Applied Soft Computing 13 (2013) 353–360

Contents lists available at SciVerse ScienceDirect

Applied Soft Computing

journal homepage: www.elsevier.com/locate/asoc

Review Article

Design of nearly perfect reconstructed non-uniform filter bank by constrained

equiripple FIR technique
A. Kumar a,∗ , G.K. Singh b , S. Anurag a
a
PDPM Indian Institute of Information Technology Design and Manufacturing, Jabalpur 482005, MP, India
b
Indian Institute of Technology Roorkee, Roorkee 247667, Uttrakhand, India

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, an efficient iterative algorithm is proposed for the design of multi-channel nearly perfect
Received 16 August 2011 reconstructed non-uniform filter bank. The method employs the constrained equiripple FIR technique
Received in revised form 14 May 2012 to design the prototype filter for filter banks with novelty of exploiting a new perfect reconstruction
Accepted 5 August 2012
condition of the non-uniform filter banks instead of using complex objective functions. In the proposed
Available online 29 August 2012
algorithm, passband edge frequency (ωp ) is optimized using linear optimization technique such that
the filter coefficients values at quadrature frequency are approximately equal to 0.707. Several design
Keywords:
examples are included to illustrate the efficacy of this methodology for designing non-uniform filter bank
Filter banks
QMF
(NUFB). It was found that the proposed methodology performs better as compared to earlier reported
Subband coding results in terms of reconstruction error (RE), number of iteration (NOI) and computation time (CPU time).
Non-uniform filter bank (NUFB) The proposed algorithm is very simple, linear in nature, and easy to implement.
Tree-structured © 2012 Elsevier B.V. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
2. Overview of non-uniform filter bank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
3. Proposed methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
4. Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
5. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359

1. Introduction which is required for several specific applications. These specific

applications include sub-band coding like audio coding, speech
Many advancements in the area of multirate filter banks in coding, data and image compression [1,2]. Uniform filter banks
conjunction with the ever increasing numerous applications have have many constraints like integer and uniform decimation in each
made multirate filter banks design an increasingly important field sub-band, and limited time frequency resolution. These constraints
of research. The research effort was first focused on design of a two- catalyze the importance of non-uniform filter banks (NUFBs). In
channel quadrature mirror filter (QMF) bank [1,2], which was later addition, NUFBs are able to provide any sort of rational decimation
extended to design of M-channel filter banks [2–5]. Since then, sev- in each channel, any extent of time–frequency resolution as per
eral techniques [6–11] were developed to enhance the performance requirement of the application, less quantization error, and low
of filter banks in different engineering fields. computational complexity. Over the past few years, a number of
Among different types of the filter banks, non-uniform filter design methods [12–15] have been proposed by different authors
banks have been elicited immense interest in the researchers in for the design of multi-channel filter banks. However, design of a
recent years due to their ability to differentiate information into linear phase multi-channel filter bank with linear optimization has
different frequency bands based on energy distribution of signal been still an issue since a very few references [16–20] provide linear
phase as well as zero aliasing error, which is very much essential
in several applications such as videos and communication systems.
∗ Corresponding author. Such an application oriented technique was presented in [21]. In
E-mail addresses: anilkdee@gmail.com (A. Kumar), gksngfee@gmail.com
this technique, the authors have been used the evolutionary pro-
(G.K. Singh), anurag1020201@iiitdmj.ac.in (S. Anurag). gramming algorithms to design the optimized prototype filter for

1568-4946/$ – see front matter © 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.asoc.2012.08.024
354 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360

H 100 (z) 2 2 G100 (z)

Processing Unit
2 G0(z)
H0(z) 2
y (n)
x(n) H 1
10 (z) 2 2 G 110 (z)

H1(z) 2 2 G1(z)

(a)

4 G0 (z) = G0 (z) G 100 (z 2)

H0 (z) =H0 (z) H 100 (z 2) 4

Processing unit
y(n)
x(n)
H1 (z) =H0 (z) H 110 (z 2) 4 4 G1 (z) = G0 (z) G 110 (z 2)

H2 (z) =H1 (z) 2 2 G2 (z) = G1 (z)

(b)

Fig. 1. A block diagram of multichannel non-uniform filter bank. (a) A block diagram of a tree structured filter bank and (b) its equivalent parallel structure.

