FINGERPRINT CLASSIFICATION USING KOHONEN TOPOLOGIC MAP

Sylvain BERNARD1,2, Nozha BOUJEMAA1, David VITALE2, Claude BRICOT2

1
INRIA Rocquencourt BP 105, F-78153 Le Chesnay, France
2
THALES Identification, 41 bd de la Republique, BP 53- 78401 Chatou Cedex, France
sylvain.bernard@inria.fr

ABSTRACT
Self Organizing Maps are efficient and usual for
dimension reduction and data clustering. In our present
work, we propose the use of Kohonen Topologic Map for
fingerprint pattern classification. The learning process
takes into account the large intra-class diversity and the
continuum of fingerprint pattern types. After a brief
introduction to fingerprint domain-specific knowledge and
the expert approach, we present an original and intuitive
description of the algorithm. For a classification based Left Loop – 32 % Rigth Loop – 32 %
on the global shape of the fingerprint, we adopted a
suitable feature space. Indeed we obtained 88% of correct
classification on a database composed of 1600 NIST
fingerprints.

1. INTRODUCTION
Most automatic systems for fingerprint comparison are
based on minutiae matching. Minutiae points [1] are
terminaisons and bifurcations of the ridge lines that Whorl – 32% Arch – 4 %
Fig. 3 – Experts' Fingerprint Classification
constitute a fingerprint pattern. To demonstrate that two
O : Core point ∆ : Delta point
fingerprints originate from the same finger or not, human
experts detect the ridge ending and bifurcation points of Class Number Number Relative
both fingerprints (Fig.1 and Fig.2) and match the two of Cores of Deltas Positions
minutiae sets by superposition to count the number of L : Left Loop 1 Core 1 Delta Delta is at the
common points. Two fingerprints are considered to be right of Core
from the same finger if the number of common points is R : Right Loop 1 Core 1 Delta Delta at the
sufficient, depending on the country’s legislation. left of Core
W: Whorl 2 Cores 2 Deltas X
A : Arch 0 Core 0 Delta X
Fig. 4 : Number of Core(s), Delta(s) and Relative Positions per
Fingerprint Class
Fig. 1 – Ridge Ending Fig. 2 – Ridge Bifurcation Several approaches have been tested for automatic
For Very Large DataBase applications, since the automatic classification. In [2][3][4], authors propose classification
comparison of fingerprints by matching their minutiae sets algorithms based on the number of cores and deltas, and
is time consuming, matching a query fingerprint with the their relative positions (Fig.4). The major problem with
entire database would be computationally intensive. We such approaches is that sometimes the delta can be out of
match the fingerprint with a subset of the database the image, especially with small sensors. In [5], Jain,
(reduced search-space) using a fast classification process. Prabhakar & Hong propose a method that uses only one
core point and a circular region around the core. The
A well-known expert classification is based on the global region of interest is divided into sectors; each sector is
shape of the fingerprints [1] statistically observed in the normalized and filtered using a bank of Gabor filters to
population (Fig. 3). produce a set of filtered images; the variance of gray
values in a sector is taken as a feature. Using a K-nearest M
neighbor classifier and a set of neural networks, the feature Using (1), D( R1 , R2 ,..., RM ) = ∑ ∑d 2
(X k , R j )
vector is classified into one of the classes. In [6], the j =1 X k ∈C j
proposed method is based on the global shape information
∂D ∂  
provided by the fingerprint direction map. The core point =  ∑ d ( X k , Ri ) = −2 ⋅ ∑ ( X k − Ri )
2
is required for alignment and a Karhunen Loeve ∂Ri ∂Ri  X k ∈Ci  X k ∈Ci
Transform is used to reduce the feature space dimension.
Then the lower-dimensional vector is send to a At time t of the descent, each Ri (t ) i =1..M moves and
Probabilistic Neural Network classifier. In [7], the lower- becomes Ri (t + 1) i =1..M :
dimensional vector is send to Fuzzy Neural Net classifier.
In this work, we present an unsupervised classifier that ∂D
uses a rotation-translation aligned direction map, as an Ri (t + 1) = Ri (t ) − ε ⋅ , ε is the learning rate
∂Ri
input. The Self Organizing Map (SOM) used takes into Ri ( t )
account not only the large intra-class diversity but also the
continuum of fingerprint pattern types.
Ri (t + 1) = Ri (t ) + 2 ⋅ ε ⋅ ∑ (X
X k ∈Ci
k − Ri (t ) )

