Periocular Recognition under Unconstrained Settings with Universal
Background Models
Jo
˜
ao C. Monteiro and Jaime S. Cardoso
INESC TEC and Faculdade de Engenharia, Universidade do Porto, Porto, Portugal
Keywords:
Biometrics, Periocular Recognition, Universal Background Model, SIFT, Gaussian Mixture Models.
Abstract:
The rising challenges in the fields of iris and face recognition are leading to a renewed interest in the area.
In recent years the focus of research has turned towards alternative traits to aid in the recognition process
under less constrained image acquisition conditions. The present work assesses the potential of the periocular
region as an alternative to both iris and face in such scenarios. An automatic modeling of SIFT descriptors,
regardless of the number of detected keypoints and using a GMM-based Universal Background Model method,
is proposed. This framework is based on the Universal Background Model strategy, first proposed for speaker
verification, extrapolated into an image-based application. Such approach allows a tight coupling between
individual models and a robust likelihood-ratio decision step. The algorithm was tested on the UBIRIS.v2 and
the MobBIO databases and presented state-of-the-art performance for a variety of experimental setups.
1 INTRODUCTION
Over the past few years face and iris have been on
the spotlight of many research works in biometrics.
The face is a easily acquirable trait with a high de-
gree of uniqueness, while the iris, the coloured part
of the eye, presents unique textural patterns resulting
from its random morphogenesis during embryonic de-
velopment (Bakshi et al., 2012). These marked ad-
vantages, however, fall short when low-quality im-
ages are presented to the system. It has been noted
that the performance of iris and face recognition al-
gorithms is severely compromised when dealing with
non-ideal scenarios such as non-uniform illumination,
pose variations, occlusions, expression changes and
radical appearance changes (Bakshi et al., 2012; Bod-
deti et al., 2011). Several recent works have tried to
explore alternative hypothesis to face this problem, ei-
ther by developing more robust algorithms or by ex-
ploring new traits to allow or aid in the recognition
process (Woodard et al., 2010).
The periocular region is one of such unique traits.
Even though a true definition of the periocular region
is not standardized, it is common to describe it as the
region in the immediate vicinity of the eye (Padole
and Proenca, 2012; Smereka and Kumar, 2013). Peri-
ocular recognition can be motivated as a middle point
between face and iris recognition. It has been shown
to present increased performance when only degraded
Figure 1: Example of periocular regions from both eyes,
extracted from a face image (Woodard et al., 2010).
facial data (Miller et al., 2010b) or low quality iris im-
ages (Bharadwaj et al., 2010; Tan and Kumar, 2013)
are made available, as well as promising results as
a soft biometric trait to help improve both face and
iris recognition systems in less constrained acquisi-
tion environments (Joshi et al., 2012).
In this work we present a new approach to peri-
ocular recognition under less ideal acquisition con-
ditions. Our proposal is based on the idea of max-
imum a posteriori adaptation of Universal Back-
ground Model, as proposed by Reynolds for speaker
verification (Reynolds et al., 2000). We evaluate the
proposed algorithm on two datasets of color periocu-
lar images acquired under visible wavelength (VW)
illumination. Multiple noise factors such as vary-
ing gazes/poses and heterogeneous lighting condi-
tions are characteristic to such images, thus represent-
ing the main challenge of the present work.
The remainder of this paper is organized as fol-
38
C. Monteiro J. and S. Cardoso J..
Periocular Recognition under Unconstrained Settings with Universal Background Models.
DOI: 10.5220/0005195900380048
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2015), pages 38-48
ISBN: 978-989-758-069-7
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
lows: Section 2 summarizes relevant works in perioc-
ular recognition; Section 3 offers a detailed analysis
of the proposed algorithm; Section 4 presents the ob-
tained results and, finally, the main conclusions and
future prospects are summarized in Section 5.
2 RELATED WORK
Periocular biometrics is a recent area of research, pro-
posed by the first time in a feasibility study by Park et
al. (Park et al., 2009). In this pioneer work, the au-
thors suggested the periocular region as a potential
alternative to circumvent the significant challenges
posed to iris recognition systems working under un-
constrained scenarios. The same authors analysed the
effect of degradation on the accuracy of periocular
recognition (Park et al., 2011). Performance assess-
ment under less constrained scenarios is also the goal
of the work by Miller et al. (Miller et al., 2010a),
where factors such as blur and scale are shown to
have a severe effect on the performance of perioc-
ular recognition. Padole and Proenc¸a (Padole and
Proenca, 2012) also explore the effect of scale, pig-
mentation and occlusion, as well as gender, and pro-
pose an initial region-of-interest detection step to im-
prove recognition accuracy.
Ross et al. (Ross et al., 2012) explored informa-
tion fusion based on several feature extraction tech-
niques, to handle the significant variability of input
periocular images. Information fusion has become
one of the trends in biometric research in recent years
and periocular recognition is no exception. Bharad-
waj et al. (Bharadwaj et al., 2010) proposed fusion
of matching scores from both eyes to improve the in-
dividual performance of each of them. On the other
hand, Woodard et al. (Woodard et al., 2010) place fu-
sion at the feature level, using color and texture infor-
mation.
