Latent Fingerprint Matching Using Descriptor-Based Hough Transform

TL;DR: A new fingerprint matching algorithm which is especially designed for matching latents and uses a robust alignment algorithm (descriptor-based Hough transform) to align fingerprints and measures similarity between fingerprints by considering both minutiae and orientation field information.

Abstract: Identifying suspects based on impressions of fingers lifted from crime scenes (latent prints) is a routine procedure that is extremely important to forensics and law enforcement agencies. Latents are partial fingerprints that are usually smudgy, with small area and containing large distortion. Due to these characteristics, latents have a significantly smaller number of minutiae points compared to full (rolled or plain) fingerprints. The small number of minutiae and the noise characteristic of latents make it extremely difficult to automatically match latents to their mated full prints that are stored in law enforcement databases. Although a number of algorithms for matching full-to-full fingerprints have been published in the literature, they do not perform well on the latent-to-full matching problem. Further, they often rely on features that are not easy to extract from poor quality latents. In this paper, we propose a new fingerprint matching algorithm which is especially designed for matching latents. The proposed algorithm uses a robust alignment algorithm (descriptor-based Hough transform) to align fingerprints and measures similarity between fingerprints by considering both minutiae and orientation field information. To be consistent with the common practice in latent matching (i.e., only minutiae are marked by latent examiners), the orientation field is reconstructed from minutiae. Since the proposed algorithm relies only on manually marked minutiae, it can be easily used in law enforcement applications. Experimental results on two different latent databases (NIST SD27 and WVU latent databases) show that the proposed algorithm outperforms two well optimized commercial fingerprint matchers. Further, a fusion of the proposed algorithm and commercial fingerprint matchers leads to improved matching accuracy.

Summary

1. Introduction

• The practice of using latent fingerprint for identifying suspects is not new.
• In 1893, the acceptance of the hypothesis by the Home Ministry Office, UK, that any two individuals have different fingerprints made many law enforcement agencies aware of the potential of using fingerprints as a mean of identification [8].
• With the invention of AFIS, fingerprint examiners identify latents using a semi-automatic procedure that consists of following stages: (i) manually mark the features (minutiae and singular points) in the latent, (ii) launch an AFIS search, and (iii) visually verify each of the candidate fingerprints returned by AFIS.
• The impressive matching accuracy reported in ELFT does not mean that the current practice of manually marking minutiae in latents should be changed.

2. Latent Matching Approach

• There are three main steps in fingerprint matching: alignment (or registration) of the fingerprints, pairing of the minutiae, and score computation.
• Given two sets of aligned minutiae, two minutiae are considered as a matched pair if their Euclidean distance and direction difference are less than pre-specified thresholds.
• Finally, a score is computed based on a variety of factors such as the number of matched minutiae and the similarity between the descriptors of the matched minutiae pairs.
• It is important to emphasize that while latents are manually encoded (namely marking minutiae), minutiae in rolled prints are automatically extracted.

2.1. Local Minutia Descriptor

• This local structure is represented as a cylinder, which contains information about the relationship between a minutia and its neighboring minutiae.
• The base of the cylinder is related to the spatial relationship, and its height is related to the directional relationship.
• Each cell in the cylinder accumulates contributions from each minutia in the neighborhood.
• The resulting cylinder can be viewed as a vector, and therefore the similarity between two minutia descriptors can be easily 1Local minutia descriptors shown in Figure 3 is from [1].
• This representation presents some advantages, such as: invariant to translation and rotation; robust against small skin distortion and missing or spurious minutiae; and of fixed length.

2.2. Fingerprint Alignment

• Fingerprint alignment or registration consists of estimating the parameters (rotation, translation and scale) that align two fingerprints.
• Since manually marking minutiae is a common practice for latent matching, their approach to align two fingerprints is based on minutiae.
• Each parameter receives “a vote” proportional to the matching score for the corresponding transformation.
• In their approach, the alignment is conducted in a very similar way, but the evidence for each parameter is accumulated based on the similarity between the local descriptors of the two involved minutiae, with the similarity and descriptor being the ones described in Section 2.1.
• Given two sets of minutiae, one from the latent and the other from the rolled print being compared, translation and rotation parameters can be obtained for each possible minutiae pair (one minutia from each set).

