7

Wavelets and
Multiresolution
Processing
All this time, the guard was looking at her, first through
a telescope, then through a microscope, and then
through an opera glass.
Lewis Carrol, Through the Looking Glass

Preview
Although the Fourier transform has been the mainstay of transform-based
image processing since the late 1950s, a more recent transformation, called the
wavelet transform, is now making it even easier to compress, transmit, and an-
alyze many images. Unlike the Fourier transform, whose basis functions are si-
nusoids, wavelet transforms are based on small waves, called wavelets, of
varying frequency and limited duration. This allows them to provide the equiv-
alent of a musical score for an image, revealing not only what notes (or fre-
quencies) to play but also when to play them. Fourier transforms, on the other
hand, provide only the notes or frequency information; temporal information
is lost in the transformation process.
In 1987, wavelets were first shown to be the foundation of a powerful new
approach to signal processing and analysis called multiresolution theory (Mallat
[1987]). Multiresolution theory incorporates and unifies techniques from a vari-
ety of disciplines, including subband coding from signal processing, quadrature
mirror filtering from digital speech recognition, and pyramidal image processing.
As its name implies, multiresolution theory is concerned with the representation
and analysis of signals (or images) at more than one resolution. The appeal of
such an approach is obvious—features that might go undetected at one resolu-
tion may be easy to detect at another. Although the imaging community’s inter-
est in multiresolution analysis was limited until the late 1980s, it is now difficult to
keep up with the number of papers, theses, and books devoted to the subject.

461
462 Chapter 7 ■ Wavelets and Multiresolution Processing

In this chapter, we examine wavelet-based transformations from a multires-

olution point of view. Although such transformations can be presented in other
ways, this approach simplifies both their mathematical and physical interpreta-
tions. We begin with an overview of imaging techniques that influenced the for-
mulation of multiresolution theory. Our objective is to introduce the theory’s
fundamental concepts within the context of image processing and simultane-
ously provide a brief historical perspective of the method and its application.
The bulk of the chapter is focused on the development and use of the discrete
wavelet transform. To demonstrate the usefulness of the transform, examples
ranging from image coding to noise removal and edge detection are provided.
In the next chapter, wavelets will be used for image compression, an application
in which they have received considerable attention.

7.1 Background
When we look at images, generally we see connected regions of similar texture
and intensity levels that combine to form objects. If the objects are small in
size or low in contrast, we normally examine them at high resolutions; if they
are large in size or high in contrast, a coarse view is all that is required. If both
small and large objects—or low- and high-contrast objects—are present simul-
taneously, it can be advantageous to study them at several resolutions. This, of
course, is the fundamental motivation for multiresolution processing.
From a mathematical viewpoint, images are two-dimensional arrays of inten-
Local histograms are sity values with locally varying statistics that result from different combinations
histograms of the pixels
in a neighborhood (see
of abrupt features like edges and contrasting homogeneous regions.As illustrated
Section 3.3.3). in Fig. 7.1—an image that will be examined repeatedly in the remainder of the

FIGURE 7.1
An image and its
local histogram
variations.
 7.1 ■ Background 463

section—local histograms can vary significantly from one part of an image to

another, making statistical modeling over the span of an entire image a diffi-
cult, or impossible task.

7.1.1 Image Pyramids

A powerful, yet conceptually simple structure for representing images at more
than one resolution is the image pyramid (Burt and Adelson [1983]). Originally
devised for machine vision and image compression applications, an image
pyramid is a collection of decreasing resolution images arranged in the shape
of a pyramid. As can be seen in Fig. 7.2(a), the base of the pyramid contains a
high-resolution representation of the image being processed; the apex con-
tains a low-resolution approximation. As you move up the pyramid, both size
and resolution decrease. Base level J is of size 2J * 2J or N * N, where
J = log2 N, apex level 0 is of size 1 * 1, and general level j is of size 2j * 2j,
where 0 … j … J. Although the pyramid shown in Fig. 7.2(a) is composed of
J + 1 resolution levels from 2J * 2J to 20 * 20, most image pyramids are trun-
cated to P + 1 levels, where 1 … P … J and j = J - P, Á , J - 2, J - 1, J.
That is, we normally limit ourselves to P reduced resolution approximations of
the original image; a 1 * 1 (i.e., single pixel) approximation of a 512 * 512
image, for example, is of little value. The total number of pixels in a P + 1 level
pyramid for P 7 0 is

1 1 1 4
N2 ¢ 1 + 1
+ 2
+Á+ P
≤ … N2
(4) (4) (4) 3

Figure 7.2(b) shows a simple system for constructing two intimately related
image pyramids. The Level j - 1 approximation output provides the images

Level 0 (apex)
a
11
b
22 Level 1
FIGURE 7.2
44 Level 2
(a) An image
pyramid. (b) A
simple system for
Level J
1 creating
N/ 2  N/ 2
approximation
Level J (base) and prediction
NN
residual pyramids.

Downsampler
(rows and columns)
Approximation Level j
1
2T
filter approximation
2c Upsampler
(rows and columns)
Interpolation
filter
Prediction
Level j
Level j

 prediction
input image
residual
464 Chapter 7 ■ Wavelets and Multiresolution Processing

needed to build an approximation pyramid (as described in the preceding

paragraph), while the Level j prediction residual output is used to build a
complementary prediction residual pyramid. Unlike approximation pyramids,
prediction residual pyramids contain only one reduced-resolution approxi-
mation of the input image (at the top of the pyramid, level J - P). All other
In general, a prediction levels contain prediction residuals, where the level j prediction residual (for
residual can be defined
as the difference
J - P + 1 … j … J) is defined as the difference between the level j approxi-
between an image and a mation (the input to the block diagram) and an estimate of the level j approx-
predicted version of the
image. As will be seen in
imation based on the level j - 1 approximation (the approximation output in
Section 8.2.9, prediction the block diagram).
residuals can often be
coded more efficiently
As Fig. 7.2(b) suggests, both approximation and prediction residual pyra-
than 2-D intensity arrays. mids are computed in an iterative fashion. Before the first iteration, the image
to be represented in pyramidal form is placed in level J of the approximation
pyramid. The following three-step procedure is then executed P times—for
j = J, J - 1, Á , and J - P + 1 (in that order):

Step 1. Compute a reduced-resolution approximation of the Level j input

image [the input on the left side of the block diagram in Fig. 7.2(b)]. This is
done by filtering and downsampling the filtered result by a factor of 2. Both
of these operations are described in the next paragraph. Place the resulting
approximation at level j - 1 of the approximation pyramid.
Step 2. Create an estimate of the Level j input image from the reduced-
resolution approximation generated in step 1. This is done by upsampling
and filtering (see the next paragraph) the generated approximation. The re-
sulting prediction image will have the same dimensions as the Level j input
image.
Step 3. Compute the difference between the prediction image of step 2
and the input to step 1. Place this result in level j of the prediction residual
pyramid.

At the conclusion of P iterations (i.e., following the iteration in which

j = J - P + 1), the level J - P approximation output is placed in the pre-
diction residual pyramid at level J - P. If a prediction residual pyramid is not
needed, this operation—along with steps 2 and 3 and the upsampler, inter-
polation filter, and summer of Fig. 7.2(b)—can be omitted.
A variety of approximation and interpolation filters can be incorporated
into the system of Fig. 7.2(b). Typically, the filtering is performed in the spatial
domain (see Section 3.4). Useful approximation filtering techniques include
neighborhood averaging (see Section 3.5.1.), which produces mean pyramids;
lowpass Gaussian filtering (see Sections 4.7.4 and 4.8.3), which produces
Gaussian pyramids; and no filtering, which results in subsampling pyramids.
Any of the interpolation methods described in Section 2.4.4, including nearest
neighbor, bilinear, and bicubic, can be incorporated into the interpolation fil-
ter. Finally, we note that the upsampling and downsampling blocks of Fig.
7.2(b) are used to double and halve the spatial dimensions of the approxima-
tion and prediction images that are computed. Given an integer variable n and
1-D sequence of samples f(n), upsampled sequence f2c (n) is defined as
 7.1 ■ Background 465
In this chapter, we will be
f(n>2) if n is even
f2c (n) = b (7.1-1) working with both
continuous and discrete
0 otherwise functions and variables.
With the notable
exception of 2-D image
where, as is indicated by the subscript, the upsampling is by a factor of 2. The f(x, y) and unless other-
complementary operation of downsampling by 2 is defined as wise noted, x, y, z, Á are
continuous variables;
i, j, k, l, m, n, Á are
f2T (n) = f(2n) (7.1-2) discrete variables.

Upsampling can be thought of as inserting a 0 after every sample in a sequence;

downsampling can be viewed as discarding every other sample. The upsampling
and downsampling blocks in Fig. 7.2(b), which are labeled 2 c and 2 T, respectively,
are annotated to indicate that both the rows and columns of the 2-D inputs on
which they operate are to be up- and downsampled. Like the separable 2-D DFT
in Section 4.11.1, 2-D upsampling and downsampling can be performed by suc-
cessive passes of the 1-D operations defined in Eqs. (7.1-1) and (7.1-2).

■ Figure 7.3 shows both an approximation pyramid and a prediction residual EXAMPLE 7.1:
pyramid for the vase of Fig. 7.1. A lowpass Gaussian smoothing filter (see Approximation
and prediction
Section 4.7.4) was used to produce the four-level approximation pyramid in
residual pyramids.
Fig. 7.3(a). As you can see, the resulting pyramid contains the original
512 * 512 resolution image (at its base) and three low-resolution approxima-
tions (of resolution 256 * 256, 128 * 128, and 64 * 64). Thus, P is 3 and levels
9, 8, 7, and 6 out of a possible log 2 (512) + 1 or 10 levels are present. Note the
reduction in detail that accompanies the lower resolutions of the pyramid. The
level 6 (i.e., 64 * 64) approximation image is suitable for locating the window
stiles (i.e., the window pane framing), for example, but not for finding the stems
of the plant. In general, the lower-resolution levels of a pyramid can be used for
the analysis of large structures or overall image context; the high-resolution im-
ages are appropriate for analyzing individual object characteristics. Such a
coarse-to-fine analysis strategy is particularly useful in pattern recognition.
A bilinear interpolation filter was used to produce the prediction residual
pyramid in Fig. 7.3(b). In the absence of quantization error, the resulting predic-
tion residual pyramid can be used to generate the complementary approxima-
tion pyramid in Fig. 7.3(a), including the original image, without error. To do so,
we begin with the level 6 64 * 64 approximation image (the only approxima-
tion image in the prediction residual pyramid), predict the level 7 128 * 128 res-
olution approximation (by upsampling and filtering), and add the level 7
prediction residual. This process is repeated using successively computed ap-
proximation images until the original 512 * 512 image is generated. Note that
the prediction residual histogram in Fig. 7.3(b) is highly peaked around zero; the
approximation histogram in Fig. 7.3(a) is not. Unlike approximation images, pre-
diction residual images can be highly compressed by assigning fewer bits to the
more probable values (see the variable-length codes of Section 8.2.1). Finally, we
note that the prediction residuals in Fig. 7.3(b) are scaled to make small predic-
tion errors more visible; the prediction residual histogram, however, is based on
the original residual values, with level 128 representing zero error. ■
466 Chapter 7 ■ Wavelets and Multiresolution Processing

a
b
FIGURE 7.3
Two image
pyramids and
their histograms:
(a) an
approximation
pyramid;
(b) a prediction
residual pyramid.

The approximation
pyramid in (a) is called a
Gaussian pyramid
because a Gaussian filter
was used to construct it.
The prediction residual
pyramid in (b) is often
called a Laplacian
pyramid; note the
similarity in appearance
with the Laplacian fil-
tered images in Chapter 3.

7.1.2 Subband Coding

Another important imaging technique with ties to multiresolution analysis is
subband coding. In subband coding, an image is decomposed into a set of
bandlimited components, called subbands. The decomposition is performed so
that the subbands can be reassembled to reconstruct the original image with-
out error. Because the decomposition and reconstruction are performed by
means of digital filters, we begin our discussion with a brief introduction to
digital signal processing (DSP) and digital signal filtering.
The term “delay” implies Consider the simple digital filter in Fig. 7.4(a) and note that it is constructed
a time-based input
sequence and reflects the
from three basic components—unit delays, multipliers, and adders. Along the
fact that in digital signal top of the filter, unit delays are connected in series to create K- 1 delayed
filtering, the input is
usually a sampled analog
(i.e., right shifted) versions of the input sequence f(n). Delayed sequence
signal. f(n - 2), for example, is
 7.1 ■ Background 467

o
f(0) for n = 2
f(n - 2) = d
f(1) for n = 2 + 1 = 3
o
As the grayed annotations in Fig. 7.4(a) indicate, input sequence f(n) =
f(n - 0) and the K - 1 delayed sequences at the outputs of the unit delays,
denoted f(n - 1), f(n - 2), Á , f(n - K + 1), are multiplied by constants
h(0), h(1), Á , h(K - 1), respectively, and summed to produce the filtered
output sequence
q If the coefficients of the
fN(n) = aqh(k)f(n - k)
filter in Fig. 7.4(a) are
indexed using values of n
k=-
between 0 and K - 1 (as
(7.1-3) we have done), the limits
= f(n)
h(n) on the sum in Eq. (7.1-3)
can be reduced to 0 to
where
denotes convolution. Note that—except for a change in variables— K - 1 [like Eq. (4.4-10)].
Eq. (7.1-3) is equivalent to the discrete convolution defined in Eq. (4.4-10) of
Chapter 4. The K multiplication constants in Fig. 7.4(a) and Eq. (7.1-3) are

f(n
0) Unit f (n
1) Unit f (n
2) Unit f (n
K  1)

f(n) ...
delay delay delay

h(0) h(1) h(2) h(K
1)

h(0)f(n) h(1)f (n
1) h(2)f (n
2) h(K
1)f (n
K  1)

  ...  fˆ(n)  f (n)
h(n)

h(0)f (n)  h(1)f (n
1)

h(0)f (n)  h(1)f (n
1)  h(2)f (n
2)

K
1
h(k)f (n
k)  f (n)
h(n)
k0

1 1
h(1)
Input sequence f (n)  (n)

h(2)
Impulse response h(n)

h(0)
h(K
1)
h(3)
0 0
...
h(4)

1
1

1 0 1 2 . . . ... K
1
1 0 1 2 . . . ... K
1
n n
a
b c
FIGURE 7.4 (a) A digital filter; (b) a unit discrete impulse sequence; and (c) the impulse response of the filter.
468 Chapter 7 ■ Wavelets and Multiresolution Processing

called filter coefficients. Each coefficient defines a filter tap, which can be
thought of as the components needed to compute one term of the sum in Eq.
(7.1-3), and the filter is said to be of order K.
If the input to the filter of Fig. 7.4(a) is the unit discrete impulse of
Fig. 7.4(b) and Section 4.2.3, Eq. (7.1-3) becomes
q
fN(n) = a h(k)d(n - k)
k = -q (7.1-4)
= h(n)

That is, by substituting d(n) for input f(n) in Eq. (7.1-3) and making use of
the sifting property of the unit discrete impulse as defined in Eq. (4.2-13), we
find that the impulse response of the filter in Fig. 7.4(a) is the K-element se-
quence of filter coefficients that define the filter. Physically, the unit impulse
is shifted from left to right across the top of the filter (from one unit delay to
the next), producing an output that assumes the value of the coefficient at
the location of the delayed impulse. Because there are K coefficients, the im-
pulse response is of length K and the filter is called a finite impulse response
(FIR) filter.
In the remainder of the Figure 7.5 shows the impulse responses of six functionally related filters. Fil-
chapter, “filter h(n)” will
be used to refer to the ter h2(n) in Fig. 7.5(b) is a sign-reversed (i.e., reflected about the horizontal
filter whose impulse axis) version of h1(n) in Fig. 7.5(a). That is,
response is h(n).

h2(n) = -h1(n) (7.1-5)

1 1 1
h1(n) h2(n) 
h1(n) h3(n)  h1(
n)

0 0 0

1
1
1
. . .
3
2
1 0 1 2 3 4 5 6 7 . . . . . .
3
2
1 0 1 2 3 4 5 6 7 . . . . . .
3
2
1 0 1 2 3 4 5 6 7 . . .
n n n

1 1 1
h4(n)  h1(K
1
n) h5(n)  (
1)nh1(n) h6(n)  (
1)nh1(K
1
n)

0 0 0

1
1
1
. . .
3
2
1 0 1 2 3 4 5 6 7 . . . . . .
3
2
1 0 1 2 3 4 5 6 7 . . . . . .
3
2
1 0 1 2 3 4 5 6 7 . . .
n n n

a b c
d e f
FIGURE 7.5 Six functionally related filter impulse responses: (a) reference response; (b) sign reversal;
(c) and (d) order reversal (differing by the delay introduced); (e) modulation; and (f) order reversal and
modulation.
 7.1 ■ Background 469

