Yanushkevich S.N., Wang P.S.P., Gavrilova M.L., Srihari S.N. (eds.) Image Pattern Recognition. Synthesis and Analysis in Biometrics

Подождите немного. Документ загружается.

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

348 Synthesis and Analysis in Biometrics

metrics should be in place during enrollment, so that the template created

is a representation that can be classiﬁed by the biometric system at a later

time.

14.3.2. 3D Face Recognition

To further evaluate face recognition and the eﬀects illumination conditions

has on performance, the authors evaluated a three dimensional face

recognition algorithm in

[

24,25

]

. According to 3D face manufacturers the

technology has advantages over 2D face as it compares the 3D shape of the

face which is invariant in diﬀerent lighting conditions and pose, as opposed

to points, features, or measured distances that are evaluated in 2D face

recognition systems.

Fig. 14.5. Six light levels and directions used for the 3D face evaluation.

The evaluation consisted of:

(a) Four frontal illumination levels, one side light condition, and

(b) One behind the head light condition.

The six illumination conditions can be seen in Fig. 14.5. To eliminate

external eﬀects on the experiment and to emphasize the sole eﬀect of

lighting on the performance of the system, the subjects position, facial pose,

and face covering artifacts were deﬁned by the test protocol, which followed

ANSI INCITS Face Recognition Format for Data Interchange standard

[

]

. Each image reﬂects the speciﬁcations for captured images as shown

in Fig. 14.6. The resulting 3D model created from two 2D face images is

shown in Fig. 14.7.

Like the 2D face evaluation, the 3D evaluation consisted of 30

individuals, of which 75% were male. The hypotheses were set up to

establish whether there was a signiﬁcant diﬀerence in the performance

of the algorithm using the enrollment level illumination 3D template

and veriﬁcation attempts using all illumination levels shown in Fig. 14.5.

A statistical analysis revealed that there was no statistically signiﬁcant

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

Issues Involving the Human Biometric Sensor Interface 349

Fig. 14.6. INCITS face recognition data format image requirement [20].

Fig. 14.7. 3D face recognition model.

diﬀerence in the performance of the algorithm when the light level was

measured

(a) Between light level 1 (220–225 lux), and

(b) The other front illumination levels (320–325 lux; 650–655 lux; 1020–

1140 lux), side light, and behind light conditions.

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

350 Synthesis and Analysis in Biometrics

So the conclusions in

[

24,25

]

showed an increase performance over prior

2D tests, but encountered a tradeoﬀ of speed for accuracy, as an average

veriﬁcation (1:1) attempt took roughly 10 seconds, which was signiﬁcantly

higher than the 2D evaluation in

[

]

To summarize the eﬀects of environmental factors on biometric system

performance, the system designer must be:

(a) Cognizant of the application,

(b) Fit the correct biometric system into that environment.

the stored enrollment template is usable in the conditions where the

biometric system is implemented.

To continue the discussion on quality, the next section of this chapter

will focus on ﬁngerprint image quality across diﬀerent age groups.

14.4. Fingerprint Image Quality

Fingerprint image quality issues have been widely discussed in the

biometrics literature

[

44,5,21

]

. Variations in the quality and attributes of

the ﬁngerprints from the same ﬁnger is a common issue for ﬁngerprint

recognition systems

[

]

. The inconsistencies in biometric sample quality

lead to an overall degraded system performance. The usefulness of quality

metrics is acknowledged in the biometrics ﬁeld, but there is no standard that

provides guidelines for objective scoring of biometric samples. The current

standardization eﬀort refers to three diﬀerent connotations of quality:

(a) Character that represents inherent features of the source from which

the biometric sample is derived.

(b) Fidelity that represents how closely the samples from the same

individual match, and

overall performance of a system.

The utility of a sample is aﬀected by the character and ﬁdelity of the

sample, and performance of a system is aﬀected by the utility of a biometric

sample.

Non-uniform and irreproducible contact between the ﬁngerprint and

the platen of a ﬁngerprint sensor can result in an image with poor utility

quality. Non-uniform contact can result when the presented ﬁngerprint is

too dry, or too wet, as shown in Fig. 14.8. Irreproducible contact occurs

when the ﬁngerprint ridges are semi-permanently or permanently changed

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

Issues Involving the Human Biometric Sensor Interface 351

due to manual labor, injuries, disease, scars or other circumstances such as

loose skin

[

]

Fig. 14.8. Image of a low quality dry ﬁngerprint (left). Image of a normal good quality

ﬁngerprint (right). Both images were acquired using an optical sensor.

These contact issues can aﬀect the sample provided to the ﬁngerprint

sensor when an elderly user presents a ﬁngerprint to the ﬁngerprint device

[

]

. Due to eﬀect of ageing, the skin becomes drier, the skin sags from

loss of collagen, and the skin becomes thinner and loses fat as a direct

result of elastin ﬁbers. All of these decrease the ﬁrmness of the skin, which

aﬀects the ability of the sensor to capture a high quality image

[

]

.The

skin of elderly individuals is likely to have incurred some sort of damage

to the skin over life of the individual. Medical conditions like arthritis

aﬀect the ability of the user to interact with the ﬁngerprint sensor. The

data collection and signal processing sub-systems are directly aﬀected by

non-uniform contact, irreproducible contact and inability of the subject to

interact with the system.

