IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO.

18, SEPTEMBER 15, 2016 4723

Nonparametric Bayesian Attributed Scattering Center

Extraction for Synthetic Aperture Radar Targets
Yulai Cong, Bo Chen, Member, IEEE, Hongwei Liu, Member, IEEE, and Bo Jiu, Member, IEEE

Abstract—As a good way to represent target backscatter target at high frequencies is well approximated as the sum of
measured by high-frequency synthetic aperture radar (SAR) responses from individual scattering centers on the target [1].
systems, the attributed scattering center (ASC) model is able to As a result, scattering center descriptions are extensively stud-
provide concise and physically relevant features of a complex ied and proposed as features for model-based automatic target
target and has played an important role in model-based automatic recognition (ATR) in high-frequency SAR target classification
target recognition (ATR). However, most existing ASC feature systems.
extraction methods suffer from imprecise image segmentation or In this paper, we adopt the attributed scattering center (ASC)
high computational cost, which greatly encumber their practical model as the representation of target backscatter [2], [3].
applications. To tackle this problem, we present a novel ASC This model includes both frequency and aspect dependence
feature extraction algorithm for SAR targets based on Lévy ran-
of scattering by retaining dominant terms of monostatic scat-
dom fields in a nonparametric Bayesian framework. Specifically,
Lévy random fields, yielding a natural sparse representation
tering solutions from both Physical Optics and the Geometric
of the unknown ASC model, are introduced to construct prior Theory of Diffraction [4]. It describes each scattering center
distributions, which lead to the specification of a joint prior by a set of parameters characterizing its location, amplitude,
distribution for the number of ASCs and the ASC associated shape, and orientation (pose), which provide a concise and
parameters. Meanwhile, the problem may be formulated as physically relevant description for the target. Therefore, the
a sparse representation problem, with regularization induced ASC model has been widely used to extract features for SAR
through the Lévy random field prior. We also develop a reversible ATR [5]–[8].
jump Markov chain Monte Carlo (RJ-MCMC) method to enable However, feature extraction based on the ASC model is chal-
relatively fast posterior inference. Experimental results confirm lenging due to the issues of its highly nonlinear characteristics
the effectiveness and efficiency of the proposed algorithm. and high-dimensional parameter space. To the best of our knowl-
edge, there are mainly two kinds of methods for this problem.
Index Terms—Attributed scattering center (ASC), Bayesian (1) Image-domain methods transform the radar backscatter into
nonparametric modeling, Lévy process, Lévy random fields, image domain and accomplish ASC feature extraction based on
feature extraction, reversible jump Markov chain Monte Carlo
the segmentation of the obtained image [2], [4], [9]. In prac-
(RJ-MCMC).
tice, SAR images are such highly intricacy that the segmenta-
tion is not precise enough to correctly isolate each ASC, which
I. INTRODUCTION may degenerate the final performance. (2) Frequency-domain
methods perform feature extraction directly in frequency-aspect
HE electromagnetic reflectivity of a target is closely re-
T lated to its geometry and orientation. Such reflectivity
data can be obtained via synthetic aperture radar (SAR) by
domain to avoid image segmentation. However, most exist-
ing frequency-domain methods suffer from high computational
complexity or storage requirement when it comes to the real-
illuminating the target with electromagnetic radiation and mea- world targets with a large number of ASCs [10], [11]. More-
suring the energy scattered back. The obtained data, named over, they usually need some ad-hoc stopping criteria (such as
as the radar backscatter, are a finite-extent, noisy set of in- a prescribed number of ASCs or residual energy) to control
formation about the scattering geometries of the target. It is the learning procedure, which are usually hard to define in the
interesting to extract features containing such information from real-world applications.
the radar backscatter, which may then be used to identify In this paper, the method that we discuss belongs to the sec-
the target. It is well known that the radar backscatter from a ond category. Considering the above drawbacks, it appeals that
the proposed approach should be able to provide more reliable
Manuscript received May 02, 2015; revised October 28, 2015 and April 05, parameter estimation with acceptable computational cost (both
2016; accepted April 26, 2016. Date of publication May 17, 2016; date of
current version July 25, 2016. The associate editor coordinating the review
time and memory), and at the same time, adaptively reach to
of this manuscript and approving it for publication was Dr. Fauzia Ahmad. an appropriate number of ASCs or residual energy in a fully
This research has been supported by the National Science Fund for Distin- data-driven way. An intuitive choice could be traditional sparse
guished Young Scholars (61525105), the Program for Thousand Youth Tal- reconstruction (SR) methods with lp -norm regularization [12].
ents by Chinese Central Government, the National Natural Science Founda- However, the size of the parameterized dictionary leads to
tion of China (61372132, 61271291), the National Defense Pre-research Fund tremendous computational issues due to the high-dimensionality
(9140A07010115DZ01015, 9140A07010115DZ01019), and the Program for
New Century Excellent Talents in University (NCET13-0945). (Corresponding
of the ASC model, and the regularization parameter compromis-
author: Bo Chen.) ing data fitting and sparsity needs to be tuned. Another choice
The authors are with the National Laboratory of Radar Signal Processing, lies in the Bayesian framework, where the specification of the
Xidian University, Xi’an, Shaanxi, China, 710071, and the Collaborative Inno- parameter priors is crucial.
vation Center of Information Sensing and Understanding at Xidian University, Lévy process is a stochastic process with great flexibility
Xi’an, Shaanxi, China, 710071 (e-mail: yulai_cong@163.com; bchen@mail. in modeling jumps, extremes and other anomalous behavior
xidian.edu.cn; hwliu@xidian.edu.cn; bojiu@xidian.edu.cn).
Color versions of one or more of the figures in this paper are available online
of phenomena [13]. Thus, it has been successfully applied in
at http://ieeexplore.ieee.org. different fields, such as statistics, finance and physics. More
Digital Object Identifier 10.1109/TSP.2016.2569463 generally, the authors in [14] exploited the corresponding

1053-587X © 2016 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.
4724 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

Lévy random fields, linear mappings from functions to random random fields, which will be adopted to induce prior distribu-
variables, and employed them to induce prior distributions for tions for our problem.
nonparametric function estimation. Motivated by their work, We start with the definition of a Gamma process:
we present a nonparametric Bayesian ASC feature extraction Definition 1: A Gamma process denotes a stochastic process
algorithm in this paper. In light of the specific parameterization Xt , t ≥ 0 with the properties:
form and underlying physical explanation of the ASC model 1) X0 = 0 almost surely;
in the real-world SAR imagery, we extend Gamma random 2) Xt has independent increments on disjoint intervals;
fields (one special class of Lévy random fields) to the complex 3) for any 0 ≤ s < t, Xt − Xs ∼ Ga (γ (t − s) , η).
domain to model the hierarchical priors for the ASC model, where Ga (·, ·) denotes the Gamma distribution, γ ( · )
including the number of ASCs, the backscattered coefficients, is the mean measure, and η is the scale parameter. If a
and the ASC associated parameters. After the derivation of Lebesgue measure is used for γ, a Gamma process Xt can be
the joint posterior distribution, we develop a reversible jump represented as
Markov chain Monte Carlo (RJ-MCMC) method [15] to enable  
relatively fast posterior inference. Xt = uN (du, ds) (3)
Compared with previous methods, our algorithm has five ad- R + ×[0,t]
vantages: 1. it does not require SAR image segmentation; 2.
the computational cost is much lower compared to most exist- with E [N (du, ds)] = ν (du) ds = γu−1 e−u η duds.
ing frequency-domain methods; 3. it provides an automatic and Similarly, we can construct a Gamma random field
adaptive learning scheme for the appropriate number of ASCs  
in a fully data-driven way; 4. it has the potential to provide Γ [ψ] = uψ (t) N (du, dt) (4)
more reliable parameter estimation; 5. the framework can be R + ×T
easily extended to deal with more complicated feature extrac-
tion problem where bistatic, 3-dimensional, or fully polarimetric where E [N (du, dt)] = γu−1 e−u η dudt.
ASC models are utilized [16]–[21]. For a generating function ψ (t; s, ϕ) with (s, ϕ) defined on
An outline of this paper is as follows. In Section II, we briefly S × Φ, if we set a real function f (t) by
introduce the theoretical background on Lévy random fields  
and the ASC model. Section III introduces related work. In f (t) = Γ [ψ] = uψ (t; s, ϕ) N (du, ds, dϕ) (5)
Section IV, we describe the proposed nonparametric Bayesian R + ×S ×Φ
ASC feature extraction algorithm by firstly providing the where N (du, ds, dϕ) defines a Poisson random field
adopted prior distributions, then the joint posterior distri- on R+ × S × Φ, with the corresponding mean measure
bution, and finally the corresponding RJ-MCMC method. E [N (du, ds, dϕ)] = γu−1 e−η u dudsπ (dϕ), where π (dϕ) de-
Section V demonstrates the effectiveness and efficiency of our fines a finite measure on Φ. The discrete nature of a Poisson
algorithm through experimental results on synthetic backscatter random field allows the following representation of (5):
data, XPATCH backscatter prediction data, and measured SAR
data. Section VI provides the conclusion. 
J
f (t) = ui ψ (t; si , ϕi ) . (6)
II. BACKGROUND THEORY AND ASC MODEL i=1

