MICROW A VE ELECTRONICS

STRATEGIES FOR INVERSE PROFILING OF

TWO DIMENSIONAL
DIELECTRIC SCATTERERS USING
ELECTROMAGNETIC ILLUMINATION
Thesis submitted 10 Ihe
CoC/'in University of Science and Technology
in partialfi'{fillmenl ( ~ ( t h e
requirements/or the degree of
Doctor of Philosophy
under the Faculty qfTechnology
By
VINUTHOMAS
DEPARTMENT OF ELECTRONICS
F ACUL TY OF TECHNOLOGY
COCHIN UNIVERSITY OF SCIENCE AND TECHNOLOGY
COCHIN 682 022, INDIA
SEPTEMBER 2008
CERTIFICATE
This is to certify that thi s thesis entitled "STRATEGIES FOR INVERSE PROFILING
OF TWO DIMENSIONAL DIELECTRIC SCATTERERS USING
ELECTROMAGNETIC ILLUMINATION" is a bona fide record of the research work
carried out by Mr. Vinu Thomas. under my supervision in the Department of Electronics.
Cochin University of Science and Technology. The results presented in this thesis or
pans of it have not been presented for any other degree.
Cochin 682 022
9,h September 2008
Dr. K. T. MAT HEW
(Supervising Teacher)
Professor Emeritus
Department of Electroni cs
Cochin University of Science and Technology
DECLARATION
I hereby declare that the work presented in this thesis entitled "STRATEGIES FOR
INVERSE PROFILING OF TWO DIMENSIONAL DIELECTRIC SeA TTERERS
USING ELECTROMAGNETIC ILLUMINATION" is based on the original work
done by me under the supervision of Dr. K. T. Mathew in the Department of
Electronics, Cochin University of Science and Technology, and that no part thereof has
been presented for any other degree.
Cochin
9th September2008
Abstract
Electromagnetic tomography has been applied to problems in
nondestructive evolution, ground-penetrating radar, synthetic aperture
radar, target identification, electrical well logging, medical imaging
etc. The problem of electromagnetic tomography involves the
estimation of cross sectional distribution dielectric permittivity,
conductivity etc based on measurement of the scattered fields. The
inverse scattering problem of electromagnetic imaging is highly non
linear and ill posed, and is liable to get trapped in local minima. The
iterative solution techniques employed for computing the inverse
scattering problem of electromagnetic imaging are highly computation
intensive. Thus the solution to electromagnetic imaging problem is
beset with convergence and computational issues. The attempt of this
thesis is to develop methods suitable for improving the convergence
and reduce the total computations for tomographic imaging of two
dimensional dielectric cylinders illuminated by TM polarized waves,
where the scattering problem is defmed using scalar equations. A multi
resolution frequency hopping approach was proposed as opposed to
the conventional frequency hopping approach employed to image large
inhomogeneous scatterers. The strategy was tested on both synthetic
and experimental data and gave results that were better localized and
also accelerated the iterative procedure employed for the imaging. A
Degree of Symmetry formulation was introduced to locate the scatterer
in the investigation domain when the scatterer cross section was
circular. The investigation domain could thus be reduced which
reduced the degrees of freedom of the inverse scattering process. Thus
the entire measured scattered data was available for the optimization of
fewer numbers of pixels. This resulted in better and more robust
reconstructions of the scatterer cross sectional profile. The Degree of
Symmetry formulation could also be applied to the practical problem
of limited angle tomography, as in the case of a buried pipeline, where
the ill posedness is much larger. The formulation was also tested
using experimental data generated from an experimental setup that was
designed. The experimental results confirmed the practical
applicability of the formulation.
Table of Contents
Acknowledgement
5
1 Introduction
1.1 Introduction 7
1.2 Imaging and Inverse Problems in 9
Electromagnetics
1.3 Effect of Electromagnetic waves on matter 15
1.4 Dielectric Polarization 16
1.5 Brief sketch of present study 17
References 19
2 Review of Literature 23
References 31
3 Formulation of the Problem
3.1 Introduction 45
3.2 Approximate solution of the forward 52
scattering problem
3.3 Numerical solution ofthe forward 54
scattering problem
3.4 Inverse scattering techniques 60
References 75
4 Multiscaled frequency hopping
4.1 Introduction 79
4.2 Problem statement and solution 81
4.3 Results and Discussions 84
References 95
5 Degree of symmetry formulation
5.l Introduction 99
5.2 Formulation of the problem 101
5.3 Degree of symmetry formulation for localizing 103
the scatterer
5.4 Simulations and discussions 108
5.5 Formulation of the buried pipeline scattering 128
problem
5.6 Degree of symmetry formulation for localizing 130
the buried pipeline
5.7 Numerical Simulations
6
7
References
Experimental setup for monochromatic inverse
scattering
6.1 Introduction
6.2 Mechanical specifications for the apparatus for
the inverse scattering experiment
6.3 Results and discussions
References
Conclusion and future directions
7.1 Introduction
7.2 Shortcomings and sources of error
7.3 Possible directions
Reference
List of publications
145
149
150
157
173
177
180
181
182
183
Acknowledgement
I thank God Almighty for his immense blessings all throughout this
endeavour.
I would like to express my deepest gratitude to Professor K. T.
Mathew for his faithful guidance, valuable advice, immense patience,
directive comments, continuous supervision and encouragement throughout
the course of this research.
I also express my heartfelt gratitude to the Head of Department of
Electronics, CUSAT, Professor K Vasudevan for his constant support and
motivation.
1 deeply thank Professors K G Balakrishnan, P R S Pillai, P
Mohanan, Tessamma Thomas, C K Aanandan, James Kurian, Supriya M H
and 0 Rajaveerappa for their continuous encouragement.
I acknowledge with gratitude the suggestions and comments
provided by my co researchers of the Microwave Tomography and Material
Research Lab, especially Dr Joe Jacob, Dr Jaimon Yohannan, Ani 1
Lonappan, Dr G Bindu, Dr V Hamsakutty, Praveen Kumar A V, Robin and
Ullas.
Special thanks are due to Dr. Thomaskutty Mathew, School of
Technology and Applied Sciences, M.G. University, Kerala, Mr Cyriac M.
Odakkal, Assistant Professor, VJCET, Muvattupuzha and my colleague and
friend Mr. C Gopakumar, for their valuable suggestions and criticism.
I also acknowledge the help rendered by my other co researchers of the
Department of Electronics of CUSAT. Thanks are also due to the technical
and office staff of the Dept of Electronics, CUSAT, for all the help they
have rendered to me.
I also acknowledge the support of my parent institution, the Institute
of Human Resource Development, Government of Kerala, for supporting me
during my research.
I would also like to thank my parents, my late grand father, my wife
and our three children and all my friends and relatives for their patience,
endurance and spiritual support.
CHAPTER J
Introduction
1.1 Introduction
Inverse problems are ubiquitous in science and engineering and have
rightfully received a great deal of attention from applied
mathematicians, statisticians and engineers. This is due in part to the
importance of applications like biomedical, geophysical and seismic
imaging that require the practical solution of inverse problems. It is
also due to the recent development of powerful computers and fast
reliable numerical methods with which to carry out the solution
process.
Inverse problems involve the estimation of certain quantities
based on indirect measurement of these quantities. For example
seismic exploration yields measurements of vibrations recorded on the
earth surface. These measurements are only indirectly related to the
subsurface geological formations that are to be determined. In medical
computerized tomography, structures within the body are sought to be
imaged form measurements of x-rays that have passed through the
body. In ground water flow modeling one estimates the material
parameters of an aquifer from measurements of pressure of a fluid that
immerses the aquifer.
Inverse problems are typically ill posed, as opposed to the
well-posed problems more typical when mode ling physical situations
where the model parameters or material properties are known (forward
problems). According to the definition given by Jacques Hadamard I,
the three conditions for a well-posed problem are existence,
uniqueness, stability of the solution or solutions. [n most inverse
problems, the condition of stability is most often violated 2. [n the
sense of functional analysis, the inverse problem is represented by a
mapping between metric spaces. While inverse problems are often
formulated in infinite dimensional spaces, limitations to a finite
number measurement, and the practical consideration of recovering
only a finite number of unknown parameters, may lead to the problems
being recast in discrete form. In this case the inverse problem will
typically be ill-conditioned.
Inverse problems can be either linear or non-linear problems 3.
In general all inverse problems maybe expressed in the form
m:::c:Gf (1.1 ),
where m is the measurement made, f is the physical quantity to
be estimated and G is the operator matrix. When G is a linear
operator, the inverse problem is a linear one. From a theoretical point
of view, many linear inverse problems are well understood 3. When
G is a non linear operator, the inverse problem is inherently more
difficult to solve. Non linear inverse problems are characterized by
multiple solutions
4
. Thus solution procedures for solving non linear
inverse problems are liable to get trapped in local optima.
The solution strategies for solving the non linear inverse
problem could be deterministic or stochastic. Deterministic methods
involve linearizing the inverse problem around a current estimate and
solving the problem iteratively, each time improving the correctness of
the estimate. The iterative procedure is gradient based, with a cost
functional minimized in every iteration. Therefore the deterministic
procedures are likely to be trapped in local minima. The stochastic
methods, such as Genetic algorithms or Simulated annealing, converge
to the global optimum of the solutions. However the stochastic
methods are extremely computation intensive, and are not employed
when the number of unknowns of the inverse problem is large 5.
1.2 Imaging and Inverse Problems in
Electromagnetics
The attempt to image an object that is invisible to the naked
eyes is an important aim of scientists in many disciplines. The imaging
problem is an in inverse scattering problem, where one attempts to
infer the profile of a scattering object from measurement data collected
at a distance away from the scattered, as shown in figure 1.1.
T -- "-
/ ~ ~ R
I ", " \
"
I ~ " - ~ . / e
1..( )-!.
\ E<x y) I
\ I
J
-V
/
/
--- -y- ~
Figure 1.1
The inverse scattering problem of imaging.
Here the cross sectional profile of the object is sought to be reconstructed
from the scattered field measurements at the receivers.
In addition to obtaining the shape of an object, a quantitative
description of its permittivity, velocity, or conductivity profile is also
obtainable from an inverse - scattering experiment. This can be
extremely valuable diagnostic infonnation 6.
The advent of X-rays brought much enthusiasm to the
scientific community even though it was discovered to have an adverse
radiation effect. An X-ray picture is a shadow cast on a photographic
film by denser media that are more opaque to X-rays compared with
less dense media. Later it was discovered how to make three-
dimensional pictures out of a series of shadows cast by X-rays. This is
known as X-ray tomography, which later developed to become
computer-aided tomography (CAT) 7,8.
Another very significant modality is the use of nuclear
magnetic resonance imaging (MRI), MRI makes use of the fact that
spin echoes of protons have different resonance frequencies depending
on the ambient static magnetic field, The stronger the magnetic field,
the higher is the resonance frequency of the proton spin. Hence, by
immersing a body in a static magnetic field that has a linear gradient in
a three-dimensional space, the spins echo with different frequencies
when excited by a time varying electromagnetic signal. By
performing the same experiment again by varying the linear gradient
over a number of directions, enough data diversity can be generated so
that a two- or three-dimensional image of an object maybe
reconstructed 9.
Another modality is ultrasonic imaging 10. This works almost
like X-ray tomography. However, due to the lower frequencies and
longer wavelengths of ultrasonic waves, they do not travel exactly in a
straight line but diffract around objects. Also there may be multiple -
scattering effects within an object when ultrasonic waves enter it.
Hence, when ultrasonic images are reconstructed based upon assuming
straight rays and no multiple scattering, image distortion arises. An
imaging scheme called diffraction tomography (DT) attempts to
correct for the diffraction effect but not for the mUltiple scattering
effects 11.
To date electromagnetic (microwave and radio wave) imaging
has not found wide medical applications due to its longer wavelength
and hence, lower resolution. The wavelength can be reduced by using
higher frequencies, but the ensuring material loss is also higher greatly
reducing the penetration depth. However, electromagnetic imaging has
found a niche in nondestructive evolution, ground-penetrating radar 12,
synthetic aperture radar 13, target identification and electrical well
logging 14. Many rugged and constrained environments are not
amenable to having X-ray or nuclear magnetic resonance equipment
on site. Also, it is often much easier to couple electromagnetic signals
into material media at an air- solid interface compared to ultrasonic
signals. At such interfaces, the acoustic impedance contrast between
the air and the solid is often huge while the electromagnetic impedance
contrast is generally lower. The high impedance contrast gives rise to a
small transmission coefficient for acoustic signals but a reasonable one
for electromagnetic signals. Electromagnetic waves can propagate long
distances with little attenuation in air or vacuum. Consequently,
electromagnetic waves are extremelv useful for remote sensin!! and
imaging using SAR techniques where the sensor - target separation is
often enormous 13.
The inverse- scattering solution is sought from the field
perturbation, or the scattered field induced by the presence of the
object in the transmitted fields. This solution is often difficult to
obtain, especially when multiple scattering effects are present within
the object. Such effects can cause the scattered field to be nonlinearly
related to the object function to be reconstructed. Object function
describes the velocity, permittivity, or conductivity distribution of the
object whose profile is to be reconstructed.
The solution to the inverse scattering problem of the
Electromagnetic/Microwave imaging problem is non unique because
high spatial frequency portions of the object give rise to evanescent
waves. These waves are exponentially small at the receiver locations,
and in practice are not measurable unless the receivers are extremely
close to the scatterer. Hence, the scattering operator is a low-pass filter
that removes high spatial-frequency information about the object
function. All low-pass filtering operations do not have a unique
inverse operation unless some assumptions are made about the
solution.
In the past, many methods have been proposed to solve
particular classes of inverse scattering problems by making certain
assumptions about the object or the scattering process. For example in
Computer Aided Tomography, a ray picture is used wherein waves are
assumed to propagate in straight lines but all diffraction and mUltiple-
scattering effects are ignored.. Diffraction TomoQ"ranhv take:.;: into
account wave diffraction but ignores multiple scattering and assumes
the object is small. Methods have been proposed to solve the inverse-
scattering problem exactly for one-dimensional objects such as planar
layered media 16. However, such methods are not possible for multi
dimensional objects.
The relationship between the scattered field and the scattering
object is a non linear one in the inverse scattering problem of
electromagnetic imaging. The non linearity increases with multiple
scattering effects within the object 17. as shown in the figure 1.2.
Assuming that when only one scatterer I is present, the scattered field
is Eis and when only scatterer 2 is present, it is Eh' When both
scatterers are simultaneously present, the scattered field is not just
El, + E
2s
' but Ell + E2.' + Ems' where Ems is the result of multiple
scattering between the scatterers. Thus linear superposition cannot be
employed to solve the scattering problem.
Figure 1.2
MUltiple scattering gives rise to non linearity in the
inverse problem that orecludes the use of linear
I fits nonlmear relationship prevents a simple solution to the
inverse - scattering problem. One way to solve this nonlinear problem
has been to cast it into an optimization format. Here, a cost functional
is minimized by applying a gradient - search technique. With such an
approach, a forward - scattering solution is sought whenever the
gradient is required. Since scattering by an arbitrary object has to be
obtained by a numerical method, the nonlinear inverse scattering
method is often computationally intensive. However, this method can
have an important technological impact if the computational burden is
reduced to an acceptable leveL Overall, some of the issues that
nonlinear inverse scattering theories should address are 18
Resolution of the reconstructed object.
Maximum object contrast that can be accurately inverted
Maximum object size that can be inverted
Algorithm speed
Robustness and numerical stability relative to nOISY or
imperfect measured data.
Validity for practical scattering measurement geometries.
Generality allowing application to other wave equations and
scattering phenomena (for example, ultrasound as well as
electromagnetic waves)
Minimizing the amount of a priori information about the
object.
One of the most pressi
l1
g problems for nonlinear inverse -
scattering theory is the speed of solution. Solving the inverse problem
numerically, often involves solving the forward - scattering problem
numerically. However, the latter solutiQll is by itself computationally
intensive. Hence there is considerable need to develop algorithms that
can either rapidly solve forward scattering also.
The ways in which microwave interacts with matter will be
discussed so as to develop a comprehensive picture of the
electromagnetic inverse scattering problem.
1.3 Effect of electromagnetic waves on matter
Electromagnetic waves propagate in uniform dielectrics according to
the relation,
(1.2)
where E is the scalar instantaneous electric field in the dielectric at a
distance z from the origin, Eo is the electric field at the origin, and y is
the complex propagation constant 19. 20, The complex propagation
constant is defined as
(1.3)
where (j) represents the angular frequency, J.1 the magnetic (complex
permeability) and & the dielectric (complex permittivity) property of
the medium. For perfect dielectrics, the magnetic properties are
assumed to be those of free space, i.e. purely real with no attenuation
or phase shift.
In lossy dielectric materials, the dielectric properties are
complex 20 when polarized by an electric field or when subjected to a
time harmonic electric field. This is due to the' formation of both
conduction and displacement currents. The conduction currents
represent the current flow that is in phase with the applied voltage
whereas the displacement currents are in phase quadrature with the
applied voltage. The complex relative permittivity is represented as,
e :::: e '-Je 11
r r r
(l.4)
where er 'is the real part of complex relative permittivity known as the
dielectric constant and er"" is the imaginary part known as the
dielectric loss. The real and imaginary parts of complex relative
permittivity represent the complimentary processes of energy storage
and dissipation respectively. Since heat production is related to the
frequency and the dielectric loss factor, these are combined to be
known as the conductivity cr as,
and loss tangent as,
(J :::: me e "
n r
e "
tanJ = _r_.
er'
1.4 Dielectric polarization
(1.5)
(1.6)
The common feature of dielectric materials is their ability t6 store
electromagnetic energy. This is accompanied by the displacement of
positive and negative charges under the influence of the applied
electric fields, against the force of atomic and molecular attraction 21.
The mechanism of charge displacement (polarization) depends on the
type of dielectric material and the frequency of.the applied field. The
electrical equilibrium is disturbed because the applied field causes
spatial separation of charges of opposite signs.
Space charge polarization occurs at low-frequency ranges such as
VLF and LF bands. It occurs when the material contains free electrons
whose displacements are restricted by obstacles such as boundaries.
When an electric field is applied, the electrons accumulate on the
obstacle, and the resulting charge separation polarizes the material.
Entire regions of the material become either positive or negative.
Polarization by dipole alignment occurs at higher frequencies of the
HF band and microwave region, at the molecular level. Dielectric
heating takes place due to this phenomenon.
Ionic polarization takes place at infrared frequencies due to the
separation of positive and negative ions in the molecule.
Electronic polarization occurs at very high frequencies close to the
ultraviolet region. The negative electronic cloud surrounding the
positive nucleus is displaced in the direction of the applied field.
1.5 Brief sketch of the present study
The mm of this thesis is to develop algorithms for
electromagnetic inverse scattering for dielectric scatterers that address
the important issues of convergence, computational speed and
robustness.
Chapter two describes the background literature about various
Inverse scattering algorithms that have been developed for
tomographic reconstruction of the scatterers.
17
Chapter three introduces the mathematical formulations for the
electromagnetic forward and inverse scattering problems and the
discretization model employed.
Chapter four discusses the multi-scaled frequency hopping
approach that was fonnulated for two dimensional dielectric scatterers
for addressing the problem of local minima.
Chapter five discusses the Degree of Symmetry
(D.o.S) fonnulations that have been developed for the tomographic
reconstruction of two dimensional dielectric pipelines of circular cross
section.
Chapter six describes an experimental set up for performing
inverse scattering experiments and the experimental verification of the
D.o.S formulations.
Chapter seven describes the conclusion arrived at and the
possible future directions that might be pursued.
References
1. Hadamard, J., Lectures on Cauchy's problem in linear partial
differential equations, New Haven, C.T, Yale University Press,
1923.
2. Vogel, C. R., Computational methods for inverse problems,
SIAM, Philadelphia, 2002.
3. Tarantola, A., Inverse problem theory and methods for model
parameter estimation. SIAM, Philadelphia, 2005.
4. Kleinman, R.E. and P.M. van den Berg, "Non linearized
approach to profile inversion", Int. J. Imag. Sys. Tech, Vol.2,
1990, pp 119-126.
5. Franchois, A., Contribution A La Tomographie Microonde:
Algorithms De Reconstruction Quantitative Et Verfications
Experimentaies, PhD Thesis, University of Paris, 1993.
6. Chew, W.C., G. P. Otto, W. H. Weedon, J. H. Lin, C. C. Lu, Y.
M. Wang and M. Moghaddam, "Non-linear diffraction
tomography, the use of inverse scattering for imaging", Int. J.
Imag. Sys. Tech, VoI.7, 1996, pp 16-24.
7. Raviv, J., F. Greenleaf and C. T. Herman, Computer Aided
Tomography and Ultrasonics in Medicine. Elsevier, North
Holland, 1979.
8. Kak, A.C., "Computerized tomography with X-Ray, emission
and ultrasound sources", Proceedings of IEEE, Vol. 67, 1979,
pp 1245-1272.
9. Krestel, E., Imaging systems for Medical Diagnosis,
Fundamentals and technical solutions: X-Ray diagnostics,
Computed Tomography, Nuclear Medical diagnostics,
Magnetic Resonance Imaging and Ultrasound technology,
New York, Wiley, 1990.
10. Lee, H. and G. Wade, Modern Acoustic Imaging, New York,
IEEE Press, 1986.
11. Devaney, A.J., "A filtered Back projection method for
diffraction tomography", Ultrasonic Imaging, Vol. 4, 1982, pp
336-360.
12. Maser, K.R., Ground penetrating radar surveys to characterize
pavement layer thickness variations at GPR sites, Strategic
highway research program, 1994.
13. Curlander, J.e. and R.N. McDonough, Synthetic Aperture
Radar: systems and signal processing, New York, 1991.
14. Chew, W.C and Q. H. Liu, " Inversion of Induction tool
measurements using Distorted Born Iterative Method and CO-
FFT", IEEE Transactions on Geo-science and Remote Sensing,
Vot. 32, 1994, pp 878-884 .
15. Chew, W.e., o. P. Otto, W. H. Weedon, 1. H. Lin, C. C. Lu, Y.
M. Wang and M. Moghaddam, "Non-linear diffraction
tomography, the use of inverse scattering for imaging", Int. J.
Imag. Sys. Tech, Vo1.7, 1996, pp 16-24.
16. Bube, K.P and R. Burridge, "The one-dimensional problem of
reflection seismology", SIAM Review, Vol. 25, 1983, pp 497-
559.
.
17. Chew, W.C., Waves and Fields in Inhomogeneous Media, New
York, Van Nostrand Reinhold, 1990.
18. Taflove, A., ed., Advances in Computational Electrodynamics,
The Finite Difference Time Domain Method, Artech House,
Boston, 1998.
19. Larsen, L.E. and J. H. Jacobi, Medical Applications of
Microwave Imaging, IEEE Press, 1985, pp 118-137.
20. Grant, E.H., R. 1. Sheppard and G. P. South, Dielectric
behavior of Biological Molecules in Solution, Oxford-
Clarendon Press, 1978.
21. Thurey, 1., Microwaves, Industrial, Scientific and Medical
Applications, Artech House, Boston, 1992.
CHAPTER 2
Review of Literature
In linear inverse scattering theories, approximations are made such that
a linear relationship exists between the measured data and the object.
Then the information about the object is obtained by solving an
associated set of linear equations. This is true for modalities such as
Computer Aided Tomography (C.A.T) where the modality employed,
X- Ray, is assumed to travel in a straight line path and traverse an
object only once. A Back projection algorithm using the projection
slice theorem is used to reconstruct the object efficiently in such cases
