Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO.

5, MAY 2019 1753

Compressive Sensing-Based Born Iterative


Method for Tomographic Imaging
Giacomo Oliveri , Senior Member, IEEE, Lorenzo Poli, Member, IEEE, Nicola Anselmi, Member, IEEE,
Marco Salucci , Member, IEEE, and Andrea Massa , Fellow, IEEE

Abstract— The problem of imaging nonweak scatterers in the applicative scenarios, including biomedical diagnosis [6], [7],
compressive sensing (CS) framework, without recurring to the ground-penetrating radar prospecting [8], [9], antenna
contrast-source formulation, is addressed by means of an inno- characterization [10], and radar imaging [11]. Such a success,
vative hybrid born iterative CS method (BI-CSM). Toward this
end, the born iterative (BI) formulation of the nonlinear inverse which is not limited to electromagnetics but also extends to
scattering equations is adopted and a customized version of the other fields such as information theory [12]–[14] and signal
CS-based inversion is implemented to retrieve the target contrast processing [15] is motivated by several concurring factors,
while estimating the electric field distribution without recurring including the availability of well-established theoretical
to time-expensive full-wave simulations. The selected numerical foundations [1], [3], [14], the robustness, the effectiveness,
and experimental results, drawn from a wide validation, are
presented to illustrate the procedural steps of the BI-CSM and and the flexibility of the arising sparseness-promoting
to assess, also comparatively, the features, the potentialities, and regularization techniques [1], [3], and the availability of
the limitations of the proposed inversion technique. several efficient and customizable SW implementations [16].
Index Terms— Born iterative (BI) method, compressive Nevertheless, the formulation of inverse problems in the
sensing (CS), inverse scattering, microwave imaging. standard CS framework still presents several theoretical and
practical challenges, especially when microwave imaging
is at hand [3]. Indeed, unlike other regularization-driven
I. I NTRODUCTION approaches [17]–[19], an effective application of CS requires
that: 1) the unknown (e.g., the target contrast or the equivalent
T HE study and the development of innovative strategies
based on the compressive sensing (CS) paradigm
are currently among the most active research areas in the
source) is sparsely encoded over a basis function (sparse-
ness condition); 2) the sampling and the sensing bases are
inverse scattering community [1]–[5]. This is proven by incoherent (incoherence condition) [3], [13], [14]; and 3) the
the number of CS-based tools for dealing with different relation between the problem data and the corresponding
unknowns is modeled as a linear function (linearity condition).
Manuscript received October 21, 2018; revised December 28, 2018; A commonly used approach to fit this latter requirement
accepted January 25, 2019. Date of publication March 11, 2019; date of
current version May 6, 2019. This work benefited from the networking is to rely on formulations based on approximations, such
activities carried out within the SNATCH Project (2017–2019) funded by the as the Born (BA) [20]–[23] or the Rytov [24] ones, but
Italian Ministry of Foreign Affairs and International Cooperation, Directorate this choice turns out to be accurate only when handling
General for Cultural and Economic Promotion and Innovation, the Project
“WATERTECH - Smart Community per lo Sviluppo e l’Applicazione di weak scatterers [20]. Alternatively, “contrast-source” formu-
Tecnologie di Monitoraggio Innovative per le Reti di Distribuzione Idrica negli lations (CSFs) have been adopted [25]. However, since an
usi idropotabili ed agricoli” (Grant no. SCN_00489) funded by the Italian unknown CSF is the equivalent source (i.e., the product
Ministry of Education, University, and Research within the Program “Smart
cities and communities and Social Innovation” (CUP: E44G14000060008), between the contrast function and the total field [25]), choos-
and the Project “Antenne al Plasma - Tecnologia abilitante per SAT- ing a suitable sensing basis to satisfy the sparseness condition
COM (ASI.EPT.COM)” funded by the Italian Space Agency (ASI) under is not trivial—unless point targets are at hand [25]—because
Grant 2018-3-HH.0 (CUP: F91I17000020005). (Corresponding author:
Andrea Massa.) of its dependence on the probing field and the direction of
G. Oliveri is with the ELEDIA Research Center (ELEDIA@UniTN - incidence of the impinging wave besides on the scatterer
University of Trento), 38123 Trento, Italy, and also with the ELEDIA Research contrast. Therefore, imaging nonweak targets with CS-based
Center (ELEDIA@L2S - UMR 8506), 91192 Gif-sur-Yvette, France (e-mail:
giacomo.oliveri@unitn.it; giacomo.oliveri@l2s.centralesupelec.fr). reconstruction algorithms are still an open challenge and no
L. Poli, N. Anselmi, and M. Salucci are with the ELEDIA Research general-purpose solutions have been proposed yet [3].
Center (ELEDIA@UniTN - University of Trento), 38123 Trento, Italy (e-mail: In this paper, a new hybrid born iterative CS method
lorenzo.poli@unitn.it; nicola.anselmi.1@unitn.it; marco.salucci@unitn.it).
A. Massa is with the ELEDIA Research Center (ELEDIA@UniTN - (BI-CSM) is proposed to address current issues in effec-
University of Trento), 38123 Trento, Italy, with the ELEDIA Research tively applying CS to handle nonweak targets within the
Center (ELEDIA@L2S - UMR 8506), 91192 Gif-sur-Yvette, France, contrast-field formulation (CFF) of the inverse scattering prob-
and also with the ELEDIA Research Center (ELEDIA@UESTC),
University of Electronic Science and Technology of China, School of Elec- lem. Toward this end, the nonlinear scattering equations are
tronic Engineering, Chengdu 611731, China (e-mail: andrea.massa@unitn.it; formulated with the BI approach [26] and, at each step of the
andrea.massa@l2s.centralesupelec.fr; andrea.massa@uestc.edu.cn). iterative process, they are solved with a CS-based technique by
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. leveraging on the sparsity of the target under test with respect
Digital Object Identifier 10.1109/TMTT.2019.2899848 to the chosen representation basis. The exploitation of such
0018-9480 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
1754 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO. 5, MAY 2019

an approach for the data inversion is mainly motivated by the is the object function that depends on the electromagnetic
following: 1) unlike BA approaches [21]–[23], the proposed properties (i.e., the relative permittivity, ε(r), and the electric
technique allows one to retrieve nonweak scatterers by means conductivity, σ (r), distributions) of the dielectric scatterers
of a CS-based method and within the CFF of the nonlinear embedded in the homogeneous lossless (σ0 = 0.0 S/m)
inverse scattering equations, which are recast in a series of lin- background medium with permittivity ε0 . Generally speaking,
ear ones; and 2) unlike contrast-source imaging strategies [25], the objective of an imaging process is the retrieval of τ (r)
the BI-CSM relies on the choice of the sensing basis in the and Tv (r), v = 1, . . . , V , within the investigation domain (i.e.,
target domain that only depends on the scatterer features, but r ∈ ) starting from the knowledge of Iv (r) everywhere (i.e.,
not on the illumination type and its position and/or angle of r ∈ ∪) and of Ev (r) in the observation domain (i.e., r ∈ )
incidence. for all views (v = 1, . . . , V ).
In short, the main innovative contributions of this paper To numerically deal with the scattering equations, the point
comprise as follows: matching version of the moment method [27], [28] is applied
1) for the first time to the best of the our knowledge, by sampling (1) in M measurement points, rm ∈ ,
the exploitation of a BI formulation for the definition m = 1, . . . , M, and (2) in P matching points, r p ∈ ,
of a CS-based inversion approach not requiring cumber- p = 1, . . . , P, to yield
some full-wave simulations; 
k2
2) the development of a CS retrieval technique, within the Ev (rm ) = j τ (r )Tv (r )H(k|rm − r |)dr (4)
4 
CFF, which is suitable for nonweak objects and that 
k2
exploits sensing bases which are independent on the Iv (r p ) = Tv (r p ) − j τ (r )Tv (r ) × H(k|r p − r |)dr
probing sources; 4 
3) the derivation of a set of operative guidelines for a (5)
reliable use of the proposed inversion strategy. v = 1, . . . , V . By choosing [3], [22], [23] a suitable sens-
This paper is organized as follows. Starting from the statement ing basis, ψ  {ψn (r); n = 1, . . . , N}, so that τ (r) can
of the imaging problem at hand, the BI-CSM inversion is be described by S (S  N) nonzero coefficients, {τ (s);
detailed in Section II. A selected set of numerical and exper- s = 1, . . . , S}, in the expansion
imental results, drawn from a wide validation, are presented
to illustrate the procedural steps of the BI-CSM and to assess, 
N

also comparatively, its features, potentialities, and limitations τ (r) = τ (n) ψ (n) (r) (6)
n=1
(see Section III). Some final remarks (see Section IV) and
conclusions (see Section V) are eventually drawn. and by expressing the total field as

