FINGERPRINT

1
FINGERPRINT

• TOPICS COVERED
• Sensors Used
• Representations
• Matching Algorithms
• State of Art
• Research Problems

2
Sensors Used
3
(No Transcript)
4
Basic Types

• Optical Sensors
• Oldest and most widely used
• Solid State Sensors
• Thermal Based Sensors
• Pressure Based Sensors
• Recent rarely used
• Ultrasonic Based Sensors
• Recent rarely used

5
Optical Sensors

• The finger is placed on a coated plate
• Charged Coupled Device (CCD) converts the image
  of the fingerprint
• It also takes a picture of the dirt, greases,
  and contamination found on the finger

6
Optical Sensors

• The process, referred to as
• Frustrated Total Internal Reflection

7
Optical Sensors

• Dirty Fingerprints cannot use system effectively
• Latent prints are leftover prints from previous
  users
• No ESD issues
• Durable to incidental damage

8
Solid State Capacitance Sensors

• The sensor uses solid-state capacitance sensing
  to capture unique fingerprint data
• Finger as one plate
• Surface of sensor as other plate
• Sensor surface - silicon chip containing an array
  of 90,000 capacitor plates with sensing circuitry
  at 500-dpi pitch

9
Solid State Capacitance Sensors

• Veridicom one of the leading players
• Easy Integration into a variety of electronics

10
Solid State Capacitance Sensors

• Very difficult to spoof.
• Immune to day-to-day fingerprint variations
• Low power
• Immune to ambient light
• High image quality
• Scratch resistant

11
Thermal Based

• Infrared to sense the temperature differences
  between the ridges and valleys of the finger to
  create a fingerprint image
• Temperature differential between the skin ridges
  and the air caught in the fingerprint valleys
• No latent prints
• Good Quality Images

12
Thermal Based

• Sweeping needs some user skill
• High power consumption ? to avoid the possibility
  of a thermal equilibrium between the sensor and
  the fingerprint surface.
• AMTEL one of the leading players

13
Pressure-Based Sensors

• Principle
• when a finger is placed over the sensor area,
  only the ridges of the Fingerprint come in
  contact with the sensor piezo array
• pressure sensors generate a 1-bit binary image

14
Pressure Based Sensors

• Works well with Dry as well as Wet skin
• Larger Sensing Area

15
Ultra Sound Based Sensors

• Use High Frequency Sound Waves
• Transmits acoustic waves and measures the
  distance based on the impedance of Finger, Plate
  and Air
• Ultrasound can penetrate through many mediums
• Considered perhaps the most accurate of the
  fingerprint technologies

16
Acquisition Problems

• Regular Scratches
• Skin Peeling due to weather conditions
• Natural Permanent creases
• Temporary Creases
• Dirty Fingers
• Long Nails
• Ethnic Trait

17
Feature Extraction
18
Fingerprint Features

• Classification
• Distinguishing Characteristics

19
Fingerprint Classification

• On the basis on ridge flow patterns
• Arch, Tented Arch, Whorl and Loop (Right/Left)

20
Distinguishing FeaturesRidge Features and their
Position
21
MINUTIAE

• Points where ridges terminate, bifurcate
• or merge with each other are called
  minutiae points
• In law enforcement 12 -16 matching
• minutiae are sufficient to match a
• person

22
Image Enhancement

• Noise in fingerprint may be due to dry or wet
  skin, dirt, cut or noise of capture device
• Enhancement operations
• Adaptive Matched Filter to enhance ridges
  oriented in the same direction as those in the
  same locality
• Adaptive Thresholding (binarization)

23
Minutiae Extraction Algorithm
24
Feature Extraction

• Original Grey level
• image
• Orientation of the ridges calculated by
• Fourier transform

25
Feature Extraction (Contd)

• Segmentation into foreground and background
• Masking out the background is done in order to
  retrieve the ridges

26
Feature extraction (Contd)

• Finally minutiae points are calculated from the
  ridge image
• Endings have 1 adjacent black pixel ( 8
  neighborhood )
• Bifurcations have more than 2 adjacent black
  pixels
• Finally the minutiae points are superimposed on
  the original image

27
Feature extraction (Contd)

• Minutiae extracted are represented by
• - Their (x,y) coordinate
• - Orientation (T)
• - Forming a 3 tuple (x, y, T)
• - Also the type of minutiae i.e. Ridge ending,
  ridge bifurcation could be stored.

28
Chain coded Ridge Extraction MethodBy Dr
Venugopal, Zhixin Shi John Schneider
29
Chain coded Ridge Extraction MethodBy Dr
Venugopal, Zhixin Shi John Schneider

• Pin vector leading to candidate point P from
  several previous neighboring contour points
• Similarly Pout
• Calculate S(Pin , Pout) lt x1y2 x2y1
• S(Pin , Pout) gt 0 Left Turn and S(Pin , Pout) lt
  0 Right Turn
• Threshold

30
Tessellated approach

• Equal sized non-overlapping windows over
• the image and normalizing pixel intensities
• within the window to constant mean and
  variance.
• Windows of size 3030
• Bank of 8 Gabbor filters is applied to each
  window
• Absolute average deviation of intensity in each
  filtered cell is treated as a feature value
• Thus 8 Feature values for each cell
• Feature values from all cells concatenated
  inorder to form feature vector of the image.
• For a 300 300 image 648d feature vector.

