Signal Processing Applications of Wavelets

Arthur Asuncion
Information and Computer Science
University of California, Irvine
[email protected]

ABSTRACT: Wavelets have great utility in the area of digital

signal processing. A digital signal can be represented
Wavelets are powerful mechanisms for analyzing and as a summation of wavelets that are fundamentally
processing digital signals. The wavelet transform identical except for the translation and dilation
translates the time-amplitude representation of a factors (or coefficients). Hence, a signal can be
signal to a time-frequency representation that is represented entirely by wavelet coefficients. These
encapsulated as a set of wavelet coefficients. These coefficients provide important frequency and
wavelet coefficients can be manipulated in a temporal information which can be used to analyze a
frequency-dependent manner to achieve various signal. Furthermore, the signal can be processed in
digital signal processing effects. The inverse wavelet the wavelet coefficient domain before being
transform can then convert the manipulated wavelet transformed back to the normal time-amplitude
coefficients back to the normal time-amplitude representation. Thus, wavelets facilitate a unique
representation in order to yield a modified signal. framework for digital signal processing.

After an overview of Fourier and wavelet transforms, FOURIER TRANSFORM

the Haar wavelet and the Daubechies wavelet are
described in this paper. Several signal processing Wavelet literature typically includes overviews of
and musical applications of wavelets, including Fourier analysis and the Fourier transform [1][3],
denoising, wavelet filtering, and data compression, since the Fourier transform is a conceptual precursor
are investigated. A Java implementation of a to the wavelet transform. Two centuries ago, Joseph
wavelet-based effects processor is also presented. Fourier showed that any periodic signal can be
decomposed into a summation of sinusoids.
KEYWORDS:
The Fourier transform converts a signal into a
Wavelet Transform – Haar Wavelet – Daubechies frequency spectrum derived from the frequencies of
Wavelet – Digital Effects the sinusoids. The Fourier transform is presented
below:
AN INTRODUCTION TO WAVELETS ∞

∫ x(t )e
− jω t
X(f ) = dt
From digital signal processing to computer vision, −∞
wavelets have been widely utilized to analyze and
transform discrete data. The concept of wavelets is The exponential factor represents the sinusoidal
rooted in many disciplines, including mathematics, component (via Euler’s Relation), f represents a
physics, and engineering [1]. The 1980s witnessed a particular frequency, and x(t) represents the input
new wave of wavelet discoveries, like multiresolution signal as a function of time [4]. Essentially, this
analysis and orthonormal compactly supported integral is an inner product which correlates the input
wavelets. These advances have revolutionized the signal with the sinusoidal component. A fast
field and have led to many novel applications of discrete method that computes the integral above is
wavelets. known as the Fast Fourier Transform.
A wavelet, which literally means little wave, is an Due to the Heisenberg uncertainty principle, a
oscillating zero-average function that is well fundamental tradeoff exists between frequency
localized in a small period of time. A wavelet resolution and time resolution [4]. In the Fourier
function, known as a mother wavelet, gives rise to a transform, a longer input time signal increases the
family of wavelets that are translated (shifted) and accuracy of frequency information at the cost of
dilated (stretched or compressed) versions of the losing temporal information. While the Fourier
original mother wavelet [2]. transform is an excellent tool for spectral analysis,
the Fourier transform is not capable of splitting the
time-frequency tradeoff into fine levels of granularity aj
based on frequency.
H G
WAVELET TRANSFORM
↓2 ↓2
The wavelet transform is a fine-grained approach that
seeks to achieve an optimal balance between
aj+1 cj+1
frequency resolution and time resolution. At higher
frequencies, the transform gains temporal H G
information in exchange for a loss in frequency
information, while at lower frequencies, the ↓2 ↓2
transform gains frequency information in exchange
…
for a loss in temporal information. This fine-grained cj+2
approach in handling the tradeoff is useful for digital
signal and music applications, since transients Figure 1. Fast Wavelet Transform Using Filters
normally occur at high frequencies (thus needing a
higher time resolution), and lower frequencies HAAR WAVELET
usually require a higher frequency resolution.
The Haar wavelet, which Alfred Haar discovered in
Like the Fourier transform, the wavelet transform can 1910, is both powerful and pedagogically simple.
be represented as an integral: The basic Haar wavelet is a piecewise constant
function that is defined as follows [5]:
∞
Wx(u , s ) = ∫ x(t )Ψ u ,s (t ) dt  1, 0 ≤ r < 12

