UNIVERSITY
OF TRENTO
DIPARTIMENTO DI INGEGNERIA E SCIENZA DELL’INFORMAZIONE
38123 Povo – Trento (Italy), Via Sommarive 14
http://www.disi.unitn.it
A MULTI-RESOLUTION TECHNIQUE BASED ON SHAPE
OPTIMIZATION FOR THE RECONSTRUCTION OF
HOMOGENEOUS DIELECTRIC OBJECTS
M. Benedetti, D. Lesselier, M. Lambert, and A. Massa
January 2009
Shape Optimization for the Re onstru tion
of Homogeneous Diele tri Obje ts
M. Benedetti
1,2
, D. Lesselier2
, M. Lambert2
, and A. Massa1
1
Departmentof InformationEngineeringandComputerS ien e,ELEDIAResear h Group,UniversityofTrento, ViaSommarive14,38050Trento-Italy,Tel. +390461 882057,Fax. +390461882093
2
DépartementdeRe her heenÉle tromagnétisme-LaboratoiredesSignauxet Systèmes,CNRS-SUPELEC Univ. ParisSud11,3rueJoliot-Curie,91192 Gif-sur-YvetteCEDEX,Fran e
E-mail: manuel.benedettidisi.unitn.it,lesselierlss.supele .fr,lambertlss.supele .fr, andrea.massaing.unitn.it
Abstra t. Intheframeworkofinverses atteringte hniques,thispaperpresentsthe
integration of a multi-resolution te hnique and the level-set method for qualitative mi rowaveimaging. Onone hand,in order to ee tively exploit thelimited amount of information olle table from s attering measurements, the iterative multi-s aling approa h (IMSA) is employed for enabling a detailed re onstru tion only where neededwithoutin reasingthenumberofunknowns. Ontheother hand,thea-priori information on the homogeneity of the unknown obje t is exploited by adopting a shape-based optimization and representing the support of the s atterer via a level set fun tion. Reliability and ee tiveness of the proposed strategy are assessed by pro essing both syntheti and experimental s attering data for simple and omplex geometries,aswell.
Key Words - Mi rowave Imaging, Inverse S attering, Level Sets, Iterative Multi-S alingApproa h,Homogeneous Diele tri S atterers.
1. Introdu tion
The non-invasive re onstru tion of position and shape of unknown targets is a topi of great interest in many appli ations, su h as non-destru tive evaluation and testing (NDE/NDT) for industrial monitoring and subsurfa e sensing [1℄. In this framework, many methodologies have been proposed based on x-rays [2℄, ultrasoni s [3℄, and eddy urrents [4℄. Furthermore, mi rowave imaging has been re ognized as a suitable methodology sin e [1℄[5℄: (a) ele tromagneti elds at mi rowave frequen ies an penetrate non-ideal ondu tor materials; (b) the eld s attered by the target is representative of itsinner stru ture and not only ofits boundary; ( )mi rowavesshow ahighsensibilitytothewater ontentofthe stru tureundertest; (d)mi rowavesensors anbeemployed withoutme hani al onta tswiththespe imen. Inaddition, ompared to x-ray and magneti resonan e, mi rowave-based approa hes minimize (or avoid) ollateralee ts in the spe imen undertest. Therefore, they an be safely employed in biomedi al imaging.
A further advan e in mi rowave non-invasive inspe tion is represented by inverse s atteringapproa hesaimedatre onstru tinga ompleteimageoftheregionundertest. Unfortunately,theunderlyingmathemati almodelis hara terizedbyseveraldrawba ks preventing their massiveemployment inNDE/NDT appli ations. Inparti ular, inverse s attering problems are intrinsi allyill-posed [6℄as well asnon-linear [7℄.
Sin e the ill-posednessis strongly relatedto the amountof olle table information andusually thenumberofindependentdataislowerthan thedimension ofthe solution spa e, multi-view/multi-illumination systems are generally adopted. However, it is well known that the olle table information is an upper-bounded quantity [8℄-[10℄. Consequently, it is ne essary to ee tively exploit the overall information ontained in the s attered eld samples fora hieving a satisfa tory re onstru tion.
Towardsthisend,multi-resolutionstrategieshavebeenre entlyproposed. Theidea isthat ofusinganenhan edspatial resolutiononlyinthoseregionswherethe unknown s atterersarefoundtobelo ated. A ordingly,Milleretal. [11℄proposeda statisti ally-based method for determining the optimal resolution level, while Baussard et al. [12℄ developed astrategybased onsplinepyramidsforsub-surfa e imagingproblems. Asfor anexample on ernedwith qualitativemi rowaveimaging,Liet al. [13℄ implementeda multis alete hniquebasedonLinearSamplingMethod(LSM)toee tivelyre onstru t the ontourofthes atterers. Unlike[11℄-[13℄,the iterativemulti-s aleapproa h(IMSA)
developed by Caorsietal. [14℄performsamulti-step,multi-resolutioninversionpro ess inwhi htheratiobetween unknowns anddataiskeptsuitablylowand onstantatea h step of theinversion pro edure, thusredu ing the riskof o urren e of lo alminima[9℄ in the arisingoptimization problem.
On the other hand, the la k of informationae ting the inverse problemhas been addressed through the exploitation of the a-priori knowledge (when available) on the s enario under test by means of anee tive representation of the unknowns. As far as many NDE/NDT appli ations are on erned, the unknown defe t is hara terized by known ele tromagneti properties (i.e., diele tri permittivity and ondu tivity) and it lieswithinaknown hostregion. Undertheseassumptions,the imagingproblemredu es toashapeoptimizationproblemaimedatthesear h oflo ationandboundary ontours ofthedefe t. Parametri te hniquesaimedatrepresentingtheunknown obje tinterms ofdes riptiveparametersofreferen eshapes[15℄[16℄andmoresophisti atedapproa hes su h as evolutionary- ontrolled spline urves [17℄[18℄, shape gradients [19℄-[21℄ or level-sets [22℄-[30℄havethenbeen proposed. Asfaraslevel-set-based methodsare on erned, the homogeneousobje tisdened asthe zero levelof a ontinuousfun tion and,unlike pixel-based or parametri -based strategies, su h a des ription enables one to represent omplex shapes ina simple way.
Inordertoexploitboth theavailablea-priori knowledgeonthes enarioundertest (e.g.,the homogeneity ofthe s atterer)and theinformation ontentfromthe s attering measurements,thispaperproposestheintegrationoftheiterativemulti-s alingstrategy (IMSA) [14℄ and the level-set (LS)representation [23℄.
The paper is stru tured as follows. The integration between IMSA and LS is detailed in Se t. 2 dealing with a two-dimensional geometry. In Se tion 3, numeri al testing and experimentalvalidationare presented, a omparison with the standard LS implementationbeing made. Finally,some on lusions are drawn (Se t. 4).
