Spatial Resolution as an Iris Quality Metric

Transcription

1 Spatial Resolution as an Iris Quality Metric David Ackerman SRI International Sarnoff Biometrics Consortium Conference Tampa, Florida September 8,

2 Iris images with varying spatial resolution high medium low Three separate instances of the same iris captured with different spatial resolution. Does spatial resolution of iris images predict biometric matching performance?

3 Spatial resolution useful Quality Metric? Background: Quality Metrics Spatial Resolution (in an image) Spatial Resolution Response (of an algorithm) Broader agenda: Cost of biometric information Experiment: Spatial resolution as an iris quality metric Question: How and over what range can spatial resolution be used as a quality metric for iris recognition systems? Opinion: A spatial resolution standard for today 3

4 biometric matching performance Iris image quality metric (QM): Definition predictive not predictive QM (relates to image) Quality metrics measure an aspect of an iris image that predicts biometric matching performance. 4

5 Image Quality Metric Hypothesis QM High Highest QM ( Best biometric matching?) Mid Mid Middle QM ( Intermediate biometric matching?) Low Lowest QM ( Worst biometric matching?) Camera resolution Generic iris matching algorithm Image Quality Metric Hypothetical biometric matching results 5

6 normalized contrast Standard iris image resolution specification Ways to specify spatial resolution of an iris image: pixels across iris pixels / mm. MTF ½ spatial frequency (cycles/mm) Modulation Transfer Function (MTF) of entire imaging system including lens + diffraction and image sensor 6

7 normalized contrast normalized intensty Spatial resolution: Definition MTF in image plane PSF in image plane lens and sensor characteristics + diffraction MTF FT x spatial frequency (cycles/mm) distance (m) MTF total system x log MTF opt F ls lens and sensor diffraction (good approx.) x log similar to uncertainty principle (True in image plane and object plane. Note: primes denote image plane) 7

8 Spatial resolution: Example x log Measure MTF / =.5 cycles/mm in object (iris) plane ( / = 4m) blur circle of 77m diameter at halfpeak (FWHM). Note: no primes object plane m To resolve this spatial frequency: x -5 77m Minimum sampling density (Nyquist criterion) requires two pixels per period: m pixel pitch x -5 Conservative over-sampling density (x Nyquist criterion), m pixel pitch x -5 8

9 Information flow in a biometric system QM Iris camera Matching algorithm Information source: Human iris under near infrared light Captures some but not all available information Uses some but not necessarily all of image information Information passes from object to biometric system. Camera captures some of available information to produce iris image. Image quality metrics are imposed on the digital image (camera output). Algorithm uses some of available image information to encode image. Encoded image matched to database templates. 9

10 HD Spatial frequency response of an iris algorithm original azimuthal frequency (periods / ) azimuthal noise added algorithm bandwidth

11 Combining components to make systems QM High Wide Expensive, high performance?? High Narrow Out-of-band image information not used. Low Wide Does not provide algorithm with adequate info. Low Narrow Less expensive, what about performance?? camera resolution algorithm bandwidth

12 What does high performance really mean? Meets biometric requirements (FMR, FNMR) But also High value (meets cost targets while satisfying requirements) Pushes envelope: Distance Speed Size/weight/power Ease-of-use Reduced subject constraints Environmental range

13 blur spot diameter (m) Cost of high spatial resolution Increasing spatial resolution Recall: MTF ½ =.5 cycles/mm 4m period or m sample (pixel) spacing total system MTF (cycles/mm) ½ ( uncertainty curve ) Demanding high resolution (more pixels/iris) increases: size of sensor or number of cameras ($$$) quality of fast lens ($$$) narrow depth of field, necessitating autofocus ($$$) 3

14 Small Exp t: Multi-resolution iris capture d o =.m L orp = 6m pitch = 6.5 px/mm.3m m m m (L orp = object referred pixel size) Lens/sensor MTF (at 85nm): 4 cycles/mm; f = 7mm; F/5.6; pixel pitch 3.75m x L FWHM orp log opt L MTF p ls sensor Actual resolution (x FWHM - ) : lens and sensor MTF F. 5 independent of object distance 9. blur spot/mm lens 4

