Beruflich Dokumente
Kultur Dokumente
Research Scholar, Department of Biomedical Engineering, National Institute of Technology, Raipur 492010, India
2
Department of Biomedical Engineering, Vel Tech Multi Tech, Chennai 600062, India
3
Department of Biomedical Engineering, National Institute of Technology, Raipur 492010, India
4
Department of Computer Science and Engineering, Arunachala College of Engineering, Tamil Nadu 629203, India
The noise suppression and edge preservation offered by Nonlocal Means (NLM) filter depends on the selection
of decay control parameter and radii of search and similarity windows. If the decay control parameter is too
small then little noise will be removed, while if it is too high, the image becomes blurry. Empirical selection of
the filter parameters are always subjective and may be far from optimal. The present study demonstrates an
automated method to compute decay control parameter from the noise estimate and an objective method to
identify the optimum values of operational parameters using modified Pratts Figure of Merit (PFOM) on Magnetic
Resonance (MR) images. The decay control parameter is computed as the product of standard deviation of the
noise and an arbitrary constant, termed as coefficient of noise variance. The radii of search, similarity windows
and coefficient of noise variance which offer maximum noise suppression with minimum edge degradation were
identified on multiple MR images with different textural and noise content. It has been observed that regardless
of the textural and noise content of the MR image, the optimum range of coefficient of noise variance lies
between four and eight, typically at seven. For MR images, the minimum radii of similarity and search windows
should be greater than three, beyond which the radii has no significant influence on the performance of NLM,
provided the radius of search window should be sufficiently larger than radius of similarity window, typically four
times. Since the radii of search, similarity windows and coefficient of noise variance is robust to textural and
noise content of the MR image, the computed decay control parameter depends only on the standard deviation
of the noise. Hence, the proposed NLM is fully noise adaptive. The experimental analysis was performed in
Matlab .
Keywords: Decay Control Parameter, Edge Preserving Filter, Nonlocal Means Filter, Pratts Figure of Merit.
1. INTRODUCTION
Non Local Means (NLM) filter is an edge preserving filter,
widely accepted in medical image restoration because of its detail
preserving and optimum smoothing characteristics. However, the
performance of NLM is prone to the selection of its operational parameters and in general they are selected empirically
through qualitative inspection of the restored image. The operational parameters of NLM include the radii of similarity, search
windows and the decay control parameter. The decay control
parameter has dominant influence on the performance of the filter
than the radii of similarity and search windows.
In NLM, the restored intensity is the weighted sum of the pixel
intensities in the whole image or in a search window of restricted
dimension. When the decay control parameter is very high, all
the pixels in the search window will have the same weight with
respect to the pixel to be restored. The restored pixel intensity
thus turns to be approximately the average of the intensities of
the pixels in the search window, leading to a strong smoothing.
When the decay control parameter is very low, only few pixels in
the search window with neighbours are very similar to the neighbourhood of the contextual pixel will have significant weights.
The restored intensity thus tends to be the weighted average of
only few of the pixels in the search window leading to a weak
smoothing and perhaps, noise may be retained in the restored
image. In short, if the decay control parameter (h) is too small
then little noise will be removed while if h is too high, the image
becomes blurry.
There had been investigations112 to find out the optimum values of operational parameters of NLM and to frame adaptive and
image driven methods for their automatic selection. Baudes et al.1
2156-7018/2016/6/001/011
doi:10.1166/jmihi.2016.1780
RESEARCH ARTICLE
T1
PD
T2
1 ( )
2 ( )
3 ( )
4 ( )
5 ( )
1.30
1.26
1.25
1.22
1.18
1.18
1.18
1.14
1.14
1.18
1.14
1.13
1.14
1.11
1.10
kernel with h2 = 100 was active only for high level noises and
kernel with h2 = 1000 was not suitable for any level of noise
implying that which is nonsensical.
