See discussions, stats, and author profiles for this publication at: [Link]

net/publication/275346786

Integrated Multi-Stage Biometric System Design

Article in International Journal of Applied Engineering Research · January 2015

CITATION READS

1 525

4 authors, including:

Ravi Lakshmanan Assoc. Prof. Dr. Sathish Kumar Selvaperumal

Asia Pacific University of Technology and Innovation Asia Pacific University of Technology and Innovation
34 PUBLICATIONS 197 CITATIONS 87 PUBLICATIONS 381 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ravi Lakshmanan on 24 April 2015.

The user has requested enhancement of the downloaded file.

International Journal of Applied Engineering Research
ISSN 0973-4562 Volume 10, Number 4 (2015) pp. 9611-9629
© Research India Publications
[Link]

Integrated Multi-Stage Biometric System Design

1
Ravi Lakshmanan,2 Dr Sathish Kumar Selvaperumal, 3 Chow Hin Mun
1,2
Senior Lecturer, School of Engineering,
Asia Pacific University of Technology and Innovation, Kuala Lumpur, Malaysia.
3
IT support Engineer, Fusionex Corp Sdn Bhd, Kuala Lumpur, Malaysia
E-mail: 2skspresearch@[Link],1ravi@[Link],3chm1990@[Link]

Abstract

In this paper, an integrated multi stage biometric system design is proposed using
speech, face and finger print along with Artificial Neural Network (ANN). The voice
recognition subsystem utilises Mel-frequency Cepstrum Coefficient (MFCC) feature
extraction algorithm to represents the voice signal of the user in the form of cepstrum
of the voice derived from Discrete Cosine Transform (DCT) on the log power
spectrum on a non-linear Mel-scale frequency. Face recognition subsystem utilises
closing algorithm image morphology that extracts the features of the face based on the
facial outline, via superimposing of the dilation and erosion of the facial image.
Fingerprint recognition subsystem utilises minutiae extraction to extract the ridge
termination and bifurcation points of the fingerprint. The integrated biometric
recognition system consists of three neural networks configured consistently with 50
hidden neurons. The simulation result yielded an overall False Rejection Rate of 4%
and False Acceptance Rate of 8%. Also, the calculated average time taken to train the
neural network is 3.848 seconds only.

Keyword- Artificial Neural Network, Fast Fourier Transform, Direct Cosine

Transform, Mel Frequency Cepstrum Coefficient.

INTRODUCTION
In the past two decades, electronic security systems have been extensively researched
and developed, due to the increasing demands for high level security protection and
encryption. With gradual advancement of technology, specifically in the field of Very
Large Scale Integration (VLSI) technology applied in embedded system, security
systems are applied widely in various areas, from software computer-based access
control to hardware-based sophisticated vault lock, replacing the traditional
mechanical key and lock system. Hence, a wide range of security systems has been
developed and implemented to prevent unauthorised access, hacking and protection of
intellectual property.
9612 Ravi Lakshmanan et al

Traditional electronic security systems are implemented using user-password

based approach, where users are required to key in their username and an assigned
password for verification. This type of security systems are usually found in software
computer-based application such as user profile access and online banking account
management.
Radio frequency identification (RFID) and token-based security systems are
applied on physical hardware security systems such as door lock access. Similar to
key and lock system, users are required to use an object such as ID cards, tags, etc. to
identify and verify themselves as the authorised user to gain access.
The next generation of security systems is based on verification of biometric
signature. Biometric refers to the measure of biological characteristics of the human
body. Biometric characteristics are unique among every human and remain permanent
to the human, so as the means to identify the user is always available, cannot be lost
unlike the previously mentioned security system (Gregory & Simon, 2008).
The main concern of all security systems is the integrity of the system towards
protection of users‘ data and assets. Problems and limitations of traditional user-
password based systems are generally associated with the lack of uniqueness of the
password. To prevent this, users are encourage to set up complicated password,
however this make the password hard to remember and easily forgotten. RFID-based
security system eliminates the problem of remembering password, where access cards
unique to each user are provided and used for identification. However, users are not
able to gain access to the system if they lost their access card. This system is also
vulnerable to unauthorised access if the access card has been stolen or reproduced by
unauthorised user.
With the advent of biometric security system, problems and limitations in
previous security systems can be solved. Biometric signatures act as a unique feature
to identify users and remain part of the users‘ body. However, the problem in the
current use of single biometric system can still be easily hacked using fake
credentials. This is known as replay attack (Gregory & Simon, 2008), where an
intruder is able to perform a correct authentication to access the system by capturing
the biometric data of the authorised user. In the case of fingerprint recognition, fake
rubber fingers can be easily imprinted with the authorised fingerprint of the user,
which is used as a faked credential to gain access of the system.
Matching flaws have been a main research problem in biometric security
system, where excessive False Acceptance Rate (FAR) and False Rejection Rate
(FRR) will cause the system to incorrectly recognise users. FAR refers to the
percentage of accepting unauthorised user, while FRR refers to the percentage of
rejecting authorised user (Vasuhi et al., 2010). A lot of researches have been focus on
developing robust and accurate biometric recognition system to improve on the FAR
and FRR values. Various intelligent systems such as fuzzy logic and neural networks
have been applied to improve biometric recognition rate.
In addition, the use of biometric recognition system is nearly universal to
everyone. However, using only a type of biometric measure will limit the scope of
usage of the system. Take an example of voice recognition system, if the person is
mute, voice recognition cannot be perform on that person. In addition, certain cases
Integrated Multi-Stage Biometric System Design 9613

