An a-contrario Approach for Face Matching
Luis D. Di Martino, Javier Preciozzi, Federico Lecumberry and Alicia Fern´andez
Instituto de Ingener´ıa El´ectrica, Universidad de la Rep´ublica, Montevideo, Uruguay
Keywords:
Face Recognition, Face Matching, a-contrario, STASM, LBP, Extended LBP, Chi-Square.
Abstract:
In this work we focus on the matching stage of a face recognition system. These systems are used to identify an
unknown person or to validate a claimed identity. In the face recognition field it is very common to innovate
in the extracted features of a face and use a simple threshold on the distance between samples in order to
perform the validation of a claimed identity. In this work we present a novel strategy based in the a-contrario
framework in order to improve the matching stage. This approach results in a validation threshold that is
automatically adapted to the data and allows to predict the performance of the system in advance. We perform
several experiments in order to validate this novel strategy using different databases and show its advantages
over using a simple threshold over the distances.
1 INTRODUCTION
Face Recognition Systems have acquired great impor-
tance in the last two decades. They are needed in a
world were automatic identification/tagging systems
as those used in security applications and human-
computer iteraction systems are major topics of re-
search and huge development in the industry. These
work by analyzing a face in an input image, they can
identify the person whose face is in the image (Identi-
fication Mode) or validate the identity claimed by the
subject (Verification / Matching Mode). In this work
we focus on the verification mode, in this the system
works with two sets of data: the gallery (G) and the
query (Q) datasets. The first contains N
G
faces of the
people registered in the system, the second is used for
testing, it contains N
Q
faces of people who claim one
of the registered identities.
A vast amount of works have been done in this
area and huge improvement can be seen from the
first face recognition systems to the current ones.
Nowadays, face recognition systems can achieve a
recognition rate over 97% by using local features
(Zhang et al., 2007; Ahonen et al., 2006; Zhang
et al., 2005; Zhang et al., 2010; Huang et al., 2011)
when identifying people and produce a very low rate
of false positives when verifying declared identities.
This performance is achieved on a database of well
controlled conditions such as the FERET database
(Phillips et al., 1998). However, some problems as
aging and large variations in lighting and pose still
have not been fully resolved and highly degrade the
performance of a face recognition system. Is still
very difficult for an automatic face recognition system
to recognize a person from a picture taken in heav-
ily uncontrolled conditions, what is called “face in
the wild”, this has brought great attention in the face
recognition field community (Kan et al., 2013), (Zhu
and Ramanan, 2012), (Masi et al., 2013). In (Marsico
et al., 2013) this issue has been studied with emphasis,
the authors introduces two indices in order to evaluate
beforehand the quality of the input image in terms of
the sample pose (index SP) and sample illumination
(index SI). This allows to reject the image or ask for a
new capture when the input image does not meet the
necessary quality. It is also very difficult to perform
the identification or verification process when the ag-
ing factor appears between images, e.g., trying to re-
cognize a person from only one picture registered 10
years ago, this matter has been treated extensively in
the literature, (Park et al., 2010), (Lanitis et al., 2002),
(Ling and Soatto, 2010). This problem has been usu-
ally addressed by constructing models that simulate
the changes in a person’s face due to aging. These are
used jointly with systems that estimate the age of a
person from a picture containing its face, the time dif-
ference between two face images is used as an input
for the aging simulation system. The problem with
using such approaches is that lot of face images of
each individual are needed in order to build a accurate
377
D. Di Martino L., Preciozzi J., Lecumberry F. and Fernández A..
An a-contrario Approach for Face Matching.
DOI: 10.5220/0004758003770384
In Proceedings of the 3rd International Conference on Pattern Recognition Applications and Methods (ICPRAM-2014), pages 377-384
ISBN: 978-989-758-018-5
Copyright
c
 2014 SCITEPRESS (Science and Technology Publications, Lda.)

model, because the aging process is highly dependent
on the person whose face is being treated.
