) have weak
correlation. Both the DCT theory and experiment results prove that the DCT coefficients of
the matrix composed of HH, HV and VV are the most concentrated and that the decoded
images have the least loss at the same coding rate
After 1D-DCT transform, the data power of the 3D-matrix is concentrated into the 1st plane
and the redundancy among three data planes decreases greatly. In order to remove the
redundancy in every data plane of the 3D matrix, 2D-DWT is chosen. According to the
definition of 2D-DWT transform, the image data can be decomposed into horizontal,
vertical, diagonal and low frequency components after horizontal and vertical filtering. The
low frequency component can be decomposed further. After many level discrete wavelet
transform, the data power is concentrated to the low frequency components. After 1D-DCT
transform and 2D-DWT transform, the power of the whole 3D-matrix is concentrated onto
the top left corner. 3 level wavelet transform is adopted, so each data plane is decomposed
into 10 subbands including LL3, HL3, LH3 HH3, HL2, LH2, HH2, HL1, LH1, HH1. Of all the
subbands, the LL3 subband of the 1st plane has the highest power.
The 3D-matrix transform of Multi-polarimetric SAR intensity images (1D-DCT among
polarimetric channel and 2D-DWT in each polarimetric plane) is illustrated in Figure.1. The
1D-DCT and 2D-DWT are linear transform, so, the operation 1DCTz and 2DWTx,y can be
inverted in sequence, that is:
1DCTz (2 DWTx , y ( f ( x , y , z)))  2 DWTx , y (1DCTz ( f ( x , y , z))) (5)
Fig. 1. Illustration of 3D-matrix transform of multi-polarimetric SAR intensity images.
3.2 Bit allocation based on differential entropy encoding

The three mixed coefficient planes have different energy, which means different importance.
Different bit allocation scheme will greatly affect the quality of decoding images even at the
same mean quantization bit rate. Our aim is to find the optimal quantization bits allocation
to make the decoding images have the least distortion, which is a constrained optimal
problem. The rate distortion function R(D) is unknown in most cases, so its solutions have
become a hot research issue. Two variant iterate search [4], Dynamic programming [5] and
R/D curve modeling [6] are proposed, but these methods yet haven’t been adopted for its
enormous computation complexity or the local optimal not the whole optimal bit allocation
searched. In this subsection, a new bit allocation method based on differential entropy is
proposed. Compared with the existing bit allocation methods, it is easy to find a better bit
allocation in whole with the proposed method. At the same time, the computation is not
very complex.
Let R1 , R2 , R3 be the optimal bit rate for the three mixed coefficient plane respectively and
RT be the mean bit rate; Let D( R ) be distortion rate function. So, the optimal bit allocation
problem can be represented as:
 min (D1 ( R1 )  D2 ( R2 )  D3 ( R3 ))
R ,R ,R
 1 2 3 (6)
subject to R1  R2  R3  3RT
Adopt the Lagrangian multiplier,
F( R1 , R2 , R3 )  D1 ( R1 )  D2 ( R2 )  D3 ( R3 )   (3 RT  R1  R2  R3 ) (7)
 F
 0
 R1
 F
From  0
 R2
 F
 0
 R3
we can acquire
 D1 ( R1 )
 
 R1
 D2 ( R2 )
  (8)
 R2
 D3 ( R3 )
 
 R3
According to the inequation of rate distortion function for non-Gauss continuous
information source, we have:
1 1 2
h(U )  log 2 eD  R(D)  log (9)
2 2 D
2
where h(U ) is the differential entropy of information source and  is its mean square
error.
At high bit rate, the lower bound of the inequation approaches the real rate distortion
function for most probability distributed information source. So, we can let the lower bound
equal to the real rate distortion function and then acquire:
1
R(D)  h(U )  log 2 eD (10)
2
Further, we can acquire
e 2( h(U ) R )
D( R )  (11)
2 e
and
D( R )  e 2( h(U ) R )
 (12)
R e
For every mixed coefficient plane:
 1
 R1  h(U 1 )  2 log(  e )

 1
 R2  h(U 2 )  log(  e ) (13)
 2
 1
 R3  h(U 3 )  2 log(  e )

According to the constrain condition R1  R2  R3  3 RT , we can acquire the final bit

allocation for every coefficient plane:
 2 h(U 1 )  h(U 2 )  h(U 3 )

 R1  RT  3

 2 h(U 2 )  h(U 1 )  h(U 3 )
 R2  RT  (14)
 3
 2 h(U 3 )  h(U1 )  h(U 2 )
 R3  RT  3

From formula (14), we can acquire the optimal bit allocation for the three mixed coefficient
planes, then the conventional SPIHT can be adopted to encode the coefficients for every
plane according to the corresponding allocated bit rate.
3.3 3D-SPIHT embedded encoding

The formula (14) provides a method to compute the bit rate allocation，but the bit allocation
accuracy strongly depends on the degree of the lower bound of formula (9) approaches the
real rate distortion function. So, in most cases, we only can acquire an approximating
optimal bit allocation. In this subsection, a new method named 3D-SPIHT embedded
algorithm is proposed, which extends the conventional SPIHT algorithm to 3D case. It
avoids bit allocation by adopting 3D-SPIHT to encode the three mixed coefficient plane
entirely, so the optimal bit rate allocation can be ensured.
During the 3D-SPIHT encoding process, the following sets and lists will be used:
H iplane (i , j ) , Diplane ( i , j ) , Oiplane ( i , j ) and Liplane ( i , j ) represent the iplaneth coefficient plane’s
root node, descendant’s nodes of node(i,j), offspring’s nodes of node(i,j) and indirect
offspring’s nodes of node(i,j) respectively. LISiplane, LIPiplane and LSPiplane represent the
iplaneth plane’s list of insignificant pixel sets, list of insignificant pixels and list of significant
pixels respectively, where iplane  1, 2, 3 .
Just as the conventional SPIHT algorithm, the set splitting procedure for every coefficient
plane can be defined as following:
Tiplane (i , j )  H iplane (i , j )  Diplane (i , j )


Diplane (i , j )  Oiplane (i , j )  Liplane (i , j ) (15)

Liplane (i , j )   Diplane ( k , l ) ( k , l )  Oiplane ( i , j )
The 3D-SPIHT process can be divided into sorting pass and refinement pass, whose basic
operations is also significance test of set just as the conventional SPIHT algorithm. That is，
1, max C
  (i , j )  2 n
 
iplane
Sn Tiplane   i , j Tiplane (16)

0, otherwise
where C iplane (i , j ) is the wavelet coefficient of coordinate (i , j ) in the iplaneth mixed

coefficient plane.
The followings are the coding process:

1. Initialization
n _ max  log max

 1
 2  
 i , j   C 1 (i , j ) 

Output    
n _ max 2  log 2 max i , j   C 2 (i , j ) 
(17)

 
n _ max3  log 2 max i , j   C 3 ( i , j ) 


 n  n _ max1
(For the HH HV VV combination, n _ max1  n _ max 3  n _ max 2 can always be satisfied)

For every coefficient plane, set list of significant pixels (LSPiplane) as NULL, then add the
coordinate (i , j )  H iplane to LIPiplane and those that have descendants to LISiplane, at the same
time, set their type to be A.
2. Sorting pass
If n _ max 3  n , only sort C 1 (i , j ) just as conventional SPIHT algorithm.
If n _ max 2  n  n _ max 3 , sort C 1 (i , j ) and then sort C 3 (i , j ) in succession.
If n  n _ max 2 , sort C 1 (i , j ) and then sort C 2 (i , j ) , C 3 (i , j ) in succession.
3. Refinement pass
If n _ max 3  n , for any entry (i,j) in LSP1 except those included in the last sorting pass,
output the nth most significant bit of C 1 (i , j ) .
If n _ max 2  n  n _ max 3 , for any entry (i,j) in LSP1 and LSP3 except those included in the
last sorting pass, output the nth most significant bit of C 1 (i , j ) and C 3 (i , j ) .
If n  n _ max 2 , for any entry (i,j) in LSP1 , LSP2 and LSP3 except those included in the last
sorting pass, output the nth most significant bit of C 1 (i , j ) , C 2 (i , j ) and C 3 (i , j ) .
4. Quantization updating
Decrement n by 1 and go to step 2 until acquire the desired compression ratio.
The code stream of embedded 3D-SPIHT encoding can be illustrated as figure.2. The
decoding process and encoding process are just the same. So, it’s easy to present the
decoding process according to the 3D-SPIHT encoding process. The code stream of 3D-
SPIHT is more compact and efficient for the three mixed coefficient planes are interleaved
during the encoding process.
Fig. 2. Code stream illustration for embedded 3D-SPIHT.

Additionally, it is necessary to be mentioned that there have been two kinds of 3D-SPIHT
algorithms before. One is proposed for video compression by extending the conventional
SPIHT algorithm to 3D case directly and encoding the 3D-DWT wavelet coefficients of video
data, so the SOF is defined as 8 splitting [7]. The other is proposed for compression of
multispectral images, which make some amendments of the conventional SPIHT by adding
one spectral child to every baseband coefficient, so its SOF is still 4 splitting [8]. The
proposed 3D-SPIHT embedded coding in this book is very different from the two existing
3D-SPIHT algorithms, which encodes 1 or 2 or 3 coefficient planes of the 3 mixed coefficient
plane sequentially by adopting 3 independent thresholds.
4. Unsymmetrical SPIHT codec

It is easy to find that the set partitioning is keeping on 4 partitioning for conventional SPIHT
codec, that is to say the same set partitioning in vertical and horizontal direction. This means
that the wavelet decomposition only can be implemented under the constraint of the same
decomposition level in vertical and horizontal direction, which will affect the redundancy
removing of image data efficiently and the final compression performance. The
unsymmetrical SPIHT codec, adopting set 2-partitioning or set 4-partitioning according the
requirements, doesn’t require the same decomposition level in vertical and horizontal
direction. So, DWT can be implemented with each highest feasible decomposition level in
vertical and horizontal direction respectively and then the spatial redundancy can be
removed efficiently.
Because the flow chart of unsymmetrical SPIHT codec is very near to conventional SPIHT
codec, only the difference is given.
Assume the image dimensions are W * H and H  h * 2 HLevel , W  w * 2 WLevel , here
HLevel , WLevel are the highest feasible decomposition level in vertical and horizontal
direction respectively. Usually, there exists HLevel  WLevel when the image size are
unsymmetrical. For the conventional SPIHT, the highest feasible decomposition level only
can be the lower one of HLevel and WLevel , so, Level  min( WLevel , HLevel ) . But for
unsymmetrical SPIHT codec, the highest wavelet decomposition levels in vertical and
horizontal direction are HLevel and WLevel respectively.
All set partitions for conventional SPIHT are set 4 partitioning just as the formula 2, but the
set partitions for unsymmetrical SPIHT are set 2-partitioning or set 4-partitioning, just as the
following formula:
 H 
 c(i , 2 j ), c(i , 2 j  1) if H 0  i  
2
 
 W 
O(i , j )   c(2 i , j ), c(2i  1, j ) if iW0  j   (18)
 2 
c(2 i , 2 j ), c(2 i  1, 2 j ), c(2 i , 2 j  1), c(2i  1, 2 j  1) otherwise 
 
 
H W
Here, H 0  HLevel  WLevel  1
, W0  . WLevel  HLevel  1
2 2
When HLevel  WLevel , the unsymmetrical SPIHT codec will completely degenerate into
conventional SPIHT codec. When HLevel  1 or WLevel  1 , only set 2-partitioning are
implemented in one direction, the coefficients can be encoded line by line independently.
The wavelet transform in another direction is only be used as compacting the image energy.
Fig 3 and Fig 4 are the illustrations of conventional SPIHT encoding and unsymmetrical
SPIHT encoding at HLevel  3 or WLevel  5 .
Fig. 3. The illustration of set partitioning using conventional SPIHT encoding for
unsymmetrical image size
Fig. 4. The illustration of set partitioning with unsymmetrical SPIHT encoding for
unsymmetrical image size
5. 1D SPIHT codec
In section 4, unsymmetrical SPIHT codec is detailed described, which also need 2D image
data or image block. But, in real time image transmission or scan display system, the image
data are usually transmitted or displayed line by line. In order to use conventional SPIHT or
unsymmetrical SPIHT, it needs many line buffers to store the previous image data. In
hardware, it is a high burden for the costly RAM. So, 1 line image data compression method
will have the precedence over other block based compression methods, such as the 1D DWT
followed 1D SPIHT codec which will be addressed in the following.
After 1D DWT, the wavelet coefficient also has the natural pyramid characteristic: every
pixel of the high frequency subband has its 2 corresponding pixels in its adjacent level high
frequency subbands in position, which means that only set 2-partitioning can adopted. The
illustration is given in fig3.
Fig. 5. The illustration of set partitioning with 1D SPIHT encoding for 1 line image.
The SOT and set partitioning can be written as formula 19 and 20.
T (i )  c(i )  D(i )

D(i )  O(i )  L(i ) (19)
L(i )  D( k ) ( k )  O(i )
 
O(i ) = {c(2i ), c (2 i  1)} (20)
From fig 5 and formula 19 and 20, we can see that 1D SPIHT use set 2 partitioning to encode
1D DWT coefficients, which only need 1 line buffer RAM but leave another dimensional
redundancy un-removed.
6. Conclusion
In this chapter, SPIHT and it’s derivatives or its modification methods, such as 3D-SPIHT,
3D-SPIHT Embedded method, Unsymmetrical SPIHT and 1D-SPIHT are described,
which can overcome the disadvantages of traditional SPIHT codec and meet the
specific requirements for the real applications. Fig.6 gives the derivative relationship of
traditional SPIHT and its several modified methods. We can see that SPIHT is the
foundation for traditional symmetrical image, but it can’t meet the requirements at some
specific applications, such as multi-channel image, strip image (unsymmetrical image),
even image line. But, its modified version can meet some specific requirements in real
applications.
* mean the corresponding dimension is limited compared other dimensions

Fig. 6. Family of SPIHT and its derivatives.
7. References
[1] Said A. and Pearlman W.A., A new fast and efficient image codec based on set
partitioning in hierarchal trees, IEEE Trans. on Circuits and Systems for Video
Technology, 1996, 6(3): 243-250.
[2] D. E. Dudgen, R. M. Mersereau, Multidimensional digital signal processing, New
Jersey：prentice-Hall, Inc.,1984.
[3] Zhu Yanqiu，Chen Hexin，Dai Yisong. Compression Coding of color image via 3-D
matrix transform [J] ACTA EI ECTRONICA SINICA，1997，25(7):16-21.
[4] Y. Shoham, A. Gersho, Efficient bit allocation for an arbitrary set of quantizers, IEEE
Transaction on Acoustics,Speech, and Signal Processing, 1988,36(9):1445-1453
[5] P. Prandoni, M. Vetterli, R/D optimal linear predication, IEEE Transaction on Speech
and Audio Processing, 2000,8(6):646-655.
[6] A. Aminlou, O. Fatemi, Very fast bit allocation algorithm, based on simplified R-D curve
modeling, IEEE ICECS 2003, 112-115.
[7] B. J. Kim, Z. Xiong, W. A. Pearlman,Low bit-rate scalable video coding with 3-D set
partitioning in hierarchical trees(3-D SPIHT), IEEE Transaction on Circuilt and
System for Video technology, 2000, 10(8): 1374-1387.
[8] P. L. Dragotti, G. Poggi and A. R. P. Ragozini, Compression of multispectral images by
three-dimensional SPIHT algorithm, IEEE Transaction on Geoscience and Remote
sensing, 2000, 38(1):416-428.
[9] W.-C. Zhang, Y.-F. Wang, G.-H. Hu,.Compression of multi-polarimetric SAR intensity
images based on 3D-matrix transform, IET- Image processing, 2008,2(4):194:202.
[10] Zhang Zhi-hui, Zhang Jun, Unsymmetrical SPIHT Codec and 1D SPIHT Codec,
International Conference on Electrical and Control Engineering (ICECE), 2010,
wuhan 2498:2501.
Part 2
Image Processing Applications

6
Multiresolution Approaches for

Edge Detection and Classification
Based on Discrete Wavelet Transform
Guillermo Palacios, J. Ramón Beltrán and Raquel Lacuesta
University of Zaragoza
Spain
1. Introduction
Detecting edges is a very well known subject in the image-processing field. Edge detection is
the process of the localization of significant discontinuities in the grey level image and the
identification of the physical phenomena that originated them. Those significant intensity
changes occur at different resolution or scales for a given image. As suggested by Rosenfeld
and Thurston (Rosenfeld & Thurston, 1976) and Marr (Marr, 1982), we can obtain a
description of the image changes at different scales combining the information given by an
edge detector applied at different resolutions. This is the aim of the work presented in this
paper.
The first aspect to be covered by multiresolution analysis is the filter chosen to accomplish
the low-pass filtering of the image at different scales. At a single resolution, low pass
filtering is imposed because differentiation is an ill-posed problem (Torre & Poggio, 1984).
The needed regularization process is implemented by means of a low pass filter. Marr
(Marr, 1982) proposed the Gaussian filter because its optimal behaviour in terms of the
smoothing and the localization in both the spatial and frequency domains. This filter has
been commonly used in edge detectors. In a multiresolution approach the first or second
directional derivatives of the low pass filtered image with Gaussians of different widths are
used to detect edges. In the Bergholm edge focusing method (Bergholm, 1987) various edge
maps extracted at different scales are integrated allowing distinguishing shadows contours
from perfect ones using Canny’s operator (Canny, 1986) with different widths. Another
possibility is to describe the image in terms of the scale space as proposed by Witkin
(Witkin, 1983) and to detect edges in terms of the zero crossing of the Laplacian operator
with different widths (Park et al., 1995) (Eklundth et al., 1982).
Other multiresolution methods have been proposed. Mallat and Zhong (Mallat & Zhong,
1992) related multiscale edge detection with the discrete wavelet transform (DWT). They
proposed a wavelet to perform edge detection and they showed that the evolution of
wavelet local maxima across scales characterizes the shape of irregular structures. In our
work we will use the wavelet-based algorithm proposed by Mallat and Zhong and we will
show the condition that must be satisfied by the Gaussian filter to be comparable with the
Mallat and Zhong’s wavelet. Our aim is to detect and classify different edge types. Various
edge profiles have been proposed. Rosenfeld (Rosenfeld & Kak, 1976) proposed the step,
ramp, pulse and stair as the basic edge types. He stated that these profiles are suited for a
first intuitive classification of the edges found in the contours of real images.
William and Shah (Williams & Shah, 1990, 1993) have studied these edge types using the
first directional derivative of the Gaussian. Some other profiles like the blurred step have
also been proposed and analyzed (Ziou & Tabbone, 1993). In a first step of this work we are
going to analyze the evolution of the modulus of the wavelet coefficients at the edge
position in order to classify the edges into four different profiles: step, ramp, stair and pulse
(Beltrán et al., 1994). Then we will propose a general schema to detect, analyze and classify
different edge types.
Due to the high pass filtering operation involved in the edge detection algorithms, the noise is
always amplified when detecting edges. Thresholding has been the most common way to
eliminate the irrelevant or noise detected edges (Canny, 1986; Marr, 1982). To validate the
proposed classification schema we will present the noise (Gaussian noise) as a new edge class
(Beltrán et al., 1998). Then, we have analyzed this contour type and we have modified the
proposed classification algorithm to include the noise as a new edge type. With the noise edges
labelled we can easily implement a noise-filtering algorithm. Finally, we have developed a
new classification algorithm including other edge profile models, such as the roof, ridge and
two non-antisymmetrical step profiles, like the ones proposed by Paillou (Paillou, 1994).
This chapter is divided as follows. Section 2 presents a survey of the wavelet formalism
introduced by Mallat and Zhong. Then, we will show the geometrical contour types: step,
ramp, pulse and stair. Section 3 presents the edge detection algorithm including the wavelet
algorithm implementation. Section 4 presents the classification algorithm implementation
details as well as the classification results obtained processing a 256x256 grey level synthetic
image with four objects: a circle, a square, a triangle and a narrow line, each one having a
different contour type. Section 5 deals with the characterization of the noise as a new type of
contour. Section 6 copes with the new contour types just introduced, the modified
classification algorithm together with the obtained results. In section 7 we present the main
conclusions and the future work.
2. Theoretical basis
Mallat and Zhong (Mallat & Zhong, 1992) introduced the relationship between the wavelet
transform and a multiresolution edge detection algorithm. We are going to briefly
summarize those results. Let be f(x,y) ∈ L2(R2) an image and ψ(x,y) a wavelet. The
bidimensional wavelet transform of f(x,y) is defined as:
+∞ +∞ 1 x−u y−v
Ws f(u, v) = ∫ ∫ f(x, y) ψ  , (1)
−∞ −∞ s  s s 
If we define the dilation by a factor s as:
1 x y
ψ s ( x, y ) = ψ  ,  (2)
s s s
ɶ ( x, y ) = ψ
and ψ ɶ ( -x, -y ) , we can rewrite (1) as a convolution:
Ws f ( u, v ) = f ∗ ψ
ɶ ( u, v )
s (3)
Multiresolution Approaches for Edge Detection
and Classification Based on Discrete Wavelet Transform 83
So, as expressed in (3), the wavelet transform can be seen as the filtering of f(x,y) by the filter
ψɶ (x,y), that is a variable width bandpass filter (Mallat, 1989). It is possible to define N
s
directional wavelets ψi(x,y) (1 ≤ i ≤ N), satisfying energy conservation properties. In such a
case the directional wavelet transform of f(x,y) is defined as:
Wsi f ( u, v ) = f ∗ ψ
ɶ i ( u, v )
s (4)
Equation (4) represents the filtering of f(x,y) by the bidimensional, directional and bandpass
ɶ i (x,y).
filter ψ s
We can define the Discrete Wavelet Transform (DWT) by selecting the scales inside a dyadic
grid; that is to say, the scale could be expressed as S=2j with j ∈ Z. Therefore, for discrete
signals, we can understand the 2-D wavelet transform as the result of filtering the 2-D signal
(the original image) with a bandpass directional FIR filter.
Mallat and Zhong (Mallat & Zhong, 1992) designed a function specially suited for edge
detection purposes, which is a wavelet. This function is not a orthogonal wavelet, so the
only condition to be satisfied by ψ(x,y) is:
+∞ +∞
∫ ∫ ψ ( u, v ) dudv = 0 (5)
−∞ −∞
We can define two functions ψ1(x,y) and ψ2(x,y) as:
∂ θ(x, y) ∂ θ(x, y)
ψ 1 (x, y) = , ψ 2 (x, y) = (6)
∂x ∂y
Where θ(x,y) is a smoothing function, that is to say, the integral over x and y is one, and it
converges to zero at infinity. With these conditions is easy to show that both ψ1(x,y) and
ψ2(x,y) satisfy (5), so they are wavelets. Following equation (3) we can say that the wavelet
transform of f(x,y) is:
Ws1f(x, y) = f ∗ ψɶ 1s (x, y)
(7)
Ws2 f(x, y) = f ∗ ψ
ɶ s2 (x, y)
From (6) and (7) it can be shown that the wavelet transform is the gradient of the image
smoothed by a factor or scale s. It can be expressed as:
 ∂ 
 W1f(x, y)   ∂ x (f ∗ θs ) 
 s  = s  = s∇(f ∗ θs ) (8)
 W2 f(x, y)   ∂
 s   (f ∗ θs ) 
∂y 
We can define the modulus M and the phase Φ of the gradient at scale s as:

( )
2 2
M ∇(f ∗ θs ) = Ws1f(x, y) + Ws2 f(x, y) (9)
 W 2 f(x, y) 
( )
Φ ∇(f ∗ θ s ) = atan  s1
 W f(x, y) 
(10)
 s 
A point (x0,y0) will be an edge point of the smoothed image f *θs (x,y) if in this point there is
a relative maximum of M in the direction addressed by Φ. The above statement is the
classical definition of edge detection proposed by Canny [5]. In the wavelet case we have a
discrete set of scales s = 2 j (0 ≤ j ≤ J), and we can calculate the edges at each scale and not at
an only one scale as described in Canny’s work. So, we can conclude that we can implement
a multiresolution edge detection algorithm by means of the 2-D wavelet transform. The
smoothing function could be a new defined function, as in Mallat and Zhong’s work, or a
Gaussian function, as in Canny’s work.
Mallat and Zhong (Mallat & Zhong, 1992) defined θ(x,y) as a separable bidimensional cubic
spline, so ψ(x,y) is a separable bidimensional quadratic spline. Their 1-D expressions are:
8 3 2 8
 3 x + 8x + 8x + 3 if − 1 ≤ x ≤ − 1 2

 −8x 3 − 8x 2 + 4 if − 1 2 ≤ x ≤ 0
 3

θ̂(x) =  3 2 4 (11)
8x − 8x + 3 if 0≤x≤1 2

 − 8 x 3 + 8x 2 − 8x + 8 if 1 2≤x≤1
 3 3
0 otherwise

8x 2 + 16x + 8 if − 1 ≤ x ≤ −1 2

 −24x 2 − 16x if − 1 2 ≤ x ≤ 0

ψ̂(x) =  24x 2 − 16x if 0≤x≤1 2
(12)

 −8x 2 + 16x − 8 if 1 2≤x≤1

0 otherwise
ˆ ˆ
In the 2-D case we can define θ(x, y) = θ(x)θ(y) and ψ(x, y) = ψ(x)ψ(y)
ˆ ˆ . We can compare the
wavelet and the Gaussian functions behaviour. In figure 1 we can observe that both
functions present a similar aspect. We have obtained a relationship between the Gaussian
width σ and the scale s of the wavelet analysis. The Gaussian normalization constant is:
1 4
N= , and the smoothing function value at the origin is θ(0) = .
2̟ ⋅ σ 3
Equating both quantities we obtain σ= 0.3. So, to obtain a similar analysis with the wavelet
and the Gaussian function, the Gaussian width and the scale should be related by the
aforementioned expression.
Now we will to present the edge profiles considered initially. We have considered that each
profile contour is a function that represents a change in the grey level with respect to the
image dimensions. In this work we will name Ui the contour height, ω the contour width, x0
the contour position (at the middle of the contour) and s the scale.
Fig. 1. Left: Solid line: Smoothing function. Dotted line: Gaussian with σ = 0.3. Right: Solid
line: Wavelet. Dotted line: First derivative of Gaussian with σ = 0.3.
2.1 Step
A step profile is shown in figure 2. We have named x0 the step location and U0 the step
height. Let be u(x) the step function.
U 0 if x ≥ x0
u(x) =  (13)
0 if x < x0
Fig. 2. Step profile. Horizontal axis represents pixels. Vertical axis represents grey level.
2.2 Ramp
The ramp profile is drawn in figure 3. The point x0 is the middle point of the ramp; U0 is the
height and ω is the ramp width. The ramp slope is m = U0/ω. Let be r(x) the ramp function.
 ω
0 if x < x0 −
2

 ω ω
r(x) = mx + r0 if x0 − ≤ x ≤ x0 + (14)
 2 2
 ω
U 0 if x > x 0 +
 2
Fig. 3. Ramp profile. Horizontal axis represents pixels. Vertical axis represents grey level.
2.3 Stair
A stair profile with two steps, named s(x), is shown in figure 4. We have named the point x0
the middle of the stair, ω is the stair width and U0 and U1 are, respectively, the steps height.
 ω
0 if x < x0 − 2

 ω ω
s(x) = U 1 if x0 − ≤ x ≤ x 0 + (15)
 2 2
 ω
U 0 if x > x0 + 2

Fig. 4. Stair profile. Horizontal axis represents pixels. Vertical axis represents grey level.
2.4 Pulse
A pulse profile, named p(x), is shown in figure 5. The point x0 is the middle of the pulse, ω is
the width. U0 and U1 are, respectively, the maximum and minimum height.
 ω
0 if x < x0 − 2

 ω ω
p(x) = U 0 if x0 − ≤ x ≤ x0 + (16)
 2 2
 ω
U 1 if x > x 0 + 2

Fig. 5. Pulse profile. Horizontal axis represents pixels. Vertical axis represents grey level.
3. Edge detection algorithm

The processing algorithm we have used is the edge detection algorithm proposed in Mallat
and Zhong’s work (Mallat & Zhong, 1992). For analysis purposes we have used six of the
eight possible scales for a 256x256 grey level image. The scale s forms a dyadic sequence (s =
2j, j = 1..6). The algorithm is summarized in figure 6, and table 1 shows the values of the
normalization coefficients λj. The algorithm outputs are twelve 256x256 grey level images
for each original one, six corresponding to the modulus of the derivative and six to the
phase.
As we have pointed out in section 2, it is possible to perform a multiresolution analysis
changing the smoothing function. The only difference between a typical Gaussian
algorithm and the wavelet one is the filtering stage. In this case, instead of using the filter
proposed by Mallat and Zhong, a Gaussian filter with a different size for each scale could
be used. It leads to a different way to obtain the derivative of the image at different scales.
A more detailed discussion about the Gaussian-based processing algorithm is given in
Beltrán (Beltrán et al., 1998). Processing the image with the wavelet filter is faster, in
computational cost terms, because the non-zero coefficients are constant independent on
scale.
In order to obtain the contour image, a top-down searching algorithm has been
implemented. For accepting a maximum at one scale to be an edge, it has been imposed that
the maxima have to be propagated to the lowest scale with no change in the gradient
direction between scales. When a maximum is found at one scale s we look for extrema in
the lower scale within an interval of 2s centered in the extrema position at the scale s, in the
direction given by the gradient. This interval is greater than the theoretical one found by
Beltrán (Beltrán, 1994), in order to cover the maxima displacement.
The maximum appearing in the lower scale has to have the same direction that the upper
one. The best edge position is given at the lowest scale. The first scale to be analyzed
depends strongly on the image. Empirically we have noticed that this scale has to be no
higher than either the 5th or the 6th. Otherwise, we have a strong blurring in the image
that gives us information of global objects rather than finer patterns, like edges. The
procedure has been iterated until first scale. A stop could be done at an upper scale,
depending on the details we are looking for. If we were looking for finer details we
should reach the first scale. A global threshold has been included in order to discard
irrelevant edges.
j λj
1 1.5
2 1.125
3 1.031
4 1.007
5 1.001
6 1
Table 1. Wavelet filtering normalization coefficients.

At the output of this block, we obtain a black and white image with the edge positions, and,
for each detected edge, the values of the modulus of the gradient in each scale. The final step
is the contour classification, which is made by analyzing the evolution of the modulus of the
gradient across scales.
Fig. 6. Wavelet Algorithm.
4. Classification schema
To obtain real evolution patterns in an image we have analyzed the evolution across scales
of the value of the wavelet transform modulus for each contour type at the point at which
the edge is located in our test image (see figure 7).
Fig. 7. Test image. Each object has a different contour profile. Circle: step. Square: ramp.
Triangle: stair. Straight line: pulse.
In figure 8 we present for each contour class both the median value (normalized to the
highest value for each pixel) and the deviation for each scale.
It can be seen a decreasing behavior in the evolution for each contour class in the upper
scales due to the interaction between the opposite contours in the image. This decreasing
pattern is not present in the 1-D case. From those results we can say that a step profile,
figure 8(a), is characterized by an almost constant evolution across scales. The evolution
pattern for a ramp, presented in figure 8(b), is a continuous growing from low to high scales.
A constant value followed by an increasing in the normalized modulus of the gradient in the
(a) (b)
(c) (d)
Fig. 8. (a) Step evolution. (b) Ramp evolution. (c) Stair evolution. (d) Pulse evolution. (See
text for details)
edge position characterizes a stair contour, figure 8(c). As shown in figure 8(d) the pulse
profile presents a sharp decreasing in the upper scales due to the interaction between the
positive and negative slopes.
Figure 9 shows a block diagram of the edge detection and classification algorithm. An
immediate conclusion is that it is possible to implement an algorithm to detect and classify the
above-characterized four profiles using the wavelet transform coefficients. Then, we are able to
distinguish one contour type from another one only by looking at the coefficients evolution at
the appropriate contour point and with no pre-processing in the coefficient values.
The classification engine is based in a second-order polynomial-fitting algorithm. We have
chosen a second order taking into account the easiness of implementation of properties such
as derivability, continuity, concavity, and convexity. Analyzing the concavity and convexity,
zero crossing and minimum abscissa and ordinate of the fitted polynomial we are able to
distinguish between the four profiles presented. Typical values of the coefficients are
presented in table 2 and the corresponding polynomial in figure 10. The second order
polynomial is of the form: f(x)=a0+a1x+a2x2. We have used a standard polynomial-fitting
algorithm with the 6 discrete values obtained for each contour at the edge location.
Fig. 9. Edge detection and classification algorithm.
a0 a1 a2
Step 0.8295 0.1026 -0.0187
Ramp -0.0249 0.3678 -0.0362
Stair 0.3562 0.0894 0.0005
Pulse 1.0087 -0.0732 -0.0131
Table 2. Typical coefficients values of the fitted polynomial.

The decision algorithm is presented in figure 11. We first analyze the convexity of the
polynomial (a2>0), in the affirmative case we have a stair edge. In case of having a concave
polynomial we analyze its slope. In case of a positive slope, we will have a ramp edge. In the
opposite case, we will have to compute the zero crossing value. If this value is greater than,
approximately, 6 the edge will be a step, otherwise it will be a pulse.
(a) (b)
(c) (d)
Fig. 10. Dotted lines: Original values. Solid lines: polynomial. (a) Polynomial fitting for the
step profile. (b) Polynomial fitting for the ramp profile. (c) Polynomial fitting for the stair
profile. (d) Polynomial fitting for the pulse profile.
Fig. 11. Decision algorithm for edge classification.

As some preliminary results we can see the correct classification made in figure 12 for our
test image. It is important to note that no post-processing (edge tracking, non-maxima
suppression, etc.) has been made in the obtained contour image.
Fig. 12. From left to right and top to bottom: step, ramp, pulse, and stair classified profiles.
5. Noise characterization and polynomial classification

As a first practical result of our processing schema we have analyzed our algorithm
behavior with respect to noise. In order to obtain a proper characterization of the noise we
have to compare the evolution in the coefficients shown by the noise with that presented by
other kind of profiles. Firstly, we have processed a simple image with only Gaussian noise.
Secondly, we have analyzed the image with the edge detection and classification algorithm
as described in section 4. The mean values of the modulus of the gradient evolution at the
edge points are shown in figure 13a. This pattern characterizes isolated Gaussian noise.
Fig. 13. (a) Mean values of Gaussian noise evolution. (b) Noise dependence on distance.
But noise is rarely presented isolated in a real image, and its evolution depends strongly on
the distance between the noise and the contour of the objects present in the image. To obtain
the noise evolution we have processed the image of a simple square with a step profile
corrupted with Gaussian noise. The evolution in the values of the noise, in terms of the
distance between the noise and the square, are presented in figure 13b. These patterns serve
us to classify and differentiate the noise from other contour types. We can see that the
evolution pattern for a noise contour located very close to a real contour is quite similar to
the stair evolution. This is the expected behavior because the stair has been defined as closed
steps. We can include this contour type in our classification engine.
The final decision algorithm is presented in figure 14. We have to distinguish between the
stair type and the noise one. If we have a convex polynomial (a2 > 0) we have a stair or noise
edge. They are differentiated by means of the minimum ordinate value. If the minimum
ordinate is close to 0 it is classified as a noise edge.
YES YES
Convex? minimun==0? NOISE
NO
NO STAIR
YES
Increasing? RAMP
NO
YES
Xm= 6? PULSE
NO
STEP
Fig. 14. Decision algorithm for edge classification.

To analyze the robustness of the algorithm we have corrupted our test image with 20 dB of
Gaussian noise. The classification results are shown in figures 15 and 16, respectively. No
thresholding has been applied to the contour image.
(a) (b) (c) (d)

Fig. 15. Test image corrupted with 20 dB Gaussian noise. (a) Detected edges without
thresholding. (b) Non noise edges. (c) Noise edges. (d) Detected edges with Canny
algorithm.
(a) (b) (c) (d)

Fig. 16. Edge profiles detected for the test image corrupted with 20 dB Gaussian noise. (a)
Step edges. (b) Ramp edges. (c) Stair edges. (d) Pulse edges.
It can be seen that the noise is perfectly classified (figure 15c) and can be eliminated by
simply removing this edge type (figure 15b). Figure 15d shows the output of the Canny edge
detector. In this case Canny operator is more sensitive to noise than our algorithm. Figure 16
shows the different edge profiles detected.
6. New contour types

The geometrical characterization of contour types gives us very promising results. To
confirm the ability of our algorithm to distinguish different contour types we present here
four new types of edge profiles that represent a more general gray-level transition than a
simple step. We have included the roof profile, the ridge profile and two non-
antisymmetrical step profile models that have been already considered previously by
Paillou (Paillou, 1994).
6.1 Roof
A roof profile, named R(x), is shown in figure 17. We have named the point x0 the middle of
the roof, ω is the roof width and U0 the roof height.
 ω
0 if x < x0 −
2

mx + k ω
if x0 − ≤ x ≤ x0
 2

r(x) = U0 if x = x0 (17)
 ω
−mx + k if x0 ≤ x ≤ x0 +
 2
 ω
0 if x > x0 +
 2
Fig. 17. Roof profile. Horizontal axis represents pixels. Vertical axis represents grey level.
6.2 Ridge
A ridge profile, named R(x), is shown in figure 18. The point x0 is the middle of the ridge, ω1
is the width of the first ramp, ω2 is the width of the plain part, while ω3 is the width of the
second ramp. U0 is the ridge height.
 ω
0 if x < x0 − ω 1 − 2
2

mx + k ω2 ω
if x0 − ω 1 − ≤ x ≤ x0 − 2
 2 2

 ω2 ω2
R(x) = U 0 if x 0 − ≤ x ≤ x0 + (18)
 2 2
 ω2 ω2
−mx + k if x0 + 2 ≤ x ≤ x0 + 2 + ω 3

0 ω
if x > x0 + 2 + ω 3
 2
Fig. 18. Ridge profile. Horizontal axis represents pixels. Vertical axis represents grey level.
6.3 First non-antisymmetrical step

The first non-antisymmetrical step profile, named nu(x), is shown in figure 19. The point x0
is the step location, ω1 is the width of the first ramp and ω1 is the width of the second ramp.
U1 and U3 are their respective heights. U2 is the step height at position x0. The first non-
antisymmetrical step function can be written as:
0 if x < x0 − ω1
mx + k if x0 − ω1 ≤ x ≤ x 0 + ω1 − ω 2

nu(x) =  (19)
mx + k if x0 + ω 1 − ω 2 ≤ x ≤ x0 + ω 1 + ω 2
U 3 if x > x0 + ω1 + ω 2
Fig. 19. First non-antisymmetrical step profile. Horizontal axis represents pixels. Vertical
axis represents grey level.
6.4 Second non-antisymmetrical step

The second non-antisymmetrical step profile, named nu2(x), is shown in figure 20. The point
x0 is the middle of the main ramp, ω1 is the width of the first ramp, ω2 is the width of central
ramp and ω3 is the width of the third ramp, respectively. U1, U2 and U3 are their respective
heights (see figure). Let be nu2(x) the second non-antisymmetrical step function.

0 if x < x0 − ω1

mx + k ω2
if x0 − ω1 ≤ x ≤ x 0 + ω1 −
 2
 ω2 ω
nu2(x) = cx + k if x0 + ω1 − ≤ x ≤ x0 + ω1 + 2 (20)
 2 2
 ω2 ω
mx + k if x0 + ω1 +
2
− ω 3 ≤ x ≤ x0 + ω1 + 2 + ω 3
2

U ω2
if x > x0 + ω1 + + ω3
 3 2
Fig. 20. Second non-antisymmetrical step profile. Horizontal axis represents pixels. Vertical
axis represents grey level.
6.5 Modified classification algorithm

Finally, we will include the four contour types in our classification algorithm. We will also
work on the basis of the polynomial-fitting algorithm, by following characteristics such as
convexity, concavity, zero crossings and minimum abscissa and ordinate as we did
beforehand. Furthermore, we have also used a standard polynomial-fitting algorithm with
the 6 discrete mean values obtained for each contour at the edge location. However, in this
case, a third order polynomial fitting will serve us to classify some of the new profile models
introduced above. In particular, the third order coefficient pattern matching will serve us to
discriminate among the ridge, roof and ramp profiles, respectively. Maximum value at
scales 5 or 6 and comparison between values at two different scales are also used. Figure 21
shows the complete decision algorithm for edge classification.
Fig. 21. Modified decision algorithm for edge classification.

The polynomial that characterizes the roof profile (as well as the ridge profile) is concave
and with a positive slope in the first three scales, in the same way as the ramp profile does.
In order to discriminate among these profile models just mentioned above, it is necessary to
calculate the coefficient pattern for several fitting polynomials, third, fourth and fifth order).
We find a remarkable difference among the ridge profile on one side, and the ramp and roof
profiles on the other side. The coefficient pattern (-,+,-,+) ≡ (a0,a1,a2,a3), characterizes the
ridge profile. In order to distinguish between the ramp and roof profiles, we have realized
that the mean value at scale 6 is greater that the one at scale 2 in a ramp profile, whereas
both are similar in a roof profile. This is the condition to put aside both profiles.
As far as antisymmetrical step profiles are concerned, we must say that they both have
some properties like the stair profile (convexity and minimum greater than zero). It can be
shown that if the mean value at scale 6 is smaller than that at scale 1, we will obtain a stair
profile, otherwise we will have the first non-antisymmetrical step profile. For the sake of
algorithm efficiency, we have included a second condition: a stair profile has the
maximum value at scales 5 or 6. In order to obtain the second non-antisymmetrical step
profile, we compute whether the mean value at scale 2 is, at least 1.5 times the mean value
at scale 1.
In order to show the results of the modified algorithm, we have created a new test image,
including the new edge profiles. Figure 22 shows the new test image, together with the
different detected edge profiles. Results with a real image are shown in figures 23 and 24.
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)
Fig. 22. (a) Test image. (b) Ramp edges. (c) Step edges. (d) Pulse edges. (e) Stair edges. (f)
Ridge edges. (g) Roof edges. (h) First non-antisymmetrical step edges. (i) Second non-
antisymmetrical step edges.
(a) (b) (c)

Fig. 23. (a) Original image. (b) Contour image. (c) Contour image after noise filtering.
(a) (b) (c)
(d) (e) (f)
(g) (h) (i)

Fig. 24. (a) Noise edges. (b) Ramp edges. (c) Step edges. (d) Pulse edges. (e) Stair edges. (f)
Ridge edges. (g) Roof edges. (h) First non-antisymmetrical step edges. (i) Second non-
antisymmetrical step edges.
7. Conclusions
In this work, we have developed a new algorithm for edge detection and classification
purposes using the coefficients given by the DWT. We have shown that Mallat’s wavelet is a
very suitable tool to perform both edge detection and contour analysis. We have presented
the results obtained with a 256x256 synthetic image with several objects, each one with a
different contour profile, obtaining a very good segmentation. At this point, we are not only
able to see the evolution across scales of the edges proposed by Rosenfeld (Rosenfeld &
Thurston, 1976) like in Williams and Shah’s work (Williams & Shah, 1990, 1993), but we are
also able to classify them.
A new edge class has been introduced: the noise. This edge type presents a particular
evolution across scales. This has allowed us to implement a simply noise filtering algorithm
based on edge classification.
The classification algorithm we have developed is based on second order polynomial fitting
of the modulus of the wavelet transform coefficients. The mathematical behaviour of the
polynomial is a robust indicator of the edge class. This kind of classification is good enough
to obtain the five different profiles analyzed: step, ramp, stair, pulse, and noise. The
robustness of the proposed classification schema has been tested including other profiles
appeared in the literature: roof, ridge and two kinds of non-antisymmetrical step models. A
third order polynomial-fitting algorithm is needed to obtain a proper classification. This
algorithm can be viewed as a new framework to classify different contour types.
Some preliminary results, like the behaviour of ramp edges, are promising to obtain a
classification of the contours appearing in real images (shadows, changes in illumination,
corners and so on). A future work to perform, which has not been covered in this paper, is
the study of the evolution across scales of these real edges. If there were some special
evolution pattern for these real edge types it would be very important information for the
next stages of an image understanding algorithm.
An edge-closing algorithm based on the edge type is under developing in this moment. The
extra information provided by the classification stage gives very good indicators to close
edges and extract objects in an image. The processing results presented in this work have
been obtained using Matlab®.
8. References
Beltrán, J. R., J. García-Lucía, J. & Navarro, J. (1994). Edge detection and classification using
Mallat’s wavelet. Proceedings of the ICIP-94, pp. 293-296
Beltrán, J. R., Beltrán, F & Estopañan A. (1998). Multiresolution edge classification: Noise
Characterization, Proceedings of the IEEE-SMC’98, pp. 4476-448
Bergholm, F. (1987). Edge Focusing. IEEE Trans. on Patt. Anal. and Machine Intell, Vol. 9, No.
6, pp. 726-471
Canny, J. F. (1986). A computational approach to edge detection. IEEE Trans. on Patt. Anal.
And Machine Intell. Vol. 8, pp. 679-698
Eklundth, J. O., Elfving, T. & Nyberg, S. (1982). Edge detection using the Marr-Hildreth
operator whith different sizes. Procedings of the. 6th. Int. Conf. on Pattern Recognition
(ICPR), Munich, Germany, pp. 1109-1112
Mallat, S. (1989). A theory for Multiresolution Signal Decomposition: The Wavelet
Representation. Trans. on Patt. Anal. and Machine Intell, Vol. 11, No. 7, pp. 674-693
Mallat, S. & Zhong, S. (1992). Characterization of Signals from Multiscale Edges. IEEE Trans.
on Patt. Anal. and Machine Intell, Vol. 14, No. 7, pp. 710-732
Marr, D. (1982). Vision. W. H. Freeman. San Francisco
Paillou, Ph. (1994). A non antisymmetrical edge profile detection. Pattern Recognition Letters
Vol. 15, pp. 595-605
Park, D. J., Nam, Kwon M. & Park, Rae-Hong. (1995). Multiresolution Edge Detection
Techniques, Pattern Recognition, Vol. 28, No. 2, pp. 211-229
Rosenfeld, A. & Thurston, M. (1971). Edge curve detection for visual scene analysis, IEEE
Transactions Computers, Vol. 20, pp. 562-569
Rosenfeld, A. & Kak, A. C. (1976). Digital Picture Processing. Academic Press
Torre, V. & Poggio, T. (1984). On edge detection. Technical Report 768, MIT
Williams, D. J. & Shah, M. (1990). Normalized Edge Detector. Proceedings of the 10th Int.
Conference On Pattern Recognition, 1 (16-21), pp. 942-946
Williams, D. J. & Shah, M. (1993). Edge Characterization Using Normalized Edge Detector.
Proceedings of CVGIP, Vol. 5, No. 4, pp. 311-318
Witkin, A. P. (1983). Scale-space filtering. Proceedings of the 4th Int. Joint Conference on Artificial
Intelligence (IJCAI), pp. 1019-1022
Ziou, D. & Tabbone, S. (1993). A Multiescale Edge Detector. Pattern Recognition, Vol 26, No.
9, pp. 1305-1314
7
Low Bit Rate Video

Compression Algorithm Using
3-D Discrete Wavelet Decomposition
Awad Kh. Al-Asmari
Electrical Engineering Dept., King Saud University, Riyadh
Saudi Arabia
1. Introduction
Video applications, like video teleconferencing, video telephones, and advanced television
(ATV), have given the field of compression and transmission of digital video signals a
significant importance. It is expected that the advances in video compression technology
will play a crucial role in the transmission and display of three-dimensional video signal.
A typical image, for example, of size 512x512 pixels with 8 bits per pixel (bpp) needs storage
capacity of about 2 Mbits. A video sequence, on the other hand, with the same frame size
with 30 frames per second and a channel transmission rate of 64 kilo bits per second (kbps)
would take about 17 minutes of transmission time. The required transmission time would
become unmanageable with the continuously increasing demand of image base application.
You can't put enough of it over a telephone line and you can't squeeze it into the broadcast
bandwidth of available channels1. Therefore, image and video compression algorithms
became a necessity to store or transmit these images.
Data compression is the science of representing information in a compact form by exploiting
the different kinds of statistical structures that may be present in the data2. This is to reduce
the number of bits per sample while keeping the distortion constant3. There is a great deal of
correlation between neighboring pixel values of an image. Therefore, removing such
redundant information and transmitting only the new information (the changes) enables us
to reconstruct the original image. For video signals, redundancy over time between
successive images can also be eliminated.
There are two types of compression: lossless and lossy. In the lossless compression the
original image can be retrieved without error, while for the lossy compression, the original
image can’t be retrieved without error; an image copy close to the original can be retrieved.
1 Realtime video compression poses challenge to designers and vendors alike, Computer Design, vol.
32, no. 7, pp. 67-70 (Child, July 1993).
2 Hybrid coding of images for progressive transmission over a digital cellular channel, CISST’99
International Conference on Imaging Science, Systems and Technology, Monte Carlo resort, Las Vegas,
Nevada, USA, PIN 128C (Al-Asmari et al.,June 28 – July 1, 1999).
3 Introduction to data compression, Morgan Kaufmann Publishers Inc., San Francisco, California
(Sayood, 1996).
Motion pictures expert group (MPEG) video standard is the most prevalent and widely used
for video compression3–6. Also, the MPEG is an application specific standard and different
versions of MPEG (Such as MPEG-1, MPEG-2, MPEG-4, and MPEG-7) are available for
different applications and bit rates. The basic algorithm for all these versions is the same and
is very similar to the other video compression standards.
The proposed algorithm is based on temporal filtering of image sequences with short
symmetric kernel filters (SSKFs)7–8, which are well known for their simplicity. In this paper,
we use four SSKFs filters each with 4-taps and with decimation factor of 4:1 instead of two
SSKFs filters each of 2-taps and with decimation factor of 2:1 used in classical 3D –
decomposition algorithms7–8. The temporal filtering removes the redundancy in temporal
domain. On the other hand, the pyramid coding (PC) is used for subband decomposition in
the spatial domain. The vector quantization (VQ) and the absolute moment block truncation
code (AMBTC) will be used to encode the spatial domain subbands.
2. 3-D decomposition of the video sequence

Practical video compression relies on techniques to reduce the amount of data required to
represent a video sequence without any appreciable loss of information that can affect the
visual quality of image. Our aim in this paper is to introduce an algorithm that compress the
video sequence with a reduced bit rate and highest fidelity while keeping the computational
complexity to a minimum to allow for easy hardware implementation. The complexity of
the proposed algorithm is less than the standard MPEG-1 and MPEG-2 algorithms.
Since our approach to the video compression problem is based on progressive transmission,
the multiresolution representation of the signals is required. The idea of multiresolution is
that a coarser approximation to the signal is refined step by step until the desired resolution
is obtained. An elaborate discussion on 3-dimensional decomposition of the video signals,
using pyramid method for spatial decomposition, is given in this section.
A group of video sequences (four frames each) is decomposed into subbands in the
temporal and spatial domains. The temporal frequencies are restricted to four subbands by
passing four consecutive frames through four band pass filters of 4-taps each and with
decimation factor of 4:1. By applying pyramid decomposition on these temporal subbands,
nine spatial subbands are produced. In the next sub-sections, the temporal and spatial
domain decomposition will be discussed in more details.
2.1 Temporal frequency decomposition

The 4-tap Haar basis functions (SSKFs filters) are used for temporal frequency
decomposition. These filters have no phase or amplitude distortion and belong to a class of
perfect reconstruction filters. The coefficients of SSKFs filters used in this research are given
4 Digital pictures representation, compression, and standard, Plenium Press (Netravali & Haskell, 1995).
5 Image and video compression standards: algorithms and architectures, Kluwer Academic Publishers
(Bhaskaran & Konstantinidides & Hewlett Packard Laboratories, 1996).
6 Digital compression of still images and video, Academic Press (Clarke, 1995).
7 Subband coding of video for packet networks, Optical Engineering, vol. 27, no. 7, pp. 574-586,
(Karlsson & Vetterli, July 1988).

8 Pyramid coding of video signals for progressive transmission, CISST’97 International Conference, pp-
392-398, (Dantwala et al., June 30 – July 3, 1997).

Low Bit Rate Video Compression Algorithm Using 3-D Discrete Wavelet Decomposition 103
in Table 1. The frequency responses of these filters are shown in Figure 1. H0(ejω) is the low
pass filter, H3(ejω) is the high pass filter, while H1(ejω) and H2(ejω) are the band pass filters.
The 3-dB bandwidth for these filters is approximately π/4.
Lowpass filter Bandpass filters Highpass filter

N h0(n) h1(n) h2(n) h3(n)
0 0.5 0.5 0.5 0.5
1 0.5 0.5 -0.5 -0.5
2 0.5 -0.5 -0.5 0.5
3 0.5 -0.5 0.5 -0.5
Table 1. SSKFs Coefficients for Haar filters.
Ho(ejω) H1(ejω)
1 0.8
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0 0
0 0.5 1 0 0.5 1
ω/pi ω /pi
H2(ejω) H3(ejω)
0.8 1
0.8
0.6
0.6
0.4
0.4
0.2
0.2
0 0
0 0.5 1 0 0.5 1
ω /pi ω /pi
Fig. 1. Frequency response of 4-tap Haar filters.

The video sequence with four consecutive frames is decomposed in the temporal domain by
passing these frames through the 4-tap Haar filters as shown in Figure 2.
The second four frames of Miss America sequence, that contains clear motion of the eyes,
are filtered using SSKFs filters. These filters decompose the sequence into four temporal
subbands. Namely, temporal low-low (TLL), temporal low-high (TLH), temporal high-low
h0 TLL
Haar Filters
h1 TLH
h2 THL
Four consecutive
frames
h3 THH
Fig. 2. Temporal filtering process using 4-tap Haar filters.

(THL) and temporal high-high (THH) subbands. Figure 3 shows the original and the filtered
frames. It can be seen from this figure that the subband TLL contains most of the image
sequence information, while the subband TLH contains most of the motion information in the
four frames. The subbands THL and THH contain the edge information, which are relatively
spares in the nature. Figure 3 (b) shows that the information of the temporal TLL subband is
highly correlated since it is only an average of the four frames, while the temporal (TLH, THL
and THH) subbands are of low correlated data. Compression is achieved since instead of
transmitting four highly correlated frames of the video at 8 bpp each, only one frame with
highly correlated data (TLL) at high bit rate will be transmitted. The other frames (TLH, THL
and THH) with low correlated information will be transmitted with very low bit rate.
The average entropy is calculated for the original and the decomposed four frames. Table 2
shows the average entropy for the original and the temporal filtered frames of Miss America
sequence. From this table, it is clear that the multiresolution decomposition process results
in lower entropy than that of the original images. From Shannon’s theory, the entropy for a
signal can be minimized by decomposing this signal into sub-signals using orthogonal
bases. Therefore, the calculated entropy is act as a lower measure for the information to be
encoded9. 4
Original frames Entropy Temporal frames Entropy

Frame 4 5.66 TLL 5.78
Frame 5 5.68 TLH 2.39
Frame 6 5.66 THL 2.09
Frame 7 5.65 THH 2.11
Average 5.66 Average 3.1
Table 2. The average entropy for the four frames before and after the decomposition.
9 A mathematical theory of communication, Bell System Tech. J.. Vol. 27, pp. 379-423, and pp. 623-656
(Shannon, July. 1948a, Oct. 1948b).
Frame 4 Frame 5
Frame 6 Frame 7
(a)
TLL TLH
THL (b) THH
Fig. 3. (a) Original and (b) Temporal filtered frames of the second four frames of Miss
America sequence.
2.2 Spatial domain decomposition

Spatial decomposition techniques aim at achieving data compression by discarding the
redundant or visually non-perceptual information in an image. Subband coding (SBC) is one
such scheme where the image spectrum is divided into a set of uncorrelated sub-spectra,
and each of the sub-spectrums is treated individually depending on the amount of
information it carries. The low-frequency bands with more energy content are given a
higher priority as compared to the high frequency bands with low energy contents10. 567
The four frames can be independently encoded using any image compression method
available for still images. Since we are interested in multiresolution decomposition, pyramid
decomposition method has been adopted for spatial domain decomposition. Pyramid
coding (PC) has been found to be more robust to channel errors as compared to SBC8. This is
so because neither data loss nor channel corruption in the error images has a serious effect
on the quality of the reconstructed image. Moreover, lower bit rates are possible for PC as
most of the energy of the pyramid structure lies in the lowest resolution baseband, which
can be encoded at high bit rates to preserve maximum information without increasing the
overall bit rate of the compressed image. Another advantage of the PC scheme is the ease of
filter design as compared to SBC 10.
Subband filters are designed under the constraints of perfect reconstruction and narrow
transition band. Both of these constraints are relaxed in PC as it makes use of bandpass
images for reconstruction 10. Because of these advantages, this research uses PC for spatial
decomposition of images, instead of the SBC.
3. Pyramid coding
Burt and Adelson have suggested a method of pyramid coding that is suitable for progressive
transmission11. In this method, the original image is filtered to be down-sampled by a factor of
2. The image thus obtained serves as a decimated image of the original. Then, the decimated
image is filtered to be interpolated by a factor of 2 to have the same size of the original image.
The difference between the original image and the interpolated one generates an error image.
This is called the first level PC decomposition. This process can be further repeated over the
decimated image to obtain higher levels. To achieve compression, the difference images and
the decimated image are bit allocated depending on the amount of information in each
subband. Those subbands with high information content will be assigned higher bit rate than
those with lower information content.
3.1 24-tap filter

In 12, a 24-tap FIR filter has been introduced with a nominal bandwidth of π/4, which would
allow a higher decimation factor than the conventional Gaussian filter used for pyramid
coding 11. The special property of this filter is that the passband ripples are designed for
passband flatness. This ensures that the filter has as little in-band ripples as possible. It has a
nominal bandwidth of ̟/2M where M is the decimation factor (M = 4).
10 Video Signal Transmission for IS-95 Environment, Electronic Letters, Vol. 36, no. 5, pp. 465-466 (Al-
Asmari et al., 2nd March 2000).

11 The Laplacian Pyramid as a compact Image Code, IEEE Trans Communication., vol. COM-31, no. 4
pp. 532-540, (Burt & Adelson, April 1983).

12 Optimum Bit Rate Pyramid Coding with Low Computational and Memory Requirements, IEEE
Trans. Circuit and Systems for video Tech., vol. 5, no. 3 pp. 182-192,(Alasmari, June 1995).
By using a pyramidal coding scheme which basically follows a rate of change of 4 instead of
2 as used in the conventional pyramid coding, the number of samples to be encoded are 20%
less than the conventional pyramidal samples12–13. Another advantage of this filtering
technique is that the 24-tap FIR filter involves 33% fewer computations as compared to the
Gaussian filter when FFT algorithm is used. The third advantage of this filter is the lower
entropy obtained when compared with that found when the Gaussian filter is applied12.8
4. The adopted spatial domain filtering techniques

The four temporal subbands (TLL, TLH, THL and THH) resulting from the temporal filtering
process are further decomposed in the spatial domain using pyramid coding as shown in
Figure 4. The temporal subband (TLL) is further decomposed into three subbands using the
Fig. 4. Spatial decomposition for the temporal subbands using pyramid coding.
13 Low complexity subband encoding for HDTV images, IEEE J. Select. Areas Commun., vol. 11, no. 1,
pp. 77-87 (Coppisetti et al., Jan. 1993).

24-tap filter. It is found that the 24-tap filter will not give a good performance when used to
decompose the temporal (TLH, THL and THH) bands due to the spares nature of the
information in these subbands10. The Gaussian filter given in11 is used to decompose these
subbands. Figures 5, 6, 7, and 8 show the spatial subbands for the temporal bands of Miss
America sequence using the pyramid coding concept.
Fig. 5 shows the spatial pyramid decomposition of the temporal TLL subband. Two levels of
pyramid coding have been applied for this temporal subband. The decimated image (band
1), which contains most of TLL subband information, is of dimension 18 × 22 pixels. The
difference images of this band (band 2 and band 3) contain edge components and are of
dimension 72 × 88 pixels and 288 × 352 pixels; respectively.
TLL band 3
band 2
band 1
Fig. 5. Spatial subbands for the temporal low-low band.

The spatial pyramid decomposition of the temporal TLH subband is shown in Figure 6. One
level pyramid coding has been applied for this temporal subband. The decimated image of
this band (band 4) is of dimension 144 × 176 pixels. The difference image (band 5) is of
dimension 288 × 352 pixels.
TLH band 5
band 4
Fig. 6. Spatial subbands for the temporal low-high band.

In Figure 7 the spatial pyramid decomposition of the temporal THL subband is shown. One
level pyramid coding has been applied for this temporal subband. The decimated image
(band 6) is of dimension 144 × 176 pixels while the difference image (band 7) is of dimension
288 × 352 pixels.
THL band 7
band 6
Fig. 7. Spatial subbands for the temporal high-low band.

