CS1AC16 BIOMETRIC RECOGNITION

Computer vision includes methods for the acquisition, processing, analysis and comprehension of
images in order to produce a numerical or symbolic information. Computer vision is about
interpretation of images from all types of sources. Processing also involves preprocessing e.g. forms of
image enhancement.
An example of analysis and comprehension is working out where people are in surveillance camera and
what they're doing.

Pattern recognition is classification based on image analysis. It's used in most machine intelligence
systems built for decision making.

Getting a computer to recognize a picture is difficult. For humans it's very easy and there is no proof
how exactly it is done so there is nothing to base on. There is some non-conclusive evidence that human
brains perform a form of feature extraction.
Humans can instinctively separate an object from the background but a computer can't do that – to a
computer a picture is just a matrix. The challenge is to get a computer to determine a set of features
that lets it discriminate between classes. There's many different systems to do that, some are based on
what's already known about computer vision, others are based on guesses about computer vision.

Human recognition methods

 Traditional techniques for personal identification
o Knowledge-based (e.g. password or pin)
o Token-based (e.g. a document, a smartcard)
 Biometric-based
o Identification by something you are rather than something you have or know (e.g.
fingerprints, retina)

Biometric recognition is based on anatomical, physiological or behavior-related characteristics.

It's a return to more natural way: "bios" for life, "metros" for measurement

A recognition system can be:

 Multibiometric - using several images if the same biometric
 Multimodalitic – using various biometrics

Recognition:
 One to many – comparing the sample with images in the database to see if it matches with any
 One to one – verification rather than recognition, for example comparing face with passport
photo

Biometric recognition phases:

 Enrolment – data is collected and stored
 Verification/identification - data is compared with the reference data
o Verification – compare face with the passport photo – is that person who they claim to
be?
o Identification – match a face against what's in the database – who is that person?
Examples of modalities include: DNA, ear shape, face, hand veins, gait, fingerprints, iris, retina, voice,
signature. They can be classified as:
 Anatomical/physiological - referring to parts of the body
 Behavioral – relating to conduct
OR
 Genotypic – genetically defined and not changing
 Fenotypic – changing over time and depending on the environment

For a biometric trait to be useful for recognition it should be:

 Universal – everyone should have it
 Unique – different for everyone
 Collectable – possible to collect
 Permanent – not changing over time

Face recognition system:

Face recognition is difficult for a computer mainly because of the variabilities:

 Short term: facial expression, pose, illumination
 Long term: aging, make up, facial hair
Face recognition is usually done by dividing the face into areas, extracting features and comparing with
the samples:

A relatively new way is periocular recognition – identifying people using the region around their eyes –
particularly usable if the full face is not visible. It uses 2D and 3D techniques:
  2D for texture – finding details on the skin
 3D for shape – more efficient against variaitons in facial expression, illumination, partial
occlusion

Iris recognition

Iris recognition is becoming increasingly popular. However, it poses problems with collectivity – it's hard
to get a good quality iris picture.
The steps of iris recognition are:
 Image capturing
 Iris segmentation
 Normalization – it's important in any biometric recognition to normalize out differences in
acquisition that are not important to recognition, e.g. the exact position in the image
 Feature extraction – results in obtaining a biometric signature
 Matching to obtain similarity value.
All of these steps are separate research topics.

The scheme for iris recognition was devised by Daugman.

Segmentation
 An integro-differential operator is used for locating inner and outer boundaries
 Circular edge detection is used for limbic and pupillary contours
Normalization
 Daugman's rubber sheet model is used to convert iris region to polar coordinates
 You sample equal amounts of points to even out the fact that the iris is not the same size
everywhere – you get a rectangular image:

Feature extraction
 Gabor filter – based on how human vision works
 Imaginary part and real part together form iris code which is 256 bytes long
Matching
 The computer looks for differences in bit pattern. A decision to accept or reject is made on the
basis of comparison of that value with a set threshold
 The problem is that if the database is very large, the iris may be falsely recognized – the same
problems arises with face recognition

Iris recognition works very well in constrained acquisition conditions:

 Infra-red illumination of the eye
 User collaboration
 Controlled distance from the camera.
It doesn't work well in unconstrained conditions:
 With background noises
 People moving or at a distance
 You need to figure out where a person is in the picture, then find their face, then the eyes and
finally capture the iris – it's practically impossible to get a good quality picture

Some mobile phones use iris recognition. The sensor is a NIR camera and infra-red illumination is used.
To help iris location, the phone also displays circles for the user to position their eyes reflection.

