Proc. SPIE 3336, pp.

243-250 (1998) LA-UR-98-1385

Physics of Medical Imaging, J. T. Dobbins III and J. M. Boone, eds.

A simplified method of estimating noise power spectra

Kenneth M. Hanson
Los Alamos National Laboratory, MS P940
Los Alamos, New Mexico 87545 USA

ABSTRACT
A technique to estimate the radial dependence of the noise power spectrum of images is proposed in which the
calculations are conducted solely in the spatial domain of the noise image. The noise power spectrum averaged over
a radial spatial-frequency interval is obtained from the variance of a noise image that has been convolved with a
small kernel that approximates a Laplacian operator. Recursive consolidation of the image by factors of two in each
dimension yields estimates of the noise power spectrum over that full range of spatial frequencies.
Keywords: noise power spectrum, Wiener spectrum, granularity, pass-band granularity, Laplacian, binomial kernels,
Gaussian pyramid, Laplacian pyramid

1. INTRODUCTION
The noise in images is recognized as an important factor in determining image quality. Noise power spectra (NPS)
provide the means of characterizing image noise and play a central role in the ultimate measure of image quality,
the number of noise-equivalent quanta, (NEQ).1 This article presents a spatial-domain method of estimating the
radial dependence of the noise power spectrum. The goal of this method is to provide an easy and robust measure of
the NPS that can serve for day-to-day measurements of image noise. It is not intended as a replacement of Fourier
methods for estimating the NPS, which remain the gold standard.
The proposed method was motivated by the granularity measure suggested in 1935 by Selwyn2 to characterize the
spatial correlations in film-grain noise. Various √
definitions of granularity have been used in the literature. However,
they are all related to the product G(A) = σA A, where σA is the rms deviation in the average over a square or
circular region of area A. When the noise in the image is uncorrelated, i.e., has a flat NPS, G(A) is independent of
A. Variations with respect to A reflect variations of the NPS with spatial frequency.
The calculation of σA may be performed by taking the square root of the variance of the result of a convolution of
the noise image with a kernel of constant amplitude. Thus, it is natural to consider a generalization of granularity is
based on other types of kernels. Such a generalization may be useful for suppressing the high-frequency components
1
of the uniform square kernel. For example, Hanson3 observed that the use of the standard kernel masked the A− 4
dependence of G(A) for large A that was expected for the noise in computed-tomographic images. The use of
two-dimensional pyramidal kernels resolved the issue.
The proposed method uses an approach similar to the granularity calculation to quantify the radial-frequency
dependence of the NPS. The calculation is based on applying to a noise image Laplacian operators, which correspond
to pass-band filters. The variance of the noise image after application of the Laplacian is proportional to the NPS
averaged over the mid-frequencies of the spectrum. A multiresolution Laplacian pyramid sweeps out a range of radial
frequencies at which the NPS is estimated. Because the calculations for the proposed method take place only in the
spatial domain, there are potentially a number of advantages over the conventional NPS.

2. NOISE POWER SPECTRUM

The noise power spectrum (NPS), also called the Wiener spectrum, is defined mathematically in terms of the Fourier
transform of noise images:

 2 
1  
S(fx , fy ) = lim  r(x, y) exp[−2πı(xfx + yfy )] dx dy  , (1)
∆x,∆y→∞ ∆x∆y  

Contact information: Email: kmh@lanl.gov; WWW: http://www.lanl.gov/home/kmh; Telephone: 505-667-1402

where r(x, y) designates a noise image, which is a function of position (x, y), and the angle brackets indicate an
average over an infinite ensemble of such noise images. The integration is over the area ∆x∆y, which ideally goes to
infinity. The noise power spectrum of an image is inherently a two-dimensional structure, i.e., an image in the 2D
frequency domain. The units of S are [r2 L2 ], namely the square of the units of r times the units of length squared.
Mathematically speaking, the noise power spectrum is a useful characterization of the noise only if the noise is
Gaussian distributed about a zero mean value and stationary, i.e., its covariance is independent of position in the
image. Of course, in real images this condition is violated to varying degrees and one never has an infinite number
of images of infinite size. The art of estimation of noise power spectra revolves around overcoming such departures
from the ideal.
The typical technique for estimating the NPS for real images is based on averaging the power of the Fourier
transform of noisy image samples taken from several images, or more routinely from a single noise image. It is
usually important to remove the slow variations in the image intensity and to window the samples, i.e., multiply the
sample by a function that tapers to zero at the edge of the image, before taking the Fourier transform. The low-
frequency portion of the NPS and its interpretation tends to be difficult to handle.4,5 The uncertainty in the estimate
of S depends on the number of samples that can be included in the average since the relative rms deviation in the
1
power in each 2D frequency bin is approximately [N ]− 2 for N averaged samples. Despite the difficulties of estimating
NPS, it is possible to achieve results with absolute normalization that agree between different laboratories.6
Since the convolution of two images in the spatial domain is equivalent to multiplication in the frequency domain
of the Fourier transforms of those images, convolving a noise image with a kernel w modifies the NPS of a noise
image by a factor of |W |2 , where W is the Fourier transform of w. Since the total power in the frequency domain is
the same as in the spatial domain, the variance in the convolved image is given by

