Fingerprint Quality Assessment Combining Blind Image Quality,

Texture and Minutiae Features

Z. Yao, J. Le bars, C. Charrier and C. Rosenberger

Universite de Caen Basse Normandie, ENSICAEN, UMR 6072 GREYC, Caen, France

Keywords:

Fingerprint, Minutiae Template, Quality Assessment, Evaluation.

Abstract:

Biometric sample quality assessment approaches are generally designed in terms of utility property due to the

potential difference between human perception of quality and the biometric quality requirements for a recog-

nition system. This study proposes a utility based quality assessment method of ﬁngerprints by considering

several complementary aspects: 1) Image quality assessment without any reference which is consistent with

human conception of inspecting quality, 2) Textural features related to the ﬁngerprint image and 3) minutiae

features which correspond to the most used information for matching. The proposed quality metric is obtained

by a linear combination of these features and is validated with a reference metric using different approaches.

Experiments performed on several trial databases show the beneﬁt of the proposed ﬁngerprint quality metric.

1 INTRODUCTION

Fingerprint systems, among biometric modalities, are

the most deployed solution due to the invariability,

usability and user acceptance of ﬁngerprints (Jain

et al., 2004). So far, the application of ﬁngerprint is

no longer limited to traditional public security area

(ofﬁcial applications), but spread into the daily life,

smart phone authentication and e-payment, for in-

stance. Because of the continuous developments, ﬁn-

gerprint quality assessment has become a crucial task

in the deployment of systems in real applications.

There is no doubt that a good quality sample during

the enrollment process can reduce recognition errors.

The good quality of a ﬁngerprint sample is also ben-

eﬁcial to matching operations (Grother and Tabassi,

2007) in addition to the clarity of human intuition

and feature extractability of the image (Chen et al.,

2005). In this case, most previously proposed ﬁn-

gerprint quality approaches have been implemented

in terms of utility of biometric sample’s quality rather

than ﬁdelity (Alonso-Fernandez et al., 2007), i.e. bio-

metric sample’s quality should be related to system

performance. Tabassi et al. (Tabassi et al., 2004)

deﬁned their quality metric as a predictor of system

performance by considering the separation of gen-

uine matching scores (GMS) and impostor matching

scores (IMS). Chen et al. (Chen et al., 2005) later

proposed one quality metric by considering authenti-

cation performance.

As we can see in the literature, features are very

important to make a reliable judgment of the qual-

ity of a ﬁngerprint. Moreover, a ﬁngerprint can be

considered as an image or a set of minutiae we could

extract many features. This study proposes a qual-

ity metric of ﬁngerprint image based on the utility

property by considering two aspects in general: 1)

the ﬁngerprint image itself and 2) the corresponding

minutiae template which is rarely taken into account

for this issue. The validation of the proposed quality

metric is carried out by using two approaches based

on the prediction of authentication performance. The

main contribution of the paper is to propose a continu-

ous quality index of a ﬁngerprint integrating different

points of view (brought by the used features) and pro-

viding a better assessment.

This paper is organized as follows: Section 2 de-

tails the features for computing the proposed quality

metric. Section 3 presents the computation approach

of the proposed quality metric. Experimental results

are given in 4. Conclusion is given in section 5.

2 QUALITY FEATURES

The general purpose of this work is to qualify original

ﬁngerprint samples and to analyze the proposed qual-

ity metric through different validation approaches.

The proposed quality metric is based on a former

method in (El Abed et al., 2013). That work evalu-

336

Yao Z., Le Bars J., Charrier C. and Rosenberger C..

Fingerprint Quality Assessment Combining Blind Image Quality, Texture and Minutiae Features.

DOI: 10.5220/0005268403360343

In Proceedings of the 1st International Conference on Information Systems Security and Privacy (ICISSP-2015), pages 336-343

ISBN: 978-989-758-081-9

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

ated altered ﬁngerprint image quality with two kinds

of quality features, one is universal (no reference im-

age quality assessment) and another is related to the

ﬁngerprint modality. We employ this framework for

the original ﬁngerprint samples.

2.1 NR-IQA and Prior Features

In (El Abed et al., 2013), 11 features have been used

to obtain the quality metric, including one derived

from a NR-IQA algorithm (Saad et al., 2012) and the

others are image-based features. Details of the fea-

ture are not presented again in this paper. A general

description is given in table 1.

