IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.

9, September 2008 87

A Wavelet based Statistical Method for De-Noising of

Ocular Artifacts in EEG Signals
P. Senthil Kumar 1, R. Arumuganathan 1, K. Sivakumar 2 , C. Vimal 3
1
Dept. of Mathematics & Computer Applications, 2 Dept. of Electronics and Communication Engg.,
3
Dept. of Bio-Medical Engineering
PSG College of Technology, Coimbatore, India.

Summary Current Independent Component Analysis (ICA) methods

This paper presents a new empirical method for de-noising of of artifact removal require a tedious visual classification of
ocular artifacts in the electroencephalogram (EEG) records. In the components. P. LeVan [1] proposed a method which
many biomedical signal processing approach, source signals are automates this process and removes simultaneously
noisy and some have kurtosis close to zero. These noise sources multiple types of artifacts. R.J. Croft [2] reviews a
increase the difficulty in analyzing the EEG and obtaining the
clinical information. To remove this artifacts a method based
number of methods of dealing with ocular artifact in the
on Donoho’s de-noising method is used. Recently Stationary EEG, focusing on the relative merits of a variety of EOG
Wavelet Transform (SWT) has been used to de-noise the correction procedures. In EEG data sets, there may be
corrupted EEG signals. In this paper, statistical empirical some specific components or events that may help the
method for removing ocular artifacts from EEG recordings clinicians in diagnosis. They may tend to be transient
through SWT is suggested. (localized in time), prominent over certain scalp regions
(localized in space) and restricted to certain ranges of
Key words: temporal and spatial frequencies (localized in scale).
EEG, de-noising, ocular artifacts, stationary wavelet transform
There has been a tremendous amount of activity and
interest in the applications of wavelet analysis to signals,
in particular methods of wavelet thresholding and
1. Introduction shrinkage [8,13] for the removal of additive noise from
The statistical analysis of electrical recordings of the brain corrupted biomedical signals and images. Wavelet
activity by an Electroencephalogram is a major problem in analysis provides flexible control over the resolution with
Neuroscience. Cerebral signals have several origins that which neuro-electric components and events are localized
lead to the complexity of their identification. Therefore, in time, space and scale. V.J. Samar [4] describes the
the noise removal is of the prime necessity to make easier basic concepts of wavelet analysis and other applications.
data interpretation and representation and to recover the V. Krishnaveni [9,10] discussed a method to automatically
signal that matches perfectly a brain functioning. A identify slow varying ocular artifact zones and applying
common problem faced during the clinical recording of the wavelet based adaptive thresholding algorithm only to the
EEG signal, are the eye-blinks and movement of the eye identified ocular artifact zones, which avoids the removal
balls that produce ocular artifacts. It has been known for of background EEG information. The fundamental
quite some time now that the Alpha rhythm of the EEG, motivation behind these approaches is that the statistics of
which is the principal resting rhythm of the brain in adults many real world signals, when wavelet transformed are
while they are awake, is directly influenced by visual substantially simplified.
stimuli. Auditory and mental arithmetic tasks with the
eyes closed lead to strong alpha waves, which are Wavelet transforms are used to analyze time varying,
suppressed when the eyes are opened. This property of non-stationary signals, and EEG fall into these category of
the EEG has been used, ineffectively, for a long period of signals. The ability of wavelet transform is to accurately
time to detect eye blinks and movements. The slow resolve EEG into specific time and frequency components
response of thresholding, failure to detect fast eye blinks lead to several analysis applications and one among them
and the lack of an effective de-noising technique forced is de-noising. The wavelet transform of the noisy signal
researchers to study the frequency characteristics of the generates the wavelet coefficients which denote the
EEG as well. correlation coefficients between the noisy EEG and the
wavelet function. Depending on the choice of mother

Manuscript received September 5, 2008.

Manuscript revised September 20, 2008.
88 IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.9, September 2008

