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

Подождите немного. Документ загружается.

Matrix X contains {x

} in its columns, and vector

u contains the peak constraints (1 for authentics,

0 for impostors).

The OTSDF ﬁlter accomplishes two goals: it pro-

duces sharp peaks for training images, by reducing

ACE, and it achieves tolerance to additive white

noise, by reducing ONV in Eq. (1). The resulting ﬁlter

gives good discrimination between authentic and

impostor patterns, even when the authentic patterns

are noisy.

Occlusion Metric

We need to have a metric for each local patch to

evaluate the probability of this local patch belonging

to an occluded region or a true iris pattern. We call

this metric as ‘‘occlusion metric’’. For every local pixel

x in a normalized iris map, we denote p(x) as the

occlusion metric.

There are some intuitive observations for estimat-

ing the occlusion metric for a particular patch.

For example, if we observe the second row of Fig. 2

(the normalized iris map), we will see that in most

cases, occlusion caused by eyelid happens in three

regions: upper-left corner, upper-center region and

the upper-right corner. Therefore, if the location of a

point is close to these three regions, ﬁrst, it is more

likely to be an occluded region. Second, an occluded

region caused by eyelids usually has higher value of

pixel intensity. This is because in most cases eyelids

look brighter than iris texture. Third, if we observe

the occluded region in Fig. 3a and b, we would ﬁnd

that usually the eyelid region contains textures of

irregular eyelashes. The line style of eyelashes in Carte-

sian coordinate is irregular. Some are straight, others

are curved. After unwrapping them into polar coordi-

nate, the shape of eyelashes become even more irregu-

lar. Therefore, they create irregular patterns on top of

the eyelids. One special thing about eyelashes is their

intensity is usually lower than iris texture. Therefore,

eyelashes on top of eyelids lead to a large standard

deviation of pixel intensity value in local region. This,

too, can be used as one feature to distinguish occlusion

from iris texture.

Based on the three observations mentioned above,

we compute four features for every pixel on a normal-

ized iris map. They are: (1) the mean intensity value

on a local patch, centered around that pixel; (2) the

standard deviation of the intensity value in that local

patch; (3) the percentage of pixels whose intensity is

greater than one standard deviation above the mean of

the entire iris plane; (4) the shortest Euclidean distance

from that pixel to one of the three ‘‘easily-occluded-

region’’. After computing these features, we train a

Fisher linear discriminant analysis (FLDA) model to

compute a single value to represent the probability of

this pixel being occluded. This FLDA model is trained

by using 30 iris texture maps whose occluded regions

are manually labeled.

Probabilistic Graphical Model

We have described the proposed probabilistic graphical

model in previous section. Now, let us think about how

to model the potential function of the nodes and

between those nodes in Fig. 3c. As described earlier,

for each patch, we have variables d

and o

, to model

the local shift and the occlusion, respectively. We can

combine the both into one variable, say h

¼ d

; o

½

and h

is a vector in three-dimensional space. For any

pair of neighboring vectors h

and h

, we can model

the possibility of observing them C

i;j

ðh

; h

Þ with the

following equation:

i;j

ðh

; h

Þ¼C

d;i;j

ðd

; d

ÞC

o;i;j

ðo

; o

Þð4Þ

The reason that we can decompose C

i;j

ðh

; h

Þ is

that we assume the local shift of the patch and

occlusion are statistically independent events. We can

further model C

d;i;j

ðd

; d

Þ and C

o;i;j

ðo

; o

Þ with the

following equation:

d;i;j

ðd

; d

Þ¼e



a d

kkþa d

þb d

d

ðÞ

ð5Þ

o;i;j

ðo

; o

Þ¼

; o

¼ o

¼ 0

; o

¼ o

¼ 1

; o

6¼ o

ð6Þ

Basically, Eq. (5) is assuming that the length of local

shift vector should be small, the larger the length of

the shift vector is, the smaller probability that it has.

