You are on page 1of 78

CHAPTER-1

Introduction
1 | P a g e
1. Introduction
Noise suppression in medical images is a particularly delicate and difficult task. A
trade-off between noise reduction and the preservation of actual image features has to be
made in a way that enhances the diagnostically relevant image content. Image processing
specialists usually lack the biomedical expertise to judge the diagnostic relevance of the De-
noising results. or example! in ultrasound images! speckle noise may contain information
useful to medical experts. Also biomedical images show extreme variability and it is
necessary to operate on a case by case basis . "his motivates the construction of robust and
versatile De-noising methods that are applicable to various circumstances! rather than being
optimal under very specific conditions.
ourier and related representations are classical methods that have been widely used
in image processing applications. "he noise removal has been done using the #iener filter!
which is derived by assuming a signal model of uncorrelated $aussian distributed
coefficients in the ourier domain and utili%es second-order statistics of the ourier
coefficients. &tatistical modelling of image features at multiple resolution scales is a topic of
tremendous interest for numerous disciplines including image restoration! image analysis and
segmentation! data fusion! etc.
In this thesis! we propose one robust method that adapts itself to various types of
image noise as well as to the preference of the medical expert' a single parameter can be used
to balance the preservation of relevant details against the degree of noise reduction. (ee
proposed digital image enhancement and noise filtering by use of local statistics. )ecently
*i%urica and co authors proposed a low-complexity joint detection and estimation method. In
particular! the method applies the minimum mean s+uared error criterion assuming that each
wavelet coefficient represents a ,signal of interest- with a probability leading to the
2 | P a g e
generali%ed likelihood ratio formulation in the wavelet domain . "hey proposed an analytical
model for the probability of signal presence! which is adapted to the global coefficient
histogram and to a local indicator of spatial activity .e.g.! the locally averaged magnitude of
the wavelet coefficients/.
In this thesis! we propose a related! but more flexible method! which is applicable to
various and unknown types of image noise. #e employ a preliminary detection of the
wavelet coefficients that represent the features of interest in order to empirically estimate the
conditional pdf0s of the coefficients given the useful features and given background noise. At
the same time! the preliminary coefficient classification is also exploited to empirically
estimate the corresponding conditional pdf0s of the local spatial activity indicator .(&AI/.
"he preliminary classification step in the proposed method relies on the persistence of useful
wavelet coefficients across the scales ! and is related to the one in ! but avoids its iterative
procedure.
"he classification step of the proposed method involves an adjustable parameter that
is related to the notion of the expert-defined ,relevant image features-. In certain applications
the optimal value of this parameter can be selected as the one that maximi%es the signal-to-
noise ratio .&N)/ and the algorithm can operate as fully automatic. 1owever! we believe that
in most medical applications the tuning of this parameter leading to gradual noise suppression
may be advantageous. "he proposed algorithm is simple to implement and fast. #e
demonstrate its usefulness for denoising and enhancement of the 2"! 3ltrasound and
4agnetic )esonance images.

3 | P a g e
Thesis organization
"his thesis is organi%ed as follows. 2hapter-5 gives the brief introduction of
dissertation ! chapter-6 discuss Noise models!chapter-7 discuss Image denoising techni+ues !
chapter-8 discuss the #avelet-9ayesian :stimators!chapter-; briefs the versatile denoising
algorithm!chapter-< deals with the implementation! results and discussion of proposed
algorithm to the 3& and 4)I images and their performance evaluated.chapter-= gives the
conclusion and future scope.
4 | P a g e
CHAPTER-2
Noise Models
5 | P a g e
2. Noise Models
2.1 Introduction
4edical images are generally of low-contrast and they often get a complex type of
noise. Noise from the instruments is further enhanced due to application of different types of
+uanti%ation! reconstruction and enhancement algorithms. or example! >-ray images will
generally get *oisson noise whereas 22D detectors include *oisson as well as $aussian
noise. Noise suppression in such images demands a careful and flexible analysis because the
diagnostic information is contained in the variations from the usual pattern which ordinarily
coexist with noise. "he majority of research activities have been devoted to modeling noise
from these sources. Although valuable information can be obtained from these models as far
as global knowledge is concerned but assumption for noise to be $aussian! &peckle!
Impulsive or any special model for distribution of noise will not lead to good results in real
life applications.
"here can be different types of noise ac+uired in any imaging system? which can be
either signal-dependent or signal-independent. In medical images! generally! the noise is
correlated. &o they have higher visibility than uncorrelated noise of the same variance. Also!
vision is less sensitive to noise in the vicinity of object boundaries! and thus some signal
dependent noise may confuse vision by producing image-like artefacts and hiding image
details.
*ractically! all types of noise can be represented as some combination of additive!
multiplicative and impulsive noise. &o any general type of noise can be modelled as
combination of $aussian additive noise! &peckle noise and &alt and *epper noise.
6 | P a g e
2.2 Gaussian Noise
$aussian noise is evenly distributed over the signal. "his means that each pixel in the
noisy image is the sum of the true pixel value and a random $aussian distributed noise value.
As the name indicates! this type of noise has a $aussian distribution! which has a bell shaped
probability distribution function given by!

6 6
. / @ 6
6
5
. /
6
z
P z e

.6.5/
where
z
represents the gray level!

is the mean or average of the function! and

is the
standard deviation of the noise .)efer ig 6.5.a//.
2. !alt and Pe""er Noise
&alt and pepper noise is an impulse type of noise! which is also referred to as
intensity spikes. "he salt and pepper noise is generally caused by malfunctioning of pixel
elements in the camera sensors! faulty memory locations! or timing errors in the digiti%ation
process. "he probability density function for this type of noise is shown in ig 6.5.b/.

! for
. / ! for
A!otherwise
a
b
P z a
p z P z b

'



.6.6/
#ig 2.1 The "ro$a$ilit% &unctions &or 'a( Gaussian '$( I)"ulse* and 'c( Ga))a
distri$utions
7 | P a g e

2.+ !"ec,le Noise
3ltrasound imaging plays crucial role in medical imaging due to its non-invasive
nature! low cost and capability of forming real time imaging. 4edical 3ltrasound imaging is
done using ultrasonic .aka ultrasound/ waves in 7 to 6A 41% range. 3ltrasound waves are
produced from the transducer and travel through body tissues and when the wave reaches an
object or surface with different texture or acoustic nature! it is reflected back. "hese echoes
are received by the apparatus .the transducer array/ and changed into electric current. "hese
signals are amplified and conditioned and shown on a display device in real time. "he image
generated using 3ltrasound imaging - commonly known as 3ltrasound &canning - is called
an Ultra-sonogram. "he resolution of the image will be better by using higher fre+uencies
but this at the same time limits the depth of the penetration.
1owever by nature 3ltrasound image contains more noise content! especially speckle
noise than any other imaging modality. Noise is introduced at all stages of image ac+uisition.
"here could be noises due to the loss of proper contact or air gap between the transducer
probe and body? there could be noise introduced during the beam forming process and also
during the signal processing stage. :ven during the scan conversion! there could be loss of
information due to the interpolation.
&peckle noise is a multiplicative noise. "his type of noise occurs in almost all
coherent imaging systems such as laser! medical! acoustics and &A) .&yn-thetic Aperture
)adar/ imagery. "he source of this noise is attributed to random inter-ference between the
coherent returns. ully developed speckle noise has the charact-eristic of multiplicative noise.
8 | P a g e
&peckle noise follows a gamma distribution and is given as

! for
A! for
5
A
. /
. 5/B
A
az
b b
a z
e z
p z
b
z

'

<
.6.7/

"he gamma distribution is given below in ig. 6.5.c/.
2.- Rician Noise
4agnetic )esonance Imaging .4)I/ is a non-invasive imaging techni+ue used
extensively for clinical diagnosis and medical research. 4) images! however! contain
varying amounts of noise of diverse origins! including noise from stochastic variation!
numerous physiological processes! eddy currents! artefacts from the differing magnetic field
susceptibilities of neighboring tissues! rigid body motion! and many others. &ome noise
components! including bulk motion from cardiac pulsation and head or body movement!
generate unusual observations! or statistical CoutliersC! that differ substantially from most 4)
data that do not contain those noise sources .at least! not to the same degree/. *revious studies
have shown that those noise components can introduce substantial bias into measurements
and estimations made from those images! such as indices for the principle direction of fiber
tracts in diffusion tensor images. Identifying and reducing these noise components in 4)
images is essential to improving the validity and accuracy of studies designed to map the
structure and function of the human body.
"he noise in 4agnetic )esonance magnitude images obeys a )ician distribution.
3nlike additive $aussian Noise! )ician noise is signal dependent and conse+uently
separating signal from noise is difficult task. )ician noise is especially problematic in low
9 | P a g e
signal to noise ratio regions! where it is only causes random fluctuations! but also introduces
a signal dependent bias to the data that reduces the image contrast.
"he most common method for 4) images is to sample the object of interest in the
fre+uency domain of the raw data ! is denoted by!
)e Im
. ! / . ! / . ! / Y u v Y u v iY u v +
.6.8/
where u! v are the hori%ontal and vertical fre+uencies!
5 i
and
)e
Y
and
Im
Y
are
the real and imaginary components of the data. "he measurements can be decomposed into a
signal and noise component
. ! / . ! / . ! / Y u v S u v N u v +
.6.;/
where
. ! / S u v
is the signal of interest and
. ! / N u v
is a complex white $aussian noise.
"he noise is primarily due to thermal noises in the patient.
"he most common reconstruction techni+ue in 4)I is to compute the inverse D" of
the raw data Y! introduces phase errors and involves complicated algorithms. #e focus here
on the more common magnitude image reconstruction.
(et
y
is the inverse D" of Y. "he magnitude image at a location
. ! / m n
!
D ! E F D ! E F x m n y m n
is real valued and can be displayed for visual inspection or used for
automated computer analysis. &ince the magnitude reconstruction is simply the s+uare root of
the sum of the two independent $aussian )andom variables! the noise in the magnitude
image
x
is described by a )ician distribution. #e can regard
x
as an observation of the
signal G
s
0 with additive noise

!

