afastandefficientclutter rejection algorithm for sar imageryash/murthy_aero.pdf · speckle noise...
TRANSCRIPT
A Fast and Efficient Clutter
Rejection Algorithm for SAR
Imagery
K. RAMACHANDRA MURTHY
SAMBHUNATH BISWAS
ASHISH GHOSH, Member, IEEE
Indian Statistical Institute
A clutter rejection scheme is proposed for synthetic aperture
radar (SAR) imagery based on two-stage two-dimensional
principal component analysis (two-stage 2DPCA) followed by
a bipolar eigenspace separation transformation (BEST) and a
multilayer perceptron (MLP). For this, we have examined and
analyzed four different algorithms. They are based on principal
component analysis (PCA) both in one dimension (conventional
PCA) and in two dimension with three different forms (2DPCA,
alternative 2DPCA and two-stage 2DPCA), followed by BEST
and MLP in each case. Feature extraction in different cases is
carried out using respective PCA scheme. Each algorithm uses
the BEST to further reduce dimensionality and enhance the
generalization capability of the classifier. Classification between
target chips and clutter chips is finally made through an MLP
classifier. Experimental results on MSTAR public release database
of SAR imagery are presented. Comparison of all the 2DPCA
algorithms with an existing technique shows improvement both
in performance and time. Moreover, all the 2DPCA algorithms
compute eigenvectors and eigenvalues of image covariance
matrix accurately and very efficiently with respect to time, and
further reduces the effect of noise better than that in the case of
PCA. Comparison reveals the fact that two-stage 2DPCA-based
algorithm is the best (both in performance and time) between all
algorithms.
Manuscript received December 6, 2010; revised November 15,
2011, August 9, 2012, and January 2, 2013; released for publication
January 30, 2013.
IEEE Log No. T-AES/49/4/944709.
Refereeing of this contribution was handled by P. Lombardo.
This work was funded by the U.S. Army through the project
“Processing and Analysis of Aircraft Images with Machine
Learning Techniques for Locating Objects of Interest,” Contract
FA5209-08-P-0241.
Authors’ address: Machine Intelligence Unit, Indian Statistical
Institute, 203 B. T. Road, Kolkata, West Bengal 700108, India,
E-mail: ([email protected]).
0018-9251/13/$26.00 c° 2013 IEEE
I. INTRODUCTION
Military surveillance systems rely on a varied
number of sensor information to locate, identify, and
track oppositional forces. This task becomes more
difficult for a large number of potential targets when
they are found over large land areas. An automatic
target recognition (ATR) [4] system appears as one
of the solutions [5—8] for this problem. A typical
ATR system is as shown in Fig. 1. One of the main
sensors used in ATR systems is the synthetic aperture
radar (SAR) [9, 10]. The SAR is an airborne or
orbital active sensor, most of which operates in
the microwave frequency range. This characteristic
plays a significant role in providing substantial
advantages in SAR, carried all day and in all weather.
However, SAR images suffer from reflections from
the background and other objects. Echoes from the
surface exhibit a noise-like behavior that might hinder
target detection and/or introduce false alarms. These
clutter returns can be moduled as the superposition
of a slowly varying, spatially correlated component
(texture) and multiplicative noise (speckle), as
illustrated in [10].
Fig. 1. Block diagram of three-stage SAR ATR system. Input
consists of SAR imagery representing many square kilometres of
terrain and potentially containing several targets of interest; output
consists of locations and classification labels for these targets.
This article proposes improved version of clutter rejection stage.
The target detection module is certainly one of
the most important components in a SAR-based
ATR system, as the whole system may not function
properly without an excellent detector [11, 12]. A
common problem of the target detection algorithm is
false alarms (clutter), which become more prominent
when high accuracy in target detection is sought,
causing computational burden and degradation in
performance of SAR ATR [13—15]. Potential targets
identified in the detection stage are passed to a clutter
rejection stage, which is intended to reject further
false alarms based on some simple properties of
the potential target, including both geometrical and
electromagnetic effects [10]. Most false alarms are
derived from man-made objects, though some may
arise from natural clutter [10]. This necessitates the
use of an effective clutter rejection scheme. Unless
the excessive false alarms are reduced substantially,
the subsequent classifier (target recognizer) in the
SAR-based ATR system may not perform well.