We model the radar backscatter as a realization of a Lévy Therefore, the Gamma random field Γ [ψ] governs the joint
random field with generating functions introduced by the ASC probability distribution of (J, {ui } , {si } , {ϕi }). In Bayesian
model. framework, Γ [ψ] serves as a natural prior distribution for the
function f (t).
A. Lévy Random Fields We can see from (3) that, for any ε > 0, a Gamma process
defined on [0, T ] with T > 0 can be viewed as jumps of size
According to the Lévy Khinchine theorem, any Lévy process u≥
can be decomposed into the sum of a Brownian motion with a  ∞ε arriving as a Poisson process with rate T ν ([ε, ∞)) =
T ε γu−1 e−η u du. Notice that ν ([0, ∞)) = ∞, which means
drift and a pure jump process [22]. Here, we consider a special there are infinitely many jumps and the corresponding number
pure jump Lévy process, namely the increasing Lévy process of components J in (6) is infinite. Such infinite components
Xt , which has the following explanation: provide an exhaustive description for f (t), but they also result
 
in high computational cost. In fact, it is unnecessary because of
Xt = uN (du, ds) (1) the widely existing noise. Note that since ν ([ε, ∞)) < ∞ for
R + ×[0,t]
any ε > 0, we can approximate the Gamma random measure
where N (du, ds) denotes a Poisson random field on R+ 2
with ν (du) by νε (du):
mean measure E [N (du, ds)] = ν (du, ds), and ν (du, ds) is the νε (du) = γu−1 e−η u 1{u ≥ε} du. (7)
Lévy measure. More generally, for a Borel measurable generat-
ing function ψ (t) , t ∈ T , we can construct a stochastic integral Let Δε denote the sum of jumps of sizes smaller than ε in a
with respect to a random measure X (dt) as follows: Gamma process, we have
    ε
X [ψ] = ψ (t) X (dt) = uψ (t) N (du, dt) (2) E [Δε ] = uγu−1 e−u η du ≤ γε
T R + ×T 0
 ε
where X [ψ] is called a Lévy random field (see [14] for a more γε2
detailed presentation). V [Δε ] = u2 γu−1 e−u η du ≤ (8)
0 2
In the following, we consider a particular type of Lévy process
and Lévy random fields, i.e., the Gamma process and Gamma where V [·] denotes the variance.
CONG et al.: NONPARAMETRIC BAYESIAN ATTRIBUTED SCATTERING CENTER EXTRACTION FOR SYNTHETIC APERTURE RADAR TARGETS 4725

Thus, we can approximate a Gamma random measure arbi- TABLE I

trarily well with arbitrarily small ε, which indicates that f (t) PHYSICAL INTERPRETATION OF PARAMETERS α AND L
can also be approximated adequately with finite number of
components.

B. ASC Model Description

Based on high-frequency approximations [1], the ASC model
approximates radar backscatter by the sum of responses from
individual ASCs. Different from the widely-used point scatter-
ing (PS) model, which assumes that all scattering centers are
isolated points with their responses identical and independent
of frequency and aspect angle, the ASC model incorporates
both frequency and aspect dependence of ASCs, and provides
a concise and physically relevant description of the target radar
signature.
The expression of the ASC model can be written as

J
E (f, φ; Θ) = Ei (f, φ; ϑi ) + n (f, φ) (9)
i=1

where E (f, φ; Θ) denotes the radar backscatter, f the fre- aspect dependence κ is close to zero in practice. Hence, its
quency, φ the aspect angle, Θ the set of all parameters, J the influence can be approximately neglected.
number of ASCs, ϑi the parameter set of the ith ASC, n (f, φ) 2) In practical SAR systems, the relative bandwidth, namely
the complex white Gaussian noise, and Ei (f, φ; ϑi ) the re- the bandwidth to center frequency ratio, is often small,
sponse of the ith ASC, i.e., and the absolute value of the frequency dependence α of
 α i canonical scattering geometries is less than one. Therefore,
f 4π f α makes little difference in the radar backscatter.
Ei (f, φ; ϑi ) = Ai j e−j c (x i cos φ+y i sin φ) Based on the above assumptions, αi and κi are set to zero to
fc
  simplify the ASC model as
2πf  
· sinc Li sin φ − φ̄i e−2π f κ i sin φ 
J
c
E (f, φ; Θ) = Ai k (f, φ; wi ) + n (f, φ)
(10) i=1
 
where ϑi = Ai , αi , xi , yi , Li , φ̄i , κi , sinc [·] = sin [·] / [·] , 4πf
fc is the radar center frequency, c the propagation velocity, k (f, φ; wi ) = exp −j (xi cos φ + yi sin φ)
c
Ai the backscattered coefficient, αi the frequency dependence,  
xi and yi the range and cross-range locations, respectively. The 2πf  
remaining three parameters Li , φ̄i , and κi determine the aspect · sinc Li sin φ − φ̄i (11)
c
dependence of the scattering. For localized ASCs like trihedral,
top hat, sphere, and corner diffraction, Li = φ̄i = 0 and κi char- where wi = xi , yi , Li , φ̄i .
acterizes the small aspect dependence. For distributed ASCs, It seems a little brutal to neglect the influence of α, since it
such as dihedral, cylinder, edge broadside, and edge diffraction, provides important information on scattering geometries. But
κi = 0, Li models the length of the ASC, and φ̄i its orientation on the other hand, a simplified ASC model gives us a clear
angle. understanding of the dominate terms in the ASC model.
The frequency dependence α and the length L distinguish
among several scattering geometries. L determines whether the
ASC is localized or distributed; while α is relevant with the D. A General Perspective: General Harmonic Wavelets
curvature of the scattering primitive [23], [24]. Several canonical
scattering geometries distinguishable by α and L are presented Given the simplified ASC model in (11), we make some
in Table I. variable substitutions as follows:

2f cos φ xi − sin φ̄i
C. ASC Model Analysis Y =− , rci = , Li = Li .
c sin φ yi cos φ̄i
The ASC model in (10) is concise in the electromagnetic scat- (12)
tering background. However, because of the high-dimensional Then, the simplified ASC model can be rewritten as
parameter space, feature extraction based on such a concise
model is still extremely difficult. Therefore, some approxima-   J    
tions are carried out in most existing methods to simplify the E Y ;Θ = Ai K Y ; Wi + n Y
ASC feature extraction problem. By considering actual SAR
i=1
system configurations, the following simplifications are usually
made [25].   ej 2π Y T (r c i + L i /2 ) − ej 2π Y T (r c i −L i /2 )
1) Usually, the imaging time is short, leading to a small span K Y ; Wi = (13)
of the aspect angle. Moreover, the value of the localized j2π Y T Li
4726 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

   
where Wi = rci , Li and the basic functions K Y ; Wi be- then, another one is extracted and added to the parameter set, and
the present parameter set is updated using a RELAX strategy;
long to 2-dimensional General Harmonic Wavelets [26]. This subsequently, the previous step is performed iteratively until
perspective indicates that distributed ASCs are manifested as some stopping criterion (i.e., a desired number of ASCs or a
lines in the parameter space whereas localized ASCs are ex- prescribed residual energy) is satisfied. Three steps are used
pressed as points, which agree with the physical explanation of to extract one ASC from the radar backscatter. Firstly, a set
the ASC model. of dictionary atoms, D(L , φ̄ ) = [d1 , . . . , dN 1 ], are constructed
with the ith atom expressed as
III. RELATED WORK   
2πf  
In general, the ASC feature extraction problem can be seen di = vec sinc Li sin φ − φ̄i (14)
as a specific issue in the dictionary learning field [27]. But, c
unlike the traditional dictionary learning methods where general
forms of dictionary atoms are used [28]–[32], the ASC-based where vec ( · ) denotes the vectorization
 of a matrix. The corre-
dictionary learning benefits from a short and physically relevant sponding parameters Lp , φ̄p are obtained by
parameter description of the atoms, which greatly reduces the    
 
dictionary search space and leads to more accurate or task- Lp , φ̄p = arg max DH ( L , φ̄ )
S (15)
friendly representations. Some methods for the ASC feature L , φ̄
extraction problem are briefly reviewed below. where S = vec (E (f, φ; Θ)) and | · | denotes the modulus op-
eration; secondly, the location parameters (xp , yp ) are got by
A. Image-Domain Methods finding the peak position of the inverse Fourier transform (IFT)
In literature [2], Gerry proposed an image-domain method of dHp S; finally, the backscattered coefficient Ap is acquired
which was developed by transforming the ASC model into im- through the least-squares estimation.
age domain. By taking the advantage that ASCs approximately The computational cost of this method increases rapidly with
decouple in image domain, this method splits the complicated the number of ASCs and the knowledge about the stopping cri-
extraction problem into smaller ones to lower computational teria is not always obtainable in practice. Improper criteria will
cost. However, to complete the transformation, a number of re- lead to undesired results like missing ASCs or extra fake ASCs
strictive assumptions (such as a small relative bandwidth etc.) modeling noise energy. Moreover, considering the problem of
must be made, which are not always practical in real applica- grid mismatch (GM)  [33],the corresponding precision is limited
tions. To tackle this issue, Koets et al. proposed another image- since parameters Lp , φ̄p are got by a parameterized dictionary
domain method [4], which consists of three steps: firstly, the and (xp , yp ) are extracted based on IFT through the iterations.
radar backscatter in frequency-aspect domain is transformed Some other frequency-domain methods take advantage of
into image domain; then, a segmentation process based on a wa- the SR techniques, such as those of [12], [34]–[37]. Most of
tershed algorithm is employed to isolate regions of high energy; these SR methods are based on a sensing matrix, which is a
finally, each isolated region is transformed into frequency-aspect parameterized dictionary under the circumstance of the ASC
domain, where an approximate maximum likelihood (ML) tech- feature extraction problem. To get precise parameter estimates,
nique is applied to get the corresponding ASC parameters. the grid of the parameterization should be as fine as possible.
We refer to this method as SEG-ML for convenience in the However, the smaller the grid spacing, the more coherent the
following. sensing matrix becomes. Such a highly coherent sensing matrix
Although SEG-ML inherits the advantage of the decoupling may violate the incoherence or Restricted Isometry Property
property in image domain, it depends too much on the SAR (RIP) conditions that are required for many SR methods to have
image segmentation process like the other image-domain meth- good performance [37]. To handle highly coherent sensing ma-
ods [2], [9]. Usually, traditional SAR image segmentation tech- trices, Fannjiang et al. proposed an algorithm called BLOOMP
niques are not precise enough to meet the requirement of the [37], which is based on Orthogonal Matching Pursuit (OMP)
ASC feature extraction problem. Specifically, SAR images usu- [34], [35] and is embedded with two techniques, band exclusion
ally have strong fluctuation characteristics caused by the speckle (BE) and local optimization (LO). BE prevents highly coherent
noise and adjacent electromagnetic coupling. Thus, when a seg- atoms of the sensing matrix from being utilized simultaneously,
mentation process is carried out on SAR images, it is highly while LO helps to find a better solution within coherence bands
possible that a distributed ASC will be segmented into several locally.
disconnected regions. Worse still, when the orientation angle Since BLOOMP is based on a parameterized dictionary from
of a distributed ASC exceeds the span of the aspect angle, the a discretization procedure of the physically continuous param-
corresponding image appears to be two separated points [9], eter space, it also suffers from the GM problem like most SR
and the fusion of such two points in practical SAR images is methods. What is more, because of the high-dimensionality of
not easy. the ASC model, a huge dictionary occupying large memory is
necessary for BLOOMP to get precise parameter estimates.
B. Frequency-Domain Methods
To avoid the deficiency of SAR image segmentation, IV. NONPARAMETRIC BAYESIAN ASC EXTRACTION
frequency-domain methods perform feature extraction directly Existing frequency-domain methods accomplish the ASC
in frequency-aspect domain. In this case, the responses of ASCs feature extraction through parameterized dictionaries or sens-
do not decouple with each other, so all parameters must be ing matrices. It means that, to get precise parameter estimation,
estimated simultaneously. the dictionary has to be large enough to cover the support of
Literature [25] proposed a frequency-domain method based the corresponding parameters, which leads to a heavy memory
on the simplified ASC model in (11). In this method, only one burden. Inspired by [14], we alternatively solve this problem by
ASC is firstly assumed and estimated from radar backscatter; developing a nonparametric Bayesian ASC feature extraction
CONG et al.: NONPARAMETRIC BAYESIAN ATTRIBUTED SCATTERING CENTER EXTRACTION FOR SYNTHETIC APERTURE RADAR TARGETS 4727

