Performance Comparison of Known ICA Algorithms to a Wavelet-ICA Merger

Janett Walters-Williams & Yan Li
Signal Processing: An International Journal, Volume (5) : Issue (3) : 2011 80
Performance Comparison of Known ICA Algorithms to a
Wavelet-ICA Merger
Janett Walters-Williams jwalters@utech.edu.jm
School of Computing & Information Technology
Faculty of Engineering & Computing
University of Technology, Jamaica
237 Old Hope Road, Kingston 6, Jamaica W.I.
Yan Li liyan@usq.edu.au
Department of Mathematics & Computing
Faculty of Sciences
University of Southern Queensland
Toowoomba, Australia
Abstract
These signals are however contaminated with artifacts which must be removed to have pure EEG
signals. These artifacts can be removed by Independent Component Analysis (ICA). In this paper
we studied the performance of three ICA algorithms (FastICA, JADE, and Radical) as well as our
newly developed ICA technique. Comparing these ICA algorithms, it is observed that our new
techniques perform as well as these algorithms at denoising EEG signals.
Keywords: Independent Component Analysis, Wavelet Transform, Unscented Kalman Filter,
Electroencephalogram
1. INTRODUCTION
The use of Electroencephalogram in the field of Medicine has had a great impact on the study of
the human brain. The signals received have several origins however that lead to the complexity of
their identification. This complexity is made of both the pure EEG signal and other non-cerebral
signals called artifacts or noise. The artifacts have resulted in the contamination of the EEG
signals, hence the removal of these artifacts has generated a large number of denoising
techniques.
One method has been Independent Component Analysis (ICA) originating from the field of Blind
Source Separation [5]. This technique calls for the separation of the EEG into its constituent
independent components (ICs) and then eliminating the ICs that are believed to contribute to the
artifact sources. It is subjective, inconvenient and a time consuming process when dealing with
large amount of EEG data. Another method employed is wavelet transformation. This technique
calls for the decomposition of the EEG signals into wavelets and artifacts removal done using
thresholding and shrinkage.
Each of the above techniques presents their own limitations. In our opinion a combination of the
two should produce a more effective technique. This is possible as each technique is used to
overcome the limitation of the other. We present in this paper therefore a new method of
extracting artifacts from EEG signals – Cycle Spinning Wavelet Transform ICA (CTICA). CTICA is
compared to other known ICA algorithms, and saving useful EEG data.

2. SUPPORTING LITERATURE
2.1 EEG Signals
The language of communication with the nervous system is electric so when the neurons of the
human brain process information, they do so by changing the flow of electrical currents across
their membranes. These changing currents generate electric and magnetic fields that can be
recorded from the surface of the scalp. The electric fields are measured by attaching small
electrodes to the scalp. The potentials between different electrodes are then amplified and
recorded as the electroencephalogram; (EEG), which means the writing out of the electrical
activity of the brain (that which is inside the head). EEG recordings therefore, show the overall
activity of the millions of neurons in the brain.
There are five basic wave types, measured in Hertz (HZ), found in EEG signals (Tab. 1). The
most prominent type is the alpha rhythm recorded mainly over the posterior regions of the scalp
close to the places in the brain that process visual information. When the eyes are open the alpha
rhythm is very small and when the eyes are closed it becomes large.
Type Frequency(Hz) Normally
Delta (δ) 0.5-4 Hz
Deep, dreamless sleep, non-REM
sleep, unconscious
Theta (θ) 4 – 8 Hz
Intuitive, creative, recall, fantasy,
imaginary, dream
Alpha (α) 8 – 13 Hz
Relaxed, but not drowsy, tranquil,
conscious
Beta (β) 13 – 30 Hz
Formerly SMR, relaxed yet
focused, integrated, Thinking,
aware of self & surroundings,
Alertness, agitation
Gamma (γ) 30 – 100+ Hz
Motor Functions, higher mental
activity
TABLE 1: Wave Types Found in EEG Signals (adapted from Neurosky Inc. 2009 Brain Wave Signal (EEG)
of NeuroSky, Inc.)
Since an EEG is used to analyzed brain function it is used in clinical practice to:
(i) Diagnose epilepsy and see what type of seizures is occurring. EEG is the most useful
and important test in confirming a diagnosis of epilepsy.
(ii) Check for problems with loss of consciousness or dementia.
(iii) Help find out a person's chance of recovery after a change in consciousness.
(iv) Find out if a person who is in a coma is brain-dead.
(v) Study sleep disorders, such as narcolepsy.
(vi) Watch brain activity while a person is receiving general anesthesia during brain surgery.
(vii)Help find out if a person has a physical problem (problems in the brain, spinal cord, or
nervous system) or a mental health problem.
Being a physical system however, EEG is subjected to random disturbance. The measurements
or observations are generally contaminated with other non-cerebral signals called artifacts or
noise caused by the electronic and mechanical components of the measuring devices. These
may include EOG (Eye-induced) artifacts (includes eye blinks and eye movements); EKG (Fig 1)
(cardiac) artifacts; EMG (muscle activation)-induced artifacts; and Glossokinetic (chewing &
sucking movement) artifacts. Artifacts sometimes mimic EEG signals and overlay these signals
resulting in distortion making analysis impossible. In clinical practice areas in the reading with
artifacts are cancelled resulting in considerable information loss, thus sometimes resulting in
misdiagnosis.

