SS symmetry

Article
Deep Learning-Based Iris Segmentation for Iris
Recognition in Visible Light Environment
Muhammad Arsalan, Hyung Gil Hong, Rizwan Ali Naqvi, Min Beom Lee, Min Cheol Kim,
Dong Seop Kim, Chan Sik Kim and Kang Ryoung Park *
Division of Electronics and Electrical Engineering, Dongguk University, 30 Pildong-ro 1-gil, Jung-gu,
Seoul 100-715, Korea; arsal@dongguk.edu (M.A.); hell@dongguk.edu (H.G.H);
rizwanali@dongguk.edu (R.A.N.); mblee@dongguk.edu (M.B.L.); mincheol9166@naver.com (M.C.K.);
k_ds1028@naver.com (D.S.K.); kimchsi9004@naver.com (C.S.K.)
* Correspondence: parkgr@dongguk.edu; Tel.: +82-10-3111-7022; Fax: +82-2-2277-8735

Received: 8 October 2017; Accepted: 1 November 2017; Published: 4 November 2017

Abstract: Existing iris recognition systems are heavily dependent on specific conditions, such as the
distance of image acquisition and the stop-and-stare environment, which require significant user
cooperation. In environments where user cooperation is not guaranteed, prevailing segmentation
schemes of the iris region are confronted with many problems, such as heavy occlusion of eyelashes,
invalid off-axis rotations, motion blurs, and non-regular reflections in the eye area. In addition,
iris recognition based on visible light environment has been investigated to avoid the use of
additional near-infrared (NIR) light camera and NIR illuminator, which increased the difficulty
of segmenting the iris region accurately owing to the environmental noise of visible light. To address
these issues; this study proposes a two-stage iris segmentation scheme based on convolutional neural
network (CNN); which is capable of accurate iris segmentation in severely noisy environments of
iris recognition by visible light camera sensor. In the experiment; the noisy iris challenge evaluation
part-II (NICE-II) training database (selected from the UBIRIS.v2 database) and mobile iris challenge
evaluation (MICHE) dataset were used. Experimental results showed that our method outperformed
the existing segmentation methods.

Keywords: biometrics; iris recognition; iris segmentation; convolutional neural network (CNN)

1. Introduction
Biometrics has two main categories: physiological and behavioral biometrics. Behavioral
biometrics considers voice, signature, keystroke, and gait recognition [1], whereas physiological
biometrics considers face [2,3], iris [4,5], fingerprints [6], finger vein patterns [7], and palm prints [8].
Iris recognition has been widely used in security and authentication systems because of its reliability
and high-security [9,10].
Most existing iris recognition algorithms are designed for highly controlled cooperative
environments, which is the cause of their failure in non-cooperative environments, i.e., those that
include noise, off-angles, motion blurs, glasses, hairs, specular reflection (SR), eyelids and eyelashes
incorporation, and partially open eyes. Furthermore, the iris is always assumed to be a circular object,
and common methods segment it as a circle, but considering intense cases of side view and partially
open eyes, the iris boundary deviates from being circular and may include skin, eyelid, and eyelash
areas. Iris recognition systems are based on the specific texture of the iris area, which is used as a base
entity for recognition and authentication purposes. Therefore, the accurate segmentation of the iris
boundary is important even in intense environments.
Algorithms for iris segmentation should be designed to reduce user cooperation to improve the
overall iris recognition performance. Many studies have attempted to reduce the error caused by the

Symmetry 2017, 9, 263; doi:10.3390/sym9110263 www.mdpi.com/journal/symmetry

Symmetry 2017, 9, 263 2 of 25

lack of user cooperation in the past two decades [11–13], but the detection of the true iris boundary is
still a challenge. To address this issue, we proposed a two-stage iris segmentation method based on
convolutional neural networks (CNN), which is capable of robustly finding the true iris boundary in
the above-mentioned intense cases with limited user cooperation. Our proposed iris segmentation
scheme can be used with inferior quality noisy images even in visible light environment.

2. Related Works
The existing schemes for iris segmentation can be broadly divided into four types based
on implementation. The first and most common type consists of the boundary-based methods.
These systems will first find the base element pupil as an inner boundary of the iris, and then find
the other parameters, such as eyelid and limbic areas, to separate them from the iris. The first type of
methods include Hough transform (HT) and Daugman’s integro-differential operator. HT finds the
circularity by edge-map voting within the given range of the radius, which is known as the Wildes
approach [14]. HT-based detection methods are also used [15,16]. Daugman’s integro-differential
operator is another scheme that finds the boundary using an integral derivative, and the advanced
method was developed in [17]. An effective technique to reduce the error rate in a non-cooperative
environment was proposed by Jeong et al. [18]. They used two circular edge detectors in combination
with AdaBoost for pupil and iris boundary detection, and their method approximated the real boundary
of eyelashes and eyelid detection. Other methods are also known to reduce noise prior to the detection
of the iris boundary to increase the segmentation accuracy [19]. All methods in the first type of
segmentation require eye images of good quality and an ideal imaging environment; therefore,
these types of methods are less accurate for non-ideal situations or result in higher error rates.
The second type of methods is composed of pixel-based segmentation methods. These methods
are based on the identification of the iris boundary using specific color texture and illumination
information gradient to discriminate between an iris pixel and another pixel in the neighborhood.
Based on the discriminating features, the iris classifier is created for iris and non-iris pixel classification.
A novel method for iris and pupil segmentation was proposed by Khan et al. [20]. They used
2-D profile lines between the iris and sclera boundary and calculated the gradient pixel by pixel,
where the maximum change represents the iris boundary. Parikh et al. [21] first approximated the iris
boundary by color-based clustering. Then, for off-angle eye images, two circular boundaries of the
iris were detected, and the intersection area of these two boundaries was defined as the iris boundary.
However, these methods are affected by eyelashes and hairs or dark skin. The true boundary of
the iris is also not identified if it includes the area of eyelids or skin, which reduces the overall iris
authentication performance.
The third type of segmentation methods is composed of active contours and circle fitting-based
methods [22,23]. A similar approach is used in the local Chan–Vese algorithm, where a mask is created
according to the size of the iris, and then an iterative process determines the true iris boundary with the
help of the localized region-based formulation [24]. However, this approach shares the drawback faced
by other active contour-based models, i.e., it is usually prevented by the iris texture during iteration,
and considers the iris pattern as the boundary, which results in erroneous segmentation. On the other
hand, active contour-based methods are more reliable in detecting the pupillary boundary because of
the significant difference in visibility.
To eliminate the drawbacks of all current segmentation methods and reduce the complexity of
intensive pre- and post-processing, a fourth type of segmentation methods evolved, which consists
of learning-based methods [25]. Among all learning-based methods, deep learning via deep CNN
is the most ideal and popular in current computer vision applications because of its accuracy and
performance. This method has been applied to damaged road marks detection [26], human gender
recognition from human body images [27], and human detection in night environments using visible
light camera [28]. Considering segmentation, CNN can provide a powerful platform to simplify the
intensive task with accuracy and reliability similar to brain tumor segmentation done by CNN [29].
Symmetry 2017, 9, 263 3 of 25

Iris-related applications are sensitive because of the very dense and complex iris texture. Therefore,
few researchers focus on CNN related to iris segmentation. Liu et al. [30] used DeepIris to solve an
intra-class variation of heterogeneous iris images where CNNs are used to learn relational features
to measure the similarity between two candidate irises for verification purposes. Gangwar et al. [31]
used DeepIrisNet for iris visual representation and cross-sensor iris recognition. However, these two
types of research focus on iris recognition instead of iris segmentation.
Considering iris segmentation using CNNs, Liu et al. [32] identified accurate iris boundaries in
non-cooperative environments using fully convolutional networks (FCN). In their study, hierarchical
CNNs (HCNNs) and multi-scale FCNs (MFCNs) were used to locate the iris boundary automatically.
However, due to the use of full input image without the definition of region of interest (ROI) into the
CNN, the eyelids, hairs, eyebrows, and glasses frames, which look similar to iris, can be considered as
iris points by the CNN model. This scheme has better performance compared to previous methods.
However, the error of the iris segmentation can potentially be further reduced.
To address these issues concerning the existing approaches, we propose a two-stage iris
segmentation method based on CNN that is robust, to find the true boundary in less-constrained
environments. This study is novel in the following three ways compared to previous works.

- The proposed method accurately identified the true boundary even in intense scenarios, such as
glasses, off-angle eyes, rotated eyes, side view, and partially opened eyes.
- The first stage includes bottom-hat filtering, noise removal, Canny edge detector, contrast
enhancement, and modified HT to segment the approximate the iris boundary. In the second stage,
deep CNN with the image input of 21 × 21 pixels is used to fit the true iris boundary. By applying
the second stage segmentation only within the ROI defined by the approximate iris boundary
detected in the first stage, we can reduce the processing time and error of iris segmentation.
- To reduce the effect of bright SR in iris segmentation performance, the SR regions within the
image input to CNN are normalized by the average RGB value of the iris region. Furthermore, our
trained CNN models have been made publicly available through [33] to achieve fair comparisons
by other researchers.

Table 1 shows the comparison between existing methods and the proposed method.

Table 1. Comparisons between the proposed and previous methods on iris segmentation.

Type Methods Strength Weakness

Integro-differential operator [17]
Iris localization by circular HT.
Boundary-based Upper and lower eyelids detection
methods by parabolic curves [14].
Using two-circular edge detector
along with AdaBoost eye detector.
In addition, eyelash and eyelid
detection is performed [18].
Drawing 2D profile line on the
iris-sclera area and calculating the
gradient to find the boundary
Pixel-based points [20].
methods Initially approximating iris
boundary by color clustering, and
then fitting the boundary by two
side curves [21].
Active contour and Iterative method by starting with
circle fitting-based mask, checking the gradient, and
methods reaching the boundary [24].
Fast processing speed using a
simple technique.
As the first practical scheme, it can
produce a good approximation of the
iris region with image smoothness.
This method provides satisfactory
results in non-ideal situation, and a
closed eye can be detected as well.
A new way to estimate iris boundary
from both side gradient.
To reduce the error of iris
segmentation, the upper and lower
eyelids and the eyelashes are removed
by average intensity analysis.
For non-ideal cases, it can segment
the iris true boundary more accurately
than those by boundary- and
pixel-based methods.
These methods are less accurate for
iris segmentation in non-ideal
situations or visible light.
This method fails in iris segmentation
in case the RGB values of the eyelids
are similar to the iris or pupil/eyelid
detection error.
The calculated gradient is affected by
the eyelashes, and the true boundary
of the iris is not found.
Some empirical threshold is set for
eyelid and eyelash detection, but the
limitations in detecting the true
boundary of the iris still exist.
These methods are better for detecting
pupil boundary because they can be
prevented by the iris texture
considering the iris pattern as
the boundary.
 Symmetry 2017, 9, 263 4 of 25
Symmetry 2017, 9, 263 4 of 25
fitting-based gradient, and reaching the than those by boundary- and by the iris texture considering
methods boundary [24]. pixel-based methods. the iris pattern as the boundary.
As they use full image into the
Table 1. Cont.
This approach has better network, the eyelids, hairs,
Using HCNNs and MFCNs
accuracy relative to existing eyebrows, and glasses frames,
Type based deep learning method
Methods Strength Weakness
segmentation methods for which look similar to iris can be
[32]. As they use full image into the
non-ideal conditions. considered as iris points by the
CNN-based Using HCNNs and MFCNs based This approach has better accuracy network, the eyelids, hairs, eyebrows,
relative to existing segmentation CNN model.
and glasses frames, which look
methods deep learning method [32].
methods This
for approach simply finds similar to iris can be considered as iris
non-ideal conditions.
CNN-based
methods the rough boundary of the irispoints by the CNN model.
Two-stage iris segmentation A large amount of data is
This approach simplyCNN
and applies finds just
the rough
within
method iris
Two-stage (Proposed method) boundary of the iris and applies CNN A large
segmentation needed
amountforofCNN training.
data is needed for
the ROI defined based on the
method (Proposed method) just within the ROI defined based on CNN training.
rough boundary.
the rough boundary.