Some works have explored the advantages of the
periocular region as an aid to more traditional ap-
proaches based on iris. Boddeti et al. (Boddeti
et al., 2011) propose the score level fusion of a tradi-
tional iris recognition algorithm, based on Gabor fea-
tures, and a periocular probabilistic approach based
on optimal trade-off synthetic discriminant functions
(OTSDF). A similar work by Joshi et al. (Joshi et al.,
2012) proposed feature level fusion of wavelet coeffi-
cients and LBP features, from the iris and periocular
regions respectively, with considerable performance
improvement over both singular traits. A recent work
by Tan et al. (Tan and Kumar, 2013) has also explored
the benefits of periocular recognition when highly de-
graded regions result from the traditional iris segmen-
tation step. The authors have observed discouraging
performance when the iris region alone is considered
in such scenarios, whereas introducing information
from the whole periocular region lead to a significant
improvement.
Two other recent and relevant works by Moreno et
al. (Moreno et al., 2013b; Moreno et al., 2013a) ex-
plore the well-known approach of sparse representa-
tion classification in the scope of the specific problem
of periocular recognition. A thorough review of the
most relevant method in recent years concerning pe-
riocular recognition and its main advantages can be
found in the work by Santos and Proenc¸a (Santos and
Proenc¸a, 2013).
On the present work we propose a new approach
to periocular recognition, based on a general frame-
work with proven results in voice biometrics. We
explore a strategy based on the adaptation of a Uni-
versal Background Model (UBM) to achieve faster
and more robust training of individual specific mod-
els. With such idea in mind we aim not only to de-
sign a high performance recognition system but also
to assess the versatility and robustness of the UBM
strategy for biometric traits other than voice.
3 PROPOSED METHODOLOGY
3.1 Algorithm Overview
The proposed algorithm is schematically represented
in Figure 2. The two main blocks - enrollment and
recognition - refer to the typical architecture of a bio-
metric system. During enrollment a new individual’s
biometric data is inserted into a previously existent
database of individuals. Such database is probed dur-
ing the recognition process to assess either the valid-
ity of an identity claim - verification - or the k most
probable identities - identification - given an unknown
sample of biometric data.
During the enrollment, a set of N models describ-
ing the unique statistical distribution of biometric fea-
tures for each individual n ∈ {1, ..., N} is trained by
maximum a posteriori (MAP) adaptation of an Uni-
versal Background Model (UBM). The UBM is a
representation of the variability that the chosen bio-
metric trait presents in the universe of all individ-
uals. MAP adaptation works as a specialization of
the UBM based on each individual’s biometric data.
The idea of MAP adaptation of the UBM was first
proposed by Reynolds (Reynolds et al., 2000), for
speaker verification, and will be further motivated in
the following sections.
The recognition phase is carried out through the
PeriocularRecognitionunderUnconstrainedSettingswithUniversalBackgroundModels
39
Figure 2: Graphical representation of the main steps in both
the enrollment and recognition (verification and identifica-
tion) phases of the proposed periocular recognition algo-
rithm.
projection of the features extracted from an unknown
sample onto both the UBM and the individual spe-
cific models (IDSM) of interest. A likelihood-ratio
between both projections outputs the final recognition
score. Depending on the functioning mode of the sys-
tem - verification or identification - decision is carried
out by thresholding or maximum likelihood-ratio re-
spectively.
3.2 Universal Background Model
Universal background modeling is a common strategy
in the field of voice biometrics (Povey et al., 2008). It
can be easily understood if the problem of biometric
verification is interpreted as a basic hypothesis test.
Given a biometric sample Y and a claimed ID, S, we
define:
H
0
: Y belongs to S
H
1
: Y does not belong to S
as the null and alternative hypothesis, respectively.
The optimal decision is taken by a likelihood-ratio
test:
p(Y |H
0
)
p(Y |H
1
)
(
≥ θ accept H
0
≤ θ accept H
1
(1)
where θ is the decision threshold for accepting or re-
jecting H
0
, and p(Y |H
i
),i ∈
{
0,1
}
is the likelihood of
observing sample Y when we consider hypothesis i to
be true.
Biometric recognition can, thus, be reduced to the
problem of computing the likelihood values p(Y |H
0
)
and p(Y |H
1
). It is intuitive to note that H
0
should
be represented by a model λ
hyp
that characterizes the
hypothesized individual, while, alternatively, the rep-
resentation of H
1
, λ
hyp
, should be able to model all
the alternatives to the hypothesized individual.
From such formulation arises the need for a model
that successfully covers the space of alternatives to
the hypothesized identity. The most common desig-
nation in literature for such a model is universal back-
ground model or UBM (Reynolds, 2002). Such model
must be trained on a large set of data, so as to faith-
fully cover a representative user space and a signifi-
cant amount of sources of variability. The following
section details the chosen strategy to model λ
hyp
and
how individual models, λ
hyp
, can be adapted from the
UBM in a fast and robust way.
3.3 Hypothesis Modeling
On the present work we chose Gaussian Mixture
Models (GMM) to model both the UBM, i.e. λ
hyp
,
and the individual specific models (IDSM), i.e. λ
hyp
.
