Professional Documents
Culture Documents
Wang Ruifang
Wang Ruifang
–TESIS DOCTORAL–
This book was typeset by the author using LATEX2e. The main body of the text was set
using a 11-points Computer Modern Roman font. All graphics and images were included
formatted as Encapsuled Postscript (TM Adobe Systems Incorporated). The final postscript
output was converted to Portable Document Format (PDF) and printed.
Copyright
c 2013 by Ruifang Wang. All rights reserved. No part of this publication may be
reproduced or transmitted in any form or by any means, electronic or mechanical, including
photocopy, recording, or any information storage and retrieval system, without permission
in writing from the author. Universidad Autonoma de Madrid has several rights in order to
reproduce and distribute electronically this document.
Department: Tecnologı́a Electrónica y de las Comunicaciones
Escuela Politécnica Superior
Universidad Autónoma de Madrid (UAM)
SPAIN
Year: 2013
Committee: President:
Secretary:
Vocal 1:
Vocal 2:
Vocal 3:
The research described in this Thesis was carried out within the Biometric Recognition Group –
ATVS at the Dept. of Tecnologı́a Electrónica y de las Comunicaciones, Escuela Politécnica Superior,
Universidad Autónoma de Madrid (from 2011 to 2013). The project was funded by the Marie Curie
Fellowship under a European Community project BBfor2 (FP7-ITN-238803).
The author was awarded with a Marie Curie Fellowship from European Community
under the project BBfor2 (Bayesian Biometrics for Forensics, FP7-ITN-238803)
between 2011 and 2013, which supported the research summarized in this
Dissertation.
The author was awarded with the Best Poster Paper Award at IEEE/IAPR
International Joint Conference on Biometrics IJCB 2011, for one publication
originated from this Dissertation: ”Latent-to-full Palmprint Comparison based on
Radial Triangulation9”, Ruifang Wang, Daniel Ramos, Julian Fierrez, Proc. of
IJCB 2011, Washington DC, USA, October 2011.
Abstract
Robust and efficient automatic forensic palmprint recognition remains a big challenge, mainly
due to the large number of creases and large non-linear distortion in high-resolution palm-
prints, and the large storage and computation capabilities required for processing and matching
palmprints. Compared to forensic fingerprint recognition, the techniques of forensic palmprint
recognition have recently received attention while palmprint evidence is increasingly important
according to surveys by law enforcement agencies.
This Ph.D. Thesis is focused on building robust and efficient palmprint recognition systems
for forensic applications. Due to the fact that few systems aimed at forensic palmprint recognition
have been developed, deeper research is needed in this field in order to improve robustness,
accuracy and computational efficiency of automated forensic palmprint recognition technologies,
aimed at realistic forensic scenarios. Moreover, in order to foster the use of palmprint recognition
systems in forensic applications, advanced algorithms with multi-system and multi-region fusion,
and a proper evidence evaluation framework, will constitute a significant advance.
In this context, the present Ph.D. Thesis pretends to contribute with novel methods to face
the challenges of forensic palmprint recognition. In particular, the Thesis takes advantage of a
better understanding of large non-linear distortion problem in forensic palmprint recognition,
proposes novel algorithms dealing with the rich types of features in a high-resolution palmprint
image and fusion of palm regions, and performs evidence evaluation using the similarity scores,
i..e, the output of forensic palmprint recognition systems.
This Ph.D. Thesis devotes to the so-called field of forensic biometrics, a new area arising
from the fields of biometrics and forensic science. Therefore, it requires background of both
fields, as well as their interaction. We start with this point and give an introduction of basic
concepts involved in forensic palmprint recognition and evaluation. To have a general view of the
tasks, we introduce the focuses of forensic science, palmprint recognition, fusion in biometrics
and evidence evaluation.
To better understand the main problems of forensic palmprint recognition, we review rela-
ted works in this field. Some state-of-the-art algorithms in high-resolution palmprint recogni-
tion can be applied to forensic applications, which are mainly based on minutiae features in
high-resolution palmprint images. As most forensic fingerprint recognition systems also follow
minutiae-based matching, novel techniques of minutiae feature representation for fingerprints
can also be applied to forensic palmprint recognition. According to the practices in the field
of forensic biometrics, approaches of likelihood ratio based evidence evaluation using biome-
tric systems, such as forensic speaker and fingerprint recognition systems, can also be used for
palmprint evidence evaluation.
A reliable experimental framework is essential for the development of the methods proposed
in this Thesis. Following the principle practices in the field of forensic biometrics, we include
protocols and measures for both performance evaluation of biometric systems and likelihood ra-
tio based evidence evaluation methods, in order to comparatively assess the experimental parts
in this Ph.D. Thesis. The experimental framework also includes a description of the databases of
palmprints and fingerprints we use in this Thesis. For both palmprints and fingerprints, public
databases and unpublished forensic databases are used. In the case of public databases, THU-
PALMLAB for palmprints, NIST SD27 and FVC 2002 for fingerprints are used. As unpublished
forensic databases from criminal investigations, a forensic fingerprint database collected by Net-
herlands Forensic Institute and forensic palmprint databases collected by Beijing Institute of
Criminal Technology in China are used.
The first forensic palmprint recognition system we propose in this Thesis is based on radial
triangulation representation for minutiae in palmprints. The proposed complete forensic palm-
print recognition system based on radial triangulation includes both novel forensic palmprint
feature extraction and comparison, namely, the M inutiaLine extractor and RT comparator.
Moreover, a combined global feature comparison component is implemented, in which weights
of centroids of radial triangulation structures and principal line energy are both considered for
finer global comparison. The M inutiaLine extractor outperforms a commercial extractor Me-
gaMatcher 4.0 with much less spurious minutiae extracted and performs much faster than the
M inutiaCode extractor. The composed M inutiaLine+RT system performs better than the
M inutiaCode-based system regarding both accuracy and computational efficiency. Moreover,
the proposed method of combined global comparison outperforms centroid-based and principal
line-based global comparison methods.
We then present the second forensic palmprint recognition system we propose in this The-
sis which is based on weighted complex spectral minutiae representation (Weighted-SMC). We
first conduct a meaningful study of distortion assessment at feature level for fingerprints and
palmprints in forensic scenarios, which can guide the design of forensic fingerprint/palmprint
recognition systems. Inspired by the study, a novel Weighted-SMC comparator is proposed and
implemented as the key component of the recognition system. With applications first to forensic
fingerprint comparison and then to forensic palmprint comparison, the Weighted-SMC compara-
tor works in a comparable way for fingerprints compared to the results of the public evaluation,
i.e., NIST ELFT-EFS Evaluation, and outperforms RT-based comparator for palmpritns.
Aimed to improve the performance of forensic palmprint recognition systems by information
fusion, fusion scheme for forensic palmprint recognition is studied. Using the two forensic palm-
print recognition systems we propose in this Thesis, the evaluation of multi-algorithm fusion at
score level is conducted as one part of our fusion scheme for forensic palmprint recognition. Anot-
her part of the fusion scheme is implemented by anthropologically-inspired regional fusion for
high-resolution palmprint recognition in full-to-full matching mode. Existing regional fusion pro-
posals in the literature are mainly based on simply-shaped or equally-divided sectors. To achieve
more meaningful regional fusion, our proposal is anthropologically motivated, dividing the palm
in three different regions according to anatomical constrains, i.e., utilizing datum points located
on principal lines of a palmprint. This is more adapted to the view of a forensic practitioner, and
is relevant as a study of the distribution of discriminating information in the different regions of
a palm, namely, regional discriminability. To obtain palm regions based on datum points, both
manual segmentation and automatic segmentation are implemented. The evaluation of regional
fusion at score level is also conducted. A significant improvement by the proposed regional fusion
is achieved for high-resolution palmprint recognition in full-to-full matching mode which is also
important in forensic applications such as preventing identity spoofing.
The research work described in this Dissertation has led to novel contributions which inclu-
de two new methods of forensic palmprint comparison and new methods on automatic region
segmentation for high-resolution palmprints and distortion assessment in forensic scenarios, na-
mely: i) local comparison based on radial triangulation according to point pattern comparison
by relaxation; ii) forensic fingerprint/palmprint comparison based on weighted complex spectral
minutiae representation (Weighted-SMC); iii) automatic region segmentation based on convex
hull comparison for high-resolution palmprint recognition; iv) distortion assessment using minu-
tiae window. Moreover, a framework for high-resolution/forensic palmprint recognition system
based on anthropologically-inspired regional fusion has been proposed first time. Besides, some
literature reviews has been derived from this Dissertation.
To my dearest brother.
To my running buddies.
This PhD Thesis summarizes the work carried out during my Ph.D. studies with the Bio-
metric Recognition Group - ATVS since 2011. This research group was established in 1994 at
the Dept. of Ingenierı́a Audiovisual y Comunicaciones (DIAC) of the Universidad Politécnica
de Madrid (UPM) and since 2004 is affiliated to the Dept. of Ingenierı́a Informática of the Uni-
versidad Autónoma de Madrid (UAM). The financial support for the Ph.D. studies came from
a European research grant Marie Curie Fellowship under a EC project Bayesian Biometrics for
Forensics (BBfor2) which had perfectly aimed at the contributions developed during the Ph.D.
studies. And partial support also came from Spanish Guardia Civil, “Cátedra UAM-Telefónica”
and various European projects.
Foremost, I would like to thank my advisors Profs. Daniel Ramos and Julián Fiérrez for their
guidance and support over the past three years. The confidence they have always shown in me
and the continuous encouragement have definitely fostered my motivation. During these years I
have benefited so much from their vision, discipline and intelligent effort, which have shaped my
working attitudes. As the supervisor of my most research work, here I must first give a big thank
to Prof. Daniel Ramos whose importance cannot be replaced. Besides the constant support and
advice in the academic field, I cannot be more lucky to have the chance of experiencing happy
working which is a concept insisted by him. Moreover, in personal life, he has dedicated so much
to the happiness of the group via his funny talk and loud music, and has definitely helped me
too much to keep positive energy in my life. And to respectable Prof. Julián Fiérrez, I owe many
thanks to his encouragement of being a better researcher and doing high-quality research. His
intelligent advices either on research topics and problems or on academic writing have definitely
boosted the improvement of my research abilities.
In the framework of the ATVS research group, I have also received support from Profs.
Javier Ortega-Garcı́a and Joaquı́n González-Rodrı́guez, feeling fortunate to learn from their
wise advices. I must also thank Prof. Doroteo Torre-Toledano for his help with administration
issues and advices.
As involved in the international research network BBfor2, I have many thanks to the big
family consisting of managers, supervisors and research fellows from more than 20 countries,
in which I have experienced inspiring training and international friendship. I first would like to
thank the project managers Prof. David van Leeuwen and Henk van den Heuvel for their efforts
in the perfection of the network and personal support. With great gratitude, I must give many
thanks for all the discussions and advices during meetings, workshops and conferences to Profs.
Josef Bigun, Hugo Van Hamme, Pietro Laface, Sebastien Marcel, Dominic Watt, Paul Foulkes,
Didier Meuwly (to whom I owe special thanks for later) and Raymond Veldhuis (to whom I
owe special thanks for later), Drs. Phil Harrison, Mitchell McLaren, Rahim Saeidi, Roy Wallace,
Manuel Gunther, Javier Castano and Denis Burnham. Moreover, I really appreciate the close
friendship with all the fellows, lovely Miranti Indar Mandassari, brilliant Ram Prasad Krish-
namoorthy, casual Vasileos Vasilakakis, a little formal Tauseef Ali, visionary Abhishek Dutta,
handsome Laurent El Shafey, rigorous Erica Gold, pretty Nathalie Fecher, ambitious Mohamed
Hasan Bahari, helpful Anna Mikaelyan, passionate Rudolf Haraksim, joyful Ahilan Kanagasun-
dram. Among these lovely people, I have special thanks for the moments we share together to
Ram Prasad Krishnamoorthy who is also my local work mate, Miranti Indar Mandassari, Rudolf
Haraksim and Anna Mikaelyan.
The opportunity of visiting foreign institutions during my Ph.D. studies was a real luck and
precious experience. I especially remember my first research stay at University of Twente. I had
the fortune to be supervised by Prof. Raymond Veldhuis there, who kindly hosted me at the
Signals and Systems (SAS) laboratory where I met a fantastic group of people, especially a
bunch of Chinese researchers. I owe many thanks to Prof. Raymond Veldhuis for his supporting
and encouragement not only lasting the two months of my stay there bust also afterwards
collaboration, and for his perseverance which has impressed me so much. Then I would like
to especially mention Drs. Luuk Spreeuwers and Haiyun Xu who have contributed important
ideas and discussions during my stay. I had the chance to meet Dr. Haiyun Xu at her home and
enjoyed the company of her two cute sons. Also, the Chinese researchers in the group, Xiaoying
Shao, Meiru Mu, Yuxi Peng, Chanjuan Liu and Jen-Hsuan Ho, were partly responsible of my
happiness during the stay, with their kind attention and help, together with their colleagues at
the laboratory who I will never forget, Rita Lopes Simoes, Chris van Dam, Pinar Santemiz.
My second stay was at the Netherlands Forensic Institute (NFI), with Dr. Didier Meuwly.
During those two months I received an enormously motivating support and confidence from Dr.
Didier Meuwly, and I could benefit from his mastery in the field of forensic science, which led
to some contributions contained in this Thesis. I had also the opportunity to meet Dr. Anko
Lubach who is the professional fingerprint examiner and has taught me a lot in a practical way
such as finger printing and manual minutiae annotation. I would also like to thank Drs. Charles
Berger and Arent de Jongh for their inspiring discussion and advices which helped some part of
the research in this work. For the happy company during my stay, I would like to thank Rudolf
Haraksim, Vikram Doshi, Sheryl Lie Kwie and Miriam Mieremet, who have made me love the
city Den Haag.
I would also like to acknowledge a number of researchers who have helped to shaped this
PhD Thesis with their support and inspiring work. These researchers include Prof. Guangshun
Shi who helped with the important forensic palmprint databases, Dr. Jianjiang Feng, Profs. Phil
Rose and Geoffrey Morrison.
And of course I want to thank all the work mates (and friends) at ATVS who have sha-
red my daily life and filled it with so many precious memories and made me feel at home.
It is unbelievable what a wonderful life you have made for me, mis queridos AMIGOS! With
great gratitude, this huge THANK YOU goes to (in no particular order): Javier Galbally,
Javier Gonzalez-Dominguez, Ruben Vera Rodriguez, Ignacio Lopez-Moreno, Miriam Moreno-
Moreno, Maria Puertas, Pedro Tome, Marta Gomez-Barrero, Ram Prasad Krishnamoorthy,
Javier Franco-Pedroso, Sara Antequera, Alvaro Dieguez, Fernando Espinoza-Cuadros, Ester
Gonzalez-Sosa, Alicia Lozano, Derlin Morocho-Checa, Javier Eslava , Silvia Gabriel, Fátima
Garcı́a-Donday, Ricardo Landriz-Lara, Sandra Uceda, Ruben Zazo, Maria Merida-Aguilera, Jai-
me Ortiz-Lopez, Alvaro Garcia-Maroto, Veronica Peña, Eugenio Arevalo Gonzalez. And some
special thanks are to a bunch of annoying members, Ruben Vera Rodriguez, Pedro Tome, Marta
Gomez-Barrero, Ignacio Lopez-Moreno, Miriam Moreno-Moreno, Maria Puertas, Javier Galbally,
Javier Gonzalez-Dominguez, Sara Antequera, who have successfully turned me into a party girl!!!
And I would also like to mention Galbally for not letting me alone in the club of Corredores
Locos. Por fin, muchas gracias a todos without ending!
Finally and most importantly, many thanks are to my dearest brother who has given un-
conditional support and encouragement no matter how much he missed me and how much he
wanted me to stay in Beijing around him. I owe him apology for not taking good care of him
during these three years. And more thanks are to my dearest friends who treated me as a family
member and shared the moments of joy, stress, laughs and bad moods with me in distance, in-
cluding Xiuli Li, Xunqiang Tao, Ying Jin and Qiujuan Gao. The last but not the least, I would
like to thanks all my running friends who kept encouraging me not to give up any chances of
running, especially my running tutor Xiaojian Li. Thanks to all!
Ruifang Wang
Madrid, October 2013
Mathematical Notation
A consistent mathematical notation has been tried throughout this Dissertation, sometimes
at the expense of usual or original conventions in other fields or works. Following symbols denote
corresponding definitions:
Abstract VIII
Acknowledgements XIV
1. Introduction 1
1.1. Palmprint Evidence in Forensic Science . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1. Forensic Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.2. Palmprint Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3. Fusion in Biometrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.1.4. Evidence Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2. Motivation of the Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3. The Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4. Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5. Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.6. Research Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2. Related Works 13
2.1. High-Resolution Palmprint Recognition . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1. Full-to-Full Palmprint Recognition . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2. Partial-to-Full/Latent-to-Full Palmprint Recognition . . . . . . . . . . . . 20
2.2. Feature Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.1. Radial Triangulation Representation . . . . . . . . . . . . . . . . . . . . . 23
2.2.2. Spectral Minutiae Representation . . . . . . . . . . . . . . . . . . . . . . . 24
2.3. Likelihood Ratio Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4. Chapter Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 29
xix
CONTENTS
3. Experimental Framework 31
3.1. Performance Evaluation of Biometric Systems . . . . . . . . . . . . . . . . . . . . 31
3.1.1. Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.1.2. Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2. Performance of Likelihood-Ratio-Based Interpretation Methods . . . . . . . . . . 36
3.2.1. Protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.2. Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.3. Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.1. Palmprint Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3.2. Fingerprint Databases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.4. Chapter Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 44
xx
CONTENTS
xxi
CONTENTS
xxii
List of Figures
1.1. Sample palmprints. (a) A sample latent palmprint, (b) a sample full palmprint
with high resolution and its feature types. Note that the three major creases in
(b), namely, heart line, head line and life line (up to down). . . . . . . . . . . . . 2
1.4. Fusion can be accomplished at various levels in a biometric system. Most multi-
biometric systems fuse information at the match score level or the decision level.
FE: feature extraction module; MM: matching module; DM: decision-making
module; FM: fusion module. [Ross et al., 2006] . . . . . . . . . . . . . . . . . . . 7
2.6. Five sectors of a full palmprint image defined in [Jain and Demirkus, 2008]. r and
c are the height and width of the original full image respectively; Point P is the
starting point of the heart line; d1 is defined as a vertical distance relative to P
to decide the sizes of sector 1, 2, 3 and 4; d2 , d3 , and d4 are all defined to decide
the sizes of sector 5, while d3 , and d4 are defined as a horizontal and vertical
distances relative to P respectively. In [Jain and Demirkus, 2008], based on the
average shape of the palm, they set d1 = 100, d2 = 1000, d3 = 100 and d4 = 100. . 18
xxiii
LIST OF FIGURES
xxiv
LIST OF FIGURES
5.1. Different types of minutiae window for a configuration of 12 minutiae. (a) Rect-
angle, (b) ellipse, and (c) circle. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.2. Window distortion assessment on the paired minutiae data of NFI DB58. (a)
Window scaling in rectangle window, (b) window scaling in circle window, (c)
window scaling in ellipse window, (d) window rotation in degrees. . . . . . . . . . 79
5.3. Window distortion assessment on the paired minutiae data of NIST SD27. (a)
Window scaling in rectangle window, (b) window scaling in circle window, (c)
window scaling in ellipse window, (d) window rotation in degrees. . . . . . . . . . 80
5.4. Point distortion assessment on the paired minutiae data of NFI DB58 (left) and
NIST SD27 (right). (a) and (b) are the relative difference of minutia position, (c)
and (d) are the difference of minutia direction. . . . . . . . . . . . . . . . . . . . 81
xxv
LIST OF FIGURES
5.5. Illustration of the complex spectral minutiae representation procedure using Cylin-
der pulse. (a) A fingerprint and its minutiae; (b) a Cylinder pulse for a minutia
with a radius of a = 10 pixels; (c) the continuous spectrum generated from all
minutiae points in a fingerprint using Cylinder pulse; (d) the Fourier spectrum in
(c) sampled on a polar-linear grid. . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.6. Weighted-SMC comparator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.7. The moving center of the relaxation window at the four closest steps to the genuine
center (red). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.8. CMC curves for Weighted-SMC comparator using Dirac, Gaussian and Cylinder
pulse respectively on NFI DB58 compared to the original SMC method using the
Dirac pulse. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.9. Rank-1 identification rates of Weighted-SMC comparator using Dirac pulse under
different moving rate values (p2 ) on NFI DB58. . . . . . . . . . . . . . . . . . . . 94
5.10. CMC curves for Weighted-SMC comparator on NIST SD27. . . . . . . . . . . . . 95
5.11. CMC curves for Weighted-SMC comparator on the forensic palmprint databases
compared to RT comparator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.1. CMC curves for multi-algorithm fusion using RT-based system and Weighted-
SMC based system on the forensic palmprint databases. . . . . . . . . . . . . . . 103
6.2. Histogram of the LR values from the fused system using logistic regression. The
upper: same source (SS); The bottom: different source (DS). . . . . . . . . . . . 104
6.3. Tippett plot for the LR values from the fused system using logistic regression. . 105
6.4. ECE plot for the LR values from the fused system using logistic regression. . . . 105
6.5. Definitions of a palmprint: principal lines (1 - heart line, 2 - head line and 3 -
life line), palmprint regions (I - interdigital region, II - thenar region and III -
hypothenar region) and datum points (a, b-endpoint, o-their midpoint) [Zhang
and Shu, 1999]. c is the intersection point of the bottom boundary of a palm and
the perpendicular bisector of the line segment ab. Here ab and oc are used for
region segmentation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.6. Segmented regions for a sample palmprint. . . . . . . . . . . . . . . . . . . . . . . 107
6.7. ROC curves for the three region-to-region comparisons using the SMC method
compared to full-to-full matching. . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.8. ROC curves for the three region-to-region comparisons using MegaMatcher 4.0
compared to full-to-full matching. . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.9. Scatter plots of scores for the three region-to-region score sets obtained by Mega-
Matcher 4.0. (a) Interdigital (S1 ) vs. thenar (S2 ), (b) interdigital (S1 ) vs. hy-
pothenar (S3 ), (c) hypothenar (S3 ) vs. thenar (S2 ). . . . . . . . . . . . . . . . . . 110
6.10. Matching strategy for regional fusion. . . . . . . . . . . . . . . . . . . . . . . . . 111
6.11. ROC curves of regional fusion using sum rule and logistic regression respectively
compared to full-to-full comparison for the SMC-based matching method. . . . . 113
xxvi
LIST OF FIGURES
6.12. ROC curves of regional fusion using sum rule and logistic regression respectively
compared to full-to-full comparison for MegaMatcher 4.0. . . . . . . . . . . . . . 113
6.13. Canny edge detection for a sample palmprint. (a) Input image, (b) smoothed
image, (c) gradient magnitudes (IG ), (d) final strong edges detected (IE ). . . . . 116
6.14. Thresholding of gradient magnitudes. (a) Gradient magnitude image after thresh-
olding (IG0 ), (b) Left differential image after thresholding (IGl ). . . . . . . . . . . 118
6.15. Detected datum points by using convex hull. (a) Convex hull on IG0 and IGl , (b)
convex hull on IG0 and IE , (c) final datum points a and b. . . . . . . . . . . . . . 119
6.16. Regional segmentation error. (a) Automatically segmented region, (b) manually
segmented region, (c) error area. . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.17. Sample images where datum point b was not detected. (a) A sample image from
Subset B without datum point b visible, (b) A failed image from Subset A with
datum point b visible. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.18. ROC curves for region-to-region comparison and regional fusion using regions
segmented automatically. (a) Region-to-region comparison, (b) sum fusion vs.
full-to-full comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
xxvii
LIST OF FIGURES
xxviii
List of Tables
3.1. Summary of palmprint databases used in this Thesis. All images present 500 ppi. 42
5.1. The numbers of minutiae in the paired minutiae data of the databases NIST SD27
and NFI DB58. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2. Window distortion assessment by different minutiae windows for fingermarks in
NFI 58 with regard to their mated prints. . . . . . . . . . . . . . . . . . . . . . . 77
5.3. Window distortion assessment by different minutiae windows for fingermarks in
NIST SD27 with regard to their mated prints. . . . . . . . . . . . . . . . . . . . . 78
5.4. Description of parameters for spectral minutiae representation and comparison. . 84
5.5. Description of parameters for Weighted-SMC comparison. . . . . . . . . . . . . . 91
5.6. Values of parameters for Weighted-SMC comparator and the original SMC method
on NFI DB58. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.7. Optimized values of parameters for Weighted-SMC comparator on NIST SD27. . 95
5.8. Performance comparison between the Weighted-SMC comparator and the best
minutiae-only based system published in NIST ELFT-EFS Evaluation on NIST
SD27. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.9. Optimized values of parameters for the Weighted-SMC comparator on forensic
palmprint databases. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.10. Comparison time of the Weighted-SMC comparator for fingerprints and palm-
prints respectively. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.1. Region-to-region comparison and full-to-full comparison results using the SMC
method with Dirac pulse. [Wang et al., 2013f,g] . . . . . . . . . . . . . . . . . . . 108
6.2. Correlation results of different regional scores. . . . . . . . . . . . . . . . . . . . . 111
6.3. The categories of the sub-database from THUPALMLAB. . . . . . . . . . . . . . 122
6.4. Error rates of automatic region segmentation compared to manual segmentation
on Subset A (i.e., datum point b is visible and successfully detected). . . . . . . . 123
xxix
LIST OF TABLES
xxx
Chapter 1
Introduction
This PhD Thesis is focused on building robust and efficient palmprint recognition systems
in forensic applications. In particular, the Thesis is intended to provide a better understanding
of the large non-linear distortion problem in forensic palmprint recognition, fusion of palm
regions, combination of rich types of features in a high-resolution palmprint image, and evidence
evaluation using the similarity scores, i.e, the output of forensic palmprint recognition systems.
