Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

objectives creates special challenges for biometric sys-

tems, because the design strategies and indeed the

core technology choices that may maximize through-

put volumes are not necessarily the same as those

that maximize identiﬁcation accuracy. This article

reviews ﬁve ways in which automated iris recognition

[1] is used within airports and at border-crossings,

with special attention to those trade-offs and design

issues.

Arriving International Passengers:

Iris Recognition Instead of Passport

Presentation at Immigration Control

The use of biometrics as living passports, removing the

need for actual passport presentation at Immigration

control, was pioneered in the UK in 2002. A 6-month

trial of the EyeTicket JetStream system allowed a total of

2,000 frequent travelers from North American to Lon-

don Heathrow Airport to enroll their

▶ IrisCodes and

thereby to bypass Immigration control upon arrival,

passing instead through an automated iris recognition

gate. The trial was deemed fully successful and led

eventually to a large-scale system deployed by the UK

Home Ofﬁce, called IRIS: Iris Recognition Immigration

System [2]. Based on the same core Daugman algo-

rithms [ 1] but with a more user-friendly interface, the

IRIS system is today deployed at most major UK air-

ports, including all ﬁve terminals at Heathrow. The

architecture incorporates a centralized database of en-

rolled IrisCodes so that travelers can use the system

regardless of their airport or terminal, although this

also makes the system vulnerable to interruptions in

communications links or reductions in bandwidth.

Such network failures in the ﬁrst year of deployment

occasionally inter rupted the service. Nonetheless, as of

May 2008, the UK Border Agency announced that

more than one million passengers had successfully

used the system, with enrollments increasing by

about 2,000 per week, and with the system handling

about 15,000 arrivals per week. By substituting for

passport presentation, the system replaces long queues

at arrivals with an expedited automated clearance at

iris camera gates within a matter of seconds (Fig. 1).

A crucial aspect of the IRIS system is that it oper-

ates in identiﬁcation mode to determine a passenger’s

identity, not in a mere veriﬁcation mode in which an

identity is ﬁrst asserted (for example by presenting a

token, passport, or smartcard) that is then simply

veriﬁed. The requirements of biometric operation in

identiﬁcation mode by exhaustively searching a large

database are vastly more demanding than one-to-one

veriﬁcation mode in which only a single yes/no com-

parison with one nominated template is required. If P

is the False Match probability for single one-to-one

veriﬁcation trials, then (1P

) is the probability of

not making a False Match in single comparisons. The

likelihood of successfully avoiding any in each of N

independent attempts is therefore (1P

)

, and so P

Iris Recognition at Airports and Border-Crossings. Figure 1 The UK Government’s IRIS program has enabled more than

a million registered travelers to enter the country via several British airports using only automatic iris recognition for

identiﬁcation, in lieu of passport presentation or any other means of asserting an identity.

820

Iris Recognition at Airports and Border-Crossings

the probability of making at least one False Match

when searching a database containing N different pat-

terns, is

¼ 1 ð1  P

ð1Þ

Observing the approximation that P

 NP

for

small P

<< 1, when searching a database of

size N an identiﬁer needs to be roughly N times better

than a veriﬁer to achieve comparable odds against

making False Matches. In effect, as the database grows

larger and larger, the chance probability of making a False

Match also grows almost in proportion. Obviously the

frequency of False Matches over time also increases

with the frequency of independent searches that are con-

ducted against the database. These considerations make it

vital that such identiﬁcation applications operating by

exhaustive search use a biometric modality and algo-

rithms that generate score distributions with extremely

rapidly attenuating tails, when different persons are com-

pared. (These issues are discussed and documented in

more detail in the article Score Normalization Rules in

Iris Recognition.) In the absence of such rapidly attenu-

ating distribution tails, the system would drown in

False Matches when the search databases become

large. In this connection, it is noteworthy that in the

UK where the IRIS program optionally replaces pass-

port presentation, the Border Control Development

and Strategy Group forecasts that by 2015, the number

of international passengers entering the UK annually

will exceed 150 million.

