A Tutorial of the Wavelet Transform

Chun-Lin, Liu

February 23, 2010

Chapter 1

Overview

1.1 Introduction
The Fourier transform is an useful tool to analyze the frequency components
of the signal. However, if we take the Fourier transform over the whole time
axis, we cannot tell at what instant a particular frequency rises. Short-time
Fourier transform (STFT) uses a sliding window to find spectrogram, which
gives the information of both time and frequency. But still another problem
exists: The length of window limits the resolution in frequency. Wavelet
transform seems to be a solution to the problem above. Wavelet transforms
are based on small wavelets with limited duration. The translated-version
wavelets locate where we concern. Whereas the scaled-version wavelets allow
us to analyze the signal in different scale.

1.2 History
The first literature that relates to the wavelet transform is Haar wavelet. It
was proposed by the mathematician Alfrd Haar in 1909. However, the con-
cept of the wavelet did not exist at that time. Until 1981, the concept was
proposed by the geophysicist Jean Morlet. Afterward, Morlet and the physi-
cist Alex Grossman invented the term wavelet in 1984. Before 1985, Haar
wavelet was the only orthogonal wavelet people know. A lot of researchers
even thought that there was no orthogonal wavelet except Haar wavelet. For-
tunately, the mathematician Yves Meyer constructed the second orthogonal
wavelet called Meyer wavelet in 1985. As more and more scholars joined in
this field, the 1st international conference was held in France in 1987.

1
 In 1988, Stephane Mallat and Meyer proposed the concept of multireso-
lution. In the same year, Ingrid Daubechies found a systematical method to
construct the compact support orthogonal wavelet. In 1989, Mallat proposed
the fast wavelet transform. With the appearance of this fast algorithm, the
wavelet transform had numerous applications in the signal processing field.
Summarize the history. We have the following table:

ˆ 1910, Haar families.

ˆ 1981, Morlet, wavelet concept.

ˆ 1984, Morlet and Grossman, ”wavelet”.

ˆ 1985, Meyer, ”orthogonal wavelet”.

ˆ 1987, International conference in France.

ˆ 1988, Mallat and Meyer, multiresolution.

ˆ 1988, Daubechies, compact support orthogonal wavelet.

ˆ 1989, Mallat, fast wavelet transform.

2
Chapter 2

Approximation Theory and

Multiresolution Analysis

2.1 A Simple Approximation Example

Consider a periodic function x(t) with period T

2|t|
x(t) = 1 − , |t| < T. (2.1)
T
Find its Fourier series coefficient.

Recall the relationship of Fourier series expansion. We can decompose a

periodic signal with period T into the superposition of its high order har-
monics, which is the synthesis equation
∞
X 2πkt
x(t) = ak exp(j ). (2.2)
k=−∞
T
According to the orthogonal property of complex exponential function,
we have the analysis equation
1 T /2
Z
2πkt
ak = x(t) exp(−j )dt. (2.3)
T −T /2 T
In this example the Fourier series coefficients are
sin2 (πk/2) 1
ak = 2
= sinc2 (k/2). (2.4)
2(πk/2) 2
What’s the physical meaning of the result? The Fourier series coefficient
indicates the amplitude and phase of the high order harmonics, indexing by

3
Figure 2.1: An illustration of the approximation example. The red line is
the original signal. The blue dash line is the approximated signal with (a)
K = 0 (b) K = 1 (c) K = 2 (d) K = 3 (e) K = 4 (f ) K = 5

4
the variable k. The higher k is, the higher frequency refers to. In general, the
power of a signal concentrates more likely in the low frequency components.
If we truncate the Fourier series coefficient in the range of [−K, K] and set
ak = 0 outside this range, we expect the synthesis version of the truncated
coefficients be
K
X 2πkt
x̃K (t) = ak exp(j ). (2.5)
k=−K
T
Intuitively, as K approaches infinity, the reconstructed signal is close to the
original one. Refer to Fig. 2.1, we can see the approximation signal is very
close to the original one as k grows higher.

5
2.2 Abstract Idea in the Approximation Ex-
ample
From the point of linear algebra, we can decompose the signal into linear
combination of the basis if the signal is in the the space spanned by the
basis. In pp. 364-365 [1], it is,
X
f (t) = ak φk (t). (2.6)
k

where k is an integer index of the finite or infinite sum, the ak are expansion
coefficients, and the φk (t) are expansion functions, or the basis. Compare
to 2.2, we know that Fourier series expansion is a special case when φk (t) =
exp(j2πkt/T ).In general, if we choose the basis appropriately, there exists
another set of basis {φ̃k (t)} such that {φk (t)} and {φ̃k (t)} are orthonormal.
The inner product 1 is
Z
< φi (t), φ̃j (t) >= φi (t)φ̃∗j (t)dt = δij . (2.7)

where {φ̃k (t)} is called the dual function of {φk (t)}. With this orthonormal
property, we can find the coefficients by
Z
< f (t), φ̃k (t) > = f (t)φ̃∗k (t)dt
Z X
= ( ak0 φk0 (t))φ̃∗k (t)dt
k0
X Z
= ak0 ( φk0 (t)φ̃∗k (t)dt)
k0
X
= ak 0 δk 0 k
k0
= ak .

Rewrite as follows:
Z
ak =< f (t), φ̃k (t) >= f (t)φ̃∗k (t)dt. (2.8)

The satisfying result comes because of the orthonormal property of the basis.
Review 2.3, the Fourier series analysis equation is a special case when φ̃k (t) =
1 ∗
R
If x(t) and y(t) are both complex signals, < x(t), y(t)
R >= x(t)y (t)dt. If they are
both real-valued, the inner product can be reduced to (x(t)y(t)dt. The latter is more
well-known.

6
(1/T ) exp(j2πkt/T ). Therefore, it is important to choose an appropriate set
of basis and its dual. For the signal we want to deal with, apply a particular
basis satisfying the orthonormal property on that signal. It is easy to find
the expansion coefficients ak . Fortunately, the coefficients concentrate on
some critical values, while others are close to zero. We can drop the small
coefficients and use more bits to record the important values. This process
can be applied to data compression while preserving the resemblance to the
original signal at the same time. In the example in the previous section,
we can approximate the original by only some critical coefficients. Data
compression can be achieved.

7
2.3 Two Examples about Multiresolutions
2.3.1 Example 1: Approximate discrete-time signals
using delta function
Consider a discrete-time signal x[n] defined as
1
x[n] = ( )|n| . (2.9)
2
Now we choose a basis {φk [n]} = {δ[n − k]} and its dual {φ̃k [n]} = {φk [n]} =
{δ[n−k]} to find the expansion coefficients of x[n]. Note that for the discrete
case, the variable is changed to n and use brackets instead of parentheses,
and the inner product is revised as summations instead of integrations. But
the concept of function expansion and inner product is the same. First, we
check whether the basis satisfies the orthonormal property:
∞
X
< φi [n], φ̃j [n] >=< δ[n − i], δ[n − j] >= δ[n − i]δ[n − j] = δij . (2.10)
n=−∞

By 2.8, we can find the expansion coefficients are

∞
X 1 1
ak =< x[n], δ[n − k] >= ( )|n| δ[n − k] = ( )|k| . (2.11)
n=−∞
2 2

Likewise, if we reserve the coefficients in the range of index k ∈ [K1 , K2 ],

obtained approximations are shown in Fig.2.2.
This example selects a particular basis {δ[n − k]} to find its coefficients.
This basis has a different interpretation: It can be viewed as translations of
a single delta function δ[n]. δ[n − k] means the position of the impulse is
located at n = k. The reconstruction signal using partial coefficients means
the position we concern. For example, if we want to analyze the signal when
n ∈ [−2, 0], we can use a−2 , a−1 and a0 to find the reconstruction version.

2.3.2 Example 2: Basis Construction by Scaling

In the previous example, we construct our basis by translating a single delta
function. Here, we use another common operation, scaling, to see its results.
Consider a continuous function φ(t) such that
(
1 0 ≤ t < 1,
φ(t) = (2.12)
0 otherwise.

8
Figure 2.2: The result of approximation using delta functions. (a) Original
signal
P0 x[n] (b) Obtained coefficients ak (c) ReconstructedP2 signal x1 [n] =
k=−2 ak δ[n − k] (d) Reconstructed signal x2 [n] = k=0 ak δ[n − k]

9
Figure 2.3: Plot of the scaled version of the basis.(a) φn (t) = φ0 (t) (b)
φn (t) = φ1 (t) (c) φn (t) = φ−1 (t) (d) φn (t) = φ2 (t).

