Contents

1
Chapter 1

More Than Two Thousand

Years Of Fingerprinting

However the very first research paper on fingerprint identification was

published in 1684, the first record of fingerprint on official documents is in-
deed dated [?] as far back as the third century B.C., by the Chinese who were
using it to sign land sales, contracts, loans or acknowledgement of debts1 .

After Dr. Nehemiah Grew published his first paper in 1684, he was closely
followed in 1686 by the publication of Professor Marcello Malpighi (De Ex-
temo Tactus Organo) [?]. However, this is not before 1798 that the first
theory on fingerprint uniqueness and permanence was given by J.C. Mayer.
Later, in the years 1896 to 1897, Sir Edward Henry, Inspector General of the
Bengal Police in India, developed the Henry Classification System [?], [?],
which was the first fingerprint matching system to find a worldwide accep-
tance because its classification achieved a significant increase of performances
by reducing the matching time from days or weeks to only few hours.

Even though fingerprint is nowadays considered the most reliable iden-

tification method, manual verification has indeed become obsolete because
of its obvious tediousness and cost. Therefore, using an automatic finger-
print identification system (AFIS) is the only chance to cope with today’s
performance requirements for both forensic and civilian applications such as
criminal identification, access control or identity verification.

1
Although historical records also relate the use of fingerprints to establish identity in
courts, it is not agreed by historians and researchers wheter or not the Chinese were aware
of the uniqueness of fingerprints

2
Chapter 2

Basis about the AFIS process

The litterature describes that two given fingerprints come from the same
person if there is a topological match [?] between their features patterns.
However, matching two fingerprints is not an easy task. Issues such ro-
tation, translation, deformation and spurious or missing minutiae are to
be addresses with the use of techniques from image processing and pattern
recognition methods.

To introduce to the typical steps carried out within the approaches de-
scribed in the review, the next two sections will be describing a simple ap-
proach of features extraction and matching, as described in [?].

2.1 Minutiae extraction

However loops, archs and whorls are not being considered for this extrac-
tion, it will be useful for the later methods to know that they are described as
global features while minutiae are local features. Out of the eighteen different
types of local ridges identified by [?], the most noticeable are the ridge end-
ings and ridge bifurcations (see figure ??), usually referred to as minutiae.
Before being matched, the minutiae have to be extracted from a fingerprint
image. [?] considers three steps for that purpose :

• Estimation of orientation field

• Ridge detection

• Minutiae detection

The orientation of the flow-like ridges can be done using the Rao’s [?]
algorithm. The image is then enhanced using image processing techniques.

3
 Figure 2.1: Ridge endings and bifurcations

Regarding the detection of ridges, [?] gives the following property: “gray
level on ridges attain their local maxima along the normal directions of the
local ridges”. This property can be applied with the use of convolution masks
that take the (previously estimated) orientation in consideration. Still after
this image enhancement step, spurious minutiae are left due to noise, breaks
and smudges in the fingerprint image. Heuristics and algorithms can then
be used to delete spurious minutiae and reconnect broken ridges, thus finally
producing a thinned ridge map where minutiae shall be easy to detect.
A technique to then locate the minutiae is to assume that a black pixel is a
ridge ending if there is exactly 1 single black pixel among his 8 neighbooring
pixels while a ridge bifurcation pixel would require strictly more than 2.

The different steps of the minutiae detection are illustrated in figure ??.
Even though the given techniques used during the following steps of the
extraction are not the most efficient, they are the basis of most of the later
methods.

2.2 Minutiae matching

Once the minutiae have been extracted from the fingerprint image, they
can be represented as points. Thus the verification process becomes a point
pattern matching. Ideally, the points of the template pattern can be aligned
completely with those of the input pattern. However, due to noise, defor-
mation, rotation and translation, a perfect alignment of the points is not
realistic. Therefore minutiae matching methods must be tolerant (to some
extent) and also cope with issues such as rotation and translation.

4
Figure 2.2: Extraction of minutiae

5
Chapter 3

Fingerprint matching methods

As explained in the previous section, a basic fingerprint matching can be

described as a point pattern matching. Such patterns can be matched using
genetic algorithms [?] and other methods of pattern recognition [?], [?]. Such
methods are slow, due to their iterative nature. Numerous research papers
propose different approach to fingerprint verification. Indeed, gathering more
information about the minutiae such as their type (bifurcation or ending),
direction and correlation but also considering the ridges and global features
and, finally, combining different methods allow to bring more discrimination
and get far better accuracy. The next sections will introduce some of these
methods and describe their working, advantages and shortcomings.