designing modulated filter banks. Advances in filter banks have for NUFB. In Section 4, design results and application of the pro-
provided a new generation of subband coders for audio, image posed method to subband coding is carried out, followed by the
and video signals, analog to digital converters, signal compression concluding remarks in Section 5.
systems, design of wavelet bases, antenna systems, digital audio
industry and biomedical signal processing [1,2]. 2. Overview of non-uniform filter bank
Recently, several design methods [22–27] have been proposed
and evaluated for designing the non-uniform filter bank based on The non-uniform filter bank with integer decimation and linear
optimization and non-optimizations. But still, there is no such iter- phase is realized with the help of tree structured techniques, which
ative technique reported in the literature which can reduce the is based on building the filter bank using a two-channel filter bank
computation time, converse in low number of iteration and also as basic building blocks [1,2,32–34]. The generalized structure of M-
reduces the peak reconstruction error which can be used for filter channel filter banks based on tree structure approach is depicted in
banks with larger taps. Therefore, the authors in [28] have pro- Fig. 1. For M-channel NUFB having decimation M0 , M1 , M2 , . . ., MM−1
posed an optimized algorithm for designing NUFB with Blackman for each band, then decimation factors are such that [1,2,35,36]
Window family based on the algorithm given in [16]. Similar to
the cut-off frequency, a suitable value of passband edge frequency 
M−1
1
(ωp ) can reduce the amplitude distortion. There are very few refer- =1 (1)
Mk
ences [16] available in which ωp has been optimized for designing k=0
a prototype filter for two-channel QMF banks and M-channel CM
and the reconstructed signal X̂(z) is
filter banks. Literature available so far on non-uniform filter banks
reveals that there is still need for a computationally efficient tech- 
M−1
1 
Mk −1
l l
nique, which shall use linear optimization technique for designing X̂(z) = Gk (z) X(zWM )Hk (zWM ) (2)
Mk k k
non-uniform linear-phase filter bank. Apart from multirate filter k=0 l=0
banks designing techniques, the authors in [29] have recently pro-
where Hk (z) and Gk (z) are the analysis and synthesis filters respec-
posed efficient optimization techniques to design two-dimensional l ) and X (zW l ) are
tively, while WL = e−j2/L . The terms Hk (zWM k M
IIR filters. This technique is based on the particle swarm intelli- k k
gence approach [30], which was initially introduced for simulating aliasing terms introduced due to decimation/interpolation. For the
human social behaviors. Later on, this particle swarm optimization perfect reconstruction (PR), these must be minimized by any strate-
(PSO) approach was improved by authors in [31] to give a new opti- gic approach. In a tree-structured approach, a two-channel QMF
mization algorithm called fitness-adaptive differential evolution bank is used as building blocks; hence the aliasing error is com-
algorithm to design QMF banks. A comparative study of modern pletely eliminated by proper selection of synthesis filters in terms
search techniques is presented in [32] for designing two dimen- of analysis filters. In a two-channel QMF bank, the aliasing error is
sional IIR filters. eliminated with
In this paper, a new improved iterative methodology is pre-
G1 (z) = −2H0 (−z) and G0 (z) = 2H1 (−z) (3)
sented for the design of non-uniform filter bank. Organization of
the paper is as follows: a brief introduction has been provided As depicted in Fig. 1(a), the input signal x(n) was decomposed
in this section on design techniques of NUFBs. Section 2 gives an into two subbands in which the sampling rate is reduced to 1/2 of
overview of NUFBs. Section 3 presents the proposed methodology the original sampling rate. These subbands can be further extended
 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360 355

to more subbands by applying the same decomposition. In general, and

2P subbands can be obtained by repeating the same decomposition
|HL (z)|4 + |HL (z)|2 |HL (−z)|2 + |HL (−z)|2 = 1. (13)
process P times and at each of the P stages of decomposition process,
the number of two-channel QMF bank structure required, is 2P−1 . In frequency domain, Eq. (13) is reduced to
Total number of two-channel systems required is 2P−1 . On the other
|HL (ejω )|4 + |HL(ejω )|2 |HL(ej(ω−) )|2 + |HL(ej(ω−) )|2 = 1 (14)
side, the process of reconstructing the original input signal can be
seen as mirror image of the decomposition process at analysis side and if it is computed at quadrature frequency (ω = 0.5), then
as illustrated in Fig. 1(a) and its equivalent parallel structure shown
2|HL (ejr/2 )|4 + |HL (ejr/2 )|2 = 1 (15)
in Fig. 1(b). After resolving the tree structured nonuniform filter
bank into its parallel forms, the following relations can be deduced Let HL (ej/2 ) = x, then Eq. (15) is simplified as
[1,2,28]: 4 2
2x + x − 1 = 0 (16)
H0 (z) = HL (z), HL (z 2 ), G0 (z) = GL (z), GL (z 2 ) (4)
Using simple factorization approach, the roots of above polynomial
H1 (z) = HL (z), HH (z 2 ), G1 (z) = GL (z), GH (z 2 ) (5) are
H2 (z) = HH (z), G2 (z) = GH (z) (6) 1
x = ± √ = ±0.7071 and x = ±i (17)
2
Performance of tree structured approach depends on the effi-
cient design of QMF bank. The computation load is reduced due to The imaginary and negative roots are not possible, so discarding
some similarities between the coefficients of high-pass and low- them and then, it leads to Eq. (18), giving a new PR condition.
pass filters of a two-channel filter bank. However, the frequency x = |HL (ej/2 )| = 0.7071 (18)
characteristics of the filters might differ from stage to stage, but
they must be the same within a particular stage. Depending on Similarly, in 4-channel NUFB with decimation factors (8, 8, 4, 2),
which of the filter bank is used in the design, there would be per- the perfect reconstruction can be achieved, if Eq. (19) is satisfied
fect reconstruction (PR) or near perfect reconstruction (NPR). But |H0 (ejω )|2 + |H1 (ejω )|2 + |H2 (ejω )|2 + |H3 (ejω )|2 = 1,
the realization of perfect reconstruction filter bank requires very

