1776

IEEE TRANSACTIONS

ON ACOUSTICS, AND SPEECH,PROCESSING, SIGNAL VOL. ASSP-35.

NO. 12, DECEMBER 1987

elements. In these circumstances, the amplitudes of the elements FFT Pruning Applied to Time Domain Interpojation of both sequences are converted to a unit magnitude before the and Beak Localization a coherence estimate is evaluated. In this section, condition is given for the distribution of the coherence estimate to be independentof SVERRE HOLM second-channel statistics in this situation. Interestingly, the independence does not rely on the notion of spherical symmetry. Assume that the sampies in sequence are statistically indepena Abstract-Theefficiency of the fast Fourier transform magbe indent. Also, supposea and b have the property that each sample has creased by removing operations on input values which are zero, and unit magnitude, i.e., x: y f = u: u: = 1 for n = 1, . * . , N . on output values which are not required. This is applied to interpolaThen theycan be written as a = (e, . . . , e i o V ) and b = ( e i 4 , tion of complex and real valued time domain functions. For real func-

Ifeach sample e@ a i s uniformly distributed on the unit cirof cle, then so is each term exp [ i( 8, - +,)I in this expansion of y 2 ( a ,6)-regardless of the distribution of the samples of b . This I. INTRODUCTION is easily proven by convolution of an arbitrary probability density PruningofthefastFouriertransform(FFT)algorithm is the function p ( & ) with a uniform density function, modulo 27r. Furelimination of operations on zeros. When the numberinput points of thermore, the terms exp [ i (8,* - & ) ] will be statistically indepenis dent of one another. Thus, in this case, the distribution of the MSC less than the number of transform points, the number of butterestimate is 1/ N 2 times the distribution of the length squared of a flies may be reduced. This is referred to as input pruning and was vector formed by adding N unit vectors in a plane which have uni- first described by Markel [l]. The efficiency and regularity of the formly distributed direction. This problem has long history as the algorithm was improved by Skinner [2] by pruning the decimationa in-time instead of the decimation-in-frequency algorithm. This alproblem of a random walk in the plane. The problem of determining this distribution was posed by Pearson in 1905 (see [7]) and its gorithm requires ( N / 2 ) . log, IVZ complex multiplications where N is the transform length and N Z is the number of nonzero input history is detailed in [8]. values, as compared to the FFTs ( N / 2 ) . log, N complex multiNote that if the elements of the a sequence are allowed to have plications. nonconstant amplitudes (i.e., A , = 1 x, iy, 1 ) that are indepenSreenivas and Rao [3] extended the algorithm by combining indently distributed, then invariance with respect to second-channel put and output pruning. This applies to the situations where both statistics is still maintained. The distribution of the coherence estimate is, in this case, related to random walk with a random step the number of input points and the number of output points are less a than the transform length. By viewing the decimation-in-frequency size. It is also important to observe that although each term of the a algorithm as the transpose of thedecimation-in-timealgorithm, Markels pruning may be applied to the output of Skinners algosequence A,eiB is spherically symmetric, the a sequence itself is rithm, or vice versa. Furthermore, they generalized the algorithm not spherically symmetric. by allowing output pruning anywhere in the output range] .Nagai [5 VII. CONCLUSIONS [7] gives an alternative algorithm for generalized output pruning. We feel the important aspect of this correspondence is the dem- It has a more regular structure than in [5], while keeping the savonstration of the utility of a geometric approach to this probability ings in the computation. In [4] and 161, pruning algorithms are expressed using matrix formulation. problem, not just for the results obtained but also for the insight The most common application of input pruning is interpolation into why these results come about. Although we have focused on in the frequency domain. By appending zeros to a sequence prior the problem of coherence estimation with no signal present, we anticipate that the geometric approach will prove useful the case to Fourier transformation, a high resolution spectrum is obtained. in This is used, for instance, in autoregressive spectral analysis. The when signal is present. input and output pruned algorithm gives additional savings comby REFERENCES puting the high resolution spectrum in only a preselected narrow frequency band. In [5], the amount of computation is compared to J . J. Gosselin, Comparative study of two sensor (magnitude-squared the direct DFT, unpruned FFT, and chirp 2 transform methods. coherence) and single sensor (square law) receiver operating chacterIn this correspondence, the application is efficient time domain istics, in Proc. IEEE Int. ConJ Acoust., Speech, Signal Processing, interpolation. It will be shown that with some modifications, the Hartford, CT, May 1977, pp. 311-314. G. C. Carter, C . Knapp, and A. H. Nuttall, Estimation of the magavailable pruning algorithms may be applied. It will also be shown nitude-squared coherence function via overlapped fast Fourier transthat a byproduct for real valued time samples is the Hilbert transform processing, IEEE Trans. Audio Eleccroacoust., vol. AU-21, pp. form. It has applications in estimation of time delay found by lo337-344,Aug.1973. calization of the peak of the cross correlation function. to A. H. Nuttall,Invarianceofdistributionofcoherenceestimate second-channel statistics, IEEE Trans. Acoust.. Speech, Signal Pro11. TIMEDOMAININTERPOLATION cessing, vol. ASSP-29, pp. 120-122, Feb. 1981. Consider a time sequence x(n) with N samples ( a = 0 , . . . , S. MacLane and G . Birkhoff, Algebra. New York: MacMillan, 1967, N - 1), and a new interpolating sequence x ( n ) with rN samples, pp. 394-295 and pp. 420-433. where r is a positive integer. This new sequence should have the W.Fleming, Functions of SeveralVariables. Reading, MA: Addison-Wesley, 1965, p. 183. propertythat x ( m ) = x ( n ) for n = 0, . . . , N - 1. This reYork: McW. Rudin, Principles of Mathematical Analysis. New quirement distinguishes interpolation by FFT pruning from interGraw-Hill, 1953, pp. 145-150 and p. 162. K. Pearson, The problem of the random walk, Nature, vol. 72, pp. Manuscript received August 7, 1986; revised June 19, 1987. 294-203, 1905. P.O. Box 265, N-1371 The author is with AIS Informasjonskontroll, J . A. Greenwood andD. Durand, The distribution of length and comAsker, Norway. ponents of the sum of n random unit vectors, Ann. Math. Stat., vol. IEEE Log Number 8716987. 26, pp. 233-246, 1955.
~