Interaction between the user and the system. The problem of

interaction between the user and the system aﬀects the sub-category of

presentation within the data collection module. Irreproducible and non-

uniform contacts aﬀect the sub-categories of feature extraction within the

signal-processing module. Low quality images aﬀect the feature extraction

process, and results in extraction of inaccurate and spurious features. The

overall utility of the ﬁngerprint image is reduced because it aﬀects the

performance of the ﬁngerprint recognition system.

Example 15. A research study was performed to analyze ﬁngerprint image

quality from a younger group (18–25 years old) and elderly group (62 years

and above). The ﬁngerprint images were collected using an optical sensor.

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

352 Synthesis and Analysis in Biometrics

Fingerprints from 60 elderly subjects and 79 young subjects were collected.

The NIST Fingerprint Image Software 2 (NFIS2) was used for the analysis

of the ﬁngerprint images. The mindtct process was used to extract the

minutiae count for each ﬁngerprint, and the nﬁq process was used to

compute the quality score for each ﬁngerprint.

A comparison of average minutiae count for the younger group and

average minutiae count for the elderly group showed that there was a

diﬀerence between the two. Figures 14.9 and 14.10 shows the frequency

count of the minutiae count for the young and the elderly groups.

min_count_elderly

Frequency

160140120100806040

Mean 90.27

StDev 27.94

N 486

Histogram (with Normal Curve) of min_count_elderly

Fig. 14.9. Minutiae count for the elderly group.

The younger group had an average minutiae count of 54.79 and the

elderly group had an average minutiae count of 90.27. Previous research

has shown that good quality ﬁngerprints have between 40–100 minutiae

[

]

. The minutiae count analysis for ﬁngerprints from the elderly group

indicates that spurious minutiae were introduced during the ﬁngerprint

capture process. The spurious minutiae could be introduced by any number

of factors mentioned previously in this section.

The quality analysis showed that:

(a) The ﬁngerprints from the young group had an average quality score of

1.81 and the elderly group had an average quality score of 4.

(b) The quality score has a range of 1 to 5, where 1 indicates the highest

quality score possible and 5 indicates the lowest quality score possible.

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

Issues Involving the Human Biometric Sensor Interface 353

min_count_young

Frequency

1059075604530

120

100

Mean 54.79

StDev 13.08

N 632

Histogram (with Normal Curve) of min_count_young

Fig. 14.10. Minutiae count for the young group.

group showed that both the groups were signiﬁcantly diﬀerent in their

quality scores.

Figures 14.11 and 14.12 show the frequency count of quality scores for

the young and elderly group. The lower quality scores for ﬁngerprint images

from the elderly group indicated that aging and drying issues can introduce

spurious minutiae and aﬀect the quality scores for ﬁngerprints. Figure 14.13

shows the correlation between age and ﬁngerprint image quality. The study

found that the elderly group had lower quality ﬁngerprint images and a

higher minutiae count compared to the younger group.

For ﬁngerprint recognition systems that use minutiae matching

algorithms, the introduction of spurious minutiae points aﬀects the overall

performance of the recognition systems. It is crucial for the system to

be designed keeping these considerations in mind. During enrollment and

identiﬁcation/veriﬁcation phase,

(a) Low quality ﬁngerprints should be rejected so that overall performance

does not degrade.

(b) Awareness about the general demographics of the target users can help

lower the False Accept Rates and False Reject Rates for the system.

to ﬁngerprint images of the elderly can have a signiﬁcant impact on the

performance of the system.

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

354 Synthesis and Analysis in Biometrics

Elderly NFIQ Score

Frequency

7654321

300

250

200

150

100

Mean 4.008

StDev 1.237

N 486

Normal

Histogram of Elderly NFIQ Score

Fig. 14.11. Quality scores for the elderly group.

Young NFIQ Score

Frequency

543210

350

300

250

200

150

100

Mean 1.813

StDev 0.7616

N 632

Normal

Histogram of Young NFIQ Score

Fig. 14.12. Quality scores for the young group.

14.5. Digital Signature Veriﬁcation

The use of written signatures to seal business and personal transactions has

a long history. Before one can implement a signature veriﬁcation system

and collect data from it, one must decide on the method that will be used

to acquire the signature signal. On-line or dynamic signature veriﬁcation

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

Issues Involving the Human Biometric Sensor Interface 355

Fig. 14.13. Pearson correlation of image quality vs. age.

has various methods of acquiring data. Each of the various devices used

in studies demonstrate diﬀerent physical and measurement characteristics.

Image and template quality are also important for system performance.

