UNIVERSITY
OF TRENTO
DIPARTIMENTO DI INGEGNERIA E SCIENZA DELL’INFORMAZIONE
38123 Povo – Trento (Italy), Via Sommarive 14
http://www.disi.unitn.it
ON THE EFFECTS OF THE ELECTROMAGNETIC SOURCE
MODELLING IN THE ITERATIVE MULTISCALING METHOD
D. Franceschini, M. Donelli, and A. Massa
June 2007
in the Iterative Multis aling Method 2 3 4 5 6
Davide Fran es hini, Massimo Donelli,and Andrea Massa 7 8 9 10 11 12 13
Department of Information and Communi ation Te hnologies, 14
University of Trento,Via Sommarive 14,38050 Trento - Italy 15
Tel. +39 0461 882057,Fax +39 0461 882093 16
E-mail: andrea.massaing.unitn.it, {davide.fran es hini,massimo.donelli}dit.unitn.it 17
Web: http://www.eledia.ing.unitn.it 18
in the Iterative Multis aling Method
20
21
Davide Fran es hini, Massimo Donelli,and Andrea Massa 22
23
Abstra t 24
Thevalidationagainstexperimentaldataisafundamentalstepintheassessment of 25
theee tiveness of a mi rowave imaging algorithm. Itis aimed at pointing out the 26
limitationsofthe numeri al pro edureforasu essiveappli ation inareal environ-27
ment. Towards this end, this paper evaluatesthe re onstru tion apabilities of the 28
Iterative Multi-S aling Approa h (IMSA) when dealing withexperimental data by 29
onsidering dierent numeri al models of the illuminating setup. Infa t, sin e the 30
in ident ele tromagneti eld is usually olle ted in a limited set of measurement 31
points and inversion methods based on the use of the state equation require the 32
knowledge of the radiated eld in a ner grid of positions, an ee tive numeri al 33
pro edure for the synthesisof theele tromagneti sour e isgenerallyneeded. Con-34
sequently,theperforman esoftheinversionpro essmaybestrongly ae tedbythe
35
numeri almodeland,insu ha ase,agreat areshouldbedevotedtothiskeyissue
36
to guarantee suitableand reliable re onstru tions.
37
38
Keywords: 39
Mi rowaveImaging,InverseS attering,IterativeMulti-s aling Method,Sour eModeling. 40
Index Terms: 41
6982RadioS ien e: Tomographyandimaging;0629Ele tromagneti s: Inverses attering; 42
0669Ele tromagneti s: S atteringanddira tion. 43
Within the framework of the medi ine [Louis, 1992℄ and biomedi al engineering (see 45
for example [Liu et al., 2003℄ and the referen es ited therein), without forgetting the 46
industrial quality ontrol in industrial pro esses [Hoole et al., 1991℄ and the subsurfa e 47
sensing [Dubey et al., 1995; Daniels, 1996℄, many dierent appli ations require a non-48
invasivesensingofina essibleareas. Towardsthisend,mi rowaveimagingmethodologies 49
[Steinberg, 1991℄ have re ently gained a growing attention sin e they allow to retrieve 50
informationontheenvironmentprobedwithele tromagneti eldsby fullyexploitingthe 51
s attering phenomena [Colton and Kress, 1992℄. 52
Unfortunately, the inverse problem to be fa ed is intrinsi ally nonlinear, ill-posed, and 53
non-unique [Denisov, 1999℄. In parti ular, the ill-posedness and the non-uniqueness arise 54
fromthelimitedamountof information olle tableduringthea quisitionofthes attered 55
eld. Thenumberofindependents atteringdataislimited[Berteroetal.,1995;Bu iand 56
Fran es hetti, 1989℄ and they an onlybeused to retrieve anite numberof parameters 57
of the unknown ontrast fun tion. Tofully exploit su h aninformation and to a hieve a 58
suitableresolutiona ura y,severalmulti-resolutionstrategieshavebeenproposed[Miller 59
and Willsky, 1996a,1996b; Bu i et al.,2000a, 2000b;Baussard et al., 2004a,2004b℄. 60
The Iterative Multi-S aling Approa h belongs to this lass of algorithms [Caorsi et al., 61
2003℄. The unknown s atterers are iteratively re onstru ted by onsidering initially a 62
rough estimate of the diele tri distribution 1
and by enhan ing su essively the spatial 63
resolution in a set of regions-of-interest (RoIs) where the obje ts have been lo alized. 64
Su ha strategyis mathemati allyformulatedby dening asuitablemulti-resolution ost 65
fun tion whose global minimum is assumed as the estimated solution. The fun tional 66
is iteratively minimized by using a onjugate-gradient-based pro edure [Kleinman and 67
Van den Berg, 1992℄, but sto hasti [Massa, 2002℄ or hybrid algorithms an be suitably 68
applied. 69
In order to validate su h an approa h, the multi-resolution algorithm has been tested 70
against experimental data [Caorsi et al., 2004a℄ olle ted in a ontrolled environment 71
1
The IMSAisinitialized by onsidering thefreespa e distribution,then noa-priori information on the s enario under test is exploited. Moreover, the initialization of the intermediate steps is obtained from the re onstru tion of thepreviousstep with asimplemappingof the retrieved prolein the new dis retization oftheRoI.