Filters h3(n) and h4(n) in Figs. 7.5(c) and (d) are order-reversed versions of Order reversal is often
called time reversal when
h1(n): the input sequence is a
sampled analog signal.
h3(n) = h1(-n) (7.1-6)
h4(n) = h1(K - 1 - n) (7.1-7)

Filter h3(n) is a reflection of h1(n) about the vertical axis; filter h4(n) is a re-
flected and translated (i.e., shifted) version of h1(n). Neglecting translation,
the responses of the two filters are identical. Filter h5(n) in Fig. 7.5(e), which is
defined as
h5(n) = (-1)nh1(n) (7.1-8)

is called a modulated version of h1(n). Because modulation changes the signs

of all odd-indexed coefficients [i.e., the coefficients for which n is odd in
Fig. 7.5(e)], h5(1) = -h1(1) and h5(3) = -h1(3), while h5(0) = h1(0) and
h5(2) = h1(2). Finally, the sequence shown in Fig. 7.5(f) is an order-reversed
version of h1(n) that is also modulated:

h6(n) = (-1)nh1(K - 1 - n) (7.1-9)

This sequence is included to illustrate the fact that sign reversal, order rever-
sal, and modulation are sometimes combined in the specification of the rela-
tionship between two filters.
With this brief introduction to digital signal filtering, consider the two-band
subband coding and decoding system in Fig. 7.6(a). As indicated in the figure,
A filter bank is a collec-
the system is composed of two filter banks, each containing two FIR filters of tion of two or more filters.
the type shown in Fig. 7.4(a). Note that each of the four FIR filters is depicted

flp(n) a
b

h0(n) 2T 2c
g0(n)
FIGURE 7.6
(a) A two-band
f(n) Analysis filter bank Synthesis filter bank  fˆ(n) subband coding
and decoding
system, and (b) its

h1(n) 2T 2c
g1(n)
spectrum splitting
properties.
fhp(n)

H0() H1()

Low band High band


0 / 2 
470 Chapter 7 ■ Wavelets and Multiresolution Processing

as a single block in Fig. 7.6(a), with the impulse response of each filter (and the
convolution symbol) written inside it. The analysis filter bank, which includes
filters h0 (n) and h1(n), is used to break input sequence f(n) into two half-
length sequences flp(n) and fhp(n), the subbands that represent the input. Note
that filters h0 (n) and h1(n) are half-band filters whose idealized transfer char-
acteristics, H0 and H1, are shown in Fig. 7.6(b). Filter h0 (n) is a lowpass filter
whose output, subband flp(n), is called an approximation of f(n); filter h1(n) is
a highpass filter whose output, subband fhp (n), is called the high frequency or
detail part of f(n). Synthesis bank filters g0 (n) and g1 (n) combine flp(n) and
fhp(n) to produce fN(n). The goal in subband coding is to select h0 (n), h1(n),
g0 (n), and g1 (n) so that fN(n) = f(n). That is, so that the input and output of the
subband coding and decoding system are identical. When this is accomplished,
the resulting system is said to employ perfect reconstruction filters.
By real-coefficient, we There are many two-band, real-coefficient, FIR, perfect reconstruction fil-
mean that the filter
coefficients are real (not ter banks described in the filter bank literature. In all of them, the synthesis fil-
complex) numbers. ters are modulated versions of the analysis filters—with one (and only one)
synthesis filter being sign reversed as well. For perfect reconstruction, the im-
pulse responses of the synthesis and analysis filters must be related in one of
the following two ways:
Equations (7.1-10) g0 (n) = (-1)nh1(n)
through (7.1-14) are (7.1-10)
described in detail in the
filter bank literature (see, g1 (n) = (-1)n + 1h0 (n)
for example, Vetterli and
Kovacevic [1995]). or

g0 (n) = (-1)n + 1h1(n)

(7.1-11)
g1 (n) = (-1)nh0 (n)
Filters h0 (n), h1(n), g0 (n), and g1(n) in Eqs. (7.1-10) and (7.1-11) are said to be
cross-modulated because diagonally opposed filters in the block diagram of
Fig. 7.6(a) are related by modulation [and sign reversal when the modulation
factor is -( -1)n or (-1)n + 1]. Moreover, they can be shown to satisfy the fol-
lowing biorthogonality condition:

8hi (2n - k), gj (k)9 = d(i - j)d(n), i, j = 50, 16 (7.1-12)

Here, 8hi (2n - k), gj (k)9 denotes the inner product of hi (2n - k) and gj (k).†
When i is not equal to j, the inner product is 0; when i and j are equal, the
product is the unit discrete impulse function, d(n). Biorthogonality will be con-
sidered again in Section 7.2.1.
Of special interest in subband coding—and in the development of the fast
wavelet transform of Section 7.4—are filters that move beyond biorthogonality
and require

The vector inner product of sequences f1 (n) and f2 (n) is 8f1, f29 = a f*1(n)f2 (n), where the * denotes
†

the complex conjugate operation. If f1 (n) and f2 (n) are real, 8f1, f29 = 8f2, f19.
n
 7.1 ■ Background 471

8gi (n), gj (n + 2m)9 = d(i - j)d(m), i, j = 50, 16 (7.1-13)

which defines orthonormality for perfect reconstruction filter banks. In addi-

tion to Eq. (7.1-13), orthonormal filters can be shown to satisfy the following
two conditions:

g1 (n) = (-1)ng0 (Keven - 1 - n)

(7.1-14)
hi (n) = gi (Keven - 1 - n), i = 50, 16

where the subscript on Keven is used to indicate that the number of filter coef-
ficients must be divisible by 2 (i.e., an even number). As Eq. (7.1-14) indicates,
synthesis filter g1 is related to g0 by order reversal and modulation. In addi-
tion, both h0 and h1 are order-reversed versions of synthesis filters, g0 and g1,
respectively. Thus, an orthonormal filter bank can be developed around the
impulse response of a single filter, called the prototype; the remaining filters
can be computed from the specified prototype’s impulse response. For
biorthogonal filter banks, two prototypes are required; the remaining filters
can be computed via Eq. (7.1-10) or (7.1-11). The generation of useful proto-
type filters, whether orthonormal or biorthogonal, is beyond the scope of this
chapter. We simply use filters that have been presented in the literature and
provide references for further study.
Before concluding the section with a 2-D subband coding example, we note
that 1-D orthonormal and biorthogonal filters can be used as 2-D separable
filters for the processing of images. As can be seen in Fig. 7.7, the separable fil-
ters are first applied in one dimension (e.g., vertically) and then in the other
(e.g., horizontally) in the manner introduced in Section 2.6.7. Moreover, down-
sampling is performed in two stages—once before the second filtering opera-
tion to reduce the overall number of computations. The resulting filtered

FIGURE 7.7

h0 (n) 2T a(m,n) A two-
Columns
dimensional, four-
(along n) band filter bank

h0 (m) 2T for subband
Rows image coding.
(along m)
h1 (n) 2T dV(m,n)

f(m,n) Columns

h0 (n) 2T dH(m,n)

Columns

h1 (m) 2T

Rows

h1 (n) 2T dD(m,n)

Columns
472 Chapter 7 ■ Wavelets and Multiresolution Processing

outputs, denoted a(m, n), dV(m, n), d H(m, n), and dD(m, n) in Fig. 7.7, are
called the approximation, vertical detail, horizontal detail, and diagonal detail
subbands of the input image, respectively. These subbands can be split into
four smaller subbands, which can be split again, and so on—a property that
will be described in greater detail in Section 7.4.

EXAMPLE 7.2: ■ Figure 7.8 shows the impulse responses of four 8-tap orthonormal filters.
A four-band The coefficients of prototype synthesis filter g0 (n) for 0 … n … 7 [in Fig. 7.8(c)]
subband coding of
are defined in Table 7.1 (Daubechies [1992]). The coefficients of the remaining
the vase in Fig. 7.1.
orthonormal filters can be computed using Eq. (7.1-14). With the help of Fig.
7.5, note (by visual inspection) the cross modulation of the analysis and synthe-
sis filters in Fig. 7.8. It is relatively easy to show numerically that the filters are

TABLE 7.1
n g0(n)
Daubechies 8-tap
orthonormal filter 0 0.23037781
coefficients for 1 0.71484657
g0(n) (Daubechies 2 0.63088076
[1992]). 3 -0.02798376
4 -0.18703481
5 0.03084138
6 0.03288301
7 -0.01059740

a b h0(n) h1(n)
c d 1 1

FIGURE 7.8
0.5 0.5
The impulse
responses of four
8-tap Daubechies 0 0
orthonormal
filters. See
0.5
0.5
Table 7.1 for the
values of g0 (n) for
1 n
1 n
0 … n … 7. 0 2 4 6 8 0 2 4 6 8

g0 (n) g1(n)
1 1

0.5 0.5

0 0

0.5
0.5

1 n
1 n
0 2 4 6 8 0 2 4 6 8
 7.1 ■ Background 473

a b
c d
FIGURE 7.9
A four-band split
of the vase in
Fig. 7.1 using the
subband coding
system of Fig. 7.7.
The four
subbands that
result are the
(a) approximation,
(b) horizontal
detail, (c) vertical
detail, and
(d) diagonal detail
subbands.

both biorthogonal (they satisfy Eq. 7.1-12) and orthonormal (they satisfy Eq. 7.1-
13). As a result, the Daubechies 8-tap filters in Fig. 7.8 support error-free recon-
struction of the decomposed input.
A four-band split of the 512 * 512 image of a vase in Fig. 7.1, based on the
filters in Fig. 7.8, is shown in Fig. 7.9. Each quadrant of this image is a subband
of size 256 * 256. Beginning with the upper-left corner and proceeding in a
clockwise manner, the four quadrants contain approximation subband a, hori-
zontal detail subband d H, diagonal detail subband dD, and vertical detail sub-
band dV, respectively. All subbands, except the approximation subband in
Fig. 7.9(a), have been scaled to make their underlying structure more visible.
See Section 4.5.4 for
Note the visual effects of aliasing that are present in Figs. 7.9(b) and (c)—the dH more on aliasing.
and dV subbands. The wavy lines in the window area are due to the downsam-
pling of a barely discernable window screen in Fig. 7.1. Despite the aliasing, the
original image can be reconstructed from the subbands in Fig. 7.9 without
error. The required synthesis filters, g0 (n) and g1 (n), are determined from
Table 7.1 and Eq. (7.1-14), and incorporated into a filter bank that roughly
mirrors the system in Fig. 7.7. In the new filter bank, filters hi (n) for i = 50, 16
are replaced by their gi (n) counterparts, and upsamplers and summers are
added. ■
474 Chapter 7 ■ Wavelets and Multiresolution Processing

7.1.3 The Haar Transform

The third and final imaging-related operation with ties to multiresolution
analysis that we will look at is the Haar transform (Haar [1910]). Within
the context of this chapter, its importance stems from the fact that its basis
functions (defined below) are the oldest and simplest known orthonormal
wavelets. They will be used in a number of examples in the sections that
follow.
With reference to the discussion in Section 2.6.7, the Haar transform can be
expressed in the following matrix form

T = HFH T (7.1-15)

where F is an N * N image matrix, H is an N * N Haar transformation

matrix, and T is the resulting N * N transform. The transpose is required
because H is not symmetric; in Eq. (2.6-38) of Section 2.6.7, the transforma-
tion matrix is assumed to be symmetric. For the Haar transform, H contains
the Haar basis functions, hk(z). They are defined over the continuous, closed
interval z H [0, 1] for k = 0, 1, 2, Á , N - 1, where N = 2 n. To generate H,
we define the integer k such that k = 2 p + q - 1, where 0 … p … n - 1,
q = 0 or 1 for p = 0, and 1 … q … 2 p for p Z 0. Then the Haar basis func-
tions are

1
h0 (z) = h00 (z) = , z H [0, 1] (7.1-16)
1N
and

2 p>2 (q - 1)>2 p … z 6 (q - 0.5)>2 p

1
hk(z) = hpq (z) = c -2 p>2 (q - 0.5)>2 p … z 6 q>2 p
1N (7.1-17)
0 otherwise, z H [0, 1]

The ith row of an N * N Haar transformation matrix contains the elements

of hi (z) for z = 0>N, 1>N, 2>N, Á , (N - 1)>N. For instance, if N = 2, the
first row of the 2 * 2 Haar matrix is computed using h0(z) with z = 0>2, 1>2.
From Eq. (7.1-16), h0 (z) is equal to 1> 1 2, independent of z, so the first row of
H 2 has two identical 1> 1 2 elements. The second row is obtained by computing
h1(z) for z = 0>2, 1>2. Because k = 2 p + q - 1, when k = 1, p = 0 and
q = 1. Thus, from Eq. (7.1-17), h1(0) = 2 0> 12 = 1> 12, h1 (1>2) = -2 0> 12
= -1> 1 2, and the 2 * 2 Haar matrix is

1 1 1
H2 = B R (7.1-18)
12 1 -1
 7.1 ■ Background 475

If N = 4, k, q, and p assume the values

k p q
0 0 0
1 0 1
2 1 1
3 1 2

and the 4 * 4 transformation matrix, H 4, is

1 1 1 1
1 1 1 -1 -1
H4 = D T (7.1-19)
14 1 2 -12 0 0
0 0 12 -12

Our principal interest in the Haar transform is that the rows of H 2 can be used
to define the analysis filters, h0 (n) and h1(n), of a 2-tap perfect reconstruction
filter bank (see the previous section), as well as the scaling and wavelet vectors
(defined in Sections 7.2.2 and 7.2.3, respectively) of the simplest and oldest
wavelet transform (see Example 7.10 in Section 7.4). Rather than concluding
the section with the computation of a Haar transform, we close with an exam-
ple that illustrates the influence of the decomposition methods that have been
considered to this point on the methods that will be developed in the remainder
of the chapter.

■ Figure 7.10(a) shows a decomposition of the 512 * 512 image in Fig. 7.1 EXAMPLE 7.3:
that combines the key features of pyramid coding, subband coding, and the Haar functions in
Haar transform (the three techniques we have discussed so far). Called the a discrete wavelet
transform.
discrete wavelet transform (and developed later in the chapter), the represen-
tation is characterized by the following important features:
1. With the exception of the subimage in the upper-left corner of Fig. 7.10(a),
the local histograms are very similar. Many of the pixels are close to zero.
Because the subimages (except for the subimage in the upper-left corner)
have been scaled to make their underlying structure more visible, the dis-
played histograms are peaked at intensity 128 (the zeroes have been
scaled to mid-gray). The large number of zeroes in the decomposition
makes the image an excellent candidate for compression (see Chapter 8).
2. In a manner that is similar to the way in which the levels of the prediction
residual pyramid of Fig. 7.3(b) were used to create approximation images
of differing resolutions, the subimages in Fig. 7.10(a) can be used to con-
struct both coarse and fine resolution approximations of the original
vase image in Fig. 7.1. Figures 7.10(b) through (d), which are of size
476 Chapter 7 ■ Wavelets and Multiresolution Processing

a
b c d
FIGURE 7.10
(a) A discrete
wavelet transform
using Haar H 2
basis functions. Its
local histogram
variations are also
shown. (b)–(d)
Several different
approximations
(64 * 64,
128 * 128, and
256 * 256) that
can be obtained
from (a).

64 * 64, 128 * 128, and 256 * 256, respectively, were generated from
the subimages in Fig. 7.10(a). A perfect 512 * 512 reconstruction of the
original image is also possible.
3. Like the subband coding decomposition in Fig. 7.9, a simple real-coefficient,
FIR filter bank of the form given in Fig. 7.7 was used to produce Fig. 7.10(a).
After the generation of a four subband image like that of Fig. 7.9, the
256 * 256 approximation subband was decomposed and replaced by four
128 * 128 subbands (using the same filter bank), and the resulting approx-
imation subband was again decomposed and replaced by four 64 * 64 sub-
bands. This process produced the unique arrangement of subimages that
 7.2 ■ Multiresolution Expansions 477

characterizes discrete wavelet transforms. The subimages in Fig. 7.10(a)

become smaller in size as you move from the lower-right-hand to upper-
left-hand corner of the image.
4. Figure 7.10(a) is not the Haar transform of the image in Fig. 7.1. Although
the filter bank coefficients that were used to produce this decomposition
were taken from Haar transformation matrix H 2, a variety of othronormal
and biorthogonal filter bank coefficients can be used in discrete wavelet
transforms.
5. As will be shown in Section 7.4, each subimage in Fig. 7.10(a) represents a
specific band of spatial frequencies in the original image. In addition,
many of the subimages demonstrate directional sensitivity [e.g., the
subimage in the upper-right corner of Fig. 7.10(a) captures horizontal edge
information in the original image].
Considering this impressive list of features, it is remarkable that the discrete
wavelet transform of Fig. 7.10(a) was generated using two 2-tap digital filters
with a total of four filter coefficients. ■

7.2 Multiresolution Expansions

The previous section introduced three well-known imaging techniques that
play an important role in a mathematical framework called multiresolution
analysis (MRA). In MRA, a scaling function is used to create a series of ap-
proximations of a function or image, each differing by a factor of 2 in resolu-
tion from its nearest neighboring approximations. Additional functions, called
wavelets, are then used to encode the difference in information between adja-
cent approximations.