Besides this, the neighboring patches should have

similar (if not exactly the same) local shift vector, so

the probability is inversely proportional to the differ-

ence between the two local shift vectors. Equation (6)

is to estimate the joint probability of the neighboring

patches to be both occluded, both un-occluded, or one

is occluded but not the other. These parameters can be

trained by using a small portion of any iris database.

840

Iris Recognition Using Correlation Filters

The next quantity we have to model is the potential

function C

ðh

; O

Þ which is the probability of observ-

ing O

when status of the node is h

. Obviously,

ðh

; O

Þ must depend on o

because if a region is

occluded then deﬁnitely it would affect the observation

. Let us deﬁne the distribution F(s) as the probabilit y

of having observed at least the true match score s, and

F(p) as the probability of having observed at least

the true occlusion metrics p. Then C

ðh

; O

Þ can be

expressed as:

ðh

; O

Þ¼

ðmðd

ÞÞ; o

¼ 0

ðp

Þ; o

¼ 1



ð7Þ

Where mðd

Þdenotes the match score when the given

shift vector is d

. Since F(s) is the probability of having

observed at least the true match score s, it can

be modeled as the cumulative distribution function

(cdfs) of s and same thing applies to F(p). Therefore,

F(s) and F(p) can be expressed as:

ðSÞ¼Pðs < SÞ¼

1

Nðs; m

; s

Þds ð8Þ

ðPÞ¼Pðp < PÞ¼

1

Nðp; m

; s

Þdp ð9Þ

Where s represents the match score observed at the

unknown true shifts for given un-occluded iris region,

and p represents the occlusion metrics observed at the

unknown true shift for given occluded iris region.

Usually the distribution P(s) and P(p) are assumed to

be normally distributed, as in Eqs. (8) and (9).

Final Score Computation

The equations we mentioned in last section are all

about estimating probability distri bution for each

sub-region (patch). In order to generate a single

value score for comparing two different irises, we

have to combine scores from all sub-regions. We deﬁne

the ﬁnal Score M as:

M ¼

i¼1

ð10Þ

Where M

is the score from each sub-region, and b

the weighting coefﬁcient deﬁned as the probability of

sub-region i to be un-occluded, given the observation:

Pðo

¼ 0jOÞð11Þ

The score of a single sub-region M

is also com-

puted based on a probabilistic point of view. Since

we have estimated the joint probability distribution

of the model parameter h

and the observation O

we should be able to compute the posterior probability

distribution

Pðd

¼ djOÞ.SoM

can be computed as

follows:

ðdÞ

Pðd

¼ djOÞð12Þ

Speciﬁcally speaking, the posterior probability

Pðd

¼ djOÞ can be estimated from

Pðd

jOÞ,as

shown in the following:

Pðd

jOÞ¼

Pðh

jOÞð13Þ

If we view Fig. 3(c) as a Markov Random Field [13],

with the potential function speciﬁed in previous sec-

tion, we can use Loopy Belief Propagation [14]to

estimate the conditional probability

Pðh

jOÞ. Assume

j!k

ðh

Þ is the message passing from node j to neigh-

boring node k at the ith iteratio n. Then we can com-

pute as shown below:

j!k

ðh

Þ¼

ðh

; O

ÞC

j;k

ðh

; h

 P

l2NðjÞk

i1

l!j

ðh

ð14Þ

Where C

ðh

; O

Þ and C

j;k

ðh

; h

Þ are deﬁned in Eqs.

(7) and (6), respectively, and N(j) denotes the set of all

neighboring nodes for j in the graphical model.

After i iterations, the conditional probability

Pðh

jOÞ can be computed as follows:

Pðh

jOÞ¼

ðh

; O

Þ P

k2NðjÞ

k!j

ðh

Þð15Þ

where the normalizing coefﬁcient z

is given by

ðh

; O

Þ P

k2NðjÞ

k!j

ðh

Þð16Þ

Result

The framework of iris deformation and estimation has

been tested on ICE database [16]. For ICE database,

there are two sub-sets. The ﬁrst one, ICE Experiment 1

contains 1425 images of right irises from 124 users,

Iris Recognition Using Correlation Filters

841

resulting in 12214 authentic comparison and 1002386

imposter comparisons. ICE Experiment 2 contains

1528 images of left irises from 120 users resulting in

14653 authentic comparison and 1151975 imposter

comparisons.

