Документ взят из кэша поисковой машины. Адрес оригинального документа : http://www.stecf.org/conferences/adass/adassVII/reprints/murtaghf.pdf Дата изменения: Mon Jun 12 18:51:47 2006 Дата индексирования: Tue Oct 2 03:16:24 2012 Кодировка: Поисковые слова: trifid nebula

Astronomical Data Analysis Software and Systems VII ASP Conference Series, Vol. 145, 1998 R. Albrecht, R. N. Hook and H. A. Bushouse, eds.

Noise Detection and Filtering using Multiresolution Transform Metho ds
Fionn Murtagh (1,2) and Jean-Luc Starck (3) (1) Faculty of Informatics, University of Ulster, BT48 7JL Derry, Northern Ireland (e-mail fd.murtagh@ulst.ac.uk) (2) Observatoire Astronomique, 11 rue de l'Universitґ, 67000 Strasbourg e Cedex, France (e-mail fmurtagh@astro.u-strasbg.fr) (3) CEA, DSM/DAPNIA, CE-Saclay, 91191 Gif-sur-Yvette Cedex, France (e-mail jstarck@cea.fr) Abstract. A new and powerful methodology for treating noise and clutter in astronomical images is described. This is based on the use of the redundant ` trous wavelet transform, and the multiresolution support a data structure. Variance stabilization and noise modeling are discussed. A number of the examples used in the presentation are shown.

1.

Introduction

Noise is of fundamental importance in astronomy. In practice we must model it. In this article we present a powerful framework for doing this. Information on the detector's noise properties may be available (Snyder et al. 1993, Tekalp and Pavloviґ 1991). In the case of CCD detectors, or digitized c photographic images, additive Gaussian and/or Poisson distributions often provide the most appropriate model. If the noise model is unknown, e.g., due to calibration and preprocessing of the image data being difficult to model, or having images of unknown detector provenance, or in a case where we simply need to check up on our preconceived ideas about the data we are handling, then we need to estimate noise. An algorithm for automated noise estimation is discussed in Starck and Murtagh (1997). 2. Multiresolution Support

This describes in a logical or boolean way if an image I contains information at a given scale j and at a given position (x, y ). If M (I )(j, x, y ) = 1 (or = tr ue), then I contains information at scale j and at the position (x, y ). M depends on several parameters: · The input image. · The algorithm used for the multiresolution decomposition. · The noise. 449


450

Murtagh and Starck

· All constraints which we want the support additionally to satisfy. The multiresolution support of an image is computed in several steps: · Step 1 is to compute the wavelet transform of the image. · Booleanization of each scale leads to the multiresolution support. · A priori knowledge can be introduced by modifying the support. The last step depends on the knowledge we have of our images. E.g., if no interesting ob ject smaller or larger than a given size in our image, we can suppress, in the support, anything which is due to that kind of ob ject. 3. Multiresolution Support from the Wavelet Transform

The ` trous algorithm is described as follows. A pixel at position x, y can a be expressed as the sum all the wavelet coefficients at this position, plus the smoothed array:
p

c0 (x, y )= cp (x, y )+

j =1

wj (x, y )

(1)

The multiresolution support is defined by: M (j, x, y )= 1 0 if wj (x, y ) is significant if wj (x, y ) is not significant (2)

Given stationary Gaussian noise, if | wj | k if | wj | < k 4. Variance Stabilization
j j

then wj is s significant then wj is not significant

(3)

We need the noise standard deviation at each scale for iid (independent identically distributed) Gaussian pixel values. In the case of additive Poisson or Poisson plus Gaussian noise, we transform the image so that this is the case. The transform we use is one for variance stabilization. e To compute the standard deviation j at each scale, simulate an image containing Gaussian noise with a standard deviation equal to 1, and take the wavelet transform of this image. e The standard deviation of the noise at a scale j of the image, j = I j , where I is the standard deviation of the noise in the original image. If the noise in the data I is Poisson, the transform t(I (x, y )) = 2 I (x, y )+ 3 8 (4)


Noise Detection and Filtering using Multiresolution Methods

451

acts as if the data arose from a Gaussian white noise model, with = 1, under the assumption that the mean value of I is large. For combined Gaussian and Poisson noise:

t(I ) =

2

3 I (x, y )+ 2 + 2 - g 8

(5)

where is the gain, and g the standard deviation and the mean of the read-out noise.

5.

Algorithm for Noise Filtering

Reconstruction of the image, after setting non-significant wavelet coefficients to zero, at full resolution provides an approach for adaptive filtering. A satisfac~ tory filtering implies that the error image E = I - I , obtained as the difference between the original image and the filtered image, contains only noise and no `structure'. We can easily arrive at this ob jective by keeping significant wavelet coefficients and by iterating. The following algorithm (Starck et al. 1997) converges quickly to a solution which protects all signal in the image, and assesses noise with great accuracy. 1. n 0. 2. Initialize the solution, I
(0)

, to zero.

