11

CHAPTER 2

2 WAVELET TRANSFORMS

2.1 INTRODUCTION TO WAVELET

A wave is defined as an oscillating function of time or space,

such as a sinusoid. A wavelet is a “small wave", which has its energy
concentrated in time and used for the analysis of non-stationary signals. It also
has the ability to analysis time and frequency components simultaneously.

2.1.1 Overview of Wavelet

The primary scheme of wavelets is to analyze according to scale.

The idea in wavelet is based on the early superposition concept. This
superposition of sine‟s and cosines was introduced by Fourier to represent
other functions. But, in wavelet analysis, the scale plays a specific role.
Generally wavelet processes the data at different scales or resolution. Hence
wavelets are useful to visualize both the gross and small features.

This makes wavelets an interesting tool in signal and image

analysis. For many years, research was going on to find an approximate
function for Fourier bases, which varied up to infinity. Wavelet is the one
solution that approximates the sharp discontinuities in finite domain.
 12

In wavelet, a mother wavelet is used to analyze both the temporal

and frequency information of data. Therefore, wavelet is the only linear
transform used to analyze non stationary signals. Wavelet represents the
original data as a linear combination of wavelet coefficients. Thus the wavelet
coefficients can be further threshold to represent the data sparsely. This sparse
representation of original data makes wavelet an efficient tool in various fields
like, denoising, compression, sub-band coding, astronomy etc.

2.1.2 History of Wavelet

The wavelet uses the concept of Fourier. In Fourier, a complex

function is approximated as a weighted sum of simpler functions using a basis
function. The basis function can be named as a building block. Sinusoidal
signal of different frequencies is used as basis function. The building blocks
give the details of the frequency content of the original signal. Thus, the
Fourier representation is used in variety of fields.

A Fourier representation provides only the frequency information

of the signal and not the time information. In other words, Fourier stretches the
signal to infinity in time and hence can‟t be used for non-stationary signals.
Hence the time domain representation of a signal does not provide any
information about the spectral content of the signal.

Therefore, to analyze non stationary signals, both time and

frequency representation is needed. Time-frequency representation (TFR) is
obtained first by modifying Fourier transform as short time Fourier transform
(STFT) by Gabor (1946). STFT segmented the signal using a time-localized
window as Fourier and analyzed each segment. He used STFT to represent
communication signals in TF plane. In 1947, Jean Ville represented the energy
of a signal in TF plane and named as Wigner-Ville transform. Following this
 13

many TFRs have been developed and they varied from one another in the
selection of the window function. The drawback of all these methods is the use
of the same window for entire signal. Another important one to note in STFT
is either narrow window is used to analyze neither high frequency
components, nor wide windows to analyze low frequency components, but not
able to analyze both frequency components. This drawback was overcome by
Morlet (1984), a geophysical engineer who used different window function to
analyze different frequency bands. These windows were generated by dilation
or compression of a prototype Gaussian.

The windows are called as wavelets since they were small and
oscillatory. Grossman & Morlet (1984) formalized the transformation and
found the inverse transformation. This wavelet transform was the rediscovery
of Alberto Calderón‟s work on harmonic analysis in 1964. The redundancy in
choice of basis function introduced by Mortlet was overcome by Meyer
(1993), a French mathematician. Meyer (1993) developed orthogonal wavelets
with better localization properties.

A German mathematician, Alfred Haar defined orthonormal

wavelets with poor frequency localization. This orthonormal basis function
was expanded by Paul Levey, in 1930. In the meantime, Daubechies (1990)
developed the wavelet frames for discretization of time and scale parameters
of the wavelet transform. Daubechies & Mallat (1989), extended continuous
wavelet transform to the discrete wavelet transform. In 1986, Mallat introduce
the concept of Multi-Resolution Analysis (MRA) for the Discrete Wavelet
Transform (DWT). In MRA the discrete signal was decomposed into its
dyadic frequency bands using a series of low pass and high pass filters. The
DWT is computed from the approximation on various scales. Mallat's work
was based on Quadrature Mirror Filter (QMF) and sub-band coding.
 14

There was research on wavelet basis functions to obtain with

modified properties and MRA algorithms. In 1992, Albert Cohen, Jean
Feauveau and Daubechies introduced bi-orthogonal wavelets, and Coifman et
al. (1992) found an extension of MRA and named as wavelet packets.