1.2. The Back projection algorithm is based on the WKB (Wentzel,
Kramer and Brillouin) approximation which states that at high
frequencies, a wave propagates mainly in the forward direction with
very little reflection 3. Here the object, whose tomographic profile is to
be reconstructed, is described by its attenuation profile 4, which can be
obtained using a radon transformation 5, which have two dimensional
as well as three dimensional forms 6.
In projection tomography, one assumes that the waves
propagate as straight line rays without diffraction. While this is a
correct assumption at the extreme short wavelengths of X-Ray
tomography, at the shorter wavelengths employed in ultrasonic and
electromagnetic tomography, diffraction is an important phenomenon.
23
Therefore back projection tomography does not give satisfactory
results at the longer wavelengths. Diffraction was incorporated in the
context of the Born and Rytov approximations and a reconstruction
algorithm was developed accordingly, called diffraction tomography 7-
15. In diffraction tomography, the Born or Rytov approximation is used
to calculate the scattered field, when a linear relationship will exist
between the scattered field and the object 16. A Fourier Transform
relationship exists between the scattered field and the object function
and the Fourier spectrum of the object function maybe recovered using
a filtered back propagation algorithm. The object can thus be
determined 7. The Born and Rytov approximations are valid when the
contrast of the scatterer is so small that only single scattering is
important. Thus when multiple scattering effects are important, as in
electromagnetic inverse profiling of objects having strong contrasts,
the Diffraction Tomography technique breaks down 17. Diffraction
Tomography is often used in ultrasonic imaging, to achieve good
reconstruction when multiple scattering effects are small 18.
Many non linear inverse scattering theories have been proposed
to deal with inverse scattering problems involving substantial multiple
scattering effects. In one dimension, the problem has been thoroughly
studied by a large number of scientists. Several algorithms have been
developed which are layer stripping algorithms, in which the
reconstruction is done in layers. An elegant algorithm for solving the
one dimensional inverse scattering problem is the Gelfand- Levitan
Marchenko Method 19. Bube and Bridge 20 and Sezginer 21 suggested
a time domain algorithm which is also a layer stripping technique,
called the method of characteristics. Coen et al 22, 23 describes an
algorithm for the inverse scattering problem of profiling layered earth.
Ge 24 describes an iterative technique for one dimensional profile
inversion. However these algorithms are suitable only for one
dimensional objects and have not been verified for higher dimensional
. . 25
lDverse scattenng .
The only general non linear inverse scattering theory valid for
higher dimensions that has been verified till date involves numerical
methods 3. These methods are usually iterative and are traditionally
expressed as optimization problems. The Born Iterative method was
proposed by Wang et al 26 which improved over the simple Born
approximation and allowed scatterers of relatively large size or high
contrast to be constructed. Moghaddam et al 27. 28 comment on the
improved performance of the Born Iterative method in solving an
inverse electromagnetic time domain problem. Lu et al 29 describe the
use of Born Iterative method to reconstruct a scatterer inside dielectric
walls. Abubakar et al suggested an improved robust iterative method
for Born Inversion 30. While the Born iterative method is simple to
implement, it has the disadvantage that it does not offer second order
convergence 25. An improved method, the Distorted Born Iterative
method (DBIM), was proposed by Chew et al 31. 32. Leone et al 33
applied the Distorted Born Iterative procedure to reconstruct two
dimensional dielectric scatterers for cylindrical geometries. Super
resolution was observed by Chen et al 34 for reconstructions employing
the Distorted Born Iterative method. The DBIM was employed by Lu
et al 35 for electromagnetic and ultrasonic inverse scattering of two
dimensional dielectric scatterers. Haddadin and Ebbini 36 employed a
multiple frequency distorted Born iteration method for imaging
strongly scattering media. A method similar to the Newton Raphson
method for solving the electromagnetic inverse scattering problem was
proposed by loachimowicz et aI 37 and Tarantola 38 and is known as
the Newton Kantarovich (NK) method. loachimowicz et al 39
performed a systematic assessment of the effect of experimental and
model noise on the NK method and concluded that the NK algorithm
is very robust. Franchois and Pichot 40 proposed a Levenberg -
Marquardt variant of the Newton Kantarovich method with a cross
validation scheme for obtaining the regularization parameter. The
DBIM and NK methods have been shown to be computationally
equivalent 41-45.
Gradient descent based algorithms have been proposed for the
electromagnetic inverse scattering problem of microwave imaging.
Though these algorithms have the advantage that the forward problem
need not be solved in the iterations, they suffer from the disadvantage
that only single profile (pennittivity or conductivity) reconstructions
are possible, or the background has to be assumed lossless 46. The
modified gradient procedure was proposed by Kleinman et al 47.
Baussard et al 48 employed a regularization scheme based on a
Markovian approach for the modified gradient procedure to
reconstruct the permittivity profile. Souriau et al 49 developed a
modified gradient approach to inverse scattering of binary objects in
stratified media. Lambert et al 50 employed a constrained modified
gradient method for the retrieval of a buried cylindrical obstacle.
Harada et al 45 applied the conjugate gradient descent method for
minimizing the cost functional in the inverse scattering problem of
diffraction tomography. Moghaddam et al
51
showed that the conjugate
gradient method can be used to minimize the cost functional. Rekanos
et al 52. 53 combined the finite-element method and the Polak-Ribiere
nonlinear conjugate gradient optimi2!ation algorithm for reconstructing
the electromagnetic properties of unknown scatterers. Rubaek et al 54
employed a technique based on conjugate gradient least squares
method for nonlinear microwave imaging for breast-cancer screening.
Reconstructions employing experimental data using a conjugate
gradient method with an edge preserving regularization scheme was
proposed by Lobel et al 55. Pastorino 56 discusses the various inversion
algorithms for microwave imaging in biomedical, subsurface detection
and nondestructive evaluation applications. The contrast source
inversion method, which is also a gradient based algorithm, was
suggested to provide more robust reconstructions of the dielectric
scatterer cross sections 57-{)0. The Local shape function T -matrix
method was developed to invert strong metallic scatterers 61-64. It was
also found out that the LSF method can reconstruct scatterers with a
large dielectric contrast for TE polarized incidence or variable density
acoustic waves 65,66.
Stochastic methods also were employed for mIcrowave
imaging. The simulated annealing technique was employed by
Franchois, Gamero et al 67, 68 for reconstructing the cross sectional
profile of a 2-D dielectric scatterer under TM Polarization. Simulated
annealing was also employed by Park et al 69 for microwave imaging
of a perfectly conducting cylinder. A Gibbs Markov random field
model was employed for active imaging at microwave frequencies by
Regazzoni and Foresti 70. Real Coded Genetic algorithm was
employed for the inverse scattering problem of imaging parallel
perfectly conducting cylinders 71. Genetic algorithms have been used
for microwave imaging for non destructive evaluation of civil
structures 72. A genetic algorithm based strategy was proposed for the
tomographic subsurface detection of defects 73. Li and Cheng proposed
an application of the genetic algorithm for microwave imaging of a
layered dielectric object 74. Noghanian et al 75 proposed a numerical
method based on genetic algorithm optimization for microwave
imaging. Qing et al 76 suggested an improved strategy for microwave
imaging of parallel perfectly conducting cylinders where the real
coded genetic algorithm was employed to provide starting estimates
for Newton Kantarovich iterations. A technique where a parallel
genetic algorithm was combined with finite domain time difference
was applied to the inverse scattering problem of imaging 3-D buried
objects by Chen et al 77. Xiao and Yabe tested the performance of the
genetic algorithm for microwave imaging of parallel perfectly
conducting cylinders on real data which gave very good results 78.
Caorsi et al 79 applied a hybrid genetic algorithm for nondestructive
testing of two dimensional structures. Huang and Mohan 80 employed
the micro-genetic algorithm (m-GA) to solve the three dimensional
inverse scattering problem of microwave imaging of dielectric objects.
However, in general, stochastic methods require a large number of
forward solutions to the scattering problem and are hard to apply
unless the scattering problem is small 81.
Neural networks have also been employed to solve the inverse
scattering problem of microwave imaging. Caorsi and Gamba
employed a two layer perceptron neural network trained with the
YPROP variant of the backpropagation algorithm which is
significantly faster than conventional backpropagation 82. Wang and
Gong 83 employed a coupled gradient neural network to reconstruct the
complex permittivity of the biological tissues illuminated by the
transverse magnetic (TM) incident waves. Lee employed a neural
network with radial-basis functions to the problem of microwave
Imagmg of cylinders. The RBF-NN was trained by some direct-
scattering data sets and thus could predict the images of the target
cylinders 84.
The frequency hopping approach has been proposed for the
imaging of large inhomogeneous bodies 85-90. The idea is to collect
measurement data over a set of frequencies starting from a low
frequency where the inverse scattering problem is solved ftrst. This is
then taken as initial guess for iterations at progressively increasing
frequencies. Thus better convergent reconstructions are possible when
compared to the results obtained from high frequency measurement
data alone.
Electromagnetic imaging has also been employed for inverse
scattering of dielectric objects buried in inaccessible media. The
Distorted Born Iterative procedure was employed by Cui et al 91 to
image two dimensional scatterers buried in lossy earth. Higher order
extended Born approximation was applied for low frequency detection
of two-dimensional buried objects 92. The finite domain time
difference method was used by Toshiyuki et al for microwave imaging
of buried objects using time domain data 93. A differential evolution
method was applied by Massa et al 94 to the inverse scattering problem
of imaging a buried two dimensional scatterer. Objects embedded in
cured concrete could be imaged by using inverse scattering techniques
95. Vertiy and Gavrilov 96 employed first-order diffraction tomography
methods for tomographic reconstructions of two-dimensional objects
in millimeter wavelengths band. Guillermin et al 97 proposed a
diffraction tomography based algorithm for imaging objects buried in
the sediment bottom of deep sea. Chaturvedi and Plumb 9R used the
constrained optimization method for electromagnetic imaging of
underground targets. Baussard et al 99 developed an adaptive
multiscale reconstruction method for imaging underground scatterers.
Helaoui et al 100 proposed a multi illumination approach for buried
object detection.
References
1. Raviv, 1.1., F. Greenleaf and G. T. Herman, eds., Computer
Aided Tomography and Ultrasonics in Medicine, Elsevier North
Holland, 1979.
2. Herman, G.T, Image reconstruction from Projections, New
York, Academic Press, 1980.
3. Chew w.e, Waves and fields in inhomogeneous media, New
York, IEEE Press, 1995.
4. Kak A.C, "Computerized tomography with X-Ray, emission and
ultrasound sources", Proceedings 0/ the IEEE, Vo!. 67, 1979, pp
1245-1272.
5. Radon, 1., "Uber die bestimmung von funktionen durch thre
intergralwerte langs gewisser mannigfalt igkait en" , Berichte
Saechsische Akademie der Vol. 69, 1917, pp
262-277
6. Deans, S. R., The Radon Transformation and some Q/ its
applications, New York, W iley, 1983.
7. Devaney, A. J., " A Filtered Back propagation algorithm for
diffraction tomography," Ultrasound imaging, Vo!. 4, 1an. 1982,
pp 336-350
8. Devaney, A. J., "Geophysical diffraction tomography," IEEE
Transaction on Geological Science: Special issue on Remote
Sensing" Vol. GE-22, Jan. 1984, pp 3-13.
9. Kak A.C, "Tomographic imaging with diffracting and non
diffracting sources", Array Processing systems, Simon Haykin
ed., Prentice Hall, 1984
to. iCaWh, M., M. Soumekh and R. K. Mueller, "Tomographic
Imaging via wave equation inversion", ICASSP 82, May 1982,
pp 1553-1556.
11. Kenue, S.K. and 1. F. Greenleaf, "Limited angle diffraction
tomography", IEEE Transactions on Sonics and Ultrasonics.
Vol. SU-29, luly 1982, pp 213-217.
12. Lu, Z. Q., M. Kaveh and R. K. Mueller, "Diffraction tomography
using beam waves: Z-average reconstruction", Ultrasonic
lmaging, Vol. 6, Jan. 1984, pp 95- 102.
13. Mueller, R. K., M. Kaveh and G. Wade, " Reconstructive
tomography and applications to Ultrasonics", Proceedings of the
IEEE, Vot. 67, 1979, pp 567-587
14. Nahamoo, D. and A. C. Kak, Ultrasonic Diffraction imaging.
T R ~ E E 82-20, School 0 Electrical Engineering, Purdue
University, 1982.
15. Nahamoo, D., S. X. Pan and A. C. Kak, "Synthetic aperture
diffraction tomography and its interpolation free computer
implementation", IEEE Transactions on Sonics and Ultrasonics,
Vo!. SU- 31, July 1984, pp 218-229.
16. Born, M. and E. Wolf, Principles of Optics, New York,
Pergamon Press, 6
th
ed., 1980.
17. Slaney, M., A. C. Kak and L. E. Larsen, "Limitations of imaging
with first order diffraction tomography", IEEE Transactions on
Microwave Theory and Techniques. Vo!. MTT-32, 1984, pp 860-
874.
18. Lee, H. and G. Wade, Modern Acoustic Imaging. IEEE Press,
N ew York, 1986.
19. Chew, W. C., Waves and fields in Inhomogeneous media, New
York, Van Nostrand Reinhold, 1990.
20. Bube, K. P and R. Burridge, ''The one dimensional problem of
reflection seismology", SIAM Review, 25(4), 497-559.
21. Sezginer, A., FOnl-'ard and Inverse Problems in Transient
Electromagnetic fields, Ph.D. Thesis, MIT, 1985.
22. Coen, S., "Inverse scattering of a layered and dispersionless
dielectric half-space, Part 1: refelection data from plane waves at
normal incidence", IEEE Transactions on Antennas and
Propagation, AP-29, 1981, pp 298-306,
23. Coen, S. and W. Vu, "The inverse problem of direct current
conductivity profile of layered earth", Geophysics, 46, 1981, pp
1702-1713.
24. Ge, D. 8., "An iterative technique for one dimensional profile
inversion", Inverse Problems, 3, pp 399-406, 1987.
25. Taflove, A., ed., Computational Electrodynamics, the Finite
Difference Time Domain Method, Artech House, Boston, 1998.
26. Wang, Y. M. and W. C. Chew, "An iterative solution of two-
dimensional electromagnetic mverse scattering problem",
International Journal of Imaging Systems Technology, VoL 1,
No. 1, 1989, pp 100-108.
27. Moghaddam, M., W. C. Chew, and M. Oristaglio, "Comparison
of the Born iterative method and Tarantola's method for an
electromagnetic time-domain inverse problem," Int. J. Imaging
Syst. Tech., voL 3,1991, pp. 318-333.
28. Moghaddam, M. and W. C. Chew, "Study of some practical
issues in inversion with the Born iterative method using time-
domain data," IEEE Trans. Antennas Propagat., vol. 41, no. 2,
1993, pp. 177-184.
29. Lu, C.C. and X.G. Zhong., "Image reconstruction of two
dimensional objects inside dielectric walls", Microwave and
Optical technology letters, Vol. 36, Issue 2, Dec. 2002, pp 91-95.
30. Abubakar A., Van den Berg P.M., Semenov S.Y, "A robust
iterative method for Born inversion", IEEE Trans Geoscience
and Remote Sensing, 2004, 42, 2, pp 342-354.
31. Chew, W.C and Y.M Wang, "Reconstruction of two-dimensional
permittivity distribution using the distorted Born interative
method", IEEE Transactions on Medical Imaging., Vo!. 9, 1990,
pp. 218-225.
32. Chew, W. C. and S.L Chuang, "Profile inversion of a planar
medium with a line source of a point source", IEEE Int.
Geoscience and Remote Sensing Society Symposium, Strasbourg,
France, 1984.
33. Leone, Giovanni., R. Persico and R. Pierri, "Inverse scattering
under the Distorted Born approximation for cylindrical
geometries", Journal of Optical Society of America, Vol. 16,
Issue 7, pp 1779-1787.
34. Chen, F. C. and W. C. Chew, "Experimental verification of super
resolution in non linear inverse scattering", Applied Physics
Letters, 72, 3080, 1998.
35. Lu, c., 1. Lin, W. C. Chew and G. Otto, "Image reconstruction
with acoustic measurement using the distorted Born iteration
method", Ultrasonic Imaging, 1996 April, 18(2), pp 140-156.
36. Haddadin, S. and E. S. Ebbini, "Imaging Strongly scattering
media using a multiple frequency distorted Born iterative
method", IEEE Transactions on Ultrasonics, Ferroelectrics and
Frequency Control, 1998, 45(6), pp 1485-1496.
37. Joachimowicz., N., C, Pichot and J. P. Hugonin, "Inverse
scattering - An iterative numerical method for electromagnetic
imaging", IEEE Transactions on Antennas and Propagation,
Vol. AP-39, 1991, pp 1742-1752
38. Tarantola, A, " The seismic reflection inverse problem", Inverse
problems of acoustic and elastic waves, F. Santosa, Y. H. Pao,
W. Syrnes and C. Holland eds., New York, Society of Industrial
and Applied Mathematics, 1984.
39. Joachimowicz, N., 1. Mallorqui, 1. C. Bolomey and A
Broquetas, "Convergence and stability assessment of Newton-
Kantorovich reconstruction algorithms for mIcrowave
tomography", IEEE Transactions on medical imaging, 1998, vol.
17, No. 4, pp. 562-570.
40. Franchois, A and C. Pichot, "Microwave Imaging - Complex
permittivity reconstruction with a Levenberg - Marquardt
method", IEEE Trans. Antennas Propagat., vol. 45, no. 2, pp
203 - 215, Feb. 1997.
41. Remis, R.F. and P.M. van den Berg, "On the equivalence of the
Newton Kantarovich and Distorted Born methods", Inverse
Problems, vol. 16, No.I, February 2004, LI-L4.
42. Leone, G., A Brancaccio and R. Pierri, "Quadratic distorted
approximation for the inverse scattering of dielectric cylinders",
Journal of Optical Society of America, Vol. 18, Issue 3, 2001, pp
600-609.
43. Yao, Y. and Yao Wang, "Frequency domain optical imaging of
absorption and scattering distributions by a Born Iterative
method", Journal of Optical Society ~ r America, Vol. 14, No. 1,
1997, pp 325-342.
44. Joachimowicz, N., 1. J. Mallorqui, J. C. Bolomey and A
Broquets, "Convergence and stability assessment of Newton
Kantarovich reconstruction algorithms for microwave
tomography", IEEE Transactions on Medical Imaging, Vol. 17,
Issue 4, August 1998, pp 562-570.
45. Harada, H., D. J. N. Wall, T. Takenaka and M. Tanaka,
"Conjugate gradient method applied to inverse scattering
problem", IEEE Transactions on Antennas and Propagation,
Vol. 43, NO. 8, Aug. 1995, pp 784-792.
46. Kleinman, R.E., P. M. van den Berg, B. Duchene and D.
Lesselier, "Location and reconstruction of objects using a
modified gradient approach", Inverse Problems of Wave
Propagation and Diffraction, Springer Berlin / Heidelberg, 1997,
pp 143-158.
47. Cui, T. J., W. C. Chew, A. A. Aydiner and S. Chen, "Inverse
Scattering of Two- Dimensional Dielectric objects buried in
Lossy Earth using the Distorted Born Iterative Method", IEEE
Transactions on Geoscience and Remote Sensing, Vol. 39, No:2,
February 2001, pp 339-346.
48. Baussard, A, K. Belkebir and D. Premel, "A Markovian
regularization approach of the modified gradient method for
solving a two-dimensional inverse scattering problem", Journal
of electromagnetic waves and applications, Vol. 17, No. 7, pp.
989-1008.
49. Souriau, L., B. Duchene, D. Lesselier, and R. Kleinman,
"Modified gradient approach to inverse scattering of binary
objects in stratified media," Inverse Problems, Vot 12, 1996, pp
463-481.
50. Lambert, M., D. Lesselier, and B. 1. Kooij, "The retrieval of a
buried cylindrical obstacle by a constrained modified
gradientmethod in the H-polarization case and for Maxwellian
materials," Inverse Problems, Vol. 14, 1998, 1265-1283.
51. Moghaddam, M., W. C. Chew and M. Oristaglio, "Comparison
of the Born iterative method and the Tarantola's method for an
electromagnetic time domain inverse problem", International
Journal of Imaging ~ y s t e m s and Technology, Vol. 3, 1991, pp
318-333.
52. Rekanos, l.T., M. S. Efraimidou and T. O. Tsiboukis,
"Microwave Imagmg: inversion of scattered near-field
measurements", IEEE Transactions on Magnetics, Volume 37,
Issue 5, Sep 2001, pp 3294 - 3297.
53. Rekanos, LT., S. M. Panas, and T.O. Tsiboukis, "Microwave
Imagmg using the finite-element method and a sensitivity
analysis approach", IEEE Transactions on Medical Imaging,
Vol. 18, Issue 1l, Nov 1999, pp 1108-1114.
54. Rubaek, T, P. M. Meaney, P. Meincke and K. O. Paulsen,
"Nonlinear microwave imaging for breast-cancer screening using
gauss-newton's method and the CGLS inversion algorithm",
IEEE transactions on Antennas and Propagation, Vol. 55, No. 8,
2007, pp. 2320-2331.
55. Lobel ,P., C. Pichot , L. Blanc-Feraud and M. Barlaud,
"Microwave imaging: Reconstructions from experimental data
using conjugate gradient and enhancement by edge-preserving
regularization", International Journal ~ r Imaging Systems and
Technology, Volume 8 Issue 4, 1998, pp 337 - 342.
56. M. Pastorino, "Recent inversion procedures for mIcrowave
imaging in biomedical, subsurface detection and nondestructive
evaluation applications", Imaging Measurement Systems, Vo!.
36, Issues 3-4, October-December 2004, pp 257-269.
57. van den Berg, P. M. and R. E. Kleinman, " A contrast source
inversion method", Inverse Problems, Vol. 13, 1997, pp 1607-
1620.
58. van den Berg, P. M and A. Abubakar, "Microwave tomography
using linear and non linear contrast source inversion",
Proceedings oJ the URSI EMTS, 2004, pp 1089 - 1091.
59. van den Berg, P. M, A. Abubakar and J. T. Fokkema,
"Multiplicative regularization for contrast profile inversion",
Radio Science, Vol. 38, No. 2, 2003, 23-1 to 23-9.
60. van den Berg, P. M and A. Abubakar, " Contrast source
inversion method- state of the art", PIER. Vol. 34,1997, pp 189-
218,2001.
61. Chew, W. C and G. P. Otto, "Microwave imaging of multiple
metallic cylinders using shape functions", IEEE Microwave
Guided Wave Letters, Vo1.2, 1992, pp 284-286.
62. Otto, G. P and W. C. Chew, "Inverse scattering of Hz waves
using local shape function imaging: A T Matrix formulation",
International Journal of Imaging Systems and Technology,
Volume 5, 1994, pp 22- 27.
63. Weedon, W. Hand W. C. Chew, "Time domain inverse
scattering using the local shape function (LSF) method",. Jnverse
Problems, Vol. 9, 1993, pp 551-564.
64.0tto, G. P and W. C. Chew, "Time Harmonic impedence
tomography using the T Matrix method", IEEE Transactions on
MedicalImaging, Vol. 13, 1994, pp 508-516.
65. Lin, J. H. and W. C. Chew, "Ultrasonic imaging by local shape
function method with CGFFT". IEEE Transactinns nn
Ultrasonics, Ferroelectrics and Frequency Control, Vol. 43,
1996, pp 956-969.
66. Lu, C. C and W. C. Chew, "Processing Ipswich data with the
local shape function method", IEEE Antennas and propagation
magazine, Vol. 38, No.3, 1996, pp 51-53.
67. Franchois, A., L. Garnero, C. Pichot and 1. P. Hugonin,
"Application of the Simulated Annealing technique to
Microwave Tomography: preliminary results", Inverse methods
in action, P. C. Sbatier Ed., Springer-Verlag, Berlin, 1990.
68. Garnero, L., A. Franchois, 1. P. Hugonin, C. Pichot, and N.
Joachimowicz, "Microwave imaging-complex permittivity
reconstruction-by simulated annealing", IEEE Transactions on
Microwave Theory and Techniques, Vol. 39, Issue 11, Nov
1991, pp1801 - 1807.
69. Park ,So K., Cheon Seok Park, Jung Woong Ra, "Microwave
imaging of a perfectly conducting cylinder with the use of a
simulated annealing approach", Microwave and Optical
Technology Letters, Vol. 11, Issue 5, pp 263 - 265
70. Regazzoni, C. S. and G. L. Foresti, "A Gibbs Markov random
field model for active imaging at microwave frequencies", Signal
Processing, Vol. 47, Issue 2 (November 1995) pp 169 - 185.
71. Qing, A., "Microwave imaging of parallel perfectly conducting
cylinders", International Journal of Imaging Systems and
Technology, Vol.ll, Issue 6, Nov. 2001, pp 365 - 371.
72. Massa, A., M. Donelli , M. Pastorino and A. Rosani,
"Microwave imaging for non-destructive evaluation of civil
structures", Insight - Non-Destructive Testing and Condition
Monitoring, Vol. 47, Issue I, January 2005, pp 11-14.
73. Benedetti, M., M. Donelli, G. Franceschini, A. Massa, and M.
Pastorino, "Evaluation study of the effectiveness of the
integrated genetic-algorithm-based strategy for the tomographic
subsurface detection of defects", Journal of Optical Society of
America, Vol. 23, Issue 6, 2006, pp. 1311-1325.
74. Li , C. Land Y. Y. Cheng, "Application of the genetic
algorithm for microwave imaging of a layered dielectric object
via the regular shape expansion technique", International
Journal of Imaging Systems and Technology, Vol. 10, Issue 4,
Jul 1999, pp 347 - 354.
75. Noghanian, S., A. Sabouni and S. Pistorius , "A numerical
approach to microwave imaging based on genetic algorithm
optimization", Proceedings of SPIE, the International Society for
Optical Engineering, Conference on Health monitoring and
smart nondestructive evaluation ~ f structural and biological
systems V, vo l. 6177, 2006, pp. I B. I-I B.12.
76. Qing, A., K. Lee and S. Yang, "An improved algorithm for
microwave imaging of parallel perfectly conducting cylinders",
IEEE Antennas and Propagation Society International
Symposium, 2000, Vol. 3, pp 1772 - 1775.
77. Ch en, X., D. Liang and K. Huang, "Microwave imaging 3-D
buried objects using parallel genetic algorithm combined with
FDTD technique", Journal of electromagnetic waves and
applications, Vol. 20, no 13, 2006, pp. 1761-1774.
78. Xiao, F. and H. Yabe, "Computational Electromagnetics.
Microwave Imaging of Perfectly Conducting Cylinders from
Real Data by Micro Genetic Algorithm Coupled with
Deterministic Method", IEICE Transactions on Electronics, Vo.
E81-C, No.12, 1998, pp 1784-.1792.
79. Caorsi, S., A. Massa, Andrea M. Pastorino and M. Donelli,
"Improved Microwave Imaging Procedure for Non-Destructive
