Denoising Gabor Transforms

Transcription

1 1 Denoising Gabor Transforms James S. Walker Abstract We describe denoising one-dimensional signals by thresholding Blackman windowed Gabor transforms. This method is compared with Gauss-windowed Gabor threshold denoising and wavelet-based denoising, and is found to be superior to those methods in many cases. For signals that are typical in audio denoising, this new method is particularly effective. It provides the first step in developing an adaptive denoising method for non-stationary noise. Index Terms Signal denoising, Gabor transforms, thresholding. EDICS: DSP-RECO INTRODUCTION The Gabor transform is a classic method of timefrequency analysis, [1], [7], [8], [10]. Recent work has focused on its advantages for discrete signal processing [2], [4], [14], [11], [12], and [13]. The work in [4], in particular, applies discrete Gabor transforms to denoising. The window function used in [4], however, is the non-compactly supported Gaussian function (the one introduced by Gabor). The non-compact support of the window requires elaborate techniques, such as periodization and the Zak transform, to be brought to bear in the discrete realm. We study in this paper a compactly supported window, the Blackman window, for discrete Gabor transforms and compare its denoising performance with the method proposed in [4] and with the wavelet-based methods of SureShrink [6] and BayesShrink [3]. We find that it outperforms the method in [4] for almost all tested signals and James S. Walker is with the University of Wisconsin Eau Claire. outperforms the wavelet methods on many tested signals (see Tables II and III). The method proposed here, of thresholding Blackman windowed Gabor transforms has several antecedents. For example, in [9] and in [15], thresholding of Gabor transforms was performed for denoising. Those workers, however, employed iterative approaches and did not use compactly supported Blackman windows. There may be applications where speed is critical and the one-step, compactly supported, approach described here would be advantageous. We also have developed a local estimator of the noise standard deviation which we believe to be new. In section I we define the discrete Gabor transform with a Blackman window and derive our denoising method, we also state a theorem on convergence of the proposed method and provide some justification for using a Blackman window. Section II contains an objective comparison, using Mean Square Error (MSE), of the performance of the proposed denoising method with the other denoising methods mentioned above. In Section III we prove the convergence theorem stated in section I. We conclude with a summary and directions for future research. I. DERIVATION OF THE METHOD We assume that the signal f(t k ), for uniformly spaced values t k, satisfies the following additive noise model: f(t k ) = g(t k ) + n k (1)

2 2 where g is the underlying signal that is corrupted by the additive noise {n k }. To keep things simple in this initial proposal of our new method, we assume that the underlying signal g is piecewise regular. By piecewise regular, we mean that g and its first two derivatives are all bounded with only a finite number of discontinuities. We shall also assume that the values {n k } are governed by independently identically distributed Gaussian normal distributions of mean 0 and standard deviation σ. A Gabor transform of f, with window function w, is defined as follows. We begin by multiplying {f(t k )} by a sequence of shifted window functions {w(t k τ m )}, producing a sequence of time localized subsignals, {f(t k )w(t k τ m )} M m=1. Uniformly spaced time values {τ m } M m=1 are used for the shifts. The windows {w(t k τ m )} are all compactly supported and overlap each other. Because w is compactly supported we treat each subsignal {f(t k )w(t k τ m )} as a finite sequence and apply an FFT to it. Denoting the FFT by F, we have the Gabor transform of {f(t k )}: {F{f(t k )w(t k τ m )}} M m=1. (2) It is important to note that when the values t k belong to a finite interval, we shall always evenly extend signal values from each endpoint, so that the full supports of all the windows are included. The Gabor transform for the proposed denoising method uses a Blackman window defined by w(t) = cos( 2πt λ ) cos(4πt λ ) for t λ/2, and w(t) = 0 for t > λ/2, for a positive parameter λ chosen to provide the width of the window where the FFT is performed. Another popular window is a Hanning window. We will focus on the Blackman window w since it generally gave slightly lower MSEs (although the differences were so insignificant that we have chosen not to report them), and because the spectrograms of some denoised signals show more leakage with Hanning windowing than with Blackman windowing. The Fourier transform of the Blackman window is very nearly positive (with negative values less than 10 4 in size), thus providing an effective substitute for a Gaussian function (which is wellknown to have minimum time-frequency support). When displaying a Gabor transform, it is standard practice to display a plot of its magnitude-squared values, with time along the horizontal axis, frequency along the vertical axis, and darker pixels representing higher square-magnitudes. We shall refer to such a plot as a spectrogram. On the left of Fig. 3, we show a spectrogram of a noisy signal. Its denoised spectrogram, using the proposed thresholding method, is shown on the right. A denoised signal is obtained from this latter data by applying the inversion process for Gabor transforms, which we now review. For the Blackman windows used in this paper, we shall always have the windows overlapping to a great extent. In fact, there is a constant A > 1/2 such that A M w 2 (t k τ m ) (3) m=1 holds for all points t k. We use (3) to invert the Gabor transform in (2). First, apply inverse FFTs to all the transforms {F{f(t k )w(t k τ m )}} M m=1 to obtain {{f(t k )w(t k τ m )}} M m=1. Then multiply each subsignal {f(t k )w(t k τ m )} by {w(t k τ m )} and sum over m, obtaining {f(t k ) M m=1 w2 (t k τ m )}. Multiplying by [ M m=1 w2 (t k τ m )] 1, which is no larger than A 1, we obtain {f(t k )}. Thus, we can always invert our Gabor transforms. We now recall the effect of applying an FFT on the Gaussian noise n. The FFT F that we employ is given by, for all sequences {a k } : {a k } F {A l = a k e i2πkl/n }