wavelet function, larger coefficients will be generated property. Instead of sub sampling, the SWT utilizes
corresponding to the noise affected zones. The larger recursively dilated filters in order to halve the bandwidth
coefficients will be an estimate of noise from level to another.
Tatjana Zikov [12] proposed a wavelet based de-noising
of the EEG signal to correct for the presence of the ocular This decomposition scheme is shown in the figure 1.
artifact. In this paper, we proposed a simple statistical
empirical de-noising formula for removing artifacts in the
EEG signals without using any reference signals. This c j −1 …
formula very much reduces the complexity and time factor. h0 (n)
h0 (n) cj
2. Wavelets for analyzing EEG signals c j +1
h1 (n) d j −1
In statistical settings we are more usually concerned with
discretely sampled, rather than continuous functions. It is
then the wavelet analogy to the Discrete Wavelet h1 (n) dj
Transform ( DWT) which is of primary interest . Wavelet
transform [6] has emerged as one of the superior technique
in analyzing non-stationary signals like EEG [3,7]. Its Fig. 1 Wavelet Decomposition Scheme
capability in transforming a time domain signal into time
and frequency localization helps to understand more the 3. A Simple De-Noising Technique
behavior of a signal.
Suppose one has to measure a signal on which an external
The DWT means choosing subsets of the scales ‘a’ and noise is superimposed. We call this EEG signal the true
positions ‘b’ of the mother wavelet ψ (t ) . signal S(t) and the external noise ε (t ) , so that the
a measured signal can be written in the form
Ψ( a ,b ) (t ) = 2 2 ψ ( 2 a t − b) (1)
x(t) = S(t) + ε (t ) (2)
Choosing scales and positions are based on powers of two,
which are called dyadic scales and positions { a j = 2 –j ; The only assumptions needed are that S(t) and ε (t ) are
bj,k = 2 –j k } ( j and k are integers ) . Equation (1) shows
uncorrelated and are stationary processes, and can be
that it is possible to build a wavelet for any function by
written as equation (2). Thresholding is a technique used
dilating a function ψ (t ) with a coefficient 2 j, and for signal and image de-noising. When we decompose a
translating the resulting function on a grid whose interval signal using the wavelet transform, we are left with a set of
is proportional to 2 –j . Contracted (compressed) versions wavelet coefficients that correlates to the high frequency
of the wavelet function match the high-frequency subbands. These high frequency subbands consist of the
components, while dilated (stretched) versions match the details in the data set. If these details are small enough,
low-frequency components. Then, by correlating the they might be omitted without substantially affecting the
original signal with wavelet functions of different sizes, main features of the data set.
the details of the signal can be obtained at several scales. The de-noising of EEG signal is carried out by using
These correlations with the different wavelet functions can different combinations of threshold limit, thresholding
be arranged in a hierarchical scheme called function and window sizes. Choice of threshold limit and
multi-resolution decomposition. The multi-resolution thresholding function is a crucial step in the denoising
decomposition algorithm [5] separates the signal into procedure, as it should not remove the original signal
“details” at different scales and a coarser representation of coefficients leading to loss of critical information in the
the signal named “approximation”. The basic DWT analyzed data
algorithm can be modified to give a Stationary Wavelet
Transform (SWT) [14] that no longer depends on the In this paper, the following thresholding ( statistical
choice of origin. As a consequence of the sub sampling empirical ) formula is used for calculating the thresholding
operations in the pyramidal algorithm, the DWT does not limits. This formula produces better de-noised results
preserve translation invariance. This means that a than [11], which is applied to the entire length of the
translation of the original signal does not necessarily imply signal.
a translation of the corresponding wavelet coefficients.
The SWT has been introduced in order to preserve this
IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.9, September 2008 89

Threshold based on Statistics of the signal 250

200

Threshold value

Amplitude (µV)
150

⎛ x −σ ⎞
Tk = N × ⎜⎜ ⎟
100

(3)
⎟
⎝ x +σ
50

⎠ 0

Window length - 10 seconds -50

0 1 2 3 4 5 6 7 8 9 10
Time (Sec)

where N is a positive integer , ranging from 100 to 150

Fig.2 EOG Contaminated EEG
x - mean of all samples
σ - standard deviation of all samples The de-noising of EEG signal is carried out by using
threshold limit, threshold function and window size.
Choice of threshold limit and thresholding function is a
The thresholding function used in this work are as
crucial step in the de-noising procedure, as it should not
follows :
remove the original signal coefficients leading to loss of
If wavelet coefficient value > threshold value
critical information in the analyzed data. Figure 3
shows a 10 second epoch of EOG contaminated EEG with
then new wavelet its corrected version using our proposed method.
coefficient value = ( - 0.7) * ( wavelet coefficient value ) 250

200
else new wavelet
coefficient value = (old) wavelet coefficient value 150
Amplitude (µV)

100
4. Methodology
50

The general goal of this study is to removal of artifacts 0

from EEG signals. Towards this, we propose the

following method. To estimate the signal x(t), we -50
0 1 2 3 4 5
Time (Sec)
6 7 8 9 10

propose the following method :

