Abstract

One of the fundamental challenges in the field of image processing and computer vision
is image denoising, where the underlying goal is to estimate the original image by suppressing noise from a noise-contaminated version of the image. Image noise may be caused
by different intrinsic (i.e., sensor) and extrinsic (i.e., environment) conditions which are
often not possible to avoid in practical situations. Therefore, image denoising plays an
important role in a wide range of applications such as image restoration, remote sensing
imaging, digital entertainment, visual tracking, image registration, image segmentation,
and image classification, where obtaining the original image content is crucial for strong
performance. While many algorithms have been proposed for the purpose of image denoising, the problem of image noise suppression remains an open challenge, especially in
situations where the images are acquired under poor conditions where the noise level is
very high.

Contents
1 Introduction

1.0.1
1.0.2

Spatial Domain Denoising Methods . . . . . . . . . . . . . . . . . .

Transform Domain Denoising Methods . . . . . . . . . . . . . . . .

2
2

1.0.3

Dictionary Learning Based methods . . . . . . . . . . . . . . . . . .

1.1

Taxonomy of state-of-the-art denoising methods . . . . . . . . . . . . . . .

1.2

Denoising Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2 Literature Survey

2.1

Applications in Image Segmentation . . . . . . . . . . . . . . . . . . . . . .

2.2

Coherent Imaging Systems . . . . . . . . . . . . . . . . . . . . . . . . . . .

3 Spatial Domain Methods and Applications

3.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8
8

3.2

Spatial Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.2.1

Mean Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.2.2
3.2.3

Median Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Weiner Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3.2.4

Lee Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Transform Domain Methods and Applications

14

4.1
4.2

Discrete Wavelet Transform (DWT) . . . . . . . . . . . . . . . . . . . . . . 14

Wavelet Thresholding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4.3

Types of Wavelet Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.3.1

Visu Shrink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.3.2
4.3.3

SURE Shrink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Bayes Shrink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

4.3.4

Oracle Shrink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

4.3.5

Neigh Shrink . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

ii

5 Conclusion
23
5.1 Spatial Domain filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.2

Wavelet Domain filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Bibliography

25

iii

List of Figures
1.1

Taxonomy of state-of-the-art denoising methods. . . . . . . . . . . . . . . .

1.2

Denoising Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

4.1

Hard Thresholding. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.2

Soft Thresholding. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.3
4.4

2-D Wavelet Decomposition. . . . . . . . . . . . . . . . . . . . . . . . . . . 18

Two level decomposition of lena image. . . . . . . . . . . . . . . . . . . . . 18

iv

List of Tables
3.1

3X3 mask . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3.2
3.3

Neighborhood of w(5,5) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Constant weight filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.4

Median values in the neighborhood of 140 . . . . . . . . . . . . . . . . . . 11

Chapter 1
Introduction
Digital images play an important role both in daily life applications such as satellite
television, magnetic resonance imaging, computer tomography as well as in areas of research and technology such as geographical information systems and astronomy. Data sets
collected by image sensors are generally contaminated by noise. Imperfect instruments,
problems with the data acquisition process, and interfering natural phenomena can all
degrade the data of interest. Furthermore, noise can be introduced by transmission errors
and compression. Thus, denoising is often a necessary and the first step to be taken before
the images data is analyzed. It is necessary to apply an efficient denoising technique to
compensate for such data corruption.
Image Denoising is an important research area serving as the actual foundation
for many applications, such as object recognition, digital entertainment, and remote sensing imaging. As the number of image sensors per unit area increases, camera devices
tend to be more sensitive to noise. Denoising techniques have become a critical step for
improving the final visual quality of images [?].
Denoising is the process of reconstructing the original image by removing unwanted noise from a corrupted image. It is designed to suppress the noise, while preserving as many image structures and details as possible. The main challenge is to design
noise reduction filters that provide a compromise between these two. Suppose that we
have such an image formation model as
v(x) = u(x) + n(x)

(1.1)

where x denotes the 2-D spatial coordinates of pixels in an image, u is the true image,
and n indicates the independent additive noise, which we assume is normally distributed
with a standard deviation and a zero mean. In real camera systems, the noise may come
1

Image Denoising Algorithms and Applications

2014-15

from different sources such as photon noise, thermal noise, and quantization noise. Thermal noise is present in the analog circuitry and follows a zero mean Gaussian distribution.
Generally, image denoising approaches can be categorized as spatial domain, transform
domain, and dictionary learning based according to the image representation. Spatial domain methods include local and nonlocal filters, which exploit the similarities between
either pixels or patches in an image. Both transform domain and dictionary learning
based methods consider transforming images into other domains, in which similarities of
transformed coefficients are employed. The difference between them is that transform
domain approaches usually apply fixed basis functions to represent images, but learningbased methods use sparse representations on a redundant dictionary.

