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

Подождите немного. Документ загружается.

Ridge Enhancement Filtering

Ridge enhancement ﬁltering is another ﬁngerprint-

speciﬁc image processing technique. In the 1960s

and 1970s, most ﬁltering techniques were labeled con-

textual. They used ﬁltering masks similar to the ones

used to estimate orientation, either ﬁxed size or pre-

determined variable frequencies. However, it was difﬁ-

cult for these techniques to ﬂexibly adapt to very

narrow or very wide ridges and spacing [6–8]. In the

1980s, a more so phisticated method based on the two-

dimensional Fourier transform was proposed [9, 10].

In the 1990s and 2000s, Gabor ﬁltering and wavelet

ﬁltering were introduced [2, 5, 13].

Conceptually, ridge enhancement ﬁltering aims to

‘‘enhance’’ ridges by generating stripe patterns from

scratch using the previously estimated orientation

and frequency. Strong enhancement is effective for

low quality images but at the risk of destroying the

original ridge structure. The strength of ﬁltering thus

needs to be controlled adapti vely and depends on the

ﬁeld in which it is used: law enforcement and non-law

enforcement. In the former case, the original ridge

structure needs to be preserved as much as possible

in order to improve compatibility with the examiners’

deﬁnition of minutiae since it still relies on manual

processing such as latent minutia coding. This is im-

portant to improve latent-print matching accuracy,

especially for fragmental latent prints with few minu-

tiae. In order to match such latent prints, even unstable

minutiae need to be incorporated to increase chances

of hit. On the contrary, in the latter case, which is fully

automatic, neither the original ridge structure has to

be preserved nor is compatibility with the examiners’

deﬁnition critical.

With respect to minutia preserving ability, there

are two types of minutiae to be considered: stable min-

utiae and unstable minutiae. The stable minutia is a

minutia that is topologically isolated from other minu-

tiae with no chance of interfering with other minutiae.

Fingerprint Image Enhancement. Figure 6 Unstable and Stable Minutiae (NIST DB#27 076T). Notes: (b) The ridge image

(a) was enhanced by one of the popular algorithms [9] with a relatively weak enhancement parameter. (c) A skeleton

image and minutiae were automatically extracted from the image in (b). (d) The image in (b) was enhanced by one of the

popular algorithms [9] with a relatively strong enhancement parameter. (e) A skeleton image and minutiae were

automatically extracted from the image in (d).

480

Fingerprint Image Enhancement

The unstable minutia is a minutia that may either re-

main unchanged or completely disappear depending on

the physical conditions of its surroundings. Crossover

minutiae are a typical example of unstable minutiae. In

Fig. 6(a ) a stab le minu tia is presented in the yellow

circle and an unstable minutia in the red circle. Notice

how the different levels of enhancement affect the ex-

traction results. If the strength of the ridge enhancement

ﬁlter is relatively mild, the crossover ridge structure and

the corresponding minutiae are maintained (Fig. 6(c) ).

On the other hand, the crossover minutiae completely

disapp ear w hen a strong ﬁlter is applied (Fig . 6(e)).

Despite this drawback, however, it is still beneﬁcial to

adopt strong ﬁltering since it is capable of consistently

extracting stable minutiae even from poor quality

images as illustrated by the following example. The

image in Fig. 7(b) represents an ideal , man ually edited

minutia e of the image Fig . 7(a) , containing a total of

76 minutiae, 55 of which are stable and 21 are unstable.

The image in Fig . 7(d) is an ar tiﬁcially produced poor

quality image by covering it with several circular

‘‘patches.’’ When a strong ﬁlter is applied to the images

in Figs. 7(a) and (d) , most of the 55 stable minutia e are

correctly extr acted as shown in Figs. 7(c) and (e) ,

respectively. It should also be noted that this method

is especially effective when the area of the overlapping

region between the two images is large enough in

which a sufﬁcient number of stable minutiae exist.

Thus, ﬁltering strength depends on the operational

strategy, requirements, and target image characteristics.

Once ﬁngerprint ridges are suitably enhanced,

▶ ﬁngerprint binarization is then conducted to pro-

duce a black and white image and, ﬁnally,

▶ ﬁngerprint

skeletonization to generate a skeleton image.

