Rye Town, New York
July 20 - 21, 2005
Audio and Video-based Biometric Person Authentication
2005 (AVBPA) is a small event held every two years. The registration
is 168. There were 200 papers submitted of which 60 were accepted.
It was announced that a number of the biometrics events would be combined
into one large event. AVBPA, ICBA, SPIE Biometrics and Sinobiometrics
have been combined into the International Conference on Biometrics.
The next one will be held in Seoul, Korea August 2007.
Audio and Video-based Biometric Person Authentication
2005 (AVBPA) covers all areas of biometrics. From a research perspective
we saw important advances in Iris recognition and facial biometrics.
One of the more interesting aspects of the research is the ways in
which biometrics can be applied – including Iris recognition
done with cell phone cameras.
Biometrics is an emerging research area. There is much
to do and significant challenges lie ahead. Most of the individuals
here are from the discipline of pattern recognition. Yet, biometrics
requires more than pattern matching. Examples include 3D technology
and even lighting. The important areas which need to be addressed and
this can be seen here, include:
Biometrics Technologies
Biometrics Performance Evaluation
Biometric System Design
Applications
In the US, much of the research agenda is being driven
by homeland security, criminal/justice or defense. As a result the
research is near term and more focused.
In the effort to take biometrics everywhere it must be
mobile. Here at AVBPA we saw a number of presentations on MoC including
TOC (Template-on-Card). This work was focused on using today’s
very limited smart cards but nonetheless the initial results were impressive.
Iris Recognition at a Distance
The Sarnoff Corporation in Princeton, New Jersey presented
results from a DARPA funded program called Human ID at a Distance.
The motivation is to capture Iris data up to 30’. It is claimed
that Iris is important due to “no false matches in 2m comparisons” and
only 2% non- matches. One of the major problems with Iris is that current
systems are highly constrained. They must be at controlled distances
and under special lighting. A result is that a system has limited throughput.
The test system used two lenses for two cameras which
operated at 5m and 10m. IR was used for illumination at 880nm. There
were 128 pixels across the diameter of the Iris. Each image was captured
at 12f/s for 10 seconds. Variables in the collection included: 5 or
10m distance, angle 0 or 30 degrees, eye movement or tracking and lighting
that was background or spotlight. The technique used was 97% successful,
including with glasses, in locating the Iris.
The system was able to discriminate between the subject
and impostor in most cases. The factors which impacted performance
were summarized as follows:
Distance had no effect (5m or 10m)
Time (1month) had only a minor effect
Angle (30 deg) had a larger effect
Glasses had the largest effect.
At the end of the talk it was stated that a Portal System
was under development – that is, a system where individuals could
walk through a portal and have their Iris captured.
If this works well it could do much to improve the application
and operations of Iris recognition based systems.
Iris Recognition from a Cell Phone
The BERC (Biometric Engineering Research Center) of Sangmyung
University of Korea presented a paper on Iris Image Restoration. The
motivation was the most important aspect of the paper:
Need for a low processing load, high speed Iris detection
algorithm within a mega pixel camera, such as in a cell phone and
Development of Iris code extraction and recognition
algorithm in mobile device considering sunlight.
Making Face Recognition Work Under Variable Lighting Environments
Toshiba and University of Tsukuba presented a technique
called Constrained Mutual Subspace Method (CMSM) which is to allow
face recognition under changes in pose and illumination. It claimed
that this technique lowered the error rate by a factor of 2 over the
more traditional methods.
Facial Object Model for Gesture Variations
One of the problems with facial recognition is the ability
of the individual to have many facial gestures. The Institute for Neurocomputing
at the University of Bochum, Germany developed a flexible object model
that allows for the recognition and synthesis of facial expressions.
This is based on bunch graphs which uses Gabor jets. What it enables
is the recognition of faces after matching which is implemented by
the rapid comparison of many faces. The technique does both gesture
and pose normalization.
Impact of Lighting on Face Reconstruction
SUNY at Stony Brook, New York, developed a morphable
model to recover facial shape. This was then used to recover both texture
and illumination. The technique is based on spherical harmonics. The
value of this approach is that Lambertian reflectance can be represented
by the combination of the first 9 spherical harmonics. It was claimed
that the average shape error was only 3.5% and the average appearance
error 2.8%. Given the variation of pose and illumination the results
were impressive.
Applications of Face Models
Simon Baker of Carnegie Mellon gave the keynote on Model-Based
Face Analysis. Face models are “black boxes” that allow
one to control model parameters that include shape and appearance and
from these create a face image. Some of the models include 2D Active
Appearance Models (AAMs) and 3D Morphable Models (3DMMs). There is
an inverse process where a facial image can be used to create via a
fitting algorithm a face model. The fitting process is where the model
parameters are found which best matches the image. A point made by
Simon is that these models need to run in real time – 30 or 60f/s – at
video rates.
The most interesting aspect of the presentation dealt
with the applications.
Mouse or Joystick replacement – used by disabled
individuals or children
Smart airbags
Windshield Display overlays
Diver Monitoring
Intention detection – such as mother baby studies on child development
Audio – Visual speaker recognition and identification
Expression transfer
Animation generation
Low bandwidth video conferencing
Fingerprints on a Cell Phone
The Center for Biometrics and Security Research, Institute
of Automation, Chinese Academy of Sciences examined how fingerprint
authentication could be done on a mobile phone. A system was implemented
on a BIRD E868 mobile phone. This worked by passing a fingerprint feature
data base, assumed small, to the phone where the actual recognition
is done on the features. Emphasis was placed on reducing the computational
time and load. Applications are seen in identity protection on the
phone and e-business from the phone.
