1
Orientation Field Estimation
for Latent Fingerprint Enhancement
Jianjiang Feng, Member, IEEE, Jie Zhou, Senior Member, IEEE, and Anil K. Jain, Fellow, IEEE
Abstract—Identifying latent fingerprints is of vital importance for law enforcement agencies to apprehend criminals and
terrorists. Compared to livescan and inked fingerprints, the image quality of latent fingerprints is much lower, with complex
image background, unclear ridge structure, and even overlapping patterns. A robust orientation field estimation algorithm is
indispensable for enhancing and recognizing poor quality latents. However, conventional orientation field estimation algorithms,
which can satisfactorily process most livescan and inked fingerprints, do not provide acceptable results for most latents. We
believe that a major limitation of conventional algorithms is that they do not utilize prior knowledge of the ridge structure in
fingerprints. Inspired by spelling correction techniques in natural language processing, we propose a novel fingerprint orientation
field estimation algorithm based on prior knowledge of fingerprint structure. We represent prior knowledge of fingerprints using
a dictionary of reference orientation patches, which is constructed using a set of true orientation fields, and the compatibility
constraint between neighboring orientation patches. Orientation field estimation for latents is posed as an energy minimization
problem, which is solved by loopy belief propagation. Experimental results on the challenging NIST SD27 latent fingerprint
database and an overlapped latent fingerprint database demonstrate the advantages of the proposed orientation field estimation
algorithm over conventional algorithms.
Index Terms—Fingerprint matching, fingerprint enhancement, latent fingerprint, orientation field, dictionary, spelling correction.
F
1 INTRODUCTION
L
ATENT fingerprints refer to the impressions uninten-
tionally left on items handled or touched by fingers.
Such fingerprints are often not directly visible unless some
physical or chemical technique is applied to enhance them
[1]. Since the early 20th century, latent fingerprints have
served as important evidence for law enforcement agencies
to apprehend and convict criminals [2].
Compared to fingerprints captured using inking or lives-
can techniques (see Fig. 1), the quality of most latent
fingerprints is very low, with unclear ridge structure, uneven
contrast, and overlapping patterns, such as printed letters,
handwriting, or even other fingerprints [3]. Because of the
poor image quality, features (such as minutiae) in latents
need to be manually marked by latent examiners so that
they can be searched against large fingerprint databases by
automated fingerprint identification systems (AFIS).
Automatic latent feature extraction is desirable for sev-
eral reasons.
1) Reducing the time spent by latent examiners in man-
ual markup. A crime scene can contain as many as
hundreds of latents. However, only a small portion
• Jianjiang Feng and Jie Zhou are with the Department of Automation,
Tsinghua University, Beijing 100084, China.
E-mail: {jfeng,jzhou}@tsinghua.edu.cn
• Anil K. Jain is with the Department of Computer Science and Engineer-
ing, Michigan State University, East Lansing, MI-48824, U.S.A. He is
also affiliated with the Department of Brain and Cognitive Engineering,
Korea University, Anamdong, Seongbukgu, Seoul 136-713, Republic of
Korea.
E-mail: jain@cse.msu.edu
of them can be processed simply because law en-
forcement agencies do not have sufficient manpower.
It can take twenty minutes or even longer to mark
the minutiae in a single latent. Automatic feature
extraction can improve the efficiency of processing
latents, leading to more identifications quickly [4].
2) Improving the compatibility between minutiae in
latent and full fingerprints. In current practice, minu-
tiae in latents are manually marked while minutiae
in full fingerprints are automatically extracted. This
can cause the compatibility problem. Although this
compatibility issue is not a severe problem for full
fingerprint matching, this problem cannot be underes-
timated in the case of latent matching, since in a tiny
and smudgy latent, every minutia plays an important
role. To address this issue, AFIS vendors usually
provide training courses to latent examiners on how
to better mark minutiae for their particular AFIS
system since different vendors’ systems are not very
consistent in extracting minutiae. However, it takes
time for fingerprint examiners to get familiar with
a system. This problem can be alleviated provided
features in latents are also extracted by automatic
algorithms.
