Content uploaded by Abhay N Gaikwad
Author content
All content in this area was uploaded by Abhay N Gaikwad on Jul 29, 2014
Content may be subject to copyright.
Progress In Electromagnetics Research B, Vol. 17, 29–48, 2009
ANALYSIS OF CLUTTER REDUCTION TECHNIQUES
FOR THROUGH WALL IMAGING IN UWB RANGE
P. K. Verma, A. N. Gaikwad, D. Singh, and M. J. Nigam
Department of Electronics and Computer Engineering
Indian Institute of Technology Roorkee
Roorkee 247667, India
Abstract—Nowadays, through wall imaging (TWI) is an emerging
topic of research in which one of the most important tasks is
to minimize the clutter through which detection accuracy can be
improved. Clutter in TWI is due to many reasons like wall coupling,
antenna coupling, multiple reflections etc. To analyze the clutter
reduction techniques, firstly we indigenously assembled a TWI system
(i.e., step frequency continuous wave radar (SFCW)) in UWB range
(freq. 3.95 GHz to 5.85 GHz), and different observations have been
taken. We have considered metallic plate and one more material
with low dielectric constant (Teflon) as a target and kept them
behind the plywood wall. A-scan and B-scan observations have been
carried out. The observed data are preprocessed for imaging and
then different types of clutter reduction techniques like Principal
Component Analysis (PCA), Independent Component Analysis (ICA),
Factor Analysis (FA) and Singular Value Decomposition (SVD) have
been applied, and results were analyzed. Signal to noise ratio (SNR)
of the final images (i.e., after clutter removal with different techniques)
has been computed to compare the results and know the effectiveness
of individual clutter removal techniques. It is observed that ICA
has better capability to remove the clutter in comparison to other
applied techniques; especially it is found that ICA has a capability
to distinguish the difference between clutter and low dielectric target
whereas other clutter removal techniques are not showing significant
result.
Corresponding author: (dharmfec@gmail.com).
30 Verma et al.
1. INTRODUCTION
Surveillance/navigation systems such as television, infrared, and other
line-of-sight surveillance hardware are extensively used nowadays.
However, these systems cannot tell what is happening or locate
persons/assets on the other side of a wall, behind bushes, in the dark, in
a tunnel or a cave, or through a dense fog. X-rays may be used for such
purposes, but since they possess health risk, they are avoided. For an
effective detection system, the radar should have a transmitted signal
at a frequency low enough that should be capable of penetrating walls
and have a very wide bandwidth so that targets behind walls may be
clearly identified. Bandwidths need to be several gigahertzes to achieve
high resolution. UWB radar systems satisfy these low frequency and
large bandwidth requirements; they are defined as those for which the
relative bandwidth is equal to or greater than 20%.
Through wall detection [1–3] is an emerging field in research
because of its application in human monitoring, disaster search
and rescue, physical security, law enforcement, and urban military
operations. In TWI system, the electromagnetic waves that are
transmitted by the radar have to propagate through the air, non
metallic wall and other objects. TWI radar has capability to detect
any objects that lie in its line of sight if the conductivity of object or
dielectric constant or permeability is different from the surrounding
medium. It is usually the contrast in the permittivity that leads to
a reflection of the electromagnetic waves radiated by the transmit
antenna and helps in detection process. The reflected signal also
depends on the ratio between the size of object and wavelength.
Most of the work in through wall detection is currently focused
on the imaging in which researchers are working on target detection
using algorithms such as back-projection [4–6] and beamforming [7, 8].
Detection of different types of target with low and high dielectric
constants is a challenging task. In imaging, the presence of metal
target will reflect more energy and appears bright while target having
low dielectric constant will reflect less energy and appear dark that
makes the detection task challenging. So, we have focused on detection
of two targets having dielectric contrast in the same scene by using
clutter reduction techniques.
Researchers are using various clutter removal techniques in ground
penetrating radar (GPR) data but still in TWI; importance of these
techniques has to be explored. Clutter reduction techniques are
classified among others as statistical signal processing [9], classical
filtering [10–12] and non linear signal processing based on neural
networks [13, 14]. Automatic clutter reduction based on combination
Progress In Electromagnetics Research B, Vol. 17, 2009 31
using statistical and multilayer perceptrons is described in [15]. Clutter
reduction based on statistical signal processing techniques such as
PCA [16–18], ICA [19–23], method of FA [24–26], and SVD [27, 28]
is considered in present paper to remove or minimize the clutter.
All these techniques have their own advantages in image processing.
For example, ICA and PCA have feature extraction property. After
processing data using these techniques, SNR of images has been
calculated, and results are compared.
The paper is organized in following order. Experimental setup and
measurement procedure are outlined in Section 2. Section 3 deals with
data preprocessing while Section 4 elaborates the principles of various
clutter removal techniques which are used in this paper. Results and
discussion are covered in Section 5 followed by conclusion in Section 6.
