Feature
Level Fusion of Face and
Fingerprint
Biometrics
A. Rattani, D. R. Kisku, M. Bicego, Member, IEEE and M. Tistarelli
Abstract- The aim of this paper is to study the fusion at
feature extraction level for face and fingerprint biometrics. The
proposed approach is based on the fusion of the two traits by
extracting independent feature pointsets from the two modalities,
and making the two pointsets compatible for concatenation.
Moreover, to handle the 'problem of curse of dimensionality', the
feature pointsets are properly reduced in dimension. Different
feature reduction techniques are implemented, prior and after
the feature pointsets fusion, and the results are duly recorded.
The fused feature pointset for the database and the query face
and fingerprint images are matched using techniques based on
either the point pattern matching, or the Delaunay triangulation.
Comparative experiments are conducted on chimeric and real
databases, to assess the actual advantage of the fusion performed
at the feature extraction level, in comparison to the matching
score level.
Index Terms- Face, Feature level fusion, Fingerprint,
Multimodal Biometrics
I. INTRODUCTION
In recent years, biometric authentication has seen
considerable improvement in reliability and accuracy, with
some of the traits offering good performance. However none
of the biometrics are 00% accurate. Multibiometric systems
[1] remove some of the drawbacks of the uni-biometric
systems by grouping the multiple sources of information.
These systems utilize more than one physiological or
behavioral characteristic for enrollment and verification/
identification. Ross and Jain [2] have presented an overview
of Multimodal Biometrics with various levels of fusion,
namely, sensor level, feature level, matching score level and
decision level.
However it has been observed that, a biometric system that
integrates information at an earlier stage of processing is
Manuscript received May 25, 2007. This work has been partially supported
by grants from the Italian Ministry of Research, the Ministry of Foreign
Affairs and the Biosecure European Network of Excellence.
A. Rattani is with DIEE, University of Cagliari, Piazza d'Armi, 09123
Cagliari- Italy( Phone: +39 348 5579308 Fax: +39 070 6755782 email:
aj ita.rattani(odiee.unica. it)
D. R. Kisku is with DCSE, Indian Institute of Technology Kanpur,
208016 Kanpur- India (email: drkisku oiitk.ac.in).
M. Bicego is with DEIR, University of Sassari, via Torre Tonda 34,
07100 Sassari- Italy (e-mail: bicego ouniss.it).
M. Tistarelli is with DAP, University of Sassari, piazza Duomo 6,
07041 Alghero (SS) - Italy (Phone: +39 079 9720410, Fax: +39 079 9720420
e-mail: tista d,uniss.it).
978-1-4244-1597-7/07/$25.00 ©2007 IEEE
expected to provide more accurate results than the systems
that integrate information at a later stage, because of the
availability of more richer information. Since the feature set
contains much richer information on the source data than the
matching score or the output decision of a matcher, fusion at
the feature level is expected to provide better recognition
performances.
Fusion at matching score, rank and decision levels have
been extensively studied in the literature [3][4]. Despite the
abundance of research papers related to multimodal
biometrics, fusion at feature level is a relatively understudied
problem. As a general comment, it is noticed that fusion at
feature level is relatively difficult to achieve in practice
because multiple modalities may have incompatible feature
sets and the correspondence among different feature spaces
may be unknown. Moreover, concatenated feature set may
lead to the problem of curse of dimensionality: a very complex
matcher may be required and the concatenated feature vector
may contain noisy or redundant data, thus leading to a
decrease in the performance of the classifier [5]. Therefore, in
this context, the state of the art is relatively poor.
Ross and Govindarajan [5] proposed a method for the
fusion of hand and face biometrics at feature extraction level.
Gyaourova et al. [6] fused IR-based face recognition with
visible based face recognition at feature level, reporting a
substantial improvement in recognition performance as
compared to matching individual sensor modalities. Recently,
Ziou and Bhanu [7] proposed a multibiometric system based
on the fusion of face features with gait features at feature
level.
Even though face and fingerprint represent the most widely
used and accepted biometric traits', no methods for feature
level fusion of these modalities have been proposed in the
literature. The possible reason is the radically different nature
of face and fingerprint images: a face is processed as a
pictorial image (holistic approach) or as composed by patches
(local analysis), while fingerprint is typically represented by
minutiae points. In this paper a recently introduced
methodology for face modeling [8] is exploited, which is
based on a point-wise representation of a face called Scale
Invariant Features Transform (SIFT), thus making the feature
level fusion of face and fingerprints possible.
l Remarkably face and fingerprint data has been adopted as biometric traits
to be included in the European electronic passport. The same traits are
currently being used at the US immigration for manual identification of
passengers.
