Biometrics Matures; Vision Gets Nod

Wouldn’t it be nice if you could never lock your keys in the house? Or if a law enforcement officer never had to be concerned about fake IDs because an ID existed that couldn’t be duplicated or changed?

Those are the premises behind biometrics. Biometrics is the process of converting unique physical features from a human being -- preferably a feature that is unchanging over time -- into a mathematic representation that is also unique, storable and verifiable. During the past decade, machine vision has been a major part of biometrics. Recently, machine vision’s involvement in biometric technology has grown as security needs in all nations have increased -- right along with displeasure from the citizenry for the additional security measures.

But rather than discouraging the acceptance of machine vision in biometric security, public safety concerns in light of escalating terrorism and violence, plus public resistance to ‘invasive’ security measures, are actually promoting the use of machine vision as a leading remote sensing technology capable of making facilities and geographic borders secure with minimal personal privacy invasion.

Putting Vision In Its Place

Like other surveillance and security systems, more and more biometric systems are using machine vision technologies to help speed, simply, and automate the enrollment, identification and verification processes*. For biometrics, machine vision has found a secure spot in the ‘‘front end’‘ or acquisition of the biometric key, during both initial enrollment and during subsequent verifications.

This article will look at the major biometric keys, how and why machine vision can play a roll, and the technical needs that drive the machine vision technical specifications for biometrics.

The best biometric keys are based on unchanging physical features unique to each person, including fingerprints, facial features, iris of the eye, hand geometry, and voice. Each has its pros and cons. Fingerprints and palm prints can become worn, and are affected by extreme dry skin, damage caused by manual labor, etc; law enforcement has more fingerprints in their databases than any other biometric, however, and law enforcement is the main customer for biometric systems. Facial features are relatively constant regardless of weight gain or hair growth, but authorities are only starting to build large facial recognition databases, and while database can be done offline with stored images, etc., the success of these systems is critically dependent on the quality of the image – as with all biometric keys. The color pattern of the iris is extremely accurate and remains essentially unchanged from shortly after birth all the way to death, but free societies resist invasive security measures that require you to stop and place your face to a microscope for identification or verification.  Hand-vein geometries also have to overcome public dislike of ‘invasive’ biometrics that require the individual to stop and touch a security device. Voice identification is not as accurate, doesn’t use imaging technologies, and is usually integrated with a second, more secure, biometric.

Of these biometric systems, fingerprint is the most prevalent because it has been in use in non-digital formats for decades. Facial recognition is the second most used, followed by hand geometry and iris.

A Smart Resolution

Whether it’s a fingerprint, face, hand, or iris, biometric systems follow the same general format. A camera on the front end of the biometric system takes a high-resolution photo of the person during enrollment, which can be active or passive as in the case of extracting facial or iris biometrics from surveillance cameras. An algorithm analyzes the features of the image looking for key minutia, and the resulting pattern is then converted into a mathematic value that can be stored in a smaller footprint than the original image. This is called enrollment. During identification and verification, the process repeats itself, and the acquired image is analyzed and checked against the stored database value to identify an individual.

Dr. Joseph Atick of L-1 Identity Solutions (Stamford, Connecticut) has spoken at several conferences on biometrics and says that the technology, standards and practices are maturing to a point where the biometric industry is able to suggest specific imaging requirements for different types of biometric systems.

‘‘Biometrics has definitely benefited from the CMOS camera revolution and the cheaper, faster cameras that CMOS delivers, as well as the advent of high-definition cameras,’‘ explains Atick. ‘‘Automated gain control is also important, but we’re looking for a closed-loop automated gain control circuit that can establish the gain specific to the face, and not just the background.

‘‘Smart cameras are also something the biometric industry needs,’‘ continues Atick. ‘‘More than just the output of the digital frame, we need the cameras to be able to just output the face, for example. We would like more image processing on the board, particularly for face and eye finding when subjects are on the move. Image processing can remove eye lashes, for example, and insure image quality. We’ve been very impressed with Texas Instrument’s DaVinci image processing card for developing these types of applications.’‘

‘‘DaVinci was specifically designed to ease the development of vision and imaging based applications,’‘ said Brooke Williams, Texas Instruments ' Vision Business Manager (Stafford, Texas). ‘‘Biometrics, automotive driver assistance systems like lane departure and blind spot detection, and machine vision smart cameras are perfect examples of applications that benefit from the holistic approach of DaVinci technology by offering leading silicon, software, hardware tools and support that are all optimized for vision and imaging to allow customers to offer lower cost, differentiated solutions to the market.’‘ As more processing finds its way into the camera, it will also reduce the need for high-bandwidth I/O and networks.

‘‘We like gigabit cameras that are offered by companies like Prosilica (Burnaby, British Columbia, Canada) says Dr. Juergen Pampus, Vice President of Sales and Marketing for Cognitec Systems (Dresden, Germany). ‘‘We need high bandwidth. It’s not only the resolution of the image, but also that we can’t introduce any artifacts by compressing the image. We needed an uncompressed image, and if that’s a 3 megapixel image at 10 to 20 fps, you need bandwidth to handle that transfer. These cameras are also easily integrated into a network, which helps, but it’s the bandwidth that we need today.’‘

Deinterlaced images are also critical, adds L-1’s Atick, for the same reason - interlacing fields can introduce artifacts in moving subjects.

