The Haar–Wavelet Transform in Digital Image Processing:

Its Status and Achievements

Piotr Porwik, Agnieszka Lisowska

Institute of Informatics, University of Silesia, ul. Bȩdzińska 39, 41-200 Sosnowiec, Poland
e-mail: [email protected]
Institute of Mathematics, University of Silesia, ul. Bankowa 14, 40-007 Katowice, Poland
e-mail: [email protected]

Abstract. Image processing and analysis based on the continuous or discrete image transforms
are classic techniques. The image transforms are widely used in image filtering, data description, etc.
Nowadays the wavelet theorems make up very popular methods of image processing, denoising and
compression. Considering that the Haar functions are the simplest wavelets, these forms are used in
many methods of discrete image transforms and processing. The image transform theory is a well
known area characterized by a precise mathematical background, but in many cases some transforms
have particular properties which are not still investigated. This paper for the first time presents graphic
dependences between parts of Haar and wavelets spectra. It also presents a method of image analysis
by means of the wavelets–Haar spectrum. Some properties of the Haar and wavelets spectrum were
investigated. The extraction of image features immediately from spectral coefficients distribution were
shown. In this paper it is presented that two–dimensional both, the Haar and wavelets functions
products man be treated as extractors of particular image features. Furthermore, it is also shown that
some coefficients from both spectra are proportional, which simplify slightly computations and analyses.

Key words: Wavelets, Haar Transform, multiresolution.

1. Introduction

The computer and video–media applications have developed rapidly the field of multi-
media, which requires the high performance, speedy digital video and audio capabilities.
Nowadays, the image processing and analysis based on continuous or discrete trans-
forms are the classic processing techniques [3, 27, 30, 36, 50]. Digital signal processing
is widely used in many areas of electronics, communication and information techniques
[1, 4, 6, 9, 14, 15, 17, 18, 20, 23, 28, 35]. In the signals compression, digital filtration,
systems identification, the commonly used transforms are based on sinusoidal basic func-
tions such as: Discrete Fourier Transform, Discrete Sine or Cosine Transform, Hartley
Transform or rectangular basic functions Slant Transform, Discrete Walsh Transform,
and Discrete Wavelet Transform (Haar, Daubechies, etc.) [3, 4, 6, 11, 12, 17, 25, 49, 50].
All these mentioned functions are orthogonal, and their forward and inverse transforms
require only additions and subtractions. It makes that it is easy to implement them on
the computer.

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
80 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

Haar functions are used since 1910. They were introduced by Hungarian mathemati-
cian Alfred Haar [1]. Nowadays, several definitions of the Haar functions and various
generalizations [39] as well as some modifications [19, 37, 50] were published and used.
One of the best modification, which was introduced, is the lifting scheme [25, 26, 29].
These transforms have been applied, for instance, to spectral techniques for multiple–
valued logic, image coding, edge extraction, etc. Over the past few years, a variety of
powerful and sophisticated wavelet–based schemes for image compression, as discussed
later, were developed and implemented. Wavelet scheme gives many advantages, which
are used in the JPEG–2000 standard as wavelet–based compression algorithms [31].
Generally, wavelets, with all generalizations and modifications, were intended to
adapt this concept to some practical applications [40, 42]. The Discrete Wavelet Trans-
form uses the Haar functions in image coding, edge extraction and binary logic design
and is one of the most promising technique today. The non–sinusoidal Haar transform
is the complete unitary transform [15, 16, 17]. It is local, thus can be used for data
compression of non–stationary ”spiky” signals. The digital images may be treated as
such ”spiky” signals. Unfortunately, the Haar Transform has poor energy compaction
for image, therefore in practice, basic Haar transform is not used in image compression.
One should remember that researches in this topic are still in progress and everyday new
solutions and improvements are found [33, 39, 41, 43, 47].
Fourier methods are not always good tools to recapture the signal [3], particularly if it
is highly non–smooth; too much Fourier information is needed to reconstruct the signal
locally. In these cases the wavelet analysis is often very effective because it provides
a simple approach for dealing with the local aspects of a signal, therefore particular
properties of the Haar or wavelet transforms allow to analyse the original image on
spectral domain effectively. These methods will be described in this paper.

2. The Discrete Haar Transform

A complete orthogonal system of functions in Lp [0, 1], p ∈ [0, ∞] which take values from
the set {0, 2j : j ∈ N} was defined by Haar [1]. This system of functions has property
that each function continuous on interval [0, 1] may be represented by a uniformly and
convergent series in terms of elements of this system. Nowadays, in the literature, there
are some other definitions of the Haar functions [16]. Those definitions are mutually
differing with respect to the values of Haar functions at the points of discontinuity. For
example the original Haar definition is as follows [4]:

1, for t ∈ [0, 21 ),

haar(0, t) = 1, for t ∈ [0, 1); haar(1, t) = (1)
−1, for t ∈ [ 12 , 1)
and haar(k, 0) = limt→0+ haar(k, t), haar(k, 1) = limt→1− haar(k, t) and at the points
of discontinuity within the interior (0, 1) haar(k, t) = 12 (haar(k, t − 0) + haar(k, t + 0)).