Thus, this paper proposes a novel approach to fuse face and
fingerprint biometrics at feature extraction level. The
improvement obtained applying the feature level fusion is
presented over score level fusion technique. Experimental
results on real and chimeric databases are reported, confirming
the validity of the proposed approach in comparison to fusion
at score level.
II. FACE AND FINGERPRINT BIOMETRICS
A. Face Recognition based on Scale Invariant Feature
Transform Features (SIFT)
The face recognition system, preliminary introduced in [8],
is based on the SIFT [9] features extracted from images of the
query and database face. The SIFT features represent a
compact representation of the local gray level structure,
invariant to image scaling, translation, and rotation, and
partially invariant to illumination changes and affine or 3D
projections. SIFT has emerged as a very powerful image
descriptor and its employment for face analysis and
recognition was systematically investigated in [8] where the
matching was performed using three techniques: (a) minimum
pair distance, (b) matching eyes and mouth, and (c) matching
on a regular grid. The present system considers spatial,
orientation and keypoint descriptor information of each
extracted SIFT point. Thus the input to the present system is
the face image and the output is the set of extracted SIFT
features s=(s1, S2 .... Sm) where each feature point si=(x ,y ,0,
k) consist of the (x, y) spatial location, the local orientation 0
and k is the keydescriptor of size 1x128.
....
B. Fingerprint Verification based on Minutiae matching
technique
The fingerprint recognition module has been developed
using minutiae based technique where fingerprint image is
normalized, preprocessed using Gabor filters, binarized and
thinned, is then subjected to minutiae extraction as detailed in
[10]. However to achieve rotation invariance the following
procedure is followed in the image segmentation module.
In order to obtain rotation invariance, the fingerprint image
is processed by first detecting the left, top and right edges of
the foreground. The overall slope of the foreground is
computed by fitting a straight line to each edge by linear
regression. The left and right edges, which are expected to be
roughly vertical, are fitted with lines of the form x = my + b
and for the top edge the form y = mx + b is applied. The
overall slope is determined as the average of the slopes of the
left-edge line, the right-edge line, and a line perpendicular to
the top edge line. A rectangle is fitted to the segmented region
and rotated with the same angle to nullify the effect of
rotation. Although the method is based on the detection of
edges, only a rough estimate of the fingerprint boundaries is
required for fitting the lines and extracting the edges. This
improves the robustness to noise in the acquired fingerprint
image. The input to this system is a fingerprint image and the
output is the set of extracted minutiae m=(Min2,m2 .... mm),
....
where each feature point mi=(x ,y ,0) consist of the spatial
location (x, y) and the local orientation 0.
FEATURE LEVEL FUSION SCHEME
The feature level fusion is realized by simply concatenating
the feature points obtained from different sources of
information. The concatenated feature pointset has better
discrimination power than the individual feature vectors. The
concatenation procedure is described in the following sections.
III.
A. Feature set compatibility and normalization
In order to be concatenated, the feature pointsets must be
compatible. The minutiae feature pointset is made compatible
with the SIFT feature pointset by making it rotation and
translation invariant and introducing the keypoint descriptor,
carrying the local information, around the minutiae position.
The local region around each minutiae point is convolved
with a bank of Gabor filters with eight different equally
spaced degrees of orientation (0, 22.5, 45, 67.5, 90, 112.5
,135, and 157.5), eight different scales and two phases (O and
w/2 ), giving a keydescriptor of size 1x128. The rotation
invariance is handled during the preprocessing step and the
translation invariance is handled by registering the database
image with the query images using a reference point location
[11]. Scale invariance is achieved by using the dpi
specification of the sensors. The keypoint descriptors of each
face and fingerprint points are then normalized using the minmax normalization technique (Snorm and Mnorm ), to scale all the
128 values of each keypoint descriptor within the range 0 to 1.
This normalization also allows to apply the same threshold on
the face and fingerprint keypoint descriptors, when the
corresponding pair of points are found for matching the fused
pointsets of database and query face and fingerprint images.
B. Feature Reduction and Concatenation
The feature level fusion is performed by concatenating the
two feature pointsets. This results in a fused feature pointset
concat=(Slnorm , S2norm, Smnorm mlnorm, m2norm, mmnorm).
Feature reduction strategy to eliminate irrelevant features can
be applied either before [7] or after [5-6] feature
concatenation.
