CHAPTER 4 STUDY OF MULTIMODAL BIOMETRIC TECHNIQUES...
Transcript of CHAPTER 4 STUDY OF MULTIMODAL BIOMETRIC TECHNIQUES...
Page 105
CHAPTER – 4
STUDY OF MULTIMODAL BIOMETRIC TECHNIQUES AND DESIGN AND
DEVELOPMENT OF MULTIMODAL BIOMETRIC AUTHENTICATION
MODEL WITH FINGERPRINT AND FACE RECOGNITION
4.1 Multimodal biometric techniques
Human has capability to identify the person based on physiological or behavioral
characteristics. But some time, more multiple modality or multiple evidences are required
to identify the person. In this section, an introduction to multibiometric system has been
given to have clear idea of such type of systems.
4.1.1 Introduction to multi biometrics
Biometric systems can be designed to recognize a person based on information acquired
from multiple biometric sources. Such systems, known as multibiometric systems, can be
expected to be more accurate due to the presence of multiple pieces of evidence.
Multibiometric systems offer several advantages over traditional unibiometric systems
listed below [1]:
1. Multibiometric systems can offer substantial improvement in the matching
accuracy of a biometric system depending upon the information being combined
and the fusion methodology adopted.
2. Multibiometric systems address the issue of non-universality or insufficient
population coverage.
3. It becomes increasingly difficult for an impostor to spoof multiple biometric traits
of a legitimately enrolled individual.
4. Multibiometric systems effectively address the problem of noisy data.
5. Multibiometric systems help in the continuous monitoring or tracking of an
individual in situations when a single trait is not sufficient.
6. A multibiometric system may be viewed as a fault tolerant system which
continues to operate even when certain biometric sources become unreliable due
to sensor or software malfunction, or deliberate user manipulation.
Page 106
4.1.2 Types of multibiometric systems
Multibiometric systems can be categorized by considering the way you are taking
multiple sources of evidence. There can be different options to have sources of evidence.
Consideration can be made for multiple instances, multiple algorithms, multiple sensors
or sometime multiple traits. Based on this, we can categorize multibiometric systems in
following categories:
1. Multi-sensor systems
2. Multi-algorithm systems
3. Multi-instance systems
4. Multi-sample systems
5. Multimodal systems
6. Hybrid systems
Let have a brief look at these systems.
1. Multi-sensor systems
With this kind of systems, multiple sensors are used to capture different images to have
different kind of evidences from single trait. For example, it is possible to use CCD
camera and range sensor to get multiple kind of information about the face. Though, the
cost will increase, but rich information set will be available. Marcialis and Roli, 2004a
discuss a scheme to fuse the fingerprint information of a user obtained using an optical
and a capacitive fingerprint sensor [2]. The authors, in their work, indicate that the two
sensors provide complementary information thereby resulting in better matching
accuracy.
2. Multi-algorithm systems
With this kind of systems, same information captured from evidence can be processed
with multiple algorithms. Fingerprint image can be processed with minutiae based and
correlation based method simultaneously.
Rose et al. made experiment with fingerprint recognition by using texture based and
minutiae based algorithm for diverse feature extraction [3].
As there is no need to have extra sensor or equipment, so cost effectiveness is there. As
well as user convenience is more.
3. Multi-instance systems
Page 107
These kinds of systems use multiple instances of same biometric trait. They are also
called multi-unit systems. For example, for fingerprint recognition, we can take two
fingers from both the hands, i.e. one from each. Neither they are using multiple sensors
not they are using multiple algorithms. So, complexity can be reduced as well as cost
effectiveness will be more. Another case can be considered for iris, where we can take
both the iris.
4. Multi-sample systems
Here, one sensor can be used to get multiple samples of the same trait of the same finger
or iris, so variations can be captured as well as completeness of the features can be
captured. For example, Front face image along with left and right face images are
captured to create a complete 3-D ace image for recognition. Hill et al. prepared similar
kind of system [4].
5. Multi-modal systems
With these systems, multiple evidences from different biometric traits to recognition the
person. For example, earlier multimodal biometric systems used face and voice trait to
identify person. The experiment was carried out by Brunelli and Falvigna [5]. The
identification accuracy can be significantly improved by utilizing an increasing number
of traits although the curse-of-dimensionality phenomenon would impose a bound on this
number. The curse-of-dimensionality limits the number of attributes (or features) used in
a pattern classification system when only a small number of training samples is available
[6]. The restrictions will be imposed on practical implementation by cost of deployment,
time, throughput time etc.
6. Hybrid systems
Hybrid refers to the systems that integrate a subset of the five type of systems discussed
above. Brunelli and Falavigna described an arrangement in which two speaker
recognition algorithms are combined with three face recognition algorithms at the match
score and rank levels via a HyperBF network [5]. This system can be considered as multi-
algorithmic as well as multimodal. Hybrid systems try to extract as much information as
possible.
