WHITE PAPER Iris Recognition The state of the art in Algorithms, Fast Identification Solutions and Forensic Applications Kirsten R. Nobel, PhD Principal Solution Engineer
Contents 2 table OF CONTENTS 3 Introduction: Why Iris Recognition? 4 What is the Iris? Where is it? How is Iris Recognition Different from Retinal Scanning? 5 Brief History of the Commercialization of Iris 6 Basics of Iris Recognition 8 Objective Assessment of Algorithm Accuracy 10 MorphoTrust Iris Technology 11 Capture Devices 12 Using Iris for Very Rapid Identification 14 Forensic Iris - New Work on Iris Image Enhancements 15 Conclusion 2
Introduction: Why Iris Recognition? While law enforcement agencies are in a relatively early stage in building iris databases, the U.S. Department of Defense and several international governments have already compiled large repositories. The accuracy afforded by, together with its remarkably fast search speed and small template make an attractive method for identification and viable for even the largest National ID programs. Thus, iris is increasingly being included in multi-modal capture programs. Iris is a very effective biometric for ascertaining identity quickly and in a manner that does not have the same criminal connotations in the U.S. as fingerprinting. For this reason, iris is also gaining traction in physical and logical access control programs, including the U.S. government s PIV card program for authentication of federal employees and contractors. Adoption of iris in civilian programs is not hindered by the cleanliness issues associated with fingerprint capture devices that concern some cultures, nor by the poor quality fingerprints associated with manual laborers and some elderly citizens. This white paper will describe what an iris is from an anatomical standpoint and will discuss the fundamentals of, focusing in particular on the Daugman algorithm upon which our technology is based. We provide a brief history of the commercialization of the technology starting from the original patent and moving quickly through to today s groundbreaking programs in which iris is used in National ID solutions. We also describe modern iris capture devices. Next we talk about applications of iris technology that may be of particular interest to the law enforcement community. Finally, we share our vision for how the forensic iris examination and mark-up software that we are developing will increase the value of automated systems. 3
What is the Iris? Where is it? How is Iris Recognition Different from Retinal Scanning? The iris is the colored portion of the eye visible around the pupil. It is covered by the cornea. The iris is the only internal part of the body visible from the outside. It is formed before birth via a process known as chaotic morphogenesis and is very stable after that time. Iris recognition begins with capture of a photograph of the eye (Figure 1). No scanning is involved; rather, iris images are acquired by taking a photograph in infrared light. Features of the iris including those created by the collagenous fibers comprising it have been referred to as contraction furrows, coronas, crypts, freckles, rifts and pits. Iris recognition is basically the matching of these features across images. Fig. 1: A. Color photo of human eye showing iris, pupil and sclera (white of the eye). B. Image of eye captured in infrared light, labeled to show iris and other anatomical structures of the eye and surrounding area. Also shown is a shadow on the iris created by the eyelid. Such areas can be removed in pre-processing the iris image. Retinal scanning is a completely different technique in which the retina, which lies at the posterior surface of the inside of the eye, is imaged. Retinal scans typically look at the pattern of vasculature (minute blood vessels), which can also be used to discriminate between people. Because retinal scans convey information that can be used to diagnose certain medical conditions (e.g. diabetes, glaucoma, degenerative diseases), there are privacy concerns associated with its use. MorphoTrust does not provide retinal scanning technology. 4
Brief History of the Commercialization of Iris The idea of using iris patterns for personal identification is fairly old. It was originally proposed in 1936 by ophthalmologist Frank Burch. In 1987, two ophthalmologists, Leonard Flom, M.D. and Aran Safir, M.D., received a patent for their Iris Recognition Technology. John Daugman, Ph.D., then at Harvard, worked with Flom and Safir to develop computer algorithms for the concept. Daugman patented the algorithm and the three founded Iridian Technologies, which was one of the early biometrics industry companies. In 2008, the Daugman patent expired and the technology has continued to advance with R&D efforts at several biometric companies. Dr. Daugman continues to consult with Morpho today. The gradual accumulation of knowledge about how to best apply iris matching technology, together with advances in sensors, has led to the adoption of iris technology in a wide range of programs including important military applications (capture of irises, among other biometrics, on untethered devices in areas of conflict in the middle east) and huge civilian programs aimed at providing unique identification numbers to every person in India and Indonesia. Recently, the Federal Bureau of Investigation announced that it will be adding iris to the finger and face biometric systems that comprise its Next Generation Identification program. After an in-depth study, Lockheed Martin has selected MorphoTrust USA s ABIS system, which the FBI has now begun to implement. MophoTrust s ABIS will