Face recognition systems, as many others biomet-
ric systems, are usually divided in three stages: pre-
processing, feature extraction and a matching stage
1
.
Generally, great efforts are dedicated to the inno-
vation on preprocessing and feature extraction tech-
niques while using classic techniques in the matching
stage. In this paper we focus on the matching stage,
performing the study, implementation and analysis
of a novel strategy based on the a-contrario frame-
work (Desolneux et al., 2003b). This approach is in-
dependent of the feature extraction process as soon as
a dissimilarity distance between samples is defined.
A-contrario models have been successfully used in
applications where the searched features are very un-
likely to occur by chance under some background
model. Therefore, the meaningful matches can be
found, in a contrario fashion, by training models that
make possible to reject these matches that are due to
chance with high precision.
Algorithms based on a-contrario framework were
first used in the detection of alignments (Desolneux
et al., 2003b), contrasted edges and grouping (Des-
olneux et al., 2003a). Later, its use has been ex-
tended to more complex tasks, e.g., the detection of
line segments (Gioi and Jakubowicz, 2010), matching
of shapes (Mus´e et al., 2006), matching of SIFT-like
descriptors (Rabin et al., 2008) and biometric iden-
tification systems based on iris templates (Mottalli
et al., 2010). In this work we extend the use of the
a-contrario framework in order to improve the per-
formance of a face recognition system working in the
verification mode where the system is used to validate
an identity claimed by a person.
The rest of the paper is organized as follows: in
Section 2 we present the proposed face matching tech-
nique based in the a-contrario strategy, in Section 3
we provide details of the used framework, databases
and conducted experiments. In Section 4 the obtained
results are presented, and finally, we conclude our
work in Section 5.
2 a-contrario Framework
The a-contrario framework is based on the Helmholtz
Principle, which in its most general form states that
whenever some large deviation from randomness oc-
curs, a structure is perceived. Thus, we can find sig-
1
This stage is commonly referred, in the pattern recognition
area, as the classification stage. In particular, in biometrics appli-
cations, it is called the matching stage as it involves the match of a
query subject with a known identity.
nificant events as those who are far from the random
or background model. This principle can been for-
malized as follows:
Definition 1 (ε-meaningful event). We say that an
event of the type “a given configuration of objects has
a property” is ε-meaningful if the expectation of the
number of occurrences of this event is less than ε un-
der the background model.
Definition 2 (Number of False Alarms - NFA). Given
an event of the type “a given configuration of objects
has a property”, the number of false alarms (NFA) is
the expectation of the number of occurrences of this
event under the background model.
Definition 1 can be rewritten in terms of the NFA
defined before:
Definition 3 (ε-meaningful event). An event E of the
type “a given configurationof objects has a property”
is ε-meaningful if the NFA is less than ε:
NFA(E) < ε (1)
Let H
0
be the background model: “a given con-
figuration of objects has a property and is produced
by some known model”. We define a random vari-
able E, and we analyze the observation E of this ran-
dom variable considering the number of false alarms.
The correct definition of this NFA is a central prob-
lem in all a-contrario methods. However, quite often
this definition can be reduced to an expression of the
following form, which gives an upper bound of the
actual NFA as defined before:
Definition 4 (Number of false alarms - NFA). The
number of false alarms (NFA) of an event E is defined
as:
NFA
′
(E) = N · P(E ≥ E|H
0
) (2)
where N is the number of all possible configurations
of the event E.
Moreover, we can often show that the expectation
of the number of occurrences of an event E satis-
fying NFA
′
(E) < ε is actually less than ε. For this
reason, defining an event as ε-meaningful, whenever
NFA
′
(E) < ε, is still consistent with Definition 1 and
ensures that the method is robust in the sense that no
more than ε “false detections” will be obtained due to
noise.