3. Proposed
3. Proposed Method
Method

3.1. 3.1. Overview

Overview of Proposed
of the the Proposed System
System
Figure
Figure 1 shows
1 shows an overall
an overall flowchart
flowchart of proposed
of the the proposed two-stage
two-stage iris segmentation
iris segmentation method.
method. In the
In the
firstfirst stage,
stage, thethe rough
rough irisiris boundary
boundary is is obtainedfrom
obtained fromthetheinput
inputimage
image to to define
define ROI
ROI for
for next
nextstage.
stage.The
Theresultant
resultantimage
imagefromfromStage
Stage1 includes
1 includes parts of of
parts thethe
upper
upperandandlower eyelids
lower and and
eyelids otherother
areas, such as
areas,
suchskin, eyelashes,
as skin, and and
eyelashes, sclera. Consequently,
sclera. Consequently,the true iris boundary
the true iris boundaryneeds to betoidentified.
needs In the
be identified. In second
the
stage, within this ROI (defined by Stage 1), CNN is applied, which can provide
second stage, within this ROI (defined by Stage 1), CNN is applied, which can provide the real iris the real iris boundary
with the
boundary help
with theof learned
help features.
of learned Considering
features. the standard
Considering information
the standard informationof the ratio
of the between
ratio between pupil
pupilcontraction
contraction andanddilation,
dilation,pupil
pupilapproximation
approximationisisperformed,
performed,and andfinally,
finally, the actual iris
the actual iris area
area can
can be
obtained.
be obtained.

Figure
Figure 1. Flowchart
1. Flowchart of proposed
of the the proposed method.
method.

3.2. Stage 1. Detection of Rough Iris Boundary by Modified Circular HT

An approximate localization of the iris boundary is the prerequisite of this study, and it is obtained
by modified circular HT-based method. As shown in Figure 1, Stage 1 consists of two phases, namely,
pre-processing and circular HT-based detection.
Symmetry 2017, 9, 263 5 of 25
Symmetry 2017, 9, 263 5 of 25
3.2. Stage 1. Detection of Rough Iris Boundary by Modified Circular HT
3.2. Stage 1. Detection of Rough Iris Boundary by Modified Circular HT
An2017,
Symmetry approximate
9, 263 localization of the iris boundary is the prerequisite of this study, and5 of it 25
is
An approximate
obtained localization
by modified circular of themethod.
HT-based iris boundary
As shown is the prerequisite
in Figure 1, Stage of this study,
1 consists of twoand it is
phases,
obtained by modified circular
namely, pre-processing HT-based
and circular method.
HT-based As shown in Figure 1, Stage 1 consists of two phases,
detection.
3.2.1. Phase 1. Pre-Processing
namely, pre-processing and circular HT-based detection.
3.2.1.InPhase
Phase 1. 1,
Pre-Processing
as shown in Figures 2 and 3c, the RGB input image is converted into grayscale
3.2.1. Phase
for further 1. Pre-Processing
processing, andinmorphological
In Phase 1, as shown Figures 2 and operation
3c, the RGB is input
applied through
image bottom-hat
is converted intofiltering
grayscalewith
for
symmetrical
In Phase
further structuring
processing, and element
1, as shown disk2ofand
in Figures
morphological size 5 for
3c, the contrast
operation RGB inputenhancement.
is applied image Finally, two
is converted
through images
intofiltering
bottom-hat ofwith
grayscale the
for
gray
furtherimage
symmetrical and
processing,resultant
structuring image by
and element bottom-hat
morphological
disk filtering
operation
of size 5 for areenhancement.
is applied
contrast added to obtain
through an enhanced
bottom-hat
Finally, image
twofiltering
images as
ofwith
the
shown in
symmetrical Figure
gray image and 3d.
structuring
resultantelement
image bydiskbottom-hat
of size 5 for contrast
filtering areenhancement. Finally,
added to obtain two images
an enhanced of the
image as
gray
shown image and resultant
in Figure 3d. image by bottom-hat filtering are added to obtain an enhanced image as
shown in Figure 3d.

Figure 2. Overall pre-processing procedure of Phase 1.

Figure 2. Overall pre-processing procedure of Phase 1.
Figure 2. Overall pre-processing procedure of Phase 1.

(a) (b)
(a) (b)

(c) (d)
(c) (d)
Figure 3. Resultant images by Phase 1 of Figure 1. (a) Original input image of 400 × 300 pixels; (b)
Figure 3. Resultant
grayscale images by Phase 1 ofimage
Figureby1.bottom-hat
(a) Originalfiltering;
input image of 400 × 300 pixels; (b)
Figure 3. converted
Resultant image;
images(c)byresultant
Phase 1 of Figure 1. (a) Original input (d)image
resultant image
of 400 by adding
× 300 pixels;
grayscale
the two converted
images of image;
(b) and (c) resultant image by bottom-hat filtering; (d) resultant image by adding
(c).
(b) grayscale converted image; (c) resultant image by bottom-hat filtering; (d) resultant image by
the two images of (b) and (c).
adding the two images of (b,c).
Symmetry 2017, 9, 263 6 of 25

3.2.2. Phase
Symmetry 2017,2.
9, ROI
263 Detection by Rough Iris Boundary 6 of 25

The overall process of Phase 2 is presented by the flowchart in Figure 4. In Phase 2, the filtered
3.2.2. Phase 2. ROI Detection by Rough Iris Boundary
image from Phase 1 is redefined as an image of 280 × 220 pixels to reduce the effect of eyebrows
The overall
in detecting process
the iris of Phase
boundary. 2 is presented
Then, by the flowchart
a 17 × 17 median in Figure
filter is applied to 4. In Phase
smooth the2,image
the filtered
from
image from Phase 1 is redefined as an image of 280 × 220 pixels to reduce
salt and pepper noises and reduce the skin tone and texture illumination as shown in Figure 5c. the effect of eyebrows in
detecting
13 the iris boundary.
× 13 symmetrical Gaussian Then, a 17 ×filter
smoothing 17 median
with σ of filter
2 isisapplied
appliedtotothesmooth
filtered the image
image tofrom salt
increase
pixel uniformity as shown in Figure 5d. Then, Canny edge detector with same σ value is used 13
and pepper noises and reduce the skin tone and texture illumination as shown in Figure 5c. 13 × to
symmetrical
detect the edgesGaussian
of the smoothing
iris boundary filter
as with
shown σ ofin 2Figure
is applied to the filtered
5e. However, imageare
the edges to not
increase pixel
very clear,
uniformity
and gammaasadjustment
shown in Figure
with γ5d. = Then,
1.90 isCanny
applied edge detector the
to enhance withcontrast
same σ of valuethe is used as
image to detect
shownthe in
edges of the iris boundary as shown in Figure 5e. However, the edges are
Figure 5f. With the gamma-enhanced image, the binarized edge image is obtained with eight neighbor not very clear, and gamma
adjustment with
connectivity 𝛾 =1.90
as shown in is applied
Figure 5g. to
Inenhance
this edge the contrast
image, of the
there are image as shown
more circular in Figure
edges along5f. theWith
iris
the gamma-enhanced image, the binarized edge image is obtained with
boundary edges, and circular HT can find all possible circles in the image. However, the incorrect eight neighbor connectivity
as shown inedges
circle-type Figureas5g. In thisinedge
shown image,
Figure 5h canthere
beare more circular
removed edgesthe
by filtering alongedgesthe whose
iris boundary
radius edges,
is out
and
of thecircular
range of HTthecan find all possible
minimum and maximumcircles inhuman
the image. However,
iris radius. thethe
Then, incorrect circle-typeedges
most-connected edgesareas
shown in Figure 5h can be removed by filtering the edges whose radius is
selected as iris edges, and the rough iris boundary is detected in Figure 5i. Considering the possibility out of the range of the
minimum
of detection and maximum
error human
of the iris iris radius.
boundary, ROI is Then, the most-connected
defined slightly larger than edges the are selectedboundary
detected as iris edges,
as
and theinrough
shown Figure iris
7a.boundary is detected in Figure 5i. Considering the possibility of detection error of
the iris
Forboundary, ROI is defined
fair comparisons, all theslightly
optimallarger than the
parameters fordetected boundary
the operation in ROIas shown in Figure
detection 7a.
including
medianForfilter,
fair comparisons, all the optimal
Gaussian smoothing parameters
filter, Canny for the operation
edge detector, in ROI detection
gamma adjustment, including
and binarization,
median filter, Gaussian smoothing filter, Canny edge detector,
etc., were empirically selected only by training data without testing data. gamma adjustment, and binarization,
etc., were empirically selected only by training data without testing data.

