Segmentation and Separation of Overlapped Latent Fingerprints Algorithms, Techniques, and Datasets

SPRINGER BRIEFS IN COMPUTER SCIENCE
Branka Stojanović
Oge Marques
Aleksandar Nešković
Segmentation
and Separation of
Overlapped Latent
Fingerprints
Algorithms,
Techniques, and
Datasets
123
SpringerBriefs in Computer Science
Series editors
Stan Zdonik, Brown University, Providence, RI, USA
Shashi Shekhar, University of Minnesota, Minneapolis, MN, USA
Xindong Wu, University of Vermont, Burlington, VT, USA
Lakhmi C. Jain, University of South Australia, Adelaide, SA, Australia
David Padua, University of Illinois Urbana-Champaign, Urbana, IL, USA
Xuemin Sherman Shen, University of Waterloo, Waterloo, ON, Canada
Borko Furht, Florida Atlantic University, Boca Raton, FL, USA
V. S. Subrahmanian, Department of Computer Science, University of Maryland,
College Park, MD, USA
Martial Hebert, Carnegie Mellon University, Pittsburgh, PA, USA
Katsushi Ikeuchi, Meguro-ku, University of Tokyo, Tokyo, Japan
Bruno Siciliano, Dipartimento di Ingegneria Elettrica e delle Tecnologie
dell’Informazione, Università di Napoli Federico II, Napoli, Italy
Sushil Jajodia, George Mason University, Fairfax, VA, USA
Newton Lee, Institute for Education, Research and Scholarships, Los Angeles,
CA, USA
SpringerBriefs present concise summaries of cutting-edge research and practical
applications across a wide spectrum of fields. Featuring compact volumes of 50 to
125 pages, the series covers a range of content from professional to academic.
Typical topics might include:
• A timely report of state-of-the art analytical techniques
• A bridge between new research results, as published in journal articles, and a
contextual literature review
• A snapshot of a hot or emerging topic
• An in-depth case study or clinical example
• A presentation of core concepts that students must understand in order to make
independent contributions
Briefs allow authors to present their ideas and readers to absorb them with
minimal time investment. Briefs will be published as part of Springer’s eBook
collection, with millions of users worldwide. In addition, Briefs will be available
for individual print and electronic purchase. Briefs are characterized by fast, global
electronic dissemination, standard publishing contracts, easy-to-use manuscript
preparation and formatting guidelines, and expedited production schedules. We
aim for publication 8–12 weeks after acceptance. Both solicited and unsolicited
manuscripts are considered for publication in this series.
More information about this series at http://www.springer.com/series/10028

Branka Stojanović • Oge Marques
Aleksandar Nešković
Segmentation and Separation

of Overlapped Latent
Fingerprints
Algorithms, Techniques, and Datasets
123
Branka Stojanović Oge Marques
Vlatacom Research and College of Engineering
Development Institute Ltd Belgrade and Computer Science
Belgrade, Serbia Florida Atlantic University
Boca Raton, FL, USA
Aleksandar Nešković
School of Electrical Engineering
University of Belgrade
Belgrade, Serbia
ISSN 2191-5768 ISSN 2191-5776 (electronic)

SpringerBriefs in Computer Science
ISBN 978-3-030-23363-1 ISBN 978-3-030-23364-8 (eBook)
https://doi.org/10.1007/978-3-030-23364-8
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG.
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
B.S.: To the memory of my father.
O.M.: For Ingrid, with love and appreciation.
A.N.: For George and Milan—my lovely kids.
Preface
The field of biometrics is well situated in the research community and has intrigued
researchers for many years. Although there are numerous biometric modalities in
use today, fingerprints remain the dominant one, because of its noninvasive nature
and ease of applicability. This is especially important when it comes to security
and forensic applications. A very challenging task in fingerprint recognition is over-
lapped latent fingerprint processing, which includes segmentation and separation
processes prior to (individual) fingerprints matching.
This book presents an overview of problems and technologies behind segmen-
tation and separation of overlapped latent fingerprints. Written from a technical
perspective, and yet using language and terminology accessible to non-experts,
it describes the technologies, introduces relevant datasets, highlights the most
important research results in each area, and outlines the most challenging open
research questions.
It is targeted at a scientific audience and enthusiasts interested in the field
of fingerprints matching, in particular, and biometrics, in general. By offering a
structured overview of the most important approaches currently available, putting
them in perspective, and suggesting numerous resources for further exploration, the
book gives its readers a clear path for learning new topics and engaging in related
research.
We expect that the book will fulfill its goal of serving as a preliminary reference
on the subject. Readers who want to deepen their understanding of specific topics
will find more than 100 references to additional sources of related information.
We want to express our gratitude to the Vlatacom Research and Development
Institute, Belgrade, Serbia—management, scientific council, and personnel—for
their encouragement and support during overlapped latent fingerprint research
process and writing this book.
This book represents a part of the doctoral dissertation of Dr. Branka Stojanović
at the School of Electrical Engineering, University of Belgrade. It is also one of the
results of a joint research project between Florida Atlantic University and Vlatacom
Research and Development Institute.
vii
viii Preface
We owe special gratitude to Dr. Borko Furht—Professor of Computer Science

and Engineering and Director of the NSF Industry/University Cooperative Research
Center (CAKE) at Florida Atlantic University, Boca Raton, FL—for his help and
support.
We would also like to thank Susan Lagerstrom-Fife and her team at Springer for
their support throughout this project.
Belgrade, Serbia Branka Stojanović

Boca Raton, FL, USA Oge Marques
Belgrade, Serbia Aleksandar Nešković
January 2019
Contents
1 Latent Fingerprints Matching Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Latent Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Manual Latent Fingerprints Matching: The ACE-V Method . . . . . . . . . . 2
1.4 Automated Latent Fingerprints Matching System . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Overlapped Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Latent Fingerprint Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Latent Fingerprint Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2.1 Publicly Available Latent Fingerprint Datasets . . . . . . . . . . . . . . . . 10
2.2.2 Specialized Overlapped Latent Fingerprint Datasets . . . . . . . . . . 11
2.3 The Vlatacom Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.2 Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.3 Dataset Structure and Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.4 Additional Aspects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.5 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3 Overlapped Latent Fingerprints Segmentation: Problem Definition . . . 21
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 The Mask Segmentation Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Evaluation Methodology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
ix
x Contents
4 Machine Learning Based Segmentation of Overlapped Latent

Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Description of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1 CNN Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.2.2 CNN Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Experimental Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
5 Overlapped Latent Fingerprints Separation: Problem Definition. . . . . . . 35
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
5.2 Orientation Field Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
5.2.1 Orientation Field Estimation Algorithms for Single
Fingerprint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2.2 Orientation Field Estimation Algorithms for
Overlapped Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
5.3 Overlapped Latent Fingerprint Separation: Representative
Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4 Evaluation Methodology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
6 Machine Learning Based Separation of Overlapped Latent
Fingerprints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.2 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.3 Initial Orientation Field Estimation and Enhancement . . . . . . . . . . . . . . . . . 47
6.3.1 Initial Orientation Field Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.3.2 Single Region Orientation Field Enhancement. . . . . . . . . . . . . . . . . 47
6.4 Orientation Separation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.5 Neural Network: Architecture and Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.6 Experimental Evaluation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Chapter 1
Latent Fingerprints Matching Systems
Abstract This chapter presents fundamental concepts and terminology associated

with latent fingerprint processing, with particular reference to overlapped latent
fingerprints. It introduces the matching process itself, including classical manual
latent fingerprints matching procedures. It also presents the design of a modern
automated latent fingerprints matching system and describes its main components
and steps.
1.1 Introduction
Fingerprints have been widely used for personal identification and criminal inves-
tigations for more than a century. The popularization of computer-based biometric
authentication in recent years has driven a significant body of research associated
with fingerprint processing and verification [16].
When used as a biometric modality, such as passport control at immigration
booths and official ID cards or badges, fingerprints can be captured using offline
(inked) or live-scan methods, following well-established protocols that ensure that
the resulting fingerprints are of good quality for future use in fingerprint verification
(matching) operations. In the context of crime scene investigation, however, the
lifted fingerprints (called latent fingerprints, or simply latents) are usually of
poor quality and contain large overlap between the foreground area (containing
the friction ridge pattern) and a noisy background. These aspects make latent
fingerprints segmentation and enhancement (for subsequent matching) a difficult
problem [17].
These challenges become even more significant when two or more fingerprints
overlap, which happens quite often in fingerprints lifted from crime scenes (the same
surface was touched by two or more fingers at different times), but might also occur
in live-scan fingerprint images when the surface of fingerprint sensors contains
residues of fingerprints of previous users [17]. Overlapped latent fingerprints
separation is an open research challenge for which there have been several recently
proposed solutions in the literature (see Chaps. 5 and 6).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 1

B. Stojanović et al., Segmentation and Separation of Overlapped
Latent Fingerprints, SpringerBriefs in Computer Science,
https://doi.org/10.1007/978-3-030-23364-8_1
2 1 Latent Fingerprints Matching Systems
1.2 Latent Fingerprints
Latent fingerprints (or simply latents) are partial fingerprints lifted from a crime
scene. They are usually produced by the transfer of sweat and/or grease accumulated
in the ridges of a person’s fingers when they come in contact with an object. In
forensic applications, they are collected using chemical development and enhance-
ment methods, such as powder dusting and iodine fuming. The resulting fingerprint
images are then matched against a database in order to arrest suspects and bring
them to justice [13].
The quality of the resulting latent fingerprint images can vary significantly among
different exemplars (Fig. 1.1).
1.3 Manual Latent Fingerprints Matching: The ACE-V

Method
Manual examination of latent fingerprints follows the ACE-V (Analysis, Compar-

ison, Evaluation, Verification) method, which consists of four sequential phases
[1, 15]:
1. Analysis: where the latent print examiner analyzes the latent impression(s) in
order to determine the sufficiency of a latent impression. This stage has two
major outcomes: (1) the fingerprint is labeled to convey a sense of its quality
for performing following steps, (2) the matching features are marked.
2. Comparison: where a manual comparative measurement (matching) is made
between the latent and the available exemplar fingerprints (i.e., fingerprints of
known subjects), based on previously marked matching features. The result of
this stage is the list of similarities and differences between the latent and available
exemplar fingerprints.
Fig. 1.1 Examples of latent fingerprint images of varying quality from NIST Special Database 27
(SD-27A): (a) good, (b) bad, and (c) ugly [9]
1.4 Automated Latent Fingerprints Matching System 3
3. Evaluation: where the examiner derives a conclusion of: (1) Identification (or
Individualization)—when there is evidence to conclude that the latent and
the exemplar record have come from the same source; (2) Exclusion—when
the latent print cannot be assigned to any known exemplar labels; or (3)
Inconclusive—when the examiner is unable to make a decision regarding the
unknown latent prints. The conclusion is made based on the previously created
list of similarities and differences.
4. Verification: where a second examiner reviews all findings in order to verify the
original conclusions.
The scientific accuracy, reliability, and consistency of manual latent fingerprints
matching have been the subject of many studies and debates. The results of multiple
studies on the consequences and implications of human performance in matching
latent fingerprints [3–6, 8, 10, 12, 19] are rather inconsistent, and sometimes even
contradictory.
1.4 Automated Latent Fingerprints Matching System
An automated latent fingerprint recognition system is designed with three main

goals in mind: (1) there should be minimal human intervention; (2) the decisions
should be deterministic, thereby eliminating subjective inconsistency; and (3) the
time required for comparison should be substantially shorter than the time taken by
the human counterpart [14].
As shown in Fig. 1.2, an automated latent fingerprint recognition system con-
sists of four main tasks, which are performed in a predetermined sequence, as
follows [17]:
Fig. 1.2 Schematic view of a typical automated latent fingerprint recognition system
1. Latent fingerprint segmentation, also known as region of interest (ROI) extrac-