Modern forensic science is highly influenced by the use of biometric systems, mostly for
source searching in large biometric databases. Amongst all the biometric traits in forensic
science, fingerprints are possibly the most popular, due to their high power to discriminate
between different identities. In recent years, palmprints have also been increasingly used [Dewan,
2003], and current commercial Automated Fingerprint Identification Systems (AFIS) used in
forensic science typically include palmprint identification functionalities. As an example of the
importance of these biometric traits, the FBI even created a Latent Hit of the Year Award to an
outstanding examiner who worked on major violent crime cases solved with the assistance of an
AFIS [FBI, 2006a]. Moreover, fingerprint comparison by human examiners has been used and
admitted in court since the beginning of the 20th century, and more recently for palmprints.
Although fingerprint identification procedures have been criticized in the last years in the context
of modern forensic science [Eric H. Holder et al., 2011; Spinney, 2010] (especially considering
the so-called Paradigm Shift towards a more scientific framework [Saks and Koehler, 2005]),
recent work is filling the gap towards scientific procedures that are reinforcing the admissibility
of fingerprint evidence in US courts [Neumann et al., 2012; Ulery et al., 2011].
In this context, palmprint evidence is increasingly important according to surveys by law
enforcement agencies, which indicate that 30 percent of the latent marks recovered from crime
scenes are from palms [Dewan, 2003]. This importance is evidenced in initiatives like the Na-
tional Palm Print System (NPPS), included in FBI’s Next Generation Identification (NGI)
project [FBI, 2006b], with the aim of creating a centralized repository for palmprint data [FBI,
2012] in order to foster source searching in palmprint databases using biometric technology.
However, despite the advances in AFIS technologies due to its extensive use in the last decades,
source searching based on latent palmprints still faces particular challenges with respect to fin-
1
1. INTRODUCTION
(a)
Singular points
Interdigital Region
Major Creases
Figure 1.1: Sample palmprints. (a) A sample latent palmprint, (b) a sample full palmprint with high
resolution and its feature types. Note that the three major creases in (b), namely, heart line, head line
and life line (up to down).
gerprints. First, the low quality of latent palmprints is worsen by the fact that they are usually
quite partial compared to full palmprints in reference databases (see Fig. 1.1 (a)). This dif-
ference in size is typically much higher than in fingerprints. Second, due to the large number
of creases and minutiae of high-resolution palmprints typical in databases, much larger storage
and computation capabilities are required for image processing and comparison as compared
to fingerprints (see Fig. 1.1 (b)). Third, minutiae extraction is much more difficult than in
fingerprints, since ridges are not so clear in some regions, and creases that are confused with
ridges can result in many spurious minutiae. Finally, sometimes blank regions appear in full
palmprints, even when they are acquired in controlled conditions, being an additional problem
when a latent palmprint has to be aligned with a full palmprint.
2
1.1 Palmprint Evidence in Forensic Science
The first type is known as recovered material, sample or mark, and is transferred to the
crime scene from a person involved (e.g., recorded speech, fingermarks, latent palmprints,
etc.) or vice-versa (e.g., glass fragments from a window, fibers from the victim’s clothes,
etc.).
The second type is called control material, in this case for samples whose origin or source
is known. Materials can be recovered from the crime scene (remains of a broken window)
or directly from a suspect (fingerprints or palmprints).
The role of forensic scientists is to examine the recovered material and control material to
assess the contribution of these findings to the choice between two opposite hypotheses [Evett,
2011]. When the assumptions are about whether the source of both samples is the same, we
are facing a problem of the attribution of sources [Ramos, 2007] or a question at the source
level [Cook et al., 1998], which attempts to answer the question: Do the recovered material and
control material come from the same source?
3
1. INTRODUCTION
Crime Transfer
Criminal
Scene Evidence
...blackmail...
? ?
Crime scene Criminal
(source) Criminal Crime Scene (source)
epidermal ridge patterns are called the palmar flexion creases. These are the firmer attachment
areas to the basal (dermis) skin structure. Flexion creases appear before the formation of friction
ridges during the embryonic skin development stage, and both of these features are immutable,
permanent and unique to an individual [Ashbaugh, 1999]. There are three main groups of
flexion creases appearing on the palm, namely major flexion creases, minor flexion creases and
secondary creases [Ashbaugh, 1999].
Generally, an automatic palmprint identification system consists of four major components:
User Interface Module, Acquisition Module, Recognition Module and External Module, as shown
in Fig. 1.3. According to the way the palmprint is acquired, automatic palmprint identification
systems can be classified into two categories: online and offline [Zhang et al., 2003]. An online
system captures palmprint images using a palmprint capture sensor that is directly connected to
a computer for real-time processing. An offline palmprint identification system usually processes
previously captured palmprint images, which are often obtained from inked or latent palmprints
that were digitalized by a digital scanner. In general, online palmprint recognition systems use
low-resolution images with a resolution of 100dpi or less usually and focuses in more general
features, such as main lines or texture [Connie et al., 2005; Zhang et al., 2003]. Though they
can achieve a very high performance, they are only suitable for commercial or civil applications,
not for forensic applications where the palmprint images are usually captured or recovered at
500 dpi at least [Allinson, 2011]. Offline systems [Cappelli et al., 2012; Dai et al., 2012; Dai
and Zhou, 2011; Tan et al., 2009; Yang et al., 2011] use high-resolution images, allowing the
extraction of detailed features such as minutiae or ridges, thus, they can be used for forensic
purposes.
However, compared to the extended use of Automated Fingerprint Identification Systems
(AFIS) in the forensic community, the development of palmprint identification systems that
4
1.1 Palmprint Evidence in Forensic Science
Figure 1.3: A framework of automatic palmprint identification system. [Lu et al., 2007]
perform as well as AFIS do is still a big challenge. This is due to the reasons explained previously:
large number of creases and large non-linear distortion in a high-resolution palmprint image,
large storage and computation capabilities required for processing and matching palmprints, etc.
In the literature, only few works [Jain and Demirkus, 2008; Jain and Feng, 2009; Liu et al., 2013]
have been found that develop palmprint identification specifically aimed to be used in forensic
applications.
5
1. INTRODUCTION
Sensor level fusion refers to the combination of raw data from the biometric sensors.
Feature level fusion refers to the combination of different feature vectors obtained either
with different sensors or by applying different feature extraction algorithms to the same
raw data.
Score level fusion refers to the combination of matching scores provided by different sys-
tems.
Decision level fusion refers to the combination of decisions taken by the individual systems.
For palmprint recognition systems in forensic applications, the match score level fusion is
achievable for most cases. Moreover, palmprint quality information can be utilized in many
score level fusion algorithms. Some works [Baker et al., 2005; Fierrez-Aguilar et al., 2005a,b]
have been done to develop quality-based score level fusion technologies.
1. Computation of the similarity score between the control and recovered palmprint
materials using a biometric system.
3. Reporting of results to the court. The reported LR value measures the value of the
forensic evidence in the inferential process for the given case [Evett, 2011]. Moreover, a
proper validation of the technique should be accessible in case that the fact finder will
question the admissibility of the procedure, or according to court demands.
The above three steps also can be classified as three abstraction levels [Ramos, 2007]: a) the
discrimination level where the systems compute the score; b) the presentation level which allows
the computation of a LR value from a score; c) the forensic level which finally form the forensic
report including the LR value.
6
1.2 Motivation of the Thesis
FM
Se n sor
Le ve l
Fu sion
D e cision
FE
Le ve l
Fu sion MM Right Eye
DM
FE MM DM FE
Match
Non-
FE MM DM Match FE
Fe a t u r e
DM
Le ve l
Fu sion Left Eye
FM
275 0.4 58
MM MM MM
Scor e
FE FE FE
Le ve l
Fu sion
Figure 1.4: Fusion can be accomplished at various levels in a biometric system. Most multibiometric
systems fuse information at the match score level or the decision level. FE: feature extraction module;
MM: matching module; DM: decision-making module; FM: fusion module. [Ross et al., 2006]
Due to the fact that few systems [Jain and Feng, 2009; Liu et al., 2013] aimed at forensic
palmprint recognition have been developed, deeper research is needed in this field in order
to improve robustness, accuracy and computational efficiency of automated forensic palm-
print recognition technologies, aimed at realistic forensic scenarios. This motivates us to
develop novel algorithms based on minutiae features.
In the search of novel minutiae-based comparison techniques, we find that the spectral
minutiae representation [Xu et al., 2009; Xu and Veldhuis, 2010] proposed recently has
shown its robustness in minutiae-based fingerprint recognition and good computational
efficiency at the matching stage. The latter is essential in large-scale forensic palmprint
recognition and fusion of systems. This motivates us to study the spectral minutiae rep-
resentation for palmprint recognition and apply the method to forensic scenarios.
7
1. INTRODUCTION
As indicated in [Jain and Feng, 2009], the three palm regions divided by three major
creases, i.e., interdigital, hypothenar and thenar, have different performance according to
matching accuracy, which meant that the thenar region had much lower accuracy than
interdigital and hypothenar regions. This property can be considered as a unique aspect
of palmprints while it does not exist in fingerprints, and motivates regional analysis for
high resolution palmprint recognition and fusion of different palmprint regions which can
be further applied to forensic scenarios.
1.4. Objectives
Given the problems and needs described in the previous sections, the following objectives
are pursued during the course of this Thesis:
Reviewing and studying existing techniques and problems of forensic palmprint recogni-
tion.
Computing likelihood ratios from the output of forensic palmprint recognition systems
as a first step towards the scientific evaluation of evidence in palmprints using biometric
systems.
8
1.5 Outline
1.5. Outline
The Dissertation is structured according to a traditional complex type [Paltridge, 2002]
with literature review, theoretical and practical methods and experimental studies in which the
methods are applied. Chapters are structured as follows:
Chapter 1 has introduced the basic concepts of palmprint recognition in forensic scenarios,
a description of the main problems, and the motivation of this Dissertation. Research con-
tributions originated from this Thesis will also be exposed at the end of this introductory
chapter.
Chapter 3 describes our experimental framework for all the systems developed in this
Thesis. It includes protocols and measures for performance evaluation of biometric systems
in commercial and forensic applications, and descriptions of the databases of palmprints
and fingerprints we use in this Thesis.
Chapter 4 describes the proposed methods and experimental evaluation of a novel forensic
palmprint recognition system based on radial triangulation.
Chapter 5 describes the proposed methods and experimental evaluation of a novel forensic
palmprint recognition system based on a weighted complex spectral minutiae representa-
tion, with applications in fingerprints and palmprints.
Chapter 6 describes fusion schemes for high-resolution palmprint recognition and presents
experimental evaluation results. Multi-algorithm fusion at score level is implemented for
forensic palmprint recognition using the two algorithms proposed in Chapters 4 and 5.
And regional fusion is implemented for high-resolution palmprint recognition.
Chapter 7 concludes the Dissertation summarizing the main results obtained and outlining
future research lines.
The dependence among the chapters is depicted in Figure 1.5.
1. Feature extraction for high-resolution palmprint recognition [Wang et al., 2011, 2013b]
(Best Poster Paper Award).
9
1. INTRODUCTION
3. Distortion assessment in latent fingerprints from forensic casework. [Wang et al., 2013e].
Theoretical frameworks.
1. Theoretical framework for distortion assessment at feature level for palmprints and finger-
prints [Wang et al., 2013a,e].
Novel methods.
2. Radial triangulation based global structure comparison [Wang et al., 2012a, 2013b].
4. Weighted complex spectral minutiae representation for forensic fingerprint and palmprint
comparison [Wang et al., 2013a].
5. Automatic region segmentation based on convex hull comparison for high-resolution palm-
print recognition [Wang et al., 2013c].
New systems.
1. A forensic palmprint recognition system based on radial triangulation [Wang et al., 2013b].
2. Study on the use of spectral minutiae representation for high-resolution palmprint recog-
nition [Wang et al., 2013g].
10
1.6 Research Contributions
3. Distortion assessment comparing public forensic fingerprint databases and realistic databases
from actual forensic cases [Wang et al., 2013e].
4. Manual segmentation of high-resolution palmprints using major creases features, i.e., da-
tum points [Wang et al., 2013d].
5. Experimental results of evidence evaluation using all the four kinds of forensic palmprint
recognition systems.
Other contributions related to the problem explored in this Thesis but not presented in the
Dissertation include:
Theoretical frameworks.
1. A new practical framework of latent fingerprint examination integrating the evidence eval-
uation using scores from matched fingerprint templates. [Krish et al., 2013a,b,c].
2. Study on the quantification of the weight of evidence with partial finger marks [Krish
et al., 2013a,c].
11
1. INTRODUCTION
Chapter 1: Chapter 2:
Introduction Related Works
Chapter 4:
RT-based Forensic Palmprint
Recognition
Chapter 3:
Experimental Framework
Chapter 5:
Weighted-SMC for Forensics:
Fingerprints and Palmprints
Chapter 6:
Fusion Schemes for High-
Resolution Palmprint Recognition
12
Chapter 2
Related Works
This chapter summarizes the works related to this PhD Thesis. We focus on: i) the state-of-
the-art in high-resolution palmprint recognition with special interest in the techniques that can
be applied to forensic applications; ii) techniques of feature representation for fingerprints that
can be applied to forensic palmprint recognition; iii) approaches of likelihood ratio computation
for evidence evaluation using biometric systems. During this literature review, we can foresee
the importance of the research conducted in this Thesis, due to the relatively recent history
of forensic palmprint recognition using biometric systems, the remaining challenges and the
potential room for performance improvement.
This chapter is organized as follows. First we give an overview of the most important works
on the study of high-resolution palmprint recognition (Sect. 2.1) describing two major matching
scenarios, i.e., full-to-full and partial-to-full. Then in Sect. 2.2 we review novel methods of
feature representation for fingerprints that can be adapted for high-resolution palmprints because
they are based on minutiae, a common feature of fingerprint and palmprint. The next section is
dedicated to a summary of likelihood ratio computation methods using biometric systems (Sect.
2.3). Finally, the summary and conclusions of this chapter are presented (Sect. 2.4).
13
2. RELATED WORKS
between palm marks and complete high-resolution palmprints. In this section, we will review
previous works for each comparison scenario, paying special attention on the techniques that
can be used for forensic scenarios, and remarking the challenges associated with each technique.
Only in recent years, there are several works proposing full-to-full palmprint recognition
systems [Cappelli et al., 2012; Dai et al., 2012; Dai and Zhou, 2011; Tan et al., 2009]. According
to the features used for matching, these systems can be classified into two categories: i) minutiae-
based, when only minutiae are used, and ii) multifeature-based, when minutiae are used as well
as other features.
Starting from 2007, the research group in the Institute of Machine Intelligence in Nankai
University published several works on minutiae-based offline palmprint recognition techniques,
including palmprint segmentation [Zheng et al., 2007a], enhancement [Zheng et al., 2007b], fea-
ture extraction and postprocessing [Li and Shi, 2008; Li et al., 2009; Yang et al., 2009]. In 2009,
they built a minutiae-based offline palmprint identification system [Tan et al., 2009] using the
techniques developed previously. The system architecture is shown in Fig. 2.2. It includes three
major modules, i.e., image processing, feature extraction and feature matching. At the matching
stage, there are four steps, i.e., local structure construction (see Fig. 2.3), local matching, global
datum mark election and global matching. Due to the simple nearest-neighbor-based approach
used for local matching, the system only obtained a rank-1 identification rate of 59.02% while
evaluated on a uni-impression high-resolution palmprint database containing 49 testing full im-
ages and 9600 reference full images. However, the techniques of image preprocessing and feature
postprocessing they developed are applicable to forensic scenarios, and are especially useful for
feature extraction in high-resolution full palmprints.
In 2008, the group from Michigan State University developed a complete full-to-full palmprint
recognition system [Jain and Demirkus, 2008] only based on minutiae. The block diagram of
that full-to-full palmprint matching system is shown in Fig. 2.4. The system consists of three
major components: (1) pre-processing, (2) feature extraction, and (3) minutiae matching. The
preprocessing phase segments the palmar area. The feature extraction phase first extracts
minutiae using a commercial minutiae extractor Neurotechnology VeriFinger 4.2 which is an
early version of MegaMatcher 4.0 [Neurotechnology, 2011], then obtains the crease map (see Fig.
2.5) and eliminates spurious minutiae located on the creases, and finally extracts the minutiae
based on the crease map. The matching phase registers the palmprint according to a common
coordinate axis, i.e., the heart line, and does a division in 5 sectors (see Fig. 2.6). Sector-wise
minutiae match scores are obtained and then fused to get the final match score between two
palmprints. For minutiae matching, Neurotechnology VeriFinger 4.2 [Neurotechnology, 2011],
is utilized to obtain both the optimal alignment between the query and template images, and a
14
2.1 High-Resolution Palmprint Recognition
Comparison scenarios
(a) Full-to-full
(b) Partial-to-full
Figure 2.1: Comparison scenarios for High-resolution palmprint recognition. (a) Full-to-full, (b) partial-
to-full. In the partial-to-full scenario, there are three kinds of partial images: i) synthetic partial palm-
print, ii) pseudo partial palmprint, and iii) latent palmprint.
15
2. RELATED WORKS
Segmentation
Figure 2.2: Minutiae-based offline palmprint identification system in [Tan et al., 2009].
Figure 2.3: k-nearest local structure for a minutia P in [Tan et al., 2009].
match score. The system achieves a genuine matching accuracy of 98.9% at a false acceptance
rate (FAR) of 0.01% for full-to-full palmprint matching on an in-house database of 100 unique
palms (50 subjects × 2 palms per subject) with 10 impressions per palm. In the same publication,
a partial-to-full palmprint matching system was also proposed , which will be described in a
later section (Sect. 2.1.2). The techniques of crease map extraction and alignment in this work,
and sector-wise minutiae matching and fusion are valuable for forensic applications. Especially
sector-wise minutiae matching and fusion inspires our work of regional fusion which will be
detailed in Chapter 6. However, their sector division is only based on one datum point, i.e., the
starting point of the heart line, while our method will use all the three datum points as defined
in [Zhang and Shu, 1999] and divide a palmprint into three characteristic regions.
In 2012, the researchers from University of Bologna developed a full-to-full palmprint recog-
nition system [Cappelli et al., 2012] by applying a novel local structure coined Minutia Cylinder-
Code (MCC) which was originally proposed for fingerprint matching [Cappelli et al., 2010]. The
16
2.1 High-Resolution Palmprint Recognition
Figure 2.4: Full-to-full palmprint matching system in [Jain and Demirkus, 2008].
Figure 2.5: Palmar crease map extraction. [Jain and Demirkus, 2008]
MCC representation associates a local structure to each minutia. This structure encodes spatial
and directional relationships between the minutia and its (fixed-radius) neighborhood and can be
conveniently represented as a cylinder whose base and height are related to the spatial and direc-
tional information, respectively (see Fig. 2.7). In the system for palmprints, minutiae matching
was performed by means of a local matching strategy based on the MCC representation applying
simple bitwise operations between the two corresponding binary vectors as described in [Cap-
pelli et al., 2010], followed by a relaxation procedure to derive a global matching score (further
details about it in [Cappelli et al., 2012]). Results on a public multi-impression high-resolution
17
2. RELATED WORKS
Figure 2.6: Five sectors of a full palmprint image defined in [Jain and Demirkus, 2008]. r and c are
the height and width of the original full image respectively; Point P is the starting point of the heart line;
d1 is defined as a vertical distance relative to P to decide the sizes of sector 1, 2, 3 and 4; d2 , d3 , and
d4 are all defined to decide the sizes of sector 5, while d3 , and d4 are defined as a horizontal and vertical
distances relative to P respectively. In [Jain and Demirkus, 2008], based on the average shape of the
palm, they set d1 = 100, d2 = 1000, d3 = 100 and d4 = 100.
palmprint database, THUPALMLAB [Tsinghua Univ., 2011b], show that the proposed system
compares very favorably with the state of the art, with a EER < 0.01% as reported in [Cappelli
et al., 2012] and high efficiency, i.e., average speeds of 1.935s for feature extraction and 0.038s
for matching. However, the system has not been evaluated on uni-impression/partial-to-full
palmprint databases more representative of forensic scenarios. Moreover, the system assumes a
full-to-full matching scenario and it is unclear how the MCC-based local matching strategy would
cope with the forensic scenario where the comparison is between a partial or latent palmprint
with few minutiae and a full palmprint with much more minutiae.
18
2.1 High-Resolution Palmprint Recognition
Figure 2.7: A graphical representation of the local structure associated to a given minutia: (a) the
cylinder with the enclosing cuboid, (b) the discretization of the cuboid into cells of size ∆S × ∆S × ∆D :
Only cells whose center is within the cylinder are shown, (c) a cell can be uniquely identified by three
indices (i, j, k), (d) the three-dimension axis, and (e) the direction of the given minutia. Note that the
cuboid is rotated so that the axis i is aligned to the direction of the given minutia. [Cappelli et al., 2010]
ever, the system presents a high computational burden, with average speeds of 67s for feature
extraction and 5.160s for matching.
Query
Cascade Filtering Segment-Based Matching and Fusion
Registration
Filter 1
N
P
Identification stage
Registration
n
Non-match Identity/Unknown
Figure 2.8: Outline of the palmprint identification system in [Dai et al., 2012].
As an improvement of the previous system, [Dai et al., 2012] built a new multifeature-based
19
2. RELATED WORKS
Figure 2.9: Relations between the three characteristics of palmprints and the three modules of the
proposed matching algorithm in [Dai et al., 2012].
palmprint identification system (see Fig. 2.8) which consists of three modules: palmprint reg-
istration, cascade filtering, and segment-based matching and fusion. The relations between the
three modules and the major characteristics of palmprints are shown in Fig. 2.9. An orientation
field based registration algorithm is designed for registering palmprints of different positions
and rotations into the same coordinate system before matching. To reduce the computational
complexity, a cascade filter is built to reject non-mated gallery palmprints in early stage by com-
paring just a small portion of the whole palmprint. To deal with the distortion and the varying
discrimination power of different palmprint regions, a segment-based palmprint matching and
fusion algorithm is proposed. The whole palmprint image is divided into small segments which
are then separately matched to deal with distortion. The similarity between two palmprints is
calculated by fusing the similarity scores of different segments using a Bayesian framework. The
system achieves a True Positive Identification Rate (TPIR) of 97.9% when the False Positive
Identification Rate (FPIR) is set as 2 × 10−3 in matching 840 full palmprints against the gallery
set of 13,736 palmprints which are inked images from 13,616 different palms (one print per palm)
provided by the Forensic Science Department of the Chinese Police. Also, the system is compu-
tationally efficient, with average matching speeds of 161ms for genuine attempts and 39ms for
impostor attempts in the full-to-full matching scenario. However, their equally divided regions
are less meaningful compared to our proposal of dividing a palmprint into three characteristic
regions, i.e., interdigital, thenar, and hypothenar regions. The system is also evaluated under
partial-to-full comparison scenario. These results will be shown in the next section.
Partial-to-full palmprint recognition deals with the matching between partial images and
complete high-resolution images. The partial images can be from two classes:
Simulated latent palmprints which are either cropped from complete high-resolution
palmprints (i.e., synthetic type) or captured from a partial area of a palm (i.e., pseudo
type) (see Fig. 2.1 (b)).
20
2.1 High-Resolution Palmprint Recognition
Figure 2.10: Partial-to-full palmprint matching system in [Jain and Demirkus, 2008].
The works proposing this kind of palmprint recognition systems [Jain and Demirkus, 2008;
Jain and Feng, 2009; Liu et al., 2013], which are mainly based on minutiae and will be detailed
below, are also quite recent aiming to be applicable to forensic scenarios. There are other works
[Singh et al., 2013, 2012] dealing with partial-to-full palmprint comparison which are not based
on minutiae and will not be detailed due to the focus of this Thesis.
In [Jain and Demirkus, 2008], a partial-to-full palmprint recognition system based on minu-
tiae and Scale Invariant Feature Transformation (SIFT) [Lowe, 2004] is presented. The block
diagram of the proposed partial-to-full palmprint matching system is shown in Fig. 2.10. The
system consists of the three major components: (1) latent region-of-interest (ROI) detection, (2)
feature extraction, and (3) feature matching. Latent ROI detection phase is only applied to the
full palmprint images. Since it is generally assumed that the latent (partial) palmprints come
from specific regions of the palm (i.e. thenar, hypothenar and interdigital), these regions are
automatically detected and features from these regions are utilized in the matching phase. The
feature extraction phase obtains the crease map and SIFT feature points (i.e., repeatable charac-
teristic points such as local extrema), and extracts the minutiae based on the crease map. In the
feature matching phase, minutiae and SIFT matchers are used in parallel to obtain two match
scores. Score-based fusion is utilized to obtain the final match score. The system achieves
a rank-1 identification rate of 96% in matching 500 synthetic latent palmprints and 100 full
palmprints, and a rank-1 identification rate of 82% on the pseudo-latent database including 240
images (10 subjects × 2 palms per subject × 12 partial print per palm). This work is valuable
as a pioneering proposal towards forensic palmprint recognition. However, the computational
21
2. RELATED WORKS
efficiency of the system is quite low. The average time taken for palmprint preprocessing, fea-
ture extraction, matching of minutiae and SIFT are 25s per image, 15s per image and 5s per
match, respectively, when tested on a PC with 2 GB of RAM and a 2.13 GHz Intel Core 2 CPU.
