A Hierarchical Fingerprint Matching System · 2018-11-27 · A HIERARCHICAL FINGERPRINT MATCHING...

A HIERARCHICAL FINGERPRINT MATCHINGSYSTEM

A thesis submitted in partial fulfillment of the

requirements for the degree of

Bachelor-Master of Technology (Dual)

by

ABHISHEK RAWAT

to the

Department of Computer Science and Engineering

Indian Institute of Technology Kanpur

July 2009

Abstract

Fingerprint Recognition is a widely popular but a complex pattern recognition prob-

lem. It is difficult to design accurate algorithms capable of extracting salient features

and matching them in a robust way. The real challenge is matching fingerprints af-

fected by: i) High displacement/or rotation which results in smaller overlap between

template and query fingerprints, ii) Non-linear distortion caused by the finger plastic-

ity, iii) Different pressure and skin condition and iv) Feature extraction errors which

may result in spurious or missing features. The information contained in a finger-

print can be categorized into three different levels, namely, Level 1 (pattern), Level

2 (minutia points), and Level 3 (pores and ridge contours). Despite their discrimina-

tive power, the Level 3 features are barely used by the vast majority of contemporary

automated fingerprint authentication systems (AFAS) which rely mostly on minutiae

features. This is mainly because, most of these authentication systems are equipped

with 500 ppi (FBI’s standard of fingerprint resolution for AFAS) scanners, and reli-

ably extracting “fine and detailed” Level 3 features require high resolution images.

While this may have been the case with many older live-scan devices, the current

devices are capable of detecting a reasonable amount of level 3 details even at the rel-

atively limited 500 ppi resolution. In this thesis the above mentioned problems have

been addressed and a new hierarchical matcher has been proposed. The hierarchical

matcher utilizes Level 3 features (pores and ridge contour) in conjunction with Level

2 features (minutiae) for matching. The aim is to reduce the error rates, namely FAR

(False Acceptance Rate) and FRR (False Rejection Rate) in the existing minutiae

based systems. The hierarchical matcher has been tested on three diverse databases

in public domain. The obtained results are promising and verify our claim.

2

Acknowledgments

I would like to express my deep-felt gratitude to my advisor, Dr. Phalguni Gupta

for giving me an opportunity to work with the Biometrics Group and for his advice,

encouragement, constant support. I wish to thank him for extending me the greatest

freedom in deciding the direction and scope of my research. It has been both a

privilege and a rewarding experience working with him.

I would also like to thank all my friends and colleagues here at IIT Kanpur for

all the wonderful times I have had with them. Their valuable comments and sugges-

tions have been vital to the completion of this work. I want to thank the faculty of

Computer Science Department and the staff for providing me the means to complete

my degree and prepare for a career as a computer scientist.

And finally, I am grateful to my parents for their love, sacrifice, understanding,

encouragement and support.

Abhishek Rawat

i

Contents

List of Figures iii

List of Tables v

1 INTRODUCTION 11.1 Fingerprint as a Biometric . . . . . . . . . . . . . . . . . . . . . . . . 3

1.1.1 History of Fingerprint Recognition . . . . . . . . . . . . . . . 41.1.2 Fingerprint Representation . . . . . . . . . . . . . . . . . . . . 5

1.2 Motivation and Problem Definition . . . . . . . . . . . . . . . . . . . 81.3 Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2 BACKGROUND AND LITERATURE REVIEW 132.1 Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142.2 Image Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.3 Fingerprint Image Enhancement . . . . . . . . . . . . . . . . . . . . . 172.4 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.4.1 Minutiae Extraction . . . . . . . . . . . . . . . . . . . . . . . 202.4.2 Pores Extraction . . . . . . . . . . . . . . . . . . . . . . . . . 222.4.3 Ridge Contour Extraction . . . . . . . . . . . . . . . . . . . . 25

2.5 Fingerprint Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . 262.5.1 Correlation-based Techniques . . . . . . . . . . . . . . . . . . 272.5.2 Minutiae-based Methods . . . . . . . . . . . . . . . . . . . . . 28

2.5.2.1 Global Matching . . . . . . . . . . . . . . . . . . . . 312.5.2.2 Local Matching . . . . . . . . . . . . . . . . . . . . . 34

2.5.3 Ridge Feature-based Matching Techniques . . . . . . . . . . . 372.5.3.1 Texture Feature based Techniques . . . . . . . . . . 372.5.3.2 Level 3 Features-based Techniques . . . . . . . . . . 39

ii

3 PROPOSED HIERARCHICAL MATCHER 423.1 Preprocessing and Enhancement . . . . . . . . . . . . . . . . . . . . . 443.2 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.3 Hierarchical Matching . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.3.1 Minutia-based Matcher . . . . . . . . . . . . . . . . . . . . . . 483.3.1.1 Local Structure Matching . . . . . . . . . . . . . . . 503.3.1.2 Consolidation step . . . . . . . . . . . . . . . . . . . 54

3.3.2 Level 3 features based Matcher . . . . . . . . . . . . . . . . . 553.3.2.1 Pore Matching . . . . . . . . . . . . . . . . . . . . . 563.3.2.2 RidgeContour Matching . . . . . . . . . . . . . . . . 573.3.2.3 Fusion of Level 2 and Level 3 features . . . . . . . . 61

4 EXPERIMENTAL RESULTS 634.1 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.2 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . 654.3 Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 674.4 Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 684.5 Experiment 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.6 Timing Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5 CONCLUSION & FUTURE WORK 825.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

Bibliography 85

iii

List of Figures

1.1 Fingerprint features at Level 1, Level 2 and Level 3 ([7, 54]) . . . . . 61.2 Open and closed pores ([29]) . . . . . . . . . . . . . . . . . . . . . . . 71.3 Characteristics of ridge contours and edges ([9]) . . . . . . . . . . . . 8

2.1 Architecture of a Fingerprint Verification System. . . . . . . . . . . . 142.2 Fingerprints at different resolutions a) 380 ppi, b) 500 ppi, c)1000 ppi

([29]). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162.3 A typical minutiae extraction process ([30]) . . . . . . . . . . . . . . 232.4 An example of two false matched local structures ([36]) . . . . . . . . 35

3.1 The proposed hierarchical matcher. . . . . . . . . . . . . . . . . . . . 443.2 Effect of differential finger pressure([36]). . . . . . . . . . . . . . . . 463.3 Ridge Contour extraction process in a image scanned with Cross Match

Verifier 300 scanner at 500 dpi. . . . . . . . . . . . . . . . . . . . . . 473.4 Pores extracted from a image scanned with Cross Match Verifier 300

scanner at 500 dpi. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473.5 The schematic description of a “star” with k = 6 neighbors. The

neighbors are chosen from all the four quadrants . . . . . . . . . . . . 493.6 Two impressions of same fingerprint at 500 dpi. . . . . . . . . . . . . 573.7 Schematic of square to circular transformation . . . . . . . . . . . . . 593.8 The basic LBP operator . . . . . . . . . . . . . . . . . . . . . . . . . 60

4.1 Sample images from Neurotechnology Database . . . . . . . . . . . . 644.2 Sample images from FVC 2004, DB3 Database . . . . . . . . . . . . . 644.3 Sample images from FVC 2006, DB2 Database . . . . . . . . . . . . . 654.4 ROC curves for proposed minutia matcher and CBFS-Kplet based

matcher, on Neurotechnology Database . . . . . . . . . . . . . . . . . 684.5 ROC curves for proposed minutia matcher and CBFS-Kplet based

matcher, on FVC 2004, DB3 Database . . . . . . . . . . . . . . . . . 694.6 ROC curves for proposed minutia matcher and CBFS-Kplet based

matcher, on FVC 2006, DB2 Database . . . . . . . . . . . . . . . . . 70

iv

4.7 ROC curves comparing the performance of LBP and ZM (Neurotech-nology Database) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

4.8 ROC curves comparing the performance of LBP and ZM (FVC 2004,DB3 Database) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

4.9 ROC curves comparing the performance of LBP and ZM (FVC 2006,DB2 Database) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

4.10 ROC curves for the minutia matcher and hierarchical matcher on Neu-rotechnology Database . . . . . . . . . . . . . . . . . . . . . . . . . . 74

4.11 ROC curves for the minutia matcher and hierarchical matcher on FVC2004, DB3 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.12 ROC curves for the minutia matcher and hierarchical matcher on FVC2006, DB2 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.13 Distribution of number of matched minutiae, for Genuine and Impostorcases, on Neurotechnology Database . . . . . . . . . . . . . . . . . . . 78

4.14 Distribution of number of matched minutiae, for Genuine and Impostorcases, on FVC 2004, DB3 Database . . . . . . . . . . . . . . . . . . . 79

4.15 Distribution of number of matched minutiae, for Genuine and Impostorcases, on FVC 2006, DB2 Database . . . . . . . . . . . . . . . . . . . 80

v

List of Tables

4.1 Equal Error Rate (EER) comparison between proposed minutia matcherand CBFS-Kplet based matcher. . . . . . . . . . . . . . . . . . . . . . 67

4.2 Equal Error Rate (EER) comparison between proposed minutia matcherand proposed hierarchical matcher. . . . . . . . . . . . . . . . . . . . 71

4.3 Comparison of matching time for Level 3 Features. . . . . . . . . . . 81

1

Chapter 1

INTRODUCTION

Biometric based recognition, or biometrics, is the science of identifying, or verifying

the identity of, a person based on physiological and/or behavioral characteristics [14].

Physiological traits are related to the physiology of the body and mainly include

fingerprint, face, DNA, ear, iris, retina, hand and palm geometry. Behavioral traits

are related to behavior of a person and examples include signature, typing rhythm,

gait, voice etc. Biometric recognition offers many advantages over traditional PIN

number or password and token-based (e.g., ID cards) approaches. A biometric trait

cannot be easily transferred, forgotten or lost, the rightful owner of the biometric

template can be easily identified, and it is difficult to duplicate a biometric trait [20].

There are a number of desirable properties for any chosen biometric characteristic

[14]. These include:

1. Universality : Every person should have the characteristic.

2. Uniqueness : No two persons should be the same in terms of the biometric

characteristic.

2

3. Permanence : The biometric characteristics should not change, or change min-

imally, over time.

4. Collectability : The biometric characteristic should be measurable with some

(practical) sensing device.

5. Acceptability : The user population and the public in general should have no

(strong) objections to the measuring/collection of the biometric trait.

A biometric system is essentially a pattern recognition system that operates by

acquiring biometric data from an individual, extracting a feature set from the acquired

data, and comparing this feature set against the template set in the database [32].

Depending on the application context, a biometric system may operate either in

verification mode or identification mode:

• In the verification mode, an individual provides his/her biometric data and

claims an identity, usually via a PIN (Personal Identification Number), a user

name, a smart card, etc. The system then verifies the individual’s identity

by comparing the acquired biometric data with the individual’s own biometric

template(s) stored in system database. Such a system basically performs a one-

to-one comparison to determine whether the claimed identity is true or not.

• In the identification mode, the system compares the given biometric data with

the templates of all the users in the database. Therefore, the system conducts

a one-to-many comparison to establish an individual’s identity (or fails if the

subject is not enrolled in the system database) without the subject having to

claim an identity.

1.1. FINGERPRINT AS A BIOMETRIC 3

The effectiveness of a biometric system can be judged by following characteristics

[35]:

1. Performance : This refers to the achievable recognition accuracy, speed, ro-

bustness, the resource requirements to achieve the desired recognition accuracy

and speed, as well as operational (work environment of individual, e.g., manual

workers may have a large number of cuts and bruises on their fingerprints) or

environmental factors (humidity, illumination etc.) that affect the recognition

accuracy and speed .

2. Scalability : This refers to the ability to encompass large number of individuals

without a significant decrease in the performance.

3. Non-invasiveness : This refers to the ease with which the information can be

captured from individuals, without damaging an individual’s physical integrity

and ideally without special preparations by/of an individual.

4. Circumvention : This refers to the degree to which the system is resistant to

spoofs or attacks.

A practical biometric system should meet the specified recognition accuracy, speed,

and resource requirements, be harmless to the users, be accepted by the intended pop-

ulation, and be sufficiently robust to various fraudulent methods and attacks to the

system.

1.1 Fingerprint as a Biometric

A fingerprint is an impression of the friction ridges, from the surface of a finger-

tip. Fingerprints have been used for personal identification for many decades, more


recently becoming automated due to advancements in computing capabilities. Fin-

gerprint recognition is nowadays one of the most important and popular biometric

technologies mainly because of the inherent ease in acquisition, the numerous sources

(ten fingers) available for collection, and the established use and collections by law

enforcement agencies. Automatic fingerprint identification is one of the most reliable

biometric technologies. This is because of the well known fingerprint distinctiveness,

persistence, ease of acquisition and high matching accuracy rates. Fingerprints are

unique to each individual and they do not change over time. Even identical twins

(who share their DNA) do not carry identical fingerprints. The uniqueness can be

attributed to the fact that the ridge patterns and the details in small areas of friction

ridges are never repeated. These friction ridges develop on the fetus in their definitive

form before birth and are known to be persistent throughout life except for perma-

nent scarring. Scientific research in areas such as biology, embryology, anatomy and

histology has supported these findings [4]. Also, the matching accuracy of fingerprint

based authentication systems has been shown to be very high. Fingerprint-based au-

thentication systems continue to dominate the biometrics market by accounting for

almost 52% of authentication systems based on biometric traits [35].

1.1.1 History of Fingerprint Recognition

Fingerprints have been found on ancient artifacts recovered from excavation sites

of various civilizations [42]. However fingerprints have been used for identification

only from nineteenth century onwards. A time-line of important events that has

established the foundation of the modern fingerprint based biometric technology can

be found in [2]. Henry Fauld [21] has first scientifically suggested the individuality

and uniqueness of fingerprints. Sir Francis Galton has published the well-known book


entitled Fingerprints [22], in which a detailed statistical model of fingerprint analysis

and identification has been discussed. Galton has introduced Level 2 features by

defining minutia points as either ridge endings or ridge bifurcations on a local ridge.

An important advance in fingerprint identification has been made by Edward Henry,

who has established a system known as “Henry system” for fingerprint classification

[65].

In [44], Locard has introduced the science of “poroscopy”, the comparison of

sweat pores for the purpose of personal identification. Locard has stated that like

the ridge characteristics, the pores are also permanent, immutable, and unique, and

are useful for establishing the identity, especially when a sufficient number of ridges

is not available. Chatterjee has proposed the use of ridge edges in combination with

other friction ridge formations to establish individualization, which is referred to as

“edgeoscopy” [9].

Over the last few years, poroscopy and edgeoscopy have received growing attention

and have been widely studied by latent fingerprint examiners [9]. It has been claimed

that shapes and relative positions of sweat pores and shapes of ridge edges are as

permanent and unique as traditional minutia points. And when understood, they

add considerable weight to the conclusion of identification [9].

