Section A
4x3 marks =12 marks
1.
Q1: WHAT ARE THE COMMON LIVELINESS TESTS USED IN INDIA [3 MARKS]
LIVENESS DETECTION:
One of the most frequently used
PAD techniques is called liveness detection. The relevant aspects of this
technique will be preliminarily elaborated and explained within this chapter.
Subsequently, the obtained knowledge will be used to acquire and compare
several different liveness detection techniques to recognize the previously
described attack scenarios against biometric systems. The general task of
liveness detection is to detect whether a biometric probe (e.g. a fingerprint)
belongs to a living subject that is present at the point of biometric capture
[Am15]. Using liveness detection techniques, a reliable recognition of dead
fingers or photographed faces can be established, for example. Consequently,
the risk of successful presentation attacks is significantly reduced. Thus, in
addition to the regular biometric recognition, liveness detection is an
important procedure aiming at an increased reliability of biometric systems. In
the global market, several different methods verifying the liveness of biometric
features are already established.
Among others these methods
include an evaluation of anatomical characteristics, physiological processes of
the human body and involuntary reactions to stimuli as well as various
predictable human behaviors [Am15].
3.1 Hardware and Software-based
Approaches
Typically, liveness detection methods are divided into hardware and
software-based approaches. Giving an example, special medical hardware can be
used to perform an electrocardiogram or pulse oximetry to detect living
subjects. For this purpose, the acquisition of additional sensors such as
devices for measuring body temperature or pulse rate ([LJ09b], pp.924-925) that
needs to be combined with the regular biometric test system is required. As a
consequence, account should be taken of additional costs for acquisition as
well as for routine maintenance. Software-based techniques, by contrast, make
use of biometric data already being captured for biometric recognition of
individuals. In general, these solutions are implemented as supplementary
algorithms being integrated into a consisting biometric system. To give an
example, these algorithms are then applied to the extracted biometric
fingerprint probe in order to detect the deformation of a living finger that is
pressed on the sensor ([LJ09b], p. 925).
3.2 Passive and active Techniques
Besides the distinction of hardware and software-based techniques another
common attempt consists of a separation between passive (non-stimulating) and
active (stimulating) automated liveness detection methods [Am15]. In general,
passive detection techniques make use of biometric probes which were recorded
through a biometric sensor. According to this, further interactions with the
data subject are not necessary. For this, a typical example would be a
temperature or pulse measurement taking place while the biometric probe is
collected [MA14]. Active detection techniques normally require additional
interaction of the biometric data subject with the biometric system. These
further interactions should be requested using challenge response procedures.
The different challenge response approaches can be read in ISO/IEC DIS 30107-1
as they cannot be discussed in this paper [Am15].
3.3 Defense against
Presentation Attacks
There are several different liveness detection techniques
that have already been evaluated on the market and successfully used for
presentation attack detection. Various techniques that behave as possible
countermeasures against previously described attack scenarios will be elaborated
first. Based on the main results of this paper, an overall evaluation of these
scenarios will be carried out in chapter 3.4 subsequently. The table below
contains liveness detection techniques that could be used as counter-measures
for detecting various presentation attacks:
Type: Fingerprint scanner
Challenges: 2D print, dead
finger, artificial finger, capacitive finger
LIVENESS TEST
Passive: pulse measurement*
[Am15], temperature measurement** [MA14], sweat detection [Am15], skin
resistance detection ** [KS13] Active (challenge response): Request of
different fingers in random order [Am15]
Type: Vein scanner
Challenges: 2D print, dead
finger, artificial finger,
Liveness test:
Passive: pulse measurement,
temperature measurement, (sweat detection), skin resistance detection Active
(challenge response): Request of different fingers in random order
Type: Face scanner
Challenges: 2D print, 3D face
mask, video attack
Liveness test:
Passive:** natural eye blinking *
[Am15], natural muscle movements while speaking [MA14] Active (challenge
response): eye closing request [Am15], voice usage request** [MA14], head
turning request** [Am15]
Type: Fingerprint, vein & face scanner
Challenges: 2D print, 3D face
mask, dead body parts, artificial or digital fingers, veins & faces
Liveness test:
Passive: Infrared &
ultraviolet light, thermal scans* [MA14], medical techniques like ECG, pulse
oximetry or blood pressure reading [KS13]
Q2: WHAT ARE THE REASONS AND
NEED FOR COMPRESSION OF DATA ON BIOMETRICS [3 MARKS]
COMPRESSSION IN BIOMETRIC DATABASE
Compression is a method for database protection used to be
in line with legal requirements internationally and in india.
While
there are hardly any legal provisions in the world that are specific to biometric data, despite the very specific
character of such data, the French Data Protection Act of 1978, officially
entitled the "Loi relative à l'informatique, aux fichiers et aux
libertés " [English title: Act on Information
Technology, Data Files and Civil Liberties] sets out specific requirements for
biometric data.