2. EXPERIMENTAL SETUP AND MEASUREMENT
PROCEDURE
We have indigenously assembled step frequency continuous wave radar
(SFCW) [29] system for scanning the wall in the frequency band of 3.95
to 5.85 GHz at 4001 points. The stepped frequency continuous wave
radar has many advantages such as wider dynamic range, higher mean
power, lower noise figure, and the most important one is the possibility
of shaping the power spectral density. SFCW radar also provides single
and multi frequencies processing, time-frequency analysis, polarimetric
processing. The main advantage of SFCW radar system is its high
resolution in downrange.
In this setup Rohde & Schwarz vector network analyzer (VNA)
ZVB8 is used, which generates a stepped frequency waveform.
A pyramidal horn antenna is used in a monostatic mode having
bandwidth 1.9 GHz for transmitting and receiving signal. Circulator
in the same band was used for separating the received signal from the
transmitted signal. The antenna was mounted on 2D-scanning frame
made of wood on which the antenna can slide along crossrange and
along height. Observations were taken for 30 antenna positions in cross
range direction by shifting the antenna by 5 cm at each scanning point
(Figure 1). The observations have been carried out for A-scan and
B-scan. A-scan is obtained by stationary measurement, transmission
and collection of a signal after placing the antenna above the position
of interest. The collected signal is presented as signal strength vs time
delay or distance, and B-scan (or two dimensional data presentation)
signal is obtained as horizontal collection from ensemble of A-scans.
The horizontal axis of the two dimension image consists of crossrange
(antenna position), and vertical axis is downrange (distance from the
32 Verma et al.
VNA
Antenna System with Scanner
Targe
Wall
t
Figure 1. Block diagram of experimental setup for TWI.
antenna along the propagation direction of wave).
After calibrating VNA by standard two port calibration process
Through Open Short Matched (TOSM), the scattering parameters S
21
was measured at 4001 frequency points with the step size of 0.475 MHz
in presence and absence of target. Data are taken in frequency domain,
so it is converted to time domain by Inverse Fast Fourier Transform
(IFFT) for imaging.
Plywood wall of thickness of 12 mm is used for observation. An
aluminum metal plate of circular shape having diameter 58 cm and a
circular Teflon plate of diameter 50 cm behind the plywood wall have
been taken as a target, and both are separated in a distance of 30 cm
in cross range. Wall is kept at a distance of 190 cm from antenna,
and targets are kept at a distance of 30 cm from wall; therefore, total
distance from antenna to target is 221.2 cm. Though the maximum
room dimension in down range is not more than 5 m; the maximum
unambiguous range is taken more so that the other irrelevant signals
do not affect target detection.
3. DATA PROCESSING
3.1. Calibration Using Metal Sheet
In order to identify the delay due to antenna system, calibration using
metallic plate was carried out. The metallic plate (reference) is kept at
a known distance, and the range profile is plotted by which delay due
to antenna system is calculated. The reflection from antenna system
Progress In Electromagnetics Research B, Vol. 17, 2009 33
which was found through calibration should be subtracted to find out
the exact distance between antenna system and wall.
First of all, frequency domain data collected at 4001 points are
converted into time domain using Inverse Fast Fourier Transform
(IFFT) given by (1)
s(t) =
N−1
X
n=0
S(f
n
) exp(j2πf
n
t) (1)
where frequency f
n
varies from f
0
to f
0
+ n∆f; f
0
is the starting
frequency which is 3.95 GHz; n is the number of discrete points varies
from 0 to 4000; ∆f is the frequency step size which is 0.475 GHz;
t varies from 0 to (N − 1)/BW with step interval of 1/BW; BW is
bandwidth of the system which is 1.9 GHz; S(f
n
) is received signal in
frequency domain at nth frequency point; s(t) is time domain signal.
If the reference plate is located at a known distance of R
ref
from
antenna then one way propagation delay t
ref
is given by (2)
t
ref
=
R
ref
c
(2)
where c is the speed of light.
If t
disp
is the time after which we are getting reflection from
reference plate then delay due to antenna system can be calculated
by (3)
t
delay
= t
disp
− t
ref
(3)
The corrected time domain signal that removes the phase dispersion
within antenna system is given by (4)
s(t, t
delay
) =
N−1
X
n=0
S(f
n
) exp(j2πf
n
(t + t
delay
)) (4)
Figure 2(a) shows the results when the metallic plate is placed at the
location of wall for calibration. In Figure 2(b), two range profiles are
shown, where range profile shown in solid line (red line) is original one,
and range profile shown in circled (blue color) is after phase correction.
It is found that due to delay whole range profile is shifted by 48 cm.
3.2. Range Selection
For range selection, time domain signal must be converted into spatial
domain given by (5)
S(z) =
N−1
X
n=0
S(f
n
) exp(j2πf
n
t) (5)
34 Verma et al.
48 cm
(a) (b)
Signal strength
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Downrange in metres
Metal Wall Peak
Range Profile
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Downrange in metres
Range profile after phase correction
Normalized signal amplitude
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
Figure 2. Range profile for calibration: (a) When metallic plate
(reference) is placed at known distance, (b) after phase correction when
plywood wall is placed and a Teflon target behind it.
where z is distance in downrange which can be calculated as z =
ct
2
.
