NIKAM et al.

: FINGERPRINT RECOGNITION

Fingerprint Recognition

Shankar B. Nikam, Suneeta Agarwal,

Department of Computer Science and Engineering, Motilal Nehru National Institute of Technology, Allahabad
Email [email protected], [email protected]

 Fingerprint acquisition: This deals with acquiring

Abstract—Fingerprint verification is one of the most fingerprints in digital format.
reliable personal identification methods. However, Representation: This deals with storing fingerprints in
manual fingerprint verification is so tedious, time- databases. An alternative to image-based representation is
consuming, and expensive that it is incapable of to store only the extracted features.
meeting today’s increasing performance requirements. Pre-processing: The quality of fingerprints is enhanced by
An automatic fingerprint verification is widely needed. applying different tools. In a local area, the ridges and
Fingerprints have been an invaluable tool for law valleys
enforcement and forensics for over a century. More
of a fingerprint have a well defined frequency and
recently fingerprints have found an application in
orientation, therefore , it is natural to use frequency
biometric systems. Fingerprints are the most common
analysis tools such as Fourier Transforms, Gabor filters
and trusted biometric for personal identification. Due
to its security-related applications and interest from and wavelets to enhance ridges[1]. Histogram
the developers of biometric systems, automated equalization, Filtering, Orientation field estimation,
fingerprint identification systems are becoming binarisation, thinning are common steps in pre-
increasingly widespread and are being extensively processing.
researched. This paper discusses extensive research Feature extraction-This is concerned with finding and
that has been done on automated fingerprint measuring important properties of the fingerprint that are
verification. useful for fingerprint matching.
Matching: This compares features extracted from input
Keywords-AFIS, Biometrics, Orientation fields, Ridge detection fingerprint with features of template fingerprint and
determines whether these fingerprints have been captured
I. INTRODUCTION from the same finger.
Fingerprints are graphical flow like ridges present on
human fingers. They have been used as a means of
personal identification for several centuries. Fingerprint Acquisition
Traditionally, the driving force behind advancements in
fingerprint technology has been law enforcement agencies Representation
and forensic scientists. Massive fingerprint databases
have been collected by law enforcement agencies around
the world. Automated systems are necessary for the Preprocessing
efficient management of these databases. Recently
fingerprints have found applications in biometric systems. Feature Extraction
Biometrics is the automatic identification of an individual
based on physiological or behavioral characteristics. Due
to the continuing needs of law enforcement and biometric Matching
systems, efficient automated fingerprint identification
systems (AFISs ) are becoming popular.
Fig. 1 The stages of a Typical AFIS
An Automated fingerprint identification system mainly
consist of following stages as given in figure 1.
The main goal of this paper is to review different types of
NIKAM et al.: FINGERPRINT RECOGNITION

feature extraction approaches and their matching

algorithms. The section Feature extraction discusses
various types of features that can represent fingerprint
and approaches to extract them. Matching algorithms are
discussed in the section Fingerprint Matching. The
section Conclusions closes with concluding remarks and
future research directions.

Fig. 2. A fingerprint image(left) and its orientation field (right)

II. FEATURE EXTRACTION
Different types of features and methods to extract them Orientation can also be estimated using the ridge-
are as follows- valley algorithm that use a 9 x 9 mask as shown in
Minutiae Features: figure 3. The algorithm is based on one originally
The most important fingerprint features are minutiae. proposed for the binarisation of fingerprint images [3].
Minutiae are ridge endings and bifurcations. Reliable The first step is the convolution of the mask with the
extraction of minutiae is vital to a system’s performance. image. The mask is comprised of 8 slits, numbered 1
Following are the main approaches of minutiae through 8. The slit sum Si for i=1,…,8 is equal to the
extraction. sum of the pixel intensity values at all spots that
Minutiae extraction from ridge skeletons- contain an i in the mask. The next step is to classify
These algorithms generally consist of following three each pixel as belonging to a ridge or a valley, and
stages: orientation field estimation, ridge detection and
thinning. Minutiae are extracted from the ridge skeleton.
1.Orientation field estimation: Orientation fields contain
information about the local directions of fingerprint
ridges. Orientation can be estimated using Jain et al
algorithm that is modified version of Rao’s algorithm [2]
which has following main steps.
1. Divide the input fingerprint into blocks of size W
X W.
2. Compute the gradients Gx and Gy at each pixel in
each block
3. estimate the local orientation of each block using
following formula: Fig. 3 The ridge valley mask

