PDF Ieeee

U
UNIVERSITY OF CINCINNATI
Date:
I, ,
hereby submit this original work as part of the requirements for the degree of:

in
It is entitled:

Student Signature:
This work and its defense approved by:
Committee Chair:
Approval of the electronic document:
I have reviewed the Thesis/Dissertation in its final electronic format and certify that it is an
accurate copy of the document reviewed and approved by the committee.
Committee Chair signature:
02/02/2009
Xin Hu
Master of Science
Electrical Engineering
An Improved 2D Adaptive Smoothing Algorithm in
Image Noise Removal and Feature Preservation
William G. Wee
Jing-Huei Lee
Chia-Yung Han

An Improved 2D Adaptive Smoothing
Algorithm in Image Noise Removal and
Feature Preservation
A thesis submitted to the
Graduate School
of the University of Cincinnati

In partial fulfillment of the requirements for the degree of

MASTER OF SCIENCE
in
ELECTRICAL ENGINEERING

in the Department of Electrical and Computer Engineering
of the College of Engineering

2009

By
Xin Hu
B.S. Tsinghua University
P.R.China, 2005
Committee Chair: Dr. William G. Wee
Abstract
We introduce an improved 2D adaptive smoothing algorithm for noise removal and
feature preservation. Comparing to the original 2D adaptive smoothing algorithm, this new
algorithm is also based on the novel idea of utilizing contextual discontinuity and local
discontinuity jointly to detect and distinguish edges and noise. The new algorithm improves the
main concept -- contextual discontinuity by introducing a novel homogeneity region definition
with a corresponding method for contextual discontinuity measurement. Comparing to the
original algorithm and other smoothing algorithms, the improved algorithm can preserve edges
more effectively while removing noise.
The improved 2D algorithm has been implemented and extensive experiments have been
carried out to compare the algorithm to the original algorithm and other smoothing strategies to
quantitatively demonstrate improvement in performance. Measurements are applied to evaluate
the noise removal and edge preservation performance. Simulation results show that this
improved algorithm has a superior performance over both the original algorithm and other
popular smoothing strategies in noise removal as well as feature preservation.

Acknowledgement
I would like to express my gratitude to all those who gave me the possibility to
complete this thesis.
I am deeply indebted to my advisor Prof. William G. Wee for his supervision and
encouragement in all the time of my research and completion of this thesis. Also, I
would like to thank Dr. Jing-huei Lee and Dr. Chia-Yung Han for their kindly help and
valuable advices.
My former colleagues from the Department of ECE supported me in my research
work. I want to thank them for all their help, support, interest and valuable suggestions.
Especially I am obliged to Xiang Cai, Hui Peng and Joe Kesler for their professional
advice in my research.
Last but not the least, my special thanks go to my parents and my fianc Wei
Huai whose patient love enabled me to complete this work.

1

Table of Content
List of Figures ................................................................................................................................. 4
List of Table .................................................................................................................................... 7
Chapter1 Introduction ................................................................................................................... 10
1.1 Overview of Image Noise and Image Smoothing ................................................................ 10
1.2 Review of Image Smoothing Algorithms ............................................................................ 12
1.3 Thesis Organization ............................................................................................................. 17
Chapter2 2D Adaptive Smoothing Algorithm and Its Deficiencies ............................................. 18
2.1 Overview ............................................................................................................................. 18
2.2 The Basic Ideas of Adaptive Smoothing Algorithm ........................................................... 19
2.2.1 The Contextual Discontinuity Measurement ................................................................ 20
2.2.2 Use Discontinuity Map to Smooth Image .................................................................... 22
2.3 The Resulting Adaptive Smoothing Algorithm ................................................................... 24
2.4 Advantages and Deficiencies of this 2D Adaptive Smoothing Algorithm .......................... 28
2.4.1 Advantages of this 2D Adaptive Smoothing Algorithm ............................................... 28
2.4.2 Deficiencies of this 2D Adaptive Smoothing Algorithm.............................................. 29
2.4.2.1 Error in Homogeneity Neighborhood Definition ................................................... 29
2.4.2.2 Error in Contextual Discontinuity Design ............................................................. 31
2.5 Conclusion ........................................................................................................................... 32
Chapter3 Research Problem, Objective ........................................................................................ 34
2

3.1 Research Problem ................................................................................................................ 34
3.2 Research Objective .............................................................................................................. 35
Chapter4 Improved 2D Adaptive Smoothing Algorithm ............................................................. 37
4.1 Overview ............................................................................................................................. 37
4.2 Proposed Improvements ...................................................................................................... 38
4.2.1 Generation of a Correct Homogeneity Neighborhood.................................................. 38
4.2.2 Incorporation of Direction Information in H Design .................................................... 41
4.3 Resulting Improved Algorithm ............................................................................................ 42
4.4 Advantages and Disadvantages of the Proposed Algorithm................................................ 44
4.5 Implementation of the Improved Smoothing Algorithm ..................................................... 45
Chapter5 Experimental Results .................................................................................................... 46
5.1 Overview ............................................................................................................................. 46
5.2 Comparison Experiment Results ......................................................................................... 46
5.2.1 Simulated Image Experiment Results........................................................................... 47
5.2.2 Real Image Experiment Results ................................................................................... 52
5.2.3 Simulated MRI Image Experiment Results .................................................................. 75
5.3 Experimental Parameter Determination .............................................................................. 81
5.3.1 Parameter Sensitivity Investigation .............................................................................. 82
5.3.2 Applicable Parameter Selection Algorithm ................................................................ 102
5.4 Conclusion ........................................................................................................................ 107
3

Chapter6 Conclusions .................................................................................................................. 109
Bibliography ................................................................................................................................ 110

4

List of Figures
Figure 1.1 Framework of applying fuzzy connectedness concept in image smoothing and
segmentation ...16
Figure 2.1 Comparison of local discontinuity and contextual discontinuity ................................ 23
Figure 2.2 Flow chart of Chens Adaptive Smoothing Algorithm ................................................ 25
Figure 2.3A Pixel near image boundary ........................................................................................ 32
Figure 2.3B Pixel in smooth area corrupted by image noise ......................................................... 32
Figure 4.1 An illustration of two HN of two pixels close to edge ................................................. 40
Figure 4.2 Flow chart of Chens Algorithm .................................................................................. 44
Figure 5.1a Simulated test image .................................................................................................. 47
Figure 5.1b Smoothing results comparison .................................................................................. 49
Figure 5.2 Noise Free Images ........................................................................................................ 53
Figure 5.3 Smoothing result comparison on Lena Image .............................................................. 56
Figure 5.4 Smoothing result on MRI simulated noisy image ........................................................ 76
Figure 5.5 MRI real brain image smoothing result example 1 ...................................................... 78
Figure 5.9 Simulated Image - SNR performance along with iteration number increase ............... 85
Figure 5.10 Simulated Image - Best SNR obtained under different parameter h .......................... 86
5

Figure 5.11 Simulated Image - Iteration number corresponding to best SNR obtained under
different parameter ........................................................................................................................ 86
Figure 5.12 Simulated Image - SNR performances under different h and Iteration number ........ 87
Figure 5.13 Lena Image - SNR performance along with iteration number increase ..................... 88
Figure 5.14 Lena Image - Best SNR obtained under different parameter h .................................. 89
Figure 5.15 Lena Image - Iteration number corresponding to best SNR obtained under different
parameter h .................................................................................................................................... 89
Figure 5.16 Lena Image - SNR performances under different h and Iteration number ................. 90
Figure 5.17 Camera man Image - SNR performance along with iteration number increase ......... 91
Figure 5.18 Camera man Image - Best SNR obtained under different parameter h ...................... 92
Figure 5.19 Camera man Image - Iteration number corresponding to best SNR obtained under
different parameter h ..................................................................................................................... 92
Figure 5.20 Camera man Image - SNR performances under different h and Iteration number .... 93
Figure 5.21 Image - SNR performance along with iteration number increase .............................. 94
Figure 5.22 Scene Image - Best SNR obtained under different parameter h ................................. 94
Figure 5.23 Scene Image - Iteration number corresponding to best SNR obtained under different
parameter h .................................................................................................................................... 95
Figure 5.24 Scene Image - SNR performances under different h and Iteration number ............... 96
Figure 5.25 Human Face Image - SNR performance along with iteration number increase ......... 97
Figure 5.26 Human Face Image - Best SNR obtained under different parameter h ...................... 98
6

Figure 5.27 Human Face Image - Iteration number corresponding to best SNR obtained under
different parameter h ..................................................................................................................... 98
Figure 5.28 Human Face Image - SNR performances under different h and Iteration number .... 99
Figure 5.29 Pepper Image - SNR performance along with iteration number increase ................ 100
Figure 5.30 Pepper Image - Best SNR obtained under different parameter h ............................. 100
Figure 5.31 Pepper Image - Iteration number corresponding to best SNR obtained under different
parameter h .................................................................................................................................. 101
Figure 5.32 Pepper Image - SNR performances under different h and Iteration number ............ 102
Figure 5.33 A 3D performance illustration for Figure 5.2a ......................................................... 104
Figure 5.34 Best stopping iteration number versus h for all 5 images in Figure 5.2 ................... 105
Figure 5.35 Illustration of obtained SNR/best SNR versus h for all images in Figure 5.2.......... 105
Figure 5.36 Best stopping iteration number versus h for all 5 images in Figure 5.2 using Chens
Algorithm .................................................................................................................................... 106
Figure 5.37 Illustration of (obtained SNR/best SNR) versus h for all images in Figure 5.2 using
Chen algorithm ............................................................................................................................ 107

7

List of Tables
Table 1.1 Classification of different non-linear smoothing algorithms ......................................... 14
Table 5.1 Smoothing results for Chens Algorithm and my algorithm (SNR) .............................. 48
Table 5.2 Smoothing results for Chens Algorithm and my algorithm (DR & FAR) ................... 51
Table 5.3 Comparison of performances for Figure 5.2(a) ............................................................. 55
Table 5.4 Smoothing results comparison for Lena image (without small region removal process)
(SNR)............................................................................................................................................ 56
Table 5.5 Smoothed results of Lena image for Chens Algorithm and our algorithm (Threshold
value =0.15 & no small region removal process) .......................................................................... 57
value =0.1 & no small region removal process) ............................................................................ 58
value =0.05 & small region removal process) ............................................................................... 60
Table 5.9 Smoothing results comparison for Camera man image (without small region removal
process) .......................................................................................................................................... 61
Table 5.10 Smoothed results of Camera Man image for Chens Algorithm and my algorithm
(Threshold value =0.07 & no small region removal process) ....................................................... 62
Table 5.11 Smoothed results of Camera Man image for Chens Algorithm and our algorithm
(Threshold value =0.07 & small region removal process) ............................................................ 63
8

Table 5.12 Smoothing results comparison for Scene image (without small region removal process)
....................................................................................................................................................... 64
Table 5.13 Smoothed results of Scene image for Chens Algorithm and my algorithm (Threshold
Table 5.15 Smoothing results comparison for Human Face image (without small region removal
process) .......................................................................................................................................... 67
Table 5.16 Smoothed results of Human Face image for Chens Algorithm and my algorithm
(Threshold value =0.05 & no small region removal process) ....................................................... 68
(Threshold value =0.05 & small region removal process) ............................................................ 69
Table 5.18 Smoothing results comparison for Pepper image (without small region removal
process) ........................................................................................................................................ 70
Table 5.19 Smoothed results of Pepper image for Chens Algorithm and my algorithm (Threshold
Table 5.21 moothing results for Chens Algorithm and my algorithm (SNR) .............................. 74
Table 5.22 Smoothing results comparison for 5 real images (without small region removal
process) .......................................................................................................................................... 74
Table 5.23 Smoothing results comparison for 5 real images (with small region removal process)
....................................................................................................................................................... 75
9

Table 5.24 Smoothed results of MRI image for Chens Algorithm and my algorithm (SNR) ..... 77

10

Chapter1
Introduction

1.1 Overview of Image Noise and Image Smoothing
No matter how good the imaging devices are, images acquired from the real world are
inevitably corrupted by noise from various sources. The term of image noise, in the image
processing, usually refers to the high spatial frequency random perturbations of pixel brightness
[1]. The most common types of noise are Gaussian noise as well as salt and pepper noise [1].
A digital image is generally encoded as a matrix of grey-level or color values. Taking a
2D gray-level digital image I(x, y) as an example, (x, y) corresponds to a pixel on a 2D grid and
I(x, y) is a real value usually raging from 0-255. In the case of 3D images, (x, y) is extended to
be (x, y, z) representing a point in 3D space.
11