2. Mathemati al Formulation
Let us onsider a ylindri al homogeneous non-magneti obje t with relative permittivity
ǫ
C
and ondu tivityσ
C
that o upies a regionΥ
belonging to an investigationdomainD
I
. Su h a s atterer is probed by a set ofV
transverse-magneti (T M
)planewaves, withele tri elddire tedalongtheaxisofthe ylindri algeometry,namely
ζ
v
(r) = ζ
v
(r)ˆz
(
v = 1, . . . , V
),r = (x, y)
. The s attered eld,ξ
v
(r) = ξ
v
(r)ˆz
,is olle ted atM(v)
,v = 1, ..., V
, measurement pointsr
m
distributed in the observation domainD
O
.In order toele tromagneti ally des ribe the investigationdomain
D
I
,let usdene the ontrast fun tionτ (r)
given byτ (r) =
τ
C
0
r ∈ Υ
otherwise (1) whereτ
C
= (ǫ
C
− 1) − j
σ
C
2πf ε
0
,
f
being the frequen y of operation(the time dependen ee
j2πf t
being implied).
The s attering problem is des ribed by the well-known Lippmann-S hwinger integral equations
ξ
v
(r
m
) =
2π
λ
2
Z
D
I
τ (r
′
) E
v
(r
′
) G
2D
(r
m
, r
′
) dr
′
,
r
m
∈ D
O
(2)ζ
v
(r) = E
v
(r) −
2π
λ
2
Z
D
I
τ (r
′
) E
v
(r
′
) G
2D
(r, r
′
) dr
′
, r
∈ D
I
(3)where
λ
is the ba kground wavelength,E
v
is the total ele tri eld, and
G
2D
(r, r
′
) =
−
j
4
H
(2)
0
2π
λ
kr − r
′
k
isthefree-spa e two-dimensionalGreen'sfun tion,
H
(2)
0
being the se ond-kind, zeroth-order Hankel fun tion.In order to retrieve the unknown position and shape of the target
Υ
by step-by-stepenhan ingthespatialresolutiononlyinthatregion, alledregion-of-interest(RoI
),R ∈ D
I
, where the s atterer is lo ated [14℄, the following iterative pro edure ofS
max
steps is arriedout.
With referen e to Fig. 1(a) and to the blo k diagram displayed in Fig. 2, at the rst step (
s = 1
,s
being the step number) a trial shapeΥ
s
= Υ
1
, belonging toD
I
, is hosen and the region of interestR
s
[R
s=1
= D
I
℄ is partitioned intoN
IM SA
equal square sub-domains,whereN
IM SA
depends onthe degrees offreedomofthe problemat hand and it is omputed a ording to the guidelines suggested in [9℄.In addition, the level set fun tion
φ
s
is initialized by means of a signed distan e fun tion dened as follows [23℄[25℄:φ
s
(r) =
−
minb=1,...,B
s
kr − r
b
k
ifτ (r) = τ
C
minb=1,...,B
s
kr − r
b
k
ifτ (r) = 0
(4)where
r
b
= (x
b
, y
b
)
is theb
-thborder- ell (b = 1, . . . , B
s
) ofΥ
s=1
.•
Problem Unknown Representation - The ontrast fun tion is represented interms of the levelset fun tionas follows
e
τ
k
s
(r) =
s
X
i=1
N
IM SA
X
n
i
=1
τ
k
i
B
r
n
i
r
∈ D
I
(5)where the index
k
s
indi ates thek
-th iteration at thes
-th step [k
s
= 1, ..., k
opt
s
℄,B
r
n
i
isare tangularbasisfun tionwhosesupportisthe
n
-thsub-domainatthei
-th resolution level [n
i
= 1, ..., N
IM SA
,i = 1, ..., s
℄, and the oe ientτ
k
i
is given byτ
k
i
=
τ
C
ifΨ
k
i
r
n
i
≤ 0
0
otherwise (6) lettingΨ
k
i
r
n
i
=
φ
k
i
r
n
i
ifi = s
φ
k
opt
i
r
n
i
if(i < s)
andr
n
i
∈ R
i
(7) withi = 1, ..., s
asin (5).•
Field Distribution Updating - On eτ
e
k
s
(r)
has been estimated, the ele tri eldE
v
k
s
(r)
is numeri ally omputed a ording toa point-mat hing version of the Methodof Moments (MoM)[31℄ ase
E
v
k
i
r
n
i
=
P
N
IM SA
p
i
=1
ζ
v
r
p
i
h
1 −
τ
e
k
i
r
p
i
G
2D
r
n
i
, r
p
i
i
−1
,
r
n
i
, r
p
i
∈ D
I
n
i
= 1, ..., N
IM SA
.
(8)•
Cost Fun tion Evaluation - Starting from the total ele tri eld distribution(8), the re onstru ted s attered eld
ξ
e
v
k
s
(r
m
)
at them
-th measurement point,m = 1, ..., M(v)
, isupdated by solving the following equatione
ξ
v
k
s
(r
m
) =
s
X
i=1
N
IM SA
X
n
i
=1
e
τ
k
i
r
n
i
e
E
v
k
i
r
n
i
G
2D
r
m
, r
n
i
(9)and the t between measured and re onstru ted data is evaluated by the multi-resolution ost fun tion
Θ
dened asΘ {φ
k
s
} =
P
V
v=1
P
M(v)
m=1
e
ξ
v
k
s
(r
m
) − ξ
v
k
s
(r
m
)
2
P
V
v=1
P
M(v)
m=1
ξ
v
k
s
(r
m
)
2
.
(10)•
MinimizationStopping -Theiterativepro essstops[i.e.,k
opt
s
= k
s
andτ
e
opt
s
=
τ
e
k
s
℄ when: (a)aset of onditionsonthestabilityof there onstru tionholds trueor(b) when the maximum number of iterations is rea hed [k
s
= K
max
℄ or ( ) when thevalueofthe ostfun tionissmallerthanaxedthreshold
γ
th
. Asfarasthestability of the re onstru tion is on erned [ ondition (a)℄,the rst orresponding stopping riterion is satised when, for a xed number of iterations,K
τ
, the maximum number of pixelswhi h varytheir value is smallerthan auser dened thresholdγ
τ
a ording to the relationshipmax
j=1,...,K
τ
N
IM SA
X
n
s
=1
|
τ
e
k
s
(r
n
s
) −
τ
e
k
s
−j
(r
n
s
)|
τ
C
< γ
τ
· N
IM SA
.
(11)These ond riterion, aboutthestabilityofthere onstru tion,issatisedwhenthe ost fun tion be omes stationarywithin a windowof
K
Θ
iterationsas follows:1
K
Θ
K
Θ
X
j=1
Θ {φ
k
s
} − Θ {φ
k
s
−j
}
Θ {φ
k
s
}
< γ
Θ
.
(12)K
Θ
beingaxed numberofiterationsandγ
Θ
beinguser-denedthresholds;. Whenthe iterative pro ess stops, the solution
τ
e
opt
s
at thes
-th step is sele ted as the one represented by the best level set fun tionφ
opt
s
dened asφ
opt
s
=
argh
minh=1,...,k
opt
s
(Θ {φ
h
})
i
.