Table 1: List of quality features in (El Abed et al., 2013).

Feature Description

NO.

NR-IQA BLIINDS (Saad et al., 2012)

1-N1

SIFT point

number

Number of SIFT keypoints

2-S1

SIFT DC coef-

ﬁcient

DC coefﬁcient of SIFT features

3-S2

SIFT Mean Mean measure related to SIFT keypoints

4-S3

SIFT STD

Standard deviation related to SIFT key-

points

5-S4

Block number Number of blocks (17×17)

6-P1

Patch RMS

Mean

Mean of blocks RMS

values.

7-P2

Patch RMS

STD

Standard deviation of RMSs

8-P3

Patch RMS

Median

Median of blocks RMSs.

9-P4

Patch RMS

skewness

Skewness of blocks RMSs.

10-P5

Median LBP 256-level MBP histogram

11-P6

1. ’RMS’ is the abbreviation of Root Mean Square.

Salient features are extracted by using Scale In-

variant Feature Transform (SIFT) operator. For

patched features, it ﬁrstly divide images into blocks of

17×17, and then the root mean square (RMS) value of

each block is computed to obtain the quality features.

2.2 Texture-based Quality Features

Texture features are widely used for image classiﬁ-

cation and retrieval applications. There is not study

observed that whether some of them are able to con-

tribute distinctive results for quality assessment of ﬁn-

gerprint image. In this study, 11 texture features have

been selected as the components for generating the

proposed quality metric, cf. 2. These features have

been classiﬁed into four classes:

1) The ﬁrst class of textural features embeds local bi-

nary pattern (LBP) features and its extensions or

transforms. LBP features have been proposed by

Ojala et al (Ojala et al., 2002) for image classiﬁ-

cation. This feature is simple yet efﬁcient so that

it is widely used for texture analysis. The idea of

LBP operator was that the two-dimensional sur-

face textures can be described by two complemen-

tary measures: local spatial pattern and gray scale

contrast (Pietik

ainen, 2011). Basic LBP operator

generates a binary string by thresholding each 3-

by-3 neighborhood of every pixel of the image.

Table 2: List of texture features.

Feature Format

NO.

LBP 256-level LBP histogram vector

1-C1

Four-patch

LBP

Descriptor code vector

2-C1

Completed

LBP

512-bit 3D joint histogram vector

3-C1

GLCM mea-

sures

8-bit GLCM vector

4-C2

LBP H-FT LBP histogram FT

vector

5-C1

2S 16O

Gabor 64-bit Gabor response vector

6-C3

4S 16O Gabor 128-bit Gabor response vector

7-C3

8S 16O Gabor 256-bit Gabor response vector

8-C3

16S 16O Ga-

bor

512-bit Gabor response vector

9-C3

LRS 81-bit LRS motif histogram vector

10-C4

Median LBP 256-level MBP histogram

11-C1

1. ’S’, ’O’ and ’FT’: abbr. of scale, orientation and Fourier Transform.

The transforms of LBP involved in this study

include four-patch LBP (FLBP), completed

LBP (CLBP), LBP histogram Fourier transform

(LBPHFT) (Nanni et al., 2012) and median LBP

(MLBP) (Haﬁane et al., 2007).

2) Second class is Haralick feature or gray level

co-occurrence matrix (GLCM) (Haralick et al.,

1973). In this study, 4 statistic measures generated

from the GLCM matrix in 4 directions combina-

tion of neighbor pixels are computed, including

energy, entropy, moment and correlation.

3) The 2D Gabor decomposition is a sinusoidal func-

tion modulated by a Gaussian window. In this

case, the basis of a Gabor function is complete

but not orthogonal. In the last few decades, it has

been widely applied to ﬁngerprint image and other

biometric data to perform classiﬁcation and seg-

mentation tasks. Shen et al. (Shen et al., 2001)

proposed using Gabor response to evaluate ﬁnger-

print image quality, in which it is said that one

or several Gabor features of 8-direction Gabor re-

sponse are larger than that of the others. Olsen et

al. (Olsen et al., 2012) proposed a quality index

based on 4-direction Gabor response and it is said

that 4-direction is sufﬁcient to qualify ﬁngerprint.

