Chapter - 3

DISCRETE WAVELET TRANSFORMS AND

HYPERANALYTIC WAVELET TRANSFORMS: THEORY

In this chapter, Discrete Wavelet Transform and Hyper analytic Wavelet

Transform are presented and its important properties are used for watermarking
outlined. First section deals theoretical aspects of Discrete Wavelet Transform and
its implementations aspects, properties, applications and limitations. Whereas Hyper
analytic wavelet transform, its implementation, applications and advantages are
described in the next section.

3.1 DISCRETE WAVELET TRANSFORM DEFINITION

Discrete wavelet transform (DWT) algorithms have widely accepted as a

standard tool for discrete-time signal and image processing in various fields in
research and industry. Wavelet analysis is a particular time- or space-scale depiction
of signal which has found a wide variety of applications in physics, signal
processing and applied mathematics in the last few years. Wavelets are
mathematical functions that split the data into diverse frequency components and
then examine each component with a resolution matched to its scale. Normally,
Wavelet Transform gives a time-frequency representation of the signal. It was
developed to conquer the limitation of the Short Time Fourier Transform (STFT),
which can also be employed to analyze the non-stationary signals. The STFT
provides a constant resolution at all frequencies, whereas the Wavelet Transform
exploits multi-resolution approach by which different frequencies are analyzed with
diverse resolutions. Among the transform domain, DWT based watermarking
techniques are receiving more popularity since DWT has more advantages than
other transforms such as progressive and low bit-rate transmission, quality
scalability and region-of-interest (ROI) coding demand more proficient and flexible
image coding that can be exploited for both image compression and watermarking
applications.

57
3.2 DISCRETE WAVELET TRANSFORM ANALYSIS

The continuous wavelet transform (CoWT) has the drawbacks of

redundancy and impracticability with digital computers. As scaling parameter (a)
and shifting parameter (b) take continuous values, the resulting CoWT is a very
redundant representation, and impracticability is the result of redundancy.
Therefore, the scale and shift parameters are evaluated on a discrete grid of time-
scale plane leading to a discrete set of continuous basis functions.

The discretization is performed by setting


a  a and       for j, k Z, ... (3.1)

where, ao > 1 is a dilated step and b0 0, is a translation step. The family of

wavelets then becomes,

 
,         3.2

and the wavelet decomposition of the function  is

     ,   ,   3.3

Where 2 Dimensional set of coefficients D x, y is called DWT of given function

xt.

The most widely used form of such discretization with a= 2, and b= 1 on a

dyadic time-scale grid is shown in figure 3.1. Such a wavelet transform is described
as the standard DWT.

Figure 3.1 Standard DWT on dyadic time-scale grid

58
 The selection of (t) is made in such a way that basis function set {j,k}
constitute an orthonormal basis of L2 () so that

'
 ,    $ ,  & < , ,  > 3.4
'

3.2.1. Multi-Resolution Analysis

Multi-Resolution Analysis (MRA) consists of a sequence of embedded

subspaces V2 V1 V0 V-1 V-2 .. of L2 (R) as shown in figure 3.2.

W2 V3 Vj: Subspaces
W1 corresponding to scaling
W0 basis (approximations)
V2
V0 Wj: Subspaces
corresponding to
V1 wavelet basis (details)

V0 W 0 W 1 W 2

Figure 3.2 Nested vector spaces spanned by scaling and wavelet basis

The MRA follows the following conditions:

1. Vj V j+1, j

2. V-= {0} and V= L2

3. x (t)Vjx (2t) Vj+1

4. V2 = V0 +W0 +W1

5. L2 = + W-2 + W-1 + W0 + W1 + W2 +. = V0 + W1 + W2 +

6. W- + .+ W-2 + W-1 = V0

59
A scaling function t is introduced such that for each fixed j, the family



. ,  2  . /2    0 ,  1 2& $ . &  1  3.5

is an orthonormal basis of the subspace Vj.

If Wj is orthonormal component of Vj in subspace Vj+1, then there exist a

function t (Mother Wavelet) such that for each fixed j the family



,  2  /2    0 ,  1 3.6

is an orthonormal basis of the subspace Wj.

Because of the nested subspaces and MRA condition (3), the scaling
function satisfies the following 2-scale (dilation or refinement) equation,

.  2 '
' [2].2  2 , n ... (3.7)

'
' 2 = 2

Where it satisfies the admissibility condition.