2.4. Score Computation

• Score computation is a very important step in the matching process.
• This is not appropriate for latent matching because the number of minutiae in different latents varies substantially.
• One solution to modify the above scoring method is to divide the number of matched minutiae by the number of minutiae in the latent, which is almost always smaller than the number of minutiae in the rolled print.
• In their approach, the authors use minutiae similarity to weigh the contribution of each pair of matched minutiae.
• Then, the matching score between the two aligned fingerprints is given by: 𝑠𝑐𝑜𝑟𝑒 = ∑ ∀𝑚𝑖∈𝑀 𝑆𝑖 𝑁 . (5) To further improve the matching performance, the authors combine the scores based on matched minutiae from two different pairing thresholds by their weighted sum; they assume equal weights.

3. Experimental Results

• Matching experiments were conducted on the NIST Special Database 27, which consists of 258 latent fingerprint images.
• Another indicator of fingerprint quality that affects the matching performance is the number of minutiae in the latent print [6].
• In order to estimate the alignment error, the authors use ground truth mated minutiae pairs, which are marked by fingerprint examiners, to compute the average distance of the true mated pairs after alignment.
• We also show their results for latents of six different quality levels (good, bad, ugly; large, medium, small) separately.the authors.the authors.
• The improvement of the proposed matcher over VeriFinger at rank-1 accuracy varies from 2.3% for latents with a large number of minutiae to 22% for latents of ugly quality.

Figures

Figure 1. Types of Fingerprints obtained by different acquisition methods. (a) Rolled (ink) (NIST SD27), (b) plain (live-scan) (FVC2002), and (c) latent (NIST SD27).

Figure 6. Overall matching performance (in terms of ROC curves) for latents with different subjective and objective qualities. Manually marked minutiae in latents are used as input for both the matchers (proposed and VeriFinger).

Figure 2. Latent fingerprints of three different quality levels in NIST SD27. (a) Good, (b) Bad, and (c) Ugly.

Figure 8. Latent prints that were not successfully matched. These two latents were matched to their true mates at ranks 1253 and 1057, respectively.

Figure 7. Latent prints correctly identified at rank-1.

Figure 3. Overview of the proposed approach.

Table 1. Rank-1 accuracies for various subjective qualities of latents in NIST SD27.

Table 2. Rank-1 accuracies for various objective qualities of latents in NIST SD27.

Figure 4. Alignment Accuracy: percentage of correctly aligned latents vs. alignment error.

Figure 5. Example of alignment error due to the small number of true mated minutia pairs in the overlapping area between a latent and its mated rolled print.