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp .79-98
Piotr Porwik, Agnieszka Lisowska 81

Instead of described relations some authors use the formula haar(k, t) = haar(k, t + 0)
where in the practice it is usually assumed that the Haar function takes zero value at
the points of discontinuity.
Discrete Haar functions may be defined as functions determined by sampling the
Haar functions at 2n points. These functions can be conveniently represented by means
of matrix form. The Haar matrices H(n) are considered in the natural and sequence
ordering which differ in the ordering of rows. Each row of the matrix H(n) includes
the discrete Haar sequence haar(w, t) (or otherwise the discrete Haar function). In this
notation, index w identifies the number of the Haar function and index t the discrete point
of the function determination interval. In this case, the Haar matrix of any dimension
may be obtained by the following recurrence relation [3]:
 
H(n − 1) ⊗[1 1]
H(n) = , H(0) = 1 (2)
2(n−1)/2 I(n − 1) ⊗[1 −1]
and H(n) 6= H(n)T for n > 1 and H(n)−1 = 2−n · H(n)T , where H(n) — matrix of
the discrete Haar functions of degree 2n , I(n) — identity matrix of degree 2n , ⊗ — the
Kronecker (tensor) product.
From (2) one can observe that, unlike the Fourier transform, H(n) matrix has only
real elements. The Haar √ matrix is non–symmetric and its elements are 1, −1 or 0,
multiplied by powers of 2. The discrete, orthogonal Haar functions, obtained from the
formula (2), are defined on [0, 1) interval.
Def. 2.1. Two–dimensional N × N = 2n × 2n forward discrete Haar transform is
defined in matrix notation as
S = a · H(n) · F · a · H(n)T . (3)
The inverse transform is defined as
F = b · H(n)T · S · b · H(n) , (4)
where F is the image in matrix form, the matrix is of dimension N × N pixels, S is the
spectrum
√ matrix and a · b = 1/N , hence parameters a or b may be defined as values 1/N ,
1/ N or 1, n = log2 N .
Because matrix (2) has many zero entries, some values of the Haar spectral coefficients
are equal to 0 too. Fig. 1 presents some known transforms of a test image. As transform
matrices H(n), Walsh–Hadamard, Haar, sine and cosine matrices were used. All figures
make up the graphic representation of spectral coefficients. Therefore each picture in
Fig. 1 may be interpreted as matrix, where axis row and col describe elements of matrix
S and axis value of coeff. indicates values of spectral coefficients in S. The test image F
is built as 8 × 8 matrix, which has zero values everywhere except the upper left element,
which has the value of eight.
The Fast Haar Transform has already been well known from many works [3, 4, 5,
17, 28] therefore it will not be present in details. The Haar transform is a symmetric,

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
82 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

separable transform that uses Haar function for its basis. It exists for N = 2n , where n is
an integer. The Haar function, which is an odd rectangular pulse pair, is the simplest and
oldest orthonormal wavelet [16, 17, 28]. Whereas the Fourier transform basis functions
differ only in frequency, the Haar functions vary in the both scales of width and position.

√ √
a) a = 1/ 8 b) a = 1/ 8

c) a = 1 d) a = 1
Fig. 1. The S transform of image containing the test impulse: a) Walsh–Hadamard, b) Haar, c) DST
(Discrete Sine Transform), d) DCT (Discrete Cosine Transform).

From Fig. 1 we can observe that all N 2 elements of these transforms are nonzero
except the Haar transform, which has only 2N nonzero entries. These features are very
important in image processing because in many cases spectral coefficients have zero
entries before next retrieval operations. This case occurs in black and white images very
often. From image data compression point of view, this case is very convenient. From
the energy distribution it can be estimated which spectral coefficients can be reduced
[15, 17, 28]. It is easy to observe from Fig. 1 that the Walsh transform gives the worst
results here: distribution of spectral energy is uniformable. In c) and d) cases distribution

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
Piotr Porwik, Agnieszka Lisowska 83

of spectral energy has sharply outlined maximum. Outside of the maximum value, one
can observe the decrease of energy. Unfortunately, the distribution of the Haar spectrum
(Fig. 1b) is not proper too, but we can treat this transform differently.
The discrete transforms, presented in Fig. 1, enable us to observe where energy
concentrations occur. Hence, we know which spectral coefficients are important in image
processing. From this representation, it is not possible to find more precisely information
about real image. For example, it is difficult to point places, which describe horizontal,
vertical, etc. details of real image. These troubles can be overcome by well known
multiresolution analysis [11, 17, 20, 34, 49].