Such models are capable of capturing the empirical
probability density function (PDF) of a given set of
feature vectors, so as to faithfully model their intrin-
sic statistical properties (Reynolds et al., 2000). The
choice of GMM to model feature distributions in bio-
metric data is extensively motivated in many works
of related areas. From the most common interpre-
tations, GMMs are seen as capable of representing
broad “hidden” classes, reflective of the unique struc-
tural arrangements observed in the analysed biometric
traits (Reynolds et al., 2000). Besides this assump-
tion, Gaussian mixtures display both the robustness
of parametric unimodal Gaussian density estimates,
as well as the ability of non-parametric models to
fit non-Gaussian data (Reynolds, 2008). This dual-
ity, alongside the fact that GMM have the notewor-
thy strength of generating smooth parametric densi-
ties, confers such models a strong advantage as gen-
erative model of choice. For computational efficiency,
GMM models are often trained using diagonal covari-
ance matrices. This approximation is often found in
biometrics literature, with no significant accuracy loss
associated (Xiong et al., 2006).
All models are trained on sets of Scale In-
variant Feature Transform (SIFT) keypoint descrip-
tors (Lowe, 2004). This choice for periocular image
description is thoroughly motivated in literature (Ross
et al., 2012; Park et al., 2011), mainly due to the ob-
servation that local descriptors work better than their
global counterparts when the available data presents
non-uniform conditions. Furthermore, the invariance
of SIFT features to a set of common undesirable fac-
tors (image scaling, translation, rotation and also par-
tially to illumination and affine or 3D projection),
confer them a strong appeal in the area of uncon-
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
40
strained biometrics.
Originally such descriptors are defined in 128 di-
mensions. However, we chose to perform a Principle
Component Analysis (PCA), as suggested in (Shinoda
and Inoue, 2013), reducing the dimensionality to 32.
Such reduction allows not only a significant reduction
in the computational complexity of the training phase,
but also an improved distinctiveness and robustness
to the extracted feature vectors, especially as far as
image deformation is concerned (Ke and Sukthankar,
2004). We computed the principle components from
the same data used to train the UBM.
3.4 H
1
: UBM Parameter Estimation
To train the Universal Background Model a large
amount of “impostor” data, i.e. a set composed of
data from all the enrolled individuals, is used, so as
to cover a wide range of possibilities in the individual
search space (Shinoda and Inoue, 2013). The training
process of the UBM is simply performed by fitting
a k-mixture GMM to the set of PCA-reduced feature
vectors extracted from all the “impostors”.
If we interpret the UBM as an “impostor” model,
its “genuine” counterpart can be obtained by adapta-
tion of the UBM’s parameters, λ
hyp
, using individ-
ual specific data. For each enrolled individual, ID,
an individual specific model (IDSM), defined by pa-
rameters λ
hyp
, is therefore obtained. The adaptation
process will be outlined in the following section.
3.5 H
0
: MAP Adaptation of the UBM
IDSMs are generated by the tuning of the UBM pa-
rameters in a maximum a posteriori (MAP) sense,
using individual specific biometric data. This ap-
proach provides a tight coupling between the IDSM
and the UBM, resulting in better performance and
faster scoring than uncoupled methods (Xiong et al.,
2006), as well as a robust and precise parameter es-
timation, even when only a small amount of data is
available (Shinoda and Inoue, 2013). This is indeed
one of the main advantages of using UBMs. The de-
termination of appropriate initial values (i.e. seeding)
of the parameters of a GMM remains a challenging is-
sue. A poor initialization may result in a weak model,
especially when the data volume is small. Since the
IDSM are learnt only from each individual data, they
are more prone to a poor convergence that the GMM
for the UBM, learnt from a big pool of individuals. In
essence, UBM constitutes a good initialization for the
IDSM.
The adaptation process, as proposed by
Reynolds (Reynolds et al., 2000), resembles the
Expectation-Maximization (EM) algorithm, with
two main estimation steps. The first is similar to the
expectation step of the EM algorithm, where, for
each mixture of the UBM, a set of sufficient statistics
are computed from a set of M individual specific
feature vectors, X = {x
1
...x
M
}:
n
i
=
M
∑
m=1
p(i|x
m
) (2)
E
i
(x) =
1
n
i
M
∑
m=1
p(i|x
m
)x
m
(3)
E
i
(xx
t
) =
1
n
i
M
∑
m=1
p(i|x
m
)x
m
x
t
m
(4)
where p(i|x
m
) represents the probabilistic alignment
of x
m
into each UBM mixture. Each UBM mixture
is then adapted using the newly computed sufficient
statistics, and considering diagonal covariance matri-
ces. The update process can be formally expressed
as:
ˆw
i
= [α
i
n
i
/M + (1 − α
i
)w
i
]ξ (5)
ˆµ
i
= α
i
E
i
(x) + (1 − α
i
)µ
i
(6)
ˆ
Σ
i
= α
i
E
i
(xx
t
) + (1 − α
i
)(σ
i
σ
i
t
+ µ
i
µ
i
t
) − ˆµ
i
ˆµ
i
t
(7)
σ
i
= diag(Σ
i
) (8)
where {w
i
,µ
i
,σ
i
} are the original UBM parameters
and { ˆw
i
, ˆµ
i
,
ˆ
σ
i
} represent their adaptation to a spe-
cific speaker. To assure that
∑
i
w
i
= 1 a weighting
parameter ξ is introduced. The α parameter is a data-
dependent adaptation coefficient. Formally it can be
defined as:
α
i
=
n
i
r + n
i
(9)
where r is generally known as the relevance factor.