Fingerprint Image Enhancement. Figure 7 Effects of Strong Ridge Enhancement Filter (NIST DB#27 076T).

Notes: (b) The ideal skeleton image was manually generated so that skeleton curves correctly coincided with the

original ridges. Then, minutiae were extracted from the ideal skeleton image. (c) The ridge image in (a) was

automatically enhanced by one of the popular algorithms [9] with a relatively strong enhancement parameter. Then, a

skeleton image and minutiae were automatically extracted from that enhanced ridge image. (e) The ridge image in

(d) was automatically enhanced by one of the popular algorithms [9] with a relatively strong enhancement parameter.

Then, a skeleton image and minutiae were automatically extracted from that enhanced ridge image.

Fingerprint Image Enhancement

481

Summary and Future Improve ment

Fingerprint image enhancement is a very effective tool

for improving ridge clarity. Undoubtedly, improve-

ment in matching accuracy reported in the past two

to three decades can be attributed to innovation in

image enhancement techniques. Unfortunately, it is

far from true if considered in terms of how close the

automated ﬁngerprint recognition got to the ability of

human perception. This is because the current techni-

ques that heavily rely on ridge orientation and fre-

quency (and whatever informati on one can think of)

are not capable of perceiving a ﬁngerprint image as a

ﬁngerprint but just a collection of gray-scale pixels,

and the circumstance has not changed in the course

of over 40 years of research. This may change in the

future if a leap forward in the computational neurosci-

ence reveals the mechanism of human pattern recogni-

tion, but for the time being, a goal pro tempore is

probably to ﬁnd a way to extract information from

unmodiﬁed gray images to avoid side effects of the

image enhancement as far as possible.

Acknowledgments

The sample images are courtesy of the NIST and the

University of Bologna [14, 15]. The author would

like to thank Amane Yoshida for proofreading and

rewriting in English. For more information on the

technical details, refer to ‘Fingerprint Analysis and

Representation’ in Handbook of Fin gerprint Recognition

by Maltoni Et al. [2].)

Related Entries

▶ Fingerprint Classiﬁcation

▶ Fingerprint Features

▶ Fingerprint Image Quality

References

1. ANSI/NIST-ITL 1-2007 Fingerprint Standard – http://ﬁnger

print.nist.gov/standard/index.html

2. Maltoni, D. et al.: Handbook of ﬁngerprint recognition. Springer

(2003)

3. Xia, X., O’Gorman, L.: Innovation in ﬁngerprint capture

devices. Pattern Recognit. 36, 361–369 (2003)

4. Hara, M.: Image Density Conversion Method, Image Enhan-

cement Processor, and Program Thereof (USP 20080050030A1

– Pending)

5. Hong, L et al.: Fingerprint image enhancement: Algorithm

and performance evaluation. IEEE Trans. Pattern Anal. Mach.

Intell. 20, 777–789(1998)

6. Stock, R.: Automatic ﬁngerprint reading. In: The 1972

Carnahan Conference, on Electronic Crime Countermeasures,

April 19–21, 1972

7. Asai, K. et al.: Automatic ﬁngerprint identiﬁcation. SPIE

vol. 182, Imaging Application for Automated Industrial Inspec-

tion & Assembly (1979).

8. Capello, R. et al.: Method and apparatus for contextual data

enhancement (USP 4,876,726)

9. Kamei, T. et al.: Image ﬁlter design for ﬁngerprint enhancement.

In:Proceedings International Symposium on Computer Vision,

109–114 (1995).