do not provide the proper environment to perform specific type of biometric measure,
such as fingerprint system cannot be implemented in dirty environment or workplace.
Therefore, there is a need to implement multiple biometric measures to provide usage
and flexibility to those situations.
The identified research problems in existing biometric security system are
vulnerability to replay attacks, matching flaws and constrained usage in single type
biometric. Therefore, by integrating multiple biometric measures into a multi-stage
system with the use of intelligent neural network pattern recognition approach, the
proposed system is able to strengthen the level of security, improve on recognition
rate and provide usage flexibility in biometric measures.
After the introduction in section I, the rest of the paper is organized as follows.
The implementation of the proposed method is detailed in section II. The
experimental results are discussed in section III and the simulated results are
discussed in section IV. Finally, the conclusion is concluded in section V.

II. PROPOSED METHOD

Figure 2.1 shows the block diagram of the proposed design for the intelligent bio-
metric security system. The approach in proposing a model for the intelligent
biometric security system is based on segmenting the overall system into three
recognition subsystems. The important design aspects of the recognition subsystem
are the feature extraction algorithms and the configuration of the neural network.
Each recognition subsystem has its own feature extraction algorithm; voice signals are
extracted based on Mel-frequency Cestrum Coefficients (MFCC) algorithm, facial
images undergo digital image processing based on closing algorithm image
morphology and fingerprint features are extracted based on minutiae points.

Figure 2.1: Operation Flowchart of the Proposed Design

9614 Ravi Lakshmanan et al

These extracted features for each of the biometric measures for the subsystems
are used as inputs to train the neural networks, which are configured consistently in
all three subsystems based on feedforward network with back propagation training.
These neural networks also act as a database to store the biometric templates of the
users. The output decision of the neural network is sent to the GUI for validation.
Before the recognition system can be utilised, user enrolment process have to
be done where the neural network is trained with multiple biometric samples of the
users, for each of the recognition subsystem as shown in the flow chart Figure 2.2.
Then, the overall recognition system can be tested.

Figure 2.2: Operation Flowchart of the Proposed Design

The authentication process is same all three of the recognition subsystems,

which begins with the acquisition of input biometric data, performing feature
extraction and using the trained neural network to authenticate the input biometric
data with the stored template. The authentication process is in a stage-by-stage basis,
where a sequential successful authentication in the order of voice, face and fingerprint
subsystems are required to unlock the door and granted access. For instance, if the
users succeed in voice and face authentications, but failed in fingerprint
authentication, the door will still remain locked and access is denied. Two attempts
are given for authentication in each recognition subsystem, when the users fail to
authenticate themselves on the second attempt, the door will remained locked and an
alarm will sound, to indicate access denied.
Integrated Multi-Stage Biometric System Design 9615

The design and implementation of the feature extraction algorithm for voice
recognition system is done by first writing the source code in MATLAB. Figure 2.3
shows the block diagram of the feature extraction processes for voice recognition
system based on Mel-Frequency Cepstrum Coefficient (MFCC) [Link]
novelty provided is that silence detection and hamming window technique used for
pre-processing is integrated with the MFCC algorithm and neural network.