Since data are collected at 4001 points, the maximum range is
calculated by Z
max
=
c(N−1)
2BW
. In our experiment it is approximately
315 m with range resolution ∆Z =
c
2N∆f
, which is 7.89 cm. Since room
dimension is small, so 5 m range is considered for displaying range
profile.
4. CLUTTER REMOVAL TECHNIQUES
Clutter reduction is the main part in through wall imaging to
accurately detect the target and remove the unwanted signals which
arise due to the first reflection from the wall and other reflections
due to unwanted objects. Once the signal is transmitted through
the antenna, it suffers attenuation due to wall and other obstacles.
Therefore, the main aim of this paper is to reduce clutter due to wall
and enhance the peak due to target using signal processing techniques.
Generally in signal processing terms, the techniques used for clutter
reductions are called blind source or signal separation methods and
concerned with the separation of a set of signals called source signals
from their mixture signals, without acquaintance of any information
(or with very little information) about mixing background and sources.
Blind source separation is the separation of a set of signals into a set of
other signals in which the regularity between the signals is minimized
(decorrelation is minimized) or the regularity between the signals is
maximized (statistical independence is maximized). For TWI system,
Progress In Electromagnetics Research B, Vol. 17, 2009 35
it is assume that the scattered response is composed of superposition of
responses from individual scatterers, i.e., linear model. Thus, mainly
three components are assumed. One is measurement noise; second
is clutter; third is reflection from desired target. All the unwanted
contributions like antenna cross talk, wall reflection and multiple
reflections are considered as clutter. Thus, using clutter reduction
techniques, source signal can be decomposed into desired target and
clutter.
Some important clutter removal techniques which are generally
used in GPR data have been applied and analyzed in the present
paper [31]. Brief descriptions of these techniques are given in following
subsections.
4.1. Singular Value Decomposition (SVD)
SVD has many applications in signal processing and image processing
that can be used for many purposes such as noise reduction,
information retrieval, compression, and patterns detection [27, 28].
The main use of SVD is to split the data matrix into complementary
subspaces called signal and noise subspaces in order to increase SNR
which is useful for clutter reduction.
For clutter reduction using SVD, B-scan data are represented
by a rectangular matrix X
ij
, whose dimension is M × N, (i =
1, 2, . . . , M; j = 1, 2, . . . , N). Here i denotes the time or distance index,
and j denotes the antenna position index. The number of discrete
distance p oints is greater than the antenna index; therefore M ≥ N
will be assumed. SVD of X is given by (6)
X = USV
T
(6)
where U and V are (M ×M) and (N ×N) unitary matrices respectively
and S = diag (σ
1
, σ
2
, . . . , σ
r
) with σ
1
≥ σ
2
≥ . . . ≥ σ
r
≥ 0. The
columns of U and V are called the left and right singular vectors
respectively. Basically U and V are the eigenvectors of
©
XX
T
ª
and
©
X
T
X
ª
. For (r = N < M), the SVD is given by (7)–(9)
X = σ
1
.
.
.
u
1
.
.
.
¡
· · · v
T
1
· · ·
¢
+σ
2
.
.
.
u
2
.
.
.
¡
· · · v
T
2
· · ·
¢
+ · · ·
+σ
N
.
.
.
u
N
.
.
.
¡
· · · v
T
N
· · ·
¢
(7)
36 Verma et al.
X =
N
X
i=1
σ
i
u
i
v
T
i
(8)
X = M
1
+ M
2
+ M
3
+ . . . + M
N
(9)
where M
i
are matrices of the same dimensions of X and called as modes
or ith eigenimage of X. X can be decomposed into two subspace, signal
and clutter respectively.
X = X
signal
+ X
clutter
=
k
X
i=1
σ
i
u
i
v
T
i
+
N
X
i=k+1
σ
i
u
i
v
T
i
(10)
After applying SVD to our experimental data and analyzing all the
eigen images obtained using (9), we found that first eigen image M
1
provides the clutter information; second eigen image M
2
provides the
target information; rest eigen images represent the noise. B-scan data
X can be split into three parts.
X = M
t
+ M
c
+ M
n
(11)
where M
t
, M
c
and M
n
are the target, background (i.e., clutter) and
noise images respectively. Clutter can be estimated by (12), target
by (13) and noise image by (14)
M
c
= M
1
= σ
1
× u
1
× v
T
1
(12)
M
t
= M
2
= σ
2
× u
2
× v
T
2
(13)
M
n
=
N
X
i=3
σ
i
u
i
v
T
i
(14)
4.2. Factor Analysis
Factor analysis is a general term for a family of statistical techniques.
It uses correlations between observed variables to estimate common
factors. It makes use of second order statistics to extract signal so
that signal to noise ratio can be increased. Also, factor analysis is
concerned with the dimensional (number of variables) reduction of a
set of observed data in terms of a small number of latent factors [24–
26]. The main application of Factor Analysis is to reduce data variables
and to classify them.