Several other countries are also deploying the same

iris recognition algorithms as a substitute for passport

presentation. One of these is The Netherlands, where

iris-based border-crossing has been used since 2003 for

frequent travelers into Amsterdam Schiphol Airport;

members of the Privium program pay an annual fee

to be able to use automated iris gates for clearing

Immigration, in lieu of waiting in queues for pass port

presentation. Another country with a similar but

larger deploym ent is Canada, where the CANPASS

program operates in all the eight international airports

(Edmonton, Winnipeg, Calgary, Halifax, Ottawa,

Montreal, Toronto, and Vancouver) with about 40 kiosks

at each [3]. Both US and Canadian citizens or perma-

nent residents are entitled to enroll in this iris-based

system for entering Canada. In addition, the NEXUS

program operated jointly by the USA and Canada

allows border-crossing in both directions across their

shared border using iris recognition for preapproved

travelers (Fig. 2).

Finally, motorcyclists who commute daily across

the border between Malaysia and Singapore for work

use iris recognition to avoid the long queues for check-

ing passports and ID papers. The Singapore Iris Border

Iris Recognition at Airports and Border-Crossings. Figure 2 At Schiphol Airport (Amsterdam NL), the Privium Program

has a membership of about 40,000 frequent travelers. They pay an annual fee to use the iris recognition system at

automated gates, thereby avoiding the queues at Immigration for passport presentation.

Iris Recognition at Airports and Border-Crossings

821

Control for Motorcycles allows 3,000 commuters to cross

the border efﬁciently using ‘‘registered iris’’ lanes with

automated gates, as may be seen in an on-line video

[4]. The motorcyclists in these lanes remain on their

bikes; the gate is equipped from the side with iris

cameras, including one for a passenger on the bike.

Riders must stop and remove their helmets, but they

do not assert their identity. Rather, identiﬁcation is

performed by exhaustive search of the enrolled iris

database linked to the fully automated gates. The sys-

tem also maintai ns a watch-list that is checked.

Departing Passengers: Expedited Check-

in, Security Screening, and Border

Controls

The US Transportation Security Administration (TSA)

and Department of Homeland Security (DHS) in 2005

began a public/private partnership known as the

Registered Traveler (RT) Program to make airport se-

curity procedures more efﬁcient for departing passen-

gers deemed to be trusted. Under this program, dozens

of US airports have deployed iris and ﬁngerprint recog-

nition systems to conﬁrm the identity of ‘‘trusted tra-

velers’’ who have been vetted by the TSA and approved

for expedited security screening. Bypassing the long

lines that have become a feature of airport security

checkpoints since September 2001 is a beneﬁt for fre-

quent travelers, who pay an annual fee of about $100 for

this privilege. It is also an enhancement for TSA security

processes which can become more focused and can take

advantage of the background vetting that was done

when a person was enrolled in the scheme by virtue of

being deemed a minimum security risk.

Although baggage X-ray and metal detection

checks remain universal, enrollees in these systems

face less intrusive screening (e.g., they can keep their

coats and shoes on and laptops in their bags), and they

enjoy access to a reserved fastlane with shorter delays.

These privileges are asserted by presenting a smartcard

credential that contains their biometric data as well as

other information, all under two layers of encr yption

and readable only by TSA card-readers. Biometric

kiosks in the departure fastlanes read the cards and

conﬁrm passengers’ identity with iris cameras or ﬁnger-

print readers. The network is interoperable across some

30 US airports, and the list is steadily expanding [5].

Beginning with Orlando Airport in July 2005, some of

the major participating US airpor ts today include JFK,

LaGuardia, Newark, Dulles, Regan, Denver, and San

Francisco International Airports. The largest such pro-

gram is called CLEAR, operated by Veriﬁed Identity

Pass, which had 175,000 enrolled members as of July

2008 [5]. Additional newer participants in the

Registered Traveler public/private partnership with the

TSA include FLO, Unisys, and Vigilant.

In Europe, for travelers who are nationals of the

25 EU countries that have entered into the Schengen

Agreement for harmonized border control, the iden-

tiﬁcation formalities for crossing into and out of the

Schengen Zone are done by iris recognition at kiosks

in certain airports. The ﬁrst such deployment was

at Frankfurt/Main Airport and is known as the

Automate d and Biometrics-based Border Checks

(ABG) initiative. This multinational project is led by

Germany’s Federal Ministry of the Interior and Federal

Border Police. The stated objectives of the scheme are

to eliminate the use of fraudulent travel documents

and multiple identities, to speed trusted travelers

across borders, and to allow greater productivity for

border ofﬁcials.