indi ations and they modelanideal s enario. 73
In dealing with real data, one of the key issue is the modeling of the ele tromagneti 74
sour eoroftherelatedradiatedeld. Ingeneral,theele tromagneti eldemittedbythe 75
probing system is measured onlyin the observation domain. However, iterative methods 76
basedonDataandState equationsrequiretheknowledgeofthein identeld(i.e.,the 77
eld without the s atterers) generated from the sour e in the investigation domain. T o-78
wards this end, ana urate but simplymodel(i.e., requiringa reasonable omputational 79
burden) of the sour eshouldbe developed. Compli atednumeri almodelsa urately re-80
produ erealdata,buttheyaredi ulttobeimplementedstartingfromalimitednumber 81
of samples of the radiatedele tromagneti eld olle ted in a portion of the observation 82
domain. On the other hand, a rough model ould introdu e erroneous onstraints to 83
the re onstru tion pro ess. Nevertheless, whatever the sour e model, an ee tive inver-84
sion pro edure should be able to re onstru t the s atterer undertest with an a eptable 85
a ura y a ording toits robustness tothe noise. 86
In this framework, toassessthe ee tiveness and the robustness ofthe IMSA, the results 87
of aset of experiments, wheredierentmodels forapproximating the illuminatingsour e 88
are onsidered,will beshown. 89
The paper is organized as follows. In Se tion 2, the statement of the inverse problem 90
and the mathemati alformulationof the IMSA will bebriey resumed, while in Se tion 91
3 the numeri al models used to synthesize the probing ele tromagneti sour e will be 92
des ribed. A numeri al validation and an exhaustive analysis of the dependen e of the 93
re onstru tiona ura yonthemodelingoftheradiatedeldwillbe arriedoutinSe tion 94
4 by onsidering some experimental test ases. Finally, some on lusions will be drawn 95
(Se t. 5). 96
2 Mathemati al Formulation 97
The inversion pro edure willbeillustrated referringto atwo-dimensionalgeometry (Fig-98
ure 1). Let us onsider an investigation domain
D
I
, where an unknown s atterer is 99supposed to be lo ated. The embeddingmedium is assumed lossless, non-magneti ,and 100
hara terized by a diele tri permittivity
ε
0
. Su h a s enario is illuminated by a set of 101V
in ident mono hromati ele tromagneti eldsE
v
inc
(x, y)
,v = 1, ..., V
, and the orre-102sponding s attered elds
E
v
scatt
x
m
(v)
, y
m
(v)
,
v = 1, ..., V
, are available ( omputed asthe 103dieren ebetween the eldwith
E
v
tot
and withoutthes attererE
v
inc
,E
v
scatt
= E
tot
v
− E
inc
v
) 104in
m
(v)
= 1, ..., M
(v)
,v = 1, ..., V
,positionsbelongingtotheobservationdomainD
M
. The 105obje t is des ribed by a ontrast fun tion
τ (x, y) = ε
r
(x, y) − 1 − j
σ(x, y)
2πf ε
0
,
(x, y) ∈ D
I
, 106ε
r
(x, y)
andσ(x, y)
beingthediele tri permittivityandthe ele tri ondu tivity, respe -107tively. 108
The arising s attering phenomena are mathemati ally des ribed through the well-known 109
Lippmann-S hwingerintegral equations [Colton and Kress, 1992℄: 110 111 112
E
v
scatt
(x
m(v)
, y
m(v)
) = k
2
0
R
D
I
G
2d
(x
m(v)
, y
m(v)
|x
′
, y
′
)τ (x
′
, y
′
)E
υ
tot
(x
′
, y
′
)dx
′
dy
′
, m
(v)
= 1, ..., M
(v)
(x
m
(v)
, y
m
(v)
) ∈ D
M
v = 1, ..., V
(1) (Data Equation) 113 114 115E
v
inc
(x, y) = E
tot
v
(x, y) − k
2
0
Z
D
I
G
2d
(x, y|x
′
, y
′
)τ (x
′
, y
′
)E
tot
v
(x
′
, y
′
)dx
′
dy
′
(x, y) ∈ D
I
(2) (State Equation) 116 117where
G
2d
denotes the Green fun tion of the ba kground medium[Jones, 1964℄. 118Sin etheproblemasso iatedwith(??)isill-posed(see[Groets h,1993℄and[Vogel,2002℄) 119
the systemmatrixafterdis retizationoftheData Equation (a ordingtotheRi hmond's 120
pro edure[Ri hmond,1965℄)ishighlyill- onditioned,and,hen etheproblemisextremely 121
sensitive to the the noise. To remedy this ill- onditioning, a regularization is needed. 122
Thus, the problem is then reformulated in nding the unknown ontrast fun tion that 123
minimizesa suitable ost fun tiongenerally dened as follows 124
Φ {τ (x
n
, y) , E
tot
v
(x
n
, y
n
) ; n = 1, ..., N; v = 1, ..., V } =
=
P
V
v=1
P
M
(v)
m
(v)
=1
E
v
scatt
x
m
(v)
, y
m
(v)
−
P
N
n=1
n
τ (x
n
, y
n
) E
tot
v
(x
n
, y
n
) G
ext
2d
A
n
, ρ
nm
(v)
o
2
+
P
V
v=1
P
N
n=1
E
v
inc
(x
n
, y
n
) −
h
E
v
tot
(x
n
, y
n
) −
P
N
u=1
{τ (x
u
, y
u
) E
tot
v
(x
u
, y
u
) G
int
2d
(A
u
, ρ
un
)}
i
2
where
G
int
2d
andG
ext
2d
indi ate the dis retized forms of the internal and external Green's 125operators [Colton and Kress, 1992℄,
ρ
nm
(v)
=
r
x
n
− x
m
(v)
2
+
y
n
− y
m
(v)
2
,ρ
un
=
126q
(x
u
− x
n
)
2
+ (y
u
− y
n
)
2
andA
n
(A
u
) is the area of then
-th (u
-th) square dis retiza-127tion domain. In parti ular,the rstterm of(??) enfor esdelitytothe s attereddata in 128
the observation domain (
E
v
scatt
(x
m
(v)
, y
m
(v)
)
,(x
m
(v)
, y
m
(v)
) ∈ D
M
) and it amounts to the 129residual errorwith respe t tothe s attered eld omputed fromthe Data Equation (??). 130
The se ond term is a regularization term equal to the residual error with respe t to the 131
in ident eld in the investigation domain (
E
v
inc
(x
n
, y
n
)
,(x
n
, y
n
) ∈ D
I
) omputed from 132the StateEquation (??). 133
However, due to the limitedamountof information ontent inthe input data [Bu i and 134
Fran es hetti, 1989℄, it would be problemati toparametrize the investigationdomain in 135
terms of a large number
N
of pixel values (in order to a hieve a satisfying resolution136
levelinthere onstru ted image). Toover omethis drawba k, aninitialuniform( oarse)
137
dis retization is used and su essively an iterative parametrization of the test domain
138
allows to adaptively in rease the resolution level only in the region-of-interest of the
139
investigationareathusa hievingtherequiredre onstru tiona ura y[Caorsietal.,2003℄.
140
To retrieve the unknown s atterer (i.e., an obje t fun tion that better ts the problem 141
data, (
E
v
scatt
(x
m
(v)
, y
m
(v)
)
,E
v
inc
(x, y)
),Eqs. (??)and(??)are dis retizeda ordingtothe 142Ri hmond's pro edure [Ri hmond, 1965℄. Moreover, to better exploit the limited infor-143
mation ontent of the s attering data, an adaptive multi-resolution strategy is adopted 144
[Caorsi et al., 2003℄. 145
More in detail, su h an adaptive multi-resolution algorithm an be briey des ribed as
146
follows. Firstly, the IMSA onsiders (
i = 0
,i
being the step index) an homogeneous147
dis retization of the investigation domain with a number of dis retization domains
N
(0)
148
equal to the essential dimension of the s attered data and omputed a ording to the
149
riterion dened in [Isernia et al., 2001℄. Then, a oarse re onstru tion of the investi-150
gation domain is yielded by minimizing (??) starting from the free-spa e onguration
151
[
τ (x
n
(0)
, y
n
(0)
) = 0.0
andE
v
tot
(x
n
(0)
, y
n
(0)
) = E
inc
v
(x
n
(0)
, y
n
(0)
)
℄ in order toassess the robust-152ness ofthe overallapproa hwithrespe t tothestartingguess inworst- ase. Afterthe 153
minimization, where a set of onjugate-gradient iterations (
k
being the iteration index)(RoI),
D
O(i)
,i = 0
, is dened. Su ha squared area is entered at 156x
RoI
c
(i)
=
x
RoI
re
(i)
+ x
RoI
im
(i)
2
, y
RoI
c
(i)
=
y
RoI
re
(i)
+ y
RoI
im
(i)
2
(4) wherex
RoI
re
(i)
,x
RoI
im
(i)
, y
RoI
re
(i)
andy
RoI
im
(i)
are dened as157 158
x
RoI
ℜ(i)
=
P
R
r=1
P
N(r)
n(r)=1
n
x
n(r)
ℜ
h
τ
x
n(r)
,y
n(r)
io
P
N(r)
n(r)=1
n
ℜ
h
τ
x
n(r)
,y
n(r)
io
, R = i
(5) 159y
ℜ(i)
RoI
=
P
R
r=1
P
N
(r)
n
(r)
=1
n
y
n
(r)
ℜ
h
τ
x
n
(r)
, y
n
(r)
io
P
N
(r)
n
(r)
=1
n
ℜ
h
τ
x
n
(r)
, y
n
(r)
io
(6) 160and its side
L
(i)
is dened asfollows161 162
L
RoI
(i)
=
L
RoI
re
(i)
+ L
RoI
im
(i)
2
(7) 163 164L
RoI
ℜ(i)
= 2
P
R
r=1
P
N
(r)
n
(r)
=1
ρ
n(r)c(i)
ℜ
h
τ
x
n(r)
,y
n(r)
i
max
n(r)=1,..,N(r)
n
ℜ
h
τ
x
n(r)
,y
n(r)
io
P
R
r=1
P
N
(r)
n
(r)
=1
ℜ
h
τ
x
n(r)
,y
n(r)
i
max
n(r)=1,..,N(r)
n
ℜ
h
τ
x
n(r)
,y
n(r)
io
(8) 165where
ℜ
standsfortherealortheimaginarypartandρ
n
(r)
c
(i)
=
r
x
n
(r)
− x
RoI
c
(i)
2
+
y
n
(r)
− y
RoI
c
(i)
2
.