D ! E D ! E D ! E x m n s m n m n +
.6.;/
10 | P a g e
"he additive noise is not a simple white noise in this model. )ather
D ! E D ! E D ! E m n x m n s m n
and the structure of this noise depends on
D ! E s m n
. 1ence we say
that the )ician noise is a signal dependent noise. In low intensity .dark/ regions of the
magnitude! the noise distribution tends to the )ayleigh distribution. In high intensity .bright/
regions! the noise tends a $aussiasn distribution.
2.. Cha"ter !u))ar%
In this chapter! we have discussed various noises models such as gaussian noise!
rician noise! speckle noise and salt and pepper noise. out which speckle noise present in
ultrasound image and )ician noise present in 4) image! we have addressed the probability
function of these noises and their occurrences possibilities briefly.
11 | P a g e
CHAPTER-
I)age denoising techni/ues
12 | P a g e
.I)age denoising techni/ues
.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 ac+uisition process! and interfering natural phenomena can all degrade the data of
interest. urthermore! noise can be introduced by transmission errors and compression. "hus!
denoising is often a necessary and the first step to be taken before the images data is
analy%ed. It is necessary to apply an efficient denoising techni+ue to compensate for such
data corruption. Image denoising still remains a challenge for researchers because noise
removal introduces artifacts and causes blurring of the images. "his chapter describes
different methodologies for noise reduction .or denoising/ giving an insight as to which
algorithm should be used to find the most reliable estimate of the original image data given
its degraded version. Noise modeling in images is greatly affected by capturing instruments!
data transmission media! image +uanti%ation and discrete sources of radiation. Different
algorithms are used depending on the noise model. 4ost of the natural images are assumed to
have additive random noise which is modeled as a $aussian. &peckle noise is observed in
ultrasound images whereas )ician noise affects 4)I images. "he scope of the paper is to
focus on noise removal techni+ues for natural images.
.2 E0olution o& I)age 1enoising Research
Image Denoising has remained a fundamental problem in the field of image
processing. #avelets give a superior performance in image denoising due to properties such
as sparsity and multiresolution structure. #ith #avelet "ransform gaining popularity in the
13 | P a g e
last two decades various algorithms for denoising in wavelet domain were introduced. "he
focus was shifted from the &patial and ourier domain to the #avelet transform domain.
:ver since Donoho0s #avelet based thresholding approach was published in 5HH;! there was
a surge in the denoising papers being published. Although Donoho0s concept was not
revolutionary! his methods did not re+uire tracking or correlation of the wavelet maxima and
minima across the different scales as proposed by 4allat . "hus! there was a renewed interest
in wavelet based denoising techni+ues since Donoho demonstrated a simple approach to a
difficult problem. )esearchers published different ways to compute the parameters for the
thresholding of wavelet coefficients. Data adaptive thresholds were introduced to achieve
optimum value of threshold. (ater efforts found that substantial improvements in perceptual
+uality could be obtained by translation invariant methods based on thresholding of an
3ndecimated #avelet "ransform. "hese thresholding techni+ues were applied to the non
orthogonal wavelet coefficients to reduce artifacts. 4ultiwavelets were also used to achieve
similar results. *robabilistic models using the statistical properties of the wavelet coefficient
seemed to outperform the thresholding techni+ues and gained ground. )ecently! much effort
has been devoted to 9ayesian denoising in #avelet domain. 1idden 4arkov 4odels and
$aussian &cale 4ixtures have also become popular and more research continues to be
published. "ree &tructures ordering the wavelet coefficients based on their magnitude! scale
and spatial location have been researched. Data adaptive transforms such as Independent
2omponent Analysis .I2A/ have been explored for sparse shrinkage. "he trend continues to
focus on using different statistical models to model the statistical properties of the wavelet
coefficients and its neighbors. uture trend will be towards finding more accurate
probabilistic models for the distribution of non-orthogonal wavelet coefficients.