L
II. S TATEMENT OF THE I MAGING P ROBLEM Tv (r) = Tv(l) γ (l) (r) (7)
AND BI-CSM S OLUTION M ETHOD l=1
Let us consider a time-harmonic tomographic γ  {γ (l) (r);l = 1, . . . , L} being the set of L (L = P)
transverse-magnetic inverse scattering problem, where rectangular basis functions [27] whose lth entry is equal to
an investigation domain  is probed by a set of V known γ (l) (r) = 1 if r ∈ (l) and γ (l) (r) = 0 otherwise, and (l) |l= p
electromagnetic sources working at the angular frequency ω1 is the square support of γ (l) centered in rl (∪ Pp=1 (l) |l= p =
and radiating incident z-polarized waves of amplitudes Iv (r), ), it turns out that the vth (v = 1, . . . , V ) discretized versions
v = 1, . . . , V . The resulting vth (v = 1, . . . , V ) “total,” of (4) and (5) are [27]
Tv (r), and “scattered,” Ev (r) [Ev (r)  Tv (r) − Iv (r)], fields

k 2  (n)  (l)
N L
comply with the so-called Data [27]

m
Ev (r ) = j τ Tv ψ (n) (r )
k2 4 
τ (r )Tv (r )H(k|r − r |)dr , r ∈  (1)
n=1 l=1
Ev (r) = j
4  ×γ (l) (r )H(k|rm − r |)dr (8)
k 2  (l)  (n)
and State [27] L N
 Iv (r p ) = Tv( p) − j Tv τ
k2 4
Iv (r) = Tv (r) − j τ (r )Tv (r )  l=1 n=1
4 
×H(k|r − r |)dr , r ∈  (2) × ψ (n) (r )γ (l) (r )H(k|r p − r |)dr . (9)

equations, where  is the observation domain ( ∩  = ∅), In the case of a weak scatterer (i.e., the product between the
H(·) is the 0th order Hankel function of the second kind, k is diameter and the refractive index of the object being below
the free-space wavenumber (k  (2π/λ)), and 0.35λ [20]), these equations can be considerably simplified
σ (r) under the first-order BA that implies the condition Tv (r) ≈
τ (r)  ε(r) − 1 − j (3) Iv (r) when r ∈  [20], [21]. More specifically, (9) turns out
ωε0
to be an identity, while (8) becomes a linear function of the
1 For notation simplicity, the time dependence exp(− jωt) is assumed and unknown sparse vector τ  {τ (n) ; n = 1, . . . , N}. Thanks to
omitted hereinafter. the sparseness of the scatterer representation and the linearity
OLIVERI et al.: CS-BASED BI METHOD FOR TOMOGRAPHIC IMAGING 1755

of the relation between data and unknowns, the inversion can by reformulating the scattering-data inversion in a Bayesian
be then efficiently performed with CS-based methods [21]. sense [1], [3], thus computing the profile distribution as
Of course, intolerable distortions arise whether using such
1  H
V
an approximation for higher contrasts/wider scatterers. In this −1 
τ q =
 Gv q Gv q + di ag(
α q ) × GvH q Ev
latter case, the inherent nonlinearity of (8) and (9) needs to be V
v=1
addressed in a more suitable and complex way. Toward this (14)
end, an innovative integration of the BI approach [26] and
of sparsity-promoting solution strategies [1], [3] is described where di ag(·) and · H denote the “diagonal” and “conjugate
hereinafter. The main motivations of such a choice can be transpose” operations, respectively. Moreover
summarized as follows: 1) BI formulations are known to be
Ev  {Ev (rm ); m = 1, . . . , M} (15)
stable and also sufficiently accurate when imaging targets
exhibiting contrast values beyond the range of validity of Gv q is the qth approximation of the discretized Green oper-
the BA [26], as well; 2) BI techniques address the nonlinear ator given by
inverse scattering problem, described by (8) and (9), as the
solution of a set of (much simpler) linear ones to find a 
k 2  (l) 
L
sequence of approximations of the unknown object function G v q  j Tv q−1
4
τ (n) q ; n = 1, . . . , N}
τ q = {
 (10)  l=1

× ψ (n) (r )γ (l) (r )H(k|rm − r |)dr ,


and of the total field coefficients 


Tv q = {Tv(l) q ; l = 1, . . . , L} (11)
m = 1, . . . , M, n = 1, . . . , N (16)
v = 1, . . . , V , q (q = 1, . . . , Q) being the iteration index [26];
and 3) by exploiting some a priori information to represent in while the shared hyperparameter vector  α q is found by
a sparse fashion the class of scatterers under test and thanks maximizing the likelihood function [1], [3]
to the linearity of the qth (q = 1, . . . , Q) scattering equations,
1
V
efficient linear solvers [1], [3] can be profitably used.

α q = arg max − [log |Uv q |
The BI-CSM is a two-task iterative inversion that works α 2
v=1
as follows. At each qth (q = 1, . . . , Q) iteration, the first  
task is devoted to estimate the contrast distribution,  τ q , +(2N + 2β1 ) log EvH Uv −1
q Ev + 2β2 ) (17)
by fitting (8) or, in an equivalent fashion, by solving the by means of an efficient local search technique such as the
following minimization problem relevance vector machine (RVM) [1], [3]. In (17), Uv q is the
qth iteration auxiliary matrix
1 
V
τ q =
 arg min
V
v=1
τ Uv q  ι + Gv q diag (α)−1 GvH (18)
⎧ 
q
⎨M 
N
k 2  (l)
L
β1 and β2 are user-defined scale priors and ι the identity
(n)
(Ev (r ) −
m
τ j Tv
⎩ 4 matrix.
m=1 n=1 l=1 q−1

Once the qth approximate contrast fitting (8) has been

(n)  (l)   
found from (14), the second task of the BI-CSM at the qth
× ψ (r ) × γ (r )H(k|r − r |)dr
m
(12) (q = 1, . . . , Q) iteration is the solution of (9) to predict the

estimated field distribution, Tv q . Toward this end, the forward
subject to the a priori knowledge on the sparseness of the scattering problem at hand is formulated as the following
scatterer with respect to the representation basis ψ, starting minimization one:
from the initialization (q = 1) of the total field coefficients  P 
Tv(l) q−1 = Tv(l) 0 (v = 1, . . . , V ) according to the BA [26]: Tv q = arg min Iv (r p ) − Tv( p)
 Tv
p=1

Tv(l)
≈ Iv (r)γ (l) (r)dr, l = 1, . . . , L. (13) 
k 2  (l)  (n)
L N
0 
+j Tv 
τ ψ (n) (r )
Owing to the linearity of (12) with respect to τ and the 4 
l=1 n=1 q
sparse nature of this latter, CS techniques appear as a natural 
choice for solving the inversion problem at hand. However, ×γ (l) (r )H(k|r p − r |)dr (19)
due to the impossibility of a priori guaranteeing either the
incoherence between the sensing and the measurement bases then solved through [26] to yield (v = 1, . . . , V )
or the compliance with the restricted isometry property of the
sensing matrix [1], [3] arising in realistic microwave imaging 
Tv q = Iv + Fq 
Tv q−1 (20)
problems, CS formulations alternative to the standard 1 -norm where
regularization need to be used. Accordingly, the qth (q = 1,
. . . , Q) estimate of the contrast function, τ q , is retrieved Iv  {Iv (r p ); p = 1, . . . , P} (21)
1756 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO. 5, MAY 2019

and Fq is the qth forward observation matrix whose


expression is
 
k 2  (n)
N
Fq  −j 
τ q ψ (n) (r )γ (l) (r )
4 
n=1

 
×H(k|r − r |)dr ; p = 1, . . . , P, l = 1, . . . , L .
p

(22)