10. Chikkerur, S. et al.: Fingerprint enhancement using STFT

analysis. Pattern Recognit. 40, 198–211 (2007), 109–114 (1995)

11. Funada, et al.: System and method for processing ﬁngerprint/

palmprint image (USP 7,027,626)

12. Hara, M.: System for recognizing ﬁngerprint image, method

and program for the same (USP 20070036401A1 – Pending)

13. Paul, A. et al.: A study of image enhancement techniques for

ﬁngerprint identiﬁcation. Proceedings of the IEEE Interna-

tional Conference on Video and Signal Based Surveillance

(AVSS’06) (2006)

14. NIST (National Institute of Standard and Technologies) Special

Database #27 – http://www.nist.gov/srd/nistsd27.htm

15. FVC2002 Second Fingerprint Veriﬁcation Competition Database

– http://bias.csr.unibo.it/fvc2002/

Fingerprint Image Quality

ELHAM TABASSI,PATRICK GROT HE R

National Institute of Standards and Technology,

MD, USA

Synonym

Expected performance or utility of ﬁngerprint image in

an automated comparison environment

Definition

The intrinsic characteristic of a biometric signal may

be used to determine its suitability for further proces-

sing by the biometric system or assess its conformance

to preestablished standards. The quality of a biometric

482

Fingerprint Image Quality

signal is a numerical value (or a vector) that measures

this intrinsic attribute. Quality score is a quantitative

expression of the utility, or predicted performance of a

biometric sample in a comparison environment. This

means that ﬁnger image quality scores should correlate

to the observed false match and

▶ false non-match

rates of the samples.

Introduction

With an increase in the need for reliable identity

authentication, biometric recognition systems have

been increasingly deployed in several different applica-

tions: government applications such as national ID

card, border control; and commercial applications,

such as physical access control, e-commerce, or mobile

phone. Among all biometric modali ties, ﬁngerprint

recognition is the most widespread due to its perma-

nence and uniqueness [1].

A ﬁngerprint is a pattern of friction ridges on the

surface of a ﬁngertip. A good quality ﬁngerprint has

distinguishable patterns and features that allow the

extraction of features, which are useful for subsequent

matching of ﬁngerprint pairs. This viewpoint may be

distinct from the human conception of quality. If, for

example, an observer sees a ﬁngerprint with clear ridges,

low noise, and good contrast then he or she might

reasonably say it is of good quality. However, if the

image contains few minutiae points then a minutiae-

based matcher would underperform. Thus, in the context

of automated matching, the term quality should not

be used to refer to the ﬁdelity of the sample, but instead

to the utility of the sample to an automated system.

Figure 1 shows examples of good and poor quality

ﬁngerprint image s.

Automatically and consistently determining the

quality of a given biometric sample for identiﬁcation

and/or veriﬁcation is a problem with far-reaching

ramiﬁcations. If one can identify low qualit y biometric

samples, this information can be used to improve the

acquisition of new data. This same quality measure can

be used to selectively improve an archival biometric

database by replacing poor quality biometric samples

with better quality samples. Weights for multimodal

biometric fusion can be sele cted to allow better quality

biometric samples to dominate the fusion. All of these

applications require that the quality of the biometric

sample be determined prior to identiﬁcation or veriﬁ-

cation. Most of these applications also require that

quality of the biometric sample be computed in real-

time during data acquisition.

Fingerprint Image Quality

Performance of an automated ﬁngerprint recognition

system is greatly affected by the degree of imperfection

present in the ﬁnger image. Accuracy of current ﬁnger-

print recognition systems is high when high-quality

samples are being compared [2] (Note that according

to Minutia Interoperability Exchange Test 2004

(MINEX04) report, best single ﬁnger proprietary ﬁn-

gerprint recognition system performed at 0.0047 false

non-match rate at 1% false match rate.). However,

performance degrades substantially as quality drops.

Fingerprint Image Quality. Figure 1 Good quality fingerprint images (a) have clear pattern of ridge and valleys;

however, poor quality ﬁngerprint images (b) do not have easily distinguishable patterns. Poor quality images result in

spurious and missed features, thus degrading the performance of the overall system. Poor quality samples can be due to

distorted source like abraded skin (b), distortion in one or more steps of the process, e.g., capture (residual ﬁngerprints on

the platen in (c)) or compression, or low character source, the sample may subjectively be assessed as ’’good’’ quality, but

a matcher may not be able to match it to its mate (d).

Fingerprint Image Quality

483

Although only a small fraction of input data are of

poor-quality, the bulk of recognition errors can be

attributed to poor-quality samples.