Another limitation of the work is that SIFT features can not be consistently detected in latent
palmprints and full prints.
The previously described multifeature-based palmprint identification system [Dai et al., 2012]
is also evaluated under partial-to-full comparison scenario, and achieves a TPIR of 91.9% when
the FPIR is set as 5 × 10−4 in matching 2,520 partial palmprints against the gallery set of
13,736 palmprints. The partial palmprints they use are cropped from the thenar, hypothenar
and interdigital regions of full palmprints. Though the system shows its identification power
in partial-to-full palmprint matching, the orientation field based registration method will face
difficulties under real forensic scenarios due to the low quality and high noise of latent palmprint
images.
In 2009, a pioneering complete latent-to-full palmprint recognition system was proposed [Jain
and Feng, 2009]. To deal with creases in palmprints, a region growing algorithm is proposed
to reliably estimate the ridge direction and frequency in the sine wave representation of a local
ridge block. This is then used for robust minutiae extraction. To reduce the computational
complexity of the minutiae matching algorithm, a fixed-length minutia descriptor, MinutiaCode
(see Fig. 2.11), is proposed, which captures information about the ridges and other minutiae
in the neighborhood of a minutia. The proposed system was evaluated by matching partial
palmprints (150 livescan partial palmprints and 100 latent palmprints from real cases) against
a background database of 10,200 full palmprints. Rank-1 identification rates of 78.7% and
69 %, respectively, were achieved in searching live-scan partial images and latents against the
background database. However, the system is quite slow, with an average feature extraction
time of 7s for partial palmprints and 22s for full palmprints, and an average matching time of
0.34s.
In 2013, as an improvement of the previous system, a coarse to fine minutiae-based latent
palmprint matching system was proposed [Liu et al., 2013]. The flowchart of the new system is
shown in Fig. 2.12. The system consists of three stages:
Registration stage: for a given template of full palmprint image, minutiae and descrip-
tors are extracted and clustered based on the obtained centroids.
Matching stage: given a query latent palmprint image, minutiae and descriptors are first
extracted and clustered as the registration stage, then a small number of initial minutiae
pairs is selected within the same cluster and fed to a match propagation procedure to get
a fine match result.
22
2.2 Feature Representation
Figure 2.11: Configuration of a MinutiaCode. The numbers of four types of neighboring minutiae, RS
(reliable and same direction), US (unreliable and same direction), RO (reliable and opposite direction),
and UO (unreliable and opposite direction), in sectors 1 and 2 are [1 0 1 0] and [0 2 0 0], respectively.
Square indicates reliable minutiae and circle indicates unreliable minutiae. [Jain and Feng, 2009]
The system achieves a rank-1 identification rate of 79.4% in searching 446 latent palmprints
against a background database of 12,489 full palmprints, and a EER of 0.11% on the THU-
PALMLAB database. The system operates quite fast, with average speeds of 141ms for genuine
match and 50ms for impostor match during latent-to-full palmprint matching. However, the
system will face challenges when small overlap exists between a mark and a full palmprint, due
to the dependence of cluster-based matching.
23
2. RELATED WORKS
Figure 2.12: Flowchart of the latent-to-full palmprint matching system in [Liu et al., 2013].
it relies on the construction of a polygon and its triangulation. This results in a structure that
is rotation invariant and more robust to fingerprint distortion, and which does not require a
large amount of storage space and computing time to be calculated and matched. Given a set of
minutia locations there is a unique centroid defined by the arithmetic mean of the coordinates.
The centroid is then used for ordering the minutia locations in a “radar” manner, which is used
to define a unique polygon. Given this polygon the centroid is also used for triangulating it: radii
are added from the vertices of the polygon to the centroid. Figure 2.13 (a) presents the polygon
and radial triangulation of a set of 9 minutiae while Fig. 2.13 (b) shows the standard Delaunay
triangulation for the same configuration of minutiae. Any vertex in the radial triangulation is
linked to two other vertices and to the centroid, while in the Delaunay triangulation the number
of connections for each vertex can vary. The local structure generated by radial triangulation
can be used for local matching in fingerprint recognition systems. The structure is particularly
useful in forensic comparison as carried out by human experts, where there is a reduced set of
minutiae [Neumann et al., 2012].
Recently, the spectral minutiae representation [Xu et al., 2009; Xu and Veldhuis, 2010] has
shown its power in minutiae-based fingerprint recognition. This feature representation can
handle minutiae translation and rotation with good computational efficiency at the matching
stage, and therefore it satisfies the requirements of high-resolution palmprint recognition as well.
As defined by [Xu et al., 2009], the method uses the minutiae locations in the spatial domain,
24
2.3 Likelihood Ratio Computation
Figure 2.13: Polygon triangulation. (a) The radial triangulation of a set of nine minutiae, (b) the
Delaunay triangulation of the same set of minutiae. [Neumann et al., 2007]
takes the Fourier transform of the coded locations and obtains the magnitude of its Fourier
spectrum in the frequency domain. Fig. 2.14 shows the general procedure of spectral minutiae
representation. There are three types of spectral minutiae representations: i) the location-based
(SML), ii) the orientation-based (SMO), and iii) the complex spectral minutiae representation
(SMC). Among them, the enhanced SMC method [Xu and Veldhuis, 2010] performs best for
fingerprints with an EER of 3.05% on FVC2002 DB2A database [University of Bologna, 2011]
and a matching speed of 8000 comparisons per second. Due to the procedures of summation and
normalization of all the spectrums generated from the minutiae set in a fingerprint, the spectral
minutiae representation can be considered as a kind of global structure which can be used for
global matching in minutiae-based systems. It also presents the important property of being
invariant to rotation and translation.
As introduced in Sect. 1.1.1, forensic science refers to the application of science or technology
in the investigation of criminal activities and the establishment of facts or evidence in court
25
2. RELATED WORKS
(a) (b)
(c) (d)
Figure 2.14: Illustration of the general spectral minutiae representation procedure (images from the
SMO case). (a) a fingerprint and its minutiae; (b) representation of minutiae points as real (or complex)
valued continuous functions; (c) the 2D Fourier spectrum of (b) in a Cartesian coordinate and a polar-
logarithmic sampling grid; (d) the Fourier spectrum sampled on a polar-linear grid. [Xu and Veldhuis,
2010]
[Champod, 2009; Dessimoz and Champod, 2007]. In a forensic case where some evidence should
be evaluated, the hypotheses, defined at the source level [Cook et al., 1998], are usually as
follows:
θd : the source of the recovered samples is not the same as the source of the control sample,
but another one from a relevant population of sources.
In order to make a decision, the judge has to accept either the hypothesis θp of the prosecution, or
the hypothesis of the defense θd . The increasingly adopted position [Evett, 2011] is that forensic
experts should provide a degree of support for either of the two hypotheses, based on their
examinations of the trace and reference data. There is a growing interest in a more objective
quantitative support for these hypotheses based on the output of biometric systems [Liu et al.,
26
2.3 Likelihood Ratio Computation
Figure 2.15: Bayesian inferential framework in LR-based evidence analysis where the responsibility of
a forensic scientist is to compute the LR using the evidence coming from the comparison of the recovered
and control samples, e.g., from a biometric system. [Ramos, 2007]
2013; Neumann et al., 2012; Wang et al., 2012a] instead of manual comparison. However, the
output of a score-based biometric system is not directly suitable for quantifying the evidential
value contained in a trace. A suitable measure that is gradually becoming accepted in the forensic
community is the likelihood ratio (LR) which is the ratio of the probability of the evidence given
θp and the probability of the evidence given θd . The use of the Bayesian framework (or the LR-
based framework) as shown in Fig. 2.15 exploits these computed LRs in a very useful manner
towards an informed decision between the two hypothesis by the court. In fact, this methodology
has been proposed as the standard for reporting evaluative opinions in the forensic laboratories
of European Network of Forensic Science Institutes (ENFSI) [ENFSI, 2013], with the support of
the ENFSI board [Evett, 2011]. A general description of this framework can be found in [Aitken
et al., 2004]. In this framework, the responsibility of a forensic scientist is to compute the LR.
The evidence coming from a biometric system can be considered essentially as an observation of
some uncertain variable, the evidence, and the LR is the ratio of the distribution of this variable
under two hypotheses, evaluated at its observed value.
The above-mentioned definition of LR can be obtained starting from the following ratio:
P (θp | E, I)
, (2.1)
P (θd | E, I)
which is the measure the court should use to base his decision between the two hypothesis.
We also use the Bayes’ theorem [Aitken et al., 2004] with the posterior probabilities of both
27
2. RELATED WORKS
hypothesis:
p(E | θp , I) · P (θp | I)
P (θp | E, I) = (2.2)
p(E | I)
p(E | θd , I) · P (θd | I)
P (θd | E, I) = (2.3)
p(E | I)
where I is the background information available in the case not related to the evidence E. This
I may include not only circumstantial information in the case (such as witness testimony), but
also the analysis of other forensic evidence apart from E (such as glass fragments, paint flakes,
etc.). Using equations 2.2 and 2.3 we then obtain:
P (θp | E, I) P (θp | I)
= LR · (2.4)
P (θd | E, I) P (θd | I)
where:
p(e | θp , I)
LR = (2.5)
p(e | θd , I) e=E
Equation 2.4 is the so-called odds form of Bayes’ theorem which, in words, means that the
posterior odds an be obtained by multiplying the prior odds by the LR (see Fig. 2.15). In this
framework, we can distinguish two terms:
The prior probabilities P (θp | I) = 1 − P (θd | I), which are province of the fact finder and
should be stated assuming only the background information (I) in the case [Cook et al.,
1998].
The LR (Equation 2.51 ), computed by the forensic scientist [Aitken et al., 2004].
The LR value (Equation 2.5) is the quotient of two probability densities. On the one hand,
the probability density function (pdf) p(e | θp , I) in the numerator in Equation 2.5 is known as
the within-source distribution, and models the variability of the scores computed with materials
coming from different acquisitions of the same source. Its evaluation in e = E gives a measure of
the similarity between the recovered and control materials. On the other hand, the pdf p(e | θd , I)
in the denominator is known as the between-source distribution, and models the variability of
the scores computed from different sources in a population and the trace. Both values in the
numerator and denominator, should be computed in a transparent way by the forensic scientist.
It is also the duty of the forensic scientist, following the background information of the case (I),
to select the population of individuals which will be proper for the case at hand.
The so-called LR approach has been applied to several biometric modalities including forensic-
DNA comparison [Buckleton, 2005], forensic speaker comparison [Champod and Meuwly, 2000;
1
Unless explicitly stated, we will use a capital E for referring to the given value of the evidence, according
to the literature on LR−based analysis of the evidence [Aitken et al., 2004; Champod and Meuwly, 2000]. Thus,
the small e will be used as the argument in likelihoods.
28
2.4 Chapter Summary and Conclusions
Gonzalez-Rodriguez et al., 2007; Meuwly and Drygajlo, 2001], forensic fingerprint comparison
[Neumann et al., 2012] and general biometric systems [Gonzalez-Rodriguez et al., 2005]. As
detailed in [Gonzalez-Rodriguez et al., 2005; Meuwly and Drygajlo, 2001], the LR values are
computed by first modeling probability density functions (pdfs) of the Within-Source Variabil-
ity (WSV) and the Between-Source Variability (BSV) scores and then finding the ratio of these
pdfs at a given score value. In [Gonzalez-Rodriguez et al., 2007], the logistic regression approach
is used to estimate the natural logarithm of the ratio of these densities instead of estimating the
pdfs of the WSV and the BSV separately.
29
2. RELATED WORKS
30
Chapter 3
Experimental Framework
This chapter describes the experimental framework used in this Thesis. The common prac-
tices in performance evaluation of biometric systems and evidence evaluation of forensic bio-
metrics are introduced, including experimental protocols and measures. The palmprint and
fingerprint databases used in the experimental chapters are also described, including databases
with high-resolution full images and latent images.
This chapter is organized as follows. First we introduce the guidelines of performance evalu-
ation used in this Dissertation (Sect. 3.1). Then we describe the common practices of likelihood
ratio based evidence evaluation (Sect. 3.2). The next section is dedicated to a summary of
palmprint and fingerprint databases used in this Dissertation (Sect. 3.3). Finally, the summary
and conclusions of this chapter are presented (see Sect. 3.4).
31
3. EXPERIMENTAL FRAMEWORK
In the context described above, and as a result of the experience gained in biometric perfor-
mance evaluation, the UK Biometrics Working Group has generated a set of best practices for
testing and reporting performance results of biometric systems [Mansfield and Wayman, 2002],
to which we adhere in this Thesis.
3.1.1. Protocols
Performance evaluations of biometric technology are divided into three overlapping categories
with increasing complexity of the uncontrolled variables considered: technology, scenario, and
operational [Phillips et al., 2000a]. A thorough evaluation of a system for a specific purpose starts
with a Technology Evaluation, followed by a Scenario Evaluation, and finally an Operational
Evaluation.
The primary goal of Technology Evaluations is to measure the performance of biometric
systems, typically only the recognition algorithm component. They are repeatable and usually
short in duration. Technology Evaluations are usually performed using standard datasets col-
lected previous to testing. Results from a Technology Evaluation generally reveal specific areas
that require future research and development (R&D) and provide performance data that is use-
ful when selecting algorithms for Scenario Evaluations (see below). Examples of Technology
Evaluations are FVC, NIST ELFT-EFS and PVC.
The primary aim of Scenario Evaluations is to measure the performance of a biometric system
operating in a particular application. For example, testing biometrics for access control purposes
at a mock doorway in a laboratory. Each tested system normally would have its own acquisition
sensor and would thus receive and produce slightly different data. For this and other reasons,
Scenario Evaluations are not always completely repeatable. Scenario Evaluations usually take a
few weeks to complete because multiple trials must be performed to ensure adequate habituation
of the end users (if the scenario calls for it) and to achieve a statistically relevant number of
samples. Results from a typical Scenario Evaluation show areas that require additional system
integration and provide performance data on systems for the application tested. An example of
a Scenario Evaluation is the UK Biometric Product Testing [Mansfield et al., 2001].
At first glance, an Operational Evaluation appears very similar to a Scenario Evaluation,
except that the test is conducted at the actual site using actual end users, a subset of the end
users, or a representative set of subjects. Rather than testing for performance (which is difficult
to do in some operational evaluations), Operational Evaluations typically aim to determine the
workflow impact caused by the addition of a biometric system [Bone and Crumbacker, 2001].
Operational Evaluations are typically not repeatable. Also, they can last from several weeks to
several months because the evaluation team must first examine workflow performance prior to
the use of the technology and again after users are familiar with the technology. An accurate
analysis of the benefit of the new technology requires a comparison of the workflow performance
before and after use of the technology.
In an ideal three-step evaluation process, Technology Evaluations are first performed on all
technologies that could conceivably meet application requirements. The technical community
32
3.1 Performance Evaluation of Biometric Systems
then uses the results to plan future R&D activities, while potential users use the results to
select promising systems for application-specific Scenario Evaluations. Results from the Scenario
Evaluations will enable users to determine the best system for their specific application and to
have a good understanding of how it will operate at the proposed location. These outcomes,
combined with workflow impact results from subsequent Operational Evaluations, will enable
decision makers to develop a solid business case for potential installations.
So for those analyzing evaluation reports, it is important to determine which type of eval-
uation occurred and its relevance to an intended application. Generally, technology evaluation
reports contain information relevant to most intended applications of a given biometric technol-
ogy, while operational evaluation reports are generally only useful if the intended application
is very closely related to what was tested. It is also important to be clear about biometric
evaluation terms such as recognition, verification and identification, which are usually referred
to as “operation modes”, described as follows:
Verification, also known as authentication or detection, occurs when the biometric system
attempts to confirm an individual’s claimed identity by comparing a submitted sample to
one or more previously enrolled templates.
Identification occurs when the biometric system attempts to determine the identity of
an individual. A biometric whose ID is unknown is collected and compared to all the
templates in a database where the IDs are known. Identification is “closed-set” if the
person is assumed to exist in the database. In “open-set” identification, the person is not
guaranteed to exist in the database, and the system must determine if the person is in the
database, and if so, which of the templates correspond to him.
Recognition is a generic term and does not necessarily imply either verification or identi-
fication. All biometric systems perform “recognition” of a person who has been previously
enrolled, either to confirm or determine his identity.
In this Thesis we carry out the performance evaluation experiments of biometric systems
as technology evaluations. We assume two operation modes for the systems proposed, either
close-set identification mode or verification mode.
3.1.2. Measures
Unlike password-based systems, where a perfect match between two alphanumeric strings
is necessary in order to validate a user’s identity, a biometric system seldom encounters two
samples of a user’s biometric trait that result in exactly the same feature set. This is due to
imperfect sensing conditions (e.g., noisy fingerprint due to sensor malfunction), alterations in the
user’s biometric characteristic (e.g., intrinsic session variability impacting speaker recognition),
changes in ambient conditions (e.g., inconsistent illumination levels in face recognition) and
variations in the user’s interaction with the sensor (e.g., occluded iris or partial fingerprints).
33
3. EXPERIMENTAL FRAMEWORK
Thus, two feature sets originating from the same biometric trait of a user rarely look exactly
the same.
The degree of similarity between two biometric feature sets is indicated by a similarity score.
The action of computing a score from two biometric samples is referred to as a trial, or a
biometric comparison. A score that is computed from two samples of a biometric trait of the
same source is known as a genuine, target, match or authentic score. It is known as an impostor
(or non-target) score if it involves comparing two biometric samples originating from different
sources. Here, source may refer to an individual if the biometric sample can be associated to an
individual (e.g., DNA). Sometimes, an individual may generate different biometric samples, like
in fingerprints, in which case the source is assumed to be the finger, originating the fingerprint
samples. An impostor score that exceeds a given decision threshold η results in a false accept
(or, a false match), while a genuine score that falls below the threshold η results in a false reject
(or, a false non-match). The False Accept Rate (FAR) (or, the False Match Rate (FMR), and
generally the False Positive Rate (FPR) ) of a biometric system can therefore be defined as the
fraction of impostor scores exceeding the threshold η. Similarly, the False Reject Rate (FRR)
(or, the False Non- match Rate (FNMR), and generally the False Negative Rate (FNR)) of
a system may be defined as the fraction of genuine scores falling below the threshold η. The
Genuine Accept Rate (GAR) (or generally the True Positive Rate (TPR)) is the fraction of
genuine scores exceeding the threshold η. Therefore, GAR = 1 − F RR. Regulating the value of
η changes the FRR and the FAR values, but for a given set of scores from a biometric system,
it is not possible to decrease both these errors simultaneously.
The FAR and FRR at various values of η can be summarized using a Receiver Operating
Characteristic (ROC) curve [Egan, 1975] that plots the FRR against the FAR at various thresh-
olds on a linear, logarithmic or semi-logarithmic scale (see Fig. 3.1 (a)). In many instances, the
ROC curve plots the GAR (rather than the FRR) against the FAR (see Fig. 3.1 (b)). When
a normal deviate scale is used to plot these error rates, then the resulting graph is known as
a Detection Error Tradeoff (DET) curve [Martin et al., 1997] (see Fig. 3.1 (c)). The primary
difference between DET and ROC curves is the use of the normal deviate scale in the former. In
DET curves, normal distribution of scores appear as straight lines. In ROC/DET curves with
FRR against FAR, the closer to the origin, the better. In ROC curves with GAR against FAR,
the closer to the up-left corner, the better. Both ROC and DET curves measure discriminating
power, and they are conceptually the same performance measure.
The performance of a biometric system may also be summarized using other single-valued
measures such as the Equal Error Rate (EER). The EER refers to the point in a ROC/DET curve
where the FAR equals the FRR; a lower EER value, therefore, indicates better performance.
In the case of identification, the input biometric sample of unknown source is compared
against many or all templates of a database in order to determine the template with higher
similarity to the input (known as the “best match”). The best match can be determined by
examining the match scores pertaining to all the comparisons and reporting the identity of the
template corresponding to the largest similarity score. The identification rate indicates the
34
3.1 Performance Evaluation of Biometric Systems
ROC ROC
100 100
90 90
60 60
50 50
40 40
30 30
20 20
10 10
0 0
0 20 40 60 80 100 0 20 40 60 80 100
False Accept Rate (%) False Accept Rate (%)
(a) (b)
DET Cumulative Match Characteristic
1
40
0.9
False Reject Rate (%)
0.8
Identification Rate
20
0.7
10 0.6
5 0.5
0.4
2
1 0.3
0.5 0.2
0.2 0.1
0.1
0 2 4 6 8 10 12 14 16 18 20
0.1 0.2 0.5 1 2 5 10 20 40
False Accept Rate (%) Rank
(c) (d)
Figure 3.1: As an example, the performance of a biometric system can be summarized using ROC, DET
and CMC curves. (a) ROC curve with FRR against FAR, (b) ROC curve with GAR against FAR, (c)
DET curve, (d) CMC curve.
proportion of times where a previously enrolled individual is successfully mapped to the correct
identity in the system. Here, we assume that the question being asked is: “in which proportion
of cases does the top match correspond to the correct identity?”. An alternate question could be:
“in which proportion does any one of the top k matches correspond to the correct identity?” (see
[Moon and Phillips, 2001]). The rank-k identification rate, Rk , indicates the proportion of times
where the correct identity is included in the top k matches as determined by the match score.
Rank-k performance can be visualized using the Cumulative Match Characteristic (CMC) curve
[Moon and Phillips, 2001] that plots Rk against k, for k = 1, 2, ..., M with M being a maximum
value, which varies among representations, and which can even be the number of enrolled users
(see Fig. 3.1 (d)). The relationship between CMC and DET/ROC curves has been discussed
by [Grother and Phillips, 2004] and [Bolle et al., 2005].
In this Thesis, we use ROC curves to show discriminating power and as a measure of perfor-
mance in verification mode. On the other hand, we use CMC curves for graphical illustration
in close-set identification tasks. We also use EER and rank-k identification rate for single-value
summarizing measures.
35
3. EXPERIMENTAL FRAMEWORK
Output
Input to (next level
Abstraction levels
each level input)
Court Forensic Forensic
Background information
adequate testing
about the case
Results from
Speech Presentation LR
databases Level value
Recovered
Crime Scene Discrimination Similarity
Control
Suspect
Level score
Figure 3.2: Abstraction levels in evidence evaluation using biometric systems (forensic speaker recogni-
tion system as an example).[Ramos, 2007]
3.2.1. Protocols
According to [Ramos, 2007], there are three abstraction levels in evidence evaluation using
biometric systems, namely, discrimination level, presentation level and forensic level (see Fig.
3.2). The discrimination level is the aim of the biometric systems, and its performance is
measured typically by ROC and DET curves and other measures as described in Sect. 3.1.2.
The presentation level yields LR computation, and its performance is measured by Cllr , ECE
and Tippett plots which will be described in the next section. A third level, the forensic level,
require the consideration of many other factors, and it is not within the scope of this Thesis
(details in [Ramos, 2007]).
In this Thesis, we will study the presentation level by computing likelihood ratios. At the
presentation level, the protocols of LR-based evidence evaluation include two aspects (see Fig.
3.3):
36
3.2 Performance of Likelihood-Ratio-Based Interpretation Methods
Figure 3.3: The presentation levels in evidence evaluation using biometric systems (forensic speaker
recognition system as an example).[Ramos, 2007]
3.2.2. Measures
In this Thesis, following the recommendations in [Brümmer and du Preez, 2006; Gonzalez-
Rodriguez et al., 2007; Ramos and Gonzalez-Rodriguez, 2013; Ramos et al., 2013], we use Tippett
and ECE plots for graphical illustration of the measures of LR accuracy, and Cllr for single-value
assessment of LR accuracy.
Tippett plots have been classically used for empirical performance assessment [Champod and
Meuwly, 2000; Dessimoz and Champod, 2007; Evett and Buckleton, 1996; Gonzalez-Rodriguez
et al., 2007]. Tippett plots are obtained in the same way as the probabilities of false acceptance
and genuine acceptance depending on a threshold η, namely F AR and GAR. However, these
values in Tippett plots are not interpreted as decision error rates, but as the proportion of
values for which log10 (LR) > η in the given LR set, conditioned to θp and θd respectively. The
relationship between FAR/FRR plots and Tippett plots can be seen in Figure 3.4. Important
values shown by these curves are the empirical cumulative distributions of the LR values in a
37
3. EXPERIMENTAL FRAMEWORK
100 1
θp Misleading
0.7
Evidence
60 0.6
40 θd
Misleading 0.4
Evidence
0.3
20
0.2
0.1
0 −4 −2 0 2 4
10 10 10 10 10 0
−4 −3 −2 −1 0 1 2 3 4
Figure 3.4: An example of Tippett plots showing the actual LR distributions and the rates of misleading
evidence when θp and θd are respectively true. (a)The Tippett plots, (b) their corresponding FAR/FRR
plots.[Ramos, 2007]
given set, and the rates of misleading evidence, defined as the proportion of LR values giving
support to the wrong hypotheses (LR > 1 when θd is true and LR < 1 when θp is true) [Dessimoz
and Champod, 2007]. Note that the rates of misleading evidence are the FAR and FRR when
the threshold η is set at log10 (LR) = 0.