2. PRESENTATION OF SOM Ri (t + 1) = Ri (t ) + 2 ⋅ ε ⋅ ∑X k − 2⋅ε ⋅ ∑ R (t ) i

Suppose we want to cluster a data point set {X } N

k k =1 of R 2 X k ∈Ci X k ∈Ci

in M classes. A simple prototype-based clustering Ri (t + 1) = Ri (t ) ⋅ [1 − 2ε ⋅ Card(C i )] + 2ε ⋅ ∑X k

algorithm called deterministic Forgy algorithm can be X k ∈Ci
used. The stochastic version allows an example by
where Card (C i ) is the class Ci cardinality.
example learning process and a neighborhood constraint
yields the SOM.
1
∑ Xk
X k ∈Ci
If we suppose that to each class Ci , is associated a unique With ε = then Ri (t + 1) =
prototype Ri and given the rule (1) : 2 ⋅ Card (C i ) Card (C i )
(1) X k ∈ C i ⇔ Ri = Argmin d 2 ( X k , R j ) , The algorithm is then : first, all the prototypes are
j
randomly initialized.
the space is then partitioned in Voronoï cells (Fig.5). The
space partition is expected to have good generalization For each prototype Ri, find all the data points that are
properties : the misclassification rate of new data points closer to Ri than all other prototypes. Those points
should be as small as possible. So, a good clustering constitute the class Ci. Ri becomes the mass-center of Ci.
algorithm should place the prototypes in the middle of Go back to until stabilization.

areas where there is a high concentration of data points. • The stochastic Forgy algorithm consists of a
minimization of the function (2) using a stochastic
Fig. 5 – Example of gradient descent technique.

[ ]
N N
D( R1 , R2 ,..., RM ) = ∑ min d 2 ( X k , R j ) = ∑ Dk*
Voronoï cells. Each point
is a class prototype and j
k =1 k =1
each cell delimits a
region of the space. A stochastic descent on D consists of a gradient descent on
each term of the sum : Dk* = d 2 ( X k , R * ) .
• The deterministic Forgy algorithm consists of a R * = Argmin d 2 ( X k , R j ) is the nearest prototype to
j
minimization of the function (2) using a deterministic
gradient descent technique. Xk . Data point by data point (this is an example by
R * (t ) moves and becomes R * (t + 1) :
[ ]
N example process),
(2) D( R1 , R2 ,..., RM ) = ∑ min d ( X k , R j ) .
2

k =1
j ∂Dk*
R * (t + 1) = R * (t ) − ε ⋅ ,
The distance D, to be minimized, representants the sum of ∂R * *
R (t )
the distances between each data point and its nearest
∂D
( )
*
prototype.
with = −2 ⋅ X k − R *
k

∂R *
so, (3) R
*
(
(t + 1) = R * (t ) + 2 ⋅ ε ⋅ X k − R * (t ) . )
The algorithm is then : first, all the prototypes are
randomly initialized and fix empirically the learning rate ε.
For each data point Xk, find the nearest prototype R*. R*


moves in the direction of Xk according to eqn.(3).

Decrease ε and go back to until stabilization.


• The SOM [8][9], is a low-dimensional map placed in

the feature space to approximate the data distribution.
Unlike the Karhunen Loeve Transform, the SOM provides Global Orientation Local Directions
a non-linear approximation (Fig.6).

Fig. 6 – 3D Distribution Approximation with a 2D Map

The map is sampled in a grid, each node being a
prototype. The previous stochastic Forgy algorithm can fit
the data but, in addition, prototypes have to respect a
neighbor constraint linking them to each other, because
they constitute the map.
The algorithm is then : first, all the prototypes are
randomly initialized and fix empirically ε and σ. Direction Map
* Fig. 7 : Fingerprint Features
R