algorithm. Our algorithm is based on stochastic expansions of which gives the following prior distributions.
continuous dictionaries [38] with Lévy random fields adopted  
to induce prior distributions. J ∼ p (J|γ) = P o J|νε+ , νε+ = γ |Ω| E1 (ε)
u−1 e−u η
u ∼ p (u|η) = 1{u η > ε}
A. Prior Distributions E1 (ε)
In our model, unknown parameters are modeled by random s ∼ p (s) = 1/|Ω|, ϕ ∼ p (ϕ) (22)
variables with priors based mainly on Gamma random fields.
Specifically, the joint prior of the number of ASCs and their com- where P o ( · ) denotes the Poisson distribution, νε+ the Pois-
plex backscattered coefficients is derived by extending Gamma son rate, |Ω|
 ∞ the volume of the space to which s belongs, and
random fields to the complex domain; other variables are given E1 (ε) = ε t−1 e−t dt the exponential integral function. Note
priors designed by the radar operator and these priors are chosen that we have assumed a (noninformative) uniform distribution
as a compromise: physical meaning and calculation feasibility. to describe the range and cross-range locations of the ASCs. ε
1) Priors Based on Gamma Random Fields: Because the is a user-defined parameter. It is introduced to truncate Gamma
backscattered coefficients {Ai } are complex numbers with non- random measure and enable the construction of practical nu-
negative magnitudes and Gamma random fields naturally model merical methods to approximate Gamma random fields within
non-negative jumps in the parameter space, we prefer to adopt specified error bounds, as mentioned in Section II.A. In fact,
Gamma random fields to induce prior distributions for the com- the random measure in (21) may be seen as inducing a sparsity
plex ASC model. As the foregoing background theory is in the penalty to the function F (f, φ; Θ) in (20), similar (or stronger
real domain, we make some substitutions in the following. actually) to the l1 -norm regularization [39]. Parameter γ > 0
First of all, we modify the ASC model as controls νε+ , namely the expected number of ASCs. Because
the mean and variance of the Poisson distribution are equal,
E (f, φ; Θ) = F (f, φ; Θ) + n (f, φ) (16) the distribution for the number of ASCs J is too concentrated
about the mean apriori [40], [41]. To ameliorate this issue, an
where the unknown mean function F (f, φ; Θ) can be additional layer is introduced by placing a conjugate Gamma
written as prior Ga (γ|aγ , bγ ) on γ, leading to the over-dispersed negative

J binomial prior N B (J|r, q) for J, i.e.,
 
F (f, φ; Θ) = ui ψ (f, φ; si , ϕi ) (17) r+J −1
i=1 J ∼ N B (J|r, q) = q r (1 − q)J (23)
J
where ui = |Ai | is the backscattered intensity of the ith ASC, where r = aγ and q = bγ / (bγ + |Ω|E1 (ε)). Given a user-
si = {xi , yi } denotes its location, and defined ε, parameter η > 0 governs the scale of the backscat-
 α i tered intensity u, and hence the range of radar backscatter. Here,
f 4π f
for a better adaptability to different radar backscatter ranges, we
ψ (f, φ; si , ϕi ) = ej ς i e−j c (x i cos φ+y i sin φ)
fc employ an additional inverse Gamma prior on this parameter,
  i.e., η ∼ IG (η|aη , bη ).
2πf   −2π f κ sin φ Parameters γ and η, as well as the truncation constant ε,
· sinc Li sin φ − φ̄i e i
c determine the expected squared truncation error (ESTE), i.e.,
(18) EST E
 
with ϕi = ςi , αi , Li , φ̄i , κi and ςi = ∠Ai + αi π2 . Then, with 
substitutions =E   uψ (f, φ; s, ϕ) N (du, ds, dϕ)
R + ×S ×Φ
Re (F (f, φ; Θ))   2 
F (f, φ; Θ) = 
Im (F (f, φ; Θ)) − uψ (f, φ; s, ϕ) Nε (du, ds, dϕ) 

R + ×S ×Φ F
Re (ψ (f, φ; si , ϕi ))  
ψ (f, φ; si , ϕi ) = , (19)
Im (ψ (f, φ; si , ϕi )) = |u|2 ψ2F γu−1 e−η u 1{η u ≤ε} dudsπ (dϕ)
R + ×S ×Φ
Equation (17) can be equivalently transformed into the real  
domain as = ψ2F γη −2 1 − (1 + ε) e−ε . (24)

J where  · 2F denotes the Frobenius norm and  ψ(f, φ;
F (f, φ; Θ) = ui ψ (f, φ; si , ϕi ) . (20) s, ϕ) 2F is simplified as  ψ 2F .
i=1 Therefore, the hyperparameters ε, aγ , bγ , aη , bη jointly con-
trol ESTE and the prior distributions for J and {ui }.
From (13), it is apparent that the dominant terms in (18) can
2) Priors for Other Parameters: Since the practical
be viewed as 2-dimensional General Harmonic Wavelets, which
backscattered coefficients in SAR images have uniform phases,
have been proven to belong to bounded measurable generating
we place a uniform distribution on ςi , i.e.,
functions in [14]. Therefore, the Gamma random field Γ [ψ] can
be adopted to induce priors. i.i.d.
ςi ∼ p (ς) = U[0,2π ] . (25)
The corresponding Gamma random measure is expressed as
In Table I, it is apparent that the frequency dependence α of
νε (du, ds, dϕ) = γu−1 e−η u 1{η u > ε} dudsπ (dϕ) , (21) canonical scattering geometries has only five possible values.
4728 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

Usually no prior knowledge about the scattering geometries is

available before observation. Thus, we also choose a uniform
prior for α, i.e.,
i.i.d.
αi ∼ p (α) = U{−1− 1 ,0, 1 ,1 } . (26)
2 2

We then consider the prior for the length parameter L. From

Section II.B, we know that there are two kinds of ASCs, local-
ized ASCs with L = 0 and distributed ASCs with L > 0. Since Fig. 1. Schematic representation of our model. Note that ω i =
a single probability density function like a Gamma distribution α i , xi , y i , L i , φ̄ i , κ i .
tends to put too much proportion on distributed ASCs, which
is inconsistent with the actual situation, we introduce the spike-
and-slab prior for L to maintain a balanced proportion between
localized ASCs and distributed ASCs [40]–[42], i.e., Then, we extend the foregoing priors based on Gamma ran-
dom fields to the complex domain and rewrite the equiva-
i.i.d.
Li ∼ p (L) = (1 − ρ) δ0 (L) + ρGa (L|aL , bL ) (27) lent but clearer hierarchical priors for ASC feature extraction
problem as
where ρ ∈ [0, 1] is the mixing coefficient.
From (18), we can see that the amplitude of the sinc function J ∼ p(J|γ) = P o(J|νε+ ), νε+ = γ|Ω|E1 (ε)
decreases as |φ − φ̄i | increases. Moreover, if an ASC has an
orientation angle exceeding the span of the aspect angle, namely γ ∼ p(γ) = Ga(γ|aγ , bγ )
the main lobe of the corresponding sinc function is not observed |A|−1 e−|A |η
i.i.d.
completely, its response is getting weaker. Hence, we design the Ai ∼ p(A|η) = 1{|A |η > ε}
following blended prior for the orientation angle φ̄ to maintain 2πE1 (ε)
some expansibility. η ∼ p(η) = IG(η|aη , bη )
⎧
⎪
2
⎨ (|φ̄ |−D φ / 2 ) 1 1
  −   i.i.d.
ωi ∼ p(ω) = p(L)p(φ̄)p(κ)
, φ̄ > Dφ /2
i.i.d. 2σ 2
φ̄i ∼ p φ̄ = D φ + √1 2π σ φ e φ
5 |Ω|
⎪
⎩  √     
1/ Dφ + 2πσφ , φ̄ ≤ Dφ /2
σn2 ∼ p σn2 , (32)
(28)
where Dφ is the variation of the aspect angle φ. with the corresponding equivalent ESTE being
Since the value of the localized aspect dependence κ is  
extremely small in high-frequency SAR systems, we EST E = g2F γη −2 1 − (1 + ε) e−ε . (33)
 2
 adopt
a narrow
 Gaussian
 prior, i.e., p (κ) = N o κ|0, σ κ , where
N o ·|μ, σ 2 denotes the Gaussian distribution with mean μ B. Posterior Inference
and variance σ 2 . Based on the ASC model in (30) and the assumed complex
Finally, we adopt the non-informative Jeffrey’s prior [43] for white Gaussian noise, the likelihood function is expressed as
the complex white Gaussian noise variance σn2 :   
p E J, {(Ai , ωi )} , σn2
  ⎛
1
p σn2 ∝ 2 . (29)  2 ⎞
σn  
J 
=
1
exp ⎝−  1
E − A i g (ω i )
 ⎠
 (34)
3) Prior Distributions in the Complex Domain: Although (πσn2 )N E σn2  
i=1 F
the foregoing priors are in the real domain, it is not difficult to
extend them to the complex domain. where NE denotes the number of elements in matrix E. Note
Firstly, we rewrite the ASC model as that E (f, φ; Θ) and g (f, φ; ωi ) are simplified as the matrix
form E and g (ωi ) for convenience, respectively.
E (f, φ; Θ) = F (f, φ; Θ) + n (f, φ) For a better understanding of our model, we construct a

