CONTRACT INSPECTION REPORT FOR AUTOMATIC IMAGE RECOGNITION BY COHERENT OPTICAL TECHNIQUES

• Document Type: CREST
• Collection: NGA Records (Formerly NIMA)
• Document Number (FOIA) /ESDN (CREST): CIA-RDP78B04770A002300020007-9
• Release Decision: RIPPUB
• Original Classification: C
• Document Page Count: 12
• Document Creation Date: December 28, 2016
• Document Release Date: March 28, 2005
• Sequence Number: 7
• Publication Date: September 23, 1965
• Content Type: FORM

ENGINEERING SECTION/CB/PD/OL
Declass R
Re new by NGA.
NAME OF CONTRACTOR
25X1
25X1
^ SECRET
C C1 UNCLASSIFIED
416.
CONTRACT NO. TASK NO.
CONTRACT INSPECTION REPORT 25X
TO:
DATE
23 5eptembe ' 1965
P
INS
ECTION REPORT NO. (If final, so state)
4
ESTIMATED COMPLETION DATE
30 member 1965
TYP
A u t ttc ImAge P.A
ticrn by Caherent Ojpti*al Ted s
THE CONTRACTOR IS ON SCHEDULE
THE CON CTOR WILL PROBABLY REMAIN WITHIN ALLOCATED
^ YES IN
FUNDS ?YES O NO IF ANSWER IS "NO" ADVISE REC.
N
o
OMMENDATION AND/OR ACTION OF SPONSORING OFFICE
ON
.
REVERSE HEREOF. IF KNOWN, INDICATE MAGNITUDE OF AD-
PER CENT OF WORK COMPLETED -
DITIONAL FUNDS INVOLVED.
PER CENT OF FUNDS EXPENDED -
HAS AN INTERIM REPORT. FINAL REPORT. PROTOTYPE, OR OTHER END ITEM BEEN RECEIVED FROM THE CONTRACTOR
DURING THE PERIOD?, ^ YES ?
NO (If yes, give details on reverse side.)
HAS GOVERNMENT-OWNED PROPERTY BEEN DELIVERED TO CONTRACTOR DURING THIS PERIOD? ^ YES NO
(If yes, indicate items, quantity, and cost on reverse side.)
INCENTIVES
IS THIS AN INCENTIVE CONTRACT 0 YES NO
NOTE:
IF YES. CHECK TYPE
USE REVERSE SIDE FOR COMMENTS.
^ COST 0 PERFORMANCE
0 DELIVERY
FINAL REPORT MUST CONTAIN INCENTIVE EVALUATION.
OVERALL PERFORMANCE OF CONTRACTOR
1. OUTSTANDING
3. ^ ABOVE AVERAGE S. El BELOW AVERAGE 7. ^ UNSATISFACTORY
2. EXCELLENT
4. ? AVERAGE B. ^ BARELY ADEQUATE
IF OVERALL PERFORMANCE OF C
ONTRACTOR IS UNSATISFACTORY OR BARELY ADEQUATE, INDICATE REASONS ON
REVERSE SIDE.