Due to the limitation of the device and the influence of the environment, the resulting
image is always contaminated with noise. In a rough approximation one can write I(x, y) = u(x, y)
+ n(x, y), where (x, y) represents a particular pixel in an image, I(x, y) is the observed gray value,
u(x, y) would be the true value at pixel (x, y), and n(x, y) is the noise perturbation.
There are numerous types of noise that exist in images. Various noise models are
designed to simulate n(x, y). The most widely used noise models are Gaussian noise and salt and
pepper noise [1]. In the case of Gaussian noise, the overall noise is generally assumed to be an
additive Gaussian distribution with a zero-mean and a constant variance. The salt and pepper
noise assumes a constant proportion of pixels are corrupted to be extreme white or black. In these
noise models, the normalized values of n(x, y) and n(i, j) at different pixels are assumed to be
independent random variables or white noise.
Generally speaking image smoothing is a method for image enhancement with the
objective of removing image noise. If the image noise can be modeled as I(x, y) = u(x, y) + n(x,
y), where (x, y) represents a particular pixel in an image, I(x, y) is the observed gray value, u(x, y)
would be the true value at pixel (x, y), and n(x, y) is the noise perturbation. The goal of an
image smoothing algorithm is to recover u(x, y) from I(x, y). However in most situations it is
impossible to find n(x, y). Thus it is impossible to completely remove noise and recover the
original image. In most cases an image smoothing algorithm will blur the image to some extent
while removing the image noise.
In image processing, smoothness is a generic assumption that many types of images have
approximately piece-wise constant gray levels [1]. Based on this widely adopted assumption, an
image can be viewed as the combination of many constant areas. Inside each of this area the gray
value is approximately constant. The conjunctions of those constant areas, or say boundary areas,
can be defined as image edges. Pixels in these boundary areas are called edge pixels. These
image edges are very important in charactering image contents.
12

To effectively enhance image quality, the image smoothing algorithm must be able to
preserve image edges while removing the image noise. Most of current image smoothing
algorithms do not have a proper image edge feature definition and therefore can not effectively
separate edge pixels from other image pixels. Thus either the smoothing result is blurred or the
image noises are not effectively removed.

1.2 Review of Image Smoothing Algorithms
Generally image smoothing algorithms can be roughly classified into two categories:
linear image smoothing algorithms and nonlinear image smoothing algorithms [1]. Linear image
smoothing algorithms apply the same smoothing operation on all image pixels and normally have
fixed smoothing filter shape and size. They are simple in design and fast in application. However,
the shortcoming is that they may blur important features.
Another type of image smoothing algorithm is nonlinear image smoothing algorithm.
The general goal of various nonlinear smoothing algorithms that have been developed is to
preserve important edge features while noise is removed during the smoothing process. The
intrinsic idea of nonlinear smoothing algorithm having better performance over linear smoothing
algorithm is that nonlinear one may vary their shapes and sizes, or modify their smoothing update
strategies accordingly and perform different filtering at different parts of an image, i.e. to apply a
smoothing operator which adapts itself to the local topography of the image [3]. Extensive
research on how to adaptively modify the smoothing operation according to image content has
been conducted in the last two decades [1, 2, 3, 4, 5].
Since nonlinear image smoothing algorithms have much better performance over linear
smoothing algorithms in image edge feature preservation and noise removal, we will only focus
on nonlinear image smoothing algorithms.
13

Nonlinear image smoothing algorithms can be further divided into different categories
based on different operations. Several common standards are listed as following:
Iterative and Non-Iterative: The adopted smoothing procedure in the image smoothing
algorithm is an iterative smoothing procedure or not.
Contextual and Non-Contextual: The image smoothing algorithm uses the neighborhood of
one pixel to decide this pixels property or not.
Scaled based and Non-Scale based: When the neighborhood is used to decide one pixels
property, the size of the neighborhood is fixed or adaptive to different pixels.
In the scale-based algorithms, different sizes of neighborhood around a pixel, which
involve in the smoothing of the pixel, are selected according to the image brightness distribution
around that pixel. While in the non-scale based methods, the size of smoothing area is pre-
selected and fixed across the whole image.
One general conclusion about these different types of algorithms is that the contextual
smoothing algorithms have better performance over non-contextual ones on edge preservation
and noise removal. The scale based algorithms are superior to the non-scale based algorithms in
the same way.
The table below is a comparison of the most popular nonlinear image smoothing
algorithms [2]. Extensive review and analysis about these smoothing algorithms are available in
[1, 2, 3]. Here we will only concentrate on the last two sets of algorithms, fuzzy-connected based
methods and its closest derivation, local and contextual discontinuities based methods.
Comparing to other smoothing algorithms, these two apply novel mathematic concepts to image
smoothing and have better performance on edge preservation and noise removal.

14

Smoothing Algorithms
Classification from
iteration point of
view
Classification from
scale point of view
Classification from
contextual information
point of view
Order statistic neighbor filters
[1,4,5]

Non-iterative

Non-scale-based

Non-contextual
K-nearest neighbor operator, the
-trimmed mean and the sigma
filter [6, 7,8]
Non-iterative Non-scale-based Non-contextual
Peak Noise Filter [1] Non-iterative Non-scale-based Non-contextual
Adaptive Template Filter [9] Non-iterative Scale-based Contextual
Adaptive Wiener Filter [1] Non-iterative Non-scale-based Contextual
SUSAN [10] Non-iterative Scale-based Contextual
0,1 Mask adaptive smoothing
filter [11]
Non-iterative Scale-based Contextual
Adaptive smoothing filter
derived from a clustering method
[12, 13]
Iterative Scale-based Contextual
Fuzzy-connectedness based
methods [14, 15,16, 17, 18]
Local & contextual
discontinuities combined method
[19, 20]

Table 1.1 Classification of difference non-linear smoothing algorithms
15

The implementation of fuzzy connectedness concepts and objection definition were first
introduced by Udupa and his collaborators [14, 15]. Their work extended the mathematic fuzzy
concepts to image processing. Fuzzy connectedness was used as a measurement of image context
closeness of two adjacent pixels in an image. The fuzzy connectedness between two pixels was
computed based on both their spatial distance and intensity difference. It was computed as the
weighted sum of the intensity and the intensity gradient in the neighborhood of the pixel in order
to capture the intensity features and patterns of intensity variations. Though any neighborhood
size was allowed for the connectedness calculation, it was not practical to perform the calculation
for an unreasonably large size. It may also be inaccurate and not robustness to perform the
calculation for a very small size. By applying previous work of object scale in the framework of
fuzzy connectedness, the size of the neighborhood was allowed to be changed dynamically in
different parts of the image. Saha and Udupa argued that scale-based connectedness was natural
in object definition and also proved that it led to more effective object segmentation [16, 30].
This fuzzy based connectedness concept was then used in many image processing
applications. Fig 1.1 shows the framework of applying fuzzy connectedness concept in image
smoothing and image segmentation.
Udupa and Saha first realized this concept in medical image segmentation [17]. In this
application, as showed in Fig 1.1, the fuzzy connectedness combined two features together, the
object-based feature and homogeneity-based feature. Then a thresholding method that accounted
for fuzzy connectedness was used to perform image segmentation.
Ke Chen follows another way to develop a novel feature preserving scale-based iterative
adaptive smoothing algorithm as shown in Fig 1.1 [19, 20, 21, 22, 23, 24, 25, 26, 27]. He only
applies the homogeneity-based feature in defining one pixels surrounding property, contextual
discontinuity, which is further used to control the smoothing procedure. The proposed concept of
contextual discontinuity was first introduced as a measurement of inhomogeneity value which
was calculated in a larger neighborhood of every pixel. Based on both contextual and local
16

discontinuities (not shown in Fig 1.1), this smoothing scheme can effectively perform noise
removal while preserving features and the contextual discontinuity was further utilized to control
the smoothing speed.

Fig 1.1 Framework of applying fuzzy connectedness concept in image smoothing and
segmentation
In addition to noise removal, some degree of success in preserving image edge feature is
achieved by these most recent algorithms. However, many problems still exist. This thesis will
17

analyze these problems and propose improvements to obtain better edge preservation and noise
removal performance.

1.3Thesis Organization
The remainder of this thesis is organized as following: In Chapter 2, Chens image
smoothing algorithm will be introduced in detail including the limitations of the algorithm. The
research problem is presented in Chapter3. A list of research objectives are proposed in Chapter3
as well. In Chapter 4, an improved 2D smoothing algorithm based on a new contextual
discontinuity computation method will be presented. The simulation experiment results and
analysis of the parameter settings and stopping criterion for the improved algorithm are presented
in Chapter 5. Conclusions and future work are presented in Chapter 6.

18

Chapter2
2D Adaptive Smoothing Algorithm and
Its Deficiencies

2.1 Overview
In this Chapter, Chens 2D Adaptive Smoothing Algorithm (Chens Algorithm in the
following) will be presented and analyzed in detail. Then we will discuss the deficiencies in
Chens Algorithm which indicate the potential improvements.
There are several reasons that make us start our research from Chens Algorithm while
the most important reason is Chens Algorithm is the best till now in this area based on the
literature review in Chapter 1. As we discussed in Chapter 1, adaptive smoothing algorithms can
provide better smoothing result than non-adaptive smoothing algorithms since the former one can
adaptively modify the smoothing operation according to the image content such that how good
one adaptive smoothing algorithm can determine the image content can almost decide the
19

goodness of the smoothing result. Chens Algorithm makes a big progress by adapting fuzzy
connectedness concept in finding the image content around one pixel. Comparative experiments
in [20] show the advances of Chens Algorithm among other regular adaptive smoothing
algorithms. Our own experiment results presented in Chapter 4 also reveal its superior over other
peer image smoothing algorithms. Thus we choose Chens Algorithm as the start point of our
research on how to smooth image and preserve the image features.
Our objective to discuss Chens Algorithm is that we want to further improve the
smoothing result with edge preservation by providing improvements targeting on deficiencies of
Chens Algorithm. With this objective in this Chapter we will analyze the original idea behind
Chens Algorithm and discuss its deficiencies. This analysis and discussion is the base of our
improvements presented in Chapter 3.
In this Chapter, Section 2.2 presents the basic ideas of Chens Algorithm. Section 2.3
introduces the detail steps of Chens Algorithm. Section 2.4 analyzes the advantages and
deficiencies of Chens Algorithm. Section 2.5 provides the conclusion of this Chapter.

2.2 The Basic Ideas of Adaptive Smoothing Algorithm
Unlike most regular adaptive smoothing algorithms which use local discontinuity (like
gradient) to determine the smoothing speed (brightness update speed) at one pixel Chens
Algorithm combined two types of discontinuity measurement, contextual discontinuity (also
termed as inhomogeneity) and local discontinuity, to determine the smoothing speed. While the
local discontinuity is used for detection of local brightness discontinuities the contextual
discontinuity is designed to detect the smoothness of the image content in a relative larger area.
These two discontinuities are combined as a multi-scale measurement to detect possible image
features and decide a proper smoothing speed.
20

The following sections will emphasize on introducing and analyzing two important
concepts/operations in Chens Algorithm which are also the core of the algorithm: 1). contextual
discontinuity measurement; 2). Use discontinuity map to smooth the image.

2.2.1 The Contextual Discontinuity Measurement
To effectively evaluate the smoothness of a larger area, local measurements like gradient
are not enough since measurements based on derivatives are sensitive to all kinds brightness
change and cant separate noise from image content change. So we need some kind Contextual
Discontinuity measurement to detect the differences between brightness changes caused by
noise and image content change. Thus there are at least two requirements for this measurement:
1). New measurement must be defined in a larger area much bigger than local 3 by 3 neighbors;
2). New measurement should not be based on the simple derivatives.
To meet the above requirements, a Homogeneity Neighborhood (HN or short for the
rest of this thesis) is defined first to generate the Contextual Discontinuity measurement in
Chens Algorithm. It is defined that each pixel has a HN within which all pixels have the
similar image property similar brightness. Or in the other way this pixel and its HN belong to
the same object which is very smooth in brightness. The HN is defined as an isotropic ring shape
(square shape in practical applications) while the radius of the ring is the only parameter to be
decided. The size of the HN grows until some defined variance of the intensities inside the HN
meets a chosen threshold. However this isotropic HN can be quite small when the pixel is close
to the 'boundary' (could be the real boundary, also could be the fake boundary generated by noise).
The mathematical definition for HN is as [20]:
21

Based on HN, the contextual discontinuity measurement or inhomogeneity, H(x, y), is
defined to represent the brightness roughness of a larger surrounding around (x, y) and used to
determine the smoothing speed at pixel (x, y). In Chens Algorithm to derive H(x, y): first a
mutual size of HN for each pair of (x, y) and one of its 3 by 3 neighbors is selected by choosing
the smaller size of HN such that two pixels among each pair have the same size HNs. Then the
difference between pixels in corresponding positions in the pair of HN are weighted averaged to
reflect the brightness change in this pixel pair direction (from (x, y) to its neighbor). At last the
eight differences are averaged to generate H(x, y). The mathematical definition of calculating
22

H(x, y) is as (9), (10), (11), (12) in [36]. Theoretically the influences caused by noise in this
comparison can be ignored due to statistically reason.
Introducing the contextual discontinuity measure H(x, y) (or inhomogeneity) is the most
important contribution of this smoothing algorithm. The H(x, y) is crucial in this adaptive
smoothing algorithm since it is used to decide the smoothing speed at different locations. It is
also the most powerful part of this algorithm to use H(x, y), a region based discontinuity measure
instead of local discontinuity measure like local derivatives comparing to some classic algorithms.
Thus H(x, y) must be carefully selected to reflect how smooth the small region centered by (x, y)
is. The current H(x, y) computation algorithm is a product of applying fuzzy theory which has
very strong mathematical fundamentals into image processing. The basic concept is to find the
differences between neighbor pixels respective homogeneity neighborhood. This concept is
novel and reliable based on much research product of fuzzy theory [14, 15, 16, 17, 18].