It is a rectangular function centered at x = 1/2 with width 1. The scaled

version is defined as φs (t) such that

φs (t) = φ(st). (2.13)

Note that s is a continuous scaling factor. As s higher, the support of

this scaled basis will become narrower. For the discrete basis we discussed
above, we let s = 2n just for simplicity, where n is an integer. Therefore, we
rewrite the functions as
φn (t) = φ(2n t). (2.14)
For consistency, this tutorial use index n to indicate the scaling operation,
defined as above. Fig. 2.3 shows the scaling operation. Note that once the
support of the function becomes narrower, the higher frequency it has. We
can approximate the high frequency component more precisely by the higher
order scaling.

10
 Here comes another question: is the family {φn (t)} a set of orthogonal
family? Check the first two functions, φ0 (t) and φ1 (t)
1
Z ∞ Z
2 1
< φ0 (t), φ1 (t) >= φ0 (t)φ1 (t)dt = dt = .
−∞ 0 2

Unfortunately, the two functions are not orthogonal to each other, so as the
others. But don’t be frustrated here, we can apply Gram-Schmidt process to
obtain a set of orthonormal basis from existing {φn (t)}. The procedure is

φ00 (t) = φ0 (t) = φ(t)

< φ1 (t), φ0 (t) >
φ01 (t) = φ1 (t) − φ0 (t)
< φ0 (t), φ0 (t) >

1/2
 0 ≤ t < 1/2,
= −1/2 1/2 ≤ t < 1,

0 otherwise.


= ψ(t)/2

We can continuously apply this process to extend the basis. But we look
at the first two basis and its Fourier transform. By inspection, φ(t) has
a nonzero mean while ψ(t) is zero in average. The φ(t) has two jumps at
x = 0, 1 while ψ(t) jumps at x = 0, 1/2, 1. Therefore, the power of φ(t)
is more compact at low frequencies while the power of ψ(t) concentrates at
relatively high frequencies. In formal φ(t) is called scaling function to do
approximation and ψ(t) is called wavelet function to find the details.

11
2.4 First Look at Multiresoltuion Analysis
In the translating example, we found a function which is orthonormal to
its translating version. Different scaled version can see different frequency
resolutions. Combined with the two properties, we can construct a basis from
the scaling function and wavelet function with two parameters: scaling and
translating, formally defined in pp. 365-372 [1] as

φj,k (t) = 2j/2 φ(2j t − k). (2.15)

ψj,k (t) = 2j/2 ψ(2j t − k). (2.16)

where j is the parameter about dilation, or the visibility in frequency and
k is the parameter about the position. In practice, we may want to see the
whole data with ”desired” resolution, i.e. for some resolution j. We define
the subspace
Vj = Span{φj,k (t)}. (2.17)

Wj = Span{ψj,k (t)}. (2.18)

With these formal definition, some requirements for multiresolution anal-
ysis are

1. The scaling function is orthogonal to its integer translates: In the ex-

ample, the scaling function has values only in [0, 1]. The integer trans-
lates of the scaling function does not overlap with the original scaling
function.

2. The subspaces spanned by the scaling function at low scales are nested
within those spanned at higher scales: From Fig.2.3, it is clear to show
that φ−1 (t) = φ0 (t) + φ0 (t − 1). In the notations defined above, it’s
Vj ⊂ Vj+1 .

3. The only function that is common to all Vj is f (x) = 0.

4. Any function can be represented with arbitrary precision.

Requirement 2 is worthy to note. For Haar scaling function (in the ex-
ample), it is
1 1
φ(t) = φ0,0 (t) = √ φ1,0 (t) + √ φ1,1 (t).
2 2

12
 Figure 2.4: The relationship between scaling and wavelet function spaces.

Expand the notation of φj,k (t), we have

1 √ 1 √
φ(t) = √ ( 2φ(2t)) + √ ( 2φ(2t − 1)). (2.19)
2 2
In general, this equation is called refinement equation, multiresolution anal-
ysis equation, or dilation equation
X √
φ(t) = hφ [n] 2φ(2t − n). (2.20)
n
√ √
It is evident to show that hφ [n] = {1/ 2, 1/ 2} for Haar scaling functions.
A physical meaning of 2.20 is that φ(2t) is a function with higher frequency
components than φ(t). It is designed properly such that we apply a discrete
low pass filter hφ [n] to have φ(t).
Similar relationship exists for the wavelet functions. It is
X √
ψ(t) = hψ [n] 2φ(2t − n). (2.21)
n
√ √
Again, for Haar wavelets, hψ [n] = {1/ 2, −1/ 2}. These two filters are
related by
hψ [n] = (−1)n hφ [1 − n]. (2.22)
Finally, a diagram shows the relationship between the sets and the func-
tions. V0 is the approximation in scale 0. We can union the set spanned by
different levels of wavelet functions to have higher order of approximation.

13
 Figure 2.5: The relationship between scaling and wavelet functions.

Once we union infinite wavelet sets, the sets are equal to the L2 (R) set2 .
This is requirement 4. In mathematical words, it is

L2 (R) = V0 ⊕ W0 ⊕ W1 ⊕ .... (2.23)

Therefore, we can decompose any function in L2 (R), which is almost in

all cases, using the scaling function and wavelet functions. An analogy of
this concept is the corss-section of a cabbage. A cabbage is made of a kernel
and a lot of leaves. A cabbage on the market can be considered as a scaling
function at the finest scale J. As we take a layer of leaves off, the remaining
cabbage is like the original cabbage in shape, but different in size. This is the
concept of the scaling function at different scales with different supports but
with similar shape. The leaves of a layer are the differences between different
scales as the wavelet function. It can be used in cookery, or in mathematical
analysis. We can peel the leaves until the leaves are all taken off just like the
multiresolution analysis.

2
L2 (R) = {f (x)| |f (x)|2 dx < ∞}
R

14
Chapter 3

Discrete Wavelet Transform

3.1 Definition
In the previous chapter, we discuss the scaling function, wavelet function,
and their properties. Suppose the scaling function and wavelet function are
given as Haar, Daubechies,... Say, the basis are known. We can approximate
a discrete signal in l2 (Z)1 by
∞
1 X 1 XX
f [n] = √ Wφ [j0 , k]φj0 ,k [n] + √ Wψ [j, k]ψj,k [n]. (3.1)
M k M j=j0 k

Here f [n], φj0 ,k [n]andψj,k [n] are discrete functions defined in [0, M − 1], to-
tally M points. Because the sets {φj0 ,k [n]}k∈Z and {ψj,k [n]}(j,k)∈Z2 ,j≥j0 are
orthogonal to each other. We can simply take the inner product to obtain
the wavelet coefficients
1 X
Wφ [j0 , k] = √ f [n]φj0 ,k [n]. (3.2)
M n

1 X
Wψ [j, k] = √ f [n]ψj,k [n] j ≥ j0 . (3.3)
M n
3.2 are called approximation coefficients while 3.3 are called detailed co-
efficients.

1 2
P∞
l (Z) = {f [n]| n=−∞ |f [n]|2 < ∞}

15
3.2 The Fast Wavelet Transform
Start from the definition, if the form of scaling and wavelet function is known,
its coefficients is defined in 3.2 and 3.3. If we can find another way to find
the coefficients without knowing the scaling and dilation version of scaling
and wavelet function. The computation time can be reduced. From 2.20, we
have

φj,k [n] = 2j/2 φ[2j n − k]

X √
= hφ [n0 ] 2φ[2(2j n − k) − n0 ]. (3.4)
n0

Let n0 = m − 2k, we have

X √
φj,k [n] = hφ [m − 2k] 2φ[2j+1 n − m]. (3.5)
m

If we combine the equation above with 3.2, it becomes

1 X
Wφ [j, k] = √ f [n]φj,k [n]
M n
1 X
= √ f [n]2j/2 φ[2j n − k]
M n
1 X X √
= √ f [n]2j/2 hφ [m − 2k] 2φ[2j+1 n − m]
M n m
X 1 X
= hφ [m − 2k]( √ f [n]2(j+1)/2 φ[2j+1 n − m])
m M n
X
= hφ [m − 2k]Wφ [j + 1, m] (3.6)
m

= hφ [−n] ∗ Wφ [j + 1, n] . (3.7)

n=2k,k≥0

Similarly, for the detail coefficients, it is


Wψ [j, k] = hψ [−n] ∗ Wφ [j + 1, n] . (3.8)

n=2k,k≥0

For the commonly used discrete signal, say, a digital image, the original
data can be viewed as approximation coefficients with order J. That is,
f [n] = Wφ [J, n] By 3.7 and 3.8, next level of approximation and detail can
be obtained. This algorithm is ”fast” because one can find the coefficients
level by level rather than directly using 3.2 and 3.3 to find the coefficients.
This algorithm was first proposed in [2].

