Design of a Configurable ATR System Using

MATLAB
Eric L. Wright, Member, IEEE and James Northern III, Member, IEEE

network was capable of scanning images, which produced a

Abstract—Highly efficient and adaptable image recognition firing rate spectrum that maximized at the target’s location
systems have become a high priority within the last decade due to within the image. The classification algorithm differs,
terrorism. This paper describes a design and development process however, from our system in that a maximum response does
for a configurable automatic target recognition system. Our novel not necessarily indicate the target’s location within the image;
template matching system employs a Matlab algorithm developed
to accurately detect object patterns within a JPEG image. After
downloading the original JPEG image, the Matlab algorithm
processes it in three sequential phases: 1) “Sobel” edge detection
of the original image 2) Grey- level template matching based on
the Squared Euclidean distance theory and 3) Statistical pattern
recognition of the resulting convoluted image. For our test cases,
the cross correlation was determined using a template library
created for the algorithm’s image recognition process. The
average runtime for our system ranges from 5-8 minutes, per test
case, with 100% correct recognition using the template matching
technique described.

Index Terms—Image recognition, signal processing, template

matching, optimization.

I. INTRODUCTION

T HE ability to quickly detect and identify people and

objects has recently become a high priority for various
departments and agencies within the U.S. government;
including the Department of Homeland Security and the
Department of Defense. Consequently, the demand for
adaptable, robust target recognition systems has stimulated
significant interest among researchers in the science Figure 1. ATR algorithm flowchart.
community. Typically, target recognition systems classify
video sequences by integrating detection, tracking and final resulting in the reduction of system accuracy. In this paper, we
image recognition. In Patil et al [1], a tracking and approach the target recognition task by integrating detection,
classification system is described that consists of five cameras template matching and object classification, Fig. 1, of the
oriented to capture partially overlapping video frames that image from a grey-level perspective; dealing only with still
were stitched together to produce a single panoramic video images. The matching robustness of our system is primarily
frame for target recognition processing. However, unlike our attributable to the ability to reconfigure each phase of our
system, it uses color blob tracking through mean shift analysis template matching scheme. Image enhancements are
instead of specific object tracking within images. In Clarkson accomplished during the first stage of our process by adjusting
[2], a method for performing target detection and classification the threshold value. The threshold, in our algorithm,
using (pRAM) neural networks is presented. Once successfully determines the sharpness level of the grayscale image and its
trained on the edge detection of infrared images, the neuron manipulation helps to eliminate distortion, improving the
overall ability of the template matching process. The algorithm
can also be configured to recognize any target through
Dr. J. Northern and Eric Wright are with the Department of Electrical and template library modification. To configure the library, an
Computer Engineering, Prairie View A&M University, Prairie View, TX
77446 USA (phone: 936-261-9915; fax: 936-261-9930; e-mail: image is downloaded from the internet and its template created
[email protected] or [email protected]). from a distinguishable pattern within the image. It is the
simplicity in configuration that allows the algorithm to be
continuously reconfigured to successfully recognize any target.

978-1-4244-2077-3/08/$25.00 ©2008 IEEE.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on August 18,2010 at 13:36:00 UTC from IEEE Xplore. Restrictions apply.
 Detection is accomplished by the system, automatically,
once the image is initially uploaded. Next, we employ the 2. Each input image will be separately compared to
“Sobel” method of edge detection to convert the image into its each template.
grey-level representation. Afterwards, the target pattern within
the binary image is tracked using a binary template matching 3. Successful target recognition will be determined by
scheme. The resulting match correlation represents the input selecting the template with the greatest amount of
for the classification stage of our matching process. Final image overlap.
image classification is accomplished with a probabilistic
method similar to the Principal Component Analysis method
(PCA) described in Clarkson [2], where match evaluation is Template filtering through the binary image generated a
accomplished at the end of each three stage process. Correct confidence measure whose correlation was primarily
recognition results in an image identification statement, motivated by the Squared Euclidean Distance theory outlined
otherwise a “no match” results and the algorithm resets itself. in Lewis [3]. If the template value matches the image value, a
We ran two test cases for the evaluation experiment, using 3 “1” was returned else a “0” was entered.
(512x512) downloaded images shown in Fig. 5. The remainder of this paper discusses our experiment in
further detail. Section 2 describes the detection and tracking
algorithm. Section 3 describes the template matching
% template library call algorithm. Section 4 describes the target classification
Load templates; algorithm. Section 5 details the evaluation and simulation of
our system. The conclusion and future work are discussed in
% “Sobel” edge detection the final section.
ATR_tgt_detect;