The overall measure of the accuracy of a probabilistic set of opinions of a forecaster is defined
as the average value of a strictly proper scoring rule over many different forecasts. For instance,
38
3.2 Performance of Likelihood-Ratio-Based Interpretation Methods
for the logarithmic scoring rule, this mean value would be the logarithmic score (LS):
1 X 1 X
LS = − log2 P (θp | ej ) − log2 P (θd | ej ) (3.2)
Np Nd
i∈targets j∈non−targets
where Np and Nd is the number of target (or genuine) and non-target (or impostor) trials
being evaluated. This average value can be viewed as an overall loss. Moreover, it is also
demonstrated in [DeGroot and Fienberg, 1983] that such a measure of accuracy can be divided
into two components:
A refinement loss component, which measures how sharp or how spread the forecasts are.
Roughly speaking, low refinement loss means that if the calibration loss of the forecaster
is low, for a given value of the forecast P (θp | E) the frequency of trials where θ = θp is
near 0 or 1.
P (θp | I) X P (θd | I) X
ECE = − log2 P (θp | Ei , I) − log2 P (θd | Ej , I) (3.3)
Np Nd
i∈targets j∈non−targets
where Ei and Ej denote the evidence in each of the comparisons (cases) in the validation set
where θp or θd are respectively true. Here a validation set of LR values will be computed from a
so- called validation database which contains the corresponding ground-truth labels indicating
the source that originated each LR value.
The value of ECE considers the discrimination performance of the LR set being assessed,
but also its calibration loss [DeGroot and Fienberg, 1983]. This is a main motivation for defining
the accuracy of a LR values set in terms of ECE. An example of ECE plots is shown in Fig.
3.5, which contains three comparative performance curves together:
1. Solid curve is the ECE of the LR values in the validation set, as a function of the prior
log-odds. The lower this curve, the better the accuracy.
2. Dashed curve is the ECE of the validation set of LR values after being transformed
using the Pool Adjacent Violators algorith (PAV) algorithm [Brummer, 2010; Brümmer
and du Preez, 2006; van Leeuwen and Brümmer, 2007], which shows the performance of
a validation set of optimally-calibrated LR values. The lower the curve, the better the
discriminating power. Moreover, the closer the curve to the solid one, the better the
calibration.
39
3. EXPERIMENTAL FRAMEWORK
1
LR values
After PAV
0.8 LR=1 always
0.4
0.2
0
-2 -1 0 1 2
Prior log (odds)
10
3.2.2.3. Cllr
In [Brümmer and du Preez, 2006] a measure of accuracy of a validation set of LR values set
is proposed, namely Cllr :
1 X 1 1 X
Cllr = log2 1 + + log2 (1 + LRj ) (3.4)
2 · Np LRi 2 · Nd
i∈targets j∈non−targets
where Np and Nd are respectively the number of target and non-target LR values in the valida-
tion set.
In Cllr it is easily seen that, for each LRi value in a given trial, a logarithmic transformation
is applied, defined by:
1
θp true : log2 1 + LRi
(3.5)
θd true : log2 (1 + LRi )
It can be seen that Cllr is based on the logarithmic scoring rule (Equation 3.1) in the partic-
ular case where P (θp ) = P (θd ) = 0.5. In fact, ECE is an information-theoretical generalization
of Cllr , and its value at log10 (LR) = 0 is the value of Cllr . Consequently, the smaller of Cllr , the
better.
40
3.3 Databases
3.3. Databases
Since databases containing latent marks and mated full palmprints are not publicly available,
in our experiments, 22 latent palmprints from real forensic cases and 8680 full palmprints from
criminal investigations captured by Beijing Institute of Criminal Technology in China, are used
to evaluate the performance of the proposed forensic automatic palmprint identification system.
The 8680 full prints are from 4340 subjects with the two palmprints from left hand and right
hand respectively, including mated images to the 22 latent prints. Image size of full palmprints
is 2304 × 2304 pixels, while the size of latent marks varies and generally is smaller than for
the full prints. All the palmprints are recovered (for the latent marks) or captured (for the full
prints) with the resolution of 500 ppi. Samples of palm marks and full palmprints from forensic
palmprint databases are shown in Fig. 3.6. The main characteristics of forensic palmprint
databases used in this Thesis are summarized in Table 3.1.
Recently, the first public high-resolution palmprint database THUPALMLAB [Tsinghua Univ.,
2011b] was released by Tsinghua University. The database contains 1,280 palmprints from 80
subjects (left and right palms of each subject, 8 impressions per palm). These images were
captured using a commercial palmprint scanner from HisignT M . The image size is 2040 × 2040
pixels, with a resolution of 500 ppi and 256 gray levels. In order to mimic forensic scenarios
using this public high-resolution palmprint database, we also obtain the testing database PV-
TEST collected by the same laboratory [Tsinghua Univ., 2011a] (i.e., PV-TEST-PARTIAL and
PV-TEST-FULL) used in Palmprint Verification Competition (PVC), which is not included in
THUPALMLAB database. The images in PV-TEST are collected from 5 subjects. For each
subject, 8 full palmprints and 4 partial palmprints of each palm are captured using a CogentT M
scanner with the resolution of 500 ppi. In total, the database contains 120 palmprint images
from 10 different palms, including 40 partial prints and 80 full prints. The image size of all
full palmprints is the same. i.e., 2500 × 2552 pixels). Partial palmprints have been simulated,
by automatically cropping them to reduce the image size, and therefore they present a variable
size. Finally, we compose the two databases THUPALMLAB and PV-TEST into the public
high-resolution palmprint databases including 40 partial palmprints and 1360 full palmprints
which are used for comparison evaluation. Samples of palmprints from public databases are
shown in Fig. 3.7. A summary of the main characteristics of public high-resolution palmprint
databases used in this Thesis is given in Table 3.1.
41
3. EXPERIMENTAL FRAMEWORK
Figure 3.6: Samples of palm marks and their mated full prints from forensic palmprint databases.
Table 3.1: Summary of palmprint databases used in this Thesis. All images present 500 ppi.
42
3.3 Databases
(a) (d)
(b) (e)
(c) (f)
Figure 3.7: Samples of palmprints from public databases.(a)-(c) full palmprints from THUPALMLAB
database where (a) and (b) are from the same palm and (c) is from a different palm; (d)-(f ) partial and
full palmprints from PV-TEST database where (d) and (e) are from the same palm and (f ) is from a
different palm.
43
3. EXPERIMENTAL FRAMEWORK
In this Thesis, some of the contributions are adapted for both forensic fingerprint and palm-
print recognition. Therefore, as some experiments have been performed with fingerprints in
Chapter 5, it is necessary that we describe the fingerprint databases used in this Thesis. We use
two forensic fingerprint databases as described below.
NFI DB58 is developed by the Netherlands Forensic Institute (NFI) in a similar way
as NIST SD27. This database is not publicly available. Overall, the database consists
of 58 cases. Each case includes the latent image, the matching fingerprint image from a
ten-print card, and four sets of minutiae that have been validated by 3 professional latent
examiners in the same way as described for the four minutiae sets of NIST SD27. Each
image is 620 × 620 pixels in size and has been scanned at 500 ppi, quantized to 256 levels
of gray, and stored in an uncompressed format. While the minutiae number in each latent
image in NIST SD27 varies, the number of minutiae annotated in each latent image in NFI
DB58 is fixed to 12. Moreover, starting from the 12-configuration minutiae set, minutiae
subsets with configurations of a minutiae number from 5 to 11 are also annotated.
44
3.4 Chapter Summary and Conclusions
evaluation. Finally we have given an overview of existing palmprint and fingerprint databases
involved in the experiments in this Dissertation.
45
3. EXPERIMENTAL FRAMEWORK
46
Chapter 4
This chapter presents the methods and experimental evaluation of a novel forensic palmprint
recognition system based on radial triangulation. The proposed radial triangulation representa-
tion for minutiae in palmprints is introduced. Then a complete forensic palmprint recognition
system based on radial triangulation is described, including novel forensic palmprint feature ex-
traction and comparison. Remarkably, this system outputs Log-likelihood ratios, and therefore
it can be used in a framework of evidence interpretation. Experiments and evaluations of system
performance and palmprint evidence are also concluded.
The main contributions of the work presented in this chapter are as follows:
1) Palmprint feature extraction with image preprocessing and postprocessing. In this pro-
cess, reliable minutiae and principal lines are obtained. As both kinds of features are used for
comparison, i.e., minutiae for local comparison and principal lines for global comparison, the
name M inutiaLine is given to our extractor.
2) Local feature comparison algorithm based on radial triangulation (RT ) structures, ac-
cording to point pattern comparison by relaxation. This is a novel comparison algorithm robust
to distortion in order to compute the matching score.
3) A combined global feature comparison algorithm is proposed, in which weights of centroids
of radial triangulation structures and principal line energy are both considered for finer global
comparison.
4) By using extended features such as principal lines which are unique in palmprints compared
to fingerprints, our algorithm can be fused with other state-of-the-art algorithms [Cappelli et al.,
2012; Liu et al., 2013] on high-resolution palmprint recognition at feature level or score level
in order to further improve the performance of forensic palmprint recognition. This is because
the fusion of different algorithms is essential especially in forensic scenarios while evidence from
results/scores obtained by automated systems should be evaluated comprehensively. Moreover,
due to the reduced computational complexity of our proposal, the fusion will not imply a high
47
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
increase in the computational burden, and therefore it will be suitable for large database searches.
5) Our algorithm is based on a sound quantitative approach to fingerprint comparison by
human experts, namely radial triangulation [Neumann et al., 2007]. This kind of approach is
starting to be used in court in opposition to procedures based on human comparison such as
the ACE-V procedure [Giannelli and Imwinkelried, 1993]. Therefore, our algorithm for forensic
palmprint recognition could be easily integrated into this framework to compute the weight of
the forensic evidence based on radial triangulation.
This chapter is organized as follows. First we introduce the proposed radial triangulation
representation for minutiae in palmprints (Sect. 4.1). Then we describe our forensic palm-
print recognition system based on radial triangulation (Sect. 4.2). The next section presents
experimental results on the evaluation of system performance (Sect. 4.3). A followed section
is dedicated to the results on palmprint evidence evaluation using the proposed forensic palm-
print recognition system (Sect. 4.4). Finally, the summary and conclusions of this chapter are
presented (see Sect. 4.5).
This chapter is based on the publications: [Wang et al., 2011, 2012a, 2013b].
In this section, we will introduce the basic concepts of radial triangulation and point pattern
comparison by relaxation, in order to understand the rest of this Chapter.
In order to execute minutiae-based comparison between latent marks and full palmprints,
the spatial arrangement of minutiae in local minutiae structures is essential not only for local
minutiae comparison but also for global minutiae comparison, as latent marks are usually partial
images that need to be aligned with full prints.
Recently, radial triangulation (RT) modelling has been applied to palmprint minutiae mod-
eling [Wang et al., 2011, 2012a], motivated by its application to minutiae modeling for forensic
evidence evaluation in fingerprint comparison [Neumann et al., 2007, 2012]. An example of a
local minutiae structure based on radial triangulation is shown in Fig. 4.1. Given a set of N
minutiae, i.e., M = {mk }Nk=1 where mk = (xk , yk , θk ), the general vector form of a local minutiae
structure based on radial triangulation is,
where PC = (xC , yC , θC ) is the centroid of the polygon, Vk is a vertex where minutia mk locates
in a polar coordinate system centered in its corresponding centroid as represented in Equ. 4.2,
Rk is the distance between minutia mk and the centroid, Lk,k+1 is the length of the polygon
side between minutia mk and minutia mk+1 , STk is the area of the triangle defined by minutiae
48
4.1 Radial Triangulation for Palmprints
where the centroid PC = (xC , yC , θC ) is the pole, rk denotes the polar radius of Vk relative to
PC , αk denotes the polar angle of Vk relative to PC , and ϕk denotes the angle of Vk relative
to PC according to their directions. Generated in this way, the minutiae structure vector LS
consists of the radial triangulation vector RT = {Vk , Rk , Lk,k+1 , STk }N
k=1 which is used for local
minutiae comparison, and the centroid PC which is used for global minutiae comparison.
Vk+1 PC
STk
Lk,k+1
Rk
Vk
49
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
in LM relative to a minutia in F M , how many pairs (LMi , F Mj ) lie closer together than a
threshold t. Here, δ is measured as the Euclidean distance between two minutiae points, and
the threshold t must be substantially smaller than the average interpoint distance in LM and
F M , since otherwise many false matches will be detected. Let LMh and F Mk be any minutiae
other than LMi and F Mj , and let δi,j (h, k) be the distance between LMh and F Mk when LMi is
mapped into F Mj . If |δi,j (h, k)| is zero, F Mk is exactly in the same position relative to F Mj as
LMh is relative to LMi , so that the pair (LMh , F Mk ) should yield maximal support to the pair
(LMi , F Mj ); while as |δi,j (h, k)| increases, this support should decrease. Let the support given
to (LMi , F Mj ) by (LMh , F Mk ) be denoted by ϕ(|δi,j (h, k)|). In the experiments described in
this work, we used ϕ(x) = 1/(1 + x2 ) as in [Ranade and Rosenfeld, 1980]. For each LMh there
may be several F Mk that lie close to it; but we only want one F Mk to correspond to LMh when
LMi is paired with F Mj . Thus it is reasonable to define the support to (LMi , F Mj ) associated
with LMh as
max[ϕ(|δi,j (h, k)|)]. (4.3)
k6=j
In this way, we compute the support provided to the pair (LMi , F Mj ) by each LMh . To
obtain the total support for the pair (LMi , F Mj ), we average contributions of all the LMh :
P
{max[ϕ(|δi,j (h, k)|)]}
h6=i k6=j
sM (LMi , F Mj ) = , (4.4)
m−1
where sM (LMi , F Mj ) denotes the similarity of minutiae pair (LMi , F Mj ).
50
4.2 Forensic Palmprint Recognition System Based on Radial Triangulation
Examination by
human experts Latent minutiae set
Initial minutiae
Postprocessing Full minutiae set
extraction
Preprocessing
Principal line energy
Full principal line energy
extraction
Full palmprint
full palmprints can be extracted in an automatic way properly. However, original full prints
usually contain blank regions, knuckle-finger print regions, unrecoverable low quality regions
and creases, which result in many spurious minutiae. As a consequence, image preprocessing,
i.e., segmentation and enhancement, and post-processing for full palmprints are essential for
extracting reliable minutiae in order to improve identification rate and computation efficiency.
Also, in this work, besides minutiae extraction, we automatically extract principal line energy as
an additional feature both in latent marks and full palmprints, and we use it as a global feature
in global comparison. Fig. 4.2 generally illustrates the forensic palmprint feature extraction
procedure. Notice that some of the techniques used in the process of palmprint preprocessing
and minutiae feature extraction are common, and therefore we only give a brief description of
those techniques. More details can be found in the related references.
For full palmprint segmentation, a block-based method proposed in [Zheng et al., 2007a] is
used. An original palmprint image is divided into four types of distinctive areas: blank regions,
knuckle-finger regions, unrecoverable low quality palm regions and high quality palm regions.
The first three kinds of areas are identified respectively as follows: 1) Blank regions are separated
according to the block white pixel proportion and the block variance. 2) Knuckle-finger regions
are separated from palm regions, which are marked as Main Connected Components (MCC).
3) Unrecoverable low quality palm regions are separated from high quality palm regions by a
feed-forward neural network classifier using orientation and ridge information. Fig. 4.3(b) shows
an example of this segmentation procedure applied to the palmprint in Fig. 4.3(a).
In order to obtain high-quality minutiae in a reliable way, palmprint enhancement is im-
plemented using the method described in [Zheng et al., 2007b], that consists of four steps: 1)
51
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
(a)
(b)
(c)
Figure 4.3: Full palmprint preprocessing. (a) Original image. (b) Segmented image and a sample of
orientation field. (c) Thinned image.
52
4.2 Forensic Palmprint Recognition System Based on Radial Triangulation
Estimate and modify the orientation field as shown in Fig. 4.3(b). 2) Remove noises in the
grey-scale image. 3) Convert the grey-scale image into a binary image. 4) Remove noises in the
binary image. The final thinned image is shown in Fig. 4.3(c).
We extract minutiae features in full palmprints and construct the initial feature set using
the feature extraction approach in [Li et al., 2009], which consists of two stages. The first
stage is based on the thinned image and obtains the position of the minutia. Given a thinned
image, we use the Rutovitz Crossing Number (CN) [Rutovitz, 1966] to detect the coordinates of a
minutia. Then the second stage obtains the direction of the minutia according to the orientation
information of its associated local ridge. Finally we record the direction of the minutia together
with its position.
Given the rough minutiae set extracted in the above section, a feature post-processing ap-
proach is implemented based on statistical and structural information, combining the structural
relationship in the minutiae subsets, and local ridges following [Li and Shi, 2008].
In order to improve the accuracy and reliability of latent-to-full palmprint comparison, we
use a minutiae filter based on line feature extraction [Yang et al., 2009] as a finer post-processing
53
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
(a) (b)
(c) (d)
Figure 4.5: Principal line extraction. (a) A segmented palmprint, (b) crease energy image, (c) crease
direction image, (d) extracted principal lines.
step to remove spurious minutiae. Minutiae extraction after post-processing is shown in Fig. 4.4.
Like the ridges in palmprints, the principal lines (i.e., the three major creases in a full
palmprint, namely, heart line, head line and life line, see Fig. 1.1 (b)) are firmly attached
to the dermis, and are immutable for the whole life [Ashbaugh, 1999]. This fact asserts the
discriminative power of principal lines, which has already been exhaustively studied in the area
of low-resolution palmprint recognition [Lin et al., 2005; Liu et al., 2007; Wu et al., 2006; Zhang
and Zhang, 2004]. In this work, we use an effective method of principal line extraction based
on Modified Finite Radon Transform (MFRAT) recently proposed [Huang et al., 2008]. Then,
the extracted principal line energy is considered as a kind of global feature which will be used
for palmprint global comparison.
54
4.2 Forensic Palmprint Recognition System Based on Radial Triangulation
Radial
Latent minutiae set
triangulation
Local minutiae comparison
Radial
Full minutiae set
triangulation
G lobal feature comparison
Minutiae
similarity
ranking
F eature 1
Global minutiae
Score
correspondence
Latent principal line energy
Feature 2
YES Weighting
Energy ≠ 0 ?
principal line energy
Full principal line energy
According to the concept that a crease line in a palmprint can be approximately regarded as
a straight line in a small local area, the MFRAT-based principal line extraction is carried out as
follows. Starting from a segmented palmprint image (Fig. 4.5(a)), we first calculate the energies
and directions of all pixels using a set of finite module grids as defined in [Huang et al., 2008]
and obtain the crease energy image (see Fig. 4.5(b)) and crease direction image (see Fig. 4.5(c)).
Then we use energy threshold and direction restrictions to remove wrinkles, and obtain the final
principal lines with energies and directions as shown in Fig. 4.5(d). Details can be found in
[Huang et al., 2008].
As latent marks are mostly partial and much smaller than full prints, position and direction
information of a minutia cannot directly be used for latent-to-full palmprint comparison without
prior known or estimated registration points in a pair of images. In order to find optimal
registration points for global minutiae correspondence, local feature comparison followed by
global comparison is a frequently used technique in minutiae-based palmprint recognition [Jain
and Feng, 2009; Liu et al., 2013; Wang et al., 2011, 2012a]. In this section, we propose local
minutiae comparison based on radial triangulation [Wang et al., 2011, 2012a], also using point
55
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
pattern comparison by relaxation for the similarity computation, which will contribute to global
minutiae correspondence. The proposed process is described below.
Step 1: Minutiae are sorted in ascending order, according to their distance from the
origin of the coordinate system.
Step 2: The minutia in the sorted set which is closest to the origin is selected as the
starting point. Then, the N − 1 minutiae closest to the starting minutia are selected
to form a local N -minutiae set. After that, the total N minutiae are deleted from
the sorted minutiae set. In our following experiments, we set N = 9 according to the
experiments performed in [Wang et al., 2011].
Step 3: We repeat steps 1 and 2 on the reduced minutiae set until the number of
remaining minutiae is less than N .
Step 1: For all pairs of radial triangulation structures RTi and RTj , their distance
vector Dij is calculated by Eq. (4.5) after their centroids are aligned. The two struc-
tures are then considered as a matched pair when the distance vector is within the
threshold D0 = {DV0 , DR0 , DL0 , DST0 }.
N N N N
1 XX 1 XX
Dij = { DVik ,jl , DRik ,jl ,
N2 N2
l=1 k=1 l=1 k=1
(4.5)
N N N N
1 XX 1 XX
DLi ,j , DSTik ,jl },
N2 k l
N2
l=1 k=1 l=1 k=1
where
q
DVik ,jl = ∆x2ik ,jl + ∆yi2k ,jl ,
56
4.2 Forensic Palmprint Recognition System Based on Radial Triangulation
LM i k FMjl
PC i PC j
RT i RT j
Dij
i =i + 1 j =j +1
Dij < D0 ?
NO
YES
k =k + 1 l =l + 1
sM (LMi k , FM j l )
s M ( LM i k , FM j l )
max[{s M ( LM ik , FM jl )}lN=1 ] NO
YES
{s M ( LM ik , FM jl )}iL=,1N, k =1
Figure 4.7: Comparison of radial triangulation structures from a latent mark and a full palmprint.
57
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
∆xik ,jl = rik cos αik − rjl cos αjl , ∆yik ,jl = rik sin αik − rjl sin αjl .
Step 2: Then the similarity sM (LMik , F Mjl ) of all minutiae pairs LMik from the
latent print and F Mjl from the full print in the pair of matched structures (RTi , RTj )
is calculated using Eq. (4.6), according to the similarity of point pattern comparison
by relaxation using Eq. (4.4).
P
{max[ϕ(|δik ,jl (p, q)|)]}
p6=k q6=l
sM (LMik , F Mjl ) = , (4.6)
N −1
where δik ,jl (p, q) and ϕ(x) are defined as in Section 4.1.2.
Step 3: A constraint is applied to choose paired minutiae: The minutia F Mjl is paired
with the minutia LMik only when the similarity sM (LMik , F Mjl ) is the maximum
value in the similarity set of minutia LMik , i.e., max{sM (LMik , F Mjl )}N
l=1 . Finally,
all paired minutiae are marked as candidates.
We propose a combined scheme for global feature comparison based on the distribution of
RT centroids and principal line energy as shown in Fig. 4.6. Two parallel procedures are carried
out to weight the two kinds of global features, and then merged into a novel weighting procedure
combining both of the global features.
Step 1: Local L-centroid sets for the full palmprint are selected using the same pro-
cedure as in Section 4.2.2.1, while only one L-centroid is set for the latent palmprint.
Note L is the number of centroids in the latent palmprint.
Step 2: Simplified radial triangulation structure sets LSCF ull = {LSCf }Z f =1 are
extracted for the full palmprint based on their L-centroid sets in accordance with the
following simplified radial structure form,
58
4.2 Forensic Palmprint Recognition System Based on Radial Triangulation
RTC RTC f
Latent
RT comparison
Iterate:
f =f +1
Centroid similarity
L, F
{sC ( P C i
, P C j )}i =1, j =1
Weight of centroids
{ p ( P C i , P C j )}iL=,1F, j =1
Figure 4.8: Weighting based on centroid distribution for global minutiae correspondence.
where PCC = (xCC , yCC , θCC ) is the centroid of the polygon formed by a L-centroid
set. The number of simplified structures in a full print is Z = bF/Lc. However, only
one simplified radial triangulation structure set is generated for the latent palmprint,
i.e., LSCLatent = [PCC , {Vk , Rk }L
k=1 ]Latent .
Step 3: For each simplified radial triangulation structure set in the full palmprint,
the similarity sC (PCi , PCj ) of centroids PCi from the latent print and PCj from the
full print is calculated by applying Eq. (4.6) to the centroids. Therefore, point pat-
tern comparison with relaxation is also applied to the centroids. Then the support
p(PCi , PCj ) of the centroid pair (PCi , PCj ) given to the structure pair (LSCLatent , LSCf ),
f = 1, 2, . . . , Z, is calculated using Eq. (4.8), and will be assigned to the minutiae
59
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
similarity sM (LMik , F Mjl ) to obtain the modified minutia similarity s0 (LMik , F Mjl )
using Eq. (4.9).
sC (PCi , PCj )
p(PCi , PCj ) = L (4.8)
P
sC (PCi , PCj )
j=1
where the minutiae LMik and F Mjl are in the radial triangulation structures related
to the centroids PCi and PCj respectively, namely RTi and RTj . Notice that all
minutiae in a structure pair (RTi , RTj ) receive the same weight p(PCi , PCj ). In that
sense, the global centroid weight boosts the importance of minutiae that belong to
structures with good centroid matching, and vise-versa.
Step 4: All minutia pairs are sorted in decreasing order of the modified similarity
s0 (LMik , F Mjl ) defined in Step 3, and each of the top-ten minutia pairs is used to
correspond the two sets of minutiae.
where IEL (x, y) and IEF (x, y) denote the principal line energies of latent and full palmprints
at pixel (x, y) respectively, and IDL (x, y) and I F (x, y) denote the principal line directions.