tedious and complex approach, which makes it practically unreal- for 0≤ω≤ (19)
5
izable. Therefore, a common interest of researchers is towards NPR,
since it is a practically realizable approach. But during realization where
of this approach, there is always an introduction of three distort- H0 (z) = HL (z)HL (z 2 )HL (z 4 ), H1 (z) = HL (z)HL (z 2 )HH (z 4 ),
ions: amplitude distortion, phase distortion and aliasing error. The
aliasing error can be removed by using proper pairs of analysis and H2 (z) = HL (z)HH (z 2 ), and H3 (z) = HH (z). (20)
synthesis filters, phase error by using linear phase filters, and ampli-
tude distortion can be minimized by using different optimization Using Eq. (20), then Eq. (19) turns out to be
techniques.
|HL (z)HL (z 2 )HL (z 4 )|2 + |HL (z)HH (z 2 )HH (z 4 )|2
3. Proposed methodology + |HL (z)HH (z 2 )| + |HH (z)|2 = 1 (21)

In multi-channel non-uniform filter bank [1,2], the perfect

reconstruction is possible if As discussed earlier, the energy of original signal and its up
sampled version has same value, therefore

M−1

|Hk (ejω )|2 = 1, for 0<ω≤ (7) |HL (z 4 )|2 = |HH (z 2 )|2 = |HL (−z)|, and |HH (z 4 )|2
M
k=0
= |HH (z 2 )|2 = |HH (z)|2 (22)
where Hk (ejω )
is kth analysis filter in equivalent parallel structure
of a non-uniform filter bank, and M is the number of bands in NUFB. Using the above equation and HH (z) = HL (− z), Eq. (21) is further
For example in Fig. 1(b), perfect reconstruction is possible if refined as

|H0 (ejω )|2 + |H1 (ejω )|2 + |H2 (ejω )|2 = 1, for 0 ≤ ω ≤ (8) |HL (z)|6 + |HL (z)|4 + |HL (−z)|2 + |HL (z)|2 |HL (−z)|2 + |HL (−z)|2 = 1
3
If Eq. (8) is further simplified using the prototype filter HL (ejω ) of a (23)
two-channel QMF bank, then it is reduced to

|HL (z)HL (z 2 )|2 + |HL (z)HH (z 2 )|2 + |HH (z)|2 = 1 (9) and in frequency domain, it leads to

Since energy of the original signal and its up-sampled version will |HL (ejω )|6 + |HL (ejω )|4 |HL (ej(ω−) )|2 + |HL (ejω )|2 |HL (ej(ω−) )|2
be same as in up sampled signal. In up sampling, L − 1 equidistant + |HL (ej(ω−) )|2 = 1 (24)
zero valued samples are inserted between two consecutive samples
of the original signal. Therefore,
Similarly, if it is also evaluated at ω = 0.5, then
|HL (z 2 )|2 = |HL (z)|2 (10)
2|HL (ej/2 )|6 + |HL (ej/2 )|4 + |HL (ej/2 )|2 = 1 (25)
similarly,
Similar to 3-channel NUFB, let HL (ej/2 ) = x, and finally the above
|HH (z 2 )|2 = |HH (z)|2 (11) equation becomes a six degree polynomial equation in terms x
Using the above equations and the quadrature mirror filter condi- 2x6 + x4 + x2 − 1 = 0 (26)
tion (HH (z) = HL (− z)), Eq. (9) is redefined as
The roots of this equation is found using simple factoriza-
|HL (z)|4 + |HL (z)|2 |HL (−z)|2 + |HL (−z)|2 = 1 (12) tion approach, these are x = ±0.7071, 0.5 ± 0.866i and 0.5 ± 0.866i.
356 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360

