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

388 Synthesis and Analysis in Biometrics

(conﬁdence interval for training). Note that the diﬀerence ∆t = t

− t

corresponds to the cost-function of training.

A-strategy of training

T(t)

Time t

Effect '=t

-t

Confidence

interval for

training

B-strategy of training

Fig. 15.9. Training function T (t) for estimation of A and B strategies of training.

The training function T (t) is useful for the estimation of the impact of

various modiﬁcations and installations of new devices and systems on the

personnel’s skills. In terms of Bayesian statistics, the eﬀect of a new device

installation can be expressed as

P (Skills) × P (New device|Skills)=P(Skills|New device),

that is,

The prior-to-data probability of the HQP skills revised with

the probability model of additional devices gives the updated

probability of skills.

The training function T (t) is useful for the estimation of the impact

of various modiﬁcations and installations of new devices on the quality of

personnel.

Example 30. Assume that two new devices have been installed in t

and t

(Fig. 15.10). Without training, personnel are not able to use these devices,

and the training function drops.

15.3.4. Evolution Function

A function that shows how additional information can improve decision

making is called an evolution function I(t) in our approach. This is a non-

decreasing function of time t. In terms of Bayesian statistics, the eﬃciency

of installation of new biometric devices can be expressed as

P (New device) × P (Data|New device)=P (New device|Data),

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

Fundamentals of Biometric-Based Training System Design 389

Acceptable level of qualification

Initial training Re-training

Re-training

T(t)

Time t

New device A

installation

New device B

Installation

Motivation

for training

Fig. 15.10. Installation of the new devices A and B at t

and t

, respectively, require

re-training of personnel.

The prior-to-data probability of the new biometric devices

revised with the probability model of any additional device

gives the updated probability of the new devices.

Example 31. The evolution function I(t) is interpreted in Fig. 15.11

as a function of the number of newly installed biometric devices. They

provide additional information about the authorized individual. The eﬀect

of installing new devices is an increase in the probability of making the

correct decision.

I(t)

Time t

New biometric devices

provide extra information

Effect '

Fig. 15.11. The function of insuﬃciency of information I(t) vs. time of installation for

new biometric devices.

Example 32. The IR image indicates that there is no drug or alcohol

consumption. However, through conversation with a person, the oﬃcer

suspects that drugs and alcohol have been consumed. His decision may

include testing of this individual’s iris, or a pupil analyzer. The evolution

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

390 Synthesis and Analysis in Biometrics

function is also useful for predictive analysis both in the PASS and the

training system.

15.4. Training System Design

Prototyping a training system is based on functional analysis, which

includes design analysis and system synthesis. Design analysis aims at

establishing requirements, speciﬁcations, and design evaluation criteria

for the designed system. System synthesis is the translation of system

requirements and functions into hardware and software speciﬁcations.

15.4.1. Desired Characteristics

T-PASS is distinguished from existing training systems by many features,

which can be grouped with respect to architectural, topological, functional,

communication, cost, and other criteria. To minimize cost, T-PASS

utilizes PASS architectural principles and related characteristics and

functionalities. Below, a sample of components and desired characteristics

of T-PASS are given:

Screening, which is widely used in applications such as airport checkpoints,

facilities access, etc.

Decision making support. A knowledge intensive technique can be useful.

However, alternative techniques can be applied to achieve the same

but more cost-eﬃcient result (we used multi-valued logic and decision

diagram techniques

[

]

)

Real-world conditions of training, which is accomplished by the reconﬁgu-

ration of PASS into the training system using minimal extra tools.

Flexibility, the extent to which the system can accommodate changes to

the tasks and environments beyond those ﬁrst speciﬁed.

Expandability and adaptibility. In the proposed training system, the basic

conﬁguration includes tools for generation of early warning data using

visible and IR images, and tools for generation of other synthetic

biometrics. The ﬁrst will model the pre-screening process, and the

second will model the authorization process at the check-desk.

Early warning, that is, the ability to get preliminary information about

an individual, and in this way to reserve extra time for analysis of data

before authorization.

Multiprocessor and distributed conﬁguration. Three main information

streams should be processing in real time: pre-screening data, screening

data, and data from local and global databases.

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

Fundamentals of Biometric-Based Training System Design 391

Mobile conﬁguration. T-PASS inherits this property from PASS which

can be deployed at any public place where authorization is needed

(important public events, exhibitions, and other social meetings or high

proﬁle settings where VIP security is required on a temporary basis).

These characteristics were the desired aims at the various phases of the

system design.