The existing automatic clutter rejection technique
in [16] used a bipolar eigenspace separation
transformation (BEST) for feature extraction and a
IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 49, NO. 4 OCTOBER 2013 2431
multilayer perceptron (MLP) for clutter rejection in
forward looking infrared (FLIR) images. Eigenvectors
are computed from the normalized, clipped, and
scaled training data for a subspace projection using
the BEST method. The resulting projected values
are then fed to an MLP for training to perform a
two-class classification. Thus, in their experiment
[16] computation needed to carry out the BEST is
more than that required for the subsequent MLP-based
classifier. Different eigenspace transforms are also
used in the field of clutter rejection and target
detection [17—20].
As clutters have similar signature and
characteristics to those of targets, the decision
boundary between these two classes will be highly
nonlinear and hence it will be much more difficult to
approximate the boundary. To simplify the decision
boundary and the further reduction of dimensionality,
one therefore needs a subspace projection on the
input data. Preprocessing of the input data with
eigenspace separation transform (EST) [2] allows
reduction of the size of a neural network while its
generalization accuracy is enhanced as a binary
classifier [2]. In the case of image data, EST may
not drastically decrease the size of the concerned
neural classifier and hence, the time complexity
may not also be considerably reduced. To get rid of
these difficulties, we have used BEST having equal
number of dominant eigenvectors corresponding
to both positive and negative eigenvalues [16].
But for SAR images, due to strong correlation and
multiplicative speckle noise, behavior of BEST is
unpredictable, i.e., it might generalize or degrade the
performance of a subsequent classifier. Moreover,
correlation difference matrix and its eigenvectors in
BEST may not be evaluated accurately because of
huge dimensional input space. In our experiments
BEST alone fails in generalizing the capability of
the classifier; whereas, by reducing the effect of
speckle noise and decorrelating the feature space,
generalization improves a lot. Therefore, one can
use a feature extraction technique like principal
components analysis (PCA) to reduce the effect of
speckle noise.
PCA appears to have one fundamental use in SAR image
series i.e., speckle noise reduction. From this perspective some
significant advantages accrue: (i) identification of dominant
spatio-temporal modes of backscattering within the scene; (ii)
partitioning of the variance into a set of discrete images that
can be submitted to segmentation and classification algorithms;
and (iii) production of a high spatial resolution, low spatial
noisy image that can serve as a template for georeferencing and
multi-sensor fusion [21].
We have used two-dimensional PCA (2DPCA)
[1] for this purpose because the classical 1D-PCA
always transforms a 2D image matrix into a 1D image
vector. This usually leads to a high-dimensional vector
space, and hence it becomes difficult to evaluate
the covariance matrix accurately. In such cases the
eigenvectors can be calculated efficiently using
the singular value decomposition (SVD) [22, 23]
technique, where the process of generating the
covariance matrix is actually avoided. However, this
does not imply that the eigenvectors can be evaluated
accurately in this way, since the eigenvectors are
determined by solving a huge set of linear equations
with a huge number of unknowns [24]. Also for a
high-dimensional matrix, SVD may be very expensive.
This problem has been reflected in our experimental
results also. 2DPCA avoids this problem and extracts
features from the image matrix directly. Furthermore,
it has been reported in [1] that 2DPCA constructs the
covariance matrix directly from 2D image matrices
rather than 1D vectors, and evaluates the covariance
matrix more accurately. Even more, in this case the
size of the covariance matrix is much smaller. But one
of the drawbacks of 2DPCA is that it only eliminates
the correlations between rows, necessitating more
features, leading to large storage requirements and
more cost (in time) for classification. One can easily
note the equivalence of 2DPCA to line-based PCA
as proved in [25]. Alternative 2DPCA (A2DPCA)
has been proposed in [26], which is also poor in
reducing dimensionality, because it works in the
column direction only. However, two-stage 2DPCA
eliminates correlation between rows as well as
columns achieving more reduction in dimensionality
and has been used for efficient face recognition
and representation [26]. Since a SAR image has
texture correlation in both directions (row and column
wise), one can think about the novel use of two-stage
2DPCA in the area of target/clutter recognition.
Feature extraction as well as recognition based
on subspace projection, might also be an option,
such as Fisher’s linear discriminant analysis (LDA)
technique [27].
In this article we propose three cascaded subspace
projection schemes, i.e., 2DPCA, A2DPCA, and
two-stage 2DPCA along with BEST. Finally, based
on experimental results, we have concluded that the
cascading of two-stage 2DPCA along with BEST
is superior to the other two cascades as well as the
method in [16]. The projected data in the subspace
are used to classify actual targets from clutters
through the well-known MLP classifier [3, 28]. To
compare with [16], we have implemented PCA,
2DPCA, A2DPCA, and two-stage 2DPCA along with
BEST and MLP in each case. The method based on
two-stage 2DPCA is seen to be the best with respect
to classification accuracy and time.
