Image Denoising Using Earth Mover's Distance and Local Histograms

Nezamoddin N. Kachouie
International Journal of Image Processing, Volume (4): Issue (1) 66
Image Denoising Using Earth Mover's Distance and Local
Histograms
Nezamoddin N. Kachouie nezam.nk@gmail.com
Department of Systems Design Engineering,
University of Waterloo,Waterloo, ON, Canada
Present Affiliation: Harvard-MIT Health Sciences and Technology
Harvard Medical School,Cambridge,MA,USA
Abstract
In this paper an adaptive range and domain filtering is presented. In the
proposed method local histograms are computed to tune the range and domain
extensions of bilateral filter. Noise histogram is estimated to measure the noise
level at each pixel in the noisy image. The extensions of range and domain filters
are determined based on pixel noise level. Experimental results show that the
proposed method effectively removes the noise while preserves the details. The
proposed method performs better than bilateral filter and restored test images
have higher signal to noise ratio than those obtained by applying popular
Bayesshrink wavelet denoising method.
Keywords: Denoising, Bilateral filtering, Local histogram, Earth mover’s distance.
1. INTRODUCTION
Noise elimination is an important concern in image processing and computer vision. Images
obtained from the real world are corrupted with noise. The image noise might decrease to some
negligible levels under ideal conditions such that denoising is not necessary, but in general to
recover the image the corrupting noise must be removed for practical purposes. Noise makes
ambiguities in the underlying signal relative to its observed form by perturbations which are not
related to the scene under study. The goal of denoising is to remove the noise and to retain the
important signal features as much as possible. Linear filters, which consist of convolving the
image with a constant matrix to obtain a linear combination of neighborhood values, have been
widely used for noise elimination in the presence of additive noise. However they can produce a
blurred and smoothed image with poor feature localization and incomplete noise suppression.
Gaussian filters are typical linear filters that have been widely used for image denoising.
Gaussian filters assume that image signals have smooth spatial variations and pixels in a
neighborhood have close values, so noise will be suppressed while signal will be preserved by
averaging pixel values over a local neighborhood. The assumption of slow spatial signal
variations works well in smooth regions; however it fails and undesirably blurs the signal where
spatial variations are high such as edges.
To overcome this shortcoming and prevent undesirable blurring in regions with high spatial signal
variations, a number of filters in spatial and spatial-frequency domain are proposed. The most
popular ones in spatial domain are anisotropic diffusion [1-3], bilateral filtering and its extensions
[4-8]. Diffusion based methods iteratively solve partial differential equations and average the

signal over spatial neighborhood whose extension is determined based on local signal variations.
Bilateral filtering also known as range and domain filtering is a non-linear filter which performs
weighted averaging in both range and domain.
Bilateral filtering was introduced by Tomasi and Manduchi [4] to smooth noisy images while
preserve edges using neighboring pixels. Bilateral filtering is a local, nonlinear, and noniterative
technique which considers both gray level (color) similarities and geometric closeness of the
neighboring pixels. In a traditional domain filter, weight of the pixels decays by distance from the
center of the filter. Low pass filters assume that spatial variations is slow over the image, so by
weighted averaging of pixel values in a neighborhood, noise will be averaged away while the
signal will be preserved. However, this assumption fails at edges where the spatial variations are
not smooth and application of the low pass filter blurs the edges. Bilateral filter overcomes this by
filtering the image in both range and domain. Pixels in a neighborhood are considered close
either based on their spatial location (domain), or based on their pixel values (range). Therefore
bilateral filter averages pixel values based on weights that decay by both distance and pixel
dissimilarity.
There are several extensions to improve bilateral filtering [5-8]. In [6], a training-based bilateral
filtering is proposed where a general degradation model is considered for degraded images. Then
a restoration algorithm is developed to restore the degraded images with unknown degradation
process. Therefore, the success of restoration process depends on the general definition of
degradation model. In [7,8], different methods to speed up the bilateral filtering have been
proposed.
Nonlinear filters in spatial-frequency domain have also been proposed to preserve detail signal
and suppress the noise. The most popular ones are wavelet based denoising techniques [9-12].
In wavelet based denoising methods, the noise is estimated and wavelet coefficients are
thresholded to separate signal and noise. Various approaches to nonlinear wavelet-based
denoising have been introduced among them Bayesshrink wavelet denoising is developed in the
Bayesian framework and has been widely used for image denoising [10-11].
In this paper, an adaptive technique is proposed to tune the extensions of range and domain
filters. In the proposed method, the distance of the local histogram from the estimated noise
histogram is measured using earth mover's distance. The measured distance at each spatial
location is then used for adaptive tuning of bilateral filter. The proposed method provides
promising results and effectively removes the noise while preserves the signal characteristics.
The proposed method is presented in the next section followed by results and conclusions.
2. The Proposed Method
Let pure signal S (here an image) be distorted by additive noise n. We can write
where I is the noisy signal. The goal of denoising is separating signal S and noise n by estimating
n such that S can be extracted from I:
This can be done by applying a filter h to the signal I

