Wavelets and Image Compression

Vlad Balan, Cosmin Condea

January 30, 2003
Abstract
In this paper we will examine the wavelet transform, one of the most recent
mathematical tools related to signal representation and illustrate its appli-
cation in the eld of image compression. The paper is divided in two main
parts. In the rst one we present the mathematical principles of multireso-
lution analysis, we illustrate them using the Haar wavelet in the one dimen-
sional case, we present the transform algorithms and we end up discussing a
number of more advanced topics. The second part starts by describing the
transition to the discrete case and then presents in a step by step manner
the general procedure for image compression using wavelets. There are four
basic steps: applying the wavelet transform, threshold detection, quantizing
and encoding the resulting data and nally applying an inverse transform.
These theoretical aspects are illustrated through a MATLAB project which
we developed using the Stanford Universitys Wavelab toolbox.
Contents
1 Wavelets 2
1.1 Principles of multiresolution analysis . . . . . . . . . . . . . . 2
1.2 A simple example: The Haar Wavelet . . . . . . . . . . . . . . 4
1.3 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Advanced topics . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Image Compression 8
2.1 Applying the transform . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Choosing a threshold . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Compression methods . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Applying the Inverse Transform . . . . . . . . . . . . . . . . . 13
1
1 Wavelets
1.1 Principles of multiresolution analysis
We dene a multiresolution analysis as a mathematical object consisting of
the following:
(a) A bilateral sequence of closed subspaces V
j
of L
2
ordered by inclu-
sion:
. . . V
2
V
1
V
0
V
1
. . . V
j
V
j1
. . . L
2
(1)
and obeying to the following axioms:

j
V
j
= 0 (separation axiom) (2)
_
j
V
j
= L
2
(completeness axiom) (3)
(b) A scaling property of the V
j
subspaces:
V
j
= D
2
(V
j
) j Z where (4)
D
2
(f) =
n

k=0
h
k
f(t k) (5)
or: f V
j
f(2) V
j1
.
(c) There exists a function L
2

L
1
such that its translates ( ( k)k Z )
form an orthonormal basis of V
0
. The function is called the scaling
function. We notice that the space V
0
uses one such basis vector per
unit length while V
j
uses 2
j
basis vectors per unit length.
We conclude that the functions
j,k
V
j
constitute a basis of V
j
. How-
ever, we cannot form a basis of L
2
just by taking the union of these since the
subspaces V
j
cannot be orthogonal as a consequence of relation (a).
We can dene another sequence of subspaces W
j
= V
j
V
j1
. These can
be proven to be pairwise orthogonal, and even more

j
W
j
= L
2
.
Looking at the formula W
0
= V
0
V
1
and bearing in mind that V
0
uses
twice as many vectors per unit length as V
1
we would be tempted to start
2
looking for a function which would constitute a basis of W
0
. Such a con-
struction is possible by going into the Fourier domain, and the reader is
invited to consult [Blatter, chap. 5] The resulting functions
j,k
constitute
an orthonormal basis of L
2
.
For our purposes it is convenient to require that:
_

(x)dx = 1,
_

|(x)|
2
dx = 1 (6)
_

|(x)|
2
dx = 1 (7)
From axiom b) the following identity holds:
(t) =
_
(2)

h
k
(2t k) for almost all t R (8)
with the coecient vector h l
2
(Z).
Furthermore we can prove that the wavelet function:
(t) =

g
k
(2t k) (9)
=

(1)
k
h
1k
(2t k) for almost all t R (10)
with the coecient vector g l
2
(Z) is orthogonal to the scaling function. If
one additional condition is met, the wavelet functions can be proved to be
orthogonal to each other, giving us a valid basis.
We can therefore dene the wavelet expansion of a function f as

j,kZ
C
jk
2
j/2
(2
j
k) (11)
with the coecients C
jk
dened by
C
jk
= f,
jk
=
_
+

f(x)2
j/2
(2
j
k) dx (12)
We can think of the coecients of each the functions belonging to each
V
j
as representing those features of a signal that have a spread of size com-
parable to 2
j
/