A. MFCC Feature Speech Recognition

1. Voice Acquisition via Sound Card
Based on Figure 2.3, the processing chain of the voice recognition system begins by
first obtaining the voice sample of the user in the form of continuous voice signal, via
the sound card of the computer.
The continuous voice signal will be converted into discrete voice signal, with
sampling frequency set to 10 kHz with duration of the recording time set to 1 second.
The sampling time is given by:

According to Nyquist limit, the sampling frequency is twice the amount of the
maximum analogue frequency, such as:

Hence, the maximum analogue frequency of the continuous voice signal that
can be obtained is 5000 Hz, which is within sufficient range that covers the voice
frequency produced by human.
9616 Ravi Lakshmanan et al

Figure 2.3: Proposed Automated Speech Recognition Method

2. Silence Detection
Next, the voice signal will undergo pre-processing to condition the signal into a
suitable format for further processing during feature extraction. Silence detection is
used to detect the silent portion of the voice signal, where the high amplitude portion
of the voice signal will be time-shifted to the starting point, at n = 0. As a result, any
silent portion of the voice signal at the beginning will be eliminated, which acts to
normalize the voice signal. The absolute value of the voice signal amplitude must be
above 0.05 to be considered as non-silent portion of the signal.

3. Hamming Window
The next pre-processing step is to apply Hamming windowing technique in time
domain to the voice signal. A Hamming window with 4000 samples is defined, and
then fit into 10000 samples of the voice signal. Hamming window is used to condition
the voice signal to suit the condition of Mel-frequency warping at the later stage of
the feature extraction process.
Integrated Multi-Stage Biometric System Design 9617

4. Fast Fourier Transform (FFT)

After all the pre-processing is done, the voice signal is transformed from time domain
to frequency domain using Fast Fourier Transform (FFT). The FFT of the voice signal
is symmetrical at the central frequency, 5000 Hz, which also corresponds to the
defined maximum analogue frequency of the voice signal.

5. Mel-Warping
Mel-warping is done to transform the spectrum of the voice signal obtained after FFT,
into Mel-spectrum of the voice signal. In order to perform Mel-warping, the Mel-
frequency scale must be defined first. After defining the Mel-frequency scale, Mel-
warping can be done by combining the spectrum of the voice signal in frequency
domain with the Mel-frequency [Link] process of Mel-warping is crucial in the
feature extraction stage, as the frequency of the voice signal is transformed into Mel-
frequency scale, in the attempt to emulate the biological hearing sense in human that
perceive the frequency contents of voice signal, in a non-linear scale similar to the
Mel-frequency.

6. Triangular Window Frame Blocking

Frame blocking serves to truncate the Mel-spectrum of the voice signal into individual
frame. To do this, a number of 20 overlapping triangular windows are defined. The
20 triangular windows overlap each other with a bandwidth of 50 Hz. For instance,
the eleventh triangular window will occupy the range of 500 Hz to 600 Hz,
overlapping 50 Hz with the tenth triangular window. This is done to prevent
discontinuities of the resultant truncated Mel-spectrum voice signal. After all the 20
triangular windows have been defined, the voice signal in the form of Mel-spectrum
can be truncated into 20 individual frames.

7. Obtain Spectrum Energy

The spectrum energy of the voice signal can be calculated by first summing the square
of the Mel-spectrum within the frame. After that, a vector of spectrum energy can be
obtained by summing the log of all the 20 spectrum energies obtained from the 20
frames.

8. Discrete Cosine Transform (DCT)

The final step of the feature extraction is to transform the log of the spectrum energy
from frequency domain back to time domain, using Discrete Cosine Transform
(DCT). Finally, the output of the MFCC feature extraction algorithm produces the
Mel-cepstrum coefficient of the voice signal, which can be used to uniquely represent
the voice of the user and train the neural network during the pattern recognition
process.
9618 Ravi Lakshmanan et al

B. Face Feature Recognition

Figure 2.4 shows the block diagram of the feature extraction process for face
recognition subsystem based on closing algorithm image morphology. The novelty
provided is that the histogram equalization technique is used for preprocessing
integrated with the Closing algorithm image morphology and Neural network.

Figure 2.4: Block diagram of the Feature Extraction Process for Face
Recognition System

The Closing algorithm image morphology algorithm is detailed as below:

Step 1: Image Acquisition via Webcam

Based on Figure 2.1, the processing chain of the face recognition subsystem begins by
first obtaining the facial image of the user via the webcam of the computer and using
the image acquisition toolbox in MATLAB to configure the acquisition parameters.
The acquired facial image can then be saved as Tagged Image File Format (TIFF) and
exported to the MATLAB workspace for pre-processing.