FIGURE 1: EEG Signal corrupted with ECG/EKG and line signals (adapted from Artifact Removal from EEG
Signals using Adaptive Filters In Cascade, A Garcés Correa et al, Journal of Physics: Conference Series 90,
2007)
Artifacts must be eliminated or attenuated to ensure correct analysis and diagnosis. Through the
years there have been different methods of denoising such as artifacts rejection, regression and
Principal Components Analysis (PCA). More recently two other methods have been discussed –
Independent Component Analysis (ICA) and Wavelet Transform (WT).
2.2 Independent Component Analysis
When a signal is contaminated it is a combination of the true signal S(t) and the artifacts ε(t)
producing equation (1) where c(t) is the contaminated signal.
( ) ( ) ( )c t S t tε= +
(1)
Researchers have been utilizing ICA to remove ε(t).
ICA is an extension of PCA which originated from the field of Blind Source Separation. It is
suitable for performing source separation where
(i) sources are independent
(ii) propagation delays of mixing medium are negligible
(iii) source are analog with pdfs not too unlike the gradient of a logistic sigmoid
(iv) the number of independent signals sources is the same as the number of sensors.
Investigations show that EEG satisfies (i) since there are statistically independent brain
processes, (ii) since the volume conduction in the brain tissue is efficiently instantaneous. The
assumption of (iii) is plausible but the assumption that EEG signals are a linear mixture of exactly
N sources is questionable since we are do not know the effective number of statistically
independent brain signals contributing to the EEG recorded from the scalp [19]. ICA can
therefore be used to performance separations on these signals. There are problems with using
ICA however
(i) Its performance depends however on the length of the dataset, because the larger
the set the more likely person will have to deal with an over complete ICA which
cannot separate artifacts from the signals.
(ii) When ICA performs separations sometimes some useful signals maybe removed as
a part of the artifacts resulting in information loss [11].