The wavelet function satisfies similar equation,

  2 '
' :2 2  2, n . . . (3.8)

; :2  0 2& :2   1= [ 2 + 1]

=

With the conditions.

Where, h[n] and g[n] can be viewed as the coefficients of low pass and
high pass filters. For a function , the wavelet coefficients < , ? ,  >
describes the information loss when going from projection of xt onto the space
Vj+1, to the projection onto the lower resolution space Vj. With MRA, any function
xt L2 described by the equation (3.3) can be modified by using both scaling
function and wavelet function as:
60

  B
A
C ' @ ,  . A, ,  + B C '  ,   A ,  
' A '
. . . (3.9)

Where, C j, k are scaling coefficients j is an arbitrary starting scale for coarsest
resolution, and G is an arbitrary finite upper limit for upper limit for high
resolutionj > jG . In practice, the selection of G and j depends on the
characteristics of the signal , range of the resolution required and the sampling
rate of the signal.

3.2.2. Filter-bank Implementation

If C j, k and  ,  are the scaling coefficients (approximations) and

wavelet coefficients (details) of the projection of the signal x onto Vj and Wj
respectively, the successive lower resolution coefficients are then recursively
derived based on equations (3.7) and (3.8) with MRA concept as:

@  + 1,   = 2  2@ , 2 3.10 )

D ( j + 1, k) = = :  2  2  , 2 3.10 b)

These equations can be implemented as a tree structured filter bank shown in

fig 3.3. Because of the orthonormal wavelet basis, this analysis filter bank also
satisfies the synthesis of high resolution scaling coefficients from the next
immediate level lower resolution scaling and wavelet coefficients as,

Cx
h(n) h (n)

x
x
Dx
g(n) g (n)

Decomposition Reconstruction

Figure 3.3 Decomposition and Reconstruction of signal using DWT

61
@ ,   =   22@  + 1, 2 + = :   22  + 1, 2 . . (3.11)

For a sampled signal, the filter bank tree is also viewed as an

implementation of 1-D DWT with initial maximum resolution component,
C (j=0, k) and its decomposition into number of details D j, n at successive
resolution scales.

For the standard DWT, the size of approximate coefficients and detail
coefficients decreases by a factor of 2 at each successive decomposition level. Thus
the standard DWT is perfectly non-redundant of O(n) representation of a given
signal in multi-resolution environment. The sparse representation with energy
compaction makes the standard DWT widely accepted for signal compression.

The reconstruction filter bank structure shown in figure (3.3) follows the
recursive synthesis similar to equation 3.11 with reconstruction filters h(n) and
g(n), which are identical to their corresponding decomposition filters h(n) and g(n)
but with time reversal.

The most important criterion with filter bank implementation of DWT is the
proper retrieval of signal, which is commonly termed as perfect reconstruction. The
perfect reconstruction imposes certain constraints on analysis and synthesis filters.
The nature of constraints relates these filters to either the orthogonal wavelet bases
or to the Bi-orthogonal wavelet basis.

3.2.3. Perfect Reconstruction

When reconstructed signal x(t) is identical to the orthogonal signal x(t) for a
simple 2 channel filter-bank structure, as shown in fig. 3.3, the associated
decomposition and reconstruction filters satisfy certain conditions.

H(z) H(z) + G(z) G(z) = 2

H(z) H(-z) + G(z) G(-z) = 0 (3.12)

where, H(z) and H(z)) are the Z transforms of h[n] and h(n) respectively.

62
 Most of the orthonormal wavelet basis associated with Perfect
Reconstruction (PR) filter-bank of fig. 3.3 has prototype wavelet with infinite
support (length). Hence all the filters require infinite taps. If the FIR filter-bank is
iterated on the low-pass channel, the overall impulse response of the iterated filter-
tree takes the form of continuous time function with compact support. With infinite
iterations over filter-tree, the impulse response converges to a smooth function
(mother wavelet). Filter having this property are called regularity. A necessary
condition for regularity is for low-pass filter to have at least one zero at the aliasing
frequency = . The number of zeroes at = determines the degree of
smoothness or differentiability of the resulting wavelet. regularity (smoothness) is
an important feature of wavelet for its application in detection of discontinuities.