3 3 where N is a fixed number of points. From Parseval s equality a k 2 = 1 N l=0 A l 2 (4) we conclude that F is N times a unitary transformation. It is well-known that for all random signals, {x k }, the real part of the FFT, {R(X l )}, and its imaginary part, {I(X l )}, satisfy (on average, in terms of expected value): l=0 R(X l ) 2 = l=0 I(X l ) 2 = 1 2 l=0 X l 2. (5) From equations (4) and (5), we conclude that both RF and IF can be expected to behave on random signals as N/2 times orthogonal transformations. Hence {RF(n k )} and {IF(n k )} will be normally distributed with mean 0 and standard deviation σ N/2. We now examine the effect of multiplying the noise signal n by the Blackman window w. By repeatedly computing the mean and standard deviation of nw, we concluded that the mean of nw is still near 0 (always less than 1/100 the standard deviation) and the standard deviation is approximately 0.55σ. Theoretically, we explain this result as follows. The expected value of the variance estimation 1 N 1 w(t k ) 2 n 2 k w(t k ) 2 n2 k N is w(t k) 2 σ 2 /N, a Riemann sum approximation of 1/2 1/2 w(t)2 σ 2 dt = (0.55) 2 σ 2. Similar calculations show that the expected value of the mean is approximately 0 as well (we made use of this fact in the preceding sentence). Thus, we expect that the standard deviation of nw is approximately 0.55σ, and its mean is approximately 0, as confirmed by our repeated numerical experiments Fig. 1. Left: Average of 5 histograms of Gaussian random noise, 512 samples with variance 1. Right: Average of 5 histograms of real part of FFT of Blackman windowing of the random noise data, calculated variances were all approximately 78, and 78 (0.55) 2 512/2. Since F converts multiplication into convolution and w is an even function, we conclude that RF(nw) and IF(nw) are convolutions of RF(n) and IF(n) with F(w). Since F(w) is of very narrow width having non-zero values of 0.42N, 0.25N, and 0.04N only at the frequency indices l = 0, ±1, and ±2, respectively with a mean over N indices of 1, we shall assume that these convolutions are still normally distributed. The means of RF(nw) and IF(nw) will be approximately 0 and their variances will be approximately (0.55) 2 σ 2 N/2. For example, in Fig. 1, we show histograms illustrating this result. The averaged histogram of the noise data approximates a normal distribution, and so does the averaged histogram of the real part of the FFTs. Furthermore, we find that the calculated variances for the real part of the FFTs of the noise data were all about 78 (0.55) 2 N/2 when N = 512. Similar results obtain for the imaginary parts of the FFTs. With these results in hand, we can derive our method of noise removal. First, we note that the normalized data RF(nw)/0.55σ N/2 is N(0,1). Hence, with probability 1 as N, all of their values will be smaller in magnitude than log N. (In 0