Evaluations of Two-Dimensional Structures", Technical Report
DIT-03-040, Informatica e Telecomunicazioni, University of
Trento, 2003.
SO. Huang, T. and A. S. Mohan, "Microwave Imaging of Three-
Dimensional Dielectric Objects", IEICE Transactions on
Communications 2005 ES8-B(6), pp 2369-2376.
81. W. C. Chew, "Imaging and mverse problems 10
electromagnetics", Advances in Computational Electrodynamics:
the Finite-Difference Time-Domain Method, Alien Tajlove
(Editor), Artech House, 2003.
82. Caorsi. Sand P. Gamba, "Electromagnetic detection of dielectric
cylinders with a neural network approach", IEEE Transactions
on Geoscience and Remote sensing, Vo!. 37, No. 2. 1999.
83. Wang , Y. and X. Gong, "A neural network approach to
microwave imaging", International Journal of Imaging Systems
and Technology, Vo!. 11 Issue 3, 2001, Pages 159 - 163.
84. Lee, K. C., "A neural-network-based model for 20 microwave
imaging of cylinders", International Journal of RF and
Microwave Computer-Aided Engineering, Volume 14 Issue 5,
August 2004, pp 398 - 403.
85. Chew, W.C. and J. H. Lin, "A frequency-hopping approach for
microwave Imagmg of large inhomogeneous bodies",
Proceedings of the Antennas and Propagation Society
International Symposium, AP-S. Digest, Volume 3, Jun 1995, pp
1610 - 1613.
86. Chew, W.C. and 1. H. Lin, "A frequency-hopping approach for
microwave imaging of large inhomogeneous bodies", IEEE
Microwave guided wave letters, Vot. 5, 1995, pp 439-441.
87. Marklein, R., Balasubramanian, K., Qing, A. and Karl J
Langenberg, "Linear and nonlinear iterative scalar inversion of
multi-frequency multi-bistatic experimental electromagnetic
scattering data", Inverse Problems, Vo1.17, 2001, pp 1597-1610.
88. Belkebir, K., R. E. Kleinman and C. Pichot, "Microwave
imaging-Location and shape reconstruction from multifrequency
scattering data", IEEE Transactions on Microwave Theory and
Techniques, Vol. 45, Issue 4, Apr 1997, pp 469 - 476.
89. Semenchik, V., V. Pahomovand S. Kurilo, "Multifrequency
microwave imaging", Proceedings of the Fifth International
Kharkov Symposium on Physics and Engineering of Microwaves,
Millimeter, and Submillimeter Waves, 2004. MSMW 04, Vol. 1,
pp 193- 195.
90. Belkebir, K. and M. Saillard, "Special section: Testing Inversion
Algorithms against Experimental Data", Inverse Problems,
Institute of Physics Publishing, vol. 17, 2001, pp 1565-1571.
91. Cui, T. J., W. C. Chew, A. A. Aydiner and S. Chen, "Inverse
Scattering of Two- Dimensional Dielectric objects buried in
Lossy Earth using the Distorted Born Iterative Method", IEEE
Transactions on Geoscience and Remote Sensing, Vol. 39, No:2,
February 2001, pp 339-346.
92. Cui, T. l., Y. Qin, G. L. Wang and W. C. Chew, " Low
Frequency detection of two-dimensional buried objects using
high-order extended Born approximation", Inverse Problems, 20,
2004, S41-S62.
93. Toshiyuki, T. , T. Takashi and H. Jia, "Microwave lmaging of
Buried Objects using Time Domain Data",
Institute of
Electronics. Information and Communication Engineers
Technical Report, Vol.l00,2001, No. 583, pp 97-103.
94. Massa, A., M. Pastorino and A. Randazzo, "Reconstruction of
two-dimensional buried objects by a differential evolution
method", Inverse Problems 20, 2004, S 135-S 150.
95. Chang ,M., P. Chou and H. Lee, "Tomographic mIcrowave
imaging for nondestructive evaluation and object recognition of
civil structures and materials", 29th Asilomar Conference on
Signals. Systems and Computers (2-Volume Set). p. 1061.
96. Vertiy, A. A. and S. P. Gavrilov, "Mode ling of Microwave
Images of Buried Cylindrical Objects", International Journal of
Infrared and Millimeter Waves, Springer New York, Vol. 19,
Number 9 September, 1998, pp 1201-1220.
97. Guillermin, R., P. Lasaygues, 1. P. Sessarego and A Wirgin,
"Imaging an object buried in the sediment bottom of a deep sea
by linearized inversion of synthetic and experimental scattered
acoustic wavefields", Inverse Problems, 16,2000, pp 1777-1797.
98. Chaturvedi. P. and R. G. Plumb, "Electromagnetic Imaging of
Underground Targets using Constrained Optimization", IEEE
Transactions on Geoscience and Remote Sensing. Vol. 33, No. 3,
pp 551-561, May 1995.
99. Baussard. A, E. L. Miller, D. Lesselier, "Adaptive Multiscale
Reconstruction of Buried Objects", Inverse Problems, 20, 2004,
SI-SIS.
100. Helaoui, L., 1. B. H. Tahar and F. Choubani, "Multi-source
illumination approach for Buried Objects Exploration",
Proceedings of the Second International Conference on the
Digital Society, 2008, pp. 146-149.
CHAPTER THREE
Formulation of the Problem
3.1 Introduction
In the inverse scattering problem of Electromagnetic (RF) and
Microwave imaging, one attempts to infer the properties of the
scatterer which is located in an investigation domain D
Obl
that includes
all possible locations of the scatterer, from the scattered field
measurements which are made in the measurement domain D\f ' which
is outside the investigation domain. The forward scattering problem,
where the object and the incident field that illuminates the
investigation domain are assumed to be known and the measured
scattered field is assumed to be unknown, maybe modeled using
integral equation formulations. The solution to this forward scattering
problem is needed in many inverse scattering techniques and also for
generating the simulated synthetic measured scattered field data for the
inversion. The formulations leading to the mathematical modeling of
the forward scattering problem is discussed next. In this work the
imaging of two dimensional dielectric scatterers is considered. Hence
the discussions are limited to the case of the unknown object whose
tomographic image is to be reconstructed being a 2-D dielectric
scatterer.
Integral equation formulations have attracted the interest of
mathematicians for a long time. Most integral equations do not have a
45
closed form solution. However they can often be discretized and
solved on a digital computer. Proof that a solution exists by
discretization of the integral equation was first presented by Fredholm
I. The volume integral method is employed for the modeling of the
forward scattering problem in microwave imaging 2. The volume
integral equation offers a good physical picture of the mechanism that
gives rise to scattering, providing an insight as to how approximate
scattering solutions can be obtained. In this method, the unknowns in
the problem are expressed in terms of volume currents flowing in the
inhomogeneity. The volume current consists of conduction as well as
displacement currents induced by the total electric field. An integral
equation can then be formulated from which the total field is solved.
Certain physical phenomena maybe expressed by the scalar
wave equation, such as the acoustic waves and the Schrodinger waves.
Certain electromagnetic problems also maybe described by the scalar
wave equation. In the case of a two dimensional scatterer in the
investigation domain, the electromagnetic scattering can be modeled
by a scalar wave equation when the incident field used for illuminating
the scatterer is TM po larized as in figure 3.1. By a two dimensional
scatterer, one refers to the fact that the complex permittivity does not
vary along one of the three Cartesian coordinate axes.
When the incident wave is TM polarized, the magnetic field is
transverse to the axis of the scatterer, or in other words, the electric
field is parallel to the cylinder axis.
11D-TMI
;;t _
~
( D ' ~ J
Figure 3.1
Two dimensional dielectric cylinder
illuminated by a TM polarized source
The scalar wave equation is considered here, for the case of the
two dimensional scatterer illuminated by TM polarized
electromagnetic wave. In the equations, a time factor of ej(l!/ is
understood. In a homogenous medium containing a source q(r) , a
scalar wave lIf(r) (electric field component parallel to the cylinder
axis), satisfies the homogenous wave equation of the form
(3.1)
where the wave number kb represents the wave number of the
background medium, i.e., the spatial frequency of the plane wave and
is a function of the wavelength Ab or
k
2;r 2
b=-=aJjibeb'
Ab
(3.2)
where jib being the permeability and e
b
being the complex
permittivity.
For imaging, an inhomogeneous medium is of interest and
hence the more general form of the wave equation is considered,
(3.3)
where k2(r)=aJ
2
(r)ji(r)8"(r) , aJ being the angular frequency of the
incident wave that illuminates the scatterer, ji being the permeability
and 8" being the complex permittivity. Here e(r) represents the
inhomogeneous medium over a fmite domain Doh} Both medium and
the scatterer are considered to be non magnetic, so ji ::: jih ::: I ,
everywhere in the investigation domain and measurement domain.
The equation (3.4) is rewritten as
(3.4)
The right hand side of equation (3.4) can be considered to be
an equivalent source. An integral formulation for finding the unknown
total field for the forward problem can now be formulated in terms of
the Greens function. The Greens function of a wave equation is
defined as the solution of the wave equation for a point source. When
the solution of a wave equation due to a point source is known, then
the solution due to a general source can be obtained by the principle of
linear superposition.
The Greens function is given by
(3.5)
with g(r,r') being the solution for the wave equation for a point
source at r'. By reciprocity, g(r,r')=g(r',r). Any arbitrary source
q(r) maybe represented as
q(r) = f dr' q(r') o(r - r') . (3.6)
This is actually a linear superposition of point sources in mathematical
terms. Hence the solution to equation (3.1) is therefore
I,I/(r)=- J dr' g(r, r')q(r - r'), (3.7)
which is an integral linear superposition of the solution of equation
(3.5). Since the Greens function corresponding to the differential
operator on the left hand side of equation (3.5) is known, by the
principle of linear superposition one can write
I,I/(r) = - J dr'g(r,r')q(r') + f dr'g(r,r')[e(r')-k/]v,/(r'). (3.8)
In the above equation the first term is just the field due to the
source in the absence of any inhomogeneity, and hence is the incident
field (i.e., the component of the incident electric field parallel to the
cylinder axis). Therefore the above equation can be rewritten as
(3.9)
In the above equation, if the total field IIf(r') inside the
investigation domain, which contains the inhomogeneity, is known,
then IIf(r) maybe calculated everywhere. However IIf(r) is unknown at
the point r'. Therefore this equation is a Fredholm integral equation of
the second kind, where the unknown quantity IIf(r) is both inside and
outside the integral operator. The second term on the right hand side of
the equation (3.9) IS the scattered field due to the
inhomogeneity e(r)-k
b
2
For the case of a two dimensional dielectric
scatterer (the axis of the cylinder being oriented in the z- direction)
illuminated by a TM polarized electromagnetic wave, the Greens
function is the solution to the equation
(V2+e (rg(r,r')== - 6(x - x')6(y - y'). (3.10)
It can be shown that the solution to the above equation is given by
(3.11 )
where Ho
2
is the Hankel's function of the zeroth order and second
kind, r=[x,y] andr'=[x',y'] 2,3.4.
In three dimensions, the Greens function is given by
eikb! r-r'!
g(r,r')=-, ,.
41l' r - r'
(3.12)
For modeling the general three dimensional electromagnetic
scattering problems, the integral equation is of the form
(3.13)
where E and Eill are the total and incident vector electric fields
respectively and G(r,r')is the dyadic Greens Function, which relates
the electric field vector E(r) to the current density J(r) via
E(r):::: jOJ,ll f G(r,r').J(r')dr'
v
and is given by the dyad
- VV
G(r,r')::::[I +-0 ]g(r,r,.
k,,-
(3.14)
(3.15)
The integral equations may thus be written for the vector integral
equations of the 2-D TE and 3-D cases as
E(r) = E
in
, (r) + f dr '(l + 2 grad jr - r 'I)[k"(r ') - k/]E(r')
(3.16)
and
(3.17)
3.2 Approximate solution of the forward
scattering problem
The solution to the volume integral equation has to be solved for
numerically. This is in general computationally intensive. For many
problems, however, when the scattering from the inhomogeneity is
weak:, it suffices to derive approximate solutions to the scattering
problem. One of the popular approximations is the Born
approximation 5. In the case when e - kb
2
is small, or when the
contrast of the scatterer is weak so that the second tenn of the equation
(3.9) or (3.13) is small when compared to the first term, the total field
maybe approximated by the incident field.
I.e.,
E(r) = EinC(r) ,
(3.18)
where the field could be either the scalar field of (3.9) or the vector
field of (3.13). The volume integral equation therefore maybe
approximated as
(3.19)
This is the first order Born approximation. It is also the first order
approximation in the Taylor's series approximation of E(r) using
(e - kb 2) as a small parameter 2.
The Born approximation assumes single scattering. As per the
equation (3.19), the incident wave enters the scatterer with no
distortion, induces a polarization current proportional to (e - kh 2)Einc
and causes a re-radiation, or scattering. Since the incident field is
unaffected when it gives rise to the scattered field, Born approximation
violates energy conservation. However, because of the symmetry of
the dyadic Greens function, reciprocity is still preserved under the
Born approximation. The constraint on the Born approximation to be
valid for the scalar wave case or the case of a TM wave impinging on
an infinite dielectric cylinder is
(3.20)
Thus for this case the Born approximation becomes exceedingly good
at low frequencies, since the constraint can be met even when fiE > I
r
if kbL 1. However this is not true for the vector electromagnetic
field 2.
3.3 Numerical solution of the forward scattering
problem
Derming k2(r) - k/ as the object contrast function c(r), the integral
equation (3.9) now maybe rewritten as
(3.21 )
fI/"(r}= Do,.$ dr' E DM
(3.22)
The first equation gives the total field in the object domain
D gflj and is called the state equation. The second equation gives the
measured scattered field in the measurement domainD
M
and is known
as the observation equation. There is no restriction to the shape or
dimension of D ob} or D
M
The only condition to be satisfied is that
D ob} should be large enough to include all possible object locations. It
is usually considered to be a square or a rectangle for the ease of
computations. The measurement domain should be outside the
investigation domain. A typical measurement setup is shown in figure
3.2. The use of the moment method can transform the integral relations
to matrix relations 6.
.-
@-
:r
Figure 3.2
Typical measurement selUp
By choosing pulse basis functions and point matching, the
object domain maybe divided into N elementary square shaped cells.
In each of the cells. the total electric field and complex permittivity are
assumed to be constant. The relation between the values of the
scattered field at the measurement domain (which is assumed to have
N
r
measurement points for a particular transmitter antenna) and the
values of the total field in the N elementary cells is
.-
iI" . (r.) = g ( r ri )c(ri)iI'.(ri ).m = 1.2 ..... N,
(3.23)
j=1
while the total electric field m the N cells is the solution of the
N
'l/
v
inC
(r,,) = - g(r
n
, rj }c(rj }]'l/V(r
j
}, n = 1,2, ... N
)=1
(3.24)
where v represents the view in a multi incidence configuration and the
object is successively illuminated from different angles by different
incident fields. This is done to collect a larger amount of measured
scattered field data, which will aid in reconstruction of better profiles
of the scatterer. '1/'\',. represents the measured scattered field for the
view v, '1/1' and 'I/"inc represent the total and incident electric fields in
o ={l n:t=J'
h
.. (b' ) d . d ni' Th
t e mvesttgatIon 0 omam an' . e contrast
= {O,n = j
vector c(r) and the total field vector '1/1' (r) are lexicographically
arranged as in figure 3.3.
! J .\'1! .\'1-1 .\'1
.\'1+1 .\'1+: SI+)
Figure 3.2
Lexicographical ordering of a 2-D array into a vector
The array is of dimension NI by N2 and the ordered
vector will be of dimension
N=Nl xN2.
The equations (3.23) and (3.24) maybe respectively rewritten
in the form of matrix equations
[Ij/"l'] = [Grv.om][ C]( Ij/v]
(3.25)
(3.26)
Here [lj/J v] is the measured scattered field vector of
dimension N
r
and [Grv.om] is the integrated Greens function matrix of
dimension N
r
x N that relates a point in the investigation domain to a
point in the measurement domain, for the view v. [Gr.",] is the
integrated Greens function Matrix when both points are in the
investigation domain. [C] is a diagonal Matrix of dimension N by N
such that C(i, i) = c(i), i = 1, 2 ... N . [Ij/,.] and [Ij/"i/I<] are total and incident
field vectors respectively, formed by lexicographically arranging the
electric field values in the investigation domain. The singularity of the
Greens function matrix [Graa] when both points coincide is integrated
by considering the cell to be disk shaped instead of a square.
The analytical expression employed for the integration at the
singularities, which is when the argument of the Greens function is
zero, is given by 7
211" a
if f
H
o\k
b
P)p'dp'drfJ,=_Ld;rk
b
aH,2(k
h
a)-2j]. (3.27)
00 2 ~
where the radius of the cell a is chosen equal to the square side of the
cell divided by Jj since the integrals over a square and a circle of any
polynomial function of order 3 are then exactly related 8.
Another case is when the object lies in an inaccessible medium
and the measurement is done in another medium, as done in the
inverse profiling of buried objects. The development of imaging
techniques for investigation of physically inaccessible objects is of
much importance in areas such as oil exploration, seismic imaging,
non destructive evaluation, buried object detection etc. In most of
these cases, reliable information about the physical properties of the
inaccessible target is desired in addition to detection of the target. All
these problems may be classified as electromagnetic inverse problems
when the characterization of the unknown target is computed by
considering electromagnetic illumination and employing the values of
the scattered electromagnetic field. Here the forward problem is
formulated employing the domain integral equations, where
If/(r) = If/inc (r) + J c(r 19 m
2
m2 (r, r 11f/(r 1dr', rE D
nb
; (3.28)
D(Jhi
is the state equation and
(3.29)
is the observation equation, where the object whose profile is to be
formed is located in the inaccessible medium with permittivity Em2
(for example, soil), and the measurements are made in the medium
with permittivity Em] (usually air). The investigation domain lies
entirely in the inaccessible medium. When the investigation domain is
selected as a rectangle (which is' the usua"l case, for the ease of
computations), and when the scatterer is a 2-D dielectric scatterer
whose axis is oriented in the z- direction and illuminated by a TM
electromagnetic source in the z- direction (i.e. the cross section of the
scatterer lying in the x-y plane), the expression for the Greens
(3.30)
where rand r' lies in Doh; and
(3.31)
T
m2m
, are reflection and transmission coefficients of a TM wave
. I . b 9 10
respectIve y, gIven y .
(3.32)
In the above equations, the rectangular investigation domain is
assumed to begin from a depth of h from the interface between media
mj and m2. The observation and state equations are discretized using
pulse basis functions and point matching method of moments. The
corresponding matrix representations for the state and observation
equations for a transmitter view v maybe respectively written as
(3.33)
and
(3.34)
where Wv , W:
nc
and c are the total electric field vector, the incident
field vector and the complex permittivity contrast vector respectively,
of dimension N x I ,all in lexicographical order, N being the number of
cells in Dob}' Grmcm, and Grv.m,m, are the matrices of properly integrated
Greens functions of dimension N x Nand N
r
x N respectively where
N
r
is the number of receivers in DM for a particular view v. C is the
diagonal matrix of dimension N x N whose entries are the elements of
the contrast vector c 11
3.4 Inverse scattering techniques
The methods to solve the inverse problem of Electromagnetic
(RF) and Microwave imaging could be broadly classified as
deterministic and stochastic. They tackle the nonexistence of the
inverse scattering problem by redefining it as the minimization of a
cost functional. The deterministic algorithms seek to solve the inverse
scattering problem by linearizing it around a current estimate and then
seeking progressively better estimates of the solution. However, due to
the nonlinearity of the inverse scattering problem, there is a risk that
the solution gets trapped in a local minimum. The deterministic
algorithms include those that seek repeated exact field computations
for the forward oroblem., such a . ~ Rorn iterative Method (RIM)
Distorted Born Iterative Method (DBIM) and Newton Kantarovich
(NK) Method, and those that are gradient based where successive
approximations of the configuration and the excited fields are obtained
simultaneously 12. However for reconstruction of objects buried in a
half space or multi layered media, gradient based techniques can
reconstruct only a single profile (either permittivity or conductivity),
or the background has to be constrained as lossless 9. The stochastic
methods, though globally convergent, are extremely computation
intensive, especially when the number of unknowns is high. In this
work, the DBIM and NK techniques have been employed for the
inverse scattering problem of electromagnetic imaging and are
discussed next.
In the inverse scattering experiment of electromagnetic
microwave imaging, only the measured scattered field vector is
available for the inverse profiling of the scatterer, given by
!jIS (r) = !jI(r) _1//iI1C (r), rE D,H , (3.35)
where the scatterer lies entirely in D obj' The observation equation
cannot be directly solved for the cross sectional profile of the
object kl (r), since the total electric field in Doh; , ljI(r) ,is also
unknown and is a function of k
2
(r) itself. Therefore the integral is a
non linear functional ofe(r). Thus there is a non linear relationship
between the scattered field and the scattering object. The physical
reason of this non linearity can be attributed to multiple scattering
effects within the scattering object. The multiple scattering, and hence
the non linearity of the inverse scattering problem, increases when the
object contrast and size increases 13,14.
The problem can be linearized by using the Born approximation as,
'l'S (r) = D"nJ dr' g (r, r ') c(r ') 'l'me (r '), rE DM
(3.36)
The above equation is now linear inc and can be solved for the object
contrast. The error in the above equation can be shown to be of the
order of c
2
However the Born approximation is poor when the
scattered field is not weak when compared to the incident field. This
occurs when the dielectric contrast of the scatterer with respect to the
surrounding medium is large or when the scatterer is large compared
to the wavelength.
One way to overcome this problem is to seek increasingly
better estimates of'l'(r). For this, the zeroth-order estimate to c(r) is
found fIfst using the Born approximation. This estimateco(r) is then
used to calculate the zeroth-order estimate of the scattered
field, 'l's o(r). The estimated object contrast is then employed in the
state equation, together with 'l'ine (r) to obtain a better estimate of the
total field If/(r) within the scatterer. The new total field is denoted
as '1'1 (r) = 'l'ine (r) + 1//
0
(r). By so doing, a new approximation for the
scattered field may be obtained, given by
'1" (r)= D",J dr' g(r,r ')c(r') '1'1 (r'), rE DM
(3.37)
With 1//
1
(r) known, equation (3.37) is a linear integral equation
in c(r) and hence, can be solved as a linear inverse problem. This
method yields a new object contrast, denoted byct(r), which is then
used to obtain a better estimate of v/(r) within the scatterer. This
procedure is repeated until convergence is reached. This procedure is
called the Born iterative method (BIM), for solving the inverse
problem. This method improves over the simple Born approximation
and allows scatterers of relatively large size or high contrast to be
constructed 15.
While the Born iterative method is simple to implement, it may
not be the best method to solve the nonlinear inverse-scattering
problem because it does not offer second-order convergence. An
improved method with second-order convergence is the distorted Born
iterative method (OBIM).With DBIM, the background medium is not
constrained to be homogeneous and is updated in each iteration
l6
. This
creates the added difficulty that the Green's function is not available in
closed form and must be sought via a numerical method. However,
DBIM usually converges faster than BIM and works better for objects
with large contrast. This advantage is tempered by the more robust
performance of BIM in the presence of noise. Nevertheless, a
convergent solution can be achieved with DBIM.
The expression for the total field is now expressed as
I/f(r) = I//inc.b (r) + D f dr 'gb (r, r')[ e (r ') - k/ ]v/ (r')
"hI
(3.38)
where V/inc.b(r)is the incident field when the background kh(r) is not
constrained to be homogeneous. It is to be noted that
f//inc,b (r) =f//inC (r) + (1J2p J dr' go (r, r ') eh (r f//inc,b (r ') (3.39)
where go(r,r')is the Greens function with the background assumed to
be homogeneous. The scattered field due to the difference between
e(r) and k/(r) in the object domain is denoted by IJIs.h(r) and is
given by
IjIS,b (r) = D".,J dr' gb (r, r ') c(r ') If/(r ,), rE D,o,-r
(3.40)
where,
(3.41)
Once e
b
is assumed known, ljIil1c.b (r) is calculated by solving
(3.39). If the total field is measured at the receivers, then 1jI"b (r) is
obtained at the receivers via (3.41). Essentially, the left-hand side of
(3.41) is known at the receivers via measurement and solving (3.39)
for lJIil1c.b (r) .
The goal of the inverse problem is to solve for c(r) from
which6(r), the cross sectional complex permittivity distribution of the
2-D dielectric scatterer, can be found. However, from (3.40), it is clear
that this is a nonlinear inverse-scattering problem because ljI(r) is also
a functional of c(r) , as is clear from (3.38). But if 6(r) is close
to 6
b
(r) , then IjIs,b(r) '1/
nc
,h(r) as is evident from (3.38),
Thereforelfl(r)is approximated with lfI
iI1C
,b(r) in (3.40) and linearized
equation is
IfIS,h(r)= D".J dr'gb (r,r')c(r')lfI
ine
.
h
(r'),r E D,\1
(3.42)
