Expert Systems With Applications 67 (2017) 178–188

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

An eye detection method robust to eyeglasses for mobile iris

recognition

Yujin Jung

a , Dongik Kim

a , Byungjun Son

b , Jaihie Kim

a , ∗a School of Electrical and Electronic Engineering, Biometrics Engineering Research Center (BERC), Yonsei University, B619, 2nd Engineering Hall, 50 Yonsei-ro,

Seodaemun-gu, Seoul 120-749, Republic of Korea b Digital Media & Communications Business, Samsung Electronics Co.,Ltd, Digital Media & Communications R&D Center, Maetan 3-dong, Yeongtong-gu,

Suwon-si, Gyeonggi-do, 443-742, Republic of Korea

a r t i c l e i n f o

Article history:

Received 10 June 2016

Revised 3 September 2016

Accepted 23 September 2016

Available online 24 September 2016

Keywords:

Eye detection

Iris recognition

Eyeglasses

Eye validation

Iris detection

Mobile

a b s t r a c t

Finding the accurate position of an eye is crucial for mobile iris recognition system in order to extract

the iris region quickly and correctly. Unfortunately, this is very difficult to accomplish when a person is

wearing eyeglasses because of the interference from the eyeglasses. This paper proposes an eye detec-

tion method that is robust to eyeglass interference in mobile environment. The proposed method com-

prises two stages: eye candidate generation and eye validation. In the eye candidate generation stage,

multi-scale window masks consisting of 2 × 3 subblocks are used to generate all image blocks possibly

containing an eye image. In the ensuing eye validation stage, two methods are employed to determine

which blocks actually contain true eye images and locate their precise positions as well: the first method

searches for the glint of an NIR illuminator on the pupil region. If this first method fails, the next method

computes the intensity difference between the assumed pupil and its surrounding region using multi-

scale 3 × 3 window masks. Experimental results show that the proposed method detects the eye posi-

tion more accurately and quickly than competing methods in the presence of interference from eyeglass

frames.

© 2016 Elsevier Ltd. All rights reserved.

f

s

c

p

1

i

i

c

t

u

i

c

i

c

a

t

e

1. Introduction

Biometric technology provides a robust level of identity veri-

fication with convenience and reliability. As a result, fingerprint

recognition and facial recognition are being extensively adopted on

mobile devices for tasks such as mobile device unlocking. However,

demands for more rigorous and reliable identity verification meth-

ods, necessitated by e-commerce and mobile payment systems,

have led to iris recognition being considered as the next biomet-

ric technology for user verification on mobile devices ( Ashbourn,

2014 ).

Most iris recognition systems adopt near infrared (NIR) cameras

with high resolution and powerful NIR illuminators ( Ghinea, Djer-

aba, Gulliver, & Pernice Coyne, 2007 ). Iris images are captured with

NIR cameras because iris images with sharper structural patterns

can be obtained even from dark-colored irises with NIR cameras

than with visible light cameras. After input images are obtained,

eye detection is processed to locate the position of the iris. Dif-

∗ Corresponding author. Fax: + 82 2 2312 4584.

E-mail addresses: yjeclat@yonsei.ac.kr (Y. Jung), godsknight15@yonsei.ac.kr (D.

Kim), bj91.son@samsung.com (B. Son), jhkim@yonsei.ac.kr (J. Kim).

i

d

w

l

t

http://dx.doi.org/10.1016/j.eswa.2016.09.036

0957-4174/© 2016 Elsevier Ltd. All rights reserved.

erent from eye detection in the face recognition, the iris with the

ize that is appropriate for iris recognition must be detected be-

ause the detected iris region should be sufficiently large that the

attern of the iris is clear enough. According to Int. Std. ISO/IEC

9,794-6 ( Commission & Standards, 2011 ), an iris image which the

ris diameter is larger than 200 pixels is considered as good qual-

ty and which the iris diameter is between 150 and 200 pixels is

onsidered as adequate quality.

In mobile iris recognition, because of the limited resolution of

he NIR camera and the limited power of the NIR illuminators,

se of a single eye image is preferred to obtain a good quality

ris image. If both eyes are used, the view angle of the camera

ould not cover both eyes without severely distorting the iris

mage. In addition, for user convenience, the first input image

ontaining a good quality iris image is chosen automatically from

mong the continuous stream of input images without asking

he user to make a selection. Thus, each input image has to be

valuated sufficiently quickly not to miss the good iris image for

ris recognition as all the input images except the selected one are

iscarded in a real-time manner. Because of the low CPU facility

ithin a mobile phone, an eye detection algorithm should have

ow computational cost in order to detect a good iris image in real

ime. Moreover, the eye position has to be determined precisely

Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188 179

a

r

c

p

c

t

t

s

d

c

o

a

Z

3

r

l

Z

o

b

K

A

2

b

s

e

s

t

o

t

e

F

r

c

d

f

t

d

d

h

g

n

f

t

t

r

g

t

v

1

(

t

a

d

o

a

t

p

m

a

f

I

t

c

i

I

i

r

o

e

g

t

T

t

p

c

V

l

n

d

c

a

c

t

e

f

(

u

b

t

a

m

t

a

(

T

g

e

Z

w

2

w

T

i

A

i

S

fl

a

f

f

t

e

(

1

d

i

r

a

t

d

t

i

F

i

b

i

s it will subsequently affect iris boundary detection in the iris

ecognition system. Therefore, fast and accurate eye detection is

rucial in mobile iris recognition system.

Throughout the past few decades, eye detection or finding the

osition of an eye has been employed in various applications in-

luding iris recognition, gaze estimation, face modeling, and in-

elligent user interface, because information about the eye posi-

ion plays important roles in most face analysis tasks ( Ji, Wech-

ler, Duchowski, & Flickner, 2005; Wildes, 1997 ). Numerous eye

etection techniques have been developed to achieve highly ac-

urate and robust performance. In general, eye detection meth-

ds can be divided into four categories according to geometric

nd photometric properties ( Hansen & Ji, 2010; Song, Tan, Chen, &

hou, 2013 ): 1) shape-based methods; 2) feature-based methods;

) appearance-based methods; and 4) hybrid modeling methods.

Shape-based methods use the model of an eye and its sur-

ounding shapes, such as iris contour, pupil contour, and eye-

ids ( Daugman, 2004; Gwon, Cho, Lee, Lee, & Park, 2013; Jafari &

iou, 2015 ; Zhang, Chen, Su, & Liu, 2013 ). Feature-based meth-

ds identify local features in and around an eye, such as lim-

us, pupil, and cornea ( Chen & Liu, 2015; Feng & Yuen, 2001;

roon, Maas, Boughorbel, & Hanjalic, 2009; Markuš, Frljak, Pandži ́c,

hlberg, & Forchheimer, 2014; Monzo, Albiol, Sastre, & Albiol,

011 ). Appearance-based methods model the variation of an eye

y filter responses or the intensity distribution of the eye and its

urrounding area ( Ge et al., 2016; Qian & Xu, 2010; Valenti & Gev-

rs, 2012; Viola & Jones, 2004 ). Hybrid modeling methods combine

hape, feature, and appearance methods ( Kim, Lee, & Kim, 2010 ).