Page 108
Figure 4.1: Various multibiometric system examples
4.1.2.1 Processing sequence
The processing sequence adopted by multibiometric system refers to the order in which
the acquired information is used in order to take a decision. The concentration is made
not in the order of acquisition, but the order in which they are processed. The samples can
be acquired sequentially, but can be processed simultaneously. In the serial or cascade
mode, the processing of information takes place sequentially. The cascading scheme can
improve user convenience as well as allow fast and efficient searches in large scale
identification tasks. A multibiometric system designed to operate in the parallel mode
generally has a higher accuracy because it utilizes more evidence about the user for
recognition. Most multibiometric systems have a parallel architecture because the
primary goal of system designers has been to reduce the error rates of biometric systems
and not necessarily the throughput and/or processing time.
In addition to these two modes of operation, it is possible to have a hierarchical (tree-like)
architecture to combine the advantages of both cascade and parallel architectures [7]. In
Page 109
this method, a subset of the acquired modalities may be combined in parallel, while the
remaining modalities may be combined in a serial fashion. This architecture can be
dynamically determined based on the quality of the individual biometric samples as well
as the possibility of encountering missing biometric data.
4.1.3 Introduction to multimodal biometrics
Multimodal biometrics refers to the use of a combination of two or more biometric
modalities in a verification / identification system. Identification based on multiple
biometrics represents an emerging trend. The most compelling reason to combine
different modalities is to improve the recognition rate. This can be done when biometric
features of different biometrics are statistically independent. There are other reasons to
combine two or more biometrics. One is that different biometric modalities might be
more appropriate for the different applications. Another reason is simply customer
preference [9].
4.1.4 Example multimodal biometric systems
Work in the area of multimodal biometric systems started in mid 1990s. Few example
multimodal biometric systems are shown here.
Modalities fused Authors Level of
fusion Fusion methodology
Face and voice Brunelli and
Falavigana, 1995
Match
score and
rank
Geometric weighted average;
HyperBF
Kittler et al.,1998 Match
score
Sum, product, min, max and
median rules
Ben-yacoub et al.,
1999
Match
score
SVM, Multilayer perceptron, C4.5
decision tree, Fisher‘s linear
discrimination, Bayesian classifier
Bigaun et al., 1997 Match
score
Statistical model based on
Bayesian theory
Face, voice and
lip movement
Frischholz and
Dieckmann, 2000
Match
score,
decision
Weighted sum rule, majority
voting
Face and
fingerprint
Hong and Jain,
1998
Match
score
Product rule
Snelick et al., 2005 Match
score
Sum rule, weighted sum rule
Page 110
Face, fingerprint
and hand
geometry
Ross and Jain,
2003
Match
score
Sum rule, decision trees,
discriminant function
Face, fingerprint
and voice
Jain et al., 199b Match
score
Likelihood ratio
Face and iris Wand et al., 2003 Match
score
Sum rule, weighted sum rule,
fisher‘s linear discrimination ,
neural network
Face and gait Shakhnarovich et
al., 2001
Match
score
Sum rule
Kale et al., 2004 Match
score
Sum and product rules
Face and ear Chang et al., 2003 Sensor Concatenation of raw images
Face and
palmprint
Feng et al., 2004 Feature Feature concatenation
Fingerprint, hand
geometry and
voice
Toh et al., 2004 Match
score
Weighted sum rule
Fingerprint and
hand geometry
Toh et al., 2003 Match
score
Reduced multivariate polynomial
model
Fingerprint and
voice
Toh and Yau,2005 Match
score
Functional link network
Fingerprint and
signature
Fierrez-aguilar et.
al.,2005c
Match
score
SVM in which quality measures
are incorporated
Voice and
signature
Krawczyak and
Jain,2005
Match
score
Weighted sum rule
Table 4.1: Various multimodal biometric systems
4.1.5 Advantages of multimodal biometric systems
Multimodal biometric systems are expected to be more reliable due to the presence of
multiple pieces of evidence. These systems are also able to meet the stringent
performance requirements imposed by various applications [10].
Multimodal systems address the problem of non-universality: it is possible for a subset of
users to not possess a particular biometric. For example, the feature extraction module of
a fingerprint authentication system may be unable to extract features from fingerprints
associated with specific individuals, due to the poor quality of the ridges. In such
instances, it is useful to acquire multiple biometric traits for verifying the identity.
Multimodal systems also provide anti-spoofing measures by making it difficult for an
intruder to spoof multiple biometric traits simultaneously. By asking the user to present a
random subset of biometric traits, the system ensures that a live user is indeed present at
Page 111
the point of acquisition. However, an integration scheme is required to fuse the
information presented by the individual modalities.
4.2 Introduction to different fusion techniques in multimodal biometric systems
The fusion can be achieved in two different ways. The first is information fusion prior to
matching and the second method is fusion after matching [8].
1. Fusion prior to matching
Fusion prior to matching can be achieved in two different ways:
1. Sensor level fusion
2. Feature level fusion
Sensor level fusion is applicable only if the multiple sources represent samples of the
single biometric trait obtained either using a single sensor or different compatible
sensors.
Feature level fusion is achieved by combining different feature sets extracted from
multiple biometric sources. Feature sets could be either homogeneous or heterogeneous.