Therefore, the above findings lead to same perfect reconstruction

Start
as defined by Eq. (18)
Like 3-channel, 4-channel NUFB, same procedure is repeated
for finding perfect reconstruction for 5-channel non-uniform filter Specify: band (M), passband edge frequency (ωP),
bank with decimation factors 16, 16, 8, 4, and 2. Just like 3-channel stopband edge frequency (ωS), filter tap length (N),
and 4-channel NUFBs, PR can be achieved if stopband attenuation (AS), and passband ripple (AP)

|H0 (ejω )|2 + |H1 (ejω )|2 + |H2 (ejω )|2 + |H3 (ejω )|2 + |H3 (ejω )|2 = 1,
 Initialize: step size (Step), Magnitude Response of a prototype
for 0 ≤ ω ≤ (27) filter (MRI) = 0.707, Tolerance (Tol) and Counter (Count).
5
where
Design prototype filter (HL (z)) using the constrained equiripple
H0 (z) = HL (z)HL (z 2 )HL (z 4 )HL (z 8 ), FIR technique and calculate magnitude response (MRC) at ω =
π /2. Also compute Error = MRI-MRD
H1 (z) = HL (z)HL (z 2 )HL (z 4 )HH (z 8 ), H2 (z) = HL (z)HL (z 2 )HH (z 4 ),

H3 (z) = HL (z)HH (z 2 ) and H4 (z) = HH (z) (28)

YES Is |Error| ≤ Tol_
Using Eq. (28), then Eq. (27) turns out to be
Design other filters
NO
|HL (z)HL (z 2 )HL (z 4 )HL (z 8 )|2 + |HL (z)HL (z 2 )HL (z 4 )HH (z 8 )|2 using prototype filter.
YES
+ |HL (z)HL (z 2 )HH (z 4 )| + |HH (z)|2 = 1 (29) Is MRC< MRI
End
Since ωp = ωp + step
NO
8 2 4 2 2 2
|HL (z ) | = |HL (z ) | = |HL (z ) | = |HL (z)2 | (30)
ωc = ωc - step
and
2 2 2
|HH (z 8 ) | = |HH (z 4 ) | = |HH (z 2 ) | = |HH (z)2 | (31)

Using the above equations and HH (z) = HL (− z), Eq. (29) is further
refined as
Step=Step/2
|HH (z)|8 + |HH (z)|6 |HH (−z)|2 + |HH (z)|4 |HH (−z)|2 and
Count =Count + 1
+ |HH (z)|2 |HH (−z)|2 + |HH (−z)|2 = 1 (32)
Fig. 2. A flowchart for the developed method.
and it leads to Eq. (33) in frequency domain

|HL (ejω )|8 + |HL (ejω )|6 |HL (ej(ω−) )|2 + |HL (ejω )|4 |HL (ej(ω−) )|2

+ |HL (ejω )|2 |HL (ej(ω−) )|2 + |HL (ej(ω−) )|2 = 1
Now, if it is evaluated at ω = 0.5, then
2|HL (ej/2 )|8 + |HL (ej/2 )|6 + |HL (ej/2 )|4 + |HL (ej/2 )|2 = 1
Let HL (ej/2 ) = x,
and finally the above equation becomes a eight
degree polynomial equation in terms x
2x8 + x6 + x4 + x2 − 1 = 0
The roots of this equation are found using simple factoriza-
tion approach and the roots are x = ±0.7071, −0.7071 ± 0.7071i,
0.7071 ± 0.7071i, and ±i. Since negative and imaginary roots are
not possible, so after discarding them
x = 0.7071 ⇒ |HL (ej/2 )| = 0.707
Therefore, the above findings lead to same perfect reconstruc-
tion as defined by Eq. (18). Now, the non-uniform filter bank
design problem is reduced to estimate the optimized values filter
coefficients so that their values at quadrature frequency are 0.707.
In the proposed method, this new PR condition is approximately
satisfied by adjusting the filter coefficients. Similar to other algo-
rithms [16], the passband edge frequency is optimized to satisfy
new perfect reconstruction. In this algorithm, instead of comput-
ing complex objective function, the filter coefficients are evaluated
at ω = /2 in each iteration, which reduces the computation com-
plexity. The desired steps required for design of prototype filter
using the proposed method are:
(33)
Step 1: Specify design specifications stopband attenuation (As ),
passband ripple (Ap ), and normalized passband (ωp ), stopband
frequency (ωs ), number of band (M), and step size.
(34) Step 2: Initialize counter and the magnitude response (MR) of the
prototype filter given by Eq. (18) and also specify tolerance (Tol).
Step 3: Design the prototype filter using constrained equiripple
finite impulse response (FIR) technique before the optimization
start. Calculate the magnitude response of designed filter (MRD)
(35)
at ω = /2. Also calculate error = MR − MRD.
Step 4: (A) If error is not comparable to tolerance (Tol), the passband
edge frequency (ωp ) is varied using the step size. It is varied in two
ways:
a. if MRD < MR, then increase ωp by step
b. Otherwise, decrease ωp by step
(36)
Step 4: (B) If error is not comparable to tolerance. Then, design the
other filters composing the parallel equivalent NUFB using Eqs.
(4)–(6) in case of 3-channel NUFB, Eq. (24) in 4-channel NUFB and
Eq. (28).
Step 5: Redesign the prototype filter using new ωp and same order.
Calculate MRD and also error.
Step 6: Increment the counter by 1 and step = step/2. Go to step 4
till error is not comparable to tolerance.
A flowchart of the complete routine is shown in Fig. 2. The pro-
posed method has been simulated in Matlab 7.6 on Genuine Intel
(R) CPU T2300 @ 1.66 GHz, 1 GB RAM.
 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360 357