C. Feature Reduction techniques
1. K-means clustering. The normalized feature pointsets
(Snorm and Mnorm) are first concatenated together (concat).
Redundant features are then removed using the "k-means"
clustering techniques [12] on the fused pointset of an
individual retaining only the centroid of the points from each
cluster. These clusters are formed using spatial and orientation
information of a point. The keypoint descriptor of each
cluster's centroid is the average of keypoint descriptors of all
the points in each cluster. The distance classifier used is
euclidean distance. The number of clusters are determined
using the PBM cluster validity index [13]. Since, the feature
poinset from the two modalities i.e., face and fingerprint are
affine invariant and moreover, they are nonnalized using
normalization technique as discussed before. They are treated
simply as a set of points belonging to an individual irrespective of whether they are extracted from face or
fingerprint thus making K-means clustering possible.
2. Neighbourhood Elimination. This technique is applied on
the normalized pointset of face and fingerprint (s1Orm and
mnorm) individually. That is, for each point of face and
fingerprint, those point that lie within the neighbourhood of a
certain radius (20 and 15 pixels for face and fingerprint
respectively on experimental basis) are removed giving
snorm'and mnorm', the reduced face and fingerprint pointsets.
Spatial information is used to determine the neighbours of
each considered point. The result of neighbourhood
elimination is shown in Fig. 1.
3. Points belonging to specifc regions. Only the points
belonging to specific regions of the face i.e., specific
landmarks like the eyes, the nose and the mouth lower portion
and the fingerprint images (the central region) are retained as
reduced pointset. Face images in BANCA Database are preregistered with respect to eyes and mouth location and the
nose tip is manually identified for the current experiments.
The corepoint in fingerprint is located using a reference point
location algorithm discussed in [11]. A radius equal to 85 and
120 pixels was set for the face and fingerprint feature points
selection as shown in Fig. 2. The SIFT points around specific
landmarks on face carry highly discriminative information as
experimented and reported in [8]. The region around core
point accounts for combating the effect of skin elasticity and
non-linear distortion due to varing pressure applied during
image acquisition as it is the least affected region.
a)
Fig. 2 Example of selected regions on a face (left) and a fingerprint (right)
The aims of the "k-means" and "neighbourhood
elimination" techniques are to remove redundant information
and at the same time retaining most of the information by
removing onlyl the points which are very near, as computed
using euclidean distance, to a specific point. As these points
may not provide any additional information because of being
in vicinity. And the aim of "points belonging to specific
region" is to consider only the points belonging to highly
distinctive region. Thus keeping only optimal sets.
D. Matching techniques
The concatenated features pointset of the database and the
query images concat and concat' respectively (in which the
feature reduction techniques have already been applied even
before or after concatenation) are processed by the matcher
which gives matching score based on the no. of matching pairs
found between the two pointsets. In this study two different
matching techniques are applied.
1. Point pattern matching. This technique aims at finding
the percentage of points "paired" between the concatenated
feature pointset of the database and the query images. Two
points are considered paired only if the spatial distance (1), the
direction distance (2) and the Euclidean distance (3) between
the corresponding key descriptors are all within some are
within a pre-determined threshold, set with 4 pixels, 30, 6
pixels for r0, O0, ko on the basis of experiments:
(1)
sd (concat>, concat) =(X Xi)2 + (Y -t)2 <r0
dd(concat', concat,) = min(6! 6t ,3600 |6 6t) < 00
-
-
-
(2)
(k
k)2 < ko
j,concat i) =
(3)
where the points i andj are represented by (x, y, 0, k) with k
k'... k'28 of the concatenated database and query pointsets
concat' and concat, sd is the spatial distance, dd is the
direction distance, and kd is the keypoint descriptor distance.
The one to one correspondence is achieved by selecting
among the candidates points lying within the threshold of
spatial, direction and Euclidean distance, the one having
mimimum Euclidean distance for the keypoint descriptor.
Since, the feature pointsets are rotation, scale and translation
invariant, in case of fingerprint, the registartion is done at
image preprocessing level as explained earlier and SIFT
features for face are already affine invariant features. This
kd (concat
b)
Fig. 1. Effects of the neighbourhood elimination on a) Fingerprint
and b) Face
obviates the need to calculate transformation parameters for
aligning the database and query fused pointsets.