Full text

Latent Fingerprint Matching using Descriptor-Based Hough Transform
Alessandra A. Paulino
Dept. of Computer Science
and Engineering
Michigan State University
East Lansing, MI, U.S.A.
paulinoa@cse.msu.edu
Jianjiang Feng
Dept. of Automation
Tsinghua University
Beijing, China
jfeng@tsinghua.edu.cn
Anil K. Jain
∗
Dept. of Computer Science
and Engineering
Michigan State University
East Lansing, MI, U.S.A.
jain@cse.msu.edu
Abstract
Identifying suspects based on impressions of fingers
lifted from crime scenes (latent prints) is extremely impor-
tant to law enforcement agencies. Latents are usually par-
tial fingerprints with small area, contain nonlinear distor-
tion, and are usually smudgy and blurred. Due to some of
these characteristics, they have a significantly smaller num-
ber of minutiae points (one of the most important features
in fingerprint matching) and therefore it can be extremely
difficult to automatically match latents to plain or rolled
fingerprints that are stored in law enforcement databases.
Our goal is to develop a latent matching algorithm that uses
only minutiae information. The proposed approach consists
of following three modules: (i) align two sets of minutiae by
using a descriptor-based Hough Transform; (ii) establish
the correspondences between minutiae; and (iii) compute a
similarity score. Experimental results on NIST SD27 show
that the proposed algorithm outperforms a commercial fin-
gerprint matcher.
1. Introduction
The practice of using latent fingerprint for identifying
suspects is not new. According to Cummins and Midlo [2],
the first publication in modern literature related to finger-
print identification appeared in Nature, in 1880. This pub-
lication was entitled “On the Skin-furrows of the Hand,”
authored by Faulds [3]. In this article, Faulds suggested
that fingerprints left on crime scenes could be used to iden-
tify criminals or to exclude suspects. Soon after this article
was published, a letter written by Herschel was published
in Nature [5] stating that he had been using fingerprint as
a method of identification in India for about 20 years, with
different applications such as to avoid personification.
∗
A.K. Jain is also with the Dept. of Brain and Cognitive Engineering,
Korea University, Anam-dong, Seongbuk-gu, Seoul 136-713, Korea.
(a) (b) (c)
Figure 1. Types of Fingerprints obtained by different acquisition
methods. (a) Rolled (ink) (NIST SD27), (b) plain (live-scan)
(FVC2002), and (c) latent (NIST SD27).
In 1893, the acceptance of the hypothesis by the Home
Ministry Office, UK, that any two individuals have different
fingerprints made many law enforcement agencies aware
of the potential of using fingerprints as a mean of identi-
fication [8]. Some law enforcement agencies started col-
lecting fingerprints from offenders so that they could iden-
tify them later in case they changed their names to evade
harsher penalties. Also, fingerprints collected from crime
scenes were compared to fingerprints collected from previ-
ous offenders so that they could identify repeat offenders,
criminals who have been previously arrested.
Fingerprint identification started as a completely manual
approach. Due to growing demands on fingerprint recog-
nition, research was initiated to automate fingerprint recog-
nition, which led to the development of Automated Finger-
print Identification Systems (AFIS).These systems are used
worldwide not only by law enforcement agencies but also
in many other government and commercial applications.
Nowadays fingerprint recognition is routinely used in civil-
ian applications that have stringent security requirements.
There are three types of fingerprints: rolled, which is a
print obtained by rolling the finger “nail-to-nail” on a paper
or the platen of a scanner; plain, which is a print obtained by
placing the finger flat on a paper or the platen of a scanner
without rolling; and latents, which are lifted from surfaces
of objects that are inadvertently touched or handled by a
1

person typically at crime scenes (see Fig. 1). Lifting of la-
tents may involve a complicated process, and it can range
from simply photographing the print to more complex dust-
ing or chemical processing.
Rolled prints contain the largest amount of informa-
tion about the fingerprint since they contain information
from nail-to-nail; latents usually contain the least amount
of information for matching or identification. Compared to
rolled or plain fingerprints, latents are smudgy and blurred,
capture only a small finger area, and have large nonlinear
distortion due to pressure variations. Due to their poor qual-
ity and small area, latents have a significantly smaller num-
ber of minutiae compared to rolled or plain prints (the aver-
age number of minutiae in NIST Special Database 27 (NIST
SD27) [12] images is 21 for latents versus 106 for the cor-
responding rolled prints). Those characteristics make the
latent fingerprint matching problem very challenging.
Manual latent fingerprint identification is performed fol-
lowing a procedure referred to as ACE-V (analysis, com-
parison, evaluation and verification) and it requires a large
amount of human intervention. Because this procedure is
quite tedious and time consuming for latent examiners, la-
tents are usually matched against full prints of a small num-
ber of suspects identified by other means. With the in-
vention of AFIS, fingerprint examiners identify latents us-
ing a semi-automatic procedure that consists of following
stages: (i) manually mark the features (minutiae and singu-
lar points) in the latent, (ii) launch an AFIS search, and (iii)
visually verify each of the candidate fingerprints returned
by AFIS. The accuracy and speed of this procedure is still
not satisfactory.
Recent studies on latent fingerprints can be classified
into two categories according to their objective: higher
matching accuracy [4, 6, 7] or higher degree of automa-
tion [15, 13]. Improved latent matching accuracy has been
reported by using extended features which are manually
marked for latents [6, 7]. However, marking extended fea-
tures (orientation field, ridge skeleton, etc.) in poor quality
latents is very time-consuming and might be only feasible
in rare cases.
NIST has been conducting a multi-phase project on Eval-
uation of Latent Fingerprint Technologies (ELFT) to eval-
uate automatic latent feature extraction and matching tech-
niques [10]. In Phase I, the most accurate system showed
a rank-1 accuracy of 80% (100 latents against 10, 000
rolled prints). In Phase II, the rank-1 accuracy of the most
accurate system was 97.2% (835 latents against 100, 000
rolled prints). These accuracies cannot be directly com-
pared since the Phase I and Phase II evaluations used dif-
ferent databases. Also, the quality of latents used in Phase
II is better compared to Phase I. Fig. 2 shows three latents of
different quality in NIST SD27. The impressive matching
accuracy reported in ELFT does not mean that the current
(a) (b) (c)
Figure 2. Latent fingerprints of three different quality levels in
NIST SD27. (a) Good, (b) Bad, and (c) Ugly.
practice of manually marking minutiae in latents should be
changed.
The goal of this work is to develop a latent fingerprint
matching algorithm that is solely based on minutiae. Since
manually marking minutiae in latents is a common practice
in the latent fingerprint community, the proposed matcher
can be directly used in operational settings.
The rest of the paper is organized as follows: in Section
2, all steps of our proposed method are described; in Section
3, our experimental results are presented and discussed; in
Section 4, we present our conclusions and future work.
2. Latent Matching Approach
There are three main steps in fingerprint matching: align-
ment (or registration) of the fingerprints, pairing of the
minutiae, and score computation. In our approach, we use
a Descriptor-based Hough Transform to align two finger-
prints. Given two sets of aligned minutiae, two minutiae are
considered as a matched pair if their Euclidean distance and
direction difference are less than pre-specified thresholds.
Finally, a score is computed based on a variety of factors
such as the number of matched minutiae and the similarity
between the descriptors of the matched minutiae pairs. Fig-
ure 3
1
shows an overview of the proposed approach. It is
important to emphasize that while latents are manually en-
coded (namely marking minutiae), minutiae in rolled prints
are automatically extracted.
2.1. Local Minutia Descriptor
Minutia Cylinder-Code (MCC) is a minutiae representa-
tion based on 3D data structures [1]. In the MCC represen-
tation, a local structure is associated to each minutia. This
local structure is represented as a cylinder, which contains
information about the relationship between a minutia and
its neighboring minutiae. The base of the cylinder is related
to the spatial relationship, and its height is related to the di-
rectional relationship. Each cell in the cylinder accumulates
contributions from each minutia in the neighborhood. The
resulting cylinder can be viewed as a vector, and therefore
the similarity between two minutia descriptors can be easily
1
Local minutia descriptors shown in Figure 3 is from [1].

