Fusion Based Multimodal Biometrics (Fingerprint and Speech)
Er.Upasana Dutta1, Er.Shikha Tuteja2, Dr.Himanshu Monga3
M.tech scholar1, Assistant professor2, H.O.D (ECE)3 at RBIEBT, Mohali
upu01.jmu@gmail.com1, ershikha_tuteja@gmail.com2, himanshumonga@gmail.com3
ABSTRACT: In this paper we have
increased the accuracy of biometrics
using two modalities one is fingerprint
and the other is speech. Fusion of these
two modalities is experimented for
authentication purpose. The features of
fingerprint and speech have been
extracted which are fused and saved as a
database. Later various samples are tested
with this database and finally various
parameters like false acceptance rate
(FAR), false rejection rate (FRR) along
with accuracy of the system has been
calculated.
KEYWORDS: False acceptance rate,
False rejection rate, Genetic algorithm,
Independent component analysis, Mel
frequency cepstrum coefficient.
1. INTRODUCTION
Authentication systems are heading
towards biometrics, which has many
demerits like handling large databases,
difficult biometric computations and
above all the results are not fast and
accurate. To minimize these problems,
multimodal biometrics has been adopted.
The systems using passwords, pin codes,
etc for authorization purposes have a
disadvantage that other persons may
know their secret numbers or even they
can guess that easily [J. K. Anil, R. Arun,
P. Salil, (2004)]. This has lead to a new
technique called biometrics in which one
of the feature modality such as iris,
fingerprint and palm-vein has been used
as a person’s identity that will check for
his/her authorization further [Chowdhary.
V, Monga. H, (2014)]. But these systems
also suffer from problems like sometimes
one’s single feature is not readable
properly like if anyone has a scar or cut
over the finger that feature would not be
appropriate for identification, also
sometimes some features could be
matching for two people like twins
having same faces. In that case
multimodal biometrics has given more
accurate results. We have worked on the
fusion of two modalities one is fingerprint
that is physiological feature other is
speech that is behavioral feature. Now
fusion of these two modalities can be
done in many ways feature level fusion,
score level fusion and decision level
fusion [U. Oleg, N. Sergey, (2006)]. We
have opted feature level fusion due to its
supremacy, as in this technique fusion is
performed on the features extracted using
appropriate algorithms and later samples
are tested with the trained fused dataset.
2. METHODOLOGY
In this proposed work a fingerprint
recognition
system
and
speaker
verification system has been combined as
these modalities are widely accepted and
natural to produce. The major issues are
the degree at which features are extracted
and the cost factor involved. As the
number of features increases the
variability of the intra-personal samples
due to greater lag times in between
consecutive acquisitions of the sample
also increases. Increase in variability of
the system will further increase FAR.
Thus to resolve these issues an effective
fusion level and fusion mode is required.
The proposed work has presented a novel
user authentication system based on a
combined acquisition of fingerprint and
speech signals with high accuracy rate.
A. Speech Feature Extraction: Speech
recognition system works on a
technology in which speech signals are
processed
by computers
to
do
authentication and other related tasks.
Today speech has an application in
almost every areas like household
appliances
(television,
radio,
communication RADAR), in education
(improving
pronunciation,
studying
online lectures), in toys (which work on
voice instructions), in information
retrieval systems and the list is endless
[G. Christos, P. Stavros, P. Maja, (2014)].
The algorithm which we have used for
extraction of speech sample features is
MFCC
(Mel
frequency
cepstral
coefficient; the other such algorithms
which can be used for the same are LPC
(linear predictive coding), LPCC (linear
predictive cepstral coefficient). But we
have opted MFCC as it gives more
accurate and superior results than other
two [K. G. Shashidhar, R. K. Sreenivasa,
(2012)]. The purpose of feature extraction
is to represent speech signal in the form
of finite number of values, so that further
comparison can be completed. These
feature values generated are very large in
number so we have used GA (genetic
algorithm) as an optimization technique.
B. Fingerprint feature extraction: The
very first modality that was adopted for
biometrics and became most popular was
fingerprint. The reason could be its easy
availability, uniqueness, ease of feature
extraction and people have no security
concern to their hygienity in case of
giving fingerprint samples [Z. Qijun, Z.
Lei, Z. David, L. Nan, (2008)]. The
fingerprint uniqueness is because of
different shapes of minutiae. There are lot
of ridges and bifurcations on the finger
tips which are unique for everyone. The
fingerprint recognition can be achieved
by minutiae or pore extraction but we
have adopted an algorithm ICA
(independent component analysis) to
reduce the complexity. ICA is a data
analysis tool derived from the "source
separation" signal processing technique
[W. Xiaoyong et al, (2009)]. The basic
idea of ICA is to represent a set of
random variables using basic component,
where the components (basic functions)
are statically as independent as possible.
