DEVELOPMENT OF AN AUTOMATIC TARGET RECOGNITION SYSTEM
Document Type:
Collection:
Document Number (FOIA) /ESDN (CREST):
CIA-RDP78B04770A000200020022-5
Release Decision:
RIPPUB
Original Classification:
K
Document Page Count:
50
Document Creation Date:
December 28, 2016
Document Release Date:
October 28, 2004
Sequence Number:
22
Case Number:
Content Type:
REQ
File:
Attachment | Size |
---|---|
![]() | 1.72 MB |
Body:
25X1 Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved, For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
1
DEVELOPMENT
r
25X1
1
Prepared by:
25X1
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
211
This unsolicited technical proposal is submitted by
This proposal, TP?288,
describes a two-year research and development program which
will, at its completion, yield an operational prototype of an
adaptive, automatic, target-recognition system.,
Phase I of the two-phase program will be a combined re-
search, investigation, and experimental program devised to
establish the basic design parameters and the related perfor-
mance characteristics of the operational system. Phase II will
be devoted to the detailed design, construction, and testing
? of the prototype system.
a prime contractor in the areas of
research, development, and fabrication, works within the realm
of the physical sciences to originate new concepts and improve
existing techniques in military weapon and countermeasure
development. Military electronics and electro-mechanical
equipment are l (major products.
Although the Company is qualified for operation in the
2 1 1
small business category, reputation in the research
and development field has been gained in the fulfillment of
contracts that were acquired competitively and were awarded
solely on the basis of technical originality and
excellence o
Approved For Release 2004/11/30 : CIA-RDP78B0477OA000200020022-5
1
11
I
t
1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
The Company is located atI f only
a few minutes from Washington, D. C. Research efforts are
greatly facilitated by proximity to both ASTIA and the
Library of Congress.
occupies a functionally designed facility with
40,000 square feet of floor space. Engineering offices and
laboratories, and fully equipped model, prototype, and fabri-
cation shops are contained within the building's six bays,
employs approximately 200 persons,
of whom more than fifty percent are professional and other
technical personnel. The Company?s organizational structure
expedites the carrying-out of difficult technical assignments
,by minimizing administrative detail and by providing the best
available engineering support services. `
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
25X1
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
PAGE
I. INTRODUCTION AND SUMMARY . . . . . . . . . .
.
1
II. RESEARCH AND DEVELOPMENT CONSIDERATIONS . . .
.
6
A. THE OPTICAL SCANNER . . . . . . . . . . .
.
6
B. THE VIDEO PROCESSING SYSTEM . . . . . . .
.
11
C. THE ADAPTIVE RECOGNITION SYSTEM . . . . .
.
15
STATUS OF CURRENT INVESTIGATIONS . . . . .
D
.
19
.
1 E. SUMMARY . . . . . . . . . . . . . . . . .
.
22
III. PROPOSED PROGRAM . . . . . . . . . . . . . . .
.
23
IV. PROGRAM PERSONNEL . . . . . . . . . . . . . .
.?
33
ILLUSTRATIONS
1
FIGURE
PAGE
1 AUTOMATIC TARGET RECOGNITION MODEL . . . . . .
.
3
2 IMPLEMENTATION OF SUCCESSIVE INTEGRAL SCANS .
.
9
3 PROPOSED TIMETABLE . . . . . . . . . . . . . .
.
32
I
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RD078B04770A000200020022-5
r
1
In the past several years, a number of investigations
have been made of the problem of automatic pattern-recognition.
One of the more difficult tasks to be accomplished in this
field, one in which the patterns to be recognized are in the
form of optical images, is that of automatically identifying
targets on aerial photography. Limited forms of image proces-
sing and subsequent adaptive recognition processes have, to
this point, produced encouraging results.
The dual tasks of the proposed program are: (1) evaluate
the techniques previously explored to establish which forms of
processing can be most successfully used and (2) make use of
state-of-the-art hardware techniques in the design and construc-
tion of an adaptive, automatic target-recognition system.
The general approach to all problems of this kind involves
accomplishing two major tasks: (1) extracting measurements
from a data source characterizing the patterns, and (2) building
a statistical model of each of the classes of interest based on
these measurements. Data from new patterns can then be com-
pared with the stored models, and, if the similarity is within
prescribed limits, a classification can be made.
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Data sources such as aerial imagery have presented a
formidable problem since the total information content for
typical scenes is extremely high. At a resolution of twenty
lines per millimeter, it is possible to define 40,000 discrete
resolution elements in one square centimeter of an aerial
.photograph.
.Because this number is larger by several orders of magni-
tude than the number of measurements that can be dealt with
by conventional pattern-recognition methods, fewer measurements
must be supplied to the system. In part, reducing the number
of measurements supplied to an adaptive recognition system is
justified by the fact that images under translation and rota-
tion keep the same classification. In successfully implementing
automatic recognition, it is essential to define which measures
'both efficiently characterize the imagery and display minimum
sensitivity to image variables such as translation and rotation.
Once the appropriate set of measurements has been made,
a relatively straightforward adaptive-learning process can be
used to construct the class models, using measurements from
examples of each class. The process consists of accumulating
and comparing data from all examples, setting the appropriate
limits of variation which can be expected and setting up special
subclasses in classes which have more than one basic mode of
input.