1.1.2 Fingerprint Representation

The types of information that can be collected from a fingerprint’s friction ridge

impression can be categorized as Level 1, Level 2, or Level 3 features as shown in

Figure 1.1.

At the global level, the fingerprint pattern exhibits one or more regions where

the ridge lines assume distinctive shapes characterized by high curvature, frequent


Figure 1.1: Fingerprint features at Level 1, Level 2 and Level 3 ([7, 54])

termination, etc. These regions are broadly classified into arch, loop, and whorl. The

arch, loop and whorl can further be classified into various subcategories. Level 1

features comprises these global patterns and morphological information. They alone

do not contain sufficient information to uniquely identify fingerprints but are used for

broad classification of fingerprints.

Level 2 features or minutiae refers to the various ways that the ridges can be

discontinuous. These are essentially Galton characteristics, namely ridge endings and

ridge bifurcations. A ridge ending is defined as the ridge point where a ridge ends

abruptly. A bifurcation is defined as the ridge point where a ridge bifurcates into

two ridges. Minutiae are the most prominent features, generally stable and robust

to fingerprint impression conditions. The distribution of minutiae in a fingerprint is

considered unique and most of the automated matchers use this property to uniquely


identify fingerprints. Uniqueness of fingerprint based on minutia points has been

quantified by Galton [22]. Statistical analysis has shown that Level 2 features, have

sufficient discriminating power to establish the individuality of fingerprints [56].

Level 3 features are the extremely fine intra ridge details present in fingerprints

[6]. These are essentially the sweat pores and ridge contours. Pores are the openings

of the sweat glands and they are distributed along the ridges. Studies [9] have shown

that density of pores on a ridge varies from 23 to 45 pores per inch and 20 to 40

pores should be sufficient to determine the identity of an individual. A pore can be

either open or closed, based on its perspiration activity. A closed pore is entirely

enclosed by a ridge, while an open pore intersects with the valley lying between

two ridges as shown in Figure 1.2. The pore information (position, number and

shape) are considered to be permanent, immutable and highly distinctive but very

few automatic matching techniques use pores since their reliable extraction requires

high resolution and good quality fingerprint images. Ridge contours contain valuable

Level 3 information including ridge width and edge shape. Various shapes on the

friction ridge edges can be classified into eight categories, namely, straight, convex,

peak, table, pocket, concave, angle, and others as shown in Figure 1.3. The shapes

and relative position of ridge edges are considered as permanent and unique.

Figure 1.2: Open and closed pores ([29])

1.2. MOTIVATION AND PROBLEM DEFINITION 8

Figure 1.3: Characteristics of ridge contours and edges ([9])

1.2 Motivation and Problem Definition

Fingerprint recognition is a complex pattern recognition problem. It is difficult

to design accurate algorithms capable of extracting salient features and matching

them in a robust way, especially in poor quality fingerprint images and when low-cost

acquisition devices with small area are adopted. There is a popular misconception

that automatic fingerprint recognition is a fully solved problem since it was one of

the first applications of machine pattern recognition. On the contrary, fingerprint

recognition is still a challenging and important pattern recognition problem. The real

challenge is matching fingerprints affected by: i) High displacement/or rotation which

results in smaller overlap between template and query fingerprints (this case can be

treated as similar to matching partial fingerprints), ii) Non-linear distortion caused

by the finger plasticity, iii) Different pressure and skin condition and iv) Feature

extraction errors which may result in spurious or missing features.

The vast majority of contemporary automated fingerprint authentication systems

(AFAS) are minutiae (level 2 features) based [46]. Minutiae-based systems gener-

ally rely on finding correspondences 1 between the minutia points present in “query”

and “reference” fingerprint images. These systems normally perform well with high-

1A minutiae in the “query” fingerprint and a minutiae in the “reference” fingerprint are said tobe corresponding if they represent the identical minutiae scanned from the same finger

1.2. MOTIVATION AND PROBLEM DEFINITION 9

quality fingerprint images and a sufficient fingerprint surface area. These conditions,

however, may not always be attainable. In many cases, only a small portion of the

“query” fingerprint can be compared with the “reference” fingerprint as a result the

number of minutiae correspondences might significantly decrease and the matching

algorithm would not be able to make a decision with high certainty. This effect is even

more marked on intrinsically poor quality fingers, where only a subset of the minutiae

can be extracted and used with sufficient reliability. Although minutiae may carry

most of the fingerprint’s discriminatory information, they do not always constitute

the best trade-off between accuracy and robustness. This has led the designers of fin-

gerprint recognition techniques to search for other fingerprint distinguishing features,

beyond minutiae, which may be used in conjunction with minutiae (and not as an

alternative) to increase the system accuracy and robustness.

It is a known fact that the presence of Level 3 features in fingerprints provides

minute detail for matching and the potential for increased accuracy. The forensic

experts in law enforcement often make use of Level 3 features, such as sweat pores

and ridge contours, to compare fingerprint samples when insufficient minutia points

are present in the fingerprint image or poor image quality hampers minutiae analysis.

That is, experts take advantage of an extended feature set in order to conduct a more

effective matching. Despite their discriminating property, level 3 features are barely

utilized in the commercial automated fingerprint authentication systems (AFAS), as

a result a large amount of fingerprint information is ignored by such systems. This

is mainly because, most of these authentication systems are equipped with 500 ppi

(pixels per inch) scanners, and reliably(or consistently) extracting ”fine and detailed”

Level 3 features require high resolution images. While this may have been the case

with many of the older live-scan devices, the current devices are capable of detect-

1.3. APPROACH 10

ing a reasonable amount of level three detail even at the relatively limited 500 ppi

resolution. Ray et al. [62] have presented a means of modeling and extracting pores

(which are considered as highly distinctive Level 3 features) from 500 ppi fingerprint

images. This study showed that while not every fingerprint image obtained with a

500 ppi scanner has evident pores, a substantial number of them do have. Thus, it is

a natural step to try to extract Level 3 information, and use them in conjunction with

minutiae to achieve robust matching decisions. In addition, the fine details of level 3

features could potentially be exploited in circumstances that require high-confidence

matches.

1.3 Approach

In this thesis an approach has been presented which addresses the various issues

and challenges (discussed in previous section) in fingerprint matching. The aim is

to reduce the error rates, namely False Acceptance Rate (FAR) and False Rejection

Error (FRR) in the existing fingerprint matching algorithms.

The proposed approach utilizes Level 3 features (pores and ridge contours) along

with Level 2 features for matching fingerprints at 500 ppi, in a hierarchical manner.

The first stage of hierarchical matcher is the minutia matching stage in which the

Level 2 minutiae points from the “query” and “reference” fingerprints are matched.

Based on the performance of the minutiae matcher the matching process either stops

if a match is found or continues to next stage where the Level 3 features are used to

decide match/non-match decision. The hierarchical matcher utilizes the fine details

of Level 3 features to decide the match/non-match decision in circumstances where

the decision can not be made solely on the basis of Level 2 features.

1.3. APPROACH 11

The proposed approach addresses the various challenges in fingerprint matching

in following way:

• The plastic nature of finger skin results in non-linear distortion in successive

acquisitions of the same finger. To deal with this problem the matching of

all feature classes (Level 2 and Level 3) are done within a local region. The

use of localized matching minimizes the effects of non-linear distortion. This

is because the effects of distortion does not significantly alter the fingerprint

pattern locally [35].

• Due to high displacement and rotation which are introduced during fingerprint

acquisition, different impressions of the same finger differ from each other. Most

of the existing minutia matching algorithms first align fingerprint images and

then find minutia correspondences. But in case of poor quality fingerprints,

global registration (alignment) parameters do not exist and as a result it is not

possible to get a correct alignment. The errors introduced during registration

steps can introduce errors in the subsequent steps. The proposed approach does

not use alignment at any stage and relies on rotationally invariant structures

(in case of minutia and pore matching) and features (in case of ridge contours)

• The high displacement/rotation introduced during fingerprint acquisition re-

sults in “small overlap” between query and reference fingerprints. Also the

noise introduced by several factors such as poor skin conditions, unclean scan-

ner surface etcetera, results in a very small portion of the fingerprint which can

actually be used for comparison. The use of Level 3 features in conjunction

with Level 2 features takes care of such situations. Studies [40, 41] have shown

that given a sufficiently high resolution fingerprint, the use of Level 3 features

1.4. THESIS OUTLINE 12

from fingerprint fragments results in same quantity of discriminative informa-

tion that can be extracted when Level 2 features are considered and the entire

image is used.

• Finally, due to the noise in the fingerprints the feature extraction techniques

often introduce errors such as missing or spurious minutiae/pores. The match-

ing technique should handle such cases. The proposed approach uses an elastic

string matching algorithm for pores matching and minutiae matching, which

can accommodate perturbations of minutiae/pores from their true locations

and can tolerate spurious and missing minutiae/pores. Also, for matching ridge

contours we have used statistical features based on Zernike moments and Lo-

cal Binary Patterns (LBP) which are known for their tolerance to noise and

gray-scale invariance, respectively.

1.4 Thesis Outline

The outline of the thesis is as follows: Chapter 2 discusses the literature on fin-

gerprint based verification systems. Chapter 3 presents the proposed hierarchical

fingerprint matching system. Chapter 4 presents the results and evaluations of the

proposed approach. Finally, in Chapter 5 we conclude and outline the future work.

13

Chapter 2

BACKGROUND AND

LITERATURE REVIEW

A fingerprint recognition system may operate either in verification mode or iden-

tification mode. In verification mode, the system verifies an individual’s identity by

comparing the input fingerprint with the individual’s own template(s) stored in the

database. In, the identification mode, the system identifies an individual by searching

the templates of all the users in the database for a match. The fingerprint classifi-

cation and indexing techniques are used to speed up the search in fingerprint based

identification systems. The fingerprint feature extraction and matching algorithms

are usually quite similar for both fingerprint verification and identification problems.

In this thesis the focus is on fingerprint based verification systems.

The various stages in a fingerprint verification system is shown in Figure 2.1.

The first stage is the data acquisition stage in which a fingerprint image is obtained

from an individual by using a sensor. The next stage is the pre-processing stage

in which the input fingerprint is processed with some standard image processing

2.1. DATA ACQUISITION 14

algorithms for noise removal and smoothening. The pre-processed fingerprint image

is then enhanced using specifically designed enhancement algorithms which exploit

the periodic and directional nature of the ridges. The enhanced image is then used to

extract salient features in the feature extraction stage. Finally, the extracted features

are used for matching in the matching stage. This chapter discusses the current state

of the art feature extraction techniques and gives a literature survey on the various

fingerprint matching algorithms.

Figure 2.1: Architecture of a Fingerprint Verification System.

2.1 Data Acquisition

Traditionally, in law enforcement applications fingerprints were acquired off-line

by transferring the inked impression on a paper. Nowadays, the automated fingerprint

verification systems use live-scan digital images of fingerprints acquired from a fin-

gerprint sensor. These sensors are based on optical, capacitance, ultrasonic, thermal

and other imaging technologies.

2.1. DATA ACQUISITION 15

The optical sensors are most popular and are fairly inexpensive. These sensors

are based on FTIR (Frustrated Total Internal Reflection) technique. When a finger

touches the sensor surface (which actually is a side of a glass prism), one side of the

prism is illuminated through a diffused light. While the fingerprint valleys that do

not touch the sensor surface reflect the light, ridges that touch the surface absorb the

light. The sensor exploits this differential property of light reflection to differentiate

the ridges (which appear dark) from the valleys.

The capacitive sensors utilize the principle associated with capacitance to form

the fingerprint images. These sensors consists of a two-dimensional array of metal

electrodes. Each metal electrode acts as one plate of a parallel-plate capacitor and

the contacting finger acts as the other plate. When a finger is pressed on the sensor

surface, it creates varying capacitance values which depends inversely on the distance

between the sensing plate and the finger surface. The ridges thus have increased

capacitance compared to valleys. This variation is then converted into an image of

the fingerprint.

The Ultrasound technology based sensors are the most accurate of the fingerprint

sensing technologies. It uses ultrasound waves and measures the distance based on

the impedance of the finger, the plate and air. The thermal sensors are made up

of pyro-electric materials, which generate a temporary electrical potential when they

are heated or cooled. When a finger is swiped across the sensor, there is differential

conduction of heat between the ridges and valleys (as skin is a better conductor than

the air in the valleys) which is measured by the sensor.

One of the most essential characteristics of a digital fingerprint image is its res-

olution, which indicates the number of dots or pixels per inch (ppi). The minimum

resolution that allows the feature extraction algorithms to locate minutiae is 250 to

2.2. IMAGE PREPROCESSING 16

300 ppi. The FBI’s standard for resolution of fingerprint sensors is 500 ppi. A large

number of automated fingerprint verification systems accept 500 ppi fingerprints. Fig-

ure 2.2 shows the fingerprints captured at different resolutions. At 500 ppi pores are

visible but in-order to extract pores reliably a significantly higher resolution (1000

ppi) of image is needed.

Figure 2.2: Fingerprints at different resolutions a) 380 ppi, b) 500 ppi, c)1000 ppi([29]).

2.2 Image Preprocessing

The preprocessing steps try to compensate for the variations in lighting, contrast

and other inconsistencies which are introduced by the sensor during the acquisition

process. The following preprocessing steps are generally used:

• Gaussian Blur : A Gaussian blur is a convolution operation which is applied

to the original fingerprint image to reduce image noise (introduced by sensor).

The Gaussian kernel used for blurring is given by:

G(x, y) =1

2πσe−

x2+y2

2σ2 (2.1)

where σ is the variance of the gaussian distribution, x is the distance from the

2.3. FINGERPRINT IMAGE ENHANCEMENT 17

origin along the horizontal axis, y is the distance from the origin along the

vertical axis.

• Sliding-window Contrast Adjustment : Sliding-window contrast adjustment is

used to compensate for any lighting inconsistencies within a fingerprint and

to obtain contrast consistency among different fingerprints. A m ×m window

is centered on each pixel of the Gaussian blurred image. The corresponding

output pixel value is then calculated by finding the minimum and maximum

intensity values within the window and by using:

O(i, j) = (I(i, j)−mini,j)(

255

maxi,j −mini,j

)(2.2)

where I(i, j) and O(i, j) are the input and output pixel intensity values respec-

tively, and mini,j and maxi,j are the minimum and maximum pixel intensity

values within the m×m window centered at pixel (i, j).

• Histogram-based Intensity Level Adjustment : This final step is used to further

enhance the ridges and valleys. The image’s histogram is examined to determine

two intensity values: a lower threshold L and an Upper threshold U . The

intensity value (I(i, j)) of each pixel is processed using these thresholds to obtain

the output pixel intensity (O(i, j)) which is given by the following equation :

O(i, j) =

255 if I(i, j) >U

0 if I(i, j) <L

(I(i, j)−L)(

255U−L

)else

(2.3)

2.3 Fingerprint Image Enhancement

The performance of fingerprint feature extraction and matching algorithms relies

heavily on the quality of the input fingerprint images. Due to various factors such as


skin conditions (e.g., wet, dry, cuts, scars and bruises), non-uniform finger pressure,

noise introduced by sensor and inherently poor-quality fingers (e.g, manual workers,

elderly people), a significant percentage of fingerprint images is of poor quality. In-

fact, a single fingerprint image may contain regions of good, medium, and poor quality.