For an orthogonal wavelet system, the conditions for analysis and synthesis
filters are given as,

h[n] = h[-n] g[n] = g[-n]

= 2I 2 + 2   J  3.13 

= :2 : I 2 + 2   J  3.13 

By allowing non-orthogonal and dual basis, a Bi-orthogonal wavelet system

is formed. Bi-orthogonal wavelet basis have the advantages of linear phase, and
more degrees of freedom in filter design. If the scaling function set{. , ,
.I , }, and the wavelet function set { , , I , } represent the dual basis

for decomposition (analysis) and reconstruction (synthesis) for Bi-orthogonal

system, then the 2 scale equations similar to the equations (3.8) and (3.9) are
given as,

.  2 '
' 2  . 2  2 3.14 

't  2 '
-'
h'n 2t-n (3.14 b)

'
  2 ; 22  2 3.14 P
'

63
 '
IQ  2 ; I 22  2  3.14&
'

Bi-orthogonal wavelet basis also satisfy the relation,

< , , I R,=  > J[  S]J[  2] 3.15

and reconstruction formula becomes,

    ; ; < , ,   > I ,   3.16

3.2.4 Two Dimensional Discrete Wavelet Transform (2D - DWT)

The DWT is extensively used in its non-redundant form known as standard

DWT. The filter bank implementation of standard DWT for images is viewed as
2D-DWT. Filter-bank structure discussed in section 3.1.2 is the simple
implementation of 1-D DWT, whereas image-processing applications requires two-
dimensional implementation of wavelet transform. Implementation of 2D-DWT is
also referred to as multidimensional wavelet transform.

Restore 2- g
D image HH
and form1-
D column
g sequence HL
h
2-D Form
image 1-D
sequenc Restore 2- g
h D image LH
and form1-
D column
sequence h LL

Can be
iterated
further. . .

Figure 3.4 Single level analysis filter-bank for 2-D DWT

64
 The implementation of an analysis filter-bank for a single level 2-D DWT is
shown in figure 3.4. This structure produces three detailed sub-band images
corresponding to three different directional-orientations i.e. The horizontal (HL),
the Vertical (LH) and the Diagonal (HH) and lower resolution sub-band image LL.
The filter-bank structure can be iterated in a similar manner on the LL channel to
provide multilevel decomposition. Multilevel decomposition hierarchy of an image
is illustrated in figure. 3.6.

Each decomposition breaks the parent image into four child images. Each of
such sub-images is one fourth of the size of a parent image. The sub-images are
placed according to the position of each sub-band in the two-dimensional partition
of frequency plane as shown in figure 3.7. The structure of synthesis filter-bank
follows the reverse implementation of analysis filter-bank but with the synthesis
filters.

LL HL

LH HH

Figure 3.5 2D DWT coefficients of image

Level 3 Decomposition
LLL Level 2

LLH
LLL Level 1

LHL LHH
LH
LL
Original
2-D image

HL
HH

Figure 3.6 Multilevel decomposition hierarchy of an image with 2-D DWT

65
 The separable wavelets are also viewed as tensor products of one-
dimensional wavelets and scaling functions. If (x) is the one-dimensional wavelet
associated with one-dimensional scaling function (x), then three 2-D wavelets
associated with three sub-images are given as,

. T  . U 3.17 

.W  .U 3.17 

. X  U 3.17 P

Figure 3.7 Frequency plane partitioning with 2-D DWT

3.2.5 Properties of the Discrete Wavelet Transform

The wavelet transform decomposes an image into three spatial directions, i.e
horizontal, vertical and diagonal. At each level of decomposition the magnitude of
the DWT coefficients is larger in the approximation sub-band LL, and smaller for
other detail sub-bands: HL, LH and HH. The most significant coefficients in a sub-
band are with large magnitudes. For an arbitrary image the high resolution sub-
bands help in locating the edge and texture patterns.

66
Advantages of DWT are much better than the DCT:

1. Wavelet coded image can be shown at different levels of resolution and can
be sequentially processed from lower resolution to higher resolution.

2. DWT splits the signal into individual bands, which can be processed
independently. Hence it is closer to the human visual system than the DCT.