4 4 fact, because of the rapid decrease to 0 of e x2 /2 it is almost certain that all N magnitudes lie below log N for finite N.) Similarly, we expect that all (certainly all as N ) of the magnitudes of IF(nw)/0.55σ N/2 lie below log N. We conclude that for a threshold of T = 0.55σ N log N we expect that all of the Gabor transformed noise values will have their magnitudes no larger than T. Based on this conclusion, our method of noise removal is the following four-step process: 1) Compute a Blackman-windowed Gabor transform of the noisy signal {f(t k )} using N = c N s length FFTs (see Table I); 2) Obtain σ, an estimate of the standard deviation σ of the noise (see Remark 1 below); 3) Apply a hard thresholding to this Gabor transform using threshold T = 0.55 σ N log N, replacing all Gabor transform values of magnitude less than or equal to T by zero-values; 4) Apply the inversion procedure to the thresholded Gabor transform, producing the denoised signal. Remark 1: We found σ by the following process. First, for each of the first 20 windowings, we calculated a median of the imaginary parts of the highest quarter frequency Gabor transform values. Second, we computed the arithmetic mean of those 20 values. Finally, we divided that average median by (N/4). (Using a median estimator on highfrequency data of orthogonal transforms has been shown to be a good method by other workers, see e.g. [5] for such a formula for wavelet transforms.) We have found that our estimator always correctly predicts the standard deviation to within ±10% for all the test functions used in the simulations described below (within ±1% for the largest value N s = 8192). Moreover, both our test denoisings and our proof of convergence show that such an estimator is perfectly adequate. We also note that this estimator is local, using only a portion of the signal, which sets the stage for further development of a denoising procedure for non-stationary noise. The proposed method does converge for a wide variety of signals as evidenced by the following theorem. Theorem 1: Let f(t k ) = g(t k ) + n k where g is piecewise regular on [0,1] and {n k } are i.i.d. normal random variables with mean 0 and standard deviation σ. The expected value of the MSE of a Gabor- Blackman thresholded denoising f converges to 0 at the rate O(1/ N s ) as the number of samples N s tends to. We will prove this theorem in Section III. For a signal containing N s sample values, the proposed method is of O(N s log 2 N s ) complexity. This follows from the fact that each FFT has log 2 N s ) complexity. Hence our proposed method is a fast procedure. The method is also practical in applications like streaming audio, since the compact support of the Blackman windows requires only that a buffer of length O( N s ) be available for computing and thresholding the spectrogram as the data streams in. O(N 1/2 s II. COMPARISON OF DENOISING METHODS In this section, we compare the proposed method, referred to as Hard Gabor (Blackman), with three other methods: the hard thresholding method of [4], which we shall call Hard Gabor (Gauss), and two wavelet based methods, SureShrink [6] and BayesShrink [3]. We compare MSEs for six test signals using data for Hard Gabor (Blackman) and SureShrink reported in [4], and data for BayesShrink reported in [3]. The six test signals are called Bumps, HeaviSine, Doppler, Blocks, QuadChirp, and MishMash in [6]. For the three sample sizes, N s = 512, N s = 2048, and N s = 8192, these signals were obtained from uniform samples over the interval 0 t < 1. For the first four of these signals, Theorem 1 applies, we

5 5 Fig. 3. Left: Spectrogram of noisy QuadChirp, 2048 points. Right: Spectrogram of its thresholded Gabor transform. Fig. 2. Top: Bumps function, 8192 points. Middle: Bumps function with additive noise. Bottom: Denoising using Hard Gabor (Blackman) method. Samples N s FFT-length N Index shift τ/ t N s c N s, c = 4 or 4 2 c N s/ TABLE I PARAMETERS FOR HARD GABOR (BLACKMAN) DENOISING. discuss how to extend Theorem 1 to cover QuadChirp and MishMash in Remark 2 below. In each case, the signals were normalized by multiplying by constants so that the standard deviation of the samples {g(t k )} was always 7. For all of the denoisings, the noisy data {f(t k )} was generated by adding Gaussian normal noise {n k } of standard deviation 1. One example of this test data is shown in Fig. 2 for the Bumps function with N s = 8192 samples. Notice how well the Hard Gabor (Blackman) method denoised this example, producing a very satisfying, noise-free reconstruction of the true signal (with MSE 0.015, a 67-fold noise reduction). A second example is shown in Fig. 3. The spectrograph at the left is for a noisy QuadChirp signal using N s = 2048 points. The spectrogram of its thresholded Gabor transform is shown on the right. Again, the proposed method provides an excellent denoising (with MSE 0.102, a 10-fold noise reduction). In Table I we show the values of the parameters for the proposed method, where N s is the number of sample points, N is the number of points used in the FFT F, and τ/ t is the magnitude of each index shift between successive starting points for the windows w(t k τ m ). To maximize speed, the parameter c was chosen to yield power-of-two FFTs. Note that the parameters in Table I yield highly oversampled transforms, by factors ranging from 8 to 15. The Hard Gabor (Gauss) method reported in [4] used only 2-times oversampling. On the other hand, it used a pre-set exact value of σ rather than an estimate from the noisy data. In Table II, we report the results of our comparison of MSEs for the three methods Hard Gabor (Gauss),