2.1 a contrario Matching Approach
As explained before, a-contrario methods are based
on the identification of events which are very improb-
able under some background model. In the case of
face recognition, given two faces, we can define the
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
378

background model H
0
: “The faces correspond to dif-
ferent persons” (we can also define H
1
: “Both faces
correspond to the same person”). Note that even if we
do not have an explicit formulation for the H
0
model,
we can build it empirically from the data, since the
number of samples is overwhelming. This is one of
the reasons that explains the popularity of a-contrario
methods with respect to classical hypothesis testing:
the model we test against is not the one that describes
the rare events but the a-contrario one that is in gen-
eral much more easier to obtain. Given the compari-
son of N
Q
query faces against N
G
faces in the gallery
dataset, it is clear that there are much more represen-
tatives that comply the H
0
hypothesis than the ones
that comply with H
1
. This explains why if we build an
empirical distribution from the data is better to work
with the a-contrario model (because of the law of
large numbers, the empirical distribution of H
0
will
be closer to the real distribution than H
1
).
Lets now formalize all these concepts: indepen-
dently from the features used to compare two faces,
we always have a distance associated to this compar-
ison. Let D(q
i
, g
j
) be the distance between two faces
q
i
and g
j
from the query and gallery datasets respec-
tively. If we compute the distance of q
i
against all the
faces in the gallery, we can build the empirical distri-
bution of the distance D(q
i
, g
j
) given the background
model H
0
. Let be p
q
i
|H
0
(x) this probability density.
Given a value of distance D(q
i
, g
j
) = δ between two
faces, the probability that the compared faces satisfies
the H
0
hypothesis is computed using Equation 3.
P(D(q
i
, g
j
) ≤ δ|H
0
) =
Z
δ
−∞
p
q
i
|H
0
(x)dx (3)
In our a-contrario framework, the meaningful event
(the one that has very low probability under the back-
ground model H
0
) is that the two compared faces cor-
respond to the same person. Thus, a matching be-
tween two faces q
i
and g
j
is considered to be correct
if the “Number of False Alarms” is smaller than a
given threshold, which in fact is a threshold on the
expectation of the improbable event. The “Number of
False Alarms” for the face matching problem can be
defined as follows:
Definition 5 (Number of false alarms - NFA). The
number of false alarms (NFA) of the event E: “faces
q
i
and g
j
correspond to the same person” is defined
as:
NFA(q
i
, g
j
) = N
Q
(N
G
− 1)P(D(q
i
, g
j
) ≤ δ|H
0
) (4)
where N
Q
and N
G
are the size of the query and gallery
datasets respectively, therefore N
Q
(N
G
− 1) corre-
sponds to all possible configurations of the event de-
scribed before.
Having defined a global threshold ε, the match
between q
i
and g
j
is said to be ε-meaningful if
NFA(q
i
, g
j
) ≤ ε, in this case the face images are val-
idated as corresponding to the same person. From
these definitions is easy to prove that the expected
NFA when testing all the possible combinations be-
tween the people in the query and gallery datasets is
smaller than ε.
The definition of a threshold ε in the a-contrario
framework, instead of thresholding the distance, has
several advantages. First, it represents an intuitive in-
dicator of the expected number of false alarms and
therefore allows to control the performance of the sys-
tem in advance. Second, this threshold is automati-
cally adapted to the database as, for each person in the
query dataset, it is applied over the trained probabil-
ity function and therefore does not need to be adjusted
each time the database is changed.
3 EXPERIMENTS
In this section we first present the techniques used in
the preprocessing and feature extraction stages. Then,
we provide details of the used databases and con-
ducted experiments.
3.1 Developed Framework
3.1.1 Preprocessing Module
The preprocessing module automatically validates the
face in an input image, discards all unnecessary in-
formation, and resizes the resultant image to a pre-
defined size with fixed eye positions. This module is
very important in any automatic face recognition sys-
tem since this is the only stage where an input image
can be rejected, e.g., if there isn’t a face in the im-
age. It is also very important because the quality in
the registration of the face affects directly the perfor-
mance of the system when local features are used. For
this purpose we use both the manually marked eyes
positions and a version of the technique Active Shape
Models called STASM (Milborrowand Nicolls, 2008).
