You are on page 1of 5

Wavelet Compression of EMG Signal

Sarina Mansor, C. Eswaran and Rosli Bin Besar


Faculty of Engineering, Multimedia University, 63100 Cyberjaya, Selangor, Malaysia

Abstract information for the assessment of


neuromuscular disorders. Generally, the
This paper presents some experimental decomposition of an EMG signal is the
results of EMG signal compression using separation of the complex signal into its
several types of wavelets. Based on these constituent motor unit action potentials
results, the optimal wavelet for compressing (MUAP) [1]. Therefore, most of the analysis
EMG signal is identified. Next, a of this signal is focused on the methods to
comparison between the performances of the identify and classify MUAPs [2]. Besides
wavelet transform and the discrete cosine that, Davies and Reisman [3] have applied
transform (DCT) for EMG signal the time-frequency analysis to the EMG
compression is carried out. The results signal, showing a good spectral compression
indicate that wavelet transforms outperform during fatigue. Lacetra and Olmo [4] have
the DCT in terms of compression ratios and proposed the use of matched wavelet
signal quality. transform as a new methodology for EMG
interpretation. The wavelet analysis is
chosen because it matches with the shape of
1. Introduction MUAP, and the resulting wavelet transform
yields the best possible energy localisation
Wavelet transforms have received in the time-scale [4]. In other words, the
significant attention recently from EMG signal can be best represented by
mathematicians, signal analysts and wavelet transform that leads to several
engineers as a new tool for signal interesting applications such as MUAP
compression. Unlike the traditional Fourier detection and data compression.
techniques, wavelets are localised both in
time and frequency domain. This feature The purpose of this paper is to
makes them suitable for the analysis of demonstrate the potential use of different
nonstationary signals. One interesting area types of wavelets for EMG signal
of wavelet applications is biomedical compression. The paper is organised as
engineering. Wavelets have been applied to follows. The wavelet theory and its fast
several problems in biomedical signals, pyramidal algorithm are summarised in
including signal analyses, feature extraction Section II. Section III describes the scheme
and data compression; in particular for EEG used to demonstrate the wavelet
and ECG signals. However, analysis of compression of EMG signal. The results
EMG signals by using wavelet transforms obtained by using eight different types of
has not been widely explored yet. wavelets are presented in Section IV. The
performances of these wavelets are also
Electromyography (EMG) signals compared with that of discrete cosine
represent the electrical activity of muscle transform (DCT). We conclude with some
during contraction, which provide useful remarks in Section V.
H ↓2 ↑2 H′
S H ↓2 ↑2 H′ S
G ↓2 ↑2 G′
G ↓2 ↑2 G′

Analysis Wavelet Coefficients Synthesis


DWT Inverse DWT

Fig. 1: The forward and inverse discrete wavelet transform

2. Wavelet Theory digital data and filtering, we need to


discretize the wavelet transform resulting in
The wavelet transforms decomposes a the Discrete Wavelet Transform (DWT).
signal f(t) into its basis functions which are The mother and scaling functions can be
generated by dilations and translations of a defined as [5]:
prototype function ψ called mother wavelet.
ψ ( t ) = ∑ 2 1/2 h( n )ψ ( 2t − n)
− i/ 2 i
ψ i, k ( t ) = 2 ψ(t 2 − k ) , k,i ∈ Ζ n

φ( t ) = ∑ 21 / 2 g ( n )φ( 2t − n)
n
The above equation shows that the mother
function is dilated by the integer i and where h(n) and g(n) are highpass and
translated by the integer k. The wavelet lowpass filters respectively.
decomposition of a signal can be represented
as below For computation of wavelet transform, we
use the pyramidal algorithm developed by
f ( t ) = ∑ ∑ c i ,k ψi , k Mallat [6]. This algorithm has two parts: the
i k
forward algorithm, which serves to compute
the wavelet transform and the backward
The coefficients ci,k are obtained through algorithm, which computes the inverse
the following inner product: transform.

