6
RETINA IDENTIFICATION
Robert âBuzzâ Hill
1
Portland, OR
buzzhill@rain.com
1. Introduction
Identification of a given person is often an essential part of transactions on a network.
While this is the goal, the fact is we often are left with substitutes for true personal
identification in such transactions such as something the person knows (password) or
has (a card, key, etc.). Retinal identification (RI) is an automatic method that
provides true identification of the person by acquiring an internal body image, the
retina/choroid of a willing person who must cooperate in a way that would be difficult
to counterfeit.
RI has found application in very high security environments (nuclear research and
weapons sites, communications control facilities and a very large transaction-
1
The author of this chapter is the original RI inventor and the founder of EyeDentify, Inc.
(1976). Although, he no longer owns stock or otherwise has an interest in EyeDentify, ha has,
at various times since 1987, served as its consultant.
Abstract
Retina based identification is perceived as the most
secure method of authenticating an identity. This chapter traces
the basis of retina based identification and overviews evolution of
retina based identification technology. It presents details of the
innovations involved in overcoming the challenges related to
imaging retina and user interface. The retinal information used
for distinguishing individuals and a processing method for
extracting an invariant representation of such information from
an image of retina are also discussed. The issues involved in
verifying and identifying an individual identity are presented. The
chapter describes performance of retina based identification and
the source of inaccuracies thereof. The limitations of the retina
based technology are enumerated. Finally, the chapter attempts to
speculate on the future of the technology and potential
applications.
Keywords:
Fundus, choroid, fundus camera, astigmatism,
ergonomics, infrared imaging, fixation.
2
Hill
processing center). The installed base is a testament to the confidence in its accuracy
and invulnerability. Its small user base and lack of penetration into high-volume price-
sensitive applications is indicative of its historically high price and its unfriendly
perception.
2. Retina/Choroid as Human Descriptor
Awareness of the uniqueness of the retinal vascular pattern dates back to 1935 when
two ophthalmologists, Drs. Carleton Simon and Isodore Goldstein, while studying eye
disease, made a startling discovery: every eye has its own totally unique pattern of
blood vessels. They subsequently published a paper on the use of retinal photographs
for identifying people based on blood vessel patterns [7].
Later in the 1950's, their conclusions were supported by Dr. Paul Tower in the
course of his study of identical twins [8]. He noted that, of any two persons, identical
twins would be the most likely to have similar retinal vascular patterns. However,
Tower's study showed that of all the factors compared between twins, retinal vascular
patterns showed the least similarity.
The eye shares the same stable environment as the brain and among physical
features unique to individuals, none is more stable than the retinal vascular pattern.
Because of its internal location, the retina/choroid is protected from variations
caused by exposure to the external environment (as in the case of fingerprints,
palmprints etc.).
Referring to Figure 6.1, the retina is to the eye as film is to camera. Both detect
incident light in the form of an image that is focused by a lens. The amount of light
reaching the retina (or film) is a function of the iris (f-stop). The retina is located on
the back inside of the eyeball. Blood reaches the retina through vessels that come
from the optic nerve. Just behind the retina is a matting of vessels called the choroidal
vasculature.
Visual Axis
Pupil
Blind Spot
Vessels
Iris
Fovea
Scan Area
(Area Band)
Figure 6.1
Eye and scan circle (area band).
The products of EyeDentify, Inc. have always used infrared light to illuminate the
retina as will be discussed later. The retina is essentially transparent to this
wavelength of light. The mat of vessels of the choroid just behind the retina reflect
Retina Identification
3
most of the useful information used to identify individuals, so the term âretinal
identificationâ is a bit of a misnomer but nevertheless useful because the term is
familiar. RI in this chapter will be used interchangeably to mean retina/choroid
identification. This area of the eye is also referred to by medical doctors as the eye
fundus.
It might seem that corrective error changes (such as becoming more near-sighted
over time) might change the image of this very stable structure. In fact, the low
resolution required to acquire adequate identification information masks any effect the
focus errors might have. The RI products of EyeDentify, Inc. take advantage of this
fact. No focusing of the RI system optics is necessary reducing cost and making the
unit easier to use.
The operational rule-of-thumb for the circular scan RI systems described here is as
follows: If the person to be identified can see well enough to drive with at least one
eye, it is highly likely that he/she can use RI successfully.
Children as young as four years of age have been taught how to use RI. Once
learned, RI is simple to use for the vast majority of the human population.
3. Background
The concept of a simple device for identifying individuals with RI was conceived in
1975. A practical implementation of this concept did not emerge for several years.
The author formed a corporation, EyeDentify, Inc. in 1976 and a full time effort began
to research and develop RI. In the late 1970s several different brands of ophthalmic
instruments called fundus cameras were modified in an attempt to obtain live images
of the retina suitable for personal identification [9,10]. Using then available fundus
cameras for the optics portion of RI had at least three major disadvantages:
1. Critical alignment was necessary requiring either extraordinary expertise or the
assistance of an operator.