2.3 Wavelet Transform
Wavelet analysis, a sub brand of applied mathematics has been used to decompose signals in
the time frequency scale plane (fig 2). It has been found to be an efficient technique for non-
stationary signal processing of which EEG falls. [1] [22]. Its capability to transform the EEG time
domain signal into time and frequency localization helps researchers understand more the
behaviour of the signals.
FIGURE 2: Demonstration of (a) a wave and (b) a wavelet. Notice that the wave has an easily discernible
frequency while the wavelet has a pseudo frequency in that the frequency varies slightly over the length of
the wavelet. (adapted from D.L. Fugal. 2009. Conceptual Wavelets in Digital Signal Processing: An in depth
Practical Approach for the Non-Mathematician, Space & Signals Technologies LLC)
There are two basic types of wavelet transform. One type of wavelet transform is designed to be
easily reversible (invertible); that means the original signal can be easily recovered after it has
been transformed. This kind of wavelet transform is used for image compression and cleaning
(noise and blur reduction). Typically, the wavelet transform of the image is first computed, the
wavelet representation is then modified appropriately, and then the wavelet transform is reversed
(inverted) to obtain a new image.
The second type of wavelet transform is designed for signal analysis for study of EEG or other
biomedical signals. In these cases, a modified form of the original signal is not needed and the
wavelet transform need not be inverted (it can be done in principle, but requires a lot of
computation time in comparison with the first type of wavelet transform). Decomposition- into
wavelets is done by a “mother and “father” wavelet function. These “mother” functions include
Haar, Daubechies and Mexican Hat. Equation (2) shows that it is possible to build a wavelet for
any function by dilating the mother wavelet function ψ(t) with a coefficient 2
j
, and translating the
resulting function on a grid whose interval is proportional to 2
–j
.
2
( , ) ( ) 2 ( 2 )
a
a
a b t t bψΨ = −
(2)
Compressed versions of the wavelet function match the high-frequency components, while
stretched versions match the low-frequency components. By correlating the original signal with
wavelet functions of different sizes, the details of the signal can be obtained at several scales or
moments. These correlations with the different wavelet functions can be arranged in a
hierarchical scheme called multi-resolution decomposition. The multi-resolution decomposition
algorithm separates the signal into “details” at different moments and wavelet coefficients [22]
[23]. These coefficients are called the Discrete Wavelet Transform (DWT) of the signal. As the
moments increase the amplitude of the discrete details become smaller however the coefficients
of the useful signals increase [27] [28].
If the details are small enough they might be omitted without substantially affecting the main
signals. This omission is done through Thresholding. There are two main ways to denoise a
signal in WT – soft and hard thresholding. Research as shown that soft-thresholding has better
mathematical characteristics [27] [28] and provides smoother results [9]

2.4 Unscented Kalman Filter
Unscented Kalman Filter (UKF) is a Bayesian filter which uses minimum mean-squared error
(MMSE) as the criterion to measure optimality [4][34]. For highly nonlinear systems, the linear
estimate of the nonlinear model does not provide a good approximation of the model, and the
Extended Kalman Filter (EKF) will not track signals around sharp turning points. Another problem
with the EKF is that the estimated covariance matrix tends to underestimate the true covariance
matrix and therefore risks becoming inconsistent in the statistical sense without the addition of
"stabilising noise". UKF was found to address these flaws. It involves the Unscented
Transformation (UT), a method used to calculate the first and second order statistics of the
outputs of nonlinear systems with Gaussian. The nonlinear stochastic system used for the
algorithm is:
1k k k k
k k k
x A x B u v
y H x w
+ = + +
= +
(3)
where A and H are the known and constant matrices respectively, xk is the unobserved state of
the system, uk is a known exogenous input, yk is the observed measurement signal, vk is the
process noise and wk is the measurement noise.
FIGURE 3: Noisy EEG and its Wavelet Transform at different scales (adapted from Weidong Z.,
Yingyuan, L. 2001. EEG Multi-resolution Analysis using Wavelet Transform, 23rd
Annual
International Conference of the IEEE Engineering in Medicine and Biology Society
(IEEE/EMBS) 2001)
UKF uses the intuition that it is easier to approximate a probability distribution function rather than
to approximate an arbitrary nonlinear function or transformation. Following this intuition, a set of
sample points, called sigma points, are generated around the mean, which are thenpropagated
through the nonlinear map to get a more accurate estimation of the mean and covariance of the
mapping results. In this way, it avoids the need to calculate the Jacobian, which for complex
functions can be a difficult task in itself (i.e., requiring complicated derivatives if done analytically
or being computationally costly if done numerically).
3. PREVIOUS RESEARCH
WT and ICA in recent years have often been used in Signal Processing. [22] [27]. Although ICA is
popular and for the most part does not result in much data loss; its performance depends on the
size of the data set i.e. the number of signals. The larger the set, the higher the probability that