7.2.1 Series Expansions

A signal or function f(x) can often be better analyzed as a linear combination
of expansion functions

f(x) = a ak wk (x) (7.2-1)

k

where k is an integer index of a finite or infinite sum, the ak are real-valued

expansion coefficients, and the wk(x) are real-valued expansion functions. If
the expansion is unique—that is, there is only one set of ak for any given f(x)—
the wk(x) are called basis functions, and the expansion set, E wk(x) F , is called a
basis for the class of functions that can be so expressed. The expressible func-
tions form a function space that is referred to as the closed span of the expan-
sion set, denoted

V = Span E wk(x) F (7.2-2)

k

To say that f(x) H V means that f(x) is in the closed span of E wk(x) F and can
be written in the form of Eq. (7.2-1).
478 Chapter 7 ■ Wavelets and Multiresolution Processing

For any function space V and corresponding expansion set E wk(x) F , there is a
set of dual functions denoted E wk(x) F that can be used to compute the ak coeffi-
'

cients of Eq. (7.2-1) for any f(x) H V. These coefficients are computed by taking
'
the integral inner products† of the dual wk(x) and function f(x). That is,

ak = 8wk(x), f(x)9 =
' '
w*k(x)f(x) dx (7.2-3)
L
where the * denotes the complex conjugate operation. Depending on the or-
thogonality of the expansion set, this computation assumes one of three possi-
ble forms. Problem 7.10 at the end of the chapter illustrates the three cases
using vectors in two-dimensional Euclidean space.

Case 1: If the expansion functions form an orthonormal basis for V,

meaning that

8wj (x), wk (x)9 = djk = b

0 j Z k
(7.2-4)
1 j = k
'
the basis and its dual are equivalent. That is, wk(x) = wk(x) and Eq. (7.2-3)
becomes
ak = 8wk(x), f(x)9 (7.2-5)

The ak are computed as the inner products of the basis functions and f(x).

Case 2: If the expansion functions are not orthonormal, but are an orthog-
onal basis for V, then

8wj (x), wk (x)9 = 0 j Z k (7.2-6)

and the basis functions and their duals are called biorthogonal. The ak are
computed using Eq. (7.2-3), and the biorthogonal basis and its dual are
such that

8wj (x), wk(x)9 = djk = b

' 0 j Z k
(7.2-7)
1 j = k
Case 3: If the expansion set is not a basis for V, but supports the expan-
sion defined in Eq. (7.2-1), it is a spanning set in which there is more than
one set of ak for any f(x) H V. The expansion functions and their duals are
said to be overcomplete or redundant. They form a frame in which‡

A7f(x)7 2 … a ƒ 8wk(x), f(x)9 ƒ 2 … B7f(x)7 2 (7.2-8)

k

The integral inner product of two real or complex-valued functions f(x) and g(x) is 8f(x), g(x)9 =
†

f*(x)g(x) dx. If f(x) is real, f*(x) = f(x) and 8f(x), g(x)9 = f(x)g(x) dx.
L L
The norm of f(x), denoted 7f(x)7, is defined as the square root of the absolute value of the inner prod-
‡

uct of f(x) with itself.

 7.2 ■ Multiresolution Expansions 479

for some A 7 0, B 6 q , and all f(x) H V. Dividing this equation by the

norm squared of f(x), we see that A and B “frame” the normalized inner
products of the expansion coefficients and the function. Equations similar
to (7.2-3) and (7.2-5) can be used to find the expansion coefficients for
frames. If A = B, the expansion set is called a tight frame and it can be
shown that (Daubechies [1992])

8wk(x), f(x)9wk(x)
1
Aa
f(x) = (7.2-9)
k

Except for the A-1 term, which is a measure of the frame’s redundancy,
this is identical to the expression obtained by substituting Eq. (7.2-5) (for
orthonormal bases) into Eqs. (7.2-1).
7.2.2 Scaling Functions
Consider the set of expansion functions composed of integer translations and
binary scalings of the real, square-integrable function w(x); this is the set
E wj, k(x) F , where
wj, k(x) = 2j>2w(2jx - k) (7.2-10)

for all j, k H Z and w(x) H L2(R).† Here, k determines the position of wj, k(x)
along the x-axis, and j determines the width of wj, k(x)—that is, how broad or
narrow it is along the x-axis. The term 2 j>2 controls the amplitude of the func-
tion. Because the shape of wj, k(x) changes with j, w(x) is called a scaling function.
By choosing w(x) properly, E wj, k(x) F can be made to span L2(R), which is the
set of all measurable, square-integrable functions.
If we restrict j in Eq. (7.2-10) to a specific value, say j = j0 , the resulting
expansion set, E wj0, k(x) F , is a subset of E wj, k(x) F that spans a subspace of L2(R).
Using the notation of the previous section, we can define that subspace as

Vj0 = Span E wj0 , k(x) F (7.2-11)

k

That is, Vj0 is the span of wj0 , k(x) over k. If f(x) H Vj0, we can write

f(x) = a ak wj0 , k(x) (7.2-12)

k
More generally, we will denote the subspace spanned over k for any j as

Vj = Span E wj, k(x) F (7.2-13)

k

As will be seen in the following example, increasing j increases the size of Vj,
allowing functions with smaller variations or finer detail to be included in the
subspace. This is a consequence of the fact that, as j increases, the wj, k(x) that
are used to represent the subspace functions become narrower and separated
by smaller changes in x.
†
The notation L2(R), where R is the set of real numbers, denotes the set of measurable, square-integrable,
one-dimensional functions; Z is the set of integers.
480 Chapter 7 ■ Wavelets and Multiresolution Processing

EXAMPLE 7.4: ■ Consider the unit-height, unit-width scaling function (Haar [1910])
The Haar scaling
function.
1 0 … x 6 1
w(x) = b (7.2-14)
0 otherwise

Figures 7.11(a) through (d) show four of the many expansion functions that
can be generated by substituting this pulse-shaped scaling function into
Eq. (7.2-10). Note that the expansion functions for j = 1 in Figs. 7.11(c) and
(d) are half as wide as those for j = 0 in Figs. 7.11(a) and (b). For a given in-
terval on x, we can define twice as many V1 scaling functions as V0 scaling func-
tions (e.g., w1, 0 and w1, 1 of V1 versus w0, 0 of V0 for the interval 0 … x 6 1).
Figure 7.11(e) shows a member of subspace V1. This function does not be-
long to V0, because the V0 expansion functions in 7.11(a) and (b) are too
coarse to represent it. Higher-resolution functions like those in 7.11(c) and (d)

a b w0, 0 (x)  w (x) w0, 1(x)  w (x
1)

c d
e f
1 1
FIGURE 7.11
Some Haar
scaling functions.
0 0

x x
0 1 2 3 0 1 2 3

w1, 0(x)  2 w (2x) w1, 1(x)  2 w(2x
1)

1 1

0 0

x x
0 1 2 3 0 1 2 3

f(x) H V1 w0, 0 (x) H V1

1 1, 1 1
w1, 1/ 2

0.25 1, 4
0 0
0.5 1, 0 w1,0/ 2
x x
0 1 2 3 0 1 2 3
 7.2 ■ Multiresolution Expansions 481

are required. They can be used, as shown in (e), to represent the function by
the three-term expansion
f(x) = 0.5w1, 0 (x) + w1, 1 (x) - 0.25w1, 4 (x)
To conclude the example, Fig. 7.11(f) illustrates the decomposition of
w0, 0 (x) as a sum of V1 expansion functions. In a similar manner, any V0 expan-
sion function can be decomposed using
1 1
w0, k(x) = w1, 2k (x) + w1, 2k + 1 (x)
12 12
Thus, if f(x) is an element of V0, it is also an element of V1. This is because all
V0 expansion functions are contained in V1. Mathematically, we write that V0 is
a subspace of V1, denoted V0 ( V1. ■

The simple scaling function in the preceding example obeys the four funda-
mental requirements of multiresolution analysis (Mallat [1989a]):
MRA Requirement 1: The scaling function is orthogonal to its integer
translates.
This is easy to see in the case of the Haar function, because whenever it has a
value of 1, its integer translates are 0, so that the product of the two is 0. The
Haar scaling function is said to have compact support, which means that it is
0 everywhere outside a finite interval called the support. In fact, the width of
the support is 1; it is 0 outside the half open interval [0, 1). It should be noted
that the requirement for orthogonal integer translates becomes harder to
satisfy as the width of support of the scaling function becomes larger than 1.
MRA Requirement 2: The subspaces spanned by the scaling function at low
scales are nested within those spanned at higher scales.
As can be seen in Fig. 7.12, subspaces containing high-resolution functions
must also contain all lower resolution functions. That is,
V- q ( Á ( V-1 ( V0 ( V1 ( V2 ( Á ( Vq (7.2-15)
Moreover, the subspaces satisfy the intuitive condition that if f(x) H Vj, then
f(2x) H Vj + 1. The fact that the Haar scaling function meets this requirement

V0 V1 V2 FIGURE 7.12
The nested
function spaces
spanned by a
scaling function.

V0
482 Chapter 7 ■ Wavelets and Multiresolution Processing

should not be taken to indicate that any function with a support width of 1
automatically satisfies the condition. It is left as an exercise for the reader
to show that the equally simple function

1 0.25 … x 6 0.75
w(x) = b
0 elsewhere

is not a valid scaling function for a multiresolution analysis (see Problem 7.11).
MRA Requirement 3: The only function that is common to all Vj is f(x) = 0.
If we consider the coarsest possible expansion functions (i.e., j = - q ),
the only representable function is the function of no information. That is,

V- q = 506 (7.2-16)

MRA Requirement 4:Any function can be represented with arbitrary precision.

Though it may not be possible to expand a particular f(x) at an arbitrarily
coarse resolution, as was the case for the function in Fig. 7.11(e), all mea-
surable, square-integrable functions can be represented by the scaling
functions in the limit as j : q . That is,

Vq = E L2(R) F (7.2-17)

Under these conditions, the expansion functions of subspace Vj can be ex-

pressed as a weighted sum of the expansion functions of subspace Vj + 1. Using
Eq. (7.2-12), we let

wj, k (x) = a an wj + 1, n (x)

n

where the index of summation has been changed to n for clarity. Substituting
The an are changed to
hw(n) because they are
for wj + 1, n (x) from Eq. (7.2-10) and changing variable an to hw(n), this becomes
used later (see Section
7.4) as filter bank
coefficients. wj, k (x) = a hw(n)2 (j + 1)>2w(2 j + 1x - n)
n

Because w(x) = w0, 0 (x), both j and k can be set to 0 to obtain the simpler non-
subscripted expression

w(x) = a hw(n)12w(2x - n) (7.2-18)

n

The hw(n) coefficients in this recursive equation are called scaling function co-
efficients; hw is referred to as a scaling vector. Equation (7.2-18) is fundamental
to multiresolution analysis and is called the refinement equation, the MRA
equation, or the dilation equation. It states that the expansion functions of any
subspace can be built from double-resolution copies of themselves—that is,
from expansion functions of the next higher resolution space. The choice of a
reference subspace, V0, is arbitrary.
 7.2 ■ Multiresolution Expansions 483

■ The scaling function coefficients for the Haar function of Eq. (7.2-14) EXAMPLE 7.5:
are hw(0) = hw(1) = 1> 12, the first row of matrix H 2 in Eq. (7.1-18). Thus, Haar scaling
function
Eq. (7.2-18) yields
coefficients.

C 12w(2x) D + C 12w(2x - 1) D
1 1
w(x) =
12 12
This decomposition was illustrated graphically for w0, 0 (x) in Fig. 7.11(f), where
the bracketed terms of the preceding expression are seen to be w1, 0 (x) and
w1, 1 (x). Additional simplification yields w(x) = w(2x) + w(2x - 1). ■

7.2.3 Wavelet Functions

Given a scaling function that meets the MRA requirements of the previous
section, we can define a wavelet function c(x) that, together with its integer
translates and binary scalings, spans the difference between any two adjacent
scaling subspaces, Vj and Vj + 1. The situation is illustrated graphically in Fig. 7.13.
We define the set 5cj, k (x)6 of wavelets

cj, k (x) = 2 j>2c(2 jx - k) (7.2-19)

for all k H Z that span the Wj spaces in the figure. As with scaling functions, we
write
Wj = Span E cj, k (x) F (7.2-20)
k
and note that if f(x) H Wj,

f(x) = a ak cj, k (x) (7.2-21)

k
The scaling and wavelet function subspaces in Fig. 7.13 are related by
Vj + 1 = Vj { Wj (7.2-22)
where { denotes the union of spaces (like the union of sets). The orthogonal
complement of Vj in Vj + 1 is Wj, and all members of Vj are orthogonal to the
members of Wj. Thus,
8wj, k (x), cj, l (x)9 = 0 (7.2-23)
for all appropriate j, k, l H Z.

V2  V1 { W1 V0 { W0 { W1 FIGURE 7.13

The relationship
between scaling
V1  V0 { W0 and wavelet
function spaces.
W1
W0
V0
484 Chapter 7 ■ Wavelets and Multiresolution Processing

We can now express the space of all measurable, square-integrable func-

tions as

L2(R) = V0 { W0 { W1 { Á (7.2-24)

or
L2(R) = V1 { W1 { W2 { Á (7.2-25)
or even

L2(R) = Á { W-2 { W-1 { W0 { W1 { W2 { Á (7.2-26)

which eliminates the scaling function, and represents a function in terms of

wavelets alone [i.e., there are only wavelet function spaces in Eq. (7.2-26)].
Note that if f(x) is an element of V1, but not V0, an expansion using Eq. (7.2-24)
contains an approximation of f(x) using V0 scaling functions. Wavelets from
W0 would encode the difference between this approximation and the actual
function. Equations (7.2-24) through (7.2-26) can be generalized to yield

L2(R) = Vj0 { Wj0 { Wj0 + 1 { Á (7.2-27)

where j0 is an arbitrary starting scale.

Since wavelet spaces reside within the spaces spanned by the next higher
resolution scaling functions (see Fig. 7.13), any wavelet function—like its scal-
ing function counterpart of Eq. (7.2-18)—can be expressed as a weighted sum
of shifted, double-resolution scaling functions. That is, we can write

c(x) = a hc(n)12w(2x - n) (7.2-28)

n

where the hc(n) are called the wavelet function coefficients and hc is the
wavelet vector. Using the condition that wavelets span the orthogonal comple-
ment spaces in Fig. 7.13 and that integer wavelet translates are orthogonal, it
can be shown that hc(n) is related to hw(n) by (see, for example, Burrus,
Gopinath, and Guo [1998])

hc(n) = (-1)n hw(1 - n) (7.2-29)

Note the similarity of this result and Eq. (7.1-14), the relationship governing
the impulse responses of orthonormal subband coding and decoding filters.

EXAMPLE 7.6: ■ In the previous example, the Haar scaling vector was defined as
The Haar wavelet hw(0) = hw(1) = 1> 12. Using Eq. (7.2-29), the corresponding wavelet
function vector is hc(0) = (-1)0hw(1 - 0) = 1> 12 and hc(1) = (-1)1hw(1 - 1)
coefficients.
= -1> 12. Note that these coefficients correspond to the second row of ma-
trix H2 in Eq. (7.1-18). Substituting these values into Eq. (7.2-28), we get
 7.2 ■ Multiresolution Expansions 485

c(x) = w(2x) - w(2x - 1), which is plotted in Fig. 7.14(a). Thus, the Haar
wavelet function is
1 0 … x 6 0.5
c(x) = c -1 0.5 … x 6 1 (7.2-30)
0 elsewhere
Using Eq. (7.2-19), we can now generate the universe of scaled and translated
Haar wavelets.Two such wavelets, c0, 2 (x) and c1, 0 (x), are plotted in Figs. 7.14(b)
and (c), respectively. Note that wavelet c1, 0 (x) for space W1 is narrower than
c0, 2 (x) for W0; it can be used to represent finer detail.
Figure 7.14(d) shows a function of subspace V1 that is not in subspace V0. This
function was considered in an earlier example [see Fig. 7.11(e)]. Although the
function cannot be represented accurately in V0, Eq. (7.2-22) indicates that it can
be expanded using V0 and W0 expansion functions. The resulting expansion is
f(x) = fa (x) + fd (x)

c(x)  c0, 0 (x) c0, 2 (x)  c(x
2) a b

c d
1 1 e f
FIGURE 7.14
Haar wavelet
0 0
functions in W0
and W1.

