Feature Extraction in FLIR Imagery Using the
Continuous Wavelet Transform (CWT)
Romain Murenzi,1 Lance Kaplan,1;2 Fernando Mujica,1;3 and Sandor Der 4
murenzi@hubble.cau.edu, lmkaplan@hubble.cau.edu, fmujica@eedsp.gatech.edu, sder@arl.mil
Center for Theoretical Studies of Physical Systems 1;2
Department of Physics 1 and Engineering 2 , Clark Atlanta University
Atlanta, GA 30314
Phone: (404) 880-8655
Center for Signal and Image Processing 3
School of Electrical Engineering, Georgia Institute of Technology
Atlanta, GA 30332
Phone: (404) 894-2969
Army Research Laboratory 4
ARL-SE-SE, MD 20783
Phone: (301) 394-0807
Abstract
This paper investigates the use of the Continuous
Wavelet Transform (CWT) for detection of targets
in cluttered environments. We rst de ne the multiple 2D CWT energy densities in the feature space.
We then use the spatial energy density as a computational core for our target-detection algorithm. Results from FLIR images are presented at the end of
the paper. We used both the Mexican Hat wavelet
and target-adapted wavelets in the simulations.
Introduction
Automatic target detection and recognition (ATD/R)
consists of computer processing to detect and recognize target signatures embedded in a cluttered
environment. Typical targets include planes and
tanks; clutter includes grass, trees, topographical features, and atmospheric phenomena (e.g., clouds and
smoke). In general, the problem can be modeled using the following equation:
s(~x) = n(~x) +
X T (~x ? ~x );
N
l=1
l
ol
(1)
where n(~x) represents an additive noise (clutter +
measurement noise), Tl (~x) are targets to be detected
and recognized, and s(~x) represents the accessible
measured signal.
Automatic or assisted target detection and identication for FLIR imagery requires the ability to extract the essential features of an object from cluttered
environments under the condition that the range to
the target is unknown. Multiscale techniques (e.g.,
the multidimensional continuous wavelet transform
(MCWT)) are highly desirable, because they can extract and normalize both the unknown scale and orientation of the target. The typical features to be
extracted from the image of a target in a cluttered
environment are the position of the target, the spatial extent (size) of the target (i.e., the scale), and
the shape of the target, including its orientation and
symmetry.
To extract such target features, it is preferable not
to work in the image space, but rather to map the
image into a feature space of position, scale, and orientation | the MCWT performs this mapping exactly, as discussed in detail in [1, 2]. The 2D CWT is
a decomposition [3, 4] of an image on a set of dilated,
rotated, and translated versions of a single function
called the mother wavelet. The scale dependence allows sensitivity to variations in sensor resolution as
well as determination of target size. Rotation dependence leads to robust behavior, under varying target
orientations.
Integration over all the parameters of the squaredmodulus of the CWT gives the total energy in the
signal, which is Parseval's energy conservation condition. Therefore, integration on a subset of the parameters gives an energy density in the remaining
variables. A total of 14 energy densities (4 1D densities, 6 2D densities, and 4 3D densities) exist. Those
densities have what we refer to as CWT features.
In the next section; we de ne 2D CWT energy densities. We then describe an algorithm for target detection in cluttered environments, using the position
energy density, and present results of the algorithm
on FLIR data from the TRIM2 database, using the
anisotropic Mexican hat wavelet, as well as targetadapted wavelets.
Continuous Wavelet Transform and Energy
Densities in Feature Space
The 2D CWT is a linear transformation from
2~
L2 (IR2 ; d2~x) to L2 (IR+ [0; 2[IR2 ; da
a3 d d b),
Z
1 ( 1 r (~x ? ~b))s(~x) ; (2)
a a
IR2
where s is the input image, S is the CWT, is the
mother wavelet, a is the dilation parameter, is the
rotation parameter (with r , the corresponding 2D
rotation matrix), ~b = [bx ; by ]0 is the translation vector, the asterisk denotes complex conjugation, and
the hat denotes the Fourier transform.
Considering the energy density js(~x)j2 , js^(~k)j2 , we
have the following energy-conservation theorem:
S (a; ; ~b) =
Z
IR2
d2 ~x
d2~xjs(~x)j2 =
=
Z d2~kjs^(~k)j2
(3)
IR
ZZZ
1
d2~b djS (a; ; ~b)j2 :
2
a3
This theorem leads to the interpretation of
jS (a; ; ~b)j2 as an energy density of the signal, s, in
position, scale, and orientation parameters. It is clear
that any partial integration of the CWT energy density in any subset of the parameter space gives an
energy density in the remaining variables. We restrict the analysis to 2D energy densities, giving four
1D dimensional densities, six 2D densities, and four
3D densities. We have thus the following 2D CWT
energy densities:
Space (range and aspect) energy density
E34 (bx ; by ) =
Z 1da Z 2 djS(a; ; b ; b )j2 (4)
0 a3 0
x
y
Scale-angle energy density
Z +1db Z +1db jS(a; ; b ; b )j2 (5)
E12 (a; ) =
x
?1
y
?1
x
y
Anisotropy angle and aspect energy density
(bx; by ) = j~bj(cos( ); sin( ))
Z 1daZ +1dj~bjj~bjjS(a; ; j~bj; )j2 (6)
E24 (; ) =
0 a3 0
Scale-range energy density
E13 (a; j~bj) =
Z 2 dZ 2 d jS(a; ; j~bj; )j2
0
0
(7)
Anisotropy angle-range energy density
E23 (; j~bj) =
Z 1da Z 2d jS(a; ; j~bj; )j2
0 a3 0
(8)
Scale-aspect energy density
E14 (a; ) =
Z 1dj~bjj~bj Z 2djS(a; ; j~bj; )j2: (9)
0
0
We are currently being investigating these energy
densities in order to build a detector based on CWT
2D features of a given target type under various conditions. These features will then be used as inputs to
a convolutional neural network (CNN) algorithm [5].