Driving Facial Recognition to the Next Level of Performance
The graduate school of Engineering at Osaka Electro-Communication
University examined facial recognition using multiple means of creating
the facial image. The technique used two images, one a thermal imager,
and another was based on multiple images, i.e., a sequence, which also
collects a color stereo image. 3D measurements were used to determine
the apparent size and position of the faces. The decision system uses
trainable classifiers and in the example presented it was with a neural
network. The data set was only 30 images. The results were impressive:
recognition rates varied from 97% to 100%. It was claimed with additional
work it could be 100%. Caution should be exercised in assuming these
results will apply to a large scale application.
MOC – Michigan State University
Michigan State University asked the question in a poster
paper “Can fingerprint template information be secured in a resource
constrained devices such as smart cards without sacrificing matching
performance?” Their system uses the scanner to do template extraction
but the card holds the template, does MOC and non-critical parts of
the matching process are done in the scanner. The matching technique
is based on triplets. The on card data includes a pruned triplet map,
ridge features map and a personal information code. The card determines
the transformation values of the triplet information which is then
sent to the scanner. It extracts the ridge feature map from the presented
image. The card then compares the ridge feature map and provides a
match score. If the match score meets a predetermined threshold a verification
signal is generated. Matching performance is claimed to similar but
more computationally demanding techniques.
MOC – Institute for Informatics and Telematics – Italy
This technique uses a template stored on the card but
the actual matching is done on a secure local host, i.e., PC. To enhance
security a PKI security module is retained on the device which is used
to enhance PIN security. A Mobile Agent Runtime Environment is implemented
on the card in Java.
A second paper proposes an asymmetric fingerprint matching
algorithm for a Java Card. One of the most challenging aspects of working
with smart cards are the limited resources. They cite:
5 – 10 MHz CPU
1k RAM, 32Kb EEPROM
JCVM does not support threads, and garbage collection
Java Card bytecode is interpreted
The technique uses local minutiae for matching and only
20 minutiae are supported per template. In order to reduce the time
to MOC the matching technique stops when fewer minutiae are “well” matched.
Not only is a low FAR required but matching time must
be reasonable. They found that a FAR of .1% was achieved in two tests
at 69% and 88% of the time within 8 seconds. The performance was also
dependent on the quality of the enrollment image. They found that most
matches were accomplished in 1 – 8 sec. Using low end smart cards
the FAR was .1% with a FRR of 7.3%.
Improving Liveness Detection with Perspiration
Pattern Matching – Clarkson
University
One of the weaknesses of fingerprint matching is the
susceptibility to spoofing. There have been a number of techniques
proposed for liveness detection and the use of perspiration patterns
is one of the most unique. Based on these early results there are indications
that fingerprint perspiration is but another biometric which can enhance
the security of fingerprint matching.
The technique uses wavelets to perform the perspiration
pattern matching. The actual perspiration patterns were found by an
inverse transform using difference coefficients between images. Examples
were shown of difference images with a cadaver and spoof and they had
no perspiration patterns.
The tests covered a range of individuals by age, ethnicity
and sex. Fingerprint images were collected over 5 seconds and 5 months.
Scanners of the optical, electro-optical and capacitive DC types were
tested. Only similarity scores were presented. But the results showed
that perspiration patterns were unique and there was good consistency.
It was recognized the future work needs to test a larger group of individuals
and consistency needs to be tested over a longer period of time.
Using Fingerprints for a Security Key to Improve Overall Security = Michigan
State and IBM
The technique uses fingerprint minutiae locations for
locking and unlocking a secure vault. The key length is 128 bits. The
upside of this technique is that the FAR is 0%. Of the 9,900 attempts
to unlock the value none were successful. The down side is that the
GAR was only .79. That is 21 of the 100 query templates could not work.
In the future they will work on better template alignment with the
expectation of lowering the FRR.
Ear Biometrics – Some Promise – University
of Southampton
Coming from left field a matching technique using “force
field convergence” from energy physics was shown to have excellent
results on ear biometrics. One of the problems with ear biometrics
is that the ear is 3D when the depth of the ear features are included.
To handle this, based on 2D images only, a force field method was proposed
which uses a convergence field to map the ear. The results were impressive – out
of 252 samples from 63 subjects the recognition rate was 99%.
Finger Surface Features as a Biometric – First
Attempt does not Make the Grade - Notre Dame
This technique uses 3D finger surface data. 233 subjects
were tested over a 4 month period. Both range and intensity images
were collected of 3 fingers – central 3. The matching technique
was based on a shape index. This index allowed for a means to classify
various shape patterns such as a spherical cup, dome, and tough. Of
the total 335 verification experiments performed the equal error rate
varied from 5.5% to 15%.. For future work it was suggested that a large
data set be collected, better templates developed and fusion techniques
tried to combine 2D with 3D date.
WAVE Comments
We continue to come away amazed at the various forms
of biometrics being considered and seriously evaluated. In essence
biometrics is about uniqueness of the individual. But the current emphasis
is on seeking ways in which technology can detect a biometric and recognize
individuals based on the metric. Thus, as suggested in this report,
the following lays out the scope of biometrics:
If it is alive, walks, talks, acts and/or is a part
of the body there can be a biometric associated with it.
AVBPA exposed the reality that biometrics can go well
beyond its security application. In particular was the paper by Simon
Baker of Carnegie Mellon which discussed many biometric applications
including as a HCI input device. As we step away from a biometric in
security applications this brings a new perspective. For example, the
role of uniqueness, as outlined above becomes much less important.
One of the stand outs from AVBPA is the role of facial models. Used
as part of developing a facial template it was shown to have much broader
applications, even including video conferencing.
The use of biometric technology beyond security has barely
been explored. This is a market opportunity that remains to be developed.