the effective number of sources will overcome the number of channels (fixed over time), resulting
in an over complete ICA. This algorithm might not be able to separate noise from the signals.
Another problem with ICA algorithms has to do with the signals in frequency domain. Although
noise has different distinguishing features, once they overlap the EEG signals ICA cannot filter
them without discarding the true signals as well. This results in data loss.
WT utilizes the distinguishing features of the noise however. Once wavelet coefficients are
created, noise can be identified. Decomposition is done at different levels (L); DWT produces
different scale effects (Fig 3). Weidong et al. [25] proved that as scales increase the WT of EEG
and noise present different inclination. Noise concentrates on scale 21, decreasing significantly
when the scale increases, while EEG concentrates on the 22-25 scales. Elimination of the smaller
scales denoise the EEG signals. WT therefore removes any overlapping of noise and EEG
signals that ICA cannot filter out.
More recently there has been research comparing the denoising techniques of both. It was found
(i) If artifacts and signals are nearly the same or higher amplitude, wavelets had difficultly
distinguishing them. ICA on the other hand looks at the underlying distributions thus
distinguishing each [29].
(ii) ICA gives high performance when datasets are large. It suffers from the trade off between a
small data set and high performance [11].
Research therefore shows that ICA and wavelets complement each other, removing the
limitations of each [29]. Since then research as been done applying a combination of both with
ICA as a per- or post- denoising tool. Inuso et al. [11] used them where ICA and wavelets are
joint. They found that their method outperformed the pre- and post- ICA models.
4. RESEARCH DATASETS
EEG data was taken from two sites
(i) http://sccn.ucsd.edu/~arno/fam2data/publicly_available_EEG_data.html. The signals
from here are contaminated with EOG. Data is sampled at a rate of 128 samples per
second recorded from 32 electrodes at 1000Hz
(ii) http://www.filewatcher.com/b/ftp/ftp.ieee.org/uploads/press/rangayyan.0.0.html. Data was
collected at a sampling rate of 1000Hz but noise free. These signals had to artificially
contaminated
These two sites produce signals of different sizes as well as 1D and 2D signals.
5. METHODOLOGY
When a signal is decomposed it is represented as a set of wavelet coefficients that correlates to
high frequency sub-bands. Artifacts are usually of low frequency and can be removed by
shrinkage or thresholding. Research has shown however that thresholding has a slow response
[22] [23]. In this paper we are presently another method to denoising using WT and ICA. Some of
the ideas appear in earlier algorithms however the main difference of CTICA is the use of cycle
spinning; the merger of Wavelet Transform and ICA into one and the improvement of denoising.
The presented method is based on decomposition by using Symmlets which is a near symmetric
extension of Daubechies. Symlets are orthogonal and its regularity increases with the increase in
the number of moments [6]. After experiments the number of vanishing moments chosen is 8
(Sym8).