14 | P a g e
Classi&ication o& 1enoising Algorith)s
"here are two basic approaches to image denoising! spatial filtering methods and
transform domain filtering methods.
. !"atial #iltering
A traditional way to remove noise from image data is to employ spatial filters. &patial
filters can be further classified into non-linear and linear filters.
..1 Non-2inear #ilters
#ith non-linear filters! the noise is removed without any attempts to explicitly
identify it. &patial filters employ a low pass filtering on groups of pixels with the assumption
that the noise occupies the higher region of fre+uency spectrum. $enerally spatial filters
remove noise to a reasonable extent but at the cost of blurring images which in turn makes
the edges in pictures invisible. In recent years! a variety of nonlinear median type filters such
as weighted median! rank conditioned rank selection! and relaxed median have been
developed to overcome this drawback.
..2 2inear #ilters
A mean filter is the optimal linear filter for $aussian noise in the sense of mean
s+uare error. (inear filters too tend to blur sharp edges! destroy lines and other fine image
details! and perform poorly in the presence of signal-dependent noise. "he wiener filtering
method re+uires the information about the spectra of the noise and the original signal and it
works well only if the underlying signal is smooth. #iener method implements spatial
smoothing and its model complexity control correspond to choosing the window si%e. "o
overcome the weakness of the #iener filtering! Donoho and Iohnstone proposed the wavelet
based denoising scheme in.
15 | P a g e
.+ Trans&or) 1o)ain #iltering
"he transform domain filtering methods can be subdivided according to the choice of
the basis functions. "he basis functions can be further classified as data adaptive and non-
adaptive. Non-adaptive transforms are discussed first since they are more popular.
.+.1 !"atial-#re/uenc% #iltering
&patial-fre+uency filtering refers use of low pass filters using ast ourier "ransform
."/. In fre+uency smoothing methods the removal of the noise is achieved by designing a
fre+uency domain filter and adapting a cut-off fre+uency when the noise components are
decorrelated from the useful signal in the fre+uency domain. "hese methods are time
consuming and depend on the cut-off fre+uency and the filter function behavior. urthermore!
they may produce artificial fre+uencies in the processed image.
.- 3a0elet do)ain
iltering operations in the wavelet domain can be subdivided into linear and nonlinear
methods.
.-.1 2inear #ilters
(inear filters such as #iener filter in the wavelet domain yield optimal results when
the signal corruption can be modeled as a $aussian process and the accuracy criterion is the
mean s+uare error .4&:/. 1owever! designing a filter based on this assumption fre+uently
results in a filtered image that is more visually displeasing than the original noisy signal! even
though the filtering operation successfully reduces the 4&:. In a wavelet-domain spatially
adaptive I) #iener filtering for image denoising is proposed where wiener filtering is
performed only within each scale and intrascale filtering is not allowed.
.-.2 Non-2inear Threshold #iltering
"he most investigated domain in denoising using #avelet "ransform is the non-linear
coefficient thresholding based methods. "he procedure exploits sparsity property of the
16 | P a g e
wavelet transform and the fact that the #avelet "ransform maps white noise in the signal
domain to white noise in the transform domain. "hus! while signal energy becomes more
concentrated into fewer coefficients in the transform domain! noise energy does not. It is this
important principle that enables the separation of signal from noise.
"he procedure in which small coefficients are removed while others are left
untouched is called 1ard "hresholding. 9ut the method generates spurious blips! better
known as artifacts! in the images as a result of unsuccessful attempts of removing moderately
large noise coefficients. "o overcome the demerits of hard thresholding! wavelet transform
using soft thresholding was also introduced in. In this scheme! coefficients above the
threshold are shrunk by the absolute value of the threshold itself. &imilar to soft thresholding!
other techni+ues of applying thresholds are semi-soft thresholding and $arrote thresholding .
4ost of the wavelet shrinkage literature is based on methods for choosing the optimal
threshold which can be adaptive or non-adaptive to the image.
'i( Non-Ada"ti0e thresholds
JI&3&hrink is non-adaptive universal threshold!which depends only on number of
data points. It has asymptotic e+uivalence suggesting best performance in terms of 4&:
when the number of pixels reaches infinity. JI&3&hrink is known to yield overly smoothed
images because its threshold choice can be unwarrantedly large due to its dependence on the
number of pixels in the image.
'ii( Ada"ti0e Thresholds
&3):&hrink uses a hybrid of the universal threshold and the &3): D&tein0s
3nbiased )isk:stimatorE threshold and performs better than JI&3&hrink. 9ayes&hrink
minimi%es the 9ayes0 )isk :stimator function assuming $enerali%ed $aussian prior and thus
yielding data adaptive threshold. 9ayes&hrink outperforms &3):&hrink most of the times.
2ross Jalidation replaces wavelet coefficient with the weighted average of neighborhood
17 | P a g e
coefficients to minimi%e generali%ed cross validation .$2J/ function providing optimum
threshold for every coefficient.
"he assumption that one can distinguish noise from the signal solely based on
coefficient magnitudes is violated when noise levels are higher than signal magnitudes.
3nder this high noise circumstance! the spatial configuration of neighboring wavelet
coefficients can play an important role in noise-signal classifications. &ignals tend to form
meaningful features .e.g. straight lines! curves/! while noisy coefficients often scatter
randomly.
.-. Non-orthogonal 3a0elet Trans&or)s
3ndecimated #avelet "ransform .3D#"/ has also been used for decomposing the
signal to provide visually better solution. &ince 3D#" is shift invariant it avoids visual
artifacts such as pseudo-$ibbs phenomenon. "hough the improvement in results is much
higher! use of 3D#" adds a large overhead of computations thus making it less feasible. In
normal hard@soft thresholding was extended to &hift Invariant Discrete #avelet "ransform. In
&hift Invariant #avelet *acket Decomposition .&I#*D/ is exploited to obtain number of
basis functions. "hen using 4inimum Description (ength principle the 9est 9asis unction
was found out which yielded smallest code length re+uired for description of the given data.
"hen! thresholding was applied to denoise the data. In addition to 3D#"! use of
4ultiwavelets is explored which further enhances the performance but further increases the
computation complexity. "he 4ultiwavelets are obtained by applying more than one mother
function .scaling function/ to given dataset. 4ultiwavelets possess properties such as short
support! symmetry! and the most importantly higher order of vanishing moments. "his
combination of shift invariance K 4ultiwavelets is implemented in which give superior
results for the (ena image in context of 4&:.
18 | P a g e
.-.+ 3a0elet Coe&&icient Model
"his approach focuses on exploiting the multiresolution properties of #avelet
"ransform. "his techni+ue identifies close correlation of signal at different resolutions by
observing the signal across multiple resolutions. "his method produces excellent output but is
computationally much more complex and expensive. "he modeling of the wavelet
coefficients can either be deterministic or statistical.
'i( 1eter)inistic
"he Deterministic method of modeling involves creating tree structure of wavelet
coefficients with every level in the tree representing each scale of transformation and nodes
representing the wavelet coefficients. "his approach is adopted in . "he optimal tree
approximation displays a hierarchical interpretation of wavelet decomposition. #avelet
coefficients of singularities have large wavelet coefficients that persist along the branches of
tree. "hus if a wavelet coefficient has strong presence at particular node then in case of it
being signal! its presence should be more pronounced at its parent nodes. If it is noisy
coefficient! for instance spurious blip! then such consistent presence will be missing. (u et
al.! tracked wavelet local maxima in scale space! by using a tree structure. Lther denoising
method based on wavelet coefficient trees is proposed by Donoho.
'ii( !tatistical Modeling o& 3a0elet Coe&&icients
"his approach focuses on some more interesting and appealing properties of the
#avelet "ransform such as multiscale correlation between the wavelet coefficients! local
correlation between neighborhood coefficients etc. "his approach has an inherent goal of
perfecting the exact modeling of image data with use of #avelet "ransform. "he following
two techni+ues exploit the statistical properties of the wavelet coefficients based on a
probabilistic model.
19 | P a g e
'a( Marginal Pro$a$ilistic Model
A number of researchers have developed homogeneous local probability models for
images in the wavelet domain. &pecifically! the marginal distributions of wavelet coefficients
are highly kurtotic! and usually have a marked peak at %ero and heavy tails. "he $aussian
mixture model .$44/ and the generali%ed $aussian distribution .$$D/ are commonly used
to model the wavelet coefficients distribution. Although $$D is more accurate! $44 is
simpler to use. proposed a methodology in which the wavelet coefficients are assumed to be
conditionally independent %ero-mean $aussian random variables! with variances modeled as
identically distributed! highly correlated random variables. An approximate 4aximum A
*osteriori .4A*/ *robability rule is used to estimate marginal prior distribution of wavelet
coefficient variances. All these methods mentioned above re+uire a noise estimate! which
may be difficult to obtain in practical applications. &imoncelli and Adelson used a two
parameter generali%ed (aplacian distribution for the wavelet coefficients of the image! which
is estimated from the noisy observations. 2hang et al. proposed the use of adaptive wavelet
thresholding for image denoising! by modeling the wavelet coefficients as a generali%ed
$aussian random variable! whose parameters are estimated locally .i.e.! within a given
neighborhood/.
'$( 4oint Pro$a$ilistic Model
1idden 4arkov 4odels .144/ models are efficient in capturing inter-scale
dependencies! whereas )andom 4arkov ield models are more efficient to capture intrascale
correlations. "he complexity of local structures is not well described by )andom 4arkov
$aussian densities whereas 1idden 4arkov 4odels can be used to capture higher order
statistics. "he correlation between coefficients at same scale but residing in a close
neighborhood are modeled by 1idden 4arkov 2hain 4odel where as the correlation between
coefficients across the chain is modeled by 1idden 4arkov "rees. Lnce the correlation is
20 | P a g e
captured by 144! :xpectation 4aximi%ation is used to estimate the re+uired parameters and
from those! denoised signal is estimated from noisy observation using well known 4A*
estimator. A model is described in which each neighborhood of wavelet coefficients is
described as a $aussian scale mixture .$&4/ which is a product of a $aussian random
vector! and an independent hidden random scalar multiplier. &trela etal described the joint
densities of clusters of wavelet coefficients as a $aussian scale mixture! and developed a
maximum likelihood solution for estimating relevant wavelet coefficients from the noisy
observations. Another approach that uses a 4arkov random field model for wavelet
coefficients was proposed by Iansen and 9ulthel . A disadvantage of 14" is the
computational burden of the training stage. In order to overcome this computational problem!
a simplified 14"! named as u14"! was proposed.
'iii( 1ata-Ada"ti0e Trans&or)s
)ecently a new method called Independent 2omponent Analysis .I2A/ has gained
wide spread attention. "he I2A method was successfully implemented in denoising Non-
$aussian data. Lne exceptional merit of using I2A is it0s assumption of signal to be Non-
$aussian which helps to denoise images with Non-$aussian as well as $aussian distribution.
Drawbacks of I2A based methods as compared to wavelet based methods are the
computational cost because it uses a sliding window and it re+uires sample of noise free data
or at least two image frames of the same scene. In some applications! it might be difficult to
obtain the noise free training data.
.. Cha"ter !u))ar%
*erformance of denoising algorithms is measured using +uantitative performance
measures such as peak signal-to-noise ratio .*&N)/! signal-to-noise ratio .&N)/ as well as in
terms of visual +uality of the images. 4any of the current techni+ues assume the noise model
to be $aussian. In reality! this assumption may not always hold true due to the varied nature
21 | P a g e
and sources of noise. An ideal denoising procedure re+uires a priori knowledge of the noise!
whereas a practical procedure may not have the re+uired information about the variance of
the noise or the noise model. "hus! most of the algorithms assume known variance of the
noise and the noise model to compare the performance with different algorithms. $aussian
Noise with different variance values is added in the natural images to test the performance of
the algorithm. Not all researchers use high value of variance to test the performance of the
algorithm when the noise is comparable to the signal strength. 3se of " in filtering has
been restricted due to its limitations in providing sparse representation of data. #avelet
"ransform is the best suited for performance because of its properties like sparsity!
multiresolution and multiscale nature. In addition to performance!issues of computational
complexity must also be considered. "hresholding techni+ues used with the Discrete #avelet
"ransform are the simplest to implement. Non-orthogonal wavelets such as 3D#" and
4ultiwavelets improve the performance at the expense of a large overhead in their
computation. 144 based methods seem to be promising but are complex. #hen using
#avelet "ransform! Nason emphasi%ed that issue such as choice of primary resolution .the
scale level at which to begin thresholding/ and choice of analy%ing wavelet also have a large
influence on the success of the shrinkage procedure. #hen comparing algorithms! it is very
important that researchers do not omit these comparison details. &everal papers did not
specify the wavelet used neither the level of decomposition of the wavelet transform was
mentioned. It is expected that the future research will focus on building robust statistical
models of non-orthogonal wavelet coefficients based on their intra scale and inter scale
correlations. &uch models can be effectively used for image denoising and compression.
22 | P a g e
CHAPTER-+
3a0elet 5a%esian Esti)ators
23 | P a g e
+.3a0elet 5a%esian Esti)ators
+.1 Introduction to $a%esian anal%sis
"he 9ayesian approach to statistical inference begins at the same place as more
conventional approaches' a probability model .also known as a data generating process/. As
opposed to the point estimators .means! variances/ used by classical statistics! 9ayesian
statistics is concerned with generating the posterior distribution of the unknown parameters
given both the data and some prior density for these parameters. As such! 9ayesian statistics
provides a much more complete picture of the uncertainty in the estimation of the unknown
parameters! especially after the confounding effects of nuisance parameters are removed.
A probability model relates observed data y to a set of unknown parameters M! with
the possible inclusion of fixed! known covariates x. Lur data are usually a collection of
observations indexed i N 5! . . . ! n' y N Oy
5
! y
6
! . . . ! y
n
P. "he observed data need not be
scalars? in fact! they can be anything supported by the probability model! including vectors
and matrices. "he probability model has k parameters! which are represented M N .M
5
! M
6
! . . . !
M
k
/. "he covariates!or independent variables! are typically a collection of column vectors' x N
Ox
5
! x
6
! . . . ! x
n
P."he probability model can be written p.yFM! x/! or! suppressing the
conditioning on the covariates! p.yFM/. It is important to stress the importance of choosing
appropriate probability models. #hile canonical models exist for certain types of dependent
variables! the choice of model is rarely innocuous! and is thus something that should be tested
for ade+uacy. "he linear regression model is! perhaps! the most commonly used probability
model in political science. Lur dependent variable y N Oy
5
! y
6
! . . . ! y
n
P is a collection of
scalars with domain y
i
). Lur independent variables can be represented as a collection of
column vectors x N Ox
5
! x
6
! . . . ! x
n
P! each of dimension .Q R 5/.
24 | P a g e
"he purpose of statistical inference is to learn about parameters that characteri%e the
data generating process given observed data. In the conventional! fre+uentist approach to
statistical inference! one assumes that the parameters are fixed! unknown +uantities! and that
the observed data y are a single reali%ation of a repeatable process! and can thus be treated as
random variables. "he goal of the fre+uentist approach is to produce estimates of these
unknown parameters. "hese estimates are denoted
S

. "he most common way to obtain these


estimates is by the method of maximum likelihood. "his method uses the same probability
model! buttreats it as a function of the fixed! unknown parameters computed using 9ayes0
"heorem
+.2 5a%e6s Theore)
"he foundation of 9ayesian statistics is 5a%es7 theore). &uppose we observe a
random variable y and wish to make inferences about another random variable M! where M is
drawn from some distribution p.M/. rom the definition of conditional probability!
. /
. /
.
!
F
/
r
r
r
P y
P y
P y


.8.5/
Again from the definition of conditional probability! we can express the joint
probability by conditioning on to give
*r.y!M/ N *r.yFM/ *r.M/
*utting these together gives 9ayes0 theorem
. / . /
. /
.
F
/
F
r r
r
r
P y P
P y
P y