In this paper, a feature-based eye detection method is applied

o the gray-scaled eye images obtained by an NIR camera installed

n a mobile phone. Because the pupil and iris regions are darker

han their surrounding sclera region, dark region detection is

mployed by using the intensity information of the eye image.

eature-based methods use the local features of an eye that are

obust to illumination changes ( Hansen & Ji, 2010 ). Erroneous eye

andidates may also be obtained from regions that have similar

arkness to the eye such as an eyebrow or an eyeglass frame in

eature-based methods. However, these can be removed by using

he eye validation method proposed in this paper.

Despite much previous research on eye detection, accurate eye

etection and locating are still challenging because of various con-

itions, such as occlusion of the eye by the eyelid, variation of

ead pose, changes in lighting conditions, and interference of eye-

lasses. Interference of eyeglasses is a crucial problem that has

ot been seriously considered to date. Interference of the eyeglass

rame with the eyes degrades the accuracy of eye detection. In par-

icular, when a person is wearing eyeglasses with a thick frame,

he frame around the eye region considerably affects the accu-

acy of eye detection. Moreover, because many adults wear eye-

lasses, the interference of eyeglass frames in an eye image has

o be resolved during the eye detection process. According to Pre-

ent Blindness America and the National Eye Institute, more than

50 million Americans were wearing corrective eyewear in 2008

Prevent Blindness America & Counsil, 2008 ); further, according to

he Vision Council of America, approximately 63.2% of American

dults were wearing eyeglasses in 2010 ( Fultz & Council, 2010 ).

Eyeglasses may cause eye detection problems in iris recognition

ue to interference from the eyeglass frame and occlusion by glints

n the glass surface. Both of these problems degrade eye detection

ccuracy. Several eye detection methods that consider the condi-

ions when the eyes are obstructed by eyeglasses have been re-

orted in the literature. Kim et al., 2010 proposed an eye detection

ethod based on the eye shape feature. In their proposed method,

window mask with 3 × 3 subblocks is used to compute the dif-

erence in intensity between an eye and its surrounding regions.

n addition, a priori geometric information of both eyes, such as

he distance between them, is used to improve the accuracy. They

onducted experiments on eyeglass wearers. However, their exper-

ments did not consider those wearing thick or dark glass frames.

n these cases, when an eyeglass frame appears in the input eye

mage, a part of the eyeglass frame may be included in the first

ow of the mask comprising 3 × 3 subblocks, causing degradation

f the eye detection performance. Gwon et al. (2013) proposed an

ye detection method that utilizes the AdaBoost and CAMShift al-

orithms to search the face area. First, they remove specular reflec-

ions that appear on the lens of the eyeglass before pupil detection.

hen, they locate the center of the pupil by circular edge detec-

ion ( Daugman, 2007 ) and binarization of the pupil area. Their pro-

osed method removes specular reflection; however, they do not

onsider the problem of obstruction caused by an eyeglass frame.

alenti and Gevers (2012) proposed a method that utilizes circu-

ar symmetry based on isophote curvatures, which are curves con-

ecting points with equal intensity ( Lichtenauer, Hendriks, & Rein-

ers, 2005 ). To locate the center of an eye, isophotes in which the

urvedness is maximal are used to vote for a center. Consequently,

fter summing the votes, the point with the highest response is

onsidered to be the location of the center of the eye. The shape of

he isophotes is invariant to lighting conditions and rotation. How-

ver, due to closed eyelids and strong glints, it may detect eyeglass

rame, eyebrows, and glints instead of an actual eye. Zhu et al.

2005) proposed an appearance-based eye detection method that

tilizes an infrared illuminator. Specifically, their method uses the

right pupil effect caused by an IR illuminator. When an illumina-

or is located very close to the camera, the pupil appears bright

s a result of the reflection occurring on the pupil. If another illu-

inator is located away from the camera, a dark pupil appears on

he image. The difference image resulting from the dark eye im-

ge and the bright eye image is used by a Support Vector Machine

SVM) ( Cortes & Vapnik, 1995 ) to detect the location of the pupil.

his method is robust to reflections that appear on the lens of eye-

lasses. However, if the reflections appear on the eyeglass frame,

rrors might occur that may result in the pupils being missed.

hang, Chen, Yao, Li, and Zhuang (2007) proposed a method in

hich they detect the eye regions using AdaBoost ( Viola & Jones,

004 ) and locate the center of the pupil by Fast Radial Symmetry,

hich detects the points with high radial symmetry in intensity.

heir results suggest that redial symmetry transform can avoid the

nterference of eyeglass frames in the eye detection stage. However,

daBoost may detect the eye region inaccurately because of the

nterference of eyeglass frames in the eye-region detection stage.

everal researches have considered the interference of specular re-

ections caused by eyeglasses. However, no research focusing on

n eye detection method that avoids the interference of eyeglass

rames has been discovered to date.

In this paper, an eye detection method that is robust to inter-

erence from eyeglasses is proposed for iris recognition. Eye de-

ection is defined here as the process used to determine both the

xistence and location of an eye in the region-of-interest image

Hansen & Ji, 2010 ). The proposed method consists of two stages:

) eye candidate generation; and 2) eye validation. In the eye can-

idate generation stage, eye candidates are generated by comput-

ng the difference in intensity between the iris region and its sur-

ounding regions using multi-scale window masks which fit within

n eyeglass frame in order not to include the eyeglass frame in

he mask area when locating the eye. False findings that are pro-

uced by the masks are removed in the eye validation process. In

he eye validation stage, two methods are employed to ascertain

f a pupil region is in the mask area found in the previous stage.

irstly, a pupil region is detected by searching the glint of the NIR

lluminator that normally appears inside the pupil region. In mo-

ile iris recognition, because the distance between the camera and

lluminators is small, the glint of NIR illuminators typically appears

180 Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188

Fig. 1. Mobile iris recognition system.

Input Image

Eye Candidate Generation

Eye Candidates

Eye Validation

Validated Image

Eye Position

Fig. 2. Block diagram of the proposed eye detection method.

2

f

l

e

t

e

f

e

i

c

c

i

a

p

r

w

(

b

i

s

d

a

i

s

3

e

t

a

m

m

e

b

c

t

a

inside the pupil region. This fact is used in the proposed eye val-

idation method. If this method fails to find the pupil of the eye,

then the next method, which computes the intensity difference be-

tween the pupil region and its surrounding regions using another

much smaller 3 × 3 window mask than the one used for finding

the eye candidates, is applied to find the pupil of the eye. The main

contributions made in this paper are as follows:

• A fast eye detection algorithm robust to eyeglasses is proposed

for implementation on mobile devices

- Multi-scale window masks consisting of 2 × 3 subblocks are

proposed to generate eye candidates in the search for an

eye. The location of the eye is quickly and accurately ob-

tained while avoiding interference from any eyeglass frame

appearing in the eye image.