The rest of the paper is organized as follows.
Section II describes the three algorithms that we have
analyzed to select the proposed method. Section III
presents the experimental results, comparisons, and
discussions. Finally conclusion is made in Section IV.
2432 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 49, NO. 4 OCTOBER 2013
Fig. 2. Block diagram of proposed algorithms.
II. DESCRIPTION OF THE PROPOSED ALGORITHMS
In this section we describe the proposed technique.
Initially, we have examined and analyzed four
different new algorithms and depending on their
performance and merits, we have chosen the best one.
These algorithms are based on 2DPCA, A2DPCA, and
two-stage 2DPCA. BEST and MLP have been used in
conjunction with each algorithm. The 2DPCA, in fact,
uses 2D matrices against the standard PCA, based
on 1D vectors, to ensure requirement of less time
for computation and better features for classification
[1]. As BEST compresses data that can be suitably
used to improve the performance of a binary classifier,
we have incorporated this in our algorithm. In one
of the cases we have implemented a 1D PCA-based
algorithm for comparison purpose. We briefly describe
all 2DPCA-based algorithms in the following sections.
The block diagram of the proposed algorithms can be
found in Fig. 2.
A. Algorithm 1
Let us first briefly describe Algorithm 1 based
on 2DPCA. Let A be an m£ n random image matrix
and X 2 Rn£d a matrix with orthonormal columns,n¸ d. Projecting A onto X, we can write the projectedmatrix Y = AX, Y 2 Rm£d. The total scatter J(X) isthe trace of the covariance matrix of the projected
samples, i.e.,
J(X) = trfE[(Y¡E[Y])(Y¡E[Y])T]g= trfE[(AX ¡E[AX])(AX ¡E[AX])T]g= trfXTE[(A¡E[A])T(A¡E[A])]Xg: (1)
Here E[:] denotes the expected value and T denotesthe transpose. The image covariance matrix G =E[(A¡E[A])T(A¡E[A])] is an n£ n nonnegativedefinite matrix. For M training chip images (potential
target chips, i.e., true target and clutter chips), Ak(k = 1,2, : : : ,M), each of size m£ n, the averageimage is
A=1
M
Xk
Ak:
Hence, G can be evaluated as
G =1
M
MXk=1
(Ak ¡ A)T(Ak ¡ A): (2)
Therefore,
J(X) = tr(XTGX): (3)
Maximization of J(X) is equivalent to minimization ofcorrelation between the features; and the maximum
value of J(X) can be obtained from Xopt, whichconsists of the orthonormal eigenvectors X1, : : : ,Xdof G corresponding to the d largest eigenvalues, i.e.,Xopt = [X1, : : : ,Xd]. As the size of G is only n£n,computation of its eigenvectors is very easy. Also, like
in PCA the value of d can be controlled by setting athreshold μ, according to (4)Pd
i=1¸iPni=1¸i
¸ μ: (4)
Note that ¸1,¸2, : : : ,¸n are the n dominant eigenvaluesof G. Algorithm 1 is the cascading of 2DPCA along
with BEST followed by an MLP. Schematically,
Algorithm 1 is described in Fig. 2 and its time
complexity is O(nm2 +M2dm), where O(nm2) isthe time complexity of 2DPCA and M2dm is thetime complexity of BEST, which is a biquadratic
time complexity. It is straightforward to observe that
m and n are fixed and M2À nm2, therefore nm2 isnegligible. PCA- and BEST-based algorithms are also
biquadratic time algorithms with time complexity
of O(M2nm)). One main advantage of Algorithm 1
is that it is efficient for both noise reduction and
simplification of decision boundary. These two criteria
are responsible to fasten the convergence of MLP. On
the other hand, as PCA does not preserve the class
discrimination information and the BEST (existing)
based algorithm does not reduce speckle noise, both
will require more time to approximate the decision
boundary. As a result, Algorithm 1 is faster than PCA-
and BEST-based algorithms.
B. Algorithm 2
Algorithm 2 uses A2DPCA and examines its
merits over Algorithm 1. A2DPCA is merely an
alternative definition of 2DPCA because 2DPCA
works in the row direction of images, whereas
A2DPCA works in the column direction of images.
