U.S. patent application number 11/472709 was filed with the patent office on 2007-12-27 for fingerprint representation using localized texture features.
Invention is credited to Rudolf Maarten Bolle, Sharat Suresh Chikkerur, Sharathchandra UmapathiRao Pankanti, Nalini Kanta Ratha.
Application Number | 20070297653 11/472709 |
Document ID | / |
Family ID | 38873616 |
Filed Date | 2007-12-27 |
United States Patent
Application |
20070297653 |
Kind Code |
A1 |
Bolle; Rudolf Maarten ; et
al. |
December 27, 2007 |
Fingerprint representation using localized texture features
Abstract
A system and method for processing fingerprints includes
representing each minutiae in a fingerprint by determining
quantized Gabor coefficients to represent texture content of the
minutiae. A distance is computed between represented minutiae and
stored minutiae. The minutiae matches are ranked based on the
distance to identify the fingerprint.
Inventors: |
Bolle; Rudolf Maarten;
(Bedford Hills, NY) ; Chikkerur; Sharat Suresh;
(Cambridge, MA) ; Pankanti; Sharathchandra
UmapathiRao; (Manhasset, NY) ; Ratha; Nalini
Kanta; (Yorktown Heights, NY) |
Correspondence
Address: |
KEUSEY, TUTUNJIAN & BITETTO, P.C.
20 CROSSWAYS PARK NORTH, SUITE 210
WOODBURY
NY
11797
US
|
Family ID: |
38873616 |
Appl. No.: |
11/472709 |
Filed: |
June 22, 2006 |
Current U.S.
Class: |
382/124 |
Current CPC
Class: |
G06K 9/00073
20130101 |
Class at
Publication: |
382/124 |
International
Class: |
G06K 9/00 20060101
G06K009/00 |
Claims
1. A method for processing fingerprints, comprising: representing
each minutiae in a fingerprint by determining quantized Gabor
coefficients to represent texture content of the minutiae;
computing a distance between represented minutiae and stored
minutiae; and ranking minutiae matches based on the distance to
identify the fingerprint.
2. The method as recited in claim 1, further comprising pooling
evidence related to the fingerprint to assist in confirming the
identity.
3. The method as recited in claim 1, wherein computing includes
computing a Hamming distance.
4. The method as recited in claim 1, further comprising listing
candidate identities based on a score.
5. The method as recited in claim 1, wherein representing includes
computing Gabor elementary function coefficients and the minutiae
are represented using the coefficients.
6. The method as recited in claim 1, further comprising optimizing
the Gabor coefficients using a gradient descent approach.
7. The method as recited in claim 1, further comprising reducing
the number of Gabor coefficients to reduce the representative size
of the print.
8. The method as recited in claim 1, wherein representing includes
representing each minutiae based on a 32 by 32 neighborhood of
pixels.
9. A method for processing fingerprints, comprising: inputting a
fingerprint image; identifying minutiae in the fingerprint image;
computing Gabor coefficients to represent texture content of each
minutiae in the fingerprint; and quantizing the Gabor coefficients
for storage in memory.
10. The method as recited in claim 9, further comprising storing
evidence related to the fingerprint to assist in confirming the
identity.
11. The method as recited in claim 9, wherein computing Gabor
coefficients includes optimizing Gabor coefficients using a
gradient descent approach.
12. The method as recited in claim 9, further comprising reducing
the number of coefficients to reduce the representative size of the
fingerprints.
13. The method as recited in claim 9, wherein each minutiae in the
image is represented by a 32 by 32 neighborhood of pixels.
14. The method as recited in claim 9, wherein the memory includes a
database and further comprising the association of an identity with
a quantized representation of the fingerprint image in the
database.
15. A computer program product for processing fingerprints
comprising a computer useable medium including a computer readable
program, wherein the computer readable program when executed on a
computer causes the computer to perform the steps of claim 9.
16. A computer program product for processing fingerprints
comprising a computer useable medium including a computer readable
program, wherein the computer readable program when executed on a
computer causes the computer to perform the steps of: representing
each minutiae in a fingerprint by determining quantized Gabor
coefficients to represent texture content of the minutiae;
computing a distance between represented minutiae and stored
minutiae; and ranking matches based on the distance to identify the
fingerprint.
