Tools for Iris Recognition Engines Martin George CEO Smart Sensors Limited (UK)
About Smart Sensors Limited Owns and develops Intellectual Property for image recognition, identification and analytics applications Primarily applied (to date) in the form of software development tools and know-how for iris biometrics (algorithms and recognition engines) We work with camera manufacturers and systems integrators, aiming to be agnostic to image source Small footprint methods give rapid performance and ability to scale to many different processor platforms (PC, PDA, Linux, DSP) MIRLIN = Monro Iris Recognition Library and INterface, based on work of Prof. Don Monro (University of Bath, UK) Very competitive performance Changing the iris biometrics business model to one that makes economic sense for developers and end-users Implementations and licence terms that open up volume markets 2
Key benefits of Iris Recognition Fundamentally simple A digital photo of the eye, using night-vision illumination (Near Infra-Red = NIR) Exceptional discrimination power Excellent for IDENTIFICATION applications De-duplication in large scale enrolment programs Non-contact, hands-free usage Non-intrusive, Convenient, Versatile Good where fingerprints are not good Low and high humidity, manual workers, elderly, 3
Myth-busting You will already be well aware of these issues, but in our experience public users and some deployers still have misperceptions of iris biometrics: Iris Scanning Doesn t it use lasers? Intrusively images the back of the eye The iris changes if you re ill Identical twins have similar eyes No scanning is involved. A simple digital photo of the eye surface is taken, using a low level of near-infra red light. No. The illuminators are Light Emitting Diodes (LEDs), and are just below the visible spectrum very similar to night vision cameras. That is for retinal recognition. Iris recognition should not be confused in any way with retinal recognition. Iris recognition biometrics looks at the fibrous tissue structure of the iris, which is fully stable a few months after birth. Iridology (the alternative health practice of recognising illness from the iris) has not been proven in any medical studies, and moreover relates to colour. The fibrous tissue of the iris develops in a completely random way, and is not genetically dependent. Identical twins have different irises (as indeed are left and right eyes). 4
Typical Iris Recognition system process Localization Normalization * Image Enhanced User passing through PASS or Divert Decision Match Watch list Extract Eye Image, Pupil/Iris segmentation Classifier Enrolled Database Feature Extraction Feature Vector (Template) Read enrolled template or interoperable image (e.g. IREX Kind 7) from Card/Credential * Please note these intermediate images are a proprietary format, and not derived from IREX Kind 16 or 48 * 5
MIRLIN - Basic Features Proprietary transform, Rapid Iris/Pupil finder, Rotation compensation, Specularity masking/scrubbing Quality metrics enable rapid image assessment before computing resource is spent on image segmentation Optimised for images acquired with NIR light sources: 720 900nm, >0.5 mw/cm 2 irradiance Iris image resolution conforming to ISO/IEC 19794-6 Minimum (in current image standard): 100 pixels Recommended for optimum performance: 200 pixels+ Fast template matching based on Hamming Distance Uses simple XOR logic Typically 500,000 matches per second with common PC server configurations. Much more with hardware accelerator 6
Target applications Physical access control (PAC): Non-contact Building Access and Security Policy Management System Time and Attendance Logical Access Control Security of access for workstations Links easily with PAC and Building Management systems Ideal for Access Hierarchy and Policy setting systems that are based on presence of particular individual(s) Border control and Frequent Traveller Programs Use with large databases and watch lists, no contact needed Military, Security, Emergency and Law Enforcement forces Mobile units with credential readers Force protection, Base access, ID on the move 7
Liveness Detection and Spoof Counter-measures Video frame rate image processing and template creation enables real-time liveness checks via pupil and template dynamics Frame-rate measurements of pupil position, size and pupil/iris ratio Synchronize illumination changes with detection of pupil behaviour Blink detection Printed pattern contact lenses if not transparent to NIR detected via anomalies in dynamic template creation Dilation drops alter iris texture behaviour as well as causing abnormal iris/pupil size ratio 8
MIRLIN Iris Analyst Import and Review iris image databases; image normalisation Create ROC and DET charts with data output to.csv Analysis of six image metrics: Average grey level Contrast Saturation Signal-noise ratio Sharpness assessment Occlusion Binning of performance data according to image metrics scores 9
New techniques A second set of data from the same iris image Different error profile compared to spatial frequency domain techniques (almost all others) Offers dispute resolution potential on very large databases or identification applications Works on segmented images (i.e. requires iris texture only) Non-match Match 10
New techniques Fast database matching For rapid matching within very large databases Exploits proprietary template indexing techniques 2,500 class look-ups/sec, on ANY database size Example: 4TB RAM holds ~50M indexed ID records, and can support 2,500 ID queries/second run by a single PC Can be used with other binary templates (e.g. fingerprints) Based on Smart Sensors US patent applications: Fast Database Matching 23 October 2006, ASN 11/585,365. Publication no. 20080097992 Fuzzy Database Matching 23 October 2006, ASN 11/585,358. Publication no. 20080097983 11
Partner examples: Stand-off Iris Capture Sarnoff Iris On the Move AOptix InSight : at 1.5-3m distance! MIRLIN compatible and demonstrated with these systems: the default option with InSight Requirements of Stand-Off Iris Capture Require minimal user co-operation Handle 20 people per minute throughput Cope with glasses, contact lenses, etc. Optics and photon budget MUST be right! Opportunity to integrate facial recognition Iris feature processing near frame rate Identification mode -no contact needed Adaptive Optics puts the icing on the cake 12
Partner examples: Datastrip Mobile ID Runs on Windows CE Typical 0.9s iris enrolment Auto iris capture New DSV2+ Turbo Multimodal mobile reader incorporates MIRLIN Iris Recognition Engine with TRIAD application software On-board capacity for 4000 data records, with local 4000:1 ID Match (all three biometrics) Wireless communication to server for unlimited external template comparison capacity 13
Partner examples: Vista Imaging USB 2.0 interface Provides video image stream Automatic iris capture mode Standard camera tripod mount Cold mirror enables easy self-acquisition of images Ultrasonic auto-range sensor MT2: Multi-modal with iris camera, face camera and optical fingerprint sensors FA2: Combination Iris and Face camera (separate sensors) Full audio interface (mic+spkr) MT2: Multi-modal biometric acquisition unit for handheld or mounted use by operator, or by user FA2: Compact iris/face acquisition module for door-mounted, tripod, or handheld operation (actual size 93 x 93 x 50mm) 14
Smart Sensors Capability Summary Iris/Pupil finder very rapid location of iris and pupil coordinate output ideal for camera developers MIRLIN SDK - fully featured software toolkit for iris recognition engines and back-end ID server systems excellent cross-platform support Bath Iris Image Database up to 800 people / 1600 eyes / 32,000 images ideal resource for test and evaluation MIRLIN Iris Analyst integrated suite of tools that generates ROC/DET curves and bins images according to quality metrics vs. performance 15
Contact Details Further information available from: Smart Sensors Limited Carpenter House Innovation Centre BATH, BA1 1UD United Kingdom Tel: +44 (0) 1225 388690 Martin George CEO mgeorge@smartsensors.co.uk 16