On Fingerprint Template Synthesis

Wei-Yun Yau, Kar-Ann Toh, Xudong Jiang, Tai-Pang Chen and Juwei Lu
Centre for Signal Processing
School of Electrical & Electronic Engineering
Nanyang Technological University

Abstract contact with the sensor. This, as compared to conven-

tional ink based rolled fingerprint impression, possesses
In this paper, we address the false rejection problem due a much smaller information area. A consequence of this
to small solid state sensor area available for fingerprint can be seen in using different partial areas of the same
image capture. We propose a minutiae data synthesis finger for matching, which causes false rejection. For this
approach to circumvent this problem. Main advantages reason, during enrollment of a person in a database, a
of this approach over existing image mosaicing approach rolled fingerprint would be preferred over a plain touch
include low memory storage requirement and low com- impression.
putational complexity. Moreover, possible overhead on Apart from requiring the individual user to ensure
the search engine (for fingerprint matching) due to data good placement of fingerprint area during image ac-
redundancy could be reduced. Extensive experiments quisition process, few automatic fingerprint identifica-
were conducted to determine the best transformation tion/verification system has addressed the problem of
suitable for minutiae alignment. We demonstrate false rejection cause by using different image regions for
the idea of synthesis with an example using physical matching. While acquiring a few separate fingerprint im-
fingerprint images. The proposed synthesis system is also ages during registration could simply handle the false re-
found to improve (lower) the number of false rejects due jection problem, much of the acquired information could
to the use of different fingerprint regions for matching. be redundant (due to much common regions) and hence
takes up unnecessarily large storage space. Moreover,
keywords:- overheads of the multi-modal search would increase due to
Biometrics, image database, fingerprint identification, the much larger number of records available for matching.
fingerprint verification, minutiae points, template match- In [6], an image mosaicing technique is developed for con-
ing. structing a rolled fingerprint from an image sequence of
partial fingerprints. The proposed fingerprint mosaicing
algorithm consists of four stages namely, (i) segmentation
of foreground and background areas in each frame; (ii)
1 Introduction weighting of each image’s contribution using a foreground
mask; (iii) stacking of the weighted gray scale frames to
In general, an automatic fingerprint identification or ver-
compute the mosaic gray scale image; (iv) stacking of
ification (see e.g. [1, 2, 3, 4]) system consists of three
the foreground masks to compute a confidence index at
main processing stages namely, image acquisition, fea-
every pixel. Although mosaicing technique possesses the
ture extraction and matching. In image acquisition, query
capability of acquiring a larger area of fingerprint image,
and template database images are acquired through var-
it is at the expense of large storage requirement for a
ious input devices. Development over the years has seen
much larger synthesized image. Moreover, as seen from
through means that mechanically scan the ink based fin-
the pixel level computation which is applied directly to
gerprints into the computer system, and means which di-
the acquired image, the computational cost is high.
rectly capture the fingerprints using more sophisticated
solid state sensors. With fingerprint images which could In this paper, we propose a minutiae based syn-
be distorted or contaminated with noise, the automated thesis method for an automatic fingerprint identifica-
system seeks to extract characteristic features which are tion/verification system. The proposed methodology not
discriminating and yet invariant with respect to image only synthesize necessary information for fingerprint iden-
orientation. The final stage of fingerprint identification or tification, but also possesses several desired features:
verification is to search and verify matching image pairs. 1. no restriction on the hardware sensor area;
The use of inkless sensors has advanced the data ac-
quisition aspects in an automatic fingerprint identifica- 2. small storage requirement since the synthesized data
tion/verification system. This includes optical and solid contains only the necessary minutiae information
state devices. By means of CCD array and laser technolo- needed for matching. This is especially useful for
gies, the optical sensors offer a cost effective solution for search within a large database in fingerprint identi-
fingerprint image capture. Conforming to the regulation fication;
by National Institute of Standard and Technology (NIST,
3. low computational complexity.
USA), conventional optical sensors have a sensing area of
1-inch by 1-inch. The solid state sensors, which adopt ca- The paper is organized as follows: in section 2 we pro-
pacitance, electric field, pressure or temperature sensing vide a brief overview on fingerprint identification and ver-
technologies, offer a more compact means for fingerprint ification problems related to the subject matter. This is
image capture with additional features to detect presence followed by section 3 where our representation system for
of fingers such as locally adjustable automatic gain con- minutiae data and our methodology for minutiae synthe-
trol [5]. However, due to manufacturing limitation and sis are presented. Section 4 provides an outline on minu-
cost factors, most solid state sensors do not come with tiae detection, minutiae alignment and several transfor-
large sensing area (e.g. Veridicom’s iTouch has a sensing mations available for minutiae alignment. In section 5,
area of 1.5cm by 1.5cm and Infineon’s FingerTip has a we perform experimental study to decide upon the best
sensing area of 11.1mm by 14.3mm). Moreover, the imag- transformation for our application. Then, on top of a