17. A system for processing fingerprints, comprising: an input
device configured to receive fingerprint images; an extraction
module configured to determine minutiae in the fingerprint; a
representation module configured to compute Gabor coefficients to
represent texture content for each minutiae to represent the
fingerprint; and a comparing module for computing a distance
between represented minutiae in the fingerprint and stored minutiae
in memory to rank matches based on the distance to associate an
identify with the fingerprint.
18. The system as recited in claim 17, wherein computing includes
computing a Hamming distance.
19. The system as recited in claim 17, wherein the Gabor
coefficients are optimized using a gradient descent approach.
20. The system as recited in claim 17, wherein the system is a
scalable 1:N identification system.
Description
BACKGROUND
[0001] 1. Technical Field
[0002] The present invention relates to fingerprint processing and
more particularly to a system and method for employing Gabor
coefficients in a texture representation for fingerprint
recognition.
[0003] 2. Description of the Related Art
[0004] Biometric identifiers, such as, fingerprints, face, iris and
voice prints, offer a way of reliable personal authentication.
Biometrics is rapidly replacing traditional token and password
methods. Of all the biometric modalities, fingerprints have emerged
as a popular choice due to their universality, distinctiveness,
permanence and acceptability. Another reason for their popularity
is the wide variety of implementations of recognition algorithms
that are already available.
[0005] Existing fingerprint matching algorithms may be broadly
classified into the following categories based on fingerprint
representation.
[0006] 1. Correlation based: In this representation, the
fingerprint image itself is used as a template. Matching is
performed by measuring the result of cross correlation between the
two images. This requires reasonably low resolution images and is
very fast, since correlation may also be implemented through
optical techniques. However, the matching is global and requires an
accurate registration of the target and reference fingerprint
images, since correlation is not invariant to translation and
rotation. The accuracy of correlation based techniques further
degrades with non-linear distortion of the fingerprint.
[0007] 2. Minutiae Representation: Minutiae represent local
fingerprint ridge discontinuities and mark the position where a
ridge comes to an end or bifurcates into two. Given target and
reference fingerprints and their corresponding minutiae features,
the process of matching is a point pattern matching problem. This
is by far the most popular approach to fingerprint recognition.
However, minutiae based matching algorithms do not perform well in
the case of small fingerprints that have very few minutiae.
Furthermore, minutiae based systems completely ignore gray scale
content of the fingerprint images. Human experts routinely utilize
the rich structural and texture cues present in the fingerprint
image during the process of matching. Minutiae based
representations do not encode this information either explicitly or
implicitly.
[0008] 3. Texture Descriptors: A fingerprint image can also be
viewed as a pattern of oriented texture formed by the gray scale
variation of the ridges. Therefore, texture descriptors provide a
good representation for the ridge content in the image. A global
texture descriptor scheme called `finger code` utilizes both global
and local ridge descriptions. The features are extracted by
measuring the responses of radial fingerprint image sectors to a
filterbank. The matching is based on measuring the Euclidean
distance between the feature vectors. A disadvantage of this
approach is that it requires that the fingerprint core is
accurately located. This is a difficult problem in itself.
[0009] Thus, algorithms based on minutiae require a point matching
algorithm, but do not measure gray scale content. On the other
hand, correlation and texture based methods measure gray scale
content but require very accurate alignment. Another issue with the
above mentioned techniques is that of scalability. It has been
shown that algorithms designed for 1:1 verification scale poorly
when used for 1:N identification tasks.
[0010] Fingerprint recognition based on localized information
selects `interesting` regions in the fingerprint image to perform
local gray scale correlation. Plain ridges do not carry any
information except their orientation, ridge frequency, ridge
endings and ridge bifurcations. The `interesting` regions (similar
to distinctive minutiae configurations) include regions around the
minutiae, regions of high curvature and regions around the singular
points such as core and delta. However, the optimal process of
selecting these `interesting` regions is very inefficient.