.., - ,

tions, analytic signal concepts may be used to get the Hilbert transform as a byproduct, and applied to the cross correlation function this gives an effcient and accurate method for peak localization.

0096-3518/87/1200-1776$01.00 0 1987 IEEE

IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-35, NO. 12, DECEMBER 1987

1777

polation with a low-pass filter [ 8 ] . The requirement is equivalent greater than the accuracy in determining the location of the peak of the correlation. A further advantage for bandpass signals is that to increasing the sampling frequency by a factor r, and corresponds to adding zeros between the old and new Nyquist frequency. Let the envelope is free from the influence of the oscillations caused X ( k ) be the N point discrete Fourier transform of ( n ), and X( k ) by the center frequency of the frequency band. Peak localization x in the envelope of the cross correlation function is therefore a good the rN point discrete Fourier transform of x ( n ) . Then the new rN frequency samples X ( k ) should be found from the N frequency method for determining the delay of narrow-band signals. samples X ( k ) via

k=O;*- ,N/2 k=N/2+ 1;-* X(k ,rN-N/2

+N

- rN),

k = rN

- N/2

+ 1,

, rN - 1 .

The pruned FFT of [ 2 ] cannot be applied directly to inverse transform X ( k ) since the zeros are in the middle of the sequence instead 111. TIMEDOMAIN INTERPOLATION IN A LIMITEDREGION of at the end. However, Nagais decimation-in-time algorithm [ 7 ] For peak finding it should not really be necessary interpolate to may be transposed to handle these input data. The transposed aigorithm is a decimation-in-frequency algorithm with frequency shift the entirecross correlation function, but only the region around the A For- peak. With some prior knowledge on the peaks location, both inat the input, when operated as an inverse Fourier transform. put and output pruning could be applied. This prior knowledge tran implementation of this algorithm may be obtained from the could be obtained from a tracking algorithm which is often a part author. Note that Sreenivas and Raos generalized output pruning of a time delay estimation system, or it could come from a prelimalgorithm [SI is not so simple to transpose into an algorithm that By will handle the input data of ( 1 ) , since the nonzero input samples inary, low resolution cross correlation estimate. using either the algorithm of Sreenivas and Rao [ 5 ] , or Nagais algorithm [ 7 ] with are in two separate bands. The transpose of Nagais algorithm; however, will handle this because of the cyclic property of the fre-input pruning [ 2 ] and equation ( 2 ) , a high resolution, limited region, correlation estimate and its Hilbert transform may be obquency shift operation. In many cases, the time domain signal to be interpolated is real. tained. In this case, the information inX ( k ) is fully contained in the FouIV. CONCLUSION rier transform of the analytic signal The pruned FFT has applications in frequency and time domain interpolation and smoothing. In this correspondence, the frequency jW), k=O domain interpolation methods have been extended to time domain i 2X(k), k = 1 , - * * ,N / 2 - 1 interpolation. A byproduct for real signals is the Hilbert transform S(k) = which aids in accurate peak localization. (N/2), k =N/2

