Content uploaded by David Menotti
Author content
All content in this area was uploaded by David Menotti on Jun 22, 2015
Content may be subject to copyright.
An Approach to Iris Contact Lens Detection
based on Deep Image Representations
Pedro Silva, Eduardo Luz, Rafael Baeta, David Menotti
Computing Department
Federal University of Ouro Preto - UFOP
Ouro Preto, MG, Brazil, 35400-000
Email: {pedroh21.silva,eduluz,rafael.baeta,menottid}@gmail.com
Helio Pedrini, Alexandre Xavier Falc
˜
ao
Institute of Computing
University of Campinas - UNICAMP
Campinas, SP, Brazil, 13083-852
Email: {helio,afalcao}@ic.unicamp.br
Abstract—Spoofing detection is a challenging task in biometric
systems, when differentiating illegitimate users from genuine
ones. Although iris scans are far more inclusive than fingerprints,
and also more precise for person authentication, iris recognition
systems are vulnerable to spoofing via textured cosmetic contact
lenses. Iris spoofing detection is also referred to as liveness
detection (binary classification of fake and real images). In this
work, we focus on a three-class detection problem: images with
textured (colored) contact lenses, soft contact lenses, and no
lenses. Our approach uses a convolutional network to build a deep
image representation and an additional fully-connected single
layer with softmax regression for classification. Experiments are
conducted in comparison with a state-of-the-art approach (SOTA)
on two public iris image databases for contact lens detection: 2013
Notre Dame and IIIT-Delhi. Our approach can achieve a 30%
performance gain over SOTA on the former database (from 80%
to 86%) and comparable results on the latter. Since IIIT-Delhi
does not provide segmented iris images and, differently from
SOTA, our approach does not segment the iris yet, we conclude
that these are very promising results.
Keywords-Iris Biometrics; Contact Lens Detection; Deep
Learning; Convolutional Networks.
I. INTRODUCTION
Biometric-based person identification systems have been
developed rapidly in the last two decades. Moreover, biometric
systems based on iris recognition have been deployed in
several applications, such as border-crossing control systems,
controlled environments, access to personal computers and
smart phones. Iris is considered as the most promising, reliable
and accurate biometric trait, providing rich texture that allows
high discrimination among subjects. Furthermore, iris is stable
along ageing of individuals [1].
The first functional iris recognition method was introduced
by Daugman in 1993 [1], whereas the first patent proposing
iris texture as biometric modality appeared in 1987 [2].
Thenceforward, several iris recognition approaches have been
proposed in the literature [3]–[5].
Due to the increasing use of iris as a source of biometric
information in the last decade, the possibility of attacks
to these systems has become more common [6]–[8]. These
attacks are usually referred in the literature to as iris spoofing
and several works for dealing with this problem have been
proposed [9]–[11]. Nonetheless, the definition of iris spoofing
detection may be confusing, where liveness and counterfeit
detection terms are used with different meanings and, in some
cases, interchangeably [12]. Several works in the literature
have addressed the problem of classifying an iris image as
real/live or as fake, in which a fake image is not a live one
(e.g., a printed image [6], [10], [13]). In addition, counterfeit
detection approaches have also been proposed in the past
years [14]–[19], in which counterfeit iris with printed color
contact lenses are considered fake images and iris images with
soft/clear or no lenses are considered real images.
Given that cosmetic contact lenses are becoming more
popular, the sort of attacks with textured contact lenses that
an iris biometric system may suffer varies in a wide range.
For instance, a person who has banned from a country or
geographical region and has been included in a watch list may
want to rejoin that region by using contact lenses to obfuscate
his/her textured irises avoiding to be identified. Similarly,
an individual may want to personify someone else by using
textured contact lens iris of an enrolled person [9]. Moreover,
transparent or prescript contact lenses used when acquiring iris
images have indeed shown to decrease the false acceptance
rate [15], [19] in iris recognition systems, demonstrating that
it is important to identify when soft/clear lenses are present.
Furthermore, the accuracy of textured contact lens detection
methods may be affected by contact lens patterns and also
sensor manufacturer as shown in [18].
In this context, we introduce the use of deep learning
techniques [20]. In the last few years, deep learning has
allowed promising and outstanding results for several and
important visual analysis tasks, such as face recognition [21]–
[24], pedestrian detection [25], character recognition [26],
[27], traffic sign classification [28], general object recognition
in large categorized databases [29], among others. Besides the
success in these areas, the use of deep representations for
spoofing detection on iris, face, and fingerprint images has
also been recently proposed [10], in which a simpler two-
class problem of detecting fake/spoof and real/live images is
addressed.
The present paper addresses a more complex three-class
image detection problem, where iris images may appear with
textured (colored) contact lenses, soft contact lenses, and no
lenses. We propose a convolutional network to build deep
image representations, followed by a fully-connected single
layer with softmax regression for image classification. Our
approach is based on the work of Krizhevsky et al. [29], in
which the weights of all layers are learned by backpropagation.
In [30], the authors present two image databases to evaluate
methods on the three-class detection problem: the 2013 Notre
Dame Contact Lens Detection database (NDCL) and the IIIT-
Delhi Contact Lens Iris database (IIIT-D). Each database
contains images from two different sensors: LG4000 and
AD100 in the NDCL database, where images come with
iris location, and Cogent and Vista in the IIIT-D database,
where iris location is not available (i.e., more challenging).
We compare our approach with the state-of-the-art algorithm
(SOTA), also proposed in [30], by taking into account images
from each sensor and from different sensors.
