Adaptive Bilateral Filter and Bayesian Threshold Based Image Denoising

G.Santhanamari1 , DR.V. R.Vijaykumar2, A.V.V. Bhaskar Rao3

1
Department of Electronics and Communication Engineering, Tamilnadu college of Engineering, Coimbatore
2
Department of Electronics and Communication Engineering, Anna University of Technology (AUT), Coimbatore
3
Research Scholar, Department of Electronics and Communication Engineering, AUT, Coimbatore
([email protected] , [email protected] , [email protected])

The bilateral filter is essentially a smoothing filter

Abstract - In this paper a hybrid denoising algorithm and it does not restore the sharpness of degraded image.
which combines adaptive bilateral Filter and Bayesian However in order to enhance sharpness, bilateral filter is
thresholding for digital images corrupted by Gaussian noise improved with some modifications lead to Adaptive
is proposed. The wavelet filter bank is used to decompose the Bilateral Filter (ABF). Nonlinear filters in spatial-
image into approximation sub band and detail sub band.
The adaptive bilateral filter is applied to approximation sub
frequency domain have also been proposed to preserve
band and Bayesian thresholding is applied to detail sub detail signal information and suppress the noise. The most
band. The proposed algorithm is tested on gaussian noise popular ones are wavelet based denoising techniques [5-
corrupted images. The observation of parameters like visual 7].In wavelet based denoising methods, the noise is
quality results and quantitative performance in terms of estimated and wavelet coefficients are thresholded to
PSNR reveals that the proposed algorithm performs better separate signal and noise. Among various approaches to
than the other existing methods in terms of noise removal nonlinear wavelet-based denoising, Bayesshrink wavelet
and edge preservation. denoising based on Bayesian framework has been widely
used for image denoising [8].
Keywords –Adaptive bilateral filter, Gaussian noise,
This paper has been organized in the following
Bayesian thresholding, Wavelet decomposition
manner. Section II deals with background works, section
III proposes algorithm steps, section IV discusses the
simulation results and finally section V gives concluding
I. INTRODUCTION
remarks.
In most applications image is degraded due to loss
of sharpness (blur) and presence of noise, during image
II. BACKGROUND WORKS
acquisition or transmission. Noise having Gaussian-like
distribution is very often encountered in image
A. Bilateral filter
acquisition, which is characterized by adding to each
image pixel a value from a zero-mean Gaussian Bilateral filtering is a local, nonlinear, and non
distribution. The zero-mean property of the distribution iterative technique which considers both gray level (color)
allows such noise to be removed by locally averaging similarities and geometric closeness of the neighboring
pixel values [1]. The goal of denoising is to remove the pixels. In a traditional domain filter (low pass), there is an
noise and to retain the important signal features as much assumption that spatial variations are slow over the image,
as possible for sharpness enhancement. Conventional so by weighted averaging of neighborhood pixel values,
linear filters such as arithmetic mean filter and Gaussian noise is averaged away while the signal is preserved.
filter smooth noises effectively but blur edges. Since However, this assumption fails at edges where the spatial
gaussian filters assume that image signals have smooth variations are not smooth and application of the low pass
spatial variations and pixels in a neighborhood have close filter blurs the edges. Bilateral filter overcomes this by
values, noise will be suppressed while signal will be filtering the image in both range and domain. Therefore
preserved by averaging pixel values over a local bilateral filter averages pixel values based on weights that
neighborhood. The assumption of slow spatial signal decay by both distance and pixel dissimilarity. As a result
variations works well in smooth region. However it fails bilateral filter preserves signal details such as edges while
and undesirably blurs the signal where spatial variations suppresses noise, with fixed parameters (d, r).
are high such as edges. Since the goal of the filtering Bilateral filter [4] combines range and domain
action is to cancel noise while preserving the integrity of filtering as given in equation (1)
edge and detail information, nonlinear approaches
generally provide more satisfactory results than linear
h( x, Pd ,V d , Pr ,V r ) hd ( x, Pd ,V d ) hr ( x, Pr ,V r ) (1)
techniques. To prevent the undesirable blurring in regions
with high spatial signal variations, one of the most ­° 1 § x  P ·2 ½°
popular non linear filtering technique called bilateral h hd ( x, Pd , V d ) exp ® ¨ d
¸ ¾ (2)
filtering introduced by Tomasi and Manduchi and its 2 © Vd ¹ °
extensions[2-4] can be used to smooth noisy images while ¯° ¿
preserve edges using neighboring pixels.