This software takes as an entry a passport style im-
age, validates a face in it, and then returns the position
of seventy two pre-determined facial landmarks, from
which two corresponds to the eyes positions. The lo-
cation of these seventy two points might seem as an
exaggeration since only two of these are finally used
for the registration of the face. But this strategy al-
lows to obtain better results finding eyes positions,
even if they are closed, gaining in robustness. The
STASM technique has been widely used in the face
Ana-contrarioApproachforFaceMatching
379

(a) Application of the classic LBP operator.
(b) The circular (8,2), (8,1), and (16,2) neighborhoods in the
ELBP approach.
(c) Result of applying the ELBP operator.
Figure 1: LBP and ELPB strategies.
recognition field, its implementation provides accu-
rate results when marking points of interest in a face
(a comparison of this method and other implementa-
tions of the ASM technique can be found in (Marsico
et al., 2013)).
3.1.2 Feature Extraction
The feature extraction module receives a normalized
image and computes a set of features (a vector) which
ideally would uniquely represent the person in the in-
put image.
We used the method presented on (Ahonen et al.,
2006) for its simplicity, this technique uses Extended
Local Binary Patterns (ELBP) to extract the so-called
“micro-patterns”. This sets of features has been suc-
cessfully used in texture and structures description,
especially in face recognition(an extensivesurveycan
be found in (Huang et al., 2011)). In the original LBP
technique, first an operator is applied to every pixel
in the image. This operator compares the gray values
of each pixel against its neighbors and assign a binary
label as shown in Figure 1(a). In the ELBP strategy
this procedure is extended in order to consider more
than only the immediate neighbors by taking the gray
values in points uniformly distributed on a circle as
shown in Figure 1(b). The pixel values are bilinearly
interpolated whenever the sampling point is not in the
center of a pixel.
The result of applying the ELBP operator is a new
image of features as shown in Figure 1(c). This image
is divided in regions from which feature histograms
are extracted. These histograms are finally concate-
nated, this way both local and global information is
preserved. The micro-patterns hold the local infor-
mation and the order in the concatenation of the local
histograms assure that when comparing two faces the
same regions of the faces will be taken into consider-
ation for the comparison.The feature vector has a size
s that can be computed as
s = N
b
N
p
(5)
Where N
b
represents the quantity of bins in the local
histograms and N
p
the total number of patches used
in the division of the image.
3.1.3 Matching
In order to measure the dissimilarity between feature
vectors we used the Chi-square distance between his-
tograms corresponding to each one of the patches, this
distance has been widely used when working with his-
tograms as feature vectors obtained with operators as
ELBP. Given a face q
i
in the input image and a face g
j
in the gallery dataset, the distance between them can
be computed as the sum of the distances measured be-
tween patches as shown in Equation 6.
D(q
i
, g
j
) =
N
p
∑
n=1
χ
2
n
(q
i
, g
j
)
=
1
2
N
p
∑
n=1
N
b
∑
m=1
(q
i,n,m
− g
j,n,m
)
2
(q
i,n,m
+ g
j,n,m
)
(6)
After the distance between samples is computed, we
perform the matching by using two different tech-
niques. The first matching strategy is based in the
nearest neighborhood (NN) technique, in which an
identity is validated just by applying a threshold in the
distance between two faces. This strategy has been
widely used in the face recognition area, and there-
fore we used as a benchmark. The second strategy,
which is the novelty presented in this work is based in
the use of a-contrario model that works as described
in Section 2.1.
3.2 Databases
In this work we use two different databases to evalu-
ate the performance of the developed face recognition
system: FERET and DNIC.