2.1.3 Wavelet Terminology

The introduction of wavelet transforms by Mallat is based on the

filter bank theory. The idea of MRA in wavelet leads to the implementation of
DWT in dyadic frequency bands. Wavelet systems have properties like perfect
reconstruction, smoothness, compact energy, symmetry. These properties
make more type of wavelet transforms to be evolved, like the Wavelet Packets
Transform, the Complex Wavelet Transform etc. Therefore, the uses of
wavelet transform in signal and image processing application have been
increased.

2.2 EVOLUTION WAVELET TRANSFORM

The need to analyze the signal in time for its frequency content
led to the evolution of wavelet transform. But the original idea is from Fourier
transform. Here the starting point Fourier transform to the ending point
wavelet transform is described.

2.2.1 Fourier Transform

The basic concept of Fourier transform is to decompose a

function into its harmonic components. It is a function of summation of sine
and cosine terms of different frequency. Generally signals are measured in
time domain as signal amplitude versus time. At the same time, frequency
content of the signal is obtained by measuring the signal amplitude as a
 15

function of frequency. This conversion from time domain to frequency domain

is called as Fourier transforms. The Forward Fourier transform converts a
temporal signal into a spectrum and reverse Fourier transform coverts a
spectrum back into the temporal signal. This can be defined in terms of the
equation as

∫ (2.1)

∫ (2.2)

In the above equations (2.1 & 2.2), denotes signal in time

domain and denotes signal in the frequency domain. Here is the frequency
and stands for time. Fourier transforms equation can be stated as the sum
over all time of the signal multiplied by a complex exponential. The
complex exponential can be broken into real and imaginary sinusoidal
components. The results of the transform are the Fourier coefficients and
graphically, the process looks like

Figure 2.1 Graphical Representation of Fourier Transform

By looking at equation (2.1) we can see that the forward Fourier

transform does not differentiate between stationary and transient parts in
signals. Hence projection of signals using Fourier transform have good
frequency localization but not time localization. Therefore the limitation of the
 16

Fourier transform is time localization that makes it not suitable for different
applications.

2.2.2 Short Time Fourier Transform

To analyze non stationary signals Short Time Fourier Transform

(STFT) is a well-known technique. STFT was introduced to overcome the
limitations of FT by Gabor in 1946. STFT can be computed using the
following steps

1. Choose a window function of finite length.

2. Place the window on top of the signal at t=0.
3. Multiply the signal using this window.
4. Compute the FT of the multiplied signal, save results.
5. Incrementally slide the window to the right.
6. Go to step 3, until window reaches the end of the signal.

The result is the two dimensional signal and represented mathematically as

{ } ∫ (2.3)

In equation (2.3) STFT is computed for window centered at zero,

is the signal to be analyzed and is the window function. The sliding
window may be a sliding band pass filter and hence STFT can be interpreted
as a filtering operation. STFT uses an arbitrary but constant window length for
analysis, with the assumption of non-stationary signals to be stationary. The
mapping of STFT of a signal is shown in figure 2.2.
 17

Figure 2.2 Mapping of Short Time Fourier Transform of a Signal

In STFT the exact time-frequency representation of a signal can‟t

be known and only we can know about at which time intervals what interval of
frequencies are present. This is called as the Heisenberg uncertainty principle.
When the energy function and its Fourier transform are centered on
zero in both time as well as a frequency will satisfy the following conditions

∫ | | (2.4)
‖‖

∫ | | (2.5)
‖‖

Here ‖ ‖ is the norm of the function . STFT uses constant window for a whole
analysis of signals, but many signals require a more flexible approach. A wide
window provides good frequency resolution and poor time resolution. A
narrow window gives good time resolution and poor frequency resolution. In
order to obtain better analysis, we can vary the window size more accurately
either in time (or) in frequency.