Fig.3 Contaminated EEG and Corrected EEG
(i) The Stationary Wavelet Transform (SWT) to the
Figure 4 shows the power spectra of the contaminated
contaminated EEG signal with Symlet (sym3) as a
EEG and the corrected EEG. From this figure it is shown
basis function and decomposes upto eight levels
that, the powers of the spectral components have been
(ii) To calculate statistical measures , mean and standard
fully retained. The cross correlation between the noisy
deviation for the entire length of the signal.
EEG and EOG is shown in figure 5. This shows how
(iii) To apply equation (3) to fix the suitable threshold
close both the signals are in terms of the shape.
value.
(iv) To apply threshold function to fix the wavelet 60
EEG with artifact
coefficients in a new position. Corrected EEG
40
(v) To apply wavelet reconstruction procedure to
Amplitude (dB)

reconstruct the EEG signal. 20

0
5. Results and Discussion
-20

EEG data with ocular artifacts are taken from http:// -40

www.sccn.ucsd.edu/~arno/famzdata/publicly_ 0 10 20 30
Frequency (Hz)
40 50 60

available_EEG_data.html for testing the proposed method.

The data is sampled at a rate of 128 samples per second. Fig. 4 Power Spectra Plot
The effect of ocular artifacts is dominant in the Frontal and
Fronto-polar channels like FP1, FP2, F7 and F8. Hence
it is sufficient to apply the method to these channels.
Consider the 10 second EOG contaminated EEG epoch
(sampled at a rate of 128 samples/second) shown in Fig.2
90 IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.9, September 2008

1 250
Corrected EEG
EEG with artifact
0.8
200
Correlation c oeffic ient

0.6
150

Amplitude (uV)
0.4
100

0.2
50

0
0 10 20 30 40 50 60
0
Frequency (Hz)

Fig. 5 Correlation Plot -50

0 1 2 3 4 5 6 7 8 9 10
Time(sec)

Fig. 8 Proposed method

The results obtained has been compared with the existing
methods. Figure 6 and 7 shows the artifact removal using
[11,12]. The proposed method shows a better result when 250
EEG With Mean+2*Std
EEG With 1.5*std
compared with [11] which is depicted in figure 8,9 and 200
EEG With Proposed Method

10. One can observed that the artifacts in EEG signals EEG with Artifacts

are considerably reduced using the proposed method and 150

shown in figure 9 and 10. Amplitude (uV)

100

250 50
Corrected EEG
EEG with artifact
200 0

150 -50
0 1 2 3 4 5 6 7 8 9 10
Amplitude (uV)

Time(sec)

100
(a)
50
500

0 0

-500
-50
0 1 2 3 4 5 6 7 8 9 10
0 1 2 3 4 5 6 7 8 9 10
Time(sec) 100

Fig.6 Modified Threshold function 0

(mean + 2*standard deviation )
-100
0 1 2 3 4 5 6 7 8 9 10
100
250
Corrected EEG 0
EEG with artifact
200 -100
0 1 2 3 4 5 6 7 8 9 10
50
150
Amplitude (uV)

0
100
-50
0 1 2 3 4 5 6 7 8 9 10
50

(b) (i) EEG with artifacts

0
(ii) Modified Threshold function
(iii) Threshold function
-50
0 1 2 3 4 5
Time(sec)
6 7 8 9 10 (iv) Proposed method
Fig. 7 Threshold function Fig. 9 (a), (b) Comparison results
( 1.5*standard deviation )
IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.9, September 2008 91

x 10
5 [4] Vincent J. Samar, Ajit Bopardikar, Raghuveer Rao and
16
EEG With Mean+2*Std
Kenneth Swartz, “ Wavelet analysis of neuroelectric
14 EEG With 1.5*std waveforms: A conceptual tutorial “ , Brain and Language,
EEG With Proposed Method 66, pp. 7 – 60, 1999.
12
[5] S.G. Mallat, “ A Theory for Multiresolution signal
10 decomposition : The Wavelet representation “, IEEE Trans.
on Pattern Analysis and Machine Intell.,Vol. 2 (7), pp. 674 –
Amplitude (uV)