15.4.2. Architecture

In the development of the training system T-PASS, most of the features

of PASS are inherited. In Fig. 15.12, the topological and architectural

properties of both of these systems are shown. It follows that the T-PASS

architecture uses the PASS platform. However, T-PASS is diﬀerent from

PASS in the following features:

(a) Biometrics of pre-screened and screened individuals are simulated by

the system,

(b) The appearance of biometric data can be modeled by the user using

modeling tools,

Phase 1

(unknown person)

Processing

preliminary

data

Phase 2

(specified

individual)

Dialogue

support

Knowledge domain

converter

Forming

the

person’s

file

Global

database

Phase 3:

Decision

making

Visible

band

camera

Visible

band

camera

IR ban

camera

IR band

camera

Voice

analyzer

On-line

processing

10-20 meters distance

Gait biometric

Individual

Officer

Pre-screened

individual

PASS

Processing

preliminary data

Dialogue

support

Knowledge domain

converter

Voice

analyzer

Trained

person

Temperature

templates

Synthetic

faces

generator

Speech

tracing

Scenario

database

T-PASS

Synthetic

biometric data

generation

(a) (b)

Fig. 15.12. The PASS (a) and T-PASS (b) architecture.

The following distribution of biometric devices is related to the screening

discipline (Fig. 15.13):

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

392 Synthesis and Analysis in Biometrics

Distance, or non-contact biometrics (facial image in visible and IR

bands), used to provide early warning information about a pre-screened

individual. Gait biometrics is a good candidate for obtaining the

additional information when individuals move from pre-screened to

screened positions.

Contact biometrics, used to provide information about an individual

(ﬁngerprints, palmprint, hand geometry, signature, stroke dynamic,

etc.) Correlation between various biometrics in semantic form provide

a useful interpretation of biometric patterns.

PRE-

SCREENING

Gait biometric

(

20 meter

)

CHECK-

POINT

Distance

(non-contact)

biometrics

Contact

biometrics

Synthetic

biometrics

SIMULATION

OFFICER

PASS

T-PASS

Gait biometric

simulatio

Synthetic

biometrics

Fig. 15.13. The topological properties of PASS and the distribution of various types

of biometric: distance (non-contact) biometrics are used to obtain early warning

information and near distance and contact biometrics are used in real time authorization.

Our approach to T-PASS design is related to virtual modeling

techniques

[

]

. Virtual environments refer to interactive systems that

provide the user with a sense of direct physical presence and direct

experience. These systems include visual and feedback techniques, and

allow natural interaction by manipulation of computer-generated objects.

In T-PASS, synthetic biometric data are used that can be understood as

virtual biometric data.

Example 33. Surgeons and security personnel do not prefer immersion-

based training systems since it makes them feel isolated from the real

operation scene.

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

Fundamentals of Biometric-Based Training System Design 393

15.4.3. Training Scenarios

Humans form mental models through experience, training and instruction.

Important human factors that are studied through the mental model are

the prioritization of events and their sequence of occurrence. In PASS,

scenarios are developed using architectural and topological characteristics.

Possible scenarios are divided into three groups: regular, non-standard, and

extreme. Most scenarios in each group are generated automatically using

computer-aided design tools.

Scenarios based on distance IR images

Let us consider an example of a training scenario, in which the system

generates the following data about the screened person:

Time 00.00.00:

Protocol of pre-screening person 45

Warning, level 04

Specification: Drug or alcohol consumption, level 03

Possible action:

1. Direct special inspection

2. Register with caution

It follows from this protocol that the system evaluates the fourth level

of warning using automatically measured drug or alcohol consumption for

the screened individual. A knowledge-based sub-system evaluates the risks

and generates two possible solutions. The oﬃcer can, in addition to an

automatic analysis, analyze the acquired images in the visible and IR

spectra.

Scenarios corresponding to the results of matching

We distinguish scenarios that correspond to the results of the matching of

an individual’s data with data in local and global databases. This process

is fully automated for some stationary conditions, for example:

Time 00.00.00:

Protocol of screening person 45

Warning, level 04

Specification: Drug or alcohol consumption, level 03

Local database matching: positive

Possible action:

1. Register

2. Clarify and register

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

394 Synthesis and Analysis in Biometrics

Note that data may not always be available in a database — this is the

worst case scenario, and knowledge-based support is vital in this case.

Scenarios based on the analysis of behavioral biometric data

The example below introduces scenarios based on the analysis of

behavioural biometric data.

Time 00.00.00:

Protocol of screening person 45

Warning, level 04

Specification: Drug or alcohol consumption, level 03

Local database matching: positive

Proposed dialog:

Question 1: Do you need medical assistance?

Question 2: Any service problems during the flight?

Question 3: Do you plan to rent a car?

Question 4: Did you meet friends on board?

Question 5: Do you prefer red wine or whisky?

Question 6: Do you have drugs in your luggage?

The results of automatically analyzed behavioral information are

presented to the oﬃcer:

Time 00.00.00:

Protocol of screening person 45

Warning, level 04

Specification: Drug or alcohol consumption, level 03

Local database matching: positive

Level of trustworthiness of Question 6 is 03: Do you have drugs in

your luggage?