tion: The goal of the segmentation stage is to separate the foreground latent
fingerprint from its background. In addition to the usual challenges of automatic
image segmentation, latent fingerprints segmentation is particularly hard because
it should not only be capable of marking out the outline boundary of a latent, but
also—ideally—leave out any smudges and structured noises inside the boundary,
producing a region of interest (ROI) that labels all the foreground regions
accurately, while including as minimum background information as possible.
2. Latent fingerprint enhancement: Once a latent fingerprint has been segmented, it
is usually post-processed to remove noise and enhance the ridge structure. The
goal of the enhancement stage is to produce at its output a fingerprint image that
is more suitable for feature extraction than the original segmented image. This
process is usually accompanied by a quality assessment step, where quality, in
this case, is a prediction of the “matchability” of the fingerprint image [13]: if
the segmented impression does not contain the minimally required information
to make a valid confident match, it should be discarded as FTE (Failure To Enroll)
or FTR (Failure To Register) fingerprints.
3. Feature extraction: The feature extraction stage encodes the contents of a
(segmented and enhanced) fingerprint into a compact and robust representation,
which should ensure fast matching while maintaining the fingerprint’s unique-
ness. Fingerprint features can be broadly classified into three categories: overall
ridge flow pattern (Level 1), minutiae points (Level 2), and extended features
(Level 3) such as dots, pores, and incipient ridges.
4. Matching: A fingerprints matching algorithm compares two fingerprint images
and returns either a degree of similarity between the two fingerprints or a binary
decision (match/no match) [13]. Fingerprints matching process is a challenging
problem, mainly due to the large intra-class variations involved, i.e., the large
variability in different impressions of the same finger [13]. The two most
prominent problems associated with automated latent fingerprints matching are:
the limited amount of available information and the presence of noise.
A comparative study of research results in each of the individual stages described
earlier—namely: segmentation, quality assessment, enhancement, automatic feature
extraction, and feature matching—has shown that, even with manual annotation of
minutiae features, a maximum accuracy of about 75% can be achieved in the NIST
SD-27 database [14], which supports the argument for additional research in every
aspect of the system [17].
1.5 Overlapped Fingerprints
Overlapped fingerprints are typically found in latent fingerprints lifted from crime
scenes as well as in live-scan fingerprint images, when the surface of fingerprint
sensors contains residues of fingerprints of previous users [17]. Figure 1.3 shows
examples of overlapped latent fingerprint images from different datasets.
1.5 Overlapped Fingerprints 5
Fig. 1.3 Examples of overlapped latent fingerprint images from different databases: (a) NIST
Special Database 27 (SD-27A) [9], (b) Tsinghua SOF [7], and (c) Tsinghua OLF [7]
An overlapped fingerprint image might contain two or more component finger-

prints; however, due to the complexity of the problem, all recent studies deal with
overlapped fingerprints containing only two component fingerprints. Processing
such overlapped fingerprints presents a challenge, which begins at the segmentation
stage: when overlapped fingerprints are present in the image, the segmentation
algorithm must—in addition to separating foreground from background—segment
and distinguish the individual component fingerprints as well [17] (see Chaps. 3
and 4).
An automated overlapped latent fingerprints matching system contains the same
basic building blocks as the automated latent fingerprints matching system described
in Sect. 1.4, with additional challenges in the two steps that take place before feature
extraction, namely: latent fingerprints segmentation (or ROI extraction) and latent
fingerprint enhancement.
A typical solution consists of the following steps (Fig. 1.4) [17]:
• Segmentation—This step involves latent fingerprints segmentation (also known
as ROI extraction), which consists of manual1 segmentation of component
fingerprints’ region masks.
• Initial orientation field estimation and enhancement—The initial orientation
field (which consists of a matrix containing information about ridge angle in
every pixel/block on fingerprint image) for the overlapped fingerprint image
is usually estimated and enhanced using a block-based approach. This step
1 Manual segmentation is still the norm in the current state of the art. In Chap. 4 we present exper-
imental results for an algorithm capable of performing automatic ROI extraction of component
fingerprint regions—background (B), single fingerprint (S), and overlapped fingerprints (O)—
from the overlapped fingerprint image.
Fig. 1.4 Example of an automated overlapped fingerprint separation solution (Neural Network
(NN) based solution, adapted from [18])
produces a labeled image containing three different regions: (1) background

region without orientation values, (2) single region with one orientation value
per block, and (3) overlapped region with two orientation values, randomly
distributed, per block.
• Orientation separation—Mixed orientation fields of overlapped regions for two
component fingerprints are separated, smoothed, and enhanced in order to correct
remaining errors. The output of this step are component fingerprints’ orientation
fields.
• Component fingerprint images extraction—Two component fingerprints are
extracted from the overlapped fingerprints image by filtering the overlapped
image with the appropriate filters—usually two different Gabor filters [2, 11],
tuned to the corresponding component fingerprints’ orientation fields.
1.6 Concluding Remarks
In this chapter we discussed the process of latent fingerprints matching, its applica-
tions, and associated problems. We have shown that latent fingerprint recognition is
a very challenging problem from the point of view of image analysis and pattern
recognition, which becomes even more complex in cases where two or more
fingerprints overlap. Overlapped latent fingerprint recognition introduces another
step in the matching workflow, namely the separation of individual component
fingerprints, which is still an open research question.
References 7
Takeaways from this chapter:

• Fingerprints matching has numerous applications, both in the civil sphere
(where fingerprints are a popular biometric identifier) and in investigations
of crimes based on fingerprints left at the crime scene.
• The problem of latent fingerprints matching is particularly challenging, due
to the fact that the fingerprints are often noisy and incomplete.
• Overlapped latent fingerprints matching represents an even more com-
plex problem—compared to single latent fingerprints matching—since it
requires fingerprint separation as an additional (and the most complex) step
in the fingerprint processing pipeline.
• The process of overlapped latent fingerprints matching is still semi-
automated, because fingerprints segmentation in most approaches is
performed manually.
• Latent fingerprints matching remains an open and very active research area.
References
1. D. Ashbaugh, Quantitative-Qualitative Friction Ridge Analysis: An Introduction to Basic and

Advanced Ridgeology (CRC Press, Boca Raton, 1999)
2. J.G. Daugman, Uncertainty relation for resolution in space, spatial frequency, and orientation
optimized by two-dimensional visual cortical filters. J. Opt. Soc. Am. A 2(7), 1160–1169
(1985)
3. I.E. Dror, D. Charlton, Why experts make errors. J. Forensic Identif. 56(4), 600 (2006)
4. I.E. Dror, A.E. Peron, S.-L. Hind, D. Charlton, When emotions get the better of us: the effect
of contextual top-down processing on matching fingerprints. Appl. Cogn. Psychol. 19(6), 799–
809 (2005)
5. I.E. Dror, D. Charlton, A.E. Péron, Contextual information renders experts vulnerable to
making erroneous identifications. Forensic Sci. Int. 156(1), 74–78 (2006)
6. I.E. Dror, C. Champod, G. Langenburg, D. Charlton, H. Hunt, R. Rosenthal, Cognitive issues in
fingerprint analysis: inter-and intra-expert consistency and the effect of a ‘target’ comparison.
Forensic Sci. Int. 208(1), 10–17 (2011)
7. J. Feng, Y. Shi, J. Zhou, Robust and efficient algorithms for separating latent overlapped
fingerprints. IEEE Trans. Inf. Forensics Secur. 7(5), 1498–1510 (2012)
8. P.A. Fraser-Mackenzie, I.E. Dror, K. Wertheim, Cognitive and contextual influences in
determination of latent fingerprint suitability for identification judgments. Sci. Justice 53(2),
144–153 (2013)
9. M.D. Garris, R.M. McCabe, Fingerprint minutiae from latent and matching tenprint images, in
Tenprint Images, (National Institute of Standards and Technology, 2000)
10. L.J. Hall, E. Player, Will the introduction of an emotional context affect fingerprint analysis
and decision-making? Forensic Sci. Int. 181(1), 36–39 (2008)
11. L. Hong, Y. Wan, A. Jain, Fingerprint image enhancement: algorithm and performance
evaluation. IEEE Trans. Pattern Anal. Mach. Intell. 20(8), 777–789 (1998)
12. G. Langenberg, Precision, reproducibility, repeatability, and biasability of conclusions result-
ing from the ACE-V process. J. Forensic Identif. 59(2), 219 (2009)
13. D. Maltoni, D. Maio, A. Jain, S. Prabhakar, Handbook of Fingerprint Recognition (Springer

Science and Business Media, London, 2009)
14. A. Sankaran, M. Vatsa, R. Singh, Latent fingerprint matching: a survey. IEEE Access 2, 982–
1004 (2014)
15. Document #9 Standard for the Documentation of Analysis, Comparison, Evaluation, and
Verification (ACE-V) in Tenprint Operations, Ver. 2.0, Scientific Working Group on Friction
Ridge Analysis, Study and Technology (SWGFAST), National Institute of Standards and
Technology
16. B. Stojanović, A. Nešković, O. Marques, Fingerprint ROI segmentation using Fourier coeffi-
cients and neural networks, in 23rd Telecommunications Forum Telfor (TELFOR), 2015 (IEEE,
Piscataway, 2015), pp. 484–487
17. B. Stojanović, O. Marques, A. Nešković, Latent overlapped fingerprint separation: A review.
Multimed. Tools Appl. 76(15), 16263–16290 (2017)
18. B. Stojanović, A. Nešković, O. Marques, A novel neural network based approach to latent
overlapped fingerprints separation. Multimed. Tools Appl. 76(10), 1–25 (2016)
19. K. Wertheim, G. Langenburg, A. Moenssens, A report of latent print examiner accuracy during
comparison training exercises. J. Forensic Identif. 56(1), 55 (2006)
Chapter 2
Latent Fingerprint Datasets
Abstract This chapter presents an overview of existing image datasets that can be
used for evaluating approaches for overlapped fingerprint separation. It gives special
attention to the Vlatacom dataset, created by the authors and publicly available,
which consists of 120,000 synthetically overlapped test images (and the associated
masks), with and without noise, processed with three different rotation angles, and
in two variations of overall brightness.
2.1 Introduction
Every field of scientific research requires datasets that can be used to evaluate new
approaches to a problem, benchmark solutions, and ensure research reproducibility.
The fields of fingerprints segmentation, enhancement, verification (matching), and
overlapped fingerprint separation are no exception to this rule. The systematic
evaluation of fingerprint processing and analysis algorithms usually requires a
large number of sample images, which ideally should be made freely and publicly
available to the research community at large.
Collecting a sufficiently large number of fingerprints of suitable quality is, how-
ever, an expensive, tedious, and delicate process. To circumvent those limitations,
competitions for collecting large number of fingerprints, such as FVC 2000 [5] and
2002 [6], have been organized in the past. Collecting overlapped fingerprints is an
even more challenging task, since, in addition to the need to find volunteers willing
to participate in fingerprint acquisition, the process also involves using forensic
techniques which require experienced examiners [10].
Until recently, there were two main datasets of overlapped fingerprints in use:
Tsinghua OLF and Tsinghua SOF, which are described in more detail in Sect. 2.2.
More recently, a new dataset, called the Vlatacom dataset (VLD) was proposed
[12] to overcome several limitations of the Tsinghua SOF dataset. The Vlatacom

10 2 Latent Fingerprint Datasets
dataset, freely and publicly available,1 consists of 120,000 synthetically overlapped

test images (and the associated masks), with and without noise, processed with
three different rotation angles, and in two variations of overall brightness. It will
be described in detail throughout this chapter.
2.2 Latent Fingerprint Datasets
Research in the field of latent fingerprints requires large publicly available datasets,
preferably acquired in real environments. Producing such latent fingerprint datasets
is a very challenging process for several reasons, summarized in [9], including:
• The need for professional expertise in collecting and lifting latent fingerprints.
• The time required to lift and collect latent fingerprints.
• The appropriate equipment and scarcity of trained practitioners associated with
several latent fingerprint lifting techniques.
• Difficulty in agreeing on the meaning of “simulating real-time environments”,
since latent fingerprints collected from crime scenes may vary widely in terms of
quality and possible backgrounds.
• The challenge of creating datasets with large enough number of samples
and enough variability (e.g., multiple sensors, multiple backgrounds, multiple
sessions, and varying quality).
2.2.1 Publicly Available Latent Fingerprint Datasets
There are three publicly available latent fingerprint datasets: NIST SD-27A2 [3],
IIIT-D Latent Fingerprint3 [7], and IIIT-D Simultaneous Latent Fingerprint
(SLF)4 [8]. The fingerprints in these datasets have been acquired at different
times, in different environments, and have significantly different characteristics,
summarized in Table 2.1.
1 If you are interested in obtaining the Vlatacom dataset, please send your request by email to
[email protected]. The archive includes all the image files (organized into folders
as described in Sect. 2.3.3), the masks for all component images, the CSV files with x and y
coordinates (and type) of singular points within the overlapped region, and a README file that
explains how to use the dataset.
2 http://www.nist.gov/itl/iad/ig/sd27a.cfm.
3 http://www.iab-rubric.org/resources.html.
4 http://www.iab-rubric.org/resources.html.
2.2 Latent Fingerprint Datasets 11
Table 2.1 Latent fingerprint datasets