1.0.1

Spatial Domain Denoising Methods

Spatial domain filters [?]-[?] exploit spatial correlations in images. The spatial filters are
classified into two categories: local and nonlocal filters. A filter is local if the candidate
selection process used for filtering is restricted by the spatial distance. A filter is nonlocal
if the candidate selection depends only on the similarity and is not restricted by the spatial
distance.

1.0.2

Transform Domain Denoising Methods

Transform domain methods have been researched in the context of image denoising for
decades [?]. Although there are a large number of variations in this category, such as
discrete cosine transform (DCT), wavelets, wedgelets, curvelets, bandlets, contourlets,
and steerable wavelets, wavelets based methods are still dominant.

1.0.3

Dictionary Learning Based methods

Since sparse modeling was proposed by Olshausen and Field, training an overcomplete
dictionary for the patch representation has been extensively explored in many research
fields . In the past decade, it has been successfully applied to denoising problems.

1.1

Taxonomy of state-of-the-art denoising methods

Vladimir et al. classified denoising filters according to local/nonlocal and pointwise/multipoint.

A novel taxonomy of the denoising methods has been proposed in [?], as shown in Fig.1.
M.Tech, Signal Processing, SIT, Tumakuru

Image Denoising Algorithms and Applications

2014-15

Learning-based approaches have been introduced, which the progress is made in the past
several years in denoising. Furthermore, the comprehensive analysis, comparative evaluation of prevailing classical methods, and recent promising techniques will serve as a good
reference and provide insights for future research in denoising. The reasons for choosing
a method are either that it is representative within a category or it is the best version in
the variants of a popular method.

Figure 1.1: Taxonomy of state-of-the-art denoising methods.

1.2

Denoising Concept

The basic idea behind denoising is the estimation of the uncorrupted image from the
distorted or noisy image, and is also referred to as image denoising. There are various methods to help restore an image from noisy distortions. Selecting the appropriate
method plays a major role in getting the desired image. The denoising methods tend to
be problem specific. For example, a method that is used to denoise satellite images may
not be suitable for denoising medical images.
The Linear operation shown in Figure 1.2 is the addition or multiplication of the noise
n(x, y) to the signal s(x, y). Once the corrupted image w(x, y) is obtained, it is subjected
to the denoising technique to get the denoised image z(x, y).
In case of image denoising methods, the characteristics of the degrading system and
the noises are assumed to be known beforehand. The image s(x, y) is blurred by a linear

M.Tech, Signal Processing, SIT, Tumakuru

Image Denoising Algorithms and Applications

2014-15

Figure 1.2: Denoising Concept

operation and noise n(x, y) is added to form the degraded image w(x, y). This is convolved
with the restoration procedure g(x, y) to produce the restored image z(x, y).

M.Tech, Signal Processing, SIT, Tumakuru

Chapter 2
Literature Survey
2.1

Applications in Image Segmentation

To find different applications in image segmentation this paper [?] presented a new
clustering-based segmentation technique. In image processing method clustering algorithm is widely used segmentation method but due to occurrence of noise during image
acquisition, this might affect the processing results. In order to overcome this drawback,
the proposed algorithm called Denoising-based(DB) clustering algorithm presented three
variations namely, Denoising based- K-means (DB-KM), Denoising-based-Fuzzy C-means
(DB-FCM), and Denoising-based-Moving K-means (DBMKM) [?]. This algorithm was
proposed to minimize the salt-and-pepper noise without degrading the fine details of the
images during the segmentation process . These methods incorporated a noise detection
stage to the clustering algorithm, producing an adaptive segmentation technique specifically for segmenting the noisy images [?]. In this paper authors introduced a version
of adaptive clustering based segmentation techniques. The basic concept behind this
technique was to remove impulsive noises i.e., Salt-and- Pepper noise. Salt-and-Pepper
noise contaminations , caused by errors in the image acquisition/recording/or transmission are detectable at the minimum and maximum intensities. It is important to eliminate
Salt-and-Pepper noise contained in the image because its occurrence severely damage the
information or data embedded in original image. Conventionally, in case of occurrence
of Salt-and-Pepper noise in images we have to apply a pre-processing task like filtering
before segmentation. To tackle this problem, they proposed an adaptive clustering-based
segmentation technique by incorporating the noise detection stage for segmenting noisy
images. The inherited noise detection behavior improved the segmentation results by
only selecting noise-free pixels for the process of segmentation. The result of this proposed algorithm was better when compared with conventional algorithms because of the

Image Denoising Algorithms and Applications

2014-15

inclusion of the noise detection stage in its process. Simulation results showed that the
proposed algorithms were able to remove low density of salt-and-pepper noise (i.e., up
to 50 percentage) during the segmentation process. This finding is proven by smaller
values of F(I), F(I), Q(I), and E(I) produced by the proposed DB-clustering algorithm
[?]. This finding concluded the DB-clustering as a good technique for segmentation of
noisy images, which could be used as pre or post-processing technique in the consumers
electronics fields.