3.2.1 Feret
The FERET database was created as part of the pro-
gram Face Recognition Technology carried on in
the years 1993 − 1997 (Phillips et al., 1998). This
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
380

database has become very popular in the area of face
recognition and is commonly used as a benchmark. It
contains a gallery f
a
containing 994 people and three
standards test sets: f
b
, dup
1
and dup
2
. The set fb in-
cludes images taken on the same day of the ones in the
gallery with only differences in expressions, the eyes
position coordinates of 849 people of this dataset are
provided. The set dup
1
contains images of 736 per-
sons and includes the aging effect when compared to
the samples in the gallery. The set dup
2
is a subset of
dup
1
that contains 228 persons with images taken, at
least, 540 days later than the images in the gallery.
3.2.2 DNIC
DNIC (Direcci´on Nacional de Identificaci´on Civil) is
the Uruguayan government organization responsible
for the emission of ID cards and passports. It has a
civil identification system, with a database containing
more than three millions identities
2
. This database
has the special feature of being used in a real ap-
plication. Its images are acquired under not com-
pletely controlled and differents environments, there-
fore, some special characteristics are to be expected
as changes in illumination and pose variations. Some
others factors are controlled like the lack of face ac-
cesories (hats, lenses, scarfs, etc) and neutral face ex-
pressions. Summarising, one could say that this is a
database of semi-controlled conditions. As the IDs
are renewed periodically
3
, this database allows us to
study the impact of the aging factor in the perfor-
mance of the system.
3.3 Experiments
In this work we perform several test of the proposed
system using the databases presented above. In all the
performed experiments the system is evaluated both
using the NN approach and the a-contrario proposed
technique. In each case, the Verification Rate (VR)
is plotted against the False Acceptance Rate (FAR).
This two indicators are commonly used when evalu-
ating the performance of a biometric system used for
identities verification. In order to present how these
indicators are computed the following functions are
introduced.
1
τ
D
(q
i
, g
j
) =
(
1 if D(q
i
, g
j
) ≤ τ
0 otherwise
(7)
2
Because of the local laws that protect privacy, sample
images of this database can not be shown in this article.
3
Normally, the IDs are renewed every 10 years, but in
special cases (for example, in the case of stolen or lost doc-
uments) the IDs are renewed with a smaller time difference.
1
id
(q
i
, g
j
) =
(
1 if id(q
i
) = id(g
j
)
0 if id(q
i
) 6= id(g
j
)
(8)
The indicator function in Equation 7 equals 1 when
the distance between the faces q
i
and g
j
is below a
predefined threshold τ
4
and therefore are declared by
the system as being faces of the same person. The
function defined in Equation 8 equals 1 when the
compared faces correspond to the same person (their
associated identity is the same). The VR and FAR in-
dicators are computed according to the equations 9
and 10 respectively.
VR(τ) =
N
Q
∑
i=1
N
G
∑
j=1
1
id
(q
i
, g
j
) × 1
τ
D
(q
i
, g
j
)
N
Q
(9)
FAR(τ) =
N
Q
∑
i=1
N
G
∑
j=1
(1− 1
id
(q
i
, g
j
)) × 1
τ
D
(q
i
, g
j
)
N
Q
× (N
G
− 1)
(10)
It could be said that the VR indicator represents the
rate of persons who, claiming their true identity, are
correctly validated by the system. On the other hand,
the value FAR indicates the rate of people who, claim-
ing a false identity, can fool the system that validates
them as genuine users.
Experiment 1. In this experiment the FERET set
f
b
is used to evaluate the system both in the manual
and automatic configurations, in the manual config-
uration the provided eyes coordinates are used while
in the automatic mode the eyes position coordinates
are found using the presented STASM technique. This
first test has two different goals. First, since the im-
ages of this set have little variation with respect to the
images in the gallery, it allows us to study the impact
of the positioning of the eyes in the performance of
the system. Second, it validates the performance of
the a-contrario proposed technique in a well known
database that is usually used as a benchmark.
Experiment 2. In Experiments 2 and 3 we perform
the evaluation with several subset from the DNIC
database. As stated before, this database allows to
evaluate the system working with a database of semi-
controlled conditions used by the DNIC in a produc-
tion environment. In experiment two, we test the de-
veloped system using a database of 5000 people, this
permits us to evaluate the system working with a mid-
size database.