Iris recognition is also used for other, nonsecurity

related enhancements for departing passengers at air-

ports such as Milan’s Malpensa and Tokyo’s Narita

Airport. Under the Simplifying Passenger Travel scheme

implemented by the Ministry of Justice in Japan, the

JAL Group offers streamlined procedures for passenger

check-in and boarding pass issuance, as well as immi-

gration control at departure and certain ‘‘e-airport’’

utilities and facilities. These services are provided at

iris-enabled automated kiosks and gates in departure

areas, as illustrated in Fig. 3.

Airport Employees: Access Control to

the Tarmac, Aircraft, and Restricted

Areas

Probably the most traditional use of biometric recog-

nition is for physical access control, to ensure that

only authorized persons enter restricted facilities.

This classical mode of biometric deployment is found

at many airports, controlling access to aircraft mainte-

nance facilities, baggage handling areas, the tarmac and

other secure zones.

The Canadian Air Transport Security Authority

uses iris recognition to verify the identities of airport

822

Iris Recognition at Airports and Border-Crossings

workers at all the 29 major airports in Canada. Iris

biometric data are embedded within an ID card called

RAIC: Restricted Area Identiﬁcation Card. Workers

must present this card and verify their identities at

iris cameras controlling automated portals. Similar

systems are deployed at Schiphol Airport (Amsterdam)

for 30,000 airport employees; at Albany Airport for

baggage handlers; and at New York JFK Airport

for access to the tarmac at two terminals. Some air-

ports such as Douglas International (Charlotte) have

also deployed iris recognition gates speciﬁcally for

pilots and other airline crew members to reach air-

side more efﬁciently.

Finally, it is noteworthy that an International Stan-

dard speciﬁcally related to biometric identiﬁcation of

airport employees was published in 2008. The ISO/IEC

24713-2 Standard gives normative requirements on

Biometric Proﬁles for Interoperability and Data Inter-

change: Physical Access Control for Employees at Airports

[6]. The scope of this Standard includes recommended

practices for enrollment, watch-list screening, preven-

tion of duplicate token issuance, and employee identity

veriﬁcation. It also describes architectures and business

processes appropriate to token-based identity manage-

ment within the secure environment of an airport.

Watch-list Screening of Arriving

Travelers

The rapid search capabilities of iris recognition, and its

robustness against making False Matches despite the

fact that large search databases create many opportu-

nities for such errors, have led to the deployment of

this technology for watch-list screening. The largest

such deployment is in the United Arab Emirates,

where visa-bearing travelers arriving at any of the

32 air, land, and sea ports of entry are processed with

iris recognition cameras as illustrated in Fig. 4.

Known as the Expellee Tracking and Border Security

Iris System, the scheme was launched in 2001 by the

UAE Ministry of Interior. A noteworthy aspect of the

UAE is that among its 5.4 million residents, about 85%

are foreign nationals [7] on work permits. Because of

this large foreign labor force drawn by economic

opportunities much better than elsewhere in the Mid-

dle East and Sou th Asia, men outnumber women by a

factor of 2.74 among persons in the 15–65 age group

[7], and the border-crossing volume of migrant work-

ers whose homeland roots are elsewhere is very high

(some 12,000 per day). In 2001 an amnesty was

granted to all foreign nationals who had overstayed

Iris Recognition at Airports and Border-Crossings. Figure 3 In the e-airport deployment at Tokyo Narita Airport, iris

recognition is used for expedited check-in of departing passengers. In dozens of US airports, Registered Travelers

approved by the Transportation Security Administration receive expedited security screening once their identities are

proved by ﬁngerprint or iris recognition.