3. The distortions introduced by wavelet domain coding with high compression

ratio are less annoying than those introduced at the same bit rate by the DCT.
In the JPEG case, block-shaped distortions are clearly visible, since image
coding based on the DCT usually operates on independent 8x8 blocks.

4. Watermarking schemes put more watermark energy into the large DWT
coefficients, its mostly effect on regions of lines and texture .The human
visual system is not sensitive to recognize the embedded watermark.

5. DWT has spatial frequency locality, which means if the watermark is

embedded into the DWT coefficients it will affect the image locally. Hence a
wavelet transforms provides both frequency and spatial description for an
image.

3.2.6 Applications of Discrete Wavelet Transforms

Wavelets are a powerful statistical tools which can be used for a wide range
of applications namely, watermarking, Data compression, Smoothing and image de-
noising, Fingerprint verification, EEG analysis, Speech recognition, Computer
graphics and multi fractal analysis etc.,

3.2.7 Limitations of Discrete Wavelet Transforms

Though the standard DWT is a powerful tool, it has three major limitations
that are described below particularly for signal and image processing tasks.

1. Shift Sensitivity

If shifting of input signal in time domain causes an unpredictable change in

transforms coefficients said to shift variant. It has been observed that the standard
DWT is seriously disadvantaged by the shift sensitivity that arises from down
67
samplers in the DWT implementation. Shift sensitivity is an undesirable property
because it implies that DWT coefficients fail to distinguish between input signal
shifts. The down samplers are responsible for shift sensitivity.

2. Poor Directionality

An m- Dimensional transform (m>1) suffers poor directionality when the

transform coefficients reveal only a few feature orientations in the spatial domain.
The separable 2-D DWT partitions are in the frequency domain into three
directional sub-bands. 2-D DWT can resolve only three spatial-domain feature
orientations: HL, LH and HH. Natural images contain number of smooth regions
and edges with random orientations; hence poor directionality affects the optimal
representation of natural images with of the separable standard 2-D DWT as shown
in figure 3.8.

Figure 3.8 Directionality of standard 2D DWT

3. Absence of Phase Information

For a complex valued signal or vector, its phase can be computed by its real
and imaginary projections. Digital image is a data matrix with a finite support in 2-
D. Filtering the image with 2D- DWT increases its size and adds phase distortion.
Human visual system is sensitive to phase distortion. Furthermore, Linear phase
filtering can use symmetric extension methods to avoid the problem of increased
data size in image processing. Phase information is valuable in many signal
processing applications such as e.g. in image compression and power measurement.

68
 Most DWT implementations use separable filtering with real coefficient
filters associated with real wavelets resulting in real-valued approximations and
details. Such DWT implementations cannot provide the local phase information. All
natural signals are basically real-valued, hence to avail the local phase information,
complex-valued filtering is essential to avail local phase.

3.3. HYPERANALYTIC WAVELET TRANSFORM

The success of wavelet techniques in many fields of signal and image

processing was proved to be highly influenced by the properties of wavelet
transforms used, mainly the shift invariance and the directional selectivity.
Unfortunately, the 2D-DWT is shift-variant and has a reduced directional
selectivity. These disadvantages can be attenuated if a complex wavelet transform is
used. The Hyper analytic Wavelet Transform (HWT) with simple implementation is
introduced here. It is quasi shift-invariant; it has a good directional selectivity, and a
reduced degree of redundancy.

3.3. 1 Hyperanalytic Wavelet Transform Implementation

The entire one dimensional wavelet transforms are simpler or more

complicated for 2D generalizations. The generalization of the analyticity concept in
2D is not obvious, because there are multiple definitions of the Hilbert transform in
this case. In the following, the definition of analytic signal, associated to a real 2D
real signal named hyper complex signal, is used. So, the hyper complex of hyper
analytic mother wavelet associated to the real wavelet x, y is defined as

Y , U  , U + ? Z {, U} +  Z[ {, U} +  Z \ Z[ {, U}]

(3.18)

^_`_, ?         ?  1, ?  ?  ,     ? 2&  ?  

