Академический Документы
Профессиональный Документы
Культура Документы
2, FEBRUARY 2019
Abstract—This letter presents a novel design method for mul- based on optimization techniques present in the literature. While
tidimensional finite-impulse response (FlR) filters based on the in [9] are shown some examples of hardware implementation of
representation of the desired frequency response by using a Gabor FIR filters.
system generated by a Gaussian function. A direct synthesis of the
desired filter frequency response is obtained from standard design In this work we present a novel methodology for the design of
specifications. Further, we report some comparative tests with the multidimensional FIR filters, based on the representation of the
well known Parks–McClellan’s methodology in order to illustrate desired frequency response by means of a Gabor system gener-
the concrete advantages of the novel design approach that we have ated by the Gaussian function [10]–[12]. The proposed design
proposed. With such an approach, the design procedure turns out approach will therefore be termed Gaussian Based Approxima-
to be especially simple to implement, as it employs closed design
formulas which directly link filter coefficients to design specifica- tion or GBA for short. GBA is innovative under two respects:
tions, and moreover, affords a better pass-band flatness than any it does not fall under any of the conventional aforementioned
known design approach. It also yields stop-band attenuations and approaches, and attempts a direct synthesis of the desired filter
grants control over critical frequencies, which turn out being very frequency response, thus avoiding the costly iterative compu-
close to those obtainable with Parks–McClellan method. The en- tations of most conventional methods [13]. As a result of this
hanced performance of the 2-D filters thus obtained are well suited
for space applications. simplicity and effectiveness, the proposed GBA methodology
is especially suitable for on board digital signal processing in
Index Terms—Multidimensional FIR filter design, closed-form space applications [14]. For example in [15] was used an adap-
FIR tap coefficients, Gabor system, equispaced Gaussian functions.
tive Wiener filtering algorithm (2D-CSAWF) for the denois-
ing of satellite images contaminated with Gaussian noise. The
I. INTRODUCTION 2D-CSAWF is a two-dimensional adaptive FIR driven by an
optimization algorithm to self-adjust the filter coefficients. In
HE conventionally used approaches to 2D FIR filter de-
T sign fall under the following categories: the window-based
methods, the frequency sampling methods, the optimal filter de-
this application the use of GBA approach to calculate the filter
coefficients can be very effective because the GBA method is
very simple to implement and employs closed design fomulas,
sign and the transfomation methods. The first three are exten-
directly linking the filter coefficients to design specifications.
sions of 1D methods which generally do not explicitly optimize
The paper is structured as follows. In the first section the GBA
the desired filter specifications. Further, in the above approaches
design procedure for 2D FIR filters is described in detail for
not all filter specifications can be directly linked to design pa-
square and circular shaped filters (the extension to generic shape
rameters [1]–[5].
is straightforward). In the results section, the perfomance of the
Recently several authors proposed a number of design meth-
2D FIR filters designed following the proposed procedure are
ods. In [6]–[8] are presented three approaches to the 2D FIR
illustrated by some significant examples and compared, where
filter design, in which the design problem is formulated as an
applicable, with those yielded by the method of [2]. Furthermore
optimization problem. The methods described in [6], [7] re-
the proposed design procedure has been compared with the
quire the property of the quadrantally even symmetric of the
recent 2D FIR-FDC method described in [6]. Finally, the authors
filter impulse response in order to mitigate the computational
conclusions are drawn.
complexity, typical of the optimization problems. This peculiar-
ity limits the applicability of these design methods to the cases
of filters whose impulse response presents a similar symmetry.
II. DESCRIPTION OF THE GBA DESIGN PROCEDURE
Instead the design method presented in [8] is a 1D FIR filter de-
sign method that provide performances similar to the methods For the sake of simplicity we describe the GBA design pro-
cedure for the case of a 2D, square-shaped, low-pass FIR fil-
Manuscript received November 11, 2018; revised January 5, 2019; accepted ter. The extension to more than two dimensions and arbitrary
January 7, 2019. Date of publication January 9, 2019; date of current version shapes is easily practicable. ln particular a 2D rectangular filter
January 21, 2019. The associate editor coordinating the review of this manuscript
and approving it for publication was Dr. Demetrio Labate. (Corresponding is straightforward to accommodate along the lines of the treat-
author: Giacomo Capizzi.) ment expounded in the sequel. The circular shape case is briefly
The authors are with the Department of Electrical, Electronics and In- reported below.
