A new image denoising method

based on Gaussian filter

Sheng Zheng，Xiaolong Li,Xiongjie Qin
Manyu Wang College of Science
College of Computer and Information Technology China Three Gorges University
China Three Gorges University Yichang,Hubei, China
Yichang,Hubei, China
wangmanyu123@126.com

Abstract—In this paper, we propose a new image denoising distance between the current pixel and the pixel set. Although
method based on Gaussian filter and Non-local means filter. The these methods can reduce the noise to some extent, they
new algorithm dealing with the image noise learns from the destroy the original image details and bring about image
weighted average thoughts of particle filter. we analyse the blurring. So we need to find a proper compromise between
difference between the proposed method and other two image denoising and keep image details. In recent years, some
algorithms, then we do multi-group experiments to compare the researchers try to restore the lost information of the noisy
denoising effect of these methods, and evaluate their image rather than smooth it.
performance mainly through visual effect and peak signal-to-
noise ratio (PSNR). The experimental results show that the This paper mainly put forward a new denoising method on
proposed new method is effective in image denoising. the basis of Gaussian filter and non-local means filter[5]. Our
new method dealing with the image noise learns from the
Keywords—Image denoising; denoising effect; visual effect; weighted average thoughts of particle filter[6]. From the
PSNR contrast of the experimental results, we can see that our method
is superior to the Gaussian filter and NL-means filter.
I. INTRODUCTION
This article is organized as follows. Section 2 gives a brief
Usually, an image will be influenced by a variety of noises introduction about the principle of Gaussian filter. Section 3
during the process of transmission, those noises will debase the introduces the central idea and innovation points of our method
quality of the image and influence our understanding and in detail. In section 4 we contrast the three methods, then
analysis about the original image information. Meanwhile, it analyse and discuss the experimental results, and the last
can make a bad effect on the whole process of image section is the conclusion of the whole paper.
processing. In order to reduce the noise and improve the image
quality, we need to smooth the image. Image denoise is an II. THE GAUSSIAN FILTER
indispensable link to the image processing, it is the
precondition of image understanding and image analysis, and it Gaussian filter is a linear smoothing filter that chooses the
is also one of the major research in image processing field. weights according to the shape of Gaussian function[7].
Whether in spatial domain, or in frequency domain, Gaussian
Generally speaking, Gaussian noise is the most common smoothing filter is a kind of effective low-pass filter, especially
noise in almost all images. In addition, there are two kinds of to remove the noises that are subject to the normal distribution.
typical noises, one is the salt &pepper noise, which have the Therefore it has a broad prospect of application in image
same amplitude but random position, the other is the random processing. For one-dimensional Gaussian function of zero
noise that distribute randomly and each pixel has amplitude. mean, it can be expressed as follow:
And there exist a wide variety of denoise methods, according
to the different domain of image processing, these methods can
be classified into the spatial domain de-noising and frequency x2

domain de-noising[1]. In the spatial domain, there are mean g ( x)  e 2
2
(1)
filter[2], median filter[3], neighborhood averaging method and so
on. And in the frequency domain, low pass filter and high pass Among them, the Gaussian distribution parameter
filter are the most common choices [4].
determines the width of Gaussian functions. And in image
Whether space domain de-noising, or frequency domain de- processing, we often use 2-d discrete Gaussian function of zero
noising, the basic idea of these methods is local average mean as smooth filter[8], the corresponding function expression
filtering, which means that we compute all pixels or part of the is:
pixels’ weighted average in a specified neighborhood. These
traditional denoising algorithms choose the pixel set to
participate in average operation according to the spatial
____________________________________
978-1-4799-3197-2/14/$31.00 ©2014 IEEE


 
i2  j 2
g ( x, y )  w( x ', y ')  u ( x ', y ') (3)
 g[i, j ]  e 2
2
(2) ( x ', y ') I

