chapter 1 introduction - information and library...
TRANSCRIPT
1
CHAPTER 1
INTRODUCTION
1.1 Introduction
Signal and system are the two major components in signal processing. A signal is a physical
quantity having the characteristics of varying w.r.t. time and space and the system is a process
whose input and output is a signal. The signal could be of any type. This chapter gives the brief
introduction about the biomedical signals and the various transformation techniques used for de-
noising of the non-stationary signals. The chapter also gives an idea about biomedical signals
and various methods used for the analysis of these signals. A wide study of literature has been
presented followed by the outcome from the literature and objectives of the thesis.
1.2. Biomedical Signals and Analysis
Biomedical signal generally represents a collective electrical signal attained from any organ,
signifying a physical variable of interest. This signal can be expressed with respect to its
amplitude, frequency and phase as well as it is on the whole a function of time. In common, the
observations gained from the physiological activities such as gene and protein sequences, neural
and cardiac rhythms, tissue and organ images of organisms are said to be biomedical signals.
Depending upon their source, application or signal characteristics, the biomedical signals are
classified. They can be either continuous or discrete. A number of signal sources may result into
a biomedical signal. Those sources are bioelectric Signals, bioimpedance signals, bioacoustic
signals, biomagnetic signals, biochemical signals and bio-optical signals.
Biomedical signal covers a wide range of signals including Electro-Oculogram (EOG) signal,
Electroneurogram (ENG) signal, Electrogastrogram (EGG) signal, Phonocardiogram (PCG)
signal, Carotid Pulse (CP) signal, Vibromyogram (VMG) signal, Vibroarthogram (VAG) signal,
Electrocardiogram (ECG), Electroencephalogram (EEG) and Electromyography (EMG) signal.
More precisely, the significant and widely applied biomedical signals are Electrocardiogram
(ECG), Electroencephalogram (EEG) and Electromyography (EMG).
Once the biomedical signals are recorded, they need to be analysed. The major operations of
signal processing include:
2
1) Signal acquisition and reconstruction,
2) Quality improvement including filtering, smoothing and digitization,
3) Feature extraction,
4) Signal compression,
5) Prediction.
In other words, the analysis includes, information gathering i.e. inferring a system by phenomena
measurement, diagnosis of malfunction or deformity and monitoring the system for continuous
or periodic information. Therapy and control which is modifying the system behaviour with
respect to the result of the above listed activities guarantees a definite result and finally the
evaluation which is to make it able to meet functional requirements, perform quality control, or
qualify the treatment effectiveness.
Biomedical signal processing aid the biologists to discover new biology and doctors in
monitoring diverse diseases. However the major problem faced by the entire signal processing
applications is noise. Noise is an unwanted signal superimposed over a pure signal [1]. A noise
can be differentiated according to its time and frequency domain properties. Types of noises are
white noise, uniform noise, and Gaussian noise. White noise is mainly hard to distinguish and to
eliminate because it is located in all frequencies. Uniform noise has a constant probability
density over a finite interval whereas Gaussian noise is defined over an infinite interval by just
two factors, average and spread. Additive white Gaussian noise is a special type of white and
Gaussian noises, which is a ubiquitous model in the context of statistical image restoration.
Fractional Gaussian noise (fGn) is the simplification of white noise. As an effect of these noises,
the information in a noisy signal will be misunderstood. Thus, in almost all signal processing and
communications applications, it is a significant task that the noise is eliminated totally from such
signals. This task is referred to as denoising.
There are various denoising techniques such as Fourier Transform, Time-Frequency analysis[2],
Wavelet Transform, Neural Network, Independent Component Analysis (ICA), Unscented
Kalman Filter (UKF) ), Empirical Mode Decomposition (EMD) [3], Canonical Correlation
Analysis (CCA) [30], Principal Component Analysis (PCA), Adaptive Impulse Correlated Filter
(AICF) [4], Time Sequence Adaptive Filter (TSAF), Signal-Input Adaptive Filter (SIF) [5] [6],
Adaptive Filters, Wiener filter, Singular Value Decomposition (SVD), FIR or IIR digital filters.
3
Wavelets commonly used for denoising biomedical signals include the Daubechies ‘db2’, ‘db8’
and ‘db6’ wavelets and orthogonal Meyer wavelet. Even though there are several techniques for
signal denoising, the common ones are the Fourier and Wavelet Transforms. Biomedical data
processing becomes more and more vital these days, because of its property of relevance and
support to the specialists for making decisions in their respective domains. The processing
techniques of biomedical signals generally employ standard algorithms for the purpose of
denoising.
Denoising using Wavelet Transform has different approaches; among them the mostly adopted
method is the one where the signals are decomposed into wavelets followed by thresholding and
shrinkage application for noise removal. The work focuses on Wavelet based denoising for the
biomedical signals ECG, EEG and EMG. Each signal has its unique feature and is adopted in a
wide range of applications. These biomedical signals are studied one by one in detail.
Denoising an EEG signal is a tricky preprocessing step prior to qualitative or quantitative
analysis. A blind source separation (BSS) problem is defined as disturbance minimization due to
muscular activity in EEG signals which consists of estimating the original sources underlying the
multichannel, without a prior knowledge about the sources and mixing process.
The noise reduction in electrocardiography signals is one of the important problems, which
appear during the analysis of this data. ECG signal is a non-stationary biological signal in nature
and plays a big role in diagnostics of human diseases [8]. One of the most serious problems in
the registration of electrocardiographic (ECG) signals is the parasite interference of muscle
active potentials – electromyography (EMG) signals. Because of EMG wide spectrum, it is
considered as white noise.Therefore the electrocardiography signals need an effective denoising.