2.2.3 Wavelet Transform

The limitations of Fourier transform and STFT are overcome by

the introduction of wavelet transform in 1984 by Grossman and Morlet.
 18

Wavelet transform divide the given signal into different scale of components.
Each scale of component can be analyzed with different resolutions. The
mapping of wavelet analysis of a signal is shown in figure 2.3. Wavelet
transform are better than Fourier transform, because signals having sharp
peaks and discontinuities are handled accurately. Hence, wavelet transform is
suitable for signals having both low-frequency components in long time
duration and high-frequency components in short time duration.

The wavelet transform of any function is obtained from a single

prototype wavelet called the mother wavelet, by dilations or contractions
(scaling) and translations (shifts). The mother wavelet is defined as

(2.6)
√

In equation (2.6) is the scale factor and is the translation factor

and the factor √ is for energy normalization across the different scales. The
wavelet transforms of at the time and scale can be given as

∫ (2.7)

The name mother wavelet is, because of two important properties

of wavelet analysis as given below

1. The term wavelet means a small wave. The smallness refers to the
condition that this (window) function is of finite length. The wave
refers to the condition that this function is oscillatory. The term mother
implies that the functions with a different region of support that are
used in the transformation process are derived from one main function.
2. The term translation is related to the location of the window, as the
window is shifted through the signal. This term, obviously, corresponds
to time information in the transform domain.
 19

Figure 2.3 Mapping of Wavelet Transform of a Signal

The comparison of different looks of time, frequency, STFT and

wavelet based view of a signal is shown in figure 2.4.

Figure 2.4 Comparison of Different Looks of Time, Frequency, STFT

and Wavelet Transform of a Signal

2.2.4 Complex Wavelet Transform

The resolution problem of STFT is overcome by the introduction

of continuous wavelet transform. In wavelet analysis the signal is multiplied
by a function similar to the window function in STFT and the transform is
 20

computed for different segments separately. The two main differences

between STFT and Complex Wavelet Transform (CWT) is

1. In STFT negative frequencies are not computed.

2. The width of the window is changed for every single spectral
component in CWT.

Theoretically, CWT can be defined as the sum of over all time of

the signal multiplied by scaled and shifted versions of the wavelet function.
The shifted version of wavelet function is shown in figure 2.5.

Figure 2.5 Shifted Wavelet Function

The results of the CWT are wavelet coefficients and they are the
functions of scale and position. Multiplying each coefficient by an
appropriately scaled and shifted wavelet, produces the constituent wavelets of
the original signal and the graphical representation is as shown below

Figure 2.6 Original Signal and Constituent Wavelets

 21

Mathematically continuous wavelet transform is represented as

( ) (2.8)
√

Here is the scaling factor and is the shifting factor. Normalization

occurs if and only if

‖ ‖ ‖ ‖ (2.9)

The mother wavelet has to satisfy the admissibility condition

| |
∫ (2.10)

In equation (2.10) is the Fourier transform of . Since

have sufficient decay, the admissibility condition reduces and the
wavelet behaves like a band pass filter, this is shown in equation (2.11)

∫ (2.11)

The CWT of a function is defined as in equation (2.12)

∫ 〈 〉 (2.12)

The obtained wavelet coefficients preserve the energy due to the

orthonormal basis decomposition. The basis function is the filter bank
impulse response. In order to focus the long term behaviour of the signal ,
the function is dilated. For the signal resolution is limited by the
frequency content present in it. But in CWT the resolution is not altered
because scale change can be reversed. Therefore the original signal can be
recovered by the reconstruction formula mentioned below
 22

∫ ∫ (2.13)

2.3 DISCRETE WAVELET TRANSFORM

The Discrete Wavelet Transform (DWT) has gained wide spread

acceptance in signal processing and image processing. Because of their
inherent multi-resolution nature, wavelet schemes are especially suitable for
applications where scalability and tolerable degradation are important.

In the discrete case, filters of different cut-off frequencies are

used to analyze the signal at different scales. The signal is passed through a
series of high pass filters to analyze the high frequencies, and it is passed
through a series of low pass filters to analyze the low frequencies. The
resolution of the signal, which is a measure of the amount of detail
information in the signal, is changed by the filtering operations. The scale is
changed by up sampling and down sampling operations. Thus the forward and
inverse DWT definitions are shown in equations (2.14) and (2.15).