However, in this study, it is observed that 2-scale

FingerprintQualityAssessmentCombiningBlindImageQuality,TextureandMinutiaeFeatures

337

4-direction Gabor ﬁlters do not bring out distinc-

tive regularity for ﬁngerprint images of a speciﬁed

database.

4) The last one concerns local relational string (LRS)

(Haﬁane and Zavidovique, 2006) which is an illu-

mination invariant operator and it reﬂects varia-

tion of local gray level of the image. The operator

is based on the local pixels relation in a speciﬁed

scale, and it uses 3 relations to generate local re-

lation motif histogram for measuring local spatial

variations of the image.

2.3 Minutiae-based Quality Features

Feng et al. (Feng and Jain, 2011) proposed to

reconstruct a ﬁngerprint image from the triplet

representation of minutia point. Such a result demon-

strates the signiﬁcance of minutiae template. In this

study, we relate the minutiae template to the quality

assessment of ﬁngerprint by deﬁning several quality

features based on minutiae number and DFT of the

three components of minutiae point, as shown in

table 3.

Table 3: Minutiae number-based measures related to ﬁnger-

print quality.

Measure Description

NO.

Minutiae number

(MN)

Minuitiae number of ﬁngerprint.

1-M1

Mean of minutiae

DFT

Deﬁned as equation (1b)

2-M1

STD of minutiae

DFT

Equation (1c)

3-M1

MN in ROI

MN in a rectangle region.

4-M1

MN in ROI 2 MN in a circle region.

5-M1

Region-based

RMS

Root mean square (RMS) value of

MN based on two blocks of the tem-

plate.

6-M1

Region-based me-

dian

Median value of MN obtained by di-

viding the template into 4 blocks.

7-M1

Block-based mea-

sure

A block-based score for the tem-

plate.

14-M1

1. region of interest.

Minutiae-based measures given in table 3 are cal-

culated based on a the template of detected minutiae

extracted by using NBIS tool (Watson et al., 2007).

This template contains a quadruple representation of

minutia point which consists of 1) the position (x, y)

of detected minutiae, 2) the orientation θ of detected

minutiae, and 3) a quality score of detected minutiae.

In the experiment, only the minutiae positions and ori-

entations are used for calculating these measures. In

the following, the details of some of the measures are

presented.

In the experiment, both measure 2 and 3 are de-

rived from the magnitude of the Fourier transform of

the linear combination of 3 minutia components after

eliminating DC component, as described in equation

F (x, y, θ) =

N−1

∑

n=0

·µ

+ y

·ν

+ θ·ω

. (1a)

where µ, ν, and ω are frequency samples. Measures

and M

are ﬁnally computed as follows:

= |F (x, y, θ)|, (1b)

∑

i=1

− M

). (1c)

DC component was eliminated when computing these

two measures because there is no valuable informa-

tion in this element.

For measure 4, the size of rectangle region is de-

termined by the maximum value of both x and y co-

ordinates of minutiae, for which there is no useful in-

formation outside the foreground of the ﬁngerprint in

this case. This choice also ensures that the region of

interest will not go over the effective area of minutiae.

An example of rectangle region is shown in ﬁgure 1

(a).

Figure 1: Example of circle region (a), rectangle region (b),

and template block partition in the size of ﬁngerprint (c).

The radius of the circle region for measure 5 is

also determined by the maximum and minimum loca-

tion value of minutiae along the horizontal direction

of ﬁngerprint, for minutiae lie around ﬁngerprint cen-

ter are said to be those who contribute most to ﬁnger-

print matching, i.e. they are considered as the most in-

formative. As the quadruple representation does not

provide information of ﬁngerprint core point, an es-

timated point was used as the center’s location of the

ﬁngerprint. A comparison has been made between the

estimated center point and a core point detected by an-

other approach, and it is found that the result does not

vary too much. The estimated center position was de-

termined by considering the maximum and minimum

ICISSP2015-1stInternationalConferenceonInformationSystemsSecurityandPrivacy

338

minutiae location as well. An example of the circular

region is shown in ﬁgure 1 (b).

For measures 6 and 7, the whole ﬁngerprint re-

gion is divided into 2 and 4 blocks, and minutiae num-

ber in each block is considered to generate a measure.