Basically Factor Analysis extracts the set of factors from data
set using correlation. Generally these factors are orthogonal and are
ordered according to the proportion of the variance of the original data.
Progress In Electromagnetics Research B, Vol. 17, 2009 37
Therefore in general, only a (small) subset of factors is considered as
relevant, and the remaining factors are considered as either irrelevant
or nonexistent. The observed variables can be written as the linear
combinations of the factors plus error terms.
For clutter reduction, B-scan data are represented by a rectangular
matrix X
ij
, whose dimension is M × N, (i = 1, 2, . . . , M; j =
1, 2, . . . , N). Here i denotes the time or distance index, and j denotes
the antenna position index. The observed variables are modeled as
linear combinations of the factors plus error terms.
x
i
=
N
X
j=1
a
ij
s
j
+ e
i
(15)
In matrix notation it can b e written by (16)
X = AS + E (16)
where X is the matrix consisting the M A-scans in each row with N
time samples; S is the N ×K matrix of factor scores (latent variables);
A is the M × N matrix of factor loading; E is a matrix of error
terms. The Factor Analysis can be modeled in terms of variances
and covariances given by (17)
Σ = AΦA
T
+ Ψ (17)
where Σ is the M × M population covariance matrix of the observed
variables; Φ is the N × N covariance matrix of the factors; Ψ is the
M × M residual covariance matrix.
The primary assumption is that factors are uncorrelated, which
implies covariance matrix should be identity matrix i.e., Φ = I, and
the M-dimensional e is distributed according to N(0, Ψ), where Ψ is
diagonal matrix. The assumption of diagonality of Ψ implies that the
observed variables are conditionally independent (given the factors).
The distribution of observed variable x must have zero mean and
covariance Σ.
Factor Analysis finds optimal A and Ψ which best describe the
covariance structure of x. The best model of A and Ψ can be
found using Expectation Maximization (EM) algorithm [32]. The
EM procedure is a two step iterative procedure for maximizing the
log likelihood. A brief explanation of generalized EM algorithm
for maximum likelihood method is discussed in this paper. Detail
explanation of EM algorithm for maximum likelihood Factor Analysis
is given in [33].
38 Verma et al.
In Expectation step, it calculates the expected value of log
likelihood function with respect to unknown variable z given by
Eq. (18)
Q
³
Y
¯
¯
¯
Y
(T )
´
= E
z
|
x,Y
(T )
[log L (Y |X, Z )] (18)
where Y is the unknown parameter to be estimated under the
conditional distribution of Z when X is given; L (Y |X, Z ) is the
likelihood function.
Maximization step finds the optimal parameter values that
maximize the expectation which is computed in expectation step given
by (19)
Y
(T +1)
= arg max
Y
n
Q
³
Y
¯
¯
¯
Y
(T )
´o
(19)
Apply these two steps iteratively until a converged solution for Y is
obtained. After applying FA on experimental data, we found that it
splits data matrix X into factor score matrix S and factor loading
matrix A given by (16). Target can be extracted by selecting the
factor score and factor loadings components which carry the target
information and given by (20). Generally the second column of A and S
gives the information of target, and first column gives the information
of clutter.
X
target
= A
T
2
S
2
(20)
4.3. Principal Component Analysis (PCA)
PCA isolates the components on the basis of high correlation due
to large size of B-scan matrix. If highly correlated components are
present then the accuracy of algorithm will increase. Remaining
uncorrelated components can be removed easily. PCA can be used
in many applications, such as signal processing, data compressing,
data visualization, image analysis and pattern recognition. PCA
can be used for noise reduction in images by using the concept of
dimensionality reduction [16, 17].
For clutter reduction using PCA, B-scan data are represented by a
rectangular matrix X
ij
, whose dimension is M ×N (i = 1, 2, . . . , M; j =
1, 2, . . . , N). Here i denotes the time or distance index, and j denotes
the antenna p osition index. N principal components of data matrix X
can be given by (21)
Y = A
T
X (21)
where X = [x
1
, x
2
, x
3
, . . . , x
n
]
T
is the zero-mean input vector; Y =
[y
1
, y
2
, y
3
, . . . , y
n
]
T
is the output vector called the vector of principal
components (PCs); A is an M × N matrix that transforms X into
Progress In Electromagnetics Research B, Vol. 17, 2009 39
Y . The purpose of PCA is to derive a relatively small number of
decorrelated linear combination (principal comp onent) of a set of
random zero-mean variables while retaining as much of the information
from the original variables as possible. Therefore, PCA expresses input
data variables into smaller number of decorrelated linear combination
of a set of zero mean random variables, while retaining as much of the
information from the original variables as possible. The basic idea in
PCA is to find the rows of the y
T
1
, y
T
2
, y
T
3
. . . , y
T
n
. PCA assumes that
A is an orthonormal matrix (A
T
i
· A
j
= δ
ij
) such that the covariance
matrix of Y ; (C
y
) is diagonalized.