FIGURE 4: Proposed Artifacts Removal System
A block diagram representation of the proposed work is shown in FIGURE 4. EEGs are acquired
and Cycle Spinning applied. Cycle Spinning utilizes the periodic time invariance of the wavelet
transform to separate noise from signals. The EEG signals are then decomposed by Forward
DWT using the Symmlet family of wavelets. The wavelet coefficients are separated into
statistically independent sources using ICA and denoising takes place. Each IC is then filtered
using UKF. Finally, the sources that are identified as non-artifacts are used to reconstruct the
artifact-free EEGs and Cycle Spinning applied again.
6. RESULTS & DISCUSSION
We conducted experiments, using the above mentioned signals, in Matlab 7.8.0 (R2009) on a
laptop with AMD Athlon 64x2 Dual-core Processor 1.80GHz. Noisy signals were generated by
adding noise to the original noise-free signals and the length of all signals, N, were truncated to
lengths of power of twos i.e. 2
x
.
FastICA Jade Radical CT-ICA
4.1954 4.1909 4.1865 4.1912
7.1276 7.1191 7.1106 7.1192
5.1226 5.1281 5.1226 5.1278
8.0569 8.0484 8.0399 8.0487
7.8736 7.8827 7.8736 7.8815
3.5646 3.5696 3.5646 3.5703
6.0995 6.1057 6.0995 6.1042
2.733 2.7364 2.733 2.7361
0.1374 0.1373 0.1372 0.1373
8.658 8.6521 8.6462 8.6499
0.284 0.2841 0.284 0.2841
0.2234 0.2235 0.2234 0.2235
3.2436 3.2395 3.2355 3.2387
TABLE 2: MSE for 13 EEG signals (x.xe+07)
6.1 Testing Against Known ICA Algorithms
We compared the performance of our method with several state-of-the art ICA algorithms -
FastICA, Radical, and Jade. All the algorithms were downloaded from the web sites of the
respective authors. In the case of FastICA a symmetrical view based on the tan score function
was used for comparison. To determine the quality of each signal the Mean Square Error (MSE),
the Peak Signal to Noise Ratio (PSNR), the Signal to Distortion Ratio (SDR), the Signal to noise
Ratio (SNR) and the Amari Performance Index were calculated.
ICA
Denoising
Decomposed
into Wavelets
Raw
EEG
Filtering
using UKF
Reconstruct
Signal
Independent
Components
Wavelets
Pure
EEG

MSE measures the average of the square of the "error" and defined as:
2
1 1
1
[ ( , ) '( , )]
M N
y x
MSE I x y I x y
MN = =
= −∑ ∑
(4)
The error is the amount by which the estimator differs from the quantity to be estimated. The
difference occurs because of randomness or because the estimator doesn't account for
information that could produce a more accurate estimate. TABLE 2 shows the MSE for 13
signals. Observations show that there is not much difference in the MSE for each algorithm. The
lower the MSE the lesser the error on the signal; it can seen that on average our method
performed better than FastICA and Jade. Radical had a lower MSE.
-18.0969 -18.0923 -18.0877 -18.0926
-20.3987 -20.3935 -20.3883 -20.3935
-18.9641 -18.9687 -18.9641 -18.9685
-20.9309 -20.9263 -20.9217 -20.9265
-20.8309 -20.836 -20.8309 -20.8353
-17.3893 -17.3954 -17.3893 -17.3962
-19.7221 -19.7266 -19.7221 -19.7255
-16.2355 -16.241 -16.2355 -16.2405
-23.2477 -23.2453 -23.243 -23.2442
-21.2434 -21.2404 -21.2375 -21.2393
-26.4025 -26.4042 -26.4025 -26.404
-25.359 -25.3609 -25.359 -25.3611
-16.9794 -16.974 -16.9686 -16.9729
TABLE 3: PSNR for 13 EEG signals
PSNR is the ratio between the maximum possible power of a signal and the power of corrupting
noise that affects the fidelity of its representation. It is defined as:
2
1010 log ( )
MAX
PSNR
MSE
= ×
(5)
Because many signals have a very wide dynamic range, PSNR is usually expressed in terms of
the logarithmic decibel scale. In this research MAX takes the value of 255. Tab 3 shows the
PSNR for 13 signals. If the PSNR is high then the ratio of signal to noise is higher and therefore
the algorithm is considered good.
After experiments it can be seen that our algorithm has the same PSNR on average. It was also
seen that it has a higher PSNR than Jade and Radical. The similar signal to noise ratio can be
seen in the SNR graph in figure 5 where only Jade has a different value.