2.2.2 Use Discontinuity Map to Smooth Image
So far we introduce two different discontinuity measurements used in Chens Algorithm:
local discontinuity and contextual discontinuity. Figure 2.1 demonstrates a noisy image as well
as its two different discontinuities map. It is obvious that the local discontinuity measurement is
very sensitive to all kinds of brightness change while the contextual discontinuity measurement is
more robust to noise. Contextual discontinuity map provides a clear clue leading to image
content change which also indicates the most important image feature -- edges.

23

Figure 2.1 Comparison of local discontinuity and contextual discontinuity
Noisy image (Upper Left), Local discontinuity measurement map (Upper Right), Contextual
discontinuity measurement map (Lower Center)

In Chens Algorithm both discontinuities are combined to decide the smoothing speed for
a particular pixel but playing different roles. Contextual discontinuity measurement reflects the
potential image edges existing so it is used to decide how fast to update this pixels brightness
while the local discontinuity reflects the difference of one pixel with its surrounding so it is used
to decide how big that the brightness of this pixel should be considered in updating its neighbors
24

brightness. Take an example that large H(x, y) indicates (x, y) may locate in a rough area with
image content change such that brightness of (x, y) should be updated less. While large gradient
in (x, y) only indicates that its brightness is very different with its neighbors due to either image
content change or noise. In either situation brightness of (x, y) should not be seriously considered
in updating other pixels brightness because of its uniqueness. The mathematical definition for
image smooth using both discontinuity measurements are listed as (13), (14), (15) in [20].

2.3 The Resulting Adaptive Smoothing Algorithm
Figure 2.2 shows the steps of Chens Algorithm. As it is shown below, for each noisy
image contextual discontinuity will be only calculated once. The brightness updated speed at (x,
y) in each iteration during this iterative smoothing process is decided by contextual discontinuity
H(x, y) while the local discontinuity is computed every iteration. As we mentioned above the
local discontinuity measurement is used to decide how big the brightness of this pixel should be
considered in updating its neighbors brightness. Keep updating gradient during the iteration in
case that some noise pixels are smoothed out and their brightness can be used to update their
neighbors brightness. This iterative smoothing will be stopped once some absolute value of the
brightness change in single iteration is smaller than pre-defined minimal threshold value or after
completing pre-selected number of iterations.
25

Figure 2.2 Flow chart of Chens Adaptive Smoothing Algorithm

The detailed steps are stated as following:
Step 1: Initialization Process.
Target noisy image is input and proper parameters are specified. Parameter S is usually
specified as S = 20. Parameter h which is selected from 0-1 and T the iteration number will be
also specified at beginning and their selection will be addressed in detail in Chapter 5.
Step 2: Compute Inhomogeneity map.
This step introduces how to generate an inhomogeneity map or map for the target noisy
image. The idea of inhomogeneity or contextual discontinuity is addressed in Section 2.2.1.
Basically it is a measurement for each pixel describing the image content in a relative larger area
centered by this pixel and this measurement is further applied in deciding the primary smoothing
26

speed at this pixel. (x, y) together with (x, y) derived from local discontinuity decide the
overall smoothing speed at pixel (x, y). Due to the high important role of inhomogneity, this step
is the most important step in Chens Algorithm and the most time consuming part.
As stated in Section 2.2.1, a homogeneity neighborhood is generated first for each pixel
in the noisy image. For each pair of adjacent neighborhoods centered by (x, y) and (i, j)
respectively the difference of the neighborhood [Nxy(Rxy;ij), Nij(Rxy;ij)] is defined as:

Rxy;ij is defined as the minimum value between Rxy and Rij. d(xy,vw) is defined as:

D
xy ;ij
+
and D
xy;ij
defined below are used together to offer a measure for detecting intra and iter-
object directional intensity variation.

27

The inhomogeneity value H(x, y) is further derived based on the defined difference
between neighborhoods . H(x, y) is defined as the average of the neighborhood differences
between HN of (x, y) and HNs of (x, y)s 8-neighbors. H(x, y) is defined as:

Bzy(1) represents pixel (x, y)s 8-neighbors and |Bzy(1)| is the number of its neighbors.
As shown in Figure 2.2, H(x, y) is further transformed to be (x, y) through a non-
increase function while local gradient is transformed to be (x, y). and at pixel (x, y) together
decide the smoothing speed at pixel (x, y).

Step 3: Iterative Smooth Process
Noisy image is smoothed through an iterative smooth procedure during which the
smoothing speed at each pixel is decided by (x, y) and (x, y) as following:

The termination of this iterative smoothing procedure is decided by pre-defined parameter T
which means the iterative steps executed.
28

2.4 Advantages and Deficiencies of this 2D Adaptive Smoothing Algorithm
2.4.1 Advantages of this 2D Adaptive Smoothing Algorithm
This section provides a brief summary of the advantages of Chens Algorithm
As introduced in the review section, Chens Algorithm is based on Fuzzy Connectedness
which is proved to be a strong mathematical tool to describe the relationship or the closeness
between image pixels. The Fuzzy Connectedness theory is first introduced into image processing
by Saha and Udupa [15, 16] to describe the possibility that two pixels belong to one single object.
The tool was proved very useful in image segmentation algorithms. Chen further develop the
usage of this tool to describe if the belonging objects changed for two adjacent objects and use
this change to decide if one pixel is close to a boundary which should be preserved or not [20].
Basically these two usages are similar and thus it is reasonable to expect good performance of
Chens Algorithm.
Another advantage in Chens Algorithm is brought by combining the two discontinuity
measurements together to control the iterative smoothing process. Unless the regular adaptive
smoothing algorithms Chens Algorithm uses contextual discontinuity which is very robust to
predict the image feature and indicate the potential edge pixel. At the same time the local
discontinuity is not abandoned but used to control the pixels used is brightness update. The
combination usage of these two measurements assures the convergence of the iterative smoothing
and makes the algorithm insensitive to the termination time [20].
Chens Algorithm also created a good framework for the adaptive smoothing algorithm.
Essentially Chens Algorithm can be viewed as a high level structure that a new smoothing
algorithm can be created by substituting the contextual discontinuity map as well as local
discontinuity map with any similar roadmap indicating the image features or image edge pixels.
From this respective Chens Algorithm is very important and worth to be carefully reviewed.
29

These three advantages are further verified by comparison experiment results presented
in [20]. Also Chapter 4 provided more experiment results in comparing Chens Algorithm with
some regular smoothing algorithms.

2.4.2 Deficiencies of this 2D Adaptive Smoothing Algorithm
Despite the advantages of Chens Algorithms there are also deficiencies or errors in both
HN definition and the design of contextual discontinuity which may lead into errors in smoothing.
Section 2.3.2.1 and 2.3.2.2 will discuss the errors in HN definition and in contextual discontinuity
design separately.

2.4.2.1 Error in Homogeneity Neighborhood Definition
For the simplicity HN is defined as an isotropic, square shape neighborhood around the
pixel in [20]. This isotropic neighborhood assumption is very sensitive to noise and not suitable
for pixels either near edge nor corrupted by serious noise. Chens Algorithm may generate
smaller or even incorrect HN for those pixels and result in inaccurate contextual discontinuity
map which is calculated based on HN result.
Basically the error using isotropic HN definition is that there is no isotropic homogeneity
neighborhood existing in the real world and its an assumption for simplicity. In most of cases
HN doesnt have an isotropic shape. Under HN definition in Chens Algorithm the size of HN
will yield to the minimum size of all sides.
Take the pixels near the boundary as an example to illustrate this problem. For pixels
near boundaries, their smallest isotropic neighborhood (3 by 3 neighborhood) may cross different
image regions such that strictly say they do not have any isotropic homogeneity neighborhoods.
This assumption is also invalid for other pixels when the image is corrupted by serious noise. In
30

such situations, current H(x, y) computation algorithm will assume 3 by 3 neighborhood as the
default homogeneity neighborhood. This 3 by 3 neighborhood is too small and it could be nnot
even homogeneity at all. Both these two situations actually violate the fundamental requirement
of H(x, y). This disadvantage of the HN used in Chens Algorithm will result in wrong
contextual discontinuity calculation which is stated in the following paragraph and finally
influence the smoothing result.
The H(x, y) in Chens Algorithm will use the small HN which could possibly be non-
homogeneity neighborhood to compute H(x, y) in some locations which directly violates the
theoretical fundamental of H(x, y) and causes inaccuracy H(x, y) value. In the basic theory, a
homogeneity region which can represent the local region property is selected to be compared with
the nearby homogeneity region to detect if this small region locates in a flat area or not. However,
if this homogeneity neighborhood shrinks to be its 3 by 3 neighbors, H(x, y) value cant pass any
information about the region but just local derivatives which is easily influenced by noise. Thus
the smoothing algorithm loses its most important power of using contextual discontinuity robust
to noise. Similarly, H(x, y) computed based on non-homogeneity neighborhoods can not reflect
the real situation either. First, differences between non-homogeneity neighborhoods may not
correctly reflect the area change. Assume 3 by 3 neighbors of pixel (x, y) in the flat area are
corrupted by noise. H(x, y) will be calculated based on 3 by 3 non-homogeneity neighborhood of
pixel (x, y) and its nearest pixels neighborhood. Since noise the H(x, y) will be wrongly
calculated high. The reason causing this is non-homogeneity neighborhood may include wrong
information into its consideration. Second, the non-homogeneity neighborhoods for pixels near
edges may cross the edges. Those neighborhoods will wrongly decrease H(x, y) in the following
computation. Besides, the most powerful part for this algorithm to smooth noise while preserving
edges is it employs statistical concepts to avoid noise influence in H(x, y) computation. However,
in the current algorithm the neighborhoods for pixels near edges and pixels serious corrupted by
31

noise, which are most concerned, can only be very small. Thus H(x, y) can be very sensitive to
noise.

2.4.2.2 Error in Contextual Discontinuity Design
Another deficiency of Chens Algorithm is no directional information is considered in
H(x, y) calculation. Chens Algorithm assumes the HN differences in all 8 directions are
equivalent the same in H(x, y) calculation. However it is reasonable that the HN differences in
different directions should be considered with the direction information instead of taking simple
average. The following example illustrates this problem.
Figure 2.3A and 2.3B illustrate two different situations of pixel *. The different
brightness values are marked with white and gray. Figure 2.3B simulates a pixel * is located in
an area corrupted by noise while the Figure 2.3A simulated a pixel * located close to the image
boundary. It is expected that H(*) in Figure 2.3A much bigger than Figure 2.3B to represent the
content change in * pixels surrounding. However based on Chens Algorithm these two Hs are
similar. The primary reason causing this is there is no direction information considered.
Contextual discontinuity calculation doesnt consider the HN change in directions close to each
other which could indicates existing of an image boundary. Thus the deficiency in failure of
considering direction information could lead to bad contextual discontinuity calculation result and
further influence the smoothing performance.
32

Figure 2.3A Pixel near image boundary

Figure 2.3B Pixel in smooth area corrupted by image noise

2.5 Conclusion
Based on our above discussion, we have the following conclusions:
Chens 2D Adaptive Smoothing Algorithm is advanced comparing to regular adaptive
smoothing algorithms. There are several deficiencies in Chens Algorithm which could be further
33

improved. Chens Algorithm provides a good framework which can be used to adopt in
developing improved smoothing Algorithm.