where traditionally h is defined as a local filter assigning higher weights to neighboring pixels
which are spatially closer to the central pixel xc of the neighborhood. A popular and simple case
of h is Gaussian filter
where µd = xc is the central pixel of the neighborhood such that d(x, µd) = |x - µd | is the Euclidean
distance between xc and a neighboring pixel x. Gaussian domain filtering by using a Gaussian
filter averages away noise and preserves the signal in smooth regions, however in the same way
it averages away and blurs signal details such as edges. A popular solution to solve this problem
is employing bilateral filter [4].
Bilateral Filter
Bilateral filter combines range and domain filtering
where the range filter averages the signal values in a neighborhood by assigning the weights
based on the similarity of the neighboring pixels and the central pixel:
where µr = I(µd) = I(xc) is the intensity value of the central pixel of the neighborhood such that
r(I(x), µr) = |I(x) - µr| is the absolute intensity difference of the central and a neighboring pixel x.
Bilateral filtering overcomes the shortcomings of linear domain filtering by combining the linear
domain filter with a nonlinear range filter. As a result bilateral filter preserves signal details such
as edges while suppresses noise, however it considers fix parameters (σd, σr) for extensions of
both domain and range filters. The performance of bilateral filter can be improved by adaptively
tuning the filter parameters over the image based on the spatial noise level.
Adaptive Range and Domain Filtering
In the proposed method, to improve the performance of bilateral filtering, spatial noise level is
locally estimated to determine the filter parameters (σd, σr).
To estimate the local spatial noise level nl, the image noise histogram ng is estimated and
compared with the local signal. To compare two probability density functions (PDFs), a number of
nonparametric models have been used including minimizing the comparison ‫א‬2
function between
two PDFs.
The ‫א‬2
distance between histograms of two delta functions δ(x1) and δ(x2) where x1 ≠ x2 is the
same regardless of the distance between x1 and x2. This is not generally suitable for many image
processing applications where different smooth regions could be represented with disjoint δ
functions.
The earth mover's distance (EMD) or the Wasserstein distance is a mathematical measure to
compare distributions (histograms). EMD was first introduced by Gaspard Monge in 1781, it was
later used as a distance measure for intensity images [13]. The EMD between two distributions is

the least work that is required to move one distribution to another such that two distributions
completely cover each other.
Let Ha and Hb be two normalized histograms with cumulative distributions Ca and Cb respectively.
EMD between Ha and Hb is defined by
Local histogram for each pixel x in image I is computed over the neighborhood w consisting pixel
x and its neighboring pixels. The EMD is then computed to compare the normalized local
histogram Hx and image noise histogram ng
where Cx and Cn are cumulative distributions of Hx and ng respectively. The extensions of domain
and range filters (σd and σr) at each pixel x are set using E(Hx, ng). The domain filter extension at
pixel x is defined as
and we have
where E(Hx, ng) is normalized EMD between noise and pixel histograms, σ is the filter extension
parameter, and d is considered to avoid domain filter extension σd to be set to zero. The range
filter extension also is tuned based on E(Hx, ng)
and we have
where r is considered to avoid range filter extension σr to be set to zero.