D
Theoretically, SE is between 0 and 1, and the larger the value, the greater the similarity
between the latent energy image and the energy image of the compared region in a full
palmprint.
60
4.2 Forensic Palmprint Recognition System Based on Radial Triangulation
EL
Pixel-to-area
comparison EF
w
Matched energy w =w + 1
{ E m }W0
w w=1
Figure 4.9: Weighting based on principal line energy for global minutiae correspondence.
Since the number of minutiae in a full palmprint is more than 20 times bigger than in a
latent palmprint, structural global comparison based on centroid sets can be poor when
there are few minutiae in a latent mark. Therefore, we propose a two-layer global compar-
ison method combining weights of centroids and principal line energy. Firstly, weighting
global information based on centroid sets is performed between a latent mark and a full
palmprint to obtain all modified local minutia similarity s0 (LMik , F Mjl ). Then we weight
global information based on principal line energies between the latent mark and its corre-
sponding rectangle region with the same size in the full palmprint as shown in Fig. 4.9, in
which the center point is the centroid of the structure LSCf including the centroid PCj ,
and obtain the similarity SEf of principal line maps for that rectangle region. Finally we
61
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
Note that SEf equally weights all the minutiae in a structure pair (RTi , RTj ). Therefore,
principal line weighting does not affect the ordering of similarity of minutiae pairs, but the
value of the overall similarity of one pair of matched minutiae candidates.
After the above combined weighting, all minutia pairs are sorted in decreasing order of the
combined similarity sb (LMik , F Mjl ) defined in Eq. (4.12), and each of the top-ten minutia
pairs (i.e., the ten pairs of minutiae with highest sb (LMik , F Mjl )) is used to correspond
the two sets of minutiae. Note that only when a latent mark does not contain principal
lines, we use the modified similarity defined in Eq. (4.9) to obtain top-ten minutia pairs.
In other words, if there are no principal lines in the mark, then SEf = 1. The latter avoids
spurious scoring. Then, minutiae are examined in turn, while a minutiae pair that is the
closest with each other according to both location and direction, is deemed as a matched
minutiae pair. After all the minutiae pairs have been examined, a set of matched minutiae
with the largest number of matched pairs among the ten corresponding trials is obtained.
The minutiae-based comparison score S between two palmprints in this Chapter is set by a
linear expression (see Eq. (4.13)) using two individual scores Sn and Sq .
S = λ0 + λ1 × Sn + λ2 × Sq (4.13)
NM NM
Sq = SD × × (4.15)
NM + NL NM + NF
where SD is the average similarity of radial triangulation structures for all the matched minutiae,
and NL and NF denote the number of unmatched minutiae in latent marks and full palmprints
respectively.
62
4.3 Experiments and Performance Evaluation
In this Chapter, to evaluate the performance of the proposed forensic automatic palmprint
identification system, we use both forensic palmprint databases and public high-resolution palm-
print databases as detailed in Sect. 3.3.1. The forensic palmprint databases contain 22 latent
palmprints from real forensic cases and 8680 full palmprints from criminal investigations [Tan
et al., 2009] captured by Beijing Institute of Criminal Technology in China. And the public high-
resolution palmprint databases consists of the first public high-resolution palmprint database
THUPALMLAB [Tsinghua Univ., 2011b] and the testing database PV-TEST [Tsinghua Univ.,
2011a] (i.e., PV-TEST-PARTIAL and PV-TEST-FULL) used in Palmprint Verification Com-
petition (PVC), both captured by Tsinghua University. Table 3.1 summarizes the main charac-
teristics of the databases used in this Chapter.
The proposed palmprint comparison algorithm requires a protocol of learning parameters
based on logistic regression for score computation as described in Sect. 4.2.2.3. As stated in
[Long, 1997], it is risky to use maximum likelihood estimates for such a learning procedure in
training samples having less than 100 elements, while samples above 500 elements should be
adequate. In our following experiments, for the forensic database we choose a training set of
1800 score pairs (Sn , Sq ), generated by comparison between 6 latent palmprints corresponding
to one subject and 300 randomly selected full palmprints. On the other hand, 1600 score pairs
generated by comparison between 20 partial palmprints and 80 full palmprints are used for
training on the public high-resolution palmprint database. In both cases, we have followed a
leave-one-out training scheme due to the small amount of palm marks in forensic palmprint
databases and partial palmprints in public high-resolution palmprint database. Notice that the
use of leave-one-out training scheme may lead to results that are overoptimistic.
In this section we evaluate the proposed forensic palmprint identification system (see Fig. 4.2
and Fig. 4.6). We compare it with a baseline MinutiaCode-based system, implemented in a demo
software [THU, 2011] released by the authors of [Jain and Feng, 2009]. This demo software
contains two main modules, namely MinutiaCode extractor and MinutiaCode comparator.
In our experiments, we combined each extractor and each comparator from the proposed
system and from the MinutiaCode baseline, leading to four palmprint identification systems, i.e.,
M inutiaLine extractor+RT comparator, MinutiaCode extractor+RT comparator, MinutiaCode
extractor and comparator, and M inutiaLine extractor+MinutiaCode comparator.
In order to evaluate the performance of the two extractors, i.e., M inutiaLine and Minutia-
Code, we first use both of them to extract minutiae of the 8680 full palmprints in the forensic
63
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
Table 4.1: Statistical results of average minutiae number on the forensic database
database. Statistical results of average minutiae number and extraction efficiency on such a
full palmprint database are shown in Table 4.1. We can see that the computational efficiency
of M inutiaLine is around 8 times higher than that of MinutiaCode while the average num-
ber of minutiae extracted by M inutiaLine is similar to MinutiaCode. Moreover, M inutiaLine
extracts principal lines, not only minutiae.
To evaluate the performance in a closed-set identification task, we use the Cumulative Match
Characteristic (CMC) curve to measure the identification performance. CMC curves for the four
combined comparison systems on the forensic database are shown in Fig. 4.10(a). Comparison
results of identification rate and computational efficiency are shown in Table 4.2. We observe
the following:
1. The MinutiaCode comparator presents a degraded performance when the number of minu-
tiae is increasing in full palmprints, while the performance of the RT comparator is stable.
This means that the RT comparator is more robust than the MinutiaCode comparator.
Especially, the increase in the number of minutiae in full palmprints may be partially from
increasing spurious minutiae extracted by the minutiae extractors. Therefore, it may also
indicate that the MinutiaCode comparator suffers of spurious minutiae more than the RT
comparator does.
3. The proposed automated forensic palmprint identification system, namely the M inutiaLine+RT
system, performs best in both accuracy and computational efficiency. A rank-1 identifica-
tion rate of 68.2% is achieved by our system, compared to 63.6% for the MinutiaCode-based
identification system. The average identification time of our system is 200ms per pair, com-
pared to 300ms of the MinutiaCode-based systems using MinutiaCode-based comparison.
A finer analysis reveals that there were 6 latent marks whose mated full palmprints were
not identified in the database by both of the two complete systems. By visually checking
64
4.3 Experiments and Performance Evaluation
the failed latent marks (see one sample in Fig. 4.11), it is seen that latent-to-full palmprint
comparison heavily depends on the quality and number of reliable minutiae in the latent
marks. Among failed searches, we found out that there were 3 latent marks with scars or
recovered with interruption, 1 latent mark with quite unclear ridges and 2 latent marks
whose manually-marked minutiae number was less than 30.
0.8
0.75
0.7
0.65
0.6
0.55
0.5
2 4 6 8 10 12 14 16 18 20
Rank
(a)
Cumulative Match Characteristic
1
Combined weighting
0.95
Centroid weighting
0.9 Principal line weighting
Identification Rate
0.85
0.8
0.75
0.7
0.65
0.6
0.55
0.5
2 4 6 8 10 12 14 16 18 20
Rank
(b)
Figure 4.10: CMC curves showing performance on the forensic palmprint database. (a) Performance
of the four combined palmprint comparison systems. (b) Performance of the MinutiaLine + RT system
using three kinds of global comparison methods.
65
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
Table 4.2: Comparison results of the four combined systems on the forensic database
We evaluate the performance of different global comparison methods using our M inutiaLine
extractor and our RT comparator. Three global comparison methods are evaluated, i.e., centroid-
based weighting, principal line-based weighting and the proposed combined global comparison.
CMC curves in Fig. 4.10(b) show that the combined method achieves a rank-1 identification
rate of 68.2% outperforming the other approaches, with 63.6% of centroid-based method and
54.5% of principal line-based method.
66
4.4 Preliminary Likelihood Ratio Performance on Forensic Databases
the comparison using CMC curves not relevant. Moreover, in public competitions (e.g., PVC),
ROC curves are typical evaluation measures. For those reasons, we use the ROC curve for
comparing the performance of both systems with respect to their discriminating power. The
ROC curves of the two systems on high-resolution palmprint databases are shown in Fig. 4.12(b).
The EER of our system is 9.2% compared to 11% of the MinutiaCode-based system. These
numbers are reasonable according to the current published results of PVC [Tsinghua Univ.,
2011a]. In order to show the influence of the kind of database on the system performance, we
compare the ROC curves of the two systems on our forensic palmprint database in Fig. 4.12(a).
The EER of our system is 13.6% compared with 20.6% of the MinutiaCode-based system in the
forensic database. These results mean that for both systems there is a significant performance
degradation on forensic databases compared to public high-resolution palmprint databases, i.e.,
an absolute increase of 4.2% in the EER for our system and 8.4% for the MinutiaCode-based
system. This is true even when the minutiae in forensic databases are manually extracted. It
reveals the importance of system evaluation on real forensic databases when developing forensic
automated palmprint identification systems [Wang et al., 2012b].
Due to the fact that logistic regression, which is one typical method used for LR computation
(see Sect. 2.3), has already been used for score computation in the proposed system, the scores
of the proposed system can be considered as log(LR) values directly. This allows the straight-
forward use of the RT system in forensic evaluation of palmprint evidence. In this section, we
present LR performance results using the scores obtained by the proposed M inutiaLine+RT
system. The distribution of the LR values from the proposed system is shown in Fig. 4.13.
Here, each same-source (SS) LR is computed from a corresponding genuine score in a biometric
system, while each different-source (DS) LR is computed from a corresponding impostor score.
Thus, 22 same-source LR values and 190,938 different-source LR values are computed.
To evaluate the accuracy of the LR values obtained from the proposed system, we present
both Tippett plots (see Fig. 4.14) and ECE plots (see Fig. 4.15). From the Tippett plots, the
rate of misleading evidence for same source is 31.82%, and for different source is 11.19%. The
results indicate high rate of misleading evidence in the forensic palmprint databases, especially
for same source. However, the maximum magnitude of misleading LR values is quite reduced,
and misleading LR values are not bigger than 10 or smaller than 0.1. From the ECE plots,
we get Cllr of 0.3389 and Cllr min of 0.1443. The distance between the solid and dashed curves
indicates that the logistic regression based LR computation is not so good even when leave-one-
out training scheme is used. This needs to be improved by exploring better calibration models
for the proposed system.
67
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
68
4.5 Chapter Summary and Conclusions
ROC Curve
1
0.9
0.8
0.7
True Positive Rate
0.6
0.5
0.4
0.3
0.2
MinutiaLine+RT,EER=13.6%
0.1
MinutiaCode,EER=20.6%
0 −5 −4 −3 −2 −1 0
10 10 10 10 10 10
False Positive Rate
(a)
ROC Curve
1
0.9
0.8
0.7
True Positive Rate
0.6
0.5
0.4
0.3
0.2
0.1 MinutiaLine+RT,EER=9.2%
MinutiaCode,EER=11%
0 −4 −3 −2 −1 0
10 10 10 10 10
False Positive Rate
(b)
Figure 4.12: ROC curves for the proposed system and MinutiaCode-based system on (a) forensic palm-
print databases, (b) public high-resolution palmprint databases.
69
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
8
SS LR
6
0
-20 -15 -10 -5 0 5 10
LR
4
x 10
6
DS LR
0
-20 -15 -10 -5 0 5 10
LR
Figure 4.13: Histogram of the LR values from the proposed system. The upper: same source (SS); The
bottom: different source (DS).
80
70
Proportion of cases (%)
60
50
40
30
20
10
0
−3 −2 −1 0 1 2 3
Log10(LR) Greater Than
Figure 4.14: Tippett plot for the LR values from the proposed system.
70
4.5 Chapter Summary and Conclusions
MinutiaeLine+RT
1
LR values
After PAV
Empirical cross entropy
0.6
0.4
0.2
0
-2 -1 0 1 2
Prior log10(odds)
Figure 4.15: ECE plot for the LR values from the proposed system.
71
4. FORENSIC PALMPRINT RECOGNITION BASED ON RADIAL TRIANGULATION
72
Chapter 5
This chapter presents the methods and the experimental evaluation of a novel forensic palm-
print recognition system based on weighted complex spectral minutiae representation (Weighted-
SMC). Due to the fact that both fingerprints and palmprints suffer of distortion, especially in
forensic scenarios, robust forensic fingerprint and palmprint recognition systems should be able
to deal with this problem. Moreover, forensic palmprint recognition includes latent-to-full palm-
print comparison, which requires focusing on smaller partial regions than a full palm. We propose
modeling distortion with minutiae windows, in order to focus the palmprint comparison process
in limited regions. This distortion modeling method is then incorporated into SMC represen-
tation which is invariant to translation and rotation, in order to deal with different types of
distortion, namely rigid transformation and non-linear distortion. The combination of the dis-
tortion modeling and SMC representation motivated our proposed Weighted-SMC comparator
with applications to both fingerprints and palmprints. In this Chapter, the proposed method
of distortion assessment at feature level for fingerprints and palmprints is introduced first, and
its corresponding experimental analysis is presented. Then the Weighted-SMC representation is
proposed, and its major components are described, including minutiae representation and match-
ing score computation. Based on the new Weighted-SMC representation, a novel comparison
scheme for forensic fingerprint and palmprint recognition is proposed, namely, the Weighted-
SMC comparator. Experiments and evaluations of the Weighted-SMC comparator are described
for both fingerprints and palmprints.
The main contributions of the work presented in this chapter are as follows:
1) A new method of distortion assessment at feature level for fingerprints and palmprints
which guides the design of our proposed forensic fingerprint and palmprint recognition systems
later. This method can also serve as inspiration for future systems developed by using distortion
73
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
5.1. Definitions
5.1.1. Minutiae Window
Minutiae window is defined as the minimum 2-dimensional window which can include all
the minutiae points in a finger mark or a fingerprint. A minutiae window is generated only
based on the positions of the minutiae, not the directions. The types of windows used for our
study are as follows:
Rectangle: defined by a center point PC , the width of the window (horizontal) w and its
height (vertical) h.
Ellipse: defined by a center point PC , the length of the semi-major axis a0 , the length of
the semi-minor axis b0 and the corresponding direction of the axes.
Given a set of N minutiae from a finger mark or a fingerprint, i.e., M = {mi }N i=1 where
mi = (xi , yi , θi ), the three types of windows from M can be obtained as follows:
2. For ellipse minutiae window, the major axis and the minor axis, including both the size
and the direction, are decided by finding the minimum volume enclosing ellipsoid (MVEE)
of the minutiae set based on Khachiyan’s algorithm [Khachiyan and Todd, 1993], and the
center point PC is decided accordingly.
3. For circle minutiae window, the diameter is also the line segment connecting two minutiae
with the largest distance, and the center point PC is decided accordingly.
74
5.1 Definitions
400 w
400
mi 400
380 mi mi
380
360
h 360
350
340
340
αi αi αi
320 320
300
PC α=0 300
PC α=0 300 PC α=0
280 280 b0 r
a0
260 260 250
240 240
220 220
200
50 100 150 200 250 50 100 150 200 250 50 100 150 200 250 300
(a) (b) (c)
Figure 5.1: Different types of minutiae window for a configuration of 12 minutiae. (a) Rectangle, (b)
ellipse, and (c) circle.
In the following sections, we denote the minutiae set from a finger mark as M T = {mTi }N i=1
(here T stands for “trace” or “testing”), where mTi = (xTi , yiT , θiT ) and the minutiae set from a
fingerprint as M R = {mR N R R R R
i }i=1 (here R stands for “reference”), where mi = (xi , yi , θi ), and
the center points of the minutiae windows as PCT = (xTC , yCT ) and P R = (xR , y R ) for distortion
C C C
assessment. Three types of windows generated for a minutiae set from a sample finger mark,
with some of the parameters related to the windows and their minutiae, are as shown in Fig.
5.1.
Window distortion is defined as a kind of global distortion which measures scaling and
rotation at global level representing rigid transformation, given the ground-truth labels of minu-
tiae pairing between a testing window for the minutiae in a mark and the reference window for
the paired minutiae in its mated print. The global measures for window distortion are defined
as follows:
where WT and WR denote the measures of minutiae window size for a finger mark and its
paired fingerprint, i.e., the width w and height h for the rectangle window, the length of
semi-major axis a0 and semi-minor axis b0 for the ellipse window, and the radius r for the
circle. Therefore, specifically for the three types of windows, window scaling measures: i)
for a rectangle window, the window size variance in percentage both horizontally using w
and vertically using h; ii for an ellipse window, the window size variance in percentage
along the direction of the major axis using a0 and the direction of the minor axis using
b0 ; iii) for a circle window, the window size variance (i.e., radius variance) in percentage
using r.
75
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
Window rotation ∆αW : the average difference of the polar angles αi of the minutiae from
a finger mark and from its paired fingerprint in degrees, calculated by Eq. (5.2):
N
1 X T
∆αW = (αi − αiR ), (5.2)
N
i=1
where αiT and αiR are the polar angles of the minutia mTi from a finger mark relative to
its minutiae window center PCT and the minutia mR i from its paired fingerprint relative to
R
its minutiae window center PC in the polar coordinate systems centered with the center
points of the two minutiae windows respectively, as shown in Fig. 5.1. We emphasize that
the polar coordinate system is the typical type with a origin located at the center of each
minutiae window and a starting polar angle α = 0 horizontally. In this way, the spatial
distribution of all minutiae points in one finger mark or fingerprint keeps the same with
that in Cartesian coordinate system, which is aimed for the later observation/assessment
mode, not for matching mode.
76
5.2 Distortion Assessment at Feature Level in Fingerprints
Table 5.1: The numbers of minutiae in the paired minutiae data of the databases NIST SD27 and NFI
DB58.
Table 5.2: Window distortion assessment by different minutiae windows for fingermarks in NFI 58 with
regard to their mated prints.
Scaling Rotation
Window
Measure of size Mean Std. Mean Std.
w -0.0174 0.1431
Rectangle
h 0.0636 0.1111
a0 0.0209 0.0965 2.6911 16.3927
Ellipse
b0 0.0116 0.1020
Circle r 0.0151 0.0792
each pair of mated mark and print. Each M atch set contains the minutiae points in common
between the finger mark and its paired fingerprint. The minutiae pairing labels for the two
M atch sets in NIST SD27 are provided by the group of Intelligent Systems Laboratory in
Halmstad Univiversity [Mikaelyan and Bigun, 2012]. The paired minutiae data of NFI DB58
are annotated by the latent examiners in Netherlands Forensic Institute (NFI). As shown in
Table 5.1, the number of minutiae annotated in a finger mark in NIST SD27 vary in a big range
from 5 to 86, while in NFI DB58 the number of minutiae annotated for each mark is fixed as
12.
The global measures for window scaling and rotation are then calculated using Eq. (5.1) and
(5.2) on both databases. As shown in Fig. 5.2, on the database NFI DB58, the circle window
obtained the least amount of distortion (i.e., with the majority in the range [−15%, 15%] for
window scaling in terms of radial variance) while window rotation on the database locates in the
range [−30◦ , 72◦ ], with the majority in the range [−40◦ , 40◦ ]. More details are given in Table 5.2
where the mean and standard deviation values of each measure for window distortion on NFI
DB58 are given.
Fig. 5.3 shows distortion measures for NIST SD27. Similar with NFI DB58 database, the
circle window obtained the least amount of distortion (i.e., with the majority in the range
[−15%, 15%] for window scaling in terms of radial variance) and the rectangle window had the
largest distortion (i.e., with the majority in the range [−30%, 30%] for window scaling in both
horizontal and vertical directions), while window rotation on the database locates in the range
[−161◦ , 103◦ ], with the majority in the range [−50◦ , 50◦ ]. More details are given in Table 5.3
77
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
Table 5.3: Window distortion assessment by different minutiae windows for fingermarks in NIST SD27
with regard to their mated prints.
Scaling Rotation
Window
Measure of size Mean Std. Mean Std.
w -0.0060 0.1845
Rectangle
h 0.0108 0.1152
a0 0.0063 0.0892 -1.1245 22.1322
Ellipse
b0 -0.0030 0.1070
Circle r 0.0024 0.0799
where the mean and standard deviation values of each measure for window distortion on NIST
SD27 are also given. Actually, it is not surprising that the circle window obtains the least
amount of window distortion both on NFI DB58 and NIST SD27. This is because circle window
is invariant to rotation. Therefore, if a mark and a print are rotated to each other, window
scaling does not increase. This makes it a proper measure of distortion in databases where there
is a strong rotation between the mark and the print, like NIST SD27 and NFI DB58.
Based on the finding in Section 5.2.1 that circle window obtained least amount of distor-
tion for both databases, we only use circle window for point distortion assessment on the two
databases. The same data of paired minutiae in the databases NIST SD27 and NFI DB58 for
window distortion assessment are used for point distortion assessment, i.e., including minutia
position and direction, and labels of minutiae pairing between the minutiae set in a finger mark
and the matched minutiae set in its mated fingerprint, namely minutiae pairing labels. We then
calculate the local measures, i.e., relative differences of minutia position and direction using Eq.
(5.3) and (5.4). As shown in Fig. 5.4, the point distortion under circle window on NIST SD27 is
much larger than the distortion on NFI DB58 regarding both the difference of relative minutia
position and direction. On NIST SD27, the average difference of relative minutia position is 140
pixels compared to 95 pixels on NFI DB58. And the difference of minutia direction is mostly in
the range [−50◦ , 50◦ ] on NIST SD27, compared to the range [−20◦ , 20◦ ] on NFI DB58. Much
larger distortion is observed on NIST SD27 than on NFI DB58. This is mainly because the
number of minutiae annotated in a finger mark in NIST SD27 vary in a big range from 5 to 86,
which results in seriously varying window sizes on the database. Meanwhile, in NFI DB58 the
number of minutiae annotated for each mark is fixed as 12, and therefore the database is stable
to the assessment using minutiae window.
78
5.2 Distortion Assessment at Feature Level in Fingerprints
(a) (b)
(c) (d)
Figure 5.2: Window distortion assessment on the paired minutiae data of NFI DB58. (a) Window
scaling in rectangle window, (b) window scaling in circle window, (c) window scaling in ellipse window,
(d) window rotation in degrees.
79
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
(a) (b)
(c) (d)
Figure 5.3: Window distortion assessment on the paired minutiae data of NIST SD27. (a) Window
scaling in rectangle window, (b) window scaling in circle window, (c) window scaling in ellipse window,
(d) window rotation in degrees.
80
5.2 Distortion Assessment at Feature Level in Fingerprints
Point distortion in position NFI DB58 Point distortion in position NIST SD27
35 200
mean=94.5755, std=56.6811 mean=140.0595, std=85.6028
180
30
160
25 140
120
20
100
15
80
10 60
40
5
20
0 0
0 50 100 150 200 250 300 350 400 0 100 200 300 400 500
Relative distance in pixels Relative distance in pixels
(a) (b)
Point distortion in direction NFI DB58 Point distortion in direction NIST SD27
30 500
mean=7.4555, std=14.3545 mean=0.6999, std=16.5421
450
25
400
350
20
300
15 250
200
10
150
100
5
50
0 0
−40 −20 0 20 40 60 80 −200 −150 −100 −50 0 50 100 150
θ − distance in degrees θ − distance in degrees
(c) (d)
Figure 5.4: Point distortion assessment on the paired minutiae data of NFI DB58 (left) and NIST
SD27 (right). (a) and (b) are the relative difference of minutia position, (c) and (d) are the difference of
minutia direction.
81
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
where m(x, y; σ 2 ) is the summation of Gaussian indicator functions at the locations of all
minutiae in a fingerprint/palmprint and σ is the standard deviation parameter. Note that the
Gaussian indicator function representing a pulse at the location of a minutia is the same for
SML and SMC methods. However, we use different notations for the parameter σ, and therefore
the parameters σL and σC denote the standard deviations for the SML representation and the
SMC representation respectively. Then, we take the Fourier transform of m(x, y; σ 2 ) and obtain
the magnitude of its Fourier spectrum, i.e., for the SML representation [Xu et al., 2009],
N
ωx2 + ωy2 X
ML (ωx , ωy ; σL2 ) =| exp(− ) exp(−j(ωx xi + ωy yi )) |, (5.6)
2σL−2 i=1
In the SML representation, only the location of a minutia is transformed, while both the location
and orientation of a minutia are transformed in the SMC representation. Especially, when σ = 0,
it means that a Dirac pulse is assigned to the location of each minutia, namely mi (x, y) =
δ(x − xi , y − yi ), and the SML representation becomes
N
X
ML (ωx , ωy ) =| exp(−j(ωx xi + ωy yi )) |, (5.8)
i=1
82
5.3 Weighted Complex Spectral Minutiae (Weighted-SMC)
Finally, the continuous spectrums SML or SMC are sampled on a polar-linear grid with the size
M0 × N0 where M0 (set to 128) samples are located in the radial direction between λl and λh ,
and N0 (set to 256) samples are located in the angular direction between 0 and π for SML or
between 0 and 2π for SMC. λl and λh are parameters denoting frequency range boundaries in
the radial direction, which can be tuned during spectrum sampling.