Furthermore, the algorithm is not robust to rotations. Localized
correlation is used in other methods, but only in conjunction with
a geometry based minutiae matcher. The localized correlation is
used to assess the quality of the pairs obtained by the minutiae
matcher. The localized correlation itself is not accurate enough to
be useful. Furthermore, the algorithms are designed for 1:1
verification and are therefore not directly scalable for large
scale identification tasks.
SUMMARY
[0011] A system and method for processing fingerprints includes
representing each minutiae in a fingerprint by determining
quantized Gabor coefficients to represent texture content of the
minutiae. A distance is computed between represented minutiae and
stored minutiae. The minutiae matches are ranked based on the
distance to identify the fingerprint.
[0012] Another method for processing fingerprints includes
inputting a fingerprint image, identifying minutiae in the
fingerprint image, computing Gabor coefficients to represent
texture content of each minutiae in the fingerprint, and quantizing
the Gabor coefficients for storage in memory.
[0013] A system for processing fingerprints includes an input
device configured to receive fingerprint images, an extraction
module configured to determine minutiae in the fingerprint, and a
representation module configured to compute Gabor coefficients to
represent texture content for each minutiae to represent the
fingerprint. A comparing module computes a distance between
represented minutiae in the fingerprint and stored minutiae in
memory to rank matches based on the distance to associate an
identify with the fingerprint.
[0014] These and other objects, features and advantages will become
apparent from the following detailed description of illustrative
embodiments thereof, which is to be read in connection with the
accompanying drawings.
BRIEF DESCRIPTION OF DRAWINGS
[0015] The disclosure will provide details in the following
description of preferred embodiments with reference to the
following figures wherein:
[0016] FIG. 1A is an image showing relative geometry used for
minutiae based matchers;
[0017] FIG. 1B is an image showing the gray scale image content
used in correlation and texture based matchers;
[0018] FIG. 2A shows gray scale images of illustrative flow pattern
configurations;
[0019] FIG. 2B shows gray scale images of illustrative highly
distinctive flow pattern configurations;
[0020] FIG. 3 shows image representations of Gabor basis functions
used to represent minutiae neighborhoods;
[0021] FIG. 4 shows gray scale images of minutiae neighborhoods
reconstructed in accordance with present principles;
[0022] FIG. 5 is a receiver operating characteristic (ROC) curve
obtained by matching pairs of single minutiae and showing the
effect of neighborhood sizes;
[0023] FIG. 6 is a diagram showing enrollment of fingerprints into
a database using Gabor features/coefficients in accordance with one
embodiment;
[0024] FIG. 7 is a block/flow diagram showing identification of
fingerprints using Gabor features/coefficients in accordance with
another embodiment;
[0025] FIG. 8 is a block/flow diagram showing a system for
enrollment and identification of fingerprints using Gabor
features/coefficients in accordance with another embodiment;
and
[0026] FIGS. 9A-9C are loose (FIGS. 9A and 9C) and strict (FIG. 9B)
Cumulative Match Curves (CMCs) for identification of fingerprints
using Gabor features/coefficients in accordance with present
principles.
DETAILED DESCRIPTION OF PREFERRED EMBODIMENTS
[0027] Embodiments in accordance with present principles provide
representations of localized texture features and outline
dimensionality reduction for use in fingerprint identification.
Fingerprint representations can be broadly divided into three
categories: gray scale image level, texture derived features and
minutiae features. Both image based and texture based
representations need accurate alignment before comparison. This
presents a problem since accurate localization of singular points
and registration of fingerprints is still a pertinent problem in
this field.
[0028] On the other hand, minutiae based matchers are invariant to
changes in orientation and position, but completely ignore the rich
gray scale content in the image. In accordance with present
principles, a localized texture based representation system and
method are provided that rely on gray scale content for
identification and at the same time do not need absolute alignment.
Techniques to efficiently compute these features and methods to
perform identification based on these features are provided.
Experimental evaluations are also presented on databases of several
sizes that show that the features are both accurate and
scalable.