J schematic representation in Fig. 1. The joint posterior distri-
= Ai g (f, φ; ωi ) + n (f, φ) (30) bution for all unknown parameters is then expressed as
 
i=1 p J, {(Ai , ωi )} , σn2 , γ, η |E
where ωi = αi , xi , yi , Li , φ̄i , κi is the ith ASC associated   
∝ p E J, {(Ai , ωi )} , σn2 p (J|γ) p (γ)
parameters,
 
 α i · p ({Ai } |J, η) p (η) p ({ωi } |J) p σn2
f 4π f
g (f, φ; ωi ) = e−j c (x i cos φ+y i sin φ)  N E +1 1 
fc γ J e−γ |Ω|E 1 (ε) 1 − 2 E− Ji= 1 A i g(ω i ) 
2

  ∝ e σn F

2πf   J
(10π) J! σn2
· sinc Li sin φ − φ̄i e−2π f κ i sin φ ,
c ! "
J
 
(31) · |Ai |−1 e−|A i |η 1{|A i |η > ε} p (Li ) p φ̄i p (κi )
i=1
and j α i is blended into Ai since it only affects the uniform
distributed phases. · Ga (γ|aγ , bγ ) IG (η|aη , bη ) . (35)
CONG et al.: NONPARAMETRIC BAYESIAN ATTRIBUTED SCATTERING CENTER EXTRACTION FOR SYNTHETIC APERTURE RADAR TARGETS 4729

TABLE II least one of the following two conditions should be satisfied,

MAIN STEPS OF THE RJ-MCMC METHOD i.e.,
    
 x − x T cos φ̄  L  − sin φ̄ T cos φ̄ 
Step 1: Initialize the parameter set {J, {(A i , ω i )}, η } and move probabilities  2 1 2  1  1 2
,
 >
sin φ̄2  2  cos φ̄1 sin φ̄2 
PV , Pb , Pd , Pu ;
Step 2: Iteration t:  y2 − y1
• With probability P V , we implement the following moves to update
    
 x − x T cos φ̄  L  − sin φ̄ T cos φ̄ 
varying-dimensional parameters {(A i , ω i )};
− With probability P b , perform birth move;  2 1 1  2  2 1
.
 >
sin φ̄1  2  cos φ̄2 sin φ̄1 
− With probability P d , perform death move;
− With probability P u , perform update move;  y2 − y1
• With probability 1 − P V , we update the fixed-dimensional parameter η ;
• Sample the nuisance parameters σ n2 , γ using (36);
(38)
Step 3: t = t + 1 and go to Step 2.
Such constraint is embedded in every step of the RJ-MCMC
method to make sure it is fulfilled all the time. Note that without
this constraint, the continuous dictionary will be so flexible that
the RJ-MCMC method gets stuck in a local minimum and results
It is apparent that the so-called nuisance parameters, namely in a large number of mixing iterations.
σn2 and γ, have the following full conditional posterior distribu- 1) Initialization: Generally speaking, there are two ways to
tions, respectively. initialize the parameter set {J, {(Ai , ωi )}}:
⎛  1) Initialize J as zero. Accordingly, coefficients {Ai } and
  2 ⎞
 2    J  ASC associated parameters {ωi } are set to null;
 
p σn |− = IG ⎝σn2 NE , E − Ai g (ωi ) ⎠ 2) Use the results of another simpler ASC feature extrac-
   tion method as the initialization, such as that of the afore-
i=1 F
mentioned image-domain methods or frequency-domain
p (γ|−) = Ga (γ |J + aγ , bγ + |Ω|E1 (ε) ) . (36) methods.
Considering that most existing ASC feature extraction meth-
Based on the research in [44], [45] where a collapsed param- ods are either based on the unreliable SAR image segmentation
eter space is preferred, we tend to integrate out these nuisance or too computational expensive, we prefer the first way to get a
parameters and focus on the marginalized joint posterior: clean start in this paper. The problem of an efficient initialization
p (J, {(Ai , ωi )} , η|E) will be considered in the future.
⎛ Given the constant ε, the parameter η controls the sparsity of
2 ⎞−N E
 J  the whole Bayesian model [14]. Empirically, we find it efficient
 
∝ ⎝E − Ai g (ωi ) ⎠ · IG (η|aη , bη ) to initialize η as a slightly larger value, like 0.6 for all of our
  experiments, to introduce strong sparsity in the beginning.
i=1 F
To focus our main effort on the inference of varying-
Γ (J + aγ ) dimensional parameters {(Ai , ωi )}, we set probability PV to a
· large value, such as 0.95. In accordance with literature [42], we
(bγ + |Ω|E1 (ε))J +a γ (10π)J J! adopt the following strategy for move probabilities Pb , Pd , Pu :
J !   " Pb = cBD · min {1, P r (J + 1) /P r (J)}
· |Ai |−1 e−|A i |η 1{|A i |η > ε} p (Li ) p φ̄i p (κi ) . (37)
i=1 Pd = cBD · min {1, P r (J − 1) /P r (J)}
It is worth noting that the marginalized joint posterior in Pu = 1 − P b − P d (39)
(37) is highly nonlinear and does not have closed form. As
we have fixed-dimensional parameters σn2 , γ, η and varying- where P r (J) is the prior probability of J in (23), and cBD
dimensional parameters {(Ai , ωi )} , i = 1, · · · , J, the dimen- is a constant that tunes the proportion of varying-dimensional
sion of the parameter space is not fixed. Thus, trans-dimensional moves.
MCMC, such as a RJ-MCMC algorithm [15], must be utilized 2) Birth Move: In birth move, a new proposal is drawn ac-
to provide samples for posterior inference. cording to a birth proposal distribution. Then, with probability
min {1, αB }, where αB denotes the Hastings ratio for birth
move, we accept the proposal and add it into the current ex-
C. RJ-MCMC tracted parameter set; otherwise, we reject the proposal and
The RJ-MCMC method contains three types of moves: Birth, keep the current extracted parameter set unchanged.
Death and Update. A birth step involves adding a new ASC to From [15], we know that an appropriate birth proposal distri-
the current parameter set and increasing J by one; a death step bution will dramatically speed up the mixing of the RJ-MCMC
involves removing an ASC from the current parameter set and method. Motivated by [39]–[41], we develop a birth proposal
decreasing J by one; an update step involves selecting an ASC distribution involving the current residual so that our method
and updating its parameters. The main steps of the RJ-MCMC favors regions where a new ASC is more probable.
method are summarized in Table II. Let the current residual denoted by Er = E −
J
Considering practical target size and high range and cross- i=1 Ai g (ωi ). To speed up the mixing, the new pro-
range resolutions of most SAR imaging systems, we make an posal (A∗ , ω ∗ ) should decrease the new residual energy
assumption that every two ASCs in the range/cross-range do-  Er − A∗ g (ω ∗ ) 2F as much as possible. In other words, we
main do not intersect with each other. In other words, we have tend to find a birth proposal distribution satisfying
the following constraint: Let ω1 = α1 , x1 , y1 , L1 , φ̄1 , κ1 and
∗
g(ω ∗ )2F
ω2 = α2 , x2 , y2 , L2 , φ̄2 , κ2 denote any two ASCs, then at qB (A∗ , ω ∗ |Er ) ∝ e−E r −A . (40)
4730 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

However, this distribution is hard to normalize and sample. current residual can be represented by
Thus we make some approximations to simplify the normaliza-
4π f ∗ ∗
tion and sampling. Based on the Bayes’ theorem, we separate Er (f, φ|Θ) = A∗ e−j c (x cos φ+y sin φ)
qB (A∗ , ω ∗ |Er ) as  
2πf ∗  ∗

· sinc L sin φ − φ̄ + n (f, φ)
qB (A∗ , ω ∗ |Er ) = q (α∗ ) q (κ∗ ) q (x∗ , y ∗ |Er ) c
  
K
·q L∗ , φ̄∗ |x∗ , y ∗ , Er + Ai e−j
4π f
c (x i cos φ+y i sin φ)
 
·q A∗ |x∗ , y ∗ , L∗ , φ̄∗ , Er (41) i=1
 
2πf  
· sinc Li sin φ − φ̄i . (45)
where parameters α∗ and κ∗ are sampled from their priors. The c
rest parts are designed as below.
a) q (x∗ , y ∗ |Er ): It is unrealistic to get a precise Given x∗ and y ∗ , we eliminate their influence by
q (x∗ , y ∗ |Er ) because it needs to integrate over all the other  
parameters. However, the traditional SAR imaging based on 4πf ∗
Er (f, φ|Θ) = Er (f, φ|Θ) ◦ exp j (x cos φ + y ∗ sin φ)
the PS model gives us some implications. Here, we use the PS c
model to get a rough yet sufficient approximation as  
2πf ∗  
= A∗ sinc L sin φ − φ̄∗
c
q (x∗ , y ∗ |Er ) ∝ e−E r −A
∗ 2
exp (−j 4 πc f (x ∗ cos φ+y ∗ sin φ) )
. F 
K
((x i −x ∗ ) c o s φ + (y i −y ∗ ) sin φ )
4π f
(42) + Ai e−j c

i= 1
For a clear deduction, let Er = vec(Er ) and B =  
vec(exp(−j 4πc f (x∗ cos φ + y ∗ sin φ))). Then, with least- 2πf  
· sinc Li sin φ − φ̄i + n (f, φ) (46)
squares estimation of the coefficient A∗ , we have c

#  where ◦ denotes the Hadamard product.

 −1 2 $
  The foregoing non-intersect constraint makes sure that terms
∗ ∗ 
q (x , y |Er ) ∝ exp −Er − B H
B B B Er 
H
xi − x∗ and yi − y ∗ in (46) cannot be zeros simultaneously.