Thus an enhancement algorithm which can improve the quality of ridge structure is

necessary.

A survey on different enhancement techniques can be found in [35]. The most

widely used technique for fingerprint image enhancement is based on contextual filters.

The parameters of these filters change according to the local context i.e. local ridge

frequency and orientation. Such a filter can capture the local information and can

use them to efficiently remove the undesired noise (i.e. fill small ridge breaks, fill

intra-ridge holes, and separate parallel touching ridges) and preserve the true ridge

and valley structure. The filters themselves may be defined in spatial or in the Fourier

domain.

This section describes a popular enhancement algorithm by Sharat et al. [16],

which uses contextual filtering in Fourier domain. The algorithm consists of two

stages. The first stage consists of STFT (Short Time Fourier Transform) analysis

and the second stage performs the contextual filtering.

The STFT analysis stage yields the ridge orientation image (O(x, y)), the ridge

frequency image (F (x, y)) and the region mask (R(x, y)). The orientation image

represents the instantaneous ridge orientation at every point in the fingerprint image.

The frequency image indicates the average inter ridge distance distance within a local

region. The region mask indicates the foreground regions of the image where ridge

structures are present. During STFT analysis, the image is divided into overlapping

windows. This is done based on the assumption that the image has a consistent


orientation and frequency within a small local region. This assumption however is

not true for regions with singularities such as core, delta etc. The Fourier spectrum

within each window is analyzed and a probabilistic approximation of the dominant

ridge orientation and frequency within each window are obtained. An energy map

E(x, y) is also obtained during STFT analysis where each value indicates the energy

content of the corresponding block. This energy map can be used as a region mask

to distinguish between the foreground and background regions (the background and

noisy regions are characterized by very little energy content in the Fourier spectrum.).

The orientation image (O(x, y)) is then used to compute the Coherence Image which

contains coherence values of the various regions within the fingerprint. The coherence

value is low in regions with points of singularities (core, delta etc.). The coherence

image is then used to adapt the angular bandwidth of the directional filter.

The resulting contextual information from STFT analysis is then used to filter

each overlapping window (B) in the Fourier domain. The filter used is separable in

radial (Equation 2.5) and angular (Equation 2.6) domain and is given by

H(r, φ) = Hr(r)Hφ(φ) (2.4)

Hr(r) =

√[(rrBW )2n

(rrBW )2n + (r2 − r2c )

2n

](2.5)

H(φ)(φ) =

cos2 π2

(φ−φc)φBW

if |φ| < φBW

0 otherwise

(2.6)

where rBW and φBW are the radial bandwidth and angular bandwidth, respectively;

rc and φc are the mean frequency and mean orientation, respectively. The enhanced

block B′

is obtained by:

F = F ∗H(r, φ) (2.7)

2.4. FEATURE EXTRACTION 20

B′= FFT−1(F ) (2.8)

Finally, the results of each analysis window is tiled to obtain the enhanced image.

2.4 Feature Extraction

Chapter 1 introduced the fingerprint features at different Levels. The feature

extraction technique for minutiae points (bifurcation and endings), pores and ridge

contours is described in this section.

2.4.1 Minutiae Extraction

Most of the existing minutia extraction techniques trace the fingerprint skeleton

to find the different type of minutia points. The ridge bifurcation and ridge endings

are the most prominent minutiae types and have been extensively used by matching

algorithms. The flowchart of a typical minutia extraction algorithm is depicted in

Figure 2.3. It consists of following stages:

• Orientation Estimation : A fingerprint image is a oriented texture pattern and

the ridge orientation at a pixel (x, y) is the angle that the ridges within a small

neighborhood centered at (x, y), form with the horizontal axis. The fingerprint

image is first divided into a number of non-overlapping blocks. An analysis of

grayscale gradients within a block is done to estimate the representative ridge

orientation within that block. Different approaches such as optimization[61],

averaging [39] or voting[47] could be used to determine the block orientation.

• Segmentation : During this stage the portions of fingerprint image depicting


the finger (foreground) is segmented. This step is useful in order to avoid

the extraction of spurious features from background and noisy regions within

a fingerprint. The foreground and background can be differentiated by the

presence of an oriented pattern in the foreground and of an isotropic pattern

(i.e., one without a dominant orientation) in the background. The simplest

approaches segment the foreground by global or local thresholding. Since the

fingerprint background is not always uniform and lighter than foreground (due

to the presence of noise such as dust, grease on the sensor surface), a simple

approach based on local or global thresholding is not effective and more robust

segmentation techniques [45, 10, 60] are required.

• Ridge Detection : An important property of the ridges in a fingerprint image is

that the gray level values on ridges attain their local maxima along a direction

normal to the local ridge orientation [30]. The ridge pixels are identified based

on this property. The resulting ridge map often contains false ridges in the

form of holes and speckles (due to presence of noise, breaks, and smudges, etc

in the input image). The ridge map is cleaned (false ridges are removed) using

a connected component algorithm [57]. Finally, the ridges are thinned using

standard thinning algorithm [50].

• Minutiae Detection : The minutia points are then extracted from the thinned

ridge map by examining the 8-neighborhood of each ridge skeleton pixel. Let

(x, y) denote a pixel on a thinned ridge, and N0, N1, ..., N7 denote its eight neigh-

bors. A pixel (x, y) is a ridge ending if(∑7

i=0Ni

)= 1 and a ridge bifurcation

if(∑7

i=0Ni

)> 2. The minutiae points thus obtained in the above step may

contain many spurious minutiae. This may occur due to the presence of noise,


ridge breaks (even after enhancement) and image processing artifacts.

• Postprocessing : A number of heuristics are used to remove spurious minutiae.

False minutia points are generally obtained at the borders as the ridges ends

abruptly. These false minutiae at the borders can be recognized by analyzing

the number of foreground pixels in a region around minutia point. If number

of foreground pixels is relatively small then the minutia point can be removed.

Too many minutiae in a small regions, very close ridge ending (with orientations

anti-parallel to each other) and two very closely located bifurcations sharing a

common short ridge may indicate spurious minutiae and could be discarded

[27].

2.4.2 Pores Extraction

Pores are extremely fine details which are lost after the enhancement stage. There-

fore, for pore extraction the enhancement stage is omitted and pores are directly ex-

tracted from the preprocessed image. The pore extraction algorithm can be broadly

classified into two classes: the first class of algorithms extract pores by tracing fin-

gerprint skeletons, the second class of algorithms extract pores directly from gray

scale image. Stosz et al. [40] and Kryszczuk et al. [67] have proposed skeletonization

based approach for pore extraction. The skeletonization based approach is reliable

for extracting pores in good quality (and high resolution) images. As the image res-

olution decreases or the skin condition is not favorable, the method does not give

reliable results. In [29] Jain et al. have proposed a pore extraction technique directly

from gray scale image. The majority of existing approaches for pore extraction con-

sider only location information of pores for matching. The pores are distributed over


Figure 2.3: A typical minutiae extraction process ([30])


ridges and using orientation detail can provide additional information for matching.

A recent study [23] by the International Biometric Group has proposed a new ap-

proach for pore extraction which utilizes orientation information of pores along with

the location information. The approach proposed in [23] is presented in this section.

The first step in pore extraction process is the estimation of ridge orientation.

This data is utilized later in the representation of pores. The Local Ridge orientation

is determined by the least square estimate method [25]. The fignerprint image is

first divided into a number of non-overlapping blocks of dimension w × w. For each

pixel (i, j) in the preprocessed image, gradients in both the horizontal and vertical

direction (Equation 2.9 and Equation 2.10) are calculated by using a Sobel operator.

For each w × w block, the gradient along both the x and y directions is summed

(Equation 2.11 and Equation 2.12) and finally the arctangent is used to obtain the

representative orientation (Equation 2.13).

δx =(I(i− 1, j − 1) + 2I(i− 1, j) + I(i− 1, j + 1))−

(I(i+ 1, j − 1) + 2I(i+ 1, j) + I(i+ 1, j + 1))(2.9)

δy =(I(i− 1, j − 1) + 2I(i, j − 1) + I(i+ 1, j − 1))−

(I(i− 1, j + 1) + 2I(i, j + 1) + I(i+ 1, j + 1))(2.10)

Vx(i, j) =

i+w/2∑u=i−w/2

j+w/2∑v=j−w/2

2δx(u, v)δy(u, v) (2.11)

Vy(i, j) =

i+w/2∑u=i−w/2

j+w/2∑v=j−w/2

(δ2x(u, v)− δ2

y(u, v)) (2.12)

θ(i, j) =1

2arctan

(Vy(i, j)

Vx(i, j)

)(2.13)


Next the pores are extracted from the preprocessed image. The pores are first

enhanced by convolving the preprocessed fingerprint image with a Mexican hat ker-

nel. The Mexican hat mother wavelet is defined in Equation 2.14 and the daughter

wavelets are defined in Equation 2.15.

Ψ(i, j) =(1− x2 − y2

)ex2+y2

2 (2.14)

Ψa,b(λ) =1√a

Ψ

(λ− ba

)(2.15)

where a is the factor by which mother wavelet is scaled (dilated), b is the factor

by which the mother wavelet is translated (or shifted) and λ specifies the center of

daughter wavelet.

After the convolution step, the fingerprint image is thresholded with a single-point

threshold (T ). After this step the pores have an intensity of 255. The pores are then

extracted by a blob detector which locates groups of connected pixels (pores) with

an intensity of 255 and with size within a pre-determined range. Each pore thus

extracted is represented by the coordinates of the central pixel and an orientation θ,

which is the ridge orientation at that particular location.

2.4.3 Ridge Contour Extraction

Ridge contours can be extracted by using classical edge detection algorithms.

However, these algorithms are sensitive to creases, pores etc. and as a result the

detected edge contours are often very noisy (especially in low resolution images).

Jain et al. [29] have proposed an algorithm to extract the ridge contours which uses

a simple filter to detect ridge contours. The algorithm can be described as follows:

First, the image is enhanced by using Gabor filters [24]. Next, a wavelet transform

2.5. FINGERPRINT MATCHING 26

is applied to the original fingerprint image. The enhancement ridge contours are

obtained by using a linear subtraction of wavelet response and gabor enhanced image.

The resulting image is then binarized using a heuristically determined threshold.

Finally, the binarized image is convolved with a filter H (Equation 2.16) and ridge

contours are extracted.

r(x, y) =∑n,m

Ib(x, y)H(x− n, y −m) (2.16)

where filter H = (0, 1, 0; 1, 0, 1; 0, 1, 0) counts the number of neighborhood edge points

for each pixel. A pixel (x, y) is classified as a ridge contour pixel if r(x, y) = 1 or 2.

2.5 Fingerprint Matching

A variety of automatic fingerprint matching algorithms have been proposed in

the pattern recognition literature. This chapter provides a survey of existing ap-

proaches for automatic fingerprint matching. Most of these algorithms have no dif-

ficulty in matching good quality fingerprint images, but matching low quality and

partial fingerprints remains a challenging problem. The main approaches proposed

in the literature for fingerprint matching can be roughly classsified into three cate-

gories and they are correlation-based matching , minutiae-based matching and ridge

feature-based matching. In correlation-based fingerprint matching, the template and

query fingerprint images are spatially correlated to estimate the degree of similar-

ity between them. The minutia-based techniques essentially consists of finding the

alignment between the query and template minutia points. The ridge feature-based

techniques rely on various features of fingerprint ridge pattern such as ridge shape,

texture information, local orientation and frequency. A very good literature survey

on fingerprint recognition can be found in [35].


2.5.1 Correlation-based Techniques

Let T and Q denote the template and query fingerprint images, respectively. The

sum of squared differences (SSD) between the intensities of corresponding pixels in

the two images, can be used as a measure of the diversity between them.

SSD(T,Q) = ||T −Q||2 = (T −Q)t(T −Q) = ||T ||2 + ||Q||2 − 2T tQ (2.17)

In the above equation, the superscript “t” denotes the transpose operation on a vector.

The cross-correlation (CC) between T and Q is given by, CC(T,Q) = T tQ. If the

terms ||T ||2 and ||Q||2 in above equation are constants, the diversity (SSD(T,Q))

becomes inversely proportional to cross-correlation, which constitutes the third term

(−2.CC(T,Q)) in Equation 3.1. The cross-correlation then becomes a measure of

similarity. Two fingerprint impressions from same finger differ due to many factors,

as discussed in the beginning of this chapter. Thus, additional steps are required

before similarity between T and Q can be calculated.

Let Q(∆x,∆y,θ) represent a transformation of the query image, where ∆x and ∆y

are translation parameters along x and y direction, and θ is the rotation parameter.

Them a similarity measure can be given by:

S(T,Q) = max∆x,∆y,θ

CC(T,Q(∆x,∆y,θ)) (2.18)

S(T,Q) only considers the rotation and translation factors and thus is not a very

accurate measure of similarity. This method fails if the images are highly distorted.

The distortion is more pronounced in global fingerprint patterns, thus considering

local regions can minimize distortion to some extent. [11, 51] presents some of the

approaches to localized correlation based matching. Also variable finger pressure and

skin condition cause image brightness, contrast and ridge thickness to vary across dif-

ferent acquisitions of same finger. The use of more sophisticated correlation measures


such as the normalized cross-correlation or the zero-mean normalized cross-correlation

may compensate for contrast and brightness variations and applying a proper combi-

nation of enhancement, binarization, and thinning steps (performed on both T and

Q) may limit the ridge thickness problem [35]. The computation of maximum corre-

lation (S(T,Q)) in spatial domain is very expensive. The computational complexity

can be reduced and translation invariance can be achieved by calculating correlation

in fourier domain [18] .

T ⊗Q = F−1(F ∗(T )× F (Q)), (2.19)

where ⊗ denotes the correlation in spatial domain, ∗ denotes the complex conjugate,

× denotes the point-by-point multiplication of two vectors, F (.) and F−1(.) denote

the Fourier transform and inverse Fourier transform, respectively. Rotation has to be

dealt with separately in this technique. Fourier-Mellin transform [68] can be used to

achieve both rotation and translation invariance.

2.5.2 Minutiae-based Methods

Let T and Q be the feature vectors, representing minutiae points, from the tem-

plate and query fingerprint, respectively. Each element of these feature vectors is

a minutia point, which may be described by different attributes such as location,

orientation, type , quality of the neighbourhood region, etc. The most common rep-

resentation of a minutia is the triplet x, y, θ, where x, y is the minutia location and θ

is the minutia angle. Let the number of minutiae in T and Q be m and n, respectively.