provide the FBI s NGI program with superior biometric matching for face and iris modalities. In the past several years there has been a proliferation of companies selling iris capture devices. Today, these cameras are considered commodity items. With India s UID program as a primary catalyst, the cost of iris cameras has dropped to the same cost as a standard digital camera and in some cases, much less. The challenging environments and tremendous number of irises captured from people ranging in age from small children to elderly people in India and Indonesia has also been a factor in improving the usability of iris cameras. They are available in a variety of form factors, one and two-eye capture models, and at prices that depend upon their sophistication. There are several multi-biometric capture devices on the market today. Other devices are simple, lightweight and meant only for iris capture. Cutting-edge devices automatically acquire irises at distances of 10 feet or more, while subjects are moving. 5
Basics of Iris Recognition Iris recognition analyzes features in the colored tissue surrounding the pupil. A typical iris has at least 200 unique, identifying features that can be used for comparison, including rings, furrows and freckles. Iris technology combines techniques from the fields of computer vision, pattern recognition, statistical inference and optics. Figure 2 illustrates at a high level the basic steps involved in turning a picture of an iris into a small binary code. Figure 2. The process explained in four basic steps. The process of begins with image capture (Step 1 in Fig. 2) and segmentation of the image of the iris from the pupil, eyelid, and eyelashes (Step 2). Segmentation involves registering the outside of the iris border and the inside edge next to the pupil, forming a donut shape. The texture of the image inside this donut is then analyzed and encoded using various methods (Step 3), many of which stem from the original Daugman technology. Finally (Step 4) this code, or template, can be compared with one or more previously stored templates. 6
october 2012 MoRphoTRusT TM IRIs RecognITIon Because the iris is not perfectly circular, we apply a technique called Active Contours (Figure 3A) which can accurately represent the non-circular aspects of the iris borders. Fig. 3: A. Adaptable Active Contour technique. Lines are drawn between iris and sclera (sclera boundary, green heavy line) and between pupil and iris (faint, light line). This step refines the initial iris segmentation. B. Gaze correction. Sometimes images are captured of eyes not looking directly into the camera. These must be remapped to frontal before the iris segmentation step shown in A. Another advanced iris pre-processing capability is the ability to correct images obtained by subjects not looking directly into the camera. If an iris image is off angle and cannot be recaptured, then gaze correction algorithms are applied prior to segmentation (Figure 3B). Once the iris image has been accurately segmented, we unwrap the iris image texture from an annular ring, as it appears in the images above, to a polar representation. Figure 4 illustrates the process of transforming the image from its original form to a linear one. The iris in the rectilinear image (Figure 4, left) is rolled out into a fixed size polar iris image (Figure 4, right). The advantage of the polar representation is that it removes iris scale and dilation, and a rotation of the iris becomes a horizontal shift in position. 7
Fig. 4. The same iris image in rectilinear (left) and polar (right) formats. The red areas in both images are the masked occlusions (eyelids, eyelashes). At this stage, we extract features and create a template. Gabor filtering and encoding of phase and amplitude information results in a 2D code like that shown in the upper left corner of Figure 2. Our code is 512 bytes total. Depending on whether the application is for one-to-one verification or one-to-many identification, an iris of interest ( probe ) can subsequently be compared to one or more previously stored iris templates ( gallery ). A match score that quantifies the similarity between each probe and gallery image is created by counting the number of bits that differ between two templates. We then convert that measure to a more meaningful score which represents the expected false accept rate of the match. This score gives a user an indicator of the confidence we have that this is an actual match. Due to the relative richness of the iris code, match score confidence levels are typically very high. Objective Assessment of Algorithm Accuracy The U.S. government has been involved in biometric vendor testing since the earliest days of the industry, in order to establish objective metrics for comparing technologies and therefore, to support users in making procurement decisions. Their objective, non-biased, scientific approach makes NIST tests the gold standard of accuracy assessment worldwide. The algorithms from all major vendors have been carefully benchmarked and results are publicly available at http://www.nist.gov/itl/iad/ig/irex.cfm. 8
