IEICE Electronics Express, Vol.1, No.

18, 575–581

A new approach to reference

point location in fingerprint
recognition
Piotr Porwika) and Lukasz Wieclawb)
Institute of Informatics, Silesian University
41–200 Sosnowiec ul. Bedzinska 39, Poland
a) [email protected]
b) [email protected]

Abstract: Fingerprint matching is one of the most important prob-

lems in Fingerprint Identification System (AFIS). In this paper a new
method of the reference point alignment has been presented. A new
approach of reference point localization is based on so-called identifi-
cation masks which have been composed on the basis of analysis of
biometric characteristic of human finger. Construction of such masks
has been presented.
Experiments show that our approach locates a unique reference
point with high accuracy for all types of fingerprints. Generally, fin-
gerprint matching consists with three steps: core (reference) point de-
tection, filter the image using a bank Gabor filters, and comparison
with imprint pattern. It seems, that today, the Gabor filtering gives
the best results in fingerprint recognition. The proposed method was
evaluated and tested on various fingerprint images, included in the
FVC2000 fingerprint database. Performed results with representative
investigations have been compared.
Keywords: Fingerprint recognition, reference point determination
Classification: Science and engineering for electronics

References

[1] L. Hong, Y. Wan, and A. K. Jain, “Fingerprint image Enhancement: Algo-

rithm and Performance Evaluation,” IEEE Trans. Pattern Anal. Machine
Intell., vol. 20, no. 8, pp. 777–789, Aug. 1998.
[2] A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti, “Filterbank-Based
Fingerprint Matching,” IEEE Trans. Image Processing, vol. 9, no. 5, May
2000.
[3] A. K. Jain, S. Prabhakar, and L. Hong, “A multichannel approach to fin-
gerprint classification,” IEEE Trans. Pattern Anal. Machine Intell., vol. 9,
no. 4, pp. 348–359, 1999.
[4] X. Jiang, M. Liu, and A. C. Kot, “Reference Point Detection for Finger-
print Recognition,” Proc. of the 17th Int. Conf. on Pattern Recognition
(ICPR’04), Cambridge, UK, pp. 540–543, Aug. 2004.
[5] K. Karu and K. Jain, “Fingerprint Classification,” Pattern Recogn.,


c IEICE 2004 vol. 29, no. 3, pp. 389–404, 1996.
DOI: 10.1587/elex.1.575
Received November 16, 2004
Accepted December 09, 2004
Published December 25, 2004

575
 IEICE Electronics Express, Vol.1, No.18, 575–581

[6] C. H. Park, S. K. Oh, D. M. Kwak, B. S. Kim, Y. C. Song, and K. H.

Park, “A New Reference Point Detection Algorithm Based on Orientation
Pattern Labeling in Fingerprint Images,” Proc. of 1st Iberian Conf. on
Pattern Recognition and Image Analysis, Puerto de Andratx, Spain, June
2003.
[7] http://www.mathworks.it/matlabcentral/fileexchange/
loadFile.do?f4addg fingerprint&objectId=4239&objectType=file

1 Introduction
The fingerprints are extremely complex. For fingerprint identification, the
most technically critical is detection of a reliable reference point. To locate
a unique reference point for all types of fingerprints, we define the reference
point as the point with maximum curvature on the convex ridge, which is
usually located in the central fingerprint area [5]. In general, a analyzed
fingerprint is a well-defined orientation field, which has to compared with
pattern. In most cases the reference point detection methods operate on the
orientation field of fingerprint. The Poincare index analysis is well known
method described in [5]. Unfortunately, this method is sensitive to noise
of orientation field. Very efficient method proposed in [2] is based on mul-
tiresolution analysis, but this method is sensitive to the fingerprint rotation.
Additionally, a fingerprint may be corrupted by noise due to finger pressure
variation, skin properties etc. during image capture. Fingerprint enhance-
ment is used to recover the topology structure of ridges and valleys from the
noisy image. Most of the fingerprint enhancement algorithms based on the
estimation of the orientation field [1]. In presented method, the orientation
field, with some additional modifications, to reference point indication has
been exploited. In our approach, base on orientation field, the special masks
are proposed, which allow to detect a unique reference point consistently and
accurately for all types of prints.

2 Direction image determination

The analyzed M × N image (fingerprint) is binarized firstly and the bina-
rization threshold has been fixed on the level:


N

M
v(pxy )
y=1 x=1
T = − 30 (1)
M ·N
where:
v(pxy ) – the value of a pixel at point pxy ,
N, M – vertical and horizontal dimension of fingerprint image, respec-
tively.

