Sunteți pe pagina 1din 9

INTERNATIONAL JOURNAL OF ELECTRONICS AND COMMUNICATION ENGINEERING & TECHNOLOGY (IJECET)

ISSN 0976 6464(Print) ISSN 0976 6472(Online) Volume 4, Issue 2, March April, 2013, pp. 315-323 IAEME: www.iaeme.com/ijecet.asp Journal Impact Factor (2013): 5.8896 (Calculated by GISI) www.jifactor.com

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

IJECET
IAEME

DESIGN OF SIGNAL MATCHED NONUNIFORM SYNTHESIS FILTER BANK: A NOVEL APPROACH


Dr. Sanjay Nalbalwar Head, Department of Electronics & Telecommunication Engineering Dr. Babasaheb Ambedkar Technological University, Lonere, M.S., India

ABSTRACT A novel approach is proposed for the design of non-uniform synthesis filter bank matched to a signal. To estimate the parameters of the synthesis filters, the output of the synthesis side is taken as delayed version of the input signal. The filter coefficients are obtained by minimizing the square error between corresponding filter output and delayed version of the input signal. This approach is not only simple but leads to FIR synthesis filters where order can be chosen according to desired level of accuracy for perfect reconstruction. In fact one could choose the synthesis filters FIR and design the system. To design synthesis filter bank for given statistics case MMSE approach is used and for that of given data case least square approach is used. The proposed approach is computationally inexpensive and mostly leads to FIR filters .By extending the theory available for analysis side of Nonuniform filter bank and using synthesis bank in the proposed fashion, the resultant filter bank will be useful to compression, code or represent the signal. When this estimated filter bank is used in the application of compression, it is observed that the results of compression are better than the uniform filter bank for audio and speech signal. Finally, the results of simulations are also presented. I. INTRODUCTION The need for modeling of stochastic processes is of fundamental importance in applications such as representation, compression, identification, etc. [10]-[11]. The process of modeling involves finding an accurate mathematical description of the underlying physical process and then estimation of the model parameters. Framework based on multirate filter bank can be efficiently used to model the general class of stochastic processes. Multirate analysis/synthesis filter bank systems are conventionally used to decompose signals into separate frequency bands. The most frequently studied cases of filter bank systems are one 315

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

with equal sampling factors, called uniform filer bank (UFB) system. The M-channel UFB is shown in fig.1. In applications such as audio, speech and image analysis and coding, the information desired lies in particular frequency band and if we are analyzing this signal with uniform filter bank(s) then it may happen that, the band of interest may get divided into two or more bands. In this situation instead of uniform filter bank, if non-uniform filter bank is chosen then we can divide the input spectrum in bands so that the band of interest will be a single band. Hence, same can be processed in the required manner. The non-uniform filter bank is best suited in this situation. NUFB is widely used in audio and speech processing application and offers considerable flexibility in decomposition of frequency bands. A proper choice of these bands will lead to efficient coding and reduced computational complexity. A number of researchers have addressed the issue of finding filter banks [7]-[11] matched to a given signal. Tsatsanis and Giannakis [8] have obtained an M-band perfect reconstruction (PR) filter bank which minimizes the mean square error between the original signal and its low-resolution version to give uncorrelated, low-resolution principal components with decreasing variance. The reconstructed output signal is not exactly the same as input signal and also, the approach is computationally expensive. Lu and Antoniou [7] have obtained an M-band signal adapted biorthogonal filter bank by solving a constrained optimization problem which minimizes coding gain related objective functions subject to perfect reconstruction condition, leading to a very complicated solution. Gupta, et al [10]-[11] presented design strategies for bi-orthogonal PR filter bank by maximizing the projection of a signal on the scaling subspace, both deterministic and random signal cases. The methods presented for the design of filters of M-band filter banks involve matrix inversions. The methods proposed in [8]-[11] are limited to uniform filter bank structure and do not have natural extensions to non-uniform case apart from being computationally expensive. The issue of finding synthesis side of NUFB for a given decimation set is open issue from many years [2-7]. This paper is an attempt towards the issue of design of synthesis NUFB matched to a given signal or its statistics, when the outputs from analysis sides are known. Many times, we need to perfectly reconstruct the given signal/signal class only and not every signal. So we propose an approach for perfect reconstruction of the given signal only. The proposed approach is not only simple but leads to FIR synthesis filter where order can be chosen according to desired level of accuracy for perfect reconstruction. The issue is addressed for the case of non-uniform filter bank which is most general case. The outline of the paper is as follows: Section-II gives the relevant results from the theory of multirate filter bank. In Section-III a novel method of finding signal matched non-uniform synthesis bank is presented. To validate the theory, simulation results and discussions are presented in sectionIV. The conclusion is presented in section -V. II. RELEVANT RESULTS FROM THE THEORY OF MULTRATE FILTER BANK Filter bank is a system consisting of analysis and synthesis banks. The analysis bank decomposes a signal into a many subbands with smaller bandwidths [1]. The synthesis bank reconstructs the signal decomposed by analysis bank. A typical M-channel uniform filter bank (UFB) system, which decomposes the signal into uniform band, is shown in Fig. 1. A signal x ( n ) is passed through all the filters H k ( z ) , 0 k M 1 , in the analysis bank simultaneously (Fig.1). The input signal x ( n ) gets split into M subbands by these analysis filters. The outputs of the filters are, then, decimated by M to produce subband signals xi ( k ) , 0 i M 1 . At the synthesis end, the subband signals, xi ( k ) s, are upsampled by 316