A can be computed using covariance matrix. Let X be the data
matrix after normalization and subtracting the mean. Then covariance
matrix C
x
of X is given by (22)
C
x
=
1
N
XX
T
(22)
The eigenvector and eigenvalue matrices of C
x
are Φ and Λ respectively
and can be computed by (23)
C
x
Φ = ΦΛ (23)
where Λ = diag(λ
1
, λ
2
, λ
3
, . . . , λ
N
) and λ
1
, λ
2
, λ
3
, . . . , λ
N
are the eigen
values. After arranging eigen values in the decreasing order, λ
1
≥ λ
2
≥
λ
3
≥ . . . ≥ λ
N
the matrix of N leading eigen vectors A is given by (24)
A = [Φ
1
, Φ
2
, Φ
3
, ..., Φ
N
] (24)
Principal component matrix S can be given by (25)
S = A
T
X (25)
It infers that PCA can be used as given in Eq. (25) for detection of the
targets or objects behind the walls. This can be done by selecting some
components that mainly carry target information, say A
p
, and rest
components represent the clutter. The reconstructed clutter-free signal
space can be extracted from the original B-scan matrix containing
target and clutter information. After calculating principal components,
target can be extracted by second column of A and S. Generally, first
eigen image represents the clutter, and second eigen image represents
the target. It means that second column of transformation matrix A
i.e., A
2
and principal component matrix S i.e., S
2
represents the target
that is given by (26).
X
target
= A
T
2
S
2
(26)
40 Verma et al.
4.4. Independent Component Analysis (ICA)
ICA is used to solve blind source separation problem. ICA divides
data into statistically independent components while other techniques
such as PCA or FA represents data into uncorrelated components.
Therefore, PCA or FA cannot separate signals efficiently because
uncorrelatedness is not enough. Statistical independence is necessary
which takes into consideration higher order moments which are
stronger statistical properties than decorrelation. Therefore, ICA is
widely used in many applications such as feature extraction and noise
reduction from the images, finding hidden factors from financial data
and mostly used in telecommunications for separating the original
source signal from interfering signals [18–22]. In ICA model, it is
assumed that the observed data X have been generated from source
data S through a linear pro cess X = AS, where both the sources S and
mixing matrix A are unknown. ICA algorithms are able to estimate
both the sources S and mixing matrix A from the observed data X
with very few assumptions [16–21].
For clutter reduction using ICA, B-scan data are represented by a
rectangular matrix X
ij
, whose dimension is M ×N (i = 1, 2, . . . , M; j =
1, 2, . . . , N). Here i denotes the time index, and j denotes the antenna
position index.
ICA assumes that every x
i
is a linear combination of each s
j
given
by (27)
x
i
=
N
X
j=1
a
ij
s
j
(27)
j = 1, 2, 3, . . . , N or in the matrix notation
X = AS (28)
Here A is an M × N basis transformation or mixing matrix, and S
is the matrix holding the N independent source signals in rows of N
samples. ICA of matrix X can be obtained by finding a full rank
separating matrix W such that output signal matrix can be defined by
Y = W X. The estimation of source signal can be given by (29)
ˆs
j
= y
j
=
N
X
i=1
w
ji
x
i
(29)
j = 1, 2, 3, . . . , N or in the matrix notation
ˆ
S = Y = W X (30)
Progress In Electromagnetics Research B, Vol. 17, 2009 41
where W is a N × M matrix which makes the outputs
_
S from the
linear transformation of the dependent sensor signals x as independent
as possible.
Formulation of ICA can be done in two steps. First one is
to formulate a contrast function G(y) that estimates the level of
statistical independence between the components of y, and second one
is the optimization of contrast function that enables the calculation
of independent components. Contrast function estimates the level of
statistical independence between the components of y i.e., optimization
of contrast function provide the independent components. To apply
ICA, some preprocessing is needed. The most basic preprocessing is
centering in which the mean is subtracted from each range profile in B-
scan matrix X. Second preprocessing is whitening in which observed
vector X is transformed into new vector
˜
X which is white i.e., its
components are un-correlated, and their variance is equal to unity.
Here we have used the FASTICA [23] algorithm which is fixed
point iteration based algorithm to calculate the separating matrix W
by finding a maximum of non Gaussianity of W
T
X. After computing
the separating matrix W , mixing matrix A can be computed by
taking inverse of it i.e., A = W
−1
. Since mixing matrix is known,
corresponding independent component S matrix can be calculated
using (28).
After applying ICA on experimental data, we are able to get the
number of independent components as much the number of sources or
A-scans. By considering each row of independent comp onent matrix S
and column of mixing matrix A, images have been generated using (28)
and observed. Image that contains target information can be chosen,
and remaining is discarded.
X
target
= A
target
S
target
(31)
5. RESULTS AND DISCUSSION
Clutter reduction techniques discussed in above sections are applied
to the experimental data, and the images have been compared on the
basis of signal to noise ratio (SNR). SNR can be given as the ratio
of average energy of the image matrix after clutter reduction to the
average energy of clutter plus noise matrix. So in the denominator
term, we will get clutter plus noise term by subtracting it from raw
B-scan image. The final image after applying clutter reduction is
supposed to have only information about the target, but actually
it is not and contains good amount of background noise that we
have obtained after applying different clutter reduction techniques.