The
"United Nations Resolution" of December 14, 1990, which sets out
guidelines for the regulation of computerized personal data files does not have
any binding force.
On
April 14, 2016, the draft General Data
Protection Regulation was
adopted by the European Parliament. Its provisions will be directly applicable
in all 27 Member States of the European Union and the UK
in May 2018. And biometric data are clearly defined.
In a
nutshell:
1. it establishes a harmonized framework
within the EU,
2. the right to be forgotten,
3. "clear" and "affirmative"
consent,
4. and, amongst other things, serious
penalties for failure to comply with these rules.
Supreme court has dealt with the same in aadhar case:
With respect to
the same, the majority in the Aadhaar Judgment has held as follows:-
(i) Any
authentication records are to be stored only for 6 months. The provision under
Regulation 27(1) of the Aadhaar (Authentication) Regulation, 2016 allowed the
Aadhaar authority to retain the transaction related data for 6 months and
archive the same for 5 years. Now, the authority is only allowed to retain the
data for 6 months.
(ii)
Creation of metabase, as provided for in Regulation 26 of the Aadhaar
(Authentication) Regulation, 2016, has been held to be impermissible in its
current form and has been suggested that the same needs suitable amendments.
(iii)
Interestingly, Section 33(2) of the Aadhaar Act has also been struck down.
Under this provision disclosure of information, including identity information
and authentication records collected was allowed, when disclosed in the
interest of national security.
(iv)
Section 57 of the Aadhaar Act, which enabled body corporate and individual to
seek authentication, has been held to be unconstitutional.
(v) Most
importantly, the importance of bringing out a robust data protection regime in
the form of an enactment, basis the Justice B.N.Krishna Committee Report (with
necessary and appropriate modifications), has been stressed upon. For further reading on aspects
of Data Protection Bill, 2018, read
our earlier blog
As rightly stated
by Justice R. F. Nariman, in the Indian context right to privacy will mainly
cover three aspects “(i) Privacy that involves person i.e. physical privacy,
(ii) informational privacy and (iii) privacy of choice. We can ground physical
privacy and privacy relating to the body in Articles 19 (1) (d) and (e) read
with Article 21; ground personal information privacy under Article 21; and the
privacy of choice in Article 19 (1)(a) to (c), 20 (3), 21 and 25 of the
Constitution of India”. We had written earlier on some aspects of this.
It is clearly
laid down under the Aadhaar judgment referring to the judgment provided by
nine-judge bench that “right to privacy cannot impinge without a just, fair and
reasonable law: It has to fulfill the test of proportionality i.e. (i)
existence of a law; (ii) must serve a legitimate State (as defined under the
Indian Constitution) aim; and (iii) proportionality”.
The first
requirement is an express requirement under Article 21, where no restrictions
on right to privacy or no individual can be deprived of a right to privacy
without a law being in existence. The second requirement of legitimate state
aim makes ensures that the law enacted by the legislature which imposes
restrictions on individual’s right to privacy, falls within the ambit of
reasonable restriction as mandated under Article 14 of the Indian Constitution.
The third requirement ensures that the means that are adopted by the
legislature are proportional in nature to the object and needs sought to be
fulfilled by the law.
In reaching a
conclusion with respect to the above, reference has been placed to the
statement of objects and reasons for Introducing the Aadhaar (Targeted Delivery
of Financial and Other Subsidies, Benefits and Services) Bill, 2016 and
preamble of the corresponding Act (2016), which emphasises on “a good governance, efficient,
transparent, and targeted delivery of subsidies, benefits and services, the
expenditure for which is incurred from the Consolidated Fund of India, to
individuals residing in India through assigning of unique identity numbers to
such individuals and for matters connected therewith or incidental thereto.”
The judgement
confirms that, there is a vital State interest in ensuring that scarce public
resources are not dissipated by the diversion of resources to persons who do
not qualify as recipients. Allocation of resources for human development is
coupled with a legitimate concern that the utilization of resources should not
be siphoned away for extraneous purposes.
Therefore, the
first two requirements of the 3-part test are met by the Aadhaar scheme and the
act i.e. existence of a law and the restrictions imposed on the fundamental
right of an individual by the law serves legitimate state interests.
The Supreme Court
also examined the third criteria of proportionality and held that the Aadhaar
act meets the criteria of proportionality, as all the aspects of
proportionality, stand satisfied.
In view of the
above, it has been held that the Aadhaar act does not lead to any loss of
privacy and is not unconstitutional, subject to certain provisions of the
Aadhaar act which were struck down. The apex court resorted to the following
grounds to establish that the Aadhaar act is constitutional. Firstly, only
minimal information of the applicant is stored, who intends to have an Aadhaar.
Secondly, there is no risk of loss of privacy as the information collected by
the enrolment agencies is transmitted to the CIDR in an encrypted form.
Finally, at the time of authentication, the information does not remain with
the requesting entity.