Iris Recognition at Airports and Border-Crossings

823

their work permits or committed other visa violations,

but a condition of the amnesty waiver of penalties was

that such persons were expelled from the Emirates for

some period of time, and their iris patterns were

registered in a database. This action enabled enforce-

ment of the ban on re-entry and defeated thousands of

attempts to return under false identities and with fake

travel documents. Over the period 2001–2007 the da-

tabase of expellees’ IrisCodes was enlarged with Iris-

Code databases of foreign nationals who had been

imprisoned for crimes such as prostitution or drugs

trafﬁcking, and of persons deemed to be security risks

or unwelcome for other reasons.

Today this iris watch-list contains 1.2 million Iris-

Codes from persons of 156 nationalities. All travelers

seeking visa entry into the UAE via any port have their

iris images acquired by cameras as shown in Fig. 4,so

that their IrisCodes can be computed and matched

exhaustively against the full database. Since on average

some 12,000 such persons arrive at the UAE each day,

about 14 billion IrisCode comparisons are performed

daily across a dedicated network. The IrisFarm archi-

tecture is a distributed host/cl ient system with a single

central database maintained by the Abu Dhabi Police,

linked over a network of comm unication channels to

clients that send IrisCode queries to it from all ports

of entry. The average turn-around time is about 2s.

Because every query IrisCode is compared exhaustively

with all on the watch-list, the total volume of such iris

comparisons performed over the years of operation

now number in many trillions [8]. Tens of thousands

of persons have been caught trying to re-enter the UAE

under false identities, who are turned away but who

often make repeated attempts, and the UAE Ministry

of Interior hails the system as a huge success. The

system is now expanding into neighboring Gulf States

including Jordan and Oman, and it will be linked with

an iris-based national identity and border-crossing

system being procured in the Kingdom of Saudi

Arabia.

System Design Contrasts and

Vulnerabilities

The most important differences among the various

systems reviewed in this article are (1) whether they

operate in identiﬁcation mode, in which no identity is

asserted but identity is determined by searching a

database, versus veriﬁcation mode in which a token

like a smartcard is used to assert a particular identity

that is then simply veriﬁed one-to-one; and (2) whether

the objectives of a valid user or an attacker are to be

matched to an identity on a database, or not.

Identiﬁcation is vastly more demanding than one-

to-one veriﬁcation, both in terms of search space and

Iris Recognition at Airports and Border-Crossings. Figure 4 In the United Arab Emirates deployment of iris

recognition at all the 32 air, land, and sea ports, travelers are screened against a watch-list of expellees, or persons

deemed to be a security risk, before being allowed to enter the Emirates.

824

Iris Recognition at Airports and Border-Crossings

comparison speeds, and in terms of the requirement to

avoid any False Matches despite what may be a huge

number of opportunities to make them if the database

is large. If a weak biometric system such as face recog-

nition is used, an attac ker would have an excellent

chance to be matched just by chance against at least

one person in a trusted traveler database, if that were

the only test and if the database were larger than a few

hundred or perhaps a thousand. For this reason,

weaker biometrics rely on smartcards or other tokens

to assert a particular identity, so that only one com-

parison must be executed successfully. But presenta-

tion of a token makes the process more cumbersome,

and in any case it has no value for watch-list screening .

Nearly all deployments of iris recognition operate

in identiﬁcation mode by exhaustive search of a data-

base, because the technology’s speed and accuracy

allow it. The exceptions to this mode are (1) the Pri-

vium system because Dutch law forbids the storage by

the State of personal data like biometrics, and so the

citizens alone retain it; and (2) the CLEAR program

because a smartcard is used for several other purposes

in the transaction anyhow. In both of these cases the

use of a storage token makes it unnecessary to perform

identiﬁcation by searching a database.

In identiﬁcation system s operating by database

search, it is necessary to combat the inevitable net

increase in the likelihood of chance False Matches as

the size of the search databases grow. This form of

probability summation is the same phenomenon as

arises when playing the game of Russian Roulette an

increasing number of times. In the case of the iris

recognition algorithms [1, 8] used in all current iris

deployments, combatting this is accomplished by min-

ute adjustments in the decision threshold with search

database growth, keeping the net False Match proba-

bility minuscule. Further details about these processes

are given in the accompanying article, Score Normali-

zation Rules in Iris Recognition.

