Chinese

Journal of
Aeronautics
Chinese Journal of Aeronautics 23(2010) 556-562 www.elsevier.com/locate/cja

High Range Resolution Profile Automatic Target Recognition

Using Sparse Representation
Zhou Nuo, Chen Wei*
School of Electronics and Information Engineering, Beijing University of Aeronautics and Astronautics, Beijing 100191, China

Received 29 November 2009; accepted 8 April 2010

Abstract
Sparse representation is a new signal analysis method which is receiving increasing attention in recent years. In this article, a
novel scheme solving high range resolution profile automatic target recognition for ground moving targets is proposed. The
sparse representation theory is applied to analyzing the components of high range resolution profiles and sparse coefficients are
used to describe their features. Numerous experiments with the target type number ranging from 2 to 6 have been implemented.
Results show that the proposed scheme not only provides higher recognition preciseness in real time, but also achieves more
robust performance as the target type number increases.

Keywords: automatic target recognition; high range resolution profile; sparse representation; feature extraction; dictionary gen-
eration

1. Introduction1 applications. The method based on HRRP amplitudes

is more popular due to its clear physical interpretation,
High range resolution profile (HRRP) is a one-di- convenient acquisition and high recognition precise-
mensional measurement of target radar reflectivity ness.
along radar line-of-sight (LOS). It provides detailed Recently, as a new method of signal analysis, sparse
structure information of ground moving targets for representation has received a great deal of attention. In
identification. Furthermore, compared with two-di- sparse representation theory, it is considered that sig-
mensional synthetic aperture radar (SAR) images, it nals are mixtures of a wide range of waveforms with
also has the advantage of easy acquisition, fast proc- different frequency properties, time-scale properties or
essing and less storage space taking. Over the recent time-frequency properties. Instead of conventional
years, high range resolution (HRR) mode has been methods which decompose signals by a single type of
used as a complementary mode to ground moving tar- waveforms, sparse representation analyzes signals over
get indicator (GMTI) and SAR imagery for ground a redundant dictionary consisting of multiple types of
moving target recognition[1-3]. waveforms. The waveforms in the dictionary are called
Feature analysis and extraction are significant parts atoms in sparse representation theory. Compared with
of HRR automatic target recognition (ATR). Their the conventional methods, local structures within sig-
missions are to discover and capture the primary sig- nals could be individually adapted by the atoms which
natures of HRRP as well as to delete plenty of unnec- take the most similar properties, and the signal de-
essary information. The obtained features benefit the composition is more concise. Consequently, sparse
improvement of recognition preciseness and running representation provides a more flexible and exact ap-
speed. Several methods such as HRRP amplitudes[3-5], proach for signal analysis. To date, sparse representa-
FFT spectrum analysis[6], S transformation[7], wavelet tion has been successfully applied in signal analysis,
coefficient fitting[8-9] and nonlinear kernel transforma- images or videos compression and reconstruction, ra-
tion[10-11] have been used extensively for HRR ATR dar signal formation, and so on[12]. In addition, because
of its excellent capability in signal analysis and signal
representation, sparse representation has been at-
*Corresponding author.Tel.: +86-10-82317203. tempted to be applied to experiments including face
E-mail address: [email protected] recognition[13-14], human movement recognition[15] and
1000-9361/$ - see front matter © 2010 Elsevier Ltd. All rights reserved. cancer diagnosis[16].
doi: 10.1016/S1000-9361(09)60254-5
No.5 Zhou Nuo et al. / Chinese Journal of Aeronautics 23(2010) 556-562 · 557 ·