RECOMMENDED ACTION
CONTINUE AS PROGRAMMED
^ WITHHOLD PAYMENT PENDING
SATISFACTORY PERFORMANCE
^ TERMINATE
OTHER (Specify)
IF TERMINATION IS RECOMMEND
ED OR IF THIS IS A FINAL REPORT PUT COMMENTS ON REVERSE IN NARRATIVE
FORM ON CONTRACTOR'S PERFORMANCE
AND CERTIFY THAT ALL DELIVERABLE ITEMS UNDER THE CONTRACT HAVE BEEN
RECEIVED. THESE INCLUDE. WHERE
APPLICABLE. THE FOLLOWING:
ITEM
RECD
DOES NOT
ITEM
REC 'D
DOES NOT
APPLY
APPLY
PROTOTYPES
MANUALS
DRAWINGS AND SPECIFICATIONS
FINAL REPORT
PRODUCTION AND/OR OTHER
SPECIAL TOOLING
END ITEMS
OTHER GOVERNMENT PROPERTY
DATE OF LAST CONTACT WITH CONTRACTOR
30 August 1965
SIGNATURE OF INSPECTOR
DIVISION
P&D8
INSPE
SIGNATURE OF APPROVER
Approved For Release 200
5/05/02 : CIA-RD
78BO477OA002300020007-9
FORM
10.64 1897 EDITIONS
EDITION
UNCLASSIFIED ? CONFIDENTIAL ^ SECRET (12-36)
 Approved For Release 2005/05/02 : CIA-RDP78BO477OA002300020007~9 7
CONRDEMflL
Automatic Image Recognition by
Coherent Image Techniques
Progress Report No. 7
Financial Status
Amount Authorized
Estimated expenditures
through 8 August 1965
Funds Committed
Funds Remaining
164-3-18
Copy of 6
Technical Status
Attached is a technical progress report which consists
of a summary report presented to the sponsor's technical
representative during his visit on August 19.
Distribution:
Copies 1 thru 4 - Technical Representative
Copy 5 - Contracting Officer
Copy 6 - File
GROUP 1
Approved For Release 2005/
ME
EXCLUDED FRMy;,+ I?, kT C
B04770A002300020 ?1 d.' T~
qt~:~ir,CAs ~:QII
25X1
 Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
Automatic Image Recognition by Coherent Optical Techniques
Summary Technical Status
Section A - Theory
The main problem considered here was to find if there were any theoretical
limitations to the recognition process in handling large amounts of highly detailed
information or recognizing objects of any degree of complexity. This was broken
into 4 problems:
A - 1 Maximum Information Content of a Photograph or Filter
Calculation shows that the upper limit of detail in the negative which the
process can handle is 2000 lines per millimeter. At the present time there are
few lenses or emulsions-which will realize as much as 200 lirges per millimeter,
and therefore the process is not a limiting factor in any photograph we can make.
There is no problem in attaining the necessary information content in the
filter. The two technical difficulties are resolution and dynamic range. The
resolution of the filter material has to be adequate to record the interference
lines produced by the two-beam process. This is usually in the range of 120 lines
per millimeter and is easily recorded on high resolution materials. Unfortunately,
these materials have a short dynamic range which makes it necessary to adjust the
exposure fairly exactly to obtain the desired selectivity in the filter, but this
is not a limitation on information content.
A - 2 Minimum Size of Detectable Objects
As the size of an object in a negative decreases the amount of light which
it passes decreases in proportion to its area. At some size the object will pass
so little light that even after it is concentrated into a recognition spot this
energy will not be appreciably higher than the surrounding noise level and the object
cannot be detected. Theory shows that for aerial negatives having an average "noise"
this minimum size is .3 mm or .012 inch. For fine-grained films the minimum size
would be even less, and for noise-free subjects such as printed copy the minimum
size is much less.
Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
 Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
A - 3 Modulation Transfer Function of the Optical System
The problem was to find if the recognition process imposed impossible
requirements on the optics. The requirements placed on each of the optical
elements in the recognition system and the filter making subsystem were analyzed
separately. Theory and experiment both show that there is no practical difficulty
in realizing a transfer function that is high enough for any practical application.
The only serious problem is keeping the lenses and other optical parts clean during
the life of the system.
A - 4 Obtainable Signal-to-Noise Ratios
Aerial photographs contain objects containing all sorts of detail as well
TAW
as grain which together create th6 optical equivalent of noise. The SNR can be
calculated only in closed form for specified signal shapes and for certain spectral
densities or noise. Making logical assumptions for the area of the object, film
graininess and other variables, it is possible to calculate a SNR of about 15 db.
Translated into optical terms this is adequate for detection and appears to agree
with experimental findings.