Fig. 8 shows the spatial pyramid decomposition of the temporal THH subband. One level
pyramid coding has been applied for this temporal subband. The decimated image (band 8)
is of dimension 144 × 176 pixels and the difference image is of dimension 288 × 352 pixels.
THH band 9
band 8
Fig. 8. Spatial subands for the temporal high-high band.

5. Local adaptive vector quantization

Once the original image has been decomposed and the redundancy in the data removed, the
next step in the image compression problem is to encode the constituent bands. It has been
found that most of the energy signal resides in the lower spatial frequency subands, namely
bands 1, 2, and 3. Subbands 4, 5, 6, 7, 8, and 9, which corresponding to the high frequencies,
carry most of the motion and edge information and acts as a motion detector. Thus, by
accurate coding of low spatial-temporal bands, the spatial details of the original image are
conserved. Unlike the majority of the works on the structuring of VQ codebooks, the
primary goal of this work is to make the codebook simple and robust to the motions which
occur in video sequences and which are seldom capture from a single training sequence.
Therefore, the local adaptive vector quantization (LAVQ) 14 is adapted to encode some of the
spatial subbands.
A simple and effective one-pass image compression algorithm is provided by the local
adaptive vector quantization (LAVQ) 15. The encoder has a codebook containing
codewords (vectors). Each of these codewords is assigned an index corresponding to its
position in the codebook. The image is scanned block by block. Each time, a block is taken
and compared to the stored codewords. If there is a codeword sufficiently close to the
image block (within the pre-decided error) the index itself is sent, and that codeword is
moved to the top of the codebook in both transmitter and receiver. If the codebook search
is complete without accepted codeword, a special index is sent and followed by the block
itself. Now, this block is considered a new codeword and is placed at the top of the
codebook. All other codewords are pushed down, and if the number of codewords
exceeds the maximum allowed, the last codeword is lost 14. The LAVQ algorithm
maintains the most recently used vectors in the codebook in order of last usage. This
allows the LAVQ algorithm to be quick and efficient for any image to be encoded without
codebook training. New codewords are generated more often in regions containing edges
and fine features, while blank regions are coded with fewer new codewords. Therefore,
LAVQ is suitable for the high bands where the correlation is low. The properties of one-
pass and high speed codebook generation and encoding are the two main advantages of
LAVQ14. 9
5.1 Adaptive dead zone for the high subbands

Before we apply the LAVQ on the spatial subbands, a dead zone for each subband based on
the number of occurrence of the pixels around zero value is selected. Then, the band is
divided into vectors. The number of zero vectors after applying the dead zone will be
increased.
So, instead of encoding and sending all vectors, we just encode and send the non-zero
vectors with its location. Figure 9 shows an example for band 5 before and after applying the
dead zone concept. This process reduces the number of vectors to be transmitted to 20% –
30% of that needed if the dead zone approach is not adopted.
14 Analysis of Coding and Compression strategies for Data Storage and Transmission, Ph. D. thesis,
california Instiute of Technology (Sayano, 1992).

15 Image compression scheme using improved basic-LAVQ and optimized VLC, J. King Saud Univ., Vol. 8,
Eng. Sci. (2), pp. 251-266 (Al-Asmari et al., 1996).

Fig. 9. Example of dead zone process for band 5.
5.2 Searching method for the LAVQ

The difference (error) between the image vectors and the codebook vectors can be calculated
using LAVQ searching method concept as follow:
1/2
 cd (V ( j ) − Vm ( j ))2 
dm =  ∑ i  (m = 1, 2,...Cs and i = 1,1 + Cd ,1 + 2Cd..., mxn) (1)
 j =1 cd 
 
Where
dm is the root mean square error (RMSE).
Vi is the image vector.
Vm is the codebook vector.
cs is the codebook size.
cd is the codeword dimension.
The RMSE (dm) will be compared with a pre-decided threshold error (Vth). If dm is less
than this threshold, then the index of the codebook vector with lowest error will be
transmitted and this vector is moved to the top of the codebook in both transmitter and
receiver. Otherwise, the image vector will be transmitted and placed at the top of the
codebook.
6. Encoding of the subbands

High bit allocation is assigned for the baseband (band 1) since most of the energy of the
decomposed TLL image is concentrated in this band. For the high bands (band 2 – band 9),
different encoding algorithms are design to be tested for these bands. The first encoding
algorithm is to test the LAVQ for all the high bands (band 2 – band 9). The second technique
is to encode some of these bands using the edge detection concept, then applay the LAVQ
on the detected information. The third approach is to encode some of these high bands
using the absolute moment block truncation coding (AMBTC) algorithm16. The overall
encoding algorithm is decided based on the highest performance in terms of the peak signal
to noise ration (PSNR) and the visual quality for the encoded subband. For the first
algorithm, those bands reconstructed with PSNR greater than 51 dB will be encoded with
LAVQ. The second algorithm is adapted for those subbands encoded with PSNR greater
than 40 dB and with bit rate less than 0.06 bpp. The overall encoding algorithm is decided
based on the best performance for each subband. 10
The second level in the pyramid has the high frequency content of the decomposed
temporal TLL subband. From the simulation results of the previous three algorithms,
it is found that AMBTC algorithm will give the best performance for this subband (band 2)
because it is highly correlated than the other subbands. The mean, absolute moment and bit
map are transmitted for each subblock.
The first difference level (band 3) in the pyramid has the minimum information most of
which is concentrated around the edges. This information is encoded by applying an edge
detection approach to find the location of pixels that are perceptually important and then
transmitting only these encoded pixels. To avoid the transmission of the position of the
encoded locations, a predicting scheme for the edges of band 3 from the encoded bands
(band 1 and band 2) is adapted at the receiver side. This is shown in Figure 10.
Band
Band 11 4
H24
h24
Band Edge
Edge-
Band 22 4 H24
h24 detection
detection
H24
Mapping
Mappingthe
the edge
–
edge Coding the pixels at
Band
Band 33 pixels to band 3 the edge- location
Fig. 10. Coding of Edge-detection for band 3.

Before encoding, the baseband is interpolated to the second level and added to band 2.
Then, the result is further interpolated to the size of band 3. Edge-detection is applied to this
up-sampled version and the corresponding pixels from band 3 are formed into vectors to be
encoded for transmission. At the receiver, similar process is repeated by upsampling the
encoded baseband to be added to the encoded band 2. Then, these two subbands are up-
16 Absolute moment block truncation coding and application to color images, IEEE Trans. on Commun.,
Vol. COM-32, pp-1148-1157 (Lemma & Mitchell, Oct. 1984).

sampled to have the same size as band 3. This partially reconstructed version of the original
image is used for edge-detection, which gives the location of the vectors of the band 3 that
are encoded. Once the locations are known, the encoded vectors are suitably placed to form
band 3. Thus, using this approach, no side information needs to be sent for the encoded
areas of the first difference image level and an average of only 4 % to 5% of this level needs
to be encoded. Thus, more compression is achieved by edge-detection instead of coding the
entire band. After edge-detection, the data is encoded by using LAVQ.
The decomposed subbands (band 4 and band 5) of the temporal (TLH) subband are encoded
using the LAVQ for band 4 and the edge-detection approach for band 5. Band 4 is interpolated
to the size of band 5. Then, the edge-detection technique is applied to the interpolated version
and the corresponding pixels from band 5 are formed into vectors to be transmitted. At the
receiver, this process is repeated by interpolating the decoded band 4 to the size of band 5.
Then, edge-detection is taken for this band to decide the location of the received vectors for
band 5. The same decoding process is adapted for bands 6 and 7 to get the temporal THL
subband. Band 8 and band 9 formed the temporal THH subband. Band 9 has extremely low
energy content and the sparse information carried in this band is not significant for the final
image reconstruction. Thus, it can be safely discarded. Band 8 is encoded using the LAVQ
algorithm. At the receiver, this band is interpolated to get the temporal THH subband.
The subbands encoded using LAVQ have different codebooks sizes and vectors (codewords)
dimension. The choice of the codebook size and the codeword dimension depends on many
factors such as the important of the information to be transmitted, the correlation between the
data in each band, and the size of the band to be encoded. Band 4, for example, is more
important than band 8 because it has some of the motion information, and also used at the
receiver as a detector to encode band 5. Therefore, some care shall be given to this band. The
codebook size is selected to be 128 codewords each of dimension 4. Band 6 and band 8 have
most of the edges information, which already emphasized by encoding band 7. Therefore, a
codebook with smaller size and a codeword of bigger dimension than that of band 4 can be
adapted. From the simulation results, we find that the codebook of size 64 with codeword
dimension of 8 will give an excellent reconstructed video sequence quality.
Since the encoded vectors corresponding to the edge-detection concept are the only
information used to reconstruct band 3, band 5, and band 7, then, these vectors shall be
quantized with lowest possible MSE. Therefore, codebooks of size 64 and codewords of
dimension 4 pixels are selected to encode those subbands. Table 3 shows the encoding
techniques, PSNR, and bit rate for Miss America sequence.
Encoding technique PSNR Avg. Bit rate (bpp)

Band 1 AMBTC 45 0.0027
Band 2 AMBTC 43.37 0.02536
Band 3 LAVQ + Edge-detection 40.3749 0.0543
Band 4 LAVQ 52.6049 0.0133
Band 6 LAVQ 54.2208 0.0123
Band 8 LAVQ 51.7706 0.0145
Average 37.2 0.137
Table 3. The average bit rates and PSNR for Miss America sequence with Vth = 9.
The compression algorithm is tested on three video sequences with different motions and
backgrounds. The first sequence is Miss America sequence with slow motion and static
background. In this sequence, the only moving objects are the lips and head. The second
sequence with moderate motion and noisy background is the Salesman sequence. The man’s
head and hand are moving faster than Miss America sequence and with noisy background.
The third sequence with fast motion than the man’s sequence is Walter sequence. All of these
sequences were 256 level gray-scale images with dimension of 288 x 352 pixels per frame at the
rate of 30 frames /s and 8 bits per pixel. These sequences are standard and are used by many
researchers. The only way to check the performance of our proposed algorithm is to test it on
such images and compare the results with other compression algorithms.
7.1 Calculating the bit rates

The baseband (band 1 with 18×22 pixels) and band 2 (with 72×88 pixels) are encoded using
AMBTC because the information in those subbands is highly correlated.
Band 1 and band 2 are divided into sub-blocks to be encoded by using AMBTC algorithm.
The average bit rate needed to encode each band is calculated as follow:
l1 l2
n
× n
( Bm + Bh − l ) × ( 4 4 )
Bd
Bitrate = (2)
l1 × l2
Where Bm is the bit-map,Bh-1 is the required bits to encode the high and the low mean, Bd is
the sub-block dimension, l1×l2 is the original image frame dimension, and n is the pyramid
level number (n = 1, for first level, and n = 2 for second level). In this paper, Bd is selected to
be 3x3 for band 1, and 4x4 for band 2. The high mean and the low mean for band 1 are
encoded at 8 bits each, while for band 2, the high mean and the low mean are encoded at 6
and 4 bits; respectively.
The bit rate calculation for the Miss America sequence according to equation (2) is shown
below for four frames of the original image sequence; namely, frame 4,5,6 and frame 7.
These frames are considered to have the highest motion among this test sequence.
18 × 22
(9 + 16) ×
Bitrate(band 1) = 3 × 3 = 0.01085 = 0.0027 bpp (3)
288 × 352 4
72 × 88
(16 + 10) ×
Bitrate(band 2) = 4 × 4 = 0.10156 = 0.0253bpp (4)
288 × 352 4
The second encoding technique (edge detection + LAVQ) is used to encode bands 3, 5, and
7. First, the edge concept is applied for those bands. Then, those pixels corresponding to the
positions represented with “1” are encoded using the LAVQ technique.
The first algorithm (LAVQ) is found to be of high performance for bands 4,6 and 8 since the
motion and edges information are presented in these bands. The average bit rate for each
band can be calculated using the following formula:
( X × b ) + (Y × ((cd × 8) + b ))
Bitrate = (5)
l1 × l2
Where X is the number of matched codewords, b is the number of bits needed to encode the
index, Y is the number of non-matched codewords, and cd is the dimension of codeword.
Since the eight subbands represent four frames, then the bit rate for each frame is given by
the sum of the bit rates for each band divided by a factor of four.
Simulation results on these sequences can be discussed based on three main factors: peak
signal to noise ration (PSNR), bit per pixel (bpp) and visual quality. The PSNR in decibel
(dB) is given by;
 2552 
PSNR = 10 log 10   (6)
 MSE 
Where MSE is the mean square error written as follow;
1 m n ~
MSE = ∑∑ ( − )2
mxn i = 1 j = 1 X ij X ij
(7)
where;
m is the number of rows.
n is the number of columns.
~
X is the reconstructed pixel value.
Table 4 shows the average bit rate and average PSNR with different pre-decided thresholds
error (Vth) for the three sequences. These thresholds are compared with the RMSE (dm) that
result from LAVQ searching method. The bigger the Vth the lower number of non-matched (Y)
codewords will be. This will reduce the average bit rate required for transmission as given in
equation (5). However, the quality of the reconstructed image will be effected. Therefore, a
compromised between the bit rate and the required visual quality shall be decided.
Miss America Salesman Walter

PSNR 37.34 36.5 35.6
Vth = 6 Bpp required 0.143 0.172 0.207
Bit rate 0.435 Mbps 0.523 Mbps 0.629 Mbps
PSNR 37.2 36.31 35.42
Vth = 9 Bpp required 0.137 0.161 0.198
PSNR 36.9 36.2 35.15
Vth = 12 Bpp required 0.13 0.154 0.191
Table 4. Performance characteristics for different test sequences.
Three different thresholds have been tested in this study. The best result in terms of PSNR
(dB) is given at Vth = 6. However, at this threshold the bit rate requirement is higher than
that at Vth > 6. At Vth = 12, the bit rate for the three sequences is very low. It has been found
from the simulation results that the visual quality of the reconstructed sequences is excellent
at Vth = 12. Figure 11 shows the original and the reconstruction of the second four frames of
Miss America sequence at Vth =12. The visual quality of the reconstructed sequence is the
same as the original. From the simulation, it can be concluded that this compression
algorithm is capable of compressing a video sequences of different motions.
Frame 4 Frame 5
Frame 6 Frame 7
(a)
Frame 4 Frame 5
Frame 6 (b) Frame 7
Fig. 11. (a) Original and (b) Reconstructed frames of the second four frames of Miss America
sequence at Vth = 12.
The PSNR (dB) via the number of frames is demonstrated for the three sequences at
different thresholds. Also the average bit rate (bpp) versus the number of frames is
presented for 16 frames of the test video sequence. Since four frames are simulated as one
group at a time, only 4 different bit rates will be observed. However, the subband TLL is
transmitted first. Then, band TLH which is the second important subband regarding the
information contents. This approach will be followed until all the subbands are transmitted.
Accumulative calculation for the bit rate is then adopted in order to plot the bit rate curves.
Figure 12 shows the overall performance in terms of PSNR (dB) and bit rate (bpp) of the
proposed algorithm for Miss America sequence.
0.2 Threshold=7
Threshold=8
Threshold=9
0.18 Threshold=10
Threshold=11
0.16
0.14
bpp
0.12
0.1
0.08
0.06
2 4 6 8 10 12 14 16
Number of frames
(a)
38
Threshold=7
Threshold=8
37.8 Threshold=9
Threshold=10
Threshold=11
37.6
PSNR in dBs
37.4
37.2
37
36.8
2 4 6 8 10 12 14 16
Number of frames
(b)
Fig. 12. Performance curves. (a) PSNR vs. number of frames for Miss America sequence. (b)
bpp vs. number of frames for Miss America sequence.
8. Performance evaluation
The performance evaluation of the proposed algorithm is done in two stages. First, it is
compared with the performance of MPEG standard algorithm. Second, it is compared with
some existing research works in this field using multiresolution decomposition concept.
8.1 Comparison with MPEG-1

In17, we have tested MPEG-1 for the monochrome Miss America sequence at 30
frames/second. The results are produced for CIF (288 × 352) image format. The PSNR is 37
dB and the bit rate is 0.343 bpp17. The proposed algorithm in this paper achieves 36.9 dB at
0.13 bpp. At approximately the same PSNR, the saving in bit rate is about 56% as shown in
Table 5. The visual quality is excellent in both MPEG-1 and the proposed technique.
However, the encoder/decoder of MPEG-1 is more complex than our scheme. MPEG-1
based on a coder uses the DCT transform for the intraframe and the motion compensation
(MC) for the interframe. Studies reveal that if there are many moving objects in the image,
each moving in a different direction, the search technique becomes computationally
complex, involving larger storage and delay problems17. 11
MPEG-1 KARL [7] NEHAL [8] Al-Asmari [10] Al-Asmari [17] This Method
Interframe Motion Temporal 2-tap Temporal 2-tap Temporal 2-tap Motion Temporal 4-tap
Coding compensation filtering SSKFs filtering SSKFs filtering SSKFs compensation filtering SSKFs
DCT –
Intraframe SBC + ADPCM PC + DPCM or PC + FSCL AMBTC and PC + AMBTC +
Transform
Coding and PCM BTC (VQ) quantizetion LAVQ
(JPEG)
Average
37 dB 36.9 dB 36.5 dB 36.52 dB 37 dB 36.9 dB
PSNR
Bpp 0.343 0.434 0.273 0.25 0.2 0.13
Relative
Complex Moderate Low Moderate Moderate Low
complexity
Excellent Excellent
Competitive Excellent Excellent
quality with Good Quality at quality with
Comments quality with quality with quality at
moderate bit high bit rate very low bit
low bit rate low bit rate Low bit rate
rate rate
Table 5. Comparison with different compression methods.
8.2 Comparing with other works

The performance of the proposed algorithm is compared with the results obtained in related
works in7-8, 10, and17 in terms of PSNR, bit rate, and complexity. The comparison between
those algorithms is presented in Table 5. Karlsson and Vetterli in7, and Nehal, et al. in8 have
suggested schemes for progressive transmission using temporal filtering with 2-tap
symmetric short kernel filters (SSKFs). The reconstruction quality is very good. For7, the
PSNR is 36.9 dB with an average bit rate of 0.434 bpp while for8 the PSNR is 36.5 dB and the
bit rate is 0.273 bpp. Al-Asmari, et al. in 10 have suggested an algorithm for video
compression. The 2-tap SSKFs filters are used for temporal domain and pyramid coding for
intraframe. The decomposed bands are encoded using VQ called frequency selective
competitive learning (FSCL). This technique used the neural network concept to design the
17 Low Complexity Video Compression Algorithm Using AMBTC, Proceeding of IEEE Military
communication conference, Atlantic city, NJ (Al-Asmari et al., 31 Oct. – 3 Nov. 1999).

codebook for the vector quantization. This algorithm gives an excellent image quality for the
Miss America sequence at an average bit rate of 0.25 bpp and PSNR 36.52 dB. This algorithm
is considered to be of higher complexity than our algorithm because of the codebook design.
In 17, the authors present a coder based on a combination of AMBTC for intraframe and MC
for interframe. They produce results for the monochrome Miss America sequence. For the
CIF format (i.e. 288 × 352) at 30 frames/sec, they achieve approximately 37 dB at 0.2 bpp.
The disadvantage of this algorithm is the use of motion compensation for fast motion video
sequence. For the same video sequence, the proposed algorithm in this paper gives a higher
PSNR (37.2 dB) and a lower bit rate (0.13 bpp) than those algorithms for a coder with lower
complexity.
9. Conclusion
The results presented here are better than other coding schemes, which are published using
almost the same coding technique concept. The 3-D decomposition does not make
unrealistic assumptions about the data, as do methods based on motion compensation (MC).
Moreover, coding and decoding for the proposed algorithm are of comparable and
relatively low complexity. The robustness obtained by adapting the LAVQ for the codebook
has been discussed. The results reported in this paper are independent of which sequence is
used to produce the codebook.
This scheme is faster than MPEG algorithms and other existing technique based their
encoder on VQ concept since no need for training set or codebook generation. This scheme
will be an optimal choice for real time transmission. It is well suited for progressive
transmission of the video sequence and for browsing moving images via the Internet.
LAVQ technique gives good performance with those bands, which are highly uncorrelated,
and with spark information. Bit rate is varying from 0.13 to 0.191 bpp depending on the
nature of the sequence. Different video sequences have been tested and show very good
image quality with PSNR in the range of 36.9 to 35.15 dB and with bit rate range from 0.395
Mbps to 0.58 Mbps as shown in table 4.
10. References
Al-asmari Awad Kh. (1995), Optimum Bit Rate Pyramid Coding with Low Computational and
Memory Requirements, IEEE Trans. Circuit and Systems for video Tech., vol. 5, no. 3
pp. 182-192, (June 1995).
Al-Asmari Awad Kh., Ahmed Abobakr & Al-Doweesh Abdullah (1996), Image compression
scheme using improved basic-LAVQ and optimized VLC, J. King Saud Univ., Vol. 8,
Eng. Sci. (2), pp. 251-266.
Al-Asmari Awad Kh., Aryai Deepali, & Kwatra Subhash C. (2000), Video Signal Transmission
for IS-95 Environment, Electronic Letters, Vol. 36, no. 5, pp. 465-466, 2nd March
2000.
Al-Asmari Awad Kh., Dave Sameep & Kawatra Subhash C. (1999), Low Complexity Video
Compression Algorithm Using AMBTC, Proceeding of IEEE Military communication
conference, Atlantic city, NJ, (31 Oct. – 3 Nov).
Al-Asmari Awad, Singh Vinay & Kawatra Subhash (June 28 – July 1 CIST 99), Hybrid coding
of images for progressive transmission over a digital cellular channel, CISST’99
International Conference on Imaging Science, Systems and Technology, Monte

Carlo resort, Las Vegas, Nevada, USA, PIN 128C.
Bhaskaran Vasudev, Konstantinidides Konstantions & Hewlett Packard Laboratories (1996),
Image and video compression standards: algorithms and architectures, Kluwer Academic
Publishers.
Burt P. & Adelson E. (1983), The Laplacian Pyramid as a compact Image Code, IEEE Trans
Communication., vol. COM-31, no. 4 pp. 532-540, (April 1983).
Child J. (July 1993), Realtime video compression poses challenge to designers and vendors alike,
Computer Design, vol. 32, no. 7, pp. 67-70.
Clarke.R. J., (1995) Digital compression of still images and video, Academic Press.
Coppisetti N., Kwatra S. C., & Al-asmari A. Kh. (1993), Low complexity subband encoding for
HDTV images, IEEE J. Select. Areas Commun., vol. 11, no. 1, pp. 77-87, (Jan. 1993).
Dantwala Nehal, Kwatra Subhash C. & Al-Asmari Awad Kh. (June 30 – July 3 - CISST’97),
Pyramid coding of video signals for progressive transmission, CISST’97 International
Conference, pp-392-398.
Karlsson .G. & Vetterli. M. (July 1988), Subband coding of video for packet networks, Optical
Engineering, vol. 27, no. 7, pp. 574-586.
Lemma Maximo D. & Mitchell O. R. (1984), Absolute moment block truncation coding and
application to color images, IEEE Trans. on Commun., Vol. COM-32, pp-1148-1157,
(Oct. 1984).
Netravali. Arun . N & Haskell Barry G. (1995), Digital pictures representation, compression, and
standard, Plenium Press.
Sayano M. (1992), Analysis of Coding and Compression strategies for Data Storage and
Transmission, Ph. D. thesis, california Instiute of Technology.
Sayood Khalid (1996), Introduction to data compression, Morgan Kaufmann Publishers Inc.,
San Francisco, California.
Shannon.C. E (Oct. 1948), A mathematical theory of communication, Bell System Tech. J.. Vol.
27, pp. 379-423, July. 1948 and pp. 623-656.
8
Low Complexity
Implementation of Daubechies
Wavelets for Medical Imaging Applications
Khan Wahid
University of Saskatchewan,
Canada
1. Introduction
The Discrete Wavelet Transform (DWT) has extensively been used in a wide range of
applications, including numerical analysis, image and video coding, pattern recognition,
medical and telemetric imaging, etc. The invention of DWT decomposition by Mallat
(Mallat, 1998) shows that the DWT can be viewed as a multiresolution decomposition of
signal. This means it decomposes the signal into its components in different frequency
bands. The Inverse DWT does the opposite, i.e. it reconstructs the signal from its octave
band components. After its inclusion in JPEG2000 compression standard (Seo & Kim, 2007),
significant research has been done to optimize the DWT implementation to reduce the
computational complexity. Among a wide range of wavelets, the Daubechies wavelets
include members ranging from highly localized to highly smooth and can provide excellent
performance in image compression (Daubechies, 1992). Among the family members, the
first two – Daubechies 4-tap (DAUB4) and Daubechies 6-tap (DAUB6) – are popular choices
in medical imaging applications.
While compressing medical images, the key here is to preserve as much critical information
as possible in the reconstructed image so that accurate diagnosis is possible. There have
been several efficient implementations of wavelet filters proposed for applications in image
processing (Lee & Lim, 2006; Martina & Masera, 2007; Acharyya et al., 2009; Shi et al., 2009;
Lai et al., 2009). But, the use of conventional fixed-point (FP) binary (or any other weighted)
representation for implementing discrete wavelet coefficients (that are irrational in nature)
introduces round-off or approximation errors at the very beginning of the process. The error
is due to the lack of exact representation of the irrational numbers that form the coefficient
basis. These errors tend to expand as the calculations progress through the architecture,
degrading the quality of image reconstruction (Wahid et al., 2003). A lossless mapping
technique, known as Algebraic Integer Quantization (AIQ), can be used to minimize the
approximation error and efficiently compute the DAUB4 and DAUB6 coefficients (Wahid et
al., 2004). The AIQ scheme is divided into two parts: the first stage is based on factorization
and decomposition of transform matrices exploiting the symmetric structure. After the
decomposition, we map the irrational transform basis coefficients using multidimensional
algebraic integers that results in exact representation and simpler implementation. As a
result, less error is introduced in the computation process that yields significantly better
reconstruction of images while keeping critical information, making the scheme suitable for
medical and telemetric imaging applications.
As a case study, we apply the scheme to several medical images, such as endoscopic,
ultrasound, x-ray, CT-scan images and evaluate the performance. The chapter is organized
as follows: Previous related works are presented next. Section 3 presents a brief introduction
to Daubechies wavelets. In Section 4, we explain the AIQ scheme applied to Daubechies
wavelets. Then the simulation and synthesized results of the case study are summarized in
Section 5. Finally, we conclude the work in Section 6.
2. Past work
Lewis and Knowles proposed an architecture for Daubechies wavelets without multipliers
(Lewis & Knowles, 1991). A major drawback was that it was heavily dependent on the
properties of only one specific wavelet, DAUB4 tap coefficients. At the same time, Aware
Inc. came out with a chip called Wavelet Transform Processor (WTP) (Aware, 1991). It
essentially consists of a 4-tap filter (4 Multiply-Accumulate cells) and some external
memory with control but no specific features that can take advantage of the DWT
structure rather it relies heavily on the software to compute the DWT. It is also a complex
design requiring extensive user control. Parhi and Nishitani proposed two architectures,
folded and digit serial, for 1D DWT (Parhi & Nishitani, 1993). These architectures do not
easily scale with the filter size and the number of octaves computed. The number of
multipliers is higher, and hence the silicon area is large. In (Vishwanath et al., 1995), the
authors proposed linear systolic array architecture. Paek and Kim in proposed recursive
and semi-recursive architectures for DWT which has several drawbacks like large area
(hardware cost), scheduling control overhead and incomplete data-bus utilization (Paek &
Kim, 1998).
Most of the research work to reduce the hardware complexity is inclined towards
multiplierless implementations by maneuvering the filter banks (Lee & Lim, 2006; Martina
& Masera, 2007; Acharyya et al., 2009) or using lifting schemes (Shi et al., 2009; Lai et al.,
2009; Huang et al., 2004). However, in these designs, the use of conventional FP binary
representation results in erroneous computation process and degrades image
reconstruction. In this chapter, we present an efficient low-cost implementation of the
DAUB filters with a demonstration of performance advantages on medical images and noisy
environment.
3. Daubechies wavelets
This section provides a brief introduction to Daubechies wavelets. This class of wavelets
includes members ranging from highly localized to highly smooth – Daubechies-2 (DAUB2
with two coefficients) to Daubechies-20 (DAUB20 with 20 coefficients) and also provides
excellent performance in image compression (Daubechies, 1992). The Daubechies wavelet
coefficients are based on computing wavelet coefficients, C n (where, n = 0, 1, 2,..., N-1 and N
is the number of coefficients) to satisfy the following conditions (Mallat, 1998):
1. The conservation of area under a finite length signal x(t ) :  C n  2
2. The accuracy conditions:  ( 1)n nmC n 0 (where m = 0, 1, 2,...,
n
p-1 and p  N )
2
3. The perfect reconstruction conditions:  C n  2 and  C nC n  2 m  0
n 2
n n
Low Complexity Implementation of Daubechies Wavelets for Medical Imaging Applications 123
Cn
Then the low-pass filter is h( n)  and the high-pass filter is g(n)  ( 1)n  1 h(n  N  1) .
2
One of the simplest and most localized members is the DAUB6 which has six coefficients:
(1  z1  z2 ) (5  z1  3 z2 ) (10  2 z1  2 z2 )
C0  C1  C2 
16 2 16 2 16 2
(1)
(10  2 z1  2 z2 ) (5  z1  3 z2 ) (1  z1  z2 )
C3  C4  C5 
16 2 16 2 16 2
Where, z1  10 and z2  5  2 10 . For an 8x8 input data, the DAUB6 forward transform
matrix (using an assumption of periodicity) is shown in Eq. (2):
C 0 C1 C 2 C3 C4 C5 0 0 
C C 4 C 3 C 2 C 1 C 0 0 0 
 5
0 0 C0 C1 C 2 C3 C4 C5 
 
0 0 C5 C 4 C 3 C 2 C1 C 0 
 6 (C )   (2)
C C5 0 0 C0 C1 C2 C3 
 4 
C 1 C 0 0 0 C5 C 4 C 3 C 2 
C C3 C4 C5 0 0 C0 C1 
 2 
C 3 C 2 C 1 C 0 0 0 C5 C 4 
C0 -C5
x0
C1 -C4
x1
C2 C3
x2
C3 -C2 LOW0 / HIGH0
x3
C4 C1
x4
C5 -C0
x5
Fig. 1. FP-based DAUB6 filter architecture.

The structure of the matrix uses the set of coefficients, {C 0 , C 1 ,..., C 5 } as a smoothing filter
(low-pass) and the set, {C 5 , C 4 ,..., C 0 } as a non-smoothing filter (high-pass). The DWT is
invertible and orthogonal - the inverse transform, when viewed as a matrix, is simply the
transpose of the forward transform matrix. So, basically we need only 2 sets of multiply-
accumulate (MAC) cells each containing 6 multipliers and 5 adders where partial
products are computed separately and subsequently added. However, it can be seen that,
by introducing additional control circuitry, the same multipliers can be used for both low-
pass and high-pass filtering. As a result, the number of multipliers can be reduced to 6
instead of 12. Fig. 1 shows the signal flow graph of the conventional finite-precision (FP)
implementation of Eq. (2), where xi is the input data vector. Since all the coefficients are
fixed, for a fixed precision, we can in fact replace all the multipliers by adders and
shifters. As a result, the total equivalent additions required to compute the 1-D DAUB6
filter is 44.
4. AIQ-based algorithm
Algebraic integer (AI) is defined by real numbers that are roots of monic polynomials with
2 j
integer coefficients (Wahid et al., 2004). As an example, let   e 16 denote a primitive 16th
root of unity over the ring of complex numbers. Then  satisfies the equation: x 8  1  0 .
The ring Z( ) can be regarded as consisting of polynomials in  of degree 7 with integer
coefficients. The elements of Z( ) are added and multiplied as polynomials, except that the
rule  8  1 is used in the product to reduce the degree of powers to below 8.
In summary, algebraic integers of an extension of degree n can be assumed to be of the
form:
a00  a11  ...  an  1n  1 (3)
Where, {0 , 1 ,..., n  1 } is called the AI basis and the coefficients ai are integers. The process
of mapping with AI is known as Algebraic Integer Quantization (AIQ).
The AIQ technique is useful in computing discrete transforms as first explored by Cozzens
and Finkelstein (Cozzens & Finkelstein, 1985). In their work, the algebraic integer number
representation, in which the signal sample is represented by a set of (typically four to eight)
small integers, combines with the Residue Number System (RNS) to produce processors
composed of simple parallel channels. The analog samples must first be quantized into the
algebraic integer representation and the final algebraic integer result converted back to an
analog or digital form. In between these two conversions, the algebraic integer
representation must be converted into and out of two levels of RNS parallelism.
4.1 AIQ-based Daubechies wavelets

Now we present the concept of AIQ to encode the DAUB6 coefficients. First of all, consider
the polynomial of two variables:
1 1
 a00 a10   1 z2 
f ( z1 , z2 )   aij zi 1 z j 2   (4)
i 0 j 0  a01 a11   z1 z1 z2 
a a10 
So the corresponding coefficients, aij , are encoded in the form  00 . Then all the
 a01 a11 
DAUB6 coefficients are exactly encoded (scaled by 16 2 ) and shown below in Eq. (5):
1 1 5 1 10 2 
C0   C1   C2  
1 0  3 0  2 0

(5)
 10 2  5 1  1 1
C3   C4   C5  
 2 0   3 0  
 1 0 
The obvious advantages of this approach are: a) Very small dynamic range (numbers
ranging from 0 to 10); b) Multiplication by a constant is very easy and efficient (only 1
addition is required in most cases; so, multiplication can be eliminated by add/shift
algorithm); and c) We have 3 parallel channels through which data flows independently
(since a11 is zero for all) and also a very simple scheduling is needed. No quantization errors
would be incurred. Fig. 2 shows the signal flow graph of the AIQ-based scheme that
requires 42 adders.
Fig. 2. AIQ-based DAUB6 filter architecture.

The encoding scheme can be easily applied to DAUB4 coefficients. In that case, we need a
2nd degree polynomial of one variable:
f ( z3 )  a0  a1 z3  a2 z 2 3 (6)
Where, z3  3 . As a result, the error-free mapping of DAUB4 coefficients can computed as

given below (scaled by 4 2 ):
C 0  1  z3 ; C 1  3  z 3 ; C 2  3  z3 ; C 3  1  z 3 (7)
4.2 Hardware implementation

The AIQ-based architecture is coded in Verilog and prototyped onto Xilinx VirtexE FPGA to
assess the performance. A precision of 8-bit is used in the multipliers to minimize the
hardware and optimize the operation, which is completed in one clock cycle (CC). Table 1
presents the comparison of the synthesized results along with the fixed-point (FP) designs. It
can be seen from the table that the AIQ scheme for both DAUB4 (D4) and DAUB6 (D6) costs
lesser hardware resources and has lower critical path delay; in case of DAUB4, the savings is
even higher.
Scheme Fixed-point AIQ Overall savings (%)

D4 D6 D4 D6 D4 D6
Datapath 4 6 3 6 25 0
Adders 32 44 16 42 50 5
LUTs 268 364 124 340 54 7
Registers 422 520 200 494 53 5
Critical path Tm+2Ta Tm+3Ta 5Ta 6Ta -- --
Frequency (MHz) 98 112 148 120 -- --
1Xilinx VirtexE (xcv300epq240-8); D4 = DAUB4; D6 = DAUB6; Tm = latency for multiply operation; Ta =
latency for addition operation
Table 1. FPGA1 implementation and hardware comparison.
A key advantage of the AIQ approach is error reduction. In order to better understand the
sources of error induction, we present in Fig. 3 the entire process of digital computation
process. Here each arrow represents one key stage in the computation. In case of FP-based
approach, there are three stages of error accumulation (or induction). In stage 1, the
quantization (or approximation) error is introduced due to the lack of finite representation
of transform basis coefficients that are irrational numbers. The error gets larger and larger in
later stages as the process continues.
However, in AIQ-based approach, there is only one stage of error induction at the end: in
stage 1, the basis coefficients are mapped with algebraic integers (that is error-free); then all
the required computations take place (that is also error-free) using the AI representation.
Finally, the data is converted back to binary (in stage 3) where some errors may be
introduced; but these errors are much less compared to FP-approach, and are introduced
very late in the computation process, and hence less affect the quality of image
reconstruction.
2 3
2 , 3  5 , cos , sin
16 16
(a)
2 3
2 , 3  5 , cos , sin z, z1 , z 2  (0,1, 2,...)
16 16
(b)
Fig. 3. Stages in computation process (error induced in shaded blocks): (a) FP approach; (b)
AIQ approach.
Moreover, the error introduced at the final stage in AIQ approach can be further minimized
using higher precision AIQ multipliers. We have performed an error analysis that shows the
error incurred for different bit-length of the AIQ multipliers (in Fig. 4). The error is
computed taking a multiplier of 16-bit width as reference. The signed digit representation
(for 8-bits) is shown below in Eq. (8):
z1  10  11.001010  3  2 3  2 5
z2  5  2 10  100.10100  4  2 1  2 3 (8)
2 5
z3  3  10.010010  2  2  2
-3
x 10
7
Z1
6 Z2
Z3
5
Computation Error
0
6 7 8 9 10 11 12 13 14
Multiplier Precision (bits)
Fig. 4. Computation error in AIQ multipliers.

5. Performance evaluation
The AIQ-based algorithm to compute the Daubechies Wavelet Transform is intended to
be used in applications where the quality of image reconstruction is critical, such as
biomedical imaging, telemedicine, capsule endoscopy (Wahid et al., 2008), etc. Due to the
error-free nature of integer mapping, the AIQ approach results in a much better
reconstruction compared to conventional binary approach. Here, we present the results of
our study, where we apply the scheme to several standard benchmark and medical
images, such as endoscopic, ultrasound, x-ray, CT-scan images and evaluate the
performance.
The section is divided into four sub-sections. In the first section, we evaluate the
performance of the AIQ scheme for standard images followed by the analysis of medical
images. Next, we show the performance of the scheme in a noisy environment. Finally, the
results are compared with existing works related to medical image compression. In all these
cases, we have used peak-signal-to-noise-ratio (PSNR) as the visual quality assessment
index which is given by Eq. (9):
255
PSNR  20  log 10 (9)
1
 n 1 m1
 xm ,n  x 'm , n 
N M 2
MN
Where, M and N are the image width and height namely; x and x’ are the original and
reconstructed component values namely.
5.1 Performance analysis on standard images

Here, we perform 1-D forward transform on the benchmark “Goldhill” image followed by
the inverse transform to get the original image back. No compression was performed, so the
image quality degradation is purely due to arithmetic quantization effects. Different
hardware precision (number of bit) is used where a full adder is considered as the unit for
the hardware cost, and making a simple assumption of a hardware cost of n for an n-bit
word (this will be a best case comparison for the fixed-point binary implementation). The
results are shown in Fig. 5 and 6.
As shown in Fig. 5(a), the PSNR of the reconstructed image gets higher with the increased
bit precision which is expected. In all cases, the AIQ performs much better that FP, i.e., the
data recovery rate is higher, especially in lower bit rate region. In other words, for a fixed
level of distortion, the number of bits required to transmit the transformed coefficients of
AIQ approach would be less than those required for FP technique.
In Fig. 5(b), we present an analysis of reconstruction quality (in PSNR) with the cost of
implementation. An interesting comparison is to select similar hardware cost and then
compare the reconstruction performance. As an example, for same hardware cost of 150,
compare the PSNR of both FP (44dB) and AIQ implementation (82dB). In this case the
difference in the PSNR is 38dB in favour of AI – that is for similar cost of implementation,
the AIQ scheme produces much better image reconstruction. On the other hand, for same
PSNR, say 60 dB, AIQ scheme (95) requires around three times less hardware than FP
implementation (270).
DAUB4, Goldhill DAUB4, Goldhill

90 90
FP AIQ
80
80
70
70
60
60
PSNR (dB)
PSNR (dB)
50
40 50
30
40
20
30
10 FP
AIQ
0 20
6 8 10 12 14 50 100 150 200 250 300 350 400 450
Bit Precision Cost of Implementation
(a) (b)
Fig. 5. Performance analysis of DAUB4 – FP vs. AIQ: (a) PSNR vs. Precision; (b) PSNR vs.
Hardware cost.
DAUB6, Goldhill DAUB6, Goldhill

90 90
FP AIQ
80
80
70
60 70
PSNR (dB)
PSNR (dB)
50
60
40
30 50
20
40
10 FP
AIQ
0 30
6 8 10 12 14 100 200 300 400 500 600 700
Bit Precision Cost of Implementation
(a) (b)
Fig. 6. Performance analysis of DAUB6 – FP vs. AIQ: (a) PSNR vs. Precision; (b) PSNR vs.
Hardware cost.
Same kind of superiority is seen for DAUB6 (Figure 6) too. So, not only an improvement in
image reconstruction quality is obtained but also hardware cost is reduced. Fig. 7 shows the
image reconstruction for a 8-bit FP vs. a 14-bit AIQ for DAUB4. The difference in PSNR is
about 45dB and the level of improvement is quite noticeable.
(a) Original (b) PSNR = 43dB (c) PSNR= 88dB

Fig. 7. (a) Original goldhill image; (b) Reconstructed image using FP scheme (8-bits); (c) AIQ
scheme (14-bits).
5.2 Performance analysis on medical images

We have performed an exhaustive simulation using several medical images, such as, X-ray,
CT-scan, Ultrasound (US) and endoscopic images, and the results are presented in Table 2
(showing for two cases of bit precision: 8-bits and 12-bits). In all cases, the AIQ-based
scheme produces a high PSNR and outperforms the conventional approach by a far margin.
Some sample original and reconstructed images are shown in Fig. 8 – 11. In most cases, the
difference in PSNR is around 8dB with noticeable level of improvement.
D4-FP D4-AIQ D6-FP D6-AIQ

8 bit 44.4 52.5 41.1 49.3
CT
12 bit 66.7 76.6 65.5 76.6
8 bit 44.8 51.8 41.4 50.8
Endoscopic
12 bit 67.1 76.0 66.1 75.9
8 bit 48.0 55.9 44.8 53.7
US
12 bit 70.4 79.9 69.0 79.9
8 bit 43.3 50.1 39.9 49.2
X-ray
12 bit 65.5 74.1 64.5 74.1
Table 2. Quality of reconstruction (in terms of PSNR in dB) for medical images.

Fig. 8. (a) Original endoscopic image; (b) Reconstructed image using FP scheme; (c) AIQ
scheme.
Fig. 9. (a) Original x-ray image; (b) Reconstructed image using FP scheme; (c) AIQ scheme.
Fig. 10. (a) Original CT-scan image; (b) Reconstructed image using FP scheme; (c) AIQ
scheme.
Fig. 11. (a) Original US image; (b) Reconstructed image using FP scheme; (c) AIQ scheme.
5.3 Performance analysis on noisy images

As a final study, the AIQ algorithm is tested under a noisy environment. The Gaussian
white noise and Poisson noise are added to the images of all types, and the performance is
compared with the FP implementation. The results are tabulated in Table 3 (all using 8-bit
precision). Like previous cases, due to less error accumulation in the computation process,
the AIQ-based approach is seen to have performed better than FP approaches even in a
noisy environment.
PSNR (dB)
Noise Algorithm
Goldhill US X-ray Endoscopic CT
D4 - FP 43.9 47.5 43.1 44.6 44.4
D4 - AIQ 51.4 55.5 50.0 51.6 52.4
Gaussian
D6 - FP 40.6 44.3 39.7 41.3 41.1
D6 - AIQ 49.9 53.2 49.3 50.6 49.7
D4 - FP 44.1 47.6 43.2 44.8 44.6
D4 - AIQ 51.6 55.6 50.0 51.8 52.5
Poisson
D6 - FP 40.7 44.5 39.8 41.5 41.3
D6 - AIQ 50.1 53.4 49.3 50.9 49.9
Table 3. Quality of reconstruction in noisy environment.
5.4 Comparative analysis

In order to show the effectiveness of the AIQ approach, the algorithm is compared with
some compression standards: JPEG, JPEG2000, and existing algorithms targeted to medical
imaging. Since there is no benchmark for medical images, we have conducted the
experiment with benchmark images like “Lena”, “Barbara” and “Goldhill”. Table 4
summarizes the comparison results (all using 8-bit precision with 0.25 bits per pixel). From
the table, it is clearly observed that the error-free algorithm performs competitively
compared to other existing compression schemes.
Algorithm Lena Barbara Goldhill

HS-HIC (Mohammed, 2008) 35.0 26.1 30.5
Hybrid (Yu & Mitra, 1997) 35.0 31.5 32.9
JPEG (Wahid et al., 2008) 32.4 27.7 29.7
JPEG2000 (Seo & Kim, 2007) 34.1 28.8 30.5
OB-HIC (Mohammed & Abd, 2010) 35.9 32.8 33.8
D4-AIQ 44.9 45.0 35.8
D6-AIQ 43.2 43.6 35.3
Table 4. Comparative analysis (in terms of PSNR in dB) of the AIQ scheme.
6. Conclusion
In this chapter, we have presented an efficient approach to compute Daubechies wavelet
transforms that is based on encoding the basis set of forward transform coefficients using
algebraic integers. The AIQ approach not only reduces the number of arithmetic operations,
but also reduces the dynamic range of the computations. Because of error-free mapping in
the earlier stages, less error is introduced in the system, as compared to FP implementation,
that results in much better data reconstruction. The performance is validated using standard
and medical images in both normal and noisy conditions. In all cases, the AIQ-based
approach outperforms the conventional FP scheme by far margin. The rate of data recovery
is very high while preserving critical information that makes the scheme suitable for
medical and telemetric imaging applications.
7. Acknowledgement
The author would like to acknowledge the Natural Science and Engineering Research
Council of Canada (NSERC) for its support to this research work. The author is also
indebted to the Canadian Microelectronics Corporation (CMC) for providing the hardware
and software infrastructure used in the development of this design.
8. References
Acharyya, A., Maharatna, K., Al-Hashimi, B., Gunn, S. (2009), Memory reduction
methodology for distributed arithmetic based DWT/IDWT exploiting data
symmetry, IEEE Trans. on Circuits and Systems II, vol. 56, no. 4, pp. 285-289.
Aware Inc., (1991) Aware Wavelet Transform Processor WTP) Preliminary, Cambridge, MA.
Cozzens, J. and Finkelstein, L. (1985) Computing the Discrete Fourier Transform using
Residue Number Systems in a Ring of Algebraic Integers, IEEE Transactions on
Information Theory, vol. 31, pp. 580-588.
Daubechies, I., (1992) Ten lectures on wavelets, SIAM, 1992.
Huang, C., Tseng, P., and Chen, L., (2004) Flipping structure: an efficient VLSI architecture
for lifting based discrete wavelet transform, IEEE Trans. Signal Process., vol. 52, no.
4, pp. 1080–1089.
Lai, Y., Chen, L., Shih, Y. (2009) A high-performance and memory-efficient VLSI architecture
with parallel scanning method for 2-D lifting-based discrete wavelet transform,
IEEE Trans. on Consumer Electronics, vol. 55, no. 2, pp. 400 – 407.
Lee, S. and Lim, S., (2006) VLSI design of a wavelet processing core, IEEE Trans. Cir. Syst.
Video Tech., vol. 16, pp. 1350-1360.
Lewis, A. and Knowles, G. (1991) VLSI Architecture for 2D Daubechies Wavelet Transform
without Multipliers, IEE Electronics Letters, vol. 27, no. 2, pp. 171-173.
Mallat, S., (1998) A wavelet tour of signal processing, New York: Academic.
Martina, M. and Masera, G., (2007) Multiplierless, folded 9/7-5/3 wavelet VLSI
architecture,” IEEE Trans. Cir. Syst. II, 54, pp. 770-774, 2007.
Mohammed, U. (2008), Highly scalable hybrid image coding scheme, Digital Signal
Processing, Science Direct, vol. 18, pp. 364–374.
Mohammed, U. and Abd-elhafiez, W. (2010), Image coding scheme based on object
extraction and hybrid transformation technique, Int. J. of Engineering Science and
Technology, vol. 2, no. 5, pp. 1375–1383.
Paek, S. and Kim, L. (1998) 2D DWT VLSI Architecture for Wavelet Image Processing, IEE
Electronics Letters, vol. 34, no. 5, pp. 537-538.
Parhi, K. and Nishitani, T. (1993) VLSI Architectures for Discrete Wavelet Transforms, IEEE
Transactions on VLSI Systems, vol. 1, no. 2, pp. 191-202.
Seo, Y. and Kim, D. (2007) VLSI architecture of line-based lifting wavelet transform for
motion JPEG2000, IEEE J. Solid-State Circuits, vol. 42, no. 2, pp. 431-440.
Shi, G., Liu, W., Zhang, L., Li, F. (2009) An efficient folded architecture for lifting-based
discrete wavelet transform, IEEE Trans. on Circuits and Systems II, vol. 56, no. 4,
pp. 290-294.
Vishwanath, M., Owens, R. and Irwin, M. (1995) VLSI Architectures for the Discrete Wavelet
Transform, IEEE Transactions on Circuits and Systems - II, vol. 42, no. 5, pp. 305-
316.
Wahid, K., Dimitrov, V., Jullien, G. (2004) VLSI architectures of daubechies wavelet
transforms using algebraic integers, J. of Circuits, Sys., and Comp., vol. 13, no.6,
pp. 1251-1270.
Wahid, K., Dimitrov, V., Jullien, G. and Badawy, W. (2003) Error-Free computation of
daubechies wavelets for image compression applications, Elect. Lett., vol. 39, no. 5,
pp. 428-429.
Wahid, K., Ko, SB. and Teng, D., (2008) Efficient hardware implementation of an image
compressor for wireless capsule endoscopy applications, Proc. of the IEEE Int. Joint
Conf. on Neural Net., pp. 2762-2766.
Yu, T., and Mitra, S. (1997), Wavelet based hybrid image coding scheme, Proc. IEEE Int
Circuits and Systems Symp, vol. 1, pp. 377–380.
9
Discrete Wavelets on Edges

Alexandre Chapiro1, Tassio Knop De Castro1, Virginia Mota2, Eder De
Almeida Perez2, Marcelo Bernardes Vieira2 and Wilhelm Passarella Freire2
1Instituto de Matemática Pura e Aplicada,
2Universidade Federal de Juiz de Fora
Brazil
1. Introduction
Human life is closely tied to signals. These signals are present everywhere - listening to music
is possible because of audible sound signals traveling through air, reading a book is feasible
due to light waves bouncing off objects and interpreted by our bodies as visual images,
electromagnetic waves allow us to communicate through the radio or wireless Internet.
Signal Processing is an area of electrical engineering and applied mathematics that deals with
either continuous or discrete signals. Particularly, Image Processing is any kind of Signal
Processing where the input is an image, such as a digital photograph. The underlying
essence of Image Processing lies in understanding the concept of what is an image and
studying techniques for the manipulation of images with the use of a computer. While these
explanations may seem quite generic, the importance of Image Processing in the modern
world is undeniable and progress in this field is very desirable.
1.1 Images
The concept of an image can initially be mathematically defined as a function f : S → C
that goes from a certain space S (such as R2 , for instance) to a space C of colors that can be
perceived by the human eye. This definition does not exhaust all of the possible meanings of
this word, but will be enough for this chapter. When working on a computer, however, both
the domain and counter-domain of the image-function must be discrete. The most common
representation of an image in Image Processing thus consists of taking a discrete subset of S -
S and a function that associates the values of S to a certain subset of C - C . In this way, an
image I can be thought of as a discrete function I : S → C .
In this work and in Image Processing in general, the kind of image we are most interested in is
a digital image, usually obtained through a digital camera or generated by a computer. As the
previous mathematical definition, digital images are discrete, that is they are composed of a
finite number of elements. A digital image can be thought of as a mosaic of colors taken form a
certain set. In mathematical terms, a digital image can be represented via a matrix M ∈ Mn,m ,
composed of numbers that represent colors that can be shown by modern electronic devices,
such as televisions, computer monitors and projectors. Each element of this matrix is called a
pixel (this name comes from the words ’picture element’).
It is important to understand the concept of color. Initially, color is a sensation produced by
the human brain when it receives certain visual stimuli. This input is given by electromagnetic
radiation (or light) in a set wavelength that is called the visible spectrum. A typical human
eye will respond to wavelengths from about 390 to 750 nm. Theoretically speaking, the space
136
of all visible colors, as given by their wavelengths is of infinite dimension, and thus not fit for
a computer. This limitation is bypassed through the study of the human vision.
Scrutiny of the human eyes shows that they contain two different kinds of photo-receptor cells
that allow vision. These cells are rods and cones. Rods are very sensitive to light, being mostly
responsible for night vision and have little, if any, role in color vision. Cones on the other
hand are of three types (Short, Medium and Long), each covered in a different photo-sensitive
pigment. These pigments respond differently to incoming light wavelengths. A chart showing
the response of each kind of cone to light can be seen below in Figure 1.
By using the knowledge above, modern visual devices are built so that they emit light at only
three different wavelengths, specifically suited to excite each cone in a known way. This allows
devices to create a wide range of visible colors. While its not possible to re-create all possible
color sensations using only these three colors, the difference when using modern technology
is mostly imperceptible. Thus we have arrived at the discretization of the color space used
for digital images. These colors can now be codified as certain finite amounts taken in small
intervals of these three primary colors. A schematic of a digital image can be seen in Figure 2.
Fig. 1. Human eye response curves. (Image in Public Domain)
Fig. 2. Raster Image. (Image in Public Domain)
1.2 Applications of Image Processing

Image Processing has seen a great variety of methods developed in the last fifty years. These
techniques are greatly diverse and are present in various aspects of human daily life, as well
as other important scientific fields.
Some typical tasks in Image Processing involve text or pattern recognition by a computer
(machine vision), like identifying individuals from photographs, for instance using their face,
retina or fingerprints. In this last case, a specialized camera is used to create a digital image
of a person’s fingers. This image is then analyzed by a computer program that searches for
patterns, which are larger characteristics of the ridges in the skin, and minutia - smaller details
Discrete Wavelets
Discrete Wavelets on Edges on Edges 1373
such as ends and bifurcations of said ridges. Figure 3 shows a program extracting information
from a finger photograph and Figure 4 shows a fingerprint recognition device being used.
Fig. 3. Human fingerprint analysis. (Image in Public Domain)
Fig. 4. Biometric reading device. (Image in Public Domain)

Other applications involve various methods of obtaining valuable data from several image
sources, such as satellites or other sensors in order to discover important characteristics.
Several software products such as Photoshop (trademark of Adobe Systems Incorporated)
and GIMP (trademark of The GIMP Development Team) rely on common Image-Processing
techniques to alter or improve the quality of images. An example of this is High Dynamic
Range imaging - a method that blends the information from several differently exposed digital
photographs in order to obtain a better view of the scene. An excellent source of information
on this topic can be found at (Max Planck Institut fur Informatik, n.d.). See Figure 5 below for
an example.
Image Processing can be used to allow cars and other machines to operate automatically by
interpreting the information of a video-camera and determining the shapes or movement of
objects on the visible scene. An example of a new technology that involves heavy use of
Image-Processing in this way is the new Kinect gaming system developed by Microsoft for the
Xbox360 console. This device is comprised of three cameras, two of which serve the purpose
of analyzing the distance of objects on the scene from the device using an infrared laser. These
images are then processed so that the system is able to separate the location of the players
from the background or other objects in a process called segmentation (Shotton et al., 2011).
An image of this device can be seen on Figure 6. An example of image segmentation can be
seen in Figure 7, where the frog is separated from the background.
These and other applications show the importance of Image Processing as a field of research.
A good overview of the whole field of Image Processing can be found in (Velho et al., 2008).
138
Fig. 5. An example of HDR creation from multiple differently-exposed images. (Source

exposures by Grzegorz Krawczyk)
Fig. 6. The Kinect gaming device. (Image in Public Domain)
Fig. 7. An example of image segmentation. The frog is being segmented from the
background.
Some more information on interesting applications of this field and otherwise can be found in
(Acharya & Ray, 2005).
2. A quick glance at wavelet transforms applied to edge detection

The first mention of wavelets appeared in (Haar, 1911). But only in the 1980s did Stephane
Mallat (Mallat, 1989) spearheaded the use of wavelets in his work with digital image
processing. Inspired by this work, Yves Meyer (Meyer, 1987) constructed the first non-trivial
wavelets, which were differentiable, unlike Haar wavelets. They did not, however, have
compact support. A few years later, Ingrid Daubechies (Daubechies, 1988) used the works
of Mallat to construct a set of orthonormal bases of wavelets with compact support. These
works of Daubechies are the foundation of the current use of wavelets in Image Processing.
More historical information on wavelets can be found in (Daubechies, 1992).
Discrete Wavelets
There are plenty of uses of wavelets in image processing. For example, in 1994 (Fröhlich &
Weickert, 1994) presented an algorithm to solve a nonlinear diffusion equation in a wavelet
basis. This equation has the property of edge enhancement, an important feature for image
processing. More applications in edge detection are shown later in this chapter. The JPEG
2000 image coding system (from the Joint Photographic Experts Group) uses compression
techniques based on wavelets. In (Walker, 2003) the author describes a wavelet-based
technique for image denoising. Applications of the wavelet transform to detect cracks in
frame structures is presented by (Ovanesova & Suárez, 2004). Wavelet transforms have an
important role in multiresolution representations in order to effectively analyze the content of
images. Multiresolution will be introduced later in this chapter.
2.1 Wavelet Transforms

While the Fourier transform decomposes a signal over sine functions with different
frequencies, the wavelet transform decomposes a signal onto translated and dilated versions
of a wavelet. Figure 8 shows both a sine wave for the Fourier transform and a wavelet for
wavelet transform.
Fig. 8. A sine wave and a wavelet (image from (Radunivić, 2009))
Unlike the Fourier transform, the wavelet transform can capture both frequency and location
information.
A wavelet is a function ψ ∈ L2 (R ) with a zero average:
+∞
ψ(t)dt = 0 (1)
−∞
This function is normalized ψ = 1, and centered in the neighborhood of t = 0. A family of
time-frequency atoms is obtained by scaling ψ by s and translating it by u:

1 t−u
ψu,s (t) = √ ψ (2)
s s
Thus, the Continuous Wavelet Transform (CWT) of a function f at a scale s > 0 and translated
by u ∈ R can be written as:
+∞
1 t−u
W f (u, s) = f (t) √ ψ∗ dt (3)
−∞ s s
In the field of image processing we are interested in wavelets which form a base of L2 (R2 )
to represent images. If we have an orthonormal wavelet basis in L2 (R ) given by ψ with the
scaling function φ, we can use
ψ1 ( x1 , x2 ) = φ ( x1 ) ψ ( x2 ),
ψ2 ( x1 , x2 ) = ψ ( x1 ) φ ( x2 ), (4)
ψ3 ( x1 , x2 ) = ψ ( x1 ) ψ ( x2 ),
140
to form an orthonormal basis in L2 (R2 ) (Mallat, 1999):

ψ1j,p , ψ2j,p , ψ3j,p (5)
[ j,p] ∈ Z3
where ψ1 corresponds to variations along rows, ψ2 corresponds to variations along columns

and ψ3 corresponds to variations along diagonals.
It is computationally impossible to analyze a signal using all wavelet coefficients. Thus, for
discrete computations, we have to use a Discrete Wavelet Transform (DWT), that is a wavelet
transform for which the wavelets are discretely sampled (Mallat, 1999).
Let f [n] = f (n) be the discrete signal of size N. Its discrete wavelet transform is computed at
scales s = a j . A discrete wavelet scaled by a j is defined by:

1 n
ψj [n] = √ ψ j
(6)
a j a
The DWT can then be written as a circular convolution ψ̄j [n] = ψ∗j [n]:
N −1
W f (n, a j ) = ∑ f [m]ψ∗j [m − n] = f ψ̄j [n] (7)
m =0
A wavelet transform computed up to a scale a J is not a complete signal representation (Mallat,

1999). We need to add the low frequencies L f [n, d] corresponding to scales larger than d. A
discrete and periodic scaling filter is computed by sampling the scaling function φ(t) defined
by:
1 n
φ J [n] = √ φ J (8)
a J a
Let φ̄j [n] = φ∗j [n]. The low frequencies are carried by:
N −1
L f [n, a J ] = ∑ f [m]φ∗J [m − n] = f φ̄ J [n] (9)
m =0
As we can see in the Equations 6 and 9, the DWT is a circular convolution. In that way, we will
have lowpass and highpass filters which form a bank of filters. Figure 9 shows the discrete
wavelet transform for 3 scales. hψ (n) is a highpass filter and hφ is a lowpass filter. This form
is known as Fast Wavelet Transform (FWT).
Fig. 9. Fast Wavelet Transform for 1 dimension Mallat (1999)

As we saw before, in the field of image processing we are interested in two dimensional
signals. For two dimensions, the DWT of a function f ( x1 , x2 ) of size M × N can be written as:
Discrete Wavelets
−1 N −1
Wφ ( j0 , m, n) = √ 1 ∑ xM1 = 0 ∑ x2 =0 f ( x1 , x2 ) φ j0 ,m,n ( x1 , x2 )
MN
(10)
−1 N −1
Wψi ( j, m, n) = √ 1 ∑ xM1 = 0 ∑ x2 =0 f ( x1 , x2 ) ψ j,m,n ( x1 , x2 )
i
MN
where i = {1, 2, 3}
Similar to Figure 9, we can express the FWT in two dimensions like the Figure 10.
Fig. 10. Fast Wavelet Transform for 2 dimensions Mallat (1999)

For more information on the theory of multiresolution and high-frequency in images, read
Section 3.
3. Multiresolution and high frequency in images

Multiresolution Analysis is a very efficient way to process different levels of detail in an image.
Detecting and assessing discontinuities of an image allows one to detect its borders, edges and
peaks.
3.1 What are high-frequencies?