The two-task loop of the BI-CSM is then iterated by incre-


menting the iteration index q (q ← q + 1) and updat-
ing the estimates of  τ q and  Tv q until (a) a maximum
number of iterations q = Q is reached or the stagna-
tion conditions (b) (
 τ q − τ q−1
2 /

τ q−1
2 ) ≤ μ or
(c) (
 Tv q − Tv q−1
2 /
 Tv q−1
2 ) ≤ ν are met, μ and ν
being user-defined convergence thresholds, when the conver-
gence distributions are returned:  τ opt = τ q and  Tv opt =

Tv q (v = 1, . . . , V ). Accordingly, it is worthwhile to remark
that the stopping criterion does not imply the knowledge of
the true electromagnetic profiles of the scatterers τ (r).
In short, the BI-CSM can be summarized with the following
sequence of algorithmic steps:
1) Initialization:Set q = 1 and initialize the contrast
function to the background value (i.e.,  τ 0 = 0). Choose
Fig. 1. Benchmark scattering distributions (τ = 2.0). Dielectric profiles for
the BCS control parameters β1 and β2 . For each vth the (a) square, (b) C-shaped, (c) rectangle, (d) multiple-squares, and (e) hollow
(v = 1, . . . , V ) view, collect the field data Ev , Iv , and targets.
compute the entries of  Tv 0 (13).
2) Hyperparameter Estimation: Compute  α q (17).
3) Contrast Update: Compute  τ q (14). by computing the corresponding (total/internal/external) aver-
4) Field update: Compute  Tv q (20) (v = 1, . . . , V ). age error figures
5) Convergence Check: Check the converge criteria (a)–(c). 
1 |
τ (r)opt − τ (r)|
If the convergence has been attained then return  τ opt = ξ  dr (23)
A  |τ (r) + 1|
τ q and 
 Tv opt =  Tv q (v = 1, . . . , V ), otherwise
update qarr owq + 1 and go to 2. where A being the area of the region ,  ∈
{tot, i nt, ext}.
It is worthwhile to remark that other CS tools (see [1], [3], and
the reference therein), alternative to the BCS one (17), may
be applied to solve (12) still keeping unaltered the BI-CSM A. Illustrative Example
scheme. On the other hand, let us point out that unlike [29] Let us consider, a representative benchmark scenario,
(and its exploitation in the CS framework [7]) the proposed a 3λ-sided square investigation domain  probed
iterative procedure does not require a full-wave solution of a by V = 27 plane waves impinging from the directions
forward problem (19) at each qth (q = 1, . . . , Q) iterative θv = (2π(v − 1)/V ), v = 1, . . . , V , with measurement points
step, but the field distribution is estimated through (20) [26] located in an external circular (ρ = 3.0 λ in radius) observation
with a nonnegligible reduction of the computational cost of domain  at the positions rm = (ρ cos(2π(m − 1)/M),
the whole inversion process. ρ sin(2π(m − 1)/M)) (m = 1, . . . , M). Without the loss
of generality [3], [22], [23], the scatterers have been
III. N UMERICAL AND E XPERIMENTAL A SSESSMENT assumed to be sparse with respect to the pulse basis (i.e.,
ψ  {ψn (r); n = 1, . . . , N} and ψn (r) = 1 if r ∈ (n) and
This section has a twofold objective. On the one hand, it is ψn (r) = 0 otherwise) and to avoid the inverse crime [30] a
devoted to derive general (i.e., a suboptimal tradeoff among PDir = 36 × 36 direct and a P = 18 × 18 (N = P) inverse
all the optimal single-inversion cases) operative guidelines for discretization grids featuring (λ/12)- and (λ/6)-sided cells
the reliable application of the proposed inversion method to (λ being the free-space wavelength), respectively, have been
handle nonweak scatterers. On the other hand, it is aimed at employed.
validating the BI-CSM also in comparison with some state- The first numerical example is aimed at illustrating, on a
of-the-art competitive techniques. Toward this end, besides step-by-step basis, the inversion process when applying the
the pictorial/visual representation of the reconstructed contrast BI-CSM. Toward this end, an off-centered λ/3-sided square
profiles, the inversion accuracy will be quantitatively estimated target with τ = 2.0 [S = 4; Fig. 1(a)] has been retrieved by
OLIVERI et al.: CS-BASED BI METHOD FOR TOMOGRAPHIC IMAGING 1757

Fig. 3. Illustrative example (“square” profile, S N R = 20 dB, τ = 2.0).


Total (ξtot ), internal (ξint ), and external (ξext ) error indices as well as inversion
time (t BI-CSM ) versus the iteration number (q = 1, . . . , Q).

Fig. 2. Illustrative example (“square” profile, S N R = 20 dB, τ = 2.0).


BI-CSM retrieved contrast when (a) q = 1, (b) q = 2, (c) q = 3, (d) q = 4,
(e) q = 5, (f) q = 6, (g) q = 7, and (h) q = Q = 8.

processing scattering data characterized by a signal-to-noise


ratio (SNR) [21] equal to S N R = 20 dB and setting the
BCS parameters to β1 = 4.0 × 101 and β2 = 5.0 × 10−2 Fig. 4. Illustrative example (“square” profile, S N R = 20 dB, τ = 2.0).
according to [31]. The plots of the evolution of the retrieved (a) Magnitude of the actual field (|Tv (r)|) and normalized difference between
profile, τ q , q = 1, . . . , Q, show a refinement of the BI- the actual and the BI-CSM estimated (q = Q) field distribution, Tv (r)
v (r)||/maxr {|Tv (r)|})], for the v = 1 view (θv = 0
[Tv (r)  (||Tv (r)| − |T
CSM reconstruction during the multi-iteration procedure (see [deg]): (b) q = 1, (c) q = 2, (d) q = 3, (e) q = 4, (f) q = 5, (g) q = 6,
Fig. 2). Indeed, although the location and the sparse nature (h) q = 7, and (i) q = Q = 8.
of the scatterer can be clearly recognized in all the itera-
tions [see Fig. 2(a)–(h)], the accuracy significantly improves
thanks to the sequence of approximations carried out within of the external errors are often equal to zero (ξext q = 0 when
the BI scheme [e.g., q = 1, Fig. 2(a), versus q = 8, q = 2, 3, 5, 7, and 8) and never above 0.13% as a result
Fig. 2(h)]. Such a feature, which is quantitatively confirmed of the sparseness-promoting solver that mitigates/avoids the
by the corresponding decrement of the total error values (e.g., background artifacts despite the noisy data and the approx-
ξtot q=1 = 7.98 ×10−3 versus ξtot q=8 = 8.71 ×10−4 , Fig. 3), imated formulation. Moreover, the monotonic decreasing of
is mainly due to the capability of the BI-CSM scheme to both the internal, ξint q , and total, ξtot q , errors (see Fig. 3)
profitably exploit, on the one hand, the reformulation of the underlines the reconstruction improvements throughout the
full nonlinear scattering problem into a set of (much simpler) iterations (e.g., (ξtot q − ξtot q+1 /ξtot q+1 ) ≈ 134% when
linear ones more and more accurate, and, on the other hand, q = 6, Fig. 3) as a result of a more and more faithful prediction
the a priori knowledge on the sparsity of the unknown with of the field distribution Tv (r) in  (see Fig. 4) thanks
respect to the pulse basis. As a matter of fact, the magnitudes to the BI-based approximation of the scattering equations.
1758 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO. 5, MAY 2019

Fig. 6. BI-CSM calibration (“C-shaped” profile, τ = 2.0). BI-CSM retrieved


contrast when processing noisy data with (a) and (b) SNR=20 dB and
opt opt
(c) and (d) SNR=10 dB by setting (a) and (c) β1 = β1 and β2 = β2
sub sub
or (b) and (d) β1 = β1 and β2 = β2 .