- An eye validation approach that employs two methods ap-

plied in sequence to select the true eye image among the

candidates produced during eye candidate generation is pro-

posed. This step enhances eye detection accuracy. • Analysis of the eye images of various races acquired by an NIR

camera is provided to formulate an eye candidate generation

threshold for different races. • A glint detection method is proposed to detect a glint that ap-

pears inside a pupil and a larger glint on the glass surface.

The remainder of this paper is organized as follows: details of

the proposed eye detection method are given in Section 2 . Exper-

imental results and performance analysis are presented in Section

3 . Finally, concluding remarks are given in Section 4 .

2. Eye detection in mobile environment

In this paper, the ultimate purpose of eye detection is for mo-

bile iris recognition; hence, only single eye images are captured

as the input images using a mobile iris system shown in Fig. 1 .

The mobile iris recognition system was implemented based on our

previous paper ( Kim, Jung, Toh, Son, & Kim, 2016 ). An NIR camera

with a resolution of 1280 × 960 pixels and two 850 nm wavelength

NIR illuminators are installed in a mobile phone. The eye-shaped

guide is displayed on the screen for user convenience.

The proposed eyeglass-robust eye detection method is illus-

trated in Fig. 2 . This method consists of two main stages: eye can-

didate generation and eye validation.

.1. Eye candidate generation

To perform iris recognition, the iris region has to be extracted

rom the input image. In order to extract the iris region, accurate

ocation of the eye position has to be obtained in advance. How-

ver, interference of the eyeglass frame with the eye could degrade

he accuracy of eye detection. Therefore, the proposed method uses

ye candidate generator which is robust to occlusion of eyeglass

rame. Eye candidate generation is the procedure that generates

ye candidates from an input eye image. An eye candidate is an

mage block that possibly contains an eye image. In this stage, the

haracteristics of eye images are explored for the detection pro-

ess. Typically, in the eye image captured by an NIR camera, the

ris region is darker than its surrounding regions such as sclera

nd skin. In the proposed method, multi-scale window masks com-

rising 2 × 3 subblocks are considered to search for the eye accu-

ately and rapidly in the mobile environment. When the mask is

indowed on the correct eye position, the upper-center subblock

B 0 ) is positioned on the iris including the pupil and the other sub-

locks correspond to the region surrounding the iris. Therefore, the

mage block, or the window mask, that has a lower average inten-

ity than its surrounding subblocks is considered as an eye candi-

ate. The intensity differences between the upper-center subblock

nd its surrounding subblocks are computed by a mask compris-

ng 2 × 3 subblocks, as depicted in Fig. 3 . Window-based methods

imilar to the proposed method have used a window mask with

× 3 subblocks to search for an eye region expecting the high-

st response when the center of 3 × 3 subblocks is positioned at

he iris of an eye; however, interference from an eyeglass frame

t the top row of the 3 × 3 mask often occurs and disturbs the

ask response resulting in inaccurate eye detection. Therefore, the

ask comprising 2 × 3 subblocks was used to avoid the interfer-

nce from an eyeglass frame. The size of a subblock is decided to

e the size of the human irises and the focused distance from the

amera to an eye. However, since the iris size differs depending on

he individual and the focused distance, masks at multiple scales

re adopted to detect the iris. In the proposed approach, because

Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188 181

Multi-scale Masks

Fig. 3. Multi-scale masks for eye candidate generation.

a

i

t

s

t

m

s

t

i

g

m

b

a

t

e

C

w

r

t

c

t

i

a

a

b

t

e

t

g

v

g

d

i

5

w

t

t

g

i

t

f

o

Fig. 4. True positive rate versus threshold.

Fig. 5. Examples of eye images from various races (ND-IRIS-0405 iris dataset): (a)

Caucasian; (b) Asian; (c) Hispanic; (d) African American.

2

g

s

d

d

d

a

e

o

f

f

r

C

C

i

t

(

e

f

c

A

f

B

E

i

n iris image is considered as adequate quality if the iris diameter

s between 150 and 200 pixels, a determination was made to use

he subblocks of 100 × 100 pixels to 250 × 250 pixels. However, the

ize of the window mask needs to be adjusted for any change in

he iris image size.

In order to be robust to different illumination conditions, nor-

alization is applied to an input image for contrast adjustment

o that the input image gains a higher contrast ( Pratt, 2007 ). This

echnique reduces the effect of illumination changes in the input

mages because the boundary between the iris and the sclera re-

ion becomes more significant. To detect the eye candidates, the

ask comprising 2 × 3 subblocks is slid from the upper left to the

ottom right of the input image, computing the feature value C pj

s in ( 1 ). The eye candidate with the highest feature value among

he eye candidates within a certain area is designated as the only

ye candidate in that area.

F = max {

C p 1 , C p 2 , · · · , C p j }

p j =

5 ∑

i =1

( E i − E 0 ) p j , p j ∈ R

E i = 1 / M

2 ·∑

( x,y ) ∈ B i I ( x, y ) (1)

here F denotes the maximum feature value within the circular

egion R , where p j is the position of the center subblock B 0 . E i is

he average intensity of block B i , and E 0 is the average intensity of

enter subblock B 0 in Fig. 3 . The feature value, C pj , is the sum of

he average intensity difference between center subblock B 0 and

ts surrounding subblocks B 1 –B 5 . I is the intensity of pixel (x,y)

nd M

2 is the number of pixels in a subblock. The integral im-

ge is used to quickly calculate the average intensity of each sub-

lock ( Crow, 1984 ; Kim et al., 2010 ). If feature value C pj is greater

han a specified threshold, center block B 0 is considered to be an

ye candidate. In general, the iris region has a higher feature value

han the other regions. However, regions such as eyebrows, eye-

lass frames, and the corners of the eyes may also have feature

alues higher than the threshold. Therefore, eye candidates are also

enerated from the regions that are not iris regions.

The threshold was decided by observing the true positive rate

uring the eye candidate generation experiments using the train-

ng dataset of 500 images wearing eyeglasses and another set of

00 images without wearing eyeglasses in our BERC dataset which

ere not used again during the test, as shown in Fig. 4 . When the

hreshold is lower, more false positives are made. The purpose of

he eye candidate generator is not to miss any true positives re-

ardless the number of many false positives. If false positive rate

s also considered, some true positives could be lost due to the

rade-off between true positive rate and false positive rate. There-

ore, only true positive rate was considered in selecting the thresh-

ld. In this paper, the threshold is determined experimentally to be

00, after which the true positive rate begins to decrease for eye-

lass wearers as well as non-wearers.

Because the feature values of neighboring eye candidates are

imilar to each other, the number of eye candidates can be re-

uced by selecting the one that represents all neighboring eye can-

idates within a certain area. The feature values of the eye candi-

ates within a certain area, e.g., a radius of 50 pixels in this case,

re compared with each other. The eye candidate with the high-

st feature value among those eye candidates is designated as the

nly eye candidate in that area. After obtaining all eye candidates

rom the input image, they are sorted in descending order by their

eature values.