.8.6/
#ith n possible outcomes .M
5
.......M
n
/!
25 | P a g e
5
. / . /
. /
. /
. /
F
F
F
. / F / .
j j
j
r r r j
r i
r n
i
r i
r
P y P
P y
P y
P y
P P b y


.8.7/
*
r
.M/ is the prior distribution of the possible M values! while *
r
.MFy/ is the posterior distribution
of M given the observed data y.
+. Ma8i)u)-2i,elihood 'M2( Esti)ation
An important class of estimators is the maximum-lieli!ood .4(/ estimators. "he 4(
parameter estimate is the one that makes the set of mutually-independent observations x NOx
5
!
x
6
! . . . ! x
n
P .which is an instance of the random sample O>
5
! >
6
! . . . ! >
n
P/ most likely to
occur. "he random sample comprises mutually independent )Js! thereby making the joint
*D e+uivalent to the product of the marginal *Ds. "his defines the lieli!ood function for
the parameter M as
. F / . F / " # P #

5 5
6 6
. ! ! ... F /
n n
p # x # x # x
N
5
. F /
i
n
x i
i
P x


.8.8/

"he 4( parameter estimate is
T
arg max . F / " #


+.+ Ma8i)u)-a-Posteriori 'MAP( Esti)ation
&ometimes we have a priori information about the physical process whose parameters
we want to estimate. &uch information can come either from the correct scientific knowledge
of the physical process or from previous empirical evidence. #e can encode such prior
26 | P a g e
information in terms of a *D on the parameter to be estimated. :ssentially! we treat the
parameter M as the value of an )J. "he associated probabilities *.M/ are called the prior
probabilities. #e refer to the inference based on such priors as $ayesian inference. 9ayesC
theorem shows the way for incorporating prior information in the estimation process

. / . /
. /
.
F
/
F # P P
P #
P y



.8.;/
"he term on the left hand side of the e+uation is called the posterior. Ln the right
hand side! the numerator is the product of the likelihood term and the prior term. "he
denominator serves as a normali%ation term so that the posterior *D integrates to unity.
"hus! 9ayesian inference produces the maximum a posteriori .4A*/ estimate
arg max . F / arg max . F / . / P # P # P



.8.</
+.- Mean !/uare Esti)ation
&uppose >
5
! >
6
! . . . ! >
n
represent a se+uence of random variables about whom one
set of observations are available! and U represents an unknown random variable. "he
problem is to obtain a good estimate for U in terms of the observations >
5
! >
6
! . . . ! >
n
.


S
. / . / # Y Y Y #
.8.=/
(et
5 6
! ! . . . . . / S ! /
n
# # # y # represent such an estimate for U. Note that
V../ can be a linear or a nonlinear function of the observation >
5
! >
6
! . . . ! >
n
.2learly
represents the error in the above estimate! and F

F
6
the s+uare of the error. &ince

is a
random variable! :OF

F
6
P represents the mean s+uare error. Lne strategy to obtain a good
estimator would be to minimi%e the mean s+uare error by varying over all possible forms of
27 | P a g e
and this procedure gives rise to the 4inimi%ation of the 4ean &+uare :rror .44&:/ criterion
for estimation. "hus under 44&: criterion! the estimator V../ is chosen such that the mean
s+uare error :OF

F
6
P is at its minimum.
+.. Cost &unctions
"he 9ayesian measure to minimi%e 9mse N
{ }
6
S
. / % w.r.t *.x!M/
&o! 9mse is :xpected value of s+uare of error! Now define a +udratic cost function
2.

/N

6
N
6
S
. /
.8.W/
then! we have that! 9mseN :O2.

/P
we have an general bayesian criteria is to define a 2ost function 2.

/!define bayes
risk with this cost function
)N :O 2.

/P w.r.t *.x!M/
).
S

/N:O2.M-
S

/P .8.H/
and 4inimi%es the bayes risk w.r.t to
S

.
S

depends on the choice of estimator.


#e have three common cost functions! they are +uadratic! absolute and 1it or miss cost
functions
+...1 /uadratic cost &unction
is defined by 2.

/N

6
and its optimal estimation is given by ).
S

/N:O.M-
S

/
6
P
where
S

N :OMFxPwhich is given by mean of *.MFx/ 2.

#ig. +.1 9uadratic cost &unction


+...2 A$solute cost &unction
is defined by 2.

/NF

F and its optimal estimation is given by ).


S

/N:OF.M-
S

/FP
28 | P a g e
where
S

N :OMFxPwhich is given by 4edian of *.MFx/


2.

/ 2.

#ig. +.2 A$solute cost &unction


#ig. +. Hit or )iss cost &unction
+... Hit or )iss cost &unction
is defined by
A!F F
. /
5!F F
&

<
'

and its optimal estimation is given by


).
S

/N:O.M-
S

/P!where
S

N :OMFxPwhich is given by 4ode of *.MFx/. which is 4aximum A


posteriori or 4A*.
If p.MFx/ is unimodal and symmetric ! then meanN medianN mode
+.: 3a0elet-5a%esian Esti)ators
2onsider an image contaminated by an additive white $aussian noise. &ince the
wavelet transform is orthonormal! the noise is also $aussian and white in the wavelet
domain. "he coefficient of the noisy image in the wavelet domain is expressed as
y ' n +
!
where
'
is the coefficient of the clean image and
n
is the coefficient of the noise which
follows the $aussian distribution.
+.:.1 2inear Esti)ator
If we assume that the coefficients of the subbands of clean images follow the
$aussian distribution! the linear estimator is defined as.
29 | P a g e

6
6 6
S . /
'
' n
' y y

+
where
6 6 6
maxOA! :. /P
' n
' . .8.5A/

+.:.2 Threshold Esti)ator
ield has shown that kurtosis of subbands of natural scenes varies with filter band
width and is maximal at roughly one octave. "he wavelet subband coefficients have highly
non-$aussian statistics. "hus! the linear estimator may not be a good estimator for
coefficients in the wavelet domain. &ince the pdf of coefficients at A ' is significantly
large! we can assume that small amplitude values of the coefficients arose from a value of
A ' . "hen the threshold estimator is defined as

! F F 7
S . /
A!
n
y y
' y
(t!er'ise
>

'

.8.55/
+.:. 5a%esian Marginal 'Coring( Esti)ator
Although the threshold estimator is based on the property that the pdf of coefficients
is significantly large! it does not exploits the whole pdf. "he 9ayesian marginal estimator
uses the whole pdf for estimation! using a pdf model for clean coefficients. "he generali%ed
(aplacian distribution model fits the histogram so well that it produces very small relative
entropy compared to the total entropy . "hen! the optimal estimate of
'
in the least mean
s+uare sense using this model is the conditional mean

. / . /
S . / :. F /
. / . /
n '
n '
'f y ' f ' d'
' y ' y
f y ' f ' d'

.8.56/
30 | P a g e
where
. /
'
f '
is the generali%ed (aplacian distribution.
+.:.+ 5a%esian 4oint Esti)ator
"he 9ayesian joint estimator uses the dependence between coefficients for estimation
where the marginal estimator ignores. "o consider the dependencies between coefficients!
&imoncelli suggested a 4arkov model! in which the density of a coefficient! 0
'
0 is
conditionally $aussian with variance! a linear function of the s+uared coefficients in a
neighborhood
6 6
. F / .A! /

f ' p N ' p b +

v
.8.57/
where the neighborhood O

p
P consists of coefficients at adjacent spatial locations!
other orientations and adjacent scales. "he weight parameters G

'
0 and the constant G b 0 are
estimated by minimi%ing the mean s+uare error

O P
6 6 6
!
6
S
min :

' b

& ' P b
1 _
+
1
, ]

.8.58/
#e can find the 9ayesian joint estimator using the noisy coefficients y! the estimated
parameters

'
and b ! and the neighboring coefficients estimated using the 9ayesian
marginal estimator
6 6
6 6 6
S
. /

' P b
& y y
' P b
+

+ +

.8.5;/
31 | P a g e
+.; Cha"ter !u))ar%
In this chapter! wavelet bayesian estimators! we have discussed basics of bayesian
analysis and bayeCs theorem and its implementation in bayesian estimators. we had brief
overview of various cost functions and some of their properties and they are significant
measure in estimation.we have discussed various wavelet bayesian estimators.

32 | P a g e
CHAPTER--
Pro"osed <ersatile 1enoising Algorith)
33 | P a g e
-.<ersatile 1enoising Algorith)
-.1 Introduction
In this section we will study the theoretical concept behind the proposed
method! and then the new practical algorithm is described.
#e start from a general noise model

y ' n
! where

'
is the unknown
noise-free wavelet coefficient! a point-wise mathematical operation .addition in the case
of additive noise and multiplication in the case of speckle noise/ and

n
an arbitrary noise
contribution. Lur wavelet domain estimation approach relies on the joint detection and
estimation theory and is related to the problem of the spectral amplitude estimation. "he
algorithm is implemented using the +uadratic spline wavelets.
-.2 5asic Theor% =& Pro"osed Algorith)
(et

#
denote a random variable! which takes values

x
from the binary label set
{ } A!5
. "he hypothesis ,the wavelet coefficient

y
represents a signal of interest- is e+uivalent
to the event
5

#
! and the opposite hypothesis is e+uivalent to
A

#
. "he wavelet
coefficients representing the signal of interest in a given sub band are identically distributed
random variables with the probability density function F
. F 5/

Y #
p '
. &imilarly! the
34 | P a g e
coefficients in the same sub band! corresponding to the absence of the signal of interest! are
random variables with the pdf F
. F A/

Y #
p '
.
3nder the model assumptions! the minimum mean s+uared error estimate .the
conditional mean/ of

'
is
S . F ! 5/ . 5F / . F ! A/

' % ' y # P # y % ' y # +

. A F /!