2.2

Coherent Imaging Systems

In study of coherent imaging systems for e.g. ultrasound and laser imaging, multiplicative noise (also known as speckle noise)models are central. With respect to the standard
Gaussian additive noise scenario these models have two additional layer of difficulties:1)
the noise is multiplied by (rather than added to) the original image; 2) the noise is not
Gaussian, with Rayleigh and Gamma being commonly used densities [?].In this paper
author performed a set of experiments and presented that their proposed method named
MIDAL (multiplicative image denoising by augmented Lagrangian), yields state-of-theart results in terms of speed as well as in denoising performance. In this paper [?], they
addressed the (unconstrained) convex optimization problem results from the look multiplicative model formulated with logarithm of the reflectance. An optimization algorithm
is proposed with these respective building blocks:
The original unconstrained optimization problem is first transformed into an equivalent constrained problem, via available splitting procedure that is described in [?];
This constrained problem is then addressed in [?] using an augmented Lagrangian
method.
Author of this paper [?] proposed a recursive filter, called the Cluster-based Adaptive
Fuzzy Switching (CAFSM)for removing impulse noise from digital images. This filter is
composed of a cascaded easy to-implement impulse detector and a noise filter (for detail
preserving restoration). During digital images acquisition process images are commonly
get contaminated with impulse noise. Hence, they focused on developing a robust filter
that provide denoising for any type of impulse noise models. The CAFSM filter operates
at a wide range of impulse noise densities without affecting image fine details and textures.
They developed a fast and automated algorithm. In their simulation results they showed
that the CAFSM filter outperforms other state-of-the art impulse noise filters in terms
of subjective and objective qualities in the filtered images when applied recursively and
iteratively and shows excellent restoration results in denoising color images [?]. The
M.Tech, Signal Processing, SIT, Tumakuru

Image Denoising Algorithms and Applications

2014-15

advantage of this proposed filter is its capability in handling realistic impulse noise model
for real world application and regarded as a universal impulse noise filter. Fuzzy reasoning
is embedded as part of its filtering mechanism. Extensive simulation results verified its
excellent impulse detection and detail preservation abilities by attaining the highest PSNR
and lowest MAE values across a wide range of noise densities shown in [?].

M.Tech, Signal Processing, SIT, Tumakuru

Chapter 3
Spatial Domain Methods and
Applications
3.1

Background

Filters play a significant role in the image denoising process. It is a technique for modifying
or enhancing an image. The basic concept behind reducing noise in noisy images using
linear filters is digital convolution and moving window principle. Linear filtering is filtering
in which the value of an output denoised pixel is a linear combination of the values of the
pixels in the input pixels neighborhood. Let be the input signal subjected to filtering, and
be the filtered output. If the applied filter satisfies certain conditions such as linearity and
shift invariance, then the output filter can be expressed mathematically in simple form as
given below
Z
z(x) =

w(t)h(x t)dt

(3.1)

Where impulse response or point is spread function and it completely characterizes

the filter. The above process called as convolution and it can be expressed as z = w h .
In case of discrete convolution the filter is as given below
z(i) =

w(t)h(i t)

(3.2)

This means that the output z(i) at point i is given by a weighted sum of input pixels
surrounding i and here the weights are given by h(t) . To create the output at the next
pixel i + 1 , the function is shifted by one and the weighted sum is computed again . The
overall output is created by a series of shift-multiply-sum operations, and this forms a

Image Denoising Algorithms and Applications

2014-15

discrete convolution. For the 2-dimensional case h(t) becomes h(t, u), and above Equation
becomes
z(i) =

w(t)h(i t)

(3.3)

Here the values of h(t, u) are referred to as the filter weights, the filter kernel, or filter
mask. For reasons of symmetry h(t, u) is always chosen to be of size m. where m and n
are both usually odd (often m=n). In physical systems, always the kernel h(t, u) must
be non-negative, which results in some blurring or averaging of the image. The narrower
the h(t, u) , then the filter gives less blurring. In digital image processing, h(t, u) maybe
defined arbitrarily and this h(t, u) gives rise to many types of filters.

