U.S. patent application number 09/985451 was filed with the patent office on 2002-05-16 for a-scan isar classification system and method therefor.
Invention is credited to Richardson, Dennis W., Ryan, Patrick S..
Application Number | 20020057216 09/985451 |
Document ID | / |
Family ID | 23724535 |
Filed Date | 2002-05-16 |
United States Patent
Application |
20020057216 |
Kind Code |
A1 |
Richardson, Dennis W. ; et
al. |
May 16, 2002 |
A-Scan ISAR classification system and method therefor
Abstract
A target recognition system and method wherein only target
amplitude data, i.e., coherent A-scan data, is interrogated for
target recognition. Target aspect angle is ignored within the
angular segmentation of the feature library without degrading
classification performance. Observed signature characteristics are
collected at various aspect angles and through unknown and
arbitrary roll, pitch and yaw motions of each anticipated target
and provided to a neural network as training sets. The neural
network forms feature vectors for each target class which are
useful for valid classification comparisons in all sea states,
especially in calm and littoral waters. These feature vectors are
useful for valid classification comparisons over at least 30
degrees of target aspect angle.
Inventors: |
Richardson, Dennis W.;
(Apalachin, NY) ; Ryan, Patrick S.; (Endicott,
NY) |
Correspondence
Address: |
Whitham, Curtis & Christofferson, P.C.
Suite 340
11491 Sunset Hills Road
Reston
VA
20190
US
|
Family ID: |
23724535 |
Appl. No.: |
09/985451 |
Filed: |
November 2, 2001 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
09985451 |
Nov 2, 2001 |
|
|
|
09434515 |
Nov 5, 1999 |
|
|
|
Current U.S.
Class: |
342/25A ;
342/189; 342/195; 342/90 |
Current CPC
Class: |
G01S 7/417 20130101;
G01S 13/9027 20190501; G01S 7/412 20130101; G01S 13/9064 20190501;
G06V 10/255 20220101 |
Class at
Publication: |
342/25 ; 342/90;
342/189; 342/195 |
International
Class: |
G01S 007/41 |
Claims
We claim:
1. A target recognition system comprising: receiving means for
receiving target amplitude data; interrogation means for
interrogating said received target amplitude data and identifying
said target as belonging to a particular target class in a feature
library, said feature library being segmented by plural aspect
angles.
2. The target recognition system of claim 1 wherein said received
amplitude data is coherently integrated over time.
3. The target recognition system of claim 2 wherein said coherently
integrated data is coherent A-scan data from an inverse synthetic
aperture radar (ISAR).
4. The target recognition system of claim 3 wherein said coherent
A-scan data is coherently integrated over the aperture time.
5. The target recognition system of claim 4 wherein said target
aspect angle is used as an index into a feature library coarsely
segmented by the aspect angle.
6. The target recognition system of claim 5 wherein said coherent
A-scan data is coherently integrated over an integration range
swath centered on the target and is limited to a largest target
defined for said system to image.
7. The target recognition system of claim 1, wherein said
interrogation means uses ISAR A-Scan data from a high range
resolution ISAR radar which has a relative range motion of an
observing platform and target removed.
8. The target recognition system of claim 7, wherein the A-Scan
data is coherently integrated over an ISAR aperture time in an
observed range swath centered on the target, removal of the
relative range motion of the observing platform and target, and
extending only to a largest target for the target recognition
system to image.
9. The target recognition system of claim 8, wherein the coherently
integrated A-Scan data forms a high signal to noise return of the
target's Radar Cross Section (RCS) for each high range resolution
radar cell along a line of sight path incident with the target over
the observed range swath.
10. A method of radar target recognition comprising the steps of:
receiving target amplitude data; interrogating said received target
amplitude data; and identifying said target as belonging to a
particular target class in a feature library, said feature library
being segmented by plural aspect angles.
11. A method of radar target recognition as in claim 10, wherein
said received target amplitude data are collected at aspect angles
through unknown and arbitrary roll, pitch, and yaw motions over a
wide range of sea states, including calm and litoral water
conditions.
12. A method of radar target recognition as in claim 11, wherein a
plurality of said feature vectors are derived from observance of a
corresponding target at a single aspect angle.