2
σw = |W |2 S(fx , fy ) dfx dfy . (2)

For discretely sampled images, the range of integration for both fx and fx is from −fN to fN , where fN = [2a]−1 is
the Nyquist frequency, a being the pixel spacing, which is assumed here to be the same in both x and y directions.
2
From Eq. (2) we see that σw is essentially a weighted average of the NPS over a region in Fourier space with the
weight function given by the power spectrum of w. In equation form,
2 2
S|W |2 = σw a Kw , (3)

where the angle brackets indicate an average, weighted by |W |2 , of the power spectrum of the convolution kernel w.
The normalization constant Kw is the reciprocal of the total power of the kernel
 
−1
Kw = |W |2 dfx dfy = 2
wij , (4)
ij

where the wij are the elements of the convolution kernel for discretely sampled images.
By appropriately selecting the convolution kernel, the variance in the convolved noise image, can be used to
estimate the NPS for specific spatial frequencies.

3. CONVOLUTION KERNELS AND MULTIRESOLUTION PYRAMIDS

In this section, I will briefly describe binomial kernels and their use to create Gaussian pyramids. A Gaussian pyramid
can be used to generate a multiscale hierarchy of Laplacian images, which effectively represent the original image
processed by a bank of pass-band filters. The objective is to use the Laplacian-convolved noise images at multiple
resolution scales to estimate the NPS at various radial frequencies.

3.1. Binomial Kernels

Binomial kernels provide a useful family of convolution kernels for smoothing images. The hierarchy begins with the
definition of a 2 × 2 kernel  
1 1 1
B1 = , (5)
4 1 1
which is normalized so the sum over all elements is unity. Subsequent orders of binomial kernels are obtained by
convolution, so, in general,
B m+n = B m B∗ n , (6)
where ∗ represents convolution. The size of Bn is (n + 1) × (n + 1). These kernels all have unit sum. They are
reasonable approximations to Gaussian kernels, particularly as n increases. As described by Jähne,7 the binomial
kernels, or filters as he calls them, can be used to construct many types of smoothing kernels with characteristics in
the frequency domain that can be tailored to one’s needs. For this work, I will use only the simple binomial filters.
An important property of the binomial kernels is that they are separable into an outer product of a row vector
times a column vector. This property translates into calculational efficiency because the convolution of Bn with an
image can be accomplished by convolving a (n + 1) × 1 kernel followed by a convolution of a 1 × (n + 1) kernel,
which involves less computation than required convolution by an arbitrary nonseparable (n + 1) × (n + 1) kernel.
The computational savings grows as n increases.
From the above binomial kernels, one can construct Laplacian operators. The lowest order Laplacian is
 
−0.0625 −0.1250 −0.0625
L2 = I − B2 =  −0.1250 0.7500 −0.1250  , (7)
−0.0625 −0.1250 −0.0625

where I is essentially a delta-function, in this case, a 3 × 3 kernel with unity in the central position and zeros
elsewhere. The next higher-order Laplacian would be
 