In a trusted traveler scheme (CLEAR, IRIS, Pri-

vium, etc.), the objective of an attacker is to imperson-

ate another person – either a particular person, or

anyone at random just by accident – who is registered

in the trusted database. The likelihood of success by

blind chance (a ‘‘zero effort attack’’) is minuscule in

the case of iris, but much higher if a printed contact

lens can be produced to mim ic a particular target

individual’s iris. In a watch-list deployment such as

the UAE one, the objective of an attacker is simply to

look like anybody other than himself (or anyone else

registered in the watch-list). Such a ‘‘concealment at-

tack’’ by means of printed contact lenses is easier than

an ‘‘impersonation attack,’’ and indeed it can even be

attempted simply by being uncooperative. Therefore,

these border securit y system s incorporate tests for the

vitality, or ‘‘liveness,’’ of iris patterns including their

motion and deformation with changes in the pupil

size, which obviously does not occur if printed on a

contact lens. Similarly, the standard algorithms per-

form biometric quality assessments to detect extremely

dilated pupils or excessively closed eyelids, as indica-

tors of possible attacks. However, the struggle between

countermeasure and new counter-countermeasure

continues and escalates relentlessly.

Related Entries

▶ Iris Encoding and Recognition Using Gabor

Wavelets

▶ Score Normalization Rules in Iris Recognition

References

1. Daugman, J.G.: How iris recognition works. IEEE Trans. Circuits

Syst. Video Technol. 14, 21–30 (2004)

2. UK Border Agency, Project IRIS website. http://www.ukba.

homeofﬁce.gov.uk/managingborders/technology/iris/

3. Canadian Border Services Agency, CANPASS website. http://

cbsa-asfc.gc.ca/publications/pub/bsf5017-eng.html

4. Singapore Iris-Based Border Control for Motorcyclists. http://

www.youtube.com/watch?v=HieaASl9sE8&feature=related

5. CLEAR Veriﬁed Identity Pass website. http://www.ﬂyclear.com

6. International Organisation for Standards: Biometric Proﬁles for

Interoperability and Data Interchange, Part 2: Physical Access

Control for Employees at Airports. ISO/IEC 24713-2 (2008)

7. Demographics of UAE, 2008. http://en.wikipedia.org/wiki/

United_Arab_Emirates

8. Daugman, J.G.: Probing the uniqueness and randomness of

IrisCodes: results from 200 billion iris pair comparisons. Proc.

IEEE 94, 1927–1935 (2006)

Iris Recognition Immigration System

(IRIS)

▶ Iris Recognition at Airports and Border-Crossings

Iris Recognition Immigration System (IRIS)

825

Iris Recognition Operational Range

The maximal transversal region along the optical axis

of an imaging system within which an iris recognition

algorithm provides accurate identiﬁcation. Typically,

iris recognition operational range is signiﬁcantly af-

fected by the system’s depth of ﬁeld.

▶ Wavefront Coded

Iris Biometric Systems

Iris Recognition Performance

Under Extreme Image Compression

JOHN DAUGMAN,CATHRYN DOW NI N G

Computer Laboratory, Cambridge University,

Cambridge, UK

Definition

The compressibility of images is usually gauged by their

subjective appearance and by metrics for the amount

of distortion that can be tolerated. In the context of

biometrics, compressibility can be gauged objectively

by measuring the impact of compression schemes on

recognition performance compared to baseline perfor-

mance. Standard biometric methodologies such as

Receiver Operating Characte ristic (ROC) curves are

perfectly suited for measuring the impact of compres-

sion on performance. It is possible for performance

actually to beneﬁt from slight image compression, as

has been seen both with ﬁngerprint and iris recogni-

tion, because high frequency noise is the ﬁrst thing lost;

but at more severe levels, compression must become

detrimental. For iris recognition, it is possible to com-

press images to as little as 2,000 bytes through a combi-

nation of methods including cropping, region-of-interest

(ROI) isolation and JPEG2000 wavelet coding, while

suffering only a little reduction in recognition perfor-

mance. This is important because Governments and

Standards organizations prefer that biometric data be

stored in a relatively raw, unprocessed form in order to

remain algorithm-neutral and future-proof. It is also

mathematically important because of insights from

information theory and complexity theory related to

minimal description length, entropy, compressibility,

and discriminability.