Step 2: Histogram Equalization

Image histogram equalization is done to enhance the contrast of the image by equally
distributing the colour contrast of the image. After performing image histogram
equalization, the resultant facial image shows a more distinctive outline of the face
with respect to the background.

Step 3: Image Binarization

The next pre-processing is to perform image Binarization to convert the grayscale
Integrated Multi-Stage Biometric System Design 9619

image to binary image. The resultant facial image will be in the form of 150 x 119
represented by binary bits of 1 and 0.

Step 4: Image Dilation

Facial image feature extraction using closing algorithm begins with first creating a
morphological structuring element. A flat, disk-shaped structuring element ‗SE‘, with
the value of radius, 3is created. This morphological structuring element is used with
the binarized facial image to perform image dilation. The resultant facial image is the
output of binary dilation through the decomposition of the defined structuring element
‗SE‘.

Step 5: Image Erosion

Similar to image dilation, the defined morphological structuring element ‗SE‘ is used
with the binarized facial image to perform image erosion. The resultant facial image is
the output of binary erosion through the decomposition of the defined structuring
element ‗SE‘. When performing binary dilation along with binary dilation, erosion is
applied by automatically using the binary image packing to speed up the dilation.

Step 6: Image Outlining

After image dilation and erosion are done, the facial image outline. Closing algorithm
image morphology involves obtaining the outline of the face shape and certain details
of the face structure such as the eyes, nose and mouth. This is done by superimposing
the dilation and erosion of the facial image. The resultant facial image illustrates the
outline of the facial image, where the details of the eyes, nose, mouth and even the
shape of the hair can be seen.

Step 7: Image Downsizing

Image downsizing serves to compress the extracted features from the facial outline, to
a smaller dimension, in order to suit the process for training the neural network for
facial recognition. The resultant downsized facial is a 30 x 30 image represented by
binary bits of 1 and 0. Finally, the downsized facial image will be reshaped into a
column vector with 900 feature points of binary number, to train the neural network.

C. Finger Print Feature Recognition

The design and implementation of the future extraction algorithm for fingerprint
recognition system is done by first writing the source code in MATLAB. Figure 2.5
shows the block diagram of the feature extraction processes for fingerprint
recognition system based on fingerprint minutiae points feature extraction algorithm:
9620 Ravi Lakshmanan et al

Figure 2.5: Block Diagram of Fingerprint Minutiae Points Feature Extraction

Algorithm

Step 1: Fingerprint Acquisition via Fingerprint Reader

Based on Figure 2.5, the processing chain of the fingerprint recognition subsystem
begins by first obtaining the fingerprint image of the user via a fingerprint reader and
using a Security Development Kit (SDK) known as UniFinger, as a software interface
to save the fingerprint image. Using the user interface, the captured fingerprint image
can be saved in the computer as a bitmap image file, where subsequent processing can
be done by loading the fingerprint image over to MATLAB.

Step 2: Image Enhancement via Filtering

The captured fingerprint image in the form of 360 x 256 bitmap file is exported to the
workspace and pre-processing is done to enhance the quality of the image. The
fingerprint image will undergo two types of noise removal filters which are 2-D
Median filter and 2-D adaptive Wiener [Link] fingerprint image after applying
Median filter contains less dotted points. While the fingerprint image after applying
Wiener filter have a smoother edge and better image quality. As a result, the ridge
pattern of the fingerprint can be clearly seen and the feature points can be properly
extracted in the following processes.

Step 3: Image Binarization

Fingerprint feature extraction begins with first converting the pre-processed
Integrated Multi-Stage Biometric System Design 9621

fingerprint image into binary. The resultant image seen will be in the form of 360 x
256 represented by binary bits of 1 and 0.

Step 4: Image Normalization

The next step in feature extraction is to normalize the dimension of the fingerprint
image, to focus on the Region of Interest (ROI). The resultant fingerprint image
shown illustrates the ROI of the fingerprint, where the more distinctive ridge pattern
of the fingerprint is focused, so that the feature points can be extracted easily and
consistently.

Step 5: Image Thinning

Image thinning, also known as skeletonisation is done to thin down the ridge pattern
into a single line. Thinning of the fingerprint image shows a single line skeletal view
of the fingerprint ridge pattern.

Step 6: Minutiae Extraction

Minutiae extraction algorithm is based on finding the termination and bifurcation
points in the ridge pattern of the fingerprint. The resultant extracted minutiae image
consists of a combination of red and blue markings. The red markings represent the
ridge termination points, while the blue markings represent the ridge bifurcation
points.