::

k =N/2

+ 1, .

, rN - 1 .

(2)

S ( k ) can be directly inverse transformed by the pruned FFT, and the result will be a complex sequence s(n), with the desired se, quence x( n ) as the real part, and its Hilbert transform as the imagREFERENCES inary part. The magnitude of s (n) is the envelope of the analytic [l] J. D. Markel, FFT pruning, IEEE Trans. Audio Electroacoust., signal, vol. AU-19, pp. 305-311, Dec. 1971. This was also usedby Markel [l] for cepstral smoothing. In this [2]D. P. Skinner,Pruningthedecimation-in-time FFTalgorithm, application, one originally has rN values of the cepstrum, but reIEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-24, pp. duces it to by applying a cepstral window that zeros out the samN 193-194, Apr. 1976. (2 ples in the middle of the sequence; Via ) and the pruned forward [3] T. V. Sreenivas and P. V . S . Rao, FFT algorithm for bothinput FFT, this gives a smoothed log spectral estimate. A difference from and output pruning, IEEE Trans. Acoust., Speech, Signal Processtime domain interpolation is that reduction instead of increase of ing, vol. ASSP-27, pp. 291-292, June 1979. [4] L. P. Jaroslavski, Comments on FFT algorithm for both input and resolution is the aim. Another difference is that the imaginary part output pruning, IEEE Trans. Acoust., Speech, Signal Processing, of the result (the Hilbert transform) is of no use. vol. ASSP-29, pp. 448-449, June 1981. In time domain interpolation, ( 2 ) and the pruned FFT have the [5] T. V. Sreenivas and P. V. S. Rao,Highresolutionnarrow-band following advantages. Better accuracy than the often-used paraspectra by FFTpruning, lEEE Trans. Acoust., Speech, Signal Probolic interpolation [ 9 ] which gives a biased estimate of the peaks cessing, vol. ASSP-28, pp. 254-257, Apr. 1980. location; and simpler and more efficient implementation than the [6] R. G. Gan, K . F. Eman, and S.M. Wu, An extended FFT algorithm I straightforward implementation in [ l o , sect. V. Here the interfor ARMA spectral estimation, IEEE Trans. Acoust., Speech, Sigpolating samples are found by using the time shift property of the nal Processing, vol. ASSP-32, pp. 168-170, Feb. 1984. Fourier transform. The cross spectrum is multiplied bya linear [7] K. Nagai, Pruning the decimation-in-time FFT algorithm with frequency shift, IEEE Trans. Acoust., Speech, Signal Processing, vol. phase shift, and inverse transformed by the short,point, discrete N ASSP-34, pp. 1008-1010, Aug. 1986. Fourier transform. This is done a total of r times. The results are [8] Y.C. Lim, An interpolation technique for computing the DFT of a interleaved to get the interpolated cross correlation. sparse sequence, IEEE Trans. Acoust., Speech, Signal Processing, A third advantage of the present method is the availability of the vol. ASSP-33, pp. 1456-1460, Dec. 1985. Hilbert transform of the cross correlation. Since it passes through [9] R. E. Boucher and J. C. Hassab, Analysis of discrete implementazero at the lag where the correlation has a peak, it is also a sensitive of generalized cross correlator, IEEE Trans. Acoust., Speech, tion [1 indicator of the peaks location. Cabot 1 1 demonstrates this propSignal Processing, vol. ASSP-29, pp. 609-61 1, June 1981. erty by giving examples of typical correlation functions and their [lo] W. H. Haas and C. S. Lindquist, A synthesis of frequency domain filters for time delayestimation, IEEE Trans. Acoust., Speech, SigHilbert transforms. Bendat [12] also shows that the accuracy in nal Processing, vol. ASSP-29, pp. 540-548, June 1981. is determining the exact zero crossing of the Hilbert transformoften

ACKNOWLEDGMENT The author would like to thank one of the reviewers and Dr.J. Endresen for helpful suggestions and comments.