166Su essively, the iterative pro ess starts (
i → i + 1
). A ording to the multi-resolution167
strategy, an higher resolution level denoted by
R
(R = i
) is adopted only for the RoI.168
D
O(i−1)
is dis retized inN
(i)
square sub-domain whi h number is always hosen equal to169
the essential dimension of the s attered data [Bu i and Fran es hetti, 1989℄. A ner
170
obje tfun tionproleisthenretrieved,startingfromthe oarserre onstru tion a hieved
171
at the (i-1)-th step, by minimizing the multi-resolution ost fun tion,
Φ
(i)
, dened as
Φ
(i)
τ
(i)
x
n
(r)
, y
n
(r)
, E
tot
v
(i)
x
n
(r)
, y
n
(r)
;
r = 1, ..., R = i;
n
(r)
= 1, ..., N
(r)
;
v = 1, ..., V
=
=
nP
V
v=1
P
M
(v)
m
(v)
=1
E
scatt
v
x
m
(v)
, y
m
(v)
−
P
R
r=1
P
N
(r)
n
(r)
=1
n
w
x
n
(r)
, y
n
(r)
τ
(i)
x
n
(r)
, y
n
(r)
E
tot
v
(i)
x
n
(r)
, y
n
(r)
G
ext
2d
A
n
(r)
, ρ
n
(r)
m
(v)
o
2
+
nP
V
v=1
P
R
r=1
P
N
(r)
n
(r)
=1
n
w
x
n
(r)
, y
n
(r)
E
v
inc
x
n
(r)
, y
n
(r)
−
h
E
tot
v
(i)
x
n
(r)
, y
n
(r)
−
P
N
(r)
u
(r)
=1
n
τ
(i)
x
u
(r)
, y
u
(r)
E
tot
v
(i)
x
u
(r)
, y
u
(r)
G
int
2d
A
u
(r)
, ρ
u
(r)
n
(r)
oi
o
2
(9) where 174w(x
n
(r)
, y
n
(r)
) =
0
if (x
n
(r)
,
y
n
(r)
) /
∈ D
O(i−1)
1
if (x
n
(r)
,
y
n
(r)
) ∈ D
O(i−1)
and
R
indi ates the resolution level andD
O(i)
denotes the area of the RoI dened at175
the i-th step of the iterative pro edure. It should be pointed out that the denition of
176
(??) requires not only the knowledge of the available s attered eld in the observation
177 domain [
E
v
scatt
x
m
(v)
, y
m
(v)
= E
v
tot
x
m
(v)
, y
m
(v)
− E
v
inc
x
m
(v)
, y
m
(v)
,x
m
(v)
, y
m
(v)
∈
178D
M
℄, but alsothat of the in ident eld inD
O(i)
[E
v
inc
(x
n
(r)
, y
n
(r)
)
,(x
n
(r)
,
y
n
(r)
) ∈ D
O(i−1)
℄.179
This latter information is generally not available from measurements [sin e, in general,
180
only the samples of
E
v
inc
x
m
(v)
, y
m
(v)
other thanE
v
tot
x
m
(v)
, y
m
(v)
are experimentally 181measured℄, thereforeitshouldbesyntheti allygenerated by meansof asuitablemodelof
182
the ele tromagneti sour e.
183
The multi-steppro ess ontinues by omputing a new RoI a ording to(??)(??) and by
184
estimating anew diele tri distribution through theminimization ofthe updated version
185
of (??) untila "stationary re onstru tion"is rea hed [Caorsi et al.,2003℄ (
i = I
opt
) .186
Su hapro edure an beextendedtomultiple-s atterersgeometriesby onsideringa
suit-187
able lustering pro edure [Caorsi etal.,2004b℄aimedatdeningthenumberofs atterers
188
Q
belonging to the investigation domain and the regionsD
(q)
O(i)
,q = 1, ..., Q
, where the189
syntheti zoomwillbe performed atea h step of the iterativepro ess.
Thein identelddataplaya ru ialroleintheimagingpro esssin etheknowledge/availability 192
of
E
v
inc
(x, y)
in the investigationdomainadds new information. Infa t, asit an be no-193ti ed in the equation dening the multi-resolution ost fun tion (??), it allows to dene 194
another onstraint (??) for the problem solution then redu ing the ill-posedness of the 195
inverseproblem[BerteroandBo a i,1998℄sin esu haterm anbealso onsideredasa 196
sort ofregularizationterm. Clearly,anerroneousorimpre iseknowledgeofthe in ident 197
eld ould onsiderably ae t the reliability of the fun tional and onsequently of the 198
overall imaging pro ess sin e (??) ontrols the minimization pro edure. As a matter of 199
fa t, inmany pra ti alsituations, the in ident eld isonly availableatthe measurement 200
points belonging to the observation domain,
E
v
inc
x
m
(v)
, y
m
(v)
,x
m
(v)
, y
m
(v)
∈ D
M
. 201Su h a situation is ommonly en ountered when dealing with real data be ause of the 202
omplexityanddi ultiesin olle tingreliableandindependentmeasures inadensegrid 203
of points. Hen e, to fully exploit the knowledge of the in ident eld and before fa ing 204
with the data inversion, it is mandatory to develop a suitable model able to predi t the 205
in ident eld radiated by the a tual ele tromagneti sour e in the investigationdomain, 206
E
v
inc
(x, y)
,(x, y) ∈ D
I
. Towards this aim, in the referen e literature (see [Belkebir and 207Saillard, 2001℄ and the referen es ited therein), dierent solutions have been proposed. 208
They aremainlybasedonplaneor ylindri alwavesexpansions, sin efar-eld onditions 209
are usually satised. In this paper, su h models willbe analyzed and a new distributed 210
modelwill beproposed. More in detail, letus onsider 211
•
thePlane-Waves Model (PW-Model)wherethein identeldismodeledasthe 212superposition of aset of
W
planewaves 213E
υ
inc
(x, y) =
W
X
w=1
A
w
e
−jwk
0
(xcosθ
v
+ysinθ
v
)
(10)θ
v
being the in ident angle,k
0
the free-spa e propagation onstant, andA
w
the 214amplitude of
w
-th wave; 215•
the Con entri -Cylindri al-Waves Model (CCW-Model) wherethe radiated 216eld is represented through the superposition of ylindri alwaves a ording tothe 217
E
υ
inc
(x, y) =
W
X
w=−W
A
w
H
w
(2)
(k
0
ρ) e
jwφ
v
(11)where
A
w
is an unknown oe ient,H
(2)
w
indi ates the se ond kindw
-th order 219Hankel fun tion,
ρ
is the distan e between the observation point lo ated at(x, y)
220
and the phase enter of the radiating system where the
w
-th line sour e is pla ed221
and
φ
v
the orresponding angle;222
•
the Distributed-Cylindri al-Waves Model (DCW-Model) where the a tual 223sour e is repla ed with alinear array of equally-spa ed line-sour es, whi h radiates
224
an ele tri eld given by
225
E
υ
inc
(x, y) = −
k
2
0
8πf ε
0
W
X
w=1
A (x
w
, y
w
) H
0
(2)
(k
0
ρ
w
)
(12)where
A(x
w
, y
w
)
is the unknown oe ientrelated tothew
-th elementandρ
w
the 226distan e between the observation point and the
w
-th line sour e. 227Su hmodelsare ompletelydenedwhenthesetofunknown oe ients,
A
w
orA(x
w
, y
w
)
, 228have been determined. Therefore, the solution of an inverse sour e problem, where the 229
known terms are the values of the in ident eld measured in the observation domain 230
E
v
inc
x
m
(v)
, y
m
(v)
, is required. More indetail, the followingsystem has tobe solved: 231
E
v
inc
(x
1
, y
1
)
...