The individual dependent adaptation parameter serves
the purpose of weighting the relative importance of
the original values and the new sufficient statistics in
the adaptation process. For the UBM adaptation we
set r = 16, as this is the most commonly observed
value in literature (Reynolds et al., 2000). Most works
propose the sole adaptation of the mean values, i.e.
α
i
= 0 when computing ˆw
i
and
ˆ
σ
i
. This simplifica-
tion seems to bring no nefarious effects over the per-
formance of the recognition process, while allowing
faster training of the individual specific models (Kin-
nunen et al., 2009).
PeriocularRecognitionunderUnconstrainedSettingswithUniversalBackgroundModels
41
3.6 Recognition and Decision
After the training step of both the UBM and each
IDSM, the recognition phase with new data from an
unknown source is somewhat trivial. As referred
in previous sections, the identity check is performed
through the projection of the new test data, X
test
=
{x
t,1
,. .. ,x
t,N
}, where x
t,i
is the i-th PCA-reduced
SIFT vector extracted from the periocular region of
test subject t, onto both the UBM and either the
claimed IDSM (in verification mode) or all such mod-
els (in identification mode). The recognition score is
obtained as the average likelihood-ratio of all key-
point descriptors x
t,i
,∀i ∈ {1..N}. The decision is
then carried out by checking the condition presented
in Equation (1), in the case of verification, or by de-
tecting the maximum likelihood-ratio value for all en-
rolled IDs, in the case of identification.
This is a second big advantage of using UBM.
The ratio between the IDSM and the UBM probabil-
ities of the observed data is a more robust decision
criterion than relying solely on the IDSM probabil-
ity. This results from the fact that some subjects are
more prone to generate high likelihood values than
others, i.e. some people have a more “generic” look
than others. The use of a likelihood ratio with an uni-
versal reference works as a normalization step, map-
ping the likelihood values in accord to their global
projection. Without such step, finding a global op-
timal value for the decision threshold, θ, presented in
Equation 1 would be a far more complex process.
4 EXPERIMENTAL RESULTS
In this section we start by presenting the datasets
and the experimental setups under which performance
was assessed. Further sections present a detailed anal-
ysis regarding the effect of model complexity and fu-
sion of color channels in the global performance of
the proposed algorithm.
4.1 Tested Datasets
The proposed algorithm was tested on two noisy color
iris image databases: UBIRIS.v2 and MobBIO. Even
though both databases were designed in an attempt to
promote the development of robust iris recognition al-
gorithms for images acquired under VW illumination,
their intrinsic properties make them attractive to study
the feasibility of periocular recognition under similar
conditions. The following sections detail their main
features as well as the reasoning behind their choice.
4.1.1 UBIRIS.v2 Database
Images in UBIRIS.v2 (Proenc¸a et al., 2010) database
were captured under non-constrained conditions (at-
a-distance, on-the-move and on the visible wave-
length), with corresponding realistic noise factors.
Figure 3 depicts some examples of these noise fac-
tors (reflections, occlusions, pigmentation, etc.). Two
acquisition sessions were performed with 261 indi-
viduals involved and a total of 11100 300 ×400 color
images acquired. Each individual’s images were ac-
quired at variable distances with 15 images per eye
and per season. Even though the UBIRIS.v2 database
was primarily developed to allow the study of uncon-
strained iris recognition, many works have explored
its potential for periocular-based strategies as an al-
ternative to low-quality iris recognition (Bharadwaj
et al., 2010; Joshi et al., 2012; Padole and Proenca,
2012).
(a) (b) (c) (d)
Figure 3: Examples of noisy image from the UBIRIS.v2
database.
4.1.2 MobBIO Database
The MobBIO multimodal database (Sequeira et al.,
2014) was created in the scope of the 1st Biomet-
ric Recognition with Portable Devices Competition
2013, integrated in the ICIAR 2013 conference. The
main goal of the competition was to compare various
methodologies for biometric recognition using data
acquired with portable devices. We tested our algo-
rithm on the iris modality present on this database.
Regarding this modality the images were captured un-
der two alternative lighting conditions, with variable
eye orientations and occlusion levels, so as to com-
prise a larger variability of unconstrained scenarios.
Distance to the camera was, however, kept constant
for each individual. For each of the 105 volunteers 16
images (8 of each eye) were acquired. These images
were obtained by cropping a single image comprising
both eyes. Each cropped image was set to a 300 ×200
resolution. Figure 4 depicts some examples of such
images.
The MobBIO database presents a face modality
which has also been explored for comparative pur-
poses in the present work. Images were acquired in
similar conditions to those described above for iris
images, with 16 images per subject. Examples of such
images can be observed in Figure 5.
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
42
(a) (b) (c) (d)
Figure 4: Examples of iris images in the MobBIO database.
(a) (b) (c) (d)
Figure 5: Examples of face images in the MobBIO
database.
4.2 Evaluation Metrics
Performance was evaluated for both verification and
identification modes. Regarding the former we ana-
lyzed the equal error rate (EER) and the decidabil-
ity index (DI). The EER is observed at the decision
threshold, θ, where the errors of falsely accepting and
falsely rejecting H
0
occur with equal frequency. The
global behavior of both types of errors is often ana-
lyzed through receiver operating characteristic (ROC)
curves. On the other hand, the DI quantifies the sep-
aration of the “genuine” and “impostor” likelihood
score distributions, as follows:
DI =
|µ
g
− µ
i
|
q
0.5(σ
2
g
+ σ
2
i
)
(10)
where (µ
g
,σ
g
) and (µ
i
,σ
i
) are the mean and standard
deviation of the genuine and impostor score distribu-
tions, respectively.