The consolidation of feature set creates problems as the feature sets originate from
different algorithm and modalities.
2. Fusion after matching
Fusion after matching can be achieved in three different ways:
1. Matching score level fusion
2. Rank level fusion
3. Decision level fusion
Matching score level fusion provides richest set of information.
Rank level fusion consolidates the ranks output by the individual subsystems in order to
derive a consensus rank of each identity. Rank level fusion provides less information with
compare to match score level fusion.
Decision level fusion is carried out at decision level when the decisions output by the
individual matcher are available. COTS (Commercial off The Shelf) matchers provide
only the final decision and those decisions are evaluated with the help of some rules like
―AND‖ or ―OR‖, majority voting, Bayesian decision fusion etc. Here there will be a
problem of least information about the features or scores of different modalities.
Page 112
Here the researcher has selected two biometric traits, so instead of fusion before
matching, which mostly deals with single modality, concentration has been given to
fusion after matching. So, here is an introduction to fusion methods after matching.
1. Match score level fusion
This fusion technique is also known as measurement level or confidence level fusion. It is
comparatively easy to consolidate the scores generated by different biometric matchers.
This method is the most commonly used method for fusion.
Here we have to identify the pattern only in two classes: genuine (Truly what something
is said to be; authentic) or impostor (A person who pretends to be someone else in order
to deceive others, esp. for fraudulent gain). In general there are three different methods to
achieve match score level fusion. They are:
1. Density based score fusion
2. Transformation based score fusion
3. Classifier based score fusion
As the match score level fusion use scores from different modalities based on different
scaling methods, the scores cannot be combined or used directly. It is required to perform
score normalization, thereby converting the scores into common domain or scale.
Score normalization can be carried out with different methods. Here are some methods of
normalization worked well with different modalities. Slobodan Ribaric and Ivan Fratric
carried out experiments for bimodal biometrics with palmprint and facial features [11].
They adopted match score level fusion for fusion of information. They discovered new
normalization technique – piecewise linear normalization. They calculated EER (Equal
Error Rate) and minimum TER (Total Error Rate) with different normalization technique.
They achieved EER of 2.79 % and min. TER of 5.15 % with this normalization. The
chart of comparison is given below:
Normalization
Technique
Piecewise
linear
median-
MAD Tanh Minmax
EER 2.79 2.79 3.05 3.12
Min. TER 5.15 5.42 5.74 6.39
Table 4.2: EER and Min. TER under different normalization techniques
Page 113
Mingxing He et al. Proposed a new method of normalization for scores in score level
fusion, Reduction of High-scores Effect normalization (RHE) [12]. They experimented
on four different databases with multi modality of fingerprint, face and fingervein. They
revealed that RHE performs better with compare to other techniques of score
normalization in score level fusion.
FAR (%) GAR (%)
Minmax Z-score Tanh RHE
0.01 97.9 98.2 97.7 99.4
0.001 96.9 97.0 95.8 98.2
Table 4.3: Performance of sum rule-based fusion on NIST-multimodal database [12]
After performing normalization, the next step is to perform fusion of the scores. Here are
some examples of different fusion models for score level fusion. Gian Luca Marcialis and
Fabio Roli suggested the following model for score level fusion of fingerprint and face
traits [13].
Figure 4.2: Score level fusion [13]
They carried out experiments on multimodal data set made up of 100 subjects with two
independent face and fingerprint data sets. With the above given scheme, they achieve
improvement in the error rate. Their results showed that fusion has improved the
reliability of the system by reducing the gap between expected and real performance.
Feifei Cui and Gongping Yang performed biometric fusion with fingerprint and finger
vein recognition [14]. They did this with score level fusion. They collected 2880
fingerprint and finger vein images from 80 fingers. With score level fusion they achieved
the following performance:
Page 114
Biometrics method Recognition rate
Fingerprint 95.3 %
Finger vein 93.72 %
Score level fusion 98.74%
Table 4.4: Recognition rate for fingerprint and finger vein fusion
These results shows that score level fusion works well with compare to unimodal
biometric traits.
Fawaz Alsaade experimented score level fusion with face and voice biometrics. He
investigated the results under three data conditions and with min-max normalization. He
used Adaptive Neuro – Fuzzy Inference System (ANFIS) for decision making [15]. He
was able to achieve 0 % EER with ANFIS approach with clean data of both face and
voice biometrics. Table 4.5 shows experimental outcomes carried out by Fawaz Alsaade.
Modality EER%
Voice (TIMIT Database) 2.55
Face (XM2VTS Database) 3.57
Fused: voice and face by BFS (Brute force search) 0.05
Fused: voice and face by SVM 0.68
Fused: voice and face by ANFIS 0
Table 4.5: Multimodal biometric verification based on clean biometric data
Sarat C. Dass et al. proposed a framework to combine the match score from multiple
modalities with the use of likelihood ratio statistic computed using generalized densities
which were estimated from genuine and impostor match scores[16]. They conducted
experiments on two different databases with different number of users. The details of
databases are shown in Table 4.6:
Database Modalities No. of Users
MSU-Multimodal Fingerprint, Face, Hand-geometry 100
NIST-Multimodal Fingerprint (Two Fingers), Face (Two
matchers)
517
Table 4.6: Details of databases used by Sarat Dass et al. [16]
Page 115
They proposed two different approaches to combine evidences based on generalized
densities:
1. Product rule
2. Copula model
With the above given database and these two score fusion methods, they achieved
consistently high performance.