3.2
3.2.1

Spatial Filters
Mean Filter

A mean filter acts on an image by smoothing it. i.e., it reduces the variation in terms
of intensity between adjacent pixels. The mean filter is a simple moving window spatial
filter, which replaces the center value in the window with the average of all the neighboring pixel values including that center value. It is implemented with a convolution mask,
which provides a result that is a weighted sum of the values of a pixel and its neighbor
pixels. It is also called a linear filter. The mask or kernel is a square. Often a 3 3
square kernel is used. If the sum of coefficients of the mask equal to one, then the average brightness of the image is not changed. If the sum of the coefficients equal to zero,
then mean filter returns a dark image. This average filter works on the shift-multiply-sum
principle . This principle in the two-dimensional image can be represented as shown below,
Let us consider a 512 512 image and 3 3 mask and let the filter mask is
h1

h2

h3

h4

h5

h6

h7

h8

h9

Table 3.1: 3X3 mask

And the neighborhood of pixel (5,5) is given in Table 3.2
Then the output filter value at pixel (5,5) is given as
h1 w44 + h2 w45 + h3 w46 + h4 w54 + h5 w55 + h6 w56 + h7 w64 + h8 w65 + h9 w66
In the above filter if all the weights are same then it is called constant weight filter. and
M.Tech, Signal Processing, SIT, Tumakuru

Image Denoising Algorithms and Applications

2014-15

w44

w45

w46

w54

w55

w56

w64

w65

w66

Table 3.2: Neighborhood of w(5,5)

If the sum of coefficients of the mask equal to one, then the average brightness of the
image is not changed. If the sum of the coefficients equal to zero, the average brightness
is lost, and it returns a dark image. For example, here the sum of coefficients equal to one.

1
9

1
9

1
9

1
9

1
9

1
9

1
9

1
9

1
9

Table 3.3: Constant weight filter

Computing the straightforward convolution of an image with the above mask carries
out the mean filtering process. This mean filter used as a low pass filter, and it does not
allow the high frequency components present in the noise. It is to be noted that larger
kernels of size 5 5 or 7 7 produces more denoising but make the image more blurred.
A tradeoff is to be made between the kernel size and the amount of denoising.

3.2.2

Median Filter

A median filter comes under the class of nonlinear filter. It also follows the moving window
principle, like mean filter. A 3 3, 5 5, or 7 7 kernel of pixels is moved over the
entire image. First the median of the pixel values in the window is computed, and then
the center pixel of the window is replaced with the computed median value. Calculation
of Median is done as first sorting all the pixel values from the surrounding neighborhood
(either ascending or descending order) and then replacing the pixel being considered with
the middle pixel value.
The below process illustrates the methodology of median filtering Let us take 3 3 mask
and the pixel values of image in the neighborhood of considered noisy pixel are
Let us consider pixel at (3, 3) i.e., pixel value of 100.Neighborhood of this pixel are
115,150,108,132,107,152,128,134.
After sorting these pixels (in ascending order) we will get
107,108,115,128,132,134,140,150,152.
M.Tech, Signal Processing, SIT, Tumakuru

10

Image Denoising Algorithms and Applications

2014-15

127

147

175

111

150

120

115

150

108

118

122

132

140

107

112

112

152

128

134

112

134

155

155

198

145

Table 3.4: Median values in the neighborhood of 140

And the median value among this is 132(5th value). So, now this pixel magnitude 140 will
replace with the value of 132 unrepresentative of the surrounding pixels. The median is
more robust compared to the mean. Since one of the neighbor value or considered pixel
used as median , this filter does not create new pixel values when the filter straddles an
edge. It shows that median filter preserves sharp edges than the mean filter.

3.2.3

Weiner Filter

The Wiener filter is a spatial-domain filter and it generally used for suppression of additive noise. Norbert Wiener proposed the concept of Wiener filtering in the year 1942 .
There are two methods: (i) Fourier-transform method (frequency-domain) and (ii) meansquared method (spatial-domain) for implementing Wiener filter. The Fourier method
is used only for denoising and deblurring. whereas the later is used for denoising. In
Fourier transform method of Wiener filtering requires a priori knowledge of the noise
power spectra and the original image. But in latter method no such a priori knowledge
is required. Hence, it is easier to use the mean-squared method for development. Wiener
filter is based on the least-squared principle, i.e. this filter minimizes the mean-squared
error (MSE) between the actual output and the desired output.
Image statistics vary too much from a region to another even within the same image.
Thus, both global statistics (mean, variance, etc. of the whole image) and local statistics
(mean, variance, etc. of a small region or sub-image) are important. Wiener filtering is
based on both the global statistics and local statistics and is given
fb(x, y) = g +

f2
(g(x, y) g)
f2 + 2

(3.4)

Where fb(x, y) denotes the restored image, g is the local mean, f2 is the local variance
and 2 is the noise variance.
Let us consider (2m + 1)(2n+1)windowthenlocalmeanisdef inedas
M.Tech, Signal Processing, SIT, Tumakuru

11

Image Denoising Algorithms and Applications

2014-15

s=m t=n
1 X X
g=
g(s, t)
L s=m t=n

(3.5)

where L is the total number of pixels in a window.

