This action might not be possible to undo. Are you sure you want to continue?

not only can we see how the frequencies change. Although the number ofways of describing a given signal are countless. by analyzing the signal in time and frequencyjointly. 1. The fundamental idea behind Fourier’s original work where Y can be of finite dimension or infinite dimension.. which clearly r 1 ~nn-.. that is. [62] and seismic signals (Fig. i. perhaps even relatively short duration. but we also can see the intensity of the frequencies as shown by the relative brightness levels of the plot. 2 is the time waveform of the word “hood” spoken by a five-year-oldboy. however. 1 ..t has been well understood that a given signal can be classical time and frequency analysis.” which are usually of finite.. Different signal representations can be used for difIn this article. as a function of both time and frequency.500 0. parametric) time-frequency analysis method. many alternatives. But when studying or designing the sys-.. MARCH 1999 IEEE SIGNAL PROCESSING MAGAZINE 53 .200 0. the square of the Fourier transform (power spectrum) is also widely used in many applications. the classical power spectrum does not reveal such important information. signal expansion and inner product. As we time and frequency is the Fourier wansfom. While the time Expansion and Inner Product domain function indicates how a signal‘s amplitude In what follows. extending from negative infinity to positive infinity in time. we introduce the basic concepts and ferent applications. Despite their simple interpretation of pure the set of elementary functions {tp R E Z .short-time Fourier transform. for a given signal s from the domain Y. [5 31.partition this article into two parts: the linear (e. we are able to better understand the mechanism of human speech.Finally. the frequency domain function tells mathematical tools to perform the joint time-frequency how often such changes take place. The bottom plot of Fig. indicates the pattern of the formants. we often llke to study signals and systems in the fre. i n contrast to me T 1 1 . For example.The seismic signals are not like the sinusoidal functions.e. The larger plot in 1. tem. Although the frequency content of the majority of signals in the real world evolves over time. was to decompose a signal as the sum ofweighted sinusoi. . the sinusoidal functions are not good models. which reveals three frequency tones. Wigner-Ville distribution). In order to 0. ” H o o d (data courtesy of Y. Seconds such as the Gabor expansion..g. The plot on the right is the standard power spectrum. .600 0. For such applications. the most important and fundamental variables in nature are time andpeguency. 1 2. From the spectrum alone. all understand. signals obtained well-tested algorithms for joint time-frequency analysis.300 0. Seismic signal. In addition to linear transformations. wavelets 1211.g. [5 11. However.700 0. One common example is the speech signal. the Beckman institute at the University of .we may write a signals in terms of a linear combination of dal functions.100 0. Zhao. we roughly fimctions of time.> opea ana wiaeiy sruaiea. T h s is because many important features of quadratic transforms (e. [501. techniques as joint time-peguency anabsis. Gabor expansion) and the quency domain. the Fourier transform is not always the best the set of integers for the Y domain.819 overcome this problem. 1). where 2 denotes frequencies.400 0. we name these new represented in an infinite number of different ways. we will first review the fundamental changes over time. From the time-dependent spectrum. from most engineering applications are usually Analogous to the classical Fourier analysis. we cannot tell how those frequencies evolve over time. In other words. .000 0. the upper left is the time-dependent spectrum. the power spectrum in general is only suitable for those signals whose spectra do not change with time. tool to analyze “real life signals. The bridge between analysis. iiiinois). the signal or system are more easily characterized in the fre. we briefly introduce the so-called model-based (or quency domain than in the time domain. Common exnsZ amples include biomedical signals [381.

=~~/. Each individual function can be considered as the tick mark of the ruler. in all cases.)=c8[n--n’] and{v. the signal‘s feature described by the coefficients {a. the resulting expansion is redundant. however. and are interchangeable. 3(b)]. which is known as an expansion or series. formulae (2) and (3) are known as transformations and $. the larger the If$.> for all vectors S with L elements. One particularly interestwhere c denotes a constant ing scenario is $ ( t ) = c ~ / (t). 54 IEEE SIGNAL PROCESSING MAGAZINE . where 6[n]= 1 n=O 0 otherwise’ (5) for continuous-time signals.). is complete for three-dimensional space. then G must be equal to I ’ T. to measure the signal under investigation. “*” in (2) and (3) denote the complex conjugate operation.. We have the matrix G with dimensionM x L and the matrix H with dimension L x M . the positions of U/. function$ .} are exactly the signal’s projection of the synthesis functions. For the tight frame. ~ }its ~ ~ .}.} n t z is unique. The goodness of the ruler is measured by the fineness of the unit.In other words. The concepts introduced above can also be understood from the matrix operation point ofview. . of any measurement largely depends on the smallest unit used. Otherwise.z are orthogonal. elementary functions for the signal expansion should be chosen such that the resulting tick mark is the finest. If G = H-’= HH.. that in general. we name {U/ nsZ as a tight frame. a family of vectors that characterizes any signal from its inner products (or scalar product) as in (2) and (3). ( t )is called the analysis function.} ns% and {$ .th~n(~.} such that the expansion coefficients can be computed by up to a constant multiplier c. yields... {U/. The expansion coefficient {a. but in orthogonal expansion. then there i!i no way to tell the patients height in terms ofcentimeters. the matrix G could not be unique. H and G are said to be biorthogonal to each other. If M = L (H and G are square matrices). (1)is t known as an inverse transform and ~ / . are exactly the signal‘s projection on the elementary functions I I MARCH 1999 {w. the expansion coefficients c{a. the closer the signal s(t) is to the dual (t). In both biorthogonal and orthogonal cases.} n s Z forms a frame [21]..} may not have obvious relations with the set of the basis function {U/ n } nsZ [Fig.taking the transpose. In this case. When the dual frame {$ .}.G. Because H G = I (6).. ( is)called the synthesis function. In this case. number. constituted by a set of functions {$ .S=S (6) inner product a. When the dual frame is not unique.. The operation of the inner product in (2) and (3) may be thought of as using a ruler. the matrix G is unique. Moreover. Suppose that the length of the signal 5 is L. For the biorthogonal. {a depicts the signal’s behavior in the Y’ domain. For a given frame { u / .} must be the signal’s projection on the basis function {U/ n t Z [Fig. either of them could be used for analysis or synthesis. In both cases. 3(a)]. ~ corresponding dual frame is generally not unique. If a physician uses a ruler whose smallest scale i:i the decimeter to measure a patient’s height.}ntZ forms a basis and there is no redundancy. there will nsZ be a dual set {$. when { u / ~ forms ~a ~ biorthogonal expansion.H a n d G are orthogonal.. We all know that the precision . then the set of {U/.If a physician uses a ruler wh smallest scale is the decimeter to measure a patient‘s height. the expansion coefficients {a.$. then we say that the space constituted by the matrix H i s complete. + .. In electrical engineering. That is. where M 2 L. For the orthogonal expansion. or (3) for dscrete-time signals. When the set of {U/. In this case. If urn H. with respect to HG. ey2 / y3 t“ (a) Biorthogonal Expansion (b) Orthogonal Expansion 3. expansion coefficients a. The superscript “H” denotes the Hermitian conjugate. .. which reflects the similarity between the signal s(t) and the dual function$ (t). Accordingly.}indicates the weight of the signal’s projection on the tick mark determined by @ . Note. So. The inner product has an explicit physical interpretation. then there is no way to tell the patient‘s height in terms of centimeters.} ntZ are } ~ basis.