Figure 4.
Figure 4. Overall
Overall region
region of
of interest
interest (ROI)
(ROI) detection
detection procedure
procedure of
of Phase
Phase 2.
2.
Symmetry 2017, 9, 263 7 of 25
Symmetry 2017, 9, 263 7 of 25

(a) (b)

(c) (d)

(e) (f)

(g) (h)

(i)
Figure 5. Resultant images of rough iris boundary detected by Phase 2 of Figure 1; (a) pre-processed
Figure 5. Resultant images of rough iris boundary detected by Phase 2 of Figure 1; (a) pre-processed
image of 400 × 300 pixels from Phase 1; (b) redefined image of 280 × 220 pixels; (c) median filtered
image of 400 × 300 pixels from Phase 1; (b) redefined image of 280 × 220 pixels; (c) median filtered
image; (d) image by Gaussian smoothing; (e) image after Canny edge detector; (f) image after gamma
image; (d) image by Gaussian smoothing; (e) image after Canny edge detector; (f) image after gamma
contrast adjustment; (g) binarized edge image; (h) resultant image by removing incorrect circles by
contrast adjustment; (g) binarized edge image; (h) resultant image by removing incorrect circles by
radius (i) final image with rough iris boundary.
radius (i) final image with rough iris boundary.
Symmetry 2017, 9, 263 8 of 25

Symmetry 2017, 9, 263 8 of 25

3.3. Iris Rough Boundary Analysis
3.3. Iris Rough Boundary Analysis
The rough iris boundary detected in Stage 1 is not the real iris boundary, but an approximation
The rough
for the next stage. iris boundary ideal
Considering detected
casesin in
Stage 1 isuser
which not the real iris boundary,
cooperation butthe
is available, an output
approximation
of the HT
for the next stage. Considering ideal cases in which user cooperation is available,
can show the accurate iris boundary. However, for non-ideal cases, such as off-angle eyes, rotatedthe output of the
eyes,
HT can show the accurate iris boundary. However, for non-ideal cases, such as off-angle
eyelash occlusions, and semi-open eyes, the HT can sometimes produce inaccurate iris boundaries eyes, rotated
as eyes,
showneyelash occlusions,
in Figure and semi-open
6. Moreover, the detected eyes,
iris the HTincludes
circle can sometimes produce
the eyelid, eyelash,inaccurate
pupil, andirisSR,
boundaries as shown in Figure 6. Moreover, the detected iris circle includes the eyelid, eyelash, pupil,
which should be discriminated from the true iris area for iris recognition. Therefore, we proposed
and SR, which should be discriminated from the true iris area for iris recognition. Therefore, we
a CNN-based segmentation method of the iris region based on the ROI defined by the rough iris
proposed a CNN-based segmentation method of the iris region based on the ROI defined by the
boundary in Stage 1.
rough iris boundary in Stage 1.

Figure 6. Examples of rough iris boundaries detected by Stage 1 for non-ideal environments.
Figure 6. Examples of rough iris boundaries detected by Stage 1 for non-ideal environments.
Symmetry 2017, 9, 263 9 of 25

Symmetry 2017, 9, 263 9 of 25

3.4. CNN-Based Detection of the Iris Region
3.4. CNN-Based Detection of the Iris Region
3.4.1. Extracting the Mask for CNN Input
3.4.1. Extracting the Mask for CNN Input
To detect the iris region accurately, the square mask of 21 × 21 pixels is extracted from the ROI of
Figure 7a Toand is used
detect asregion
the iris input accurately,
to CNN. The themask
squareismask
extracted
of 21 within the is
× 21 pixels ROI to reduce
extracted fromthe
thenumber
ROI of
of Figure
objects to be7a and is used
classified. as input toin
Specifically, CNN.
many The maskiris
cases, is extracted
color canwithin the ROI
be similar to reduce
to the the number
eyebrows and eyelids.
of objects toin
Furthermore, benon-ideal
classified. cases,
Specifically, in many
the skin cases,similar
can have iris color cantobeiris.
color similar to the eyebrows
Therefore, and the
by extracting
eyelids. Furthermore, in non-ideal cases, the skin can have similar color
mask only within the ROI, we can reduce the iris segmentation error by CNN. This mask is scannedto iris. Therefore, by in
extracting the mask only within the ROI, we can reduce the iris segmentation error
both horizontal and vertical directions as shown in Figure 7b. Based on the output of CNN, the center by CNN. This
mask is scanned in both horizontal and vertical directions as shown in Figure 7b. Based on the output
position of the mask is determined as an iris or non-iris pixel. Figure 7c shows the examples of the
of CNN, the center position of the mask is determined as an iris or non-iris pixel. Figure 7c shows the
collected masks of 21 × 21 pixels for CNN training or testing. As shown in Figure 7c, the mask from
examples of the collected masks of 21 × 21 pixels for CNN training or testing. As shown in Figure 7c,
thethe
irismask
region hasthe
from theiris
characteristics
region has thewhere most pixels
characteristics where of most
the mask
pixelsare
of from
the maskthe iris texture,
are from the whereas
iris
thattexture,
from the non-iris region has the characteristics where most pixels are from the
whereas that from the non-iris region has the characteristics where most pixels are from theskin, eyelid, eyelash,
or sclera.
skin, eyelid, eyelash, or sclera.

(a) (b)

(c)
Figure 7. Extracting the mask of 21 × 21 pixels for training and testing of CNN. (a) ROI defined from
Figure 7. Extracting the mask of 21 × 21 pixels for training and testing of CNN. (a) ROI defined
Stage 1 of Figure 1; (b) extracting the mask of 21 × 21 pixels within the ROI; (c) examples of the
from Stage 1 of Figure 1; (b) extracting the mask of 21 × 21 pixels within the ROI; (c) examples of the
extracted masks (the 1st (4th) row image shows the masks from the boundary between the upper
extracted masks (the 1st (4th) row image shows the masks from the boundary between the upper (lower)
(lower) eyelid and iris, whereas the others represent those from the left boundary between iris and
eyelid
sclera). iris, whereas the others represent those from the left boundary between iris and sclera).
and

However,
However, if the
if the mask
mask is isextracted
extractedfrom
fromthe
thebright
bright SR
SR region,
region, the
the characteristics
characteristicsofofthe
thepixels
pixelsofof the
thecan
mask maskbecan be changed
changed as shown
as shown in Figure
in Figure 8a, which
8a, which cancan increase
increase thethe error
error of of
irisiris segmentationby
segmentation byCNN.
CNN. To solve this problem, if the bright pixels whose gray level is higher than 245 exist
To solve this problem, if the bright pixels whose gray level is higher than 245 exist inside the mask, inside the
mask, they are replaced by the average RGB value of the iris area (the red dotted box in Figure 8a) as
they are replaced by the average RGB value of the iris area (the red dotted box in Figure 8a) as shown
shown in Figure 8b. The threshold of 245 was experimentally found with training data. Here, we can
in Figure 8b. The threshold of 245 was experimentally found with training data. Here, we can reduce
the effect of SR on the iris segmentation by CNN. The position of the red dotted box is determined
Symmetry 2017, 9, 263 10 of 25
Symmetry 2017, 9, 263 10 of 25

reduce the effect of SR on the iris segmentation by CNN. The position of the red dotted box is
based on the rough iris boundary detected in Stage 1. If the dark pixel exists inside the box because the
determined based on the rough iris boundary detected in Stage 1. If the dark pixel exists inside the
darkboxpupil can be
because theincluded
dark pupil incan
thebebox, the box
included position
in the is adaptively
box, the box positionmoved into the
is adaptively lower
moved direction
into the
untillower
it does not include
direction until itthe
doesdark pixel, and
not include the the
darkaverage RGB
pixel, and thevalue is extracted
average RGB valuein this box. in this
is extracted
box.

(a) (b)
Figure 8. Example of replacing specular reflection (SR) pixels by the average RGB value of the iris
Figure 8. Example of replacing specular reflection (SR) pixels by the average RGB value of the iris
region. (a) The mask including the SR region; (b) the mask where SR pixels are replaced.
region. (a) The mask including the SR region; (b) the mask where SR pixels are replaced.
3.4.2. Iris Segmentation by CNN
3.4.2. Iris Segmentation by CNN
The mask of 21 × 21 pixels is used as input to CNN, and the output of CNN is either iris or non-
The
iris mask
area. of 21
In this × 21
study, pixels is used
a pre-trained as input
VGG-face modelto [34]
CNN, and by
is used thefine-tuning
output ofwith CNNtheishelp
either iris or
of our
training
non-iris area.images.
In thisThe VGG-face
study, model was
a pre-trained pre-trained
VGG-face with[34]
model about 2.6 million
is used face images
by fine-tuning of 2,622
with the help
different
of our trainingpeople. Detailed
images. explanations
The VGG-face of thewas
model configuration
pre-trained of VGG-face
with about are2.6
shown in Table
million 2 and of
face images
Figure 9. To obtain an accurate boundary and its difference from other objects, the
2,622 different people. Detailed explanations of the configuration of VGG-face are shown in Table 2ROI is selected
and with slightly
Figure 9. To increased
obtain anrough iris boundary
accurate boundarydetected by HT.
and its difference from other objects, the ROI is selected
with slightly increased rough iris boundary detected by HT.
Table 2. Configuration of the VGG-face model used in the proposed method.

Configuration No.
of the of
VGG-face model used Kernel No. of No. of
Table
Type of2.
Layer Feature Map Sizein the proposed method.
Filter Size Stride Padding
224 × 224 × 3 (height
Image input layer No. of No. of
Type of Layer No. of Filter × width
Feature × channel)
Map Size Kernel Size
Stride Padding
Conv-1
64 224××3224
224 × 224 × 64 ×
(height 3×3 1×1 1×1
Image (1st convolutional
input layer layer)
width × channel)
Relu-1 224 × 224 × 64
Group 1 Conv-1
Conv-2 64 64 × 224
224224 × 64
× 224 × 64 33 ××33 1 1× ×
1 1 11××1 1
(1st convolutional layer)
(2nd convolutional layer)
Relu-1 224 × 224 × 64
Group 1 Relu-2 224 × 224 × 64
Conv-2
Pool-1 1
64 224112 × 112
× 224 × 64
× 64 23 ××23 2 1× ×
2 1 01××0 1
(2nd convolutional layer)
Conv-3
Relu-2 × 224
224112 × ×64128
128 × 112 3×3 1×1 1×1
(3rd convolutional
Pool-1 layer) 1 112 × 112 × 64 2×2 2×2 0×0
Relu-3 112 × 112 × 128
Conv-3
Group 2 Conv-4 layer) 128 112 × 112 × 128 3×3 1×1 1×1
(3rd convolutional 128 112 × 112 × 128 3×3 1×1 1×1
(4th convolutional
Relu-3 layer) 112 × 112 × 128
Group 2 Relu-4 112 × 112 × 128
Conv-4
Pool-2 layer) 128 1 112 × × 56××128
56112 128 23 ××23 2 1× ×
2 1 01××0 1
(4th convolutional
Conv-5
Relu-4 112 × 112 × 128
256 56 × 56 × 256 3×3 1×1 1×1
(5th convolutional
Pool-2 layer) 1 56 × 56 × 128 2×2 2×2 0×0
Group 3 Relu-5
Conv-5 56 × 56 × 256
256 56 × 56 × 256 3×3 1×1 1×1
Conv-6layer)
(5th convolutional
256 56 × 56 × 256 3×3 1×1 1×1
Relu-5
(6th convolutional layer) 56 × 56 × 256
Group 3
Conv-6
256 56 × 56 × 256 3×3 1×1 1×1
(6th convolutional layer)
Relu-6 56 × 56 × 256
Symmetry 2017, 9, 263 11 of 25