16
Figure 3.1: The schematic diagram to realize discrete wavelet transform.
Here the filter names are changed.

3.3 Subband Coding

The satisfying result in 3.7 and 3.8 is much simpler to implement than 3.2 and
3.3. Convolve Wφ [j, n] with different filters, hφ [−n] and hψ [−n]. Downsample
by the factor of 2. We will find the next-level coefficients. Multiresolution
analysis can be achieved by cascading the structure above.
A schematic diagram of this algorithm is shown in Fig. 3.1. The synthesis
process does the inverse operation of the analysis process. First upsample
by a factor of 2 and then convolve the signal with a known inverse filter. We
expect the reconstructed signal, x0 [n], to be exactly the same as the original
signal, x[n]. Analyze each node in the subband coding schematic diagram in
both time domain and z-domain, and we have
1. Time domain: x(1) [x] = x[n],
z-domain: X(1) (z) = X(z),
2. Through a LTI system with impulse response h[n],
Time domain: x(2) [n] = x[n] ∗ h[n].
z-domain: X(2) (z) = X(z)H(z).
3. Time domain: x(3) [n] = x[n] ∗ g[n].
z-domain: X(3) (z) = X(z)G(z).
4. Time domain: x(4) [n] = x(2) [2n] = {..., x(2) [−4], x(2) [−2], x(2) [0], x(2) [2], x(2) [4], ...}.
z-domain:
By definition of X(2) (z), we have

X(2) (z) = ...+x(2) [−2]z 2 +x(2) [−1]z 1 +x(2) [0]+x(2) [1]z −1 +x(2) [2]z −2 +....

Let z = −z, we have

X(2) (−z) = ...+x(2) [−2]z 2 −x(2) [−1]z 1 +x(2) [0]−x(2) [1]z −1 +x(2) [2]z −2 +....

17
 Add the two equations above

X(2) (z)+X(2) (−z) = 2(...+x(2) [−4]z 4 +x(2) [−2]z 2 +x(2) [0]+x(2) [2]z −2 x(2) [4]z −4 +...).
1 1
X(2) (z 2 )+X(2) (−z 2 ) = 2(...+x(2) [−4]z 2 +x(2) [−2]z 1 +x(2) [0]+x(2) [2]z −1 x(2) [4]z −2 +...).
so 1 1
X(2) (z 2 ) + X(2) (−z 2 )
X(4) (z) = .
2
5. Time domain: x(5) [n] = x(3) [2n]
z-domain: 1 1
X(3) (z 2 ) + X(3) (−z 2 )
X(5) (z) = .
2
6. Time domain: x(6) [n] = x(4) [n/2] for n=multiple of 2, zero for others.
z-domain:
∞ ∞ ∞
X X m=n/2 X
X(6) (z) = x(6) [n]z −n = x(4) [n/2]z −n = x(4) [m]z −2m = X(4) (z 2 ).
n=−∞ n=−∞ m=−∞

7. Similarly, the z-domain representation at node 7 is

X(7) (z) = X(5) (z 2 ).

8. Convolve x(6) [n] with a LTI filter with impulse response h1 [n]. The
z-domain representation is

X(8) (z) = H1 (z)X(6) (z).

9. z-domain function at node 9 is found to be

X(9) (z) = G1 (z)X(7) (z).

10. The reconstructed signal x0 [n] is the sum of node 8 and 9. Finally, the
z-domain representation is

X0 (z) = X(8) (z) + X(9) (z)

1
= (H(z)H1 (z) + G(z)G1 (z))X(z)
2
1
+ (H(−z)H1 (z) + G(−z)G1 (z))X(−z). (3.9)
2

18
 3.9 gives us some constraint to make the reconstructed signal lossless.
For any signal, we let X(z) = X0 (z). The following two conditions can be
derived under the perfectly reconstruction assumption,
H(z)H1 (z) + G(z)G1 (z) = 2 (3.10)
H(−z)H1 (z) + G(−z)G1 (z) = 0. (3.11)
Rewrite 3.10 and 3.11 into matrix form, we obtain
" #" # " #
H(z) G(z) H1 (z) 2
= . (3.12)
H(−z) G(−z) G1 (z) 0
Solve the linear system, the relationship between the decomposition and
reconstruction filters is
" # " #
H1 (z) 2 G(−z)
= (3.13)
G1 (z) det(Hm ) −H(−z)
where " #
H(z) G(z)
Hm = . (3.14)
H(−z) G(−z)
3.13 shows once the decomposition filters is determined the correspond-
ing reconstruction filters are unique. For most case, Hm is invertible so
det(Hm ) 6= 0. Another sufficient and necessary condition of 3.13 is the
biorthogonal relationship. It consists of four equations:
H(z)H1 (z) + H(−z)H1 (−z) = 2 (3.15)
G(z)G1 (z) + G(−z)G1 (−z) = 2 (3.16)
H(z)G1 (z) + H(−z)G1 (−z) = 0 (3.17)
G(z)H1 (z) + G(−z)H1 (−z) = 0. (3.18)
Rewrite in the time domain, we will clearly see the biorthogonality:
< h[k], h1 [2n − k] > = δ[n] (3.19)
< g[k], g1 [2n − k] > = δ[n] (3.20)
< g1 [k], h[2n − k] > = 0 (3.21)
< h1 [k], g[2n − k] > = 0. (3.22)
The biorthogonality is an important guideline to design the subband coding
algorithm. A regular requirement is to make the four filter FIR filter for
simplicity. From 3.13, if we let det(Hm ) = constant or αz −(2k+1) , H1 (z) and
G1 (z) are in FIR form. The reconstruction filter can be discussed in three
cases:

19
 1. det(Hm ) = 2:

h1 [n] = (−1)n g[n] (3.23)

g1 [n] = (−1)n+1 h[n]. (3.24)

2. det(Hm ) = −2:

h1 [n] = (−1)n+1 g[n] (3.25)

g1 [n] = (−1)n h[n]. (3.26)

3. det(Hm ) = 2z −(2k+1) :

h1 [n] = (−1)n+1 g[n + 2k + 1] (3.27)

g1 [n] = (−1)n h[n + 2k + 1]. (3.28)

A wavelet interpretation of this result is that the set of filters, {h[n], g[n]},
corresponds to an arbitary set of scaling and wavelet functions, {φ(t), ψ(t)},
while another set of filters, {h1 [n], g1 [n]}, maps another set of wavelet func-
tions, {φ̃(t), ψ̃(t)}. These filters are tied with biorthogonal relations. This
result gives us more flexibility to design the filters.

20
3.4 2D Wavelet Transform
Recall the 2D case in the Fourier transform, the basis are modified into

exp(j(ω1 t1 + ω2 t2 )) (3.29)

instead of exp(jωt). The transformed coefficient becomes two variable func-

tions so as the 2D wavelet transform. In pp. 386-388 [1], the scaling and
wavelet function are two variable functions, denoted φ(x, y) and ψ(x, y) here.
The scaled and translated basis functions are defined as

φj,m,n (x, y) = 2j/2 φ(2j x − m, 2j y − n), (3.30)

i
ψj,m,n (x, y) = 2j/2 ψ i (2j x − m, 2j y − n), i = {H, V, D}. (3.31)

There are three different wavelet functions, ψ H (x, y), ψ V (x, y) and ψ D (x, y).
Conceptually, the scaling function is the low frequency component of the
previous scaling function in 2 dimensions. Therefore, there is one 2D scaling
function. However, the wavelet function is related to the order to apply the
filters. If the wavelet function is separable, i.e. f (x, y) = f1 (x)f2 (y). These
functions can be easily rewritten as

φ(x, y) = φ(x)φ(y), (3.32)

H
ψ (x, y) = ψ(x)φ(y), (3.33)
ψ V (x, y) = φ(x)ψ(y), (3.34)
ψ D (x, y) = ψ(x)ψ(y). (3.35)

If we define the functions as separable functions, it is easier to analyze

the 2D function and we can focus on the design of 1D wavelet and scaling
functions. The analysis and synthesis equations are modified to
M −1 N −1
1 X X
Wφ (j0 , m, n) = √ f (x, y)φj0 ,m,n (x, y), (3.36)
M N x=0 y=0
M −1 N −1
1 X X
Wψi (j, m, n) = √ f (x, y)φij,m,n (x, y), i = {H, V, D}(3.37)
M N x=0 y=0
1 XX
f (x, y) = √ Wφ (j0 , m, n)φj0 ,m,n (x, y)
MN m n
∞ XX
1 X X
+ √ Wψi (j, m, n)φij,m,n (x, y).
(3.38)
M N i=H,V,D j=j0 m n

21
Figure 3.2: Schematic diagram of 2D wavelet transform

Figure 3.3: Original Lena image

22
 Figure 3.4: Lena image after wavelet decomposition

This is the general form of 2D wavelet transform. If the scaling and

wavelet functions are separable, the summation can be decomposed into two
stages. First step is along the x-axis and then calculate along the y-axis. For
each axis, we can apply fast wavelet transform to accelerate the speed. A
schematic diagram is shown in Fig. 3.4. The two dimensional signal (usu-
ally image) is divided into four bands: LL(left-top), HL(right-top), LH(left-
bottom) and HH(right-bottom). The HL band indicated the variation along
the x-axis while the LH band shows the y-axis variation. Fig. 3.3, 3.4, and
3.5 show the decomposition of a image. The power is more compact in the LL
band. In the point of coding, we can spend more bits on the low frequency
band and less bit on the high frequency band or even set them to zero. A
famous algorithm, named Embedded Zerotree Wavelet (EZW) proposed by
Shapiro [3] and some modified versions in [4] and [5] are popular.