−0.0039 −0.0156 −0.0234 −0.0156 −0.0039
 −0.0156 0.0000 0.0312 0.0000 −0.0156 
 
L4 = B•2 − B•4 = 
 −0.0234 0.0312 0.1094 0.0312 −0.0234  ,
 (8)
 −0.0156 0.0000 0.0312 0.0000 −0.0156 
−0.0039 −0.0156 −0.0234 −0.0156 −0.0039

where the dots indicate that the central elements of the kernels involved should be aligned, which is required for
clarity because B 2 is 3 × 3 and B 4 is 5 × 5.
The 2D power spectrum of L4 is shown in Fig. 1. The mean radial frequency (weighted by f df to obtain the
proper radial mean) is 0.540 fN along the frequency axes and 0.583 fN along the diagonals, where FN is the Nyquist
frequency. From the 2D spectrum, it can be seen that L4 samples the spatial frequency domain in a nearly isotropic
manner. The mean radial frequency over all angles is 0.559 fN . The width of the radial profile is about 0.38 fN
FWHM and the normalization for Eq. (3) is K = 49.8. The 2D power spectrum of L2 is decidedly nonisotropic
because it extends right up to the corners of the 2D Fourier domain delineated by ±fN . The mean radial frequency
is 0.812 fN along the frequency axes and 1.088 fN along the diagonals, with a mean over all angles of 0.917 fN .
The normalization is K = 1.561. A useful feature of the Laplacian convolution is that it effectively rejects the
contributions from frequencies that are much lower than the mid frequencies of the corresponding filter.
One approach to proceeding toward lower frequencies would be to recursively convolve the image with broader
and broader binomial kernels and compute the variance of the Laplacians formed by taking differences between the
images of different resolution so formed. This calculation is not as inefficient as it might seem at first because of the
separability of the binomial kernels. One characteristic of this approach is that the noise image stays the same size.
Alternatively, similar results can be achieved with a multiresolution representation in which the smoothed images
at each resolution scale are consolidated, or subsampled, to produce a progression of smaller and smaller images that
can be convolved with fixed-size kernels.

3.2. Gaussian Pyramid

The idea of repeatedly consolidating an image to obtain successively smaller images of coarser and coarser resolution
is captured by the Gaussian pyramid scheme.7,8 This concept also is the basis of multiscale, multigrid, and multires-
olution methods and filter banks. The zeroth level of the Gaussian pyramid is the initial full-resolution image, which
is designated as G0 . Subsequent degrees of lower resolution are obtained by smoothing the image with a Gaussian
Figure 1. The two dimensional frequency contour plot of the square of the Fourier transform of L4 (left), renormal-
ized to make its total power unity, and its radial frequency dependence (right) along the axis (solid line) and along
the diagonal (dashed). Zero frequency of the 2D frequency plot on the left is in middle of the image.

kernel G and decimating, or consolidating, the result by a factor of two by taking every second pixel and every second
row, a process represented by R. The image at the kth level of the Gaussian pyramid is, therefore,

Gk+1 = RGGk . (9)

If the image Gk possesses components at frequencies above the Nyquist frequency for the new pixel spacing, those
components will appear in Gk+1 at lower frequencies because of the consolidation process, a phenomenon known as
aliasing. For noise power spectra, the result of this aliasing is to add to the true power spectrum below the Nyquist
frequency contributions from above. The essential trick is to consolidate the noise image only after convolving it
with a sufficiently broad kernel that there remains little or no noise power above the new Nyquist frequency.
A Laplacian image at the kth level can be formed as follows:

Lk = Gk − GGk . (10)

A computational advantage of this form is that the second term can be used to obtain Gk+1 by (9). The hierarchy
of Laplacian images formed in this way is called a Laplacian pyramid.
I will use this Laplacian pyramid to estimate NPS at frequencies below one half the Nyquist frequency of the
original image. For G, I propose the fourth-order binomial kernel B 4 discussed above. This kernel corresponds to
a filter that drops nearly to zero by one half the Nyquist frequency of the image on which it operates. Thus, the
consolidation operation results in little aliasing. Jähne suggests the conservative approach of using B8 to construct
Gaussian pyramids to avoid aliasing effects. However, this approach may not be needed because we are dealing here
with the power spectrum, not the Fourier amplitudes and the fraction of aliasing contributions should be reduced by
their square.
Figure 2 shows the power spectrum of the effective filter at the first level (k = 1) for the Laplacian (10). This
power spectrum is the product of the power spectrum for I − B4 times that for B 4 at the zeroth level. The variance
of the noise image obtained at this point in the Gaussian pyramid has contributions from radial frequencies with a
mean value of 0.67 fN,1 (the Nyquist frequency for the pixel spacing at this level) with relatively small contributions
from above fN,1 caused by aliasing. The normalization constant for Eq. (3) for this kernel is K = 6.24, including a
Figure 2. The radial frequency dependence of the power spectrum of the Laplacian at the first level of the Gaussian
pyramid along the frequency axis (solid line) and along the diagonal (dashed) of the 2D spectrum. The Nyquist
frequency at this level is one-half that of the original image.