This approximation is called the Distorted Born approximation or the
Distorted-wave Born approximation. Ifc(r) is small, it can be easily
shown that the error in this approximation is of order c
2
(i.e., second
order) 17.
The linearized integral equation can be numerically solved to
obtain c(r), the cross sectional contrast of the scatterer with respect to
the background. However, the profile thus obtained is not accurate
because of the distorted Born approximation. This
approximate c(r) can be used to improve on the approximated object
cross sectional profilet:(r) . Thent:(r) can be used as the new t:b(r).
With this new t:b(r) ,the If/S.b(r) , lfIine.b(r) and gb(r,r') in (3.42) are
recalculated and updated. Next, a new c(r) is obtained and the process
is repeated until convergence is attained. Since the error in (3.42) is of
second order, this method has second-order convergence. This method
of solving the inverse-scattering problem is called the distorted Born
iterative method. It is to be noted that in the DBlM, the left side of
equation (3.41) is updated in every iteration as If/inc.b(r) is recalculated
in each iteration step. Therefore If/,.h (r) becomes smaller as
k/(r) approaches k2(r) as the iteration step increases. If the
measurement data is contaminated with noise, so that
\If(r) = \If(r)+n(r)
where n(r) is the noise, then
_ s,b
\If (r)=\If(r)-\lfinc.b(r)+n(r)
(3.43)
(3.44)
is the updated scattered fieId in each iteration step. Thus as the
iteration step improves on v/nc.h(r) so that it is closer to \If(r) ,
S ,b
\If (r) is swamped by the noise n(r). Thus precautionary steps should
be taken to prevent the iteration steps from diverging, when the
scattered field data is corrupted by large measurement noise.
Thus the difference between DBIM and BIM is that in DBIM,
the Greens function that propagates the field from a point in the object,
r 'to a point in the measurement domain r, is also updated at each
step, in addition to the total field. The error estimate for the DBIM is
of the second order, while that of the BIM is of the first order. So the
DBIM will converge faster than BIM, and will converge to the global
solution in cases where BIM does not.
The state and observation equations are discretized using pulse
basis function and point matching method of moments. For a
transmitter view v , the observation equation may also be written as,
(3.45)
where [IjI,1 is an N x N diagonal matrix such that
66
If/ y, i) :;: V/Ji), i = 1,2, .... N . (3.46)
Here [V/
v
] and [cl are lexicographically ordered vectors of
dimension N . Collecting the measured scattered field from all the
transmitter views, we can write
[V/
I
] = [P][c]
(3.47)
where [P] = and [P;] = [GIj,()m ][If/)-
P;
P
M,
[P;] is of dimension N
r
x N. If we denote the total number of
measurement points in the measurement domain DM as M , where
M :;: N, N
r
, N, being the number of transmitter antennas, [P] is of
dimension M x N .
The solution of(3.47) for[c] is not advised since the matrix[P]
is ill-conditioned and might not have a unique inverse. There are two
main reasons for this ill-conditioning.
1. [P] is an M x N matrix. It is possible for M <N ,if the
number of measurement data points is less than the number
ofpixels used to approximate the object function, then, the
range space of [P] would have a smaller dimension than the
domain, implying that the inverse does not exist.
2. The inverse problem is faced with fundamental scattering
physics that contributes to its nonuniqueness. [P] maps the
object function [c], to the measurement data, If'. Many
high spatial frequency components ofs(r) are not mapped
to If' because these components generate evanescent
waves that do not reach the receiver location. Therefore,
these fine-grained details are lost in the inversion process.
A technique to overcome this problem is to use the method of
regularization to solve the equation 18 In this method, a functional is
defined as
1=111f' - p.cll' +allcll'
(3.48)
In the above expression, the first term is the error between the
measurement data and that predicted by the linearized model. The
second term is the norm of the contrast vector multiplied by a
constant, a , the Tikhonov regularization parameter. An optimal
solution for c(r) is sought by minimizing the cost functional. The
regularization procedure is employed since the electromagnetic inverse
scattering problem is ill-posed in the Hadamard sense 19. The
Tikhonov regularization parameter is a tuning parameter, chosen to
weigh the relative importance of the first term and the second term. In
68
this manner, the optimal solution will minimize the error between the
measurement data and norm of c(r). Because of this the values of
c(r) obtained will be bounded. If a is chosen too small, the numerical
stability of the solution is compromised. However, if a is too large,
useful information about the solution maybe filtered out. Therefore the
regularization parameter should be carefully selected.
The cost functional maybe expanded as
1 = Ill')" _c'P')(III
S
-Pc)+ac'c (3.49)
where * indicates the conjugated transpose. Taking the frrst variation
of 1 with respect to the first variation in c and then setting it to zero,
the following equation is obtained,
(3.50)
where I is an Identity matrix of dimension N x N . The contrast vector
maybe thus be found out as
(3.51)
The Newton Kantarovich (NK) Procedure is discussed next.
The procedure allows for the easy insertion of a priori information that
can be used to remove unrealistic solutions 20, The NK and DBIM
procedures have been shown to be computationally equivalent 8,12,17.
From the observation and state equations, considering small variations
of the field, the equations become
and
=
Expressing the quantity as
Introducing (3.53) into (3.54), we get
Since [GroJ is a symmetric matrix, we have
(3.52)
(3.53)
(3.54)
(3.56)
This will simplify the calculations as [J - GrnnCr
l
has already been
calculated in the state equation. Substituting (3.55) in (3.52) and
employing the equation (3.56), the variation of the scattered field in
the measurement domain maybe expressed as
(3.57)
Collecting the quantity for every transmitter view, leads to the
fo Howing relation,
(3.58)
where
D=
(3.59)
DN'
and
(3.60)
Here [DJ is of dimension N,. x Nand [D] of dimension
M x N where M = N, NI" The above equation is solved iteratively for
[ ~ c ] using the following relation,
(3.61)
In the above equation, the Tikhonov regularization procedure is
employed since the inverse scattering problem is ill-posed. By
introducing this relationship into an iterative procedure
(Here k indicates the iteration number), the initial non linear relation of
the inverse scattering problem can be transformed into a sequence of
linear ones. Hence, starting at an a priori initial contrastc(I' succession
of intermediate c
k
are generated by minimizing the difference between
the computed scattered field according to the observation equation and
the measured scattered field. The iterative process is summarized as
follows,
1. Computation of the total internal field by inverting the state
equation
2. Estimation of the scattered field at the receiver locations
3. Computation of the error between the scattered field computed
in step 2 and the measured scattered field, ~ f//k
4. Estimation of the error in contrast C
k
by solving (3.61).
5. Updating the contrast function CH = C
k
+ C k
6. Go to step I as long as ~ f//k is larger than an acceptable error.
The linearized matrix equations maybe solved for the object
contrast (in the case of Newton Kantorovich procedure, the update to
the object contrast) using Singular Value Decomposition Technique 21.
As example, consider equation (3.47). Here the matrix [p* P] is ill
conditioned because of the multiple scattering. Since [p* P] is
Hermitian 2, we can decompose it as
[p* P] = S.A.S' (3.62)
where S is a unitary matrix ( i.e. S' = S-I) and A is a diagonal matrix
containing the eigen values of [p* P]. However because of the ill-
conditioning of [p* P], it has eigen values too small or zero.
Therefore in
[p* pr
l
= S.,,1,-I.S', (3.63)
A-I is unbounded. When a regularization procedure is employed,
instead of seeking the inverse of [p* P], the inverse of[P* P +a/] is
sought. The Tikhonov Regularization parameter therefore will pad the
zero eigen values of [p* P]with a non-zero value. It is to be noted that
numerical instability of the inverse process is related to the large
condition number of [P * P]. To decrease the condition number, it is
necessary to filter out the high frequency components which are
affected by the measurement noise, via the regularization procedure.
On the other hand, a strong regularization process may remove useful
components and decrease the spatial resolution because of the low pass
filtering effect. As a consequence, the regularization factor is to be
fixed in order to accommodate a convenient spatial resolution-stability
compromise. In this work, an empirical formula for the regularization
parameter is employed 8
a = P Trace[P* P] IIF(ck ) - wJ
N IIw
s
112
(3.64)
Here F(c
k
) is the computed scattered field with the current estimate of
the t t
II
F
(c
k
)-wsl!2. h I' d l' 'd 1
con ras c
k
, IS t e norma lze re attve rest ua
IIVfJ
error Err ( IfIS) and f3 is a constant that is detennined empirically. The
value of P is initially chosen and is updated in each iteration according
to
j
fJ"i
l
Err(w'J- Err(v/_) < -O.lErr(vr',) .
/3+ = S Err(IIf'J- Err(".'I'_) S O.lErr(IIf',) .
2/
f
e,if Err(v/J - Err(v/_) > O.lErr('1/',)
(3.65)
In the above expression, P+ is the next update and Pc is the current
value. Err ( '1/ c) and Err ( IfIs _ ) are the current and previous normalized
residuals respectively. The constraint is thus weakened when the
process converges too slowly or when the error increases slightly,
which will aid the process to escape from getting trapped in a local
minimum. Even so, an iterative deterministic procedure may fail to
reach a convergent solution because of the non linearity and ill posed-
ness of the inverse scattering problem. The ill-posedness is even more
severe when scattered field measurements are possible only from a
limited number of transmitter views, as in the case of imaging of
buried objects.
References
1. Fredho1m, I. "Su rune classe d'equations fonctionalles" Acta
Math. 27: 365-390, 1903).
2. Chew W.C, Waves and fields in inhomogeneous media, New
York, IEEE Press, 1995.
3. Umashankar K., Taflove A, Computational Electromagnetics,
Norwood, Artech House Inc., 1993
4. Slaney M, Azimi M, Kak A.C, Larsen L E, "Microwave
imaging with first order diffraction tomography", Medical
applications of microwave imaging, pp 184-211, IEEE Press
New York, 1986)
5. Born, M. and E. Wolf. 1980. Principles of Optics. 6
th
edition.
New York: Pergamon Press. First edition 1959.
6. Booton Richard C. Junior, Computational method.. for
Electromagnetics and Microwaves, Wiley lnterscience, New
York, 1992.
7. Richmond, J. H, "Scattering by a dielectric cylinder of arbitrary
cross section shape", IEEE Transactions on Antennas and
Propagation, Vol. AP- 13, pp 334-341, 1965.
8. Franchois A and C. Pichot, "Microwave Imaging-comp1ex
permittivity reconstruction with a Levenberg-Marquardt
Method", IEEE Transactions on Antennas and Propagation,
Vol. 45, no.2, pp 203-215, February 1997.
9. Cui, T. 1., W. C. Chew, A. A Aydiner and S. Chen, "Inverse
Scattering of Two- Dimensional Dielectric objects buried in
Lossy Earth using the Distorted Born Iterative Method", IEEE
Transactions on Geoscience and Remote Sensing, VoL 39,
N (l'') nn T ~ Q - 1 4 n Ff'hnlltrv ?()() 1
10. Cui, T. l., Y. Qin, G. L. Wang and W. C. Chew, " Low
Frequency detection of two-dimensional buried objects using
high-order extended Born approximation", Inverse Problems,
20, S41-S62, 2004.
11. Baussard. A, E. L. Miller, D. Lesselier, "Adaptive Multiscale
Reconstruction of Buried Objects", Inverse Problems, 20, S 1-
S15,2004.
12. Tijhuis, A. G., K. Belkebir, A. C. S. Litman and B. P. de Hon,
" Theoretical and Computational Aspects of 2-D Inverse
Profiling", IEEE Transactions on Geoscience and Remote
Sensing, Vol. 39, No. 6, pp 1316-1329,June 2001.
13. Belkebir, K., R. E. Kleinman and C. Pichot, " Microwave
lmaging- Location and Shape reconstruction from
Multifrequency Scattering Data", IEEE Transactions on
Microwave Theory and Techniques, Vol. 45, No. 4, pp 469-476,
April 1997.
14. Chew, W.e. and l. H. Lin, "A Frequency Hopping Approach
for Microwave Imaging of Large Inhomogeneous Bodies",
IEEE Microwave and Guided Wave Letters, Vol. 5, No. 12, pp
439-441, December 1995.
15. Wang, Y. M. and W. C. Chew, "An iterative solution of two-
dimensional electromagnetic inverse scattering problem",
International Journal C?f Imaging Systems Technology, Vol. 1,
No. 1, pp 100-108, 1989.
16. Chew, W. C. and Y. M. Wang, "Reconstruction of Two-
Dimensional Permittivity Distribution Using the Distorted
Born Iterative Method", IEEE Transactions on Medical
Imaging, Vol. 9, No. 2, pp ~ 1 8 - 2 2 5 , June 1990.
17. Teflove, A., Advances in Compuational Electrodynamics,
Boston, London, Artech House, 1998.
18. Tikhonov, A. N. and V. Y. Arsenin, Solution of ill-posed
problems, Washington D. C., V. H. Winston and Sons, 1977.
19. Hadamard, J., Lectures on Cauchy's problem in linear partial
differential equations, New Haven, C.T, Yale University Press,
1923.
20.10achimowitz, N., C. Pichot and 1. P. Hugonin, "Inverse
Scattering: An Iterative Numerical Method for Electromagnetic
Imaging", IEEE Transactions on Antennas and Propagation,
Vol. 39, No. 12, pp 1742-1751, December 1991.
21. Press, W. H., S. A. Teukolsky, W. T. Vellering and B. P.
Flannery, Numerical Recipes in C, Cambridge University Press,
1992.
F
CHAPTER 4
Multi-scaled frequency hopping
4.1 Introduction
A large class of image processing problems such as deblurring, high
resolution rendering, image recovery, image segmentation, motion
analysis and tomography require the solution of inverse problems.
Often, the numerical solution to inverse problems is beset with several
computational issues, particularly when the problem is highly non-
linear as in the case of Microwave/RF imaging. The non-linearity and
ill-posedness of the inverse scattering problem of microwave imaging
is attributed to multiple scattering, which increases with object size
and contrast I. The algorithms employed for this inverse scattering
problem maybe broadly classified into deterministic or stochastic. All
of them tackle the nonexistence of the inverse scattering problem by
redefining it as the minimization of a cost functional 2-5 However, due
to the nonlinearity of the inverse scattering problem, there is a risk that
the solution gets trapped in a local minimum. The stochastic methods 2
.6.7, though globally convergent, are extremely computation intensive,
especially when the number of unknowns is high. One way to
minimize the risk of local minima is to use prior knowledge about the
scatterer, so that the nonlinear problem is linearized about a different
background 8. However such prior knowledge is not always available.
A technique that has been suggested to improve the convergence of
79
this inverse scattering problem is to collect multi-frequency data
instead of monochromatic data for the inversion 9-13. The inverse
problem is solved at a lower frequency where it is more linear and thus
less prone to local minima, and this solution is used as an initial
estimate for inversion at a higher frequency. The final solution will be
much less prone to local minima than in the monochromatic case of
the highest frequency alone. The convergence of the solution at the
higher frequencies is determined by the initial estimate provided by the
iterations at the low frequencies.
To solve the inverse scattering problem of electromagnetic
imaging, most algorithms work by performing all computations using
a fixed discretization grid. While progress has been made in reducing
the computational complexity of the fixed grid methods 14-1 R,
computational cost is still a big concern. More importantly fixed grid
methods employ optimization techniques that perform a local search of
the cost function, and are therefore more susceptible to being trapped
in local minima that can result in poorer quality reconstructions.
Multi resolution techniques have been widely investigated to
reduce computational burden of inverse problems and to ensure their
converging globally. Simple multi resolution approaches such as
initializing fine resolution iterations with coarse resolution have been
shown to be effective in many imaging problems 19-23. Multi resolution
methods have been employed in the inverse scattering problem of
microwave imaging, which considerably reduces the risk of the trial
solution getting trapped in a local minimum with respect to the
deterministic algorithm employed 24,25. However, since the solution to
the inverse electromagnetic scattering probleQl is found by minimizing
a cost functional, it is possible that local minima may again be
reintroduced, even though the solution is less likely to be trapped there
compared to fixed grid methods.
The approach employed in this work is to combine frequency
hopping with a multi-scale method. At each frequency hop, the inverse
problem is solved at a coarse resolution. This solution is used as an
initial estimate for the next finer scale and so on until the inverse
problem is solved at the finest scale for the frequency employed. The
final solution for this frequency is used as the initial estimate for the
next higher frequency hop. The process is repeated until the solution is
found for the finest scale in the highest frequency employed. The
method employed for inverse profiling is the Newton Kantorovich
algorithm 5,8, but the strategy is independent of the deterministic
optimization method employed. The strategy has been tested on
synthetic data and the Marseille experimental data 26 and the results
are promising.
4.2 Problem statement and solution
A cylindrical object of arbitrary cross section surrounded by a
homogenous exterior medium of known complex permittivity E
exl
'
under TM harmonic excitation is located in a rectangular imaging
domain I . Usually water or saline solution has been employed as the
medium for microwave medical imaging for enhancing the coupling of
electromagnetic energy. Recently, sodium meta silicate gel, which is
significantly less lossy than water or saline in the ISM band, has been
proposed as a suitable coupling medium for microwave medical
imaging 27 A cylindrical geometry for the imaging system is
employed, with M line sources equispaced on a circle in the
measurement domain. At a time one of the antennas will be emitting
and the others will be receiving. The scattered field is measured for
different views. A time convention of e
jM
is employed. The integral
equations for modeling the forward and inverse scattering are
reproduced here.
lfI(r)=II/
nC
(r)+ D"rJ dr'g(r,r')c(r')lI
l
(r'),r E D
ob
) (4.1)
1fI'(r}= D"bJ dr'g(r,r')c(r')II
I
(r'),rE DM (4.2)
Nonexistence of the inverse solution is coped by redefining the
solution as a minimizer of the least square error with respect to data. In
practice, one looks for a spatially discretized vectorc(r), the
lexicographically ordered representation of the two dimensional image
contrast. This alleviates non uniqueness to a good extent. Thus it is
evident that the stability of the solution will be better when the
discretization cell size is not too small 8. Hence, non uniqueness is less
prominent at a coarse resolution. However, the use of a cost functional
may again introduce local minima. The integral equations are
discretized using the Richmond's procedure 28. The imaging region is
discretized into N cells of dimension corresponding to the particular
scale. The coupled equations (4.1) and (4.2) are solved for the current
estimate of c .A regularization procedure is employed since the matrix
employed in the inversion is ill conditioned 29. An empirical formula
for the regularization parameter has been employed 5,8 as per equation
(3.65). The inverse problem is first. solved at a coarse spatial sampling
rate where it is more stable. This solution is chosen as the initial
estimate for solving the problem at progressively finer sampling rates.
The fmal convergent solution is used as the initial estimate for the next
frequency hop. The process is continued until a convergent solution is
found for the highest frequency hop. The entire procedure employed is
illustrated in the following flow chart.
B
I o w ~ s t scale
i. t.nffS fJftH .,,,,,.I.1IgI1I
Figure 4.1
Flow chart illustrating the multi scale frequency
hopping strategy
4.3 Results and Discussion
The strategy was tested on synthetic data as well as
experimental data provided by the Centre Commun de Ressourses
Micro-ondes (CCRM) of Marseille France. The coupled equations
(4.1) and (4.2) were used to generate synthetic broadband data for the
inversion. The following frequencies have been employed for the
incident field: 10Hz, 20Hz and 40Hz. The problem is solved at a
coarse resolution of A-
S
at 10Hz and then at finer resolutions of ~
10
A-
and as per the multiscale optimization scheme. Thereafter, it is
20
solved at resolutions ~ and ~ at 20Hz and 40Hz. For frequency
10 20
hopping without muitiscaling, the inverse scattering problem is solved
at the highest resolution of ~ at every frequency. In the frequency
20
hopping technique, the solution from the lower frequency
measurements is used as the initial estimate for solving the problem at
a higher frequency. The stability of the solution is better when the
discretization of the inverse scattering problem is coarse. This when
employed as the initial estimate for solving the problem at a smaller
discretization cell size, will give a more stable solution. Thus the
multiscaling approach reduces the risk of the solution from the lower
frequency measurements getting trapped in a local minimum, with
respect to the deterministic optimization method employed.
At each stage, the iterations are stopped when the residual, the
error between measured and computed scattered fields, is below a
predetermined tolerance level of 2 x 10'2. The proposed scheme shows
84
an improvement over frequency hopping for some of the test object
configurations, as illustrated in the examples in figures.
In the first example the results of the strategy when tested on
the CCRM experimental data are shown. Here the scatterer
configuration consists of two dielectric cylinders of diameter 3 cm
spaced apart by 9 cm. The scatterer is illuminated by TM polarized
EM wave. For the reconstructions, the scattered field values at I, 2 and
4 GHz were considered. The incident field is modelled by an infinite
line source. The set of images in figure 1 shows the results obtained by
employing frequency hopping alone. The iterations at each step are
terminated when the normalized residual error Err ( ifS) reduces below
2 x 10-
2
. The results in figure 2 show the reconstructions when the
proposed multiscale frequency hopping strategy was employed.
Comparing the final convergent images in figure 1 and figure 2, it is
observed that the two dielectric scatterers are better localized with the
proposed strategy. [t is also seen that the number of computational
steps required for the normalized residual to decrease below the set
value is lesser. The final convergent image was obtained after 20
iterations of the Newton Kantarovich procedure at 1 GHz where the
imaging region is discretized into a 32 x 32 grid, 10 iterations at 2
GHz where the imaging region is discretized into a 64 x 64 grid and 8
iterations at 4 GHz where the imaging region is discretized into a 128
x 128 grid. For the multiscaled frequency hopping strategy, for the
frequency 1 GHz, the number of iterations at A is 18, those at ~ IS
5 10
16 and a t ~ t h e number of iterations is 16. At 2 GHz, the number of
10 '
iterations at . ~ is 10 and at ~ ~ the number of iterations is 9. Likewise
at 4 GHz, the Newton Kantarovich procedure converges in 6 iterations
at ~ and in 6 iterations at ~ The number of computations per
10 20
iteration is of the order of N
3
when the Singular Value Decomposition
procedure is adopted to solve the Matrix formulations for equations
4.1 and 4.2. For the frequency hopping scheme, at 1 GHz, the order of
computations is 0(20 x (32 x 32)3) = 0(2.147 e I 0), whereas for the
proposed strategy, this IS
0(18x(8x 8)3 + 16x(16x 16)3 + 16x(32 x32)3) = 0(1.7453el0). The
corresponding values at 2 GHz are 0(6.871gell) and
O(6.2814ell) respectively. Similarly the corresponding values at 4
GHz are O(3.5184e13)and O(2.680Ie13)respectively. Comparing
the above values, it is observed that the computations are reduced at
every frequency hop with the proposed strategy.
The synthetic data for inversion were generated by using the
same configuration for the imaging setup as were employed for the
CCRM experimental data. The figure 3 shows a much stronger
scatterer than in the previous example, which resembles the human
ann 27. It is observed that the reconstructions with the proposed
strategy are better than the frequency hopping scheme in
reconstructing the features of the scatterer, as shown by final
convergent images in the figures 4 and 5. The cylinders are better
localized in the MSFH scheme compared to the frequency hopping
scheme. This is because the multiscaling strategy at the lower
frequencies succeeded in providing a better initial estimate for the
iterations at the higher frequencies for this object, than when the
solution was computed at the frnest scales alone.
- -----.
F ....
". :i!:: I
'i::. ..
,
' 1 ...
. :.::!' .... -
... . ... ' .. .. ....... .
:; I =
.. . . .' .. --.
... : -: - ' j IJ:
_ .. .. .. __ .... --
1 , & '-'\, .
:: ... : . , I ' . "
00- - . _ . ,-,
: .",," ::
....
... ..
-r
... _ .... ..
Figure l.a
The reconstruction after the
first Newton Kantarovich
iteration al I GHz
Figure Le:
The reconstruction after the
third Newton Kantarovich
iteralion al 2 GHz
;,
Figure I.b
The final convergent image
with the Newton Kantarovich
iteration at I GHz
Figure l.d
The final convergent image
with the Newton Kantarovich
iteration at 2 GHz
Figure t.e
The reconstruction after
the fourth Newton
Kantarovich iteration al
4GHz
Figure I.e
The final convergent image at the eight
iteration of the Ne\\10n Kantarovich
procedure at 4 G Hz
Figure 2.a
The final convergent image at
..l
-aI
5
the IS!h NK iteration at I GHz
.... ,,.,.,..
Figure 2.c
The final convergent image at
the 16!b NK iteration at I GHz
..
..l
-at
20
Figure 2.e
The final convergent image at
..l
- at the 9!h
20
Figure 2. b
. ..l
The final convergent Image at - at
10
the 16!h NK iteration at I GHz
Figure 2.d
The final convergent image at
the lO!" NK iteration at 2 GHz