If a function has a short time duration.s(t +2/2)s* (t-2l2)dt. The balance of the time and the frequency concentration is controlled by the parameter a. For example. R(T) =jls(t)s*(t-T)dt =J. it is the Gaussian-type signal. (9) which indicates the similarity between the signal and a set of harmonically related complex sinusoidal functions {eiznnr'*}. the tick mark adopted by the classical Fourier transform does not contain time information. (13) which indicates that either H or G can be used as analysis (or synthesis).} nsZ How to compute the corresponding set of dual functions {Cx}sez. precisely describe the signal's behavior at the frequency 2nn/T. According to the Wiener-Khinchin theorem. i. They are linked by the Fourier transform. we can't find a function that has an arbitrarily short time duration and narrow frequency bandwidth at the same time.G.we have s (t) = -I--1 27c S(o)eitod o (10) and the Fourier transform Conventionally.e. The two most fundamental questions for the expansion are A How to build a meaningful set of functions {U/ . the square of Fourier transform is named the power spectrum. The smaller the value of a. the measurements. the finest frequency resolution). the Fourier series is defined as Note that the averaging process removes all time information from R(2).e. such as Because the elementary functions ej2m r / T in (S).~=(G. ) ~ s=s (7) where the correlation function R ( T ) the average of the is instantaneous correlation s(t)s(t. For a non-periodic signal s ( t ) in L2. the Fourier coefficients { f i n } . the narrower the frequency bandwidth (longer time duration). From the uncertainty principle point of view. the reader should bear in mind that a function's time and frequency properties are not independent. and are not associated with a particular instant in time. that achieves the optimal joint time-frequency concentration. are orthonormal with respect to the scalar product. the classical Fourier series is unsuitable for characterizing signals that only last for a short time or whose frequency contents change over time. Because eizNlt'Tcorrespon& impulses in the to frequency domain. we can readily obtain the expansion coefficients of the Fourier series in (8) by the regular inner product operation ( 2 ) . Cabor expansion sampling grid. MARCH 1999 T 4. One of the most well-known expansions is the Fourier series.. Consequently. harmonically related complex sinusoidal functions.)~ ( H . H. Cabor Expansion As mentioned in the preceding section. its frequency bandwidth must be wide. The reason for this is that the elementary fimctions employed in the Fourier series extend from negative infinite time to positive infinite time. In other words.T). For a periodic signal T ( t ) with a period T. and are not associated with a particular instant in time. Hence. the dual function and the elementary function have the same form. Finally. or vice versa. i.e. the ruler used for the Fourier transform possesses the finest frequency tick marks (that is.The elementary functions employed in the Fourier series extend from negative infinite time to positive infinite time. the power spectrum can also be considered as the Fourier transform of the correlation function.. 55 IEEE SIGNAL PROCESSING MAGAZINE . or vice versa..g.

The smaller the value of a..+(t)} in general where Cm. Dennis Gabor. the STFT grid. Many important aspects of the Gabor expansion were not known during his lifetime [24].(t) are concentrated in [mT.Gabor Expansion and Short-Time Fourier Transform In 1946. and it was to a great extent. The set of elare won’t be orthogonal unless h. such as the exponential. {h. 5 ) . it is oversampling. product of TQ determines the density of the s a m p h g While Gabor focused on the expansion.because it is optimally c o n havior jointly? Before we answer these questions.~(t)} in general does ration of y(t). that are concentrated in both the time and frequency do-. dense enough.(t).) Then. mains [26J. the better the frequency resolution (poorer time resolution). as long as the sampling grid i:. (l6. Gabor suggested that the coefficient Cm. When the product TQ = 2n. The parameters T and Q are (STFT).. Gabor was awarded the Nobel Physics Prize for his dscovery of the principles underlying the science of holography.n called the Gabor coefficients. In fact. (In 1970. it is considered critical sampling. Because h..1 Gabor coefficients C. then the estimation of the b.n (t)=h(t-mT)eQm.+(t)}are not orthogonal.quency resolution).Short-time Fourier transform. Note that the set of {h. let’s recentrated in the joint time-frequency domain. Moreover.. the short-time Fourier transform the uncertainty principle.n reflected a signal’s where the analysis window function y ( t ) balances the behavior in the vicinity of [mT. However. which In Gabor‘s original paper. Consequently.n(t) are poorly concenementary functions { L I % . in terms of view another approach. the Gabor expansion is more transform can then be thought of as a signal‘s frequency suitable for characterizing transients and signals with behavior during the time period covered by the data sharp changes (Fig. If quency-shifted function h ( t ) .i. Gabor had the clear insight to see the potential of this expansion. a Hungarian-born British physi. [29]-[31].nQ] . or vice versa.%are not a signal’s projection onh. time and frequency resolutions. Therefore. Gabor expansion and the Gabor expansion-sampling STFT takes the Fourier transform (FT) on a block by grid. suggested expanding a signal into a set of function:. The resulting representation is now known as the Gabor expansion. the Gabor coeffi. for instance.%won’t be trivial. h ( t )is the normahzed Gaussian one can better describe a signal’s time and frequency befunctiong(t). The smaller the product. The blocks could be overlapped or cients C.. can be used in the Gabor expansion.nQ]. The smaller the time duthat is not really true. Because the frequency-shifted Gaussian functions block basis (Fig. 56 I ‘ 5. The troduction of the Gabor expansion [321. respectively. ~ ( ~ consists of a time. 4 illustrates the elementary functions used in the and e@‘““). use the coefficients as the description of the signal’s local property.cist.The time and the frequency resoblock.which appeared at about the same time as the intime and frequency sampling steps. because {~?.. 1). Time IEEE SIGNAL PROCESSING MAGAZINE MARCH 1999 ..and fre)} trated in either the time or the frequency domain. As a matter of fact. rectangular.. We shall disjointed. Instead of processing the entire signal at once. the denser the sampling. the better the time resolution (poorer frenot form a tight frame.e. c U 3 TQ 227~ (17) t all those commonly used analysis fbnctions. claims that spurred much of the subsequent work on joini: time-frequency analysis. modified the Fourier transform (inner products of s(t) Fig. he did not actually contribute to the mathematical theory of the Gabor e x pansion. his work and his claims that spurred much of the subsequent work on joint time-frequency analysis. defined in (14).. his work and hi:. Gabor limitedh(t) to be the Gaussian func tion and TQ = 2n.. the resulting Fourier have a short time duration. great extent. The percentage of overlap between each block elaborate on this issue later. and Gaussian-type functions. A natural question at this point is “How do we compute the Gabor coefficients?”If the Gabor coefficients are not unique.. The procedure can be described by lutions of b m . When TQ< 2n. n ( ~ ) be adjusted by the parameter a in can (14).