[0029] Embodiments of the present invention can take the form of an
entirely hardware embodiment, an entirely software embodiment or an
embodiment including both hardware and software elements. In a
preferred embodiment, the present invention is implemented in a
combination of hardware and software. The software may include but
is not limited to firmware, resident software, microcode, etc.
[0030] Furthermore, the invention can take the form of a computer
program product accessible from a computer-usable or
computer-readable medium providing program code for use by or in
connection with a computer or any instruction execution system. For
the purposes of this description, a computer-usable or computer
readable medium can be any apparatus that may include, store,
communicate, propagate, or transport the program for use by or in
connection with the instruction execution system, apparatus, or
device. The medium can be an electronic, magnetic, optical,
electromagnetic, infrared, or semiconductor system (or apparatus or
device) or a propagation medium. Examples of a computer-readable
medium include a semiconductor or solid state memory, magnetic
tape, a removable computer diskette, a random access memory (RAM),
a read-only memory (ROM), a rigid magnetic disk and an optical
disk. Current examples of optical disks include compact disk-read
only memory (CD-ROM), compact disk-read/write (CD-R/W) and DVD.
[0031] A data processing system suitable for storing and/or
executing program code may include at least one processor coupled
directly or indirectly to memory elements through a system bus. The
memory elements can include local memory employed during actual
execution of the program code, bulk storage, and cache memories
which provide temporary storage of at least some program code to
reduce the number of times code is retrieved from bulk storage
during execution. Input/output or I/O devices (including but not
limited to keyboards, displays, pointing devices, etc.) may be
coupled to the system either directly or through intervening I/O
controllers.
[0032] Network adapters may also be coupled to the system to enable
the data processing system to become coupled to other data
processing systems or remote printers or storage devices through
intervening private or public networks. Modems, cable modem and
Ethernet cards are just a few of the currently available types of
network adapters.
[0033] Referring now to the drawings in which like numerals
represent the same or similar elements and initially to FIGS. 1A
and 1B, images are illustratively shown to demonstrate the
differences between minutiae based matchers (FIG. 1A) and
correlation and texture based matchers (FIG. 1B). In FIG. 1A,
minutiae are depicted. Minutiae based matchers rely on relative
geometry but do not need to be fully aligned. Correlation and
texture based matchers rely on gray scale content as shown in FIG.
1B, but require alignment.
[0034] Embodiments in accordance with the present principles
provide a local texture based representation scheme that relies
solely on gray scale content and at the same time does not require
absolute alignment. Using these features, a system and method for
identification is provided. It is to be noted, that the method
performs 1:N (scaling) matching.
[0035] Localized texture features: To obtain position and rotation
invariance, localized texture features are relied upon instead of
relying on global descriptions. The texture is quantified within
the local vicinity of each minutiae, and this evidence is collected
across all minutiae during identification. While this information
is most useful when combined with the geometrical information given
by the minutiae, as a first step and also to measure the relative
merit of the new features, identification is performed exclusively
on texture features. While most minutiae are not distinctive, there
are highly distinctive minutiae that convey a lot of information
about the identity of the fingerprint itself (See, e.g., FIG. 1A
and FIG. 1B). Experimental evidence to be described hereinafter
proves that such an approach is effective.
[0036] Texture Representation: A multi-resolution representation
using a Gabor expansion in accordance with present principles is
employed to represent a minutiae neighborhood. The bases are
derived using self-similar Gabor elementary functions computed at
multiple scales and orientations.
[0037] The present principles represent fingerprint images with
rotationally invariant indices derived from local function
expansions around fingerprint image landmarks. An example of such
function expansions includes Gabor function expansions. The
function expansions may be optionally transformed to make the
representation revocable and anonymous to preserve privacy. This
includes that the coefficients, and therefore, the biometrics are
transformed with a non-invertible, revocable transform.
[0038] If for some reason the coefficients are compromised: (1) the
real (untransformed) coefficients cannot be recovered because the
transform is non-invertible (and moreover secret). Therefore, the
true biometric cannot be reconstructed. But also (2) the
coefficients can be transformed with a new non-invertible,
revocable transform and the subject can be re-enrolled. It is
similar to, e.g., the cancellation of a credit card. A new number
(biometric) is given and at the same time your true biometric is
not compromised since it was previously transformed anyway.