4
The threshold τ is the one applied directly over the dis-
tance between samples or in the probability computed using
the proposed a-contrario technique.
Ana-contrarioApproachforFaceMatching
381

10
−4
10
−3
10
−2
10
−1
10
0
0.8
0.85
0.9
0.95
1
1.05
VR vs FAR for FERET set f
b
FAR
VR
A−Contrario (Manual)
NN (Manual)
A−Contrario (Automatic)
NN (Automatic)
Figure 2: Experiment 1: VR vs FAR for automatic and
semi-automatic configurations, FERET f
b
dataset.
Experiment 3. In this experiment we analyze the
impact of aging in the face recognition process by
using datasets composed of 200 people with a time
difference between the query images and the samples
in the gallery of 2, 3, 4, 5 and 9 years. In all these
tests we evaluate the system using both the proposed
a-contrario framework and the NN approach.
In all the experiments we used normalized images
of size 128×128 pixels with fixed eyes positions. We
used the ELBP technique considering 8 neighbors in
a circle of radius 3 pixels, using 256 bins in each local
histogram and dividing the image with a 9× 9 grid.
4 RESULTS
The results of Experiment 1 are shown in Figure 2.
Clearly the positioning of the eyes affects the over-
all performance of the system in both the NN and
proposed a-contrario approaches. The performance
is worse when the automatic eyes positions finder
technique is used instead of providing the manually
marked eyes coordinates. This behaviour is as ex-
pected considering that the features are extracted by
regions on the face. Therefore, any difference in the
registration of a person’s face between the images in
the gallery and query datasets could result in the com-
parison of features that are extracted from differents
face regions negatively impacting the performance.
From this first experiment is clear that the a-
contrario proposed technique outperforms the classic
NN approach in both the manual and automatic con-
figurations of the system. The performance when us-
ing both strategies is slighty different for the higher
values of the FAR indicator as shown in Figure 2. On
the other hand, when we pursue lower values of the
10
−3
10
−2
10
−1
10
0
0.75
0.8
0.85
0.9
0.95
1
1.05
VR vs FAR for DNIC
FAR
VR
A−Contrario
NN
Figure 3: Experiment 2: VR vs FAR for DNIC dataset, sub-
set containing 5000 people.
FAR index, the a-contrario technique remains more
robust than the NN approach leading to significantly
higher values of the verification rate indicator. This is
very important, as in these applications usually is con-
venient to work with a very low FAR value while not
affecting the system’s genuine users. In the NN strat-
egy the decision threshold needs to be more restric-
tive in order to achieve these FAR values leading to a
degradation in the verification rate. The a-contrario
strategy is more robust to this change as its thresh-
old is trained for each person taking many samples of
distances between different people’s faces.
The results of Experiment 2 are shown in Figure 3,
it can be seen once more that the proposeda-contrario
technique outperforms the NN approach. In this case
the difference in the performance maintains for all the
range of FAR values. This behaviouris to be expected
as, in this case, there are many more samples to train
the probability density function (pdf) used in the a-
contrario technique. A better estimation of the pdf
allows for a better adjustment of the threshold when
comparing two persons increasing the discrimination
capability.
We present in Table 1 the relationship between
the selected ε threshold for the application of the a-
contrario technique and the obtained VR and FAR
values when testing the system with the DNIC big
dataset. As stated before, the setting of the ε threshold
allows to control the expected number of false alarms,
this represents a great advantage as it allows to con-
trol the performance of the system in advance. From
the presented data, is clear that the FAR indicator that
measures the total number of false alarms remains
bounded by the choosen epsilon value. This shows
empirically that the setting of an arbitrary threshold ε
allows to control the performance of the system be-
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
382

Table 1: Experiment 2: FAR in the a-contrario approach
versus ε threshold.
ε FAR VR
0.005 0.004 0.905
0.010 0.009 0.913
0.050 0.047 0.934
0.100 0.097 0.945
0.500 0.495 0.989
forehand.