∑ (2.14)

∑ ∑ (2.15)

2.3.1 Multi-Resolution Analysis

We say that is the space of all square-summable sequences

if . Then a Multi-Resolution Analysis (MRA) consists of a
sequence of embedded closed spaces as shown in figure 2.7
 23

Figure 2.7 Multi-resolution Closed Space

MRA is designed to give good time resolution and poor

frequency resolution at high frequencies and good frequency resolution and
poor time resolution at low frequencies. This approach makes sense, especially
when the signal at hand has high frequency components for short durations
and low frequency components for long durations.

MRA Requirement is

1. The scaling function is orthogonal to its integer translates.

2. The subspaces spanned by the scaling function at low scales are nested
within those spanned at higher scales as shown in figure 2.8.
 24

Figure 2.8 Subspace Scanned by Scaling Function

3. The only function that is common to all is , { }.

4. Any function can be represented with arbitrary precision { }

The scaling function for each fixed is an orthonormal basis of

the subspace and it is given as

∫ (2.16)

If is the orthonormal component of in the subspace ,

then the wavelet function is

(2.17)

Since MRA have nested subspaces the scaling function satisfies

the following two conditions

√ ∑ [ ] (2.18)

∑ [ ] √ (2.19)

Similarly the wavelet function satisfies the conditions

 25

√ ∑ [ ] (2.20)

∑ [ ] and [ ] [ ] (2.21)

In the above equations [ ] and [ ] are the coefficients of low

pass and high pass filters. For any function using MRA the function can
be modified with the help of scaling and wavelet function as

∑ ∑ ∑ ∑ (2.22)

Here is the starting resolution scale and is the higher resolution scale
with .

2.3.2 Perfect Reconstruction

A discrete two-channel multi-rate filter bank convolves a signal

with a low-pass filter and a high-pass filter and then subsamples the output.
The reconstructed signal is similar to the original signal by satisfying the
following conditions

̃ ̃ (2.23)

̃ ̃ (2.24)

In above equations and are z-transform of and

respectively. The figure 2.9 illustrates the reconstruction process.
 26

Figure 2.9 Reconstruction Process

From the reconstruction filter bank it is clear that the prototype

wavelet has infinite length and all filters require infinite tap. Hence, for this
requirement, finite tap Finite Impulse Response (FIR) filter can be used. The
FIR filter with finite iterations makes the impulse response converge to a
smooth mother wavelet. A filter having this property is called a regular filter.
Regularity is an important feature in wavelet to detect discontinuities. The
wavelet system has the following condition for analysis and synthesis filters

̃ [ ] [ ] (2.25)

̃ [ ] [ ] (2.26)

∑ [ ]̃ [ ] (2.27)

FIR linear phase filters are often used in signal processing

applications. Similarly bi-orthogonal wavelet basis has linear phase and
degrees of freedom. Therefore the scaling and wavelet function can be defined
using 2 scale equation as

√ ∑ [ ] ̄ √ ∑ ̃ [ ] (2.28)

√ ∑ [ ] ̃ √ ∑ ̃ [ ] (2.29)

The bi-orthogonal wavelet also satisfies the following condition

 27

〈 ̃ 〉 [ ] [ ] (2.30)

Then the reconstruction formula is given as

∑∑ 〈 〉̃ (2.31)

The constraints of analysis and synthesis filter is

̃ [ ] [ ] (2.32)

[ ] ̃ [ ] (2.33)

∑ [ ]̃ [ ] (2.34)

2.3.3 DWT Filter Bank

The discrete wavelet transform is implemented by a binary tree of

one dimensional (1D) filters as shown in figure 2.10.
 28

Figure 2.10 Discrete Wavelet Transform 1D Filter Bank

The input signal x is down sampled by filters H0 and H1 into a

low pass component x0 and a high pass component with a ratio of 2:1. The
low pass component can be further split into x00 and x01, which are again
decimated by 2:1. The outputs of the DWT are the band pass coefficients x1,
xo1, xo01, xoo01, and the final low pass coefficients xoo ... oo. Here the
output sample rate is equal to the input sample rate and hence there is no
redundancy in the transform. At first level of split, the procedure computes