(13)•
Iteration Update - The iterationindex is updated [k
s
→ k
s
+ 1
℄;•
Level Set Update - Thelevelsetis updateda ordingtothe followingHamilton-Ja obi relationship
φ
k
s
(r
n
s
) = φ
k
s
−1
(r
n
s
) − ∆t
s
V
k
s
−1
(r
n
s
) H {φ
k
s
−1
(r
n
s
)}
(14)where
H {·}
isthe Hamiltonianoperator[32℄[33℄ given asH
2
{φ
k
s
(r
n
s
)} =
max2
n
D
x−
k
s
; 0
o
+
min2
n
D
x+
k
s
; 0
o
+
+
max2
n
D
y−
k
s
; 0
o
+
min2
n
D
y+
k
s
; 0
o
ifV
k(s)
r
n(s)
≥ 0
min2
n
D
x−
k
s
; 0
o
+
max2
n
D
x+
k
s
; 0
o
+
+
min2
n
D
y−
k
s
; 0
o
+
max2
n
D
y+
k
s
; 0
o
otherwise (15) andD
x±
k
s
=
±φ
ks
(x
ns±1
,y
ns
)∓φ
ks
(x
ns
,y
ns
)
l
s
,D
y±
k
s
=
±φ
ks
(x
ns
,y
ns±1
)∓φ
ks
(x
ns
,y
ns
)
l
s
.∆t
s
is the time-step hosen as∆t
s
= ∆t
1
l
s
l
1
with
∆t
1
to be set heuristi allya ording to the literature [23℄,l
s
being the ell-side atthes
-thresolution level.V
k
fun tion omputedfollowingthe guidelines suggested in[23℄ by solving theadjoint problemof (8) inorder todetermine the adjointeld
F
v
k
s
. A ordingly,V
k
s
(r
n
s
) = −ℜ
( P
V
v=1
τ
C
E
v
ks
(
r
ns
)
F
ks
v
(
r
ns
)
P
V
v=1
P
M(v)
m=1
|
ξ
v
ks
(
r
m
)|
2
)
,
n
s
= 1, ..., N
IM SA
(16)where
ℜ
standsfor the real part.When the
s
-th minimization pro ess terminates, the ontrast fun tion is updated [τ
e
opt
s
(r)=
τ
e
k
s−1
(r)
,r ∈ D
I
(5)℄ aswell astheRoI
[R
s
→ R
s−1
℄. Todo so, the following operations are arriedout:•
Computation of the Bary enter of the RoI - the enter ofR
s
of oordinates(
x
e
c
s
,
y
e
s
c
)
is determinedby omputing the enter of mass of the re onstru ted shapes as follows [14℄ [Fig. 1(b)℄e
x
c
s
=
R
D
I
x
τ
e
opt
s
(r) B (r) dx dy
R
D
I
τ
e
opt
s
(r) B (r) dx dy
(17)e
y
c
s
=
R
D
I
y
τ
e
opt
s
(r) B (r) dx dy
R
D
I
τ
e
opt
s
(r) B (r) dx dy
;
(18)•
Estimation of the Size of the RoI -thesideL
s
ofR
s
is omputedbyevaluatingthe maximum of the distan e
δ
c
(r) =
q
(x −
x
e
c
s
)
2
+ (y −
y
e
c
s
)
2
in order to en lose the s atterer, namelye
L
s
=
maxr
(
2 ×
τ
e
opt
s
(r)
τ
C
δ
c
(r)
)
.
(19)On e the
RoI
has been identied, the level of resolution is enhan ed [k
s
→ k
s−1
℄ only in this region by dis retizingR
s
intoN
IM SA
sub-domains [Fig. 1( )℄ and by repeating the minimization pro ess untilthe syntheti zoombe omes stationary(s = s
opt
), i.e.,(
|Q
s−1
− Q
s
|
|Q
s−1
|
× 100
)
< γ
Q
,
Q =
x
e
c
,
y
e
c
,
L
e
(20)γ
Q
being a threshold set as in [14℄, oruntil a maximum number of steps (s
opt
= S
max
)is rea hed.
Atthe end ofthe multi-steppro ess (
s = s
opt
),the problemsolutionis obtainedase
τ
opt
r
n
i
=
τ
e
opt
s
r
n
i
,n
i
= 1, ..., N
IM SA
,i = 1, ..., s
opt
.3. Numeri al Validation
In order to assess the ee tiveness of the IMSA-LS approa h, a sele ted set of representativeresults on ernedwithbothsyntheti and experimentaldataispresented herein. Theperforman esa hievedareevaluatedbymeansofthefollowingerrorgures:
•
Lo alization Errorδ
δ|
p
=
r
e
x
c
s
|
p
− x
c
|
p
2
−
y
e
c
s
|
p
− y
c
|
p
2
λ
× 100
(21) wherer
c
|
p
=
x
c
|
p
, y
c
|
p
is the enter of the
p
-th true s atterer,p = 1, ..., P
,P
being the number of obje ts. The average lo alizationerror< δ >
is dened as< δ >=
1
P
P
X
p=1
δ|
p
.
(22)•
Area EstimationError∆
∆ =
I
X
i=1
1
N
IM SA
N
IM SA
X
n
i
=1
N
n
i
× 100
(23) whereN
n
i
isequal to1
ifτ
e
opt
r
n
i
= τ
r
n
i
and0
otherwise.As far as the numeri al experiments are on erned, the re onstru tions have been performedbyblurringthe s atteringdatawithanadditiveGaussiannoise hara terized by a signal-to-noise-ratio(
SNR
)SNR = 10
logP
V
v=1
P
M
(v)
m=1
|ξ
v
(r
m
)|
2
P
V
v=1
P
M(v)
m=1
|µ
v,m
|
2
(24)µ
v,m
being a omplex Gaussian randomvariable with zero mean value.
3.1. Syntheti Data - Cir ular Cylinder
3.1.1. Preliminary Validation In the rst experiment, a lossless ir ular o- entered
s atterer of known permittivity
ǫ
C
= 1.8
and radiusρ = λ/4
is lo ated in a square investigationdomain of sideL
D
= λ
[23℄.V = 10 T M
plane waves are impingingfrom the dire tionsθ
v
= 2π (v − 1)/ V
,v = 1, ..., V
, and the s attering measurements are olle ted atM = 10
re eivers uniformlydistributed ona ir leof radiusρ
O
= λ
.As faras the initializationofthe IMSA-LS algorithmis on erned,the initialtrial obje t
Υ
1
is a disk with radiusλ/4
and entered inD
I
. The initial value of the time step is set to∆t
1
= 10
−2
stopping riteria,the following ongurationof parametershas been sele ted a ording to a preliminary alibration dealing with simple known s atterers and noiseless data:
S
max
= 4
(maximum number of steps),γ
e
x
c
= γ
e
y
c
= 0.01
andγ
e
L
= 0.05
(multi-step pro ess thresholds),K
max
= 500
(maximum number of optimization iterations),γ
Θ
= 0.2
andγ
τ
= 0.02
(optimization thresholds),K
Θ
= K
τ
= 0.15 K
max
(stabilityounters), and
γ
th
= 10
−5
(threshold on the ost fun tion).