Step 7: Reshape Minutiae Points

The position of the minutiae points is described by their coordinates in the form of x-
axis and y-axis relative to the origin. Hence, these minutiae points will be reshaped
into 60 feature vectors. Pythagoras‘ Theorem is used to determine the exact point of
the ridge termination and bifurcation, based on their coordinates on the x-axis and y-
axis. Finally, these points are combined into a column vector of 60 feature vectors,
which will be used in training the neural network.

D. Neural Network:
After obtaining and compiling all the extracted biometric data from the users, the next
procedure is to configure and train the neural network using the neural network
toolbox utility in MATLAB. However, due to the limited number of biometric
samples which are 25 samples, the number of samples is duplicated 4 times, which
gives a total of 100 samples. This would be sufficient to properly train the neural
network.
In supervised training of the neural network, the biometric feature vectors are
the inputs to the neural network, where the neural network will be trained to map the
corresponding targets of the inputs. The 100 input data containing the biometric
feature vectors will be randomly divided into 70% for training data, 15% for
validation data and 15% for testing data. Training data constitute the major samples,
as they are presented to the network during training, where the value of the weight
and bias are adjusted to reduce the error rate in recognising the pattern. Validation
data are used to measure the generalisation of the neural network. The training of the
9622 Ravi Lakshmanan et al

neural network will be stopped when generalisation stops improving, as the neural
network starts to memorise the pattern, instead of learning to recognise. Testing data
have no effect on the training, only to provide independent measure on the
performance of the network during and after training.
The structure of the neural network is be configured as feedforward network.
The number of hidden neurons in the hidden layer can be defined. For the purpose of
simulation test and analysis, three different neural networks are configured with 10,
50 and 100 hidden neurons. Table 2.1 shows a summary on the number of neuron for
each layer of the neural networks for each of the recognition subsystem:

Table 2.1: Number of Neuron in Each Layer

Recognition Subsystem Input Neuron Hidden Neuron Output Neuron

Voice 20 10, 50, 100 5
Face 900 10, 50, 100 5
Fingerprint 60 10, 50, 100 5

The number of input neuron will correspond to the number of input biometric
feature vector. For voice recognition subsystem, there are 20 MFCC feature vectors
and hence there are 20 input neurons, same goes for the face and fingerprint
recognition subsystem. The output layer consists of 5 output neurons which
correspond to the number of users enrolled to the system. Hence, the output layer is
acting as similar to the database that stores the template of the enrolled users. The
transfer function used in these feedforward neural networks is tangent-sigmoid, which
is commonly used for pattern [Link] the neural network is properly
configured, training can be commenced, with the use of a training algorithm known as
scaled conjugate gradient back-propagation.

III. EXPERIMENTAL RESULTS

Database
For the proposed voice recognition system, testing is done on two different scenarios
which are word independent recognition and word dependent recognition. 5 voice
samples of 5 users have been obtained, which provide a total of 25 voice samples,
where these voice samples act as the voice biometric data that will undergo MFCC
feature extraction. The extracted MFCC feature vectors of these voice samples are
used to train the neural network.
For the first case scenario for simulating word independent recognition, the
voice recognition system is tested in recognizing the users when all the uttered
passwords are same; hence it is a word independent recognition basis. Table 3.1
shows the name of the users, user ID and the uttered password ‗Activate‘, which is
same for all users.
Integrated Multi-Stage Biometric System Design 9623

Table 3.1: Users Details for Word Independent Simulation

User ID Username Uttered Password

1 Ivan Activate
2 Paul Activate
3 Ian Activate
4 Esther Activate
5 Steven Activate

For the second case scenario for simulating word dependent recognition, the
voice recognition system is tested in recognizing the users when the uttered passwords
are different; hence it is a word dependent recognition basis. Table 3.2 shows the
name of the users, user ID and the different uttered password assigned to each user.