In addition to decompose the image in two axis, Yu-Si Zhang [6] intro-
duced a method to decompose the image along the natural edge of the image.
But in general, the structure of Fig.3.4 is used due to the implementation
complexity.

23
Figure 3.5: Reconstructed Lena image. The border is black due to the con-
volution.

24
Chapter 4

Continuous Wavelet Transform

4.1 Introduction
In this chapter, we talk about continuous wavelet transform. This transform
works when we use a continuous wavelet function to find the detailed coef-
ficients of a continuous signal. Like the concept in chapter 2, we have to
establish a basis to do such analysis. First, we give the definition of continu-
ous wavelet transform and do some comparison between that and the Fourier
transform.

25
4.2 Definition
We define a mother wavelet function ψ(t) ∈ L2 (R), which is limited in time
domain. That is, ψ(t) has values in a certain range and zeros elsewhere.
Another property of mother wavelet is zero-mean. The other property is
that the mother wavelet is normalized. Mathematically, they are
Z ∞
ψ(t)dt = 0 (4.1)
−∞
Z ∞
2
kψ(t)k = ψ(t)ψ ∗ (t)dt = 1. (4.2)
−∞
As the dilation and translation property states, the mother wavelet can
form a basis set denoted by

n 1 t − u o
ψs,u (t) = √ ψ( )  . (4.3)
s s u∈R,s∈R+

u is the translating parameter, indicating which region we concern. s is the

scaling parameter greater than zero because negative scaling is undefined.
The multiresolution property ensures the obtained set {ψu,s (t)} is orthonor-
mal. Conceptually, the continuous wavelet transform is the coefficient of the
basis ψu,s (t). It is
W f (s, u) = < f (t), ψs,u >
Z ∞
∗
= f (t)ψs,u (t)dt
−∞
Z ∞
1 t−u
= f (t) √ ψ ∗ ( )dt. (4.4)
−∞ s s
Via this transform, one can map an one-dimensional signal f (t) to a two-
dimensional coefficients W f (s, u). The two variables can perform the time-
frequency analysis. We can tell locate a particular frequency (parameter s)
at a certain time instant (parameter u).
If the f (t) is a L2 (R) function. The inverse wavelet transform is
Z ∞Z ∞
1 1 t−u ds
f (t) = W f (s, u) √ ψ( )du 2 , (4.5)
Cψ 0 −∞ s s s
where Cψ is defined as
∞
|Ψ(ω)|2
Z
Cψ = dω < ∞. (4.6)
0 ω
Ψ(ω) is the Fourier transform of the mother wavelet ψ(t). This equation is
also called the admissibility condition.

26
4.3 An example of wavelet transform
In the previous section, we assume the mother wavelet is given with some
satisfactory properties. Here we illustrate a famous mother wavelet function,
called Mexican hat wavelet with
2 t2 t2
ψ(t) = √ ( − 1) exp(− ). (4.7)
π 1/4 3σ σ 2 σ2

The complicated form is derived from the second derivative of a Gaus-

sian function, exp(−t2 /(2σ 2 )). The messy constant multiplied is for the
normalization property of mother wavelets. The nickname, ”Mexican-hat,”
is because the shape of the function is like an upside-down Mexican hat. Due
to the fast decay of the Gaussian function, this function drop to zero very
fast. We can expect it is time-limited in a certain range, say [−5, 5]. This
meets the property of mother wavelets.
The corresponding Fourier transform of the mother wavelet function is
√
− 8σ 5/2 π 1/4 2 σ2ω2
Ψ(ω) = √ ω exp(− ). (4.8)
3 2

It is still in a complicated form. The illustration of these two functions is

in Fig. 4.1. There is a Gaussian exponential term in its Fourier spectrum.
Therefore, we expect its spectrum is roughly band-limited in a certain range,
too. Note that the low frequency component is relatively small compared
to that at ω = ±1.5. This result corresponds to that in the multiresolution
theory: the wavelet function is a function with high frequency components.
As the perform scaling on the wavelet function, the mainlobe of its spectrum
will shift to higher frequency region to have finer frequency resolution.

27
Figure 4.1: Mexican-hat wavelet for σ = 1 and its Fourier transform.

28
4.4 Comparison Among the Fourier Trans-
form, Short-time Fourier Transform(STFT)
and Wavelet Transform
In this section, we talk about features among the three transforms.

4.4.1 Forward transform

Fourier transform
Z ∞
F (ω) = f (t) exp(−jωt)dt. (4.9)
−∞

As we know, Fourier transform convert signal in time domain to frequency

domain by integrating over the whole time axis. However, if the signal is
not stationary, that is, the frequency composition is a function of time, we
cannot tell when a certain frequency rises.

STFT
Z ∞
Sf (u, ξ) = f (t)w(t − u) exp(−jξt)dt. (4.10)
−∞

The STFT tries to solve the problem in Fourier transform by introducing a

sliding window w(t−u). The window is designed to extract a small portion of
the signal f (t) and then take Fourier transform. The transformed coefficient
has two independent parameters. One is the time parameter τ , indicating
the instant we concern. The other is the frequency parameter ξ, just like
that in the Fourier transform. However another problem rises. The very low
frequency component cannot be detected on the spectrum. It is the reason
that we use the window with fixed size. Suppose the window size is 1. If
there is a signal with frequency 0.1Hz, the extracted data in 1 second look
like flat (DC) in the time domain.

Wavelet transform
∞
t−u
Z
1
W f (s, u) = f (t) √ ψ ∗ ( )dt. (4.11)
−∞ s s
Wavelet transform overcomes the previous problem. The wavelet function
is designed to strike a balance between time domain (finite length) and fre-
quency domain (finite bandwidth). As we dilate and translate the mother
wavelet, we can see very low frequency components at large s while very high
frequency component can be located precisely at small s.

29
4.4.2 Inverse transform
Fourier transform
Z ∞
1
f (t) = F (ω) exp(jωt)dt. (4.12)
2π −∞

STFT
Z ∞ Z ∞
1
f (t) = Sf (u, ξ)w(t − u) exp(jξt)dξdu. (4.13)
2π −∞ −∞

Wavelet transform
∞ ∞
t−u
Z Z
1 1 ds
f (t) = W f (s, u) √ ψ( )du 2 , (4.14)
Cψ 0 −∞ s s s
Z ∞
|Ψ(ω)|2
Cψ = dω < ∞. (4.15)
0 ω

4.4.3 Basis
Here we discuss the different basis for each transforms. Fig. 4.4.2 shows the
results.

Fourier transform
Complex exponential function with different frequencies:

exp(jωt). (4.16)

STFT
Truncated or windowed complex exponential function:

w(t − u) exp(jξt). (4.17)

Wavelet transform
Scaled and translated version of mother wavelets:
1 t−u
√ ψ( ). (4.18)
s s

30
Figure 4.2: Different basis for the transforms. (a) Real part of the basis for
Fourier transform, exp(jπt). (b) Basis for different frequency, exp(j4πt). (c)
Basis for STFT, using Gaussian window of σ = 1. It is exp(−t2 /2) exp(jπt).
(d) Basis for different frequency, exp(−t2 /2) exp(j4πt). (e) Mexican-hat
mother wavelet function and (f ) s = 4.

31
Figure 4.3: Different time-frequency tile allocation of the three transforms:
(a) Fourier transform, (a) STFT and (a) wavelet transform.

32
4.4.4 Time-frequency tiling
In quantum physics, the Heisenberg uncertainty principle states that cer-
tain pairs of physical properties, like position and momentum, cannot both
be known to arbitrary precision. The same principle holds in signal pro-
cessing. We cannot locate both time and frequency very precisely. The
product of variation in time and variation in frequency is greater than 1/2,
i.e. σt σω ≥ 1/2. It can be viewed as a rectangle with constant area and
different transform adjusts the width and height of the rectangle. The three
transforms are illustrated in Fig. 4.4.4.

Fourier transform
The time information is completely lost. Frequency axis is divided uniformly.
Frequency resolution can be very precise if we integrate along the whole time
axis.

STFT
Add a window to take the time domain information into consideration. The
frequency resolution depends on the time resolution, or the size of the win-
dow. We cannot zoom in a particular frequency range because the box is
uniformly placed.

Wavelet transform
The s parameter is inversely proportional to the frequency. As we see, if we
want to focus on low frequencies, larger s is used while higher frequencies
uses small s. This flexibility increases the time-frequency analysis.