Figure 3 shows samples of re onstru tions with the IMSA-LS. At the rst step [Fig. 3(a) -
s = 1
℄, the s atterer is orre tly lo ated, but its shape is only roughly estimated. Thankstothemulti-resolutionrepresentation,thequalitativeimagingofthe s attererisimproved inthenext step[Fig. 3(b)-s = s
opt
= 2
℄as onrmedbytheerror indexesinTab. 1. For omparisonpurposes,theproleretrievedbythesingle-resolution method [23℄ (indi ated in the following as Bare-LS), whenD
I
has been dis retized inN
Bare
= 31 × 31
equalsub-domains, isshown [Fig. 3( )℄. In general,the dis retizationof the Bare-LS has been hosen in order to a hieve in the whole investigation domain a re onstru tion with the same level of spatial resolution obtained by the IMSA-LS in the
RoI
ats = s
opt
.Although the nal re onstru tions [Figs. 3(b)( )℄ a hieved by the two approa hes aresimilarandquite losetothetrues atterersampledatthespatialresolutionof Bare-LS [Fig. 3(d)℄ and IMSA-LS [Fig. 3(b)℄, the IMSA-LS more faithfully retrieves the symmetryofthea tualobje t,eventhoughthere onstru tionerrorappearstobelarger thanthe one of the Bare-LS (Fig. 4). During the iterativepro edure, the ost fun tion
Θ
opt
= Θ {φ
opt
s
}
isinitially hara terizedbyamonotoni allyde reasingbehavior. Then,Θ
opt
⌋
IM SA
be omesstationaryuntilthe stopping riteriondened by relationships(11)and (12) is satised (Fig. 4 -
s = 1
). Then, after the update of the eld distribution indu ing the error spike whens = s
opt
= 2
andk
s
= 1
,Θ
opt
⌋
IM SA
settles to a value of6.28 × 10
−4
whi hisoftheorder oftheBare-LS error(
Θ
opt
⌋
Bare
= 1.42 × 10
−4
). Su ha slightdieren ebetween
Θ
opt
⌋
IM SA
andΘ
opt
⌋
Bare
dependsonthedierentdis retization [i.e., the basis fun tionsB
r
n(i=2)
,n(i) = 1, ..., N
IM SA
are not the same as those ofBare-LS℄, but it does not ae t the re onstru tion in terms of both lo alization and area estimation, sin e
δ⌋
IM SA−LS
< δ⌋
Bare−LS
and∆⌋
IM SA−LS
< ∆⌋
Bare−LS
(Tab. 1). Fig. 4 alsoshows that the multi-step multi-resolutionstrategy is hara terized by a lower omputationalburden be ause of the smallernumber of iterationsfor rea hing the onvergen e (Fig. 4 -k
tot
⌋
IM SA
= 125
vs.k
tot
⌋
Bare
= 177
, beingk
tot
the total numberofiterationsdenedask
tot
=
P
s
opt
redu ed number of oating-pointoperations. As a matterof fa t,sin e the omplexity of the LS-based algorithms is of the order of
O (2 × η
3
)
,
η = N
IM SA
, N
Bare
(i.e., the solutionof twodire t problems is ne essary for omputing anestimate of the s attered eld and for updating the velo ity ve tor), the omputational ost of the IMSA-LS at ea hiteration istwoorders in magnitudesmaller than that of the Bare-LS.3.1.2. Noisy Data As for the stability of the proposed approa h, Figure 5 shows
the re onstru tions with the IMSA-LS [Figs. 5(a)( )(e)℄ ompared to those of the Bare-LS [Figs. 5(b)(d)(f)℄ with dierent levels of additive noise on the s attered data [
SNR = 20 dB
(top);SNR = 10 dB
(middle);SNR = 5 dB
(bottom)℄. As expe ted, when theSNR
de reases, the performan es worsen. However, as outlined by the behavior of the error gures in Tab. 2, blurred data and/or noisy onditions ae t moreevidentlythe Bare implementationthan themulti-resolutionapproa h. For ompleteness, the behavior ofΘ
opt
⌋
IM SA
versus the iteration index is reported in Fig. 6 for dierent levels ofSNR
. As it an be noti ed, the value of the error atthe end of the iterative pro edure de reases as theSNR
in reases.In the se ond experiment, the same ir ular s atterer, but entered at a dierent position within a larger investigation square of side
L
D
= 2λ
(ρ
O
= 2λ
), has been re onstru ted. A ording to [9℄,M = 20
;v = 1, ..., V
re eivers andV = 20
views are onsidered andD
I
isdis retized inN
IM SA
= 13 × 13
pixels.Figure 7(a) shows the re onstru tion obtained at the onvergen e (
s
opt
= 3
) by IMSA-LS whenSNR = 5 dB
. The result rea hed by the Bare-LS (N
BARE
= 47 × 47
) is reported in Fig. 7(b) as well. As it an be noti ed, the multi-resolution inversion is hara terizedby abetterestimationof the obje t enter and shapeas onrmed by the values ofδ
and∆
(δ⌋
IM SA−LS
= 0.59
vs.δ⌋
Bare−LS
= 2.72
and∆⌋
IM SA−LS
= 0.48
vs.∆⌋
Bare−LS
= 0.64
). As for the omputationalload,the same on lusions frompreviousexperiments hold true.