Similarly, consider (2m + 1)(2n+1)windowthenlocalvarianceisdef inedas
g2

s=m
t=n
X X
1
(g(s, t) g)2
=
L 1 s=m t=n

(3.6)

The local signal variance f2 , is used in (3.4) is calculated from g2 with a priori knowledge of noise variance, 2 simply by subtracting 2 from g2 with the assumption that the
signal and noise are not correlated with each other.
From (3.4) it may be observed that the filter-output is equal to local mean, if the current
pixel value equals local mean. Otherwise, it outputs a different value. the value being
somewhat different from local mean. If the input current value is more (less) than the
local mean, then the filter outputs a positive (negative) differential amount taking the
noise variance and the signal variance into consideration. Thus, the filter output varies
from the local mean depending upon the local variance and hence tries to catch the true
original value as far as possible. In statistical theory, Wiener filtering is a great land
mark. It estimates the original data with minimum mean-squared error and hence, the
overall noise power in the filtered output is minimal. Thus, it is accepted as a benchmark
in 1 D and 2 D signal processing.

3.2.4

Lee Filter

The Lee filter, developed by Jong-Sen Lee, is an adaptive filter which changes its characteristics according to the local statistics in the neighborhood of the current pixel. The
Lee filter is able to smooth away noise in flat regions, but leaves the fine details (such as
lines and textures) unchanged. It uses small window (3 3, 5 5, 7 7). Within each
window, the local mean and variances are estimated.
The output of Lee filter at the center pixel of location (x, y) is expressed as:
fb(x, y) = k(x, y)[g(x, y) g] + g

(3.7)

Where
k(x, y) = 1

n2
g2

(3.8)

The parameter k(x, y) ranges between 0 (for flat regions) and 1 (for regions with high
signal activity). The distinct characteristic of the filter is that in the areas of low signal

M.Tech, Signal Processing, SIT, Tumakuru

12

Image Denoising Algorithms and Applications

2014-15

activity (flat regions) the estimated pixel approaches the local mean, whereas in the areas
of high signal activity (edge areas) the estimated pixel favours the corrupted image pixel,
thus retaining the edge information. It is generally claimed that human vision is more
sensitive to noise in a flat area than in an edge area. The major drawback of the filter
is that it leaves noise in the vicinity of edges and lines. However, it is still desirable to
reduce noise in the edge area without sacrificing the edge sharpness. Some variants of Lee
filter available in the literature handle multiplicative noise and yield edge sharpening.

M.Tech, Signal Processing, SIT, Tumakuru

13

Chapter 4
Transform Domain Methods and
Applications
4.1

Discrete Wavelet Transform (DWT)

Wavelets are simply mathematical functions and these functions analyze data according
to scale or resolution. They aid in studying a signal at different resolutions or in different windows . The wavelet transform was borne out of a need for further developments
from Fourier transforms. Wavelets transform signals in the time domain (rather, assumed
to be in the time domain) to a joint time-frequency domain. The main weakness that
was found in Fourier transforms was their lack of localized support, which made them
susceptible to Heisenbergs Uncertainty principle. In short, this means that we could
get information about the frequencies present in a signal, but not where and when the
frequencies occurred. Wavelets, on the other hand, are not anywhere as subject to it.
Wavelets provide some more advantages over Fourier transforms. For example, they do
a better job in approximating signals with sharp spikes or signals having discontinuities
whereas Fourier transform does not give efficient results. Mainly, Wavelets can be used
in image compression, turbulence, radar, human vision, earthquake prediction, etc.
A wavelet is, as the name might suggest, a little piece of a wave. The finite scale
multi resolution representation of a discrete function can be known as a discrete wavelet
transform (DWT).It is a fast linear operation on a data vector, whose length is an integer
power of 2. Discrete wavelet transform is invertible and orthogonal, where the inverse
transform expressed as a matrix is the transpose of the transform matrix. The wavelet
basis or wave let function is quite localized in space. But individual wavelet functions are

14

Image Denoising Algorithms and Applications

2014-15

localized in frequency similar to sines and cosines in Fourier transform. The orthonormal
basis or wavelet basis is defined as
(j,k) (x) = z j/2 (z j x k)

(4.1)

And the scaling function is given as

(j,k) (x) = z j/2 (z j x k)

(4.2)

Where is wavelet function and j and k are integers that scale and dilate the wavelet
basis or function. The factor j in the above equations is known as the scale index and it
indicates the wavelets width. The factor k provides the position. The wavelet function
is dilated by powers of two and it is translated by the integer k. In terms of the wavelet
coefficients, the wavelet equation is

(x) =

N
1
X

gk

p
2(2x k)

(4.3)

Here are high pass wavelet coefficients. Scaling equation in terms of the scaling coefficients is given as shown below,

(x) =

N
1
X

gk

p
2(2x k)

(4.4)

Where the function is scaling function and the coefficients are low pass coefficients.
The wavelet and scaling coefficients are related by quadrature mirror relationship as given
below
gn = (1)n h1n+N

(4.5)

Where N is the number of vanishing moments. The wavelet equation produces different types of wavelet families like Daubechies, Haar, symlets, coiflets, etc. Wavelets are
classified into a family by the number of vanishing moments N . Within each family of
wavelets there are wavelet subclasses distinguished by the number of coefficients and by
the level of iterations.