One can also construct the covariance matrix Gconsidering the outer product between column vectors
of images. This only needs a straightforward and
minor modification of (1). For this we consider Ak =
[(A(1)k )(A(2)k ) : : :(A
(m)k )] and A= [(A(1))(A(2)) : : : (A(m))],
where A(j)k and A(j) denote the jth column vector of Akand A, respectively. An alternative definition for imagecovariance matrix G therefore becomes
G =1
M
MXk=1
nXj=1
(A(j)k ¡ A(j))(A(j)k ¡ A(j))T: (5)
Let Z 2 Rm£q be a matrix with orthonormalcolumns. The orthonormal columns are eigenvectors
of the matrix G. Projection of the random matrix Aonto Z yields a matrix B = ZTA of size q by n.
MURTHY, ET AL.: A FAST AND EFFICIENT CLUTTER REJECTION ALGORITHM FOR SAR IMAGERY 2433
The optimal projection matrix Zopt can be obtainedby computing the eigenvectors Z1, : : : ,Zq of G, derivedfrom (5), corresponding to the q largest eigenvaluesof G, i.e., Zopt = [Z1, : : :Zq]. The value of q can bedetermined by setting a threshold μ in (4). Because theeigenvectors of the matrix G reflect the informationbetween columns of images, the A2DPCA is found
to work in the column direction of images. Thus,
Algorithm 2 differs from Algorithm 1 mainly in the
construction of covariance matrix G. Fortunately,this change in construction of the covariance matrix
leads to a significant positive gain in time, on our
data particularly. This is clearly demonstrated in our
experimental results. Schematically Algorithm 2 is
similar to Algorithm 1 (Fig. 2). Similarly, Algorithm 2
also has biquadratic time complexity of (O(M2nq+n2m)) and is faster than PCA and BEST (existing)based algorithms.
C. Algorithm 3
Algorithm 3 is based on two-stage 2DPCA.
2DPCA and A2DPCA work in the row and column
direction of images, respectively. Let us assume that
the two projection matrices X and Z are obtainedfrom Algorithm 1 and Algorithm 2, respectively. By
projecting an arbitrary input image A onto X and Z,we get an output C as shown in (6):
C = ZTAX: (6)
Algorithm 3 is similar to the above algorithms, except
2DPCA/A2DPCA is replaced by Two-Stage 2DPCA
(Fig. 2). The time complexity of Algorithm 3 is
O(M2dq+ n2m+ nm2), which is also biquadratic.Even though two-stage 2DPCA has to compute two
projection matrices X and Z, better dimensionalityreduction will be achieved. So the subsequent BEST
and MLP are faster than that of Algorithms 1 and 2.
Therefore, Algorithm 3 is expected to be faster than
Algorithms 1 and 2.
These projected features undergo the BEST
for adequate separation between two classes in the
subspace, e.g., the target and clutter. Finally, an MLP
classifies the targets. Algorithm 3 is also described
schematically in Fig. 2.
Note that both 2DPCA [1] and BEST [16] are
feature extraction methods, but they are used with
different objectives in our case. In the case of the
2DPCA family, the objective is to minimize the
energy loss to reduce noise, while the objective
of BEST is to further reduce dimension and
preserve adequate separation between classes in the
projected subspace. This adequacy in separation is
achieved through the consideration of both positive
and negative eigenvalues in BEST, contrary to
either positive or negative eigenvalues as done in
EST [2].
Fig. 3. A few samples of target and clutter chips. (a) Targets.
(b) Clutters.
III. EXPERIMENTAL RESULTS AND DISCUSSION
In order to examine the performance of the
proposed algorithm, and the existing (BEST-based)
algorithm mentioned in [16], we have used standard
SAR image data collected by DARPA and Wright
Laboratory as a part of the Moving and Stationary
Target Acquisition and Recognition (MSTAR)
program [29]. The targets in the MSTAR database
contain three T72 main battle tanks (MBT), three
BMP2 armored personnel carriers (APC), a BTR70
and a SLICY canonical target. The database also
contains clutter scenes at the Redstone Arsenal,
Huntsville. The data consists of X-band SAR images
with 1 ft by 1 ft resolution measured in the spotlight
mode. The targets have been measured over the full
360± azimuth angles with an increment from 1±—5±
considering depression angles of 15±, 17±, 30±, and45±. The target data in the database is presented assubimage chips centered about the target and each of
these chips has a size of 128£ 128 pixels.From the MSTAR database we extracted clutter
chips of our interest and targets manually. Figures 3(a)
and (b) show a sample of clutter and target chips,
respectively. The potential target images, obtained
from the target detection stage, generally consist of
either a part of shadow, or a part of target and shadow,
or only a part of target, or background objects etc.