Also, the
Authority has mandated the use of a registered device (RD) for all
authentication requests. The RD rules out any possibility of the
arbitrary use of the stored biometric, as the biometric captured is encrypted
within a few seconds. The Authority gets only the registered device code for
authentication and does not obtain any information relating to the IP address
or the GPS location from where such information is collected. Further, the
Authority or any entity is barred from collecting, keeping or maintain information
for the purpose of the authentication under Section 32 (3) of the Act. Above
all, there is an oversight by the Technology and Architecture Review Board
(TARB) and Security Review Committee to all the information flow through the
Aadhaar system. These committees endeavor to provide safeguards.
Lossy and lossless are the two methods for compression of
images.
Lossy compression transform an image into another domain and
also quantize and encode its coefficients. It is used to compress multimedia
data such as video, audio and still images. It eliminates the redundant or
unnecessary data.
Lossless compression is used to reconstruct original images
from the compressed data and is used in source code, text documents and
executable programs. Images can be compressed based on some transformations.
Discrete Cosine Transform (DCT) and Discrete Wavelet
Transform (DWT) are the two common transformations.
DCT expresses a sequence of data points in terms of a sum of
cosine functions oscillating at different frequencies. A DCT is a fourier
related transform similar to the discrete fourier transform (DFT), but using
only real numbers.
The discrete wavelet transform (DWT) uses a discrete set of
the wavelet scales and translations obeying some defined rules. This transformation
decompose the signal into mutually orthogonal set of wavelets. The wavelet is
constructed from a scaling function which describes its scaling properties.
Wavelet Scalar Quantization is a DWT based algorithm. WSQ is
an algorithm to compress gray scale fingerprint images. It is a Federal Bureau
of Investigation (FBI) standard for the compression of fingerprint images.
In addition to WSQ, Contourlet Transform is another
fingerprint compression algorithm.
Bezier curves are used in many fields such as industrial and
computer-aided design, vector-based drawing, font design and 3D modeling and
also used in computer graphics to model smooth curves.
Since the curve is in the convex hull of its control points
and it is possible to graphically display the points and also the control
points can be used to manipulate the curve intuitively.
In biometric data processing, data
compression is the process of encoding information using fewer bits (or other
information-bearing units) than an unencoded representation would use, through
use of specific encoding schemes (compression). As with any communication,
compressed data exchange only works and the message makes sense when both the
sender and receiver of the information understand the encoding scheme applied
to the message. Similarly, compressed biometric information can only be
understood if the encoding and decoding method is agreed upon and known by all
parties involved in the transaction.
The
criminal justice communities throughout the world traditionally exchange
fingerprint (friction ridge) imagery data primarily in 8-bit gray-scale and at
500 pixels per inch (19.7 pixels per millimeter) using the Wavelet
Scalar Quantization (WSQ) fingerprint image compression algorithm.
An
emerging class of next-generation systems now support friction ridge imagery at
the resolution of 1000 pixels per inch (39.4 pixels per millimeter).
The NIST
compression study formalizes a JPEG 2000 compression profile for 1000 ppi
friction ridge imagery that can be used by stakeholders to ensure compatibility
and utility of these images.
Q3. WHAT ARE DIFFEREENT TYPES OF
FP SENSORS IN USE [3 marks]
Fingerprint
Sensor Types
There are many types of Fingerprint Sensor Types, but
most commonly used are Optical Fingerprint sensors and Capacitive fingerprint
sensors.
Capacitive sensors
Capacitive sensors use an array capacitor plates to image
the fingerprint. Skin is conductive enough to provide a capacitive coupling
with an individual capacitive element on the array.
Ridges, being closer to the detector, have a higher
capacitance and valleys have a lower capacitance.
Capacitive sensors can be sensitive to electrostatic
discharge (shock) but they are insensitive to ambient lighting and are more
resist contamination issues than some optical designs.
Capacitive sensors use a difference between skin-sensor and
air-sensor contact in terms of capacitive values.
When a finger is placed on the sensor, an array of pixels
measures the variation in capacity between the valleys and the ridges in the
fingerprint.
Optical Sensors
Optical sensors use arrays of photodiode or phototransistor
detectors to convert the energy in light incident on the detector into
electrical charge.
The sensor package usually includes a light-emitting-diode
(LED) to illuminate the finger.
With optical sensors, the finger is placed on a plate and
illuminated by LED light sources. Through a prism and a system of lenses, the
image is projected on a CMOS image sensor. Using frame grabber techniques, the
image is stored and ready for analysis.