Generally, adequate ECG denoising algorithms and procedures should have the following
properties :a) Improve signal-to-noise ratio (SNR), b) Preserve the original shape of the signal
and especially the sharp Q, R, and S peaks, without distorting the P and T waves and the smooth
transition of the ST-T segment. The noise presence problem is partially avoided by low-pass
(LP) filtering, FIR or IIR digital filters. Ensemble Averaging (EA) for the extraction of small
cardiac components from the noise contaminated ECG. As an improvement over EA, classical
Adaptive Filter (AF) with varying impulse response of the signal for the noise cancellation of
ECGs containing baseline wander, power line interference, EMG noise, and motion artifacts.
4
Wavelet analysis is often very effective because it provides a simple approach for dealing with
the local aspects of a signal. Electromyography (EMG) signals can be used for
clinical/biomedical applications, Evolvable Hardware Chip (EHW) development, and modern
human computer interaction. EMG signals acquired from muscles require advanced methods for
detection, decomposition, processing, and classification.
EMG signal denoising is mainly based on wavelet transform since it is the most useful tool to
remove noises in myoelectric recognition system. Wavelet denoising in analysis of EMG signal
is studied in the last decade, particularly in the engineering application such as the control of
upper-limb and lower-limb prostheses. The WT decomposes a signal into several multi-
resolution components according to a basic function which is called a wavelet function. As
discussed before, filters are one of the most widely used signal processing functions. The
resolution of the signal, which is a measure of the amount of detail information in the signal, is
determined by the filtering operations, and the scale is determined by up sampling and down
sampling operations. Wavelet function, level of wavelet decomposition, estimation function of
threshold value, and transformation (shrinkage) function of threshold value with wavelet
coefficients are the four important things that are considered for achieving the optimal wavelet
denoising algorithm for EEG signal.
1.3. Wavelet Transforms
The earlier method of ECG signal analysis was based on time domain method. But this is not
always sufficient to study all the features of ECG signals. So, the frequency representation of a
signal is required. To accomplish this, FFT (Fast Fourier Transform) technique is applied. But
the unavoidable limitation of this FFT is that the technique failed to provide the information
regarding the exact location of frequency components in time [11] [13] [14]. A method for ECG
denoising based on Wavelet Shrinkage approach using Time-Frequency Dependent Threshold
(TFDT) has been proposed in[93]. Generally speaking, the TFDT is high for the non-informative
wavelet coefficients, and low for the informative coefficients representing the important signal
features [69]. Donoho and Johnstone proposed Wavelet thresholding de-noising method based on
discrete wavelet transform (DWT) is suitable for non-stationary signals [12].
5
The Fourier transform is less useful in analyzing non-stationary signal (a non-stationary signal is
a signal where there is change in the properties of the signal) i.e., there is no repetition within the
region sampled. Fourier transform is only localized in frequency domain. The main drawback of
Fourier transform is that we lose our time information which is very important [10]. Fourier
transform cannot provide any information about the spectrum changes with respect to time.
Fourier transform assumes the signal to be stationary, but speech signal is always non-stationary.
To overcome this deficiency, a modified method-Short Time Fourier transform allows
representing the signal in both time and frequency domain through time windowing functions.
The usage of windowing with the Fourier Transform is called the Short Time Fourier Transform,
(STFT). The problem with this is that adequate understanding of the contents of the signal is
required to make appropriate windowing. This is hardly ever the case and many times, these
assumptions lead to problems. Moreover the STFT is time and frequency localized; there are
issues with the frequency time resolution [10]. Although the Fourier transforms tells how much
its frequency exists in a signal, it does not tell when in time these frequency components occur.
This information is required when the signal is non-stationary. All the real world signals are not
stationary, since their frequency changes in time. So what happen to a non-stationary signal
when it is processed, to view it in a frequency domain? All these are the major issues in signal
processing using Fourier transform techniques. Because of all the above said limitations, Fourier
Transform is not applied for denoising signals in the thesis. Moreover all these drawbacks are
overcome by Wavelet Transform technique and thus make wavelet a logical choice of denoising
technique in the thesis.
All types of signal transmission are based on transmission of a series of numbers. For signal
transmission or signal storage, the first step is to convert the given information to a series of
numbers. To do this the coefficients of the series need to be stored and only the coefficients are
sent. The wavelet tool is the new tool to perform this function. In the wavelet transform one
does not lose the time information, which is useful in many contexts [17]. In wavelet transform,
a signal is analyzed and expressed as a linear combination of the sum of the product of the
wavelet coefficients and mother wavelet. Wavelets are localized in both the frequency and the
time. This makes it possible to better localize properties of the analyzed signal. The result is a
well known ability of the wavelet transforms to pack the main signal information into a very
6
small number of large wavelet coefficients. A wavelet is simply a small wave which has energy
concentrated in time to give a tool for the analysis of transient, non stationary or time-varying
phenomena. A Wavelet transform uses a set of transform basis function called wavelets to
decompose a signal.Proper selection of wavelet basis function plays a vital role in denoising
[18].
Wavelets are a powerful statistical tool which can be used for a wide range of applications,
namely signal processing, data compression, smoothing and de-noising, fingerprint verification,
biology for cell membrane recognition, blood-pressure, heart-rate and ECG analysis etc.
There are many types of wavelets in the family. The Daubechies wavelet is described by a
maximal number of vanishing moments for some given support. Haar wavelet is an order of
rescaled square shaped function. Symlet wavelets are an improved version of Daubechies
wavelets with increased symmetry. Coiflets wavelets have scaling functions with vanishing
moments. The biorthogonal family wavelets are signed as bior. In Legendre wavelet, the
Legendre function have common applications in which spherical coordinate system are suitable.
Among the various wavelet families are defined in the literature, Daubechies wavelets are the
most popular wavelets. The Daubechies wavelets are used in different applications. The Haar,
Daubechies, Symlets and Coiflets are compactly supported orthogonal wavelets. These wavelets
along with Meyer wavelets are capable of perfect reconstruction. The Meyer, Morlet and
Mexican Hat wavelets are symmetric in shape. The wavelets filters are selected based on their
ability to analyze the signal and their shape in an application [20] [23] [27] [29].
