Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

Alternatively, a large number of candidate minutiae

can ﬁrst be obtained, e.g., by detecting lack of linear

symmetries during the direction ﬁeld estimation, then

a gray-value model of the minutiae, e.g., the parabolic

appearance of terminations and bifurcations, can be

enforced the candidates to retain the true minutiae

[19], Fig. 4.

Minutia Direction When matching or registering

two ﬁngerprints the

▶ minutia direction is a valuable

discriminative information. The minutiae directions

can be either extracted from the direction ﬁeld directly

or from the direction of the binarized and thinned

ridges, corresponding to minutiae locations, Fig. 2.

The directions along with the type information (ter-

mination or bifurcation) are attached to minutiae

coordinates.

Spatial Frequency Another descriptive feature which

can be attached to minutiae positions is the spatial

frequency information in the vicinity of minutiae. The

spatial frequency is usually deﬁned in terms of a direc-

tion in ﬁngerprints and has different implementations

[20]. One implementation is to use the average

frequency of the ridge or the count of ridges in a

ﬁxed line segment orthogonal to the minutia direction.

Another implementation of the frequency measure is

to count ridges or the average frequency along the line

joining a pair of minutiae. Because pairs as well as

triplet constallations of minutiae are commonly used

in ﬁngerprint matching, the frequency measures are

attached as a descriptive feature to the corresponding,

pairs or triplets.

Singular Points

Singular points are landmarks that are deﬁned in large

image patches (1–5 mm) compared to the size of

minutiae. There are typically 1–2 singular points in a

ﬁngerprint though they may occasionally be missing or

may be difﬁcult to identify in a ﬁngerprint. Three basic

types can be discerned, loop (also known as core),

whorl, and delta.

Amajoruseofthemistoclassifyaﬁngerprint

typically into one of the six categories, (Left-loop,

Right-Loop, Double-Loop, Arch, Tented-Arch, Whorl)

which are different constellations of loops and deltas,

Fig. 1. Such rough categorizations are employed to

match, and to organize massive amounts of ﬁngerprits

data efﬁciently.

Loops can provide a unique intrinsic global orien-

tation and position for a ﬁngerprint, allowing an orien-

tation and translation normalization of the ﬁngerprint

only on the basis of itself. Most whorls and deltas can

provide a direction too, though these are in general not

unique. Two singular points in the same image provide

always a unique

▶ intrinsic direction of ﬁngerprint.

This normalization is a practical alternative to registra-

tion by minutiae or can be complementary. Every

Fingerprint Features. Figure 4 Illustration of a use of direction field. From the left, the original image, the direction

vector ﬁeld color coded, the original superimposed with minutiae locations and the loop singularity [19, 24] are shown,

respectively. On the far right the Fingercode grid placed, on the loop singularity, is shown [10]. The direction ﬁeld

image (second) represents the complex quantities I

(7), where the argument of I

is mapped to the Hue (color)of

the HSV color model (same color indicates common direction) and the magnitude representing the quality of the

direction ﬁt is mapped to the Value (intensity).

470

Fingerprint Features

ﬁngerprint (the query, as well as every ﬁngerprint in

the database) is rotated and translated such that a

reference point and a half-line that is well deﬁned w.r.

t. a singular point of the ﬁngerprint become the origin

and the positive x-axis. Two translation and rotation

normalized ﬁngerprints are then more efﬁciently

matched – with minutiae or other features, because

no rotation or translation compensation speciﬁc to the

considered pair will be necessary.

Finally, singular points can function as anchors to

extract other descriptive features, e.g., the spatial fre-

quency. One can count the ridges along a line joining

two singular points, or along a line joining a minutia

and a singular point, etc. The spatial frequency estima-

tion issue is analogous to the one that has been dis-

cussed in conjunction with minu tiae.