(a) 50 (b) 100

Magnitude ->(dB)
Magnitude in dB:
0
-100
-50
-200
-100
-300

-150 -400
0 1 2 3 4 5 0 0.1 0.2 0.3 0.4 0.5
Normalized Freq. by 2pi: Normalized freq by 2pi:

(c) 1.003 (d) 0.02

1.002

Magnitude ->(dB)
0.01
Magnitude ->

1.001
0
1
-0.01
0.999

0.998 -0.02
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Normalized freq by 2pi: Normalized freq by 2pi:

Fig. 3. Performance of 3-channel non-uniform filter bank (4, 4, and 2) designed using the proposed method. (a) Magnitude responses of the prototype filter in dB, (b)
magnitude responses of the analysis filters in dB, (c) variation of peak reconstruction error and (d) variation of peak reconstruction error in dB.

4. Results and discussion this method is evaluated in terms of number of iterations (NOI),
computation time (CPU time), and reconstruction error (PRE) given
In this section, the proposed methodology has been imple- by
mented on MATLAB and employed for designing non-uniform
M−1  M−1 
filter bank via constrained equiripple FIR technique. To illus-    
trate the efficacy of the developed method, many examples with PRE = max  jω 2
Hk (e ) − min  jω 2
Hk (e ) (37)
different design specifications are included. The performance of k=0 k=0

(a) 50 (b) 50

0
Magnitude in dB:

Magnitude ->(dB)

-50 -50

-100 -100

-150 -150
0 1 2 3 4 5 0 0.1 0.2 0.3 0.4 0.5
Normalized Freq. by 2pi: Normalised freq. by 2pi:

(c) 1.003 (d) 0.02

1.002
Magnitude ->(dB)

0.01
Magnitude ->

1.001
0
1

-0.01
0.999

0.998 -0.02
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Normalised freq. by 2pi: Normalised freq. by 2pi

Fig. 4. Performance of 4-channel non-uniform filter bank (8, 8, 4 and 2) designed using the proposed method. (a) Magnitude responses of the prototype filter in dB, (b)
magnitude responses of the analysis filters in dB, (c) variation of peak reconstruction error and (d) Variation of peak reconstruction error in dB.
358 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360

a 50 b 50

Magnitude ->(dB)
Magnitude in dB:
0 0

-50 -50

-100 -100

-150 -150
0 1 2 3 4 5 0 0.1 0.2 0.3 0.4

Normalized Freq. by 2pi: Normalized freq. by 2pi:

0.02
c 1.003 d

Magnitude ->(dB)
1.002 0.01
Magnitude ->

1.001
0
1
-0.01
0.999

0.998 -0.02
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Normalized freq. by 2pi: Normalized freq. by 2pi:

Fig. 5. Performance of 5-channel non-uniform filter bank (16, 16, 8, 4, and 2) designed using the proposed method. (a) Magnitude responses of the prototype filter in dB, (b)
magnitude responses of the analysis filters in dB, (c) variation of peak reconstruction error and (d) Variation of peak reconstruction error in dB.

Example-I: A 3-channel linear-phase non-uniform filter bank
with decimation factors 4, 4, and 2 is designed by the pro-
posed method. Design specifications for the prototype filter
are N + 1 =96, stopband attenuation (As ) = 80 dB, passband ripple
(Ap ) = 0.000521 dB, ωs = 0.51, and ωp = 0.41. The peak reconstruc-
tion error (PRE), computational time (CPU time) and number of
iterations (NOI) obtained are 2.8 × 10−3 , 0.826 s, and 11, respec-
tively (Fig. 3).
Example-II: A 4-channel linear-phase non-uniform filter bank
(8, 8, 4, 2) with the prototype filter design specifications: N + 1 =48,
As = 80 dB, Ap = 0.0005 dB, ωs = 0.61 and ωp = 0.41 is designed
with this method. In this case, PRE and computational time are
3.11 × 10−3 and 0.733 s, respectively, while NOI = 15 and the simu-
lation results is depicted in Fig. 4.
Example-III: A 5-channel NUFB with decimation factors 16, 16,
8, 4, and 2 and same design specifications as used in example II is
designed with the proposed method. The design results obtained in
this case are shown in Fig. 5. The proposed method is also exploited
for other design examples, and the simulation results obtained
are tabulated in Table 1. As it can be observed from the simula-
tion results, the proposed method gives better performance in all
aspects. The peak reconstruction error is reduced appreciably. The
average peak reconstruction error obtained in 3-channel and 4-
channel NUFBs is 4.9 × 10−3 , and 4.81 × 10−3 , respectively, while in
5-channel NUFBs is 4.85 × 10−3 . The computation time and number
of iterations required in the proposed method are extremely low
even if the number of filter taps is more. Hence, this technique can
be effectively utilized for designing non-uniform filter bank with
larger filter taps.
Results of the comparative studies made are given in Table 2,
which clearly show the superiority of the proposed methodol-
ogy over the exiting algorithms [28,38,39]. For this, non-uniform

Table 1
Performance of the proposed method for designing nonuniform filter bank.