Figure 2.f
The final convergent image at

..l
- at
10
..l
- al the 6' 1>
10
Figure 2.g
A
The final convergent image at - at the 6
111
NK iteration at 4
20
GHz. The total number of computations has becn reduced for
the relative error to reduce below the specified tolerance. The
reconstructed image is found to be slightly better with the
proposed scheme, compared to figure l .c.
figure 3.a
Real part of the comple"
permitti vity profile of the
hand like scatterer
Figure 3.b
Imaginary part of the comple"
permittivity profile or the
hand like scatterer
Figure 4.8
Reconstructed rcal part of thc complex
permittivity profile of the scatterer using the
frequency hopping scheme at the highest scale
Figure 4.b
Reconstructed imaginary part of the complex
permittivity profile of the scatterer using the
frequency hopping scheme at the highest scale
Figu ... e S.a
Reconstructed real part of the complex permittivity profile
of the scatterer using the proposed scheme. The features of
the scatterer are better reconstructed compared to figure
4.>
Figure S.b
Reconstructed imaginary part of the complex permittivi.ty
profile of the scatterer using the proposed scheme. The
features of the scatterer are better reconstructed cQmpared
to f i ~ u r e 4.b. The reconstruction artifacts have also been
elerences
1. B010mey, lC. and C. Pichot, "Microwave tomography, from
theory to practical imaging systems", International Journal of
Imaging systems Technology, Vol.2, 1990, pp 144-156.
2. Garnero, L., A. Franchois, 1. P. Hugonin, C. Pichot and N.
Joachimowicz, "Microwave Imaging, Complex permittivity
reconstruction by simulated annealing", IEEE Trans.
Microwave Theory Tech., Vo\. 39, pp 1801- 1807, Nov. 1991.
3. Moghaddam, M. and W.e. Chew, "Non linear two-dimensional
velocity profile inversion in the time domain", IEEE
Transcations on Geoscience and Remote Sensing, Vol. GE-30,
1992, pp 147-156
4. Chew, W.c., and Y. M. Wang, "Reconstruction of two-
dimensional permittivity distribution using the distorted Born
iterative method", IEEE Trans. Med. Imag., vol. 9, pp 218 -
225, June 1990.
5. Joachimowicz, N., C. Pichot and J. P. Hugonin, "Inverse
Scattering: An Iterative Numerical Method for Electromagnetic
Imaging", IEEE Trans. Antennas Propagat., vol. 39, no. 12, pp
1742 - 1751, Dec. 1991
6. Pastorina, M., A. Massa and S. Caorsi, "A Microwave inverse
scattering technique for image reconstruction based on a
genetic algorithm", IEEE Trans. Meas. Instrum., Vol. 49, pp
573 - 578, June 2000.
7. Chen, X., D. Liang and J. Huang, "Microwave imaging 3-D
buried objects using parallel genetic algorithm combined with
FDTD technique", Journal of Electromagnetic waves and
applications, Vol. 20, No. 13, December 2006
8. Franchois. A. and C. Pichot, "Microwave lmaging - Complex
permittivity reconstruction with a Levenberg - Marquardt
method". IEEE Trans. Antennas Propagat., vol. 45, no. 2, pp
203 -215, Feb. 1997.
9. Chew, W.e. and 1. C. Lin, "A Frequency Hopping approach
for microwave imaging of large inhomogeneous bodies", IEEE
Microwave and Guided wave letters, vol. 5, no.12, pp 439 -
441, Dec 1995.
10. Belkebir, K., R. E. Kleinman and e. Pichot, "Microwave
imaging - Location and Shape Reconstruction from
Multifrequency Scattering Data", IEEE Trans. Microwave
Theory Tech., vot. 45, noA, pp 469 - 475, April 1997.
11. Crocco, L., "Inverse scattering with real data: detecting and
imaging homogeneous dielectric objects", Inverse Problems,
17, 1573-1583,2001.
12. Semenchik, V., V. Pahomov and S. Kurilo, "Multifrequency
Microwave Imaging", Proceedings of the Fifth International
Kharkov Symposium on Physica and Engineering of
Microwaves, Millimeter and Sub-millimeter waves, Vol. I, pp
193-195, 2004.
13. Kim, 1. H. and Ra, Jung-Woong, "Multifrequency microwave
imaging of a lossy dielectric cylinder in a lossy medium",
Microwave and Optical technology letters, Vol. lO, issue I, pp
26-31, 1995.
14. Lin, 1. H. and W.e. Chew, "Three dimensional mIcrowave
imaging with local shape function method with CGFFT", IEEE
Antennas and Symposium AP-S Digest, Vo!. 3, Issue 21-26, pp
2148-2151, July 1996.
15. Lu, C.C. and X.G. Zhong., "Image reconstruction of two
dimensional objects inside dielectric walls", Microwave and
Optical technology letters, Vol. 36, Issue 2, Pages 91-95,
December 2002.
16. Abubakar, A. and P. M van den Berg, "The contrast inversion
method for location and shape reconstruction", Inverse
Problems, Vol. 18, pp 495-510, 2002.
17. Mia, J., W. C. Chew, C.C. Lu, and 1. Song, "Image
reconstruction from TE scattering data using equation of strong
permittivity fluctuation" , IEEE Transactions on Antennas and
Propagation, Vol. 48, Issue 6, pp 860-867, June 2000.
18. Harada H., M. Tanaka and T. Tanaka, "Image reconstruction of
a three dimensional object using gradient based optimization",
Microwave and Optical technology letters, Vol. 31, No.6, pp
332-336, December 2001
19. Ranganath, M.V., A. P. Dhavan and N. Mullani, "A Multigrid
expectation maximisation algorithm for positron emission
tomography", IEEE Transactions on Medical Imaging, Vol. 7,
NoA, pp 273-278, Dec 1988.
20. Pan. T. and A. E. Yagle, "Numerical study of multigrid
implementations of some iterative image reconstruction
algorithms", IEEE Transactions on Medical Imaging, Vol. 10,
NoA, pp 572-588, Dec 1991.
21. Milstein, A.B., S. Oh, 1. S. Reynolds, K .1. Webb, C. A.
Bouman and R. P. Millane, "Three dimensional optical
diffusion tomography using experimental data", Optical letters,
Vol. 27, pp 95-97, January 2002.
22. Oh, S., A.B Milstein, R. P. MiIlane, C. A. Bouman and K .1.
Webb, "Source detector calibration in three dimensional
Bayeslan optical diffusion tomography", Journal of Optical
Society America, Vo1.l9, no. 10, pp 1983-1993, October 2002.
23. Sauer, K. and C. A. Bouman, "A local update strategy for
iterative reconstruction from projections", IEEE Transactions
on Signal Processing, Vol. 41, no. 2, pp 534-548, February
1993.
24. Tanaka, M., K. Yano, H. Yoshida and A. Kusunoki, "Multigrid
optimization method applied to electromagnetic inverse
scatteing problem", IEICE Transactions on Electronics, E90-
C(2), pp 320-326, 2007.
25. Caorsi, S., M. Donelli, D. Franceschini and A. Massa, A New
methodology based on an Iterative Multiscaling for Microwave
lmaging, IEEE Trans. Microwave Theory Tech., vol. 51, no. 4,
April 2003.
26. Belkebir, K. and M. Saillard, "Special section: Testing
inversion algorithms against experimental data", Inverse
Problems VoLl7, pp1565-1571, 2001.
27. V. Hamsakutty, A. Lonappan, V. Thomas, G. Bindu, J. Jacob,
J. Yohannan and K. 1. Mathew, Coupling medium for
microwave medical imaging applications, IEE Electronic
Letters, vol. 39, no. 21, Oct. 2003.
28. 1. H. Richmond, Scattering by a dielectric cylinder of Arbitrary
Cross Section Shape, IEEE Trans. Antennas Propagat., vol. AP
-13, pp 334 - 341, May 1965.
29. A. N. Tikhonov and V. Y. Arsenine, Solutions of III Posed
Problems, Washington, DC: Winston, 1977
[
Degree of Symmetry Formulation
5.1 Introduction
Several methods, both stochastic and deterministic have been
developed to solve the inverse scattering problem of electromagnetic
imaging. Typically, all these methods solve a large-scale, non-linear
optimization problem by generating values for every pixel covering
the investigation domain. The electromagnetic inverse scattering
problem is ill posed in the Hadamard sense and is also highly non
linear. Due to its non linearity, there are multiple solutions to the
inverse scattering problem and therefore the solutions are liable to get
trapped in local minima I.
The investigation domain has to be chosen sufficiently large to
include all possible locations of the scatterer whose cross sectional
profile is to be inverted from the measured scattered field data 2 - 4.
When the investigation domain is large, the number of unknowns of
the inverse scattering problem increases and an iterative procedure for
solving the inverse problem may fail to reach a convergent solution
and could get stuck in a local minimum, in addition to increasing the
computation time. The algorithms employed for the electromagnetic
inverse scattering problem generate values also for those regions in the
investigation domain that might not contain any useful information
about the object 4. Therefore if a localization of the object could be
done in the investigation domain nrior to the th ..
available information about the scatterer known VIa the measured
scattered field can now be employed for the optimization of fewer
numbers of cells in the investigation domain, which will help in the
global convergence of the solution
For a 2-0 Dielectric scatterer of circular cross section, with the
complex permittivity distribution symmetric with respect to the centre
of the cross section, a custom defined Degree of Symmetry
formulation is employed here to localize the scatterer in the
investigation domain, prior to the inverse scattering. This a priori
information about the scatterer is valid for structures such as dielectric
posts and pipe lines. The degree of symmetry computed for each
transmitter position is a measure of the symmetry of the scattered field
vector in the measurement domain for each transmitter view. The
degrees of symmetries thus computed for a scatterer together forms a
Degree of Symmetry vector.
It is seen that the Degree of Symmetry vectors exhibit unique
features for the geometric properties of the scatterer. This makes it
possible to localize the object in the investigation domain. This will
reduce the number of unknowns of the inverse scattering problem,
thereby aiding the global convergence of the solution.
The Degree of Symmetry formulation was also applied to the
practical case of limited angle tomography, when scattered field
measurements around the object are not possible, as in the case of
buried objects. The development of imaging techniques for
investigation of physically inaccessible objects is of much importance
m areas such as oil exploration, seismic Imagmg, non destructive
evaluation, buried object detection etc. In most of these cases, reliable
information about the physical properties of the inaccessible target is
desired in addition to detection of the target. All these problems may
be classified as electromagnetic inverse problems when the
characterization of the unknown target is computed by considering
electromagnetic illumination and employing the values of the scattered
electromagnetic field. The ill posedness of the inverse scattering
problem is more severe here, due to the availability of only aspect
limited data. Two dimensional models have been employed to model
electromagnetic scattering from buried dielectric pipelines 4 - 8. When
TM polarization of the EM wave is employed to illuminate the
scatterer, the scattering problem reduces to a scalar one 9. The
reduction in the investigation domain reduces the degrees of freedom
of the inverse scattering problem. This improves the convergence of
the solution and also results in a reduction in computation time of the
inverse scattering methodology employed.
5.2 Formulation of the Problem
The problem is formulated first when the scatterer and the
measurement setup lies in the same background medium. When
measurements of the scattered field maybe done around the object, a
circular geometry is assumed for the measurement setup in the
simulations as shown in figure 5.1.
x
---.
.--.....
e-


mf?a.51lremell'


(iomaill D

"
.\1



~
'"
(;)
..... -
r

@ . ~ .

- ~ , , ~ , ~
- .