Degradation in ﬁngerprint image quality reduces

the amount of identi ﬁable information in a ﬁnger-

print. Poor quality images cause spurious and missed

features which decrease the likelihood of a correct

veriﬁcation and/or identiﬁcation, while extremely

poor quality samples might be impossible to verify

and/or identify. The variation in performance for dif-

ferent quality levels is shown in Fig. 2. The ﬁve traces

of Detection Error Tradeoff (DET) curves correspond

to ﬁve different levels of quality as measured by NIST

Fingerprint Image Quality (NFIQ) [3, 4]. NFIQ is an

integer between 1 and 5 where 1 represents the highest

quality and 5 the lowest (unusable) quality.

Several factors affect the quality of ﬁngerprint images:

user’s skin condition, improper ﬁnger placement, scan-

ner limitation or imperfection, impurities on the scanner

surface and others. The cause of these imperfections can

be classiﬁed in four groups: (1) impairments in the

source of

▶ Biometric character istics: like scars, blisters,

skin conditions such as wet or dry, age, occupation,

etc.; (2) user behavior: such as improper ﬁnger place-

ment, e.g., rotating ﬁnger or placing only tip of a ﬁnger

which cause capturing insufﬁcient area of ﬁnger image;

(3) imaging: e.g., low contrast, distortion, sampling

error, insufﬁcient dynamic range, etc.; and (4) environ-

ment: such as temperature, humidity, or unclean

platen.

Fingerprint Image Quality Measures

It is w idely accepted that a statement of a biometric

sample’s quality should be related to its recognition

performance. That is, a quality measurement algo-

rithm takes a signal or image, x, and produces a scalar,

q ¼ Q(x), which is predictive of error rates associated

with the veriﬁcation or identiﬁcation of that sample.

This predictive value of qu ality measures may be im-

perfect but valuable nevertheless. It should be noted

that operat ionally the requirement for a scalar is not

necessary: a vector could be stored and could be used.

The fact that quality has historically been conceived of

as scalar is a widely manifested restriction [5].

International Standards Organization (ISO) has

recently established a biometric sample quality draft

standard [6], in which quality score of a biometric

sample is deﬁned as predicted performance of the

Fingerprint Image Quality. Figure 2 Quality ranked detection error trade-off characteristics. Five traces correspond

to five NFIQ levels. Fingerprint images with NFIQ¼1 (highest quality) cause lower recognition error than images with

NFIQ¼5 (lowest quality).

484

Fingerprint Image Quality

sample in a comparison environment. It considers

three components of quality: (1) character, which refers

to quality of inherent physical features of the source,

for example, a ﬁngerprint with a scar has low character;

(2) ﬁdelity, which is the degree to which a sample is an

accurate representation of its source, for example, dis-

tortion degrades ﬁdelity; and (3) utility, which refers to

contribution of a sample to the overall biometric rec-

ognition error rates and is related monotonically to the

performance of biometric matchers. Character and

ﬁdelity of a sample positively or negatively impact the

utility of the sample.

There are several ﬁngerprint analysis approaches

that gauge character and ﬁdelity of ﬁngerprint images.

These measures are then summarized into a scalar (or a

vector) quality score that is indicative of utility of the

sample. Broadly ﬁngerprint image analysis can be

divided into local and global analysis methods [7].

Fingerprint local structure constitutes the main texture-

like pattern of ridges and valleys within a local region

while valid global structure puts the ridges and valleys

into a smooth ﬂow for the entir e ﬁngerprint. The quality

of a ﬁngerprint image is determined by both its local and

global structures. Local feature analysis methods parti-

tion an image into nonoverlapping blocks and assign a

quality score to each block which indicates the amount of

useful information in that block for subsequent match-

ing. Final image quality score can be computed by com-

bining quality scor es of the blocks. Global feature analysis

examines continuity and uniformity of ridge–valley

structure of a ﬁngerprint image in a holistic manner

and computes a global measure of ﬁngerprint quality.

Global and local quality measures could be com-

bined to obtain ﬁnal quality score of a ﬁngerprint

image such that the overall quality score is a measure

of match ability of the sample in an automated match-

ing process, i.e., the derived quality score should be

related to the biometric error rates that is likely to be

realized when the sample is matched.