3.2. Syntheti Data - Re tangular S atterer
These ondtest asedealswith amore omplexs attering onguration. Are tangular o- entered s atterer (
L = 0.27λ
andW = 0.13λ
) hara terized by a diele tri permittivityǫ
C
= 1.8
islo atedwithinaninvestigationdomainofL
D
= 3λ
asindi ated bythe reddashedlineinFig. 8. Insu ha ase,theimagingsetupismadeup ofV = 30
sour es andM = 30
measurement points for ea h viewv
[9℄.D
I
is partitioned intoN
IM SA
= 19 × 19
sub-domains (whileN
Bare
= 33 × 33
)and∆t
1
isset to0.06
.3.2.1. Validation of the Stopping Criteria Before dis ussing the re onstru tion
apabilities,let usshowa result on erned with the behaviorof the proposedapproa h when varying the user-dened thresholds (
γ
Θ
,γ
τ
,γ
e
x
c
,γ
e
y
c
,γ
e
L
)of the stopping riteria. Figure8 displays the re onstru tions a hieved by using the sets of parameters given in Tab. 3[Γ
1
- Fig. 8(a);Γ
2
-Fig. 8(b);Γ
3
-Fig. 8( );Γ
4
-Fig. 8(d)℄while thebehaviors of the ost fun tion are depi ted in Fig. 9. As it an be noti ed, the total number of iterationsk
tot
in reases as the values of the thresholdsγ
Θ
andγ
τ
de rease. However, in spite of a largerk
tot
, using lowerthreshold values does not providebetter results,as shown by the omparison between settingsΓ
2
andΓ
4
[Figs. 8(b)-(d),and Fig. 9℄. The sets of parameters hara terized byγ
Θ
= 0.2
andγ
τ
= 0.02
provide a good trade-o between the arising omputational burden and the quality of the re onstru tions. As far as the stopping riterion of the multi-resolution pro edure is on erned, Figure 9 also shows two dierent behaviors of the ost fun tion when usingΓ
2
andΓ
3
(lettingγ
Θ
= 0.2
andγ
τ
= 0.02
). In parti ular, the proposed approa h stops ats
opt
= 3
,insteadof
s
opt
= 4
,whenin reasing by adegreeofmagnitudethe values ofγ
e
x
c
,γ
e
y
c
,andγ
L
e
. Although with a heavier omputational burden, the hoi eγ
e
x
c
= γ
e
y
c
= 0.01
andγ
L
e
= 0.05
results more ee tive [see Fig. 8(b)vs. Fig. 8( )℄.3.2.2. Noisy Data Figures 10-12 and Table 4 show the results from the omparative
study arriedoutin orresponden ewithdierentvaluesofsignal-to-noiseratio[
SNR =
20 dB
- Fig. 10(a) vs. Fig. 10(b);SNR = 10 dB
- Fig. 10( ) vs. Fig. 10(d);SNR =
5 dB
- Fig. 10(e) vs. Fig. 10(f)℄. They further onrm the reliability and e ien yof the multi-resolution strategy in terms of qualitative re onstru tion errors (Fig. 11), espe iallywhen the noise levelgrows. In parti ular,the Bare implementationdoes not yield either the position or the shape of the re tangular s atterer when
SNR = 5 dB
, whereas the IMSA-LS properly retrieves both the bary enter and the ontour of the target. As forthe omputational ost, itshouldbenoti edthatalthoughtheIMSA-LS requires a greater number of iterationsfor rea hing the onvergen e (Fig. 12,Tab. 4), the total amountof omplex oating-pointoperations,f
pos
= O (2 × η
3
) × k
tot
, usually results smaller(Tab. 4).3.3. Numeri alData - Hollow Cylinder
The third test ase is on erned with the inversion of the data s attered by a higher permittivity (
ǫ
C
= 2.5
) o- entered ylindri al ring, lettingL
D
= 3λ
. The external radius of the ring isρ
ext
=
2
3
λ
, and the internal one isρ
int
=
λ
3
. By assuming the same arrangement of emitters and re eivers as inSe tion 3.2, the investigation domain isdis retized withN
IM SA
= 19 × 19
andN
Bare
= 35 × 35
square ells forthe IMSA-LS and the Bare-LS,respe tively. Moreover,∆t
1
is initializedto0.003
.As it an be observed from Fig. 13,where the proles when
SNR = 20 dB
[Figs. 13(a)(b)℄andSNR = 10 dB
[Figs. 13( )(d)℄ re onstru ted by means of the IMSA-LS [Figs. 13(a)( )℄ and the Bare-LS [Figs. 13(b)(d)℄ are shown, the integrated strategy usually over omes the standard one both in lo ating the obje t and in estimating the shape. In parti ular, whenSNR = 20 dB
, the distribution in Fig. 13(a) is a faithful estimate of the s atterer under test (δ⌋
IM SA−LS
= 1.25
and∆⌋
IM SA−LS
= 3.13
). On the ontrary, the re onstru tion with the Bare-LS is very poor (δ⌋
Bare−LS
= 65.2
and∆⌋
Bare−LS
= 34.39
). Certainly, a smallerSNR
value impairs the inversion as shownin Fig. 13( ) [ ompared to Fig. 13(a)℄. However, in this ase, the IMSA-LS is able to properly lo ate the obje t (
δ⌋
IM SA−LS
= 1.7
vs.δ⌋
Bare−LS
= 65.9
) giving rough but useful indi ations about its shape (∆⌋
IM SA−LS
= 7.6
vs.∆⌋
Bare−LS
= 34.55
).3.4. Syntheti Data - Multiple S atterers
The lastsyntheti test ase is aimedatillustratingthe behavior of the IMSA-LS when dealingwith
P = 3
s atterers(ǫ
C
= 2.0
)distan edfromone another. Thetestgeometry is hara terizedby anellipti o- entered ylinder,a ir ularo- entereds atterer, and a square o- entered obje t lo ated in a square investigation domain hara terized byL
D
= 3λ
. By adopting the same arrangement of emitters and re eivers as in Se tion3.3, the investigationdomainisdis retized with
N
IM SA
= 23 × 23
andN
Bare
= 31 × 31
square ells for the IMSA-LS and the Bare-LS, respe tively. Moreover,∆t
1
is set to0.03
.Figures 14 and 15 show the results from the omparative study arried out in orresponden e with dierent values of signal-to-noise ratio. As shown by the re onstru tions (Fig. 14) andas expe ted,the multi-resolutionapproa hprovidesmore a urate results and appears to be more reliablethan the Bare-LS espe ially with low
(Fig. 15), for whi h the IMSA-LS a hievesa lowerlo alizationerror as well asa lower area error than those of Bare-LS, espe ially for
SNR = 5 dB
. On the other hand, both algorithmsprovide goodestimates of the s atterer undertest when inverting data ae ted by lownoise [SNR = 20 dB
- Fig. 14(a)vs. Fig. 14(b);Fig. 15(a)and (b)℄.3.5. Laboratory-Controlled Data
In order to further assess the ee tiveness of the IMSA-LS also in dealing with experimental data, the multiple-frequen y angular-diversity bi-stati ben hmark provided by Institut Fresnel in Marseille (Fran e) has been onsidered. Withreferen e to the experimental setup des ribed in [34℄, the dataset dielTM_de 8f.exp has been pro essed. The eld samples [
M = 49
,V = 36
℄ are related to an o- entered homogeneous ir ular ylinderρ = 15
mm in diameter, hara terized by a nominal value of the obje t fun tion equal toτ (r) = 2.0 ± 0.3
, and lo ated atx
c
= 0.0
,y
c
= −30
mm within an investigation domain assumed in the following of squaregeometry and extension
20 × 20
m2
.
By setting
ǫ
C
= 3.0
, the re onstru tions a hieved are shown in Fig. 16 (left olumn) ompared to those from the standard LS (right olumn) atF = 4
dierent operation frequen ies. Whatever the frequen y, the unknown s atterer isa urately lo alizedand both algorithmsyield,at onvergen e, stru turesthat o upy alarge subsetof the true obje t. Su hasimilarityofperforman es,usuallyveriedinsyntheti experimentswhen the value ofSNR
is greaterthan20 dB
,seems to onrmthe hypothesisof alow-noise environment as already eviden ed in[35℄.Finally, also in dealing with experimental datasets, the IMSA-LS proves its e ien y sin e the overall amount of omplex oating point operations still remains two orders inmagnitude lowerthan the one of the Bare-LS (Tab. 5 -Fig. 17).