In this article, we explore the use of sparse repre- quency properties. Moreover, because the freedom of
sentation into HRRP feature analysis and expect its the selection of s, u and [ is unlimited, the number of
performance in HRRP recognition. gc is infinite. That is, the created dictionary can be as
redundant as it is needed. However, in practical
2. Sparse Representation application, it is sometimes hard to decide the suitable
number of gc for a dictionary. Furthermore, as all at-
2.1. Theory oms in the dictionary derive from the same base func-
tion, the adaptability in signal component representa-
Sparse representation is originally introduced by S. tion is relatively limited.
G. Mallat and Z. F. Zhang in 1993[17], and developed Another category for dictionary generation is a
by D. L. Donoho[18-20] later. It is a method to analyze training sample-based method[14]. The dictionary is
signals under a redundant dictionary. combined with training samples from several target
types. These training samples are collected under a
ª x1 º wide scope of real scenes or their simulations where
«x » N the measured signals, or also called test samples re-
y AX [ A1c A2c " ANc ] « 2 » ¦ Aicxi (1) lated to training samples, probably take place. Due to
« # » i 1
« » the similar collecting conditions, training samples have
¬ xN ¼ similar characteristics with their associated test sam-
ples. The dictionary generated by these training sam-
where the measured signal y=[y1 y2 … yd]TęRd ples is able to represent the test samples more effi-
is a d-dimensional vector, the redundant dictionary ciently and precisely. Additionally, since the real
A [ A1c A2c " ANc ] ęRdhN is a matrix with size of scenes where signals take place are known or predict-
dhN, d˘N, the atom Aic ęA is a d-dimensional vec- able, the atom number in the dictionary can be decided
tor, i=1,2,…,N, the coefficient vector X=[x1 x2 … easily.
xN]TęRN is a N-dimensional vector obtained by y ex-
2.3. Approaches
panding over A. As d˘N, equations y=AX are under
determined. There are infinite X satisfying y=AX. The Finding the sparse representation by the minimum l0
freedom of selection of X facilitates the adaption to norm in Eq.(2) is complex and extremely hard[18].
various practical applications. Several algorithms have been designed to find subop-
In all the possible solutions, X with fewest nonzero timal or approximate solutions instead. There are three
coefficients, X0, is the sparse representation[21]: approaches widely used in sparse representation over
X0 min X 0
subject to y AX (2) redundant dictionaries. They are matching pursuit
X
(MP), basis pursuit (BP) and least absolute shrinkage
In fact, X0 is greatly attractive in actual demands and selection operation (LASSO) [17-23].
since it always means fewer sensors in a radar system MP[17-21], proposed by S. G. Mallat, is a greedy pur-
or less prerequisites for a machine. suit algorithm. It finds the sparse solution by an itera-
tive matching process. Compared with BP and LASSO,
2.2. Dictionary MP has the advantage of easy implementing and short
running time. However, since the solutions produced
Dictionary design is crucial for signal component by MP are just local optimal, they are sometimes not
analysis in sparse representation. Generally, atoms in a precise enough.
dictionary should meet two basic requirements: con- D. L. Donoho[19] simplified problems minimum l0
taining a wide range of types of atoms to better adapt norm into problems minimum l1 norm:
to different local structures of signals, and having suf- y AX
X1 min X subject to (4)
ficient number of atoms to make the dictionary redun- X 1

dant. Furthermore, he gave evidence that X1 in Eq.(4) is

The available methods for dictionary establishment equal to X0 in Eq.(2) when the sparse representation of
can be classified into two categories. One category is y exists. Recently, l1 minimization has been considered
that the dictionary consists of a collection of gc: as an effective and general method for underdeter-
1 t  u  ʌ[ t 2 mined equations. Based on the idea in Eq.(4), D. L.
gc g( )e (3) Donoho proposed BP[18-21] whose solutions are much
s s
sparser and more precise than the solutions MP pro-
where the collection of gc are transformations of a base duces. Nevertheless, computation of BP is still too
function g(t)[17], s, u and [ are dilation, translation and complicated to be used in applications, especially in
frequency modulation parameters, respectively. gc the situation that the dimension of y is large.
produced by different s, u and [ takes various fre- LASSO is also a method dealing with sparse prob-
quency properties, time-scale properties or time-fre- lem based on Eq.(4). But it tackles l1 minimization
· 558 · Zhou Nuo et al. / Chinese Journal of Aeronautics 23(2010) 556-562 No.5

problem by a regression process subject to constraint and the dictionary redundancy.