3. The Discrete Haar Wavelet Transform

An outstanding property of the Haar functions is that except function haar(0, t), the
i–th Haar function can be generated by the restriction of the (j − 1)–th function √ to
be half of the interval where it is different from zero, by multiplication with 2 and
scaling over the interval [0, 1]. These properties give considerable interest of the Haar
function, since they closely relate them to the wavelet theory. In this setting, the first
two Haar functions are called the global functions, while all the others are denoted
as the local functions. Hence, the Haar function, which is an odd rectangular pulse
pair, is the simplest and oldest wavelet. The motivation for using the discrete wavelet
transform is to obtain information that is more discriminating by providing a different
resolution at different parts of the time–frequency plane. The wavelet transforms allow
the partitioning of the time-frequency domain into nonuniform tiles in connection with
the time–spectral contents of the signal. The wavelet methods are strongly connected
with classical basis of the Haar functions; scaling and dilation of a basic wavelet can
generate the basis Haar functions.
Def. 3.1. Let Ψ: R → R, the Haar wavelet function is defined by the formula [43]:
 1, for t ∈ [0, 12 ),


Ψ(t) = −1, for t ∈ [ 12 , 1), (5)

0, otherwise.


Taking into account the Definition 3.1, any Haar function (except function haar(0, t))
from basis (2) may be generated by means of the formulas:
√
Ψji (t) = 2j Ψ(2j t − i), i = 0, 1, . . . , 2j − 1 and j = 0, 1, . . . , log2 N − 1 . (6)
√
The constant 2j is chosen so that the scalar product < Ψji , Ψji >= 1, Ψji ∈ L2 (R). If
one considers the wavelet function on other intervals than [0, 1), the normalisation con-
stant will be different. For example: Ψ00 = haar(1, t), Ψ10 = haar(2, t), Ψ11 = haar(3, t),

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
84 The Haar–Wavelet Transform in Digital Image Processing: ItsStatus and Achievements

Ψ20 = haar(4, t), Ψ21 = haar(5, t), Ψ22 = haar(6, t), Ψ23 = haar(7, t). Generally Ψji =
haar(2j + i, t). From this example follows that functions Ψji (t) are orthogonal to one an-
other. Hence, we obtain linear span of vector space W j = span{Ψji }i=0,...,2j −1 . A collec-
tion of linearly independent functions {Ψji (t)}i=0,...,2j −1 spanning W j is called wavelets.
Def. 3.2. Let Φ: R → R, the Haar scaling function is defined by the formula [43]:

1, for t ∈ [0, 1),
Φ(t) = (7)
0, for t 6∈ [0, 1).

Similarly to the properties of the wavelet function, for scaling function one can define
the family of functions:
√
Φji (t) = 2j Φ(2j t − i), i = 0, 1, . . . , 2j − 1 and j = 0, 1, . . . , log2 N − 1 . (8)
√
The constant 2j is chosen so that the scalar product < Φji , Φji >= 1, Φji ∈ L2 (R). The
index j refers to dilation and index i refers to translation [11, 17]. Hence, we obtain
linear span of vector space V j = span{Φji }i=0,...,2j −1 . The basic functions from the
space V j are called scaling functions.
In multiresolution analysis the Haar basis has important property: V j = V j−1 ⊕
j−1
W , where ⊕ stands for orthogonality of V j and W j spaces [16].
From Definitions 3.1 and 3.2 follows, that vector space W j can be treated as the or-
thogonal complement of V j in V j+1 . In other words, let W j be the space of all functions
in V j+1 that are orthogonal to all functions in V j . Therefore, the basis functions Ψji of
W j together with the basis functions Φji of V j form a basis for V j+1 and every basis
function Ψji of W j is orthogonal to every basis function Φji of V j . From the properties of
the Haar functions, described above, follows that basic wavelet is progressively narrowed
(reduced in scale) by powers of two. Each smaller wavelet is then translated by incre-
ments equal to its width, so that the complete set of wavelets at any scale completely
covers the interval. From mentioned equations one can conclude, that the basic √ wavelet
is scaled down by powers of 2, but its amplitude is scaled up by powers of 2.

4. The Haar and Wavelet Basic Images

Due to its low computing requirements, the Haar transform has been mainly used for
image processing and pattern recognition. From this reason two dimensional signal
processing is an area of efficient applications of Haar transforms due to their wavelet–
like structure.
Because H(n) and H(n)T are the square matrices, their product is commutative,
therefore equations (3) and (4) can be rewritten and expressed as:

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
Piotr Porwik, Agnieszka Lisowska 85

N
X −1 N
X −1
s(k, m) = f (x, y) × haar(k, x) × haar(m, y) , (9)
x=0 y=0

N
X −1 N
X −1
f (x, y) = s(k, m) × haar(k, x) × haar(m, y) , (10)
k=0 m=0

where S = [skm ], F = [fxy ], x, y, k, m ∈ {0, 1, . . . , N − 1}.

Basing on equation of analysis (3) we can conclude that in 2D spectral domain the
values of coefficients depend on appropriate product of the two Haar functions. Fig. 2
presents an example of multiplication the arbitrary selected Haar functions.
y

<0
haar(1,y)

>0

=0

haar(4,x)

Fig. 2. The example of multiplication of the discrete Haar functions.