Another block-based measure is calculated by divid-

ing the whole ﬁngerprint region into several blocks in

the size of 64×64. The blocks are classiﬁed into 3

classes, reasonable block, vague block, and unreason-

able block. At last, a quality index is assigned to each

block in terms of the minutiae number in the block,

for which a threshold is used for determining the in-

dex of each block. An example of block partition is

shown by ﬁgure 1 (c). In addition, features proposed

in (Ross et al., 2005) are calculated in terms of minu-

tiae distribution and orientations, and they are rotation

and translation invariant.

We analyse in section 4.2 the behavior of these

quality features. Based on all these quality features,

we generate a quality metric using the method de-

scribed in the next section.

3 QUALITY METRIC

DEFINITION

The quality metric of ﬁngerprint (QMF) image in this

study is computed by an approach using a Genetic Al-

gorithm (GA) proposed in (El Abed et al., 2013). It

uses a weighted linear combination of the quality fea-

tures, formulated as

Q =

∑

i=1

, (2)

where N is the number of quality features F

(i =

1, ··· , N), α

are the weighted coefﬁcients. The

weighted coefﬁcients are computed via optimizing

a ﬁtness function of GA which is composed by the

Pearson correlation between combined quality results

deﬁned by equation 2 and corresponding GMS of ﬁn-

gerprints samples. This approach realizes a learn-

ing of quality assessment and optimizes the weighted

coefﬁcients to generate a continuous quality metric

combining different features.

4 EXPERIMENTAL RESULTS

In order to validate the behavior of the quality met-

ric (denoted as QMF) of this study, an analysis of the

proposed features is realized. The validation of QMF

is implemented by observing the evaluation results of

both QMF and NFIQ (Tabassi et al., 2004).

4.1 Protocol and Databases

In this study, three FVC databases (Maio et al., 2004)

have been used for experiments: FVC2002DB2A,

FVC2004DB1A, and FVC2004DB3A. The ﬁrst two

databases are established by optical sensor and the

last one is thermal sweeping sensor. The resolu-

tions and image dimensions of all 3 databases are

500dpi, 500dpi, 512dpi, and 296×560, 480×640, and

300×480, respectively. Each database involves 100

ﬁngertips, and 8 samples for each ﬁngertip. The intra-

class and inter-class matching scores involved in the

experiment have been calculated by using NBIS tool

(Watson et al., 2007) namely Bozorth3 and a commer-

cial SDK (IDt, ). Minutiae template used in the ex-

periment were also extracted by using the correspond-

ing MINDTCT and the SDK. The minutiae-based fea-

tures involve in only the location (x, y) and the orien-

tation o of minutia so that we don’t consider the qual-

ity value of the points (generated by MINDTCT) and

the type of them (given by SDK).

4.2 Feature Analysis

Fernandez et al. (Alonso-Fernandez et al., 2007)

and Olsen (Olsen et al., 2012) respectively calcu-

lated Pearson and Spearman correlation coefﬁcients

between different quality metrics to observe the

behaviour of them. We use the same approach in this

study by computing the Pearson correlation between

several quality metrics and the QMF. Quality metrics

used for this analysis include OCL (Lim et al., 2002),

orientation ﬂow (OF) (Chen et al., 2004), standard

deviation (STD) (Lee et al., 2005), Pet Hat’s wavelet

(PHCWT) (Nanni and Lumini, 2007) and NFIQ.

Figure 6, 8, 7 presents the correlation results obtained

from the trial databases for all quality features.

In table 6, highlighted columns (with yellow)

demonstrated a relatively stable correlation for all the

three databases, and some others marked with green

illustrated their feasibility for certain data sets. Ac-

cording to this observation, we could make an attempt

to reduce some redundant features in next study. Ta-

ble 8 presents only 11 of the minutiae-based features,

for the correlation of them is not very distinctive.

Some of them demonstrate good correlated behavior

with the quality metrics, but greatly vary among the

data sets and even not correlated with any of the qual-

ity metrics. Likewise, the correlation results of the

image-based features are given in table 7. We use all

these features to calculate the quality metric which

enables qualifying ﬁngerprint samples with comple-

mental information. Note that the last columns denote

FingerprintQualityAssessmentCombiningBlindImageQuality,TextureandMinutiaeFeatures

339

relatively good correlation between the NR-IQA and

the quality metrics.