3.1 Topology-based matching

An important key to fingerprint matching is the scheme being used to
classify the extracted features. The Henry classification system [?], which
was the first effective system, is based on the global features of the ridges
and has been used during many years [?] as a mean to partition fingerprint
databases. Considering the local features of the fingerprint shall substan-
tially improve the classification. However, the fact that distortions of the
pattern can not be avoided is a critical issue for such a classification.

With this regard, [?] proposes an topology-based classification where the

structure of the fingerprint is represented by a geometrical configuration of
its minutiae, thus reflecting the general spatial structure in the neighbor-
hood of each minutiae. After the minutiae have been extracted, a specific
neighborhood, called the minutiae window, is determined for each of them.
In each minutiae window, the following information are collected:

6
 • Total number of surrounding minutiae

• Number of ridges between the central minutiae of the window and the
surrounding ones

• Orientation of the surrounding minutiae

• Angle between the direction of the central minutiae ridge and the ridge
of surrounding minutiae

• Distance between the central minutia and the surrounding ones

Afterwards, the minutiae windows are ordered according to the number of

minutiae they contain. At this point we could argue that this ordering scheme
does not take enough characteristics of the minutiae window for ordering
them (two windows may have the same amount of surrounding minutiae). We
could also mention that spurious and missing minutiae (due to already cited
factors) will further degrade the order. However, it appears that this is not a
problem at all when using a tree matching. Furthermore, if noise issues can
not be avoided, it must be noticed that the features located within a minutiae
window do not suffer of rotations or translations. The tree matching can be
described as follows:

• On a one-to-one basis, line up the surrounding minutiae in the pair of

initial window.

• Starting at the root, generate a tree via a breadth-first search scheme.

• Continue the tree expansion if similarly constructed counterpart can be

found in the template (classified as describe before), until all minutiae
have been used.

• If no counterpart is found, re-iterate until considering a non-match if

non sufficiently large tree can be generated.

The scheme relies on the assumption that surrounding features in a close

neighborhood are supposely invariant to rotation and translation. However
it appears that minutiae associated to global features such as whorls or lo-
cated in noisy regions are quite troublesome. An advantage of this matching
method might be his classification scheme. Since features are partitioned in
the database, it should increase the efficiency of one-to-many queries in a
database. However, the tolerance required to cope with spurious or missing
minutiae would substantially affect the efficiency of queries.

7
3.2 Orientation-based Minutia Descriptor
Keeping in mind that considering only the minutiae location and direc-
tion does only exploit a very low amount of the rich information content
that is present in fingerprint patterns, [?] introduces a minutiae descriptor
that relies on describing the orientation information of the fingerprint pat-
tern with regards to each minutiae details. Fingerprint are here classified
as “weakly-order textures exhibiting a dominant ridge orientation at each
point”. To compare with the method in [?], the description around each
minutiae is orientation-based, rather than minutiae-based. In other words,
not the surrounding minutiae, but the ridge orientation of each of them are
considered. Their orientation are even invarient to rotation and translation
with noisy inputs, thus bringing a substantial advantage over the minutiae-
based method. The minutiae descriptor is further said to be “independant
with respect to any other minutiae.

The matching is then done by performing the following steps:

• Detect the minutiae of the input pattern and pair as many of them
with those of the template.

• For each corresponding pair of minutiae, compute their match (ridges

direction distance) to end up with a total matching score normalised
with the number of overlapping minutiae.

• The match is then determined by evaluating the score against an em-

pirical treshold.

The experimental results report that the rejection rates increases with factors
such as errors in minutiae detection, unreliable orientation. Poor quality
images are also an issue because they reduce the number overlapping (because
not detected) minutiae to be compared.

3.3 Alignment-based matching

As it has already been explained, due to deformations and consequences of
noises, it is not realistic to consider that the minutiae of overlapping patterns
will perfectly match. In [?], alignment-based recognition [?] is used in the
matching algorithm because of its simplicity, efficiency of discrimination and
speed. The algorithm can be divided into two main steps:

8
 • Alignment: transformations due to rotation, translation and scaling are
estimated and then used to realign the minutiae of the input pattern
with the template (see figure ??)

• Matching: conversion of the input and template minutiae into polygons

in the polar coordinate system and use of an elastic matching algorithm
to match the polygons

Figure ?? illustrates the alignment during the matching process. In the first
step the input pattern is aligned with the template by rotation and transla-
tion. The elastic point pattern algorithm concatenates each of the minutiae
(then represented as polygons in the polar coordinate system) into two strings
in the increasing order of radial angles. The distance between the two strings
is calculated and used to compute the match score.