The paper is organized as follows. In Section II, we present
a brief review of relevant works directly related to contact
lens spoofing detection. In Section III, the databases used in
our experiments are described. The methodology proposed
to cope with spoofing detection is detailed in Section IV.
Experimental results are described and discussed in Section V.
Finally, conclusions and directions for future work are outlined
in Section VI.
II. RELATED WORK
In this section, we review some relevant works directly
related to the three-class iris image problem addressed in this
paper, that is, those that propose to classify iris images in
(color) textured contact lenses, soft (prescript or clear) contact
lenses and non lenses.
The first step of a recognition system is to capture the iris
images. Due to the difficulty in identifying the iris textures in
the color images, sensors have to operate under near infrared
(NIR) illumination. However, cosmetic contact lenses can
change the pattern of the iris and its presence could be very
difficult to detect on images taken under NIR illumination.
Such undesirable feature is used against iris recognition sys-
tems, which makes spoofing attacks with textured lenses easier
and also increases false negative matching even for prescript
lenses [15], [17], [31].
Lee et al. [32] propose a new method for detecting fake
iris based on the Purkinje image. To acquire the data, a
conventional USB camera is used with modified CCD sensor
and special illumination. To accomplish the experiments, the
dataset is built with 300 live irises and 15 fake ones. The
authors report a false accept rate (FAR) of 0.33% and false
reject rate (FRR) of 0.33% on the dataset, however, a more
robust evaluation, on a larger and diverse dataset, should be
considered to properly validate the method.
Wei et al. [14] present three methods for detecting textured
lenses: measurement of iris edge sharpness, application of iris-
texton for characterizing the visual primitives of iris textures,
and use of selected features based on co-occurrence matrix. To
perform experiments, two datasets are built using CASIA [33]
and BATH [34] datasets for live irises, whereas fake irises are
collected by the authors. The reported correct classification
rate (CCR) is up to 100% for experiments using co-occurrence
matrix features.
In [35], a method based on Local Binary Pattern (LBP)
encoding and Adaboost learning together with Gaussian kernel
density estimation achieves FAR of 0.67% and FRR of 2.64%
on discriminating fake iris texture from live iris. The method is
evaluated on CASIA-Iris-V3 [36] and ICE v1.0 [37] with the
addition of 600 custom fake iris images. Among the used 600
fake images, there are 20 different types of textured contact
lenses.
In [16], it is proposed a contact lens detection algorithm
based on Scale-Invariant Feature Transform (SIFT), weighted
LBP and Support Vector Machines (SVM). According to
the authors, the combination of SIFT and LBP improves
its variance of scale illumination and local affine distortion.
The authors claim that their method achieves state-of-the-art
performance in contact lens detection. They build a custom
dataset of 5000 fake iris images with 70 different types of
textured lens.
After Daugman [38] presented a method for allowing easy
detection of contact lens patterns, many other authors have
reported accuracy rates over 98% [14], [16], [35]. How-
ever, since the contact lens technology is under constant
development, robust detection has become more difficult [9].
Combined to this fact, some studies found in the literature
are favored by their methodology due to the use of datasets
containing contact lenses from a single manufacturer among
both training and test data [9], [14]. According to [18], in
more realistic scenarios, methods whose accuracy are close to
100% could decrease to below 60%.
To avoid this situation, two datasets are built in [18]
with textured contact lenses from three major manufacturers.
Multiple colors are selected for each manufacturer and some
lenses are also designed to correct astigmatism. Authors claim
that textured lens detection can drop dramatically when tested
on a manufacturer of lenses not seen in the training data
and when the iris sensor is different between training and
test data. An extension of this work is presented in [30],
where the datasets are well described and made available upon
request. Additionally, state-of-the-art results are reported by a
modified LBP feature extraction method and compared with 17
different classifiers. The databases are tested with techniques
available in the literature, such as textural features based on
co-occurrence matrix, weighted LBP approach, as well as
other techniques based on LBP and SVM. Finally, the authors
suggest that the development of a fully general approach to
textured lens detection is a problem that still requires attention.
In a recent work [19], a new contact lens detection method
based on binarized statistical image features reports accuracies
near to optimality for the NDCL database. However, in that
work, the authors deal with a two-problem classification, that
is, soft/clear lens iris images are considered the same class as
non lens iris images.
Fig. 1. Samples of images in the 2013 Notre Dame Contact Lens
Detection (NDCL) database. In the first and second columns, we show
images acquired with AD100 and LG4000 sensors, respectively. The first,
second and third rows present samples with textured/cosmetic contact lenses,
soft/clear/prescript contact lenses, and no contact lenses, respectively.
III. DATABASES
In this section, we describe the databases used in our
experiments. Both are publicly available upon request and
were specifically developed for evaluation of contact lens
iris detection in a three-class way [30]. We summarize the
main characteristics of each database in Table I and present
additional details in the following subsections. Note that all
images of these databases are grayscale with 640 ×480 pixels.
A. Notre Dame Contact Lens Database
The 2013 Notre Dame Contact Lens Detection (NDCLD’13
or simply NDCL) database consists of 5100 images [39]. All
640 × 480 pixel images of this database were acquired under
near-IR illumination using two types of cameras, LG4000
and IrisGuard AD100. This database is divided into two
subsets: LG4000 with 3000 images for training and 1200 for
verification; AD100 with 600 images for training and 300
for verification. These subsets are indeed used as primary
databases for intra-camera evaluation.