These densities will also be used to design an ATR algorithm for detection, classi cation, and recognition
of targets in FLIR and SAR imagery.
Detection Algorithm Using the Position
(Range-Aspect) Energy Density
Consider the FLIR images shown in gures 1 and 2:
they contain a certain number of targets embedded
in a cluttered environment. To enhance these images,
we compute a 2D CWT energy density over position,
while suppressing clutter at the same time.
First, we compute the CWT in the position representation at all relevant scales a = aj and angles
= j , that is, S (aj ; j ; bx ; by ). For detection, we
take the image obtained for each xed a = aj and
= j , threshold it, and add all the images together.
Thresholding is performed in an adaptive way, becoming more severe for smaller a. The e ect of this
adaptation is to suppress the clutter information,
while preserving the target information. Note that
other nonlinear transformations (e.g., enhancement,
morphological operators) may also be applied. After
wavelet image integration, the target information is
??
s(~
x)
??
-?
@ @@
@@
R
S (a1; 1; ~b)j2
j
..
.
S (an ; n; ~b)j2
j
- 6Threshold-
+f
@@
@@R
- 6Threshold-.
??
?
- 6Threshold ?
-
-
E3;4(~b)
Figure 3: Detection algorithm.
sulting composite image. These centroids correspond
to the positions of potential targets; one controls the
false-alarm rate by adjusting the thresholds to eliminate ambiguous targets.
50
100
150
200
Application to Detection of Target in FLIR
Images
250
300
We apply the CWT-based algorithm to the two images shown in gures 1 and 2 taken from the TRIM2
database. We then compute the composite images as
described above.
Figures 4 and 5 show the results for the anisotropic
Mexican Hat wavelet (with anisotropy parameter =
Figure 1: A FLIR image consisting of targets (tanks M1) 5). Figures 6 and 7 show the results using a mother
and clutter, from the TRIM2 database.
wavelet adapted to the targets. We performed all
calculations at scales a = 1=8; 1=4; 1=2; 1 and orientations = 0; 4 ; 2 .
350
400
450
500
100
200
300
400
500
50
100
150
50
200
100
250
150
300
200
350
250
400
300
450
350
400
500
100
200
300
400
500
450
500
Figure 2: A FLIR image consisting of targets (helicopters)
and clutter, from the TRIM2 database.
100
200
300
400
500
Figure 4: Results for gure 1, using an anisotropic Mexican Hat wavelet with anisotropy parameter = 5:0 .
reinforced and becomes visually enhanced. Figure 3
shows a block diagram of the proposed target detec- As the gures show, the CWT-based detection altion algorithm.
gorithm makes the targets stand out from their backTo complete the steps of the detection algorithm, grounds with more de nition than the original imwe compute the centroids ~b = ~bi ; i = 1; :::; L in the re- ages. This suggests that the CWT may be used
50
50
100
100
150
150
200
200
250
250
300
300
350
350
400
400
450
450
500
500
100
200
300
400
500
100
200
300
400
500
Figure 5: Results for gure 2, using an anisotropic Mexi- Figure 7: Results for gure 2, using a target-adapted
can Hat wavelet with anisotropy parameter = 5:0 .
wavelet.
References
50
100
150
200
250
300
350
400
450
500
100
200
300
400
500
Figure 6: Results for gure 1, using a target-adapted
wavelet.
for target detection, as described in earlier, and that
CWT-based features might be appropriate as input
to a convolutional neural network [5].
Conclusions
In this paper, we presented a family of wavelet features based on the CWT, and we examined the e ectiveness of the CWT in separating targets from background in synthetic FLIR imagery. Target signatures
extracted from a training set can be formed into the
CWT feature space, allowing targets to be detected
through the nonlinear matching algorithm described
in [5]. Our future work will include automated learning of wavelet features, using arti cial neural nets on
a large set of real FLIR and SAR images.
Acknowledgments
The authors wish to thank Drs. Marvin Cohen and
Mark J. T. Smith for their valuable assistance in our
research.
View publication stats
[1] Jean-Pierre Antoine, K. Bouyoucef, P. Vandergheyst and R. Murenzi.,"Target Detection and
Target Recognition Using 2 Dimensional Continuous Isotropic and Anisotropic Waveletes," in Automatic Object Recognition V, SPIE-1995, Vol. 2485,
20-31, (1995).
[2] Sang-II Park, Mark J.T. Smith and R. Murenzi,
\Multidimensional Wavelets for Target Detection
and Recognition," in SPIE International Symposium on Aerospace/Defense Sensing and Controls,
Orlando, Florida, April 8-12, (1996).
[3] J-P. Antoine and R. Murenzi, \Two Dimensional
Directional Wavelets and the Scale-Angle Representation," Signal Processing 52, 259-281, (1996).
[4] J-P. Antoine, R. Murenzi and P. Vandergheynst,
\Two Dimensional Directional Wavelets in Image Processing," International Journal for Imaging
Systems and Technology, Vol.7, 152-165, (1996).
[5] Syed A. Rizvi and Vincent Mirelli, \Automatic
Target Recognition Using a Multi-layer Convolution Neural Network," preprint (1996).