Contactless Fingerprint Recognition:
a Neural Approach
for Perspective and Rotation Effects Redu ction
Ruggero Donida Labati, Angelo Genovese, Vincenzo Piuri, Fabio Scotti
Departmen t of Computer Science
Universit`a degli Studi di Milano
Milano, 201 22, Italy.
{ruggero.donida, angelo.genovese, vincenzo.piuri, fabio.scotti}@unimi.it
Abstract—Contactless fingerprint recognition systems are be-
ing researched in order to reduce intrinsic limitations of tra-
ditional biometric acquisition technologies, encompassing the
release of latent fingerprints on th e sensor platen, non-linear
spatial distortions in the captured samples, and relevant image
differences with respect to the moisture level and pressure of t he
fingertip on the sensor surface.
Fingerprint images captured by single cameras, however, can
be affected by perspective d istortions and deformations due to
incorrect alignments of the finger with respect to the camera
optical axis. These non-idealities can modify the ridge pattern and
reduce the visibility of th e fingerprint details, thus decreasing the
recognition accuracy. Some systems in the literature overcome
this problem by computing three-dimensional models of the
finger. Unfortunately, such approaches are usually based on
complex and expensive acquisition setups, which limit their
portability in consumer devices like mobile phones and tablets.
In this paper, we present a novel approach able to recover
perspective deformations and improper fingertip ali gnments in
single camera systems. The approach estimates the orientation
difference between two contactless fingerprint acquisitions by
using neural networks, and permits to register t he considered
samples by applying the estimated rotation angle to a synthetic
three-dimensional model of the finger surface. The generalization
capability of neural networks offers a significant advantage
by allowing processing a robust estimation of the orientation
difference with a very l imited need of computational resources
with respect to traditional techniques. Experimental results show
that the approach is feasible and can effectively enhance the
recognition accuracy of single-camera biometric systems. On the
evaluated dataset of 800 contactless images, the proposed method
permitted to decrease t he equal error rate of the used biometric
system from 3.04% to 2.20%.
I. INTRODUCTION
Fingerprint recognition systems usually adopt acquisition
proced ures that require the contact of the finger with a sen sor.
Contact-based acquisition techniques, however, can produce
samples affected by non-linear sp atial distortions and low
contrast regions due to improper pressures of the finge r on
the sensor platen. Moreover, these technologies suffer from
an impo rtant security lack since every biometric acquisition
releases a latent fing erprint on the sensor surface.
In order to avoid these problems and to improve the usability
and user acceptance of fingerprint-based biometric systems,
Fig. 1. Possible rotations of the finger with respect to the camera optical
axis in contactless recognition systems.
contactless reco gnition techniques based on CCD cameras
are researched. These techniques can also permit to increase
the possible applicative contexts of fingerprint biom e trics. For
example, contactless biometr ic systems can be adopted in mo-
bile devices with integrate d CCD cameras without introducing
additional hardware costs. Moreover, they are more ro bust to
dust and dirt with respect to contact-based technologies.
One of the most important non-idealities of the fingerprint
samples acquired by contactless recognition systems consists
in the presence of perspective distortions due to rotations of the
finger with respect to the camera optical axis (Fig. 1), which
are particularly relevant in fingerprint images captured without
using finger placem e nt guides [1]. Perspective distortions can
drastically re duce the accuracy of the matching algorithms
since most of them require samples with a constant resolution.
The majority of the fingerprint matching techniqu es, in fact,
is based on the evaluation of the metric distances between
minutia points [2,3].
In m any applicative contexts, like consumer devices, it is not
possible to adopt complex and expensive acquisition systems.
Contactless fingerprint rec ognition systems based on single
© 2013 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future
media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or
redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
CCD cameras theref ore re sult m ore suitable. Examples of this
kind of biometric systems are described in [4,5, 7–10] . Most
of the existing technologies, however, are n ot able to obtain
accuracy levels co mparable to contact-based systems since
they do not use feature extraction and matching techniques
specifically d esigned to overcome the typical non-idealities of
contactless samples.
In this paper, we pr opose an approach fo r perspective
deformation and roll rota tion registration in contactless fin-
gerprint recognition systems b a sed on a single CCD camera.
In particular, th e method is designed to work in uncontrolled
applications, like recognition systems integrated in mobile
devices. The c ontribution of the paper is threefold: we propose
a specific set of features f or the estimation of the roll angle
of the finger, we present a deformation recovery strategy
based on synthetic three-dim ensional m odels of the finger,
and w e describe the results obtained by applying the proposed
approa c h in a co mplete biometric system.