the literature outcome [31] on the dependence of the optimal


opt
prior configuration on the noise level [e.g., β1 SNR=50 dB ≈
2.0 × 10−1 versus β1 SNR=10 dB ≈ 6.0 × 10−1 ]. Therefore,
opt

an operative rule to set β1 , when no information on the actual


Fig. 5. BI-CSM calibration (“C-shaped” profile, τ = 2.0). Total retrieval noise level is available, is to choose the optimal tradeoff value
error of the BI-CSM reconstruction versus (a) β1 (β2 ≡ β2 ) and
opt  
50
opt opt
(b) β2 (β1 ≡ β1 ). β1  arg min (ξtot SNR )dS N R = 5.0. (24)
β1 5

The BI-CSM loop has been terminated when q = Q = 8 Analogously, the tradeoff setup for β2 turns out to be
since no significant changes in  τ q have been detected fur-  
50
(ξtot SNR )dS N R = 2.0 × 10−2 . (25)
opt
ther iterating the inversion process. As for the computational β2  arg min
costs, the BI-CSM β2
 Q turns out to be numerically very efficient 5
(TB I −C S M = q=1 tq ≈ 1.34 × 101 s), despite its iterative By fixing those thresholds as default values for the control
nature,2 because of the well-known speed of CS solvers based parameters, the accuracy of the arising BI-CSM inversion has
on Bayesian formulations [1], [3] when exploiting the a priori been then investigated through some numerical experiments.
information on the target sparsity at each qth (q = 1, . . . , Q) For instance, the plots of the retrieved profiles when
inversion (i.e., tq ∈ [0.31, 3.27] s, Fig. 3). S N R = 20 dB [see Fig. 6(a)] or SNR = 10 dB [see Fig. 6(c)]
indicate that, notwithstanding the presence of the noise
B. BI-CSM Calibration (i.e., S N R ∈ {10, 20} dB, Fig. 6) and the complexity
of the target [i.e., S = 8, Fig. 1(b)], the shape and the
It is worth noticing that the scattering data of the previ- position of the scatterer are correctly identified, while
ous illustrative example have been processed by setting the only minor background artifacts [see Fig. 6(c)] are present
scale priors of the BI-CSM according to the state-of-the-art as quantitatively confirmed by the corresponding error
guidelines for the BCS [31]. It is then of interest for the figures [i.e., ξtot SNR=20 dB = 3.74 × 10−3 , Fig. 6(a);
potential BI-CSM users to understand whether those setting ξtot SNR=10 dB = 9.50 × 10−3 , Fig. 6(c)]. On the
rules are still effective when integrating the BCS within the other hand, the plots in Fig. 5 also give some useful
BI-CSM loop as well as the dependence of the reconstruction insights on the robustness of the BI-CSM technique on
on the values of those control parameters. Toward this end, possible misconfigurations of its control parameters. As a
the retrieval of a “C-shaped” profile with τ = 2.0 [S = 8, matter of fact, the retrieval error is always below 3.5%
Fig. 1(b)] has been carried out by varying the values of β1 regardless of the choice of the scale priors values within
and β2 . In Fig. 5(a), the behavior of ξtot versus β1 confirms the admissibility ranges [31] and the noise level [i.e.,
2 For the sake of fairness, nonoptimized MATLAB implementations of all ξtot ≤ 3.1% when β1 ∈ [1, 102 ] and S N R ∈ [5, 50]
the methods have been run hereinafter on a laptop equipped with a single-core dB, Fig. 5(a); ξtot ≤ 3.5% when β2 ∈ [10−3 , 10−1 ] and
2.16-GHz CPU. S N R ∈ [5, 50] dB, Fig. 5(b)]. As an example, let us
OLIVERI et al.: CS-BASED BI METHOD FOR TOMOGRAPHIC IMAGING 1759

Fig. 8. Numerical assessment (“rectangle” profile, τ = 2.0). (a) Reconstruc-


tion errors versus SNR and retrieved profiles with the (b) and (d) BI-CSM
Fig. 7. BI-CSM calibration (“C-shaped” profile, τ ∈ {1.5, 2.5}). and the (c) and (e) BA-CSM when (b) and (c) SNR=20 dB and
(a) Total (ξtot ), (b) internal (ξint ), and (c) external (ξext ) errors. (d) and (e) SNR=10 dB.

consider the dielectric distributions found with the scale 4) the BI-CSM minimizes the occurrence of background
opt
priors set to β1sub  8 × β1 and β2sub  2.5 × β2
opt artifacts whatever the operative conditions [i.e., ξext ≤
[see Fig. 6(b) and (d)]. Although the accuracy reduces 1.96 × 10−2 , Fig. 7(c)].
[i.e., (ξtot (β sub ,β sub ) /ξtot (β opt ,β opt ) )SNR=20 dB ≈ 2.11,
1 2 1 2
Fig. 6(b); (ξtot (β sub ,β sub ) /ξtot (β opt ,β opt ) )SNR=10 dB ≈ 1.65, C. Numerical Assessment
1 2 1 2
Fig. 6(d)], the BI-CSM still gives a quite faithful qualitative The third numerical example deals with an off-centered
reconstruction of the target [see Fig. 6(b) and (d)]. “rectangle” with τ = 2.0 [see Fig. 1(c)] and, besides further
In order to assess the generality of the choice of the checking the reliability of the deduced BI-CSM guidelines
tradeoff scale priors (24) and (25), different contrasts have when imaging differently shaped targets, it is concerned with
been handled next. More specifically, the same “C-shaped” a comparison with a state-of-the-art BA-based CS method
profile [S = 8, Fig. 1(b)] has been imaged, but now setting (i.e., the BA-CSM) [21] to investigate on the benefits of
τ = 1.5 or τ = 2.5 (see Fig. 7). The outcomes from the exploiting the BI procedure. By analyzing the behavior of
behavior of the error figures versus the noise level are that: the error indices versus the SNR, it turns out that 1) the
1) the inversion gets better with the SNR regardless of BI-CSM always improves the error figures of the single-
the contrast (e.g., (ξtot SNR=50 dB /ξtot SNR=5 dB τ =1.5 ≈ step BA-CSM no matter what the noise level [i.e., ξtot ∈
6.97 × 10−2 [see Fig. 7(a)] and (ξint SNR=50 dB / [35%, 75%], ξint ∈ [11%, 70%], and ξext ∈ [25%, 100%],
ξint SNR=5 dB )τ =1.5 ≈ 1.44 × 10−1 [Fig. 7(b)]); being
2) the accuracy improves as τ reduces [e.g., (ξtot τ =2.5 / BA-CSM − ξ BI-CSM
ξ 
ξtot τ =1.5 )SNR=20 dB ≈ 6.41, Fig. 7(a)] because of the ξ  (26)
lower degree of nonlinearity of the scattering equations ξ
BA-CSM

when weaker targets are at hand [20];  ∈ {tot, i nt, ext}, Fig. 8(a)], 2) thanks to the combination
3) the total retrieval error is always below 4% [i.e., ξtot ≤ of the BI formulation and the constraint on the sparsity of
3.8 × 10−2 , Fig. 7(a)]; the targets at hand forced by the CS, the empty background
1760 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO. 5, MAY 2019