13. The method of claim 11, wherein said step of receiving target
amplitude data further develop radar target amplitude data and
aspect angle data, said step of interrogating said target amplitude
data further determines features thereof.
14. The method of claim 13, further comprising: comparing said
features to feature data corresponding to anticipated targets in a
feature library, said feature library being segmented by target
aspect angle ranges and accessible in accordance with said aspect
angle data, and classifying said target based upon said comparing
step.
15. A method of operating a radar system utilizing radar amplitude
data and aspect angle data therein, said method comprising steps of
interrogating said radar amplitude data to obtain features thereof,
and comparing said features to feature data corresponding to
anticipated targets in a feature library, said feature library
being segmented by target aspect angle ranges.
16. A method of radar target recognition based on a set or group of
feature vectors formed from a library of a plurality of anticipated
targets comprising the steps of: obtaining a signature
characteristic of said plurality of anticipated targets, each said
signature characteristic including a radar cross section for a
corresponding anticipated target at each of a plurality of range
resolution cells along said radar's line of sight path; comparing
said signature characteristic of a target against each of said
feature vector groups to classify said target.
17. A target recognition system based on a set or group of feature
vectors formed from a library of a plurality of anticipated targets
comprising: means for obtaining a signature characteristic of said
plurality of anticipated targets, each said signature
characteristic including a radar cross section for a corresponding
anticipated target at each of a plurality of range resolution cells
along said radar's line of sight path; and means for comparing said
signature characteristic of a target against each of said feature
vector groups to classify said target.
Description
BACKGROUND OF THE INVENTION
[0001] 1. Field of the Invention
[0002] The present invention generally relates to radar systems and
radar signature identification methods and, more particularly, to
methods and systems for automatically classifying targets imaged on
Inverse Synthetic Aperture Radar systems.
[0003] 2. Background Description
[0004] Synthetic Aperture Radars (SARs) produce a two-dimensional
(2-D) image (with intensity as a pseudo non-spatial third image
dimension) where one dimension in the image is a range and the
other dimension is an azimuth (or cross range). Range is a measure
of the "line-of-sight" distance from the radar to the target and is
determined by measuring the time between the transmission of a
radar signal to the receipt of the reflected signal (e.g., echo)
from a target. The range resolution may be limited by the
transmitted pulse width so that, for example, a narrow pulse width
may be required to yield fine range resolution.
[0005] In order to produce a relatively fine azimuth resolution, a
large antenna is needed to focus the transmitted and received
energy into a sharp beam such that the sharpness of the beam
defines the azimuth resolution. However, in many instances, a large
antenna cannot be used due to the physical constraints of the radar
platform. In these instances when a large antenna is required but
cannot be used, a synthetic antenna (e.g., aperture) may be used in
order to provide the needed fine azimuth (or cross range)
resolution. In order to provide a synthetic aperture, a radar
platform collects data while moving a certain known distance such
that the distance moved is the dimension of the required antenna,
thereby "synthesizing" the antenna (e.g., synthetic aperture).
Thus, the term "synthetic aperture" is generally applied to the
motion of the radar platform to effectively synthesize a relatively
large array of transmitters and receivers (e.g., antennas).
[0006] Similar to the SAR, Inverse Synthetic Aperture Radars (ISAR)
also produce a two-dimensional image (e.g, range and cross range,
and intensity). However, the ISARs use the motion of the target to
synthesize a large aperture antenna and not the motion of the radar
platform itself. Thus, the term "inverse synthetic aperture" is
generally applied to motion of the target which allows derivation
of information concerning the shape and size of the target.
[0007] More specifically, ISARs are used to obtain information of a
target at long range using the Doppler effect to produce an image
by exploiting the relative motion of a target (typically pitching
of a ship on an ocean surface) to develop an approximation of a
visual profile of the target. In typical ISAR systems, the image
produced is a two dimensional image (i.e., a 2-D Doppler versus
range "picture") of the target, with each image being 512 pixels by
128 Doppler cells. Each range/Doppler pair has six bits of
amplitude (intensity) data for pixel display. The profile image is
developed as shown in FIG. 1 by assuming a center of rotation of
the target and developing height versus range information based on
the Doppler shift of the frequency of the returned signal.