Romaissaa Mazouni and Abdellatif Rahmoun fused face and speech modalities with five
different methods of score level fusion: Particle Swarm Optimization (PSO), Adaptive
Neuro Fuzzy Interface Systems (ANFIS), Genetic Algorithm (GA), Brute Force Search
(BFS), and Support Vector Machine (SVM) [17]. They did their experiments with three
kinds of datasets: Clean data, varied data, and degraded data. They derived the conclusion
that Genetic algorithm (GA) and Particle Swarm Optimization performed best among all
five method even in worst conditions.
2. Decision level fusion
Now a day, if you are using commercial off the shelf tools for biometric verification, then
decision level fusion is the only option for fusion, as they don‘t provide the data about the
scores or feature neither they provide details about the ranking of different users after
comparison. Decision level fusion is also referred as abstract level fusion. They only
provide the result of matching in the form of whether the user is genuine or imposter.
With decision level fusion, there are different rules that can be used to authenticate the
user. Lam and Suen proposed majority voting rule [18].They also proposed behavioral
knowledge space method. Xu et al. proposed weighted voting based on Dempster -
Shafer theory [19]. Daugman proposed AND/ OR rules for deciding the decision [20].
The general and mostly used approach for decision level fusion is majority voting. Here
the input sample is given the identity for which the majority of the matchers are agreed.
AND and OR rules are used rarely, because as they combine two different matchers, so
sometimes degradation of performance could be there with this method [20]. The main
benefit of the majority voting method is that neither you require prior knowledge about
the matcher nor the training is required for final decision making [8]. Domingos and
Page 116
Pazzani suggested that naïve Bayesian decision fusion works very well even if the
matchers are dependent to each other [21].
3. Rank level fusion
The rank level fusion is generally adopted for the identification of a person rather than
verification. In verification, as we have to compare the template only with one template
in the database, here we have to generate rank of identities in sorted order with all
modalities. Then after with the help of one method of fusion, we have to fuse the ranking
for each person available for different modalities. Then the identity with lowest score is
identified as the correct person. This method provide more accuracy with compare to just
a identifying best match with one modality. But the only thing is that, it provides less
information for fusion purpose. With compare to match score level fusion, here you can
easily compare the ranking from different modalities. So the decision making is easy.
Md. Maruf Monwar and Marina L. Gavrilova carried out rank level fusion with face,
signature and ear biometric traits [22]. They performed experiments with PCA and
fisher‘s LDA. The rank of individual matchers was combined with highest rank, Borda
count, and logistic regression approaches. With this approach, the performance was
improved performance even with low quality of data. Table 6 shows performance of the
experiment.
Systems Biometric identifiers Fusion level and approach EER
Md. Maruf Monwar et al. Face, Ear, Signature Rank ; logical regression 1.12%
Garcia – Salicetti et al. Signature, Voice Match score; 1.88%
Nandkumar et al. Fingerprint Match score 3.39%
Table 4.7: Comparison of different multibiometric systems [22]
Ajay kumar and Sumit Shekhar suggested combination of multiple palmprint
representations to achieve improvement in the performance with compare to individual
performance [23]. They performed various rank level combinations like, Borda Count,
Logistic Regression, Highest rank method and Bucklin majority voting approach. With
this approach, they performed experiments with NIST BSSR database. The nonlinear
Page 117
fusion approach gave best results for first – rank recognition rates. Average rank one
recognition rate was of 99%.
With rank level fusion, three most common approaches are Borda count method, Logistic
Regression method and Highest rank method. Out these three methods, Borda count and
highest rank method do not use statistical information of the classifier performance. But
with Logistic regression method, statistical information is required and weights are
assigned to classifiers. These weights depend on the data.
Abaza and Ross performed experiments with two modalities: Fingerprint and Face [24].
They evaluated results with two databases: WVU and NIST. With Modified Highest
Rank method, they achieved rank - 1accuracy of ~ 99 % on WVU dataset. They proposed
Q-based rank algorithms for rank level fusion. They were able to improve the
performance by ~4 %.
4.2.1 Score normalization in biometric systems
Score normalization refers to changing the location and scale parameters of the match
score distributions at the outputs of the individual matchers, so that the match scores of
different matchers are transformed into a common domain. When the parameters used for
normalization are determined using a fixed training set, it is referred as fixed score
normalization. Here, the set of match scores available for training the fusion module of a
multibiometric system is examined and a suitable statistical model is chosen to fit to the
data. Based on the model, the score normalization parameters are determined. In adaptive
score normalization, the normalization parameters are estimated based on the match score
of the current test sample. This approach has the ability to adapt to variations in the input
data such as the changes in the duration of the speech signals in speaker recognition
systems [1].