We may summarize the basic approach by the simple model
shown in Figure 1. The input or data source (which may be film)
is scanned, and the video data is supplied to a special elec-
tronic processing system. In the processor, the video data is
Approved For Release 2004/11/30 : CIA-RDP78BO4'770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
INPUT
( FILM )
SZ
OPTICAL
SCANNER
VIDEO
SIGNAL
VIDEO
PROCESSOR
IMA
ADAPTIVE
RECOGNITION
SYSTEM
Ivi
E
TORS
OUTPUT
(CLASSIFICATIONS)
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
11
transformed into a significantly smaller number of measure-
ments describing the imagery. The adaptive recognition system
will provide, once it is trained with measurements made on
suitable examples of classes of imagery, the required classi-
fication data on new inputs.
The diverse methods of scanning, processing, and recogni-
zing the aerial imagery will be evaluated during the design
effort of the first phase of this program. The data resulting
from programs under current sponsorship, as well as from a
number of previous programs carried out at will serve
as the data base for this study.
25X1
has carried 25X1
out studies on pattern recognition and wideband video-signal
recognition under the following contracts:
CONFLEX I, a large capacity, conditioned-reflex system,
was constructed under the second of these contracts. Our.inves-
tigations of automatic target recognition on aerial photography
have been carried out under the following contracts:
These contracts have been devoted primarily to testing the
capabilities of the CONFLEX system and the use of preprocessing
methods for measuring properties of imagery which are not sen-
sitive to translation and rotation. The first phase of research
Approved For Release 2004/11/30 : CIA-RDP78BO4770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
t
t
will also draw upon the investigations carried out at other
laboratories,
25X1
25X1
The combined results of our programs and those of others
will be used to full advantage in insuring a practical and
successful approach to the design of the operational system.
Sufficient experimental work will be included in the first
phase to remove any doubts regarding the essential performance
characteristics of portions of the system.
In the paragraphs which follow, some of the important
facets of research and development associated with each basic
subsystem will be discussed. Finally, a detailed program will
'be laid out showing the areas of concentrated effort for both
phases of the program.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
The optical scanner will reduce input data into the form
used by the processor, and the data will be transduced onto
various roll film formats. At least two forms of optical
scanning will be considered for use in the operational system --
a conventional TV raster scan and the line-integral scans under
current study.
A typical point-by-point raster scan of pictorial data
clearly supplies all of the information necessary to describe
the scene under process. Were this the only consideration,
,',the conventional TV scan would suffice in accomplishing the
initial data pickup. We must look ahead, however, and realize
that measurements of picture information must be through the
summation of appropriate combinations of individual picture
elements. This summation may be accomplished in several ways,
using the conventional TV scan.
In one case, we may buffer the point-scan information on
the entire scene and, through serial or parallel processing,
effect the necessary weighted combinations necessary for mea-
suring the pictured information. At the other extreme, we may
choose to operate directly upon the input image, using optical
spatial filtering or area-matching masks. to detect various
measurements. In what might be termed an intermediate form of
transducing process, the scene can be scanned with a multipoint
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
scan having sample apertures arranged to sum image activity
over limited areas. The line-integral scan is one form of
multipoint scan that will be of interest here.
Edge enhancement is the simplest means of processing
point-scan data. Here, an electronic filter with known impulse
response recombines local picture elements to generate a new
video signal. A high-pass filter will produce a new video
signal which carries. information principally on the edges of
images where abrupt changes in contrast take place. The time-
sequence weighting of point-scan data in this manner does not
permit the simple union of adjacent data points which are not
oriented in the direction of scan. Thus, a conventional TV
scan will permit edge enhancement for one orientation only.
Various forms of isotropic scans and rotating TV rasters have
been used to circumvent the direction sensitivity of simple
point-scan processes.
Although the optical scanner will be a physically sepa-
rate operating unit, the design of this unit will strongly
depend upon the video processing methods selected for use in
the system. Since the literature on the more common forms of
point scanning is relatively abundant, no further details of
their structure will be given here. Their applicability will
be discussed in the next section on video processing. Line-
integral scanning will prove to be of interest, and a brief
discussion of this scanning method is included. The reader
is referred to the final report prepared by
under for details concerning the
development of the integral scanning system described here.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
25X1
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Figure 2 shows a schematic view of a line-integral scan-
ning process. Backlighted sample slits are positioned at
regular intervals around a cylinder, and, in the center of
the cylinder, a 45-degree rotating mirror directs the light
from successive slits through the objective lens. An image
of each slit is thereby swept across the photographic trans-
parency under process. Each successive sample slit is rotated
by an amount equal to its angular separation on the sample
slit cylinder. A condensing lens collects the light and
focuses it onto a multiplier phototube which transduces the
light signal. The video waveform obtained is a series of video
pulses; each pulse is made up of the series of line integrals
sampled as the slit sweeps across the picture. The simple
implementation shown here can be replaced with flying spot or
vidicon hardware with appropriately programmed scans.
When a line-integral scan is used, a portion of the image
combining process is carried out quite easily in the initial
stage of processing. Further advantages of this form of scan
will be discussed in the next section.
