You are on page 1of 11

Design of Time-Frequency Localized Filter Bank

Using Modified Particle Swarm Optimization

Swati P. Madhe1 and Amol D. Rahulkar2 Raghunath S. Holambe2


1
Department of Instrumentation and Control,
Cummins College of Engineering for Women, Pune-411052, India
swatimadhe@gmail.com,
2
Department of Electrical and Electronics Engineering,
National Institute of Technology, Goa-403401,India
3
Department of Instrumentation Engineering,
SGGS Institute of Engineering and Technology, Nanded-431606, India

Abstract. This paper presents a design of an optimized time-frequency


localized, perfect reconstruction filter bank (FB) based on modified Par-
ticle Swarm Optimization (PSO). The PSO scheme is modified to add
the constraints as vanishing moments (VMs) and perfect reconstruction
(PR) during the design of the FB. First, VMs are imposed in order to
initialize the coefficients of linear-phase, low-pass-analysis filter. Next,
the initialized coefficients, satisfying VM constraint are used in modified
PSO (MPSO) scheme to obtain optimized time-frequency localized low-
pass-analysis filter. Similarly, a linear-phase, low-pass-synthesis filter is
designed by adding PR and VM constraints. The proposed FBs are illus-
trated with numerical examples and their performances are validated by
comparing their time-frequency localization and frequency band errors
with those of existing FBs. The proposed MPSO scheme minimizes the
frequency band errors and optimizes the time-frequency localization of
FBs. This simultaneous time-frequency localization is useful in extract-
ing most effective features from the signal.

Keywords: filter bank, Particle Swarm Optimization, perfect recon-


struction, time-frequency localization, vanishing moments

1 Introduction
Wavelets and filter banks (FBs) have revolutionized multirate signal process-
ing applications as it provides more precise information about signal data than
other signal analysis techniques (Vaidyanathan, 1992), (Vetterli and Kovacevic,
1995). FBs are designed by considering various properties of filters such as en-
ergy compaction, flatness, regularity, orthogonality, linear phase, pass-band and
stop-band errors and frequency selectivity depending on applications. In this pa-
per, we are focusing time-frequency localization as the design parameter because
it gives better joint time-frequency analysis of a signal (Daubechies, 1992). In lit-
erature, two channel FB based on time-frequency localization is designed mainly
2 Swati P. Madhe et al.

using polynomial factorization, lifting scheme, and complementary filter design


method. Morris et al. (Morris and Peravali, 1997) designed orthogonal time-
frequency localized FBs using optimization of the lattice parameters. Sharma
et al. (Sharma et al., 2010) have designed time-frequency localized FB using
two-parameter parametrization technique. However, these techniques (Morris
and Peravali, 1997), (Sharma et al., 2010) do not guarantee optimal solutions.
Tay (Tay, 1999) has designed a special class of linear phase biorthogonal fil-
ter banks called half-band pair filter banks (HBPF) by employing a measure
called balanced-uncertainty metric, which is a weighted summation of time and
frequency variances of the filter to be designed. The author uses Parametric
Bernstein Polynomial to optimize filter coefficients. But various parameters and
constants are involved in this approach. Eigenfilter-based approach was first used
by Patil B. D. et al. for filter bank design (Patil, Patwardhan, and Gadre, 2008)
to minimize the pass-band and stop-band errors. Sharma et al. (Sharma, Gadre,
and Porwal, 2015) have designed optimal time-frequency localized biorthogonal
filter bank using this eigenfilter approach. They have used a convex combination
of Gabor’s uncertainty-based time and frequency variances of the filters as an
optimality criterion. Rafi et al. (Rafi, Kumar, and Singh, 2013) suggested the ef-
fectiveness of Particle Swarm Optimization (PSO) during the design of Multirate
FB. In their design, the objective function was linear combination of pass-band
and stop-band errors. This design is computationally simple but these FB are
not regular and do not address the issue of optimal time-frequency localization.
Moreover this design has not considered the constraints in optimization.
PSO is a computational intelligence-based evolutionary algorithm used as a
robust tool for solving the linear and non-linear equations (Kennedy and Eber-
hart, 1995). In this paper, we have designed optimal filter bank using PSO
scheme instead of parametrization or eigenfilter-based approach because of the
following reasons (Del Valle et al., 2008):

1. Easier to implement.
2. Requires fewer parameters to adjust.
3. More efficient in providing the diversity of solution to set global solution.
4. Used in multiobjective and constrained optimization environment (Banks,
Vincent, and Anyakoha, 2008).