2. A bright illumination light was necessary.
3. They were too complex and therefore too expensive.
The early RI experiments used visible light to illuminate the retina. This proved
undesirable since the amount of light required for a sufficient signal-to-noise ratio was
often uncomfortable to the user. An experiment was tried using near infrared as the
illuminating source. This wavelength is invisible to the human eye and eliminates the
bright illuminating light that can be annoying to the subject and cause his/her pupils to
constrict (lowering the detected light). Inexpensive light sources and detectors existed
for the near IR providing a cost saving advantage as well.
The first practical working prototype of RI was built in 1981. An RI camera using
an infrared light was connected to a general-purpose desktop computer for analyzing
the reflected light waveforms. Several forms of feature- extraction algorithms were
evaluated. Simple correlation proved to be a superior matching technique however.
Four years of refinement led to the first production RI system built by EyeDentify,
Inc. (then of Portland, OR). It was called the EyeDentification System 7.5 and
performed three basic functions:
4
Hill
1. Enrollment - where a person's reference eye signature is built and a PIN number
and text (such as the person's name) is associated with it.
2. Verification - a person previously enrolled claims an identity by entering a PIN
number. The RI scans the ID subjectâs eye and compares it with the reference eye
signature associated with the entered PIN. If a match occurs, access is allowed.
3. Recognition - RI scans the ID subjectâs eye and âlooks-upâ the correct, if any,
reference eye signature. If a match occurs, access is allowed.
System 7.5 performed a circular scan of the retina, reducing the circular fundus
image composed of 256 twelve bit logarithmic samples into a reference signature for
each eye of 40 bytes. The contrast pattern was coded in the frequency domain. An
additional 32 bytes per eye of time-domain information was added to speed up the
Recognition mode.
Patents
State-of-the-art RI is covered by at least nine active U.S. patents. The RI
implementation described here is covered by at least four major U.S. patents dating
back to 1978 [2,3,4,5]. The patent with broad first claim to the method of RI [2]
expired in 1995 and is thought to have discouraged others from developing RI
technology. Now that the method of identifying individuals by their retinal patterns is
no longer protected (as opposed to the apparatus to identify), we may see more
interest by others in developing RI technology that is not protected by the active
patents whose claims are less general than the expired patent. EyeDentify, Inc. either
owns or has exclusive license to the three aforementioned patents that have not, as
yet, expired. These patents deal with the alignment/fixation and user interface
subsystems of the RI technology.
4. Technology
The three major subsystems of the RI technology are:
Imaging, signal acquisition, and signal processing
: An RI Camera that
translates a circular scan of the retina/choroid into a digital waveform.
Matching
: A computer that verifies or recognizes the acquired eye pattern with a
stored template.
Representation
: The eye (retina) signature reference templates with the
corresponding identification information; storage issues.
Sections 6, 7, 8, 9, and 10 describe in more detail the entire RI system. Section 5
discusses representation issues.
Retina Identification
5
5. Eye Signature (Reference Template)
The representation of retina is derived from a retina scan composed of an annular
region of retina, scan circle (Figure 6.1). The spot size (width of annular band) and
scan circle size are chosen to return sufficient light and contrast detail in the worst
case (very small eye pupil) to support the performance specification of the RI.
Two major representations for the RI eye signature have emerged. The original
representation consisted of 40 bytes of contrast information encoded as real and
imaginary coordinates in the frequency domain and was generated with the fast
fourier transform.
The second representation, while slightly larger at 48 bytes, leaves the contrast
data in the time domain. The primary advantage of the time domain representation of
the eye signature is computing efficiency resulting in lower computer cost and/or
higher processing speed.
Taking the ratio of the brightness at any point to the average regional brightness
removes artifacts that are due to non-uniformity of the beam at the point where it
enters the eye. This also normalizes the identifying signal for varying pupil sizes that
greatly influence the total light returned to the detector.
The fully processed digital eye signature can be described as a normalized contrast
waveform of the entire scan circle. Average RMS contrast averages approximately 1.5
to 4% of the total light detected. The contrast maximum is the brightest reflection
from the scan circle and the contrast minimum is the darkest reflection from the scan
circle. The waveform is normalized so that the maximum or the minimum is at the 4
bit limit (either +7 or â8, respectively) to fully utilize digital dynamic range.
The simplest form of the RI reference eye signature is an array of 96 four-bit
contrast numbers for each of 96 equally spaced scan circle positions for a time-
domain pattern of 48 bytes per eye. An optional 49th byte carries the AC RMS value
of the waveform to be used for equalizing the RMS values of the acquired and
reference waveforms in the correlation (match) routine.
6. RI Camera
Most of us, at one time or another, have gone to an optometrist or ophthalmologist to
have our eyes examined. As part of the exam, the doctor uses an instrument called a
retinascope. The RI camera accomplishes the same thing as a retinascope. Its light
source is projected onto the subject's retina and (the doctor in case of the retinascope)
detects the return light. The light coming from the retinascope is in a collimated beam
so that the eye lens focuses it to a spot on the retina.