such as ||X ||1” p. As radar bandwidth increases over the years, radar
2 resolution in range has been greatly improved. Radar
min y  AX 2
subject to || X ||1 d p (5) is no longer a blob detector, but a sophisticated system
X
with the abilities to acquire detailed information for
In the regress process, sparse solution is obtained by
target type recognition[26]. Point scattering models are
optimizing the parameter p value and the coefficients
universally adopted to describe target scattering prop-
at each chosen value of p. The available literature has
erties which are produced by wideband radars[27].
demonstrated that solutions obtained by LASSO are
Assume there are L scattering centers for a given
sensible approximation to y, or even equal to X0, and
target. Distance of each scattering center to the radar
the performance of LASSO is fast, efficient and stable
no matter the scale of dimension of y is large or receiver along radar LOS is ri, i=1,2,…,L. The corre-
small[19-23]. For these merits, LASSO is selected to sponding intensity and initial phase are Vi and \i indi-
solve the sparse representation in this article. vidually. L scattering centers are divided into N con-
tiguous range bins according to ri.
3. HRR ATR Assume further that there are Ln scattering centers in
nth range bin, n=1,2,…,N, L=L1+L2+…+LN. Their
3.1. Data description distances to the radar receiver, intensities and initial
phases are rni, Vni and \ni in sequence, i=1,2,…,Ln.
Due to a lack of real HRRP database obtained from HRRP in nth, hn, is the sum of echoes from Ln scatter-
moving ground vehicles, the raw HRRP used in our ing centers:
experiments are formed from a SAR image database Ln
4S
collected by an X-band HH-polarization SAR sen- hn ¦V ni exp[ j( O rni \ ni )] (6)
sor.[24] This SAR image database contains several i 1

classes of stationary ground vehicles which are no where O is radar wavelength.

more than 3 m in height, 4 m in width and 8 m in It is seen that there is a close relationship between
length. Each SAR image in the database is about 128 point scattering models and HRRP. However, point
pixel×128 pixel with 1 foot by 1 foot resolution. Each scattering models are strongly dependent on target
vehicle corresponds to two groups of SAR images. types and the aspect and depression angles where tar-
These two groups of images are obtained under com- gets stay ahead. For targets with different types, since
parable conditions except that one group of SAR im- the sizes, shapes, materials and scattering properties
ages are collected at 15q depression angles over the are distinct, their scattering properties are highly dif-
360q coverage in aspect while 17q depression angles ferent from each other[28]. Various point scattering
for the other. At the same depression angle for each models are required to represent them individually. For
vehicle, the interval between two neighboring SAR a given target, a relatively large rotation in aspect or
images is less than 3q aspect angles, and the number of depression, about 10q[29], also results in the changes of
SAR images is more than 120. According to the in- target scattering properties. Similarly, a series of point
struction of the database[24], SAR images from 17q scattering models thus are also needed to describe the
depression angles are used as training samples in a scattering properties in continuous angle ranges. In
recognition experiment while SAR images from 15q fact, although the point scattering model is constant for
depression angles are used as test samples. The de- a given target, a small rotation in aspect or depression
tailed transformation process about the raw HRRP equally produces the changes of scattering center dis-
generation from SAR images is explicitly described in tribution in range bins. It is called migration through
Ref. [25]. resolution cells (MTRC). Assume the range resolution
Due to the formed raw HRRP containing amount of is R, the maximum size of the given target in height
background information which mainly expands in sev- and the azimuth plane is no more than Z. If the target
eral first and last range bins and is unnecessary for an rotation GI satisfies
identification task, only the information in the central R 360
GI d u (7)
60 range bins of raw HRRP is kept as the HRRP data Z 2S
used in the following experiment. MTRC can be avoided [29-30]
. Therefore, it is observed
that all of the changes caused by target types, large
3.2. Dictionary generation rotation in aspect or depression and MTRC can cause
the changes of L, ri, Vi, \i and Ln, rni, Vni, \ni, and fi-
Considering the superiorities in decomposition ac- nally result in the variability of HRRP local structures.
curacy and predictability of the dictionary size, dic- According to the discussion above, it is known that
tionary generation based on training samples is at- if a dictionary is combined with training samples from
tempted to be used in this experimental study. In the multiple target types, the diversity of atom types in the
remainder of this section, its feasibility will be dis- dictionary can be satisfied easily. Furthermore, since
cussed from two aspects: the variability of atom types the scale of experimental HRRP is 60 and the number
No.5 Zhou Nuo et al. / Chinese Journal of Aeronautics 23(2010) 556-562 · 559 ·