Table 2. Cont.
Symmetry 2017, 9, 263 11 of 25
No. of No. of
Type of Layer No. of Filter Feature Map Size Kernel Size
Stride Padding
Relu-6 56 × 56 × 256
Conv-7
Conv-7 256 256 56 ×5656× ×
56256
× 256 33 ××33 1 ×1 1× 1 1 1× ×
1 1
(7th (7th
convolutional layer)
convolutional layer)
Relu-7 56 × 56 × 256
Relu-7 56 × 56 × 256
Pool-3 1 28 × 28 × 256 2×2 2×2 0×0
Pool-3 1 28 × 28 × 256 2×2 2×2 0×0
Conv-8
Conv-8 512 28 × 28 × 512 3×3 1×1 1×1
(8th convolutional layer) 512 28 × 28 × 512 3×3 1×1 1×1
(8th convolutional layer)
Relu-8 28 × 28 × 512
Relu-8 28 × 28 × 512
Conv-9
Conv-9 512 28 × 28 × 512 3×3 1×1 1×1
Group 4 (9th convolutional layer) 512 28 × 28 × 512 3×3 1×1 1×1
(9th convolutional layer)
Group 4 Relu-9 28 × 28 × 512
Relu-9 28 × 28 × 512
Conv-10
Conv-10 512 512 28 ×2828× ×
28512
× 512 33 ××33 1×1
1×1 1 1× ×
1 1
(10th convolutional layer)
(10th convolutional layer)
Relu-10 28 × 28 × 512
Relu-10
Pool-4 1 28 × 28
14 × 14 × 512× 512 2×2 2×2 0×0
Pool-4 1 14 × 14 × 512 2×2 2×2 0×0
Conv-11
Conv-11 512 14 × 14 × 512 3×3 1×1 1×1
(11th convolutional layer) 512 14 × 14 × 512 3×3 1×1 1×1
(11thRelu-11
convolutional layer) 14 × 14 × 512
Relu-11 14 × 14 × 512
Conv-12
Conv-12 512 14 × 14 × 512 3×3 1×1 1×1
Group 5 (12th convolutional layer) 512 14 × 14 × 512 3×3 1×1 1×1
(12th convolutional layer)
Group 5 Relu-12 14 × 14 × 512
Relu-12 14 × 14 × 512
Conv-13
Conv-13 512 14 × 14 × 512 3×3 1×1 1×1
(13th convolutional layer) 512 14 × 14 × 512 3×3 1×1 1×1
(13th convolutional layer)
Relu-13 14 × 14 × 512
Relu-13
Pool-5 1 7 ×147 ××14 × 512
512 2×2 2×2 0×0
Pool-5 1 7 × 7 × 512 2×2 2×2 0×0
Fully connected layer-6 (1st FCL) 4096 × 1
Fully connected
Relu-6 layer-6 (1st FCL) 40964096
× 1× 1
Relu-6
Dropout-6 4096 × 1× 1
4096
Dropout-6 4096 × 1
Fully connected layer-7 (2nd FCL) 4096 × 1
Fully connected layer-7 (2nd FCL) 4096 × 1
Relu-7 4096 × 1
Relu-7
Dropout-7 40964096
× 1× 1
Dropout-7 4096 × 1
Fully connected layer-8 (3rd FCL) 2×1
Fully connected layer-8 (3rd FCL) 2×1
Softmax layer 2×1
Softmax layer 2×1
Output layer 2×1
Output layer 2×1

Figure 9. CNN architecture used in the proposed method.

Figure 9. CNN architecture used in the proposed method.
Symmetry 2017, 9, 263 12 of 25

The VGG-face model consists of 13 convolutional layers and 5 pooling layers in combination with
3 fully connected layers. The filter size, rectified linear unit (Relu), padding, pooling, and stride are
explained in Table 2 and Figure 9. A total of 64 3 × 3 size filters are adopted in the 1st convolutional
layer. Therefore, the size of the feature map is 224 × 224 × 64 in the 1st convolutional layer. Here,
224 and 224 denote the height and width of the feature map, respectively. They are calculated based
on (output height (or width) = (input height (or width) − filter height (or width) + 2 × the number
of padding)/the number of stride + 1 [35]). For example, in the image input layer and Conv-1 in
Table 2, the input height, filter height, number of padding, and number of strides are 224, 3, 1, and 1,
respectively. As a result, the height of the output feature map becomes 224 (= (224 − 3 + 2 × 1)/1 + 1).
There are three common activation functions, such as sigmoid, tanh function, and Relu [35].
The sigmoid-based activation function forces the candidate input value between the range of 0 and 1
as shown in Equation (1), which means that for the negative inputs, the output becomes zero, whereas
it becomes 1 for large positive inputs.
1
σ( x) = (1)
1 + e− x
The tanh function is slightly different from the sigmoid activation function because it keeps fitting
the input value in the range of −1 and 1 as shown in Equation (2).

e x − e− x
tan h( x ) = (2)
e x + e− x

Relu performs faster compared to these two nonlinear activation functions, and it is useful to
wipe off the gradient problem in the back propagation at the time of training [36,37]. In [38], they also
showed that the speed of training by Relu with the CIFAR-10 dataset based on the four-layered CNN
is six times faster than the tanh function with same dataset and network. Therefore, we use Relu for
faster training with simplicity and to avoid the gradient issue in CNN. The Relu was initially used in
Boltzmann machines, and it is formulated as follows.
(
xi , xi ≥ 0
yi = , (3)
0, xi < 0

where yi and xi are the corresponding outputs and inputs of the unit, respectively. The Relu layer
exists after each convolutional layer, and it maintains the size of each feature map. Max-pooling
layers can provide a kind of subsampling. Considering pool-1, which performs max pooling
after the convolutional layer-2 and Relu-2, the feature map of 224 × 224 × 64 is reduced to that
of 112 × 112 × 64. In the case that the max-pooling layer (pool-1) is executed, the input feature map
size is 224 × 224 × 64, the filter size is 2 × 2, and the number of strides is 2 × 2. Here, 2 × 2 for the
number of strides denotes the max-pooling filter of 2 × 2 where the filter moves by two pixels in both
the horizontal and vertical directions. Owing to the lack of overlapped area due to filter movement,
the feature map size is reduced to 1/4 (1/2 horizontally and 1/2 vertically). Consequently, the feature
map size after passing pool-1 becomes 112 × 112 × 64 pixels. This pooling layer is used after Relu-2,
Relu-4, Relu-7, Relu-10, and Relu-13 as shown in Table 2. For all cases, the filter of 2 × 2 and stride of
2 × 2 are used, and the feature map size diminishes to 1/4 (1/2 horizontally and 1/2 vertically).
When an input image of 224 × 224 × 3 is given to the CNN shown in Figure 9, it passes through
13 convolutional layers (Conv-1 to Conv-13), 13 Relu functions, and 5 pooling layers (pool-1 to
pool-5). Therefore, after Group 5 (Conv-11, Relu-11, Conv-12, Relu-12, Conv-13, Relu-13, and pool-5),
the feature-map of 7 × 7 × 512 is obtained. Then, this feature-map passes through three fully connected
layers (fully connected layer-6 to fully connected layer-8), from which two outputs can be obtained.
These two final outputs correspond to the two classes of iris and non-iris regions. The number of
Symmetry 2017, 9, 263 13 of 25

output nodes of the 1st, 2nd, and 3rd FCLs are 4096, 4096, and 2, respectively. After the fully connected
layer-8, the softmax function is used as follows [39].

eqj
σ( q ) j = (4)
∑Sn=1 eqn

In Equation (4), given that the array of the output neurons is q, the probability of neurons of the
jth class can be obtained by dividing the value of the jth component by the summation of the values of
all the components.
CNN usually has an over-fitting issue, which can reduce the testing accuracy. Therefore, to solve
this issue, the dropout method and data augmentation have been considered. Dropout methods are
important during training to avoid the same type of neurons representing the same feature repeatedly.
Which can cause overfitting, wastage of network capacity, and computational resources. Therefore,
the solution is to drop out these cases at random by using the dropout ratio to obtain specific benefits,
such as the dropped-out neurons’ contribution in forward or back propagation. Consequently, it can
reduce co-adoption during the process by decreasing dependencies over other neurons and it can force
the network to learn strong features with different neurons [38,40]. In this study, we use the dropout
ratio of 0.5. The dropout layer was used twice, that is, after the 1st FCL with Relu-6 and after the 2nd
FCL and Relu-7, as shown in Table 2.
Compared to the elements and parameters of the original VGG-face model, three parts of initial
learning rate, the momentum value, and the size of the mini-batch were modified, and their optimal
values were experimentally found with training data. The detailed explanations of these values are
included in Section 4.2. In addition, the last part modified is the number of CNN output as 2 as shown
in Figure 9 because the number of classes in our research is 2 as iris and non-iris pixels.

3.5. Pupil Approximation by the Information of the Ratio of Pupil Contraction and Dilation
When applying CNN, false positive errors (non-iris pixel is incorrectly classified into iris pixel)
exist in the pupil area. However, the noisy iris challenge evaluation part -II (NICE-II) database used
in our experiment includes inferior quality images, where it is very difficult to segment pupillary
boundary accurately. Therefore, we used anthropometric information provided by Wyatt [41].

Pd
dp = (5)
Id

where Pd represents the pupil diameter and Id shows the iris diameter. This anthropometric
information [41] provides the details of pupil dilation ratio (dp ) in Equation (5) that varies from 12% to
60%, and we use the minimum value in our experiment. Specifically, the iris pixel that is extracted by
CNN belongs to the region whose dp is less than 12% is determined as non-iris pixel (pupil).

4. Experimental Results

4.1. Experimental Data and Environment

In this study, we used the NICE-II training database, which was used for NICE-II competition [42].
The database includes extremely noisy iris data of the UBIRIS.v2. This database contains severely
noisy 1000 image of 171 classes. The size of the image is 400 × 300 pixels. The images of the iris
were acquired from people walking 4–8 m away from a high-resolution visible light camera with
visible light illumination [43]. Therefore, this database includes the difficulties, such as poorly focused,
off-angle, rotated, motion blur, eyelash obstruction, eyelids occlusions, glasses obstructions, irregular
SR, non-uniform lighting reflections, and partial captured iris images as shown in Figure 10.
[42]. The database includes extremely noisy iris data of the UBIRIS.v2. This database contains
severely noisy 1000 image of 171 classes. The size of the image is 400 × 300 pixels. The images of the
iris were acquired from people walking 4–8 m away from a high-resolution visible light camera with
visible light illumination [43]. Therefore, this database includes the difficulties, such as poorly
focused,2017,
Symmetry off-angle,
9, 263 rotated, motion blur, eyelash obstruction, eyelids occlusions, glasses obstructions,
14 of 25
irregular SR, non-uniform lighting reflections, and partial captured iris images as shown in Figure 10

Symmetry 2017, 9, 263 14 of 25

Figure 10. Examples of noisy iris images of NICE-II database.