this is done by comparing the slit sums. Let Smin be the

lowest slit sum at a pixel and Smax be the highest. The
pixel is classified as a “valley” pixel (white) if:
The quality of estimation field can be improved by
finding consisting level. A consistency level is
calculated based on the variation of orientations in a
The intuition behind this is that if the pixel belongs to a
block. If the consistency level is below a given
valley, one of the slits will lie along that valley, and
threshold, the orientations for this region are re-
consequently have a high sum. Valleys are white, have a
estimated at a lower image resolution This process is
high intensity. The other slits will cross ridges and
continued until the consistency level is above the
valleys, giving them similar (and lower) slit sums.
threshold. An example of orientation field is shown in
Therefore, the average of Smin and Smax is expected to be
figure 2.
higher than the average of all the slit sums. If equation is
false, the pixel is classified as a “ridge” pixel. Valley
pixels are assigned the orientation of Smax because the slit
sum with the highest sum is the most likely to lie along
NIKAM et al.: FINGERPRINT RECOGNITION

the valley. Similarly, the ridge pixels are assigned the determine whether the window contains minutiae.
orientation of Smin. This process assigns each pixel an 4)Marco Gamassi et. al. proposed fingerprint local
orientation (from 8 angles). The method is simple and analysis for high performance minutiae extraction. Based
fast. However, the orientation field is coarse due to the on the analysis of the local properties, the typical patterns
limit of 8 possible directions. of minutiae are identified by studying the intensity along
2 Ridge detection: The most salient property squared paths in the image. This method works both on
corresponding to ridges is that gray level values on ridges grey-level and binarized images.
attain their local maxima along the normal directions of Non-Minutiae features:
local ridges. Therefore, pixels can be identified to be Minutiae are computationally expensive to extract,
ridge pixels based on this property. Jain et al. convolve unreliable for low quality images, and vulnerable to
fingerprint image with two masks. The masks are capable nonlinear deformations. Consequently, there has been
of adaptively accentuating the local maximum gray level some research into features not based on minutiae.
values along the normal direction of the local ridge Minutiae Descriptor-Recently, Tico and Kuosmannen
direction After applying the masks, thresholding is used [5] built a minutiae descriptor for each minutiae, which
to determine whether a pixel belongs to a ridge or valley. consists of the original minutiae point and a set of local
The ridge valley algorithm can also be used for ridge orientation values uniformly sampled around this point.
detection. The matching algorithm based on these minutiae
3.Thinning: This gives a skeleton of ridge map where descriptors was reported to have better performance than
ridges are one pixel thick. Due the presence of noise, before.
breaks, and smudges, etc in the input image, the resulting FingerCode- Jain et al. [6] proposed a novel feature
binary ridge map often contains holes and speckles. called FingerCode. They first detected the reference
Therefore, a hole and speckle removal procedure is needs point and extracted the region of interest around it, then
to be applied after thinning. filtered the image with a bank of Gabor filters with
4.Minutiae extraction: Minutiae detection is a trivial task different orientations. The filtered results are discretely
when an ideal thinned ridge map is obtained. Any black coded as the feature, and combined with the minutiae for
pixel that has only one black (eight-connected) neighbor final decision. This method implicitly used the global
is a ridge ending, and any black pixel with more than two orientation information. To avoid its sensitivity to the
black neighbors is a bifurcation. However, due to poor reference point detection, Ross et al. [7] proposed to use
the entire filtered images, called “ridge feature map,” as
quality of images spikes and broken ridges will create
the features, and combined them with the minutiae as a
spurious minutiae. Heuristics are used to remove
hybrid matcher. All three of these algorithms need large
spurious minutiae.
storage for saving the additional features in the
Direct grey scale minutiae extraction- fingerprint template, especially the hybrid matcher and
Following methods extract minutiae directly from grey the minutiae descriptor.
scale images. Orientation Field-One important kind of features in
1) Maio and Maltoni propose a minutiae extraction based fingerprints is the orientation field which is already
on following ridge lines in the grey scale image [4]. discussed. It is defined as a matrix whose elements are
Ridges are traced using orientation field, and the the ridge direction at the corresponding pixel (block) in
locations of ridge endings and bifurcations are recorded. the original image. The direction is defined in [0,Π)
2) Liu et al. proposed another method of ridge line instead of [0,2Π).. The orientation field directly describes
tracing. Instead of tracing only ridges, the authors the global structure of the fingerprint ridge pattern. It has
propose to trace both ridges and valleys, and observe been used for image enhancement in the preprocessing
their structural relationship. A ridge and its two stage [8] and fingerprint type classification (indexing)
neighboring valleys are traced, and if the two valleys join [9]. As a kind of global feature, the orientation field has
a ridge ending is found. Similarly, if the distance between many advantages over the minutiae: 1) it is almost
the two valleys grows, a bifurcation is found. continuous and smooth everywhere except those regions
3) The application of Fuzzy logic to minutiae extraction near the singular points, which makes it more robust to be
from grey scale images has been investigated by Sagar et extracted and less sensitive to the noise; 2) compared with
the minutiae, it is less sensitive to the image deformation
al. A small window is scanned over the fingerprint image,
due to the skin condition, the pressure, etc. Few works, as
and fuzzy rules based on pixel intensities are used to
far as the authors know, utilized its discriminant
NIKAM et al.: FINGERPRINT RECOGNITION