FIGURE 5: SNR results for 13 signals
The accuracy of the separation for each algorithm in terms of the signals can be calculated by the
total SDR defined as:
( )
2
1
2
1
( )
( , ) 1,...
( ) ( )
L
i
n
i i L
i i
n
x n
SD R x y i m
y n x n
=
=
= =
−
∑
∑ (6)
where xi(n) is the original source signal and yi(n) is the reconstructed signal. When SDR are
calculated any found below 8-10dB are considered to fail separation. Fig 5 shows that all four
algorithms had SDR above 8dB. It also shows that CTICA had SDR very close to the other four
so that there was no differentiation in the graph.
FIGURE 6: SDR results for 13 signals

The global accuracy of the separation of each algorithm was tested using the Amari performance
index defined as:
, 1
| | | |1
1
2 max | | max | |
m
ij ij
err
i j k ik k kj
p p
P
m p p=
 
= + −  
 
∑
(7)
where pij = (BA)ij. It assesses the quality of the de-mixing matrix W for separating observations
generated by the mixing matrix A. The lower the Amari index, the more accurate the separation
is. We have normalized all values of the Amari index to be between 0 and 1 (the max). The Amari
indexes obtained for the different algorithms and for different sample sizes are presented in
TABLE 4. Observations show that the Amari indexes for our method is very similar to those of
Jade and FastICA. On average however it has a lower Amari than both FastICA and Jade but not
Radical.
1238 1237 1236 1237
1583 1582 1581 1582
1363 1363 1363 1363
1669 1668 1667 1668
1652 1653 1652 1653
1140 1141 1140 1141
1477 1478 1477 1478
989 990 989 990
2069 2068 2068 2068
1720 1720 1719 1720
2683 2683 2683 2683
2471 2471 2471 2471
1085 1085 1084 1084
TABLE 4: Amari Test Results for 13 EEG signals (x.xe-05)
6.2 Testing against Known WT-influenced Algorithms
Zhou et al. [28] in 2004 found that a combination of wavelet threshold de-noising and ICA
resulted in the removal of electromyogram (EMG) and electrocardiograph (ECG) artifacts from
EEG signals. Further research in 2007 by Inuso et al. [11] resulted in the creation of a new
technique for EEG artifact removal, based on the joint use of Wavelet transform and Independent
Component Analysis (WICA). After comparison to pre- and post- ICA and wavelet denoising
using artificial artifact-laden EEG datasets they found that this combination had the best artifact
separation performance for every kind of artifact also allowing for the minimum information loss.
These show that a merger of WT and ICA is more effective.
Pre-WT Post-WT WT-UKF WT-ICA CT-ICA
33.9443 1.1158e3 1.1025 1.1051 1.0947
29.0936 1.0438 1.0499 1.0379 1.0372
23.9498 1.0058 997.4019 982.4991 979.2423
TABLE 5: Sample MSE for 3 EEG signals
Sameni et al. [21] experimented with denoising using EKF on ECG data. They found that the
results show that the EKF may be used as a powerful tool for the extraction of the ECG signals
from noisy measurements. Jacob and Martin [12] tested a combination of WT and Weiner Filter.
They concluded that this combination basic denoising using only WT

