ELEKTROTEHNI ˇ SKI VESTNIK 79(1-2): 1–6, 2012
ENGLISH EDITION
Robust 3D Face Recognition
Janez Kriˇ zaj, Vitomir
ˇ
Struc, Simon Dobriˇ sek
University of Ljubljana, Faculty of Electrical Engineering, Trˇ zaˇ ska 25, 1000 Ljubljana, Slovenia
E-mail: janez.krizaj@fe.uni-lj.si
Abstract. Face recognition in uncontrolled environments is hindered by variations in illumination, pose,
expression and occlusions of faces. Many practical face-recognition systems are affected by these variations.
One way to increase the robustness to illumination and pose variations is to use 3D facial images. In this paper
3D face-recognition systems are presented. Their structure and operation are described. The robustness of such
systems to variations in uncontrolled environments is emphasized. We present some preliminary results of a
system developed in our laboratory.
Keywords: biometric systems, face recognition, 3D images, features
1 INTRODUCTION
Systems for the biometric recognition of individuals
assess the identity of these individuals on the basis
of their physiological or behavioral characteristics, like
fingerprint, face, speech, gait and iris patterns. The scope
of these systems includes various kinds of access control
(border crossing, access to personal information), foren-
sics, law enforcement and others. Face-recognition sys-
tems are some of the most popular among all biometric
systems. This is mostly due to their non-intrusive nature,
as the data acquisition can be performed from a distance,
even without the subject’s cooperation. Special attention
in such biometric systems is being paid to developing
the so-called smart surveillance technologies, especially
to developing portals for the automatic control of border
crossings [1].
Although people recognize faces without any spe-
cial effort, automatic face recognition with a computer
presents a considerable difficulty and challenge, espe-
cially if the images are acquired in an uncontrolled
environment. The main factors that affect the accuracy
of face-recognition systems are the variability in the
illumination and the pose of the faces, expressions,
occlusions (scarf, beard, glasses), time delay (signs
of aging), makeup and similar occurrences during the
image-acquisition task. The presence of these factors in
the facial images can lead to a diminished recognition
reliability.
In order to improve the robustness and reliability
of face-recognition systems, various data-acquisition
techniques were employed, including video, infra-red
images, multiple consecutive shots from different angles
and 3D images. The benefits of using 3D images for
Received Januar 12, 2012
Accepted Februar 2, 2012
face recognition encompass the invariance of 3D data to
illumination conditions and the ability to rotate the 3D
facial data to a normal pose [2]. Despite this, most of
the 3D face-recognition systems are affected by facial
expression, occlusion and time delay.
This paper first discusses the basic structure and op-
eration of 3D face-recognition systems. Some examples
of popular methods used in these systems are also
presented - from the 3D image-acquisition techniques
to the classification methods. The preliminary results of
our own 3D face-recognition system developed in our
laboratory are also introduced.
2 STRUCTURE OF 3D FACE-RECOGNITION
SYSTEMS
A typical 3D face-recognition system is built from
the following units (Fig. 1): an image-acquisition and
pre-processing unit, a feature-extraction unit and a
similarity-measure and classification unit. In the sub-
sequent sections, the units are presented in detail with
examples of implementations that have emerged in the
literature.
image
acquisition and
pre-processing
feature
extraction
similarity
measure and
classification
facial
features of
the users
ID
feature
vector
localised and
normalised
3D image
3D
image
Figure 1. Block diagram of a 3D face-recognition system.

2 KRI
ˇ
ZAJ,
ˇ
STRUC, DOBRI
ˇ
SEK
3 IMAGE ACQUISITION AND
PRE-PROCESSING
It is assumed that the representation of faces with 3D im-
ages has several advantages over 2D images. However,
advantages like invariance to illumination and robustness
to the rotation of faces in 3D space do not hold com-
pletely in reality [3]. Rotation and scale normalization
can be computationally quite expensive and the existing
methods are not always convergent. Similarly, the fact
that the 3D data is illumination invariant is not always
true - strong light sources and reflective surfaces can
significantly affect the 3D sensor reading. Therefore, the
raw 3D images usually contain some degree of noise,
which can be recognized as spikes (reflective regions
- oily skin) and holes (missing data in the transparent
regions - eyes, eyebrows, hair, beard) in the scanned
image. Examples of the above distortions are given in
Fig. 2.