of training samples for each target is more than 120 as are

described in Section 3.1, the dictionary generated by c mk
the concatenation of training samples of various types
of targets is sufficiently redundant. Meanwhile, the
y AX ¦ Bi X ic | Bk X kc ¦ Akc , j xk , j (8)
i 1 j 1
dictionary size can be correctly predicted by summing
mi
up the number of training samples.
In addition, according to Section 3.1, it can also be ¦ Aic, j xi, j  y , izk (9)
j 1
seen that the maximum size of targets in height and the
azimuth plane is no more than 8 m, and thus the rota- The recognition process can be divided into three
tion GI avoiding MTRC should be no more than 2.18q. steps:
That is, if a rotation in aspect or depression is less than 1)Sum up the arithmetic product of Bi and X ic
2.18q, L, ri, Vi, \i and Ln, rni, Vni, \ni are relatively sta-
mi
ble. The components of HRRP can be viewed as con-
stant. But if the rotation is larger than 2.18q, L, ri, Vi,
sum i Bi X ic ¦ Aic, j xi, j , i  [1, c] (10)
j 1
\i and Ln, rni, Vni, \ni will vary. For the training or test
samples of a given target collected at a fixed depres- 2)Find the minimum distance d between y and sumi
sion angle, since the angle increment in aspect between min d min y  sum i , i  [1, c ] (11)
contiguous samples is between 1q and 3q, the local
structures of samples vary following the change of 3)Recognize the test sample to the type of targets
aspect angles over the 360q aspect coverage. However, which the minimum d corresponding to.
the training samples and the test samples gained at the
same aspect angle consist of similar local structures in 4. Results and Discussion
that the difference between them is just 2q in depres-
sion. Therefore, atoms combined with training samples 4.1. Sparse decomposition
of multiple target types not only meet the two basic
acquirements of dictionary generation, but also take Numerous test samples have been experimented to
superiorities to adapt the local structures of individual validate deduction in Section 3.2 about the feasibility
test samples more flexibly and exactly. of dictionary generation based on training samples and
the assumption in Section 3.3 that the decomposition
3.3. Feature sparsity results are sparse. The results show that the obtained
decomposition results are sparse indeed and primary
According to Section 3.2, it is observed that training sparse coefficients assemble within the range of the
samples are highly similar to local structures with their corresponded target types.
corresponding test samples. If a test sample is given Ranges at [1, 135], [136, 283], [284, 421], [422,
from one of the target types, the ideal signal compo- 658], [659, 795], and [796, 1 032] are related to atoms
nent analysis in theory should approximately locate on belonging to 1st, 2nd, 3rd, 4th, 5th, and 6th target
the atoms related to this target type, especially the ones types, respectively. As examples, Fig.1 shows LASSO
obtained at the same aspect angle or around it. There- decomposition results of test samples which belong to
fore, the decomposition result is sparse. It will be util- 1st, 2nd and 3rd type of targets around 0q, 90q, 180q
ized as features in recognition experiments.

