Wavelet Transform and Signal Denoising using

Wavelet Method
Çiğdem POLAT DAUTOV
Mehmet Siraç ÖZERDEM
Electrical and Electronics Engineering
Electrical and Electronics Engineering
Bitlis Eren University
Dicle University
Bitlis, Turkey
Diyarbakır, Turkey
cigdembilmuh@hotmail.com
sozerdem@dicle.edu.tr

Abstract—Over the last decade, a great progress has been different sections of frequency. Thus, detecting the corruption
made in the signal processing field. Especially new signal within the signal and eliminating it becomes a relatively easy
processing methods such as Wavelet Transform (WT) allowed process.
researchers to solve diverse and complicated signal processing
issues. The paper provides answers to several questions related to There are two types of WT defined as Continues Wavelet
WT technique such as what WT is, how and why WT emerged, Transform (CWT) and Discrete Wavelet Transform (DWT).
what WT types currently available. The main advantages like CWT’s mathematical representation is given as:
noise reduction and compression of WT are also explained in this
study. A set of MATLAB experiments were carried out in order CWT ( = d (1)
to illustrate the use of WT as a signal denoising tool. Analysis on where x(t) is the original signal , ψ(t) is the analyzing function
different signals contaminated with noise are performed. (wavelet), a is a scale parameter and b is a position in time.
Different types of thresholding and mother wavelets were applied
and the outcome of the experiments indicate that Daubechies CWT uses inner products of analyzing function and the
family along with the soft thresholding technique suited our original CWT signal and measures the similarity between these
application the most. The study proves that choosing the right two functions by integration [2]. Changing scale parameter and
thresholding technique and wavelet family is vital for the success position parameter, CWT allows us to shift and compress or
of signal denoising applications. stretch the mother wavelet in order to capture multiple
Keywords — Daubachies; signal denoising; thresholding;
frequency components at different locations (Figure 1).
Wavelet transform.

I. INTRODUCTION
Despite the fact that WT is relatively new method, it is
important to state that a considerably important growth in the
theory and practice of wavelets occurred in last 10 years.
Wavelets are mathematical tools that divide the original signal
into different frequency components and examine each
component individually. Depending on the frequency, the basis
functions of Wavelet Transform (WT) are scaled. There are Figure 1. Scaling and positioning by CWT.
different small waves (also known as mother wavelets) that can
be used for the implementation of WT [1]. The mother wavelet Considering ψ(t) as the bandpass impulse response, scaling
is actually a window function that moves forward in time. the wavelet varies the bandwidth of the bandpass. Contrary to
Unlike STFT, it is possible to use different types of window STFT, CWT allows changing the support of the wavelet to get
functions by using these mother wavelets and performing the better resolution in frequency domain. In Figure 2, we can
decomposition of the signal x(t) into weighted set of scaled observe the resolution difference between CWT and STFT. As
wavelet functions y(t). we see there are overlaps in the magnitude scalogram of STFT
while there is a clear separation between [3].
Dubachies, Haar, Symlet, Coiflet, Mexican Hat, Morlet
wavelets can be given as some of the different types of
wavelets. Each of them comes from different wavelet families
and includes different properties [1]. When we are applying
these wavelets, we decide which one to use depending on the
requirements of the application.
The main idea behind the WT is decomposing a signal into
different frequency level of coefficients. Fundamental
capability of analyzing signal in different frequency bands and
scales can be helpful determining signals behaviors’ in
 A. Algorithm
The algorithm that we used for de-noising our sample
signals consist of three different fundamental iterations [4].
The scheme of the Wavelet de-noising algorithm is given in the
Figure 3.

Figure 2. Magnitude scalograms of CWT and STFT.