1
1
x x
0 1 2 3 0 1 2 3

c1, 0 (x)  2 c(2x) f (x) H V1  V0 { W0

1 1

0 0

1
1
x x
0 1 2 3 0 1 2 3

fa(x) H V0 fd (x) H W0

3 2/4 w0, 0
1 1
2/8 c0, 2

0 0

1
2/8 w0, 2
1
2/4 c0, 0

x x
0 1 2 3 0 1 2 3
486 Chapter 7 ■ Wavelets and Multiresolution Processing

where
312 12
fa (x) = w0, 0 (x) - w (x)
4 8 0, 2
and
- 12 12
fd (x) = c0, 0 (x) - c (x)
4 8 0, 2
Here, fa (x) is an approximation of f(x) using V0 scaling functions, while fd (x)
is the difference f(x) - fa (x) as a sum of W0 wavelets. The two expansions,
which are shown in Figs. 7.14(e) and (f), divide f(x) in a manner similar to a
lowpass and highpass filter as discussed in connection with Fig. 7.6. The low
frequencies of f(x) are captured in fa (x)—it assumes the average value of
f(x) in each integer interval—while the high-frequency details are encoded in
fd (x). ■

7.3 Wavelet Transforms in One Dimension

We can now formally define several closely related wavelet transformations:
the generalized wavelet series expansion, the discrete wavelet transform, and
the continuous wavelet transform. Their counterparts in the Fourier domain
are the Fourier series expansion, the discrete Fourier transform, and the inte-
gral Fourier transform, respectively. In Section 7.4, we develop a computation-
ally efficient implementation of the discrete wavelet transform called the fast
wavelet transform.

7.3.1 The Wavelet Series Expansions

We begin by defining the wavelet series expansion of function f(x) H L2(R) rel-
ative to wavelet c(x) and scaling function w(x). In accordance with Eq. (7.2-27),
f(x) can be represented by a scaling function expansion in subspace Vj0
[Eq. (7.2-12) defines such an expansion] and some number of wavelet func-
tion expansions in subspaces Wj0, Wj0 + 1, Á [as defined in Eq. (7.2-21)]. Thus,

q
f(x) = a cj0(k)wj0, k (x) + a a dj(k)cj, k (x) (7.3-1)
k j = j0 k

where j0 is an arbitrary starting scale and the cj0(k) and dj (k) are relabeled ak
from Eqs. (7.2-12) and (7.2-21), respectively. The cj0(k) are normally called
approximation and/or scaling coefficients; the dj (k) are referred to as detail
and/or wavelet coefficients. This is because the first sum in Eq. (7.3-1) uses scal-
ing functions to provide an approximation of f(x) at scale j0 [unless f(x) H Vj0
so that the sum of the scaling functions is equal to f(x)]. For each higher scale
j Ú j0 in the second sum, a finer resolution function—a sum of wavelets—is
added to the approximation to provide increasing detail. If the expansion
 7.3 ■ Wavelet Transforms in One Dimension 487

functions form an orthonormal basis or tight frame, which is often the case, the
expansion coefficients are calculated—based on Eqs. (7.2-5) and (7.2-9)—as

cj0(k) = 8f(x), wj0, k (x)9 = f(x)wj0, k (x) dx (7.3-2) Because f is real, no con-
L jugates are needed in the
inner products of Eqs.
and (7.3-2) and (7.3-3).

dj (k) = 8f(x), cj, k (x)9 = f(x)cj, k (x) dx (7.3-3)

L
In Eqs. (7.2-5) and (7.2-9), the expansion coefficients (i.e., the ak) are defined
as inner products of the function being expanded and the expansion functions
being used. In Eqs. (7.3-2) and (7.3-3), the expansion functions are the wj0, k and
cj, k; the expansion coefficients are the cj0 and dj. If the expansion functions
are part of a biorthogonal basis, the w and c terms in these equations must be
' '
replaced by their dual functions, w and c, respectively.

■ Consider the simple function EXAMPLE 7.7:

The Haar wavelet
series expansion
x2 0 … x … 1
y = b of y = x2.
0 otherwise

shown in Fig. 7.15(a). Using Haar wavelets—see Eqs. (7.2-14) and (7.2-30)—
and a starting scale j0 = 0, Eqs. (7.3-2) and (7.3-3) can be used to compute the
following expansion coefficients:

` =
1 1
x3 1 1
c0(0) = x 2w0,0 (x) dx = x2 dx =
L0 L0 3 0 3

1 0.5 1
1
d0(0) = x2c0,0 (x) dx = x 2 dx - x2 dx = -
L0 L0 L0.5 4

0.25 0.5
1
12
d1(0) = x2c1,0 (x) dx = x 2 12 dx - x 2 12 dx = -
L0 L0 L0.25 32

1 0.75 1
312
d1(1) = x 2c1,1 (x) dx = x 2 12 dx - x2 12 dx = -
L0 L0.5 L0.75 32

Substituting these values into Eq. (7.3-1), we get the wavelet series expansion
12
w0,0(x) + c - c0,0(x) d + c - d
1 1 312 c (x) + Á
y = c1,0(x) - 1,1
3
¯˘˚ ˙ ¯˚ 4
˚˘˚˙ 32
¯˚˚˚˚ 32
˚˚˘˚˚˚˚˚˚˙
¯˚˚V0 ˚˚˚˘˚˚˚˚˙ W0 W1
V1 = V0 { W0
¯˚˚˚˚˚˚˚˚˚˚˚˚˘˚˚˚˚˚˚˚˚˚˚˙
V2 = V1 { W1 = V0 { W0 { W1
488 Chapter 7 ■ Wavelets and Multiresolution Processing

a b 1 1
c d y  x2 V0
1/ 3 w0, 0
e f
0.5 0.5
FIGURE 7.15 1/3
A wavelet series
0 0
expansion of
y = x2 using Haar
wavelets.
0.5 x
0.5 x
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

1 1
W0 V1
7/12
0.5
1/4 c0, 0
1/4
1/12
0

1/4

0.5
0.5 x
x
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

1 1
W1 37/48 V2

3 2/ 32 c1, 1

2/ 32 c1, 0
19/48
3/16
7/48
1/16 1/48

1/16

3/16

0.5
x x
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1

The first term in this expansion uses c0 (0) to generate a subspace V0 approxima-
tion of the function being expanded. This approximation is shown in Fig. 7.15(b)
and is the average value of the original function. The second term uses d0 (0) to
refine the approximation by adding a level of detail from subspace W0. The
added detail and resulting V1 approximation are shown in Figs. 7.15(c) and
(d), respectively. Another level of detail is added by the subspace W1 coeffi-
cients d1(0) and d1(1). This additional detail is shown in Fig. 7.15(e), and the
resulting V2 approximation is depicted in 7.15(f). Note that the expansion is
now beginning to resemble the original function. As higher scales (greater lev-
els of detail) are added, the approximation becomes a more precise represen-
tation of the function, realizing it in the limit as j : q . ■

7.3.2 The Discrete Wavelet Transform

Like the Fourier series expansion, the wavelet series expansion of the previous
section maps a function of a continuous variable into a sequence of coeffi-
cients. If the function being expanded is discrete (i.e., a sequence of numbers),
the resulting coefficients are called the discrete wavelet transform (DWT). For
example, if f(n) = f(x0 + n ¢x) for some x0 , ¢x, and n = 0, 1, 2, Á , M - 1,
 7.3 ■ Wavelet Transforms in One Dimension 489

the wavelet series expansion coefficients for f(x) [defined by Eqs. (7.3-2) and
(7.3-3)] become the forward DWT coefficients for sequence f(n):

1
1M a
Ww (j0, k) = f(n)wj0, k (n) (7.3-5)
n
1
1M a
Wc (j, k) = f(n)cj, k (n) for j Ú j0 (7.3-6)
n

The wj0, k (n) and cj, k (n) in these equations are sampled versions of basis func-
tions wj0, k (x) and cj, k (x). For example, wj0, k (n) = wj0, k (xS + n¢xs) for some
xs, ¢xs, and n = 0, 1, 2, Á , M - 1. Thus, we employ M equally spaced sam-
ples over the support of the basis functions (see Example 7.8 below). In accor-
dance with Eq. (7.3-1), the complementary inverse DWT is

q
1 1
f(n) = a
1M k
Ww (j0, k)wj0, k (n) +
1 M ja a Wc (j, k)cj, k (n) (7.3-7)
= j0 k

Normally, we let j0 = 0 and select M to be a power of 2 (i.e., M = 2J) so

that the summations in Eqs. (7.3-5) through (7.3-7) are performed over
n = 0, 1, 2, Á , M - 1, j = 0, 1, 2, Á , J - 1, and k = 0, 1, 2, Á , 2j - 1. For
Haar wavelets, the discretized scaling and wavelet functions employed in the
transform (i.e., the basis functions) correspond to the rows of the M * M
Haar transformation matrix of Section 7.1.3. The transform itself is composed
of M coefficients, the minimum scale is 0, and the maximum scale is J - 1. For
reasons noted in Section 7.3.1 and illustrated in Example 7.6, the coefficients
defined in Eqs. (7.3-5) and (7.3-6) are usually called approximation and detail
coefficients, respectively.
The Ww(j0, k) and Wc(j, k) in Eqs. (7.3-5) to (7.3-7) correspond to the
cj0(k) and dj (k) of the wavelet series expansion in the previous section. (This
change of variables is not necessary but paves the way for the standard nota-
tion used for the continuous wavelet transform of the next section.) Note that
the integrations in the series expansion have been replaced by summations,
and a 1> 1M normalizing factor, reminiscent of the DFT in Section 4.4.1, has
been added to both the forward and inverse expressions. This factor alternate-
ly could be incorporated into the forward or inverse alone as 1>M. Finally, it
should be remembered that Eqs. (7.3-5) through (7.3-7) are valid for ortho-
normal bases and tight frames alone. For biorthogonal bases, the w and c
' '
terms in Eqs. (7.3-5) and (7.3-6) must be replaced by their duals, w and c,
respectively.

■ To illustrate the use of Eqs. (7.3-5) through (7.3-7), consider the discrete EXAMPLE 7.8:
function of four points: f(0) = 1, f(1) = 4, f(2) = -3, and f(3) = 0. Because Computing a one-
M = 4, J = 2 and, with j0 = 0, the summations are performed over dimensional
discrete wavelet
x = 0, 1, 2, 3, j = 0, 1, and k = 0 for j = 0 or k = 0, 1 for j = 1. We will use transform.
the Haar scaling and wavelet functions and assume that the four samples of
490 Chapter 7 ■ Wavelets and Multiresolution Processing

f(x) are distributed over the support of the basis functions, which is 1 in width.
Substituting the four samples into Eq. (7.3-5), we find that

1 3
2 na
Ww(0, 0) = f(n)w0, 0 (n)
=0

C1 # 1 + 4 # 1 - 3 # 1 + 0 # 1D = 1
1
=
2
because w0, 0 (n) = 1 for n = 0, 1, 2, 3. Note that we have employed uniformly
spaced samples of the Haar scaling function for j = 0 and k = 0. The values
correspond to the first row of Haar transformation matrix H4 of Section 7.1.3.
Continuing with Eq. (7.3-6) and similarly spaced samples of cj, k (x), which cor-
respond to rows 2, 3, and 4 of H4, we get

C 1 # 1 + 4 # 1 - 3 # (-1) + 0 # (-1) D = 4
1
Wc(0, 0) =
2

C 1 # 12 + 4 # A - 12 B - 3 # 0 + 0 # 0 D = -1.5 12
1
Wc(1, 0) =
2

C 1 # 0 + 4 # 0 - 3 # 12 + 0 # A - 12 B D = -1.5 12
1
Wc(1, 1) =
2
Thus, the discrete wavelet transform of our simple four-sample function rela-
tive to the Haar wavelet and scaling function is E 1, 4, -1.512, -1.512 F ,
where the transform coefficients have been arranged in the order in which
they were computed.
Equation (7.3-7) lets us reconstruct the original function from its transform.
Iterating through its summation indices, we get

C Ww (0, 0) w0, 0 (n) + Wc (0, 0)c0, 0 (n) + Wc (1, 0) c1, 0 (n)

1
f(n) =
2
+ Wc(1, 1)c1, 1 (n) D

for n = 0, 1, 2, 3. If n = 0, for instance,

C 1 # 1 + 4 # 1 - 1.512 # A 12 B - 1.512 # 0 D = 1
1
f(0) =
2
As in the forward case, uniformly spaced samples of the scaling and wavelet
functions are used in the computation of the inverse. ■

The four-point DWT in the preceding example is an illustration of a two-

scale decomposition of f(n)—that is, j = 50, 16. The underlying assumption
was that starting scale j0 was zero, but other starting scales are possible. It is
left as an exercise for the reader (see Problem 7.16) to compute the single-
scale transform E 2.512, -1.512, -1.512, -1.512 F , which results when the
starting scale is 1. Thus, Eqs. (7.3-5) and (7.3-6) define a “family” of transforms
that differ in starting scale j0.
 7.3 ■ Wavelet Transforms in One Dimension 491

7.3.3 The Continuous Wavelet Transform

The natural extension of the discrete wavelet transform is the continuous
wavelet transform (CWT), which transforms a continuous function into a highly
redundant function of two continuous variables—translation and scale. The re-
sulting transform is easy to interpret and valuable for time-frequency analysis.
Although our interest is in discrete images, the continuous transform is cov-
ered here for completeness.
The continuous wavelet transform of a continuous, square-integrable func-
tion, f(x), relative to a real-valued wavelet, c(x), is defined as

q
Wc(s, t) = f(x)cs,t(x) dx (7.3-8)
L- q
where

1 x - t
cs,t(x) = c¢ ≤ (7.3-9)
1s s

and s and t are called scale and translation parameters, respectively. Given
Wc(s, t), f(x) can be obtained using the inverse continuous wavelet transform

q q
1 cs,t(x)
f(x) = Wc(s, t) dt ds (7.3-10)
Cc L0 L- q s2
where
q
ƒ ° (m) ƒ 2
Cc = dm (7.3-11)
L- q ƒmƒ

and °(m) is the Fourier transform of c(x). Equations (7.3-8) through (7.3-11)
define a reversible transformation as long as the so-called admissibility criterion,
Cc 6 q , is satisfied (Grossman and Morlet [1984]). In most cases, this sim-
ply means that °(0) = 0 and °(m) : 0 as m : q fast enough to make
Cc 6 q .
The preceding equations are reminiscent of their discrete counterparts—
Eqs. (7.2-19), (7.3-1), (7.3-3), (7.3-6), and (7.3-7). The following similarities
should be noted:
1. The continuous translation parameter, t, takes the place of the integer
translation parameter, k.
2. The continuous scale parameter, s, is inversely related to the binary scale
parameter, 2j. This is because s appears in the denominator of
c A (x - t)>s B in Eq. (7.3-9). Thus, wavelets used in continuous transforms
are compressed or reduced in width when 0 6 s 6 1 and dilated or ex-
panded when s 7 1. Wavelet scale and our traditional notion of frequency
are inversely related.
492 Chapter 7 ■ Wavelets and Multiresolution Processing

3. The continuous transform is similar to a series expansion [see Eq. (7.3-1)]

or discrete transform [see Eq. (7.3-6)] in which the starting scale
j0 = - q . This—in accordance with Eq. (7.2-26)—eliminates explicit scal-
ing function dependence, so that the function is represented in terms of
wavelets alone.
4. Like the discrete transform, the continuous transform can be viewed as a
set of transform coefficients, E Wc(s, t) F , that measure the similarity of f(x)
with a set of basis functions, E cs,t(x) F . In the continuous case, however, both
sets are infinite. Because cs,t(x) is real valued and cs,t(x) = c*s,t(x), each
coefficient from Eq. (7.3-8) is the integral inner product, 8f(x), cs,t(x)9, of
f(x) and cs,t(x).