T = m1,m2, .......,mm, mi = xi, yi, θi, i = 1...m

Q = m′

1,m′

2, .......,m′

n, m′

j = x′

j, y′

j, θ′

j, j = 1...n


A minutia mi in T and m′j in Q are considered matching, if following conditions are

satisfied:

sd(m′

j,mi) =√

((x′j − xi)2 + (y

′j − yi)2) ≤ r0 (2.20)

dd(m′

j,mi) = min(|θ′

j − θi|, 360− |θ′

j − θi|) ≤ θ0 (2.21)

Here, r0 and θ0 are the parameters of the tolerance window which is required to com-

pensate for errors in feature extraction and distortions caused due to skin plasticity.

The number of “matching” minutia points can be maximized, if a proper align-

ment (registration parameters) between query and template fingerprints can be found.

Correctly aligning two fingerprints requires finding a complex geometrical transfor-

mation function (map()), that maps the two minutia sets (Q and T ). The desirable

characteristics of map() function are: it should be tolerant to distortion, it should re-

cover rotation, translation and scale1 parameters correctly. Let match() be a function

defined as:

match(m′′

j ,mi) =

1 if m′′j and mi satisfy (2.20), (2.21)

0 otherwise(2.22)

where, map(m′j) = m

′′j . Thus, the minutia matching problem can be formulated as

[35]:

maxP

m∑i=1

match(map(m′

P (i)),mi) (2.23)

where P () is the minutia correspondence function that determines the pairing between

the minutia points in Q and T . A minutiae-based matching algorithm thus attempts

to find the appropriate mapping function (map()) and correspondence function (P ()),

so that the number of matching minutia points, between T and Q, can be maximized.

1scale is considered when the fingerprints are captured at different resolutions


If either of P () or map() is known then solving Equation 2.23 becomes a very trivial

task. But, in practice, neither map() nor P () is known apriori. The errors in minu-

tia extraction (spurious, missing minutia and measurement errors) further make the

minutia matching problem very hard.

A number of minutia matching algorithms have been proposed in literature. These

algorithms can be broadly classfied as:

• Global Matching : This approach tries to simultaneously align all the minu-

tia points. The alignment could be either implicit or explicit. The Implicit

alignment technique tries to find the point correspondences and in the process

optimal alignment is obtained. The explicit alignment technique on the other

hand explicitly aligns the minutia sets first, before finding the point correspon-

dences.

• Local Matching : This approach tries to match local minutia structures; local

structures are characterized by attributes that are invariant with respect to

global transformations.

The local versus global matching is a trade-off among simplicity, low computational

complexity, high distortion tolerance (local matching), and high distinctiveness (global

matching). Local matching techniques are more robust to non-linear distortion and

partial overlaps when compared to global approaches. Matching local minutiae struc-

tures relax global spatial relationships which are considered to be highly distinctive,

and therefore reduce the amount of information available for discriminating finger-

prints.


2.5.2.1 Global Matching

The minutia matching problem has been addressed as point pattern matching

problem in literature. A number of approaches to point pattern matching exists in

literature. Some of the approaches are discussed here.

The relaxation approach [58] iteratively adjusts the confidence level of each cor-

responding pair based on its consistency with other pairs until a certain criterion is

satisfied. Let pij be the probability that point i corresponds to point j, and c(i, j;h, k)

be a compatibility measure between the pairing (i, j) and (h, k). At each iteration r,

pij is incremented if it increases the compatibility of other points and is decremented

otherwise.

p(r+1)ij =

1

m

m∑h=1

[maxk=1..n

c(i, j;h, k).p(r)ij

], i = 1..m, j = 1..n (2.24)

At convergence, each point i is associated with a point j such that pij = maxs pis.

The iterative nature of this approach makes it slow and unsuitable for automatic

matching.

Hough transform based approach [66] are quite popular for minutia matching.

This approach converts point pattern matching to a problem of detecting the highest

peak in the hough space of transformation parameters. The hough space of transfor-

mation parameters consists of all the possible values of parameters under the assumed

distortion model. Ratha et al. [59] have used a transformation space consisting of

quadruples (∆x,∆y, θ, s), representing translation along x direction, translation along

y direction, rotation and scale parameters, respectively. To avoid computation com-

plexity the search space is discretized into a finite set of values:

∆x ∈ ∆x1,∆x2, ....,∆xa

∆y ∈ ∆y1,∆y2, ....,∆yb


θ ∈ θ1, θ2, ...., θc

s ∈ s1, s2, ...., sd

A four dimensional accumulator A of size (a × b × c × d) is maintained. Each cell

A(i, j, k, l) represents the likelihood of the transformation parameters (∆xi,∆yj, θk, sl).

The following algorithm is used to accumulate evidence:

for each mi in T

for each m′j in Q

for each θ ∈ θ1, θ2, ...., θc

if dd(θ′j + θ, θi) ≤ θ0

for each s ∈ s1, s2, ...., sd

∆x

∆y

=

xi

yi

− s. cos θ − sin θ

sin θ cos θ

x

′j

y′j

Let (∆x+,∆y+) be the quantization of (∆x,∆y) to the nearest bin

A [∆x+,∆y+, θ, s] = A [∆x+,∆y+, θ, s] + 1

At the end of the accumulation process, the best alignment transformation is

obtained as

(∆x∗,∆y∗, θ∗, s∗) = argmax(∆x+,∆y+,θ,s)A[∆x+,∆y+, θ, s

]A hierarchical Hough transform-based algorithm [38] may be used to reduce the size

of the accumulator array by using a multi-resolution approach.


It can be shown that if a perfect pre-alignment could be achieved, the minutiae

matching problem could be reduced to a simple pairing problem. Most minutia-based

matchers first transform (register) the input and template fingerprint features into

a common frame of reference. The parameters of alignment are typically estimated

either by (i) superimposing singular points in the fingerprints, e.g., core and delta

segments; (ii) correlating the orientation images; or (iii) by correlating ridge features

(e.g., length and orientation of ridges). Jain et al [34] have proposed an alignment

based minutia matching approach that uses ridge features for alignment and an adap-

tive elastic string matching algorithm [19] for matching the pre-aligned minutia sets.

In their approach, each minutia is associated with the ridge on which it resides. The

ridge is represented as a planar curve with origin coincident with the minutia location

and x-axes along the minutia direction. The global transformation parameters (∆x,

∆y and θ) are caculated from a pair of matching ridges. Finding such a pair involves

iteratively matching pairs of ridges, untill a pair is found whose matching degree ex-

ceeds a certain threshold. The pair found is then used for aligning the query and

template minutia sets. The minutia corresponding to the matching ridges are used as

reference minutia points during matching stage. Each minutia in Q and T is converted

to a polar coordinate system with respect to the reference minutia in its set. Both

Q and T are transformed into a string of minutia points, ordered in increasing order

of radial angle. The strings are matched using a dynamic programming technique

to find their edit distance. The use of an adaptive tolerance window for matching

minutia points tolerates local distortion. And matching minutiae representations by

their edit distance tolerates missing and spurious minutiae. However, an exhaustive

reordering and matching is required to deal with the problem of “order flips”, which

may be caused due to measurement errors introduced during feature extraction.


The alignment based approaches are not very accurate, as reliably extracting

singular points from low quality images is difficult. They suffer when image is of poor

quality or is highly distorted. In such cases global registration parameters does not

exist and errors introduced during alignment step lead to errors in further steps.

2.5.2.2 Local Matching

The local matching approaches rely on evidence accumulated from matching local

structures in neighbourhood of minutia points. These local structures are generally

characterized by properties that are invariant with respect to global transformation,

and therefore are suitable for matching without any apriori alignment. However, local

neighborhoods do not sufficiently capture the global structural relationships thereby

making false accepts very common. Thus, it is possible that local minutia structures

in two non-matching fingerprints might match. Figure 2.4 shows an example of two

false matched local structures. They are similar at the local structures but conflict

with each other in global context (at very different locations with respect to the core

and delta points).

To deal with this problem, the local matching algorithms use an additional consol-

idation step to check whether the locally matched minutia points match at global level

or not. A large number of local matching techniques has been proposed in literature.

An overview of some of the algorithms is discussed here.

Jiang and Yau [37] use a local structure formed by a central minutia and its two

nearest-neighbour minutiae; the feature vector vi associated with the minutia mi,

whose nearest neighbors are mj and mk is :

vi = [dij, dik, θij, θik, φij, φik, nij, nik, ti, tj, tk]

where dab is the distance between minutiae ma and mb, φab is the direction difference


Figure 2.4: An example of two false matched local structures ([36])

between the angle θa of ma and the direction of the edge connecting ma to mb, nab is

the ridge count between ma and mb, and ta is the minutia type of ma . For all minutia

pairs (mi,m′j), mi ∈ T and m

′j ∈ Q, a weighted distance between their vectors vmi

and vm′j

is calculated. An additional consolidation step is used to enforce the result of

local matching. The best matching pair (with least distance ) is used for registering

the two minutia sets. The feature vectors of the remaining aligned pairs are matched

and a final score is computed by taking into account contributions from first stage

and consolidation stage.

Ratha et al. [26] have proposed a “star” representation for capturing local minutia

structure in the form of a minutia adjacency graph (MAG). The star associated with

a minutia mi is the graph Gi = (Vi, Ei) consisting of the set of vertices Vi containing

all minutia points mj whose distance (dij) from mi is less than a predefined threshold,

and the set of edges Ei contining edges from mi to all vertices in Vi. Each edge eij ∈ Ei

is labelled with a 5-tuple (mi,mj, dij, rcij, φij), where rcij is the ridge count between


mi and mj and φij is the angle subtended by the edge with the x-axis. During local

matching, each “star” in Q is matched with each star in T , the matching is performed

by clockwise traversing the corresponding graphs in an increasing order of radial angle.

At the end of local matching stage, a set TOP of best matched star pairs is returned

and these pairs are further checked for consistency during the consolidation stage. A

pair of stars is consistent if their spatial relationships (distance and ridge count) with

a minimum fraction of the remaining stars in TOP are consistent.

Sharath et al. [15] define a local structure representation called K − plet that is

invariant under translation and rotation. The K − plet consists of a central minutiae

mi and K other minutiae m1,m2...mK chosen from its local neighbourhood. Each

neighbourhood minutiae mj is defined as a 3-tuple (φij, θij, rij), where rij represents

the Euclidean distance between mi and mj, θij is the relative orientation of mj with

respect to the central minutiae mi, and φij represents the direction of the edge con-

necting mj and mi, with respect to the orientation of central minutia. A given finger-

print is represented as a directed graph G(V,E). V is the set of vertices containing

all minutia points and E is the set of directed edges containing all possible (mi,mj)

pairs, where mi and mj are neighbouring minutia points. Each vertex v is denoted

by a 4-tuple (xv, yv, θv, tv) representing the co-ordinate, orientation and type of minu-

tiae. Each directed edge (u, v) is labelled with corresponding K − plet co-ordinates

(ruv, φuv, θuv). Let G(V,E), H(V′, E

′) be the graphical representation for T and Q,

respectively. The matching algorithm is based on matching a local neighbourhood

(K-plets) and propagating the match to the K-plet of all the minutiae in the neigh-

bourhood successively. The minutia points in K-plet are arranged in increasing order

of radial distance rij and a dynamic programming based string alignment algorithm

[19] is used for local matching. The consolidation of the local matches is done by a


Coupled Breadth First Search (CBFS) algorithm that propagates the local matches

simultaneously in both the fingerprints. The CBFS algorithm requires two vertices

vi ∈ T and vj ∈ Q as the source nodes from which to begin the traversal. The CBFS

algorithm is executed for all possible correspondence pairs (vi, vj). The pair with

maximum number of matches is used to compute the final matching score.

2.5.3 Ridge Feature-based Matching Techniques

There are various reasons which induce designers of fingerprint recognition tech-

niques to look for features, beyond minutiae. Minutiae based matching algorithms do

not perform well in case of small fingerprints that have very few minutiae. Reliable

extraction of minutiae from poor quality fingerprints is very difficult. Furthermore,

registering minutiae representation is very challenging. The most commonly used al-

ternative features are i) global and local texture information, and ii) Level 3 features

2.5.3.1 Texture Feature based Techniques

Global and local texture information are important alternatives to minutiae. Tex-

tures are defined by spatial repetition of basic elements, and are characterized by

properties such as scale, orientation, frequency, symmetry, isotropy, and so on. Fin-

gerprint ridge lines are mainly described by smooth ridge orientation and frequency,

except at singular regions. These singular regions are discontinuities in a basically

regular pattern and include the loop(s) and the delta(s) at a coarse resolution and the

minutiae points at a high resolution. Global texture analysis fuses contributions from

different characteristic regions into a global measurement and, as a result, most of

the available spatial information is lost. Local texture analysis has proved to be more

effective than global feature analysis; although most of the local texture information


is carried by the orientation and frequency images, most of the proposed approaches

extract texture by using a specialized bank of filters.

Jain et al. [31] have proposed a filter bank based local texture analysis technique.

The fingerprint image is tesselated into 80 cells (5 bands and 16 cells) around the

core point and a feature vector representing texture information is obtained from the

tesselation. The feature vector consists of an ordered enumeration of the features

extracted from the local information contained in each sector. Thus the feature

elements capture the local texture information and the ordered enumeration of the

tesselation captures the global relationship among the local contributions. A bank of

8 gabor filters (8 orientations and 1 scale = 1/10) is used to obtain texture informtion

from each sector. Thus each fingerprint is represented by a 640 (80 × 8) fixed-size

feature vector called the FingerCode. The generic element Vij of the vector (i = 1..80

is the cell index, j = 1..8 is the filter index) denotes the energy revealed by the filter

j in the cell i, and is computed as the average absolute deviation (AAD) from the

mean of the responses of the filter j over all the pixels of the cell i. Matching two

fingerprints is then performed by computing the Euclidean distance between their

Finger-Codes. The disadvantage of this approach is that it uses a core points as a

reference. When the core points cannot be reliably detected, or it is close to the border

of the fingerprint area, the FingerCode of the input fingerprint may be incomplete or

incompatible with the template. Also, fingerCodes are found to be not as distinctive

as minutiae. However, they carry complementary information which can be combined

with minutiae to yield higher accuracy. In [28] and [64] a variant of above method

has been proposed where the two fingerprints to be matched are first aligned using

minutiae and tessellation is performed over a square mesh grid.

Nanni et al.[52] have proposed a hybrid approach where fingerprints are pre-


aligned using minutiae, and then texture features are extracted by invariant local

binary patterns (LBP) from the fingerprint image convolved with Gabor filters. The

fingerprint image is first divided into several sub-windows, then a segmentation step

is performed in order to discard the background, finally each sub-window is convolved

with a bank of Gabor filters and LBP are calculated. The comparison between the

unknown fingerprint and the stored template is performed by the average Euclidean

distance calculated among the correspondent couples of foreground sub-windows.