The colors of iris and skin are not the same among different

aces. For example, Asians have black or dark brown irises, whereas

aucasians have light brown, blue, or green irises. The skin color of

aucasians is brighter than that of Asians. However, the colors of

ris and skin do not show much difference among different races in

he images obtained by NIR cameras. The ND-IRIS-0405 iris dataset

Phillips et al., 2010 ), created from images obtained by an NIR cam-

ra, was used to observe the difference in eye images between dif-

erent races. The ND-IRIS-0405 iris dataset contains 64,980 images

orresponding to 250 Caucasians, 82 Asians, 11 Hispanics, and two

frican Americans, as shown in Fig. 5 . The average intensity dif-

erence between the center block ( B 0 ) and its neighboring blocks,

1 –B 5 , of each race dataset was calculated, as shown in Table 1 .

ach image in the dataset was normalized in advance. The average

ntensity difference of feature values of each block, B –B , between

1 5

182 Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188

Table 1

Comparison of average intensity difference for various ethnic groups in the

ND-IRIS-0405 database.

Race B 1 B 2 B 3 B 4 B 5 E 1 −E 0 E 2 −E 0 E 3 −E 0 E 4 −E 0 E 5 −E 0

Caucasian 105 .13 88 .567 175 .20 142 .27 152 .37

Asian 93 .484 77 .715 169 .50 145 .94 151 .77

Hispanic 90 .911 70 .734 167 .17 144 .40 145 .32

African American 96 .163 83 .349 168 .86 144 .89 152 .29

Fig. 6. Examples of eye candidates.

Fig. 7. Illustration of eye validation by glint.

M

w

g

r

o

t

o

t

t

a

d

p

t

F

s

e

c

a

a

a

p

p

t

l

r

g

s

r

t

e

t

t

a

a

m

t

o

i

i

2

d

s

t

t

each race dataset is less than 15. This means that the intensities of

iris and skin are similar for each race in a normalized NIR image.

Thus, the threshold can be set equally for every race.

2.2. Eye validation

Typically, two or more eye candidates are found from an input

image during the eye candidate generation process. However, some

false eye candidates which contain parts other than the iris such

as the eyebrow, eye corner, and eyeglass frame are produced as

shown in Fig. 6 . Eye validation is therefore needed to select the eye

candidate containing the true eye image among all eye candidates.

Two eye validation methods are proposed in this paper: 1) eye

validation by glint; and 2) eye validation by pupil detection. Eye

validation by glint is used for mobile iris recognition where a glint

typically appears inside the pupil region because of the small dis-

tance between a camera and illuminators on a mobile phone. How-

ever, when a glint is not found inside the pupil, the alternative

eye validation process in which the pupil is detected by a window

mask is adopted, as described in the next section.

2.2.1. Eye validation by the glint from illuminators

Eye validation by glint is a method that detects any glint that

appears inside the pupil region. Because the NIR illuminator is nor-

mally placed near the NIR camera in a mobile iris recognizer, a

small glint typically appears inside the pupil region, as shown in

Fig. 6 . However, to avoid the red-eye effect appearing on an eye

image, the illuminator should not be placed too close to the cam-

era ( Morimoto, Koons, Amir, & Flickner, 20 0 0 ). NIR illuminators

were positioned 1.1 cm away from the NIR camera to avoid the red-

eye effect with a stand-off distance of 20 cm in this paper.

Each of the eye candidate images is checked to determine

whether there is a bright region matching the glint appearing in

the pupil. Because white glints have the highest intensity in the

image, the eight-connectivity based connected-component label-

ing ( Dillencourt, Samet, & Tamminen, 1992; Samet & Tamminen,

1988 ) is used to detect the bright region. In order to determine

the threshold for binarization, bright regions which correspond to

the illumination glints have been checked manually from normal-

ized images in the training dataset. According to the observation,

the lowest intensity was approximately 220. Therefore, every pixel

whose intensity is greater or equal to 220 was used to obtain

connected-components. After normalizing the image, the intensi-

ties of illumination glints would be close to 256. However, it would

be slightly different depending on a dataset.

After the bright regions are detected, the area ( M 00 ) and cen-

troid ( { x, y } = { M 10 / M 00 , M 01 / M 00 } ) are calculated by raw mo-

ments ( Nguyen & Shwedyk, 2009 ). The raw moment is defined as

i j =

∑

x

∑

y

x i y j I ( x, y ) (2)

here I ( x,y ) is pixel intensities. i and j are the number of orders.

However, a bright region in the pupil may also be a glint on a

lass surface or an oversaturated skin region, shown as the dashed

egion in Fig. 7 . To verify that the glint is in the pupil, the size

f the bright region is measured as the glint in a pupil is rela-

ively much smaller than the others. If the width, height, and area

f a bright region are within certain predetermined ranges, then

he region is considered as a glint candidate in a pupil. The prede-

ermined range is decided by the minimum and maximum size of

glint. The minimum size of a glint was 2 pixels for the training

ataset. The maximum size of a glint was decided as the smallest

upil size. The pupil size is normally at least 2 mm in diameter and

he average size of the iris is 12 mm in diameter ( Spector, 1990;

orrester, Dick, McMenamin, Roberts, & Pearlman, 2015 ). Therefore,

ince the block size of a window mask of the eye candidate gen-

rator is similar to an iris size, the maximum size of a glint was

onsidered as 1/6 of the block size. The width and height are same

s the size because the glint is in a circle shape. The range of the

rea of the glint is calculated by multiplying the determined width

nd height. After selecting only the small candidate glint, one more

rocess is performed to check if it lies in the pupil region. The

upils are generally dark and have low intensity. Thus, the intensi-

ies of pixels surrounding the glint in a pupil were approximately

ess than 40 in the normalized image of the training dataset. The

egion surrounding the bright region is defined as the dotted re-

ion in Fig. 7 . The ratio of the number of dark pixels in the region

urrounding the bright region (dotted region) is measured. If the

atio is greater than the predetermined threshold, it is considered

hat the pupil with the glint inside is an eye candidate. Thus, the

ye candidate is validated as it contains a true eye. The center of

he glint is defined as the eye position. Therefore, even if reflec-

ions and eyeglass frame disturb detecting eye in the eye image,

n accurate eye position could be found as long as the glint inside

pupil is detected.

60% of pixels in the region surrounding the glint were deter-

ined to be dark for the images with a glint on the boundary of

he pupil in the training set. Therefore, the predetermined thresh-

ld was decided as 60. However, the threshold could vary depend-

ng on a dataset because the size and power of the installed NIR

lluminator are different.

.2.2. Eye validation by pupil detection

Eye validation by pupil detection is an alternative to eye vali-

ation by glint. It is applied when validation by glint has failed as

hown in Fig. 8 or as a substitute for eye validation by glint for

he other iris recognition when there is no glint in the pupil. In

his paper, eye validation by pupil detection is used when a glint

Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188 183

Fig. 8. Block diagram of the eye validation process.