As we mentioned earlier, another type of WT exists in

literature called DWT which is an implementation of WT using
mutually orthogonal set of wavelets defined by carefully
chosen scaling and translation parameters (a and b). This leads
to a very simple and efficient iterative scheme for doing the
transformation [3]. The scaling function should meet several
requirements such as orthogonality, the normalized area Figure 3. The algorithm steps based on reference study [4].
between the functions etc… The translation equation is given In order to de-noise any signal, we need to put the noisy
as: signal into the decomposition process by applying wavelet
transform. Wavelet transform allows us to decompose signal
into groups of coefficients at different frequency levels [5].
After successfully applying the first stage, it is possible to
(2) obtain relevant information related to the signal characteristic.
Understanding how signal behaves in different frequency
where n is delay parameter, N is the length of signal, ψ is the segments will allow us to select the most suitable threshold in
discretized mother wavelet. the next stage. The next step is determining the best threshold
Since in DWT, computation performed on discrete set of values and applying threshold values to these set of coefficients
wavelets, it provides a significant yield in terms of so it is possible to eliminate unwanted data. The thresholding
computational time. Therefore, DWT is useful when technique will be covered in the next section. Final step of the
compression of the signal is needed. However, this property algorithm is where we use these filtered coefficient sets to
comes at a price, especially when the main purpose is to recompose our signal. For this to be possible an inverse
analyze data to discover hidden information. discrete Wavelet transform (IDWT) is done.

II. WAVELET DENOISING B. Thresholding

Two different experiments were conducted in order to show The most important thing in applications of thresholding is
how wavelet denoising works. For the first experiment, we the determination of threshold value. There are four different
analyzed a male speech using the given different forms of estimator of threshold value given as heasure, minimax, rigsure
audio samples: desired (original) signal, signal containing and sqtwolog. Minimax and SURE threshold selection rules are
keyboard sound as a noise and noise sound itself. For the more convenient when small details of the signal lie near the
second experiment, a sample female sound is analyzed. For this noise range [4] [6]. For the other cases, the rest two give
experiment, we had only noisy signal form and no other outcomes that are more efficient.
information. For the experiments the sampling frequency is Thresholding can be also applied in two different methods:
defined as Fs=1600 Hz. DWT method is used as WT hard and soft thresholding. Hard threshold can be explained as
algorithm in order to eliminate the noise with adaptive setting elements to zero where their absolute values are lower
thresholding. By running the algorithm steps iteratively, we than the threshold. Hard threshold gives sharper results. On the
established the ideal threshold technique, type and mother other hand, soft thresholding softens the coefficients exceeding
wavelets for each experimental sample. It should be noted that the threshold by lowering them as much as threshold value.
even though this method provides good performance, it Soft thresholding requires more computations but gives better
requires great amount of computations. denoising performance [4] [7].
 III. MATLAB EXPERIMENTS represent the data after thresholding, we also need to move
forward and finish the last stage where we reconstruct our
A. Experiment I: A male subject typing and reading signal as a denoised version of the original signal. In the last
In the first experiment, we used a sample male speech step, the results of the de-noising function were represented as
record where the subject reads and types on the computer original, noisy and de-noised signal given in Figure 6.
keyboard. The desired output of the first analysis and de-
noising process is also known as record with no keyboard
sound. The keyboard sound itself is determined as the noise.
The first thing we want to examine is plotting all the records
defined as desired, noisy signal (x) and noise (Figure 4 and 5).

Figure 6. Comparision of original, noise and denoised signals.

B. Experiment II: A female speech with unknown noise

In the second experiment, we used a sample female speech
Figure 4. Original and noisy signals in time domain. record where the subject counts several digits. The desired
output of the first analysis and de-noising process is unknown.
The first thing we want to do is plotting the record, the noisy
signal (x) (Figure 7).

Figure 5. Noise in time domain. Figure 7. Noisy signals in time domain.