In addition, Gaussian function has five important properties Where w( x ', y ') is the normalized weight calculated by the
as follow: exponential function as formula (4), and satisfies the conditions
1) 2-d Gaussian function is rotational symmetry. 0 w( x ', y ') 1 and w( x ', y ')  1 . It depends on the
( x ', y ') I
Generally, we don’t know the images’ edge direction and not
sure about whether we should do more smooth in a direction similarity of pixel ( x ', y ') and the current point, those pixels
than the other one before filtering. Rotational symmetry with similar grey level neighborhood and edge structure to the
means Gaussian filter will not turn to any direction in the current pixel will get a greater weight when calculating their
average value.
subsequent image processing.
2) Gaussian function is a single value function. Gaussian
filter uses the weighted mean of the pixel neighbourhood to d ( x ', y ')2  m ( x ', y ')2
1 
replace the current pixel, and the weight is monotone w( x ', y ')  e
2
gf (4)
z ( x ', y ')
decreasing with the distance to the centre point become far
away. So it will have little effect on the pixels that are far And parameter
controls the decay of the exponential
away from the centre, and the image will not distort. function. z ( x ', y ') is a normalized factor.
3) The Fourier transform spectrum of Gaussian function is
single. And this property can help the images get rid of high
d ( x ', y ')2  m ( x ', y ')2
frequency signals pollution and retain most useful signals. 
4) The width of Gaussian filter is characterized by the z ( x ', y ')  e
2
gf (5)
parameter
. The bigger
is, the better the smoothness is,
and the frequency band of Gaussian filter becomes wider. By Here we set current pixel as the centre of the fixed window,
adjusting the smoothness parameters, we can get a then d ( x ', y ') is the grey variance difference between current
compromise between excessive smoothing and not smoothing pixel and any other pixels in the fixed window with the radius
thoroughly. of (ksize  1) / 2 . m( x ', y ') is the grey mean difference of
5) The Gaussian function is separate, so a bigger Gaussian current pixels and any other pixel in the fixed window with the
filter can be realized effectively. The convolution of 2-d radius of (ksize  1) / 2 . gf is a Gaussian filter with the size of
Gaussian function can be divided into two steps, and the
ksize and variance of
. And we choose different ksize and
calculated amount of 2-d Gaussian filtering grows linearly

to do the contrast experiments.
with the width of filter template increasing.
III. OUR METHOD
In essence, the proposed method is the improvement of
Gaussian filtering algorithm. In comparison, traditional
Gaussian filter only consider grey information of the images,
but our method considers both the grey information and image
edges structure information. It can remove noises while
retaining lots of useful details.
Meanwhile, our new method adds the thought of particle
filter into the weighting filter which is based on similarity
choosing. When select the weighted average pixels, it learns
from the importance resampling[11], each pixel’s weight is Fig. 1. Filtering window.
analogy to the importance density function.
Suppose that we choose the filtering window of 11 11
Specifically, first we calculate the grey mean and variance pixels, and set the template size as 5  5 . In order to calculate
of the original image, and then select parts of pixels that are the grey value of current pixel u ( x, y ) after denoising, we only
near around the current pixel’s neighborhood to calculate
current point’s grey value. need to figure out the similarity between the current point and
every pixel in the filtering window with the radius
For an image, the denoising grey value of any pixel of (ksize  1) / 2 . To calculate weight, first calculate the mean
g ( x, y ) is calculated by the weighted summation of the pixels and variance of the current pixel, and then subtract it with
within its ksize  ksize neighborhood. Suppose that the discrete every pixel in the filtering window. Next, put the
noise image is u  u ( x ', y ') | ( x ', y ') I , I is the fixed filtering corresponding result d ( x ', y ') and m( x ', y ') into formula (4) to
window area, the computational formula is as follow: calculate its weight. At last multiply each pixel within the
filtering window by its corresponding weight, and use the
weighted summation to instead of the original pixel grey value.