42 Verma et al.
Therefore, in the numerator term, we have not taken average energy
of image giving target information b ecause this contributes more to
the numerator term of SNR and difficult to compare different clutter
reduction techniques in this paper. Therefore, we have considered the
peak signal to noise ratio (PSNR) [34] in which numerator term of
SNR is taken as the p eak value of normalize image matrix. It infers
that PSNR may be used as a first hand indicator to compare the
results of various clutter reduction techniques. PSNR can be calculated
using (32) and (33).
PSNR (dB) = 10 log {1/MSE } (32)
MSE =
1
M × N
N
X
i=1
M
X
j=1
{g (i, j) − f (i, j)}
2
(33)
where g(i, j) is our original B-scan image; f(i, j) is the reconstructed
image after clutter reduction; MSE is mean square error; M and N
are dimensions of image.
Results after applying clutter reduction techniques are shown in
Figure 3 that includes A-scan (range profile) and B-scan images. In
A-scan, solid line shows original target with wall, while dashed line
represents target signal estimated using clutter removal techniques.
B-scan images are obtained after clutter reduction.
Figure 3(a) shows the range profile of raw data (without clutter
reduction) in which target peak at 221 cm and wall peak at 190 cm
are observed, and B-scan image is shown in Figure 3(b) in which both
metal and Teflon targets are displayed, but Teflon target is suppressed
because of clutter.
Range profiles are observed for both targets. It is found that only
metal target peak is visible, while Teflon target is not detected clearly
in SVD, FA and PCA techniques whereas in ICA it is observed. It may
be due to the reason that the reflected signal strength with metal target
will be high in comparison to Teflon target. Figure 3(c) shows the
range profile of the metal target signal with ICA. The signal after ICA
is shown in dashed line and compared with original signal containing
wall and target reflections peak shown in solid line. From this figure,
we can observe that only target peak is enhanced, and noise floor is
very low. Figure 3(d) shows the B-scan image after clutter reduction
using ICA in which both targets are clearly visible at the distance
about 221 cm in down range. This image has very low background
noise (0.0 to 0.2 normalized values).
Figure 3(e) shows the range profile of metal target with SVD
clutter reduction technique. In this figure, it is observed that target
Progress In Electromagnetics Research B, Vol. 17, 2009 43
peak is enhanced; clutter is suppressed in one hand, and in the other
hand noise floor is also increased. This can be more visualized in B-
scan image (Figure 3(f)), where metal target is clearly visible, but
second target (Teflon) is mixed with clutter. In this figure, noise level
is also high (in between 0.3 to 0.4 normalized values).
Figures 3(g) and 3(h) show A-scan and B-scan images after PCA
Metal Tar
g
e
t
Teflon Tar
g
e
t
(a) (b)
Metal Tar
g
e
t
Teflon Tar
g
e
t
(c) (d)
Metal Tar
g
e
t
(e) (f)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Downrange in metres
Wall Peak
Range profile
Target Peak
Normalized signal amplitude
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Raw B-scan image
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
0 0.5 1 1.5
crossrange in metres
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Metal Target Teflon Target
Raw and estimated source (target) signal using ICA
0 1 2 3 4 5
Downrange in metres
Normalized signal amplitude
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Downrange in metres
B-scan image after clutter removal using ICA
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1 1.5
crossrange in metres
Metal Target Teflon Target
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
Downrange in metres
B-scan image after clutter removal using SVD
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
crossrange in metres
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
Downrange in metres
0 0.5 1 1.5
Metal Target
Raw and estimated source (target) signal using SVD
Normalized signal amplitude
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5
Downrange in metres
44 Verma et al.
Metal Tar
g
e
t
(g) (h)
Metal Tar
g
e
t
(i) (j)
Raw and estimated source (target) signal using PCA
Normalized signal amplitude
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5
Downrange in metres
B-scan image after clutter removal using PCA
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
Downrange in metres
crossrange in metres
0 0.5 1 1.5
Metal Target
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
B-scan image after clutter removal using FA
crossrange in metres
0 0.5 1 1.5
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
Downrange in metres
Metal Target
Raw and estimated source (target) signal using FA
Normalized signal amplitude
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 1 2 3 4 5
Downrange in metres
Figure 3. (a) Range profile of raw data (without clutter reduction),
(b) B-scan of raw data (without clutter reduction), (c) Range profile
with and without clutter reduction using ICA, (d) B-scan after using
ICA, (e) Range profile with and without clutter reduction using SVD,
(f) B-scan after using SVD, (g) Range profile with and without clutter
reduction using PCA, (h) B-scan after using PCA, (i) Range profile
with and without clutter reduction using FA, (j) B-scan after using
FA.
clutter reduction technique, whereas Figures 3(i) and 3(j) show A-
scan and B-scan images after FA clutter reduction techniques. In
all these figures i.e., 3(h) and 3(j), it is clearly observed that the
detection of metal target is enhanced, while background noise is not
so much suppressed by these techniques and secondly, it is difficult
to detect the second low dielectric target i.e., Teflon. If we compare
the results of various clutter reduction techniques (i.e., Figures 3(c)
to 3(j)), it is clearly observed that ICA based technique outperforms
in comparison to other techniques for extracting target feature of low
dielectric material from the raw clutter data.