for all data points, find the nearest prototype x, y ,

having (x,y) coordinates in the topologic map. Because two fingerprint images from the same finger
differs in terms of translation and rotation, we need to
R x*, y and its neighbors move in the map : ∀m, n , normalize the database direction maps using reference
(
R x*+ m, y + n ← R x*+ m, y + n + 2ε ⋅ Gσ (m, n) ⋅ X k − R x*+ m, y + n ) points and a global orientation.
From the direction map, we automatically detect the core
Gσ being a gaussian of variance σ. and delta points. Indeed, we observe that the variation of
Decrease ε and σ and go back to  until stabilisation of directions around core and delta points is high. So, the
the prototypes. calculated variance of directions has to be superior than a
thresold to be in presence of a core or delta point. We
3. AUTOMATIC CLASSIFICATION combine this information with the Poincaré index [4] to
allow a better detection and differ a core from a delta.
• Feature Space For rotation compensation, we compute the global
In a suitable feature space, fingerprints from the same orientation of the fingerprint using a Principal Component
class are expected to be close, creating clusters. Analysis of directions around the core point.
The expert classification is based on the global shape of
• SOM for Automatic Classification :
the fingerprint. Using a Gabor Wavelet filter bank having
8 different orientations and 3 different frequencies [10], If the 4 classes were clearly separated, the Forgy algorithm
we calculate the local direction of ridges and valleys in with 4 prototypes would be sufficient. In reality, with Very
each pixel neighborhood. The direction map (Fig.7) which Large Databases, we observe :
is a low-resolution image of local directions, provides the - a large diversity of fingerprints within the same class.
global shape information. Its calculation consists of Several prototypes per class are required to capture the
dividing the image of local directions into blocks of 16x16 intra-class diversity : we use a SOM having 10x10
pixels, and computing the mean direction in each block. prototypes.
The NIST image size is 512x512 pixels, so the dimension - the classes are not clearly separated because some
of the feature space is 32x32. fingerprints belong to two classes. This is the reason why
 we put a map in the feature space : we want to capture bidimensional approximation of a direction-map
the continuity of patterns between two classes. distribution in 32x32 feature space. It takes into account
the large intra-class diversity and the continuum of
• Results :
fingerprint pattern types. We obtained 88% correct
We ran the algorithm on 1600 fingerprints from the NIST classification on a NIST database composed of 1600
database. Fig.8 presents a manual classification of the prints. This result is encouraging but we have to reach
topologic map prototypes after stabilisation of the 98% of correct classification to be integrated in the
algorithm. The unclassified prototypes (X) are due to the THALES Identification products. In the short term, we
presence of poor quality prints in the learning database. will investigate a first space dimension reduction using a
Karhunen Loeve Transform to place the SOM in the
R R R R R R A A L L reduced feature space. Moreover, for a fast and efficient
R R R R R R R A L L pruning, we could combine the previous classification with
R R R R R R W W W L orientation and frequency histograms [10] and the Galton
R R R R R W W L L L number. The Galton number counts the number of ridges
R R R R R W W L L L that crosses a line between the core and delta. It is a very
R R R R R R R L L L discriminating criteria but it can not be used when the
A A R R L L L L W W delta is out of the image. A suitable combinaison of all this
X A A A L L L W W W information will provide an efficient index that constitute a
X A L L L L W W W W major part of a real-time Automated Fingerprint
A A A L L L W W W W Identification System for fingerprint image retrieval in
Fig. 8 : Manual Classification of Prototypes
Very Large DataBases.
We observe on Fig.8 that :
- classes are grouped in the map 5. REFERENCES
- the clustering process has effectively organised the
[1]- The Science of Fingerprints
groups into a continuum of fingerprint pattern types. For
US Department of Justice – FBI.
example, on the first line of Fig.8, the class of each
prototype changes slowly from Right loop to Left Loop [2]- Byoung-Ho Cho & Jeung-Seop Kim & Jae-Hyung Bae &
In-Gu Bae & Kee-Young Yoo, "Core-based Fingerprint
with an intermediate Arch class.
Image Classification", ICPR 2000, vol. 2, pp. 863-866.
- the distribution of the 4 groups in the map correspond
[3]- Meltemp Ballan & F. Ayhan Sakarya & Brian L.Evans, "A
approximately to the distribution observed in reality
Fingerprint Classification Technique Using Directional
(Fig.3, L:32% R:32% W:32% A:4%). Images".
So a bidimensional SOM having 100 prototypes provides [4]- K. Karu & A. Jain, "Fingerprint Classification", Pattern
a good approximation of a fingerprint direction-map Recognition, vol. 29 n°3, pp. 389-404, 1996.
distribution in a 32x32 feature space. [5]- A. Jain & S. Prabhakar & L. Hong, "A Multichannel
Approach to Fingerprint Classification", Michigan State
For automatic classification, we applied the following rule University internal report.
according to eqn.(1) : the attributed class of a given
[6]- G.T Candela & P.J Grother & C.I Watson & R.A
fingerprint is the class of its nearest prototype in the map.
Wilkinson & C.L Wilson, "PCASYS - A Pattern
On a 1600 NIST fingerprint database, 88% of fingerprints Classification Automation System for Fingerprints", NIST
that originate from the same finger are classified in the Internal Report n° 5647.
same class. [7]- Zheng Jason Geng & Weicheng Shen, "Fingerprint
This results could be rapidly enhanced by : Classification Using Fuzzy Cerebellar Model Arithmetic
Computer Neural Networks", Journal of Electronic
- eliminating poor quality fingerprint present in the Imaging, vol. 6(3), pp. 311-318, 1997.
learning database. Indeed, the feature extraction process
[8]- Teuvo Kohonen, "Self-Organizing Maps", Springer,
[11] provides a low quality score in case of poor quality Second Edition, 1997.
prints.
[9]- F. Blayo & M. Verleysen, "Les réseaux de Neurones
- giving a more important weight to directions that are Artificiels", Que Sais-Je, PUF.
situated around the core point, because this region is
[10]-S. Bernard & N. Boujemaa & D. Vitale & C. Bricot,
highly class informative.
"AFIS as Domain Specific CBIR System", IEEE
Workshop on Signal Processing Applications, Brisbane,
4. CONCLUSION Australia, 2000.
Self Organizing Maps are efficient for fingerprint
classification. Indeed, it provides a good non-linear