Capacitive
·
Greater miniaturization
·
Newer technology
·
Can be embedded into small devices
·
Prone to dirt etc since finger
touches silicon
·
Relatively cheap
Optical Sensors
·
Larger sensing area since
manufacturing large pure silicon chips is expensive
·
More robust. Longer life
·
More expensive
·
Better image quality and higher
resolution
Fingerprint Sensor Types
|
A
Comparison : Optical vs. Capacitive (Semiconductor) Fingerprint Sensors
|
||
|
Optical Sensors
|
Capacitive Sensors
|
|
|
Sensor Type
|
Optical
|
Capacitive
(Semiconductor or chip) |
|
Sensor Surface
|
No special treatments or maintenance required
|
Usually needs surface treatments, including ESD and other
protective coatingsCoatings may be uneven, wear out over time, degrade
performance, and shorten product lifetime
|
|
Overall Durability
|
Scratch-proof, unbreakable glass platen made of material
as hard as quartz. Resistant to shock, ESD, and extreme weather
|
Corrodes easily from repeated handling and everyday
exposureSusceptible to damage by electrostatic discharge
Thin silicon chips are inherently
fragile and susceptible to damage by hard external impact and scratches
|
|
Imaging Area, Resolution, and Size
|
Large imaging area (0.5 inch x 0.6 inch)High resolution
(500 dpi)
Large image size (78,000 pixels)
|
Usually smaller imaging area, image size, and resolution
due to greater cost of manufacturing larger, high-quality chips
|
|
Cost-Effectiveness
|
Low manufacturing cost, long life, no maintenance required
|
Consistent quality surface coatings may be expensive to
produceReplacement, maintenance, and downtime costs can add up
|
Q4: WRITE ABOUT
SIGNATURE GAIT AND KEYSTROKES [3 marks]
Signature
3.10 Signature
Signature defines the way
in which one individual writes a specific word (mostly her name or a symbol).
It is one of the oldest forms of biometric and has gained acceptance for
several application scenarios. With the progress of technology, two different
kinds of signature biometric emerged. Offline signature considers geometrical
features of the signature biometric. Online signature provides cue about
dynamic features of hand-written signature. These dynamic features include the
speed and the acceleration of the pen movement, pen-up and pen-down times, etc
Two different feature
extraction approaches are present for handwritten signatures as they aim to
capture the static or the dynamic features. Geometrical features of the
signature are considered for the static approach. Dynamic features of
hand-written signature includes the speed and the acceleration of the pen
movement, penup and pen-down times,
3.11Gait
The posture and the
way a person walks maintain uniqueness about the person. It is non-invasive and
hard to conceal. It can be easily captured even at public places at a low
resolution. Unlike other biometric systems, the individual is not required to
pay any attention while her gait is being captured. As the gait can be captured
from a distance, it becomes very useful for security. It requires detection of
the subject, silhouette extraction, extraction of features, selection of
features and classification. A lot of research has been carried out in gait
recognition system. Body part (specially, feet and head) trajectories were
introduced for extracting gait features.
A gait energy image will lead to a set of view intensive
features [Liu and Sarkar, 2006]. Recognition is carried out by a match of
similar patterns in the gait energy image. Silhouette quality quantification
has a key role in the method in [Han and Bhanu, 2006]. A one-dimensional
foreground sum signal modeling is used in [Han and Bhanu, 2006] to analyse the
silhouettes. Segmenting the human into components and subsequent integration of
the results to derive a common distance metric has lead to an improved
performance for gait recognition [Vera-Rodriguez et al, 2013]. Using a
population based generic walking model, a gait recognition system attempts to
solve the challenges which the system encounters due to surface time, movement
speed, and carrying condition [Liu and Sarkar, 2005]. The concept of point
cloud registration has been proposed in [Lee et al, 2009] in order to avoid the
problem of occlusion during gait recognition. Moreover, face and gait
characteristics have been extracted using principal component analysis from a
side face image and gait energy image, respectively. A set of synthetic
features have been attained by integration of these features and application of
multiple discriminant analysis. Performance improvement is achieved in comparison
to individual biometric features [Tan et al, 2006].
3.13 Key Stroke
It is believed that there is a pattern about how a person
types on a keyboard. This behavioral biometric, which is referred as key stroke
dynamics, shows traits to identify or verify a person. But it is not considered
to be as unique as several other biometric traits.
Section
B 9
marks x 2 questions = 18 marks
5a: WHAT ARE THE CHALLENGES IN Facial
recognition
Facial
recognition is the most natural means of
biometric identification. The face recognition system does not require any
contact with the person.The 1000 million electronic passports in
circulation in mid 2017 provide a huge opportunity to implement face
recognition at international borders. Guidelines to improve the quality of the
reference picture embedded in the epassport micro-chip are
provided by the ISO/IEC 19794-5 standard and used by the International
Civil Aviation Organization 9303 standard for passport photographs.
According to the Keesing Journal of
documents & Identity (March 2017) , 2 complementary topics have been
identified by standardization groups.
·
Make sure the captured image has been
done from a person and not from a mask, a photograph or a video screen,
(liveliness check or liveness
detection)
·
Make sure that facial images (morphed
portraits) or two or more individuals have not been joined into a reference
document such as a passport.