−∞ Ψ [ 0,1[ (r ) =  − 1, 12 ≤ r < 1
 0, otherwise
In the integral above, the input signal x(t) is 
correlated with the wavelet with translation
parameter u and dilation parameter s [3]. This
transform converts a signal into coefficients that
represent both time and frequency information, with 1
more time resolution at high frequencies, and more
frequency resolution at low frequencies. The dilation
of the wavelet enables the fine-grained tradeoff to
occur. As with the Fourier transform, there are fast
1/2 1
discrete ways of computing the wavelet transform. -1
One fast way of computing a wavelet transform is Figure 2. The Standard Haar Wavelet
with a cascade of filters [3]. The input signal is fed
into two filters, H and G. The filters produce two The Haar wavelet transform recursively replaces
sets of coefficients which are both down-sampled by adjacent pairs of steps in the signal with a wider step
a factor of 2. As shown in Figure 1, this procedure is and a wavelet [5]. A step φ is a function that is 1 in a
recursively applied to the set of coefficient that continuous region and zero everywhere else.
comes out of the H filter. One assumption of the
wavelet transform is that the number of samples in Consider a simple signal f of two samples: {7, 1}.
the input signal is a power of 2. If the number of The Haar wavelet transform calculates the average
samples is not a power of 2, the signal can be zero- value coefficient, (7 + 1)/2, and the change
padded to achieve this criterion. coefficient (7-1)/2. The average value is the
coefficient for the wider step, while the change value
is the coefficient for the standard Haar wavelet [5].
The transform is presented below, and a graphical
depiction is shown in Figure 3.
 (7 + 1) (7 − 1)
f = ϕ [ 0,1[ + Ψ[ 0,1[ {3, -1, 4, 8, 0, -2, 7, 1}
2 2

{1, 6, -1, 4, 2, -2, 1, 3}

7

1 {3.5, 1.5, -2.5, -2.5 2, -2, 1, 3}

1/2 1

{2.5, 1, -2.5, -2.5 2, -2, 1, 3}

= Figure 4. Derivation of Haar wavelet transform
coefficients, ordered from low to high frequencies

4 One difference between the Fourier transform and the

wavelet transform is that the frequency channels (or
bins) of the Fourier transform are equally spaced,
1/2 1 while the frequency channels of the wavelet
transform are logarithmically spaced [4].

In Figure 4, the frequency of the rightmost channel,

which has coefficients of 2, -2, 1, 3, is twice as great
+ as the frequency of the middle channel, which has
coefficients of -2.5, -2.5. Likewise, the frequency of
the middle channel is twice the frequency of the
3 lowest channel, which has a coefficient of 1. The
left-most result of the Haar wavelet transform, 2.5, is
1/2 1 the average value of the whole signal.
-3
DAUBECHIES WAVELET

Figure 3. Haar Wavelet Transform on Signal with 2 Named after Ingrid Daubechies, the Daubechies
Samples wavelet is more complicated than the Haar wavelet.
Daubechies wavelets are continuous; thus, they are
Consider another signal f that has 8 values: {3, -1, 4, more computationally expensive to use than the Haar
8, 0, -2, 7, 1}. The Haar wavelet transform on this wavelet, which is discrete [5].
signal follows the procedure shown in Figure 1. The
wavelet transform needs to undergo log(8)=3 sweeps,
with the recursion being applied to the average value
coefficients. Figure 4 details the derivation of the
wavelet transform of signal f.

Figure 5: Daubechies Wavelet

The Daubechies 4 filter can be used to perform the

Daubechies wavelet transform. The inverse filter can
then reconstruct the signal out of the wavelet
coefficients. The following ‘lifted’ filter equations different ways to visually represent wavelet
encapsulate the Daubechies wavelet transform [6]: coefficients. A time-frequency plane can be used to
depict the distribution of energy among time and
1. a[n] = S [2n ] + 3S [2n+ 1] frequency [6]. This distribution of energy is
determined by wavelet coefficients.
3 3−2
2. c[n] = S [2n ] + a[n ] − a[n − 1] AWE can visualize the wavelet transform
4 4 coefficients in two different ways. Like a spectrum
3. a[n ] = a[n] − c[n + 1] plot, the first visualization simply graphs the wavelet
transform coefficients in order from low to high
3 −1
4. a[n] = a[n] frequencies, with logarithmically spacing. One
consequence of the logarithmic spacing is that the
2
highest frequency occupies the entire right half of the
3 +1
5. c[n] = c[n ]
plot. The second visualization is similar in concept
to the sonogram [4] and the time-frequency plane.
2 The wavelet coefficients are mapped onto a triangular
shape. The apex of the triangle represents the
These equations are performed in sequential order for coefficient for the lowest frequency, while the bottom
each iteration of the wavelet transform. S stores the of the triangle represents the coefficients for the
signal array, a stores the low-pass coefficients, and c highest frequency. Each row in the triangle
stores the high-pass coefficients. Succeeding represents a different frequency channel. This
iterations of the Daubechies wavelet transform would triangular representation of the wavelet coefficients
then apply these filter equations to the low pass inversely corresponds to the wavelet domain of
coefficients. This transform follows the same general influence in time, which is also a triangle [4].
process that was depicted in Figure 1.