Previous research shows that this type of potential
2434 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 49, NO. 4 OCTOBER 2013
target image, obtained from the detection stage, is the
main cause of a substantially large number of false
alarms. This ultimately results in a poor performance
of an ATR system on real-life data. In our analysis
we have used different types of clutters, namely
the SLICY canonical target, shadows (due to their
different rotations), shadows and a part of target,
only a part of target (MBT, APC, and BTR70), and
background objects (houses, buildings etc.) from
clutter scenes at the Redstone Arsenal, Huntsville,
from the MSTAR database. Thus, our analysis
includes a wide variety of clutter samples. The size of
these image chips is fixed at 51£ 41 pixels. The sizeof the target as well as the clutter chips are also the
same. We have considered 4000 target chips and 5945
clutter chips. The choice of the size of target/clutter
chips at 51£ 41 pixels has been made because themaximum size of the target present in the extracted
images is 51£41 pixels.To perform the experiment we have randomly
chosen 5% to 15% samples from each class and
extracted features through different techniques, already
mentioned in our algorithms, to train the standard
MLP, and we have computed the success rate.
Training has been carried out 50 times for a particular
sample size, while testing has been made with the rest
of the samples, and in each case the success rate has
been computed. All the three proposed algorithms are
compared with the existing (BEST-based) algorithm.
Note that we have also included 1D PCA in order to
check its performance against those of 2DPCA. Thus,
the entire PCA family members (1D as well as 2D)
are taken into consideration.
To extract the features efficiently, we have
preserved 95% energy information within the data of
all PCA family members. Dimensionality reduction,
on an average, was seen to be 963, 51£ 35, 41£ 30,and 30£ 35 features with PCA, 2DPCA, A2DPCA,and two-stage 2DPCA, respectively, from 51£41 features. Note that there is more reduction of
dimensionality in the case of A2DPCA compared
with 2DPCA because of different resolution in
row and column directions. To carry out the BEST
on extracted features, one can consider the most
dominant positive as well as the negative eigenvalues
with the corresponding eigenvectors for the projection
matrix. Figure 4(a) shows the plot of eigenvalues
of the correlation difference matrix in increasing
order for the existing (BEST-based) method, whereas
Fig. 4(b) shows the same for Algorithm 3. We have
considered the eigenvalues which are greater than
0.01 and less than ¡0:01, because the huge numberof eigenvalues in the interval (¡0:01,0:01) (Fig. 4)increases the network size and may subsequently
degrade the performance [16] and does not contribute
anything in enhancing the generalization capability of
a binary classifier. An MLP was selected with eight
hidden layers with ten nodes in each of them. The
Fig. 4. Eigenvalue numbers versus eigenvalues for existing
(BEST-based) and proposed algorithms. (a) Existing (BEST-based)
method. (b) Algorithm 3.
number of hidden layers and nodes in each layer was
decided through experimental observations for the best
results. The number of input nodes depends on the
number of features, i.e., the number of eigenvalues
considered through BEST.
Figure 5 shows the average success rate on
training data with varied sample size. It is observed
that in all the algorithms success rate (learning) does
not improve with increase in training size. For the
existing (BEST-based) algorithm, training performance
is good, but the test performance is found to be very
poor (Fig. 6). This is because the computation of a
huge correlation difference matrix and its eigenvectors
may involve loss of information, and the BEST
feature space is affected by speckle noise and so is
the subsequent classification. Clearly in the existing
(BEST-based) algorithm there is a downward trend
on increasing training data size and the same is also
reflected on test data. In the case of PCA, training
MURTHY, ET AL.: A FAST AND EFFICIENT CLUTTER REJECTION ALGORITHM FOR SAR IMAGERY 2435
Fig. 5. Average success rate for different training sample sizes.
is found to be oscillatory in nature, but for other
cases it is not prominent. The performance on the
test data, in the case of PCA, shows only marginal
oscillation (almost unnoticeable). The reason for
this is the inaccuracy in computation of eigenvectors
for a huge covariance matrix and noise in the data.
Since the eigenvectors of a huge covariance matrix
are not computed correctly, it introduces some error.