..
'-e_
-.- .
... -
im't'\/;glllioll
domaill DObj
Figure 5.1
Measurement setup for electromagnetic inverse scattering experiment
Circular geometry is employed for the setup
A 2-D dielectric scatterer of circular cross section and complex
permittivity distribution symmetric with respect to the centre of the
cross section is located in an investigation domain DOh!' which is
usually a square or a rectangle. The maximum diameter possible for
the unknown circular object to be imaged is also assumed to be known.
This a priori is valid for structures such as dielectric posts and pipe
lines. However, since the location of the object to be imaged is not
known, the investigation region has to be chosen sufficiently large.
The background and the object are non magnetic. To simplify the
implementation, TM polarization of the incident field is considered.
For generating synthetic scattered data for inversion, the incident fields
for the different transmitter view are generated using infinite line
source expressions with N, three dimensional line sources (two
dimensional point sources) equi-spaced on a circle in the measurement
domain. The incident field for a transmitter view v is given by the
electric field expression for a line source,
(5.1 )
where rr = (xp Yr) is the transmitter antenna coordinate
andr = (x,y) 10. A time convention of e
jM
is assumed. At a given time
one of the antennas will be emitting and the others will be receiving.
i.e., the number of receivers per view is N
f
-1. The scattered field is
measured for different views. The measurement domain is outside the
investigation domain. The circular geometry is selected for the
following reasons:
a) Scattered information is collected all around the object for
each incidence 11.
b) It helps in the computation of the degree of symmetry for
the different transmitter positions, as will be shown in the following
discussion.
5.3 Degree of Symmetry Formulation for
localizing the scatterer
Synthetic scattered field for inversion is generated Via the
coupled Matrix equations (3.25) and (3.26). The degree of symmetry
( D.o.S) for a transmitter position v is defined as (assuming even
number of receivers N
f
-1 per view)
[
CN'_1)/2 2]
D.D.S(v) = abs tt 1I('1k)-'1N
1
-k))1I ' v = 1,2 ... N,
(5.2)
and
when there are odd numbers of receivers per view, i.e., N, -1 is odd.
Thus the D.o.S for a transmitter position is a function of the
Euclidean distance between the fIrst half and the spatially reflected
second half of the measured scattered field vector, and the transmitter
v is at the centre of the measurement points employed for computing
the degree of symmetry value. The D.o.S value will be equal to zero
when the two halves are exactly similar and maximum when they are
most dissimilar.
When circular geometry of measurement is employed and if
the dielectric cylinder is circularly symmetric and off centered, the
measured scattered field vector If/"s exhibits symmetry with respect to
its centre, only for the views VI and v
2
that are diametrically opposite,
as shown in figure 5.2.
I lnIDCeMr I
measurement domain
I----+__ imagiJlg domain
---t-- reduced imaging domain
'----- circular scadew-er
(Js
Figure 5.2
VI and v ~ are the maximum symmetric views, while G
I
and a
z
are
the maximum asymmetric views.
The symmetry plots also exhibit two maxima, a
l
and a
2
which
correspond to the transmitter positions where the symmetry of the
measured scattered field vector is least. The plot of D.o.S with
respect to V, exhibits two significant minima, at VI and v
1
' and two
significant maxima at a
l
and G
2
as shown in the example in figure 5.3.
This means that the object centre is located in a line joining the
views VI and V
1
It is clear that G
I
and a
z
will be symmetrically located
with respect to both VI and V
1
When the centre of the circularly
symmetric scatterer is much closer to V
1
than to VI' a
l
and a
2
will be
closer to v
2
than to VI'
(
0.035
(
<;!
0.03
(
DD.S
( il
0.025
r<
(
il
I ~
/l<
;)
0.02
-(
(
0.015
(
0.01 (
0.005
I?J
0
I? "" ~ Q
0 5 10 15 20 25 30 35 40 45 50
Transmitter number
Figure 5.3
The degree of symmetry values are plotted for various transmitter positions.
The minimum values indicate the maximum symmetric positions and the
maximum values indicate the maximum asymmetric positions. In the above
example, there are 49 line sources in D!v/ . The centre of the cylinder lies in
the line connecting transmitter 16 and 40.
This is explained as follows: the scattered field at a
measurement point depends on the distance to the scatterer via the
Greens function, and the field inside the object which is a function of
the object contrast and the incident field. When the transmitter is at
position VI or v
2
' the field inside the object is symmetric about the line
joining VI and v
2
Since the receivers are symmetrically located with
respect to the scatterer, the measured scattered field vector will be
symmetric about its centre. However when the transmitter position is
moved to either side of V
2
or VI' the distances from the adjacent
symmetrically located receivers on either side of the transmitter, to the
circularly symmetric scatterer are different. This difference is more
pronounced when the transmitter position is moved away from v
2
' than
when the transmitter position is moved away from VI
Also the incident field at the position of the scatterer will
exhibit a marked difference in symmetry about the line joining VI
and v
2
' the maximum symmetry direction, when the transmitter
position is moved away from v
2
than VI. This is because the field is
inversely related to the distance. The distance from the transmitter to
the position of the scatterer exhibits a larger variation when the
transmitter is moved away from position v, than when moved away
from position VI along the measurement domain. Therefore the
asymmetric points in the symmetry plots will be closer to V
z
than VI'
when the scatterer is much closer to v
2
than VI. The values of D.o.S at
Q
I
and a
2
will be much larger when the scatterer is farther from the
centre of the investigation domain than when it is closer. Also when
the scatterer is close to the centre of the investigation domain, the
maximum asymmetry positions a
l
and a
2
will be almost equidistant
from the symmetry positions VI and v
z
. However if the object centre
and the centre of the investigation domain coincide, the measured
scattered field vector for all the views will be symmetric with respect
to its centre and hence the degree of symmetry values for all the views
will be very small. Since the maximum diameter possible for the
unknown object is assumed to be known, a reduced image may be
defined at the centre of the investigation domain in this case.
Thus by measunng the maximum value of D.o.s and the
distance between the maxima, and also noting the direction of
symmetry of the scattered field, it is possible to localize the 2-D
circularly symmetric dielectric scatterer in the investigation domain.
Thus a new reduced investigation domain I, ~ 1 may be defined.
This is il1ustrated in the following numerical example. Since a
reduced investigation region is employed, the number of degrees of
freedom is reduced and there is a larger data redundancy from which
better reconstructions may be obtained. The computation time per
iteration for the direct solution of the scattering integral equations
is D(N'), N being the number of cells into which the investigation
domain is reduced. This is now reduced to D(N,') where N, is the
number of pixels in the reduced imaging domain.
5.4 Simulations and Discussions
Some typical symmetry plots are considered in figure 5.4. Here,
49 line sources are equi-spaced in the measurement domain for
generating the synthetic scattered field data using the coupled state and
observation equations. The incident field frequency employed for the
simulations is 3 GHz. The transmitter I is at the zero degree position
with the antennas numbered in the ascending order in the
anticlockwise direction as shown in figure 5.1. For the forward
problem, the investigation domain is discretized at a fine resolution
o f ~ , i.e., each cel1 is a 0.5cmxO.5cm square. This is done because
20
the forward problem needs to be solved accurately 12. The cross
sectional profiles for the dielectric scatterer considered are a
homogenous profile and a profile consisting of two concentric circles,
108
Wliich are usually the practical cases considered for the dielectric
pipeline.
The synthetic scattered field in the measurement domain for
inversion is generated by employing the following parameters for the
object profiles:
The diameter is varied in the range between 6 and 11 sampling
intervals.
The relative permittivity of the scatterer is varied in the range
[1:16].
The offset of the centre of the cylinder cross section from the
centre of D "bi is varied in steps of one sampling interval.
In practice the scattered field data is corrupted by measurement
nOlse. Therefore random white Gaussian noise is added to the
synthetic scattered data to generate the measurement data for inversion.
The Signal to Noise Ratio is defined as
11 11
2
S 1.//
- = IOlog --dB
N 10 II
n
l1
2
(5.4)
The symmetry plots have been generated from nOISY
measurement data with an SNR of 10 dB. The symmetry positions are
19 and 44, which are the diagonal transmitter views, with the
transmitter view I at the 0 degree position. The distances between the
asymmetry positions through 19 and through 44 along the
measurement circle are not as different in figure 5.4 a, as in figure 5.4
b. The peak values at the asymmetry positions are much smaller in
figure 5.4 a than in figure 5.4 b, indicating that the cylindrical scatterer
is very close to the origin in the first example and farther away from
the origin and closer to the transmitter position 44 in the second case .
In the case of figure 5.4 c, the degree of symmetry values for the entire
transmitter positions are very small, indicating that the object centre
and the centre ofthe imaging domain coincide.
One way to choose the reduced imaging domains in the three
cases in figure 5.4 is indicated in figure 5.5. The reduced regions in
figures 5.5 a and 5.5 b are chosen as overlapping by the maximum
radius of the scatterer. This is to avoid the possibility of the scatterer
not being fully located in either of the reduced regions. The
dimensions of the reduced investigation strip are dictated by the
maximum possible diameter of the circular scatterer. It is to be noted
that the reduced regions are always oriented in the VI to V
2
direction. As
an example consider the circular dielectric scatterer of circular cross
section located in the investigation domain as shown in figure 5.6. The
figure 5.7 shows the reconstructed image when the entire imaging
region is employed. The Newton Kantorovich (NK) Procedure is
stopped after the tenth iteration as the relative residual error Err ( If/' ) is
seen to increase afterwards. It is obvious that the NK procedure has
not converged.
D.oS
D.OS
U.IJII
0.035
0.03
0.025 9
0.02
0.015
9
0.01
0.005
00
Transmitter Number
Figure 5.4 a
The centre of the cylinder cross section li c in the line
joining the diagonallransmitters 19 and 44 and is close 10
the centre of the investigation domain.
0.045
0.04
0.035
0.03
0.025
0.02
0.015
0.01
9
0005
00
09 Id 19
5 10 15
'"
25
"
35
'"
45
Transmitter Number
Figure 5.4 b
The cemre of the cyl inder cross section lie in the line
joining the diagonal transmittcrs 19 and 44 and is close
10 the transmitter 44
50
figure S.4 cc
The centre of the cylinder cross section coincides with the
centre of the imaging domain. 3S indicated by the extremely
low D.O.S values
figure S.S a figure 5.5 b fi gure S.S c
3 . reduced imaging region for the synuneny plots in figure 5.4 a
b. reduccd imaging region for the symmetry plots in ligure 5.4 b
c. reduccd imaging region for the symmetry plots in figure 5.4 c
Figure 5.6
The actual profile of a two di mensional
circularly symmetric scaHcrer
5.7
The reconstructed permitti vit y proli lc using the
Newton Kantorovich Procedure . The ":Illirc
investigation domain
\
0.9
0.8
; 0.7
10.8
i
! 05
0.4
0.3
0.20
2 4

8 10 12
iterIIIons
Figure 5.8
The Relative Residual error starts
increasing after the loth iteration
14 18 18
The D.o.s values for the different transmitter positions are plotted in
figure 5.9. The centre of the scatterer cross section lies on the diagonal
of the investigation domain, i.e., the line joining the transmitters 19
and 44, and is closer to the transmitter 19. The variation of the relative
residual error E"(lf/S) with the iterations is shown in figure 5.10. The
iterations are stopped when Err(lf/S) reduces below 2.0 x 10-
2
The
reconstructed image when the reduced imaging region is employed is
shown in figure 5.11. The result with the proposed technique is seen to
be much better, in addition to the reduction in the computational time
per iteration from O(N
3
) toO(N/).
0.'"
0.00'
0.113
0.025
0.02
DD.S
0.015
O.ot
0.005
00 ,
10
15 '"
25 lJ 35 40 45 50


"
i
i ..
i
.,
.2
. ,
'.
Transmitter Number
Figure 5.9
The O.o.S values for the different transmitter positions for the
scatterer in figure 5.6. The centre of the scatterer lies on the
leading diagonal and is closer to the transmitter 19 .
,



" "
..-
Figure S.10
The iterations are stopped when R.R.E reduces below
the selected thresbold, in the J O'h iteration
figure 5.11
The reconstructed image when the reduced
investigation region is employed.
owever the strategy has the disadvantage in that the reduced imaging
Ciomain is no longer a square or rectangle. Ln addition, since the
reduced regions are always oriented in the v. to v
2
direction, the shapes
of the reduced investigation domains will depend on the direction of
symmetry from the D.OS plot. Due to this, the ordering of the cells in
the reduced domain and updating of the contrast vector in all iterations
of tbe inverse scattering algorithm employed becomes more involved.
It is also observed that the D.oS vector is also influenced by the
radius of the cylinder. As an example consider the D.oS plots for the
scatterers shown in figure 5. 12. It is seen that as the radius of the
scatterer increases, the valley between tbe maximum asymmetric
points of the D.oS vector widens. This means tbat a further
localization of the investigation domain can be achieved if the radius
information can be extracted in addition to the offset from the centre
of D <>Ill ' Therefore the problem was recast as a pattern classificat ion
problem, where the classes indicated the reduced investigation
domains in which the scatterer is located.
Figure S.12 a
Perminivity profiles of
three scatterers of
increasing radii located
in the investigation
domain
D.O.s
D.D.S
D.O.S
D.DS
0.ll4
D.03
0.02
O.DS
o.oe
0.01
DD
om
aol5
~
lUllS
0
0.011
o.oe
0.01
D.035
0.03
0.02!i
0.02
0.015
0.01
oms
00
It
I lIf
Ir
Transmitter Number
IrJ 11 In 11
:Ill l5 3D ;I
Transmitter Number
IrJ
!f
Transmitter Number
40 SI
L
Figure S.1l b
The valley between the
maximum asymmetric
points of the
D.O.S plots widens
with increasing radius
For this purpose, a neural network based classifier was employed.
Neural networks otTer an advantage in that they can work as a black
box system capable of solving the physical problem without directly
considering the physical relationship between the input and the output
values of the problem itself 13. Neural Networks have been employed
for the solution of electromagnetic inverse problems 14. 15. However
Neural Networks are used here to localize the scatterer in the
investigation domain as a preprocessing step prior to the inverse
scattering.
The primary requirement of the classifier here is fast training
and easy retrainability. The paradigms such as Back Propagation
Neural Networks are characterized by slow training times. They are
also not easily retrained. i.e., when a new training pattern is to be
incorporated into the training set, the network will have to be retrained
again with the updated set of training data, from the start 16.
The Probabilistic Neural Network (PNN) paradigm was chosen
to classify the training set of DD.S vectors into classes that indicate
the reduced investigation domains in which the cylinder could be
localized. The PNN was chosen because of its rapid training and easy
retraining ability. The PNN provides a general solution for pattern
classification problems by following an approach developed in
statistics, called Bayesian classifiers 17. It also gives an evidence
estimate based on which the classification decision is made. The
network paradigm also uses Parzen Estimators, which were developed
to construct the probability density functions required by Bayes'
theory. The PNN is guaranteed to converge to the Bayesian classifier,
which is the optimal classifier, with enough training data. It is also
ust In t e presence of noise and generalizes well. The architecture
of the PNN is shown in figure 5.13 .
--
Figure 5.13
PNN architcelUre
.......... '
-
-
-
-
--
-
Tbe inputtcst 0.05 vector is classified into onc
ofthc target output classes that indicate tbe
localized invcstigation domain
The network has four layers: an input distribution layer, a
pattern layer, a summation layer and an output decision layer. The
distribution layer has as many elements as there are separable
parameters needed to describe the objects to be classified. Pattern layer
organizes the training set such that each input vector is represented by
an individual processing element. The pattern layer weight matrix is
set to the transpose of the matrix formed from the input training set, in
this case, the training set of D.o.S vectors. The pattern layer transfer
function is the Gaussian Probability function centred at each training
case. When the standard deviation of the Gaussian function is
approaching zero, the PNN approximates a nearest neighbor classifier
and when it approaches infinity, the classifier is limited to functions
that are linearly separable. Some intermediate value of the spread is
usually employed. For the simulations a spread value of 0.1 was
employed.
The training set is minimized by considering only diagonal
displacements along the leading diagonal of the centre of the circular
symmetric scatterer from the centre of the imaging domain. Each
output target class indicates the offset from the centre and the range of
radii the unknown circularly symmetric scatterer belongs to. The
weight matrix of the summation layer is derived from the target classes.
Since the training data consists of scatterers that are displaced
only along the main diagonal, the test degree of symmetry vector is
circularly shifted in the anticlockwise direction by n elements, so that
the two minima of the degree of symmetry are aligned with the views
VI and V
2
of the main diagonal. The PNN classifies the test degree of
symmetry vector into one of the target classes that indicates the offset
and range of radii the circularly symmetric scatterer belongs to. The
reduced investigation domains are indicated by the output classes. The
reduced domains are chosen to be overlapping by the maximum radius
of the scatterer in them so as to avoid the possibility of the scatterer
not being fully located in either of the reduced regions. For the radius
ofthe scatterers chosen as mentioned earlier, the reduced investigation
domains are chosen as squares of dimensions 14 x 14,18 x 18 and
24 x 24 cells containing circular scatterers of outer radii in the range
3 - 5, 6 - 8 and 9 -11 cells respectively.
--"-l AD example is shown in Figure 5.14. The synthetic data is
generated from the profile shown in Figure 5.14 3. The degree of
svrnmetrv for the scatterer is shown in Fie:ure 5. 14 b.
D.o.S
Figure 5. 14 a
The actual profile of the scatterer
employed 10 generate synthetic scattered
data
Transmitter Nwnbcr
Figure 5.14 b
The computed D.O.S vcctor for
the above profile
Since the training data consists of scatterers that are displaced
only along the main diagonal, the test degree of symmetry vector is
circularly shifted in the anticlockwise direction by n elements, so that
the two minima of the degree of symmetry are aligned with the views
VI and VI of the main diagonal. The rotated D.OS is shown in figure
5.14 c.
DD.S
Transminer Nwnber
Figure 5.14 (!
ThcD.o.S vector is circularly shifted
so that it is oriented in the direction of
the main diagonal.
The reconstructed image in the lOth iteration of the Newton
Kantorovich procedure, considering the entire investigation domain, is
shown in figure 5.14 d. The relative residual error Err(I//S) varies with
the iterations as shown in figure 5.14 e.
Figul'"e 5.1 4 d
The reconstructed image considering the entire
investigation domain after the 10
th
iteration of the
NK procedure
"

0.7

i 0.'
i 0.5

! 0.'

0.3
0>
0.'
00 ,

"
. .-.
figul'"e 5. 14 c
Err(lj/S) is found to increase after the 10
111
iteration
Therefore the NK procedure is stopped after the 10
th
iteration.
"
..
The PNN classifier classified the circularly shifted D.O.S vectol'" in
figure 5. 14 c into a reduced tar-get domain of 18 x 18 cells where the
scatterer IS located. The final reconstructed image is rotated in the
clockwise direction about the centre of the invest Igation domain by
360' . n Th . . d b k' h 1 k .
"-'-':-C-'-. e reconst ructed Image IS rotate ac III t e c oc wise
N,
direction by when the reduced investigation domain is considered is
shown in the figure 5. 14f. The NK procedure converges in the loth
iteration indicated by the plot of 1'r(I/I ' ) vs. iterations as shown in
figure 5.14 g
Figure 5.14 f
. TI1C reeonstruclcd ilTl.1ge cons idering a reduced
invesligation domain of 18 x 18 cell s
0.9
0.8
0.7
106
I
..
0.5
It:
i 0.4
It:
0.3
0.2
0.1
0
0
\
\
\
\
2 4 6 8 10
~
Figure 5.14 g
The NK procedure is stopped after the tenth
iteration as the Err( vr ) decreases below the
specified threshold of2.0 x 10-6
12 14
The D.O.s formulation for circular scatterers was successfully
applied to the case when only aspect limited data is available as in the
case of inverse profiling of buried dielectric cylinders. The ill
posedness of the inverse scattering problem is more severe here, due to
the availability of only aspect limited data 5,6.
Many algorithms have been designed to tackle the
electromagnetic inverse scattering problem of imaging buried
scatterers. For the detection of buried targets and land mines, a simple
model has been proposed by using the Ground penetrating radar (GPR)
technique, without accounting for the air-earth interface 18.19. To build
up more accurate models, a half space problem must be considered to
represent the air-earth interface. Several methods that include the
modified gradient approach 20, 21, Born iteration and constrained
optimization 5, Distorted Born Iterative method 6 and high-order
extended Born approximations 7 have been investigated for the buried
object problem.
The cross sectional profiles considered for the cylinder are a
homogenous profile and a cross sectional profile of two concentric
circles, which are the practical cases for a buried pipeline. The
investigation domain IS illuminated with TM polarized
electromagnetic radiation so that the scattering problem reduces to a
scalar one. Scattered data can be measured only in a limited angle, as
the imaging domain is inaccessible. The following situation is
investigated: the pipeline is buried in wet soil in which case a half
space problem is considered to account for the air-soil interface. A
localization of the dielectric pipeline in the investigation domain is
done prior to the cross sectional profiling, which will reduce the
number of unknowns. For this, the custom defmed scalar
D.o.S parameter is computed for every transmitter view. Thus for a
particular object configuration, the D.o.S vector, which comprises of
the scalar values computed for each transmitter view maybe formed.
The PNN is trained with the degree of symmetry vectors computed for
a training set of object configurations, which classifies the degree of
symmetry vector of the unknown scatterer presented to it into one of
the classes that indicate the localized region in the investigation
domain, prior to the inverse scattering.
5.5 Formulation of the buried pipeline scattering
problem
The investigation domain D
ob
} is chosen to be a square or rectangle
that possibly contains the buried dielectric pipeline. This should be
large enough to include all possible locations of the buried pipeline.
The axis of the pipeline is oriented in the z direction so that the cross
section lies in the x - y plane. Both media and the scatterer are
assumed linear, isotropic, non magnetic and penetrable.
When the transmitters and receivers are mounted at the same
height, the Greens functions gm"ml and gmlml maybe simplified as 6
and
(5.6)
The transmitters are ideal time harmonic electric current line sources,
set parallel to the z-axis (TM polarized), and mounted at a height Yo
from the origin in the measurement domain D
M
The incident field is
given by
(5.7)
where,
+4l
(
') - f d'k (k.) i
k
.. " (2h-y-y') jk,(x-x')
g, r,r - xgo < e e
(5.8)
and
(5.9)
r
T
are the receiver and transmitter coordinates respectively. The y-
axis passes through the central transmitter. The receivers are also ideal
and the number of receivers is one more than that of the transmitters,
as there will be a receiver antenna on either side of a transmitter
antenna. The measurement setup IS assumed movable in the
x direction. The scattering problem IS scalar, and only the
z component of the scattered field is needed.
The geometry of the problem is as shown in figure 5.15, where
the homogenous half spaces, air and wet soil, are separated by a planar
interface.
. : ,- ................. .. ---_.
'l777777?777J?)JJ7/7777?7?7T/7777777777777777
_ ...
1
D ..
10,01
Figure 5.15
The measurement setup for imaging the buried
pipeline Transmitters and receivers are ideal tine
sources mounted at the same height above tbe
air-soil interface
5.6 Degree of Symmetry Formulation for
localizing the buried pipeline

)
The degree of symmetry (D.O.s l . for the transmitter view v
is defined as (assuming even number of receivers N ~ per view)
[
N " ')
D.o.S(vl=abs b ivt;(kl-II': (N.+I-kll .v= I.2.3 ...... .N.-1
(5.10 1
In equation 5.10, the degree of symmetry for any gIven
transmitter view v is computed with all the measured scattered values
employed for the calculation, with the central transmitter numbered
as N
r
Thus the D.o.S for a transmitter position is a function of the
2
Euclidean distance between the flfst half and the spatially reflected
second half of the measured scattered field vector .The D.o.S value
will be equal to zero when the two halves are exactly similar and
maximum when they are most dissimilar. The two halves of the exact
measured scattered field vector are identical only for case of the
central view v = N
r
and when the circularly symmetric pipeline cross
2
section lies symmetric with respect to the measurement array with its
centre on the y-axis as in figure 5.15. The measurement setup is
moved in the x direction until the D.o.S for the central transmitter is
minimum, which is when the centre of the pipeline cross section
coincides with the y -axis. For this purpose a modified degree of
symmetry is computed for each transmitter position as follows:
N
for v =1:_' ,
2
1//,'1- = 11< (r),r = 1,2, ... v
1//I'R =1//:' (r), r = 2v,2v-l. 2v - 2, .... v+ 1
N
for v>_r ,
2
'1/"L =11< (r),r = 2v - N
r
+ l, ... v
(5.11)
Thus the lengths of the ftrst half vector 'I/\'L and the spatially
reflected second half vector 'I/.R are equal for a given transmitter view
v. In (5.10), the degree of symmetry for any given transmitter view v is
computed with all the measured scattered values employed for the
calculation, whereas in (5.11), the transmitter v is always at the centre
of the measurement points employed for computing the modified
degree of symmetry. If the buried 2-D cylindrical pipeline is located
under a transmitter offset from the central transmitter as shown in
figure 5.16 a, the D.o.S
mcxI
value will be equal to zero for that view as
shown in figure 5. 16 b. The measurement setup is now moved in the
x direction such that the minimum value of D.o.S
mod
now coincides
with the central transmitter N
r
, which is when the centre of the
2
pipeline cross section coincides with the y -axis as in figure 5.15. The
new D.o.S
mcxI
vector is plotted in figure 5.16 c where the minimum
value is for the central transmitter. The scalar D.o.s values are then
computed for each transmitter position, which together constitutes the
D.o.S vector. In this case, D.o.S vector will be symmetric with
respect to its centre, where it has its minimum value.
Figure 5. 16 a
A 2-D dielectric pipeline with circular cross section is located wit h its
di splaced to the right of the centre of the investigation doma in
.. ..
,.
1



u
transmittcr
Figure 5.16 b
"
The modi tied degree of symmetry is ploued for the pipdine in tigure 2. a.
For the simulation J I transmitters have been employed. The pl ot shows that the
pipeline centre is located under the transmitter 20. The measurement setup is to be
moved by 4 transmitter spacing: to the right so that the minimum uf o.o.S"hl is now
aligned with the central transmitter 16.
,
"
,
..
'.

iu!

transminer
Figure 5. 16 c
The modified degree of symmetry D.O.Smod is ploued from the computed
measured scattered field for the new position of tbe measurement array.
The centre of the pipeline is now located under the central transmitter, as
indicated by the minimum value of D.O.SrncxI coinciding wit h the central
transmincr 16.
However when the transmitter view is changed to either side of
the central view, the scanered field vector Ill: wi ll nol be symmetric
with respect to its centre as the receiver array is not symmetric with
respect to the transmitter location. Therefore the D.o.s will inc_Tease
either side of the central view. There wi ll be two views symmetrically
located OD either side of the central view where '1/: is mosl
asymmetric and hence the D.o.S is maximum.
The measured scattered field vector is dependent on the
distance of the receivers from the scatterer via the Greens
function G .... ",l and the total field inside the object fII v which is a
function of the object contrast c. the incident field fII!fJC . It is seen that
the geometril: properties namel y the inner and outer radii and the depth
at which the scalterer is located in D"/'I and the dielectric contrast of
the scatterer influence the values and features of the D.o.S vector. For
example, consider the pipeline located at the top of the invest igation
domain, as shown in figure 5.17 a. The IJ.o.s for this configuration is
ploued in figure 5.17 b. The same pipeline is shown buried at a greater
depth in figure 5.17 c and the corresponding D.o.s is shown plotted in
figure 5.17 d. In figure 5. 17 e, the pipeline li es buried at the bottom of
o and its D.o.S is plotted in figure 5.17 f. it is observed from figure
5. I 7 that the D.o.s values show a faster increase from the central
minimum value to the maximum values on either side, when the object
is located at a shallower depth in The features of the D.o.5 vector
maybe employed 10 loca lize the dielectric pipeline in D"hi as shown
earlier. for which the PNN classi fier is employed.
Figure ;)
The pipel ine is al the top of the
invest igat ion dOlTLl in
o.o.s
t ~ '0'
..

