DSpace at VNU: An Efficient Cascaded System for Latent Fingerprint Recognition

2013 IEEE RIVF International Conference on Computing & Communication Technologies Research, Innovation, and Vision for the Future (RIVF)

An Efficient Cascaded System
for Latent Fingerprint Recognition
Nguyen Thi Huong Thuy1, Hoang Xuan Huan2, Nguyen Ngoc Ky1 and Le Minh Khoi2
1

General Department of Technique - Logistic, Vietnamese Ministry of Public Security
,

2

Faculty of Information Technology, Vietnam National University – Hanoi, University of Engineering and Technology
,

Abstract - This paper proposes a cascaded scheme to improve
the efficiency of latent fingerprint identification system. In this
scheme, the feature set of latent fingerprint such as finger codes,
basic patterns, ridge counts and minutia with their local
structures are sequentially exploited in four cascaded layers. In
the first layer, possible finger codes of latent fingerprint are
recognized based on basic pattern features and then reordered to
determine their database access priority. In the second layer, its
minutiae are extracted and assessed to affine matching in the
third layer. In the fourth layer, any case having too many
candidates in the previous layer is further matched based on
exploiting local structure in order to downsizing the result list.
On the verification layer, the minutiae information and local
structure of corresponding pairs are presented to human experts
for further verification. Experimental results on C@FRIS
database show that our proposed method obtains high matching

accuracy and considerably low identification time.
Keywords: AFIS, C@FRIS, latent fingerprint identification,
fingerprint verification, cascading, minutiae.
I. INTRODUCTION
Fingerprint-based identifications have been applied
efficiently in forensics applications over a century [4].
However, there are many open problems attracting researchers
in this field [1]-[5], [7]-[9]. While verification and automatic
authentication for rolled and plain fingerprints have achieved
tremendous progress, latent fingerprint recognition faces
difficult problems [1], [7]. Latent fingerprints from relative
careless inadvertent individuality on objects are usually of
poor quality because of noise and non-linear distortion.
Therefore, it is difficult to match them. In addition, in
Vietnam and many other countries, along with latent
fingerprints collected at crime scenes, databases mainly store
paper-thin fingerprints or scanned paper-thin fingerprints
which are more difficult to process than sensor fingerprints.
In Automatic Fingerprint Identification System (AFIS),
matching two fingerprint images is one of the main tasks. It
decides the performance of the system [7]. In the literature,
many matching techniques have already been proposed [2],
[7]. Nevertheless, the techniques yielding high performance
usually require large amount of search time.
One of the prospective solutions to reduce the identification
time is to apply cascading technique which integrates many
algorithms from simple to complex into an AFIS [3], [13].
This paper proposes a 4-layers cascaded architecture for latent
fingerprint identification system. In the first layer, latent
fingerprint is recognized by its possible fingers [10] based on

basic pattern features and then reordered (to determine the
database access priority on the third layer); In the second
layer, minutiae and their local ridge-valley structure will be
extracted and assessed for affine matching in the third layer.
If the matching result is doubtful, matching algorithm based

978-1-4799-1350-3/13/$31.00 ©2013 IEEE

on P-TPS model improved from [9] will be applied to decide
before verification. P-TPS matching helps eliminate the nonlinear distortion effectively. In order to decrease identification
time, fingerprints in database are organized and indexed by
finger codes, basic fingerprint pattern. Moreover, the minutiae
matching process is parallelized on a computer cluster.
Experimental results on the database C@FRIS DB show
that our new proposed system provides better outcomes than
that of the earlier version of C@FRIS (built by research group
of the Vietnamese Ministry of Public Security, awarded
VIFOTEC 2008 and upgraded in 2009).
The rest of the paper is organized as follows. Section 2
briefly introduces latent fingerprint recognition and some
basic techniques for cascaded architecture. The method
applying P-TPS technique is described in Section 3. The
scheme of the cascaded system and the organization of the
database for parallelized searching are highlighted in Section
4. Section 5 presents experimental results compared with
C@FRIS system. Finally, Section 6 concludes the paper.
II. LATENT FINGERPRINT RECOGNITION AND RELATED WORKS
This section briefly introduces latent fingerprint
identification system. It also describes some techniques such
as finger recognition, fingerprint classification, matching