The retina reflects some of the light back towards the eye lens, which once again
collimates the light. This light leaves the eye at the same angle that it enters the eye, a
process called retro-reflection. The light reflected from the retina is observed by the
examining doctor who holds the instrument to his own eye. In the case of RI, a light
detector replaces the examining doctor's eye.
If the doctor were to examine the eye from a number of points 10 degrees off the
visual axis of the patient's eye, it would simulate the action of the fovea centered RI
scan we will discuss here.
6
Hill
Old Camera
}
}
34
26
26
62
57
60
64
66
10
28
30
20
14
14
10
12
32
17
32
18
5
68
59
56
34
50
48
52
54
82
74
72
76 78 80
38 36
37
35
33
70
46
44
40
39
42
Scanner
Light Transmitter
Reflected light receiver
Figure 6.2
Old camera.
The first products of EyeDentify used a camera disclosed in detail in US Patent
#4620318 [4] (Figure 6.2). This design used a rotating mirror assembly to generate
the scan circle on the retina. Hot mirrors (reflecting infra red while transmitting
visible light) are used to combine the Scanner optical path and the align/fixate target
optical path. What follows is a description of the operation of the relevant portions of
the camera described in the patent as they relate to the EyeDentification System 7.5.
A fixation target (33) allows the RI subject to properly focus his/her eye (5) and
align its visual axis (10) with an optical axis (34) of the scanner. Fixation target (33)
includes a visible light emitting diode (35) positioned in a mounting structure (36)
having a pinhole (37). LED (35) illuminates the fixation reticle (38).
US Patent#4923297 [1] describes an improved fixation targeting system that
replaced the system described above in production 7.5's. This patent describes the 7.5
fixation target system as a quasi-reticle that generates enhanced multiple ghost
reflections of a single pinhole. It is a simple plate of glass with a partially silvered
mirror on one surface and an opaque mirror surface with a pinhole on the other
illuminated by a light emitting diode.
Alignment is a critical requirement of RI and this so-called âghosticleâ
alignment/fixation system accomplishes its function elegantly. It is simple and
intuitive - just line up the dots - and both alignment and fixation are assured. Yet it is
inexpensive and easier to align in production than previous RI alignment/fixation
systems.
Once alignment and fixation are accomplished the scan can be initiated either
manually by pressing a button or automatically when the RI is placed in the Auto-
Acquire mode (a feature introduced in the model 8.5 product).
An IR source (39) provides a beam of IR radiation for scanning fundus (12) of eye
(5). IR source (39) includes an infrared light emitting diode (the drawing shows a
tungsten bulb (40) as the light source) that produces light that passes through a spatial
filter (42) and is refracted by a lens (44). An IR filter (46) (not used in the IR LED
version) passes only the IR wavelength portion of the beam, which then passes
through a pinhole (48). The beam is then reflected by a mirror (50) onto a beam
splitter (52) that is mounted to coincide with the fixation target optical axis (34).
The scanner directs the beam into the fixated eye from an angle of 10 degrees
offset from the optical axis. The scanner includes a rotatable housing (57) and scanner
Retina Identification
7
optics that rotate with the housing as indicated by a circular arrow (58). As the
scanner rotates, the 10 degree offset beam rotates about the optical axis.
The scanner optics include a hot mirror (59) (one that reflects IR radiation while
passing visible light), located in the path of the source beam and the fixation beam.
The visible wavelength fixation beam is passed through hot mirror (59), while the IR
source beam is reflected away from the housing (57) at a point spaced apart from
optical axis (34) and is oriented to direct the IR beam through an IR filter (62) and
into the eye (5) as housing (57) rotates. Hot mirror (59) causes a displacement of the
fixation beam so an offset plate (64) is positioned to compensate for the displacement.
When housing (57) rotates, the IR beam is directed into the eye (5) in an annular
scanning pattern centered on the fovea as represented by circular locus of points (32).
Light reflected from fundus (12) of the eye (5) varies in intensity depending on the
structures encountered by the scan. The reflected light is re-collimated by the lens
(30) of the eye (5) directed out pupil (28), back through objective lens (66) and IR
filter (62), and reflected off scanner mirror (60) and hot mirror (59). The reflected
beam is then focused by objective lens (56) on to beam splitter (52) which passes a
portion of the reflected scanning beam to a hot mirror (70) that reflects the beam
through a spatial filter (72). The beam is next reflected by a mirror (74), refracted by a
lens (76) and passed through another spatial filter (78) to a detector (80).