The constant 30 which occurs in equation (1) was fixed experimentally,

c IEICE 2004
DOI: 10.1587/elex.1.575 and allows us to get the better binarization effect for scanned fingerprint im-
Received November 16, 2004
Accepted December 09, 2004
Published December 25, 2004

576
 IEICE Electronics Express, Vol.1, No.18, 575–581

ages with dark background. In presented implementation, fingerprint images

have had dimension 512 × 512 and with density 500 dpi have been scanned.
To decide the ridge direction of each pixel in the image, the ridge direction
of a given raw pixel pxy into directions i (i = 0, 1, . . . , 7) has been divided
in a 9 × 9 window with the pixel in the center [3] (Fig. 1 b). The values
of pixels in eight directions (at positions marked by numbers 0,1,. . . ,7) are
added together to obtain sums s0 , s1 , . . . , s7 . The sums si are equivalent to
convolving the image with 9 × 9 masks mi , where each mask has value of 1
at positions where block shown in Fig. 1 b has values i, and 0 elsewhere.
In next stage, for indices i min/max values are computed:
⎧
⎪
⎪ {max(si )} for pbin
⎨ sd = arg xy = 1
i={0,1,..,7}
i= (2)
⎪
⎪ {min(si )} for pbin
⎩ sl = arg xy = 0
i={0,1,..,7}

where: pbin
xy – is a pixel value in binarized fingerprint.
The direction at pixel is defined by means of sd value if the central pixel
is located on a ridge (black area), and by the sl value if the central pixel
is located in a valley (white area). From equation (2), indices image can
be prepared. Such image can also be called direction image. Unfortunately,
obtained values treated as direction for each pixel are usually noisy, therefore
they should be smoothed and averaged in a local neighborhood. In our
application, as smoothing operation the mode function has been applied.
The function mode computes the mode of the given data and the mode is
defined as observation with the highest frequency. In other words, mode
function can be treated as measure of central tendency. For example d =
mode(a, b, c, d, d, d, e, f, g). If there is more than one observation with the
modal frequency, then choice is arbitrary. Results of mentioned operations
are shown on Fig. 1.

Fig. 1. a) a given fingerprint, b) the 9×9 mask to compute

the ridge directions, c) 3 × 3 mode filter mask, d)


c IEICE 2004 directions image.
DOI: 10.1587/elex.1.575
Received November 16, 2004
Accepted December 09, 2004
Published December 25, 2004

577
 IEICE Electronics Express, Vol.1, No.18, 575–581

For a given fingerprint (Fig. 1 a), directions image has been computed
(Fig. 1 d). Such image is additionally filtered by means of 3 × 3 averaging
filter mask, which works like the mode function (Fig. 1 c).

3 A new method of reference point determination

From directions image, the orientation field has been prepared [2, 4]. The
orientation represents the local orientation of the ridges contained in the fin-
gerprint. Directions image from Fig. 1 d into 9 × 9 blocks (windows) has been
divided. In each block, dominant direction (the most frequently occurring
direction) like previously was calculated. From the gradient information the
orientation angle is estimated. The 8 directions (see Fig. 1 b) in each block
is stored as table. This table can be presented as image (although it is not
necessary), where each direction is performed as appropriate line. For ex-
ample, for angle 00 horizontal line is performed, for angle 900 − vertical, and
so one by step 22, 50 for another direction lines. In next stage background
is eliminated and only foreground is used. If in a given 9 × 9 block at least
one black pixel can be found, then whole block create part of orientation
field, otherwise such block is treated as background and is rejected. The
orientation field of fingerprint shown Fig. 2 a.

Fig. 2. a) the appropriate orientation fields obtained from

directional image (see Fig. 1 d), b) features of
identification mask, c,d) the identification masks
for 00 and +450 , respectively

After background elimination the orientation field is smoothing with the

aid of mode mask, similar as previously.
In the last stage the reference point is determined. Because print of


c IEICE 2004
DOI: 10.1587/elex.1.575 finger can be impressed in different manner, influence of rotation should
Received November 16, 2004
Accepted December 09, 2004
Published December 25, 2004

578
 IEICE Electronics Express, Vol.1, No.18, 575–581

be eliminated. In our method orientation field image is filtered by means

of appropriates masks. That masks have special construction, what allows
to collect fingerprints located in 00 , ±450 , ±900 . Proposed masks are called
identification masks. Fig. 2 a presents identification features of each mask. A
special arrangement of short lines in each cell characterizes ridges in a human
fingerprint. Such ridges distribution has been established experimentally.
Such approach is very efficient and it is equivalent to minutiae detection
method [5], but unlike another methods [1, 4, 6] our solution gives faster
reference point detection. By means of identification masks, which explore
whole orientation field, fingerprint features represent by appropriate mask are
sought. The masks for angles 00 and +450 presents Fig. 2 c. Because it is not
known, how fingerprint was impressed, orientation field is filtered in turn by
all masks. Finally, each identification mask more than one reference point can
indicate. All identification masks into cells are split. The cell S(x, y) indicate
direction at point x, y in orientation field. The 8 directions in orientation
field are represented by means of numbers 0, 22.5, 45, 67.5,. . . ,157.5 (see
Fig. 1 b).
From orientation field (Fig. 2 a) the reference point for the angle 00 can
be determined if the global condition is fulfilled:

If (−22, 5◦ ≥ S(x, y) ≥ 22, 5◦ ) and not (−22, 5◦ ≥ S1a ≥ 22, 5◦ )

and not (−22, 5◦ ≥ S1b ≥ 22, 5◦ ) and (−22, 5◦ ≥ S2a ≥ 22, 5◦ )
and (−22, 5◦ ≥ S2b ≥ 22, 5◦ ) and (0◦ ≥ S3a ≥ 90◦ )
and (90◦ ≥ S3b ≥ 180◦ ) and (22, 5◦ ≥ S4a ≥ 90◦ )
and (90◦ ≥ S4b ≥ 157, 5◦ ) and (0◦ ≥ S5a ≥ 90◦ )
and (90◦ ≥ S5b ≥ 180◦ ); (3)
Similarly, for angle +450 , the reference point can be determined if the global
condition of orientation field can be fulfilled:
If (22, 5◦ ≥ S(x, y) ≥ 67, 5◦ ) and (90◦ ≥ S1a ≥ 180◦ )
and (90◦ ≥ S1b ≥ 180◦ ) and (0◦ ≥ S2 ≥ 90◦ )
and (45◦ ≥ S3a ≥ 157, 5◦ ) and (112, 5◦ ≥ S3b ≥ 225◦ )
and (45◦ ≥ S4a ≥ 157, 5◦ ) and (112, 5◦ ≥ S4b ≥ 225◦ )
and (45◦ ≥ S5a ≥ 157, 5◦ ) and (112, 5◦ ≥ S5b ≥ 225◦ )
and (0◦ ≥ S6a ≥ 90◦ ) and (0◦ ≥ S6b ≥ 90◦ ) ; (4)

For remained masks, performed conditions (4) and (5) should be modified by
appropriate rotation of the masks. In the worst case, each identification mask
can point different reference points, but it is well known that for fingerprint
only one reference point can be indicated. From this reason, for all potential
reference points, detected by means of identification masks (Fig. 2 c), so-
called influence coefficients have been estimated. The influence coefficients
(inco) is calculated as follow:


c IEICE 2004 For angle 00 :
DOI: 10.1587/elex.1.575
Received November 16, 2004 If (S(x, y) = S2a) then inco := inco + 1;
Accepted December 09, 2004
Published December 25, 2004

579
 IEICE Electronics Express, Vol.1, No.18, 575–581

If (S5a <> 90◦ ) or (S5b <> 90◦ ) then inco := inco + 1;

If (|S1a−S(x, y)| = 90◦ ) or (|S1b−S(x, y)| = 90◦ ) then inco := inco+1;

For angle +450 :

If (S(x, y) = S2) then inco := inco + 1;

If (S5a ≤ 22, 5◦ ) or ((S5b ≥ 90◦ ) and (S5b ≤ 135◦ )) then inco := inco +1;
If (|S1a−S(x, y)| = 90◦ ) or (|S1b−S(x, y)| = 90◦ ) then inco := inco +1;

Remained the influence coefficients are calculated similarly, and only ap-
propriate values of angles should be changed: for −450 , ±900 , respectively.
Additionally, the next principles are considered:
– if two (or more) references points is located in a local neighborhood, then
value of the inco coefficient is increased of 4,
– for the lowest located reference point, its inco value is increase of 5.
– potential reference point which lies at a distance less than 8 pixels from
edge background is rejected.
Finally, the point which has the largest inco coefficient, will be classified
as reference point.
Proposed detection of reference point together with the 2D Gabor filtering
method [2] can be used to fingerprint matching.

4 Conclusions
In this paper, a new method to locate a unique reference point has been
presented.
Since human experts may not be able to locate the pixel wise accurate
reference point, we propose the new method which allow to determine such
point. The our method base on so-called identification masks, which was
designed on the basis of human finger print analysis. For proposed masks,
the influence coefficients have been stated. Proposed method with complete
algorithm described in [2, 7] has been compared. Complete results could
not be included as over 100 fingerprints from database were tested. The
described method with the aid of the FVC2000 fingerprint database has been
tested. Mentioned data collection includes varying quality fingerprint images.

Fig. 3. a) The perfect reference point location, b,c) refer-

c IEICE 2004 ence point determination for different fingerprints
DOI: 10.1587/elex.1.575 quality from the FVC2000 database.
Received November 16, 2004
Accepted December 09, 2004
Published December 25, 2004

580
 IEICE Electronics Express, Vol.1, No.18, 575–581

Taking into account paper length restrictions, only main differences between
algorithms have been stated. Fig. 3 presents obtained results of reference
point location. In experiments the reference points by signs ‘+’ or ‘x’ have
been marked. Our reference point location by means of ‘x’ sign is indicated.
In any fingerprint can be shown points on the basis of which reference
point can be detected. Such objects can be delta points, the Galton line or
the set of minutiaes. In our approach, the perfect reference point should be
placed at the point with maximum curvature on the convex of fingerprint
ridge (Fig. 3 a) [4, 5, 6].
From Fig. 3 follows, that reference points are located very close each
other, though clearly at different places. From conducted investigations dif-
ferences between reference points location are always not greater that 9 pix-
els, where difference is determined between middles of appropriate markers.
From observation follows, that our reference point location in most cases
more precisely corresponds to perfect reference point definition.

c IEICE 2004
DOI: 10.1587/elex.1.575
Received November 16, 2004
Accepted December 09, 2004
Published December 25, 2004

581