minutiae, TPS warping model.
A. Latent fingerprint recognition and identification
1) Latent fingerprint matching problem
The latent fingerprint matching problem is described as
follows: Given a query fingerprint Iq (latent) and a fingerprint
database, it is to determine whether the database contains the
genuine fingerprint of this query fingerprint or not. If yes, the
system will display it.
Latent fingerprints in nature are often of poor quality and
not complete as rolled/plain fingerprint. Thus, it is difficult to
use them as inputs for an automatic recognition system.
Therefore, the identification process is usually divided into
two layers: identification by computer and visual verification
by human being [7]. The identification layer aims at finding
the fingerprint images in the database which are most similar
to Iq. This layer is usually done by AFIS. After having the
images outputted by the first layer, the verification layer is to
identify among them which one is genuine with Iq. This layer
is often performed by human experts, either with or without
computer assistance. Examiners often stop once getting the
first match.
2) Latent fingerprint identification system
While automatic fingerprint identification systems work
well with rolled/plain fingerprints, latent fingerprint
identification still remains a challenging task and attracts

123


attention from the research community [1], [2]. In order to

solve the latent fingerprint problem with high accuracy and
decreased search time, it is necessary to combine different
techniques (see [13]). Classification and cascading is one of
the most efficient methods which help downsize the matched
fingerprint list from database. Moreover, fast matching
method helps remove most dissimilar fingerprints with Iq to
focus on most similar fingerprints. Specifically, a matching
method needs to effectively process non-linear distortion.
This paper proposes an AFIS using cascaded architecture
with the following components: fingerprint classification,
finger recognition, parallelized minutiae matching based on
reasonable organizing database for matching and verification
assistance. In the following sections, these components are
going to be briefly introduced.
B. Finger recognition based on fingerprint trace
In order to determine the finger or the order of possible
fingers of suspect which left the latent trace, K. Nguyen Ngoc
[10] proposed a statistic method using maximum a posteriori
probability (MAP) for finger recognizing based on latent
fingerprint trace.
Based on this method, the identification process can be
conducted first on the suspected fingerprint instead of all
fingerprints. This significantly reduces the search time.
C. Fingerprint classification
In order to accelerate the identification process, fingerprint
images are often classified into basic patterns according to
local ridges and relative position of singular points [4], [6],
[11], [12]. Identification only applies to fingerprints of the
same type with Iq in the database. The FBI [7] proposed
classifying fingerprints into three basic patterns: the arch, the

loop and the whorl.
In order to improve fingerprint classification performance
according to the FBI’s standard, Karu [4], [7] proposed a
solution using Poincare index for detecting minutiae.
However, the drawback of this method is that it requires
complete fingerprint and local ridge orientation of core region
and delta region which need to be clear.
To eliminate the limitations, Wang [11] proposed a
classification algorithm which is only based on core points
and directions around the core point. Nevertheless, an exact
identification of the core points is required.
Our proposed system combines the two methods of Karu
[4] and Wang [11] which are going to be described in Section
4.1 to classify a fingerprint.
D. Minutiae based fingerprint matching
Given a query fingerprint Iq and a template fingerprint It, it
is to find out whether they originate from the same finger or
not. In all fingerprint matching algorithms [7], matching based
on minutiae is simple but yet efficient and therefore is widely
used.
1) Minutiae based method

In a fingerprint image, points representing discontinuities
of fingerprint local structure such as end points, bifurcation
points are called minutiae.
In order to match fingerprints using sets of minutiae, two
fingerprint images Iq, It have to be pre-processed by extracting
and assessing features.
2) Matching scheme based on minutiae
Let nt and nq be the number of minutiae on the query