Singularities by Poincare

Index

One of the oldest singular point detection techniques

used in ﬁngerprint processing is the Poincare

index

[21]. The index is deﬁned for a path in a vector ﬁeld

and represents the total amount of angle change of the

vectors along the curve. Assuming that the curve is

closed and it is in the gradient ﬁeld of a ﬁngerprint

then the Poincare

index, P, is given by

P ¼

dx þ

dy ¼

@x@y



@y@x

Þdxdy

ð11Þ

where the function y(x, y) represents the argument

(angle) of the gradient vectors and the last expression

is obtained by Green’s Lemma. It is worth noting that

even though the original ﬁngerprint image is assumed

differentiable (continuous) the gradient angle is not

continuous, (p and p) though its partial derivatives

are. By laying the closed curve around a loop, a whorl,

a regular (non-singular) point, and a delta, it can be

concluded that P will assume2p,p, 0 and p

radians, respectively. In Fig. 5 stylistic models of such

ﬁngerprint patch es are shown along with segments of

iso-curves (to which the gradients are orthogonal but

are not shown for convenience). When one walks the

dashed circle in full, the direction of iso-curves, and

thereby the gradient angles change with the Poincare

index. This obser vation is used, t ypically along with

the curve integral of Eq. (11), to determine if a candi-

date point is a whorl, loop, regular, or delta ty pe. It is

also possible to compute P according to the right hand

side of the equation, by a double integral applied to the

interior patch of the curve. By using the directions of

linear-symmetry vector ﬁeld, as opposed to those

of the gradient ﬁeld and the double integral [22]

Fingerprint Features. Figure 5 The top row shows the harmonic functions that generate the iso-curves of the patterns in

the second row. The iso-curves (their linearized examples are shown as red line segments) are given by a weighted

combination of the real and the imaginary parts of the respective harmonic functions, with a certain ratio between the

weights, deﬁning the direction parameter ’ of each pattern [23]. The third row shows the ﬁlters, n where g

is an

ordinary Gaussian (4), to detect the singularity points and ’ by a (complex) convolution applied to the direction ﬁeld (12).

The third row shows the symmetry order of the ﬁlters. The last row shows the Poincare

index of gradients.

Fingerprint Features

471

suggested an alternative way of computing P. In this

case the angles of the used vector ﬁeld are continuous

from the beginning so that no special care needs to be

taken to achieve continuity at angles around p andp.

The resulting P must be divided by 2 to correspond to

the gradient based Poincare

index.

Singularities by the Generalized Structure

Tensor

A singular point can also be detected by use of

the Generalized Structure Tensor (GST), which is

an extension of the structure tensor to curvilinear

(harmonic) coordinates [9, 23]. The fundamental

idea is the same as that of the structure tensor – to

ﬁnd an (unknown) angle such that the patch remains

invariant to a small translation along the found angle

direction but in the curvilinear coordinates. It turns

out that in this model, a singularity can be detected

by complex ﬁltering of the direction ﬁelds, already in

the complex representation (7).

ðD

þiD



¼ðl

max

l

max

Þe

i2’

max

ð12Þ

Here n

is a ﬁlter specialized to detect a loop, a delta

or a whorl, Fig. 5. The magnitude of a ﬁlter response,

which is complex valued, encodes the likelihood

that a location represents a singularity exactly in the

same way as the ordinary structure tensor, but now

the coordinates are harmonic, representing a pattern

of a singularity, and the l

max

, and l

min

are the error

extrema due to translation in curvilinear trajectories

having a certain direction. Likewise, its argument

(angle) encodes the intrinsic orientation of the

singularity (for loops and deltas their global inclina-

tion, for whorls the amount of chirality). The singu-

larity ﬁlters can be implemented by derivatives of

Gaussians which are separable, making them 1D ﬁlters.

Because the complex feature space obtained from

such ﬁlter responses are continuous both in their

arguments and positions, scale-space ﬁltering , e.g.,

coarse-to-ﬁne reﬁnement, is possi ble [24]. That the

symmetry axes (intrinsic orientati on) are available

in the GST method is useful, because the obtained

angle information can be used as a descriptive feature

attached to the singular point coordinates, much like

the use of minutiae orientations in ﬁngerprint match-

ing. Additionally, loop orientations alone allow a

normalization/registration of a ﬁngerprint pair even

if other singular points lack, and no minutiae are

available.

Singularities by Other Methods

The methods discussed earlier can ﬁnd singular points

by modeling direction variations on closed curves (in

practice a circle) or in regions containing a singularity.

Methods which do not use closed paths are exempliﬁed

as follows. Such a method to obtain singulariti es is the

early suggestion by [25] which models the direction

variations along the horizon tal scan lines. Information

deﬁning the location and the type of the singularity is

contained in the direction information around the

singular point and the horizontal lines contain only a

part of this. This information is instead injected into

the model in terms of orientation-change rules be-

tween scan lines. In [26], gradient vectors model half

a circle, like ‘‘n.’’ Then generalized Hough transform is

used to ﬁnd a peak, suggesting the location of a loop.