% 1st T.M. convolution II. TARGET DETECTION

ATR_tgt_recogntn(dimtmp);
Edge detection is the first and most important step in our
% target recogntn tracker variable target detection algorithm. Neighboring pixel intensities are
Z(1) = mtch1; compared within the original image in order to pronounce the
images maximum changes in intensity. Extreme changes in
% 2nd T.M.convolution intensity within an image correspond to the boundary where
ATR_tgt_recogntn(Nickletmp);
Z(2) = mtch2; object meets background and the resulting silhouette
represents a filtered version of its predecessor in binary
% 3rd T.M. convolution format.
ATR_tgt_recogntn(Pentmp); To accomplish image edge detection, the “Sobel” method
Z(3) = mtch3;
for edge detection was employed. The “Sobel” method,
% final recognition process considered to be a “gradient filter” edge detection method,
For u = 1:3 detects the edges of an image by locating the maximum and
If Z(u) == 1 minimum within the first derivative of the image. The “Sobel”
Out = u; % obj class var
End for
operator performs a 2-D spatial gradient measurement, on the
input grayscale image, to approximate the absolute gradient
% determine template match magnitude at each point within the image. Our algorithm uses
If (out == 1) two (3x3) “Sobel” matrix kernels; one for horizontal intensity
Disp(‘dime’); changes and another for vertical intensity changes (Fig. 3).
Elseif (out ==2)
Disp(‘nickel’); Filtering the image in both directions produces two well
Elseif (out == 3) defined image contours.
Disp(‘penny’);
Else disp(‘not in our database’)
End if;

Figure 2. Matlab pseudo-code for ATR main function

A Template library consisting of three pattern-templates was

created by extracting specific details from three of the images
downloaded. The library was created based on the following
assumptions: Figure 3.“Sobel” matrix kernels for horizontal and vertical
intensity changes.
1. Each template will be created from an identical
image feature extracted from each predetermined The Sobel masks were filtered through the original RBG
system target. image one window at a time computing the weighted average

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on August 18,2010 at 13:36:00 UTC from IEEE Xplore. Restrictions apply.
 of all neighboring pixels creating two separate images; one
containing vertical lines and the other containing horizontal
lines. The addition of the images produces a single (510x510) (1)
edge detected image due to the size of the matrix kernels. A
threshold technique is then applied to the new image, thus
completing the edge detection process. If the pixel intensity where f is the image, t the template and summation is over
value was above our pre-determined threshold value then its positions (x,y) under the template positioned at (u,v).
position carried a “1” within the binary matrix else a “0” was Expanding the squared version of (1) yields:
inserted into its position. The result of our detection algorithm
is a binary representation of the original image downloaded,
shown in Figure 4, which becomes the prerequisite to the
tracking algorithm.
(2)

With the first and third terms being approximately constant,

the remaining term represents the measure of similarity
between the image and the template. The cross correlation is
given by:

(3)

A larger correlation value (c) results in a greater similarity

Figure 4: “Sobel” edge detected image between the image and template. For our system, a maximum
correlation exists when the image pixel and the corresponding
template pixel have matching values. The pseudo-code for the
III. TEMPLATE MATCHING template matching algorithm is shown in
Fig. 6.
This section describes our template matching algorithm. Once
the binary signature of the downloaded image is created, we
began tracking the target through template matching. Template load template from library
open file from previous
matching is a standard cross correlation technique commonly
function
used to locate objects or patterns within an image. The loop_img: scan entire image
template can represent the smallest detail within an image or it loop_templ: scan templ window
can represent an entire object. Our templates were created by compare pixel values
extracting image specific details from prospective tracking count correlation matches
targets. In order to track the target within the image, the calculate probability
template is filtered through the edge detected image using a end loop_templ
“nearest neighbor” technique, where a mathematical end loop_img
correlation is determined during the ”pixel by pixel”
convolution.