Figure 3. Overview of the proposed approach.
computed as a vector correlation measure. A more detailed
description of the cylinder generation and of the similarity
between two cylinders can be found in [1]. This representa-
tion presents some advantages, such as: invariant to trans-
lation and rotation; robust against small skin distortion and
missing or spurious minutiae; and of fixed length.
2.2. Fingerprint Alignment
Fingerprint alignment or registration consists of estimat-
ing the parameters (rotation, translation and scale) that align
two fingerprints. There are a number of features that may
be used to estimate alignment parameters between two fin-
gerprints, including orientation field, ridges and minutiae.
There are also a number of ways of aligning two finger-
prints: Generalized Hough Transform, local descriptors, en-
ergy minimization, etc.
In the latent fingerprint case, singularities are not al-
ways present, making it difficult to base the alignment of
the fingerprint on singular points alone. To obtain manu-
ally marked orientation field is expensive, and to automati-
cally extract orientation field from a latent image is a very
challenging problem. Since manually marking minutiae is a
common practice for latent matching, our approach to align
two fingerprints is based on minutiae.
Ratha et al. introduced an alignment method for minu-
tiae matching that estimates rotation, scale, and transla-
tion parameters using a Generalized Hough Transform [14].
Given two sets of points (minutiae), a matching score is
computed for each transformation in the discretized set of
all allowed transformations. For each pair of minutiae, one
minutia from each set, and for given scale and rotation pa-
rameters, unique translation parameters can be computed.
Each parameter receives “a vote” proportional to the match-
ing score for the corresponding transformation. The trans-
formation that gives the maximum score is considered the
best one. In our approach, the alignment is conducted in
a very similar way, but the evidence for each parameter is
accumulated based on the similarity between the local de-
scriptors of the two involved minutiae, with the similarity
and descriptor being the ones described in Section 2.1.
Given two sets of minutiae, one from the latent and the
other from the rolled print being compared, translation and
rotation parameters can be obtained for each possible minu-
tiae pair (one minutia from each set). Let {(𝑥
𝑙
,𝑦
𝑙
,𝜃
𝑙
)}
and {(𝑥
𝑟
,𝑦
𝑟
,𝜃
𝑟
)} be the minutiae sets for latent and rolled
prints, respectively, centered at their means. Then, for each
pair of minutiae, we have
𝜃 =min(∥𝜃
𝑙
− 𝜃
𝑟
∥, 360 −∥𝜃
𝑙
− 𝜃
𝑟
∥), (1)
(
Δ𝑥
Δ𝑦
)
=
(
𝑥
𝑟
𝑦
𝑟
)
−
(
cos 𝜃 sin 𝜃
− sin 𝜃 cos 𝜃
)(
𝑥
𝑙
𝑦
𝑙
)
. (2)
Since it is not necessary to consider the scale parame-
ter in fingerprint matching, unique translation parameters