With this aspect, PSO scheme is used for the design of two channel, bi-orthogonal
FB. The optimality criteria is time-frequency localization. In the proposed de-
sign, regularity and perfect reconstruction (PR) constraints are added in op-
timization problem by modifying PSO scheme (MPSO). The performance of
proposed FBs is validated by comparing with existing FBs designed in (Sharma,
Gadre, and Porwal, 2015) and (Rafi, Kumar, and Singh, 2013). It is observed
that, performance of the FBs is superior than others in terms of time-frequency
localization and frequency band errors.
Rest of the paper is organized as follows: Section 2 explains the PSO and
MPSO schemes as it applies to the design of linear phase filters with constraints.
The formulation of constraints and objective function are explained in Section
Filter Bank Design Using Modified PSO 3

3. Section 4 explained the detail design methodology using MPSO. Several de-
sign examples and the comparative study have been included in Section 5. The
conclusion is given in Section 6.

2 Proposed Modified Particle Swarm Optimization


Particle Swarm Optimization is a population based stochastic optimization tech-
nique originated by Kennedy and Eberhart (Kennedy and Eberhart, 1995). De-
velopment of PSO was based on the concepts and rules that methodize socially
organized populations in nature, such as bird flocks, fish schools, and animal
herds. PSO relies on the exchange of information between individuals, called
particles, and the population, called swarm. Particles tend to move towards its
local best position Pbest found by them so far. They also keep the track of global
best Gbest solution, the best (shortest) path found by any particle at particular
instance. Each particle is correlated with a velocity, through which it gets moved
faster towards local and global best path, the position in ‘n′ dimension space
and the current position of particle with respect to Gbest and Pbest . Velocity of
particles at iteration i + 1 can be calculated using velocity and particle position
at previous iteration. Velocity and position of the particle can be expressed as
given in (Kennedy and Eberhart, 1995):

→ −

V (i + 1) = w V (i) + c1 ϕ1 (Pbest − P (i)) + c2 ϕ2 (Gbest − P (i)) (1)
and

→ −
→ −

P (i + 1) = P (i) + V (i + 1) (2)
where w is inertia weight factor ranging linearly from 0 to 1, c1 and c2 are called
cognitive and social acceleration factors, respectively. ϕ1 and ϕ2 are uniformly
distributed random numbers having range between 0 to 1. This PSO scheme is
modified (MPSO) by adding the constraints in the optimization problem. There
are numerous methods available in literature for handling constraints. However,
we have used simple method to handle the constraints using PSO used in (Hu
and Eberhart, 2002). In this, the particles are initialized and the constraints
are checked. If it satisfies the constraint, they are considered to generate the
swarm for optimization. Hence, each particle searches whole space but only keeps
tracking the values which satisfies constraints, in order to find the optimum
solution. This initialization of particles continues till it satisfies the constraints
of the objective function. During the design of time-frequency localized PR FB,
the objective function is convex combination of time and frequency variances,
the constraints are vanishing moment (VM) and PR.

3 Problem Formulation
Consider two channel FB system architecture as shown in Figure 1 to formulate
objective function and the constraints for FB design.
Here X(z) is the signal to be decomposed and Y (z) is the signal to be re-
constructed. The filters H0 (z) and H1 (z) are analysis low-pass filter (LPF) and
high-pass filter (HPF) respectively, and G0 (z) and G1 (z) are the synthesis LPF
and HPF respectively.
4 Swati P. Madhe et al.

Analysis Side Synthesis Side

H0 (z) 2 2 G0 (z)

X(z) Y (z)

H1 (z) 2 2 G1 (z)

Fig. 1: Two-channel 1-dimensional filter bank

3.1 VM Constraint
In certain applications such as pattern recognition regularity of FB improves
the accuracy of the system (Tang, 2000). The filters need to obey regularity for
the iterative convolution to converge (Vaidyanathan, 1992). This regularity or
smoothness can be added as a constraint in the PSO scheme by using VMs. Let
the filters h0 (n) and g0 (n) have 2VA and 2VS , VMs respectively. The regularity
can be forced by adding 2VA zeros at z = −1 for LPF. The regularity constraint
of LPF having order of 2N , given in (Sharma, Gadre, and Porwal, 2015) can be
expressed as ,

X
N
nK (−1)n h0 (n) = 0, for K = 0, 1, 2, . . . , 2VA − 1 (3)
n=−N

With this, VM constraint can be written in linear form as:

Vh0 = 0 (4)

where, the vector h0 contains filter coefficients h0 (n) and is defined as

h0 = [h0 (0) h0 (1) . . . h0 (N )]T (5)

V is a matrix of size VA × (N + 1) formed using equality (3). The same VM


constraint can be considered during the design of synthesis LPF.