New Camera
Current RI camera technology is based on an active US Patent [5]. It is a much
simpler design that also takes advantage of the concentric nature of the RI's fixation
and scanning to reduce labor intensive alignment of camera parts and the part count
The current RI camera is shown schematically in Figure 6.3. It includes a rotating
scanner disk (116) that integrates a multi-focal Fresnel fixation lens (114), a Fresnel
optical scanner (122,124) and an angular position encoder (140) into a unitary,
inherently aligned, compression-molded scanner disk. An RI subject views through
the multi-focal Fresnel lens, an image of a pinhole (108) illuminated by a krypton
bulb (104). The multi-focal lens is centered on the disk and creates a multiple in- and
out-of-focus images (180 182, 184, 186) of the pinhole image. By setting the focal
distances of these images along a range that includes corrective errors of from -7
diopters (very near sighted) to +3 diopters (very far sighted), at least one of the
pinhole images will be in relatively sharp focus for virtually everyone in the RI
subject population. The images will appear concentric when the RI subject is properly
aligned with the scanner disk and associated optics.
Once aligned, the subject initiates scanning which causes the scanner disk to rotate.
The Fresnel optical scanner receives IR light from the krypton bulb light source and
creates an IR scanning beam (126). IR light reflected by the eye fundus (12) of the RI
subject returns via a reciprocal path, by way of a beam splitter (112) and into a
detector (134) to generate eye waveform data. Rotational position information from
the encoder instructs the signal acquisition system when to sample the detector's
output.
The key feature of this new RI camera design is that it integrates and inherently
aligns multiple optical elements greatly reducing both the material and labor costs of
8
Hill
the original RI camera. Overall costs of the camera yields to manufacturing
economies of scale much more so than with the original RI camera.
Both camera types share the same subsystem functions:
Figure 6.3
New camera
Target - Align & Fixate
To insure that the circular scan of the RI is centered on the fovea and that the subject
is in the scanner beam throughout the scan, an alignment/ fixation target is presented
to the ID subject. One form of this target is an optical system that presents say four
simple reticles at focal distances of -7, -3, 0 and +3 diopters. For virtually all of the ID
population, at least one of the reticles will be in focus regardless of corrective error
(near-sightedness through far-sightedness). When the ID subject âfocusesâ (fixates)
on the target, the RI is angularly aligned to subject's eye, centering the RI's scan circle
on the eye fundus. When he/she aligns two or more of the reticle patterns nulling their
parallax, the RI illumination beam is centered on the eye pupil. Translation along the
optical (Z) axis is not critical and is achieved by providing a rest for some part of the
face (forehead, eye socket, etc.). Rotation about the Z-axis caused by head tilt is
addressed by the Rotator algorithm, discussed later.
It is important to note that Fixation/Alignment is an absolute requirement for this
method of RI to work. It would be prohibitively difficult to identify someone using
RI without his or her cooperation in performing this function. Depending on one's
perspective, this requirement can be seen as a benefit (usually to the ID subject) or a
negative (covert ID). This does, however, prevent RI's use in identifying an
individual against his or her will which may make RI appear more acceptable to the
ID subject population.
Retina Identification
9
Transmitter (Light Source)
The light source is ideally near infrared and is not visible to the identification subject.
The illumination spot projected on the retina must be uniform. A suitable diffuser is
required to achieve spot uniformity when using a light source that, when projected on
the retina/choroid, is not homogenous. This is usually the case with an IR Light
Emitting Diode and can also be true of other light sources.
A tungsten lamp is considerably brighter than an IRED and can produce better S/N
figures. The disadvantages of such a lamp compared to an IRED is the need for filter,
turn-on time and lamp life. Retinal identification systems have been proposed that
would use a laser (preferably solid state). The author is not familiar with any
commercially available system that uses a laser, however. Further, lasers can be
considered dangerous by the RI subject population.
Receiver
The light receiver is composed of a silicon photodetector, a high gain pre-amplifier
and a sharp cut-off low-pass filter. The filter is necessary to sharply reduce high
frequency noise generated by the detector/preamp that is outside of the useable
passband which is determined by the spot and scan circle sizes and the scanner speed.
With the selected parameters a good choice is an 8th order switched capacitor elliptic
filter with a corner frequency of approximately 220 Hz.
Scanner
The scanner has to deal with the light noise arising from (i) corneal reflections, (ii)
other scattered light sources, and (iii) ambient light. Reflected noise in the RI comes
from essentially four sources, the front and back surfaces of the cornea, and the front
and back surfaces of crystalline lens. Extensive spatial filtering that is conjugate to
the retina and the scan angle reduces the light noise to insignificance.
Corneal reflections of the scanner light source is one of the primary reasons for
using a circular rather than a raster scan of the eye fundus. The reflections would
render the center pixels of a raster scan of the retina useless unless an annular
illumination requiring very critical alignment is used. The scanner consists of the
following components comprising the signal acquisition and processing subsystem:
7. Signal Acquisition Subsystem
The signal acquisition subsystem consists of the following components:
Detector/Preamplifier
The silicon photodetector operating in the photo-ampuric mode receives the light
collected from the RI camera. It is converted to a voltage by a low noise op-amp
configured as a trans-impedance amplifier. With a carefully selected op-amp, the
primary sources of electrical noise are the thermal noise of the feedback resistor and
10
Hill
quantum noise. A second op-amp brings the signal level up to a level sufficient to
drive the contrast processor.