34

Chapter 3
Research Problem, Objective

3.1 Research Problem
From the result of our review and our study on the most recent Chens Algorithm which
has the best performance so far, there are still some important problems needed to be addressed in
this area though much of the previous research has been done in this area. Two most important
research problems which will be addressed here in this thesis are stated below.
The first research problem is Chens Adaptive Smoothing Algorithm should be improved
by correcting the errors and deficiencies revealed in Chapter 2 to provide an improved smoothing
algorithm with better performance of noise removal and image feature preservation. By applying
fuzzy connectedness theory and using combination of two discontinuity measurements in image
smoothing task Chens Algorithm provides a very nice smoothing result at the same time
preserves the important image features. However there are still many deficiencies and errors in
35

Chens Algorithm which prevent it from better performance especially on real images. Most of
these deficiencies/errors are caused by inaccurate assumptions and insufficient information usage.
It is expected that the performance of image smoothing with feature preservation could be further
improved by providing more accurate definitions and designs considering more possible
information.
Another problem preventing the practical use of Chens Algorithm is the selection of
parameter. The parameter used in Chens Algorithm is very sensitive to different image contents,
iteration termination number especially for the real images smoothing task. Without solving the
parameter selection problem Chens Algorithm cant be implemented to automatically solve
image smoothing tasks. Without automation Chens Algorithm will lose its most strength. Based
on this we list parameter selection as one important research problem for our improved smoothing
algorithm.

3.2 Research Objective
Our major objective is to provide an improved adaptive smoothing algorithm with better
noise removal and image feature preservation together with a practical way of parameter selection
in order to enable the practical use of the improved smoothing algorithm. Based on the research
problems proposed to solve we have the detailed itemized research objectives as following:

1. Our first research objective is to provide an improved adaptive smoothing algorithm
to provide better noise smoothing result with preservation of image features.

2. Provide a practical parameter selection method which can direct the use of our
algorithm in real image smoothing tasks and provide acceptable good result.
36

3. To demonstrate the improvement of our algorithms and the parameter selection
method a set of complete comparison and practical experiment plan should be
designed and executed. The experiment results can also be used as a reference of the
practical implementation.

37

Chapter 4
Improved 2D Adaptive Smoothing
Algorithm

4.1 Overview
As we stated in the previous Chapter, the primary objective of this thesis is to improve
Chens Algorithm to provide better noise smoothing result with preservation of image features by
correcting the errors occurring in Chens Algorithm. To achieve this objective, the major
problem to solve is to correct the errors and deficiencies revealed in Chens Algorithm and
provide improved smoothing algorithm.
For the first research problem presented, we believe a good H map depends on a correct
HN and correct HN depends on having a correct shape of homogeneity neighborhood which is
not a square in general. Our research starts from correcting the wrong the wrong assumption
made in defining homogeneity neighborhood in Chens Algorithm, considering more information
38

in estimating H map to provide more accurate edge estimation. By providing the improvements
in these two areas we believe an improved adaptive smoothing algorithm can be obtained.
In this Chapter, section 4.2 presents the proposed improvements to Chens Algorithm.
Section 4.3 introduces the resulting improved smoothing algorithm. Section 4.4 analyzes the
advantages and disadvantages of this improved smoothing algorithm. Section 4.5 provides some
implementation details.

4.2 Proposed Improvements
4.2.1 Generation of a Correct Homogeneity Neighborhood
A good H map is critical in deciding performance of Chens Algorithm and it depends on
a correct HN with a correct homogeneity neighborhood. In general the shape of HN would not be
square. The assumption of isotropic HN is made for simplicity but not reflecting the situation of
the real image. Take pixels close to the image boundary as an example. The image content
changes sharply in the direction towards boundary while the image could remain smooth along
other directions. It is hard to assume an isotropic or square shape HN existing for those pixels.
However Chens Algorithm assumes each pixel has such a square shape HN. With this
assumption either size of HN will be limited to be very small (determined by the minimum length
of HN among all direction) or HN could violate the homogeneity requirement in some directions.
From this aspect, Chens Algorithm has deficiency in its fundamental definition which brings bad
smoothing actions.
A correct neighborhood has the following properties: it takes the shape of homogeneous
neighborhood and is not necessarily a square one, meets a selected smoothness criterion, and
finally, has the largest neighborhood area (larger than 3 by3). The larger is the HN size the better
(more accurate) is the statistics on the intensity calculations.

39

HN is the neighborhood containing the pixels which belong to the same image object
with its center pixel. Each HN represents the local image content and are the further combined to
compare the difference to determine the image smoothness in a relative larger area. So it is
critical that the HN must be homogeneity or in the other way only contains pixels with the similar
property. The reason to use HN instead of single pixel in comparison is to avoid noise influence.
From the statistic aspect, when the homogeneity is guaranteed, the larger the size of HN is the
better the estimate of the smoothness is derived. So the basic requirements for good HN are: 1).
Homogeneity; 2). Size should be large enough.
Since the primary problem preventing HN in Chens Algorithm meeting the above
requirements is that the isotropic HN assumption does not exist for many pixels, we try to find an
anisotropic neighborhood for those pixels. In this proposed idea, each pixel is assumed to have a
local homogeneity neighborhood region and there are no limits on the neighborhood region shape.
This anisotropic homogeneity region can be used as the new HN. To locate this anisotropic HN,
complex mathematic tool like deformable contour algorithms can be applied. These tools can
adapt their shape to the image content while the size and speed can be controlled by setting
proper parameters. Later these new anisotropic HNs can be further processed to generate H map.
Here, we propose the following improvement using level set method to define HN which
meets the above HN requirements.
Level Set method is one of the most popular deformable contour algorithms [28, 29, 31,
32]. It is an iterative algorithm and grows by steps until reaching the high contrast image
boundaries. Level Set algorithm can produce different shapes of boundaries while the grow
speed and sensitivity to image contrast are controlled by preset parameters. These features
provide us convenience to find anisotropic HN by using Level Set algorithm.
The level set method is used to search for the HN since the method is an iterative process
derived from a selected smoothing criterion and guarantees a connected region from an initial
starting local region around a pixel. The smoothing criterion is embedded in the deforming speed
40

function in deforming the HN. The speed function used in our implementation is the commonly
adopted inverse of a gradient function. The initial contour is placed around the interested pixel
and consists of its 4 neighbors. To reduce the computation time, the maximum number of
iterations is set to be 5, and the maximum size of HN is set to be 150 pixels. A morphological
erosion operation is applied to the resulting close region to produce a smoother HN. H of each
pixel is the average brightness of the pixels in its HN. Obviously with this modification, the
problem of having 3 by 3 HN in Chens Algorithm is overcome.
In this proposed improvement, the homogeneity of HN can be guaranteed by set the step
of the Level Set which controls the tolerance of the non-smoothness. The size of the HN is
controlled by setting the maximum number of iteration. Under the normal situation the Level Set
boundary can grow at least in some directions and this can assure the resulting HN is larger than 3
by 3.
Figure 4.1 is a demonstration of HN formed using this Level Set HN definition. Blue
boundary and red boundary represent the anisotropic HN for blue pixel and red pixel respectively.

Figure 4.1.An illustration of two HN of two pixels close to edge

41

4.2.2 Incorporation of Direction Information in H Design
The H map is derived from the HN map to extract the inhomogeneity (context
discontinuity) of a pixel and its 8 neighboring pixels. As stated earlier, it reflects whether a pixel
is adjacent to an edge element or not.
The basic idea behind this is the difference between HN of one pixel and its adjacent
pixel can reflect the image content change along this direction. In Chens Algorithm H(x, y)
which represents the roughness of its neighborhood is derived by taking the average of the HN
differences between its HN and its 8-neighbors HNs. The directional information is not
considered in this average calculation. However the directional information is an important
knowledge needed in the adaptive method to identify the neighboring edge pixels as illustrated in
Figure 4.1. For the pixel close to the boundary, the image content change caused by the existence
of image boundary should happen in several adjacent directions. In the other way similar image
content change in several adjacent directions strongly indicates possible existence of image
boundary while the image content change caused by noise should be random and is less possible
to happen in adjacent directions. In this aspect, the direction information is important and should
be considered in H map deriv.
Here we provide an improvement of H map calculation with incorporation the directional
information.
A large average intensity difference between two adjacent pixels HNs in a particular
direction is an indication of either the presence of a large noise or an adjacent edge segment. To
distinguish these two situations, a threshold value is used to compare with these directional
differences. If three or more adjacent pixels differences exceed this threshold, then the present of
an edge segment is declared and smoothing weights are change accordingly. To accomplish this,
HN(i) is computed as the average intensity difference of the center pixels (*) HN and the HN
of its ith adjacent pixel with i being from 1 to 8.
42

The HN(i) indicates the directional intensity difference, and referring back to Figure 4.1
illustration of having an edge as shown, the HN(i) in the diagonal direction will be larger than
all other direction. The weight i will be set larger to emphasize the directional discontinuity.
The detail steps to decide the weight i are: 1). All i (i=1~8) is set to be 1/8. 2)
Differences between center pixels HN and its 8-neighbors HN are calculated and only the
directions on which the differences are larger than the threshold are recorded. 3) If the number of
the recorded directions is equal or bigger than three and these directions are adjacent, the weight
i for these directions are doubled. The other weighs are adjusted by taking average of (1-2*
1/8*number of the directions recorded).
There are two major advantages brought by considering direction information in H map
calculation. First it can strengthen the image edge feature. For the pixels close to boundary area,
comparing to the original H map calculation, the weights are adjusted larger on the directions
towards the boundary and results in larger H value. Thus the original brightness of these pixels
can be better preserved such that image edge feature is better preserved. While the image feature
is better preserved larger H value caused by noise is not enlarged since the weighs are not
adjusted assuming noise distribute in several random directions instead of in adjacent directions.
Thus the noise removal performance is kept.

4.3 Resulting Improved Algorithm
All the improvements are concentrated in producing an improved H map so that a better
smoothing result can be obtained with edge features being properly preserved. The rest of the
adaptive iterative smoothing algorithm stays the same as Chens original algorithm. Figure 4.2
. 8 ~ 1 ; 1 ; ) ( (*)

i i HN H
i i

43

shows the flow chart of Chens original algorithm. Our improvements are focused only in the red
block which is part generating the in-homogeneity (contextual discontinuity) or H map. Basically
we develop a new H map generation algorithm. Then the derived H map will used into Chens
original smoothing framework as shown in Figure 4.2.
The detail steps of the improvements on H map are stated above in section 4.2.
Comparing to the original Chens Algorithm, our improved algorithm is using an improved H
map generation algorithm while keeping the major frame work. There are two reasons why we
believe applying this improvement will help improved the smoothing performance with better
edge preservation. First, the in-homogeneity (contextual discontinuity) or H map is the key
measurement which decides the performance of smoothing and edge preservation to a large
extent and Chens Algorithm has deficiencies in this part. Second the overall smoothing
framework of Chens Algorithm is good especially the combining use of local discontinuity and
contextual discontinuity in deciding the smoothing speed.
Due to the adopting of the overall strategy of Chens Algorithm, our improved algorithm
has the similar issue in deciding parameters for practical use. This problem will be addressed in
Chapter 5.

44

Figure 4.2 Flow chart of Chens Algorithm

4.4 Advantages and Disadvantages of the Proposed Algorithm
The major advantage of this improved algorithm is it can provide better H map and
finally help in providing better smoothing performance with edge preservation. In saying better
we mean higher H value for pixels near the image edge which are the pixels we want to preserve
and higher H value assure less smoothing action applied. At the same time the improved H
derive algorithm wont increase the H value for those pixels located in smooth but noised area
thanks to the bigger HN used.
In the other side, the major disadvantage for this improved algorithm is the computation
burden. The primary reason causing the computation burden is the wide adopt of level set
method in locating HN. Though the grow of level set is limited to be 5 for in calculating HN for
each pixel the computation amount is huge considering the number of pixels grow in square of
the image size.
45

4.5 Implementation of the Improved Smoothing Algorithm
This improved smoothing algorithm is implemented in Matlab and C++.
In the consideration of computation time the most computation burden part, level set
algorithm, is implemented in C++ while the rest part is implemented in Matlab.

46

Chapter 5
Experimental Results

5.1 Overview
In this Chapter rich experiment results using Chens Algorithm, our improved smoothing
algorithm as well as some other popular smoothing algorithms are presented in Section 5.2. Then
we focus on experimental determination of parameter settings of h and T in Section 5.3.