2 Hence, only the response of the ASC located at (x∗ , y ∗ ) has
      
1 
constant phases, while the remainings have phases varying with
 H 2  2
= exp B E r exp −Er  . the frequency f . If we sum Er (f, φ) along f in each aspect angle
NE 2 2 φ, approximately, the energy of the ASC located at (x∗ , y ∗ ) will
(43) accumulate, while those of the remaining ASCs will counteract.
Therefore, we adopt the following proposal procedures for L∗
and φ̄∗ : 
Since Er can be seen as constant for now, we further simplify 1) M (φ) = Z1 f Er (f, φ) ≈ sinc( 2πcf c L∗ (φ − φ̄∗ )),
2  
(43) as q (x∗ , y ∗ |Er ) ∝  B H Er 2 . Apparently, this approxi- where assumption sin φ − φ̄ ≈ φ − φ̄ is made thanks
mation is close to 2-dimensional IFT of Er . Therefore, we use to the practical small variation of the aspect angle, and Z
the following strategies to get q (x∗ , y ∗ |Er ): means maximum normalization;
1) Given the current residual Er , we first accomplish 2- 2) Then, we calculate (LT , φ̄T ) = arg minL , φ̄  |M |2 ◦
dimensional IFT to get a discrete image; 2
2) Based on the obtained image, we put a uniform distribu- (|M |2 − |sinc( 2πcf c L(φ − φ̄))| ) 22 ;
tion in each resolution cell to get a continuous distribution 3) Finally, to give a comprehensive search, we sample L∗ and
q (x∗ , y ∗ |Er ). φ̄∗ from N o(L∗ |LT , σL2 ) and N o(φ̄∗ |φ̄T , σφ̄2 ), respectively.
 
Hence, the proposal distribution q (x∗ , y ∗ |Er ) is expressed as c) q A∗ |x∗ , y ∗ , L∗ , φ̄∗ , Er : From [4], we know that
General Harmonic Wavelets are independent to each other as
long as they have no intercross in frequency domain. Therefore,
∗ ∗
[IF T 2 [Er ]]2(x ∗ ,y ∗ ) 1 based on Section II.D and the foregoing non-intersect constraint,
q (x , y |Er ) = (44) we assume the approximate independence
IF T 2 [Er ]2F pr pa  between two ASCs
in the RJ-MCMC method and present q A∗ |x∗ , y ∗ , L∗ , φ̄∗ , Er
as follows. 4π f ∗ ∗
where IF T 2 [Er ] denotes the 2-dimensional IFT of Er . [·](x ∗ ,y ∗ ) 1) Firstly, we perform B = vec(e−j c (x cos φ+y sin φ)
denotes the magnitude of the resolution cell to which (x∗ , y ∗ ) sinc( 2πc f L∗ sin(φ − φ̄∗ )));
belongs. pr and pa are the range and cross-range resolutions, 2) Then, we calculate the least-squares estimation of the co-
−1
respectively. It is worth noting that such proposal distribution is efficient, i.e., AT = (B H B) B H Er ;
easy to compute
 and sample. 3) Finally,  ∗ a complex  Gaussian distribution A∗ ∼
b) q L∗ , φ̄∗ |x∗ , y ∗ , E
 r : Based on the previous obtained CN o A |AT , σA is utilized to complete this pro-
2

x∗ and y ∗ , we design q L∗ , φ̄∗ |x∗ , y ∗ , Er in the following. posal distribution.

Similarly, we would like this distribution to focus on the regions So far, we have completed the hierarchical birth proposal
where the residual energy is large. For simplicity, we use the scheme in (41), with which a new ASC can be calculated ef-
simplified ASC model to design this distribution. Suppose the ficiently. Based on this birth proposal distribution, we get the
CONG et al.: NONPARAMETRIC BAYESIAN ATTRIBUTED SCATTERING CENTER EXTRACTION FOR SYNTHETIC APERTURE RADAR TARGETS 4731

following Hastings ratio αB for birth move, i.e., TABLE III

SIMULATED SYSTEM CONFIGURATIONS
⎛  2 ⎞ N E
   
E − Ji= 1 Ai g (ωi ) p (L∗ ) p φ̄∗ p (κ∗ )
⎜ F ⎟
αB = ⎝ J ∗ 2 ⎠ Center Frequency f c 30 GHz
  10πJ ∗ (bγ + |Ω|E1 (ε))
E − i = 1 A∗i g (ωi∗ ) Bandwidth B 3 GHz
F Variation of Aspect Angle D φ 5.7296◦
∗
Γ (J + aγ ) −(lo g |A ∗ |+ |A ∗ |η ) Pd qD (A∗ , ω ∗ )
· e (47)
Γ (J + aγ ) Pb qB (A∗ , ω ∗ |Er )
where the death proposal distribution qD (A∗ , ω ∗ ) is given in The Hastings ratio αU for update move can be written as
the following death move. ⎛  2 ⎞N E
   
3) Death Move: In death move, we first choose an ASC from E − Ji=1 Ai g (ωi ) p (L∗k ) p φ̄∗k p (κ∗k )
the current extracted parameter set with a death proposal dis- ⎜ F ⎟  
αU = ⎝  J ∗ 2 ⎠
tribution. Then, with probability min {1, αD }, we remove this   p (Lk ) p φ̄k p (κk )
E − i=1 A∗i g (ωi∗ )
ASC; otherwise, we keep the extracted parameter set unchanged. F
Here, we tend to remove the ASCs with small coefficients · exp ((log |Ak | + |Ak | η) − (log |A∗k | + |A∗k | η)) .
[40], [41]. Namely, we adopt the following death proposal dis-
tribution, i.e., (50)

J 5) Update Fixed-Dimensional Parameter η: The fixed-
qD (Ak , ωk ) = |Ak |−1 / |Ai |−1 . (48) dimensional parameter η is updated using the Metropolis Hast-
i=1 ings method, where RW is adopted as the proposal distribution,
i.e.,
Accordingly, the Hastings ratio αD for death move becomes   
⎛   2 ⎞ N E η ∗ ∼ q (η ∗ |η) = N o η ∗ η, ση2 . (51)
 
E − Ji= 1 Agi (ωi ) 10πJ (bγ + |Ω|E1 (ε))
⎜ ⎟   However, the joint posterior distribution in (37) is indistinct
αD = ⎝ J ∗ F2 ⎠
 ∗ ∗  p (Lk ) p φ̄k p (κk ) on η because of its eternal coupling with coefficients {Ai }. To
E − i = 1 Ai g (ωi )
F present a clear expression, we introduce new variables βi =
Γ (J ∗ + aγ ) (lo g |A k |+ |A k |η ) Pb qB (Ak , ωk |E∗r ) Ai η. Subsequently, we rewrite the joint posterior distribution in
· e . (49) (37) to give an explicit invariant distribution on η as
Γ (J + aγ ) Pd qD (Ak , ωk )
4) Update Move: In update move, we randomly sample one p (J, {(βi , ωi )} , η |E )
ASC from the current extracted parameter set and update its pa- ⎛ 2 ⎞−N E
rameters using an update proposal distribution. Subsequently,   J 
 1 
with probability min {1, αU }, we accept the updating; other- ∝ ⎝E − βi g (ωi ) ⎠ · IG (η |aη , bη )
 η i=1 
wise, we keep the extracted parameter set unchanged. F
The update proposal distribution qU ((A∗ , ω ∗ ) | (Ak , ωk )) is
of great importance to the mixing of the RJ-MCMC method. Γ (J + aγ )
·
To maintain a reasonable accept rate, we prefer to update one (bγ + |Ω|E1 (ε))J +a γ (10π)J J!
parameter at each step using the Random Walk (RW) proposal
distribution. Considering that there are two kinds of ASCs: lo- J !   "
calized ASCs and distributed ASCs, which have the physical · η|βi |−1 e−|β i | 1{|β i |> ε} p (Li ) p φ̄i p (κi ) . (52)
meanings of points and lines respectively, we present the fol- i=1
lowing update strategies:
Thus, we get an ad-hoc modified Hastings ratio αU F with the
1) If the chosen ASC is localized, randomly perform one of
following form:
the following RW updates:
a) Update αk∗ ∼ U{−1,0,1/2,1} ⎛  2 ⎞N E
   
b) Update x∗k ∼ N o  x∗k |xk , σx2 E − η1 Ji=1 βi g (ωi ) η∗
J
⎜ F ⎟
αU F = ⎝  J  ⎠
c) Update yk∗ ∼ N o yk∗ |yk , σy2  2 η
  E − η1∗ i=1 βi g (ωi )
d) Update L∗k ∼ N o+ L∗k |Lk , σL2 F
e) Update κ∗k ∼ N o κ∗k |κk , σκ2 
∗
IG (η |aη , bη )
f) Update A∗k ∼ CN o A∗k |Ak , σA2 × . (53)
IG (η |aη , bη )
where N o+ ( · ) denotes the truncated normal distribution
on R+ . D. Utilizing the Bayesian Outputs
2) If the chosen ASC is distributed, randomly perform one of
the following RW updates: Following the procedures in Table II, we run the Markov chain
a) Update αk∗ ∼ U{−1/2,0,1/2,1}
 
for Nr iterations with the first Nb iterations discarded as burn-
b) Update x∗k ∼ N o  x∗k |xk , σx2 in. The remaining samples are collected for posterior inference.
Since the dimensionality may change in different samples and
c) Update yk∗ ∼ N o yk∗ |yk , σy2
  the label switching problem may potentially occur [46], it is
d) Update L∗k ∼ N o+ L∗k |Lk , σL2 improper to directly average the collected samples as the final
e) Update φ̄∗k ∼ N o φ̄∗k |φ̄k , σφ̄2 solution as is conventionally done. Here, following the strategy
  in [47], we select the ML sample among the collected samples
f) Update A∗k ∼ CN o A∗k |Ak , σA2 as our solution for simplicity.
4732 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