Introduction

A watershed event in biometric informatics occurred in

1993 when the FBI digitized a vast library of ﬁngerprint

cards that had been stored in acres of ﬁling cabinets

and adopted the Wavelet Scalar Quantization (WSQ)

protocol [1] to compress these images, achieving com-

pression ratios in the range of 10:1 or 15:1 without

detectable loss of detail. Compressibility is a fundamen-

tal issue for biometrics, not only for mundane practical

reasons related to storage requirements and data trans-

mission times, but also for abstract mathematical rea-

sons related to information content and pattern

recognition. Governments, regulatory bodies, and inter-

national standards organizations often mandate the

storage of relatively raw data rather than processed

biometric templates, hoping thereby to preserve inter-

operability and to keep biometric data ‘‘vendor-neutral

and future-proof’’ while the algorithms for pattern de-

scription and recognition inevitably evolve and im-

prove. But raw images as data objects are almost a

thousand times larger than the biometric templates

ultimately computed from them. Hence, the conﬂicting

goals of reducing the size yet preserving the information

content of biometric data make it important to under-

stand how various possible compression schemes im-

pact on biometric recognition performance. This article

studies these questions in the context of iris recognition,

and reviews data [2] showing that it is possible to

compress iris images to within about a factor of two

of the standard iris biometric template sizes with al-

most no impact on recognition performance.

Information Theory and Data

Compressibility

Data compression is one of several disciplines rooted

in information theory having relevance to biometric

technologies for identifying persons, and its signi-

ﬁcance extends beyond the practical matters of storage

requirements and data transmission times. One

of Shannon’s fundamental insights in formulating

826

Iris Recognition Operational Range

information theory [3] was that the entropy of a ran-

dom variable measures simultaneously its information

content (expressed in bits) and its compressibility with-

out loss (to the same number of bits). This link between

entropy, informativeness, and compressibility extends

also to other measures that apply to biometrics. For

example, the relative entropy between two distributions

is one way to measure how well a biometric technique

separates samples from same versus different persons.

The amount of variability in a given biometric system

across a population, or in different samples from the

same source, is also captured by conditional entropies,

with larger entropy signifying greater randomness. Fi-

nally, the similarity between pairs of biometric tem-

plates is reﬂected by their mutual information: the

extent to which knowledge of one sample predicts the

other. All of these properties are deeply connected with

the compressibility of biometric data.

An extreme variant of Shannon’s insight was

expressed by Kolmogorov [4] in his notion of minimal

description length, which deﬁned the complexity of a

string of data as the length of the shortest binary

program that could generate the data. Creating that

program ‘‘compresses’’ the data; executing that pro-

gram ‘‘decompresses’’ (generates) the data. Fractal

image compression is based on this idea; and a data

string is said to be Kolmogorov incompressible if the

shortest program that can generate it is essentially a

data statement containing it, so the data are then their

own shortest possible description. Within biometrics,

this notion has appeared implicitly under a different

rubric in work on syntheti c biometrics, seeking method s

for artiﬁcially synthesizing a biometric image that is

indistinguishable in practice from some actual biomet-

ric image. Pioneering work in this direction was done

by Terzopoulos and Waters [5] for facial images and

sequences, by Cappelli et al. [6] for ﬁngerprints, and by

Cui et al. [7] and by Zuo et al. [8] for iris images. In

future, such programs for generating particular bio-

metric images might therefore serve as ways to ‘‘com-

press’’ them in Kolmogorov’s sense; and one might

even anticipate biometric recognition by comparison

of the synthesizing programs. The present article

explores a combination of image compression meth-

ods applied to iris images, speciﬁcally probing the

question of how aggressively they can be compressed

without impairing iris recognition. A convergence be-

tween compressed image size and biometric descrip-

tion length begins to emerge.

Schemes for Iris Image Compression

Iris templates (e.g., ▶ IrisCodes) are usually computed

from a polar or pseudo-polar coordinate mapping of

the iris, after locating its inner and outer boundaries

[9, 10, 11]. However, if simple circular boundary mod-

els are imposed, polar mappings depend strongly upon

the choice of origin of coordinates which may be prone

to error, uncertainty, or inconsistency, especially since

the true iris boundaries are often not actually circular.