Progress In Electromagnetics Research B, Vol. 17, 2009 45
Table 1. Performance of clutter reduction algorithms on basis of
PSNR.
S. No. Clutter reduction Algorithm PSNR (dB) Results
1.
Independent Component
Analysis (ICA)
36.25
Both metal and
Teflon target
detected
2. Factor Analysis (FA) 26
Only metal
target detected
3.
Singular Value
Decomposition (SVD)
23.6
Only metal
target detected
4.
Principal Component
Analysis (PCA)
23.42
Only metal
target detected
The performance of these clutter reduction techniques is compared
using the PSNR of B-scan images. PSNR of the images is computed
by using (32) and (33). Table 1 shows comparison of clutter reduction
techniques, and it is clearly observed that ICA has better PSNR in
comparison to other three techniques. Another observable point is
that we are unable to detect the Teflon target using PCA, SVD and
Factor Analysis. Only ICA is capable of extracting that target as we
can see from Figure 3(d).
6. CONCLUSION
A TWI system in UWB range has been assembled, and various data
for A-scan and B-scan have been collected. The focus of present paper
is to explore the possibility of application of various existing clutter
removal techniques for TWI data and also to check the possibility of
detection of low dielectric target. For this purpose, different clutter
removal algorithms have been implemented and compared. ICA based
technique gives a better result than other clutter removal techniques
like SVD, PCA and FA. PSNR for B-scan image is quite high in
case of ICA compared to other techniques. It is also observed that
ICA based clutter removal technique have a better potential to detect
low dielectric constant target like Teflon b ehind the plywood wall.
In future, analysis will be carried out for different walls like brick
wall, asbestos wall etc. All the clutter reduction techniques which are
considered in this paper are based on the assumption of linear mixture
of signals. Thus, non linear model may be studied in future as it is
more realistic.
46 Verma et al.
REFERENCES
1. Baranoski, E. J., “Through-wall imaging: Historical perspective
and future directions,” J. Franklin Inst., Vol. 345, 556–569,
Jan. 2008.
2. Farwell, M., J. Ross, R. Luttrell, D. Cohen, W. Chin, and
T. Dogaru, “Sense through the wall system development and
design considerations,” J. Franklin Inst., Vol. 345, 570–591,
Jan. 2008.
3. Dehmollaian, M., M. Thiel, and K. Sarabandi, “Through-the-wall
imaging using differential SAR,” IEEE Trans. Geosci. Remote
Sens., Vol. 47, No. 5, 1289–1296, May 2009.
4. Yegulalp, A. F., “Fast backprojection algorithm radar for
synthetic aperture,” The Record of the 1999 IEEE Radar
Conference, 60–65, Waltham, MA, USA, Apr. 1999,
5. Cui, G., L. Kong, and J. Yang, “A back-projection algorithm
to stepped-frequency synthetic aperture through-the-wall radar
imaging,” IEEE 1st Asian and Pacific Conference on Synthetic
Aperture Radar, APSAR 2007, 123–126, 2007.
6. Lundgren, W., U. Majumder, M. Backues, K. Barnes, and
J. Steed, “Implementing SAR image processing using backprojec-
tion on the cell broadband engine,” IEEE Conference on Radar,
1–6, May 2008.
7. Ahmad, F., M. G. Amin, S. A. Kassam, and G. J. Frazer,
“A wideband synthetic aperture beamformer for through-the-
wall imaging,” IEEE International Symposium on Phased Array
Systems and Technology, 187–192, Oct. 2003.
8. Xu, X. and R. M. Narayanan, “Enhanced resolution in SAR/ISAR
imaging using iterative sidelob e apodization,” IEEE Trans. Image
Processing, Vol. 14, No. 4, 537–547, Apr. 2005.
9. Xu, X., E. L. Miller, C. M. Rappaport, and G. D. Sower,
“Statistical method to detect subsurface objects using array
ground-penetrating radar data,” IEEE Trans. Geosci. Remote
Sensing, Vol. 40, 963–976, Apr. 2002.
10. Merwe, A. V. and I. J. Gupta, “A novel signal processing technique
for clutter reduction in GPR measurements of small, shallow land
mines,” IEEE Trans. Geosci. Remote Sensing, Vol. 38, 2627–2637,
Nov. 2000.
11. Zoubir, A. M., I. J. Chant, C. L. Brown, B. Barkat, and
C. Abeynayake, “Signal processing techniques for landmine
detection using impulse ground penetrating radar,” IEEE Sensors
J., Vol. 2, No. 1, 41–51, Feb. 2002.