 IV. CONTRAST EXPERIMENT AND ANALYSIS
In our experiment, we discuss four kinds of image de-
noising methods, respectively they are mean filter, Gaussian
filter, NL-means filter and the proposed new method.
NL-means filter is an emerging and effective image
denoising filter which was proposed by Alexei Efros and
Thomas Leung in 1999[9,10]. It makes full use of a large number
of the similar phenomena in natural images filtering and has
made impressive results. The main idea of NL-means method
is to calculate the average grey scale of all the similar pixels.
Namely find out the pixels within the whole image which are Fig. 3. The PSNR value under different i0 (sigma)
similar to the current pixel, and calculate weighted summation
of these pixels to instead of the current pixel’s grey value. If we choose a much bigger variance to smooth the noise
Though NL-means can smooth the image well while retain image, then we will get a better smooth effect, and the
enough details and get a good robustness, but it also costs a corresponding PSNR becomes bigger. The following result
large computation and has less efficiency. also shows that when the variance of the smooth filter is small,
Gaussian filter’s PSNR is the largest one, and with variance
There we make a contrast among three methods for increasing slowly, our method stays ahead again.
Gaussian noise images, and use the standard grey image Lena
of size 256  256 as our test image, our templates is 5  5 .The
original size of Gaussian filter is set as 11, that is to say we fix
the search window as 11 11 pixels, so the corresponding
complexity of our algorithm is about 25  121 256  256 . The
initial parameter of the experiment is set as: i0 (sigma) = 0.01,
ksize = 11, gf (sigma) = 1.8, pd (sigma) = 8, besides, we did
the dot product between pd and gf. In figure 2, the images from
left to right and from the top to bottom are: Gaussian noise
image, the variogram, the denoising results of mean filter, our
method, Gaussian filter and NL-means filter.
Fig. 4. The PSNR value under different gf (sigma)

When choosing a bigger size for Gaussian filter, the smooth

scale and search window becomes larger. Accordingly, the
running time is increasing. So we need to get a balance
between the two factors in practical application, and try to use
different parameters for different images. In contrast with the
above methods, the PSNR of our denoising method always
rank first in almost all situations. But for the same method, the
change in filter size has little effect on the amplitude variation;
the change is not that significant as figure 5 shows.

Fig. 2. The initial experiment results

In addition to perceive the denoising results directly

through human eyes, we also use one of the commonly used
criterions PSNR to measure these denoising algorithms. PSNR
estimate the variance of the image noise, the bigger the PSNR
is, the better denoising effect we get.
In our experiment, we use the control variable method to
adjust different parameters, and can get the PSNR value of the Fig. 5. The PSNR value under different ksize
four methods under the control of different parameters.
In addition, when do a dot products between the weights pd
The experimental results show that when the variance and gf, we can get a much bigger PSNR, that is to say, filtering
sigma of the added Gaussian noise is bigger, the PSNR get the weight will get a better effect. At the same time, we will get
smaller, but under the same variance, in figure 3 we can see a much bigger PSNR value if we set a big value for the
that our method’s PSNR is obviously bigger than other variance of the weight function, and our method is still the best
methods. one of all the four methods.


Fig. 6. The PSNR value under different Pd (sigma)