TABLE 1: Comparison of the proposed method with bilateral and
Bayesshrink wavelet filtering methods.
Clearly there is a tradeoff here to choose the domain filter extension σd: as the filter extension σd
expands the number of neighborhood elements grows, allowing for greater noise reduction in the
computation but at the same time causing greater spatial blurring by fusion of values from more
distant locations. Moreover, the range filter essentially compresses the image histogram by fusion
of pixel values and is set by σr.
In the proposed method the maximum of σd and σr are set by σ based on equations (9) and (11).
As 0 ≤ (1 - E) ≤ 1, for pixels which are contaminated with high noise, the distance between noise
and the pixel histograms E is small, therefore (1 - E) will be large. Considering σ is fixed, σd will
be large allowing the neighborhood to be extended for greater noise reduction while σr will be
large based on (11) to allow significant histogram compression.
On the other hand for pixels which are contaminated with low noise, E is large, therefore (1 - E) is
small, and in turn σd will be small avoiding the neighborhood to be extended which in turn it
allows less blurring. Considering that the pixel either is not contaminated with noise or is
contaminated with low noise, σd will be small. Pixels have close values in small neighborhood,
therefore σr will be small avoiding significant histogram compression.
Noise Histogram Estimation
To estimate the noise histogram (ng), the local variance is first computed. Considering the local
neighborhood w, the local variance of pixel x is defined as:
where Nw is the number of pixels in the neighborhood w and
Noise histogram ng is estimated by computing the histogram of the local variance image. Further,
we set σ = σn where the noise power σn
2
is estimated by obtaining the mean of local variance
image:

Finally, the local histogram Hx is computed for each pixel x and EMD is used to measure the
distance between noise histogram ng and local histogram Hx. The schematic of noise histogram
estimation is depicted in Fig. 1.
Figure 1: Noise histogram estimation
3. Results and CONSLUSION
To test the proposed method five test images were used. Test images were corrupted by additive
Gaussian noise with standard deviation of 15 and 25. The proposed method, the original bilateral
filter, and a popular wavelet denoising method so called Bayesshrink wavelet denoising were
applied to the corrupted test images. The recovered images applying aforementioned three
methods were compared both based on PSNR and visual quality. The results are summarized in
Tab. 1.
As we can observe in Tab. 1 for additive Gaussian noise with standard deviation of 15, the
proposed method performs better than the original bilateral filtering method. It gains higher PSNR
than both the original bilateral filtering and Bayesshrink wavelet denoising methods for all of the
test images. The recovered images applying the proposed method have also better visual quality.
The recovered images applying Bayesshrink wavelet and the proposed method are depicted in
Fig. 2. The proposed method performs better and the restored image gains higher PSNR (Tab.
1). It has also a better visual quality than that of Bayesshrink method which can be observed by a
closer look. Fig. 3 shows the Boat noisy image and EMD computed for the noisy image. The
denoised Boat image using the original bilateral filtering, Bayesshrink wavelet, and the proposed
method are depicted in this figure.

Figure 2: Restored Lena test image: (a) Bayesshrink wavelet. (b) The proposed method.
Figure 3: Boat test image: (a) Original image. (b) Noisy. (c) EMD computed for (b).
(d) Bilateral. (e) Bayesshrink wavelet. (f) The proposed method.
Fig. 4 shows the application of bilateral filter and the proposed method to the Cameraman test
image where the image is corrupted with additive Gaussian noise with σn = 25. The application of
Bilateral filtering and the proposed method to Goldhill test image which is corrupted with additive
Gaussian noise with σn = 15 is depicted in Fig. 5. Fig. 6 shows the comparison of the
Bayesshrink, bilateral, and the proposed method where they are applied to the Lena test image
corrupted with additive Gaussian noise with σn = 15. As we can observe in Fig. 4-6, the proposed