ing area for the finger is further restricted to the area in minutiae synthesis example, we provide an evaluation of
the synthesis system in aspects of false rejection. Some we can write for the kth minutiae data set as:
concluding remarks are drawn in section 6. Mk = {(xi , yi , ϕi , ti )}k , i = 1, 2, ...nk , k = 1, 2, ..., m.
(2)
Among these minutiae data sets, there would be com-
2 Fingerprint Identifica- mon regions whereby information is redundant. If it is
to search through each individual minutiae data sets for
tion/Verification matching, it would not be cost effective since these redun-
dant information are being searched through more than
While fingerprint identification refers to the process of
once. Moreover, the geometrical relationships among
matching a query fingerprint against a template finger-
these minutiae data sets are no longer preserved since
print database to establish the identity of an individual,
these data sets are treated as separate entity. In order
fingerprint verification refers to determination of whether
to save data storage space with respect to redundancy as
two fingerprints are from the same finger or not. Since
well as to provide a good overall picture about the minu-
verification is a comparison of a query fingerprint against
tiae sets, a synthesis with consideration to relationship
an enrolled template fingerprint, it is also termed as one-
between data sets is needed.
to-one matching. Identification, on the other hand, is
For fingerprint images with common regions, we can
termed as one-to-many matching.
express the resultant synthesized information as:
In view of effective search within a huge database in
m
!
an identification problem, fingerprint matching is usually
M∪m = {fk (xi , yi , ϕi , ti )}k , i = 1, 2, ...nk (3)
carried out at two levels namely, the coarse level and the 1
k=1
fine level. The coarse level matching is also referred to as
classification (see e.g. [7, 8, 9, 10]) where fingerprints are where fk , k = 1, 2, ..., m denote the necessary topological
grouped into few classes so that a fine level search can be transformations for aligning the different sets of minutiae
performed within the matched class. data. Suppose o1 is the number of overlapping points
As it is difficult to use the raw digital fingerprint image between M1 and M2 . Then the number of minutiae points
directly for matching in an automatic fingerprint identifi- in M∪2 can be expressed as n(M∪2 ) = n1 + n2 − o1 . Now
1 1
cation/verification system, a suitable computer represen- let o2 be the number of overlapping points between M∪2
1
tation is essential. Several desirable properties for such and M3 . And the number of minutiae points in M∪3 is
1
representation can be identified as: n(M∪3 ) = n1 + n2 + n3 − o1 − o2 . In short, the total
1
1. retention of the discriminating power of each finger- number of minutiae points in the synthesized minutiae
print (information content), data set M∪m can be written as:
1
2. stable with respect to noise and distortion, m
" m−1
"
n(M∪m )= nk − ok , (4)
3. small in data storage size, and 1
k=1 k=1
4. ease in manipulation (e.g. adding new data points). where ok is the number of overlapping minutiae points
Here, we note that large information content may pose a between M∪k and Mk+1 . Hence, if each ok is of consid-
1
constrain to small data storage size requirement. erable size, the total number of minutiae in the synthe-
According to [11], eighteen different types of finger- sized data set (n(M∪m )) can be significantly smaller than
#m 1
print features can be identified. These features include
k=1 nk .
ridge endings, ridge bifurcations, short ridges and ridge Upon acquiring two images for synthesis, the task im-
crossovers which are collectively termed as minutiae. It mediately after minutiae detection is to find correspon-
has been widely accepted that local ridge structures (a dence between these two images so that the global minu-
collection of minutiae details) from two fingerprints match tiae information between the two images can be aligned.
each other if the fingerprints are from the same source A match between these two images is performed utilizing
[11, 12]. Hence, the problem of fingerprint verification both local and global minutiae information. Two minu-
can be reduced to a point pattern matching problem when tiae data sets are considered matched if a weighted score
these local structures are considered. between the local and the global information exceeds a
certain threshold value. The match shall reject the input
image if the intersecting region is too small for good cor-
3 Minutiae Data Synthesis respondence. We shall discuss our alignment method in
a separate section.
Our representation for the fingerprint consists of a global Let Mj and Mk , j "= k, j, k ∈ {1, 2, ..., m}, be two fin-
structure and a local structure [13]. The global struc- gerprint minutiae data sets to be synthesized. Suppose
ture consists of positional and directional information of there are p corresponding points (minutiae coordinates)
ridge endings and ridge bifurcations. The local structure between the two images. Denote this set of p correspond-
consists of relative information of each detected minutia ing points by C. Then, a topological transformation f
with other neighboring minutiae. Since the local struc- can be determined relating Mj and Mk from
ture contains relative information which is insensitive to xj = f (xk ) (5)
rotation and translation, the main issue concerning minu-
tiae data synthesis is to establish the relationship between where xj = {(xi , yi )}j and xk = {(xi , yi )}k for all i ∈ C.
the global structures of two fingerprints acquired with Since the transformation will be used for aligning those
common regions. non-corresponding minutiae points, a careful study on its
Let sensitivity with respect to noise and deformation is nec-
essary. We shall discuss various transformation models
M = {(xi , yi , ϕi , ti )}, i = 1, 2, ...n (1) for image points alignment in the following section.