1. Local Analysis To locally analyze a ﬁngerprint

image, it is divided into grids of blocks (Fig. 3).

For each block, local features such as directional

ﬂow of ridges are computed which are then sum-

marized into a quality score representing quality of

the block. Each block should be large enough to

contain sufﬁcient ridge–valley information, at least

two ridges per block. For example, for a ﬁngerprint

with a resolution of 500 ppi, each block could be

32  32 pixels. An overview of existing local analy-

sis methods follows.

a. Orientation certainty ﬁeld: A ﬁngerprint image

within a small block generally consists of ridges

(dark pixels) separated with valley (light pixels)

lines along the same orientation. High-quality

blocks of a ﬁngerprint image contain consistent

ridge (or valley) orientation. Local angle infor-

mation in each block can be used to compute

local features. Lim et al. [8] computed energy

concentration along the dominant direction of

Fingerprint Image Quality. Figure 3 Local analysis consists of partitioning a fingerprint image into small blocks. Local

features such as orientation consistency or directional flow are extracted from each block. These features convey

information useful for comparison of the image and therefore indicate quality of the block.

Fingerprint Image Quality

485

ridges by computing the ratio between two eigen-

values of the covariance matrix of a block’s

gradient vector. It gives an indication of how

strong the energy is concentrated along the

ridge–valley orientation. Chen et al. [9] measured

orientation coherence in each block using gra-

dient of the gray level image.

b. Ridge–valley structure: Well-formed and clearly

visible ridges are essential to the reliable detec-

tion of ridge endings and bifurcations, also

known as minutia points. Ridges that are too

close or too far apart, or ridges that are unrea-

sonably thick or thin indicate that the ﬁnger

image may not have captured properly, due to,

e.g., pressing too hard or too soft (Fig. 4). Shen

et al. [10] applied Gabor ﬁlter to image sub-

blocks, to identify blocks with clear repetition of

ridge and valley pattern as good quality blocks.

c. Pixel intensity or Directional contrast: Region of

good quality exhibits high directional contrast,

which means that the ridges and the valleys are

well separate d with regard to gray values. High-

quality blocks will exhibit large variance in gray

levels while low-quality blocks will show small

variance. [11–13] assess quality of each block

based on its pixel intensity. Bolle et al. [14]

used ratio of directional area to other nondirec-

tional area as a quality measure.

d. Power Spectrum: Ridge and valley structure in a

high-quality block forms a periodic signal,

which can be approximated either by a square

wave or a sinusoidal wave with its frequency

lie in certain range. In frequency domain, a

square wave exhibits a dominant frequency

with sideband frequency components (sinc

function), and a sinusoidal wave consists of

one dominant frequency and minimum com-

ponents at other nondominant frequencies.

Therefore, existence of a dominant frequency

component plus its frequency are indicative of

high- quality blocks of ﬁngerprint image. Poor

quality blocks will not exhibit a dominant fre-

quency or it will be out of the normal range of

ridge frequency [12]. Hong et al. [15] modeled

the ridge and valley pattern as sine wave, and

computed the amplitude, frequency as well as

the variance of the sine wave to decide the

quality of the ﬁng erprint. Nill and Bouzas [16]

propose an objective image quality based on the

digital image power of normally acquires

scenes. Their system is designed to assess the

quality of digital images and can be applied to

ﬁngerprint as well.

2. Global Analysis A good qualit y ﬁngerprint exhibit

smooth changes in ridge orientation across the

entire ﬁngerprint image except when a core or

delta point occurs. Ratio of ridge to valley thickness

should also be fairly constant throughout the whole

image. [8] used local angle information in each

block to assess continuity in orientation ﬁeld be-

tween neighboring blocks and uniformity of ridge

to valley thickness ratio. Chen et al. [13] computed

a block’s absolute difference in the orientation

angle with its neighboring blocks as a measure of

Fingerprint Image Quality. Figure 4 Examples of (a) good, (b) thin, and (c) thick ridge structure. (b) and (c) pose

challenge to automated matching system and hence are of lower quality than (a).