N indicates the number of frequency channels. both are time. for an arbitrary 2D function. Although there were some discrete versions of the inverse STFT [40] suggested before. STFT[mT.. i.Except for a [22]. c -Ec. derive the discrete Gabor expansion we as s [i]= which is known as a formal identity. [2S]. in general. MARCH 1999 Z [i+nL]= 0 Oli<L L<i<2L-N 2L O<q<--l. now allow us to map time-domain functions into the joint time-frequency domain.n.[63]. we investigate signals and systems in the time or the frequency domain separately.e. A . A A4 . The key issue of the Gabor expansion is how to compute the dual function y ( t ) for a given h ( t ) .. we require that L is evenly divisible by N and AM.shifted versions of a single prototype function..> a subspace is of the entire set of 2D functions (Fig..[34]. (20) Note that (20) and (16)have exactly the same form. there may be no corresponding signal s ( t ) and window function y(t). The dual function can be solved by AM independent linear systems [47].. 6 ) . we are always able to find the joint time-frequency function C.. (t)=y(t -mT)einRt. Recently.+ = STFT[mT.e. Traditionally. where O<p<L/AM. s(t) -m ..andL denotesthelengthofh[i] and y[i]. ..[37].and frequency. The Gabor expansion pair.2 . such as the relationship between oversampling and perfect reconstruction were not clear. andp are defined as .(t)dt y. i.. By using digital computers. the anafew functions at critical sampling [ 4 ] . Let us rewrite (18) in the regular inner product form.11 h [i-mAM]W.. these relationships were not widely understood until the early ‘80s [3]-[SI. The ratio N / M 4 is considered as the oversampling rate. The relationship between y ( t ) and h ( t )is lytical solutions of (21). the oversampling rate must be more than or equal to one.. then (15) and (19) form a pair of Gabor expansions. The periodic auxiliary function Z[i]is given by 6. However. we can now readily compute the Gabor expansion and apply it to real-world problems. Some important aspects.1. This fact inspired many researchers to seek the discrete version of the Gabor expansion.%. Conversely. which is not directly suitable for numerical implementation. was generally unbounded. The Cabor coefficients are a subset of the set of 2 0 functions. = cs[iB*[ic ”]W.1 (24) where the elements of the matrixA. Usually. The set of { C. N-1 n z . However. the summation of computing the dual function derived previously. (25) 57 IEEE SIGNAL PROCESSING MAGAZINE . Importantly. Discrete Cu6or Expansion Applying the sampling theorem and the Poisson-sum formula [44].. nr (23) where AM denotes the discrete time sampling interval. [63]. For a stable reconstruction. This has been found extremely helpful when the system to be analyzed is time variant. (19) and (lS).. do not exist. one of the major advances in the study of the Gabor expansion is the discovery of the discrete version [33]. An arbitrary 20 time-frequency function may not be a valid C . in fact.._- ky*b-pk k = 0.nQ] where =jw y* .[44]. and vectors 7. In particular. the Gabor coefficient. The number of frequency bins N is a power of two. Let C. nQ]. Equations (19) and (15) indicate that the STFT is.”” n-I) where and c. they were not motivated by expansion and inner product operations. [721. for a given time functions(t) and analysis window function y(t).is determined by the time sampling step T and the length of the analysis window function y( t ). the Gabor expansion can be thought of as the inverse of the STFT.

are inner products ofs[i] and x. the signal energy is usually concentrated in a relatively small region.[i] is inversely proportional to the oversampling rate. For the orthogonal-like Gabor expansion. For critical sampling.. 6 " " 0 0. Note that the difference between h[i] and yo. 7 illustrates the dual function y[i] of the normalized Gaussian functiong[i] in (14) at critical sampling. Moreover. [58].[i] becomes closer to h[z]... the solution of y[i] almost always exists! But. Because (22) and (23) can be realized by the FFT and inverse FFT. As the oversampling rate increases.n] as long as h. the Gabor coefficients C. then the Gabor coefficients C.0 .6 " " 0. the Gabor coefficients C. n]. we have to make y[i] concentrated in both the time and frequency domains. will fail to describe the signal's property in the vicinity of [mAM. . the natural solutions of y[i] are those which are similar to the predetermined h [ i ] .. the Gabor-expansion-based filter banks have been found to be superior to wavelet-based filter banlcs [35]..2 0.. . Dual function y(t) (solid line) for the normalized Gaussian function g(t) (dotted line) at critical sampling. According to (23).5 0. the samples have to be periodically extended.. we can also simply let Z[i+nL]=h[i] In many applic ansion-based filter banks have been found to be superior to wavelet-based r banks.. y[i] is concentrated neither in time nor in frequency. .which reflect the similarity If between s[i] and yWL. Finally..[i]. is the detection and estimation of noise-corrupted signals [70].e. whereb[i] is assumed to be normalized.. The optimization in (27) can also be considered a process to pursue the near tight frame. the discrete Gabor expansion. can be thought of as a set of perfect reconstruction filter banks [54]. using (25)..n]. y." Although the set of {h.7 0.. a=o N-1 .n].y[i] =cb[i].. -E c.8 0. L N (26) In this case... such as subband echo cancellation. the Gabor coefficients C. thenwe can rewrite (22) and (23) as s [i] = m o<q<-. (22) and (23). are unique).2.[i].3 0.[i] is concentrated in [mAM. To ensure that the Gabor coefficients C't. the solution of y[i] is unique. the error in (27) is a function of variance a and a sampling pattern determined by AM and N . The dual relation is described by (24).2[i]}. Note that in this case. It is observed that the minimum error occurs 1441.% indeed characterize a signal's properties around [mAM. Hence.. Once the functionh[i] is selected to be optimally concentrated (as it always is)..for instance. 8 demonstrates h[i] and when a = K/(NAM) corresponding y. On the other hand.9 1 7..When the signal length is equal to the length of h[i]ancl y[i] [63]. Fig.. we can impose any type of e x tension on the samples. y[i] is badly localized. In many applications. Consequently. the length of the original window.[i] . orthogonal-like Gabor expansion.. the regional signal-to-noise ratio (SNR) MARCH 1999 IEEE SIGNAL PROCESSING MAGAZINE .. Ify[i] isclosertob[i].[i]} may not be linearly independent.. It is interesting to linow that the optimal solution of (27) is the minimum norm of (24)! When h[i] is a discrete version of the normalized Gaussian function in (14). the solution of y[i] in (24) is unique (thereby the Gabor coefficients C ..1 0. 58 Time-Varying Filtering One of the most important applications of joint time-frequency representations such as the Gabor expansion. Orthogonal-Like Cabor Expansion When N = AM (critical sampling)....11 h[i-maM]W. While random noise tends to spread evenly into the entire joint time-frequency domain. it is only suitable for finite and small number of samples. no matter how h[i] is concentrated.are indeed the orthogonal projection on {h. t h e i m p l e m e n t a t i o n of t h e Gabor-expansion-based filter banks is very efficient. i. which enables us to perform online processing..4 " 0. a = N/AM. Fig. we name it as an So.will well characterize the signal's behavior in the vicinity of [ m M .L. this method guarantees that: the length of the dual window will always be the same a:.