Fig. 2 shows the schema of the bio metric recognitio n
process based on the proposed approach for the perspective
and r otation effects reduction. I n the enrollment phase, the
system estimates and stores perspective features of the sample,
and then it computes a fixed numbe r of minutiae templates
related to fingerprint images obtained by simulating acquisi-
tions performed with different roll angles. In th e verification
phase, the system fir st estimates a set of perspective features
from the f resh sample. These characteristics are related to
the silhouettes of the compare d finge r images and to char-
acteristics estimated using Gabor filters. Then, a feature set is
computed from the perspective characteristics of the compared
acquisitions, and the r oll angle difference between the two
fingerpr int images is estimated b y usin g neural networks.
Finally, the matching score is computed by comparin g the
minutiae template obtained from the fresh sample and the
stored template associated to the estimated angle. In a similar
manner, the proposed approach can be applied in identification
systems.
The results of the proposed approach are evaluated on
biometric samples captured with a great variability of roll an -
gles. The accuracy increa se obtained by a complete biometric
system b a sed on the proposed approach is then analyzed.
The paper is structured as follows. Section II briefly de-
scribes the state of the art related to contactless fingerprint
recogn ition systems. The proposed biometric recognition ap-
proach is then presented in Section III. The obtained results
and a comparison with a well-known recognition technique in
the literature are discussed in Section IV. Finally, Section V
proposes conclusions and final remar ks.
II. RELATED WORKS
Contactless fingerprint recognition technologies can be di-
vided in to systems based on three-dimensional models, and
systems based on two-dimensional samples. The systems
appertaining to the first class permit to obtain good quality fin-
gerprint samples, overcoming problems related to persp e ctive
distortions and finger rotations. These technologies, however,
(a)
(b)
Fig. 2. Biometric recognition process based on the proposed approach for
the perspective and rotation effects reduction: (a) enrollment; (b) verification.
It is also possible to use the proposed approach in identification systems.
can be adopted in a limited set of applicative scenarios since
they require complex and expensive acq uisition methods, such
as multiple view techniques [ 1,11–14] and structured light
setups [15–17].
Contactless fingerprint recognition systems based on two-
dimensional samples usually adopt less expensive acquisition
techniques, and can therefore be applied in a wider range of
scenarios, like consumer devices. The biometric recognition
process per formed by contactless systems b ased on two-
dimensional samples can usually be divided into the sequ ent
steps: acquisition; computation of a contact-equivalent finger-
print image; feature extraction and matching.
The simplest acquisition techn ique adopted by two-
dimensional systems consists in the use o f a low-cost CCD
camera in uncontrolled light conditions. A biometric system
that captures fingerprint images using a webcam in natur al
light con ditions is presented in [4] , and studies on the use of
mobile phone cameras are described in [5,18].
The images captured in unc ontrolled light conditions, how-
ever, present poor contrast between ridges and valleys. For
this reason , most of the contactless r e cognition systems in the
literature use illu mination techniques to improve the visibility
of the ridge pattern, like a point light source [8– 10] or ring
illuminators [19]. In the literature, there are also studies on
the light wavelengths that permit to enhance the visibility
of the ridge p a ttern [1 1,20]. These studies report that lon g
wavelength rays, like white and infrared light tend to penetrate
the skin , and to be absorbed by the epidermis. Differently, a
blue light with a wavelength of 500 nm permits to obtain better
quality images. For this reason, some contactless fingerprint
acquisition systems use illumination techniques based on blue
led lamps [15,21]. The work presented in [20] compares
illumination setups based on different light wavelengths, light
positions, polarizations, and diffusion techniques. Other ac-
quisition systems adopt transmission-based illumination tech-
niques, like the system described in [22], which uses a red
light source placed on the fingerprin t side to focus the light
transmitted through the finger onto a CCD.
The finger print images obtain e d by single views, however,
can present a non-uniform resolution and out of focus regions
caused by the mapping of the finger shape into a two-
dimensional image. In order to overcome this problem, some
systems compute the mosaicking of multiple views [23] or use
ring-shaped mirrors [15].