69
The Hyper analytic Wavelet Transform (HWT) of an image f(x, y) is

Zab{c , U}  < c, U, , U > +< c, U, ?Z {, U} > +

< c, U, Z[ {, U} > +< c, U, Z \Z[ {, U}] > 

< c, U, , U > +? < Z { c, U}, , U > +

< Z[ { c, U}, , U > + < Z \Z[ {, U}] , , U

> Zab{ c , U}

 ab{c , U} + ? ab dZ {c, U}e + ab \Z[ {c, U}]

+ ab fZ \Z[ {c, U}]g

 < cY , U, , U >  ab { cY , U} 3.19

From the equation 3.19, it is obvious that, HWT of the image fx, y can be
computed with the aid of two dimensional- Discrete Wavelet Transforms of its
associated hyper complex image. In consequence the HWT implementation uses
four trees, each one implementing a 2D-DWT, thus having a redundancy of four.
The first tree is applied to the input image. The second and the third trees are
applied to one dimensional Hilbert transforms computed across the lines (Hx) or
columns (Hy) of the input image. The fourth tree is applied to the result obtained
after the computation of the two 1D Hilbert transforms of the input image. The
HWT implementation is presented in figure 3.9.

In the reconstruction, a possible method would be to reverse the Hilbert

transform applied at the input for each of the three trees mentioned above, and the
final result to be the mean of the four images computed on the four trees.

The reconstruction of original image from the Hyper analytic Wavelet

Transformed Image can be constructed by doing reverse process of construction of
HWT. The reverse of first tree shown in the Fig 3.9 is accomplished by 2D Inverse
Discrete Wavelet Transform (2D-IDWT) of real part, for the second the tree, the
70
reverse process is, 2D- IDWT of imaginary ith part, followed by inverse HX (inverse
Hilbert transform along rows), similar technique can be applied to the remaining
trees.

f (x,y)
2D DWT
+
Z r+
-
Hx {f} 2D DWT
+
Z i+
+
+ Z i-
-
Hy {f} 2D DWT +

+ Z r-
Hx {f}

Hy {Hx {f}} 2D DWT

Figure 3.9 Implementation of HWT

3.3.2. Directional Selectivity

The enhancement of the directional selectivity of the HWT is made through

linear combinations of detail coefficients belonging to each sub band of each of
the four 2D DWTs as shown in the Fig. 3.9.

From Fig 3.9,

1jk  &_?lm Pn_cc?_2m ab {c, U}  ab fZ \Z[ {c, U}]g

1j  &_?lm Pn_cc?_2m ab {c, U} + ab fZ \Z[ {c, U}]g

1ok  &_?lm Pn_cc?_2m ab{Z {c , U}} + ab{Z[ {c, U}}

1o  &_?lm Pn_cc?_2m pabdZ {c, U}e  ab \Z[ {c, U}]q

. . . (3.20)
71
 1j  1jk +  1j 1o  1ok +  1o . . . (3.21)

The coefficients of 1jk and 1ok are oriented only in the positive directions
G r
such as, b2 G  , s
and b2 G 2 for horizontal, diagonal and vertical details

respectively, and The coefficients of 1j an& 1o are oriented only in the negative
G r
directions such as, b2 G  , s , and b2 G 2 for horizontal, diagonal and

vertical details respectively. The preferred directions of HWT are summarized in the
table 3.1.
From the table 3.1, it is obvious that that 1jk and 1ok rejects the negative
directions and 1j an& 1o rejects the positive directions.

Table 3.1 Preferred directions of HWT

Coefficient Sub-bands tuk tu tvk tv

1 1 1 1
LH (horizontal) b2 G   b2G   b2 G   b2G  
2 2 2 2

HL (vertical) b2 G 2 b2G 2 b2 G 2 b2G 2

w w w w
 
4 4 4 4
HH (diagonal)

3.3.3. Applications of Hyper analytic Wavelet Transform (HWT)

As HWT is only two dimensional wavelet transform, it is only applicable for

images only, not for signals. In the case of images, the increased directional
selectivity in addition to shift-invariance leads to better performances especially
when it comes to images containing significant diagonal edges. Due directional
selectivity feature of HWT is used in applications like texture analysis and
synthesis, watermarking, image segmentation and classification etc.,

72
3.4 CHAPTER SUMMARY

In this chapter theoretical back ground of mathematical tools are used in the
proposed work for digital image watermarking such as DWT and HWT are
explained. The mathematical tools explained in this chapter are used to develop
watermark embedding and watermark extraction algorithms and details of which are
discussed in following chapters.

73