The risks of error are related to very
different factors.
·
Take the example of a person with their
biometric characteristics. We have noted that particular biometric techniques
were more or less well suited to certain categories of persons. The difficulties
are related to ergonomic factors of which we do not yet have a firm grasp or
understanding. A certain system may work for women, but less well for men, or
for young people but not for older people, for people with lighter skin, but
less well for those with darker skin.
·
Other difficulties arise in particular
with facial recognition, when the person dyes or cuts their hair, changes the
line of their eyebrows or grows a beard. We can imagine cases of "false
acceptance" when the photo taken modifies distinctive character traits in
such a way that they match another item of biometric data stored in the
database.
·
Other errors are also possible
depending on the technologies used during the biometric enrollment phase. A
verification photo taken with a low-quality model of camera can noticeably
increase the risk of error. The accuracy of the identification relies entirely
on the reliability of the equipment used to capture data.
·
The risk of error also varies depending
on the environment and the conditions of application. The light may differ from
one place to another, and the same goes for the intensity or nature of
background noise. The person's position may have changed.
In the laboratory, under perfect
conditions, in a controlled environment and using adapted technologies, the
rate of error in detection of a face varies between 5 and 10 %.
In addition, in a biometric control application,
the rejection or acceptance rate are intertwined and can be tuned according to
an acceptable level of risk. It is not possible to modify one without impact
the other one.
In the case of a nuclear plant access
control application, the rate of false acceptance will be extremely reduced.
You don't want ANYONE to enter by chance.
This will also impact the rate of
false rejections because you will tune the system to be extremely accurate. You
will probably use several authentication factors including a valid ID in
addtional to biometrics (single mode or multimodal).
5b] WRITE ABOUT IRIS
ACQUISITION, SEGMENTATION,BOUNDARY DETECTION AND POLAR MAPPING[ 9 marks]
3.2 Iris
Iris is considered to be another
reliable biometric trait. The iris is the muscle in the eye which controls the
pupil size in order to regulate the amount of light rays entering the eye. The
iris can be identified as an annular region in between the sclera (white
portion of the eye) and the pupil (Figure 3). The pattern of iris of an
individual is also unique [Jain et al, 2004]. A set of twins also possess
distinguishing iris patterns. The speed and accuracy of iris recognition have
also caused widespread adoption of iris biometric. Registration of iris takes
relatively longer time as it needs several iris images.
A test template is
generated upon scanning of an individual’s iris. Subsequently, the produced
template is matched with the existing templates which were produced at the time
of registration. Zero crossing representation of the one-dimensional wavelet
transform has been proposed in [Radu et al, 2012] to encode the texture in
iris. In [Sun et al, 2005], an integration of Local Feature Based Classifier
(LFC) and an iris blob matcher increases the accuracy of iris recognition. The
noisy images are detected by the iris blob matcher.
Therefore, it helps in the
situations where the LFC does not guarantee the performance. In [Dong et al,
2011], a set of training images is used to learn classspecific weight map for
iris matching technique. Experiments have revealed the effectiveness of the
technique. Moreover, a preprocessing method has been suggested to enhance the
performance of iris biometric on mobile phones which usually are constrained by
computing power. The system has demonstrated its capability to decide whether
the person is dead or alive.
Question
6a] What
should biometric vendors do to accelerate the deployment of biometric
technologies? 9 marks
Answer
Biometric vendors must continue to integrate closely with three As solution vendors such as IBM, RSA, Netegrity and CA.
Biometric vendors must continue to integrate closely with three As solution vendors such as IBM, RSA, Netegrity and CA.
Biometric
vendors should partner with online information solutions, including portals,
online exchanges and other intranet infrastructure providers.
Targeting
financial services and health care verticals will lead to early wins and
long-term recurring revenue.
Targeting
universities and health clubs will help educate and broaden the acceptance of
biometric technologies among a diverse user base.
These
systems should be sold at deep discounts in order to provide the technology as
an educational and viral marketing activity.
Vendors
should also focus on wireless technology providers. They will be integrating
voice biometrics into cellphones to enable strong authenticated purchasing.
Question 6b discuss FP FUSION
TECHNIQUES BASED ON MINUTIAE, RIDGES, OR OTHER FEATURE EXTRACTION TECHNIQUES
With the increasing emphasis on the
automatic personal identification applications, biometrics especially
fingerprint identification is the most reliable and widely accepted technique.
In this paper Fingerprint Verification based on fusion of Minutiae and Ridges
using Strength Factors (FVMRSF) is presented.
In
the preprocessing stage the Fingerprint is Binarised and Thinned.
The Minutiae Matching Score is determined
using Block Filter and Ridge matching score is estimated using Hough Transform.
The strength factors Alpha (α) and Beta (β)
are used to generate Hybrid matching score for matching of fingerprints.