In addition to the constraints imposed on the optical
transducer by the form of processing selected, the format of
the incoming data must be considered. It is anticipated that
the input will be rolls of aerial film ranging from five to
nine inches in width. Our scanning process must accommodate
these formats and maintain a reasonably fast processing rate.
A system is envisioned in which the roll film is in continuous
motion, providing a coarse x-scan while an oscillating or rota-
ting mirror(s) sweeps the center of attention back and forth
across the film in the y direction.
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Ml Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
OPAQUE
BACKLIGHTED
CYLINDER
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
An optical system for integral scanning, as shown in
Figure 2, would perhaps be employed in conjugate form. An
image of the scene under process would be brought via the
coarse scan mirror just described to the rapidly rotating
mirror required for producing the video signals. Phototubes
behind each of the sampling slits (or an optically abbreviated
equivalent) could be used to generate the line-integral scan
data as the image sweeps by. In such an implementation, there
is available around the sample cylinder a moving image which
makes one full rotation as it sweeps around the cylinder. If
a sample slit is replaced with a small aperture, then a point
scan can be had which scans in any desired direction on the
field.
The film transport system required will be quite simple.
The usual loose coupling drive will be provided for feed and
take-up spools while the film speed is governed by friction
drive on the film itself. It may prove desirable to sense the
interframe spacings for purposes of frame count and to inhibit
attempts to classify the severed imagery. No difficulty is
expected in implementing a scanning system meeting the require-
ments dictated by the study program.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
The heart of the proposed automatic target-recognition
system will be the video processing system. Here, the thou-
sands of data points defining the actual scene must be re-
duced to a much smaller group of measurements which efficiently
describe the imagery. As pointed out in an earlier discussion,
edge enhancement is one of the simpler forms of processing used
on imagery; however, the transformation is one-to-one and does
not reduce the total number of measurements that define the
imagery. We have also said that a useful set of measurements
would be minimally sensitive to translation and rotation;
this, in fact, implies a nonlinear transformation of a many-
to-one nature. That is, the image of an airplane, say, would
yield the same set of measurements regardless of position or
,orientation on the scene.
The most promising form of processing thus far uncovered
to minimize the effects of image translation and rotation re-
lates to the integral scanning process described in the pre-
vious paragraph. If we imagine that the video signal is obtained
as a long, narrow slit sweeps across some image, it is evident
that the signal waveshape will be unchanged as the image is
translated on the field. Let us further assume that our video
signal is obtained by a sequence of integral scans in which
the slit orientation is advanced through small angles between
each scan until a complete circle is traversed. Under these
conditions, the effect of image rotation is to shift the
phase of the periodic sequence of video signals without chang-
ing the individual waveshapes. The net result of the sequence of
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
integral scans is a video signal relatively insensitive to
translation and rotation of the imagery.
The development of a video processor which can extract
data from the integral scans has been the subject of the cur-
rent research efforts. In principle, the techniques under
investigation are similar; each attempts to establish a set of
primitive waveshapes with which the incoming signals can be
compared. The integral scan sequence provides a number of
discrete video pulses (equal to the number of sample slits
used) with which all primitive waveshapes can be compared. The
resulting set of measurements can be used to provide a first-
order description of the scene. By using a subsequent level
of processing, we can take into account the time sequence of
measurements obtained from each type of comparison; the time
sequence will be periodic with a period equal to the time re-
quired for a complete circle of scans. This signal can be
characterized independent of its phase, or equivalently, inde-
pendent of rotation of the imagery. The primitive filters
under study include bandpass filters which make measurements
of spatial frequency, and tapped delay lines to implement
weighting functions which can be used to measure correlation
with arbitrary waveshapes.
Through the use of line integral scans and filters with
arbitrary weighting functions operating on the video signals,
the necessary two-dimensional integrals are implemented on the
aerial scene. By neglecting certain phase characteristics in
our measurements, the resulting parameters can be made essen-
tially indifferent to translations and rotations of the imagery.
Approved For Release 2004/11/30 : CIA-RDP78B0477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
1
The video signals derived from point scans are not so
easily processed for complex two-dimensional integrals. They
can, however, be useful in deriving the characteristics of
various isotropic density distributions such as those corre-
sponding to wooded areas and other natural areas which are
free of structure that has a preferred direction.
The specifications of the video processing portion of
the automatic target recognition system will dictate the
classes of two-dimensional weighting functions to be applied
to the scene under process. The particular measurements repre-
sented by these integrals must efficiently describe the imagery,
must permit the required separation of classes of interest,
and must incorporate the types of many-to-one transformations
of image data which minimize the effects of translation and
rotation. Although explicit mention of other image variables
such as aspect, illumination, seasonal changes, contrast and
so forth has not been made, it is important that the scanner-
processor combination maintain descriptive measures of the
imagery in the presence of such changes.
The total number of'measurements which the processor must
generate and send to the adaptive recognition system will
depend upon the signal-to-noise ratio expected on the incoming
imagery. By signal-to-noise ratio in this context, we mean the
ratio of signal from the object being sought to the total sig-
nal of the surround within the purview of the sample aperture.
Noise can also be associated with unexpected structural varia-
tions in the object being sought which distort the set of
measurements describing that object. The complexity of image
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
description for signal and noise through various transforma-
tions permits only qualitative estimates of these values.