3.2 PR constraint
The output of two channel FB shown in Figure 1, is expressed as (Vaidyanathan,
1992),
Y (z) = τ0 (z)X(z) + τ1 (z)X(−z) (6)
where τ1 (z) is called alias system function and has to be equal to zero to make
the system linear shift invariant (LSI), and can be expressed as:

1
τ1 (z) = {H0 (−z)G0 (z) + H1 (−z)G1 (z)} (7)
2
Filter Bank Design Using Modified PSO 5

Therefore, for τ1 (z) = 0,

H1 (z) = z −1 G0 (−z) and G1 (z) = zH0 (−z). (8)

Next, PR system function, τ0 (z) is given as:

τ0 (z) = {H0 (z)G0 (z) + H1 (z)G1 (z)} (9)

In order to satisfy PR property,

τ0 (z) = C0 z −D (10)

where, C0 is a constant and D is the delay. Assuming C0 = 2, D = 0 and using


conditions in (8), the filter bank can be written as:

H0 (z)G0 (z) + H0 (−z)G0 (−z) = 2 (11)

where, H0 (z)G0 (z) is called product filter P (z) and it is expressed as:

P (z) = H0 (z)G0 (z) (12)

P (z) is a symmetric (zero phase) polynomial in z, whose coefficients correspond-


ing to even powers of z are zero, except for origin which carries a coefficient 1
after normalization. Hence, the design of two-channel PR FB reduces to the
design of the half band filter P (z). In complementary design method, a valid
analysis filter is first designed independently. For given analysis filter, the syn-
thesis filter is designed by imposing PR condition. So for PR, after designing
the analysis LPF, h0 (n), the complementary synthesis LPF g0 (n) of order 2M
should satisfy the condition in linear form as,

Pg0 = 0 (13)

where, vector g0 contains filter coefficients of g0 (n) and is defined as,

g0 = [g0 (0) g0 (1) . . . g0 (M )]T (14)

P is a real matrix formed using coefficients of the analysis LPF h0 (n). Thus,
PR FB using MPSO, (4) is used as a constraint for VM and (13) is used as a
constraint for PR.

3.3 Formulation of Objective Function for Proposed MPSO Scheme


Let h(n) be the impulse response of a linear phase, low pass filter of the order
2N . The frequency response of the filter is expressed as:

X
N
H(w) = h(0) + 2h(n) cos(wn) = aT c(w) (15)
n=1

where vectors a, c(w) ∈ RN+1 are defined as,

a = [h(0) h(1) h(2)... h(N − 1) h(N )]T (16)


6 Swati P. Madhe et al.

c(w) = [1 2 cos(2w) ... 2 cos(N w)]T (17)


The time and frequency variance of h(n) in l2 (Z) normalized to unit energy are
defined as (Tay, 1999),
X Z
1 π ′
σn 2 = n2 |h(n)|2 = |H (w)|2 dw (18)
π 0
Z
1 π 2
σw 2 = w |H(w)|2 dw (19)
π 0
where H ′ (w) is the derivative of frequency response of h(n) and can be written
as:
d d
H ′ (w) = H(w) = aT c(w) (20)
dw dw
Let vector f is obtained by differentiating the vector c(w) and is given as,
d
f(w) = c(w) = [0 − 2 sin(w) · · · − 2 sin(N w)]T (21)
dw
Using (20) and (21), the time variance can be written as:
Z
1 π T
σn 2 = a f(w)fT (w)a dw = aT Ta (22)
π 0
where T is a real, symmetric, positive-definite matrix of size (N + 1) × (N + 1)
and is given by, Z
1 π
T= f(w)fT (w)dw (23)
π 0
In the same way using (15), the frequency variance can be written as:
Z
1 π T 2
σw 2 = a w c(w)cT (w)a dw = aT Fa (24)
π 0
where F is also a real, symmetric, positive-definite matrix of size (N +1)×(N +1)
and is as follows, Z
1 π 2
F= w c(w)cT (w)dw (25)
π 0
Considering time and frequency variance, time-frequency localized objective func-
tion φ which is to be minimized can be written as a convex combination of the
variances of the filter as (Sharma, Gadre, and Porwal, 2015),