Finally, we analyze how the aging process im-
pacts in the performance of the developed system us-
ing both the a-contrario technique and NN approach.
The obtained results are shown in Figure 4. Clearly,
the age gap between the images in the gallery and
query dataset highly degrades the performance of the
system. There is an improvement in system perfor-
mance when there is a gap of three years between
images with respect to the case where the separation
is of two years. This behavior is not as expected
as it is normal that the performance worsens as the
time difference between images increases. This re-
sult can only be explained by the nature of the data,
it is important to remark that this database is com-
posed by semi-controlled conditions (lighting, pose,
expressions), these conditions could affect the system,
in some cases, more than the time difference between
images.
While the a-contrario technique produces better
results than the NN-based strategy, both of them start
to fail as the age difference between images increase.
This behaviour can be explained by looking at the fea-
ture extraction stage. The selected features are very
simple to implement and fast to compute when pro-
cessing images but not robust to the changes produced
in a face as a result of the aging process. The a-
contrario technique does not solve this problem as it
relies in the extracted features used when describing
a face.
5 CONCLUSIONS
The face recognition field is rapidly growing and there
is a huge need for new techniques and algorithms that
could enable to identify a subject or verify an iden-
tity even in very uncontrolled conditions as those pre-
sented in pictures taken from surveillance cameras.
In this work we introduce a matching strategy
based in the a-contrario models that were applied
widely in the pattern recognition area but not in the
particular problem of face recognition. This tech-
nique does not present any constraints for its appli-
cation and is well suited even in the case there is
2 3 4 5 6 7 8 9
0.3
0.4
0.5
0.6
0.7
0.8
VR vs Age difference for FAR = 0.01
Age difference (years)
VR at FAR = 0.01
A−Contrario
NN
Figure 4: Experiment 3: VR vs Age difference for FAR =
0.01 in DNIC database.
only one sample per person in the gallery dataset.
It is also independent from the extracted features as
soon as a dissimilarity measure between samples is
defined, this allows to adapt the matching stage to
different feature extraction scenarios. The conducted
experiments show that this algorithm outperforms the
classic NN approach widely used in the face recogni-
tion field as it is adjusted to the input image and the
database. It is also shown that it allows to predict with
high precision the total number of false alarms by set-
ting a threshold ε, thus enabling to control in advance
the performance of the system.
Several points remains as future work. For in-
stance, we are planning to compare the performance
of the proposed technique against others matching al-
gorithms taking in consideration performance, scala-
bility and speed. In particular we are interested in the
scalability factor in order to validate how well is this
technique suited to the verification process in a citizen
database as the one from the DNIC. We are also plan-
ning to test the proposed technique with different sets
of extracted features in order to knowif any advantage
can be obtained when using some particular features
that can, for example, overcome the aging issue.
ACKNOWLEDGEMENTS
The authors would like to thank Rafael Grompone
and Pablo Mus´e for their valuable contributions, the
DNIC agency for its collaboration and for granting
us the permission to access their valuable data. This
work has been supported by an investigation grant
provided by the ANII (Uruguay Agency of Investi-
gation and Innovation).
Ana-contrarioApproachforFaceMatching
383

REFERENCES
Ahonen, T., Hadid, A., and Pietik¨ainen, M. (2006). Face
description with local binary patterns: application to
face recognition. IEEE transactions on pattern analy-
sis and machine intelligence, 28(12):2037–41.
Desolneux, a., Moisan, L., and Morel, J. (2003a). A group-
ing principle and four applications. IEEE Transac-
tions on Pattern Analysis and Machine Intelligence,
25(4):508–513.
Desolneux, A., Moisan, L., and Morel, J. (2003b). Maximal
meaningful events and applications to image analysis.
The Annals of Statistics, 31(6):1822–1851.
Gioi, R. V. and Jakubowicz, J. (2010). LSD: A fast line
segment detector with a false detection control. Pat-
tern Analysis and Machine Intelligence, IEEE Trans-
actions on, 32:722–732.