The proposed algorithm is compared against a

baseline algorithm, which does not contain the proba-

bilistic framework for iris deformation. This model

assumes that there is only a possible global shift in

iris image. And the match score between two irises

can be computed with the following equation:

mðxÞ¼

y2s

ðyÞc

ðy  xÞð17Þ

where x is the global shift vector, c

ðyÞ and c

ðyÞ

denotes the unshifted template and query iris codes,

respectively, s

is the support of the template iris code

and s

is the total number of element in template.

This method is similar to the method that Daugman

proposed.

For each experiment, we computed the False Reject

Rate (FRR) at different levels of False Accept Rate (FAR).

Speciﬁcally, FRR is measured at FAR equals to 1%, 0.1%

0.01%, and 0.001%. The results are shown in Table 1.

From Table 1, we can see that the proposed frame-

work of iris deformation and estimation has a much

better performance than the baseline, which has the

same power as the Daugman’s algorithm. It shows that

proposed framework is indeed effective and very usef ul

in iris recognition problems.

Summary

Iris pattern deformations and occlusions are two most

prominent problems in the ﬁeld of iris recognition. In

the past, some methods have been proposed for the

problem of occlusion detection, however, few have

addressed the problem of local deformation of iris tex-

tures. In this article, we introduce one novel technique

which combines the most advanced theories in machine

learning and pattern recognition to solve this problem.

By the framework of deformation and occlusion esti-

mation, two iris images (from the same class) can be

matched successfully even when one or both of them

suffers the in-plane, local deformation. Experimental

results have shown the power of the novel framework.

From the point of view of computational efﬁciency,

the proposed framework only requires a few iterations

of probability estimation, which consists of operations

of simple addition and multiplication for a few vari-

ables. It does not take much longer time than

the conventional iris recognition algorithm, which

makes it suitable for iris recognition system running

in real-time.

Related Entries

▶ Identiﬁcation and Authentication

▶ Iris Encoding and Recognition

▶ Iris Recognition, Overview

References

1. Thornton, J.: Matching Deformed and Occluded Iris Patterns:

A Probabilistic Model Based on Discriminative Cues. Disserta-

tion, Dept. of Electrical and Computer Engineering, Carnegie

Mellon University (2007)

2. Thornton, J., Savvides, M., Vijaya Kumar, B.V.K.: Enhanced iris

matching using estimation of in-plane nonlinear deformations.

In: Proceedings of SPIE Defense and Security Symposium, Vol.

6202, 62020E (2006)

3. Kerekes, R., Narayanaswamy, B., Thornton, J., Savvides, M.,

Vijaya Kumar, B.V.K.: Graphical Model Approach to Iris Match-

ing Under Deformation and Occlusion. In: IEEE Conference on

Computer Vision and Pattern Recognition, 2007 (CVPR ‘07).

vol. June 2007, pp. 1–6, 17–22

Iris Recognition Using Correlation Filters. Table 1 False

reject rates (FRRs) of the proposed algorithm and

baseline algorithm in two iris database

Methods

CASIA

FAR = 1% 0.1% 0.01% 0.001%

Baseline 1.0% 2.3% 4.8% 9.6%

With deformation

model

0% 0.1% 0.7% 1.9%

Methods ICE Experiment 1

FAR = 1% 0.1% 0.01% 0.001%

Baseline 2.35% 3.45% 4.82% 6.26%

With deformation

model

0.17% 0.33% 0.82% 1.56%

Methods ICE Experiment 2

FAR = 1% 0.1% 0.01% 0.001%

Baseline 2.16% 3.02% 3.92% 5.25%

With deformation

model

0.64% 0.94% 1.26% 1.91%

842

Iris Recognition Using Correlation Filters

4. Vijaya Kumar, B.V.K., Savvides, M., Venkataramani, K., Xie, C.:

Spatial frequency domain image processing for biometric recog-

nition. In: Proceedings of International Conference on Image

Processing, September 2002, pp. 53–56

5. Savvides, M., Vijaya Kumar, B.V.K.: Efﬁcient design of advanced

correlation ﬁlters for robust distortion-tolerant face recognition.