2. From this point of view a multiresolution analysis process

can be imagined such as the work of a common lter bank.
3
1.2 A simple example: The Haar Wavelet
The mathematician Alfred Haar was the rst to describe, in 1910, an or-
thonormal system for the Hilbert space L
2
and proved it to be isomorphic to
the space
l
2
:= {(c
k
|k N)|

k=0
|c
k
|
2
< } (13)
Since the union of all step functions of step 2
j
, j Z is dense in L
2
and
the other conditions are obviously satised we conclude that 1
[0,1)
constitutes
a valid scale function for a multiresolution analysis.Starting from the scaling
function = 1
[0,1)
we obtain the Haar Wavelet which has the form of the
following step function:
(x) =
_
_
_
1 (0 x <
1
2
)
1 (
1
2
x < 1)
0 otherwise
We see that:
(x) = (2x) (2x 1) (14)
(x) = (2x) (2x 1) (15)
Intuitively, considering that at the level V
j
we are left with a step func-
tion f
j
of step 2
j
from two neighbor steps corresponding to the intervals
[2k

2
r
, (2k +1)

2
r
) and [(2k +1)

2
r
, (2k +2)

2
r
) we obtain the step correspond-
ing to f
j+1
s interval [2k

2
r
, (2k+2)

2
r
) as their mean value and the coecient
of the corresponding
jk
function as the dierence between the mean value
and the value of the function, as seen in the gure (from [Blatter, p. 24]):
1.3 Algorithms
For the analysis of a signal having as support the interval [0, 1) we dene
the scaling function and the wavelet function on this interval. Consid-
ering that we have sampled the function in 2
N
equally distanced points, we
can approximate it by a step function. We can start applying the wavelet
transform from the subspace V
N
.
4
Figure 1: Haar Wavelets action
Considering A
N
as a 2
N
size vector containing the coecients a
Nk
we can
apply the following operator H
N
in order to obtain a vector B
N
containing
the coecients a
(N1)k
and C
(N1)k
in an interlaced form:
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
h
0
h
1
h
2
. . . h
n
g
0
g
1
g
2
. . . g
n
h
0
h
1
h
2
. . . h
n
g
0
g
1
g
2
. . . g
n
. . .
. . .
h
0
h
1
h
2
. . . h
n
g
0
g
1
g
2
. . . g
n
h
2
. . . h
n
h
0
h
1
g
2
. . . g
n
g
0
g
1
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
_
By applying a permutation matrix:
P
N
[a
N1
1, C
N1
1, . . . , a
N1
N
2
1, C
N1
N
2
1]
T
= [a
N1
1, . . . , a
N1
N
2
1, . . . , C
N1
1, . . . , C
N1
N
2
1]
T
we move the coecients a
N1
k to the front, therefore we take the rst
half of B
N
for A
N1
.
We conclude that we can perform a full wavelet transform by applying
a series of PH operations and an inverse transform by applying a series of
H
1
P
1
operations. Knowing that we are multiply by the coecients h
k
in
order to obtain the next sequence of as and with g
k
in order to obtain the
next sequence c
k
we can represent the transform process by the following
diagram:
a
j

h
a
j1

h
a
j2
. . .
h
a
0

g
c
j1
c
j2
c
0
5
while the inverse transform can be represented as:
a
j
a
j1
a
j2
. . . a
0

c
j1
c
j2
c
0
Since the matrixes involved are sparse the complexity of the multiplica-
tions is O(n). The complexity of the whole series is therefore

N
k=1
O(n)
2
k
=
O(2n) = O(n).
1.4 Advanced topics
Having exposed the basic principles of wavelet analysis we proceed now to
describing some properties of dierent common wavelets. Since our wavelet
functions are centered around their set of coecients h
k
we are interested in
nding the minimal set of coecients satisfying some constraint equations.
We have from (7) just by integrating both sides:
_
dx =

h
k
_
(2x k)d(2x k) (16)

h
k
=

2 (17)
and by integrating with respect to the L
2
norm:
_

2
dx =

(h
k
_
(2x k)d(2x k))
2
(18)

|h
k
|
2
= 1. (19)
Signals that are smooth present a high degree of linearity. Their Taylor series
expansion around each point tends to decay very fast. We would be interested
in a wavelet function whose scalar product with a given polynomial vanishes:
_

x
j
(x)dx = 0 forj = 0, 1, . . . , L 1 (20)
We say in this case that has its rst L moments equal to 0.
Combining equation (9),(20) we obtain