An image is composed by the sum of its components of low and high frequencies.
Low frequencies are responsible for the general smooth areas, while high frequencies are
responsible for details, like edges and noise Gonzalez & Woods (2006).
A filter that attenuates high frequencies is called a lowpass filter. A filter that has the opposite
characteristic, i.e., highlights high frequencies, is called highpass filter. As we saw on previous
sections, in a Discrete Wavelet Transform we have a filter hφ that corresponds a lowpass filter
and a filter hψ that corresponds a highpass filter.
The Figure 11 shows an example of applying a lowpass filter and a highpass filter on a image.
Therefore, high frequencies on images can be used for several applications which need the
details of an image, such as detecting edges, corners and textures.
3.2 Multiresolution analysis

A multiresolution analysis of the space L2 (R ) consists of a sequence of nested subspaces such
that
{0} · · · ⊂ V0 ⊂ V1 ⊂ · · · ⊂ Vn ⊂ Vn+1 ⊂ · · · ⊂ L2 (R )
with some important properties. The most important characteristics that we consider in the
context of image processing for high frequency assessment are:
142
Fig. 11. Results of lowpass and highpass filters. The first image is the original, the second is
the result of a lowpass filter and the third is the result of a highpass filter
• Regularity
The subspace V0 is generated as the linear combination of integer shifts of one or a
finite number of generating functions φ1 ,. . . ,φr . These generating functions are called
scaling functions. Usually those functions must have compact support and be piecewise
continuous.
• Completeness
those nested subspaces fill the whole space L2 (R ), and they are not too redundant. So, the
intersection of these subspaces should only contain the zero element.
This concept, applied to image processing and wavelets, justifies the successful use of image
pyramids in the context of high frequency detection.
3.3 Image pyramids

A simple, but powerful, structure to represent images at more than one resolution is the image
pyramid Burt & Adelson (1983). Basically, it is a collection of decreasing resolution images
arranged in the shape of a pyramid (Figure 12 ).

Fig. 12. Image Pyramid.

The idea behind image pyramids is to generate a number of images corresponding to the
response of a bank of filters at different scales. There are many different types of filters that
can be used for this purpose.
One special family of filters consists of Wavelets. They are constructed from a mother wavelet.
A family is constructed by dilating and translating the mother wavelet by different quantities.
The main advantage of using this family of functions over the Fourier transform is that
wavelets respond very well to discontinuities, i.e, high frequencies. The most know wavelet
families are the Haar, Daubechies, Coiflet, and Symmlet.
Discrete Wavelets
The Daubechies family is of particular interest because it is fractal in nature, and the Haar
family, although very simple, can be very useful in many applications.
In practical terms, the base of the pyramid is the image which we want to filter in various
scales, and each level of the pyramid above the base is produced by filtering it and generating
an image with half of its width and height.
Using wavelet and scale functions, the nested subspaces of scale and detail are produced. The
horizontal, vertical and diagonal details of a subspace Vi+1 are the information that cannot be
represented in Vi (Figure 13).
⊂ ⊂ ⊂
Fig. 13. Nested subspaces in the context of image processing. The details are represented in
the grey regions (the contrast was enhanced for better visualization).
Now is easy to understand how the discrete wavelet transform can be applied for images.
As we saw in Equation 10, the Discrete Wavelet Transform in two dimensions captures the
variations on rows, columns and diagonals. Figure 14 shows an example of a DWT applied
for an image in 3 scales.
Fig. 14. The result decomposition of a blank image using a discrete Wavelet Transform for 1
and 2 scales Gonzalez & Woods (2006)
Section 4.2 describes a method which produces a pyramid of a chosen image and processes the
correspondent details in every scale. This allows us to detect discontinuities in a very precise
and adaptative approach.
144
3.4 Edge detectors using multiresolution and discrete wavelet transform

Many works use multiresolution as a step to gather specific image information on a single
scale. The idea is to combine the information present at several scales, as appearance is related
to the scale of the observation, so a scene should be described at multiple scales.
The first ones to formalize this concept were Witkin (1983) and Koenderink (1984) with the
idea of scale-space linear filtering. The principle is to convolve the original image by a family
of Gaussians of increasing variance related to the studied scale, and then to progressively
eliminate the smallest structures in the image.
However, this approach suffers from several drawbacks such as blurred edges and the edges
at the coarse scale are shifted. Multiple nonlinear diffusion filters have been suggested to
overcome these drawbacks. More elaborated approaches have been suggested to accelerate
the resolution, such as wavelet-based ones.
Recent works still use the idea of convolution by a family of Gaussians (Sumengen &
Manjunath (2005), Zhang et al. (2010)) and nonlinear diffusion filters(Tremblais & Augereau
(2004)). Other works are wavelet-based, as can be seen in (Belkasim et al., 2007), (Shih &
Tseng, 2005), (Han & Shi, 2007), Brannock & Weeks (2006) and Heric & Zazula (2007).
Sumengen & Manjunath (2005) create an Edgeflow vector field where the vector flow is
oriented towards the borders at either side of the boundary. To create this vector field, they
use a fine to coarse strategy. In that way, the proposed method favors edges that exist at
multiple scales and suppress edges that only exist at finer scales. The strength of the edges are
represented by the strength of the vectors at the edge location where the vector field changes
its direction. This method is also used to multi-scale image segmentation.
Tremblais & Augereau (2004) present a new explicit scheme to the linear diffusion filtering
which preserves edges. A fast filtering algorithm is then combined with a simple multiscale
edge detection algorithm.
For Zhang et al. (2010), the one pixel width edge is more accurate than other edge detection.
So, they explore the zero-crossing edge detection method based on the scale-space theory.
For image segmentation, Belkasim (Belkasim et al., 2007) uses a wavelet-based image analysis
scheme based on extracting all objects in the image using their borders or contours. The size
of the contour can then be used to define the level of resolution and hence the extent of the
analysis.
Shih (Shih & Tseng, 2005) argue that edge extraction based only on a gradient image will
produce a bad result with noise and broken edges. In order to solve this problem, they
combine an edge detector with a multiscale edge tracker based on the discrete wavelet
transform.
In Han (Han & Shi, 2007), the wavelet transform plays an important role in the task of
decomposing a texture image into several levels. Once a decomposition level is chosen,
textures are then removed from the original image by the reconstruction of low frequencies
only.
The problem for Brannock & Weeks (2006) is to automatically detect edges. To determine
its efficacy, the 2D discrete wavelet transform is compared to other common edge
detection methods. They conclude that the discrete wavelet transform is a very successful
edge-detection technique, especially when utilizing auto-correlation.
Heric & Zazula (2007) present a novel edge detector based on the wavelet transform and signal
registration. The proposed method provides an edge image by a time-scale plane based edge
detection using a Haar wavelet. Then, this edge image is used in a registration procedure
in order to close the edge discontinuities and calculate a confidence index for the detected
contour points.
Discrete Wavelets
11
4. The DWT applied to high-frequency assessment from multiresolution analysis

In this section, we present a practical use of wavelets for visualization of high frequency
regions of a multiresolution image. Our approach combines both multiresolution analysis
and orientation tensor to give a scalar field representing multiresolution edges. Local maxima
of this scalar space indicate regions having coincident detail vectors in multiple scales of
wavelet decomposition. This is useful for finding edges, textures, collinear structures and
salient regions for computer vision methods. The image is decomposed into several scales
using the DWT. The resulting detail spaces form vectors indicating intensity variations which
are adequately combined using orientation tensors. A high frequency scalar descriptor is then
obtained from the resulting tensor for each original image pixel.
4.1 Orientation tensor

One way of estimating salient regions in image processing is to use multiresolution to capture
global and local brightness variations. Even in a non-redundant wavelet decomposition, local
and global borders occurring in the same region may carry useful information. The problem
lies in combining this global information into a single image. In this way, we can capture
the multivariate information of several scales and color channels using orientation tensors
(Knutsson, 1989).
A local orientation tensor is a special case of non-negative symmetric rank 2 tensor, built
based on information gathered from an image. As shown by Knutsson (Knutsson, 1989), one
can be produced by combining outputs from polar separable quadrature filters. Because of
its construction, such a tensor has special properties and contains valuable information about
said image.
From definition given by Westin (Westin, 1994), orientation tensors are symmetric, and thus
an orientation tensor T can be decomposed using the Spectral Theorem as shown in Equation
11, where λi are the eigenvalues of T.
n
T= ∑ λi Ti (11)
i =1
If Ti projects onto a m-dimensional eigenspace, we may decompose it as
m
Ti = ∑ es esT (12)
s =1
where {e1 , ..., em } is a base of R m . An interesting decomposition of the orientation tensor T
(Westin, 1994) is given by
n −1
T = λn Tn + ∑ (λi − λi+1 ) Ti (13)
i =1
where λi are the eigenvalues corresponding to each eigenvector ei . This is an interesting
decomposition because of its geometric interpretation. In fact, in R3 , an orientation tensor T
decomposed using Equation 13 can be represented by a spear (its main orientation), a plate
and a ball
T = (λ1 − λ2 ) T1 + (λ2 − λ3 ) T2 + λ3 T3 (14)

A R3 tensor decomposed by Equation 14, with eigenvalues λ1 ≥ λ2 ≥ λ3 , can be interpreted
as following:
• λ1 ≫ λ2 ≈ λ3 corresponds to an approximately linear tensor, with the spear component
being dominant.
146
• λ1 ≈ λ2 ≫ λ3 corresponds to an approximately planar tensor, with the plate component

being dominant.
• λ1 ≈ λ2 ≈ λ3 corresponds to an approximately isotropic tensor, with the ball component
being dominant, and no main orientation present
Consider two orientation tensors A and B and its summation T = A + B. After the
decomposition of T using Equation 14, the component (λ1 − λ2 ) T1 is an estimate of the
collinearity of the main eigenvectors of A and B.
4.2 Proposed method

The method proposed in (de Castro et al., 2009) uses high frequency information extracted
from wavelet analysis. Given an input image I, for each scale j and position p ∈ I, we create
a vector v j,p as follow:
v j,p = [ I · ψ1j,p , I · ψ2j,p , I · ψ3j,p ] T (15)

This vector contains the high frequency value at vertical, horizontal and diagonal directions
of the image I at the position p and scale j. Simmetric rank 2 tensors are then created as
M j,p = v j,p v Tj,p (16)

We find the final tensor M0,p for each pixel of the original image using
nj
M0,p = ∑ k j Mj,p (17)
j =1
to combine the tensors obtained at each scale j, where n j is the number of scales and k j ∈ R is
the weight assigned to each scale, given by
np

∑n=1 Trace M j,n
k j = nj np
(18)
∑k=1 ∑n=1 Trace Mk,n
where n p is the number of pixels and Trace( M j,p ) is the sum of the eigenvalues of M j,p .
The trace represents the amplification driven by the tensor to the unit sphere and is a good
estimator of its importance. Thus, the tensor sum is weighted by the proportion of energy of
each scale in the multiresolution pyramid.
In order to find M j,p in Equation 17, we use bilinear interpolation of the tensor values, relative
to each position p in the initial image, at the subsampled image at scale j to find the resulting
tensor M j,p for each pixel of the initial image. This is depicted in Figure 15, where tensors are
represented as superquadric glyphs whose longer axis shows the main direction.
Note that the tensor presented in Equation 17 is a 3x3 positive symmetric matrix with
real coefficients, and thus we may apply Equation 14. We then find the main orientation
component (spear) of the final orientation tensor for each pixel of the input image. This
component indicates the collinearity of the interpolated tensors and provides interesting
results.
4.2.1 Implementation
The proposed algorithm consists of three main steps: a discrete wavelet transform (Barnard,
1994; Mallat, 1999), a tensor field computation and a weighted sum of the computed tensors.
The whole process is illustrated in Figure 16.
Discrete Wavelets
13
Fig. 15. A tensor is computed for each pixel in original image by a weighted sum of
corresponding tensors in each scale. In this example, two wavelet decompositions are
performed.
Fig. 16. Example of the proposed algorithm using Daubechies1 to decompose the image into
two scales.
The number of scales to be used is a parameter of the algorithm. The DWT splits the image
into three detail components and one scale component in the beginning of each iteration. In
the next iteration, the same process is applied, using the resulting scale component as the
input image.
For each pixel of the input image, its correspondent position at the current scale is computed
with subpixel precision for each resolution. The four nearest pixels in a given resolution are
used to compute the final tensor. The vectors v j,p described in Equation 15 are computed for
each of these pixels and then used to compute four spear type tensors. The final tensor for the
subpixel position is obtained by combining these four tensors with bilinear interpolation. The
pixel tensor is computed by combining the n j tensors as showed in Equation 17.
The pixel tensors are decomposed and their eigenvalues are then extracted. The values
λ1 − λ2 are computed and normalized to form the output image. Color images are split
148
into three monochromatic channels (Red, Green and Blue) and the proposed algorithm is
applied to each channel separately. The tensors for each color channel are summed before
eigen decomposition.
The complexity of the whole process is O(n j · n p ), where n j is the number of analyzed scales
and n p the amount of input pixels. Thus, this is an efficient method that can be further
parallelized.
4.2.2 Experimental results

The proposed method was tested with several images and using several wavelets functions
(de Castro et al., 2009). A piece of the experiments is shown in Figure 17. The DWT is applied
with different analyzing Daubechies filters and number of scales. The Church’s ceiling is
formed by coincident frequencies on its geometric details. These details can be better observed
in Figure 17c.
(a)
(b) (c)
(d) (e)
Fig. 17. (a) input image. (b) λ1 − λ2 with Daubechies1 and 1 scale. (c) Daubechies1 and 3
scales. (d) Daubechies3 and 1 scale. (e) Daubechies3 and 3 scales.
Discrete Wavelets
15
A better estimation of soft edge transitions is obtained by changing the analyzing filter from
Daubechies1 to Daubechies3. Figures 17b and 17d illustrate this behavior.
In general, it can be noted that this method highlights high frequencies occurring in the same
region at different scales. We used thermal coloring with smooth transition from blue to red,
where blue means absence of high frequencies, and red means presence of high frequencies.
The green regions also indicate high frequencies, but less intense than those indicated by red
regions. The red regions provide the better higher frequencies estimation tensors.
4.2.3 Conclusion
We presented an overview of discrete wavelets and multiresolution applied to edge detection.
We also presented a method for high frequency assessment visualization using these powerful
tools. The method is based on the DWT decomposition and detail information merging
using orientation tensors. This multiresolution analysis showed to be suitable for detecting
edges and salient areas in an image. The experimental results show that the high frequency
information can be inferred by varying the DWT filters and number of scales. Coincident
frequencies in space domain are successfully highlighted. By tuning the number of scales, one
may infer texture feature regions. The λ1 − λ2 scalar field is one of the most used orientation
alignment descriptors. However, other relations can be extracted from final tensors. This
method can be easily parallelized, the use of technologies like GPGPUs and multicore CPUs
turns it attractive for high performance applications.
5. References
Acharya, T. & Ray, A. K. (2005). Image Processing - Principles and Applications, First Edition,
Wiley InterScience.
Barnard, H. J. (1994). Image and Video Coding Using a Wavelet Decomposition, PhD thesis, Delft
University of Technology, Department of Electrical Engineering, Information Theory
Group, P.O.Box 5031, 2600 GA, Delft.
Belkasim, S., Derado, G., Aznita, R., Gilbert, E. & O’Connell, H. (2007). Multi-resolution
border segmentation for measuring spatial heterogeneity of mixed population
biofilm bacteria, Computerized Medical Imaging and Graphics 32.
Brannock, E. & Weeks, M. (2006). Edge detection using wavelets, Proceedings of the 44th annual
Southeast regional conference, ACM-SE 44, ACM, New York, NY, USA, pp. 649–654.
Burt, P. J. & Adelson, E. H. (1983). The laplacian pyramid as a compact image code, IEEE
Transactions on Communications 31: 532–540.
Daubechies, I. (1988). Orthonormal bases of compactly supported wavelets, Communications
on Pure and Applied Mathematics 41(7): 909–996.
Daubechies, I. (1992). Ten lectures on wavelets, Society for Industrial and Applied Mathematics,
Philadelphia, PA, USA.
de Castro, T. K., de A. Perez, E., Mota, V. F., Chapiro, A., Vieira, M. B. & Freire, W. P. (2009).
High frequency assessment from multiresolution analysis., ICCS (1), Vol. 5544 of
Lecture Notes in Computer Science, Springer, pp. 429–438.
Fröhlich, J. & Weickert, J. (1994). Image processing using a wavelet algorithm for nonlinear
diffusion.
Gonzalez, R. C. & Woods, R. E. (2006). Digital Image Processing (3rd Edition), Prentice-Hall, Inc.,
Upper Saddle River, NJ, USA.
Haar, A. (1911). Zur theorie der orthogonalen funktionensysteme, Mathematische Annalen
71: 38–53. 10.1007/BF01456927.
Han, Y. & Shi, P. (2007). An adaptive level-selecting wavelet transform for texture defect
detection, Image Vision Comput. 25(8): 1239–1248.
150
Heric, D. & Zazula, D. (2007). Combined edge detection using wavelet transform and signal
registration, Image Vision Comput. 25: 652–662.
Knutsson, H. (1989). Representing local structure using tensors, The 6th Scandinavian
Conference on Image Analysis, Oulu, Finland, pp. 244–251. Report LiTH–ISY–I–1019,
Computer Vision Laboratory, Linköping University, Sweden, 1989.
Koenderink, J. J. (1984). The structure of images, Biological Cybernetics 50(5): 363–370–370.
Mallat, S. (1999). A Wavelet Tour of Signal Processing, Second Edition (Wavelet Analysis & Its
Applications), Academic Press.
Mallat, S. G. (1989). A theory for multiresolution signal decomposition: the wavelet
representation, IEEE Transactions on Pattern Analysis and Machine Intelligence
11: 674–693.
Max Planck Institut fur Informatik, M. (n.d.). http://www.mpi-inf.mpg.de/resources/hdr/.
Meyer, Y. (1987). Principe d’incertitude, bases hilbertiennes et algèbres d’opérateurs. (The
uncertainty principle, Hilbert base and operator algebras)., Sémin. Bourbaki, 38ème
année, Vol. 1985/86, Exp. Astérisque 145/146, 209-223 (1987).
Ovanesova, A. V. & Suárez, L. E. (2004). Applications of wavelet transforms to damage
detection in frame structures, Engineering Structures 26(1): 39 – 49.
Radunivić, D. (2009). WAVELETS from MATH to PRACTICE, Springer.
Shih, M. & Tseng, D. (2005). A wavelet-based multiresolution edge detection and tracking,
IVC 23(4): 441–451.
Shotton, J., Fitzgibbon, A., Cook, M., Sharp, T., Finocchio, M., Moore, R., Kipman, A. & Blake,
A. (2011). Real-time human pose recognition in parts from a single depth image.
Sumengen, B. & Manjunath, B. S. (2005). Multi-scale edge detection and image segmentation,
European Signal Processing Conference (EUSIPCO).
Tremblais, B. & Augereau, B. (2004). A fast multi-scale edge detection algorithm, Pattern
Recogn. Lett. 25: 603–618.
Velho, L., Frery, A. & Gomes, J. (2008). Image Processing for Computer Graphics and Vision, First
Edition, Springer.
Walker, J. S. (2003). Tree-adapted wavelet shrinkage, Vol. 124 of Advances in Imaging and
Electron Physics, Elsevier, pp. 343 – 394.
Westin, C.-F. (1994). A Tensor Framework for Multidimensional Signal Processing, PhD thesis,
Department of Electrical Engineering Linköping University.
Witkin, A. P. (1983). Scale-Space Filtering., 8th Int. Joint Conf. Artificial Intelligence, Vol. 2,
Karlsruhe, pp. 1019–1022.
Zhang, M., Li, X., Yang, Z. & Yang, Y. (2010). A novel zero-crossing edge detection
method based on multi-scale space theory, Signal Processing (ICSP), 2010 IEEE 10th
International Conference on, pp. 1036 –1039.
10
Discrete Wavelet Transform and

Optimal Spectral Transform Applied to
Multicomponent Image Coding
Isidore Paul Akam Bita1, Michel Barret2, Florio Dalla Vedova1,
Jean-Louis Gutzwiller2 and, Dinh-Tuan Pham3
1LUXSPACE Sarl, Chateau de Betzdorf, Betzdorf
2SUPELEC, Information Multimodality and Signal Team, 2, rue E. Belin, Metz
3Jean Kuntzmann Laboratory, 51 rue des Mathématiques, Grenoble Cedex 9
1Luxembourg
2,3France
1. Introduction
These last years, research activities on multicomponent image compression have been
expanded, due to the development of multispectral and hyperspectral image sensors which
supply larger and larger amount of data. The end-users of such images become also more
numerous and have various needs and various applications. The future earth observation
systems, for instance, will use multi-, super- and hyperspectral image sensors with higher
resolutions leading to bigger amount of transmitted data. However the channel bandwidth for
transmission is limited and therefore there is an interest of conceiving compression systems
(onboard and on the ground) of multicomponent images which are not application dependent
and which are compatible with the diversity of end-users’ needs. The components of a
multicomponent image generally represent the same scene with different views depending
on the wavelength. For data from different sensors, a preliminary step of image registration
is therefore required as there is a high degree of dependence (or redundancies) between
the various components: the usual spatial redundancy (between different pixels in each
component) and the spectral redundancy (between the components).
During the past two decades, different solutions have been proposed for multicomponent
image coding. A solution currently adopted consists of using two different transformations,
each with the goal of reducing only one of the two redundancies. In (Dragotti et al., 2000),
a 2-D discrete wavelet transform (DWT) is used to reduce the spatial redundancies in each
component while the Karhunen Loève transform (KLT) is applied to reduce the spectral
ones. In that paper, the quantization and entropy coding are achieved thanks to the well
known SPIHT (Set Partitioning in Hierarchical Trees) codec by Said and Pearlman (Said
& Pearlman, 1996) in its original version and in a modified version including VQ (vector
quantization). In the same way, with the use of the 2-D DWT of (Antonini et al., 1992)
(usually called the Daubechies 9/7), the authors of (Vaisey et al., 1998) use a lattice VQ
with a stack run coder as quantization and entropy coding. More recently in (Rucker et al.,
2005), the KLT associated with the Daubechies 9/7 2-D DWT and with EBCOT (Taubman,
2000; Taubman & Marcellin, 2002) for quantizing and entropy coding has been tested on
152
hyperspectral images with different bit-allocations between components. It is shown that the
Post Compression Rate-Distortion (PCRD) optimizer of EBCOT applied across multiple bands
gives the best rate-distortion performance. Another solution consists of using a 3-D DWT for
reducing both the spatial and spectral redundancies with only one transform. This approach
is generally applied to hyperspectral images as in (Christophe et al., 2006). An overview of
3-D wavelet-based techniques and more can be found in (Fowler & Rucker, 2007). The two
above mentioned solutions are compatible with the JPEG2000 Part 2 standard. The JPEG2000
standard is well known and well spread today. Moreover the KLT used in JPEG2000 Part 2
is considered as the best existing lossy compression techniques for hyperspectral images at
medium and high bit rates (Du & Fowler, 2007; Penna et al., 2007). The KLT consists in a
Principal Component Analysis (PCA), well known of statisticians, where all the components
are kept. However, the rather great computational complexity of the KLT hinders its adoption
in practice — specially on satellite platforms — and recent works propose different solutions
in order to pass round this problem. One approach consists in reducing the complexity of
the covariance matrix computation. This is done by randomly sampling the entire image in
order to obtain a small sample of the pixels’ population on which the covariance matrix is
computed (Du & Fowler, 2008; Penna et al., 2007). Another approach consists in computing
a kind of KLT average on a set of images (the learning basis) issued from only one sensor
and using it on other images obtained with the same sensor. This sub-optimal transform
is called exogenous KLT in (Thiebaut et al., 2006) and the computational complexity of the
second approach is compatible with satellite platforms. Both approaches are fruitful: the
rate-distortion performance sacrifice compared with the true KLT is very slight, whereas the
computational burden is significantly reduced. In the second approach, the exogenous KLT
matrix is known by the decoder, hence there is no need to transmit it.
It is well known that the KLT can be suboptimal in transform coding when the data are not
Gaussian. Now, under only the high resolution quantization hypothesis, nearly everything
is known about the performance of a transform coding. Nevertheless, the optimal transform
computation is generally considered as a difficult task and the Gaussian assumption is then
used in order to simplify the calculation. Recently, the problem of computing the optimal
coding transform associated with scalar variable-rate quantizers for still images was resolved
under high-resolution quantization hypothesis, with mean square error as distortion and
without the Gaussian assumption (Narozny et al., 2005; 2008). However, for the JPEG2000
Part2 compression scheme, the previous optimal transform computation cannot be directly
applied to obtain the optimal spectral transform, because of the 2D DWT presence—see the
criterion (15) in Section 4, which depends on subband statistics—. In (Akam Bita et al.,
2010a), the authors solved both the problems of computing an optimal spectral transform
(OST), with the constraint of orthogonality and without any constraint but invertibility, for
that compression scheme, when the 2D DWT has fixed coefficients and under the only
high resolution quantization hypothesis. They showed that on hyperspectral images, the
orthogonal OST, called OrthOST, performs slightly but significantly better than a KLT at
low, medium and high bit-rates and that the gain obtained by removing the orthogonality
constraint in the computation of the OST is not significant. Further, it is not widely
known that even when the input data are Gaussian, the KLT is not optimal in the above
mentioned compression scheme. Indeed, after the 2D DWT, the variance of the wavelet
coefficients depends on the subband they belong to (even for Gaussian data) and the KLT
does not capture these various variances, while the EBCOT coder with its PCRD optimizer
performing simultaneously across all the codeblocks from the entire image take them into
Discrete Wavelet Transform
and
DiscreteOptimal Spectral
Wavelet Transform Transform
and Optimal Applied
Spectral Transform toMulticomponent
Applied to MulticomponentImage CodingImage Coding 1533
account. In (Akam Bita et al., 2010b), the authors introduced an orthogonal spectral transform
(called JADO for Joint Approximate Diagonalization under Orthogonality constraint) using
only second order statistics that has not this shortcoming, and that is optimal at high
bit-rates for the JPEG2000 Part 2 compression scheme, when the data are Gaussian. They
showed on natural hyperspectral images that JADO (resp. OrthOST) performs slightly but
significantly better than the KLT (resp. JADO). The main drawback of the OSTs is their heavy
computational cost, which is much higher than the one of a KLT or JADO (which both have
roughly the same complexity).
In order to reduce the complexity of a codec based on OrthOSTs, the authors of (Akam Bita
et al., 2008; 2010c; Barret et al., 2009) used the same strategy as in (Thiebaut et al., 2006):
they replaced the OrthOST, which must be computed for each new encoded image, with an
exogenous quasi optimal spectral transform. This last transform is an OrthOST computed once
and for all on a learning basis constituted of images from only one spectrometer and which
is then applied to any image to be coded stemming from the same spectrometer. Using either
the JPEG2000 codec called Verification Model version 9 (JPEG2000, 2001) or the Bit Plane
Encoder (BPE (CCSDS-1, 2007)) recommended for satellite image compression (Yeh et al.,
2005) by the CCSDS (Consultative Committee for Space Data Systems), they showed that this
strategy yielded good performances, sometimes better than the (non exogenous) KLT ones,
in terms of bit-rate versus distortions. Four different distortions were considered: Signal to
Noise Ratio (SNR), Maximum Absolute Difference (MAD), Mean Absolute Error (MAE) and
Maximum Spectral Angle (MSA). Indeed, it is well-known that providing the mean square
error as one distortion only is not sufficient to assess the quality of a codec for hyperspectral
images (Christophe et al., 2005). However in the simulations presented in (Akam Bita et al.,
2008; 2010c; Barret et al., 2009) when the VM9 is used, the computational complexity of the
EBCOT coder associated with its PCRD optimizer is very high, and when the BPE is applied
to encode each component of the transformed image, the complexity of the algorithm for
optimal allocation between components is also very high. In both cases, the computational
complexity is too high for a compression system on-board a satellite. In (Barret et al., 2011), the
authors present a low complexity hyperspectral image coder based on exogenous OrthOST
and zerotrees well adapted to OrthOST.
It is important to note that the point of view presented in this chapter — i.e., a compression
scheme for hyperspectral images that is independent of the end-user application — is no
longer justified at very low bit-rates (lower than 0.5 bits per pixel and per band). For more
details on low-bit rates hyperspectral compression see (Chang et al., 2010c).
In this chapter, we study the question of an optimal linear transform for reducing spectral
redundancies under high resolution and variable rate constrained quantization hypothesis,
when a 2-D DWT — with fixed coefficients — is applied to each component to reduce
spatial redundancies and one scalar quantizer per subband and per component is used.
This compression scheme, described in Section 2, is compatible with the JPEG2000 Part 2
standard. The asymptotic expression of the mean square error distortion associated with that
compression scheme is given in Section 3. In Section 4, we clarify the criterion minimized by
such an optimal spectral transform with mean square error distortion and we show the link
between the criterion and the mutual information contrast used in Independent Component
Analysis (ICA). In Section 5, we derive a criterion minimized by an OrthOST under Gaussian
data assumption. Moreover, we describe in Section 6 the quasi-Newton algorithms used for
the minimization of the criterion, either with the constraint of an orthogonal transform or
with no constraint but invertibility or with the constraint of an orthogonal transform and the
154
assumption of Gaussian data. The two first algorithms are derived from an algorithm by Pham
ICAinf described in (Pham, 2004) that performs ICA. Then in Section 7, performances of these
transforms and comparisons with the KLT are given for multi- and hyper-spectral satellite
images, with the four above mentioned different measures of distortion. Finally, in Section 8
we introduce quasi-optimal OrthOSTs, called exogenous, that have not the main drawback of
heavy computational cost and we compare their performances in lossy coding with OrthOSTs.
2. Description of the separable compression scheme

2.1 Conventions and notations
We consider a multicomponent image X with N components X1 , . . . , X N . Each component Xi
is a 2-D image with Nr rows and Nc columns. To simplify the notations and the mathematical
expressions, we assume that each component is written as a row vector by scanning all its
pixels row by row (for example). Then X is a N × L matrix, with L = Nr Nc . In the following,
depending on the context, we shall interpret Xi as a 2-D image or as a row vector of dimension
L. For a square matrix M, the expressions det M, tr M and diag(M) denote respectively its
determinant, its trace and the diagonal matrix obtained with its diagonal elements.
In the following compression scheme, the 2-D DWT has fixed coefficients (in our tests, the
Daubechies 9/7 DWT is always used), but the spectral linear transform is adapted to the data.
We denote W the invertible L × L matrix associated with the 2-D DWT.
2.2 The separable scheme

The separable scheme is compatible with the JPEG2000 Part 2 standard. It can be described as
follow:
• Coding. The same 2-D DWT is applied to each component Xi in order to reduce the spatial
redundancies and a linear transform A is applied between the components in order to
reduce the spectral redundancies. The result of the 2-D DWT applied to the entire image
X is XW T and the transformed coefficients are the elements of the matrix Y = AXW T . For
each component, the wavelet coefficients of each subband are regrouped according to a
fixed scan that does not depend on the component. This re-ordering corresponds to the
right multiplication of XW T by a permutation matrix P T . We can suppose without loss of
generality that P is the identity, otherwise we could replace W with PW. This partitioning
can be written XW T = [(XW T )(1) . . . (XW T )( M) ], where M is the number of subbands.
Then, the transformed coefficients Y = [Y(1) . . . Y( M) ] (where Y(i) = A(XW T )(i) ) are
quantized and entropy coded with one quantizer per subband and per component (see
§ 7.1).
• Decoding. Let Yq denote the matrix with the same dimension as Y containing the
dequantized transformed coefficients. The mathematical inverse transforms are applied
= A−1 Yq W−T of the original image X.
to Yq in order to reconstruct an approximation X
We can remark that the order of the transformations (i.e., applying first the DWT then A, or
first A then the DWT) has no effect on the result, since Y = A(XW T ) = (AX)W T . This is why
that scheme is called separable.
3. Expression of the distortion

In lossy or quasi-lossless coding, the quantization leads to irreversible loss of information,
therefore the decoded image X is an approximation of the original image and in order to
and
and Optimal Applied
quantify the quality of the reconstructed image, it is necessary to introduce a measure of

distortion. In this section we give, under various hypotheses, the relation that links the
distortion between X and X to the quantizers distortions, when the distortion is the mean
square error:
N L
) = 1 X − X
Da (X, X 2 with 2 =
X − X ∑ ∑ (Xi (k) − Xi (k))2 . (1)
NL i =1 k =1
We begin by recalling the solution of the problem in a simple general case (Gersho & Gray,
1992; Taubman & Marcellin, 2002).
3.1 A simple general case

Lemma 3.1. Let X be a real random vector with N components and A be an invertible matrix of order
N. The transformed vector Y = AX is quantized and dequantized in Yq . The original vector X is
= A−1 Yq and let b = Y − Yq be the quantization noise. Then, the end-to-end
approximated by X
distortion D = N E(X − X
1 2 ), where E denotes the mathematical expectation, satisfies the relation
− −

D = N tr E(bb )A A 1 .
1 T T
Proof: We have X − X = A−1 b and A−1 b2 = b T A−T A−1 b = tr[A−1 bb T A−T ] =

tr[bb A A ], therefore D = N1 E(A−1 b2 ) = N1 tr E(bb T )A−T A−1 .
T − T − 1
Further, we may need the following assumption, that can be deduced from high resolution
quantization hypothesis (Gersho & Gray, 1992) (this point is recalled in Subsection 3.2).
H1 : The components of the quantization noise are zero mean and uncorrelated.
Theorem 1. 1. With the hypotheses of Lemma 3.1 and assuming H1 , the distortion becomes
1 N
N i∑
D= α i Di , (2)
=1
where Di = E(bi2 ) is the quantizer distortion of the ith component Yi of Y and, with ei the ith
canonical vector of R N and (A−1 )ij the element of A−1 located on row i and column j, we have
N
αi = ∑ (A−1 )2ji = A−1 ei 2 . (3)
j =1
2. The assertion 1. holds without the assumption H1 if A−T A−1 is diagonal, e.g. if A is orthogonal.
Proof: The assumptions in 1. or 2. state that at least one of the two matrices A−TA−1 and
E(bb T ) is diagonal. Hence the trace of their product is equal to the sum of the products of
their diagonal elements.
3.2 Justification of the assumption H1

We recall here well known results that can be found e.g. in (Gersho & Gray, 1992). The
assumption H1 can be justified under the following conditions. C1 : the random vector
Y = (Y1 , . . . , YN ) T has a continuous probability density function (pdf) f Y ; C2 : separable
high-rate quantization is achieved, meaning that the quantization steps h = (hi )1≤i≤ N of
the N components are small with respect to the variations of f Y (i.e. f Y (y + h) f Y (y),
∀y ∈ R N ) and C3 : for any cell S of the separable N-D quantizer, the dequantized value Yq
156
associated with S is the iso-barycenter of S. Indeed, if the three conditions C1 , C2 and C3 hold,
then the pdf f Y can be considered as quasi constant in the hypercube Yq + ∏iN=1 [− hi /2 , hi /2].
Further, the conditional law of the quantization noise b = Y − Yq knowing the dequantized
value Yq satisfies f b|Yq (u) 1/ ∏iN=1 hi if u ∈ ∏iN=1 [− hi /2 , hi /2], 0 otherwise. We see that
the conditional pdf f b|Yq does not depend on the quantized value Yq , hence it is equal to f b ,
the pdf of b. Further the components of b are zero mean and (quasi) independent since their
joint density is approximatively equal to the product of their marginal densities.
3.3 The separable subband scheme

In the following, the symbols X, Y and Yq refer again to the matrices defined in Section 2 and
A denotes the matrix of the linear transform that associates Y with X. We are going to apply
the formulae of the general simplified case to the separable scheme. The actual distortion Da
given in relation (1) is an estimation of the distortion
) = E[ Da (X, X
)] = 1 2 ].
D (X, X E[X − X (4)
NL
Now, in order to express the relation (3) in terms of the DWT W and the spectral transform
A, it is important to note first that the canonical basis of the space of matrices of dimension

N × L is the family of matrices ei,k = ei ekT (1 ≤ i ≤ N, 1 ≤ k ≤ L), with ei (resp. ek ) the ith
(resp. kth ) vector of the canonical basis of R N (resp. R L ). Therefore, the weighting factor αi in
relation (3) depends here on the two indices i and k: αik = A−1 ei,k 2 . Then, let
wi = A −1 e i 2 (1 ≤ i ≤ N ), (5)
T T
we have A−1 ei,k = A−1 ei ek W−T and A−1 ei,k 2 = tr[A−1 ei ek W−T W−1 ek eiT A−T ] =

eiT A−T A−1 ei ekT W−T W−1 ek and finally

αik = A−1 ei,k 2 = wi W−1 ek 2 . (6)
Therefore, according to Theorem 1, under assumption H1 we have
N L
) = 1 ∑ ∑ wi W−T e 2 E[(Yi (k) − Y q (k))2 ].
D (X, X (7)
NL i=1 k=1 k i
Now, for any subband m (1 ≤ m ≤ M ), let Km be the number of columns in Y corresponding

to that subband and let
Km
πm = (1 ≤ m ≤ M ) (8)
L
be the ratio of wavelets coefficients that belong to it. If k (1 ≤ k ≤ L) refers to a column indice
of the matrix Y located in that subband and if we assume that
q (m)
H2 : for any component i (1 ≤ i ≤ N ), the distortion E[(Yi (k) − Yi (k))2 ] = Di does not
depend on the spatial position k in the subband m,
(which is the case under high resolution quantization hypothesis), then equation (7) becomes
1 N M (m) 1
) =
D (X, X
N i∑ ∑ π m w i ω m Di with ωm =
Km ∑ W − T e k 2 , (9)
=1 m =1 k
where in the last summation, the range of k consists in the columns of Y with the subband m.
Or, by adopting a different perspective, if we assume that
and
and Optimal Applied
H3 : the weight W−1 ek 2 = ωm does not depend on the spatial position k in the subband m,
then equation (7) becomes
) = 1 N M (m) (m) 1 q
N i∑ ∑ π m ω m w i Di ∑ E[(Yi (k) − Yi (k))2 ],
D (X, X with Di = (10)
=1 m =1
Km k
where in the last summation, the range of k consists in the columns of Y with the subband m.
Remark 1. The condition H3 is satisfied by dyadic wavelets having Finite Impulse Response (FIR)
synthesis filters, when edge effects are neglected (for more details see e.g. (Usevitch, 1996; Woods &
Naven, 1992)).
Lastly, we can notice that the actual distortion Da given in equation (1) satisfies
) = 1 tr[(X − X
Da (X, X ) T ] = 1 tr[A−1 (Y − Yq )W−T W−1 (Y − Yq ) T A−T ],
)(X − X
NL NL
therefore if we assume
H4 : the DWT is orthogonal, i.e. WW T = I L , with I L the identity matrix of dimension L,
then Da (X, X ) = tr[A−1 (Y − Yq )(Y − Yq ) T A−T ] = −1 (m) − Yq(m) )( Y(m) −
=1 tr[ A ( Y
1 1 M
NL NL ∑m
Y q ( m ) T −
) A ].
T
Remark 2. The hypothesis H4 is roughly satisfied with the approximately orthogonal Daubechies
9/7 DWT (indeed, a simulation shows that the infinity norm of the diagonal, and respectively the off
diagonal, elements of W T W − I L is worth 0.42 and 0.16, for five levels of decomposition on a 1-D
signal of length 512).
Now, K1m (Y(m) − Yq(m) )(Y(m) − Yq(m) ) T is the actual autocorrelation matrix of the m-th
subband quantization noise. If we assume

H1 : in each subband, the actual autocorrelation matrix of the quantization noise is
(m) (m) (m)
diagonal, i.e., Km ( Y
1
− Yq(m) )(Y(m) − Yq(m) ) T = diag( D1 , . . . , D N ) (1 ≤ m ≤ M),
then we have
N
(m) (m) (m)
tr[A−1 diag( D1 , . . . , D N )A−T ] = ∑ w i Di
i =1
1 M (m) (m)
) = πm tr[A−1 diag( D1 , . . . , D N )A−T ]
N m∑
Da (X, X
=1
N M
) = 1 ∑ ∑ π m wi D ( m ) .
Da (X, X (11)
N i =1 m =1 i
Moreover, if we assume H4 and

H5 : the spectral transform A is orthogonal, i.e. AA T = I N ,
then
) = Da (Y, Yq ).
Da (X, X (12)
Let us state these results in the following theorem.
158
Theorem 2. With the notations of Section 2.2, the end-to-end distortion of the separable scheme is
given by:
• equation (9) under the assumptions H1 and H2 ;

• equation (12) under the assumptions H4 and H5 .

Remark 3. 1. The assumptions H1 and H1 are consequences of high resolution quantizations (see
Subsection 3.2). They can also be deduced from the condition of statistical independence of the
transformed components, since if the components of Y are independent, then the components of the
quantization noise Y − Yq , which is generally centered, are uncorrelated.
2. A method for the computation of the weighting wavelet coefficients ωm (1 ≤ m ≤ M) can be found
in (Usevitch, 1996; Woods & Naven, 1992).

3. Since the assumptions H1 , H1 , . . . , H4 , are only approximatively satisfied, the equalities (9–13)
are only approximations. However, we observed on many experiments that these approximations
are very good for bit-rates greater than 0.25 bits per pixel and per band.
We search the optimal spectral transform (that is the one which minimizes the total bit-rate
for a given end-to-end distortion) which adapts to the data, assuming high resolution
quantizations hypotheses and 2-D DWT with fixed coefficients, i.e., which do not adapt to
the data. As already mentioned, in our tests we always used the Daubechies 9/7 DWT. First,
we derive the criterion minimized by an optimal spectral transform. We emphasize the fact
that we do not assume Gaussian data and that generally in the literature this assumption
is made in order to clarify the criterion (coding gain) maximized by the optimal transform.
However, the Bennett’s formula and the optimal bit allocation between quantizers formula on
which our criteria are based are well-known and therefore it is straightforward to deduce these
criteria from well-known results. Our major innovation consists especially in the computation
of the optimal transforms, since this computation is generally presented as a difficult task in
classical transform coding and has never been done in the case of the separable scheme which
is JPEG2000 compatible.
4. Criteria for optimal transforms under high resolution quantizations

We recall the extension of the Bennett’s formula which can be stated as follows: if X is
a real random variable quantized under the high resolution hypothesis, then the bit-rate
of quantized variable X q is well approximated by H ( X ) − 12 log2 (cD ), where H ( X ) is the
differential entropy of X, D is the distortion (expected mean square error) introduced by
the quantization and c is a constant depending on the quantization, e.g., for uniform scalar
(m)
quantization c = 12 (Gray & Neuhoff, 1998). Hence, if Ri denotes the quantizer bit-rate
associated with component i and subband m, the Bennet’s approximation gives
(m) (m) 1 (m)
Ri H (Yi )− log2 (cDi )
2
(m)
and the total bit-rate R = 1
N ∑iN=1 ∑m
M
=1 π m R i satisfies
1 N M 1
(m) (m)
R ∑ ∑
N i =1 m =1
πm H (Yi ) − log2 (cDi ) .
2
(13)
and
and Optimal Applied
The problem now consists in minimizing R under the constraint (given by Theorem 2)
1 N M (m)
N i∑ ∑ π m ω m w i Di ≤ D t (14)
=1 m =1
for a given end-to-end distortion Dt . In other words, for a target end-to-end distortion Dt ,
(m)
how can the quantizer distortions Di be distributed in each subband of each component
in order to minimize the total bit-rate? It is a classical problem in compression, called
optimal bit allocation (Gersho & Gray, 1992), that can be solved as follows. According to
relation (13), when the spectral and spatial transforms A and W are given, the differential
(m)
entropies H (Yi ) and the factors wi and ωm are given. Then, the total bit-rate is minimized
(m) πm
∏iN=1 = 1 ( Di )
M
if and only if ∏m N is maximized, that is if and only if

1
N M πm N N M N M πm
(m) πm (m)
∏∏ ∏ wi ∏ =∏ ∏
N N
Di ωm ω m w i Di
i =1 m =1 i =1 m =1 i =1 m =1
is maximized. Now the mean inequality states the last expression (which is a geometric mean)
is not greater than the arithmetic mean corresponding to the left member of inequality (14),
with equality if and only if all the terms in the summation are equal. Hence, the minimization
(m) −1 w−1 for all m and i. That leads to
holds when Di = Dt ω m i

M N
1 (m) 1 1 1
R ∑ πm ∑ H (Yi ) + 2 log2 wi + 2 log2 ωm − 2 log2 (cDt )
m =1
N i =1
and since wi is the ith diagonal element of A−T A−1 , the other terms ωm do not depend on A,
we obtain the following theorem.
Theorem 3. For the separable scheme when the 2-D DWT has fixed coefficients, if high resolution
quantizations hypotheses are assumed, then the optimal spectral transform A is an N × N matrix that
minimizes the criterion:
N M
(m) 1
C2 (A) = ∑ ∑ πm H (Yj )+ log2 det diag(A−T A−1 ). (15)
j =1 m =1
2
=1 πm = 1, the criterion C2 ( A ) can be expressed as

M
Remark 4. Since ∑m

M N
(m) 1 det diag A−T A−1
C2 (A) = ∑ πm ∑ H (Yi ) − log2 | det A| + log2
m =1 i =1
2 det(A−T A−1 )
M
(m)
= ∑ πm C ICA (A) + CO (A), (16)
m =1
(m) (m)
where, for 1 ≤ m ≤ M, C ICA (A) = ∑iN=1 H (Yi ) − log2 | det A| is the criterion to minimize when
performing only ICA to the N components of the transformed coefficients that belong to the subband
m. Pham (Pham, 2004) used that criterion to perform the algorithm ICAinf.
160
det diag(A−T A−1 )

Remark 5. It results of Hadamard’s inequality, that the term CO (A) = 12 log2 det(A−T A−1 ) is
always positive or null (Narozny et al., 2008) and vanishes if and only if A is a matrix whose columns
are pairwise orthogonal, therefore it can be seen like a kind of measure of deviation to orthogonality.
The relation (16) shows that the criterion C2 (A) takes into consideration the fact that one
quantizer per subband and per component is allocated. It is also important to notice that
the criterion C2 (A) involves the transformed coefficients Y. Therefore, even for the separable
scheme (where the order of processing between the 2-D DWT and the spectral transform does
not matter), the search of the optimal spectral transform must be done after the 2-D DWT.
Note that the separable compression scheme does not take into account the difference of
statistics between subbands, since the same spectral transform is applied to all the subbands.
Moreover it is well known that after a DWT some redundancies remain between adjoining
wavelets coefficients. In (Akam Bita et al., 2010a), the authors introduced the subband
compression scheme, that uses as many optimal spectral transforms as subbands in order to
capture the difference of statistics between subbands, and the mixed subband compression
scheme, that captures both redundancies between adjacent wavelet coefficients and the
difference of statistics between subbands. Their experiments on hyperspectral images showed
that these variants of the separable scheme, which are not JPEG2000 compatible, perform
finally worse than the separable scheme because of the increasing of memory size occupied
by the optimal spectral transforms in the bit stream.
Lastly, note that the algorithm that computes a KLT is customarily applied first to the image
before the DWT, but this would be equivalent to applying it after the DWT (i.e. to the DWT
coefficients) if the DWT is orthogonal (as is often the case or at least nearly so in practice1 ). But
then it will not distinguish subbands: the DWT coefficients are considered as coming from a
same (Gaussian) distribution, regardless of the subband they belong to. We feel that the higher
performance — shown in § 7 — of criterion (15) over the criterion 12 log2 ∏ N j=1 var(Yj ), which
leads to the KLT, is due primarily to the fact that it treats each subband separately rather than
that treating the distribution in each subband as non Gaussian. This is logical since, after
any DWT, the energy in each subband depends on the power spectrum of the input signal.
It is important to notice that there is no contradiction in the fact that the criterion (15) treats
each subband separately, while the same spectral transform A is applied to all the subbands.
The idea is then to introduce the distinction between subbands but retain the (approximate)
Gaussian assumption used by the KLT. The distribution of all the wavelet coefficients (with no
distinction between subbands) is a mixture of distributions of the coefficients in the subbands.
It can be shown that the kurtosis of the mixed distribution is higher than the average kurtosis
of the individual distributions. In particular, mixture of Gaussian distributions has always
a positive kurtosis, unless all the individual distributions are the same. Thus the wavelet
coefficients, regardless of the subband they belong to, have a positive kurtosis even if in
each subband their distribution is Gaussian. The above consideration suggests modifying the
criterion (15) by treating the transformed coefficients in each subband m as having a Gaussian
distribution with differing variance for different m. The transformation minimizing this modified
criterion is no longer optimal, but can be nearly so if the distribution in each subband is not too
far from Gaussian. This is not an unrealistic situation: the wavelet coefficients in a subband
is the (decimated) output of a bandpass filter which tends to produce more Gaussian output
than input, due to the reasoning (given e.g. in (Papoulis, 1984) section 8-5) that yields to the
proof of the Central Limit Theorem. The advantage of the modified criterion is that it avoids
1 See Remark 2.
and
and Optimal Applied
11
the entropy estimation and uses only second order statistics. Thus its minimization requires
much less computer resources than using (15).
5. A simplified criterion using only second order statistics

Let H − ( Z ) = log2 var( Z )2πe − H ( Z ) denote the negentropy of Z (which is the difference
of entropy between a Gaussian distribution with variance var( Z ) and the distribution of Z), it
is non negative and vanishes if and only if Z is Gaussian. The criterion (15) can be rewritten
for orthogonal2 matrices
N M
(m) 1 N M (m)
C⊥ (A) = − ∑ ∑ πm H − (Yi
2 i∑ ∑ πm log2 [var(Yi )2πe].
)+ (17)
i =1 m =1 =1 m =1
An analysis of criterion (17) shows that it takes into account two phenomena: 1) the non
(m)
Gaussianity of the transformed coefficients Yi for 1 ≤ m ≤ M and 1 ≤ i ≤ N — this
is controlled by the first term — and 2) the inhomogeneity of the variances in the subbands
— this is controlled by the second term. It is natural to explore the case where the second
phenomenon is the most important, since the DWT tends to render the variables more
Gaussian. In practice, this condition is generally roughly satisfied, except in the LL subband (a
subband of lowest resolution) for which the weighting coefficient πm is generally small. Thus,
if we neglect the variation, induced by the spectral transform A, of the first term in the right
member of equation (17), and if we consider only orthogonal matrices A, then the optimal
transform minimizes the new criterion
1 N M (m)
C (A) =
2 i∑ ∑ πm log2 [var(Yi )]. (18)
=1 m =1
Furthermore if we assume in each component the transformed coefficients have all the
same variance,
regardless
of the subband they belong to, then the criterion (18) becomes
1
2 log2 ∏ N
i =1 var ( Yi ) , leading to the KLT.
In the following, we express criterion (18) in terms of the covariance matrices of the wavelets
coefficients XW T = (XW T )(1) (XW T )(2) · · · (XW T )( M) located in the same subband. The
matrix (XW T )(m) is of dimension N × πm L. Its columns can be considered as different
realizations of a random vector of dimension N whose covariance matrix is denoted C(m) .
Now, Y = AXW T can be written Y = [Y(1) · · · Y( M) ], where Y(m) = (AXW T )(m) is a
matrix whose columns can also be considered as different realizations of a random vector
(m)
having AC(m) A T as covariance matrix. With these notations, we have ∏ N j=1 var(Yj ) =
det diag(AC(m) A T ) and hence the new criterion becomes
1 M
C (A) = πm log2 det diag(AC(m) A T )
2 m∑
(19)
=1
to be minimized with respect to A, under the constraint that it is orthogonal.
2 The orthogonality constraint will be justified in § 7 in which we find that minimizing (15) with and
without this constraint yields almost the same performances. With the orthogonality constraint, the
second term in (15) vanishes.
162
The FG algorithm in (Flury & Gautschi, 1986) can be used to minimize the above criterion.
We have developed a slightly different algorithm (called JADO) which is briefly described in
Appendix 6.3.
6. Minimization of the criteria for the separable scheme

We explain now three algorithms that minimize the criterion (15), one with no constraint but
invertibility, another with the constraint of orthogonality and the third with the constraints
of orthogonality and Gaussian data. To simplify some mathematical expressions we shall use
the Neperian logarithm instead of the base two logarithm until the end of this section.
6.1 The algorithm OST

As in (Pham, 2004) and (Narozny et al., 2008), the algorithms of minimization are based on a
quasi-Newton method with the relative gradient and a simplified relative Hessian. Starting
with a current estimator A, the method consists of expanding C2 (A + E A) with respect to the
matrix E = [Eij ] up to the second order, in a neighborhood of E = 0 N (the null matrix), and
then minimizing the resulting quadratic form in E to obtain a new estimate. Using the results
of (Pham, 2005) it is straightforward to deduce that the Taylor expansion up to the second
(m)
order of C ICA (A + E A) can be approximated as follows
(m) (m) (m) (m)
C ICA (A + E A) = C ICA (A) + ∑ E[ ψ
Yi
(m) (Yi )Yj ]Eij
1≤ i = j ≤ N
1 (m) ( m )2 2
2 1≤i∑
+ {E[ψ2 (m) (Yi )] E[Yj ]Eij + Eij E ji } + · · · , (20)
Yi
= j≤ N
(m) (m)
where the function ψ (m) is equal to the derivative of − log p(yi ) — p ( yi ) denoting the
Yi
(m)
probability density function of Yi — and is known as the score function. Let M = A−T A−1 .
In (Narozny et al., 2008), the Taylor expansion of CO (A + E A) is given up to the second order,
however it is quite involved and it is simplified into

M ji 1 M jj 2
CO (A + E A) ≈ CO (A) − ∑ E ji + ∑ E ji + E ji Eij + · · · (21)
1≤ i = j ≤ N
Mii 2 1≤i = j≤ N Mii
by neglecting the non diagonal elements of M = [ Mij ] in the second order terms of the Taylor
expansion.
Using the approximation (21), the equality (20) and the relation (16) we obtain

M M
(m) (m) ij
C2 (A + E A)=C2 (A) + ∑ ∑ πm E Yj ψYi(m) Yi − E
1≤ i = j ≤ N m =1
M jj ij

1 M M
( m )2 (m)
2 1≤i∑ ∑ πm Eij E Yj
2 2 ii 2
+ E ψ (m) Yi + E + 2Eij E ji . (22)
= j ≤ N m =1
Yi M jj ij
The quadratic form associated to this last expansion is positive definite. One iteration of the
algorithm is first to solve the following equation

Ψij 2 Eij Φij
= , (23)
2 Ψ ji E ji Φ ji
and
and Optimal Applied
13
Mij (m) (m) (m) ( m )2

with Φij = M jj − ∑m
M
=1 πm E [ ψY (m) (Yi )Yj ] and Ψij = ∑m
M
=1 πm E [ ψ (m) (Yi
2 )] E[Yi ]+
i Yi
Mii
M jjand then to replace the current solution A with A + E A. Since the diagonal elements of
E are undetermined, they are arbitrarily fixed to zero. For the practical computation of the
(m) that is described in (Pham, 2005) as well
algorithm, we replace ψ (m) with its estimator ψ
Y i Y i
as the estimator of the differential entropy. The mathematical expectations are replaced with
simple empirical means. We call OST (Optimal Spectral Transform) the algorithm described
above and OST the optimal transform returned by this algorithm.
6.2 The algorithm OrthOST

To minimize the criterion (15) with the constraint that the solution is an orthogonal matrix, it
is important to note, as in (Narozny et al., 2008), that if A is orthogonal, then A + E A remains
orthogonal when I + E is also orthogonal. This condition is satisfied up to the first order if E
is an antisymmetrical matrix, since then (I + E ) T (I + E ) = I + E T E . Using that condition, the
expansion (22) becomes
M
(m) (m) (m) (m)
C (A + E A) = C (A) + ∑ ∑ πm E[Yj ψ
Y
(m) (Yi )] − E[Yi ψ
Y
(m) (Yj )] Eij +
m =1 1≤ i < j ≤ N i j
M
1 ( m )2 (m) ( m )2 (m)
2 1≤i∑ ∑ πm E[Yj ] E[ψ2 (m) (Yi )] + E[Yi ] E[ψ2 (m) (Yj )] − 2 Eij2 . (24)
Y Y
< j≤ N m =1 i j
The matrix E is calculated in that case according to

(m) (m) (m) (m)
M
∑m =1 π m E [ Y i ψ
Yj
( m ) ( Yj )] − E [ Y j ψ
Yi
( m ) ( Yi )]
Eij = . (25)
( m )2 (m) ( m )2 (m)
M
∑m =1 π m E [ Yj ] E [ ψ 2
(m) ( Y i )] + E [ Yi ] E [ ψ 2
(m) ( Yj )] − 2
Yi Yj
Actually, A + E A obtained in this way is not a true orthogonal matrix. This can be overcome
by replacing A + E A with eE A = (I + E + E 2 /2! + · · · )A, which is orthogonal and differs from
A + E A only by second order terms. We call OrthOST (Orthogonal Optimal Spectral Transform)
this algorithm and OrthOST the orthogonal transform returned by the algorithm. The case
where the spectral transform is constrained to be orthogonal is particularly interesting because
the weightings which depend on the linear transform are all equal to one.
6.3 The JADO (Joint Approximate Diagonalization under Orthogonality constraint) algorithm
Given K positive definite (complex) matrices C1 , . . . , CK associated with positive weights w1 ,
. . . , wK , the JADO algorithm aims to find a unitary matrix B which minimizes
K
C (B) = ∑ wk log det diag(BCk B∗ ) (26)
k =1
where ∗ denotes the hermitian operator. This algorithm differs only slightly from FG
algorithm in (Flury & Gautschi, 1986). However, its derivation in (Flury & Gautschi, 1986)
is complex and difficult to understand. Here we provide briefly a much simpler derivation.
164
The idea is to make successive Givens rotations, each time on a pair of rows of B, the ith row
Bi· and the jth row B j· , say:

Bi · Bi ·
← Tij , (27)
B j· B j·
where Tij is a 2 × 2 unitary matrix, chosen so that the criterion is decreased. The processing
K ( K −1)
of all the 2 pairs is called a sweep. The algorithm consists of repeated sweeps until
convergence is achieved.
The decrease of the criterion (26) induced by (27) is

K Bi ·
∑ wk log (Bi· Ck Bi∗· )(B j· Ck B∗j· ) det diag Tij Ck Bi∗· B∗j· Tij∗ .
k =1 B j·
A natural idea is to chose Tij to maximize this decrease, but there is no closed form formulae
for that. Our idea is to maximize a lower bound of it instead. Since for a > 0, b ≥ 0, log( a/b) ≥
1 − b/a, the above decrease can be seen to be bounded below by
∗ ∗
2(w1 + · · · + wK ) − Tij;1· PTij;1 · − Tij;2· QTij;2· , (28)
where Tij;1· and Tij;2· are the first and second rows of Tij and