The entire database, i.e., the fusion of images acquired by
the LG4000 & AD100 cameras, is proposed as a multi-camera
training set of 3600 images and a verification (testing) set
of 1500 images. The images are equally divided into three
classes: (1) wearing cosmetic contact lenses, (2) wearing clear
soft contact lenses, and (3) wearing no contact lenses. Fig. 1
illustrates some samples of the NDCL and its cameras and
classes.
Fig. 2. Samples of images in the IIIT-Delhi Contact Lens Iris (IIIT). In the
first and second columns, we show images acquired with Cogent and Vista
sensors, respectively. The first, second and third rows present samples with
textured/cosmetic contact lenses, soft/clear/prescript contact lenses, and no
contact lenses, respectively.
All images in the database are annotated with the following
information: an ID to the subject it belongs, eye (left and
right), the subject’s gender, race, the type of contact lenses
used, and the coordinates of pupil and iris. These coordinates
allow us to perform experiments considering perfect iris seg-
mentation. More specific details for this database can be found
in [39, Section II.B]
B. IIIT-D Contact Lens Iris Database
The Indraprastha Institute of Information Technology (IIIT)-
Delhi Contact Lens Iris (IIIT-D CLI or simply IIIT) database
contains 6583 iris images of 101 subjects. For each individual:
(1) both left and right eyes were captured generating 202 iris
classes (different iris); (2) images were captured without lens
and with soft and textured lens - the three classes considered
here; (3) the textured lenses were captured by using variations
in iris sensors and lenses (colors and manufacturers). Images
in this database are illustrated in Fig. 2.
The used iris sensors are Cogent dual and VistaFA2E single.
Although this database offers a large variation of textured
contact lenses, the iris location information is not provided.
1
More specific details for this database can be found in [39,
Section II.A].
1
We only conducted experiments on this database using the entire eye
image. The perfect iris segmentation or annotation is planned as future work.
TABLE I
MAIN FEATURES OF THE DATASETS CONSIDERED HEREIN AND INTRODUCED IN [30].
Database Sensor
# Training # Testing/Verification # Full
Text. Soft No Total Text. Soft No Total Text. Soft No Total
NDCL
IrisGuard AD100 200 200 200 600 100 100 100 300 300 300 300 900
LG4000 iris camera 1000 1000 1000 3000 400 400 400 1200 1400 1400 1400 4200
Multi-camera 1200 1200 1200 3600 500 500 500 1500 1700 1700 1700 5100
IIIT
Cogent Scanner 589 569 563 1721 613 574 600 1787 1202 1143 1163 3508
Vista Scanner 535 500 500 1535 530 510 500 1540 1065 1010 1000 3075
Multi-scanner 1124 1069 1063 3256 1143 1084 1100 3327 2267 2153 2163 6583
IV. DEEP REPRESENTATIONS
In this section, we present the proposed method for iris
contact lens detection based on deep image representations.
Initially, we briefly describe the structure of the deep learning
techniques used to build deep representations for the problem,
which involves a combination of convolutional network [40],
for deep image representations, and a fully-connected [41]
three-layered network for classification. Then, we detail the
methodology to choose the network topology and learn its
parameters, by using the domain-knowledge from the liter-
ature. The activation operations used here are the rectified
linear units (ReLU) [29], which have demonstrated to be
essential to learn deep representations. Based on gain control
mechanisms found in cortical neurons [42], the normalization
operation promotes competition among filter outputs such
that high and isolated responses are further emphasized [10].
Spatial pooling is a foundational operation in convolutional
networks [40] that aims at bringing translational invariance to
the features by aggregating activations from the same filter
in a given region. The order of these last two operations,
i.e., normalization and pooling, in a convolutional layer is an
open problem and is application dependent. As we expect to
achieve higher discrimination power with deep representations,
the convolutional network stacks several layers for final image
representation. All these operations and layers demand the
determination of several parameters. Instead of performing
random search on the hyperparameter space [24], [43] or even
applying some specific search algorithm [44], we preferred
to empirically analyze a set of parameters at a time to build
the final network structure (topology), and to learn the filter
weights by backpropagation. The idea of learning the network
architecture by using random weights [10], [24], [43], [44]
certainly deserves more attention and we will leave this
approach for a future work. The idea here is to first evaluate
how far one can go with the domain-knowledge from previous
works for object classification [29], in CIFAR-10 database
2
and spoofing [10], to establish a preliminary topology network
and explore its parameters according to the perception of
the problem. These steps are explained in Section IV-A and
employed in Section V.
The final layer of the convolutional network outputs a
deep image representation. For classification, we use a fully-
connected three-layered network [41]. We discard the use
2
http://www.cs.toronto.edu/
∼
kriz/cifar.html
of unshared local layers, since the literature [10] has shown
that they are inappropriate to problems in which the object
structure is irrelevant. The last network contains only three
neurons (one for each class) and classification is performed
by softmax regression. Then, the weights of each layer in
both networks are learned by the well-known backpropagation
algorithm.
The framework described above appears in CUDA-convnet
library implemented in C++ / CUDA by Krizhevky
3
. It is
important to highlight that such networks are a longstanding
approach, but it has recently enabled significant advances in
computer vision and pattern recognition fields, due to the
availability of more data and processing power, as well as
a better understanding of the learning process [21], [29], [44].