formatics Engineering, University of Catania, 95125 Catania, Italy (e-mail:,
gcapizzi@diees.unict.it; glosciuto@dii.unict.it). ln the design of a 2D, square-shaped, low-pass filter with pass-
Digital Object Identifier 10.1109/LSP.2019.2891965 band and stop-band edges ωc and ωs respectively, the frequency
1070-9908 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
CAPIZZI AND SCIUTO: NOVEL 2-D FIR FILTER DESIGN METHODOLOGY BASED ON A GAUSSIAN-BASED APPROXIMATION 363
c
definition domain can be constructed as a linear combination of N1 = (6)
bidimensional Gaussian functions. I.e., we let, for some appro- Δμ
priate domain D over which the integer indices i, j range: It should be noted that this amounts to using a grid as the
domain D occurring in the expression (2) for the synthesized
Hr (ωx , ωy )
frequency response Hr (ωx , ωy ); this is a result of the filter
= ci,j Gσ x i ,σ y i (ωx , ωy ) δ(ωx − μx i, ωy − μy i ) (2) shape being a square. Other filter geometries would be reflected
i,j ∈D
by a change in D.
Finally, the coefficients ci,j in (2) are set to a common value
denotes the convolution operator, δ(ωx − μx i , ωy − μy i ) is a 1/M, chosen so as to normalize the filter gain. M depends on the
bidimensional Dirac delta centred upon the (μx i , μy i ) point in design specifications through δ, σ and N1 .
the R2 space ranged over by (ωx , ωy ), and
1
N1
M= 1 + 2 (1 − δ)4∗i
2
the coefficients. I.e., the indices i and j in the double sum are
constrained to fall within a circle.
⎛
1 ⎜
N1
M= ⎜1 + 2 (1 − δ)4∗i
2
2πσ 2 ⎝
i=1
⎞
N1
N1
2⎟
(1 − δ)4∗i (1 − δ)4∗j ⎟
2 2
+2 (1 − δ)4∗j + 4 ⎠
j =1 i,j=1
i 2 +j 2 ≤N 12
(9)
1 (nσ)2 (nσ)2
h(n, m) = exp − · exp −
4π 2 M 2 2
⎛ Fig. 2. Frequency response for a 2D, 11 × 11 order, square-shaped, low-pass
FIR filter: PM design.
⎜
N1
N1
×⎜
⎝1 + 2 cos (n · μx i ) + 2 cos (m · μy j )
i=1 j =1
⎞
N1
⎟
+4 cos (n · μx i ) · cos (m · μy j )⎟
⎠
i,j=1
i 2 +j 2 ≤N 12
(10)
Gσ x ,σ y ,σ z (ωx , ωy , ωy )
ωx2 ωy2 ω2
= A · exp − 2 · exp − 2 · exp − z2 (11)
2σx 2σy 2σz
V. CONCLUSION
ln this work we have described the GBA 2D FIR filter design
Fig. 5. Frequency response for a 2D, 21 × 21 order, square-shaped, low-pass
FIR filter: GBA 2-D design.
methodology, an approach based on Gaussian approximation.
Filter shapes different from the square considered above can
be dealt with, by a simple and direct extension of the present
treatment.
Unlike other approaches, the GBA design procedure turn out
to be very simple to implement, especially since it employs
closed (rather than iterative) design fomulas, directly linking fil-
ter coefficients to design specifications. Furthermore, it exhibits
a very good control over the precision of the critical frequencies
of the synthesized filter. As a result, it succeeds in implement-
ing 2D, real coefficients FIR filters with better efficiency and
performance than other known methods.
At least two advantages of GBA design deserve to be high-
lighted as far as space comunications are concerned.
r Since the design procedure is fast and direct, it is easy for
an application to reconfigure a filter at run-time, in order
to adapt to changing (e.g. environmental) requirements.
r Experiments carried out show that GBA design generally
Fig. 6. Frequency response for a 2D, 21 × 21 order, circular-shaped, low-pass
FIR filter designed with the GBA method. affords the desired performance figures with a lower order
filter than other methods.
While these properties are generally desirable, they are def-
An example of design of circular 2D FIR filter with the pro- initely crucial in instances, like space applications, where the
posed method is illustrated in Fig. 6. Filter performances are the operation environment is particularly hard and computing time
same as in the square case. and storage space are at a premium.