(a) (b) (c)
Figure 2. Examples of scans with distortions: (a) nose absence,
(b) noise around the eye region, (c) distortion due to face
motion during image acquisition.
3.1 3D sensors
The current 3D sensors fall into two categories, i.e.,
active sensors and passive sensors, of which active
sensors are more suitable for the 3D face-recognition
task. The passive 3D sensors reconstruct the 3D images
indirectly from the 2D images or video. The active 3D
sensors or scanners use laser light or structured light
patterns to capture the 3D data. Among the active 3D
sensors used to capture face scans structured-light, the
3D scanners are the most popular. These scanners use
projected light patterns and a camera to measure the
3D shape of an object. The projector emits multiple
light stripes (usually infra-red light or laser light) onto a
3D-shaped surface, while the camera acquires the light
patterns that are distorted from other perspectives than
that of the projector. The 3D shape of an object can be
reconstructed from the differences between the projected
and acquired patterns. This technique is used in a large
number of commercially available 3D sensor systems:
Konica Minolta Range 5 / Vivid9i / Vivid910 (examples
of scans acquired with Konica Minolta Vivid910 can
be seen in Fig. 2), Cyberware PX, 3DFaceCam, Face-
SCAN, FastSCAN, IVP Ranger M50. Recently, some
low-cost alternatives for the 3D data acquisition have
emerged in the market, such as the Microsoft Kinect sen-
sor and Asus Xtion PRO. Although these sensors have
numerous limitations, such as a low depth resolution and
depth of field, it is possible to obtain an adequate 3D
model of an object with a suitable pre-processing step.
A representative example is the method in [4], where a
reconstructed 3D face model (Fig. 3b) can be obtained
with an iterative adaptation of the average face model
to the 3D scan acquired by the Kinect sensor (Fig. 3a).
(a) (b)
Figure 3. 3D face image acquired by the Kinect sensor: (a)
raw image, (b) adapted 3D face model.
3.2 3D image pre-processing
The output of a common 3D sensor is a set of 3D
points of a scanned surface, with the values of the x, y
andz components at each point. The 3D data is usually
presented as a point cloud (Fig. 4(a)) or a range image
(Fig. 4b). The point cloud is a set of (x;y;z) coordinates
of scanned points from the object surface. A range image
(or a depth image) can be obtained by the projection
of scanned 3D points onto the (x;y) plane. Therefore,
the range image is formatted in a similar way to a 2D
intensity image, but with the difference that in the range
image the pixel intensities are proportional to the depth
components of a scanned object (z coordinates).
(a) (b) (c)
Figure 4. 3D data representation: (a) point cloud, (b) depth
image, (c) shaded depth image.
In the acquired 3D image, the face detection and lo-
calization are usually performed first. Detection denotes
a process where the presence and the number of faces
in the image are determined. Assuming that the image
contains only one face, the localization task is to find the
location and size (and sometimes also the orientation) of
a facial region. Most methods for face localization in 3D
images are based on an analysis of the local curvedness

ROBUST 3D FACE RECOGNITION 3
of the facial surface [5–7]. This gives us a set of possible
points for the locations of the characteristic facial parts,
such as the location of the nose, eyes and mouth, through
which the exact location, size and orientation of the
facial area can be determined. Based on the locations
of these points, the face area can be cut from the rest
of the image and eventually re-scaled and rotated to the
normal pose (Fig. 5).
Since the raw images generally contain some degree
of noise, the images are usually filtered before sub-
sequent processing. Usually, low-pass filters are used
to filter out high-frequency noise (spikes), while the
missing data are substituted by the interpolation of
adjacent points on the facial surface [5, 8].
(a) (b)
Figure 5. 3D image pre-processing: (a) raw image, (b) local-
ized and filtered facial region.
4 FEATURE EXTRACTION
The purpose of feature extraction is to extract the
compact information from the images that is relevant for
distinguishing between the face images of different peo-
ple and stable in terms of the photometric and geometric
variations in the images. One or more feature vectors
are extracted from the facial region. Depending on how
the facial region is treated, the existing feature-extraction
methods can be divided into the groups described below.
(a) (b)
Figure 6. Conceptual presentation of feature types: (a) global
features, (b) local features.