In: Proceedings of IEEE Conference on Advanced Video and

Signal Based Surveillance, July 2003, pp. 45–52

6. Thornton, J., Savvides, M., Vijaya Kumar, B.V.K.: Robust iris

recognition using advanced correlation techniques. In: Proceed-

ings of International Conference on Image Analysis and Recog-

nition, September 2005, pp. 1098–1105

7. Vijaya Kumar, B.V.K., Mahalanobis A., Juday, R.D.: Correlation

Pattern Recognition. Cambridge University Press, Cambridge,

UK (2005)

8. Vijaya Kumar, B.V.K.: Tutorial survey of composite ﬁlter designs

for optical correlators. Appl. Opt. 31, 4773–4801 (1992)

9. Vijaya Kumar, B.V.K., Mahalanobis, A., Takessian, A.: Optimal

tradeoff circular harmonic function corr1elation ﬁlter methods

providing controlled in-plane rotation response. IEEE Trans.

Image Process. 9(6), 1025–1034 (2000)

10. Kerekes, R., Vijaya Kumar, B.V.K.: Correlation ﬁlters with con-

trolled scale response. IEEE Trans. Image Process. 15(7),

1794–1802 (2006)

11. Mahalanobis, A., Vijaya Kumar, B.V.K., Casasent, D.: Minimum

average correlation energy ﬁlters. Appl. Opt. 26, 3630–3633

(1987)

12. Vijaya Kumar, B.V.K., Carlson, D.W., Mahalanobis, A.: Optimal

trade-off synthetic discriminant function ﬁlters for arbitrary

devices. Opt. Lett. 19, 1556–1558 (1994)

13. Perez, P.: Markov random ﬁelds and images. CWI Quarterly

11(4), 413–437 (1998)

14. Felzenszwalb, P., Huttenlocher, D.: Markov random ﬁelds and

images. Int. J. Comput. Vis. 70(1), 41–54 (2006)

15. ICE iris database. National Institute of Standards and Technolo-

gy. http://iris.nist.gov/ICE/, 2005. Accessed 1 March 27, 2009

16. Daugman, J.: How Iris Recognition Works. IEEE Trans. Circ.

Syst. video Tech. 14(1), 21–30 (2004)

Iris Recognition with Deformation

and Occlusion Estimation

▶ Iris Recognition Using Correlation Filters

Iris Retina Biometric Fusion

▶ Simultaneous Capture of Iris and Retina for

Recognition

Iris Sample Synthesis

NATALIA A. S CHMID

Department of Computer Science and Electrical

Engineering, West Virginia University, Morgantown,

WV, USA

Synonym

Synthetic iris images

Definition

Iris image synthesis is a process of creating images of

an iris by means of statistical and stochastic models,

computer graphic tool s or through manipulating,

modifying or transforming parts or complete images

collected from real irises. Since the size of synthesized

datasets can be made arbitrary large, these data are

suggested to be used for the purpose of extensive

testing of the performance and efﬁciency of newly

designed iris recognition algorithms.

Introduction

Iris as a biometric has been known for a long time.

However, only in the recent years it has gained a sub-

stantial attention of both the research community and

governmental organizations resulting in the develop-

ment of a large number of new iris encoding and proces-

sing algorithms. Most of the designed systems and

algorithms are claimed to have exclusively high recogni-

tion performance. However, since there are no publicly

available large-scale and even medium-size datasets, only

very few newly designed algorithms have undergone

extensive testing. There are several datasets of frontal

view iris images presently available for public use. A

brief summary of these databases is provided in Table 1.