Unlike rectilinear coordinates, for which a shift error

has no more effect than a shift, in polar coordinate

mappings a shift error in the choice of coordinate

origin can cause large distortions in the mapped data,

with no way to recover from the deformed sampling.

A pioneering study of iris compressibility was

undertaken by Rakshit and Monro [11] showing that

if segmented and normalized iris data were extracted in

polar form, this ‘‘unwrapped’’ polar data structure

could be compressed to 2,560 bytes or even less with-

out impairing recognition performance. However, be-

cause this approach stores the iris image in polar form,

it is not robust against errors in assigning the origin of

coordinates or the loss of iris data when circular bor-

ders are inappropriately enforced. It also suffers from

making the accurate detection of eyelid boundaries for

exclusion of eyelid regions very difﬁcult, since the

available pixel data is tightly cropped around the iris

and ignores valuable shape-constraining data from the

boundary betw een sclera and eyelid. Finally, the polar

unwrapping has the consequence of highly nonuni-

form sampling, with pixels near the outer perimeter

of the iris sampled with a density 2.5 times lower than

those near the pupil. Preservation of the original recti-

linear format of an iris image is more veridical than

polar unwrapping methods because pixels retain con-

stant size and spac ing.

Avoiding the problems introduced by polar unwrap-

pings of iris images, the authors [2] investigated three

compression schemes that retain the native rectilinear

image array format but compress it to as little as

2,000 bytes while still allowing very good recognition

performance on the difﬁcult Iris Challenge Evaluation

(ICE-1) iris database available from the National Insti-

tute of Standards and Technology (NIST) [12] which

includes many poor quality images. Interoperability

was documented between those images when subjected

to the three compression regimes and their uncom-

pressed form, and it was found that on average only

Iris Recognition Performance Under Extreme Image Compression

827

2–3% of the bits within the computed IrisCodes are

affected even when the net image data reduction factor

reaches 150:1.

Tools for Iris Image Compression:

JPEG, JPEG2000, Region-of-Interest

Extraction

Clearly, a ﬁrst step in image data reduction is to crop

the iris images from the stand ard format of 640  480

pixels with 8 bits grayscale data per pixel, consuming

307,200 bytes, to a smaller region of 320  320 pixels

centred on the iris. This was done by running the eye-

ﬁnding part of the standard algorithms [10] that are

used in all current public deployments of iris recogni-

tion, on all images in the publicly available NIST [ 12]

ICE1Exp1 datab ase, which contains 1,425 iris images

from 124 Subjects with ‘‘ground-truth’’ information

given about which images were taken from the sam e

iris. The algorithms correctly localized the iris in all

images and produced from each one a new cropped

image of 320  320 pixels with the iris centred in it. For

those NIST images in which the iris was partly outside

of the original image frame, the missing pixels were

automatically replaced with black ones. For those in

which the algorithms detected that the gaze was direct-

ed away from the camera, as gauged by projective

deformation of the eye shape, a corrective afﬁne trans-

formation was automat ically applied which effectively

‘‘rotated’’ the eye in its socket into orthographic per-

spective on-axis with the camera. The ﬁrst column of

Fig.1 shows three examples of iris images cropped as

described earlier.

This new gallery of simply cropped images was

subjected to three different compression schemes: (1)

▶ JPEG compression with quality factors (QF) of 70,

30, and 20; (2) JPEG compression with the same QFs

but after

▶ Region-of-Interest (ROI) segmentation;

and (3)

▶ JPEG2000 compression after ROI segmenta-

tion with compression factors (CF) of 20, 50, and 60, as

illustrated in the second column of Fig.1 for the case of

CF¼50.

The use of cropping and JPEG compression alone

(with QFs of 70, 30, and 20) produced image ﬁle sizes

averaging 12,400, 5,700, and 4,200 bytes, respectively,

but with large variability around these means. Includ-

ing the initial threefold reduction in ﬁle size due merely

to cropping the imag es to 320  320 pixels, these net

data reduction factors relative to the original full-size

images therefore average 25:1, 54:1, and 72:1, respec-

tively. But fur ther signiﬁcant reductions in image data

size can be achieved through the use of ROI segmenta-

tion of the iris image.