Iris acquisition challenges:

 Accuracy problems – a small region needs to be captured from a large distance
 The target is moving
 Ensuring enough illumination without harming the eyes
 Occlusions: eyelids, eyelashes, reflections, hair, contact lenses

Fingerprint recognition
The ideal situation is where you enroll and verify on the same sensor – sensor characteristics may vary
across sensors
Methods of acquisition:
 Offline – ink technique
 Live-scan – optical and ultrasound sensors
Fingerprint features
The recognition process as a whole is based on the outcome of consecutive steps:

The initial steps (direction map and foreground mask estimation) influence all the following steps.
Direction map estimation is determining the ridges flow direction and foreground mask estimation is
distinguishing foreground and background regions

Matching methods:
 Correlation based – computing the correlation between corresponding pixels for different
alignments (comparing the image with database pixel by pixel)
 Minutiae based – finding the alignment between the template and input minutiae sets that
results in the maximum number of pairings
 Ridge feature based – comparing in terms of features

Fingerprint recognition is often used in smartphones.

There are different sensor types:

 Touch – whole area scan with no motion – bigger but more reliable
 Swipe – small and portable but less accurate
 Touchless – quality may not be as good but there is much less maintenance needed (no
possibility of transferring dirt onto the sensor) - captures the reflections of light from the
fingerprint
Fingerprint sensing technology
 Optical sensor – relatively big in size, requires camera calibration; used for border control
 Solid state fingerprint sensor – can be compact and integrated to different products (laptops,
doors etc.)
 Multi-spectral imaging sensor – captures surface and deep tissue data
 3D scanner

Fingerprint image quality

Based on the image quality, 3 different feature levels can be observed. The minimum resolution for level
1 and 2 features is 250-300 ppi. 500 ppi is the standard used by FBI. The minimum resolution for level 3
features is 1000 ppi.

Fingerprint acquisition challenges:

Vulnerability to noise and distortion by dirt, perspiration, etc.
Some people have altered or damaged fingerprints
Optical sensors are vulnerable to illumination

Unimodal system performs the recognition based on a single source of biometric information – a single
trait. It can be affected by:
 Noise in the acquired biometric data
 Non-universality of a certain trait
 Lack of individuality
It is also susceptible to circumvention.

Multimodal biometrics combines multiple biometric traits – addressing the problem of non-universality.
It also increases the resistance against spoofing.

The fusion of biometrics can be done at different levels – sensor, feature, score or decision.
An example of multimodality is a system that combines 3 biometrics – face, iris and voice recognition.
How to evaluate efficiency of a biometric recognition system?
Every identity is a class.
 Intra-class comparison – comparing the images of the same person – produces the genuine
score
 Inter-class comparison – comparing images of different people – produces the imposter score
If the score is higher than some decision threshold, then the match is accepted. Otherwise it is rejected.

There are 2 types of error that can occur:

 False match – accepting different people as the same
 False non-match – rejecting the image of the same person

Imposter/genuine distribution

To set the decision threshold you need to decide what type of error is more important to avoid – this
depends on the application of the system. If the threshold is low, it will have a high rate of false
matches. If it is high, it will have a high rate of false rejections.

Confusion matrix

False acceptance rate: FAR = FP / (FP + TN) - the number of false

acceptances divided by all inputs that should have negative outcome
False rejection rate: FRR = FN / (FN + TP) - the number of false rejections divided by all values that
should be positive.
Attacks on biometric system:
 Direct – on the sensor level – presenting a fake sample
 Indirect – everything else in the system, for example tampering with databases or
communication channels

Spoofing methods evolve together with the development of technology. There are many materials to
replicate biometrics, some of which are cheap and easy to make. In addition, people leave their
biometrics in public all the time: fingerprints, images, voice...
Some spoofing examples are: silicon fingers to cheat fingerprints scanner or masks to cheat face
recognition.

Anti-spoofing techniques

Presentation attack detection – detecting attacks at sensor level made by presenting fake samples
Key points:
 Find a set of discriminant features
 Build an appropriate classifier
 Predict the probability of sample vitality given the extracted set of features.

Biometrics is more reliable than methods based on passwords or access cards, as these can be stolen,
lost or forgotten. It is also faster and more convenient.

Biometric recognition application

Examples of biometrics applications: crime investigation, border control, CCTV surveillance, personal
device authentication, office clock-in.

Properties of a biometric trait:

Universality – all people should possess it (some exceptions such as loss of irises due to genetic defect or
illness are acceptable)
Uniqueness – there should be no 2 people with the same characteristic
Permanence – the trait should not change with time
Measurability – a trait should be easy to quantify
Performance – accuracy of identification, time of identification, effects of technology on identification
Acceptability – people's willingness to accept a biometric systems – certain groups may resist using
certain biometric traits
Circumvention – easiness of falsification – a good biometric should be difficult to fake

Border control – Fast Pass project

The current automated border control process it slow – it requires a close-up face capture for good
quality picture. It also requires the person to be standing still in front of the camera.
The challenge is to be able to locate and capture a person's face on the move and from a distance,
reducing the motion blur. There are also additional challenges for outdoor border control – reflections,
sun light, bad weather etc.

Open challenges in real-life biometric recognition

Aging of biometric traits

 As people age, the biometric traits change:
 Face – changes most dramatically
 Fingerprint – the quality of the imprint gets worse
 Iris – changes every 4 years