4.2

Wavelet Thresholding

Donoho and Johnstone have done the lot of work on filtering of additive Gaussian noise
using wavelet thresholding. Wavelets play a major role in image compression and image
denoising. These Wavelet coefficients calculated by a wavelet transform represent change
in the time series at a particular resolution. By considering the time series at various
M.Tech, Signal Processing, SIT, Tumakuru

15

Image Denoising Algorithms and Applications

2014-15

resolutions, it is then possible to filter out the noise. After applying wavelet transform
small coefficients are dominated by noise, while coefficients with a large absolute value
carry more signal information than noise. Replacing the smallest, noisy coefficients by
zero and a backwards wavelet transform on the result may lead to a reconstruction with
the essential signal characteristics and with less noise. For thresholding three observations
and assumptions:
1. The decorrelating property of a wavelet transform creates a sparse signal: most
untouched coefficients are zero or close to zero.
2. Noise is spread out equally over all coefficients.
3. The noise level is not too high, so that we can recognize the signal and the signal
wavelet coefficients.
So, choosing of threshold level is important task.the coefficients having magnitude
greater than threshold are considered as signal of interest and keep the same or modified
according to type of threshold selected and other coefficients become zero. The image is
reconstructed from the modified coefficients. This process is also known as the inverse
discrete wavelet transform (IDWT).
Selection of threshold is an important point of interest. Care should be taken so as
to preserve the edges of the denoised image. There exist various methods for wavelet
thresholding, which rely on the choice of a threshold value. Some typically used methods
for denoising image are Visu Shrink, Sure Shrink, Bayes Shrink, Neigh shrink, oracle
Shrink, Smooth Shrink and Fuzzy based Shrink. Prior to the discussion of these above
methods, it is important to know about the two general categories of thresholding. They
are hard- thresholding and soft-thresholding types. The hard-thresholding can be defined
as
TH = xf or|x| > t

(4.6)

Here t is threshold value.plot for this is as shown below

In This, all coefficients whose magnitude is greater than the selected threshold value t
remain same and the others whose magnitude is smaller than t are set to zero. It creates a
region around zero where the coefficients are considered negligible. In Soft thresholding ,
The coefficients whose magnitude is greater than the selected threshold value are become
shrinks towards zero and others set to zero . The Soft-thresholding can be defined as
TS = sgn(x)(|x| t)f or|x| > t

(4.7)

Plot as shown below,

M.Tech, Signal Processing, SIT, Tumakuru

16

Image Denoising Algorithms and Applications

2014-15

Figure 4.1: Hard Thresholding.

Figure 4.2: Soft Thresholding.

In practice, it can be seen that the soft method is much better and yields more visually
pleasant images. This is because the hard method is discontinuous and yields abrupt
artifacts in the recovered images. Also, the soft method yields a smaller minimum mean
squared error compared to hard form of thresholding.
Now let us focus on the all methods of thresholding mentioned earlier. For all these
methods the image is first subjected to a discrete wavelet transform, which decomposes
the image into various sub-bands.
Graphically wavelet decomposition is shown as below,
The sub-bands HHk,HLk,LHk,k=1,2,..J are called the details, where K is the scale and
J denotes the largest or coarsest scale in decomposition. Note, LLk is the low resolution
component. Thresholding is now applied to the detail components of these sub bands to
remove the unwanted coefficients, which contribute to noise. And as a final step in the

M.Tech, Signal Processing, SIT, Tumakuru

17

Image Denoising Algorithms and Applications

2014-15

Figure 4.3: 2-D Wavelet Decomposition.

denoising algorithm, the inverse discrete wavelet transform is applied to build back the
modified image from its coefficients.

Figure 4.4: Two level decomposition of lena image.

4.3
4.3.1

Types of Wavelet Denoising

Visu Shrink

VisuShrink was introduced by Donoho[?]. The threshold value t in this type is derived
from the standard deviation of the noise. It uses hard thresholding rule. It is also called
M.Tech, Signal Processing, SIT, Tumakuru

18

Image Denoising Algorithms and Applications

2014-15

as universal threshold and is defined

t=

2logn

(4.8)

2 is the noise variance present in the signal and n represents the signal size or number
of samples. An estimate of the noise level was defined based on the median absolute
deviation given by
median({|gjk,k | : k = 0, 1, . . . 2j1 1})
0.6745
is refers to the detail coefficients in the wavelet transform.

b=

Where gj1,k

(4.9)

The main drawback of VisuShrink is it does not deal with minimizing the mean squared
error. However, VisuShrink gives the images that are overly smoothed. This is because
VisuShrink removes too many coefficients. Another disadvantage is that it cannot remove speckle noise, which is multiplicative noise. It can only deal with an additive noise.
VisuShrink follows the global thresholding scheme; here global threshold means a single
value of threshold applied globally to all the wavelet coefficients.