TABLE I
N UMERICAL ASSESSMENT (“ RECTANGLE ” PROFILE ,
τ = 2.0)—C OMPUTATIONAL I NDICES

is perfectly retrieved (i.e., ξext = 0) when S N R ≥ 10 dB


BA-CSM ≥ 2.5 × 10−3 [see Fig. 8(a)],
[see Fig. 8(a)], while ξext
3) the BI-CSM turns out to be robust to the noise since
BI-CSM 
ξtot SNR=5 dB ≤ 2% [see Fig. 8(a)]. As a representative
and illustrative proof of the improved performance enabled
by the iterative scheme, the reconstructions yielded when
SNR = 20 dB [see Fig. 8(b) and (c)] and SNR = 10 dB
[see Fig. 8(d) and (e)] are reported. As it can be observed,
the BI-CSM correctly locates, sizes, and shapes the unknown
profile in both noisy cases [i.e., SNR = 20 dB, Fig. 8(b);
SNR = 10 dB, Fig. 8(d)], while the first-order Born approach
misses both the shape and size of the target [see Fig. 8(e)]
also when SNR = 20 dB [see Fig. 8(c)]. Such a considerable
improvement, which will be confirmed also in the subsequent
comparative examples, is physically motivated by the capa-
bility of the proposed strategy to profitably address a fully
nonlinear formulation of the scattering problem, unlike the
BA-CSM technique [21], while also encoding the a priori
knowledge on the unknown profile sparsity within the solution
process (unlike standard state-of-the-art inversion methodolo-
gies). Of course, the price-to-pay for such finer approxi- Fig. 9. Numerical assessment (“multiple-squares” profile). Reconstruction
errors versus SNR when (a) τ ∈ {1.0, 2.0} and dielectric profiles retrieved
mation of the BI-CSM (see Section II) with respect to the with (b) and (d) BI-CSM and (c) and (e) BA-CSM when processing the noisy
coarser linearization of the BA-CSM [21] is the computational data with (b) and (c) SNR=10 dB and (d) and (e) SNR=5 dB scattered from
time. Indeed, the one-shot BA-CSM [21] minimizes the CPU a contrast τ = 2.0.
inversion time (i.e., 28% ≤ (t BA-CSM /t BI-CSM ) ≤ 33%,
Table I). However, it is worth pointing out that 1) the BA-
CSM method is known to computationally outperform most deterministic method that exploits a conjugate-gradient (CG)
state-of-the-art inversion strategies [21] and that 2) the time solver [33], [34] to minimize the data-misfit function of the
required for a BI-CSM reconstruction is of the same order BA formulation of the inverse scattering problem [34]. From
of magnitude of the noniterative BA-CSM (see Table I). This the maps of ξtot yielded when imaging the “C-shaped” profile
outcome is also theoretically supported by the comparison of [see Fig. 1(b)] in different noisy conditions and for various
the computational complexity, χ, of the BI-CSM and BA-CSM contrasts, one can infer the following:
techniques (i.e., χ BI-CSM ∝ O(2 × Q × V × N × S 2 ) versus 1) the BI-CSM outperforms the CG-based approach regard-
χ BA-CSM ∝ O(2 × V × N × S 2 ) - [32]). less of the contrast and the SNR [Fig. 10(a) versus
The problem of retrieving multiple targets is addressed by Fig. 10(b)];
considering a “multiple-squares” scattering scenario [S = 12, 2) a nonnegligible degradation of the CG-inversion arises
Fig. 1(d)]. The comparison of the plots of ξtot versus the as τ increases even slightly blurring the scattering data
SNR for different τ values [see Fig. 9(a)] confirms both the [e.g., ξtot > 2.5×10−2 when τ ≥ 2.2 ∀ SNR, Fig. 10(b)],
BI-CSM enhanced accuracy [see (ξtot BA-CSM /ξ BI-CSM )
τ =1.0 ∈ while the BI-CSM exhibits a retrieval error ξtot ≥ 2.5 ×
tot
[270%, 1420%], (ξtot BA-CSM /ξ BI-CSM )
τ =1.0 ∈ [215%, 515%], 10−2 only in very heavy noisy conditions and strong
tot
Fig. 9(a)] and its capability to reliably image scatterers fea- scatterers [i.e., SNR≤ 1.843 × exp(0.3992 × τ ) + 1.03 ×
BA-CSM 
turing higher contrasts [e.g., ξtot τ =1.0 > ξtot
BI-CSM 
τ =2.0 10−12 × exp(7.714 × τ ) dB, Fig. 10(a)].
when SNR = 5 dB, Fig. 9(a)]. Such a quantitative inference For illustrative purposes, the plots of the profiles retrieved
is visually highlighted by the images of the reconstructions when imaging the “C-shaped” scatterer with contrast
when SNR = 10 dB [Fig. 9(b) versus Fig. 9(c)] and SNR = τ = 2.0 by processing low-noise [i.e., S N R = 50 dB,
5 dB [Fig. 9(d) versus Fig. 9(e)]. Fig. 11(a) and (b)] and high-noise [i.e., S N R = 5 dB,
The next numerical experiment is aimed at extend- Fig. 11(c) and (d)] scattering data visually assess the superior
ing the comparative assessment of the BI-CSM to a capabilities of the BI-CSM with also a considerable time
OLIVERI et al.: CS-BASED BI METHOD FOR TOMOGRAPHIC IMAGING 1761

TABLE II
N UMERICAL A SSESSMENT (“C-S HAPED ” P ROFILE ,
τ = 2.0)—C OMPUTATIONAL I NDICES

Fig. 10. Numerical performance assessment (“C-shaped” profile). Map of


total reconstruction error versus τ and SNR when processing the scattering
data with (a) BI-CSM and (b) CG.

Fig. 12. Numerical assessment (“hollow” profile). Reconstruction errors ver-


sus (a) target contrast τ and dielectric profiles retrieved by (b) and (d) BI-CSM
and (c) and (e) GA when processing noisy data with SNR=10 dB scattered
from an object with (b) and (c) τ = 1.0 and (d) and (e) τ = 2.0.

also considering the CG computational complexity—χ CG ∝


O(4 × N 2 × V ) [35]—and it is due to the feature of CS-based
approaches of implicitly and effectively exploiting the prior
information on the sparse nature of the class of scatterers under
test [1], [3], [21].
For completeness, the last test of the comparative study
is concerned with an “hollow” profile with τ = 2.0
[S = 8, Fig. 1(e)] taken from [36] and there retrieved
with a state-of-the-art stochastic technique, which is based
Fig. 11. Numerical assessment (“C-shaped” profile, τ = 2.0). Retrieved on a standard genetic algorithm (GA), formulated within
contrast with (a) and (c) BI-CSM when processing noisy data with the BA [36] for a fair comparison with the BI-CSM. The
SNR = 50 dB and (b) and (d) CG when processing noisy data with
(c) and (d) SNR = 5 dB. quantitative evaluation of the inversion performance ver-
sus the contrast value [see Fig. 12(a)] shows, also in this
saving with respect to the CG-based inversion (i.e., (t CG / case, that the proposed sparsity-promoting technique yields
t BI-CSM )SNR=50 dB ≈ 253%, (t CG /t BI-CSM )SNR=5 dB ≈ a smaller error whatever the value of τ and the noise level
238%, Table II). Such a numerical efficiency is not surprising [e.g., 170% ≤ (ξtot GA /ξ BI-CSM )
tot SNR=20 dB ≤ 1100%,
1762 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO. 5, MAY 2019

TABLE III
N UMERICAL A SSESSMENT (“H OLLOW ” P ROFILE ,
τ = 2.0)—C OMPUTATIONAL I NDICES

150% ≤ (ξtot GA /ξ BI-CSM )