4.1 Global-feature extraction methods
These methods extract the feature vector from the
whole face region (Fig. 6a). The majority of the global
3D facial-feature-extraction methods have been derived
from methods originally used on 2D facial images,
where 2D gray-scale images are replaced by range
images. Principal component analysis (PCA) is the most
widespread method for global-feature extraction. PCA
was first used for feature extraction from 2D face images
and later also for feature extraction from range im-
ages [9]. Other popular global-feature extraction meth-
ods, such as linear discriminant analysis (LDA) [10] and
independent component analysis (ICA) [11], were also
used on range images.
The advantages of global features are: a considerable
reduction of the data dimensionality, and the spatial rela-
tionship among the different parts of the face is retained
(in the case of local features this information is generally
lost). The main disadvantage of global-feature methods
is that these methods require the precise localization and
normalization of the orientation, scale and illumination.
Changes in these factors can affect the global facial
features, resulting in a decreased recognition rate. In
global-feature-based recognition systems, localization
and normalization are often performed by the manual la-
beling of characteristic points on the face (which makes
the whole process semi-automatic). Automatic localiza-
tion and normalization is generally achieved using the
iterative closest-point algorithm (ICP) [12]. However,
the ICP algorithm is computationally expensive and does
not always converge to a global maximum. The global-
feature-based approaches are normally also sensitive to
facial expression and occlusion.
4.2 Local-feature extraction methods
Local-feature extraction methods extract a set of fea-
ture vectors from a face, where each vector holds the
characteristics of a particular facial region. The use of
global features is prevalent in face-recognition systems
based on images acquired in a controlled environment.
The local features are at an advantage over the global
features in uncontrolled environments, where the vari-
ations in facial illumination, rotation, expressions and
scale are present, since a local analysis of facial parts
provides a better basis to deal with such variations.
The process of local-feature extraction can be divided
into two parts. In the first part, the interest points on the
face region are detected. In the second part, the interest
points are used as locations at which the local feature
vectors are calculated.
There are several methods for interest-points detec-
tion. The interest points can be detected as extrema in the
scale-space, resulting in the invariance of features to the
scale. This approach of interest-points detection is used
in the scale-invariant feature transform (SIFT) [13, 14]
and the speeded-up robust features (SURF) [15]. The
interest points can also be detected as follows: on the
basis of the local curvedness analysis [16]; by the
alignment of faces with a face model in which interest-
point locations are marked a priori [17]; by the elastic
bunch graph method (EBGM) [18] as nodes of the
elastic graph; and as nodes of a rectangular grid covering

4 KRI
ˇ
ZAJ,
ˇ
STRUC, DOBRI
ˇ
SEK
the facial region [8, 19, 20]. The latter approach is
equivalent to detecting the local features on a block
basis, where the feature vectors are extracted by the
sliding-block technique.
Interest-point detection is followed by the extrac-
tion of local feature vectors at the locations of the
interest points. In the earlier approaches, local features
were generally defined from the geometric relations
among the interest points (location of the points, dis-
tances and angles between the points, distance ratios,
geodesic distances). For the description of the local
surface around the interest points, the latter approaches
normally use: differential geometry descriptors (mean
curvature, Gaussian curvature, shape index) [8, 21],
point signatures [22], Gabor filters [18], coefficients of
discrete cosine transform (DCT) [19, 20] and orientation
histograms [13].
Local features have several advantages over global
features. Due to the nature of the local-feature-based
approaches, the recognition performance is less affected
by the imprecise face localization and normalization than
in the case of global features. Therefore, some local-
feature-based approaches do not require the normaliza-
tion of illumination, rotation and scale variations. The
local approaches are also less sensitive to expression
variations.
4.3 Hybrid-feature extraction methods
Hybrid-feature-based approaches use both of the
above-mentioned feature types. Data from the global and
local approaches can be fused at the feature level [23],
similarity level [24] or decision level [25].
5 SIMILARITY MEASURE AND
CLASSIFICATION
The last step of the 3D face-recognition process presents
a similarity measure between the test face and the
faces from the system’s database. The decision that
follows depends on the purpose of the recognition sys-
tem. Recognition systems can be divided according to
their purpose into verification systems and identification
systems. Verification systems validate the identity claim
of the test person by comparing their features to the
template associated with the claimed identity. If the
similarity with the template is high enough, the system
recognizes the person as a client, or if the similarity is
too low, the person is recognized as an impostor. The
identification system conducts a one-to-many compari-
son by searching for the maximum similarity between
the test-person features and all the templates stored in
the database. The output of the identification system is
the person’s identity or the answer that the person does
not fit to any model from the database.