In contrast to GST and Poincare

index methods, the

(loop) inclination is assumed to be approximately

vertical, or a separate model is des igned for alternative

loop inclinations.

Summary

Descriptive features are used to match ﬁngerprints.

They include the locatio ns of minutiae points, and

the singular points. The location information can be

enhanced with additional descriptive measurements

including the local direction of the ridges and valleys

at minutiae locations, the intrins ic orientation of sin-

gular points, the type of the singular points, ridge

counts or average frequencies between minutiae as

well as singular points. To extract such descriptive

information direction maps are computed. Being tex-

ture measures, structure tensor representations of di-

rection maps can also be used as descriptive features on

their own if anchor points are available or in addition

to minutiae based features. Similarly, Gabor ﬁlters can

be used to obtain descriptive features if anchor points

are available. Commonly used anchors for registration

as well as descriptive features are the three basic singu-

larity types, loops, whorls, and deltas. They can be

detected and described independent of minutiae

information.

472

Fingerprint Features

Related Entries

▶ Fingerprint Enhancement

▶ Fingerprint Matching

▶ Fingerprint Quality

▶ Fingerprint Registration

▶ Fingerprint Sensing

References

1. Locard, A.: L’Identiﬁcation des Re

cidivistes. A. Maloine, Paris

(1909)

2. Bigun, J., Granlund, G.: Optimal orientation detection of linear

symmetry. In: First International Conference on Computer Vi-

sion, ICCV, London, June 8–11, pp. 433–438. IEEE Computer

Society, London (1987)

3. Kass, M., Witkin, A.: Analyzing oriented patterns. Comput.

Vision Graph. Image Process. 37, 362–385 (1987)

4. Bigun, J., Granlund, G., Wiklund, J.: Multidimensional orienta-

tion estimation with applications to texture analysis and optical

ﬂow. IEEE-PAMI 13(8), 775–790 (1991)

5. Granlund, G.: In search of a general picture processing operator.

Comput. Graph. Image Process 8(2), 155–173 (1978)

6. Ratha, N.K., Chen, S., Jain, A.K.: Adaptive ﬂow orientation-

based feature extraction in ﬁngerprint images. Pattern

Recogn. 28(11), 1657–1672 (1995). URL http://dx.doi.org/

10.1016/0031-3203(95)00039-3

7. Grother, P., Tabassi, E.: Performance of biometric quality mea-

sures. IEEE Trans. Pattern Anal. Mach. Intell. 29(4), 531–543

(2007). URL http://dx.doi.org/10.1109/TPAMI.2007.1019

8. Fronthaler, H., Kollreider, K., Bigun, J., Fierrez, J., Alonso-

Fernandez, F., Ortega-Garcia, J.: Fingerprint image quality esti-

mation and its application to multi-algorithm veriﬁcation. IEEE

Trans. Inform. Forens. Security 3(2): 331–338 (2008)

9. Bigun, J.: Vision with Direction. Springer, Heidelberg (2006)

10. Jain, A.K., Prabhakar, S., Hong, L., Pankanti, S.: Filterbank-

based ﬁngerprint matching. IEEE Trans. Image Process. 9(5),

846–859 (2000). URL http://dx.doi.org/10.1109/83.841531

11. Maio, D., Maltoni, D.: Direct gray-scale minutiae detection in

ﬁngerprints. IEEE Trans. Pattern Anal. Mach. Intell. 19(1),

27–40 (1997). URL http://www.computer.org/tpami/tp1997/

i0027abs.htm

12. Maltoni, D., Maio, D., Jain, A.K., Prabhakar, S.: Handbook of

Fingerprint Recognition. Springer, Berlin (2003). URL http://

bias.csr.unibo.it/maltoni/handbook/

13. Hong, L., Wand, Y., Jain, A.: Fingerprint image enhancement:

Algorithm and performance evaluation. IEEE-PAMI 20(8),

777–789 (1998)