Since overall system performance must also be based on the
requirements of the recognition system, continuation of this
discussion is deferred to the latter part of the next section.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
I
The sets of measurements which define the various classes
of imagery presented to the scanner-processor must be stored
in a form appropriate for later comparison with test imagery;
this is the role. of the adaptive recognition system. It is
a special-purpose computer which takes in sequence the sets of
measurements on.each class, compares them with previously
received data, and modifies or adds to its memory the infor-
mation on the new example.
The CONFLEX I, a general-purpose pattern-recognition
system, was built nearly three years ago. New ideas have been
introduced since that time, and the application of some of
the concepts utilized in CONFLEX to a specific problem warrants
the use of different processing techniques. For the reader
who is not familiar with the existing CONFLEX system, a brief
description has been included in an appendix.
The structure envisioned for the adaptive portion of
the automatic target recognition system is similar to that of
CONFLEX. The measurements from the processor will be combined
in various ways in summing circuits and clipped to simple
binary or ternary values by threshold circuits (D-cells).
The most likely departures from the CONFLEX design will be in
the manner of selecting the combinations of measurements prior
to the clipping process, in the introduction of special pro-
cessing of the bits forming the D-field as originally defined,.
and in the sequence of processing used during learning.
Approved For Release 2004/11/30 : CIA-RDP78BO4770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
In CONFLEX, the combinations of inputs to the D-cell are
selected arbitrarily; a pseudo-random number-generator sets
switches in a cross-bar matrix to which the input signals are
attached. As.CONFLEX was designed for general-purpose pattern
recognition functions, no a priori assumptions of the pattern
structure were made. However, if we are given the structure
of the preprocessing hardware and examples of typical imagery
to be encountered in operation, the statistics of the measure-
ments to be obtained can be estimated empirically. From these
data, it is possible to design the sampling process to be in
some sense optimized for the sets of measurements to be classi-
fied. For instance, a general rule regarding the information
content of a set of measurements is based on the mean, variance,
and covariance statistics over the range of inputs to be ex-
pected. If, over the gamut of possible input imagery, the
measurements have strong covariance terms, then a sampling
procedure may be selected which will minimize the effect of
these terms. In the simplest case, this may be the removal
of the DC term of unipolar measurements such as those asso-
ciated with the density spectrum.
The details of the sampling process just described can
be related to the manner in which the measurement space is
initially partitioned. Since a D-cell output in its simplest
form is a one or zero output according to the weighted sum of
its inputs, the output tells us in which half of hyperspace
the input pattern is located with respect to a hyperplane
defined by the weighted sample. Prescribing the D-cell inputs
can partition measurement space so that a uniform separation
capability is obtained for the range of the expected inputs.
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
t
The special D-field processing prior to storage in the
adaptive learning process can carry out two functions. The
first of these is essentially a recoding of the partitions
of measurement space established by the D-field. In the limit
it is possible to assign one of a set of orthogonal codes to
each possible D-field representation. This will make possible
the arbitrary classification of a group of inputs provided
only that distinct inputs produced distinct D-field responses.
The drawback of such a system is that every expected partition
must be preassigned its correct classification; this preassign-
ment is similar to constructing an AND gate for every D-field
and ORing the gates associated with the same class. By judi-
ciously choosing the initial form of D-field processing,
recognition of the various classes of imagery throughout
reasonable ranges of variation and on the basis of small num-
bers of training examples can be achieved.
The second important process is the allocation of memory
for statistical models which represent classes and subclasses
of imagery. In practice, the D-fields for each of several
classes of imagery will have common vector components. The
rules by which a complete separating function are constructed
must be formulated or adapted according to the characteristics
of the problem. If one class of imagery is represented by sets
of measurements with two distinct groupings or modes, the indi-
vidual modes, in general, will each require a portion of memory
for the statistical model. In a particular problem, however,
the vector sum of these two modes may be distinct and nonambi-
guous, thus requiring a single vector representation in memory.
It is evident that an ideal solution can not be reached until
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
1
r
the complete set of examples has been examined. The purpose
of the final stage of processing in the adaptive recognition
system will be to recheck the criterion for class separation
with the addition of each example. In this way, the system
can adapt to classify patterns of measurements with both wide
and narrow separation as long as the separations are within
the partitioning of the D-field structure.
in the final analysis, we are concerned with the perfor-
mance of the system in a real problem. Performance can be pre-
dicted when the statistics of the measurements are known for
the classes of imagery to be recognized and for the noise or
surround environment in which they must be found. We must
finally arrive at the number of modes of input to be recog-
nized, the variations to be expected and the resultant separa-
tion which can be achieved in the D-field representations.
The first phase study program will be concerned in large mea-
sure with the accumulation of empirical data on which design
parameter and performance characteristic data can be based.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
25X1
1
2111
25X1
I
1
has studied two facets of image processing, the
line integral scan (formerly termed "conical transform") and
the effects of image variability on recognition.
