Iris Recognition - PowerPoint

					Iris Recognition




    Steve Bruce
    Tammy Tran
                    Introduction




   Iris: colored portion of the human eye
   Muscle that regulates pupil size
   Unique coloring based on melatonin pigment
   Formed prior to birth and structurally distinct
           Definition and Usage
   Process of recognizing a person by analyzing the
    random pattern of the iris
   Allows for one-to-many comparison
   Automated method first patented in 1994
   United Arab Emirates – 12 Billion comparisons
    of arriving travelers done daily
   U.K. – Iris Recognition Immigration System:
    travelers bypass customs (200,000+ by 2008)
                     History
   Iris divination dates back to early Western-Civ
   Alphonse Bertillon (1885): First to suggest the
    iris as a means of human identification
    James Doggart (1949): Suggested iris patterns
    sufficiently unique to serve like fingerprints
   Flom and Safir (1987): Patented Doggart’s
    concept, but had no algorithm or method to
    make it possible
                History (Cont.)
   John Daugman (1994): Awarded patent for his
    automated iris recognition algorithms
   Commercialization efforts stumble (1993-2006):
     Inferior cameras used
     Disputes over licensing terms lead to lawsuits

     Inability to develop new algorithms

   Patent on Daugman’s algorithms expires this
    year
           Expanding Technology
   Best Biometric  Database accuracy and search speed
   National programs for Biometric ID cards/passports
   NIST Iris Challenge Evaluation – 42 research groups
   Biometric Data Interchange Format Standards
      Databases of iris images for algorithm development and
       testing
   International conferences and books on iris recognition
   Popular futurism and movies (i.e. Minority Report)
   Cultural iconography associated with the eye
      “Window to the Soul”

      Non-verbal communication through eye contact

   Draw of challenging intellectual pursuits
                     Iris Scanners
   High-quality digital
    cameras
   Typically use infrared
    light to illuminate the iris
   Image then filtered and
    mapped into vectors that
    determine orientation,
    spatial frequency, and
    position
   Generally more reliable
    than fingerprint scanners
Iris Recognition: How?
     To Begin Iris Recognition…
   Find the iris in an image
   Determine inner and outer boundaries
   Detect occluding upper and lower eyelid
    boundaries
   Detection and exclusion of…
     Superimposed eyelashes
     Reflections from the cornea or eyeglasses

   Overall process known as segmentation
Importance of Accuracy

              Iris mapping critically
               dependent on finding
               true boundaries

              Inaccuracies in detection,
               modeling, or
               representation can lead
               to misidentification
                    Challenges
   Iris as an annulus… not really
      Inner and outer boundaries usually not concentric

      Boundaries are also typically non-circular

      Cannot be mapped to a traditional polar-coordinate
        system
   Significant improvement in performance when initial
    assumptions are relaxed
      More disciplined methods for detection and
        modeling of boundaries
      More flexible and generalized coordinate system
               More Challenges
   Boundaries often occluded…
     Outer boundaries by eyelids
     Inner boundary by reflections from illumination

     Both by reflections from eyeglasses

   Goals:
     Fit flexible contours
     Boundary models must form closed curves

     Smoothness constraint
              Occlusion Solution
   Describe iris boundaries in terms of “Active Contours”
   Fourier analysis and low-pass filtering
     Correction for Off-Axis Gaze
   Current iris recognition cameras require on-axis
    image of an eye
   “Stop and Stare” interface which aligns optical
    axis with camera’s optical axis
    – NOT FLEXIBLE
   Off-axis deformation corrections done by
    “reliably estimating actual parameters of gaze”
     “Fourier-based Trigonometry”
-Reverses projective geometric deformation caused
   by gaze deviation by an affine transformation
   of the off-axis image.

1.   Determine parameterized coordinate vectors of
     pupil boundary:

     where:
     “Fourier-based Trigonometry”
2.   Compute direction of gaze deviation:




3.   Compute the “magnitude” of gaze deviation:
   Off-Axis Gaze Transformation
Result:                       Before Transformation:
 Converts deviated
  images into apparent
  orthographic form

Limitation:                   After Transformation:
 Method of
  transformation assumes
  iris is a planar surface;
  not curvature
          Occlusion by Eyelashes
   Random & complex
    shapes makes it difficult
    to be detected by
    “elementary shape
    models”
   Can be strongest signals
    in iris image; hence can
    hinder IrisCode with
    false data
            Statistical Inference
   Solution: Statistical estimation methods used to
    determine distribution of iris pixels
   Based on significant Z-score deviation testing
   Tests acceptance of hypothesis:
    H0: iris contains superimposed eyelashes
   Once eyelashes are detected:
     Eyelashes marked with white pixels
     Positions recorded to prevent from influencing data
    Test of Statistical Independence
   Iris recognition works by performing a test of
    statistical independence between IrisCodes
   Test equivalent to “tossing a coin”
      50-50 outcomes = independent irises
      Great number of corresponding bit pairs =

       strong evidence IrisCodes came from same iris
           Score Normalization
   Interpretation of outcome must take into
    account total amount of comparison data
    available.
   Benefit: “to prevent False Matches arising by
    chance due to few bits being compared”
   When only a few bits available, normalization
    may reject same-eye matches due to HDnorm
    (standard score) being above acceptance
    threshold from low demanding False Match
    Rates
            Score Normalization
   NIST ICE-1 iris database
    consists of a few
    thousand iris images
   False Reject Rates vs.
    False Accept Rates
   Equal Error Rate where
    FRR = FAR
   Best performance
    achieved WITHOUT
    score normalization
                                  Algorithm 1 = with normalization
                               Algorithm 2 = without normalization
             Score Normalization
   Based on 200 billion iris
    cross-comparisons of
    632,500 IrisCodes
   No normalization =
    (no account of # of bits)
   SQRT normalization =
    2,000 times better
            Score Normalization
   Normalization most applicable when involving
    “astronomic numbers of cross-comparisons”
   Relatively small datasets (& irrelevance of Equal
    Error Rate) achieve better performance without
    score normalization due to the number of
    penalties on good matches when few bits are
    compared

				
DOCUMENT INFO