Professional Documents
Culture Documents
W. A. Barrett, cmpe dept., SJSU vs. 2.0 Theory in a nutshell Segmentation Recognition Verification Fingerprints/Face/Iris/Speaker recognition Logface matching
Theory in a Nutshell
Capture images of objects (usually persons) Segment a view Compress views to biometric codes. Compare two biometric codes, yielding a biometric difference. When two differences are small enough (less than some threshold), the corresponding objects are considered the same. Otherwise the objects are considered different.
A Sample Database
NAME Bill Barrett Dave Matthews Mike Sanders Fred Friendly Bill Clinton BIOCODE 5.9.6.30.7.6 3.9.4.25.9.7 5.8.4.33.6.5 8.4.2.28.7.3 2.6.3.30.6.7
distance j = square(D[j,i] - C[i])/var[i] where D[j,i] = database js component i, 1 <= j <= 5 (rows) C[i] = candidate component i, 1 <= i <= 6 (columns) var[i]= variance from database, component i
Distance Calculation
NAME Bill Barrett Dave Matthews Mike Sanders Fred Friendly Bill Clinton BIOCODE 5.9.6.30.7.6 3.9.4.25.9.7 5.8.4.33.6.5 8.4.2.28.7.3 2.6.3.30.6.7 AVERAGE VARIANCE Candidate: BIOCOD ES 5 9 3 9 5 8 8 4 2 6 4.6 7.2 5.3 4.7 4.8 7.9 6 4 4 2 3 3.8 2.2 3.2 30 25 33 28 30 29.2 8.7 31 7 9 6 7 6 7 1.5 6.2 6 7 5 3 7 5.6 2.8 5.1
Bill Barrett Dave Matthews Mike Sanders Fred Friendly Bill Clinton
DISTANCE D ISTANCES 4.66 0.01 0.26 11.81 0.61 0.26 0.79 0.01 0 8.86 1.93 3.24 3.7 1.48 0.77
Segmentation
Image typically contains background noise Segmentation is isolating a biometric view from the image
Motion segmentation uses video to reject static background pixels Two or more cameras yield distance measures Given a static image, segmentation requires heuristic methods
Static segmentation may be the most difficult design challenge of a biometric system
Recognition
Form an enrolled database of biometric codes each entry represents a different candidate each candidate is associated with a biometric code, name, address, etc. Capture a view of a candidate and compute its biometric code C. Compare C with all candidates in the database. Form a list of database candidates, ordered by increasing biometric distance. Front of the list should be the matching candidates.
Recognition (2)
If the top candidate has a small-enough biometric distance, we say that we have recognized the candidate. If the top candidate's biometric distance is too large, then the candidate has not been recognized. This implies a threshold level has been determined for biometric differences
(good) Top candidate's biocode is small enough, and is the correct person. (bad) Top candidate's biocode is small enough, but is the wrong person (false acceptance) (good) Top candidate's biocode is too large, and this is the wrong person. (bad) Top candidate's biocode is too large, yet this is the correct person (false rejection)
Recognition Goals
Maximize correct matching of a candidate to the database Minimize false acceptance and false recognition
Verification
Candidate presents biometric image PLUS identification information, such as a credit card plus PIN System locates candidate in the database through the credit card/PIN data One biometric distance is computed -- if small enough, the candidate is verified. Can still have a false acceptance or false rejection!