Fig. 9. (a). Eye candidate image; (b) Binarized image; (c) Mask used for pupil de-

tection.

i

f

t

m

p

i

i

w

p

o

a

o

d

s

d

g

a

d

b

F

p

v

t

f

c

C

w

a

t

l

a

i

b

w

e

e

b

a

i

t

t

h

T

C

w

i

o

a

i

m

i

t

a

t

o

T

c

a

A

2

s

t

t

3

1

w

s not found in the pupil. A pupil is at the center of the iris. There-

ore, if a pupil exists in the eye candidate image, it can be inferred

hat the eye candidate contains a true eye image. In this section, a

ethod that detects a pupil from eye candidates is presented.

Because the eye images are obtained by an NIR camera, the

upil is darker than other regions such as iris, sclera, and skin. The

ntensity is also lower than the other regions. Image binarization

s carried out to convert an eye candidate image to a black and

hite image, as shown in Fig. 9 . A threshold value was chosen ex-

erimentally as 40 so that pupils are converted to black and the

ther regions are converted to white. The intensities of pixels in

pupil were approximately less than 40 in the normalized image

f the training dataset; the parameter might be slightly different

epending on a dataset. Following image binarization, the inten-

ity difference between pupil and the other regions becomes very

istinct. Consequently, the pupil is detected by searching the re-

ion where the intensity differences between the center block ( B )

0

nd its surrounding blocks ( B 1 –B 8 ) are considerable. The intensity

ifferences between the center subblock and its surrounding sub-

locks are computed by a mask comprising 3 × 3 subblocks, as in

ig. 9 (c). The size of a subblock is decided to be the size of the

upil. However, since the pupil size differs depending on the indi-

idual and lighting condition, masks at multiple scales are adopted

o detect the pupil.

To detect a pupil, the mask comprising 3 × 3 subblocks is slid

rom the upper left to the bottom right of an eye candidate image

omputing the feature value, as in ( 3 ):

1 =

8 ∑

i =1

( E i − E 0 )

E i = 1 / M

2 ·∑

( x,y ) ∈ B i I ( x, y ) (3)

here C 1 is the feature value of the intensity difference, E i is the

verage intensity of block B i , and E 0 is the average intensity of cen-

er block B 0 in Fig. 9 (c). I is the intensity of pixel x and M is the

ength of the block.

Here, again the integral image is used to quickly calculate the

verage intensity of each subblock ( Crow, 1984 ). Feature value C 1 s the sum of the average different intensities between center sub-

lock B 0 and its surrounding subblocks B 1 –B 8 . Center subblock B 0 ith the highest feature value, C 1 , is considered a pupil. In gen-

ral, a pupil has a higher feature value than other regions. How-

ver, after image binarization of an eye candidate image, the eye-

row, eyelashes, and eyeglass frame also convert to black. Thus,

dditional information on the intensity difference is needed. The

ntensity variance of surrounding subblocks is used for pupil de-

ection. Because eyelashes may be included in the subblocks be-

ween B 1 and B 5 , the intensity variances of subblocks B 1 –B 5 are

igh, whereas the intensity variances of subblocks B 6 –B 8 are low.

he feature value of the intensity variance is computed as in ( 4 ):

2 =

8 ∑

i =6

V i

V i = 1 / M

2 ·∑

( x,y ) ∈ B i ( I ( x, y ) − E i )

2 (4)

here C 2 is the feature value of the intensity variance, V i is the

ntensity variance of subblock B i , and E i is the intensity average

f subblock B i . Feature value C 2 is the sum of the intensity vari-

nces of subblocks B 6 –B 8 . I is the intensity of pixel (x,y) and M

s the length of the block. Therefore, an eye candidate with the

aximum C 1 is considered a true eye image that contains a pupil

mages if feature values C 1 and C 2 satisfy the conditions of both

hresholds. The center of subblock B 0 , the pupil region, is defined

s the eye position. The thresholds were obtained by using the

raining dataset of 500 images wearing eyeglasses and another set

f 500 images without wearing eyeglasses in our BERC dataset.

he thresholds for C 1 ( Eq. 3 ) and C 2 ( Eq. 4 ) were selected by ROC

urves ( Fig. 10 ) showing the trade-off between true positive rate

nd false positive rate with varying threshold for both C 1 and C 2 .

ccording to the ROC curve, the threshold of C 1 was decided as

00 and C 2 was decided as 5000 which have the lowest EER. C 1 hould be greater than the threshold and C2 should be less than

he threshold in order for an eye candidate to be considered as a

rue eye. The parameters vary depending on a dataset.

. Experimental results

Experiments were conducted under three different conditions:

) without eyeglasses, 2) with eyeglasses, and 3) both without and

ith eyeglasses. The detection rate, accuracy, and processing time

184 Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188

Fig. 10. ROC curves for eye validation threshold: (a) C 1 ; (b) C 2 .

Fig. 11. Examples of database images: BERC mobile iris database (a) without eyeglasses (b) with eyeglasses, CASIA 4 - Iris Thousand database (c) without eyeglasses (d) with

eyeglasses.

i

o

t

e

9

w

d

g

l

i

o

s

s

t

of the proposed method were then compared with existing meth-

ods.

3.1. Experimental iris image data

The proposed method was tested on two databases: the BERC

mobile iris database collected for this work and the publicly avail-

able CASIA 4 – Iris Thousand database. For the BERC database, im-

ages were captured by an iris recognition system which was imple-

mented on a mobile phone. 2007 images were collected from 50

subjects not wearing eyeglasses and 1004 images were collected

from 25 subjects wearing eyeglasses by using the iris recogni-

tion system. The subjects captured the images holding the mobile

phone by herself/himself. Various types of eyeglass frames with

varying thicknesses were used, as shown in Fig. 11 (b). Continuous

nput stream images were obtained via an NIR camera installed

n the mobile phone and the user adjusted the camera holding

he mobile phone. The operating distance from the camera to the

ye was between 15 cm and 20 cm, and the image resolution was

60 × 1280 pixels.

The second database was the CASIA 4 - Iris Thousand (2015) ,

hich contains 19,988 images collected from 10 0 0 subjects. The

atabase consists of 14,668 images captured without wearing eye-

lasses and 5320 images captured with eyeglass on. However, un-

ike the BERC eyeglass data, CASIA eyeglass data include predom-

nantly thin or bright eyeglass frames rather than thick or dark

nes, as shown in Fig. 11 (d). Thus, the difference between the re-

ult of images captured with and without eyeglasses was not con-

iderable for the CASIA 4 – Iris Thousand database compared to

he BERC mobile iris database.

Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188 185

Table 2

Eye detection rates on BERC database.

Method Detection rate (%) within 50 pixels

Without glasses With glasses Without glasses + with glasses

AdaBoost 91 .95 79 .44 85 .70

RED 97 .81 86 .99 92 .40

Isophote 99 .30 89 .37 94 .34

MEC 96 .82 79 .64 88 .24

EC + EV (Proposed) 99 .31 94 .42 96 .87

3

p

a

a

e

i

c

f

c

d

t

c

n

b

g

M

e

c

i

q

d

d

t

a

E

a

s

t

g

o

B

f

l

p

o

e

a

A

e

s

m

p

T

b

p

m

a

s

t

t

t

a

r

r

d

b

A

o

a

E

t

F

a

r

d

e

a

p

u

s

t

w

b

.2. Eye detection results and analysis

For the first experiment, the detecting rate of EC + EV was com-

ared with AdaBoost, RED, Isophote , and MEC on the BERC database,

nd the results are displayed in Table 2.