c i ,k = f ,ψ~i ,k The forward version consists of two linear


filters, lowpass (G) and highpass (H) which
The function ψ~ is the algebraic dual of decompose the signal into low and high
frequency components respectively.
the mother wavelet ψ. In the orthogonal
Convolving the signal with H gives us a set
case, ψ~ and ψ are identical. of wavelet coefficients ci,k while the
convolution with G gives us the
In a multiresolution analysis, one really approximation coefficients dk. Because of
has two functions: the mother wavelet ψ and the redundancy of information, these filters
the scaling function φ. It is helpful at this are down-sampled, throwing away every
point to view the scaling function as a low other sample at each operation, thus halving
pass filter and wavelet function as a high the data each time. The approximation
pass filter. For application involving with coefficients dk are then convolved again
with the filters G and H to form the next similar to the one applied for speech signal
level of decomposition. [9].

The backward algorithm simply inverts the Original Reconstructed


data data
process. It combines two linear filters, H′ DWT Truncation IDWT
and G′ with up-sampling operation, which
inserts a zero between data values, doubling Fig.2: The simple scheme of signal compression
the data at each operation. Fig. 1 shows the
operations involved in the wavelet As a sample signal, we use an EMG signal
decomposition and synthesis of the signal. vector of length 4096, representing the
The impulse responses of the decomposition activity of tibialis anterior muscle during
and synthesis filters, h(n) and g(n) are walking. Fig. 3 shows the EMG signal used
related as: for experimentation.
1.5

g ′( n ) = ( −1) n −1
h( n − 1)
1
h′( n ) = (−1) g( n − 1)
n

0.5

Signal Amplitude
Unlike Fourier series which has a single 0
set of basis function, wavelets have infinity
set of basis functions; provided that ψ(t) ∈ -0.5

L2 and -1

+∞ -1.5

∫ ψ(t )dt = 0 -2
−∞ 0 500 1000 1500 2000 2500 3000 3500 4000 4500
Number of data

There are different trade-offs between how Figure 3: The input EMG signal
compactly the basis functions are localised
in space and how smooth they are when After wavelet transformation, all wavelet
using different types of wavelets. For signal coefficients with magnitudes larger than the
compression, it is preferable for ψ to have threshold value are retained with full
compact support so that the signal features accuracy and the remaining coefficients are
in both time and scale can be discriminated truncated (ie: set to zero). Then, the
[7]. However, at present, there exist no compression ratio (CR) is defined as [10]:
theoretical results that can predict which
wavelet is suitable for a particular type of CR =
signal [8]. Usually, the optimal wavelet is Original number of wavelet coefficients
chosen by comparing the performances of Number of wavelet coefficients after
several types of wavelets. thresholding

3. Compression of EMG Signal The threshold value can be adjusted to


alter the number of preserved coefficients,
To demonstrate the potential of wavelets thereby changing the compression ratio. The
for EMG signal compression, we use a retained coefficients are used to reconstruct
simple scheme of signal compression as the signal. In this experiment, we use the
shown in Fig. 2 (truncation approach) percent root-mean-square difference (PRD)
to measure the numerical distortion between compression ratio is now defined as the ratio
the original and the reconstructed EMG of the number of bits required before and
signal. after encoding. In this experiment, Huffman
coding is employed to improve the
compression ratio. Fig. 5 illustrates the rate-
distortion curves obtained when we encode
the coefficients. From Figs. 4 and 5, it is
clear that the compression ratio is improved
by using the encoder.

where f(i) is the original signal and f$ ( i ) is


40

the reconstructed signal.


35

For comparison, we choose the discrete


cosine transform (DCT) which is the most
30
commonly used transform for data

PRD
compression. A detail description of DCT Db6

Db10
can be found in [11]. 25 Bior3.5

Bior6.8

4. Results and Discussion Coif3

20 Coif5

Sym4

A number of experiments were performed Sym7

to evaluate the ability of wavelets to 15


DCT

10 15 20 25 30 35
compress the EMG signal and to identify the Compression Ratio

optimal wavelet for this type of signal. The Fg. 4: The performance curves of the eight wavelets
wavelet bases used in the experiments are and DCT
Daubechies (Db6 and Db10), Biorthogonal
(Bior3.5 and Bior6.8), Coiflets (Coif3 and
40
Coif5) and Symlets (Sym4 and Sym7).