14. Chen, Y., Dass, S.C., Jain, A.K.: Fingerprint quality indices for

predicting authentication performance. In: Audio- and Video-

Based Biometric Person Authentication, p. 160 (2005). URL

http://dx.doi.org/10.1007/11527923_17

15. Xiao, Q., Raafat, H.: Fingerprint image postprocessing:

A combined statistical and structural approach. Pattern

Recogn. 24 (10), 985–992 (1991). URL http://dx.doi.org/

10.1016/0031-3203(91)90095-M

16. Hung, D.C.D.: Enhancement and feature puriﬁcation of ﬁnger-

print images. Pattern Recogn. 26(11), 1661–1671 (1993). URL

http://dx.doi.org/10.1016/0031-3203(93)90021-N

17. Shih, F.Y., Pu, C.C.: A skeletonization algorithm by maxima

tracking on Euclidean distance transform. Pattern Recogn. 28

(3), 331–341 (1995)

18. Farina, A., Kovacs Vajna, Z.M., Leone, A.: Fingerprint minutiae

extraction from skeletonized binary images. Pattern Recogni-

tion 32(5), 877–889 (1999). URL http://www.sciencedirect.

com/science/article/B6V14-3WMK59F-D/2/bf21218ba 618c9f6

3efb1663ea24a6f6

19. Fronthaler, H., Kollreider, K., Bigun, J.: Local feature extraction

in ﬁngerprints by complex ﬁltering. In: S.Z.Li et al. (ed.) Inter-

national Workshop on Biometric Recognition Systems – IWBRS

2005, Beijing, Oct. 22–23, LNCS 3781, pp. 77–84. Springer,

Heidelberg (2005)

20. Maio, D., Maltoni, D.: Ridge-line density estimation in digital

images. In: International Conference on Pattern Recognition,

vol I, pp. 534–538 (1998). URL http://dx.doi.org/10.1109/

ICPR.1998.711198

21. Kawagoe, M., Tojo, A.: Fingerprint pattern classiﬁcation. Pattern

Recogn 17, 295–303 (1984)

22. Bazen, A., Gerez, S.: Systematic methods for the computation

of the directional ﬁelds and singular points of ﬁngerprints.

IEEE-PAMI 24 (7), 905–919 (2002)

23. Bigun, J., Bigun, T., Nilsson, K.: Recognition by symmetry deri-

vatives and the generalized structure tensor. IEEE-PAMI 26,

1590–1605 (2004)

24. Nilsson, K., Bigun, J.: Localization of corresponding points in

ﬁngerprints by complex ﬁltering. Pattern Recogn. Lett. 24,

2135–2144 (2003)

25. Wegstein, J.H.: An automated ﬁngerprint identiﬁcation system.

Tech. Rep. Special Publication 500-89, National Bureau of Stan-

dards, NBS (1982). URL http://www.itl.nist.gov/iad/894.03/ﬁng/

Special_Publication_500-89.pdf

26. Novikov, S., Kot, V.: Singular feature detection and classiﬁcation

of ﬁngerprints using Hough transform. In: E. Wenger, L. Dimi-

trov (eds.) Proc. of SPIE, vol. 3346, pp. 259–269 (1998)

27. Garcia, J.O., Aguilar, J.F., Simon, D., Gonzalez, J., Zanuy, M.F.,

Espinosa, V., Satue, A., Hernaez, I., Igarza, J.J., Vivaracho, C.,

Escudero, D., Moro, Q.I.: MCYT baseline corpus: a bimodal bio-

metric database. IEE Proc. Vision Image Signal Process. 150 (6),

395–401 (2003). URL http://ieeexplore.ieee.org:80/xpls/abs_all.

jsp?isNumber=2825%2&prod=JNL&arnumber=1263277&arSt=

+395&ared=+401&arNumber=1263277

Fingerprint Identification

▶ Fingerprint Indexing

▶ Fingerprint Recognition, Overv iew

Fingerprint Identification

473

Fingerprint Image Compression

▶ Fingerprint Compression

Fingerprint Image Enhancement

MASANORI HARA

NEC Corporation, Tokyo, Japan

Synonyms

Fingerprint contrast enhancement; Ridge enhance-

ment; Ridge extraction

Definition

Fingerprint image enhancement is the process of

applying techn iques to emphasize ﬁngerprint images

in order to facilitate the identiﬁcation of ridge valley

structures and hence their features.