...
E
v
inc
(x
m
(v)
, y
m
(v)
)
...
...
E
v
inc
(x
M
(v)
, y
M
(v)
)
=
G
11
...
G
1s
...
G
1S
...
...
...
...
...
...
...
...
...
...
G
m1
... G
ms
... G
mS
...
...
...
...
...
...
...
...
...
...
G
M1
... G
M s
... G
M S
I
1
...
...
I
s
...
...
I
S
(13)or ina more on iseform 232
where (a) for the PW-model
G
ms
= e
−jsk
0
d
m
,
d
m
= x
m
cosθ
v
+ y
m
sinθ
v
, andI
s
=
233A
s
,s = 1, ..., S
,S = W
; (b) for the CCW-modelG
ms
= H
(2)
s
(k
0
ρ
m
) e
jsφ
v
,ρ
m
=
234q
(x
m
− x
source
)
2
+ (y
m
− y
source
)
2
,(x
source
, y
source
)
being the lo ation of the sour e, and 235I
s
= A
s−1−W
,s = 1, ..., S
,S = 2W +1
;( )fortheDCW-modelG
ms
= −
k
2
0
8πf ε
0
H
(2)
0
(k
0
ρ
ms
)
, 236ρ
ms
=
q
(x
m
− x
s
)
2
+ (y
m
− y
s
)
2
, andI
s
= A (x
s
, y
s
)
,s = 1, ..., S
,S = W
. 237Unfortunately, (??)involvesthe limitationstypi al of an inverse-sour e problem(see for 238
example, [Devaney and Sherman, 1982℄). In parti ular,
[G]
is ill- onditioned and the 239solution isusually non-stable and non-unique. Now, the problem of determining
[I]
from 240the knowledge of the in ident eld an be re ast as the inversion of the linear operator 241
[G]
through the SVD-de omposition [Natterer, 1986℄ 242 243[I] = [G]
+
[E]
(15) where 244[G]
+
= [V] [Γ]
−1
[U]
∗
(16) and 245 246[Γ]
−1
=
1/γ
1
...
0
...
1/γ
s
...
0
...
1/γ
S
(17)Owing of the propertiesof
[G]
, the sequen e of singularvalues{γ
s
}
S
s=1
willbe de reasing 247and onvergenttozero. Consequently,thesolutionofequation(??)doesnot ontinuously 248
depend on problemdata and the unavoidable presen e of the noise, due to measurement 249
errors as well as to an ina urate model of the experimental setup, ould produ e an 250
unreliable sour e synthesis. 251
In the next se tion, anexhaustivenumeri alanalysiswill be arried out toassess the ro-252
bustness oftheIMSAagainsttheerrorinthein identelddataandtobetterunderstand 253
how and how mu h the model of the a tual ele tromagneti sour e ae ts the IMSA 254
performan es. 255
In this se tion,su hanassessmentwillbe performedby onsideringdierenttargetsand 257
starting from experimental data. The s attered data refers to the dataset available at 258
the Institute Fresnel - Marseille, Fran e. As des ribed in[Belkebir and Saillard, 2001; 259
Testorf and Fiddy, 2001; Marklein et al., 2001℄ and sket hed in Figure 2, the bistati 260
radar measurement system onsists of an emitting antenna pla ed at
r
s
= 720 ± 3mm
261
from the enter of the experimental setup and a re eiver whi h olle ts equally-spa ed
262 (
5
◦
)measurements ofE
v
tot
x
m
(v)
, y
m
(v)
andE
v
inc
x
m
(v)
, y
m
(v)
ona ir ularinvestigation 263domainofradius
r
m
= 760±3mm
. Notethepresen eofablind-se torofθ
l
= 120
◦
around 264
the emittingantenna (Figure 2). The s atterers onsidered in the following experiments 265
are shown inFigs 2(a)-( )for referen e. 266
In the rst example [Fig. 2(a)℄, we will onsider the ir ular diele tri prole (
L
ref
=
26730 mm
in diameter) positioned about30 mm
from the enter of the experimental setup 268(
x
c
ref
= 0.0, y
c
ref
= −30 mm
)and hara terizedbyahomogeneouspermittivityε
r
(x, y) =
2693.0 ± 0.3
[τ (x, y) = 2.0 ± 0.3
℄. The square investigation domain,L
DI
= 30 cm
sided, 270is partitioned in
N = 100
homogeneousdis retization domains and the re onstru tion is 271performed by exploiting all the available measures (
M
(v)
= 49
,v = 1, ..., V
) and views 272(
V = 36
), but using mono-frequen y data (f = 4 GHz
). 273Theperforman esoftheIMSAintermsofquantitativeaswellasqualitativeimaginghave
274
beenassessed onsideringne essarilytheStateTerm
2
duringtheminimizationofthe ost
275
fun tion (??) and thusintrodu ingthe information- ontent of the in ident ele tri eld.
276
To do this, two simple models for the eld emitted by the probing antenna have been 277
preliminary taken into a ount. The rst one represents the radiated eld with a plane 278
wave(
W = S = 1
),theotherwitha ylindri alwave(W = 0
,S = 1
). Theamplitudesof 279the modeledin identwavesare estimateda ordingtotheSVD-basedpro eduredetailed 280
in Se t. 3startingfromthe knowledgeof the values of the in identeld measured inthe 281
forwarddire tionand availabledire tly fromtheexperimentaldataset. They turn out to 282 be
A
(P W −M odel)
w=1
= 1.23
andA
(CCW −M odel)
w=0
= 17.27
, respe tively. 283In spite of the ina ura y inreprodu ing the values of the in ident eld olle ted at the 284
2
Some examples of algorithms employing only the Data Term an be found in the spe ial se tion
to lo alize the unknown targetwith asatisfa tory degree of a ura y as shown in Fig. 4
286
and onrmed by the geometri parameters reportedin Tab. I.