EXAMPLE 7.9: ■ The Mexican hat wavelet,

A one-
dimensional
≤ (1 - x2)e -x >2
2 -1>4 2
continuous c(x) = ¢ p (7.3-12)
wavelet 13
transform.
gets its name from its distinctive shape [see Fig. 7.16(a)]. It is proportional to
the second derivative of the Gaussian probability function, has an average
value of 0, and is compactly supported (i.e., dies out rapidly as ƒ x ƒ : q ). Al-
though it satisfies the admissibility requirement for the existence of continuous,
reversible transforms, there is not an associated scaling function, and the com-
puted transform does not result in an orthogonal analysis. Its most distinguish-
ing features are its symmetry and the existence of the explicit expression of
Eq. (7.3-12).
The continuous, one-dimensional function in Fig. 7.16(a) is the sum of two
Mexican hat wavelets:

f(x) = c1, 10 (x) + c6, 80 (x)

Its Fourier spectrum, shown in Fig. 7.16(b), reveals the close connection be-
tween scaled wavelets and Fourier frequency bands. The spectrum contains
two broad frequency bands (or peaks) that correspond to the two Gaussian-
like perturbations of the function.
Figure 7.16(c) shows a portion (1 … s … 10 and t … 100) of the CWT of
the function in Fig. 7.16(a) relative to the Mexican hat wavelet. Unlike the
Fourier spectrum in Fig. 7.16(b), it provides both spatial and frequency infor-
mation. Note, for example, that when s = 1, the transform achieves a maxi-
mum at t = 10, which corresponds to the location of the c1, 10 (x) component
of f(x). Because the transform provides an objective measure of the similarity
between f(x) and the wavelets for which it is computed, it is easy to see how it
can be used for feature detection. We simply need wavelets that match the fea-
tures of interest. Similar observations can be drawn from the intensity plot in
Fig. 7.16(d), where the absolute value of the transform ƒ Wc(s, t) ƒ is displayed
as intensities between black and white. Note that the continuous wavelet
transform turns a 1-D function into a 2-D result. ■
 7.4 ■ The Fast Wavelet Transform 493

f(x) F(m) a b
1 0.035 c d
FIGURE 7.16
0.03
The continuous
wavelet transform
(c and d) and
Fourier spectrum
(b) of a
continuous 1-D
function (a).
0

0.005

0.4 x 0 m
0 10 80 100 0

20
Wc(s, t)

1
0.8
0.6
0.4
0.2
0 s

0.2

0.4

0.6 10

0.8
10
8 100
6 80
4 60
s 40
2 20 t
1
0 0 0 10 t 80 100

7.4 The Fast Wavelet Transform

The fast wavelet transform (FWT) is a computationally efficient implementa-
tion of the discrete wavelet transform (DWT) that exploits a surprising but
fortunate relationship between the coefficients of the DWT at adjacent scales.
Also called Mallat’s herringbone algorithm (Mallat [1989a, 1989b]), the FWT
resembles the two-band subband coding scheme of Section 7.1.2.
Consider again the multiresolution refinement equation

w(x) = a hw(n) 12w (2x - n) (7.4-1) Equation (7.4-1) is Eq.

(7.2-18) of Section 7.2.2.
n
j
Scaling x by 2 , translating it by k, and letting m = 2k + n gives

w(2 jx - k) = a hw(n)12w A 2(2 jx - k) - n B

n

= a hw(n)12w (2 j + 1x - 2k - n)
m

= a hw(m - 2k)12w (2 j + 1x - m) (7.4-2)

m
494 Chapter 7 ■ Wavelets and Multiresolution Processing

Note that scaling vector hw can be thought of as the “weights” used to expand
w(2jx - k) as a sum of scale j + 1 scaling functions. A similar sequence of
operations—beginning with Eq. (7.2-28)—provides an analogous result for
c(2jx - k). That is,

c(2 jx - k) = a hc(m - 2k)12w(2 j + 1x - m) (7.4-3)

m

where scaling vector hw(n) in Eq. (7.4-2) corresponds to wavelet vector hc(n)
in Eq. (7.4-3).
The wavelet series Now consider Eqs. (7.3-2) and (7.3-3) of Section 7.3.1. They define the
expansion coefficients
become the DWT wavelet series expansion coefficients of continuous function f(x). Substituting
coefficient when f is Eq. (7.2-19)—the wavelet defining equation—into Eq. (7.3-3), we get
discrete. Here, we begin
with the series expansion
coefficients to simplify
the derivation; we will be dj (k) = f(x)2 j>2 c(2 jx - k) dx (7.4-4)
able to substitute freely L
from earlier results (like
the scaling and wavelet
function definitions). which, upon replacing c(2jx - k) with the right side of Eq. (7.4-3), becomes

dj (k) = f(x)2 j>2 B a hc(m - 2k)12w(2 j + 1x - m) R dx (7.4-5)

L m

Interchanging the sum and integral and rearranging terms then gives

dj (k) = a hc(m - 2k) B f(x)2 (j + 1)>2 w(2 j + 1x - m) R (7.4-6)

m L

where the bracketed quantity is cj0(k) of Eq. (7.3-2) with j0 = j + 1 and

k = m. To see this, substitute Eq. (7.2-10) into Eq. (7.3-2) and replace j0 and k
with j + 1 and m, respectively. Therefore, we can write

dj (k) = a hc(m - 2k)cj + 1(m) (7.4-7)

m

and note that the detail coefficients at scale j are a function of the approxima-
tion coefficients at scale j + 1. Using Eqs. (7.4-2) and (7.3-2) as the starting
point of a similar derivation involving the wavelet series expansion (and
DWT) approximation coefficients, we find similarly that

cj (k) = a hw(m - 2k)cj + 1(m) (7.4-8)

m

Because the cj (k) and dj (k) coefficients of the wavelet series expansion be-
come the Ww(j, k) and Wc(j, k) coefficients of the DWT when f(x) is discrete
(see Section 7.3.2), we can write
 7.4 ■ The Fast Wavelet Transform 495

Wc(j, k) = a hc(m - 2k)Ww(j + 1, m) (7.4-9)

m

Ww(j, k) = a hw(m - 2k)Ww(j + 1, m) (7.4-10)

m

Equations (7.4-9) and (7.4-10) reveal a remarkable relationship between

the DWT coefficients of adjacent scales. Comparing these results to Eq. (7.1-7),
we see that both Ww(j, k) and Wc(j, k), the scale j approximation and the de-
tail coefficients, can be computed by convolving Ww(j + 1, k), the scale j + 1
approximation coefficients, with the order-reversed scaling and wavelet vec-
tors, hw(-n) and hc(-n), and subsampling the results. Figure 7.17 summarizes If hw(m - 2k) in
Eq. (7.4-9) is rewritten as
these operations in block diagram form. Note that this diagram is identical to hw(-(2k - m)), we see
the analysis portion of the two-band subband coding and decoding system of that the first minus sign
is responsible for the
Fig. 7.6, with h0 (n) = hw(-n) and h1(n) = hc(-n). Therefore, we can write order reversal [see
Eq. (7.1-6)], the 2k is

Wc(j, k) = hc(-n)
Ww(j + 1, n) `

responsible for the
(7.4-11) subsampling [see Eq.
n = 2k, k Ú 0 (7.1-2)], and m is the
and dummy variable for

Ww(j, k) = hw(-n)
Ww(j + 1, n) `

convolution [see
(7.4-12) Eq. (7.1-7)].
n = 2k, k Ú 0

where the convolutions are evaluated at instants n = 2k for k Ú 0. As will be

shown in Example 7.10, evaluating convolutions at nonnegative, even indices
is equivalent to filtering and downsampling by 2.
Equations (7.4-11) and (7.4-12) are the defining equations for the computa-
tion of the fast wavelet transform. For a sequence of length M = 2J, the num-
ber of mathematical operations involved is on the order of O(M). That is, the
number of multiplications and additions is linear with respect to the length of
the input sequence—because the number of multiplications and additions in-
volved in the convolutions performed by the FWT analysis bank in Fig. 7.17 is
proportional to the length of the sequences being convolved. Thus, the FWT
compares favorably with the FFT algorithm, which requires on the order of
O(M log2 M) operations.
To conclude the development of the FWT, we simply note that the filter
bank in Fig. 7.17 can be “iterated” to create multistage structures for computing
DWT coefficients at two or more successive scales. For example, Fig. 7.18(a)
shows a two-stage filter bank for generating the coefficients at the two highest
scales of the transform. Note that the highest scale coefficients are assumed to
be samples of the function itself. That is, Ww(J, n) = f(n), where J is the highest

FIGURE 7.17

hc(
n) 2T Wc(j, n) An FWT analysis
bank.

Ww( j  1, n)

hw(
n) 2T Ww(j, n)
496 Chapter 7 ■ Wavelets and Multiresolution Processing

a

hc(
n) 2T Wc(J
1, n)
b
FIGURE 7.18
f (n) 
(a) A two-stage or
hc(
n) 2T Wc(J
2, n)
Ww(J, n)
two-scale FWT
analysis bank and
(b) its frequency
hw(
n) 2T Ww(J
1, n)
splitting
characteristics.

hw(
n) 2T Ww(J
2, n)

|H(v)|

VJ
VJ
1

VJ
2 WJ
2 WJ
1

v
0 p/4 p/2 p

scale. [In accordance with Section 7.2.2, f(x) H VJ, where VJ is the scaling
space in which f(x) resides.] The first filter bank in Fig. 7.18(a) splits the origi-
nal function into a lowpass, approximation component, which corresponds to
scaling coefficients Ww(J - 1, n); and a highpass, detail component, correspond-
ing to coefficients Wc(J - 1, n). This is illustrated graphically in Fig. 7.18(b),
where scaling space VJ is split into wavelet subspace WJ - 1 and scaling subspace
VJ - 1. The spectrum of the original function is split into two half-band compo-
nents. The second filter bank of Fig. 7.18(a) splits the spectrum and subspace
VJ - 1, the lower half-band, into quarter-band subspaces WJ - 2 and VJ - 2 with
corresponding DWT coefficients Wc(J - 2, n) and Ww(J - 2, n), respectively.
The two-stage filter bank of Fig. 7.18(a) is extended easily to any number of
scales. A third filter bank, for example, would operate on the Ww(J - 2, n) co-
efficients, splitting scaling space VJ - 2 into two eighth-band subspaces WJ - 3
and VJ - 3. Normally, we choose 2J samples of f(x) and employ P filter banks
(as in Fig. 7.17) to generate a P-scale FWT at scales J - 1, J - 2, Á , J - P.
The highest scale (i.e., J - 1) coefficients are computed first; the lowest scale
(i.e., J - P) last. If function f(x) is sampled above the Nyquist rate, as is usu-
ally the case, its samples are good approximations of the scaling coefficients at
the sampling resolution and can be used as the starting high-resolution scaling
coefficient inputs. In other words, no wavelet or detail coefficients are needed
at the sampling scale. The highest-resolution scaling functions act as unit dis-
crete impulse functions in Eqs. (7.3-5) and (7.3-6), allowing f(n) to be used as
the scaling (approximation) input to the first two-band filter bank (Odegard,
Gopinath, and Burrus [1992]).
 7.4 ■ The Fast Wavelet Transform 497

■ To illustrate the preceding concepts, consider the discrete function f(n) EXAMPLE 7.10:
= 51, 4, -3, 06 from Example 7.8. As in that example, we will compute the Computing a 1-D
transform based on Haar scaling and wavelet functions. Here, however, we will fast wavelet
transform.
not use the basis functions directly, as was done in the DWT of Example 7.8.
Instead, we will use the corresponding scaling and wavelet vectors from
Examples 7.5 and 7.6:

1> 12 n = 0, 1
hw(n) = b (7.4-13)
0 otherwise

and
1> 12 n = 0
hc(n) = c -1> 12 n = 1 (7.4-14)
0 otherwise

These are the functions used to build the FWT filter banks; they provide the filter
coefficients. Note that because Haar scaling and wavelet functions are orthonor-
mal, Eq. (7.1-14) can be used to generate the FWT filter coefficients from a single
prototype filter—like hw(n) in Table 7.2, which corresponds to g0(n) in Eq. (7.1-14):
Since the DWT computed in Example 7.8 was composed of elements
E Ww(0, 0), Wc(0, 0), Wc(1, 0), Wc(1, 1) F , we will compute the corresponding
two-scale FWT for scales j = 50, 16. That is, J = 2 (there are 2J = 22 sam-
ples) and P = 2 (we are working with scales J - 1 = 2 - 1 = 1 and
J - P = 2 - 2 = 0 in that order). The transform will be computed using the
two-stage filter bank of Fig. 7.18(a). Figure 7.19 shows the sequences that re-
sult from the required FWT convolutions and downsamplings. Note that func-
tion f(n) itself is the scaling (approximation) input to the leftmost filter bank.
To compute the Wc(1, k) coefficients that appear at the end of the upper
branch of Fig. 7.19, for example, we first convolve f(n) with hc(-n). As ex-
plained in Section 3.4.2, this requires flipping one of the functions about the ori-
gin, sliding it past the other, and computing the sum of the point-wise product of
the two functions. For sequences 51, 4, -3, 06 and E -1> 12, 1> 12 F , this
produces

E -1> 12, -3> 12, 7> 12, -3> 12, 0 F

where the second term corresponds to index k = 2n = 0. (In Fig. 7.19, under-
lined values represent negative indices, i.e., n 6 0.) When downsampled by

n hW(n) TABLE 7.2

Orthonormal
0 1> 12 Haar filter
1 1> 12 coefficients for
hw(n).
498 Chapter 7 ■ Wavelets and Multiresolution Processing

{
1/ 2,
3/ 2, 7/ 2,
3/ 2, 0}

{
1/ 2, 1/ 2 } 2T Wc(1, n)  {
3/ 2,
3/ 2 }

Ww(2, n)  f(n) Ww(I, n)  {5/ 2,
3/ 2 }
{
1/ 2, 1/ 2} 2T Wc(0, 0)  {4}

 {1, 4,
3, 0}

{
2.5, 4,
1.5}

{1/ 2, 1/ 2 } 2T

{1/ 2, 5/ 2, 1/ 2,
3/ 2, 0}
{1/ 2, 1/ 2 } 2T Ww(0, 0)  {1}

{2.5, 1,
1.5}

FIGURE 7.19 Computing a two-scale fast wavelet transform of sequence 51, 4, -3, 06 using Haar scaling and
wavelet vectors.

taking the even-indexed points, we get Wc(1, k) = E -3> 12, -3> 12 F for
k = 50, 16. Alternatively, we can use Eq. (7.4-12) to compute

Wc(1, k) = hc(-n)
Ww(2, n) ` = hc(-n)
f(n) `

n = 2k, k Ú 0 n = 2k, k Ú 0

= a hc(l - 2k)x(l) `
l k = 0, 1

x(2k + 1) `
1 1
= x(2k) -
12 12 k = 0, 1

Here, we have substituted 2k for n in the convolution and employed l as a dummy

variable of convolution (i.e., for displacing the two sequences relative to one
another). There are only two terms in the expanded sum because there are only
two nonzero values in the order-reversed wavelet vector hc(-n). Substituting
k = 0, we find that Wc(1, 0) = -3> 12; for k = 1, we get Wc(1, 1) = -3> 12.
Thus, the filtered and downsampled sequence is E -3> 12, -3> 12 F , which match-
es the earlier result. The remaining convolutions and downsamplings are per-
formed in a similar manner. ■

As one might expect, a fast inverse transform for the reconstruction of f(n)
from the results of the forward transform can be formulated. Called the
inverse fast wavelet transform (FWT -1), it uses the scaling and wavelet vectors
employed in the forward transform, together with the level j approximation
 7.4 ■ The Fast Wavelet Transform 499

FIGURE 7.20
Wc( j, n) 2c
hc(n)
The FWT -1
synthesis filter
bank.
 WW( j  1, n)

Ww( j, n) 2c
hw(n)

and detail coefficients, to generate the level j + 1 approximation coefficients.

Noting the similarity between the FWT analysis bank in Fig. 7.17 and the two-
band subband analysis portion of Fig. 7.6(a), we can immediately postulate the
required FWT-1 synthesis filter bank. Figure 7.20 details its structure, which is
identical to the synthesis portion of the two-band subband coding and decod-
ing system in Fig. 7.6(a). Equation (7.1-14) of Section 7.1.2 defines the relevant
synthesis filters. As noted there, perfect reconstruction (for two-band ortho-
normal filters) requires gi (n) = hi (-n) for i = 50, 16. That is, the synthesis
and analysis filters must be order-reversed versions of one another. Since the
FWT analysis filters (see Fig. 7.17) are h0(n) = hw(-n) and h1(n) = hc(-n),
the required FWT -1 synthesis filters are g0(n) = h0(-n) = hw(n) and g1(n)
= h1(-n) = hc(n). It should be remembered, however, that it is possible also
to use biorthogonal analysis and synthesis filters, which are not order-reversed
versions of one another. Biorthogonal analysis and synthesis filters are cross-
modulated per Eqs. (7.1-10) and (7.1-11).
The FWT -1 filter bank in Fig. 7.20 implements the computation

Ww(j + 1, k) = hw(k)
Ww2 c (j, k) + hc(k)
Wc2 c (j, k) ` (7.4-15)

kÚ0
where W 2 c signifies upsampling by 2 [i.e., inserting zeros in W as defined by Remember that like in
pyramid coding (see
Eq. (7.1-1) so that it is twice its original length]. The upsampled coefficients are Section 7.1.1), wavelet
filtered by convolution with hw(n) and hc(n) and added to generate a higher transforms can be com-
puted at a user-specified
scale approximation. In essence, a better approximation of sequence f(n) with number of scales. For a
greater detail and resolution is created. As with the forward FWT, the inverse 2J * 2J image, for exam-
ple, there are 1 + log2 J
filter bank can be iterated as shown in Fig. 7.21, where a two-scale structure for possible scales.
computing the final two scales of a FWT -1 reconstruction is depicted. This co-
efficient combining process can be extended to any number of scales and guar-
antees perfect reconstruction of sequence f(n).