P # y
where
../ %
stands for the expected value. If the signal of interest is surely
absent in a given wavelet coefficient! then )

*+ and %,'

-y

.#

/+0*+. In the case where the


signal of interest is surely present! we approximate %,'

-y

.#

/10*y
!
which accounts for the
fact that vast majority of the coefficient magnitudes representing the signal of interest are
highly above the noise level. Applying 9ayeCs rule! one can express
. 5F /

P # y
as a
generali%ed likelihood ratio! and our estimate becomes
S
5



' y

+
.;.5/
where
( )
( )
F
F
. F 5/
5F
!
. F A/ AF
Y #


Y #

p y
P #
p y P #


P
P
.;.6/
and P symbolically denotes the prior knowledge that is used to estimate the probability of
signal presence.
*i%urica proposed a method to estimate this probability for each wavelet coefficient
from its local surrounding! using a chosen indicator

e

of the local spatial activity. In
particular! since our estimate of the probability of signal presence is a function of

e
! we
write
35 | P a g e
( ) ( ) 5F 5F

P # P # e P
! and replace

in .;.6.5/ by

( )
( )
. F 5/
5F
F
S
. F A/
AF
F

Y #

Y #

p e
P # e
r
p e
P # e

.;.7/
where r is the ratio of unconditional prior probabilities
( )
( )
5
A

P #
r
P #

.;.8/
or a given type of noise! one can derive the complete estimator analytically. In such
approaches where the re+uired conditional densities need to be expressed analytically! the
choice of the local spatial activity indicator is usually restricted to simple forms' even when

e

is defined simply as the locally averaged coefficient magnitude! certain simplifying
assumptions about the statistical properties of the wavelet coefficients are needed in order to
derive
F
. F /

% #
p e x

analytically. "he algorithm that we propose in this paper is applicable
to various noise types! and allows an arbitrary choice of

e
.
-. The <ersatile denoising Techni/ue
"he idea behind the proposed algorithm is to empirically estimate the probabilities
and the probability density functions that specify the estimator. (et N denote the number of
wavelet coefficients in a detail image. or each detail image
{ }
! ..
! 5!
!
2 2 2
j N j j
y y >
! we first
estimate the mas
{ }
! ..
! 5!
S
!
2 2 2
j N j j
x x ?
which indicates the positions of significant wavelet
coefficients .representing the signal of interest/. As usual! we relate the notion of significant
wavelet coefficients to the standard deviation of the noise . Also! we rely on the persistence
36 | P a g e
of significant wavelet coefficients across resolution scales. In particular! we extend our robust
coarse-to-fine classification method from as follows'

6
! !
!
6
! !
S S A! if F FF F . /
S
S S 5! if F FF F . /
2 2 2
j j j
2
j
2 2 2
j j j
y ' 3
x
y ' 3

<

'

.;.;/
where
S
2
j
is an estimate of the noise standard deviation in the detail image
2
j
>
! and
3

is a heuristic! tuneable parameter that controls the notion of the signal of interest. #e
estimate the standard deviation of the input noise S

as the median absolute deviation of the
wavelet coefficients in the 44 sub band at the finest resolution scale! divided by A.<=8; .
In estimating
S
2
j
! we follow
6 6
S S . /
2 2
j j
S
! where for each sub band the constant
2
j
S

is calculated from the filter coefficients of the high pass filter g

and the low pass filter
h

of the discrete wavelet transform! as
( ) ( )
! 6 6 "4 4"
j l
l
S g !

and
( ) ( )
6 6. 5/
6 6
j
44
j l l
S g !


. "o initiali%e the classification ! we start from
S
2 2
5 5
3 > !
where 5 is the coarsest resolution level in the wavelet decomposition.
Now we address the estimation of the wavelet coefficients
2
5
>
using the estimated
mask
S
2
j
8
. "he estimator .;.7.5/ re+uires the conditional densities F
. F /

Y #
p y x
and
F
. F /

% #
p e x
. &ince F
. F /

Y #
p y x
is usually highly symmetrical around A ! in practice
we shall rather estimate the conditional pdf0s F
. F /

6 #
p m x
of the coefficient
magnitudes
F F

m y
. As the local spatial activity indicator

e
! we use the averaged energy
37 | P a g e
of the neighbouring coefficients of

y
where the neighbours are the surrounding coefficients
in a s+uare window at the same scale and the ,parent- .i.e.! the coefficient at the same spatial
position at the first coarser scale/. 1aving the estimated mask 5
S S S x O .. P
N
x x
! let
A
S O '

S x

AP
! and
5
S O ' 5P

S x
. "he empirical estimates F
S . F A/

6 #
p m
and F
S . F A/

% #
p e
are
computed from the histograms of
A
O ' P

m S
and
A
O ' P

e S
respectively .by
normali%ing the area under the histogram/. &imilarly! F
S . F 5/

6 #
p y
and F
S . F 5/

% #
p e
are
computed from the corresponding histograms for
5
S
.
Lur estimation approach still re+uires the probability ratio . )easoning that
. 5/

P #

can be estimated as the fractional number of labels for which
S 5

x
! we estimate the
parameter r

from previous e+uation as
5
5
S
S
S
N

N

x
r
N x

.;.</
"hen the final estimation is defined as
S
S S
S
S
S S 5



r
' y
r

+
.;.=/
where

F
F
S . F 5/
S
S . F A/


6 #

6 #
p m
p m

and
F
F
S . F 5/
S
S . F A/


% #

% #
p e
p e

.;.W/
38 | P a g e
In ig.;.7.5! we show an example of the empirical densities F
S . F /

6 #
p m x
and
F
S . F /

% #
p e x
. "he direct computation of the ratios
S

and
S

from the normali%ed


histograms shown in ig. is not appropriate due to errors in the tails. Lne solution is to first
fit a certain distribution to the histogram. 1ere we apply a simpler approach! observing that
both ( )
S
log

and ( )
S log

can be approximated well by fitting a piece-wise linear curve as


illustrated in ig. ;.7.5. ormally! we approximate
( )
5 5
6 6
S
! 5
S
log
S
! 5


a b m
a b m

+ <

'
+

;
.;.H/
( )
5 5
6 6
S ! 5
S log
S ! 5


c d m
c d m

+ <

'
+

;
.;.5A/
"he fitting in is done as follows. or the index set
F
S O ' . F A/
6 # l
7 l p m



F
S O ' . F A/ A!
6 # l
7 l p m



F
S A! . F 5/ AP
6 # l
p m
we compute the log-ratios
( ) ( ) { } log F F
S S log . F 5/ log . F A/ '
6 # l 6 # l
p m p m l 7


.
urther! let { }
!5 log!
' ! A
l
7 l l 7

<
and { }
!6 log!
' ! A
l
7 l l 7


. #e look for the
polynomial coefficients
5 5
! a b
! which minimi%e
( )
!5
6
5 5 log! l l l 7
a m b

39 | P a g e


#ig -.1 E8a)"les o& the e)"irical "d&7s and &itted log-ratios in the "ro"osed
)ethod* &or the to" le&t ultrasound i)age.
and the polynomial coefficients
6 6
! a b
! which minimi%e
( )
!6
6
6 6 log! l l l 7
a m b


.;.55/

40 | P a g e

i.e.! which are optimal in the least s+uare sense. "he e+uivalent procedure is applied to find
the coefficients
5 5
! c d

and
6 6
! c d

. "his completes the specification of the proposed method.
"he summary of the proposed algorithm is described below and we demonstrate its
applications to 3ltrasound and 4) images in next section
-.+ !u))ar% o& Pro"osed Algorith)
!te" 1. &ompute t!e non decimated 'avelet transform 'it! 5 resolution levels.
!te" 2. 8nitialize !
S
# O ! ! P
2 2
5 5
Y 2 4" "4 44 .
!te" . 9or eac! orientation 2and for eac! scale
6 ! 5!...! 5
j
j 5
. and
.a/ 9or all t!e spatial positions
5!...! N
.
.i/ %stimate
6
! !
!
6
! !
S S A! if F FF F . /
S
S S 5! if F FF F . /
2 2 2
j j j
2
j
2 2 2
j j j
y ' 3
x
y ' 3

<

'

.
.ii/ &ompute t!e local spatial activity indicator

e
using averaged energy of
t!e neig!bouring coefficients of

y
'!ere t!e neig!bours are t!e surrounding coefficients
in a s:uare 'indo'.
.b/ &ompute .

( )
5 5
S S S @
N N
l l
l l
r x N x



.c/ 2efine { }
A
S ' A
l
S l x
and { }
5
S ' 5
l
S l x
.
%stimate
F
S . F A/!
6 #
p m

F
S . F 5/!
6 #
p m

F
S . F A/!
6 #
p e

and
F
S . F 5/
6 #
p e
! from t!e
corresponding !istograms of F F
l l
m W

and
l
e over
A
l S

and over
5
l S .
41 | P a g e
.d/ 9it t!e log-ratios ( )
log

and ( )
log

using ( )
5 5
6 6
S
! 5
S
log
S
! 5


a b m
a b m

+ <

'
+

;
and
( )
5 5
6 6
S ! 5
S log
S ! 5


c d m
c d m

+ <

'
+

;
.e/ 9or
5!...! ! N
estimate
S
S S
S
S
S S 5



r
' y
r

+

!te" +. ;pply t!e inverse 'avelet transform.
-.- Cha"ter !u))ar%
In this chapter! versatile denoising algorithm! we have addressed basic theory of
proposed algorithm and then its theoretical implementation to derive the versatile denoising
algorithm! which implemented to ultrasound images and 4) images in next chapter. 9y
which its performance in terms of &N) ! time lapse and contrast of image can be evaluated
and compared with other filters.