31
Matching Algorithms

• Fingerprints represented by Minutiae points
• Simplest Method Point Pattern Matching
• Requirement
• Correspondence between Template and Input
• No Deformations
• Every Minutiae Localized
• Not Realistic

32
Matching Algorithms

• Requirement of the Matching Model
• Different Locations
• Different Orientations
• Different Pressure
• Spurious Minutiae
• Missing Genuine Minutiae
• Linear / Non-linear perturbation of pair of
  minutiae

33
Matching Algorithms

• Different Approaches
• Image Based
• Graph Based
• Ridge Based
• Minutiae Based

34
Point Based Matching

• 1 . Relaxation Method
• Iteratively adjust confidence level
• Inherently slow due to Iterative property
• 2. Hough Transform Method
• Detecting Peaks in Transformation parameter Space
• If only a few minutiae points, difficult to
  accumulate enough evidence for a match

35
Point Pattern Matching

• 3. Energy Minimization Approach
• Correspondence between pair of points by using
  an energy function
• Energy function based on initial set of possible
  correspondences
• Very Slow ? unsuitable for real-time applns.
• 4. Tree-pruning Approach
• Search over a tree of possible matches
• Strict requirements equal number of points
• Impractical requirements

36
Point Pattern Matching

• Alignment Based
• Alignment Stage
• Transformations determined for alignment
• Matching Stage
• Elastic String Matching Algorithm

37
Alignment Based Matching

• ALIGNING
• Corresponding point pairs
• Exhaustive test
• Large Number of tests
• Impractical though Feasible
• Aligning Minutiae by aligning Ridges

38
Ridge Alignment
39
Post Alignment Matching

• Counting the number of overlapping points if
  exact overlap
• Elastic Algorithm tolerating deformation
• Bounding Box
• Minutiae Points as Strings
• Dynamic Programming approach for String Matching
  ( edit distances )
• Distance measure ? penalty for a mismatch
• Adaptive Bounding Box

40
Ridge Based Matching

• Correlation Based ? compare the global patterns
  Ridge and Furrows
• Dont perform very well due to noisy Images
• Invariant Representation needed
• Strength of Ridges at various orientations
• 2D Gabor wavelets

41
Ridge Based Matching
Parameters f -gt Frequency ? Ridge
Frequency Sx, Sy -gt Standard Deviations Theta
-gt Orientation
42
Ridge Based Matching

• Each of 8 Gabor Filters applied
• Standard Deviation Map for each of 8 Images
• For Alignment,
• Weighted Correlation
• Euclidean Distance measure

43
Graph Based Matching

• Clustering Techniques used
• Homogeneous Regions
• Regions with similar Direction
• Using these regions, develop Relational Graphs
• invariant with respect to translation and
  rotation
• Tolerates Partial Matches

44
(No Transcript)
45
Multilevel Matching

• Text Based
• Textual Fields
• age range / color of hair and eye
• Class Based
• 5 classes of Fingerprints
• Ridge Density Based
• Count of the ridges
• Elastic Matching

46
Performance Evaluation

• FVC 2004 Fingerprint Verification Competition
• 4 databases 2 optical, 1 thermal sweeping
  sensor and 1 synthetic
• REJ, FMR, FNMR, ROC, Genuine/Imposter
  distribution
• Enrollment time, Matching time, average and
  maximum template size, memory allocated

47
Best Algorithm

• Winner of FVC2002 Bioscrypt Inc.
• Ridge patterns not ridge endings
• Pattern based templates not minutiae based
• correlation of ridge patterns
• Heavy weights to areas where images are clear and
  highly complex
• Incompatible with minutiae based systems

48
Pressure based Systems

• Pressure sensitive
• Wet or dry fingers
• Captures print of the finger not just image of
  the print
• By Elform OEM Inc.

49
Ultrasonic Fingerprint Technology

• Sound waves reflecting off ridges and valleys on
  the finger
• Oblivious to dirt, grease, ink, moisture, grime,
  or other substances routinely found on fingers
  which cause the most false readings
• Fingerprints of children

50
Ultrasonic Fingerprint Technology
51
Ultrasonic Fingerprint Technology
52
Fingerprinting Children
53
Research Problems
54
Research Problems (1)

• 1 Acquisition Problems
• Image acquisition susceptible to noise
• SOLUTION
• Sensors capable of capturing Fingerprint Image
  invariant of noise

55

• 2 Enhancement Problems
• The Gray Scale Image obtained has to be enhanced
  for further processing
• SOLUTION
• Better Binarization Algorithms
• More Effective Representation Schemes of
  FingerPrint Images

56

• 3 Features to be Extracted
• Deciding the exact features for matching
• Only Global or Local or both
• SOLUTION
• A comparative study of each Feature combinations
  for determining Individuality

57

• 4 Feature Extraction
• The feature Extraction Algorithm should be robust
  to noise
• Should detect false features
• Should capture Maximum possible features

58

• 5 Partial Matches
• Only a few Feature Points captured
• SOLUTION
• Matching Algorithm Based upon trying to Match
  using a subset of actual Feature points

59
Fingerprint Classification

• To search large databases efficiently
• exclusive classification
• 90 in three classes
• Continuous Classification
• Fingerprints not classified into non overlapping
  classes
• Instead as a numerical vector (by K-L Transform)

60

• E- Commerce applications
• Fingerprint generation
• multimodal biometrics (e.g., combination of
  fingerprints and faces),
• combination of multiple matchers
• digital watermarking of fingerprints