let us write (22) in matrix form. In this case.e.4 0. such as the ionosphere.5 (b)L=48. However.. err = 0. The problem here is that the Gabor coefficients are a subset of the set of 2D functions. In this case. we can immedately identify the presence of the 1. the impulse signal becomes a nonlinear chirp signal...1 0.8 0.} from the noise background and then reconstruct the time waveform via the Gabor expansion.9 1 0 0.6 0. I f H and G satisfy the dual relation determined by (24). Department of Energy ALEXIS/BLACKBREAD satellite.fl} are no longer in the set of valid Gabor coefficients. Oversampling = 4/3..8 0. N = 8. (31) (a) L = 56. N = 8 . IEEE SIGNAL PROCESSING MAGAZINE 59 MARCH 1999 . AM= 5.6 0. Oversampling= 8/7.4 0..4 0.. when looking at the plot of I C.1 0. 9. i. Fig. there may not be an existing signal that corresponds to it. N = 8.9 (d) L = 32. neither the time waveform nor the power spectrum indicatethe existence of the impulse signal.5 (C) 0.4018 0. 9 depicts the impulse signal received by the U.[i] becomes more and more close to h[i])..can be substantiallyimproved in the joint time-frequency domain [71]. for a given modified {C. Oversampling = 8/5.. Z =E=@GF . AM = 2. F=HS (28) where T.2 0. 1 Oversampling = 2. This observation suggests that we can use a mask to filter out the desired Gabor coefficients {C. are not square matrices except for critical sampling. N = 8. again).6 0... chirp-type signal arching across the joint time-frequency domain. H .1 0 0. (30) Note thatH and G. h[i] (dotted line) and yJi] 0. the modified {C.1628 8. and Gabor coeacients vector.7 0. For the sake of presentation simplicity.2 0.. AM= 7.0865 (solid line) (as the oversampling rate increases.5 L = 40. synthesis matrix. The matrix form of (23) is s=G 5 (29) where G denotes the analysis matrix.3 0. then HG = I and S= H 5= HG S. denote the modified and desired Gabor coefficients. After passing through dispersive media.. that is. Let 5. y. err = 0.3 0. in general. err = 0.S.}.2598 0. Usually..3 0.1 0 -0. While the time waveform is severely corrupted by random noise.7 0. the power spectrum is dominated mainly by radio carrier signals that are basically unchanged over time (Fig. A M = 6. and E denote the signal vector.2 0. err = 0.

[65]. To reduce the computational burden in real applications.where @ is a diagonal matrix whose diagonal elements are either one or zero. that is. 9. =@GF. t o solve t h e LSE. S. which is computationally intensive and memory demanding. it would be difficult to apply an LSE algorithm with conventional PCs. Los Alamos National Laboratory. The result obtained by the first iteration is closer to that computed by the LSE method. Fig. In this example. Fig.. S ~ ~ ~ = ( G ~ G ) . which shows the iteration exponentially converging. and. Zd The sufficient condition for the iteration above to converge can be found in [69]. that is. h[i] = y[i] or critical sampling ({hPn+} forms a basis). the LSE solution often is not the best one. The LSE has been widely used and extensively stud ied for many years [7]. =@GHEl =(@GH)’Ed C.013 0. A common approach is to seek a S opt such that the distance between the Gabor coefficients of Soprand the desired Gabor coefficients zd is minimum.the performance of the orthogonal-like. we n e e d to c o m p u t e t h e pseudoinverse. tt is well known that the solution of (32) is the pseudoinverse of G. SI=WEd. =HE. 11( depicts the reconstructed signal after b) 30 iterations. The procedure can be summarized as follows: I Gabor Coefficients Spectrum t 0 i t I I _I . Under that condition.040 0. But it may not be the best solution for many applications. it is less interesting in applications. However.!) which is known as the least square error (LSE) method. When the data includes over 10. =@G?.027 0. l l ( a ) depicts the error vs. the ionized impulse signal cannot be recognized in either time or frequency domains. minj/zd. h[i] = ~ [ z ] usually implies high oversampling and thereby heavier computation. GF. Although y[i] is hardly exactly equal to h [i] ..) E. Map of LSE filtering. satisfy the sufficient condition [69]. which is impossible to be solved by the LSE de- (33) Fig. In this case. Due to the low SNR. Fig.G S ~ ~ ~ I1 (3.I G ~ E ~ .053 0. GF. =(@GH)h . critical sampling often leads to bad localization of the dual function y[i].J we can readily distinguish it. are not the same as the desired Gabor coefficients Ed . as pointed out earlier. [23]. MARCH 1999 IEEE SIGNAL PROCESSING MAGAZINE . the modified Gabor coefficients are always valid Gabor coefficients. A more effcien t method developed recently is an iterative algorithm in which we continuously apply the same mask and reconstruct the time waveform. . =@GHEd s. In particular. the number of iterations. But. from lC. is equivalent to rapt computed by the LSE method. the iteration converges at the first step. 0. z converges to inside of @. HOWever. 1 0 illustrates the LSE algorithm. Because the modified Gabor coefficientsare generally not valid Gabor coefficients. 11 plots the simulation results for the example illustrated in Fig. = E d . [27]. 60 I- I 2-D Functions I (34) IO. The result of the first iteration S. (Data courtesy of the Nonproliferation and international security division. that is. #Ed (31) which says that the joint time-frequency properties of T. the number of samples is 9000.060 (ms) =HEd El =@GF. they are generally not equal unless the convergence condition [69] is truly met. For critical sampling. In many applications.000 samples. Two trivial cases. Based on Ed we compute a time waveform s. Gabor-expansion-based algorithm has been found to be good enough for most applications. we employ the orthogonal-like Gabor expansion introduced in the preceding section. On the other hand. Note that there is a substantial improvement between the first and second iterations. 9.

a straightforward approach o f c o m p u t i n g t h e time-dependent spectrum is to take the square of the STFT. Obviously. Error (a) and reconstruction (b). the counterpart of the classical Fourier transform. By averaging the instantaneous correlations. ( 3 7 ) can better characterize a signal's properties in the joint time-frequency domain.. The short time duratian window (left) leads to good time resolution and poor frequency resolution. in fact. The main problem of the STFT-based spectrogram is that it suffers from the so-called window effect. i. there is no window effect. as we shall see later.053 0. The iterative algorithm not only yields a better SNR.060 (ms) (b) Reconstruction(After 30 Iterations) 1 1. Noise TF 1 5 10 15 20 25 (a) Number of Iterations vs.. whereas the right plot uses a long time duration window. Seconds 12. the STFT-based spectrogram indicates how the linear chirp signal's frequencies change over time.040 0.e. I I I 1 1 0 0.027 0. Wigner-Ville Distribution A mentioned in the preceding section (12). The goal here is to seek a representation that describes a signal's power spectrum changing over time. On the other hand. The STFT-based spectrogram for a linear chirp signal (The results are subject to the selection of the analysis window function. another important method for the joint time-frequency analysis is the time-dependent spectrum. it has been found inadequate for the applications where both high time and frequency resolutions are required. the left plot has better time resolution (poor frequency resolution). ICk-Ck-. 61 MARCH 1999 IEEE SIGNAL PROCESSING MAGAZINE . Formula (37) is known as the Wigner-Ville distribution (WVD). s the conventional correlation. such as the Gabor expansion and wavelets. or vice versa). a short time duration window function is used.013 0. 12 illustrates the STFT-based spectrogram for a linear chirp signal. While the time waveform (bottom plot) and classical power spectrum only tell a part of the signal's behavior. which was originally developed in the area of quantum mechanics by a Hun- Time-Dependent Spectrum Besides the linear transforms. As shown in Fig. In the left plot. Obviously. i. Although the STFT-based spectrogram is simple and easily implemented.I2 30 We call it the STFT-based spectrogram to distinguish it from the Gabor-expansion-based spectrogram introduced later.z)to replace R ( T ) For instance. 12. the right one has poor time resolution (better frequency resolution). Compared to the STFT-based spectrogram. Analogous to the conventional power spectrum. is the average of instantaneous correlations s(t+z/2)s*(t-~/2). Fig. but also is practical for online implementation.e.scribed by ( 3 3 ) due to the computational complexity and memory limitations. the width of y ( t ) governs the resulting time and frequency resolutions. we lose the time information. How about us? ing R(t.