Li et al.8 decomposed the noisy image into wavelet sub-bands
and the coefficients in each sub-band were modelled by Generalized Gaussian Distribution (GGD), whose parameters were
used to estimate the noise variance. The adequate
decay control
parameter setting was approximated as h = 2 by using maximum PSNR as the objective function. An adaptive selection of
the radius of the similarity window inversely proportional to the
local variance was proposed by Alhosainy and Badran.9 Optimum value of decay control parameter for different values of
radius of similarity window (1 RS 18) was observed on multiple test images, corrupted with additive white Gaussian noise
of zero mean and standard deviation equal to 20. The optimum
value of decay control parameter was exponentially decreasing with the increase in radius of similarity window on all test
images. Similarly on a single test, image corrupted at different
noise level (10 30), the optimum value of decay control
parameter was observed for different values of radius of the similarity window.
In contradictory to the exponential decrease of optimum value
of the decay control parameter with respect to the increase in
radius of the similarity window (previous observation), the optimum value of the decay control parameter remained constant for
low noise levels. But maximum possible value of SNR, used as
the objective function does not ensure that the edges are preserved and the noise is fully eliminated and hence the decay
control parameter is optimum.
Zia et al.10 used different sets of operational parameters for
low and high frequency regions in the MR image. Beuchers
morphological gradient operators were used to separate low and
high frequency regions. The optimum parameter settings for these
separate regions were identified empirically. Sticking on to the
observation in Manjon,2 the optimum decay control parameter
h = 12, by fixing h = 14 and h = 11 for radius of the similarity window equal to 1 and 2, the radius of the search window
was continuously varied (3 Rsearch 7) and average improvement in RMSE was observed for different noise level (2
20) for low and high frequency region separately. Similarly, by
maintaining Rsearch = 5 (Rsearch at which maximum improvement
in RMSE was observed), the radius of the similarity window
was varied (1 RS 4) for (2 20), at h = 11 and h =
14, for low and high frequency regions in the image. The optimum parameter setting identified for high frequency region was
Rsearch = 5, RS = 1 and h = 11 and for low frequency region
the optimum setting was Rsearch = 5, RS = 1 and h = 14. The
literature used two separate kernels for computing the weights
corresponding to the pixels falling in the low and high frequency
regions. But in NLM, conceptually, the Gaussian weighting kernel is meant to offer spatially decaying weights to the grey level
differences between the neighbourhood or similarity window of
the pixel to be restored and a contextual pixel in the search
window. RMSE was the objective function used to identify the
optimum radii of similarity and search windows.
Farmer et al.11 observed that for added white Gaussian noise
( = 30) at h = 2 and radius of the search window equal to
10, the value of the radius of the similarity window equal to 1,
offers the minimum Mean Squared Error (MSE). Similarly, for
h = 2 and the radius of the similarity window equal to 1, the
RESEARCH ARTICLE
2. METHODOLOGY
Many of the image restoration schemes like bilateral filter13 average the similar pixels in the image to estimate the restored pixel
intensity. The similarity may be in terms of the closeness of grey
levels alone and sometimes the spatial proximity too. In contrast to bilateral filter, in NLM the measure of similarity between
pixels is more robust as the regions surrounding the pixels are
compared rather than the individual pixel comparison. Pixels far
from the pixel being filtered are not penalized due to its distance
to the current pixel, as happens with the bilateral filter. The pattern redundancy is not restricted to be local therefore; the scheme
is referred as non-local.
Given the corrupted image X, the restored intensity at any
pixel location Y i j, is the weighted sum of all the pixels in the
corrupted image X is given as
Y i j =
N
M
m=1 n=1
i = 1 2 M
and
j = 1 2 N
(1)
Constrained by,
0 W i j m n 1 and
N
M
m=1 n=1
j
i
1
1
Rs p=i q=j 2i + 1 2j + 1
and
j = 1 2 Rs
(4)
RESEARCH ARTICLE
n =
0.0068 0.0201 0.0572 0.0572 0.0572 0.0201 0.0068
G = 0.0068 0.0201 0.0572 0.0572 0.0572 0.0201 0.0068
0.0068 0.0201 0.0572 0.0572 0.0572 0.0201 0.0068
0.0068 0.0201 0.0201 0.0201 0.0201 0.0201 0.0068
0.0068 0.0068 0.0068 0.0068 0.0068 0.0068 0.0068
Fig. 1. Weighting kernel.