TABLE IV
PARAMETER TRUTH AND FINAL PARAMETER ESTIMATES

ASC # Estimate Type Detected Shape Parameters

α x y L φ̄ κ A

1 Truth Dihedral 1 −1.5 1.1 0.2 −0.045 0 92

1 Final (10 dB) Dihedral 1 −1.5003 1.1000 0.1992 −0.0450 0 92.3225
1 Final (−5 dB) Dihedral 1 −1.4997 1.1004 0.1977 −0.0450 0 90.8578
2 Truth Cylinder 0.5 −2.01 2.01 1.998 0.02 0 230.86
2 Final (10 dB) Cylinder 0.5 −2.0098 2.0092 1.9981 0.0200 0 228.3566
2 Final (−5 dB) Edge Broadside 0 −2.0100 2.0010 2.0161 0.0200 0 230.6065
3 Truth Edge Broadside 0 −2.5 1.1 0.3 0.015 0 90
3 Final (10 dB) Edge Broadside 0 −2.4998 1.0998 0.3003 0.0150 0 89.1561
3 Final (−5 dB) Edge Broadside 0 −2.5003 1.1014 0.2988 0.0147 0 85.9096
4 Truth Edge Diffraction −0.5 −2.6 −0.8 1.2 0.009 0 120
4 Final (10 dB) Edge Diffraction −0.5 −2.6000 −0.7999 1.1998 0.0090 0 118.9820
4 Final (−5 dB) Edge Broadside 0 −2.5991 −0.7972 1.2017 0.0091 0 121.5527
5 Truth Trihedral 1 −2.8 0.3 0 0 -3.00E-11 40
5 Final (10 dB) Trihedral 1 −2.8002 0.3000 0 0 −2.95E−11 39.8721
5 Final (−5 dB) Trihedral 1 −2.8000 0.3004 0 0 −2.76E-11 38.3840
6 Truth Top Hat 0.5 −2.1 0.85 0 0 1.00E-11 25
6 Final (10 dB) Top Hat 0.5 −2.1000 0.8501 0 0 1.12E-11 24.7697
6 Final (−5 dB) Sphere 0 −2.1022 0.8491 0 0 3.95E-12 22.1432
7 Truth Sphere 0 −1.9 0.83 0 0 −7.00E-11 26
7 Final (10 dB) Sphere 0 −1.8997 0.8301 0 0 −7.11E-11 25.7460
7 Final (−5 dB) Corner Diffraction −1 −1.8990 0.8296 0 0 −7.20E-11 25.3850
8 Truth Corner Diffraction -1 −1.9 −1.6 0 0 4.00E-11 51
8 Final (10 dB) Corner Diffraction −1 −1.9001 −1.6001 0 0 4.12E-11 50.8510
8 Final (−5 dB) Sphere 0 −1.9000 −1.5995 0 0 3.99E-11 50.1849
9 Truth Trihedral 1 −2.7 −1.4 0 0 5.70E-11 32
9 Final (10 dB) Trihedral 1 −2.7001 −1.4001 0 0 5.70E-11 31.9308
9 Final (−5 dB) Trihedral 1 −2.7008 −1.3999 0 0 5.91E-11 30.9006
10 Truth Sphere 0 −2.1 −0.2 0 0 3.00E-11 20
10 Final (10 dB) Sphere 0 −2.1000 −0.1997 0 0 2.85E-11 19.9619
10 Final (−5 dB) Sphere 0 −2.1015 −0.1992 0 0 2.82E-11 18.1984

E. Complexity domain method SEG-ML and the frequency-domain method

Since the operation of 2-dimensional IFT dominates the BLOOMP as the counterparts in this paper.
We run all experiments on a computer with Intel(R) Xeon (R)
costs of trans-dimensitional moves of our algorithm (i.e., birth
E5-2680 CPU working at 2.7 GHz, 64 GB memory and software
and death moves), the corresponding one-step complexities are
O (NE log NE ) time and O (NE ) space. It is obvious that the MATLAB (R2012a). In all experiments, a 0.6-coherence band
is adopted in BLOOMP and the grid spacing is chosen as fine
one-step complexities for the rest moves are O (NE ) time and
as possible.
O (NE ) space. Accordingly, for an MCMC stream of length
M , the complexity of our algorithm is O (M NE log NE ) time
and O (NE ) space. Note that in practice, the rates of trans- A. Synthetic Backscatter Results
dimensitional moves are limited, and moreover, it is unneces-
sary to update the birth proposal in every trans-dimensitional Firstly, based on the simulated system configurations shown
moves. For example, we could even form the birth proposal in Table III, which gives only 0.1 relative bandwidth and
in the beginning and then fix it through the MCMC stream, in 0.05m range and cross-range resolutions, we synthesize the
which case the complexity of our algorithm drops to O(M NE ) radar backscatter of 10 ASCs, including all the canonical scat-
time and O (NE ) space. Therefore, the practical computational tering geometries listed in Table I. We further add complex
cost of our algorithm is much lower. white Gaussian noise to the synthetic backscatter data, and
choose a noise variance σn2 with 10 dB signal to noise  ratio

(SNR). Here, SNR is defined as SN R = 10log10 PE /σn2 ,
where PE = Es 2F /NE is the pixel-wise average power of the
V. EXPERIMENTAL RESULTS noiseless backscatter Es . The parameter truths of the simulated
This section describes experiments with a variety of simu- ASCs are given in Table IV, and a conventional Fourier image
lated and measured SAR data. Three types of data have been of the synthetic backscatter data, formed by 2-dimensional IFT,
used: firstly, synthetic backscatter data generated directly from is shown in Fig. 2.
the ASC model are used to validate our algorithm when both Based on the Fourier image, we take the location parameters
the data and the noise fit the ASC model; secondly, XPATCH (x, y) ∈ [−3.5, 1] × [−2.5, 4.5], leading to |Ω| = 31.5. We
backscatter prediction data of a backhoe are used to further test choose hyperparameters ρ, aL , bL to achieve a 99% prior on in-
our algorithm; finally, measured spotlight SAR X-band image terval 0 ∪ [0.1, 4] for L to ensure an appropriate support for the
chip of a T72 tank from the Moving and Stationary Target Ac- sinc function and a good compromise between localized ASCs
quisition and Recognition (MSTAR) program are used to test and distributed ASCs. Hyperparameters σφ and σκ are chosen
the practical performance of our algorithm. We take the image- as Dφ /4 and 1/fc , respectively. Our choice of hyperparameters
CONG et al.: NONPARAMETRIC BAYESIAN ATTRIBUTED SCATTERING CENTER EXTRACTION FOR SYNTHETIC APERTURE RADAR TARGETS 4733

Fig. 4. Results at different SNR. (a) Average RSE. (b) Correct rate of α by
Fig. 2. Conventional Fourier image of synthetic backscatter data. our algorithm at different SNR. Note that correct rate is defined as the average
proportion of the correctly extracted parameter.
TABLE V
HYPERPARAMETERS USED IN SIMULATION

ε aη bη aγ bγ ρ aL bL
0.05 10 6 0.8 3.6 0.5 5 7.3

Fig. 5. Backhoe data. (a) Data set. (b) Corresponding Fourier image.

dependence α is given in Fig. 4(b). Since both SEG-ML and

BLOOMP may not always offer the right number of ASCs, we
only present the corresponding results of our algorithm. It is
apparent that the correct rate of α decreases with the decreasing
SNR. The main reason is that the small relative bandwidth leads
Fig. 3. RSE against iteration number of our algorithm at SNR = 10 dB. to different α values making little difference in radar backscatter.
Note that a larger relative bandwidth will result in the estima-
ε, aη , bη , aγ , bγ is guided by three objectives: achieving a 95% tion of α more robust to noise. From Fig. 4, we can see that the
prior predictive interval of [1,100] for J, introducing strong average RSEs of our algorithm are stably less than 0.01 even
sparse property in the beginning, and obtaining a limit on ESTE, with a low correct rate of α, which implies that: (1) given a 0.1
i.e., EST E ≤ 0.05  g 2F . After numerous implementations, relative bandwidth, α does make little influence, which makes it
we find that the choices of these hyperparameters make little even harder to extract; (2) the remaining parameters are robust
influence on the final results as long as they satisfy the above to noise and close to their truths (refer to Table IV where final
objectives. The adopted ones are given in Table V. parameter estimates from one experiment at SNR = −5 dB are
In this simulation, we implement Nr = 6000 samplings, and exhibited). Therefore, we think that, at high SNR, our algorithm
take the first Nb = 3000 iterations as the burn-in period. is able to provide reliable parameter estimation; at low SNR, the
The relative squared error (RSE), defined as RSE = extracted α is not credible, but other parameters, e.g., the length

Es − Ji=1 Ai g (ωi ) 2F / Es 2F , of our algorithm is shown in of the distributed ASC, can serve as reliable features.
Fig. 3. We can see that RSE stabilizes at about 1500 iterations.
This is mainly attributable to the fact that the designed proposal
distributions greatly increase the acceptance ratio of moves. The B. XPATCH Backhoe Results
final ASC features of our algorithm are shown in Table IV for In this section, we conduct experiment to evaluate our al-
a clear comparison. It is obvious that these inferred parameters gorithm using the full public release data set, referred to as
are close to their truths. the Backhoe Data Dome, Version 1.0, provided by DARPA
To further test our algorithm’s robustness against noise, we IXO and AFRL/SNAS. We choose a data set with 512 × 128
re-perform the simulation at SNR = 15 dB, 10 dB, 5 dB, 0 size as shown in Fig. 5(a), in which the frequency varies
dB, and −5 dB, respectively. SEG-ML and BLOOMP are per- from 7.0472 GHz to 12.9528 GHz and the aspect angle from
formed on same data correspondingly. At each SNR, we repeat −4.5714◦ to 4.5000◦ . The corresponding Fourier image is given
the experiments 10 times. The average RSEs of three meth- in Fig. 5(b) for a better understanding of the data.
ods are shown in Fig. 4(a). We can observe that, compared We implement Nr = 10000 samplings of the parameters, and
with SEG-ML and BLOOMP, our algorithm robustly gives the take the first Nb = 5000 iterations as the burn-in period.
lowest average RSEs (i.e., the most accurate reconstructions) Fig. 6 shows the RSE result and posterior histogram for J.
at different SNRs. This is mainly because the nonparametric Here, as Es is unknown, we use E instead to calculate RSE.
Bayesian framework based on Lévy random fields provides ex- The final parameter estimates of the backhoe data are shown in
tra freedom for an adaptive estimation of noise variance. Note Fig. 7, represented by a series of iconic representations of canon-
that the average RSE of our algorithm at SNR = 10 dB is close ical ASCs in Table I. Referring to Fig. 5(b), it is apparent that
to zero, implying that RSE curves as shown in Fig. 3 is stable in our algorithm provides more reasonable parameter estimates.
all Monte-Carlo experiments. The correct rate of the frequency For example, thanks to the ability of adaptively estimating noise
4734 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