IIIT-D simultaneous
latent fingerprint
Dataset NIST SD-27A IIIT-D latent fingerprint (SLF)
Size 291 1046 1080
Description Consists of grayscale Consists of latent Consists of latent
fingerprint images, fingerprints of 15 subjects fingerprints of 15
corresponding minutiae and their mated optical slap subjects and their
and selected latent fingerprints. Contains mated optical slap
fingerprints multiple instances for every fingerprints.
corresponding to fingerprint, enabling latent to Contains images of
fingerprint images in the latent fingerprint 500 and 1000 ppi
data set. Contains images comparison. Contains resolutions
of 500 and 1000 ppi images of 500 and 1000 ppi
resolutions resolutions
Table 2.2 Overlapped latent fingerprint datasets

Dataset Tsinghua OLF Tsinghua SOF
Size 100 100
Description Consists of: (1) grayscale overlapped Consists of: (1) grayscale simulated
fingerprint images obtained by pressing overlapped fingerprint images
two of twelve different fingers on a synthesized from the Db1_b
white paper, enhanced with forensic (impressions no. 3 and no. 4) [2] subset
dust and scanned; (2) corresponding of the Db1 fingerprint dataset from
component fingerprint masks; and (3) FVC2002a [6]; (2) corresponding
corresponding template fingerprints. component fingerprint masks; and (3)
Contains images of 500 ppi resolution corresponding template fingerprints.
Contains images of 500 ppi resolution
a The FVC 2002 dataset is publicly available (http://bias.csr.unibo.it/fvc2002/). The Db1 subset of
FVC 2002 is divided into two parts, Db1_a and Db1_b: Db1_a contains 800 samples (8 impressions
of 100 different fingerprints), while Db1_b contains 80 samples (8 impressions of 10 different
fingerprints); it contains grayscale single fingerprint images of size 388 × 374 pixels (500 ppi),
taken from the subjects in organized sessions [6]
2.2.2 Specialized Overlapped Latent Fingerprint Datasets
For overlapped fingerprint verification, until recently, there were only two publicly
available datasets: the Tsinghua Overlapped Latent Fingerprint dataset (Tsinghua
OLF)5 [2] and Tsinghua Simulated Overlapped Fingerprint dataset (Tsinghua
SOF)6 [2]. Their main characteristics are summarized in Table 2.2.
The main differences between the Tsinghua OLF and Tsinghua SOF datasets
are [12]: (1) the overlapped fingerprints in OLF are obtained by forensic means,
while the fingerprints in SOF are artificially overlapped; (2) the samples in the
5 http://ivg.au.tsinghua.edu.cn/.
6 http://ivg.au.tsinghua.edu.cn/.
Fig. 2.1 Overlapping ratios histogram for the Tsinghua OLF (left) and Tsinghua SOF (right)
datasets (adapted from [12])
OLF dataset are of lower quality than their SOF counterpart, since the former
contains background noise; and (3) the size of the overlapped region (expressed as
overlapping ratios—defined as the ratio between the size of the overlapped region
and the size of the whole fingerprint region) varies significantly between the two, as
indicated by the corresponding histograms (Fig. 2.1).
2.3 The Vlatacom Dataset
This section describes in detail the processes followed by the authors for the creation
and organization of the Vlatacom dataset (VLD) [12].
2.3.1 Motivation
The decision to create the VLD stemmed in great part from the realization
that existing datasets for research on overlapped fingerprints have shown some
limitations, among them [12]:
• Both Tsinghua SOF and Tsinghua OLF datasets contain a small number of
samples: only 100 overlapped fingerprints each.
• The Tsinghua SOF dataset consists of synthetically overlapped images of good
quality whereas real latents contain noise caused by forensic dust.
• The Tsinghua SOF dataset does not include variations in component fingerprints’
rotation angles.
• Both datasets lack information about the samples’ complexity factors, such as the
number of singular points (and their location) and the image’s overall brightness.
2.3 The Vlatacom Dataset 13
The creation of the Vlatacom dataset was motivated by the overlapped fingerprint
separation research community’s need for large and comprehensive overlapped
fingerprint dataset, which has enough samples to be used for training, validation,
and testing purposes, and overcomes the aforementioned limitations.
2.3.2 Requirements
The Vlatacom dataset was created with the following requirements in mind [12]:
• Size: 10,000 or more images, to enable a broad range of combinations for
training, cross-validation, and test partitions.
• Variability: images should contain a diverse range of variation in parameters
such as amount of noise, angle of rotation of individual fingerprints, and overall
brightness.
• Access: the dataset should be made free and publicly available for academic
research.
• Additional features: besides the images themselves, the dataset should provide
supplemental information about the number and the position of the singular
points contained within the overlapped region, as a metric for distinguishing
between “easy” and “hard” samples—in general, the more singular points are
present in the overlapped area, the harder the separation problem.7
2.3.3 Dataset Structure and Organization
The process of building the Vlatacom dataset started from 200 fingerprint images
from a subset of the Db1 fingerprint dataset from FVC2002 [6]. Each image is
388 × 374 pixels (500 ppi8 ), grayscale, and contains single fingerprints. We chose
to use impressions no. 3 and no. 4 of the Db1_a dataset (200 images) because these
impressions contain differences in fingerprint rotation.
We implemented the following transformations on the images9 [12]:
• Rotation (3 variations): each individual fingerprint image is rotated by one of the
following angles: 0◦ , 45◦ , or 90◦ .
7 Singular points are points where fingerprint ridges show discontinuity. Since overlapped finger-
print separation approaches usually rely, at least partially, on the continuity of ridges, they are more
prone to errors when singular points appear in the overlapped area [11].
8 500 ppi is a standard resolution for all latent fingerprint datasets.
9 The MATLAB code for producing such variations is available upon request.
Fig. 2.2 Examples of

overlapped fingerprint images
from the Vlatacom dataset: (l)
without noise, (r) with added
Gaussian noise that resembles
the noise from forensic dust,
usually present in latent
images
• Noise (2 variations): each overlapped fingerprint image is available in two

variations: (1) noiseless; and (2) with added Gaussian noise that resembles the
noise from forensic dust, usually present in latent images (Fig. 2.2).
• Overall brightness (2 variations): each overlapped fingerprint image is available
in brighter and darker versions. Since we have no control over the original
brightness of each original component image, we use the histogram matching
technique as follows: before overlapping, we compute the average brightness of
each component image, and use the brighter (cf. darker) image histogram as a
reference against which the histogram of the other component image must be
matched.
The Vlatacom dataset is organized in four main folders, as follows [12]:
• overlap_nn: This folder contains 60,000 overlapped fingerprints without back-
ground noise. It is organized into three subdirectories (overlap_00, overlap_45,
and overlap_90), corresponding to three different rotation angles between com-
ponent images (0◦ , 45◦ , and 90◦ , respectively). Each of these folders is further
divided in two subfolders (simply named ‘1’, and ‘2’), corresponding to the
brighter or darker version of each overlapped image, respectively.
• overlap_wn: This folder contains 60,000 overlapped fingerprints with back-
ground (Gaussian noise) noise. Its subdirectories follow the same convention as
the overlap_nn folder, namely three angles of rotation and—for each of them—
two levels of average brightness.
• mask: This folder contains 60,000 manually marked binary mask images
(Fig. 2.3) for the overlapped fingerprints (two masks per image).10 It is orga-
nized into three subfolders (overlap_m_00, overlap_m_45, and overlap_m_90)
according to the rotation angle between component images (0◦ , 45◦ , and 90◦ ,
respectively).
• template: The first impressions of 100 fingers (used for creating the overlapped
fingerprints) in FVC2002 Db1_a (Fig. 2.4).
10 Note that average brightness and presence or absence of noise have no impact on the masks.
Fig. 2.3 Examples of mask images for the overlapped fingerprint images from Fig. 2.2
Fig. 2.4 Examples of templates for the overlapped fingerprint images from Fig. 2.2
The main motivation behind the naming convention adopted in the Vlatacom
dataset was to assign to each sample a name that uniquely identifies it and
encapsulates as much data about the included variations as possible. Hence, the
filename convention concatenates the title of the first and the second compo-
nent fingerprints, information about noise, amount of rotation, and brightness,
and—as a bonus, for the convenience of quick reference—the number of sin-
gular points within the overlapped region.11 For example, a file whose name is
image_3_image_4_nn_90_2_06 is a combination of image_3 and image_4
from the Db1 fingerprint dataset from FVC2002 [6], without noise, with a total of
90◦ rotation (i.e., image_3 was rotated by −45◦ and image_4 was rotated by 45◦ ).
It is the darker of the two variants, and it contains six singular points within the
overlapped area.
Singular points are points where fingerprint ridges show discontinuity. Fin-
gerprint separation approaches based on the continuity of ridges can show poor
11 Please refer to the README file for more details about the naming convention used in the
dataset, including the naming convention for the mask images.
Fig. 2.5 Examples of singular points locations on the overlapped fingerprint images from (left to
right) Vlatacom dataset, SOF dataset and OLF dataset
performance when singular points appear within the overlapped area—especially at

the very edge of the overlapped area—because it is quite possible for some error
propagation to occur, causing unreliable separation results. Figure 2.5 shows typical
overlapped fingerprints examples, with marked singular points.
2.3.4 Additional Aspects
These are some additional aspects behind the creation of the Vlatacom dataset that
are worth mentioning [12]:
• The Vlatacom dataset was designed to provide a broad range of overlapping
ratios, thereby circumventing a limitation of both Tsinghua SOF and OLF
datasets, mentioned in Sect. 2.2. In contrast to the histograms shown in Fig. 2.1,
the proposed dataset shows a Gaussian-like distribution—see, for example, the
histogram of the size of the overlapped region (overlapping ratios) for the subset
with 90◦ rotation between component images, presented in Fig. 2.6.
• In response to a requirement listed in Sect. 2.3.2, the images in the Vlatacom
dataset contain information about the number of the singular points in the
overlapped region (encoded as the last two digits in the name of each sample).
The vast majority of the images in the dataset contain between 2 and 3 singular
points in the overlapped region.
• The Vlatacom dataset also includes three CSV files with information about
the position (x and y coordinates) and type of singular points (delta or core)
contained in the overlapped region: SP_00.csv, SP_45.csv, SP_90.csv, for the
subsets with 0◦ , 45◦ , and 90◦ rotation between component images, respectively.
Fig. 2.6 Overlapping ratios histogram for the Vlatacom dataset (adapted from [12])
2.3.5 Experiments and Results
Preliminary experiments, originally reported in [12], have demonstrated the fea-