[ I l l R. C. Cabot, A note on the application of the Hilbert transform to time delay estimation, ZEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-29, pp. 607-609, June 1981. [12] J. S. Bendat, The Hilbert Transform and Applications to Correlation Measurements, Bruel Kjaer and Handbook BT0008-11, Naerum,

where

Denmark,1986. and

Comments on Two-Dimensional Interpolation by Generalized Spline Filters Based on Partial Differential Equation Image Models
N. B. KARAYIANNIS
AND

A . N. VENETSANOPOULOS

The second-order B-spline corresponding to the operator

Abstract-The generalized spline formula corresponding to the separable semicausal Partial Differential Equation (PDE) image model is given in its correct form. In addition, the correct formula for the second-order B-spline is derived. Finally, it is shown that there is no reason to consider separable PDE image models only, at least when describing the concept of generalized splines.

L(D,) = D,Z -

0 1

(DX CY)(0,- a ) +

is obtained by convolving B , (x, a ) with B , (x, - a ) . In the paper, the second-order B-spline is given by (31), Le.,

Inthiscorrespondenceweintroducecorrectionsinsomeformulas of the above paper which, regardless of their origin, may cause problems in the understanding of the paper. For the convenienceoffuturereaders,weidentifytheerrors first and subsequently provide the correct formulas. The generalized spline formula for the semicausal PDE model, as given in the paper by (6), should be corrected. In this case

where

(exp [ a ( 4 R + x)]
-

+Cax)

+ (1,a) (e-cux
and

exp [ a ( 4 R + x)])]

where

+ 2 exp [ a ( 2 R + x ) ]
LZ(Di) L z ( D z )s,(z)
= 0,

= x,

and L* is the formal adjoint operator of L. Thefunction s y ( . ) iscorrectlygiven by (18) of the paper. However, s( ), as given in (18), is incorrect. The adjoint operator ,. forL,( is
e )

- 2 exp [ a ( 2 R - x ) ] - e - a x - (x + h ) 2 exp [ a ( 2 + x)] ~

+ 2 exp [ a ( 2 R

x)]

L,*(D,)

& ( D X )= Di

+ a:. +
0.

- x( exp [ a ( 4 R + x] )

Therefore, sx( . ) is the general solution of

(0:

+ a:)s,(x)

The second-order B-spline should be a continuous function of x E [ -2R, 2R] and,therefore,becontinuousatpoint x = -R. Hence, B2 ( . ) should satisfy the equation

That is, s , ( x ) = k l cos a l x

+ k2 sin alx + k3x cos a l x + k4x sin alx.

In addition. the second-order B-spline given (3 1) of the paper in is also incorrect. The first-order B-spline corresponding to the operator

It can be easily verified that ( 2 ) is not satisfied by (31) regardless of what constant h, undefined in the paper, is. In the sequel, the correct expression of B2( . ) is obtained. For X E [ -2R, - R ] ,

L(D,) = DX + a
is given by (30), i.e.,

B , ( x , a , - a ) = &(x, a , - a )

Manuscript received May 22, 1986; revised June 8, 1987. The authors are with the Department of Electrical Engineering, University of Toronto, Toronto, Ont., Canada MSS 1A4. IEEE Log Number 8717268. IT. C. Chen and R . J. P. deFigueiredo, ZEEE Trans. Acoust., Speech, Signal Processing, vol. A S P - 3 3 , pp. 631-642, June 1985.

After some algebra

&(x, a , - a )

2eZaR (e2aR- I)

[(x

+ 2 R ) cosh [ a ( . + 2 R ) ]
+ 2R)]]
(3)

- ( l / a ) sinh

[a(x