Because this product is multiplied by image matrix, the result of such multiplication
may be treated as a particular extractor, which allows us to characterise the features
in real image at specified point. In other words, that product can be used to locate
the specific lines (edges) hidden in image. By looking for all coefficients in the spectral
space, one can find all important edge directions in the image. This information can
further be used in image (object) analysis [10].
Taking into account the Mallat algorithm [7, 8, 9] multiresolution image analysis
can be applied to the classical Haar basis, described by matrix–equation (2). In this
case, we must find decomposition matrices of matrix . Based on Fig. 2, construction of
decomposition matrices can be as follows:
Step 1. According to the formula V n = V n−1 ⊕ W n−1 , the matrix M1 has a form
M1 = [Φn−1
j=0,...,2n−1 −1 ⊂ V
n−1
, Ψn−1
j=0,...,2n−1 −1 ⊂ W
n−1 T
] .

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
86 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

Step 2. In second step, because V n−1 = V n−2 ⊕ W n−2 ⊕ W n−1 , the matrix M2 can
be constructed as follows
M2 = [Φn−2
j=0,...,2n−2 −1 ⊂ V
n−2
, Ψn−2
j=0,...,2n−2 −1 ⊂ W
n−2
, Ψn−1
j=0,...,2n−1 −1 ⊂ W
n−1 T
] .
...
Step n. Finally, after n steps of calculations V 1 = V 0 ⊕ W 0 ⊕ W 1 ⊕ W 2 ⊕ . . . ⊕ W n−1 ,
hence the matrix Mn has a structure
Mn = [Φ00 ⊂ V 0 , Ψ00 ⊂ W 0 , Ψ1j=0,1 ⊂ W 1 , . . . , Ψn−1
j=0,...,2n−1 −1 ⊂ W
n−1 T
] .
Exmp. 4.1. Let n = 3 then:
V 3 = V 2 ⊕ W2

M1 = [Φ20 , Φ21 , Φ22 , Φ23 , Ψ20 , Ψ21 , Ψ22 , Ψ23 ]T =

 
2 2 0 0 0 0 0 0

 0 0 2 2 0 0 0 0 


 0 0 0 0 2 2 0 0 


 0 0 0 0 0 0 2 2 


 2 −2 0 0 0 0 0 0 


 0 0 2 −2 0 0 0 0 

 0 0 0 0 2 −2 0 0 
0 0 0 0 0 0 2 −2
V 2 = V 1 ⊕ W1 ⊕ W2

M2 = [Φ10 , Φ11 , Ψ10 , Ψ11 , Ψ2j=0,...,3 ⊂ W 2 ]T =

 √ √ √ √ 
2 2 2 2 √0 √0 √0 √0
 0 0 0 0 2 2 2 2
 √ √ √ √


 2 2 − 2 − 2 √0 √0 0 0 

 0 √ √ 
 0 0 0 2 2 − 2 − 2 

 2 −2 0 0 0 0 0 0 
 
 0 0 2 −2 0 0 0 0 
 
 0 0 0 0 2 −2 0 0 
0 0 0 0 0 0 2 −2
V 1 = V 0 ⊕ W0 ⊕ W1 ⊕ W2

M3 = [Φ00 , Ψ00 , Ψ1j=0,1 ⊂ W 1 , Ψ2j=0,...,3 ⊂ W 2 ]T =

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
Piotr Porwik, Agnieszka Lisowska 87

 
1 1 1 1 1 1 1 1
√1 1
√ 1
√ √1 −1 −1 −1 −1
 
 

 2 2 − 2 − 2 √0 √0 0
√ √0 


 0 0 0 0 2 2 − 2 − 2 


 2 −2 0 0 0 0 0 0 


 0 0 2 −2 0 0 0 0 

 0 0 0 0 2 −2 0 0 
0 0 0 0 0 0 2 −2
For last decomposition level, it can be noticed
√ that Mn = H(n). If each orthogonal ma-
trix Mi , i = 1, 2, 3 will be multiplied by 1/ 2 factor, then the procedure of calculations
will be according to the classical Mallat algorithm.
The product of the decomposition levels for all 2D Haar functions (for case N = 8)
is shown in Fig. 3. Taking into account equation (9), these products can be treated as
extractors of image features.

Fig. 3. The 2D Haar functions product treated as extractors. Decomposition levels: (left) first, (center)
second, (right) third.

The pictures presented in Fig. 3 have been composed on the basis of Mi , i = 1, 2, 3

matrices and method shown in Fig. 2.
From Fig. 3 we can conclude that the classical Haar transform gives different spectral
coefficients on different decomposition levels. One advantage of the method presented
above is that often a large number of the detail coefficients turn out to be very small in
magnitude, as in the example of Fig. 1. Truncating, or removing, these small coefficients
from the representation introduces only small errors in the reconstructed image, giving
a form of lossy image compression. Additionally, we can control which coefficients will
be removed, because its distribution is known (Fig. 3). In other words, that distribution