978-1-4244-5967-4/10/$26.00 ©2010 IEEE

 ­° 1 § I ( x)  P ·2 ½° Gaussian Distribution (GGD) prior used on wavelet co-
hr ( x, Pr , V r ) exp ® ¨ r
¸ ¾ (3)
efficient. It is an adaptive threshold which is determined
2 © V ¹ °¿ by the parameters , n as in [8] where , n2, y2 represent
¯° r
signal variance, noise variance and measurement
Where x is neighboring pixels, d, r are standard variance. If 2  y2 then TB = max (|Yij|).
deviation of range and domain filters and μr, μd are mean
value of range and domain filters respectively.
Where r = I(d) = I(xc) is the intensity value of III: PROPOSED ALGORITHM
the central pixel of the neighborhood such that r(I(x), r)
= |I(x) - r| is the absolute intensity difference of the Step1: Let Xij be the pixel intensity value at a position i, j
central and a neighboring pixel x. in an original image and nij is a noisy value at i, j which is
drawn from zero mean Gaussian distribution and then
B. Adaptive Bilateral Filter (ABF)
noisy image Yij is given by equation (6).
The performance of bilateral filter can be Yij = Xij + nij (6)
improved by adaptively tuning the filter parameters (d, Step2: Take discrete wavelet transform on noisy image as
r) over the image based on the spatial noise level leads to given in equation (7).

¦ ¦ x(k)2
a new sharpening and smoothing filter called an adaptive
bilateral filter [9] whose impulse response and filtering  W( j, k) j k
M(2 j n k)
 j/2
(7)
operation are given in the equations (4) and (5).
It is followed by correction step in which wavelet co
m0 N §  mm0 2  nn0 2 · efficient is rectified according to set of mapping function
h[m0, n0;m, n] r ¦ ¦n n N exp¨
1 n0 N
¸
0 ¨ 2Vd2 ¸ given in equations (8) and (9).
m m0 N
© ¹ ___
§  g[m, n]  g[m0 , n0 ] ] [m0 , n0 ] ·
2 MW [ M W1 .M W2 .....] (8)
*exp¨  ¸ (4) š
¨ 2 V 2
[m , n ] ¸ ___
© d 0 0 ¹
[m,n]:m0 n0 XW M W { yw} (9)
š

š X W is a vector of scalar mapping function (shrinkage

f  m, n  h > m0 , n : m, n @ g > m, n @ (5) function to transform coefficient).
Step3: The ABF whose impulse response defined in
Adaptive bilateral filter retains general form of
equation (4) is applied to the low frequency components
bilateral with  = 0, and fixed r. In ABF domain filter is a
low pass Gaussian filter with d =1.0 and range filter with (LL). The Bayesian soft thresholding determined as in
locally adapted variables , r are adapted for achieving equations (10) - (12) is applied to the high frequency
both sharpening and smoothening in denoised image. The components (LH, HL, and HH) as shown in Figure.1.
parameters , r are estimated as in [9]. Vˆ 2
TˆB Vˆ X  (10)
C. Bayes shrink threshold Vˆ X
Where
Wavelet shrinkage is a method of removing noise
M e d i a n | Y i j | 
from images. In wavelet shrinkage, an image is subjected Vˆ , Y i j   S u b b a n d  H H 1 (11)
to the wavelet transform, the wavelet coefficients are 0 .6 7 4 5
found, the components with coefficients below a Vˆ X max Vˆ Y2  Vˆ 2  (12)
threshold are replaced with zeros. The image is then
reconstructed. Bayes shrink gives the soft threshold which Step4: The restored image is obtained by taking inverse
minimizes the Bayesian risk. There are two reasons for wavelet transform to the modified coefficients as given in
preferring soft threshold. First one is soft-thresholding equation (13).
has been shown to achieve near-optimal minimax rate  Xˆ W  1 Xˆ w (13)
over a large range of Besov spaces and the second one is
for the generalized gaussian prior, the optimal soft-
thresholding estimator yields a smaller risk than the
optimal hard-thresholding estimator.
Soft-thresholding method yields more visually
pleasant images over hard-thresholding because the
hard-thresholding is discontinuous and yields abrupt
artifacts in the recovered images especially when the
noise energy is significant. Bayes shrink threshold [8] is
derived from Bayesian frame work with Generalized Figure.1 Block diagram of proposed algorithm
 TABLE 1
PSNR FOR VARIOUS FILTERS FOR GRAY SCALE IMAGES AT DIFFERENT NOISE DENSITIES

Standard Bayes Bilateral Multi resolution Proposed

Image
Deviation Shrink Filter Bilateral Filter Method
10 33.38 33.65 33.71 33.91
20 30.27 30.33 31.28 31.56
Lena 30 28.60 28.54 28.71 28.91
40 26.25 26.79 26.81 26.98
10 30.25 30.37 30.79 30.99
20 27.32 27.02 27.74 27.89
Barbara
30 25.34 25.69 25.61 25.78
40 21.09 22.34 22.45 22.77
10 31.98 32.02 32.58 32.65
20 28.55 28.40 29.25 32.44
boats
30 26.71 26.57 27.24 32.16
40 24.12 23.91 23.32 31.71

(a) (b) (c) (d) (e) (f)

Fig 2. (a).Original image (b).Noisy image with noise standard deviation =30 (c).Result of Bilateral filter (d). Result of Bayesian threshold
(e). Result of Multi resolution bilateral filter. (f). Result of proposed method