can be obtained for each pair based on the rotation dif-
ference between the minutiae in the pair. The translation
and rotation parameters need to be quantized to the clos-
est bins. After the quantization, evidence is accumulated in
the correspondent bin based on the similarity between the
local minutiae descriptors. The assumption here is that true
mated minutiae pairs will vote for very similar sets of align-
ment parameters, while non-mated minutiae pairs will vote
randomly throughout the parameter space. As a result, the
set of parameters that presents the highest evidence is con-
sidered the best one. For robustness, more than one set of
alignment parameters with high evidence are considered.
In order to make the alignment computationally efficient
and also more accurate, we use a two-stage approach for
the Descriptor-based Hough Transform. We first perform
the voting in a relatively coarse parameter space. Based on
the peaks in the Hough space, we repeat the voting i nside a
neighborhood around the peaks, but with a more refined set
of parameter range. We also keep track of the points that
contribute to the peaks and then compute a rigid transfor-
mation matrix from those points.
2.3. Minutiae Pairing
After aligning two sets of minutiae, we need to find the
minutiae correspondences between the two sets, i.e. minu-
tiae need to be paired. The pairing of minutiae consists of
finding minutiae that are sufficiently close in terms of loca-
tion and direction. Let 𝑚
𝑖
=(𝑥
𝑖
,𝑦
𝑖
,𝜃
𝑖
) be a minutia from
the aligned latent and 𝑚
𝑗
=(𝑥
𝑗
,𝑦
𝑗
,𝜃
𝑗
) be a minutia from
the rolled print. Then, 𝑚
𝑖
and 𝑚
𝑗
are considered paired or
matched minutiae if
𝑑(𝑚
𝑖
,𝑚
𝑗
)=
√
(𝑥
𝑖
− 𝑥
𝑗
)
2
+(𝑦
𝑖
− 𝑦
𝑗
)
2
≤ 𝑑
0
(3)
𝜃
𝑖𝑗
=min(∥𝜃
𝑖
− 𝜃
𝑗
∥, 360 −∥𝜃
𝑖
− 𝜃
𝑗
∥) ≤ 𝜃
0
, (4)
In aligning two sets of minutiae, this is the most natural
way of pairing minutiae. We use a one-to-one matching,
which means each minutia in the latent can be matched to
only one minutia in the rolled print. Ties are broken based
on the closest minutia.
2.4. Score Computation
Score computation is a very important step in the match-
ing process. A straightforward approach to compute the
matching score consists of the number of matched minutiae
divided by the average number of minutiae in the two finger-
prints. This is not appropriate for latent matching because
the number of minutiae in different latents varies substan-
tially. One solution to modify the above scoring method is
to divide the number of matched minutiae by the number of
minutiae in the latent, which is almost always smaller than
the number of minutiae in the rolled print.
In our approach, we use minutiae similarity to weigh
the contribution of each pair of matched minutiae. Given
a search fingerprint (latent) and a template fingerprint
(rolled), and considering that the fingerprints are already
aligned, let 𝑀 be the set of 𝑛 matched minutiae pairs be-
tween the two fingerprints, {𝑚
𝑖
}
𝑛
𝑖=1
be matched minutiae
pairs in 𝑀 , {𝑆
𝑖
}
𝑛
𝑖=1
be their respective similarities, and 𝑁
be the number of minutiae in the latent. Then, the matching
score between the two aligned fingerprints is given by:
𝑠𝑐𝑜𝑟 𝑒 =
∑
∀𝑚
𝑖
∈𝑀
𝑆
𝑖
𝑁
. (5)
To further improve the matching performance, we com-
bine the scores based on matched minutiae from two dif-
ferent pairing thresholds by their weighted sum; we assume
equal weights. Since we perform 10 different alignments,
we compute 10 different matching scores between two fin-
gerprints; the final score between the two fingerprints is the
maximum among the 10 scores computed from different hy-
pothesized alignments.
3. Experimental Results
Matching experiments were conducted on the NIST Spe-
cial Database 27, which consists of 258 latent fingerprint
images. The background database consists of 258 mated
rolled prints from NIST SD27, and the first 2, 000 rolled
impressions from NIST SD14 [11]. So, the total number
of background prints is 2, 258. NIST SD27 contains latent
prints of three different qualities, termed “good”, “bad”, and
“ugly”, which were classified by latent examiners. Some
examples of latents from those three qualities are shown
in Fig. 2. Although this classification of latent prints as
“good”, “bad”, and “ugly” is subjective, it has been shown
that such a classification is correlated with the matching per-
formance [6].
Another indicator of fingerprint quality that affects the
matching performance is the number of minutiae in the la-
tent print [6]. Based on the number of minutiae 𝑛 in la-
tents in NIST SD 27, Jain and Feng [6] classified latents in
NIST SD 27 into three groups: large (𝑛>21), medium
(13 <𝑛<22), and small (𝑛 ≤ 13), containing 86, 85, and
87 prints, respectively. We present our experimental results
for each of the six quality groups. We also show results
of the commercial matcher VeriFinger [9] for the purpose
of performance comparison. Although VeriFinger was not
designed specifically for latent matching case, it should be
noted that there is no latent fingerprint matcher SDK nor
forensic AFIS available for individual use. VeriFinger is
widely used as a benchmark in fingerprint publications.
We use manually marked minutiae (provided with NIST
SD 27) as features in latent fingerprints. For rolled finger-
print images, only minutiae are needed for matching and
they are automatically extracted using VeriFinger SDK.
Minutia Cylinder Code (MCC) is used as local descrip-
tors for minutiae. MCC parameters are set as suggested in