The signals are reconstructed using Inverse DWT. The IDWT is applied as a reverse process to
the decomposed signal and the original signal is reconstructed. The IDWT is obtained by the
quadrature filter bank [32]. On the other hand a classifier achieves its objective by making a
classification decision based on some characteristics. The application of Artificial Neural
Network enables the selection of the best wavelet type for each type of biomedical signal [33].
The denoised signal is compressed by a hybrid wavelet Shannon-Fano coding for reducing its
storage size [41].
Now a brief introduction about signal processing, biomedical signals and the nature of various
signal denoising techniques has been given. In this section, several studies proposed especially
for denoising are to be discussed. Moreover, the nature and applications of wavelet techniques
7
and different classifiers, which were exploited in a wide area of research other than denoising are
to be presented. A detailed survey about the various techniques designed for denoising and the
methods proposed for signal compression in the literature referred earlier is given in the next
section.
1.4. State of Art/ Literature Survey:
In this section, some of the studies based on signal processing have been presented. It is stated in
the previous sections that denoising is applied extensively for biomedical signals, images, and
for audio, video signals. So, a detailed review of literature has been done in which various
methods for denoising, compression, reconstruction and classification have been studied.
Stephane G. Mallat (1989) studied the properties of the operator which approximates a signal at a
given resolution. He showed that the difference of information between the approximation of a
signal at the resolutions 2’+’
and 2j can be extracted by decomposing this signal on a wavelet
orthonormal basis of L2(R
n). This decomposition defines an orthogonal multiresolution
representation called a wavelet representation. [3]
Metin Akay (1995) compared various methods for biomedical signal processing only to conclude
that wavelet transform is the best among all these methods. He gave the advantages and
disadvantages of all the techniques used in the signal processing and proposed wavelet transform
for biomedical signals as they are non-stationary in nature. [11]
Michael Unser (1996) emphasized on the statistical properties of the wavelet transform and
discusses some recent examples of application in medicine and biology. The redundant forms of
the transform (CWT and wavelet frames) are well suited for detection tasks (e.g., spikes in EEG,
or micro-calcifications in mammograms). The CWT, in particular, can be interpreted as a pre-
whitening multi-scale matched filter. Redundant wavelet decompositions are also very useful for
the characterization of singularities as well as the time-frequency analysis of non-stationary
signals. Some examples of applications in Phonocardiography, ECG and EEG are also discussed
in the paper. [14]
Nikolay Nikolaev and Atanas Gotchev (1998) investigated a method for ECG denoising based
on wavelet shrinkage approach. They proposed a shrinkage threshold which was high for the
non-informative wavelet coefficients and low for the informative coefficients. [23]
8
Arthur Petrosian et al., (2000) applied recurrent neural networks (RNN) combined with signal
wavelet decomposition to the problem of predicting the onset of epileptic seizure which is an
important and difficult biomedical problem. [33].
Tom Chau (2001) presented a review paper in which, part 1 explored applications of fuzzy,
multivariate statistical and fractal methods to gait data analysis and part 2 extended this critical
review to the applications of ANN and wavelets to gait data analysis. The review concluded with
a practical guide to the selection of alternative gait data analysis methods. [39]
Claude Robert et al. (2002) presented more than 100 current neural network applications
dedicated to EEG processing. They demonstrated the importance of neural network in medicine
and biology involving EEG signal processing. Positive results obtained in most applications had
shown relevance for processing electroencephalograms. Works were categorized according to
their objective. [45]
Thierry Blu and Michael Unser (2002) showed that the wavelets and radial basis functions are
the two types of representation which were closely linked together through fractals. They
identified and characterized the whole class of self-similar radial basis functions that could be
localized to yield conventional multi-resolution wavelet bases. They also proved that for any
compactly supported scaling function, there existed a one-sided central basis function, which
spanned the same multi-resolution subspaces. [80]
Alka Yadav et al. (2003) proposed a new approach to filter the ECG signal from noise using
Multi resolution Technique based on Wavelet Transform . They proposed decomposition method
using the Stationary Wavelet Transform and by selecting db wavelet, the noisy signal had been
decomposed, in the 4th decomposition level. As a result approximate coefficients aj and detail
coefficients dj were obtained. They showed that their method gave better results than the other
technique applied in this field. [172]
Andrew P. Bradley (2003) reviewed a number of approaches to reduce, or remove, the problem
of shift variance in the discrete wavelet transform (DWT). He described a generalization of the
critically sampled DWT and the fully sampled algorithm ‘A TROUS’ that provided approximate
shift-invariance with an acceptable level of redundancy. [49]
9
Minos Garofalakis and Amit Kumar (2004) proposed a novel, computationally efficient schemes
for deterministic maximum-error wavelet thresholding in one and multiple dimensions. For one-
dimensional wavelets, they introduced an optimal, low polynomial-time thresholding algorithm
based on a new Dynamic Programming formulation that can be used to minimize either the
maximum relative error or the maximum absolute error in the data approximation [52]
Pawel Costka and Ewaryst Tkacz (2004) presented Wavelet-neural systems (WNS) which
inherited the properties of neural networks, belong to the class of universal approximators of
unknown functions. Classifier structures described in their work fulfilled the role of
approximators of functions, which were able to assign the input signal to a particular class with a
given accuracy. [54]
S.A. Chouakri et al. (2005) presented an algorithm of filtering the noisy real ECG signal in this
paper. The classical wavelet denoising process, based on the Donoho et al. algorithm, at the 4th
level, appears clearly the P and T waves whereas the R waves undergo considerable distortion.
This is due to the interference of the WGN and the free noise ECG detail sequences at level 4. To
overcome this drawback, the key idea is to estimate the corrupted WGN and consequently
remove the noise interfering R waves at the 4th level detail sequence. [65]
Mohammad Pooyan et al. (2005) presented a novel approach for wavelet compression of
electrocardiogram (ECG) signals based on the set partitioning in hierarchical trees (SPIHT)
coding algorithm. SPIHT algorithm had achieved prominent success in image compression. They
used a modified version of SPIHT for one dimensional signal. [64]
Vangelis P. Oikonomou and Dimitrios I. Fotiadis (2006) proposed the Bayesian approach for
biomedical signal denoising. This approach mainly deals with the biomedical signals affected by
white Gaussian noise. To obtain a meaningful solution, they introduced many constraints in the
problem. They selected the desired signal to belong to the class of smooth signals. The
introduction of constraints led them to a bayesian formalism of the problem.