In the case of global-feature-based approaches, where
each face is represented by one feature vector, the
similarity between two faces is defined on the basis of
a certain distance measure (L
1
norm, L
2
norm, cosine
distance, Mahalanobis distance) between the feature rep-
resentations of these two faces. Local-feature approaches
may require a different similarity-measure procedure,
since each face is generally represented by a variable
number of local-feature vectors, and therefore we do not
have a unified representation of faces. When comparing
two faces represented by the local features, we usually
also do not know which local-feature vectors belong to
the same face regions of the two faces. For the reasons
outlined above, the unified encoding of faces represented
by local features is often utilized with the parameters
of the Gaussian mixture model (GMM) [26] or the
hidden Markov model (HMM) [19, 20]. The comparison
between two faces represented by local features can also
be performed directly by comparing each local-feature
vector of one face to each local-feature vector of the
other face. In this case, the similarity measure is based
on the number of the matched feature vectors or on
the sum of the distance measures among the matching
feature vectors.
The classification process normally utilizes the result
of the similarity measure. The most common and the
simplest classification technique is the nearest-neighbor
classifier (1-NN), while other popular classification tech-
niques include support vector machine (SVM) [27, 28]
and likelihood ratio [20, 26].
6 SYSTEM IMPLEMENTATION AND
EXPERIMENTS
We implemented a 3D face-verification system based
on the local features and GMM models (see Fig. 7).
This system represents a popular procedure used for
2D face recognition as well as for 3D face recognition.
The system will serve as a starting point for further
research in which we will try to justify the applicability
of local features in robust 3D face recognition. The
system performance was compared to the global PCA
method.
Gallery
picture
EM
UBM
2D DCT
2D DCT
Training
pictures Pre-processing &
feature extraction
GMM
supervector
Claimant
picture
2D DCT
MAP
GMM
supervector
UBM
UBM
supervectors
SVM
Decision
border
Position
to the decision
border?
Accept or
reject the
claimant
Pre-processing &
feature extraction
Pre-processing &
feature extraction
MAP
Figure 7. Block diagram of the implemented system.
In the pre-processing step, a low-pass filter was used
to remove the high-frequency noise, while the missing
data were compensated by the interpolation of adjacent
points on the facial surface. Only a rough automatic face

ROBUST 3D FACE RECOGNITION 5
localization was performed, based on the work in [16].
The local-feature vectors were extracted on a block-by-
block basis. From each block, 2D DCT coefficients were
calculated and the first ten low-frequency coefficients
were used to build the feature vector. The distribution
of feature vectors from each face image was described
by the GMM model. A GMM model, consisting of
K Gaussian components, is defined by the following
parameters: weights f k
g
K
k=1
, mean vectors f k
g
K
k=1
and covariance matricesf k
g
K
k=1
. The expectation max-
imization algorithm (EM) [29] was utilized to set these
parameters. Due to the small number of feature vectors
for each face, the parameters of the so-called universal
background model (UBM) were determined first. The
UBM is a GMM trained on all the feature vectors from
the training set. The GMM of each person was adapted
from the UBM by the maximum a posteriori estima-
tion (MAP) [30], where only the mean vectors were
adapted. GMM-based verification systems normally use
the likelihood ratio test for the classification task, while
in our system the SVM-based classifier is employed.
An unified representation of the images has to be made
to utilize the SVM classification. For this purpose, the
mean vectors from each face were stacked one over the
other to form the so-called supervector for each face.
In the enrollment phase, when the person is introduced
to the system, the SVM constructs the decision border
between the supervector of the enrolled person and the
supervectors from all the training images. In the test
phase, the claimant is accepted or rejected with respect
to the position of the claimant’s supervector to the
decision border.
The experiments were performed on the face recogni-
tion grand challenge (FRGC) data set [9]. Fig. 8 shows
the verification performance in the form of a receiver
operating characteristic curve or ROC. This curve plots
the false-acceptance rate against the true-acceptance
rate for all possible operating points, i.e., thresholds.
10
-4
10
-3
10
-2
10
-1
10
0
0.4
0.5
0.6
0.7
0.8
0.9
1
FAR
TAR
ROC
 
 
GMM/UBM
PCA
Figure 8. ROC curve of the proposed system compared to the
global PCA approach.