15 Small Exp t: HD results HD HD matrix: 3 subjects, two irises/subject, 5 distances per iris, trials/distance All authentics matched to all. Even the lowest resolution cases m No false positives. Subject, left (top), right (bottom) 5

16 Medium sized numerical experiment Enrollment database: irises from 33 subjects. 3 instances of each iris. Sample density = px/mm. Correlate reduction in resolution with matching performance numerically reduce resolution match all x all DET 6

17 5 Resolution reduction methodology N x N block average e.g., N = x 48 pristine image down-sampling irreversibly destroys iris information up-sampling does not add back information but, can use any method that produces useful images, e.g., bicubic interpolation, etc. 7

18 5 5 Resolution reduction: Example, N = 5 down-sample 5 up-sample x 48 pristine image x 48 block averaged N = x 48 bicubic interp. up-sampled image detail: px/mm 35 detail: single block of size x L p detail: px/mm 8

19 false reject rate % (FRR) false reject rate % (FRR) DET curves for x and x cases decreasing sample density false accept rate (FAR) false accept rate (FAR) enrollment block avg: N = probe block avg: N = TAR (@ FAR = -6 ) = 98.7% enrollment block avg: N = probe block avg: N = TAR (@ FAR = -6 ) = 9.% 9

20 true accept rate FRR FAR = -6-6 ) TAR (@FAR = -6 ) vs. spatial resolution 98 Results matching pristine enrollment images to block averaged reduced resolution probe images: For sampling densities greater than 5/mm, there appears to be no perceptible variation in biometric performance sampling density (mm - ) But, the standard is intended to be symmetric between enrollments / probes

21 true accept rate FRR FAR = -6-6 ) Symmetric enrollment and probe standard 98 Results matching N-block averaged enrollment images to N-block averaged probe images (dark blue): (information cartoon) sampling density (mm - ) Results are almost indistinguishable from pristine vs. block averaged case. Again,for sampling densities greater than 5/mm, there appears to be no perceptible variation in biometric performance. But, so far, have only looked at a single iris matching algorithm

22 true accept rate FRR FAR = -6-6 ) Different algorithms, same behavior Results matching N-block averaged enrollment images to N-block averaged probe images using two different commercial algorithms (green and red): 94 Results are almost indistinguishable from previous cases sampling density (mm - ) Again,for sampling densities greater than 5/mm (and maybe 4/mm), there appears to be no perceptible variation in biometric performance as a function of algorithm. Will very large databases resolve dependence between 99% and %?

23 Iris information band: Schematic view high resolution information low resolution information iris matching algorithm filter bandwidth Iris matching algorithms that only use modest spatial frequency (in-band) information would explain the apparent lack of correlation between spatial resolution beyond sample densities of ~ 5/mm. 3

24 Remarks and conclusions Spatial resolution is more than pixel count and relates to optics and sampling. Numerical experiments suggest that spatial resolution does not correlate to biometric performance beyond sampling densities of ~ 5/mm and thus is only a good quality metric for 5/mm and below. (For modest database size so far.) This result holds up for asymmetric and symmetric enrollment and probe resolution. This result holds up for 4 different commercial algorithms. Need tests on large databases with more algorithms. If you still want high spatial frequency information just in case, be prepared to pay for it in terms of $$$ and system complexity. Opinion: For an in-focus image, need > 5 pixels / mm. An iris spatial resolution standard for today might use 7 pixels/mm or 8 pixels/iris diameter. 4

25 Acknowledgments to: Jim Bergen Ann-Marie Lanzillotto Ansley Jessup Ray Kolczynski Oleg Naroditsky Thank you for your attention. Headquarters: Silicon Valley SRI International 333 Ravenswood Avenue Menlo Park, CA Washington, D.C. SRI International Wilson Blvd., Suite 8 Arlington, VA Princeton, New Jersey SRI International Sarnoff Washington Road Princeton, NJ Additional U.S. and international locations