2.5.3.2 Level 3 Features-based Techniques

The use of Level 3 features in an automated fingerprint identification system has

been studied by only a few researchers. Existing literature is exclusively focused

on the extraction of pores in order to establish the viability of using pores in high

resolution fingerprint images to assist in fingerprint identification.

Stosz and Alyea [67] have presented a technique which utilizes pore information,

extracted from high-resolution fingerprint images, to augment matching processes

that use only Level 2 data. The images were taken by a custom built optical/electronic

sensor in a controlled environment. An enrollment process is required in which singu-

lar points (used as origin), minutia points, and characteristic regions containing pores

are selected manually and stored in template. Matching is initiated by determining

the origin of the input fingerprint which is defined as the position of maximum correla-

tion between the origin stored in the template and the input binary fingerprint. Next,

the remaining binary image segments from the template are compared to the input

fingerprint. Two pores are considered matched if they lie within a certain bounding

box. Finally, experimental results based on a database of 258 fingerprints from 137

individuals showed that by combining minutia and pore information, a lower FRR


of 6.96 percent (compared to 31 percent for minutiae alone) can be achieved at a

FAR of 0.04 percent [67]. Later, Roddy and Stosz [63] conducted a statistical anal-

ysis of pores and predicted the performance of a pore-based automated fingerprint

system. The study demonstrated the efficacy of using pores, in addition to minutiae,

for improving the recognition performance.

Kryszczuk, et al. [40, 41] have conducted research to determine whether Level 3

features could compensate for decreased number of Level 2 features when attempting

to match with fingerprint fragments, given sufficiently high-resolution data. The ex-

periments are done on fingerprints acquired from a high resolution (2000 dpi) custom-

built scanner. The size of database used was small (12 genuine and 6 impostor com-

parisons). Pores and ridge structure are used in conjunction with Level 2 features for

comparison of fragmentary fingerprint images and comparison has been done based

on a geometric distance criterion. The authors have presented two hypotheses: i) as

the size of the fingerprint fragment decreases, or the number of minutiae decreases,

the usefulness of Level 3 data increases, and ii) given a sufficiently high resolution,

the same quantity of discriminative information could be extracted from fingerprint

fragments by using Level 3 data, as could be when considering Level 2 data extracted

from the complete fingerprint image.

Jain, et al. [29, 33] have proposed a hierarchical matching system that utilizes

all the three levels extracted from 1000 ppi fingerprint images. The level 3 features

(pores and ridge contours) are locally matched, in windows associated with matched

minutiae points, using the Iterative Closest Point (ICP) [13] algorithm. There is

a relative reduction of 20 percent in the EER when Level 3 features are employed

in combination with Level 1 and 2 features. This significant performance gain was

consistently observed across various quality fingerprint images. The iterative nature


of ICP algorithm makes the approach unsuitable for automated matching. Also, the

Level 3 matching (ICP algorithm) relies on an initial alignment based on Level 2

features and in case of distorted fingerprints such an alignment is not reliable.

In [70] Vatsa et al. have presented a score-level fusion technique that combines

level-2 and level-3 match scores to provide high accuracy. The match scores obtained

from Level 2 and Level 3 classifiers are first augmented with a quality score that

is quantitatively determined by applying redundant discrete wavelet transform to a

fingerprint image. The quality augmented match scores are then fused using Dezert-

Smarandache theory. The proposed fusion method provided better results than other

existing fusion techniques. The proposed algorithm is able to perform well even in

the presence of imprecise, inconsistent, and incomplete fingerprint information.

42

Chapter 3

PROPOSED HIERARCHICAL

MATCHER

Fingerprint Matching is the most important stage in fingerprint recognition pro-

cess. A fingerprint matching algorithm compares two sets of features originating from

two fingerprints and determines whether or not they represent the same finger. Fin-

gerprint matching is an extremely difficult problem, mainly due to the large intra-class

variations that exists in different impressions of the same finger. The main factors

responsible for these intra-class variations are [35]: i) Displacement and Rotation :

The same finger may be placed at different locations and at different orientation on

the sensor during different acquisitions resulting in a (global) translation and rotation

of the fingerprint area, ii) Partial overlap : Finger displacement and rotation often

cause part of the fingerprint area to fall outside the sensor’s “field of view”, resulting

in a smaller overlap between different impressions of the same finger, iii) Non-linear

distortion : The act of sensing maps the three-dimensional shape of a finger onto the

two-dimensional surface of the sensor. This mapping results in a non-linear distortion

43

in successive acquisitions of the same finger due to finger skin plasticity, iv) Pressure

and skin condition: The ridge structure of a finger would be accurately captured if

ridges of the part of the finger being imaged were in uniform contact with sensor sur-

face. However, finger pressure, dryness of the skin, skin disease, sweat, dirt, grease,

and humidity in the air all confound the situation, resulting in a non-uniform contact.

As a result, the acquired fingerprint images are very noisy, and v) Feature extraction

errors : The feature extraction algorithms are imperfect and often introduce measure-

ment errors. For example, in low-quality fingerprint images, the minutiae extraction

process may introduce a large number of spurious minutiae and may not be able to

detect all the true minutiae.

Chapter 2 has provided an overview of different fingerprint matching approaches.

The correlation based method requires the complete image to be stored (large tem-

plate sizes). The texture based methods are less accurate than minutiae based match-

ers since most regions in the fingerprint carry low textural content. Both types of

methods requires accurate alignment of fingerprints. The minutia based techniques

on the other hand are more accurate and they very closely resemble the manual ap-

proach as used by forensic experts. Studies [31], [70, 67, 29, 33] have shown that

by combining additional information in the form of texture features, level 3 features

with minutia based matcher, higher accuracy can be achieved. The minutia based

approaches like other approaches cannot give a high confidence match when the im-

ages are of poor quality or when there is a very small overlap i.e. very few minutia

points are available for matching. Level 3 features are known to carry discriminative

information and forensic examiners often make use of Level 3 features when insuffi-

cient minutia points are present. A new hierarchical matching system which utilizes

additional information in form of Level 3 features (pores and ridge contours) has

3.1. PREPROCESSING AND ENHANCEMENT 44

been proposed in this chapter. Figure (3.1) illustrates the architectural design of the

proposed hierarchical system. A similar architecture (hierarchical) was first proposed

by Jain et al. [29].

Figure 3.1: The proposed hierarchical matcher.

3.1 Preprocessing and Enhancement

To compensate for the variations in lighting, contrast and other inconsistencies

three preprocessing steps are used: Gaussian blur, sliding window contrast adjust-

ment, and histogram based intensity level correction. Gaussian blurring is used to

remove any noise introduced by the sensor. The lighting inconsistencies are ad-

justed by using sliding-window contrast adjustment on the Gaussian blurred image.


To further enhance the ridges and valley a final intensity correction is made by us-

ing Histogram-based Intensity Level Adjustment. These techniques are described in

Chapter 2.

The preprocessed image is enhanced using a popular fingerprint enhancement

technique by Sharat et al.[16] which uses contextual filtering in fourier domain. The

technique is discussed in Chapter 2. The enhanced fingerprint image is not suitable

for extracting pores as the pore information is lost during enhancement. Thus for

pore extraction the preprocessed image is used.

3.2 Feature Extraction

The minutiae points are extracted from the enhanced image by using NIST’s 1

minutia extraction software (NFIS). The method first generates the image quality

maps by checking the regions with high curvature, low flow and low contrast. And

then, a binary representation of the fingerprint is constructed. Minutiae are generated

by comparing each pixel neighborhood with a family of minutiae templates. Finally,

spurious minutiae are removed by using a set of heuristic rules. The NFIS also counts

the neighbor ridges and assigns each minutia point a quality (in the range 0 to 100)

determined from image quality maps. The minutia representation generated by NFIS

consists of the location information, orientation, minutia type (bifurcation or ending)

and minutia quality. The proposed minutia matcher does not differentiate between

different minutiae types. This is because the minutia types are difficult to distinguish

when the applied finger pressure during acquisition varies. Figure 3.2 shows one such

example where the same minutiae extracted from two different impressions appears

1NIST stands for National Institute of Standards and Technology and is a federal technologyagency that develops and promotes measurement, standards, and technology.

3.3. HIERARCHICAL MATCHING 46

as a bifurcation in one image and as a ridge ending in other.

Figure 3.2: Effect of differential finger pressure([36]).

The Level 3 features are extracted only when the minutiae based matching fails

to match the query fingerprint image with the template image. The ridge contours

are extracted from the enhanced fingerprint image by using the approach proposed

by Jain et al. [29]. The extracted ridge contours by using this approach is shown

in Figure 3.3. For extracting pores the technique proposed in [23] by International

Biometric Group is used. Figure 3.2 show the extracted pores by using this approach.

The pore information thus extracted contains the location information and orientation

information of pores. The pore extraction and ridge contour extraction techniques

are described in Chapter 2.

3.3 Hierarchical Matching

Dealing with Non-Linear distortion is the major problem faced by fingerprint

matching algorithms. The effect of non-linear distortion is more pronounced when

global structures with global parameters are considered for matching. Localized

matching can minimize the effect of non-linear distortion. Throughout our approach

we have used localized matching in-order to tolerate the effects of non-linear distor-

tion. Also we have used rotation invariant structures (in case of pores and minutiae)


(a) Sample 500 dpi fin-

gerprint

(b) enhanced fingerprint (c) Traced Ridge Contours

Figure 3.3: Ridge Contour extraction process in a image scanned with Cross MatchVerifier 300 scanner at 500 dpi.

(a) Sample 500 dpi finger-

print

(b) extracted pore

Figure 3.4: Pores extracted from a image scanned with Cross Match Verifier 300scanner at 500 dpi.


and features (in case of ridge contours) for matching, thus at any stage any type of

alignment is not required.

Let T and Q be the template and query images which are to be matched. At first

the minutia-based matcher matches T and Q and returns the matching minutia pairs,

which is a measure of degree of similarity between the two representations (finger-

prints). Let S2 be the number of matched minutia pairs returned by the minutia-based

matcher. If S2 > τ2 (τ2 is a threshold whose value is chosen as 12), the two finger-

prints are considered “matched” and the matching terminates. The threshold τ2 is

set to be 12 because typically, a match consisting of 12 minutia points (the 12-point

guidelines) is considered as sufficient evidence for making positive identification in

many courts of law [56]. However, If S2 ≤ τ2 then the matching continues and Level 3

based matcher further matches the two fingerprints and returns the final match/non-

match decision. The matched minutiae at Level 2 are further examined and the Level

3 features in their neighborhood are matched.

3.3.1 Minutia-based Matcher

We propose a local structure based minutia matcher. The local structure (“star”)

is similar to the “kplet” used by Sharat et al [15]. It consists of a central minutia

point mi and its k neighboring minutia points (mi1,mi2, ....mik) within a local neigh-

borhood. The local neighborhood is defined by a region within distance dmax from

the central minutia mi. The k neighbors are chosen amongst all the minutiae within

dmax distance and from all four quadrants . Choosing neighboring minutia from

all four quadrants captures local structure information accurately and during exper-

iments we found that it is more effective than choosing k nearest neighbors of the

central minutia. Figure 3.5 demonstrates a star with k = 6 neighbors. There exists a


trade-off while choosing the size of local neighborhood (dmax). dmax should be small

enough to tolerate distortion, however keeping dmax too small might not accumulate

enough evidence and hence may lead to false matches. k represents the number of

neighboring minutia points which forms the “star” and its value can vary between

a predefined range bounded by kmin and kmax i.e. kmin ≤ k ≤ kmax. If for a given

minutia k < kmin then that minutia is not considered for matching. On the other

hand, if k > kmax then kmax neighbors are considered. The matching is performed in

two stages:

Figure 3.5: The schematic description of a “star” with k = 6 neighbors. The neighborsare chosen from all the four quadrants

1. Local matching : the “stars” corresponding to central minutia points are matched.

At the end of this stage pairs of matching minutia points are returned.


2. Consolidation step : During this step the matched minutia pairs are further

checked for consistency at global level.

3.3.1.1 Local Structure Matching

During this stage the “stars” in Q are matched with the “stars” in T . If two

stars are found to be matching then the minutiae representing their centers are also

considered to be matching. A star associated with a central minutia m and with k

neighboring minutiae is represented by a k element feature vector vm. Each element

vmi ∈ vm is a 3-tuple (ri, θi, φi) representing the radial parameters of ith neighboring

minutia point with respect to the central minutia m.

vm = (r1, θ1, φ1), ....., (rk, θk, φk)

where ri is the distance of ith neighboring minutia from m, θi is the orientation

of ith neighboring minutia with respect to m’s orientation and φi is the direction

difference between m’s orientation and the direction of the edge connecting m and

mi. Throughout this chapter we shall refer to a “star” by its feature vector. The

“star” structure is invariant to rotation and translation and thus any type of alignment

is not required. The local structure matching problem thus reduces to matching two

ordered sequences

vm = (vm1, ....vmk) = (r1, θ1, φ1), .., (rk, θk, φk) and

vm′ = (vm′1, ....vm′k′ ) = (r′

1, θ′

1, φ′

1), .., (r′

k, θ′

k, φ′

k)

where vm represents a star in T and vm′ represents a star in Q.

The edit distance is used as a similarity measure between two “stars”. Edit

Distance between two sequences indicates a measure of similarity between them. The


edit distance is capable of capturing the impression variations (deletion of genuine

minutiae, insertion of spurious minutiae, and perturbation of minutia) associated with

different impressions of same finger. The algorithm considers a query minutiae (in

the query “star”) and a template minutiae (in the template “star”) to be a mismatch

if their attributes (radius, radial angle and minutia direction with respect to the

central minutiae) are not within a tolerance window. A penalty is associated with

every match and mismatch (Equation 3.2). The penalty associated with a match is

proportional to the disparity in the values of the attributes related to the two matched

minutiae. On the other hand, a predefined huge penalty is associated with every mis-

match. The sum of all penalties associated with match and mismatch between two

sequences defines the edit distance. Among several possible sets of transformation of

the query sequence into the template sequence, the string matching algorithm chooses

the transform associated with the minimum cost (edit distance) based on dynamic

programming.

An unmatched “star”, vm′ ∈ Q is matched with all the unmatched “stars” in

T and the “star” which returns the minimum edit distance is chosen as the best

match. Given two “stars”, vm = (vm1, ....vmk) ∈ T and vm′ = (vm′1, ....vm′k′ ) ∈ Q, of

lengths k and k′, respectively, the edit distance, D(k, k

′), is recursively defined with

the following equations :

D(i, j) =

0 if i = 0, or j = 0

min

D(i− 1, j) + Ω

D(i, j − 1) + Ω

D(i− 1, j − 1) + w(i, j)

0 < i ≤ k, and 0 < j ≤ k′

(3.1)


w(i, j) =

α∆r + β∆θ + γ∆φ if∆r < tr, ∆θ < tθ and ∆φ < tφ

Ω otherwise

(3.2)

where, ∆r = (|ri− r′j|/min(ri, r

′j)), ∆θ = |θi− θ

′j| and ∆φ = |φi−φ

′j|; α, β and γ are

weights associated with each component, respectively; tr, tθ and tφ are the parameters

of the tolerance window ; and Ω is a pre-specified penalty for mismatch.