[0039] Every person in the database could have his/her biometric
transformed so that the true biometric is not exposed. It is better
to keep the transform secret, but even if the transform is
compromised, because of the non-invertibility, the biometric cannot
be reconstructed. If the transform is invertible, the transform has
to be kept secret. The transform is revocable because a subject can
at any time be re-enrolled with a new transform.
[0040] Existing fingerprint indexing solutions use more global
invariant features such as fingerprint pattern class or geometric
primitives such as line segments and triangles with associated
ridge counts. Drawbacks of these methods include that they offer
small penetration rates (i.e., fraction of the prints filtered out
due to indexing is small) or are overly complex.
[0041] Referring to FIGS. 2A and 2B, minutiae configurations are
shown for common minutiae configurations (FIG. 2A) and for highly
distinctive minutiae configurations (FIG. 2B).
[0042] The present embodiments compose invariant indices from novel
minutiae features where the features may be distorted for privacy
reasons. The system and method includes two phases, e.g.,
enrollment and identification. These two phases may be
illustratively summarized as follows.
[0043] Enrollment:
[0044] 1. Process the fingerprint image(s) of an individual to
obtain the minutiae.
[0045] 2. For each (one or more) minutiae: a. Obtain (Gabor)
function expansion features from the minutiae neighborhood(s). b.
Optionally distort the expansion coefficients with a secret,
revocable transformation. c. Prepare an index from the (distorted)
minutiae features. d. Store the person identity (and possibly other
minutiae information) in the entry indicated by the index of an
enrollment table.
[0046] Identification:
[0047] a. Initialize evidence for the enrolled identity or
identities to zero. b. Obtain the minutiae from query
fingerprint(s) of an unknown individual to be recognized.
[0048] c. For each (one or more) minutiae: 1. Obtain (Gabor)
function expansion features from the minutiae neighborhood(s). 2.
Optionally distort the expansion coefficients with the same secret
transformation. 3. Prepare an index from the (distorted) minutiae
features. 4. Retrieve the identity or identities stored in the
entry indicated by the index of the enrollment table. 5. Obtain the
match score of the query and indexed minutiae.
[0049] d. Determine the N minutiae with the highest match scores
and use these to accumulate evidence for the associated identities.
e. One or more identities corresponding to the highest evidence are
returned as candidate identities.
[0050] Referring to FIG. 3, some of the Gabor basis functions used
to derive local feature representations are visually depicted.
These representations are chosen since Gabor elementary functions
have important signal properties such as optimal joint space
frequency resolution. Gabor elementary functions form a very
intuitive representation of fingerprint images since they capture
the periodic, yet non-stationary nature of the fingerprint regions.
Furthermore, apart from being a good representation for texture
information, Gabor elementary functions also form a very good
sparse representation for image representation. Gabor elementary
functions have been employed as basis functions to represent
generic 2D images.
[0051] The following form of the 2D-Gabor elementary function, may
be given:
G(x,y)=exp(-.pi.[(x-x.sub.0).sup.2.alpha..sup.2+(y-y.sub.0).sup.2.beta..-
sup.2])exp(-2.pi.[u.sub.0(x-x.sub.0)+v.sub.0(y-y.sub.0)]) (1)
[0052] Here x.sub.0 and y.sub.0 represent the center of the
elementary function in the spatial domain. u.sub.0 and v.sub.0
represent the modulation frequencies. .beta..sup.2 and .beta..sup.2
represent the variance along the major and minor axes respectively,
and therefore the extent of support in the spatial domain. A
32.times.32 gray scale region is represented around a fingerprint
minutiae using basis functions that span two scales and four
orientations resulting in 272 coefficients (25+9=34 positions and
four orientations for each position and 2 (real+imaginary)
components for each coefficient). The gray scale image can
therefore approximately be represented using:
I ( x , y ) = n = 1 136 a n G n ( x , y ) ( 2 ) ##EQU00001##
[0053] The basis function at scale m, displacement
(p,q).epsilon.(0, 2.sup.m, 2.sup.2m . . . , N) and orientation
.theta. is given by:
.PSI..sub.mpq.theta.(x, y)=2.sup.-mG(x.sup.l,y.sup.l) (3)
x.sup.l=2.sup.-m[x cos(.theta.)+y sin(.theta.)]-p (4)
y.sup.l=2.sup.-m[-x sin(.theta.)+y cos(.theta.)]-q (5)
[0054] Referring to FIG. 4, examples are shown of normalized gray
scale minutiae neighborhoods and their approximate reconstructions
using Gabor expansion. In FIG. 4, a left image (L) shows the
original minutiae neighborhood and a right image (R) shows the
reconstructed minutiae neighborhood. Three pairs of images are
illustratively shown. The almost perfect reconstruction supports
the choice of Gabor basis functions for the gray scale image
content representation.