Figure 10. Examples of noisy iris images of NICE-II database.

In this study, among the total 1000 iris images, 500 iris images are used for training, and the
otherIn500 thisiris
study, among
images aretheusedtotalfor
1000testing
iris images,
purposes. 500 iris
Theimages
averageare used
valueforoftraining, and the other
two accuracies was
500 iris images
measured are usedcross
by two-fold for testing purposes.
validation. CNN The average
training andvalue
testing ofare
twoperformed
accuracieson was measured
a system by
using
two-fold cross validation. CNN training and testing are performed on a system using Intel ® Core™
Intel Core™ i7-3770K CPU @ 3.50 GHz (4 cores) with 28 GB of RAM, and NVIDIA GeForce GTX
®

i7-3770K
1070 (1920 CPUCuda @ 3.50 GHz
cores) with (4 cores)
graphicswith 28 GB ofofRAM,
memory 8 GB and NVIDIA
(NVIDIA, GeForce
Santa Clara,GTXCA,1070
USA) (1920
[44].Cuda
The
cores)
training withandgraphics memory
testing are done with of 8 GB (NVIDIA,
Windows Caffe Santa Clara,1)CA,
(version [45].USA) [44]. The training and testing
are done with Windows Caffe (version 1) [45].
4.2. CNN Training
4.2. CNN Training
In our proposed method, we fine-tuned the VGG-face model using 500 iris images for
In our proposed method, we fine-tuned the VGG-face model using 500 iris images for classification
classification of two classes (iris and non-iris). For two-fold cross validation, we used the first 500
of two classes (iris and non-iris). For two-fold cross validation, we used the first 500 images and the
images and the other 500 images for testing. From the first 500 iris images for training, we have 9.6
other
million 500 imagesoffor
images 21 testing.
× 21 pixels From the first
from 500ROI
the iris iris images for training,
for training, and in we thehave
second 9.6 training
million images of
for cross
21 × 21 pixels from the iris ROI for training, and in the second training for cross validation, we have
validation, we have 8.9 million images of 21 × 21 pixels from the iris ROI. During the CNN training,
8.9 milliongradient
stochastic 21 × 21(SGD)
images ofdescent pixelsisfromused thetoiris ROI. During
minimize the CNN between
the difference training, thestochastic
calculatedgradient
and
descent (SGD) is used to minimize the difference between the calculated and desired outputs with the
desired outputs with the help of the gradient derivative [46]. The number of the whole training set
help
dividedof theby gradient
a mini-batchderivative
size is [46].
definedTheas number
iteration. of the
Thewhole training
total time taken setfor
divided by a mini-batch
the complete training
size is defined as iteration. The total time taken for the complete training
including all the iterations is set as 1 epoch, and the training was executed several times as per a pre- including all the iterations
is set as 1 epoch,
determined epoch.and Inthe
thistraining
study, CNNwas executed
was trained severalby times as perFor
10 epochs. a pre-determined
the fine-tuning epoch. In this
of VGG-face,
study, CNN was trained by 10 epochs. For the fine-tuning of VGG-face,
the optimum fine-tuning model was experimentally found based on the optimal parameters of initial the optimum fine-tuning
model
learning wasrateexperimentally
of 0.00005, the found momentum based value
on theofoptimal
0.9, andparameters
the size of theof initial
mini-batchlearning rate
of 20. The of detailed
0.00005,
the momentum value of 0.9, and the size of the mini-batch of 20.
explanations of these parameters can be found in the following literature [47]. Figure 11a,b show The detailed explanations of these
the
parameters can be found in the following literature [47]. Figure
curves of average loss and accuracy with training data in case of two-fold cross validation. Epoch 11a,b show the curves of average
loss
count and accuracy with
is represented ontraining
the X-axis,data in case the
whereas of two-fold
right Y-axis crossrepresents
validation. theEpoch count
training is represented
accuracy and the
on the X-axis, whereas the right Y-axis represents the training accuracy
left Y-axis shows the training loss. Depending on the learning rate and mini-batch size, the loss and the left Y-axis shows the
varies.
training loss. Depending on the learning rate and mini-batch
While training, it is important to reach the minimum training loss (maximum training accuracy);size, the loss varies. While training, it
is important
therefore, to reach
learning theshould
rate minimum trainingcarefully.
be decided loss (maximum The losstraining accuracy);dramatically
value decreased therefore, learning
with a
rate
higher value of the learning rate, which can deteriorate the loss value with reaching to avalue
should be decided carefully. The loss value decreased dramatically with a higher of the
minimum.
learning rate, which
In our proposed can deteriorate
method, the finest modelthe loss value the
in which with reaching
training losstocurve
a minimum.
converges Intoour
0%proposed
(training
method, the finest model in which the training loss curve converges
accuracy of 100%) is used for testing as shown in Figure 11. To have fair comparisons by to 0% (training accuracy of 100%)
other
is used for testing as shown in Figure 11. To have fair comparisons
researchers, we have made our trained CNN models publicly available through [33]. Figure 12 shows by other researchers, we have made
our trained CNN
the examples models
of trained publicly
filters of 1st available
convolutional through layer[33]. Figure
of Table 12case
2 in shows the examples
of two-fold of trained
cross validation.
filters of 1st convolutional layer of Table 2 in case of two-fold cross validation.
 Symmetry2017,
Symmetry 2017,9,9,263
263 15
15ofof25
25
Symmetry 2017, 9, 263 15 of 25

(a)
(a)

(b)
(b)
Figure 11. Average loss and accuracy curves for training. Average loss and accuracy curve from (a)
Figure 11. Average loss and accuracy curves for training. Average loss and accuracy curve from (a) 1st
Figure 11.cross
1st fold Average loss and
validation; andaccuracy curves
(b) 2nd fold crossfor training. Average loss and accuracy curve from (a)
validation.
fold cross validation; and (b) 2nd fold cross validation.
1st fold cross validation; and (b) 2nd fold cross validation.

(a) (b)
Figure 12. Examples of trained filters of the 1st convolutional layer of Table 2. (a) 1st fold cross
(a) (b)
validation; (b) 2nd fold cross validation.
Figure 12. Examples of trained filters of the 1st convolutional layer of Table 2. (a) 1st fold cross
Figure 12. Examples of trained filters of the 1st convolutional layer of Table 2. (a) 1st fold cross
4.3.validation;
Testing of (b)
the 2nd
Proposed CNN-Based
fold cross Iris Segmentation
validation.
validation; (b) 2nd fold cross validation.
The performance of the proposed method for iris segmentation is evaluated based on the metrics
4.3. Testing of thecompetition
Proposed CNN-Based Iris Segmentation
4.3. of the NICE-I
Testing of the Proposed to compare
CNN-Based Iristhe accuracy with that of the teams participating in NICE-I
Segmentation
competition
The performance of the proposed methodthe
[48]. As shown in Equation (6), forclassification error (𝐸
iris segmentation ) is measured
is 𝑖evaluated by on
based comparing
the metrics
The performance of the proposed method for iris segmentation is evaluated based on the metrics
of the NICE-I competition to compare the accuracy with that of the teams participating in NICE-I
of the NICE-I competition to compare the accuracy with that of the teams participating in NICE-I
competition [48]. As shown in Equation (6), the classification error (𝐸𝑖 ) is measured by comparing
competition [48]. As shown in Equation (6), the classification error ( Ei ) is measured by comparing the
Symmetry 2017, 9, 263 16 of 25

resultant image (Ii (m0 , n0 )) by our proposed method and the ground truth image (G (m0 , n0 )) with help
Symmetry
of the 2017, 9, 263(XOR) operation.
exclusive-OR 16 of 25

the resultant image (𝐼𝑖 (𝑚′ , 𝑛′ )) by our

1 proposed method
truth image (𝐺(𝑚′ , 𝑛′ )) with

and the ground
help of the exclusive-OR (XOR) Ei =operation.∑ Ii m , n ⊗ Gi m , n
0 0 0 0
(6)
m×n (m0 ,n0 )
1
𝐸𝑖 = ∑ 𝐼𝑖 (𝑚′ , 𝑛′ )⨂ 𝐺𝑖 (𝑚′ , 𝑛′ ) (6)
𝑚 of
where m and n are the width and height × 𝑛the ′image,
′
respectively. To evaluate the proposed method,
(𝑚 ,𝑛 )
the average segmentation error (E1) is calculated by averaging the classification error rate ( Ei ) of the
whereas𝑚shown
whole images and 𝑛inare the width
Equation (7).and height of the image, respectively. To evaluate the proposed
method, the average segmentation error (E1) is calculated by averaging the classification error rate
(𝐸𝑖 ) of the whole images as shown in Equation (7).
1
k 1∑
E1 = Ei (7)
i
𝐸1 = ∑ 𝐸𝑖 (7)
𝑘
where k represents the total number of testing images. 𝑖E1 varies between [0,1], which denotes that “0”
where kthe
represents represents the and
least error total“1”
number of testing
represents theimages.
largest E1 varies between [0,1], which denotes that “0”
error.
represents the least error and “1” represents the largest error.
4.3.1. Iris Segmentation Results by the Proposed Method
4.3.1. Iris Segmentation Results by the Proposed Method
Figure 13 shows the examples of good segmentation results by our proposed method. In our
experiment, Figurewe13canshows the examples
consider two typesof good segmentation
of error, results
such as false by our and
positive proposed
false method.
negativeInerrors.
our
experiment, we can consider two types of error, such as false positive and false negative
The former denotes that the non-iris pixel is incorrectly classified into an iris, whereas the latter denotes errors. The
that former
the irisdenotes
pixel is that the non-iris
incorrectly pixelinto
classified is incorrectly classified
non-iris. In intothe
Figure 13, anfalse
iris, whereas
positive the
andlatter
false denotes
negative
that
errors arethe iris pixel
shown is incorrectly
in green classified into
and red, respectively. Thenon-iris. In Figure
true positive case13,
(iristhe false
pixel is positive
correctlyand false
classified
negative errors are shown in green and red, respectively. The true positive case
into iris one) is shown in black. As shown in Figure 13, our proposed method can correctly segment (iris pixel is correctly
classified
the iris regioninto iris one) is of
irrespective shown in black.
various noisesAsinshown in Figure
eye image. 13, our14proposed
Figure method can
shows examples of correctly
incorrect
segment the iris region irrespective of various noises in eye image. Figure 14 shows examples of
segmentation of the iris region by our proposed method. False positive errors occur in the eyelash area
incorrect segmentation of the iris region by our proposed method. False positive errors occur in the
whose pixel values are similar to that of the iris region whereas false negative errors happen in case of
eyelash area whose pixel values are similar to that of the iris region whereas false negative errors
the reflection noises from glasses surface or severely dark iris area.
happen in case of the reflection noises from glasses surface or severely dark iris area.