5.2 Comparison Experiment Results
In this section we will first illustrate the advantage of adaptive smoothing algorithms over
some of the more common non-adaptive algorithms on a single simulated image in section 5.2.1.
Then section 5.2.2 presents comparisons of performances of our proposed algorithm and Chens
Algorithm based on SNR, DR(edge feature detection rate) and FAR(edge feature false alarm rate)
47

on both a simulated image and a set of 5 real images contaminated with various degrees of
additive noise. Section 5.2.3 presents another set of experiment on simulated human brain MRI
images.

5.2.1 Simulated Image Experiment Results
Simple simulated image is usually adopted in comparison test due to its simplicity and
clear boundaries. A simulated image as shown in Figure 5.1a with four different brightness
intensity values is selected as our first experiment target. The initial comparison test is to give a
simple illustration on the improvements provided by our improved 2D adaptive smoothing
algorithm comparing with Chens 2D adaptive smoothing algorithm.
In this initial experiment our improved adaptive smoothing algorithm and Chens
Algorithm are both applied to noise added simulated image. Both smoothed results are evaluated
using traditional SNR measurement and FAR\DR measurement which describes the image edge
preservation performance. Results are compared using both measurements and the comparisons
show that our improved smoothing algorithm is generally more effective than Chens Algorithm
in increasing SNR and edger preservation of an image.

Figure 5.1a Simulated test image
48

Table 5.1 shows the comparison result of our improved smoothing algorithm and Chens
Algorithm in SNR measurement. A different level of Gaussian noise specified in first row is
added to original simulated image in Figure 5.1a as test target for each column. The SNR result
for noisy image, smoothed results by Chens Algorithm and our improved algorithm are
presented in the following rows. SNR used here is defined as:

For both Chens Algorithm and our improved smoothing algorithm the combination of
parameters under which best SNR achieved is listed under the SNR figure.

Noise
Added
Std=10,mean=0 Std=20,mean=0 Std=30,mean=0 Std=40,mean=0 Std=50,mean=0
Noisy
Image SNR
168.9 49.5 24.1 14.6 9.8
Smoothed
Result
SNR(Chen)
434
(h=0.3,S=20)
328
(h=0.09,S=20)
248
(h=0.1,S=20)
193
(h=0.12,S=30)
162
(h=0.14,S=31)
Smoothed
Result
SNR(me)
447
(h=0.1,S=20)
341
(h=0.11,S=20)
256
(h=0.11,S=20)
201
(h=0.13, S=20)
157
(h=0.1,S=20)

Table 5.1 Smoothing results for Chens Algorithm and my algorithm (SNR)
49

The comparison shows our improved smoothing algorithm generally has better
performance than Chens Algorithm while both algorithms can increase SNR sharply from the
noisy image. Figure 5.1b shows a comparison of smoothed results. One thing needs to be
pointed out is SNR is not a good measurement of image feature preservation but a measurement
in signal processing area. In the other way, the image feature preservation may not be as good as
reflected by SNR. Figure 5.1b shows this point too.
In order to effectively compare the image feature preservation performances, we
developed another measurement.

Figure 5.1b Smoothing results comparison
Lower the original noisy image, upper left smoothed result by Chens Algorithm and
upper right is the smoothed result by our smoothed algorithm.
50

The similar comparison experiments as Table 5.1 are presented using different
performance measurement DR/FAR. DR/FAR measurement is designed to reflect how image
edges are recovered from noisy image by applying smoothing algorithm. FAR and DR is defined
as following:
For DR\FAR definition, we have devised the following procedure:

i) Use the Sobel edge operator with a threshold to extract all edge pixels from an
image.
ii) Use the resulting edge pixels from an original image as the ground truth.
iii) Calculate detection rate (DR) and false-alarm rate (FAR) as measures of the edge
preservation capability.

DR = (Total number of edge pixels matched between the smoothed image and the
original image) / (Total number of edge pixels of the original image)
FAR = (Total number of edge pixels found in the smoothed image and not in the original
image) / (Total number of edge pixels of the original image).
While DR reflects how many edge pixels are recovered the FAR reflects the noise
existence in the smoothed result. It makes no sense to just use one measurement to evaluate the
smoothed result. Take an example for the noisy image it has a very high DR but at the same time
the FAR is extreme high too which makes high DR meaningless. Similar to signal to noise ratio
only considering DR together with FAR can evaluate the image edge preservation performance.

Table 5.2 shows comparison smoothed results using DR/FAR measurement.

51

Noise
Added
and SNR
Result
Test 1

Test 2 Test 3 Test 4 Test 5
Noise
Added
Std=10,mean=0 Std=20,mean=0 Std=30,mean=0 Std=40,mean=0 Std=50,mean=0
Noisy
Image
DR=99.7%
FAR=9.2
DR=99.7%
FAR=19.5
DR=99.1%
FAR=20.9
DR=98.5%
FAR=21.0
DR=98.4%
FAR=21.7
Smoothed
Result
(Chen)
DR=99.5%
FAR=0.02
DR=99%
FAR=0.1
DR=98.9%
FAR=0.11
DR=98.7%
FAR=0.17
DR=98.4%
FAR=0.23
Smoothed
Result
(me)
DR=99.6%
FAR=0.02
DR=99.2%
FAR=0.07
DR=99%
FAR=0.08
DR=98.9%
FAR=0.1
DR=98.7%
FAR=0.14

Table 5.2 Smoothing results for Chens Algorithm and my algorithm (DR & FAR)

Table 5.2 further proved the superior of our improved algorithm to Chens Algorithm in image
edge preservation.
The performance comparison on 5 cases of different degrees of added noise is shown in
Table 5.3. The results show slight improvements in SNR when comparing our algorithm with
that of Chens with the exception in one case of very high additive noise. Most importantly, these
results indicate that the application of an adaptive smoothing algorithm significantly reduces the
FAR while preserving the DR of the extracted contour in the smoothed image. Our algorithm
shows further improvements of having lower FARs when compares to that of Chens Algorithm.
52

This simulated image may not be the best image to compare performances using SNR of
two algorithms although the exact edge pixels are known. A reason is that the total number of
edge pixels is very small in general compared to the total size of the image, and also the image
consists of a number of large but flat brightness regions and thus reducing the significance of the
SNR figure. While not shown in this writing due to space limitation, our experience shows that a
high SNR in many cases of using simulated images does show visibly significant improvement in
image quality.

5.2.2 Real Image Experiment Results

Since the simulated image only contains limited brightness value with clear image edges,
it is quite different with actual real image. To more effectively evaluate one image smoothing
algorithm, it is necessary to examine its performance on real images. To the end, simulated
image test is to provide a quick idea on the performance while improving the smoothing
performance on the real noisy image is the final target.
In this section a set of comparison experiments of our improved adaptive smoothing
algorithm and Chens Algorithm are presented to show the strength of our improved smoothing
algorithm in noise removal and image edge preservation.
The 5 images used in our experiment are selected to reflect a broader class of images so
as to provide a more critical comparison of performance. These images are carefully selected to
reflect different image data collection situations from an object with smooth background image to
a highly texture image, from a single person image to a very dense crowd, and from an in-door
image to an out-door scene. It is expected that the experimental results derived from these
images can provide a preliminary and practical conclusion on the applicability of the two
algorithms.
53

The 5 selected images are showed in Figure 5.2. Here is a brief explanation on why each
image is selected here. Figure 5.2.a Lena is a popular image adopted extensively in image
processing area due to this image contains very detailed image feature like hairs and brightness
changing in eyes as well as clear object and image boundary like the background. Figure 5.2.b
and Figure 5.2.c is selected due to the similar reason combinations of detailed foreground and
simple background. Figure 5.2.d is a good represent of image with extensive image details. It is
a very big challenge for image smoothing algorithm since small unclear image features are very
difficult to be separated from noise. Figure 5.2.e is an example of images with sharp brightness
change and shadow.

The actual size for these 5 selected images is 512 by 512.

Figure 5.2 Noise Free Images: a) Lena; b) Cameraman; c) Scene; d) Human face; e) Peppers
In most of the following real image tests two different levels of Gaussian noise are added
to the noise free selected images as the tested noisy images. Additive Gaussian noise of STD of
10 and 20 are considered like most others reported in the literature.
a b c
d e
54

In this real image experiment our improved adaptive smoothing algorithm and Chens
Algorithm are both applied to noise added simulated image. Both smoothed results are evaluated
using traditional SNR measurement and FAR\DR measurement. The two measurements used are
the same as defined in section 5.2.1.
The ground truth edge pixels are extracted using the Sobel operator with a threshold on
the original image. Note that different threshold settings give different total number of edge
pixels. Our experiments have indicated that the comparison results of performances are
consistence using different threshold values. Different threshold values with different DR\FAR
results for each image are presented to provide a more complete result in the following real image
tests.
Noted that in our DR/FAR measurement definition Sobel operator is used to extract edge
pixels. Theoretically Sobel operator preserves all high frequency texture edge features including
edge pixels and noise pixels. However in the aspect of image feature extraction, small & isolated
noise existence can be easily removed by morphology algorithms. So in order to demonstrate the
image edge preservation result, another further set of experiment result with smaller region
removal is presented. The step is after Sobel threshold result foreground pixels are further
examined by calculating the number of pixels of the 8-neighbors connected area it located in. If
one foreground pixel locates in an 8 neighbors connected area with the number of pixels lower
than 4, this pixel is considered as noise pixel and removed. Then the final result is applied to
normal DR\FAR measurement calculation.

1. Lena Image
The first experiment on Lena image (Figure 5.2.a) is to give: First, a simple illustration
that adaptive smoothing algorithms are generally more effective than some popular non-adaptive
algorithms in increasing SNR of an image; Second, our improved smoothing algorithm can
55

improve SNR better than Chens Algorithm. From Table 5.3, the SNR figures produced by two
adaptive algorithms, Chens Algorithm and the improved algorithm are much higher than those
from median filter, average filter and Wiener filter (another adaptive algorithm) under different
additive noise conditions while generally our improved algorithm can produce better SNR result
than Chens Algorithm.
Image 1 Image 2
Added
Gaussia
n Noise
Mean (Gray Level) 0 0
Variance (Gray
level)
10 5
Added
Salt and
pepper
Noise
density (percentage) 1% 5%
Noisy image SNR 21.4 17.5
Smooth
ed
Image
SNR
Median Filter 136 119
Average Filter 67 44
Wiener Filter 103 89
Chen Adaptive
Smoothing
197 178
Improved Adaptive
Smoothing
247 206

Table 5.3 Comparison of performances for Figure 5.2(a)
56

Figure 5.3 Left Noisy Lena Image with Gaussian noise added (mean=0, STD=10) Center
smoothed result with Chens Algorithm, Right Smoothed result by our improved algorithm.
In the following Table 5.4 only different levels of Gaussian Noise are added to generate
test noisy images. The SNR results show our improved algorithms superiority.
Image Name Lena(Fig. 5.2.a)
Test Setting Test 1
[0,10]*
Test2
[0,20]
Noisy
Image
SNR 125 32
Smoothed
Image
(Chen)
SNR 220 139
Smoothed
Image
(Our)
SNR 248 149
*: [0, 10]: Gaussian noise with 0 mean and std 10;
Table 5.4 Smoothing results comparison for Lena image (without small region removal
process)
57

From Table 5.5 to 5.8 the very same smoothed results as being evaluated in Table 5.4 are
reevaluated using DR/FAR measurement. To provide a complete test result different threshold
values (0.05, 0.1 and 0.15) and small region removal process are applied to generate different
comparison results.

Noise Added
and SNR Result
Test 1

Test 2
Noise Added Std=10,mean=0 Std=20,mean=0
Noisy Image DR=0.92
FAR=0.25
DR=0.89
FAR=0.34
Chens
Smoothed
Algorithm
DR=0.77
FAR=0.12
DR=0.73
FAR=0.14
Our Smoothed
Algorithm
DR=0.81
FAR=0.1
DR=0.79
FAR=0.12

value =0.15 & no small region removal process)

58

Noise Added
and SNR Result
Test 1

Test 2
Noisy Image DR=0.92
FAR=0.91
DR=0.9
FAR=0.72
Chens
Smoothed
Algorithm
DR=0.7
FAR=0.19
DR=0.69
FAR=0.21
Our Smoothed
Algorithm
DR=0.76
FAR=0.1
DR=0.72
FAR=0.12


59

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.93
FAR=1.7
DR=0.91
FAR=2.17
Chens
Smoothed
Algorithm
DR=0.68
FAR=0.11
DR=0.58
FAR=0.14
Our Smoothed
Algorithm
DR=0.79
FAR=0.09
DR=0.63
FAR=0.10


60

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.94
FAR=1.76
DR=0.93
FAR=1.9
Chens
Smoothed
Algorithm
DR=0.73
FAR=0.09
DR=0.6
FAR=0.11
Our Smoothed
Algorithm
DR=0.81
FAR=0.07
DR=0.67
FAR=0.08

value =0.05 & small region removal process)

From these comparison results, the general conclusion is our improved algorithm
provides a better DR/FAR performance over Chens Algorithm or in the other words the
smoothed result by improved algorithm has a higher DR value with a smaller FAR.