the WVD is often thought of as a signal’s energy d i s t r i b u t i o n in t h e j o i n t time-frequency domain. the energy ofthe WVD is the Same as the energy content in the signal s ( t ) . in 1932.in atomic nucleus established a method for discovering arid applying the principles of symmetry to the behavior of physical phenomena. Therefore.It is the crossterm interference that prevents the W distribution from b real applications. It is interesting to note that all useful properties of the WVD [i. or vice versa).000 0. there is nothing in the vicinity of 2 kHz during the time interval from 0.) Let S(t)=A(t)e’“’). he was awarded the Nobel Physics Prize.We can prove that the mean frequency of the WVD at time t is truly equal to the signal‘s weighted average instantaneous frequency. Traditionally. we can easily distinguish those two components. Ville 11591. While the Cohen’s class can be thought of as 2D linear filtering of the Wigner-Ville distribution. (38) and (39)]are associated with certain types of averages.015 0.0. it was introduced into the signal processing area by a French scientist J..e. The STFT-based spectrogram for a linear chirp signal (The results ore subject to the selection of the analysis window function.o)dtdo =J. Fig. - WVD(t . the first derivative Cohen’s CIass of the phase is called the instantaneous frequency. let’s look at two time. Wigner [64]. the Cohen’s class and the Gabor spectrogram. an intuitive approach of removing is a single-valued fimction.though it possesses many desirable properties for signal analysis. i. the crossterm is almost always is actually incorrect for several reasons. it is named the crossterm. the first derivative of the phase @(t) represents the weighted average of the instantaneous frequency. (Wigner’s pioneering application of group theory to . The short time duration window (left) leads to good time resolution and poor frequency resolution. Fifteen years later. it does not contribute energy to the signal). In what follows. So.005 0. The problem of the WVD is the so-calledcrossterm interference. 13 illustrates the WVD of the linear chirp signal.e. From the 62 0. i. which As mentioned previously. $’(t) strongly oscillated.01 . the WVD indeed describes the signays joint time-frequency behavior. Although the average of the crossterm in this example is near to zero (in other words. 14. Moreover. the WVD has much better time and frequency resolutions.025 Seconds 12. The STFT-based spectrogram possesses neither (38) nor (39). As an example.e. the Gabor spectrogram is a truncated Wigner-Ville Istribution. 1 I 0.010 0. time waveform and the power spectrum.(s(t)l d t =-j1 2% -- - p(o)/’dw (39) I5 n As a result of this property. Then. However. It is the crossterm interference that prevents the Wigner-Ville distribution from being used for real applications. Eugene P. Because this term reflects the correlation of two signal components [47].025 Seconds IEEE SIGNAL PROCESSING MAGAZINE MARCH 1999 . where both A(t) and $(t)are real-valued time functions.. In 1963. garian-born American physicist.000 0. the magnitude of the crossterm can be proven to be twice as large as that of the signal terms. to reduce the crossterm interference with limited affects on the useful properties. the WVD plot shows a strongly oscillated term in that area. though it possesses many desirable properties for signal analysis. whereas at each time instant:.and frequency-shifted Gaussian functions in Fig.005 0. Essentially. we shall introduce two alternative methods.015 s. 12. generay there are multiple frequencies.020 0.010 0. Compared to the STFT-basedspectrogram in Fig..015 0. For example. 7 .020 0.

The parameter U in (43) denotes the order of the GS. [i9]. Formula (41) was first developed in the field of the quantum mechanics by Leon Cohen [17] and. The amount of energy contained in each individual atom is inversely proportional to the rate of oscillation.n‘ l m . has a closed form that can be saved as a table..nC*wt.(t. Then take the Wigner-Ville distribution on both sides to obtain m(t.and frequency-shifted Gaussian functions defined in (16). Cohen’s class greatly facilitates the selection of our desired transformation. 15.. and [66]..o) concentrated.U)= 1.The prominent members of Cohen’s class include the STFT-based spectrogram. It is worth noting that the energy atom. the Choi-Williams distribution [ 141. the GS only contains those terms in which m = m and n ’ = n’. T) denotes the Fourier transform of @(p.o) in (43). ~ ) denotes the WVD of two time. [47]. [47]. As the order D increases.o) OO + 1 . (t.OE+O 5. . it is named as Cohen’s class. It is interesting to note that almost all previously known time-dependent spectra [38]. Im-m‘l+(n-n’li 11 - MARCH 1999 IEEE SIGNAL PROCESSING MAGAZINE 63 . the GS converges to the WVD. [28].OE-1 1. [46] as introduced in the preceding section. The contribution of each individual energy atom to the signal‘s energy not only depends on its weight. In addition to the true signals. 17 illustratesthe STFT-based spectrogram and Gabor spectrogram for the same bat sound.(t.. we first expand the signal into. symis metrical. 15.o) an energy atom..6E-1 O. the cone-shaped distribution [741. increasingly more less-correlatedcomponents are included.. T).o) Crossterm where WVD. [49] can be written in the form (41).5E-1 2. Based on the closeness of hIn. (t.=(t) hllc. and oscillated.nC*m’.. Comprehensive discussions of Cohen’s class can be found in [16]. For example. @..OE+2 Cm.. but have negligible influence to the signal energy [46l. WVD of the sum of two Gaussian functions. When D goes to infinity. T) [or +(t. Cn2. In other words.e. WVD. EO i 8 I 2. The signal energy in the joint time-frequency domain can be represented in terms of an infinite number of elementary energy atoms.. [18].. WTD.the crossterm is to apply a 2D low-pass filter @( to the o. we only consider the correlation of two identical components.OE-2 1. Fig. they are much closer to the real value in the Gabor spectrogram cases. All those atoms are concentrated. ( t . j .. we rewrite (42) as GS. When D = 0.t) Wigner-Ville distribution.. Because the kernel hnction @(t. The best selection usually is D = 3 4. It can be shown that the energy of WTD. As illustrated in Fig. the resolution improves. If we consider WVD. (43) which is known as the Gabor spectrogram (GS). Cabor Spectrogram Another way of separating a signal‘s components in time and frequency is to apply the Gabor expansion [42]. While the instantaneous frequencies computed by the STFT-based spectrogram are subject to the selection of the analysis window function. then (42) indicates that a signal’s energy can be thought of as the sum of an infinite number of energy atoms. in this case.. Obviously. 16 depicts the WVD of an echo-location pulse emitted by the large brown bat.Fig. When D gets larger. (t... because it is a Gabor expansion-based-spectrogram. Expanding the term of W D in the above equation yields where the function Q(t. o ) ] determines the property of (41).OE-1 14.o)is inversely proportional to the rate of as oscillation. the GS converges to the WVD. and the radial Butterworth kernel representation [67] and [68]. as the order increases. in the and (t) time and frequency domains. the WVD also exhibits strong crossterm interference. symmetrical..n. The highly but also oscdlated atoms are directly associated with crossterm interference.. the adaptive kernel representation [61. Eptesicusfiscus. i. on the oscillation rate. and oscillated.