Reza Sameni et.al (2006) proposed a nonlinear Bayesian filtering framework for the filtering of
single channel noisy ECG recordings. Within this framework several suboptimal filtering
schemes were developed. The necessary dynamic models of the ECG were based on a modified
nonlinear dynamic model, previously suggested for the generation of a highly realistic synthetic
10
ECG. A modified version of this model was used in several Bayesian filters, including the
Extended Kalman Filter, Extended Kalman Smoother, and Unscented Kalman Filter. An
automatic parameter selection method was also introduced, to facilitate the adaptation of the
model parameters to a vast variety of ECGs. [78]
Jong Yong A. Foo (2006). In this paper it is shown that Photoplethysmography (PPG) can be
used in time-related measurements such as heart rate (HR) and pulse transit time (PTT)
estimations in the medical fields. This paper compares the capabilities of two signal processing
techniques; digital adaptive filtering and discrete wavelet transformation, in restoring artifact-
induced PPG signals during two regulated mild movements. [68]
P. Kukharchik et al. (2007) presented an initial study of feature extraction based on wavelets and
pseudo wavelets in the task of vocal pathology diagnostic. A new type of feature vector, based
on continuous wavelet and wavelet-like transform of input audio data was proposed. Support
vector machine had been used as a classifier for testing the feature extraction procedure. The
proposed scheme revealed, that features based on continuous wavelet transform and continuous
pseudo wavelet transform had potential for future usage for vocal fold pathology detection when
working with realistic records, that were not previously processed or prepared. [86]
Omid Sayadi and Mohammad B. Shamsollahi (2007) presented a new modified wavelet
transform, called the multiadaptive bionic wavelet transform (MABWT), that could be applied to
ECG signals in order to remove noise from them under a wide range of variations for noise. This
proposed thresholding rule had been applied and worked successfully in denoising the ECG. [76]
Orlando Jos´e Ar´evalo Acosta and Matilde Santos Pe˜nas (2007). This contribution consists of
the application of a hybrid technique of signals digital processing and artificial intelligence, to
classify two kinds of biomedical spectra, normal brain and meningioma tumor. Each signal is
processed to extract the relevant information within the range of interest. Then, a Haar 4 wavelet
transform is applied to reduce the size of the spectrum without losing its main features. This
signal approximation is coded in a binary set which keeps the frequencies that could have
representative amplitude peaks of each signal. [81]
Laurent Brechet et.al (2007) proposed a novel scheme for signal compression based on the
discrete wavelet packet transform (DWPT) decomposition. The mother wavelet and the basis of
11
wavelet packets were optimized and the wavelet coefficients were encoded with a modified
version of the embedded zero tree algorithm. This signal dependant compression scheme was
designed by a two-step process. The first (internal optimization) was the best basis selection that
was performed for a given mother wavelet. The second (external optimization) was the selection
of the mother wavelet based on the minimal distortion of the decoded signal given a fixed
compression ratio. [82]
Ilker Bayram, Ivan W. Selesnick (2007) The author discusses the 2-band discrete wavelet
transform (DWT) which provides an octave-band analysis in the frequency domain, but might
not be ‘optimal’ for a given signal. In this paper, a method to implement a dual-tree complex
wavelet packet transform (DTCWPT) has been discussed. To find the best complex wavelet
packet frame for a given signal, the author adapts the basis selection algorithm by Coifman and
Wickerhauser, providing a solution to the basis selection problem for the DT-CWPT. [83]
Alain de Cheveigne and Jonathan Z. Simon (2008) introduced a denoising method based on
spatial filtering for removing unwanted components of biological origin from neurophysiologic
recordings such as Magnetoencephalography (MEG), electroencephalography (EEG), or
multichannel electrophysiological or optical recordings. A spatial filter was designed to partition
recorded activity into Stimulus-related and stimulus-unrelated components, based on a criterion
of stimulus-evoked reproducibility. Components that are not reproducible were projected out to
obtain clean data. [96]
Yannis Kopsinis and Stephen (Steve) McLaughlin (2008), used Empirical Mode Decomposition
(EMD) based denoising techniques of the major wavelet thresholding principle in the
decomposition modes resulting from applying EMD to a signal. They showed that although a
direct application of this principle in the EMD case was not feasible, it could appropriately
adapted by exploiting the special characteristics of the EMD decomposition modes. [94]
Leandro Aureliano da Silva et al. (2008) proposed the technique used for noise reduction during
the reconstruction of speech signals, particularly for biomedical applications. They implemented
and compared two algorithms for speech denoising: the Kalman’s filter in the time domain
(FKT) and in the frequency domain (FKF). Comparison with discrete Kalman filter in the
frequency domain showed better performance of the proposed technique. [98]
12
Guoshen Yu et al. (2008) introduced a block thresholding estimation procedure which adjusted
all the parameters adaptively to signal property by minimizing a Stein estimation of the risk
calculated from the data. The resulting algorithm was robust to variations of signal structures
such as short transients and long harmonics. [95]
Omid Sayadi and Mohammad Bagher Shamsollahi (2008) proposed a study in which they
presented efficient denoising and lossy compression schemes for electrocardiogram (ECG)
signals based on a modified extended Kalman filter (EKF) structure. The new EKF structure was
used not only for denoising, but also for compression, since it provided estimation for each of the
new 15 model parameters. [97]
Hariharan Nalatore et al. (2009) attempted to apply a state-space smoothing method, based on
the combined use of the Kalman filter theory and the Expectation–Maximization algorithm, to
denoise two datasets of local field potentials recorded from monkeys. Their main goal was to
establish that the denoising procedure based on a simple data model works on actual neural data.