method performs better and produces smoother results while the details are better preserved in
comparison with bilateral filter and the Bayesshrink. It also gains higher PSNR (Tab. 1).
In this paper an adaptive range and domain filtering method based on local histograms was
introduced. The noise histogram is estimated and the extensions of range and domain filters are
tuned at each spatial location by measuring the distance between the pixel's and noise
histograms using earth mover's distance. The proposed method was applied to several test
images and its performance was compared with the original bilateral filtering and Bayesshrink
wavelet denoising methods. The experimental results obtained by the proposed method showed
the improvement of the visual image quality and increase of PSNR in comparison with the
bilateral filtering and Bayesshrink wavelet.
Figure 4: Cameraman test image: (a) Original image. (b) Noisy image (PSNR = 20.65).
(c) Bilateral filtering (PSNR = 24.70). (d) The proposed method (PSNR = 26.00).

Figure 5: Goldhill test image: (a) Original image. (b) Noisy image (PSNR = 24.71).

Figure 6: Lena test image: (a) Original image. (b) Bayesshrink (PSNR = 29.06).
4. REFERENCES
1. P. Perona and J. Malik, “Scale-space and edge detection using anisotropic diffusion”, IEEE
Tran. on PAMI, 12(7), pp. 629-639, 1990.
2. G. Sapiro and D. L. Ringach, “Anisotropic diffusion of color images”, in Proc. Society of Photo-
Optical Instrumentation Engineers (SPIE) Conference, 2657, pp. 471-482, 1996.
3. M. Ceccarelli, V. D. Simone, and A. Murli, “Well-posed anisotropic diffusion for image denois-
Ing”, IEE Proc. on VISP, 149(4), pp. 244-252, 2002.
4. C. Tomasi and R. Manduchi, “Bilateral filtering for gray and color images”, in Proceedings of
Intl Conference on Computer Vision (ICCV), pp. 836-846, 1998.

5. J. Xie and P. A. Heng, “Color image diffusion using adaptive bilateral filter”, in 27th Annual
EMBS International Conference, pp. 3433-3436, 2005.
6. B. Zhang and J. P. Allebach, “Adaptive bilateral filter for sharpness enhancement and noise
Removal”, IEEE Tran. on Image Processing, 17(5), pp. 664-678, 2008.
7. S. Paris and F. Durand, “A fast approximation of the bilateral filter using a signal processing
Approach”, MIT technical report, MIT-CSAIL-TR-2006-073, 2006.
8. M. Elad, “On the bilateral filter and ways to improve it”, IEEE Tran. on Image Processing,
10(11), pp. 1141-1151, 2002.
9. D. L. Donoho and I. M. Johnstone, “Ideal spatial adaptation via wavelet shrinkage”, Biometrika,
81(1), pp. 425-455, Sept 1994.
10. S. G. Chang, B. Yu, and M. Vetterli, “Adaptive wavelet thresholding for image denoising and
Compression”, IEEE Tran. on Image Processing, 9(9), pp. 1532-1546, 2000.
11. S. G. Chang, B. Yu, and M. Vetterli, “Spatially adaptive wavelet thresholding with context
modeling for image designing”, IEEE Tran. on Image Processing, 9(9), pp. 1522-1531, 2000.
12. M. N. Do and M. Vetterli, “The finite ridgelet transform for image representation”, IEEE
Transactions on Image Processing, 12(1), pp. 16-28, 2003.
13. S. Peleg, M. Werman, and H. Rom, “A unified approach to the change of resolution: Space
and gray-level”, IEEE Tran. on PAMI, 11(7), pp. 739-742, 1989.

Image Denoising Using Earth Mover's Distance and Local Histograms

More Related Content

What's hot (19)

Similar to Image Denoising Using Earth Mover's Distance and Local Histograms (20)

Recently uploaded (20)

Image Denoising Using Earth Mover's Distance and Local Histograms