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
88 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

may be used to locate the specific lines (edges) hidden in image. By looking for all large
coefficients in the spectral space, it can be found all important edge directions in the
image.
There are many ways of forming the Haar spectrum. The simplest method consists of
multiplying the data by the Haar matrices, then setting some of the resulting coefficients
equal to zero (Fig. 1). A widely used method involves the specifying threshold [12, 15, 17].
All coefficients whose magnitudes lie bellow this threshold are set equal to 0. This method
is frequently used for noise removal, where coefficients whose magnitudes are significant
only, because of the added noise will often lie bellow a well–chosen threshold [15]. A
second method keeps only the largest magnitude coefficients, while setting the rest equal
to zero. This method is convenient for making comparison when it is known in advance
how many terms are needed. A third method is called the energy method [15, 17]. This
one involves specifying a fraction of the signal’s energy, where the energy is the square
root of the sum of squares of the coefficients. Then we retain the least number of the
largest magnitude coefficients whose energy exceeds this fraction of the signal’s energy
and set all other coefficients equal to 0. Instead of the above–described methods of
forming the Haar spectrum, we can apply the presented in this paper method of the 2D
Haar function analysis. This new method allows very often obtaining the similar results
through selection of the appropriate coefficients.
Problem of image analysis described above can be solved differently by applying the
same functions and spaces. As it is known, there are two ways of forming the Haar
coefficients: standard and non–standard method [16]. The standard approach has been
presented above. In this paper is applied non–standard method because, as it will be
shown, it has a strong connection with method of the Haar functions analysis described
above. Additionally, non–standard method is slightly more efficient to compute spectral
coefficients [49].
Basing on Definitions 3.1 and 3.2, on properties of the Haar functions and the facts
that V j = span{Φji }i=0,...,2j −1 , W j = span{Ψji }i=0,...,2j −1 we can express functions Φ
and Ψ as a linear combination of the basis functions from V and W spaces respectively.
In the Haar wavelet case Φ and Ψ functions can be written as
√ √
Φ(t) = h(0) 2Φ(2t) + h(1) 2Φ(2t − 1) , (11)
√ √
Ψ(t) = g(0) 2Φ(2t) + g(1) 2Φ(2t − 1) , (12)

where {h(0), h(1)} and {g(0), g(1)} define the low–pass and high–pass filters respectively.
For this case, h(k), k = 0, 1 coefficients are known [5]:

1 1 1 1
h(0) = √ , h(1) = √ , g(0) = √ , g(1) = − √ . (13)
2 2 2 2
Let us denote F as an image in matrix form and define the operators

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp. 79-98
Piotr Porwik, Agnieszka Lisowska 89

1 1
A(i) = √ F(2i) + √ F(2i + 1) , (14)
2 2
1 1
D(i) = √ F(2i) − √ F(2i + 1) , (15)
2 2
where F(i) — vector of size N , containing row or column of matrix F, i ∈ {0, 1, . . . , N2 −
1}, A(i) — vector of size N/2, containing approximation coefficients, D(i) — vector of
size N/2, containing detail coefficients.
To get non–standard wavelet decomposition on the first level of an image F (the
spectrum matrix called S1 ) we first apply the operators (14), (15) to all columns of the
matrix and then to all rows [16]. To get the second level of non–standard decomposition
(that is matrix S2 ) one can apply similar analysis to upper left sub–matrix of size N2 × N2
of matrix S1 . And generally, to get k–th level — matrix Sk , one can apply this analysis
N N
to upper left sub–matrix of size 2k−1 × 2k−1 of matrix Sk−1 , where k ∈ {1, . . . , log2 N }.
Note, that applying filters (14), (15) to an image, give the same results as multiplying
matrices S1 = 18 M1 · F · MT1 , where matrix M1 is taken from Ex. 4.1. Therefore, S1
may be treated as extractor of image features on the first level of wavelet decomposition,
similar as above in Haar decomposition case. Because on the second and next levels only
the part of a matrix is transformed (opposite to Haar decomposition) these extractors
on these levels are different. For example, for N = 8 the products of the non–standard
wavelet decomposition levels are shown in Fig. 4.

Fig. 4. The 2D wavelet functions product treated as extractors. Decomposition levels: (left) first,
(center) second, (right)third.

All considerations, which have been presented until now for the classical of Haar
functions, have applications in that case too, with the exception of extractors distribution
(Fig. 4). Additionally, it is common knowledge that coefficients h(k), k = 0, 1 (in other

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
90 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

words so–called filters) may be described by means of scalar pro duct of functions Φ and
Ψ respectively:
√
h(k) =< Φ(t), 2Φ(2t − k) > (16)

√
g(k) =< Ψ(t), 2Ψ(2t − k) > (17)
Presented both the standard and non–standard approaches are applied very often in
practical solutions [5, 15, 16, 17, 28, 46]. Each of them has some advantages. The stan-
dard decomposition can be simply implemented as matrix multiplication. This presents
equation (3). The non–standard algorithm is a bit more efficient. More compact de-
scription of the non–standard method allows us even increase efficient computation.

5. Experimental Results

In order to test Haar matrix–based method and wavelet approach the well known image–
benchmarks were used. Mentioned benchmarks one can find in many databases. As test
images we used some standard images such as Barbara, Bridge, Lena, Flowers, Mandrill,
Baboon etc. Each of these images was of size a × a × 8 bits, where a ∈ {32, 64, 128, 256},
respectively.
By analysing the Figs. 3–4 we can divide area of each figure into 4 sub–sampled
spaces. Each piece has dimension (N/2)×(N/2) and is called A, H, V and D respectively.
Location of these sub–areas presents Fig 5.