The standard lossy JPEG coding scheme [13, 14]

effectively allocates bytes on an ‘‘as needed’’ basis,

meaning that the cost of encoding uniform regions of

an image is almost nil, whereas image areas containing

busy textures such as eyelashes may consume much of

the available informati on budget. In uniform regions,

the only nonzero DCT (discrete cosine transform)

coefﬁcient in each block of 64 frequency components

that encode an 8  8 pixel block (a data unit) is the DC

coefﬁcient specifying their average gray value; all other

coefﬁcients are 0 if the data unit is a truly uniform

region, or else become 0 after coarse quantization, and

so their cost in the zeroes run-length coding stage is

essentially nil. Therefore JPEG encoding of iris image s

can be made much more efﬁcient if all noniris parts of

the image are replaced with a uniform gray value. This

was accomplished for the image gallery automat ically

using the standard algorithms [10] for eyelids detec-

tion and ﬁtting, and iris boundary localization, as seen

in the second column of Fig.1.

JPEG coding schemes lend themselves well to ROI

differential assignment of the coding budget. Indeed

the JPEG2000 standard [15, 16, 17] and even the Part 3

extension of the old JPEG standard [13, 14], support

variable quantization for explicitly specifying different

quality levels for different image regions. In JPEG2000,

the MAXSHIFT tool allows speciﬁcation of an ROI of

arbitrary shape. This utility was explored for biometric

face recognition by Hsu and Grifﬁn [18] who demon-

strated that recognition performance was degraded by

no more than 2% for ﬁle sizes compressed to the range

of 10,000–20,000 bytes with ROI control.

In the approach to ROI segmentation presented

here, noniris regions are encoded in a way that distin-

guishes sclera from eyelids or eyelashes regions, so that

postcompression algorithms can still determine both

types of iris boundaries. Therefore two different sub-

stitution gray levels are used: a darker one signifying

eyelids, and a brighter one for the sclera, computed as

an average of actual sclera pixels and blending into

actual sclera pixels near the iris outer boundary. Since

the substitution gray levels are uniform, their coding

cost is minimal and could be further reduced by using

larger data units. JPEG compression of such ROI seg-

mented iris images typically yields a further twofold

reduction in ﬁle size for each of the QFs studied, while

828

Iris Recognition Performance Under Extreme Image Compression

maintaining a simple rectilinear image format and easy

localization of eyelid boundaries in later stages.

In 2000 a more powerful version of JPEG coding

offering more ﬂexible modes of use, and achieving

typically a further 20–30% compression at any given

image quality, was adopted as the JPEG2000 Stan-

dard [16, 17]. Mathematically based on a Discrete

Wavelet Transform (DWT) onto Daubechies wavelets

rather than the Discrete Cosine Transform (DCT),

JPEG2000 does not suffer as badly from the block

quantization artifacts that bedevil JPEG at low bit-

rates, which are due to the fact that the DCT simply

chops cosine waves inside box windows with obvious

truncation consequences when they are sparse and

incomplete. Moreover, the different levels within the

multiresolution DWT wavelet decomposition allow

local areas with in each image data unit to be encoded

using different subbands of coefﬁcients [17] as need ed.

The net superiori ty of JPEG2000 over JPEG in terms of

image quality is especially pronounced at very low bit-

rates corresponding to severe compression, as investi-

gated here, in the range of 0.15 bits/pixel (bpp).

Several mechanisms exist within JPEG2000 for het-

erogeneous allocation of the coding budg et, including

tile (data unit) deﬁnition, code-block selection allow-

ing different DWT resolution levels in different tiles,

Iris Recognition Performance Under Extreme Image Compression. Figure 1 Cropping of iris images (first column),

followed by region-of-interest isolation of the iris (second column) to achieve greater compressibility while retaining a

rectilinear image array format. Substitution of noniris regions by uniform gray levels prevents wasting wavelet coding

budgets on costly irrelevant structures such as eyelashes. All images in the second column have been JPEG2000

compressed to a data size of only 2,000 bytes.

Iris Recognition Performance Under Extreme Image Compression

829