sibility of the proposed dataset for scientific research in the field of overlapped
fingerprint separation. The performance of the algorithm described in [11] was
evaluated for the following four datasets: (1) Tsinghua SOF dataset; (2) Tsinghua
OLF dataset; (3) Vlatacom dataset variant A: 90◦ rotation, lighter brightness,
without noise; and (4) Vlatacom dataset variant B: 90◦ rotation, lighter brightness,
with noise.
These experiments include matching of separated component fingerprints with
the corresponding template fingerprints using a commercial fingerprint matcher
(VeriFinger 6.7 SDK) and plotting of the corresponding receiver operating char-
acteristic (ROC) and cumulative match characteristic (CMC) curves. The ROC
curve plots the true acceptance rate (TAR) against the false acceptance rate (FAR)
for different possible FAR values. It is a widely used measure of verification
performance, based on aggregate statistics of match scores corresponding to all
biometric samples [1]. The CMC curve measures identification performance based
on the relative ordering of match scores corresponding to each biometric sample [1].
We followed the same approach as [13] to plot the CMC curves.
Fig. 2.7 CMC curve comparing the approach described in [11] across four datasets (adapted
from [12])
The comparative CMC and ROC curves are shown in Figs. 2.7 and 2.8, respec-
tively. They demonstrate that the two variants of the proposed dataset offer “harder”
test cases for existing algorithms, which is a welcome change,12 especially in the
case of the CMC curves, whose values for the Tsinghua OLF and SOF datasets were
approaching the (illusory) perfect 100% identification rate. Moreover, as intuitively
expected, the case labeled as variant B (with noise) leads to lower performance than
variant A (without noise) [12].
In summary, the Vlatacom dataset has several distinctly unique and useful
properties, among them: larger number of images than its predecessors, meaningful
variants (with or without noise, different brightness levels, different rotation angles),
and a Gaussian-shaped distribution of overlapped ratios among its images. It offers
an easy way to use subsets of the proposed dataset for specific needs (such as
brightness/contrast correction or noise removal, for instance) and/or “levels of
difficulty” (e.g., based on the amount of overlap and/or the number of singular
points within the overlapped region). We expect that it will be widely used by the
fingerprint processing research community.
12 The transition from Tsinghua datasets to the proposed Vlatacom dataset mirrors, somehow, the
creation of the Caltech 256 object category dataset (to replace its predecessor, Caltech 101) in the
field of object recognition [4].
2.4 Concluding Remarks 19
Fig. 2.8 ROC curve comparing the approach described in [11] across four datasets (adapted
from [12])
In this chapter we introduced publicly available datasets in the field of overlapped

latent fingerprints and discussed their essential role in benchmarking and evaluating
algorithms in the field of fingerprint processing and matching. Until recently, there
were two main datasets of overlapped fingerprints in use: Tsinghua OLF and
Tsinghua SOF. We have highlighted some of their limitations and have shown that
a new dataset, called the Vlatacom dataset (VLD), was recently created to overcome
them.

• Every field of scientific research requires datasets that can be used to
evaluate new approaches to a problem, benchmark solutions, and ensure
research reproducibility.
• The field of overlapped latent fingerprints research still lacks large publicly
available databases collected in real-life conditions.
(continued)
• There are many challenges associated with the creation of a large latent
fingerprint dataset, among them: the need for professional expertise, large
amounts of time, proper equipment, trained personnel, and a standardized
procedure for data collection, annotation, and organization.
• A newly released, publicly available, and more comprehensive dataset
for overlapped fingerprint processing—the Vlatacom dataset (VLD)—is
described.
• It has been experimentally confirmed that the VLD dataset can be used
to develop and evaluate fingerprinting methods including segmentation,
separation, and other related processes.
References
1. B. DeCann, A. Ross, Relating ROC and CMC curves via the biometric menagerie, in 2013
IEEE Sixth International Conference on Biometrics: Theory, Applications and Systems (BTAS)
(2013), pp. 1–8
4. G. Griffin, A. Holub, P. Perona, Caltech-256 Object Category Dataset (2007)
5. D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, A.K. Jain, FVC2000: fingerprint verification
competition. IEEE Trans. Pattern Anal. Mach. Intell. 24(3), 402–412 (2002)
6. D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, A.K. Jain, FVC2002: Second fingerprint
verification competition, in Proceedings of the 16th International Conference on Pattern
Recognition, 2002, vol. 3 (IEEE, Piscataway, 2002), pp. 811–814
7. A. Sankaran, T.I. Dhamecha, M. Vatsa, R. Singh, On matching latent to latent fingerprints, in
International Joint Conference on Biometrics (IJCB), 2011 (IEEE, Piscataway, 2011), pp. 1–6
8. A. Sankaran, M. Vatsa, R. Singh, Hierarchical fusion for matching simultaneous latent
fingerprint, in IEEE Fifth International Conference on Biometrics: Theory, Applications and
Systems (BTAS), 2012 (IEEE, Piscataway, 2012), pp. 377–382
9. A. Sankaran, M. Vatsa, R. Singh, Latent fingerprint matching: a survey. IEEE Access 2, 982–
1004 (2014)
10. B. Stojanović, O. Marques, A. Nešković, Latent overlapped fingerprint separation: A review.
overlapped fingerprints separation. Multimed. Tools Appl. 76(10), 1–25 (2016)
12. B. Stojanovic, O. Marques, A. Neskovic, A novel synthetic dataset for research in overlapped
fingerprint separation, in Proceedings of the IPTA, Seventh International Conference on Image
Processing Theory, Tools and Applications (IPTA2017) (2017)
13. N. Zhang, Y. Zang, X. Yang, X. Jia, J. Tian, Adaptive orientation model fitting for latent
overlapped fingerprints separation. IEEE Trans. Inf. Forensics Secur. 9(10), 1547–1556 (2014)
Chapter 3
Overlapped Latent Fingerprints
Segmentation: Problem Definition
Abstract This chapter describes the problem of segmentation of overlapped

fingerprints, which is a required prerequisite step in the fingerprint processing
pipeline, performed before the processes of fingerprint separation and subsequent
verification. Overlapped fingerprints segmentation is usually performed manually,
and only recently there have been (semi-)automatic approaches proposed in the
literature. The evaluation procedure to assess the quality of such approaches is also
discussed.
3.1 Introduction
Overlapped fingerprints can be found in latent fingerprints lifted from crime scenes
as well as in live-scan fingerprint images, when the surface of fingerprint sensors
contains residues of fingerprints of previous users. Overlapped fingerprints must be
extracted out of the image (i.e., segmented from the background) and separated from
one another before they can be processed by contemporary commercial fingerprint
matchers (Fig. 3.1).
The overlapped fingerprints matching procedure typically consists of two
steps [14]:
1. Segmentation of component fingerprints regions, in order to distinguish the
region of the image that belongs to both component fingerprints (overlapped
region, i.e. the region of the image where component fingerprints are overlapped)
from the regions of the image belong to separate component fingerprints (single
regions, i.e. regions where only one fingerprint is present).
2. Separation of the resulting overlapped region into component fingerprint images,
a process by which individual pixels can be labeled as belonging to the ridges of
either overlapped fingerprint, which will be described in Chap. 5.
After these steps, each of the separated component fingerprints can be processed
by a typical automated fingerprints matching system.

22 3 Overlapped Latent Fingerprints Segmentation: Problem Definition
Fig. 3.1 Typical overlapped latent fingerprints matching procedure. See text for explanation
Figure 3.2 presents an overview of the overlapped latent fingerprints segmenta-

tion process, where the input is an image containing overlapped fingerprints and
the outputs are region masks corresponding to pixels that belong to each individual
fingerprint (M1S , M2S ) or the overlapped region between them (MO ).
The topic of single latent fingerprints segmentation has received a lot of attention
in recent years, and many successful approaches have emerged as a result of these
efforts [3, 4, 9, 10, 12, 15, 18–21]. The segmentation of overlapped fingerprints,
however, remains a technically challenging problem, usually performed manually
using a drawing tool (such as Photoshop or Gimp) to create (“paint”) region masks.
The resulting region masks, along with the overlapped fingerprint images, are
then used as inputs to overlapped fingerprint separation algorithms. Region masks
are an integral part of all existing overlapped fingerprints datasets (see Chap. 2),
since those masks are necessary to ensure reproducibility of results and meaningful
comparisons among competing fingerprint separation methods. The creation of such
masks is, therefore, a manual process, with only few recent attempts to develop fully
automated segmentation of overlapped fingerprints (see Chap. 4).
3.2 The Mask Segmentation Problem
The original problem of (individual) latent fingerprints segmentation comprises two

steps: (1) segmentation of the region of interest (ROI), i.e. the fingerprint area and
(2) removal of background noise from the fingerprint area.
3.2 The Mask Segmentation Problem 23
Fig. 3.2 Overlapped latent fingerprints segmentation process: the input is an image containing
two overlapping fingerprints and the outputs are region masks corresponding to pixels that belong
to each individual fingerprint (M1S , M2S ) or the overlapped region between them (MO )
Fig. 3.3 Examples of latent fingerprint images of varying quality from NIST Special Database 27
(SD-27A) [7]. (a) good, (b) bad, (c) ugly
The examples in Fig. 3.3—classified as good, bad, and ugly by the creators of the
NIST Special Database 27 (SD-27A) [7]—provide compelling examples of some of
the challenges in individual latent fingerprints segmentation.
Fig. 3.4 Examples of overlapped latent fingerprint images from different databases, with illustra-
tively marked component fingerprints. (a) NIST Special Database 27 (SD-27A) [7], (b) Tsinghua
SOF [6], (c) Tsinghua OLF [6]
Fig. 3.5 Overlapped fingerprints mask segmentation. See text for details
The segmentation problem is even more challenging when two or more latents
overlap (Fig. 3.4), since now the segmentation algorithm must—in addition to
separating ROI from the background—segment and distinguish the overlapped
region or individual component fingerprints as well.
Automatic segmentation of single latent fingerprints is a challenging problem, for
which several solutions have been proposed in recent years [1, 3, 4, 8–12, 15, 18–
21], some of which dealing with difficult databases, but none of them addressing
the segmentation of overlapped latent fingerprints, the special case where the
overlapping fingerprint area should distinct from single fingerprint area, in addition
to distinguishing between fingerprint area and background.
Segmentation of overlapped fingerprints involves the following tasks (Fig. 3.5):
(1) segmentation of fingerprint area (ROI) from the background, regardless of
which component fingerprint that area belongs to; (2a) segmentation of single (S)
3.3 Evaluation Methodology 25
and overlapped (O) regions (blocks), regardless of which component fingerprint

single area the blocks belongs to; (2b) segmentation of the individual component
fingerprints ROIs; and (3) removal of unstructured background noise, caused by
forensic dust.
Both segmentation options (Fig. 3.5, (2a) and (2b)) can be used as input for
the subsequent separation tasks, depending on the separation algorithm. The core
difficulty of overlapped fingerprints segmentation includes the fact that fingerprints
are not opaque objects (with consistent—and easily distinguishable—color/texture
patterns), for example. This is in contrast with the usual case of (natural scene)
semantic segmentation [16, 17].
Algorithms capable of performing automatic ROI extraction of component
fingerprint regions—background, single region, and overlapped region—from the
overlapped fingerprint image have just recently started to appear in the literature
[13, 14] and will be described in Chap. 4.
3.3 Evaluation Methodology
The performance evaluation of single fingerprints segmentation methods is usually

done indirectly, by computing metrics related to fingerprints matching performance,
under the following rationale: the better the performance of the fingerprints
matching algorithm, the better (presumably) the segmentation method. There are
several problems associated with relying on indirect indicators of segmentation
quality, among them the fact that the performance of the segmentation algorithm
can be masked by the quality of the fingerprint matcher [13].
In order to evaluate the quality of the segmentation approach directly, one
can treat the segmentation problems as binary classification problems, and adopt
commonly used figures of merit for the performance evaluation of binary classifiers,
such as confusion matrix (Fig. 3.6) and ROC (receiver operating characteristic)
curve, which is a widely used measure of performance of supervised classification
rules [2, 13].
The metrics used to compare the automatically segmented fingerprint regions
against manually created masks, considered as the ground truth, are [14]:
• T P R—True Positive Rate, also known as T AR (True Acceptance Rate),
sensitivity, hit rate, or recall, depending on the context (ideal value = 1);
• F P R—False Positive Rate, equivalent to F AR (False Acceptance Rate) or fall-
out (ideal value = 0);
• ACC—Accuracy (ideal value = 1).
• P P V —Positive Predictive Value, equivalent to precision (ideal value = 1).
• F 1 score—The harmonic mean of precision and recall (ideal value = 1).
• AU C—Area Under the ROC Curve (ideal value = 1).
They are mathematically defined as follows:
Fig. 3.6 Typical confusion

matrix
TP
TPR = (3.1)
T P + FN
FP
FPR = (3.2)
FP + T N
TP +TN
ACC = (3.3)
T P + FP + FN + T N
TP
PPV = (3.4)
T P + FP
2T P
F1 = (3.5)
2T P + F P + F N
where: TP, TN, FP, and FN stand for true positives, true negatives, false positives,
and false negatives, respectively.
The ROC curve plots TPR against FPR and provides a graphical representation of
the relative tradeoffs between true positives and false positives [5]. The area under
the ROC curve (AUC) represents the probability that a randomly chosen positive
example is correctly ranked with greater suspicion than a randomly chosen negative
example [2].
The special case of overlapped fingerprints can be represented as a multi-class
classification problem. Image blocks/pixels should be classified in one of three
classes—background (B), single (S), and overlapped (O). The evaluation of such a
3.4 Concluding Remarks 27
Fig. 3.7 Evaluation cases. (a) Overlapped fingerprints sample, (b) Segmentation
system can be performed in two evaluation cases, which are essentially two separate
binary classification problems [14]:
• Region of Interest segmentation, in which we measure the algorithm’s ability
to segment both fingerprints (painted white and gray in Fig. 3.7b) from the
background (painted black in Fig. 3.7b);
• Overlapped region segmentation, in which the area of overlap (painted white in
Fig. 3.7b) is segmented from the remaining of the ROI (painted gray in Fig. 3.7b).
In this chapter we discussed the problem of latent fingerprints segmentation, with

particular reference to overlapped latent fingerprints, since this step is performed
manually in the most of the published approaches. We also introduced and explained
the typical evaluation methodology and associated figures of merit.