Introduction

Computerized ﬁngerprint feature extractors more or

less require some sort of image pre-processing or en-

hancement to improve perceptibility. In doing so, they

need to contend with two major types of problems: one

is associated with image contrast such as insufﬁcient

dynamic range, and the other is associated with adverse

physical factors such as scars, blurs, creases, sweat

pores, and incipient ridges. Fingerprint image enhance-

ment aims to minimize the undesired effects caused

by such elements in order to extract a sufﬁcient

number of reliable features, namely, minutiae and

▶ ﬁngerprint singularities (cores and deltas). Broadly

speaking, ﬁngerprint image enhancement encom-

passes, but is not limited to, the intermediate steps

such as contrast enhancement, pore and incipient

ridge removal, ridge orientation and frequency estima-

tion, foreground segmentation, and ridge enhance-

ment ﬁltering.

The focus here is on the performance and limita-

tions of current image enhancement techniques rather

than on their algorithmic details. For this purpose,

many samples including problematic images and

their corresponding enhanced images are presented.

Fingerprint Image Digitalization and

Density

Fingerprint images are digitized through either inked-

print scanning or live scanning, most often with a

resolution of 500 dpi and a depth of 8 bits (i.e., 256

gray levels) in compliance with the NIST standard [1].

The gray level is a value associated with each pixel

representing its intensity or luminance. However, the

term density, the degree of ink thickness on the paper

surface, has been used throughout this section for the

sake of illustration. Thus, the higher the density, the

darker the ridges, and vice versa.

Ideally, the density of pores should be hi gher than

that of valleys and the density of incipient ridges

should be lower than that of true ridges. In fact, this

is precisely what some feature extractors traditionally

use to distinguish pores and incipient ridges from true

ridges and valleys. Although most inked-print scanned

images have continuous density, some live scanned

images exhibit a sparse nature. It has been reported

that the effective bit depth of some live scanners is only

2 or 3 bits [2, 3]. Obviously, such loss of information

makes the subsequent processes virtually impossible

to distinguish the key features.

Recognition of Fingerprint Ridge

Since ridge orientation and frequency characterize the

local ridge valley structure in the region of interest, the

problem of ﬁngerprint ridge recognition essentially

simpliﬁes to the task of estimating these two pieces of

information. Therefore, local ridge orientation and

frequency estimation play a key role in ﬁngerprint

ridge recognition.

Local orientation estimation, taking advantage of

the fact that ridge orientation does not change sudden-

ly when viewed locally, can ‘‘interpolate’’ ridge orien-

tation even in obscured regions. However, this is not

the case in frequency estimation; frequency can be-

come rather unstable when there is a sudden change

474

Fingerprint Image Compression

in ridge spacing even in a clear, well-deﬁned region, as

illustr ated in Fig . 1 . The ridge s in the region marked by

the red square (region A) in Fig . 1(a ) are nicely aligned

in the vertical direction, whereas frequency signiﬁcantly

changesdue tothe presenceof a spur,alternativelycalled a

whisker, in theregionmarkedby thered oval(region B)in

Fig. 1(b) . An average inter-ridge spacing is 8.7 pixels

(0.435 mm) in region A whereas it is only 5.7 pixels

(0.285 mm) in region B, which is narrower than the

neighboring region by 35%. For this image, some feature

extractions are able to correctly extract this narrow spur

as show n in Fig. 1(c) but some others fail (Fig . 1(d) ).

Even if the spur in region B is invisible as

shown in Fig . 1(e) , it is still easy to estimate ridge

orientation in the region with a high degree of certain-

ty. However, if this small region is contaminated with

noise, most feature extractions incorrectly estimate

the frequency in region B to be the same as that of

its neighbor, which results in a failure to detect the

spur (Fig . 1(f )). An ideal feature extractor should be

able to mark this region as ‘‘indeterminate’’ because it

is difﬁcult even for human examiners to identify the

spur conﬁdently.

Intermediate Steps in Fingerprint Image

Enhancement

A typical set of intermediate steps in ﬁngerprint image

enhancement includes:

1. Contrast enhancement or normalization.

2. Pore and incipient ridge removal.

3. Ridge orientation estimation.

4. Frequency estimation.

5. Foreground segmentation.

6. Ridge enhancement ﬁltering.

Contrast Enhancement

Whatever features or structures there may be, either

local or global, distinctiveness is important to appro-

priately separate one from the other. The conditions

that are preferably satisﬁed may include uniform

background density and a sufﬁciently w ide dynamic

range between ridges and valleys/background. If these

conditions are fulﬁlled, a simple stretching and/or

Fingerprint Image Enhancement. Figure 1 An Example of Unstable Frequency Estimation (NIST DB#27 002T).