3.4. Recognition

Refer to the definition in Eq.(1) and assume that

there are c types of targets T1,T2,…,Tc. The dictionary
is A=[B1 B2 … Bc], Bi [ Aic,1 Aic,2 Aic,m ] , i=1,2,…,c,
i

mi is the number of training samples associated with

ith target type. The resolved sparse coefficient vector
is X i [ X1c X 2c " X cc ]T ęRM, X ic =[xi,1 xi,2 …
xi ,m ]T , M=m1+ m2+…+mc. y is a test sample from kth
i

target type, kę[1,c].

According to the analysis in Section 3.2 and Section
3.3, the nonzero elements should mainly locate on
X kc =[xk,1 xk,2 … xk ,mk ]T , and few locate on X ic =
[xi,1 xi,2 … xi ,mi ]T , ię[1,c], iĮk. Therefore, there
· 560 · Zhou Nuo et al. / Chinese Journal of Aeronautics 23(2010) 556-562 No.5

ranges from 2 to 6 sequentially.

In order to make the results of SRBR and TBR
comparable, test samples used in both approaches are
produced as described in Section 3.1; templates em-
ployed in TBR and the dictionary for sparse represen-
tation are combined by training samples formed in
Section 3.1 as well. Once the number of target types in
a sort of experiments is fixed, numerous recognition
experiments are implemented, and the target types in
each experiment are selected randomly from the data-
base; the probability of correct classification (PCC) of
each sort of experiments is measured by the statistical
mean of the results obtained in the corresponding
experiments.
Fig.2 depicts the performance of TBR and SRBR
during 2 to 6 recognition experiments. It is seen that,
when the target type number is fixed, SRBR keeps
higher PCC than TBR does. The improvement in 2 to 6
recognition experiments is 2.371 8%, 3.682 8%,
4.680 1%, 5.368 4% and 5.981 7%, respectively. It can
also be observed that as the number of target types
increases, both of PCC in SRBR and TBR drop. The
rate of PCC declining in SRBR is about 14.809 6%,
from 93.508 2% to 78.698 6% during 2 to 6 recogni-
tion experiments, while that in TBR is more than 18%,
from 91.136 4% to 72.716 9%.

Fig.1 Sparse coefficients obtained by sparse analysis. Fig.2 Performance comparison of SRBR and TBR in 2 to 6
recognition experiments.
and 270q aspect angles. It is seen that there are few
Experiments are implemented using MATLAB on a
sparse coefficients appearing in Fig.1. Significant
PC with 2.8 GHz CPU. The average running time for
sparse coefficients concentrate on the atoms related to
SRBR and TBR is 4.7 s and 2.7 s, respectively.
the type of test samples while sparse coefficients of
atoms corresponding to other target types are relatively 4.3. Discussion
much smaller, or even equal to zero.
From Fig.2, it is known that both of TBR and SRBR
4.2. Recognition comparison provide satisfied recognition results. However, SRBR
is superior to TBR throughout 2 to 6 recognition ex-
Two approaches are considered in recognition com- periments. The improvement in SRBR primarily bene-
parison. One is the currently widely used template fits from the insight of sparse representation into signal
based recognition (TBR) using HRRP magnitudes[3-5]. decomposition. For example, Fig.3 contains a template
The other is the sparse representation based recogni- matching result that a test sample of 2nd target type at
tion (SRBR) employing sparse coefficients proposed 14q aspect angle. The maximum correlation coefficient
in this article. Both approaches are implemented in 5 appears at the range between 659th template and 795th
sorts of recognition experiments. The target type template which corresponds to 5th target type. Hence,
number in these 5 sorts of recognition experiments the test sample is judged as the 5th target type accord-
No.5 Zhou Nuo et al. / Chinese Journal of Aeronautics 23(2010) 556-562 · 561 ·

ingly. Obviously, this recognition result is wrong. The

main reason of the mistake is that the method based on
template matching emphasizes on the rough similarity
in general between two samples, but ignores the dif-
ference in local structures in detail. The judgment
lacks accuracy.