The presented method expectedly outperforms the
PCA approach. This stems mainly from the fact that in
the PCA approach, global-feature vectors are used, while
the GMM models utilize the local-feature vectors, result-
ing in an improved robustness to imprecise localization,
expression variations and occlusions. These factors are
comprised in most of the images from the FRGC data
set employed in our experiments.
7 CONCLUSION
In this paper we present some of the 3D face-recognition
systems. The whole recognition process is described
from the image-acquisition stage to the classification
task. Operation of such systems in an uncontrolled
environment is highlighted, where the recognition per-
formance can be affected by numerous variations during
the image acquisition. The recognition systems based
on local features are generally more robust to these
variations, as can also be seen from the results of our
experiments.
ACKNOWLEDGEMENT
The research leading to the above results has received
funding from the European Union’s Seventh Framework
Programme (FP7-SEC-2010-1) under grant agreement
number 261727 and the bilateral programme named Fast
and reliable 3D Face Recognition (BI-BG/11-12-007).
REFERENCES
[1] C. Busch and A. Nouak, “3d face recognition for unat-
tended border control.” in Security and Management’08,
2008, pp. 350–356.
[2] G. Medioni and R. Waupotitsch, “Face modeling and
recognition in 3-d,” in Proceedings of the IEEE Inter-
national Workshop on Analysis and Modeling of Faces
and Gestures, ser. AMFG ’03, 2003, pp. 232–240.
[3] S. Huq, B. Abidi, S. G. Kong and M. Abidi, “A survey
on 3d modeling of human faces for face recognition,” in
3D Imaging for Safety and Security, ser. Computational
Imaging and Vision, 2007, vol. 35, pp. 25–67.
[4] M. Zollh¨ ofer et al., “Automatic reconstruction of person-
alized avatars from 3d face scans,” Computer Animation
and Virtual Worlds (Proceedings of CASA 2011), vol. 22,
pp. 195–202, 2011.
[5] A. Mian, M. Bennamoun and R. Owens, “Face recogni-
tion using 2d and 3d multimodal local features,” ISVC
’06, vol. 860, pp. 860–870, 2006.
[6] A. Colombo, C. Cusano and R. Schettini, “3d face detec-
tion using curvature analysis,” Pattern Recogn., vol. 39,
pp. 444–455, 2006.
[7] S. Mehryar, K. Martin, K. Plataniotis and S. Stergiopou-
los, “Automatic landmark detection for 3d face image
processing,” in CEC ’10, 2010, pp. 1–7.
[8] T. Inan and U. Halici, “3-d face recognition with local
shape descriptors,” Information Forensics and Security,
IEEE Transactions on, vol. 7, no. 2, pp. 577–587, 2012.
[9] P. J. Phillips et al., “Overview of the face recognition
grand challenge,” in CVPR ’05, 2005, pp. 947–954.
[10] T. Heseltine, N. Pears and J. Austin, “Three-dimensional
face recognition: A fishersurface approach,” in Image

6 KRI
ˇ
ZAJ,
ˇ
STRUC, DOBRI
ˇ
SEK
Analysis and Recognition, ser. LNCS, 2004, vol. 3212,
pp. 684–691.
[11] C. Hesher, A. Srivastava and G. Erlebacher, “A novel
technique for face recognition using range imaging,” in
Proc. Seventh Int Signal Processing and Its Applications
Symp, vol. 2, 2003, pp. 201–204.
[12] B. Amor, M. Ardabilian and L. Chen, “New experiments
on icp-based 3d face recognition and authentication,”
in Pattern Recognition, 2006. ICPR 2006. 18th Interna-
tional Conference on, vol. 3, 2006, pp. 1195–1199.
[13] T. R. Lo and J. P. Siebert, “Local feature extraction and
matching on range images: 2.5d sift,” Comput. Vis. Image
Underst., vol. 113, pp. 1235–1250, 2009.
[14] C. Maes et al., “Feature detection on 3d face surfaces for
pose normalisation and recognition,” in BTAS ’10, 2010,
pp. 1–6.
[15] H. Bay, A. Ess, T. Tuytelaars and L. Van Gool, “Surf:
Speeded up robust features,” in Computer Vision and
Image Understanding (CVIU), 2008, pp. 346–359.
[16] P. Szeptycki, M. Ardabilian and L. Chen, “A coarse-to-
fine curvature analysis-based rotation invariant 3d face
landmarking,” in BTAS ’09, 2009, pp. 32–37.