4. Con lusions
In this paper, a multi-resolution approa h for qualitative imaging purposes based on shape optimization has been presented. The proposed approa h integrates the multi-s ale strategy and the level set representation of the problem unknowns in order to protablyexploittheamountofinformation olle tablefromthes atteringexperiments as wellas the available a-priori information onthe s atterer under test.
•
innovative multi-level representation of the problem unknowns in theshape-deformation-based re onstru tion te hnique;
•
ee tive exploitation of the s attering data through the iterative multi-stepstrategy;
•
limitationof theriskof beingtrappedinfalsesolutionsthankstotheredu ed ratiobetween data and unknowns;
•
useful exploitationofthe a-priori information(i.e.,obje thomogeneity) about thes enariounder test;
•
enhan ed spatial resolution limitedto the region of interest.From the validation on erned with dierent s enarios and both syntheti and experimental data, the following on lusions an be drawn:
•
the IMSA-LS usually proved more ee tivethan thesingle-resolutionimplementa-tion, espe ially whendealing with orrupted data s attered fromsimple as wellas omplex geometries hara terized by one or several obje ts;
•
the integratedstrategy appearedless omputationally-expensivethanthe standardapproa h in rea hing a re onstru tion with the same level of spatial resolution within the support of the obje t.
Referen es
[1℄ P.J.Shull,Nondestru tive Evaluation: Theory,Te hniques andAppli ations.CRCPress,2002.
[2℄ J. Baru hel, J.-Y. Buère, E. Maire, P. Merle, and G. Peix, X-Ray Tomography in Material S ien e.HermesS ien e,2000.
[3℄ L. W. S hmerr, Fundamentals of Ultrasoni Nondestru tive Evaluation: A Modeling Approa h. Springer,1998.
[4℄ B.A. Auldand J.C. Moulder,Review ofadvan es in quantitativeeddy urrentnondestru tive evaluation, J.Nondestru tive Evaluation,vol.18,no.1,pp.3-36,Mar.1999.
[5℄ R. Zoughi, Mi rowave Nondestru tive Testing and Evaluation. Dordre ht, The Netherlands: KluwerA ademi Publishers,2000.
[6℄ O. M. Bu i and T. Isernia, Ele tromagneti inverse s attering: retrievable information and measurementstrategies, Radio S i.,vol.32, pp.2123-2138,Nov.-De .1997.
[7℄ M. Berteroand P. Bo a i, Introdu tion to Inverse Problems in Imaging. IOP Publishing Ltd, Bristol,1998.
[8℄ O. M.Bu i and G. Fran es hetti, On the spatial bandwidth of s attered elds, IEEE Trans. AntennasPropagat.,vol.35,no.12,pp.1445-1455,De .1987.
[9℄ T.Isernia,V.Pas azio,andR.Pierri,Onthelo alminimainatomographi imagingte hnique, IEEETrans.AntennasPropagat.,vol.39,no.7,pp.1696-1607,Jul.2001.
[10℄ H. Haddar, S. Kusiak, and J. Sylvester, The onvexba k-s attering support, SIAM J. Appl. Math.,vol.66,pp.591-615,De .2005.
[11℄ E. L. Miller,and A. S. Willsky, Amultis ale, statisti ally basedinversions hemefor linearized inverses atteringproblems, IEEETrans.AntennasPropagat.,vol.34,no.2,pp.346-357,Mar. 1996.
[12℄ A.Baussard,E.L.Miller,andD.Lesselier,Adaptivemulis alere onstru tionofburiedobje ts, InverseProblems, vol.20,pp.S1-S15,De .2004.
[13℄ J. Li, H. Liu, and J. Zou, Multilevel linear sampling method for inverse s attering problems, SIAMJ.Appl. Math.,vol.30,pp.1228-1250,Mar.2008.
[14℄ S. Caorsi,M.Donelli, andA. Massa,"Dete tion,lo ation, andimagingof multiples atterersby meansoftheiterativemultis alingmethod,"IEEETrans.Mi rowaveTheoryTe h.,vol.52,no. 4,pp.1217-1228,Apr.2004.
[15℄ S. Caorsi,A. Massa, M.Pastorino,and M.Donelli, Improvedmi rowaveimagingpro edure for nondestru tiveevaluationsoftwo-dimensionalstru tures, IEEETrans.AntennasPropag.,vol. 52,pp.1386-1397,Jun. 2004.
[16℄ M.Benedetti, M.Donelli, andA. Massa,Multi ra kdete tionin two-dimensionalstru turesby meansof GA-based strategies, IEEE Trans. Antennas Propagat., vol. 55, no. 1,pp. 205-215, Jan.2007.
[17℄ V. Cingoski, N. Kowata, K. Kaneda, and H. Yamashita, Inverse shape optimization using dynami ally adjustable geneti algorithms, IEEE Trans. Energy Conversion, vol. 14, no. 3, pp.661-666,Sept.1999.
[19℄ J. Cea, S. Garreau, P. Guillame, and M. Masmoudi, The shape and topologi al optimizations onne tion, Comput.Methods Appl.Me h. Eng.,vol.188,no.4,pp.713-726,2000.
[20℄ G.R.Feijoo,A.A.Oberai,andP.M.Pinsky,Anappli ationofshapeoptimizationinthesolution ofinversea ousti s atteringproblems, Inverse Problems,vol.20,pp.199-228,Feb.2004.
[21℄ M.Masmoudi,J.Pommier,andB.Samet,Thetopologi alasymptoti expansionfortheMaxwell equationsandsomeappli ations, InverseProblems, vol.21,pp.547-564,Apr.2005.
[22℄ F.Santosa,Alevel-set approa hforinverseproblems involvingobsta les, ESAIM: COCV, vol. 1,pp.17-33,Jan.1996.
[23℄ A.Litman,D.Lesselier,andF.Santosa,"Re onstru tionofatwo-dimensionalbinaryobsta leby ontrolledevolutionofalevel-set,"Inverse Problems,vol.14,pp.685-706,Jun. 1998.
[24℄ O. Dorn, E. L. Miller, C. M. Rappaport, A shape re onstru tion method for ele tromagneti tomographyusingadjoint eldsand levelsets, Inverse Problems, vol. 16,pp.1119-1156,May 2000.
[25℄ C.Ramananjaona, M. Lambert, D. Lesselier,and J.P. Zolésio, "Shapere onstru tion ofburied obsta les by ontrolled evolution of a level-set: from a min-max formulation to a numeri al experimentation,"Inverse Problems,vol.17, pp.1087-1111,De .2001.
[26℄ R. Ferrayé, J.-Y. Dauvigna , and C. Pi hot, An inverse s attering method based on ontour deformationsbymeansofalevelsetmethodusingfrequen yhoppingte hnique, IEEETrans. AntennasPropagat.,vol.51,no.5,pp.1100-1113,May2003.