Fig.4 Process of sparse analysis for the test sample from

2nd target type at 14q aspect angle.
Fig.3 Template matching result of a test sample from 2nd
target type at 14q aspect angle.
5. Conclusions
In contrast, sparse representation, like a microscope,
provides an insight into the internal local structure (1) For the experimental data with less dimension
composition. During the signal approximation process, but large number of training samples, dictionary gen-
sparse analysis keeps adding atoms adapted to signal eration based on training samples is a simple but effi-
local structures, evaluating their significance and re- cient means.
moving atoms with less similarity. Fig.4 shows the (2) Sparse representation provides a more accurate
approximation process of the test sample of 2nd target way for HRRP analysis. The decomposition results are
type at 14q aspect angle in sparse representation. sparse and exact.
Similarly with template matching, at the beginning, (3) Compared with TBR, SRBR produces an obvi-
sparse analysis firstly fixes the atom related to 5th tar- ous improvement in recognition precision when the
get type as the original and rough analysis result, as target types range from 2 to 6. Furthermore, perform-
shown in Fig.4(a). During the successive process, at- ance obtained by SRBR is more robust and stable.
oms are added or deleted according to their signifi- (4) Compared to TBR, the running time of SRBR is
cance at each step of approximation, shown as little longer than TBR. But the difference is within
Figs.4(b)-(g). The final result is a perfect approxima- several seconds, so it is still able to meet the demands
tion of the test sample with atoms as few as possible in real applications.
(see Fig.4(h)).
References

[1] Westerkamp J J, Williams R L, Palomino A P, et al.

Synthetic moving target HRR profile generation using
measured and modeled target data. Proceedings of
SPIE Algorithms for Synthetic Aperture Radar Imagery
VII. 2000; 4053: 631-641.
[2] Blasch E P. Modeling intent for a target tracking and
identification scenario. Proceedings of SPIE Signal and
Data Processing of Small Targets. 2004; 5428: 260-
269.
[3] Nguyen D H, Kay J H, Orchard B J, et al. Classifica-
tion and tracking of moving ground vehicles. Lincoln
Laboratory Journal 2002; 13(2): 275-308.
[4] Cong S, Hong L, Blasch E. Performance model for
joint tracking and ATR with HRR radar. Proceedings of
SPIE Algorithms for Synthetic Aperture Radar Imagery
XV. 2008; 6970: 1-12.
[5] Turnbaugh M A, Bauer K W, Oxley M E, et al. HRR
· 562 · Zhou Nuo et al. / Chinese Journal of Aeronautics 23(2010) 556-562 No.5