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

M and then passed through synthesis filters, Fi ( z ) s, and then added to produce the ( n ) . The subband signals, xi (n) s, can, mathematically, be written reconstructed signal, x as:

xi ( k ) =

x(n)h ( Mk n),
i n = i i

i = 0,1,K M 1.

(1)

( n ) , can be expressed as: The reconstructed signal, x


M 1

( n) = x

x (k ) f (n kM ) .
i = 0 k =

(2)

Fig. 1 M-channel Maximally Decimated Uniform Filter Bank.

Fig .2 Typical M-channel Nonuniform Filter Bank


( n) is taken as cx[ n n0 ] [1]. In the case of perfect reconstruction (PR) filter bank, x In Fig. 1, in each channel of the analysis bank, the signal x( n) is first convolved with filter and then every M th sample of the filter output is retained. Similarly at the synthesis side, in each channel subband signals are first upsampled by M and then passed through the synthesis filters. The extra computations in the filter bank can be reduced by employing polyphase decomposition. This decomposition also permits great simplification of theoretical results [1]. There are two types of polyphase decompositions, namely, TypeI decomposition and Type-II decomposition. Typically, Type I representation is used for the simplification of analysis filter bank and TypeII representation is used for the simplification of synthesis filter bank. The polyphase representation of the M -channel filter bank is shown in Fig. 3.

317

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

Fig. 3 Polyphase Representation of M-channel Uniform Filter Bank The signal matched analysis filter bank is single-input, multiple-output systems with channel outputs are white within and across the channels. In short it is single-input, multiple-output whitening filter. The signal matched synthesis filter bank can be seen as multiple-input, single output inverse filter. Filter bank is said to be signal matched if all its analysis and synthesis filters estimated from signal or its statistics [7]. III. DESIGN OF M-CHANNEL SIGNAL MATCHED SYNTHESIS FILTER BANK In this section we address the issue of finding the synthesis side of the nonuniform filter bank. Clearly one of the ways could be by use of the theory of PRFB to obtain the synthesis side. This approach requires inverse of the analysis polyphase matrix [1] which, in general, leads to IIR synthesis filters. Also this approach is very tedious as we have to first convert NUFB into the equivalent UFB. Then analysis polyphase matrix is to be formed and then inverse is taken to obtain synthesis filters. Also, we need to perfectly reconstruct the given signal/signal class only and not every signal. So we propose an approach for perfect reconstruction of the given signal only. The proposed approach is not only simple but leads to FIR synthesis filter where order can be chosen according to desired level of accuracy for perfect reconstruction. In fact one could choose the synthesis filters FIR as well as IIR and design the system. Since the analysis filter bank is already at our disposal, we make the following assumptions: (i) The input signals to each channel of the synthesis filter bank are known; (ii) each channel input is WSS (When dealing with the stochastic processes). This assumption will be clearly valid if the underlying process is WSS, since filtering and down sampling retain the stationarity. Second the assumption will be valid even if the underlying process is WSCS provided the respective channel decimators are integral multiple of the period of cyclostationarity. Since a self-similar process passed through a band pass filter produces WSS process, the assumption would be valid for such processes also in all the channels but the low pass one. So the inputs to the synthesis side are assumed to be uncorrelated, band pass, white signals, which are the outputs from the estimated analysis side. To estimate the filters of the synthesis side for PRFB we choose the output of synthesis side as just delayed version of input signal. Precise problem statement is as follows: Given M different band pass, white inputs (fig. 4), find the corresponding M-synthesis filters such that output is close to the desired signal in some optimum sense.

318

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

A.