287
As far as the single-plane-wave model is on erned, it should be pointed out that the 288
re onstru ted ontrast 3
is hara terized by an average value of the obje t fun tion equal
289
to
τ = 2.1
, then very lose to the a tual value of the real target. However, several290
pixelsbelongingtotheareaof thereferen eprolepresent alargerobje tfun tionvalues
291
[
τ (x
n
, y
n
) = 2.5
℄and theretrievedobje t ontourdoesnot a uratelyreprodu ea ir ular292
shape.
293
With respe t to the PW model, a better re onstru tion is obtained when a little more
294
omplexsour emodel(i.e.,thesingleCW-Model)isusedasit anbeobservedinFig. 4(b) 295
and inferred from the values of the error gures (whi h quantify the qualitative imaging 296
of the s atterer undertest) given in Tab. II and dened as follows 297
ρ
(q)
=
rh
x
(q)
c
(Iopt)
− x
(q)
c
ref
i
2
+
h
y
c
(q)
(Iopt)
− y
(q)
c
ref
i
2
λ
q = 1, ..., Q
(I
opt
)
(18) 298∆
(q)
=
L
(q)
(I
opt
)
− L
(q)
ref
L
(q)
ref
× 100 q = 1, ..., Q
(I
opt
)
(19) where the sub-s ript ref refers tothe a tualprole.299
A ording to the indi ations drawn fromthese experiments,whi h point out that even a 300
rough representation of the in ident eld signi antly benets the inversion of the s at-301
tered eld data, the su essive pro edural step will be aimed at rening the numeri al 302
modelof the ele tromagneti sour e to further improve the ee tiveness of the retrieval 303
pro ess. However, it should be noti ed out that using a wrong, even though omplex, 304
model might a tually degrade the re onstru tion, thus great are is needed in dening 305
the most suitable omplex model. In order to point out su h a on ept, the problem 306
has beenstudied onsideringthe previouss atteringgeometry, butusingnumeri al mea-307
sured data with a ontrollable degree of noise. More in detail, the following analysis 308
has been arried out. Dierent ele tromagneti sour es have been onsidered to illumi-309
nate the s enarioundertest (i.e.,PW-Sour e, CCW-Sour e,and DCW-Sour e)and 310
3
domain
E
v
inc
x
m
(v)
, y
m
(v)
,x
m
(v)
, y
m
(v)
∈ D
M
,varioussour emodels (i.e.,PW-Model, 312CCW-Model, and DCW-Model) have been synthesized. Then, a noise hara terized 313
bya
SNR = 20 dB
hasbeensuperimposedtothedataandthere onstru tionpro esshas 314been arried out starting from the dierent sour e models previously determined. The 315
obtainedresultsintermsofqualitative(??)-(??)andquantitativeerrorgures
ξ
(j)
dened 316 as 317ξ
(j)
=
R
X
r=1
1
N
(r)
(j)
N
(r)
(j)
X
n
(r)
=1
(
τ (x
n
(r)
, y
n
(r)
) − τ
ref
(x
n
(r)
, y
n
(r)
)
τ
ref
(x
n
(r)
, y
n
(r)
)
)
× 100
R = S
opt
(20) whereN
(j)
(r)
an range over the whole investigation domain (j ⇒ tot
), or over the area 318where the a tuals atterer is lo ated (
j ⇒ int
), or over the ba kground belonging tothe 319investigationdomain(
j ⇒ ext
), arereportedinTab. III.Asexpe ted,the useofamodel 320orresponding to the a tual sour e turns out to be the most suitable hoi e and more 321
omplex modeling ause larger errors. As an example, let us onsider the PW-sour e. 322
When theproleretrievalisperformedusingthePW-model thenthere onstru tion error
323
is equal to
ξ
tot
= 0.30
. Otherwise,ξ
(DCW −M odel)
tot
= 13.30
andξ
(CCW −M odel)
tot
= 20.53
.324
Similar on lusions hold true also for other illuminations and sour e models in terms of
325
quantitative error gures,as well.
326
Consequently,the more omplexsour e ongurations des ribed inSe tion3, whi h
on-327
siderthesuperpositionofplanewavesorof ylindri alwaves, havebeentakenintoa ount
328
in ordertodene the mostsuitablesour e model. Insu h aframeworksin e the
numeri-329
aldes ription ofthe a tualsour e inthe real measurementsetup isonlypartiallyornot
330
generally available, the optimalmodelhas tobe dened by lookingfor the most suitable
331
numberof the unknown sour e oe ients,
S
, and orrespondingvalues,A
s
,s = 1, ..., S
.332
For ea h of the sour e models,
S
has been hosen by lookingfor the onguration that333
provides a satisfa tory mat hing between measured and numeri ally- omputed values of
334
the in ident eld in the observation domain. Su h a mat hing has been evaluated by
335
omputing the following parameter
µ = (V M
(v)
)
−1
P
V
v=1
P
M
(v)
m
(v)
=1
h
Re
n
E
v
inc
x
m
(v)
, y
m
(v)
o
− Re
n
E
g
v
inc
x
m
(v)
, y
m
(v)
oi
2
+
h
Im
n
E
v
inc
x
m
(v)
, y
m
(v)
o
− Im
n
E
g
v
inc
x
m
(v)
, y
m
(v)
oi
2
1
2
(21) 338where
Re {·}
andIm {·}
stand for the real and imaginary part, respe tively, and the339
super-s ript
e
indi ates anumeri ally-estimated quantity.340
In Fig. 5, the behavior of the mat hing parameter is displayed for dierent sour e
341
models. As an be observed,
µ
redu es whenS
in reases. Thus, the optimal number of342
sour e oe ients,
S
opt
,hasbeenheuristi ally-denedasthevaluebelongingtoastability343
region. Consequently, the optimal values have been set to:
S
(P W −M odel)
opt
= 20
(where 344µ ≃ 4 × 10
−4
)andS
(CCW −M odel)
opt
= 11
(whereµ ≃ 10
−4
). Theamplitudesofthe weighting
345
sour e oe ients are shown in Fig. 6. The magnitudes of the CCW-Model oe ients
346
[Fig. 6(b)℄areverylargewhen omparedtothoseofthesinglePW-Model orsingle
CCW-347
Model. Asexpe ted,the orrespondingradiated-elddistributionsinsidetheinvestigation
348
domain
D
I
[Figs. 7( ),(d)℄ turn out to be una eptable (for omparison purposes, the349
plot of the in ident ele tri eld omputed by means of the single CCW-Model is given
350
in Figs. 7(e),(f)). Moreover, Figs. 7(a),(b) show howeven the in identeld synthesized
351
bymeansofthe PW-Model presentsratherhighvalues withrespe ttothe distributionof
352
Figs. 7(e),(f). Sin ethein identeldistheguessvaluefortheoptimizationoftheinternal
353
eld, a ompletelywrongstartingdistributionmay onsiderablyae tthe wholeretrieval
354
pro edure. A ordingly, the adopted inversion strategy is not able to orre tly estimate
355
neitherthe shapenor thediele tri distributionofthe unknown s atterer (Fig. 8). As far
356
as the ase related to the PW-Model is on erned, it should be noted that the iterative
357
pro ess isstopped atthe fourth step(Tab. I)and the qualityof the re onstru ted prole
358
(Fig. 8(a))turns out to be strongly redu ed (if ompared to that of Fig. 4(a)) in terms
359
of qualitativeas wellas quantitativeimaging. Similarindi ations an be drawn fromthe
360
analysis of the retrieved distribution obtained with the CCW-Model. However, redu ing
361
the number ofterms inthe expansion ouldlead tobetter resultslike,forexample,those
362
presented in the spe ial se tion [Belkebir and Saillard, 2001℄ and those obtained in this
used in the proposed experiments.