signature classification using syntactic pattern recogni- [20] Donoho D L. For most large underdetermined systems
tion. IEEE Aerospace Conference. 2008: 1-9. of linear equations the minimal 1-norm solution is also
[6] Liu J, Zhang J Y, Zhao F. A new time-shift invariant the sparsest solution. Communications on Pure and
feature of radar HRRPs. Journal of electronics & In- Applied Mathematics 2006; 59(6): 797-829.
formation Technology 2008; 30(8): 1949-1953. [in [21] Yang J Y, Peng Y G, Xu W L, et al. Ways to sparse
Chinese] representation: an overview. Science in China Series F:
[7] Wang D W, Li Y J, Zhang K. Time-frequency signature Information Sciences 2009; 52(4): 695-703.
extraction and recognition of target based on [22] Amaldi E, Kann V. On the approximability of mini-
S-transform. Acta Aeronautica et Astronautica Sinica mizing nonzero variables or unsatisfied relations in
2009; 30(2): 305-309. [in Chinese] linear systems. Theoretical Computer Science 1998;
[8] Morris H C, de Pass M M. Wavelet feature extraction 209(1-2): 237-260.
of HRR radar profiles using generalized Gaussian dis- [23] Osborne M R, Presnell B, Turlach B A. A new ap-
tributions for automatic target recognition. Proceedings proach to variable selection in least squares problems.
of SPIE Signal Processing, Sensor Fusion, and Target IMA Journal of Numerical Analysis 2000; 20(3):
Recognition XIV. 2005; 5809: 165-175. 389-403.
[9] Liu H W, Yang Z, He K, et al. Radar high range resolu- [24] Ross T D, Worrell S W, Velten V J, et al. Standard SAR
tion profiles recognition based on wavelet packet and ATR evaluation experiments using the MSTAR public
subband fusion. IEEE International Conference on release data set. Proceedings of SPIE Algorithms for
Acoustics, Speech, and Signal Processing. 2005; 5: Synthetic Aperture Radar Imagery V. 1998; 3370:
445-448. 566-573.
[10] Chen B, Liu H W, Bao Z. An efficient kernel optimiza- [25] Gross D, Oppenheimer M, Kaler B, et al. Preliminary
tion method for high range resolution profile recogni- comparison of high range resolution signatures of
tion. Radar International Conference. 2006: 1-4. moving and stationary ground vehicles. Proceedings of
[11] Liu H W, Su H T, Bao Z. Radar high range resolution SPIE Algorithms for Synthetic Aperture Radar Imagery
profiles feature extraction based on kernel PCA and IX. 2002; 4727: 205-212.
kernel ICA. Advances in Neural Networks 2005; 3496: [26] Skolnik M l. Introduction to radar systems. 3rd ed.
913-918. Beijing ˖ Publishing House of Electronics Industry,
[12] Wang J Y, Yin Z K, Zhang C M. Sparse analysis and 2007.
preliminary application in signal and image. Sichuan: [27] Huang P K, Yin H C, Xu X J. Characteristics of radar
Southwest Jiaotong University Press, 2006. [in Chi- targets. Beijing˖Publishing House of Electronics In-
nese] dustry, 2005. [in Chinese]
[13] Wright J, Ganesh A, Zhou Z H, et al. Demo: robust [28] Shirman Y D. Computer simulation of aerial target
face recognition via sparse representation. IEEE Inter- radar scattering, recognition, detection, and tracking.
national Conference on Automatic Face and Gesture Boston: Artech House, 2002.
Recognition. 2008: 1-2. [29] Du L. Study on radar HRRP target recognition. PhD
[14] Wright J, Yang A Y, Ganesh A, et al. Robust face rec- thesis, Xi’an:Xidian University, 2007. [in Chinese]
ognition via sparse representation. IEEE Transactions [30] Li H J, Yang S H. Using range profiles as feature vec-
on Pattern Analysis and Machine Intelligence 2009; tors to identify aerospace objects. IEEE Transactions
31(2): 210-227. on Antennas and Propagation 1993; 41(3): 261-268.
[15] Gkalelis N, Tefas A, Pitas I. Sparse human movement
representation and recognition. IEEE 10th Workshop
Biographies:
Multimedia Signal Processing. 2008: 165-169.
[16] Hang X Y. Cancer classification by sparse representa-
tion using microarray gene expression data. IEEE In- Zhou Nuo Born in 1981, she is currently a Ph.D. candidate
ternational Conference on Bioinformatics and in Beijing University of Aeronautics and Astronautics. Her
Biomeidcine Workshops. 2008: 174-177. research interests include signal processing and automatic
[17] Mallat S G, Zhang Z F. Matching pursuits with target recognition.
time-frequency dictionaries. IEEE Transactions on Sig- E-mail: [email protected]
nal Processing 1993; 41(12): 3397-3415.
[18] Chen S S, Donoho D L, Saunders M A. Atomic de- Chen Wei Born in 1963ˈhe is an associate professor in
composition by basis pursuit. SIAM review 2001; Beijing University of Aeronautics and Astronautics. His
43(1): 129-159. research interests include signal and information processing,
[19] Donoho D L, Tsaig Y. Fast solution of l1-norm information fusion, digital image processing and target rec-
minimization problems when the solution may ognition.
be sparse. <http://www-dsp.rice.edu/files/cs/FastL1. E-mail: [email protected]
pdf >. 2006.