Under recently completed and current contracts,
The line integral scan was studied under Air Force con
recently completed, program effort was
directed toward effecting preliminary evaluations of the inte-
gral scan, bandpass-filter form of processing system. In these
experiments, 30 sets of prenormalized data that represented
TITAN site imagery were used. Although the crude breadboard
constructed during the first study program was used in taking
,,the experimental data, the experiments produced encouraging
results. Correct identifications were made of all the training
examples presented to the CONFLEX system and of 87 percent of
Under a current Air Force contract, is continuing
experimentation related to the line integral scan. Ten tapped
delay line filters for measuring the integral scan signatures
have been constructed. The hardware has just been completed
and will soon provide valuable information regarding the pro-
cessing capabilities of the filters. Each filter is equipped
with variable threshold gates and gated counters to tally the
statistics of the resulting measurements.
Including unknowns at orientations different from those used
in training.
25X1
25X1
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
1
The other facet of image processing, the effects of
image variables which affect recognition, is being considered
under a CONFLEX study program We
have, through controlled experiments, tested the effects of
the following variables in addition to the variables of trans-
lation and rotation:
1. Scale
2. Background Noise
3. Aspect Angle
4. Illumination Angle
5. Image Contrast
6. Background Contrast
7. Resolution
The recognition performance characteristics of the CONFLEX
system were tested with each variable under independent control.
Over 400 examples of each test scene were prepared for use in
this program. In general, the results demonstrated that trans-
lation and rotation have the strongest detrimental effect on
system performance. In these experiments, the CONFLEX was
tested with the imagery directly. We expect the prenormalized
data to be even less sensitive to some of these variables, but
the experimental work has yet to be carried out. Part of this
work will be executed under the current sponsor program. It is
expected that further testing will be required in the proposed
program to establish and evaluate the operational system speci-
fications.
25X1
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Under the current sponsor program, the large scale pre-
normalizing system has just been assembled. During the pre-
liminary phases of this design, our confidence in the proposed
scheme was reaffirmed by the use of simple modeling procedures
using computer simulation. The system, as it now stands, will
execute 51 integral scans on the viewed area and generate 400
prenormalized variables which interface with the CONFLEX I
through an electronic sensory field. Testing of the system will
include some of the image variable work just described which
progresses from simple simulated imagery to very low signal-to-
noise target detection tests. Thus far in the program, the
principal gains. have been experienced along hardware lines,
especially that for implementing the scanner processor system.
At least three months of testing will be conducted which will
give broad insight into the performance of the prenormalized
variables.
In each of the programs described above, an independent
study effort has been made with the inclusion of experimental
tests. The important study phase of the proposed program will
be used to combine and extend this work.
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
The major factors to be considered in the design and
development of an automatic target recognition system have
been briefly discussed in this technical proposal. The basic
elements of the system include the optical scanner for trans-
ducing data from the input film, the video processor for re-
ducing the number of measurements to be processed,'and the
adaptive recognition system.
The scanner-processor portion of the system will be two
physically separate units with the joint functions of trans-
ducing data from the input film and providing the set of measure-
ments which describe the imagery. The principal design consi-
derations have to do with the exact form of the aperture or
apertures which scan the film image and with the types of
filters used to combine picture elements for purposes of making
the measurements.
The first or study phase of the proposed program will be
concerned with the utility of available techniques and with
selecting suitable ones. During the second phase, the design
and construction of the operational system will be carried out,
using off-the-shelf hardware components, insofar as it is
possible.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
The goal of this program is to develop a prototype of
an operational, automatic target-recognition system. The
proposed system will automatically process all standard width
films, and the system's output will include the frame (and
location thereon) in which classes of interest are found. The
system will have the capacity to search on 10 classes, simul-
taneously, and the classes will be easily and quickly program-
mable.
We believe that a successful program to develop an opera-
tional target-recognition prototype must include a thorough
study of current image-analysis technology. For this reason,
the proposed program is divided into two distinct phases. The
first phase (which will consist of research, evaluation, experi-
mentation, and establishment of preliminary design) will be
conducted to fix the design parameters and verify the associated
performance characteristics. The second phase will be devoted
to the design, construction, and testing of the prototype sys-
tem. Further details are given in the following paragraphs.
A two-year program to execute the entire development effort
is contemplated. The timetable for the tasks described below
is given at the end of this section.
t
Approved For Release 2004/11/30 : CIA-RDP78B0477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
1
The nine-month study phase will be divided into five
tasks that will be carried out partly in series and partly
in parallel. These areas can be termed (1) research, (2) ex-
perimentation, (3) technique evaluation, (4) preliminary
design, and (5) theoretical system evaluation.
Task 1
The first task will be to collect detailed information
on current work pertinent to the automatic target recognition
problem so that the overall design decision can be made with
full knowledge of all available techniques. The principal
objects of study will be the transformation(s) implemented
by the scanner-processor and the various organizations of adap-
tive learning systems. In the case of the scanner-processor
combination, attention will be given to the classes of two-
dimensional integrals afforded by various systems. Point and
aperture scans will both be considered in the light of various
processing schemes. Strong consideration will be given to the
line integral scan being tested under.a current sponsor pro-
gram. A rotating TV-type scan used in investigations by
Litton also represents a means of sampling the aerial scene
in a manner which lends itself to meaningful types of proces-
sing. The subsequent filtering and detection techniques which
will be considered are the bandpass filter scheme under current
study, time domain filtering using tapped delay lines, and
other easily implemented descriptors such as those under inves-
tigation by other research groups.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
The structure of the adaptive recognition system will
be based on the CONFLEX system and on recent innovations that
deal with problems not solvable by normal linear separation
techniques. A principal research task in this area is the
generation of algorithms for creating subclasses to handle
multimodal class distributions. These may be necessary when
objects of the same class are represented by sets of measure-
ments which occupy separate and disjunct regions of the deci-
sion space. General classes of vehicles, light manufacturing
areas, ships, and aircraft are among those which can have dis-
tinct subclass categories.