3) Improving repeatability/reproducibility of latent iden-
tification. The minutiae in the same latent marked
by different latent examiners or even by the same
examiner (but at different times) may not be the same.
This is one of reasons why different latent examiners
or even the same examiner (but at different times)
make different matching decisions on the same latent-
exemplar pair [5], [6]. The Daubert standard, which
To appear in IEEE TPAMI, 2012.
2
(a) (b) (c)
Fig. 1. Fingerprints obtained using three types of tech-
niques. (a) Inked fingerprint, (b) live-scan fingerprint,
and (c) latent fingerprint.
specifies the admissibility of scientific testimony in
United States courts, requires that the error rate of
latent matching should be known. However, lack
of repeatability/reproducibility makes estimating the
error rates of latent examiners very difficult [7]. Even
if an error rate can be estimated by a “black box”
test
1
[5], it cannot apply to a different examiner’s
decision on a new latent-exemplar pair. The only
viable solution appears to be to keep improving
automated fingerprint systems’ performance so that
the role of latent examiners is limited to very difficult
latents. Hence, automating latent feature extraction is
an indispensable step towards this long-term goal.
To enable reliable feature extraction, a latent fingerprint
image, which is often of very poor quality, needs to
go through an image enhancement stage, which connects
broken ridges, separates joined ridges, and removes over-
lapping patterns. After the latent is enhanced, conventional
minutiae extraction algorithm can be used [8]. Contex-
tual filtering (or directional filtering) is the most widely
used fingerprint enhancement technique [9]–[11]. Although
different contextual filters differ in details, the intended
behavior is the same: (i) performing low-pass filtering along
the ridge in order to fill gaps and pores, and (ii) performing
bandpass filtering across the ridges in order to separate
joined ridges [8].
The contextual filtering techniques require reliable esti-
mation of local ridge orientation, which is not a trivial task
for poor quality fingerprints. That is why orientation field
estimation and a very related topic, singularity detection,
are two of the most active topics in the fingerprint recog-
nition literature [10]–[37]. However, all these algorithms
were developed for plain or rolled fingerprints. As shown
in Fig. 2, the performance of representative orientation field
estimation algorithms on latents is far from satisfactory.
Realizing the gap between human and machine’s perfor-
mance in extracting orientation field for latents, a few recent
studies have focused on latent orientation field estimation
[38], [39]. However these algorithms require manually
marked singular points in order to obtain a reasonable
1. In a black box test [5], the latent matching process of examiners is
treated as a black box and only the final accuracy in making the decisions
is studied.
performance.
In this paper, a robust orientation field estimation al-
gorithm is proposed to process poor quality fingerprints,
especially latents. Given prior knowledge of fingerprint
structure, which is represented by a dictionary of reference
orientation patches and compatibility constraints between
adjacent orientation patches, the proposed algorithm obtains
better performance for latents than published algorithms
(see Fig. 2). For some latents, the match scores using
minutiae automatically extracted from latents enhanced by
the proposed algorithm are even higher than the match
scores using manually marked minutiae.
The rest of the paper is organized as follows. In section
2, published orientation field estimation algorithms are
reviewed. In section 3, the motivation of the proposed al-
gorithm is discussed. The details of the proposed algorithm
are presented in section 4. Experimental results are reported
and analyzed in section 5. Finally, we conclude the paper
and suggest future research directions for this topic.
2 RELATED WORK
In this section, we review published algorithms for orien-
tation field estimation, which are coarsely classified into
three categories.
2.1 Local Estimation
Local estimation approaches compute a local ridge orienta-
tion at pixel x = (x, y) using only the neighborhood around
x, which is typically 32×32 pixels for 500 ppi fingerprints.
The most well-known local estimation approach is
gradient-based [13], [14], [41], [42]. Since gradient opera-
tors, such as Prewitt or Sobel operators [43], are sensitive to
noise and pores (regularly placed on the ridges), a dominant
orientation is computed using the gradients in the local
neighborhood.