With the lack of data, two major solutions to the

problem of algorithm testing are possible: (1) physically

collecting a large number of iris images or (2) syntheti-

cally generating a

▶ large scale dataset of iris images.

This article addresses the second approach. The

following sections summarize a number of techniques

to synthesize iris images and characterize methods to

evaluate the performance of generated images.

Iris Sample Synthesis

843

The purpose of a synthesized dataset is to provide an

option to compare efﬁciency, limitations, and capabil-

ities of newly designed iris recognition algorithms

through their testing on a large scale dataset of generated

irises. Although adding synthetic data to the test set, or

adding artiﬁcial noise to data for a scenario testing may

introduce a

▶ performance bias (see recommendations

by Mansﬁeld and Wayman [5]), when used with cau-

tion, a large synthesized dataset may evaluate efﬁciency

and point to drawbacks of testing algorithms.

Available Models

The ﬁrst methodology for generating synthetic iris

images has been proposed by Cui et al. [1], where a

sequence of small patches from a set of iris images

was collected and encoded by applying Principal

Component Analysis (PCA) method. Principal compo-

nents were further used to generate a number of low

resolution iris images from the same iris class. The low

resolution images were combined in a single high reso-

lution iris image using a superresolution method. A

small set of random parameters was used for the gener-

ation of images belonging to different iris classes.

Another method for the generation of synthetic iris

images based on the application of Markov Random

Field (MRF) has been recently developed by Makthal,

Shah, and Ross [4, 6]. The generation technique relies

on texture primitives to synthesize new iris images.

Texture primitives are small portions of iris texture

selected from a host iris image. The generation tech-

nique is deterministic in the sense that it regenerates

texture primitives locally, but positions them at loca-

tions speciﬁed by a frequency of their occurrence

map. The map speciﬁes the probability of occurrence

of a texture primitive at various locations in a generated

image. Thus generated image globally exhibits a differ-

ent structure compared to host images. A few images

generated using this approach are shown in Fig. 1.

Lefohn et al. [3] developed an ocularist’s approach

using the computer vision technolog y for the purpose

of both the ocular prosthetics and entertainment

industries. In their work, a set of textured layers was

used to render each iris.

Wecker et al. [7] combined characteristics of real

irises to augment existing real iris databases. In their

work a multiresolution technique known as reverse sub-

division was used to capture the necessary characteristics.

Zuo et al. [8, 2] took a model based, anatomy based

approach for the generation of iris images. The work

makes a number of observations on common visual

characteristics of irises such as (1) radial ﬁbers, radially

arranged iris vessels, constitute the basis for iris tissues

and dominate the structure information; (2) a large

part of iris is covered by a semitransparent layer with a

bumpy look and few furrows which are caused by

retractor muscles; (3) the irregular edge of the top

layer contributes to the iris pattern; (4) the collarette

part is raised due to the overlap of sphincter and

dilator muscles. Thus, the main frame of the iris pattern

in this work is formed by the radial ﬁbers, the raised

collarette, and the partially covered semitransparent

layer with the irregular edge. At the same time, the

difference of pixel values in an infrared iris image is

not only the result of the iris structure information.

It is also related to the type of muscles, vessels,

and cells that the iris is composed of, surface color,

and lighting conditions. Involving all those visual and

anatomical characteristics makes each synthetic iris look

similar to a real iris.

To simulate the stochastic nature of individual iris

patterns, the process of iris image generation is reduced

to the generation of a dense ﬁber structure controlled

by a set of random parameters and postprocessed using

a variety of image processing techniques. The inﬂu-

ence of controlled parameters is carefully researched

and the operational range of parameters is evaluated.