42 | P a g e
CHAPTER-.
I)"le)entation* results and discussions
43 | P a g e
..I)"le)entation* Results and 1iscussion
..1 Introduction
3ltrasound images and 4) images have different medical images with different types
of Noise. #e impliment the versatile denoising algorithm and compare the performance of
the proposed method to one conventional approach in speckle filtering' the homomorphic
#iener filter . In particular! we apply 4atlab0s spatially adaptive #iener filter to the image
logarithm and subse+uently perform the exponential transformation on the filtered out put.
"he window si%e of the #iener filter was experimentally optimi%ed to produce the maximum
output &N) for each test image and for each amount of noise used in the simulations.
..2 A""lication To @ltrasound I)ages
3ltrasound images are corrupted by speckle noise! which affects all coherent imaging
systems. 4ore details about the speckle noise are given in ig. <.5 and <.6 illustrates the
examples of gradual speckle suppression using the proposed method. "he results in this
figure correspond to the window si%e ; > ; and different values of the tuning parameter 3.
44 | P a g e
"he results demonstrate that the increase of 3 leads to a stronger suppression of the
background texture and to the enhancement of sharp intensity variations.
ig <.7! ig <.8 and ig <.; shows the visual comparison for 3ltrasound 2hest!
$allbladder and &ynthetic images. ig <.< and ig <.= enables us to make a visual
comparison of the results of the proposed method and the homomorphic #iener filter on
realistic noisy images. ig <.W shows the hori%ontal line intensity profile comparison for
synthetic image..
)egarding the noise suppression performance! the proposed method shows a stable
behaviour with respect to the tuning parameter 3. ig. <.H demonstrates that for different
noise levels and for different test images the same value of this parameter .3 N 7/ can be
chosen to provide a nearly maximum output &N). It can also be seen that the window si%e
7>7 is optimal under the assumed speckle model.
"able 5 shows the +uantitative comparison of widely used metrics! signal to noise
ratio .&N)/! and peak signal to noise ratio .*&N)/. "he computation time for each algorithm
is also included in this table. #e can notice that our proposed filter exhibits more than 6d9
improvement in both &N) and *&N) over 1omomorphic #iener filter. "he results clearly
demonstrate that the proposed filter outperforms the homomorphic spatially adaptive #iener
filtering both in terms of &N) and *&N) .ig.<.5A/
45 | P a g e
#ig. ..1 Gradual noise su""ression in ultrasound Aidne% i)age* using the
"ro"osed )ethod
46 | P a g e
#ig. ..2 Gradual noise su""ression in ultrasound Chest i)age* using the "ro"osed
)ethod
#ig. .. <isual Co)"arison &or @ltrasound Chest i)age 'a( =riginal I)age '$(
s"ec,led Nois% i)age 'c( Ho)o)or"hic s"atiall% ada"ti0e 3iener &ilter 'd( "ro"osed
)ethod* &or K B and CindoC size -8-
47 | P a g e
#ig. ..+ <isual Co)"arison &or @ltrasound Gall$ladder i)age 'a( =riginal I)age '$(
arti&iciall% s"ec,led i)ages* the results o& the ho)o)or"hic s"atiall% ada"ti0e 3iener
&ilter* and the results o& the "ro"osed )ethod* &or K B and CindoC size -8-.
48 | P a g e
49 | P a g e
#ig...- Results &or s%nthetic Test i)age. 'a( !%nthetic test i)age '$( arti&iciall% s"ec,led
i)ages* 'c( the results o& the ho)o)or"hic s"atiall% ada"ti0e 3iener &ilter* and 'd(the
results o& the "ro"osed )ethod* &or K B : and CindoC size 8
#ig. ... A""lication to the real nois% i)age-1 'a( Real s"ec,le noised @ltrasound i)age1
'$( 3iener &ilter 'c( "ro"osed &ilter.
50 | P a g e
#ig. ..: A""lication to the real nois% i)age-2 'a( Real s"ec,le noised @ltrasound i)age2
'$( 3iener &ilter 'c( "ro"osed &ilter.
#ig. ..; Horizontal "ro&ile co)"arison &or Ds%nth7 i)age
51 | P a g e
#ig. ..E The "er&or)ance o& the "ro"osed )ethod as a &unction o& the 0alue o& the
"ara)eter K. 'a( Test i)age 1 &ro) #ig...2.1* in"ut !NRB1.. d5. '$( Test i)age 2 &ro)
#ig...2.1* in"ut !NRB11.. d5.
#ig. ..1F Co)"arison $etCeen the "ro"osed &ilter and the ho)o)or"hic 3iener &ilter.
'a( The s%nthetic test i)age '$( The Test I)age 1.
Ta$le ..1 !ignal to Noise Co)"arisons &or the three @ltra sound Test I)ages.
52 | P a g e
Reconstruction
Method
Testi)age1*-?-*AB Testi)age2*-?-*AB Testi)age*-?-*AB:
!N
R
'd5(
P!NR
'd5(
Co)".
Ti)e
'!ec(
!NR
'd5(
P!NR
'd5(
Co)".
Ti)e
'!ec(
!NR
'd5(
P!NR
'd5(
Co)".
Ti)e
'!ec(
Nois% i)age 5=.=; 6;.A< - - - 5=.=5 6=.75 - - - 67.WH 76.<A - - -
Ho)o)or"hic3ie
ner #ilter
5=.W< 6;.5< A.A=W 5H.A; 6W.<; A.A=W 6<.H< 7;.<=
A.A5<
Pro"osed #ilter 5H.;5 6<.W5 6.;<6 6A.=7 7A.78 6.756 6H.A6 7=.=6 A.=;
.. A""lication To Magnetic Resonance I)ages
In magnetic resonance imaging the practical limits of the ac+uisition time impose a
trade-off between the signal to noise ratio and the image resolution. "he 4)I image is
commonly reconstructed by computing the inverse discrete ourier transform of the raw data.
4ost commonly! the magnitude of the reconstructed image is used for visual inspection and
for computer analysis. Noise in the 4)I image magnitude is )ician ! having a signal
dependent mean .
It was noted that! due to the signal-dependent mean of the )ician noise! both wavelet
and scaling coefficients of a noisy 4)I image are biased estimates of their noise-free counter
parts. It was shown that one can efficiently overcome this problem by filtering the s:uare of
the 4)I magnitude image in the wavelet domain. In the s+uared magnitude image! data are
non-central chi-s+uare distributed! and the wavelet coefficients are no longer biased estimates
of their noise-free counterparts. "he bias still remains in the scaling coefficients! but is not
signal dependent and it can be removed easily at the resolution scale
6
j
! from each scaling
coefficient
5
6
j
c

+
should be subtracted! where
6
c


is the under lying complex $aussian
noise variance. #e therefore apply our method to the s+uared magnitude of the 4)I image!
subtract the constant bias from the scaling coefficients! and subse+uently compute the s+uare
root of the denoised s+uared magnitude image.
"hree clinical 4) Images! a *elvic 4) Image of si%e <88><6<! a 9rain 4) Image
of si%e 8=5>785! and a &pine 4) Image of si%e 8HA>8W< are used for the experimental
evaluation purpose. In simulations! complex %ero mean white $aussian noise with standard
deviation
7A
c