φ = α σn 2 + (1 − α) σw 2 , α ∈ [0, 1] (26)

where α is the trade-off factor between the variances. The equality (26) can be
expressed as a convex quadratic form as:

φ = αaT Ta + (1 − α)aT Fa
= aT {αT + (1 − α)F} a = aT Ra (27)
Filter Bank Design Using Modified PSO 7

where, R = αT + (1 − α)F is a real, symmetric, positive-definite matrix of order


of (N + 1) × (N + 1) and a is a unit norm vector associated with the filter
coefficients. Thus in the proposed MPSO scheme, the optimization problem is
casted as minimize objective function φ, subjected to the VM constraint (4) and
PR constraint (13).

4 Design Methodology
The parameters required for MPSO scheme are initialized with values as, mini-
mum error = 1e-10, swarm size = 200, range for position and velocity parameters=[-
1,1], c1 = c2 = 2 and iterations = 1000. The design of optimized time-frequency
localization PR FB consists of designing filters h0 (n) and g0 (n). The detailed
procedure is as follows:
Design of analysis filter h0 (n) of order 2N
1. Specify design parameters of the filter as,
– particle size = N
– vanishing moment = VA
– trade off factor = α.
2. The constrained minimization problem to design the analysis LPF h0 (n) can
be expressed as,
minimize φ = hT 0 Rh0
h0 (28)
subject to Vh0 = 0
where φ is the objective function as given in (26) subjected to the VM
constraint defined in(4). R ∈ R(N +1)×(N +1) is a real, symmetric, positive-
definite matrix calculated using (27) and h0 is used to find coefficients of
analysis LPF h0 (n).
3. Initialize h0 as particles and check the VM constraint. Repeat initialization
till h0 satisfies the VM constraint.
4. Randomly generate the swarm with these particle position and velocity vec-
tor using the initialized parameters.
5. Compute new velocities and new particle positions by (1),(2) for all particles.
6. Find the value of objective function φ for each particle positions.
7. If the value of φ is less than Pbest then go to next step, otherwise increase
the iterations counter by 1 and go to step (9).
8. Check the iteration counter limit and the minimum error condition, if any
one of the condition is satisfied directly follow step (10).
9. Assign Pbest with present position and update Gbest
10. The optimized vector h0 = Gbest . The vector h0 contains N + 1 coefficients
of the desired analysis LPF filter h0 (n) for 0 ≤ n ≤ N . The remaining N
coefficients can be deduced by symmetry.
Design of synthesis LPF g0 (n) of order 2M
1. Specify design parameters of the filter as:
– particle size = M
– vanishing moment = VS
8 Swati P. Madhe et al.

– trade off factor = α.


2. For the designed analysis LPF, the design of the synthesis LPF can be for-
mulated as a constrained minimization problem,

minimize φ = gT
0 Rg0
g0
(29)
subject to Pg0 = 0, Vg0 = 0

where φ is the objective function as given in (26) subjected to the VM


constraint and PR constraint. Matrix R ∈ R(Q+1)×(Q+1) is calculated using
(27) and g0 is used to find coefficients of synthesis LPF g0 (n).
3. Initialize g0 (n) as particles and check VM constraint as defined in (4) along
with VM constraint as defined in (13). Repeat initialization till g0 (n) satisfies
both the constraints.
4. Follow the previous steps from (4)-(9).
5. The optimized vector g0 = Gbest is obtained using MPSO scheme. The
vector g0 contains M + 1 coefficients of the desired analysis LPF g0 (n) for
0 ≤ n ≤ M . The remaining M coefficients can be deduced by symmetry.

5 Design Examples and Results


5.1 Design Examples
This section illustrates the proposed MPSO scheme with few numerical ex-
amples. We have compared the proposed work with related FBs designed in
(Sharma, Gadre, and Porwal, 2015) and (Rafi, Kumar, and Singh, 2013). The
performance is compared using Time-frequency localization (∆) and Frequency
Band errors (E).

