Title Goes Here Algorithms for Biometric Authentication

Transcription

1 Title Goes Here Algorithms for Biometric Authentication February 2003 Vijayakumar Bhagavatula 1

2 Outline Motivation Challenges Technology: Correlation filters Example results Summary 2

3 Motivation Recognizing the identity of a person can improve security of access to physical and virtual spaces Continuous recognition prevents unauthorized access when a legitimate user forgets to log off. Most current methods rely on passwords, ID cards that can be easily forgotten or stolen Vision: identity recognition based on Biometrics (e.g., Fingerprints, face, voice, iris, etc.) Intelligent fusion of information from multiple biometrics Face Fingerprint Voiceprint Iris 3

4 Challenges in Face Recognition Pose Illumination Expression Occlusion Time lapse Individual factors: Gender 4

5 Facial Recognition Vendor Test Media Expr. Resol. Illum. Temporal Pose Distance 5

6 c25 c09 c31 c22 c02 c37 c05 c27 c29 c11 c14 c34 c07 6

7 Recognition Performance Commercial Product Probe Gallery Source: S. Baker 7

8 Correlation Filters Test Image FFT IFFT Analyze Decision Training Correlation Filter Filter Design Correlation output Training Images Recognition For Authentic For Impostor 8

9 Train on 3, 7, 16, -> > Test on 10. 9

10 Using same Filter trained before, Perform cross-correlation on cropped-face shown on left 10

11 Using same Filter trained before, Perform cross-correlation on cropped-face shown on left. 11

12 CORRELATION FILTERS ARE SHIFT-INVARIANT Correlation output is shifted down by the same amount of the shifted face image, PSR remains SAME! 12

13 Using SOMEONE ELSE S Filter,. Perform cross-correlation on cropped-face shown on left. As expected very low PSR. 13

14 Peak to Side Lobe Ratio 1. Locate peak 2. Mask 5x5 pixel region PSR Peak mean 3. Compute the mean and in a 20x20 region centered at the peak 14

15 Features of Correlation Filters Shift-invariant; no need for centering the test image Graceful degradation Can handle multiple appearances of the reference image in the test image Closed-form solutions, no need for iterations Methods to provide tolerance to expected distortions Common algorithms for many biometrics 15

16 13 cameras 21 Flashes The 3D Room (v. 2.0) 16

17 PIE Database Illumination Variations Simulations using 65 people from the Pose, Illumination and Expression (PIE) Database from the Robotics Institute at CMU. Each person has 21 face images with different illuminations. We show results of using the CF filter and the individual eigen-space method (IESM) on this database. 17

18 49 Faces from PIE Database illustrating the variations in illumination 18

19 Training Image selection We used three face images to synthesize a correlation filter and an individual eigenspace to perform verification and provide a comparison of the performance of the two methods. The three selected training images consisted of 3 extreme cases (dark left half face, normal face illumination, dark right half face). n = 3 n = 7 n = 16 19

20 EER using IESM Equal Error Rate using Individual Eigenface Subspace Method on PIE Database with No Background Illumination Equal Error Rate Average Equal Error Rate = 30.8 % Person 20

21 EER using Correlation Filters Authenticate Threshold Reject 21

22 Fingerprint Verification with Correlation Filters NIST Database 24 - Digital Video of Live-scan Fingerprint Data Optical sensor 500 dpi resolution Image Size : 720x Classes 300 images per finger 1 filter per class using 15 training images Tested on all 300 images from all 10 classes Equal error rate of 0.15% 22

23 Iris Verification Correlation filters applied to the 9 iris images yield zero verification errors Challenge is to acquire good-quality iris images Source: National Geographic Magazine Source: Dr. J. Daugman s web site 23

24 Speaker Verification Front-end processing Speaker model Impostor model + - Λ, Accept Λ, Reject 24

25 Multi-Biometric Authentication Improved authentication performance obtained via intelligent management of information from multiple biometric sensors Individual authentication modules must provide more information than just the verification decision (e.g., confidence levels about verification decisions) Biometric 1 h 1 Biometric 2 Biometric 3 h 2 h 3 Joint Decision Biometric N h N 25

26 Summary Correlation filters attractive for biometric verification Shift-invariant Allow design of distortion tolerance Closed-form solutions Graceful degradation No verification errors on PIE illuminations database with 1 correlation filter trained from 3 face images About 0.5% EER on NIST Fingerprint Database 24 No verification errors among 9 iris images Started a speaker verification effort 26

27 Research Plan Extend face recognition work to handle Expression changes Pose variations Face recognition from video Improved fingerprint verification algorithms Iris image database Develop and evaluate iris verification algorithms Develop improved speaker verification methods Demonstrate the advantages of multi-modal biometrics 27