• The process of individual latent fingerprints segmentation consists of
outlining the region of interest and removing any kind of background noise.
• Segmentation of overlapped latent fingerprints requires an additional step,
namely the segmentation of different component fingerprints regions.
• Overlapped latent fingerprints segmentation is still performed manually
in most of the published overlapped fingerprint separation and matching
approaches, which is a way such techniques are referred to as semi-
automated.
• Overlapped latent fingerprints segmentation remains an open, challenging,
and active research area.
References
1. W. Bian, S. Ding, W. Jia, Collaborative filtering model for enhancing fingerprint image. IET
Image Process. 12(1), 149–157 (2017)
2. A.P. Bradley, The use of the area under the ROC curve in the evaluation of machine learning
algorithms. Pattern Recogn. 30(7), 1145–1159 (1997)
3. K. Cao, E. Liu, A. Jain, Segmentation and enhancement of latent fingerprints: a coarse to fine
ridgestructure dictionary. IEEE Trans. Pattern Anal. Mach. Intell. 36(9), 1847–1859 (2014)
4. H. Choi, M. Boaventura, I.A. Boaventura, A.K. Jain, Automatic segmentation of latent
fingerprints, in IEEE Fifth International Conference on Biometrics: Theory, Applications and
Systems (BTAS) 2012 (IEEE, Piscataway, 2012), pp. 303–310
5. T. Fawcett, An introduction to ROC analysis. Pattern Recogn. Lett. 27(8), 861–874 (2006)
8. M. Ghafoor, I.A. Taj, W. Ahmad, N.M. Jafri, Efficient 2-fold contextual filtering approach for
fingerprint enhancement. IET Image Process. 8(7), 417–425 (2014)
9. S. Karimi-Ashtiani, C.-C.J. Kuo, A robust technique for latent fingerprint image segmentation
and enhancement, in 15th IEEE International Conference on Image Processing, 2008. ICIP
2008 (IEEE, Piscataway, 2008), pp. 1492–1495
10. A. Sankaran, A. Jain, T. Vashisth, M. Vatsa, R. Singh, Adaptive latent fingerprint segmentation
using feature selection and random decision forest classification. Inf. Fusion 34, 1–15 (2017)
11. P. Schuch, S. Schulz, C. Busch, Survey on the impact of fingerprint image enhancement. IET
Biom. 7(2), 102–115 (2017)
12. N.J. Short, M.S. Hsiao, A.L. Abbott, E.A. Fox, Latent fingerprint segmentation using ridge
template correlation, in 4th International Conference on Imaging for Crime Detection and
Prevention 2011 (ICDP 2011) (IET, Stevenage, 2011), pp. 1–6
13. B. Stojanović, O. Marques, A. Nešković, Latent overlapped fingerprint separation: a review.
14. B. Stojanović, O. Marques, A. Nešković, Deep learning-based approach to latent overlapped
fingerprints mask segmentation. IET Image Process. 12(11), 1934–1942 (2018)
15. D.H. Thai, C. Gottschlich, Global variational method for fingerprint segmentation by three-part
decomposition. IET Biom. 5(2), 120–130 (2016)
16. Q. Wang, J. Gao, Y. Yuan, Embedding structured contour and location prior in Siamesed fully
convolutional networks for road detection. IEEE Trans. Intell. Transp. Syst. 19(1), 230–241
(2017)
17. Q. Wang, J. Gao, Y. Yuan, A joint convolutional neural networks and context transfer for street
scenes labeling. IEEE Trans. Intell. Transp. Syst. 19(5), 1457–1470 (2017)
18. X. Yang, J. Feng, J. Zhou, S. Xia, Detection and segmentation of latent fingerprints, in
2015 IEEE International Workshop on Information Forensics and Security (WIFS) (IEEE,
19. J. Zhang, R. Lai, C.-C.J. Kuo, Latent fingerprint segmentation with adaptive total variation
model, in 2012 5th IAPR International Conference on Biometrics (ICB) (IEEE, Piscataway,
2012), pp. 189–195
20. J. Zhang, R. Lai, C.-J. Kuo, Latent fingerprint detection and segmentation with a directional
total variation model, in 2012 19th IEEE International Conference on Image Processing (ICIP)
(IEEE, Piscataway, 2012), pp. 1145–1148
21. J. Zhang, R. Lai, C.-C.J. Kuo, Adaptive directional total-variation model for latent fingerprint
segmentation. IEEE Trans. Inf. Forensics Secur. 8(8), 1261–1273 (2013)
Chapter 4
Machine Learning Based Segmentation
of Overlapped Latent Fingerprints
Abstract This chapter describes a convolutional neural network (CNN)-based

approach for overlapped fingerprint mask segmentation. The CNN classifies each
image block within the overlapped fingerprint image into three classes—background
(B), single fingerprint (S), and overlapped fingerprint (O). The proposed segmenta-
tion method has been successfully tested on three different datasets.
4.1 Introduction
This chapter presents a machine learning based approach for overlapped fingerprint
mask segmentation. As discussed in Chap. 3, the segmentation of overlapped
fingerprints (and creation of associated masks) is usually performed manually, with
only few recent attempts to develop fully automated segmentation of overlapped
fingerprints.
The method presented in this chapter (and described in greater detail in [7]) is
based on previously developed neural network-based methods for single fingerprint
image ROI segmentation [4, 5], which were trained and tested on fingerprint images
of good quality (i.e., not latents).
The proposed method is based on convolutional neural networks (CNNs) [2, 3],
which—in addition to an ever-growing number of applications in computer vision—
have also shown good performance in single fingerprints processing, for different
applications [1, 6, 8].
4.2 Description of the Algorithm
The CNN-based method for overlapped fingerprints segmentation—represented

schematically in Fig. 4.1—consists of the following steps [7]:
1. The fingerprint image is divided into overlapping blocks.
2. The blocks are sequentially presented to the input layer of the CNN.

30 4 Machine Learning Based Segmentation of Overlapped Latent Fingerprints
Fig. 4.1 Overlapped fingerprints mask segmentation based on convolutional neural networks
3. The CNN classifies each block in one of three classes—background (B), single
(S), or overlapped (O) region.
4. Results for each block are finally merged into a region mask.
The proposed method uses a simplified AlexNet (SAlNet) CNN architecture,
designed to determine the orientation field of latent fingerprints [1]. The optimal
block size (40 × 40 pixels) was determined experimentally (see [7] for details).
The VLD dataset (described in Chap. 2) was chosen for algorithm training
purposes. Additionally, the Tsinghua SOF and OLF datasets (also described in
Chap. 2) were used for algorithm verification purposes.
4.2.1 CNN Architecture
The SAlNet CNN architecture (Fig. 4.2) consists of three convolutional layers
(conv1–3) and one fully connected layer (fc4). The input of the CNN is a fingerprint
patch of size 40 × 40 pixels. The first layer (conv1) contains convolutional sub-
layer with 96 filters of size 11 × 11 and stride of 4. The output of the neurons are
modeled by the ramp activation function (ReLU), f (x) = max(0, x). This layer
also contains max-pooling sub-layer, which takes the maximum over 3 × 3 spatial
neighborhoods with a stride of 2. The third sub-layer of conv1 is a local response
normalization. Similar definitions are used for layers conv2 and conv3. The fully
connected layer (fc4) is applied after the last convolutional layer (conv3). A dropout
regularization (with probability of 0.5) is performed after fc4 layer, in order to avoid
overfitting [7].
4.2 Description of the Algorithm 31
256 256
128 4¥4
40x40 96 5¥5
11¥1
3
-- Background
-- Single
-- Overlapped
classes
conv1
conv2 fc4
conv3
Input Convolution Convolution Convolution
Max pooling Max pooling ReLU
Batch norm. Batch norm. Fully
ReLU ReLU connected
Hidden layers Classification
Fig. 4.2 SAlNet architecture. See text for details
4.2.2 CNN Training
Two separate CNNs are involved: (1) CNN-SO: a network designed for segmen-
tation of synthetically overlapped fingerprints (Tsinghua SOF dataset) and (2)
CNN-OL: a network designed for segmentation of latent overlapped fingerprints
(Tsinghua OLF dataset).
For CNN-SO training purposes, 30,000 overlapped fingerprint images’ blocks
were randomly chosen from synthetically overlapped fingerprints (500 ppi) con-
tained in the VLD dataset. As described in Chap. 2, the VLD dataset contains
significantly larger amount of fingerprints and provides additional variations com-
pared to the Tsinghua SOF dataset in terms of overlapping parameters, as well
as the availability of samples containing Gaussian noise. Training blocks were
selected such that half of the samples contain Gaussian noise, and all three classes
(background, single, and overlapped region) were equally presented [7].
For CNN-OL training purposes, 30,000 overlapped fingerprint images’ blocks
were randomly chosen from the 60 latent overlapped fingerprints, created for this
purpose in a similar manner as the prints from Tsinghua OLF dataset, using forensic
methods and using different component fingerprints. Training blocks were also
selected in a manner that all three classes (background, single, and overlapped
region) are equally presented [7].
The 30,000 blocks from the training dataset were randomly divided into two
subsets according to the following percentages: (1) 60% of a dataset was used for
training and (2) 40% of the dataset was used for validation. In order to avoid any
correlation between training and testing datasets, testing of the CNN-SO and CNN-
OL was performed on the Tsinghua OLF and SOF datasets, created from different
component fingerprints comparing to VLD dataset, and without any correlation
between them [7].
4.3 Experimental Evaluation
Two different CNNs were trained [7]:

• CNN-OL, designed for segmentation of latent overlapped fingerprints (with
noisy background). This CNN can be used in investigative applications without
additional network training, where samples usually contain noise from forensic
dust residues.
• CNN-SO, designed for segmentation of synthetically overlapped fingerprints
(without background noise). This CNN can be used in civil applications, where
the source of overlapped images is a fingerprint scanner with residues of the
previous subject fingerprints on its scanning surface.
Figures 4.3 and 4.4 show examples of the CNN-OL and the CNN-SO in action,
respectively.
Both networks, CNN-OL and CNN-SO, were tested extensively and demon-
strated convincing numerical results, substantially better than previously published
methods (see [7] for details). For an example of the occasionally produced incorrect
results, see Figs. 4.5 and 4.6. The CNN-OL failure case (Fig. 4.5) shows how
segmentation becomes more difficult when the input image’s fingerprint ridges are
not clear, and both single and overlapped area contains smudges. The smudgy area
is incorrectly detected by the network as overlapped area. The CNN-SO failure
case (Fig. 4.6) shows how segmentation can be problematic when the input image
contains one or more areas of poor quality, in this case, that fingerprint ridges are
not continuous lines but more dot-like. The area with dots is incorrectly detected by
the network as the overlapped area [7].
Fig. 4.3 CNN-OL segmentation results examples 1 (top) and 2 (bottom): input image (left),
generated mask (center), and masked image (right)
4.3 Experimental Evaluation 33
Fig. 4.4 CNN-SO segmentation results examples 1 (top) and 2 (bottom): input image (left),
generated mask (center), and masked image (right)
Fig. 4.5 Typical segmentation failure cases. CNN-OL exemplar (left), corresponding mask
(center), and masked fingerprints (right)
Fig. 4.6 Typical segmentation failure cases. CNN-SO exemplar (left), corresponding mask
(center), and masked fingerprints (right)
In this chapter we discussed a new method for segmentation of overlapped

fingerprints based on machine learning techniques, specifically convolutional neural
networks (CNNs). The proposed method gives good results and opens the field for
new research and development of fully automated methods for the separation of
overlapped fingerprints.