[ ] ∑ [ ] [ ] (2.35)

[ ] ∑ [ ] [ ] (2.36)
 29

In order to reconstruct the signal, a pair of reconstruction filters

has to be used. Perfect reconstruction is achieved if the output signal Y(z) is
identical to the input signal X(z). Therefore perfect reconstruction depends on
the design of filters. Hence, xooo may be reconstructed from Xoooo and
xooo1 and so on back to x, using an inverse tree of G filters. The
reconstruction of the original signal can be mathematically defined as

[ ] ∑ [ ] [ ] [ ] [ ] (2.37)

2.3.4 DWT Decomposition

DWT decomposes a signal into a set of mutually orthogonal

wavelet basis functions. The wavelet basis function differs from sinusoidal
basis function in terms of spatial localization. The wavelet functions are
dilated, translated and scaled versions of the mother wavelet . The DWT is a
set of transforms with a different set of wavelet basis functions. The two most
important wavelets are the Haar and Daubechies. Some properties of wavelets
are

1. Wavelet functions are spatially localized.

2. Wavelets are dilated, translated and scaled versions of mother wavelet.
3. Each set of wavelet functions forms an orthogonal set of basis function.

2.3.5 DWT in Two Dimensions

The two dimensional DWT can be computed by repeated

application of one dimensional DWT. The figure 2.11 shows how the one
dimensional DWT is applied to an image. From the figure we can see first the
one dimensional DWT is applied along the rows and secondly applied along
the columns. The DWT results in four sub-band regions, namely LL, LH, HL,
 30

and HH respectively. Here L represents low pass coefficients and H represents

high pass coefficients. The LL sub-band corresponds to the down sampled
factor of the original image. The LH and HL bands denote the horizontal and
vertical features of the original image. The HH band point the high frequency
features in the image. The decomposition can be scaled further by
transforming the LL sub-band to obtain the very low frequencies in the image.

Figure 2.11 1D Discrete Wavelet Transform on an Image

2.3.6 Application of Wavelet Transform

The principles of wavelet are similar to those of the Fourier

developed in 19th century. A wavelet is a mathematical tool used in various
fields as mentioned below.

2.3.6.1 Signal processing

In signal processing, wavelet helps to recover weak signals from

noisy observation. This finds application in X-ray, magnetic-resonance images
and other digital images. Here images are decomposed into four parts by
 31

DWT. The decomposition continues until the size of the sub-image is as small
as we want. The noise removal is done by removing the information not
needed and images are cleaned without blurring the details of the image.

2.3.6.2 Internet communications

In Internet communications, wavelets are used to compress

images using the sub-band coding technique. Wavelets have good energy
compaction property and thus the wavelet coefficients are very small. The
unwanted wavelet coefficients can be discarded with affecting the
reconstructed image. Hence the compressed image is as small as about
25 percent the size of the original image.

2.3.6.3 Source and channel coding

Wavelets help in source coding to produce a compact

representation of information to ease the transmission. In channel coding
wavelets controls the amount of redundancy to reduce the channel noise.

2.3.6.4 Biomedical

Wavelets have achieved great application in biomedical, since all

biological signals are non-stationary signals. Wavelets are used in
electrocardiogram for diagnosing cardiovascular disorders,
electroencephalogram for diagnosing Neurophysiological disorders, etc.
Wavelets have also been used in mammograms, computer tomography and
magnetic resonance images.
 32

2.3.6.5 Nondestructive evaluation

Wavelets are used for flaw detection in nuclear power plant,

aircraft, gas pipelines etc.

2.3.6.6 Partial differential equations

Wavelet basis function discretizes PDE‟s and solves them numerically.

Further, it is extended to finite element analysis.

2.3.7 Limitations of Wavelet Transform

In spite of its efficient sparse representation, the wavelet

transform has some disadvantages, they are listed below.

2.3.7.1 Oscillations

The wavelet coefficients in wavelet transform tend to oscillate

positive and negative around singularities. This makes wavelet transform
unsuitable for many applications. Since the oscillating function passes through
zero, this yields small (or) zero wavelet coefficients.

