Vous êtes sur la page 1sur 4

International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882

Volume 5, Issue 2, February 2016

Design of Quadrature Mirror Filter Bank by using Levenberg-Marquardt


Algorithm
D. S. L. Satya Vani 1, K. Ponnika2, K. Amulya3, M. Divya4, K. Akhila 5
1, 2, 3, 4, 5
B. Tech. Student, Electronics and Communication Engineering,
Bapatla Engineering College, A. P, India

Abstract
This paper presents an improved and efficient
method for the design of a two-channel quadrature
mirror filter (QMF) bank. In the proposed method, the
filter bank design problem is formulated as a low-pass
prototype filter design problem, whose responses in the
passband and stopband are ideal and their filter
coefficients value at quadrature frequency is 0.707. A
new method is developed for the design of a low-pass
prototype filter which minimizes the objective function
by optimizing the filter taps weights using the
LevenbergMarquardt method. QMF has wider
applications such as speech processing, image
processing, biomedical signal processing and in
communication systems, modems, data transmissions,
speech and audio coding, video coding. Hence
minimizing the error at the output of QMF is very
essential.
Keywords: Filter banks; LevenbergMarquardt
method; QMF bank; Subband coding.

I. INTRODUCTION
In signal processing, sub-band coding (SBC) is any
form of transform coding that breaks a signal into a
number of different frequency bands and encodes each
one independently. This decomposition is often the
first step in data compression for audio and video
signals. QMF bank was the first type of filter bank
used in signal processing applications for separating
signals into subbands and reconstructing them from
individual subbands. But the original subband coding
systems were suffered from aliasing distortion,
amplitude distortion and phase distortion.
The design of quadrature mirror filter (QMF)
banks has been studied widely in recent years because
of the wide applications of QMF banks in many
engineering fields [1]. Initially, the concept of QMF
bank was introduced to confiscate aliasing distortion in
speech coding. Subsequently, a considerable progress
has been occurred in many engineering fields such as
subband processing of audio, image and video signals,
analog to digital converters, signal compression
systems, design of wavelet bases, antenna systems,
digital audio industry and biomedical signal
processing. The general theory of QMF banks and
their designs was developed by many researchers [4].
Neural network based architecture is presented to

design the QMF bank without any convergence


problem. A method based on eigenvector computation
in each iteration is proposed to obtain the optimum
quantized filter weights [3]. In comparison to earlier
band pass filter based subband coding systems, the
QMF bank based systems have many advantages as
followed:
(a) Aliasing distortion is eliminated in QMF bank
based subband coding systems; therefore, the transition
width of the filters is not much important. Lower order
filters with wider transition band can be used.
(b) Computation complexity is reduced in case of
subband coding system based on QMF banks.
(c) Lower bit rates are possible, without degrading the
quality of decoded speech signals.
(d) QMF based subband coders provide more natural
sounding, pitch prediction, and wider bandwidth than
earlier subband coders
In a two-channel QMF bank, input signal x(n)
is divided into two overlapping subbands having equal
bandwidth, using the low-pass and high-pass analysis
filters H0(z) and H1(z), respectively shown in Figure 1.
Several techniques have been reported for synthesizing
the prototype filter for two-channel QMF banks in
frequency domain and in time-domain [2]. These
subband signals are decimated by a factor of two to
achieve signal compression or to reduce processing
complexity. The decimated signals are typically coded
and transmitted.

Figure (1)
At the receiver, the two subband signals are
decoded and then interpolated by a factor of two and
finally passed through low-pass and high-pass
synthesis filters, 0() and 1(),respectively. The

www.ijsret.org

116

International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 5, Issue 2, February 2016

outputs of the synthesis filters are combined to obtain


the reconstructed signal y(). The reconstructed
signal
() suffers from three types of errors: aliasing
distortion (ALD), amplitude distortion (AMD), and
phase distortion (PHD), due to the fact that the filters
0 (), 1 (), 0 (), and 1 () are not ideal.
At the receiver, the two subband signals are
decoded and then interpolated by a factor of two and
finally passed through low-pass and high-pass
synthesis filters, 0() and 1(),respectively. The
outputs of the synthesis filters are combined to obtain
the reconstructed signal y(). The reconstructed
signal
() suffers from three types of errors: aliasing
distortion (ALD), amplitude distortion (AMD), and
phase distortion (PHD), due to the fact that the filters
0 (), 1 (), 0 (), and 1 () are not ideal. In most
of applications, a common requirement is that
reconstructed signal x() should be as close to ()
as possible. Therefore the main stress of most of the
researchers while designing filters for the QMF bank
has been on the elimination or minimization of the
three distortions to obtain a perfect reconstruction (PR)
or nearly perfect reconstruction (NPR) system.
The paper is structured as follows: Sect. II defines the
Design of QMF bank. The algorithm for implementing
Levenberg-Marquardt method is in Sect. III and a
performance analysis of this method is given in Sect.
IV and conclusion, simulation results are shown in
Sect. V respectively.