Levels of Fusion in Multimodal Biometric
System
A detailed classification of
various fusion techniques for multimodal biometric can be found in [Dinca and
Hancke, 2017]. Fusion in multimodal biometric can occur at various levels –
such as sensor, feature, matching score, rank and decision level fusion.
Each of these is explained in
this section.
5.1 Sensor Level Fusion
This fusion strategy directly
mixes raw data from various sensors (Figure 6) – for example, from the iris and
fingerprint sensors.
Raw information is captured at
several sensors to fuse at the very first level to generate raw fused
information.
Sensor level fusion strategies
can be put into following three groups: (i) single sensor multiple instances,
(ii) intra-class multiple sensors, and (iii) inter-class multiple sensors.
In the case of single sensor
multiple instances, a single sensor captures multiple instances of the same
biometric trait. For example, a fingerprint sensor may capture multiple images
of the same finger to reduce the effect of noise. Simple or weighted averaging,
and mosaic construction are some of the common fusion methods in this case
[Yang et al, 2005].
Multiple sensors are used to
acquire multiple instances of the same biometric trait in the intraclass
multiple sensors category [Yang et al, 2005]. For example, a 3D face image is
obtained by using multiple face images taken from various cameras.
In the case of inter-class
multiple sensors, two or more different biometric traits are used together. For
example, images of palmprint and palm vein can be fused together for biometric
recognition. Mosaicing is a nice application of sensor level fusion.
Several researchers have proposed
fingerprint mosaicing. It provides a good recognition accuracy as it combines
multiple images of the same fingerprint. Therefore, it can handle the
difficulty in recognition due to data quality. The fingerprint mosaicing
technique uses a modified Iterative Closest Point (ICP) [Jain and Ross, 2002]
algorithm to generate 2D or 3D surfaces by considering the inputs from multiple
instances.
In [Fatehpuria et al, 2006], a
touchless fingerprint system is developed using a 3D touchless setting with
multiple cameras and structured light illumination (SLI) to generate 2D
fingerprint images and 3D fingerprint shape. This kind of set up is expensive
due to deployment of multiple cameras. Alternatively, usage of a single camera
and two mirrors are suggested in [Choi et al, 2010]. Two mirrors have been used
to obtain finger side views.
Sensor level fusion is generally
applied for the same trait. There are also instances of applying sensor level
fusion for different traits. Few of these are mentioned here. Face and
palmprint images are combined in [Jing et al, 2007]. Pixel level fusion is
preceded by Gabor transform of the images. Infrared images of palmprint and
palm vein are fused in [Wang et al, 2008]. At first, image, registration is
carried out on these images. Subsequently, a pixel level fusion takes place.
5.2 Feature Level Fusion
Features extracted from several
biometric traits are integrated into a single vector. According to this fusion
strategy, biometric sensor signals (from camera or microphone) are preprocessed
and feature vectors are derived from them independently. Then a composite
feature vector is generated by combining these individual feature vectors .
The feature level fusion exhibits
a better performance than score level and decision level fusion techniques as
feature level fusion techniques directly deals with the unique biometric
features. Normalization and selection of features are two important processes
in feature level fusion. Min-max technique and media scheming based
normalization is carried out to change the scale and location of feature
values. Scale invariant feature transform is also carried out from the
normalized images.
Dimensionality reduction through
appropriate feature selection also enhances the accuracy of the techniques.
Sequential forward selection, sequential backward selection, and partition
about medoids are standard feature selection techniques. Particle Swarm
Optimization (PSO) is applied on the feature vector for dimensionality
reduction. The multimodal biometric techniques in [Raghavendra et al, 2009;
2011] uses this concept while combining face and palmprint features.
Incompatibility of the feature sets among different biometric traits and
non-linearity of the joint feature set of different biometric traits poses
challenges for feature level fusion. A feature vector can be generated using
weighted average of multiple feature vectors if those vectors correspond to
same biometric. For example, this becomes possible if all of these vectors are
obtained from fingerprint images of an individual. If these vectors correspond
to different biometrics, then they are concatenated to obtain a single vector.
Another example of feature level fusion can be found in [Kim et al, 2011].
Simultaneous use of time-of-flight (ToF) depth camera and near infrared (NIR)
camera acquires face and hand vein images in a touchless acquisition set up.
Several multimodal system also
combines face and ear, as ear is considered to be one of the most unchangeable
feature of the human traits. Unlike face, human ear is not generally affected
by age. PCA based feature extraction and a sparse representation method for
feature level fusion is proposed in [Huang et al, 2013]. Experimental results
reveal that this technique performs better than their unimodal components.
Experiments also show that the performance is similar to that of the unimodal
systems even if one of the modality is corrupted. Local 3D features (L3DF) are
generated from ear and frontal face images in [Islam et al, 2013]. Feature
level fusion is applied in these cases. A matrix interleaved concatenation
based new approach is presented in [Ahmad et al, 2016] for face and palmprint
biometrics. Discrete Cosine Transform (DCT) is used here to extract the
features. Then, these features are concatenated in an interleaved matrix which
estimates the parameters of the feature concatenation and exhibits their
statistical distribution.