The final matching score is calculated on the basis of the
ratio of the number of matched pairs to the total number of
feature points found in the database and query sets, for both
monomodal traits and for the fused feature pointset.
a)
b)
Fig. 3. Triangulation of pointset: a)Voronoi diagram b)Delaunay triangulation
2. Matching using the Delaunay Triangulation technique. In
this case, instead of considering individual points, triplet of
points are grouped together as new features. Given a set S of
points p, P2, ..., pN, the Delaunay triangulation of S is obtained
by first computing its Voronoi diagram [14] which
decomposes the 2D space into regions around each point such
that all the points in the region around Pi are closer to Pi than
delaunay triangulation is computed by connecting the centers
of every pair of neighboring Voronoi regions.
The Delaunay triangulation technique [15] is applied
individually on the face and the fingerprint normalized
pointset Snorm and Mnorm and then on the concatenated feature
pointsets concat=(Snorm, Mnorm). Five features are computed
from the extracted triplet of points. (a) The minimum and
median angles (amin amed) of each triangle (b) The triangle side
(L) with the maximum length (c) The local orientation (0) of
the points at the triangle vertexes (d) The ratio (11/12) of the
smallest side to the second smallest side of each triangle (e)
The ratio (12/13) of the second smallest side to the largest side
of each triangle.
All these parameters compose the feature vector fv=(tl, t2
t.,), where ti= (amin,, amed, L, 0, 11/12, 12/13) is the triangle
computed by the Delaunay triangulation. The process is
repeated for the database and the query pointsets to getfv and
fv '. The final score is computed on the basis of the number of
corresponding triangles found between the two feature vectors
fv and ft'. Two triangles are correctly matched if the
difference between the attributes of the triangles ti and ti ' are
within a fixed threshold. As the fused poinset contain affine
invariant and pre-normalized points thus making the
application of delaunay triangulation possible.
IV. EXPERIMENTAL RESULTS
The proposed approach has been tested on two different
databases: the first consists of 50 chimeric individuals
composed of 5 face and fingerprint images for each individual.
The face images are taken from the controlled sessions of the
BANCA Database [ 16] and the fingerprint images were
collected by the authors. The fingerprint images were acquired
using an optical sensor at 500 dpi.
The following procedure has been established for testing the
mono-modal and multimodal algorithms:
Training: one image per person is used for enrollment in
the face and fingerprint verification system; for each
individual, one face-fingerprint pair is used for training the
fusion classifier.
Testing: four remaining samples per person are used for
testing, generating client scores. Impostor scores are generated
by testing the client against the first sample of all other
subjects. For the multimodal testing, each client is tested
against the first face and fingerprint samples of the rest of the
chimeric users. In total 5Ox4=200 client scores and
5Ox49=2450 imposters scores for each of the uni-modal and
the multimodal systems are generated.
Evaluation: The best combination of feature reduction and
matching strategy has been further tested on a real multimodal
database acquired by the authors. The database consists of 100
individual with four face and fingerprint images per person.
The first face and fingerprint combination is used for training
and the rest three image pairs are used for testing, providing
lOOx3=300 client scores. Each individual is subject to
imposter attack by ten random face and fingerprint pairs for a
total of lOOxlO=1000 impostor scores. The experiments were
conducted in four sessions recording False Acceptance Rate
(FAR), False Rejection Rate (FRR) and Accuracy (which is
computed at the certain threshold, FAR and FRR where the
performance of the system is maximum ie., max (1-(FAR +
FRR)/2).
A. The face and the figerprint recognition systems were
tested alone, without any modification in the feature sets, i.e.
SIFT features (x, y, 0 , k) and minutiae features (x, y, 0). The
matching score is computed using point pattern matching
independently for face and fingerprint. The individual system
performance was recorded and the results were computed for
each modality as shown in table 1.
B. In the second session, the effect of introducing the
keydescriptor around each minutiae point is examined. Once
the feature sets are made compatible, the keypoint descriptors
of SIFT and the minutiae points are normalized using the
min-max normalization technique. The normalized feature
pointsets are then concatenated and the k-means feature
reduction strategy is applied on each fused pointset.
From the presented results (table 2), it is evident that the
introduction of the keydescriptor for the fingerprints increased
the recognition accuracy by 1.64%, and the feature level
fusion outperformed both single modalities, as well as the
score level fusion, with an increase in the accuracy of 2.64%
in comparison to score level. The score level fusion is
performed scores independently for face and fingerprint are
computed independently for face and fingerprints which are
then normalized and added using sum of scores technique.
C. In the third session, to remove redundant features, two
feature reduction strategies are applied prior to concatenation.