K
wk Bi · K
wk Bi ·
P= ∑ ∗ Ck [ Bi∗· B∗j· ]; Q= ∑ ∗ Ck [ Bi∗· B∗j· ].
k =1
B i · C k Bi · B j· k =1
B j · C k B j· B j·
Since Tij;2· has unit norm and is orthogonal to Tij;1· , it must be of the form eiα Tij;1· J where
α is some phase angle, x denotes the complex conjugate of x and J is the 2 × 2 matrix with 0
∗
on the diagonal and 1, −1 on the anti-diagonal. Thus Tij;2· QTij;2 ∗ ∗
· = Tij;1· JQJ Tij;1· , but since
the above left hand side is real (as Q is hermitian), it also equals Tij;1· JQJ∗ Tij;1
∗ . Therefore
·
expression (28) can be rewritten as 2(w1 + · · · + wK ) − Tij;1· (P + JQJ∗ )Tij;1
∗ . Maximizing it
·
with respect to the unitary matrix T thus amounts to minimizing Tij;1· (P + JQJ∗ )Tij;1 ∗
· with
respect to the vector of unit norm Tij;1· . The solution is that Tij;1· is (up to a factor of unit
modulus) the normalized left eigenvector of the smallest eigenvalue of P + JQJ∗ . Since Tij;2·
is orthogonal to Tij;1· it is the other eigenvector. Finally, Tij is the matrix formed by the left
eigenvectors of P + JQJ∗ . Its elements can be computed explicitly in closed form as follows.
We note that the off diagonal elements of JQJ∗ is the negative of those of Q while the diagonal
elements are those of Q in reverse order. Thus JQJ∗ = tr(Q)I − Q where tr denotes the trace.
Since the addition of a multiple of the identity matrix does not change the eigenvectors, Tij
is also the matrix formed by the left eigenvectors of P − Q. One can now recognize that
the rotation (27) is the same as an iteration in the G loop of the FG algorithm. However, it
differs from our JADO algorithm in that it repeats (27) with the same pair i, j (but with the
newly computed Bi· and B j· ) until convergence (the G loop) and only then another pair i, j
is considered. We feel that this is not efficient since the decrease of the criterion will be very
small near the end of the G loop. We call JADOST the transform returned by the algorithm.
and
and Optimal Applied
15
6.4 Computational complexity of the optimal transforms

We give here a rough estimation of the number of operations required for the computation
of the two first algorithms described above, taking into account only multiplications and
divisions. The differential entropies and the score functions are calculated according to a
method explained in (Pham, 2005). The computational complexity of each of these quantities
is O(NrL), where r is the number of bins in the binned kernel density estimation. In
general r L and for most cases r belongs to the interval [30 , 60]. At each iteration, the
criterion and the matrix E must be computed. The complexity of the criterion computation is
O(NrL + N 3 ). For the calculation of the matrix E , we first need to compute the score function.
The complexity of the matrix E computation (including the score function computation) is
O(NrL + N 2 L). Finally, the complexity of one iteration is O(NrL + N 2 L + N 3 ). In practice,
the convergence of the algorithm is usually obtained after p iterations, p ∈ [20 , 60]. Generally
N L and the total computational complexity is O( p( NrL + N 2 L)). The computational
complexities of OrthOST is the same. We recall that for the computation of a KLT, this
complexity is O( LN 2 ). The JADOST and KLT computation complexities are roughly the same.
In this section we present the performances in image compression of the optimal transforms
described in the previous sections.
7.1 Description of the tests
Fig. 1. From up to down Moissac, Vannes, Toulouse, Port-de-Bouc

We tested two kinds of multicomponent images: multispectral ones and hyperspectral ones.
The multispectral images are3 PLEIADES simulations of French cities with N = 4 components
and coded on Nb = 12 bpppb: Moissac with Nc × Nr = 320 × 3152, Port-de-Bouc with
3 These images have been given by the French Space Agency CNES (Centre National d’Etudes Spatiales).
They are described on the web site http://smsc.cnes.fr/PLEIADES/.
166
Nc × Nr = 320 × 1376, Toulouse with Nc × Nr = 352 × 3816, Vannes with Nc × Nr = 352 × 3736,
. . . The hyperspectral images are4 AVIRIS images (Moffett, Cuprite and Jasper) with N = 224
components from the visible to the infrared and coded on Nb = 16 bpppb. They are
originally acquired with Nr × Nc = 512 × 624, but for the simulations we kept only the 512
leftmost columns. Some images used in our tests are shown in figures 1 and 2. As already
Fig. 2. From left to right: Moffett, Jasper and Cuprite

mentioned, the 2-D DWT used in all our experiments is the Daubechies 9/7 which proved
to be efficient in lossy image compression (Antonini et al., 1992; Taubman & Marcellin, 2002).
For simplicity, we used only uniform scalar quantizers with a dead zone twice as large as
the quantization step. The performances are evaluated in terms of bit-rate versus end-to-end
distortion. For hyperspectral images, we considered four distortions. A first one is the mean
square error (MSE) expressed in terms of the Signal to Noise Ratio, SNR = 10 log10 (σ2 /D )
where D is the actual end-to-end MSE distortion and σ2 = ∑iN=1 ∑nL=1 ( Xi (n) − μ)2 /( NL)
is the empirical variance of the initial image with the empirical mean of the image μ =
∑iN=1 ∑nL=1 Xi (n)/( NL). A second distortion is the maximal absolute difference (MAD =
max{| Xi (n) − X i (n)| : 1 ≤ i ≤ N and 1 ≤ n ≤ L}), a third one is the maximum spectral

∑ i = 1 Xi ( n ) Xi ( n )
N
angle MSA = max acos N 2
: 1 ≤ n ≤ L and the last one is the mean
∑i=1 Xi2 (n) ∑i=1 X i (n)
N
absolute error (MAE = ∑iN=1 ∑nL=1 | Xi (n) − X i (n)|/( NL)). With these four
distortions, one
can estimate the performances of a codec on usual applications of hyperspectral images, like
classifications and targets detections (Christophe et al., 2005). For multispectral images, we
considered only the MAD and the MSE distortions, the last one being expressed in terms
of Peak of Signal to Noise Ratio, PSNR = 10 log10 [(2 Nb − 1)2 /D ], where D is the actual
end-to-end MSE distortion and Nb is the number of bits per pixel and per band (bpppb) of
the initial image. The bit-rate, expressed in bpppb, was measured on the actual bit stream
obtained with the JPEG2000 coder EBCOT (Taubman, 2000) and its PCRD optimizer applied
across components for optimal bit allocation. We used the Verification Model version 9.1
(VM9 (JPEG2000, 2001)) codec developed by the JPEG2000 group. The coefficients of A−1 (the
inverse matrix of the optimal spectral transform) and the mean of each component are stored
in the bitstream as float32 data (this costs 32( N + 1)/L bpppb). A difference exists between
the aimed bit-rate and the actual bit-rate obtained with the VM9. In our tests, this difference
does not exceed ± 0.001 bpppb and thus the precision of the PSNR is about ± 0.05 dB.
4 These images have been downloaded from the NASA web site http://aviris.jpl.nasa.gov/.
and
and Optimal Applied
17
7.2 Bit-rate versus distortion performances

In this subsection, we discuss and compare the bit-rate versus distortion performances of
different spectral transforms. Table 1 presents the bit-rate of different transforms versus the
PSNR (dB) MAD
bit-rate 0.25 0.50 0.75 1.00 1.50 2.00 2.50 3.00 0.25 0.50 0.75 1.00 1.50 2.00 2.50 3.00
Moissac
Id 36.37 39.59 41.93 43.89 47.22 50.16 52.93 55.66 691 366 253 187 108 68 49 38
KLT 38.61 42.39 45.24 47.63 51.51 54.49 56.98 59.44 716 381 214 135 79 48 32 25
JADOST 38.54 42.30 45.14 47.52 51.39 54.42 56.98 59.47 700 357 298 137 79 46 31 24
OrthOST 38.67 42.50 45.35 47.72 51.55 54.55 57.11 59.60 818 399 229 145 78 47 33 24
OST 38.69 42.55 45.43 47.80 51.62 54.59 57.15 59.65 745 496 215 138 78 48 32 23
Port-de-Bouc
Id 30.36 33.68 36.14 38.25 41.93 45.27 48.43 51.52 1198 653 544 361 198 135 85 64
KLT 33.47 37.74 40.88 43.45 47.53 50.89 53.82 56.53 922 513 297 230 139 74 50 35
JADOST 33.26 37.56 40.73 43.32 47.47 50.91 53.93 56.68 922 504 324 256 135 75 52 32
OrthOST 33.42 37.80 41.05 43.71 47.90 51.31 54.28 56.99 885 513 305 237 122 77 51 31
OST 33.46 37.85 41.12 43.78 48.00 51.40 54.36 57.06 866 557 351 256 129 82 53 35
Vannes
Id 39.25 42.89 45.67 47.99 51.77 54.80 57.51 60.11 603 269 178 109 63 42 29 21
KLT 41.36 45.71 48.78 51.11 54.38 56.82 59.24 61.79 482 219 148 86 51 33 24 18
JADOST 41.83 46.15 49.16 51.42 54.61 57.09 59.53 62.06 368 214 134 84 48 29 24 19
OrthOST 41.90 46.27 49.29 51.54 54.71 57.18 59.62 62.16 354 190 135 91 46 30 25 18
OST 41.94 46.34 49.35 51.59 54.74 57.22 59.68 62.20 393 204 138 88 45 33 25 16
Strasbourg
Id 30.82 34.19 36.73 38.91 42.70 46.09 49.20 52.13 1357 877 546 353 205 118 86 60
KLT 32.51 36.59 39.77 42.49 46.99 50.58 53.51 56.08 1041 927 438 403 184 90 52 38
JADOST 32.47 36.51 39.65 42.33 46.78 50.36 53.33 55.92 1082 872 543 371 189 85 56 45
OrthOST 32.51 36.59 39.78 42.50 47.01 50.61 53.55 56.11 1010 948 449 404 178 87 50 38
OST 32.49 36.59 39.79 42.53 47.07 50.67 53.60 56.17 1149 904 455 289 162 81 55 42
Montpellier
Id 32.17 35.23 37.59 39.62 43.17 46.30 49.17 51.95 1216 630 406 292 168 117 77 54
KLT 34.09 37.75 40.60 43.03 47.20 50.69 53.63 56.18 747 488 340 248 143 75 49 34
JADOST 34.09 37.72 40.55 42.99 47.15 50.62 53.55 56.13 782 501 323 245 124 81 51 36
OrthOST 34.08 37.90 40.92 43.46 47.72 51.16 54.01 56.55 681 454 338 255 127 68 47 32
OST 34.14 37.99 41.01 43.56 47.79 51.21 54.06 56.60 704 483 332 239 127 68 46 33
Perpignan
Id 33.71 36.90 39.34 41.43 45.04 48.17 51.04 53.78 984 526 332 230 158 89 62 42
KLT 36.51 40.44 43.29 45.60 49.33 52.36 54.99 57.52 726 435 245 172 84 54 41 29
JADOST 36.55 40.51 43.37 45.69 49.43 52.44 55.07 57.60 715 388 234 172 90 58 39 30
OrthOST 36.59 40.59 43.48 45.83 49.61 52.66 55.30 57.82 721 371 232 165 94 54 37 30
OST 36.60 40.60 43.49 45.84 49.62 52.67 55.32 57.85 645 383 292 164 94 55 38 28
Table 1. Bit-rate (in bpppb) versus PSNR (in dB) and versus MAD of different spectral
transforms on multispectral images (best results are bolded). The bit-rate was computed with
the VM9.
two distortions PSNR and MAD on six multispectral images and Tables 2 and 3 show the
bit-rate of different transforms versus the four distortions SNR (in dB), MAE, MAD and
MSA (expressed in degree ◦ ) on three hyperspectral images. All the 2-D DWT was applied
with five levels of decomposition. We observe the well-known fact that spectral transforms
perform significantly better than the identity matrix (i.e., no spectral transform), especially
for hyperspectral images. Indeed, on six multispectral images (see Table 1) the average gains
168
SNR (dB) MAE

bit-rate 0.25 0.50 0.75 1.00 1.50 2.00 2.50 3.00 0.50 0.75 1.00 1.50 2.00 2.50 3.00
Moffett
Id 25.45 30.37 33.97 36.94 41.78 45.76 49.15 52.01 24.32 16.93 12.52 7.62 5.02 3.48 2.51
KLT 44.21 47.68 50.08 51.97 54.76 57.10 59.21 61.04 3.83 3.03 2.49 1.82 1.39 1.07 0.85
JADOST 45.13 48.39 50.70 52.50 55.17 57.47 59.53 61.30 3.54 2.83 2.35 1.74 1.33 1.03 0.82
OrthOST 45.31 48.57 50.87 52.61 55.28 57.57 59.62 61.37 3.47 2.78 2.32 1.72 1.31 1.02 0.81
OST 45.32 48.56 50.87 52.62 55.30 57.64 59.77 61.63 3.47 2.78 2.32 1.72 1.30 1.00 0.79
Cuprite
Id 29.99 33.48 36.12 38.41 42.44 45.99 49.19 52.11 26.07 19.85 15.60 10.13 6.89 4.83 3.47
KLT 47.79 50.46 52.55 54.16 56.76 59.07 61.26 63.27 3.96 3.23 2.73 2.04 1.55 1.19 0.92
JADOST 48.22 50.85 52.86 54.42 56.97 59.27 61.44 63.43 3.80 3.13 2.65 1.99 1.51 1.16 0.90
OrthOST 48.25 50.88 52.89 54.44 56.99 59.29 61.46 63.44 3.79 3.12 2.65 1.98 1.51 1.16 0.90
OST 48.26 50.89 52.89 54.44 57.01 59.34 61.56 63.60 3.79 3.12 2.65 1.98 1.50 1.14 0.88
Jasper
Id 21.34 24.83 27.56 29.92 34.01 37.67 41.09 44.33 64.84 34.39 26.82 17.23 11.52 7.89 5.49
KLT 42.93 46.49 48.61 50.37 53.18 55.56 57.72 59.66 4.04 3.27 2.72 1.99 1.51 1.16 0.91
JADOST 43.56 46.89 48.97 50.67 53.43 55.78 57.91 59.83 3.87 3.15 2.63 1.94 1.47 1.13 0.89
OrthOST 43.66 46.94 49.02 50.73 53.47 55.81 57.94 59.85 3.85 3.13 2.62 1.93 1.46 1.13 0.88
OST 43.70 46.96 49.05 50.74 53.50 55.87 58.03 60.01 3.84 3.12 2.61 1.92 1.45 1.11 0.87
Table 2. Bit-rate (in bpppb) versus SNR (in dB) and versus MAE of different spectral
transforms on hyperspectral images. The bit-rate was computed with the VM9.
of the KLT, JADOST, OrthOST and OST on Identity are respectively 3.6 dB, 3.6 dB, 3.8 dB and
3.8 dB. On three hyperspectral images (see Table 2) the average gains of the KLT, JADOST,
OrthOST and OST on Identity are respectively 15.9 dB, 16.3 dB, 16.3 dB and 16.4 dB. Moreover,
we can notice that the optimal transforms OrthOST and OST perform always a little better
than the KLT at medium and high bit-rates: on six multispectral (resp. three hyperspectral)
images the average gains of OrthOST and OST on KLT are about 0.23 dB and 0.28 dB (resp.
0.43 dB and 0.49 dB). On the multispectral images, we observed that JADOST performs
roughly as the KLT for MSE distortion, sometimes slightly better, sometimes slightly worse, at
any rate. On six images, the average gain of JADOST on the KLT is negligible (about 0.02 dB)
at medium and high bit-rates (from 0.25 to 3 bpppb), whereas the average gain of OrthOST on
JADOST is about 0.21 dB at the same rates. Nevertheless, on hyperspectral images, JADOST
performs slightly but significantly better than the KLT for the four distortions tested at
medium and high bit-rates (see Tables 2 and 3) and nearly reaches the OrthOST scores with a
significantly lower computational complexity. The average gain of JADOST on the KLT (resp.
OrthOST on JADOST) is 0.37 dB (resp. 0.07 dB) on the range [0.25 bpppb , 3 bpppb]. Further,
we can remark that there is an insignificant difference of performances between OrthOST and
OST. This can be explained by the fact that transforms minimizing the criterion (16) must have
a small value for CO (A), i.e., they must be close to orthogonality (see Remark 5). Therefore
there is no advantage to use OST rather than the orthogonal transform OrthOST. In examining
the MAD distortion we observe that on the multispectral images tested, at medium bit-rates
(i.e. between 0.25 and 1.5 bpppb), OrthOST performs worse than the KLT (see Table 1). On
the other hand, on the three hyperspectral (AVIRIS) images tested, for all the distortions
measured, at medium and high bit-rates, JADOST and OrthOST perform always better than
the KLT (see Tables 2 and 3). This is a nice finding, since the optimality of OrthOST is justified
only for the MSE distortion and at high bit-rates.
and
and Optimal Applied
19
MSA (◦ ) MAD
bit-rate 0.25 0.50 0.75 1.00 1.50 2.00 2.50 3.00 0.25 0.50 0.75 1.00 1.50 2.00 2.50 3.00
Moffett
Id 12.12 6.82 3.94 2.66 1.29 0.85 0.52 0.36 1676 781 492 1259 183 62 32 20
KLT 1.43 0.87 0.57 0.37 0.20 0.15 0.12 0.10 392 211 119 67 24 14 8 7
JADOST 1.15 0.59 0.42 0.27 0.19 0.14 0.11 0.09 279 120 67 44 18 12 8 6
OrthOST 0.96 0.47 0.31 0.25 0.18 0.14 0.11 0.09 261 77 49 33 18 10 8 6
OST 0.86 0.50 0.32 0.25 0.18 0.14 0.11 0.09 207 101 46 37 19 12 7 6
Cuprite
Id 5.30 2.81 2.20 1.57 1.01 0.59 0.40 0.26 659 360 253 185 110 62 61 40
KLT 0.42 0.25 0.22 0.15 0.12 0.08 0.07 0.06 154 135 100 54 26 16 10 8
JADOST 0.33 0.25 0.16 0.14 0.10 0.08 0.07 0.05 112 109 61 39 20 11 9 7
OrthOST 0.32 0.25 0.17 0.14 0.10 0.08 0.07 0.05 113 110 61 37 22 11 9 7
OST 0.35 0.24 0.16 0.14 0.10 0.08 0.07 0.05 113 109 58 42 17 11 9 7
Jasper
Id 18.20 12.53 7.88 5.70 3.87 2.14 1.41 1.01 1907 1220 732 559 241 160 84 55
KLT 0.91 0.53 0.43 0.34 0.26 0.20 0.15 0.12 225 151 82 57 30 15 10 7
JADOST 0.87 0.51 0.44 0.33 0.24 0.19 0.15 0.12 157 91 56 51 20 11 9 7
OrthOST 0.83 0.51 0.40 0.33 0.24 0.19 0.15 0.12 157 84 46 34 23 13 9 7
OST 0.79 0.51 0.41 0.32 0.24 0.18 0.14 0.11 156 86 48 34 22 14 8 6
Table 3. Bit-rate (in bpppb) versus MSA (in degree ◦ ) and versus MAD of different spectral
transforms on hyperspectral images for the separable scheme. The bit-rate was computed
with the VM9.
As already mentioned, the main drawback of the OrthOSTs returned by JADO and OrthOST
algorithms is their heavy computational costs. In the next section we introduce quasi-optimal
orthogonal spectral transforms.
8. Performances of exogenous quasi-optimal spectral transforms

8.1 Exogenous quasi-optimal spectral transforms
When one gets a set of images coming from one (and only one) spectrometer sensor, it is
possible to compute an exogenous OrthOST from a learning basis extracted from this set.
Generally, images from one spectrometer have the same number of bands and the same
number of rows. However, the number of columns may vary. To compute an exogenous
OrthOST, we first split the set of all images in two disconnected sets, one consisting of several
images and which becomes the learning basis L, the other constituted of the remaining images
and which becomes the test subset. Then, all the images of the learning basis are connected
band per band and row per row to construct a single virtual large image X having the same
numbers of bands and rows as any image from the spectrometer and a large number of
columns. This image is used as input of the OrthOST algorithm described in (Akam Bita
et al., 2010a) and the output is the exogenous OrthOST associated with the learning basis L.
The exogenous KLT and exogenous JADOST are calculated similarly.
8.2 Performance comparison between exogenous and non exogenous OrthOSTs

In our tests, we used 10 images5 (shown in Figure 3) from the imaging spectrometer MERIS
on-board the satellite ENVISAT. This fifteen spectral bands spectrometer operates in the solar
5 The images were acquired via the Data Disseminated System (http://dwlinkdvb.esrin.esa.it/DDS/)
thanks to ESA/ESRIN.
170
Fig. 3. Fifth component (corresponding approximately to the band [555 nm, 565 nm]) of the
hyperspectral images MERIS, numbered 1–4, 6, 8, 10, 13, 15–16, from left to right
reflective spectral range of visible and near infrared light. Each band has a programmable
width and a programmable location in the 390 nm to 1040 nm spectral range. As mentioned
in (MERIS, 2006) the instrument scans the Earth’s surface by the push-broom method, CCD
arrays provide spatial sampling in the across-track direction, while the satellite’s motion
provides scanning in the along-track direction. The scene is imaged simultaneously across
the entire spectral range, through a dispersing system, onto the CCD array. Therefore there
is no problem of deregistration on the MERIS images. The ten images of our tests have all
the same dimensions: Nr = 128, Nc = 1121 and N = 15. They are originally coded on
Nb = 16 bpppb and they were acquired with the same fifteen spectral bands. To construct
exogenous KLT and exogenous OrthOST, we split the ten MERIS images in two disconnected
sets, one constituted of seven images (the learning basis), the other constituted of the three
remaining images (the test subset). We considered 13 various learning bases, denoted Li
(1 ≤ i ≤ 13) which are presented in Table 4. The bit-rate is computed with the Verification
Model version 9 (VM9 (JPEG2000, 2001)). Note that the exogenous transforms are fixed (i.e.,
they do not adapt to the encoded image), hence they are known by the decoder and have not
to be transmitted. However, in the lossy compression results given in Table 5 with the VM9,
the inverse of the spectral transform is coded in the bit-stream (it costs less than 0.001 bpppb).
and
and Optimal Applied
21
L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11 L12 L13

Meris1 × × × × × × × × ×
Meris2 × × × × × × × × ×
Meris3 × × × × × × × × ×
Meris4 × × × × × × × × × ×
Meris6 × × × × × × × × ×
Meris8 × × × × × × × × ×
Meris10 × × × × × × × × ×
Meris13 × × × × × × × × ×
Meris15 × × × × × × × × ×
Meris16 × × × × × × × × ×
Table 4. Various learning bases, denoted Li (1 ≤ i ≤ 13) and constituted of seven MERIS
images each
In Table 5 we present the performances obtained with two images when the learning basis
varies. Among all the tests we made, we chose to show the best and worst cases obtained
with an exogenous OrthOST. For this, the PSNR of a spectral transform is compared to that
obtained with the KLT by subtracting, and we considered that difference of PSNR at 1 bpppb.
The best and worst cases correspond respectively to the tested images MERIS2 and MERIS8.
We can see that for MERIS2, the exogenous OrthOST performs significantly better than the
KLT at all rates for both MSE and MAE global distortions. Whereas for both MAD and
MSA local distortions, exogenous OrthOST and KLT have roughly the same performance, the
winner depending on the bit-rate. A more interesting result is the worst case: at bit-rates
not greater than 1 bpppb, the worst exogenous OrthOST performs worse than the worst
exogenous KLT and this trend is reversed for bit-rates larger than 1.0 dB. Moreover, the loss
of PSNR compared to the KLT is 4.3 dB at 1 bpppb, however, the difference of PSNR between
the KLT and identity (i.e., no spectral transform) is particularly high here (30 dB). For the
other eight tested images, the loss of PSNR of the worst exogenous OrthOST with respect to
the KLT, is not greater that 2.5 dB, at all bit-rates. Moreover, it is always smaller than the loss
of PSNR of the best exogenous KLT with respect to the KLT. An example is shown in Fig. 4,
where the bit-rate is computed either with the VM9 or the Bit Plane Encoder (BPE) (CCSDS-1,
2007) recommended by the CCSDS (Yeh et al., 2005). In order to compute the bit-rate with
the BPE, we proceeded as follows: first, for each transformed component we computed a few
hundred points of the graph that links mean square error to bit-rate, then we applied the
algorithm by Shoham and Gersho (Shoham & Gersho, 1988) to optimally allocate distortions
between components for given maximal total bit-rates.
In average on the 10 images, the loss of PSNR of the worst exogenous OrthOST with respect
to the KLT is significantly smaller than the one of the best exogenous KLT (see Table 6). It
is no longer the case for exogenous JADOST. We observed the importance of the learning
basis, whose influence can range from 0 dB to −4 dB. In other words, when the learning basis
is well chosen (depending on the scene and not only on the spectrometer), one can expect
a loss of PSNR of an exogenous OrthOST with respect to the KLT not greater than 0.4 dB.
Whereas, when it is badly chosen, the same loss of PSNR should be limited to 4 dB. However,
these values are only indicative and should not be considered definitive, because they were
obtained on a set of 10 MERIS images, which was not proven to be statistically significant.
We observed good performances of exogenous OrthOST used with the VM9 and in (Akam Bita
et al., 2010c) the authors observed that, associated with the BPE and the optimal bit allocation
172
PSNR (dB) MAD

bit-rate 0.25 0.50 0.75 1.00 1.50 2.00 3.00 0.25 0.50 0.75 1.00 1.50 2.00 3.00
MERIS2
Id 39.40 42.13 44.20 45.99 49.22 52.26 58.26 5878 3271 2369 2226 1258 850 443
KLT 54.87 63.81 69.60 73.66 79.29 83.02 88.39 1407 564 229 140 65 27 13
JADOST 55.65 65.04 70.53 74.49 80.02 83.66 88.92 1309 465 197 134 56 26 13
OrthOST 55.65 65.41 71.15 75.23 80.66 84.17 89.33 1354 397 189 119 48 22 12
exo3_KLT 54.11 62.36 67.79 71.85 77.77 81.86 87.53 1392 511 256 159 72 37 16
exo3_JADOST 55.11 64.01 69.53 73.46 78.82 82.78 88.23 1313 434 228 121 191 37 15
exo3_OrthOST 55.35 64.48 70.15 74.18 79.71 83.40 88.72 1267 476 240 134 53 28 13
exo5_KLT 54.08 62.37 67.89 72.04 78.05 82.11 87.69 1326 536 325 154 68 41 15
exo5_JADOST 55.12 64.01 69.54 73.48 79.08 82.88 88.30 1289 398 228 117 58 28 13
exo5_OrthOST 55.36 64.46 70.11 74.13 79.70 83.42 88.74 1314 497 257 134 49 26 12
exo7_KLT 54.69 62.92 68.09 72.00 77.79 81.84 87.52 1226 602 265 183 66 43 16
exo7_JADOST 55.22 63.97 69.14 72.88 78.32 82.26 87.86 1236 524 244 165 64 39 14
exo7_OrthOST 55.61 64.94 70.32 74.18 79.57 83.25 88.60 1418 504 238 128 56 28 13
exo12_KLT 54.13 62.37 67.83 71.95 77.99 82.07 87.66 1432 513 312 152 68 42 15
exo12_JADOST 55.13 64.02 69.54 73.50 79.08 82.87 88.30 1289 435 258 136 52 30 14
exo12_OrthOST 55.35 64.48 70.14 74.15 79.72 83.42 88.73 1305 481 221 139 47 30 13
MERIS8
Id 35.22 38.02 40.17 42.11 45.67 49.06 55.71 9723 5652 4123 3068 2050 1313 626
KLT 53.85 62.69 68.09 72.11 77.71 81.48 87.18 2638 1192 411 228 87 37 17
JADOST 54.21 63.68 69.39 73.44 78.83 82.35 87.91 2838 709 262 148 50 31 15
OrthOST 54.2 63.86 69.61 73.7 79.07 82.58 88.12 2773 705 305 130 51 26 15
exo2_KLT 52.22 59.49 64.16 68 73.95 78.34 84.79 3501 1504 608 400 149 51 24
exo2_JADOST 51.19 58.64 63.66 67.83 73.93 78.83 85.2 3536 1608 646 321 501 59 21
exo2_OrthOST 50.54 58.19 63.46 67.8 74.39 78.95 85.29 3446 1471 592 326 133 46 20
exo6_KLT 52.62 59.99 64.57 68.16 73.81 78.24 84.77 2813 973 471 255 115 59 23
exo6_JADOST 52.3 59.46 63.86 67.4 73.16 77.43 84.44 2670 1077 547 312 138 229 23
exo6_OrthOST 51.57 59.7 64.95 68.96 75.01 79.36 85.6 2871 1195 482 290 91 51 20
exo7_KLT 52.49 59.68 64.18 67.83 73.63 78.06 84.59 2887 1013 469 287 126 70 23
exo7_JADOST 52.28 59.45 63.85 67.44 73.31 77.92 84.6 2689 980 543 328 135 70 22
exo7_OrthOST 51.71 59.29 64.16 68.03 74.14 78.69 85.11 2768 1147 451 326 115 51 22
exo11_KLT 52.56 59.87 64.43 68.06 73.76 78.16 84.67 2885 876 459 255 104 68 24
exo11_JADOST 52.28 59.49 63.94 67.6 73.49 78.08 84.7 2610 961 525 290 126 67 23
exo11_OrthOST 51.89 59.59 64.64 68.73 75.01 79.44 85.68 2909 1130 441 346 131 47 19
MSA (in ◦ ) MAE
MERIS2
Id 34.94 24.67 19.57 14.91 10.70 8.054.37 503.5 377.8 302.7 248.5 173.7 123.3 62.17
KLT 8.63 2.91 1.62 1.36 0.42 0.230.14 87.50 31.02 15.88 10.03 5.41 3.58 1.93
JADOST 10.41 2.66 1.54 0.83 0.38 0.220.11 80.46 27.18 14.45 9.22 4.99 3.33 1.82
OrthOST 9.47 2.55 1.57 1.07 0.44 0.210.11 80.22 25.87 13.43 8.48 4.65 3.15 1.73
exo3_KLT 10.26 4.19 2.05 1.16 0.54 0.280.13 95.35 36.85 19.79 12.44 6.43 4.09 2.13
exo3_JADOST 8.42 3.33 1.75 1.19 0.51 0.300.13 85.27 30.56 16.17 10.35 5.66 3.69 1.97
exo3_OrthOST 9.07 3.27 1.63 1.10 0.38 0.250.11 83.01 28.91 15.06 9.52 5.17 3.44 1.86
exo5_KLT 10.12 3.89 2.06 1.38 0.54 0.300.13 95.60 36.72 19.50 12.18 6.24 3.98 2.10
exo5_JADOST 8.61 3.59 1.96 1.00 0.49 0.280.12 85.21 30.49 16.12 10.31 5.55 3.64 1.95
exo5_OrthOST 9.03 3.25 1.49 1.08 0.41 0.240.12 82.71 28.94 15.10 9.56 5.18 3.43 1.86
exo7_KLT 8.59 3.79 1.89 1.37 0.53 0.320.13 89.55 34.46 19.04 12.20 6.40 4.10 2.14
exo7_JADOST 9.72 2.95 1.67 1.29 0.47 0.270.13 84.59 30.68 16.86 11.01 6.03 3.91 2.06
exo7_OrthOST 10.33 2.78 1.56 0.90 0.43 0.240.12 80.62 27.32 14.70 9.49 5.24 3.49 1.89
exo12_KLT 9.41 4.17 2.46 1.23 0.54 0.290.13 95.20 36.77 19.67 12.30 6.28 3.99 2.10
exo12_JADOST 8.45 3.23 1.89 1.25 0.45 0.260.13 85.00 30.51 16.15 10.30 5.55 3.65 1.95
exo12_OrthOST 8.97 3.20 1.57 1.25 0.39 0.260.12 82.91 28.88 15.07 9.54 5.17 3.43 1.86
MERIS8
Id 35.69 29.24 28.75 19.75 15.56 10.93 6.22 796.1 584.5 461.4 373.3 250.6 170.83 80.01
KLT 8.34 2.81 1.54 1.02 0.5 0.28 0.14 95.77 33.84 18.52 11.93 6.47 4.25 2.22
JADOST 8.39 2.28 1.29 0.74 0.36 0.24 0.13 92.49 30.45 16.06 10.28 5.71 3.86 2.04
OrthOST 7.71 2.96 1.25 0.77 0.36 0.25 0.11 92.67 29.82 15.68 10.02 5.56 3.76 1.99
exo2_KLT 8.2 3.92 2.53 1.5 0.63 0.34 0.17 113.61 46.49 27.87 18.44 9.75 6.04 2.93
exo2_JADOST 10.53 4.81 3.14 1.44 0.68 0.39 0.17 128.47 52.80 29.90 18.97 9.61 5.71 2.79
exo2_OrthOST 12.27 4.97 2.61 1.53 0.81 0.34 0.16 138.68 56.09 30.87 19.22 9.33 5.65 2.76
exo6_KLT 8.31 4.16 1.84 1.29 0.71 0.41 0.17 109.55 44.79 26.67 18.02 9.84 6.09 2.94
exo6_JADOST 8.75 4.22 2.17 1.46 0.85 0.45 0.18 113.05 46.89 28.35 19.25 10.47 6.55 3.05
exo6_OrthOST 10.18 3.47 1.87 1.24 0.55 0.35 0.17 123.22 47.09 25.79 16.51 8.67 5.39 2.67
exo7_KLT 8.84 3.97 2.06 1.35 0.71 0.46 0.19 110.39 45.8 27.74 18.67 10.06 6.21 3
exo7_JADOST 8.10 3.58 2.00 1.76 0.71 0.45 0.17 113.27 46.72 28.58 19.26 10.32 6.30 3.00
exo7_OrthOST 9.89 4.05 2.28 1.76 0.68 0.38 0.19 120.79 48.94 28.22 18.41 9.52 5.8 2.82
exo11_KLT 8.18 4.12 1.95 1.3 0.68 0.46 0.18 110.01 45.14 27.14 18.32 9.93 6.15 2.97
exo11_JADOST 8.93 3.42 2.07 1.69 0.70 0.38 0.19 113.56 46.76 28.29 18.97 10.13 6.20 2.96
exo11_OrthOST 9.8 4.25 1.97 1.56 0.6 0.33 0.15 118.26 47.54 27.19 17.15 8.69 5.34 2.64
Table 5. Bit-rate (in bpppb) vs PSNR (in dB), vs MAD, vs MSA and vs MAE of various
spectral transforms on two images. The exoi_KLT, exoi_JADOST and exoi_OrthOST
correspond respectively to exogenous KLT, JADOST and OrthOST computed with the
learning basis Li . The best (resp. worst) results of exogenous transforms at 1.0 bpppb are
bolded (resp. in italics).
and
and Optimal Applied
23
Fig. 4. PSNR (in dB) versus bit-rate (in bpppb) for various spectral transforms (KLT, JADOST
OrthOST and (left) best exogenous KLT, best exogenous JADOST, best exogenous OrthOST
or (right) worst exogenous KLT, worst exogenous JADOST, worst exogenous OrthOST). The
image is MERIS15 and the bit-rate is computed with first row: the VM9, second row: the BPE.
bit-rate (in bpppb) 0.25 0.50 0.75 1.00 1.50 2.00 3.00
mean (in dB) {PSNR(KLT)− worst exogenous PSNR(OrthOST)} 0.67 1.19 1.56 1.68 1.49 1.15 0.86
mean (in dB) {PSNR(KLT) − worst exogenous PSNR(JADOST)} 0.56 1.32 1.99 2.29 2.25 1.89 1.40
mean (in dB) {PSNR(KLT) − best exogenous PSNR(KLT)} 1.02 1.83 2.34 2.51 2.27 1.82 1.4
Table 6. Comparison of the averaged losses of PSNR with respect to the KLT for the worst
exogenous OrthOST, the worst exogenous JADOST and the best exogenous KLT. The worst
and best exogenous transforms are selected at 1.00 bpppb. The averages are computed on the
ten images.
174
algorithm by Shoham and Gersho (Shoham & Gersho, 1988) for quantization and entropy
coding, exogenous OrthOST still performs well (see Fig. 4). However, the VM9 and the
Shoham and Gersho algorithm both have a too high computational complexity for a coder
on-board a satellite. In (Gutzwiller et al., 2009), the authors propose an extension to
multicomponent images of the well-known 2-D SPIHT encoder that has not the shortcoming
of a high computational cost for bit-rate allocation.
9. Conclusion
In this chapter, we have studied the problem of finding optimal spectral transforms associated
with fixed 2D discrete wavelet transforms in coding of multi- and hyper-spectral images, for
a compression scheme that is compatible with the JPEG2000 Part 2 standard. We clarified
the criterion that gives, when minimized, an optimal transform under high-rate entropy
constraint scalar quantization hypothesis and when one scalar quantizer per subband and per
component is applied. We showed the link between the criterion and the mutual information
contrast used in independent component analysis. We derived a criterion minimized by an
orthogonal optimal transform when the data are Gaussian. Then we gave three algorithms
that return the spectral transforms that minimize the JPEG2000 compatible criterion, two
under the constraint of orthogonality — one of which assuming Gaussian data — and the
third with no constraint, but invertibility. Finally, we have tested the optimal transforms
on satellite multi- and hyper-spectral images and found that for hyperspectral images the
orthogonal optimal transform OrthOST and JADOST performs a little better than the KLT
for four distortion measures that permit to evaluate the performances of the codec in
applications of hyperspectral images like classifications or target detections. However the
computational complexity of the optimal transform is too heavy for actual applications. Last
we have presented the exogenous orthogonal quasi-optimal spectral transforms, that have a
significantly smaller complexity, and their performances in lossy coding. In future works, we
will study the problem of designing optimal spectral filters (i.e. a convolutive rather than an
instantaneous mixture) in lossy compression of multi- and hyper-spectral images.
10. References
Akam Bita, I. P., Barret, M., Dalla Vedova, F. & Gutzwiller, J.-L. (2008). Onboard com-
pression of hyperspectral images using an exogenous orthogonal quasi-optimal
transform, Proceedings of On-Board Payload Data Compression Workshop, Noordwijk
(The Netherlands).
Akam Bita, I. P., Barret, M. & Pham, D.-T. (2010a). On optimal transforms in lossy compression
of multicomponent images with JPEG2000, Signal Processing Vol. 90(No. 3): 759–773.
Akam Bita, I. P., Barret, M. & Pham, D.-T. (2010b). On optimal orthogonal transforms at high
bit-rates using only second order statistics in multicomponent image coding with
JPEG2000, Signal Processing Vol. 90(No. 3): 753–758.
Akam Bita, I. P., Barret, M., Vedova, F. D. & Gutzwiller, J.-L. (2010c). Lossy and lossless
compression of MERIS hyperspectral images with exogenous quasi-optimal spectral
transforms, Journal of Applied Remote Sensing Vol. 4: 1–15.
Antonini, M., Barlaud, M., Mathieu, P. & Daubechies, I. (1992). Image coding using wavelet
transform, IEEE Transactions on Image Processing Vol. 1(No. 2): 205–220.
and
and Optimal Applied
25
Barret, M., Akam Bita, I. P., Gutzwiller, J.-L. & Dalla Vedova, F. (2009). Lossy hyperspectral
image coding with exogenous quasi optimal transforms, Proceedings Data Compression
Conference, Snowbird (USA), pp. 411–419.
Barret, M., Gutzwiller, J.-L. & Hariti, M. (2011). Low complexity hyperspectral image coding
using exogenous orthogonal optimal spectral transform (OrthOST) and degree-2
zerotrees, IEEE Transactions on Geoscience and Remote Sensing Vol. 49(No. 5).
CCSDS-1 (2007). Image data compression, Report concerning space data system standards, CCSDS
120.1-G-1, Green book .
Chang, C.-I., Ramakrishna, B., Wang, J. & Plaza, A. (2010). Low-bit rate exploitation-based
lossy hyperspectral image compression, Journal of Applied Remote Sensing Vol. 4: 1–24.
Christophe, E., Léger, D. & Mailhes, C. (2005). Quality criteria benchmark for hyperspectral
imagery, IEEE Transactions on Geoscience and Remote Sensing Vol. 43(No. 9): 2103–2114.
Christophe, E., Mailhes, C. & Duhamel, P. (2006). Best anisotropic 3-D wavelet decomposition
in a rate-distortion sense, Proceedings of International Conference on Acoustic, Speech and
Signal Processing, Toulouse (France), pp. II–17–20.
Dragotti, P. L., Poggi, G. & Ragozini, A. R. P. (2000). Compression of multispectral images
by three-dimensional SPIHT algorithm, IEEE Transactions on Geoscience and Remote
Sensing Vol. 38(No. 1): 416–428.
Du, Q. & Fowler, J. E. (2007). Hyperspectral image compression using JPEG2000 and principal
component analysis, IEEE Geoscience and Remote Sensing Letters Vol. 4: 201–205.
Du, Q. & Fowler, J. E. (2008). Low-compexity principal component analysis for hyperspectral
image compression, International Journal of High Performance Computing Applications
Vol. 22: 438–448.
Flury, B. N. & Gautschi, W. (1986). An algorithm for simultaneous orthogonal transformation
of several positive definite symmetric matrices to nearly diagonal form, SIAM Journal
on Scientific and Statistical Computing Vol. 7(No. 1): 169–184.
Fowler, J. E. & Rucker, J. T. (2007). chap 14: 3D wavelet-based compression of hyperspectral
imagery, in C.-I. Chang (ed.), Hyperspectral Data Exploitation: Theory and Applications,
John Wiley & Sons, Hoboken.
Gersho, A. & Gray, R. M. (1992). Vector quantization and signal compression, Kluwer Academic
Publisher.
Gray, R. M. & Neuhoff, D. L. (1998). Quantization, IEEE Transactions on Information Theory Vol.
44(No. 6): 2325–2384.
Gutzwiller, J.-L.and Hariti, M., Barret, M., Christophe, E., Thiebaut, C. & Duhamel, P. (2009).
Extension du codeur SPIHT au codage d’images hyperspectrales, Proceedings of
Colloquium COmpression et REprésentation des Signaux Audiovisuels (CORESA), Tou-
louse (France).
URL: http://liris.cnrs.fr/Documents/Liris-4087.pdf
JPEG2000 (2001). JPEG2000 verification model 9.1 (technical description), ISO/IEC JTC 1/SC
29/WG 1 WG1 N2165 pp. 1–213.
MERIS (2006). MERIS detailed instrument description, European Space Agency .
URL: http://envisat.esa.int/instruments/meris/
Narozny, M., Barret, M., Pham, D.-T. & Akam Bita, I. P. (2005). Modified ICA algorithms
for finding optimal transforms in transform coding, Proceedings of IEEE International
Symposium on Image and Signal Processing and Analysis, Zagreb (Croatie), pp. 111–116.
176
Narozny, M., Barret, M. & Pham, D.-T. (2008). ICA based algorithms for computing optimal
1-D linear block transforms in variable high-rate source coding, Signal Processing Vol.
88(No. 2): 268–283.
Papoulis, A. (1984). Probability, Random Variables, and Stochastic Processes, McGraw-Hill.
Penna, B., Tillo, T., Magli, E. & Olmo, G. (2007). Transform coding techniques for lossy
hyperspectral data compression, IEEE Transactions on Geoscience and Remote Sensing
Vol. 45(No. 5): 1408–1421.
Pham, D.-T. (2004). Fast algorithms for mutual information based independent component
analysis, IEEE Transactions on Signal Processing Vol. 52(No. 10): 2690–2700.
Pham, D.-T. (2005). Entropy of random variable slightly contaminated with another, IEEE
Signal Processing Letters Vol. 12(No. 7): 536–539.
Rucker, J. T., Fowler, J. E. & Younan, N. H. (2005). JPEG2000 coding strategies for hyperspectral
data, Proceedings of International Geoscience and Remote Sensing Symposium, Seoul
(Korea), pp. 128–131.
Said, A. & Pearlman, W. A. (1996). A new, fast and efficient image codec based on set
partitioning in hierarchical trees, IEEE Transactions on Circuits and Systems for Video
Technology Vol. 6(No. 3): 243–250.
Shoham, Y. & Gersho, A. (1988). Efficient bit allocation for an arbitrary set of quantizers, IEEE
Transactions on Acoustics, Speech, and Signal Processing Vol. 36(No. 9): 1445–1453.
Taubman, D. S. (2000). High performance scalable compression with EBCOT, IEEE Transactions
on Image Processing Vol. 9(No. 7): 1158–1170.
Taubman, D. S. & Marcellin, M. W. (2002). JPEG2000: Image Compression Fundamentals,
Standards and Practice, Kluwer Academic, Place of publication.
Thiebaut, C., Lebedeff, D., Latry, C. & Bobichon, Y. (2006). On-board compression algorithm
for satellite multispectral images, Proceedings of Data Compression Conference, Snow-
bird (USA), pp. 28–30.
Usevitch, B. (1996). Optimal bit allocation for biorthogonal wavelet coding, Proceedings of Data
Compression Conference, Snowbird (USA), pp. 387–395.
Vaisey, J., Barlaud, M. & Antonini, M. (1998). Multispectral image coding using lattice VQ and
the wavelet transform, Proceedings of IEEE International Conference on Image Processing,
Chicago (USA), pp. 307–311.
Woods, J. W. & Naven, T. (1992). A filter based bit allocation scheme for subband compression
of HDTV, IEEE Transactions on Image Processing Vol. 1(No. 3): 436–440.
Yeh, P. S., Armbruster, P., Kiely, A., Masschelein, B., Moury, G. & Schafer, C. (2005). The
new CCSDS image compression recommendation, Proceedings of IEEE Aerospace
Conference, Big Sky (USA), pp. 1–8.
Part 3
Discrete Wavelet Transforms for Watermaking

11
Watermarking-Based
Image Authentication System in the
Discrete Wavelet Transform Domain
Clara Cruz Ramos, Rogelio Reyes Reyes, Mariko Nakano Miyatake and
Héctor Manuel Pérez Meana
SEPI ESIME Culhuacan, National Polytechnic Institute of México, México City,
México
1. Introduction
Nowadays, digital images and video are gradually replacing their conventional analog
counterparts. This is quite understandable because digital format is easy to edit, modify,
and exploit. Digital images and videos can be readily shared via computer networks and
conveniently processed for queries in databases. Also, digital storage does not age or
degrade with usage. On the other hand, thanks to powerful editing programs, it is very
easy even for an amateur to maliciously modify digital media and create "perfect"
forgeries. It is usually much more complicated to tamper with analog tapes and images.
Tools as digital watermarks help us establish the authenticity and integrity of digital
media and can prove vital whenever questions are raised about the origin of an image and
its content.
A digital watermarking technique embeds an invisible signal with an imperceptible form for
human audio/visual systems, which is statistically undetectable and resistant to lossy
compression and common signal processing operations. So far there some content
authentication of digital image methods, which can be classified in two groups:
watermarking based technique (Hsu & Wu, 1999) and digital signature based technique
(Friedman, 1993). Some authors had written about digital image authentication systems
(Wong, 1998; Holiman & Memos, 2000; Wong & Memon 2001; Celik, et al, 2002; Monzoy, et
al, 2007; Cruz, et al, 2008; Cruz, et al, 2009; Hernandez, et al, 2000; Lin & Chang 2001; Maeno,
2006; Hu & Chen, 2007; Zhou, et al, 2004; Lu & Liao 2003) and are classified in three
categories: complete authentication, robust authentication and content authentication (Liu &
Steinebach, 2006). Complete authentication refers to techniques that consider the whole
piece of multimedia data and do not allow any manipulation (Yeung & Mintzer, 1997; Wu &
Liu, 1998). Because the non-manipulable data are like generic messages, many existing
message authentication techniques can be directly applied. For instance, digital signatures
can be placed in the LSB of uncompressed data, or the header of compressed data. Then,
manipulations will be detected because the hash values of the altered content bits may not
match the information in the altered digital signature.
We define robust authentication as a technique that treats altered multimedia data as
authentic if manipulation is imperceptible. For example, authentication techniques, that
tolerate lossy compression up to an allowable level of quality loss and reject other
manipulations, such as tampering, belong to this category.
Content authentication techniques are designed to authenticate multimedia content in a
semantic level even though manipulations may be perceptible. Such manipulations may
include filtering, color manipulation, geometric distortion, etc. We distinguish these
manipulations from lossy compression because these perceptible changes may be
considered as acceptable to some observers but may be unacceptable to others.
A common objective for authentication is to reject the crop-and-replacement process that
may change the meaning of data. Many robust watermarking techniques in literature are
designed to be robust to all manipulations for copyright protection purpose. They usually
fail to reject the crop-and–replacement process so that they are not suitable for robust
authentication and content authentication.
An authentication system can be considered as effective if it satisfies the following
requirements:
1. Sensibility: The authenticator is sensitive to malicious manipulations such as crop-and-
replacement.
2. Robustness: The authenticator is robust to acceptable manipulations such as lossy
compression, or other content-preserving manipulations.
3. Security: The embedded information bits cannot be forged or manipulated. For
instance, if the embedded watermarks are independent of the content, then an attacker
can copy watermarks from one multimedia data to another.
4. Portability: Watermarks have better portability than digital signatures because the
authentication can be conducted directly from only received content.
5. Identification of manipulated area: Users may need partial information. The
authenticators should be able to detect location of altered areas, and verify other areas
as authentic.
Regardless of security issues, watermarking capacity is determined by invisibility and
robustness requirements. There are three dimensions shown in Figure 1. If one parameter is
determined, the other two parameters are inversely proportional. For instance, a specific
application may determinate how many bits of message are needed. After the embedded
amount is decided, it always exists a trade-off between visual quality and robustness which
must be considered. Robustness refers to the extraction of embedded bits with an error
probability equal to or approaching zero. Watermark imperceptibility (invisibility) represents
the quality of watermarked image respect to the original one. In general, if we want to make
our watermark more robust against attacks then a longer codeword or larger codeword
amplitudes will be necessary to provide better error-resistence. However, visual quality
degradation cannot be avoided. Another scenario may be that with a default visual quality,
there exists a trade-off between the information quantity of embedded message and
robustness. For instance, the fewer the message bits are embedded, the more redundant the
code word can be. Therefore, the code word has better error correction capability against noise.
It is difficult for an authenticator to know the purpose of manipulation. A practical approach
is to design an authenticator based on the manipulation method. In this work, we design an
authenticator which accepts format transformation and lossless compression (JPEG). The
authenticator rejects replacement manipulations because they are frequently used for
attacks. Our authenticator does not aim to reject or accept, in absolute terms, other
manipulation methods because the problem of whether they are acceptable or not depends
on applications.
Watermarking-Based Image Authentication System in the Discrete Wavelet Transform Domain 181
Fig. 1. Parameters of watermarking: Robustness, information quantity of embedded message

and invisibility.
2. Previous techniques for robust authentication and content authentication

In Paquet, Ward & Pitas, 2003, a novel watermarking scheme to ensure the authenticity of
digital images is presented. Their authentication technique is able to detect malicious
tampering of images even if they have been incidentally distorted by common image
processing operations. The image protection is achieved by the insertion of a secret author’s
identification key in the wavelet coefficients by their selective quantization. Their system
uses characteristics of the human visual system to maximize the embedding energy while
keeping good perceptual transparency and develop an image-dependent method to
evaluate, in the wavelet domain, the optimal quantization step allowing the tamper proofing
of the image. The nature of multiresolution discrete wavelet decomposition allows the
spatial and frequency localization of image tampering. Experimental results show that
system can detect unauthorized modification of images.
Kundur & Hatzinakos (Kundur & Hatzinakos, 1999), presented a fragile watermarking
technique for the tamper proofing of still images. A watermark is embedded in the discrete
wavelet domain by the quantization of the corresponding wavelet coefficients. The Haar
wavelet is used for the image decomposition and a pseudo-random binary sequence is
generated by a secret identification key. The rounding of the DWT coefficients to even or
odd quantization steps embeds the zeros or ones of the watermark. The embedding
locations are stored in the coefficient selection key, ckey. In addition, an image-dependent
quantization key, qkey, is introduced to improve security against forgery and monitor
specific changes to the image.
In the same line a digital image authentication procedure that allows the detection of
malicious modifications, while staying robust to incidental distortion introduced by
compression is presented in Yu, et al., 2000. A binary watermark is embedded in the wavelet
transform domain. The insertion is again done by the even or odd quantization of selected
wavelet coefficients. To increase the robustness of the scheme to image processing
operations, the authors proposed to make the embedded watermark more robust by
rounding the mean value of weighted magnitudes of wavelet coefficients to quantization
levels specified by the predetermined function Q(x,q). The same function is also used in the
blind detection process to retrieve the watermark privately by reversed quantization. In
order to distinguish malicious tampering from incidental distortion, the amount of
modification on wavelet coefficients introduced by incidental versus malicious tampering is
modeled as Gaussian distributions with small vs. large variance. The probability of
watermark detection error due to incidental alterations is shown to be smaller than the
probability of watermark detection error due to malicious tampering because they produce
comparatively smaller variance difference with the embedded marks. The authors argue
that this grants a certain degree of robustness to the system and show that their method is
able to authenticate JPEG compressed images without any access to the original unmarked
image. However, the degree of image compression allowed by the detection procedure is
not stated and the selection procedure of quantization parameters is not explained either.
In this work we develop a content authentication technique using imperceptible digital
watermarking which is robust to malicious and incidental attacks for image authentication,
embedding a digital signature as watermark. A digital signature is a set of features extracted
from an image, and these features are stored as a file, which will be used later for
authentication. To avoid the extra bandwidth needed for transmission of the signature in a
conventional way; having extracted the digital signature we applied the discrete wavelet
transform (DWT) to the image to embed the watermark in the sub band of lowest frequency,
because we want the watermark insertion to be imperceptible to the Human Visual System
and robust to common image processing such as JPEG compression and noise
contamination. The proposed system is able to extract the watermark in full blind detection
mode, which does not have access to the original host signal, and the watermark extracted
has to be re-derived from the watermarked signal, this process increases the system security.
In the security community, an integrity service is unambiguously defined as one which
insures that the sent and received data are identical. Of course, this binary definition is also
applicable to image, however it is too strict and not well adapted to this type of digital
document. Indeed, in real life situations images will be transformed, their pixel values will
therefore be modified but not the actual semantic meaning. In other words, the problem of
image authentication is released on the image content, for example: when modifications of
the document may change its meaning or visually degrade it. In order to provide an
authentication service for still images, it is important to distinguish between malicious
manipulations, which consist of changing the content of the original image (captions, faces,
etc.) and manipulations related to the usage of an image such as format conversion,
compression, noise, etc.
Unfortunately this distinction is not always clear; it partly depends on the type of image and
its usage. Indeed the integrity criteria of an artistic master piece and a medical image will
not be the same. In the first case, a JPEG compression will not affect the perception of the
image, whereas in the second case it may discard some of the fine details which would
render the image totally useless. In the latter case, the strict definition of integrity is
required. We applied the proposed algorithms in to grayscale and color no medical images.
3. Proposed watermarking algorithm

The figure 2(a) shows a general block diagram to the watermark insertion where we can see
that original image is divided in non-overlapping blocks, we extracted a digital signature
from each block then we insert a signature as watermark in the same block, finally all the
watermarked blocks form the watermarked image. Figure 2(b) shows a general block
diagram to the watermark extraction process from the watermarked block where we can see
that is not necessary to now the original image to extract the digital signature. Finally in the
verification process we compare the extracted watermark and the digital signature to
determine if the image has been modified, or not.
Fig. 2. (a) Watermark insertion system; (b) Watermark extraction system.
3.1 Digital signature generation

The algorithm used to extract the digital signature was proposed in Fridrich, 1999, and used
by Chen, et al., 2001. The goal of this algorithm is to make a method for extracting bits from
image blocks so that all similarly looking blocks, whether they are watermarked or attacked,
will produce almost the same bit sequence of length N. Method is based on the observation
that if a low-frequency DCT coefficient of an image is small in absolute value, it cannot be
made large without causing visible changes to the image. Similarly, if the absolute value of a
low-frequency coefficient is large, we cannot change it to a small value without influencing
the image significantly. To make the procedure key-dependent, we replace DCT modes with
low-frequency DC-free (i.e., having zero mean) random smooth patterns generated from a
secret key (with DCT coefficients equivalent to projections onto the patterns). For each
image, we calculate a threshold Th so that on average 50% of projections have absolute
value larger than Th and 50% are in absolute value less than Th. This will maximize the
information content of the extracted N bits.
Given an image I, we divide it into blocks of 16x16 pixels (for large images, larger block
sizes could be used) as showed in Figure 3. Using a secret key K (a number uniquely
associated with an author, movie distributor, or a digital camera) we generate N random
matrices with entries uniformly distributed in the interval [0, 1]. Then, a low-pass filter is
repeatedly applied to each random matrix to obtain N random smooth patterns. All patterns
are then made DC-free by subtracting the mean value from each pattern. Considering the
block and the pattern as vectors, the image block B is projected on each pattern Pi, 1< i < N
and its absolute value is compared with a threshold Th to obtain N bits bi:
| . | < ℎ = 0
| . | ≥ ℎ = 1 (1)
Since the patterns Pi have zero mean, the projections do not depend on the mean gray value
of the block and only depend on the variations in the block itself. The distribution of the
projections is image dependent and should be adjusted accordingly so that approximately
half the bits bi are zeros and half are ones. This will guarantee the highest information
content of the extracted N-tuple. This adaptive choice of the threshold becomes important
for those image operations that significantly change the distribution of projections, such as
contrast adjustment.
Fig. 3. Digital signature extraction process.
3.2 Wavelet transform for image signals

Two-dimensional DWT leads to a decomposition of approximation coefficients at level j in
four components: the approximation at level j + 1, and the details in three orientations
(horizontal, vertical, and diagonal).
Figure 4 describes the basic decomposition steps for images.
Fig. 4. Subband decomposition using 2D-DWT.

The subbands labeled LH1, HL1, and HH1 represent the finest scale wavelet coefficients. In
the present work, the wavelet transform is realized with Daubechies Wavelets of order 2.
Using this wavelets, the image is decomposed into four subbands: LL1, LH1, HL1 and HH1.
3.3 Watermark embedding algorithm

Because we want the embedded watermark to be imperceptible to the Human Visual
System (HVS) and robust to common image processing such as JPEG compression and
contamination, we implement the algorithm proposed by Inoue, et al. 2000. In this method
information data can be embedded in the lowest frequency components of image signals by
using controlled quantization process. The data is then extracted by using both the
quantization step-size and the mean amplitude of the lowest frequency components without
access to the original image.
Once the digital signature is extracted, we applied the discrete wavelet transform (DWT) to
embed the watermark, the subband LL1(i,j) is divided into small subblocks Bk with the size
of bx×by and calculate the mean Mk of the wavelet coefficients of Bk. A quantization step-size
which is called the embedded intensity Q=5 is used, then we calculate the mean of the
wavelet coefficients of Bk. The watermark information is embedded into the subblock Bk
modifying the quantization value q and adds δMk to the wavelet coefficients of Bk, as
described in detail (Inoue, et al. 2000). Finally we construct the watermarked image using
the inverse wavelet transform.
Figure 5 illustrates the embedding process; the data wk =0 or 1 into a subblock Bk when
bx=by =2 and Q =5.
Fig. 5. Watermark insertion process.