[0008] While such systems have been quite successful in presenting
images to trained personnel for interpretation, several major
drawbacks have been encountered. Specifically, while rotational
target motion need not be large for acceptable results to be
obtained (i.e. about 1.degree. of pitching motion is often
sufficient), ISAR systems cannot produce a profile image if the
target is stationary. Further, accuracy of the profile image
development is also complicated and errors occasioned by other
motions of the target such as roll and yaw. Therefore, the ISAR
systems cannot be calibrated to provide any form of standardization
of the profile images produced.
[0009] Moreover, automatic classification schemes require a
computational "collapsing" of the profile image containing in
excess of sixty thousand pixels, each represented by a six bit
intensity value. As can be understood, the processing of the image
is thus very computationally intensive and relatively slow in
providing a response. For example, one proposed classification
technique requires processing of several hundred images per
classification with each of 512.times.128 pixels requiring 250
operations with 10 instructions per operation or roughly fifty
billion computer operations per performed classification. It can be
appreciated that classification can thus require several minutes of
processing time on a relatively powerful high-speed data processor,
and that processing times of much less than thirty seconds cannot
be provided by processors currently practical for shipboard or
airborne use.
[0010] Trained ISAR operators, using this technology, are able to
identify surface combatants at large standoff ranges from the
produced profile images. Unfortunately, this recognition skill is
not trivial and a human operator must learn how to recognize
distinct classes of ships in an ISAR image, requiring long training
times. Typical training courses consist of an intensive course of
several weeks duration to learn features of an ISAR image on which
a classification may be based. Further, this recognition skill is
easily lost because of the subtleties of differences of ISAR images
and must be constantly reinforced and refreshed in order for the
operator to retain such recognition skills. Some studies have
suggested that the average operator only retains 20% of the ISAR
recognition skill obtained during training after two weeks of
training completion.
[0011] Consequently, automatic ISAR classification techniques have
been proposed. These techniques process the ISAR image data,
applying any number of pattern matching techniques against a
library of known target types. These prior art approaches commonly
apply image processing techniques (such as segmentation, noise
rejection, major axis definition, edge detection, etc.) to enhance
image recognition. However, pattern matching, whether of the entire
image or particular features thereof, is particularly sensitive to
the aspect of the target (i.e., the angle of the target relative to
the direction of observation). Therefore, many images of each type
of target to be classified or recognized must be first obtained,
then maintained in an image library and processed against a given
ISAR image. Thus, both data acquisition and database management
present substantial difficulty in addition to data processing
requirements.
[0012] State-of-the-art ISAR image classification requires
processing data from hundreds of ISAR image frames. Typically, ISAR
image classification requires several billion computations for
several seconds for a single image while classification accuracy is
known to be greatly improved by the processing of a plurality of
images. Currently, there are no off-the-shelf mission computers
with the processing capability to provide reasonably accurate
classification in substantially real time. Consequently, available
resources, in current off-the-shelf mission computers, fall far
short of what is necessary to arrive at satisfactory recognition
results. Thus, there is a need for reducing the number of
computations performed in target classification on state of the art
mission computers.
SUMMARY OF THE INVENTION
[0013] It is therefore an object of the present invention to reduce
the number of computations necessary in target classification.
[0014] It is another object of the present invention to reduce the
time necessary for target classification.
[0015] It is still another object of the present invention to
increase the useable data to the classifier by reducing sensitivity
to other target motions, for example, roll and yaw, and reducing
sensitivity to the absence of target motion.
[0016] The present invention is a target recognition system and
method wherein target amplitude data, known in the art as A-scan
data, alone, is interrogated for target recognition. In the
preferred embodiment method target recognition is on coherent
A-Scan data, i.e., coherently integrated (over the ISAR aperture
time) amplitude versus range data with aspect angle and other
non-pitch target motions generally ignored within a 30 degree range
without significantly compromising classification accuracy. Thus,
the required computer resources for the preferred embodiment target
recognition system are dramatically reduced over prior art systems
while providing greatly enhanced speed of recognition with accuracy
comparable to prior art systems. Also, the preferred embodiment of
the present invention requires substantially reduced computer
resources since the method of the present invention uses the
coherent A-Scan data.