Progress In Electromagnetics Research B, Vol. 17, 2009 47
12. Kempen, V. L. and H. Sahli, “Signal processing techniques
for clutter parameters estimation and clutter removal in GPR
data for landmine detection,” Statistical Signal Processing, 2001
Proceedings of the 11th IEEE Signal Processing Workshop, 158–
161, May 2001.
13. Vicen-Bueno, R., R. Carrasco-lvarez, M. Rosa-Zurera, and
J. C. Nieto-Borge, “Sea clutter reduction and target enhancement
by neural networks in a marine radar system,” Sensors, Vol. 9,
1913–1936, 2009.
14. Chan, L. A., N. M. Nasrabadi, and D. Torrieri, “Eigenspace trans-
formation for automatic clutter rejection,” Optical Engineering,
Vol. 40, No. 4, 564–573, Apr. 2001.
15. Xie, N., H. Leung, and H. Chang, “A multiple-model prediction
approach for sea clutter modeling,” IEEE Trans. Geosci. Remote
Sensing, Vol. 41, 1491–1502, Jun. 2003.
16. Yu, C. Q., X. S. Guo, A. Zhang, and X. J. Pan, “An improvement
algorithm of principal component analysis,” International Con-
ference on Electronic Measurement and Instruments, 2529–2534,
Jul. 2007.
17. Diamantaras, K. I., “PCA neural models and blind signal
separation,” International Joint Conference on Neural Networks,
Vol. 4, 2997–3002, Jul. 2001.
18. Abujarad, F. and A. Omar, “GPR data processing using
the component-separation methods PCA and ICA,” IEEE
International Workshop on Imaging Systems and Techniques, 60–
64, Apr. 2006.
19. Karlsen, B., J. Larsen, H. B. D. Sorensen, and K. B. Jakobsen,
“Comparison of PCA and ICA based clutter reduction in GPR
systems for anti-personal land-mine detection,” Proceedings of
the 11th IEEE Signal Processing Workshop on Statistical Signal
Processing, 146–149, Aug. 2001.
20. Comon, P., “Independent component analysis, a new concept,”
Signal Processing, Vol. 36, 287–314, Apr. 1994.
21. Baloch, S. S. H., H. Krim, and M. G. Genton, “Robust
independent component analysis,” 13th IEEE Workshop on
Statistical Signal Processing, 61–64, Jul. 2005.
22. Hyvarinen, A., J. Karhunen, and E. Oja, Independent Component
Analysis, John Wiley and Sons, USA, 2001.
23. Hyvarinen, A., “Fast and robust fixed-point algorithms for
independent component analysis,” IEEE Trans. Neural Net.,
Vol. 10, No. 3, 626–634, 1999.
48 Verma et al.
24. Chen, C. H. and W. Zhenhai, “ICA and factor analysis
application in seismic profiling,” IEEE International Conference
on Geoscience and Remote Sensing Symposium, 1560–1563,
Aug. 2006.
25. Wu, N. and J. Zhang, “Factor analysis based anomaly detection,”
Information Assurance Workshop, 2003, IEEE Systems, Man and
Cybernetics Society, 108–115, Jun. 2003.
26. Hong, S., “Warped image factor analysis,” 1st IEEE International
Workshop on Computational Advances in Multi-sensor Adaptive
Processing, 121–124, Dec. 2005.
27. Abujarad, F., A. Jostingmeier, and A. S. Omar, “Clutter
removal for landmine using different signal processing techniques,”
Proceedings of the Tenth IEEE International Conference on
Ground Penetrating Radar, GPR 2004, 697–700, Jun. 2004.
28. Wall, M. E., A. Rechtsteiner, and L. M. Rocha, “Singular value
decomposition and principal component analysis,” A Practical
Approach to Microarray Data Analysis, Chapter 5, 91–109,
Boston, MA, USA, 2003.
29. Chandra, R., A. N. Gaikwad, D. Singh, and M. J. Nigam,
“An approach to remove the clutter and detect the target for
ultra-wideband through-wall imaging,” Journal of Geophysics and
Engineering, Vol. 5, 412–419, Oct. 2008.
30. Cois Cardoso, J., “Blind signal separation: Statistical principles,”
IEEE Proc., Vol. 86, No. 10, 2009–2025, Oct. 1998.
31. Liu, J. X., B. Zhang, and R. B. Wu, “GPR ground bounce removal
methods based on blind source separation,” PIERS Online, Vol. 2,
No. 3, 256–259, 2006.
32. Saul, L. K. and M. G. Rahim, “Maximum likelihood and
minimum classification error factor analysis for automatic speech
recognition,” IEEE Trans. Speech Audio Process., Vol. 8, No. 2,
115–125, Mar. 2000.
33. Rubin, D. and D. Thayer, “EM algorithms for factor analysis,”
Psychometrika, Vol. 47, No. 1, 69–76, Mar. 1982.
34. Wei, G. W., “Generalized Perona-Malik equation for Image
restoration,” IEEE Signal Proc. Letters, Vol. 6, No. 7, 165–167,
Jul. 1999.