Through comparing the different PSNR under the control Fig. 8. The magnified contrast experimental result
of different parameters, we can combine the best parameters of
each group, select the parameters which can get the largest In our experiment we only discuss and compare the
PSNR in each group, there we choose i0(sigma)=0.005, situation of Gaussian noise images, because the proposed
gf(sigma)=2.7,pd(sigma)=10,ksize=13, and finally get an method is mainly aimed at noises that are obey to the normal
experimental result as the first row of figure 7 shows. In figure distribution. So it doesn’t make any sense to do any contrast
7, from the left to the right, the denoising methods respectively with other three methods.
are: our method, Gaussian filtering and NL-means filtering.
The PSNR of these three methods are: 28.1995, 22.4545 and V. CONCLUSIONS
28.0254. From the graph, we can see that our method and NL- This paper proposes a new denoising algorithm based on
means filtering can keep enough edge details. In comparison, Gaussian filter and NL-means filter, but it is quite different
the denoising effect of NL-means filter is not good as our from them. Gaussian filter choose weight to smooth the image
method, but it can keep better detail texture than our method. noises according to the shape of Gaussian function. The closer
from the current point, the bigger the weight accounts. NL-
means filter uses the information of image itself to select pixels
that are similar to the current pixel to do weighted average
within the whole image, and replace the current point’s grey
value with the weighted average result, so as to remove the
noise effectively. And our method selects pixels that are
similar to the current pixel to do weighted average within a
fixed small window but not range in the whole image. And the
weight is not only relates to image grey value, but also uses the
variance of its edge structure. So it can both remove image
noises and maintain well texture structure.
In general, the experimental results show our method is
superior to Gaussian filter and NL-means filter, it neither like
Gaussian filter which is easy to fuzzy edge, nor as the NL-
Fig. 7. The two groups contrast experiments
means filter which can’t eliminate noise completely in order to
keep good structure. Our method can make a good compromise
Relative to the best parameters combination, we get another between get rid of noises and reserve details. And it can get a
idea. We only choose the set of parameters which are good denoising effect when dealing with the noises that are
correspond to the largest PSNR, and get result as the second obey to normal distribution, but the result is not that good
row of figure 7 shows. The corresponding PSNR are 28.5621, when facing with other types of noises, so we need to use
24.2761 and 28.0408. different denoising methods for different noise images as to get
a better denoising effect.
In figure 8, we can clearly see that the first group is much
fuzzier than the second group, and does not have a good effect ACKNOWLEDGMENT
in detail. So the final parameter combination is:
i0(sigma)=0.005, gf(sigma)=1.8, pd(sigma)=8, ksize=11. This work was supported in part by funds from China
National Natural Science Fund (Project No.60875009). And
finally, I would like to express my sincere thanks to my
supervisor, who has given me so much useful advices on my
paper writing, without his illuminating suggestions and careful
revisions the present paper should not have been attempted.

REFERENCES


 [6] Liang jun. Research on Particle Filter Algorithm and Its Application.
[1] Wang ying, Guang-yu Zeng. Research on Image De-noising Algorithm Harbin Institute of Technology.2009
[J].Computer and Information Technology.2011 (04) [7] Qin-lan Xie. Adaptive Gaussian smoothing filter for image denoising [J].
[2] S.Rakshit, A.Ghosh, B.Uma Shankar. Fast mean filtering technique. Computer Engineering and Applications. 2009(16)
Pattern Recognition.2007 [8] Ito,K. Gaussian filter for nonlinear filtering problems. In Proceedings of
[3] Xiang-ju Wang. The Algorithm Research of Image Denosing Based on the 39th IEEE Conference on Decision and Control, 1218 - 1223 vol.2.
Median Filter and Wavelet Transform [D].Xi'an University of Science 2000
and Technology.2008 [9] A.Efros and T.Leung. Texture synthesis by non-parametric sampling. In
[4] Rafael C Gonzalez, Richard E Woods. Digital Image Processing. 2002 Proc. Int. Conf. Computer Vision, volume 2, pages 1033-1038, 1999.
[5] A.Buades, B.Coll, and J.M.Morel. A non-local algorithm for image [10] A.Buades, B.Coll,and J.M.Morel. A review of image denoising
denoising. In Proc.of the IEEE Int.Conf.on Computer Vision and Pattern algorithms with a new one. Multiscale Modeling and
Recognition, 2005. Simulation,4(2):490-530,2005.
[11] A Doucet, J F G de Freitas and N J Gordon, An introduction to
sequential Monte Carlo methods, Springer-Verlag: New York .2001