Least Squares Estimation of M-channel Nonuniform Synthesis Filter Bank Let us consider the estimation of k th channel synthesis filter with upsampling factor n k . Let y (n) be the output signal which is a delayed version of (the input to the analysis

side) the given input signal, and let x k (n) be the band pass, white input to the kth branch, as shown in fig. 4. In order to estimate the synthesis filter f k (n) we consider the following set of equations: p nr ) , (3) e k (n) = y ( n) f k ( r ) x k ( n = i nk and l , i Z . nk r =0
r = l nk

Notice here that error is computed with respect to y (n) , in place of y k (n) which is k th channel output. Since each output is assumed to have mutually independent inputs, this assumption is clearly justified. By writing the above equations in vector form for all n , we get: ek = y X k f k . (4) By minimizing the error ek in the above equation in the Least squares sense, the estimate of the filter parameters are obtained. Least squares estimate of f k can be written as: = X+y. f
k

(5)

where X + is the Moore-Penrose inverse of X . However, as can be easily seen, these are only those coefficients which are multiple of n k . In order to find all the coefficients of the filter we need to obtain and solve the equations for n = i n k 1, i n k 2,..., i n k ( nk 1) .By using the same approach the remaining M 1 filters of synthesis side can be estimated. B. MMSE Estimation of M-channel Nonuniform Synthesis Filter Bank The synthesis filters for the given statistics case can be estimated using the following approach: For obtaining the synthesis filters of the PR FB we can choose output statistics to be same as the input statistics. Let R y ( m, n) be the statistics of the output signal. Let
R xk ( m, n) , k = 0, 1, M 1 be the second order statistics available at the input of kth channel. Consider the equation (3), again. nl By multiplying the equation (1) by x k ( ) on both the sides & taking expectation and then nk applying MMSE criteria, we get:
E{ y ( n) x k (
p nr nr nl . )} = f k (r ) E{x k ( ) xk ( )} r =0 nk nk nk r =l nk

(6)

But we know from the analysis side that,


p

x k (n) = hk (r ) x(n k n r ) , k = 0,1,.., M 1 .


r =0

(7)

319

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

By putting n =

nl in equation (7) we get the following: nk

xk (

p nl ) = hk (r ) x(n l r ) , n = qM , l = qM , q Z . nk r =0

(8)

Substituting the equation (8) in the equation (6), we get: p p nr nl E{ hk (r ) y (n) x(n l r )} = f k (r ) E{x k ( ) xk ( )} , n = qn k , l = qnk , q Z (9) nk nk r =0 r =0
r = l nk p p

hk (r ) Rx (n n0 , n l r ) =
r =0

f
r =0 r =lM

( r ) R xk (

nr nl , ) , n = qnk , l = qnk , q Z .(10) nk nk

The last equation above follows from the fact that for PR we require, y ( n) = x( n n0 ) .
th By solving the equation (10), we get the nk coefficients of filter f k (n) . Similarly, by writing equation for remaining n (i.e. n = qn k 1, qn k 2 ,.., qn k ( n k 1) ), as was done in the

least squares case, we can find the remaining coefficients the filter f k (n) . In this fashion all M synthesis filters are estimated from the statistics of the signal. The resultant synthesis filter bank, obtained in this manner, will be matched to the statistics of the stochastic processes. As a special case, by taking n k = M , 0 k M 1 , we can obtain the uniform synthesis bank. This means that, if we known the process or its statistics, we can find the analysis side [6-7] and then by delayed version of the process and input signals from the analysis side, filters of synthesis side can be estimated.

Fig. 4 M-channel nonuniform synthesis bank IV. SIMULATION RESULTS The proposed method is applied on audio clip for the given data case as well as the given correlation case. To obtain the signal matched synthesis bank, first, the theory presented in [10-11] is used to get the output of the analysis side. Then by using outputs from the analysis side and delayed version of the input, synthesis filters are estimated by using proposed theory. In given correlation case, the results are obtained by assuming clip as: 1/ f b 320

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

processes. Magnitude frequency spectrums of the estimated analysis and synthesis filters are shown in Fig. 5(a) for the given statistics (Self-similar case) and in Fig. 5 (b) for the given data case. The plots of original and reconstructed signals are shown in Fig. 6(a)-(b) respectively. From the plots it is clearly seen that the signal is perfectly reconstructed.
1.5 1 0.8 1 0.6 0.4 0.5 0.2 0 0 500 1000 0 0 500 1000 0 0 500 1000 0.5 1 1.5

800 600

15

15

10 400 5 200 0 0 500 1000 0 0 500 1000

10

0 0 500 1000

Fig. 5 (a) Analysis ( h0 , h1 & h2 : first row) and synthesis filters ( f 0 , f1 & f 2 : second row) Magnitude responses. For given data case

1 0.8

1.5

1.5

1 0.6 0.5 0.4 0.2 0 500 1000 0 0 500 1000

0.5

0 0 500 1000

8 6 4

5 4

15

10 3 2 5 2 0 0 500 1000 1 0 0 500 1000 0 0 500 1000

Fig. 5 (b) Analysis ( h0 , h1 & h 2 : first row) and synthesis filters ( f 0 , f 1 & f 2 : second row) Magnitude responses. For given correlation case.