[0017] More specifically, a target recognition system comprises a
receiver for receiving target echoed amplitude data and an
interrogation apparatus (e.g., transmitter) for interrogating the
target for echoed amplitude data and identifying the target as
belonging to a particular target class. The received amplitude data
is coherently integrated over the aperture time, and is preferably
coherent A-scan data from an inverse synthetic aperture radar
(ISAR). In the embodiments of the present invention, A-scan data is
coherently integrated per range cell over an integration range
swath (comprised of range cells of resolution commensurate with the
radar's fidelity) centered on the target (where centering the range
swath on the target involves removal and/or compensation for the
effects of the relative range motion of the observing platform and
target) and is limited to the largest target defined for the system
to image. Moreover, the target aspect angle is used as an index
into a feature library, where features correspond to target
classes, are coarsely segmented by the aspect angle. The A-Scan
data forms a feature vector (or signature) where comparisons to the
segmented feature library yields target class.
BRIEF DESCRIPTION OF THE DRAWINGS
[0018] The foregoing and other objects, aspects and advantages will
be better understood from the following detailed description of a
preferred embodiment of the invention with reference to the
drawings, in which:
[0019] FIG. 1 schematically illustrates the basic principles of
inverse synthetic aperture radar imaging,
[0020] FIG. 2 is a flow diagram of an image processing system
required for classification of ISAR images,
[0021] FIG. 3 is a flow diagram of the image classification system
in accordance with the present invention,
[0022] FIG. 4 is an example of a two-dimensional ISAR image created
by pitch motion, and
[0023] FIG. 5 is an illustration of the differences between ISAR
images and A-Scan data.
DETAILED DESCRIPTION OF A PREFERRED EMBODIMENT OF THE INVENTION
[0024] Referring now to the drawings and, more particularly to FIG.
1, there is illustrated, in highly schematic form, the basic
environment of an inverse synthetic aperture radar system. All
radar systems 1 basically include a transmitter 5 for generating an
electromagnetic wave, and an antenna for radiation and receipt of
the electromagnetic wave reflected by the target, and a receiver 15
for receiving the antenna collected energy and processing the same
in a selective manner. It is, of course, preferred that the
receiver 15 preserve as much information as possible that may be
included in the reflected energy so as to accurately image the
target 20.
[0025] Still referring to FIG. 1, the target 20 is depicted as a
seaborne vessel such that the natural motion of the vessel such as,
for example, pitching about an assumed center of rotation due to
encountering waves, causes differences in relative motion of
portions of the vessel. Thus, structural features farther from the
center of rotation (e.g. higher) will have a higher velocity and
cause greater Doppler shift in the reflected signal such that the
frequency shift in the reflected signal is a relative measure of
the height of the feature reflecting that signal in that range
cell. More generally, the doppler frequency yields resolution along
an axis normal to the axes of target rotation and the radar's
line-of-sight to the target. Thus, the cross axis resolution is
equal to the radar's wavelength .lambda. divided by two times the
angular rotation viewed over the aperture time. Thus, mapping both
the intensity encoded frequency and the range of the returned
signal in two dimensions can provide a relatively accurate image of
the physical profile of the vessel and thus provide for
identification or classification of the vessel at very long range.
This is depicted in FIG. 4.
[0026] However, the accuracy of the profile is only approximate and
other target motions, i.e., roll and yaw and other viewing angles
degrade image quality. Furthermore, substantial experience is
required to recognize configurations of features of the image
produced which, in combination, allow recognition of a target or a
class of possible targets. As discussed above, training sufficient
for this purpose must be intensive while the capability of
recognizing targets is substantially lost over a period of time
comparable to that during which the capability was acquired.
[0027] In order to overcome the shortcomings of insufficient
training and/or skill retention after training, application of
feature extraction or image recognition techniques to ISAR images
have been developed. However, current systems require numerous
processes to be applied to a relatively large number of image
values and thus require an extremely large number of individual
data processing operations.
[0028] More specifically, FIG. 2 shows a generalized flow diagram
of a related system to the present invention that includes image
processing operations capable of successful image recognition using
ISAR data. It should be understood that FIG. 2 may equally
represent a high level block diagram of an image
processing/classification system which implements the steps of the
flow diagram now depicted in FIG. 2. It is further to be understood
that the processing operations of FIG. 2 require large available
computer resources which are beyond the capability of off-the-shelf
mission computers to provide reasonably accurate classification in
substantially real time image recognition and verification.