Table 3.2: Users Details for Word Dependent Simulation

User ID Username Uttered Password

1 Ivan Omega
2 Paul Alpha
3 Ian Beta
4 Esther Delta
5 Steven Hello

For face recognition subsystem, 5 facial image samples are collected from 5
users, summing up to a total of 25 facial image samples, where these facial image
samples acting as the face biometric data will undergo closing algorithm image
morphology feature extraction. The extracted facial feature vectors will be used to
train the neural network. The facial images are taken with a slight variation from each
other, in terms of the capture angles, facial expressions and lightings. This is done so
that the neural network can be trained to recognise the face of the users at different
conditions.
For fingerprint recognition subsystem, 5 fingerprint samples taken are
collected from 5 users, summing up to a total of 25 fingerprint samples, where these
fingerprint samples acting as the fingerprint biometric data will undergo minutiae
points feature extraction. The extracted minutiae point feature vectors of these
fingerprint samples will be used to train the neural network.

Graphical User Interface

Graphical User Interface (GUI) is designed to replace the MATLAB command line
interaction when operating the system. The GUI maintains the same function and
processing flow, but it is design to provide user-friendly interaction when using the
system. The layout of the GUI is design using GUI Development Environment
(GUIDE) in MATLAB and a script file is generated containing the source code to
implement all the recognition function into the GUI. Four GUIs have been design, one
9624 Ravi Lakshmanan et al

main GUI and three GUIs for the individual recognition [Link] 3.1 shows
the design of the Main GUI:

Figure 3.1: Main GUI

The Main GUI represents the input stage of the overall system, where it
functions to activate the peripheral door lock system through serial port
communication, user enrolment by training the neural networks and username input.
In addition, the date and time of access is displayed for reference. A notification bar is
also included at the bottom of the GUI, which display a text that instructs and
provides guidance to user in using the system. The notification bar remains present in
the all the [Link] that, message box will pop up to provide additional options
for user to select, such the option to retrain neural networks as or as an error message
when invalid action is done, such as an invalid username is entered. Figure
3.2illustrates then mentioned two types of message boxes:

Figure 3.2: Option and Error Boxes

When the system is activated, all neural networks have been trained and a
valid username is entered, the system will prompt the user to proceed to the next GUI,
where the authentication stage begins. The authentication flow is the same as
designed using the command line interface, which starts with voice authentication,
followed by face and fingerprint authentication.
Integrated Multi-Stage Biometric System Design 9625

Results
After the multi-stage biometric recognition system has been designed and
implemented, simulation testing is done to evaluate the overall performance of the
system. The simulation setup will be the same as the setup for the recognition
subsystem, where the overall FRR and FAR is used to evaluate the overall recognition
rate of system with respect to raw biometric test samples. Mean Squared Error (MSE),
number of epoch and percentage error (%E) are used to evaluate the training
performance of the neural networks. Table 3.3 shows the training performance of
neural networks configured with 50 hidden neurons for the individual recognition
subsystem:

Table 3.3: Neural Networks Performance for Multi-Stage Biometric Recognition

system

Recognition Subsystem MSE Epoch Percentage Error

Voice 40 0%
Face 31 0%
Fingerprint 49 0%

Table 3.4shows the result tabulation of FRR test on the overall system when
tested with 25 raw biometric samples from 5 users:

Table 3.4(a): Result Tabulation of FRR Test

Raw Test Biometric Samples

Users Ivan1 Ivan2 Ivan3 Ivan4 Ivan5 Paul1 Paul2 Paul3 Paul4 Paul5
Ivan 1 1 1 1 1 0 0 0 0 0
Paul 0 0 0 0 0 1 0 1 1 1
Ian 0 0 0 0 0 0 0 0 0 0
Esther 0 0 0 0 0 0 0 0 0 0
Steven 0 0 0 0 0 0 1 0 0 0
False Rejection No No No No No No Yes (Voice) No No No

Table 3.4(b): Result Tabulation of FRR Test

Raw Test Biometric Samples

Users Ian1 Ian2 Ian3 Ian4 Ian5 Esther1 Esther2 Esther3 Esther4 Esther5
Ivan 0 0 0 0 0 0 0 0 0 0
Paul 0 0 0 0 0 0 0 0 0 0
Ian 1 1 1 1 1 0 0 0 0 0
Esther 0 0 0 0 0 1 1 1 1 1
Steven 0 0 0 0 0 0 0 0 0 0
False Rejection No No No No No No No No No No
9626 Ravi Lakshmanan et al

Table 3.4(c): Result Tabulation of FRR Test

Raw Test Biometric Samples

Users Steven1 Steven 2 Steven 3 Steven 4 Steven 5
Ivan 0 0 0 0 0
Paul 0 0 0 0 0
Ian 0 0 0 0 0
Esther 0 0 0 0 0
Steven 1 1 1 1 1
False Rejection No No No No No