The comparison of various normalization techniques can be done on the basis of
robustness and efficiency. Here is the comparison of features of different score
normalization techniques:
Page 118
Normalization technique Robustness Efficiency
Min-max No High
Decimal scaling No High
Z-score No High
Median and MAD Yes Moderate
Double sigmoid Yes High
Tanh estimators Yes High
Table 4.8: Characteristics of various normalization techniques
Here, in the case of face recognition, as the Euclidean distances have large values which
requires to be normalized, while in case of fingerprint recognition, the scores are
generated in the range of 0.0 to 1.0, which don‘t required to be normalized. The
researcher has defined his own normalization technique to normalize face scores. The
steps of normalization technique are shown below.
1. Calculate sum of all vector values (Euclidian distance).
2. Calculate average Euclidian distance
3. Calculate difference between average Euclidian distance and all other distances.
4. Divide each difference with average Euclidian distance and get value in the range
of [-1,1].
4.3 Deciding method of fusion
Based on the discussion in 4.2, it is clear that match score level fusion provides rich
information compared to rank level fusion or decision level fusion. Specially, when it is
required to normalize scored available with different modalities, there are many
mechanisms through which it is easier to normalize the score. At the same time
performance of the match score level fusion is also acceptable. Decision level fusion
provides very less information i.e. only the results of modalities in terms of final
decision. Rank level fusion provided ranks of different matches. It is possible to assign
some weights to these ranks. But considering the richness of information, the researcher
has adopted match score level fusion for multimodal biometric authentication system
[25].
Page 119
4.4 Building multimodal biometric system using face and fingerprint recognition
The model of proposed basic multimodal biometric system is shown in figure 4.3
.
Figure 4.3: Basic multimodal biometric system model
As shown in basic model, initially system will capture the face and fingerprint sample of
the person. The face sample will be compared with master face DB and calculate
Euclidean distance. The fingerprint sample will be compared with master fingerprint DB
and minutiae features will be matched. Both the scores will be combined to generate
match score. Then system will identify maximum score and will be displayed along with
respective person index.
The elaborated view of the system is shown in figure 4.4. The system captures face and
fingerprint weights. These weights will be multiplied with respective face and fingerprint
matching scores. Then both will be added to get final score. In the next step, scores of all
the persons are compared and maximum match score will be identified. Person index of
maximum match score will be find out and both values i.e. maximum score and person
index will be displayed on GUI.
Master
face DB
Master FP
DB
Calculate Euclidean
distance
Calculate minutiae
score
Display maximum
matching score and person
index
Page 120
Figure 4.4: Multimodal biometric system model
The detail about the creation of master databases is mentioned in section 3.2. The
important aspect is the fusion method which has been used here. The researcher has used
match score level fusion. With this fusion method, the scores of both modalities for
respective person will be combined. One more important aspect is the use of
normalization technique used for score normalization. The researcher has designed novel
normalization technique to normalize face score. Normalization is not required for
Applying normalization on 30
face scores
Capture weight for
face score
Generate combined score multiplying
weights with match score and
Euclidean distance and summation
Capture weight for
fingerprint score
Identify person index with maximum
score and display with text control
Master face
database
Master fingerprint
database db.mat
Comparison engine to calculate
Euclidean distances from all database
samples and generate file with 30
Euclidean distances
Comparison engine calculates match
score from each database sample
template and generate file with 30
match scores
Page 121
fingerprint score as it is generated in the range of 0.0 to 1.0. Another important aspect is
to assign weights for both modalities. The sum total of weights of both modalities must
be 1.0. The used has to decide the weights of both modalities. The flow chart
representation is shown in figure 4.5.
4.4.1 Flowchart representation of multimodal biometric system
Figure 4.5: Flow chart representation of Multimodal biometric system
Master face
database
Execute Multimodal IDE
and select Multimodal face
and fingerprint option
Calculate Euclidean
distance from test
sample and compare
with database samples‘
Euclidean distance
Load test face
sample from
test database
End process
Start
Load test
fingerprint
sample from
test database
Master fingerprint
database db.mat
Identify minutiae
features from test
sample and compare
with db.mat templates Apply normalization
process with 30
comparison scores
Enter weight for
fingerprint score
Enter weight for
face score
Calculate combined score by
multiplying weights with
respective face and fingerprint
scores
Display person index with
maximum matching score
Page 122
4.4.2 GUI representation of multimodal biometric system
The steps for operating with multimodal biometric system are shown here.
Step 1: Run multimodal_ide.m to load GUI of multimodal biometric system.
Execute multimodal_ide.m file to load GUI of multimodal biometric system. The GUI
contains Unimodal and Multimodal menu options.
Step 2: Select „Multimodal‟ menu option and „Multimodal with face and fingerprint‟
submenu option
Initially, it is required to run multimodal_ide.m file to load GUI of the system. After
loading GUI, next step is select ‗Multimodal‘ menu option. On selection of this menu
option, system will show submenu containing two options. Out of these two options,
select ‗Multimodal with face and fingerprint‘ option.