A - 5 Relative Merits of Phase (PPR) and Amplitude (Photographic) Filters
So far as theory is concerned, there is little difference between the two
kinds of filters. Ideally, the matched filter should consist of two parts, the
signal-passing portion and the noise-rejection portion. It is possible to calculate
that the signal-passing portion is identical in operation for a filter made by either
the photoplastic or photographic process. It has not been possible to calculate the
effect of noise-rejection portion, but this is rarely used anyway.
The conclusion to be drawn from the five theoretical studies is that the processes
is capable of much more than we need, the real problems are materials and technique.
Section B - Equipment
Several practical operating problems were considered:
B - 1 Effect of Glass Plates
It is sometimes convenient to use object transparencies or spatial filters on
Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
glass photographic plates, will irregularities in the glass be a problem? For small
 Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
areas, such as one-inch square spatial filters, the errors found in most glass
photographic plates have little effect on the recognition system. Over large
areas commercial photographic plates contain surface irregularities that result in
wavefront deformations which reduce the S/N ratio and cause a loss of detection.
This difficulty can be eliminated by placing the film or plates in a liquid gate
or by making the transparencies on microflat glass.
B - 2 Effect of Vibration
The 20-foot optical bench was vibrated in a rather severe manner and the
recognition system operated throughout the tests with no change in adjustments
necessary afterwards. Vibration does not appear to be a problem if the system
is properly designed.
B - 3 Reducing the Amplitude of the Reference Beam
Several means of beam attenuation appear equally satisfactory if the
attenuators are placed at the focal point of the second lens. Satisfactory
attenuation was obtained with a series of gelatin filters, a circular wedge and
two crossed polarizers mounted on flat glass plates.
B - 4 Liquid Gates
For film we have tested, equally good results were obtained with or without
liquid gates.
In review, none of the equipment problems appear serious and simple solutions
were found for the four studied.
Section C - Automation
If the system is ever to be used economically it must be automated to perform
the search and recognition mechanically. There are many problems to automation
and only the following were studied:
C - 1 Accuracy of Filter Alignment
If spatial filters are to be changed mechanically we must know the tolerance
on their location. Small errors of alignment, less than .001 inch had negligible
Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
 Approved For Release 2005/05/02 : CIA-RDP78BO477OA002300020007-9
4
effect. For average size objects, an error of ? .002 inch reduced the recognition
signal to 80 percent of its normal value. These values are within the tolerance
of automatic equipment.
C - 2 Orientation Tolerance
A rotation error of one-half degree between the image and filter results in
negligible signal loss. A two-degree angle error reduces the signal to 80 percent
of its normal value.
C - 3 Scale Tolerance
The curve of magnification error between the object and spatial filter as a
~ft: function of signal intensity has a practically flat plateau at zero error and
essentially linear slopes. A size difference of 2 percent results in little loss
of signal. Another way of describing the magnification or scale tolerance is that the
signal drops to 80 percent of maximum when the magnification is varied ? 3 percent
and falls to 50 percent with a change in size of ? 6 percent. For variations of less than
1 percent the loss of signal strength is negligible.
C - 4 Time Constant for a Scale and Orientation Search
This is based on the time to perform the scale (magnification) and orientation
(rotation) search of the photograph and indicates the time to process one negative.
Using the rotational and scale tolerances derived earlier, it appears that the search
time for a typical aerial negative is about two minutes including the time to change
transparencies.
C - 5 Use of Photoplastic Film for Filters
The signal-to-noise ratio of the recognition signal is equal for equivalent
silver and PPR spatial filters. The red-sensitive PPR material has low transmission
for red light and the signal is attenuated by this absorption. The processing time
as well as alignment of the finished filter is much faster for the PPR. The PPR
(red-sensitive) requires less exposure time than a satisfactory silver material and
it is reusable.
Approved For Release 2005/05/02.: CIA-RDP78BO477OA002300020007-9
 Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
C - 6 Real-Time Filter Generation
We redefined "real-time" to mean a short time rather than instantaneous.