i = m & j = n
(5)
Rsearch
Rsearch
m=Rsearch n=Rsearch
Xi m j n
(6)
Rsearch is the radius of the search window. In this study the decay
control parameter (h) is considered as a linear function of the
standard deviation of the noise as demonstrated in Eq. (3). But
devising a noise estimator for computing the noise variance in
highly textured MR images which contain morphological edges
as well is a difficult task. Both the edge and noise pixels are
equally sensitive to second order derivative and both of them
register zero crossings. Usually, it is advisable to estimate the
noise from homogenous regions devoid of edges. Hence, a noise
estimator kernel is designed to be insensitive at high frequency
regions as the difference between two kernels L1 and L2 , each
approximating the Laplacian such that,14
1
H = 2L2 L1 2
1
where L1 = 1 4
0
and L2 = 1/2 0 4
1
4
1
0
1
0
1
0
N
M
1
Xi j H
2 6M 2N 2 i=1 j=1
(7)
Ns
1
1
MaxN0 Ns k=1 1+ d2 k
(8)
f0 = x y =
0 else
(10)
1 if Gs x y GT
and fs x y =
0 else
where the threshold GT is the mean of local gradients of the original image. X = 1 2 M
and y = 1 2 N
. For an image
of size M N the equation is
GT =
N
M
1
G x y
MN x=1 y=1 0
(11)
RESEARCH ARTICLE
N
M
f0 x y
and
x=1 y=1
Ns =
N
M
fs x y
(a)
(b)
(12)
x=1 y=1
The magnitude gradient is computed from gradient along horizontal and vertical directions17 such that,
Gx y = gx2 + gy2
(13)
where the gradient along x and y directions,
f x y
x
and
gy =
f x y
y
Fig. 2.
(14)
For a smoothing kernel with ideal edge preservation, the number of dominating edge pixels in the smoothened and original
image would be equal, with their spatial locations overlapping.
For a perfect match of maps of dominating edges in the original and smoothened image, the PFOM is unity and a value of
0 indicates abrupt mismatch. In other words a value of PFOM
equal to unity ensures that none of the dominating edge pixels in
the original MR image is degraded or misplaced during smoothing. The commonly used value of arbitrary scaling constant
is 1/9,1820 , 215 and 0.2.21 There is no hard and last constrain for
setting
and appreciably
> 0 as recommended by Evans and
Liu.21 The value of
is set to unity to make the PFOM completely dependent over the number of degraded edge pixels and
the accuracy of the location of the preserved edge pixels during
restoration.
3. RESULTS
The NLM proposed in this study has three operational parameters, coefficient of noise variance, Rs and Rsearch . The optimum
values of these parameters are identified using modified PFOM
as an objective function. MR images which are different in textural and noise content are used to identify the optimum values of
coefficient of noise variance, equally suitable for MR images of
different series and planes. Two specimen MR images are shown
in Figures 2(a and b). The specimens used for the experimental analysis are MR images of pre-operative and biopsy-proven
GBM-edema complex, collected from Hind Labs, Government
Medical College Kottayam, Kerala. The specimen images are of
study, MRS, under T2 FLAIR, GRE, DWI, 1000b ASSET and
T1 FS-ECE series. The specification of MR equipment is; Manufacturer: GE Medical Systems, Model Name: SignaHDxt, Acquisition Type: 2D and 1.5T field strength. All the experimental
analysis is performed in Matlab .
Before employing modified PFOM as the objective function to
identify the optimum values of operational parameters of NLM,
the feasibility of minimum/maximum PSNR as the objective
function is analysed, because majority of the work follow PSNR
based protocol. The variation of PSNR between the noisy and
restored images with respect to coefficient of noise variance for
different values of radius of similarity window (Rs = 3, 5 and 7)
for Rsearch = 9 on the MR image 1 is depicted in Figure 3. The
restored images, gradient of restored images and corresponding
binary edge maps for coefficient of noise variance between 2 and
20 for Rs = 3 and Rsearch = 9 are furnished in Figures 4 to 13.