3. Determine the multiresolution support of the image. 4. Estimate the significance level (e.g., 3-sigma) at each scale. 5. Keep significant wavelet coefficients. Reconstruct the image, I these significant coefficients. 6. Determine the error, E filtered).
(n) (n)

, fr o m

= I-I

(n)

(where I is the input image, to be
(n)

7. Determine the multiresolution transform of E

.

8. Threshold: only retain the coefficients which belong to the support. ~ 9. Reconstruct the thresholded error image. This yields the image E taining the significant residuals of the error image. 10. Add this residual to the solution: I 11. If | (
E
(n-1)

(n)

con-

(n)

I

(n)

~ +E

(n)

.

-

E

(n)

)/

E

(n)

|>

then n n +1 and go to step 4.


452 6.

Murtagh and Starck Example 1: Simulation

A simulated image containing stars and galaxies is shown in Figure 1 (top left). The simulated noisy image, the filtered image and the residual image are respectively shown in Figure 1 top right, bottom left, and bottom right. We can see that there is no structure in the residual image. The filtering was carried out using the multiresolution support.

Figure 1. (a) Simulated image, (b) simulated image and Gaussian noise, (c) filtered image, and (d) residual image.

7.

Example 2: Spectrum Filtering

Figure 2 shows a noisy spectrum (upper left, repeated lower right). For the astronomer, the spectral lines ­ here mainly absorption lines extending downwards


Noise Detection and Filtering using Multiresolution Methods

453

­ are of interest. The continuum may also be of interest, i.e. the overall spectral tendency. The spectral lines are unchanged in the filtered version (upper center, and upper right). The lower center (and lower right) version shows the result of applying Daubechies' coefficient 8, a compactly-supported orthonormal wavelet. This was followed by thresholding based on estimated variance of the coefficients, as proposed by Donoho (1990), but not taking into account the image's noise properties as we have done. One sees immediately that a problem- (or image-) driven choice of wavelet and filtering strategy is indispensable.

Original
10000

Thresholded (see text)
10000

Superimposed

9800

10000

9600

9800

9400

9600

9200

9400

0

100

200

300

400

500

9200

0

100

200

300

400

500

9200

9400

9600

9800

0

100

200

300

400

500

Original
10000 10200

Daubechies8/DonoJohnUniv
10000

Superimposed

9800

10000

9600

9800

9400

9600

9200

9400

0

100

200

300

400

500

9200

0

100

200

300

400

500

9200

9400

9600

9800

0

100

200

300

400

500

Figure 2. Top row: original noisy spectrum; filtered spectrum using the multiresolution support method; both superimposed. Bottom row: original; filtered (using Daubechies coefficient 8, and Donoho and Johnstone "universal" thresholding); both superimposed.


454 8. The this seco tion

Murtagh and Starck Example 3: X-Ray Image Filtering galaxy cluster A2390 is at a redshift of 0.231. Figure 3 shows an image of cluster, obtained by the ROSAT X-ray spacecraft. The resolution is one arc nd per pixel, with a total number of photons equal to 13506 for an integratime of 8.5 hours. The background level is about 0.04 photons per pixel.

It is obvious that this image cannot be used directly, and some treatment must be performed before any analysis. The standard method consists of convolving the image with a Gaussian. Figure 4 shows the result of such processing. (The Gaussian used had full-width at half maximum equal to 5", which is approximately that of the point spread function). The smoothed image shows some structures, but also residual noise, and it is difficult to give any meaning to them. Figure 5 1997a). The structures in phenomena, shows an image filtered by the wavelet transform (Starck and Pierre noise has been eliminated, and the wavelet analysis indicates faint X-ray emission, allowing explanation of gravitational amplification observed in the visible domain (Starck and Pierre 1997b).

Figure 3.

ROSAT Image of the cluster A2390.


Noise Detection and Filtering using Multiresolution Methods

455

Figure 4. ROSAT image of the cluster A2390 filtered by the standard method (convolution by a Gaussian).

Figure 5. ROSAT image of the cluster A2390 filtered by the method based on wavelet coefficients. 9. Conclusion

The methods described in this article have been applied to many practical problems over the last few years. A comprehensive treatment of these methods can


456

Murtagh and Starck

be found in Starck et al. (1998). A large software package will be available from early 1998. References Donoho, D.L. & Johnstone, I.M. 1993, "Ideal spatial adaptation by wavelet shrinkage", Stanford University, Technical Report 400 (available by anonymous ftp from playfair.stanford.edu: /pub/donoho) Snyder, D.L., Hammoud, A.M. and White, R.L. 1993, J. Optical Soc. Am. 10, 1014 Starck, J.-L. Murtagh, F. & Bijaoui, A. 1998, Image and Data Analysis: the Multiscale Approach, Cambridge University Press, in press Starck, J.-L. & Murtagh, F. 1997, "Automatic noise estimation from the multiresolution support", PASP, in press Starck, J.-L. & Pierre, M. 1997a, "X-ray structures in galaxy cores", to appear in A&A Starck, J.-L. & Pierre, M. 1997b, "Structure detection in low intensity X-ray images", to appear in A&A Tekalp, A.M. and Pavloviґ, G. 1991, in A.K. Katsaggelos, ed., Digital Image c Restoration, Springer-Verlag, New York, 209