After denoising, the discrepancy between the two subjects was significantly reduced. [107]
Slavy G. Mihov et al. (2009) investigated the use of wavelet transform for denoising speech
signals contaminated with common noises. They showed the basic principles of wavelet
transform as an alternative to the Fourier transform. The practical results obtained were based on
processing a large dedicated database of reference speech signals contaminated with various
noises in several SNRs. [112]
G. Umamaheswara Reddy et al. (2009) proposed a new thresholding technique for denoising of
ECG signal. This new denoising method was called as improved thresholding denoising method
and could be regarded as a compromising between hard- and soft thresholding denoising
methods. The proposed method selected the best suitable wavelet function based on DWT at the
decomposition level of 5, using mean square error (MSE) and output SNR. The advantage of the
improved thresholding denoising method was that it retained both the geometrical characteristics
of the original ECG signal and variations in the amplitudes of various ECG waveforms
effectively. [111]
Sana Ktata et al. (2009) proposed a wavelet ECG data codec, based on the Set Partitioning in
Hierarchical Trees (SPIHT) compression algorithm. The SPIHT algorithm had achieved notable
13
success in still image coding. They modified the algorithm for the one-dimensional (1-D) case
and applied it to compression of ECG data. By this compression method, small percent root
mean square difference (PRD) and high compression ratio with low implementation complexity
were achieved. [114]
M. Murugesan and R. Sukanesh (2009) presented an effective system for classification of
electroencephalogram (EEG) signals that contain credible cases of brain tumor. The
classification technique support vector machine (SVM) was utilized in the proposed system for
detecting brain tumors. Initially, the artifacts present in the EEG signal were removed using
adaptive filtering. Then the spectral analysis method was applied for extracting generic features
embedded in an EEG signal. The key advent of the proposed approach was that it enabled early
detection of brain tumors initiating quicker clinical responses. [109]
I. Omerhodzic et al. (2009) discussed a wavelet-based neural network (WNN) classifier which
was implemented and tested for recognizing three sets of EEG signals. First, the DWT with the
MRA was applied to decompose EEG signal at resolution levels of the components of the EEG
signal and the Parseval’s theorem were employed to extract the percentage distribution of energy
features of the EEG signal at different resolution levels. Second, the neural network (NN)
classified those extracted features to identify the EEGs type according to the percentage
distribution of energy features. The results showed that the proposed classifier had the ability of
recognizing and classifying EEG signals efficiently. [110]
Mojtaba Bandarabadi et al. (2010) proposed a method for electrocardiogram signal (ECG)
denoising. The basis of this method was achieved through filtering Singular Values (SV) of the
signal. It was based on enhancing and optimizing the SV for omitting noise from ECG. The
advantage of this proposed method was its capability in enhancing the signal very well. Using
different quantitative and qualitative parameters, the efficacy of the method was evaluated. [131]
Wasim Ahmad et al. (2010) proposed a shift-invariant analysis scheme which was non
redundant. This scheme combined minimum-phase (MP) reconstruction with the DWT so that
the resultant scheme provided a shift-invariant transform. The detailed properties of MP signal
and different methods to reconstruct it were explained. The proposed scheme could be used for
the analysis-synthesis, classification, and compression of transient sound signals. [127]
14
A.N. Akansu et.al (2010) highlighted inherently built-in approximation errors of discrete-time
signal processing techniques employing WT framework. Then, they presented an overview of
emerging analog signal processing applications of wavelet transform along with its still active
research topics in more matured discrete-time processing applications. It was shown that analog
wavelet transform was successfully implemented in biomedical signal processing for design of
low-power pacemakers and also in ultra-wideband (UWB) wireless communications. [123]
Dimitri Van De Ville et al. (2010) introduced a family of elementary singularities that were
point-Holder-regular. These singularities were self-similar and were the Green functions of
fractional derivative operators; i.e., by suitable fractional differentiation, one retrieved a Dirac
function at the exact location of the singularity. They showed that the wavelet coefficients of the
(non-redundant) decomposition could be fitted in a multiscale fashion to retrieve the parameters
of the underlying singularity. They proposed an algorithm based on stepwise parametric fitting
and the feasibility of the approach to recover singular signal representations. [129]
Abdel Rahman et al. (2010), proposed a new approach to filter the ECG signal from noise using
Wavelet Transform. Different ECG signals were used and the method had been evaluated using
MATLAB software. The aim of the paper was to adapt the discrete wavelet transform (DWT) to
enhance the ECG signal. The presented method showed good results when compared to
conventional methods particularly in ECG signal case. This method had better performance than
Donoho’s discrete wavelet thresholding coefficients and FIR filter. [121]
A. Phinyomark et al. (2010) proposed the selection of the modified wavelet shrinkage functions
which was based on the improving of recognition in EMG signal from upper-limb motions.