be the set of minutiae containing the positional informa-

tion (x, y), directional information (ϕ) and minutiae type
information (ti = 0 indicates a ridge ending and ti = 1
4 Minutiae Detection and
indicates a bifurcation) for n minutiae elements in the Alignment
global structure.
Suppose we have a total of m number of minutiae data For minutiae detection, we adopt an adaptive ridge trac-
sets from m partial fingerprints of the same finger, then ing algorithm which is evolved from [14]. Our approach
adaptively traces the gray-level ridges of the fingerprint for this transformation, a further match combining both
image and applies adaptive oriented filters to the image local and global information is adopted.
only at those regions that require to be smoothed. A long Generally, transformation means can be classified in
tracing line will be obtained when there is little variation linear form and nonlinear form. Consider two sets of
in contrast and when the bending level of the ridge is image points: x = (x, y) and X = (X, Y ). The prob-
low. Main advantage of our approach is that tracing is lem here is to find the best transformation f that relates
by adaptive piece wise linear approximation of the ridges these two sets of image points, i.e. x = f (X). For linear
which speeds up the process of ridge detection as com- transformations, we have x = TX where T denotes the
pared to other methods which adopt either pixel wise or transformation matrix.
fixed step tracing [15]. The tracing is only performed Affine geometry compares distances only on the same
within the region of interest. The region of interest is line or on parallel lines. As compared to the Euclidean
segmented based on the local certainty level c(x, y) at geometry, affine geometry relaxes the requirement on per-
pixel (x, y) on image I. pendicularity. Hence transformation under affine geome-
Several stopping criteria for ridge tracing which deter- try is more general as compared to that under Euclidean
mines detection of minutiae are adopted as: geometry.
1. Tracing exits from region of interest. In this case, Projective geometry is the result of relaxing the restric-
minutiae extraction will not be performed. tions preserving parallel lines but require that straight
lines remain straight lines for any changes we might im-
2. Tracing ridge line intersects another already traced pose on the figure.
skeleton ridge line. Under this condition, a bifurca- Topology encompasses the projective, affine and
tion minutiae is detected. Euclidean geometries. An even smaller set of properties is
3. Tracing ridge line ends when the tracing line is invariant under topological transformation (e.g. preserv-
shorter than a threshold value and when the next ing only closed curves, order and connectivity). Here,
traced point lies on another ridge. only the quadratic type of topological transformation is
investigated.
In addition to minutiae detection, post processing is per-
formed to remove spurious minutiae.
Having the minutiae extracted for two images, it is
necessary to align the two sets of data so that they form
5 Experiments
a larger picture of the fingerprint. Several problems which
are inherent to this alignment process are enumerated as 5.1 Transformation study
follows:
In this section, we perform experimental study, using
1. Translation and rotation variance between the two physical fingerprint data, to determine a suitable trans-
fingerprint images. formation for minutiae synthesis. We collect 5 images
2. Some minutiae may be dropped and some spurious corresponding to 5 different areas (centre, top-left, top-
minutiae may be detected. right, bottom-left and bottom-right) for each finger for
the experiment. A total of 200 images were captured us-