information for fingerprint recognition. This is partly due the fingerprint for the purpose of sampling ridge features
to the difficulty of storing the entire orientation field in as in figure 5. These lines are referred as fiducial lines.
the feature template. Finally, for each intersection of a fiducial line and a
Sweat Pores-Michael Ray et al. [10] has presented a ridge, four features are calculated: (1) the distance since
approach to extract the locations of sweat pores from the the last intersection, (2) the angle of intersection, (3) the
gray scale image. They used a modified minimum change in angle since the last intersection, and (4) the
squared error approach and a slightly modified 2- curvature of the ridge at the intersection. When calculated
dimensional Gaussian function as given below: for each intersection, these features capture information
M(x,y)=1 – e- |x2+y2| about the orientation, spacing and curvature of the
They have shown that, majority of pores within a 500 dpi fingerprint ridges.
image can be approximated using above function. These
locations of sweat pores can be used as additional
discriminatory information in fingerprint representations.
Ridge patterns-Some approaches extracted and used the
ridge patterns (shape and frequency) as features for
matching. Rui Ma et al. presented a ridge line extraction
approach by way of ridge line tracing. Fingerprint ridge
line in a grey scale image is viewed as a track of a ridge Fig. 5. Fingerprint ridges with horizontal fiducial lines
segment moving along the ridge. The curve tracing for sampling
problem is solved by the target tracking technique in
computer vision. The model of fingerprint ridge line DCT Features- DCT features can be used for
segments is formulated first and then a target tracking fingerprint matching [12]. In the feature extraction
method is applied to trace each of the ridge lines. In process, a fingerprint image is quartered and transformed
addition, a feedback technique is adopted to correct the by the Discrete Cosine Transform (DCT). The standard
fingerprint directional image in each tracing step in order deviation of the DCT coefficients in predefined areas is
to improve tracing accuracy. By connecting all the traced then calculated and used for matching. Similarly,
ridge line segments, a polyline extraction of the ridge line matching can be done based on FFT and wavelet
can be obtained. features.
A geometric framework for analyzing the global ridge Learned Features- Fingerprint classification approach
structure of a fingerprint image has been developed by based on a novel feature learning algorithm is presented
Chong et al. [11]. Their method approximates the shape [13]. Approach is based on Genetic Programming (GP),
of each ridge using a B-spline curve. Another which learns to discover composite operators and features
mathematical model for representing fingerprint ridge that are evolved from combinations of primitive image
structures based on hierarchical kernel fitting has been processing operations. This approach can find good
proposed by Jain and Minut. For each fingerprint class a composite operators to effectively extract useful features.
fingerprint kernel is defined that represents the ridge Bayesian classifier is used for classification.
structure of fingerprints belonging to that class. The Singularities- A singularity is a local region of a
fingerprint kernels are defined using polynomial splines fingerprint where the ridge pattern has special
as shown in figure 4. properties making it visually prominent. There are two
types of fingerprint singularities: cores and deltas (see
Fig. 6). A core is the turning point of an inner- most ridge
and a delta is a place where ridges running side-by-side
diverge. Singularities are mainly used for fingerprint
Fig. 4. The kernels (from left to right) for arch, left loop, classification. They also have applications such as
right loop and whorl. alignment landmarks for fingerprint matching.
Singularities are found using Poincare index. The
Fiducial Lines- Senior has developed a novel method for Poincare index is defined as the rotation of the vectors
extracting fingerprint ridge features. The first step is to along a curve in the orientation field. All points of a
obtain a thinned binary representation of the fingerprint fingerprint can be classified as either core points, deltas
ridges. Next, a set of horizontal lines are arranged across or normal points depending on their Poincare index. The
NIKAM et al.: FINGERPRINT RECOGNITION