A/D Conversion
The raw unprocessed analog signal derived from the camera photodetector can span at
least two orders of magnitude due to the range of pupil sizes encountered in normal
operation of the RI. Performing the conversion at this point requires close to 16 bits
of resolution to accommodate absolute signal variations, contrast figures and
sufficient resolution left to quantize the âcontrastâ portion of the signal. A more
economical scheme is to perform the contrast processing function ahead of the
conversion. An 8 bit analog-to-digital converter is all that is required in this case
Contrast Processor
8. The function of this stage in the signal chain is to reduce the raw camera signal
into salient contrast information that has both human descriptor qualities of
invariance and discrimination. It can be done in hardware or software and both
methods have been used successfully in EyeDentifyâs commercial products. The
far less expensive modality is hardware because it dramatically reduces the
resolution required of the analog /digital converter. The contrast processor
removes the redundant or variable content from the acquired scanner waveform
while retaining sufficient information to yield a unique eye signature.
8. Computing Subsystem
The computing subsystem could be explained in terms of its hardware and software
components.
Hardware
EyeDentify's System 7.5, the first widely available RI, used a Motorola 68000
microprocessor as both the controller and signal processor. By moving contrast
processing to hardware and coding correlation in the time domain in the late 1980s, it
was possible to move to a 68HC11 micro-controller to replace most of the
functionality of the 68000 based System 7.5. The cost of the computing elements of
RI have been and currently are insignificant compared to the opto-mechanical portion
of the system.
Software
The software performs the following two functions: phase correction and matching.
Phase Correction
Each time the RI subject looks into the RI camera, his or her head may be slightly
tilted (rotated) from the position it was before. The rotator algorithm (phase corrector)
Retina Identification
11
shifts the acquired waveform through the equivalent of several degrees of rotation or
head tilt. This is done while correlating it with the stationary reference eye signature
to find the best match (highest correlation).
Matching
Comparison of the acquired contrast waveform is done with a routine that performs
the following steps:
1. Sample rate converts the reference eye signature into an array with the same
number of elements as the acquired array.
2. Normalize both arrays to have a RMS value of 1.0.
3. Correlate arrays using the time domain equivalent of Fourier-based correlation.
The quality of match is indicated by the correlation value, where the time shift is
zero. It ranges from +1.0, a perfect match, to -1.0, a perfect mismatch. Experience has
shown that scores above 0.7 can be considered a match (see Performance discussion
below).
8. System Operation
Taking an Eye Reading
Central to every RI transaction is the process wherein the camera scans the RI
subject's eye. We present here the detailed user instructions below to give an idea of
the user involvement and training needed for retina based identification. The subject is
instructed as follows:
If you wear glasses, take them off (does not apply to contacts).
If the system requires PIN (Personal Identification Number), enter it (recognition
does not require a PIN).
Position camera at eye level (or eye to camera)
The target consists of a number of softly illuminated dots. Moving the head in
relation to the eye lens opening, without tilting or skewing the head centers the
target. Do this until all of the dots move one behind the other. The smaller dots
will then appear inside the larger dots.
Both eyes should be wide open. Squinting or closing one eye can cause eyelashes
to be included in the reading.
Be sure that your eye is about three-quarters of inch from the eye lens.
Press the scan button (or wait for scanner to stop if in the Auto-Acquire mode).
Hold your head steady during the reading.
12
Hill
Although it is important to fixate on the center of the target during the reading,
you should not fixate for more than a couple of seconds before pressing the
button. Otherwise, the eye may drift.
Various incarnations of RI cameras and systems have different user requirements,
but the steps above apply generally to all of them.
Alignment/Fixation
To use RI, it is important for the subject to be aligned with the RI camera and fixated
on its target. After peering into the camera, the subject achieves alignment by lining
up the dots of the target so they appear as one. At that point fixation is also
accomplished since the virtual dot image is then focused on the fovea of the subject's
eye (Figure 6.1). This process assures that the subject's eye pupil is within the
âacceptance diamondâ which is the cross sectional shape of the volume where the
entire scan's beam will fill the eye pupil. This volume is essentially like two cones
placed together at their bases with their centers along the eyeâs optical axis (the Z
axis). A larger volume means less critical alignment. The volume is a function of the
exit/entrance aperture, which is determined by the size of the RI camera's objective
lens.
Scanning
Eyeglasses must be removed for the RI camera to work reliably. There are two
reasons: 1) Reflections from the lens surface may interfere with the scanner signal,
and 2) Distortion of the retina/choroid image may occur if eye glasses are not in the
same position on the face from use to use such as when they slide down the ID
subject's nose. If an attempt is made to enroll an individual with eyeglasses, it is
possible that the eye glass reflection will be enrolled, not the retina/choroid, resulting
in a very simple eye signature that might be duplicated.