Band (M) Filter taps (N) As (dB) PRE NOI CPU time (s)
−3
48 0.00050 3.11 × 10 15 0.733
66 0.00017 2.51 × 10−3 17 0.782
72 0.00016 4.91 × 10−3 16 0.846
Three band (4, 4, 2)
80 0.00077 8.11 × 10−3 17 0.923
90 0.00027 8.11 × 10−3 16 0.867
96 0.00051 2.80 × 10−3 11 0.826

48 0.00050 3.11 × 10−3 15 0.733

66 0.00017 5.70 × 10−3 9 0.702
72 0.00016 5.70 × 10−3 18 1.056
Four band (8, 8, 4, 2)
80 0.00077 7.40 × 10−3 16 0.923
90 0.00027 3.50 × 10−3 18 1.162
96 0.00051 3.50 × 10−3 15 1.157

48 0.00050 3.1 × 10−3 11 0.585

66 0.00017 5.70 × 10−3 9 0.686
72 0.00016 5.70 × 10−3 17 1.216
Five band (16, 16, 8, 4, 2)
80 0.00077 7.31 × 10−3 17 1.232
90 0.00027 3.72 × 10−3 18 1.211
96 0.00051 3.59 × 10−3 17 1.102
 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360 359

Table 2
Comparison of the proposed methodology with earlier published results.

Algorithms Band (M) Filter taps (N) As (dB) PRE NOI CPU time (s)

Algorithm in [39] Three band (4, 4, 2) 63 110 7.80 × 10−3

Algorithm in [38] Three band (4, 4, 2) 64 60 7.80 × 10−3

63 75 8.60 × 10−3 153 0.625

Algorithm in [28] Three band (4, 4, 2) 63 110 3.85 × 10−3 158 2.35
63 70 1.10 × 10−3 158 2.28

48 80 3.11 × 10−3 15 0.733

Three band (4, 4, 2) 66 80 2.51 × 10−3 17 0.782
72 80 4.91 × 10−3 16 0.846
Proposed method 48 80 3.11 × 10−3 15 0.733
Four band (8, 8, 4, 2)
66 80 5.70 × 10−3 9 0.702
96 80 3.59 × 10−3 17 1.102
Five band (16, 16, 8, 4, 2)
90 80 3.72 × 10−3 18 1.211

Table 3
Fidelity assessment parameters in the proposed algorithm.

Signal PRD MSE ME SNR

Speech, M, Eng.: M 45 0.970 × 10−1 1.327 × 10−11 7. 231 × 10−6 62.27

Speech, F, Eng.: M 45 1.035 × 10−1 4.316 × 10−11 1.016 × 10−5 59.61
Speech, L, Eng.: M 45 0.977 × 10−1 1.620 × 10−11 9.251 × 10−4 60.21
Speech, O, Eng.: M 45 0.935 × 10−1 1.358 × 10−11 8.029 × 10−4 60.44

filter banks with similar design specifications as published ear-
lier have been designed and compared. As, it can be seen that the
performances of the proposed method are significantly improved
as compared to the earlier known techniques in terms of peak
reconstruction error, computational time (CPU time) and number of
iterations. The percentage difference in peak reconstruction error
provided by this method relative to other methods [28,38,39] is
15.31%, 51.03%, and 51.03%, respectively. The computational time
and number of iterations are significantly reduced as in each iter-
ation only filter coefficients value at ω = /2 are computed and
compared with the ideal values. Therefore, this technique can be
successfully employed for various applications which are carried
out in real time and quasi real time.
In practice, almost all the real signals have uneven distribution of
energy in different bands or energy is dominantly concentrated in a
particular region of frequency. This uneven distribution of energy in
different bands provides the basis for signal or source compression.
The speech signal has also non-uniform frequency distribution.
Since speech coding is widely used in numerous applications such
as mobile satellite communications, cellular telephony, video tele-
conferencing systems and storage of speech for speech synthesis
and playback, or for the transmission of voice at a later time, hence
Origional Signal
1 verges very fast in low number of iterations. It can be effectively
Magnitude->
it is necessary to compress the speech signal. Here, the proposed
method has been used for subband coding of speech signal. For
this, 3-channel NUFB is designed with same design specifications
as in Example I, and is exploited for subband coding. Several fidelity
parameters such as percent root mean square difference (PRD),
mean square error (MSE), maximum error (ME), and signal to noise
ratio (SNR) discussed in [6,20,37–39] are computed and summa-
rized in Table 3. Fig. 6 shows the original ECG signals (MIT/BIH-100)
and its reconstructed version.
As it can be observed from Table 3, all the fidelity parameters
obtained with this methodology are much better than the accept-
able range in practice. Thus, it is evident that the proposed method
can effectively used for subband coding of the signal.
5. Conclusions
A new algorithm for the design of linear-phase non-uniform fil-
ter bank is presented. In this method, the passband frequency is
optimized so that the magnitude response of the prototype filter
is approximately close to 0.707 at quadrature frequency. A com-
parative study of the performance has also been carried out. From
the study, it is evidenced that the peak reconstruction error is less
as compared to earlier published result. It is also observed that
the proposed algorithm requires less computational time and con-