Iris Sample Synthesis. Table 1 Public iris databases

Database

name

Database

size

#of

classes

#of

Images per

class

Color or

gray

scale

CASIA-I 756 108 7 gray

CASIA-III-

device1

1,200 60 20 Gray

CASIA-III-

device2

1,200 60 20 Gray

CASIA-III-

Interval

2,655 396 NA Gray

CASIA-III-

Lamp

16,213 819 NA Gray

CASIA-III-

Twins

3,183 400 NA Gray

WVU-off-

angle

560 140 4 Gray

WVU 2453 359 NA Gray

UBIRIS 1,877 241 NA Color

UPOL 384 128 3 Color

BATH 1,000 50 20 Gray

844

Iris Sample Synthesis

The value of each parameter is limited to a certain

range to keep the common iris features. However, with-

in the range it is allowed to vary maximally to increase

the randomness of the iris pattern. A four step proce-

dure for the generation of iris images was developed:

(1) The generation of a continuous ﬁber structure in

a cylindrical coordinate system; (2) Projection of the

three-dimensional structure onto a two-dimensional

image space based on the depth of ﬁbers in the structure;

(3) Modeling and blurring the top layer; and (4) Adding

visual effects adherent to various parts of the iris. An

example of a three-dimensional ﬁber structure is dis-

played in Fig. 2. The results of the procedures described

in steps 2–4 are demonstrated in Fig. 3. A small gallery

of gen erated iris images synthesized by following the

approach by Zuo et al. is displayed in Fig. 4.

Iris Sample Synthesis. Figure 1 A gallery of synthetic iris images generated using MRF approach. The regular structure

imposed by MRF is smoothed by incorporating additional synthesized features such as furrows and crypts through a

line integral convolution (Published with approval of S. Shah and A. Ross).

Iris Sample Synthesis. Figure 2 An example of a three-dimensional fiber structure generated using the methodology

described in Zuo et al. [2].

Iris Sample Synthesis

845

Quantifying Performance

When generating synthetic iris images, the problem

that one faces is to deﬁne a measure of ‘‘realism.’’

What is the set of requirements that a synthetic iris has

to satisfy to be recognized and treated as a physically

collected iris image? One can make various conclusions:

(1) it should look like a real iris. (2) it should have

Iris Sample Synthesis. Figure 3 Shown are the steps 2–4 of iris image generation.

Iris Sample Synthesis. Figure 4 A gallery of synthetic iris images generated using model based, anatomy based

approach. Iris 4 is a real iris image, a sample from CASIA dataset.

846

Iris Sample Synthesis

the statistical characteristics of a real iris. Real iris

patterns are so anatomically complex that it is nearly

impossible to mathematically describe any particular

one. Thus, standards of realism will be limited to some

degree.

The approaches that various research groups took

in order to prove the validity of generated data varied

substantially. For example, Makthal and Ross com-

pared generated iris patterns against non stochastic nat-

ural patterns from Brodatz library by invoking k-mean

clustering technique. The data were clustered in two

groups, iris and non-iris based on four distinct features

associated with texture analysis and derived from

the spatial grey level co-occurrence matrix, a second

order statistics of images. Zuo et al. analyzed the veriﬁ-

cation performance of generated iris images by extra-

polating the tails of genuine and imposter probability

density functions ﬁtted into genuine and imposter

histograms of generated and real iris images. They

also performed

▶ sensitivity analysis to conclude on

importance of various parameters involved in generat-

ing iris images.

The common feature of all approaches to evaluating

performance of generated iris data is that they focus on

evaluating the similarity between real and synthetic iris

images at three different levels: (1) global layout,

(2) features of ﬁne iris texture and (3) recognition

performance.

Summary

Multiple approaches to generating sample iris image s

exist in the literatu re. These approaches vary in terms

of tools and models used to synthesize images. Since

synthetic data are known to introduce a bias that is

impossible to predict, the data have to be used with

caution. However, in the absence of a large- or even

medium-scale dataset of real iris images, the generated

data provide an option to compare efﬁciency, limita-

tions, and capabilities of newly designed iris recogni-

tion algorithms through their testing on a large scale

dataset of generated irises.