33
Chapter 5

Some Common Wavelet

Functions

In this chapter, we discuss how to design a mother wavelet. The design

equations determine the coefficients of the scaling/wavelet functions and the
corresponding lowpass/highpass filters. We will go through some popular
wavelet functions and illustrate how they are designed. Original lectures can
be found in [7].

5.1 Design Equations

5.1.1 Scaling equation
First, we examine (2.20) and (2.21), which can be considered as the basic
definition of scaling and wavelet functions. Setting t = t/2 and we have
1 t X
√ φ( ) = hφ [n]φ(t − n), (5.1)
2 2 n
1 t X
√ ψ( ) = hψ [n]φ(t − n). (5.2)
2 2 n

Taking the Fourier transform on both equations, we obtain

1
√ Φ(2ω) = Hφ (ω)Φ(ω), (5.3)
2
1
√ Ψ(2ω) = Hψ (ω)Φ(ω), (5.4)
2
where Φ(ω) and Ψ(ω) are the Fourier transform of the scaling and wavelet
functions, respectively. Hφ (ω) and Hψ (ω) are the discrete-time Fourier trans-

34
form of the discrete filters. If we substitute the recursive equation continu-
ously, we obtain
∞
Y Hφ (ω)
Φ(ω) = √ Φ(0). (5.5)
p=1
2
By (5.5), once we have the lowpass filter, we can calculate the correspond-
ing scaling function by its Fourier transform. If we know the scaling function
in advance, 5.3 helps us to find the lowpass filter.

5.1.2 Conjugate mirror filter

The scaling equation only gives us a connection between the scaling and the
corresponding filters. There is still another relationship holds for the filters.
It is
|Hφ (ω)|2 + |Hφ (ω + π)|2 = 2, (5.6)
and √
Hφ (0) = 2. (5.7)
The filters that satisfy this relationship is called conjugate mirror filters. Its
proof is technical. One can refer to pp. 271-276 [8] to see the details.
The similar equation for wavelet highpass filter is

|Hψ (ω)|2 + |Hψ (ω + π)|2 = 2, (5.8)

and
Hψ (ω)Hφ∗ (ω) + Hψ (ω + π)Hφ∗ (ω + π) = 0. (5.9)
These equations might not strange to you because similar equations appear
in (3.15), using subband coding approach. We usually set

Hψ (ω) = − exp(−jω)Hφ∗ (ω + π). (5.10)

It is a trivial solution to the two constraints. Calculating the inverse trans-

form yields
hψ [n] = (−1)n hφ [1 − n]. (5.11)

5.1.3 Real filters

For simplicity, we want the obtained filters to be real-coefficient. This yields

Hφ (ω) = Hφ∗ (−ω) (5.12)

Hψ (ω) = Hψ∗ (−ω) (5.13)

35
5.1.4 Design lowpass filter
Here we have all the design equations and we can focus on the design of the
lowpass filter. Note that we only have to design the frequency band ω ∈
[0, π/2]. the value in [π/2, 3π/2] can be obtained using (5.6) and [3π/2, 2π]
band is the reflection version due to the real filter property.

36
5.2 Some Design Issues
Besides the above design equations, we want other properties which are useful
in the general case. Here, we mention the vanishing moments, the size of
support, support versus moments, and regularity. These properties are more
strongly related to the practical use of these functions.

5.2.1 Vanishing moments

The vanishing moment is a criterion about how a function decays toward
infinity. For example, the function sin t/t2 decays at a rate of 1/t2 as t
approaches to infinity. We can estimate the rate of decay by the following
integration, Z ∞
tk f (t)dt. (5.14)
−∞
The parameter k indicates the rate of decay. Take sin t/t2 for an example
again, if k = 0, the function tk f (t) behaves like a function decays at a rate
of 1/t2 and the sin t term makes the function oscillating between 1/t2 and
−1/t2 . These properties ensured that the integral converges to zero for k = 0.
For k = 1, the well-known sine integral gives the result to be π. For higher
k, the integral diverges. Therefore, we can indicate the decay rate from the
parameter k. We say the wavelet function ψ(t) has p vanishing moments if
Z ∞
tk ψ(t)dt = 0 for 0 ≤ k < p. (5.15)
−∞

The definition seems not useful in design of the wavelet functions because
it involves a continuous integration. A theorem (pp. 284 in [8]) shows the
four statements are equivalent if |ψ| = O((1 + t2 )−p/2−1 ) and |φ(t)| = O((1 +
t2 )−p/2−1 ),
1. The wavelet ψ has p vanishing moments.
2. The Fourier transform of φ(t), ψ(t) and its first p − 1 derivatives are
zero at ω = 0.
3. The Fourier transform of hφ [n], Hφ (ejω ) and its first p − 1 derivatives
are zero at ω = π.
4. For any 0 ≤ k < p,
∞
X
qk (t) = nk φ(t − n) (5.16)
n=−∞

is a polynomial of degree k.

37
If we combine statement 3 with the conjugate mirror filter condition (5.6),
we can write the lowpass filter as

jω
√ 1 + ejω p
Hφ (e ) = 2( ) L(ejω ) (5.17)
2
Where L(x) is a polynomial. This result simplifies the process of wavelet
design.

5.2.2 Size of support

The size of support indicates the filter length. Note that we prefer the discrete
filter to be FIR due to the stability and implementation issue. A theorem
(pp. 286 in [8]) states that if the support of φ(t) and hφ [n] is [N1 , N2 ], the
support of ψ(t) is [(N1 − N2 + 1)/2, (N2 − N1 + 1)/2].

5.2.3 Support versus Moments

Due to (5.17), if we choose a high order vanishing moment, Hφ (ejω ) is a high
order polynomial of ejω . The corresponding hφ [n] must have longer filter size.
This is a trade-off between the vanishing moments and the filter length.

5.2.4 Regularity
The regularity of ψ has significant influence on the error introduced by thresh-
olding or quantizing the wavelet coefficients. When reconstructing a signal
from wavelet coefficients
∞
X ∞
X
f= < f, ψj,n > ψj,n , (5.18)
j=−∞ n=−∞

the reconstruction error is related to the wavelet function we choose.

38
5.3 Daubechies wavelets
Based on these equation, Daubechies [9], designed a type of wavelet for a
given vanishing moment p and find the minimum size discrete filter. The
conclusion is that if we want the wavelet function with p vanishing moments,
the minimum filter size is 2p. The derivation first starts from 5.17, rewrite
as
√ 1 + e−jω p
Hφ (ejω ) = 2( ) R(ejω ). (5.19)
2
The absolute-square of this function is
|Hφ (ejω )|2 = Hφ (ejω )Hφ∗ (ejω )
1 + e−jω 1 + ejω p
= 2( ) R(ejω )R∗ (ejω )
2 2
2 + ejω + e−jω 2p
= 2( ) |R(ejω )|2
4
ω 2p ω
= 2(cos ) P (sin2 ). (5.20)
2 2
The last step makes P (sin2 ω2 ) = |R(ejω )|2 . Recall (5.6), we can determine
the form of P (x). Let y = sin2 ω2 . We have
(1 − y)p P (y) + y p P (1 − y) = 1. (5.21)
A theorem in algebra, called Bezout theorem, can solve this equation. The
unique solution is
p−1
!
X p−1+k
P (y) = yk . (5.22)
k=0
k
The polynomial P (y) is the minimum degree polynomial satisfying (5.21).
Once we have P (y), the polynomial R(ejω ) can be derived. First we decom-
pose R(ejω ) according to its roots
m
X m
Y
−jkω
jω
R(e ) = rk e = r0 (1 − ak e−jω ). (5.23)
k=0 k=0

Let z = ejω , the relation between P and R is

m
2 − z − z −1 Y
P( 2
) = r0 (1 − ak z −1 )(1 − ak z). (5.24)
4 k=0
−1
By solving the roots of P ( 2−z−z
4
) = 0, we have the roots of R, {ak , 1/ak }k=0,1,...,m
p−1
and r0 = 2 . Usually, we choose ak lies in the unit circle to have minimum
phase filter.

39
 Taking p = 2 for a example. The obtained polynomial P (y) is
1
!
X 1+k
P (y) = y k = 1 + 2y. (5.25)
k=0
k

2 − z − z −1 1 1
P( ) = 2 − z − z −1 . (5.26)
4 2 2
√ √
The roots are 2 + 3 and 2 − 3 . After factorization, we have the lowpass
filter to be
√ √ √ √ √ √ √ √
jω 2 + 6 3 2 + 6 −jω 3 2 − 6 −j2ω 2 − 6 −j3ω
Hφ (e ) = + e + e + e .
8 8 8 8
(5.27)
The discrete-time domain representation is
√ √ √ √ √ √ √ √
2+ 6 3 2+ 6 3 2− 6 2− 6
hφ [n] = δ[n]+ δ[n−1]+ δ[n−2]+ δ[n−3].
8 8 8 8
(5.28)
The result is the minimum size filter with 2 vanishing moments and the
corresponding filter size is 4. Recall the conclusion mentioned above, the filter
size is two times the vanishing moment. Higher order Daubechies wavelets
are derived at similar way. The coefficient and the plot are shown in the
appendix A.