3.4 Watermark extracting algorithm

We can extract the embedded data w by using the parameters n (decompose level), bx, by, Q
and LM’. Let I’ be the watermarked image, we decompose I’ for the scale 1 and obtain the
lowest frequency components LL1’(i,j). Then we divide LL1’(i,j) into subblocks Bk with the
size of bxxby and compute the mean Mk’ of Bk and find the quantization value S from
= [( ) ⁄ ] (2)
Then, we extract the embedded binary data wk as follows: if S is an even number, then wk
=0, otherwise wk =1.
3.5 Authentication process

After the watermark wk and the digital signature sequences are extracted from the
watermarked image I’, we determines a threshold (Thv) to decide using an XOR operation if
the block is tampered or not, which is expressed in equation (3).
∑ ⨂ ≤ ℎ ℎ
(3)
∑ ⨂ ≥ ℎ
Threshold Thv was determined through trial and error; resulting value of Thv was 4, it
means that if bits number of digital signature extracted of the block authenticated has at
least 12 of 16 bits equal, the block is consider as authentic else it is consider as modified.
Although the block is considered modified, sometimes you do not get the same 16-bit digital
signature extracted with respect to the original signature can be caused by any intentional
modification, which is why we proposed the following process check.
3.6 Verification process

After the watermark wk from the watermarked image I’ is extracted, we compare it with the
digital signature extracted from I’. If they have some different blocks we make an
“difference image”(Idif).
According to evaluation carried out using 200 images, in authentication process, the
following conclusion was reached: when error blocks are present in regions non intentional
modified, these blocks are presented in isolation, as shown in figures 6(a,b), however in the
case of images modified intentionally error blocks are detected in concentrated form as
shown in figures 6(c,d), so when error blocks are detected isolated, means that region is
authentic otherwise it is non-authentic. Therefore to establish a criterion to determine
whether the change at a block is intentional or unintentional, we define the following rule:
If there are more than three consecutive error blocks in the region of Idif the image was
intentionally modified, otherwise the change was made by common signal processing as
JPEG compression or noise. Applying the concept of connectivity between the 8 neighbors
of error blocks, it can help us to identify intentionally modified regions of which are not.
This criterion is represented mathematically by the equation (4).
ℎ ≤ 3
(4)
> 3
were represents an error block, so if there are more than three consecutive error blocks in
the region, it has been intentionally modified.
(a) Isolated error blocks (b) Isolated error blocks
(c) Concentrated error blocks (d) Concentrated error blocks
Fig. 6. (a,b) Non intentional modified image; (c,d) Intentionally modified image.
4.1 Digital signature robustness
To evaluate the robustness of the bit extraction procedure, we subjected the test image
"Barbara" with 512x512 pixels and 256 gray levels to various image processing operations
available in specialized commercial image manipulation software (we used Photoshop). The
test image "Barbara" had 1024 blocks of 16x16 pixels. We extracted N=16 bits from each
block for the original image and the manipulated image and calculated the average number
of error over all 1024 blocks. The results are shown in Table 1.
Image name Shine (%) Average number of

error bits
Barb_100 10 0
Barb_200 20 0
Barb_300 30 0
Barb_400 40 0
Barb_500 50 0
Contrast (%)
Barb_10 10 0
Barb_20 20 0
Barb_30 30 0
Barb_40 40 0
Barb_50 50 0
Ecualization 0.015
JPEG compression
Quality factor Average number of
error bits
20, 25, 30, 35, 40, 45, 50, 55, 60, 65, 70, 2.15, 1.98, 1.78, 1.69,
75, 80, 85, 90, 95 and 100 1.51, 1.35, 1.30, 1.22,
1.11, 1.08, 0.98, 0.91,
0.75, 0.67, 0.5, 0.36
and 0.07
Impulsive noise
Intensity PSNR Average number
of error bits
0.0010 35.8258 0.58
0.0020 32.3782 0.98
0.0030 30.8060 1.16
0.0040 29.6593 1.23
0.0050 28.6105 1.87
0.0060 27.8483 2.01
0.0070 27.1725 2.22
0.0080 26.5314 2.53
0.0090 26.0121 2.85
0.0100 25.6005 2.92
Table 1. Average number of error recovered bits out of 16 bits after some image processing
operations.
4.2 Semi-fragile watermark system performance

In order to confirm that the proposed digital watermark system is effective, we
implemented some numerical experiments with attacks such as JPEG compression,
impulsive and Gaussian noise and photomontage. Experimental results show that the
algorithm is capable to determine whether the image has been altered. The algorithm was
evaluated using 200 standard images. These images are 8 and 24 bits per pixel (bpp)
grayscale and color images, which were 512x512 and 128x128 pixels in size showed in figure 7.
Another advantage of this algorithm is that the size and texture of the image doesn´t affect
on the correct operation of the system.
(a) 8 bits per pixel (bpp) grayscale image (b) 8 bits per pixel (bpp) grayscale image
(c) 24 bits per pixel (bpp) color image (d) 24 bits per pixel (bpp) color image
Fig. 7. Some images used in the experimental process.
4.2.1 Watermarked image quality

In our system we use the peak signal to noise ratio (PSNR) to mesure the degradation of the
image quality caused by watermarking, this value is given by (5),
= 10 (5)
where is the mean square of the difference between the original image and the
watermarked one.
Figure 8 shows some examples of original images (in grayscale and color) together with
their respective watermarked images and PSNR values, where we can see that watermarked
images are to perceptually very similar to the original version. In table 2 PSNR values of
some grayscale and color images are shown, where we can observe that the average PSNR
value in the grayscale image is 45 dB’s and in the color image is 50 dB’s, so we can conclude
that degradation in the watermarked image is not perceptible.
Original grayscale image Watermarked grayscale image PSNR=45 dB
Original color image Watermarked color image PSNR=49.80 dB

Fig. 8. Watermarked image quality.
Grayscale PSNR Color PSNR

watermarked image (dB´s) watermarked image (dB´s)
Barbara 45.059944 Plane 49.806491
Boat 44.966452 Mountain 49.848597
Bridge 45.007931 Lake 49.810409
Camera 45.056509 Chiles 49.853756
Chiles 45.003896 People 49.848703
Goldhill 44.959921 Lena 49.815038
Lena 44.958274 Home 50.342928
Baboon 45.041577 Girl 49.568472
Bird 44.962013
Table 2. PSNR values of some grayscale and color watermarked tested images.
4.2.2 Robustness against JPEG compression

The authenticator is sometimes expected to pass only those images that are compressed by
JPEG up to a certain compression ratio or quality factor (fc). For example, if the image is
JPEG compressed below to image quality 75 (The Mathworks, 2008), the image is acceptable,
otherwise, if it is more compressed, it will fail the test. The argument for failing highly
compress images is that such images usually have poor quality and should not be
considered as authentic. To satisfy this need, we calculate the increase of the number of the
“different” signature bits after compression (error blocks). The number of the error blocks
increases if the image is more compressed. We can set a threshold on this change to reject
those images that have too many error blocks.
If the error blocks are isolated, we apply equation (4) to determinate if those blocks are
result of a JPEG compression, however, if they are concentrated we are talking about an
intentional attack. We called to this process “verification” and it helps us to differentiate
between an intentional or non intentional attack.
Figure 9 shows the extracted results from the authentication JPEG compressed watermarked
images with quality factors higher than 75 and their corresponding verified image; we can
see that compressed images with quality factors higher than 75 have their error blocks
(white blocks) isolated; consequently, before the verification process they are considered as
not attacked.
Error blocks of “chiles”, with fc=75 Verified image

Authenticated image (not attacked)
Error blocks of “boat”, with fc=80 Verified image

Authenticated image (not attacked)
Fig. 9. Tampered regions detection of the JPEG compressed images.
Table 3 shows some compression ratio where the JPEG compressed watermarked image is
considered as authentic by the system. In this table we can see that in grayscale
watermarked images were considered as authentic when their quality factor of JPEG
compression was higher than 75 and in the color compressed images with a quality factor
higher than 70.
4.2.3 Robustness against additive and Gaussian noise

We contaminate watermarked image with different levels of additive and Gaussian noise to
simulate the communication channel noise. Tables 4 and 5 show the highest density and
variance value of additive and Gaussian noise in grayscale and color images before the
system considers the error blocks detected as intentionally tampered, these results indicate
that the system is efficient in front of impulsive noise attacks because it supports a density=
0.002 which produces a PSNR average value equal to 32 dB between watermarked image
and contaminated watermarked image; a similar case occurs whit the Gaussian noise; the
highest variance that the system accepts is 0.00011 before it considered watermarked
contaminated image as intentionally modified.
Quality Error Original size/ Bits/

factor blocks compresion pixel
Grayscale
watermarked images
Boat 80 5 257k/38.7k 1.20
Bridge 75 18 65k/14.6k 1.79
Camera 80 14 65k/10.3k 1.26
Chiles 75 18 257k/31.4k 0.97
Lena 75 4 65k/10.5k 1.29
Baboon 80 24 257k/72.1k 2.24
Bird 80 11 65k/7.56k 0.93
Color
watermarked images
Plane 70 13 193k/11k 0.45
Home 70 5 193k/9k 0.37
Girl 75 17 193k/10k 0.41
Chiles 65 17 193k/14k 0.58
Lake 70 14 193k/14k 0.58
Lena 65 2 193k/11k 0.45
Mountain 75 12 193k/12k 0.49
People 70 11 193k/9k 0.37
Table 3. Compression ratio of some JPEG compressed images considered as authentics.
Grayscale Density Error PSNR Color Density Error PSNR

watermarked blocks (dB´s) Watermarked Blocks (dB´s)
image image
Barbara 0.002 27 32.6765 Plane 0.0016 9 33.0454
Boat 0.002 30 32.9256 Home 0.0016 15 33.5461
Bridge 0.002 17 32.1300 Girl 0.0015 18 32.4259
Camera 0.002 36 32.4250 Chiles 0.0024 8 31.4738
Chiles 0.002 26 32.0939 Lake 0.0018 7 32.0180
Goldhill 0.002 11 32.3100 Lena 0.0025 12 31.1327
Lena 0.002 14 32.0448 Mountain 0.0015 18 33.0446
Baboon 0.002 24 32.7793 People 0.0015 26 32.2085
Bird 0.0009 18 35.7967
Table 4. Test to resistance to impulsive noise from grayscale and color watermarked images.
Grayscale Variance Error PSNR Color Variance Error PSNR

watermarked blocks (dB´s) Watermarked blocks (dB´s)
image image
Barbara 0.00011 48 39.5594 Plane 0.00031 14 35.2396
Boat 0.0001 36 40.0032 Home 0.00027 15 35.5625
Bridge 0.00014 22 38.8350 Girl 0.00027 20 35.9495
Camera 0.00011 29 39.5884 Chiles 0.00033 21 35.1859
Chiles 0.00011 37 39.5761 Lake 0.00027 10 35.5499
Lena 0.00031 16 35.2449
Mountain 0.00027 24 35.5431
People 0.00027 23 35.8596
Table 5. Test to resistance to gaussian noise from grayscale and color watermarked images.
4.2.4 Robustness against photomontage

Of course, an important aspect of our system is its ability to localize tampered regions into
the image. For that reason, we have tampered the previously watermarked Bird and lake
images and evaluated the ability of our system to detect. We found that the ability of our
system to detect tampering is excellent (Figure 10) because our system detected correctly
which group of blocks were modified intentionally and which were not into the
watermarked image, based on the assumption explained in section 2.5. To tamper the
images we used Photoshop. Figures 10(a) to 10(d) show the results of this evaluation in
grayscale images and figures 10(e) to 10(g) the results of color images. Figures 10(c) and
10(g) show by white blocks the tampered detected by our system where we can see that its
location is correct comparing 10(a) vs. 10(b) and 10(e) vs. 10(f) where the first are the
watermarked images and the others are the tampered watermarked images. Finally in figure
10(d) we see that the verification is working well because it eliminates the isolated error
blocks which were caused by the processing image.
5. Conclusion
The transition from analog to digital technologies is widely used, with the higher capacity of
storage devices and data communication channels, multimedia content has become a part of
our daily lives. Difital data is now commonly used in many areas such as education,
entertainment, journalism, law enforcement, finance, health services, and national defense.
The low cost of reproduction, storage, and distribution has added an additional dimension
to the complexity of the problem. In a number of applications, multimedia needs to be
protected for several reasons. Watermarking is a group of complementary technology that
has been identified by content provider to protect multimedia data.
In this paper we have successfully developed a robust digital signature algorithm which is
used as a semi-fragile watermarking algorithm for image authentication. The highest
advantage of this combination besides the digital signature robustness and the watermark
image imperceptibility, is that is not necessary an additional band width to transmit the
digital signature, since this is embedded in the host image as a watermark. Besides to the
extraction and authentication process, we propose a verification process, which helps us to
differentiate between an intentional or non intentional modification applying the concept of
connectivity between the 8 neighbors of error blocks.
(a) Watermarked image (b) Tampered image
(c) Authentication of the altered image (d) Verification of the authenticated image
(e) Watermarked image (f) Tampered image
(g) Authentication of the altered image

Fig. 10. Authentication and verification process of a tamper watermarking grayscale and
color image.
Numerical experiments show that this algorithm is robust to JPEG lossy compression, the
lowest acceptable JPEG quality factor is 75 for grayscale images and 70 for color images. In the
case of impulsive noise, verification system determines that a watermarked image has no-
intentional modification if its density value is less than 0.002 which produce a PSNR average
value equal to 32 dB between watermarked image and contaminated watermarked image; a
similar case occurs with the Gaussian noise; the highest variance that the system accept is
0.00011 before it consider watermarked contaminated image as intentionally modified.
An important characteristic of this system besides its robustness against common signal
processing is its capacity to detect the exact tampered locations, which are intentionally
modified. Several watermarking systems using digital signature had been reported but they
aren’t robust to JPEG compression neither to modifications caused by common signal
processing.
Finally it is important to mention that the watermarked images generated by the proposed
algorithm are secure because the embedded watermarks are dependent on their own
content.
6. Acknowledgment
This work is supported by the National Polytechnic Institute of México.
7. References
Celik, M.; Sharma, G.; Saber, E. & Tekalp, A. (2002), Hierarchical Watermarking for Secure
Image Authentication with Localization, IEEE Transactions on Image Processing, Vol.
11, No. 6, pp. 585–595.
Chen, T.; Wang, J. & Zhou, Y. (2001), Combined Digital Signature and Digital Watermark
Scheme for Image Authentication, Info-tech and Info-net, 2001. Proceedings. ICII 2001 -
International Conferences on, Vol. 5, pp. 78-82, Print ISBN: 0-7803-7010-4.
Cruz, C. ; Reyes, R.; Nakano M. and Pérez, H. (2009), Image Authentication Scheme Based
on Self-embedding Watermarking, CIARP '09 Proceedings of the 14th Iberoamerican
Conference on Pattern Recognition: Progress in Pattern Recognition, Image Analysis,
Computer Vision, and Applications, ISBN: 978-3-642-10267-7.
Cruz, C.; Reyes, R.; Mendoza, J.; Nakano, M. and Pérez, H. (2008), A Novel Verification
Scheme for watermarking based Image Content Authentication Systems,
Telecommunications and Radio Engineering, vol. 67, no. 19, pp. 1777-1790, 2008,
ISSN:0040-2508, http://begelhouse.com
Fridrich, J. (1999), Robust Bit Extraction from Images, Proceedings of IEEE International
Conference on Multimedia Computing and Systems (ICMCS’99), Vol. 2, pp. 536-
540, ISBN:0-7695-0253-9.
Friedman, G.L. (1993), The Trustworthy Digital Camera: Restoring Credibility to the
Photographic Image, IEEE Transactions on Consum. Elec., Vol. 39, pp. 905–910.
Hernández, V. ; Cruz, C. ; Nakano M. and Pérez, H. (2000), Algoritmo de Marca de Agua
Basado en la DWT para Patrones Visualmente Reconocibles, IEEE Latin America
Transactions, Vol. 4, No. 4, June 2006.
Holiman, M. & Memos N. (2006), Counterfeiting Attacks on Oblivious Block-Wise
Independent Invisible Watermarking Scheme, IEEE Transactions on Image Processing,
Vol. 9, No. 3, pp. 432-441.
Hsu, C. T. & Wu, J.I. (1999). Hidden Digital Watermarks in Images, IEEE Transactions on
Image Processing, Vol. 8, pp. 58–68.
Hu, Y. & Chen, Z. (2007), An SVD-Based Watermarking Method for Image Authentication,
Proceedings of the Sixth International Conference on Machine Learning and Cybernetics,
pp. 1723-1728, Hong Kong, 19-22 August (2007).
Inoue, H.; Miyazaki, A. & Katsura, T. (2000), A Digital Watermark for Images Using the Wavelet
Transform, Journal Integrated Computer-Aided Engineering, Vol. 7, No. 2, pp. 105-115.
Kundur, D. & Hatzinakos, D. (1999). Digital watermarking for telltale tamper proofing and
authentication, Proceedings of the IEEE, Vol. 87, No.7, pp. 1167–1180.
Lin, C. & Chang S. (2001), A Robust Image Authentication Method Distinguishing JPEG
Compression from Malicious Manipulation, IEEE Transactions on Circuits and
systems of Video Technology, Vol. 11 No. 2, pp. 153-168.
Liu, H. & Steinebach, M. (2006), Semi-Fragile Watermarking for Image Authentication with
High Tampering Localization Capability, Proc. of Int. Conf. Automated Production
of Cross Media Content for Multi-Channel Distribution, ISBN:0-7695-2625-X.
Lu, C. & Liao, H. M. (2003), Structural Digital Signature for Image Authentication: An
Incidental Distortion Resistant Scheme, IEEE Transactions on Multimedia, Vol. 5, No.
2, pp. 161-173.
Maeno, K.; Sun, Q.; Chang, S. & Suto, M. (2006), New Semi-Fragile Image Authentication
Watermarking Techniques Using Random Bias and Nonuniform Quantization,
IEEE Trans. on Multimedia, Vol. 8, No. 1, pp. 32-45.
Monzoy, M.; Salinas, M.; Nakano, M. & Pérez, H. (2007), Fragile Watermarking for Color
Image Authentication, 4th Int. Conf. Electrical and Electronic Engineering (ICEEE
2007), pp. 157-160.
Paquet, H. A.; Ward, R. K. & Pitas, I. (2003). Wavelet packets-based Digital Watermarking
for Image Verification and Authentication, Journal Signal Processing - Special section:
Security of data hiding technologies archive, Vol. 83 Issue 10, Amsterdam, The
Netherlands.
The Mathworks. Inc. (2008), Imwrite: Functions (Matlab functions references), Matlab help,
Ver. 7.6.0.324.
Wong, W. P. (1998), A Public Key Watermark for Image Verification and Authentication,
Proceedings of the IEEE Int. Conf. Image Processing, pp. 425-429.
Wong, W. P. & Memon, N. (2001), Secret and Public Key Image Watermarking Schemes for
Image Authentication and Ownership Verification, IEEE Transactions on Image
Processing, Vol. 10, No. 10, pp. 1593-1601.
Wu, M. & Liu, B. (1998), Watermarking for image authentication, Image Processing, 1998. ICIP 98.
Proceedings. International Conference on. Vol. 2, pp. 437–441, Print ISBN: 0-8186-8821-1
Yeung, M. & Mintzer, F., (1997), An Invisible Watermarking Technique for Image
Verification, Image Processing, International Conference on, Vol. 2, pp. 680, ISBN: 0-
8186-8183-7.
Yu, G.J. ; Lu, C.-S. ; Liao, H.-Y. M. & Sheu, J.-P. (2000). Mean quantization blind
watermarking for image authentication, IEEE International Conference on Image
Processing (ICIP’2000), Vol. III, pp. 706–709, Vancouver, BC, Canada.
Zhou, X.; Duan, X. & Wang, D. (2004), A Semi-Fragile Watermark Scheme for Image
Authentication, Proc. of Int. Conf. Multimedia Modeling Conference, pp. 374 – 377,
Print ISBN: 0-7695-2084-7.
12
Application of Discrete
Wavelet Transform in Watermarking
Corina Nafornita and Alexandru Isar
“Politehnica” University of Timisoara,
Romania
1. Introduction
Proliferation of multimedia data on the Internet and the ease of copying this data have
brought an interest for copyright protection (Cox et al., 2002). During transmission, data can
be protected using encryption; however after decrypting it, it is no longer protected. As an
alternative to encryption, watermarking has been proposed as a means of identifying the
owner, by secretly embedding an imperceptible signal into the host signal (Cox, 2005) – see
Fig. 1.
Cover work X0 Watermarked

Watermark work Xw
embedding ε
Watermark W
0100100010...
Key K Data embedding algorithm
Fig. 1. Watermark embedding. The watermark is embedded using a secret or public key,
making invisible changes to the cover work.
The main properties of a watermarking system are perceptual transparency, robustness,
security, and data hiding capacity (Cox et al., 1997). Some of the terms used in
watermarking are (Cox et al., 2002):
- The original data where the watermark is to be inserted is referred to as host or cover
work.
- The hidden information is called payload.
- Visible watermarks are visual patterns (images, logos) inserted or overlaid on
images/video. Visible watermarks are applied to photos publicly available on the web,
to prevent commercial use of such images. One example of visible watermarking has
been implemented by IBM for the Vatican library (Braudaway et al., 1996).
- Most watermarking systems involve making the watermark imperceptible.
- The key is required for embedding the watermark. If the same key is used for retrieving
the watermark, the system is private, while if another key is used to retrieve it, the
system is known as public.
- If the cover work is required at the detector, the system is informed (non-blind); if it’s
not required at the detector, the system is blind.
- Watermarking systems are robust or fragile. Robust watermarks should resist any
modifications and are designed for copyright protection. Fragile watermarks are
designed to fail whenever the cover work is modified and to give some measure of the
tampering. Fragile watermarks are used in authentication.
Most of existing watermarking systems proposed in the literature can be classified
depending on the watermarking domain, where the embedding takes place: spatial domain
techniques (Nikolaidis & Pitas, 1998), where the pixels are directly modified, or transform
domain techniques.
The majority of watermarking algorithms operate based on the spread spectrum (SS)
communication principle. A pseudorandom sequence is added to the host signal in some
critically sampled domain and the watermarked signal is obtained by inverse transforming
the modified coefficients. Typical transform domains are the Discrete Wavelet Transform
(DWT), the Discrete Cosine Transform (DCT) and the Discrete Fourier Transform (DFT). The
DWT based algorithms usually produce watermarked images with the best balance between
visual quality and robustness due to the absence of blocking artefacts (Nafornita, 2008).
Watermarks can be robust or fragile, depending on the application. For copyright
protection, robustness is required. This can be assured with encoding of the watermark
using a repetition code or an error correcting code. Robustness is increased with the increase
of the correction capacity of the code. Despite of their efficient use in telecommunications,
turbo codes have been rarely used in watermarking (Abdulaziz et al., 2002, Serdean et al.,
2003, Balado & Perez-Gonzalez, 2001, Nafornita et al., 2009).
At the embedding side, the watermark can be added to coefficients of known robustness
(large valued coefficients) or perceptually significant regions (Cox, 2005), such as contours
and textures of an image. This can be done empirically, selecting larger coefficients (Cox et
al., 1997) or using a thresholding scheme in the transform domain (Podilchuk & Zeng, 1998,
Nafornita et al., 2005). Another approach is to insert the watermark in all coefficients of a
transform, using a variable strength for each coefficient (Barni et al., 2001). Hybrid
techniques, based on compression schemes, embed the watermark using a thresholding
scheme and variable strength (Podilchuk & Zeng, 1998). The performance of such a system
depends on the quality of the wavelet transform.
This chapter will focus on the application of the wavelet transforms in robust watermarking
for static images. We will present the classical techniques of watermarking; starting with the
spread spectrum DCT based watermarking system proposed by Cox et al. (Cox et al., 1997)
and continuing with those proposed in the wavelet domain.
Other wavelet transforms as the Double Tree Complex Wavelet Transform (DTCWT)
(Selesnick et al., 2005) or the Hyperanalytic Wavelet Transform (HWT) (Nafornita et al.,
2008, Firoiu et al., 2009) could also be considered. The advantages of such transforms
compared to DWT are: quasi-shift invariance and enhanced directional selectivity. The data
hiding capacity increases with the increase of redundancy (4x for DTCWT and HWT). We
will compare the efficiency of those wavelet transforms in watermarking.
2. Watermarking methods
Most techniques embed the watermark in a transform domain as mentioned before. Early
techniques have used the Discrete Cosine Transform. One of the most influential
Application of Discrete Wavelet Transform in Watermarking 199
watermarking works is a spread spectrum approach proposed in (Cox et al., 1997). They
argue that the watermark be placed explicitly in the perceptually most significant
components of the data, and that the watermark be composed of random numbers drawn
from a Gaussian distribution N ( 0,1 ) , in order to make it invisible and robust to attacks:
v′ ( i ) = v ( i ) ( 1 + α w ( i ) ) (1)
where v(i) is the DCT coefficient to be watermarked, w(i) is the watermark bit, α is the
embedding strength and v’(i) is the watermarked coefficient. Detection is made using the
similarity between the original W and extracted Ŵ watermarks:
ˆ
ˆ = W ⋅W
(
sim W , W
Wˆ ⋅W
ˆ
) (2)
The fact that the transform is performed over the entire image increases the computation
time. Other methods have been proposed that use the block-based DCT transform, just like
in the JPEG compression (see for example Podilchuk & Zeng, 1998).
Other authors have proposed the use of the Discrete Fourier Transform or its variant – the
Fourier-Mellin transform. This is useful in order to perform phase modulation between the
watermark and the original signal (Ó Ruanaidh et al., 1996). The phase is more important
than the amplitude; hence it will be difficult for an attacker to remove the watermark. Phase
modulation often possesses superior noise immunity in comparison with amplitude
modulation. Many watermarking techniques use DFT amplitude modulation because the
watermark will be translation invariant. The DFT is more often used in its derived forms
such as the Fourier-Mellin transform. This Fourier-Mellin transform approach has arisen out
of the need for Rotation, Scale and Translation invariant (RST-invariant) watermarking
techniques. It involves creating a Log Polar map of the DFT amplitudes of the image, where
the embedding takes place. This method is said to be extremely RST invariant and uses a
RST invariant watermark (Lin et al., 2001, Ó Ruanaidh & Pun, 1998).
3. Watermarking using wavelets

3.1 Discrete wavelet transform methods
There are different approaches to embed the watermark in the wavelet domain. Almost all
methods rely on masking in some way the watermark, either by selecting a few coefficients,
or using adaptive embedding strength.
Podilchuk & Zeng, 1998 propose an image-adaptive (IA) approach. They use the just
difference noticeable difference (JND) to determine the image dependent perceptual mask
for the watermark. They applied this method in both DCT and wavelet domain:
 I u , v + JNDu , v × wu , v , if I u , v > JNDu , v

I *u,v =  (3)
Iu ,v , otherwise
I u , v are the coefficients of the original image, wu , v are the watermark bits, and JNDu , v are
the JND values computed using visual models. In the case of DCT, they are computed using
Watson’s perceptual model; for the wavelet domain, the weight is computed for each
frequency band based on typical viewing conditions. Detection is made using correlation
between the image difference and the watermark sequence. This method is more robust than
the spread-spectrum method by Cox et al., 1997. Although more robust than IA-DCT, the
IA-W method does not take into account perceptual significant regions, so the watermark
can be erased from perceptually insignificant coefficients. For example, low-pass filtering
will affect the watermark inserted in high frequency components.
Xia et al., 1998 propose a watermarking algorithm using the Haar mother wavelet, and two
levels of decomposition. A pseudo-random sequence is added to the highest coefficients not
located in the lowest resolution:
β
f ′ ( m , n ) = f ( m , n ) + α ⋅ f ( m , n ) wi (4)
where α is the watermark strength, and β is the amplification for large coefficients. This
algorithm concentrates most of the energy in edges and textures, which are the coefficients
in detail subbands. This increases the invisibility of the watermark, because human
observers are less sensitive to change in edges and textures compared to changes in smooth
areas of an image. More watermarks are inserted in each subband, and detection is done
hierarchically, for each resolution level, using intercorrelation between original watermark
and the difference of the two images. The method is robust to a series of distortions, but
low-pass and median filtering affect the watermark.
Kundur & Hatzinakos, 1998 use the Daubechies wavelet family to compute the DWT on
three levels of decomposition. The watermarking algorithm selects in a pseudo-random
manner the embedding locations from the detail subbands. The authors state that the
spread-spectrum technique is not appropriate for transmitting the watermark because the
correlator used for watermark detection is not effective in the presence of fading. Hence,
they use quantization for embedding the watermark bits. To increase robustness, they use a
reference watermark in order to estimate if the watermark bit has been embedded (Kundur
& Hatzinakos, 2001).
One of the popular methods is the one proposed by Barni et al., 2001. The watermark is
masked according to the characteristics of the human visual system (HVS), taking into
account the texture and the luminance content of all the image subbands. For coefficients
corresponding to contours of the image a higher strength is used, for textures a medium
strength is used and for regions with high regularity a lower strength is used, in accordance
with the analogy water-filling and watermarking (Kundur, 2000).
The image I, of size 2M×2N, is decomposed into 4 levels using Daubechies-6 wavelet
mother, where I θl is the subband from level l∈{0, 1, 2, 3}, and orientation θ∈{0, 1, 2, 3}
(horizontal, diagonal and vertical detail subbands, and approximation subband). A
pseudorandom binary (±1) sequence is casted into 2D binary watermarks, each of size
MN/4l, xθl . The watermark is embedded in all coefficients from level l=0 by addition
Iɶlθ ( i , j ) = I θl ( i , j ) + α wθl ( i , j ) xθl ( i , j ) (5)
where α is the embedding strength and wθl ( i , j ) is half of the quantization step:
qθl ( i , j ) = Θ ( l ,θ ) Λ ( l , i , j ) Ξ ( l , i , j )
0.2
(6)
as it is presented in the following figure.

Watermark
IDWT
DWT
Original Marked
Mask
Fig. 2. Watermark embedding in the wavelet domain (Barni et al., 2001). The watermark is
embedded in the first resolution level using a perceptual mask.
This is a product of three factors: sensitivity to noise, local brightness and texture activity
around a pixel. They are computed as follows:
1.00 l = 0
 
 2 , θ = 1   0.32 l = 1
Θ ( l ,θ ) =  ⋅  (7)
 1 otherwise  0.16 l = 2
0.10 l = 3 
Λ ( l, i , j ) = 1 + L '( l, i , j ) (8)
(
L ( l , i , j ) = I 33 1 + i 2 3 − l  ,1 +  j 2 3 − l  256 ) (9)
3−l 2 1 2
Ξ ( l , i , j ) = ∑ 16 − k ∑ ∑  Iθ ( y + i
k +l 2 k , x + j 2 k ) 
k =0 θ =0 x ,y =0 (10)
{
⋅ Var I ( 1 + y + i 2
3
3
3−l
,1 + x + j 2 3−l
)} x = 0 ,1
y = 0 ,1
The texture activity around a pixel is composed by the product of two contributions; the first
is the local mean square value of the DWT coefficients in all detail subbands and the second
is the local variance of the 4th level approximation image. Both are computed in a small 2×2
neighborhood corresponding to the location (i, j) of the pixel. The first contribution is the
distance from the edges, and the second one is the texture. This local variance estimation is
computed with a low resolution.
Detection is made using the correlation between the marked DWT coefficients and the
watermarking sequence to be tested for presence (the original image is not needed):
l l
2 M /2 − 1 N /2 − 1
ρ ( l ) = 4l ∑ ∑ ∑ Iɶlθ ( i , j ) xlθ ( i , j ) ( 3MN ) (11)
θ=0 i =0 j =0
The correlation is compared to a threshold Tρ(l), computed to grant a given probability of

false positive detection, using the Neyman-Pearson criterion. For example, for Pf ≤ 10 −8 , the
threshold is Tρ ( l ) = 3.97 2σ ρ2 ( l ) , with σρ(l)2 the variance of the wavelet coefficients, if the
image was watermarked with a code Y other than X,
l l
2 M /2 − 1 N /2 − 1
σ ρ2(l ) ≈ ( 4 l ( 3 MN ) ) ∑ ∑ ∑ ( Iɶ ( i , j ) ) .
2 2
θ
l (12)
θ=0 i =0 j =0
Barni’s method is quite robust against common signal processing techniques like filtering,
compression, cropping and so on. However, because embedding is made only in the last
resolution level, the watermark information can be easily erased by an attacker. Nafornita,
2008 proposed a pixel-wise mask allowing insertion of the watermark in lower resolution
levels. The third factor of the texture is estimated using the local standard deviation of the
original image computed on a rectangular moving window W(i,j) of WS×WS pixels, centered
on each pixel I(i,j). This criterion of segmentation finds its contours, textures and regions
with high homogeneity. The local mean is:
µˆ ( i , j ) = WS−2 ∑ I ( m, n ) (13)
I ( m , n )∈W ( i , j )
The local variance is given by:
σˆ 2 ( i , j ) = WS−2 ∑ ( I ( m, n) − µˆ ( i , j ))
2
(14)
I ( m , n )∈W ( i , j )
The local standard deviation is the square root of this local variance. The texture for a
considered DWT coefficient is proportional with the local standard deviation of the
corresponding pixel from the host image. We denote this local standard deviation image
with S, and the local mean image with U. Embedding is made in the subband s, level l; the
size of the texture matrix must agree with the size of the subband. Hence, the approximation
image at the lth decomposition level is used. This compression can be realized exploiting the
separation properties of the DWT. To generate the mask required for the embedding into the
detail subimages corresponding to the lth decomposition level, the DWT of the local
standard deviation image is computed (making l+1 iterations). The required mask will be
the approximation subimage from level l, denoted Sl3, normalized to the local mean, also
compressed in the wavelet domain, Ul3. This is illustrated in Fig. 3. One difference between
the watermarking method proposed by Nafornita, 2008 and the one proposed by Barni et
al., 2001, is given by the computation of the local variance – the second term – in (10). To
obtain the new values of the texture, the local variance of the image to be watermarked is
computed, using the relations (13) and (14). The local standard deviation image is
decomposed using one iteration wavelet transform, and only the approximation image is
kept. Relation (10) is then replaced with:
3 −l 2 1
 Iθk + l ( y + i 2 k , x + j 2 k ) 
2
Ξ ( l , i , j ) = ∑ 16− k ∑ ∑  
k =0 θ =0 x , y =0 (15)
⋅S ( i , j ) U ( i , j )
3
l
3
l
Sθ l
normalization
Local DWT S3 0 S0 0
Original standard S3 0 NS30
image deviation
I S2 0 S1 0
S
Uθl
DWT
U30 U00
Local
U30
mean
2 1
U U 0 U 0
Fig. 3. Watermark embedding. The watermark is embedded using a secret or public key,
making invisible changes to the cover work.
The second difference is that the luminance mask is computed on the approximation image
from level l, where the watermark is embedded. The DWT of the original image using l
decomposition levels was computed and the approximation subimage corresponding at
level l was separated, obtaining the image I l3 . The luminance content is computed using:
L ( l , i , j ) = I l3 ( i , j ) 256 (16)
Since both factors are more dependent on the resolution level in the method proposed by
Barni, the noise sensitivity function becomes:
 2 , θ = 1  1.00 l ∈ {0,1}
Θ ( l ,θ ) =   . (17)
 1, otherwise  0.66 l = 2 
It was considered the ratio between the correlation ρ(l) in Eq. (11) and the image dependent
threshold Tρ(l), hence the detector was viewed as a nonlinear function with a fixed
threshold. In Nafornita, 2007a, three detectors are used, to take advantage of the wavelet
hierarchical decomposition. The watermark presence is detected,
1. from all resolution levels, “all_levels”,
2. separately from each resolution level, considering the maximum detector response from
each level, “max_level”,
3. separately from each subband, considering the maximum detector response from each
subband, “max_subband”.
Evaluating the correlations separately per resolution level or subband can be sometimes
advantageous. In the case of cropping, the watermark will be damaged more likely in the
lower frequency than in the higher frequency, while lowpass filtering affects more the
higher frequency than lower ones. Layers or subbands with lower detector response are
discarded. This type of embedding combined with new detectors is more attack resilient to a
possible erasure of the three subbands watermark. The detector “all_levels” evaluates the
watermark’s presence on all resolution levels:
d1 = ρ d 1 Td 1 (18)
where the correlation ρ d 1 is given by:
l l
2 M /2 − 1 N /2 − 1
2
 2

ρ d 1 = ∑∑ ∑ ∑ Iɶlθ ( i , j ) xlθ ( i , j )  3MN ∑ 4 − l  (19)
l =0 θ =0 i =0 j =0  l=0 
The threshold for Pf ≤10-8 is Td1 = 3.97 σ ρd1

2
, with:
l l
2 M /2 − 1 N /2 − 1 2
2
 2
−l 
∑ ∑ ( Iɶ ( i , j ) )
2
σ ρd1
2
≈ ∑∑ θ
l  3MN ∑ 4  (20)
l=0 θ=0 i =0 j =0  l=0 
The second detector “max_levels” considers the responses from different levels, as
d(l)=ρ(l)/T(l), with l∈{0, 1, 2}, and discards the detector responses with lower values:
d2 = max {d ( l )} (21)
l
The third detector considers the responses from different subbands and levels, as d(l,θ) the
ratio ρ(l,θ)/T(l,θ), with l,θ∈{0, 1, 2}, and discards the detector responses with lower values,
d3 = max {d ( l ,θ )} (22)
l ,θ
The correlation and threshold are computed with the same rationale on one subband,
indicated by its orientation and level.
3.2 Complex wavelet transform methods

The discrete wavelet transform is useful to embed the watermark because the visual quality
of the images is very good. However, it has three main disadvantages (Kingsbury, 2001):
lack of shift invariance, lack of symmetry of the mother wavelets and poor directional
selectivity. Caused by the lack of shift invariance of the DWT, small shifts in the input signal
can produce important changes in the energy distribution of the wavelet coefficients. Due to
the poor directional selectivity for diagonal features of the DWT the watermarking capacity
is small. The most important parameters of a watermarking system are robustness and
capacity. These parameters must be maximized. These disadvantages can be diminished
using a complex wavelet transform (Kingsbury, 2000, 2001).
A very simple implementation of the Hyperanalytic Wavelet Transform, (HWT), recently
proposed (Adam et al., 2007) has a high shift-invariance degree versus other quasi-shift-
invariant wavelet transforms (WT) at same redundancy. It has also an enhanced directional
selectivity. All the WTs have two parameters: the mother wavelets (MW) and the primary
resolution (PR), (number of iterations). The importance of their selection is highlighted in
Nason, 2002. Another appealing particularity of those transforms, coming from their
multiresolution capability, is the interscale dependency of the wavelet coefficients.
We present in the next paragraphs a new implementation of HWT (Adam et al., 2007) and
its application to watermarking (Nafornita et al., 2008). The watermark capacity was studied
in Moulin & Mihcak, 2002, where an information-theoretic model for image watermarking
and data hiding is presented. Models for geometric attacks and distortion measures that are
invariant to such attacks are also considered. The lack of shift invariance of the DWT and its
poor directional selectivity are reasons to embed the watermark in the field of another WT.
To maximize the robustness and the capacity, the role of the redundancy of the transform
used must be highlighted first. An example of redundant WT is represented by the tight
frame decomposition. In Hua & Fowler, 2002 are analyzed the watermarking systems based
on tight frame decompositions. The analysis indicates that a tight frame offers no inherent
performance advantage over an orthonormal transform (DWT) in the watermark detection
process despite the well known ability of redundant transforms to accommodate greater
amounts of added noise for a given distortion. The overcompleteness of the expansion,
which aids the watermark insertion by accommodating greater watermark energy for a
given distortion, actually hinders the correlation operator in watermark detection. As a
result, the tight-frame expansion does not inherently offer greater spread-spectrum
watermarking performance. This analytical observation should be tempered with the fact
that spread-spectrum watermarking is often deployed in conjunction with an image-
adaptive weighting mask to take into account the human visual model (HVM) and to improve
perceptual performance. Another redundant WT, the DTCWT, was already used for
watermarking (Loo & Kingsbury, 2000). The authors of this paper prove that the capacity of a
watermarking system based on a complex wavelet transform is higher than the capacity of a
similar system that embeds the watermark in the DWT domain. Many authors (e.g. Daugman,
1980) have suggested that the processing of visual data inside our visual cortex resembles
filtering by an array of Gabor filters of different orientations and scales. The proposed
implementation of HWT is efficient, has only a modest amount of redundancy, provides
approximate shift invariance, has better directional selectivity than the 2D DWT and it can be
observed that the corresponding basis functions closely approximate the Gabor functions. So,
the spread spectrum watermarking based on the use of an image adaptive weighting mask
applied in the HWT domain is potentially a robust solution that increases the capacity.
3.2.1 A new implementation of the Hyperanalytic Wavelet Transform

The hypercomplex mother wavelet associated to a real mother wavelet ψ ( x , y ) is:
ψ a ( x , y ) = ψ ( x , y ) + iHx {ψ ( x , y )} +
(23)
{
+ jHy {ψ ( x , y )} + kHx Hy {ψ ( x , y )} }
where i 2 = j 2 = − k 2 = −1, and ij = ji = k (Davenport, 2008). The HWT of the image f ( x , y ) is:
HWT { f ( x , y )} = f ( x , y ) ,ψ a ( x , y ) . (24)
The 2D-HWT of the image f ( x , y ) can be computed using the 2D-DWT of its associated
hypercomplex image:
HWT { f ( x , y )} = DWT { f ( x , y )} +
x { y }
iDWT H { f ( x , y )} + jDWT H { f ( x , y )} + { }
{ }}
(25)
{
+ kDWT H H { f ( x , y )} =
y x
f ( x , y ) ,ψ ( x , y ) = DWT { f ( x , y )} .
a a
HWT uses four trees, each implemented by 2D-DWT, being adequate to a multi-wavelet
environment (Firoiu et al., 2009). Hx is the Hilbert transform computed across lines and Hy
across columns (Fig. 4). The HWT coefficients are organized in two sequences of complex
coefficients separated by the sign of their preferential orientation, with 6 subbands, 3 of
positive orientations and 3 of negative orientations ±atan(1/2), ±π/4 and ±atan(2):
z± = z± r + jz± i

=  f D1,2 ,3 ∓
 { }
Hy Hx { f }

D1,2 ,3  + j

( Hx D1,2 ,3 ± Hy )
D1,2 ,3 .
(26)
Fig. 4. The new HWT implementation architecture.
3.2.2 Watermarking using the Hyperanalytic Wavelet Transform

Adapting the strategy already described in the previous paragraph to the case of HWT, a
new method was proposed in Nafornita et al., 2008. The first three wavelet decomposition
levels are used and the watermark is embedded into the real coefficients with positive and
negative orientations, z+ r and z− r , respectively. In this case the relations already described
in the previous paragraph were used independently for each of these two images. The same
message was embedded in both images, using the mask from Nafornita, 2007a. The
difference is that the orientations or preferential directions are in this case: atan(1/2), π/4,
atan(2) (respectively for θ = 0, 1, 2), for the image z+ r and -atan(1/2), -π/4, -atan(2), (θ=0, 1,
2) for the image z− r . At the detection side, we consider the pair of images ( z+ r , z− r ), thus
having twice as much coefficients than the standard approach, and θ takes all the possible
values, ±atan(1/2), ±π/4, ±atan(2).
3.3 Results and comparisons

We will compare in the following watermarking systems based on DWT with the ones
based on complex WTs, namely the HWT.
3.3.1 Results for methods based on the discrete wavelet transform

In Nafornita et al., 2006a, the system proposed by Barni et al. was modified, using the
texture mask in (15). The image Barbara is watermarked with various values of the
embedding strength α. The binary watermark is embedded in all the detail wavelet
coefficients of the first resolution level. Watermarked Barbara for α=1.5 is shown in Fig. 5.
Fig. 5. Original and watermarked Barbara images with α = 1.5.
Fig. 6. Left: The ratio ρ/T as a function of the PSNR between the marked and the original
images, for different quality factors, JPEG compression. Right: Ratio ρ/T as a function of
embedding strength, for different quality factors, JPEG compression. Pf is set to 10-8.
Fig. 6 shows results for JPEG compression, for different quality factors: the ratio ρ/T is
plotted as a function of the peak signal-to-noise ratio (PSNR) between the marked (un-
attacked) image and the original one, and respectively as a function of α. The probability of
false positive detection is set to 10-8. If this ratio is greater than 1 then the watermark is
positively detected. Generally, for a PSNR higher than 30 dB, the original image and
watermarked one are considered indistinguishable. For compression quality factors higher
or equal than 25 the distortion introduced by JPEG compression is tolerable. For PSNR in
the range of 30-35 dB, of practical interest, the watermark is detected for all significant
compression quality factors. Increasing the embedding strength, the PSNR of the
watermarked image decreases, and the ratio ρ/T increases. The watermark is still detectable
even for very small values of α. For the quality factor Q=5 (or a compression ratio CR=32), the
watermark is still detectable even for α=0.5. Fig. 7 shows the detection of a true watermark for
various quality factors, in the case of α=1.5; the threshold is well below the detector response.
In Table 1 we give a comparison between the two methods, for the Lena image, α=1.5 in the
case of JPEG compression with a quality factor of 5 (compression ratio of 46).
Fig. 7. Left: Detector response ρ, threshold T, as a function of different quality factors (JPEG
compression). The watermark is successfully detected. Pf is set to 10-8. Right: Highest
detector response, ρ2, corresponding to a fake watermark and threshold T. The threshold is
above the detector response.
Nafornita et al., 2006a Barni et al., 2001

ρ 0.3199 0.038
T 0.0844 0.036
ρ2 0.0516 0.010
Table 1. A comparison for JPEG compression with a compression ratio CR = 46.
The detector response for the original embedded watermark ρ, the detection threshold T,
and the second highest detector response ρ2 are given. Pf was set to 10-8 and 1000 marks
were tested. The detector response is higher than in Barni’s case.
Fig. 8. Original image Lena; mask from Nafornita et al., 2006b and Barni’s mask for level l=0.
The masks are the complementary of the real ones.
In Nafornita et al., 2006b, Barni’s method is modified, using the texture mask in (15), as well
as the luminance factor in (16). The masks obtained are shown in Fig. 8. The improvement is
clearly visible around edges and contours. The method is applied in two cases, when the
watermark is inserted in level 0 only and when it’s inserted in level 1 only. JPEG
compression is again considered. The image Lena is watermarked at level l=0 and
respectively at level l=1 with α ranging from 1.5 to 5. The binary watermark is embedded in
all the detail wavelet coefficients of the resolution level, l as previously described. For α=1.5,
the watermarked images, in level 0 and level 1, as well as the image watermarked using
Barni’s mask, are shown in Fig. 9. Obviously the quality of the watermarked images are
preserved using the new pixel-wise mask. The PSNR values are 38 dB (level 0) and 43 dB
(level 1), compared to Barni’s method, with a PSNR of 20 dB.
Fig. 9. Watermarked images, α =1.5, for Nafornita et al., 2006b, level 0 (PSNR = 38 dB); level 1
(43 dB); for Barni et al., 2001, level 0 (20 dB).
Fig. 10. Left: PSNR as a function of α. Embedding is made either in level 0 or in level 1.Right:
Detector response ρ, threshold T, highest detector response, ρ2, corresponding to a fake
watermark, as a function of different quality factors (JPEG compression). The watermark is
successfully detected. Pf is set to 10−8. Embedding was made in level 0.
PSNR values are shown in Fig. 10(left) as a function of the embedding strength. The
watermark is still invisible, even for high values of α. Fig. 11 gives the results for JPEG
compression. In all experiments, the probability of false positive detection is set to 10−8. The
watermark is successfully detected for a large interval of compression quality factors. For
PSNR values higher than 30 dB, the watermarking is invisible. For quality factors Q≥10, the
distortion introduced by JPEG compression is tolerable. For all values of α, the watermark is
detected for all the significant quality factors (Q≥10). Increasing the embedding strength, the
PSNR of the watermarked image decreases, and ρ/T increases. For the quality factor Q = 10
(or a compression ratio CR = 32), the watermark is still detectable even for low values of α.
Fig. 10(right) shows the detection of a true watermark from level 0 for various quality
factors, for α=1.5; the threshold is below the detector response. The selectivity of the
watermark detector is also illustrated, when a number of 999 fake watermarks were tested:
the second highest detector response is shown, for each quality factor. False positives are
rejected.
In Table 2 a comparison between Nafornita et al., 2006b and Barni et al., 2001, can be seen
for JPEG compression with Q=10 (compression ratio of 32). The detector response for the
original watermark ρ, the detection threshold T, and the second highest detector response
ρ2, when the watermark was inserted in level 0 are given. The detector response is higher
than for Barni et al.
Fig. 11. Ratio ρ/T as a function of the embedding strength α. The watermarked image is
JPEG compressed with different quality factors Q. Pf is set to 10−8. Embedding was made in
level 0 (left), and in level 1 (right).
Nafornita et al., 2006b Barni et al., 2001

ρ 0.0750 0.062
T 0.0636 0.036
ρ2 0.0461 0.011
Table 2. A comparison for JPEG compression with a compression ratio CR = 32.
The method in Nafornita, 2007a allows embedding of the watermark in all resolution levels,
except the last one (low resolution). Three types of detectors are used, as described in
paragraph 3.1. Various images of size 512x512, have been watermarked at levels l∈{0, 1, 2}
using the new mask. The embedding strength is α=1.5. Based on human observation and the
peak-signal-to-noise ratio, PSNR, the images are indistinguishable from the original ones.
For Barni et al. method, a watermark is embedded in all the detail wavelet coefficients of the
first resolution level, l=0, for α=0.2, that results in a similar image quality (see Fig.12). This
has been concluded in Nafornita, 2007b, where by limiting the watermark strength such that
the PSNR is 35 dB and in average the percentage of affected pixels is less than 25%, the
quality of the images is greatly improved. Girod’s model has been used for determining the
location and number of affected pixels (Girod, 1989). For instance, in Barni’s case, the
watermarked image with α=0.2 has a PSNR of 36.39 dB, 11.84% affected pixels, compared to
the one watermarked with α=1.5 has a PSNR of 20 dB, and all pixels are affected. What is
kept constant for comparison are the 2D watermarks embedded in the first level, and the
image quality. The method Nafornita, 2007a cannot be compared with the one in Barni et al.,
2001 when the watermark is embedded in all resolution levels, simply because their mask
isn’t suited for embedding in other levels than the highest resolution level. Results for some
of the standard images from the USC SIPI Image Database are given.
Fig. 12. (left) Original image Lena, (middle) Watermarked images for Nafornita, 2007a,
α=1.5, PSNR=36.86 dB, (right) Barni et al., 2001, α=0.2, PSNR=36.39 dB.
Table 3 includes PSNR values for the two cases. For the first detector, an estimate of the false
positive probability is shown for the image Lena, before and after JPEG compression attack,
with quality factor Q=10, as a function of the detection thresholds, Tρ1. The threshold values
have been computed using as estimate the variance of the ρ1 obtained from experiments.
The mean PSNR for the twelve images is 34.16 dB for the proposed method (Nafornita,
2007a) and 34.06 dB for Barni’s method.
Nafornita, 2007a
Detector response vs. attack Barni’s method
1-All levels 2-Max level 3-Max subband
JPEG compression, Q=10 2.38 1.98 1.44 1.75
Median filtering, M=5 1.32 1.12 1.46 0.25
Scaling, 50% 4.06 5.21 5.76 1.85
Cropping, 512x512 -> 32x32 0.68 0.98 1.73 1.48
Gamma correction, γ=2 20.32 29.19 28.06 32.54
Motion blur, L=31, θ=11 1.98 5.48 8.04 6.14
Table 3. Resistance to different attacks, for Nafornita, 2007a method. The detector response
is a mean value of different responses.
Tests were made for JPEG compression, median filtering, cropping, resizing, gamma
correction and blurring. Table 3 shows the mean values of the detector responses for each
attack. A particular attack parameter is chosen where the watermark is still detectable by at
least one detector. For compression, the method in Nafornita, 2007a successfully detects the
watermark at Q=10. The 1st detector is better in all cases. This new method has better results
than Barni’s technique. The watermark of both methods survived in all images for median
filtering with kernel sizes up to 3. For kernel size 5, the watermark of Nafornita, 2007a using
the first and third detector is detectable; Barni’s method fails to detect the watermark. In the
case of scaling to 50%, the watermark was successfully detectable in both cases, with better
results for Nafornita, 2007a. The third detector has the best performance in detecting the
mark. The watermark of Nafornita, 2007a was successfully detected in the cropped image of
32x32, only with the third detector, which proves its efficiency. Barni’s method detects the
watermark with similar detector responses as in the case of the third detector. As expected
for normalized correlation detection, both methods are practically insensitive to gamma
correction adjustment. For the motion blur attack, both methods have successfully detected
the watermark in all cases. Detector 3 has slightly better results than the others.
Fig. 13. Experimentally evaluated probability of false positive Pf vs. Tρ1/σρ1, the ratio
between the detection threshold and standard deviation of the correlations in the case where
an incorrect watermark was embedded. The theoretical trend is also shown (‘o’ marker).
Tests were made on Lena, before and after JPEG compression with quality factor 10, using
5×104 different watermarks.
For the first detector, the probability of false positive was estimated by searching many
different watermarks into one watermarked image, Lena. Each threshold Tρ1 was set in such
a way to grant a given value of Pf. The trial was repeated for values of Pf ranging from 10-1
through 10-4. In total 5x104 watermarks per image have been tested. The estimation has been
done before any type of manipulation and after JPEG compression, with quality factor 10.
The estimated Pf is plotted in Fig. 13 versus the ratio Tρ1/σρ1 between the detection
thresholds and standard deviations of correlations for the case corresponding to certain
estimates of this probability of false positive. This case corresponds to the situation where
the image is watermarked with a code Y other than X.
Surprisingly, the estimated false alarm Pf, is lower in the case of compression than in the
case of no attack, for the same detection threshold. This can be explained by the fact that
before compression, the empirical pdf of the correlations in the case for an incorrect
watermark is embedded, was not Gaussian. Although the two empirical pdf’s are closer
after the attack, they are still very good separated and the empirical pdf for an incorrect
watermark has the mean below zero, compared to the equivalent one before – which is
centered on zero. Thus setting a particular threshold can indeed result in a lower false alarm
after attack. Similar results were obtained for Barbara, and for the same attack.
For the first detector, the obtained probability of false positive is close to the expected one.
The assumption that the wavelet coefficients from different levels and subbands are i.i.d. is
thus reasonable and the detector has a good performance.
3.3.2 Results for methods based on the Hyperanalytic Wavelet transform

In Nafornita et al., 2008 the watermark is embedded in the HWT domain, in all levels (0, 1
and 2) and all orientations (positive and negative). The test image is Lena, of size 512x512.
For α=1.5, the watermarked image has a PSNR of 35.63 dB. The original image, the
corresponding watermarked image and the difference image are presented in Fig. 14.
Fig. 14. Original and watermarked images with method (Nafornita et al., 2008), for α=1.5,
PSNR=35.63 dB; Difference image, amplified 8 times.
The watermarked images have been exposed at some common attacks: JPEG compression
with different quality factors (Q), shifting, median filtering with different window sizes M,
resizing with different scale factors, cropping with different areas remaining, gamma
correction with different values of γ, blurring with a specified point spread function (PSF)
and perturbation with AWGN with different variances.
Resistance to unintentional attacks, for watermarked image Lena, can be compared to the
results obtained using the watermarking methods in Barni et al., 2001 and Nafornita, 2007a
analyzing Table 4. For the method in Nafornita, 2007a, the same watermark strength, 1.5 is
used and the watermark is embedded in all three wavelet decomposition levels, resulting in
a PSNR of 36.86 dB. For the method in Barni et al., 2001, the watermark strength 0.2 is used
and the embedding is made only in the first resolution level, resulting in a similar quality of
the images (PSNR=36.39 dB).
DWT-Nafornita, 2007a HWT-Nafornita et al., 2008