tot SNR=10 dB ≤ 350%, Fig. 12(a)].
More specifically, the GA inversion does not allow a cor-
rect detection of the “hollow” nature of the scatterer [see
Fig. 12(c) and (e)] when the SNR is low, unlike the BI-CSM
[see Fig. 12(b) and (d)]. Moreover, thanks to the exploitation of
a fast RVM solver, the BI-CSM is considerably more efficient
than the stochastic inversion since there is a difference of
more than 3 orders of magnitude in reaching the conver-
gence solution (i.e., (t GA /t BI-CSM )τ =1.0 ≈ 1.82 × 103 ,
(t GA /t BI-CSM )τ =2.0 ≈ 1.90 × 103 , Table III), as also
theoretically expected by the corresponding computational
complexity (i.e., χ GA ∝ O(4 × N 2 × V × η × ), η being
the number of GA individuals per generation and  the total
number of generations [36], [37]).
D. Experimental Assessment
The final experiment is devoted to analyze the BI-CSM
inversion of experimental data related to nonBorn targets [38].
Toward this end, the benchmark data provided by the Frésnel
Institute of Marseille, France, [38], [39] have been employed. Fig. 13. Experimental assessment (“FoamDielInt” and “FoamDielExt”
Such experiments, carried out in the 14×6.5×6.5 m3 faradized profiles). (a) Actual and (b) retrieved profiles with (c) and (d) BA-CSM and
anechoic chamber of the Center Commun de Ressources en (e) and (f) BI-CSM when processing data (a), (c), and (e) “FoamDielInt” and
(b), (d), and (f) “FoamDielExt” at the frequency of 4.0 GHz.
Micro-ondes [39], have been performed using V = 8 and
M = 241 positions of transmitting and receiving ridged horn
antennas (ARA DRG-118), located at ρ = 1.67 m with an TABLE IV
angular positioning accuracy of 0.01 deg and connected to an E XPERIMENTAL A SSESSMENT (“F OAM D IEL I NT ” AND
“F OAM D IEL E XT ” P ROFILES )—C OMPUTATIONAL I NDICES
Agilent HP 8530 vector network analyzer [39]. The targets,
consisting of cylinders with a length of 1.5 m, have been
supported by a Polyamide 6 tube with 0.1 m diameter [39].
More specifically, the “FoamDielInt” [see Fig. 13(a)] and the
“FoamDielExt” profiles [see Fig. 13(b)] have been considered
and, because of the single-frequency formulation of the BI-
CSM, the data collected at 4 GHz [see Fig. 13(c)–(f)] have
been processed. It is worth pointing out that such test cases deduced for a wide set of targets, contrasts, and noise
are very challenging for a CS-based approach since here levels (see Fig. 6). On the other hand, the proposed
the objects under test are not sparse in the adopted pixel inversion method turns out to be robust, even though
representation [i.e., S ≈ 24, Fig. 13(a); S ≈ 28, Fig. 13(b)]. suboptimally performing, to possible misconfigurations
Therefore, although the BI-CSM reconstructions are better of its control parameters (see Fig. 5);
than the BA-CS ones, they are not so accurate as in the 2) The BI-CSM outperforms single-step BA-formulated CS
previous pixel-sparse test cases. On the other hand, one can imaging strategies [21] in terms of accuracy, flexibility,
observe that the position, the size, and the inhomogeneous and reliability (see Fig. 8) when processing experimental
nature of the unknown targets [i.e., Fig. 13(c) versus Fig. 13(e) data sets (see Fig. 13), as well;
and Fig. 13(d) versus Fig. 13(f)] are roughly estimated. 3) The BI-CSM positively competes with the state-
For completeness, the inversion times for both BI-CSM and of-the-art deterministic (see Fig. 10) and stochastic
BA-CSM have been reported in Table IV. (see Fig. 12) inversion techniques formulated within the
Born approximation for a fair comparison;
IV. R EMARKS 4) Despite its iterative nature, the BI-CSM turns out much
The numerical assessment (see Section III) has shown the to be more computationally efficient than standard
following. CG- and GA-based ones (see Tables II and III) with
1) Effective and reliable tradeoff setups of the calibration an inversion time of the same order of magnitude of the
parameters of the BI-CSM [i.e., (24) and (25)] can be BA-CSM (see Table I).
OLIVERI et al.: CS-BASED BI METHOD FOR TOMOGRAPHIC IMAGING 1763

Future works, beyond the scope of this paper, will be [11] V. M. Patel, G. R. Easley, D. M. Healy, Jr., and R. Chellappa, “Com-
aimed at extending the proposed imaging scheme to mul- pressed synthetic aperture radar,” IEEE J. Sel. Topics Signal Process.,
vol. 4, no. 2, pp. 244–254, Apr. 2010.
tilayered [40], [41] and 3D [42] scenarios as well as at [12] E. J. Candes and T. Tao, “Decoding by linear programming,” IEEE
generalizing the BI-CSM to process multifrequency scatter- Trans. Inf. Theory, vol. 51, no. 12, pp. 4203–4215, Dec. 2005.
ing data. Moreover, the exploitation of alternative sensing [13] R. G. Baraniuk, “Compressive sampling,” IEEE Signal Process. Mag.,
bases [22], [23], [43] as well as the use of alternative iterative vol. 24, no. 4, pp. 118–124, Jul. 2007.
[14] E. J. Candes and M. B. Wakin, “An introduction to compressive
linearization schemes, such as the BIM [7], [29], [44] or the sampling,” IEEE Signal Process. Mag., vol. 25, no. 2, pp. 21–30,
DBIM [45], [46], is currently under investigation. Mar. 2008.
[15] M. Cossalter, G. Valenzise, M. Tagliasacchi, and S. Tubaro, “Joint com-
pressive video coding and analysis,” IEEE Trans. Multimedia, vol. 12,
V. C ONCLUSION no. 3, pp. 168–183, Apr. 2010.
[16] J. A. Tropp and S. J. Wright, “Computational methods for sparse solution
An innovative inversion method has been proposed to image of linear inverse problems,” Proc. IEEE, vol. 98, no. 6, pp. 948–958,
nonweak targets in the CS framework. To avoid CSFs, a Jun. 2010.
BI formulation has been considered and the retrieval of the [17] K. Xu, Y. Zhong, R. Song, X. Chen, and L. Ran, “Multiplicative-
regularized FFT twofold subspace-based optimization method for inverse
contrast distribution has been carried out with an iterative scattering problems,” IEEE Trans. Geosci. Remote Sens., vol. 53, no. 2,
procedure. The key methodological advancements include: pp. 841–850, Feb. 2015.
1) the integration of a sparsity-constrained regularizer in [18] X. Y. Wang, M. Li, and A. Abubakar, “Acceleration of 2-D multi-
plicative regularized contrast source inversion algorithm using paralleled
the BI formulation without requiring cumbersome full-wave computing architecture,” IEEE Antennas Wireless Propag. Lett., vol. 16,
simulations; 2) the derivation of a CS-based inversion to effec- pp. 441–444, 2017.
tively handle nonweak scatterers while avoiding CSFs; and [19] K. Xu, Y. Zhong, and G. Wang, “A hybrid regularization technique
3) the deduction/validation of a set of operative guidelines its for solving highly nonlinear inverse scattering problems,” IEEE Trans.
Microw. Theory Techn., vol. 66, no. 1, pp. 11–21, Jan. 2018.
reliable exploitation. The reported results have confirmed the [20] M. Slaney, A. C. Kak, and L. E. Larsen, “Limitations of imaging with
effectiveness of the proposed BI-CSM, as well as highlighted first-order diffraction tomography,” IEEE Trans. Microw. Theory Techn.,
its computational efficiency. vol. MTT-32, no. 8, pp. 860–874, Aug. 1984.
[21] L. Poli, G. Oliveri, and A. Massa, “Microwave imaging within the
first-order born approximation by means of the contrast-field Bayesian
ACKNOWLEDGMENT compressive sensing,” IEEE Trans. Antennas Propag., vol. 60, no. 6,
pp. 2865–2879, Jun. 2012.
Author A. Massa wishes to thank E. Vico for her never- [22] N. Anselmi, M. Salucci, G. Oliveri, and A. Massa, “Wavelet-based
ending support and help. compressive imaging of sparse targets,” IEEE Trans. Antennas Propag.,
vol. 63, no. 11, pp. 4889–4900, Nov. 2015.
[23] N. Anselmi, G. Oliveri, M. A. Hannan, M. Salucci, and A. Massa,
R EFERENCES “Color compressive sensing imaging of arbitrary-shaped scatterers,”
IEEE Trans. Microw. Theory Techn., vol. 65, no. 6, pp. 1986–1999,
[1] A. Massa, P. Rocca, and G. Oliveri, “Compressive sensing in Jun. 2017.
electromagnetics—A review,” IEEE Antennas Propag. Mag., vol. 57, [24] G. Oliveri, L. Poli, P. Rocca, and A. Massa, “Bayesian compressive
no. 1, pp. 224–238, Feb. 2015. optical imaging within the Rytov approximation,” Opt. Lett., vol. 37,
[2] P. Shah, U. K. Khankhoje, and M. Moghaddam, “Inverse scattering no. 10, pp. 1760–1762, 2012.
using a jointL1 − L2norm-based regularization,” IEEE Trans. Antennas [25] G. Oliveri, P. Rocca, and A. Massa, “A Bayesian-compressive-sampling-
Propag., vol. 64, no. 4, pp. 1373–1384, Apr. 2016. based inversion for imaging sparse scatterers,” IEEE Trans. Geosci.
[3] G. Oliveri, M. Salucci, N. Anselmi, and A. Massa, “Compressive sensing Remote Sens., vol. 49, no. 10, pp. 3993–4006, Oct. 2011.
as applied to inverse problems for imaging: Theory, applications, current [26] S. Caorsi, A. Massa, and M. Pastorino, “Iterative numerical computation
trends, and open challenges,” IEEE Antennas Propag. Mag., vol. 59, of the electromagnetic fields inside weakly nonlinear infinite dielectric
no. 5, pp. 34–46, Oct. 2017. cylinders of arbitrary cross sections using the distorted-wave Born
[4] Y. Rodríguez-Vaqueiro, Y. Á. López, B. González-Valdes, approximation,” IEEE Trans. Microw. Theory Techn., vol. 44, no. 3,
J. Á. Martínez, F. Las-Heras, and C. M. Rappaport, pp. 400–412, Mar. 1996.
“On the use of compressed sensing techniques for improving multistatic [27] X. Chen, Computational Methods for Electromagnetic Inverse Scatter-
millimeter-wave portal-based personnel screening,” IEEE Trans. ing. Singapore: Wiley, 2018.
Antennas Propag., vol. 62, no. 1, pp. 494–499, Jan. 2014. [28] J. Richmond, “Scattering by a dielectric cylinder of arbitrary cross
[5] L. Pan, X. Chen, and S. P. Yeo, “A compressive-sensing-based phaseless section shape,” IEEE Trans. Antennas Propag., vol. AP-13, no. 3,
imaging method for point-like dielectric objects,” IEEE Trans. Antennas pp. 334–341, May 1965.
Propag., vol. 60, no. 11, pp. 5472–5475, Nov. 2012.
[29] Y. M. Wang and W. C. Chew, “An iterative solution of the
[6] X. Zhu, Z. Zhao, J. Wang, J. Song, and Q. H. Liu, “Microwave-induced two-dimensional electromagnetic inverse scattering problem,” Int. J.
thermal acoustic tomography for breast tumor based on compressive Imag. Syst. Technol., vol. 1, no. 1, pp. 100–108, 1989.
sensing,” IEEE Trans. Biomed. Eng., vol. 60, no. 5, pp. 1298–1307,
May 2013. [30] F. Cakoni and D. Colton, Qualitative Methods in Inverse Scattering
Theory, New York, NY, USA: Springer-Verlag, 2006.
[7] L. Guo and A. M. Abbosh, “Microwave imaging of nonsparse domains
using born iterative method with wavelet transform and block sparse [31] L. Poli, G. Oliveri, P. Rocca, and A. Massa, “Bayesian compressive
Bayesian learning,” IEEE Trans. Antennas Propag., vol. 63, no. 11, sensing approaches for the reconstruction of two-dimensional sparse
pp. 4877–4888, Nov. 2015. scatterers under TE illuminations,” IEEE Trans. Geosci. Remote Sens.,
[8] A. C. Gurbuz, J. H. McClellan, and W. R. Scott, “A compressive sensing vol. 51, no. 5, pp. 2920–2936, May 2013.
data acquisition and imaging method for stepped frequency GPRs,” IEEE [32] S. Ji, D. Dunson, and L. Carin, “Multitask compressive sensing,” IEEE
Trans. Signal Process., vol. 57, no. 7, pp. 2640–2650, Jul. 2009. Trans. Signal Process., vol. 57, no. 1, pp. 92–106, Jan. 2009.
[9] M. Salucci, A. Gelmini, L. Poli, G. Oliveri, and A. Massa, “Progressive [33] W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery,
compressive sensing for exploiting frequency-diversity in GPR imaging,” Numerical Recipes: The Art of Scientific Computing, 3rd ed. New York,
J. Electromagn. Waves Appl., vol. 32, no. 9, pp. 1164–1193, Oct. 2018. NY, USA: Cambridge Univ. Press, 2007.
[10] M. D. Migliore, “A simple introduction to compressed sensing/sparse [34] H. Harada, D. J. N. Wall, T. Takenaka, and M. Tanaka, “Conjugate
recovery with applications in antenna measurements,” IEEE Antennas gradient method applied to inverse scattering problem,” IEEE Trans.
Propag. Mag., vol. 56, no. 2, pp. 14–26, Apr. 2014. Antennas Propag., vol. 43, no. 8, pp. 784–792, Aug. 1995.
1764 IEEE TRANSACTIONS ON MICROWAVE THEORY AND TECHNIQUES, VOL. 67, NO. 5, MAY 2019