3. Deformation of the fingerprint images which induces ing the Veridicom Sensor for 40 fingers.
location errors. Matching was first performed to obtain the correspond-
Together with an indication of ridge ending and ridge ing coordinates between two images which are to be syn-
bifurcation, the notation in (1) provides a global descrip- thesized. We used the centre area as the base image to
tion of the minutia. Since this feature vector is not rota- match with one of the other areas (top-left, top-right,
tion and translation invariant, we construct a local feature bottom-left and bottom-right) of the same finger. The
vector for our alignment purpose. In what follows, only a matched image pairs with 10 or more corresponding minu-
brief outline on the local structure for alignment match- tiae coordinates were then used for the following trans-
ing will be provided. The interested reader is referred to formation study. As a result, only 50 matched pairs were
[13] for greater details. found to have 10 or more matched points.
Let mj , j = 1, 2, ..., l be the j th nearest neighbour with To assess the accuracy of each transformation dis-
respect to m0 . The distance between m0 and mj can then cussed in previous section, 3/4 of the matched points
be expressed as were used for identifying the transformation parameters
$ (fitting) and the rest of 1/4 were used for extrapolation
dj0 = (xj − x0 )2 + (yj − y0 )2 . (6) test (testing). The distribution of the sum of squared er-
rors (SSE) for these matched pairs are plotted in Fig. 1
Denote by ϕ0 the direction of m0 , the relative radial angle and Fig. 2 for fit data and test data respectively. The
for mj with respect to m0 is given by continuous line (‘ ’) corresponds to SSE distribution for
% & affine transformation. The dashed line (‘- -’) and the
yj − y0 dotted line (‘...’) correspond to projective transformation
θj0 = tan−1 − ϕ0 , −π ≤ θj0 ≤ π. (7)
xj − x0 and topological transformation respectively. The mean
value and the standard deviation (STD) for these errors
Let cj0 , j = 1, 2, ..., l be the ridge count between m0 and are also tabulated in Table 1 to reflect an overall view of
mj , then together with corresponding minutiae type tj0 these results.
we pack the local feature vector as As seen from Fig. 1, the topological (quadratic) trans-
Flj = [dj0 , θj0 , cj0 , tj0 ]T , j = 1, 2, ..., l. (8) formation provides the best fit since the dotted curve falls
below the other two curves for all samples. This is also
It is obvious that this local structure is rotation and trans- reflected in Table 1 since the mean SSE and the stan-
lation invariant since it contains only relative information. dard deviation (STD) are the smallest among the three
Hence, it can be used directly for preliminary local align- transformations. As for test data not included in the fit-
ment matching. ting process, results from Fig. 2 and Table 1 show that
A match weighting the similarity between the local fea- affine transformation gives the best result, in the sense
ture vectors from the two images is performed so that of lowest mean SSE and lowest STD. It is important to
a common reference can be established. Once the pre- note that both the mean SSE and STD for the other two
liminary correspondence between these local features is transformations (projective and topological) are consider-
established, the transformation required for global align- ably huge as compared to those by affine transformation.
ment can be found. To validate good correspondences Main reason being that coordinate warping according to
 SSE distribution
600
Table 1: Sum of Squared Errors for fit and test
data
500
Sum of squared error for fit data
Affine Projective Nonlinear
400
Mean 31.1304 62.2254 16.4531
STD 25.3905 87.7418 19.4167
SSE