II. Design of QMF Bank


Consider a two-channel QMF bank with system
architecture as shown in Fig. 1. The reconstructed
output signal is
Y (z) = T (z)X(z) + A(z)X(z)
where the first term
T (z) = [H0(z)G0(z) + H1(z)G1(z)]
is the distortion transfer function, and the second term
A(z) = [H0(z)G0(z) + H1(z)G1(z)]
is the aliasing distortion transfer function. In the above
equations, X(z) is the original input signal. By setting
G1(z) =H0(z)
and
G0(z) = H1(z)
the aliasing
error is completely eliminated, and giving the input
output relation:
Y (z) = [ H0(z)H1(z) H1(z)H0(z)]X(z)
Since, the mirror image relation exists between H1(z)
and H0(z) at quadrature frequency, the alias-free
reconstructed output signal is
Y (z) = [H02(z) H02 (z)] X (z)
Now, T (z) is reduced to
T (z) = [H02(z) H02(z)]
Let H0(z) be a low-pass finite impulse response (FIR)
filter with frequency response:
H0(e j)= ejN/2H0()

where H0() is the amplitude response of the filter.


Then, the overall frequency response of QMF bank
becomes
T(e j)= ejN [|H0(e j)|2 (1)N|H0(e j()|2 ] /2
If the order of filter N is odd, the above equation gives
T (e j) = 0 at = /2,
implying amplitude distortion. Therefore, N must be
chosen even to avoid this distortion and Eq. (8) is
rewritten as if the order of filter N is odd, the above
equation gives T (e j) = 0 at
=/2, implying
amplitude distortion. Therefore, N must be chosen
even to avoid this distortion and Eq. is rewritten as
T(e j)= ejN [|H0(e j)|2 +|H0(e j()|2 ] /2
If H0(z) is assumed to have ideal response in passband
and stopband, the reconstruction error exists in
transition band. If the amplitude response of QMF
bank in transition band is optimized, the overall
performance can be improved. The overall amplitude
response of the filter bank is
|T ()|=|H0()|2 +|H0( )|2
If the prototype filter H0(e j) has ideal characteristics,
then its amplitude response H0() must satisfy
H0() = H0(0)
for[0, p]
= 0,
for [s, ]
Thus, for perfect reconstruction, the overall amplitude
response of QMF bank must be equal to square of the
frequency response of prototype filter at zero
frequency and at
= 0.5, Eq. (10) is reduced to
T(0.5)|2=|H0(0.5)|2+|H0(0.5 )2
=|H0(0)|2+2|H0(0.5)|2
=|H0(0.5)|
= 0.707|H0(0)|
which is a perfect reconstruction condition, and is also
the prototype filter response at quadrature frequency.
Thus, the design problem of filter reduced to design a
low-pass filter whose response is
D()=H0(0)
for[0,wp]
=0
for [s, ]
=0.707H0(0)
at = /2
Therefore, a new method is developed for designing a
low-pass prototype filter by minimizing the quadratic
measure of errors in desired response of filter. The
objective function () has been formulated using the
weighted sum of errors in desired response
= t + p + (1 ) s,
for (0, 1]
where p , s and t are respectively error in passband,
stopband and transition band given by
t = [H0() 0.707H0(0)]2
at = /2
p = 1/ [H0(0) H0()]2 d
for[0,wp]
s = 1/[H0()]2 d
for[ws, ]
In Eq. (15), is a constant, which controls the relative
accuracies of approximation in passband and stopband.
The amplitude response of the prototype filter is given
by
H0() = [2h0(n) cos(n + )]

www.ijsret.org

117

International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 5, Issue 2, February 2016

for n=0 to [N/2]-1


= [b(n) cos(n + )]
for n=0 to M-1
where M = N/2 and b(n) = 2h0(n).The above equation
can be rewritten as
H0() = bT c()
Where
b = [b 0 b1 b2 . . . bN/21] T
and
c() =[cos(/2) cos(3/2) . . . cos((N 1)/2)]T
Using the above equations , the design problem in
transition band (t ) can be implemented as
t =(bT v 0.707H0(0))2
where v is a vector equal to vector c(), when it is
evaluated at = /2 and H0(0) is equal to
H0(0) = bT c(0) = bT 1
where 1 is a vector with all (N/2) elements equal to
unity. The error in passband p can be realized as:
p = bT Pb
where P is a real valued, symmetric matrix, given by
P = 1/ (1 c())(1 c())T d for [0,wp]
Similarly, the error in stopband s can be rewritten as
s = bT Sb
where S is also a real valued, symmetric matrix, which
is equal to
S = 1/ccT d
for [ws, ]
From the above equations, can be redefined as
= (bT v 0.707H0 (0)) 2 + bT Pb + (1) bT Sb
Now, the design problem of prototype filter is deduced
to estimate vector b such that for given value of it, the
objective function is reduced to minimum value.
The input/output relation of a typical twochannel QMF bank (as shown in Fig. 1) can be
formulated as [17]