486

Fingerprint Image Quality

smoothness of the change in orientation angles

among blocks. As mentioned earlier, the ridges of

a ﬁnger image can be locally approximated by one

sine wave with its frequency in a cer tain range.

A region of interest (ROI) of the spectrum is de-

ﬁned as an annular region with radius ranging

between the minimum and maximum typical

ridge frequency values. For a more robust ridge

structure (i.e., the better image quality) the energy

will be more concentrated within the ROI. [9]

measured the ener gy concentration in ring-shaped

regions of the ROI by employing bandpass ﬁlters to

extract the energy in each frequency band. Good

quality images will have the energy concentrated in

few bands while poor quality ﬁngerprints will have

a more diffused distribution.

3. Overall Fingerprint Image Quality: prediction of

performance: It is desirable to combine local and

global quality features into one scalar or a vector of

quality such that the overall ﬁngerprint image qual-

ity is related to the expected false match and false

non-match of the image. The summarization can

simply be the percentage of blocks classiﬁed as

‘‘good’’ or ‘‘bad’’ quality after a local analysis, or

more elaborate combination methods such as

weighted average of local qualities. For example,

higher weights could be assigned to blocks closer

to the centroid of a ﬁngerprint since features

extracted from blocks near the centroid have

more useful and reliable information [9, 11]. Use

of a classiﬁer to nonlinearly combine local and

global features was ﬁrst proposed by Tabassi et al.

[3, 4]. The method called NIST Fingerprint Image

Quality NFIQ [3, 4] was developed to predict how

far a genuine score would lie from its impostor

distribution and is thus effective at improving

false rejections while suppressing false acceptance

errors.

NFIQ extrac ts minutia, assigns a quality value

to each minutia point, and measures orientation

ﬁeld, pixel intensity, and directional map to com-

pute the following local and global features: num-

ber of foreground blocks, number of minutia,

number of minutia that have quality value better

than certain thresholds, percentage of foreground

blocks of excellent, good, fair, and poor quality.

A neural network was trained to classify the com-

puted feature vectors into ﬁve levels 1–5 where

NFIQ = 1 is the best quality and NFIQ = 5 is the

lowest quality. Figure 2 shows that the highest

recognition performance is achieved for the best

quality samples (NFIQ¼1), and samples with low-

est quality (NFIQ¼5) have the lowest perfor-

mance. The plots of Fig. 5 show, respectively, the

genuine and impostor distributions for NFIQ

values 1 (excellent quality), 3 (average quality),

and 5 (poor quality). The overlapping of genuine

and impostor for the poorest NFIQ (i.e., NFIQ = 5)

means higher recognition errors for that NFIQ

level while the almost complete separation of the

two distributions for the best quality samples (i.e.,

NFIQ = 1) indicates lower recognition error. Source

code for

NFIQ algorithm can be found in [17].

Applications of Biometric Quality

Values

This section describes the roles of a sample quality

measure in the various contexts of biometric opera-

tions. The quality value here is simply a scalar sum-

mary of a sample that is taken to be some indicator of

matchabilit y. These uses of biometric sample quality

are not ﬁngerprint speciﬁc and can be generalized to

other modalities like face or iris.

1. Enrollment Phase Quality Assessment Enrollment is

usually a supervised process, and it is common to

improve the quality of the ﬁnal stored sample by

acquiring as many samples as are needed to satisfy

either an automatic quality measurement algorithm,

a human inspector (a kind of quality algorithm), or a

matching criterion (by comparison with a second

sample acquired during the same session). Our

focus on automated systems’ needs is warranted

regardless of analyses of these other methods, but

the authors do contend that naive human judg-

ment will only be as predictive of a matcher’s

performance as the human visual system is similar

to the matching system’s internals, and it is not

evident that human and computer matching are

functionally comparable.

Speciﬁcally, human inspectors may underesti-

mate performance on overtly marginal samples.

Certainly human inspectors’ judgment may be im-

proved if adequate training on the failure modes

and sensitivities of the matcher is given to the

Fingerprint Image Quality

487

inspector, but this is often prohibitively expensive

or time consuming and not scalable. Immediate

matching also might not be predictive of perfor-

mance over time because same-session samples

usually produce unrealistically high match scores.