FIGURE 7.21
Wc(J
1, n) 2c
hc(n)
A two-stage or
two-scale FWT -1
synthesis bank.
Wc(J
2, n) 2c
hc(n)  Ww(J, n)
Ww(J
1, n)

 2c
hw(n)

Ww(J
2, n) 2c
hw(n)
500 Chapter 7 ■ Wavelets and Multiresolution Processing

EXAMPLE 7.11: ■ Computation of the inverse fast wavelet transform mirrors its forward coun-
Computing a 1-D terpart. Figure 7.22 illustrates the process for the sequence considered in Example
inverse fast
7.10.To begin the calculation, the level 0 approximation and detail coefficients are
upsampled to yield 51, 06 and 54, 06, respectively. Convolution with filters
wavelet

g0(n) = hw(n) = E 1> 12, 1> 12 F g1(n) = hc(n) = E 1> 12, -1> 12 F
transform.
and
produces E 1> 12, 1> 12, 0 F and E 4> 12, -4> 12, 0 F , which when added give
Ww(1, n) = E 5> 12, -3> 12 F . Thus, the level 1 approximation of Fig. 7.22, which
matches the computed approximation in Fig. 7.19, is reconstructed. Continuing in
this manner, f(n) is formed at the right of the second synthesis filter bank. ■

We conclude our discussion of the fast wavelet transform by noting that

while the Fourier basis functions (i.e., sinusoids) guarantee the existence of the
FFT, the existence of the FWT depends upon the availability of a scaling func-
tion for the wavelets being used, as well as the orthogonality (or biorthogonal-
ity) of the scaling function and corresponding wavelets. Thus, the Mexican hat
wavelet of Eq. (7.3-12), which does not have a companion scaling function,
cannot be used in the computation of the FWT. In other words, we cannot con-
struct a filter bank like that of Fig. 7.17 for the Mexican hat wavelet; it does not
satisfy the underlying assumptions of the FWT approach.
Finally, we note that while time and frequency usually are viewed as different
domains when representing functions, they are inextricably linked. When you
try to analyze a function simultaneously in time and frequency, you run into the
following problem: If you want precise information about time, you must accept
some vagueness about frequency, and vice versa. This is the Heisenberg
uncertainty principle applied to information processing. To illustrate the princi-
ple graphically, each basis function used in the representation of a function can
be viewed schematically as a tile in a time-frequency plane. The tile, also called a
Heisenberg cell or Heisenberg box, shows the frequency content of the basis
function that it represents and where the basis function resides in time. Basis
functions that are orthonormal are characterized by nonoverlapping tiles.
Figure 7.23 shows the time-frequency tiles for (a) an impulse function (i.e.,
conventional time domain) basis, (b) a sinusoidal (FFT) basis, and (c) an FWT
{
3/ 2, 0,
3/ 2, 0}
{
1.5, 1.5,
1.5, 1.5, 0}
Wc(1, n)  {
3/ 2,
3/ 2 } 2c
{1/ 2,
1/ 2 }
{4, 0}
{4/ 2,
4/ 2, 0}
f (n)  Ww(2, n)
Wc(0, 0)  {4} 2c
{1/ 2,
1/ 2 } 
Ww(1, n)  {5/ 2,
3/ 2 }  {1, 4,
3, 0}

 2c
{1/ 2, 1/ 2 }
{2.5, 2.5,
1.5,
1.5, 0}

{5/ 2, 0,
3/ 2, 0}
Ww(0, 0)  {1} 2c
{1/ 2, 1/ 2}
{1/ 2, 1/ 2, 0}
{1, 0}

FIGURE 7.22 Computing a two-scale inverse fast wavelet transform of sequence E 1, 4, -1.512, -1.512 F
with Haar scaling and wavelet functions.
Frequency 7.5 ■ Wavelet Transforms in Two Dimensions 501

Time Time Time

a b c
FIGURE 7.23 Time-frequency tilings for the basis functions associated with (a) sampled
data, (b) the FFT, and (c) the FWT. Note that the horizontal strips of equal height
rectangles in (c) represent FWT scales.

basis. Each tile is a rectangular region in Figs. 7.23(a) through (c); the height
and width of the region defines the frequency and time characteristics of the
functions that can be represented using the basis function. Note that the stan-
dard time domain basis in Fig. 7.23(a) pinpoints the instants when events occur
but provides no frequency information [the width of each rectangle in Fig. 7.23(a)
should be considered one instant in time]. Thus, to represent a single frequency
sinusoid as an expansion using impulse basis functions, every basis function is
required. The sinusoidal basis in Fig. 7.23(b), on the other hand, pinpoints the
frequencies that are present in events that occur over long periods but pro-
vides no time resolution [the height of each rectangle in Fig. 7.23(b) should be
considered a single frequency]. Thus, the single frequency sinusoid that was
represented by an infinite number of impulse basis functions can be represented
as an expansion involving one sinusoidal basis function. The time and frequency
resolution of the FWT tiles in Fig. 7.23(c) vary, but the area of each tile (rec-
tangle) is the same.At low frequencies, the tiles are shorter (i.e., have better fre-
quency resolution or less ambiguity regarding frequency) but are wider (which
corresponds to poorer time resolution or more ambiguity regarding time). At
high frequencies, tile width is smaller (so the time resolution is improved) and
tile height is greater (which means the frequency resolution is poorer). Thus,
the FWT basis functions provide a compromise between the two limiting cases
in Fig. 7.23(a) and (b). This fundamental difference between the FFT and FWT
was noted in the introduction to the chapter and is important in the analysis of
nonstationary functions whose frequencies vary in time.

7.5 Wavelet Transforms in Two Dimensions

The one-dimensional transforms of the previous sections are easily extended to
two-dimensional functions like images. In two dimensions, a two-dimensional
scaling function, w(x, y), and three two-dimensional wavelets, c H(x, y),
cV(x, y), and cD(x, y), are required. Each is the product of two one-
dimensional functions. Excluding products that produce one-dimensional results,
like w(x)c(x), the four remaining products produce the separable scaling function
w(x, y) = w(x)w(y) (7.5-1)
502 Chapter 7 ■ Wavelets and Multiresolution Processing

and separable, “directionally sensitive” wavelets

cH(x, y) = c(x)w(y) (7.5-2)
cV(x, y) = w(x)c(y) (7.5-3)
D
c (x, y) = c(x)c(y) (7.5-4)
These wavelets measure functional variations—intensity variations for images—
along different directions: cH measures variations along columns (for exam-
ple, horizontal edges), cV responds to variations along rows (like vertical
edges), and cD corresponds to variations along diagonals. The directional sen-
sitivity is a natural consequence of the separability in Eqs. (7.5-2) to (7.5-4); it
does not increase the computational complexity of the 2-D transform dis-
cussed in this section.
Given separable two-dimensional scaling and wavelet functions, extension
of the 1-D DWT to two dimensions is straightforward. We first define the
scaled and translated basis functions:

wj, m, n(x, y) = 2 j>2w(2 jx - m, 2 jy - n) (7.5-5)

cij, m, n(x, j>2 i j j
y) = 2 c (2 x - m, 2 y - n), i = 5H, V, D6 (7.5-6)

where index i identifies the directional wavelets in Eqs. (7.5-2) to (7.5-4).

Rather than an exponent, i is a superscript that assumes the values H, V, and
Now that we are dealing D. The discrete wavelet transform of image f(x, y) of size M * N is then
with 2-D images, f(x, y)
is a discrete function or
sequence of values and x 1 M-1 N-1
1MN xa a f(x, y)wj0, m, n (x, y)
and y are discrete Ww(j0 , m, n) = (7.5-7)
variables. The scaling and =0 y=0
wavelet functions in
1 M-1 N-1
a f(x, y)cj, m, n (x, y), i = 5H, V, D6
Eq. (7.5-7) and (7.5-8)
Wci (j, m, n) = i
1MN xa
are sampled over their (7.5-8)
support (as was done in
the 1-D case in
=0 y=0
Section 7.3.2).
As in the one-dimensional case, j0 is an arbitrary starting scale and the
Ww(j0 , m, n) coefficients define an approximation of f(x, y) at scale j0. The
Wci (j, m, n) coefficients add horizontal, vertical, and diagonal details for scales
j Ú j0. We normally let j0 = 0 and select N = M = 2 J so that j = 0, 1, 2, Á ,
J - 1 and m = n = 0, 1, 2, Á , 2 j - 1. Given the Ww and Wic of Eqs. (7.5-7) and
(7.5-8), f(x, y) is obtained via the inverse discrete wavelet transform

1
1MN a a Ww(j0, m, n)wj0, m, n(x, y)
f(x, y) =
m n
q
1 i i
1MN i =a a a a W c(j, m, n)cj, m, n (x, y)
+ (7.5-9)
H,V,D j = j0 m n

Like the 1-D discrete wavelet transform, the 2-D DWT can be implemented
using digital filters and downsamplers. With separable two-dimensional scaling
and wavelet functions, we simply take the 1-D FWT of the rows of f(x, y), fol-
lowed by the 1-D FWT of the resulting columns. Figure 7.24(a) shows the
 7.5 ■ Wavelet Transforms in Two Dimensions 503

hc(
m) 2T
D
Wc (j, m, n)
Rows

hc(
n) 2T (along m)

Columns
V
(along n)
hw(
m) 2T Wc(j, m, n)

Ww(j  1, m, n) Rows

H

hc(
m) 2T Wc (j, m, n)
Rows

hw(
n) 2T
Columns

hw(
m) 2T Ww(j, m, n)

Rows

H
Ww( j, m, n) Wc ( j, m, n)

Ww( j  1, m, n)

V D
Wc( j, m, n) Wc ( j, m, n)

D
Wc (j, m, n) 2c
hc(m)

Rows
(along m)  2c
hc(n)

Columns
V
Wc(j, m, n) 2c
hw(m) (along n)

Rows
 Ww(j  1, m, n)

H
Wc (j, m, n) 2c
hc(m)

Rows
 2c
hw(n)

Columns
Ww(j, m, n) 2c
hw(m)

Rows
a
b
c
FIGURE 7.24 The 2-D fast wavelet transform: (a) the analysis filter bank; (b) the
resulting decomposition; and (c) the synthesis filter bank.
504 Chapter 7 ■ Wavelets and Multiresolution Processing

process in block diagram form. Note that, like its one-dimensional counterpart
in Fig. 7.17, the 2-D FWT “filters” the scale j + 1 approximation coefficients
to construct the scale j approximation and detail coefficients. In the two-
dimensional case, however, we get three sets of detail coefficients—the hori-
zontal, vertical, and diagonal details.
The single-scale filter bank of Fig. 7.24(a) can be “iterated” (by tying the ap-
proximation output to the input of another filter bank) to produce a P scale
transform in which scale j is equal to J - 1, J - 2, Á , J - P. As in the one-
dimensional case, image f(x, y) is used as the Ww(J, m, n) input. Convolving
its rows with hw( -n) and hc(-n) and downsampling its columns, we get two
subimages whose horizontal resolutions are reduced by a factor of 2. The high-
pass or detail component characterizes the image’s high-frequency informa-
tion with vertical orientation; the lowpass, approximation component contains
its low-frequency, vertical information. Both subimages are then filtered
columnwise and downsampled to yield four quarter-size output subimages—
Note how Ww, WH V
c , Wc , Ww, W H V D
c , W c , and W c . These subimages, which are shown in the middle of
and WD c are arranged in
Fig. 7.24(b). For each Fig. 7.24(b), are the inner products of f(x, y) and the two-dimensional scaling
scale that is computed, and wavelet functions in Eqs. (7.5-1) through (7.5-4), followed by downsam-
they replace the previous
scale approximation on pling by two in each dimension. Two iterations of the filtering process pro-
which they were based. duces the two-scale decomposition at the far right of Fig. 7.24(b).
Figure 7.24(c) shows the synthesis filter bank that reverses the process just
described. As would be expected, the reconstruction algorithm is similar to the
one-dimensional case. At each iteration, four scale j approximation and detail
subimages are upsampled and convolved with two one-dimensional filters—
one operating on the subimages’ columns and the other on its rows. Addition
of the results yields the scale j + 1 approximation, and the process is repeated
until the original image is reconstructed.

EXAMPLE 7.12: ■ Figure 7.25(a) is a 128 * 128 computer-generated image consisting of 2-D
Computing a 2-D sine-like pulses on a black background. The objective of this example is to
fast wavelet illustrate the mechanics involved in computing the 2-D FWT of this image.
transform.
Figures 7.25(b) through (d) show three FWTs of the image in Fig. 7.25(a). The
2-D filter bank of Fig. 7.24(a) and the decomposition filters shown in Figs. 7.26(a)
and (b) were used to generate all three results.
The scaling and wavelet Figure 7.25(b) shows the one-scale FWT of the image in Fig. 7.25(a). To
vectors used in this
example are described
compute this transform, the original image was used as the input to the filter
later. Our focus here is bank of Fig. 7.24(a). The four resulting quarter-size decomposition outputs (i.e.,
on the mechanics of the
transform computation,
the approximation and horizontal, vertical, and diagonal details) were then
which are independent of arranged in accordance with Fig. 7.24(b) to produce the image in Fig. 7.25(b). A
the filter coefficients
employed.
similar process was used to generate the two-scale FWT in Fig. 7.25(c), but the
input to the filter bank was changed to the quarter-size approximation subim-
age from the upper-left-hand corner of Fig. 7.25(b). As can be seen in Fig.
7.25(c), that quarter-size subimage was then replaced by the four quarter-size
 7.5 ■ Wavelet Transforms in Two Dimensions 505

a b
c d
FIGURE 7.25
Computing a 2-D
three-scale FWT:
(a) the original
image; (b) a one-
scale FWT; (c) a
two-scale FWT;
and (d) a three-
scale FWT.

(now 1 > 16th of the size of the original image) decomposition results that were
generated in the second filtering pass. Finally, Fig. 7.25(d) is the three-scale
FWT that resulted when the subimage from the upper-left-hand corner of Fig.
7.25(c) was used as the filter bank input. Each pass through the filter bank pro-
duced four quarter-size output images that were substituted for the input from
which they were derived. Note the directional nature of the wavelet-based
subimages, WH V D
c , Wc , and Wc , at each scale. ■

The decomposition filters used in the preceding example are part of a well-
known family of wavelets called symlets, short for “symmetrical wavelets.” Al- Recall that the compact
though they are not perfectly symmetrical, they are designed to have the least support of a function is
asymmetry and highest number of vanishing moments† for a given compact the interval in which the
function has non-zero
support (Daubechies [1992]). Figures 7.26(e) and (f) show the fourth-order values.

†
The kth moment of wavelet c(x) is m(k) = 1 xkc(x) dx. Zero moments impact the smoothness of the
scaling and wavelet functions and our ability to represent them as polynomials. An order-N symlet has
N vanishing moments.
506 Chapter 7 ■ Wavelets and Multiresolution Processing

a b h0(n)  hw(
n) h1(n)  h c (
n)

c d 1.2 0.5
e f 1
g
0.8 0
FIGURE 7.26 0.6
Fourth-order
0.4
symlets: (a)–(b)
0.5
decomposition 0.2
filters; (c)–(d) 0
reconstruction
0.2 n
1 n
filters; (e) the 0 2 4 6 8 0 2 4 6 8
one-dimensional
wavelet; (f) the g0(n)  hw(n) g1(n)  h c (n)
one-dimensional 1.2 0.5
scaling function; 1
and (g) one of 0.8
three two- 0
0.6
dimensional
wavelets, cV(x, y). 0.4

0.5
See Table 7.3 for 0.2
the values of 0
hw(n) for
0.2 n
1 n
0 … n … 7. 0 2 4 6 8 0 2 4 6 8

c(x) w(x)
2 1.4
1.5 1.2
1 1
0.8
0.5
0.6
0
0.4

0.5 0.2

1 0

1.5 x
0.2 x
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7

c(x) w( y)
1.5

0.5

0.5

1
8
6 8
4 6
y 2 4 x
2
0
 7.5 ■ Wavelet Transforms in Two Dimensions 507

n hW(n) TABLE 7.3

Orthonormal
0 0.0322 fourth-order
1 -0.0126 symlet filter
2 -0.0992 coefficients for
3 0.2979 hw(n).
4 0.8037 (Daubechies
5 0.4976 [1992].)
6 -0.0296
7 -0.0758

1-D symlets (i.e., wavelet and scaling functions). Figures 7.26(a) through (d)
show the corresponding decomposition and reconstruction filters. The co-
efficients of lowpass reconstruction filter g0(n) = hw(n) for 0 … n … 7 are
given in Table 7.3. The coefficients of the remaining orthonormal filters are
obtained using Eq. (7.1-14). Figure 7.26(g), a low-resolution graphic depiction
of wavelet cV(x, y), is provided as an illustration of how a one-dimensional
scaling and wavelet function can combine to form a separable, two-dimensional
wavelet.
We conclude this section with two examples that demonstrate the useful-
ness of wavelets in image processing. As in the Fourier domain, the basic ap-
proach is to

Step 1. Compute a 2-D wavelet transform of an image.