Example 1: In this example, 7-tap filter of CDF-9/7 FB is considered as


analysis LPF. We have designed the complementary synthesis filter of length
13 by choosing the design parameters as M = 6, VS = 1, and α = 0.5. FB
in (Sharma, Gadre, and Porwal, 2015) has used same design parameters. QMF
bank is designed using the PSO scheme in (Rafi, Kumar, and Singh, 2013) with
the parameters as, length = 13, pass-band frequency = 0.4π and stop-band fre-
quency= 0.6π. The comparison of frequency responses of analysis and synthesis
of proposed filter with existing methods and (Sharma, Gadre, and Porwal, 2015)
are shown in figures 2(a) and 2(b) respectively. As we have taken standard 7 tap
analysis LPF for designing the FB, the frequency responses are overlapping in
figure 2(a). It is observed from figure 2(b), that the designed synthesis filter re-
sponse is having less frequency band errors as compared to (Rafi, Kumar, and
Singh, 2013) and is more localized as compare to (Sharma, Gadre, and Porwal,
2015).

Example 2: In this example 11/25 FB with two VMs for analysis and synthesis
filters each is designed. So the design parameters are as follows: N = 5, VA = 2,
and α = 5/6 for analysis filter and M = 12, VS = 2, and α = 10/11 for synthesis
filter. The value of trade-off factor α is different for analysis and synthesis filters,
Filter Bank Design Using Modified PSO 9

Plot of Analysis LPF Plot of Synthesis LPF


1 1
Proposed approach Proposed approach
Rafi approach [11] Rafi approach [11]
0.8 Sharma approach [9] 0.8 Sharma approach [9]

Magnitude

Magnitude
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Frequency (×π rad/samples) Normalized Frequency (×π rad/samples)

Fig. 2: Magnitude plots of (a)Analysis-low-pass filter and (b)Synthesis-low-pass


filter of Example 1

which indicates that time and frequency localization is not given same weightage
in the objective function. For the QMF bank using PSO algorithm in (Rafi,
Kumar, and Singh, 2013) we have chosen length as 25, pass-band frequency =
0.4π and stop-band frequency = 0.6π. Figures 3(a) and 3(b) shows the frequency
responses of the designed analysis and synthesis LPF using the proposed method
along with other methods in (Rafi, Kumar, and Singh, 2013) and (Sharma,
Gadre, and Porwal, 2015), respectively. It is observed from figure 3, that the
designed filter responses are having less frequency band errors as compared to
(Rafi, Kumar, and Singh, 2013) and are more localized as compare to (Sharma,
Gadre, and Porwal, 2015).

Plot of Analysis LPF Plot of Synthesis LPF


1 1
Proposed approach Proposed approach
Rafi approach [11] Rafi approach [11]
0.8 Sharma approach [9] 0.8
Sharma approach [9]
Magnitude

Magnitude

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized Frequency (×π rad/samples) Normalized Frequency (×π rad/samples)

Fig. 3: Magnitude plots of (a)Analysis-low-pass filter and (b)Synthesis-low-pass


filter of Example 2

5.2 Comparison with related filter banks


The proposed MPSO scheme of optimized time-frequency localization FB is
compared with the FB designed in (Sharma, Gadre, and Porwal, 2015) and
(Rafi, Kumar, and Singh, 2013). The comparison of filter properties of Example
1 and Example 2 are shown in Table 3.
10 Swati P. Madhe et al.

Table 1: Relative properties of proposed MPSO scheme with respect to other


algorithms of Example 1 and 2
FB by FB by
Property Measures Proposed FB
M Sharma et al Rafi et al
Example 1 2 1 2 1 2
Analysis LPF
Time-frequency localization, ∆ 0.356 0.292 3.339 1.031 0.3268 0.219
Frequency band error, E 19.666 41.759 0.828 0.0071 19.666 15.323
Synthesis LPF
Time-frequency localization, ∆ 0.564 0.490 3.339 1.031 0.8005 0.451
Frequency band error, E 19.349 10.306 0.828 0.0071 20.237 8.209

The FB designed using methodology in (Sharma, Gadre, and Porwal, 2015)


gives better time-frequncy localization but the frequency band error is more
than the FB designed using methodology in (Rafi, Kumar, and Singh, 2013).
On the contrary, frequency band error in (Rafi, Kumar, and Singh, 2013) are
less than in (Sharma, Gadre, and Porwal, 2015) but these FB are not time-
frequency localized. Thus, it is observed from Table 1 that, the time-frequency
localization of analysis and synthesis filters are better than the FB designed in
(Rafi, Kumar, and Singh, 2013). Frequency band errors are less than the FB
designed in (Sharma, Gadre, and Porwal, 2015).