A. Methodology
The development of a network architecture for the three-
class detection problem, involving textured, soft, and no
contact lens images, starts from the Spoofnet — a network
specially developed to address the two-class detection problem
of fake and live images [10]. From this network, we determine
the range of the parameter values to evaluate and understand
their influence on the performance of the contact lens detection
method. These parameters are related to four groups: (i) the
training methodology; (ii) the network architecture; (iii) the
input image size; (iv) the database annotation. These groups
are described in more details next.
Training methodology: We follow the training methodol-
ogy established in [29] and described in
4
. An initial learning
rate (LR) must be chosen. It is set to 10
−3
in [29] and to
10
−4
in [10]. We analyze both values.
Given an initial number of epochs, we develop the following
steps in order to train a network: (1) train 100% of epochs in
three out of four batches of the training data, using the fourth
one as a validation set; (2) train more 40% of epochs in all
four batches with the same learning rate; (3) train more 10% of
epochs in all training batches by decreasing the LR by a factor
of 10; (4) finally, train more 10% of epochs in all training
batches by decreasing the LR again by a factor of 10. In [29],
this initial number is set to 100, whereas in [10], it is set to 200.
The authors in [10] argue that this parameter is both data and
problem dependent. Then, here, besides evaluating 100 and
3
https://code.google.com/p/cuda-convnet/
4
https://code.google.com/p/cuda-convnet/wiki/Methodology
200 for the initial number of epochs, we propose to evaluate
higher numbers while overfitting is not achieved. After those
steps, we compute the accuracy of the trained network using
the verification data.
Network architecture: Once the training methodology
parameters are defined, we focus on the network topology
definition. In the specification of the network architecture,
one can use several layer and operation details
5
, although
here we evaluate: the number of convolutional layers of the
networks: {1, 2, 3}, that is, networks with one, two, or three
convolutional layers – the number of fully-connected layers
is fixed in only one layer per class in order to reduce the
number of possibilities to be evaluated; the use or not of
normalization operation on top of each layer is also evaluated;
the number of filters in each layer is also evaluated – com-
bination of {16, 32, 64} filters are evaluated for one, two and
three layers. The window sizes of the convolutional, pooling
and normalization operations are kept identical to the ones of
Spoofnet.
Input image dimension: After finding the best network
architecture, we investigate the influence of the input image
size. We evaluate different image sizes, i.e., 64×64, 128×128
and 256 ×256 pixels, given that for lower values than 64×64,
the contact lens details are not visible, whereas for higher
values than 512×512, oversampling is performed and memory
issues arise. To obtain images with the proposed dimensions,
we resize them.
A very important aspect, that also affects the input image
size, is data augmentation. It is strongly recommended to
reduce overfitting. In Krizhevsky’s framework [29], given an
input image, it is possible to define a window size such that
five image patches are cropped out from the original image.
We define the border in pixels to be cropped out from the
image. For instance, for a 64 × 64 input image, we consider
the cropped image (a window) with 56×56 pixels at its center
and we also slide this central window of 4 pixels horizontally
and vertically to get cropped images from the four corners of
the original image. We also apply reflections on each of the
five images such that this procedure on each original image
ends up to 10 training images. Here, we propose to evaluate
crop border values of {2, 4, 6, 8} for 64×64, {4, 8, 12, 16} for
128×128, and {8, 16, 24, 32} for 256 ×256 image sizes. Note
that the crop border values respect a proportion regarding the
image size.
Database annotation: As previously mentioned in Sec-
tion III, the NDCL database (with images from the AD100 and
LG4000 sensors) come with annotations for the pupil and iris
locations, i.e., the x and y coordinates and the ratio, allowing
a perfect iris segmentation or, in our case, only a perfect iris
location, since we use squared region crops. For these datasets,
through these annotations, we consider to use the iris image
region plus a percentage of the background and define the
following value: 0% (without), 10%, 20%, 30%, and 40%, in
order to evaluate the importance of background addition.
5
https://code.google.com/p/cuda-convnet/wiki/LayerParams
Fig. 3. Spoofnet - initial network topology used here. Source: [10].
V. EXPERIMENTS AND RESULTS
In this section, we present the experiments performed in
this work. We start by evaluating the groups of parameters
established in the previous section to study their behavior
and to obtain a performing network topology for contact lens
detection called CLDnet (see Fig. 4). Then, we compare the
effectiveness of our proposed approach with state-of-the-art
results in different scenarios.
A. Parameter Evaluation
As established in Section IV-B, we have to first evaluate
parameters in order to analyze their influence in the effective-
ness of the proposed method and also design a robust network
topology. These experiments were conducted separately only
on NDCL database, namely AD100 and LG4000 sensors, since
the iris location is available.
As initial network topology, we consider the one used
in [10], i.e., Spoofnet. Its configurations are illustrated in Fig 3.
We also consider image size of 128 × 128 and crop border of
8 pixels (indeed, input images are 112 × 112 pixels) – values
used in Spoofnet. Furthermore, 10% of background addition
was selected, before cropping for generating the initial input
images. The 10% value was decided through visual inspection
on the images and verifying that this amount suffices to include
the contact lens borders in the cropped iris image.
The first evaluation is on the training methodology. We
verified that, for initial learning rate of 10
−3
, the framework
caught/crashed in early iterations/epochs. This probably oc-
curred because the learning rate was too aggressive. Then, for
all remaining experiments, an initial learning rate of 10
−4
was
used. We start the initial number of epochs in 100, however,
we also tested 200, 300 and 400 epochs. When 400 epochs
were evaluated, we observed that the learning process was
overfitting in the validation batch of the training set, then we
decide to take 300 as the initial number of epochs, since the
learning process was still achieving generalization. Therefore,
this is defined as our evaluation protocol for the remaining
experiments.