The first term of the above equation represents a linear


shift-invariant response between X(z) and X(z);
whereas the second term represents the aliasing error
because of change in sampling rate.
By setting

e) The reconstructed signal will be the


combination of the outputs from both the
filters of synthesis section.
f) To find out the error subtract the original
signal from reconstructed signal and also
consider the delay occurred during filter
process. The other way to find the error is to
compare the magnitude spectra of original
signal with that of reconstructed signal by
plotting both spectras on the same plane using
PSD command.
g) Create a network object and initialize it with
the help of newff command.
h) The network is trained for K epochs (where K
is a constant value).
i) After initializing the network, the network
training is originated using train command.
j) To compare results compute the output of the
network with training data and validation data
with the help of sim command.
k) Now repeat the steps from (a) to (f) and then
compare the PSD of signals in the first case
(i.e. without training the network by using LM
method) and in the second case (i.e. after
training the network using LM method).
In the proposed algorithm the value for the order
of filter is assumed to be 99 and the number of epochs
is considered as 500.

IV. Results and Discussion


In this section, the proposed method has been used for
the design of the prototype filter for a two-channel
QMF banks. A design example is included to
demonstrate the usefulness of this algorithm. The
performance of this method can be observed from
PSDs of the signals with and without implementing
Levenberg-Marquardt method and it is observed that
the error has reduced after the implementation of the
proposed algorithm.

and
the aliasing error in the above equation can be
eliminated.

III. Levenberg-Marquardt Method


Steps to implement this method in the design
procedure are as follows:
a) The order of the filter must be an odd number.
b) Design the analysis and synthesis filters and
see the response.
c) Examine the signal properties using
periodogram command.
d) Simulate the reconstruction of the signal and
filter that signal by using the filter command.

Figure (a) PSD of error without using LevenbergMarquardt Method

www.ijsret.org

118

International Journal of Scientific Research Engineering & Technology (IJSRET), ISSN 2278 0882
Volume 5, Issue 2, February 2016

We also express our sincere thanks to Dr. B. Chandra


Mohan, M.Tech., Ph.D., Head of the Electronics and
Communication Engineering Department for his
encouragement and providing the necessary facilities.
We express our deep sense of gratitude and sincere
appreciations to our project guide SK. M. Subhani,
M.Tech, for his esteemed guidance and constant
encouragement throughout the project. We are
indebted for his instruction guidelines that proved to be
very much in completing our project successfully in
time.
Figure (b) PSD of error signal after implementing
Levenberg-Marquardt method

V. Conclusion
In this paper, we have presented Levenberg-Marquardt
algorithm for the design of Perfect reconstruction
QMF filter bank. The method is based on the ideal
response of low-pass prototype filter in passband,
stopband, and filter coefficients value at quadrature
frequency. The simulation results clearly reveal that
the proposed method yields the smallest reconstruction
error relative to that achieved with designs obtained
using other methods. The amount of computational
time required with this method is low and the
algorithm is simple to implement compared to other
optimization methods. Hence, the result for proposed
method shows better performance when compared to
the approaches of previous papers.

References
[1] Chen, C.K., Lee, J.H.: Design of quadrature mirror
filters with linear phase in the frequency domain. IEEE
Trans. Circuits Syst.II 39(9), 593605 (1992)
[2] Croisier, A., Esteban, D., Galand, C.: Perfect
channel
splitting
by
use
of
interpolation/decimation/tree
decomposition
techniques. In: International Conference on
Information Sciences and Systems, Patras, pp. 443
446 (1976)
[3] O.P. Sahu, M.K. Soni, I.M. Talwar, Marquardt
optimization method to design two-channel quadrature
mirror filter banks, Digital Signal Process. 16 (6)
(2006) 870879
[4] Jou, Y.D.: Design of two channel linear phase
QMF bank based on neural networks. Signal Process.
87(5), 10311044 (2007)

Acknowledgements
We are thankful to the sanctum BAPATLA
ENGINEERING COLLEGE for giving us the
opportunity to fulfil our aspirations.
We take the opportunity to express our heartfelt
gratitude to our beloved Principal
Dr. N. Sudhakar,
M.Tech., Ph.D., for providing necessary facilities in
the department.

www.ijsret.org

119

Vous aimerez peut-être aussi