Usually, the samples captured by contactless sensors cannot
be directly elaborated by recognition methods designed for
contact-based fingerprint im ages. In [18], the p erformances
obtained by a commercial fingerpr int recognition software
on contactless images a re evaluated. The paper reports that
sufficient results have been obtained only on the best quality
images. In order to perform biometric recognitions ba sed
on well-known methods in the literature, the computation of
contact-equivalent images is usually performed. Th is step can
be divided into two tasks: image enhancement, and resolution
normalization. I n the literature, there are different image
enhancement techniques. The method described in [4] first
performs a prepr ocessing task based on the Lucy-Richardson
algorithm and on the deconvolution of the input image by a
Wiener filter. Then, a background subtraction algorithm and a
cutoff filter tuned according to the mean ridge freque ncy are
applied. Differently, the image enha ncement method proposed
in [8–10] adopts a contextual filtering technique based on the
STFT analysis [24]. Similarly to the technique presented in
[25], the method described in [7] applies Gabor filters tuned
accordin g to the local ridge frequency and orientatio n, but
it c omputes the r idge orientation map by using an iterative
regression algorithm designed to be more robust to the noise
present in contactless fin gerprint images.
The resolution normalization is then applied in order to
obtain contact-equivalent fingerprint images that can be used
by matching techniques based on minutia features. Some
contactless recognition systems that require the placement of
the finger at a fixed position o btain this result by evaluating
the in formation relate d to the focal length and the distance
between the finger and the camera [11,19]. Systems that do
not impose this constraint can only perform an approximated
normalization, like the method presented in [4], which is based
on the evaluation of the finger silhouette.
Most of the biometric technologies based on contactless
fingerpr int images perform the recogn ition task by using
methods based on minutia features since they usually permit
to obtain more accuracy with respe ct to algorithms b ased
on global featu res. These systems usually adop t matching
algorithm s design ed for con ta c t-based images [4,5,7–11,23].
These algorithms however, require fingerprin t images with
known resolution and acquired with controlled rotations. In
order to perform the recognition in scenarios that do not
impose constraints on the finger placement, some systems
use matc hing methods based on adimensional features. The
matching technique proposed in [26] is based on a feature
set similar to the Fingercode [27]. This method uses the
principal component analysis (PCA) to search the most distinc-
tive featu res and supp ort vector mach ines (SVM) to perform
the template comparison. The technique described in [28]
compare s local features centered in the m inutia points by using
neural classifiers.
In this paper, we present an approach designed to reduce
problems related to finger rotations and perspective distortions
in contactless finge rprint images. In order to obtain accurate
biometric recognitions, the approach is designed to be inte-
grated with state of the art minutia-based techniques.
III. THE PROPOSED APPROACH
The proposed approach permits to reduce problems related
to finger rotations and perspective distortions in contactless
fingerpr int samples captured by a single CCD camera. It uses
neural networks and specifically designed features to estimate
the roll angle difference between biometric acquisitions. The
estimated angle permits to compute a recovered fingerprint
image by rotating a synthetic three-dimensional model of the
finger sur face.
In the enrollment phase, the biometric recognition system
based on the proposed approach stores n
θ
rotated templates
in the biometric database. Every template consists in minutiae
features extracted using the software MINDTCT [29] of the
National Institute of Standards and Technology (NIST). Ad-
ditional data describing per spective deformations of the finger
silhouette and ridge pattern are also stored in order to be u sed
in the verification phase for searching the best template to be
compare d with the fresh acquisition. These data consist in an
array of 18 real numb e rs extracted from the fin ger silhouette
and the ridge pa ttern.
In the verification phase, the set of features describing the
perspective deformation of the fresh sample are estimated.
These features and the ones stored in the compared identity
representation are u sed to estimate a feature set describing
the angular difference between the considered fingerprint
acquisitions. Neural networks are then adopted to numerically
estimate this angle. The obtained value is finally used to select
the best rotate d minutiae template to be compared with the
fresh fing e rprint sample.
The biometric recognitio n process based on the proposed
approa c h is sh own in Fig. 2, and can be divided into the
sequent steps:
1) contactless acquisition ;
2) image preprocessing;
3) simulation of finger rotations;
4) feature extraction;
5) rotation estimation with neural networks;
6) template computation;
7) matching.
A. Contac tless acquisition
Fingerprint images ar e captured contactless by a single CCD
camera placed at a distance of more than 20 cm from the
finger.
The acquisition setup d oes not use finger placement guid es,
but requ ires that the finger is placed on a surface with a fixed
distance to the cam e ra in order to control the lens focus.