The various representations utilized in this paper are abbrevi-

ted as follows:

• AdaBoost : AdaBoost method ( Viola & Jones, 2004 ) using Harr-

like features, which is used to find the location of an eye.

This approach is selected because it is a widely employed and

adapted method for detecting an eye from a face image in iris

recognition ( Rathgeb, Uhl, & Wild, 2012 ). • RED (Rapid eye detection) : Mask comprising 3 × 3 subblocks

method, which is adopted for eye candidate generation. ( Kim

et al., 2010 ). An eye is detected by comparing the registered

iris shape features to the feature on every position of the input

image. • Isophote (Isophote based method) : Method that utilizes cir-

cular symmetry based on isophote curvatures ( Valenti & Gev-

ers,2012 ). To locate the center of an eye, isophotes in which

the curvedness is maximal are used to vote for a center. Con-

sequently, after summing the votes, the point with the highest

response is considered to be the location of the center of the

eye. • MEC (Maximum eye candidate) : Mask comprising 2 × 3 sub-

blocks method, which is adopted for the proposed eye candi-

date generation method without eye validation. The eye candi-

date with the highest feature value is considered to contain the

true eye image. • EC (Eye candidate generator) + EV (Eye validation) : Mask com-

prising 2 × 3 subblocks method, which is adopted for the pro-

posed eye candidate generation method followed by the pro-

posed eye validation method.

In the iris recognition process, after finding the position of an

ye, the boundary between the iris and pupil is searched start-

ng from the determined eye location. For fast iris segmentation,

oarse pupil localization is adopted to reduce the search space, and

or fast pupil localization, the determined eye location should be

lose to the pupil center ( Rathgeb et al., 2012 ). Therefore, the eye

etection was considered to be a success when the distance be-

ween the determined eye location and the manually marked pupil

enter was less than 50 pixels.

The proposed method achieved the detection rate of 99.31% for

on-glass wearers and 94.42% for eyeglass wearers on BERC mo-

ile iris database. The detection rate was 2.49% higher for non-

lass wearers and 14.78% higher for eyeglass wearers than that of

EC , the method without eye validation. The result shows that the

ye detection rate improved when eye candidate generation was

ombined with eye validation. MEC selected regions other than the

ris, such as the eyebrow, eye corner, and eyeglass frame more fre-

uently than EC + EV ; however, in EC + EV , the proposed eye vali-

ation helped to show better performance in selecting the eye can-

idate containing the true eye image. RED and Isophote achieved

he detection rate of 97.81% and 99.30% for non-glass wearers

nd eyeglass wearers, respectively, which is a little lower than

C + EV. However, for eyeglass wearers, the detection rate of RED

nd Isophote was 7.43% and 5.05% lower than that of EC + EV , re-

pectively. In the case of eyeglass wearers, RED may have failed

o detect the true eye from images with various kinds of eye-

lasses because only one iris feature is used for registration. More-

ver, Isophote may have failed to detect the true eye because the

ERC database contains images with strong glint on the glass sur-

ace. A strong glint on the glass surface may change the circu-

ar eye pattern, which leads to the weak contribution of iris and

upil in center voting ( Valenti & Gevers, 2012 ). The detection rate

f EC + EV was higher than that of AdaBoost on both datasets of

yeglass wearers and non-wearers because strong glints on glasses

nd eyeglass frame may have interfered with the eye detection of

daBoost .

The proposed method performed the best on both datasets of

yeglass wearers and non-wearers because strong glints on a glass

urface do not distract the eye detection unless the glints cover

ost of the iris region. In addition, the proposed validation process

revents the system from selecting eyeglass frames as the true eye.

he outstanding performance of EC + EV is attributable to the com-

ination of eye candidate generation using the 2 × 3 mask and the

roposed eye validation method.

The locating accuracy of the detected eye by the proposed

ethod was compared with other methods on the BERC database,

s shown in Table 3 . The accuracy was measured by the root mean

quare (RMS) error between the manually marked pupil center and

he location of the eye position. The error was calculated only for

he samples for which eye detection was successful. Table 3 shows

hat the position error is the lowest in EC + EV . Because the im-

ges in the BERC database are obtained by a mobile phone iris

ecognition system, the iris region that is illuminated by LED di-

ectly is bright; however, the iris region on the opposite side is

ark due to the shadow created by the eye’s protruding part that

locks the light ( Kim et al., 2016 ). The eye positions detected by

daBoost, RED , and MEC are not accurate due to the interference

f glint, eyeglass frame, and the shadowy part of the input im-

ge. However, the results show that the best accuracy comes from

C + EV for both the dataset with and without eyeglasses because

he eye validation method in EC + EV reduces the position error.

ig. 12 shows the examples of the results obtained on different im-

ges of the BERC database and Fig. 13 shows the examples of the

esults for eyes occluded by a reflection or eyeglass frame.

Similar experiments were also conducted using the CASIA

atabase. The images of the CASIA database are obtained in differ-

nt illumination compared to the BERC database. Two glints gener-

lly appear in the pupil region and sometimes they are not in the

upil region. Thus, only the eye validation by pupil detection was

sed for the CASIA database.

The results for EC + EV on the CASIA dataset with eyeglasses

how less improvement compared to the other methods than for

he BERC dataset owing to the fact that the dataset of eyeglass

earers in the CASIA dataset are primarily users with thin or

right colored glass frames, which have little effect on eye detec-

186 Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188

Table 3

Eye detection RMS error for BERC database.

Method RMS error (pixels) within 50 pixels

Without glasses With glasses Without glasses + with glasses

Mean Mean Mean

AdaBoost 23 .48 22 .71 23 .10

RED 18 .42 24 .80 21 .61

Isophote 9 .73 9 .68 9 .70

MEC 20 .52 28 .20 24 .35

EC + EV (Proposed) 6 .54 7 .64 7 .09

Fig. 12. Examples of detection results for eyes: BERC database (a) without eyeglasses (b) with eyeglasses.

Fig. 13. Examples of detection results for occluded eyes: BERC database (a) by reflection (b) by eyeglass frame.

Table 4

Eye detection rate for CASIA database.