[27℄ E. T. Chung, T. F. Chan, X. C. Tai, Ele tri al impedan e tomography using level set representationand totalvariationalregularization, J. Comput.Phys.,vol.205,no.2,pp. 707-723,May.2005.
[28℄ K.vandenDoelandU.M.Asher,Dynami levelsetregularizationforlargedistributedparameter estimationproblems, Inverse Problems,vol.23, pp.1271-1288,Jun. 2007.
[29℄ J.Strain, Three methods formoving interfa es, J. Comput. Phys., vol.151, pp.616-648,May 1999.
[30℄ O.Dorn and D. Lesselier, Level set methods for inverse s attering, Inverse Problems, vol. 22, pp.R67-R131,Aug.2006.
[31℄ J.H.Ri hmond,S atteringbyadiele tri ylinderofarbitrary ross-se tionshape,IEEETrans. AntennasPropagat.,vol.13,pp.334-341,May1965.
[32℄ S.OsherandJ.A.Sethian,Frontspropagatingwith urvature-dependentspeed: algorithmsbased onHamilton-Ja obiformulations, J. Comput.Phys.,79,pp.12-49,Nov.1988.
[33℄ J. A. Sethian, Level Set and Fast Mar hing Methods. Cambridge University Press, Cambridge, UK,2nded.,1999.
[34℄ K. Belkebir and M. Saillard, Testing inversion algorithms against experimental data, Inverse Problems,vol.17, pp.1565-1702,De .2001.
[35℄ M.TestorfandM.Fiddy,Imagingfromreals atteredelddatausingalinearspe tralestimation te hnique, Inverse Problems,vol.17,pp.1645-1658,De .2001.
a
( )
b
( )
c
( )
Actual Scatterer
Υ
1
(e
x
c
s=1
, e
y
s=1
c
)
r
n
s=1
r
n
s=1
R
s=2
φ
k
opt
s=1
φ
k
s=1
s = 1
,
k = 1
s = 1
,
k = k
opt
s = 2
,
k = 1
D
I
DI
D
I
e
L
s=1
r
n
s=2
φ
k
s=2
Figure 1. Graphi alrepresentationofthe IMSA-LS zoomingpro edure. (a)Firststep (
k = 1
): the investigationdomainisdis retizedinN
sub-domainsand a oarse solution is looked for. (b) First step (k = k
opt
): the region of interest that ontains the rst estimateof the obje tis dened. ( )Se ond step (
k = 1
): anenhan ed resolution level is used onlyinside the region of interest.Field Distribution Updating
Determine
Cost Function Evaluation
Level Set Update
Problem Unknown Representation
FALSE
TRUE
TRUE
Stop
Initialization
FALSE
IMSA
Stopping Criterion
Stopping Criteria
LS
Compute e
E
v
k
i
r
n
i
r
n
i
∈ D
I
e
x
c
s
,e
y
c
s
,e
L
s
Determine e
ξ
v
k
s
r
m
v
and compute
Θ {φ
k
s
}
r
m
v
∈ D
O
Compute
φ
k
s
from
φ
k
s
−1
e
τ
k
s
(r) =
P
s
i=1
P
N
I M SA
n
i
=1
τ
k
i
B r
n
i
r
n
i
∈ D
I
s-th resolution level
k
s
= k
s
+ 1
Υ
s
opt
Υ
s
= Υ
1
k
s
= 0
s = s + 1
γ
x
e
c
,γ
e
y
c
,γ
e
L
γ
τ
, γ
Θ
, γ
th
, K
max
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
R
(2)
1
τ (x, y)
0
1
τ (x, y)
0
(a) (b)x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
1
τ (x, y)
0
1
τ (x, y)
0
( ) (d)Figure 3. Numeri al Data. Cir ular ylinder (
ǫ
C
= 1.8
,L
D
= λ
, Noiseless Case). Re onstru tions with IMSA-LS at (a)s = 1
and (b)s = s
opt
= 2
, ( ) Bare-LS. Optimalinversion (d).10
-4
10
-3
10
-2
10
-1
10
0
0
50
100
150
200
Θ
Iteration, k
s=1
s=2
IMSA (
Θ
opt
)
IMSA (
Θ
)
BARE
Figure 4. Numeri al Data. Cir ular ylinder (
ǫ
C
= 1.8
,L
D
= λ
, Noiseless Case). Behavior of the ost fun tion.R
(2)
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
1
τ (x, y)
0
1
τ (x, y)
0
(a) (b)(3)
R
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
(2)
R
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
1
τ (x, y)
0
1
τ (x, y)
0
( ) (d)(3)
R
R
(2)
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
1
τ (x, y)
0
1
τ (x, y)
0
(e) (f)Figure 5. Cir ular ylinder (
ǫ
C
= 1.8
,L
D
= λ
, Noisy Case). Re onstru tions with IMSA-LS (left olumn) and Bare-LS (right olumn) for dierent values ofSNR
[SNR = 20 dB
(top),SNR = 10 dB
(middle),SNR = 5 dB
(bottom)℄.10
-3
10
-2
10
-1
10
0
10
1
0
20
40
60
80
100
Θ
opt
Iteration, k
SNR = 20 dB
SNR = 10 dB
SNR = 5 dB
s = 1
s = 2
s = 1
s = 2
s = 2
s = 1
Figure 6. Numeri alData. Cir ular ylinder (
ǫ
C
= 1.8
,L
D
= λ
). Behaviorof the ost fun tion versus the noise level.(2)
R
(3)
R
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
1
τ (x, y)
0
(a)x
y
λ
λ
−0.5
0.5
0.5
−0.5
1.0
−1.0
−1.0
1.0
1
τ (x, y)
0
(b)Figure 7. Numeri al Data. Cir ular ylinder (
ǫ
C
= 1.8
,L
D
= 2λ
,SNR = 5 dB
). Re onstru tions with (a) IMSA-LS and (b)Bare-LS.x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
(a) (b)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
( ) (d)Figure 8. Numeri alData. Re tangular ylinder(
ǫ
C
= 1.8
,L
D
= 3λ
,NoiselessCase). Re onstru tions with IMSA-LS for the dierent settings of Tab. 3[(a)Γ
1
, (b)Γ
2
, ( )10
-4
10
-3
10
-2
10
-1
10
0
10
1
0
200
400
600
800
1000
1200
1400
Θ
opt
Iteration, k
Γ
3
Γ
2
Γ
4
Γ
1
Γ
1
Γ
2
,
Γ
3
Γ
4
Figure 9. Numeri alData. Re tangular ylinder(
ǫ
C
= 1.8
,L
D
= 3λ
,NoiselessCase). Behavior of the ost fun tion of IMSA-LS forthe dierent settingsof Tab. 3.x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