Let R and T be the two sampled minutiae spectrums of dimensions M0 ×N0 obtained from the
minutiae sets in a reference fingerprint/palmprint and a test fingerprint/palmprint respectively.
The matching score between them is calculated as the maximum correlation coefficient between
R and T as follows:
1 X
S (R,T ) = max | R(m, n)T (m − p, n − q) |, p = 0, q ∈ [−n0 , n0 ], (5.10)
p,q M0 N0 m,n
where T (m − p, n − q) denotes a cyclic shift version of T (m, n) with a shift of p in the radial
direction for scaling and a shift of q in the angular direction for rotation. In practice, we set p = 0
and assume that there is no scaling difference between the fingerprints/palmprints in a same
database since the images in a database are either collected using a single scanner or recovered in
the same way regarding image resolution. Note that in this Dissertation we use different angular
shift range [−n0 , n0 ] for different scenarios according to the database property in terms of image
rotation. The values of [−n0 , n0 ] will be assigned in the experimental parts of the related sections
(Sects. 5.5, 6.2 and 6.3). Originally in [Xu et al., 2009], a range [−15, 15] in steps of 3 units is
used for fingerprints in civil applications, which is equivalent to a range [−20◦ , 20◦ ] in steps of
4◦ for SMC, and a range [−10◦ , 10◦ ] in steps of 2◦ for SML. In our experiments for palmprints,
e.g., for high-resolution palmprint matching on the database THUPALMLAB [Tsinghua Univ.,
2011b], we use the full angular shift range [−128, 128] in steps of one unit, equivalent to a range
[−180◦ , 180◦ ] in steps of 1.4◦ for SMC, and a range [−90◦ , 90◦ ] in steps of 0.7◦ for SML. This
is because the palmprint images in the database we use have much larger rotation compared to
the fingerprint databases used in [Xu et al., 2009; Xu and Veldhuis, 2010]. The parameters in
Equations (5.5) - (5.10) are summarized in Table 5.4.
83
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
Table 5.4: Description of parameters for spectral minutiae representation and comparison.
deducing the Fourier transform of the cylinder function only have general mathematic meaning,
and will not be used after the Fourier transform of the cylinder function is obtained. Given a
cylinder centered at the origin with radius a, the cylinder function is defined as
( p
1 f or x2 + y 2 6 a
C(x, y) = p (5.12)
0 f or x2 + y 2 > a
kx = k cos α
(5.13)
ky = k sin α
and
x = r cos θ
(5.14)
y = r sin θ.
Then
F (k, a) = Fx,y [C(x, y)](k, a)
R 2π R a
= 0 0 ei(kr cos α cos θ+kr sin α sin θ) rdrdθ (5.15)
R 2π R a
= 0 0 eikr cos(α−θ) rdrdθ
Let b = θ − α, so db = dθ. Then
R 2π−α R a ikr cos b
F (k, a) = −α e rdrdb
R 2π R a 0ikr cos b
= 0 0 e rdrdb
Ra
= 2π 0 J0 (kr)rdr (5.16)
= 2πa
k J1 (ka)
= 2πa2 J1ka
(ka)
.
84
5.3 Weighted Complex Spectral Minutiae (Weighted-SMC)
a = 10
1
0.8
0.6
0.4
0.2
0
20
10 20
0 10
0
−10 −10
−20 −20
(a) (b)
(c) (d)
Figure 5.5: Illustration of the complex spectral minutiae representation procedure using Cylinder pulse.
(a) A fingerprint and its minutiae; (b) a Cylinder pulse for a minutia with a radius of a = 10 pixels; (c)
the continuous spectrum generated from all minutiae points in a fingerprint using Cylinder pulse; (d) the
Fourier spectrum in (c) sampled on a polar-linear grid.
N
J1 (a(ωx + jωy )) X
ML (ωx , ωy ; a) =| 2πa2 exp(−j(ωx xi + ωy yi )) |, (5.17)
a(ωx + jωy )
i=1
and the SMC representation using Cylinder pulse for each minutia is defined as,
N
J1 (a(ωx + jωy )) X
MC (ωx , ωy ; a) =| 2πa2 exp(−j(ωx xi + ωy yi ) + jθi ) | . (5.18)
a(ωx + jωy )
i=1
Fig. 5.5 shows the the general procedure of complex spectral minutiae representation using
Cylinder pulse.
85
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
where d0 and ∆θ0 are the estimated maximum values for minutia point distortion in position
and direction, which can be tuned or trained on a training dataset. The first part of the complex
weight, exp(− | (di − d0 )/d0 |), is assigned to the real part of the spectra of each minutia to give
compensation in position. The second part of the complex weight, exp(−j(| (∆θi −∆θ0 )/∆θ0 |)),
is assigned to the imaginary part of the spectra of each minutia to give compensation in direction.
N
2
ωx2 + ωy2 X
MCw (ωx , ωy ; σC ) =| exp(− −2 ) wi exp(−j(ωx xi + ωy yi ) + jθi ) | . (5.21)
2σC i=1
N
J1 (a(ωx + jωy )) X
MCw (ωx , ωy ; a) =| 2πa wi exp(−j(ωx xi + ωy yi ) + jθi ) |, (5.22)
ωx + jωy
i=1
86
5.4 Minutiae Comparison Based on Weighted-SMC
where wi , the complex weight as defined in Eq. (5.19), is assigned to the original SMC represen-
tation for each minutia, representing a measure of relative minutia quality. In other words, with
respect to the original SMC representation as in [Xu and Veldhuis, 2010], our Weighted-SMC
representation weights the contribution of each minutia according to the estimated distortion in
position and direction.
In order to compute the score between two minutiae windows from a mark and a print
respectively by incorporating the information of the estimated relative minutia quality, we gen-
erate the minutiae spectrum for the reference minutiae window in the way of Weighted-SMC
representation and the minutiae spectrum for the testing minutiae window in the way of SMC
representation. The two continuous spectrums are then sampled on a polar-linear grid with
the size M0 × N0 where M0 (set to 128) samples are located in the radial direction between λl
and λh , and N0 (set to 256) samples are located in the angular direction between 0 and 2π.
Finally the matching score for comparing the two sampled spectrums R and T is calculated as
the maximum correlation coefficient between them in the same way as defined in Eq. (5.10).
In the following sections, we describe the procedures in the proposed Weighted-SMC com-
parator applied to forensic fingerprint comparison, while the same procedures can be applied to
forensic palmprint comparison.
87
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
K - mean Clustering
Density Clustering
2. Spectral comparison
Spectral
Comparison
Top1Window ? NO
Similarity
YES
Score
88
5.4 Minutiae Comparison Based on Weighted-SMC
refinement of reference minutiae window and calculation of the weight of relative minutia quality
for each minutia in a candidate window.
If the minutiae number of the initial reference window is less than 3, the window will
be discarded.
Otherwise, if in the initial reference window the number of qualifying minutiae ac-
cording to the position and direction conditions when compared to the minutiae in
the testing window, i.e., d√i ≤ d0 and ∆θi ≤ ∆θ0 , is also less than 3, the window will
be discarded. Here d0 = 22 t0 + dt where dt denotes minutia point distance in pixels,
and ∆θ0 denotes minutia direction difference in degrees. Note that d0 has a part
of value depending on the moving step t0 in pixels. This is because the maximum
distance between the moving center of the relaxation window
√
at the four closest steps
2
and the genuine center is estimated as no bigger than 2 t0 in the case the genuine
center locates at the center point of the rectangle formed by the four closest steps, as
illustrated in Fig. 5.7.
Otherwise, the refined reference window with preserved minutiae is then compared
to the testing window under a K-means clustering condition, i.e., their within-cluster
sums of point-to-centroid distances should be close as (1−p3 )sumdT ≤ sumdR ≤ (1+
p3 )sumdT where sumdT and sumdR denote the sums of point-to-centroid distances
for the minutiae set in the testing window and in the reference window respectively,
and p3 ∈ (0, 1). Here the minutiae set in the refined reference window is treated as one
cluster, and the minutiae set in the testing window is another cluster. If the refined
reference window does not qualify the K-means clustering condition, the window will
be discarded.
89
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
t0
Figure 5.7: The moving center of the relaxation window at the four closest steps to the genuine center
(red).
Otherwise, the refined reference window with preserved minutiae is then compared to
the testing window under a probability density estimate condition, i.e., their density
sums fT and fR should be close as (1−p4 )fT ≤ fR ≤ (1+p4 )fT where p4 ∈ (0, 1). Here
a typical Kernel Density Estimation (KDE) [Wand and Jones, 1995] is performed for
the minutiae set in the refined reference window and the minutiae set in the testing
window respectively. If the refined reference window does not qualify the probability
density estimate condition, the window will be discarded.
Otherwise, the refined reference window is preserved for the next step.
After candidates of reference minutiae window are selected, we then implement spectral
transform and comparison using Weighted-SMC representation as defined in Eqs. (5.20) - (5.22).
For the minutiae set enclosed in each candidate window, a weighted complex spectrum is gener-
ated and sampled in the same way as described in Sect. 5.3.1. The weighted complex spectrum
is then compared to the original complex spectrum generated from the minutiae set in the mark.
A comparison score calculated as the maximum correlation coefficient between the two sample
spectrums is then obtained for each candidate window using Eq. (5.10). The highest score
among the comparisons between those candidates of reference minutiae window and the testing
window is the final score.
In Table 5.5, we summarize the parameters required for the Weighted-SMC comparison.
This table extends the parameters described in Table 5.4.
90
5.5 Experiments and Performance Evaluation
In this Chapter, we use forensic fingerprint databases NIST SD27 and NFI DB58 as described
in Chapter 3 for performance evaluation of the Weighted-SMC comparator for fingerprints. NIST
SD27 contains 258 pairs of finger marks and fingerprints, and four minutiae sets that have been
validated by a professional team of latent examiners. We will use the less constrained minutiae
sets, i.e., one set of minutiae containing all minutiae points on the finger mark, and the second
set containing all minutiae points on the fingerprint mate. In other words, we use the NIST
SD27 database as it would be used in operational conditions for database search, for instance
by the use of an AFIS. At the comparison stage, each minutiae set of a finger mark is compared
to all minutiae sets of the 258 fingerprints. Therefore, we obtain 258 genuine scores and 66306
(= 258 × 257) impostor scores on the database NIST SD27. Similarly, NFI DB58 contains 58
pairs of finger marks and fingerprints, and two minutiae sets that have been validated by latent
examiners. While each minutiae set of a finger mark is compared to all minutiae sets of the 58
fingerprints, we obtain 58 genuine scores and 3306 (= 58 × 57) impostor scores on the database
NFI DB58.
91
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
The proposed Weighted-SMC comparator as described in Sect. 5.4 is then used for forensic
fingerprint comparison on the two databases NIST SD27 and NFI DB58. In order to evaluate
the closed-set identification task, we use the Cumulative Match Characteristic (CMC) curve.
We first compare Weighted-SMC comparator using three different pulse functions for minu-
tiae, i.e., Dirac, Gaussian and Cylinder pulse. CMC curves for the the comparator based on
three pulse functions on NFI DB58 are shown in Fig. 5.8. The parameters are tuned on this
database and will be used as trained parameters for the test on other databases, i.e., NIST SD27
and the forensic palmprint database. On NFI DB58, a rank-1 identification rate of 94.83% is
achieved for the Dirac pulse, while 87.93% and 79.3% for the Cylinder and Gaussian pulse re-
spectively. The Dirac pulse performs much better than the Cylinder and Gaussian pulse, which
indicates that a pulse allowing more distortion in spatial domain, e.g., a Cylinder pulse or Gaus-
sian pulse, does not have to better model distortion in frequency domain. The parameters used
in this evaluation on NFI DB58 are described in Table 5.6. Moreover, the proposed Weighted-
SMC comparator using either of the three pulse functions for minutiae outperforms the original
SMC method [Xu and Veldhuis, 2010] which obtained a rank-1 identification rate of 56.9% on
NFI DB58 as it can be seen in Fig. 5.8. Here the original SMC method has been adapted to
forensic fingerprint comparison, and does not consider minutiae distortion or refinement in the
comparison process. This comparison between the proposed Weighted-SMC comparator and
the original SMC method has shown significant effectiveness of weighting the relative minutiae
quality as described in Sect. 5.3.3 and justifies the distortion compensation scheme as described
92
5.5 Experiments and Performance Evaluation
0.95
Identification Rate
0.9
0.85
0.8
0.75
0.7
0.65
Weighted-SMC, Dirac pulse
0.6 Weighted-SMC, Cylinder pulse
Weighted-SMC, Gaussian pulse
0.55
Original SMC, Dirac pulse
0.5
2 4 6 8 10 12 14 16 18 20
Rank
Figure 5.8: CMC curves for Weighted-SMC comparator using Dirac, Gaussian and Cylinder pulse
respectively on NFI DB58 compared to the original SMC method using the Dirac pulse.
in Sect. 5.4. Due to the fact that the moving rate p2 effects both the moving step t0 of the
center of the relaxation window and the minutia point distance constraint d0 , we carefully choose
the parameter p2 by several trials to observe the corresponding performance, mainly the rank-1
identification rate. Results of the trials using Dirac pulse are shown in Fig. 5.9, which illustrates
the rank-1 identification rates of Weighted-SMC comparator using Dirac pulse under different
moving rate values (p2 ). And p2 = 0.1 has been chosen for the optimized value due to that its
corresponding rank-1 identification rate is the highest among all the trials performed.
Table 5.6: Values of parameters for Weighted-SMC comparator and the original SMC method on NFI
DB58.
Then using the trained parameters from NFI DB58, we evaluate the Weighted-SMC com-
93
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
0.94
0.92
0.9
0.88
0.86
0.84
0.82
0.8
0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2
p 2 - moving rate
Figure 5.9: Rank-1 identification rates of Weighted-SMC comparator using Dirac pulse under different
moving rate values (p2 ) on NFI DB58.
parator based on Dirac pulse on the public forensic fingerprint database NIST SD27. Another
set of optimized parameters from NIST SD27, as detailed in Table 5.7, is also used to obtain
comparative results of the Weighted-SMC comparator, and to evaluate the robustness of the
system to the selection of the parameters. As shown in Fig. 5.10, a rank-1 identification rate of
68.22% is achieved by the Weighted-SMC comparator using the trained parameters from NFI
DB58, while 72.87% using the optimized parameters. The decreasing performance of the com-
parator on NIST SD27 using the trained parameters from NFI DB58 is mainly due to the fact
that much larger rotation and non-linear distortion exist in NIST SD27 than in NFI DB58, as ob-
served in Sect. 5.2. It is remarkable that, even in this situation, the performance decrease of the
system is very controlled. This indicates that, although the amount of parameters to determine
is considerable, the system is robust to their selection. This also indicates that the difference
in rotation distortion amongst databases is not a critical issue regarding the robustness of the
system. Moreover, the proposed system works in a comparative way compared to the recent
results of [NIST ELFT-EFS Evaluation, 2011], specifically the minutiae-only based evaluation,
since the proposed Weighted-SMC comparator works only using minutiae feature. A detailed
performance comparison between the Weighted-SMC comparator and the best minutiae-only
based system with a rank-1 identification rate of 64.34% published in NIST ELFT-EFS Evalua-
tion is shown in Table 5.8. Note that in NIST ELFT-EFS Evaluation, the gallery size is enlarged
to 857 rolled full fingerprints based on the original NIST SD27, which means 600 images are
added for the minutiae-only based evaluation. To implement a fairer performance comparison,
94
5.5 Experiments and Performance Evaluation
0.95
Identification Rate
0.9
0.85
0.8
0.75
0.7
0.65
0.6
Table 5.7: Optimized values of parameters for Weighted-SMC comparator on NIST SD27.
further experiments should be conducted using a gallery with a size no smaller than 857 as
future work.
95
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
Table 5.8: Performance comparison between the Weighted-SMC comparator and the best minutiae-only
based system published in NIST ELFT-EFS Evaluation on NIST SD27.
Table 5.9: Optimized values of parameters for the Weighted-SMC comparator on forensic palmprint
databases.
the same. Compared to the RT comparator on the forensic palmprint databases, the Weighted-
SMC comparator performs better at all ranks when using optimized parameters. However, when
using the trained parameters from NFI DB58, the Weighted-SMC comparator achieves the same
identification rate as the RT comparator at several ranks such as rank-1, rank-3, rank-4, rank-5,
while it performs better at the rest ranks. Overall, the Weighted-SMC comparator performs
better than the RT comparator. Still we emphasize that the trained parameters from NFI
DB58 do not fit the forensic palmprint databases reasonably due to the different data properties
between fingerprints and palmprints such as minutiae number in one image and image size. A
better training dataset for palmprints can be explored for further evaluation. There are two
facts to be highlighted here. First, the difference between trained and optimized performance is
not high, and therefore we can say that the Weighted-SMC system is robust to variation in the
parameter setting. Second, if we would have a forensic palmprint database, we expect results
that would be even better than the results using trained parameters from NFI DB58. Moreover,
the different performance of the RT comparator and the Weighted-SMC comparator at different
ranks indicates that the two comparators are not extracting the same kind of information, and
therefore they can benefit each other via fusion, which will be further discussed in Chapter 6.
We also evaluate the efficiency of the proposed Weighted-SMC comparator. For fingerprints,
the average comparison time of the proposed Weighted-SMC comparator for one comparison
between a finger mark and a fingerprint using different pulses is shown in Table 5.10, i.e., 0.82s by
Dirac pulse, 0.93s by Gaussian pulse and 0.97s by Cylinder pulse. The difference in comparison
time between the three pulses is mainly due to the different complexity of spectral transform
based on different pulses, where the spectral transform based on Dirac pulse takes least time.
The original SMC method needs to compare one testing window with all the reference minutiae
96
5.5 Experiments and Performance Evaluation
0.95
0.9
Identification Rate 0.85
0.8
0.75
0.7
0.65
0.6
Weighted-SMC, Dirac pulse (parameter optimized)
0.55 Weighted-SMC, Dirac pulse (parameter trained)
MinutiaLine + RT
0.5
2 4 6 8 10 12 14 16 18 20
Rank
Figure 5.11: CMC curves for Weighted-SMC comparator on the forensic palmprint databases compared
to RT comparator.
windows at all the steps during window searching without minutiae refinement, consequently, it
performs much slower, with an average comparison time of 3.2s.
For palmprints, the average comparison time of the proposed Weighted-SMC comparator
for one comparison between a palm mark and a full palmprint using Dirac pulses is 1.82s as
shown in Table 5.10. It indicates that the Weighted-SMC comparator performs much slower
for palmprints than for fingerprints, which is mainly because the average number of minutiae
enclosed in a testing or reference window for a palm mark or a full palmprint is above 2 times
of that for a finger mark or a fingerprint. The Weighted-SMC comparator also performs 8 times
slower than the RT comparator for palmprints due to the high complexity of minutiae window
searching. This can be improved by a better procedure of the pre-selection of reference minutiae
Table 5.10: Comparison time of the Weighted-SMC comparator for fingerprints and palmprints respec-
tively.
97
5. FORENSIC PALMPRINT RECOGNITION BASED ON WEIGHTED COMPLEX SPECTRAL MINUTIAE
REPRESENTATION
window candidates such as minutiae clustering according to local characteristics. The latter is
proposed as a future improvement at the end of this Thesis.
98
Chapter 6
This chapter presents the methods and experimental analysis of several fusion schemes for
high-resolution palmprint recognition with application to forensic scenarios. One fusion scheme
based on multi-algorithm fusion is motivated by the achievable performance improvement using
different palmprint recognition systems according to information fusion, which can be applied
to forensic scenarios to improve the performance of forensic palmprint recognition. This is
also motivated by the fact that one of the systems is computationally light, and therefore the
performance improvement due to fusion does not imply a huge increase in computational burden.
Then, another fusion scheme based on regional fusion is proposed for high-resolution palmprint
recognition, due to the fact that systems can benefit of the combination of different regions in
a palm to improve the performance. Although most approaches of regional fusion previously
proposed divide the palm into regions in order to improve performance by the combination of
different regions, those previously proposed methods have not an anthropological interpretation.
Our regional fusion is inspired anthropologically using the three palm regions with different
discriminability, namely interdigital, thenar and hypothenar, and therefore it can even benefit
of accurate segmentation by human examiners who can easily segment a palm into the three
regions without previous training or knowledge about the system operation.
Based on the above motivations, multi-algorithm fusion at score level is implemented for
forensic palmprint recognition using the two algorithms proposed in Chapter 4 and 5, namely
RT and Weighted-SMC. Based on anatomic manual region segmentation, differences of the dis-
criminability of palmprints in different regions, i.e., interdigital, thenar and hypothenar, are
studied. Then, regional fusion for high-resolution palmprint recognition using palm regions seg-
mented manually is presented and evaluated. After that, in order to automate the segmentation
of the palmprints into regions, a method of automatic region segmentation for high-resolution
palmprints based on datum points is introduced and evaluated regarding segmentation accuracy.
Consequently, regional fusion using the palm regions segmented automatically is evaluated.
99
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
The main contributions of the work presented in this chapter are as follows:
1) Multi-algorithm fusion at score level for forensic palmprint recognition using the RT-based
algorithm and Weighted-SMC based algorithm.
2) A method of automatic region segmentation for high-resolution palmprints based on datum
points.
3) A new matching strategy based on regional fusion for high-resolution palmprint recogni-
tion which can improve the performance compared to region-to-region matching and full-to-full
matching.
This chapter is organized as follows. First we present multi-algorithm fusion at score level
for forensic palmprint recognition using the RT-based algorithm and Weighted-SMC based al-
gorithm (Sect. 6.1). Then we describe the study on regional discriminability using palm regions
segmented manually (Sect. 6.2). In a following section we present the methodology and exper-
imental analysis of regional fusion for high-resolution palmprint recognition based on manual
segmentation (Sect. 6.3). The next section is dedicated to a description of regional fusion
for high-resolution palmprint recognition based on automatic region segmentation (Sect. 6.4).
Finally, the summary and conclusions of this chapter are presented (Sect. 6.5).
As deeply discussed in [Ross et al., 2006], the match score is a measure of similarity between
the testing and reference biometric samples. When match scores output by different biometric
matchers are combined in order to arrive at a final recognition decision, fusion is said to be done
at the match score level. This is also known as fusion at the measurement level or confidence level.
Apart from the raw data and feature vectors, the match scores contain the richest information
about the input pattern. Also, it is relatively easy to access and combine the scores generated
by different biometric matchers. Consequently, information fusion at the match score level is
the most commonly used approach in multibiometric systems.
It must be noted that the match scores generated by the individual matchers may not be
homogeneous. For example, one matcher may output a distance or dissimilarity measure (a
smaller distance indicates a better match) while another may output a similarity measure (a
larger similarity value indicates a better match). Furthermore, the outputs of the individual
matchers need not be on the same numerical scale (range). Finally, the match scores may follow
different probability distributions. These three factors have been the main concerns of match
score level fusion.
Score level fusion generally follows classic classifier combination rules, such as Product Rule,
Sum Rule, Max Rule, Min Rule and Median Rule. Among these rules, Sum Rule has been
proved to be robust to errors in the estimation of the posteriori probabilities during classification
100
6.1 Multi-Algorithm Fusion at Score Level for Forensic Palmprint Recognition
[Kittler et al., 1998]. Therefore, the sum decision rule usually works quite well in practice and
is commonly used in multibiometric systems. In our multi-algorithm fusion at score level for
forensic palmprint recognition, we will also use sum rule. Another fusion method we use is based
on logistic regression where parameters for a linear combination of scores from different systems
is obtained using logistic regression learning and a training procedure is required. This method
has been used in score computation for the RT comparator in Chapter 4, and outputs scores
which can be used for evidence evaluation directly. This motivates us to fuse scores from the
RT comparator and the Weighted-SMC comparator using logistic regression for further evidence
evaluation.
The experimental part of this section is based on the outputs of two forensic palmprint
comparators described in previous Chapters 4 and 5, namely, RT comparator and Weighted-
SMC comparator. The two comparators have outputted two sets of scores for the performance
evaluation on the forensic palmprint databases as described in Sect. 3.3.1.1, which contains
22 palm marks from real forensic cases and 8680 full palmprints from criminal investigations
captured by Beijing Institute of Criminal Technology in China. Among the 8680 full palmprints,
there are 11 full palmprints corresponding to the 22 palm marks examined by human examiners.
The number of palm marks corresponding to one palm is varying, and at most 6 palm marks are
corresponding to one full palmprint. The comparison protocol has been that each minutiae set
of a palm mark is compared to all minutiae sets of the 8680 full palmprints. Therefore, in total
we obtain 22 genuine scores and 190938 (= 22 × 8679) impostor scores on the forensic palmprint
databases for each comparator. Then both genuine scores and impostor scores are fused using
sum rule and logistic regression. The procedures of the two fusion methods are then described
below.