10 12 14 16 18 20 22 24 26 28 30
50
55
60
65
70
75
80
85
90
95
100
Alignment Accuracy
Average error in alignment (pixels)
Percentage of Correctly Aligned Latents
Verifinger
Most similar minutia pair (MCC)
Descriptor−based Hough Transform Alignment
Figure 4. Alignment Accuracy: percentage of correctly aligned
latents vs. alignment error.
[1], with the number of cells along the cylinder diameter
as 8 (𝑁
𝑠
). However, we consider all cells in a cylinder as
valid cells. For Euclidean distance pairing, we use two dif-
ferent thresholds, 15 and 25 pixels, and direction difference
threshold of 20 degrees.
In order to estimate the alignment error, we use ground
truth mated minutiae pairs, which are marked by finger-
print examiners, to compute the average distance of the true
mated pairs after alignment. If the average Euclidean dis-
tance for a given latent is less than a pre-specified number of
pixels in at least one of the ten best alignments (peaks of the
DBHT), then we consider it a correct alignment. This per-
formance is shown in Figure 4. The x-axis shows the align-
ment error, and the y-axis shows the percentage of correctly
aligned latent fingerprints in at least one of the ten align-
ments. For comparison, we also show VeriFinger alignment
accuracy, as well as the accuracy of aligning the minutiae
sets based on the most similar minutiae pair (according to
the MCC similarity) - in this case, each alignment is based
on one of the ten most similar minutiae pairs.
There are very few errors in alignment if we consider
the average alignment error of less than 25 pixels. The
main reason for these failure cases is there are a very small
number of true mated minutia pairs in the overlapping area
between the latent and mated rolled print, so there are not
many true mated pairs voting for the correct alignment pa-
rameters. The absence of true mated pairs is due to limited
number of minutiae in latents and the poor quality region in
the rolled print. One such example is shown in Fig. 5.Blue
squares are manually marked minutiae in the latent print
(left) and automatically extracted minutiae in the rolled
print (right). Red triangles indicate ground truth minutiae
pairs, and the yellow lines indicate true mated pairs.
Although the minutiae pairing based on Euclidean dis-
Figure 5. Example of alignment error due to the small number of
true mated minutia pairs in the overlapping area between a latent
and its mated rolled print.
Table 1. Rank-1 accuracies for various subjective qualities of la-
tents in NIST SD27.
Quality VeriFinger (%) Proposed Matcher (%)
All 51.2 62.4
Good 75.0 78.4
Bad 47.0 55.3
Ugly 30.6 52.9
Table 2. Rank-1 accuracies for various objective qualities of latents
in NIST SD27.
Quality VeriFinger (%) Proposed Matcher (%)
All 51.2 62.4
Large 79.0 81.4
Medium 50.6 67.0
Small 24.1 39.0
tance and direction difference is relatively simple, it works
well after the fingerprints are aligned using the Descriptor-
based Hough Transform. Our approach performs better than
the commercial matcher VeriFinger on manually marked
minutiae; performance of both these matchers are shown
in Fig. 6. We also show our results for latents of six differ-
ent quality levels (good, bad, ugly; large, medium, small)
separately. The rank-1 accuracies for the proposed matcher
and VeriFinger are shown in Table 1 for the three subjective
qualities and Table 2 for the three objective qualities.
Figure 6 shows that the advantage of our algorithm over
the commercial matcher is consistent throughout the match-
ing ranks. We can also notice that the improvement is
more clearly observed on latents of poor quality and with
small number of minutiae. The improvement of the pro-
posed matcher over VeriFinger at rank-1 accuracy varies
from 2.3% for latents with a large number of minutiae to
22% for latents of ugly quality. Figure 7 shows examples
of latent prints of good (medium) and ugly (small) qualities
correctly identified at rank-1, and Fig. 8 shows examples of
latent prints incorrectly identified at higher ranks because