.The elementary function hk(t)is selected such that the residue is minimum.OE-2 35E-2 4. Moreover. Comparison of GS and STn-based spectrogram... Assume that the given data are samples of the signal s ( t ) . It has been found.OE-2 25E-2 3. WvD. .(t.. When the analyzed signal is discrete.. [45]. (45) which has unit energy. indicates the signal projection on the elementary function hk(t). -< Parametric Time-Frequency Analysis The joint time-frequencp analysis methods presented so far are essentially nonparametric in that no model is assumed to be a priori.OE-2 best performance-a result 16. In principle. i. s h ( t ) in (44) denotes the residue after the k t h approxima- tion. however. such as the Choi-Williams distribution.OE-3 1.o) in (42) being optimally concentrated in the joint time-frequency domain. Consequently. t h a t t h e W i g n e r Ville-distribution-based decomposition (42) has the OOE+O 5. 17. where the coefficientA.. i..9.Jjt(t. there are several parametric approaches. we can also decompose other transforms. the time and frequency resolutions of all those methods are subject to the number of data samples. we don't need to compute it online. weshall brieflyintroduce the signal adaptive approximation and adaptive spectrogram. . where hk(t) denotes a desired signal model. as previhas ously mentioned.5E-2 2.o) closed form. Inwhatfollows. [41].. In contrasr with the nonparametric methods. WVD ofecho-location pulse (Bat data provided by Curtis Condon. [43]. For example.Consequently. 64 IEEE SIGNAL PROCESSING MAGAZINE MARCH 1999 .OE-2 1. via a signal's Gabor expansion. once we obtain the Gabor coefficients C*t.e. the energy atom 17VZl. such as a chirplet.he energy a t o m the Beckman Institute at the University of Illinois). resulting in a tremendously improved computation speed.. the rest of the process is nothing more than repeatedly using the --look-up table. such as the matching pursuit and adaptive spectrogram that have been proposed recently [36].e. Ken White. that is. and AI Feng of o f .z (which can be efficiently computed by the 3%windowed FFT)..

Shie Qiun is a Senior Research Scientist in the Advanced Signal Processing Department at National Instruments Co. But their high performance is at the cost of computational complexity. [45]. In real applications. and finally. Mahesh Chugani. Because the Wigner-Ville distribution conserves energy (39) and is a using relationship (55). When high resolution is a must. Durka. Moreover.we have the adaptive spectrogram as (57) Formula (54) is a signal adaptive approximation.(v4=C14J2 ( C P ) m. which was independently developed by the authors of this article and Mallat et a1 [36]. Like the classical Fourier analysis.com). the mathematical tools of joint time-frequency analysis are also inner product and expansion. The first summation represents the set of autoterms.. to rewrite (44) as (54) and (53) as (55) Taking the WVD on the both sides of (54) yields m. Compared to the nonparametric methods. they are only suitable for certain types of signals. Feichtinger. The material presented in this article appears to have reached a level of maturity for real applications [91. 65 IEEE SIGNAL PROCESSING MAGAZINE . MARCH 1999 Acknowledgments The authors wish to express their deep appreciation to Professors Piotr J. 1431. the Gabor spectrogram is usually a favorite because it is relatively robust and computationally efficient.chen@natinst. The key issue of signal adaptive approximation and chirplet-based adaptive spectrogram [73] is the estimation of the optimal elementary function hb(t)in (50). Summary In this article. Hans G. It can be shown that the residue (t)lJ2 converges as the number of terms Nincreases.~ 5 1~521. Shidong Li.. The chirplet-based adaptive spectrogram has very high resolution without crossterm interference. [551.[601. Many applications may start with the STFT-based spectrogram when a signal‘s frequencies do not change dramatically because the STFT-based spectrogram is simple and easily implemented. and (49) yields (44).(dapang. Dr. Dupang Chen is Manager of the Advanced Signal Processing Department at National Instruments C O . we have .. . [111-[131. they could perform extremely well if a parametric model indeed fits the analyzed signal [57].[41]. Based on (46) and (48). However.we usually neglect the residue when it is small enough.It is easy to see that adding (47). whereas the second indicates the set of crossterms. (48). . [60]. and XangGen Xia for their many constructive comments as well as advice. the parametric algorithms are computationally intensive. (shie. the joint time-frequency algorithms also fall into two categories: h e a r and quadratic. ~ 1 Note that all methods introduced in this section are mainly designed for signals whose frequency contents change rapidly over time. [561. While the STFT and Gabor expansion provide vehicles to map a signal between the time domain and time-frequency domain. the time-dependent spectrum is a powerful tool for understanding the nature of those signals whose power spectra change with time. Usually.qian@natinst. Flemming Pedersen.. we name (57) as a chirplet-based adaptive spectrogram. Dr.(t) has a form (45). we briefly introduced the concepts and methods of joint time-frequencyanalysis. (53) While the STFT and Gabor expansion provide vehicles to map a signal between the time domain and time-frequency domain. The authors would also like to thank their colleague. for his careful reading of manuscripts and valuable suggestions. . the time-dependent spectrum is a powerful tool for understandmg the nature of those signals whose power spectra change with time.. When the elementary function h.com).