Due to the ranges of the scores obtained by the two comparators are different, a min-max
normalization for each score in a score set obtained by a matcher/comparator is performed
before sum rule fusion as following
si − smin
snorm = , (6.1)
smax − smin
where snorm is the normalized value of a raw score si in a score set S = {si }N t
i=1 with Nt score
values, smin = min(S) and smax = max(S).
It is not proper to use logistic regression on the scores directly from the two comparators,
due to the fact that the RT comparator alerady outputs the final score with logistic regression,
also trained with a leave-one-out procedure. If we would perform logistic regression again with
101
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
the Weighted-SMC score, and leave-one-out procedure would be used again, the results would
not be honest, since training of the fusion would be done on already seen data. Moreover, as we
have a small forensic palmprint database (i.e., only 22 palm marks), another scheme for fusion
other than leave-one-out seems to lead to a too sparse training set, and the results would not
be reliable. However, the RT comparator generates two individual scores for each comparison
before performing logistic regression, one quantitative score Sn and one qualitative score Sq as
described in Sect. 4.2.2.3. These two scores can be fused with the score from the Weighted-SMC
comparator generated by the same comparison using logistic regression as
Slr = λ0 + λ1 × Sn + λ2 × Sq + λ3 × Sw (6.2)
102
6.1 Multi-Algorithm Fusion at Score Level for Forensic Palmprint Recognition
0.95
0.9
Identification Rate
0.85
0.8
0.75
0.7
0.65
Weighted-SMC, Dirac pulse (parameter trained)
0.6
MinutiaLine + RT
0.55 Logistic Regression fusion
Sum fusion
0.5
2 4 6 8 10 12 14 16 18 20
Rank
Figure 6.1: CMC curves for multi-algorithm fusion using RT-based system and Weighted-SMC based
system on the forensic palmprint databases.
values directly due to the use of logistic regression during score fusion. The distribution of the
LR values from the proposed system is shown in Fig. 6.2. Again, each same-source (SS) LR is
computed from a corresponding genuine score in a biometric system, while each different-source
(DS) LR is computed from a corresponding impostor score. Thus, 22 same-source LR values
and 190,938 different-source LR values are computed.
To evaluate the accuracy of the LR values obtained from the fused system, we present
both Tippett plots (see Fig. 6.3) and ECE plots (see Fig. 6.4). From the Tippett plots, the
rate of misleading evidence for same source is 9.98%, and for different source is 13.91%. The
results indicate significant improvement of misleading evidence for same source in the forensic
palmprint databases, compared to 31.82% of misleading evidence for same source using the
RT -based system. This also shows that multi-algorithm fusion at score level can improve the
performance of forensic palmprint cognition systems, even when same source samples are so
min of 0.0504. C min
few, e.g., 22 in our case. From the ECE plots, we get Cllr of 0.3672 and Cllr llr
obtained by the fused system is much smaller than the RT -based system with a Cllr min of 0.1442.
However, the distance between the solid and dashed curves is larger than that obtained by the
RT -based system. This indicates that better calibration models for the fused system needs to
be explored other than the logistic regression based LR computation. According to [Ramos and
Gonzalez-Rodriguez, 2013], this may be due to the fact that, with a higher discriminating power,
the likelihood ratios obtained should be higher in magnitude. However, the fused system seems
to be under-confident in its own scores, and the LRs are limited in value. The good news is
that this leads to a very small magnitude of misleading evidence, i.e., LR values that supports
103
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
4
SS LR
3.5
2.5
1.5
0.5
0
-20 -15 -10 -5 0 5 10
4
x 10
9
DS LR
8
0
-20 -15 -10 -5 0 5 10
Figure 6.2: Histogram of the LR values from the fused system using logistic regression. The upper:
same source (SS); The bottom: different source (DS).
104
6.1 Multi-Algorithm Fusion at Score Level for Forensic Palmprint Recognition
80
Proportion of cases (%)
70
60
50
40
30
20
10
0
-3 -2 -1 0 1 2 3
Log 10 (LR) Greater Than
Figure 6.3: Tippett plot for the LR values from the fused system using logistic regression.
Log.Reg.Fusion
1
LR values
After PAV
Empirical cross -entropy
0.6
0.4
0.2
0
-2 -1 0 1 2
Prior log (odds)
10
Figure 6.4: ECE plot for the LR values from the fused system using logistic regression.
the wrong proposition. As it can be seen, misleading LR values are not bigger than 10 and
lower than 0.1, which means that errors are not important in a Bayesian decision framework.
The reason of such under-confidence can be found in the sparsity of the database. As logistic
regression is trained with features (scores) of bigger dimension than RT (3 scores in the former
instead of 2 scores in the latter), the requirements of the amount of training data increase, and
therefore the training dataset is even relatively smaller in size than for the RT system. This
highlights the need of the availability of a wealth of data in order to perform fusion, or the
development of algorithms that require less training data to work properly. These issues are
proposed as future work in the Thesis.
105
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
Interdigital
Inner Outer
side side
Thenar c
Hypothenar
Figure 6.5: Definitions of a palmprint: principal lines (1 - heart line, 2 - head line and 3 - life line),
palmprint regions (I - interdigital region, II - thenar region and III - hypothenar region) and datum
points (a, b-endpoint, o-their midpoint) [Zhang and Shu, 1999]. c is the intersection point of the bottom
boundary of a palm and the perpendicular bisector of the line segment ab. Here ab and oc are used for
region segmentation.
As observed from Fig. 1.1 and Fig. 6.5, the three palm regions have different feature prop-
erties: 1) the interdigital region contains significant singular points and the heart line which
could improve its discriminability, 2) the thenar region contains many more minor creases and
wrinkles which deteriorate its discriminability, 3) the hypothenar region contains more regular
ridges which improve its discriminability. To check these observations, we implemented region
segmentation based on datum points [Zhang and Shu, 1999] marked manually, and obtained
region-to-region matching results on a subset of the database THUPALMLAB [Tsinghua Univ.,
2011b]. The subset of THUPALMLAB includes the palmprint images from the last 50 subjects
with 100 palms, i.e. 800 (= 50 × 2 × 8) images. The reason for using a subset is that regions
are manually segmented in the database, as will be explained below, and therefore the dataset
should be reduced in order to make this time-consuming task feasible. In the following sections,
we use this testing subset of THUPALMLAB for experimental analysis. In order to check the
generalization of the observations, we use two systems to perform the region-to-region matching
and full-to-full matching, i.e., the SMC method which has been applied to full-to-full palmprint
comparison in [Wang et al., 2013f,g] and a commercial SDK MegaMatcher 4.0 [Neurotechnology,
2011].
As shown in Fig. 6.5, due to the stability of the principal lines, the endpoints a and b of the
106
6.2 Regional Discriminability of High-Resolution Palmprints
Region I - Interdigital
life line and the heart line which intersect both sides of the palm, and their midpoint o are also
stable according to their locations in the full palmprint. They were defined as datum points
in [Zhang and Shu, 1999]. Some significant properties of datum points can be used for region
segmentation: (1) the locations of the endpoints and their midpoint are rotation invariant in a
palmprint; (2) a palm can be divided into three regions: interdigital region (I), thenar region (II)
and hypothenar region (III) by the connections between the endpoints and their perpendicular
bisector, i.e., line segments ab and oc (see Fig. 6.5).
Based on the properties of the datum points, we segment each palmprint into three regions
manually. Firstly, we manually choose endpoints a and b according to their definition and
obtain their position with x and y axis values. Then we calculate the position of their midpoint
o. Finally we divide each palmprint into those three regions by treating the line segments ab
and oc as boundary lines. One example of segmented regions for a palmprint is shown in Fig.
6.6. Three subsets are generated from the testing database (800 palmprints) of THUPALMLAB
after manual region segmentation and named as Subset I, II and III corresponding to interdigital
region (I), thenar region (II) and hypothenar region (III) respectively.
In our experiments, we first use the commercial SDK MegaMatcher 4.0 to implement feature
extraction for the three regional subsets, and then we perform region-to-region comparison using
spectral minutiae representation (see Sect. 5.3.1) on the three subsets I, II and III separately.
For each subset, there are 2800 (= (100 × 8 × 7)/2) genuine comparison attempts and 4950
107
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
(= (100 × 99)/2) impostor comparison attempts. Since the SMC method outperforms the SML
method [Wang et al., 2013g; Xu and Veldhuis, 2010], we only use SMC for spectral minutiae
representation in this section. The best results with parameter configurations for the three
regions are shown in Table 6.1, and the ROC curves are shown in Fig. 6.7. The EER results are
11.11%, 17.5% and 5.46% corresponding to interdigital region, thenar region and hypothenar
region respectively. As it can be observed, the thenar region shows much lower discriminating
power. This is consistent with the observation that the thenar region contains many more minor
creases and wrinkles which deteriorate its discriminability. It is also consistent with the result
reported in [Jain and Feng, 2009] when the matching is based on minutiae features. The results
also show that the hypothenar region outperforms the interdigital region on the THUPALMLAB
database, which is possibly because that the hypothenar region has more minutiae with better
quality due to much less minor creases in this region than in the interdigital region. Furthermore,
the interdigital and hypothenar regions outperform full-to-full comparison with EER of 14.2%
using the SMC method, but not the thenar region. This evidences that not all the regions should
receive equal treatment, like it happens in full-to-full comparison. This observation reinforces
the motivation of regional fusion, explained below.
Table 6.1: Region-to-region comparison and full-to-full comparison results using the SMC method with
Dirac pulse. [Wang et al., 2013f,g]
In order to generalize the experimental study, we also use the commercial SDK MegaMatcher
4.0 [Neurotechnology, 2011] to implement region-to-region matching on the three subsets I, II and
III separately. Similarly, for each subset, there are 2800 genuine comparison attempts and 4950
impostor comparison attempts. The EER results are 1.18%, 6.57% and 0.89% corresponding to
interdigital region, thenar region and hypothenar region respectively. The ROC curve is shown
in Fig. 6.8. It also indicates that on the THUPALMLB database, the thenar region shows
much lower matching accuracy, which is consistent with the observation at the beginning of
this section and the result reported in [Jain and Feng, 2009]. It can be seen that the results
are consistent with the SMC method, reinforcing the conclusions about regional discriminability.
Moreover, the hypothenar region outperforms interdigital region on the THUPALMLB database,
and also outperforms full-to-full comparison with EER of 1% using MegaMatcher 4.0. Again,
the evidence that the different regions should receive different treatment is confirmed, and the
motivation of regional fusion is supported.
In addition to regional discriminability, we also calculated correlation coefficients between
different region-to-region score sets obtained by MegaMatcher 4.0. The correlation results are
108
6.2 Regional Discriminability of High-Resolution Palmprints
ROC Curve
1
0.8
0.4
Figure 6.7: ROC curves for the three region-to-region comparisons using the SMC method compared to
full-to-full matching.
ROC Curve
1
0.995
True Positive Rate
0.99
0.985
EER=1%, Full−to−full
0.98 EER=1.18%, Interdigital(I)
EER=6.57%, Thenar(II)
EER=0.89%, Hypothenar(III)
0.975 −4 −3 −2 −1 0
10 10 10 10 10
False Positive Rate
Figure 6.8: ROC curves for the three region-to-region comparisons using MegaMatcher 4.0 compared to
full-to-full matching.
shown in Table 6.2. The results indicate that there is moderate correlation between genuine
scores from different regions, while impostor scores from different regions are practically un-
correlated. Fig. 6.9 shows scatter plots of three region-to-region comparisons to show the
relationship between different regional comparison score sets. As it is known that one of the
109
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
4000
S2
3000
2000
1000
0
0 500 1000 1500 2000 2500
S1
3500
3000
2500
3
S
2000
1500
1000
500
0
0 500 1000 1500 2000 2500
S1
4000
S2
3000
2000
1000
0
0 500 1000 1500 2000 2500 3000 3500 4000 4500
S3
Figure 6.9: Scatter plots of scores for the three region-to-region score sets obtained by MegaMatcher
4.0. (a) Interdigital (S1 ) vs. thenar (S2 ), (b) interdigital (S1 ) vs. hypothenar (S3 ), (c) hypothenar (S3 )
vs. thenar (S2 ).
indicators that fusion improves performance is the lack of correlation of the features to fuse,
these observations motivate us to implement regional fusion using region-to-region comparison
110
6.3 Regional Fusion at Score Level Based on Manual Region Segmentation
Region-to-region comparison
s
I I
s
II II
III III
s
Regional fusion
Fused score
111
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
sf 1 = s1 + s2 + s3 , (6.3)
where sf 1 and sf 2 are the fused output scores, and w = [w0 , w1 , w2 , w3 ] is a vector of real-valued
weights trained using logistic regression. The training procedure is described below. Note that
no normalization is needed for the scores before fusion because the system for obtaining the
regional scores is the same for the three region-to-region comparisons, and therefore the ranges
of scores in the three regions are assumed to be similar.
112
6.3 Regional Fusion at Score Level Based on Manual Region Segmentation
ROC Curve
1
0.9
0.8
0.7
True Positive Rate 0.6
0.5
0.4
0.3
Figure 6.11: ROC curves of regional fusion using sum rule and logistic regression respectively compared
to full-to-full comparison for the SMC-based matching method.
ROC Curve
1
0.9995
0.999
0.9985
True Positive Rate
0.998
0.9975
0.997
0.9965
Figure 6.12: ROC curves of regional fusion using sum rule and logistic regression respectively compared
to full-to-full comparison for MegaMatcher 4.0.
comparisons as reported in Section 6.2.2. This also indicates that regional fusion outperforms
region-to-region comparison as well as full-to-full comparison. Therefore, the hypotheses which
motivated this fusion method are confirmed. It is surprising that logistic regression based fusion
performs worse than sum fusion. The reason is that logistic regression training scheme may be
113
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
suffering of data sparsity, especially for EER values which are very low, where sample sizes need
to be even bigger.
The purpose of edge detection in general is to significantly reduce the amount of data in an
image, while preserving the structural properties to be used for further image processing. In our
research, we need to preserve strong edges in palmprints, i.e., principal lines and contour lines
for the next-stage processing since datum points are normally located on those strong edges. To
this purpose, we use Canny edge detector [Canny, 1986] for palmprint edge detection as it has
become one of the standard edge detection methods still used in research commonly.
114
6.4 Regional Fusion at Score Level Based on Automatic Region Segmentation
Generally, the Canny edge detection algorithm runs five steps as follows.
Step 1: Smoothing. To remove noise, Gaussian filter with the form in Equations (6.5)
and (6.6) is used for blurring the image.
hg (n1 , n2 )
h(n1 , n2 ) = PN1 PN2 , (6.5)
n1 =1 n2 =1 hg (n1 , n2 )
−(n1 2 +n2 2 )
hg (n1 , n2 ) = e 2σ 2 , (6.6)
where N1 and N2 denote the filter size in width and height, and σ is the standard deviation.
In our experiments, the filter size N1 = N2 = 15 and standard deviation σ = 8 are chosen
empirically for palmprints according to [Jain and Demirkus, 2008]. The effect of smoothing
a sample palmprint (Fig. 6.13 (a)) with the filter is shown in Fig. 6.13 (b).
Step 2: Finding gradients. By calculating gradients at each pixel in the smoothed image
in the x- and y-direction respectively, the gradient magnitudes | G | (also known as the
edge strengths) can then be determined as an Euclidean distance measure by applying
the law of Pythagoras as shown in Equation (6.7),
q
| G |= Gx 2 + Gy 2 , (6.7)
Step 3: Non-maximum suppression. In this step, the smoothed edges in the image of
gradient magnitudes (see Fig. 6.13 (c)) are converted to sharp edges. Basically this is done
by preserving all local maxima in the gradient image, and setting other pixels to black
pixels, i.e., pixel value = 0. To determine all local maxima, an 8-connected neighbourhood
is used for each pixel in the gradient image.
Step 4: Double thresholding. Potential edges are determined by two thresholds, i.e., a
low threshold µl and a high threshold µh calculated by
115
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
(a) (b)
(c) (d)
Figure 6.13: Canny edge detection for a sample palmprint. (a) Input image, (b) smoothed image, (c)
gradient magnitudes (IG ), (d) final strong edges detected (IE ).
Step 5: Edge tracking. Final edges are determined by suppressing all edges that are not
connected to a very certain (strong) edge, as shown in Fig. 6.13 (d).
The source code of Canny edge detector we used can be found in [Senapati, 2011], and the
images in this section are generated using that code.
116
6.4 Regional Fusion at Score Level Based on Automatic Region Segmentation
It is interesting that the datum point b and contour line on the outer side of a palm appear
on both the gradient magnitude image IG (see Fig. 6.13 (c)) and the final edge image IE (see
Fig. 6.13 (d)). However, the convex hulls of strong pixels in the two images can be different
due to much less details in the edge image. Inspired by this finding, we propose a novel method
of datum point detection based on convex hull comparison. It contains three steps as described
below.
Here we choose the differential step t = 20 after observing the ridge width in the palm-
prints. Both thresholds gl and gr are close to g0 in order to preserve the details of contour
line on the outer side of a palm where the datum point b exists, and their empirical values
are 1.7 and 1.6 respectively. Taking the left palm image used in Section 6.4.1.1 as an ex-
ample, we show its gradient magnitude image after thresholding and left differential image
after thresholding in Fig. 6.14.
117
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
(a) (b)
Figure 6.14: Thresholding of gradient magnitudes. (a) Gradient magnitude image after thresholding
(IG0 ), (b) Left differential image after thresholding (IGl ).
where D (p, p0 ) is the Euclidean distance between two pixels p and p0 , and the threshold
D0 is 12 pixels empirically. Finally, the datum point candidate be is located at the most
upper-left point in KGb with the direction assigned by the line segment from the candidate
to the secondary most upper-left point for a left palm image, or the most upper-right point
with the direction assigned by the line segment from the candidate to the secondary most
upper-right point for a right palm image. Similarly, we perform convex hull comparison
between the two images IG0 and IGl for a left palm image or IGr for a right palm image,
to locate another candidate bl or br . For a left palm image, the detected candidates of
datum point b are shown in Fig. 6.15 (a) and (b).
For each candidate of datum point b detected in Step 2, we estimate the position of
its corresponding candidate of datum point a according to the following rules: (1) The
direction of a is perpendicular to the direction of b; (2) a is on the right side of b for a left
palm image, and on the left side of b for a right palm image; (3) the distance between b and
a falls in a relatively small range and then an average value of the distance can be used to
determine a. Note also that a small variation in the distance of point a to point b may be
of little consequence in order to do the segmentation of the palm in regions, because the
segmentation using ab is mostly dominated by the perpendicular direction of b, and the
distance between a and b may suffer small variation without a significant impact in the
segmentation. In our experiments, we use a statistical average value of 1272 in pixel as
118
6.4 Regional Fusion at Score Level Based on Automatic Region Segmentation
2000 2000
1800 1800
1600 bl 1600 be ae
al
1400 1400
1200 1200
1000 1000
800 800
600 600
400 400
200 200
0 0
0 200 400 600 800 1000 1200 1400 1600 1800 2000 0 200 400 600 800 1000 1200 1400 1600 1800 2000
(a) (b)
b
a
(c)
Figure 6.15: Detected datum points by using convex hull. (a) Convex hull on IG0 and IGl , (b) convex
hull on IG0 and IE , (c) final datum points a and b.
the length of ab estimated from manual selected datum points on the palmprint database
THUPALMLAB. After applying the above rules, two candidates ae and al or ar can be
located from both be and bl /br candidates. The final datum points a and b are determined
as the midpoints of ae al and be bl for a left palm or the midpoints of ae ar and be br for a
right palm. For a left palm image, the detected candidates of datum point a are shown in
Fig. 6.15 (a) and (b), and the final detected datum points a and b are shown in Fig. 6.15
(c).
119
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
After datum points a and b are located, a palm can be divided into three regions, including
interdigital region (I), thenar region (II) and hypothenar region (III) by the connections between
the endpoints and their perpendicular bisector, i.e., line segments ab and oc as shown in Fig.
6.5. Firstly, we calculate the position of the midpoint o of line segment ab. Then the endpoint
c is located as the intersection of the perpendicular bisector of ab and the bottom boundary of
the palmprint in most cases. Due to pose effect when capturing palmprints under pressure, the
endpoint c can also locate on the bottom part of left/right boundary. Finally we divide each
palmprint into those three regions by treating the line segments ab and oc as boundary lines.
The under detected pixels which are pixels that belong to the manually segmented region
and not belong to the automatically segmented region.
The over detected pixels which are pixels that belong to the automatically segmented
region and not belong to the manually segmented region.
120
6.4 Regional Fusion at Score Level Based on Automatic Region Segmentation
(a)
(c)
(b)
Figure 6.16: Regional segmentation error. (a) Automatically segmented region, (b) manually segmented
region, (c) error area.
where NDi is the number of different pixels for each region, i = 1, 2, 3, corresponding to Region
I, II, III respectively. NF is the total number of pixels in the original full palmprint, simply
determined by the full image size. In our experiments, NF = 2040 × 2040.
Finally, evaluating on a database, we will get 3 regional error rates represented by the average
values of ejR from all palmprints in each of the three regions respectively. Similarly, we will also
get one total error rate represented by the average value of eF from all palmprints.
121
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
The database used in this section is still the subset of THUPALMLAB as described in Sect.
6.2 with 800 palmprint images since this is the subset for which manual segmentation labels
are available. We first evaluate our automatic region segmentation method compared to manual
segmentation on the subset. Then we implement region-to-region comparison and regional fusion
using regions segmented automatically. And the two fusion methods as described in Sect. 6.3.1
are used as well, i.e., sum rule and logistic regression.
As our automatic segmentation method mainly works by assuming that datum point b exists
in the palmprint, we first divide the database into two categories manually, i.e., Subset A with
visible datum point b and Subset B without visible datum point b (see one sample image in Fig.
6.17 (a)), and count the number of images on each subset. We emphasize that it is better that a
further procedure for automatic detection of the visibity of datum point b is developed, however,
this can be jutified as in forensic scenarios ten-print cards are collected under control to ensure
complete full palmprints are preserved in the reference database. We then implement automatic
region segmentation on Subset A. After the automatic segmentation, we also count the number
of failed images (see one sample image in Fig. 6.17 (b)) due to low image quality which means
bad contrast between the palm region and the background. The categories are shown in Table
6.3. Among the 800 palmprint images, there are 70 images belonging to Subset B without visible
datum point b . Among the 730 images belonging to Subset A, there are 702 images segmented
successfully and 28 images failed in segmentation due to the failure of detecting datum point b.
Using the successfully segmented images in Subset A, we obtained results regarding regional
segmentation error and total segmentation error as shown in Table 6.4. The regional error rates
of interdigital (I), thenar (II) and hypothenar (III) regions are 15.72%, 17.05% and 21.38%
respectively. And the total error rate is 19.54% relative to full palmprint images.
As shown in Table 6.3, there are 702 images segmented successfully using the proposed
segmentation method, among which there are 85 palms obtaining successfu segmentation for
their complete 8-image sets. We then use 680 images from the 85 palms and their automatically
122
6.4 Regional Fusion at Score Level Based on Automatic Region Segmentation
(a) (b)
Figure 6.17: Sample images where datum point b was not detected. (a) A sample image from Subset B
without datum point b visible, (b) A failed image from Subset A with datum point b visible.
Table 6.4: Error rates of automatic region segmentation compared to manual segmentation on Subset
A (i.e., datum point b is visible and successfully detected).
segmented regions to obtain regional fusion results. Therefore, for each region-to-region compar-
ison, we obtain 2380 (= (85 × 8 × 7)/2) genuine scores and 3570 (= (85 × 84)/2) imposter scores
using MegaMatcher 4.0 by comparing automatically segmented regions. Then, we compute
regional fused scores with sum rule and logistic regression respectively using region-to-region
comparison scores. We also obtain full-to-full comparison scores for the 680 images using Mega-
Matcher 4.0 for performance reference. As shown in Fig. 6.18, evaluated using automatically
segmented regions, the ROC curve is also improved by regional fusion based on sum rule and
logistic regression compared to full-to-full comparison. The EERs of regional fusion based on
sum rule and logistic regression are both 0.17%, compared to 0.5% of full-to-full comparison and
0.46%, 4.79%, 0.42% corresponding to the three individual region-to-region comparisons. This
also indicates that regional fusion outperforms region-to-region comparison as well as full-to-full
comparison. Again as seen in Fig. 6.18, the logistic regression based fusion performs slightly
worse than sum rule fusion, which may be due to the logistic regression training scheme suffering
of data sparsity when EER values are very low and not big enough sizes of samples used for
123
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
training.
ROC Curve
1
0.99
0.98
True Positive Rate
0.97
0.96
0.95
0.94
0.93
ROC Curve
1
0.9995
0.999
True Positive Rate
0.9985
0.998
0.9975
0.997
0.9965
0.996
Full palm, EER=0.5%
0.9955
LR fusion, EER=0.17%
Sum fusion, EER=0.17%
0.995
-4 -3 -2 -1 0
10 10 10 10 10
False Positive Rate
(b)
Figure 6.18: ROC curves for region-to-region comparison and regional fusion using regions segmented
automatically. (a) Region-to-region comparison, (b) sum fusion vs. full-to-full comparison.