Slit-based approach is another widely used orientation
field estimation method [30]. This approach explicitly uti-
lizes 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.
Ridge pattern in a local area of a finger can be ap-
proximated by a 2D sine wave [44]. Thus the magnitude
spectrum of the Fourier transform of a local fingerprint
image will contain a pair of peaks whose location corre-
sponds to the parameters of the sine wave. The magnitude
spectrum can be mapped to the polar coordinate system.
The normalized magnitude spectrum can be viewed as a
probability distribution [11]. The best orientation can be
estimated as the most probable orientation or the mean.
Orientation fields obtained by local estimation ap-
proaches for poor quality fingerprints are usually very
noisy. To deal with this problem, two types of algorithms
have been adopted to regularize the noisy orientation field,
namely, orientation field smoothing and global parametric
model fitting. Typically, some constraints or knowledge
about the fingerprint orientation field is utilized in the
regularization algorithm.
3
(a) (b)
(c) (d) (e)
Fig. 2. A latent fingerprint (a), its mated rolled fingerprint (b) with the corresponding region marked by a
green box, and its enhanced latents using three different orientation field estimation algorithms: (c) FOMFE
[25], (d) STFT [11], (e) proposed. Minutiae in (a) are manually marked by latent examiners, minutiae in (b) are
automatically extracted using VeriFinger SDK 6.2 [40], while minutiae in (c, d, e) are automatically extracted from
the enhanced images using VeriFinger. The minutiae match scores (computed by VeriFinger) between (a, c, d,
e) and the mated rolled fingerprint (b) are 39, 35, 24, and 54, respectively.
2.2 Smoothing
Many orientation field regularization techniques have been
proposed to deal with noise present in the fingerprint. The
most commonly used smoothing method is based on low-
pass filtering [14]. Although the low-pass filtering method
is simple and effective, the size of the filtering window
is a critical parameter. A large window can suppress the
noise better while a small window can preserve the true
orientation in high curvature region. Several authors have
suggested using multi-resolution orientation fields to ad-
dress this problem [9], [21], [30], [45]. However, when the
noise is severe as in latents, smoothing techniques are not
able to recover the true orientation field.
Several researchers have implemented orientation field
smoothing by using the Markov random field (MRF) model
or energy minimization approach [16], [19], [31]. A well
known limitation of these algorithms is that the orientation
variable corresponds to a very small image region so that
it can be represented by a single dominant orientation.
However, a MRF model with small neighborhood or context
is able to exploit only limited prior knowledge about
fingerprint structure [46], [47] and thus cannot deal with
fingerprints of very poor quality.
2.3 Global Parametric Models
Researchers have proposed several mathematical models to
represent the whole fingerprint orientation field. Some of
the models are quite general, such as polynomials [22] and
Fourier series [25], while the others are more specific to
fingerprints [12], [20], [29]. Without invoking constraints
on the parameters [22], [25], general models tend to have
over-fitting (e.g., if the order of the polynomial is high) or
under-fitting problems (e.g., if the order of the polynomial
is low) especially when the initial orientation field is very
noisy. Models which explicitly consider singular points
[12], [20], [29] rely on reliable extraction of singular points.
However, extracting singular points in latents is a very
challenging problem itself. That is why the orientation
field estimation approaches in [38], [39] require manually
marked singular points as input.
3 MOTIVATION
Although conventional orientation field estimation algo-
rithms can satisfactorily process most live-scan and inked
fingerprints, their performance on most of the latents are
far from satisfactory (see Fig. 2). We believe that a major
limitation of conventional algorithms is that they do not
4
(a) (b)
(c) (d) (e)
Fig. 3. A fingerprint image in (a) is divided into a
number of non-overlapping blocks of 16 × 16 pixels in
(b). A patch contains 10 × 10 blocks and neighboring
patches are overlapped as shown in (b). Given a
subimage in (c), the associated blocks and or ientation
patch is shown in (d) and (e). The orientation element
represents the dominant ridge flow in a block.
adequately incorporate prior knowledge of fingerprints. It is
now widely recognized that representing and learning prior
knowledge is of fundamental importance in many natural
language processing and computer vision tasks [47], [48].