8
693, 1989.
6
[6] C. Sidney Burrus, Ramesh A Gopinath and Haitao Guo,
4 “ Introduction to Wavelets and Wavelet Transforms “,
Prentice-Hall international Inc, 1998.
2
[7] W.J. Tompkins, “ Biomedical digital signal processing“,
0 New Jersey, Prentice Hall, 1993
-2 [8] D.L. Donoho, “ De-noising by Soft Thresholding”, IEEE
Trans. Info.Theory, Vol. 41, pp. 613-627, 1995
-4
0 1 2 3 4 5 6 7 8 9 10 [9] V. Krishnaveni, S. Jayaraman, S. Aravind, V. Hariharasudhan
Time(sec) and K. Ramadoss, “ Automatic identification and removal of
Fig. 10 Cross Correlation Plot ocular artifacts from EEG using Wavelet transform“,
Measurement Science Review, Vol. 6 (4), 2006.
[10].V. Krishnaveni, S. Jayaraman, L. Anitha and K. Ramadoss,
“ Removal of Ocular Artifacts from EEG using Adaptive
6. Conclusion thresholding of Wavelet coefficients “, Journal of Neural
Engineering, Vol.3, pp.338-346, 2006
The accuracy of the technique has been checked on several [11] V. Krishnaveni, S. Jayaraman, N. Malmurugan,
artifical signals. In this paper, a method to remove ocular A.Kandasamy and D. Ramadoss,“Non adaptive thresholding
methods for correcting ocular artifacts in EEG“, Academic
artifacts using a new threshold formula and threshold
Open Internet Journal, Vol.13, 2004
function is given. Our method gives a better result without [12] T Zikov, S Bibian, G.A. Dumont , M. Huzmezan and
any complexity and also retains the original information C.R.Ries, “ A wavelet based de-nosing technique for
contained in the EEG signal. Power Spectral Density ocular artifact correction of the electroencephalogram“,
plot and Correlation plot are used as performance metrics. 24th Annual conference of the IEEE Engineering in
We conclude that our proposed statistical method gives Medicine and Biology, Vol.1, pp.98-105, 2002
lesser complexity and easier to remove the artifacts with [13] D.L. Donoho and I.M. Johnstone, “ Adapting to
the help of wavelet decomposition. It is an efficient unknown smoothness via wavelet shrinkage”, Journal of
technique for improving the quality of EEG signals in the American Stat. Association, 90(432), pp. 1200-1224,
1995
biomedical analysis.
[14] G.P. Nason and B.W. Silverman, “ The Stationary Wavelet
Transform and some Statistical Applications “, Tech. Rep.
BS8 1Tw, University of Bristol, (1995).
Acknowledgement

The authors would like to express their deep gratitude to P. Senthil Kumar received the
M.Sc.,and M.Phil degrees in
Prof. S. Jayaraman, Head, Dept. of Electronics and
Mathematics from Bharathiar Univ.
Communication Engg., PSG College of Technology, India in 1987 and 2001 respectively.
Coimbatore. He is working as a Senior Lecturer in
Department. of Mathematics and
References Computer Applications, PSG College
[1] P. LeVan, E. Urrestarrazu and J. Gotman, “A system for of Technology, Coimbatore. India. He
automatic artifact removal in ictal scalp EEG based on is pursuing his Ph.D. under the
independent component analysis and Bayesian classifi- guidance of Prof. R. Arumuganathan.
-cation”, Clinical Neurophysiology, Vol. 117(4), pp. 912- His research interests include wavelet transforms and linear
927, 2006. algebra. He is a life member of ISTE.
[2] R.J. Croft, R.J. Barry, “ Removal of ocular artifact from the
EEG : a review “, Clinical Neurophysiology, Vol.30(1),
pp. 5 – 19, 2000.
[3] M. Akay, “Time Frequency and Wavelets in Biomedical
Signal Processing ”, IEEE Press series in Biomedical
Engineering, 1998
92 IJCSNS International Journal of Computer Science and Network Security, VOL.8 No.9, September 2008

Dr. R.Arumuganathan
received the M.Sc., M.Phil., degrees in
Mathematics from Madurai Kamaraj
Univ., India. He received Doctoral
degree from Bharathiar Univ. India in
1998. Now he is working as a
Professor in the Dept. of Mathematics
and Computer Applications., PSG
College of Technology, Coimbatore,
India. His research interests include
queueing theory and wavelet transforms.

K. Sivakumar received the

B.E and M.E. degrees in Electronics
and Communication Engineering from
Anna Univ. India in 2003 and 200
respectively. His area of interests are
digital signal processing and
bio-medical signal processing.

C. Vimal is a student of dept. of Bio-Medical Engineering,

PSG College of Technology, Coimbatore. India. His area of
interests are bio-medical signal processing and medical imaging.