Finally, under Task 1, we will, in collaboration with the
sponsor, investigate the classes of imagery of interest and
select those most pertinent to operational objectives. We
anticipate that the usual strategic classes of imagery will
be of interest, including airfields, industrial or urban areas,
missile sites, radar complexes, tank farms, shipyards and areas
involved in light manufacturing. Under the tactical class of
imagery will fall tanks, artillery emplacements, troop emplace-
ments, trains, vehicles and ships at sea. We recognize the
importance of having special classes which can be rapidly adap-
ted to any arbitrary object or complex feature of interest,
and this capability will be included in the system design.
In summary, the first task will consist of compiling the
available techniques and operational objectives, selecting
those techniques which hold promise in application to the sys-
tem, and ferreting out special problem areas or gaps which need
attention. Also, the classes of imagery to be treated will be
examined.
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Task 2
Task 2 will consist of using available scanner and
prenormalizing equipment and a general-purpose computer to
gather empirical data on typical types of imagery. This
data will be used to produce the statistical data necessary
for the evaluation of specific scanner-processor systems.
Requirements for additional hardware during this study
phase will be limited to some digital recording equipment
for rapid data acquisition from the existing prenormalizer and
scanner. The general-purpose computer will be used: (1) to
simulate the algorithms for learning set forth in the first
task, (2) to generate statistical data on prenormalized measure-
ments, (3) to develop performance characteristics, and (4) to
simulate any other system processes deemed reasonable in the
research program.
This task will have as its goals the testing of the trans-
formations implemented by the prenormalizer-processor combina-
tion and the performance of various adaptive learning structures.
Task 3
During and after completion of the experimental work, an
evaluation will be made to determine the exact forms of scan-
ning, processing, and learning procedures to be used. This
evaluation will be based, in part, on the empirical data ob-
tained during the latter phases of our current programs, partly
on our experimental work in the proposed program, partly on the
experimental work of others, and partly on the results of theo-
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
retical considerations. Regarding the prenormalized measure-
ments, the statistical analysis will permit an evaluation of
their efficiency both in characterizing the imagery and per-
mitting discrimination of the various classes.
In all evaluations, both the theoretical and practical
aspects of implementing the available methods will be compared
with the performance data. We believe that our previous ex-
perience in building scanners, analog processors, and adaptive
learning systems will insure that our choice of methods will
have excellent implementation potential.
Task 4
On the basis of the methods evaluation, a preliminary
design of the entire system will be prepared. At this time,
the scanning method and the forms of processing to be used will
be chosen. Also, the character and number of measures required
to achieve a specific level of performance will be specified.
At the end of this task, the block diagram and functional
specification for the system will be completed. These will
include the necessary system parameters such as optical resolu-
tion, video signal-to-noise ratio, bandwidth characteristics,
number of measurements, adaptive system capacity, etc.
Task 5
Task 5 will consist of the preparation of performance
characteristics which will be obtained with the system design.
All empirical and theoretical measures of performance will be
drawn upon to provide a fair estimate of the system performance.
All trade-offs between time and hardware or performance and
-27-
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
system complexity tentatively set in previous tasks can be
finally established at this time.
Task 5 will conclude the work in Phase I. The results
of this work will be fully documented, in the form of an interim
technical report, at the end of this nine-month period.
The main result of Phase I will be a complete system
specification and a functional design of the system to serve
as the basis for detailed system design and development. The
specification and functional design will represent the inte-
grated results of efforts in automatic processing and
the efforts of other research groups involved in this area of
investigation.
Because a study has not yet been made of the desired
operational system, the critical design parameters can only
be estimated at this time; however, a reasonable judgment
based on previous experience places some very tentative require-
ments on the operational system. We envision a scanner with
a scanned area variable between 0.1 and 1.0 inches at the film,
with approximately three to five hundred line resolution on
the area within the purview of the scanner. Approximately five
hundred well-defined measurements on the scene under process
should be adequate to achieve probability of detection of approxi-
mately 0.9 with a false alarm probability under .01 for a ten
class search. Provisions will be made for approximately ten
modes of subclass storage for each class. The processing speed
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 CIA-RDP78BO477OA000200020022-5
will be approximately one linear inch of film per second, and
video bandwidths on the order of two megacycles will provide
this system speed. Servo speed control may be necessary to
accommodate varying film densities and widths. A small printer
will type the output classifications,associated film position,
and the confidence level of the decision. We must repeat that
these estimates are tentative and cannot be determined accu-
rately before the system study; they are meant only to convey
our present best estimate of these parameters.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
The detailed design, construction, and testing of the
operational prototype system will be carried out in a fifteen-
month program. Five specific tasks must be carried out in
serial fashion. .These tasks include (1) detailed system design,
(2) fabrication and assembly, (3) debug, (4) testing, and
(5) documentation.