[30] A. Acoust. Raz. Lu and J. vol. “ W i p e r distribution decomposition aiid cross-term deleted representation. “Feature analysis of Doppler ultrasound signals obtained from mammalian arteries. 36. 594-598.J. San Francisco. [ 121 V. lnfinn.” in Time-FrequencySignal Analysis: Methods and Applications. February 1980.M. 33. 6.M. no. pp.J. no. pp. 2429-2439.M. pp. 1998. September 1990. 20. “A parametric class of discrcte Gabor expansions. 1952.” Proc. 2 pp. pp. Lett. s. Morris. pp. [ 3 11 A. Rabiner.. no.”IEEE Trans. Chen and S. “A general theory of discrete Gabor expansion. pp.” IEEE Trans. Thesis. 7. Gabor. Birkhauser. and L. [ 171 L. “On thc sliding-window reprcscntation in digital signal p r o cessing. April 1996. pp. vol. August 1985. pp. [27] F.Appl. Part 11: Discrete . Speech. 2. 1 1 pp.” OpticalErgineeiing. Qian aiid J.” 1’h.B. 19-49. [39] F. MD. Signall’rocessifig. 41. 3. vol. vol. 1-11. [45] S. Chen and S.” lEEE Tram Signal Processing. Bastiaans. November 1946. [ 161 T. 3070. 1. [ 131 V. [SI V. 40. 377-394. 41.Qian and D. August 1996. ‘Theory of communication. pp. vol. 538-539. Signal Processing. [33] S. 1990.” IEEE Trans.” Pruc. Theoiy. [ 111 V. 43-54. no. March 1993. 44. no.. 509-5 12. 41.” to appear inJ. pp. pp. Math. Signal Processirg. Hlawatsch and G. 4. and signal analysis. 11. 125-144.R.L. Williams. June 1996. NJ: Prentice Hall.J.” OpticalEngineering. S&nul Pmcessing. vol. vol. 40. “Matching pursuit with time-frequency dictionaries. Chen. vol. Acoust.E. 1980.M. vol.” Proc. 1946.pp. 1994. “Joint time-frequency transform for radar range-doppler imaging. 1997. 2. “Time-frequency-based ISAR image formation technique.Qian.”IEEE Trans. “Gabor representation of generalized functions. “Discrete Gabor transform. vol. vol. 961-1005. 12. June 1989. 1966. “Time-frequency representation of digital signal and systems based on short-time Fourier analysis. no. Porat. IEEE-SP Intemational Symposium on Time-Frequency and Time-ScaleAnalysis. Part 1 1 Relations with other time-frequency signal 1: transformations. July 1993. vol. Lacy. “Time-frequency transform vs. 3397-3415. 7.” IEEE Trans. pp.”/. Jones.. no. 68. Janssen. “Bilinear signal synthesis.F.a tool for time-frequency signal analysis . “Time-frequency distribution-a review.R. 77.” Proc. pp. 1261 D. Rensselaer Polytechnic Institute. 941-981. Signal Processing. “Instantaneous power spectrum. Chugani.K. [ 2I M.Acoust. p p . pp. [43] S. 201-211. D. and optimal wavelet correlator. pp. Rabiner. vol. Page. pp. “Detection of transient signal by thc Gabor representation. ICASSP’98. 2.” Philips/. Res. her 1993. vol. pp. IEEE. vol. RoudreaiwBartcls and T.. Koenig. Boashash. 1983. May 1992. Paris.” Shnal Processing. vol.E. WA. 4.J. 37. Speech. 23. Stromas. Farkash and S. no. CA.” IEEE Signal Processing Magazine. 208-232. [41] S. [6] R. Friedlander and B.W. B. Morris.”J.. “The wavelet transform. Hlawatsch and W. 486-499. vol. April 1998. 1994.. Deccm- [9] V. 2. May 1998. and K. “Gabor Expansion for Echo Cancellation. [20] R.” IEEE Tram. “Optimality property of the Gaussian window spectrogram.M. Speech. “Signal representation in adaptive Gaussian func) tions and adaptive spectrogram. Speecb. “Signal approximation via data-adaptive normalized Gaussian functions and its applications for speech processing. pp. Soc.” IEEE Trans. SHE un WaveletApplications. 37. Acoust. [42] S. 352-363. Choi and W. Speech. 4.155-181. Boudreaux-Bartels. vol. Chen.D. Ed. Baltimore.Aerosp. Englewood Cliffs. “Applications of time-frequency processing to radar imaging. Mulctrate Digital Signal Prucessing.” IEEE Trans. vol. Time-FrequencyAnalysis.Li. [ 14) H . ‘Time-variant filtering via the Gabor expansion.G. Chen. Mecklenbrauker. Janssen. SPIE on AboYithmJ fov Synthetic Aperture k d a r Imapvy. February 1992. [ 191 L.J. Fourier transform for radar imaging. “The Gabor-expansion-based time-frequency spectra. “Radar ambiguity function. no. Appl. February 1989. 34. Amin. vol. 1.. [23] S. vol.. 1152-1161. 24. Daubechies. June 1986. 1977. pp. [35] Y. IEE.” Signal Processing. ASSP-33. January 1991.”IEEE Trans. and Gabor’s expansion of a signal in Gaussian elementary signals. Li and D. Chen. no. IEEE. 66 IEEE SIGNAL PROCESSING MAGAZINE MARCH 1999 .” Signal Processing Magazim. ICASSP’92. 217-250. time-varying matched filter. 2212-2217.” Pruc. [3] M. [ 181 L. March 1999. vol. Crochiere aiid L.. 141-144.” Signal Prucessing I/: Tbeories and Applications. Phys. vol. [38] C. Amer.” Optical Ergineeriizg. Krattcnthaler. time-frequency localization. 2.” IEEE Trans. Bastiaans.. H. [34] S. no.R. Claasen and W. Zhang. vol. Qian. vol. Cohen. GaburAnalysysband Algorithms: Theoiy and Applications. pp.Qian and D. 38. 1589-1602. Chcn. vol. 25. “A signal-dependent time-frequency representation.” Pruc. 665-666. pp. 1558-1564. CA. Signall’rocessing. Chcn. [29] A. “Discrete Gabor expansion of discrete-time signals in 12(Z) via frame theory”. 1995. “Improved time-frequency representation of multicomponent signals using exponential kernels. 18. “Gabor’s expansion of a signal into Gabor clcmentary signals. 389-392. IEEE. pp. Appl. March 1992.”]. pp. IEEE S&al Processing Magazine. New York: Elsevier. 2. 35. 39. Allen and L..Part I: Continuous-time signals.451.D. 9. pp. P ~ J ~vol. Mallat and 2. “CFAR detection and extraction of unknown signal in noise with time-frequency Gabor transform. “A unified approach to short-time Fourier transform analysis and synthesis”. vol. 169-180. [37] J. no. Seattle. Lu. [21] I. 93.E.”/.H. Einzigcr. Pruc. 1992. 1. pp.”IEEE Trans. no. no. SPIE’94 Mathematical Imaging: WaveletApplications. Morris and Y. Syst. 1981. no. July 1989. 36. July/August 1981. 2762. pp. Qian and 1..E.M. pp. March 1992. no. pp. “Signal representation using adaptive normalized Gaussian functions. [36] S. Protnoff. Englcwood Cliffs. Math. no. FourierAnal. Bastiaans. Cohen.References [ I ] J. pp. 21-67. Signal Processing. 781-806. [22] P.Y. 55-69. pp. Acoust. I’edersen. January 1994. 1980. 862-871. February 1996. Chcn. vol. [4] M. Baraniuk and D.J.San Diego. no.M. ‘The sound spectrograph. 6. 59-65. “Generalized phase-spacc distribution functions. 285-294. Healy Jr. [ 101 V. “Duality and biorthogonality for Weyl-Heisenbcrg frame. 1. 34. March. no. Apr. Chen. NJ: Prentice-Hall. July 1994.” IEE Electrun. “Linear and quadratic timc-frequency signal representations. pp. “Time-frequency spectrum analysis and estimation for tionstationary random processes. [44] S. vol. no. “Time-varying filtering and signal estimation using Wigiier distribution synthesis techniques.” Pruc.” l’roc. [SI 1M. 7. Australia: Wiley Halstcd Press. “Gabor expansion of an aperture field in exponential elementary beams. France. 202-204.” l’ruc. no.” IEEE Trans. [24] H. vol.Acoust. [28] F. Janssen. [ 71 G.”/. 65. [25] B. 429-457. Signal Processing. pp. 1. 7. Anal. 1988. The 27th Annual Conference on Information Sciences and Systems. no. March 1994. Parks.J.F. Chen. “A sampling theorem for the complex spectrogram. Signal Processing. [32] R. Elecwun. Signal Processing. 442. [40] M. Dunn. 103-106. no. [ 151 M. pp.Qian.time signals. Cohen. 868-873. 28. “The Wigner-distribution . April 1997. Feichtiiiger and T.