300
Sum of squared error for test data
Affine Projective Nonlinear
200
Mean 28.2344 227.7379 464.9472
STD 29.7245 428.2317 1559.9000
100

0
0 5 10 15 20 25 30 35 40 45 50
sample number The ‘circles’ in the figure indicates the original detected
minutiae points from Fig. 3, whereas the ‘plus’ and ‘stars’
Figure 1: Sum of squared error distribution for indicate those additional minutiae points transferred from
Fig. 4 and Fig. 5 respectively. As seen from Fig. 6, these
fit data additional minutiae points have found correct correspon-
dences on the fingerprint image (Fig. 3) which are not
detected in the original capture. A match comparing a
SSE distribution
2500 query image data with minutiae data from Fig. 6 will have
a higher matching count.

2000

1500
50
SSE

1000
100

500

150

0
0 5 10 15 20 25 30 35 40 45 50
sample number
200

Figure 2: Sum of squared error distribution for

test data 250

300
the fit data (interpolation) may not necessary fit well the 50 100 150 200 250 300
test data (extrapolation). Base on this study, the affine
transformation is adopted for alignment in our minutiae
synthesis system. Figure 3: Fingerprint sample 1 with detected
minutiae
5.2 A minutiae synthesis example
In this part, we show an example of synthesizing three
fingerprint images. As shown in Fig. 3 through Fig. 5,
three fingerprint images are captured from three differ-
5.3 Performance evaluation
ent portions of the same finger. Minutiae points (shown In this experiment, we show that the fingerprint synthesis
in circles in Figures 3-5) are detected from these finger- method can improve performance in terms of False Rejec-
print images using the ridge tracing algorithm. A visual tion caused by using different regions of fingerprints for
examination on these figures shall reveal that the minu- matching. A test sample consisting of 115 query images
tiae information extracted in each image contains similar and 6 × 115 template images were used for matching eval-
points (found in common regions) and dissimilar points uation. The query images were randomly acquired from
(found outside common regions). It is also observed that different partial regions of a finger of each individuals.
even within the common region, some minutiae detected The first five sets (labeled as (a)-(e)) of template images
in one image may not be detected in another image due to consist of different partial regions (i.e. centre, top-left,
different image qualities. Due to these reasons, when any top-right, bottom-left, bottom-right) from each enrolled
two of these three images are used for matching in a fin- finger. The last set (label as (f)) of templates consists
gerprint identification or verification system, false rejec- of synthesized data which are obtained by merging those
tion would occur when the threshold related to the total corresponding data from the same finger of the first five
number of matched minutiae pairs is set rather high. sets. As such, the last set of templates contains the same
Fig. 6 shows the synthesized minutiae points collected number of records as those in the first five sets, but with
from Figures 3-5, using Fig. 3 as the background image. richer information.
 50 50

100 100

150 150

200 200

250 250

300 300
50 100 150 200 250 300 50 100 150 200 250 300

Figure 4: Fingerprint sample 2 with detected Figure 5: Fingerprint sample 3 with detected
minutiae minutiae