Other such well known algorithm is PCA.
Pseudo code of ICA:
start
upload image
im= image
u=0; where u is initial point of selection,
bit_count=1
while (bit_count<im bits.count)
threshold=ICAselection(im(bitcount:bitco-unt+6))
bit_count= bit_count+6;
if bit.value<threshold_value
ignore
else
independent points= ICA bit.points
end if
end while
C. Fusion strategy and classification:
When
implementing
multimodal
authentication system fusion has an
important role. The features of all the
single biometrics are fused [R. Arun, J.
Anil, (2003)], that fused database is used
further for testing of the sample. There
are certain types of fusion, [E. Youssef,
E. Zakaria, A. Reda, (2014)] out of them
we have opted feature level fusion i.e.,
the features of two independent
modalities are extracted and then fused
using sum rule to give the final dataset.
The classification technique used at the
testing stage is Euclidean distance. At the
end FAR, FRR and Accuracy of the
system performance has been calculated.
Step2: The testing panel has speech and
fingerprint sample upload, fusion and
testing button.
Step3: If the result is matched fusion
graph shows straight line otherwise the
difference is plotted.
Figure 2 Result Graph if sample does not match
3. RESULTS AND DISCUSSION
We have used GUI (graphical user
interface) platform for implementing this
proposed work.
Step1: A new GUI has been designed
having training and testing panel, the
training panel has speech and fingerprint
processing option along with fusion
button and then their fused results have
been saved in the database as record sets.
The speech GUI panel has data upload,
MFCC and GA options likewise
fingerprint panel has ICA option.
Figure 1 GUI
Figure 3 Result Graph if sample is matched
Step4: At the end parameters such as
FAR, FRR and Accuracy of the
authentication system has been calculated
using the following formulae.
Error=
(sqrt((sum(sum(fused_value))sum(sum(test_fused)))^2/(rs*cs*35)));
FAR= (error/(i*rs*cs));
(1)
FRR= (error-FAR)/(i*rs*cs);
(2)
Accuracy= (1-(FAR+FRR))*100;
(3)
Figure 4 Parameter evaluated (FRR)
Figure 5 Parameter evaluated (FAR)
Figure 4 Parameter evaluated (Accuracy)
4. CONCLUSION
The experimental work on proposed
model with multimodal biometric
authentication based on fusion technique
showed better results compared to
unimodal biometric technique in terms of
high accuracy, low false acceptance and
false rejection errors. The obtained results
are clearly in support of fusion based
multimodal biometric system for our
proposed model.
REFERENCES
[1] J. K. Anil, R. Arun, P. Salil, (2004)
"An
introduction
to
biometric
recognition", IEEE Transactions on
Circuits and Systems for Video
Technology, Special Issue on Image- and
Video-Based Biometrics, Vol. no.14,
Issue no.1, pp: 1-29.
[2] Chowdhary. V, Monga. H, (2014)
“Human Identification Using Palm-Vein
Images Using Gabor Filter and Laplacian
Palm”,
International
Journal
of
Engineering Sciences and Research
Technology, Vol. no.3, issue no.8, pp:
116-121.
[3] U. Oleg, N. Sergey, (2006)
“Biometric Fusion: Robust Approach”,
Article at Biolink Technologies, pp: 4-9.
[4] G. Christos, P. Stavros, P. Maja,
(2014)
"Visual-only
discrimination
between native and non-native speech",
IEEE International Conference on
Acoustics, Speech and Signal Processing
(ICASSP), pp: 4828-4832.
[5] K. G. Shashidhar, R. K. Sreenivasa,
(2012) “Emotion recognition from
speech: review”, International Journal of
Speech Technology springer science+
business media, Vol. no.15, pp: 99-117.
[6] Z. Qijun, Z. Lei, Z. David, L. Nan,
(2008) “Adaptive Pore Model for
Fingerprint Pore Extraction”, 9th
International conference on Pattern
Recognition (IEEE), pp:1-4.
[7] W. Xiaoyong et al, (2009) “A novel
approach of fingerprint recognition based
on multilinear ICA”, IEEE conference on
Network infrastructure and digital
content, pp: 740:744.
[8] R. Arun, J. Anil, (2003) "Information
fusion in biometrics", Pattern recognition
letters Elsevier, Vol. no.24, Issue no.13,
pp: 2115-2121.
[9] E. Youssef, E. Zakaria, A. Reda,
(2014) “Score Level Fusion Based
Multimodal Biometric Identification
(Fingerprint & Voice)”, IEEE Systems
Journal, Vol. no.3, Issue no.4, pp: 146150.