6 6 SIGNAL SAMPLES GAB-G SURSHR GAB-B Bumps HeavySine Doppler Blocks QuadChirp MishMash TABLE II AVERAGE MSES ON DENOISING SIX TEST SIGNALS. GAB-G IS HARD GABOR (GAUSS), SURSHR IS SURESHRINK, AND GAB-B IS HARD GABOR (BLACKMAN). SureShrink, and the proposed Hard Gabor (Blackman). Averages were taken using 100 realizations of noisy signals. These results indicate the value of the proposed method, it outperforms both Hard Gabor (Gauss) and SureShrink on two-thirds of the data. There is a second wavelet-based denoising method, BayesShrink [3], which has been shown to be generally superior (in terms of smaller MSEs) to SureShrink. In Table III we compare MSEs of BayesShrink and the proposed method. Averages were taken using 1000 realizations of noisy signals. Here we see that the two methods are roughly comparable; for half the data BayesShrink provides lower MSEs, while for the other half Hard Gabor (Blackman) provides lower MSEs. It should be noted that our proposed method is extremely simple it does not employ advanced Bayesian statistical modelling like Signal BayesShrink Hard Gabor (Blackman) Bumps HeaviSine Doppler Blocks TABLE III AVERAGE MSES FOR DENOISINGS (N s = 1024) BayesShrink hence there is room for improvement. III. PROOF OF THEOREM 1 For simplicity of notation, we assume that the constant c (see Table I) is 1. Hence N = Ns 1/2. Since Equation (7) below shows that the upper quarter of the Gabor transform coeefficients tend to 0 at an inverse square rate, it follows that the estimator for σ described in Remark 1 converges rapidly to σ as N s. Therefore, we shall use the exact value σ in the calculations that follow. For values of t k separated from any discontinuities of g, g and g, as N s the inversion process is adding bounded multiples of the centered value t k = τ m. The signal g has Gabor transform values F{g(t k )w(t k τ m )} satisfying F{g(t k )w(t k τ m )}[l] t N s F s {g( )w(t)}[l]. Ns Here F s denotes Fourier series coefficients over the interval [τ m 1/2,τ m +1/2] and the window w on the right is using λ = 1. This formula was obtained by viewing the FFT as a trapezoidal rule approximation of a Fourier series integral [with a tolerable error of O(1) after thresholding]. We use t g( ) =. g( τ m ) + 1 g ( τ m )(t τ m ) Ns Ns Ns Ns

7 7 to obtain for l 2 F{g(t k )w(t k τ m )}[l] g(τ m )F s (w)[l]o(n 1/2 s ) (6) and for l > 2 F{g(t k )w(t k τ m )}[l] O(l 2 ). (7) On the other hand F{n k m } satisfy RF{n k m }[l] = N(0,0.55σNs 1/4 / 2)[l] IF{n k m }[l] = N(0,0.55σNs 1/4 / 2)[l]. Combining these last two equations with (6), we see that with probability 1 as N s, all Gabor transform values with l > 2 are asymptotically noise-dominated. Therefore, a hard thresholding, using threshold T = 0.55σNs 1/4 log(ns 1/2 ), is sufficient for large enough N s to remove all noise values (and signal transform values) with l > 2. [Note: Also all the O(1) errors for l > 2 in our first approximation are removed. Thus, at most, an O(1/N s ) contribution to MSE results from the remaining O(1) errors.] Thus, the magnitude of signal error at τ m due to losing Gabor transform values of the signal for l > 2 is bounded by (when performing an inverse FFT): 1 Ns O(l 2 1 ) = O( ). Ns Ns l >2 The MSE due to this loss is then bounded by 1 N s N s k=1 1 O( ) 2 = O( 1 ). Ns N s For l 2, we proceed as follows. Using (7), and assuming that all signal transform values are used [we just showed that neglecting them for l > 2 contributes only O(1/N s ) to MSE], we need only estimate for l 2 the contribution to MSE of the noise values {ñ k m } defined by ñ k m = 1 Ns 2 F{n k m }[l]e i2π(k m)l/ N s. l= 2 They all have standard deviation O(1/Ns 1/4 ), because RF{n k m } and IF{n k m } have standard deviations of 0.55σNs 1/4 / 2. Hence their contribution to expected value is E( 1 1 ñ2 N k m ) = O( ). s Ns Finally, the perturbation due to the finite number of transform values ignored by considering only the O( N s ) values t k away from the any discontinuities of g, g, and g, contributes at most O[(log N s ) 2 /N s ] to MSE by similar considerations. That completes the proof of Theorem 1. Remark 2: (a) The proof reveals that the greatest contribution to MSE comes from the extremely low frequency data for l 2. More research is needed to find a way to reduce MSE in that case. (b) Although Theorem 1 does not apply to the QuadChirp and MishMash data (which provide discrete models of distributions), a straightforward modification of its proof shows that the expected value of MSE is also O(1/ N s ) for those signals. CONCLUSION We found that our new method for denoising onedimensional signals, Hard Gabor (Blackman), outperforms the method introduced in [4] and SureShrink for two-thirds of tested data and outperforms BayesShrink on a significant set of data. The easy interpretation of the spectrograms, which can be created as part of the denoising method, is an added feature. Future research will focus on several questions: (1) examining soft thresholding (also called shrinkage