Spatial resolution with few if any artifacts or distortions is critical to success in biometrics. According to Uzi Moshe, CEO of Micro Utility Ltd., (Lev-Hashron, Israel), a biometric integrator, a facial recognition pilot project in the city of Jerusalem had a success rate of 15% using standard 640x480 pixel surveillance cameras for enrollment and verification. The government agency told Moshe that if he could achieve 40%, they would buy ‘‘as many as they could.’‘ By moving to 3.2 megapixel cameras for enrollment, and 2 megapixel cameras for biometric acquisition at border crossings, the success rate soared to more than 90%. The client was happy.

‘‘Going to 5 to 6 megapixels for enrollment would be even better, but cost becomes an issue when you’re looking at rolling out 650 workstations for the final project,’‘ explains Moshe.

Fingerprints: A Rolling Problem

Several years ago, inexpensive capacitive sensors were thought of as the best solution for collecting digital images of fingerprints. Unfortunately, these sensors proved to be too susceptible to damage and had difficulty collecting images of damaged skin common to manual laborers.

Today’s fingerprint biometric systems often want to be able to collect both finger prints and palm prints, along with hand geometry. At resolutions of 1000 dots per inch for an entire hand, the resulting image can approach 20 megapixels. Because of this high-res requirement, says Atick, customers prefer CMOS imagers to contact sensors. The cameras need to be able to image up to 30 fps because law enforcement needs a fingerprint ‘‘from nail to nail.’‘ This means the subject has to roll the finger across the sensor. While palm and hand may only need 10 fps, fingerprints need higher speeds so that the system will have multiple images to choose from when selecting the best possible image displaying key regions of interest.

Despite representing the largest biometric database on the globe (think FBI fingerprint database and multiply that by the number of industrialized countries), fingerprints are typically viewed as an invasive biometric. Today, fingerprints are used as a back up biometric whenever possible, typically supporting facial recognition as a primary identification method.

Face to Face

At first blush, the spatial requirements for facial recognition would seem to be more relaxed. ‘‘The minimum standard measure that we use is 50 pixels between the eyes for facial recognition,’‘ says Cognitec’s Pampus. ‘‘But that becomes more challenging depending on how far the face is from the camera. You could do it with a standard CCTV camera, but it’s not likely to be close enough to the subject to work.’‘

Facial recognition is also branching into 3D. Older systems used laser scanners to create a 3D map of a person’s face at a distance, but these systems cost several tens of thousands of dollars, which is too much for the number of deployments government agencies require today. Newer systems use infrared (IR) LEDs projecting a grid onto a person’s face and imaged by a silicon-based CMOS camera. Atick claims this approach allows the 3D image of the face to be collected at ‘‘arm’s length’‘, but adds that newer cameras optimized for light in the 760 nm spectral band would significantly improve these systems’ operations. The addition of stronger IR LEDS would also increase the acquisition distance and utility of the systems. Companies are investing in pulsed IR sources that can provide more light without posing a risk to the IR-sensitive retina.

Look Into My Eyes

Near infrared (NIR) sensitive cameras would also help biometric systems that are based on the iris. Eyes are covered in a thin layer of water that makes the surface highly reflective, while the iris pattern is actually located a few nanometers below the surface of the eye. Having cameras optimized for the NIR would greatly improve iris biometric acquisition because IR radiation penetrates tissue to the depth of a few millimeters.

Moving to large pixel microbolometer, or other inexpensive IR cameras isn’t an option, however. Not just because of the cost of non-CMOS IR imagers, but because of losses in resolution and speed.

According to Atick, iris biometrics require 200 pixels across the iris, which is only about 1 cm wide. The eye also constantly moves, even when we think it’s still. This ‘‘psychotic’‘ movement allows the brain to process information, but plays the devil with high-resolution imaging. The problem becomes more pronounced when you try to design a system that doesn’t require the subject to place their eye to a microscope objective. U.S. trips in Afghanistan are using handheld iris biometric systems to check for insurgents, but to move the system even farther away will require some unique combinations of high-quality, high resolution imager, with NIR sensitivity and motion filtering to provide high-resolution iris images at a distance greater than arms length. Finally, although telephoto lenses can zoom to provide 200 pixels on an iris from a distance, they suffer from a small depth of field, which complicates passive acquisition of iris biometrics from surveillance and similar data sources. L-1 is working on adaptive optics solutions that will correct for wavefront errors and extend the depth of field for iris biometric systems.’‘

According to Atick, when that combination hits the market, iris recognition will grow significantly in market size because of its high accuracy and passive collection capabilities.

In today’s security conscious world, biometrics offer a convenient and affective answer, as well as an exciting opportunity for machine vision cameras, image processors, optics and lights.

‘‘We need to be in the sweet spot of machine vision,’‘ notes Atick, ‘‘and that’s a challenge because we’re on the cutting edge of many technologies, and the cutting edge is never the sweet spot. Sometimes, we have to wait for the technologies to be developed that catch up with our requirements. I can say that I’ve been very impressed with the machine vision community’s responsiveness, especially those represented by the Automated Imaging Association. I’m sure that over time they’ll be able to meet our needs.’‘

*Editor’s Note: Biometric systems can ‘identify’ a person from large stored data base, or ‘verify’ a person based on a code located on a portable storage device, such as a smart card or passport.