Attacks vs. DWT-Barni
max max
detector response all levels max level et al., 2001 all levels max level
subband subband
Before attack 21.57 39.12 33.60 44.31 24.78 43.18 26.30
JPEG, Q=50 5.45 6.76 5.02 6.22 6.25 7.87 4.85
JPEG, Q=25 3.02 3.67 2.60 3.03 3.23 4.19 2.62
JPEG, Q=20 2.55 3.08 2.09 2.38 2.72 3.58 2.33
Shift, li=128, co=128 21.57 39.12 33.59 44.31 24.78 43.18 26.30
Median filter, M=3 4.29 4.58 4.87 1.57 4.59 5.42 4.37
Median filter, M=5 1.66 1.24 2.27 0.59 1.61 1.64 1.49
Resizing, 0.75 9.53 15.86 15.64 14.09 10.93 19.34 14.67
Resizing, 0.50 4.21 5.72 5.75 2.31 4.56 6.14 8.71
Cropping, 256x256 7.40 12.14 17.10 18.08 8.68 15.20 13.82
Cropping, 128x128 3.11 4.66 8.31 8.01 3.53 6.04 6.86
Cropping, 64x64 1.10 1.72 4.45 3.92 1.32 2.47 3.71
Gamma correction, γ=1.5 22.18 39.76 33.74 43.04 25.31 43.61 26.45
Gamma correction, γ=2 22.59 39.70 32.98 42.43 25.62 43.24 25.88
Blur, L=31, β=11 2.69 7.81 9.56 9.05 3.05 9.18 7.55
Table 4. Resistance to different attacks, for HWT based method compared to DWT based
methods.
Special attention was paid to the shifting attack. First the watermarked image was circularly
shifted with li lines and co columns, obtained the attacked image ( Iɶt ) . Supposing that the
numbers li and co are known, the messages at level l are circularly shifted with li/2l lines
θ
and co/2l columns obtaining the new messages ( xt )l . Next the watermark was detected
θ
using the image ( I )t and the messages ( xt )l . The values obtained for li=128 and co=128 are
ɶ
presented in Table 4.
From the results, it is clear that embedding in the real parts of the HWT transform yields in
a higher capacity at the same visual impact and robustness. In fact the results obtained in
Nafornita et al., 2008 are slightly better than the results obtained with the DWT-based
methods in Nafornita et al., 2008 and Barni et al., 2001 for JPEG compression, median
filtering with window size M=3, resizing and gamma correction. For the other attacks the
results obtained are similar with the results of the watermarking methods based on DWT.
The case of the shifting attack is very interesting. In this case the robustness of the
watermarking method is given by two properties: the shift invariance degree of the WT
used and the masking ability. All the methods compared in Table 4 are very robust against
the shifting attack. The values of the ratios between the correlations and the image
dependent thresholds obtained before and after the shifting attack are equal for all the
methods compared in Table 4. So, the ability of masking seems to be more important than
the shift invariance degree of the WT used for the conception of counter-measures against
the shifting attack, when the numbers of lines and columns used for the attack are already
known. Of course, the detection of these numbers must also be realized, for the
implementation of a strategy against the shifting attack.
4. Conclusion
In a watermarking system, robustness evaluation should be made if invisibility criteria are
satisfied. For this purpose, perceptual watermarks are being used to overcome the issue of
robustness against invisibility. In the literature, there was proposed a blind spread spectrum
technique that uses a perceptual mask in the wavelet domain, taking into account the noise
sensitivity, texture and the luminance content of all image subbands. We described new
techniques proposed by the authors, based on the modifications of this perceptual mask, in
order to increase robustness, while still maintaining imperceptibility. Moreover, using the
new mask, information is successfully hidden in the lower frequency levels, thus increasing
the capacity and making the watermark more robust to common attacks that affect both
high frequencies and low frequencies of the image. A good balance between robustness and
invisibility of the watermark is achieved when embedding is made in all detail subbands for
all resolution levels, except the coarsest level; this can be particularly useful against erasure
of high frequency subbands containing the watermark in Barni’s system.
A nonlinear detector with fixed threshold – as ratio between correlation and the image
dependent ratio – has been used; three watermark detectors were proposed in Nafornita,
2007a that take advantage of the hierarchical wavelet decomposition: 1) from all resolution
levels, 2) separately from each level, considering the maximum detector response for each
level and 3) separately from each subband, considering the maximum detector response for
each subband. This has been advantageous for cropping, scaling and median filtering where
the 3rd detector shows improved performance. We tested our methods against different
attacks, and found out that it is better than Barni’s method. The behavior of our methods can
be explained by the fact that we have used a better estimate of the mask and we took
advantage of the diversity of the wavelet decomposition. The effectiveness of the new
perceptual mask is appreciated by comparison with Barni’s method. Simulation results
show the superiority of the proposed methods (Nafornita et al., 2006a, b, Nafornita,
2007a).
The HWT is a very modern WT as it has been formalized only two years ago. A very simple
implementation of this transform has been used, which permits the exploitation of the
mathematical results and of the algorithms previously obtained in the evolution of wavelets
theory. It does not require the construction of any special wavelet filter. It has a very flexible
structure, as we can use any orthogonal or bi-orthogonal real mother wavelets for the
computation of the HWT. The presented implementation leads to both a high degree of
shift-invariance and to an enhanced directional selectivity in the 2D case. An ideal Hilbert
transformer was considered. A new type of pixel-wise masking for robust image
watermarking in the HWT domain has been presented (Nafornita et al., 2008). Modifications
were made to two existing watermarking technique proposed in Barni et al., 2001 and
Nafornita, 2007a, based on DWT. These techniques were selected for their good robustness
against the usual attacks. The method is based on the method in Barni et al., 2001, with some
modifications. The first modification is in computing the estimate of the variance, which
gives a better measure of the texture activity. An improvement is also owed to the use of a
better luminance mask. The third improvement is to embed the watermark in the detail
coefficients at all resolutions, except the coarsest level, making the watermark more attack
resilient. The HWT embedding exploits the coefficients z+ r and z− r .
The simulation results illustrate the effectiveness of the proposed algorithms. The methods
were tested against different attacks (in terms of robustness). The HWT based watermarking
method is similar and in some cases outperforms the DWT based methods, but it has a
superior capacity than the DWT based methods.
As a future research direction, the statistical properties of the HWT will be used to improve
the watermark detection.
5. References
Abdulaziz, N.; Glass, A.; Pang, K.K. (2002). Embedding Data in Images Using Turbo-
Coding, 6th Symposium on DSP for Communication Systems, 28-31 Jan. 2002, Univ. of
Wollongong, Australia.
Adam, I.; Nafornita, C.; Boucher, J.-M. & Isar, A. (2007). A New Implementation of the
Hyperanalytic Wavelet Transform, Proc. of IEEE Symposium ISSCS 2007, Iasi,
Romania, ’07, 401-404.
Balado F. & Perez-Gonzalez, F. (2001). Coding at the Sample Level for Data Hiding: Turbo
and Concatenated Codes, SPIE Security and Watermarking of Multimedia Contents,
San Jose CA, 22-25 Jan. 2001, San Jose CA , USA, 2001, 4314, 532-543.
Barni, M.; Bartolini, F. & Piva, A (2001). Improved wavelet-based watermarking through
pixel-wise masking, IEEE Trans. Image Processing, 10, 5, May 2001, 783 – 791.
Braudaway, G.W.; Magerlein, K.A. & Mintzer, F. (1996). Protecting publicly available images
with a visible watermark, Proc. SPIE – Int. Soc.Opt. Eng., vol. 2659, pp.126 – 133,
1996.
Cox, I. (2005). Robust watermarking, ECRYPT Summer School on Multimedia Security,
Salzburg, Austria, Sept. 22, 2005
Cox, I.; Killian, J.; Leighton, T. & Shamoon, T. (1997). Secure Spread Spectrum
Watermarking for Multimedia, IEEE Trans. Image Processing, 6, 12, 1997, 1673-1687
Cox, I.; Miller, M. & Bloom, J. (2002). Digital Watermarking, Morgan Kaufmann Publishers,
2002
Daugman, J. (1980). Two-dimensional spectral analysis of cortical receptive field profiles,
Vision Res., 20, ’80, 847-856.
Davenport, C. (2008). Commutative Hypercomplex Mathematics, Available from:
http://home.comcast.net/~cmdaven/hyprcplx.htm.
Firoiu, I.; Nafornita, C.; Boucher, J.–M. & Isar, A. (2009). Image Denoising Using a New
Implementation of the Hyperanalytic Wavelet Transform, IEEE Transactions on
Instrumentation and Measurements, vol. 58, Issue 8, August 2009, pp. 2410-2416.
Girod, B. (1989). The information theoretical significance of spatial and temporal masking in
video signals, Proc. SPIE Human Vision, Visual Processing, and Digital Display, vol.
1077, pp. 178–187, 1989.
Hua, L. & Fowler, J. E. (2002). A Performance Analysis of Spread-Spectrum Watermarking
Based on Redundant Transforms, Proc. IEEE Int. Conf. on Multimedia and Expo,
Lausanne, Switzerland, ’02, vol. 2, 553–556.
Kingsbury, N. (2001). Complex Wavelets for Shift Invariant Analysis and Filtering of
Signals, Applied and Comp. Harm. Anal. 10, ’01, 234-253.
Kingsbury, N. (2000). A Dual-Tree Complex Wavelet Transform with improved
orthogonality and symmetry properties, Proc. IEEE Conf. on Image Processing,
Vancouver, ’00, paper 1429.
Kundur, D. (2000). Water-filling for Watermarking?, Proc. IEEE Int. Conf. On Multimedia and
Expo, NY, 1287-1290, Aug. 2000.
Kundur, D. & Hatzinakos, D. (1998). Digital Watermarking using Multiresolution Wavelet
Decomposition, Proc. IEEE Int. Conf. On Acoustics, Speech and Signal Processing,
Seattle, Washington, Vol. 5, pp. 2969-2972, May 1998.
Kundur, D. & Hatzinakos, D. (2001). Diversity and Attack Characterization for Improved
Robust Watermarking, IEEE Transactions on Signal Processing, Vol. 49, No. 10, 2001,
pp. 2383-2396.
Lin, C. Y.; Wu, M.; Bloom, J. A.; Cox, I. J.; Miller, M. L. & Lui, Y. M. (2001). Rotation, Scale,
and Translation Resilient Watermarking for Images, IEEE Trans. On Image
Processing, 10, 5, May 2001
Loo, P. & Kingsbury, N. (2000). Digital Watermarking Using Complex Wavelets, ICIP 2000.
Moulin, P. & Mihcak M.K. (2002). A Framework for Evaluating the Data-Hiding Capacity of
Image Sources, IEEE Trans. Image Processing, 11(9), ’02, 1029-1042.
Nafornita, C.; Isar, A. & Borda, M. (2005). Image Watermarking Based on the Discrete
Wavelet Transform Statistical Characteristics, Proc. IEEE EUROCON 2005, Serbia &
Montenegro, 943-946.
Nafornita, C. (2008). Contributions to Digital Watermarking of Still Images in the Wavelet
Transform, Ph.D. thesis, Feb. 2008, Technical University of Cluj-Napoca, Romania.
Nafornita, C.; Isar, A.; Kovaci M. (2009). Increasing Watermarking Robustness using Turbo
Codes, IEEE International Symposium on Intelligent Signal Processing WISP 2009,
Budapest, Hungary, 26-28 August 2009.
Nafornita, C.; Firoiu, I.; Boucher, J.-M. & Isar, A. (2008). A New Watermarking Method
Based on the Use of the Hyperanalytic Wavelet Transform, Proc. SPIE Europe:
Photonics Europe, vol. 7000: Optical and Digital Image Processing 70000W, pp.70000W-
1-70000W-12, ISBN 97808194 71987, Strasbourg, April, 2008.
Nafornita, C. (2007). A New Pixel-Wise Mask for Watermarking, Proc. of ACM Multimedia
and Security Workshop, 2007, Dallas, TX, USA.
Nafornita, C.; Isar, A. & Borda, M. (2006). Pixel-wise masking for watermarking using local
standard deviation and wavelet compression, Scientific Bulletin of the Politehnica
Univ. of Timisoara, Trans. on Electronics and Communications, 51(65), 2, pp. 146-151,
ISSN 1583-3380, 2006.
Nafornita, C.; Isar, A. & Borda, M. (2006). Improved Pixel-Wise Masking for Image
Watermarking, Multimedia Content Representation, Classification and Security,
September 11-13, 2006, Istanbul, Turkey, Lecture Notes in Computer Science,
Springer-Verlag, 2006, pp. 90-97.
Nafornita, C. (2007). Robustness Evaluation of Perceptual Watermarks, IEEE Int. Symposium
on Signal, Circuits and Systems ISSCS 2007, 12-13 July 2007, Iasi, Romania.
Nason, G.P. (2002). Choice of wavelet smoothness, primary resolution and threshold in
wavelet shrinkage, Statistics and Computing, 12, ’02, 219-227.
Nikolaidis, N. & Pitas, I. (1998). Robust Image Watermarking in the Spatial Domain, Trans.
Signal Processing, Vol. 66, No. 3, pp. 385-403, 1998.
Ó Ruanaidh, J.J.K. & Pun, T. (1998). Rotation, Scale and Translation Invariant Spread
Spectrum Digital Image Watermarking, Signal Processing, 66(1998), pp. 303-317.
Ó Ruanaidh, J.J.K; Dowling, W.J.; Boland, F.M. (1996). Phase watermarking of digital
images, Proc. IEEE Int. Conf. Image Processing, 1996, pp. 239-242.
Podilchuk, C. & Zeng, W. (1998). Image-Adaptive Watermarking Using Visual Models, IEEE
Journal on Selected Areas in Communications, 16, 4, May 1998, 525-539
Selesnick, I. W.; Baraniuk, R. G. & Kingsbury, N. (2005). The Dual-tree Complex Wavelet
Transform - A Coherent Framework for Multiscale Signal and Image Processing,
IEEE Signal Processing Magazine, 22(6):123-151, November 2005.
Serdean, C.V.; Ambroze, M.A.; Tomlinson, M. & Wade, J.G. (2003). DWT based high-
capacity blind video watermarking, invariant to geometrical attacks, IEE Proc.-Vis.
Image Signal Process., 150, 1, Feb. 2003.
Xia, X.; Boncelet, C. G. & Arce, G. R. (1998). Wavelet Transform Based Watermark for Digital
Images, Optics Express, Vol. 3, No. 12, 1998, pp. 497-505.
Part 4
Discrete Wavelet Transform Algorithms

13
Shift Invariant Discrete Wavelet Transforms

Hannu Olkkonen and Juuso T. Olkkonen
Department of Applied Physics, University of Eastern Finland, Kuopio,
VTT Technical Research Centre of Finland, VTT,
Finland
1. Introduction
The discrete wavelet transform (DWT) has an established position in processing of signals
and images in research and industry. The first DWT structures were based on the compactly
supported conjugate quadrature filters (CQFs) (Smith & Barnwell, 1986; Daubechies, 1988).
However, a drawback in CQFs is related to the nonlinear phase effects such as image
blurring and spatial dislocations in multi-scale analyses. On the contrary, in biorthogonal
discrete wavelet transform (BDWT) the scaling and wavelet filters are symmetric and linear
phase. The biorthogonal filters (BFs) are usually constructed by a ladder-type network
called lifting scheme (Sweldens, 1988). The procedure consists of sequential down and
uplifting steps and the reconstruction of the signal is made by running the lifting network in
reverse order. Efficient lifting BF structures have been developed for VLSI and
microprocessor environment (Olkkonen et al. 2005; Olkkonen & Olkkonen, 2008). The
analysis and synthesis filters can be implemented by integer arithmetics using only register
shifts and summations. Many BDWT-based data and image processing tools have
outperformed the conventional discrete cosine transform (DCT) -based approaches. For
example, in JPEG2000 Standard (ITU-T, 2000), the DCT has been replaced by the lifting BFs.
One of the main difficulties in DWT analysis is the dependence of the total energy of the
wavelet coefficients in different scales on the fractional shifts of the analysed signal. If we
have a discrete signal x[n] and the corresponding time shifted signal x[n − τ ] , where
τ ∈ [0,1] , there may exist a significant difference in the energy of the wavelet coefficients as
a function of the time shift. Kingsbury (2001) proposed a nearly shift invariant method,
where the real and imaginary parts of the complex wavelet coefficients are approximately a
Hilbert transform pair. The energy (absolute value) of the wavelet coefficients equals the
envelope, which provides smoothness and approximate shift-invariance. Selesnick (2002)
observed that using two parallel CQF banks, which are constructed so that the impulse
responses of the scaling filters have half-sample delayed versions of each other: h0 [ n] and
h0 [n − 0.5] , the corresponding wavelets are a Hilbert transform pair. In z-transform domain
we should be able to construct the scaling filters H 0 ( z) and z −0.5 H 0 ( z) . For design of the
scaling filters Selesnick (2002) proposed a spectral factorization method based on the half
delay all-pass Thiran filters. As a disadvantage the scaling filters do not have coefficient
symmetry and the nonlinearity interferes with the spatial timing in different scales and
prevents accurate statistical correlations. Gopinath (2003) generalized the idea for N parallel
filter banks, which are phase shifted versions of each other. Gopinath showed that
increasing N the shift invariance of the wavelet transform improves. However, the greatest
advantage comes from the change N = 1 to 2 .
In this book chapter we review the methods for constructing the shift invariant CQF and BF
wavelet sequences. We describe a dual-tree wavelet transform, where two parallel CQF
wavelet sequences form a Hilbert pair, which warrants the shift invariance. Next we review
the construction of the BF wavelets and show the close relationship between the CQF and
BF wavelets. Then we introduce a novel Hilbert transform filter for constructing shift
invariant dual-tree BF banks.
Fig. 1. The analysis and synthesis parts of the real-valued CQF DWT bank.
2. Design of the shift invariant CQF

The CQF DWT bank consists of the H 0 ( z) and H 1 ( z) analysis filters and G0 ( z) and G1 ( z)
synthesis filters for N odd (Fig. 1)
H 0 ( z) = (1 + z −1 )K P( z)
H 1 ( z) = z− N H 0 ( − z −1 )
(1)
G0 ( z) = H 1 ( − z)
G1 ( z) = − H 0 ( − z)
where P( z ) is a polynomial in z −1 . The scaling filter H 0 ( z ) has the Kth order zero at ω = π .
The wavelet filter H 1 ( z) has the Kth order zero at ω = 0 , correspondingly. The filters are
related via the perfect reconstruction (PR) condition
H 0 ( z )G0 ( z) + H 1 ( z)G1 ( z) = 2 z − N
(2)
H 0 ( − z )G0 ( z) + H 1 ( − z )G1 ( z ) = 0
The tree structured implementation of the real-valued CQF filter bank is described in Fig. 2.
Let us denote the frequency response of the z-transform filter as
H ( z) = ∑ hn z− n ⇒ H (ω ) = ∑ hn e − jωn (3)
n n
Correspondingly, we have the relations
H ( − z ) ⇒ H (ω − π )
(4)
H ( − z −1 ) ⇒ H ∗ (ω − π )
Shift Invariant Discrete Wavelet Transforms 223
where * denotes complex conjugation. In M-stage CQF tree the frequency response of the
wavelet sequence is
M
WM (ω ) = H 1 (ω / 2)∏ H 0 (ω / 2 k ) (5)
k=2
Fig. 2. The tree structured implementation of the real-valued CQF DWT, which yields the
wavelet sequences w1[n], w2 [n] ... wM [ n] and one scaling sequence sM [n] .
Next we construct a phase shifted parallel CQF filter bank consisting of the scaling filter
H 0 ( z) and the wavelet filter H 1 ( z) . Let us suppose that the scaling filters in parallel CQF
trees are related as
H 0 (ω ) = e − j φ (ω ) H 0 (ω ) (6)
where φ (ω ) is a 2π periodic phase function. Then the corresponding CQF wavelet filters are
related as
H 1 (ω ) = e − jω N H 0* (ω − π ) (7)
and
H 1 (ω ) = e − jω N H 0* (ω − π ) = e − jω N e jφ (ω −π ) H 0* (ω − π ) = e j φ (ω −π ) H 1 (ω ) (8)
We may easily verify that the phase shifted CQF bank (6,8) obeys the PR condition (2).
Correspondingly, the frequency response of the M-stage CQF wavelet sequence is
M M
WM (ω ) = H 1 (ω / 2) ∏ H 0 (ω / 2 k ) = e j φ (ω /2 −π ) H 1 (ω / 2)∏ e − j φ (ω /2 ) H 0 (ω / 2 k )
k
k=2 k =2
M (9)
− j ∑ φ (ω /2 )
k
M
j φ (ω /2 −π ) jθ
=e e k =2
H 1 (ω / 2)∏ H 0 (ω / 2 ) = e WM (ω )
k
k=2
where the phase function
M
θ = φ (ω / 2 − π ) − ∑ φ (ω / 2 k ) (10)
k =2
If we select the phase function φ (ω ) in (6) as
φ (ω ) = ω / 2 (11)
the scaling filters (6) are half-sample delayed versions of each other. By inserting (11) in (10)
we have
ω /2 −π M
1 π ω
θ= −ω∑ = − + M +1 (12)
2 k =2 2 k+1 2 2
The wavelet sequences (5,9) yielded by the CQF bank (1) and the phase shifted CQF bank
(6,8) can be interpreted as real and imaginary parts of the complex wavelet sequence
W MC (ω ) = W M (ω ) + jW M (ω ) (13)
The requirement for the shift-invariance comes from
W M (ω ) = H {ψ M (ω )} (14)
where H denotes the Hilbert transform. The frequency response of the Hilbert transform
operator is defined as
H(ω ) = − j sgn(ω ) (15)
where the sign function is defined as

 1 for ω ≥ 0
sgn(ω ) =  (16)
 −1 for ω < 0
In this work we apply the Hilbert transform operator in the form
H(ω ) = e − jπ /2 sgn(ω ) (17)
Our result (12) reveals that if the scaling filters are the half-sample delayed versions of each
other, the resulting wavelet sequences are not precisely Hilbert transform pairs. There
occurs a phase error term ω / 2 M + 1 , which depends both in frequency and the stage M of the
wavelet sequence. In sequel we describe a novel procedure for elimination this error. We
move the phase error in front of the phase shifted CQF tree using the equivalence described
in Fig. 3. Then the error term reduces to ω / 2 . The elimination of the error term can be made
by prefiltering the analyzed signal by the half-sample delay operator D( z) = z−1/2 , which has
the frequency response D(ω ) = e − j ω /2 . The total phase function is then for −π ≤ ω ≤ π
θ (ω ) = ∠D(ω ) − π / 2 + ω / 2 = −π / 2 (18)
which warrants that the M-stage CQF wavelet sequence and the phase error corrected
sequence are a Hilbert transform pair.
Fig. 3. The two equivalents for transferring the phase function in front of the phase shifted
CQF tree.
3. Biorthogonal discrete wavelet transform

The first DWT structures were based on the compactly supported conjugate quadrature
filters (CQFs) (Smith & Barnwell, 1986), which have unavoided nonlinear phase effects in
multi-resolution analyses. On the contrary, in biorthogonal discrete wavelet transform
(BDWT) the scaling and wavelet filters are symmetric and linear phase. The two-channel
biorthogonal filter (BF) bank is of the general form
H 0 ( z) = (1 + z −1 )L Q( z)
H 1 ( z) = (1 − z−1 )M R( z)
(19)
G0 ( z ) = H 1 ( − z)
G1 ( z ) = − H 0 ( − z)
where the scaling filter H 0 ( z) has the Lth order zero at ω = π . The wavelet filter H 1 ( z) has
the Kth order zero at ω = 0 , correspondingly. Q( z) and R( z) are polynomials in z −1 . The
low-pass and high-pass reconstruction filters G0 ( z ) and G1 ( z ) are defined as in the CQF
bank. For two-channel biorthogonal filter bank the PR relation is
H 0 ( z)G0 ( z) + H 1 ( z)G1 ( z) = 2 z− D
(20)
H 0 ( − z)G0 ( z) + H 1 ( − z)G1 ( z) = 0
4. Relationships between CQF and BF wavelet transforms

In the following treatment we use a short notation for the binomial term
BK ( z) = (1 + z−1 )K (21)
which appears both in the CQF and BF banks. Using the binomial term the CQF bank can be
written as
H 0 ( z) = BK ( z)P( z)
H 1 ( z) = z − N ( − z)K BK ( − z)P( − z−1 )
(22)
G0 ( z ) = z − N zK BK ( z)P( z −1 )
G1 ( z ) = BK ( − z)P( − z)
For the PR condition of the CQF bank ( ) the following is valid for K odd
B2 K ( z)P( z)P( z −1 ) − B2 K ( − z)P( − z)P( − z −1 ) = 2 z − N (23)
On the other hand, the PR condition of the BF bank gives
BL + M ( z)Q( z)R( − z) − BL + M ( − z)Q( − z)R( z) = 2 z− D (24)
Both PR conditions are identical if we state 2K = L + M . Then we have
P( z)P( z −1 )z− N + D = Q( z)R( − z) (25)

The above relation (25) gives a novel way to design of the biorthogonal wavelet filter bank
based on the CQF bank and vice versa. The polynomials Q( z) and R( − z) can be found by
factoring P( z)P( z−1 ) , which is a symmetrical polynomial. The roots of the product filter
P( z)P( z−1 ) should be optimally divided so that both Q( z) and R( − z) are low-pass. Then
R( z) is high-pass. If the BF bank is known it is easy to factor Q( z)R( − z) into
P( z) and P( z−1 ) using some spectral factorization method. An important result is related to
the modification of the BF bank (Olkkonen & Olkkonen, 2007a).
Lemma 1: If the scaling filter H 0 ( z) , the wavelet filter H 1 ( z) and the reconstruction filters
G0 ( z ) and G1 ( z ) in FB bank (19) have a perfect reconstruction property (20), the following
modified FB bank obeys also the PR relation
H 0 ( z ) = F ( z )H 0 ( z )
H 1 ( z) = F −1 ( − z)H 1 ( z)
(26)
G0 ( z) = F −1 ( z)G0 ( z)
G1 ( z) = F( − z)G1 ( z)
−1
where F( z ) is any polynomial in z . Proof is yielded by direct insertion (26) to PR condition
(20).
5. Hilbert transform filter for construction of shift invariant BF bank

In BF bank the shift invariance is not an inbuilt property as in CQF bank. In the following
we define the Hilbert transform filter H( z ) , which has the frequency response
H(ω ) = e − jπ /2 sgn(ω ) (27)
where sgn(ω ) = 1 for ω ≥ 0 and sgn(ω ) = −1 for ω < 0 . We describe a novel method for
constructing the Hilbert transform filter based on the half-sample delay filter D( z) = z−0.5 .
The classical approach for design of the half-sample delay filter D( z) is based on the Thiran
all-pass interpolator
c k + z −1 z− N A( z −1 ) c N + c N − 1 + ⋯ + z− N
p
D( z) = z−0.5 = ∏ −1
= = (28)
k =1 1 + c k z A( z) 1 + c1 z−1 + ⋯ + c N z− N
where the ck coefficients are optimized so that the frequency response follows approximately
D(ω ) = e − jω /2 (29)
In this work we define the half-sample delay filter more generally as
A( z)
D( z) = (30)
B( z)
The quadrature mirror filter D( − z) has the frequency response
D(ω − π ) = e − j(ω −π )/2 (31)

The frequency response of the filter D( z)D−1 ( − z) is, correspondingly
D(ω )
= e − jω /2 e j (ω −π )/2 = e − jπ /2 (32)
D(ω − π )
Comparing (27) and using the IIR filter notation (30) we obtain the Hilbert transform filter as
A( z )B( − z )
H( z ) = (33)
A( − z )B( z )
The Hilbert transform filter is inserted in the BF bank using the result of Lemma 1 (26). The
modified prototype BF filter bank is
H 0 ( z ) = H( z )H 0 ( z )
H 1 ( z ) = H −1 ( − z ) H 1 ( z )
(34)
G0 ( z ) = H −1 ( z )G0 ( z )
G1 ( z ) = H( − z )G1 ( z )
The BF bank (34) can be highly simplified by noting the following equivalents concerning on
(33)
H −1 ( − z ) = H( z )
(35)
H −1 ( z ) = H( − z )
By inserting (35) in (34) we obtain a highly simplified FB bank
H 0 ( z ) = H( z ) H 0 ( z )
H 1 ( z ) = H( z ) H 1 ( z )
(36)
G0 ( z ) = H( − z )G0 ( z )
G1 ( z ) = H( − z )G1 ( z )
The modified BF bank (36) can be realized by the Hilbert transform filter H( z ) , which works
as a prefilter for the analysed signal. The Hilbert transform filter H( − z) works as a postfilter
in the reconstruction stage, respectively. The wavelet sequences yielded by the two parallel
BF trees can be considered to form a complex wavelet sequence by defining the Hilbert
transform operator
Ha ( z ) = 1 + j H( z) (37)
By filtering the real-valued signal x[n] by the Hilbert transform operator results in an
analytic signal
x a [ n] = x[n] + j H{ x[ n]} (38)
whose magnitude response is zero at negative side of the frequency spectrum
 2 X (ω ) 0≤ω <π
X a (ω ) =  (39)
 0 −π ≤ω < 0
The wavelet sequence is obtained by decimation of the high-pass filtered analytic signal
1
W (ω ) = [ X a (ω ) H 1 (ω )]↓ 2 = W a (ω )↓ 2 = X a (ω / 2) H 1 (ω / 2) (40)
2
The result (40) means that the decimation does not produce aliasing but the frequency
spectrum is dilated by two. The frequency spectrum of the undecimated wavelet sequence
Wa (ω ) contains frequency components only in the range 0 ≤ ω < π , but the frequency
spectrum of the decimated analytic signal has the frequency band 0 ≤ ω < 2π . Hence, the
decimation does not produce overlapping and leakage (aliasing) to the negative frequency
range. A key feature of the dual-tree wavelet transform is the shift invariance of the decimated
analytic wavelet coefficients. The Fourier transform of the decimated wavelet sequence of the
1 − jωτ /2
fractionally delayed signal x[n − τ ] is e Wa (ω / 2) and the corresponding wavelet
2
sequence is w[n − τ / 2] . The energy (absolute value) of the decimated wavelet coefficients is
1
W (ω / 2) , which does not depend on the fractional delay τ . If the wavelet filter has linear
2
phase the wavelet coefficients are shift invariant in respect to their energy content.
An integer-valued half-delay filter D( z) = A( z) / B( z) is obtained by the B-spline transform
(see details Olkkonen & Olkkonen, 2007b). Table I gives the polynomial coefficients for the
B-spline orders K=4, 5 and 6. The frequency response of the Hilbert transform filter
constructed by the fourth order B-spline (Fig. 4) shows a maximally flat magnitude
spectrum. The phase spectrum corresponds to an ideal Hilbert transformer (15).
K A( z ) B( z )
−1 −2
1 + 6z +z 1 + 4 z −1 + z − 2
4
8 6
1 + 76 z −1 + 230 z −2 + 76 z −3 + z −4 1 + 11 z −1 + 11 z −2 + z −3
5
384 24
1 + 237 z −1 + 1682 z −2 + 237 z −3 + z −4 1 + 26 z −1 + 66 z −2 + 26 z −3 + z −4
6
3840 120
Table I. The half-delay filter polynomials for the B-spline transform order K=4, 5 and 6.
Fig. 4. Magnitude and phase spectra of the Hilbert transform filter yielded by the fourth
order B-spline transform.
6. Conclusion
It is well documented that the real-valued DWTs are not shift invariant, but small fractional
time-shifts may introduce significant differences in the energy of the wavelet coefficients.
Kingsbury (2001) showed that the shift invariance is improved by using two parallel filter
banks, which are designed so that the wavelet sequences constitute real and imaginary parts
of the complex analytic wavelet transform. The dual-tree discrete wavelet transform has
been shown to outperform the real-valued DWT in a variety of applications such as
denoising, texture analysis, speech recognition, processing of seismic signals and
neuroelectric signal analysis (Olkkonen et al. 2006; Olkkonen et al. 2007b).
Selesnick (2002) made an observation that a half-sample time-shift between the scaling
filters in parallel CQF banks is enough to produce the shift invariant wavelet transform. In
this work we reanalysed the condition and observed a phase-error term ω / 2 M + 1 (12)
compared with the ideal phase response θ (ω ) = −π / 2 . The phase error attains s highest
value at high frequency range and small stage M of the wavelet sequence. Fortunately, we
showed in this book chapter that the phase error term can be cancelled by adding a half-
delay prefilter in front of the CQF chain. For this purpose the half-delay filter
D( z) = A( z) / B( z) (30, Table I) constructed by the B-spline transform (Olkkonen & Olkkonen,
2007a) is well suited. In addition, there exists many other design methods for half-delay
filters (see e.g. Laakso et al. 1996; Johansson & Lowenborg, 2002; Pei & Tseng, 2003; Pei &
Wang, 2004; Tseng, 2006).
In multi-scale DWT analysis the complex wavelet sequences should be shift invariant. This
requirement is satisfied in the Hilbert transform-based approach (Olkkonen et al. 2006,
Olkkonen et al. 2007b), where the signal in every scale is Hilbert transformed yielding
strictly analytic and shift invariant transform coefficients. The procedure needs FFT-based
computation which may be an obstacle in many digital signal processor realizations. To
avoid this we conducted the novel shift invariant dual-tree BF bank (36) based on the
Hilbert transform filter (33). This highly simplified BF bank is yielded by Lemma 1 and the
equivalence (35) of the Hilbert transform filter (33). In many respects the BF bank (36)
outperforms the previous nearly shift invariant DWT approaches.
7. References
Daubechies, I. (1988). Orthonormal bases of compactly supported wavelets. Commmun. Pure
Appl. Math., Vol. 41, 909-996.
ITU-T (2000) Recommend. T.800-ISO DCD15444-1: JPEG2000 Image Coding System.
International Organization for Standardization, ISO/IEC JTC! SC29/WG1.
Johansson, H. & Lowenborg, P. (2002). Reconstruction of nonuniformy sampled bandlimited
signals by means of digital fractional delay filters, IEEE Trans. Signal Process., Vol.
50, No. 11, pp. 2757-2767.
Kingsbury, N.G. (2001). Complex wavelets for shift invariant analysis and filtering of
signals. J. Appl. Comput. Harmonic Analysis. Vol. 10, 234-253.
Laakso, T., Valimaki, V., Karjalainen, M. & Laine, U.K. (1996). Splitting the unit delay. Tools
for fractional delay filter design, IEEE Signal Processing Magazine, pp. 30- 80.
Olkkonen, H., Pesola, P. & Olkkonen, J.T. (2005). Efficient lifting wavelet transform for
microprocessor and VLSI applications. IEEE Signal Process. Lett. Vol. 12, No. 2, 120-
122.
Olkkonen, H., Pesola, P., Olkkonen, J.T. & Zhou, H. (2006). Hilbert transform assisted
complex wavelet transform for neuroelectric signal analysis. J. Neuroscience Meth.
Vol. 151, 106-113.
Olkkonen, H. & Olkkonen, J.T. (2007a). Half-delay B-spline filter for construction of shift-
invariant wavelet transform. IEEE Trans. Circuits and Systems II. Vol. 54, No. 7, 611-
615.
Olkkonen, H., Olkkonen, J.T. & Pesola, P. (2007b). FFT-based computation of shift invariant
analytic wavelet transform. IEEE Signal Process. Lett. Vol. 14, No. 3, 177-180.
Olkkonen, H. & Olkkonen, J.T. (2008). Simplified biorthogonal discrete wavelet transform
for VLSI architecture design. Signal, Image and Video Process. Vol. 2, 101-105.
Pei , T. S.C. & Tseng, C.C. (2003). An efficient design of a variable fractional delay filter
using a first-order differentiator, IEEE Signal Processing Letters, Vol. 10, No. 10, pp.
307-310.
Pei , S.C. &and Wang, P.H. (2004). Closed-form design of all-pass fractional delay, IEEE
Signal Processing Letters, Vol. 11, No. 10, pp. 788-791.
Selesnick, I.W. (2002). The design of approximate Hilbert transform pairs of wavelet bases.
IEEE Trans. Signal Process. Vol. 50, No. 5, 1144-1152.
Smith, M.J.T. & Barnwell, T.P. (1986). Exaxt reconstruction for tree-structured subband
coders. IEEE Trans. Acoust. Speech Signal Process. Vol. 34, 434-441.
Sweldens, W. (1988). The lifting scheme: A construction of second generation wavelets.
SIAM J. Math. Anal. Vol. 29, 511-546.
Tseng,C.C. ( 2006). Digital integrator design using Simpson rule and fractional delay filter,
IEEE Proc. Vision, Image and Signal Process., Vol. 153, No. 1, pp. 79-85.
14
Condition on Word Length of Signals and

Coefficients for DC Lossless Property of DWT
Masahiro Iwahashi1 and Hitoshi Kiya2
1Nagaoka University of Technology, Niigata,
2Tokyo Metropolitan University, Tokyo,
Japan
1. Introduction
A discrete wavelet transform (DWT) has been widely applied to various digital signal
processing techniques. It has been designed under a certain condition such as perfect
reconstruction, aliasing cancellation, regularity, vanishing moment, etc. This article
introduces a new condition referred to “DC lossless”. It guarantees lossless reconstruction of
a constant input signal (DC signal) instead of rounding of signal values and coefficient
values inside a transform. The minimum word length of the values under the new condition
is theoretically derived and experimentally verified.
Since JPEG 2000 algorithm based on the discrete wavelet transform (DWT) was adopted as
an international standard for digital cinema video coding [1], high speed and low power
implementation of a DWT has been becoming an issue of great importance [2,3]. In
designing a DWT, its coefficient values and signal values are assumed to be real numbers.
However, in implementation, they are rounded to rational numbers so that they are
expressed with finite word length representation in binary digit. Therefore it is inevitable to
have rounding errors inside a DWT processing unit.
In this article, we derive a condition on word length of coefficient values and that of signal
values of a DWT such that the transform becomes lossless for a DC signal. Under this
condition (DC lossless condition), it is theoretically guaranteed that an output signal
contains no error in spite of rounding of coefficients and signals inside the DWT. We treat
the irreversible 9-7 DWT adopted by the JPEG 2000 for lossy coding of image signals as an
example.
In case of the 5-3 DWT in JPEG 2000 for lossless coding, benefiting from its lifting structure
[4-6], lossless reconstruction of any signal is guaranteed even though signals and coefficients
are rounded. On the contrary, it does not hold for the 9-7 DWT because of scaling for
adjusting DC gain of a low pass filter in a forward transform [7]. However, we have pointed
out that it became possible to be lossless for a DC signal under a certain condition on word
length of coefficients and signals [8].
This DC lossless condition is a necessary condition for the regularity which has been
analyzed by numerous researchers to improve coding performance of a transform. When
the regularity is not satisfied, the DWT has some problems such as a checker board artifact
which is observed in a reconstructed signal as unnecessary high frequency noise in flat or
smooth region of a signal [9]. It also brings about DC leakage which decreases the coding
gain of a transform [10].
The regularity has been structurally guaranteed for a two channel quadrature mirror filter
bank (QMF) [9] and the DCT [10] respectively. However, since these previous methods were
based on the lattice structure, these are not directly applicable to the lifting structure of the
9-7 DWT. Beside these relations to the regularity, the DC lossless condition itself is also
considered to be important for white balancing of a video system in which the DC signal is
used as a reference input for calibration [11].
This article aims at deriving the DC lossless condition theoretically and clarifying the
minimum word length of signals and coefficients. In conventional analysis, errors due to
shortening of word length of signals (signal errors) were described as 'additive' to a signal
[7,12]. They were treated as independent and uniformly distributed white noise. On the
other hand, errors due to rounding of coefficients (coefficient errors) were described as
'multiplicative' to a signal and evaluated with the sensitivity [13-15]. It should be noted that
the signal error and the coefficient error have been treated independently. Unlike those
conventional approaches, we utilize mutual effect between rounding of signals and that of
coefficients. Introducing a new model which unifies the coefficient error and the signal
error, we define tolerance for those errors as a parameter to simultaneously control both of
word length of signals and that of coefficients.
As a result of our theoretical analysis, the minimum word length of signals and that of
coefficients inside the lifting 9-7 DWT are derived under the DC lossless condition. We
confirm that the minimum word length derived by our analysis is shorter than that
determined by a conventional approach. We also confirm that the DWT under the condition
does not have the checker board for a DC signal.
This article is organized as follows. Chapter 2 defines a rounding operation and a rounding
error, describes their basic properties in algebraic approach, and derives 'addition' formula
and 'multiplication' formula of the rounding (modulo) operation. Application of these
formulas to scaling of a signal value is introduced in chapter 3. Chapter 4 introduces the DC
lossless DWT. Its usefulness is also described. Derivation process of conditions on word
length of signals and coefficients is described in chapter 5. The new condition derived from
the basic properties in chapter 2 is summarized in chapter 6. Other related condition derived
from a conventional approach is also summarized. Theoretical results are verified and the
minimum word length of the DC lossless DWT is clarified in chapter 7. This article is
concluded in chapter 8.
2. Rounding operation and its basic formulas

This chapter introduces basic properties of the rounding operation focusing on 'quotient',
rather than 'remainder', in modulo operation. So far, 'remainder' had been attracted
numerous mathematicians' attention and various basic properties were found such as the
Chinese remainder theorem in the commutative algebra (commutative ring theory). On the
contrary, 'quotient' plays an important role as a 'practical' value in finite word length
implementation in modern computer systems. This chapter introduces an algebraic
approach of expressing 'quotient' as a practical value, and 'remainder' as a rounding error,
so that it can be applied to analyzing exact behavior of rounding errors in a complex
calculation procedure.
Condition on Word Length of Signals and Coefficients for DC Lossless Property of DWT 233
2.1 Definition of rounding operation and rounding error

In a digital calculation system, all the values of both of signals and coefficients are calculated
and stored as a binary digit with finite word length. In this article, we treat a case such that a
value x is expressed with a fixed point binary expression as
I 1
x b
p  F
p 2 p , bp  {0,1}, I  1, F  0, I  Z , F  Z (1)
where bp, p  {-F,  ,I-1), is a set of binary digit for a value x. It has I bit integer part
including one sign bit and F bit fraction part. Hereinafter, F is referred to as word length of a
value x. This F bit value x has a range expressed as
x  [ 2 I  1 , 2 I  1  2  F ]  [ 2 I  1 , 2 I  1 ) . (2)
For example, in case of I=1 and F=2, the maximum value is x=0.75 for [b0 b-1 b-2]=[0 1 1], and
the minimum value is x=-1.00 for [b0 b-1 b-2]=[1 0 0].
When an F bit signal value is multiplied with a coefficient value, in a convolution of a
filtering process in DWT for example, a resulting signal value has longer word length than
its original value. Therefore it is rounded to F bit again. So far there are various types of
rounding operations [16]. In this article, we deal with the rounding operation defined by
R0 [ x ]   x  2 1  or R0 [ x ]  x ' ( x ' mod 1) for x '  x  2 1 (3)
as an example. This rounding operation generates a rounding error. We denote it as
 0 [ x ]  x  R0 [ x ] or  0 [ x ]   ( x  2 1 ) mod 1   2 1 . (4)
Expanding these expressions to an F bit case, we can define the rounding operation and the
rounding error as
 RF [ x ]  R0 [ x 2 ]2
F F
 F
(5)
  F [ x ]   0 [ x 2 ]2
F
for an F bit word length implementation case.

Fig.1 illustrates rounding operations expressed by these equations. The term RF[x] is a
quotient, and ΔF[x] (= x - RF[x]) is related to a remainder. The former is an actual value
treated in a digital system under a finite word length implementation, and the latter is a
rounding error. We are now trying to develop an algebraic expression approach to exactly
trace a practical value and a rounding error in a convolution processing inside a DWT.
x 0 y  R0 [ x ] x F y '  RF [ x ]

 x2 1
  R0 [ x 2 F ]2  F
=
0
2F 2-F
(a) integer (b) F bit fraction
Fig. 1. Definition of the rounding operation and the rounding error. (a) An integer
implementation case. (b) An F bit word length implementation case.
2.2 Basic properties of the rounding operation

Since a convolution includes additions and multiplications, we should know behavior of an
addition of two values x and y. Resulting value is RF[x+y] and its rounding error is ΔF[x+y].
A multiplication result RF[xy] and its error ΔF[xy] should be also investigated.
First of all, let's derive basic properties of the rounding operation starting with an obvious
property;
y  Z  R0 [ x  y ]  R0 [ x ]  y for xR . (6)
It represents that only a real number x can be rounded if y is an integer to calculate a

rounded value of x+y. In this case, its rounding error becomes
y  Z  0[x  y ]  0[x ] for x R . (7)
It suggests that an integer y can be ignored when only the rounding error is considered in an
analysis. There is another obvious property;
R0 [ x ]  0  x  [ 2 1 , 2 1 ) . (8)
Since the range of a rounding error is
 0 [ x ]  [ 2 1 , 2 1 ) , (9)
we can add two more identities;
 R0   0 [ x ]  0,
 (10)
 0   0 [ x ]   0 [ x ].
The equations above for F=0 can be straightforwardly extended to an F≠0 case as follows.
 RF [ x  y ]  RF [ x ]  y
y 2F  Z   for x R (11)
 F [ x  y ]   F [ x ]
RF [ x ]  0  x  [ 2 1  F , 2 1  F ) (12)
 F [ x ]  [ 2 1  F , 2 1  F ) (13)
 RF   F [ x ]  0
 (14)
 F   F [ x ]   F [ x ]
In addition, Eq.(12) can be extended to a more general case with an integer n as
RF [ x ]  n2  F  x 2 F   2 1  n , 2 1  n  for n  Z . (15)
2.3 Basic formulas of the rounding operation

Utilizing the basic properties in Eqs.(11)-(14), we can derive an addition formula and a
multiplication formula of a practical value (quotient) of the rounding operation as follows.
Addition formula
RF [ x  y ]  RF [ x ]  RF  y   F [ x ] for x, y  R (16)
Proof:
RF  x  y 
 RF  RF [ x ]   F [ x ]  y   (4)
 RF [ x ]  RF   F [ x ]  y   (11)
Q.E.D.
Multiplication formula
RF [ xy ]  RF  xRF [ y ]  RF  x F [ y ]   F  xRF [ y ]  for x, y  R (17)

 
Proof:
RF [ xy ]
 RF  x F [ y ]  xRF [ y ]  (4)
 RF  x F [ y ]   F  xRF [ y ]  RF  xRF [ y ]   (4)
 
 RF x F [ y ]   F  xRF [ y ]  RF  xRF [ y ]
   (11)
 
Q.E.D.
Formulas for a rounding error (remainder) can be also derived as
 F [ x  y ]   F   F [ x ]   F [ y ]
 (18)
  F [ xy ]   F   F [ x ]RF [ y ]  RF [ x ] F [ y ]   F [ x ] F [ y ]
for real numbers x and y. These formulas have following variations;

Addition formula
 RF [ x  y ]  RF [ x ]  y

y 2 F  Z   RF [ x  y ]   F [ x ]  x  y (19)
 [ x  y ]   [ x ]
 F F
Multiplication formula
 RF [ xy ]  RF   F [ x ]y   RF [ x ]y

y 2  Z   RF [ xy ]   F   F [ x ]y   xy
F
(20)

 F [ xy ]   F   F [ x ]y 
Especially when two kinds of word lengths are mixed in a signal processing, the following
variation of the multiplication formula is conveniently applied to analyzing behavior of
signals and errors in a pair of encoder and decoder [17].
RF2  xRF1 [ y ]  RF2 [ xy ]  RF2   x F1 [ y ]   F2 [ xy ] (21)
Proof:
RF2  xRF1 [ y ]

 R0   x F1 [ y ]2 F2  xy 2 F2  2  F2
 R0   x F1 [ y ]2 F2   0 [ xy 2 F2 ]  R0 [ xy 2 F2 ] 2  F2
 R0   x F1 [ y ]2 F2   0 [ xy 2 F2 ] 2  F2  R0 [ xy 2 F2 ] 2  F2
 RF2   x F1 [ y ]   F2 [ xy ]  RF2 [ xy ]
Q.E.D.
3. Application of the formulas to basic signal processing

This chapter applies the formulas to some basic signal processing cases.
3.1 Mapping invariant condition

Fig.2 illustrates a scaling of a signal value x with a coefficient value h. As illustrated in
Fig.2(a), this processing maps an input value x to an output value y* with an ideal (infinite
word length) coefficient value h. Note that x has F bit word length. Output value of the
multiplication is also rounded to F bit (y* has F bit word length). In implementation, as
illustrated in Fig.2(b), a coefficient value h is also rounded to W bit word length (h' has W bit
word length). We aim at finding the minimum word length W of a coefficient h' such that
the mapping is invariant (y -y* =0).
x y* x y
F F F F
h h'=RW[h]
h∈real number h'∈rational number

W → ∞ [bit] W → min. [bit]
(a) assumption (b) implementation
Fig. 2. Scaling of a signal value x with a coefficient value h. (a) This processing maps x to y*
with h under a given F. (b) A mapped y should be equal to y* even though h is rounded to h'.
In case of Fig.2(b), an input value x is multiplied by a rounded value RW[h] (=h') of a given
coefficient h. The result RW[h]x is rounded to RF[RW[h]x] (=y). When it is the same as RF[hx]
(=y* ), the mapping of x is invariant. It means that effect of rounding of h is nullified. This
mapping invariant case is expressed as
Em  RF  RW [ h ]x   RF [ hx ]
Em  0 for  . (22)
 RW [ h ]  h   W [ h ]
From the basic properties, the mapping invariant condition is derived as
 W [ h ]x   F [ hx ]   2 1  F , 2 1  F  . (23)
Proof:
RF  RW [ h ]x   RF [ hx ]
 RF  hx   W [ h ]x   RF [ hx ]
 RF   F [ hx ]   W [ h ]x   RF [ hx ]  RF [ hx ]
 RF   F [ hx ]   W [ h ]x 
0
  F [ hx ]   W [ h ]x   2 1  F , 2 1  F 
Q.E.D.
The Eq.(23) also means
   F [ hx ]  2 1  F [ hx ]  2 1 
 W [h]   , , x  0
  x x 
 (24)
   F [ hx ]  2 1
 F [ hx ]  2 
1
 W [h]   x
,
x
, x  0
  
which gives tolerance to the rounding error of a coefficient [8]. This is the mapping invariant
condition on word length W of a coefficient under a given word length F of signals. It
represents exact (not approximated) behavior of rounding errors.
Unlike the condition above, a sufficient condition can be derived by substituting the upper
bound of errors and signals;
 W [ h ]  2 1  W ,  F [ hx ]  2 1  F , x  2 I 1 , (25)
to Eq.(23). It results in the condition described as
W F  I 1. (26)
This condition is too strict and requires too long word length to guarantee the mapping
invariance. In both cases of Eq.(24) and Eq.(26), the mapping invariant condition determines
the minimum of word length W of a coefficient under a given word length F of signals.
3.2 Lossless condition on a scaling pair

Fig.3 illustrates a pair of two multipliers. In Fig.3(a), an input signal x has F2 bit word length.
It is scaled with a coefficient h1, and its output value y is rounded to F1 bit. It is re-scaled
with a coefficient h2 (=1/h1), and its final output value w is rounded to F2 bit. This scheme is
embedded in a forward transform and a backward transform of DWT for example. It is
required to regain w exactly the same as x, under a given word length set of F1 and F2. Note
that rounding errors due to finite word length expression of coefficients h1 and h2 can be
ignored as far as the mapping invariant condition is satisfied.
x y w x y w
F2 F1 F2 F2 F1 F2
h1 h2 (=1/h1) h1 h2
x, y, w ∈ real number x, y, w ∈ rational number
F1 → ∞ [bit] F1 → min. [bit]
(a) assumption (b) implementation
Fig. 3. Scaling pair has two coefficients h1 and h2 (=1/h1). (a) Output w is exactly the same as
its original x. (b) This lossless property is guaranteed under a condition on F1 and F2.
We apply the formulas and the properties to derive the condition on F1 and F2. The lossless
case in Fig.3(b) is described as
Ep  RF2  h2 RF1 [ h1 x ]  x

Ep  0 for  . (27)
h1 h2  1
From the basic properties, the lossless condition on a scaling pair is derived as
h2  F1 [ h1 x ]   2 1  F2 , 2 1  F2  . (28)
Proof:
RF2  h2 RF1 [ h1x ]  x
 R0  h2  F1 [ h1x ]2 F2  h2 h1 x 2 F2  2  F2  x
 R0  h2  F1 [ h1x ]2 F2  2  F2  x  x
 RF2  h2  F1 [ h1x ]

0
  h2  F1 [ h1 x ]   2 1  F2 , 2 1  F2 
Q.E.D.
This condition determines the word length F1 and F2 of signals for an input value x. It
represents exact condition such that total accumulated rounding error is nullified by the
rounding just after the final multiplier with h2. As a result, the original value x is recovered
as the final output without any loss.
Unlike the exact condition above, a sufficient condition can be derived by analyzing the
upper bound as follows.
h2  F1 [ h1 x ]  h2 2 1  F1  2 1  F2 (29)
As a results, when the sufficient condition given by
F1  F2  log 2 h2 (30)
holds, the scaling pair becomes lossless. However this condition is too strict and requires too
long word length of signals.
4. Application of the formulas to DWT

This chapter introduces the DC lossless DWT [8, 18]. Definition and its usefulness are also
described. The algebraic approach based on the formulas is applied to derive conditions on
word length of signals and coefficients. Derivation process is described in chapter 5.
4.1 DWT and its word length of signals and coefficients

Fig.4 illustrates the irreversible 9-7 DWT of the JPEG 2000 standard [1]. The forward
transform in Fig.4(a) decomposes an input signal x(n), n N, N={n | 1,2,  , L} into band
signals y1(m) and y2(m), m M, M={m | 1,2,  , L/2}. The backward transform in Fig.4(b)
reconstructs the signal w(n) from the band signals. In the figure, z-1 and ↓2 indicate the delay
and the down sampler respectively.
x(n)
FX ↓2 FB
+ + c6 y1(m)
FS FS
1+z
1+z
c2 c4
z
c1 c3 1+z-1
1+z-1
FS FS c5 y2(m)
+ +
↓2 FB
(a) Forward transform
FS FX ↑2
y1(m) c5 - - + w(n)
FS FS
1+z
1+z
c4 c2
z-1
c3 c1
-1
1+z-1
1+z
FS FS
y2(m) c6
- -
FS FX ↑2
(b) Backward transform
Fig. 4. The irreversible 9-7 DWT of the JPEG 2000 standard.
The multiplier coefficients ci, i I, I={i | 1,2,  , 6} are designed under the word length long
enough to be treated as real numbers. When the DWT is implemented, coefficient values are
rounded to the length as short as possible to minimize total hardware complexity. Similarly,
signal values are also rounded. In the figure, fraction part of each signal is shortened to FS,
FB or FX [bit] by a rounding operation illustrated as a circle.
Denoting the integer part as IS [bit], total word length WS [bit] of a signal s is defined as
WS  I S  FS  1 (31)
including 1 [bit] for the sign part. Similarly, total word length WC [bit] of a coefficient c is
defined as
WC  IC  FC  1 . (32)
In Fig.4, fraction part of the input signal x(n) is given as FX [bit]. Inside the DWT, fraction
part of the signals are rounded to FS [bit] just after each of all the multiplications with ci, i I.
Output signals from the forward and backward transforms are rounded to FB [bit] and FX
[bit] respectively. Note that we do not truncate integer part of signals and that of
coefficients. We are determining FS and FC such that the DC lossless property is satisfied.
4.2 Definition of DC lossless property and its necessity

In this article, we define the DC lossless as the conjunction of the following two
propositions:
n  N, m  M  x(n)  d  y1 ( m)  d  y 2 (m)  0  (33)
n  N, m  M  y 1 (m)  d  y 2 (m)  0  w(n)  d  (34)
for a given constant value d with FX [bit] fraction part. When the proposition in Eq.(33)
holds, the DWT has no DC leakage for the DC input signal with value d. Similarly, when the
proposition in Eq.(34) is true, the reconstructed signal w(n) contains no checker board
artifact for the DC input signal. In the following chapters, we investigate the minimum
fraction part of signals FS [bit] which guarantees the DC lossless for given FX and FB [bit].
We also investigate the minimum fraction part FCi [bit] of a coefficient ci, i I with flexibility
of trading off the signal error and the coefficient error.
Fig.5(a) illustrates an example of a video system. It contains an encoder and a decoder which
are composed of a forward DWT and a backward DWT. In white balancing, a camera and a
display are calibrated with a constant valued input signal (DC signal) [11,19]. Therefore, it is
useful for this calibration if the forward DWT and its backward do not generate any error. In
this case, the camera and the display can be calibrated ignoring existence of the encoder and
the decoder as illustrated in Fig.5(b). Namely, the DC lossless condition provides a low
complexity DWT useful for the white balancing.
input output
DC signal DC signal
Camera Display
Camera Display
x(n) w(n)
adjust adjust
Encoder Decoder
(a) video system (b) calibration
Fig. 5. The DC lossless property is useful for white balancing in a video system.
In addition, the DC lossless condition is a necessary condition for the regularity which
controls smoothness of basis functions and coding performance of a transform. A DWT
under the regularity does not generate the checker board artifact or the DC leakage. Harada
et. al. analyzed a condition for the regularity of a two channel quadrature mirror filter bank
(QMF) [9]. They confirmed that a QMF under the condition has reduced checker board
artifact for an input step signal. It is expanded to a multirate system under short word
length expression [20]. The regularity was structurally guaranteed for a biorthogonal linear
phase filter bank [21,22] and the DCT [10] respectively. However, since these previous
methods are based on factorization of a transfer function including (1+z-1) or (1-z-1) in the
lattice structure, these are not directly applicable to the lifting structure of the 9-7 DWT in
Fig.4.
In this article, we derive the DC lossless condition theoretically in chapter 5, and determine
the minimum word length of signals and that of coefficients under the condition in
following chapters.
5. Derivation of condition for the DC lossless DWT

This chapter describes derivation process of the DC lossless condition.
5.1 New model for error analysis

Fig.6(a) illustrates a multiplier in the DWT circuit. An input value s has FS [bit] fraction part
and multiplied by a coefficient c'. The coefficient is originally designed as a real number c. It
is rounded to a rational number c' in implementation. It produces the coefficient error:
c  c  c ' . (35)
Just after the multiplication, the signal is rounded to s' with FS [bit] fraction part as
s '  RFS [c ' s]  c ' s  e ' (36)
where e' is the signal error. From Eq.(35) and (36), the final output becomes
s '  cs  cs  e ' . (37)

where cs is the ideal output. This conventional model, illustrated in Fig.6(b), describes the
coefficient error c as multiplicative to the signal s [13-15], and the signal error e' as
additive [7,12]. In addition, these errors are treated independently as mutually uncorrelated
noises.
Unlike these existing approaches, as illustrated in Fig.6(c), we describe the coefficient error
e'' as
s '  RFS [cs ]  e '',

 (38)
 e ''  RFS   FS [cs ]   FC [c ]s  .
From Eq.(15), we utilize the fact that e'' is observed as a 'particle';
e '' 2 FS  p (39)
where p is an integer. Given the tolerable maximum to an integer p, word length of the
coefficient c can be controlled independently of other coefficients in other sections inside the
DWT. Furthermore, denoting the signal error as e' similarly to Eq.(36), the output value is
described as
s '  cs  e
 (40)
 e  e ' e ''
where
 e '   FS [cs]
 (41)
 
 e ''  RFS   FS [cs ]   FC [c ]s 
as illustrated in Fig.6(d). In this new model, both of the coefficient error e'' and the signal
error e' are unified to the error e. Utilizing Eqs.(13) and (15), its absolute value is limited
to
|e| ( p  2 1 )2  FS . (42)
Note that the parameter p to control word length of a coefficient c is included in this
equation. It is equivalent to
2  FC  FS  IS  1  p (43)
where its proof is given in appendix.