5.4 Symlets
Take a look at the discrete filters and the scaling/wavelet functions of Daubeches
wavelets. These functions are far from symmetry. That’s because Daubechies
wavelets select the minimum phase square root such that the energy concen-
trates near the starting point of their support. Symmlets select other set of
roots to have closer symmetry but with linear complex phase. The coeffi-
cients of these filters are list in the Appendix B.

5.5 Coiflets
For an application in numerical analysis, Coifman asked Daubechies [9] to
construct a family of wavelets ψ that have p vanishing moments, minimum-
size support and Z ∞
φ(t)dt = 1, (5.29)
−∞

40
 Z ∞
tk φ(t)dt = 0 for 1 ≤ k < p. (5.30)
−∞

The equation above can be taken as some requirement about vanishing mo-
ments of the scaling function. The resulting coiflets has a support of size
3p − 1. These coefficient of the filters are shown in Appendix C.

41
Chapter 6

Applications of Wavelet
Transform

6.1 Introduction
In the previous sections, we mention the fundamental concepts of the wavelet
transform. More than a theory, the wavelets are widely used in many appli-
cations. Its ability of multiresolution outperforms Fourier-based transform
such as discrete cosine transform in terms of compression and quality. Here,
we talk more about JPEG200, which is wavelet-based, and some applications
of the wavelet transform.

6.2 Image Compression: JPEG 2000

Before JPEG 2000, we mention the JPEG image compression standard.
JPEG is based the discrete cosine transform (DCT). However, this standard
divides the original image into 8-by-8 blocks and do DCT. Use different quan-
tization tables for luminance and chrominance. We have a good compression
ratio and maintain the quality of the original image. If the compression ratio
is pretty high, the block effect is perceivable. This effect originates from the
artificial division of the blocks. The neighboring blocks might exhibit simi-
lar color, but after quantization and inverse quantization, the reconstructed
blocks differ in color. The discontinuity between blocks emerges. Fig. 6.2
shows the JPEG coding system.
JPEG 2000 adapts the wavelet transform. The algorithm dynamically
selects the high power part and records them. At low compression ratio,
the performance of JPEG and JPEG 2000 is almost the same. At high
compression ratio, the performance of JPEG 2000 is much better than that

42
 Figure 6.1: The simple system diagram of JPEG

of the JPEG because it has no block effect. See Fig. (6.2).

JPEG 2000 is a wavelet-based image compression standard. It was pro-
posed by the ISO JPEG committee in the year 2000 with the intention of
replacing their original DCT-based JPEG standard which was proposed in
the year 1991. Different from JPEG, JPEG 2000 regards the whole image
as a unit and take 2D wavelet transform many levels. Due to the discussion
in the previous chapters, the most important coefficients concentrates in the
LL band. We can focus ourselves on how to encode the coefficients by its
significance. The coding method [3], [4], [5], [10] are good references.
JPEG 2000 uses two different wavelet transforms, one is biorthogonal
Daubechies 5/ 3 for lossless compression and a Daubechies 9 /7 wavelet for
lossy compression. The Daub 9/7 wavelet is an irreversible transform because
it has quantization noise that depends on the precision of the decoder. The
Daub 5/ 3 is reversible because the coefficients are integers, we do not have to
round the results. At low compression ratio, the performance of JPEG and
JPEG 2000 is almost the same. At high compression ratio, the performance
is much better than the JPEG because it has no block effect.
JPEG 2000 requires much computation time than JPEG on the simi-
lar compression rates. [11] proposes a new calculation way to compute the
convolution and it has much less computation complexity than the usual

43
Figure 6.2: Comparison between JPEG and JPEG 2000. CR stands for
compression ration and RMSE means root mean square error.

convolution.
An overview of the JPEG 2000 system can be found in [12], [13].

6.3 Other Applications

6.3.1 Edge and Corner Detection
An important issue in pattern recognition is to locate the edges. With wavelet
transform, we can anaylze a portion of a signal with different scales. We can
distinguish the noise and actual corner more precisely.

44
6.3.2 Pattern recognition
Pattern recognition depends on computing the features of the data. If we
first condense the signal by the wavelet transform, the computation effort
can be reduced. Once the pattern is probably recognized, we can use wavelet
transform to do finer analysis.

6.3.3 Filter design

As the wavelet transform can distinguish the noise and signal edge. We can
design an anti-noise filter without distortion of the original signal.

6.3.4 Analysis of the Electrocardiogram (ECG)[14]

An ECG is not smooth with many sudden transitions. If we analyze the
spectrum, noises and desired signal transitions cannot be separated using or-
dinary filters. Here wavelet transform can also remove the noise and preserve
the features of ECG.
There are still another applications which are not noted. In short, the
wavelet transform is a powerful tool solving the problems beyond the scope
of Fourier transform.

45
Chapter 7

Conclusions

In this tutorial, we explore the world of wavelets. From the abstract idea
in approximation, multiresolition theory, we generalize the Fourier transform
and start the journey of wavelets. The discrete wavelet is more useful in real-
ization. We often use 2D wavelets to do image compression. The continuous
wavelet transform analyze the continuous-time signal in a different perspec-
tive. By the advantage of multiresolution, we can locate time and frequency
more accurately. The wavelet design is more complicated in mathematics but
the design procedure completes the existence of the wavelets. The application
chapter mentions the nowadays JPEG and JPEG2000 standard.
[8] and [9] give a thorough approach to the wavelets but in purely math-
ematically words. [1] and [15] illustrate some examples on the wavelet and
then the abstract concepts. Other references mainly focus on some parts of
the materials.
The wavelets bring us to a new vision of signal processing. It tactically
avoids the problem that Fourier analysis encounters. Its implementation
is simple. We need some designed filters to do the task. Although the
implementation is quite easy, the filter design includes lots of mathematical
originality and this is the research topic. Once we establish wavelets with
more ideal properties, lots of difficult problems might be solved.

46
Appendix A

Daubechies wavelets

A.1 Coefficients
A.1.1 p=2
1

n LoD HiD LoR HiR

0 -0.129409523 -0.482962913 0.482962913 -0.129409523
1 0.224143868 0.836516304 0.836516304 -0.224143868
2 0.836516304 -0.224143868 0.224143868 0.836516304
3 0.482962913 -0.129409523 -0.129409523 -0.482962913

A.1.2 p=3

n LoD HiD LoR HiR

0 0.035226292 -0.332670553 0.332670553 0.035226292
1 -0.085441274 0.806891509 0.806891509 0.085441274
2 -0.13501102 -0.459877502 0.459877502 -0.13501102
3 0.459877502 -0.13501102 -0.13501102 -0.459877502
4 0.806891509 0.085441274 -0.085441274 0.806891509
5 0.332670553 0.035226292 0.035226292 -0.332670553
1
The filters are abbreviated in Lo/Hi (lowpass/highpass) and D/R (decomposi-
tion/reconstruction) in subband coding.

47
A.1.3 p=4

n LoD HiD LoR HiR

0 -0.010597402 -0.230377813 0.230377813 -0.010597402
1 0.032883012 0.714846571 0.714846571 -0.032883012
2 0.030841382 -0.630880768 0.630880768 0.030841382
3 -0.187034812 -0.027983769 -0.027983769 0.187034812
4 -0.027983769 0.187034812 -0.187034812 -0.027983769
5 0.630880768 0.030841382 0.030841382 -0.630880768
6 0.714846571 -0.032883012 0.032883012 0.714846571
7 0.230377813 -0.010597402 -0.010597402 -0.230377813

A.1.4 p=5

n LoD HiD LoR HiR

0 0.003335725 -0.160102398 0.160102398 0.003335725
1 -0.012580752 0.60382927 0.60382927 0.012580752
2 -0.00624149 -0.724308528 0.724308528 -0.00624149
3 0.077571494 0.138428146 0.138428146 -0.077571494
4 -0.03224487 0.242294887 -0.242294887 -0.03224487
5 -0.242294887 -0.03224487 -0.03224487 0.242294887
6 0.138428146 -0.077571494 0.077571494 0.138428146
7 0.724308528 -0.00624149 -0.00624149 -0.724308528
8 0.60382927 0.012580752 -0.012580752 0.60382927
9 0.160102398 0.003335725 0.003335725 -0.160102398