Then, we evaluate the network architecture parameters.
We evaluate some configurations by varying the number of
layers and the number of filters in each layer. The results in
correct classification rate (CCR) are shown in Table II.
TABLE II
NETWORK ARCHITECTURE EVALUATION FOR AD100 AND LG4000
SENSORS ON NDCL DATABASE - VARYING THE NUMBER OF LAYERS AND
THE NUMBER OF FILTERS IN EACH LAYER.
Sensor N. Filters CCR N. Filters CCR
AD100
16 72.33 16-16-16 73.67
32 68.67 16-16-32 76.00
64 70.00 16-16-64 77.00
16-16 75.67 16-32-16 72.33
16-32 75.00 16-32-32 76.33
16-64 74.67 16-32-64 71.00
32-32 76.00 32-32-16 75.00
32-64 76.00 32-32-64 79.67
LG4000
16 79.50 16-16-16 77.59
32 77.34 16-16-32 83.34
64 80.84 16-16-64 81.17
16-16 84.34 16-32-16 82.92
16-32 84.82 16-32-32 81.75
16-64 84.17 16-32-64 76.92
32-32 85.59 32-32-16 81.34
32-64 85.00 32-32-64 83.75
Note that the use of three convolution layers does not
increase significantly the method effectiveness and the network
using a single layer does not present promising results. The
best result for the AD100 sensor (79.67%) is obtained using
a three layers of convolutions, while only two layers yielded
the best result for LG4000 sensor (85.59%). For our CLDnet,
we kept the configuration of two layers using 32 and 64 filters
for the first and second layers, respectively, since the results
seemed more stable for both sensors of the NDCL database.
We also evaluate whether or not to use the normalization
operation, but the results demonstrated that the method ef-
fectiveness is insensitive to this operation in the contact lens
detection problem. Thus, this operation was removed from the
CLDnet.
Finally, we evaluate the input image size and database
annotations parameters simultaneously. The results of these
experiments are shown in Table III. By observing these results,
we can conclude that, in general, the results achieved by the
larger input image size, i.e., 256 × 256 pixels, correspond
to the worst CCRs for both sensors, AD100 and LG4000.
Additionally, on average, the results reported when using input
image dimensions of 64×64 and 128 ×128 pixels are slightly
Fig. 4. CLDnet - network for Contact Lens Detection proposed here.
similar. As the image size might be a constraint in some
applications, we prefer the smallest one for image input in
our CLDnet. Moreover, the best results presented in Table III
are obtained by networks with 64 × 64 pixel images as input,
4 pixels for crop border, and 10% of background addition,
such that the final designed CLDnet shown in Fig. 4. Despite
that fact, there is no strong claim to be stated for crop border
and background addition parameters. That is, the results vary
significantly when analyzing these two parameters.
B. Results
In this section, we compare the results obtained with our
method against the state-of-the-art (SOTA) results in [30].
Tables IV, V and VI present CCRs for no (N), textured
(T) and soft (S) contact lens classes and the overall (O)
CCR when analyzing intra, inter, and multi-sensor evaluations,
respectively. These results are analyzed as follows.
It is important to note that for the experiments run in the
sensors of the IIIT-D database, we use the same network,
CLDnet, however, we had to adjust the initial learning rate to
10
−3
, because 10
−4
was not sufficient for effective learning.
All the remainder parameters and procedures were maintained
as the sensors of the NDCL database.
1) Intra-sensor evaluation: It is possible to observe that
the proposed method outperformed SOTA for AD100 &
LG4000 sensors in the NDCL database, in which iris loca-
tion is available, therefore, establishing new SOTA results.
A marginal improvement is observed for the AD100 sensor
images, however, for the LG4000 sensor, the CCR raises from
approximately 80% to 86%, an improvement of 30%. We can
TABLE III
NETWORK ARCHITECTURE EVALUATION FOR AD100 AND LG4000 SENSORS ON NDCL DATABASE - EVALUATING THE INPUT IMAGE SIZE, THE CROP
BORDER PARAMETER USED IN THE DATA AUGMENTATION, AND THE BACKGROUND ADDITION FROM THE DATABASE ANNOTATIONS.
Input image size & Crop borders
Sensor
Background 64 × 64 128 × 128 256 × 256
addition (%) 2 4 6 8 4 8 12 16 8 16 24 32
AD100
0 74.67 73.67 71.00 71.00 74.00 78.00 70.67 70.00 70.33 70.67 71.33 63.33
10 74.67 78.33 74.00 73.33 73.33 76.00 72.67 65.33 71.33 73.67 68.00 62.33
20 71.33 76.67 76.00 67.33 69.67 75.33 76.33 68.00 71.33 71.00 73.00 68.33
30 69.00 70.00 72.67 75.00 68.33 72.33 73.00 75.33 67.33 70.00 72.00 67.00
40 66.33 69.67 72.67 69.67 73.33 71.67 71.00 68.33 66.67 69.67 75.67 68.33
LG4000
0 82.50 81.92 82.75 82.08 84.25 83.83 84.17 82.08 77.25 76.50 76.00 77.00
10 83.25 86.00 84.25 82.75 84.58 84.58 85.25 82.92 72.25 75.58 76.25 75.08
20 81.25 82.83 84.08 80.58 84.75 84.83 85.58 84.33 72.17 74.92 75.83 73.58
30 82.00 81.92 82.50 80.00 83.42 82.83 84.33 84.42 71.08 70.42 74.50 71.33
40 80.25 81.42 81.50 82.08 82.17 82.92 84.92 82.67 68.00 70.58 72.08 71.08
also see comparable results to SOTA for the Cogent & Vista
sensor images in the IIIT-D database. In this case, iris location
is not provided and the entire eye image was used as input to
our method. Nonetheless, our method achieves higher results
than the second best performing methods reported in [30].