321

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME
1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 0 500 1000 1500 2000 2500 3000

Fig. 6(a) Simulation results of signal matched filter bank for given statistic case (original signal)
1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 0 500 1000 1500 2000 2500 3000

Fig. 6 (b) Simulation results of signal matched filter bank for given statistic case (reconstructed signal)

V. CONCLUSION A novel approach for the design of signal matched synthesis filter bank is presented. The synthesis filter bank design in this fashion is computationally efficient and lead to FIR filter in most of the cases. Using proposed approach it is possible to find corresponding synthesis filter bank for non-uniform filter bank case, which is open issue from many years. From simulation results, it is observed that, input signal can be perfectly reconstructed at the output of the synthesis bank. By extending the theory [10] for analysis side of NUFB and using synthesis bank obtained in proposed fashion, the resultant filter bank will be useful to compression, code or represent the signal. When this estimated filter bank is used in the application of compression, it is observed that the results of compression are better than the uniform filter bank for audio and speech signal.

322

International Journal of Electronics and Communication Engineering & Technology (IJECET), ISSN 0976 6464(Print), ISSN 0976 6472(Online) Volume 4, Issue 2, March April (2013), IAEME

REFERENCES [1] P.P. Vaidyanathan, Multirate systems and filter Banks (Englewood Cliffs, NJ; Prentice Hall, 1993). [2] P.P. Vaidyanathan, Multirate digital filter banks, polyphase networks and applications: A tutorial, Proceedings of the IEEE, Vol. 78, pp.56-93, January 1990. [3] Ashish Pandharipande and Soura Dasgupta, Biorthogonal Nonuniform Filter Banks and Tree Structures, IEEE Transactions on Circuits and Systems, 49 (10) , 2002, 1457-1467. [4] J. Kovacevic and M. Vetterli, PR Filter Banks with Rational Rate changes, IEEE Trans. on Signal Processing, 41(6), 1993, 2047-2060. [5] Sony Akkarakaran and P..P. Vaidyanathan, New Results and Open Problems on Nonuniform Filter Banks, Proceeding IEEE Int. Conference Acoust. Speech and Signal Processing, 1999, 1501-1504. [6]. S.L. Nalbalwar, S.D. Joshi and R.K. Patney, Signal Matched NUFB, Proceedings of National Conference Communication , IIT Kanpur , 26-28, Jan. 2007. [7]. W. S. Lu and Antoniou, Design of signal adapted biorthogonal filter banks, IEEE transactions circuits Syst. I, Funam. Theory Appl., vol. 488, no. 1, pp.90-102, Jan. 2001. [8]. M.K.Tsatsanis and G.B. Giannakis, Principal component filter bank for optimal multiresolution analysis, IEEE Trans. Signal Processing, vol. 43, no.8, pp. 1766-177, Aug. 1995. [9]. G. W. Wornell, A Karhunen-Loeve like expansion for 1/f processes via Wavelets, IEEE Trans. Inform. Theory, 36(4):859-861,1990. [10]. A. Gupta, S.D. Joshi and S. Prasad, A new approach for estimation of statistically matched wavelet, IEEE Trans. Signal Processing, vol. 53, no.5, pp. 1778-1792, May 2005. [11]. A. Gupta, S.D. Joshi and S. Prasad, A New Method of Estimating wavelet with desired features from a given signal, Elsevier Signal processing, 85(7), 2005, 147-161. [12]. B.E. Usevitch, A tutorial on modern lossy wavelet image compression: Fundamentals of JPEG 2000, IEEE Signal Processing Magazine, vol. 18, no.5, pp. 22-35, Sept. 2001. [13]. B.E. Usevitch, JPEG 2000 still image compression: Fundamentals of JPEG 2000, IEEE Signal Processing Magazine, vol. 3, no.5, Sept. 2001. [14]. S.L. Nalbalwar. S.D. Joshi and R.K. Patney, A Novel Approach for Design of Signal Matched Nonuniform Synthesis Filter Bank proceedings of NCC-2009, IIT, Guwahati, Jan 16-18, 2009 [15] Vinod Karar, SmarajitGhosh, Effect of Varying Contrast Ratio And Brightness Nonuniformity Over Human Attention And Tunneling Aspects In Aviation International Journal Of Electronics And Communication Engineering &Technology (IJECET) Volume 3, Issue 2, 2012, pp. 400 - 412, Issn Print : 0976- 6464, Issn Online : 0976 6472 [16] S.Anandanarayanan and Dr.S.K.Srivatsa, A High Performance Novel Image Compression Technique Using Huffman Coding With Edge Dection International Journal of Computer Engineering & Technology (IJCET) Volume 4, Issue 2, 2013, pp. 17 - 22, Issn Print: 0976 6367, Issn Online: 0976 6375

323

S-ar putea să vă placă și