Frequently asked questions

Q1. What are the contributions mentioned in the paper "Latent fingerprint matching using descriptor-based hough transform" ?

The proposed approach consists of following three modules: ( i ) align two sets of minutiae by using a descriptor-based Hough Transform ; ( ii ) establish the correspondences between minutiae ; and ( iii ) compute a similarity score. 

Q2. What future works have the authors mentioned in the paper "Latent fingerprint matching using descriptor-based hough transform" ?

The authors plan to improve the score computation by applying learning methods. The authors plan to incorporate extended features which are automatically extracted from the image into the current matcher to further improve the matching accuracy. 

Q3. What is the common way to align two fingerprints?

Since manually marking minutiae is a common practice for latent matching, their approach to align two fingerprints is based on minutiae. 

Q4. What are the features that may be used to estimate alignment parameters between two fingerprints?

There are a number of features that may be used to estimate alignment parameters between two fingerprints, including orientation field, ridges and minutiae. 

Q5. What is the process of calculating the parameters that align two fingerprints?

Fingerprint alignment or registration consists of estimating the parameters (rotation, translation and scale) that align two fingerprints. 

Q6. What is the way to mark latent features?

marking extended features (orientation field, ridge skeleton, etc.) in poor quality latents is very time-consuming and might be only feasible in rare cases. 

Q7. How do the authors estimate the distance between the mated minutiae pairs?

In order to estimate the alignment error, the authors use ground truth mated minutiae pairs, which are marked by fingerprint examiners, to compute the average distance of the true mated pairs after alignment. 

Q8. How many minutiae can be computed for each set?

For each pair of minutiae, one minutia from each set, and for given scale and rotation parameters, unique translation parameters can be computed. 

Q9. What was the purpose of using fingerprints to identify criminals?

fingerprints collected from crime scenes were compared to fingerprints collected from previous offenders so that they could identify repeat offenders, criminals who have been previously arrested. 

Q10. How many rolled impressions are in the background database?

The background database consists of 258 mated rolled prints from NIST SD27, and the first 2, 000 rolled impressions from NIST SD14 [11]. 

Q11. Why do latents have a smaller number of minutiae than plain prints?

Due to their poor quality and small area, latents have a significantly smaller number of minutiae compared to rolled or plain prints (the average number of minutiae in NIST Special Database 27 (NIST SD27) [12] images is 21 for latents versus 106 for the corresponding rolled prints). 

Q12. How many pixels are used for the distance pairing?

For Euclidean distance pairing, the authors use two different thresholds, 15 and 25 pixels, and direction difference threshold of 20 degrees.