Iris Recognition Part I: Patrick Gampp Part II: Andrea Sereinig - PowerPoint PPT Presentation

Advanced Signal Processing Seminar Iris Recognition Part I: Patrick Gampp Part II: Andrea Sereinig Advanced Signal Processing Seminar Graz, 12/05/2007 Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 1/22

Advanced Signal Processing Seminar Content • Part I: – Basics – The Daugman Iris Recognition System • Part II: – Wildes Iris Recognition – Iris on the Move – Iris Recognition Systems Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 2/22

Advanced Signal Processing Seminar Part I: Content • Motivation • Anatomy of the Human Iris • Image Aquisition • Iris Localization • Wavelet Transformation • Iris Feature Encoding by 2D Wavelets Demodulation • Pattern Matching • Recognizing Irises Regardless of Size, Position and Orientation • Decision Environment • Countermeasures Against Subterfuge • Performance, Execution Speed Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 3/22

Advanced Signal Processing Seminar Motivation • Noninvasive • Covert evaluation possible • Iris patterns variability is enormous • Iris is well protected from environment • Iris pattern is stable over time • Insensitive to angle of illumination • Insensitive to changes in viewing Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 4/22

Advanced Signal Processing Seminar Anatomy of the Human Iris (1) [Wildes 1997] [Wildes 1997] Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 5/22

Advanced Signal Processing Seminar Anatomy of the Human Iris (2) • Iris patterns largely complete by eighth month • Pigmentation accretion into adolescence • Average pupil size increases until adolescence • Advanced age: slight depigmentation and shrinking of pupillary opening • Patterns are stable with age • General structure genetically determined, but minutiae dependend from initial conditions • NIR wavelengths reveal deeper stromal features Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 6/22

Advanced Signal Processing Seminar Image Aquisition (1) • Need sufficient resolution, sharpness, contrast • Iris is relatively small • Iris is dark object • Human operators are sensitive about their eyes • Well-centered image while remaining noninvasive • Eliminate artifacts like reflectations, aberrations Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 7/22

Advanced Signal Processing Seminar Image Aquisition (2) • LED point light source • NIR illumination 700nm- 900nm • Sensitive CCD camera • 100- 200 pixels monochromatic image • Distance 15- 46cm • Video rate capture • Self- positioning of object • Focus assessment: maximizing high-frequency power of 2D Fourier spectrum Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 8/22

Advanced Signal Processing Seminar Iris Localization 1. Cirular path of contour integration for pupil edge 2. Circular path for limbic boundaries 3. Arcuate path with fitted splines for eyelid boundaries ∂ I ( x , y ) ∫ ∗ max G ( r ) ds ∂ ( r , x , y ) σ r 2 π r 0 0 ( r , x , y ) 0 0 ⎛ ⎞ 2 − ( r r ) ⎜ ⎟ 0 − 1 ⎜ ⎟ 2 = 2 σ G ( r ) e ⎝ ⎠ σ σ 2 π I ( x , y )....., Image [Daugman Website] r ....., Radius ( x , y )....., Center Coordinate s 0 0 G ( r )....., Gaussian smoothing function σ Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 9/22

Advanced Signal Processing Seminar Why (Gabor) Wavelets? • Advantage over Fourier Transform representing functions with discontinuities and sharp peaks • Compact representation of images, eg. jpeg2000 [Ulrich Günther 2001] • Simple cells in visual cortex can be modelled • „Quantum principle“ in information � good time- vs. frequency- resolution trade-off Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 10/22

Advanced Signal Processing Seminar Complex 2D Gabor Wavelets • Mother function: ⎡ ⎤ − 2 − 2 ( x x ) ( y y ) − 0 + 0 π ⎢ ⎥ [ ] 2 2 − − + − ⎢ α β ⎥ = 2 π i u ( x x ) v ( y y ) ⎣ ⎦ Ψ ( x , y ) e e 0 0 0 0 • Generating function: − = 2 m Ψ ( x , y ) 2 Ψ ( x ' , y ' ) mpq θ [ ] − = m + − x ' 2 x cos( θ ) y sin( θ ) p [ ] q = − m − + − y ' 2 x sin( θ ) y sin( θ ) x , y ....., Location of peak of gaussian envelope 0 0 x, y....., Visual space variables α , β ..... , Width and Length of envelope u , v ....., Spatial frequency variables of sinusoidal carrier 0 0 [Daugman 1988] θ ....., Discrete rotation of envelope m....., Dilation parameter p, q....., Translatio n in position; shift parameters Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 11/22