Method Detection rate (%) within 50 pixels

Without glasses With glasses Without glasses + with glasses

AdaBoost 96 .48 92 .81 95 .50

RED 98 .64 95 .27 97 .74

Isophote 94 .41 82 .87 91 .34

MEC 94 .24 89 .58 93 .00

EC + EV (Proposed) 99 .67 98 .48 99 .35

l

d

o

m

e

f

tion. Tables 4 and 5 respectively show the detection rate and accu-

racy of the eye detection methods. The proposed method achieved

the detection rate of 99.67% for non-glass wearers and 98.48% for

eyeglass wearers on BERC mobile iris dataset. Owing to the fact

that the dataset of eyeglass wearers in the CASIA database are pri-

marily users with thin or bright colored glass frames, which have

ittle effect on eye detection, the results for EC + EV on the CASIA

atabase with eyeglasses show less improvement compared to the

ther methods than for the BERC database. However, the proposed

ethod still performed the best on both datasets of eyeglass wear-

rs and non-wearers. Other existing methods except Isophote, per-

ormed better on the CASIA database than the BERC dataset; how-

Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188 187

Table 5

Eye detection RMS error for CASIA database.

Method RMS error (pixels) within 50 pixels

Without glasses With glasses Without glasses + with glasses

Mean Mean Mean

AdaBoost 16 .37 22 .69 18 .05

RED 15 .26 17 .84 15 .95

Isophote 28 .87 30 .62 29 .34

MEC 28 .27 29 .74 28 .66

EC + EV (Proposed) 15 .16 16 .24 15 .44

Fig. 14. Examples of detection results for eyes: CASIA 4 - Iris Thousand database (a) without eyeglasses (b) with eyeglasses.

Table 6

Processing times for eye detection on BERC database (PC).

Method Processing time (ms)

Without glasses With glasses Average

AdaBoost 24 .85 29 .85 27 .35

RED 4 .85 6 .18 5 .51

Isophote 11 .93 12 .73 12 .33

EC + EV (Proposed) 5 .54 6 .93 6 .23

e

B

S

w

c

n

M

d

a

E

a

r

s

a

t

2

p

T

t

T

s

p

r

a

Table 7

Processing times for eye detection on BERC database (mobile phone).

Method Processing time (ms)

Without glasses With glasses Average

EC + EV (Proposed) 25 .27 27 .34 26 .31

a

m

T

g

g

4

r

g

b

d

i

t

e

i

t

m

c

n

H

s

u

N

M

i

ver, Isophote performed worse on the CASIA database than the

ERC database because a lot of images of eyeglass wearers in CA-

IA database contain strong glints on eyeglasses and Isophote has

eakness for strong glints. In addition, Isophote selected an eye

orner as the position of maximum relevance in many images of

on-eyeglass wearers in the CASIA database.

The detection rate of EC + EV was 5.43% and 8.9% higher than

EC. This proves that performing eye validation following eye can-

idate generation reduces the errors for both the datasets with

nd without eyeglasses as in the BERC case. The proposed method,

C + EV , showed the best performance for both detection rate and

ccuracy on the CASIA database. Fig. 14 shows the examples of the

esults obtained on different images of the CASIA database.

The processing times for the eye detection methods on PC are

hown in Table 6 . The average processing times were measured by

computer with 3.07 GHz CPU and 6GB RAM. The results show

hat the processing time of the proposed method is approximately

2.8% that of AdaBoost and 50.5% that of Isophote , because the pro-

osed method has lower complexity than AdaBoost and Isophote .

he proposed method was less than 1 ms slower than RED because

he proposed method includes the two eye validation processes.

he eye validation processes increased the computation, which re-

ulted in the increased processing time. However, because the pro-

osed method yielded the highest accuracy, considering detection

ate and processing time, the proposed method gave the best over-

ll performance.

The processing time for the proposed eye detection method on

mobile phone is shown in Table 7 . The processing times were

easured by a mobile phone with 1.5 GHz CPU and 512 MB RAM.

he processing time for the images captured without wearing eye-

lasses was 25.27 ms and that for the images captured with eye-

lasses on was 27.34 ms.

. Conclusion

This paper proposed an eye detection method for mobile iris

ecognition that is fast and less affected by interference from eye-

lasses than conventional methods. The proposed method com-

ines eye candidate generation with eye validation. The eye can-

idate generator finds eye candidates that possibly contain an eye

mage, and the eye validator decides which eye candidate contains

he true eye image. By using a mask comprising 2 × 3 subblocks,

yeglass frames around the eye in the input image can be avoided

n the eye candidate generation process. Eye validation improves

he detection rate and accuracy of the eye detector. The proposed

ethod also has low computational complexity. These advantages

onfirm that the proposed method is feasible for mobile iris recog-

ition.

This paper focused on eye detection for mobile iris recognition.

owever, the proposed method can be adapted to iris recognition

ystems that find both eyes and other applications that detect eyes

sing images obtained from NIR cameras.

The performance of an iris recognition system which uses an

IR camera generally degrades in unconstrained environments.

oreover, since the user holds the mobile device to capture the

mage, off-angle and rotational issues have to be solved. In the fu-

188 Y. Jung et al. / Expert Systems With Applications 67 (2017) 178–188

K

K

M

M

M

N

P

P

Q

R

S

S

S

V

V

P

W

Z

ture, we plan to implement the iris recognition system which is

robust to unconstrained environments, off-angle, and rotations.

Acknowledgments

Portions of the research in this paper use the CASIA-IrisV4 col-

lected by the Chinese Academy of Sciences’ Institute of Automation

(CASIA), and the ND-IRIS-0405 Iris Image Dataset collected by the

University of Nortre Dame (UND).

References

Ashbourn, J. (2014). Biometrics in the new world . New York: Springer. doi: 10.1007/978- 3- 319- 04159- 9 .

CASIA Iris Image Database (2015). Retrieved January 15, 2015, from http://biometrics.idealtest.org/ .

Chen, S., & Liu, C. (2015). Eye detection using discriminatory Haar features and a

new efficient SVM. Image and Vision Computing, 33 , 68–77. doi: 10.1016/j.imavis.2014.10.007 .

Commission, I. E., & Standards, I. (2011). Int. Std. ISO/IEC 19794-6. InformationTechnology—Biometric Data Interchange Formats—Part 6: Iris Image Data, Pub. L.

No. ISBN 0-580-84261-0. Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20 (3),

273–297. doi: 10.10 07/BF0 0994018 .

Crow, F. C. (1984). Summed-area tables for texture mapping. ACM SIGGRAPH Com-puter Graphics, 18 (3), 207–212. doi: 10.1145/964965.808600 .

Daugman, J. (2004). How Iris Recognition Works. Circuits and Systems for VideoTechnology, IEEE Transactions on, 14 (1), 21–30. doi: 10.1016/B978- 0- 12- 374457- 9.

0 0 025-1 . Daugman, J. (2007). New methods in iris recognition. IEEE Transactions on Systems,

Man, and Cybernetics. Part B, Cybernetics : A Publication of the IEEE Systems, Man,

and Cybernetics Society, 37 (5), 1167–1175. doi: 10.1109/TSMCB.2007.903540 . Dillencourt, M. B., Samet, H., & Tamminen, M. (1992). A general approach to

connected-component labeling for arbitrary image representations. Journal ofthe ACM, 39 (2), 253–280. doi: 10.1145/128749.128750 .

Feng, G. C., & Yuen, P. C. (2001). Multi-cues eye detection on gray intensity image.Pattern Recognition, 34 , 1033–1046. doi: 10.1016/S0 031-3203(0 0)0 0 042-X .