The implementation of the inverse Daubechies

wavelet transform is straightforward. The inverse
algorithm can be achieved by reversing the order of
the equations above and reversing the operations
within those equations. Thus, the signal S can be
reconstructed from a and c coefficients. Figure 6: First visualization in AWE

APPLICATIONS OF WAVELETS

This section outlines several major applications of

wavelets in signal processing. A Java program called
AWE (Art’s Wavelet Effects) has been implemented
by the author in order to test the utility of wavelet Figure 7: Second visualization in AWE
transforms in signal processing.
DENOISING
ANALYSIS & VISUALIZATION
Noise removal is a commonly cited application of
While excellent methods for spectral analysis already wavelets [2][5][6]; however, denoising appears to be
exist (like Fourier analysis), wavelets can also be a used more in image processing than in sound
powerful tools for analyzing signals. The wavelet processing. Denoising is essentially accomplished by
transform facilitates multiresolution analysis, since amplifying or reducing certain frequency channels in
the wavelet transform can be recursively applied to order to mitigate noise.
the signal to achieve any level of accuracy. In
essence, the wavelet transform allows for multiple If the signal is known to be within a certain range of
perspectives on the signal, from a coarse-grained frequencies, then all of the frequency channels
overview of the signal, to a detailed accurate view of outside of this range can be turned off in order to
the signal. reduce noise. If random noise is in the signal, the
removal of small variations within the signal can help
The visualization of wavelet coefficients is an to denoise the signal [6]. Small signal variations can
integral aspect of analyzing a signal. There exist
be reduced by setting the coefficients of the highest wavelet coefficients can be achieved. AWE also
frequency channel to zero. provides a demonstration of cross-synthesis. In
AWE, the wavelet coefficients of one signal act as
AWE facilitates the denoising of signals. Frequency thresholds for the wavelet coefficients of the other
channels can be both strengthened and weakened signal.
using AWE’s filter interface.
Many other applications of wavelets, like pitch
WAVELET FILTERING shifting and the comb wavelet transform, are
described in the literature [4].
Using the same approach as denoising, the wavelet
coefficients can be manipulated to achieve a simple COMPUTATIONAL PERFORMANCE ISSUES
form of equalization. Channels can be amplified to
highlight certain frequencies or reduced to suppress While the wavelet transforms may not be as
certain frequencies [4]. Thus, filtering can be done in computationally efficient as the Fast Fourier
the wavelet domain. AWE allows for the Transform, the transforms can still be accomplished
manipulation of ten frequency channels. near real-time. AWE implements both the Haar and
Daubechies transforms along with various signal
COMPRESSION effects, and the sound still proceeds smoothly. The
visualization of the wavelet coefficients is the most
Another widely cited application of wavelets is the computationally expensive aspect in AWE. Thus, the
data compression of signals. Wavelets are being used program provides a mechanism to stop the
in new JPEG and MPEG standards to compress data. visualization in order to ensure the efficient
processing of sound.
Compression can be accomplished through several
ways. One way is to pass the wavelet coefficients CONCLUSION
through a threshold function. If a wavelet coefficient
is above a specified threshold, then this coefficient is Wavelet-based signal processing can be achieved by
important, since it provides some measurable manipulating wavelet coefficients. Denoising,
contribution to the signal. This coefficient would be filtering, and data compression are all possible in the
kept. However, if the coefficient is below the wavelet domain. Different wavelets can be used in
specified threshold, the compression scheme would the forward and inverse wavelet transforms, and the
turn the coefficient to zero. particular wavelet that is used influences the sonic
characteristics of the effects.
Data compression can also be accomplished
removing the high frequencies from the signal. Since ACKNOWLEDGEMENTS
the frequencies are logarithmically spaced, the higher
frequencies need much more memory space than the DSP discussions with Professor Dobrian were helpful
lower frequencies. Assuming that frequencies are in writing this paper. Professor Meyer in EECS also
spaced by powers of two, the removal of only the top provided some useful insight into the Haar wavelet.
two frequency channels reduces the storage space
needed to ¼ of the original space. REFERENCES

AWE provides a simulation of wavelet data [1] Debnath, Lokenath. Wavelet Transforms &
compression. Since the Daubechies wavelet is Their Applications. Birkhauser, 2002. P. 12.
continuous and the Haar wavelet is discrete, data [2] Teolis, Anthony. Computational Signal
compression is more sonically pleasing with the Processing with Wavelets. Birkhauser, 1998. P. 60-
Daubechies wavelet than with the Haar wavelet. 61, 173-174.
[3] Mallat, Stephane. A Wavelet Tour of Signal
REAL-TIME MUSICAL EFFECTS Processing. 2nd ed. Academic Press, 1999. P. 5, 257.
[4] Roads, Curtis. The Computer Music Tutorial.
Other interesting sound effects can be generated by MIT Press, 1996. P. 446, 581-589, 1087.
manipulating the wavelet coefficients. Modulation [5] Nievergelt, Yves. Wavelets Made Easy.
effects can be produced by multiplying the Birkhauser, 1999. P. 3-60.
coefficients by a cosine wave. AWE provides a [6] Jensen, A, and Cour-Harbo, A. Ripples in
simple demonstration of this modulation effect. Mathematics: The Discrete Wavelet Transform.
Furthermore, cross-synthesis among two sets of Springer, 2001. P. 102, 155-158.