The total elapsed time to make a silver-bearing filter on Kodak 649 material was
2 minutes exposure plus one-hour processing. The total elapsed time to make a PPR
filter of the same subject was 10 seconds exposure plus 20 seconds charging time and
a fraction of a second development time. The film could be precharged reducing the
time to practically the exposure time.
C - 7 Television Readout of Recognition Signals
A mechanical readout is essential and a television type system appears most
suitable. This work not yet completed.
In review, answers or at least partial answers have been found to several
of the problems of automating the system and adapting it to operate in real time.
Considerably more work on automation remains to be done, but no serious problems
have been uncovered.
Section D - Problems of the Photograph
This section was concerned with problems such as how well low-contrast and
camouflaged targets could be detected, also how well a filter made from a vertical
photograph would detect an oblique image of the same object.
D - 1 Effect of Target Size
The optimum signal-to-noise ratio for aerial photographs was obtained with
objects close to 5 mm size using our 20-foot bench, as predicted by the "f-400 rule".
Objects much smaller than 2 mm gave poor recognition on this bench, but when processed
on another instrument with shorter focal length lenses gave good recognition. The
entire transparency could not be covered at one time on the smaller instrument.
D - 2 Effect of Multiple Identical Targets
Multiple, identical targets in the same photograph can result in loss of
detection signal for some of these targets. If vehicles are oriented at certain
Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
 Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
small angles to each other resulting in phase interference in the most important
parts of the diffraction pattern, then one or more of these vehicles will produce
a weak signal. Close-packed vehicles lose part of their outline and the inner
vehicles show some loss of detection. In none of the cases tested did the relative
signal strength fall below 47 percent of that of a single vehicle.
D - 3 Multiple-Object Filter
For simple objects such as characters on a noise-free background, 20 or 30
can be used in one filter, certainly as many as 10. For more complex objects
such as vehicles on a noisy background, the present limit appears to be three or
four. The signal-to-noise ratio decreases as the number of objects in the filter
increases. The number of false-alarms for similar objects increases as the number
of objects in the filter increases. The magnification and orientation tolerances
are tightened slightly as the.number of objects increases. The use of a filter
containing four objects should reduce the search time to nearly one-quarter that
of using four separate filters.
D - 4 Effect of Shadows
Small, sharp shadows around a vehicle outline it, increasing the contrast
with its surroundings and improving recognition. In general, recognition is better
in photographs taken on sunny days than cloudy days when there are no shadows. On
the other hand, a low sun will produce long distorting shadows which will reduce
the recognition. We have no good numerical results on this effect at present.
D - 5 Effect of Target Aspect
This has not been completed, but the results indicate that variation in aspect
angle up to 20 degrees is not serious for the recognition of vehicles.
D - 6 Target Obscuration - Effect of Low-Contrast
With negatives showing no scratches or flaws, images of objects at extremely
low-contrast are located at the same signal-to-noise ratio as obtained in normal
Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
 Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
contrast negatives. Good recognition was obtained for images that had a density
difference from the background of 0.15 or less. The intensity of the recognition
signal is less in low-contrast pictures, but the noise also is smaller. Scratches
and defects become more serious as the contrast of the negative is reduced. If
the contrast is low due to camouflage, then the signal will be weak and the noise
normal. Recognition under these conditions could be difficult.
D - 7 Dazzle Camouflage
Dazzle camouflage produces strong diffraction noise. If the average
frequency of this noise is close to that of the chief spatial frequency of the
target, then the target may be obscured by noise and be undetectable by this process.
The presence of confusing background actually reduces the intensity of the recognition
signals. This is not understood at present.
D - 8 Overlap or Partial Concealment
Good recognition was obtained when up to 60 percent of the target was obscured.
Some recognition was obtained for targets 80 percent obscured. These recognitions
are possible only when the target has average or high contrast and the background
is the same. Low-contrast targets with noisy camouflage becomes more difficult.