5% reduction to account for aliasing. As one proceeds down the pyramid, K is slightly affected by the B 4 smoothing
from higher levels. It needs to multiplied by a factor of 1.370.23(k−1) for each level k.

3.3. Algorithm
This algorithm is based on calculating the variance of a noise image after convolution with Laplacians. For the NPS
estimates in the two highest-frequency intervals, the two Laplacian operators (7) and (8) are used. Estimates for
lower-frequency intervals are obtained using the Laplacian pyramid.
Designating the original image as G0 , the proposed algorithm is as follows:
(1) subtract from G0 the result of its convolution with B 2 to obtain L2 G0 ;
the variance of this result times 1.56 a2 is S at radial frequency of 0.917/2a
(2) subtract from B 2 G0 the result of its convolution with B 2 to obtain L4 G0 ;
the variance of the result times 49.8 a2 is S at radial frequency of 0.559/2a
The image G0 is the beginning of the Gaussian pyramid. Begin loop, starting with k = 0 (a0 = a):
(3) consolidate B 4 Gk by factor of two putting every other pixel and row into Gk+1 and increment k;
(new pixel spacing is twice previous one , i.e., ak+1 = 2ak )
(4) subtract from Gk the result of its convolution with B 4 ;
variance of the result times 6.24 a2k Πk (1.37)0.23(k−1) is S at radial frequency of 0.67/2ak
Repetition of steps (3) and (4) many times generates estimates of S at geometrically decreasing frequencies.
1
The standard deviation in these estimates is approximately c [2np ]− 2 , where np is the number of pixels in the
convolved image
 that can 2be used to calculate the variance and c is scaling constant that depends on the convolution
∗ 2 /|w| .
kernel; c = |w w|

4. EXAMPLE
The noise image was taken from a storage-phosphor screen (Fuji BAS 3), 20 cm × 40 cm in size, that was scanned
with a Fuji model BAS 1500 at the normal pitch of 100 µm. The screen was exposed to a bremstrahlung beam
produced by a microtron running at 15 MeV after being filtered through 4 cm of depleted uranium. The dose on the
 a b

Figure 3. (a) A 768 × 768 pixel2 noise image obtained from a storage-phosphor screen exposed to high-energy x
rays and (b) the Laplacian of Eq. (10) at the first level of the Gaussian pyramid. The actual linear dimension of
(b) in pixels is half that of (a), but it is displayed here with the same size to emphasize the pass-band nature of the
Laplacian pyramid.

screen is estimated to have been about 400 mR (4 mGy), Co60 equivalent, yielding scanner readings of about half
of the saturation level. The full image is flattened by dividing by a fit to the original image that is cubic in x and
y. Thus the analyzed noise image is dimensionless, its noise fluctuation values are relative to the mean amplitude of
the original image. A 7.68-cm-square subsample of the full-size image is used in the present analysis, which is shown
in Fig. 3.
The NPS estimates obtained using the algorithm outlined above are presented in Fig. 4. The error bars shown
indicate the estimated rms uncertainties. The lowest-frequency point comes from a 24×24 pixel2 image after consol-
idation of the original 7682 noise image by a factor of 25 = 32.
For comparison, also shown is the NPS obtained using a standard Fourier-transform technique. This calculation
commences by removing the very low frequencies from the same 7682 noise image by zeroing out the first three
frequency components of the Fourier transform of the image, i.e., eliminating spatial frequencies equal to and below
0.04 mm−1 . Eleven overlapping 128×128-pixel2 subsamples are extracted from the full image in each direction. Each
subsample is multiplied by a 2D Hanning window and Fourier transformed. The final NPS is the average of all 121
Fourier power spectra, normalized to account for the effect of the Hanning window. The match between the Fourier
results shown in Fig. 4 and those of the proposed algorithm is very good, except for the point at 0.106 mm−1 . The
strong disagreement there is most likely caused by the fact in the Fourier analysis, the spatial frequency components
below 0.04 mm−1 are removed.
The spatial-domain calculation for this image takes 45 s on a PC with an AMD K6 processor running at 166
MHz, whereas the Fourier transform method takes about five times longer.
Figure 5 shows the full two-dimensional NPS for the sample noise image obtained with the Fourier-transform
technique. It is readily seen that this NPS is not isotropic. In particular, there are sharp peaks in the spectrum
at mid frequencies. These peaks have been observed to be reproducible from image to image and are essentially
eliminated when difference images between successive scans are analyzed. These observations indicate that these
anomalies are probably produced by the scanner. These results show how the 2D NPS may be useful to diagnose an
imaging system, as was also demonstrated in Ref. 3.
Figure 4. Noise power spectrum for the noise image shown in Fig. 3 as a function of radial spatial frequency
estimated using the proposed spatial-domain technique (crosses) compared to that obtained using a conventional
Fourier transform technique (line).