Notes: (c) and (f) A skeleton image was extracted by one of the traditional algorithms [7] using the contextual ridge

enhancement ﬁlter with a narrow spacing, which corresponds to higher frequencies. (d) A skeleton image was

extracted by the same algorithm as in (c) and (f) with a wide spacing, which corresponds to lower frequencies.

Fingerprint Image Enhancement

475

thresholding should sufﬁce. In reality, however, more

elaborate and rigorous approaches are needed. Some

real issues related to dynamic range are outlined in the

following:

1. Uneven dynamic range.

A sample of uneven dynamic range is presented

in Fig . 2(a) . The ridges on the left (surrounded by

the red oval) are substantially lighter than the ones

on the rig ht.

2. Uneven valley density.

A sample of uneven valley density is shown in

Fig . 2(b) . This is a late nt imag e that is lifte d from

paper. The latent print is impressed on across the

regions where letters (‘‘O’’, ‘‘A’’, ‘‘N’’) are printed.

Here, the ridges and valleys cannot be recognized

easily because their actual density considerably

deviates from their original deﬁnition in which

‘‘the ridges are dark and the valleys are light’’; the

density of the valley on the letters is contrarily high

and local dynamic range is extremely narrow,

whereas the valley density in the plain region is low.

3. Noisy background.

The background containing leftover ﬁngerprint

images or stripe patterns resembling ﬁngerprints

makes it difﬁcult to isolate true ﬁngerprint pat-

terns. Problematic live scanned images and an

inked imag e are presented in Fig . 2(c, e) and 2(d) ,

respectively.

Contrast enhancement is a technique to accommo-

date such problems by expanding the dynamic range of

ridges and valleys. Adaptive histogram equalization is a

popular contrast enhancement technique. Other popu-

lar techniques include a simple linear contrast stretching

that uses the local minimum and maximum densities,

and a density normalization that uses the local density

mean and variance [2, 5]. Contrast-enhanced images of

the sample problematic imag es are prese nted in Fig . 2

) through (e

Although contrast enhancement is a powerful tool,

it has a drawback, i.e., it boosts background noise at

the same time since it cannot selectively enhance only

the targeted region unless some additional information

is given. As shown in Fig . 2(a

), (c

), (d

), and (e

), the

distinguishability of the foreground and background is

lower than the original.

However, it should be stressed that it is still

imperative to employ contrast enhancement when

dealing with po or quality images, mainly latent images

(Fig. 2(b

), for example). Once the ridge valley struc-

ture becomes visible, it essentially boils down to the

problem of identifying and analyzing ridge continuity.

Pore and Incipient Ridge Removal

Sweat pores are major obstacles in frequency estima-

tion. There are a variety of methods for removing pores

or at least for reducing their side effects. Some methods

do not remove pores but they remove false minutiae

that possibly originated from the pores. Other meth-

ods rely on the fact that pores are enclosed by darker

pixels as shown in Fig . 3(a) , and they can rem ove

ty pical pores but not problematic pores such as con-

tinuou s po res and swollen pores as show n in Fig . 3(b) .

Ridge structures such as lakes and spurs are easily

confused with pores, leading to miscalculation of

frequency if they are false ly ﬁlled in.

Incipient ridg es are another obstacle in frequency

estimation. The signiﬁcant incipient ridges as shown

in Fig . 3(c) can easily fool frequency estimati on

algorithms.

These two factors have not yet been fully explored,

and their distinguishability plays an impor tant role

in improving ﬁngerprint matching accuracy.

Ridge Orientation Estimation

Ridge orientation estimation is a ﬁngerprint-speciﬁc

image processing technique. A ridge orientation esti-

mation algorithm was developed for a FBI system

in the 1960s. In the 1960s and 1970s, many ridge

orientation estimation algo rithms set ‘‘slits’’ of prede-

termined orientations (8, 12, or 16 quantized orienta-

tions) and analyzed the density response [6–8]. The

orientation slit having a higher amount of density

change is indicative of the slit running perpendicular

to the direction of the ridge ﬂow. Similarly, the slit

with a lower amount of density change is indicative

of the slit running parallel to the direction of the

ridge ﬂow.

In the 1980s, more sophisticated methods were

introduced to extract ridge orientation such as a method

based on the gradient of two-dimensional vectors whose

components are derivatives of densities at horizontal and

vertical orientations [2, 5], and a method based on the

two-dimensional Fourier transform [9, 10].