Using this hardware configuration, the finger can there fore be
placed with uncon trolled yaw orientations. In future works,
the reference surface used for the finger placement should
be removed by adopting techniqu e s f or the e stima tion of the
best quality frames in frame sequences repre senting a finger
moving toward the camera [30].
The hardware setup also uses a uniform blue light in order
to enhance the visibility of the ridge pattern.
B. Image preprocessing
This step aims to compute a contact-equivalent image f rom
the captured contactless fingerprint sample. The preprocessing
step can be divided into two tasks: image enhancement, and
resolution no rmalization.
1) Image enhancemen t: first, the region of interest (ROI)
is estimated by usin g the Otsu’s method [31] and r efined by
using a morp hological filling operator.
The enhancement of the ridge visibility is then performed.
First, the background image I
B
is estimate d by applying a
morphological o pening operation with a mask s to the image I.
Then, the background is rem oved, o btaining the imag e I
R
. In
order to incre a se the visibility of the ridge pattern, we p e rform
a nonlinear e qualization as I
L
(x, y) = log (I
R
(x, y)). A noise
reduction is then performed by applying a 8-ord er Butterworth
low pass filter [31] with frequency f
f
and size d
f
× d
f
. The
values of f
f
and d
f
have been empirically estimated on the
used dataset.
Finally, the enhanceme nt and binarization of the ridge
pattern is performed by using the rid ge following technique
implemented by the NIST MINDTCT software [29]. This
algorithm d irectly com putes the binar y image of the ridge
pattern I
B
by evaluating the shape of every ridge of the image
I
L
.
2) Resolution normalization: the proposed resolution nor-
malization technique assumes that the contactless fing erprint
images are captured at a constant distance ∆
H
from the
camera.
This method first e stima te s the resolution of the captured
image by evaluating the size of the plain P captured at a
distance ∆
H
from the camera, an d then it normalizes the
contactless image to a resolutio n of 500 ppi.
Considering the plain P , r
x
inch along the horizontal direc-
tion of this plain correspon d to i
x
pixel along the horizontal
diction of the captured im ages. The normalization factor is
then estimated as:
n
f
= i
x
/ (r
x
· PPI) , (1)
where PPI is equal to 500. The value of n
f
is computed offline
by measu ring the c haracteristics of the acquisition setup.
The normalized ridge pattern I
N
and the normalized ROI
image R
N
are then computed.
Finally, the images I
N
and R
N
are cropped along the y
axis in order to re move the regions that do not describe the
(a)
(b) (c)
Fig. 3. Image preprocessing: (a) contactless fingerprint image I ; (b) ridge
pattern image I
N
; (c) ROI image R
N
.
last phalanx. Starting from of the max imum y coordinate of
the ROI, the height of the images I
N
and R
N
is reduced to
h
c
pixel, where h
c
is a value empirically estimated on the
considered dataset.
An example of contactless fingerprint image I, the corre-
sponding ridge pattern image I
N
, and the ROI image R
N
are
shown in Fig. 3. It is possible to observe that the obtained im-
age I
N
presents a non-uniform resolution due to the ir regular
three-dimensional shape of the finger.
C. Simulation of fin ger rotations
The proposed technique for the simulation of finger rota-
tions uses a synthetic thr e e-dimensional finger model obtained
from the ROI image R
N
, and then it computes a rigid
transformation in the three-dimen sio nal space.
The synthetic three-dimensional model of the finger consists
in a depth map Z computed propor tionally to the finger
silhouette. The curvature of the finger is co nsidered as a third
order polynomial with height proportional to the width of
every column o f R
N
.
First, the po lynomial p approximating the finger curvature
is defined as the third order po lynomial passing by the x and y
coordinates (x
min
, 0), (x
m
−x
m
· c
W
, c
H
), (x
m
+x
m
·c
W
, c
H
),
(x
max
, 0). Where x
min
is the minimum x coordinate of R
N
,
x
max
is the maximum x coordinate of R
N
, x
m
= (x
max
+
x
min
)/2, c
W
and c
H
are parameters empirically estimated o n
the conside red dataset.
A vector C representing the finger curvature is obtained by
fitting the polynomial p in the interval from x
min
to x
max
.
Every column i of Z is then define d as:
Z(i) = R
N
(i) × C × (X
min
(i) − X
max
(i)), (2)
where X
min
and X
max
are vectors representing the minimum
and maximum x coordinates of the ROI image R
N
at every
(a)
(b) (c)
Fig. 4. Simulation of finger rotations: (a) synthetic three-dimensional model
of the finger shape; (b) rotated fingerprint model; (c) resulting image.
column i. An example of obtain ed depth map is shown in
Fig. 4 a.