4.3 Metric Validation

The validation involves in two sections, one is the im-

pact of enrollment selection (YAO et al., 2014) and

another is a utility evaluation (Chen et al., 2005).

4.3.1 Impact on the Enrollment Process

Authors in (Grother and Tabassi, 2007) discussed on

quality values used for three different cases, including

enrollment phase, veriﬁcation task and identiﬁcation.

Enrollment is generally a supervised task for getting

relatively good quality samples, and one main differ-

ence between veriﬁcation and identiﬁcation tasks is

the existence of enrollment which directly impacts on

how FNMR and FMR acts. However, if the purpose

is to validate a quality metric without considering the

testing type (i.e. algorithm testing, scenario testing

and etc.), the variation of enrollment samples quality

would generate distinctive impacts on matching per-

formance and the result is repeatable in the experi-

ments.

We computed the EER values of 3 databases by

choosing the best quality samples as the reference

(by using NFIQ and QMF). A good quality metric

for the choice of the references should reduce match-

ing error rate. The ROC curves and EER values of

FVC2004DB1A based on this strategy are presented

as an illustration, see ﬁgure 2.

Figure 2: Enrollment selection result of FVC2004DB1A.

The EER values by using NFIQ (for the enroll-

ment process) is 14.8%, and 14.1% with the QMF

metric. For FVC2002DB2A and FVC2004DB3A, the

EER values are 13.2% (NFIQ), 10.6% (QMF), 8.3%

(NFIQ) and 6.7% (QMF). These results show the ben-

eﬁt of QMF face to NFIQ as it permits to optimize the

enrollment process.

In addition, such EER values are calculated via

the commercial SDK, results obtained via NFIQ

are 3.99% (02DB2A), 9.39% (04DB1A) and 4.76%

(04DB3A), while the values obtained by using QMF

are 3.39% (02DB2A), 5.35% (04DB1A) and 4.64%

(04DB3A), respectively. This result demonstrates

whether the QMF is possible for dealing with inter-

operability. However, in practical, this property relies

on both the performance of matching algorithm and

the quality metric. We employ this result simply for

validating the QMF.

4.3.2 Quality and Performance Evaluation

The second approach is based on the isometric bins of

samples sorted in an ascending order the quality val-

ues and is more strict for the distribution of the quality

values. In order to validate the QMF by referring to

NFIQ, instead of dividing quality values of NFIQ into

5 isometrics bins, we divided them into 5 bins which

correspond to its 5 quality labels. The reason for do-

ing so is that NFIQ fails to satisfy the isometric-bin

evaluation criteria, as given in ﬁgure 3.

Figure 3: Example of 5-bin evaluation for NFIQ on

FVC2002DB2A.

Then, the EER values of the divided bins are cal-

culated. For NFIQ-based quality values, it is easier to

calculate the EER values of the 5 label bins, as it is

depicted in ﬁgure 4.

We are able to observe that the matching perfor-

mances on FVC2002DB2A and FVC2004DB3A are

monotonically increased by pruning bad quality sam-

ples gradually. NFIQ generated quality levels from 1

to 4 for FVC2002DB2A, and no samples of level 5

were ﬁgured out for this database. This might be due

to the minutiae points detected on the images of this

database, because NFIQ algorithm involves in minu-

tia quality of the ﬁngerprint image. This situation was

ICISSP2015-1stInternationalConferenceonInformationSystemsSecurityandPrivacy

340

Table 5: Inter-class Pearson correlation for textural features. 02DB2A (top), 04DB1A (middle) and 04DB3A (bottom).