For identification we analyze cumulative match
curves (CMC). These curves represent the rate of cor-
rectly identified individuals, by checking if the true
identity is present in the N highest ranked identities.
The N parameter is generally referred to as rank. That
allows us to define the rank-1 recognition rate as the
value of the CMC at N = 1.
4.3 Experimental Setups
Our experiments were conducted in three distinct
experimental setups, two of them regarding the
UBIRIS.v2 database and the remaining one the Mob-
BIO database:
1. In the first setup, for the UBIRIS.v2 images, six
samples from 80 different subjects were used,
captured from different distances (4 to 8 meters),
with varying gazes/poses and notable changes in
lighting conditions. One image per individual was
randomly chosen as probe, whereas the remaining
five samples were used for the UBM training and
MAP adaptation. The results were cross-validated
by changing the probe image, per subject, for each
of the six chosen images.
2. Many works on periocular biometrics evaluate
their results using a well-known subset of the
UBIRIS.v2 database, used in the context of the
NICE II competition (Proenc¸a, 2009). This
dataset is divided in train and test subsets, with
a total of 1000 images from 171 individuals. In
the present work we choose to use test subset,
composed by 904 images from 152 individuals.
Only individuals with more than 4 available im-
ages were considered, as 4 images were randomly
chosen for training and the rest for testing. Results
were cross-validated 10-fold. The train dataset
composed by the remaining 96 images from 19
individuals was employed in the parameter opti-
mization step described in further sections.
3. Concerning the MobBIO database, 8 images were
randomly chosen from each of the 105 individ-
uals for the training of the models, whereas the
remaining 8 were chosen for testing. The process
was cross-validated 10-fold. For comparative pur-
poses a similar experiment was carried out on face
images from the same 105 individuals, using the
same 8 + 8 image distribution.
As both databases are composed by color images,
each of the RGB channels was considered individu-
ally for the entire enrollment and identification pro-
cess. For the parameter optimization described in
the next section images were previously converted to
grayscale.
4.4 Parameter Optimization
A smaller dataset, for each database, was also de-
signed to optimize the the number of GMM mixtures
of the trained models. For the UBIRIS.v2 we chose
to work with the well-known train dataset from the
NICE II competition (Proenc¸a, 2009), composed by
96 images from 19 individuals. For the MobBIO
database we chose a total of 50 images from 10 indi-
viduals to perform the previously referred optimiza-
tion. The obtained performance was cross-validated
using a leave-one-out strategy. The chosen metric to
evaluate performance was the rank-1 recognition rate
(R
1
). The evolution of performance with the opti-
mization parameter can be observed in Figure 6. With
such results in mind, the recognition performance for
the experimental setups presented in the last section
was assessed for a number of mixtures M = 128 and
PeriocularRecognitionunderUnconstrainedSettingswithUniversalBackgroundModels
43
M = 64 for the UBIRIS.v2 and MobBIO databases re-
spectively. We choose both these values as the values
where a performance plateau is achieved in the graph
of Figure 6. We chose the lowest possible values for
the parameter M so as to minimize the computational
complexity of the UBM training, which constitutes
the limiting step of the process, without a significant
loss in performance.
Figure 6: Recognition rates obtained with the optimization
subset for variable values of parameter M, number of mix-
tures in the trained GMMs.
4.5 Recognition Results
The results obtained for both databases and experi-
mental setups are represented through ROC and CMC
curves on Figures 7(a) to 7(f). A comparison with
some state-of-the-art algorithms in the UBIRIS.v2
database is also presented in Table 1. In this table re-
sults are grouped according to the experimental setup
of each reported work and also the studied trait: P -
Periocular, I - Iris or P + I - Fusion of both traits.
Besides testing each of the RGB channels indi-
vidually, a simple sum-rule score-level fusion strat-
egy (Kittler et al., 1998) was also considered. It is eas-
ily discernible, from the observation of Figure 7, that
the fusion of information from multiple color chan-
nels brings about a significant improvement in per-
formance for all the tested datasets. When compar-
ing the results obtained with this approach with some
state-of-the-art algorithms a few points deserve fur-
ther discussion. First, the proposed algorithm is capa-
ble of achieving and even surpassing state-of-the-art
performance in multiple experimental setups. Con-
cerning the most common of such setups (2), it is in-
teresting to note that a few works attempted to ex-
plore the UBIRIS.v2 dataset for iris recognition. The
obtained performance has been considered “discour-
aging” in the work by Kumar et al. (Kumar and Chan,
2012). Comparing the rank-1 recognition rate ob-
tained with our algorithm (88.93%) with the 48.1%
reported in the former work, we conclude that the
periocular region may represent a viable alternative
to iris in images acquired under visible wavelength
(VW) illumination. Such acquisition conditions are
known to increase light reflections from the cornea,
resulting in a sub-optimal signal-to-noise ratio (SNR)
in the sensor, lowering the contrast of iris images and
the robustness of the system (Proenc¸a, 2011). More
recent works have explored multimodal approaches,
using combined information from both the iris and
the periocular region. Analysis of Table 1 shows that
none of such works reaches the performance reported
in the present work for the same experimental setup.