Figure 5. The two-dimensional noise power spectrum obtained with the Fourier-transform method displayed on a
logarithmic intensity scale. Zero frequency is in middle of this 2D frequency-domain image.
 5. DISCUSSION
I have proposed a spatial-domain method to estimate the radial dependence of the noise power spectrum from a noise
image. The method is based on the convolution of a noise image with Laplacian kernels, used in conjunction with
a multiresolution analysis afforded by a Gaussian pyramid. Properly scaled, the variances in the convolved images
provide estimates of the noise power spectrum at radial spatial frequencies that differ by roughly powers of two.
Compared with Fourier methods, the proposed algorithm is simple to understand and reasonably easy to imple-
ment. This algorithm has a statistical efficiency comparable to that of the Fourier tranform method. Its disadvantages
are poorer frequency resolution and lack of a 2D frequency display of the NPS, which can be helpful in understanding
an imaging system.3
Because the calculations are done in the spatial domain, the proposed algorithm can be easily adapted to visually
(and quantitatively) check whether the noise in the analyzed image is stationary. For example, it would be easy to
display the local variance at each resolution level. By directly observing the Gaussian or Laplacian images, as in
Fig. 3b, one can view the spatial correlations of noise, which may vary with position, intensity, etc., and perhaps
capture the spirit of Ref. 9. With this algorithm, one can remove the effects of identified artifacts and nonstationary
behavior from the NPS estimates.
The B 4 binomial kernel used to construct the Gaussian pyramid clearly has its limitations. Jähne7 describes a
variety of smoothing kernels that could be used to improve the frequency resolution of the NPS estimates and to
better suppress spurious low-frequency contributions estimates, as well as reduce aliasing effects. Other extensions of
the present technique would be to use other types of kernels to restrict the NPS to selected directions, which might
be important for images that are digitzed by asymmetric scanners, for instance.
The spatial-domain algorithm presented here is available on the web at http://bayes.lanl.gov/∼kmh/nps as
an IDL10 procedure.

ACKNOWLEDGEMENTS
I would like to thank Todd Kauppila for the noise image used in my example and for his NPS estimate of that image
obtained with a Fourier transform technique. This work supported by the United States Department of Energy
under contract W-7405-ENG-36.

REFERENCES
1. J. C. Dainty and R. Shaw, Image Science, Academic, London, 1974.
2. E. W. H. Selwyn Photog. J. 75, pp. 571–580, 1935.
3. K. M. Hanson, “Detectability in computed tomographic images,” Med. Phys. 6, pp. 441–451, 1979.
4. P. C. Bunch, R. Shaw, and R. V. Metter, “Signal-to-noise ratio measurements for a screen-film system,” Proc.
SPIE 454, p. 154, 1984.
5. P. C. Bunch and R. V. Metter, “Noise power spectrum analysis of a scanning microdensitometer,” Applied
Optics 27, pp. 3468–3474, 1988.
6. J. M. Sandrik, R. F. Wagner, and K. M. Hanson, “Radiographic screen-film noise power spectrum: calibration
and intercomparison,” Appl. Opt. 21, pp. 3597–3601, 1982.
7. B. Jähne, Digital Image Processing: Concepts, Algorithms, and Scientific Applications, Springer, Berlin, 1995.
8. D. H. Ballard and C. M. Brown, Computer Vision, Prentice-Hall, Englewood Cliffs, 1982.
9. H. H. Barrett, R. F. Wagner, and K. J. Myers, “Correlated point processes in radiological imaging,” in Physics
of Medical Imaging,R. L. VanMetter and J. Beutel, eds., Proc. SPIE 3032, pp. 110–124, 1997.
10. Interactive Data Language, Research Systems, Inc., 2995 Wilderness Place, Boulder, CO 80301.