Seven kinds of modified shrinkage functions and two traditional shrinkage functions were
compared by calculating classification accuracy of denoised EMG signal estimated from the
shrinkage functions. In addition, the denoising effects of the different shrinkage functions for
different levels of noise were proposed. The experimental results showed that the highest
classification accuracy can be reached using the firm wavelet shrinkage function. [122]
Minfen Shen et al. (2010) proposed a new local spatio-temporal prediction method based on
support vector machines (SVMs). Combining with the local prediction method, the sequential
minimal optimization (SMO) training algorithm, and the wavelet kernel function, a local SMO-
15
wavelet SVM (WSVM) prediction model was developed to enhance the efficiency,
effectiveness, and universal approximation capability of the prediction model. This method
greatly increased the computational speed and more effectively captured the local information of
the signal. [126]
Andreas Spanias et al. (2010), considered features extracted from the Fourier, wavelet and Walsh
power spectra of the ion-channel signals. They compared the performance of all the three sets of
features using support vector machines. They performed classification of signals from simulated
and real ion-channels and presented the results. Results obtained showed that the transform
domain features achieved high classification rates in addition to high sensitivity and specificity
rates. [74]
Corina Sararu et al. (2010) described a new classification methodology based on the use of
Independent Component Analysis and Wavelet decomposition (ICAW) techniques. An ensemble
system of classifiers was built such that each classifier independently decided the assignation of
the test examples on several representations resulted by taking projections computed by wavelets
and Independent Component Analysis (ICA). [119]
Rui Hou et al. (2011) proposed a quantitative evaluation model of denoising methods for surface
Plasmon resonance imaging signal. Their model allowed one to get the optimized denoising
method. The method can be used to suppress the noise in SPRI signals effectively. The wavelet
transform based denoising methods was used to process SPRI signals constructed from
theoretical simulated kinetic curves of bio molecular interactions. Application of the optimized
denoising method obtained from the model to SPRI signals helped to improve the resolution of
SPRI instrument. [142]
Jeremy Terrien et al. (2011) proposed an algorithm for the automatic selection of the modes
containing the signal of interest. This algorithm was based on statistical analysis describing the
noise repartition between IMFs. This algorithm used an estimate of the signal noise content from
the energy of the first IMF, which was supposed to contain a specific part of the total noise and
to contain noise only. They proposed to use mode mixing detection based on a stationary test
applied to the first IMF. [143]
16
Anil Chacko and Samit Ari (2011), proposed a denoising technique for ECG signals based on
Empirical Mode Decomposition (EMD). The noisy ECG signal was initially decomposed into a
set of Intrinsic Mode Functions (IMFs) using EMD method. In the proposed technique, the IMFs
which are dominated by noise were automatically determined using Spectral Flatness (SF)
measure and then filtered using butterworth filters to remove noise. [144]
Ibrahim Missaoui and Zied Lachiri (2011) addressed the problem of blind separation of speech
mixtures. They proposed a new blind speech separation system, which integrated a perceptual
filter bank and independent component analysis (ICA) and using kurtosis criterion. Their
proposed technique consisted on transforming the observations signals into an adequate
representation using UWPD and Kurtosis maximization criterion in a new preprocessing step in
order to increase the non Gaussianity which was a pre-requirement for ICA. [149]
B. Mohan Kumar and R. Vidhya Lavanya in 2011 used CL Multiwavelet with soft thresholding
by universal threshold selection rule for denoising the real time signals. The approach was
incorporated with time domain and frequency domain analysis. The objective of the work was to
enhance the noisy speech signal using CL Multiwavelet with soft threshold method in real time
mobile speech communication. The noise was estimated and thresholding was done according to
the estimated noise. Denoising was done for both stationary and non-stationary signals with
different noise levels. [147]
N. M. Sobahi (2011) applied wavelet transform for removing noise from the surface EMG and
provided a brief introduction of the wavelet transform in EMG signals processing. Wavelet
denoising method was expected to offer a powerful compliment to conventional filtering
techniques like notch filters and frequency domain filtering methods, which would have been
very efficient for EMG signal analysis. [133]
Om Prakash Yadav et al. (2011) presented Wavelet Based Encoder/Decoder for Compression of
ECG Signal and three compression algorithms. In EZW algorithm, 3-level decomposition was
performed to the original ECG samples, and the wavelet coefficients at different sub-band
representing the same spatial location in the ECG samples were loaded into a spanning tree.
[148]
17
Catalina Monica Fira et al. (2011). An electrocardiographic signal (ECG) compressed sensing
(CS) method, its reconstruction using specific dictionaries of cardiac pathologies and method
evaluation testing using classical measures as well as by classification error of the reconstructed
patterns based on the K-Nearest Neighbor classifier (KNN) were presented in the paper. For
compressed sensing, a random matrix with standard normal distribution had been used, followed
by a classification of compressed signals in one of eight possible pathological classes. [136]
Mingsheng Liu et al. (2011) presented a risk assessment method which combined wavelet neural
network (WNN) and entropy-grey correlation, created a WNN model and described a simulation
experiment by Matlab 7. In addition, comparisons were made in terms of convergent speed,
training precision and forecasting effect between WNN and other traditional estimation methods
such as BP-NN (Back Propagation Neural Network), FCM (Fuzzy Clustering Method) and SPR
(Statistical Pattern Recognition). [150]
Oumar Niang et.al (2012) presented a new signal denoising method based on the classical three
step procedure analysis-threshold- synthesis and the Spectral Intrinsic Decomposition (SID).