Such observation might indicate that most discrimina-
tive biometric information from the UBIRIS.v2 im-
ages might be present in the periocular region, and
that considering data from the very noisy iris regions
might only result in a degradation of the performance
obtained by the periocular region alone.
Concerning the MobBIO database, an alternative
comparison was carried out to analyze the potential of
the periocular region as an alternative to face recogni-
tion. The observed performance for periocular images
was considerably close to that using full-face infor-
mation, with rank-1 recognition rates of 98.98% and
99.77% respectively. These results are an indication
that, under more ideal acquisition conditions, there is
enough discriminative potential in the periocular re-
gion alone to rival with the full face in terms of recog-
nition performance. In scenarios where some parts of
the face are purposely disguised (scarves covering the
mouth for example) this observation might indicate
that a non-corrupted periocular region can, indeed,
overperform recognition with the occluded full-face
images. Such conditions were not tested in the present
work but might be the basis for an interesting follow-
up. Even though the observed results are promising,
it must be noted that the noise factors present in the
MobBIO database are still far from a highly uncon-
strained scenario.
The robustness of the likelihood-ratio decision
step was also assessed. We compared the perfor-
mance observed for the scores obtained with Equa-
tion 1 and the scores obtained using only its numera-
tor, i.e. only the likelihood of each test image without
the UBM normalization. For the experimental setup
(2) we obtained an average rank-1 recognition rate
of 43.6%, whereas the MobBIO experimental setup
(3) resulted in 90.5% for the same metric. It is eas-
ily noted that performance is less compromised in the
MobBIO database. Considering only the numerator
of Equation 1 is the same as considering a constant
denominator value for every tested image. As the de-
nominator represents the projection of the tested im-
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
44
(a) (b)
(c) (d)
(e) (f)
Figure 7: ROC and CMC curves for: (a-b) UBIRIS.v2 database, setup (1); (c-d) UBIRIS.v2 database, setup (2) and (e-f)
MobBIO database, setup (3);. ROC curves present the average results of cross-validation, whereas CMCs present the average
value and error-bars for the first 10 ranked IDs in each setup.
ages on the UBM, this alternative decision strategy
might be interpreted as assuming a constant back-
ground for every tested image. From the observed
results we might conclude that such assumption fits
better the images from the MobBIO database. We
also note that for more challenging scenarios, where
the constant background assumption fails, the use of
background normalization produces a significant im-
provement in performance.
A few last considerations regarding the discrim-
inative potential of the proposed algorithm may be
taken from the observation of Figure 8. On each
row we analyze the 4 highest ranked models for the
images presented in the first column. The first two
PeriocularRecognitionunderUnconstrainedSettingswithUniversalBackgroundModels
45
Table 1: Comparison between the average obtained results with both experimental setups for the UBIRIS.v2 database and
some state-of-the-art algorithms.
Work Setup Traits R
1
EER D
i
Proposed 1 P 97.73% 0.0452 4.9795
Proposed 2 P 88.93% 0.0716 3.6141
Moreno et al. (Moreno et al., 2013a) 1 P 97.63% 0.1417 –
Tan et al. (Tan and Kumar, 2013) 2 P + I 39.4% – –
Tan et al. (Tan et al., 2011) 2 P + I – – 2.5748
Kumar et al. (Kumar and Chan, 2012) 2 I 48.01% – –
Proenc¸a et al. (Proenc¸a and Santos, 2012) 2 I – ≈ 0.11 2.848
rows depict correct identifications. It is interesting
to note how each of the 4 highest ranked identities
in the second row correspond to individuals wearing
glasses. Such observation seems to indicate that the
proposed modeling process is capable of describing
high-level global features, such as glasses. Further-
more, the fact that the correct ID was guessed also
demonstrates its capacity of distinguishing between
finer details separating individual models. The third
and fourth rows present some test images whose ID
was not correctly assessed by the algorithm. In the
third row we present a case where even though the
correct ID and the most likely model were not cor-
rectly paired, the correct guess still appears in the top
ranked models. We note that even a human user ana-
lyzing the four highest ranked models would find it
very difficult to detect significant differences. The
fourth row presents the extreme case where none of
the top ranked models correspond to the true ID. It
is worth noting how the test images presented in the
third and fourth rows are very similar to a large num-
ber of images present in other individual’s models.
This observation leads to the hypothesis that some
users are easier to identify than others inside a given
population, an effect known as the Doddington zoo
effect (Ross et al., 2009). It also shows that the pro-
posed algorithm is capable of narrowing the range of
possible identities to those subjects who “look more
alike”.
4.6 Implementation Details
The proposed algorithm was developed in MATLAB
R2012a and tested on a PC with 3.40GHz Intel(R)
Core(TM) i7-2600 processor and 8GB RAM. To
train the GMM’s we used the Netlab toolbox (Nab-
ney, 2004), whereas SIFT keypoint extraction and
description was performed using the VLFeat tool-
box (Vedaldi and Fulkerson, 2010). For a single
identity check, using the UBIRIS.v2 database and
M = 128, we observed a processing time of 0.0586 ±
0.0088s.