[35] J. R. Shewchuk, “An introduction to the conjugate gradient method Lorenzo Poli (GS’10–M’12) received the M.S.
without the agonizing pain,” Carnegie Mellon Univ., Pittsburgh, PA, degree in telecommunication engineering from the
USA, Tech. Rep., 1994. University of Trento, Trento, Italy, in 2008, and
[36] F. Viani, L. Poli, G. Oliveri, F. Robol, and A. Massa, “Sparse scat- the Ph.D. degree from the International Doctoral
terers imaging through approximated multitask compressive sensing School in Information and Communication Technol-
strategies,” Microw. Opt. Technol. Lett., vol. 55, no. 7, pp. 1553–1558, ogy, Trento, in 2012.
Jul. 2013. He is currently a member of the ELEDIA Research
[37] P. Rocca, M. Benedetti, M. Donelli, D. Franceschini, and A. Massa, Center, University of Trento. His current research
“Evolutionary optimization as applied to inverse scattering problems,” interests include the solution of antenna array
Inverse Problems, vol. 25, no. 12, pp. 1–41, Nov. 2009. synthesis and electromagnetic inverse scattering
[38] K. Belkebir and M. Saillard, “Testing inversion algorithms against problems.
experimental data,” Inverse Problems, vol. 21, no. 6, p. S1, Dec. 2005.
[39] J.-M. Geffrin, P. Sabouroux, and C. Eyraud, “Free space experimental
scattering database continuation: Experimental set-up and measurement
precision,” Inverse Problems, vol. 21, no. 6, p. S117, Nov. 2005.
[40] Y. Zhong, M. Lambert, D. Lesselier, and X. Chen, “Electromagnetic
response of anisotropic laminates to distributed sources,” IEEE Trans.
Antennas Propag., vol. 62, no. 1, pp. 247–256, Jan. 2014.
[41] Y. Hu, Y. Fang, D. Wang, Y. Zhong, and Q. H. Liu, “Electromagnetic
waves in multilayered generalized anisotropic media,” IEEE Trans.
Geosci. Remote Sens., vol. 56, no. 10, pp. 5758–5766, Oct. 2018.
[42] Y. Ren, Q. H. Liu, and Y. P. Chen, “A hybrid FEM/MoM method for
Nicola Anselmi (GS’13–M’16) received the mas-
3-D electromagnetic scattering in layered medium,” IEEE Trans. Anten-
ter’s degree in telecommunication engineering from
nas Propag., vol. 64, no. 8, pp. 3487–3495, Aug. 2016.
the University of Trento, Trento, Italy, in 2012, and
[43] M. Li, O. Semerci, and A. Abubakar, “A contrast source inversion
the Ph.D. degree from the International Doctoral
method in the wavelet domain,” Inverse Problems, vol. 29, no. 2,
School in Information and Communication Technol-
Jan. 2013, Art. no. 025015.
ogy, Trento, in 2018.
[44] D. Ireland, K. Bialkowski, and A. Abbosh, “Microwave imaging for
His current research interests include synthesis
brain stroke detection using Born iterative method,” IET Microw.,
methods for unconventional antenna array architec-
Antennas Propag., vol. 7, no. 11, pp. 909–915, Aug. 2013.
tures, tolerance analysis of antenna systems, and
[45] X. Ye and X. Chen, “Subspace-based distorted-born iterative method
electromagnetic inverse scattering techniques with
for solving inverse scattering problems,” IEEE Trans. Antennas Propag.,
interest on compressive sensing methodologies for
vol. 65, no. 12, pp. 7224–7232, Dec. 2017.
microwave imaging applications.
[46] W. C. Chew and Y. M. Wang, “Reconstruction of two-dimensional
Dr. Anselmi has been a member of the ELEDIA Research Center, University
permittivity distribution using the distorted Born iterative method,” IEEE
of Trento, since 2012, and a member of the IEEE Antennas and Propagation
Trans. Med. Imag., vol. 9, no. 2, pp. 218–225, Jun. 1990.
Society since 2014. In 2016, he was a recipient of the Giorgio Barzilai Award
for Young Researchers by the Italian Electromagnetic Society. He serves
as a Reviewer for different international journals, including the IEEE
T RANSACTIONS ON A NTENNAS AND P ROPAGATION, IEEE A NTENNAS AND
W IRELESS P ROPAGATION L ETTERS , and IET Microwaves, Antennas &
Propagation.