Step 2. Alter the transform.
Step 3. Compute the inverse transform.

Because the DWT’s scaling and wavelet vectors are used as lowpass and high-
pass filters, most Fourier-based filtering techniques have an equivalent
“wavelet domain” counterpart.

■ Figure 7.27 provides a simple illustration of the preceding three steps. In EXAMPLE 7.13:
Fig. 7.27(a), the lowest scale approximation component of the discrete wavelet Wavelet-based
edge detection.
transform shown in Fig. 7.25(c) has been eliminated by setting its values to
zero. As Fig. 7.27(b) shows, the net effect of computing the inverse wavelet
transform using these modified coefficients is edge enhancement, reminiscent
of the Fourier-based image sharpening results discussed in Section 4.9. Note
how well the transitions between signal and background are delineated, de-
spite the fact that they are relatively soft, sinusoidal transitions. By zeroing the
horizontal details as well—see Figs. 7.27(c) and (d)—we can isolate the verti-
cal edges. ■
508 Chapter 7 ■ Wavelets and Multiresolution Processing

a b
c d
FIGURE 7.27
Modifying a DWT
for edge
detection: (a) and
(c) two-scale
decompositions
with selected
coefficients
deleted; (b) and
(d) the
corresponding
reconstructions.

EXAMPLE 7.14: ■ As a second example, consider the CT image of a human head shown in
Wavelet-based Fig. 7.28(a). As can be seen in the background, the image has been uniformly
noise removal.
corrupted with additive white noise. A general wavelet-based procedure for
denoising the image (i.e., suppressing the noise part) is as follows:
Step 1. Choose a wavelet (e.g. Haar, symlet, Á ) and number of levels
(scales), P, for the decomposition. Then compute the FWT of the noisy
image.
Step 2. Threshold the detail coefficients. That is, select and apply a thresh-
old to the detail coefficients from scales J - 1 to J - P. This can be ac-
complished by hard thresholding, which means setting to zero the elements
whose absolute values are lower than the threshold, or by soft threshold-
ing, which involves first setting to zero the elements whose absolute values
are lower than the threshold and then scaling the nonzero coefficients to-
ward zero. Soft thresholding eliminates the discontinuity (at the threshold)
that is inherent in hard thresholding. (See Chapter 10 for a discussion of
thresholding.)
Step 3. Compute the inverse wavelet transform (i.e., perform a wavelet recon-
struction) using the original approximation coefficients at level J - P and the
modified detail coefficients for levels J - 1 to J - P.
 7.5 ■ Wavelet Transforms in Two Dimensions 509

a b
c d
e f
FIGURE 7.28
Modifying a DWT
for noise removal:
(a) a noisy CT of a
human head; (b),
(c) and (e) various
reconstructions
after thresholding
the detail
coefficients;
(d) and (f) the
information
removed during
the reconstruction
of (c) and (e).
(Original image
courtesy
Vanderbilt
University
Medical Center.)

Figure 7.28(b) shows the result of performing these operations with fourth-
order symlets, two scales (i.e., P = 2), and a global threshold that was de-
termined interactively. Note the reduction in noise and blurring of image
edges. This loss of edge detail is reduced significantly in Fig. 7.28(c), which
510 Chapter 7 ■ Wavelets and Multiresolution Processing

was generated by simply zeroing the highest-resolution detail coefficients

(not thresholding the lower-resolution details) and reconstructing the
image. Here, almost all of the background noise has been eliminated and
Because only the highest the edges are only slightly disturbed. The difference image in Fig. 7.28(d)
resolution detail
coefficients were kept shows the information that is lost in the process. This result was generated
when generating by computing the inverse FWT of the two-scale transform with all but the
Fig. 7.28(d), the inverse
transform is their contri- highest-resolution detail coefficients zeroed. As can be seen, the resulting
bution to the image. In image contains most of the noise in the original image and some of the edge
the same way, Fig. 7.28(f)
is the contribution of all information. Figures 7.28(e) and (f) are included to show the negative effect
the detail coefficients. of deleting all the detail coefficients. That is, Fig. 7.28(e) is a reconstruction
of the DWT in which the details at both levels of the two-scale transform
have been zeroed; Fig. 7.28(f) shows the information that is lost. Note the
significant increase in edge information in Fig. 7.28(f) and the correspond-
ing decrease in edge detail in Fig. 7.28(e). ■

7.6 Wavelet Packets

The fast wavelet transform decomposes a function into a sum of scaling and
wavelet functions whose bandwidths are logarithmically related. That is, the
low frequency content (of the function) is represented using (scaling and
wavelet) functions with narrow bandwidths, while the high-frequency content
is represented using functions with wider bandwidths. If you look along the
frequency axis of the time-frequency plane in Fig. 7.23(c), this is immediately
apparent. Each horizontal strip of constant height tiles, which contains the
basis functions for a single FWT scale, increases logarithmically in height as
you move up the frequency axis. If we want greater control over the partition-
ing of the time-frequency plane (e.g., smaller bands at the higher frequencies),
the FWT must be generalized to yield a more flexible decomposition—called
a wavelet packet (Coifman and Wickerhauser [1992]). The cost of this general-
ization is an increase in computational complexity from O(M) for the FWT to
O(M log2 M) for a wavelet packet.
Consider again the two-scale filter bank of Fig. 7.18(a)—but imagine the de-
composition as a binary tree. Figure 7.29(a) details the structure of the tree, and
links the appropriate FWT scaling and wavelet coefficients [from Fig. 7.18(a)] to
its nodes. The root node is assigned the highest-scale approximation coefficients,

a b Ww(J, n)  f (n) VJ
FIGURE 7.29
An (a) coefficient
Ww(J
1, n) Wc(J
1, n) VJ
1 WJ
1
tree and
(b) analysis tree
for the two-scale
Ww(J
2, n) Wc(J
2, n) VJ
2 WJ
2
FWT analysis
bank of Fig. 7.18.
 7.6 ■ Wavelet Packets 511

which are samples of the function itself, while the leaves inherit the transform’s
approximation and detail coefficient outputs. The lone intermediate node,
Ww(J - 1, n), is a filter bank approximation that is ultimately filtered to become
two leaf nodes. Note that the coefficients of each node are the weights of a linear
expansion that produces a band-limited “piece” of root node f(n). Because any
such piece is an element of a known scaling or wavelet subspace (see Sections
7.2.2 and 7.2.3), we can replace the generating coefficients in Fig. 7.29(a) by the
corresponding subspace.The result is the subspace analysis tree of Fig. 7.29(b).Al-
though the variable W is used to denote both coefficients and subspaces, the two
quantities are distinguishable by the format of their subscripts.
These concepts are further illustrated in Fig. 7.30, where a three-scale FWT
analysis bank, analysis tree, and corresponding frequency spectrum are depicted.
Unlike Fig. 7.18(a), the block diagram of Fig. 7.30(a) is labeled to resemble the
analysis tree in Fig. 7.30(b)—as well as the spectrum in Fig. 7.30(c). Thus, while
the output of the upper-left filter and subsampler is, to be accurate,
Wc(J - 1, n), it has been labeled WJ - 1—the subspace of the function that is
generated by the Wc(J - 1, n) transform coefficients. This subspace corre-
sponds to the upper-right leaf of the associated analysis tree, as well as the right-
most (widest bandwidth) segment of the corresponding frequency spectrum.
Analysis trees provide a compact and informative way of representing mul-
tiscale wavelet transforms. They are simple to draw, take less space than their
corresponding filter and subsampler-based block diagrams, and make it relatively

a

hc(
n) 2T WJ
1
b c
FIGURE 7.30
f(x)H VJ
hc(
n) 2T WJ
2 A three-scale
FWT filter bank:

hw(
n) 2T VJ
1
hc(
n) 2T WJ
3 (a) block diagram;
(b) decomposition
space tree; and

hw(
n) 2T VJ
2 (c) spectrum
splitting
characteristics.

hw(
n) 2T VJ
3

VJ |H(v)|

VJ
VJ
1
VJ
1 WJ
1 VJ
2

VJ
2 WJ
2 VJ
3 WJ
3 WJ
2 WJ
1

VJ
3 WJ
3 v
0 p/8 p/4 p/2 p
512 Chapter 7 ■ Wavelets and Multiresolution Processing

easy to detect valid decompositions. The three-scale analysis tree of Fig. 7.30(b),
for example, makes possible the following three expansion options:

VJ = VJ - 1 { WJ - 1 (7.6-1)
VJ = VJ - 2 { WJ - 2 { WJ - 1 (7.6-2)
VJ = VJ - 3 { WJ - 3 { WJ - 2 { WJ - 1 (7.6-3)

They correspond to the one-, two-, and three-scale FWT decompositions of

Section 7.4 and may be obtained from Eq. (7.2-27) of Section 7.2.3 by letting
j0 = J - P for P = 51, 2, 36. In general, a P-scale FWT analysis tree supports
P unique decompositions.
Analysis trees also are an efficient mechanism for representing wavelet packets,
which are nothing more than conventional wavelet transforms in which the details
are filtered iteratively. Thus, the three-scale FWT analysis tree of Fig. 7.30(b)
becomes the three-scale wavelet packet tree of Fig. 7.31. Note the additional sub-
scripting that is introduced. The first subscript of a double-subscripted node
identifies the scale of the FWT parent node from which it descended. The
second—a variable length string of As and Ds—encodes the path from the par-
ent to the node.An A designates approximation filtering, while a D indicates de-
tail filtering. Subspace WJ - 1, DA, for example, is obtained by “filtering” the scale
J - 1 FWT coefficients (i.e., parent WJ - 1 in Fig. 7.31) through an additional de-
tail filter (yielding WJ - 1, D), followed by an approximation filter (giving WJ - 1, DA).
Figures 7.32(a) and (b) are the filter bank and spectrum splitting characteristics
of the analysis tree in Fig. 7.31. Note that the “naturally ordered” outputs of the
filter bank in Fig. 7.32(a) have been reordered based on frequency content in
Fig. 7.32(b) (see Problem 7.25 for more on “frequency ordered” wavelets).
The three-scale packet tree in Fig. 7.31 almost triples the number of decompo-
sitions (and associated time-frequency tilings) that are available from the three-
scale FWT tree. Recall that in a normal FWT, we split, filter, and downsample the
lowpass bands alone. This creates a fixed logarithmic (base 2) relationship be-
tween the bandwidths of the scaling and wavelet spaces used in the representa-
tion of a function [see Figure 7.30(c)]. Thus, while the three-scale FWT analysis
tree of Fig. 7.30(a) offers three possible decompositions—defined by Eqs. (7.6-1)
to (7.6-3)—the wavelet packet tree of Fig. 7.31 supports 26 different decomposi-
tions. For instance, VJ [and therefore function f(n)] can be expanded as

FIGURE 7.31 VJ
A three-scale
wavelet packet
analysis tree. VJ
1 WJ
1

VJ
2 WJ
2 WJ
1, A WJ
1, D

VJ
3 WJ
3 WJ
2, A WJ
2, D WJ
1, AA WJ
1, AD WJ
1, DA WJ
1, DD

 7.6 ■ Wavelet Packets 513

hc(
n) WJ
1, DD
a
2T
b

hc(
n) 2T WJ
1, D

hw(
n) 2T WJ
1, DA FIGURE 7.32
The (a) filter bank

hc(
n) 2T WJ
1 and (b) spectrum

hc(
n) 2T WJ
1, AD splitting

hw(
n) 2T WJ
1, A characteristics of

hw(
n) 2T WJ
1, AA a three-scale full
f(x)H VJ wavelet packet
analysis tree.

hc(
n) 2T WJ
2, D

hc(
n) 2T WJ
2

hw(
n) 2T WJ
2, A

hw(
n) 2T VJ
1

hc(
n) 2T WJ
3

hw(
n) 2T VJ
2

hw(
n) 2T VJ
3

|H(v)|

VJ
VJ
1 WJ
1
VJ
2 WJ
2 WJ
1, D WJ
1, A

VJ
3 WJ
3 WJ
2, D WJ
2, A WJ
1, DA WJ
1, DD WJ
1, AD WJ
1, AA

v
0 p/8 p/4 p/ 2 p

VJ = VJ - 3 { WJ - 3 { WJ - 2, A { WJ - 2, D { WJ - 1, AA Recall that { denotes

the union of spaces (like
the union of sets). The 26
{ WJ - 1, AD { WJ - 1, DA { WJ - 1, DD (7.6-4) decompositions associated
with Fig. 7.31 are
whose spectrum is shown in Fig. 7.32(b), or determined by various
combinations of nodes
(spaces) that can be
combined to represent
VJ = VJ - 1 { WJ - 1, A { WJ - 1, DA { WJ - 1, DD (7.6-5) the root node (space) at
the top of the tree.
whose spectrum is depicted in Fig. 7.33. Note the difference between this last Eqs. (7.6-4) and (7.6-5)
define two of them.
spectrum and the full packet spectrum of Fig. 7.32(b), or the three-scale FWT

|H(v)| FIGURE 7.33

The spectrum of
the decomposi-
tion in Eq. (7.6-5).

VJ
1 WJ
1,DA WJ
1,DD WJ
1, A

v
0 p/ 2 5p/ 8 3p/4 p
514 Chapter 7 ■ Wavelets and Multiresolution Processing

spectrum of Fig. 7.30(c). In general, P-scale, one-dimensional wavelet packet

transforms (and associated P + 1-level analysis trees) support

D(P + 1) = C D(P) D + 1
2
(7.6-6)

unique decompositions, where D(1) = 1. With such a large number of valid

expansions, packet-based transforms provide improved control over partition-
ing the spectrum of the decomposed function. The cost of this control is an in-
crease in computational complexity [compare the filter bank in Fig. 7.30(a) to
that of Fig. 7.32(a)].
Now consider the two-dimensional, four-band filter bank of Fig. 7.24(a). As
was noted in Section 7.5, it splits approximation Ww(j + 1, m, n) into outputs,
Ww(j, m, n), WH V D
c (j, m, n), W c (j, m, n), and W c (j, m, n). As in the one-
dimensional case, it can be “iterated” to generate P scale transforms for scales
j = J - 1, J - 2, Á , J - P, with Ww(J, m, n) = f(m, n). The spectrum re-
sulting from the first iteration [i.e., using j + 1 = J in Fig. 7.24(a)] is shown in
Fig. 7.34(a). Note that it divides the frequency plane into four equal areas. The
low-frequency quarter-band in the center of the plane coincides with trans-
form coefficients Ww(J - 1, m, n) and scaling space VJ - 1. (This nomenclature
is consistent with the one-dimensional case.) To accommodate the two-
dimensional nature of the input, however, we now have three (rather than one)
wavelet subspaces. They are denoted WH V D
J - 1, W J - 1, and W J - 1 and correspond
to coefficients W c (J - 1, m, n), Wc (J - 1, m, n), and WD
H V
c (J - 1, m, n),
respectively. Figure 7.34(b) shows the resulting four-band, single-scale quaternary
FWT analysis tree. Note the superscripts that link the wavelet subspace desig-
nations to their transform coefficient counterparts.
Figure 7.35 shows a portion of a three-scale, two-dimensional wavelet packet
analysis tree. Like its one-dimensional counterpart in Fig. 7.31, the first subscript of
every node that is a descendant of a conventional FWT detail node is the scale of
that parent detail node.The second subscript—a variable length string of As,Hs,Vs,
and Ds—encodes the path from the parent to the node under consideration. The
node labeled WH J - 1, VD , for example, is obtained by “row/column filtering” the

a b v vertical
p
FIGURE 7.34
The first D V D
WJ
1 WJ
1 WJ
1
decomposition of
a two-dimensional
FWT: (a) the
spectrum and H H
(b) the subspace WJ
1 VJ
1 WJ
1 vhorizontal

p p
analysis tree.
VJ
D V D
WJ
1 WJ
1 WJ
1
H V D
VJ
1 WJ
1 WJ
1 WJ
1

p
 7.6 ■ Wavelet Packets 515

VJ

H V D
VJ
1 WJ
1 WJ
1 WJ
1

H H
VJ
2 WJ
2 WJ
1, V

H V D H H H H H H H H
VJ
3 WJ
3 WJ
3 WJ
3 WJ
2, A WJ
2,H WJ
2, V WJ
2, D WJ
1, VA WJ
1, VH WJ
1, VV WJ
1, VD

FIGURE 7.35 A three-scale, full wavelet packet decomposition tree. Only a portion of the tree is provided.

scale J - 1 FWT horizontal detail coefficients (i.e., parent WH J - 1 in Fig. 7.35)

through an additional detail/approximation filter (yielding W H J - 1, V), followed by a
H
detail/detail filter (giving W J - 1, VD). A P-scale, two-dimensional wavelet pack-
et tree supports
D(P + 1) = C D(P) D + 1
4
(7.6-7)

unique expansions, where D(1) = 1. Thus, the three-scale tree of Fig. 7.35 of-
fers 83,522 possible decompositions. The problem of selecting among them is
the subject of the next example.