Figure 8: Identification results for rank-4 in the UBIRIS.v2
database. The first column depicts the tested images while
the remaining 4 images exemplify representative images
from the 4 most probable models, after the recognition is
performed. The blue squares mark the true identity.
5 CONCLUSIONS AND FUTURE
WORK
In the present work we propose an automatic model-
ing of SIFT descriptors, using a GMM-based UBM
method, to achieve a canonical representation of in-
dividual’s biometric data, regardless of the number of
detected SIFT keypoints. We tested the proposed al-
gorithm on periocular images from two databases and
achieved state-of-the-art performance for all experi-
mental setups. Periocular recognition has been the
focus of many recent works that explore it as a viable
alternative to both iris and face recognition under less
ideal acquisition scenarios.
Even though we propose the algorithm for peri-
ocular recognition, the framework can be easily ex-
trapolated for other image-based traits. To the extent
of our knowledge, GMM-based UBM methodologies
were solely explored for speaker recognition so far.
The proposed work may, thus, represent the first of a
series of experiments that explore its main advantages
in the scope of multiple trending biometric topics. For
example, the fact that any number of keypoints trig-
gers a recognition score may be relevant when only
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
46
partial or occluded data is available for recognition.
Scenarios, like the one described in the last section,
where faces are purposely occluded may be an inter-
esting area to explore.
Besides from the conceptual advantages of the
proposed algorithm, a few technical details may be
improved in further works. Exploring further color
channels besides the RGB space could bring benefits
to the proposed algorithm. Regarding fusion, explor-
ing individual specific parameters instead of a global
parametrization, would enable the algorithm to be
trained to counter the Doddington zoo effect. As not
all people are as easy to identify, fitting the properties
of the designed classification block to adapt to dif-
ferent classes of individuals seems like an interesting
idea.
Finally, and regarding the training setup, some
questions might be worthy of a more thorough re-
search. In the case of voice recognition it is com-
mon to train two separate UBMs for male and fe-
male speakers. Extrapolating this idea to image-based
traits, multiple UBMs trained on homogeneous sets
of equally or similarly zoomed images might improve
the results when more realistic and dynamic condi-
tions are presented to the acquisition system. In a re-
lated topic it is also not consensual whether the left
and right eyes, due to the intrinsic symmetry of the
face, should be considered in a single model or as
separate entities. All the aforementioned questions
demonstrate how much the present results can be im-
proved, leaving some promising prospects for future
works.
REFERENCES
Bakshi, S., Kumari, S., Raman, R., and Sa, P. K. (2012).
Evaluation of periocular over face biometric: A case
study. Procedia Engineering, 38:1628–1633.
Bharadwaj, S., Bhatt, H. S., Vatsa, M., and Singh, R.
(2010). Periocular biometrics: When iris recognition
fails. In 4th IEEE International Conference on Bio-
metrics: Theory Applications and Systems, pages 1–6.
Boddeti, V. N., Smereka, J. M., and Kumar, B. V. (2011). A
comparative evaluation of iris and ocular recognition
methods on challenging ocular images. In 2011 Inter-
national Joint Conference on Biometrics, pages 1–8.
IEEE.
Joshi, A., Gangwar, A. K., and Saquib, Z. (2012). Per-
son recognition based on fusion of iris and periocu-
lar biometrics. In Hybrid Intelligent Systems (HIS),
2012 12th International Conference on, pages 57–62.
IEEE.
Ke, Y. and Sukthankar, R. (2004). Pca-sift: A more
distinctive representation for local image descriptors.
In Computer Vision and Pattern Recognition, 2004.
CVPR 2004. Proceedings of the 2004 IEEE Computer
Society Conference on, volume 2, pages II–506. IEEE.
Kinnunen, T., Saastamoinen, J., Hautamaki, V., Vinni, M.,
and Franti, P. (2009). Comparing maximum a poste-
riori vector quantization and gaussian mixture models
in speaker verification. In IEEE International Con-
ference on Acoustics, Speech and Signal Processing,
pages 4229–4232. IEEE.
Kittler, J., Hatef, M., Duin, R. P., and Matas, J. (1998). On
combining classifiers. Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 20(3):226–239.
Kumar, A. and Chan, T.-S. (2012). Iris recognition using
quaternionic sparse orientation code (qsoc). In 2012
IEEE Computer Society Conference on Computer Vi-
sion and Pattern Recognition Workshops, pages 59–
64.
Lowe, D. G. (2004). Distinctive image features from scale-
invariant keypoints. International journal of computer
vision, 60(2):91–110.
Miller, P. E., Lyle, J. R., Pundlik, S. J., and Woodard, D. L.
(2010a). Performance evaluation of local appearance
based periocular recognition. In IEEE International
Conference on Biometrics: Theory, Applications, and
Systems, pages 1–6.
Miller, P. E., Rawls, A. W., Pundlik, S. J., and Woodard,
D. L. (2010b). Personal identification using periocular
skin texture. In 2010 ACM Symposium on Applied
Computing, pages 1496–1500. ACM.
Moreno, J. C., Prasath, V., and Proenc¸a, H. (2013a). Ro-
bust periocular recognition by fusing local to holis-
tic sparse representations. In Proceedings of the 6th
International Conference on Security of Information
and Networks, pages 160–164. ACM.