However non-linear transformations might have remained after the first

step, the elastic algorithm in polar coordinates has the great advantage to
achieve a certain amount of tolerance, thus coping with that issue. On the
other side, if this scheme is tolerant to inexact localizations and non-linear
modifications, it does not indeed compensate (which would be more clever).
Also, string matching algorithms are not the most efficient in term of speed
and this might be an issue when using the scheme for identification on large
databases.

Figure 3.1: Alignment of an input ridge and a template ridge

9
Figure 3.2: (a) Input minutiae set, (b) template minutiae set, (c) alignment,
(d) pairing of the minutiae

3.4 Bank of Gabor filtering

[?] proposes a method capturing both the local and global features into a
so-called FingerCode, using a bank of Gabor1 filters. The matching is then
reduced to finding the euclidean distance between these FingerCodes and
hence it shall be very fast. The method performs the following steps:

• Determine a reference frame for the image and tesselate this region into
8 sectors

• Using the bank of Gabor filters, filter the image in 8 different directions

• Compute the grayscale variance (standard deviation) in each sector and

thus define the feature vector of the FingerCode

The verification is achieved by measuring the euclidean distance of the query

fingerprint and of the one of the template. The flow diagram of the method
for fingerprint verification is illustrated in figure ??.

The first obvious advantage of this method is that the rotation invariance
can be very easily achieved. Thanks to its circular shape, the tesselated re-
gion (see figure ??(a)) can be simply rotated like a disc. The FingerCodes
are 640-dimensional (8 filtered images tesselated into 80 cells). The matching
1
Gabor filterbank is an efficient technique to perform the capture and decomposition
of useful information in a particular bandpass

10
is very fast because it is done by measuring their euclidean distance. Due to
the nature of the data and because the calculation of the euclidean distance
involves bit comparison, the matching scheme is suitable for hardware im-
plementation. Furthermore, because of their small size, the templates could
easily be embeded in the chip of a smart card.
This methods also brings its collection of shortcomings. The first one is

Figure 3.3: Tesselations

the determination of the reference frame. The detection of the core point
is not a trivial task. The region of interest may also be partially omited
during acquisition or it could even happen that the core point (center of the
reference frame) is not even present in the acquired image. Occlusion and
obliteration have the same consequences. It is true that the circular tessela-
tion offers the great advantage to achieve very easily the rotation invariance,
but on the other side it does not cover the whole image, thus reducing dis-
crimination and therefore accuracy. This gets even worse if the core point
is determined to be close to the boundary, because we obtain not only a
partial image but partial image which is partially void. In this case, the
features in the Fingercode can not very discriminative enough. The experi-
ments also show that about 99% of the process is spent on the Gabor filtering

[?] offers a hybrid matching method, developed by members of [?] and

aiming at addressing these issues. The framework remains basically the same.
The changes can be resumed by the following points:

• The fingerprint image is enhanced prior to filtering, thus offering more

accurate information on the ridges.

• The tesselation is a square (see figure ??(b)) that does not consider a
landmark point, thus covering the entire image.

• The minutiae are extracted with the method of [?] and are indeed used
to align the squared tesselation and add more discrimination.

11
 • The Gabor filter computation is achieved in the frequency space.

Figure 3.4: Flow diagram of the FingerCode verification system

Covering the entire image rather than only a limited region further to
align the tesselation with the extracted minutiae instead od detecting an
empirical core point, is far more accurate. The dimension of the feature
vector is substantially increased (1352 dimensions for a 13x13 tesselation
with 8 Gabor-filtered images). The matching code of the FingerCodes (their
euclidean distance) and of the minutiae-based are merged into a single match-
ing score. In the previous implementation, the Gabor filtering was the most
CPU-expensive step of the process is now fairly reduced by the new com-
putation scheme. Experiments claim better results with this hybrid match-
ing method. However, the method could be improved by first performing a

12
minutiae-based match and only continue the verification is the match score
is below a given treshold, thus reducing the processing time.

13
Chapter 4

Conclusion

We have been through several research papers, presented the working

of their approached methods and critically discussed their advantages and
shortcomings. Regarding the methods efficiency, it must be noticed that
false rejection and false acceptance rates given in the experimental results
section have not been taken in consideration as a comparison because they
are not to be considered relevant since experiments have not been performed
on the same training and test sets.

Several other approaches to fingerprint matching have been proposed in

the litterature such as optical correlation [?], [?] or graph matching [?]. It
appears very clearly that a lot of research is being done to perform better
fingerprint verification. However, there is a popular misconception that fin-
gerprint recognition system is a fully solved problem since it has been around
for a quite long period of time. In fact, fingerprint recognition remains a
very challenging and important pattern recognition problem. Finally, with
the increasing needs of identity verification for security purposes in today’s
world, the number of publications on fingerprint-related topics is obviously
not expected to decrease soon.