Based on the tabulation of the simulation results shown in Table 3.4, the FFR is
calculated as shown in the following:

Table 3.5 shows the result tabulation of FAR test on the overall system when
each user is asked to access the system as the other users, but using their own
biometric data:
Table 3.5(a): Result Tabulation of FAR Test

Raw Test Biometric Samples

Users Ivan1 Ivan2 Ivan3 Ivan4 Ivan5 Paul1 Paul2 Paul3 Paul4 Paul5
Users Ivan NA NA NA NA NA 0 0 0 0 0
Access as: Paul 0 0 1 0 0 NA NA NA NA NA
Ian 0 0 0 0 0 0 0 0 0 0
Esther 0 0 0 0 0 0 0 0 0 0
Steven 0 0 0 0 0 0 0 0 0 0
False Acceptance No No Yes (Face) No No No No No No No

Table 3.5(b): Result Tabulation of FAR Test

Raw Test Biometric Samples

Users Ian1 Ian2 Ian3 Ian4 Ian5 Esther 1 Esther 2 Esther 3 Esther 4 Esther 5
Users Ivan 0 0 0 0 1 0 0 0 0 0
Access as: Paul 0 0 0 0 0 0 0 0 0 0
Ian NA NA NA NA NA 0 0 0 0 0
Esther 0 0 0 0 0 NA NA NA NA NA
Steven 0 0 0 0 0 0 0 0 0 0
False Acceptance No No No No Yes (Voice) No No No No No
Integrated Multi-Stage Biometric System Design 9627

Table 3.5(c): Result Tabulation of FAR Test

Raw Test Biometric Samples

Users Steven1 Steven 2 Steven 3 Steven 4 Steven 5
Users Access as: Ivan 0 0 0 0 0
Paul 0 0 0 0 0
Ian 0 0 0 0 0
Esther 0 0 0 0 0
Steven NA NA NA NA NA
False Acceptance No No No No No

Based on the simulation results shown in Table 3.5, the FAR is calculated as
shown in the following:

The simulation results have shown that the overall multi-stage biometric
recognition system is able to achieved FRR of 4% and FAR of 8%. Thus, the system
has been design, implemented and tested to fulfill the research objective of achieving
overall FRR and FAR of less than 10%.
The time taken to train three neural networks has been taken into
consideration. Therefore, an additional test is conducted to record the average time
taken to train the neural networks. Table 3.6 shows results of 10 simulation tests to;
calculate the time taken to train three neural networks:

Table 3.6: Neural Networks Training Time

No. of Training Time for Training Time Training Time for Total
Test Voice Neural for Face Neural Fingerprint Neural Training
Network (sec) Network (sec) Network (sec) Time (sec)
1 1.09 1.21 1.33 3.63
2 2.03 1.33 1.01 4.37
3 1.01 2.10 1.06 4.17
4 1.23 1.16 1.17 3.56
5 1.67 1.32 1.05 4.04
6 1.17 1.43 1.00 3.60
7 1.03 1.11 1.22 3.36
8 1.02 1.35 1.62 3.99
9 1.35 1.08 1.12 3.55
10 1.04 1.61 1.56 4.21
Total 12.64 13.70 12.14 38.48
9628 Ravi Lakshmanan et al

The average time taken for neural networks training is calculated as follows:

IV. CONCLUSION
Thus an appropriate feature extraction algorithm for voice, face and fingerprint
recognition were integrated into a multi-stage biometric recognition system with
artificial neural network, in which the voice recognition system utilises Mel-
frequency Cepstrum Coefficient (MFCC) feature extraction algorithm to represents
the voice signal of the user in the form of cepstrum of the voice derived from Discrete
Cosine Transform (DCT) on the log power spectrum on a non-linear Mel-scale
frequency. Face recognition subsystem utilises closing algorithm image morphology
that extracts the features of the face based on the facial outline, via superimposing of
the dilation and erosion of the facial image. Fingerprint recognition subsystem utilises
minutiae extraction to extract the ridge termination and bifurcation points of the
fingerprint. Results show that 50 hidden neurons are sufficient to provide optimum
performance for the neural networks to recognise voice, face and fingerprint biometric
patterns. Therefore, the integrated biometric recognition system consists of three
neural networks for configured consistently with 50 hidden neurons. An overall FRR
and FAR below 10% is achieved, the outcomes for the simulation test on the
integrated biometric recognition system have shown that, the calculated FRR value is
4% and FAR value is 8%.Also, the calculated average time taken to train the neural
network is 3.848 seconds only.