48
A.1.5 p=6

n LoD HiD LoR HiR

0 -0.001077301 -0.111540743 0.111540743 -0.001077301
1 0.004777258 0.49462389 0.49462389 -0.004777258
2 0.000553842 -0.751133908 0.751133908 0.000553842
3 -0.031582039 0.315250352 0.315250352 0.031582039
4 0.027522866 0.226264694 -0.226264694 0.027522866
5 0.097501606 -0.129766868 -0.129766868 -0.097501606
6 -0.129766868 -0.097501606 0.097501606 -0.129766868
7 -0.226264694 0.027522866 0.027522866 0.226264694
8 0.315250352 0.031582039 -0.031582039 0.315250352
9 0.751133908 0.000553842 0.000553842 -0.751133908
10 0.49462389 -0.004777258 0.004777258 0.49462389
11 0.111540743 -0.001077301 -0.001077301 -0.111540743

A.1.6 p=7

n LoD HiD LoR HiR

0 0.000353714 -0.077852054 0.077852054 0.000353714
1 -0.001801641 0.396539319 0.396539319 0.001801641
2 0.000429578 -0.729132091 0.729132091 0.000429578
3 0.012550999 0.469782287 0.469782287 -0.012550999
4 -0.016574542 0.143906004 -0.143906004 -0.016574542
5 -0.038029937 -0.224036185 -0.224036185 0.038029937
6 0.080612609 -0.071309219 0.071309219 0.080612609
7 0.071309219 0.080612609 0.080612609 -0.071309219
8 -0.224036185 0.038029937 -0.038029937 -0.224036185
9 -0.143906004 -0.016574542 -0.016574542 0.143906004
10 0.469782287 -0.012550999 0.012550999 0.469782287
11 0.729132091 0.000429578 0.000429578 -0.729132091
12 0.396539319 0.001801641 -0.001801641 0.396539319
13 0.077852054 0.000353714 0.000353714 -0.077852054

49
A.1.7 p=8

n LoD HiD LoR HiR

0 -0.000117477 -0.054415842 0.054415842 -0.000117477
1 0.000675449 0.312871591 0.312871591 -0.000675449
2 -0.00039174 -0.675630736 0.675630736 -0.00039174
3 -0.004870353 0.585354684 0.585354684 0.004870353
4 0.008746094 0.015829105 -0.015829105 0.008746094
5 0.013981028 -0.284015543 -0.284015543 -0.013981028
6 -0.044088254 -0.000472485 0.000472485 -0.044088254
7 -0.017369301 0.128747427 0.128747427 0.017369301
8 0.128747427 0.017369301 -0.017369301 0.128747427
9 0.000472485 -0.044088254 -0.044088254 -0.000472485
10 -0.284015543 -0.013981028 0.013981028 -0.284015543
11 -0.015829105 0.008746094 0.008746094 0.015829105
12 0.585354684 0.004870353 -0.004870353 0.585354684
13 0.675630736 -0.00039174 -0.00039174 -0.675630736
14 0.312871591 -0.000675449 0.000675449 0.312871591
15 0.054415842 -0.000117477 -0.000117477 -0.054415842

50
A.1.8 p=9

n LoD HiD LoR HiR

0 3.93E-05 -0.038077947 0.038077947 3.93E-05
1 -0.000251963 0.243834675 0.243834675 0.000251963
2 0.000230386 -0.604823124 0.604823124 0.000230386
3 0.001847647 0.657288078 0.657288078 -0.001847647
4 -0.004281504 -0.133197386 0.133197386 -0.004281504
5 -0.004723205 -0.293273783 -0.293273783 0.004723205
6 0.022361662 0.096840783 -0.096840783 0.022361662
7 0.000250947 0.148540749 0.148540749 -0.000250947
8 -0.067632829 -0.030725681 0.030725681 -0.067632829
9 0.030725681 -0.067632829 -0.067632829 -0.030725681
10 0.148540749 -0.000250947 0.000250947 0.148540749
11 -0.096840783 0.022361662 0.022361662 0.096840783
12 -0.293273783 0.004723205 -0.004723205 -0.293273783
13 0.133197386 -0.004281504 -0.004281504 -0.133197386
14 0.657288078 -0.001847647 0.001847647 0.657288078
15 0.604823124 0.000230386 0.000230386 -0.604823124
16 0.243834675 0.000251963 -0.000251963 0.243834675
17 0.038077947 3.93E-05 3.93E-05 -0.038077947

51
A.1.9 p=10

n LoD HiD LoR HiR

0 -1.33E-05 -0.026670058 0.026670058 -1.33E-05
1 9.36E-05 0.1881768 0.1881768 -9.36E-05
2 -0.000116467 -0.527201189 0.527201189 -0.000116467
3 -0.000685857 0.688459039 0.688459039 0.000685857
4 0.001992405 -0.281172344 0.281172344 0.001992405
5 0.001395352 -0.249846424 -0.249846424 -0.001395352
6 -0.010733175 0.195946274 -0.195946274 -0.010733175
7 0.003606554 0.12736934 0.12736934 -0.003606554
8 0.033212674 -0.093057365 0.093057365 0.033212674
9 -0.029457537 -0.071394147 -0.071394147 0.029457537
10 -0.071394147 0.029457537 -0.029457537 -0.071394147
11 0.093057365 0.033212674 0.033212674 -0.093057365
12 0.12736934 -0.003606554 0.003606554 0.12736934
13 -0.195946274 -0.010733175 -0.010733175 0.195946274
14 -0.249846424 -0.001395352 0.001395352 -0.249846424
15 0.281172344 0.001992405 0.001992405 -0.281172344
16 0.688459039 0.000685857 -0.000685857 0.688459039
17 0.527201189 -0.000116467 -0.000116467 -0.527201189
18 0.1881768 -9.36E-05 9.36E-05 0.1881768
19 0.026670058 -1.33E-05 -1.33E-05 -0.026670058

A.2 Function Plot

The function plots are made from MATLAB Wavelet Toolbox. The φ(t) and
ψ(t) are derived from (5.5).

52
A.2.1 p=2

A.2.2 p=3

53
A.2.3 p=4

A.2.4 p=5

54
A.2.5 p=6

A.2.6 p=7

55
A.2.7 p=8

A.2.8 p=9

56
A.2.9 p=10

57
Appendix B

Symlets

B.1 Coefficients
B.1.1 p=2

n LoD HiD LoR HiR

0 -0.129409523 -0.482962913 0.482962913 -0.129409523
1 0.224143868 0.836516304 0.836516304 -0.224143868
2 0.836516304 -0.224143868 0.224143868 0.836516304
3 0.482962913 -0.129409523 -0.129409523 -0.482962913

B.1.2 p=3

n LoD HiD LoR HiR

0 0.035226292 -0.332670553 0.332670553 0.035226292
1 -0.085441274 0.806891509 0.806891509 0.085441274
2 -0.13501102 -0.459877502 0.459877502 -0.13501102
3 0.459877502 -0.13501102 -0.13501102 -0.459877502
4 0.806891509 0.085441274 -0.085441274 0.806891509
5 0.332670553 0.035226292 0.035226292 -0.332670553

58
B.1.3 p=4

n LoD HiD LoR HiR

0 -0.075765715 -0.032223101 0.032223101 -0.075765715
1 -0.029635528 -0.012603967 -0.012603967 0.029635528
2 0.497618668 0.099219544 -0.099219544 0.497618668
3 0.803738752 0.297857796 0.297857796 -0.803738752
4 0.297857796 -0.803738752 0.803738752 0.297857796
5 -0.099219544 0.497618668 0.497618668 0.099219544
6 -0.012603967 0.029635528 -0.029635528 -0.012603967
7 0.032223101 -0.075765715 -0.075765715 -0.032223101

B.1.4 p=5

n LoD HiD LoR HiR

0 0.027333068 -0.019538883 0.019538883 0.027333068
1 0.029519491 -0.021101834 -0.021101834 -0.029519491
2 -0.039134249 0.17532809 -0.17532809 -0.039134249
3 0.199397534 0.016602106 0.016602106 -0.199397534
4 0.72340769 -0.633978963 0.633978963 0.72340769
5 0.633978963 0.72340769 0.72340769 -0.633978963
6 0.016602106 -0.199397534 0.199397534 0.016602106
7 -0.17532809 -0.039134249 -0.039134249 0.17532809
8 -0.021101834 -0.029519491 0.029519491 -0.021101834
9 0.019538883 0.027333068 0.027333068 -0.019538883

59
B.2 Function Plot
B.2.1 p=2

B.2.2 p=3

60
B.2.3 p=4

B.2.4 p=5

61
Appendix C

Coiflets

C.1 Coefficients
C.1.1 p=1

n LoD HiD LoR HiR

0 -0.015655728 0.07273262 -0.07273262 -0.015655728
1 -0.07273262 0.337897662 0.337897662 0.07273262
2 0.384864847 -0.85257202 0.85257202 0.384864847
3 0.85257202 0.384864847 0.384864847 -0.85257202
4 0.337897662 0.07273262 -0.07273262 0.337897662
5 -0.07273262 -0.015655728 -0.015655728 0.07273262