51
gx =
50
49
48
47
46
45
44
10
12
14
16
18
20
RESEARCH ARTICLE
(a)
(b)
(c)
Fig. 4. (a) Restored MR image for K = 2, Rs = 3, Rsearch = 9, PFOM = 0889 and PSNR = 5177. (b) Gradient of restored image. (c) Binary edge map of
restored image.
(a)
(b)
(c)
Fig. 5. (a) Restored MR image for K = 4, Rs = 3, Rsearch = 9, PFOM = 0853 and PSNR = 505. (b) Gradient of restored image. (c) Binary edge map of
restored image.
suppressed and only true edge pixels remain. This region is at the
midway between the regions of optimum noise suppression and
maximum edge preservation. This useful region lies in between
the values of coefficient of noise variance 4 and 8 for both MR
images in Figure 2. The restored images, gradient and binary
edge map of denoized images corresponding to this region, are
furnished in Figures 6(ac) and Figures 8(ac). It is also apparent
from Figures 14(a, b) that the radius of similarity window has
(a)
(b)
(c)
Fig. 6. (a) Restored MR image for K = 6, Rs = 3, Rsearch = 9, PFOM = 0726 and PSNR = 4944. (b) Gradient of restored image. (c) Binary edge map of
restored image.
RESEARCH ARTICLE
(a)
(b)
(c)
Fig. 7. (a) Restored MR image for K = 8, Rs = 3, Rsearch = 9, PFOM = 0675 and PSNR = 4848. (b) Gradient of restored image. (c) Binary edge map of
restored image.
(a)
(b)
(c)
Fig. 8. (a) Restored MR image for K = 10, Rs = 3, Rsearch = 9, PFOM = 0638 and PSNR = 4764. (b) Gradient of restored image. (c) Binary edge map of
restored image.
suppressed. Beyond this optimum value of the radius of the similarity window, Rs has no influence on noise suppression or edge
preservation.
Variation of PFOM with respect to the radius of search window on MR image 1 and MR image 2 for the optimum range
of coefficient of noise variance (k = 4 to k = 8) and Rs = 1 are
depicted in Figures 16(a, b). It has been already made evident
that the radius of similarity window should be greater than 3.
But the radius of search window should be always considerably
(a)
(b)
(c)
Fig. 9. (a) Restored MR image for K = 12, Rs = 3, Rsearch = 9, PFOM = 0609 and PSNR = 4691. (b) Gradient of restored image. (c) Binary edge map of
restored image.
RESEARCH ARTICLE
(a)
(b)
(c)
Fig. 10. (a) Restored MR image for K = 14, Rs = 3, Rsearch = 9, PFOM = 0584 and PSNR = 4628. (b) Gradient of restored image. (c) Binary edge map of
restored image.
(a)
(b)
(c)
Fig. 11. (a) Restored MR image for K = 16, Rs = 3, Rsearch = 9, PFOM = 0564 and PSNR = 4573. (b) Gradient of restored image. (c) Binary edge map of
restored image.
blurring. Hence, it can be concluded that regardless of the information content, texture and noise content, while denoising MR
images Rs and Rsearch should be greater than 3 and the decay
control parameter should be in the range of 48.
The variation of PFOM with respect to the variation in coefficient of noise variance at the optimum values of radii of similarity and search window Rs = 3 and Rsearch = 9 for multiple
MR images with different textural content and noise levels (1 =
0.736, 2 = 0253 and 3 = 0268) is depicted in Figure 16(c).
(a)
(b)
For all the MR images used for the experiment, the optimum
value of coefficient of noise variance which offers maximum
edge preservation and noise supression is with in the range of
48 typically at 7.