Related Entries

▶ Face Sample Synthesis

▶ Fingerprint Sample Synthesis

References

1. Cui, J., Wang, Y., Huang, J., Tan, T., Sun, Z.: An iris image

synthesis method based on pca and super-resolution. In: Pro-

ceedings of 17th International Conference on Pattern Recogni-

tion (ICPR’04), Vol. 4, pp. 471–474. Cambridge, UK (2004)

2. Zuo, J., Schmid, N.A., Chen, X.: On generation and analysis of

synthetic iris images. IEEE Trans. Inform. Forensics Security,

2(1), 77–90 (2007)

3. Lefohn, A., Budge, B., Shirley, P., Caruso, R., Reinhard, E.: An

ocularist’s approach to human iris synthesis. IEEE Comput.

Graph. Appl. 23(6), 70–75 (2003)

4. Makthal, S., Ross, A.: Synthesis of iris images using markov

random ﬁelds. In: Proceedings of 13th European Signal Proces-

sing Conference (EUSIPCO’05). Antalya, Turkey (2005)

5. Mansﬁeld, A.J., Wayman, J.L.: Best practices in testing and

reporting performance of biometric devices. http://www.cesg.

gov.uk/site/ast/biometrics/media/BestPractice.pdf

6. Shah, S., Ross, A.: Generating synthetic irises by feature agglom-

eration. In: Proceedings of the International Conference Image

Processing. Atlanta, GA (2006)

7. Wecker, L., Samavati, F., Gavrilova, M.: A reverse subdivision

application. In: Proceedings of International Conference on

Computer Graphics and Interactive Techniques in Australasia

and Southeast Asia (GRAPHITE’05), pp. 121–125. Dunedin,

New Zealand (2005)

8. Zuo, J., Schmid, N.A.: A model based, anatomy based method

for synthesizing iris images. In: International Conference on

Biometrics (ICB’2006). pp. 428–435. Hong Kong, China (2006)

Iris Scan

▶ Biometric Sample Acquisition

Iris Scanner

▶ Iris Acquisition Device

▶ Iris Device

Iris Segmentation

Iris segmentation involves ﬁnding the pupillary and

limbic boundaries of the iris within the image allowing

Iris Segmentation

847

the iris region to be separated from the rest of the iris

image. This step is necessary prior to the creation of an

iris template and has signiﬁcant impact on iris recog-

nition performance.

▶ Iris Databases

Iris Segmentation Using Active

Contours

SUNG W. PARK,MARIOS SAVVIDES

Carnegie Mellon University, Pittsburgh, PA, USA

Synonym

Iris segmentation using snakes

Definition

Iris segmentation using active contours is ﬁnding an

iris region in an image using snakes which are curves

such as inner and outer boundaries at pupil and sclera.

Iris segmentation is a key task for iris recognition.

However, it is challenging to detect and exclude the

true inner and outer boundaries of an iris. First of all,

the occlusion caused by lower and upper eyelids and

eyelashes makes it difﬁcult to detect the iris region.

Even though eyelids and eyelashes make no occlusion,

the inner and outer boundaries are not exact circles, so

circle detection such as the Hough transform is not

enough to be applied. To obtain the iris boundaries,

snakes have been successfully applied in literature.

Snakes are energy-minimizing parametric closed curves

guided by external forces.

Introduction

Among various biometric techniques, iris recognition

can provide stable and accurate recognition rates, since

irises have highly unique pattern consisting of complex

tissue [1]. Also, iris pattern forms early, and remains

the same throughout life, so iris recognition does

not have aging or wearing problems unlike face or

ﬁngerprint recognition. However, for iris recognition,

it is required to segment iris regions successfully, and

moreover, iris segmentation is challenging in many

reasons. First of all, the different lighting conditions

often give shadows and specular reﬂections, and also,

the radius of the pupil changes because of pupil dila-

tion and contraction. Thus, we need an iris segmenta-

tion method robust on these changes in iris images.