Pre-WT Post-WT WT-UKF WT-ICA CT-ICA
32.8231 17.655 17.7071 17.6966 17.7377
33.4928 17.9446 17.9192 17.9693 17.9722
34.3378 18.1058 18.1421 18.2075 18.2219
TABLE 6: Sample PSNR for 3 EEG signals
As stated before ICA and WT complement each other, removing the limitations of each [29];
researchers have shown that the combination of WT and ICA is more effective than ICA or WT
alone supporting this theory. They have also shown that the performance of WT improves with
the addition of Filters. In our research investigations have shown that when compared to the
post- and pre- ICA models, a combination of WT with (i) ICA, or (ii) UKF we have found as seen
in Tables 5 and 6 that the merger of all three outperformed all except the Pre-ICA model. This
conforms to the findings of researchers.
7. CONCLUSION
In recent years researchers have used both ICA algorithms and WT to denoise EEG signals. In
this paper we propose a new method – Cycle Spinning Wavelet Transform ICA (CTICA). From
the experiments we can conclude the following for CTICA
(I) It can be seen from the experiments that it can successfully separate noise from EEG
signals.
(II) It has outperformed FastICA and JADE as far as MSE was concerned,
(III) It has outperformed JADE and Radical with PSNR.
(IV) It has the similar in SDR and Amari index
(V) It outperforms different WT model designs except for the Pre-ICA model.
Based on these results it can be concluded that CTICA has an overall performance which is
better than all three ICA algorithms and most WT model, i.e. it is the most consistent and robust
denoising method.
8. REFERENCES
[1] M. Alfaouri and K. Daqrouq. “ECG Signal Denoising By Wavelet Transform Thresholding”.
American Journal of Applied Sciences 5 (3), pp. 276-281, 2008.
[2] Z. Chen, “Bayesian Filtering: From Kalman Filters to Particle Filters, and Beyond”. Adaptive
Systems Lab., McMaster University., Hamilton, Ontario, Canada. Internet:
http://users.isr.ist.utl.pt/~jpg/tfc0607/chen_bayesian.pdf. 2003, [Jun 20, 2009]
[3] S. Choi, A. Cichocki, L. Zhang, and S. Amari. “Approximate Maximum Likelihood Source
Separation Using the Natural Gradient”. In Proc. IEEE Workshop on Signal Processing
Advances in Vireless Communications, (Taoyuan, Taiwan), 2001, pp. 235-238.
[4] R.R. Coifman, and D.L. Donoho. “Translation Invariant Denoising”. Lecture Notes in
Statistics: Wavelets and Statistics, pp. 125-150. 1995.
[5] P. Comon. “Independent Component Analysis, a new concept?” Signal Processing,
Elsevier, 36(3), pp. 287-314, 1994.
[6] B. Ferguson and D. Abbott. “Denoising Techniques for Terahertz Response of Biological
Samples”. Microelectronics Journal 32, pp. 943-953, 1991.
[7] N. Gadhok and W. Kinsner. “Robust ICA for Cognitive Informatics”. International Journal of
Cognitive Informatics and Natural Intelligence (IJCINI) 2(4), pp. 86-92, 2008.

[8] Y.M. Hawwar, A.M. Reza, and R.D. Turney. “Filtering(Denoising) in the Wavelet Transform
Domain”. Unpublished, Department of Electrical Engineering and Computer Science,
University of Wisconsin-Milwaukee, 2002.
[9] A. Graps. “An Introduction to Wavelets”. IEEE Journal of Computational Science and
Engineering 2(2), pp. 1-17, 1995.
[10] G.G. Herrero, and K. Egiazarian. “Independent Component Analysis by a Resampling
Strategy”. Internet: http://www.cs.tut.fi/~gomezher/projects/bss/rica/rica.pdf, 2005 [Sep 18,
2009]
[11] G. Inuso, F. La Foresta, N. Mammone, and F.C. Morabito. “Wavelet-ICA methodology for
efficient artifact removal from Electroencephalographic recordings”. In Proc International
Joint Conference on Neural Networks, 2007, pp. 1524-1529.
[12] N. Jacob, and A. Martin. “Image Denoising in the Wavelet Domain Using Wiener Filtering”.
Unpublished course project, University of Wisconsin, Madison, Wisconsin, USA, 2004.
[13] S. Julier, and J.K. Uhlmann. “Unscented Filtering and Nonlinear Estimation”. In Proc IEEE
92(3), 2004, pp. 401-421.
[14] S. Julier and J.K. Uhlmann. “A New Extension of the Kalman Filter to Nonlinear Systems”.
In Proceedings of AeroSense: In 11th International Symposium on Aerospace/Defense
Sensing, Simulation and Controls, 1997, pp. 182-193.
[15] A. Kallapur, S. Anavatti, and M Garratt. “Extended and Unscented Kalman Filters for
Attitude Estimation of an Unmanned Aerial Vehicle”. In Proc 27
th
IASTED International
Conference on Modelling, Identification, and Control (MIC 2008), 2008.
[16] Kaur, L., Gupta, S., & Chauhan, R.C., “Image Denoising using Wavelet Thresholding”. In
Proc 3
rd
Indian Conference on Computer Vision, Graphics & Image Processing (ICVGIP
2002), 22(14), 2002.
[17] V. Krishnaveni, S. Jayaraman, A. Gunasekaran and K. Ramadoss. “Automatic Removal of
Ocular Artifacts using JADE Algorithm and Neural Network”. International Journal of
Intelligent Systems and Technologies 1(4), pp. 322-333, 2006.
[18] V. Krishnaveni, S. Jayaraman, S. Aravind, V. Hariharasudhan, and K. Ramadoss.
“Automatic Identification and Removal of Ocular Artifacts from EEG using Wavelet
Transform”. Measurement Science Review 6(2, 4), pp. 45-57, 2006
[19] S. Makeig, J. Anthony, A.J. Bell, T. Jung, and T.J. Sejnowski. “Independent Component
Analysis of Electroencephalographic data”. Advances in Neural Information Processing
Systems 8, MIT Press Cambridge MA 8, pp. 145-151, 1996.
[20] N. Nikolaev and A. Gotchev. “ECG Signal Denoising using Wavelet Domain Wiener
Filtering”. In Proc 10
th
European Signal Processing Conference (EUSIPCO 2000), 2000.
[21] R. Sameni and M.B. Jutten. “Filtering Electrocardiogram Signals using the Extended
Kalman Filter”. In Proc 27th
IEEE Engineering in Medicine and Biology (EMBS) Annual
Conference 2005, pp. 5639-5642.
[22] P. Senthil Kumar, R. Arumuganathan, K. Sivakumar, and C. Vimal. “A Wavelet based
Statistical Method for De-noising of Ocular Artifacts in EEG Signals”. International Journal
of Computer Science and Network Security (IJCSNS). 8(9), pp. 87-92, 2008.