The results on the IIIT-D database can be better understood,
when we consider that SOTA counts with an iris segmentation
algorithm.
TABLE IV
INTRA-SENSOR RESULTS FOR THE NDCL AND IIIT-D DATABASES.
Sensors
AD100 LG4000 Cogent Vista
Ours SOTA Ours SOTA Ours SOTA 2nd Ours SOTA 2nd
N 73.00 81.00 84.50 76.21 35.50 66.83 59.73 60.80 76.21 49.49
T 97.00 100.00 99.75 91.62 73.00 94.91 91.87 55.88 91.62 99.42
S 65.00 52.00 73.75 67.52 98.21 56.66 52.84 98.30 67.52 59.32
O 78.33 77.67 86.00 80.04 69.05 73.01 68.57 72.08 80.04 69.84
2) Inter-sensor evaluation: Again, our method achieved
new SOTA results in this scenario for the NDCL database,
improving the CCR in 18% and 15%. In one sense, this
result highlights how robust deep representations can be when
learning features directly from the data. In contrast, disastrous
results for the IIIT-D database were achieved due to the
absence of iris location — a feature that comes with SOTA.
TABLE V
INTER-SENSOR RESULTS FOR THE NDCL AND IIIT-D DATABASES.
Sensors
Train AD100 LG4000 Cogent Vista
Test LG4000 AD100 Vista Cogent
Ours SOTA Ours SOTA Ours SOTA Ours SOTA
N 75.00 62.25 80.00 74.00 6.00 62.10 48.67 65.99
T 94.00 88.50 97.00 93.00 89.61 92.95 38.15 80.81
S 65.00 29.50 49.00 17.00 45.47 75.44 42.25 48.31
O 78.00 60.08 75.33 61.33 45.51 77.79 43.08 65.29
3) Multi-sensor evaluation: Finally, we observe that the
CCRs obtained by our method outperforms the SOTA results
in almost 10% in the multi-sensor scenario for the NDCL
database and, even though the iris location is not provided (for
the IIIT database), a comparable performance is achieved.
TABLE VI
MULTI-SENSOR RESULTS FOR THE NDCL AND IIIT-D DATABASES.
Databases
NDCL IIIT
Ours SOTA Ours SOTA
N 77.40 72.60 47.55 62.14
T 99.60 97.00 61.07 94.74
S 71.40 50.00 97.99 61.63
O 82.80 73.20 69.28 72.96
C. Architecture learning and processing times
In our experiments, we used six PCs with 32GB RAM, Intel
Core i7 CPUs, and NVIDIA GPUs (Tesla K40 with 12GB
or GTX GeForce Titan Black with 6GB). The framework
(CUDA-convnet) clearly relied on GPUs and the processing
time of the different GPUs was not significant. The training
time taken by the convolutional networks is highly dependent
on the input image size, number of layers, and other parame-
ters. For image sizes 256 × 256, 128 × 128, and 64 × 64, the
average training time was less than 172, 49, and 11 minutes,
respectively, for the LG4000 sensor – the one with the highest
number of training samples.
Although we did not measure the classification time of a
single sample, our approach is quite suitable for real-world
applications. Indeed, there is an optimized framework, Jetpacs
iOS Deep Belief image recognition framework [45], that
implements the convolutional network architecture described
in [29], which can classify a 256 × 256 image in one among
1k categories in less than 300ms on an iPhone 5S. That
architecture is significantly larger and more complex than
the ones that we propose here. Our architectures comprise
fewer operations and layers, and use lower resolution images,
64 × 64 pixels. Therefore, contact lens detection systems with
architectures developed by using [45] should be suitable for
real-world applications.
VI. CONCLUSIONS AND FUTURE WORK
In this paper, we proposed the use of deep image represen-
tations, by means of learning weights in convolutional network
followed by a classification network, for the iris contact lens
detection problem. The conducted experiments validate our
method, which could achieve a 30% performance gain over
the state-of-the-art approach, SOTA, on the NDCL database
and comparable results on the IIIT-D database. In NDCL,
iris location is available, which allows to create deep image
representations of regions of interest with mostly iris pixels.
This becomes a problem in the IIIT-D database, where neither
iris segmentation nor location is available. SOTA performs
iris segmentation, but our approach is not prepared yet to
preprocess images and segment/locate the iris. We intend to
add this feature in future work and also to evaluate deep
learning techniques in which the architecture of the network
is first learned by using filters with random weights. Once
the architecture is learned, the weights can be improved by
backpropagation.
Effective comprehension and exploitation of representations
built through deep learning techniques, such as the convolu-
tional networks, are still open problems in the literature. We
also plan to put more effort on this subject to clarify such
points.
ACKNOWLEDGMENTS
We thank UFOP, Brazilian National Research Council –
CNPq (Grants 307010/2014-7, 302970/2014-2, 479070/2013-
0, 307113/2012-4), S
˜
ao Paulo Research Foundation – FAPESP,
(Grants 2011/22749-8 and 2013/04172-0). D. Menotti thanks
NVIDIA for donating two GeForce GTX Titan Black with
6GB each.