was added to these images. ig.<.56 to ig <.58 shows the De-noising
result of the proposed method comparison with the spatially adaptive wiener filter. "he
53 | P a g e
results show that our proposed algorithm out performs with that of the spatially adaptive
#iener filtering. "able .6 shows the +uantitative comparison of widely used metrics! signal to
noise ratio .&N)/! and peak signal to noise ratio .*&N)/. "he computation time for each
algorithm is also included in this table. #e can observe that ;.7; d9 improvement in both
&N) and *&N).
"he +uantitative performance of the method! for the 7x7 window si%e is illustrated in
ig.<.5;. or different noise levels the optimal value of 3 is in the range D5.W! 6E! and the
algorithm shows a stable behaviour with respect to 3. In our experiments! on a number of
different reference 4)I images! the improvement over the spatially adaptive #iener filtering
was at least A.; d9 .for relatively clean image/ and more than 7 d9 for low &N) images. In
ig.<.5;.b/! we illustrate such a comparison between the two filters for the reference 9rain
4) image. or the #iener filter! the window si%e was optimi%ed for each noise level! to
produce maximum output &N). "he application of the proposed method to real noisy 4)I
images is demonstrated in ig.<.55. "he 4)I images were provided by the $overnment
$eneral hospital! Jijayawada. "he noise suppression in these images is expected to facilitate
further automatic processing! like e.g.! segmentation.
54 | P a g e
55 | P a g e
#ig. ..11 A""lication to real nois% MRI i)ages. 'a( and 'c( original i)ages* '$( and 'd(
the corres"onding results o& the "ro"osed )ethod &or K B 2 and a 8 CindoC size.
#ig ...12 <isual co)"arison &or a Pel0ic MR I)age* 3indoC ?* AB2 and
7A
'a(
=riginal I)age '$( Nois% I)age 'c( !"atiall% Ada"ti0e 3iener &ilter 'd( Pro"osed &ilter
Ta$le ..2 !ignal to Noise Ratio co)"arison &or three i)ages. Pel0ic MR I)age* 5rain
MR I)age and !"ine MR I)age* CindoC size ?* A B2.
56 | P a g e
Reconstruction
Method
Pel0ic MR I)age!ize
.++?.2. GBF
5rain MR I)age!ize
+:1?+1 GBF
!"line MR I)age!ize
+EF?+;. GBF
!NR
'd5(
P!NR
'd5(
Co)".
Ti)e
'!ec(
!NR
'd5(
P!NR
'd5(
Co)".
Ti)e
'!ec(
!NR
'd5(
P!NR
'd5(
Co)".
Ti)e
'!ec(
Nois% i)age 55.<< 5W.;H - - - AW.76 5=.<W - - - 57.AA 5W.W< - - -
!"atiall%
ada"ti0e
3iener #ilter
5<.67 67.5< A.76; 5A.HH 6A.7; A.5W= 5H.56 68.HW
A.6H=
Pro"osed
#ilter
6A.77 6=.6< 55.788 5<.78 6;.=A ;.<=6 5H.8H 6;.7; H.W57
#ig ..1 <isual co)"arison &or a 5rain MR I)age o& size +:1?+1* 3indoC ?* AB2
and
7A
. 'a( =riginal I)age '$( Nois% I)age 'c( !"atiall% Ada"ti0e 3iener &ilter 'd(
Pro"osed &ilter.
57 | P a g e
#ig. ..1+ <isual co)"arison &or a !"ine MR I)age o& size +EF?+;.* 3indoC ?* AB2
and
7A
. 'a( =riginal I)age '$( Nois% I)age 'c( !"atiall% Ada"ti0e 3iener &ilter 'd(
Pro"osed &ilter.
#ig. ..1- 9uantitati0e "er&or)ance o& the "ro"osed )ethod. 'a( In&luence o& the
"ara)eter K &or di&&erent 0alues
6
c
o& the underl%ing co)"le8 Gaussian noise. '$(
Noise su""ression "er&or)ance '&or K B 2( in co)"arison to the s"atiall% ada"ti0e
3iener &iltering. The original MRI i)age &ro) #ig.12 is used in the si)ulations.
58 | P a g e
Ta$le .. I)"ortant &eatures <ersatile denoising algorith)s &or -12 ? -12 Pel0ic i)age
Cith
6A
.
..+ Cha"ter !u))ar%
In this chapter! we applied versatile denoising algorithm to 3ltra sound images and
4) images! and compared its performance with homomorphic filter! although it takes high
computation time it shows gain in the *&N) ! it had ;.57 d9 and W.<= d9 improvement !
when ;56x;56 pelvic image with 6A
.it adapts itself to various noise types particularly
speckle and rician noise reduction.
59 | P a g e
Algorith)
3a0elet
used and
Noise t%"e
I)"ro0e)ent
in !NR 'd5(
Co)"utation
Ti)e '!ec(
#eatures
E&&icient
Medical
i)age
denoising
algorith)
Xuadratic
spline
wavelet
&peckle
noise and
)ician
noise
3ltra &ound'
;.57
4)I' W.<=
6.756
55.788
2oefficient classification
techni+ue. Adapts itself to
various types of noise and
preference to expert.
Applications' &peckle and
)ician noise reduction.
CHAPTER-:
Conclusion and &uture sco"e
60 | P a g e
:. Conclusion and #uture !co"e
:.1 Conclusion and &uture sco"e
In this thesis! we have developed an :fficient De-noising algorithm using the joint
statistics of the wavelet coefficients and consider the dependencies between the coefficients.
we have addressed the development of new algorithms for improvement in image
reconstruction by suppressing noise that commonly arise in 3ltra sound and 4) images
using advanced wavelets such as +uadratic spline wavelet transform. and demonstrated to be
effectively preserving the local features such as edges! corners and local orientations. "hese
techni+ues are important in medical diagnosis applications. "he +uantitative comparison of
performance of denoising algorithm by widely used metrics! signal to noise ratio .&N)/! and
peak signal to noise ratio .*&N)/.
#e have Implemented the proposed algorithm to the three 3ltra sound test images
with ;x; window si%e and with variable tuning parameter Q. #e can notice that our proposed
filter exhibits more than 6d9 improvement in both &N) and *&N) over 1omomorphic
#iener filter and the proposed algorithmCs computation time is 6 seconds more than the
1omomorphic wiener filterCs computation time. "he results clearly demonstrate that the
proposed filter outperforms the homomorphic spatially adaptive #iener filtering both in
terms of &N) and *&N).
we have implemented the proposed algorithm to pelvic! brain and spline clinical 4)
images with si%es <88x<6<!8=5x785 and 8HAx8W< respectively with standard deviation YN7A.
#e can observe that ;.7; d9 improvement in both &N) and *&N). "he computation time
for proposed algorithm is some wards higher than the spatially adaptive filter. In our
experiments! on a number of different reference 4)I images! the improvement over the
61 | P a g e
spatially adaptive #iener filtering was at least A.; d9 .for relatively clean image/ and more
than 7 d9 for low &N) images! or the window si%e 7x7.
A wide range of values can be used for tuning parameter Q. for ultrasound images
QN7 is optimal and for 4) images QN6 is optimal.
#e have demonstrated and implemented the proposed algorithm to 3ltrasound
Images and 4)I Images! and :valuated their performance with respect to 1omomorphic
filter. we have compared the &N) and "ime :lapsed for execution of proposed method with
1omomorphic filter. we observed that the proposed algorithm has a better performance in
&N)! which is a very useful factor in denoising by which a versatile denoising algorithm has
been constructed and Implemented.
urther work focuses on the extension of multiresolution based denoising
schemes to 7D-voxel data. Additionally! a detailed comparison of the denoising schemes
before and after a 2"-reconstruction is needed. "hough the applications are different! the
various denoising schemes perform within their limit. "here must be a techni+ue which can
be applied globally for all types of noisy medical images irrespective of the applications. "he
future research gives the scope for such medical image denoising algorithm which also helps
in preserving the necessary sharp details of the medical image.
62 | P a g e
)::):N2:&
D 5 E Abramovich .! &apatinas ".! and &ilverman 9.!,#avelet thresholding via a bayesian
approach!- 5. of t!e 7oyal Statist. Society $ <A! pp. =6;Z=8H! 5HHW.
D 6 E Abramovich .!9esbeas ".! and &apatinas ".! ,:mpirical 9ayes approach block
wavelet function estimation!- &omput. Statist. 2ata ;nal.! vol. 7H! pp. 87;-8;5! 6AA6.
D 7 E Achim A.! 9e%erianos A.! and "sakalides *.! ,Novel 9ayesian multiscale method for
speckle removal in medical ultrasound images!- 8%%% <rans. on 6edical 8maging! vol. 6A!
no. W! pp. ==6-=W7! Aug 6AA5.
D 8 E 9anham!4.).KQatsaggelos!A.Q.,Digital Image )estoration!-I::: &ignal processing
4ag%ine 58!68-85!5HH=.
D ; E 9eylkin! $.,(n t!e 7epresentation of (perators in $ases of &ompactly Supp-orted
)avelets!- &chlumbergerZDoll )esearch! )idgefield! 2"! 5HH6.
D < E 9harath A.A. and Ieffry. Ng! ,A &teerable 2omplex #avelet 2onstruction and
its Application to Image Denoising!-I::: "rans. on. Image *rocessing!Jol.58! No.=! Iuly
6AA;.
D = E 9hatia 4! Qarl #.2.! and #illsky A.&.!-A #avelet based method for 4ulti scale
"opographic )econstruction!- I::: "ran. on Image *rocessing! Jol.5;! No. 5! pp. H6-5A5.
5HH<.
D W E 9ottema!4.!4oran!9.!and&uvorova! &.!,An Application of #avelets in "omography!-
2igital Signal Processing ; .5HHW/! 688Z6;8
D H E 2ai [.!2heng ".1.!(u 2.!and &ubramanian Q. ).!,:fficient wavelet based image
denoising algorithm!-%lectron. "ett.! vol. 7=! no.55!pp.<W7-<W;! 4ay 6AA6.
D 5A E 2hambolle A.!DeJore ). A.! (ee N.!and (ucier 9. I.! ,Nonlinear wavelet image
processing' Jariational problems!compression! and noise removal through wavelet
shrinkage!- I::: "rans. Image *rocessing! vol. =! pp. 75H-77;! 4ar.5HHW
D55E 2hang &.$.! Uu 9.! and Jetterli 4.! ,Adaptive wavelet thresholding for image
denoising and compression!-I::: "rans.Image *roc.!vol. H! no. H! pp. 5;76-5;8<!&ept. 6AAA.
D56E Dr.4.Jenugopal )ao! Advanced wavelet techni+ues for improved image
reconstruction in 2omputer Aided "omography! osmania university! *hd "hesis!6A55.
63 | P a g e
D57E *i%urica A.!,Image denoising using wavelets and spatial context modeling!- *hD
thesis! $hent 3niversity! $hent! 9elgium! 6AA6.
D58E Jenu $oapal )ao 4.! and Jathsal &.! ,#avelet methods for Noise reduction in
2omputed "omography!- at 9elarus-Indian &cientific and technical workshop! New Delhi!
68-6<! eb 6AA7.
D5;E Jenu $oapal )ao 4.! and Jathsal &.! -(ocal Adaptive 9ivariate &hrinkage
Algorithm for 4edical Image Denoising!- International Iournal of :lectronics :ngineering
.II::/! 5.5/! pp!;H-<;! Ian 6AAH.
64 | P a g e
A""endi8
65 | P a g e
A""endi8-A
A.3A<E2ET TRAN!#=RM
A.1 Introduction
#avelets are mathematical functions defined over a finite interval and having an
average value of %ero that transform data into different fre+uency components! representing
each component with a resolution matched to its scale.
"he basic idea of the wavelet transform is to represent any arbitrary function as a
superposition of a set of such wavelets or basis functions. "hese basis functions or baby
wavelets are obtained from a single prototype wavelet called the mother wavelet! by dilations
or contractions .scaling/ and translations .shifts/. "hey have advantages over traditional
ourier methods in analy%ing physical situations where the signal contains discontinuities and
sharp spikes. 4any new wavelet applications such as image compression! turbulence! human
vision! radar! and earth+uake prediction are developed in recent years. In wavelet transform
the basis functions are wavelets. #avelets tend to be irregular and symmetric. All wavelet
functions! '.6t - m/! are derived from a single mother wavelet! '.t/. "his wavelet is a small
wave or pulse like the one shown in ig. A.5 .
#ig. A.1 Mother Ca0elet C't(
Normally it starts at time t N A and ends at t N <. "he shifted wavelet '.t - m/ starts at
t N m and ends at t N m \ <. "he scaled wavelets '.6t/ start at t N A and end at t N <@6. "heir
graphs are '.t/ compressed by the factor of 6 as shown in ig. A.6. or example! when N
66 | P a g e
5! the wavelet is shown in ig A.6 .a/. If N 6 and 7! they are shown in .b/ and .c/!
respectively.
'a(C'2t( '$(C'+t( 'c(C';t(
#ig. A.2 !caled Ca0elets
"he wavelets are called orthogonal when their inner products are %ero. "he smaller
the scaling factor is! the wider the wavelet is. #ide wavelets are comparable to low-
fre+uency sinusoids and narrow wavelets are comparable to high-fre+uency sinusoids.
A.2 !CA2ING
#avelet analysis produces a time-scale view of a signal. &caling a wavelet simply
means stretching .or compressing/ it. "he scale factor is used to express the compression of
wavelets and often denoted by the letter a. "he smaller the scale factor! the more
,compressed- the wavelet. "he scale is inversely related to the fre+uency of the signal in
wavelet analysis.
A. !HI#TING
&hifting a wavelet simply means delaying .or hastening/ its onset. 4athematically!
delaying a function f.t/ by is represented by f.t-k/ and the schematic is shown in fig. A.7