The classifier, therefore, learns the decision boundary
under noisy environment and as a result, the decision
boundary can be assumed to be oscillatory. Because
of these reasons the performance of PCA on the test
data is poor and the small oscillation in the test phase
is due to the huge number of test data points. Note
that the existing algorithm (BEST-based) has better
performance compared with that of the PCA-based
algorithm for consideration of the bipolar nature of
eigenvalues.
Figure 6 shows the average success rate on test
data with varied sample size. The performance is
more or less the same for Algorithm 1, Algorithm 2,
and Algorithm 3, but always superior to that of the
PCA-based algorithm and the existing (BEST-based)
algorithm. Performances of Algorithm 1, Algorithm 2,
and Algorithm 3 are better because for each of them
computation of eigenvectors is efficient for smaller
size of the covariance matrix, that ultimately helps
to reduce speckle noise effectively. The performance
of the existing (BEST-based) algorithm is only
comparable to that of the PCA-based algorithm. The
performance of the PCA-based algorithm is poorer
than any other algorithm because of poor learning
(Fig. 5) and evaluation of eigenvectors of the huge
covariance matrix.
Clearly, there is a downward trend on test
data (Fig. 6) for all the algorithms because of
fixed network size and huge training leading to
overlearning. A simple linear regression model has
been used to assess the downward tendency in the
performance of various algorithms with training
Fig. 6. Average success rate for different training sample sizes.
sample size. It has been found that with the increase
in training sample size the performance of all
algorithms is decreasing significantly. However, the
rate of decrement in the performance is relatively
small in the case of the proposed algorithms compared
with the existing (BEST-based) and PCA-based
algorithms. The rate of decrease in the performance
of Algorithms 1, 2, and 3 is ¡0:242, ¡0:228,and ¡0:223, respectively, and for the existing(BEST-based) and PCA-based algorithms is ¡0:357and ¡0:276, respectively. Note that there is nosignificant difference in the slopes of average success
rate curve for Algorithms 1, 2, and 3. It is understood
that performance may increase with the network
size but unfortunately this will also increase the
computational time. Therefore, to maintain a trade off
between the performance and time complexity, the
number of eigenvalues (for BEST) and the network
size should remain fixed.
From the time complexity point of view, one can
easily observe that Algorithm 3 is the best. Time
required for training and testing with various training
sample sizes is shown in Fig. 7. From Fig. 7 it is clear
that the existing (BEST-based) method is superior
only initially to the PCA-based algorithm and as
the training size increases, it becomes poorer and
poorer, and finally takes almost the same time as that
of the PCA-based algorithm. It is also seen that the
2DPCA-based algorithm (Algorithm 1) is inferior to
the A2DPCA-based algorithm (Algorithm 2) for all
training sample sizes. Algorithm 3, i.e., the two-stage
2DPCA-based algorithm, on the other hand, is found
to take the least amount of time for all training sample
sizes, and hence is the most superior among all the
algorithms.
IV. CONCLUSION
To find a good clutter rejection scheme for SAR
images, we have examined algorithms from PCA
2436 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 49, NO. 4 OCTOBER 2013
Fig. 7. CPU time in seconds for different training sample sizes.
family members. All the 2DPCA-based algorithms
are found to be superior, both with respect to time
and recognition rate as well as with respect to the
existing (BEST-based) method and the PCA-based
method. Clearly, preprocessing based on the 2DPCA
family reduces the effect of noise in SAR images
and hence produces good results. Again, of all the
methods discussed and analyzed, Algorithm 3 (i.e.,
two-stage 2DPCA-based algorithm) has been found to
be the fastest with its performance comparable to that
of other 2DPCA-based algorithms, and is by far the
best relative to the existing (BEST-based) algorithm.
The success rate as well as speed in Algorithm 3 is
mainly due to a significantly reduced number of noisy
features by the properties of two-stage 2DPCA and
the subspace separation capability of the BEST on
less noisy data obtained through two-stage 2DPCA.
Algorithm 3 is found to be the most suitable one for
clutter rejection in SAR imagery.
In a future work, based on our algorithmic
concept, the 2DPCA family will be replaced by a
supervised dimensionality reduction technique for
real-time clutter rejection in FLIR imagery and target
recognition.
ACKNOWLEDGMENT
Ramachandra Murthy is grateful to the Council
of Scientific & Industrial Research (CSIR), India
for providing him a Senior Research Fellowship
[9/93(0138)/2011-EMR-I].
REFERENCES
[1] Yang, J., et al.