fingerprint and sample fingerprint, respectively. Assume that
there are n corresponding minutiae pairs found from two
images. The similarity of two fingerprint images is
characterized by the measurement S(It,Iq) and given by the
following formula:
S(It,Iq) = n2/(nt×nq). (1)
The simplest and most general transformation used in
matching methods to align two images is the affine
transformation. However, due to the nonlinear distorted nature
of latent fingerprints, the efficiency of this method is
insufficient and is usually employed to determine initial
corresponding minutiae pairs for advanced warping methods
[2], [7]. One of widely used warping transformations is ThinPlate-Spline (TPS) deformation model in [5].
E. Thin-Plate-Spline deformation model
After having determined the n pairs of corresponding
minutiae by using affine transformations for creating an initial
set of landmark points, our system warps the image by the
TPS model [5], [7].
III. PARTIAL TPS FINGERPRINT MATCHING METHOD
In order to deal with the non-linear distortion problem, the
authors in [9] proposed a partial TPS warping method using
an additional technique to enrich the set of landmark points by
appending more pseudo-minutiae belonging to the associated
ridge-valley pair. Our experimental results verified that this
matching based on local point model helps solve the nonlinear distorted problem efficiently. In this paper, P-TPS
technique is used for building improved matching algorithm
for doubtful fingerprint image pairs which are the outcomes of
the affine matching.
IV. CASCADED ARCHITECTURE AND PROCESSING DATA
This section introduces the constituent components of the

cascaded architecture and the identification. It also highlights
the organization of the fingerprint database.
A. Components of the new system
The system is made of four linked components/modules as
described in Fig. 1:
i) Finger classification: Recognizing finger priority and
classifying to basic patterns to match in the third and the
fourth layer.
ii) Feature extraction: Extracting minutiae, ridge-valleys
structures to be used for matching in the next layers.
iii) Affine matching: Performing minutiae matching by
affine alignment as in Section 2.4.

124


iv) P-TPS matching: Performing P-TPS matching
according to the algorithm in [9].
Among the above components, it is necessary to present
more details about classification method in the first
component.
Fingerprint classification
Forensic fingerprints and registered fingerprints in database
are classified into 10 types by combining the two methods of
Karu [4] and Wang [11]. If this procedure produces
ambiguous outcomes, Karu’s method will be used. Then, if
the outcomes are still ambiguous, basic ridge will be taken
into consideration. If ridges at the core region are in good
quality, Wang’s method is applied [11]. Otherwise, delta
points will be searched using Karu’s method [4]. When the

core region and the delta points are unclear, we use basic
ridge to validate ambiguous cases. This method helps
eliminate incomplete and unclear minutiae.
Identifying basic ridge is quite straightforward by analyzing
curves which are represented in vectorized form. By doing
this, the classification accuracy is increased up to 95%.
Nevertheless, it is still insufficient to apply the cascaded
combination technique.
In order to improve the reliability, in ambiguous cases,
fuzzy recognition technique type II will be employed. If the
system cannot classify into a specific class, it will display a
list which is sorted in decreasing order of the classification
reliability. By doing this, the classification layer always
provides the next layer with a controllable input.
B. Cascaded identification scheme
With the above components, the identification process is
performed sequentially as below:
Step 1: After the acquisition phase, latent fingerprint Iq is
pre-processed and classified. Ridge counts are calculated and
the finger order is determined. As latent fingerprints normally
have a poor quality, it is difficult to extract minutiae
automatically. Therefore, the fingerprint image will be
interactively edited by human experts to improve image
quality. This is done by the assistance of a graphic tool. Based
on this finger order, classified basic pattern and automatic preextracted features of the template fingerprint It are then
retrieved from database for matching (in Step 4).
Step 2: In the classification module, basic pattern of Iq is
determined either automatically or manually. If there are
ambiguous outcomes, they are used for searching for the
match by the priority of the reliability. Afterwards, Iq is passed

to the feature extraction module.
Step 3: In the feature extraction module [7], [14], the features
of Iq including minutiae points and ridge-valley associated
structure with quality map are extracted by interactive editing
method. The results are then passed to the affine matching
module (Step 4) and the P-TPS matching module (Step 5). If Iq
has such poor quality that the system cannot perform minutiae
extraction on it, the process stops and delivers notification.
Step 4: In the affine matching module, the features of
fingerprint It in the database are sequentially retrieved to
match with the features of Iq (taken from Step3) for