Possible action:

1. Direct to special inspection

2. Continue clarification by questions

15.4.4. Decision Making Support

In the training system, decisions are generated by special tools and

various scenarios in PASS, instead of the local biometric device response

(Fig. 15.14).

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

Fundamentals of Biometric-Based Training System Design 395

Scenario 1

Scenario 2

Scenario 3

SIMULATION OF

AUTOMATIC DECISION

MAKING OF

BIOMETRIC DEVICES

OFFICER

Decision

making

KNOWLEDGE

BASED

SUPPORT

Local

level

Global

level

Analysis

Fig. 15.14. Automatic generation of decisions at local levels is replaced in the training

system by a simulator.

15.5. Training System Prototype

Prototyping of T-PASS is aimed at proving the basic concept of training

system design. In this section, we introduce some details of the platform,

data ﬂows and their representations, computing of indirect characteristics,

conversion patterns from signal to knowledge domain, and estimations.

15.5.1. Platform

In prototyping we used an extended platform, that is, we implemented the

hardware and software tools without constraints on the cost and criteria of

optimal conﬁguration. The following facilities were utilized in prototyping

[

]

Hardware platform: a network of four computers connected with cameras

in the IR and visible bands,

Software platform: about a dozen packages on biometric data processing

and simulation,

Local and global communication resources, and

Biometric devices: about a dozen contact and non-contact devices for the

acquisition and recognition of behavioral and physical biometrics.

15.5.2. Data Flow

In prototyping, we distinguish several levels of data ﬂows. First of all, the

information ﬂows are divided into the following streams:

Static document record and biometric data from local and global

databases.

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

396 Synthesis and Analysis in Biometrics

Dynamic biometric data obtained from surveillance facilities and conversa-

tion with an individual at the desk, including behavioural biometric

data (voice, facial expressions, signature).

We distinguish four forms of the individual’s information:

(a) File data from databases which are well structured and standard;

(b) Appearance, or visual data, acquired mainly in real-time;

speciﬁc skills for understanding and analysis;

(d) Various forms of behavioural biometrics (facial expressions, voice,

signature) that are acquired in real-time (during pre-screening or check-

in procedure).

15.5.3. Indirect Characteristic Computing

In our experimental study, we focus on various aspects of indirect

characteristic computing, in particular (Fig. 15.15): interpretation of

biometric patterns in semantic form, initialization of questionnaires,

reliability of biometric data, and their improvement using correlation in

the knowledge domain.

Temperature

Alcohol and dru

consumption

Blood pressure

Pulse

Artificial

accessories

Plastic surgery

Information

source

from

3-12

Tension

Intonation

Phonetic level

 Phonological

level

Morphological

level

Correlation with

physiological

parameters

 Facial

expressions

Iris

Brow and lips

movements

Correlation with

speech and

physiological

parameters

Information

source

from

400-700

Information

source

from

70-600Hz

C o r r e l a t I o n

Authorized

individual

Fig. 15.15. Prototyping indirect biometric characteristics using correlation technique.

A fragment of this design is given in Fig. 15.16. The oﬃcer manipulates

and eﬃciently uses biometric data to make a decision according to various

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

Fundamentals of Biometric-Based Training System Design 397

scenarios. Many factors render decision making a complex procedure. To

simplify this procedure, decision making is divided into several simple steps

with respect to time and levels.

Health problems

Alcohol or drug

consumption

Unexpected

question

Supporting

questions

BIOMETRIC

FEATURES IN

VISIBLE,

INVISIBLE, AND

ACOUSTIC

BANDS

Microphone

Camera

in IR

band

Camera

in visible

band

Non-

invasive

automatic

biometric

measures

Supporting

dialog

algorithms

Signal

domain

Knowledge

domain

Conversation

features to

knowled

Question

Answer

Officer

Authorized

individual

(a) (b)

Fig. 15.16. Prototyping of T-PASS: interpretation of temperature regions in an IR

image (a) and generation of supporting questions in PASS (b).

15.5.4. Conversion of Data

The key component of a system is the converter from the signal domain

to the knowledge domain. The eﬃciency of the converter is critical in

decision making support. The converter implements a knowledge-based

technology. In the development of the converter and its prototyping, we

utilize properties of the system such as

(a) Limited sets of supported questions (biometric patterns of a given class

initialize 3–5 questions),

(b) Limited types of biometrics (about 10), and

The example below explains our approach.

Example 34. Let us assume that there are three classes of samples

assigned to “Disability,” “Alcohol intoxication,” and “Normal,” and

denoted as values of the three-valued logic variable {0}, {1},and{2},

respectively. Let us also assume that feature patterns from facial IR

image and gait correspond to the above classes. Using a logic function

representation and decision diagram technique

[

35,49

]

, the following

semantic constructions can be generated by the system