D - 9 System Evaluation
This term is intended to include all of the factors in this section and
the problems of design and automation. As seen in this section, it is possible
to have objects too small to detect in a particular equipment and some kinds of
camouflage are effective in confusing the instrument. On the other hand, the
process has shown good performance on low-contrast and partially hidden objects.
There is every indication at present that the process should find any object a
human operator would normally find. An exception to this is deliberate camouflage
which may alert a human operator to look for additional clues. Human beings have
intelligence to cope with all sorts of difficult situations when necessary and the
recognition process has no intelligence.
Approved For Release 2005/05/02 : CIA-RDP78B04770A002300020007-9
 Approved For Release 2005/05/02 : CIA-RDP78BO477OA002300020007-9
The present evaluation is that the recognition process appears adequate
to handle many search and detection problems.
Approved For Release 2005/05/02 : CIA-RDP78BO477OA002300020007-9
 q.773$ 1)
Approved For Release 2005/05/02 : CIA-R P78B6770A0023000 0 07-9
MEDIA A 100
144-92.2
Copy of 6
25X1
v
25X1
Automatic Image Recognition by
Coherent Image Techniques
Progress Report No. 6
Financial Status
Amount Authorized
Estimated expenditures
thru 4 July 1965
Funds Committed
Funds Remaining
Technical Status
The engineering progress report for the period
13 June 1965 to 9 July 1965 is attached.
Distribution:
Copies 1 thru 4 - Technical Representative
Copy 5 - Contracting Officer
Copy 6 - File
GROUP Z
IXCLUDED FROM AUTOMATIC
DOWN 5 R A
Approved For Release 2INNIM P 78B04770A0023000 000
F'`il R
D CLAS ,C:i- 10,E
25X1
 Approved For Release 2005/05/02 : CIA-RDP78BO477OA002300020007-9
Automatic Image Recognition by Coherent
Optical Techniques
Progress Report No.6
Experimental studies conducted during this reporting period include work with
liquid gates on the 20-foot optical bench, multiple object filter tests, test of
first aerial films supplied and additJonal studies on the effects of obscuration.
The liquid gate problem has been solved to our satisfaction and the large gate
is now usable at least over the central 2 1/2 inch diameter area. Some liquids tested
exhibited extreme temperature sensitivity which produced refractive index stratification
and resulted in an effect exactly like astigmatism. A suitable liquid was located.
The alignment tolerances for a multiple object filter were described previously
as being quite critical. We have discovered some errors in the manner in which these
experiments were conducted and a new test was made on the alignment tolerance of the
entire pattern of ten objects with the filter and not the alignment of one object with
vwo the filter. The conclusions now reached are that the x and y displacement tolerance
of multiple object filters is the same as single order filters. The rotation and
magnification alignment tolerances are the same if the filter is made in the same
diffraction order. In general, multiple object subjects will require a higher order
filter. Since the higher order filter has "spots" at greater radial distances from the
axis, the rotation and magnification toerances are reduced proportionately.
The first aerial photographs supplied were duplicate negatives and were apparently
produced by the reversal process. These films lacked the relief characteristic found
in most photographic images. The assumption that the reversal process was probably
employed would account for the loss of this feature. Recognition obtained on the
submitted film was poor. Recognition on contract prints from the aerial negatives was
considerably improved, but still less than anticipated. Spatial filters made from
sharp originals such as models and cutouts gave slightly better recognition than filters
made from images taken from the aerial film. Some original negatives were supplied
and upon inspection they appear to not only have the desired relief effect, but are
also somewhat sharper in detail.. Filters are now being made and recognition studies
will be well underway during the early part of the next reporting period.
If negatives are in perfect condition, with no scratches or flaws, this recognition
process will locate images of objects at extremely low-contrast and at the same signal-
to-noise ratio as obtained in normal contrast negatives. Good recognition was obtained
for images that had a density difference from the background of .15 or less.
During the next reporting period, it is planned to work with the supplied aerial
negatives and begin planning the experiments with the image orthicon read-out.
Approved For Release 2005/05/02 : CIA-RDP78BO477OA002300020007-9