"
..
~
..
"
..,

1. .1
'.
,
..
"
~

transmitter
Figure 5.17 b
The DD.S ploHed for the scaltcrcr in fi AutC 5. 17 a
Figure 5.17 c
Tbe same pipel ine is located at a
b'Tcatcr dcpth in thl.: investi gat ion
domain

3S .10 '
,
"
D.D.S
,
...
(ransmittcr
figure 5. 17 d
The D.O.S plotted for the scauercr in figure
5. 17 c. The maxima of the DD.S arc spaced
fanher apart than in figure 5. 17 b.
Figure 5.17 c
The pipeline is locat ed at the bottom uf
the im'csligalion domain
D.o.S

..

.,
III
,rl
'.

" "
M
transmitter
Figure 5. 17 r
The D.o.s ploued for the scatterer in figure 5.17 e. The
max.i ma of the D.OS arc spaced fanber apart than in
figure 5.17 d
5.7 Numerical Simulations
The upper region considered here is free space (e .. = 1,0'" = 0)
and the lower region is wet soil ( & ~ ;4,0', =O.005s l m) . The
observation equation (3) and state equation (4) are employed for
generating synt hetic scattered field data ID the measurement
domain D....." for various known circular object profiles buried in wet
soil. For the simulation of the experiment, 31 transmitters have been
employed in the measurement domain. For each transmitter view. 32
measurements are made at the receiving antenna positions as
illustrated in figure 5. 15. The investigation domain D obj is chosen to be
a square of 2 meters side. The top of the investigation domain is
located at a depth of 2.25 meters from the air-soil interface. The
measurement domain is located at a heigh:t of O. 7S meters from the air-
soil interface. The investigation domain is discretized into
64x64 = 4096 pixels. The working frequency employed is 100 MHz.
Though a higher frequency would have improved the imaging
resolution, the depth of penetration of the EM wave will reduce with
increasing frequency. The measurement domain is 4 meters long with
the transmitters equally spaced on the measurement line and receivers
similarly arranged in between the transmitter positions as in figure
5.15. The measurement setup is assumed movable in the x direction.
The cross sectional profiles for the dielectric scatterer considered are a
homogenous profile and a profile consisting of two concentric circles,
which are usually the practical cases considered for the dielectric
pipeline. The measurement setup is moved in the x direction until
the DD.S value for the central transmitter is minimum, which is when
the centre of the pipeline cross section lies on the y -axis. The exact
scattered fields in the measurement domain are generated with the
following parameters of the cylindrical dielectric scatterer:
For the homogenous profile, the diameter is varied in the range
15 centimeters to 45 centimeters. For the concentric circle
cross section, the diameter of the outer cylinder is varied in the
range 30 to sampling intervals to 45 sampling intervals, and the
inner diameter between 5 sampling intervals and 10 sampling
intervals.
Two ranges of the relative permittivity of the dielectric pipeline
are considered. A loss less scatterer is first considered having
relative permittivity in the range 2 to 10. A lossy scatterer is
also considered with the relative permittivity varied in the
range 5 to 10 and conductivity in the range 0.001 to 0.010 S/m.
The relative permittivity is varied in steps of 1 and the
conductivity is varied in steps of 0.002 S/m.
The depth at which the cylinder is located in DOhj is varied in
steps of2 sampling intervals.
The Distorted Born Iterative method (DBIM) is employed for the
inverse scattering. The DD.S vectors for the various scatterer
configurations are generated from the exact scattered data as per (5.10).
A probabilistic neural network is trained with the training set of
DD.S vectors. The training is virtually instantaneous and involves
nothing more than reading in the training vectors and storing them as
rows of the pattern layer matrix.
As an example consider the cylindrical dielectric scatterer
(e = 10,0" =O.Ols / m) of diameter 15 centimeters located at a depth of
0.5 meters from the top of the investigation domain, and the centre of
the cross section offset to the right from the centre of the imaging
domain by 38 centimeters as shown in figure 5.18 a. For the measured
scattered field, an SNR of 30 dB is assumed. The DD.S
mod
values
computed from the measured scattered field are shown ploUed in
figure 5.18 b. The plot indicates that the pipeline is located under the
transmitter 19. The measurement setup is moved by three transmitters
spacing to the right so that the pipeline now lies under the central
transmitter 16.The DD.S , computed from the measured scattered field
is plotted in figure5.18 c. The calculated DD.S vector is presented to
the PNN, which classifies it into one of the target classes that indicate
the location of the centre of the scatterer and its radius. The
reconstructed permittivity and conductivity images, after the 20
th
DBIM iteration considering the entire imaging domain are shown in
figure 5.19.
Figure 5.18 a
The actual cross sectional profile of the cylindrical dielectric pipeline
(s =10,0" =O.Ols/m)
lii? fl
1.6
0.8
1.2
x10"
1.4
0 2
04
D.Q.S
mod
1 ~ ?l1
transmitter
Figure 5. 18 b
The D.O.S
moo
vector for the scatterer in figure 5.a is plotted above. It indicates that
the measurement domain is to be moved to the right in the x -direction by three
transmitter spacing so that the DD.S minimum will be aligned with the central
transmitter 16.
141
,
7

D.o.S
5

J
,
00
t.oT
5 10 15
'" " "
transmitter
Figure 5.18 c
The D.O.s ve<:tor for the pipeline after the measurement anay has
been shifted by three units to the right. This is now presented as the
lest vector to the PNN for classification.
When the D.O.S vector is employed for the localization of the
cylindrical scatterer, the PNN classifier classifies the scatterer as
belonging to a localized region of 20 x 20 square centimeters. The
reconstructed profiles after the 8
th
DBlM iteration are shown in figure
5.20. It is to be noted that the images obtained with the reduced
investigation domain are always centered about the y -axis, as the
centre of the pipeline is now aligned with the central transmitter. The
reconstructed images are found to be much better when the localized
imaging domain is employed than when the entire imaging domain is
employed, in addition to reducing the computation time.
Figure 5.19 a
The reconstructed permini vity considering the entire investigation domain
Figure 5.19 b
The reconstructed conductivity considering the entire investigation domain
Figure 5.20 a
The reconstructed permittivity wit h a reduced investigation domai n
The coordinate system has been shi fted so that the centre of the pipeline cross section lies on
tbe yaxis
Figure 5.20 b
The reconstructed condw.:ti vity with a reduced investi gation domain
The coordinate systcm has been shifted so that the centre of the pipeline cross section
lies on the yaxis
References
I. Qing, A. and C. K. Lee, "Microwave lmaging of Parallel
Perfectly Conducting Cylinders using Real Coded Genetic
Algorithm coupled with Newton Kantarovich Method",
Progress in Electromagnetic Research, PIER 28, pp 275-294,
2000.
2. Rekanos, I. T., M S. Efraimidou and T. D. Tsiboukis,
"Microwave Imaging: Inversion of Scattered Near Field
Measurements", IEEE Transactions on Magnetics, Vol. 37,
No.5, pp 3294 - 3297, September 2001.
3. Guillermin, R., P. Lasaygues,1. P. Sessarego and A Wirgin,
"Imaging an object buried in the sediment bottom of a deep
sea by linearized inversion of synthetic and experimental
scattered acoustic wave fields", Inverse Problems, 16, 1777-
1797,2000.
4. Baussard. A, E. L. Miller, D. Lesselier, "Adaptive
Multiscale Reconstruction of Buried Objects", Inverse
Problems, 20, S l-S 15, 2004.
5. Chaturvedi. P. and R. G. Plumb, "Electromagnetic lmaging
of Underground Targets using Constrained Optimization",
IEEE Transactions on Geoscience and Remote Sensing, Vol.
33, No. 3, pp 551-561, May 1995.
6. Cui, T. J., W. C. Chew, A. A. Aydiner and S. Chen, "Inverse
Scattering of Two- Dimensional Dielectric objects buried in
Lossy Earth using the Distorted Born Iterative Method",
IEEE Transactions on Geoscience and Remote Sensing, Vol.
39, No:2, pp 339-346, February 2001
7. Cui, T. J., Y. Qin, G. L. Wang and W. C. Chew, " Low
Frequency detection of two-dimensional buried objects using
high-order extended Born approximation", Inverse Problems,
20, S41-S62, 2004.
8. Etlis, G.A and I. C. Peden, "An Analysis Technique for
Buried Inhomogeneous Dielectric Objects in the presence of
an Air-Earth Interface", IEEE Transactions on Geoscience
and Remote Sensing, Vol. 33, No. 3, pp 535-540, May 1995.
9. Chew W.c., Waves andfields in inhomogeneous media, New
York, IEEE Press, 1995.
10. Slaney M, Azimi M, Kak AC, Larsen L E, "Microwave
imaging with first order diffraction tomography", Medical
applications of microwave imaging, pp 184-211, IEEE Press
New York, 1986)
11. Franchois A and C. Pichot, "Microwave lmaging-complex
permittivity reconstruction with a Levenberg-Marquardt
Method", IEEE Transactions on Antennas and Propagation,
Vol. 45, no.2, pp 203-215, February 1997.
12. Tijhuis, A G., K. Belkebir, A. C. S. Litman and B. P. de
Hon, " Theoretical and Computational Aspects of 2-D
Inverse Profiling", IEEE Transactions on Geoscience and
Remote Sensing, Vol. 39, No. 6, pp 1316-1329,June 2001.
13. Wasserman, P. D., Advanced Methods in Neural Computing,
Van Nostrand Reinhold, New York, 1993.
14. Caorsi, S. and P. Gamba, Electromagnetic detection of
dielectric cylinders by a neural network approach, IEEE
Trans. Geosci. Remote sensing, vol. 37, pp 820-827, March
1999.
15. Bermani, E., S. Caorsi and M. Rafetto, A microwave object
recognition approach based on neural networks, IEEE
instrumentation and measurement technology conference
proceedings, pp 1582-1585, Venice, May 1999.
16. Bose, N. K. and P Liang, Neural Network Fundamentals with
Graphs, Algorithms and Applications, Tata McGraw-Hill,
New Delhi, 1996.
17. Specht, D. F., Probabilistic Neural Networks for
Classification, Mapping or Associative Memory, Proceedings
of the IEEE International Conference on Neural Networks,
vol.1, pp. 525 - 532, IEEE Press, New York, 1988.
18. Witten, A. 1., J.E. Molyneux, and lE. Nyquist, Ground
Penetrating Radar Tomography: Algorithm and case studies,
[EEE Trans. Geosci. Remote sensing, vol. 32, pp 461-467,
1994.
19. Deming R. and A.l Devaney, Diffraction tomography for
multi-monostatic ground penetrating radar imaging, lnv.
Problems, vol. 13, pp 29-45, 1997.
20. Souriau, L., B. Duchene, D. Lesselier, and R. Kleinman,
Modified gradient approach to inverse scattering of binary
objects in stratified media, lnv. Problems, vol. 12, pp 463-
481, 1996.
21. Lambert, M., D. Lesselier and B. 1. Kooij, The retrieval of a
buried cylindrical obstacle by a constrained modified
gradient method in the H-polarization case and for
Maxwellian materials, Jnv. Problems, vol. 14, pp 1265-1283,
1998.
CHAPTER 6
Experimental setup for monochromatic
inverse scattering
6.1 Introduction
Experimental verification of the algorithms for the inverse
scattering problem of microwave imaging can be found in literature.
Karg described an experimental setup that employs multiple frequency
illumination for microwave imaging ] . Larsen and Jacobi reported
microwave scattering parameter imaging results on an isolated canine
kidney 2. Chang et al designed a tomographic microwave imaging
system for nondestructive evaluation and object recognition of civil
structures and materials 3. Franchois et al described a 2.45 GHz planar
microwave camera for qualitative imaging 4. Semenov et al developed
experimental prototypes for two dimensional and three dimensional
microwave imaging for biological purposes 5, 6, Belkebir et al
described an experimental setup for the two dimensional microwave
imaging problem of infinite cylinders illuminated by TM polarized
incident field 7. A 434 MHz circular complex permittivity scanner was
employed for biomedical imaging experiments by Franchois and
Tijhuis lI, Meaney et al developed a laboratory-scale microwave
imaging system for non-invasive temperature monitoring 9. A
microwave imaging system for experimental investigation of tumor
detection in multilayer breast phantoms was proposed by Li et al 10.
Tomohisa et al proposed a reflection type system with a two way
antenna for microwave imaging 11. The experimental data generated at
the Ipswich test site ofthe Rome Laboratory, Hanscom Air Force Base
have been used by some authors to verify the accuracy of the
algorithms for the inverse scattering problem of microwave imaging 12
15
6.2 Mechanical specifications of the apparatus for
the inverse scattering experiment
An apparatus for monochromatic inverse scattering of two
dimensional scatterers illuminated by TM polarized microwaves for
experimental verification of the Degree of Symmetry (DD.S)
formulation is described here. The apparatus is as shown in figure 6.1.
transmitting
antenna
helical
gears
cylindrical
scatterer
receiving
antenna
stepper
motor!!:

Figure 6.1
Schematic of the
measurement apparatus
The body of the apparatus was built almost exclusively of
polymethyl methacrylate (PMMA), which is the synthetic polymer of
methyl methacrylate 16. It is highly transparent and goes under the
trade names Plexiglass, Perspex, Acrylex, Lucite etc. Perspex is a
nearly lossless material 17, 18 The only metallic parts of the
measurement setup were the bipolar stepper motors employed for
rotating the transmitter and receiver antennas. Thus the scattering from
the measurement setup was minimized. A circular geometry was
employed for the apparatus since scattered information was collected
all around the object for each incidence 19, The transmitter and receiver
antennas were rotated by means of 10 kg cm bipolar stepper motors
operated in full step. The torque was improved by reducing the
rotation speed by means of helical gears, which had a gear ration of
1 :4. Helical gears were preferred to spur gears because of their ability
to take more torque than spur gears and to operate smoothly 20. The
gears were fabricated on Perspex disks.
Dipole antennas were employed as transmitting and receiving
antennas. The transmitting and receiving antennas were moved in a
circular measurement domain. The radius of the transmitting and
receiving measurement domains were 90 and 80 centimeters
respectively. The transmitting and receiving antennas were stepped
every 9 degrees. For every transmitter position the receiving antennas
stepped through 360 degrees. The investigation domain was chosen as
a square centrally located within the measurement domain. The plots
of the reflection parameter SII for the transmitting and receiving
antennas with frequency were as shown in the figures 6.2 and 6.3, The
frequency employed for imaging was 1.8 GHz, where the antennas
have the maximum radiation.
o . - - - - - - - - - - - - - - - - - ~ - - - - - - - - - - - - - - - - - - ~ - - - - - - - - - - - - ~ - - - -
-10
en -15
'0
......
~ -20
-25
-30
-35
1 1.5 2
2.5
frequency in GHz
Figure 6.2
Plot of SI) with frequency for the
transmitting antenna
III
'tI

/
;,'
;'
-\ /
\ (
\ /
I
\ !
-10
.5 -15
\ I
....
....
III
\ I
\ I
\ l
\ I
IJ
-20
-25

1 1.5 2 2.5
frequency in GHz
Figure 6.3
Plot of SL L with frequency for the
receiving antenna
The incident field was mode led using the TM polarized infinite line
source expression of equation S.I.The data acquisition was done
employing the Agilent 8714 ET network analyzer, interfaced to a
workstation via HPIB interface. The controlling of the stepper motor
movements and making of the measurements were done by using an
efficient program written in HP Basic. The measurements were made
in the anechoic chamber facility at the Department of Electronics,
CUSAT, so that spurious signals ansmg due to scattering from
structures other than the object under investigation could be
minimized. The measurement setup was as in figure 6.4

..
Dimensions in cm
I.. 100 ~ I .
pl)"olo"OOO
Wed2e
Aluminium
360
Figure 6.4
1= I = ~
360
The measurement setup with the measurement apparatus in the
ane<:hoic chamber The measurements are made using the
Agilen! 8114 ET network analyzer.
The figure 6.5 shows a comparison between the computed scattered
field and the measured scattered field in the measurement domain. The
computation assumes a hollow dielectric cylinder of circular cross
section with diameter 2 inches and relative permittivity &, =3, centered
in the investigation domain. whereas for the experim!:nt, a hollow PVC
( which is nearly lossless and has a permittivity of roughly 3) cylinder
of 2 inches diameter was placed at the centre of the investigation
domain. The antenna gam was measured by comparing the on-axis
measured incident field to the normalized incident field computed
using the infinite line source expression.
--
--
Figure 6.5 a
Rea.l computed scancrcd field
Figure 6.5 b
Imaginary computed scattered field
OM,
o

--
Figure 6.5 c
Real measured scattered field

--
Figure 6.5 d
Imaginary measured scattered field

6.3 Results and Disc ... ssions
The applicability of the Degree of Symmetry fonnulations for the
practical case of Limited Angle Tomography was verified
experimentally. Different die lectric scatterer configurations of circular
cross section were employed for the experimental verification of the
D.o.s formulations. Synthet:ic scattered data were generated with the
scatterer configurations, employing the equations 3.25 and 3.26, for
generating the Degree of S y.nmetry plots, which were then used to
train a Probabilistic Neural Network to localize the scatterer. The
sampling rate was 0.05,1. which at the selected frequency of 1.8 GHz,
had a value of 0.84 centimet:ers. The investigation domain was chosen
as a square of side 125 pixe Is, i.e., 105 centimeters. Equations 5.11
and 5.12 were employed for calculating the DD.S and Modified
Degree of Symmetry (D.O,S""od) Vectors respectively. For employing
the DD.S formulations, the transmitter and receiver antennas were
rotated so that the minimuI'I"I value of DD.S
mod
coincided with the
central transmitter, as illustrat:ed in figure 6.6.
~ . . . - - . - t - - ~ .. ..... -
a n . ~ d .. IDoUn
ILcJ.-------t--- cinolal daeleclric

[0.0)
sratt.J .... r
T,. . Figure 6.6 . .. .
he scatterer cross sectIOn centre. centre of the mvestlgatton domam
and the central transmitter and receiver are radially aligned.
Thus the scatterer cross section centre was now radially offset from the
central transmitter. The Probabilistic Neural Network was trained with
the D.o.S vectors computed from the synthetic scattered field. For
generating the synthetic scattered field vectors, the scatterer radii were
varied from 5 sampling intervals to 10 sampling intervals. The relative
permittivity of the cylinder was varied in the range 1 to 10. The offset
of the centre of the scatterer from the centre of the imaging domain
was varied in steps of 1 sampling interval. The pattern layer weight
matrix was derived from the D.o.s vectors, with a radial basis
function centered at each training case. The weight matrix of the
summation layer was derived from the target class. The reduced
investigation domains were chosen as squares of dimension 32 x 32
pixels. The reduced domains were chosen to be overlapping by the
maximum radius of the scatterer in them so that the possibility of the
scatterer not being fully located in either of the reduced regions was
avoided.
Hollow poly vinyl chloride pipes were chosen as the scatterers
to be imaged. The experiments were repeated for the stronger
scattering case of the poly vinyl chloride pipes filled with water. The
incident field at the measurement domain was measured by performing
the experiment without the scatterers in the investigation domain. The
following results illustrate the advantages offered by the
DD.S formulations in reconstructing the scatterer cross section. Here a
PVC pipe of 2 inch diameter was the object under test. 19 transmitter
antenna positions were employed for the reconstructions, with the
receiving antenna stepping through 19 measurement points for each
transmitter position, as in figure 6.6. The figures 6.7 to 6.11 show the
reconstructions obtained by employing the DBIM procedure on the
entire investigation domain.
Figure 6.7
First iteration oftbe DBIM Procedure on
the entire domain
Figure 6.8
IO'b iteration of the DBIM procedure on
the entire domain
"'igure 6.11
RccOIl$Iructcd image aficr 90 DBJ:vt iterat ions on the entire domain
Relative
residual
crror
..
~
u
0.'
u
0.'
0.'
itcrations
Figure 6.12
The relati ve residual error IS sccn to increasc alicr the
90'" iteration.
The DBIM iterations were sought to be tenninated when the relative
residual error Err(I//") fell below 0.02. However Err(I//") was seen to
increase after the 90
th
iteration as shown in figure 6.12. Hence the
iterations were terminated after the 90
th
DBIM step even though the
termination criterion was not met. The reconstruction shows that the
relative permittivity was poorly reconstructed, especially the value (of
air) inside the hollow cylinder, as illustrated in figure 6.11.
The Modified Degree of Symmetry D.o.S
mod
vector for the
scatterer were computed and are shown plotted in the figure 6.13.
D.O.S_
, ~
l
,-,
, ~ .
1
.:.
!
I
?
I
I I
l
I
I
,
I
I
'f )
I i :
I
I
I
I
I
i
2 4 6 8 10 12 14 18 18 20
Transmitter number
Figure 6.13
D.O.S
mod
Vector indicates that the measurement setup is to be
moved to the left horizontally with respect to the dielectric
pipe so that the scatterer, centre of the imagiog domain and
the central transmitter are radially aligned.
The O.o.Smod vector shows that the measurement setup is to be moved
horizontally to the left with respect to the imaging domain, so that the
dielectric pipeline and the centre of the imaging domain are radially
aligned to the central transmitter. In the experiment, the scatterer was
moved to the right so that it was aligned with the central transmitter.
The o.o.Svector was computed with the new position of the dielectric
scatterer and is shown plotted in figure 6. 14.
. 2 ~ '
"
,
D.D.S
i
o

o

o
. ,
00
r,
,

"
,
..
,
..
,

'"
Transmitter number
Figure 6.14
O.D.s Vector plotted for the new position of the dielectric pipe.
The Probabilistic neural network which was trained with the
simulatedV.D.S vectors classified the computed D.o.s vector into a
reduced imaging region of 32 x 32 pixels. The figures 6. 15 to 6.19
show the results obtained by employing the reduced imaging domain.
Figure 6.15
Reconstructed image after the ftrst
DBIM iteration for tbe reduced
imal!intl domain
Figure 6.16
Reconstructed image after tbe IOIll DBIM
iteration for the reduced imaging domain
Figure 6. 17
R t t d "ft"" th" 2() ccons rue C linage ..... , .. DB(M
iteration for the reduced imaging domain
Figure 6.18
Reconstructed im..ge after the 25
1h
DBIM iteration for
the reduced imaging domain
Figure 6.19
Final recorr;tructed image after the 32f'1d DBIM iteration for the
reduced imaging domain. The relative permittivity (that of air) inside
the dielectric cylinder has been reconstructed well .