• The recent popularity of machine learning and deep learning methods for
image analysis tasks has opened new avenues for the development of a
fully automated algorithm for separating overlapped fingerprints.
• The overlapped fingerprints separation algorithms published so far include
manual segmentation of the component fingerprints region.
• The proposed CNN-based method represents a significant advance towards
the development of a fully automated system for overlapped fingerprints
separation.
References
1. K. Cao, A. Jain, Latent orientation field estimation via convolutional neural network, in 2015
International Conference on Biometrics (ICB) (2015), pp. 349–356. https://doi.org/10.1109/
ICB.2015.7139060
2. I. Goodfellow, Y. Bengio, A. Courville, Deep Learning (MIT Press, Cambridge, 2016)
3. Y. LeCun, Y. Bengio, G. Hinton, Deep learning. Nature 521(7553), 436–444 (2015)
4. B. Stojanovic, A. Neskovic, Z. Popovic, V. Lukic, ANN based fingerprint image ROI segmenta-
tion, in 2014 22nd Telecommunications Forum Telfor (TELFOR) (IEEE, Piscataway, 2014), pp.
505–508
5. B. Stojanović, A. Nešković, O. Marques, Fingerprint ROI segmentation using Fourier coeffi-
cients and neural networks, in 2015 23rd Telecommunications Forum Telfor (TELFOR) (IEEE,
6. B. Stojanović, O. Marques, A. Nešković, S. Puzović, Fingerprint ROI segmentation based on
deep learning, in 2016 24th Telecommunications Forum (TELFOR) (IEEE, Piscataway, 2016),
pp. 1–4
7. B. Stojanović, O. Marques, A. Nešković, Deep learning-based approach to latent overlapped
fingerprints mask segmentation. IET Image Process. 12(11), 1934–1942 (2018). https://doi.org/
10.1049/iet-ipr.2017.1227
8. R. Wang, C. Han, Y. Wu, T. Guo, Fingerprint Classification Based on Depth Neural Network
(2014). ArXiv:1409.5188
Chapter 5
Separation: Problem Definition
Abstract Overlapped fingerprints are often found in latent fingerprints lifted from
crime scenes and in live-scan fingerprint images when the surface of fingerprint sen-
sors contains residues of fingerprints of previous users. Such overlapped fingerprints
must be separated before they can be processed by a fingerprint matcher, which has
led to the creation of several different methods designed to separate the overlapped
fingerprints. This chapter describes the problem of latent overlapped fingerprint
separation and presents a brief overview of selected contemporary techniques for
overlapped fingerprint separation in the context of latent overlapped fingerprints
matching.
5.1 Introduction
The problem of overlapped latent fingerprint separation consists in separating and

enhancing component fingerprints in order to make them suitable candidates for
later feature extraction and matching procedures. This process usually includes four
major steps [3, 8, 27, 30, 35] (Fig. 5.1):
• Segmentation—This step involves latent fingerprints segmentation (also known
as ROI extraction), whose output produces the component fingerprints’ region
masks. See Chaps. 3 and 4 for details.
• Initial orientation field estimation and enhancement—The initial orientation
field (which consists of a matrix containing information about ridge angle in
every pixel/block on fingerprint image) for the overlapped fingerprint image
is usually estimated and enhanced using a block-based approach. This step
produces a labeled image containing three different regions: (1) background
• Orientation separation—Mixed orientation fields of overlapped regions for two
component fingerprints are separated, smoothed, and enhanced in order to correct

36 5 Overlapped Latent Fingerprints Separation: Problem Definition
Fig. 5.1 Overlapped latent fingerprint enhancement: a block diagram
remaining errors. The output of this step are component fingerprints’ orientation
fields.
• Component fingerprint images extraction—Two component fingerprints are
In some approaches (for example, [36]), the second and third steps are combined,
i.e. there is no initial orientation field estimation and separation, but orientation
fields of separate component fingerprints are estimated in one step.
5.2 Orientation Field Estimation
A fingerprint orientation field is a matrix, whose value at (x, y) denotes the

dominant ridge orientation at point (x, y) [20]. The main difference between the
orientation field of an overlapped fingerprint image and the orientation field of a
single fingerprint image is that the former contains one dominant orientation in the
nonoverlapped regions and two dominant orientations in the overlapped region. The
state of the art in overlapped fingerprint separation relies on manually marked region
masks outlining the component fingerprints; hence, the overlapped region can be
obtained by finding the intersection of two fingerprint region masks [3].
The problem of initial orientation field estimation consists of estimating one
dominant ridge orientation in the nonoverlapped fingerprint regions and two
dominant ridge orientations in the overlapped region. The initial orientation field,
together with region masks, will then be used as input by the subsequent orientation
field separating algorithm [3].
5.2 Orientation Field Estimation 37
Traditional orientation field estimation algorithms consist of two steps: (1)

initial estimation (i.e., using a gradient-based method) and (2) orientation field
regularization. Regularization may be done by a simple averaging filter or global
model-based methods. However, for overlapped fingerprints containing ridges of
two different orientations in the overlapped area, the initial orientation field obtained
by gradient-based methods may be a random mix of the orientation fields of the
two component fingerprints which cannot be resolved by existing regularization
algorithms [3].
5.2.1 Orientation Field Estimation Algorithms for Single

Fingerprint
Orientation field estimation algorithms can be classified into three categories [9].
1. Local Estimation
Local estimation approaches compute a local ridge orientation at pixel x =
(x, y) using only the neighborhood around x, which is typically 32 × 32 pixels
for 500 ppi fingerprints.
Widely used local estimation approaches include:
• Gradient-based approaches [1, 2, 17, 25], which use classical gradient opera-
tors, such as Prewitt or Sobel, to compute the dominant orientation in the local
neighborhood.
• Slit-based approach [23], which explicitly exploits the fact that the variation
of intensity is the smallest along the ridge orientation and largest along
the orthogonal orientation. By testing such a hypothesis along a number of
different orientations, the best orientation is chosen [9].
• Fourier transform-based approach, which approximates the ridge pattern in
a local area of a fingerprint by a 2D sine wave, whose Fourier transform’s
magnitude spectrum will contain a pair of peaks whose location corresponds
to the parameters of the sine wave. The magnitude spectrum can then be
mapped to the polar coordinate system and normalized; the resulting nor-
malized magnitude spectrum can be viewed as a probability distribution [4].
The best orientation can then be estimated as the most probable orientation
according to the resulting distribution [9].
These approaches perform very well for fingerprint images of good quality,
but their performance is quite poor for latent images that contain structural
background noise [29].
2. Smoothing
Orientation fields obtained by local estimation approaches for poor quality
fingerprints are usually very noisy and need to be regularized, which can be
achieved using an image smoothing algorithm, such as the well-known low-pass
filtering [1]. Low-pass filtering results depend critically on the size of the filtering
window: larger windows are better at reducing noise whereas smaller windows
are capable of preserving the true orientation in a high curvature region [9].
Many alternatives to the standard low-pass filter have been proposed in the
literature, such as the implementation of orientation field smoothing by using the
Markov random field (MRF) model [5, 16, 24] and the use of multiresolution
orientation fields [18, 19, 22, 23]. Fingerprints of very poor quality still offer a
significant challenge to any of these approaches [9].
3. Global Parametric Models
Global parametric model fitting methods constitute another category of
regularization algorithms that use mathematical models to represent the whole
fingerprint orientation field. The models can be general (e.g., polynomials [12]
and Fourier series [31], or specific to fingerprints [14, 26, 37]. General models
may be prone to overfitting or underfitting problems, especially when the initial
orientation field is very noisy [9].
There are models which explicitly consider singular points [14, 26, 37], which
cannot be easily extracted from latents. Consequently, certain orientation field
estimation approaches in the literature (e.g., [32] and [33]) require manually
marked singular points as input [9].
5.2.2 Orientation Field Estimation Algorithms for Overlapped

Fingerprints
Orientation field estimation for the overlapped region is a very challenging problem
for traditional fingerprint orientation estimation methods [8]. In overlapped latent
fingerprint processing approaches most authors [3, 7, 27, 30] have adopted the local
Fourier analysis method proposed in [15] for initial orientation field estimation.
The process of orientation field estimation for the overlapped region consists of
the following steps (Fig. 5.2) [29]:
1. The fingerprint image is divided into non-overlapping image segments (blocks),
typically of size 16 × 16 pixels.
2. Fourier coefficients are calculated for these segments, as follows [15]:
(a) The local window of size 64 × 64 pixels, centered at each block, is multiplied
by a bivariate isotropic Gaussian function;
(b) For the resulting image, the discrete Fourier transform (DFT) is calculated;
(c) The amplitude of low-frequency components (points within three pixels from
the center in the frequency domain) is set to zero.
Local maxima points in the frequency domain correspond to 2D sine waves
(stripes) on the image block. For single region blocks (S blocks) one estimates
orientation as the orientation (direction) of the 2D sine wave that corresponds to
the brightest local maxima. For overlapped region blocks (O blocks) one estimates
two orientations as the orientations of the 2D sine waves that correspond to the two
brightest local maxima.
5.3 Overlapped Latent Fingerprint Separation: Representative Approaches 39
Fig. 5.2 Initial orientation estimation for one overlapped region block. (a) Overlapped (O) block;
(b) O block multiplied by Gaussian function; (c) two local maxima points in the frequency domain;
(d) two dominant orientations; (e, f) corresponding 2D sine waves (Adapted from [30])
Besides the local Fourier analysis method, there are other noteworthy approaches
in literature. Zhao and Jain in [36] proposed reconstruction of the orientation
fields of component fingerprints via modeling orientation fields and then predicting
unknown orientation fields based on a certain number of manually marked orien-
tation cues in fingerprints. Zhang et al. in [35] proposed block-based orientation
field estimation from the overlapped fingerprint image by Gabor filters approach.
In this approach K Gabor filters corresponding to K different orientations are
applied to filter the image, and the orientations are obtained according to the filters’
responses [29].
5.3 Overlapped Latent Fingerprint Separation:

Representative Approaches
Table 5.1 presents a summary of some of the most significant latent overlapped
fingerprint separation approaches proposed during the past few years. The table
contains information on the separation method, the initial orientation field estima-
tion method, the level of manual work, and whether the approach was tested on real
overlapped latents (Tsinghua OLF database) or not.1
1 For additional information on each method listed in Table 5.1, please see [29] or refer to the
original publications.
40
Table 5.1 Overlapped latent fingerprint separation approaches (adapted from [30])
Zhao and Zhang Stojanović
Approach Singh et al. [28] Chen et al. [3] Shi et al. [27] Feng et al. [8] Jain [36] et al. [34, 35] et al. [30]
Separation method Independent Relaxation Constrained Constrained Model-based Adaptive Neural networks
Component labeling relaxation relaxation separation orientation
Analysis (ICA) labeling labeling model fitting
Initial orientation n/a Local Fourier Local Fourier Local Fourier Other Other Local Fourier
estimation method analysis analysis analysis analysis
Level of manual work n/a Region masks Region masks Region masks Region masks Region masks Region masks
and singular and orientation
points cues
Tested on real overlapped No No No Yes No Yes Yes
latents (OLF)
Impact First published Additional – State-of-art for The highest level – State-of-art for
approach manual work SOF dataset of manual work OLF database
required required
5 Overlapped Latent Fingerprints Separation: Problem Definition
5.4 Evaluation Methodology 41
5.4 Evaluation Methodology
The performance of overlapped (latent) fingerprint separation approaches is usually

evaluated indirectly, based on the performance evaluation of the matching process.
Matching experiments consist of matching separated component fingerprints with
the corresponding template fingerprints and plotting of the receiver operating char-
acteristic (ROC) curves and the cumulative match characteristic (CMC) curves, as
typical biometric systems matching accuracy indicators. For matching experiments,
typically a commercial fingerprint matcher, such as the popular VeriFinger [21], is
often used [29].
The ROC curve is a widely used measure of verification performance, based on
aggregate statistics of match scores corresponding to all biometric samples [7]. It
plots the true acceptance rate (TAR) against the false acceptance rate (FAR) for
different possible FAR values. To plot the ROC curves, only genuine matches are
executed because the output scores of the VeriFinger matcher are linked to the FAR.
For each database, 200 genuine matches are executed [29].
The CMC curve measures identification performance based on the relative
ordering of match scores corresponding to each biometric sample [7]. To plot the
CMC curves, the approach used in [35] consists of 2000 rolled fingerprints in the
NIST SD4 database combined with the 12 template fingerprints in the OLF database
(10 template fingerprints in the case of the SOF database), which are used as the
background database [29].
The NIST SD4 database contains 2000 gray scale fingerprint images, of resolu-
tion 512 × 512 pixels, and is widely used for evaluating fingerprint systems on a
statistical sample of fingerprints because it is evenly distributed over the five major
classifications (Arch, Left and Right Loops, Tented Arch, Whirl) [29].
Since both the ROC (aggregate-based) and CMC (rank-based) curves are esti-
mated from the same set of match scores, they are, not surprisingly, correlated to
some degree. Recently, the notion that the ROC and CMC are directly related has
been challenged by the work of Gorodnichy who first presented an argument stating
that aggregate based metrics such as the ROC fail to appropriately evaluate opera-
tional systems characterized by large sample size and non-static populations[10, 11].
Furthermore, Gorodnichy argues that verification systems should be evaluated
(and developed) as 1:N identification systems [11], and states that “measures
for identification (i.e., ranked statistics) reveal more information regarding the
relationships between users involved in a biometric system [7]”. These arguments
suggest that CMC curves are more meaningful representations of performance
measures for biometric systems than ROC curves [29].
Figure 5.3 shows examples of ROC and CMC curves. The curves in red are from
a representative approach in the literature [30], whereas the dashed curves in blue
represent ideal values.
Fig. 5.3 Examples of ROC (left) and CMC (right) curves
In this chapter we discussed the problem of overlapped latent fingerprints separa-

tion, as the most challenging task in overlapped latent fingerprints matching process.
We paid special attention to approaches published so far and their limitations. We
also introduced and explained the typical evaluation methodology and associated
figures of merit.

• Overlapped latent fingerprints separation is the most challenging task in
the overlapped latent fingerprints matching pipeline.
• Overlapped latent fingerprints separation is still a semi-automated process
due to the adoption of manual overlapped fingerprints segmentation.
• Comparative evaluation of competing approaches is still limited by the lack
of datasets with real-life samples.
• This is an active research area, with many open problems, such as
separation of more than two overlapped fingerprints and handling other
types of (structured) background noise.
References
1. A.M. Bazen, S.H. Gerez, Systematic methods for the computation of the directional fields and
singular points of fingerprints. IEEE Trans. Pattern Anal. Mach. Intell. 24(7), 905–919 (2002)
2. J. Bigun, Optimal orientation detection of linear symmetry, in Proceedings of the IEEE First
International Conference on Computer Vision, London, June 8–11 (1987), pp. 433–438
3. F. Chen, J. Feng, A. Jain, J. Zhou, J. Zhang, Separating overlapped fingerprints. IEEE Trans.
Inf. Forensics Secur. 6(2), 346–359 (2011). ISSN 1556-6013. https://doi.org/10.1109/TIFS.
2011.2114345
References 43
4. S. Chikkerur, A.N. Cartwright, V. Govindaraju, Fingerprint enhancement using STFT analysis.

Pattern Recogn. 40(1), 198–211 (2007)
5. S.C. Dass, Markov random field models for directional field and singularity extraction in
fingerprint images. IEEE Trans. Image Process. 13(10), 1358–1367 (2004)
(1985)
7. B. DeCann, A. Ross, Relating ROC and CMC curves via the biometric menagerie, in 2013
IEEE Sixth International Conference on Biometrics: Theory, Applications and Systems (BTAS)
(2013), pp. 1–8. https://doi.org/10.1109/BTAS.2013.6712705
fingerprints. IEEE Trans. Inf. Forensics Secur. 7(5), 1498–1510 (2012). ISSN 1556-6013.
https://doi.org/10.1109/TIFS.2012.2204254
9. J. Feng, J. Zhou, A. Jain, Orientation field estimation for latent fingerprint enhancement. IEEE
Trans. Pattern Anal. Mach. Intell. 35(4), 925–940 (2013). ISSN 0162-8828. https://doi.org/10.
1109/TPAMI.2012.155
10. D.O. Gorodnichy, Multi-order analysis framework for comprehensive biometric performance
evaluation, in SPIE Defense, Security, and Sensing (International Society for Optics and
Photonics, Bellingham, 2010), p. 76670G
11. D. Gorodnichy, Multi-order biometric score analysis framework and its application to design-
ing and evaluating biometric systems for access and border control, in 2011 IEEE Workshop
on Computational Intelligence in Biometrics and Identity Management (CIBIM) (2011), pp.
44–53. https://doi.org/10.1109/CIBIM.2011.5949204
12. J. Gu, J. Zhou, C. Yang, Fingerprint recognition by combining global structure and local cues.
IEEE Trans. Image Process. 15(7), 1952–1964 (2006)
14. S. Huckemann, T. Hotz, A. Munk, Global models for the orientation field of fingerprints: an
approach based on quadratic differentials. IEEE Trans. Pattern Anal. Mach. Intell. 30(9), 1507–
1519 (2008)
15. A. Jain, J. Feng, Latent palmprint matching. IEEE Trans. Pattern Anal. Mach. Intell. 31(6),
1032–1047 (2009). ISSN 0162-8828. https://doi.org/10.1109/TPAMI.2008.242
16. T. Kamei, Image filter design for fingerprint enhancement, in Automatic Fingerprint Recogni-
tion Systems (Springer, New York, 2004), pp. 113–126
17. M. Kass, A. Witkin, Analyzing oriented patterns. Comput. Vis. Graph. Image Process. 37(3),
362–385 (1987)
18. M. Liu, X. Jiang, A.C. Kot, Fingerprint reference-point detection. EURASIP J. Adv. Sig.
Process. 2005(4), 1–12 (2005)
19. P.Z. Lo, Y. Luo, Method and apparatus for adaptive hierarchical processing of print images
(2006). US Patent App. 11/456,622
20. D. Maltoni, D. Maio, A. Jain, S. Prabhakar, Handbook of Fingerprint Recognition (Springer
Science & Business Media, New York, 2009)
21. Neurotechnology, VeriFinger SDK: Fingerprint Identification for Stand-Alone or Web Solu-
tions (2016). http://www.neurotechnology.com/verifinger.html
22. L. O’Gorman, J.V. Nickerson, An approach to fingerprint filter design. Pattern Recogn. 22(1),
29–38 (1989)
23. M. Oliveira, N.J. Leite, A multiscale directional operator and morphological tools for
reconnecting broken ridges in fingerprint images. Pattern Recogn. 41(1), 367–377 (2008)
24. S. Prabhakar et al., Probabilistic orientation field estimation for fingerprint enhancement and
verification, in BSYM’08 Biometrics Symposium, 2008 (IEEE, Piscataway, 2008), pp. 41–46
25. N.K. Ratha, S. Chen, A.K. Jain, Adaptive flow orientation-based feature extraction in
fingerprint images. Pattern Recogn. 28(11), 1657–1672 (1995)
26. B.G. Sherlock, D.M. Monro, A model for interpreting fingerprint topology. Pattern Recogn.
26(7), 1047–1055 (1993)
27. Y. Shi, J. Feng, J. Zhou, Separating overlapped fingerprints using constrained relaxation
labeling, in Proceedings of the International Joint Conference on Biometrics (2011)
28. M. Singh, D.K. Singh, P.K. Kalra, Fingerprint Separation: An Application of ICA, vol. 6982
(2008), pp. 69820L–69820L-11. https://doi.org/10.1117/12.777541
29. B. Stojanović, O. Marques, A. Nešković, Latent overlapped fingerprint separation: a review,
in Multimedia Tools and Applications (2016), pp. 1–28. https://doi.org/10.1007/s11042-016-
3908-y
overlapped fingerprints separation. Multimed. Tools Appl. 1–25 (2016). https://doi.org/10.
1007/s11042-016-3696-4
31. Y. Wang, J. Hu, D. Phillips, A fingerprint orientation model based on 2D Fourier expansion
(FOMFE) and its application to singular-point detection and fingerprint indexing. IEEE Trans.
Pattern Anal. Mach. Intell. 29(4), 573–585 (2007)
32. S. Yoon, J. Feng, A.K. Jain, On latent fingerprint enhancement, in SPIE Defense, Security, and
Sensing (International Society for Optics and Photonics, Bellingham, 2010), p. 766707
33. S. Yoon, J. Feng, A. Jain, Latent fingerprint enhancement via robust orientation field estima-
tion, in 2011 International Joint Conference on Biometrics (IJCB) (2011), pp. 1–8. https://doi.
org/10.1109/IJCB.2011.6117482.
34. N. Zhang, X. Yang, Y. Zang, X. Jia, J. Tian, Overlapped fingerprints separation based
on adaptive orientation model fitting, in 2014 22nd International Conference on Pattern
Recognition (ICPR) (2014), pp. 678–683. https://doi.org/10.1109/ICPR.2014.127
overlapped fingerprints separation. IEEE Trans. Inf. Forensics Secur. 9(10), 1547–1556 (2014).
ISSN 1556-6013. https://doi.org/10.1109/TIFS.2014.2340573
36. Q. Zhao, A. Jain, Model based separation of overlapping latent fingerprints. IEEE Trans. Inf.
Forensics Secur. 7(3), 904–918 (2012). ISSN 1556-6013. https://doi.org/10.1109/TIFS.2012.
2187281
37. J. Zhou, J. Gu, A model-based method for the computation of fingerprints’ orientation field.
IEEE Trans. Image Process. 13(6), 821–835 (2004)
Chapter 6
Machine Learning Based Separation of
Abstract This chapter describes a machine learning based approach for overlapped
fingerprint separation. The algorithm works in a block-based fashion: after produc-
ing an initial estimation of the orientation fields present in the overlapped fingerprint
image, it uses a neural network to separate the mixed orientation fields, which
are then post-processed to correct remaining errors and enhanced using the global
orientation field enhancement model. The proposed separation method has been
successfully tested on two different datasets.
6.1 Introduction
This chapter presents a neural network based approach for overlapped fingerprint
mask separation. As discussed in Chap. 5, the problem of overlapped latent fin-
gerprint separation consists in separating and enhancing component fingerprints in
order to make them suitable candidates for later feature extraction and matching
procedures. This process usually includes four steps [1, 3, 9–11] (Fig. 6.1):
1. Segmentation—This step involves latent fingerprints segmentation (also known
as ROI extraction), whose output produces the component fingerprints’ region
masks. See Chaps. 3 and 4 for details.
2. Initial orientation field (OF) estimation and enhancement—The initial ori-
entation field (which consists of a matrix containing information about ridge
angle in every pixel/block on fingerprint image) for the overlapped fingerprint
image is usually estimated and enhanced using a block-based approach. This
step produces a labeled image containing three different regions: (1) background
3. Orientation separation—Mixed orientation fields of overlapped regions for
two component fingerprints are: (1) separated using a neural network-based
approach, (2) smoothed in order to correct remaining errors, and (3) additionally