(1)
k
h
k
= 0,

k(1)
k
h
k
= 0 (21)
6
and in taking for example L = 2 the minimal set of coecients satisfying all
four conditions is:
h
0
=
1+

3
4

2
, h
1
=
3+

3
4

2
h
2
=
3

3
4

2
, h
3
=
1

3
4

2
(22)
which denes the Daubechies D
4
wavelet.
Having these coecients we can represent the graph of the function. We
start from the relation
j
(x) =

c
k

j1
(2x k), noticing its similarity
to the equation of Iterated Function Systems (fractals). We can set the
box function as
0
and by iteration draw its graph while using the wavelet
coecients we represent in the same manner the wavelet function [Strang 2]:
Figure 2: Daubechies D4 Scaling Function (left) and Wavelet Function (right)
An alternate method would be to compute using the same relation from
the values at points x = 2
j
the values at points x = 2
j+1
. For obtaining the
initial values at the points x = 1 and x = 2 we use the fact that the function
has as compact support the interval (0, 3) and we solve the equation:
(1) =

2h
2
(1) +

2h
1
(2) (23)
(2) =

2h
4
(1) +

2h
3
(2) (24)
which gives as (1) and (2) as eigenvectors for the eigenvalue = 1.
A nal question that remained unanswered during our presentation is under
what circumstances the construction in the formula (9) gives us a wavelet
that is orthogonal to its translates. We are not going to tackle this problem
here, since it is only relevant in the construction of a wavelet function but the
interested reader is can consult the article [Strang 2] or the book [Blatter].
As we can see the strong point of the wavelet transform is a good localization
7
both in terms of scale and position, which gives a signal good localization in
time, since the frequency depends on the scale, and in space. This capacity
to detect local features and features spreading over a larger distance makes
the wavelet transform a suitable candidate for image compression since it is
capable to retain and to evidentiate redundant information which is specic
to a natural signal.
We have to bear in mind that when using wavelets such as Daubechies we
are faced with a compromise between the length of the wavelet coecients
set to which the processing time is proportional and the speed of decay of
the Taylor series of the processed signal. While an image signal has a slow
decay due to many local irregularities (that is, there will always be trees in
the background) and the lters are quite short, in audio applications where
the signal is much cleaner they tend to be much longer.
The wavelet transform should not be seen as the universal solution for com-
pressing discrete time signals. For example, when compressing a signal that
is composed of sinusoidal functions a Fourier transform is guaranteed to give
a much smaller set of meaningful coecients.
Finally, anticipating the topic of the second part of this paper, let us say
that an easy method to construct bi-dimensional wavelets is to start from an
one-dimensional wavelet and take the cross products , , , , which
give clearly orthogonal functions. Although this method can be easily used,
dierent genuine bi-dimensional wavelets have been invented.
2 Image Compression
One area where wavelets have incontestably proven their applicability is im-
age processing. As you know high resolution images claim a lot of disk space.
In the age of information highway, the amount of data that needs to be stored
or transmitted is huge. Therefore, compression greatly increases the capacity
of the storage device, on one hand, and on the other, it also reduces costs.
To illustrate the use of compression take the simplest example: an image
of 256 x 256, which takes approximatively 0.2 MB. On a simple oppy disk
one can therefore store 7 such images. But think if this image can be com-
pressed at a 25:1 ratio. The result is 175 images stored on the same oppy
disk.
In this part of the paper, we describe the compression algorithm step
by step, using the Lenna image (g.3) for illustrations. Finally we will
present the MATLAB code we wrote. We used the WaveLab v802 toolbox,
downloaded from Stanford Universitys web site.
8
Figure 3: The Well-Known Lenna Image
2.1 Applying the transform
The compression algorithm starts by transforming the image from data space
to wavelet space. This is done on several levels. We start with our data
applying the bi-dimensional transform matrix and we get in the resulting
image the coecients grouped into four zones, like in the gure, where H
symbolizes high frequency data and L symbolizes low frequency data, like in
the gure:
Figure 4: The Discrete Wavelet Transform Frequency Quadrants
9
The LL quadrant of the resulting image is the input of the next iteration.
Usually for image compression purposes 4 or 5 iterations will suce. In the
next gures the result of the transform on 1 and on 5 levels.
Figure 5: The Discrete Wavelet Transform (1 level and 5 levels)
The 5-level transforms data is also presented in a mesh form in order
to visualize better the dierent intensities of the coecients.An interesting
aspect to notice is that the majority of the DWT coecients are positioned
in the upper left quadrant.
Figure 6: 3D View on the Discrete Wavelet Transform
10
2.2 Choosing a threshold
The next step in the algorithm is to neglect all the wavelet coecients that
fall below a certain threshold. We select our threshold in such a way as to
preserve a certain percent of the total coecients - this is known as quantile
thresholding.
The small values of the DWT coecients retain little detail of the pic-
ture. Therefore they can, up to a limit, be neglected. The key notion is
here the perceptual distortion. Of course some details of the picture are
consequently lost after applying the threshold but the question is to what
extent the human eye can detect the dierence between the original and the
reconstructed image. In this direction, a human visual perception model has
been created and its use in image compression has been studied. This model
still remains an ongoing research project at the current time.
In what thresholding is concerned, besides the quantile one, there exist
another 2 main types of thresholding:
Hard Thresholding eliminates all the coecients c
i
that are smaller than
an absolute threshold T. If we denote with c