Advanced Signal Processing Seminar Iris Pattern Encoding (1) ⎧ ⎧ ⎫ 2 − θ − φ 2 ( ) − − ρ ( ) ⎪ r ⎪ 0 0 ⎪ ∫∫ 2 ρ φ − ω θ − φ β ρ ρ φ ≥ ( ) 2 1 , Re ( , ) α 0 ⎨ i ⎬ if I e 0 e e d d ⎪ ⎪ ⎪ ⎪ ⎩ ⎭ ρ φ = ⎨ h Re ⎧ ⎫ 2 − θ − φ − − ρ 2 ( ) ⎪ ( ) r ⎪ ⎪ 0 0 ∫∫ 2 ρ φ − ω θ − φ β ρ ρ φ ( ) 2 0 , Re ( , ) α 0 ⎨ i ⎬ p ⎪ 0 if I e e e d d ⎪ ⎪ ⎪ ⎩ ⎭ ⎩ ρ φ ⎧ ⎧ ⎫ 2 − θ − φ 2 ( ) − − ρ ( ) ⎪ r ⎪ 0 0 ⎪ ∫∫ ρ φ − ω θ − φ β 2 ρ ρ φ ≥ ( ) 2 1 , Im ( , ) α 0 ⎨ i ⎬ if I e 0 e e d d ⎪ ⎪ ⎪ ⎪ ⎩ ⎭ ρ φ = ⎨ h Im ⎧ ⎫ 2 − θ − φ − − ρ 2 ( ) ⎪ ( ) ⎪ r ⎪ 0 0 ∫∫ 2 ρ φ − ω θ − φ β ρ ρ φ ( ) 2 0 , Im ( , ) α 0 ⎨ i ⎬ p ⎪ 0 if I e e e d d ⎪ ⎪ ⎪ ⎩ ⎭ ⎩ ρ φ [Daugman Website] I( ρ , φ )....., Raw iris image in polar coordinate system α , β ....., Multi - scale wavelet size parameters ; 8 - fold range from 0.15mm to 1.2mm ω ....., Wavelet frequency; spanning 3 octaves inverse proportion to β r , θ ....., Iris region 0 0 Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 12/22

Advanced Signal Processing Seminar Iris Pattern Encoding (2) • Cyclic phase- quadrant code: Gray code • Coarse phase quadrant quantization: � Ignore imaging contrast, illumination, camera gain • Many wavelet sizes, frequencies, orientations � 256 Byte per iris • Masking bits: ignore obscured data Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 13/22

Advanced Signal Processing Seminar Pattern Matching • Test of statistical independence • Fractional HD (Hamming Distance) ⊗ ( code A code B ) mask A mask B I I = HD mask A mask B I • Obscured data: set mask bit ‚0‘ • XOR, AND in parallel wordlength- size chunks � single machine instruction • 300MHz CPU: 100 000 iris comparisons per second Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 14/22

Advanced Signal Processing Seminar Distribution of Hamming Distances • 9 million comparisons of different irises • 4258 iris images acquired in UK, USA, Japan, Korea • Bernoulli trial, but correlation between „coin tosses“ • Only small subsets of bits are mutually independent • Same distribution for genetically identical irises ⎛ ⎞ N ⎜ ⎟ − = − m ( N m ) f ( m ) p ( 1 p ) ⎜ ⎟ m ⎝ ⎠ − p ( 1 p ) = σ 2 (for correlated Bernoulli trials) N ⇒ = N 249 Degrees of freedom [Daugman 2004] Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 15/22

Advanced Signal Processing Seminar Homogeneous Rubber Sheet Model • Problem: Pupil size change • Solution: ( ) → I x ( r , θ ), y ( r , θ ) I ( r , θ ) [ ] ∈ r 0 , 1 [ ] ∈ θ 0 , 2 π = − + x ( r , θ ) ( 1 r ) x ( θ ) rx ( θ ) p s = − + y ( r , θ ) ( 1 r ) y ( θ ) ry ( θ ) p s ( ) x ( θ ), y ( θ ) ....., Limbus boundary points s s ( ) x ( θ ), y ( θ ) ....., Pupillary boundary points p p Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 16/22

Advanced Signal Processing Seminar Cyclic Scrolling • Problem: Iris orientation depends upon head tilt etc • Solution: Compare iris code at many orientations • Preserve only best match [ ] n = − − F ( x ) 1 1 F (x) n 0 d [ ] − n 1 = = − f ( x ) F ( x ) nf ( x ) 1 F (x) n n 0 0 dx f .......... , PDF for comparison in 1 orientatio n 0 F ( x )..... , CDF 0 x..... , HD Acceptance Criterion x = ∫ F ( x ) f ( x ) dx . .... , Probabilit y of getting a false match 0 0 0 − 1 F ( x ) ..... , Probabilit y of not getting false match 0 [ ] n − 1 F ( x ) ..... , n independen t tests at n relative orientatio ns 0 [Daugman 2004] Patrick Gampp Iris Recognition Professor Horst Cerjak, 19.12.2005 17/22