Subsequently, the image of the ridge patter n I
N
is super-
imposed on the ma trix Z.
The coordinates of the depth map Z are then rotated by an
angle θ as:
Z
θ
= Z
1 0 0
0 cos θ − sin θ
0 sin θ cos θ
, (3)
where θ is the clockwise angle from the x axis.
The rotated image I
θ
is obtained by applying a resampling
with a con stant step equal to 1 to the image I
N
in the new x
and y coordinates of Z
θ
. This task considers the image I
N
as
a gray scale image, and it is based on a bilinear interpolation.
An example of rotated fingerprint three-dim ensional model
is shown in Fig. 4 b, and the corre sponding rotated image I
θ
is shown in Fig. 4 c.
D. Featu re extraction
A set of perspective features P is extracted for each sample
during the enrollment and verification phases. The set P is
composed by the finger silhouette asimmetry δ and the matrix
G representing the ridge pattern characteristics extracted by
applying Gabor filters with different or ie ntations to the image
I
N
.
1) Finger silhouette characteristics: in order to detec t the
fingertip orientation with particular referen ce to the roll angle ,
experiments showed that it is possible to measure the horizon-
tal asymmetry of the final part of the finger. For roll angles
close to zero, the shape of the finger silhouette tends to be
mostly symmetrical and, on the contrary, a rotated fingerprint
shows relevant difference s in the left and right parts of the
silhouette.
The silhouette asymmetry can b e processed more robustly
after a rotation compensation of the pan angle. With this aim,
we process a rotation of the ROI of the finge rtip (the binary
mask R
N
) by applying a bilinear interpolation rot() along its
centroid and minimizing the horizontal asymmetry:
(
∆(φ, y) = R
edge
(rot(R
N
, φ), y) − L
edge
(rot(R
N
, φ), y)
ˆ
φ = argmin
φ
(
P
b
y=a
|∆(φ, y)|)
(4)
where ∆(y) represents the size differences for each y coo r-
dinate between the right R
edge
(y) and left L
edge
(y) edges of
the rotated ROI mask R
N
; a and b represent the minimum and
maximum rows of the the ROI mask;
ˆ
φ is the optimal ang le
minimizing the displacement of the finger silhouette.
Once the ROI image has been rotated, the remaining hor-
izontal asymmetr ic contribution is processed in the first part
of the silhouette as:
δ =
c
X
y=a
∆(
ˆ
φ, y)
(5)
where c = ⌊κb⌋, with c > a, which is a floored integer
representin g a row in the middle of the rotated ROI images.
Proper values of the param eter κ are the ones allowing the
processing of the horizontal asymmetric contribution only in
the first third of the fingertip.
2) Ridge pattern characteristics: in contactless fin gerprint
images, perspective deformations can be detected by using
global fing erprint character istics (Level 1 analysis [2]). For this
reason, a set of characteristics are computed by using Gabor
filters with different orientations ψ. This set is composed by
n
G
= 32 values. Starting from the image I
N
, two images G
ψ
are computed by apply ing two Gabor filters with orientations
(0
◦
, 90
◦
). Each image G
ψ
is then divided into 4×4 rectangular
regions with the same size, and the absolute average distance
(AAD) of the intensity is computed for each region. In the
spatial domain, a symmetric Gabor filter can be described as:
G(x, y; f, ψ) = exp
(
−
1
2
"
x
′
2
σ
2
′
x
′
+
y
′
2
σ
2
′
y
′
#)
cos 2πf x
′
,
x
′
= x sin ψ + y cos ψ, (6)
y
′
= x cos ψ − y sin ψ,
where f is the frequency of the sinusoidal plane wave along
the directio n ψ from the x-axis, and σ
′
x
and σ
′
y
are the spac e
constants of the Gaussian envelope along the x
′
and y
′
axes,
respectively.
The ob ta ined AAD values are sto red in a 4 × 4 × 2 matrix
G representing the results obtained by applying Gabor filters
with orientations 0
◦
and 90
◦
in spatial ord er.
E. Rota tion estimation with neural networks
This step estimates a discrete value representing th e roll
angle difference ∆
theta
between two biometric a c quisitions A
and B by evaluating the p erspective feature sets P
A
and P
B
.
First, a feature set F is computed as:
F = d
f
(P
A
, P
B
), (7)