Task 1
The detailed design of the system will be made, using the
preliminary design data developed during the study phase. The
design will use commercially available components to the ful-
lest extent. Unnecessary hardware development will be kept to
a strict minimum.
Task 2
The necessary parts procurement and shop fabrication will
be undertaken as soon as the final design has been determined
in Task 1. The assembly and wiring of the system will be com-
pleted under Task 2.
Task 3
A complete checkout of optical and electronic systems will
be made under Task 3. All processing steps will be checked to
make sure that they function as intended in the final design.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release. 2004/11/30 :. CIA-RDP78B04770A000200020022-5
Task 4
During Task 4, the system capabilities will be checked
out, using test aerial imagery. Experimental imagery having
a wide range of variation will be tested to determine the
range over which satisfactory performance can be expected.
Task 5
The results of the program will be brought together in
a final report prepared at the end of the two-year interval.
The emphasis at this time will be on reporting the details of
the system performance and on making recommendations regarding
operational use-of-the system.
Figure 3 is a timetable for execution of the separate
tasks in the two-year program. Slight adjustments may be
necessary in the future to accommodate variations in the anti-
cipated timing. The entire program will be PERT monitored to
assure an efficient and coordinated execution of the tasks.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
SYSTEM EVALUATION
FIGURE 3
PROPOSED TIMETABLE
DEVELOPMENT OF AN AUTOMATIC TARGET RECOGNITION SYSTEM
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release-2004/11/30 : CIA-RDP78BO477OA000200020022-5
Specialized engineering teams at have
developed, through long-term collaboration on numerous research
and development programs, an unusual capacity for cooperative
team effort. This capability is especially valuable on quick-
reaction programs, as is indicated by corporate his-
tory of strict adherence to contracted schedules.
has an exceptionally. low rate of personnel turnover and, once
a technical team is selected, the customer can be assured of
a reasonably 'firm personnel commitment.
25X1
25X1
25X1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
25X1 Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Next 8 Page(s) In Document Exempt
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
applications, including laboratory breadboard models as well
as rugged equipment designed to meet severe environmental con-
ditions. Microwave work, ranging from the small filter cavity
to large antennas, is our specialty.
fabricates hardware for a variety of
I
2#1
I
PRODUCTION CAPABILITIES AND FACILITIES
production facility occupies 10,000 square feet.
The facility consists of four discrete sections: the Machine
Shop, the Wiring Shop, the Photo Laboratory, and the Quality
Control Department. The company maintains all of the test
equipment necessary to conduct government-specified test pro-
cedures, including audio, VHF and UHF testing, as well as a
50-foot antenna test range. Each section in the production
facility is excellently equipped to meet delivery schedules
with a normal safety margin.
WELDING AND SOLDERING CERTIFICATION
0
shop employees are thoroughly indoctrinated in the
specifications governing the fabrication of military electronic
chassis and components. Our shop foreman,
attended and is a certified graduate of the National Aero-
nautics-and Space Administration's Welding and Soldering
Instructors' Course at Huntsville, Alabama. All of our elec-
tronics technicians and assemblers have completed the two-
week certification course that he conducts at and they
are certified to do welding and soldering in accordance with
MSFC Proc. 158B.
25X1
25X1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
This certification is mandatory for all contractor per-
sonnel doing fabrication for NASA; consequently, I Ialso 25X1
conducts classes at NASA's Goddard Space Flight Center for
the personnel of other contractors at GSFC. These classes
are sponsored by the participating contractors with NASA
approval and encouragement.
t
1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
1
I. MACHINE SHOP
211
t
achine Shop performs welding, sheetmetal, and
assembly work, as well as machining. Shop capabilities include
a variety of highly complex setup and operational tasks on
any machine tool or accessory, or on any sheetmetal working
tool or accessory.
Machining (turning, milling, grinding, drilling, and
boring) of both metals and plastics is done to specified dimen-
rances that range from a fairly exacting (?0.005 inches) to
an extremely exacting (?0.0002 inches) precision, with minimum
finishes ranging from 63 to 8 microinches.
sions, planes, or angles. does machine work to tole-
Sheetmetal work (shearing, forming, notching, punching,
and drilling) is done on ferrous and nonferrous metals, when
most of the parts in the assembly or detail are one-eighth
inch thick or less, to minimum tolerances. IIsheetmetal
operations include general sheet layout and handwork for the
production of intricate assemblies, and require a high degree
of manual dexterity.
25X1
25X1
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
211
Arc welding, including heliarc (certified), inert-gas,
and tungsten-arc welding, is performed on ferrous and non-
ferrous materials at0 with dimensions held to exacting
tolerances. Brazing and silver soldering are also available.
II
applies chemical, organic, and metal finishes to
various ferrous and nonferrous materials. Organic finishes
are applied by brush and spray, and are both air dried and
baked dry. Metal finishes, in compliance with MIL specifica-
tions, include magnesium conversion coating (6 x 6 x 6 inches,.
Dow 7 and 9), aluminum conversion coating (etch or nonetch
cleaner, 24 x 24 x 24 inches), and silver plating (immersion)
for copper and copper alloys (6 x 6 x 6 inches).