The rate-distortion curves (PRD versus 35


compression ratio) obtained for eight
wavelets and DCT are shown in Fig. 4. It
can be inferred from this figure that the 30

performance of the wavelet transform is


PRD

Db6

superior to that of DCT with regard to Db10


25 Bior3.5
compression efficiency and reconstruction
Bior6.8
quality. In terms of optimal wavelet, the Coif3

Bior6.8 wavelet performs slightly better than 20 Coif5

the others. Sym4

Sym7

DCT
Wavelet coefficients are encoded in order 15
10 15 20 25 30 35
to reduce the number of bits required for Comp Ratio

transmission/storage. Encoding schemes


such as arithmetic, Huffman, zerotree Fig. 5: The performance curves of the eight wavelets
coding can be used for this purpose. The and DCT when combining with Huffman encoder
The truncation approach used here is simply [3] Michelle R Davies, Stanley
based on the magnitude of each individual S.Reisman, “Time frequency
coefficient, ignoring the similarity of analysis of the electromyogram
neighbouring coefficients. The similarity during fatigue,” 20th Annual
properties may result in a better quantization Northeast Bioengineering
scheme as they provide a more suitable Conference 1994, Proceedings of
threshold criterion for truncation. For 1994.
achieving better results, we can use several [4] F.Laterza and G.Olmo, “Analysis
wavelet coders which exploits the of EMG signals by means of the
correlation between wavelet coefficients; matched wavelet transform,”
such as the wavelet compression algorithm Electronics letters, Vol 33, No. 5,
based on embedded zerotree wavelet (EZW) 1997
developed by Michael Hilton [7]. [5] Marc Antonini, Michel Barlaud,
Pierre Mathieu, Ingrid Daubechies,
5. Conclusions “Image Coding Using Wavelet
Transform,” IEEE Trans. on Image
Our experiments show the promising use Processing, vol 1, no. 2, April1992.
of wavelet transforms for EMG signal [6] S.G. Mallat, “A theory for
compression. It has been found that the multiresolution signal
results obtained with EMG signal are quite decomposition: The wavelet
similar to the one obtained in [7] for the case representation,” IEEE Trans.
of ECG signal. Wavelets also give better Pattern Anal. Machine Intell., vol.
compression efficiency and signal quality 11, no. 7, pp. 674-693, July 1989.
compared to DCT. Better results may be [7] Michael L. Hilton, “Wavelet and
obtained if a sophisticated coder such as the wavelet packet compression of
embedded zerotree wavelet (EZW) coding electrocardiograms,” IEEE Trans.
or the discrete wavelet transform coding on biomedical engineering, vol 44,
with vector quantization is used. no. 5, May 1997.
[8] John D. Villasenor, Benjamin
Belzer and Judy Liao, “Wavelet
REFERENCES Filter Evaluation for Image
Compression,” IEEE Trans. on
[1] B Thompson, P Picton and N B
Image Processing, vol 4, no. 8,
Jonest, “A comparison of neural
August 1995.
network and traditional technique
[9] W.Kinsner and A.Langi, “Speech
in the classification of EMG
and image signal compresson with
signals,” Artificial Intelligence
wavelets,” IEEE Trans. Signal
Methods for Biomedical Data
Processing, 1993.
Processing, IEEE Colloquium
[10] I.Singh, P.Agathoklis, Antoniou,
1996.
“Compression of Bathymetric Data
[2] Christodoulus I. Christodoulu,
Using Wavelets,” IEEE 1995.
Constantinos S. Pattichis, “ A new
[11] C.R Gonzalez and P.Wintz, Digital
technique for the classification and
Image Processing, 2nd ed., Addison
decomposition of EMG signals,”
Wesley, 1987.
IEEE International Conference on
Neural Networks 1995,
Proceedings of 1995.

You might also like