Poincare index is computed by considering a small curve noise and deformations. Each minutia in a fingerprint is
around a point in the orientation field and summing the associated with a ridge. Therefore, here alignment is
changes in the direction of vectors along the curve. achieved by aligning corresponding ridges (see fig. 7).
Next, a string representation of both minutiae sets is
constructed based on the positions of the minutiae. Each
character in the string corresponds to a single minutia,
and they are ordered by increasing polar coordinates (the
candidate minutiae pair are defined to be at the origin of
the coordinate system). A dynamic programming string
matching algorithm is used to define an “edit distance”
between the strings corresponding to the test minutiae set
and reference minutiae set. Finally, matching score is
Fig. 6. Core point (left) and Delta Point (right)
calculated with following formula:
100Npair
III. FINGERPRINT MATCHING Matching Score =
There are two distinct (but related) fingerprint matching Max{M,N}
problems. Fingerprint identification is the problem of
matching a query fingerprint against a database to
identify its owner. Fingerprint verification (or
authentification) is concerned with determining whether
or not two prints are from the same finger.
Various approaches for fingerprint matching are as
follows-
Minutiae pattern matching-
Generally, an automatic fingerprint
verification/identification is achieved with point pattern
matching (minutiae matching) instead of a pixel-wise
matching or ridge pattern matching of fingerprint images. Fig. 7 An alignment using ridges
Because a general point matching problem is essentially
intractable, features associated with each point and their Since the fingerprint distortions are not modeled, it is
spatial properties such as the relative distances between necessary to allow for some tolerance when matching
points are often used in these algorithms to reduce minutiae pairs. This is accomplished by using bounding
exponential number of search paths. The relaxation boxes whose size grow as the distance from the reference
approach, The Hough transform-based approach, energy minutiae increases. This is a powerful matching
minimization are some of the ways for point matching. algorithm, but has a high computational complexity.
Jain et al. [14] has presented alignment-based matching The system proposed by Jinwei Gu et al. [15] use both
algorithm. The first step is to align point patterns. The minutiae and model-based orientation field. Their
output of the minutiae extraction stage is typically a list fingerprint representation contains three parts: the
of minutiae locations and orientations, represented by the orientation field model, the minutiae (and the singular
3-tuple (x,y,θ). The associated ridges are also recorded. points if available), and the effective region where the
Aligning the minutiae sets is known as registration, and fingerprint ridge pattern exhibits. (see Fig. 8). Fingerprint
is essentially a point pattern matching problem. The goal matching can be done by combining the decisions of the
is to find a translation, rotation (and possibly scaling) matchers based on the global structure (orientation field)
that aligns two point sets. Due to different number of and the local cue (minutiae) (see Fig. 9). The system is
minutiae in each set, missing and spurious minutiae and more robust, fast and accurate than conventional minutiae
nonlinear deformations of points, fingerprint point based methods. In this system, the feature template takes
matching is a difficult problem. Due to large number of 420 bytes, and the feature extraction and matching
possible correspondences between point sets, an procedures can be done in about 0.30 s.
alignment by corresponding point pairs is not practical Structural Matching-
even though it is feasible. It is well known that Structural information in pattern recognition encodes the
corresponding curve segments are capable of aligning two interrelationship of low-level features. This is a natural
point patterns with a high accuracy in the presence of approach for minutiae matching because the relationship
NIKAM et al.: FINGERPRINT RECOGNITION