Benefitting from this inequality, it becomes possible to consider mutual effect of the
coefficient error and the signal error.
e'
s c' s' s c s'
FS FS FS
c'  c  c -Δc e'  2 1 FS
(a) multiplier (b) conventional model
c e'' c e
s s' s s'
FS FS FS
e' '  p 2  FS e  ( p  2 1 )2  FS
(c) new model I (d) new model II
Fig. 6. A multiplier in the DWT and its models for error analysis.
Inside the forward DWT, the error e is propagated and added up with other errors from
other multipliers. When its maximum absolute value is less than 2 1  FB , the total error is
nullified by the rounding at the final output of the forward DWT. In this article, we utilize
this nullification of errors at output of the DWT to derive a condition on word length such
that the DC lossless defined by Eq.(33) and (34) is satisfied.
5.2 DC equivalent circuit

When the input signal is restricted to a DC signal, x(n) can be described as a scalar x
independent of n. The delay z-1 can be treated as 1 and (1+z-1) can be replaced by 2.
Therefore, instead of the circuits in Fig.4, we can use their equivalent circuits for a DC input
signal in Fig.7 to derive the condition.
In Fig.7(a), a scalar x with FX [bit] fraction part is multiplied by the rational numbers ci, i I
and rounded to FS [bit]. Finally, the signals are rounded to FB [bit] at its output to produce
two scalars [y1 y2]. The unified errors inside the circuit are described as
ei  ei ' ei '' , i  I (44)
where
ei '  FS [ci si ]



ei ''  RFS FS [cisi ]  FC [ci ]si 
s1 s3 s5   2x 2(x  s '2 ) s4 

s s s   2(x  s ' ) 2(s  s ' ) s  s ' 
 2 4 6  1 2 3 3 4
s 'i  ci si  ei  RFS [ci si ] .
Similarly, for the backward transform in Fig.7(b), errors are described as
f i  f i ' f i '' , i  I (45)
where
 f i '  FS [citi ]



 f i ''  RFS FS [citi ]  FC [ci ]ti 
t1 t3 t5  2(t3  t '2 ) 2(t '5  t '4 ) y1 

t t t   2(t  t ' ) 2t '6 y2 
 2 4 6  4 3
t 'i  citi  fi  RFS [citi ] .
Similarly to Eq.(42), these errors are described with the parameters pi and qi to control word
length of coefficients as
ei  ( pi  21 )2 FS , i  I (46)
fi  (qi  21 )2FS , i  I (47)
for a given word length Fs [bit] of signals.

x(n)
FX ↓2 FB
x s1 s'2 + s3 s'4 + s6 c e x1 y1
6 6
2c1 e2 2c3 e4
z e1 2c2 e3 2c4
x s'1 + s2 s'3 + s4 s5 c 5 e5 x2 y2
↓2 FB
(a) Forward transform

w(n)
FB FX ↑2
y1 c f t'5 t'4 - t3 t'2 - t1 w1 +
5 5
f4 2c3 f2 2c1
2c4 f3 2c2 f1 z-1
y2 c6 f6 t'6 t4 t'3 - t2 t'1 - w2

FB FX ↑2
(b) Backward transform
Fig. 7. Equivalent circuits of the DWT for a DC input signal.
5.3 Nullification of accumulated errors

In Fig.7(a), the unified errors in Eq.(46) are propagated and accumulated inside the circuit.
When the accumulated errors are nullified by the rounding at output of the forward
transform, Eq.(33) is satisfied. In the figure, Y12=[y1 y2]T is described as
Y12  RFB  IU e6  I L e5  K

 IU e4  H 4  I L e3
(48)
 H 3  IU e2  H 2  I L e1  H1IUL x    


where
IU  [1 0]T , I L  [0 1]T , IUL  IU  I L
1 0 1 2 c j  c6 0
Hi{1,3}   , H j{ 2 ,4}   , K  0
1 c 5 
.
2 c i 0 1  
It is described with the unified error matrices E1 and E2 as
Y12  RFB (H e 1E1  H e 2 E 2 )  KH 4321 x  (49)
where
H e 1  IU KIU KH 43 IU 

H e 2  I L KH 4 I L KH 432 I L  ,

H 43  H 4H 3 
and
E 1  [ e6 e4 e2 ]T

E 2  [ e5 e3 e1 ]T .
Similarly, output values W12=[w1 w2]T from the backward transform in Fig.7(b) are
W12  RFX  ( H e 3E 3  H e 4 E 4 )  (KH 4321 )1 Y12  (50)
where
H e 3  [H11IU H123 1
IU H1234 1
IU ]
 1 1
H
 e4  [ I L H I
12 L H I
1234 L ]
 1 1 1
H
 12  H 1 H 2 
and
E 3  [ f 2 f4 f 5 ]T

E 4  [ f 1 f3 f 6 ]T .
When the DWT is DC lossless, output values of the transforms are
 Yˆ 12   KH 4321 x   IU 
      x. (51)
ˆ 1 ˆ
 W12  (KH 4321 ) X 12  IUL 
Using this equation, the accumulated errors are defined as
ˆ 
 E y 12   Y12   Y
E    W    ˆ  .
12 (52)
 w 12   12   W12 
Substituting Eqs.(49), (50), (51) and using the property in Eq.(6), we have
 E y 12   RFB ( H e 1E 1  H e 2 E 2 ) 
(53)
  .
 E w 12   RFX (H e 3E 3  H e 4 E 4 ) 

Applying Eq.(12), it becomes clear that when the conditions;
 H e 1E1  H e 2 E 2  IUL 2 1  FB
 1  F
(54)
 H e 3E 3  H e 4 E 4  IUL 2 X
are satisfied, the accumulated errors are nullified by the rounding operations at the final
output of each of the forward transform and the backward transform.
6. Derived conditions on word length for the DC lossless DWT

This chapter summarizes the new condition derived from the basic properties in chapter 2,
and other related condition derived from a conventional approach.
6.1 Critical condition on word length

Finally, we derive the condition on word length of coefficients and signals such that Eq.(54)
is satisfied. Since the unified errors in E1, E2, E3 and E4 have the maximum in Eqs.(46) and
(47) described with the parameters pi and qi, the DC lossless condition is also described with
the parameters by substituting
 E1   [ p6 p4 p2 ]T  I 3  2 1  2  FS


 E 2   [ p5 p3 p1 ]T  I 3  2 1  2  FS
 (55)
 E 3   [ q2 q4 q5 ]T  I 3  2 1  2  FS

 E 4   [q1 q3 q6 ]T  I 3  2 1  2  FS
for
I 3  [1 1 1]T
into Eq.(54). This is the condition we derived based on the new model described in section
5.1. We investigate the fraction part FCi [bit] of a coefficient ci , i  I as the minimum word
length under the condition for a DC value x at the word length Fs [bit] of signals.
6.2 Sufficient condition on word length

As an example, in case of all the parameter in Eq.(55) are given as pi = qi = p and FCi = FC for
 i  I, the condition in Eq.(54) becomes
  H e 1  He 2 2  FS
( p  2 1 )  IUL 2 1  FB
L1 L1
 (56)
  H e 3 2  FS
L1
 He 4 L1
( p  2 1 )  IUL 2 1  FX
where H L1 denotes a column vector whose component is a sum of absolute value of all
components in each row. Substituting coefficients of the 9-7 DWT [1] into Eq.(56), we
have
 p  2 S E  2
1  F  G 1
 (57)
GE  2.66 [bit]
for FX = FB = 0. As a result, the DC lossless condition on the word length is given as
 log 2 (2 WC  2 WS )  GE (58)
where
[ WC WS ]  [FC  I S FS ]
and GE is the lower bound. This means a sufficient condition for the DC lossless. Since it is
too strict, the word length under this condition is redundant. Unlike this sufficient
condition, our critical condition given as Eq.(54) under Eq.(55) determines the word length
minimum and necessary for the DC lossless.
This chapter verifies theoretically derived conditions, and clarifies the minimum word
length of the DC lossless DWT.
7.1 Word length under the sufficient condition

Utilizing the sufficient condition in section 6.2, we calculated the optimized word length
under the cost function defined as J=2-1(FC +FS). The cost J is minimized for three examples.
Ex.1 trades the word length between FC and FS, namely FC + FS = constant. Ex.2 and Ex.3 are
FC = FS and WC = WS respectively. Results are summarized in table 1. Table 2 summarizes
word length of signals and coefficients for an 8 bit system with Wx=8 (Ix=7 and Fx=0). Ex.1
requires (FS, FC)=(4, 12) [bit] for signals and coefficients respectively. Ex.2 and Ex.3 require
FS =FC =11 [bit] and WS =WC =14 [bit] respectively. The condition in Eq.(58) is plotted as a
solid line in Fig.8. According to the sufficient condition, it is impossible to be DC lossless for
FC  GE  I S
 (59)
FS  GE
and it is also confirmed by the figure. It guarantees the DC lossless, however the condition is
too strict. Therefore the word length is redundant and there is room for further reduction.
Ex.1 Ex.2 Ex.3

FC+FS = const. FC = FS WC = WS
FC GE+1+IS GE+IC*+IS -IC
FS GE+1 GE+IS* GE+IC*
WC GE+2+IS +IC GE+IS*+1+IC
WS GE+2+IS GE+IS*+1+IS GE+IC*+IS +1
J GE+1+IS/2 GE+IS* GE+IC*+(IS -IC) /2

IS*=log2(2IS +1), IC*=log2(2IC +1)
Table 1. Theoretically derived word length under the sufficient condition for DC lossless. FS
and FC denote fraction part of signals and coefficients. IS and IC denote integer part of signals
and coefficients. WS=IS+FS+1, WC=IC+FC+1, J is a cost function.
Ex.1 Ex.2 Ex.3

FC+FS = const. FC = FS WC = WS
FC 11.66 11.25
FS 3.66 10.67 4.25
WC 13.66 12.67
WS 12.66 19.67 13.25
J 7.66 10.67 7.75
Table 2. Word length calculated with equations in table 1 for Wx=8 [bit].
7.2 Word length under the critical condition

In Fig.8, a cross "  " indicates a pair (FS, FC) which satisfies the critical condition in section
6.1 for any 8 bit integer x with WX=IX=8 and FX=0. Here in after, we denote an input DC
value to a video system as
xin  x  2 I X  1 (60)
where x is an input value to the DC equivalent circuit in Fig.7. The minimum of FC for each
FS is indicated as a broken line. It is clear that the word length derived by the critical
condition is shorter than that determined by the sufficient condition. For example in
Fig.8(a), the fraction part FS (= FC) is reduced from 11 [bit] to 9 [bit] for Ex.2. The word
length is not shortened for Ex.1 and Ex.3. In case of Fig.8(b), FS (= FC) is reduced from 13 [bit]
to 12 [bit] for Ex.2. (FS, FC) is reduced from (14, 4) to (13, 3) or (12, 4) for Ex.1. WS (= WC) is
reduced from 16 [bit] to 15 [bit] for Ex.3. It is confirmed that the word length is shortened
due to the analysis in this article.
20
15
Fc [bit]
Ex.2 Sufficient
10
Ex.1
Minimum
Ex.3
5
x∈[-27, 27)
A
0 5 10 15
Fs [bit]
(a) xin  [0, 2 8 ), WX  8 [bit]
20
15
Sufficient
Fc [bit]
10 Minimum
5
x∈[-29, 29)
A
0 5 10 15
Fs [bit]
(b) xin  [0, 2 10 ), WX  10 [bit]
Fig. 8. Word length under the two conditions. "  " indicates (FS, FC) such that the DWT
becomes DC lossless.
7.3 Word length for a specific value

Fig.9(a) and Fig.9(b) illustrate the word length under the conditions for the black value "16"
and the white value "235" respectively. These specific values are utilized in white balancing
of an 8-bit video system [19]. For example, (FS, FC) is reduced from (4, 12) in Fig.8(a) to (2, 9)
in Fig.9(a) for Ex.1. Table 3 summarizes the minimum word length for these specific input
DC values [23]. It is observed that the word length can be reduced by limiting input DC
signals to a specific value. Fig.10 indicates the minimum word length FC of coefficients for
an input value x at a given word length FS of signals. This is an example at (FS, WX)=(3, 8).
The sufficient condition gives the same word length for any of input values. Unlike this
conventional statistical analysis, our analysis gives the minimum word length shorter than
that determined by the sufficient condition for each of input DC values.
20
15
Fc [bit]
Sufficient
10
Minimum
5
(FS ,FC)=(2, 9) for xin=16
0 5 10 15
Fs [bit]
(a) black value
20
15
Fc [bit]
Sufficient
10
Minimum
5
(FS ,FC)=(3, 9) for xin=235
0 5 10 15
Fs [bit]
(b) white value
Fig. 9. Word length under the two conditions for a specific value used in white balancing.
20
Fs=3
15 Sufficient
min.Fc [bit]
10
Minimum
5
x∈[1,128]
20 40 60 80 100 120
input x
Fig. 10. The minimum word length of coefficients for each of input DC values at (FS, WX)=(3,
8). According to the sufficient condition, the word length is too long.
forward transform and backward transform

signals coefficients
input DC values integer fraction integer fraction
IS [bit] FS [bit] IC [bit] FC [bit]
sufficient 4 12
WX= 8 bit
any xin∈[0,28) 2 12
8 2
xin= 16 (black) 2 9
xin=235 (white) 3 9
sufficient 4 14
WX=10 bit
any xin ∈[0,210) 2 13

10 2
xin= 64 (black) 0 8
xin=940 (white) 0 12
Table 3. The minimum word length for a specific value for white balancing of a video
system.
7.4 Optimum word length assignment

Since we described tolerance for the unified errors as parameters pi and qi in Eq.(46) and (47),
it becomes possible to simultaneously control both of word length of signals and that of
coefficients. Table 4 summarizes these parameters for an input value 16 and the word length
of signals at FS=2 [bit] as an example. It indicates that [p1 p2  p6] in Eq.(46) are [0 0 0 0 0 1]
for the word length of coefficients at FC=9 [bit]. In this case, all the coefficients ci in the
forward transform have the same length. It is worth paying attention to the fact that the
parameter p1 is the same for FC=9, 8 and 7 [bit] for example. It means that word length of the
coefficient c1 can be reduced from 9 to 7 [bit] without any influence to the errors. Therefore,
word length [FC1 FC2  FC6] of coefficients [c1 c2  c6] can be reduced from [9 9 9 9 9 9] to [7 9
7 4 6 4] according to the table.
Table 5 summarizes results of this optimum word length assignment for the forward
transform. Comparing to table 3, it is observed that word length of coefficients is reduced
from 9.00 [bit] to 6.17 [bit] on average for an input value xin=16. Table 6 summarizes results
for the backward transform. In this case, the word length is furthermore shortened. It is
observed that c6 and c4 can be omitted since y2 is equal to zero under the DC lossless. Fig.11
illustrates image signals reconstructed by the DWT which does not satisfy the DC lossless
condition. It demonstrates the checker board artifact for reference. It is confirmed that total
word length is furthermore shortened utilizing the tolerance parameters pi and qi introduced
in this article.
FC p1 p2 p3 p4 p5 p6 y1-x y2
9 0 0 0 0 0 1 0 0
8 0 -3 0 0 0 0 -1 -1
7 0 -3 0 0 0 0 -1 -1
6 7 12 -8 0 0 2 6 6
5 7 -18 10 0 1 0 -7 -5
4 -21 -18 9 0 -2 1 -10 -12
3 35 107 7 25 9 -28 63 70
Table 4. Tolerance parameters in Eq.(46) and (47) for xin=16 and FS=2 as an example.
forward transform
input signals coefficients

values
FS FC1 FC2 FC3 FC4 FC5 FC6 ave.
xin= 16 (B) 2 7 9 7 4 6 4 6.17
WX=8
xin=235 (W) 3 9 7 9 9 1 4 6.50
xin= 64 (B) 0 7 9 7 1 1 4 4.83

WX=10
xin=940 (W) 0 7 11 -1 1 1 4 3.83
Table 5. The minimum word length of coefficients in the forward transform for a specific
values xin for a given word length of signals FS.
backward transform
signals coefficients
input
values
FS FC1 FC2 FC3 FC4 FC5 FC6 ave.
xin= 16 (B) 2 9 9 7 0 8 0 5.50

WX=8
xin=235 (W) 3 9 9 7 0 8 0 5.50
xin= 64 (B) 0 7 9 7 0 8 0 5.17

WX=10
xin=940 (W) 0 7 12 8 0 8 0 5.83
Table 6. The minimum word length of coefficients in the backward transform for a specific
values xin for a given word length of signals FS.
(a) 1 stage (b) 2 stages
(c) 3 stage (d) 4 stages

Fig. 11. Example of reconstructed images for 1282 pixel DC input image with x=10. Intensity
is multiplied by 16.
8. Conclusions
Introducing a new model which unifies the coefficient error and the signal error, and
utilizing the nullification of the accumulated errors, this article theoretically derived a
condition on word length of signals and coefficients such that the 9-7 DWT of JPEG 2000
becomes lossless for a DC input signal. It was confirmed that the minimum word length
derived by the newly introduced 'critical' condition was shorter than that determined by a
conventionally well known 'sufficient' condition. It was also confirmed that the DWT under
the condition does not have the checker board for a DC signal. Analysis in this article
contributes to build a low complexity DC lossless DWT.
9. Appendix
Proof of Eq.(43)
Eq.(39) with Eq.(41) means
 FS [cs ]   FC [c ]s  ( p  2 1 )2  FS (A.1)
according to Eqs.(13) and (15). Applying the triangle inequality to the left hand side, we have
 FS [cs ]   FC [c ]s   FS [cs ]   FC [c ] s . (A.2)
According to Eq.(13), each terms in the right hand side are described as
  FS [cs ]  2 1  FS

 (A.3)
1  F
  FC [c ]  s  2 C 2 S
I
Therefore (A.2) and (A.3) under (A.1) means
2 1  FS  2 1  FC 2 IS  ( p  2 1 )2  FS
and finally we have Eq.(43) as
2  FC  FS  IS  1  p .
Q.E.D.
10. References
[1] ISO/IEC FCD15444-1, "JPEG2000 image coding system," March 2000.
[2] Descampe, et.al., "A flexible hardware JPEG 2000 decoder for digital cinema," IEEE Trans.
circuits, systems on video technology, vol.16, issue 11, pp.1397-1410, Nov. 2006
[3] Bing-Fei Wu, Chung-Fu Lin, "Memory-efficient architecture for JPEG 2000 coprocessor
with large tile image, IEEE Trans. circuits and systems II, vol.53, issue 4, pp.304-
308, April 2006.
[4] H. Kiya, M. Yae, M. Iwahashi, "Linear phase two channel filter bank allowing perfect
reconstruction", IEEE Proc. international symposium on circuits and systems
(ISCAS), no.2, pp.951-954, May 1992.
[5] W. Sweldens, "The lifting scheme: A custom-design construction of biorthogonal
wavelets," Technical Report 1994:7, industrial mathematics initiative, department of
mathematics, university of South Carolina, 1994.
[6] M. L. Bruelers, A. W. M. van den Enden, "New Networks for Perfect Inversion and
Perfect Reconstruction," IEEE Journal of selected areas in communications, vol.10,
no.1, pp.130-137, Jan.1992.
[7] M. Reza, Lian Zhu, "Analysis of error in the fixed-point implementation of two-
dimensional discrete wavelet transforms," IEEE Trans. circuits and systems,
fundamental theory and applications, vol.52, issue 3, pp.641-655, March 2005.
[8] Kiya , M. Iwahashi, O. Watanabe, "A new class of lifting wavelet transform for
guaranteeing losslessness of specific signals," IEEE international conference,
acoustics, speech, and signal processing (ICASSP), pp.3273-3276, March 2008.
[9] Y. Harada, S. Muramatsu, H. Kiya, "Two channel QMF bank without checker board
effect and its lattice structure," IEICE Trans. on fundamentals, vol.J80-A, no.11,
pp.1857-1867, Nov. 1997.
[10] Wei Dai, T. D. Tran, , "Regularity-constrained pre- and post- filtering for block DCT-based
systems," IEEE Trans. signal processing, vol.51, Issue 10, pp.2568- 2581, Oct. 2003.
[11] Hirakawa, T. W. Parks, "Chromatic adaptation and white balancing problem," IEEE
Proc. international conference, image processing (ICIP), vol. III, pp.984-987, Nov.
2005.
[12] Grangetto, et.al., "Optimization and implementation of the integer wavelet transform
for image coding," IEEE Trans. Image Processing, vol.11, Issue 6, pp. 596-604, June
2002.
[13] Xiao, et.al., "Coefficient sensitivity and structure optimization of multidimensional
state-space digital filters," IEEE Trans. circuits, systems I, vol.45, issue 9, pp.993-
998, 1998.
[14] S. Yamaki, M. Abe, M. Kawamata, "A closed form solution to L2-sensitivity
minimization of second-order state-space digital filters subject to L2-scaling
constraints," IEICT Trans. fundamentals of electronics, communications and
computer sciences, vol.E91A, no.7, pp.1697-1705, July 2008.
[15] Y. Tonomura, S. Chokchaitam, M. Iwahashi, "Minimum hardware implementation of
multipliers of the lifting wavelet transform," IEEE Proc. international conference,
image processing (ICIP), pp.2499-2502, Oct. 2004.
[16] IEEE Standard 754-1985, IEEE standard for binary floating-point arithmetic.
[17] M. Iwahashi, H. Kiya, "Finite word length error analysis based on basic formula of
rounding operation", the international symposium on intelligent signal processing
and communication systems (ISPACS), no.86, pp.49-52, Dec. 2008.
[18] M. Iwahashi, H. Kiya, "Word length condition for DC Lossless DWT," Asia pacific
signal and information processing association (APSIPA) annual summit and
conference, no.TA-P2-6, pp.469–472, Oct. 2009.
[19] The society of motion picture and television engineers, "Standard for television,
1920x1080 image sample structure, digital representation and digital timing
reference sequences for multiple picture rates", SMPTE 274 M-2005, Feb.2005.
[20] H. Iwai, M. Iwahashi, K. Kiya, "Methods for avoiding the checkerboard distortion
caused by finite word length error in multirate system", IEICE Trans.
fundamentals, vol. E93-A, no.3, pp.631-635, March, 2010.
[21] S. Oraintara, T.D. Tran, T.Q. Nquen, "A class of regular biorthogonal linear-phase
filterbanks: Theory, structure, and application in image coding," IEEE Trans. signal
processing, vol.51, no.12, pp.3220-3235, Dec.2003.
[22] Y. Tanaka, M. Ikehara, "First order linear phase filter banks with regularity constrains
for efficient image coding," IEICE Trans. fundamentals, vol. J91-A, no.2, pp.192-201,
Feb. 2008.
[23] M. Iwahashi, H. Kiya, "A lossless condition of lifting DWT for specific DC values",
IEEE International Conference on Acoustics, Speech, and Signal Processing
(ICASSP), pp.1458-1461, March 2010.
15
Wavelet-Based Analysis and

Estimation of Colored Noise
Bart Goossens, Jan Aelterman, Hiêp Luong, Aleksandra Pižurica
and Wilfried Philips
Ghent University - TELIN-IPI-IBBT
Belgium
1. Introduction
Digital imaging devices inevitably produce images corrupted with noise. The noise originates
from the sensors and analogue circuitry in the camera. In order to have better and sharper
images and also for commercial reasons, there is a recent tendency to further increase the
image resolution. Nowadays, cameras with more than 20 megapixels are not uncommon. To
reach such a high number of megapixels, the area of the sensor elements must be decreased
and correspondingly the elements become more sensitive to noise, resulting in a lower image
quality due to noise.
During the last decades, the use of image processing techniques has become widespread. The
increasing processing power of computers allows for more sophisticated techniques that are
better adapted to the classes of images under consideration (e.g. photographic images or
medical images). This also allows for new classes of techniques that alleviate the physical
limitations of the sensor elements by means of post-processing such as denoising. Because
of power and hardware complexity constraints, the post-processing techniques implemented
by camera manufacturers are based on simplistic assumptions with respect to the assumed
noise model: for example, while it is well known that photon signals are Poisson distributed,
the techniques most often rely on a white Gaussian noise model. In practice, such model
mismatches generally lead to inferior denoising results. Also, many factors cause the noise
in practice to be colored instead of white (i.e. with a flat power spectrum). For example, the
image formation is often a reconstruction process based on an insufficient number of samples,
and missing samples need to be estimated using interpolation techniques (e.g. Bayer pattern
demosaicing). Doing so, the noise becomes colored. A technique that is designed to remove
white Gaussian noise may offers a image quality: either some noise artifacts may be left in the
image, or the noise is suppressed too much, leading to an overblurred image.
The obvious solution to this problem is to adapt existing techniques to use a colored noise
model that is well matched to the underlying sensor characteristics and/or reconstruction.
Therefore, estimation of the noise statistics is indispensable. Stationary colored noise (or
correlated noise) is completely described by its Power Spectral Density (PSD). The noise PSD
describes the power distribution of the noise in frequency space and can be estimated by
using the Discrete Fourier Transform (DFT). However, noisy images also contain information
other than noise (e.g. edges and textures), and directly estimating the PSD through the DFT
256
2 Discrete Wavelet Transforms: AlgorithmsDiscrete
and Wavelet
Applications
Transforms
will yield seriously biased estimates caused by the signal presence. Alternatively, the PSD
could be estimated from noise-only patches in the image. However, not all images contain
such patches and also the number of noise samples that can be used for this task is often too
limited to yield reliable PSD estimates. Hence, more specialized techniques are needed.
The discrete wavelet transform (DWT) is an important tool for developing such techniques.
The DWT provides a non-uniform partitioning of the space-frequency plane, which allows
positional information of structures to be included in the estimation. This is not possible with
the DFT, since the DFT cannot recover information at specified positions in the image.
In this chapter, we investigate the estimation of colored noise. First, we discuss a number of
origins for colored noise in images. Next, we explain the importance of wavelets in solving
the estimation problem. To proceed, it is necessary to know how the wavelet-domain and
spatial-domain autocorrelation functions are related to each other, since we are aiming at
estimating the wavelet-domain autocorrelation function. Because the wavelet transform in
general does not fully decorrelate signals as we will explain, noise-free wavelet coefficients
with significant magnitudes can still be found near high-frequent transitions in the signals (for
example, near edges in images). To benefit from prior knowledge in a statistical estimation
approach, we will discuss a number of wavelet domain prior models. Two iterative EM-based
techniques will be presented, to estimate the wavelet-domain autocorrelation function. Next,
we will explain how the parameters of a parametric noise PSD can be estimated using the
presented tools. Finally, we will give a number of experimental results for the proposed
techniques.
1.1 From white noise to colored noise

Throughout this chapter, we will consider a stationary additive Gaussian noise process:
y (p) = x (p) + w (p) (1)
where x (p) is a pixel intensity of a noise-free image at position p ∈ Z2 , y(p) is the

corresponding observed pixel intensity and w(p) is a zero-mean additive noise component.
w(p) and x (p) are mutually statistically independent. We will further assume that the samples
w(p) are generated by a (wide-sense) spatial stationary process w, in which the correlation
between two noise samples only depends on the position difference between the two noise
samples, but not on their absolute position. Consequently, w can be completely described by
the mean and the autocorrelation function.
A wide-sense stationary random process w obeying the above conditions is called white if its
autocorrelation function is a Dirac delta function:

R w (p) = E w (p ) w (p + p ) = δ ( p) . (2)
For colored noise, neighboring noise samples are not statistically independent, hence spatial
dependencies exist between these samples. Their dependencies can be characterized by the
autocorrelation function of the noise, which is - for colored noise - different from the Dirac
delta function.
The PSD is a related descriptor of colored noise. More specifically, the PSD describes how the
noise energy is distributed in frequency space. According to the Wiener-Khinchin theorem,
the power spectral density is the (discrete time) Fourier transform of the autocorrelation function
Wavelet-Based Analysis
Wavelet-based Analysis and Estimation of and
ColoredEstimation
Noise of Colored Noise 2573
1 0 1
Normalized vertical frequency ωy/π
20
0.5 −5 0.5
10
0 −10 0 0
y
−10
−0.5 −15 −0.5
−20
−1 −20 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
Normalized horizontal frequency ωx/π x
(a) (b)
Fig. 1. Noise in PAL broadcasting. (a) Power Spectral Density [dB], (b) Noise signal
(containing horizontal stripe patterns due to correlations).
R w (p):
Pw (ω ) = ∑ Rw (p) exp − jω T p . (3)
p ∈Z 2
White noise has a flat PSD: Pw (ω ) = 1. Suppose a filter with frequency response H (ω ) = 1 is

applied to the noise signal, then the resulting PSD Pw (ω ) becomes Baher (2001):

Pw (ω ) = Pw (ω ) | H (ω )|2 . (4)

Clearly, the PSD P (ω ) is subjected to the filter magnitude response | H (ω )|. Hence one can
think of correlated noise as white noise subjected to linear filtering. In analogy with the term
“white noise” the resulting term is called “colored noise” (or correlated noise, because the filtering
introduces correlations in the noise samples).
In practical circumstances, there are a number of origins of colored noise in images:
• Phase Alternating Line (PAL) television: the noise in PAL television images is a good example
of colored noise. The correlations between the noise samples are caused by several
mechanisms, such as deinterlacing Kwon et al. (2003), demodulation and filter schemes.
In Figure 1, the PSD of a noise patch from a PAL broadcast is shown. Here, there is a high
concentration of energy in the lower horizontal frequencies, leading to horizontal stripes
and artifacts.
• Color interpolation (demosaicing): modern digital cameras use a rectangular arrangement of
photosensitive elements. In this matrix arrangement, photosensitive elements of different
color sensitivity are placed in an interleaved way. This allows sampling of full color
images without the use of three arrays of photosensitive elements. One popular example
is the Bayer pattern Bayer (1976). Color interpolation (or demosaicing) is the process of
estimating the values of missing photosensitive elements.
• Post-processing techniques: image noise often becomes correlated by the use of
post-processing techniques, e.g., image quality enhancement techniques, sharpening
filters, digital zoom functions of cameras, JPEG compression...
258
and Wavelet
Applications
Transforms
π 1
0.8
50
0.6
100
ωy
0.4
150
0.2
200
−π 0
250 −π π
ωx
50 100 150 200 250
(a) (b)
Fig. 2. (a) Image corrupted with colored noise caused by demosaicing (b) PSD of the noise in
the green color channel of (a).
• Thermal cameras: images captured by thermal cameras of the push broom or whisk broom
type often exhibit streaking noise artifacts, mainly caused by detector and sampling
circuitry Aelterman, Goossens, Pižurica & Philips (2010). This kind of noise can be
approximated using a 1/ f frequency characteristic (called pink noise) Borel et al. (1996).
Pink noise also frequently arises in image sensors that acquire pixel data in time.
• Medical imaging: in computed tomography (CT), noise correlations are introduced
by the specific reconstruction technique that is being used. Noise created by the
backprojection algorithm (without reconstruction filter) is called ramp-spectrum noise,
and has an 1/ f frequency characteristic. Noise in magnetic resonance imaging
(MRI) is traditionally considered white Nowak (1999); Pižurica et al. (2003), although
many MRI scanner manufacturers have included a wide range of techniques to
allow for shorter scanning times (mainly to avoid patient motion artifacts in the
images). To name a few: K-space subsampling, partial Fourier, elliptical filtering
Aelterman, Deblaere, Goossens, Pižurica & Philips (2010). The use of these techniques
results in correlated noise in the reconstructed MRI images.
In Figure 3 another example is shown of an image corrupted with colored noise. The colored
noise
was artificially generated by subjecting white noise to a filter with magnitude response
P (ω ) and subsequently by adding the filtered noise to the images.
2. Wavelets for the estimation of colored noise

Spatially stationary colored noise can be directly specified through its mean and
autocorrelation function and/or power spectral density. Given an observed noise signal w(p),
the estimation of these parameters is then a relatively simple task by, e.g., using the sample
mean and sample autocovariance estimates. However, in practice, it often happens that the
observed signal also contains information other than noise, this underlying signal is unknown
and it is the signal that we eventually want to estimate. Hence, we are observing y(p) instead
of w(p). The estimation of the noise statistics from the signal y(p) is then considerably more
difficult.
ColoredEstimation
ωy
−π
−π ωx π
(a) (b)
Fig. 3. Illustration of the noise PSD: (a) Image with correlated noise, (b) The noise PSD (in
frequency domain, the center of the image is the origin of frequency space, white
corresponds with low noise powers, black with high noise powers).
70 20
60
Time domain 15
Wavelet domain
50 10
40 5
30 0
20 -5
70 20
10 -10
60 15
0 -15
50 10
-10
0 200 400 600 800 1000 1200 1400 1600 y(p)
-20
0 100 200 300 400 500 600 700 800 ~
y(p)
x(p)
40
~
x(p)
5
+ 30 + 0
w(p) ~
w(p)
20 -5
10 20
10 -10
8
15
0 -15
6
10
4 -10 -20
0 200 400 600 800 1000 1200 1400 1600 0 100 200 300 400 500 600 700 800
5
2
0 0
-2
-5
-4
-10
-6
-15
-8
-10 -20
0 200 400 600 800 1000 1200 1400 1600 0 100 200 300 400 500 600 700 800
(a) (b)
Fig. 4. Example of a piecewise linear signal with correlated noise. Our goal is to estimate the
noise power spectrum from the corrupted signal y(p). (a) The signals in time domain, (b)
The finest scale of the wavelet transform of the signals (Daubechies’ wavelet with 2
vanishing moments was used).
This problem is illustrated in Figure 4 for a piecewise linear signal corrupted with correlated
Gaussian noise. While the noise statistics can be easily estimated from w(p), we only have
the degraded signal y(p) at our disposal, which also contains an unknown signal component.
A straightforward solution is then to first estimate the signal x̂ (p), to subtract it from y(p)
and finally to estimate the noise statistics from the difference y(p) − x̂ (p). However, optimal
estimation of x (p) from y(p) requires knowledge of the noise statistics on its own, so we have
a chicken-and-egg problem. The common approach is then to use iterative techniques, which
first estimate x̂ (p) and then later refine this estimate x̂ (p) when better estimates for the noise
parameters become available.
In this chapter, we will take a different approach by relying on the properties of wavelets. The
wavelet transform Daubechies (1992); Mallat (1999) analyzes signals according to different
260
and Wavelet
Applications
Transforms
scales and at different points in time. Starting from a fixed mother wavelet ψ (t), the input
signal is correlated with time-shifted and time-stretched (dilated) versions of this wavelet.
Correlations with wavelets with a large dilation factor then give the coarse features of the
signal, while correlations with wavelets with small dilation factors give the fine signal details.
Because the wavelet basis functions are well localized in time or space (this is in contrast to
the basis functions of e.g., the Fourier transform), wavelets are ideal candidates for analyzing
non-stationary signals, having statistical properties that vary in time (or space).
The Daubechies wavelets are a class of orthogonal wavelets for which the number of vanishing
moments for a given support is maximal. More specifically, the n-th moment of a real-valued
wavelet function ψ (t) is defined by:
ˆ +∞
μn = tn ψ (t)dt. (5)
−∞
The Daubechies wavelet of support 2N (with N vanishing moments) will have moments μ n =
0 for 0 ≤ n < N. Now, let us denote the time-shifted and dilated basis functions of ψ (t) by:

1 t−b
ψa,b (t) = √ ψ (6)
a a
√
where a is the dilation factor, b is a time shift, and the constant 1/ a is an energy
normalization factor. The continuous wavelet transform of a signal f ∈ L2 (R ) is defined
by:
ˆ +∞
W f ( a, b ) = f (t)ψa,b (t)dt. (7)
−∞
Now, suppose that a signal is linear on a region larger than the support S ( a) of the wavelet
function ψa,b (t):
f (t) = c · t if |t − b | ≤ S ( a).
For Daubechies wavelets with at least two vanishing moments (N ≥ 2), the corresponding
wavelet coefficient W f ( a, b ) will be zero:
ˆ +∞
W f ( a, b ) = c · tψa,b (t)dt
−∞
ˆ +∞
c t−b
= √ tψ dt
a −∞ a
ˆ +∞
√

=c a at + b ψ t dt
−∞
ˆ +∞ ˆ +∞

= ca3/2 t ψ t dt + cba1/2 ψ t dt

−∞ −∞
=0
In the remainder of this chapter, for the ease of notation, we will consider one particular
wavelet subband (with scale a) at a time and we will denote the corresponding wavelet
coefficients by a tilde: for example x̃ (p) are the wavelet coefficients for that particular scale
of x (p). The process can then be repeated for other subbands as well. Let us now apply a
ColoredEstimation
Daubechies wavelet transform to the piecewise linear signal from Figure 4(a). The result is
shown in Figure 4(b) for the finest scale of the DWT1 : because of the vanishing moments of
the wavelet, the wavelet coefficients x̃ (p) are zero, except at the positions where the derivative
of x (p) does not exist. At these positions, the wavelet coefficients have a negligibly small
magnitude. This nicely illustrates the sparsifying properties of the DWT for this type of signal.
Correspondingly, the wavelet coefficients ỹ(p) are (approximately) w̃(p), which means that
the chicken-and-egg problem is solved: the noise statistics can be directly estimated from
ỹ (p)! More specifically, the wavelet domain autocorrelation function of w(p) can in this case
be estimated based on the following relationship:

Rw̃ (p) ≈ Rỹ (p) = E ỹ(p )ỹ(p + p ) . (8)
It then suffices to compute the sample autocorrelation function of ỹ(p). There are now two
issues remaining, which we will explain in the remainder of this Chapter:
1. The autocorrelation function of a signal in the wavelet domain (e.g. a for particular wavelet
subband) is not the same as the autocorrelation function of a signal in time domain.
Nevertheless, there exists a simple relation between both, as we will explain in Section
3.
2. Most real-life signals are not piecewise linear functions or piecewise polynomials. For such
signals, the wavelet coefficient magnitudes may become non-negligible, causing serious
biases to the final noise estimates. An example of a frequency modulated signal with
maximal frequency at half length of the signal, is given in Figure 5. Because of the high
local bandwidth of the signal at this time position, the wavelet is not able to cancel out
the signal, resulting in wavelet coefficients with a large magnitude. Consequently, the
approximation ỹ (p) ≈ x̃ (p) does not hold anymore. However, it can be seen in Figure
5(b) that this phenomenon is well localized in time, hence, because the noise process is
assumed to be stationary, a plausible solution would be to estimate the noise statistics from
the wavelet coefficients ỹ(p) that have a small underlying components x̃ (p) (ignoring the
outliers in Figure 5(b)). In Section 4 we will discuss solutions that generalize this idea by
using a statistical prior model for wavelet coefficients.
So far, we discussed the estimation of colored noise for one dimensional signals. The reasoning
can also be extended to higher dimensional signals, such as images. To illustrate this, a
noisy image together with its DWT are shown in Figure 6. It can be seen that the wavelet
subbands (LH, HL and HH in Figure 6) predominantly contain information on the noise,
with exception in the areas of textures and edges (the fine hairs of the mandrill). In these
areas, the (noise-free) wavelet coefficients x̃ (p) still have a relatively large magnitude, but this
phenomenon is localized - in the surrounding smooth regions the wavelet coefficients ỹ(p)
mostly consist of noise.
For higher dimensional signals, the DWT is usually computed by using basis functions that
are tensor products of one dimensional wavelets and one dimensional scaling functions.
While this approach can efficiently deal with point-wise singularities (e.g. bumps, dots,
...), most structures in images are line-like singularities with a given direction. However,
the DWT can not well adapt to the arbitrary direction of the singularity: for example, the
1 Note that for other scales the plots are similar.

262
and Wavelet
Applications
Transforms
70 20
60
Time domain 15
Wavelet domain
50 10
40 5
30 0
20 -5
70 20
10 -10
60 15
0 -15
50 10
-10
0 200 400 600 800 1000 1200 1400 1600 y(p)
-20
0 100 200 300 400 500 600 700 800 ~
y(p)
x(p)
40
~
x(p)
5
+ 30 + 0
w(p) ~
w(p)
20 -5
10 20
10 -10
8
15
0 -15
6
10
4 -10 -20
0 200 400 600 800 1000 1200 1400 1600 0 100 200 300 400 500 600 700 800
5
2
0 0
-2
-5
-4
-10
-6
-15
-8
-10 -20
0 200 400 600 800 1000 1200 1400 1600 0 100 200 300 400 500 600 700 800
(a) (b)
Fig. 5. Example of a non-piecewise linear signal with correlated noise. Our goal is to estimate
the noise power spectrum from the corrupted signal y(p). (a) The signals in time domain, (b)
The finest scale of the wavelet transform of the signals (Daubechies’ wavelet with 2
vanishing moments was used).
transform can not make a distinction between features oriented at +45° and -45°. This
is known as the checkerboard problem of the DWT: due to the separability of the higher
dimensional wavelets, these wavelets appear as a checkerboard pattern which does not have
a dominant direction. Consequently, many nonzero wavelet coefficients may be needed to
represent a line singularity at an arbitrary orientation. To overcome this limitation there has
recently been a lot of interest in transforms that offer a better directional selectivity. Examples
are steerable pyramids Simoncelli et al. (1992), dual-tree complex wavelets Selesnick et al.
(2005a), Marr-like wavelet pyramids Van De Ville & Unser (2008), 2-D (log) Gabor transforms
Fischer et al. (2007); Lee (1996), contourlets Do & Vetterli (2005), ridgelets Candès (1998);
Do & Vetterli (2003), curvelets Candès et al. (2006) and shearlets Guo & Labate (2007). These
transforms are designed to have better sparsifying properties so that our outlier problem in
Figure 5(b) is alleviated (but not solved).
In the next subsections we will focus on the DWT as a primary multiresolution decomposition
tool, however, the same reasoning can also be applied to more recently developed transforms.
3. From time-domain to wavelet-domain autocorrelation functions

Because our goal is to estimate the autocorrelation function of noise in the wavelet
domain, it is very useful to know how the wavelet-domain and time-domain autocorrelation
functions are related to each other. When the autocorrelation function of the input signal
is known, a simple Monte-Carlo based technique is to generate colored noise with this
given autocorrelation function, then to transform the noise to the wavelet domain (or other
multiresolution transform domain) and subsequently to estimate the autocorrelation function
in this domain Portilla et al. (2003). While such computational method is attractive from an
implementation point of view, it does not bring a direct analytical relationship between both
autocorrelation functions. We will see in Section 6 that an analytical relationship will prove to
be very useful when estimating parametric noise PSDs.
Let us consider the wavelet analysis filterbank shown in Figure 7(a), where a signal with
z-transform F̃1 (z) is filtered by a wavelet filter G (z) and a scaling filter H (z). Both signals
ColoredEstimation
LL2 HL2
LH2 HH2 HL1
LH 1 HH1
(a) (b)
Fig. 6. (a) Baboon image with noise, (b) DWT of the image.
are subsequently decimated by a factor of two. The analysis is iterated on the scaling
coefficients F2 (z). Now, the input signal has an autocorrelation function in the z-domain
defined by R̃1 (z) = E F̃1 (z) F̃1 (z−1 ) . The filtered signals then have autocorrelation functions
respectively R̃1 (z) G (z) G (z−1 ) and R̃1 (z) H (z) H (z−1 ). Decimating the resulting signals by a
factor 2 leads to the signal with autocorrelation function Goossens et al. (2010):

R1 ( z) = E F1 ( z) F1 ( z−1 )
1 1 1 −1 1 1 1

= R̃1 z 2 G z 2 G z 2 + R̃1 − z 2 G − z 2 G − z− 2 ,
2

R2 ( z) = E F2 ( z) F2 ( z−1 )
1 1 1 −1 1 1 1

= R̃1 z 2 H z 2 H z 2 + R̃1 − z 2 H − z 2 H − z− 2 . (9)
2
Hence, the wavelet-domain autocorrelation function R1 (z) can be directly computed from the
autocorrelation function of the input signal R̃1 (z) and the wavelet and scaling filters. This
involves two simple convolutions and a decimation operation of the input autocorrelation
function R̃1 (z). For subsequent decompositions (coarser scales of the wavelet transform), this
process can be iterated by re-inserting R̃1 (z) = R2 (z) in (9).
To show that this reasoning also applies to other wavelet transforms, we will briefly discuss
the adaptation to the dual-tree complex wavelet transform (DT-CWT) Kingsbury (2001) in
one dimension. Extension to higher dimensions is then straightforward. The 1D DT-CWT is
implemented using two parallel DWT filter banks, the first filter bank uses the real parts of
the complex wavelet and scaling filters (respectively G1 (z) and H1 (z)), while in the second
filter bank, the imaginary parts of the wavelet and scaling filters (respectively G2 (z) and
H2 (z)) are applied. Finally, the output of both filter banks are mixed together (see the right
square in Figure 7(b)), applying a 45° rotation in the complex plane. This last step is in fact
only necessary in 2D (or higher dimensions), where complex wavelets are constructed using
tensor products of 1D complex wavelets. The translation of the resulting complex-valued
filter banks to parallel real-valued filter banks then automatically results into this phase
modulation in the complex plane (for more details, see Selesnick et al. (2005b)). Defining
264
and Wavelet
Applications
Transforms
F1 ( z) +
F̃1 ( z) - G1 (z) - ↓n
2 - √1 - +n- F1 (z)
2

F2 ( z)
F̃2 ( z) - G2 (z) - ↓n
2 - √1 -U +n- F2 (z)
2
−
F̃1 ( z) - G ( z) - ↓n
2 - F1 ( z)
- H1 (z) - ↓n
2 -
Computation of
F3 ( z) complex coefficients
- H ( z) - ↓n
2 - F2 ( z) - H2 (z) - ↓n
2 - F4 ( z)
Recursive part Recursive part

(a) (b)
Fig. 7. Analysis filterbank for (a) the DWT, (b) the DT-CWT.

R̃2 (z) = E F̃2 (z) F̃2 (z−1 ) , application of (9) to the DT-CWT leads to the following equations:
1 1 1 1 1 1 1
R1 ( z) = E F1 ( z) F1 ( z−1 ) = R̃1 z 2 G1 z 2 G1 z− 2 + R̃1 -z 2 G1 -z 2 G1 -z− 2
2
1 1 1 1 1 1 1
R2 ( z) = E F2 ( z) F2 ( z−1 ) = R̃2 z 2 G2 z 2 G2 z− 2 + R̃2 -z 2 G2 -z 2 G2 -z− 2
2
1 1 1 1 1 1 1
R3 ( z) = E F3 ( z) F3 ( z−1 ) = R̃1 z 2 H1 z 2 H1 z− 2 + R̃1 -z 2 H1 -z 2 H1 -z− 2
2
1 1 1 1 1 1 1
R4 ( z) = E F4 ( z) F4 ( z−1 ) = R̃2 z 2 H2 z 2 H2 z− 2 + R̃2 -z 2 H2 -z 2 H2 -z− 2
2
1 1 1 1 1 1 1
S1,2 ( z) = E F1 ( z) F2 ( z−1 ) = S̃1,2 z 2 G1 z 2 G2 z− 2 + S̃1,2 -z 2 G1 -z 2 G2 -z− 2 (10)
2
where S̃1,2 (z) is the cross-power spectrum between F̃1 (z) and F̃2 (z): S̃1,2 (z) =

E F̃1 (z) F̃2 (z−1 ) . The final autocorrelation functions (after the complex phase modulation)
are computed from R1 (z), R2 (z) and S1,2 (z), as follows:

1 1
R1 ( z) = E F1 ( z) F1 ( z−1 ) = ( R1 ( z) + R2 ( z)) + S1,2 ( z) + S1,2 ( z−1 ) ,
2 2

1 1
R2 ( z) = E F2 ( z) F2 ( z−1 ) = ( R1 ( z) + R2 ( z)) − S1,2 ( z) + S1,2 ( z−1 ) . (11)
2 2
In Algorithm 1, an O CTAVE /M ATLAB program is given for computing the autocorrelation
functions in case of the DWT and DT-CWT, according to (9) and (10)-(11). In this program,
the variables lo and hi respectively signify the scaling and wavelet coefficients. It can
be seen that all operations are linear operations, which makes it possible to express the
conversion from time-domain to wavelet-domain as a matrix multiplication applied to the
input autocorrelation coefficient vector.
In Figure 8, an example of a parametric autocorrelation function and its DWT decomposition,
according to (9), is shown. Due to the cone of influence (Mallat, 1999, p. 174), the support
size of the autocorrelation function decreases when increasing the wavelet scale (i.e., when
analyzing finer scales). Interesting to note is the envelope of the noise variance in the wavelet
domain: the noise variance is identical to the noise autocorrelation function evaluated in
the origin (which is in this case also the maximum of the autocorrelation function). When
one modifies the center band frequency of the noise PSD in Figure 8(b), this also directly
ColoredEstimation
11
Algorithm 1 O CTAVE /M ATLAB program for computing wavelet domain autocorrelation

functions.
f = [1 2 1]; % input autocorrelation function
% d i s c r e t e w av elet t r a n s f o r m (DWT)
l o = conv ( f , conv ( h , h ( end : − 1 : 1 ) ) ) ;
h i = conv ( f , conv ( g , g ( end : − 1 : 1 ) ) ) ;
l o = l o ( 1 : 2 : end ) ; h i = h i ( 1 : 2 : end ) ;
% dual −t r e e complex w av elet t r a n s f o r m ( DT−CWT)

l o 1 = conv ( f , conv ( h1 , h1 ( end : − 1 : 1 ) ) ) ;
h i 1 = conv ( f , conv ( g1 , g1 ( end : − 1 : 1 ) ) ) ;
l o 2 = conv ( f , conv ( h2 , h2 ( end : − 1 : 1 ) ) ) ;
h i 2 = conv ( f , conv ( g2 , g2 ( end : − 1 : 1 ) ) ) ;
c r1 = conv ( f , conv ( h1 , h2 ( end : − 1 : 1 ) ) ) ; % c ros s− c o r r e l a t i o n
c r2 = conv ( f , conv ( h2 , h1 ( end : − 1 : 1 ) ) ) ;
l o 1 = l o 1 ( 1 : 2 : end ) ; l o 2 = l o 2 ( 1 : 2 : end ) ; % d e c i m a t i o n s
h i 1 = h i 1 ( 1 : 2 : end ) ; h i 2 = h i 2 ( 1 : 2 : end ) ;
c r1 = c r1 ( 1 : 2 : end ) ; c r2 = c r2 ( 1 : 2 : end ) ;
h i 1 _ o u t = 0 . 5 * ( h i 1 + h i 2 ) + 0 . 5 * ( c r1 + c r2 ) ; % complex phase m odulat ion
h i 2 _ o u t = 0 . 5 * ( h i 1 + h i 2 ) − 0. 5 * ( c r1 + c r2 ) ;
1
β=1
0.9 β=1.1
β=1.2
β=1
β=1.3
Wavelet − scale 4 β=1.1 0.8
β=1.2
β=1.3
β=1.4 0.7
Wavelet − scale 3
0.6
|R(ejω)|/β
0.5
Wavelet − scale 2
0.4
0.3
Wavelet − scale 1
0.2
0.1
Time domain
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
ω/pi [Normalized frequency]
(a) (b)
Fig. 8. (a) Wavelet analysis
ofthe
autocorrelation function (in z-domain)
β2
R(z) = ∑n π ( n2 − β2 ) 1 + cos πn β zn across different scales and for different values of β.
Daubechies’ wavelet with two vanishing moments was used. (b) Power spectral density
R(e jω ) for different values of β.
influences the noise variances of the individual wavelet subbands (see Figure 8(a)), due to
the frequency-selective behavior of the wavelets at different scales. For example, increasing
the parameter β has as effect that the noise variance at wavelet scale 4 decreases. This also
suggests that, when a thresholding strategy (e.g. soft/hardthresholding) would be used to
suppress the colored noise process, the thresholds would need to be level-dependent, e.g., as
proposed by Johnstone and Silverman Johnstone & Silverman (1997).
266
and Wavelet
Applications
Transforms
4. Statistical priors for noise estimation

As already illustrated in Figure 5, the DWT will in general not fully suppress the signal.
Consequently, wavelet-based noise estimation techniques need to take into account that the
wavelet coefficients contain a non-negligible signal component. One of the earliest and
well-known wavelet-based noise estimation techniques is the MAD estimator from Donoho,
which estimates the noise standard deviation as follows Donoho & Johnstone (1995):
Medianp (|ỹ (p)|)

σ̂ = . (12)
0.6745
The estimator gives level dependent estimates of the noise standard deviation in every
wavelet subband. Based on robust statistics, the non-zero signal coefficients are considered to
be outliers. By computing a median instead of a more traditional mean, the outlier influences
in the end result are significantly reduced.
In this chapter, we are interested in estimating the noise correlations or covariances (next to
the noise standard deviation), therefore the estimator (12) can not directly be used. For this
purpose, a general class of robust S estimators for the covariance (see, e.g., Campbell et al.
(1998); Pena & Prieto (2001)) can be used. These estimators detect outliers after finding
projections that maximize the kurtosis of the data. An illustration of such a technique is given
in Figure 9: the robust S estimators attempt to estimate the covariance of the noise (the black
dots in Figure 9). In this case this is equivalent to determining the sizes of the axes and the
orientation of the ellipse shown in the figure (the ellipse can be seen as an isocontour of the
probability function of the data). Because of the presence of outliers (the crosses in Figure 9),
this is not a trivial task. The robust estimation techniques then try to identify the outliers, in
an iterative process.
While robust S estimators are unfamiliar with the structure of the data they are applied to,
in our application, we have some more information on the data that we can take into our
advantage. In particular, due to the sparsifying properties of the chosen multiresolution
transform, the identification of the outliers (signal components) is somewhat easier: the
multiresolution transform already performs a projection to maximize the kurtosis. Instead
of relying on robust statistics, we will incorporate prior knowledge on the noise-free wavelet
coefficients to further improve the estimation performance using Bayesian techniques. Our
noise estimation approach will then consist in 1) specifying a statistical prior distribution for
the noise-free signal coefficients, 2) maximum likelihood estimation of the unknown noise
covariance matrix.
In the next subsections, we will briefly review a number of statistical models for noise-free
wavelet coefficients and we will explain how these models can be used to perform noise
estimation.
4.1 The generalized Laplace distribution

It has been found in several studies Field (1987); Mallat (1989) that histograms of wavelet
coefficients (or generally coefficients of bandpass filtered images) have a highly kurtotic
shape. An example is shown in Figure 10(a)-(b) for the Baboon image: the wavelet
coefficient histogram reveals a sharp peak and a heavy tail. The sample kurtosis of
the wavelet coefficients (6.98) is much higher than the theoretical kurtosis of a Gaussian
distribution (which is 3). Several authors Antonini et al. (1992); Chang et al. (1998); Mallat
ColoredEstimation
13
50
y(p+1)
0
−50
−50 0 50
y(p)
Fig. 9. Joint histogram of neighboring wavelet coefficients for Figure 5(b). Black dots are
noise coefficients, crosses are the outliers due to signal presence.
kurtosis=6.982752
8000
6000
4000
2000
0
−2000 −1000 0 1000 2000
(a) (b) (c)

Fig. 10. (a) wavelet subband LH1 of the Baboon image (black corresponds to a large coefficient
magnitude, white to small magnitudes, the contrast of the image was enhanced to better
reveal the details), (b) histogram of the wavelet coefficients in (a), (c) multivariate Gaussian
Scale Mixture distribution.
(1989); Moulin & Liu (1999); Simoncelli & Adelson (1996) proposed to use a generalized Laplace
distribution (GLD, also known as generalized Gaussian distribution) to model the kurtotic
behavior of wavelet coefficients. The GLD is defined as:
ν
ν x̃
f x̃ ( x̃ ) = exp − , (13)
2sΓ (1/ν) s
´ +∞
where Γ ( x ) = 0 t x −1 e−t dt is the Gamma function. The parameter s is scale parameter of
the distribution, which controls the variance of the distribution. The parameter ν is a shape
parameter that is related to the kurtosis of the distribution, given by:
Γ (5/ν)Γ (1/ν)
κ= − 3. (14)
Γ2 (3/ν)
The shape parameter ν is typically in the range [0.5, 1]. Because in practice, the actual value
of this parameter is unknown, the parameter value is usually estimated from the observed
data. This may be done using the maximum likelihood method or the method of moments
Srivastava et al. (2003).
268
and Wavelet
Applications
Transforms
4.2 Elliptically symmetric distributions and Gaussian Scale Mixtures

The GLD from (13) is a univariate distribution that can well model highly kurtotic histograms
of wavelet coefficients, however this distribution does not allow capture correlations between
different observations x̃. This can be achieved by using multivariate distributions, where
dependencies between neighboring wavelet coefficients can be modeled. For these densities,
a neighborhood of a fixed size (e.g. 3 × 3 in 2D) is defined around every wavelet coefficient.
Next, every neighborhood2 can be represented by a vector, e.g., by using column stacking.
In the following, we will use bold letters x̃(p), w̃ (p), ỹ (p) to denote neighborhood vectors
extracted by column stacking. Statistical studies Portilla et al. (2003); Srivastava et al. (2003)
have indicated that, next to the kurtotic behavior, the noise-free wavelet coefficients are
typically symmetric around the mode and the joint histograms have elliptical contours. This
suggests the use of elliptically symmetric distributions (ESD) to model these characteristics.
The ESD is defined by Kotz & Kozubowski (2001):
1/2

f x̃ (x̃) = k d |C x |−1/2 g (x̃ − m) C x−1 (x̃ − m) , (15)
where m is the mean of the distribution (typically m = 0), g(u ) is a real-valued function
(called density generator function), d is the length of x̃ and k d is a proportionality constant.
A multivariate extension of the GLD is obtained

by using the following density generator
function Kotz et al. (2000): g(u ) = exp − |u |ν . The resulting distribution is known as
the multivariate exponential power distribution (EPD). For our modeling task, the EPD
has a number of practical limitations: 1) the marginal densities of the distribution are not
EPD-distributed and 2) for estimation purposes, the exponential power ν often leads to
integral expressions that are analytically intractable.
Wainwright & Simoncelli (2000) noted that when the wavelet filter responses are normalized
by dividing by the square root of the local variance, the statistics of the normalized coefficients
are approximately Gaussian. The Gaussian Scale Mixture (GSM), see Figure 10(c), was then
proposed to account both for the correlations and the variability in local variance of the
wavelet coefficients. A random variable x̃ is GSM distributed if it can be written as the
product of a zero mean Gaussian random vector ũ and a scalar positive random variable
√
z Andrews & Mallows (1974):
d √
x̃ = zũ (16)
d
where = denotes equality in distribution. The scalar variable z is not observed and is therefore
√
also called ’hidden’ multiplier or mixing variable. Because of scaling ambiguity between z
and ũ, the hidden multiplier is often assumed to be normalized such that E [ z] = 1. Prior
distributions for z include Jeffrey’s non-informative3 prior Portilla et al. (2003), the log-normal
prior Portilla & Simoncelli (2001), the exponential distribution Selesnick (2006) and the Gamma
distribution Fadili & Boubchir (2005); Srivastava et al. (2002).
2 Quite often, the neighborhoods are chosen to be overlapping, despite of the fact that this destroys the
mutual independence of the different neighborhood vectors. This is done to arrive at a sufficiently large
number of neighborhood vectors (for example, for a 3 × 3 neighborhood, the number of vectors will be
multiplied by 9), which will generally result in more reliable estimates.
3 Note that in this case, the mathematical expectation E [z] does not exist.
ColoredEstimation
15
The GSM also belongs to the family of ESDs. The density generator function is given by:
ˆ +∞
1
f z (z)z− 2 exp − x2 dz.
d
g( x ) = (17)
0 2z
For some hidden multiplier densities f z (z) a closed-form expression can be found for
g( x ), although most often, numerical integration is performed over a closed interval. In
Gómez et al. (2008) it has been shown that the EPD is also a GSM distribution, for some
values of the shape parameter ν ∈]0, 1]. However, the distribution f z (z) depends on d and
has a complicated analytical expression (see Gómez et al. (2008)).
4.3 Other prior distributions

In literature, several other prior distributions for noise-free wavelet coefficients have
been proposed. For example, the Student-T distribution Tzikas et al. (2007), Alpha-stable
distributions Achim et al. (2001); Nikias & Shao (1995) and the Cauchy distribution
Rabbani et al. (2006). All these heavy tailed distributions have a GSM representation,
hence studying general GSMs automatically covers all of these distributions. Next, a
complex extension of the Gaussian Scale Mixture density has been proposed for modeling
complex-valued wavelet coefficients in Vo et al. (2007). This complex GSM distribution is a
special case of the GSM distribution, with a special condition imposed to the covariance matrix
of the distribution. Next to GSMs, mixtures of a Gaussian distribution and a point mass at zero
were used in Abramovich et al. (1998); Clyde et al. (1998), mixtures of two Gaussian distributions
in Crouse et al. (1998); Fan & Xia (2001); Romberg et al. (2001) and mixtures of truncated or
quasi-Laplace distributions in Pižurica & Philips (2006); Shi & Selesnick (2006).
5. Noise covariance estimation techniques

In this Section, we will use the GSM prior distribution from Section 4 to design a noise
covariance estimation technique. We therefore start from the signal-plus-noise model from
equation (1). The assumed addititivity of the signal and noise leads to an equivalent
expression in the wavelet domain:
ỹ (p) = x̃(p) + w̃ (p), (18)
where w̃ (p) is spatially stationary Gaussian distributed vector of length d with mean 0
and covariance Cw̃ . Due to the assumed noise stationarity, the covariance matrix Cw̃ has
dimensions d × d and is directly related to the noise autocorrelation function Rw̃ (p): the
covariance between two coefficients at positions p and q only depends on the difference in
location between both positions:
(Cw̃ )p,q = Rw̃ (q − p) (19)
where vector-valued indices in (Cw̃ )p,q are used as a short notation for their respective
column-stacked ordening. By (19), the estimation of the noise autocorrelation function is
equivalent to the estimation of the covariance Cw̃ . Next, the noise-free coefficients are GSM
distributed with covariance matrix C x̃ . For the GSM model, we have x̃| z ∼ N (0, zCũ ) .
Consequently, the density of ỹ is a specific case of a Gaussian mixture model:
270
and Wavelet
Applications
Transforms
ỹ | z ∼ N (0, zCũ + Cw̃ ) (20)
where the signal covariance is also unknown. We remark that this matrix can be eliminated
relying on Cũ + Cw̃ = Cỹ (this directly follows from (1), when E [z] = 1):

ỹ | z ∼ N 0, zCỹ + (1 − z)Cw̃ . (21)
The signal-plus-noise covariance matrix can be estimated using the method of maximum
likelihood: Cỹ = 1 ∑p ỹ (p)ỹ T (p), with N the number of coefficients in the considered
N
wavelet subband.
5.1 Generalized Expectation-Maximization algorithm

In Portilla (2004), a Generalized Expectation-Maximization (GEM) algorithm is given to
estimate the noise covariance matrix. Based on an initial estimate of the noise covariance
( 0)
(typically chosen as Cw̃ = cCỹ , with 0 < c < 1 a constant), the noise covariance matrix is
iteratively updated according to the following rule:

< | ỹ ( p ) Θ ( i ) ỹ (p)ỹ T (p)
( i +1)
∑ p P z z 0 ,
Cw̃ = , (22)
∑p P z < z0 |ỹ (p), Θ ( i)
where i is the iteration index and Θ ( i) denotes the GSM model parameters at iteration i and
where z0 is a small positive constant. Equation (22) can be motivated by the observation that
for z sufficiently small, Cỹ|z = Cw̃ . The posterior probability that z < z0 , conditioned on an

observation vector ỹ (p), i.e., P z < z0 |ỹ (p), Θ ( i) is then used as a weight in the averaging

process. We can understand this as follows: P z < z0 |ỹ (p), Θ ( i) represents the probability
that a given observation vector contains a negligible signal component. The estimated noise
covariance is then the average over all sample covariances y (˜p)ỹ T (p), weighted by the
probability that the considered sample contains a negligible signal component.
Because the updating rule (22) is not guaranteed to increase the likelihood of the data, at
every iteration it is checked if this new covariance estimate results in a higher likelihood:
Q(Θ ( i) , Θ ( i+1) ) > Q(Θ ( i) , Θ ( i) ), with Q(Θ ( i) , Θ ) the expected log-likelihood function of the
data:

Q(Θ ( i) , Θ ) = E log f z|ỹ (z|ỹ, Θ ) |ỹ, Θ ( i)
ˆ +∞
= ∑ f z|ỹ z|ỹ (p), Θ ( i) log f z|ỹ (z|ỹ (p), Θ ) dz. (23)
p 0
In case the expected log-likelihood (23) decreases, it is proposed in Portilla (2004) to perform
a gradient ascent step:
ColoredEstimation
17

( i +1) ( i) ∂Q(Θ ( i) , Θ )
Cw̃ = Cw̃ +λ
∂Cw̃ ( i)
Cw̃ =Cw̃
ˆ +∞
( i) λ
= Cw̃ + N f z (z) (1 − z)Cz−1 (I − C
z Cz−1 )dz, (24)
2 0
where
Cz = zCỹ + (1 − z)Cw̃ , (25)

∑p f z|ỹ z|ỹ (p), Θ ( i) ỹ (p)ỹ T (p)
z =
C . (26)
∑p f z z|ỹ (p), Θ ( i)
Although a good fitting to the data was reported in Portilla (2004), the technique requires the
relatively costly evaluation of the expected log-likelihood function (23). Another issue is the
choice of the constant z0 . In Portilla (2004), this was solved by using a discrete GSM mixture
for the hiddenmultiplier density f z (z). By assigning a non-zero probability mass at z = 0, the
probability P z = 0|ỹ (p), Θ ( i) is guaranteed to be non-zero.
5.2 Constrained EM algorithm using augmented Lagrangian optimization

In this subsection, we present a novel, alternative estimation method that does not need
evaluation of the expected log-likelihood function. First, we assume a discrete hidden
multiplier density P (z = zk ) = αk , with k = 1, ..., K. The parameters can be initialized in
a manner similar to Portilla et al. (2003):4
zk = exp (−3 + 7(k − 1)/(K − 1)) , k = 1, ..., K

αk = 1/K. (27)
In contrast to the GEM algorithm, where Cw̃ is optimized directly, we take a slightly
different approach. We rely on the fact that the density f ỹ (ỹ ) corresponds to a Gaussian
mixture model. This allows us to use the EM algorithm for Gaussian mixtures, with some
modifications that we will describe next. Let us denote by Ck the covariance matrices of
the mixture components. Because of (20), the mixture covariance matrices should be subject
to the constraint zk Cũ + Cw̃ = Ck . Our method now consists of optimizing the expected
log-likelihood function (as in a regular EM algorithm Dempster et al. (1977)), but now subject
to the GSM constraint:

max E log f z|ỹ (z|ỹ, Θ ) |ỹ, Θ ( i) s.t. zk Cũ + Cw̃ = Ck (28)
Θ
To solve this constrained problem, we use the augmented Lagrangian (AL) method. In the AL
method, a constrained problem is translated to an unconstrained problem with a Lagrange
4
Here, values zmin and zmax from (Portilla et al., 2003, p. 1343) are slighly modified to have a good
sampling of the continuous pdf f z ( z) with a small number of components K (for example, K = 6).
272
and Wavelet
Applications
Transforms
multiplier and an extra penalty term. In our case, the unconstrained problem is given by:
K K
max E log f z|ỹ (z|ỹ, Θ ) |ỹ, Θ ( i) − 2 ∑ Vec [ak ] T Vec [Ck -zk Cx̃ -Cw̃ ] - ∑ λk Ck -zk Cx̃ -Cw̃ 2F
Θ k =1 k =1
(29)
where ak , k = 1, ..., K are d × d matrices of Lagrange multipliers, λk are penalty factors, Vec [·]
converts a matrix to a column vector (e.g., using column stacking) and · F is the matrix
Frobenius norm. Taking the derivatives of (29) with respect to C x and Cw and setting to zero
leads to a linear system of equations, in block matrix form:
⎛ ( i +1) ⎞ ⎛ K ⎞
( i)
μ2 I μ1 I C x̃ ∑ k =1 z k λ k C k + a k
⎝ ⎠=⎝ ⎠ (30)
( i +1) ( i)
μ1 I I Cw̃ ∑K k =1 λ k C k + a k
with μ1 = ∑K k =1 λ k z k and μ2 = ∑k =1 λ k z k . Similarly, maximizing (29) with respect to Ck leads

K 2
to the following update equation:

( i) ( i)
( i +1)
∑p P z = zk |ỹ (p), Θ ( i) ỹ (p)ỹ T (p) − 2λk zk C x̃ + Cw̃ − ak
Ck = , k = 1, ..., K (31)
∑p P z = zk |ỹ (p), Θ ( i) − 2λk
Additionally, the Lagrange multipliers are updated in every iteration:
( i +1) ( i) λ k ( i +1) ( i +1) ( i +1)

ak = ak + Ck − zk Cx̃ − Cw̃ . (32)
2
This process
is repeatediteratively until a given convergence criterion has been reached (for
( i +1) ( i)
example Cw̃ − Cw̃ < , with a small positive number). The penalty weights λk are
F
chosen in order to speed up the convergence of the algorithm. In our method, we choose λk
inversely proportional to zk : λk = z1 /zk , with z1 < z2 < · · · < zK . The complete algorithm is
summarized in Algorithm 2.
Important to mention is that the above algorithm may fail, if the matrix in the update formula
(30) is singular, i.e. if μ21 = μ2 . It is worthful to note that the kurtosis of the wavelet subband
coefficients is given by 3μ2 /μ21 − 3 and becomes zero if μ21 = μ2 . In this case, the probability
density function f ỹ (ỹ ) is Gaussian, and every component of the GSM model will have the
same hidden multiplier value zk = μ1 , such that also f x̃ (x̃) is Gaussian. Consequently, it
becomes impossible to separate the signal from the noise: the highly kurtotic behavior of the
noise-free coefficients x can not be exploited. By our specific initialization (27), we actually
avoided the latter problem.
The elegance of this algorithm lies in the fact that simple update formulas are being used
and that the complete algorithm is guaranteed to converge (albeit to a local maximum of the
objective function, as with nearly all EM type of algorithms).
6. Estimation of a parametric noise PSD

In the previous Section, two methods were presented to estimate the noise covariance matrix
in the wavelet domain. Although these covariance matrices can be directly used in, e.g., blind
ColoredEstimation
19
Algorithm 2 Constrained EM algorithm for estimating the noise covariance matrix Cw̃ .
( 0) ( 0) ( 0) ( 0) ( 0) ( 0)
Cỹ = 1
N ∑ỹ (p)ỹ T (p), Cw̃ = 9
10 Cỹ , C x̃ = 0.1Cỹ , Ck = zk Cx̃ + Cw̃ , αk = 1
K, λk = z1
zk .
p
repeat
( i +1)
α̂k = 1
N ∑p P ( z = z k | y (p), Θ ) , for k = 1, ..., K

∑P ( z = z k |ỹ(p),Θ ( i) )ỹ (p)ỹ T ( i) ( i) ( i)
(p)−2λ k z k Cx̃ +Cw̃ −ak
( i +1) p
Ck = , for k = 1, ..., K
∑P (z=z |ỹ(p),Θ( ))−2λ
k
i
k
p
⎛ ⎞
( i +1) ( i +1) ( i)
C x̃ I − μ1 I ∑K = z k λ k Ck + ak
= 1 ⎝ k 1 ⎠
( i +1) μ2 − μ21 − μ1 I μ2 I ( i +1) ( i)
Cw̃ ∑Kk =1 λ k C k + ak

( i +1) ( i) λk ( i +1) ( i +1) ( i +1)
ak = ak + 2 Ck − zk Cx̃ − Cw̃ for k = 1, ..., K
i ← i+1

( i +1) ( i)
until convergence (Cw̃ − Cw̃ < ).
F
Input signal Wavelet transform
Wavelet subbands
Estimate of the
Parameter noise autocorr.
estimation function
Parametric autocorr. function
for each wavelet subband
Wavelet-based
Parametric noise
autocorrelation
autocorrelation function
decomposition
Fig. 11. Overview of the proposed algorithm for the estimation of a parametric noise PSD.
denoising approaches (see Portilla (2004)), the covariance matrices are not directly related to
the noise PSD (in the sense that, after estimation of the covariances matrices the noise PSD is
still unknown). We here present a novel approach to estimate the parameters of a parametric
noise PSD based on the covariance matrix estimation methods. As far as the authors are
aware of, such a technique does not yet exist. This approach also combines all the different
techniques discussed in this Chapter. An overview of our algorithm is given in Figure 11.
First, the noise is assumed to have a PSD with an unkown set of parameters β. Consequently,
by the Wiener-Khinchin theorem, the noise autocorrelation function Rw,β (p) is known. The
wavelet-domain noise autocorrelation functions can be computed from Rw,β (p), as explained
in Section 3. Using the formula (19), the parametric wavelet domain noise covariance matrix
274
and Wavelet
Applications
Transforms

Cw̃ (β ) can be found. Defining Rw (β ) = Rw,β (p) , the noise covariance matrix can be
expressed in terms of Rw (β ) by using a matrix multiplication:
Cw̃ (β ) = QRw (β ) (33)
Then, the parameter β can be estimated iteratively in every iteration of Algorithm 2.