365
The obtained dis ouraging results an be properly motivated by observing the
singular-366
valuesspe trum(Fig. 9)andby omputingthe onditionnumber
η
ofthelinearmatrixop-367
erator
[G]
(denedasfollowsη =
max
p
{σ
p
}
min
p
{σ
p
}
),whi h learly pointoutanintrinsi instability
368
of the system and the ill- onditioningof the problem. Inmore detail, the ill- onditioning
369
index turns out to be equal to
η
(P W −M odel)
= 41.07
and toη
(CCW −M odel)
= 5.62 × 10
7
, 370 respe tively. 371Apossiblesolutionforsuitablydeningthesour emodeland, onsequently,forimproving
372
the resolution a ura y of the retrieval pro ess (alternative to employ a trun ated-SVD
373
regularizationalgorithmassuggestedbythestep-likebehaviorofthesingular-values
spe -374
trum), isto dene aspatially-distributed line-sour e modelas des ribed in Se t. 3.
375
A ordingtothepro edurefor hoosingthenumberaswellasthemagnitudeofthesour e
376
weights previously des ribed, a reasonable onguration is
S
(DCW −M odel)
opt
= 15
(Fig. 5)377
with the oe ients distributed as shown in Fig. 10(a). For ompleteness, in order to
378
givean idea of the tting between measured and omputed data, Figs. 10(b)-( )display
379
the values of the amplitude and phase of the radiated-eld omputed in the observation
380
domain. Moreover, Fig. 11 gives a gray-level representation of the in ident ele tri eld
381
synthesized in the investigationdomain.
382
The use of su h a model for the in ident eld allows a signi ant improvement in the
383
re onstru tion. Su haresult anbeappre iatedinFig. 12wherethegray-level
represen-384
tation of the obje tfun tion isgiven. In parti ular, forthis representative onguration,
385
also the intermediate re onstru tions [Figs. 12(a)-( )℄ of the multi-s aling pro ess are
386
reported in order to show how the prole improves during the iterative pro edure. As
387
it an be noti ed, even though the omputationaldomain isnot nely dis retized atthe
388
st step [Fig. 12(a)℄,the IMSA iterativelyin reases the resolution inthe RoI inorder to
389
obtain an a urate dis retization at the onvergen e step [Fig. 12( )℄ where a
meaning-390
ful prole is obtained. As a matter of fa t, the lo alization as well as the dimensioning
391
error of the onvergen e step [Fig 12( )℄redu es with respe t to the other sour emodels
392
(
ρ
(DCW −M odel)
= 0.045λ
0
,∆
(DCW −M odel)
≈ 9
-Tab. II)and thehomogeneityofthe a tual
su h anapproa hwith respe t tothe othersour e-synthesis modalities is on erned,it is
395
mainly motivated by the faithful and stable reprodu tion [Figs. 10(b)-( )℄ of the a tual
396
values of the eld measured inthe observation domain.
397
To further assess the robustness and the ee tiveness of the IMSA, by validating the
398
radiated-eldsynthesisaswell,these ondexample onsidersamultiple-s attererss enario
399
(twodielTM_8f.exp -[Belkebir andSaillard,2001℄). Under thesame assumptionsofthe
400
previousexampleintermsofmeasures,radiationfrequen y,andviewsaswellasextension
401
and partitioning ofthe investigationdomain,two diele tri (
τ
(q)
= 2.0 ± 0.3
,
q = 1, ..., Q
,402
Q = 2
) ir ular(L
(q)
ref
= 30 mm
indiameter) ylinders are pla ed90 mm
from ea h other403
[Fig. 2(b)℄.
404
Fig. 13 shows the results of the re onstru tion pro ess in orresponden e with dierent
405
sour e models. As an be seen, whatever the stable sour e synthesis the two targets
406
are orre tly lo ated and dimensioned with a satisfa tory a ura y. Certainly, the more
407
sophisti atedsynthesis approa h(DCW-Model -
S = 15
)allows toobtain abetterre on-408
stru tion as onrmed by the geometri parameters of the retrieved proles resumed in
409
Tab. IV. In order toshow the apabilities of the IMSA in estimating the lossless nature
410
ofthe diele tri s atterers, there onstru tion orrespondingtothe DCW-Model hasbeen
411
run using a blind inversion s heme, that is without a-priori informationof its
hara ter-412
isti s. Su h assumption does not exploit the alternative denition of the solution spa e,
413
whi h allows to re onstru t only the real part of the obje t fun tion. A ordingly, Fig.
414
13(d) points out that the minimum of the imaginary part of the obje t fun tion is
0.08
415
( orresponding to
σ = 1.78 × 10
−3 S
m
).416
Finally, inorder to omplete the validationof the approa h,the lastexample deals with
417
a metalli stru ture. The s atterer is an U-shaped metalli ylinder [Fig. 2( )℄ and the
418
re onstru tion is performed starting from the omplete data olle tion of the dataset
419
uTM_shaped.exp [Belkebir and Saillard, 2001℄at the working frequen y of
f = 4 GHz
.420
A ordingtothestrategyproposedin[VandenBergetal.,1995℄,onlytheimaginarypart
421
of the obje t fun tion has been retrieved onsidering a lower bound in the re onstru ted
422
ontrast and if at some iterationthe estimated
Im {τ (x, y)}
is lowerthanτ
max
Im
= −15.0
,then the ontrast is repla ed by
τ
max
Im
. As a result, the imaginary part of the retrieved424
proleinthe ongurationwith theDCW-Model for thesynthesis ofthe radiatedeld, is
425
depi ted inFig. 14. Atthe onvergen e step(
I
opt
= 4
),the re onstru tion learly reveals426
that we are dealing with a U-shaped target. The outer and the inner ontour of the U
427
are well reprodu ed (even though little artifa ts appear) onrming the ee tiveness of
428
approa h inshaping and lo ating diele tri aswell metalli s atterers.