Two dimensional PCA: A new approach to
appearance-based face representation and recognition.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 26, 1 (Jan. 2004), 131—137.
[2] Torrieri, D.
The eigenspace separation transform for neural network
classifiers.
Neural Networks, 12, 3 (Apr. 1999), 419—427.
[3] Haykin, S.
Neural Networks: A Comprehensive Foundation (3rd ed).
Upper Saddle River, NJ: Pearson Education, 2009.
[4] Zhao, Q. and Principe, J. C.
Support vector machines for SAR automatic target
recognition.
IEEE Transactions on Aerospace and Electronic Systems,
37, 2 (Apr. 2001), 643—654.
[5] Bhanu, B.
Automatic target recognition: State of the art survey.
IEEE Transactions on Aerospace and Electronic Systems,
AES-22, 4 (July 1986), 364—379.
[6] Bhanu, B. and Jones, T.
Image understanding research for automatic target
recognition.
IEEE Aerospace and Electronic Systems Magazine, 8, 10
(Oct. 1993), 15—23.
[7] Shaik, J. and Iftekharuddin, K.
Detection and tracking of targets in infrared images using
Bayesian techniques.
Optics and Laser Technology, 41, 6 (Sept. 2009), 832—842.
[8] Jacobs, S. P. and O’Sullivan, J. A.
Automatic target recognition using sequences of high
resolution radar range-profiles.
Signal Processing, 36, 2 (Aug. 2002), 364—381.
[9] Perry, R. P., DiPietro, R. C., and Fante, R. L.
SAR imaging of moving targets.
IEEE Transactions on Aerospace and Electronic Systems,
35, 1 (Jan. 1999), 188—200.
[10] Oliver, C. and Quegan, S.
Understanding Synthetic Aperture Radar Images.
Norwood, MA: Artech House, 1998.
[11] Kreithen, D., Halversen, S., and Owirka, G.
Discriminating targets from clutter.
Lincoln Laboratory Journal, 6, 1 (Spring 1993), 25—52.
[12] Ross, T. D., et al.
SAR ATR: So what’s the problem? An MSTAR
perspective.
Proceedings of SPIE Conference on SAR, vol. 3721, Aug.
1999, pp. 662—672.
[13] Novak, L. M., et al.
Effect of polarization and resolution on SAR ATR.
IEEE Transactions on Aerospace and Electronic Systems,
33, 1 (Jan. 1997), 102—116.
[14] Gao, G., et al.
Fast detecting and locating groups of targets in
high-resolution SAR images.
Pattern Recognition, 40, 4 (Apr. 2007), 1378—1384.
[15] Bhanu, B. and Lin, Y.
Genetic algorithm based feature selection for target
detection in SAR images.
Image and Vision Computing, 21, 7 (July 2003), 591—608.
[16] Chan, L. A., Nasrabadi, N. M., and Torrieri, D.
Bipolar eigenspace separation transformation for
automatic clutter rejection.
Proceedings of the IEEE International Conference on
Image Analysis and Recognition, vol. 1, 1999, pp.
139—142.
[17] Chan, L. A., Nasrabadi, N. M., and Torrieri, D.
Clutter rejection using eigenspace transformation.
In F. A. Sadjadi (Ed.), Proceedings of SPIE, vol. 3718,
Automatic Target Recognition IX, 1999, pp. 67—78.
[18] Young, S., et al.
Adaptive target detection in forward-looking infrared
imagery using the eigenspace separation transform and
principal component analysis.
Optical Engineering, 43, 8 (Feb. 2004).
MURTHY, ET AL.: A FAST AND EFFICIENT CLUTTER REJECTION ALGORITHM FOR SAR IMAGERY 2437
[19] Camps-Valls, G. and Bruzzone, L. (Eds.)
Kernel Methods for Remote Sensing Data Analysis.
Hoboken, NJ: Wiley, Dec. 2009.
[20] Javidi, B. (Ed.)
Smart Imaging Systems.
Bellingham, WA: SPIE Press, 2001.
[21] Henebry, G. M.
Advantages of principal components analysis for land
cover segmentation from SAR image series.
In T-D. Guyenne and D. Danesy (Eds.), Third ERS
Symposium on Space at the Service of our Environment
(ESA special publication), vol. 414, Mar. 1997, pp.
175—178.
[22] Sirovich, L. and Kirby, M.
Low-dimensional procedure for characterization of human
faces.