calculating the initial corresponding minutiae set and the
similarity S(It,Iq):
4.1. If S(It,Iq) < Smin then the two fingerprints are not
matched. The system proceeds with the next fingerprint It in
the database.
4.2. If S(It,Iq) >Smax the system appends It to the output list.
Otherwise, that means S(It,Iq) ∈ [Smin, Smax], the system passes
the corresponding minutiae pairs set and their associated
ridge-valley pairs to P-TPS matching.
Step 5: P-TPS matching. Continue with Step 4 until all
features in the database that have the same finger code and
pattern with those of Iq are matched.
Step 6: Sorting the search result list according to priority:
finger code, basic pattern code, ridge count, similarity. The
system verifies based on the priority and displays the results.
In the output list, corresponding minutiae pairs and
associated ridge-valley pairs are displayed sequentially; each
pair of fingerprint image Iq, It is sorted in decreasing order of

the similarity of sub-patterns (based on finger code, basic
pattern, ridge count) on display screen. There are many useful
tools for assisting human experts in verifying the results.
With the identification process above, the first three steps
could run in parallel. However, that would not improve the
running time efficiently. Therefore, this paper only proposes
parallelized matching method for Step 4 and Step 5. These are
the two steps that account for the most time. So they are
implemented by a computer cluster with many nodes process
in parallel. The cascaded architecture is described in Fig. 1.

Fig. 1: The cascaded architecture scheme

C. Organizing the database
In applications such as identity card and criminal card, each
fingerprint is represented as one record having the following
basic fields:
- Identity Card number;
- Personal information fields (full name, date of birth, sex,
address);
- Finger code;
- Basic pattern;
- Left ridge count, middle ridge count, right ridge count,
ridge density;
- Minutiae set;
- Fingerprint image (standard resolution 500 dpi, about 5 MB
for a 10-finger set);
To accelerate the searching and matching processes, it is
necessary to organize the fingerprint database in a sensible
fashion. In the database, fingerprints are indexed and

organized hierarchically according to the following fields:
fingerprint code, basic pattern, ridge counts and ridge density.
Information about minutiae points and associated ridge-valley

125


pair is stored with the corresponding fingerprint image for
faster parallel searching.
The extraction of basic attributes from fingerprint with
higher reliability for indexing remains an open research
problem [2]. To reduce the false rejection rate (FRR), fuzzy
search method combining major code and minor codes for
coding ambiguous attributes has been proposed. For example,
a ridge has whorl pattern with low ridge count can be
classified incorrectly into a loop ridge. Hence, we need to
search both whorl ridge and loop ridge with major code
corresponds former searched whorl pattern and minor code
corresponds later searched loop pattern.
D. Parallel matching
For parallel matching solution, we setup a computer cluster
system with the following components:
i. The server receives matching requests and performs
searching according to the basic attributes. It then splits the
list into small pieces and distributes them to parallel
processing nodes for minutiae matching.
ii. Parallel processing nodes receive task and perform
matching. They deliver results in terms of a search result list
to the server.
iii. Workstations receive the search results from server.

They display the results to human experts who then perform
the final verification.
V. EXPERIMENTAL RESULTS
Our experiment compares the performance of the new
system with C@FRIS version 2009. A criminal identity
database of C@FRIS system applied at the Police Office of
Hanoi City contains 2.500.000 one-finger fingerprint cards
with the standard resolution of 500 dpi. The hardware system
for performing the experiments consists of one mid-range
server, five PCs linked together by the star network topology
(parallelize with k=5). The performance of the system is
determined by the length of identification result list and the
search time. In practice, the search time of C@FRIS system
and that of the new system when they use sequential
processing are equal, therefore we only compare the search
time when both systems are running in the parallel processing
mode.
Sixty-four latent fingerprints with average or quite good
quality have been matched against the database.
The result list is sorted by finger priority order, basic
pattern code order, ridge count order and the descending
similarity (in the same group). The length of the search result
list is the number of records on this list. In practice, only 710% can be found in the database. Hence, the length of the
real list is equal to the number of records on searched
fingerprint cards. The remainder 90-93% unfound fingerprint
traces are from people without a criminal record. In those
cases, examiners need to spend more time to verify until the
end of the list. Thus, sorting the result list and determining the
maximum length of the real list for both found and unfound
cases play an important role.