[17] M. O. Irfanoglu, B. Gokberk and L. Akarun, “3d shape-
based face recognition using automatically registered
facial surfaces,” in ICPR’04, 2004, pp. 183–186.
[18] M. Husken, M. Brauckmann, S. Gehlen and C. V on der
Malsburg, “Strategies and benefits of fusion of 2d and
3d face recognition,” in CVPR’05, 2005, pp. 174–182.
[19] F. Cardinaux, C. Sanderson and S. Bengio, “User authen-
tication via adapted statistical models of face images,”
Signal Processing, IEEE Transactions on, vol. 54, pp.
361–373, 2006.
[20] C. McCool, J. Sanchez-Riera and S. Marcel, “Feature dis-
tribution modelling techniques for 3d face verification,”
Pattern Recogn. Lett., vol. 31, pp. 1324–1330, 2010.
[21] N. Aly¨ uz, B. G¨ okberk and L. Akarun, “Regional regis-
tration and curvature descriptors for expression resistant
3d face recognition,” in Signal Processing and Communi-
cations Applications Conference, 2009. SIU 2009. IEEE
17th, 2009, pp. 544–547.
[22] Y . Wang, C. S. Chua and Y . K. Ho, “Facial feature
detection and face recognition from 2d and 3d images,”
Pattern Recognition Letters, vol. 23, no. 10, pp. 1191–
1202, 2002.
[23] F. Al-Osaimi, M. Bennamoun and A. Mian, “Integration
of local and global geometrical cues for 3d face recogni-
tion,” Pattern Recognition, vol. 41, no. 3, pp. 1030–1040,
2008.
[24] J.-P. Vandeborre, V . Couillet and M. Daoudi, “A practical
approach for 3d model indexing by combining local and
global invariants,” in 3D Data Processing Visualization
and Transmission. Proceedings, 2002, pp. 644–647.
[25] B. G¨ okberk, A. Salah and L. Akarun, “Rank-based
decision fusion for 3d shape-based face recognition,” in
Audio- and Video-Based Biometric Person Authentica-
tion, ser. LNCS, 2005, vol. 3546, pp. 1019–1028.
[26] C. McCool, V . Chandran, S. Sridharan and C. Fookes,
“3d face verification using a free-parts approach,” Pattern
Recogn. Lett., vol. 29, no. 9, pp. 1190–1196, 2008.
[27] H. Bredin, N. Dehak and G. Chollet, “Gmm-based svm
for face recognition,” in ICPR ’06, vol. 3, 2006, pp. 1111
–1114.
[28] A. Moreno, A. Sanchez, J. Velez and J. Diaz, “Face
recognition using 3d local geometrical features: Pca vs.
svm,” in ISPA ’05, 2005, pp. 185–190.
[29] T. Moon, “The expectation-maximization algorithm,”
Signal Processing Magazine, IEEE, vol. 13, pp. 47–60,
1996.
[30] D. A. Reynolds, T. F. Quatieri and R. B. Dunn, “Speaker
verification using adapted gaussian mixture models,” in
Digital Signal Processing, 2000, pp. 19–41.
Janez Kriˇ zaj received his B.Sc. degree in 2008 from the Faculty
of Electrical Engineering of the University of Ljubljana. Currently
he is a Ph.D. student and junior researcher at the Laboratory of
Artificial Perception, Systems and Cybernetics at the same faculty.
His research interests include computer vision, image processing and
pattern recognition.
Vitomir
ˇ
Struc received his B.Sc. and Ph.D. degrees in electrical
engineering from the University of Ljubljana in 2005 and 2010,
respectively. He is currently working as a PostDoc at the Laboratory
of Artificial Perception, Systems and Cybernetics at the same faculty.
His research interests include pattern recognition, machine learning
and biometrics.
Simon Dobriˇ sek received his B.Sc., M.Sc. and Ph.D. degrees in
electrical engineering in 1990, 1994 and 2001, respectively, from
the Faculty of Electrical Engineering of the University of Ljubljana.
In 1990 he became a Research Staff Member in the Laboratory of
Artificial Perception, Systems and Cybernetics at the same faculty,
where he is presently a teaching assistant and research fellow. His
research interests include pattern recognition and artificial intelligence.
He participates in research projects on spoken language technology
and biometric security systems.