A fingerprint and iris based
multimodal biometric recognition technique has been proposed in [Gawande et al,
2013]. Minutiae and wavelet features are extracted from fingerprint images.
Haar wavelet and block sum techniques produce features from iris images. A
feature level fusion of these four feature vectors exhibit better performance
than a unimodal fingerprint or iris biometric. A feature level fusion of
fingerprint and palm biometric traits has been proposed in [Mohi-ud-Din et al,
2011].
Another example of feature level
fusion can be found for hand geometry recognition in the contactless
multi-sensor system in [Svoboda et al, 2015] using an Intel RealSense 3D
camera. This technique carries out foreground segmentation of the acquired hand
image to determine the hand silhouette and contour. Then, the fingertips and
the valleys are located alongside determination of the wrist line from the
identified contour. Subsequently, two features vectors have been formed as
follows: (i) comprising finger length, width, and wrist valley distances and
(ii) finger widths as computed using a traversal of the overall hand surface
and median axis to surface distances. A finger and finger-vein based system has
been proposed in [Yang and Zhang, 2012]. Gabor features are extracted and the
feature fusion strategy is based on a Supervised Local-Preserving Canonical
Correlation Analysis (SLPCCAM). In [Yan et al, 2015], feature level fusion has
also been used for a contactless multi-sample palm vein recognition technique.
Automated access control systems in buildings and other secure premises are
based on the capability of identifying an individual from a distance. Because
of its importance in securing important establishments, it is emerging as an
area of interest to the research community. Improper lighting condition or
not-so-high resolution surveillance cameras pose a constraint for recognizing
an individual based on her face. Use of multimodal biometric involving face and
gait exhibits better performance [Ben et al, 2012; Huang et al, 2012].
Unlike traditional methods of face
and gait multimodal biometric recognition, [Xing et al, 2015] fuses the
features without normalization using coupled projections. A Robust Linear
Programming (RLP) method [Miao et al, 2014] for multi-biometric recognition
exhibits good results in noisy environments in spite of using less training
data. It uses uncertain constraints and concatenates heterogeneous features
from different biometric traits. Each biometric modality has been assigned a
weight to specify its degree of contribution in the fusion. More weight is a
greater relevance of the corresponding biometric trait. In the multimodal
biometric recognition system by [Chang et al, 2003], features are extracted
from face and ear using Principle Component Analysis (PCA). Subsequently,
fusion takes place at the feature level. A method of liveliness detection to
prevent spoofing attack in [Chetty and Wagner, 2005] also uses a feature level
fusion. In a recent development, sparse-based feature-fusion [Huang et al,
2015] of physiological traits has drawn sufficient interest of researchers due
to robust performance.
5.3 Matching Score Level Fusion
In the case of matching score
level fusion, matching scores are separately obtained for each biometric trait
and subsequently fused to arrive at an overall matching score. Matching score
level fusion is also referred as measurement level fusion.
There exists three different
approaches for matching score based fusion – density based, classification
based, and transformation based.
The density based scheme is on
the basis of distribution of scores and its application to popular models like
Naive Bayesian and Gaussian Mixture Model [Murakami and Takahashi, 2015].
In the classification based
approach, the matching scores of individual matching modules are concatenated
to obtain a single feature vector. The decision to either accept or reject an
individual is based on the classification of this feature vector.
According to the transformation
based approach, the scores of individual matching module are, at first,
transformed (normalized) into a pre-decided range. This transformation changes
the position and scale parameters of the matching score distribution so that
these normalized scores can be combined to obtain a single scalar score [Murakami
and Takahashi, 2015].
These normalization techniques to
handle the dissimilarities in matching score also draw attention of
researchers. In another noted development of recent time, an order-preserving
score fusion method has been proposed in [Liang et al, 2016].
5.4 Rank Level Fusion
Ranking the potential matches
between the query template and the templates in the database generates an
ordered list of all templates in the database. The first choice is the match.
These ranked are obtained for every biometric trait. In a multimodal biometric
recognition system, these rank orders are fused to generate a final ranking of
each template in the database. Unlike score level fusion, normalization is not
required for rank level fusion.