REFERENCES
[1] J. G. Daugman, “High Confidence Visual Recognition of Persons by a
Test of Statistical Independence,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 15, no. 11, pp. 1148–1161, 1993.
[2] L. Flom and A. Safir, “Iris Recognition System,” U.S. US Patent 4 641
394, 1987.
[3] K. W. Bowyer, K. Hollingsworth, and P. J. Flynn, “Image Understanding
for Iris Biometrics: A Survey,” Computer Vision and Image Understand-
ing, vol. 110, no. 2, pp. 281–307, 2008.
[4] Y. Song, W. Cao, and Z. He, “Robust Iris Recognition using Sparse
Error Correction Model and Discriminative Dictionary Learning,” Neu-
rocomputing, vol. 137, pp. 198–204, 2014.
[5] A. F. M. Raffei, H. Asmuni, R. Hassan, and R. M. Othman, “Feature
Extraction for Different Distances of Visible Reflection Iris using
Multiscale Sparse Representation of Local Radon Transform,” Pattern
Recognition, vol. 46, no. 10, pp. 2622–2633, 2013.
[6] J. Galbally, S. Marcel, and J. Fierrez, “Image Quality Assessment for
Fake Biometric Detection: Application to Iris, Fingerprint, and Face
Recognition,” IEEE Trans. on Image Processing, vol. 23, no. 2, pp.
710–724, 2014.
[7] P. Gupta, S. Behera, M. Vatsa, and R. Singh, “On Iris Spoofing using
Print Attack,” in 22nd Int. Conf. on Pattern Recognition. IEEE, 2014,
pp. 1681–1686.
[8] Z. Sun and T. Tan, “Iris Anti-Spoofing,” in Handbook of Biometric Anti-
Spoofing, ser. Advances in Computer Vision and Pattern Recognition,
S. Marcel, M. S. Nixon, and S. Z. Li, Eds. Springer London, 2014,
pp. 103–123.
[9] K. W. Bowyer and J. S. Doyle, “Cosmetic Contact Lenses and Iris
Recognition Spoofing,” Computer, vol. 47, no. 5, pp. 96–98, 2014.
[10] D. Menotti, G. Chiachia, A. Pinto, W. Schwartz, H. Pedrini, A. Falc
˜
ao,
and A. Rocha, “Deep Representations for Iris, Face, and Fingerprint
Spoofing Detection,” IEEE Trans. on Information Forensics and Secu-
rity, vol. 10, no. 4, pp. 864–879, 2015.
[11] R. Raghavendra and C. Busch, “Robust Scheme for Iris Presentation
Attack Detection Using Multiscale Binarized Statistical Image Features,”
IEEE Trans. on Information Forensics and Security, vol. 10, no. 4, pp.
703–715, 2015.
[12] Z. Sun, H. Zhang, T. Tan, and J. Wang, “Iris Image Classification Based
on Hierarchical Visual Codebook,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 36, no. 6, pp. 1120–1133, Jun. 2014.
[13] A. Sequeira, H. Oliveira, J. Monteiro, J. Monteiro, and J. Cardoso,
“MobILive 2014 - Mobile Iris Liveness Detection Competition,” in IEEE
Int. Joint Conf. on Biometrics, Sept 2014, pp. 1–6.
[14] Z. Wei, X. Qiu, Z. Sun, and T. Tan, “Counterfeit Iris Detection based on
Texture Analysis,” in Int. Conf. on Pattern Recognition. IEEE, 2008,
pp. 1–4.
[15] S. E. Baker, A. Hentz, K. W. Bowyer, and P. J. Flynn, “Degradation of
Iris Recognition Performance due to non-Cosmetic Prescription Contact
Lenses,” Computer Vision and Image Understanding, vol. 114, no. 9,
pp. 1030–1044, 2010.
[16] H. Zhang, Z. Sun, and T. Tan, “Contact Lens Detection based on
Weighted LBP,” in Int. Conf. on Pattern Recognition, 2010, pp. 4279–
4282.
[17] N. Kohli, D. Yadav, M. Vatsa, and R. Singh, “Revisiting Iris Recognition
with Color Cosmetic Contact Lenses,” in Int. Conf. on Biometrics, 2013,
pp. 1–7.
[18] J. S. Doyle, K. W. Bowyer, and P. J. Flynn, “Variation in Accuracy of
Textured Contact Lens Detection based on Sensor and Lens Pattern,”
in IEEE Int. Conf. on Biometrics: Theory, Applications, and Systems,
2013, pp. 1–7.
[19] J. Komulainen, A. Hadid, and M. Pietikainen, “Generalized Textured
Contact Lens Detection by Extracting BSIF Description from Cartesian
Iris Images,” in IEEE Int. Joint Conf. on Biometrics, 2014, pp. 1–7.
[20] Y. Bengio, A. Courville, and P. Vincent, “Representation Learning: A
Review and New Perspectives,” IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 35, no. 8, 2013.
[21] F. Schroff, D. Kalenichenko, and J. Philbin, “FaceNet: A Unified
Embedding for Face Recognition and Clustering,” in IEEE Conf. on
Computer Vision and Pattern Recognition (CVPR), 2015, pp. 132–142,
to appear.