Figure 6: Psuedo-code for template matching.

The result of our target detection algorithm is a new image

 matrix containing “regions of interest” (ROI), which represent
possible target locations within the original image. Each value
Figure 5: Template library images.
within the new ROI matrix represents the calculated
probability of a match between the binary image and the
template as it passes over the image in that location. Table 1
This correlation is based on the theory of Euclidean Distance, shows an exact template match at row 144 column 298 of the
given by: ROI matrix, which indicates target recognition in that location
of the binary image. The ROI matrix is the prerequisite to the
classification stage of our target recognition system.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on August 18,2010 at 13:36:00 UTC from IEEE Xplore. Restrictions apply.
 system. The “Sobel” edge detection algorithm filtered the
Table 1: Probability calculations for ROI matrix image of the penny returning its grey-level representation.

IV. TARGET CLASSIFICATION

The final stage for our ATR system is object classification.

Object classification is typically implemented according to the
matching methodology of a particular template matching Figure 7: Maximum ROI within ROI matrix
system. Adaptive target recognition distinguishes its target
using a statistical model of the background [11] built by using After which, the template matching algorithm initiated a pixel
a set of texture filters. Each pixel is filtered at the sensor input by pixel convolution of the “binary” penny. The image was
where a statistical test is applied to determine whether it is a filtered by each template in the library. The template matching
part of the background or a potential target. In [8] the processes resulted in a new correlation matrix. The new image
approach to classification performs target recognition by matrix represents various locations within the penny, which
representing target models and images as a set of oriented edge contain “Regions of Interest” (ROI). The ROI’s within the
pixels. The use of edge maps to model objects allows small image represents the accumulated match probability, which
irregular shaped targets to be accurately modeled; reducing the helps to determine the object’s identity. Finally, the target
probability of “false-hits.” Classification can be accomplished classification algorithm determined which region of interest
for both 2-D and 3-D objects using this system. We developed contained the highest probability. A target identification
our classification algorithm to recognize the target based on a statement, “This is a Penny,” was displayed after verification.
maximum likelihood estimation of the regions of interest The total processing time per uploaded image was
(ROI) within the new matrix. These probability based regions approximately 5-8 minutes.
are determined by the matching correlation, which was
calculated one template window at a time during the template
matching stage. This idea is based on a ROI method presented
in Patil et al [1], where potential target frames were
overlapped and merged to generate particular regions of
interest within a sequence of images. A maximum ROI
probability results when a match exists between the image and
template value for each pixel location within the convolution
window. An example of a maximum probability (%100) ROI,
which resulted in object identification, is shown in Figure 7.
When this occurs, the image is identified as the intended target
and the user receives a “match” alert. Otherwise, a “no-match”
alert is displayed and the system resets for the next image
download. Figure 8: Template extraction (penny-nose)