Therefore, we minimize the squared matrix Frobenius norm:
2
( i +1)
β ( i+1) = arg min Cw̃ − QRw (β ) . (34)
β F
Because Rw (β ) is not a linear function in general, this is a non-linear optimization problem,

which can be solved using gradient descent or Gauss-Newton techniques. The gradient
descent step is given by:
T ∂R
( i +1)
Q
w
β ( i+1) = β ( i) + γ Cw̃ − QRw β ( i) . (35)
∂β β =β (i)
Note that in practice this equation may be iterated several times until convergence in
an inner iteration, before the other model parameters are updated. As an example,
consider the autocorrelation function from Figure 8, corresponding to the PSD P (ω ) =
β sin ( β | ω |) I [ β | ω | < π ], with I [·] the indicator function. Application
of the inverse DTFT
πn

gives the spatial autocorrelation function Rw,β (n ) = β 1 + cos β 2 / π (n2 − β2 ) . Its

derivative with respect to β is given by:

∂Rw,β (n ) n πn πn
= sin π n −β
2 2
+ 2βn 1 + cos . (36)
∂β π ( n 2 − β2 ) 2 β β
Substitution of (36) into (35) then gives the desired update step.
An interesting special case is the estimation of white Gaussian noise, with autocorrelation
function Rw,β (n ) = sδ(n ), with s the unknown noise variance. In this case, (34) comprises a
least-squares problem, with a linear solution.
In this Section, we will compare the performances of the noise estimation methods from
Section 5. For this task, both iterative algorithms (the GEM algorithm and the constrained EM
algorithm), are initialized using the same set of parameters. The initial values used are given
in Algorithm 2 and in (27). The number of mixture components used is 6: K = 6. Five images
(Barbara, Baboon, Lena, Boats and Peppers) are transformed to the wavelet domain, using the
Daubechies wavelet with two vanishing moments. Artificial Gaussian noise with a known
(ground-truth) autocorrelation function is added to each LH1 -subband, which allows us to
compute the estimation
error
afterwards.
Thisground-truth noise autocorrelation function is
πy
given by: σ2 β4 1 + cos πx β 1 + cos β / π ( x 2 − β2 )( y2 − β2 ) , with β = 3/2 and
with σ ∈ {1, 5, 10, 15, 25, 50}. Then, after every iteration of both algorithms, the log-likelihood
2

function log f ỹ |Θ (ỹ | Θ ) and the quadratic error C w̃ − Cw̃ are computed, which allows us
F
to compare the performances of both algorithms as function of the iteration number i. Both
ColoredEstimation
21
6 average log−likelihood function average error

x 10
−1.74 35
Constrained EM algorithm
−1.76 30 GEM algorithm
10 log10(||Cw−Cw,true||2)
−1.78
25
log fy(y)
−1.8
20
−1.82
15
−1.84
−1.86 Constrained EM algorithm 10

GEM algorithm
−1.88 5
0 10 20 30 40 0 10 20 30 40
Iteration i Iteration i
Fig. 12. Comparison of the performance of the GEM algorithm Portilla (2004) and the
constrained EM algorithm Subsection 5.2, as a function the iteration number i. Results are
averaged over 5 images and 6 noise levels. (left) average log-likelihood log f ỹ|Θ (ỹ | Θ ), (right)
2

average estimation error in logarithmic scale 10 log10 C w̃ − Cw̃ .
F
Table 1. Comparison of the performance of the GEM algorithm Portilla (2004) and the
constrained EM algorithm (CEM) from Subsection 5.2, for 5 images and 6 noise levels. Shown
2

is the estimation error in logarithmic scale 10 log10 Cw̃ − Cw̃ after 40 iterations.
F
σ=1 σ=5 σ = 10 σ = 15 σ = 25 σ = 50
Image CEM GEM CEM GEM CEM GEM CEM GEM CEM GEM CEM GEM
Barbara 14.25 14.56 -12.98 -11.78 -26.63 -23.21 -29.29 -28.08 -36.50 -31.83 -37.93 -35.50
Baboon 24.02 28.74 -3.79 2.13 -14.84 -7.23 -21.88 -12.82 -25.60 -18.60 -30.75 -28.38
Lena 9.56 14.42 -16.77 -12.23 -23.25 -23.34 -29.54 -29.68 -37.99 -37.17 -38.38 -38.93
Boats 7.72 9.66 -17.77 -16.55 -30.35 -26.31 -30.09 -28.59 -35.83 -32.65 -37.90 -37.06
Peppers 11.28 17.06 -14.34 -9.92 -24.71 -21.25 -30.10 -27.86 -31.84 -34.05 -40.29 -36.51
Average 13.37 16.89 -13.13 -9.67 -23.96 -20.27 -28.18 -25.41 -33.55 -30.86 -37.05 -35.28
algorithms maximize the log-likelihood function, note however that this does not necessarily
results in minimizing the quadratic error. The results are shown in Figure 12 and Table 1. It can
be seen that while the GEM algorithm converges to its final value, on average the constrained
EM algorithm is able to reach a solution with a higher log-likelihood function and a lower
error. We remark that the objective function is non-convex, such that both algorithms can get
trapped in local maxima. Although both algorithms use the same initialization, in most of
the experiments (see Table 1) the constrained EM gives a more accurate estimate of the noise
covariance matrix.
In Figure 13 and Figure 14, we used the noise estimation method based on the constrained
EM algorithm in combination with the BLS-GSM Portilla et al. (2003) denoising method, in
order to perform blind noise removal. An undecimated wavelet transform of 3 levels with the
Daubechies wavelet with eight vanishing moments was used. The PSD of the Gaussian noise
is in the captions of Figure 13 and Figure 14. Clearly, the combined method is well able to
distinguish signal information from noise information, leading to a succesful removal of the
noise while preserving signal structures.
276
and Wavelet
Applications
Transforms
(a) Original image (b) With artificial noise (PSNR=20.17dB) (c) Denoised (PSNR=41.21dB)
Fig. 13. Blind denoising results (using the BLS-GSM denoising method and the proposed
constrained EM noise estimation technique). Noise PSD
P (ω ) ∼ exp(−4000((ω x /π − 0.1)2 + (ω y /π − 0.12)2 )).
(a) Original image (b) With artificial noise (PSNR=17.25dB) (c) Denoised (PSNR=29.00dB)
Fig. 14. Blind denoising results (using the BLS-GSM denoising method and the proposed
constrained EM noise estimation technique). Noise PSD P (ω ) ∼ exp(−2000((ω x /π −
0.1)2 + (ω y /π − 0.12)2 ))+ exp(−3000((ω x /π + 0.15)2 + (ω y /π − 0.22)2 )) + 10−3 .
8. Conclusion
In this chapter, we investigated the estimation of stationary colored noise, which is most
efficiently described in a Fourier basis using the power spectral density (PSD). Because of
the time or spatial locality of signal structures, estimation of colored noise is best performed
in a transform domain that allows to adapt to the signal locality. We have shown that
wavelets are very good candidates for this task: their vanishing moment properties allow
us to complete suppress smoothly varying signals, such that efficient noise estimation can
directly be performed on a single wavelet subband. However, in practice, signals are not
smoothly varying and may contain transitions (such as edges and textures in images). To take
ColoredEstimation
23
this into account, we have presented several prior models for noise-free wavelet coefficients.
These prior models are then used in an expectation-maximization algorithm, which gives us
an estimate of the noise covariance matrix for a given wavelet subband. We have further
shown how this covariance matrix is related to the noise autocorrelation function in spatial or
time domain. This relationship can then be used, e.g., to estimate parameters of parametric
PSDs, yielding reliable and accurate estimates for noise PSDs. Because noise is present
in most real-life signals and images, many signal and image processing methods can be
further improved by taking advantage of estimated noise characteristics using techniques as
described in this chapter.
9. References
Abramovich, F., Sapatinas, T. & Silverman, B. (1998). Wavelet thresholding via a Bayesian
approach, J. of the Royal Statist. Society B 60: 725–749.
Achim, A., Bezerianos, A. & Tsakalides, P. (2001). Wavelet-based ultrasound image denoising
using an alpha-stable prior probability model, Proc. International Conference on Image
Processing, Vol. 2, pp. 221–224.
Aelterman, J., Deblaere, K., Goossens, B., Pižurica, A. & Philips, W. (2010). Dual Tree Complex
Wavelet-Based Denoising of correlated noise in 3D Magnetic Resonance Imaging.
Under revision.
Aelterman, J., Goossens, B., Pižurica, A. & Philips, W. (2010). Recent Advances in Signal
Processing, IN-TECH, chapter Suppression of Correlated Noise.
Andrews, D. & Mallows, C. (1974). Scale mixtures of normal distributions, J. Royal Stat. Stoc.
36: 99–102.
Antonini, M., Barlaud, M., Mathieu, P. & Daubechies, I. (1992). Image coding using wavelet
transform., IEEE Trans. Image Process. 1(2): 205–220.
Baher, H. (2001). Analog and Digital Signal Processing, Wiley, Chichester.
Bayer, B. (1976). Color imaging array, United States Patent 3971065.
Borel, C., Cooke, B. & Laubscher, B. (1996). Partial Removal of Correlated noise in Thermal
Imagery, Proceedings of SPIE, Vol. 2759, pp. 131–138.
Campbell, N. A., Lopuhaä, H. P. & Rousseeuw, P. J. (1998). On the calculation of a robust
s-estimator of a covariance matrix., Stat Med 17(23): 2685–2695.
Candès, E. (1998). Ridgelets: Theory and Applications, PhD thesis, Departement of Statistics,
Stanford University.
Candès, E., Demanet, L., Donoho, D. & Ying, L. (2006). Fast Discrete Curvelet Transforms,
Multiscale modeling and simulation 5(3): 861–899.
Chang, S. G., Yu, B. & Vetterli, M. (1998). Spatially adaptive wavelet thresholding with context
modeling for image denoising, Proc. IEEE Internat. Conf. on Image Proc., Chicago, IL,
USA.
Clyde, M., Parmigiani, G. & Vidakovic, B. (1998). Multiple shrinkage and subset selection in
wavelets, Biometrika 85(2): 391–401.
Crouse, M. S., Nowak, R. D. & Baranuik, R. G. (1998). Wavelet-based statistical signal
processing using hidden Markov models, IEEE Trans. Signal Proc. 46(4): 886–902.
Daubechies, I. (1992). Ten Lectures on Wavelets, Society for Industrial and Applied Mathematics,
Philadelphia.
278
and Wavelet
Applications
Transforms
Dempster, A., Laird, N. & Rubin, D. (1977). Maximum likelihood from incomplete data via
the EM algorithm, Journal of the Royal Statistical Society, Series B 19(1): 1–38.
Do, M. N. & Vetterli, M. (2003). The finite ridgelet transform for image representation, IEEE
Trans. Image Processing 12(1): 16–28.
Do, M. N. & Vetterli, M. (2005). The contourlet transform: An efficient directional
multiresolution image representation, IEEE Trans. Image Process. 14(12): 2091–2106.
Donoho, D. L. & Johnstone, I. M. (1995). Adapting to unknown smoothness via wavelet
shrinking, Journal of the American Statistical Association 90(432): 1200–1224.
Fadili, J. M. & Boubchir, L. (2005). Analytical form for a Bayesian wavelet estimator of images
using the Bessel K form densities, IEEE Trans. on Image Process. 14(2): 231–240.
Fan, G. & Xia, X. (2001). Image denoising using local contextual hidden Markov model in the
wavelet domain, IEEE Signal Processing Letters 8(5): 125–128.
Field, D. J. (1987). Relations between the statistics of natural images and the response
properties of cortical cells, J. Opt. Soc. Am. A 4(12): 2379–2394.
Fischer, S., Šroubek, F., Perrinet, L., Redondo, R. & Cristobal, G. (2007). Self-Invertible 2D
Log-Gabor Wavelets, International Journal of Computer Vision 75(2): 231–246.
Gómez, E., Gómez-Villegas, M. A. & Marín, J. M. (2008). Multivariate exponential
power distributions as mixtures of normal distributions with Bayesian applications,
Communications in Statistics - Theory and Methods 37(6): 972–985.
Goossens, B., Aelterman, J., Pižurica, A. & Philips, W. (2010). A Recursive Scheme for
Computing Autocovariance functions of complex wavelet subbands, IEEE Trans.
Signal Processing 58(7): 3907–3912.
Guo, K. & Labate, D. (2007). Optimally Sparse Multidimensional Representation using
Shearlets, SIAM J Math. Anal. 39: 298–318.
Johnstone, I. M. & Silverman, B. W. (1997). Wavelet threshold estimators for data with
correlated noise, Journal of the Royal Statistical Society B 59(2): 319–351.
Kingsbury, N. G. (2001). Complex wavelets for shift invariant analysis and filtering of signals,
Journal of Applied and Computational Harmonic Analysis 10(3): 234–253.
Kotz, S., Kozubowski, T. J. & Podgorski, K. (2000). An asymmetric multivariate laplace
distribution, Computational Statistics 4: 531–540.
Kotz, S. & Kozubowski, T.and Podgorski, K. (2001). The Laplace Distributions And
Generalizations: A Revisit with Applications to Communications, Economics, Engineering,
Finance, Birkhäuser, Boston.
Kwon, O., Sohn, K. & Lee, C. (2003). Deinterlacing using Directional Interpolation and Motion
Compensation, IEEE Trans. Consumer Electronics 49(1): 198–203.
Lee, T. (1996). Image Representation Using 2D Gabor Wavelets, IEEE Trans. Pattern Analysis
and Machine Intelligence 18(10): 1.
Mallat, S. (1989). Multifrequency channel decomposition of images and wavelet models, IEEE
Trans. Acoust., Speech, Signal Proc. 37(12): 2091–2110.
Mallat, S. (1999). A Wavelet Tour of Signal Processing, Academic Press.
Moulin, P. & Liu, J. (1999). Analysis of multiresolution image denoising schemes using
generalized-gaussian and complexity priors, IEEE Trans. Info. Theory, Special Issue on
Multiscale Analysis 3(3): 909–919.
Nikias, C. L. & Shao, M. (1995). Signal Processing with Alpha-Stable Distributions and
Applications, Wiley-Interscience.
ColoredEstimation
25
Nowak, R. (1999). Wavelet-based rician noise removal for magnetic resonance imaging., IEEE
Trans Image Process 8(10): 1408–1419.
Pena, D. & Prieto, F. (2001). Multivariate outlier detection and robust covariance matrix
estimation, Technometrics 43(3): 286–310.
Pižurica, A. & Philips, W. (2006). Estimating the probability of the presence of a signal of
interest in multiresolution single- and multiband image denoising., IEEE Trans. Image
Process. 15(3): 654–665.
Pižurica, A., Philips, W., Lemahieu, I. & Acheroy, M. (2003). A versatile wavelet domain noise
filtration technique for medical imaging, IEEE Trans. Medical Imaging 22(3): 323–331.
Portilla, J. (2004). Full blind denoising through noise covariance estimation using Gaussian
Scale Mixtures in the wavelet domain, IEEE Int. Conf. on Image Process. (ICIP)
2: 1217–1220.
Portilla, J. & Simoncelli, E. (2001). Adaptive Wiener Denoising using a Gaussian Scale Mixture
Model in the Wavelet Domain, IEEE Int. Conf. on Image Process. (ICIP) 2: 37–40.
Portilla, J., Strela, V., Wainwright, M. & Simoncelli, E. (2003). Image denoising using scale
mixtures of gaussians in the wavelet domain, IEEE Transactions on image processing
12(11): 1338–1351.
Rabbani, H., Vafadust, M., Gazor, S. & Selesnick, I. W. (2006). Image Denoising Employing
a Bivariate Cauchy Distribution with Local Variance in Complex Wavelet Domain,
12th Digital Signal Processing Workshop - 4th Signal Processing Education Workshop,
pp. 203–208.
Romberg, J., Choi, H. & Baraniuk, R. G. (2001). Bayesian tree structured image modeling using
wavelet-domain Hidden Markov Models, IEEE Trans. Image Process. 10(7): 1056–1068.
Enter text here.
Selesnick, I. W. (2006). Laplace random vectors, Gaussian noise, and the generalized
incomplete Gamma function, Proc. IEEE Int. Conf. on Image Process., pp. 2097–2100.
Selesnick, I. W., Baraniuk, R. G. & Kingsbury, N. G. (2005a). The Dual-Tree Complex Wavelet
Transform, IEEE Signal Processing Magazine 22(6): 123–151.
Selesnick, I. W., Baraniuk, R. G. & Kingsbury, N. G. (2005b). The Dual-Tree Complex Wavelet
Transform, IEEE Signal Processing Magazine 22(6): 123–151.
Shi, F. & Selesnick, I. W. (2006). Multivariate Quasi-Laplacian Mixture Models for
Wavelet-based Image Denoising, Proc. Int. Conf. on Image Processing (ICIP),
pp. 2097–2100.
Simoncelli, E., Freeman, W. T., Adelson, E. H. & Heeger, D. J. (1992). Shiftable Multi-scale
Transforms, IEEE Trans. Information Theory 38(2): 587–607.
Simoncelli, E. P. & Adelson, E. H. (1996). Noise removal via Bayesian wavelet coring, Proc.
IEEE Internat. Conf. Image Proc. ICIP, Lausanne, Switzerland.
Srivastava, A., Lee, A. B., Simoncelli, E. & Zhu, S.-C. (2003). On Advances in Statistical
Modeling of Natural Images, Journal of Mathematical Imaging and Vision 18: 17–33.
Srivastava, A., Liu, X. & Grenander, U. (2002). Universal Analytical Forms for Modeling Image
Probabilities, IEEE Trans. Pattern Analysis and Machine Intelligence 24(9): 1200–1214.
Tzikas, D., Likas, A. & Galatsanos, N. (2007). Variational bayesian blind image deconvolution
with student-t priors, Proc. IEEE International Conference on Image Processing ICIP 2007,
Vol. 1, pp. I–109–I–112.
Van De Ville, D. & Unser, M. (2008). Complex Wavelet Bases, Steerability, and the Marr-like
pyramid, IEEE Trans. Image Processing 17(11): 2063–2080.
280
and Wavelet
Applications
Transforms
Vo, A., Nguyen, T. & Oraintara, S. (2007). Image denoising using shiftable directional pyramid
and scale mixtures of complex gaussians, Proc. IEEE International Symposium on
Circuits and Systems ISCAS 2007, pp. 4000–4003.
Wainwright, M. J. & Simoncelli, E. P. (2000). Scale mixtures of Gaussians and the statistics of
natural images, Adv. Neural Information Processing Systems (NIPS 1999) 12: 855–861.
16
An Adaptive Energy Discretization of

the Neutron Transport Equation
Based on a Wavelet Galerkin Method
D. Fournier and R. Le Tellier
CEA, DEN, DER/SPRC/LEPh, Cadarache, F-13108 Saint-Paul-lez-Durance
France
1. Introduction
The time-independent neutron transport equation derived from the Boltzmann equation with
a linear collision kernel models the neutron population in the six dimensional space defined
by r ∈ D the space variable, Ω ∈ S2 the direction of motion variable and E ∈ B =] EG +1, E1 [
the energy variable.
It represents
the balance between the neutrons entering the hypervolume
3 2
d rd ΩdE about r, Ω, E by fission or scattering and those leaving by streaming or any kind
of interactions. The unknown is the so-called neutron flux φ(r, Ω, E ) = v( E )n (r,
Ω, E ) with

n (r, Ω, E ) the neutron density and v( E ) the neutron velocity. The problem is defined in terms
of the neutron interaction properties of the different materials i.e. the cross sections.
The solution of this equation in a deterministic way proceeds by the successive discretization
of the three variables: energy, angle and space. The treatment of the energy variable invariably
consists in a multigroup discretization which considers the cross sections and the flux to be
constant within a group (i.e. a cell of the 1D energy mesh). A pre-homogenization of the cross
sections is performed at the library processing level using a spatially independent weighting
flux (e.g. 1/E spectrum in the epithermal range).
With a broad group structure (≈ 100 to 2000 energy groups), this prior homogenization is
unsufficient to take into account the case-specific, spatially-dependent, self-shielding effect i.e.
the flux local depression in the vicinity of resonances that largely affects the neutron balance.
As a consequence, a neutron transport calculation has to incorporate a so-called self-shielding
model to correct the group cross sections of resonant isotopes. This homogenization stage
of a neutron transport calculation is known to be a main source of errors for deterministic
methods; as a consequence, an important work has been carried out to improve it. An
optimized energy mesh structure (Mosca et al., 2011) in addition to an advanced self-shielding
model (Hébert, 2007) is incorporated in state-of-the-art transport codes.
A different treatment for the energy variable based on a finite element approach is the basis
of the present work. Such an avenue was proposed in the past by (Allen, 1986) but seldom
used in practice. Indeed, finite element methods are commonly based on polynomial function
bases which are not appropriate for non-smooth behavior.
Recently, two independent works by (Le Tellier et al., 2009) and (Yang et al., 2010) have
proposed wavelet-Galerkin methods to overcome this issue. In this chapter, after a review
282
of these two approaches, we will focus on the development in this framework of adaptive
algorithms with a control (at least, partial) of the discretization error. Such algorithms have
been partially presented in a previous conference presentation by (Fournier & Le Tellier, 2009)
but this book chapter gives a more in-depth presentation and updated numerical results for
algorithms that may be of interest for other applications of wavelet-based finite elements.
Such algorithms are analyzed in a limited framework (the fine structure flux equation for
a single isotope diluted in a mixture of non-resonant isotopes in an infinite homogeneous
medium) but the relevant issues regarding their extension in the general case are discussed.
2. Wavelet-Galerkin based energy discretization of the neutron transport equation

2.1 Generalized weak multigroup neutron transport equation
As in (Allen, 1986), the transport equation is discretized starting from the Sobolev spaces
W21 (D × S2 ) = {φ ∈ L2 (D × S2 ), all the weak derivatives of φ ∈ L2 (D × S2 )} , (1)

W21† (A) = φ ∈ L2 (A), φ ∈ W21 (D × S2 ) . (2)
with A = D × S2 × B . In particular, the energy variable is discretized as follows. An energy

mesh consisting of G groups such that E1 > E2 · · · > EG +1 (Ig =] Eg+1, Eg [) is selected and the
finite-dimension space W h (A) ⊂ W21† (A) considered is

G N
2
E ) = ∑ Π I ( E ) f gT ( E )φg (r, Ω) with φ g ∈ W 1 (D × S2 ) g ,
W h (A) = φ ∈ W 1† (A), φ(r, Ω, 2
g
g =1
(3)
g Ng
where Π Ig is the characteristic function of group g, f ∈ L2 ( Ig ) is an orthonormal set
of wavelet
functions on Ig and the group flux unknowns are the flux wavelet modes i.e.
φ g (r,
Ω) = E ).
dE f g ( E )φ(r, Ω,
Ig
Within this framework, a Ritz-Galerkin procedure casts the transport equation (written with
isotropic scattering and an external source Q(r,
Ω, E )) in a generalized (weak) multigroup
form: ∀ g ∈ [1, G ],
1 G

Ω · ∇
+ Σt g (r ) φ g (r,
Ω) = ∑ Σs g← g (r ) Φ g (r ) + Q g (r,
Ω ), (4)
4π
g =1

where Φ g (r ) = d2 Ωφ g (r,

Ω) and the source vector is Qg (r,
Ω) = dE f g ( E ) Q(r,
Ω, E ).
S2 Ig
The matrices coupling the flux modes within a group are defined in terms of the total Σt (r, E )
and scattering transfer Σs (r, E → E ) cross sections as

Σt g (r ) = dE f g ( E )Σt (r, E ) f gT ( E ), (5)

Ig

Σs g (r ) = dE f g ( E ) dE Σs (r, E → E ) f gT ( E ). (6)

Ig I g
Note that Eq. 5 introduces a coupling between the different modes within a group on the left
hand side of the transport equation i.e. a coupling of the angular flux projections φ g (r,
Ω) that
is not present for the standard multigroup approach.
An Adaptive Energy Discretization
of the Neutron
An Adaptive Transport
Energy Discretization Equation
of the Neutron TransportBased on a
Equation based on Wavelet Galerkin
a Wavelet Galerkin Method Method 2833
Two different approaches by (Le Tellier et al., 2009) and (Yang et al., 2010) based on compactly
supported Daubechies wavelets (Daubechies, 1992) have been proposed so far to deal with
this coupling:
1. in (Yang et al., 2010), a dilation order is fixed and the basis consists in the translates of the
associated scaling function; in this case, Σt g (r ) is a band matrix and the mode coupling is
limited in such a way that a Richardson iterative scheme can be employed to resolve this
coupling.
2. in (Le Tellier et al., 2009), both dilates and translates of the mother wavelet functions are
retained in the basis according to a thresholding procedure applied to the discrete wavelet
transform of either the total cross section Σt or an approximate flux. In this case, the basis
selection can be optimized but the modes are tightly coupled; a procedure based on a
change of basis through a matrix diagonalization have been proposed to explictly decouple
the equations.
This second approach proceeds as follows. Let us consider
that the nuclear data are known
g
by their projections on a set of orthonormal functions gk in each group e.g.
k ∈[1,Ng ]
gT
Σt (r, E ) = Σ̂t (r ) g g ( E ). (7)
At this stage, g g is assumed to be spatially uniform. This condition is satisfied for example if
the same set of functions is considered for all the isotopes of a givenconfiguration.

g g
Considering the isomorphism between the Hilbert space Fg = span g1 . . . , g Ng and R Ng , we
g
can construct an orthonormal basis ( f n )n∈[1,Ng ] of Fg in such a way that the different functions
g
fn are Σt -orthogonal. Indeed,

g
Σ̃t (r ) = dEg g ( E )Σt (r, E ) g gT ( E ) (8)
Ig
is unitary similar to a diagonal matrix (see (Le Tellier et al., 2009) for more details) i.e.
g g
Σ̃t (r ) = C g (r )Σt (r )C gT (r ), (9)
with
g
• C g (r ) = a unitary matrix containing the eigenvectors of Σ̃t (r ),
• Σt g (r ) = a diagonal matrix containing its eigenvalues.
Thus, f g (r, E ) = C gT (r ) g g ( E ).

The problem at this stage is that the diagonalization of this operator depends on the spatial
position through Σt (r, E ) i.e. f g (r, E ) depends on r and in the general case the discretized
streaming operator is no longer diagonal. However, in most of the practical cases, the total
cross section is defined as a step function with respect to the space variable i.e. a set of uniform
media is defined and used to represent the spatial distribution of the nuclear data. Let us
consider that the spatial
domain D is split into a set of non-overlapping uniform medium
domains i.e. D = Di . The total cross section (along with the other nuclear data) is
i
represented as
284
gT
Σt (r, E ) = ∑ Π I (E) ∑ ΠD (r)Σ̂ti
g i
(r ) g g ( E ), (10)
g i
where ΠD i is the characteristic function of Di and the flux is expanded as
φ(r, ∑ Π I (E) ∑ ΠD (r) f igT (E)φi (r, Ω).

g
Ω, E ) = g i
(11)
g i
g
For a given i, f g is uniform on Di and is obtained by diagonalizing Σ̃ti as previously
i
described.
For r belonging to a uniform medium domain Di , Eq. 4 can be written without any
complication of the streaming term. In fact, this formulation of the transport equation is
similar to the standard multigroup form. In this case, the mode coupling only appears for
the conditions at the interface Γ ij between two uniform medium domains Di and D j along

Ω. The continuity of φ(r,
Ω, E ) at r ∈ Γ ij implies directly the continuity of φ g (r,
Ω) in the
standard multigroup case:
φj (r,
Ω) = φi (r,
g g
Ω ), (12)
while, in our case, it translates into
φ j (r,
Ω) = C j Ci φi (r,
g gT g g
Ω ). (13)
When crossing an interface between two uniform media domain, a change of basis with
respect to the energy expansion has to be performed in order to maintain a diagonal group
transport operator over the whole domain.
2.2 Case of study

The numerical study of the proposed algorithms will be limited to the fine structure flux
equation for a single isotope diluted in a mixture of non-resonant isotopes in an infinite
homogeneous medium in such a way that only the energy variable has to be discretized. The
total cross section is written as Σ+ ∗ ∗ +
t + N σt ( E ) considering that Σ t is constant; ∗ refers to the
g ∗g
resonant isotope. Considering f ( E ), the σt -orthogonal and orthonormal basis of Fg , the
weak form of the fine structure flux equation is written as
G

σt ∗ g + σd φ g = ∑ σs ∗ g← g φ g + σd dE f g ( E ), (14)
g =1
Ig
In matrix-vector form, this linear system is summarized as
HΦ = SΦ + Q. (15)
We will also consider that the source-flux coupling in Eq. 15 is solved by a simple Richardson
iterative scheme under the form
HΦ n+1 = SΦ n + Q. (16)
H −1 will be denoted A in the remainder.

of the Neutron
2.3 Wavelet-based elements

Let θ be some function in L2 (R ). We consider the translates and dilates of θ denoted θ j,k such

that θ j,k ( x ) = 2 j/2 θ (2 j x − k)( j ∈ Z, k ∈ Z ) and Vj = span θ j,k , k ∈ Z the generated linear
spaces.
θ iscalled the father wavelet or scaling function and is constructed in such a way that
Vj , j ∈ Z is a multiresolution analysis (MRA) i.e.

• θ0,k , k ∈ Z is an orthonormal system in L2 (R ),
• Vj ⊂ Vj+1 , ∀ j ∈ Z,

• Vj is dense in L2 (R ).
j ≥0

Moreover, for convenience, we consider that θ is normalized in such a way that dxθ ( x ) = 1.
∞

In this case, defining Wj by Vj+1 = Wj ⊕ Vj ( j ∈ Z ), we obtain L2 (R ) = V0 ⊕ Wj . The next
j =0

step is to find a function γ ∈ W0 (γ j,k is defined in a same way as θ j,k ) such that γ0,k , k ∈ Z
is an orthonormal basis of W0 . The existence of such a function is guaranteed but it is not
unique; in anycase, it verifies
dxγ ( x ) = 0 . This function is called the mother wavelet.
Consequently, γ j,k , k ∈ Z is an orthonormal basis of Wj . Note that the mother wavelet is
always orthogonal to the father wavelet.
Within such a framework, any function φ ∈ L2 (R ) has a unique representation in terms of an
L2 -convergent series: (see (Hardle et al., 1997))
∞
φ( x ) = ∑ α0,k θ0,k (x) + ∑ ∑ β j,k γ j,k (x), (17)
k j =0 k
where α j,k and β j,k correspond to the orthogonal projection of φ on θ j,k and γ j,k respectively.
In the present work, we consider for the basis functions g g in each group Ig a subset of

θ0,k k , γ j,k obtained by the sampling, discrete wavelet transform and thresholding
j,k
∗g
of σt ( E ) or an approximate flux restricted to Ig . This is to be distinguished from the work
of (Yang et al., 2010)
where the basis is composed of the scaling functions for a given dilation
order j i.e. g g = θ j,k .
k
In the following, we restrict ourselves to compactly supported wavelets introduced by
(Daubechies, 1992) constructed starting from a function m0 (ξ ) = √1 ∑ hk e−ikξ where hk
2
k
are real-valued coefficients such that only a finite number M (the support length) of hk are
non-zero. In this context, the MRA obeys
θ j−1,l = ∑ hk−2l θ j,k , (18)

k
γ j−1,l = ∑ gk−2l γ j,k , (19)
k
and the decomposition of a sampled N −length signal is obtained efficiently by the discrete
wavelet transform (DWT) based on the cascade algorithm proposed in (Mallat, 1989).
286
Following such a wavelet decomposition, the thresholding consists in replacing Eq. 17 by

J
φ( x ) = ∑ α0,k θ0,k (x) + ∑ ∑ β̃ j,k γ j,k (x), (20)
k j =0 k
where ( β̃ j,k ) j,k is obtained from ( β j,k ) j,k and #( β̃ j,k ) j,k #( β j,k ) j,k
A natural criterion is to discard coefficients lower than a given cut-off ε i.e.

0 if | β j,k | ≤ ε max j,k ( β j,k ),
β̃ j,k = (21)
β j,k otherwise.
This method is called hard thresholding. We refer the interested reader to (Le Tellier et al.,
2009) for a comparison of different wavelet filters and thresholding strategies in this context.
3. Adaptivity
In the context of Eq. 16, adaptive algorithms aim at improving the operators discretization
during the iterative process by dynamically selecting the basis functions and consequently,
optimizing the computational cost and control (at least partially) the error on the final solution.
The proposed algorithms aim at reducing the computational cost defined as the sum of the
supports size at each iteration:
nbIter
cost = ∑ #ΛiA + #ΛSi , (22)
i =1
where ΛSi (resp. ΛiA ) represents the support of operator S (resp. A) at iteration i. Actually,
the computational cost required to solve Eq. 16 is directly linked to the size of the operators
manipulated: ΛSi for the construction of matrix Si and ΛiA the order of the system used for
iterations. It justifies the use of Eq. 22 as a measure of the algorithm computational cost.
Our work differs from the approach in (Cohen, 2003) where the goal was to minimize the
final support. Here, the purpose is to find a balance between the number of iterations and
the support size. In the following, two different algorithms are presented and tested. Both are
based on a decomposition of the error in terms of the Richardson iterations residual (δres ) and
the errors due to the discretization of A and S operators (denoted δ A and δS respectively):
n +1
Φ − Φ 1
≤ δ A
+ δ S
+ δ res
= NB. (23)
Φ n +1 1 − AS
Sections 3.2 and 3.3 explicit this bound for both algorithms. The first version, inspired from
(Cohen, 2003), uses two levels of iterations: one in order to increase the support and one to
converge the residual. The single-loop algorithm is proposed as a simplification of the first
one and a way to correlate the errors on the operators and the residual is detailed.
3.1 Numerical cases of study

As AS plays an important role in both algorithms presented in Sections 3.2 and 3.3, tests
are performed on different isotopes and energy ranges (the energy mesh used for this study
contains 172 groups) as presented in Table 1.
of the Neutron
Isotope AS Energy range (eV) Energy groups

238 U 0.26 6.16 - 7.52 88
56 Fe 0.10 1018 - 1230 56
16 O 0.01 273.2E3 - 498.9E3 26-29
Table 1. Numerical cases of study for the two adaptive algorithms
3.2 Two-loop algorithm

In this algorithm, an outer iteration loop (index j) is added. At a given iteration j, the following
system is solved:
Φ nj++11 = A j+1 S j+1 Φ nj+1 + Q , (24)
with A j+1 (resp. S j+1 ) representing matrix A (resp. S) restricted to Λ jA+1 (resp. ΛSj+1 ) support.
The error is given by:

Φ nj++11 − Φ = A j+1 S j+1 Φ nj+1 + Q − A(SΦ + Q)

= A j+1 − A S j+1 Φnj+1 + Q + A S j+1 − S Φnj+1 + AS Φnj+1 − Φ .
(25)
It follows that the relative error can be expressed by Eq. 23 with

S j+1 − S Φ nj+1
δS = A , (26)
n +1
Φ j+1

A j+1 − A S j+1 Φ nj+1 + Q
δ =
A , (27)
n +1
Φ j+1

n +1
Φ j+1 − Φ nj+1
δ = AS
res . (28)
n +1
Φ j+1
A main issue is the choice of the matrices S j+1 and A j+1 or, in other words, the selection of the
wavelet supports. The idea in the remainder is to monitor the errors related to the operator
discretizations using the numerical residual in order to obtain a relation of the type:

n +1 n +1
Φ j+1 − Φ Φ j+1 − Φ nj+1

n +1 ≤ K n +1 , (29)
Φ j+1 Φ j+1
where K is a given constant. The error on the flux is thus controlled by the residual at each
iteration.
The error on δS (resp. δ A )can be practically controlled by a thresholding on the product
SΦ n (resp. A S j+1 Φ nj+1 + Q ) ensuring:

(S j+1 − S )Φ nj+1 ≤ j+1 Φ nj+1 , (30)

A j+1 − A S j+1 Φ nj+1 + Q ≤ j+1 Φ nj+1 . (31)
288
Remaining coefficients give the new supports ΛSj+1 and Λ jA+1 such that #ΛSj+1 #ΛS and
#Λ jA+1 #Λ A where ΛS and Λ A are the support of S and A operators approximated by
a large number of coefficients. The localization property of wavelets ensure that these two
supports slowly increase when j+1 and j+1 decrease (see (Cohen, 2003) for more details).
By applying the procedures exposed above to A and S, Eq. 23 becomes:
⎛ ⎞
n +1 n n n +1 n
Φ j+1 − Φ 1 Φ j +1 Φ j +1 Φ j +1 − Φ j +1
⎝ ⎠.
n+1 ≤ 1 − AS j+1 n+1 + j+1 A n+1 + AS n +1
Φj +1 Φ
j +1 Φ j +1 Φ j +1
(32)
Note however that the thresholding procedure described for operator A in Eq. 31 cannot
be applied in the general context of the spatially-dependent transport equation (Eq. 4). A
possibility is to use the same support for operators A and S. Such a solution has been tested in
(Fournier & Le Tellier, 2009). Even if the convergence is deteriorated compared to the solution
with two different supports for S and A, results are interesting and show that the adaptive
algorithms proposed in this book chapter are extensible to the general
problem.
As proposed in (Cohen, 2003), a geometrical decreasing sequence j is fixed and iterations
on n are performed until the residual becomes inferior to the value imposed by this sequence.
To link j and j , we ensure that the first two terms defined in Eq. 32 decay at the same rate by
imposing:
j +1
j+1 = . (33)
A
At a given iteration j, Richardson iterations are carried out in order to ensure:

Φ j+1 − Φ j j +1
≤ . (34)
AS
Φ j +1
Combining Eqs. 33 and 34 with the bound of Eq. 32 guarantees the convergence of the error:

Φ j+1 − Φ 3 j+1
. (35)
1 − AS
Φ j +1
The devised algorithm

is written in pseudocode in Algorithm 1.
The choice of j is arbitrary and some numerical tests have been performed with different
values. A possible choice is
j +1

= = A j S j ,
j
the rate of convergence of Richardson method.
Indeed, two asymptotic behaviours can be observed depending on the value with respect to
ρ = AS as presented in Figure 1 for 16 O where ρ = 0.01:
• ρ (case = 1/2 in Figure 1): Richardson iterative scheme converges rapidly (and
even in one iteration in the presented case) and the error decreases linearly at the same
rate than the sequence ( j ) but it needs many outer iterations. In our example, the slope of
the straight line is equal to 0.3 = log(1/2) = log().
of the Neutron
Result: Solve HΦ = SΦ + Q thanks to an adaptive procedure

Input : Matrix H, S and vector Q calculated on an “infinite” support
given accuracy tol
Output: Φ: flux solution of Φ = H −1 (SΦ + Q) at accuracy tol
Data: Λ: support, : accuracy
Φ0 = 0, Λ0 = ∅, 0 = 1 ;
j=0;
err = 1 ;
1− AS
while j ≥ tol 3 do
j ← j+1;
j ← j−1 ;
Φ0j ← Φ j−1 ;
n←1;

while err ≥ ASj do
tmp = SΦ nj−−11 ;
prod = Thresholding(tmp, j ) ;

% remove smallest coefficients of tmp, guarantee tmp − prod ≤ j Φ nj−−11
Rnj = prod + Q ;
n
ΛS j = Support( Rnj ) ;
Φ nj = H −1 Rnj ;

Φ nj = Thresholding(Φ nj , H −j 1 ) ;
n
Λ A j = Support(Φ nj ) ;

n
Φ j − Φ nj −1
err ←
n ;
Φ j
n ← n+1;
end
Φ j+1 = Φ nj ;
end
Algorithm 1: two-loop adaptive algorithm
• AS : the number of coefficients kept increases rapidly and several Richardson

iterations are necessary to converge at a given support.
= ρ seems a good compromise between increasing too slowly the support causing useless
iterations and keeping too many coefficients which implies the resolution of a uselessly large
linear system.
Figure 2 presents the L2 −error as a function of the cost for 238 U and 16 O. Showing this two
cases is interesting because they exhibit a different spectral radius ( AS = 0.26 for 238 U and
0.01 for 16 O). As decreases, the cost decreases to a minimum value ( = 18 for 238 U), and
then increases again. As decreases, less iterations are performed which improves the cost;
below a given value too large systems are solved and the cost increases (these are the two
290
1
ε = 1/2, NB
ε = 1/2, L2 error
ε = ρ, NB
0 ε = ρ, L2 error
−1
−2
error criterion (log) −3
−4
−5
−6
−7
−8
0 5 10 15 20
number of iterations
Fig. 1. L2 −error and numerical bound for different values on groups 26 to 29 of 16 O

−1 −1
ε = 1/16 ε = 1/512
ε = 1/12 ε = 1/256
ε = 1/8 ε = 1/128
−1.5 −2
ε = 1/4 ε = 1/64
ε = 1/2 ε = 1/32
ε=ρ ε = 1/16
−2 −3
ε=ρ
error criterion (log)
−2.5 −4
−3 −5
−3.5 −6
−4 −7
−4.5 −8
−5 −9
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 200
cost cost
Fig. 2. Relative error versus cost for different values for group 88 of 238 U (left) and for
groups 26 to 29 of 16 O (right)
behaviours illustrated in Figure 1). Obtaining the value of this minimum

is not possible in the

general case but let us mention that the use of the spectral radius A j S j ensures a reasonable
cost. Figure 3 further illustrates the evolution
of the cost as a function of the parameter for
16 O and confirms the choice of S to minimize the cost.
A j j
of the Neutron
11
1050
1000
950
900
cost
850
800
750
700
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07
parameter value
Fig. 3. Cost versus for a given accuracy of 10−6 for groups 26 to 29 of 16 O
3.3 Single-loop algorithm

The previous algorithm was directly inspired from Cohen (2003) and uses two levels of
iterations which complicate the source iterations. Besides, the choice of the series ( j ) is not

obvious even if a geometrical sequence with a common ratio equal to A j S j gives good
results. As a simplification of this algorithm, a one-loop version is proposed, i.e. the iterative
system is written as:
Φ n +1 = A n +1 S n +1 Φ n + Q . (36)
A single loop means that the residual is no longer directly controlled and a strategy to handle
this point has to be devised. At a given iteration, the residual is given by:

Φ n +1 − Φ n = A n +1 S n +1 Φ n + Q − A n S n Φ n −1 + Q

= A n +1 − A n S n +1 Φ n + Q + A n S n +1 − S n Φ n + A n S n Φ n − Φ n −1 .
(37)
And the same relationship as the one for the two-loop algorithm holds for the actual error:

( I − AS ) Φn+1 − Φ = An+1 − A S n+1 Φn + Q + A S n+1 − S Φn − AS Φn+1 − Φn .
(38)

Substituting Φ n+1 − Φ n as given by Eq. 37 in Eq. 38 leads to an error bound given by Eq. 23
292
with
n+1
A S − S − SAn S n+1 − S n Φn
δS = , (39)
Φ n +1
n+1 n+1 n
A − A − AS An+1 − An S Φ +Q
δ A = , (40)
Φ n +1
n
Φ − Φ n −1
δres = ASAn S n
Φ n +1 . (41)
Such a bound for the operator-related error δ A (resp. δS ) is interesting because it takes
into account both An+1 − A (resp. Sn+1 − S ), the distance between the current operator
and the complete one, and An+1 − An (resp. Sn+1 − Sn ), the distance between two
successive operators. The direct control of the numerical residual with Richardson iterations
in the previous algorithm is now “replaced” by the introduction of the distance between two
successive operators in the error bounds on A and S. As the first term decreases with n until 0,
the
second one until A −
increases An (resp. S − Sn ). Depending on the value of AS ,
An+1 − A + AS An+1 − An can be strictly decreasing or presents a minimum or a
maximum (Figure 4).
−4
|| A − A || x 10
n+1
18 ||An+1 − An ||
8
||An+1 − A||
|| An+1 − A || + ρ || An+1 − An ||
||An+1 − An||
16 7 ||A − A|| + ρ ||A − A ||
n+1 n+1 n
14
6
12
5
10
4
8
3
6
4 2
2 1
0
0
0 20 40 60 80 100 120 140 0 100 200 300 400 500 600 700
number of coefficients number of coefficients
Fig. 4. Comparison of error terms defined in Eq. 40 for group 88 of 238 U with AS = 0.26
(left) and with AS artificially increased to 0.8 (right)
Even if the general behaviour is not known, the initial and final bounds are given by:
δ(SSn+1 =S) = δSfin = ASAn (S − S n ) Φ n , (42)

n +1 n
δ(AAn+1= A) = δ Af in = AS ( A − A ) S
n
Φ + Q , (43)
δ(SSn+1=Sn ) = S
δini = A (S − S ) Φ ,
n n
(44)

δ(AAn+1= An ) = δini
A
= ( A − A n ) S n +1 Φ n + Q . (45)
of the Neutron
13
As AS < 1 (ensuring the convergence of Richardson iterations), it guarantees that δSfin <
S and δ A < δ A . These error bounds are at the basis of our algorithm. Three different
δini f in ini
cases are considered:
• δres ∈ [ δSfin , δini
S ]. It is possible to decrease the error due to operator S discretization to
the numerical residual so S n+1 is chosen to ensure δS ≈ δres .

• δres < δSfin . Numerical residual is too small to be reached directly. Error on operator S is
reduced to
δS = αδini
S
+ (1 − α)δSfin , (46)
with α fixed in [0, 1].
• δres > δini
S . The numerical residual is not yet enough converged so the support of
operator S is not modified, S n+1 = S n .

The same approach is used to treat δ A .
Figure 5 presents the behaviour of the three error terms and the numerical bound defined
by Eq. 23. When AS is low (Figure 5 (left)), Richardson iterations converge rapidly and
do not slow the convergence of other terms. When AS tends to 1 (Figure 5 (right)),
more Richardson iterations are needed in order to converge the numerical residual and the
operators support grows slowly and stepwise. It explains the decay by step observed for the
operator discretization errors in Figure 5 (right).
δεA 1
δεS
δεres A
0 NB δε
δεS 0 δεres
NB
−1
−1
−2 −2
−3
−3
−4
−4
−5
−5
−6
−6 −7
1 2 3 4 5 6 7 8 9 0 2 4 6 8 10 12
number of iterations number of iterations
Fig. 5. Comparison of error terms on group 88 of 238 U with AS = 0.26 (left) and with AS
artificially increased to 0.8 (right)
The only remaining parameter is α. A numerical study is performed to give us some
information about the optimal value.
Figure 6 shows that the choice of this parameter is important regarding the cost of the
algorithm. If not enough coefficients are kept at each iteration, the rate of convergence is
low which causes an important cost. On the opposite, if a large number is kept, large systems
294
1800 700
650
1600
600
1400
550
1200
cost
cost
500
1000
450
800
400
600
350
400 300
0.4 0.5 0.6 0.7 0.8 0.9 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
parameter value parameter value
Fig. 6. Cost of the algorithm depending on α for a given accuracy = 10−5 on group 56 of
56 Fe (left) and = 10−4 on group 88 of 238 U (right)
have to be solved. An interesting compromise seems to keep coefficients in order to reduce

the error by about half.
3.4 Comparison of the two algorithms

A comparison of the two algorithms and of the non-adaptive strategy is done in this section.
All tests are performed by doing hard thresholding on an approximated flux and using
symmlets of the 6th order. All strategies are compared as a regard of the number of kept
coefficients but also the cost defined by Eq. 22. To make non-adaptive and adaptive strategies
comparable, non-adaptive Richardson iterations are stopped when δres is of the same order
as δS + δ A in such a way that the cost of the non-adaptive algorithm is nearly optimal.
Figure 7 (resp. Figure 8) presents results obtained on 238 U (resp. 56 Fe).
Figures 7 and 8 present coherent results and clearly highlight the interest of the two adaptive
algorithms. The use of the spectral radius in the two-loop algorithm and the construction
of our single-loop strategy make the convergence nearly independent of the case of study.
Moreover, let us recall that the non-adaptive algorithm used in this study exhibits a nearly
optimal cost and requires the control of the different error terms (δS , δ A and δres ) as
explained at the beginning of this section.
While both adaptive algorithms exhibit similar performances, the single-loop algorithm
presents some advantages. First, the treatment of source iterations is easier with only one
level of iteration. Then, the choice of the decreasing series ( j ) is problem-dependent and
more difficult to compute compared to the choice α = 0.5 in Eq. 46 for the one-loop algorithm.
4. Conclusion
Considering a wavelet-based Galerkin discretization for treating the energy variable in the
neutron transport equation, this chapter has proposed two adaptive algorithms for the
Richardson iterative scheme that is commonly used to solve the source-flux coupling. While
of the Neutron
15
non−adaptive algorithm
single−loop algorithm non−adaptive algorithm
−1
10 double−loop algorithm −1 single−loop algorithm
10 double−loop algorithm
−2
−2 10
10
error criterion
error criterion
−3
10
−3
10
−4
10
−4
10
−5
10
−5
10
50 100 150 200 500 1000 1500 2000 2500 3000 3500
number of coefficients cost
Fig. 7. Algorithms comparison in terms of the convergence (left) and the cost (right) for
group 88 of 238 U
non−adaptive algorithm −1 non−adaptive algorithm
−1 10
10 single−loop algorithm single−loop algorithm
double−loop algorithm double−loop algorithm
−2
−2 10
10
error criterion
error criterion
−3
−3 10
10
−4
−4 10
10
−5
−5 10
10
100 200 300 400 500 600 700 200 400 600 800 1000 1200 1400 1600
number of coefficients cost
Fig. 8. Algorithms comparison in terms of the convergence (left) and the cost (right) for
group 56 of 56 Fe
the first algorithm based on two nested loops is a modification of an algorithm previously
proposed in the literature, the second one has been devised as a simplification that retains
the same convergence properties. Both approaches are based on a formal decomposition of
the error into three terms: two of them are related to the operators discretization while the
third one is the Richardson residual. The algorithms then consist in a strategy to monitor and
relate these three terms in such a way that error can be controlled by the Richardson iterations
296
residual. As a benefit of these algorithms, the accuracy of the final solution is known and the
cost to obtain it has been decreased by adapting the size of the system during iterations. The
performances of these algorithms have been demonstrated in the restricted framework of the
fine structure flux equation in an homogeneous infinite medium. In the context of neutron
transport calculations, the modifications necessary for spatially-dependent cases have been
mentioned.
5. References
Allen, E. J. (1986). A finite element approach for treating the energy variable in the numerical
solution of the neutron transport equation, Transport Theory and Statistical Physics
15(4): 449–478.
Cohen, A. (2003). Numerical Analysis of Wavelet Methods, Vol. 32 of Studies in Mathematics and
its Application, North Holland.
Daubechies, I. (1992). Ten Lectures on Wavelets, CBMS-NSF Regional Conference Series in
Applied Mathematics, SIAM.
Fournier, D. & Le Tellier, R. (2009). Adaptive algorithms for a self-shielding treatment using
a wavelet-based Galerkin method, Proc. of Int. Conf. on Mathematics, Computational
Methods & Reactor Physics M&C 2009, ANS, Saratoga Springs, USA.
Hardle, W., Kerkyacharian, G., Picard, D. & Tsybakov, A. (1997). Wavelets, approximation and
statistical applications, Seminar Paris-Berlin.
Hébert, A. (2007). A review of legacy and advanced self-shielding models for lattice
calculations, Nuclear Science and Engineering 155(2): 310–320.
Le Tellier, R., Fournier, D. & Ruggieri, J. M. (2009). A wavelet-based finite element method
for the self-shielding issue in neutron transport, Nuclear Science and Engineering
163(1): 34–55.
Mallat, S. G. (1989). A theory for multiresolution signal decomposition: the wavelet
representation, IEEE Transactions on Pattern Analysis and Machine Intelligence
11: 674–693.
Mosca, P., Mounier, C., Sanchez, R. & Arnaud, G. (2011). An adaptive energy mesh constructor
for multigroup library generation for transport codes, Nuclear Science and Engineering
167(1): 40–60.
Yang, W., Wu, H., Zheng, Y. & Cao, L. (2010). Application of wavelets scaling function
expansion method in resonance self-shielding calculation, Annals of Nuclear Energy
37(5): 653–663.

Discrete Wavelet Transforms - Algorithms and Applications PDF

Uploaded by

Copyright:

Available Formats

Discrete Wavelet Transforms - Algorithms and Applications PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Discrete Wavelet Transforms - Algorithms and Applications PDF

Uploaded by

Copyright:

Available Formats

DISCRETE WAVELET

Copyright © 2011 InTech

Publishing Process Manager Ivana Lorkovic

First published August, 2011

A free online edition of this book is available at www.intechopen.com

Discrete Wavelet Transforms: Algorithms and Applications, Edited by Hannu Olkkonen

Part 1 Discrete Wavelet Transform Based Hardware Algorithms 1

Chapter 1 Discrete Wavelet Multitone Modulation

Chapter 2 A Scalable Architecture for Discrete

Chapter 3 VLSI Architectures of Lifting-Based

Chapter 4 Simulation of Models and BER

Chapter 5 Several Kinds of Modified SPIHT Codec 67

Part 2 Image Processing Applications 79

Chapter 6 Multiresolution Approaches for Edge Detection and

Chapter 7 Low Bit Rate Video Compression Algorithm

Chapter 8 Low Complexity Implementation of Daubechies

Chapter 9 Discrete Wavelets on Edges 135

Chapter 10 Discrete Wavelet Transform and Optimal Spectral

Part 3 Discrete Wavelet Transforms for Watermaking 177

Chapter 11 Watermarking-Based Image Authentication

Chapter 12 Application of Discrete

Part 4 Discrete Wavelet Transform Algorithms 219

Chapter 13 Shift Invariant Discrete Wavelet Transforms 221

Chapter 14 Condition on Word Length of Signals and

Chapter 15 Wavelet-Based Analysis and

Chapter 16 An Adaptive Energy Discretization

Hannu Olkkonen, Professor

Discrete Wavelet Transform Based

Discrete Wavelet Multitone Modulation for

GIK Institute of Engineering Sciences and Technology, Topi

2. Basics of wavelet filter banks & multirate signal processing systems

2.1 Analysis and synthesis filter banks

component ( ) and a low frequency component ( ). The input signal ( ) is passed

Fig. 1. Two-channel Analysis filter bank.

Fig. 3. Two-channel Synthesis filter bank.

2.2 Quadrature mirror filter bank

This condition may be satisfied by choosing ( ) = (− ) and ( )= (− ), then the

Fig. 4. Two-channel QMF bank.

Fig. 6. Two-channel Transmultiplexer.

3. Discrete multitone modulation technique

Fig. 7. Functional block diagram of a DMT transceiver.

3.1 Evolution of discrete wavelet multitone modulation

4. Discrete Wavelet Multitone (DWMT) in Digital Subscriber Line (DSL)

4.1 System model of DWMT

Fig. 8. Functional Block diagram of DWMT system.

4.1.1 Water filling bit loading

4.2 ADSL channel

difficulties in achieving the objective of high-speed and reliable communication (Cook, et

Fig. 11. Frequency response of telephone line FIR channel.

Fig. 12. DSL channel impulse response.

Parameter WPT-DWMT DFT-DMT

Fig. 14. (b) PSD-FEXT for downstream ADSL in G.992.1/G.992.2 standard.

4.3 Time domain equalization

4.3.1 ZF finite length equalizer

4.3.2 MMSE criterion

4.4 Simulation results

Parameter WPT-DWMT DFT-DMT

Fig. 16. Equalizing Zero-Forcing filter frequency response.

5. Pros & cons of applying DWT in multicarrier modulation techniques