(a) (b)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
( ) (d)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
(e) (f)Figure 10. Numeri al Data. Re tangular ylinder (
ǫ
C
= 1.8
,L
D
= 3λ
, Noisy Case). Re onstru tions with IMSA-LS (left olumn)and Bare-LS (right olumn)for dierent values ofSNR
[SNR = 20 dB
(top),SNR = 10 dB
(middle),SNR = 5 dB
(bottom)℄.10
0
10
1
10
2
10
3
5
10
15
20
δ
SNR [dB]
IMSA
BARE
(a)2.5
2.0
1.5
1.0
0.5
0.0
5
10
15
20
∆
SNR [dB]
IMSA
BARE
(b)Figure 11. Numeri al Data. Re tangular ylinder (
ǫ
C
= 1.8
,L
D
= 3λ
, Noisy Case). Values of the error guresversusSNR
.10
-3
10
-2
10
-1
10
0
10
1
0
200
400
600
800
1000
1200
1400
Θ
opt
Iteration, k
IMSA
BARE
(a)10
-2
10
-1
10
0
10
1
0
100
200
300
400
500
600
Θ
opt
Iteration, k
IMSA
BARE
(b)10
-1
10
0
10
1
0
100
200
300
400
500
600
Θ
opt
Iteration, k
IMSA
BARE
( )Figure 12. Numeri al Data. Re tangular ylinder (
ǫ
C
= 1.8
,L
D
= 3λ
, Noisy Case). Behavior of the ost fun tion versus the iteration index when (a)SNR = 20 dB
, (b)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
2
τ (x, y)
0
2
τ (x, y)
0
(a) (b)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
2
τ (x, y)
0
2
τ (x, y)
0
( ) (d)Figure 13. Numeri al Data. Hollow ylinder (
ǫ
C
= 2.5
,L
D
= 3λ
, Noisy Case). Re onstru tions with IMSA-LS (left olumn)and Bare-LS (right olumn)for dierent values ofSNR
[SNR = 20 dB
(top),SNR = 10 dB
(bottom)℄.x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
R
(3)
R
(3)
R
(3)
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
(a) (b)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
R
(3)
R
(3)
R
(3)
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
( ) (d)x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
R
(3)
R
(3)
R
(3)
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
1
τ (x, y)
0
1
τ (x, y)
0
(e) (f)Figure 14. Numeri al Data. Multiple s atterers (
ǫ
C
= 2.0
,L
D
= 3λ
, Noisy Case). Re onstru tions with IMSA-LS (left olumn)and Bare-LS (right olumn)for dierent values ofSNR
[SNR = 20 dB
(top),SNR = 10 dB
(middle),SNR = 5 dB
(bottom)℄.2.0
2.5
3.0
3.5
4.0
5
10
15
20
δ
SNR [dB]
IMSA
BARE
(a)1.0
2.0
3.0
4.0
5.0
6.0
5
10
15
20
∆
SNR [dB]
IMSA
BARE
(b)Figure 15. Multiple s atterers (
ǫ
C
= 2.0
,L
D
= 3λ
, Noisy Case). Values of the error gures versusSNR
.x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
x
y
λ
λ
−0.25
0.25
0.25
−0.25
0.50
−0.50
−0.50
0.50
3
τ (x, y)
0
3
τ (x, y)
0
(a) (b)x
y
λ
λ
−0.5
0.5
0.5
−0.5
1.0
−1.0
−1.0
1.0
x
y
λ
λ
−0.5
0.5
0.5
−0.5
1.0
−1.0
−1.0
1.0
3
τ (x, y)
0
3
τ (x, y)
0
( ) (d)Figure 16(I). Experimental Data (Dataset Marseille [34℄). Cir ular ylinder (dielTM_de 8f.exp). Re onstru tionswithIMSA-LS(left olumn)andBare-LS (right olumn)at dierent frequen ies
f
[f = 1 GHz
(a)(b);f = 2 GHz
( )(d)℄.x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
x
y
λ
λ
−0.75
0.75
0.75
−0.75
1.50
−1.50
−1.50
1.50
3
τ (x, y)
0
3
τ (x, y)
0
(e) (f)x
y
λ
λ
−1.0
1.0
1.0
−1.0
2.0
−2.0
−2.0
2.0
x
y
λ
λ
−1.0
1.0
1.0
−1.0
2.0
−2.0
−2.0
2.0
3
τ (x, y)
0
3
τ (x, y)
0
(g) (h)Figure 16(II). Experimental Data (Dataset Marseille [34℄). Cir ular ylinder (dielTM_de 8f.exp). Re onstru tionswithIMSA-LS(left olumn)andBare-LS (right olumn)at dierent frequen ies
f
[f = 3 GHz
(e)(f);f = 4 GHz
(g)(h)℄.M. Benedetti et al.
10
-2
10
-1
10
0
10
1
10
2
0
200
400
600
800
1000
Θ
opt
Iteration, k
IMSA
BARE
10
-3
10
-2
10
-1
10
0
10
1
10
2
0
200
400
600
800
1000
Θ
opt
Iteration, k
IMSA
BARE
(a) (b)10
-3
10
-2
10
-1
10
0
10
1
10
2
0
200
400
600
800
1000
Θ
opt
Iteration, k
IMSA
BARE
10
-3
10
-2
10
-1
10
0
10
1
10
2
0
200
400
600
800
1000
Θ
opt
Iteration, k
IMSA
BARE
( ) (d) 17. Exp erimen tal Data (Dataset Marseille [34 ℄). Cir ular ylinder 8f.exp ). Beha vior of the ost fun tion v ersus the n um b er of iterations ( a )f
=
1
G
H
z
, ( b )f
=
2
G
H
z
, ( )f
=
3
G
H
Z
, and ( d )f
=
4
G
H
z
.IMSA − LS
Bare − LS
s = 1
s = 2
δ
6.58 × 10
−6
2.19 × 10
−6
5.21 × 10
−1
∆
2.36
0.48
0.64
M.
Benedetti
et
al.
SNR = 20 dB
SNR = 10 dB
SNR = 5 dB
IMSA − LS
Bare − LS
IMSA − LS
Bare − LS
IMSA − LS
Bare − LS
δ
5.91 × 10
−1
2.72
2.28
2.45
6.78 × 10
−1
1.63
∆
0.98
1.28
1.07
1.80
1.50
2.07
2. Numeri al Data. Cir ular ylinder (ǫ
C
=
1.
8
, Noisy Case ). V alues of the indexes for dieren t v alues ofS
N
R
.Set of Parameters
γ
Θ
γ
τ
γ
x
e
c
, γ
e
y
c
γ
e
L
Γ
1
0.5
0.05
0.01
0.05
Γ
2
0.2
0.02
0.01
0.05
Γ
3
0.2
0.02
0.1
0.5
Γ
4
0.02 0.002
0.01
0.05
Table 3. Numeri al Data. Re tangular ylinder (
ǫ
C
= 1.8
,L
D
= 3λ
, Noiseless Case). Dierent settings forthe parametersof the stopping riteria.M.
Benedetti
et
al.