The matching is performed with the point pattern matching
technique. From the experimental results, presented in table 3,
it is evident that the application of the neighborhood removal
technique does not increase the accuracy of the system. On the
other hand, the reduction of points belonging to specific
regions increased the recognition accuracy by 0.31%, while
the FRR is dropped to 000. Some statistics regarding the
number of points retained in the fused poinsets, for all the
three feature reduction techniques applied to one subject, are
listed in table 4 and the performances are depicted in table 3.
TABLE 1. THE FAR, FRR AND ACCURACY VALUES OBTAINED FROM THE
MONOMODAL TRAITS
Algorithm
experiments, with the increase in recognition accuracy of
0.35%. Finally, the combination of restricting the points to
those belonging to specific regions and the Delaunay
triangulation further enhanced the recognition accuracy by
0.4400.
This last configuration was further tested on the multimodal
database acquired by the authors with multimodal fusion at
score level and feature level. The results, presented in table 6,
also demonstrate that the feature level fusion outperforms the
score level fusion of 0.67%, also for the real multimodal
database. The ROC curve obtained from the best strategy
applied to the chimeric and the real multimodal databases is
shown in Fig. 6.
Face SIFT
Fingerprint
Face
TABLE 2. FAR, FRR AND ACCURACY VALUES OBTAINED FROM THE
G
MULTIMODAL FUSION
Algorithm
FRR (%)
FAR (%)
Accuracy
Fingerprint
(Face+Finger)
score level
(Face+Finger)
Feature Level
5.384
10.97
91.82
5.66
4.78
94.77
1. 98
3.18
97.41
FII e r
56S.
mat
e
-N
X
Eg
scie
eI
J
I
"
[m
70
TABLE 3. FAR, FRR AND ACCURACY VALUES FOR THE FEATURE REDUCTION
1
20 i
Si
0
40
50
T es lool
:
"I
70
80
90
1i
Fig. 4. The Accuracy Curve for Delaunay Triangulation of face, fingerprint,
fusion at matching score and feature level
TABLE 4. STATISTICS REGARDING THE NUMBER OF POINTS RETAINED IN THE
THREE FEATURE REDUCTION TECHNIQUES I.E., K-MEANS, NEIGHBOURHOOD
ELIMINATION AND POINTS BELONGING TO SPECIFIC LOCATIONS
Algorithm
Face
Finger
Fused
(Minutiae)
pointset
Algorithm
(SIFT)
Extracted features
K-means clustering
145
50
195
145
50
89
technique
Neighbourhood
removal technique
Points belonging
to specific regions
73
25
98
47
20
67
D. In the fourth session, the matcher based on the Delaunay
triangulation of the poinsets is introduced. The reported results
are computed for monomodal modalities, and multimodal
fusion at matching score and feature extraction level. In the
first case, all the feature points were included for triangle
computation, in a second case only the reduced set of points
was used. The results presented in Fig. 4, Fig. 5 and table 5,
show that the application of the Delaunay triangulation
enhances the performance of the face and fingerprint
modalities alone by 5.0500 and 0.82%, respectively.
Moreover, the multimodal feature level fusion using the
Delaunay triangulation outperforms all the feature level fusion
TABLE 5. FAR, FRR AND ACCURACY VALUES FOR THE DELAUNAY
TRIANGULATION TECHNIQUE
Algorithm
FRR
FAR
Accuracy
Face SIFT
2.24
9.85
93.95
Fingerprint
13.63
3.07
92.64
Face+Finger at Matching level
2.95
8.07
94.48
Face+Finger at Feature Level
Face+Finger at Feature level
using points belonging to
2.95
0.89
98.07
1.95
1.02
98.51
specific_region strategy
TABLE 6. FAR, FRR AND ACCURACY OF THE BEST MATCHING AND FEATURE
REDUCTION STRATEGIES
Algorithm
Best strategy at
score
fusion
Best strategy at feature fusion
FRR(%)
FAR(%)
Accuracy
2.54
5.48
95.99
1.12
4.95
96.66
independent/ uncorrelated sources (face and fingerprint) at the
feature level fusion increases the performance as compared to
score level. This further demonstrates that ensemble of
classifiers operating on uncorrelated features increases the
performance in comparison to correlated features.
Further experiments, on "standard" multimodal databases,
will allow to better validate the overall system performances.