used for filters with larger taps used in audio coding.

0
References
-1 [1] G.J. Dolecek, Multirate Systems: Design and Applications, Idea Group of Pub-
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
lishing, Hershey, USA, 2002.
4
x 10 [2] P.P. Vaidyanathan, Multirate and Filter Banks, Prentice-Hall, Englewood Cliffs,
NJ, 1993.
Reconstructed Signal:
[3] S.S. Lin, S.S. Chan, K.M. Tsui, On the design of nearly-PR and PR FIR cosine mod-
1 ulated filter banks having approximate cosine-roll off transition band, IEEE
Magnitude:

Transactions on Circuits and System-II: Express Briefs 55 (6) (2008) 571–575.

[4] F. Cruz-Roldan, P. Martin-Martin, J. Saez-Landete, M. Blanco-Velasco, T. Sara-
0
maki, A fast windowing based technique exploiting spline function for
designing cosine modulated filter banks, IEEE Transaction on Circuits and
System-I 56 (1) (2009) 168–178.
-1
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 [5] S.W.A. Bergen, A design for cosine modulated filter banks using weighted con-
strained least squares filters, Digital Signal Processing 18 (2008) 282–290.
4
x 10 [6] A. Kumar, G.K. Singh, R.S. Anand, An improved closed form design method for
the cosine modulated filter banks using windowing technique, Applied Soft
Fig. 6. Subband coding of speech signal: sound ‘F’. Computing 11 (3) (2011) 3209–3217.
360 A. Kumar et al. / Applied Soft Computing 13 (2013) 353–360
[7] A. Jain, R. Saxena, S.C. Saxena, An improved and simplified design of cosine-
modulated pseudo-QMF filter banks, Digital Signal Processing 16 (1) (2006)
225–232.
[8] F. Cruz-Roldan, A.M. Bravo-Santos, P.M. Martin, R. Jimenez Martinez, Design
of multi-channel near perfect reconstruction transmultiplexers using cosine
modulated filter banks, Signal Processing 83 (1) (2003) 1079–1091.
[9] K. Mayyas, T. Aboulnasr, A fast exact weighted subband adaptive algorithm
and its application to mono and stereo acoustic echo cancellation, Journal of
the Franklin Institute 342 (3) (2005) 235–253.
[10] S.W.A. Berger, A. Antoniou, An efficient closed form design method for cosine
modulated filter banks using window function, Signal Processing 87 (2007)
811–823.
[11] J.-Jy. Shyu, S.-Ch. Pei, Y.-Da. Huang, An iterative approach for minimax design
of multidimensional quadrature mirror filters, Signal Processing 91 (2011)
1730–1740.
[12] X. Gao, Z. He, X.G. Xia, The theory and implementation of arbitrary length linear
phase cosine modulated filter bank, Signal Processing 80 (5) (2000) 889–896.
[13] O.G. Ibarra-Manzano, Y. Shkvarko, R. Jaime-Rivas, J.A. Andrade-Lucio, G.J. Dole-
cek, Optimization of perfect reconstruction of cosine modulated filter banks:
fast algorithm, Telecommunication and Radio Engineering 56 (12) (2001)
56–63.
[14] W.-S. Lu, T. Saramaki, R. Bregovic, Design of practically perfect reconstruction
cosine modulated filter banks: a second order cone programming approach,
IEEE Transactions on Circuits and System-I: Regular Paper 51 (1) (2004)
552–563.
[15] V. Harish, P. Sommen, K.M.M. Prabhu, Perfect reconstruction of uniform sam-
ples from K-th order non uniform samples, Signal Processing 91 (2011)
2677–2684.
[16] C.D. Creusere, S.K. Mitra, A simple method for designing high quality prototype
filters for M-band pseudo QMF banks, IEEE Transactions on Signal Processing
43 (1995) 1005–1007.
[17] F. Cruz-Roldan, P.A. Lopez, S.M. Bascon, S.S. Lawson, An efficient and sim-
ple method for designing prototype filters for cosine modulated pseudo QMF
banks, IEEE Signal Processing Letters 9 (1) (2002) 29–31.
[18] D.C. Chang, D.L. Lee, Prototype filter design for a cosine-modulated filter bank
transmultiplexers, in: Proceeding of IEEE International Conference APCCAS
2006, 2006, pp. 454–457.
[19] A. Datar, A. Jain, P.C. Sharma, Design of Kaiser window based optimized pro-
totype filter for cosine modulated filter bank, Signal Processing 90 (2010)