[0055] Computing Gabor Expansion: Unlike Fourier bases or discrete
cosine bases, Gabor elementary functions have the following
problems. (1) From a signal processing point of view, Gabor
elementary functions do not form a tight frame. This means that the
gray scale image is not represented as a linear superposition of
Gabor elementary functions with coefficients derived by projecting
the gray scale image onto a set of basis functions. However,
conditions have been established under which a set of self-similar
Gabor basis functions form a complete and approximately orthonormal
set of basis functions. (2) Gabor elementary functions form
biorthogonal bases. This means that the basis functions used to
derive the coefficients (analysis functions) and the basis
functions used to reconstruct the image (synthesis functions) are
not identical or orthogonal to each other. However, there are
several ways in which an optimal solution in the mean square sense
can be obtained.
[0056] Gradient Descent: A simple optimization approach has been
devised to obtain the coefficients (See Daugman, "Complete Discrete
2D Gabor Transform by Neural Networks for Image Analysis and
Compression, Transactions on Acoustics", Speech and Signal
Processing, 36:1169-1179, 1988, incorporated herein by reference).
This approach tries to minimize the squared difference between the
reconstructed and the original image. In other words, the approach
tries to minimize the function:
( a 1 , a 2 , a N ) = x , y { I ( x , y ) - k N a k G k ( x , y ) }
2 ( 6 ) ##EQU00002##
[0057] The optimal coefficients may be obtained by a gradient
descent approach as suggested by Daugman. At each iteration the
coefficient is incremented by an amount:
.DELTA. a n = G n ( x , y ) . I ( x , y ) - G n ( x , y ) . ( k G k
( x , y ) I ( x , y ) ) ( 7 ) ##EQU00003##
[0058] Once the coefficients are computed, the coefficients are
quantized using a=sign(a), which returns the sign of the
corresponding element a. This is inspired by a similar
representation used to quantify texture in iris images (See, e.g.,
Daugman, "High Confidence Visual Recognition of Persons by a Test
of Statistical Independence", Trans. on PAMI, 15(11):1148-1161,
1993.
[0059] However, this representation is further optimized in
accordance with present principles based on studies in visual
perception. In our preliminary experiments, we manually identified
corresponding minutiae in a small set of pairs of fingerprints and
computed the Hamming distance between corresponding and
non-corresponding minutiae.
[0060] Referring to FIG. 5, a Receiver Operating Characteristic
(ROC) curve shows that the proposed features are very
discriminative. It is to be noted that this ROC is based on
comparison between single minutiae. It is therefore very likely
that good discrimination will be achieved by pooling evidence from
multiple minutiae within the fingerprint. As seen from FIG. 5, the
neighborhood size plays a role in the feature discriminatory
powers. Curves for 32.times.32, 48.times.48 and 64.times.64
neighborhoods are depicted. 32.times.32 neighborhoods are
preferably selected based on this preliminary observation.