Contacts do not need to be removed. Certain types of contacts can prove
problematical. Lenses can cause improper signatures if any part of the edge of the lens
is inside the eye pupil while the eye is being scanned. Generally, the effect of contacts
on eye signatures is so slight that it is not necessary to enroll a given eye both with
and without them, except possibly in cases of severe or unusual correction (extreme
near- or far-sightedness and/or astigmatism).
RI at a Distance
Just as the eye doctor can use a retinascope at a distance from the patient, a suitably
designed RI can be used at a distance from the ID subject. However, the size of the RI
must increase proportionate to the scan distance in order to support the RI's scan circle
diameter. Working RI systems with an operating distance of 12 inches have been
demonstrated in the laboratory. Other considerations in such systems include ambient
light conditions and Fixation/Alignment issues. Light shields sizes have to grow in
proportion to the operating distance. A long distance universal focus target's
requirements change when the operating distance exceeds a certain threshold. Scanner
beam size will need to be larger as well.
Retina Identification
13
Enrollment
RI enrollment is the process of acquiring the reference eye signature. Each eye
signature is built from several eye readings. The person responsible for enrolling a
new RI subject, the enroller guides that person through the following steps:
Instruction on camera use. Enroller instructs the enrollee on correct RI camera
use. Enroller usually demonstrates this by scanning his/her own eye and then
describes what the Enrollee must do to align the camera. Fixation is automatic
when the enrollee achieves alignment.
Several scans until correct fixation/alignment is verified. Out of beam condition
(meaning the subject has not achieved alignment) is detected when the raw signal
drops off in some part of the waveform. Both manual and automatic modes have
proven effective for this purpose. Fixation can only be verified when scans are
compared. Correlation scores of a scan with the reference eye signature greater
than, say, 0.75 to 0.8 indicate that correct fixation has been achieved.
Several scans averaged. Scans that have a correlation within a certain range
(such as 0.75 to 0.8 are added to a waveform average. The impact on correlation
scores of variant features such as a choroidal vessel that is substantially tangent to
the scan circle is reduced with averaging.
Optional Recognize - verifies whether or not the new enrollment eye signature
already exists in the database either because the new enrollee has already enrolled
the eye or the database is large enough to include a sufficiently similar eye
signature to cause a false accept error in the Recognition mode. This step assures
the new enrollment eye signature is unique to the database.
Assignment of linked data (Name, Pin #, etc.).
Store enrollment data.
Automatic RI enrollment techniques have been studied wherein eye scans that do
not match any eye signature in the database (using the Recognition algorithm
described below) but appear repeatable are given a label indicating such and stored to
indicate intrusion attempts. The RI system can alert an administrator when
unrecognized but repeatable eye signatures occur by displaying/printing that label.
It is important to note that enrolling is somewhat of an art as well as a science.
The enroller, through experience, learns how best to train each new enrollee and to
interpret correlation scores during the enrollment process. An enroller should
remember several key points. Correlation scores should get progressively better as
enrolment progresses. It is important for the enrollee to look away between scans to
insure that the averaging process creates a true average of variations in head position.
A person's âdominant eyeâ can be easier to enroll. If one eye is difficult to enroll, try
the other.
14
Hill
Verify
Subsequent to enrollment an enrollee can authenticate his/her identity by entering a
code (such as a PIN number). An eye scan is taken and compared with the eye
signature associated with the PIN number. If the eye scan matched against the eye
signature produces a correlation score above the match threshold (typically a
correlation score of 0.7), the person scanned is said to be the person enrolled with the
given PIN and an appropriate action is taken.
Recognize
A scan is taken and using the recognize algorithm, a match to the entire database
above the correlation threshold identifies the person requesting access. Any eye
signature recognize algorithm is considerably more compute intensive than verify
algorithm. The simplest form of recognize would take the verify time and multiply it
by the number of people in the eye signature database. Several multi-level techniques
have been developed that reduce the time it takes for the recognition mode on given
computing resources. In some cases, execution time has been reduced as much as two
orders of magnitude. The down side of the methods tried is that they sometimes
eliminated good candidates, producing false reject errors.
Today's fast microprocessors and time domain correlators have nearly eliminated
the need for multi-step recognition routines for databases of medium size (hundreds to
tens of thousands) and, in the process, have virtually eliminated false reject errors
produced by multi-level recognition algorithms.
Large Database Recognition
Recognition is identification where the ID subject does not claim an identity (with a
PIN number, card etc.) as part of the process. The acquired scan waveform is
compared to an entire database of eye signatures to find the best match.
Currently available parallel processing computers can perform high accuracy RI
recognition of databases composed of millions of eye signatures at very low relative
cost. Indeed, some RI recognition mode feasibility has been studied on massively
parallel processing supercomputers with very promising results. Simply dividing up
the database and having each processor correlate the acquired scan with its portion of
the database is the simplest method. RI's small signature size, uniqueness and small
variance gives it a significant competitive advantage in terms of cost, speed and
reliability for large database recognition mode over other biometric ID methods.