■ As noted in the above discussion, a single wavelet packet tree presents nu- EXAMPLE 7.15:
merous decomposition options. In fact, the number of possible decompositions is Two-dimensional
often so large that it is impractical, if not impossible, to enumerate or examine wavelet packet
decompositions.
them individually.An efficient algorithm for finding optimal decompositions with
respect to application specific criteria is highly desirable. As will be seen, classical
entropy- and energy-based cost functions are applicable in many situations and
are well suited for use in binary and quaternary tree searching algorithms.
Consider the problem of reducing the amount of data needed to represent
the 400 * 480 fingerprint image in Fig. 7.36(a). Image compression is dis-
cussed in detail in Chapter 8. In this example, we want to select the “best”
three-scale wavelet packet decomposition as a starting point for the com-
pression process. Using three-scale wavelet packet trees, there are 83,522
[see Eq. (7.6-7)] potential decompositions. Figure 7.36(b) shows one of
them—a full wavelet packet, 64-leaf decomposition like the analysis tree of The 64 leaf nodes in
Fig. 7.35 correspond to
Fig. 7.35. Note that the leaves of the tree correspond to the subbands of the the 8 * 8 array of 64
8 * 8 array of decomposed subimages in Fig. 7.36(b). The probability that this subimages in Fig. 7.36(b).
Despite appearances,
particular 64-leaf decomposition is in some way optimal for the purpose of they are not square. The
compression, however, is relatively low. In the absence of a suitable optimality distortion (particularly
noticeable in the
criterion, we can neither confirm nor deny it. approximation subim-
One reasonable criterion for selecting a decomposition for the compression age) is due to the
program used to produce
of the image of Fig. 7.36(a) is the additive cost function the result.

E(f) = a ƒ f(m, n) ƒ (7.6-8)

m, n
516 Chapter 7 ■ Wavelets and Multiresolution Processing

a b
FIGURE 7.36 (a) A scanned fingerprint and (b) its three-scale, full wavelet packet decomposition. (Original
image courtesy of the National Institute of Standards and Technology.)

Other possible energy This function provides one possible measure of the energy content of two-
measures include the
sum of the squares of
dimensional function f. Under this measure, the energy of function f(m, n) = 0
f(x, y), the sum of the for all m and n is 0. High values of E, on the other hand, are indicative of func-
log of the squares, etc.
Problem 7.27 defines one
tions with many nonzero values. Since most transform-based compression
possible entropy-based schemes work by truncating or thresholding the small coefficients to zero, a cost
cost function.
function that maximizes the number of near-zero values is a reasonable criterion
for selecting a “good” decomposition from a compression point of view.
The cost function just described is both computationally simple and easily
adapted to tree optimization routines. The optimization algorithm must use
the function to minimize the “cost” of the leaf nodes in the decomposition
tree. Minimal energy leaf nodes should be favored because they have more
near-zero values, which leads to greater compression. Because the cost func-
tion of Eq. (7.6-8) is a local measure that uses only the information available at
the node under consideration, an efficient algorithm for finding minimal energy
solutions is easily constructed as follows:
For each node of the analysis tree, beginning with the root and proceeding
level by level to the leaves:

Step 1. Compute both the energy of the node, denoted EP (for parent
energy), and the energy of its four offspring—denoted EA, EH, EV, and
ED. For two-dimensional wavelet packet decompositions, the parent is
a two-dimensional array of approximation or detail coefficients; the
 7.6 ■ Wavelet Packets 517

offspring are the filtered approximation, horizontal, vertical, and diago-

nal details.
Step 2. If the combined energy of the offspring is less than the energy of the
parent—that is, EA + EH + EV + ED 6 EP—include the offspring in the
analysis tree. If the combined energy of the offspring is greater than or
equal to that of the parent, prune the offspring, keeping only the parent. It
is a leaf of the optimized analysis tree.
The preceding algorithm can be used to (1) prune wavelet packet trees or (2)
design procedures for computing optimal trees from scratch. In the latter case,
nonessential siblings—descendants of nodes that would be eliminated in step 2
of the algorithm—would not be computed. Figure 7.37 shows the optimized
decomposition that results from applying the algorithm to the image of
Fig. 7.36(a) with the cost function of Eq. (7.6-8). The corresponding analysis
tree is given in Fig. 7.38. Note that many of the original full packet decompo-
sition’s 64 subbands in Fig. 7.36(b) (and corresponding 64 leaves of the analy-
sis tree in Fig. 7.35) have been eliminated. In addition, the subimages that are
not split (further decomposed) in Fig. 7.37 are relatively smooth and com-
posed of pixels that are middle gray in value. Because all but the approxima-
tion subimage of this figure have been scaled so that gray level 128 indicates a
zero-valued coefficient, these subimages contain little energy. There would be
no overall decrease in energy realized by splitting them. ■

The preceding example is based on a real-world problem that was solved

through the use of wavelets. The Federal Bureau of Investigation (FBI) cur-
rently maintains a large database of fingerprints and has established a wavelet-
based national standard for the digitization and compression of fingerprint

FIGURE 7.37
An optimal
wavelet packet
decomposition for
the fingerprint of
Fig. 7.36(a).
518 Chapter 7 ■ Wavelets and Multiresolution Processing

VJ

V D
VJ
1 WJ
1 WJ
1

H V D V V V V
VJ
2 WJ
2 WJ
2 WJ
2 WJ
1,A WJ
1,H WJ
1,V WJ
1,D

H V D H H H H V V V V D D D D V V V V
VJ
3 WJ
3 WJ
3 WJ
3 WJ
2,A WJ
2,H WJ
2,V WJ
2,D WJ
2,A WJ
2,H WJ
2,V WJ
2,D WJ
2,A WJ
2,H WJ
2,V WJ
2,D WJ
1,AA WJ
1,AH WJ
1,AV WJ
1,AD

H
WJ
1

H H H H
WJ
1,A WJ
1,H WJ
1,V WJ
1,D

H H H H H H H H
WJ
1,AA WJ
1,AH WJ
1,AV WJ
1,AD WJ
1,DA WJ
1,DH WJ
1,DV WJ
1,DD

FIGURE 7.38 The optimal wavelet packet analysis tree for the decomposition in Fig. 7.37.

images (FBI [1993]). Using biorthogonal wavelets, the standard achieves a typical
compression ratio of 15:1. The advantages of wavelet-based compression over
the more traditional JPEG approach are examined in the next chapter.
The decomposition filters used in Example 7.15, as well as by the FBI, are
part of a well-known family of wavelets called Cohen-Daubechies-Feauveau
biorthogonal wavelets (Cohen, Daubechies, and Feauveau [1992]). Because
the scaling and wavelet functions of the family are symmetrical and have simi-
lar lengths, they are among the most widely used biorthogonal wavelets.
Figures 7.39(e) through (h) show the dual scaling and wavelet functions.
Figures 7.39(a) through (d) are the corresponding decomposition and recon-
struction filters. The coefficients of the lowpass and highpass decomposition
filters, h0(n) and h1(n) for 0 … n … 17 are shown in Table 7.4. The corre-
sponding coefficients of the biorthogonal synthesis filters can be computed
using g0(n) = (-1)n + 1h1(n) and g1(n) = (-1)n h0(n) of Eq. (7.1-11). That is,
they are cross-modulated versions of the decomposition filters. Note that zero
padding is employed to make the filters the same length and that Table 7.4 and
Fig. 7.39 define them with respect to the subband coding and decoding system
of Fig. 7.6(a); with respect to the FWT, hw(-n) = h0(n) and hc(-n) = h1(n).

TABLE 7.4 n h0(n) h1(n) n h0(n) h1(n)

Biorthogonal
Cohen- 0 0 0 9 0.8259 0.4178
Daubechies- 1 0.0019 0 10 0.4208 0.0404
Feauveau filter 2 -0.0019 0 11 -0.0941 -0.0787
coefficients 3 -0.017 0.0144 12 -0.0773 -0.0145
(Cohen, 4 0.0119 -0.0145 13 0.0497 0.0144
Daubechies, and 5 0.0497 -0.0787 14 0.0119 0
Feauveau [1992]). 6 -0.0773 0.0404 15 -0.017 0
7 -0.0941 0.4178 16 -0.0019 0
8 0.4208 -0.7589 17 0.0010 0
 7.6 ■ Wavelet Packets 519

h0(n) h1(n) a b
1.2 0.6 c d
1 0.4 e f
g h
0.8 0.2
FIGURE 7.39
0.6 0
A member of the
0.4
0.2 Cohen-
0.2
0.4 Daubechies-
Feauveau
0
0.6
biorthogonal

0.2 n
0.8 n wavelet family:
0 4 8 12 16 0 4 8 12 16
(a) and (b)
g0(n) g1(n) decomposition
0.8 0.5 filter coefficients;
(c) and (d)
0.6 reconstruction
0 filter coefficients;
0.4 (e)–(h) dual
wavelet and
0.2 scaling functions.

0.5
See Table 7.3 for
0 the values of
h0(n) and h1(n)

0.2 n
1 n
0 4 8 12 16 0 4 8 12 16 for 0 … n … 17.
~
w(x) w(x)
1.4 1.5
1.2
1 1
0.8
0.6 0.5
0.4
0
0.2
0
0.5

0.2

0.4 x
1 x
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
~
c(x) c(x)
1.2 1.5
1
1
0.8
0.6 0.5

0.4 0
0.2

0.5
0

0.2 x
1 x
0 2 4 6 8 10 12 14 16 18 0 2 4 6 8 10 12 14 16 18
520 Chapter 7 ■ Wavelets and Multiresolution Processing

Summary
The material of this chapter establishes a solid mathematical foundation for under-
standing and accessing the role of wavelets and multiresolution analysis in image pro-
cessing. Wavelets and wavelet transforms are relatively new imaging tools that are
being rapidly applied to a wide variety of image processing problems. Because of their
similarity to the Fourier transform, many of the techniques in Chapter 4 have wavelet
domain counterparts. A partial listing of the imaging applications that have been ap-
proached from a wavelet point of view includes image matching, registration, segmen-
tation, denoising, restoration, enhancement, compression, morphological filtering, and
computed tomography. Since it is impractical to cover all of these applications in a sin-
gle chapter, the topics included were chosen for their value in introducing or clarifying
fundamental concepts and preparing the reader for further study in the field. In
Chapter 8, we will apply wavelets to the compression of images.

References and Further Reading

There are many good texts on wavelets and their application. Several complement our
treatment and were relied upon during the development of the core sections of the
chapter. The material in Section 7.1.2 on subband coding and digital filtering follows
the book by Vetterli and Kovacevic [1995], while Sections 7.2 and 7.4 on multiresolu-
tion expansions and the fast wavelet transform follow the treatment of these subjects in
Burrus, Gopinath, and Guo [1998]. The remainder of the material in the chapter is
based on the references cited in the text. All of the examples in the chapter were done
using MATLAB (see Gonzalez et al. [2004]).
The history of wavelet analysis is recorded in a book by Hubbard [1998]. The early
predecessors of wavelets were developed simultaneously in different fields and unified
in a paper by Mallat [1987]. It brought a mathematical framework to the field. Much of
the history of wavelets can be traced through the works of Meyer [1987] [1990] [1992a,
1992b] [1993], Mallat [1987] [1989a–c] [1998], and Daubechies [1988] [1990] [1992]
[1993] [1996]. The current interest in wavelets was stimulated by many of their publica-
tions. The book by Daubechies [1992] is a classic source for the mathematical details of
wavelet theory.
The application of wavelets to image processing is addressed in general image pro-
cessing texts, like Castleman [1996], and many application specific books, some of
which are conference proceedings. In this latter category, for example, are Rosenfeld
[1984], Prasad and Iyengar [1997], and Topiwala [1998]. Recent articles that can serve
as starting points for further research into specific imaging applications include Gao
et al. [2007] on corner detection; Olkkonen and Olkkonen [2007] on lattice implemen-
tations; Selesnick et al. [2005] and Kokare et al. [2005] on complex wavelets; Thévenaz
and Unser [2000] for image registration; Chang and Kuo [1993] and Unser [1995] on
texture-based classification; Heijmans and Goutsias [2000] on morphological wavelets;
Banham et al. [1994], Wang, Zhang, and Pan [1995], and Banham and Kastaggelos
[1996] on image restoration; Xu et al. [1994] and Chang, Yu, and Vetterli [2000] on
image enhancement; Delaney and Bresler [1995] and Westenberg and Roerdink [2000]
on computed tomography; and Lee, Sun, and Chen [1995], Liang and Kuo [1999], Wang,
Lee, and Toraichi [1999], and You and Bhattacharya [2000] on image description and
matching. One of the most important applications of wavelets is image compression—
see, for example, Brechet et al. [2007], Demin Wang et al. [2006], Antonini et al. [1992],
Wei et al. [1998], and the book by Topiwala [1998]. Finally, there have been a number of
special issues devoted to wavelets, including a special issue on wavelet transforms and
multiresolution signal anaysis in the IEEE Transactions on Information Theory [1992],
a special issue on wavelets and signal processing in the IEEE Transactions on Signal
 ■ Problems 521

Processing [1993], and a special section on multiresolution representation in the IEEE

Transactions on Pattern Analysis and Machine Intelligence [1989].
Although the chapter focuses on the fundamentals of wavelets and their application
to image processing, there is considerable interest in the construction of wavelets them-
selves. The interested reader is referred to the work of Battle [1987] [1988], Daubechies
[1988] [1992], Cohen and Daubechies [1992], Meyer [1990], Mallat [1989b], Unser,
Aldroubi, and Eden [1993], and Gröchenig and Madych [1992]. This is not an exhaus-
tive list but should serve as a starting point for further reading. See also the general ref-
erences on subband coding and filter banks, including Strang and Nguyen [1996] and
Vetterli and Kovacevic [1995], and the references included in the chapter with respect
to the wavelets we used as examples.

Problems
7.1 Design a system for decoding the prediction residual pyramid generated by the
encoder of Fig. 7.2(b) and draw its block diagram. Assume there is no quantiza-
tion error introduced by the encoder.

7.2 Construct a fully populated approximation pyramid and corresponding predic-
tion residual pyramid for the image

1 2 3 4
5 6 7 8
f(x, y) = D T
9 10 11 12
13 14 15 16
Use 2 * 2 block neighborhood averaging for the approximation filter in Fig. 7.2(b)
and assume the interpolation filter implements pixel replication.

7.3 Given a 2 J * 2 J image, does a J + 1-level pyramid reduce or expand the
amount of data required to represent the image? What is the compression or ex-
pansion ratio?
7.4 Is the two-band subband coding filter bank containing filters h0(n) = E 1> 12,
1> 12 F , h1(n) = E -1> 12, 1> 12 F , g0(n) = E 1> 12, 1> 12 F , and g1(n)
= E 1> 12, -1> 12 F orthonormal, biorthogonal, or both?
7.5 Given the sequence f(n) = 50, 0.5, 0.25, 16 where n = 0, 1, 2, 3, compute:
(a) The sign-reversed sequence.
(b) The order-reversed sequence.
(c) The modulated sequence.
(d) The modulated and then order-reversed sequence.
(e) The order-reversed and then modulated sequence.
(f) Does the result from (d) or (e) correspond to Eq. (7.1-9)?
7.6 Compute the coefficients of the Daubechies synthesis filters g0(n) and g1(n)
for Example 7.2. Using Eq. (7.1-13) with m = 0 only, show that the filters are
orthonormal.

7.7 Draw a two-dimensional four-band filter bank decoder to reconstruct input
f(m, n) in Fig. 7.7.
7.8 Obtain the Haar transformation matrix for N = 8.
7.9 (a) Compute the Haar transform of the 2 * 2 image

3 -1
F = B R
6 2