Once the best match for vm′ is found, say vn = (vn1, ....vnl), then a dynamic

programming based sequence alignment technique ([53]) is used to find the number

of matching edges between vm and vn′ . Sequence alignment tries to find the best

alignment between an entire sequence S1 and another entire sequence S2. Consider

two sequences S1 = [GAATTCAGTTA] and S2 = [GGATCGA]. If required gaps( )

are inserted in S1 and T1 and finally the aligned strings have same length. A penalty

is associated with each gap and mismatch and a reward is associated with each match

in the final alignment. Based on these “rewards” and “penalty” a final alignment cost

is calculated. The optimal global alignment is one having the maximum cost.

S1′= GAATTCAGTTA

S2′= GGA TC G A

There are two phases in the alignment process:

• Forward phase: The optimal alignment cost for every substring is computed

(by using a recurrence relation) and stored in C(a (k′+ 1)× (l+ 1) dimensional

matrix). C[i, j] denotes the best cost for aligning two subsequences v′m[1...j]

(1 ≤ j ≤ k′) and vn[1...i] (1 ≤ i ≤ l).The base conditions are:

C[0, 0] = 0, C[0, j] = C[0, j − 1] + gap score


and C[i, 0] = C[i− 1, 0] + gap score

The recurrence relation is:

C[i, j] = max

C[i− 1, j − 1] + score(v

′m[j], vn[i]),

C[i− 1, j] + gap score,

C[i, j − 1] + gap score

(3.3)

score(v′

m[j], vn[i]) =

match score if∆r < tr, ∆θ < tθ and ∆φ < tφ

nonmatch score otherwise

where ∆r = (|ri − r

′j|/min(ri, r

′j)), ∆θ = |θi − θ

′j| and ∆φ = |φi − φ

′j| ;

match score is the reward when a match is found , nonmatch score and gap score

are penalties when mismatch is found and gap is introduced, respectively; tr, tθ

and tφ are the parameters of the bounding box.

In other words if we have an optimal alignment up to vm′ [1..j] and vn[1..i] then

there are only three possibilities of what could happen next: i) v′m[j] and vn[i]

match, ii) a gap is introduced in vn, and iii) a gap is introduced in vm′ . The use

of ”gaps” takes care of the cases when there are spurious and missing minutiae.

• Traceback : This step constructs the optimal alignment by tracing back in the

matrix C any path from C[l, k′] to C[0, 0] that maximizes the cost.

Let Qmap be the Hash Map containing the mapping of minutiae m ∈ Q and their

corresponding stars vm and Tmap be the Hash Map containing the mapping of minutiae

m′ ∈ T and their corresponding stars vm′ .

Qmap = (m1, vm1), ....., (mx, vmx)

Tmap = (m′

1, vm′1), ....., (m

′

y, vm′y)


where x, y denote the number of minutia points which are considered for matching

in Q, T respectively. The matching technique is summarized in algorithm 1.

Algorithm 1 Minutia-based Matching

Require: Qmap and Tmap1: let M be the Hash Map containing matched minutia pairs (initially empty).

2: for all (m, vm) ∈ Qmap do

3: for all (m′, vm′ ) ∈ Tmap do

4: if m′

is already matched then

5: continue

6: end if

7: calculate edit distance between vm and v′m

8: end for

9: choose vm′ having minimum edit distance with vm10: find the number of matching edges (nummatch) between vm and vm′ by using

a sequence alignment technique.

11: if nummatch > minmatch then

12: Insert (m,m′) in M minmatch is a predefined constant

13: end if

14: end for

15: return M

3.3.1.2 Consolidation step

During this step the matched minutiae pairs are validated at global level. The

matched minutiae should be consistent in terms of global characteristics like orien-

tation. Thus all valid matches must have similar orientation differences if they come

from same finger. We approximate the orientation difference between fingerprints by

plotting a histogram with bins representing orientation differences and each bin 36

degree wide. The bin with maximum height and its immediate neighboring bins are

considered. The matched minutiae pairs in other bins are removed from the matched

list M . Let N2 be the number of matched pairs in M , the Level 2 match score (S2)


is calculated using a very famous approach:

S2 =2 ∗N2

NQ +NT

(3.4)

where NQ and NT are the number of minutia points in Q and T , respectively. If

N2 ≥ 12, the two fingerprints (Q and T ) are considered to be from same finger,

otherwise the matching continues to Level 3.

3.3.2 Level 3 features based Matcher

The matched minutiae pairs at Level 2 are further examined and Level 3 features

in their local neighborhood are compared. Thus for a given pair of matched minutiae,

we compare the level 3 features in their neighborhood and the minutia correspondence

is recomputed based on the agreement of Level 3 features. The ridge contours and

pores are separately compared in the neighborhood. While comparing the Level 3

features, we need to consider the fact that the detected features will vary in different

impressions of the same finger, due to the degradation of image quality (because

of noise, skin deformation etc.). Thus, a decision based OR rule is used to verify

the minutiae correspondences, i.e if either of pores or ridge contours agree then the

minutia correspondence is verified.

Localized matching is done to tolerate the effect of non-linear distortion. In [40]

it has been observed that given a sufficiently high resolution, the same quantity of

discriminative information could be extracted from fingerprint fragments by using

Level 3 data, as could be when considering Level 2 data extracted from the complete

fingerprint image. Thus by using localized window sufficient information can be

obtained.


3.3.2.1 Pore Matching

Each pore is represented by a 3-tuple (x, y, θ), where x,y denote its location and

θ is the direction of the ridge, at the location where it lies. [70], [67] and [29] have

used only location information of pores for matching. Similar to minutiae, pores

maintain the same relative orientation to other pores within a fingerprint between

different impressions. In a fingerprint pores are distributed over ridges and associating

direction information with pores provides an additional information for matching.

We propose a novel approach for matching pores. Pores within a circular region

around the matched minutia points, obtained from minutia-based matcher, are used

for matching. Given a minutia m as the origin, the pores in a circular region Cm

around it are arranged in order of radial distance first and then pores with same radial

distance are ordered by the increasing radial angle (the angle between the direction

of line segment joining a pore with the central minutia and the central minutia’s

orientation). Suppose (m,m′) is a matching minutia pair, the pore information related

to them can be represented by mp, m′p respectively:

mp = (p1, ..., pi) = ((r1, θ1, φ1), ..., (ri, θi, φi))

m′

p = (p′

1, ..., p′

j) = ((r′

1, θ′

1, φ′

1), ..., (r′

j, θ′j, φ

′

j))

where i,j are the number of pores in Cm, Cm′ respectively; rk, θk, φk denotes the radial

distance, orientation difference, radial angle of kth pore with the central minutia mp;

r′

k, θ′

k and φ′

k have similar relationship with m′p.

A dynamic programming based approach is used to find the edit distance, dpores

between the ordered pore strings associated with two matching minutiae. The same

approach is used for finding the edit distance which is used in case of minutiae. The

edit distance gives a measure of similarity and is tolerant to spurious and missing


pores.

3.3.2.2 RidgeContour Matching

Ridge contour are observed to be more reliable than pores at low resolution (500

dpi). Figure (3.6) shows two fingerprints at 500 ppi having prominent ridge structure.

However, pores are not found to be consistent. We concentrate on localized matching

for tolerating non-linear distortion. Thus ridge contours are matched in a w × w

window around the matched minutiae pairs. Non-linear distortion is a major concern

when matching smaller structures, such as points along ridges (ridge contours). Lo-

calized matching can only tolerate non-linear distortion to an extent, but to deal with

the adverse effects of distortion (caused by skin plasticity) more efforts are needed.

Infact, the matching technique should be robust to non-linear distortion.

(a) (b)

Figure 3.6: Two impressions of same fingerprint at 500 dpi.

A novel approach for ridge contour matching is proposed which uses Zernike mo-


ments as shape features and Local Binary Patterns (LBP) as texture features. The

proposed approach combines the Zernike moments and LBP to form a set of features

suitable for texture shape features.

Shape Feature extraction using Zernike Moments

Zernike moments costitute a powerful shape descriptor in terms of robustness and

description capability and have been employed in a wide range of applications like

character recognition, object recognition, palm-print recognition, iris recognition and

face recognition [71, 8, 48, 17]. This shape descriptor has proved its superiority over

other moment functions [12, 43], regarding to its description capability and robustness

to noise or deformations. Zernike moments are based on a set of complex polynomials

(Zernike polynomials) that form a complete orthogonal set over the interior of the

unit circle [72]. The Zernike function of order p and repetition q are defined in the

polar coordinate system (r, θ) as

Vpq(r, θ) = Rpq(r).eiqθ (3.5)

where i =√−1 and Rpq is the orthogonal real valued radial polynomial defined as

Rpq(r) =

(p−|q|)/2∑n=0

(−1)n(p− n)!

n!(p+|q|2− n)!(p−|q|

2− n)!

rp−2n (3.6)

where p− |q| is even and |q| ≤ p

Zernike moments of an image are the projections of the image function onto these

orthogonal basis functions. The Zernike moments of order p and repetition q for an

image intensity function f(r, θ) over the polar coordinate space is:

Zpq =p+ 1

Π

∫ 2Π

0

∫ 1

0

V ∗pq(r, θ)f(r, θ)rdrdθ (3.7)

where V ∗pq denotes the complex conjugate of Vpq. If N is the number of pixels along


each axis of the image, then Equation (3.7) can be written in the discrete form as

Zpq =p+ 1

Π(N − 1)2

N∑x=1

N∑y=1

V ∗pq(r, θ)f(x, y) (3.8)

where r =√

(x2 + y2)/N , and θ = tan−1(y/x). The polar form of Zernike moments

suggests a square-to-circular image transformation [49], so that the Zernike polyno-

mials need to be computed only once for all pixels mapped to the same circle. This

transformation from square to circular region is shown in Figure 3.7. If the image

coordinate system (x,y) is defined with the origin at the center of the square pixel

grid, then the pixel coordinates of the transformed circular image can be represented

by γ, ξ where γ denotes the radius of the circle and ξ the position index of the pixel

on the circle. The normalized polar coordinates r, θ of the pixel (γ, ξ) are given by

r = 2γ/N, θ = πξ/(4γ) (3.9)

Figure 3.7: Schematic of square to circular transformation

From Equation 3.7 and 3.5, Zernike moments of a pattern rotated by an angle a

around its center of mass are given in polar coordinates as

Zapq = Zpqe

iqa (3.10)


From Equation 3.10 we have |Zpq| = |Zpqeiqa|, thus magnitude of Zernike moments

are invariant to rotation. Due to the property of orthogonality the contribution of

each moment is unique and independent. The Zernike moments are calculated in a

w × w window around the matched minutia pairs. The Zernike feature vector with

moment order n and repetition m is given by:

fZernike = [|ZM00|, |ZM01|, ..., |ZM(n−1)(m−1)|] (3.11)

where ZMmn denotes Zernike moments of order n and repetition m.

Texture Feature Extraction using Local Binary Patterns

Local Binary Patterns (LBP) [55] is a very simple, yet efficient, multi-resolution

approach to gray-scale and rotation invariant texture description and has been ex-

tensively used as a state of the art feature extractor in Face Recognition [69]. The

basic LBP operator, introduced by Ojala et al. [55], labels the pixels of an image by

thresholding the n× n (n = 3) neighborhood of each pixel with the center value and

considering the result as a binary number. Then the histogram of the labels can be

used as a texture descriptor. The basic LBP operator has been illustrated in Figure

3.8. Rotation invariance can be achieved by selecting the minimum possible value of

the binary pattern as the label.

Figure 3.8: The basic LBP operator

Given an image f(x,y), a histogram of the labeled image fl(x, y) can be defined as


Hi =∑x,y

Ifl(x, y) = i, i = 0, ..., n− 1 (3.12)

in which n is the number of different labels produced by the LBP operator and

IA =

1 A is true,

0 A is false.

(3.13)

This histogram contains information about the distribution of the local micro-

patterns, such as edges, spots and flat areas, over the whole image. The rotation

invariant LBP histogram, fLBP is calculated from a w × w symmetric (square) local

region around both the matched minutia pairs.

Fusion of LBP-Zernike features

The LBP-Zernike feature vector is obtained by combining the LBP and Zernike

features vectors as follows:

• Normalize the LBP feature and Zernike Feature respectively by z score normal-

ization.

• LBP-Zernike feature is given by:

fLBP−Zernike = fLBPU fzernike (3.14)

Finally the similarity between two minutia points m,m′

is the Euclidean distance

(dridges) between their respective fLBP−Zernike features.

3.3.2.3 Fusion of Level 2 and Level 3 features

The distance measure obtained by matching pores and ridges in a local neigh-

borhood around matched minutia pairs are used to reverify the minutia correspon-

dence. Two heuristically determined thresholds τpores and τridges are used for deciding


whether the pores and ridge contours, respectively, from two localized regions, agree

or not. Thus, a minutia pair m,m′

matched at Level 2, is considered matched at

Level 3, if either of the following two conditions hold.

dridges < τridges

dpores < τpores

Thus at Level 3 minutia correspondences are recomputed and the number of matched

minutiae N3 at level 3 is then used to calculate Level 3 match score

S3 =2 ∗N3

NQ +NT

(3.15)

where NQ and NT are the number of minutia points in Q and T , respectively. If

S3 > τ3 (τ3 is heuristically determined threshold) then the fingerprints are considered

to be matched other wise they are declared as non-matched fingerprints.

63

Chapter 4

EXPERIMENTAL RESULTS

This chapter describes the various experiments conducted and discusses the ob-

tained results. The proposed approach has been evaluated on three different databases.

4.1 Database

The evaluations and testing of the proposed approach has been done on three di-

verse fingerprint databases: Neurotechnology Database [3], FVC2004 [5] DB3 Database

and FVC2006 [1] DB2 Database. They are explained in detail below.

• Neurotechnology Database: The database consists of 51 different fingers with

8 impressions per finger resulting in 408 images. The fingerprint samples are

scanned with an optical scanner (Cross Match Verifier 300) at 500 dpi. The

sample images from this database are shown in Figure 4.1.

• FVC 2004 DB3 Database: The database consists of 100 different fingers with

8 impressions per finger resulting in 800 images. The fingerprints are scanned

with a thermal sweeping sensor (FingerChip FCD4B14CB by Atmel) at 512 dpi.

4.1. DATABASE 64

Figure 4.1: Sample images from Neurotechnology Database

The FVC 2004 databases are known to be difficult because of the perturbations

which were deliberately introduced during database collection [5]. The sample

images from this database are shown in Figure 4.2.