between a fingerprint’s minutiae contains much of a

fingerprint’s individuality. Graphs are used to represent
structural relationships, and are consequently a common
tool for structural pattern recognition.
Ching-Han Chen et al. [16] presented an Embedded
fingerprint authentication system with reduced hardware
resources requirement. They have estimated the
displacement of the fingerprint pattern by means of
directional image matching, and then circular profile
features of the fingerprint are extracted for verification.
Their system is implemented in an 8-bits microcontroller
running at 25MHz.
A Fingerprint matching algorithm based on Delaney
triangulation net is presented by Ning Liu et al. [17]. On Fig. 9 Flowchart of the proposed combination system
the basis of Delaney triangulation in computational
geometry, they used DT net to find reference minutiae Shenglin et al. presented secure fingerprint verification
pairs (RMPs). Using DT on the topological structure of system based on Fuzzy vault scheme to address a major
minutiae set, a DT net is formed with minutiae as secutity hole currently existing in most biometric systems
vertices. From the nets of input minutiae set and the [18]. The construction of the fuzzy vault during the
template minutiae set, certain pairs of minutiae that have enrollment phase is automated by aligning the most
similar structures as RMPs are selected for reliable reference points between templates, based on
alignment. Finally, matching is done using point pattern. which the converted features are used to form the lock
set. Their system has high unlocking complexity for
attackers with an acceptable unlocking accuracy for the
legal users.
Ridge based fingerprint matching using Hough transform
is presented by Aparecido Nilceu Marana et al. [19]. This
technique extracts the major straight lines using Hough
transform. These lines match the fingerprint ridges and
are used to estimate rotation and translation parameters
necessary for registration of two fingerprint images. After
the registration, matching score is computed based on the
ridge alignment. The combination of ridge-based
matching and minutia based matching improves overall
performance.
A Hidden Markov model based approach is presented by
Hao Guo [20]. A fingerprint has different orientation
angle structure in different local areas and has a texture
pattern correlation among the neighboring local areas.
Fig. 8 Fingerprint representation with global and local This can be viewed as a Markov stochastic field. Here,
features. (a) original image, (b) minutiae and singular fingerprint matching is based on embedded Hidden
points, (c) polygon-based effective region, and (d) model- Markov Model (HMM) that is used for modeling the
based orientation field. fingerprints orientation field. System is developed by
matching embedded HMM parameters which are built
after processing of extracted features, forming the
samples of observation vectors and training the embedded
HMM parameters. Correlation based matching, fuzzy
neural network based approach, Cellular Neural Network
(CNN) based verification are some of the other
approaches for fingerprint verification.
NIKAM et al.: FINGERPRINT RECOGNITION