This method consists of an iterative thresholding of the SID components. The SID-based
removal method reduced noise and could retain useful discontinuities of the signal as effectively
as the wavelet techniques based on soft thresholding. [166]
Chinmay Chandrakar et al. (2012) considered adaptive filters to reduce the ECG signal noises
like PLI and Base Line Interference. Recursive Least Squares (RLS) algorithm was proposed for
removing artifacts preserving the low frequency components and tiny features of the ECG. [114]
L. N. Sharma et al. (2012) applied Multiscale Principal Component Analysis (MSPCA) for
quality controlled de-noising of Multichannel Electrocardiogram (MECG) signals. Collecting
wavelet coefficients of all ECG channels at a wavelet scale multivariate data matrices were
formed. Principal Component Analysis (PCA) was performed on these matrices for signal
denoising. [167]
Akanksha Mishra et al. in 2012 introduced a comparison of the reconstructed 10 ECG signals
based on different wavelet families, by evaluating the performance measures as MSE (Mean
Square Error), PSNR (Peak Signal To Noise Ratio), PRD (Percentage Root Mean Square
Difference) and CoC (Correlation Coefficient). L1 minimization was used as the recovery
18
algorithm. The reconstruction results were comprehensively analyzed for three compression
ratios. [171]
Hossein Rabbani and Saeed Gazor (2012) investigated the local probability density function
(pdf) of natural signals in sparse domains. The statistical properties of natural signals were
characterized more accurately in the sparse domains. Their experiments on 3D data in 3D
discrete complex wavelet transform (DCWT) domain showed that a conditionally (given locally
estimated variance and shape) independent Bessel K-form distribution (BKFD) locally fitted the
sparse domain’s coefficients of natural signals, accurately. [159]
Zoltan Germans Saollo and Calin Ciufudean, (2012) dealt with the design of waveform-adapted
analyzing function in order to have a good wavelet decomposition of the analyzed signal. The
proposed procedure led to obtain discrete sequences as discrete wavelet function to perform
denoising, those met certain mathematical criteria. Discrete Wavelet Transform based denoising
was performed. That study introduced a waveform-adapted wavelet transform based noise
suppression procedures and presented the obtained results. [161]
Md. Ashfanoor Kabir and Celia Shahnaz, (2012) Comparison of ECG signal denoising
algorithms in EMD and wavelet domains presented a detail analysis on the Electrocardiogram
(ECG) denoising approaches based on noise reduction algorithms in Empirical Mode
Decomposition (EMD) and Discrete Wavelet Transform (DWT) domains. This study provided
the performance analyses of ECG signal denoising algorithms in EMD and wavelet domains thus
compared the effectiveness in reducing the noise. [163]
Hari Mohan Rai and Anurag Trivedi (2012) dealt with the noise removal of ECG signal using
three different wavelet families. The different noise structure (unscaled white noise, scaled white
noise and non white noise) had been selected for ECG signals and was compared their statistical
parameter to find out the best result. The wavelet families used for De-noising were Haar,
Daubechies and Symlets. They decomposed the ECG signal into 5 levels. The experiment
showed that the Daubechies4(Db4) of level 5 for scaled white noise structure gave the best result
as compared to other wavelet family and Haar wavelet gave the worst result for Unscaled white
noise structure. [174]
19
P. Karthikeyan (2012) They considered that the Discrete Wavelet Transform (DWT) based
wavelet denoising had incorporated using different thresholding techniques to remove three
major sources of noises from the acquired ECG signals namely, power line interference, baseline
wandering, and high frequency noises. Three wavelet functions ("db4", "coif5" and "sym7") and
four different thresholding methods were used to denoise the noise in ECG signals. [170]
Maedeh Kiani Sarkaleh and Asadollah Shahbahrami (2012), proposed an expert system for ECG
arrhythmia classification. Discrete wavelet transform was used for processing ECG recordings,
and extracting some features, and the Multi-Layer Perceptron (MLP) neural network performed
the classification task. Two types of arrhythmias could be detected by the proposed system.
Some recordings of the MIT-BIH arrhythmias database had been used for training and testing
our neural network based classifier. [162]
Prajakta S. Gokhale (2012) showed the effect of the wavelet thresholding on the quality
reconstruction of an ECG signal. By applying IIR notch filter directly to the non-stationary
signal like ECG gives ringing effect which can be eliminated through wavelet transform among
which Db4 performed better than other methods to de-noise the noisy ECG signal. [175]
Apoorv Gautam and Maninder kaur (2012) proposed algorithm which utilizes morphological
filtering and continuous wavelet transform with a dedicated wavelet. They showed that the multi-
resolution analysis based on the CWT can enhance small differences when the signal is
simultaneously observed at the most appropriate scales. [172]
L. Senhadji et al. (2013) introduced a model-based Bayesian denoising framework for
phonocardiogram (PCG) signals. The denoising framework was founded on a new dynamical
model for PCG, which was capable of generating realistic synthetic PCG signals. The extended
Kalman smoother (EKS) is the Bayesian filter that was used in their study. The results of the
EKS demonstrated better performance than WD over a wide range of PCG SNRs. [187]
Md. Mamun et al. (2013) employed discrete wavelet transform to remove noise from EEG
signal. Root mean square difference had been used to find the usefulness of the noise
elimination. Four different discrete wavelet functions had been used by them to remove noise
from the EEG signal gotten from two different types of patients (healthy and epileptic) to show
the effectiveness of DWT on EEG noise removal. [177]
20
E. Castillo et al. (2013) One-step wavelet-based processing for wandering and noise removing in
ECG signals technique illustrated the application of the Discrete Wavelet Transform (DWT) to
the processing of electrocardiogram (ECG) for wandering and noise suppression in this paper.
The proposed scheme allowed reducing the computational complexity, while its fixed-point
modeling showed the expected performance of possible future portable hardware
implementations. The system had been tested using synthetic ECG signals, which allowed to
accurately measuring the effect of the proposed processing. [182]
Maryam Ahmadi and Rodrigo Quian Quiroga (2013) presented an automatic denoising method
based on the wavelet transform to obtain single trial evoked potentials. The method was based on
the inter and intra-scale variability of the wavelet coefficients and their deviations from baseline
values. The performance of the method was tested with simulated event related potentials (ERPs)
and with real visual and auditory ERPs. The proposed method provided a simple, automatic and
fast tool that allowed the study of single trial responses and their correlations with behavior.
[180]
Sandeep Sharma et al. (2013) developed a technique to automatically detect and mark the basic
waveforms of ECG signal. Recently developed PCA had been used for this purpose. [181]
Amita A. Shinde and Pramod M. Kanjalkar (2013) presented an algorithm for wavelet based
ECG signal compression, where db7 was selected as the mother wavelet for analysis.
Thresholded wavelet coefficients were coded with RLC. One of the main advantages of this
method was lower calculation complexity in comparison with other methods. This algorithm was
tested for different records from MIT–BIH arrhythmia database. [185]
Vahid Majidnezhad and Igor Kheidorov (2013). In this paper, an ANN-Based method for vocal
fold pathology diagnosis was proposed so that in the proposed scheme, Mel-Frequency-Cepstral-
Coefficients along with the wavelet packet decomposition were used for feature extraction phase.
Also PCA method for the feature reduction phase was used. And finally the Artificial neural
network (ANN) was used for the classification phase. [186]
Mandavi1 et al. (2013). An efficient composite method had been developed for data compression
of ECG signals in this paper. After carrying out detailed studies of different data compression
algorithms, they used back propagation algorithm to analyze the artificial neural networks.