Qian. and R. no. [66] W. Scher. Zhou. October 1994. [57] L. 317-318.110. 2. Feichtinger.. 45. NJ. Math. L. and A.P. M. July 1990. Vagnucci. and R. A. Dehl. vol. CA. pp. Ryan. Sun.H. Jasiukaitis.J. [63] J. Williams. “Time-frequencyanalysis of high-frequencyactivity at the start of epileptic seizures. andMS.” IEEE Trans. 1992. 8. Boashash. [73] Q. 43.” Proc. 1302-1311.” Proc. Englewood Cliffs. [68]D. “Joint time-frcquencyISAR using adaptive processing. vol. 2836-2841. S.” IEEE Trans.. [67] D. [53] R. May 1998.M. and M. Wellesley.” Ph. no. [70] X. 7. Krieger. 42.J. “Discrete Gabor structure and optimal representation. N. Zhao. no. B. Signal Processing.”IEEE Trans. vol. Sun. vol. C. [49] W.. pp. 1968. ISSP’90. Yao.” IEEE Trans. 2258-2268. 45. Chugani. and D. Wang. Radio Engineers. 2. X.pp. Atlantic City. %a. Chugani. ‘Time-frequencyanalysis of the EEG signal. [69] X.E. Raz. 1173. MARCH 1999 IEEE SIGNAL PROCESSING MAGAZINE 67 . M. Wilcox. to appear. “Time-frequencyrepresentations using a radial Buttenvorth kernel. vol. 61-74.L. Macek. 2. 21.Antennas Propajat. “Complete Gabor transformation for signal representation. Electron. “ISAR motion compensation via adaptive joint time-frequencytechnique”.November 1996. Rev. pp. 1084-1091. Philadelphia. 9. vol. 3. October 1995. [62] T.vol.B. SignalProcessinj. S. 2. Summay Rep.” Proc. IEEE.-G. M. 1990. pp. Image ProcessinJ.England: Wiley Halsted Press. [71] X. NJ: Prentice-Hall. and V. “Digital signal processing aids cholesterolplaque detection. Li and A. pp.” Proc.D. April 1993. pp. 228-230. [65] C. Sclabassi. vol.D. Yan. 10. Wang and J.” IEEE Trans. “A quantitative analysis of SNR in the short-time Fourier transform domain for multicomponent signals. on Sgnal Processing. 369-374. “Detection of stenotic plaque in arteries. Vaidyanathan. Z. Sun. 1948. “Signal energy distribution in time and frequency. [52] R.” in Time-Frequency Signal Analysis:Methods and Applications. Australia. (Chinese). Thesis.Wang. pp. ICASSP’95. 498-519. no. April 1960.C. I511 M.. “Time-frequencysignal analysis via Coheil’s class of distributions and implementation algorithms. pp. rium. M. P.Q. Waveletand Filter Banks. [50] M. “Chirplet-basedsignal approximation for strong earthquake ground motion model. Rihaczek. IEEE-SI]Zntemational Sympon’nm on Time-Freyuencyand Time-ScaleAnalysis. Res. to be published.S. 25. August 1997. 10th International Conferenceon Pattem Recognition. Husain. Yin.”Proc. “The use of cone-shaped kernels for generalized time-frequencyrepresentations of nonstationary signals.N. vol. NJ: Prentice Hall. S&al P m ~ ~ e ~ ~ i n g .Qian and D. Sclabassi. Scheuer. Sgnal l’messing. 749. pp. 1591 J. Sclabassi. Ville. Wigner.J. 15th Annual International Conference ofrbe lEEE Engineering in Medicine and Siolay Society. pp. Savic. M. 60-63. Universityof Maryland at Baltimore. Chicago. 52-58. June 1990. S. 94. January 1998. X:G. “Localizingfunctional activity in the brain through time-frequencyanalysis and synthesis of the EEG. Baumann. pp. 38. vol. S. October 1994. D. IEEE. IEEE EMBC’97. Macek.[46] Qian and D. Sknal Processing.N. 46. 1993. ChenJoint Time-FreyuencyAna&. vol. Wu and J. and time-frequencydistribution series’’ IEEE Trans. D. 1995. no. pp.” in Proc. “System identificationusing chirp signals and time-varyingfilters in the joint time-frequencydomain. Qiu and H.P. May 1995. 84.Implementations and Applications. P. Longman-Cheshire. Sun. November 1990. H. M.Acoust. 1641 E.C. Wexler and S. Sclabassi.P. Strang and T. 1264-1280. University of Wisconsin at Madison. Nguyen. vol. 10. “Decomposition of the Wigner-Villedistribution S. [74] Y. 157. Inst. Tang. Center.J. Peabody. [60] J. MI. Krieger. IL. October 1993. Xia.vol. Wu. M. [72] J. “On the quantum correction for thermodynamic equilib. February 1997.Multirate Systems and fiilter Banks.” this issue.Ling.-G. Marks. 221-227.”]. Tang. Chen.G. 200-203. Englewood Cliffs. “Discrete Gabor expansions. pp. Scher. September 1996. pp. and R. [47] S.4. Atlas. P. Trintinalia and H.” IEEE Trans.” Ttch. S a .1176. Theories. no. 935-942.-G. pp. April 1997.” Phys. Qian. 207-221. 2072-2084. [54] G. AES-34. Baumann. Husain. Ni. Jasiukaitis. [61] Y. “Reduced interferencedistributions: biological applications and interpretations. “Time-frequencydomain problems in the neurosciences. rr-14. Gold Coast. “Classification of abnormal cortisol patterns by features from Wigner spectra.Chen. vol. “Thovrie et applicationsde la notion de signal analylique. pp. Savic. 85.”Signal I’rocessing. 9. Qian.” (in French) cables et Transmission. Qian. no.” IEEE Trans. Xia and S . R. pp. and Z. “Adaptive oriented orthogonal projective decomposition. October 1997. [56] M. Sun. 1932. Signal Processing. pp.J. IEEE Trans. 40. Ling. [55] M. PA.” in Proc. 152-159.”IEEE Trans. M.” l’roc. Detroit. “The synthesis problem for radar ambiguity functions. Speech. MA: Wellesley-Cambridge Press. Ed. San Diego. [48] S. no. [58] P. Chen. pp. S. 1996. K. Adelson. pp. “An iterative algorithm for time-varyingfiltering in the discrete Gabor transform domain. Morris.

- Rfc 1928
- Klopf - Brain Function and Adaptive Systems
- AM Thesis Kurt Peek
- Bedienungsanleitung Fahrradcomputer SX4
- Dalfsen et al. (2007) - VELMOD-2 (2007).pdf
- Albertin - The Time for Depth Imaging
- Castagna - Principles of AVO Crossplotting
- 50-Years Groningen - Symposium Handout
- Claerbout - Dip and Offset Together
- Assignment 13
- 6 Permian Triassic
- SCHWIND Diagnostic Systems 2013-E
- Davosklostersguide Map Feb10
- Davos Klosters Piste Map One
- Davos Klosters Piste Map Two
- 8. Geiss TOTAL - Late Charge in the K5 Area
- Gentile 40
- Apr04 Seismic Forward Modeling
- Veldhoven Aug 2005 Jansen
- TeXworks Manual r814
- Compress Ives Amp Ling

Sign up to vote on this title

UsefulNot useful- An Introduction to Time-Frequency Analysis
- OMKAR (AUM) Chanting-Time Frequency Analysis
- Principles of Underwater Acoustics
- Hot Topics in Animal Bioacoustics
- Time-Frequency Toolbox GNU Tutorial
- Time frequency analysis of speech signals
- Underwater Wireless
- A Human Error Analysis of Commercial Aviation Accidents Using the Human Factors Analysis and Classification System (HFACS)PDF
- Aviation Quotes
- Itroduction to Discrete Time Signals
- Airbus-Commercial_Aviation_Accidents-April2014.PDF
- 16references.pdf
- Bioacoustic Characteristic of Male Dolphins Bottle Nose (Tursiops aduncus )
- Underwater Wireless Communication Networks Seminar
- Introduction to Underwater Acoustics
- MIMO-OFDM-ThesisPdf
- Federal Aviation Administration (FAA) 2006 General Aviation Accident Data.
- Airbus Commercial Aviation Accidents 1958 2014
- USFS Aviation Accident History
- Adaptive Filter Theory Slide6 1
- An introduction to signal detection and estimation 1988
- Underwater Acoustics
- Seminar on Underwater Wireless Communication
- Aviation Accident Investigation
- OSD_T1001_-An_Introduction_to_Underwater_Acoustics-Updated-1.pdf
- Introduction to Underwater Comms
- Joint Time-frequency Analysis