2.3.7.2 Shift variance

The wavelet transform is shift variant means the change in time

of the input signal causes unpredictable change in output signal. This can be
better understood by considering a piecewise smooth step function as

{ (2.38)
 33

The step function is decomposed up to four levels with “db5” and

shown in figure 2.12. The figure shows the variations in reconstructing signal
at various levels. This shift variance is due to down sampling operation.

Figure 2.12 Decomposition of a Step Function at Four Levels using

“db5”

2.3.7.3 Aliasing

Since the wavelet coefficients are obtained by iterated down sampling with
high pass and low pass filters, they are affected by aliasing. Thus the obtained
wavelet coefficients affect the forward and inverse transforms. The
reconstructed signal has artifacts.
 34

2.3.7.4 Lack of directionality

Wavelet transform produces only three directional features,

horizontal (HL), vertical (LH) and diagonal (HH), but natural images have
random orientation. Hence, for such natural images oriented in several
direction wavelets produces checkerboard pattern. This limits the use of
wavelets to determine the geometric features.

2.4 DUAL TREE DISCRETE WAVELET

Complex wavelets are well suited in image and video processing

application because it has properties like near shift invariance, good
directionality, orthogonality and symmetry. The first generation of complex
wavelets is called as, single tree complex wavelet introduced in 1990 to
overcome the drawback of DWT. But single tree has not gained much
attention, because of lack of good frequency selectivity. This was overcome by
the introduction of dual tree complex wavelet transform by Kingsbury in 1998.
This dual tree complex wavelet has good directionality, good frequency
selectivity and approximate shift invariance properties.

Complex wavelet transform use an analytic filter to decompose

the complex signal into real and imaginary coefficients. Complex wavelets
describe the energy localization of the wavelet basis by computing the
amplitude and phase information with the help of real and imaginary
coefficients. The amplitude describes the strength of the singularities and
phase gives the location of the singularity. The correct value of the amplitude
and phase is determined by analytic (or) quadrature representation of a signal
using Hilbert transform. Thus, the complex wavelet transform is a good
choice, because we can process both magnitude and phase simultaneously.
 35

2.4.1 Analytic Filter

Hilbert transform was introduced by Gabor; the real signal

was converted into a complex signal as in equation (2.39)

(2.39)

In above equation is the Hilbert transform of and it can be written as

{ } (2.40)

The real signal is converted into analytic signal , which

has both the real signal and complex signal . The analytic signal is
also called as a hard space projection of the original signal . The imaginary
part will be orthogonal to i.e., is shifted version of and
is shown in figure 2.13a. In time domain is defined as

{ } ∫ (2.41)

Figure 2.13 a. Imaginary Part is Orthogonal to Real Part

b. Time Domain Representation of Complex Signal
 36

The Hilbert transform relation in frequency domain can be written as

{ { }} (2.42)

In the above equation is the Fourier transform of and

is the Fourier transform of . Is the signum function and
shown in figure 2.13b. One of the main advantage of analytic filter is we can
find the magnitude and phase information simultaneously for the signal as

√ (2.43)

[ ] (2.44)

The analytic representation is a non-negative spectral

representation with half bandwidth utilization. In multi-rate signal processing,
this reduced bandwidth avoids aliasing in filter banks, which is the key feature
of shift invariance property of complex wavelet transform. The spectral
representation of original and analytical signal is shown in figure 2.14.

Figure 2.14 Spectrum of Original and Analytical Signal

The complex wavelet transform is obtained by applying analytic

filter bank structure to the standard discrete wavelet transform. The real value
is converted to complex value and further decomposed into two real valued
 37

filters. These real valued filters have a quadrature impulse response and forms
Hilbert transform pair. This combined pair of filters is called as analytic filter
and is shown in figure 2.15.