Forrester, J. V. , Dick, A. D. , McMenamin, P. G. , Roberts, F. , & Pearlman, E. (2015). The

eye: Basic sciences in practice . Elsevier health sciences . London: Saunders . Fultz, G., & Council, T. V. (2010). 2nd Quarter 2010 US Optical Market Overview and

Outlook, 1–55. Ge, S., Yang, R., He, Y., Xie, K., Zhu, H., & Chen, S. (2016). Learning multi-channel

correlation filter bank for eye localization. Neurocomputing, 173 , 418–424. doi: 10.1016/j.neucom.2015.03.125 .

Ghinea, G., Djeraba, C., Gulliver, S., & Pernice Coyne, K. (2007). Introduction tospecial issue on eye-tracking applications in multimedia systems. ACM Trans-

actions on Mutimedia Computing, Communications and Application, 3 (4), 2–5.

doi: 10.1145/1314303.1314304 . Gwon, S. Y., Cho, C. W., Lee, H. C., Lee, W. O., & Park, K. R. (2013). Robust eye

and pupil detection method for gaze tracking. International Journal of AdvancedRobotic Systems, 10 (98), 1–7. doi: 10.5772/55520 .

Hansen, D. W., & Ji, Q. (2010). In the eye of the beholder: A survey of models foreyes and gaze. IEEE Transactions on Pattern Analysis and Machine Intelligence,

32 (3), 478–500. doi: 10.1109/TPAMI.2009.30 .

Jafari, R., & Ziou, D. (2015). Eye-gaze estimation under various head positions andiris states. Expert Systems with Applications, 42 (1), 510–518. doi: 10.1016/j.eswa.

2014.08.003 . Ji, Q., Wechsler, H., Duchowski, A., & Flickner, M. (2005). Special issue: Eye detection

and tracking. Computer Vision and Image Understanding, 98 , 1–3. doi: 10.1016/j.cviu.20 04.07.0 06 .

im, B., Lee, H., & Kim, W.-Y. (2010). Rapid eye detection method for non-glassestype 3D display on portable devices. IEEE Transactions on Consumer Electronics,

56 (4), 2498–2505. doi: 10.1109/TCE.2010.5681133 . Kim, D., Jung, Y., Toh, K.-A., Son, B., & Kim, J. (2016). An empirical study on iris

recognition in a mobile phone. Expert Systems With Applications, 54 , 328–339.doi: 10.1016/j.eswa.2016.01.050 .

roon, B., Maas, S., Boughorbel, S., & Hanjalic, A. (2009). Eye localization in low andstandard definition content with application to face matching. Computer Vision

and Image Understanding, 113 (8), 921–933. doi: 10.1016/j.cviu.2009.03.013 .

Lichtenauer, J., Hendriks, E., & Reinders, M. (2005). Isophote properties as featuresfor object detection. In 2005 IEEE computer society conference on computer vision

and pattern recognition (CVPR’05): 2 (pp. 649–654). doi: 10.1109/CVPR.2005.196 . arkuš, N., Frljak, M., Pandži ́c, I. S., Ahlberg, J., & Forchheimer, R. (2014). Eye pupil

localization with an ensemble of randomized trees. Pattern Recognition, 47 , 578–587. doi: 10.1016/j.patcog.2013.08.008 .

onzo, D., Albiol, A., Sastre, J., & Albiol, A. (2011). Precise eye localization using

HOG descriptors. Machine Vision and Applications, 22 (3), 471–480. doi: 10.1007/s00138-010-0273-0 .

orimoto, C., Koons, D., Amir, A., & Flickner, M. (20 0 0). Pupil detection and trackingusing multiple light sources. Image and Vision Computing, 18 (4), 331–335. doi: 10.

1016/S0262-8856(99)0 0 053-0 . guyen, H. , & Shwedyk, E. (2009). A first course in digital communications . New York:

Cambridge University Press .

hillips, P. J., Scruggs, W. T., O’Toole, A. J., Flynn, P. J., Bowyer, K. W., Schott, C. L.,et al. (2010). FRVT 2006 and ICE 2006 large-scale experimental results. IEEE

Transactions on Pattern Analysis and Machine Intelligence, 32 (5), 831–846. doi: 10.1109/TPAMI.2009.59 .

revent Blindness America, & Counsil, N. V. (2008). Vision Problems in the U.S.America , 1–42.

ian, Z., & Xu, D. (2010). Automatic eye detection using intensity filtering and K-

means clustering. Pattern Recognition Letters, 31 (12), 1633–1640. doi: 10.1016/j.patrec.2010.05.012 .

athgeb, C. , Uhl, A. , & Wild, P. (2012). Iris biometrics: from segmentation to templatesecurity : Vol. 59. Springer Science & Business Media .

amet, H., & Tamminen, M. (1988). Efficient Component Labeling of Images of Ar-bitrary Dimension Represented By Linear Bintrees. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 10 (4), 579–586. doi: 10.1109/34.3918 .

ong, F., Tan, X., Chen, S., & Zhou, Z. H. (2013). A literature survey on robust andefficient eye localization in real-life scenarios. Pattern Recognition, 46 (12), 3157–

3173. doi: 10.1016/j.patcog.2013.05.009 . pector, R. H. (1990). The Pupils. In W. HK, H. WD, & H. JW (Eds.), Clinical methods:

the history, physical, and laboratory examinations (3rd ed.). Boston: Butterworths.doi: 10.1016/B978- 1- 4557- 3984- 4.00205- 0 .

alenti, R., & Gevers, T. (2012). Accurate eye center location through invariant

isocentric patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence,34 (9), 1785–1798. doi: 10.1109/TPAMI.2011.251 .

iola, P., & Jones, M. (2004). Robust real-time face detection. International Journal ofComputer Vision, 57 (2), 137–154. doi: 10.1023/B:VISI.0 0 0 0 013087.49260.fb .

ratt, W. K. (2007). Digital image processing (4th ed.). New Jersey: Wiley-Inter-science .

ildes, R. P. (1997). Iris recognition: An emerging biometric technology. Proceedingsof the IEEE, 85 (9), 1348–1363. doi: 10.1109/5.628669 .

Zhang, G., Chen, J., Su, G., & Liu, J. (2013). Double-pupil location of face images.

Pattern Recognition, 46 (3), 642–648. doi: 10.1016/j.patcog.2012.08.013 . hang, W., Chen, H., Yao, P., Li, B., & Zhuang, Z. (2007). Precise eye localization with

AdaBoost and fast radial symmetry. In 2006 international conference on computa-tional intelligence and security, ICCIAS 2006: 1 (pp. 725–730). doi: 10.1109/ICCIAS.

2006.294229 . Zhu, Z., & Ji, Q. (2005). Robust real-time eye detection and tracking under variable

lighting conditions and various face orientations. Computer Vision and Image Un-

derstanding, 98 (1), 124–154. doi: 10.1016/j.cviu.2004.07.012 .