67 | P a g e
'a( 3a0elet &unction 't( '$( !hi&ted Ca0elet &unction 't-,(
#ig. A. !hi&ted Ca0elets
A.+ !CA2E AN1 #RE9@ENC>
"he higher scales correspond to the most ,stretched- wavelets. "he more stretched the
wavelet! the longer the portion of the signal with which it is being compared! and thus the
coarser the signal features being measured by the wavelet coefficients. "he relation between
the scale and the fre+uency is shown in ig. A.8.
2oC scale High scale
#ig. A.+ !cale and &re/uenc%
"hus! there is a correspondence between wavelet scales and fre+uency as revealed by
wavelet analysis'
(ow scale a 2ompressed wavelet )apidly changing details 1igh fre+uency.
1igh scale a &tretched wavelet &lowly changing! coarse features (ow fre+uency.
A.- 1I!CRETE 3A<E2ET TRAN!#=RM
2alculating wavelet coefficients at every possible scale is a fair amount of work! and
it generates an awful lot of data. If the scales and positions are chosen based on powers of
two! the so-called dyadic scales and positions! then calculating wavelet coefficients are
efficient and just as accurate. "his is obtained from discrete wavelet transform .D#"/.
68 | P a g e
A.-.1 =NE-!TAGE #I2TERING
or many signals! the low-fre+uency content is the most important part. It is the
identity of the signal. "he high-fre+uency content! on the other hand! imparts details to the
signal. In wavelet analysis! the approximations and details are obtained after filtering. "he
approximations are the high-scale! low fre+uency components of the signal. "he details are
the low-scale! high fre+uency components. "he filtering process is schematically represented
as in ig. A.;.
#ig. A.- !ingle stage &iltering
"he original signal! &! passes through two complementary filters and emerges as two
signals. 3nfortunately! it may result in doubling of samples and hence to avoid this! down
sampling is introduced. "he process on the right! which includes downsampling! produces
D#" coefficients. "he schematic diagram with real signals inserted is as shown in ig. A.<.

#ig. A.. 1eco)"osition and deci)ation
69 | P a g e
A.-.2 M@2TIP2E-2E<E2 1EC=MP=!ITI=N
"he decomposition process can be iterated! with successive approximations being
decomposed in turn! so that one signal is broken down into many lower resolution
components. "his is called the wavelet decomposition tree and is depicted as in ig. A.=.

#ig. A.: Multile0el deco)"osition
A.-. 3A<E2ET REC=N!TR@CTI=N
"he reconstruction of the image is achieved by the inverse discrete wavelet transform
.ID#"/. "he values are first upsampled and then passed to the filters. "his is represented as
shown in ig. A.W.
#ig. A.; 3a0elet Reconstruction
70 | P a g e
"he wavelet analysis involves filtering and downsampling! whereas the wavelet
reconstruction process consists of upsampling and filtering. 3psampling is the process of
lengthening a signal component by inserting %eros between samples as shown in ig. A.H.

#ig. A.E Reconstruction using u"sa)"ling
A.-.+ REC=N!TR@CTING APPR=?IMATI=N! AN1 1ETAI2!
It is possible to reconstruct the original signal from the coefficients of the
approximations and details. "he process yields a reconstructed approximation which has the
same length as the original signal and which is a real approximation of it.
"he reconstructed details and approximations are true constituents of the original
signal. &ince details and approximations are produced by downsampling and are only half the
length of the original signal they cannot be directly combined to reproduce the signal. It is
necessary to reconstruct the approximations and details before combining them. "he
reconstructed signal is schematically represented as in ig. A.5A.
#ig. A.1F Reconstructed signal co)"onents
71 | P a g e
A.-.- 1-1 3A<E2ET TRAN!#=RM
"he generic form for a one-dimensional .5-D/ wavelet transform is shown in ig.
7.55. 1ere a signal is passed through a low pass and high pass filter! h and g! respectively!
then down sampled by a factor of two! constituting one level of transform.
#ig. A.11 11 3a0elet 1eco)"osition
)epeating the filtering and decimation process on the low pass branch outputs make
multiple levels or ,scales- of the wavelet transform only. "he process is typically carried out
for a finite number of levels Q! and the resulting coefficients are called wavelet coefficients.
"he one-dimensional forward wavelet transform is defined by a pair of filters s and t that are
convolved with the data at either the even or odd locations. "he filters s and t used for the
forward transform are called analysis filters.

6
n
l
i j i j
j n
l
l s x
+

and 6 5
n
!
i j i j
j n
!
! t x
+ +


.A.5/
Although l and h are two separate output streams! together they have the same total
number of coefficients as the original data. "he output stream l! which is commonly referred
to as the low-pass data may then have the identical process applied again repeatedly. "he
other output stream! h .or high-pass data/! generally remains untouched. "he inverse process
72 | P a g e
expands the two separate low- and high-pass data streams by inserting %eros between every
other sample! convolves the resulting data streams with two new synthesis filters s0 and t0!
and adds them together to regenerate the original double si%e data stream.

0 0 0 0

n n
! l
i j i j j i j
j n j n
! !
y t l s !
+ +


+ where l
0
6i
N l
i
! l
0

6i\5
N A .A.6/
h
0
6i\5
N h
i
! h
6i
0
N A
"o meet the definition of a wavelet transform! the analysis and synthesis filters s! t! s0
and

t0 must be chosen so that the inverse transform perfectly reconstructs the original data.
&ince the wavelet transform maintains the same number of coefficients as the original data!
the transform itself does not provide any compression. 1owever! the structure provided by
the transform and the expected values of the coefficients give a form that is much more
amenable to compression than the original data. &ince the filters s! t! s
0
and t
0
are chosen to be
perfectly invertible! the wavelet transform itself is lossless. (ater application of the
+uanti%ation step will cause some data loss and can be used to control the degree of
compression. "he forward wavelet-based transform uses a 5-D subband decomposition
process? here a 5-D set of samples is converted into the low-pass subband .(i/ and high-pass
subband .1i/. "he low-pass subband represents a down sampled low-resolution version of the
original image. "he high-pass subband represents residual information of the original image!
needed for the perfect reconstruction of the original image from the low-pass subband
A.-.. 2-1 TRAN!#=RM HEIRARCH>

73 | P a g e
((5 1(5
(15 115
#ig. A.12 !u$ $and 2a$elling !che)e &or a one le0el* 2-1 3a0elet Trans&or)
"he 5-D wavelet transform can be extended to a two-dimensional .6-D/
wavelet transform using separable wavelet filters. #ith separable filters the 6-D transform
can be computed by applying a 5-D transform to all the rows of the input! and then repeating
on all of the columns.
"he original image of a one-level .QN5/! 6-D wavelet transform! with
corresponding notation is shown in ig. A.56. "he example is repeated for a three-level .Q
/7/ wavelet expansion in ig. A.57. In all of the discussion 3 represents the highest level of
the decomposition of the wavelet transform
.
#ig. A.1 !u$$and la$elling !che)e &or a Three 2e0el* 2-1 3a0elet Trans&or)
"he 6-D subband decomposition is just an extension of 5-D subband decomposition.
"he entire process is carried out by executing 5-D subband decomposition twice! first in one
74 | P a g e
((
5
1(
5
1(
6
1(
7
(1
5
11
5
(1
6
11
6
(1
7
11
7
direction .hori%ontal/! then in the orthogonal .vertical/ direction. or example! the low-pass
subbands .(i/ resulting from the hori%ontal direction is further decomposed in the vertical
direction! leading to ((i and (1i subbands.
&imilarly! the high pass subband .1i/ is further decomposed into 1(i and 11i. After
one level of transform! the image can be further decomposed by applying the 6-D subband
decomposition to the existing ((i subband. "his iterative process results in multiple
,transform levels-. In ig. A.57 the first level of transform results in (15! 1(5! and 115! in
addition to ((5! which is further decomposed into (16! 1(6! 116! ((6 at the second level!
and the information of ((6 is used for the third level transform. "he subband ((i is a low-
resolution subband and high-pass subbands (1i! 1(i! 11i are hori%ontal! vertical! and
diagonal subband respectively since they represent the hori%ontal! vertical! and diagonal
residual information of the original image.
"o obtain a two-dimensional wavelet transform! the one-dimensional transform is
applied first along the rows and then along the columns to produce four subbands' low-
resolution! hori%ontal! vertical! and diagonal. ."he vertical subband is created by applying a
hori%ontal high-pass! which yields vertical edges./ At each level! the wavelet transform can
be reapplied to the low-resolution subband to further decorrelate the image. ig. 7.58
illustrates the image decomposition! defining level and subband conventions used in the
A#I2 algorithm. "he final configuration contains a small low-resolution subband. In
addition to the various transform levels! the phrase level A is used to refer to the original
image data. #hen the user re+uests %ero levels of transform! the original image data .level A/
is treated as a low-pass band and processing follows its natural flow.
2oC Resolution !u$$and
75 | P a g e
8
7
(evel 6
(evel 5
Jertical subband
1(
8 8
7 7
(evel 6 (evel 6
(evel 5
1ori%ontal &ubband
(1
(evel 5
Diagonal &ubband
11
#ig. A.1+ I)age 1eco)"osition @sing 3a0elets
#avelet transform is first performed on each source images! then a fusion decision
map is generated based on a set of fusion rules. "he fused wavelet coefficient map can be
constructed from the wavelet coefficients of the source images according to the fusion
decision map. inally the fused image is obtained by performing the inverse wavelet
transform.
76 | P a g e
A""endi8-5
"u$lications $% the candidate

77 | P a g e
78 | P a g e

You might also like