IREX III, the latest and largest scale test of commercially available iris recognition technology, is publicly available from the link above. IREX III is the first independent test of one-to-many identification using a large, real-world dataset. It is considered the definitive benchmark of iris recognition technologies available from around the world. The exciting main message of the test was that iris is a viable biometric for large scale identification and an order of magnitude more accurate than face. Not surprisingly, IREX III also concluded that enrollment and search of both eyes offers better accuracy than a single eye. IREX III was intended to allow an objective, quantitative comparison between vendors. The MorphoTrust technology (then L-1) was shown to be the most accurate at the most demanding operating points while being much faster and smaller (more efficient) than all nearest competitors (see Figure 5). Our technology was also shown to have an increasing accuracy advantage as the database grows and to have stable scoring that allows users to set reliable thresholds that do not require adjustment as the database grows. Fig. 5: Summary graph of IREX III results. Speed, accuracy and template size were combined to compute an efficiency metric. Like other vendors, we submitted multiple algorithms. MorphoTrust submissions are all coded with a U and cluster at the left side of the graph. The NIST test showed that our Daugman-based iris technology is the best rated, with the ideal balance of speed, size and accuracy. 9
MorphoTrust Iris Technology Here are a few quick iris performance statistics that illustrate the power of the technology today: False Non-Match Rate is less than 1% using two eyes, at extremely low False Match Rate (e.g. 0.0001%) using existing commercial iris capture devices. Using a single eye, results are nearly as good. Iris match speed of 120M eyes/sec available now in ABIS 7 Because of their small size, about 100M iris templates can be stored on a single 64 Gb RAM server. (1.5M templates/1 Gb RAM) 2-eye iris accuracy is much more accurate than face and is also measurably more accurate than 10-print fingerprint. Furthermore, because of the efficiency of the iris template, 10-fingerprint level accuracy can be achieved using iris on a mobile device that stores data for a very large population. In addition to progress in algorithm accuracy and speed, there has been equally impressive progress in biometric platform architecture over the past several years. Today, the best search recognition engines perform multi-biometric (face, fingerprint, and iris) searches in the same software application, thus reducing system complexity. They run on commercial off-the-shelf (COTS) server hardware and use commercial databases for data storage. Scalability has been proven even in identification programs that include hundreds of millions of records. Finally, interfacing with mature COTS search engines is relatively simple due to the open standards-based service oriented architectures. For example, the Electronic Biometric Transmission Specification (EBTS) standard can be used for input/output transactions in order to seamlessly integrate with other law enforcement or DoD IT systems. 10
Capture Devices A variety of iris capture devices provided by Morpho and other vendors take good quality images with little or no user training and output the images compatible with use of our algorithms (JPEG2000 and PNG are the recommended formats.) Some cameras simultaneously capture images of both the left and right eye. Some capture iris images at a distance. There are iphone applications for iris image capture. There is at least one device in development for simultaneous capture of iris and face images. Iris cameras can be handheld, wall-mounted, mounted on an adjustable height pole or tripod, or for covert applications, mounted behind a sign or within a display that draws the attention of subjects of interest. 11
Using Iris for Very Rapid Identification For law enforcement and military personnel, it is often critically important to identify an individual as quickly as possible. For example, police need to know if the person standing in front of them is wanted for murder. Significant time and effort is wasted when officers fill out forms for every arrest of repeat criminals. Sometimes criminals provide aliases that lead to wasted time in the booking process. Another problem faced in law enforcement is movement of criminals from one physical location to another (e.g. local precinct, courthouse, prison). Ensuring that only inmates qualified for the work release program exit the facility and get on the bus, for example, is not a trivial problem. Dispensing medications to appropriate individuals is another challenge in a correctional facility. All of these problems could be addressed by using iris to rapidly identify a person. Basically, an iris is quickly captured and searched against a database and a result is returned within seconds. (Fingerprints, the most commonly used ID method in law enforcement, can take several minutes, maybe as many as 30 minutes, to acquire and minutes to hours to search, depending on the size and location of the AFIS.) Rapid iris-based identification can be performed at multiple points in the process of moving a person from location to location, as required. Several law enforcement jurisdictions have successfully used iris matching technology to rapidly and accurately identify criminals in this way. In one state, we deployed 64 multi-biometric booking workstations at major arrest and arraignment locations (sheriff s offices, police departments, correctional facilities) to begin building an iris identification database that will allow for rapid identification in the criminal arrest, arraignment and incarceration process. For example, within seconds, law enforcement officers will be able to know with certainty whether the correct individual is being released. 12