[22] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, “DeepFace: Closing
the Gap to Human-Level Performance in Face Verification,” in IEEE
Int. Conf. on Computer Vision and Pattern Recognition, 2014, pp. 1701–
1708.
[23] G. Chiachia, A. X. Falc
˜
ao, N. Pinto, A. Rocha, and D. Cox, “Learning
Person-Specific Representations From Faces in the Wild,” IEEE Trans.
on Information Forensics and Security, vol. 9, no. 12, pp. 2089–2099,
Dec. 2014.
[24] D. Cox and N. Pinto, “Beyond Simple Features: A Large-Scale Feature
Search Approach to Unconstrained Face Recognition,” in IEEE Int.
Conf. on Automatic Face Gesture Recognition and Workshops. IEEE,
2011, pp. 8–15.
[25] P. Sermanet, K. Kavukcuoglu, S. Chintala, and Y. LeCun, “Pedestrian
Detection with Unsupervised Multi-Stage Feature Learning,” in IEEE
Conf. on Computer Vision and Pattern Recognition. IEEE, 2013, pp.
3626–3633.
[26] D. Menotti, G. Chiachia, A. Falc
˜
ao, and V. Oliveira Neto, “Vehicle
License Plate Recognition with Random Convolutional Networks,” in
27th SIBGRAPI Conf. on Graphics, Patterns and Images, 2014, pp. 298–
303.
[27] D. C. Ciresan, U. Meier, L. M. Gambardella, and J. Schmidhuber, “Deep
Big Simple Neural Nets For Handwritten Digit Recognition,” Neural
Computation, vol. 22, no. 12, pp. 3207–3220, 2010.
[28] D. Cires¸an, U. Meier, J. Masci, and J. Schmidhuber, “Multi-Column
Deep Neural Network for Traffic Sign Classification,” Neural Networks,
vol. 32, pp. 333–338, 2012.
[29] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet Classification
with Deep Convolutional Neural Networks,” in Advances in neural
information processing systems, 2012, pp. 1097–1105.
[30] D. Yadav, N. Kohli, J. Doyle, R. Singh, M. Vatsa, and K. Bowyer,
“Unraveling the Effect of Textured Contact Lenses on Iris Recognition,”
IEEE Trans. on Information Forensics and Security, vol. 9, no. 5, pp.
851–862, 2014.
[31] S. E. Baker, A. Hentz, K. W. Bowyer, and P. J. Flynn, “Contact
Lenses: Handle with Care for Iris Recognition,” in IEEE Int. Conf. on
Biometrics: Theory, Applications, and Systems, 2009, pp. 1–8.
[32] E. C. Lee, K. R. Park, and J. Kim, “Fake Iris Detection by using Purkinje
Image,” in Advances in Biometrics. Springer, 2005, pp. 397–403.
[33] Chinese Academy of Sciences (CASIA), Insti-
tute of Automation, “CASIA Iris Image Database,”
http://biometrics.idealtest.org/findTotalDbByMode.do?mode=Iris,
2010, accessed 26 Mar, 2015 [online].
[34] University of Bath, Department of Electronic and Electrical Engineering,
“University of Bath Iris Image Database,” 2008.
[35] Z. He, Z. Sun, T. Tan, and Z. Wei, “Efficient Iris Spoof Detection via
Boosted Local Binary Patterns,” in Advances in Biometrics. Springer,
2009, pp. 1080–1090.
[36] Chinese Academy of Sciences (CASIA), Institute of
Automation, “CASIA-IrisV3 Image Database [Online],”
http://biometrics.idealtest.org/dbDetailForUser.do?id=3, 2010, accessed
26 Mar, 2015 [online].
[37] National Institute of Standards and Technology (NIST), “Iris Challenge
Evaluation (ICE),” http://www.nist.gov/itl/iad/ig/ice.cfm, 2008, accessed
26 Mar, 2015 [online].
[38] J. Daugman, “Demodulation by Complex-Valued Wavelets for Stochastic
Pattern Recognition,” Int. Journal of Wavelets, Multiresolution and
Information Processing, vol. 1, no. 1, pp. 1–17, 2003.
[39] J. Doyle and B. Kevin, “Notre Dame Image Database for Con-
tact Lens Detection In Iris Recognition-2013: README,” Available:
http://www3.nd.edu/ cvrl/papers/CosCon2013README.pdf, 2014.
[40] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based
Learning Applied to Document Recognition,” Proceedings of the IEEE,
vol. 86, no. 11, pp. 2278–2324, 1998.
[41] C. Bishop, Neural Networks for Pattern Recognition, ser. Advanced
Texts in Econometrics. Clarendon Press, 1995.
[42] W. S. Geisler and D. G. Albrecht, “Cortical Neurons: Isolation of
Contrast Gain Control,” Vision Research, vol. 32, no. 8, pp. 2429–2454,
1992.
[43] N. Pinto, D. Doukhan, J. J. DiCarlo, and D. D. Cox, “A High-
Throughput Screening Approach to Discovering Good Forms of
Biologically-Inspired Visual Representation,” PLoS Computational Bi-
ology, vol. 5, no. 11, 2009.
[44] J. Bergstra, D. Yamins, and D. D. Cox, “Making a Science of Model
Search: Hyperparameter Optimization in Hundreds of Dimensions for
Vision Architectures,” in Int. Conf. on Machine Learning), 2013.
[45] P. Wardern, “The SDK for Jetpac’s iOS Deep Belief image recognition
framework,” 2014. [Online]. Available: https://github.com/jetpacapp/
DeepBeliefSDK