Huang, D., Shan, C., Ardabilian, M., Yunhong, W., and
Liming, C. (2011). Local Binary Patterns and Its Ap-
plication to Facial Image Analysis: A Survey. Sys-
tems, Man, and Cybernetics, Part C: Applications and
Reviews, IEEE Transactions on, 41(6):765–781.
Kan, M., Xu, D., Shan, S., Li, W., and Chen, X. (2013).
Learning Prototype Hyperplanes for Face Verification
in the Wild. Image Processing, IEEE Transactions on,
22:3310–3316.
Lanitis, A., Taylor, C., and Cootes, T. (2002). Toward auto-
matic simulation of aging effects on face images. Pat-
tern Analysis and Machine Intelligence, IEEE Trans-
actions on, 24(4):442–455.
Ling, H. and Soatto, S. (2010). Face verification across
age progression using discriminative methods. Infor-
mation Forensics and Security, IEEE Transactions on,
5(1):82–91.
Marsico, M. D., Nappi, M., Riccio, D., and Wechsler, H.
(2013). Robust Face Recognition for Uncontrolled
Pose and Illumination Changes. IEEE Transactions
on Systems, Man, and Cybernetics, 43(1):149–163.
Masi, I., Lisanti, G., Bagdanov, A., Pala, P., and Bimbo,
A. D. (2013). Using 3D Models to Recognize 2D
Faces in the Wild. In Computer Vision and Pattern
Recognition Workshops (CVPRW), 2013 IEEE Con-
ference on, pages 775–780.
Milborrow, S. and Nicolls, F. (2008). Locating facial fea-
tures with an extended active shape model. Computer
VisionECCV 2008, pages 504 – 513.
Mottalli, M., Tepper, M., and Mejail, M. (2010). A con-
trario detection of false matches in iris recognition,
volume 6419. Springer Berlin Heidelberg.
Mus´e, P., Sur, F., Cao, F., Gousseau, Y., and Morel, J.-M.
(2006). An A Contrario Decision Method for Shape
Element Recognition. International Journal of Com-
puter Vision, 69(3):295–315.
Park, U., Tong, Y., and Jain, A. K. (2010). Age-invariant
face recognition. IEEE transactions on pattern analy-
sis and machine intelligence, 32(5):947–54.
Phillips, P. J., Wechsler, H., Huang, J., and Rauss, P. J.
(1998). The FERET database and evaluation proce-
dure for face-recognition algorithms. Image and Vi-
sion Computing, 16(5):295–306.
Rabin, J., Delon, J., and Gousseau, Y. (2008). A contrario
matching of SIFT-like descriptors. 2008 19th Interna-
tional Conference on Pattern Recognition, pages 1–4.
Zhang, B., Gao, Y., Zhao, S., and Liu, J. (2010). Lo-
cal derivative pattern versus local binary pattern:
face recognition with high-order local pattern descrip-
tor. IEEE transactions on image processing : a
publication of the IEEE Signal Processing Society,
19(2):533–44.
Zhang, B., Shan, S., Chen, X., and Gao, W. (2007). His-
togram of Gabor phase patterns (HGPP): a novel
object representation approach for face recognition.
IEEE transactions on image processing : a publica-
tion of the IEEE Signal Processing Society, 16(1):57–
68.
Zhang, W., Shan, S., Gao, W., Chen, X., and Zhang, H.
(2005). Local Gabor binary pattern histogram se-
quence (LGBPHS): a novel non-statistical model for
face representation and recognition. Tenth IEEE Inter-
national Conference on Computer Vision (ICCV’05)
Volume 1, pages 786–791 Vol. 1.
Zhu, X. and Ramanan, D. (2012). Face detection, pose es-
timation, and landmark localization in the wild. Com-
puter Vision and Pattern Recognition (CVPR), 2012
IEEE Conference on, pages 2879–2886.
ICPRAM2014-InternationalConferenceonPatternRecognitionApplicationsandMethods
384