[0029] The steps of FIG. 2 show a system using a two-dimensional
profile image of the target for classification (i.e., intensity
encoded doppler (height) versus range). At step S22, image
statistics are provided on the incoming digital image received from
an imaging radar (step S20), preferably an ISAR. These statistics
may be, for example, mean and standard deviation of pixel
intensities. At step S24, the image is segmented from the receiver
noise (or background picture noise) based on the image statistics
and, at step S26, is filtered with a threshold routine to reduce
image fluctuations, typically due to sea clutter. As is well known
to one of ordinary skill in the art, image noise and clutter are
the random or regular interfering effects in the data which degrade
its information-bearing quality. In the extreme case, noise and
clutter can eliminate the true radiometric information present in
an image. Thus, noise and clutter removal are important and are
used to improve the accuracy of the extracted data and the
interpretation based thereof At step S28, the extracted target
profile image is then transformed to a new pattern by feature
extraction. Through prior data collections and observances of known
targets, numerous feature vectors are extracted and clustered (a
mathematical procedure for organizing data into homogeneous groups
such that the between group variance of the specified number of
groups is maximized) to form the basis of the classification
library. At step S32, the new patterned feature and cluster data
from a ship class library provided at step S34 are then correlated
to provide a best matching pattern at step S36. As is commonly
known in the industry, correlation is the ability to locate or
match a pattern, sequence, image and the like with a corresponding
pattern, sequence, image and the like which can be taken at a
different time at a different viewing angle, or in this case, from
a ship library.
[0030] As can be seen, many processing steps must be performed in
order to obtain the best matching pattern. These processing steps
typically are resource intensive, thus requiring robust processing
systems. However, as discussed above, off-the-shelf mission
computers currently available do not possess the necessary
processing capability to provide reasonably accurate classification
in substantially real time
Solution Provided by the Method and System of the Present
Invention
[0031] In order to solve the problems of available computer
resources needed for real time ISAR classification, the method of
the present 5 invention uses coherently integrated ISAR A-Scan data
(e.g., one dimensional) instead of ISAR image data (e.g., two
dimensional) in an automatic classification system. By using the
one dimensional A-Scan data, the method and system of the present
invention is capable of decreasing the computational resource
loading by an order of magnitude over prior art systems, and
further increase the speed of classification allowing for
substantial real time classification of targets using processors of
relatively modest processing power suitable and currently available
for shipboard or airborne use.
[0032] In particular, the method of the present invention uses ISAR
A-Scan data from a high range resolution ISAR radar which has the
relative range motion of the observing platform and target removed
(where the same range slice of the target is maintained in the same
relative range cell in the range swath), and is thus coherently
integrated over the ISAR aperture time in a range swath centered on
the target and extending only to the largest target for the system
to image. The use of the coherently integrated A-Scan forms a high
signal to noise return of the target's Radar Cross Section (RCS)
for each high range resolution radar cell along the radar's line of
sight path incident with the target over the observed range
swath.
[0033] Thus, relative radar cross sections (amplitude) along the
line of incidence forms a target signature for that given line of
incidence through the target. This is otherwise known as the
target's "aspect angle" and is the angle of the target's main axis
with the radar's line of incidence. This single unique signature,
when compared to a database of previously observed signature
characteristics provides for target type classification such as, in
the case of ship targets, the ship's vessel class.
[0034] Observed signature characteristics using the method and
system of the present invention are collected at various aspect
angles and through unknown and arbitrary roll, pitch and yaw
motions of each anticipated target. These signature characteristics
are then used as training sets for a neural network, which may be
any generic neural network currently known to those skilled in the
art, such as, for example, MIT's Adaptive Clustering Network.
[0035] The neural network forms feature vectors for each target
class and are useful for valid classification comparisons in all
sea states, especially calm and littoral waters. Feature vectors
may be collected over a range of aspect angles with a given set of
feature vectors being derived from observance of a target at a
single aspect angle. These feature vectors can be used for valid
classification comparisons over at least 30 degrees of aspect
angle. Thus, tracking data is used to index an appropriate 30
degree range of stored sets of feature vectors in the
classification library. Hence, the preferred embodiment technique
is robust in that as little as a single signature at a given aspect
angle can be compared to the feature vectors defined for a
corresponding range of aspect angles to provide accurate
classification results within current state-of-the-art processing
constraints.