[0061] Referring to FIG. 6, an overview of the enrollment process
is illustratively depicted. A fingerprint 102 to be enrolled is
preprocessed using an enhancement scheme known in the art to make
the features robust to variable acquisition conditions. Then,
minutiae detection is performed to identify the location and
orientation of all the points of interest 104 within the
fingerprint 102. Subsequently, Gabor expansions 105 in 32.times.32
neighborhoods around the minutiae are extracted and normalized
based on the minutiae's orientation. Fingerprint distortion
transformations may be applied for security to the Gabor
features/coefficients. The Gabor features/coefficients are
quantized 106 and stored in a database 108 along with the
fingerprint's associated identity 110.
[0062] Identification: During identification, the quantized Gabor
expansion coefficients/features 106 are extracted similar to the
enrollment process. To identify the candidate fingerprint in the
database, evidence is pooled from results of all the minutiae from
the query print. The method is outlined below with reference to
FIG. 7.
[0063] Referring to FIG. 7, an identification method in accordance
with the present principles is illustratively depicted. For each
minutiae in a query print Q: in block 202, extract the Gabor
coefficients/features a.sub.1, a.sub.2, . . . a.sub.N from the
print. This may include fingerprint distortion transformations by
altering the Gabor features/coefficients. The print may be
distorted in accordance with the same transformation employed in
the enrollment phase. In block 204, quantize the Gabor
coefficients/features.
[0064] In block 205, perform a feature or coefficient reduction to
reduce the number of coefficients. In block 206, compute a distance
(e.g., the Hamming distance) between the extracted features and all
the minutiae in the database. In block 208, rank the minutiae
matches in increasing order of distance, i.e., a candidate minutiae
with the smallest Hamming distance has the rank of 1. In block 210,
for each candidate minutiae with rank K (for simplicity consider
only K.ltoreq.10 candidates), pool and/or use evidence about the
candidate's identity. If the candidate minutiae have associated
identity ID(K) increment the evidence for this ID by (10-K+1).
Thus, the top most candidate's ID gets a score of 10 and the 10th
best ID gets a score of 1. In block 212, the candidate identities
based on total scores are listed.
[0065] The identification method of FIG. 7 has a complexity of O(mN
log N) where m is the number of minutiae in the query print and N
is the number of minutiae in the database. While this may seem very
complex, the binary features and simple Hamming distance metric
make it computationally very feasible.
[0066] Feature Reduction: Gabor responses are computed from every
minutiae region. In one illustrative embodiment, every 32.times.32
pixel minutiae window may have 9 large filter (low frequency)
positions and 25 small filter (high frequency) components. For each
filter response, four orientations exist. Each response has a real
and imaginary component. Thus, a 272-bit binary representation is
generated. A typical fingerprint has on average 50-80 minutiae.
This can result in a very large representation when the features
are accumulated over a large database of fingerprints.
[0067] As in block 205, the number of coefficients may be
investigated to reduce the number without loss of accuracy. In one
embodiment, the features are computed at two resolution levels. A
first experiment focused on reducing the levels to only one
resolution level. Note that the coefficients are distributed
unevenly in the two levels. Even in one level, we focused further
to reduce the features by limiting the focus it to peak response in
a particular orientation. The results of these experiments are
presented below in terms of cumulative match curves (CMCs) on a
public database. Thus, 72-bit, 54-bit and 18-bit representations
are provided in addition to the original 272-bit
representation.
[0068] Referring to FIG. 8, a system 400 creating a database of
fingerprint recognition (enrollment 401) and identification of
fingerprints (identification 410) is illustratively shown in
accordance with present principles. Portions 401 and 410 may be
separate systems, but are depicted together for simplicity. System
400 includes an input device 402 for enrollment. Input device 402
may include a fingerprint scanner or other device for inputting
fingerprints into a database 404. A processor module or server 406
may be employed along with database 404 to preprocess fingerprints
to perform known fingerprint enhancements and quantizations in
accordance with present principles as described for enrollment
hereinabove.
[0069] An identification module or system 410 includes a
fingerprint reader 412. Reader 412 may be located on any device
where security access is desirable. Alternately, fingerprint images
may be input electronically to module 410 or using input 402.
System 410 may be included on a device 412 such as a computer
system, an electronic device, a keypad or at an entranceway to a
location or anywhere else fingerprint recognition would be useful.