Step 3: System hide unnecessary controls from GUI on selection of submenu option
After selection of ‗Multimodal with face and fingerprint‘ option, system will hide
unnecessary components from GUI. The system will show four buttons, 2 axes controls
and 4 edit controls required for this system.
Figure 4.6: Multimodal IDE with all components
Menu selection
options
Axes control 1 &
Axes control 2
Edit controls
Page 123
Figure 4.7: Selection of ‗Multimodal‘ menu option
Figure 4.8: Load necessary components of multimodal system
Page 124
Step 4: Click on „Load face image‟ button to select face test sample
Figure 4.9: Selection of test face sample
The next step is to load test samples. First, to load face test sample, click on ‗Load face
image‘ button. The system will show file dialog box to select face test image. Select face
test sample image. The system will load that image in axes control 1.
Step 5: Click on „Load fingerprint image‟ to load fingerprint test sample
After loading face test sample, it is required to load fingerprint test image. So, click
‗Load fingerprint image‘ button in the next step as shown in figure 4.10.
Step 6: Select fingerprint test image from test database
On clicking ‗Load fingerprint image‘, system will show file open dialog to select
fingerprint test sample. Select fingerprint test sample from test database. The GUI is
shown in figure 4.11.
Page 125
Figure 4.10: Click ‗Load fingerprint image‘ button to load fingerprint test image
Step 7: Enter the weights of face and fingerprint traits in edit controls
After loading face and fingerprint test samples, next step is to enter face and fingerprint
weights. Face and fingerprint scores should be multiplied with appropriate weights. The
sum of both weights should be 1.0. So, weights can be there in the range of 0.0 to 1.0.
Step 8: Press „Match face and fingerprint‟ button to start process of identification
After entering face and fingerprint weights, next step is to press ‗Match face and
fingerprint‘ button. On clicking this button, test face and fingerprint samples will be
compared with master face and fingerprint database and match scores are generated.
After generating match scores, both the scores will be first normalized by user defined
method of normalization. After normalization, face and fingerprint scores are multiplied
with respective weights and summation of this multiplied scores will be done. After this
summation, the score with maximum will be identified and respective person index will
also be identified. Both the highest score and person identity are loaded in edit control
box.
Page 126
Figure 4.11: Selection of fingerprint test sample
Step 9: On completion of identification process, click on „Clear all‟ to clear the content
of axis controls and edit controls
After completion of identification process, next step is to click ‗Clear all‘ button. On
pressing the button, the content of axis controls and edit controls will be cleared. Now,
the user can start for another identification process.
Page 127
Figure 4.12: Displaying test fingerprint image in axis control 2
Figure 4.13: Enter face and fingerprint weights in edit controls
Page 128
Figure 4.14: Press ‗Match face and fingerprint‘ button to start process of identification
Figure 4.15: Click ‗Clear all‘ button to clear content of all controls
Page 129
4.5 Experimentation results
The researcher has performed experiments with above mentioned multimodal biometric
system model using face and fingerprint recognition. Total 30 samples have been selected
from each modality i.e. face and fingerprint. For each person, one sample has been
selected. The experiment was carried with different weight combinations. Range of
weights starting from 0.1 and 0.9, respective for face and fingerprint modality. It goes up
to 0.9 and 0.1, respective for face and fingerprint modality. Score normalization was
applied to have normalized face scores. The results are shown in the table 4.9. The
performance representation is shown in Figure 4.16.
Here is the evaluation of performance using 30 test samples.
faceweight fpweight Success Failure Success rate
(GAR)%
Failure rate
(FRR)%
0.1 0.9 28 2 93.33 6.67
0.2 0.8 27 3 90.00 10.00
0.3 0.7 27 3 90.00 10.00
0.4 0.6 28 2 93.33 6.67
0.5 0.5 26 4 86.67 13.33
0.6 0.4 25 5 83.33 16.67
0.7 0.3 25 5 83.33 16.67
0.8 0.2 25 5 83.33 16.67
0.9 0.1 24 6 80.00 20.00
Table 4.9: Results of experiments of multimodal biometric system with face and
fingerprint recognition
Matlab code representation for evaluation of performance of the system is shown here:
-----------------------------------------------------------------------
%% multimodal_evaluation.m
%% Setting database for face recognition TrainDatabasePath='D:\phd_dtm_practical\Multimodal
Biometrics\TrainDatabase_Face';
load('db.mat');
Path='D:\phd_dtm_practical\Multimodal Biometrics\Testsamples2\'; imagecounter=100; successrate=0; failurerate=0; personidentity=0;