TABLE VII
FINAL PARAMETER ESTIMATES FOR THE BACK HOE

Detected Shape Parameters

α x y L φ̄ κ

Edge Broadside 0 2.9222 0.2529 0.2548 −0.0161 0

Edge Broadside 0 2.9224 −0.2543 0.2602 0.0125 0

Fig. 6. XPATCH backhoe results. (a) RSE against iteration number. (b) Pos- TABLE VIII
terior histogram for J (bars) and the prior distribution (solid line). FINAL PARAMETER ESTIMATES FOR THE EXSERTED BARREL

Detected Shape Parameters

α x y L φ̃ κ

Dihedral 1 0.0865 −3.2867 2.9730 −0.0134 0

Fig. 7. Backhoe data results. (a) CAD model. (b) Iconic representation by our
algorithm. (c) Iconic representation by SEG-ML. (d) Iconic representation by
BLOOMP.
TABLE VI
FINAL PARAMETER ESTIMATES FOR THE FRONT SHOVEL

Detected Shape Parameters

α x y L φ̄ κ

Edge Diffraction −0.5 −2.2430 −0.0003 2.1231 0.0000 0

Edge Broadside 0 −2.5032 −0.0101 2.0686 0.0001 0
Fig. 8. MSTAR data results. (a) SAR image of T72 tank. (b) Iconic repre-
sentation by our algorithm. (c) Iconic representation by SEG-ML. (d) Iconic
representation by BLOOMP.
statistics, our algorithm is more robust to weak ASCs than both
SEG-ML and BLOOMP. What is more, the iconic representa-
tion by our algorithm is much “cleaner” than that by BLOOMP of 1.9898 m for the front shovel and fails to provide the esti-
in the region of the front shovel, and this phenomenon may be mate for the back hoe as shown in Fig. 7(c); BLOOMP gives
attributed to three reasons: 1. our algorithm is based on stochas- estimates of 2.1616 m and 0.5639 m for the front shovel and
tic expansions of continuous dictionaries with which the GM the back hoe, respectively. From Fig. 7, we may conclude that,
problem is highly alleviated; 2. prior distributions from Lévy thanks to the nonparametric Bayesian framework based on Lévy
random fields provide a suitable compromise between model random fields, our algorithm is able to provide more reasonable
complexity and data fitting ability; 3. practical constraints like and reliable parameter estimation.
the one in (38) could be easily embedded in our framework. For
a quantitative verification, we exhibit our parameter estimates
for the front shovel and the back hoe in Tables VI and VII, C. MSTAR Target Results
respectively. According to these estimates, the extracted length We further test our algorithm using a measured spotlight mode
of the front shovel is 2.0959 m with its truth being 2.0320 m; SAR image chip of a T-72 tank. The data are collected such that
and the extracted length of the back hoe is 0.7647 m, whereas the barrel of the tank is broadside to the radar. The corresponding
its truth is 0.7107 m. By contrast, SEG-ML gives an estimate image is shown in Fig. 8(a).
CONG et al.: NONPARAMETRIC BAYESIAN ATTRIBUTED SCATTERING CENTER EXTRACTION FOR SYNTHETIC APERTURE RADAR TARGETS 4735

The proposed algorithm models some critical information,

including data likelihood and sparse prior information, into a
joint posterior distribution. Based on the RJ-MCMC inference
method, our algorithm can provide a more reliable sparse so-
lution for the ASC feature extraction problem, and at the same
time, it can automatically and adaptively reach to an appropriate
number of ASCs and residual energy. The actual computational
time is acceptable and the required memory is limited, which
ease the practical implementation of our algorithm.
In practice, the relative bandwidth of SAR systems is usually
small, leading to the frequency dependence parameter sensitive
Fig. 9. MSTAR target results. (a) RSE against iteration number. (b) Posterior to noise. But, this parameter is crucial to the recognition of actual
histogram for J (bars) and the prior distribution (solid line). scattering geometries. To alleviate this issue, other information
must be jointly considered. For example, polarization charac-
teristics are closely related to scattering geometries, which may
In terms of the complexity of the measured SAR data, we im- provide the information to tackle with the noise sensitive issue.
plement Nr = 20000 iterations, and take the first Nb = 15000 We plan to work on a full polarization Bayesian framework as
iterations as the burn-in period. the future work. Thanks to the flexible framework, the proposed
The RSE and posterior histogram for J are shown in Fig. 9. algorithm can be easily extended to other more complicated
Similarly, E is used for the calculation of RSE. Fig. 8 shows the models, such as a 3-dimensional ASC model.
corresponding iconic representation of the parameter estimates
of the T72 tank. Here, because of the strong ground clutter and
noise, reflected in Fig. 8(a), the estimation of the frequency de- ACKNOWLEDGMENT
pendence α is not reliable referring to the previous simulation.
For instance, the extracted scattering geometry of the barrel is The authors would like to thank the anonymous reviewers for
dihedral, whereas its truth is cylinder. Similarly, neither SEG- their constructive suggestions and insightful comments.
ML nor BLOOMP can give correct scattering geometries. But
as mentioned previously, the other parameters of our algorithm,
such as the length of ASCs, are robust to noise, implying that REFERENCES
these parameters can be used as the physically relevant features. [1] J. B. Keller, “Geometrical theory of diffraction,” J. Opt. Soc. Amer.,
Therefore, based on the parameter estimates for the exserted bar- vol. 52, no. 2, pp. 116–130, 1962.
rel shown in Table IX, we get the extracted length equal to 2.9730 [2] M. J. Gerry, “Two-dimensional inverse scattering based on the GTD
m, which is more close to the ground truth of 3.035 m than model,” Ph.D. dissertation, The Ohio State Univ., Columbus, OH, USA,
that of the image-domain method in [4] and BLOOMP, being 1997.
[3] M. J. Gerry, L. C. Potter, I. J. Gupta, and A. Van Der Merwe, “A parametric
0.9301 m and 2.8438 m respectively. Moreover, from Fig. 8(c), model for synthetic aperture radar measurements,” IEEE Trans. Antennas
we can observe that SEG-ML suffers a lot from imprecise SAR Propag., vol. 47, no. 7, pp. 1179–1188, 1999.
image segmentation in such a highly cluttered and noisy sit- [4] M. A. Koets and R. L. Moses, “Feature extraction using attributed scat-
uation. On the other hand, Fig. 8(d) indicates that BLOOMP tering center models on SAR imagery,” in Proc. SPIE, Algorithms for
employs several ASCs jointly in high energy regions for better Synthetic Aperture Radar Imagery VI, vol. 3721, pp. 104–115, Apr. 1999.
data fitting because of the GM problem. It also indicates that [5] K. E. Dungan and L. C. Potter, “Classifying vehicles in wide-angle
radar using pyramid match hashing,” IEEE J. Sel. Topics Signal Process.,
BLOOMP prefers distributed ASCs to “cover” more energy on vol. 5, no. 3, pp. 577–591, 2011.
the body of the tank. This may be attributed to the following [6] K. E. Dungan and L. C. Potter, “Classifying sets of attributed scattering
three facts: 1. high-dimensional parameter space of the ASC centers using a hash coded database,” in Proc. SPIE Defense and Security
model leads to many local minimums; 2. distributed ASCs can Symp., Algorithms for Synthetic Aperture Radar Imagery XVII, vol. 7699,
“cover” more regions than localized ASCs, probably leading pp. 76990Q, Apr. 2010.
to better coherence with the radar backscatter in such a highly [7] E. Ertin and R. L. Moses, “Through-the-wall SAR attributed scattering
center feature estimation,” IEEE Trans. Geosci. Remote Sens., vol. 47,
cluttered and noisy situation; 3. the greedy search strategy may no. 5, pp. 1338–1348, 2009.
end in a bad local minimum. By contrast, it is apparent from [8] Y. He, S.-Y. He, Y.-H. Zhang, G.-J. Wen, D.-F. Yu, and G.-Q. Zhu, “A
Fig. 8(b) that our algorithm can give a more reasonable local forward approach to establish parametric scattering center models for
result in practice. known complex radar targets applied to SAR ATR,” IEEE Trans. Antennas
From the foregoing experiments, we can see that, by con- Propag., vol. 62, no. 12, pp. 6192–6205, 2014.
sidering the joint posterior distribution of all the ASC relevant [9] Y. Akyildiz and R. L. Moses, “Scattering center model for SAR imagery,”
in Proc. SPIE, SAR Image Analysis, Modeling, Techniques II, vol. 3869,
parameters and the noise variance, our algorithm can automati- pp. 76–85, Sep. 1999.
cally and properly balance the target signal and noise, and thus [10] L. C. Potter and R. L. Moses, “Attributed scattering centers for SAR ATR,”
provide a more reasonable and reliable sparse solution. IEEE Trans. Image Process., vol. 6, no. 1, pp. 79–91, 1997.
[11] C. Gilavert, S. Moussaoui, and J. Idier, “Efficient Gaussian sampling for
solving large-scale inverse problems using MCMC,” IEEE Trans. Signal
VI. CONCLUSION Process., vol. 63, no. 1, pp. 70–80, 2015.
[12] L. C. Potter, E. Ertin, J. T. Parker, and M. Cetin, “Sparsity and compressed
In this paper, we develop a nonparametric Bayesian ASC fea- sensing in radar imaging,” Proc. IEEE, vol. 98, no. 6, pp. 1006–1020, 2010.
ture extraction algorithm. Our algorithm employs Lévy random [13] C. Tu, “Bayesian nonparametric modeling using Lévy process priors with
fields to induce prior distributions for the number of ASCs and applications for function estimation, time series modeling and spatio-
temporal modeling,” Ph.D. dissertation, Duke Univ., Durham, NC, USA,
the parameters governing their features. After constructing the 2006.
joint posterior distribution for all unknown parameters, a RJ- [14] R. L. Wolpert, M. A. Clyde, and C. Tu, “Stochastic expansions using
MCMC method is established to enable relatively fast posterior continuous dictionaries: Lévy adaptive regression kernels,” Ann. Statist.,
inference for the whole unknowns. pp. 1916–1962, 2011.
4736 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 64, NO. 18, SEPTEMBER 15, 2016