Figure 2.15 Analytic Filter

2.4.2 Dual Tree Discrete Wavelet Transform

The key advantage of Dual Tree Discrete Wavelet Transform

(DTDWT) over DWT is its shift invariance and directional selectivity. The
magnitude of complex wavelet coefficient provides accurate scale, space and
orientation. There are two versions of the dual tree wavelet transform, one is
real 2D dual tree and another one is a complex dual tree. The real dual tree is
two times expansive and complex is four times expansive. Both dual tree
wavelets are oriented in six distinct directions.
 38

2.4.3 Real oriented Dual Tree Discrete Wavelet Transform

Real dual tree can be implemented using two separable 2D

discrete wavelet transforms in parallel. The oriented wavelet is produced by
considering the 2D wavelet . is a complex wavelet
and it is given as

(2.45)

Substituting the above equation we have

[ ][ ] (2.46)

[ ] (2.47)

We know that the spectrum of analytic 1D filter is on one side of frequency

plane, but complex wavelet is in only one quadrant and hence complex
wavelet is oriented. The real part of complex wavelet will be the sum of two
separable wavelets and it is given as

{ } (2.48)

The real part is the HH wavelets. The spectrum of real wavelet is

free of checker board artifact since the real wavelet is oriented at . This
orientation depends on the complex wavelet high pass filter which is given as

(2.49)

The real oriented wavelet at is obtained by considering the

complex wavelet

̅̅̅̅̅̅ (2.50)
 39

In the above equation ̅̅̅̅̅̅ is the complex conjugates of

and is the analytic wavelet , then the equation (2.50) can
be expressed as

[ ][̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅] (2.51)

[ ] (2.52)

Now the real part of the wavelet is

{ } (2.53)

By repeating the above procedure on the following wavelets

, , ̅̅̅̅̅̅ and ̅̅̅̅̅̅ where
and , we can obtain four more wavelets.
Therefore a total of six oriented wavelets can be obtained. These six wavelets
can be represented as

(2.54)
√

(2.55)
√

In above equations by substituting we can get the

following six wavelets

(2.56)

(2.57)

(2.58)
 40

(2.59)

(2.60)

(2.61)

The six wavelets oriented in a distinct direction are shown in

figure. Disparate the critically-sampled separable DWT, the real oriented
wavelets are free of checker board artifact. Hence, each sub-band of real
oriented wavelet transforms corresponds to a specific orientation as shown in
figure 2.16. These six oriented wavelets can be implemented using two
separable 2D wavelets in parallel. The six sub-bands obtained are named as
two LH, two HL and two HH sub-bands. The real oriented 2D wavelet
transform is two times expansive.

The inverse 2D oriented wavelet transform can be obtained from

the transpose of forward transform. Hence the transform satisfies Parseval‟s
theorem. Even though the real oriented wavelet is non-separable it does not
have implementation complexity because, implementation requires only
addition and subtraction of respective sub-bands.

This transform is not a complex transform and hence this

transform is oriented but not approximately shift invariant.
 41

Figure 2.16 Six Oriented Wavelets of Real Dual Tree Wavelet transform

2.4.4 Complex 2D Dual Tree Discrete Wavelet Transform

Complex wavelet transform is implemented by considering the

imaginary part of the equation (2.62)

{ } (2.62)

The spectrum of the imaginary part is same as the real part

explained in the previous section. In equation (2.62) and
are HH sub-bands obtained by using row and column filters.
Similar to real parts applying the same procedure on these
wavelets ̅̅̅̅̅̅̅ ̅̅̅̅̅̅ ̅̅̅̅̅̅̅, where

and , we can obtain the six

oriented wavelets like

(2.63)
√

(2.64)
√
 42

By substituting in above equation, we can obtain the

six oriented wavelets as

(2.65)

(2.66)

(2.67)

(2.68)

(2.69)

(2.70)

Therefore, in complex wavelet transform we have six real and six

imaginary parts of wavelets that lead to a total of six pairs of complex
wavelets. The six pairs of complex wavelets are shown in figure 2.17. In this
figure, first row indicates the real part and second row indicates the imaginary
parts of complex wavelets. Third row displays the magnitude of complex
wavelets. Therefore, from the figure it is clear that the real and imaginary parts
are oriented at the same angle and the magnitude of each complex wavelet is a
circular bell shape.
 43

Figure 2.17 Six Oriented Wavelets complex Dual Tree Wavelet

transform. First Row is the Real Part, Second Row is The
Imaginary part and Third row is the magnitude.