for l=1:30 imagecount=imagecounter+l; testface=[num2str(imagecount) '.pgm'];
Page 130
testface= strcat(Path ,testface); S1 = imread(testface); % imshow(S1); imwrite(S1,'D:\phd_dtm_practical\Multimodal
Biometrics\temp_face.pgm','pgm');
TestImage='D:\phd_dtm_practical\Multimodal
Biometrics\temp_face.pgm';
T = CreateDatabase(TrainDatabasePath); [m, A, Eigenfaces] = EigenfaceCore(T); [OutputName] = Recognition(TestImage, m, A, Eigenfaces);
%% Identifying normalized scores for face recognition facescore = []; facescore1 = []; fid1=fopen('face_score.txt','r'); j=1; facescore=fscanf(fid1,'%f',60);
fclose(fid1);
for i=1:30 j=(i*1)+(i-1); k=j+1; if(facescore(j)<facescore(k)) facescore1 = [facescore1 facescore(j)]; else facescore1 = [facescore1 facescore(k)]; end end
totalscore=0; for j=1:30 totalscore=totalscore+facescore(j); end
%average score avgscore=totalscore/30; disp(avgscore);
%minimum score minscore=min(facescore); disp(minscore);
deviation=avgscore-minscore;
facescore2 = []; for i=1:30 deviation=avgscore-facescore1(i); temp=deviation/avgscore; disp(temp); facescore2 = [facescore2 temp];
Page 131
end
fid2=fopen('finalfacescore.txt','wt');
for i=1:30 fprintf(fid2,'%f \n',facescore2(i)); end
fclose(fid2);
%% Fingerprint matching from FP database testfinger=[num2str(imagecount) '.jpg']; testfinger= strcat(Path ,testfinger); S2 = imread(testfinger); imwrite(S2,'D:\phd_dtm_practical\Multimodal
Biometrics\temp_fp.jpg','jpg');
filename='D:\phd_dtm_practical\Multimodal Biometrics\temp_fp.jpg'; img = imread(filename); if ndims(img) == 3; img = rgb2gray(img); end % Color Images disp(['Extracting features from ' filename ' ...']); ffnew=ext_finger(img,1);
%% FOR EACH FINGERPRINT TEMPLATE, CALCULATE MATCHING SCORE IN
COMPARISION WITH FIRST ONE S=zeros(30,1); hscore=0.0; Identity=''; count=100;
%Open file to write fp matching score fid2=fopen('fp_score.txt','wt');
for i=1:30 second = count+i; S(i)=match(ffnew,ff{i}); %Write fp match score in file fprintf(fid2,'%f \n',S(i)); if S(i)> hscore hscore=S(i); Identity=second; end fprintf([num2str(S(i)) '\n']); end
%close file fclose(fid2); %% Find final scores and indetify person % Assign weights for face and fingerprint
faceweight= 0.5; fpweight= 0.5;
finalscore = [];
Page 132
%Calculate final score by addition of face score and fingerprint score
for %each sample
fid3=fopen('totalscore.txt','wt');
for i=1:30 totalscore=(faceweight*facescore2(i))+(fpweight*S(i)); fprintf(fid3,'%d -- %f \n',imagecount,totalscore); finalscore=[finalscore totalscore]; display(totalscore); end
fclose(fid3);
%Identify maximum score and decide person index for i=1:30 if(i==1) maxscore=finalscore(i); index=1; else if(maxscore<finalscore(i)) maxscore=finalscore(i); index=i; end end end
personidentity =100+index;
if(personidentity == imagecount) successrate=successrate+1; else failurerate=failurerate+1; end
fprintf(['Imagecount is ' num2str(imagecount) '\n']); fprintf(['Personidentity ' num2str(personidentity) '\n']);
fprintf(['successrate is ' num2str(successrate) '\n']); fprintf(['failurerate is ' num2str(failurerate) '\n']);
pause(5); end
-----------------------------------------------------------------------
%%multimodal_weight.m
fid1=fopen('finalfacescore.txt','r'); fid2=fopen('finalfpscore.txt','r'); fid3=fopen('multimodalweightresults.txt','a'); facescore = []; fpscore = []; fcount=1;
Page 133
count=100; tscore=0.0; success=0; failure=0;
faceweight=0.9; fpweight=0.1;
tempface = []; tempfp = [];
facescore=fscanf(fid1,'%f',900); fpscore=fscanf(fid2,'%f',900);
for i=1:30 maxscore=0; person=count+i; totalscore = []; index=0;
for j=1:30 tempface = [tempface facescore(fcount)]; tempfp = [tempfp fpscore(fcount)]; tscore=(facescore(fcount)*faceweight) +
(fpscore(fcount)*fpweight); totalscore = [totalscore tscore]; fcount=fcount+1; if (j==1) maxscore=totalscore(j); index=1; else if(maxscore<totalscore(j)) maxscore=totalscore(j); index=j; end end end index=index+100; if(person==index) success=success+1; else failure=failure+1; end % pause(2); end successrate=(success/30)*100; fprintf(['Successful identification :' num2str(success) '\n']); fprintf(['Failed identification :' num2str(failure) '\n']); fprintf(fid3,'%f \t %f \t %d \t %d \t
%f\n',faceweight,fpweight,success,failure,successrate);
fclose(fid1); fclose(fid2); fclose(fid3); -----------------------------------------------------------------------
Page 134
Figure 4.16: Performance representation of multimodal biometric system
4.6 Conclusion
The results have been shown in section 4.5. Result shows that with more weightage of
face scores, success rate will be lees. But with more weight assignment with fingerprint
scores, success rate can be achieved more. With fingerprint weights of 0.9 and 0.6,
success rate can be achieved up to 93.33 %, which is quite more compared to unimodal
biometric system performance.