Figure 13. Cont.

Symmetry 2017, 9, 263 17 of 25
Symmetry 2017, 9, 263 17 of 25

Symmetry 2017, 9, 263 17 of 25

(a) (b) (c) (d)

Figure 13. Examples
(a) of of good segmentation results by our proposed
(b)results method. (a) And (c) (d) segmentation
Figure 13. Examples good segmentation by our proposed(c)method. (a,c) segmentation results
results with corresponding 𝐸𝑖 ; (b) and (d) ground truth images (the false positive and false negative
with corresponding Ei ; (b,d)ofground
Figure 13. Examples truth images
good segmentation (thebyfalse
results positivemethod.
our proposed and false(a) negative errors are shown
And (c) segmentation
errors are shown in green and red, respectively. The true positive case (iris pixel is correctly classified
in green and red,
results respectively.
with corresponding 𝐸𝑖 ; true
The positive
(b) and case
(d) ground (irisimages
truth pixel(the
is correctly classified
false positive and falseinto iris one) is
negative
into iris one)are
is shown in black).
shown in errors
black). shown in green and red, respectively. The true positive case (iris pixel is correctly classified
into iris one) is shown in black).

Figure 14. Cont.

Symmetry 2017, 9, 263 18 of 25
Symmetry 2017, 9, 263 18 of 25

(a) (b) (c)

Figure 14. Examples of incorrect segmentation of the iris region by our proposed method. (a) Original
Figure 14. Examples
input images;of incorrect segmentation
(b) segmentation of the
results with iris region𝐸by
corresponding our proposed method. (a) Original
𝑖 ; (c) ground truth images (the false
input images; (b) segmentation results with corresponding E ; (c)
positive and false negative errors are shown in green and red, respectively.
i ground truth
The true images
positive case (the
(iris false
pixel is correctly classified into iris one) is shown in black).
positive and false negative errors are shown in green and red, respectively. The true positive case
(iris pixel is correctly classified into iris one) is shown in black).
4.3.2. Comparison of the Proposed Method with Previous Methods
In Table
4.3.2. Comparison 3, we
of the show the Method
Proposed comparisons of the
with proposed
Previous method with previous methods based on
Methods
E1 of Equation (7). As shown in these results, our method outperforms the previous methods in iris
In Table 3, we show
segmentation the comparisons of the proposed method with previous methods based on
error.
E1 of EquationIn addition,
(7). As shown we compared
in thesetheresults,
results byour support
method vector machine (SVM)
outperforms thewith those bymethods
previous our CNN. in iris
For fair
segmentation error. comparison, we used the same training and testing data (obtained from the stage 1 of Figure
1) of two-fold cross validation for both SVM and CNN. From the two-fold cross validation, we
In addition, we compared the results by support vector machine (SVM) with those by our CNN.
obtained the average value of E1 of Equation (7). As shown Table 3, E1 by proposed method (using
For fair comparison,
CNN for thewe used
stage 2 ofthe same
Figure 1) istraining
lower than and thattesting data (obtained
by SVM-based from SVM
method (using the stage
for the1stage
of Figure
2 1)
of two-foldofcross
Figurevalidation for both
1), from which we SVM and CNN.
can confirm that ourFrom the two-fold
proposed CNN-based crossmethod
validation, wethan
is better obtained
the averagetraditional
value of machine learning (7).
E1 of Equation method of SVM.Table
As shown The reason
3, E1 why the error by
by proposed CNN-based
method (using method
CNNisfor the
lower than that by SVM-based method is that optimal features by optimal filters can be extracted by
stage 2 of Figure 1) is lower than that by SVM-based method (using SVM for the stage 2 of Figure 1),
the 13 convolutional layers (Table 2) of our CNN compared to SVM.
from which weThe cangoalconfirm that our proposed CNN-based method is better than traditional machine
of our research is to detect the accurate positions of iris pixels based on pixel-level labels.
learning method
For this purpose, we can reason
of SVM. The considerwhy the error
fast R-CNN [49] orbyfaster
CNN-based
R-CNN [50], method
but onlyis thelower
area ofthan
squarethat by
SVM-basedbox method is thatonoptimal
shape (based box-levelfeatures
labels) canbybe optimal
detectedfilters
by thesecan be extracted
method by the 13inconvolutional
[49,50]. Therefore, order to
detect
layers (Table 2) oftheour
accurate
CNNpositions
compared of all
tothe iris pixels (not square box-level, but pixel-level), we should
SVM.
consider the different type of CNN such as semantic segmentation network (SSN) [51,52]. By using
The goal of our research is to detect the accurate positions of iris pixels based on pixel-level labels.
this SSN, the accurate positions of all the iris pixels can be obtained. In details, for SSN, whole image
For this purpose, we can consider fast R-CNN [49] or faster R-CNN [50], but only the area of square box
(instead of the mask of 21 × 21 pixels extracted from the stage 1 of Figure 1) was used as input without
shape (based on box-level
the stage 1 of Figure labels) can
1. For fairbecomparison,
detected by wethese
used method
the same [49,50].
training andTherefore, in order
testing images to detect
(from
the accurate positions
which the mask ofofall21the
× 21iris pixels
pixels were (not square
extracted box-level,
for our method) for butthepixel-level),
SSN. From the we should
two-fold consider
cross
validation,
the different type of we CNN obtained
suchthe as average
semantic value of E1 of Equation
segmentation (7). As (SSN)
network shown Table 3, E1
[51,52]. Bybyusing
proposed
this SSN,
method (using CNN for the stage 2 of Figure 1) is lower than that by SSN-based method (using SSN
the accurate positions of all the iris pixels can be obtained. In details, for SSN, whole image (instead
for the stages 1 and 2 of Figure 1), from which we can confirm that our proposed method is better
of the maskthanof 21 × 21 pixels
SSN-based method. extracted
The reason from
why the stageby1SSN-based
the error of Figuremethod1) wasisused higheras input
than that without
by our the
stage 1 of Figure 1. For fair comparison, we used the same training and testing
method is that the classification complexity of iris and non-iris pixels is increased by applying the images (from which the
mask of 21 × 21 pixels were extracted for our method) for the SSN. From the two-fold cross validation,
SSN to whole image instead of the ROI detected by the stage 1 of Figure 1.
we obtained the average value of E1 of Equation (7). As shown Table 3, E1 by proposed method (using
Table 3. Comparison of the proposed method with previous methods using NICE-II dataset.
CNN for the stage 2 of Figure 1) is lower than that by SSN-based method (using SSN for the stages
Method E1
1 and 2 of Figure 1), from which we can confirm that our proposed method is better than SSN-based
Luengo-Oroz et al. [53]
method. The reason why the error(8th byrankSSN-based method is higher 0.0305
than that by our method is that the
in NICE-I competition)
classification complexity of iris and non-iris Labati etpixels
al. [54] is increased by applying the SSN to whole image
0.0301
(7th rank in NICE-I competition)
instead of the ROI detected by the stage 1 of Figure 1.
Chen et al. [55]
0.029
(6th rank in NICE-I competition)
Table 3. Comparison of the proposed method with previous
Jeong et al. [18]
methods using NICE-II dataset.
0.028
(5th rank in NICE-I competition)
Method E1
Luengo-Oroz et al. [53]
0.0305
(8th rank in NICE-I competition)
Labati et al. [54]
0.0301
(7th rank in NICE-I competition)
Chen et al. [55]
0.029
(6th rank in NICE-I competition)
Jeong et al. [18]
0.028
(5th rank in NICE-I competition)
Symmetry 2017, 9, 263 19 of 25

Table 3. Cont.

Method E1
Peihua et al. [56]
0.022
(4th rank in NICE-I competition)
Tan et al. [57] 0.019
Proença [58] 0.0187
de Almeida [59]
0.0180
(3rd rank in NICE-I competition)
Tan et al. [60] 0.0172
Sankowski et al. [61]
0.016
(2nd rank in NICE-I competition)
Tan et al. [19]
Symmetry 2017, 9, 263 0.0131 19 of 25
(1st rank in NICE-I competition)
Haindl et al.
Peihua et al [56]
[62] 0.0124
0.022
(4th rank in NICE-I competition)
Zhao et al. [16] 0.0121
Tan et al. [57] 0.019
SSN-based method
Proença [58] [51,52] 0.02816
0.0187
de Almeida
Proposed [59]
method 0.0180
0.03852
(3rd
(Using rankfor
SVM in NICE-I
the stagecompetition)
2 of Figure 1)
Tan et al. [60] 0.0172
Proposed method
Sankowski et al. [61] 0.0082
0.016
(Using
(2ndCNNrankfor the stage
in NICE-I 2 of Figure 1)
competition)
Tan et al. [19]
0.0131
(1st rank in NICE-I competition)
4.3.3. Iris Segmentation Error with Another Haindl etOpen
al [62] Database 0.0124
Zhao et al. [16] 0.0121
We performed additional experiments
SSN-based method with other open database
[51,52] 0.02816 of the mobile iris challenge
evaluation (MICHE) data [63,64]. There are various
Proposed method databases of the CASIA datasets and iris challenge
0.03852
(Using SVM for the stage 2 of Figure 1)
evaluation (ICE) datasets. The goal of this method
Proposed study is to obtain correct iris segmentation with the iris
0.0082
image captured by the visible(Using
lightCNN environment. However,
for the stage 2 of Figure 1) very few open iris databases by visible
light environment exist, and we used the MICHE datasets for this reason. They were collected by three
4.3.3. Iris Segmentation Error with Another Open Database
mobile phones, such as iPhone 5, Galaxy Samsung IV, and Galaxy Tablet II in both indoor and outdoor
We performed additional experiments with other open database of the mobile iris challenge
environments. The ground
evaluation (MICHE) truth
data images are not
[63,64]. There are provided; therefore,
various databases of theamong the whole
CASIA datasets and images,
iris we used
the images where
challengethe ground
evaluation truth
(ICE) irisThe
datasets. regions
goal of can be obtained
this study is to obtain by their
correct iris provided
segmentationalgorithm
with [56,58]
the iris image captured by the visible light environment. However, very few open iris databases by
according to the instruction of MICHE. Same procedure of two-fold cross validation was also used for
visible light environment exist, and we used the MICHE datasets for this reason. They were collected
the MICHE by data asmobile
three that for NICE-II
phones, such asdataset.
iPhone 5, Galaxy Samsung IV, and Galaxy Tablet II in both indoor
Figuresand15 outdoor
and 16environments.
show the examples of good
The ground truth images segmentation
are not provided; and incorrect
therefore, amongsegmentations
the whole by our
images, we used the images where the ground truth iris regions can be obtained by their provided
proposed method, respectively. Like Figures 13 and 14, the false positive and false negative errors
algorithm [56,58] according to the instruction of MICHE. Same procedure of two-fold cross validation
are shown inwas green andforred,
also used respectively.
the MICHE The
data as that true positive
for NICE-II dataset. case (iris pixel is correctly classified into
iris one) is shown in black. As shown in Figure 15, our method
Figures 15 and 16 show the examples of good segmentation and can correctly
incorrect segment
segmentations by ourthe iris region
proposed method, respectively. Like Figures 13 and 14, the false positive and false negative errors
with the images captured by various cameras and environment. As shown in Figure 16, the false
are shown in green and red, respectively. The true positive case (iris pixel is correctly classified into
positive errors are iscaused
iris one) shown inby theAseyelid
black. shown region
in Figurewhose pixel can
15, our method values aresegment
correctly similar theto
iristhose
region of iris region.
On the other hand, false negative errors happen in case of the reflection noises from glasses
with the images captured by various cameras and environment. As shown in Figure 16, the false surface or
positive errors are caused by the eyelid region whose pixel values are similar to those of iris region.
environmental sunlight.
On the other hand, false negative errors happen in case of the reflection noises from glasses surface