2. Cameraman Experiment Result
61

Two level Gaussian noises are added into original noise free Camera man image to
generate noisy test images. Both Chens Algorithm and our improved algorithm are applied. The
SNR results are presented in Table 5.9. Our improved algorithm can produce better SNR result
than Chens Algorithm.

Image Name Cameramen (Fig 5.2.b)
Test Setting Test 1
[0,10]
Test2
[0,20]
Noisy Image SNR 185 48
Smoothed
Image (Chen)
SNR 456 185
Smoothed
Image (Our)
SNR 462 199

Table 5.9 Smoothing results comparison for Camera man image (without small region removal
process)

The smoothed results are reevaluated by DR/FAR measurements in the following Table
5.10 and Table 5.11. For both evaluation results with or without small region removal process
62

smoothed result by our improved algorithm has higher DR and lower FAR values than result
produced by Chens Algorithm.

Noise Added
and SNR Result
Test 1

Test 2
Noisy Image DR=0.85
FAR=0.07
DR=0.83
FAR=0.65
Chens
Smoothed
Algorithm
DR=0.69
FAR=0.05
DR=0.58
FAR=0.13
Our Smoothed
Algorithm
DR=0.75
FAR=0.02
DR=0.62
FAR=0.03

Table 5.10 Smoothed results of Camera Man image for Chens Algorithm and my algorithm
(Threshold value =0.07 & no small region removal process)

63

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.93
FAR=0.17
DR=0.88
FAR=0.82
Chens
Smoothed
Algorithm
DR=0.8
FAR=0.11
DR=0.64
FAR=0.2
Our Smoothed
Algorithm
DR=0.86
FAR=0.06
DR=0.72
FAR=0.06

Table 5.11 Smoothed results of Camera Man image for Chens Algorithm and our algorithm
(Threshold value =0.07 & small region removal process)

3. Scene Image Experiment Result
Two level Gaussian noises are added into original noise free Scene image to generate noisy
test images. Both Chens Algorithm and our improved algorithm are applied. The SNR results
are presented in Table 5.9. Our improved algorithm can produce better SNR result than Chens
Algorithm.

64

Image Name Scene(Fig. 5.2.c)
Test Setting Test 1
[0,10]
Test2
[0,20]
Chens Smoothed
Algorithm
SNR 121 82
Our Smoothed
Algorithm
SNR 128 87

Table 5.12 Smoothing results comparison for Scene image (without small region removal process)


65

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.83
FAR=0.12
DR=0.84
FAR=0.74
Chens
Smoothed
Algorithm
DR=0.68
FAR=0.1
DR=0.5
FAR=0.08
Our
Smoothed
Algorithm
DR=0.7
FAR=0.04
DR=0.52
FAR=0.05


66

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.89
FAR=0.25
DR=0.86
FAR=1.04
Chens
Smoothed
Algorithm
DR=0.77
FAR=0.09
DR=0.56
FAR=0.12
Our Smoothed
Algorithm
DR=0.79
FAR=0.08
DR=0.58
FAR=0.1


4. Human Face Image Experiment Result
Two level Gaussian noises are added into original noise free Human face image to generate
noisy test images. Both Chens Algorithm and our improved algorithm are applied. The SNR
results are presented in Table 5.15. Our improved algorithm can produce better SNR result than
Chens Algorithm.
67

Image Name Human Face
(Fig. 5.2.d)
Test Setting Test 1
[0,10]
Test2
[0,20]
Chens
Smoothed
Algorithm
SNR 224 117
Our Smoothed
Algorithm
SNR 243 129

Table 5.15 Smoothing results comparison for Human Face image (without small region
removal process)


68

Noise Added
and SNR Result
Test 1

Test 2
Noisy Image DR=0.89
FAR=0.17
DR=0.84
FAR=0.24
Chens
Smoothed
Algorithm
DR=0.68
FAR=0.08
DR=0.63
FAR=0.1
Our Smoothed
Algorithm
DR=73%
FAR=0.03
DR=69%
FAR=0.04

(Threshold value =0.05 & no small region removal process)

69

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.93
FAR=0.17
DR=0.88
FAR=0.37
Chens
Smoothed
Algorithm
DR=0.75
FAR=0.1
DR=0.70
FAR=0.2
Our
Smoothed
Algorithm
DR=0.79
FAR=0.04
DR=0.74
FAR=0.09

(Threshold value =0.05 & small region removal process)

5. Pepper Image Result
Two level Gaussian noises are added into original noise free Pepper image to generate noisy
test images. Both Chens Algorithm and our improved algorithm are applied. The SNR results
are presented in Table 5.18. Our improved algorithm can produce better SNR result than Chens
Algorithm.
70

Image Name Peppers
(Fig. 5.2.e)
Test Setting Test 1
[0,10]
Test2
[0,20]
Chens Smoothed
Algorithm
SNR 443 267
Our Smoothed
Algorithm
SNR 454 281

Table 5.18 Smoothing results comparison for Pepper image (without small region
removal process)

The smoothed results are reevaluated by DR/FAR measurements in the following Table 5.19
and Table 5.20. For both evaluation results with or without small region removal process

71

Noise Added
and SNR Result
Test 1

Test 2
Noisy Image DR=0.95
FAR=0.9
DR=0.92
FAR=1.1
Chens
Smoothed
Algorithm
DR=0.81
FAR=0.12
DR=0.73
FAR=0.32
Our Smoothed
Algorithm
DR=0.86
FAR=0.06
DR=0.74
FAR=0.09


72

Noise Added
and SNR
Result
Test 1

Test 2
Noisy Image DR=0.96
FAR=0.81
DR=0.93
FAR=0.98
Chens
Smoothed
Algorithm
DR=0.84
FAR=0.1
DR=0.75
FAR=0.21
Our Smoothed
Algorithm
DR=0.88
FAR=0.05
DR=0.77
FAR=0.08


The overall SNR performance comparison on the five selected images is shown in Table
5.21. The overall DR/FAR (without small region removal) performance comparison on the five
selected images is shown in Table 5.22 while Table 5.23 shows the DR/FAR comparison result
with small region removal.
The 5 images are selected to reflect a broader class of images so as to provide a more
critical comparison of performance. These images are carefully selected to reflect different image
data collection situations from an object with smooth background image to a highly texture image,
73

from a single person image to a very dense crowd, and from an in-door image to an out-door
scene. It is expected that the experimental results derived from these images can provide a
preliminary and practical conclusion on the applicability of the two algorithms. Additive
Gaussian noise of STD of 10 and 20 are considered like most others reported in the literature.
The ground truth edge pixels are extracted using the Sobel operator with a threshold
on the original image. Note that different threshold settings give different total number of edge
pixels. Our experiments have indicated that the comparison results of performances are
consistence using different threshold values. Be noted that Sobel operator preserves high
frequency texture edge features. The results are presented in Table 5.21, 5.22 and 5.23. The
results show slight improvements on SNRs and significant improvements on FARs in all 5 cases
when comparing the two algorithms. While our algorithm provides lower DRs compared to that
of original images, they are consistently higher than using the Chens Algorithm. Also the DRs
are reduced by 10% to 20% due to the smoothing operation with the order of magnitude
improvements in the FARs which maybe significant to any further image segmentation operations.
Note that some high frequency texture pixels are also eliminated in the process in producing
lower DRs, and therefore, the DRs are even lower for more complicated or busy images like
Figures 5.2d and 5.2e. These Low DR numbers should not affect the true edge pixels for most
object segmentation applications since they are non-texture data.

74

Noise Added
Mean = 0
Std =
10 20 30 40 50
Noisy Image SNR=168.9
DR=99.7%
FAR=9.2
SNR=49.5
DR=99.7%
FAR=19.5
SNR=24.1
DR=99.1%
FAR=20.9
SNR=14.6
DR=98.5%
FAR=21.0
SNR=9.8
DR=98.4%
FAR=21.7
Smoothed
Result(Chen)
SNR=434
DR=99.5%
FAR=0.02
SNR=328
DR=99%
FAR=0.1
SNR=248
DR=98.9%
FAR=0.11
SNR=193
DR=98.7%
FAR=0.17
SNR=162
DR=98.4%
FAR=0.23
Smoothed
Result (Our)
SNR=447
DR=99.6%
FAR=0.02
SNR=341
DR=99.2%
FAR=0.07
SNR=256
DR=99%
FAR=0.08
SNR=201
DR=98.9%
FAR=0.1
SNR=157
DR=98.7%
FAR=0.14
Table 5.21 Smoothing results for Chens Algorithm and my algorithm (SNR)
Image Name Lena(Fig. 3a) Cameramen (Fig.3b) Scene(Fig. 3c) Human Face
(Fig. 3d)
Peppers
(Fig. 3e)
Test Setting Test 1
[0,10]*
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Noisy
Image
SNR 125 32 185 48 63 17 170 43 272 71
DR 0.93 0.91 0.85 0.83 0.83 0.84 0.89 0.84 0.95 0.92
FAR 1.7 2.17 0.07 0.65 0.12 0.74 0.17 0.24 0.0 1.1
Smoothed
Image
(Chen)
SNR 220 139 456 185 121 82 224 117 443 267
DR 0.68 0.58 0.69 0.58 0.68 0.5 0.68 0.63 0.81 0.73
FAR 0.11 0.14 0.05 0.13 0.1 0.08 0.08 0.1 0.12 0.32
Smoothed
Image (Our)
SNR 248 149 462 199 128 87 243 129 454 281
DR 0.79 0.63 0.75 0.62 0.7 0.52 0.73 0.69 0.86 0.74
FAR 0.09 0.10 0.02 0.03 0.04 0.05 0.03 0.04 0.06 0.09
Table 5.22 Smoothing results comparison for 5 real images (without small region
removal process)
75

Image Name Lena(a) Cameramen(b) Scene(c) Human Face(d) Peppers(e)
Test Setting Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Test 1
[0,10]
Test2
[0,20]
Noisy Image SNR 125 32 185 48 63 17 170 43 272 71
DR 0.94 0.93 0.93 0.88 0.89 0.86 0.93 0.88 0.96 0.93
FAR 1.76 1.9 0.17 0.82 0.25 1.04 0.17 0.37 0.81 0.98
Smoothed
Image
(Chen)
SNR 220 139 456 185 121 82 224 117 443 267
DR 0.73 0.6 0.8 0.64 0.77 0.56 0.75 0.70 0.84 0.75
FAR 0.09 0.11 0.11 0.2 0.09 0.12 0.1 0.2 0.1 0.21
Smoothed
Image
(Own)
SNR 248 149 462 199 128 87 243 129 454 281
DR 0.81 0.67 0.86 0.72 0.79 0.58 0.79 0.74 0.88 0.77
FAR 0.07 0.08 0.06 0.06 0.08 0.1 0.04 0.09 0.05 0.08

Table 5.23 Smoothing results comparison for 5 real images (with small region removal
process)

5.2.3 Simulated MRI Image Experiment Results

We also have the results on an MRI brain image with various degrees of added noise with
significant improvements in SNR.

As shown in Figure 5.4 we have a simulated MRI brain image. It can be considered as noise
free image. Four different levels of noise are added to this simulated image to generate noisy
image. Both Chens Algorithm and our improved algorithm are applied to smooth the noisy
image. The SNR experiment results are shown in Table 5.24. The parameter combination is
76

selected based on the best SNR ever obtained. It is shown that our improved algorithm provides
better SNR smoothed result.

Due to the complexity of the MRI brain image, it is hard to use simple edge detection
operator like Sobel filter to generate trustable image edges. So we do not have DR and FAR
figures in this case.

Figure 5.4 Left noisy Image with 5% noise added, right smoothed result by our improved smooth
algorithm

77

Noise Added and
SNR Result
Test 1 Test 2 Test 3 Test 4
Noise Added 3% 5% 7% 9%
Noisy Image SNR 178 73 43 24
Smoothed Result
SNR(Chen)
208
(h=0.3, S=20)
91
(h=0.28, S=20)
62
(h=0.29,S=20)
38
(h=0.26,S=20)
Smoothed Result
SNR(me)
222
(h=0.36, S=20)
97
(h=0.04, S=20)
66
(h=0.5, S=20)
43
(h=0.4, S=20)

Table 5.24 Smoothed results of MRI image for Chens Algorithm and my algorithm (SNR)

To further illustrate the strength of our improved smoothing algorithm, another set of
noisy image acquired under real situations are processed by our improved smoothing algorithm.
These images are acquired by actual MRI equipment and considered containing noises. Since we
do not have ground truth for those real images, we cant calculate smoothed results SNR or
DR/FAR. So the smoothed results can be only evaluated by observation.