4. DISCUSSIONS
Non local means filter being an edge preserving restoration
scheme, while fixing the value its operational parameters,
(c)
Fig. 12. (a) Restored MR image for K = 18, Rs = 3, Rsearch = 9, PFOM = 0546 and PSNR = 4524. (b) Gradient of restored image. (c) Binary edge map of
restored image.
RESEARCH ARTICLE
(a)
(b)
(c)
Fig. 13. (a) Restored MR image for K = 20, Rs = 3, Rsearch = 9, PFOM = 0529 and PSNR = 448. (b) Gradient of restored image. (c) Binary edge map of
restored image.
(a)
(b)
0.9
Similarity window=3
Similarity Window=5
Similarity Window=7
0.8
0.75
0.7
0.65
0.85
0.8
0.75
0.7
0.6
0.65
0.55
0.6
0.5
10
12
14
16
18
Similarity window=3
Similarity Window=5
Similarity Window=7
0.9
0.85
20
0.55
10
12
14
16
18
20
Fig. 14. (a) Variation of PFOM with respect to the variation in coefficient of noise variance on MR image 1. (b) Variation of PFOM with respect to the variation
in coefficient of noise variance on MR image 2.
(a)
(b)
0.84
K=4
K=6
K=8
0.82
0.8
0.78
0.76
0.74
0.72
0.82
0.8
0.78
0.76
0.74
0.7
0.72
0.68
0.7
0.66
K=4
K=6
K=8
0.84
0.68
1
10
11
10
11
Fig. 15. (a) Variation of PFOM with respect to the radius of similarity window on MR image 1. (b) Variation of PFOM with respect to the radius of similarity
window on MR image 2.
RESEARCH ARTICLE
(a)
0.9
K=4
K=6
K=8
0.85
0.8
0.75
0.7
0.65
10
11
(b)
0.9
K=4
K=6
K=8
0.88
0.86
0.84
0.82
5. CONCLUSION
0.8
0.78
0.76
0.74
0.72
10
11
0.95
Image 1
Image 2
Image 3
0.9
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
2
10
12
14
16
18
20
Conflict of Interest
None of the authors have any conflict of interest to declare.
RESEARCH ARTICLE
11. B. Farmer, A. S. Govindarajan, and R. T. Suryaprakash, Investigation
of
the
non-local
means
algorithm,
Available:
wwwpersonal.umich.edu/arunsg/NLMeans.pdf.
12. P. Coupe, P. Yger, S. Prima, P. Hellier, C. Kervrann, and C. Barillot, An optimized blockwise nonlocal means denoising filter for 3-D magnetic resonance
images. IEEE T. Med. Imaging 27, 425 (2008).
13. C. Tomasi and R. Manduchi, Bilateral filtering for gray and color images,
Proceedings of Sixth International Conference on Computer Vision, Bombay,
India, January (1998).
14. J. Immerkear, Fast noise variance estimation. Comput. Vis. Image Und.
64, 300 (1996).
15. W. K. Pratt, Digital Image Processing, Wiley, New York (1978).
16. J. J. Ranjani and S. J. Thiruvengadam, Dual-tree complex wavelet transform
based SAR despeckling using inter-scale dependence. IEEE T. on Geosci.
Remote. 48, 2723 (2010).
17. R. C. Gonzalez and R. E. Woods, Digital Image Processing, Addison-Wesley,
Reading, Mass (1992).
18. S. S. Agaian, K. A. Panetta, S. C. Nercessian, and E. E. Danahy, Boolean
derivatives with application to edge detection for imaging systems. IEEE T.
Syst. Man. Cyb. B 40, 371 (2010).
19. P. Bao, L. Zhang, and X. Wu, Canny edge detection enhancement by scale
multiplication. IEEE T. Pattern Anal. 27, 1485 (2005).
20. P. Trahanias and A. Venetsanopoulos, Vector order statistics operators as
color edge detectors. IEEE T. Syst. Man. Cyb. B 26, 135 (1996).
21. A. N. Evans and X. U. Liu, A morphological gradient approach to color edge
detection. IEEE Trans. on Image Process 15, 1454 (2006).
11