6 Conclusion
This paper presents a new scheme to design two-channel optimized time-frequency
localized PR FB based on artificial intelligence technique, Particle Swarm Op-
timization (PSO). The traditional PSO scheme is modified to impose vanishing
moment (VM) and Perfect reconstruction (PR) constraints. The illustrated ex-
amples show that the designed filters give better frequency response than existing
filters. In addition, the proposed filter bank gives more regularity and better fre-
quency selectivity as compared to existing filter banks. It is evident that the
proposed MPSO scheme can be a good alternative approach for designing time-
frequency localized PR FB as it is simple and easy to implement. In future, this
PSO approach can be extended in multiobjective optimization environment to
enhance FB properties simultaneously.
Bibliography

Banks, Alec, Jonathan Vincent, and Chukwudi Anyakoha. 2008. “A review of


particle swarm optimization. Part II: hybridisation, combinatorial, multicrite-
ria and constrained optimization, and indicative applications.” Natural Com-
puting 7 (1): 109–124.
Daubechies, Ingrid. 1992. Ten lectures on wavelets. SIAM.
Del Valle, Yamille, Ganesh Kumar Venayagamoorthy, Salman Mohagheghi,
Jean-Carlos Hernandez, and Ronald G Harley. 2008. “Particle swarm opti-
mization: basic concepts, variants and applications in power systems.” IEEE
Transactions on evolutionary computation 12 (2): 171–195.
Hu, Xiaohui, and Russell Eberhart. 2002. “Solving constrained nonlinear op-
timization problems with particle swarm optimization.” In Proceedings of
the sixth world multiconference on systemics, cybernetics and informatics,
Vol. 5203–206. Citeseer.
Kennedy, J, and R Eberhart. 1995. “Particle Swarm Optimization. proced-
dings of IEEE International Conference on Neural Networks,(pp. 1942–1948).”
Perth, WA, Australia .
Morris, Joel M, and Ravindra Peravali. 1997. “Optimum duration discrete-time
wavelets.” Optical Engineering 36 (4): 1241–1248.
Patil, Bhushan D, Pushkar G Patwardhan, and Vikram M Gadre. 2008. “Eigen-
filter approach to the design of one-dimensional and multidimensional two-
channel linear-phase FIR perfect reconstruction filter banks.” IEEE Transac-
tions on Circuits and Systems I: Regular Papers 55 (11): 3542–3551.
Rafi, SM, A Kumar, and GK Singh. 2013. “An improved particle swarm op-
timization method for multirate filter bank design.” Journal of the Franklin
institute 350 (4): 757–769.
Sharma, Manish, Vikram M Gadre, and Saket Porwal. 2015. “An eigenfilter-
based approach to the design of time-frequency localization optimized two-
channel linear phase biorthogonal filter banks.” Circuits, Systems, and Signal
Processing 34 (3): 931–959.
Sharma, Manish, Ritesh Kolte, Pushkar Patwardhan, and Vikram Gadre. 2010.
“Time-frequency localization optimized biorthogonal wavelets.” In 2010 In-
ternational Conference on Signal Processing and Communications (SPCOM),
1–5. IEEE.
Tang, Yuan Yan. 2000. Wavelet theory and its application to pattern recognition.
Vol. 36. World Scientific.
Tay, David BH. 1999. “Balanced-uncertainty optimized wavelet filters with pre-
scribed regularity.” In Circuits and Systems, 1999. ISCAS’99. Proceedings of
the 1999 IEEE International Symposium on, Vol. 3532–535. IEEE.
Vaidyanathan, P. P. 1992. ’Multirate Systems and Filter Banks. 2nd ed. Engle-
wood Cliffs, NJ, Prentice Hall.
Vetterli, M., and J. Kovacevic. 1995. ’Multirate Systems and Filter Banks. 2nd
ed. Englewood Cliffs, NJ, Prentice Hall.

You might also like