1742–1749.
[20] A. Kumar, G.K. Singh, R.S. Anand, A simple design method for the cosine mod-
ulated filter banks using weighted least square technique, Journal of Franklin
Institute 348 (1) (2011) 606–621.
[21] S. Salcedo-Sanz, F. Cruz-Roldàn, C. Heneghan, X. Yao, Evolutionary design of
digital filters with application to sub-band coding and data transmission, IEEE
Transactions on Signal Processing 55 (April (4)) (2007) 1193–1203.
[22] X.M. Xie, S.C. Chan, T.I. Yuk, Design of perfect-reconstruction nonuniform
recombination filter banks with flexible rational sampling factors, IEEE Trans-
actions on Circuits and Systems-I: Regular Paper 52 (9) (2005) 1965–1981.
[23] X.M. Xie, S.C. Chan, T.I. Yuk, Design of linear-phase recombination nonuni-
form filter banks, IEEE Transactions on Signal Processing 54 (7) (2006)
2806–2814.
[24] A. Eghbali, H. Johansson, P. Löwenborg, Reconfigurable nonuniform transmul-
tiplexers using uniform modulated filter banks, IEEE Transactions on Circuits
and Systems-I: Regular Paper 58 (3) (2011) 539–547.
[25] L. Liang, G. Shi, X. Xie, Nonuniform directional filter banks with arbitrary
frequency partitioning, IEEE Transactions on Image Processing 20 (1) (2011)
283–289.
[26] V.J. Manoj, E. Elias, Artificial bee colony algorithm for the design of multiplier-
less nonuniform filter bank transmultiplexers, Information Sciences (2011),
http://dx.doi.org/10.1016/j.ins.2011.02.023.
[27] E. Elias, P. Lowenborgb, H. Johanssonb, L. Wanhammar, Tree-structured IIR/FIR
uniform-band and octave-band filter banks with very low-complexity analysis
or synthesis filters, Signal Processing 83 (1) (2003) 1997–2009.
[28] R.K. Soni, A. Jain, R. Saxena, An optimized design of non-uniform filter bank
using Blackman Window family, International Journal of Signal and Image
Processing 1 (2010) 18–23.
[29] S. Das, A. Konar, A swarm intelligence approach to the synthesis of two-
dimensional IIR filters, Engineering Applications of Artificial Intelligence 20
(8) (2007) 1086–1096.
[30] A. Ghosh, R. Giri, A. Chowdhury, S. Das, A. Abraham, Two-channel quadra-
ture mirror bank filter design using a fitness-adaptive differential evolution
algorithm, in: NaBIC, 2010, pp. 634–641.
[31] S. Das, A. Konar, Two-dimensional IIR filter design with modern search heuris-
tics: a comparative study, International Journal of Computational Intelligence
and Applications 6 (3) (2006) 329–355.
[32] S.C. Chan, X.M. Xie, T.I. Yuk, Theory and design of a class of cosine-modulated
non-uniform filter bank, in: Proc. of IEEE Int. Conf. on Acoustics, Speech and
Signal Processing, vol. 1, 2000, pp. 504–507.
[33] J.S. Akkarakaran, P.P. Vaidyanathan, New results and open problems on nonuni-
form filter banks, in: Proc. of IEEE, ICASSP, vol. 3, 1999, pp. 1501–1504.
[34] G. Shi, X. Xie, W. Zhong, Recent advances and new design method in nonuni-
form filter banks, in: Proc. of IEEE Int. Conf. on Communications, Circuits and
Systems, vol. 1, 2006, pp. 211–215.
[35] V.S. Sheeba, E. Elias, Design of signal-adapted nonuniform filter banks using
tree structure, Proceeding of ISCAS 2006 (2006) 887–890.
[36] S. Nalbalwar, S.D. Joshi, R.K. Patney, On characterization of linear phase nonuni-
form filter banks, Signal Processing 88 (2008) 1980–1989.
[37] S.M.S. Jalaleddine, C.G. Hutchens, R.D. Strattan, W.A. Coberly, ECG data com-
pression techniques: a unified approach, IEEE Transactions on Biomedical
Engineering 37 (4) (1990) 329–343.
[38] J. Li, T.Q. Nguyen, S. Tantaratana, A simple design method for nonuniform
multirate filter banks, IEEE Transactions on Signal Processing 45 (1997)
2105–2109.
[39] X.M. Xie, X.Y. Chen, G.M. Shi, A simple design method of linear-phase nonuni-
form filter banks with integer decimation factors, in: Proc. of Int. Symp. on
Circuit and Systems, vol. 1, 2005, pp. 724–727.