62
C.1.2 p=2

n LoD HiD LoR HiR

0 -0.000720549 -0.016387336 0.016387336 -0.000720549
1 -0.001823209 -0.041464937 -0.041464937 0.001823209
2 0.005611435 0.067372555 -0.067372555 0.005611435
3 0.023680172 0.386110067 0.386110067 -0.023680172
4 -0.059434419 -0.812723635 0.812723635 -0.059434419
5 -0.076488599 0.417005184 0.417005184 0.076488599
6 0.417005184 0.076488599 -0.076488599 0.417005184
7 0.812723635 -0.059434419 -0.059434419 -0.812723635
8 0.386110067 -0.023680172 0.023680172 0.386110067
9 -0.067372555 0.005611435 0.005611435 0.067372555
10 -0.041464937 0.001823209 -0.001823209 -0.041464937
11 0.016387336 -0.000720549 -0.000720549 -0.016387336

63
C.1.3 p=3

n LoD HiD LoR HiR

0 -3.46E-05 0.003793513 -0.003793513 -3.46E-05
1 -7.10E-05 0.007782596 0.007782596 7.10E-05
2 0.000466217 -0.023452696 0.023452696 0.000466217
3 0.001117519 -0.065771911 -0.065771911 -0.001117519
4 -0.002574518 0.06112339 -0.06112339 -0.002574518
5 -0.009007976 0.405176902 0.405176902 0.009007976
6 0.015880545 -0.793777223 0.793777223 0.015880545
7 0.034555028 0.428483476 0.428483476 -0.034555028
8 -0.082301927 0.071799822 -0.071799822 -0.082301927
9 -0.071799822 -0.082301927 -0.082301927 0.071799822
10 0.428483476 -0.034555028 0.034555028 0.428483476
11 0.793777223 0.015880545 0.015880545 -0.793777223
12 0.405176902 0.009007976 -0.009007976 0.405176902
13 -0.06112339 -0.002574518 -0.002574518 0.06112339
14 -0.065771911 -0.001117519 0.001117519 -0.065771911
15 0.023452696 0.000466217 0.000466217 -0.023452696
16 0.007782596 7.10E-05 -7.10E-05 0.007782596
17 -0.003793513 -3.46E-05 -3.46E-05 0.003793513

64
C.1.4 p=4

n LoD HiD LoR HiR

0 -1.78E-06 -0.000892314 0.000892314 -1.78E-06
1 -3.26E-06 -0.001629492 -0.001629492 3.26E-06
2 3.12E-05 0.007346166 -0.007346166 3.12E-05
3 6.23E-05 0.016068944 0.016068944 -6.23E-05
4 -0.000259975 -0.0266823 0.0266823 -0.000259975
5 -0.000589021 -0.0812667 -0.0812667 0.000589021
6 0.001266562 0.056077313 -0.056077313 0.001266562
7 0.003751436 0.415308407 0.415308407 -0.003751436
8 -0.005658287 -0.782238931 0.782238931 -0.005658287
9 -0.015211732 0.434386056 0.434386056 0.015211732
10 0.025082262 0.066627474 -0.066627474 0.025082262
11 0.039334427 -0.096220442 -0.096220442 -0.039334427
12 -0.096220442 -0.039334427 0.039334427 -0.096220442
13 -0.066627474 0.025082262 0.025082262 0.066627474
14 0.434386056 0.015211732 -0.015211732 0.434386056
15 0.782238931 -0.005658287 -0.005658287 -0.782238931
16 0.415308407 -3.75E-03 3.75E-03 0.415308407
17 -0.056077313 1.27E-03 1.27E-03 0.056077313
18 -0.0812667 0.000589021 -0.000589021 -0.0812667
19 0.0266823 -0.000259975 -0.000259975 -0.0266823
20 0.016068944 -6.23E-05 6.23E-05 0.016068944
21 -0.007346166 3.12E-05 3.12E-05 0.007346166
22 -0.001629492 3.26E-06 -3.26E-06 -0.001629492
23 0.000892314 -1.78E-06 -1.78E-06 -0.000892314

65
C.1.5 p=5

n LoD HiD LoR HiR

0 -9.52E-08 0.000212081 -0.000212081 -9.52E-08
1 -1.67E-07 0.00035859 0.00035859 1.67E-07
2 2.06E-06 -0.002178236 0.002178236 2.06E-06
3 3.73E-06 -0.004159359 -0.004159359 -3.73E-06
4 -2.13E-05 0.010131118 -0.010131118 -2.13E-05
5 -4.13E-05 0.023408157 0.023408157 4.13E-05
6 0.000140541 -0.028168029 0.028168029 0.000140541
7 0.00030226 -0.091920011 -0.091920011 -0.00030226
8 -0.000638131 0.052043163 -0.052043163 -0.000638131
9 -0.001662864 0.421566207 0.421566207 0.001662864
10 0.002433373 -0.774289604 0.774289604 0.002433373
11 0.006764185 0.437991626 0.437991626 -0.006764185
12 -0.009164231 0.062035964 -0.062035964 -0.009164231
13 -0.019761779 -0.105574209 -0.105574209 0.019761779
14 0.032683574 -0.041289209 0.041289209 0.032683574
15 0.041289209 0.032683574 0.032683574 -0.041289209
16 -0.105574209 1.98E-02 -1.98E-02 -0.105574209
17 -0.062035964 -9.16E-03 -9.16E-03 0.062035964
18 0.437991626 -0.006764185 0.006764185 0.437991626
19 0.774289604 0.002433373 0.002433373 -0.774289604
20 0.421566207 1.66E-03 -1.66E-03 0.421566207
21 -0.052043163 -6.38E-04 -6.38E-04 0.052043163
22 -0.091920011 -3.02E-04 3.02E-04 -0.091920011
23 0.028168029 1.41E-04 1.41E-04 -0.028168029
24 0.023408157 4.13E-05 -4.13E-05 0.023408157
25 -0.010131118 -2.13E-05 -2.13E-05 0.010131118
26 -0.004159359 -3.73E-06 3.73E-06 -0.004159359
27 0.002178236 2.06E-06 2.06E-06 -0.002178236
28 0.00035859 1.67E-07 -1.67E-07 0.00035859
29 -0.000212081 -9.52E-08 -9.52E-08 0.000212081

66
C.2 Function Plot
C.2.1 p=1

C.2.2 p=2

67
C.2.3 p=3

C.2.4 p=4

68
C.2.5 p=5

69
References

[1] R. C. Gonzalez and R. E. Woods, Digital Image Processing (2nd Edi-

tion). Prentice Hall, January 2002.

[2] S. G. Mallat, “A theory for multiresolution signal decomposition: the

wavelet representation,” vol. 11, pp. 674–693, July 1989.

[3] J. M. Shapiro, “Embedded image coding using zerotrees of wavelet co-

efficients,” vol. 41, pp. 3445–3462, Dec. 1993.

[4] A. Said and W. A. Pearlman, “A new, fast, and efficient image codec
based on set partitioning in hierarchical trees,” vol. 6, pp. 243–250, June
1996.

[5] D. Taubman, “High performance scalable image compression with

ebcot,” in Proc. International Conference on Image Processing ICIP
99, vol. 3, pp. 344–348, Oct. 24–28, 1999.

[6] Y.-S. Zhang, “Multiresolution analysis for image by generalized 2-d

wavelets,” Master’s thesis, 2008.

[7] I. Daubechies, Ten lectures on wavelets, vol. 61 of CBMS-NSF Regional

Conference Series in Applied Mathematics. Philadelphia, PA: Society
for Industrial and Applied Mathematics (SIAM), 1992.

[8] S. Mallat, A Wavelet Tour of Signal Processing, 3rd ed., Third Edition:
The Sparse Way. Academic Press, 3 ed., December 2008.

[9] I. Daubechies, “Orthonormal bases of compactly supported wavelets,”

Communications on Pure and Applied Mathematics, vol. 41, no. 7,
pp. 909–996, 1988.

[10] P.-S. T. Tinku Acharya, JPEG2000 Standard for Image Compression,

ch. Coding Algorithms in JPEG2000, pp. 163–196. 2005.

70
[11] N.-C. Shen, “Sectioned convolution for discrete wavelet transform,”
Master’s thesis, 2008.

[12] M. Rabbani and R. Joshi, “An overview of the jpeg 2000 still image
compression standard,” pp. 3–48, Elsevier, January 2002.

[13] C. Christopoulos, A. Skodras, and T. Ebrahimi, “The jpeg2000 still

image coding system: an overview,” vol. 46, pp. 1103–1127, Nov. 2000.

[14] P. M. A. K. Louis and A. Rieder, Wavelets Theory and Applications.

John Wiley & Sons, 1997.

[15] Z.-N. Li and M. S. Drew, Fundamentals of Multimedia. Prentice Hall,

October 2003.

71