References
[1] Arun Ross, Karthik Nandkumar, Anil Jain, ―Handbook of Multibiometrics‖, Springer,
2006
[2] Marcialis, G. L. and Roll, F. ―Fingerprint Verification by Fusion of Optical and
Capacitive Sensors‖, Pattern Recognition Letters, 25(11):1315-1322, 2004
[3] Ross, A., Jain, A. K., and Reisman, J, ―A Hybrid Fingerprint Matcher‖, Pattern
Recognition, 36(7): 1661-1673, 2003
[4] Hill, H., Schyns, P. G., and Akamatsu, S., ― Information and Viewpoint Dependence
in Face Recognition‖, Cognition, 62(2):201-222, 1997
0
10
20
30
40
50
60
70
80
90
100
Successrate
Successrate
Page 135
[5] Brunelli, R. and Falavigna, D., ― Person Identification Using Multiple Cues‖, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 17(10):955-966, 1995
[6] Jain, A. K. and Chandrasekaran, B., ―Dimensionality and Sample Size Considerations
in Pattern Recognition Practice‖, In Krishnaiah, PR. and Kanal, L. N., editors. Handbook
of Statistics, volume 2, pages 835-855, 1982
[7] Maltoni, D., Maio, D., Jain, A. K., and Prabhakar, S., ―Handbook of Fingerprint
Recognition‖, Springer-Verlag, 2003
[8] Anil Jain, Karthik Nandakumar, Arun Ross, ―Score Normalization in Multimodal
Biometric Systems‖, Pattern Recognition, 2005
[9] Ashish Mishra, ―Multimodal Biometrics it is: Need for Future Systems‖, IJCA,
Volume 3 – No.4, June 2010
[10] Hong, L., Jain, A.K., ―Integrating faces and fingerprints for personal identification‖,
IEEE Trans. PAMI 20 (12), 1295–1307, 1998
[11] Slobodan Ribaric, Ivan Fratric, ―A matching score normalization technique for
multimodal biometric systems‖.
[12] M. He, et al., ―Performance evaluation of score level fusion in multimodal biometric
systems‖, Pattern Recognition, 2009
[13] Gian Luca Marcialis, Fabio Roli, ―Score-level fusion of fingerprint and face
matchers for personal verification under ‗stress‘ conditions‖, 14th IEEE International
Conference on Image Analysis and Processing ICIAP 2007, IEEE, pp. 259- 264, 2007
[14] F. Cui, et al., Score Level Fusion of Fingerprint and Finger Vein Recognition,
Journal of Computational Information Systems 7: 16, 5723-5731, 2011
[15] Fawaz Alsaade, ―Neuro-Fuzzy Logic Decision in a Multimodal Biometrics Fusion
System‖, Scientific Journal of King Faisal University (Basic and Applied Sciences)
Vol.11 No.2 1431, 2010
[16] Sarat C. Dass, et al., A Principled Approach to Score Level Fusion in Multimodal
Biometric Systems, AVBPA 2005: 1049-1058
[17] Romaissaa Mazouni, et al., On Comparing Verification Performances of Multimodal
Biometrics Fusion Techniques, International Journal of Computer Applications, Volume
33– No.7:24-29
Page 136
[18] Lam, L. and Suen, C. Y., Optimal Combination of Pattern Classifiers, Pattern
Recognition Letters, 16:945-954.
[19] Xu, L., Krzyzak, A., and Suen, C. Y., Methods for Combining Multiple Classifiers
and their Applications to Handwriting Recognition, IEEE Transactions on Systems, Man,
and Cybernetics, 22(3):418-435.
[20] Daugman, J., Combining Multiple Biometrics,
http://www.cl.cam.ac.uk/users/jgdlOOO/combine/combine.html.
[21] Domingos, P. and Pazzani, M., On the Optimality of the Simple Bayesian Classifier
under Zero-One Loss, Machine Learning, 29(2-3): 103-130.
[22] Md. Maruf Monwar, Marina L. Gavrilova, Multimodal Biometric System Using
Rank-Level Fusion Approach, IEEE Transactions on Systems, Man and Cybernatics –
Part B: Cybernatics, Vol. 39, No. 4, August 2009, pp. 867-878
[23] Ajay Kumar, Sumit Shekhar, Personal Identification Using Multibiometrics Rank-
Level Fusion, IEEE Transactions on Systems, Man and Cybernatics- Part C: Applications
and reviews
[24] Ayman Abaza, Arun Ross, Quality Based Rank-Level Fusion in Multibiometric
Systems, Proc. of 3rd IEEE International Conference on Biometrics: Theory,
Applications and Systems (BTAS), (Washington DC, USA), September 2009.
[25] Divyakant T Meva, C K Kumbharana, ―Comparative study of different fusion
techniques in multimodal biometric authentication‖, International Journal of Computer
Applications, Volume 66, No. 19, March 2013, pp. 16-19