Module 410 may also be employed in a law enforcement capacity for
identifying fingerprints found at a crime scene, or to simply match
a fingerprint with an identity for other purposes.
[0070] Module 410 may include a processing device 414 for
processing the input fingerprint for comparison to the database
404. The processing device 414 includes an extraction module 416
configured to identify minutiae in the print image. Device 414
further includes a representation module 418 configured to
determine Gabor coefficients for each minutiae in the print (for
texture content, e.g., gray scale, for identification), and
quantize the Gabor expansion for comparison to prints
representations stored in memory or database 404. Device 414
further includes a comparison module 420 which may be included in
the database 404 to determine candidate matches.
[0071] In one embodiment, database 404 may be replaced by a memory
device where fewer prints are stored for comparison. In addition,
processing device 414 may be omitted if processing can be performed
by processor module 406. Optional distortions of the expansion
coefficients with a transformation may be applied in both the
enrollment and identification phases.
[0072] Experimental Evaluation: Recall that for an identification
query, the n-bit (n=272, 72, 54 or 18) representations of the
minutiae of the query fingerprint are computed. Each of these bit
strings is compared against all the bit strings associated with all
the minutiae of the enrolled fingerprint database. A candidate list
is constructed by ranking the enrolled fingerprint database based
on the cumulative score of all the sets often ranked identities
(e.g., fingerprints in the enrollment database).
[0073] Experiments have been conducted on the FVC2002 DB1 and DB2
databases. Both databases are obtained from 110 fingers with 8
impressions each. Thus, each database includes 880 fingerprint
images. The performance is measured using the Cumulative Match
Curve (CMC), focusing on the top 10 ranking identities. To test the
performance on a large dataset, all the fingerprints in DB1 and DB2
were combined in one database. This combined dataset hence includes
1760 fingerprints. The CMC of FIG. 9 plots the probability of
identification against a candidate list of size K. The plot
indicates the probability that the correct identity appears in
different sized candidate lists. The faster the CMC approaches 1,
say at K.sub.0, the better the identification algorithm is. This
means that the correct identity always appears in the candidate
list of size K.sub.0. However, since there are 8 impressions of
each finger enrolled in the database, the CMC needed to be modified
as follows.
[0074] Let n be the number of impressions of a finger in the
enrollment database, where n does not include the query print. Let
TOP-K be the candidate list of size K. Further, for a given
identity denoted as P, let n impressions of P in the database be
P.sub.1, . . . , P.sub.n. Then we define:
[0075] 1. A Loose CMC curve: A finger P appears in the TOP-K if and
only if there exists a j such that 1.ltoreq.j.ltoreq.n with
P.sub.j.epsilon.TOP-K. (The traditional CMC curve is a loose CMC
curve with n equal to 1.)
[0076] 2. A strict CMC curve: Given some Q, with
1.ltoreq.Q.ltoreq.min(n, K), an identity P appears in the TOP-K if
and only if at least Q impressions of P, P.sub.j1, . . . , P.sub.jq
appear in the top K.
[0077] Referring to FIGS. 9A-9C, the accuracy performance of the
presentation and the identification methods described above are
illustratively shown. FIG. 9A shows a loose CMC plot for the
identification method. FIG. 9B shows a strict CMC plot for the
identification method.
[0078] In FIG. 9C, 18-bit, 54-bit, 72-bit and 272-bit
representations are shown for the identification method in
accordance with the present principles. Note that the 72-bit and
the 54-bit representations are quite good considering that the
representation is a mere fraction of the original 272-bit
representation. However, the 18-bit representation has poorer
results with significant loss of accuracy.
[0079] Having described preferred embodiments of a system and
method for fingerprint representation using localized texture
features (which are intended to be illustrative and not limiting),
it is noted that modifications and variations can be made by
persons skilled in the art in light of the above teachings. It is
therefore to be understood that changes may be made in the
particular embodiments disclosed which are within the scope and
spirit of the invention as outlined by the appended claims. Having
thus described aspects of the invention, with the details and
particularity required by the patent laws, what is claimed and
desired protected by Letters Patent is set forth in the appended
claims.
* * * * *