In order to eliminate the noise in the signal the first step is In next step, we used db4 since it is one of the most suitable
decomposition of signal into approximation and detail wavelets. In this experiment different from the previous one,
coefficients. In this step, we could do one level wavelet we did not need to do manually multiple level tresholding.
transform but we decided to move forward using three Since the built-in auto denoising function was enough to
(level=3) level wavelet transform. As a form of wavelet, we calculate and remove the noise from the signal, we just used
used db8 since it is one of the most suitable wavelets in this this function to denoise our signal at level 5. The threshold
case. As an outcome of the process, the last approximation technique used by the auto denoising function is defined as
coefficients and detailed coefficients at level one, two and three modwtsqtwolog.
were plotted as an illustration. Next step will be de-noising the
signal by using threshold technique. The most important thing Due to necessity of evaluation of the results, we used SNR
we should focus on is selecting the right threshold method and as a metric for assessing the outcomes. For that, we need to
calculation. We move forward with soft thresholding because know pre SNR values. The SNR values gathered after the
the signal and noise exist in similar frequency levels and applications of different wavelets families are given in Table 1.
similar amplitudes. For not losing much of an information, we As we can observe, the best SNR outcome is provided by the
apply soft thresholding. The threshold calculation method is Daubachies Wavelets.
picked as hearsure for the same reasoning. In order to
 [7] Patil, R. (2015). Noise Reduction using Wavelet Transform and
TABLE 1. OUTPUT SNR VALUES OF SEVERAL WAVELETS Singular Vector Decomposition. Procedia Computer Science,
54, 849–853. https://doi.org/10.1016/j.procs.2015.06.099
Family Wavelet SNR
13.817
Daubechies db20
13.828
Daubechies db8
13.711
Coif coif2
13.741
Coif Coif5
11.503
Bior bior 3.1
Bior bior 3.2 11.573

IV. CONCLUSIONS
In this study, we have provided information related to
Wavelets theory, continuous and discrete Wavelet Transform.
Even though it is possible to find similar review studies on
WT, it is aimed to show that WT is a powerful tool for signal
analysis with the target of noise elimination. At the end, a set
of experiments were held where WT denoising with various
underlying wavelet functions and thresholding techniques were
applied in two different settings; first with a male subject
reading a book out loud and simultaneously typing on a
keyboard where keyboard stroke sound was treated as noise. In
another one, female was saying digits with additive white noise
in the background. In both cases DWT successfully clear the
signal from the known/unknown noise. For the second case, the
simplest universal soft thresholding technique under db4 of
level 3 was sufficient to get adequate results. At the same time,
this method did not work well for the first case. We found that
db8 of level 3 with adaptive thresholding gave the best results
among the applied wavelet families in reducing keyboard
strokes sound in the background. An important lesson taken in
this study is unique set of threshold and wavelet choice
depending on the nature of application is vital when
eliminating noise from a signal
REFERENCES
[1] Cengiz, Y., Doç, Y., & Arıöz, U. (2016). Ayrık Dalgacık Dönü ¸
sümü Kullanarak Konu ¸ sma Sinyallerinin Gürültüden
Arındırılması için Uygulama An Application for Speech
Denoising Using Discrete Wavelet Transform, 1–4.
[2] Federico, A., & Kaufmann, G. H. (2009). Wavelet Transform,
34(15), 2336–2338.
[3] Misiti, M., Misiti, Y., Oppenheim, G., & Poggi, J. (2007).
Wavelets and their Applications.
https://doi.org/10.1002/9780470612491
[4] Aggarwal, R., Rathore, S., Singh, J. K., Tiwari, M., India, M. P.,
Gupta, V. K., & Khare, A. (2011). Noise Reduction of Speech
Signal using Wavelet Transform with Modified Universal
Threshold. International Journal of Computer Applications,
20(5), 975–8887.
[5] Patil, S. S., & Pawar, M. K. (2012). Quality advancement of
EEG by wavelet denoising for biomedical analysis. Proceedings
- 2012 International Conference on Communication,
Information and Computing Technology, ICCICT 2012, 1–6.
https://doi.org/10.1109/ICCICT.2012.6398151
[6] Yadav, T. (2016). Denoising and SNR Improvement of ECG
Signals Using Wavelet Based Techniques, (October), 678–682.