In Table 2, the percentages of match between the query was adopted for minutiae data synthesis. The synthesized
image data set with each of the six template data sets template data set was found to improve matching perfor-
are shown. It is seen that the synthesized data set (f) mance in the sense of reducing false rejection which was
has provided the highest percentage of match for similar caused by using different fingerprint regions of the same
fingers. finger for matching.
The number of matched minutiae pairs are also plot-
ted in Fig. 7 for all test samples. In this figure, the solid
line represents the distribution of the number of matched References
minutiae pairs corresponding to the synthesized template [1] Anil Jain, Lin Hong, and Ruud Bolle, “On-line finger-
set (f) while all other dotted lines represent the distrib- print verification”, IEEE Trans. Pattern Analysis and
ution corresponding to template sets (a)-(e). It is seen Machine Intelligence, vol. 19, no. 4, pp. 302—313, 1997.
from this figure that the number of matched minutiae [2] Anil K. Jain, Lin Hong, Sharath Pankanti, and Ruud
pairs (solid line) provides almost a covering envelope for Bolle, “An identity-authentication system using finger-
the synthesize template set over all other template sets. prints”, in Proceedings of the IEEE, 1997, pp. 1365—
This indicates that most of the synthesized data set has 1388.
successfully captured the required minutiae information [3] Nalini K. Ratha, Kalle Karu, Shaoyun Chen, and
from individual template data set. Those ‘uncovered’ Anil K. Jain, “A real-time matching system for large
cases corresponds to much distorted information due to fingerprint databases”, IEEE Trans. Pattern Analysis
incorrect as well as inaccurate transformation as a result and Machine Intelligence, vol. 18, no. 8, pp. 799—812,
1996.
of the matching process. It is thus noted that obtaining
as much matching minutiae pairs before synthesizing the [4] U. Halici, L. C. Jain, and A. Erol, “Introduction to
fingerprint recognition”, in Intelligent Biometric Tech-
data could possibly help to improve the situation. niques in Fingerprint and Face Recognition, L. C. Jain
and et. al., Eds., pp. 3—34. The CRC Press international
series on computational intelligence, 1999.
Table 2: Percentage of match for each template [5] Lawrence O’Gorman, “Fingerprint verification”, in Bio-
set metrics: Personal Identification in Networked Society,
Anil K. Jain, Ruud Bolle, and Sharath Pankanti, Eds.,
1999, pp. 43—64.
Data sets
(a) (b) (c) (d) (e) (f) [6] Nalini K. Ratha, Jonathan H. Connell, and Ruud M.
57.39% 50.88% 54.95% 41.12% 36.79% 80.00% Bolle, “Image mosaicing for rolled fingerprint construc-
tion”, in 14th International Conference on Pattern
Recognition, 1998, vol. 2, pp. 1651—1653.
[7] Raffaele Cappelli, Alessandra Lumini, Dario Maio, and
Davide Maltoni, “Fingerprint classification by di-
rectional image partitioning”, IEEE Trans. Pattern
Analysis and Machine Intelligence, vol. 21, no. 5, pp.
6 Conclusion 402—421, 1999.
[8] Anil K. Jain and Lin Hong, “A multichannel approach
In view of the limitation in solid state image sensor area, to fingerprint classification”, IEEE Trans. Pattern
we propose, in this paper, a method to synthesize fin- Analysis and Machine Intelligence, vol. 21, no. 4, pp.
gerprint data. The method is advantages over existing 348—359, 1999.
mosaicing technique in terms of low computational cost [9] Kalle Karu and Anil K. Jain, “Fingerprint classifica-
and low memory storage requirements. Several transfor- tion”, Pattern Recognition, vol. 29, no. 3, pp. 389—404,
mation models were compared for minutiae points align- 1996.
ment. The affine transformation, which was found to [10] Ugur Halici and Güclü Ongun, “Fingerprint classifica-
provide good interpolation and extrapolation capabilities, tion through self-organizing feature maps modified to
 80

70

50
60

number of matched minutiae

50
100
40

30
150

20

200 10

0
0 20 40 60 80 100 120
250 sample number

Figure 7: Distribution of number of matched

300 minutiae for each test sample
50 100 150 200 250 300

Figure 6: Fingerprint sample with synthesized

minutiae (o: from samlpe 1; +: from sample 2;
*: from sample 3)

treat uncertainties”, in Proceedings of the IEEE, 1996,

pp. 1497—1512.
[11] U.S. Government Printing Office, Washingtion, D. C.,
The Science of Fingerprints: Classification and Uses,
1984.
[12] H. C. Lee and R. E. Gaensslen, Eds., Advances in Fin-
gerprint Technology, Elsevier, New York, 1991.
[13] Xudong Jiang and Wei Yun Yau, “Fingerprint minutiae
matching based on the local and global structures”, in
15th Internaional Conference on Pattern Recognition,
1999.
[14] D. Maio and D. Maltoni, “Direct gray-scale minutiae
detection in fingerprints”, IEEE Trans. Pattern Analy-
sis and Machine Intelligence, vol. 19, no. 1, pp. 27—40,
1997.
[15] Xudong Jiang, Wei Yun Yau, and Wee Ser, “Minutiae
extraction by adaptive tracing the gray level ridge of the
fingerprint image”, in IEEE Internaional Conference
on Image Processing (ICIP’99), 1999, pp. 852—856.