[0036] It has been recognized that the above-described processing
of an ISAR image using the method and system of the present
invention substantially accomplishes a collapsing, RSS of the
doppler intensities per range cell, of the Doppler/height
information of the ISAR image into a one dimensional "signature",
and it has been further discovered that target amplitude/intensity
versus range data, alone, may be interrogated for target
recognition as illustrated in FIG. 5. Further, performing target
recognition or classification from amplitude data is substantially
less sensitive to unwanted rotational motions for a given
observance of the target, i.e., roll and yaw, in a pitch imaging
scenario. Moreover, only a relatively small fraction of the data
required for ISAR imaging is required for target recognition.
Accordingly, a recognition or classification result is reached much
more quickly with accuracy at least as good and in some instances
better than ISAR image processing systems.
[0037] Having thus discovered that full image data is unnecessary,
the preferred embodiment target recognition method and system uses
the coherently integrated amplitude versus range data, known in the
art as A-scan data and often available within a radar system having
ISAR capability, for example, for the purpose of centering of the
radar target for Doppler-range tracking. Further, in the preferred
method and system of the present invention, aspect angle may thus
be ignored within a 30 degree range without significantly
compromising classification accuracy. Thus, by reducing the amount
of data to be processed, i.e., A-Scan versus image, the required
computer resources for target recognition are dramatically reduced
over prior art systems. Also, the method and system of the present
invention is preferably used with the APS-147 radar developed for
the LAMPS Mark III Block II Upgrade Program. However, any high
range resolution ISAR radar with available coherent A-scan data
would be suitable for use with the method and system of the present
invention.
[0038] Now referring to FIG. 3, a flow diagram of the image
classification system in accordance with the present invention is
shown. It is well understood by one of ordinary skill in the art
that FIG. 3 may equally represent a high level block diagram of an
image processing system which implements the steps of the flow
diagram of FIG. 2. It is also well understood that the high level
block diagram represents a preferred architecture and provides
enhanced performance to known ISAR systems. Further, as seen
compared to FIG. 2, many processing steps have been eliminated in
FIG. 3. This is due to the use of the coherent A-Scan data which
uses less computation resources thereby decreasing processing
time.
[0039] Still referring to FIG. 3, prior to ISAR imaging mode, the
operator selects the desired target track at step S40 to begin the
ISAR classification process. Once the radar is in the ISAR mode, at
step S41, the A-Scan data begins to be delivered to the ISAR
classifier of the present invention. At step S42, the A-Scan is
filtered with a threshold routing to reduce the fluctuations in the
A-Scan, typically due to sea clutter. In the preferred embodiments,
the noise rejection filter will also provide a coarse target length
estimation used to index into the correct partition of the feature
library at step S45.
[0040] The mission computer communicating with the method and
system of the present invention will use the available track data
and the platform's current position to determine the target aspect
angle (i.e., the view angle of the target) at step S43. Once the
aspect angle has been determined, the information is passed to the
ISAR A-Scan classifier library for library selection (step S45) and
for correlation processing (step S44). Since the track heading may
not be precisely known, the angular resolution of the library
partitions are preferably as large as the expected target heading
errors. The track heading may thus be used as the aspect angle
within the recognition latitude provided by the invention and the
aspect angle used to index access to the feature library further
reducing comparisons to be processed.
[0041] At step S44, the library selection data and the filtered
A-Scan data are correlated, and best match ship class is thus
obtained at step S46. By correlating with the filtered A-Scan data
with the partition of the library that corresponds to the aspect
angle, the search for the best match ship class is reduced.