Moreover, unfortunately, we often have troubles to

obtain a whole iris region because of the occlusion by

eyelids and eyelashes. In particular, the upper outer

boundaries are often partly or severely occluded by

upper eyelids and eyelashes.

To get successful results of iris recognition in spite

of these occlusion problems, the occlusion regions

should be detected and excluded during segmentation.

Active contours [2, 3] have been successfully applied to

detect true iris regions by detecting the boundaries

created by both an iris and occlusion.

Snakes for Curve Detection

In computer vision, edges or curves in an image are

commonly used as features, since features give a strong

presence in the images. For the low-level

▶ feature

detection, the art of feature detection has been much

studied, such as image ﬁltering by digital convolution

and simple thresholds. However, the results of these

image ﬁlters dramatically change by thresholds, so

more robust ways are required for reliable feature

detection. To overcome these limitations of previous

feature detectors, active shape modes with a variety of

forms, principally snakes [3], have been proposed [4–7].

Snakes are energy-minimizing parametric closed curves

guided by external forces. The aim of the snake is to ﬁnd

a location that minimizes energy for curve detection.

Iris segmentation is dramatically enhanced by active

contours, because active contours can ﬁt noncircular

boundaries in an iterative way. The basic concept

of the snake algorithm is brieﬂy summarized in this

section.

Using snakes, an active contour is an ordered col-

lection of points in an image:

P ¼fp

; p

; ; p

g; ð1Þ

where p

¼ (x

) is a point in the contour. The snake

method enables the points in the contour to approach

848

Iris Segmentation Using Active Contours

the boundary of an object iteratively through energy

minimization. For each point in the neighborhood of

, an energy function is computed:

E ¼ E

int

þ E

ext

; ð2Þ

where E

int

is the internal energy formed by the snake

conﬁguration, and E

ext

is the external energy formed

by external forces affecting the snake. First, E

int

dependent on the shape of the contour, and it is the

sum of the contour continuity energy E

cont

and the

contour curvature energy E

curv

. Next, E

ext

is depende nt

on the image properties, such as the gradient, and is

the sum of the image energy E

img

and the energy of

additional constraints E

con

. a and b are constants

providing the relative weighting of the two energy

terms.

Generally, regardless of a variant of external con-

straint described in [3], the energy at every point can

be written as

¼ aE

cont;i

þ bE

curv;i

þ gE

img;i

ð3Þ

where a, b, and g are the weights of every kin d of

energy. The full snake energy is the sum of all the

points. As a is bigger, snake points are more evenly

spaced. Also, as b for a certain point increases, the angle

between snake edges becomes more obtuse. Lastly, g

is responsible for making the snake point more sensi-

tive to the image energy, rather than to continuity or

curvature.

Figures 1 and 2 show two sets of points in the

outer and inner boundaries. The initial points in

Figs. 1a and 2a are set in the circles detected by a circle

detector [7]. The snake algorithm updates the points

iteratively and approach the boundaries. Commonly,

a boundary is easier to be detected than an outer

boundary, since the latter is often occluded more by

eyelids or eyelashes than the former. So, the in itial

points in an outer boundary are selected in a more

advanced way than those in an inner boundary. As

shown in Fig. 2a, the initial points in an inner bound-

ary are uniformly selected in the circle detec ted at

pupil. On the other hand, the initial points in an

outer boundary in Fig. 1(a) can be generated in either

of the two ways shown in Fig. 3. The two rows in Fig. 3

provide two different templates for outer boundary

detection. The ﬁrst row in Fig. 3b shows the template

when outstanding horizontal edges are detected

around the upper region of the iris by the Sobel edge

detector. In this case, we assume that an upper eyelid

occludes the iris partially, and initializes the snake

Iris Segmentation Using Active Contours. Figure 1 Snakes at Kth iteration for detecting an outer boundary. (a) initial

points; (b) K =1;(c) K =2;(d) K =5;(e) K = 12; (f) K = 30.

Iris Segmentation Using Active Contours

849