Figure 15. Cont.

Symmetry 2017, 9, 263 20 of 25

Symmetry 2017, 9, 263 20 of 25

(a) (b) (c) (d)

Figure 15. Examples of good segmentation results by our method. (a) And (c) segmentation results
Examples
Figure 15.with of good
corresponding segmentation
𝐸𝑖 . (b) and (d) ground results by our
truth images method.
(the false (a)and
positive Andfalse(c) segmentation
negative errors results
are shown in E
with corresponding i . (b)and
green and (d)
red, ground truth
respectively. images
The true (the
positive false
case (iris positive and false
pixel is correctly negative
classified into errors are
shown iniris one) is
green andshown
red,inrespectively.
black). The true positive case (iris pixel is correctly classified into iris
one) is shown in black).
Symmetry 2017, 9, 263 21 of 25
Symmetry 2017, 9, 263 21 of 25

(a) (b) (c)

Figure 16. Examples of incorrect segmentation of the iris region by our method. (a) Original input
Figure 16. Examples of incorrect segmentation of the iris region by our method. (a) Original input
images; (b) segmentation results with corresponding 𝐸 ; (c) ground truth images (The false positive
images; (b) segmentation results with corresponding E𝑖i ; (c) ground truth images (The false positive
and false negative errors are shown in green and red, respectively. The true positive case (iris pixel is
and false negative errors are shown in green and red, respectively. The true positive case (iris pixel is
correctly classified into iris one) is shown in black).
correctly classified into iris one) is shown in black).
In the next experiments, we compared the error of the iris segmentation by our proposed method
In that
with the next experiments,
by previous we compared
methods in terms ofthe
E1 error of the iris
of Equation (7). segmentation by our
As shown in these proposed
results method
of Table 4,
with
ourthat by previous
proposed methodmethods in terms
outperforms of E1 ofmethods
the previous Equation in(7).
iris As shown in error.
segmentation these results of Table 4,
our proposed method outperforms the previous methods in iris segmentation error.
Table 4. Comparison of the proposed method with previous methods using MICHE dataset.
Table 4. Comparison of the proposed method with previous methods using MICHE
Method E1 dataset.
Sub-dataset by iPhone 5 0.0193
Hu et al. [65]
Method E1 0.0193
Sub-dataset by Galaxy Samsung IV 0.0192
Sub-dataset
Sub-datasetbybyiPhone 55
iPhone 0.00368
0.0193
Hu etmethod
al. [65] 0.0193
Proposed Sub-dataset by Galaxy Samsung IV 0.00297 0.00345
Sub-dataset by Galaxy Samsung IV 0.0192
Sub-dataset by Galaxy Tablet II 0.00352
Sub-dataset by iPhone 5 0.00368
5. Conclusions Sub-dataset by Galaxy Samsung IV 0.00297 0.00345
Proposed method

In this study, we proposed aSub-dataset by Galaxy

robust two-stage Tablet II method,
CNN-based 0.00352
which can find the true iris
boundary within noisy iris images in a non-cooperative environment. In the first stage, a rough iris
5. boundary
Conclusions is found via modified circular HT, which defines the ROI by the slightly increased radius
of the iris. In the second stage, CNN is applied by VGG-face fine-tuning to the data obtained from
In this study, we proposed a robust two-stage CNN-based method, which can find the true iris
the ROI. The CNN output layer provides two output features. Therefore, based on these features, the
boundary within noisy iris images in a non-cooperative environment. In the first stage, a rough iris
iris and non-iris points are classified to find the true iris boundary. Experiments with NICE-II and
boundary
MICHE is found via
databases modified
show circular
that the HT, method
proposed which defines thehigher
achieved ROI by the slightly
accuracies increased
of iris radius of
segmentation
thecompared
iris. In the second stage, CNN is applied
to the state-of-the-art methods. by VGG-face fine-tuning to the data obtained from the
ROI. The CNN output layer provides two output features. Therefore, based on these features, the iris
and non-iris points are classified to find the true iris boundary. Experiments with NICE-II and MICHE
Symmetry 2017, 9, 263 22 of 25

databases show that the proposed method achieved higher accuracies of iris segmentation compared
to the state-of-the-art methods.
Although our method shows the high accuracy of iris segmentation, the traditional image
processing algorithms should be used in the stage 1 of Figure 1. In addition, it is necessary to
reduce the processing time for CNN-based classification with the window masks extracted from the
stage 1. To solve these problems, we can consider the other type of CNN such as semantic segmentation
network (SSN) which can use whole image as input (without the stages 1 and 2 of Figure 1). However,
as shown in Table 3, its performance is lower than proposed method. As future work, we would
research the method of using this SSN with appropriate post-processing so as to get high accuracy and
fast processing speed. In addition, we would apply our proposed method to various iris datasets in
NIR light environments, or other biometrics, such as vein segmentation for finger-vein recognition or
human body segmentation in thermal image.

Acknowledgments: This research was supported by the Bio & Medical Technology Development Program of
the NRF funded by the Korean government, MSIP (NRF-2016M3A9E1915855), by the Basic Science Research
Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education
(NRF-2017R1D1A1B03028417), and by the National Research Foundation of Korea (NRF) grant funded by the
Korea government (MSIP; Ministry of Science, ICT & Future Planning) (NRF-2017R1C1B5074062).
Author Contributions: Muhammad Arsalan and Kang Ryoung Park designed the overall system for iris
segmentation. In addition, they wrote and revised the paper. Hyung Gil Hong, Rizwan Ali Naqvi, Min Beom Lee,
Min Cheol Kim, Dong Seop Kim, and Chan Sik Kim helped to design the CNN and comparative experiments.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Li, P.; Zhang, R. The evolution of biometrics. In Proceedings of the IEEE International Conference on
Anti-Counterfeiting Security and Identification in Communication, Chengdu, China, 18–20 July 2010;
pp. 253–256.
2. Larrain, T.; Bernhard, J.S.; Mery, D.; Bowyer, K.W. Face recognition using sparse fingerprint classification
algorithm. IEEE Trans. Inf. Forensics Secur. 2017, 12, 1646–1657. [CrossRef]
3. Bonnen, K.; Klare, B.F.; Jain, A.K. Component-based representation in automated face recognition. IEEE Trans.
Inf. Forensics Secur. 2013, 8, 239–253. [CrossRef]
4. Daugman, J. How iris recognition works. IEEE Trans. Circuits Syst. Video Technol. 2004, 14, 21–30. [CrossRef]
5. Viriri, S.; Tapamo, J.R. Integrating iris and signature traits for personal authentication using user-specific
weighting. Sensors 2012, 12, 4324–4338. [CrossRef] [PubMed]
6. Jain, A.K.; Arora, S.S.; Cao, K.; Best-Rowden, L.; Bhatnagar, A. Fingerprint recognition of young children.
IEEE Trans. Inf. Forensic Secur. 2017, 12, 1501–1514. [CrossRef]
7. Hong, H.G.; Lee, M.B.; Park, K.R. Convolutional neural network-based finger-vein recognition using NIR
image sensors. Sensors 2017, 17, 1297. [CrossRef] [PubMed]
8. Yaxin, Z.; Huanhuan, L.; Xuefei, G.; Lili, L. Palmprint recognition based on multi-feature integration.
In Proceedings of the IEEE Advanced Information Management, Communicates, Electronic and Automation
Control Conference, Xi’an, China, 3–5 October 2016; pp. 992–995.
9. Daugman, J. Information theory and the IrisCode. IEEE Trans. Inf. Forensics Secur. 2016, 11, 400–409.
[CrossRef]
10. Hsieh, S.-H.; Li, Y.-H.; Tien, C.-H. Test of the practicality and feasibility of EDoF-empowered image sensors
for long-range biometrics. Sensors 2016, 16, 1994. [CrossRef] [PubMed]
11. He, Z.; Tan, T.; Sun, Z.; Qiu, X. Toward accurate and fast iris segmentation for iris biometrics. IEEE Trans.
Pattern Anal. Mach. Intell. 2009, 31, 1670–1684. [PubMed]
12. Li, Y.-H.; Savvides, M. An automatic iris occlusion estimation method based on high-dimensional density
estimation. IEEE Trans. Pattern Anal. Mach. Intell. 2013, 35, 784–796. [CrossRef] [PubMed]
13. Matey, J.R.; Naroditsky, O.; Hanna, K.; Kolczynski, R.; LoIacono, D.J.; Mangru, S.; Tinker, M.; Zappia, T.M.;
Zhao, W.Y. Iris on the move: Acquisition of images for iris recognition in less constrained environments.
Proc. IEEE 2006, 94, 1936–1947. [CrossRef]
14. Wildes, R.P. Iris recognition: An emerging biometric technology. Proc. IEEE 1997, 85, 1348–1363. [CrossRef]
Symmetry 2017, 9, 263 23 of 25