476

Fingerprint Image Enhancement

In this process, the conﬁdence level of ridge orien-

tation is calculated. The difference in density ﬂuc-

tuation between the estimated orientation and its

orthogonal orientati on can be a base for conﬁdence,

and the power spectrum is another in the case of the

Fourier analysis.

Since all these techniques estimate ridge orienta-

tion locally, the inﬂuence of adverse factors such as

scars and smudges are not negligible and often lead

to wrong estimation. In order to correct such anomaly,

local orientation is examined for validity and re-

estimated from its neighbor. This process is called

Fingerprint Image Enhancement. Figure 2 Examples of Contrast-related Problematic Images. Notes: (a’) through (e’)

The images were contrast-enhanced with one of the local adaptive stretching methods specialized for ﬁngerprints [4].

Fingerprint Image Enhancement

477

ridge orientation smoothing, and several such techni-

ques have been proposed [2, 5, 9, 11].

Ridge orientation smoothing also has a drawback.

The orientation in the region where the ridge ﬂow is

not stable (e.g., in the proximity of the core and delta)

cannot be properly corrected due to its interpolative

nature. It also fails and propagates errors if the overall

estimation quality is low because it is based on the

assumption that the majority of orientations of neigh-

boring regions are indeed correct.

Examples of problematic images in orientation es-

timat ion are presented in Fig . 4 . Orien tations in the red

oval in Fig . 4(a

) and (b

) are incorrectly estimated

because of smudges and fragmented ridges.

One of suggested methods to improve estimation

accuracy is to use global pattern types and prior

knowledge of ridge ﬂow. Once the core and delta

have been extracted with high conﬁdence, the global

pattern shape can be estimated. This information can

help estimate and adjust local ridge orientation more

accurately.

Frequency Estimation

Frequency estimation is another ﬁngerprint-speciﬁc

image processing technique. Frequency is deﬁned as

the number of ridge s per unit length and is often

interchangeably referred to as the inverse of the inter-

ridge distance. It is far more difﬁcult to estimate than

orientation, and that explains why most feature extrac-

tions in the 1960s and 1970s did not fully exploit

this information.

In the 1980s, frequency analysis such as the

two-dimensional Fourier transform was proposed

to estimate frequency [9, 10]. Another techn ique

of frequency estimation analyzes peak intervals from

gray-level proﬁle orthogonal to the ridge orienta-

tion [2, 5].

In Fig . 5(a ) an examp le of a problematic image

with a sudden frequency change, denoted by the red

oval is presented. In Fig . 5(b ), the true frequency is

reﬂected via some manual correction, and Fig. 5(c)

shows an example of automatically estimated frequency

image using one of the latest algorithms [12 ]. Dark

density pixels correspond to the region where the

inter-ridge spacing is narrow, that is, frequency is

high. It can be observed that the frequency of the

area with very narrow inter-ridge spacing is falsely

estimated to be halved from its true value.

It is known that the presence of a minutia affects

the structure of its surroundings and hence the

corresponding local frequency. This often becomes a

problem in frequency estimation where a strong fre-

quency smoothing aimed to facilitate the estimati on

process can adversely eliminate true minutiae.

Foreground Segmentation

It is natural to conduct minutia extraction only in the

foreground region to minimize the possibility of

extracting false minutiae. Foreground segmentation

aims to distinguish the ﬁngerprint ridge region from

the background. Some methods rely heavily on the

conﬁdence of the ridge orientation to deﬁne the fore-

ground, whereas others rely on gray-level statistics as

well. As already explained, gray-level analysis is not an

ideal approach when dealing with very low qu ality

images such as the ones shown in Fig. 2.

Fingerprint Image Enhancement. Figure 3 Pores and Incipient Ridges.

478

Fingerprint Image Enhancement

Fingerprint Image Enhancement. Figure 4 Problematic Images in Ridge Orientation Estimation. Notes: (a’) and (b’)

A ridge orientation image was extracted by one of the traditional algorithms [7].

Fingerprint Image Enhancement. Figure 5 Problematic Images in Frequency Estimation (NIST DB#27 073T).

Notes: (b) True frequencies were calculated from an ideal skeleton image, which was manually generated so that

skeleton curves correctly coincided with the original ridges. (c) Frequencies were calculated from an automatically

extracted skeleton image using one of the recent algorithms.

Fingerprint Image Enhancement

479