78

Figure 5.5 MRI real brain image smoothing result: left original noisy image, center smoothing
result after 5 iterations, right smoothing result after 15 iterations

79

Figure 5.6 MRI real brain image smoothing result: left original noisy image, center smoothing
result after 5 iterations, right smoothing result after 15 iterations

80

\

Figure 5.7 MRI real brain image smoothing result: Upper left original noisy image, upper right
smoothing result after 5 iterations, lower left smoothing result after 15 iterations

81

Figure 5.8 MRI real brain image smoothing result: Upper left original noisy image, upper right
smoothing result after 5 iterations, lower left smoothing result after 15 iterations

5.3 Experimental Parameter Determination

To enhance the applicability of our proposed smoothing algorithm, the parameter
sensitivity must be deeply analyzed and an applicable parameter selection method must be
82

proposed. Thus the proposed smoothing algorithm can be effectively applied in a real image
analysis task. A typical example in an image analysis project is one or several noisy images are
available for smoothing purpose while the noise free images are not available. Questions need to
be answered before apply one smoothing algorithm requiring parameter specification are: How
we can choose parameter for this\these unknown noisy image? How we know if particular
parameter selection is good enough remembering we dont have noise free image such we cant
compare smoothing results using pre-defined measurements like SNR or DR/FAR?
This section emphasize on answering the above two questions using experimental
methods. Subsection 5.3.1 will focus on presenting the different parameter selections and
resulting smoothing performances on one simulated image and five selected real images. With
these experimental results Subsection 5.3.2 continues to propose an applicable parameter
selection method for real image noise removal and get further verified by experimental results on
the very same images.

5.3.1 Parameter Sensitivity Investigation
Since our improved algorithm adopts the same iterative smoothing procedure as Chens
Algorithm, important parameters whose selection would bring big influence on smoothing
performance is h and T (the total iteration number) [35]. Due to this reason the following
experiment results as well as most analysis are also valid for Chens Algorithm. These two
parameters will be briefly introduced and followed by a set of experiment results. During the
following experiment other non-important parameters like S will be fixed to particular value
S=20.
As stated in Chapter 2 parameter h is from 0-1 and used to control the transform format
from the inhomogeneity value H to the smoothing speed as shown in the following formula:
83

g(.) is a monotonically decreasing function, like:

Giving the same H value, larger h will increase the smoothing speed at particular pixel
(x, y). Thus very small h results in no smoothing action at all while big h close to 1 results image
feature removed.
Parameter T, an integer selected from1 to infinite, is the stop time for the iterative smoothing
procedure. Similar to h, small T results in little smoothing action while big T results in image
feature removal.
The following experiments are designed to investigate parameter sensitivity and stopping rule
of the improved smoothing algorithm. Test target images are the same as section 5.2 including
one simulated image and 5 selected real images. For each tested real image and simulated image,
a set of results are presented and analyzed. Experiments for each image include:
84

1) SNR performance vs. iterative smoothing procedure under different parameter h.
2) Best SNR obtained under particular parameter choice h.
3) Corresponding iterative number under which the best SNR is obtained.
4) 3D illustration of the relationship between parameter choice h, iteration number and SNR
performance.

Simulated image in Fig 5.1 is added with Gaussian noise (mean =0, STD = 50) and used
as the test noisy image. Fig. 5.9 shows the SNR performance along the iteration number &
different h values. Fig 5.10 shows the best SNR ever obtained under different h values. From
these two results there are no obvious SNR performance dropping as the iteration number
increase while h is chosen between 0.1 and 0.16. It also shows some extent SNR sensitivity to
parameter selection h. Figure 5.11 shows the iteration number where the best SNR is obtained
under specified h. It further illustrates the relationship between h value and iteration number: the
larger h the smaller iteration number. It is obvious since large h and big iteration number
basically are the same in increasing smoothing speed. Fig. 5.12 is an overall 3D illustration of
SNR performance under different selection combination of h and iteration number.

85

Fig 5.9 Simulated Image - SNR performance along with iteration number increase
(different h for different curves, h for curves from smooth curve to sharp curve
corresponding to h 0.05, 0.1, 0.16, 0.2, 0.3, 0.4)

0 100 200 300 400 500 600
0
20
40
60
80
100
120
140
160
h
S
N
R
86

Fig 5.10 Simulated Image - Best SNR obtained under different parameter h

Fig 5.11 Simulated Image - Iteration number corresponding to best SNR obtained under
different parameter
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
80
90
100
110
120
130
140
150
160
h
S
N
R
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0
50
100
150
200
250
300
350
400
450
500
h
B
e
s
t

S
t
o
p

I
n
e
r
a
t
i
o
n

N
u
m
b
e
r
87

Fig 5.12 Simulated Image - SNR performances under different h and Iteration number.
Lena image in Fig 5.2.a is added with Gaussian noise (mean =0, STD = 20) and used as
the test noisy image. Fig. 5.13 shows the SNR performance along the iteration number &
different h values. Fig 5.14 shows the best SNR ever obtained under different h values. SNR
performance drops as the iteration number increases while at the same time SNR is kind non
sensitive to h selection. It shows an important point that unlike simulated image which is simple
and would be stable after some iterations real image is complicate and the stopping iteration
number is important in preventing the image feature are removed. Figure 5.15 shows the iteration
number where the best SNR is obtained under specified h. As specified before it illustrates
relationship between h value and iteration number: the larger h the smaller iteration number. Fig.
5.16 is an overall 3D illustration of SNR performance under different selection combination of h
and iteration number.

0
100
200
300
400
500
600 0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0
20
40
60
80
100
120
140
160
h
Iteration Number
S
N
R
88

Fig 5.13 Lena Image - SNR performance along with iteration number increase (different
h for different curves, h for curves from smooth curve to sharp curve corresponding to h
0.03, 0.05, 0.08, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7)

0 20 40 60 80 100 120
20
40
60
80
100
120
140
160
Iteration Numbers
S
N
R
89

Fig 5.14 Lena Image - Best SNR obtained under different parameter h

Fig 5.15 Lena Image - Iteration number corresponding to best SNR obtained under
different parameter h
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
90
100
110
120
130
140
150
h
S
N
R
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
10
20
30
40
50
60
70
80
90
h
B
e
s
t

S
t
o
p

I
t
e
r
a
t
i
o
n

N
u
m
b
e
r
90

Fig 5.16 Lena Image - SNR performances under different h and Iteration number.
Camera man image in Fig 5.2.b is added with Gaussian noise (mean =0, STD = 20) and
used as the test noisy image. Fig. 5.17 shows the SNR performance along the iteration number &
different h values. Fig 5.18 shows the best SNR ever obtained under different h values. Unlike
simulated image and Lena image SNR performance drops as the iteration number increases while
at the same time SNR is also kind sensitive to h selection. Figure 5.19 shows the iteration
number where the best SNR is obtained under specified h. . As specified before it illustrates
relationship between h value and iteration number: the larger h the smaller iteration number. Fig.
5.20 is an overall 3D illustration of SNR performance under different selection combination of h
and iteration number.

0
20
40
60
80
100
120 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
20
40
60
80
100
120
140
160
h
Iteration Number
S
N
R
91

Fig 5.17 Camera man Image - SNR performance along with iteration number increase
(different h for different curves)

0 20 40 60 80 100 120
40
60
80
100
120
140
160
180
200
h
S
N
R
92

Fig 5.18 Camera man Image - Best SNR obtained under different parameter h

Fig 5.19 Camera man Image - Iteration number corresponding to best SNR obtained
under different parameter h
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
160
165
170
175
180
185
190
195
200
S
N
R
h
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
10
20
30
40
50
60
70
B
e
s
t

S
t
o
p

I
t
e
r
a
t
i
o
n

N
u
m
b
e
r
h
93

Fig 5.20 Camera man Image - SNR performances under different h and Iteration number.

Scene image in Fig 5.2.c is added with Gaussian noise (mean =0, STD = 20) and used as
different h values. Fig 5.22 shows the best SNR ever obtained under different h values. Like
Lena image SNR performance drops as the iteration number increases while at the same time
SNR is not sensitive to h selection. Figure 5.23 shows the iteration number where the best SNR
is obtained under specified h. . As specified before it illustrates relationship between h value and
iteration number. Fig. 5.24 is an overall 3D illustration of SNR performance under different
selection combination of h and iteration number.
0
50
100
150 0
0.1
0.2
0.3
0.4
0.5
0
50
100
150
200
h
Iteration Number
S
N
R
94

Fig 5.21 Scene Image - SNR performance along with iteration number increase (different
h for different curves)

Fig 5.22 Scene Image - Best SNR obtained under different parameter h
0 20 40 60 80 100 120
10
15
20
25
30
35
40
45
50
55
60
Iteration Number
S
N
R
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
36
38
40
42
44
46
48
50
52
54
56
h
S
N
R
95

Fig 5.23 Scene Image - Iteration number corresponding to best SNR obtained under

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0
10
20
30
40
50
60
B
e
s
t

S
t
o
p

I
t
e
r
a
t
i
o
n

N
u
m
b
e
r
h
96

Fig 5.24 Scene Image - SNR performances under different h and Iteration number.

Human face image in Fig 5.2.d is added with Gaussian noise (mean =0, STD = 20) and
used as the test noisy image. Fig. 5.25 shows the SNR performance along the iteration number &
different h values. Fig 5.26 shows the best SNR ever obtained under different h values. It is a
very complicate image and SNR performance drops as the iteration number increases while at the
same time SNR is also very sensitive to h selection. Figure 5.27 shows the iteration number
where the best SNR is obtained under specified h. . As specified before it illustrates relationship
between h value and iteration number. Fig. 5.28 is an overall 3D illustration of SNR performance
under different selection combination of h and iteration number.

0
20
40
60
80
100
120
0
0.1
0.2
0.3
0.4
0.5
10
20
30
40
50
60
h
Iteration Number
S
N
R
97

Fig 5.25 Human Face Image - SNR performance along with iteration number increase

0 20 40 60 80 100 120
0
20
40
60
80
100
120
140
Iteration Number
S
N
R
98

Fig 5.26 Human Face Image - Best SNR obtained under different parameter h

Fig 5.27 Human Face Image - Iteration number corresponding to best SNR obtained
under different parameter h
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
60
70
80
90
100
110
120
130
h
S
N
R
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
5
10
15
20
25
30
h
B
e
s
t

S
t
o
p

I
t
e
r
a
t
i
o
n

N
u
m
b
e
r
99

Fig 5.28 Human Face Image - SNR performances under different h and Iteration number

Pepper image in Fig 5.9.e is added with Gaussian noise (mean =0, STD = 20) and used as
different h values. Fig 5.30 shows the best SNR ever obtained under different h values. SNR
performance drops as the iteration number increases while at the same time SNR is not sensitive
to h selection. Figure 5.31 shows the iteration number where the best SNR is obtained under
specified h. . As specified before it illustrates relationship between h value and iteration number.
Fig. 5.32 is an overall 3D illustration of SNR performance under different selection combination
of h and iteration number.
0
20
40
60
80
100
120
0
0.1
0.2
0.3
0.4
0.5
20
30
40
50
60
70
80
90
100
110
120
h
Iteration Number
S
N
R
100

Fig 5.29 Pepper Image - SNR performance along with iteration number increase

Fig 5.30 Pepper Image - Best SNR obtained under different parameter h
0 20 40 60 80 100 120
0
50
100
150
200
250
300
Iteration Number
S
N
R
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
120
140
160
180
200
220
240
260
280
h
S
N
R
101

Fig 5.31 Pepper Image - Iteration number corresponding to best SNR obtained under

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
10
20
30
40
50
60
70
80
h
B
e
s
t

S
t
o
p

I
t
e
r
a
t
i
o
n

N
u
m
b
e
r
102

Fig 5.32 Pepper Image - SNR performances under different h and Iteration number

From the above experiments, we have the following several conclusions:
1. Iteration number T and h has a close relationship;
2. Sensitivity of proposed smoothing algorithm to iteration number T and h highly depends
on the image content complexity, especially the selection of T;
3. The smoothing performance is to some extent stable over a period of T and h selection.

5.3.2 Applicable Parameter Selection Algorithm
Based on the experiment results derived in section 5.3.1, an applicable parameter
selection algorithm is proposed here and corresponding experiments results are presented.