"
Relative
,
residual
\
error

..
\
\
..
\
\
"
"
"
,
.
~ .
-,
'.
.. .. ~

iterations
Figure 6.20
The relative residual error versus DBlM
iterations
Figure 6.20 shows the variation of the relative residual error with the
DBIM iterations. The termination criterion was met at the end of the
32
nd
iteration, where the DBIM procedure was stopped. The final
reconstructed image is shown in the figure 6.19. The dielectric pipe
was better localized and the permittivity values within the pipe
reconstructed better, when compared to the result obtained with the
entire domain in figure 6.11, thus illustrating the advantages of the
D.o.s formulations. The DBIM procedure with the reduced domain
converged much faster, when compared to the entire domain iterations.
The experiment was also repeated for the strong scattering case
of the dielectric cylinder filled with water. The final real and
imaginary parts obtained with the DBIM procedure on the entire
domain figures 6.21 a and 6.21 b respectively.
Figure 6.21 a
Reconstructed real part of the scatterer
permiuivity cross section al the 51<1
DBLM iteration.
Figure 6.21 b
Reconstructed imaginary part of the
scatterer permittivity cross section at the
51 SI DBlM iteration
The DBIM iterations are stopped after the 51
st
iterations as the relative
residual error was found to increase as shown in figure 6.22. Evidently
the solution has been trapped in a local minimum.
Relative
resi dual
error
" " r - - ~ - - - - - - - - - - - - - - - '
..

'.
"
iterations
Figure 6.22
Relative residual error is found to
Increase after the 5, 01 iteration. Tbe
DBlM procedure is terminated after the
51 "' iteration
The reconstructed images obtained with the reduced imaging domain
employing the D.o.S formulations are shown in figure 6.23.
.. .




"


. "
1" ; ; ; ; :1
","



,. : ..


no' . 1 ' I ,
"...... ..'
... .... ....
.L __
a-,
..
"
. " ".' '1."

..


'0

"
.


..

r,


,
..

..



Figure 6.23a



..
" ..

. , ..
..


" ..

The reconstructed real part of the
profile of the scatterer cross section after the 50'
iteration

'.
"
,." ... ,", .'-'
, "'0_'
, ., .. ;"" ...
I., .,' ..
.. .... ... ,
"'.
H _

.. ' -
... ,
"
.' .
... " . '. A,
......... ....
...... ', ....... ",
. . .. ... .
1-.,- . . -.',,. -,.
"'-- ... ' . . .. .. ..
,
t '-'- .-- '. . ."
'. ". '" .
'" '.' .. .... " ...
-,-." ,., .... ..
- , .... ' w
.. ..
",.. . . , _., .... .. r
,.... ." _ L _
. ,.H., .. .. .... ,
, .. u . ,. _. , . ........
... '. .. .,," .. .. .. , .
. '-' ...... . .. ...... .
.-.. ... .-.... .. -... '
...... _. u,. ": ,
.

..

Figure 6.13b
;,
The reconstructed imaginary part of the
profile of the scatterer cross section after the 50'
iteration
J'\
0.'
\
0.7
'.
Relative
0.6
residual \
,
error
0.5
0.4
0.3
0.2
0.1
0
0 10 20 30 40 50 &0
iterations
Figure 6.24
The relative residual error versus DBlM iterations
The reconstructions obtained by considering the entire domain in
figure 6.20 are not acceptable. The reconstructions obtained employing
the reduced imaging domain in figure 6.23, though not ideal, clearly
show the location of the scatterer in the investigation domain. The
iterations are stopped after the 50
th
iteration as there was no significant
decrease in the relative residual error after that, as illustrated in figure
6.24. Thus the Images reconstructed employing the
DD.S formulations were found to be much superior, in addition to
significantly decreasing the computation time.
A photograph of the measurement setup m the anechoic
chamber is shown in figure 6.25.
Figure 6.25
The measurement setup in the anechoic
chamber facility ofDOE, CUSA T
References
1. Karg, R., "Experiments with a multi frequency mIcrowave
imaging system", Nachrichtentechnische Zeitschrifi, vol. 30, p.
651-654, Aug. 1977.
2. Larsen, L. E. and J. H. Jacobi, "Microwave scattering
parameter imaging on an isolated canine kidney", Medical
Physics, Vo!. 6, pp 394-403, 1979.
3. Chang, M., P. Chou, H. Lee, "Tomographic mIcrowave
imaging for nondestructive evaluation and object recognition
of civil structures and materials", 29th Asilomar Conlerence on
Signals, Systems and Computers (2-Volume Set), p. 1061,
1995.
4. Franchois, A, A. Joisel, C. Pichot and 1. C. Bolomey,
"Quantitative Microwave Imaging with a 2.45 GHz Microwave
camera", IEEE Transactions on Medical Imaging, Vo\. 17,
issue 4, pp 550-561, August 1998.
5. Semenov, S. Y., R. H. Svenson, A E. Boulyshev, A E.
Souvorov, V. Y. Borisov, Y. Sizov, A N. Starostin, K. R.
Dezern, G. P. Statsis and V. Y. Baranov, "Microwave
Tomography: Two dimensional system for biological
imaging", IEEE Transactions on Biomedical Engineering, Vol.
43, [ssue 9, pp 869-877, September 1996.
6. Semenov, S. Y., R. H. Svenson, A E. Boulyshev, A E.
Souvorov, A G. Nazarov, Y. Sizov, A E. Pavlovsky, V. Y.
Borisov, B. A Voinov, G. r. Simonova, A. N. Starostin, V. G.
Posukh, G. P. Statsis and V. Y. Baranov, "Three dimensional
Microwave Tomography: experimental prototype of the system
and Vector Born Reconstruction method", IEEE Transactions
VII DlOlIXCUICal cngUleermg, VOl. 4C>, Issue H, pp 937-946,
August 1999.
7. Belkebir, K. and M. Saillard, "Special section: Testing
Inversion Algorithms against Experimental Data", Inverse
Problems, Institute of Physics Publishing, vot. 17, pp 1565-
1571,2001.
8. Franchois, A. and A. G. Tijhuis, "A quasi-Newton
reconstruction algorithm for a complex microwave imaging
scanner environment", Radio Science, VOL. 38, NO. 2,2003
9. Meaney P.M., K. D. Paulsen, M. W. Fanning, D. Li, Q. Fang,
"Image accuracy improvements in microwave tomographic
thermometry: phantom experience", International Journal of
Hyperthermia, Taylor and Francis publications, Volume 19,
Number 5, pp. 534-550(l7)September-October 2003.
10. Li,X., S. K. Davis, S. C. Hagness, D. W. van der Weide and B.
D. Van Veen, "Microwave Imaging via Space-Time
Beamforming: Experimental Investigation of Tumor Detection
in Multilayer Breast Phantoms", IEEE Transactions on
Microwave Theory and Techniques, Vo\. 52, No. 8, pp 155 -
181, August 2004.
11. Tomohisa, 0., M. Takashi, S. Tsuyoshi, 1. Takashi,
"Microwave imaging with a two-way antenna", Papers of
Technical Meeting on Instrumentation and Measurement, lEE
Japan, Vol. IM-06, No.16-23, pp 1-6, 2006.
12. van den Berg, P., " Reconstruction of media posed as an
optimization problem", Wavejield Inversion, Springer
Publications, 1999.
13. Lu, C. C. and W. C. Chew, "Processing Ipswich data with
Local Shape Function Method", IEEE Antennas and
Propagation Magazine, Vol. 38, No. 3, pp 511-53, 1996.
14. Lobel. P, C. Pichot, L. Blanc-Fearud and M. Barlaud,
"Microwave imaging: Reconstructions from experimental data
using conjugate gradient and enhancement by edge preserving
regularization", International Journal of lmaging Systems
Technology, Vol. 8, pp 337-342, 1997.
IS. Fiddy, M. A, M. Testorf and U. Shahid, " Minimum Phase
based inverse scattering method applied to {PS 008", Image
Reconstruction from incomplete data /II, Proceeding of SPIE,
5562, pp 188-195, 2004.
16. Curtis, J. W., "The effect of pre orientation on the fracture
properties of glassy polymers", Journal of Physics D: Applied
Physics, Vol. 3, pp 1413-1422,1970.
17. Jeffries, D. J. and A. Kouloris, "Dielectric loading of ADR
Antennas, Experimental results", antenneX, online issue 69,
January 2003.
18. Jamwal, K. K. S and A Dhar, " Simple and fast technique for
measuring dielectric constant of polymers at microwave
frequencies", Review of Scientific Instruments, Vol. 52, Issue 5,
pp 767-769, May 1981.
19. Franchois A and C. Pichot, "Microwave lmaging-complex
permittivity reconstruction with a Levenberg-Marquardt
Method", IEEE Transactions on Antennas and Propagation,
Vol. 45, no.2, pp 203-215, February 1997.
20. Doughty, V.L and A. Vallance, Design of machine members,
4th edition, McGraw Hill, 1964.
CHAPTER 7
Conclusion and future directions
7.1 Introduction
The attempt to find out the internal structure of objects usmg
electromagnetic waves is an important aim of scientists in many
disciplines ranging from medical diagnostics to non destructive
testing. These problems involve the estimation of certain quantities
such as dielectric permittivity, conductivity etc based on measurement
of the scattered fields. The inverse problem of electromagnetic
imaging is highly non linear and the nonlinearity increases with the
dielectric contrast and size of the scatterer, as well as the frequency of
illumination. These inverse problems are ill-posed and the ill-
posedness increases when only a limited angle measurement of the
scattered data is possible, as in the case of imaging buried objects.
Therefore the electromagnetic mverse scattering problem IS
characterized by multiple minima. Thus any procedure to solve the
inverse problem of electromagnetic imaging is liable to get trapped in
local minima. The computation per iteration for finding the solution of
the inverse scattering process is of the order of N
J
where N is the
number of unknowns of the inverse scattering process. Thus
algorithms designed for computing the solution of the inverse
scattering problem of electromagnetic imaging are beset with
convergence and computational issues.
The thesis focused on methods for improving the
computational and convergence issues of inverse scattering algorithms
employed for electromagnetic imaging.
A multi-scale frequency hopping technique was proposed for
the purpose of obtaining globally convergent solutions using any
detenninistic iterative procedure. Since the deterministic procedures
employ Iinearized search methods to minimize a cost functional, they
are liable to get trapped in local minima. The Distorted Born Iterative
method and the Newton Kantarovich method were employed for the
detenninistic iterations in this work. The strategy is independent of the
exact detenninistic search technique employed. The multi-scaled
frequency hopping technique employed a multi-resolution search at a
frequency hop for the solution of the inverse scattering problem of
electromagnetic imaging. The stability of the solution is larger at a
coarser sampling rate. Hence its non uniqueness is less prominent at a
lower sampling rate. This solution was taken as the starting point for
iterations at increasing sampling rates. Thus the multi-resolution
strategy will be less prone to local minima problems than fixed grid
methods. However the employment of the cost functional may again
introduce local minima. Therefore the multi-scaling strategy was
combined with the frequency hopping method suggested by W C
Chew et a1. The combined strategy was tested on synthetic data as well
as experimental data provided by the Centre Commun de Ressourses
Micro-ondes (CCRM) of Marseille France. The results were compared
against those obtained using frequency hopping alone. The proposed
strategy yielded results that were closer to the ground truth and were
found to be better localized in the imaging domain. It was also
observed that the total computations per iteration were also
significantly reduced. This is attributed to the fact that the multi-
scaling strategy at the lower frequencies succeeded in providing a
better initial estimate for the iterations at the higher frequencies for
this object, than when the solution was computed at the finest scales
alone.
The Degree of symmetry formulation was introduced in
Chapter 5. The investigation domain where the scatterer is searched
for has to be sufficiently large to include all possible locations of the
scatterer. However when the investigation domain is large, the number
of unknowns of the inverse scattering problem of electromagnetic
imaging increases and a deterministic search to image the scatterer
may fail to reach a convergent solution. The total computation time
also increases when the number of elements to be reconstructed
increases. It was observed that the symmetry of the measured scattered
field vector could be used to localize the scatterer in the investigation
domain when the scatterer was symmetric, leading to the
D.o.S formulations. Synthetic scattered field data was generated to
verify the applicability of the formulations. A Probabilistic neural
network, trained with the D.o.S vectors, classified the scatterer as
belonging to a significantly reduced investigation domain. The entire
measured scattered field data was therefore available for the
reconstruction of fewer number of investigation domain elements.
Thus the degrees of freedom of the inverse scattering problem were
reduced, which aided the global convergence of the iterations. The
D.o.S formulations also resulted in more robust reconstructions of the
scatterer. The fonnulations were also applied to the problem of
_ __ _ ., .. "" .--...uuu, wnlcn gave
encouraging results.
An experimental setup was designed for perfonning the
verification of the D.Q.S formulations. The experimental results
confirmed the applicability ofthe formulations.
7.2 Shortcomings and Sources of Error
The sources of error and accuracy conditions encountered in the work
are detailed below.
).> The Degree of Symmetry formulation is strictly a two
dimensional formulation. The two dimensional model where a
2-D scatterer is illuminated by a TM polarized source, which
reduces the scattering equations to scalar ones, has to be
applicable. Therefore the D.Q.S formulations cannot be applied
for the reduction of imaging domain when the scatterer is 3
dimensional.
". Modeling the incident field with a line source: The
experimental verification employed a dipole antenna for
generating the incident field. More accurate representations
should use the field expressions of dipoles, which however,
would make the computations more involved.
;. Reflections from the measurement setup: Ideally the measured
scattered field should be contributed by the scattering from the
dielectric object whose cross section is to be profiled. To
minimize the scattering from the experimental setup, it is built
exclusively using PERSPEX, which is nearly lossless.
);- Less than ideal performance of the anechoic chamber: The non
ideal performance of the anechoic chamber would mean that
reflections from the chamber walls also have to be [aetored in.
7.3 Possible Directions
Some of the possible extensions that maybe pursued are listed
below
-" Stoehastie search algorithms are extremely slow compared
to deterministic search algorithms even though they
converge globally and hence were not used in this work. It
is suggested that stoehastic methods such as Genetic
algorithms or Simulated annealing be employed, at least at
coarse resolutions, to yield more convergent solutions.
,. Variable grid methods such as non linear multi-grid
relaxation based methods maybe tried out for multi-
resolution search of the solution of the inverse scattering
problem of electromagnetic imaging.
,. Proper selection of the regularization parameter is critical
for the quality of the electromagnetic inverse scattering
solution. Zeroth order regularization was employed in this
work, with an empirical formula used for computing the
value of the regularization parameter. More sophisticated
methods based on the stochastic model of the
reconstruction process maybe tried.
Reference
1. Agilent 8714 ET Network Analyzer operating and
programming manual
Publications
International Journals
1. Vi nu Thomas, C. Gopakumar, Anil Lonappan, G.Bindu, V
Hamsakutty, K.T Mathew " Microwave Irnaging of Two-
Dimensional Dielectric Cylinders with a Multiscaled Frequency
Hopping Approach" , Microwave and Optical Technology Letters
(USA) , Vo!. 43, No. 4, pp 353-355, 20
th
November 2004
2. Vinu Thomas, C. Gopakumar, A. V. Praveen Kumar, V.
Hamsakutty, Anil Lonappan, G. Bindu, K. T. Mathew "A novel
technique for reducing the imaging domain in microwave imaging of
two dimensional circularly symmetric scatterers" , Microwave and
Optical Technology Letters (USA) Vol. 44, No. 5, pp 423-427, 5
1h
March 2005
3. Vinu Thomas, C. Gopakurnar, Jaimon Yohannan, Anil Lonappan,
G. Bindu, A. V. Praveen Kumar, V. Hamsakutty, K. T. Mathew_"A
Novel Technique for Localizing the Scatterer in Inverse Profiling of
Two Dimensional Circularly Symmetric Dielectric Scatterers Using
Degrec of Symmetry and Neural Networks" Journal of"
Electromagnetic Waves and Applications(USA) Vo!. 19, No.15, pp
2113-2121,2005
4. Vinu Thomas, J. Yohannan, A. Lonappan, B. Bindu and
K.T.Mathew, "Localization of the investigation domain in inverse
profiling of buried 2-D Dielectric pipelines with circular cross
section using electromagnetic scattercd data", Progress in
Electromagnetic research, PIER 61, 111-131, 2006
5. V Hamsakutty, Anil Lonappan, Joe Jacob, Jaimon Yohannan, Vinu
Thomas, G.Bindu, K.T Mathew " A Novel Coupling medium for
Microwave Medical Imaging" , lEE Electronic Letters, Vo!. 39, No.
21, pp 1498-1499, 16
th
October 2003
6. G.Bindu , Anil Lonappan,V Hamsakutty, Vi nu Thomas, c.K.
Aanandan, K.T Mathew " Microwave Characterisation of Breast
Phantom Materials" , Microwave and Optical Technology Letters
(USA) Vol. 43, No.6, pp 506-508, 20
th
December 2004
7. G.Bindu, Vinu Thomas, Anil Lonappan, A. V. Pravccn Kumar, V
Hamsakutty, C.K. Aanandan, K.T Mathew " Two Dimensional
Microwave Tomographic Tmaging of Low Water Content Tissues" ,
Microwave and Optical Technology Letters (USA) Vol. 46, No.6, pp
599-60 I, 20
th
September 2005
8. G.Bindu, Santhosh John Abraham, Anil Lonappan, Vinu Thomas,
C.K. Aanandan, K.T Mathew "A pulsed confocal microwave
technique for the detection of dielectric contrast of breast tissue" ,
Microwave and Optical Technology Letters (USA) Vol. 47, No.3, pp
209-212, 5
th
November 2005
9. G.Bindu, Santhosh John Abraham, Anil Lonappan, Vinu Thomas,
C.K. Aanandan, K.T Mathew "Effects of reduced contrast coupling
medium in microwave breast imaging" , Microwave and Optical
Technology Letters (USA) Vol. 47, No.5, pp 443-446, 5
th
December
2005
10. G.Bindu, Santhosh John Abraham, Anil Lonappan, Vinu Thomas,
C.K. Aanandan, K.T Mathew, "Active microwave imaging for breast
cancer detection", Progress in Electromagnetics Research (USA),
PIERS 58, pp.149-169, 2006
11. G.Bindu, Anil Lonappan, Vinu Thomas, c.K. Aanandan, K.T
Mathew "Dielectric studies of corn syrup for application microwave
breast imaging", Progress in Electromagnetics Research (USA),
PIER 59, pp.175- 186,2006
12. G.Bindu, Vinu Thomas, Anil Lonappan, C.K. Aanandan, K.T
Mathew "Microwave studies of a Poly Vinyl Acetate (PV A)- Based
Phantom for Applications in Medical Imaging" , Microwave and
Optical Technology Letters (USA) Vol. 48, No.l, pp 180-183,
January 2006
13. Anil Lonappan, Vinu Thomas, G.Bindu, C. Rajasekharan, K.T
Mathew "Analysis of Human Cerebro Spinal Fluid at the ISM Band
of Frequencies" , Journal of Electromagnetic Waves and
Applications(USA) Vo\. 20, No.6, pp 773-779, 2006
14. G.Bindu, Ani! Lonappan, Vinu Thomas, c.K. Aanandan, K.T
Mathew, "Dielectric studies of polyvinyl acetate based phantom for
application in medical imaging", Journal of Material Science, Vo!.
41, No. 22, pp 7419-7424, November 2006.
IS. G. Bindu, Santhosh John Abraham, Anil Lonappan, Vinu Thomas,
C K Aanandan and K. T. Mathew, " Two dimensional microwave
tomogaraphic imaging of breast tissues", International Journal of
Cancer Research 2(1), 57-68, 2006.
16. G.Bindu, Santhosh John Abraham, Anil Lonappan, Vinu Thomas,
C.K. Aanandan, K.T Mathew, "Detection of dielectric contrast of
breast tissues using confocal microwave technique", Microwave and
Optical Technology Letters (USA) Vo\. 48, No.6, pp 1187-1190,
June 2006
International Conference Papers
I. Vi nu Thomas, C. Gopakumar, Jaimon Yohannan, Anil Lonappan,
G. Bindu, AV. Praveen Kumar, V. Hamsakutty, K. T. Mathew_"A
Novel Technique for Localizing the Scatterer in Inverse Profiling of
Two Dimensional Circularly Symmetric Dielectric Scatterers Using
Degree of Symmetry and Neural Networks" Electromagnetics
Research Symposium, PIERS 2005, August 22-26, 2005, Hangzhou,
China
2. Vinu Thomas, C. Gopakumar, A V. Praveen Kumar, V.
Hamsakutty, Jaimon Y ohannan, K. T. Mathew, " Imaging Domain
reduction in Microwave Imaging of two dimensional circularly
symmetric dielectric scatterers using degree of symmetry vector",
2004 December, Seventh International Conference on
Optoelectronics, Fiber optics and Photonics, Cochin, India.
3. V Hamsakutty, Anil Lonappan, Joe Jacob, G.Bindu , Vinu Thomas,
A V. Praveen Kumar, K.T Mathew " Biomedical Applications of
Sodium Meta Silicate Gel as Coupling medium for Microwave
Medical Imaging" , 2004 IEEE AP-S International Symposium and
USNCIURSI National Radio Science Meeting, APSIURSI 2004, June
20-26, 2004, Monterey, California, USA
4. Jaimon Yohannan, A V. Praveen Kumar, Vinu Thomas, V.
Hamsakutty, Anil Lonappan, K. T. Mathew "Half- Split Cylindrical
Dielectric Resonator Antenna" , 5
'11
Conference on Ferroeiectrics,
Ferroelectrics UK 2005, April 26- 27, 2005, University of Paisley,
Scotland, UK
5. A. V. Praveen Kumar, Jaimon Yohannan, Anil Lonappan, G. Bindu,
Vinu Thomas, V. Hamsakkutty, K. T. Mathew, "Microstripline Fed
Circular Sector Dielectric Resonator Antenna", IEEE AP-S
International Symposium and USNC I URSI National Radio Science
meeting, APSIURSI2005, July 3-8, 2005, Washington DC, USA
6. Anil Lonappan, Vi nu Thomas, G. Bindu, C. Rajasekharan, K. T.
Mathew," Human Colostrum at S-band of Microwave Frequencies"
Progress in Electromagnetics Research Symposium, PIERS 2007,
March 26 - 30, 2007, Beijing, China
7. Jaimon Yohannan, A. V. Praveen Kumar, Vi nu Thomas, V.
Hamsakutty and K. T. Mathew, "Synthesis of Dielectric Resonators
for Microwave Filter Designing", Progress in Electromagnetics
research symposium PIERS 2005, August 2005, Hangzhou, China.
National Conference Papers
1. C. Gopakumar, Vinu Thomas and K. T. Mathew, HA Reduction
Technique for imaging domain in microwave imaging of two
dimensional circularly symmetric scatterers", Proceedings of the
National Seminar on Information, Communication and Intelligent
Systems, February 2008, pp 86-91, Cochin.
2. C. Gopakumar, Vinu Thomas and K. T. Mathew, "A Multiscaled
frequency hopping approach in the microwave imaging of two
dimensional dielectric cylinders", Proceedings of the National
Seminar on Information, Communication and Intelligent Systems,
February 2008, pp 92-97, Cochin.
3. G.Bindu, Anil Lonappan,V Hamsakutty, A. V. Pravccn Kumar,
Vinu Thomas, c.K. Aanandan, K.T Mathew "Microwave imaging
of wax cylinders", December 2004, APSYM 2004, Antennas and
Propagation Symposium, Department of Electronics, Cochin
University of Science and Technology, Cochin, India