Figure 4.2: Sample images from FVC 2004, DB3 Database

• FVC 2006 DB2 Database: The database is 140 fingers wide and 12 samples per

4.2. PERFORMANCE EVALUATION 65

finger in depth (total 1680 fingerprint). The fingerprints were scanned with an

optical sensor at 569 dpi. A heterogeneous population which includes manual

workers and elderly people was used to create the database [1]. The volunteers

were simply asked to put their fingers naturally on the acquisition device, but

no constraints were enforced to guarantee a minimum quality in the acquired

images. Figure 4.3 displays the sample images from this database.

Figure 4.3: Sample images from FVC 2006, DB2 Database

4.2 Performance Evaluation

• Each sample in a Database is matched against the remaining samples of the same

finger to compute the False Rejection Rate(FRR). The FRR is the fraction of

4.2. PERFORMANCE EVALUATION 66

genuine fingerprints which are rejected and is calculated as follows

FRR =Number of genuine fingerprints rejected

Total number of genuine tests(4.1)

If the fingerprint g is matched to fingerprint h, the symmetric match (i.e., h

against g) is not executed to avoid correlation in the scores. The total number

of genuine tests are 1428 ,2800 and 9240 for Neurotechnology database, FVC

2004 Database and FVC 2006 Database respectively.

• For all databases the first sample of each finger is matched against the first

sample of the remaining fingers in the same database to compute the False

Acceptance Rate (FAR). The FAR is the fraction of impostor fingerprints which

are accepted and is calculated as follows

FAR =Number of impostor fingerprints accepted

Total number of impostor tests(4.2)

If the matching g is performed, the symmetric one (i.e., h against g) is not

executed to avoid correlation. The total number of false acceptance tests are

1275, 4950 and 9730 for the Neurotechnology database, FVC 2004 Database

and FVC 2006 Database respectively.

• For each algorithm and for each database, EER (equal-error-rate) and ROC(t)1

curve are used as performance indicators. The ROC(t) curve is a plot between

FAR and FRR for different values of the threshold, t. Given two algorithms

and their ROC curves, the better algorithm is one whose ROC curve is lower

than other’s ROC curve over a suitably large range of threshold values t. The

EER is the rate at which the FAR is equal to the FRR. The lower the EER,

1t denotes the acceptance threshold

4.3. EXPERIMENT 1 67

the better is the system. For each algorithm the Impostor and Genuine score

distributions are also reported.

4.3 Experiment 1

The first stage of the hierarchical matcher is the minutiae matching stage and the

performance of the next stage is dependent upon this stage. In this experiment the

performance of proposed minutia matcher is compared with a very popular minutiae

matching approach by Sharat et al. [15]. Sharat et al. have used a graph based

matching, Coupled BFS (CBFS) to match local minutia structures called Kplets.

Figure 4.4, 4.5 and 4.6 compare the performance of proposed minutiae matcher with

the CBFS-Kplet based matcher on the three databases mentioned above. From the

graphs it is evident that the proposed minutiae matcher performs better than CBFS-

Kplet based approach on all the three databases. Table 4.3 shows the performance

gain, in terms of EER, of the proposed minutiae matcher over the CBFS-Kplet based

matcher.

Database Proposed minutia CBFS-Kplet based % improvement in EER (%)matcher (EER %) matcher (EER %)

Neurotechnology 1.76 3.78 53.44FVC 2004, DB3 10.20 12.70 19.68FVC 2006, DB2 2.44 3.80 35.79

Table 4.1: Equal Error Rate (EER) comparison between proposed minutia matcherand CBFS-Kplet based matcher.


0

1

2

3

4

5

6

7

8

9

10

0 1 2 3 4 5 6 7 8 9 10

Fal

se R

ejec

tion

Rat

e (F

RR

) %

False Acceptance Rate (FAR) %

CBFS Kplet, EER = 3.78%minutia matcher, EER=1.76%

Figure 4.4: ROC curves for proposed minutia matcher and CBFS-Kplet basedmatcher, on Neurotechnology Database

4.4 Experiment 2

This experiment compares the capacity of both LBP and Zernike moments (ZM)

to retrieve information from ridge contours. Moment order (n) and repetition (m)

are significant parameters when ZM are used as features. With high order, ZM can

carry more fine details of an image but become more susceptible to noise. After

balancing the computational complexity and retrieval performance, the moments of

order n = 4, 8, 12, 16 are considered in our experiment.

The ROC curves for the three databases are shown in Figure 4.7, 4.8 and 4.9. The

best performance achieved by using LBP, ZM and combination of both LBP and ZM

for each database is shown in the graphs. From the graphs it is evident that there

is performance gain (in terms of EER) when ridge information is used in matching.


0

5

10

15

20

25

30

35

40

0 5 10 15 20 25 30 35 40

Fal

se R

ejec

tion

Rat

e (F

RR

) %


minutia matcher (EER=10.20%)CBFS Kplet (EER = 12.70%)

Figure 4.5: ROC curves for proposed minutia matcher and CBFS-Kplet basedmatcher, on FVC 2004, DB3 Database

The LBP performs much better than both ZM and even LBP-ZM combined. Also

the computational cost of LBP operator is minimal while that for ZM is very high

and it increases as the order of ZM increases. The timing analysis is shown in Table

4.3.


0

2

4

6

8

10

0 2 4 6 8 10

Fal

se R

ejec

tion

Rat

e (F

RR

) %


CBFS Kplet (EER = 3.80%)minutia matcher (EER=2.44%)

Figure 4.6: ROC curves for proposed minutia matcher and CBFS-Kplet basedmatcher, on FVC 2006, DB2 Database

4.5 Experiment 3

This experiment evaluates the proposed hierarchical matching algorithm on the

above mentioned databases. Of the three databases only Neurotechnology database

was found to be suitable for extracting pores. In the other two databases (FVC 2004

DB3 and FVC 2006 DB2) only ridge contours have been used as Level 3 Features.

The ROC curves for the three databases are shown in Figure 4.10, 4.11 and 4.12.

Each figure compares the ROC curve of the proposed minutia matcher with the

proposed hierarchical matcher. As expected, the hierarchical matcher performs better

than the minutia matcher, for all the three databases. The improvement in EER is

shown in Table 4.5.


0

1

2

3

4

5

6

7

8

9

10

0 1 2 3 4 5 6 7 8 9 10

Fal

se R

ejec

tion

Rat

e (F

RR

) %


minutia matcher, EER=1.76%hierarchical matcher (minutia + LBP), EER=1.05%

hierarchical matcher (minutia + LBP + ZM(4)), EER=1.54%hierarchical matcher (minutia + ZM(4)), EER=1.72%

Figure 4.7: ROC curves comparing the performance of LBP and ZM (Neurotechnol-ogy Database)

Database Proposed minutia Proposed hierarchical matcher (EER %)matcher (EER %) minutia minutia minutia, pores

& ridges & pores & ridgesNeurotechnology 1.76 1.05 3.77 1.05FVC 2004, DB3 10.20 8.80 - -FVC 2006, DB2 2.44 2.40 - -

Table 4.2: Equal Error Rate (EER) comparison between proposed minutia matcherand proposed hierarchical matcher.


0

5

10

15

20

0 5 10 15 20

Fal

se R

ejec

tion

Rat

e (F

RR

) %


minutia matcher, EER=10.20%hierarchical matcher (minutia + ridges(LBP)), EER=8.80%

hierarchical matcher (minutia + ridges(ZM16)), EER=9.60%hierarchical matcher (minutia + ridges (LBP+ZM12)), EER=9.61%

Figure 4.8: ROC curves comparing the performance of LBP and ZM (FVC 2004, DB3Database)


100

101

10-2 10-1 100 101

Fal

se R

ejec

tion

Rat

e (F

RR

) %


minutia matcher, EER=2.44%hierarchical matcher (minutia + ridges(LBP)), EER=2.40%

hierarchical matcher (minutia + ridges(ZM16)), EER=2.45%hierarchical matcher (minutia + ridges (LBP + ZM16)), EER=2.44%

Figure 4.9: ROC curves comparing the performance of LBP and ZM (FVC 2006, DB2Database)


0

1

2

3

4

5

6

7

8

9

10

0 1 2 3 4 5 6 7 8 9 10

Fal

se R

ejec

tion

Rat

e (F

RR

) %


minutia matcher, EER=1.76%hierarchical matcher (minutia + ridges), EER=1.05%hierarchical matcher (minutia+ pores), EER=3.77%

hierarchical matcher (minutia+pores+ridges), EER=1.05%

Figure 4.10: ROC curves for the minutia matcher and hierarchical matcher on Neu-rotechnology Database


0

5

10

15

20

25

30

35

40

0 5 10 15 20 25 30 35 40

Fal

se R

ejec

tion

Rat

e (F

RR

) %


minutia matcher (EER=10.2%) hierarchical matcher (EER=8.80%)

Figure 4.11: ROC curves for the minutia matcher and hierarchical matcher on FVC2004, DB3 Database


100

101

10-2 10-1 100 101

Fal

se R

ejec

tion

Rat

e (F

RR

)%

False Acceptance Rate (FAR)%

minutia matcher (EER=2.44%) hierarchical matcher (EER=2.40%)

Figure 4.12: ROC curves for the minutia matcher and hierarchical matcher on FVC2006, DB2 Database

4.6. TIMING ANALYSIS 77

The hierarchical matcher results in a performance gain (in terms of EER) of

∼ 40%, ∼ 14% and ∼ 2% for Neurotechnology, FVC 2004 DB3, and FVC 2006

DB2 database, respectively. From Figure 4.10, it is observed that the contribution

of “pores” in improving(lowering) the EER is insignificant. The hierarchical matcher

performs better when ridges are used along with minutiae as compared to when

pores are used along with minutiae. In-fact, the hierarchical matcher performs with

comparable accuracy when ridges are considered at Level 3 as compared to when

ridges and pores both are considered at Level 3.

The lowering of EER can be attributed to the fact that both FAR and FRR are

lowered when a new threshold (τ3) is chosen at Level 3. This can be observed from the

distribution of - number of matched minutiae for genuine and impostor cases. Figure

4.13, 4.14 and 4.15 compares the number of minutiae matched by minutia matcher

and hierarchical matcher, for both impostor and genuine cases, for Neurotechnology,

FVC 2004 DB3, and FVC 2006 DB2 databases, respectively. From all the graphs it

is evident that, matching at Level 3 causes the number of matched minutia pairs to

decrease significantly for impostor cases when compared to that for genuine cases.

4.6 Timing Analysis

The above experiments show that the use of Level 3 Features provide additional

information which can be used to lower the error rates. But using additional features

also requires additional time. And if the time required is very large then the approach

might not be suitable for automatic matching. In this experiment we evaluate the

time required for matching Level 3 features. Table 4.3 provides a comparison of the

matching time for Level 3 Features. The time required is least when LBP is used


0 1 2 3 4 5 6 7 8 9

10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Fre

quen

cy %

Number of Matching Minutiae

# of matches for genuine (minutiae matcher)# of matches for impostor (minutiae matcher)

# of matches for genuine (hierarchical matcher)# of matches for impostor (hierarchical matcher)

Figure 4.13: Distribution of number of matched minutiae, for Genuine and Impostorcases, on Neurotechnology Database


0

5

10

15

20

25

30

35

40

45

50

55

60

65

0 5 10 15

Fre

quen

cy %




Figure 4.14: Distribution of number of matched minutiae, for Genuine and Impostorcases, on FVC 2004, DB3 Database


0

10

20

30

40

50

60

70

0 2 4 6 8 10 12 14

Fre

quen

cy %




Figure 4.15: Distribution of number of matched minutiae, for Genuine and Impostorcases, on FVC 2006, DB2 Database


for ridge matching. The LBP operator is well known for its fast, robust and efficient

feature extraction technique. Also, the time complexity for matching ridges increases

as the order of Zernike moments increases. From this experiment we can conclude

that LBP and low order Zernike Moments are suitable for matching as far as the time

for match is concerned.

In [29] Jain et al. have proposed a hierarchical matching system which uses

an iterative closest point algorithm (ICP) [13] for matching Level 3 features. The

iterative nature of ICP makes the use of this technique unsuitable for automatic

matching. Jain et al. have reported an average time of ∼ 45 seconds for matching

Level 3 features. The system configuration used by Jain et al. is comparable to

the system configuration used in our experiments. But still a comparison cannot be

made between the match time for proposed approach and the approach by Jain et al.,

since the database used are different. However, it gives an overview of the average

match time. The system configuration on which the experiments were done is 1.6

GHz Pentium 4 CPU and 1 GB of RAM.

Table 4.3: Comparison of matching time for Level 3 Features.

Matching Technique Average time for matching Level 3Features, in seconds

Proposed Approach (ZM4) 1.28Proposed Approach (ZM8) 4.73Proposed Approach (ZM12) 10.47Proposed Approach (ZM16) 19.01Proposed Approach (LBP) 1.21Proposed Approach (pores) 0.1

82

Chapter 5

CONCLUSION & FUTURE

WORK

5.1 Conclusion

A new Hierarchical Fingerprint matcher which utilizes Level 3 features(pores &

ridges) in conjunction Level 2 Features (minutiae) has been proposed in this thesis.

The novelty lies in the matching technique used for matching Level 2 features (minu-

tiae) and Level 3 features (pores and ridge contours). We have concentrated on 500

ppi fingerprints as the Federal Bureau of Investigation’s (FBI) standard of fingerprint

resolution for Automated Fingerprint Identification System (AFIS) is 500 ppi. The

testing of our approach was done on three diverse fingerprint databases and based on

our observations and obtained results the following conclusions can be made:

• The use of Level 3 features (pores and ridge contours) provide complementary

information which can be used along with Level 2 features (minutiae) to lower

the error rates, namely FAR and FRR.

5.1. CONCLUSION 83

• At 500 ppi, the ridge contours were observed to be more reliable and prominent

features than the pores and the results obtained from our experiments also

justify this.

• The Local binary Patterns are effective means of representing the shape and

texture information of ridge contours even in poor quality images.

• The proposed hierarchical matcher has a matching time suitable for automated

fingerprint verification systems.

• Finally, We have tried to overcome the real challenges in fingerprint matching

namely, non-linear distortion, small overlap between query and template images,

error and noise introduced by feature extraction algorithms, error introduced

in registration and due to unfavorable skin conditions. Localized matching was

used for matching all feature types, in-order to minimize the effects of distor-

tion. Also, rotationally invariant structures (pores and minutia) and features

(ridge contours) are used and as a result any type of alignment (registration)

is not required at any stage. The use of Level 3 features is beneficial in de-

ciding match/nonmatch, with increased accuracy, in case of fingerprints with

small overlap. The pores and minutiae are matched using an elastic string

matching algorithm which is capable of overcoming the errors introduced by

feature extraction algorithms. The experiments are conducted on standard

public databases- FVC 2004 DB3 (containing manually perturbed fingerprints),

FVC 2006 DB2 (containing fingerprints from a heterogeneous population which

included manual workers and elderly people) and Neurotechnology Database

(a relatively good quality database). The three databases captures most of

the problems that we have mentioned. And the fact that our proposed minu-

5.2. FUTURE WORK 84

tia matcher performs better than a popular CBFS-Kplet [15] based minutia

matcher, proves that our approach is capable of overcoming the mentioned

problems to a large extent.