8 8 [5]); (2) adapting our method to handle non-stationary noise (the Gabor transform is highly local, for each windowing only a partial sequence of the data is transformed; thus allowing for adaption to changes in the noise probability distribution); (3) extending the proposed method to non-uniformly sampled signals (the description of the Gabor transform and its inverse in Section I applies to non-uniformly sampled signals); (4) development of a model for separating noise from signal at extremely low frequencies (see Remark 2); (5) extending our convergence theorem to other signal data (such as discrete models of distributions). [13] S. Qian, K. Chen, and S. Li, Optimal biorthogonal functions for finite discrete-time Gabor expansion. Signal Processing, 27 (1992), [14] J. Wexler and S. Raz, Discrete Gabor expansions. IEEE Trans. Signal Processing, 21 (1990), [15] X.-G. Xia, System identification using chirp signals and time-variant filters in the joint time-frequency domain. IEEE Trans. on Signal Processing, 45 (1997), ACKNOWLEDGEMENTS We thank Peter Balazs, Karim Drouiche, and Truong Nguyen for their encouragement and helpful suggestions. REFERENCES [1] M.J. Bastiaans, Gabor s expansion of a signal into Gaussian elementary signals. Proc. IEEE, 68 (1980), [2] M.J. Bastiaans and M.C.W. Geilen, On the discrete Gabor transform and the discrete Zak transform. Signal Processing, 49 (1996), [3] H.A. Chipman, E.D. Kolaczyk, and R.E. McCulloch. (1997). Adaptive Bayesian wavelet shrinkage. J. American Statistical Association, Vol. 92, [4] O. Christiansen, Time-frequency analysis and its applications in denoising. Thesis, Department of Informatics, University of Bergen. (2002). Available at [5] D. Donoho et al. (1995). Wavelet shrinkage: asymptopia? J. Royal Stat. Soc. B, Vol. 57, [6] D. Donoho and I. Johnstone, Adapting to unknown smoothness via wavelet shrinkage. American Statistical Association, 90 (1995), [7] H. Feichtinger and T. Srohmer [Eds.], Gabor Analysis and Algorithms. (1998) Birkäuser, Boston, MA. [8] D. Gabor, Theory of communication. Journal of the Institute for Electrical Engineers, 93 (1946), [9] D.W. Griffin and J.S. Lim, Signal estimation from modified short-time Fourier transform. IEEE Trans. on Acoustics, Speech, and Signal Processing, 32 (1984), [10] K. Gröchenig, Foundations of Time-Frequency Analysis. (2001) Birkhäuser, Boston, MA. [11] S. Qian and D. Chen, Discrete Gabor transform. IEEE Trans. Signal Processing, 41 (1993), [12] S. Qian and D. Chen, Optimal biorthogonal analysis window function for discrete Gabor transform. IEEE Trans. Signal Processing, 42 (1994), James S. Walker received his DA from the University of Illinois, Chicago, He has been a Professor of Mathematics at the University of Wisconsin Eau Claire since He has published papers on Fourier analysis, wavelet analysis, complex variables, and logic, and is the author of four books on Fourier analysis, FFTs, and wavelet analysis.