OCL -0.6826 0.3002 -0.7037 -0.7895 -0.4462 -0.3294 0.3806 0.5864 0.6832 0.8699 -0.7593

OF -0.1938 0.1783 0.0098 -0.0396 -0.0452 -0.0700 0.1685 0.2016 0.1590 -0.0012 0.0593

PHC -0.6926 0.2864 -0.6665 -0.8805 -0.3391 -0.1957 0.3552 0.6329 0.7507 0.8476 -0.7807

STD -0.6230 0.3958 -0.5590 -0.8796 -0.3016 -0.3037 0.5620 0.8066 0.8940 0.7668 -0.7438

NFIQ 0.3919 0.1240 0.3483 0.4675 0.1617 -0.0057 0.0401 -0.0676 -0.1307 -0.4569 0.2731

OCL -0.6899 -0.7979 -0.7798 0.2582 0.7151 0.0456 0.4071 0.6708 0.7223 -0.7416 0.7125

OF -0.2642 -0.3263 -0.3057 0.1580 0.2073 0.1087 0.3968 0.4206 0.4539 -0.2281 0.2057

PHC -0.7060 -0.8206 -0.8416 0.2832 0.7535 0.0373 0.4722 0.7548 0.7964 -0.7701 0.7426

STD -0.5920 -0.7066 -0.7286 0.2249 0.6471 0.0554 0.4669 0.6930 0.7264 -0.6646 0.6297

NFIQ 0.1634 0.1607 0.1775 -0.0683 -0.2101 -0.0412 0.0897 -0.0254 -0.0157 0.2295 -0.2143

OCL -0.5001 -0.6394 -0.7460 0.0406 0.5144 0.0842 0.5301 0.6505 0.6948 -0.3814 0.5536

OF -0.2510 -0.1842 -0.1539 0.1097 0.0814 -0.2304 -0.1348 -0.1148 -0.0539 -0.1537 0.1566

PHC -0.1648 -0.2758 -0.4495 0.1015 0.1439 0.1660 0.6947 0.7992 0.7450 -0.1928 0.1726

STD -0.2401 -0.3447 -0.5029 0.0839 0.2221 0.1201 0.6398 0.7359 0.7037 -0.2161 0.2550

NFIQ -0.0532 -0.0886 0.0316 0.0183 0.0518 -0.2360 -0.3640 -0.4005 -0.2608 -0.0805 0.0907

Figure 4: Monotonic increasing matching performance val-

idation of FVC2002DB2A for NFIQ, calculated by dividing

quality values into 5 isometric bins (no sample of quality 5

for this database).

Figure 5: Monotonic increasing matching performance val-

idation of FVC2002DB2A for QMF, calculated by dividing

quality values into 5 isometric bins.

Table 4: 5 Bins EER values based on QMF and NFIQ of

FVC2004DB1A and FVC2004DB3A.

Bin No. B1 B2 B3 B4 B5

. (04DB1) 22.2% 16.6% 17.2% 17.8% 13.3%

. (04DB1) 15.8% 18.1% 17.7% 23.2% 26.5%

Q. (04DB3) 14.2% 8.9% 7.4% 5.8% 4.2%

N. (04DB3) 7.5% 8.1% 13.4% 12.9% 29.8%

1. ’Q’ and ’N’ are abbreviation of ’QMF’ and ’NFIQ’, respectively.

observed when calculated the correlation between

14 minutiae quality features and genuine matching

scores in the experiment of this study. It shows

a relatively higher correlation on FVC2002DB2A,

while the values of two other databases are relatively

lower. For FVC2004DB1, both the proposed qual-

ity metric and the reference algorithm showed cer-

tain difﬁculties. Here, only the graphical results on

FVC2002DB2A are presented, while the 5 bins’ EER

values based on proposed approach and NFIQ of

FVC2004DB1A and FVC2004DB3A are given in ta-

ble 4. The quality values of QMF are normalized into

[0, 100] on each database where small value denotes

bad quality (bin 1). The NFIQ has 5 quality levels

where level 1 represents the best quality and level 5 is

the worst one.

5 CONCLUSION

This study ﬁrst propose a ﬁngerprint quality metric by

considering image-based quality features and those

derived from minutiae template. Second, the quality

metric has been validated by using different valida-

tion approaches. In the study, the proposed quality

metric was evaluated on 3 different FVC databases,

FVC2002 DB2 A, FVC2004 DB1 A, and FVC2004

DB3 A. Among the validation result, it can be ob-

served that the performance of quality metric shows

FingerprintQualityAssessmentCombiningBlindImageQuality,TextureandMinutiaeFeatures

341

Table 6: Inter-class Pearson correlation for textural features. 02DB2A (top), 04DB1A (middle) and 04DB3A (bottom).