Fig. 6: Offender ID solution components. Workstations running our Offender ID software and connected to iris and face image capture devices (Livescan fingerprint can also be captured.) These might be located in Sheriff s offices, police departments and/or jails. The Offender ID software sends ABIS a file in a standard NIST format (NIST ITL-1-2011). ABIS creates records organized by individual and arrest (Identity Manager) and creates templates that are added to biometric search galleries (Search Engine). TPE Management Server manages software updates, keeps track of user-defined metadata and performs various other management functions. The Offender ID system (Figure 6) consists of enrollment software running on a workstation, together with cameras for capturing iris and face and potentially a scanner for capturing fingerprint. After an iris is captured and compared, one-to-many, against the gallery of irises in the back-end ABIS system, a result is returned. If no match to the newly acquired probe iris is found under a different name, then a new record is created for the individual. If a match is found under a different name, then the booking process proceeds as an update to an existing record. A new face image is captured, new fingerprints may be captured, and the centrally stored record is updated with the new biometric and encounter (arrest) data. The non-biometric information is stored in the ABIS IdentityManager database, while biometric templates are stored in the ABIS Search Engine. A Management Server provides automated software updates to other components in the solution, performs user management functions, and plays a role in importing legacy data. 13
The Offender ID solution described above allows users to do a quick identity search at booking and also allows prisoners to be tracked as they move from one facility to the next (during work release, medical care outside of the correctional system or in an emergency situation). Iris matching can also be performed on a handheld device to verify identity before medications are dispensed. Importantly, this solution provides high-level administrators with a means of ascertaining the total number of criminals in a facility or in the state, during a given time period. Offender ID s ability to report the number of persons in a particular location can be a valuable feature in the event of an emergency. It can also provide usage metrics relevant for funding purposes, such as total number of criminals booked and number of inmates currently housed per facility or system-wide. Using Offender ID software in conjunction with ABIS enables fast, accurate identification and the ability to track an individual from one point in time or one physical location to the next. Iris matching technology, implemented as part of a solution like the one described above, is the only biometric that can provide this capability. Forensic Iris - New Work on Iris Image Enhancements Iris images acquired in challenging environments or from uncooperative subjects may suffer from image problems. These images may not provide good match results. In this case, various techniques can be used to enhance the image to make it usable for automated matching. MorphoTrust is developing a number of image enhancement tools for improving performance. One of these is a prototype Iris Examiner Workstation, which is analogous to a Latent Examiner or Tenprint Examiner Workstation. Low quality images are never guaranteed to produce a hit. However, the tools provided in this software improve the likelihood. Such a tool could serve as a front-end to the ABIS search engine, where biometric matching takes place. One of the most commonly encountered issues with iris images is that the iris is registered incorrectly. The tool allows for full markup of the registration data needed to enroll and match an iris. 14
Conclusion Iris recognition technology provides unparalleled accuracy and speed and has been field proven in very challenging environments (e.g. DoD military operations overseas, India s UID program, Indonesia s National ID program). Because of its reliance on very small templates, is also attractive from a storage and data transmission perspective. The confidence in the technology following the successes of the DoD and large National ID programs, together with the fact that iris cameras are now commodity items, may be factors in the increased rate of adoption we are seeing. MorphoTrust s iris algorithm is leading the industry in accuracy and speed. Templates created from this Daugman-based algorithm are increasing in number rapidly in programs around the world and number in the hundreds of millions already. From law enforcement and military applications, to civil programs such as travel document issuance and border crossing systems and even commercial applications (e.g. banking), iris solutions are definitely gaining traction. Iris solutions may include image capture with any number of commerciallyavailable cameras and enrollment/booking software like our Offender ID or Multi-Biometric Capture software. The software interfaces with a biometric search engine, such as MorphoTrust s ABIS platform. The scalable search engine performs identification and verification. New work in the area of forensic iris is showing promise in improving the quality of otherwise unusable images. About the Author Dr. Kirsten R. Nobel is Principal Solution Architect at MorphoTrust USA. She designs large scale identity management systems, staying close to the latest research and development on core algorithms. Kirsten was involved in the formation of the IBIA and has participated in biometrics standards development at both the national and international levels. She earned her doctorate in Brain and Cognitive Science from the University of Rochester where she conducted research on neural mechanisms of vision. 15