[0042] Once the correlation has been performed, simple distance
checks may be implemented to verify the ship choice in the
embodiments of the present invention. These distances are
statistical distances between the returned signatures and the
library data, and are not referring to range. If these distances
are too large (outside statistical range), the target is classified
as "UNKNOWN". If the distances are normative, i.e., statistically
consistent minimum distances, the class is added to the histogram,
and after several images have been classified, the class histogram
is presented to the operator in a tabular display. In the preferred
embodiment, a statistical check using multidimensional chi-square
comparisons is contemplated for use with the present invention. It
is well understood by one of ordinary skill in the art that the
normative distance may be any predefined distance; however, it is
realized that the lesser distance from the cluster to the feature
vector (A-Scan data) results in a more accurate classification
(e.g., reduction in misclassifications).
[0043] By way of illustrative example, utilizing the above method,
several clusters and distances from cluster to feature vector may
be determined as represented in Table 1.
1 Distance from Cluster Cluster Number to Feature Vector 1 273 2
221 3 134 4 119 5 252 6 318
[0044] In this example, the method as described above will
determine if any of the above distances are too large in order to
provide an accurate classification. Taking cluster 4 with a minimum
distance of 119 as an example, the method of the present invention
will employ a statistical check (e.g., a multidimensional
chi-squared comparison) to determine if that minimum distance is
consistent with a match. For example, if the distance of 119 is
within the 90.sup.th percentile for the chi-squared distribution
with 512 degrees of freedom, the class that is associated with
cluster 4 would be added to the class histogram. However, if this
distance is too large (e.g., within the 80.sup.th percentile), it
would be classified as "UNKNOWN" even though there is a class
associated with the 4.sup.th cluster. It is to be understood that
the above example is provided for illustrative purposes only and
that other percentile for the chi-squared distributions may equally
be used with the present invention.
[0045] It should be well understood that the method of the present
invention can be implemented using a plurality of separate
dedicated or programmable integrated or other electronic circuits
or devices (e.g., hardwired electronic or logic circuits such as
discrete element circuits, or programmable logic devices such as
PLDs, PLAs, PALs, or the like). A suitably programmed general
purpose computer, e.g., a microprocessor, microcontroller or other
processor device (CPU or MPU), either alone or in conjunction with
one or more peripheral (e.g., integrated circuit) data and signal
processing devices can be used to implement the invention. In
general, any device or assembly of devices on which a finite state
machine capable of implementing the flow charts shown in the
figures can be used as a controller with the invention.
Example of the Method and System of the Preferred Embodiment of the
Present Invention
[0046] The most important metric which quantifies the performance
of an automatic classification algorithm is classification
accuracy, or more precisely, probability of correct classification.
In order to provide for correct classification in the present
invention, a performance baseline was computed by using the MIT
Adaptive Clustering Network (ACN) (See, An Automatic Ship
Classification System for ISAR Imagery, The Lincoln Laboratory
Journal, Volume 6, Number 2, 1993) on a data set representing 22
ship classes containing approximately 7500 training images. A set
of approximately 500 testing images was then used to verify the
accuracy of the algorithm and set the classification performance
benchmark.
[0047] The data was then decomposed into A-Scans and the ACN was
re-trained and retested. Overall the MIT approach yielded a
probability of correct classification of 0.57 and the A-Scan
approach was 0.64. The following table compares the two approaches
(e.g., MIT ACN vs. A-Scan) from a computational complexity
standpoint:
2 Training Testing # of RISC Time to Vectors Vectors Clusters
Instructions Execute MIT 7214 449 38 Unknown 2.806 ACN A- 7214 449
61 11, 767, 575 0.277 Scan
[0048] The number of RISC instructions is expressed as
# RISC instructions=14Nt+26NcNt+60(Nc).sup.2+31Nc+24.
[0049] Nt=number of training vectors
[0050] Nc=number of clusters
[0051] While there is essentially no statistically significant
performance difference between the two methods with respect to
accuracy, an order of magnitude less time is obtained using the
A-Scan computations performed on the same data sets in accordance
with the invention. The advantage of the A-Scan method is thus the
ability to process a larger number of images in the same amount of
time allowing the use of off-the-shelf mission computers.
Specifically, and as a comparison, in the above example, 2.806
seconds were needed to process one image using the MIT approach.
However, the method of the present invention using the A-Scan may
processes 11 images in the same time period.
[0052] While the invention has been described in terms of preferred
embodiments, those skilled in the art will recognize that the
invention can be practiced with modification within the spirit and
scope of the appended claims.
* * * * *