14
Bibliography

[NSW Police] The Thin Blue Line, The History of Fingerprinting. New
South Wales Australia, Unofficial Police Website. Retrieved
March 23, 2004, from http://www.policensw.com/info/
fingerprints/finger01.html

[Moore] G. Moore, Brief History of Fingerprint Identification. Fin-

gerprint Consulting Service. Retrieved March 23, 2004, from
http://www.gregandsheila.com/consult/xx4.htm

[LPE, 2003] Latent Print Examination, The History of Fingerprints. Novem-

ber, 2003. Retrieved March 23, 2004, from http://onin.com/
fp/fphistory.html

[IBG, 2003] International Biometric Group, The Henry Classification

System. New York, 2003. Retrieved March 23, 2004, from
http://www.biometricgroup.com/Henry\%20Fingerprint\
%20Classification.pdf

[Henry, 1900] E.R. Henry, “Classification and uses of finger prints”, Rout-
ledge, London, 1900

[FBI, 1984] Federal Bureau of Investigation, The Science of Fingerprints:

Classification and Uses. Washington, D.C.: US Government
Printing Office, 1984.

[Lee, 1991] H.C. Lee and R.E. Gaensslen, eds., Advances in Fingerprint
Technology. New York: Elsevier, 1991.

[Jain, 1997] A. Jain, L. Hong and R. Bolle. “On-line fingerprint verification”.

IEEE Trans. Pattern Analysis and Machine Intell., 19(4):302-
3013, Apr. 1997.

[Rao, 1990] A.R. Rao, A taxonomy for Texture Description and Identifica-
tion. New Yor: Springer-Verlag, 1990

15
[Mallesware, 1976] T.C. Mallesware, “Feature Extraction for Fingerprinting
Classification”, Pattern Recognition, Vol 8, 1976, pp 181-192
[Ansari, 1992] N. Ansari, M.H. Chen, and E.S.H. Hou, “A Genetic Algo-
rithm for Point Pattern Matching”, Chapt. 13, B. Soucek and
the IRIS Group, eds., Dynamic, Genetic, and Chaotic Program-
ming. New York: John Wiley & Sons, 1992
[Ranade, 1993] A. Ranade and A. Rosenfeld, “Point Pattern Matching by
Relaxation”, Pattern Recognition, vol. 12, no. 2, pp. 269-275,
1993
[Starink, 1995] J.P.P Starink and E. Backer, “Finding Point Correspondance
Using Simulated Annealing”, Pattern Recognition, vol 28, no. 2,
pp. 231-240, 1995
[Tico, 2003] M. Tico, P. Kuosmanen, “Fingerprint Matching Using an
Orientation-Based Minutiae Descriptor”, Trans. on Pattern
Analysis and Machine Intelligence, Vol 25, pp. 1009-1014
[Isenor, 1986] D. K. Isenor, S. G. Zaky, “Fingerprint Identification Using
Graph Matching”, Pattern Recognition, Vol. 19, pp. 111-112,
1986
[Chen, 1991] Z. Chen, C. H. Kuo, “A Topology-Based Matching Algorithm
for Fingerprint Authentication”, Proc. of 25 Annual IEEE In-
ternational Carnahan Conference on Security Technology, pp.
84-87, 1991
[Huttenlocher, 1987] D.P. Huttenlocher und S. Ullman, “Object Recognition
Using Alignment”, Proc. First Int’l Conf. Computer Vision, pp.
102-111, London, 1987
[Jain, 2000] A.K. Jain, S. Prabhakar, L. Hong and S. Pankanti. Filter-bank
based Fingerprint Matching. IEEE Transactions on Image Pro-
cessing, 9(5):846-859, May 2000
[Ross, 2002] A. Ross, A. Jain, J. Reisman. A Hybrid Fingerprint Matcher.
Proc. of International Conference on Pattern Recognition, Que-
bec City, 2002
[Wilson, 2000] C. Wilson, C. Watson and E. Paek, “Effect of Resolution
and Image Quality on Combined Optical and Neural Network
Fingerpritn Matching”, Pattern Recognition, Vol 33, no. 2, pp.
317-331, 2000

16
[Gamble, 1992] F. Gamble, L. Frye and D. Grieser, “Real Time Fingerprint
Verification System”, Applied Optics, Vol 31., no. 5, pp. 652-655,
1992

[Isenor, 1986] D.K. Isenor and S.G. Zaky, “Fingerprint Identification Using
Graph Matching”, Patter Recognition, Vol. 19, no. 2, pp. 113-
122, 1986

17