429
5 Con lusions 430
The Iterative Multi-S aling Approa h has been tested against experimentally-a quired 431
data byfo usingtheattentiononitsrobustness asregardsdierentmathemati almodels 432
usedtosynthesizethein identele tri eld. Theee tivenessoftheiterativeminimization
433
of the ost fun tional in re onstru ting unknowns s atterers presents a ertaindegree of
434
sensitivity to the model of the in ident eld used to formalize the onstraint stated by
435
the StateEquation. By onsideringamore omplexapproximation model(DCM-Model),
436
satisfa tory lo alizations and re onstru tions have been arried out by indi ating the
437
positiveee t of asuitable synthesis methodology on the inversion pro ess. 438
However, even thoughana urate approximationmodelgenerallymightresultina more 439
a urate re onstru tion, whi h omplex model is more appropriate for the in ident eld 440
may depend on the measurement setup, espe ially the mi rowave sour e onguration. 441
For example, for simple plane-wave in ident eld, using the PW-model might redu e 442
artifa ts whi h result from measurement noise. So future investigations are needed by 443
onsideringotherexperimentaldatasets( urrently not-available,but underdevelopment) 444
to generalize the on lusions of su han analysis. 445
Moreover,theresultsofthenumeri alanalysis arriedoutinthepaperandthe omparison
446
with there onstru tions obtainedinthe relatedliteraturesuggestthatimproved imaging
447
te hniques (e. g., multi-frequen y te hniques) or additional regularization terms may
448
probablydiminishtheimpa tofthein identeldmodel. Sin ethispointhasnot dire tly
449
investigated other resear hes will be aimed at further improving the ee tiveness of the
450
IMSA by onsidering multi-frequen y strategies, further regularization terms and more
451
ee tiveoptimizationalgorithmsfortheminimizationofthemulti-resolution ostfun tion
455
Baussard, A., E. L. Miller, and D. Lesselier(2004a),Adaptivemultis aleapproa h
456
for2Dmi rowavetomography,URSI - InternationalSymposium onEle tromagneti
The-457
ory,Pisa, Italia,pp. 1092-1094.
458
Baussard, A., E. L. Miller, and D. Lesselier (2004b), Adaptive multis ale
re on-459
stru tion of buriedobje ts, Inverse Problems,20,S1-S15.
460
Belkebir, K., and M. Saillard (2001), Spe ial se tion: Testing Inversion Algorithms 461
against Experimental Data, InverseProblems, 17,1565-1702. 462
Bertero, M., C. De Mol, and E. R. Pike (1995), "Linear inverse problems with 463
dis retedata. I: Generalformulationand singularsystem analysis,"Inverse Problems, 1, 464
301-330. 465
Bertero, M., and P. Bo a i (1998), Introdu tion to Inverse Problem in Imaging, 466
IoP Publishing, Philadelphia. 467
Bu i, O. M., and G. Fran es hetti(1989),On the DegreesofFreedomofS attered 468
Fields, IEEE Trans. AntennasPropagat.,37, 918-926. 469
Bu i, O. M., L. Cro o, T. Isernia, and V. Pas azio (2000a), Wavelets in non-470
linear inverse s attering, Pro . IEEE Geos ien e and Remote Sensing Symp., IGARSS-471
2000, 7, 3130-3132. 472
Bu i, O. M., L. Cro o, and T. Isernia (2000b), An adaptive wavelet-based ap-473
proa h for non destru tive evaluation appli ations, Pro . IEEE Antennas Propagation 474
Symp., APS-2000, 3,1756-1759. 475
Caorsi, S., M. Donelli, D. Fran es hini, and A. Massa (2003), A new methodol-476
ogy based on an iterative multi-s aling for mi rowave imaging, IEEE Trans. Mi rowave 477
Theory Te h., 51,1162-1173. 478
Caorsi,S., M.Donelli,andA.Massa(2004a),Analysisofthestabilityandrobustness
479
ofthe iterativemultis alingapproa hformi rowaveimagingappli ations,Radio S i.,39,
480
1-17.
481
Caorsi, S., M. Donelli, and A. Massa (2004b),Lo ation, dete tion, and imaging of 482
multiples atterersbymeansoftheiterativemultis alingmethod,IEEETrans. Mi rowave 483
Colton, D., and R. Krees (1992), Inverse a ousti s and ele tromagneti s attering 485
theory, Berlin, Germany: Springer-Verlag. 486
Daniels, D. J. (1996), Surfa e penetrating radar, IEE Ele tron. Comm. Eng. J., 8, 487
165-182. 488
Denisov, A. M. (1999), Elements of theory of inverse problems, Utre ht, The Nether-489
lands: VSP. 490
Devaney, A. J., and G. C. Sherman (1982), Nonuniqueness in inverse sour e and 491
s attering problems, IEEE Trans. Antennas Propagat., 33, 1034-1037. 492
Dubey,A.C.(1995),Dete tionte hnologyforminesandmineliketargets,Eds. Orlando, 493
FL. 494
Groets h, C. W. (1993),Inverse Problems in Mathemati al S ien es,Wiesbaden, Ger-495
many: Vieweg. 496
Hoole, S. R. H. (1991),Inverse problemmethodologyand nite elementsinthe identi-497
ations of ra ks, sour es, materials,and their geometry inina essible lo ations,IEEE 498
Trans. Magn., 27, 3433-3443. 499
Isernia, T., V. Pas azio, and R. Pierri (2001), On the lo al minima problem in a 500
tomographi imagingte hnique, IEEE Trans. Geos i. Remote Sensing, 39, 1596-1607. 501
Jones, D. S. (1964),The Theory of Ele tromagnetism,Oxford,U.K.: PergamonPress. 502
Kleinman, R. E., and P. M. Van den Berg(1992), A modiedgradient method for 503
two-dimensionalproblems in tomography, J. Comput. Appl. Math., 42,17-35. 504
Liu, Q. H., Z. Q. Zhang, T. T. Wang, J. A. Bryan, G. A. Ybarra, L. W. 505
Nolte, and W. T. Joines (2003),A tive mi rowaveimaging 1-2D forward and inverse 506
s attering methods, IEEE Trans. Mi rowave Theory Te h., 50, 123-133. 507
Louis, K. (1992), Medi al imaging: state of the art and future development, Inverse 508
Problems,8, 709-738. 509
Marklein, R., K. Balasubramanian, A. Quing,and K. J. Lagenberg(2001), Lin-510
ear andnonlinear iteratives alarinversionof multi-frequen ymulit-bistati experimental 511
ele tromagneti s attering data, Inverse Problems, 17,1565-1702. 512
Massa, A.(2002),Geneti algorithmbased te hniques for2Dmi rowaveinverse s atter-513
Pandalai,Transworld Resear h Network Press, Trivandrum, India. 515
Miller, E. L., and A. S. Willsky (1996a), A multis ale, statisti ally based inversion 516
s heme for linearized inverse s attering problems, IEEE Trans. Geos i. Remote Sensing, 517
34, 346-357. 518
Miller, E.L., andA.S.Willsky(1996b),Wavelet-basedmethodsfornonlinearinverse 519
s attering problem using the extended Born approximation, Radio S i., 31,51-65. 520
Natterer, F. (1986), Numeri al treatment of ill-posed problems, in Inverse Problems, 521
Ed. G.Talenti, Le ture Notes in Mathemati s, p. 1225. 522
Ri hmond, J. H. (1965), S attering by a diele tri ylinder of arbitrary ross se tion 523
shape, IEEE Trans. Antennas Propagat., 13, 334-341. 524
Steinberg, B. D. (1991), Mi rowave imaging te hniques, New York: Wiley. 525
Testorf,M., and M.Fiddy (2001),Imagingfromreals atteredelddatausingalinear 526
spe tral estimation te hniques, Inverse Problems, 17,1565-1702. 527
Vogel, C. R. (2002), Computational Methods for Inverse Problems, Philadelphia, PA: 528
SIAM. 529
•
Figure 1. Geometry of the problem.532
•
Figure 2. Numeri al Experiments: (a) o- entered homogeneous ir ularylin-533
der (Real dataset Marseille [Belkebir and Saillard, 2001℄ - dielTM_de 8f.exp),
534
(b)twohomogeneous ir ular ylinders(RealdatasetMarseille [Belkebirand
Sail-535
lard,2001℄-twodielTM_8f.exp),and( )U-shapedmetalli ylinder(Realdataset
536
Marseille [Belkebir and Saillard,2001℄ - uTM_shaped.exp).