Rank level fusion is applied in
[Kumar and Shekhar, 2011] for combining various methods for palmprint
identification. A Nonlinear Weighted Ranks (NWR) method aggregates the ranks as
obtained from individual matching modules. Rank level fusion may not always
perform well in noisy conditions having low quality data. Though it has been
applied on low quality fingerprints [Abaza and Ross, 2009]. It applies a
derivation of the Borda count method, which involves image quality. This
approach has a similarity with logical regression. But unlike logical
regression, this approach does not need a training phase. Image quality is
considered in this approach instead of weights. As ranks are assigned to only
few of the stored templates with possible match, the rank level fusion may create
a challenge for large databases. Ranks do not cover every template in the
database. In this context, a Markov chain based method has been proposed in
[Monwar and Gavrilova, 2011] for rank level fusion. Markov chain is used to
represent a stochastic series of events, where the present or the preceding
states determine the next state. A graph is used to formally model the Markov
chain. A vertex in the graph represents a state or an event. An edge in the
graph denotes the transition from one state to another state. At first, ranks
are generated for each biometric trait. If the matching module creates partial
ranking (for example, the first three ranking results), elements are inserted
randomly to complete the list. The state transition probabilities are computed
and the stationary distribution of the Markov chain is obtained. The templates
in the database are ranked based on a decreasing order of the scores of the
stationary distribution starting from the highest score. This fusion strategy
is applied on a multimodal biometric recognition system involving iris, ear,
and face. [Monwar and Gavrilova, 2009] proposes another method to solving this
discussed problem of rank level fusion. The multimodal biometric recognition
system has three matchers for each of signature, face and ear. However, fusion
is carried out between the identities put out by at least two matchers.
5.5 Decision Level Fusion
In the case of a decision level
fusion, each individual matcher, at first, takes its own decision.
Subsequently, the fusion of various biometric modalities takes place by
combining the decisions of these individual matchers. Hence, each biometric
trait is independently pre-classified and the final classification is based on
the fusion of the outputs of the various modalities (Figure 9). The simplest
forms of decision level fusion uses logical operations such as ‘AND’ or ‘OR’.
Some advanced fusion strategies at this level also use behavior knowledge
space, the Dempster-Shafer theory of evidence, and Bayesian decision fusion.
When every individual decision
module supplies positive outcome, then the ‘AND’ rule positively recognizes a
query template. Otherwise, the ‘AND’ rule rejects the query. Hence, ‘AND’ rule
is generally reliable with extremely low false acceptance rate (FAR). But false
rejection rate (FRR) is higher than that of individual trait. On the contrary,
the ‘OR’ rule provides positive output about a query template when at least one
decision module gives positive response about it. As a result, FRR is extremely
low and FAR is higher than individual trait. In [Tao and Veldhuis, 2009], an
optimized threshold method has been proposed using the ‘AND’ and ‘OR’ rule. The
thresholds of the classifiers are optimized during the training phase.
Majority voting is another common
approach for decision fusion. If the majority of the individual traits decide
positively about the query template, then final decision is positive. Majority
voting method gives equal importance to each individual decision modules.
Otherwise, a weighted majority voting can be applied. In this method, higher
weights are assigned to decision modules which perform better.
A multi-algorithm decision level
fusion is used in [Prabhakar and Jain, 2002] for fingerprints. This method
considers four distinct fingerprint matching algorithms. These are based on
Hough transform, string distance, 2D dynamic programming, and texture. This
method selects appropriate classifiers prior to applying decision fusion. A
threshold for each individual classifier influences the outcome of a decision
level fusion. Here, threshold specifies a minimum score to decide whether the
sample is genuine or an impostor. If the matching score of the sample is higher
than the threshold, then the sample is considered as genuine. On the contrary,
if the matching score of the sample is less than the threshold, then it belongs
to an imposter. The classifiers are assumed to be independent of one another in
some biometric systems. However, other works have assumed the dependency among
the classifiers. A verification system has been introduced in [Veeramachaneni
et al, 2008] based on two fusion strategies for correlated threshold
classifiers. Between these two strategies, Likelihood Ratio Test (LRT) still
depends on the threshold of each individual classifier. The Particle Swarm
Optimisation (PSO) based decision strategy is considered effective in
comparison. Even the PSO strategy performs better than some of the score level
fusion methods. A real time sensor management using PSO is suggested in [Veeramachaneni
et al, 2005] for a multimodal biometric management. This method performs a real
time search of the optimal sensor configuration and optimal decision rule. A
similar concept is proposed in [Kumar et al, 2010] which uses Ant Colony
Optimization (ACO) technique, for a multimodal biometric system involving
palmprint and hand vein. [Kumar and Kumar, 2015] extends this experiment on
multiple multimodal databases involving palmprint and iris, fingerprint and
face, and face and speech. Another decision level fusion for multimodal
biometric recognition is proposed in [Paul et al, 2014]. In this multimodal
system, signature, face and ear biometric features are combined with social
network analysis. The Fisher image feature extraction, which is a combination of
PCA and Linear Discriminant Analysis (LDA).
5.6 Hybrid Fusion Model
A hybrid fusion model which uses
both pixel level fusion and score level fusion demonstrates good performance in
[Kusuma and Chua, 2011]. This multi-sample face recognition (both in 2D and 3D)
in [Kusuma and Chua, 2011] recombines images using principal component analysis
(PCA). Two recombined images are fused using a pixel level fusion scheme.
Additionally, score level fusion is also applied to produce good result.
No comments:
Post a Comment