Counterfeiting the Scan
A counterfeit eye must have the following characteristics.
The same optical system to simulate retina/choroid reflectivity.
A lens to substantially focus the incoming collimated beam and to re-collimate
the reflected beam.
Retina Identification
15
An alignment/fixation system that angulary orients the counterfeit eye about it's
X and Y axes, translates the counterfeit eye along its X and Y axes, positions the
eye at the correct distance from RI camera (translation along Z) and rotates the
eye about its Z axis within the tilt range of the Rotation algorithm.
The last item is the most difficult to counterfeit. A well-designed RI provides as
little information about the correct alignment/fixation as possible to the would-be
counterfeiter. Variable fixation displays could also require a counterfeiter to perform
an interpretation of the target in order to correctly interact with the acquisition
process. For example, the ID subject could be instructed to remember a random three
digit sequence that is displayed in the RI fixation target and to key it in later. This
would force the counterfeiting system to see, interpret and output to the RI keypad the
three digit sequence.
9. Performance
100
90
80
70
60
50
-10
40
-20
30
-30
20
-40
10
-50
0
51 1E-03
58 1E-04
64 1E-05
70 1E-06
Figure 6.4
Mismatch distribution in retina based identification.
Many tests of performance of the retina/choroid scanning technology described have
taken place, some with databases of several hundred individual eyes. Sandia National
Laboratory has tested the products of EyeDentify and reported no false accepts and a
three-attempt false reject error rate of less than 1.0% [11].
Mismatch Frequency Distribution
A frequency distribution of each eye signature compared against all others matches
very closely with an ideal guassian distribution with a mean of 0.144 and a standard
deviation of 0.117 as shown in Figure 6.4. The corresponding right tail probability of
guassian distribution with this mean and standard deviation at a threshold score of 0.7
is approximately one in one million.
16
Hill
Source of Errors
The retina/choroid contrast waveform has a low variability when acquired under
correct conditions. The conditions under which the variability could increase and
cause false reject (Type II) errors are:
Lack of Fixation or sustained fixation
Out of scanner beam condition
Incorrect eye distance to RI camera lens
Insufficient pupil size
Obstruction and distortion of the optical path from:
dirty camera window
contact lens edges
subject neglects to remove eyeglasses
Ambient light interference
Small pupils can cause false rejects. The purpose of the eye pupil is to regulate the
amount of light reaching the retina. Bright environments such as those encountered
outdoors in the daytime can cause pupils to constrict to a very small size compared to,
say, indoor lighting conditions. Because light must pass the eye pupil twice (once
entering and once exiting the eye), the return light to the RI camera varies inversely
with the fourth power of the pupil diameter. In the worst case (smallest pupil size),
resulting retina/choroid signals can be attenuated by as much as four orders of
magnitude. The signal can be so low that system noise swamps the acquired eye
signature data, lowering correlation scores.
Outdoor environments can also be less conducive to reliable RI performance
because of the potential for high ambient light noise entering the camera and
interfering with the scanned waveform. Because of the uniqueness of the
retina/choroid contrast circle characteristic, false accept (Type I) errors tend to be
limited to large database recognition.
10. RI Subject Motivation
An important and enduring observation of the use of RI to enroll and identify several
thousand individuals over a period of two decades is the importance of motivation to
have the enrollment and identification transactions succeed. Many of these
observations can be said, in a general sense, of other biometrics as well.
Enrollment is the subject's first hands-on use of RI. The subject should not fear
harm from the RI camera before using it the first time. Learning to align and fixate
the RI camera, while a simple process, can be impeded willfully or subconsciously by
a suspicious subject. Several scans are necessary and depending on the quality of the
Retina Identification
17
scans, this procedure can take several minutes. Because RI requires cooperation from
the subject, sabotage at this stage is very easy. If a subject is difficult to enroll, the
subject's motivation can deteriorate as time passes during the enrollment process.
The identification transaction (verify or recognize) is less susceptible to fear based
motivational problems simply because if a subject has been successfully enrolled
he/she must have overcome considerable fear or reluctance already. But subtle
sabotage can be a factor here as well. Deliberate false reject transactions with an
accompanying complaint such as âit gives me a headacheâ can diminish confidence in
the system. It is very difficult to ascertain whether subjective comments of this kind
are truthful yet the result is the same - RI is less attractive.
Many user's have naively assumed that the lack of negative consequences (I can't
work here because I can't/won't use RI) is sufficient to gain acceptance of a RI system
by the ID subject population. Experience has taught users of RI that a perceived
personal benefit (I am better off than people who can't/won't use RI) to the ID subject
has a dramatically positive effect on RI enrollment and identification speed and
acceptance.
11. Limitations
Perceived Health Threat
While the low light level is harmless to the eye, there is a widely held perception that
retinal identification can hurt the retina. This appears to be less true in information
access applications since ID subjects are generally less fearful of new technology.