4D
REFERENCES
m
[1]
Is
10
[2]
5
[3]
1)
IIl
1
20
2
30
to
V
Fig. 5. The ROC Curve for Delaunay Triangulation of face, fingerprint, fusion
at matching score and feature level
50 ''
-----A ohimerio database
PA nultimodaI
[4]
[5]
[6]
4.0
[7]
30
cr
[8]
26
20~F
[9]
10
10
[10]
1).
0
I z=
5
4FI--R
10
-
-t ----1
15
20
25
30
35
FAR
4011
45
50
Fig. 6. ROC Curve for the best strategy in chimeric and multimodal database
[11]
[12]
[13]
V. CONCLUSION
A multimodal biometric system based on the integration of
face and a fingerprint traits at feature extraction level was
presented. These two traits are the most widely accepted
biometrics in most applications. There are also other
advantages in multimodal biometric systems, including the
easy of use, robustness to noise, and the availability of lowcost, off-the-shelf hardware for data acquisition.
In this paper a novel approach has been presented where
both fingerprint and face images are processed with
compatible feature extraction algorithms to obtain comparable
features from the raw data. The reported experimental results
demonstrate remarkable improvement in the accuracies by
properly fusing feature sets. This preliminary achievement,
does not constitute an end in itself, but rather suggests to
attempt a multimodal data fusion as early as possible in the
processing pipeline. In fact, the real feasibility of this
approach, in a real application scenario, may heavily depend
on the physical nature of the acquired signal. The
experimental results demonstrate that fusing information from
View publication stats
[14]
[15]
[16]
A. K. Jain and A. Ross, "Multibiometric systems", Communications of
the ACM, 47 (1), pp. 34-40, 2004.
A. Ross and A.K. Jain, "Information Fusion in Biometrics", Pattern
Recognition Letters, 24, pp. 2115-2125, 2003.
J. Fierrez-Aguilar, J. Ortega-Garcia and J. Gonzalez-Rodriguez, "Fusion
strategies in Biometric Multimodal Verification", Proceedings of
International Conference on Multimedia and Expo, ICME 2003.
] L. Hong and A. Jain, "Integrating Faces and Fingerprints for Personal
Identification", IEEE Transactions on Pattern Analysis and Machine
Intelligence, 20(12), pp. 1295-1307, December 1998.
A. Ross and R. Govindarajan, "Feature Level Fusion Using Hand and
Face Biometrics", Proceedings of SPIE Conference on Biometric
Technology for Human Identification II, Orlando, USA, pp. 196-204,
March 2005.
S. Singh, G. Gyaourova and I. Pavlidis, "Infrared and visible image
fusion for face recognition", SPIE Defense and Security Symposium,
pp.585-596, 2004.
X. Zhou and B. Bhanu, "Feature fusion of face and Gait for Human
Recognition at a distance in video", International Conference on Pattern
Recognition, Hongkong, 2006.
M. Bicego, A. Lagorio, E. Grosso and M. Tistarelli, "On the use of
SIFT features for face authentication", Proceedings of CVPR
Workshop, New York, 2006.
Lowe and G. David, "Object recognition from local scale invariant
features," International Conference on Computer Vision, Corfu, Greece,
pp. 1150- 1157, September 1999.
A. Jain, L. Hong, and R. Bolle, "On-line fingerprint verification", IEEE
Transactions on Pattern Analysis and Machine Intelligence, 19(4), pp.
302-314, 1997.
A. Jain, S. Prabhakar, L. Hong, and S. Pankanti, "Filterbank-based
fingerprint matching," IEEE Transactions on Image Processing, 9(5), pp.
846- 859, 2000.
A. Jain and R. Dubes, " Algorithms for clustering data", Prentice Hall,
1988.
M. K. Pakhira, S. Bandyopadhyay and U. Maulik (2004). "Validity
index for crisp and fuzzy clusters". Pattern Recognition, vol. 37, pp.
487-50 1
N. Ahuja, "Dot pattern processing using voronoi neighborhoods", IEEE
Pattern Analysis and Machine Intelligence, 4, pp. 336-343, 1982.
G. Parziale and A. Niel. "A fingerprint matching using minutiae
triangulation". Proceedings of ICBA, pp. 241-248, 2004.
E. Bailly-Baillire, S.Bengio, F. Bimbot, M. Hamouz, J. Kittler, J.
Marithoz, J. Matas, K. Messer, V. Popovici, F. Pore, B. Ruiz, and J. P
Thiran, "The BANCA database and evaluation protocol", In Proceedings
of International Conference on Audio- and Video-Based Biometric
Person Authentication (AVBPA03), Springer-Verlag , pp. 625-638,
2003.