124
6.5 Chapter Summary and Conclusions
palm regions segmented manually. Then we have presented the methodology and experimen-
tal analysis of regional fusion for high-resolution palmprint recognition based on manual region
segmentation. We have also described a method of automatic region segmentation for high-
resolution palmprints based on datum points and evaluated the method regarding segmentation
accuracy. Finally, we have presented experimental results of regional fusion using the palm
regions segmented automatically. We have shown that the regional fusion indeed improves per-
formance very significantly, either for manual or automatic segmentation. Therefore, regional
fusion aids full-to-full comparisons in typical forensic tasks such as massive database searches for
suspect identification, victim identification, or database coherence. As it is not expected that
the full-to-full comparison scheme will be used for evidence evaluation in court, the performance
of likelihood ratio (LR) is not presented here, although this fusion scheme outputs scores that
can be interpreted as log −LR values.
125
6. FUSION SCHEMES FOR HIGH-RESOLUTION PALMPRINT RECOGNITION
126
Chapter 7
This Thesis has addressed some of the challenges of forensic palmprint recognition, show-
ing how novel methods can deal with them. These methods include new palmprint recognition
algorithms, the analysis of large non-linear distortion, and fusion techniques using different sys-
tems and different regions of the palm. After a background description of the so-called arising
field of forensic biometrics, a detailed vision of the state-of-the-art in high-resolution palmprint
recognition which can be applied to forensic applications has been presented, together with
novel techniques of minutiae feature representation for fingerprints and approaches of likeli-
hood ratio based evidence evaluation using biometric systems. We have proposed two novel
minutiae-based comparison algorithms for forensic palmprint recognition, namely, radial tri-
angulation (RT) and weighted complex spectral minutiae representation (Weighted-SMC). A
new method of distortion assessment using minutiae window has conducted for a quantitative
study of distortion in fingerprints and for the design of a robust forensic palmprint comparator,
namely the Weighted-SMC comparator. Using the two proposed algorithms, we then have ana-
lyzed the performance improvement by multi-algorithm fusion at score level. Based on a careful
study of regional discriminability for full high-resolution palmprint images, an anthropologically-
inspired regional fusion scheme has been proposed, and achieved a significant improvement for
high-resolution palmprint recognition in full-to-full matching mode. A new automatic region
segmentation method based on datum points and convex hull comparison for high-resolution
palmprints has been proposed. Then, regional fusion using automatically segmented regions
have also shown a significant improvement of performance in full-to-full matching. Inherent in
the different chapters, contributions of this thesis has been detailed and properly evaluated in
rigorous experimental set-ups.
7.1. Conclusions
Chapter 1 described the background of the so-called field of forensic biometrics, a new area
arising from the fields of biometrics and forensic science. Then, basic concepts involved in
forensic palmprint recognition and evaluation were introduced, with emphasis on forensic science,
127
7. CONCLUSIONS AND FUTURE WORK
128
7.2 Future Work
The proposed RT-based system is novel for palmprint comparison regarding accuracy and
efficiency, and can be used for multi-algorithm fusion and evidence evaluation.[Wang et al.,
2011, 2012a,b, 2013b]
The proposed Weighted-SMC-based system is novel and robust for both fingerprint and
palmprint comparison with distortion compensation, and can be used for multi-algorithm
fusion.[Wang et al., 2013a,e]
The proposed distortion assessment is novel for the assessment at feature level and can
guide the design of palmprint comparison systems.[Wang et al., 2013a,e]
The proposed scheme of regional fusion can significantly improve the performance in full-
to-full palmprint matching mode.[Wang et al., 2013d,f,g]
The proposed automatic region segmentation can achieve good segmentation error, and
can be used for the purpose of regional fusion.[Wang et al., 2013c]
Exploring new local comparison method incorporating clustering theory to group minutiae
with similar local characteristics, due to the fact that a main problem in palmprints is
related to the different behaviours of features in full palmprints. A most recent work [Liu
129
7. CONCLUSIONS AND FUTURE WORK
et al., 2013] has contributed with a coarse to fine matching strategy based on minutiae
clustering and minutiae match propagation.
Exploring new hierarchical matching propagation in which local comparison and global
comparison can be well combined and integrated to avoid the dependence of the perfor-
mance of either part. A recent work [Cappelli et al., 2012] has dedicated to that direction
and made efforts to improve global comparison by a relaxation procedure.
Exploring new reliable distortion assessment at feature level in forensic scenarios. Although
the proposed distortion assessment method using minutiae window can observe both rigid
transform and non-linear distortion in fingerprint and palmprints, new improved versions
could achieve more reliable results, e.g., ways of generating minutiae windows and measures
of window distortion and point distortion can be modified by considering the complete
properties of realistic forensic biometric data such as big rotation and displacement range.
Exploring new automatic region segmentation methods which can deal with high-resolution
palmprint images with various quality levels, and also adapt to the protocols used by
forensic practitioners.
Exploring hierarchical fusion schemes for multi-algorithm fusion at feature level. This
requires a deep study on specific algorithms deployed for fusion and fusion techniques at
feature level, which happens to be an open research line. Some guide can be found in [Ross
et al., 2006].
Studying the application of LR-based evidence evaluation framework for palmprint evi-
dence. A most relevant work [Neumann et al., 2012] for fingerprint evidence evaluation
using radial triangulation can provide some insight towards this purpose.
130
References
C. G. Aitken, F. Taroni, and J. Wiley. Statistics and the evaluation of evidence for forensic scientists, volume 10.
Wiley Online Library, 2004. 6, 27, 28, 36
N. M. Allinson. Fingerprint compression. In S. Li, editor, Encyclopedia of Biometrics, chapter 6, pages 446–452.
Springer, New York, second edition, 2011. 4, 13
J. P. Baker, D. E. Maurer, and M. Laurel. Fusion of biometric data with quality estimates via a bayesian belief
network. In Proceedings of the Biometric Symposium, Arlington, VA, pages 21–22, 2005. 6
R. M. Bolle, J. H. Connell, S. Pankanti, N. K. Ratha, and A. W. Senior. The relation between the roc curve and
the cmc. In Automatic Identification Advanced Technologies, 2005. Fourth IEEE Workshop on, pages 15–20.
IEEE, 2005. 35
M. Bone and C. Crumbacker. Facial recognition: Assessing its viability in the corrections environment. Corrections
Today, 63(4):62–65, 2001. 32
N. Brummer. Measuring, refining and calibrating speaker and language information extracted from speech. PhD
thesis, Stellenbosch: University of Stellenbosch, 2010. 39
N. Brümmer and J. du Preez. Application-independent evaluation of speaker detection. Computer Speech &
Language, 20(2):230–275, 2006. 37, 39, 40
J. Buckleton. A framework for interpreting evidence. Forensic DNA evidence interpretation, pages 27–63, 2005.
28
131
REFERENCES
R. Cappelli, M. Ferrara, and D. Maio. A fast and accurate palmprint recognition system based on minutiae.
IEEE Trans. Syst., Man, Cybern. B, 42(3):956–962, 2012. 4, 14, 16, 17, 18, 47, 130
R. Cappelli, M. Ferrara, and D. Maltoni. Minutia cylinder-code: A new representation and matching technique for
fingerprint recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 32(12):2128–2141,
2010. xxiv, 16, 17, 19
R. Cappelli, D. Maio, D. Maltoni, J. L. Wayman, and A. K. Jain. Performance evaluation of fingerprint verification
systems. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 28(1):3–18, 2006. 31
C. Champod. Forensic applications, overview. In S. Li, editor, Encyclopedia of Biometrics. Springer, New York,
2009. 3, 26
C. Champod and D. Meuwly. The inference of identity in forensic speaker recognition. Speech Communication,
31(2):193–203, 2000. 28, 36, 37
J. Christmas, J. Kittler, and M. Petrou. Structural matching in computer vision using probabilistic relaxation.
IEEE Trans. on Pattern Analysis and Machine Intelligence, 17(8):749–764, 1995. 49
T. Connie, A. T. B. Jin, M. G. K. Ong, and D. N. C. Ling. An automated palmprint recognition system. Image
and Vision computing, 23(5):501–515, 2005. 4
R. Cook, I. W. Evett, G. Jackson, P. Jones, and J. Lambert. A hierarchy of propositions: deciding which level to
address in casework. Science & Justice, 38(4):231–239, 1998. 3, 26, 28
J. Dai, J. Feng, and J. Zhou. Robust and efficient ridge-based palmprint matching. IEEE Trans. on Pattern
Analysis and Machine Intelligence, 34(8):1618–1632, 2012. xxiv, xxiv, 4, 14, 19, 20, 22
J. Dai and J. Zhou. Multifeature-based high-resolution palmprint recognition. IEEE Trans. on Pattern Analysis
and Machine Intelligence, 33(5):945–957, 2011. 4, 14, 18, 60
A. P. Dawid. The well-calibrated bayesian. Journal of the American Statistical Association, 77(379):605–610,
1982. 38
M. H. DeGroot and S. E. Fienberg. The comparison and evaluation of forecasters. The statistician, pages 12–22,
1983. 38, 39
D. Dessimoz and C. Champod. Linkages between biometrics and forensic science. In A. J. et al., editor, Handbook
of Biometrics. Springer, New York, 2007. 3, 26, 37, 38
S. Dewan. Elementary, watson: Scan a palm, find a clue. The New York Times, 2003. URL http://www.nytimes.
com/. 1
J. Eric H. Holder, L. O. Robinson, and J. H. Laub. The Fingerprint Sourcebook. U.S. Department of Justice
Office of Justice Programs, Washington, DC 20531, 2011. 1
I. Evett and J. Buckleton. Statistical analysis of str data. In 16th Congress of the International Society for Forensic
Haemogenetics (Internationale Gesellschaft für forensische Hämogenetik eV), Santiago de Compostela, 12–16
September 1995, pages 79–86. Springer, 1996. 37
I. W. Evett. Expressing evaluative opinions: a position statement. Science & Justice, 51(1):1-2, 2011. several
signatories. 3, 6, 26, 27
132
REFERENCES
FBI. Latent hit of the year award videos, 2006a. URL http://www.fbi.gov/about-us/cjis/fingerprints_
biometrics/iafis/iafis_latent_hit_of_the_year. 1
J. Feng, J. Zhou, and A. Jain. Orientation field estimation for latent fingerprint enhancement. Pattern Analysis
and Machine Intelligence, IEEE Transactions on, 35(4):925–940, 2013. ISSN 0162-8828. 50
T. Gneiting and A. E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American
Statistical Association, 102(477):359–378, 2007. 38
A. Goumeidane and M. Khamadja. Error measures for segmentation results: Evaluation on synthetic images. In
Proc. IEEE International Conference on Electronics, Circuits, and Systems (ICECS), pages 158–161, Athens,
Greece, Dec. 2010. 120
P. Grother and P. J. Phillips. Models of large population recognition performance. In Computer Vision and Pattern
Recognition, 2004. CVPR 2004. Proceedings of the 2004 IEEE Computer Society Conference on, volume 2,
pages II–68. IEEE, 2004. 35
D. Huang, W. Jia, and D. Zhang. Palmprint verification based on principal lines. Pattern Recognition, 41:
1316–1328, 2008. 54, 55
A. Jain and M. Demirkus. On latent palmprint matching. Technical Report MSU-CSE-08.8.2008, Michigan State
University, MI, May 2008. xxiii, xxiii, xxiii, xxiii, xxiv, 5, 13, 14, 17, 18, 21, 115
A. Jain and J. Feng. Latent palmprint matching. IEEE Trans. on Pattern Analysis and Machine Intelligence, 31
(6):1032–1047, 2009. xxiv, 5, 7, 8, 13, 21, 22, 23, 49, 55, 62, 63, 108, 128
A. Jain, P. Flynn, and A. Ross, editors. Handbook of Biometrics. Springer, New York, 2007. 3
133
REFERENCES
A. Jain, A. Ross, and S. Prabhakar. An introduction to biometric recognition. Circuits and Systems for Video
Technology, IEEE Transactions on, 14(1):4–20, 2004. 31
A. K. Jain and J. Feng. Latent fingerprint matching. Pattern Analysis and Machine Intelligence, IEEE Transac-
tions on, 33(1):88–100, 2011. 50
L. G. Khachiyan and M. J. Todd. On the complexity of approximating the maximal inscribed ellipsoid for a
polytope. Mathematical Programming, 61(1):137–159, 1993. 74
J. Kittler, M. Hatef, R. P. Duin, and J. Matas. On combining classifiers. 20(3):226–239, 1998. 101, 111
D. G. Kleinbaum, M. Klein, and E. R. Pryor. Logistic Regression: A Self-Learning Text. Springer, New York,
third edition, 2010. 62, 102
R. P. Krish, J. Fierrez, D. Ramos, R. Veldhuis, and R. Wang. Towards quantification of the weight of evidence with
partial fingermarks. In 6th Pacific Rim Symposium on Image and Video Technology (PSIVT), Guanajuato,
Mexico, Oct. 2013a. 11
R. P. Krish, J. Fierrez, D. Ramos, and R. Wang. On the importance of rare features in afis-ranked latent finger-
print matched templates. In Proc. 47th International Carnahan Conference on Security Technology (ICCST),
Medellı́n, Colombia, Oct. 2013b. 11
R. P. Krish, J. Fierrez, D. Ramos, and R. Wang. Towards quantification of the weight of evidence with partial
fingermarks on real forensic casework. In Proc. BBfor2 Conference on Biometric Technologies in Forensic
Science (BTFS), Nijmegen, the Netherlands, Oct. 2013c. 11
J. Li and G. Shi. A novel palmprint feature processing method based on skeleton image. In Proc. IEEE Inter-
national Conference on Signal Image Technology and Internet Based Systems (SITIS’08), pages 221–228, Bali,
Indonesia, Nov. 2008. 14, 53
J. Li, G. Shi, Y. Zheng, and Y. Liu. The research on offline palmprint identification. In Proc. IEEE World
Congress on Computer Science and Information Engineering (CSIE’09), pages 587–590, Los Angeles, CA,
USA, Mar. 2009. 14, 53
C. Lin, T. Chuang, and K. Fan. Palmprint verification using hierarchical decomposition. Pattern Recognition, 38
(12):2639–2652, 2005. 54
E. Liu, A. Jain, and J. Tian. A coarse to fine minutiae-based latent palmprint matching. Pattern Analysis and
Machine Intelligence, IEEE Transactions on, 35(10):2307–2322, 2013. ISSN 0162-8828. xxiv, 5, 7, 21, 22, 24,
26, 47, 49, 55, 129
L. Liu, D. Zhang, and J. You. Detecting wide lines using isotropic nonlinear filtering. IEEE Trans. on Image
Processing, 16(6):1584–1595, 2007. 54
J. S. Long. Regression Models for Categorical and Limited Dependent Variables. SAGE Publications, Thousand
Oak, 1997. 63
D. G. Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision,
60(2):91–110, 2004. 21
G. Lu, D. Zhang, W. K. Kong, and M. Wong. A palmprint authentication system. In A. J. et al., editor, Handbook
of Biometrics. Springer, New York, 2007. xxiii, 5
A. J. Mansfield and J. L. Wayman. Best practices in testing and reporting performance of biometric devices.
Centre for Mathematics and Scientific Computing, National Physical Laboratory Teddington, Middlesex, UK,
2002. 32
134
REFERENCES
T. Mansfield, G. Kelly, D. Chandler, and J. Kane. Biometric product testing final report. Contract, 92(4009):
309, 2001. 32
A. Martin, G. Doddington, T. Kamm, M. Ordowski, and M. Przybocki. The det curve in assessment of detection
task performance. Technical report, DTIC Document, 1997. 34
D. Meuwly and A. Drygajlo. Forensic speaker recognition based on a bayesian framework and gaussian mixture
modelling (gmm). In 2001: A Speaker Odyssey-The Speaker Recognition Workshop, 2001. 29
D. Meuwly and R. Veldhuis. Forensic biometrics: From two communities to one discipline. In Biometrics Special
Interest Group (BIOSIG), 2012 BIOSIG-Proceedings of the International Conference of the, pages 1–12. IEEE,
2012. 3
A. Mikaelyan and J. Bigun. Ground truth and evaluation for latent fingerprint matching. In Computer Vision
and Pattern Recognition Workshops (CVPRW), 2012 IEEE Computer Society Conference on, pages 83–88,
Providence, Rhode Island, June 2012. 44, 77
H. Moon and P. J. Phillips. Computational and performance aspects of pca-based face-recognition algorithms.
Perception-London, 30(3):303–322, 2001. 35
C. Neumann, I. Evett, and J. Skerrett. Quantifying the weight of evidence from a forensic fingerprint comparison:
a new paradigm. Journal of the Royal Statistical Society: Series A (Statistics in Society), 175(2):371–415,
2012. 1, 8, 24, 27, 29, 48, 74, 130
H. Ogawa. Labeled point pattern matching by fuzzy relaxation. Pattern Recognition, 17(5):569–573, 1984. 49
B. Paltridge. Thesis and dissertation writing: an examination of published advice and actual pratice. English for
specific purposes, 21:125–143, 2002. 9
P. J. Phillips, H. Moon, S. A. Rizvi, and P. J. Rauss. The feret evaluation methodology for face-recognition
algorithms. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 22(10):1090–1104, 2000b. 31
S. Pigeon, P. Druyts, and P. Verlinde. Applying logistic regression to the fusion of the nist99 1-speaker submissions.
Digital Signal Processing, 10(1-3):237–248, 2000. 111
D. Ramos. Forensic Evaluation of the Evidence Using Automatic Speaker Recognition Systems. PhD thesis,
Universidad Autonoma de Madrid, Madrid, Nov. 2007. xxiii, xxiv, xxiv, xxiv, xxiv, 3, 4, 6, 27, 36, 37, 38
D. Ramos and J. Gonzalez-Rodriguez. Reliable support: Measuring calibration of likelihood ratios. Forensic
science international, 2013. 37, 38, 39, 103
S. Ranade and A. Rosenfeld. Point pattern matching by relaxation. Pattern Recognition, 12(4):269–275, 1980.
49, 50
A. A. Ross, K. Nandakumar, and A. K. Jain, editors. Handbook of Multibiometrics. Springer, 2006. xxiii, 5, 7,
100, 130
135
REFERENCES
M. J. Saks and J. J. Koehler. The coming paradigm shift in forensic identification science. Science, 309:892–895,
2005. 1
L. J. Savage. Elicitation of personal probabilities and expectations. Journal of the American Statistical Associa-
tion, 66(336):783–801, 1971. 38
S. Singh, P. Correia, and L. Soares. Improved rotation-invariant degraded partial palmprint recognition technique.
In Biometrics and Forensics (IWBF), 2013 International Workshop on, pages 1–4. IEEE, 2013. 21
S. Singh, M. Ramalho, P. L. Correia, and L. D. Soares. Pp-rider: A rotation-invariant degraded partial palmprint
recognition technique. In Signal Processing Conference (EUSIPCO), 2012 Proceedings of the 20th European,
pages 1499–1503. IEEE, 2012. 21
Z. Tan, J. Yang, Z. Shang, G. Shi, and S. Chang. Minutiae-based offline palmprint identification system. In Proc.
IEEE Global Congress on Intelligent Systems (GCIS’09), pages 466 – 471, Xiamen, China, May 2009. xxiii,
xxiii, 4, 14, 16, 63
ENFSI. Enfsi standard for the formulation of evaluative reports in forensic science, 2013. 27
Neurotechnology. Megamatcher 4.0 sdk, 2011. URL http://www.neurotechnology.com. 14, 64, 106, 108
NIST ELFT-EFS Evaluation. Elft-efs evaluation 1 final report (nistir 7775), 2011. URL http://biometrics.
nist.gov/cs_links/latent/elft-efs/NISTIR_7775.pdf. 31, 94, 96
B. T. Ulery, R. A. Hicklin, J. Buscaglia, and M. A. Roberts. Accuracy and reliability of forensic latent fingerprint
decisions. Proceedings of the National Academy of Sciences of USA, 108(19):7733–7738, 2011. 1
M. M. P. Wand and M. C. Jones. Kernel smoothing, volume 60. Crc Press, 1995. 90
H. Wang and E. R. Hancock. Probabilistic relaxation labelling using the fokker-planck equation. Pattern Recog-
nition, 41(11):3393–3411, 2008. 49
R. Wang, D. Meuwly, R. Veldhuis, D. Ramos, J. Fierrez, and R. Haraksim. Weighted complex spectral minutiae
representation for forensic fingerprint recognition. Pattern Recognition, 2013a. to be submitted by October,
2013. 10, 129
136
REFERENCES
R. Wang, D. Ramos, and J. Fierrez. Latent-to-full palmprint comparison based on radial triangulation under
forensic conditions. In Proc. IAPR/IEEE International Joint Conference on Biometrics (IJCB), Washington,
DC, USA, Oct. 2011. Best Poster Paper Award. 9, 10, 48, 55, 56, 129
R. Wang, D. Ramos, and J. Fierrez. Improving radial triangulation-based forensic palmprint recognition according
to point pattern comparison by relaxation. In Proc. IAPR/IEEE International Conference on Biometrics (ICB),
pages 427–432, New Delhi, India, Mar. 2012a. 10, 27, 48, 55, 129
R. Wang, D. Ramos, and J. Fierrez. Performance of semi-automatic biometric palmprint recognition comparing
databases of simulated and forensically realistic marks. In Proc. 6th European Academy of Forensic Science
Conference (EAFS), The Hague, The Netherlands, Aug. 2012b. 10, 67, 129
R. Wang, D. Ramos, and J. Fierrez. Forensic palmprint recognition based on radial triangulation. Pattern
Analysis and Applications, 2013b. submitted. 9, 10, 48, 129
R. Wang, D. Ramos, J. Fierrez, and R. P. Krish. Automatic region segmentation for high-resolution palm-
print recognition: Towards forensic scenarios. In Proc. 47th International Carnahan Conference on Security
Technology (ICCST), Medellı́n, Colombia, Oct. 2013c. 10, 129
R. Wang, D. Ramos, J. Fierrez, and R. P. Krish. Towards regional fusion for high-resolution palmprint recognition.
In Proc. XXVI SIBGRAPI conference on Graphics, Patterns and Images, Arequipa, Peru, Aug. 2013d. 10, 11,
129
R. Wang, D. Ramos, D. Meuwly, R. Veldhuis, J. Fierrez, and R. Haraksim. Assessing latent fingerprint distortion
using forensic databases and minutiae paring by human experts. In Proc. BBfor2 Conference on Biometric
Technologies in Forensic Science (BTFS), Nijmegen, the Netherlands, Oct. 2013e. 10, 11, 129
R. Wang, D. Ramos, R. N. Veldhuis, L. Spreeuwers, J. Fierrez, and H. Xu. Regional fusion for high-resolution
palmprint recognition using spectral minutiae representation. IET Biometrics, 2013f. submitted. xxix, 10,
106, 108, 129
R. Wang, R. N. Veldhuis, D. Ramos, L. Spreeuwers, J. Fierrez, and H. Xu. On the use of spectral minutiae in
high-resolution palmprint recognition. In Proc. International Workshop on Biometrics and Forensics (IWBF),
Lisbon, Portugal, Apr. 2013g. xxix, 10, 106, 108, 129
G. N. Watson. A treatise on the theory of Bessel functions. Cambridge university press, 1995. 85
X. Wu, D. Zhang, and K. Wang. Palm line extraction and matching for personal authentication. IEEE Trans.
on Systems, Man and Cybernetics, Part C, 36(5):978–987, 2006. 54
H. Xu, R. Veldhuis, A. Bazen, T. Kevenaar, T. Akkermans, and B. Gokberk. Fingerprint verification using
spectral minutiae representations. 4(3):397–409, 2009. 7, 24, 82, 83
H. Xu and R. N. Veldhuis. Complex spectral minutiae representation for fingerprint recognition. In Proc. IEEE
Computer Society Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1–8,
San Francisco, USA, June 2010. xxiv, 7, 24, 25, 26, 82, 83, 87, 92, 108
J. Yang, G. Shi, S. Chang, Z. Tan, and Z. Shang. A novel method of minutiae filtering based on line feature
extraction. In Proc. IEEE International Conference on Intelligent Human-Machine Systems and Cybernetics
(IHMSC’09), pages 343–346, Hangzhou, China, Aug. 2009. 14, 53
X. Yang, J. Feng, and J. Zhou. Palmprint indexing based on ridge features. In Proc. IEEE International Joint
Conference on Biometrics (IJCB), pages 1–8, Washington, DC, USA, Oct. 2011. 4
137
REFERENCES
D. Zhang, W. K. Kong, J. You, and M. Wong. Online palmprint identification. Pattern Analysis and Machine
Intelligence, IEEE Transactions on, 25(9):1041–1050, 2003. ISSN 0162-8828. 4
D. Zhang and W. Shu. Two novel characteristics in palmprint verification: Datum point invariance and line
feature matching. Pattern Recognition, 32(4):691–702, 1999. xxvi, 16, 106, 107, 114
L. Zhang and D. Zhang. Characterization of palmprints by wavelet signatures via directional context modeling.
IEEE Trans. on Systems, Man and Cybernetics, Part B, 34(3):1335–1347, 2004. 54
Y. Zheng, Y. Liu, G. Shi, J. Li, and Q. Wang. Segmentation of offline palmprint. In Proc. IEEE International
Conference on Signal-Image Technologies and Internet-Based System (SITIS’07), pages 804–811, Shanghai,
China, Dec. 2007a. 14, 51
Y. Zheng, G. Shi, L. Zhang, Q. Wang, and Y. Zhao. Research on offline palmprint image enhancement. In Proc.
IEEE International Conference on Image Processing (ICIP’07), pages 541–544, San Antonio, Texas, USA,
Sept. 2007b. 14, 51
138