However, in the fingerprint recognition area, it has received
little attention.
We can draw an analogy between fingerprint orientation
field and a sentence in a natural language. A sentence is
comprised of words which are further comprised of letters.
Similarly, a fingerprint orientation field is comprised of
orientation patches which are further comprised of orienta-
tion elements. Hence, a fingerprint orientation field can be
viewed as a sentence, an orientation patch can be viewed
as a word, and an orientation element can be viewed as a
letter. These definitions are illustrated in Fig. 3.
Spelling correction [48] in a sentence is possible because
not all possible combinations of letters are valid to form
words and not all possible combinations of words form
a valid sentence. Similarly, error correction in orientation
fields is possible because not all possible combinations of
orientation elements are valid and not all possible combina-
tions of orientation patches are valid for a fingerprint. For
example, Fig. 4 shows 60 orientation patches, which are
generated by sampling an independent uniform distribution
(namely, each orientation element has the same uniform
distribution and the elements are assumed to be statistically
independent). None of these orientation patches is likely to
appear in real fingerprints.
Spelling correction techniques use dictionary (or lexicon,
word list) and context information to detect and correct
spelling errors [48]. While dictionary can be used to detect
and correct most non-word errors, contextual information
is required to resolve ambiguity when there are multiple
candidate words. For example, without context, ater can
be explained as after, later, water, alter, or ate.
The proposed orientation field estimation algorithm is
inspired by the above spelling correction method. We first
build a dictionary of reference orientation patches using
a set of orientation fields extracted from real fingerprints.
Given an input fingerprint, we estimate an initial orien-
tation field using traditional orientation field estimation
approaches. For poor quality fingerprints, such as most
latents, the initial orientation fields are very noisy. Errors
in the initial orientation field need to be corrected using
dictionary as well as context information. Specifically, for
each initial orientation patch, we find a list of candidates
from the dictionary which might be the true orientation
patch. Contextual information is then used to determine a
single candidate for each patch.
4 PROPOSED ALGORITHM
4.1 Overview
The proposed orientation field estimation algorithm consists
of an off-line dictionary construction stage and an on-line
orientation field estimation stage. In the off-line stage, a
set of good quality fingerprints of various pattern types
(arch, loop, and whorl) are manually selected and their
orientation fields are used to construct a dictionary of
orientation patches. In the on-line stage, given a fingerprint
image, its orientation field is automatically estimated using
the following steps:
1) initial estimation: The initial orientation field is ob-
tained using a local orientation estimation method,
such as local Fourier analysis [44].
2) dictionary lookup: The initial orientation field is
divided into overlapping patches. For each initial
orientation patch, its six nearest neighbors in the
dictionary are viewed as candidates for replacing the
noisy initial orientation patch.
3) context-based correction: The optimal combination of
candidate orientation patches is found by considering
the compatibility between neighboring orientation
patches.
In the following subsections, we first describe the off-line
dictionary construction and then present the three steps in
the on-line orientation field estimation algorithm.
4.2 Dictionary Construction
The dictionary consists of a number of orientation patches
of the same size. An orientation patch consists of b × b
orientation elements and an orientation element refers to
the dominant orientation in a block of size 16 × 16 pixels.
We construct a dictionary of orientation patches from
a set of high quality fingerprints (referred to as reference
fingerprints). The orientation fields (defined on blocks of
size 16×16 pixels) of these fingerprints are estimated using
5
Fig. 4. Orientation patches sampled from a uniform distribution of orientation element. None of these orientation
patches is likely to appear in real fingerprints. An orientation patch contains 10 × 10 orientation elements and an
orientation element represents the dominant direction in a block of 16 × 16 pixels.
Fig. 5. The proposed system consists of an off-line dictionary construction stage and an on-line orientation field
estimation stage.