Authentics-Imposters
Biometric quality is measured statistically by acquiring two distributions -Authentics -- distribution of biometric distances of the same persons, but with different images Imposters -- distribution of biometric distances of images of pairs of different persons These should be widely separated, but often aren't
Authentics - Imposters
Authentics-Imposters
The two distributions will overlap in general The extent of the overlap relative to the two areas provides a measure of the quality of this biometric measure Small overlap -- good biometric Large overlap -- poor biometric Best viewed through the accumulated distribution
Authentics-Imposters
Choice of Threshold
Assumes that the relative number of attempts is balanced Moving the threshold to the left means more false rejections, but fewer false acceptances Moving the threshold to the right means fewer false rejections, more false acceptances
Quality Measure
The quality of a biometric measure can be estimated from these two curves
use a good representative sample of measurements (not easily done!) find the crossover point FARR = % at crossover point
View Compression
Fast Fourier transform Gabor wavelet transform Legendre moments Chebyshev moments pseudo-Zernike moments eliminate unwanted view variations (scale, rotation, translation, avg intensity, etc.) produce maximum discrimination, i.e. smallest possible FARR
Legendre Moments
(2 p 1)(2q 1) Lpq 11Pp ( x) Pq ( y) f ( x, y) dx dy 4
1 1
(2n 1) x Pn 1 ( x) (n 1) Pn 2 ( x) Pn ( x) n
P0(x) = 1, P1(x) = x
Legendre Moments
the view can be reconstructed, given enough (p,q) pairs the translation component is in (0,0) face: need to rescale to a normal view, typically done by finding the eyes, etc. face: measure degrades with rotation
pseudo-Zernike Moments
Much more complex set of polynomials Are orthogonal and complete Not scale or translation invariant Certain functions of the moments are rotation invariant
Face Recognition
eigenfaces (Alex Pentland, MIT) feature extraction (Joseph Attick, Identix) some are proprietary uniform lighting conditions full frontal face -- no side views plain expression no attempt at disguise good segmentation, centering the eyes
Face Recognition
verification under controlled conditions (disguise can be used to evade detection, but difficult to fake a verification trial) sifting out a small number of candidates from a larger set recognition critical applications
Fingerprints
For digital prints, the FBI routinely finds persons in their large national database from prints sent through the internet (AFIS) Statistics are unknown, but believed to have a FARR less than 1E-5
Fingerprint analysis for forensic purposes has a much smaller FARR Small or smudged prints (typical of crime scenes) are likely to result in identification errors.
Iris Scanning
Iris Scanning
Locate sclera surrounds pupil Locate upper and lower eyelids Form biocode from iris patterns
Daugman uses 8 bands and a Gabor filtering to yield a 256 byte code
Daugman uses a Hamming distance measure
Distance measure
Iris Issues
Pupil finding is difficult Background light sources reflected in pupil Eyelashes sometimes obscure iris Eyes may be partly closed Eye movements are rapid, may cause image capture failure Telephoto centering and autofocus important Capture system can be expensive
Sensar, Inc.
A New Jersey startup, 1990-2000 period Used the Daugman iris patent Developed extraordinary optics system
two cameras, one wide-angle, the other a telephoto with autofocus and angular tracking system could accurately identify a person as he/she approached an ATM machine
tested in a Fort Worth bank system Sensar failed for various reasons
Speaker Recognition
Starts with an audio sample of a human voice Typically, person is prompted to repeat certain phrases Speech fragment compressed by FFT or wavelet transforms Identification/verification similar to other biometrics FARR ~ 1E-2 at best
Goal -- Match a cut log face to its mating stump U. S. Forest Service interested in combating theft of timber from national forests
start with photo of stump face find stump face in a collection of photographs of faces taken at various sawmills use biometrics to filter out the most likely candidates use forensic tools to indict and prosecute thieves
Color images input by digital camera, many supported image formats Semi-automatic segmentation of log faces
Uses pseudo-Zernike polynomials to obtain a rotation-invariant biometric code Database mysql employed under Linux Friendly user environment for locating matching faces from a database
Logface Results
Selected Bibliography
http://www.biometrics.org -- Biometrics web site http://www.identix.com -- Face recognition, fingerprint vendor http://www.iritech.com -- Daugmans iris scanning company, patent holder John Daugman, patent no. 5,291,560, Iris scanning patent Wechsler et al, editors, Face Recognition Maltoni, Maio, Jain & Prabhakar, Handbook of Fingerprint Recognition, Springer, 2003. Mukundan & Ramakrishnan, Moment Functions in Image Analysis, World Scientific, 2003 Duda & Hart, Pattern Classification and Scene Analysis, Wiley Interscience Fukunaga, Introduction to Statistical Pattern Recognition, Academic Press Theodoridis & Koutroumbas, Pattern Recognition, Academic Press
Summary
compressing an image into a biocode comparing pairs of biocodes with a distance measure d(I1, I2) forming a database of enrollees locating or verifying a candidate against the database with the distance measure
Summary
FARR = equal false acceptance and false rejection ratio Most popular human biometrics
digital fingerprints, with FARR ~ 1E-5 forensic fingerprints (non-digital), FARR < 1E-7 face, with FARR ~ 1E-2 at best iris, with FARR < 1E-7 speaker recognition, with FARR < 1E-2