Journal of the Optical Society of America, 4, 3 (1987),
519—524.
[23] Kirby, M. and Sirovich, L.
Application of the Karhunen-Loeve procedure for the
characterization of human faces.
IEEE Transactions on Pattern Analysis and Machine
Intelligence, 12, 1 (1990), 103—108.
[24] Strang, G.
Introduction to Linear Algebra (4th ed.).
Wellesley, MA: Wellesley Cambridge Press, 2009.
[25] Wang, L., et al.
The equivalence of two dimensional PCA to line based
PCA.
Pattern Recognition Letters, 26, 1 (Jan. 2005), 57—60.
[26] Zhang, D. and Zhou, Z.
(2D)2PCA: Two-directional two-dimensional PCA for
efficient face representation and recognition.
Neurocomputing, 69, 1—3 (2005), 224—231.
[27] Camps-Valls, G. and Bruzzone, L.
Kernel based methods for hyperspectral image
classification.
IEEE Transactions on Geoscience and Remote Sensing, 43,
6 (2005), 1351—1362.
[28] Roth, M. W.
Survey of neural network technology for automatic target
recognition.
IEEE Transactions on Neural Networks, 1, 1 (Mar. 1990),
28—43.
[29] Ross, T. D., et al.
Standard SAR ATR evaluation experiments using the
MSTAR public release data set.
Proceedings of SPIE Conference on Algorithms for
Synthetic Aperture Radar Imagery V, vol. 3370, Sept.
1998, pp. 566—573.
2438 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 49, NO. 4 OCTOBER 2013
K. Ramachandra Murthy received his B.S. in computer science from the
S. V. University, India, in 2002, M.S. in mathematics from the University of
Hyderabad, India, in 2005, and M.Tech. in computer science from the Indian
Statistical Institute in 2008.
He worked as a senior research fellow in the U.S. Army project at Machine
Intelligence Unit, Indian Statistical Institute during 2009—2011. He is now a
CSIR-SRF at Machine Intelligence Unit, Indian Statistical Institute from 2011
to the present. His area of research includes manifold learning, distance metric
learning, pattern recognition, artificial neural networks, fuzzy theory, image
processing, and optimization techniques.
Sambhunath Biswas obtained his M.Sc. degree in physics, and Ph.D. in radio
physics and electronics, from the University of Calcutta, India, in 1973 and 2001,
respectively. In 1983—1985 he completed courses of M.Tech. (computer science)
from Indian Statistical Institute, Calcutta.
He was a UNDP Fellow at MIT, USA, to study machine vision in 1988—1989
and was exposed to research in the Artificial Intelligence Laboratory. He visited
the Australian National University at Canberra in 1995 and joined the school
on wavelets theory and its applications. In 2001 he became a representative of
the Government of India to visit China for holding talks about collaborative
research projects in different Chinese universities. He visited Italy and France in
June, 2010. He, at present, is a system analyst, GR-I (in the rank of professor)
at the Machine Intelligence Unit of the Indian Statistical Institute, Calcutta,
where is engaged in research and teaching. He is also an external examiner
of the Department of Computer Science at Jadavpur University in Calcutta.
He has published several research articles in international journals of repute
and is a member of IUPRAI (Indian Unit of Pattern Recognition and Artificial
Intelligence). He is the principal author of Bezier and Splines in Image Processing
and Machine Vision (Springer, London).
Ashish Ghosh is a professor with the Machine Intelligence Unit, IndianStatistical Institute. He has already published more than 150 research papers in
internationally reputed journals and refereed conferences, and has edited eight
books. His current research interests include pattern recognition and machine
learning, data mining, image analysis, remotely sensed image analysis, video
image analysis, soft computing, fuzzy sets and uncertainty analysis, neural
networks, evolutionary computation, and bioinformatics. Dr. Ghosh received the
prestigious Young Scientists Award in Engineering Sciences from the Indian
National Science Academy in 1995, and in computer science from the Indian
Science Congress Association in 1992. He was selected as an Associate of
the Indian Academy of Sciences, Bangalore, India, in 1997. He is a member
of the founding team that established the National Center for Soft Computing
Research at the Indian Statistical Institute, Kolkata, in 2004, with funding from
the Department of Science and Technology, Government of India, and is currently
in charge of the Center. He is acting as a member of the editorial boards of
various international journals.
MURTHY, ET AL.: A FAST AND EFFICIENT CLUTTER REJECTION ALGORITHM FOR SAR IMAGERY 2439