21
Twelve significant features were extracted from an echocardiogram (ECG). The features of
samples were used as input to the neural network. [184]
1.5. Outcome from Literature / Gaps
From the extensive literature survey, it is observed that different denoising techniques give
significantly varying performance for each of the biomedical signals, and the main problems
encountered are as follows.
In the previous studies, while denoising the biomedical signals such as ECG, EEG and
EMG, the trained system is unable to automatically detect the best wavelet suitable for
denoising.
The Fourier transform analysis is inadequate and is localized only in the frequency band.
The major drawback of Short Term Fourier Transform for signal denoising is that the
time frequency precision is not optimal.
Digital filters and adaptive methods can be applied only to signals whose statistical
characteristics are stationary in many cases and cannot be applied to non-stationary
signals because of loss of information. The best method that offers efficient denoising for
each of the biomedical signals such as ECG, EEG and EMG had not been predefined in
the literature.
Although, some of the authors exploited different wavelet for denoising different signals,
their trained system was unable to automatically detect the best wavelet suitable for
denoising. Even though much of the schemes discussed in literature survey provided
better performance in denoising, they compromised either space or time.
For compression of biomedical signals several studies have been addressed in the existing
literature. EZW, MEZW, SPIHT coding, RLC algorithm, Wavelet transform with SPHIT,
Wavelet and huffman, JPEG 2000, Shannon fano are the few compression techniques
addressed in the literature.
1.6. Problem Formulation / Objectives
From the literature it is obvious that there is a strong need of an efficient classifier for deciding
the optimal wavelet family for denoising different biomedical signals. Since each biomedical
22
signal has its unique nature and characteristics, the classifier must be trained in an efficient way
with the nature of each signal for accurate classification of optimal wavelet. Moreover we are in
need of an effective biomedical signal compression algorithm achieving best CR and PRD
values.
Based upon the above listed directions, the problem for this work is formulated as:
To find the optimal wavelet for biomedical signal.
To focus on reducing time and storage space, however at the same time signals must be
denoised efficiently with the optimal wavelet.
Once the optimal wavelet is found, biomedical signals will be denoised using this wavelet with
focus on reducing time and storage space. Biomedical signals considered will be ECG, EEG and
EMG.
1.7 Methodology Used to Achieve Objectives
The main motive of the thesis lies on denoising signals with the most suitable wavelet, which is
classified using Artificial Neural network. It results in less overhead time, since the classification
algorithm is already trained for almost all of the members of wavelet family. Moreover a hybrid
compression algorithm is applied after the denoising process and thus a less space is needed for
storing compressed information and is used for reconstruction of the original signal. Thus the
objectives of minimum space and less time are achieved in the thesis only by using simple
methods for denoising, classification and compression. The following steps will be carried out to
achieve the objectives:
Discrete Wavelet Transformation for the signals will be applied, which will be performed
by means of a low pass and a high pass filter, yielding approximation and detailed
coefficients respectively. This process continues for a fixed number of decomposition
levels.
The signal will be decomposed using the Shift Invariant method.
Then for each level of decomposition, the wavelet frequency thresholding for the detailed
and approximation coefficients will be applied resulting in denoised signal.
23
The artificial neural network will be trained initially with the properties of each wavelet
and the nature of the biomedical signals. By using the parameters the optimized wavelet
transform will be selected which is best suitable for denoising using the neural network
classifier.
As there is a strong need of an effective biomedical signal compression algorithm for
achieving the best CR and PRD values. For this purpose, both Shannon Fano algorithm
and wavelet transform will be combined. The signals obtained after classifications will
then be compressed using hybrid wavelet Shannon-Fano coding algorithm.
These steps can be represented in the form of Block Diagram shown in Figure 1.1 whereas
detailed representation by flow chart in Figure 1.2
Figure 1.1: Main steps of the proposed system
Decomposition
Denoising Signal
Compression
Input
Signal
ANN
24
Figure 1.2: Flow chart of the proposed system
Wavelet
reconstruction
TRAINING SYSTEM
Input Biomedical
Signals (ECG, EEG
and EMG)
Selection of the best
Wavelet using Artificial
Neural Network
Wavelet Frequency
Thresholding
Apply hybrid
wavelet Shannon
Fano compression
Denoised signal
Apply Discrete
Wavelet Transform
(DWT)
Train the features in
Artificial Neural
Network
TESTING SYSTEM
Input Biomedical Signal
(ECG, EEG or EMG)
25
1.8. Organization of the Thesis
The thesis is organized as follows.
Chapter 1 presents an introduction of the proposed system model followed by the introduction
of biomedical signals and wavelet transform. Chapter also deals with a detailed survey of the and
outcome from the literature studied. The problems of the existing studies have been presented
followed by the main objectives of this thesis. The methodology used to achieve the objectives
has also been discussed in the chapter.
Chapter 2 gives the detail description about the biomedical signals used for analysis in the
thesis. The origin and recording of the ECG, EEG and EMG signals has been studied in this
chapter. The Wavelet transformation scheme will also be explained with detailed mathematical
representation followed by their types. Types of wavelet transform such as Continuous Wavelet
Transforms (CWT) and the Discrete Wavelet Transforms (DWT) will be given. Followed by
DWT, wavelet filters will conclude the second chapter.
Chapter 3 deals with the various decomposition and denoising techniques in mathematical and
diagrammatic representations. The chapter explains the different thresholding techniques for
denoising. Wavelet frequency thresholding based denoising is used in the thesis and its details
will be presented in this chapter. The chapter concludes with the graphical results of denoising
for the three biomedical signals.
Chapter 4 Signal Reconstruction and compression techniques are given in this chapter.
Reconstruction by IDWT is explained in detail followed by the signal compression by Shannon
Fano algorithm. The evaluation criteria for compression (PRD and CR) have been demonstrated
through Graphical representation.
Chapter 5 presents the general conclusions of the thesis and proposes possible improvements
and directions of future research work followed by references and Appendix A and Appendix B.