IV. DISCUSSION
REFERENCES
The comparisons has been made by considering different [1] Y. Jiajing, N. Yardi, and W. K. Jenkins, “Computationally efficient 2-D
filter lengths (11 × 11 for low order filter, 21 × 21 for medium adaptive filters based on essentially separable FIR structures,” in Proc.
order filter, 55 × 55 for higth order filter) and different transition IEEE 59th Int. Midwest Symp. Circuits Syst., Oct. 2016, pp. 1–4.
[2] J. McClellen, “The design of two-dimensional digital filters by trans-
bands (0.1, 0.04 and 0.01, respectively). formations,” in Proc. 7th Annu. Princeton Conf. Inf. Sci. Syst., 1973,
For a low filter order and a large transition bands the two pp. 247–251.
design methods are comparable with respect to the stop-band [3] U. K. Shahanas, T. Bindima, and E. Elias, “Design of reconfigurable direc-
tional filters using McClellan transformation and spectral approximation
attenuation while the PM method yields a slightly smaller rip- in 2D space,” in Proc. IEEE 30th Can. Conf. Elect. Comput. Eng., Apr.
ple. With medium transition bands the two design methods are 2017, pp. 1–4.
comparable while for narrow transition bands (lesser than 0.02) [4] B. Dumitrescu, Design of FIR Filters. Cham, Switzerland: Springer, 2017,
pp. 149–177.
PM fails to converge, whereas GBA satisfies the specifications [5] S. Ghosh, P. Nair, and K. N. Chaudhury, “Optimized Fourier bilateral
in an acceptable way. filtering,” IEEE Signal Process. Lett., vol. 25, no. 10, pp. 1555–1559, Oct.
For a medium filter order and transition bands lesser than 0.05 2018.
[6] A. Aggarwal, M. Kumar, T. K. Rawat, and D. K. Upadhyay, “Optimal
and greater than 0.03, the two design methods are comparable. design of 2D FIR filters with quadrantally symmetric properties using
Finally for an high filter order the PM method fails to con- fractional derivative constraints,” Circuits, Syst., Signal Process., vol. 35,
verge when the transition bands are very narrow, whereas the no. 6, pp. 2213–2257, 2016.
366 IEEE SIGNAL PROCESSING LETTERS, VOL. 26, NO. 2, FEBRUARY 2019
[7] A. Aggarwal, M. Kumar, and T. Rawat, “Design of 2-dimensional FIR [12] G. Capizzi, S. Coco, G. L. Sciuto, and C. Napoli, “A new iterative FIR filter
filters with quadrantally symmetric properties using the 2-D L1 -method,” design approach using a Gaussian approximation,” IEEE Signal Process.
IET Signal Process., 2018. Lett., vol. 25, no. 11, pp. 1615–1619, Nov. 2018.
[8] L. M. San-José-Revuelta and J. I. Arribas, “A new approach for the design [13] M. Sharma et al., “A review of comparison of various linear phase FIR
of digital frequency selective FIR filters using an FPA-based algorithm,” filter algorithms to design an optimum filter,” i-Manager’s J. Digit. Signal
Expert Syst. Appl., vol. 106, pp. 92–106, 2018. Process., vol. 5, no. 1, pp. 39–45, 2017.
[9] V. Gupta, D. Mohapatra, A. Raghunathan, and K. Roy, “Low-power digital [14] Z. Lin, M. Lin, J. Ouyang, W. Zhu, and S. Chatzinotas, “Beamforming
signal processing using approximate adders,” IEEE Trans. Comput.-Aided for secure wireless information and power transfer in terrestrial networks
Des. Integr. Circuits Syst., vol. 32, no. 1, pp. 124–137, Jan. 2013. coexisting with satellite networks,” IEEE Signal Process. Lett., vol. 25,
[10] O. Christensen, A. J. E. M. Janssen, H. O. Kim, and R. Y. Kim, “Ap- no. 8, pp. 1166–1170, Aug. 2018.
proximately dual Gabor frames and almost perfect reconstruction based [15] S. Suresh and S. Lal, “Two-dimensional CS adaptive FIR Wiener filter-
on a class of window functions,” Adv. Comput. Math., vol. 44, no. 5, ing algorithm for the denoising of satellite images,” IEEE J. Sel. Top-
pp. 1519–1535, Oct. 2018. ics Appl. Earth Observ. Remote Sens., vol. 10, no. 12, pp. 5245–5257,
[11] O. Christensen, et al., An Introduction to Frames and Riesz Bases. New Dec. 2017.
York, NY, USA: Springer, 2016.