A H

V D

Fig. 5. Principle of spectra partitioning.

Each mentioned piece (A, H, V or D) for N = 8 includes sixteen (4 × 4) appropriate

sub–squares from Fig. 3–4. According to presented previously arguments, mentioned ar-
eas possesses different features: A (Approximation Area) — includes information about
the global properties of analysing image. Removing spectral coefficients from this area
leads to the biggest distortion in origin image. H (Horizontal Area) — includes informa-
tion about the vertical lines hidden in image. Removing spectral coefficients from this

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
Piotr Porwik, Agnieszka Lisowska 91

area excludes horizontal details from origin image. V (Vertical Area) — contains infor-
mation about the horizontal lines hidden in image. Removing spectral coefficients from
this area eliminates vertical details from origin image. D (Diagonal Area) — embraces
information about the diagonal details hidden in image. Removing spectral coefficients
from this area leads to the smallest distortions in origin image.
Fig. 6 presents Baboon — the one of 256 × 256 × 8 grey–level test images together
with its wavelet and Haar spectra. The spectra images are different what directly follows
from Figs. 3–4. Taking into account mentioned above features of areas some differences
between spectres can be shown.

Fig. 6. Original image and its wavelet and Haar spectra, respectively.

Each left image from Figs. 7–10 presents place where spectrum was reduced. In
this place, all spectral coefficients equal to zero. Two remained images show differences
between origin image and the compressed one for wavelet and Haar method of analysis
respectively.

Fig. 7. Horizontal details elimination and lost information after applied wavelet and Haar matrix–based
method, respectively.

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
92 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

Fig. 8. Vertical details elimination and lost information after applied wavelet and Haar matrix–based
method, respectively.

Fig. 9. Diagonal details elimination and lost information after applied wavelet and Haar matrix–based
method, respectively.

Fig. 10. Horizontal and vertical details elimination and lost information after applied wavelet and Haar
matrix–based method, respectively.

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
Piotr Porwik, Agnieszka Lisowska 93

The other cases of removing the spectral coefficients can be applied as well. These
entire processes base on the fact, that appropriate selection and modification of the
spectral coefficients may preserve the contents of the image. Depending on our needs
we can shape compression ratio by means of natural selection of spectral coefficients
directly from areas A, H, V, D or from their combinations. The exact information about
distribution of spectral coefficients allows us to match easily up the compression ratio
with the type of image, but this problem was not considered in the paper. In particular
cases it is possible to obtain the large image compression ratio. Obtained results for
Haar matrix–based method and wavelet method were compared by means of PSNR
coefficients (hence there are some other similarity coefficients, for example like in [45]).
These results of investigations are collected in Table 1. From this table one can see that
Haar reconstructed images have slightly better quality.

Horizontal Vertical Diagonal Horiz.+Vert.

Method (H) (V) (D) (H+V)
Wavelet decomp. 29, 7254 27, 3697 31, 4822 25, 3813
Haar decomp. 29, 7269 27, 3702 31, 4827 25, 3813
Tab. 1. The PSNR of reconstructed images after appropriate details elimination.

Between Haar matrix–based method and wavelet approach the quantitative and
graphic relationship may be observed. Fig. 11 presents those relations. The image from
Fig. 11 was constructed as a quotient between wavelet and Haar spectra respectively.
These relationships will be obviously identical for any image objects.

Fig. 11. Areas of the proportional coefficients (white squares) for Haar and wavelet spectres.

Moreover, it can be proven that for all diagonal coefficients on all levels of decompo-
sition the following proposition is true.

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
94 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

Propos. 5.1. Let dH stands for diagonal coefficients from Haar spectrum matrix
and dW stands for diagonal coefficients from wavelet spectrum matrix, both of degree
2n . Then
dH = 2n dW . (18)
The proof of this fact, describing graphic dependences, follows instantly from Figs. 3–
4 and arguments mentioned above, as well as appropriate proportionality of sub–matrices
from the Haar and wavelet matrices on the same levels of decomposition.
This relationship is well visible in Fig. 11. The white areas appear just due to
the proportionality of diagonal coefficients. Furthermore, note from Table 1 (the last
column), that after removing all horizontal and vertical details on the first level of
decomposition we get exactly the same PSNR of both methods reconstructed images
because of the proportionality of the diagonal detail coefficients.
The amount of proportional coefficients can be easily computed. It is easy to see
that
Pn−1thei number of diagonal detail coefficients in spectrum matrix of size 2n × 2n equals
i
i=0 2 · 2 . Hence, after simple calculations we get the following corollary.
Corollary 5.1. The number of proportional coefficients in Haar and wavelet matri-
ces of size 2n × 2n equals
Pn−1 i i
i=0 2 · 2 1
n n
≈ . (19)
2 ·2 3
It means that as much as 1/3 of all Haar matrix coefficients are proportional to wavelet
matrix coefficients; for all of them the formula (18) is true.