Marco Salucci (GS’13–M’15) received the M.S.


Giacomo Oliveri (GS’06–M’09–SM’13) received degree in telecommunication engineering from the
the B.S. and M.S. degrees in telecommunications University of Trento, Trento, Italy, in 2011, and
engineering and the Ph.D. degree in space sciences the Ph.D. degree from the International Doctoral
and engineering from the University of Genoa, School in Information and Communication Technol-
Genoa, Italy, in 2003, 2005, and 2009, respectively. ogy, Trento, in 2014.
He has been a Visiting Researcher with the Lab- From 2014 to 2016, he was a Post-Doctoral
oratoire des Signaux et Systèmes (L2S), Centrale- Researcher with CentraleSupélec, Gif-sur-Yvette,
Supélec, Gif-sur-Yvette, France, in 2012, 2013, and France. From 2016 to 2017, he was a Post-Doctoral
2015. He was an Invited Associate Professor with Researcher with the Commissariat à l’énergie atom-
the University of Paris-Sud, Orsay, France, in 2014. ique et aux énergies alternatives, Gif-sur-Yvette.
He is currently an Associate Professor with the He is currently a Researcher with the Department of Information Engineering
Department of Information Engineering and Computer Science, University and Computer Science, University of Trento, where he is also a member of
of Trento, Trento, Italy, and a member of the ELEDIA Research Center, the ELEDIA Research Center. His current research interests include inverse
L2S. He is also an Adjunct Professor with CentraleSupélec. He has authored scattering, GPR microwave imaging techniques, antenna synthesis, and com-
or co-authored over 330 peer-reviewed papers in international journals and putational electromagnetics with a focus on system-by-design methodologies
conferences. His current research interests include electromagnetic direct and integrating optimization techniques and learning-by-examples methods for
inverse problems, system-by-design and metamaterials, and antenna array real-world applications.
synthesis. Dr. Salucci was a member of the COST Action TU1208 Civil Engineering
Dr. Oliveri is the Chair of the IEEE AP/ED/MTT North Italy Chapter. Applications of Ground Penetrating Radar. He is a member of the IEEE
He serves as an Associate Editor for IEEE A NTENNAS AND W IRELESS Antennas and Propagation Society. He serves as a Reviewer for different
P ROPAGATION L ETTERS , the IEEE J OURNAL ON M ULTISCALE AND M UL - international journals, including the IEEE T RANSACTIONS ON A NTENNAS
TIPHYSICS C OMPUTATIONAL T ECHNIQUES , the International Journal of AND P ROPAGATION , IEEE A NTENNAS AND W IRELESS P ROPAGATION L ET-
Antennas and Propagation, the International Journal of Distributed Sensor TERS , the IEEE J OURNAL ON M ULTISCALE AND M ULTIPHYSICS C OMPU -
Networks, and the Microwave Processing journal. TATIONAL T ECHNIQUES , and IET Microwaves, Antennas & Propagation.
OLIVERI et al.: CS-BASED BI METHOD FOR TOMOGRAPHIC IMAGING 1765

Andrea Massa (M’03–SM’16–F’18) received the University of Singapore, Singapore. He has been appointed as the IEEE
Laurea degree in electronic engineering and the AP-S Distinguished Lecturer from 2016 to 2018. He has authored or
Ph.D. degree in electrical engineering and computer co-authored more than 700 scientific publications (partial list available
science from the University of Genoa, Genoa, Italy, at: http://eledia.science.unitn.it/index.php/ricerca/pubblicazioni), among which
in 1992 and 1996, respectively. about 300 appear in international journals and more than 450 in international
From 1997 to 1999, he was an Assistant Professor conferences, where he presented more than 150 invited contributions. He has
of electromagnetic fields with the Department of organized more than 70 scientific sessions in international conferences and
Biophysical and Electronic Engineering, University has participated to several technological projects in the European framework
of Genoa. From 2001 to 2004, he was an Associate (20 EU Projects) and at the national and local levels with national agen-
Professor with the University of Trento, Trento, Italy, cies (more than 150 Projects/Grants). His current research interests include
where he has been a Full Professor of electromag- inverse problems, analysis/synthesis of antenna systems and large arrays,
netic fields since 2005 and currently teaches electromagnetic fields, inverse radar systems synthesis and signal processing, cross-layer optimization and
scattering techniques, antennas and wireless communications, wireless ser- planning of wireless/RF systems, semantic wireless technologies, system-by-
vices and devices, and optimization techniques. He is currently the Director of design and material-by-design (metamaterials and reconfigurable materials),
the ELEDIA Research Center, Network of Federated Laboratories (ELEDIA, and theory/applications of optimization techniques to engineering problems
Universiti Teknologi Brunei, Bandar Seri Begawan, Brunei, ELEDIA, Univer- (tele-communications, medicine, and biology).
sity of Cassino, Cassino, Italy, ELEDIA, University of Electronic Science and Dr. Massa is an IET Fellow and an Electromagnetic Academy Fellow.
Technology of China (UESTC), Chengdu, China, ELEDIA, a the Universidad He was appointed in 2011 by the National Agency for the Evaluation of
San Ignacio de Loyola, Lima, Peru, ELEDIA, UniNAGA, Nagasaki, Japan, the University System and National Research as a member of the Recognized
ELEDIA at Laboratoire des Signaux et Systèmes, Paris, France, ELEDIA Expert Evaluation Group (Area 09, “Industrial and Information Engineering”)
at Czech Technical University in Prague, Prague, Czech, ELEDIA, Aristotle for the evaluation of the research at the Italian University and Research Center
University of Thessaloniki, Tessaloniki, Greece, ELEDIA at the Univer- for the period 2004–2010. Furthermore, he has been elected as the Italian
sity of Trento, Trento, ELEDIA, Tsinghua University, Beijing, China, and member of the Management Committee of the COST Action TU1208 Civil
ELEDIA, Innov’COM, Tunis, Tunisia). He is also an Adjunct Professor Engineering Applications of Ground Penetrating Radar. He serves as an Asso-
with Pennsylvania State University, State College, PA, USA, a Professor ciate Editor for the IEEE T RANSACTIONS ON A NTENNAS AND P ROPAGA -
with CentraleSupélec, Paris, France, a Guest Professor with UESTC, and TION and the International Journal of Microwave and Wireless Technologies.
the UC3M-Santander Chair of Excellence at the Universidad Carlos III de He is a member of the Editorial Board of the Journal of Electromagnetic
Madrid, Madrid, Spain. He has been the Holder of a Senior DIGITEO Waves and Applications, a permanent member of the PIERS Technical
Chair at L2S, CentraleSupélec, and CEA LIST, Saclay, France, a Visiting Committee and the EuMW Technical Committee, and an ESoA member.
Professor with the Missouri University of Science and Technology, Rolla, He has been appointed on the Scientific Board of the Società Italiana di
MO, USA, Nagasaki University, Nagasaki, Japan, the University of Paris-Sud, Elettromagnetismo and elected on the Scientific Board of the Interuniversity
Orsay, France, Kumamoto University, Kumamoto, Japan, and the National National Center for Telecommunications.

You might also like