V. ATR SIMULATION AND EVALUATION

VI. CONCLUSION
The simulation of our Configurable Automatic Target
Recognition system begins with the creation of the template Each simulation yielded expected results. Our configurable
library. Four JPEG images were downloaded from the internet; Automatic Target Recognition system accurately determined a
dime, nickel, penny, and a quarter. Three of the four images %100 match for each experiment, while successfully
(nickel, penny and dime) were subjected to “Sobel” edge determining a “no-match” for objects unrecognized. We have
detection, which produced a 510 x 510 binary image for each developed a configurable automatic target recognition system
JPEG. A 147 x 73 template matrix is extracted from a specific
for recognizing objects and patterns that utilizes various
location (the nose) of each binary image, shown in Figure 8,
techniques (Sobel edge detection, grey-level template
creating the template library. To test our algorithm, two of the
matching, and maximum likelihood estimation) to accomplish
four images (penny, quarter) were, separately, uploaded to the
robust target recognition for all instances of still imagery.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on August 18,2010 at 13:36:00 UTC from IEEE Xplore. Restrictions apply.
 VII. FUTURE WORK Transactions on Neural Networks, Vol.13, No.6, November
2002
We intend to continue with the development of this [8] Clark F. Olson, Daniel P. Huttenlocher, Automatic Target
approach, using a more complex template regeneration scheme Recognition by Matching Oriented Edge Pixels, IEEE
that will improve template matching efficiency and accelerate Transactions on Image Processing 6, January 1997.
image recognition. To expedite classification, the matching
efficiency of the template will be continuously re-evaluated as [9] W. M. Brown, C. W. Swonger, A Prospectus for Automatic
Target Recognition, IEEE Transactions on Aerospace and
the template is filtered through the image. If the match Electronic Systems, Vol. 25, No. 3, May 1989.
correlation of the new template is greater than its predecessor
it becomes the search template. Each new template will [10] R. Brunelli and T. Poggio, “Face Recognition: Features versus
represent a better search pattern, which will result in more Templates”, IEEE Trans. Pattern Analysis and Machine
accurate object recognition in less time. We expect that this Intelligence, vol. 15, no. 10, pp. 1042-1052, 1993.
will lead to even better recognition rates. We also intend to
[11] K.Messer, D. deRidder, J.Kittler, “Adaptive texture
investigate the incorporation of a conversion application that representation methods for Automatic Target Recognition,”
can convert Matlab to VHDL, which would allow our system Center for Visions, Speech and Signal Processing, Department
to be implemented on a Field Programmable Gate Array of Electrical and Electronic Engineering, University of Surrey,
system. Guildford, Gu2 5xh, United Kingdom.

ACKNOWLEDGEMENTS
This material is based upon work supported by the
USAF/AFRL at Wright Patterson AFB Sensors Directorate
under Contract No. FA8650-05-D-1912. Any opinions,
findings and conclusions or recommendations expressed in this
material are those of the author(s) and do not necessarily
reflect the view of the USAF/AFRL, Wright Patterson AFB

REFERENCES
[1] R. Patil, P.E. Rybski, F. de la Torre, C. Vallespi, M. Vallespi,
and B. Browning, “People Detection and Tracking in High
Resolution Panoramic Video Mosaic,” in Proc. IEEE Conf. on
Intelligent Robots and Systems, 2004, pp 1323-1328.

[2] T. Clarkson, “Automatic Target Recognition using Neural

Networks,” Neural Network World, vol 5, No. 6, pp 861-871,
1995.

[3] J.P. Lewis, “Fast Normalized Cross-Correlation,” 1995

[Online] Available:
idiom.com/~zilla/work/invisionInterface/nip.pdf

[4] A. Goshtasby, S. H. Gage, and J. F. Bartholic, ``A Two-Stage

Cross-Correlation Approach to Template Matching'', IEEE
Trans. Pattern Analysis and Machine Intelligence, vol. 6, no. 3,
pp. 374-378, 1984.

[5] J.Saraiva, L.P.C. Bandeira and P.Pina,”A Structured Approach

to Automated Crater detection,” CVRM/Centro de Geo-
Sistemos, Instituto Superior Tecnico Av. Rovisco Pais, 1049-
001 Lisboa, Portugal.

[6] C.F Olson,” Maximum Likelihood Template Matching,” Jet

Propulsion Laboratory, California Institute of Technology, Proc.
Of IEEE Conf. on Computer Vision of Pattern Recognition,
2000.

[7] M. S. Bartlett, J. R. Movellan, T. J. Sejnowski, Face

Recognition by Independent Component Analysis, IEEE

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY WARANGAL. Downloaded on August 18,2010 at 13:36:00 UTC from IEEE Xplore. Restrictions apply.