6. Concluding Remarks

Wavelets analysis is a particular time–space scale representation of signals, which has

found a wide range of applications in physics, signal processing and applied mathematics
in the last years. The purpose of this paper was to compare the wavelets and the Haar
functions in two–dimensional space. In many cases, these methods are similar. Presented
in this paper the Haar matrix–based method and wavelet analysis can be used in many
areas of image processing such as denoising, edge detection, edge preserving smoothing
or filtering. In the paper is shown the graphic way of presentation of decomposition levels
for both the Haar matrixbased method and wavelets. As it was shown both methods
may be modelled on the basis of the wavelets theorem.
The 2D Haar matrix method of calculations like the 2D Fast Fourier Transform has
complexity O(4N 2 log2 N ) [14], classical 2 × 1D fast non–standard wavelet method of
calculations has complexity O(16/3N 2 ) only [28]. This complexity can be decreased to
O(14/3N 2 ) by suitable organization of calculations [49]. Described complexity factors
are determined as number of additions and multiplications in computation process. The
graphic distribution of the Haar–wavelet spectral coefficients also was presented. Ad-
ditionally, graphic presentation of spectra distribution allows us to point appropriate

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
Piotr Porwik, Agnieszka Lisowska 95

selection or modification (reduction) of the Haar–wavelet coefficients. For example it

may be removed vertical, horizontal or diagonal details of a given image.
The authors believe that this survey and problem description can be useful to re-
searchers working in different disciplines where the discrete Haar and wavelets transforms
are used.

References

1910
[1] Haar A.: Zur Theorie der orthogonalen Funktionensysteme. Mathematische Annalen, 69, 331–371.
1973
[2] Rudin W.: Functional Analysis. McGraw–Hill, New York.
1975
[3] Ahmed N., Rao K.R.: Orthogonal Transforms for Digital Signal Processing. Springer–Verlag, Berlin.
1977
[4] Harmuth H.F.: Sequency Theory. Foundations and applications. Academic Press, New York.
1978
[5] Pratt W. K.: Digital Image Processing. John Wiley and Sons, New York.
1988
[6] Besslich P. W., Trachtenberg L. A.: The sign transform an invertible non–linear transform with quan-
tized coefficients. In: Moraga C., editor, Theory and application of spectral techniques, University
Dortmund Press.
[7] Jain A.K.: Fundamentals of Digital Image Processing. Prentice Hall.
1989
[8] Mallat S.: Multifrequency channel decompositions of images and wavelet models. IEEE Transaction
in Acoustic Speech and Signal Processing, 37, 2091–2110.
[9] Mallat S. A.: Theory for Multiresolution Signal Decomposition: The Wavelet Representation. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 11(7), 674–693.
[10] Yankowitz D., Bruckstein A.M.: A New Method for Image Segmentation. Computer Vision, Graphics
and Image Processing, 46/1, 82–95.
1992
[11] Daubechies I.: Ten lectures on wavelets. Philadelphia PA, SIAM.
[12] Gröchenig K., Madych W. R.: Multiresolution Analysis, Haar bases and self–similar tilings of Rn .
IEEE Transactions on Information Theory, 38(2), 556–568.
1994
[13] Odegard J.: Image Enhancement by Nonlinear Wavelet Processing. Rice University CML Technical
Report.
1995
[14] Bhaskaran V., Konstantinides K.: Image and video compression standards: algorithms and archi-
tectures. Kluwer, Boston.
[15] Stollniz E. J., DeRose T. D., Salesin D. H.: Wavelets for Computer Graphics: A Primer, Part 1.
IEEE Computer Graphics and Applications, 76–84.
[16] Fournier A.: Wavelets and their Applications in Computer Graphics. SIGGRAPH’95 Course Notes,
University of British Columbia.
1996
[17] Castleman K. R.: Digital Image Processing. Prentice–Hall, New Jersey.
1997
[18] Calderbank A. R., Daubechies I., Sweldens W., Yeo B. L.: Lossless image compression using integer
to integer wavelet transforms. Proceedings of the International Conference on Image Processing, 1,
596–599.
[19] Claypoole R., Davis G., Sweldens W., Baraniuk R.: Adaptive Wavelet Transforms for Image Coding.
Asilomar Conference on Signals, Systems and Computers.

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
96 The Haar–Wavelet Transform in Digital Image Processing: Its Status and Achievements