5.2 Future Work

The proposed hierarchical matcher does not require registration or alignment of

fingerprints at any stage. This makes it suitable for fingerprint fragment (partial

fingerprint) comparison, as the major difficulty in partial fingerprint comparison is

that it is difficult to align the partial fragment with a template fingerprint. Also a

partial fingerprint has insufficient information if only minutiae features are considered,

thus using additional information in form of Level 3 features (especially ridge contour

information) can provide additional information.

85

Bibliography

[1] Fourth International Fingerprint Verification Competition. http://bias.csr.

unibo.it/fvc2006.

[2] The History of Fingerprints. http://www.onin.com/fp/fphistory.html.

[3] Neurotechnology Sample Database. http://www.neurotechnology.com/

download.html.

[4] Ridges and Furrows - history and science of fingerprint identification, technologyand legal issues. http://ridgesandfurrows.homestead.com/fingerprint.

html.

[5] FVC2004 - Third International Fingerprint Verification Competition. http:

//bias.csr.unibo.it/fvc2004, 2004.

[6] CDEFFS: The ANSI/NIST Committee to Define an Extended Fingerprint Fea-ture Set. http://fingerprint.nist.gov/standard/cdeffs/index.html, Oc-tober 2006.

[7] The Thin Blue Line. http://www.policensw.com/info/fingerprints/

finger06.html, October 2006.

[8] E. M. Arvacheh and H. R. Tizhoosh. Pattern Analysis using Zernike Moments.In Instrumentation and Measurement Technology Conference, pages 1574–1578,May 2005.

[9] D.R. Ashbaugh. Quantitative-Qualitative Friction Ridge Analysis: An Introduc-tion to Basic and Advanced Ridgeology. CRC Press, 1999.

[10] A.M. Bazen and S.H. Gerez. Segmentation of fingerprint images. In ProR-ISC 2001 Workshop on Circuits, Systems and Signal Processing, pages 276–280.Citeseer, 2001.

[11] A.M. Bazen, G.T.B. Verwaaijen, S.H. Gerez, L.P.J. Veelenturf, and B.J. van derZwaag. A Correlation-Based Fingerprint Verification System. In Proc. of 11th

http://bias.csr.unibo.it/fvc2006


http://www.onin.com/fp/fphistory.html

http://www.neurotechnology.com/download.html

http://www.neurotechnology.com/download.html

http://ridgesandfurrows.homestead.com/fingerprint.html

http://ridgesandfurrows.homestead.com/fingerprint.html



http://fingerprint.nist.gov/standard/cdeffs/index.html

http://www.policensw.com/info/fingerprints/finger06.html

http://www.policensw.com/info/fingerprints/finger06.html

BIBLIOGRAPHY 86

Annual Workshop on Circuits Systems and Signal Processing (ProRISC), pages205–213, 2000.

[12] S. O. Belkasim, M. Shridhar, and M. Ahmadi. Pattern recognition with momentinvariants: a comparative study and new results. Pattern Recogn., 24(12):1117–1138, 1991.

[13] Paul J. Besl and Neil D. McKay. A Method for Registration of 3-D Shapes.IEEE Trans. Pattern Anal. Mach. Intell., 14(2):239–256, 1992.

[14] Ruud Bolle, Jonathan Connell, Sharanthchandra Pankanti, Nalini Ratha, andAndrew Senior. Guide to Biometrics. SpringerVerlag, 2003.

[15] Sharat Chikkerur, Alexander N. Cartwright, and Venu Govindaraju. K-pletand Coupled BFS: A Graph Based Fingerprint Representation and MatchingAlgorithm. In ICB, volume 3832 of Lecture Notes in Computer Science, pages309–315. Springer, 2006.

[16] Sharat Chikkerur, Alexander N. Cartwright, and Venu Govindaraju. Fingerprintenhancement using STFT analysis. Pattern Recogn., 40(1):198–211, 2007.

[17] Anant Choksuriwong, Helene Laurent, Christophe Rosenberger, and ChoubeilaMaaoui. Object Recognition Using Local Characterisation and Zernike Moments.In Advanced Concepts for Intelligent Vision Systems, pages 108–115, 2005.

[18] Louis Coetzee and Elizabeth C. Botha. Fingerprint recognition in low qualityimages. In Pattern Recognition, volume 26, pages 1441–1460, 1993.

[19] Thomas H. Cormen, Charles E. Leiserson, Ronald L. Rivest, and Clifford Stein.Introduction to Algorithms. MIT Press and McGraw-Hill, 2001.

[20] S.C. Dass and A.K. Jain. Fingerprint-based recognition. Technometrics,49(3):262–276, 2007.

[21] H. Faulds. On the skin-furrows of the hand. Nature, 22:605, 1880.

[22] Francis Galton. Fingerprints. Macmillan, London, 1892.

[23] International Biometric Group. Analysis of Level 3 Features at High Resolutions.http://level3tk.sourceforge.net/, 2008.

[24] L. Hong, Y. Wan, and A. Jain. Fingerprint image enhancement: algorithm andperformance evaluation. IEEE Transactions on Pattern Analysis and MachineIntelligence, 20(8):777–789, 1998.

http://level3tk.sourceforge.net/

BIBLIOGRAPHY 87

[25] Lin Hong, Yifei Wan, and Anil Jain. Fingerprint Image Enhancement: Algo-rithm and Performance Evaluation. IEEE Transactions on Pattern Analysis andMachine Intelligence, 20(8):777–789, 1998.

[26] Nalini Ratha Ibm, Nalini K. Ratha, Vinayaka D. Pandit, Ruud M. Bolle, andVaibhav Vaish. Robust Fingerprint Authentication Using Local Structural Sim-ilarity. In In Workshop on applications of Computer Vision, pages 29–34, 2000.

[27] A. Jain and S. Pankanti. Fingerprint classification and matching. Michigan StateUniversity, 1999.

[28] A. K. Jain, A. Ross, and S. Prabhakar. Fingerprint Matching Using Minutiaeand Texture Features. In Proc. Int. Conf. on Image processing, pages 282–285,2001.

[29] A.K. Jain, Y. Chen, and M. Demirkus. Pores and Ridges: High-ResolutionFingerprint Matching Using Level 3 Features. PAMI, 29(1):15–27, January 2007.

[30] AK Jain, L. Hong, S. Pankanti, and R. Bolle. An identity-authentication systemusing fingerprints. Proceedings of the IEEE, 85(9):1365–1388, 1997.

[31] A.K. Jain, S. Prabhakar, Lin Hong, and S. Pankanti. FingerCode:A filterbankfor fingerprint representation and matching. In Proc. of IEEE Computer SocietyConference on Computer Vision and Pattern Recognition, volume 2, pages –193,1999.

[32] AK Jain, A. Ross, and S. Prabhakar. An introduction to biometric recogni-tion. IEEE Transactions on circuits and systems for video technology, 14(1):4–20,2004.

[33] Anil Jain, Yi Chen, and Meltem Demirkus. Pores and Ridges: FingerprintMatching Using Level 3 Features. In ICPR ’06: Proceedings of the 18th Inter-national Conference on Pattern Recognition, pages 477–480, Washington, DC,USA, 2006. IEEE Computer Society.

[34] Anil Jain, Lin Hong, and Ruud Bolle. On-Line Fingerprint Verification. IEEETransactions on Pattern Analysis and Machine Intelligence, 19(4):302–314, 1997.

[35] Anil K. Jain and David Maltoni. Handbook of Fingerprint Recognition. Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2003.

[36] Tsai-Yang Jea. Minutiae-based partial fingerprint recognition. PhD thesis, Buf-falo, NY, USA, 2005. Adviser-Govindaraju, Venugopal.

BIBLIOGRAPHY 88

[37] Xudong Jiang and Wei-Yun Yau. Fingerprint minutiae matching based on thelocal and global structures. In Proceedings of 15th International Conference onPattern Recognition, volume 2, pages 1038–1041, 2000.

[38] K. Karu and A.K. Jain. Fingerprint Classification. Pattern recognition,29(3):389–404, 1996.

[39] M. Kawagoe and A. Tojo. Fingerprint pattern classification. Pattern Recognition,17(3):295–303, 1984.

[40] K. Kryszczuk, A. Drygajlo, and P. Morier. Extraction of Level 2 and Level3 Features for Fragmentary Fingerprints. In Proc. Second COST Action 275Workshop, pages 83–88, 2004.

[41] K. Kryszczuk, P. Morier, and A. Drygajlo. Study of the Distinctiveness of Level2 and Level 3 Features in Fragmentary Fingerprint Comparison. In Proc. ofBiometric Authentication Workshop, pages 124–133, May 2004.

[42] Miroslav Krlk and Ladislav Nejman. Fingerprints on artifacts and historicalitems: examples and documents. Journal of Ancient Fingerprints, August 2007.

[43] Simon X. Liao and Miroslaw Pawlak. On Image Analysis by Moments. IEEETrans. Pattern Anal. Mach. Intell., 18(3):254–266, 1996.

[44] E. Locard. Les Pores et L’Identification Des Criminels. Biologica: Revue Scien-tifique de Medicine, 2:357–365, 1912.

[45] D. Maio and D. Maltoni. Direct gray-scale minutiae detection in fingerprints.IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(1):27–40,1997.

[46] D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, and A.K. Jain. FVC2000:Fingerprint Verification Competition. IEEE Transactions on Pattern Analysisand Machine Intelligence, 24(3):402–412, 2002.

[47] BM Mehtre and B. Chatterjee. Segmentation of fingerprint imagesa compositemethod. Pattern Recognition, 22(4):381–385, 1989.

[48] S. Mohammed, L. Sylvie, V. Vincent, and B. El-Mouldi. Face Detection byneural network trained with Zernike moments. In International Conference onSignal Processing, Robotics and Automation, pages 36–41, 2007.

[49] R. Mukundan and K. R. Ramakrishnan. Zernike Moments. Moment functionsin image analysis-theory and applications, pages 57–69.

BIBLIOGRAPHY 89

[50] N.J. Naccache and R. Shinghal. An investigation into the skeletonization ap-proach of Hilditch. Pattern Recognition, 17(3):279–284, 1984.

[51] K. Nandakumar and A. K. Jain. Local correlation-based fingerprint matching. InIndian Conference on Computer Vision, Graphics and Image Processing, pages503–508, 2004.

[52] Loris Nanni and Alessandra Lumini. Local binary patterns for a hybrid finger-print matcher. Pattern Recogn., 41(11):3461–3466, 2008.

[53] S. B. Needleman and C. D. Wunsch. A general method applicable to the searchfor similarities in the amino acid sequence of two proteins. J Mol Biol, 48(3):443–453, March 1970.

[54] H.v.d Nieuwendijk. Fingerprints. http://www.xs4all.nl/~dacty/minu.htm,October 2006.

[55] Timo Ojala, Matti Pietika Inen, and Topi Maenpa A. Multiresolution gray-scaleand rotation invariant texture classification with local binary patterns. IEEETransactions on Pattern Analysis and Machine Intelligence, 24:971–987, 2002.

[56] Sharath Pankanti, Salil Prabhakar, and Anil K. Jain. On the Individuality ofFingerprints. IEEE Trans. Pattern Anal. Mach. Intell., 24(8):1010–1025, 2002.

[57] T. Pavlidis. Algorithms for Graphics and Imag. WH Freeman & Co. New York,NY, USA, 1983.

[58] S. Ranade and A. Rosenfeld. Point Pattern Matching by Relaxation. PatternRecognition, 12(4):269–275, 1980.

[59] Nalini K. Ratha, Kalle Karu, Shaoyun Chen, and Anil K. Jain. A Real-TimeMatching System for Large Fingerprint Databases. IEEE Transactions on Pat-tern Analysis and Machine Intelligence, 18(8):799–813, 1996.

[60] N.K. Ratha, S. Chen, and A.K. Jain. Adaptive flow orientation-based featureextraction in fingerprint images. Pattern Recognition, 28(11):1657–1672, 1995.

[61] NK Ratha, K. Karu, S. Chen, and AK Jain. A real-time matching system forlarge fingerprint databases. IEEE Transactions on Pattern Analysis and MachineIntelligence, 18(8):799–813, 1996.

[62] M. Ray, P. Meenen, and R. Adhami. A novel approach to fingerprint poreextraction. Southeastern Symposium on System Theory, pages 282–286, 2005.

http://www.xs4all.nl/~dacty/minu.htm

BIBLIOGRAPHY 90

[63] A. R. Roddy and J. D. Stosz. Fingerprint Features - Statistical Analysis andSystem performance Estimates. In Proc. IEEE, volume 85, pages 1390–1421,1997.

[64] A. Ross, A. K. Jain, and J. Reisman. A hybrid fingerprint matcher. In Proceedingsof Int. Conf. on Pattern Recognition, volume 3, pages 795–798, 2002.

[65] G. S. Sodhi and Jasjeet Kaur. The forgotten Indian pioneers of fingerprintscience. Current Science, 88(1), January 2005.

[66] G.C. Stockman, S. Kopstein, and S. Benett. Matching Images to Models forRegistration and Object Detection via Clustering.

[67] Jonathan D. Stosz and Lisa A. Alyea. Automated System for Fingerprint Au-thentication Using Pores and Ridge Structure. In Proc. of Automatic Systemsfor the Identification and Inspection of Humans, volume 2277, pages 210–223,1994.

[68] Vivek A. Sujan and Michael P. Mulqueen. Fingerprint identification using spaceinvariant transforms. Pattern Recogn. Lett., 23(5):609–619, 2002.

[69] Abdenour Hadid Timo Ahonen and Matti Pietikainen. Face description withlocal binary patterns: Application to face recognition. IEEE Transactions onPattern Analysis and Machine Intelligence, 28:2037–2041, 2006.

[70] Mayank Vatsa, Richa Singh, Afzel Noore, and Max M. Houck. Quality-augmented fusion of level-2 and level-3 fingerprint information using DSm theory.Int. J. Approx. Reasoning, 50(1):51–61, 2009.

[71] P. Ying-Han, T.B.J. Andrew, N.C.L David, and F. S. Hiew. Palmprint Verifica-tion with Moments. Journal of Computer Graphics, Visualization and ComputerVision (WSCG), 12(1-3):325–332, Febuary 2003.

[72] F. Zernike. Beugungstheorie des schneidenverfahrens und seiner verbessertenform, der phasenkontrastmethode. Physica, 1:689–704, 1934.

Date post:	10-May-2020
Category:	Documents
Upload:	others
View:	13 times
Download:	0 times

A Hierarchical Fingerprint Matching System · 2018-11-27 · A HIERARCHICAL FINGERPRINT MATCHING...

Documents