i
the new coecients:
c

i
=
_
c
i
if c
i
> T
0 otherwise
Soft Thresholding again sets an absolute limit reducing to zero all the
coecients that fall under it but at the same time it shrinks toward 0
if c
i
> T. Keeping the notations, the relation for soft thresholding is:
c
i

= sgn(c
i
) max(|c
i
| T, 0)
Coming back to the compression algorithm and the threshold step, in the
next gure we represent the non-zero distribution in the DWT after we have
chosen to use 5% of the coecients - the greatest in absolute value (g.7).
2.3 Compression methods
Coding We have designed a very simple compression scheme for sparse
matrices in order to test the eciency of the algorithm. We traverse the
thresholded wavelet datas matrix line by line and we copy all the nonzero
values to a vector. When we encounter a zero value, we start counting the
11
Figure 7: Distribution of Non-zero in the DWT (5% of the coecients)
length of the sequence of zeroes to which it belongs. Every such sequence
we replace with a zero value followed by its length. After encoding our
MATLAB code prints out the length of the resulting vector.
It is easy to see that the data can be reconstructed from this vector. Our
compression scheme is not very ecient since we obtain about three times
more data than the one belonging to the selected nonzero coecients. This
mean that when we preserve 5% of the coecients we only compress the
picture to 15% of its original data.
In order to eciently store our data it is preferable to work with integers
rather than oats. We can just round our data to the nearest integer or we
can scale it rst. This process is known as quantization. We did not use it in
our MATLAB code since we just measure the length of the resulting vector,
and consider that we could encode each value using two bytes.
More popular compression codes include Entropy coding, Human coding
and specialized algorithms for coding wavelet transformed image data such
as those created by R. Wells and J. Tian or by J.S. Walker.
In the next paragraph we will briey discuss the entropy coder. The bot-
tom line idea is that the coder should take advantage of long strings of 0 -
which, after thresholding and quantization, they are mostly placed into the
high frequency quadrants. This is done by scanning.
12
Figure 8: Scanning Method for the Entropy Coder [SN]
The 3-level DWT in the above picture illustrates how scanning is per-
formed. If,for example, the shaded area in the 2
nd
quadrant is found to be
0 - most likely to be so - then is can be assumed that the shaded areas in
quadrants 5 and 8 also have zero coecients. This idea can also be explained
by comparing the DWT with a tree - where each parent has 4 children. If a
parent is found to be 0, then all his children contain zero values.
Also to be noticed is the scanning pattern for dierent frequency quadrant
sequences. Vertical for (2,5 and 8), horizontal for (3,6 and 9) and diagonal
for (4,7 and 10). The obtained AC sequences are encoded in a standard
Human way. The DC sequence is encoded based on the image continuity