Outdoors vs. Indoors
Small pupils can reduce the Type II (False Reject) performance. Because light must
pass the eye pupil twice (once entering and once exiting the eye), the return light can
be attenuated by as much as four orders of magnitude when the ID subjects pupils are
small. The signal can be so low that quantum and feedback resistor noise swamp the
eye signature data lowering correlation scores. Further, outdoor environments are less
conducive to reliable RI performance than indoor environments because of ambient
light conditions.
Ergonomics
The need to bring the RI device to an eye or the eye to the device makes the RI more
difficult to use in some applications than other biometric identification technologies.
For instance, it is quite easy for a subject, regardless of his height to reach a hand to a
fingerprint or hand geometry. The eye is much less easily manipulated. Bringing the
RI camera to the eye seems more practical in âworkstationâ applications while the
opposite is true in physical access control applications.
18
Hill
Severe Astigmatism
Because eyeglasses must be removed in order to use RI systems reliably, people with
severe astigmatism may have trouble aligning the dots in the camera's align/fixate
target. To these individuals, what appears to them can be quite different than dots.
This can result in ambiguous feedback during the alignment step of RI camera use,
causing the eye pupil to be outside the âacceptance diamondâ for part of the scan.
That part of the scan will therefore be invalid.
High Sensor Cost
The camera requirement of RI puts a lower limit on the cost of the system.
Manufacturing economies of scale can mitigate this problem, but RI is likely to
always be more expensive than some other biometrics such as fingerprint (using
chips) or speaker recognition (telephone hand-set as sensor).
12. Future
The inherent simplicity of the RI means that in mass production the cost of the entire
unit could come below, say, $100. This is still considerably more expensive than
some competing technologies which have a much cheaper scan component (such as
fingerprint chips). The trade off is accuracy. If accuracy is important to the ID
application, perhaps the additional cost of RI can be justified.
With the proliferation of e-commerce applications, RI might reach a critical mass.
Because of the RI's accuracy and small signature size it fits more naturally with the
encryption that is needed for e-commerce security than competing biometric ID
technologies. Public key encryption systems are only as secure as their private keys
and a high performance biometric identifier like RI is ideal for keeping private keys
secret.
13. Conclusions
RI is a highly accurate and secure biometric identification method. The example RI
system presented utilizes a small reference data size that makes it attractive in large
population networked systems in both verification and recognition modes. RI,
currently, is both image and performance based. The performance aspect restricts
successful use to those who are motivated to see the ID transaction successful.
RI's weakness are:
The cost of the signal acquisition hardware
ID subject's unfounded fear that it is harmful
Unfriendly access
The future will likely bring the cost of RI down dramatically if a sufficiently large
demand is created to achieve manufacturing economies of scale especially as it
Retina Identification
19
applies to the RI camera optics and mechanics. Fear becomes less of an issue as the
computer/internet age expands and raises the level of technological awareness and
acceptance. Lack of a sufficient level of friendly access may prevent RI from
becoming a truly ubiquitous method of identification.
References
[1] J. H. Arndt, âOptical alignment System,â
US Patent No. 4923297
, 1990.
R. B. Hill, âApparatus and method for identifying individuals through their retinal
vasculature patterns,â
US Patent No. 4109237
, 1978.
[3] R. B. Hill, âRotating beam ocular identification apparatus and method
,
â
US Patent No.
4393366
, 1983.
[4] R. B. Hill, âFovea-centered eye fundus scanner,â
US Patent No. 4620318
, 1986.
[5]
Eye fundus optical scanner system and method,â
US
Patent No. 5532771
, 1990.
[6] J. R. Samples and R. V. Hill, âUse of infrared fundus reflection for an identification
device,â
American Journal of Ophthalmology
, Vol. 98, No. 5, pp. 636-640, 1984.
[7] C. Simon and I. Goldstein, âA New Scientific Method of Identification,â
New York State
Journal of Medicine
, Vol. 35, No. 18, pp. 901-906, September, 1935.
[8] P. Tower, âThe fundus Oculi in monozygotic twins: Report of six pairs of identical twins,â
Archives of Ophthalmology
, Vol. 54, pp. 225-239, 1955.
[9] S. Yamamoto, H. Yokohuchi, and T. Suzuki, âImage Processing and Automatic Diagnosis
of Color Fundus Photographs,â
Proceedings 2nd International Joint Conference on
Pattern Recognition,
Copenhagen, pp. 268-269, August 13-15, 1974
.
[10] H. Yokouchi, S. Yamamoto, T. Suzuki, M. Matsui, and K. Kato, âFundus pattern
recognition,â
Japanese Journal of Medical Electronics and Biological Engineering
, Vol.
12, No. 3, pp. 123-130, June 1974.
[11]
A Performance Evaluation of Biometric Identification Devices
, Technical Report
SAND91-0276, UC-906, Sandia National Laboratories, Albuquerque, NM 87185 and
Livermore, CA 94550 for the United States Department of Energy under Contract DE-
AC04-76DP00789, 1991.
20
[12]