Moreno, J. C., Prasath, V. B. S., Santos, G. M. M., and
Proenc¸a, H. (2013b). Robust periocular recognition by
fusing sparse representations of color and geometry
information. CoRR, abs/1309.2752.
Nabney, I. T. (2004). NETLAB: algorithms for pattern
recognition. Springer.
Padole, C. N. and Proenca, H. (2012). Periocular recog-
nition: Analysis of performance degradation factors.
In 5th IAPR International Conference on Biometrics,
pages 439–445.
Park, U., Jillela, R. R., Ross, A., and Jain, A. K. (2011).
Periocular biometrics in the visible spectrum. IEEE
Transactions on Information Forensics and Security,
6(1):96–106.
Park, U., Ross, A., and Jain, A. K. (2009). Periocular bio-
metrics in the visible spectrum: A feasibility study.
In IEEE 3rd International Conference on Biometrics:
Theory, Applications, and Systems, pages 1–6.
Povey, D., Chu, S. M., and Varadarajan, B. (2008). Univer-
sal background model based speech recognition. In
IEEE International Conference on Acoustics, Speech
and Signal Processing, pages 4561–4564.
Proenc¸a, H. (2009). NICE:II: Noisy iris challenge evalua-
tion - part ii. http://http://nice2.di.ubi.pt/.
Proenc¸a, H. (2011). Non-cooperative iris recognition: Is-
sues and trends. In 19th European Signal Processing
Conference, pages 1–5.
PeriocularRecognitionunderUnconstrainedSettingswithUniversalBackgroundModels
47
Proenc¸a, H. and Santos, G. (2012). Fusing color and shape
descriptors in the recognition of degraded iris images
acquired at visible wavelengths. Computer Vision and
Image Understanding, 116(2):167–178.
Proenc¸a, H., Filipe, S., Santos, R., Oliveira, J., and Alexan-
dre, L. A. (2010). The ubiris.v2: A database of visi-
ble wavelength iris images captured on-the-move and
at-a-distance. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 32(8):1529–1535.
Reynolds, D. (2008). Gaussian mixture models. Encyclo-
pedia of Biometric Recognition, pages 12–17.
Reynolds, D., Quatieri, T., and Dunn, R. (2000). Speaker
verification using adapted gaussian mixture models.
Digital signal processing, 10(1):19–41.
Reynolds, D. A. (2002). An overview of automatic speaker
recognition technology. In Acoustics, Speech, and
Signal Processing (ICASSP), 2002 IEEE International
Conference on, volume 4, pages IV–4072. IEEE.
Ross, A., Jillela, R., Smereka, J. M., Boddeti, V. N., Ku-
mar, B. V., Barnard, R., Hu, X., Pauca, P., and Plem-
mons, R. (2012). Matching highly non-ideal ocular
images: An information fusion approach. In Biomet-
rics (ICB), 2012 5th IAPR International Conference
on, pages 446–453. IEEE.
Ross, A., Rattani, A., and Tistarelli, M. (2009). Exploit-
ing the doddington zoo effect in biometric fusion.
In Biometrics: Theory, Applications, and Systems,
2009. BTAS’09. IEEE 3rd International Conference
on, pages 1–7. IEEE.
Santos, G. and Proenc¸a, H. (2013). Periocular biometrics:
An emerging technology for unconstrained scenarios.
In Computational Intelligence in Biometrics and Iden-
tity Management (CIBIM), 2013 IEEE Workshop on,
pages 14–21. IEEE.
Sequeira, A. F., Monteiro, J. C., Rebelo, A., and Oliveira,
H. P. (2014). MobBIO: a multimodal database cap-
tured with a portable handheld device. In Proceedings
of International Conference on Computer Vision The-
ory and Applications (VISAPP).
Shinoda, K. and Inoue, N. (2013). Reusing speech tech-
niques for video semantic indexing [applications cor-
ner]. Signal Processing Magazine, IEEE, 30(2):118–
122.
Smereka, J. M. and Kumar, B. (2013). What is a” good”
periocular region for recognition? In Computer Vision
and Pattern Recognition Workshops (CVPRW), 2013
IEEE Conference on, pages 117–124. IEEE.
Tan, C.-W. and Kumar, A. (2013). Towards online iris
and periocular recognition under relaxed imaging con-
straints. Image Processing, IEEE Transactions on,
22(10):3751–3765.
Tan, T., Zhang, X., Sun, Z., and Zhang, H. (2011). Noisy
iris image matching by using multiple cues. Pattern
Recognition Letters.
Vedaldi, A. and Fulkerson, B. (2010). Vlfeat: An open
and portable library of computer vision algorithms. In
Proceedings of the international conference on Multi-
media, pages 1469–1472. ACM.
Woodard, D. L., Pundlik, S. J., Lyle, J. R., and Miller, P. E.
(2010). Periocular region appearance cues for bio-
metric identification. In IEEE Computer Society Con-
ference on Computer Vision and Pattern Recognition
Workshops, pages 162–169.
Xiong, Z., Zheng, T., Song, Z., Soong, F., and Wu, W.
(2006). A tree-based kernel selection approach to ef-
ficient gaussian mixture model–universal background
model based speaker identification. Speech communi-
cation, 48(10):1273–1282.
BIOSIGNALS2015-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
48