For instance, Fig. 6 shows an example of two same-

session ﬁngerprint images that were matched suc-

cessfully by three commercial vendors despite their

obvious poor quality.

In any case, by viewing sample acquisition as a

measurement and control problem in which the

control loop is closed on the quality measure, a

system gains a powerful means of improving over-

all sample quality.

2. Quality Assurance Finger image quality assessment

algorithms may be used to monitor quality across

multiple sites or over time. This is useful to signal

possible performance problems ahead of some

subsequent matching operation. Quality values

may be aggregated and compared with some his-

torical or geographic baselines. Use of quality

values in this role has been documented in [18].

The National Institute of Standards and Tech nol-

ogy (NIST) has published a technical guidance

toward quality summarization [19]. Quality sum-

marization addresses the important issue of enter-

prise quality-assurance surveying by providing

tools on how to combine quality scores of individ-

ual samples into one scalar representing quality of

the whole database. Such a function would support

identiﬁcation of, e.g., defective sensors, underper-

forming sites, and seasonal or secular trends.

3. Ver iﬁcation Quality Assessment During a veriﬁca-

tion transaction, quality can be improved by clos-

ing an acquire–reacquire loop on either a match-

score from comparison of new and enrollment

samples or on a quality value generated without

matching. Indeed it is common to implement an

Fingerprint Image Quality. Figure 5 Probability density of impostor scores is shown in blue and probability density

of genuine scores is shown in gray. There is a higher degree of separation between the genuine and impostor

distribution for better quality samples as measured by NFIQ. (a) Best. (b) Middle. (c) Worst.

488

Fingerprint Image Quality

‘‘up to three attempts’’ policy in which a positive

match is a de facto statement that the sample was of

good quality – even if the individual happens to be

an impostor. Depending on the relative computa-

tional expenses of sample matching, reacquisition,

and quality measurement, the immediate use of a

matcher may not be the best solution. The key

difference here (as compared with the enrollment-

phase) is that quality values of bot h the enrollment

and veriﬁcation samples can be used to predict

performance. This two-dimensional problem is

distinct from the enrollment case where only one

quality value is used.

4. Identiﬁcation Quality Assessment Quality measure-

ment in identiﬁcation systems is important for at

least three reasons. First, many users often do not

have an associated enrollment sample. So a one-to-

many match will be an inefﬁcient and inconclusive

method of stating whether the authentication sample

had high quality. Second, in negative identiﬁcation

systems where users with an enrolled sample are

motivated to evade detection, quality measurement

can be used to detect and prevent submission of

samples likely to perform poorly [20], which may

help prevent attempts at spooﬁng or defeating de-

tection. Third, identiﬁcation is a difﬁcult task: it is

imperative to minimize both the false non-match

rate (FNMR) and the false match rate (FMR). To

the exte nt that consistently high-quality samples

will produce high genuine scores, a high matching

threshold can be used and this will collaterally

reduce FMR. But in large po pulations FMR

becomes dominant, and this raises the question:

can a quality apparatus be trained to be directly

predictive of false match likelihood?

5. Differential Processing Quality measurement algo-

rithms can be used toalter the subsequent pro-

cessing of a sample. Such conditional activity

arecategorized as follows.

a. Pre-processing Phase

An identiﬁcation system might apply image res-

toration algorithms or invoke different

▶ fea-

ture extraction algorithms for samples with

some discernible quality problem.

b. Matching Phase

Certain systems may invoke a slower but more

powerful matching algorithm when low-quality

samples are compared.

c. Decision Phase

The logic that renders acceptance or rejection

decisions may depend on the measured quality

of the original samples. This might involve

changing a veriﬁcation system’s operating thresh-

old for poor quality samples. For example, in

multi-modal biometrics, the relative qualities of

samples of the separate modes may be used to

augment a fusion process [21, 22].

d. Sample Replacement

To negate the effects of template aging, a quality

measurement may be used to determine whether

Fingerprint Image Quality. Figure 6 Example of same session captures of single finger that despite their poor quality

(NFIQ¼5) were matched correctly by three leading commercial matchers.

Fingerprint Image Quality

489