that is the dierences of color are stored.
2.4 Applying the Inverse Transform
After decoding the data, the last step of the algorithm is that of applying the
inverse DWT to the doctored image matrix. In the following we include
some pictures where we set dierent quantile thresholds.
13
Figure 9: The Restored Lenna Image (with 10% of the coecients)
Figure 10: The Restored Lenna Image (with 5% of the coecients)
14
Figure 11: The Restored Lenna Image (with 3% of the coecients)
References
[Blatter] Blatter, Christian - Wavelets, A Primer, A.K. Peters, Ltd. 1998
[Kaiser] Kaiser, Gerald - A Friendly Guide to Wavelets, Birkhauser 1994
[SN] Strang, Gilbert and Nguyen, Truong - Wavelts and Filter Banks,
Wellesley-Cambridge Press 1996
[Walker] Walker, James S. - Fourier analysis and wavelet analysis. Notices
of the AMS, vol. 44, No. 6, pp. 658-670, 1997
[Strang 1] Strang, Gilbert - Wavelets, American Scientist 82 (April 1994)
250 - 255
[Strang 2] Strang, Gilbert - Wavelets and Dilation Equations, Siam Review
31 (1989) 613-627
15
Appendix
This part contains the MATLAB code we have written for the image com-
pression application of wavelets. We worked with the WaveLab v802 toolbox
designed by the Statistics Department of Stanford University. We found it
very helpful and we inspired our program from the examples it contained in
this respect.
For reference, we oer the site where we downloaded it:
<http://www-stat.stanford.edu/ wavelab/>
ncoef= input(Enter the percentage of the DWT coefficients that
you want to keep:)
ncoef = (100-ncoef)/100;
%Presenting the image of Lenna
x=readimage(Lenna);
autoimage(x);
title(The Well-Known Lenna
Image);
uiwait;
%Presenting the image transform of Lenna
qmf=MakeONFilter(Daubechies,8);
wlenna=FWT2_PO(x,3,qmf);
wlenna_1=FWT2_PO(x,7,qmf);
y_1 = abs(wlenna_1);
y=abs(wlenna);
subplot(121);
autoimage(y_1);
title(Wavelet Transform - 1 Level)
subplot(122);
autoimage(y);
title(Wavelet Transform of Lenna - 5
Levels);
16
uiwait;
mesh(wlenna);
title(3D View of the Wavelet Transform);
uiwait;
%Ilustrating the non-zero elements of the WTransform matrix
coef_sort = sort(abs(wlenna(:)));
treshold = coef_sort(floor(ncoef*65536));
new_wlenna=wlenna.*(abs(wlenna)>treshold);
[i,j,v]=find(new_wlenna);
sp_lenna=sparse(i,j,v,256,256);
spy(sp_lenna);
title(Distribution of non-zero in the WT of
Lenna);
uiwait;
%Finding out how much space we need using
%a simple compression scheme
comp = zeros(1);
sz = size(x);
s = sz(1);
comp(1) = s;
n = 1;
onzero = 0;
for i=1:s
for j=1:s
if abs(new_wlenna(i,j)) > 0
if onzero == 1
comp(n) = 0;
n = n+1;
comp(n) = nzero;
n = n+1;
end
comp(n) = new_wlenna(i,j);
n = n+1;
else
17
if onzero == 1
nzero = nzero+1;
else
onzero = 1;
nzero = 1;
end
end
end
end
disp(We compress in a vector of length);
disp(2*n);
disp(while the initial size is);
disp(2*s*s);
disp(and the compression ratio is)
disp((s*s)/n);
%Getting the image back
result = IWT2_PO(new_wlenna,3,qmf);
autoimage(result);
title(Lenna Image Restored (with 5\% of the WT coefficients));
uiwait;
18