(a) (b) (c) (d) (e) (f)

Fig 3. (a).Original image (b).Noisy image with noise standard deviation =40 (c).Result of Bilateral filter (d). Result of Bayesian threshold
(e). Result of Multi resolution bilateral filter. (f). Result of proposed method

(a) (b) (c) (d) (e) (f)

Fig 4. (a).Original image (b).Noisy image with noise standard deviation =50 (c).Result of Bilateral filter (d). Result of Bayesian threshold
(e). Result of Multi resolution bilateral filter. (f). Result of proposed method
 IV.SIMULATION RESULTS AND DISCUSSION IV. CONCLUSION

In this section we discuss about experimental The proposed filter integrates Adaptive bilateral
results on some test images. The proposed algorithm is filtering and Bayes shrink wavelet thresholding for
tested using gray scale images such as Lena, Barbara, and effective removal of Gaussian noise. The performance of
Boats, of size 512 × 512 of 8 bits/pixel and restored proposed algorithm tested for various images at different
images are shown in fig.2 - 4. The performance of noise variance interms of PSNR. The subjective and
proposed algorithm is tested quantitatively based on the objective quality of proposed filter reveals that it performs
parameter PSNR using the equation (14) for various better than the existing algorithms namely bilateral filter,
standard deviations of Gaussian noise level. Bayes shrink wavelet threshold, Multi resolution bilateral
filter. The performance of proposed filter can further be
ª º improved by fine tuning of various parameters in ABF
« 2552 » and Bayes wavelet threshold and can also be applied to
PSNR 10 log10 « »
1
«
¬ MN
¦ i ¦ j (Yij  X ij ) »¼
2 real images corrupted by Gaussian noise.
(14)
Where Yi,j and Xi,j denotes pixel values of restored
image and original image respectively and M, N are the REFERENCES
size of the image. The restoration results are compared
with bilateral filter Bayes shrink wavelet thresholding and [1] Gonzalez and Woods, “Digital Image Processing”, Prentice
Multi resolution filter based on which the proposed hall, 2nd Edition, 2001.
algorithm is developed. The quantitative measure based [2] J. Astola and P. Kuosmanen, “Fundamentals of Nonlinear
Digital Filtering”, Boca Raton, FL : CRC, 1997.
on PSNR values is shown in table 1 and also the same is
[3] Nezamoddin. N. Kochouie “Image Denoising Using Earth
plotted as a graph. The proposed method shows Mover's Distance and Local Histograms”, International
improvements in PSNR values for all test images. journal of Image Processing, Vol. 4, no. 1, pp. 66-75,
In addition to quantitative measure, the March 2010.
performance of proposed algorithm is qualitatively [4] C. Tomasi and R. Manduchi, “Bilateral filtering for gray
measured based on visual inspection carried out on the and color images” in Proc. Intl Conference on Computer
filtered image, since image is subjective to human eyes. Vision (ICCV), pp. 836-846, 1998.
The reconstructed image based on Bayes shrink wavelet [5] Soman K.P., “Insight into wavlets: From theory to
threshold is smoothed and visually appealing. In Multi practice”, Prentice Hall, 2nd Edition, 2005-2012.
[6] S.Grace Chang, Bin Yu, Martin Vetterli, “Adaptive
resolution bilateral filter based image denoising [10],
Wavelet Thresholding for Image Denoising and
bilateral filter is applied to approximation sub band of Compression”, IEEE Transaction on Image processing, vol.
wavelet decomposed image and hence edges are having 9, no. 9, pp. 1532-1546, September 2000.
the same level blurriness as in original degraded image, [7] Hel-or, Y.Shaked, D.Efi Arazi, “A Discriminative approach
although the noise is greatly reduced. The denoised for wavelet denoising”, IEEE Transaction on Image
image based on bilateral filtering, Bayes shrink wavelet processing, vol. 17, no. 4, pp. 443-457, April 2008.
denoising, Multi resolution bilateral filtering and [8] Weiling Wang Yufeng Li “ Bayesian Denoising for Remote
proposed algorithm for lena, Barbara, and boats are Sensing Image Based on Undecimated Discrete Wavelet
shown from fig.2.c - 2.f to 4.c – 4.f and comparison of Transform” , in Proc. Conf. Information Engineering and
Computer Science, ICIECS 2009, pp. 1-4.
PSNR values are plotted as a graph shown in fig.5
[9] Buyue Zhang , Jan P. Allebach “Adaptive Bilateral Filter
for Sharpness Enhancement and Noise Removal” IEEE
Trans. On image processing, vol. 17, no. 5, pp. 664-678,
May 2008.
[10] Ming Zhang and Bahadir K. Gunturk, “Multiresolution
Bilateral Filtering for image denoising” IEEE Transactions
on Image Processing, vol. 17, no. 12, pp. 2324-2333, Dec
2008.

Fig 5. PSNR of various denoising methods at different noise

standard deviation