15. Uhl, A.; Wild, P. Weighted adaptive hough and ellipsopolar transforms for real-time iris segmentation.
In Proceedings of the 5th IEEE International Conference on Biometrics, New Delhi, India, 29 March–1 April
2012; pp. 283–290.
16. Zhao, Z.; Kumar, A. An accurate iris segmentation framework under relaxed imaging constraints using total
variation model. In Proceedings of the IEEE International Conference on Computer Vision, Santiago, Chile,
7–13 December 2015; pp. 3828–3836.
17. Roy, D.A.; Soni, U.S. IRIS segmentation using Daughman’s method. In Proceedings of the IEEE International
Conference on Electrical, Electronics, and Optimization Techniques, Chennai, India, 3–5 March 2016;
pp. 2668–2676.
18. Jeong, D.S.; Hwang, J.W.; Kang, B.J.; Park, K.R.; Won, C.S.; Park, D.-K.; Kim, J. A new iris segmentation
method for non-ideal iris images. Image Vis. Comput. 2010, 28, 254–260. [CrossRef]
19. Tan, T.; He, Z.; Sun, Z. Efficient and robust segmentation of noisy iris images for non-cooperative iris
recognition. Image Vis. Comput. 2010, 28, 223–230. [CrossRef]
20. Khan, T.M.; Khan, M.A.; Malik, S.A.; Khan, S.A.; Bashir, T.; Dar, A.H. Automatic localization of pupil using
eccentricity and iris using gradient based method. Opt. Lasers Eng. 2011, 49, 177–187. [CrossRef]
21. Parikh, Y.; Chaskar, U.; Khakole, H. Effective approach for iris localization in nonideal imaging conditions.
In Proceedings of the IEEE Students’ Technology Symposium, Kharagpur, India, 28 February–2 March 2014;
pp. 239–246.
22. Shah, S.; Ross, A. Iris segmentation using geodesic active contours. IEEE Trans. Inf. Forensic Secur. 2009, 4,
824–836. [CrossRef]
23. Abdullah, M.A.M.; Dlay, S.S.; Woo, W.L.; Chambers, J.A. Robust iris segmentation method based on a new
active contour force with a noncircular normalization. IEEE Trans. Syst. Man Cybern. 2017, 1–14, in press.
[CrossRef]
24. Chai, T.-Y.; Goi, B.-M.; Tay, Y.H.; Chin, W.-K.; Lai, Y.-L. Local Chan-Vese segmentation for non-ideal visible
wavelength iris images. In Proceedings of the IEEE Conference on Technologies and Applications of Artificial
Intelligence, Tainan, Taiwan, 20–22 November 2015; pp. 506–511.
25. Li, H.; Sun, Z.; Tan, T. Robust iris segmentation based on learned boundary detectors. In Proceedings of the
International Conference on Biometrics, New Delhi, India, 29 March–1 April 2012; pp. 317–322.
26. Vokhidov, H.; Hong, H.G.; Kang, J.K.; Hoang, T.M.; Park, K.R. Recognition of damaged arrow-road markings
by visible light camera sensor based on convolutional neural network. Sensors 2016, 16, 2160. [CrossRef]
[PubMed]
27. Nguyen, D.T.; Kim, K.W.; Hong, H.G.; Koo, J.H.; Kim, M.C.; Park, K.R. Gender recognition from human-body
images using visible-light and thermal camera videos based on a convolutional neural network for image
feature extraction. Sensors 2017, 17, 637. [CrossRef] [PubMed]
28. Kim, J.H.; Hong, H.G.; Park, K.R. Convolutional neural network-based human detection in nighttime images
using visible light camera sensors. Sensors 2017, 17, 1–26.
29. Pereira, S.; Pinto, A.; Alves, V.; Silva, C.A. Brain tumor segmentation using convolutional neural networks in
MRI images. IEEE Trans. Med. Imaging 2016, 35, 1240–1251. [CrossRef] [PubMed]
30. Liu, N.; Zhang, M.; Li, H.; Sun, Z.; Tan, T. DeepIris: Learning pairwise filter bank for heterogeneous iris
verification. Pattern Recognit. Lett. 2016, 82, 154–161. [CrossRef]
31. Gangwar, A.; Joshi, A. DeepIrisNet: Deep iris representation with applications in iris recognition and
cross-sensor iris recognition. In Proceedings of the IEEE International Conference on Image Processing,
Phoenix, USA, 25–28 September 2016; pp. 2301–2305.
32. Liu, N.; Li, H.; Zhang, M.; Liu, J.; Sun, Z.; Tan, T. Accurate iris segmentation in non-cooperative environments
using fully convolutional networks. In Proceedings of the IEEE International Conference on Biometrics,
Halmstad, Sweden, 13–16 June 2016; pp. 1–8.
33. Dongguk Visible Light Iris Segmentation CNN Model (DVLIS-CNN). Available online: http://dm.dgu.edu/
link.html (accessed on 27 July 2017).
34. Parkhi, O.M.; Vedaldi, A.; Zisserman, A. Deep face recognition. In Proceedings of the British Machine Vision
Conference, Swansea, UK, 7–10 September 2015; pp. 1–12.
35. CS231n Convolutional Neural Networks for Visual Recognition. Available online: http://cs231n.github.io/
neural-networks-1/ (accessed on 26 July 2017).
Symmetry 2017, 9, 263 24 of 25

36. Glorot, X.; Bordes, A.; Bengio, Y. Deep sparse rectifier neural networks. In Proceedings of the 14th
International Conference on Artificial Intelligence and Statistics, Fort Lauderdale, FL, USA, 11–13 April 2011;
pp. 315–323.
37. Nair, V.; Hinton, G.E. Rectified linear units improve restricted boltzmann machines. In Proceedings of the
27th International Conference onachine Learning, Haifa, Israel, 21–24 June 2010; pp. 807–814.
38. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks.
In Advances in Neural Information Processing Systems 25; Curran Associates, Inc.: New York, NY, USA, 2012;
pp. 1097–1105.
39. Heaton, J. Artificial Intelligence for Humans. In Deep Learning and Neural Networks; Heaton Research, Inc.:
St. Louis, USA, 2015; Volume 3.
40. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A simple way to prevent
neural networks from overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958.
41. Wyatt, H.J. A ‘minimum-wear-and-tear‘ meshwork for the iris. Vision Res. 2000, 40, 2167–2176. [CrossRef]
42. NICE.II. Noisy Iris Challenge Evaluation-Part II. Available online: http://nice2.di.ubi.pt/index.html
(accessed on 26 July 2017).
43. Proença, H.; Filipe, S.; Santos, R.; Oliveira, J.; Alexandre, L.A. The UBIRIS.v2: A database of visible
wavelength iris images captured on-the-move and at-a-distance. IEEE Trans. Pattern Anal. Mach. Intell. 2010,
32, 1529–1535.
44. Geforce GTX 1070. Available online: https://www.nvidia.com/en-us/geforce/products/10series/geforce-
gtx-1070/ (accessed on 4 August 2017).
45. Caffe. Available online: http://caffe.berkeleyvision.org (accessed on 4 August 2017).
46. Stochastic Gradient Descent. Available online: https://en.wikipedia.org/wiki/Stochastic_gradient_descent
(accessed on 26 July 2017).
47. Caffe Solver Parameters. Available online: https://github.com/BVLC/caffe/wiki/Solver-Prototxt (accessed
on 4 August 2017).
48. NICE.I. Noisy Iris Challenge Evaluation-Part I. Available online: http://nice1.di.ubi.pt/ (accessed on
26 July 2017).
49. Girshick, R. Fast R-CNN. In Proceedings of the International Conference on Computer Vision, Santiago,
Chile, 11–18 December 2015; pp. 1440–1448.
50. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards real-time object detection with region proposal
networks. arXiv 2016, 1–14. [CrossRef] [PubMed]
51. Semantic Segmentation Using Deep Learning. Available online: https://kr.mathworks.com/help/vision/
examples/semantic-segmentation-using-deep-learning.html (accessed on 23 October 2017).
52. Badrinarayanan, V.; Kendall, A.; Cipolla, R. SegNet: A deep convolutional encoder-decoder architecture for
image segmentation. arXiv 2015, 1–14. [CrossRef] [PubMed]
53. Luengo-Oroz, M.A.; Faure, E.; Angulo, J. Robust iris segmentation on uncalibrated noisy images using
mathematical morphology. Image Vis. Comput. 2010, 28, 278–284. [CrossRef]
54. Labati, R.D.; Scotti, F. Noisy iris segmentation with boundary regularization and reflections removal.
Image Vis. Comput. 2010, 28, 270–277. [CrossRef]
55. Chen, Y.; Adjouadi, M.; Han, C.; Wang, J.; Barreto, A.; Rishe, N.; Andrian, J. A highly accurate and
computationally efficient approach for unconstrained iris segmentation. Image Vis. Comput. 2010, 28, 261–269.
[CrossRef]
56. Li, P.; Liu, X.; Xiao, L.; Song, Q. Robust and accurate iris segmentation in very noisy iris images.
Image Vis. Comput. 2010, 28, 246–253. [CrossRef]
57. Tan, C.-W.; Kumar, A. Unified framework for automated iris segmentation using distantly acquired face
images. IEEE Trans. Image Process. 2012, 21, 4068–4079. [CrossRef] [PubMed]
58. Proença, H. Iris recognition: on the segmentation of degraded images acquired in the visible wavelength.
IEEE Trans. Pattern Anal. Mach. Intell. 2010, 32, 1502–1516. [CrossRef] [PubMed]
59. De Almeida, P. A knowledge-based approach to the iris segmentation problem. Image Vis. Comput. 2010, 28,
238–245. [CrossRef]
60. Tan, C.-W.; Kumar, A. Towards online iris and periocular recognition under relaxed imaging constraints.
IEEE Trans. Image Process. 2013, 22, 3751–3765. [PubMed]
Symmetry 2017, 9, 263 25 of 25

61. Sankowski, W.; Grabowski, K.; Napieralska, M.; Zubert, M.; Napieralski, A. Reliable algorithm for iris
segmentation in eye image. Image Vis. Comput. 2010, 28, 231–237. [CrossRef]
62. Haindl, M.; Krupička, M. Unsupervised detection of non-iris occlusions. Pattern Recognit. Lett. 2015, 57,
60–65. [CrossRef]
63. MICHE Dataset. Available online: http://biplab.unisa.it/MICHE/index_miche.htm (accessed on
4 August 2017).
64. Marsico, M.; Nappi, M.; Ricco, D.; Wechsler, H. Mobile iris challenge evaluation (MICHE)-I, biometric iris
dataset and protocols. Pattern Recognit. Lett. 2015, 57, 17–23. [CrossRef]
65. Hu, Y.; Sirlantzis, K.; Howells, G. Improving colour iris segmentation using a model selection technique.
Pattern Recognit. Lett. 2015, 57, 24–32. [CrossRef]

© 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).