0
20
40
60
80
100
120
0
0.1
0.2
0.3
0.4
0.5
50
100
150
200
250
300
h
Iteration Number
S
N
R
103

The main purpose of the experiment is to enhance the applicability of our proposed smoothing
algorithm and Chens Algorithm. Both algorithms require the selections of two significant
parameter values h and T (the total iteration number). Experiments have shown the
resulting smoothed image is not sensitive to selection of S value as long as it stays within the
range between 10 and 40. If we can show experimentally that the smoothing results do not
change very much in SNRs, i.e., SNR figures are within a narrow band, for ranges of h and T will
improve the applicability of our proposed algorithm. Be noted that on real images, the SNRs drop
off rather rapidly after T iterations as illustrated in Figure 5.33. Our experimental approach is to
do an exhaustive search of h and T in obtaining the best SNR figures of all 5 images under the
contamination of an additive Gaussian noise.
For all 5 images in Figure 5.2, SNRs are computed for different values of h and T with
each incoming image being contaminated with Gaussian noise of standard deviation of 20. As an
illustration, for Figure 5.2(a), the 3D perspective plot of SNR versus h and T is shown in Figure
5.33.The figure shows that the SNR peaks for higher h values with smaller T values, and they
drop off rather rapidly for larger h and larger T. Tests on all other 4 images give similar 3D
results. For parameter value range selection results, we have Figure 5.34 and Figure 5.35. Figure
5.34 shows the plots of the T (number of iterations for the best SNR) versus the h. One can see
that T is less than 8 iterations in all 5 images for h ranging from 0.2 to 0.6. Figure 5.35 shows the
plot ratios of the obtained SNR versus best SNR figure (maximum of 1) for each image versus h
ranging from 0.3 to 0.5. There is only one best SNR for each image. While the resulting SNRs
are better than 80% of the best SNR, and they stay rather close range the SNR performances for
the whole range of h. Therefore, one can draw a conclusion that for these 5 images, and for T
ranges from 3 to 6 and h ranges from 0.3 to 0.5, our smoothing algorithm will provide SNRs of
over 80% of the best SNRs. After the computation of the H map, one needs only a very few
iteration to obtained a smoothed image in practice as shown in these 5 images. We also conduct
104

the same experiments using Chens Algorithm and very similar results for h and T are shown in
Figures 5.36 and 5.37.

Figure 5.33 A 3D performance illustration for Figure 5.2a
0
20
40
60
80
100
120 0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
20
40
60
80
100
120
140
160
h
Iteration Number
S
N
R
105

Figure 5.34 Best stopping iteration number versus h for all 5 images in Figure 5.2

Figure 5.35 Illustration of obtained SNR/best SNR versus h for all images in Figure 5.2

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
10
20
30
40
50
60
70
80
90
i
t
e
r
a
t
i
o
n

n
u
m
b
e
r
h
lena
pepper
humanface
cameraman
scene
0.3 0.32 0.34 0.36 0.38 0.4 0.42 0.44 0.46 0.48 0.5
0.75
0.8
0.85
0.9
0.95
1
h
O
b
t
a
i
n
e
d

S
N
R
/
B
e
s
t

S
N
R
lena
pepper
humanface
cameraman
scene
106

Figure 5.36 Best stopping iteration number versus h for all 5 images in Figure 5.2 using Chens
Algorithm

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
0
20
40
60
80
100
120
h
B
e
s
t

S
t
o
p

I
t
e
r
a
t
i
o
n

N
u
m
b
e
r

lena
pepper
humanface
cameraman
scene
107

Figure 5.37 Illustration of (obtained SNR/best SNR) versus h for all images in
Figure 5.2 using Chens Algorithm

5.4 Conclusion
Our modified algorithm shows slight improvements in SNR when compared to that of
Chens Algorithm as can be seen in Table 5.21 for a simulated image with different degrees of
added noise and in Tables 5.22 and 5.23 for 5 selected images shown in Figure 5.2 with different
degrees of added noise. These tables show also large improvements in preserving edge features
when comparing the same two algorithms. Furthermore, experimental results shown in Figures
5.33, 5.34 and 5.35 allow us to select the two key parameters of h ranging from 0.3 to 0.5 and T
ranging from 3 to 6 that provides comparable smoothing results (higher than 80% from the best
SNR) in all 5 images shown in Figure 5.2 for both our algorithm and Chens Algorithm. The best
0.3 0.32 0.34 0.36 0.38 0.4 0.42 0.44 0.46 0.48 0.5
0.75
0.8
0.85
0.9
0.95
1
h
O
b
t
a
i
n
e
d

S
N
R
/
B
e
s
t

S
N
R

lena
pepper
humanface
cameraman
scene
108

combination set for any specific application can visually be determined experimentally using a
small number of combinations of different h and T. For practical applications without the
knowledge of the noise free image and/or a viewer, we suggest the set of h=0.4 and T=4.

109

Chapter 6
Conclusions

In this thesis an improved adaptive smoothing algorithm is proposed and related
experiment results are presented. Our improved algorithm shows slight improvements in SNR
when compared to that of Chens Algorithm as can be seen in Table 5.21 for a simulated image
with different degrees of noise added and in Tables 5.22 and 5.23 for 5 selected images shown in
Figure 2.2 with different degrees of noise added. These tables show also large improvements in
preserving edge features when comparing the same two algorithms. Furthermore, experimental
results shown in Figures 5.33, 5.34 and 5.35 allow us to select the two key parameters of h
ranging from 0.3 to 0.5 and T ranging from 3 to 6 that provides comparable smoothing results
(higher than 80% from the best SNR) in all 5 images shown in Figure 4.1 for both our algorithm
and Chens Algorithm. The best combination set for any specific application can visually be
determined experimentally using a small number of combinations of different h and T. For
practical applications without the knowledge of the noise free image and a viewer, we suggest the
set of h=0.4 and T=4.
110

Bibliography
[1] R.C. Gonzalez and R.E. Woods, Digital Image Processing, second ed., Upper Saddle River,
N.J.: Prentice Hall, 2002.
[2] H. Peng, Modification and Implementation of a 2D Context-Sensitive Adaptive Smoothing
Algorithm and its Extension to 3D, master thesis in University of Cincinnati, 2006.
[3] P. Saint- Marc, J.S. Chen, G. Medioni, Adaptive Smoothing: A General Tool for Early Vision,
IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 13, no.6, P.514-529, June 1991.
[4] I. Scollar, B. Weidner, T.S. Huang, Image enhancement using the median and the
interquartile distance, Computer Vision, Graphics and Image Processing, Vol.25, pp.236--251,
1984.
[5] R. H. Vu, Fuzzy algorithms: Application to adipose tissue quantification on MR images,
Biomedical Signal Processing and Control, Volume 2, Issue 3, P. 239-247, July 2007.
[6] L.S. Davis and A. Rosenfeld, Noise cleaning by iterated local averaging, IEEE
Trans. on Systems, Man and Cybernetics, SMC-8(9) pp.705--710, 1978.
[7] S.Luo, J.Han, Filtering Medical Image Using Adaptive Filter, Proceedings of the
23rd Annual EMBS International Conferences, pp. 2727-2729, 2001.
[8] D.R.K. Brownrigg, The weighted median filter, Commun. ACM, Vol. 27, pp.807--818, 1984
[9] Jerry L. Prince, Jonathan Links, Medical Imaging Signals and Systems, Prentice Hall, Apr,
2005.
[10] K. Zhang, 3D Medical Image Applications of A 2D Constrained Deformable Contour
Method, master thesis in University of Cincinnati, 2004.
[11] D. Tsai, N. Tajagasgu, Y. Lee, An Adaptive Enhancement Algorithm for CT Brain
Images, Proceedings of the 2005 IEEE, Engineering in Medicine and Biology 27th
111

Annual Conference, pp. 3398-3401, Shanghai, China, 2005.
[12] S.M.Smith and J.M.Brady, SUSAN- A New Approach to Low Level Image
Processing, Int. J. Computer Vision Vol. 23, pp.45-78,1997
[13] B. Migeon, Serfaty, V. Gorkani, M. Marche, An adaptive smoothing filter
for URTURlP images, Engineering in Medicine and Biology Magazine, IEEE,vol 14,
Nov/Dec, pp.762-765,1995
[14] P.K. Saha and J.K. Udupa, Scale-Based Diffusive Image Filtering Preserving Boundary
Sharpness and Fine Structures, IEEE Trans. Medical Imaging, vol. 20, pp. 1140-1155, 2001.
[15] P.K. Saha, J.K. Udupa, and D. Odhner, Scale-Based Fuzzy Connected Image Segmentation:
Theory, Algorithms, and Validation, Computer Vision and Image Understanding, vol. 77, pp.
145-174, 2000.
[16] P. K. Saha and J. K. Udupa, Scale-based filtering of medical images, Proc. SPIE: Medical
Imaging, vol. 3979, pp. 735746, 2000.
[17] P.K. Saha and J.K. Udupa, Optimum Image Thresholding via Class Uncertainty and
Region Homogeneity, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 23, pp. 689-
706, 2001.
[18] A.S. Pednekar, I.A. Kakadiaris, Image Segmentation Based on Fuzzy Connectedness Using
Dynamic Weights, IEEE Transaction on Image Processing, vol.15, no. 6, pp.1555-1562, 2006.
[19] K. Chen, A feature preserving adaptive smoothing method for early vision, Journal of
Pattern Recognition Society, 13, 2000
[20] K. Chen, Adaptive smoothing via Contextual and Local Discontinuities, IEEE Trans. On
Pattern Analysis and Machine Intelligence, Vol. 27, Issus 10, pp. 1552-1567, 2005.
[21] J. Weickert, Theoretical Foundations of Anisotropic Diffusion in Image Processing,
Computing, vol. 11, pp. 221-236, 1996.
[22] P. Perona and J. Malik, Scale-Space and Edge Detection Using Anisotropic Diffusion,
IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 12, pp. 629-639, 1990.
112

[23] M.J. Black, G. Sapiro, D.H. Marimont, and D. Heeger, Robust Anisotropic Diffusion, IEEE
Trans. Image Processing, vol. 7, pp. 421-432, 1998.
[24] G.T. Herman and B.M. Carvalho, Multiseeded segmentation using fuzzy connectedness,
IEEE Trans. on Pattern Anal. Mach. Intell.23, pp. 460474, 2001.
[25] I.Bloch, Fuzzy spatial relationships for image processing and interpretation: a review,
Image and Vision Computing, Volume 23, Issue 2, P.89-110, February 2005.
[26] G. Gilboa, N. Sochen, Y. Zeevi, Forward-and-backward diffusion processes for adaptive
image enhancement and denoising, IEEE Transactions on Image Processing, Vol. 11, Issue 7,
pp. 689-703, 2002.
[27] G. Gilboa, Y. Y. Zeevi, and N. Sochen, Resolution enhancement by forward-and-backward
nonlinear diffusion processes, Proc. Nonlinear Signal and Image Processing Baltimore, June
2001.
[28] J.A. Sethian, Level Set Methods and Fast Marching Methods: Evolving I nterfaces in
Computational Geometry, Fluid Mechanics, Computer Vision and Materials Science, Cambridge
University Press, 1999.
[29] Y. Hou, Application of a 3D Level Set Method in MRI Surface Segmentation, master thesis
in University of Cincinnati, 2005.
[30] J. Weickert, A Review of Nonlinear Diffusion Filtering, Scale-Space Theory in Computer
Vision, P. 3-28, Berlin: Springer, 1997.
[31] T. McInerney, D. Terzopoulos, Deformable models in medical image analysis:
a survey, Medical Image Analasis, P.1:91108, 1996.
[32] M. Kass, A. Witkin, and D. Terzopoulos, Snakes: active contour models, Intl J. Computer
Vision, vol. 1, no. 4, P. 321331, 1987.
[33] K. Siddiqi, Y. B. Lauzi`ere, A. Tannenbaum and S. W. Zucker, Area and length
minimizing flows for shape segmentation, IEEE Trans. Image Processing, vol. 7, P. 433 443,
1998.
113

[34] L. Alvarez, P.L. Lions, and J.M. Morel, Image Selective Smoothing and Edge Detection by
Nonlinear Diffusion I I , SIAM J. Numerical Analysis, vol. 29, P. 845-866, 1992.
[35] S.K. Mitra and G.L. Sicuranza, eds. Nonlinear Image Processing, San Diego, Calif.:
Academic Press, 2001.
[36] J. Polzehl, V. G. Spokoiny, Adaptive weights smoothing with applications to image
restoration, Journal of the Royal Statistical Society: Series B , Vol. 62, Issue 2, P. 335-354, 2002.

PDF Ieeee

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PDF Ieeee

Uploaded by

Copyright:

Available Formats

U

You might also like