4.3.2

SURE Shrink

A threshold chooser based on Steins Unbiased Risk Estimator (SURE) was proposed by
Donoho and Johnstone [?] and is called as SureShrink. It is determined from the both
universal threshold and the SURE threshold. It is subband dependent threshold. A
threshold value for each resolution level in the wavelet transform which is referred to as
level dependent thresholding . The main advantage of SureShrink is to minimize the mean
squared error, unlike Visu Shrink, defined as
n
1 X
(z(x, y) s(x, y))2
M SE = 2
n x,y=1

(4.10)

Where z(x, y) is the estimate of the signal while s(x, y) is the original signal without
noise and n is the size of the signal. SureShrink suppresses noise by thresholding the
empirical wavelet coefficients. The SureShrink threshold ts is defined as
ts = min(t,

p
2logn)

(4.11)

Where t denotes the value that minimizes Steins Unbiased Risk Estimator, is the
noise variance computed , and n is the size of the image. It follows the soft thresholding
rule. The thresholding employed here is adaptive, i.e., a threshold level is assigned to
each resolution level by the principle of minimizing the Steins Unbiased Risk Estimator

M.Tech, Signal Processing, SIT, Tumakuru

19

Image Denoising Algorithms and Applications

2014-15

for threshold estimates. It is smoothness adaptive i.e., if the unknown function contains
abrupt changes or boundaries in the image, the reconstructed image also does.

4.3.3

Bayes Shrink

BayesShrink was proposed by Chang, Yu and Vetterli . The goal of this method is to
minimize the Bayesian risk, and hence its name, BayesShrink. It uses soft thresholding
and it is also subband-dependent, like Sure Shrink, which means that threshold level is
selected at each band of resolution in the wavelet decomposition. The Bayes threshold,
tb , is defined as
tb =

2
s

(4.12)

Where 2 is the noise variance and s2 is the signal variance without noise. The noise
variance 2 is estimated from the sub band HH1 in the decomposition of wavelet by the
median estimator. From the definition of additive noise we have
w(x, y) = s(x, y) + (x, y)

(4.13)

Since the signal and noise are independent of each other it can be stated that
w2 = s2 + 2

(4.14)

w2 can be calculated as shown below

w2 =

n
1 X 2
w (x, y)
n2 x,y=1

(4.15)

The variance of the signal, 2 is computed as shown below

s =

max(w 2 2 , 0)

(4.16)

With these 2 and w2 the Bayes threshold is computed from the below equation
2
s
the wavelet coefficients are thresholded at each band.
tb =

M.Tech, Signal Processing, SIT, Tumakuru

(4.17)

20

Image Denoising Algorithms and Applications

4.3.4

2014-15

Oracle Shrink

Oracle Shrink is wavelet thresholding method Used in image denoising. This method is
implemented with the assumption that the wavelet coefficients of original decomposed
image are known. The OracleShrink uses the threshold denoted as TO S Mathematically
they are represented by:
TOS =

.
argTmin
h

n
X

(T h (Yij ) Fij )

(4.18)

i,j1

Here, {Fij } are wavelet coefficients of the original decomposed image. T h (.) is the
soft thresholding function and T h (.) is hard thresholding function
T h (x) = sgn(x).max(|x| T, 0)

(4.19)

T h (x) = x.1 {|x| > T }

(4.20)

and

equation 4.20 keeps the input if it is larger than the threshold T ; otherwise, it is set
to zero.

4.3.5

Neigh Shrink

This wavelet-domain image thresholding scheme was proposed by Chen et al. and it
incorporating neighboring coefficients, namely Neigh Shrink. It thresholds the wavelet
coefficients according to the magnitude of the squared sum of all the wavelet coefficients,
i.e., the local energy, within the neighborhood window. The neighborhood window size
may be , etc. The shrinkage function for Neigh Shrink of any arbitrary window centered
at (i, j) is expressed as:
ij = [1

TU2
]
Sij2

(4.21)

Where TU is the universal threshold Sij2 and is the squared sum of all wavelet coefficients in the given window

Sij =

j+1
i+1
X
X

2
Ym,n

(4.22)

n=j1 m=i1

here very important consideration is + sign at the end of the formula it means keep the
positive values while setting it to zero when it is negative. The estimated center wavelet
coefficient Fij is then calculated from its noisy counterpart of Yij as:

M.Tech, Signal Processing, SIT, Tumakuru

21

Image Denoising Algorithms and Applications

Fij = ij .Yij

M.Tech, Signal Processing, SIT, Tumakuru

2014-15

(4.23)

22

Chapter 5
Conclusion
5.1

Spatial Domain filters

Noise can be removed using Linear spatial domain filters as well transform domain filters.
Linear techniques possess mathematical simplicity but have the disadvantage that they
introduce blurring effect. To reduce this blurring effect we can use non-linear filters like
median filter etc. Here the considered are AWGN and Speckle Noise. Speckle Noise uses
the advantage of logarithmic transform. Among the spatial domain filters Lee filter is
good edge preserving filter. From chapter 3, it is cleared that wiener and Lee filter having
3x3 window size are giving efficient results under low noise variance. Anisotropic diffusion
is also a powerful filter where local image variation is measured at every point, and pixel
values are averaged from neighborhoods whose size and shape depend on local variation,
it uses partial differential equations .this is iterative filter. More iteration may leads
to instability where, in addition to edges noise becomes prominent. mean and median
3x3 filter giving good results under low noise variance conditions and for medium noise
variance conditions mean and median 5x5 window sized filters are giving efficient results.

5.2

Wavelet Domain filters

Image denoising, using wavelet techniques are effective because of its ability to capture
the energy of signal in a few high transform values, when natural image is corrupted
by Gaussian noise. wavelet thresholding, an idea that noise is removed by killing coefficient relative to some threshold. Out of various thresholding techniques soft-thresholding
proposed by Donoho and Johnstone is most popular. The use of universal threshold to
denoise images in wavelet domain is known as VisuShrink , In addition, sub band adaptive systems have superior performance, such as SureShrink BayesShrink, NeighShrink.
23

Image Denoising Algorithms and Applications

2014-15

As Speckle noise is inherent property of ultrasound images. From the above simulated
values ,neigh shrink yields good performance under low variances of noise. And in case
of high noise variance , neigh shrink , Bayes shrink and fuzzy based wavelet denoising
technique give the good filtering performance. As execution time is another important
image metric it is observed that Bayes shrink and neigh shrink are taking more time(in
seconds) than fuzzy based wavelet filter.

M.Tech, Signal Processing, SIT, Tumakuru

24

Bibliography
[1] J. H. Hong, S. B. Cho, and U. K. Cho, A novel evolutionary approach to image
enhancement filter design: Method and applications, IEEE Trans. Syst., Man, Cybern.
B, Cybern., vol. 39, no. 6, pp. 14461457, Dec. 2009.
[2] L. Shapiro and G. Stockman, Computer Vision. Englewood Cliffs, NJ, USA: PrenticeHall, 2001.
[3] N. Wiener, Extrapolation, Interpolation, and Smoothing of Stationary Time Series.
New York, NY, USA: Wiley, 1949.
[4] B. Widrow and S. Haykin, Least-Mean-Square Adaptive Filters. New York, NY, USA:
Wiley-IEEE, 2003.
[5] Ling Shao, Ruomei Yan, Xuelong, and Yan Liu, From Heuristic Optimization to
Dictionary Learning: A Review and Comprehensive Comparison of Image Denoising
Algorithms, IEEE TRANSACTIONS ON CYBERNETICS, VOL. 44, NO. 7,pp.10011013, JULY 2014.
[6] Sit i Noraini Sulaiman, Nor Ashidi Mat Isa,Denoising-based Clustering Algorithms for
Segmentat ion of Low Level Salt-and-Pepper Noise-Corrupt ed Images,IEEE Transact
ions on Consumer Elect ronics, Vol. 56, No. 4, pp. 2702-2710, November 2010.
[7] Jos M. Bioucas-Dias, Mrio A. T. Figueiredo, Mult iplicat ive Noise Removal Using
Variable Splitt ing and Const rained Opt imizat ion, IEEE Transact ions on Image
Processing, vol. 19, No. 7, pp. 1720-1730, July 2010.
[8] Kenny Kal Vin Toh, Nor Ashidi Mat Isa, Cluster-Based Adapt ive Fuzzy Swit ching
Median Filt er for Universal Impulse Noise Reduct ion, IEEE Transact ions on Consumer Elect ronics, Vol. 56, No. 4, pp. 2560- 2568, November 2010.
[9] D. Donoho, Wedgelets: Nearly-minimax estimation of edges, Ann. Stat., vol. 27, no.
3, pp. 859897, Jun. 1999.

25

Image Denoising Algorithms and Applications

2014-15

[10] E. Candes and D. L. Donoho, Recovering edges in ill-posed inverse problems: Optimality of curvelet frames, Ann. Stat., vol. 30, no. 3, pp. 784842, Jun. 2002.

M.Tech, Signal Processing, SIT, Tumakuru

26