[23] P. Senthil Kumar, R. Arumuganathan, K. Sivakumar, and C. Vimal. “Removal of Ocular
Artifacts in the EEG through Wavelet Transform without using an EOG Reference Channel”.
International Journal of Open Problems in Computer Science & Mathematics (IJOPCM)
1(3), pp. 189-198, 2008.
[24] P. Shui and Y. Zhao. “Image Denoising Algorithm using Doubling Local Wiener Filtering
with Block Adaptive Windows in Wavelet Domain”. Signal Processing 87(7), pp. 1721-1734,
2007.
[25] Z. Weidong and L. Yingyuan.. 2001. “EEG Multi-resolution Analysis using Wavelet
Transform”, In Proc 23rd
Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (IEEE/EMBS) 2001.
[26] G. Welch and G. Bishop. “An Introduction to the Kalman Filter”, Technical Report TR 95-
041, University of North Carolina, Chapel Hill, North Carolina, USA. Internet:
http://www.cs.unc.edu/~welch/media/pdf/kalman_intro.pdf 2006, [Aug 29, 2009]
[27] J. Zhao, Q. Xi, X. Wang, and Y. Yang. “Image Denoising via Wavelet Domain Wiener filter
based on Coorelation Model”. In 16th
World Conference on Nondestructive Testing
(WCNDT 04), 2004.
[28] W. Zhou and J. Gotman. “Removal of EMG and ECG Artifacts from EEG Based on Wavelet
Transform and ICA”. In Proc 26
th
Annual International Conference on the IEEE EMBS,
2004, pp. 392-395.
[29] W. Zhou and J. Gotman. “Removing Eye-movement Artifacts from the EEG during the
Intracarotid Amobarbital Procedure”. Epilepsia 46(3), pp. 409-411, 2005.
[30] G. Zouridakis and D. Iyer. “Comparison between ICA and Wavelet-based Denoising of
single-trial evoked potentials”. In Proc 26th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, 2004, pp. 87-90.

Performance Comparison of Known ICA Algorithms to a Wavelet-ICA Merger

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Performance Comparison of Known ICA Algorithms to a Wavelet-ICA Merger (20)

Recently uploaded (20)

Performance Comparison of Known ICA Algorithms to a Wavelet-ICA Merger