[20] Creusere C. D.: A new method of robust image compression based on the embedded zerotree wavelet
algorithm. IEEE Transactions on Image Processing, 6, 1436–1442.
[21] Przelaskowski A., Kazubek M., Jamrógiewicz T.: Optimalization of the Wavelet–Based Algorithm
for Increasing the Medical Image Compression Efficiency. Proceedings of the TFTS’97 — 2nd IEEE
UK Symposium on Applications of Time–Frequency and Time–Scale Methods, 177-180.
[22] Walker J. S.: Fourier Analysis and Wavelet Analysis. Notices of the American Mathematical Society,
44(6), 658–670.
[23] Wojtaszczyk P.: A Mathematical Introduction to Wavelets. Cambridge University Press, Cambridge.
1998
[24] Calderbank A. R., Daubechies I., Sweldens W., Yeo B. L.: Wavelet transforms that map integers to
integers. Applied and Computational Harmonics Analysis, 5(3), 332–369.
[25] Daubechies I.: Recent results in wavelet applications. Journal of Electronic Imaging, 7(4), 719–724.
[26] Daubechies L., Sweldens W.: Factoring wavelet transforms into lifting steps. J. Fourier Anal. Appl.,
4(3), 247–269.
[27] Davis G., Nosratinia A.: Wavelet–Based Image Coding: An Overview. Applied and Computational
Control, Signals, and Circuits, 1(1).
[28] Sonka M., Hlavac V., Boyle R.: Image processing, Analysis and Machine Vision. Brooks/Cole Pub-
lishing Comp.
1999
[29] Davis G., Strela V., Turcajova R.; Multivwavelet Construction via the Lifting Scheme. Wavelet
Analysis and Multiresolution Methods, T. X. He (editor), Lecture Notes in Pure and Applied Math-
ematics, Marcel Dekker.
[30] Smith S. W.: The Scientist and Engineer’s Guide to Digital Signal Processing. California Technical
Publishing, San Diego.
2000
[31] Christopoulos C., Skodras A., Ebrahimi T.: The JPEG2000 Still Image Coding System: an Overview.
IEEE Transactions on Consumer Electronics, 46(4), 1103–1127.
[32] Fernandes F., van Spaendonck R., Burrus C.: Directional Complex–Wavelet Processing. Wavelet
Applications in Signal and Image Processing.
[33] Jahromi O. S., Francis B. A., Kwong R. H.: Algebraic theory of optimal filterbanks. Proceedings of
IEEE International Conference on Acoustics, Speech and Signal Processing, 1, 113–116.
[34] Romberg J., Choi H., Baraniuk R., Kingsbury N.: Multiscale Classification using Complex Wavelets.
IEEE International Conference on Image Processing, Vancouver, Canada.
2002
[35] Addison P. S., Watson J. N., Feng T.: Low–Oscillation Complex Wavelets. Journal of Sound and
Vibration, 254(4), 733–762.
[36] Blu T., Unser M.: Wavelets, Fractals and Radial Basis Functions, IEEE Transactions on Signal
Processing, 50(3), 543–553.
[37] Munoz A., Ertle R., Unser M.: Continuous wavelet transform with arbitrary scales and O(N ) com-
plexity. Signal Processing, 82, 749–757.
[38] Resnikoff H., Wells R. O. Jr.: Wavelet Analysis. Springer–Verlag, New York.
[39] Zeng L., Jansen C. P., Marsch S., Unser M., Hunziker R.: Four–Dimensional Wavelet Compression of
Arbitrarily Sized Echocardiographic Data. IEEE Transactions on Medical Imaging, 21(9), 1179–1188.
2003
[40] Drori I., Lischinski D.: Fast Multiresolution Image Operations in the Wavelet Domain. IEEE Trans-
actions on Visualization and Computer Graphics, 9(3), 395–411.
[41] Fernandes F., Selesnick I., van Spaendonck R., Burrus C.: Complex Wavelet Transforms with Allpass
Filters. Signal Processing, 83(8), 1689–1706.
[42] Jorgensen P.: Matrix Factorizations, Algorithms, Wavelets. Notices of the American Mathematical
Society, 50(8), 880–894.
[43] Jahromi O. S., Francis B. A., Kwong R. H.: Algebraic theory of optimal filterbanks. IEEE Transac-
tions on Signal Processing, 51, 442–457.
[44] Lisowska A.: Nonlinear Weighted Median Filters in Dyadic Decomposition of Images. Annales UMCS
Informatica AI, 1, 157–164.
[45] Prieto M. S., Allen A. R.: A Similarity Metric for Edge Images. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 25(10), 1265–1273.

Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004, pp.79-98
 Piotr Porwik, Agnieszka Lisowska 97

[46] Romberg J., Wakin M., Baraniuk R.: Multiscale Geometric Image Processing. SPIE Visual Com-
munications and Image Processing, Lugano, Switzerland.
[47] Yitzhaky Y., Peli E.: A Method for Objective Edge Detection Evaluation and Detector Parameter
Selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(8), 1027–1033.
[48] Zhang J. K., Davidson T. N., Wong K. M.: Efficient design of orthonormal wavelet bases for signal
representation. IEEE Transactions on Signal Processing.
2004
[49] Lisowska A., Porwik P.: New Extended Wavelet Method of 2D Signal Decomposition Based on Haar
Transform. Mathematics and Computers in Simulation, Elsevier Journal, (to appear).
[50] Porwik P., Lisowska A.: The New Graphic Description of the Haar Wavelet Transform. Lecture
Notes in Computer Science, Springer–Verlag, Berlin, Heidelberg, New York, 3039, 1–8.

. pp .79-98
Machine GRAPHICS & VISION vol. 13, no. 1/2, 2004,

View publication stats