46 6 Machine Learning Based Separation of Overlapped Latent Fingerprints
Fig. 6.1 Overlapped latent fingerprint separation: a block diagram
enhanced using the global orientation field enhancement model. The output of
this step are component fingerprints’ orientation fields.
4. Component fingerprint images extraction—Two component fingerprints are
The approach described in this chapter (and described in greater detail in [10])
focuses on step 3—Orientation separation—where we propose a neural network
based method for separating overlapped orientations. A secondary contribution of
the proposed method relates to step 2—Initial orientation field (OF) estimation
and enhancement—due to the fact that our method was the first method to utilize
a global-based orientation field enhancement model of initially estimated single
region orientation field prior to separating overlapped region orientations. This was
done in order to minimize influence of local-based orientation estimation model and
to produce a more lifelike orientation field as an input for subsequent processing
steps [10].
6.2 Segmentation
In this stage, segmentation masks are manually created from pre-processed and size-
normalized input images (Fig. 6.2).
6.3 Initial Orientation Field Estimation and Enhancement 47
Fig. 6.2 Segmentation of an exemplar overlapped fingerprint image: (a) Overlapped fingerprint
image exemplar (I ); (b) component fingerprints’ region masks (M1 , M2 ); (c) single region masks
(M1S , M2S ); (d) overlapped region mask (MO )
6.3 Initial Orientation Field Estimation and Enhancement
The initial orientation field for the overlapped fingerprint image is estimated
and enhanced using a block-based approach, which consists of two steps: (1) a
Fourier analysis-based estimation of orientation field for the whole overlapped
fingerprint image and (2) enhancement of single region orientation field using a
global dictionary-based orientation field enhancement approach [4].
6.3.1 Initial Orientation Field Estimation
The initial orientation field consists of three different regions: (1) background region
without orientation values, (2) single region with one orientation value per block,
and (3) overlapped region with two orientation values, randomly distributed, per
block.
We adopted the local Fourier analysis method proposed in [7] and used by several
other authors in their work on similar problems [1, 3, 9]. See Sect. 5.2.2 for details.
6.3.2 Single Region Orientation Field Enhancement
The method used for initial orientation field estimation (Fourier analysis) is block-
based and does not take into account information from the rest of the fingerprint
image. In order to correct occasional orientation estimation errors caused by the
local nature of the estimation process (and ensure more accurate input data for
the subsequent stages of our algorithm) we applied Feng et al.’s publicly available
global dictionary based orientation field enhancement method [4] to the single
region orientation field of component fingerprints. This enhancement step mini-
mizes the influence of the local-based orientation estimation model. Consequently,
the resulting single region orientation field becomes more realistic and accurate [10].
6.4 Orientation Separation
The orientation separation process consists of three stages (Fig. 6.3): (1) over-
lapped region orientation field estimation based on neural networks (Sect. 6.5);
(2) estimated overlapped region orientation field smoothing, which provides error
correction to results from the previous step; and (3) component fingerprints’
orientation field enhancement (see [10] for details).
6.5 Neural Network: Architecture and Training
The method described in this chapter uses two neural networks, one for each type
of segment (square or T segment), trained with single blocks’ orientation values in
order to calculate one overlapped block orientation value. For this type of problem,
multilayer feed-forward neural networks with error back-propagation as a learning
rule, sigmoid hidden neurons and linear output neurons, are an appropriate choice
[5]. The chosen neural network structure consists of the input layer, two hidden
Fig. 6.3 Orientation separation process

6.6 Experimental Evaluation 49
layers, and the output layer. The number of perceptrons in the two hidden layers
was determined experimentally to be 50 and 20 perceptrons, respectively (see [10]
for details).
A subset of the Db1 fingerprint database from FVC2002 [8] was used for training
purposes. The Db1 database is divided into two parts, Db1_a and Db1_b: Db1_a
contains 800 samples (8 impressions of 100 different fingerprints), while Db1_b
contains 80 samples (8 impressions of 10 different fingerprints). Each image is
388 × 374 pixels (500 ppi), grayscale, and contains single fingerprints taken from
the subjects in organized sessions [8].
We used impressions no. 3 and no. 4 of the Db1_a database (200 fingerprint
images), from which 10,000 suitable image segments of outer size of 3×3 blocks
were randomly chosen. In order to be suitable for training, the selected image
segments must belong to the fingerprint area (Region of Interest-ROI) of that
fingerprint image. There are four possible variants for each type of image segment
patterns used in the proposed algorithm. We ensure that the 10,000 training samples
contain 2500 samples of each of the four variations [10].
The proposed method was implemented in MATLAB. To ensure reproducibility
of results, the critical parameters related to the orientation separation process—and
more specifically to neural network training—have also been published (see [10] for
details).
6.6 Experimental Evaluation
The performance of the algorithm was evaluated and compared against the equiv-
alent curves from the most prominent approaches from the literature as well as
two baseline cases: (1) experiments of direct matching of overlapped fingerprint
images with the template fingerprints and (2) experiments of matching of overlapped
fingerprint images segmented with component fingerprints masks with the template
fingerprints (see [10] for details).
Experimental results demonstrate that the proposed method shows better or
comparable performance to the most prominent approaches in the literature for both
(Tsinghua OLF and Tsinghua SOF) datasets and both figures of merit (ROC and
CMC curves) (see [10] for details).
Figure 6.4 presents some typical failure cases. The first failure case (Fig. 6.4a) is
mostly caused by poor manual mask segmentation—the segmented mask contains
only part of the fingerprint area, and the resulting fingerprint image contains too little
useful data to perform the matching procedure successfully. This problem is related
to any overlapped fingerprint separation approach and highlights the need for a fully
automated mask generation system. The second failure case (Fig. 6.4b) is caused by
singular points placed at the very edge of the overlapped area. Singular points are
points where fingerprint ridges show discontinuity. Since our approach separates
orientations partially based on the continuity of ridges, when singular points appear
at the very edge of the overlapped area it is very possible for some error to occur
Fig. 6.4 Failure cases, caused by: (a) poor manual mask segmentation; (b) singular points (marked
with red dots) placed at the edge of overlapped area
in the first iterations of orientation separation process. The error propagates to the
whole overlapped area, causing unreliable separation results (see [10] for details).
In this chapter we discussed a new method for separation of overlapped fingerprints

based on machine learning techniques, specifically neural networks. This method
was the first to use neural networks to address the overlapped fingerprint separation
problem. It was also the first method in the literature that utilizes a global-based
orientation field enhancement model of single region orientation field prior to
separating overlapped region orientations in order to minimize influence of local-
based orientation estimation model. As a result, the orientation field becomes more
lifelike and produces a more accurate input data for the subsequent processing
steps. Similarly to all previous approaches in this field, the proposed method does
not use additional manual intervention except for component fingerprint masks
markup [10].
References 51

• The overlapped fingerprint orientation separation method described in this
chapter consists of three stages: (1) overlapped region orientation field
estimation based on neural networks; (2) estimated overlapped region
orientation field smoothing; and (3) component fingerprints’ orientation
field enhancement.
• Experimental results show that the described method is competitive with
the most prominent approaches to the same problem in the literature.
• The proposed neural network based method represents a significant
advance towards the development of a fully automated system for over-
lapped fingerprints separation.
References
1. F. Chen, J. Feng, A. Jain, J. Zhou, J. Zhang, Separating overlapped fingerprints. IEEE Trans.
Inf. Forensics Secur. 6(2), 346–359 (2011). ISSN 1556-6013. https://doi.org/10.1109/TIFS.
2011.2114345
(1985)
fingerprints. IEEE Trans. Inf. Forensics Secur. 7(5), 1498–1510 (2012). ISSN 1556-6013.
https://doi.org/10.1109/TIFS.2012.2204254
4. J. Feng, J. Zhou, A. Jain, Orientation field estimation for latent fingerprint enhancement. IEEE
Trans. Pattern Anal. Mach. Intell. 35(4), 925–940 (2013). ISSN 0162-8828. https://doi.org/10.
1109/TPAMI.2012.155
5. M.H. Hassoun, Fundamentals of Artificial Neural Networks (MIT Press, Cambridge, 1995)
7. A. Jain, J. Feng, Latent palmprint matching. IEEE Trans. Pattern Anal. Mach. Intell. 31(6),
1032–1047 (2009). ISSN 0162-8828. https://doi.org/10.1109/TPAMI.2008.242
8. D. Maio, D. Maltoni, R. Cappelli, J.L. Wayman, A.K. Jain, FVC2002: second fingerprint
verification competition, in Proceedings of the 16th International Conference on Pattern
Recognition, 2002, vol. 3 (IEEE, Piscataway, 2002), pp. 811–814
9. Y. Shi, J. Feng, J. Zhou, Separating overlapped fingerprints using constrained relaxation
labeling, in Proceedings of the International Joint Conference on Biometrics (2011)
overlapped fingerprints separation. Multimed. Tools Appl. 1–25 (2016). https://doi.org/10.
1007/s11042-016-3696-4
overlapped fingerprints separation. IEEE Trans. Inf. Forensics Secur. 9(10), 1547–1556 (2014).
ISSN 1556-6013. https://doi.org/10.1109/TIFS.2014.2340573

Segmentation and Separation of Overlapped Latent Fingerprints Algorithms, Techniques, and Datasets

Uploaded by

Copyright:

Available Formats

Segmentation and Separation of Overlapped Latent Fingerprints Algorithms, Techniques, and Datasets

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Segmentation and Separation of Overlapped Latent Fingerprints Algorithms, Techniques, and Datasets

Uploaded by

Copyright:

Available Formats

SPRINGER BRIEFS IN COMPUTER SCIENCE

More information about this series at http://www.springer.com/series/10028

Segmentation and Separation

ISSN 2191-5768 ISSN 2191-5776 (electronic)

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019

We owe special gratitude to Dr. Borko Furht—Professor of Computer Science

Belgrade, Serbia Branka Stojanović

1 Latent Fingerprints Matching Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

4 Machine Learning Based Segmentation of Overlapped Latent

Abstract This chapter presents fundamental concepts and terminology associated

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 1

1.2 Latent Fingerprints

1.3 Manual Latent Fingerprints Matching: The ACE-V

Manual examination of latent fingerprints follows the ACE-V (Analysis, Compar-

1.4 Automated Latent Fingerprints Matching System

An automated latent fingerprint recognition system is designed with three main

1. Latent fingerprint segmentation, also known as region of interest (ROI) extrac-

1.5 Overlapped Fingerprints

An overlapped fingerprint image might contain two or more component finger-

produces a labeled image containing three different regions: (1) background

1.6 Concluding Remarks

Takeaways from this chapter:

1. D. Ashbaugh, Quantitative-Qualitative Friction Ridge Analysis: An Introduction to Basic and

13. D. Maltoni, D. Maio, A. Jain, S. Prabhakar, Handbook of Fingerprint Recognition (Springer

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 9

dataset, freely and publicly available,1 consists of 120,000 synthetically overlapped

2.2 Latent Fingerprint Datasets

2.2.1 Publicly Available Latent Fingerprint Datasets

Table 2.1 Latent fingerprint datasets

Table 2.2 Overlapped latent fingerprint datasets

2.2.2 Specialized Overlapped Latent Fingerprint Datasets

2.3 The Vlatacom Dataset

2.3.3 Dataset Structure and Organization

Fig. 2.2 Examples of

• Noise (2 variations): each overlapped fingerprint image is available in two

performance when singular points appear within the overlapped area—especially at

2.3.4 Additional Aspects

2.3.5 Experiments and Results

Preliminary experiments, originally reported in [12], have demonstrated the fea-

2.4 Concluding Remarks

In this chapter we introduced publicly available datasets in the field of overlapped

Takeaways from this chapter:

Abstract This chapter describes the problem of segmentation of overlapped

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 21

Figure 3.2 presents an overview of the overlapped latent fingerprints segmenta-

3.2 The Mask Segmentation Problem

The original problem of (individual) latent fingerprints segmentation comprises two

and overlapped (O) regions (blocks), regardless of which component fingerprint

3.3 Evaluation Methodology

The performance evaluation of single fingerprints segmentation methods is usually

Fig. 3.6 Typical confusion

3.4 Concluding Remarks

In this chapter we discussed the problem of latent fingerprints segmentation, with

Takeaways from this chapter:

Abstract This chapter describes a convolutional neural network (CNN)-based

4.2 Description of the Algorithm

The CNN-based method for overlapped fingerprints segmentation—represented

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2019 29

4.2.1 CNN Architecture