[15] P. J. Green, “Reversible jump Markov chain Monte Carlo computation and [41] J.-H. Chu, “Bayesian function estimation using overcomplete dictionaries
Bayesian model determination,” Biometrika, vol. 82, no. 4, pp. 711–732, with application in genomics,“ Ph.D. dissertation, Duke Univ., Durham,
1995. NC, USA, 2007.
[16] J. A. Jackson, B. D. Rigling, and R. L. Moses, “Parametric scattering [42] C. Andrieu, N. De Freitas, and A. Doucet, “Robust full Bayesian learning
models for bistatic synthetic aperture radar,” in Proc. IEEE Radar Conf., for radial basis networks,” Neural Comput., vol. 13, no. 10, pp. 2359–2407,
Rome, Italy, pp. 26–30, May 2008. 2001.
[17] J. A. Jackson and R. L. Moses, “3D feature estimation for sparse, nonlinear [43] H. Jeffreys, “An invariant form for the prior probability in estimation
bistatic SAR apertures,” in Proc. IEEE Radar Conf., Washington, DC, problems,” in Proc. Roy. Soc. Lond. A, vol. 186, pp. 453–461, 1946.
USA, pp. 298–303, May 2010. [44] Y. W. Teh, D. Newman, and M. Welling, “A collapsed variational Bayesian
[18] J. A. Jackson, B. D. Rigling, and R. L. Moses, “Canonical scattering inference algorithm for latent Dirichlet allocation,” in Proc. Adv. Neural
feature models for 3D and bistatic SAR,” IEEE Trans. Aerosp. Electron. Inf. Process. Syst., pp. 1353–1360, 2006.
Syst., vol. 46, no. 2, pp. 525–541, 2010. [45] Y. W. Teh, K. Kurihara, and M. Welling, “Collapsed variational inference
[19] J. A. Jackson, “Analytic physical optics solution for bistatic, 3D scattering for HDP,” in Proc. Adv. Neural Inf. Process. Syst., pp. 1481–1488, 2007.
from a dihedral corner reflector,” IEEE Trans. Antennas Propag., vol. 60, [46] A. Jasra, C. C. Holmes, and D. A. Stephens, “Markov chain Monte Carlo
no. 3, pp. 1486–1495, 2012. methods and the label switching problem in Bayesian mixture modeling,”
[20] J. Jackson et al., “Synthetic aperture radar 3D feature extraction for Statist. Sci., pp. 50–67, 2005.
arbitrary flight paths,” IEEE Trans. Aerosp. Electron. Syst., vol. 48, no. 3, [47] B. Chen, G. Polatkan, G. Sapiro, D. Blei, D. Dunson, and L. Carin, “Deep
pp. 2065–2084, 2012. learning with hierarchical convolutional factor analysis,” IEEE Trans.
[21] J. A. Jackson, “Three-dimensional feature models for synthetic aperture Pattern Anal. Mach. Intell., vol. 35, no. 8, pp. 1887–1901, 2013.
radar and experiments in feature extraction,” Ph.D. dissertation, The Ohio
State Univ., Columbus, OH, USA, 2009.
[22] J. Jacod and A. Shiryaev, Limit Theorems for Stochastic Processes. Berlin,
Germany: Springer, 1987. Yulai Cong was born in Shandong Province, China,
[23] M. Plonus, R. Williams, and S. Wang, “Radar cross section of curved 1988. He received the B.S. degree in electronic en-
plates using geometrical and physical diffraction techniques,” IEEE Trans. gineering from Xidian University, Xi’an, China, in
Antennas Propag., vol. 26, no. 3, pp. 488–493, 1978. July 2011. He is currently working toward the Ph.D.
[24] R. A. Ross, “Radar cross section of rectangular flat plates as a function of degree at Xidian University. His research interests
aspect angle,” IEEE Trans. Antennas Propag., vol. 14, no. 3, pp. 329–335, include statistical machine learning and radar auto-
1966. matic target recognition.
[25] J. Duan, L. Zhang, M. Xing, Y. Wu, and M. Wu, “Polarimetric target
decomposition based on attributed scattering center model for synthetic
aperture radar targets,” IEEE Geosci. Remote Sens. Lett., vol. 11, no. 12,
pp. 2095–2099, 2014.
[26] D. E. Newland, “Harmonic and musical wavelets,” in Proc. Roy. Soc.
Lond. A, vol. 444, pp. 605–620, 1994.
[27] I. Tošić and P. Frossard, “Dictionary learning,” IEEE Signal Process. Mag., Bo Chen (M’13) received the B.S. and Ph.D. de-
vol. 28, no. 2, pp. 27–38, 2011. grees in electrical engineering from Xidian Univer-
[28] P. Schmid-Saugeon and A. Zakhor, “Dictionary design for matching pur- sity, Xi’an, China, 2003 and 2008, respectively. His
suit and application to motion-compensated video coding,” IEEE Trans. Ph.D. degree thesis received the honorable mention
Circuits Syst. Video Technol., vol. 14, no. 6, pp. 880–886, 2004. for National Excellent Doctoral Dissertation of P.R.
[29] M. Aharon, M. Elad, and A. Bruckstein, “The KSVD: An algorithm China in 2010. From 2008 to 2013, he was a research
for designing overcomplete dictionaries for sparse representation,” IEEE scientist with the Department of Electrical and Com-
Trans. Signal Process., vol. 54, no. 11, pp. 4311–4322, 2006. puter Engineering, Duke University. From 2014, he
[30] B. A. Olshausen, C. F. Cadieu, and D. K. Warland, “Learning real and com- has been selected to Young Thousand Talents Pro-
plex overcomplete representations from the statistics of natural images,” in gram and worked as a professor at Xidian Univer-
Proc. SPIE 7446, 74460S, 2009. [Online]. Available: http://proceedings. sity. His research interests include statistical machine
spiedigitallibrary.org/proceeding.aspx?articleid=786976 learning, statistical signal processing, and radar automatic target recognition.
[31] Q. Zhang and B. Li, “Discriminative K-SVD for dictionary learning
in face recognition,” in Proc. Comput. Vis. Pattern Recognit. (CVPR),
pp. 2691–2698, 2010.
[32] J. Mairal, F. Bach, and J. Ponce, “Task-driven dictionary learning,” IEEE Hongwei Liu (M’00) received the M.S. and Ph.D.
Trans. Pattern Anal. Mach. Intell., vol. 34, no. 4, pp. 791–804, 2012. degrees, both in electronic engineering, from Xidian
[33] Y. Chi et al., “Sensitivity to basis mismatch in compressed sensing,” IEEE University, Xi’an, China, in 1995 and 1999, respec-
Trans. Signal Process., vol. 59, no. 5, pp. 2182–2195, 2011. tively. He worked at the National Laboratory of Radar
[34] S. S. Chen, D. L. Donoho, and M. A. Saunders, “Atomic decomposition Signal Processing, Xidian University, after that. From
by basis pursuit,” SIAM J. Scientif. Comput., vol. 20, no. 1, pp. 33–61, 2001 to 2002, he was a visiting scholar at the Depart-
1998. ment of Electrical and Computer Engineering, Duke
[35] Y. C. Pati, R. Rezaiifar, and P. S. Krishnaprasad, “Orthogonal matching University, Durham, NC. He is currently a Professor
pursuit: Recursive function approximation with applications to wavelet and the director of the National Laboratory of Radar
decomposition,” in Proc. 27th IEEE Asilomar Conf. Signals, Syst., Com- Signal Processing, Xidian University. His research
put., pp. 40–44, 1993. interests are radar automatic target recognition, radar
[36] D. Needell and J. A. Tropp, “CoSaMP: Iterative signal recovery from signal processing, adaptive signal processing, and cognitive radar.
incomplete and inaccurate samples,” Appl. Comput. Harmon. Anal.,
vol. 26, no. 3, pp. 301–321, 2009.
[37] A. Fannjiang and W. Liao, “Coherence pattern-guided compressive sens-
ing with unresolved grids,” SIAM J. Imag. Sci., vol. 5, no. 1, pp. 179–202, Bo Jiu (M’13) was born in Henan Province, China,
2012. 1982. He received the B.S., M.S., and Ph.D. de-
[38] F. Abramovich, T. Sapatinas, and B. Silverman, “Stochastic expansions grees in electronic engineering from Xidian Univer-
in an overcomplete wavelet dictionary,” Probabil. Theory Relat. Fields, sity, Xi’an, China, in July 2003, March 2006, and
vol. 117, no. 1, pp. 133–144, 2000. June 2009, respectively. He is currently a Professor
[39] M. A. Clyde and R. L. Wolpert, “Nonparametric function estimation using with the National Laboratory of Radar Signal Pro-
overcomplete dictionaries,” Bayesian Statist., vol. 8, pp. 91–114, 2007. cessing, Xidian University. His research interests are
[40] J.-H. Chu, M. A. Clyde, and F. Liang, “Bayesian function estimation radar signal processing, cognitive radar, radar auto-
using continuous wavelet dictionaries,” Statistica Sinica, vol. 19, no. 4, matic target recognition and radar imaging.
pp. 1419–, 2009.