Many vendors publish speciﬁcation sheets that outline some of the

features of the device, including resolution. Other factors that will aﬀect

the quality of the image presented to the sensor will be:

(a) The cleanliness of the sensor,

(b) Ergonomics (size of the sensor and how it interacts with the body), and

the signature under the pen in the same way as in a pen/paper

interaction).

Detailed device speciﬁcation is also importantdynamic signature

veriﬁcation uses a digitizer tablet, the quality of which will vary. Although

some academic research papers do not list the speciﬁcations of the digitizer,

it is of fundamental importance, aﬀecting the quality and performance of

the device within an e-commerce transaction environment. As technology

improves, there is now no “standard” digitizer, or hardware, to collect DSV

variables. There have been a number of dynamic signature veriﬁcation

studies that have centered on capturing these traits as a pen moves across

a digitizer tablet

[

3,49

]

. These devices include both table-based digitizers

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

356 Synthesis and Analysis in Biometrics

and mobile computer (PDA type) digitizers, both of which are becoming

more sophisticated, capturing an increasing number of measures such as x,

y coordinates, velocity, and pressure. Furthermore, the evolution of mobile

electronic commerce will increase the requirement to sign documents online

with wireless mobile computing devices.

As hardware progresses towards providing solutions to mobile and

electronic commerce, there may be a shift away from table-based digitizers.

Hardware properties can aﬀect the variables collected in the data

acquisition process, and therefore the quality and performance of the device.

Diﬀerent digitizers and software enable the capture of diﬀerent parameters

of the signature, such as diﬀering resolutions and speed.

According to

[

]

, there is no consensus within the research community

on the best method of capturing this information, although the table-based

digitizer tablet is by far the most popular. As hardware manufacturers

develop other solutions to mobile and electronic commerce, there may be

a shift away from these table-based digitizers. Hardware properties can

aﬀect the variables collected in the data acquisition process; as a result,

the quality and performance of the device may also be aﬀected. Diﬀerent

digitizers and software enable the capture of diﬀerent parameters of the

signature, at diﬀering resolutions and speed. Typical acquisition features

include:

• Equivalent time interval samples,

• X and Y co-ordinates of pen-tip movements,

• Pen pressure,

• Pen altitude (the height of the pen from the digitizer),

• Pen azimuth (the angle between the reference line and the line connecting

a point with the origin)

[

]

Example 16. The central focus of

[

]

was to examine whether there

are statistically signiﬁcant diﬀerences in the measurable variables across

devices. The volunteer crew was consisted of 203 individuals whose

demographics tended toward the 19- to 26-year-old age groups, due to the

location of the testing environment. Although not representative of the

U.S. population except for gender, it is representative of the population

found in college and university environments. Males accounted for 66% of

the volunteer group. Females accounted for 35% of the volunteer group and

were typically older than the males. Right-handed individuals accounted

for 91% of the volunteer group, and left-handed individuals accounted for

9%.

April 2, 2007 14:42 World Scientiﬁc Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB

Issues Involving the Human Biometric Sensor Interface 357

The system in this study accepted all the signatures, and the failure to

enroll rate was 0.0%. In

[

]

and

[

]

they deﬁne the “failure to acquire” rate

as the “proportion of attempts for which the system is unable to capture

or locate an image of suﬃcient quality.” Typically, this is the case when

an individual is unable to present a biometric feature to the device due to

either an injury or to insuﬃcient image quality. As such, this does not apply

to dynamic signature veriﬁcation because the biometric feature (signature)

is presented by a stylus rather than direct interface with the biometric

sensor. Therefore, a paradox arises between a signature that is repeatedly

signed and acquired by the sensor yet provides an image diﬀerent to a

normal paper based signature. One subject did have diﬃculty in providing

a signature that is similar to a pen and paper version; although the system

acquired the image, the signature was thrown out, as it was a statistical

outlier with over 15,000 segments as opposed to 2 to 19 segments exhibited

by other crew members. The failure to acquire rate in this study was

0.0%.

When examining area of the signature across two diﬀerent digitizers,

the number of pen-up, pen-down sequences should not change signiﬁcantly

when the signing space is restricted. However, in

[

]

the uptime variable

was signiﬁcantly diﬀerent at an α =0.01 level. Durations of the signature,

i.e. the time for individual 0.01 mm groupings, were also signiﬁcantly

diﬀerent across these two devices. When the variables associated with the

graphical nature of the signature were examined, there was no signiﬁcant

diﬀerence in the number of variables relating to the signing area of the

device. Area was not a factor in

[

]

, as it was held constant on both

devices; and those variables measuring area, such as distance (p-value =

0.181), net area (p-value = 0.676), and rubber area (p-value = 0.344)

show that there is no statistically signiﬁcant diﬀerence in the signing areas

of the two devices. When two other digitizers were chosen (one a PDA,

and the other a table based digitizer), there was a statistically signiﬁcant

diﬀerence between the two devices for 61 variables, including uptime,

distance, and event variables. Even though the signing area between the

two devices was constant, and the variables related to area measurement

such as measuring:

Area,

Distance (p-value = 0.000),

Net area (p-value = 0.000), and

Rubber area (p-value = 0.000),