The experimental results show that the proposed method
gains a high performance. It increases search speed
significantly. In addition, it reduces on average 66.2%
verification time per request compared with the case when the
cascaded configuration is not applied.
VI. CONCLUSION
This paper proposes a cascaded architecture for latent
fingerprint recognition system. By applying combined
fingerprint classification techniques, finger code recognition,
ridge count, we were able to dramatically shorten the
matching list. Moreover, applying partial TPS matching
method helps reduce the effect of non-linear distortion
phenomenon.
Thanks to the parallel matching process, search time has
been significantly decreased. The experimental results show
that the new system gains higher performance and better,
response time in large database.
ACKNOWLEDGEMENT
This work is partly supported by Vietnam National
Foundation for Science and Technology Development.
REFERENCES
[1]
[2]
[3]

[4]
[5]

[6]

[7]
[8]
[9]

[10]
[11]
[12]
[13]

[14]

126

A.K. Jain, J. Feng, “Latent fingerprint matching”, IEEE Trans. Pattern.
Anal Intell, pp. 88–100 (2011)
T. Y. Jea (2005), “Minutiae based partial fingerprint recognition”. PhD
thesis of the University at Buffalo, the University of NewYork.
Q. Jin, Z. Shi, X. Zhao and Y. Wang (2004), “Cascading a couple of
registration methods for a high accurate fingerprint verification
system“, Proceedings of SINOBIOMERTRICS, Guangzhou, China,
pp. 490-497.
K. Karu and A. Jain (1996), “Fingerprint classification“, Pattern
Recognition, Vol.29, no.3, pp. 389-404.
J. Li, S. Tulyakov, Z. Zhang, V. Govindaraju (2008), “Fingerprint
Matching Using Correlation and Thin-Plate Spline Deformation
Model”, 2nd IEEE Conference on Biometrics: Theory, Applications,
and Systems (BTAS 08), Washington, pp. 1 – 4.
S. U. Maheswari, Dr. E. Chandra (2012), “A Review Study on
Fingerprint Classification Algorithm used for Fingerprint Identification
and Recognition”, IJCST Vol. 3 no. 1, pp. 739-744.

D. Maltoni, D. Maio, A. K. Jain, S. Prabhakar (2009), “Handbook of
fingerprint recognition”, Second ed, Springer-Verlag.
M. A. Medina-Pérez, M. García-Borroto, A. E. Gutierrez-Rodriguez, L.
Altamirano-Robles (2012), “Improving Fingerprint Verification Using
Minutiae Triplets”, Sensors, vol. 12, pp. 3418-3437.
T.H.T. Nguyen, H. Hoang Xuan and K. Nguyen Ngoc (2013), “An
Efficient Method for Fingerprint Matching Based on Local Point
Model”, Proc. of the International Conference on Computing,
Management and Telecommunications (ComManTEL2013), January
21-24, 2013, in Ho Chi Minh City, Vietnam, pp. 334-339.
K. Nguyen Ngoc (1997) “A Fingerprint-based Method for predicting
finger”, Journal of Public Security, No 1, pp. 25-27 (in Vietnamese).
S. Wang, W. W. Zhang, Y. S. Wang (2002), “Fingerprint Classification
by Directional Fields“, Fourth IEEE International Conference
Multimodal Interface, Pittsburgh, PA, pp. 395-399.
Q. Zhang, K. Huang and; H. Yan (2001), “Fingerprint Classification
Based on Extraction and analysis of Singularities and Pseudo ridges“,
Proceedings Selected papers from VIP2001, Sydney, Australia, pp. 83-87.
S. Zia, S. K. Soni, S. Sweta, P. Mokal (2011), “A Casscaded
Fingerprint Quality Assessment Scheme for Improved System
Accuracy”, International Journal of Computer Science Issues, Vol. 8,
Issue 2, pp. 449-455.
Neurotechnology,
Inc,
Verifinger
4.2
SDK
.