2.4.5 Filter bank Structure of Dual-Tree CWT

The 1D filter bank structures of analysis and synthesis section

spanned over three levels are shown in figures 2.18 and 2.19. From the figure
2.18 we can say that the filter bank structure is similar to standard DWT
having twice the complexity. It has two DWT tree in parallel. One tree is the
real and another tree is the imaginary one. The real tree is also called as tree-a
and the imaginary tree is called as tree-b.

The conjugate filters used in 1D DTDWT can be written as , where

and has a set of filters { } and { } in x-direction. Here and
are low pass and high pass filters of a real tree. Similarly and are filters
of imaginary tree. The filters and have different numerical value at
different decomposition levels and the same is true for filters and . In
synthesis filter bank structure, the filter pairs ̃ ̃ and ̅ ̅ form the
orthonormal pairs with respective to analysis filters.
 44

Figure 2.18 Analysis filter bank for 1D DT-DWT

Figure 2.19 Synthesis filter bank for 1D DT-DWT

 45

The filter bank structure of 2D Dual Tree Complex Wavelet

Transform (DTCWT) is shown in figure 2.20. It has four trees for analysis and
synthesis structure. The complex conjugate of two dimension filter bank
structure is given as

( ) ( ) ( ) (2.71)

From the figure 2.20 we can see that tree-a and tree-b is the real
part, while tree-c and tree-d is the imaginary part of the analysis filter bank.
Similarly, tree- ̃ and tree- ̃ are the real part, while tree- ̃ and tree- ̃ is the
imaginary part of the synthesis filter bank structure. The row and column
filtering for tree-a spanned over three levels is shown in figure 2.21. All other
trees have similar structures.

Figure 2.20 Filter bank structure of 2D-DTCWT

 46

Figure 2.21 Filter bank structure of tree-a

DTCWT has shift invariance property. To achieve this property

the sampling rate has to be doubled at each level of the tree. To double the
sampling rate, the samples have to be placed evenly and down sampler have to
be eliminated after level-1. The next thing is to maintain a uniform interval
between the samples of both the trees after level-1. To obtain this, the delays
of one tree should be of half sample different from the other tree.

The shift invariance property is achieved either by using an odd-

even length filters (or) Quarter (Q) shift filter. The analysis structure of odd-
even and Q-shift filters are shown in figures 2.22 and 2.23. In Q-shift filter
bank structure, the brackets indicate the delay of filter with samples.
 47

Figure 2.22 Analysis tree using odd-even filters

Figure 2.23 Analysis tree using Q-shift filters

 48

2.4.6 Properties of DTCWT

2.4.6.1 Shift invariance

Shift invariance means the magnitudes of DTCWT are invariant

to the shift of the input image. This property can be defined by means of an
example by considering a circular disc as shown in figure 2.24. The circular
disc is light in color and present at dark background. The input image after
undergoing DWT and DTCWT up to level 4, the reconstructed wavelet
coefficients using both transform is shown in figure 2.25. The upper row of the
figure 2.25 shows the DTCWT image while the lower row shows the DWT
image. By comparing both the outputs we can say the lower row has aliasing
artifacts, irregular edges, but these are absent in the upper row. Thus the
smooth and continuous image shows the shift invariance property of DTCWT.

Figure 2.24 Input Circular Disc image

 49

Figure 2.25 Reconstructed disc at levels 1 to 4 of an image using the 2D-

DTCWT (upper row) and 2D-DWT (lower row).

2.4.6.2 Directional selectivity

DTCWT exhibit directional selectivity property and this can be

seen when DTCWT gives six band pass sub-images of complex coefficients at
each level. These sub-images are oriented at angles is
shown in figure 2.26. These directional responses are obtained by interpreting
the low pass coefficients as complex pairs and combining them with high pass
coefficients. Therefore DTCWT can separate positive and negative frequency
components in both one and two dimensions.
 50

Figure 2.26 DTCWT

2.4.6.3 Rotational invariance

The DTCWT is also rotation invariant in addition to shift

invariant. Already we have seen in the figure 2.25 that the reconstructed image
by DWT has artifacts while DTCWT image have a sharp edge on a hyperbolic
trajectory. This is due to the fact that all the six directional sub-bands are used
to reconstruct the image and also the diagonal sub-band has a high centre
frequency.