IV. CONCLUSION [8] L. Hong, Y. Wan, and A, K, Jain, “Fingerprint image

Due to incomplete or poor-qualitied fingerprint images, enhancement: Algorithm and performance
non-linear deformations, fingerprint verification is still a evaluation,” IEEE Trans. Pattern Anal. Mach. Intell.,
complex problem, not fully solved and there is need of vol. 20, no. 8, pp. 777-789, Aug. 1998.
fast, robust, accurate approaches to fingerprint [9] R. cappelli, A. Lumini, D. Maio, and D. Maltoni,
verification. There are several areas where future “Fingerprint classification by directional image
research needs to be focused. Minutiae give rich partitioning,” IEEE Trans. Pattern Anal. Mach.
information about fingerprint topology. Fast accurate Intell., vol. 21, no. 5, pp. 402-421, May 1999.
and reliable minutiae extraction and their matching needs [10] Michael Ray, Peter, Reza Adhami, “A Novel
further investigation. Non-minutiae feature extraction approach to fingerprint pore extraction,” Proc. IEEE,
methods need performance improvement. Hybrid systems 2005
incorporating several independent feature sets (including [11] Neil Yager, Adnan Amin, “Fingerprint
minutiae and non-minutiae features) can be used to classification: a review,” in Pattern Analysis
increase robustness, and this approach will likely become Applications (2004) 7: 77-93
more common in future. Security of fingerprint [12] S. Tachaphetpiboon, T. Amornraksa,” A Fingerprint
verification systems is one important aspect. Fingerprint matching method using DCT features,” Proc. of
scanners may be susceptible to spoofing using artificial ISCIT2005IEEE, 2005
materials, or in the worst case, dismembered fingers. An [13] Xuejun Tan, Bir Bhanu, Lin, “Fingerprint
anti-spoofing methods based on liveness detection should classification based on learned features,” IEEE
be incorporated in fingerprint verification systems. Since Trans. Systems, Man, and Cyber., vol. 35, no. 3,
there is a great demand of fingerprint based biometric pp.287-300, Aug. 2005.
systems in the market and there are more opportunities [14] Anil Jain, Lin, Rudd, “On- line fingerprint
for innovations in this area, automatic fingerprint verification, “IEEE Trans. Pattern Anal. Mach.
verification continues to be an important area for Intell., vol 19, no. 4, pp. 302-314, April 1997.
academicians and developers of biometric systems. [15] Jinwei Gu, Zhou, Yang, “ Fingerprint recognition by
combining global structure and local cues,” IEEE
REFERENCES Trans. Image Process., vol 15, no. 7, pp. 1952-1964,
July 2006.
[1] Neil Yager, Adnan Amin, “Fingerprint verification [16] Ching-Han Chen, Jia-Hong Dai, “An Embedded
based on minutiae features: a review,” in Pattern fingerprint authentication with reduced hardware
Analysis Applications (2004) 7: 94-113 resources requirement,” Proc. IEEE, April 2005
[2] Rao A (1990) A taxomy for texture description and [17] Ning Liu, Yilong Yin, Hongwei Zhang, “ A
identification. Springer, Berlin Heidelberg, New Fingerprint matching algorithm based on Delaunay
York. triangulation net,” in IEEE Proc. of Fifth
[3] Stock RM, Swonger CW “Development and International Conference on Computer and
evaluation of a reader of fingerprint minutiae”, Information Technology(CIT’05), 2005
Cornell Aeronautical Laboratory, 1969, Technical [18] Shenglin Yang and Ingrid Verbauwhede, “
Report CAL No. XM-2478-X-1:13-17,Ithaca, NY. Automatic secure fingerprint verification system
[4] Maio D, Maltoni D, “ Direct grey-scale minutiae based on fuzzy vault scheme,” Proc. IEEE ICASSP
detection in fingerprints. IEEE Transactions on 2005, pp. v-609-612.
Pattern Anal. Mach. Intell.,1997, 19(1):27-40. [19] Aparecido Nilceu Marana, A. K. Jain, “Ridge based
[5] M. Tico and P. Kuosmannen, “ Fingerprint matching fingerprint matching using Hough transform,” Proc.
using an orientation-based minutia descriptor,” IEEE of the 18th Brazilian symposium on Computer
Trans.on Pattern Anal. Mach. Intell., vol. 25, no. 8, graphics and Image processing (SIBGRAPI’05),
pp.1009-1014, Aug. 2003. 2005
[6] A. K. Jain, Salil Prabhakar, Hong, Pankanti, “ [20] Hao Guo, “ A Hidden Markov Model fingerprint
Filterbank-based fingerprint matching,” IEEE Trans. matching approach,” IEEE Proc. of the Fourth Inter.
Image Process.,vol. 9, no. 5, pp.846-859, May 2000. Conf. Mach. Learn. And Cybernetics, Aug. 2005.
[7] A. Ross, A. Jain and Reisman, “A hybrid fingerprint
matcher,” Pattern Recognit., vol 36, no.7, pp.1661-
1673, 2003.