REFERENCES

[1] Anusuya.M.A., &Katti.S.K., ―Speech Recognition by Machine‖, International

Journal of Computer Science and Information Security, Vol.6 (3), 2009, pp.
181-205.
[2] Azam,S.M., Mansoor,Z.A., Mughal,M.S., & Mohsin,S, ―Urdu Spoken Digits
Recognition Using Classified MFCC and Backpropagation Neural Network.
Computer Graphics, Imaging and Visualisation‖, Bangkok. 14th to 17th
August 2007. Bangkok: CGIV. pp. 414 – 418.
[3] ChadawanIttichaichareon, SiwatSuksri& Thaweesak Yingthawornsuk,‖Speech
Recognition using MFCC‖, in the International Conference on computer
Graphics, Simulation and Modeling, July 28-29, 2012, pp.135-138.
[4] Chen, [Link] al., ―Incremental Feedback Learning Methods For Voice
Recognition Based On DTW. 2012 Proceedings of International Conference
on Modeling, Identification & Control‖, Wuhan, 24th to 26th June 2012.
Wuhan: ICMIC. pp.1011 – 1016.
Integrated Multi-Stage Biometric System Design 9629

[5] DavoodZabihzadeh, & Mohammad Moattar, ―Manifold learning based

speaker dependent dimension reduction for robust text independent speaker
verification‖, Inetrnational Journal of Speech technology, springer, 2014,
pp.1-10.
[6] DebmalyaChaakrabarty, MahadevaPrasana& Rohan Kumar Das,
―Development and evaluation of online text-independent speaker verification
system for remote person authentication‖, International Journal of Speech
Technology, Springer, 2013,pp. 75 – 88.
[7] DimitriosVerveridis, & Constantine Kotropoulos, ―Emotional speech
recognition: resources, features and methods‖, Aritificial Intelligence and
Information Analysis Laboratory, Aristotle university pf Thessaloniki, Greece,
pp. 1 – 22.
[8] Gandhiraj,R&Sathidevi,P.S., ―Auditory-based Wavelet Packet Filterbank for
Speech Recognition using Neural Network‖, International Conference on
Advanced Computing and Communications. Guwahati, 18th to 21st December
2007. Guwahati: ADCOM. pp. 666 – 673.
[9] Gregory , P. & Simon, M. A.,―Biometric For Dummies. Indianapolis:‖ Wiley
Publishing, Inc.,2008.
[10] Ibrahim Patel, &[Link] Rao, ― Speech Recognition using HMM with
MFCC-An analysis using frequency spectral decomposition technique‖,
International Journal of Signal & Image Processing, Vol.1(2), December 2010,
pp. 101- 110.
[11] Muzaffar, F., Mohsin, B. & Naz, F, ― DSP Implementation of Voice
Recognition Using Dynamic Time Warping Algorithm‖, Student Conference
on Engineering Sciences and Technology. Karachi, 27th August 2005.
Karachi: SCONEST. pp. 1 – 7.
[12] [Link], &Gajanan [Link], ―An overview of Technical
Progress in Speech recognition‖, International Journal of Advanced research
in computer science and software engineering, Vol.3,Issue. 3, March 2013, pp.
488 – 497.
[13] Suzuki, H., Zen,H., Nankaku,Y., & Miyajima,C., ―Speech recognition using
voice characteristic dependent acoustic models‖, in the IEEE International
conference on Acoustics, Speech, and Signal Processing, 6-10 April 2—3, pp.
740-743.
[14] Tsenov,G.T&Mladenov,V.M., ― Speech recognition using neural netwroks‖,
in 10th Symposium on Neural Network Applications in Electrical Engineeirng,
23-25 Sept.2010,pp. 181-186.
[15] Vasuthi, S, Vaidehi, V, Babu, N.T.N, and Treesa, T.M. , ―An Efficient Multi-
Modal Biometric Person Authentication System Using Fuzzy Logic‖,2010
Second International Conference on Advanced Computing. Chennai, 14th to
16th December 2010. Chennai: ICoAC., 2010, pp. 74–81.
[16] Yang Liu, Elizabeth Shriberg, & Andreas Stolcke, ―Enriching Speech
recognition with automatic detection of sentence boundaries and disfluencies‖,
IEEE transactions on speech & audio processing, 2013, pp. 1- 15.
 9630 Ravi Lakshmanan et al

View publication stats