537
•
Figure 3. Comparisonsbetween the in ident eld measured inD
M
and the values538
synthesized by means of the PW-Model ((a) amplitude and (b) phase), and
CCW-539
Model (( ) amplitudeand (d) phase).
540
•
Figure 4. Re onstru tions of an o- entered homogeneous ir ular ylinder (Real541
dataset Marseille [Belkebir and Saillard, 2001℄ - dielTM_de 8f.exp) a hieved
542
at the onvergen e step of the inversion pro edure by modeling the radiated eld
543
through (a)the single PW-Model and (b) the single CCW-Model.
544
•
Figure 5. Fitting between omputed and measured values of the radiated eld545
in the observation domain versus various numbers of sour e oe ients,
S
,and for546
dierentsour e models.
547
•
Figure 6. Behaviorofweighting sour e oe ientsasafun tionofthe indexw
for548
(a) the PW-Model (
S = 20
) and for(b) the CCW-Model (S = 11
).549
•
Figure7. Plotsoftheradiatedelds(V = 1
) omputedbymeansofthePW-Model550
(
S = 20
) (amplitude (a) and phase (b) distributions), the CCW-Model (S = 11
)551
(amplitude ( ) and phase (d) distributions), and the single CCW-Model (
S = 1
)552
(amplitude (e) and phase (f) distributions).
553
•
Figure 8. Re onstru tions of an o- entered homogeneous ir ular ylinder (Real554
dataset Marseille [Belkebir and Saillard, 2001℄ - dielTM_de 8f.exp) a hieved
555
at the onvergen e step of the inversion pro edure by modeling the radiated eld
556
through (a)the PW-Model (
S = 20
) and (b) the CCW-Model (S = 11
).•
Figure 9. Normalizedbehavior of the singularvalues of[G]
for (a) the PW-Model558
(
S = 20
) and for (b)the CCW-Model (S = 11
).559
•
Figure 10. Radiated-eld modeling: DCW-Model (S = 15
). (a) Behavior of560
weighting sour e oe ientsasa fun tionof the index
w
. Comparison between the561
in identeld measuredin
D
M
andthe numeri ally- omputedvalues ((b) amplitude562
and ( ) phase).
563
•
Figure 11. Plots of the radiated eld (V = 1
) omputed by means of theDCW-564
Model (
S = 15
) (amplitude(e)and phase(f) distributions).565
•
Figure 12. Re onstru tion of ano- entered homogeneous ir ular ylinder(Real566
dataset Marseille [Belkebir and Saillard, 2001℄ - dielTM_de 8f.exp)a hieved at
567
(a) i=1, (b) i=2and ( ) at the onvergen e step (i=3) of the inversion pro edure
568
by modelingthe radiated eld through the DCW-Model (
S = 15
).569
•
Figure 13. Re onstru tions of two homogeneous ir ular ylinders (Real dataset570
Marseille [Belkebir andSaillard,2001℄-twodielTM_8f.exp)a hievedatthe
on-571
vergen e step ofthe inversionpro edurebymodeling theradiatedeld through (a)
572
the single PW-Model, (b) the single CCW-Model and the DCW-Model (
S = 15
)573
[( ) real part and (d)imaginary part℄.
574
•
Figure 14. Re onstru tion of an U-shaped metalli ylinder (Real datasetMar-575
seille [Belkebir and Saillard, 2001℄ - uTM_shaped.exp) a hieved at the
onver-576
gen e step of the inversion pro edure by modeling the radiated eld through the
577
DCW-Model (
S = 15
).•
Table I. Re onstru tion of an o- entered homogeneous ir ular ylinder (Real 580datasetMarseille [Belkebir and Saillard,2001℄ -dielTM_de 8f.exp)-Estimated 581
geometri al parameters. 582
•
Table II. Re onstru tion of an o- entered homogeneous ir ular ylinder (Real 583dataset Marseille [Belkebir and Saillard, 2001℄ - dielTM_de 8f.exp)- Error g-584
ures. 585
•
TableIII.Re onstru tionofano- enteredhomogeneous ir ular ylinder(SNR =
58620 dB
) for dierent illuminationsand onsidering variousele tromagneti sour es -587Quantitativeerror gures [(a)
ξ
tot
, (b)ξ
int
and ( )ξ
ext
℄.588
•
Table IV. Re onstru tion of two homogeneous ir ular ylinders (Real dataset 589Marseille [Belkebirand Saillard,2001℄-twodielTM_8f.exp)-Estimated geomet-590 ri al parameters (
d
(I
opt
)
=
rn
x
(1)
c
(Iopt)
− x
(2)
c
(Iopt)
o
2
+
n
y
(1)
c
(Iopt)
− y
c
(2)
(Iopt)
o
2
). 591 592594 595 596
(x , y )
m
m
D
M
D
I
φ
v
θ
inc
x
y
Observation Domain
Investigation Domain
597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612Figure 1 - D. Fran es hini et al., On the Ee ts of the Ele tromagneti ...
(x , y )
m
m
r
m
r
s
x
y
30 mm
θ
l
(a)(x , y )
m
m
r
m
r
s
x
y
θ
l
90 mm
(b)(x , y )
m
m
r
m
r
s
x
y
θ
l
40 mm
50 mm
80 mm
( ) 614 615 616 617Figure 2 - D. Fran es hini et al., On the Ee ts of the Ele tromagneti ...
0
0.2
0.4
0.6
0.8
1
1.2
1.4
0
5
10
15
20
25
30
35
40
45
50
|E
v
inc
(x
m
,y
m
)|,v=1
m
PW-Model (S=1)
PW-Model (S=20)
Actual Values
(a)-4
-3
-2
-1
0
1
2
3
4
0
5
10
15
20
25
30
35
40
45
50
Phase[E
v
inc
(x
m
,y
m
)],v=1
m
PW-Model (S=1)
PW-Model (S=20)
Actual Values
(b) 619 620 621 622 623 624Figure 3(I) - D. Fran es hiniet al., On the Ee ts of the Ele tromagneti ...
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
0
5
10
15
20
25
30
35
40
45
50
|E
v
inc
(x
m
,y
m
)|,v=1
m
CCW-Model (S=1)
CCW-Model (S=11)
Actual Values
( )-4
-3
-2
-1
0
1
2
3
4
0
5
10
15
20
25
30
35
40
45
50
Phase[E
v
inc
(x
m
,y
m
)],v=1
m
CCW-Model (S=1)
CCW-Model (S=11)
Actual Values
(d) 626 627 628 629 630 631Figure 3(II) - D. Fran es hiniet al., On the Ee ts of the Ele tromagneti ...
x
y
2.6
Re {τ (x, y)}
0.0
(a)x
y
2.6
Re {τ (x, y)}
0.0
(b) 634 635 636 637 638Figure 4 - D. Fran es hini et al., On the Ee ts of the Ele tromagneti ...
641 642 643 644
0
0.005
0.010
0.015
0.020
0.025
0.030
0.035
0.040
0
5
10
15
20
25
µ
S
DCW-Model
PW-Model
CCW-Model
645 646 647 648 649 650 651 652 653 654 655 656 657 658 659Figure 5 - D. Fran es hini et al., On the Ee ts of the Ele tromagneti ...