Quantities of complex electromechanical equipment, such
as flush-circuit switching, commutator-brush assemblies, and
continuous belt-driven, photoelectric, tone-generating disc
units, are assembled in lI Machine Shop.
The remaining pages in Part 1 contain photographs of
our Machine Shop and of typical products fabricated there.
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
II. WIRING SHOP
Wiring Shop, fully experienced in the wiring of
electronic and electromechanical assemblies and subassemblies,
specializes in rugged, subminiature circuit modules. Routine
tasks include the wiring of special-purpose computers and data-
processing systems designed and constructed at Other
Wiring Shop functions include the assembly of harnessing and
cables, and encapsulating. The Shop is equipped with such .
special facilities as ovens, inert-gas chambers, special vacuum
chambers, and pressurized encapsulant guns.
Our technicians are especially experienced in applying
resin encapsulants, foams, and conformal coatings to protect
equipment against stringent environmental conditions. Common
materials used for this purpose are Ecco-foam, sty-cast,
Scotch-cast, RTV silastic, and polyurethane. The capabilities
of this.Shop also include the bonding techniques, encompassing
adhesive bonds such as glass-to-metal and teflon-to-most materials.
Refined. mechanical cleaning and chemical cleaning are
effected by liquid honing and vapor degreasing, respectively.
After preparation by these cleaning and surface treatment,'
small components are routinely painted in-house to military
specifications.
25X1
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
211
211
a
III. PHOTO LABORATORY
Photo Laboratory is equipped to process all types
of black-and-white film and to make 16 x 20 inch contact
prints. The'Photo Lab also processes and reproduces art
work for circuit boards and for tone generator discs and
nameplates, and prepares stencils for silk-screening. Advanced
techniques of controlling the exposure of Ektachrome film have
during the past two years.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
IV. QUALITY CONTROL DEPARTMENT
211
The Quality Control Department at is responsible
25X1
for insuring that products comply with the customer's
specifications. The manager of this department reports
directly to the Executive Vice President. Our rigid quality
control system is actively enforced and is approved as meeting
the requirements of MIL-Q-9858 by the Cognizant Government
Inspection Agency, INSMAT, Baltimore, Maryland. I IQuality 25X1
Control is further certified by Lockheed Missiles and Space
Company and General Electric Company.
Specific procedures for accomplishing the requirements
of MIL-Q-9858 are set forth in Quality Assurance Manual. 25X1
.These procedures provide for incoming and in-process inspec-
tions; calibration of test equipment, tools, and gauges; hand-
ling of defective material; vendor surveillance; and for the-
shipping, receiving, and handling of materials. In-house
workmanship standards are established by the Engineering Stan-
dards Manual and are controlled by Quality Control inspections
and tests.
Inspection and test parameters are determined from the
contract specifications and task statement, and in-process
inspections are performed to assure conformance. Non-conforming
parts discovered during inspection are reviewed by. Engineering
Shop Services and Quality Control. Corrective action is then
taken to eliminate recurring discrepancies. Drafting room
Approved For Release 2004/11/30 : CIA-RDP78B04770A000200020022-5
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
1
procedures strictly dictate that all manufacturing, production,
and engineering changes be incorporated into the basic draw-
ings and require prompt distribution of changes to all acti-
vities concerned.
Quality Control also maintains liaison with both vendors
and customers in order to direct the quality-control effort
in such a manner that a quality product will be delivered at
an economical cost.
I as an extensive inventory of standard
2#1
1
test equipment, including oscilloscopes, signal generators,
and voltmeters.
The fifty-foot antenna radiation pattern test range main-
tained by has the capability of measuring radiation
patterns over a 30 db dynamic range. In addition, has 25X1
the capacity to completely design and test antenna voltage
standing wave ratio (VSWR), and to conduct Government-specified
test procedures, including audio, VHF and UHF areas.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5
Approv4d For Release 2004/11/30: CIA-RDP78B04770A000200020022 !
RESEARCH 8 DEVELOPMENT
IN THE PHYSICAL SCIENCES
Subject:
TP-288, Development of an
Automatic Target Recognition System
25X1
25X1
25X1
25X1
I I is pleased to submit the attached technical
proposal or your consideration. Also attached is our cost
proposal for the complete program. The work described in our
technical proposal represents a logical succession to the work
presently being brought to conclusion under a current program
with the Sponsor.
The total program described in our technical proposal involves
a technical effort of 24 months. This effort is divided into
two phases. Phase I covers a study effort of nine (9) months
and Phase II covers the design and construction of the proto-
type system. The total estimated cost for the entire program
is I land is bid on a CPFF basis. The bid is valid
for ninety (90) days from date of this letter. The total
cost for the program has been broken down into Phase I and
Phase II as indicated on the attached Cost Analysis Sheets.
We look forward to your review of our proposal and invite
any questions or discussion you may have concerning it.
If you have questions concerning technical aspects of the
program, please call F_ I
Falls Church, Virginia. Questions concerning contractual
matters should be addressed to lat the
same number.
Sincerely,
25X1
25X1
25X1
Director, Marketing and
Contracts
VB:HLF:ms
Encls: Cost Analysis Sheets, 5 cys.
Technical Proposal, TP-288, 5 cys.
Approved For Release 2004/11/30 : CIA-RDP78BO477OA000200020022-5