You are on page 1of 49

Feature Extraction using F-SIFT

Hunny Mehrotra

24 September 2013

Department of Computer Science and Engineering National Institute of Technology Rourkela


1/32 Hunny Mehrotra F-SIFT

Outline

Introduction State-of-the-art Iris Localization Scale Invariant Feature Transform F-SIFT Conclusions

2/32

Hunny Mehrotra

F-SIFT

Introduction

Iris Biometrics
Iris is externally-visible, owery ring around the pupil The texture pattern is complex, unique and highly stable Provides real-time recognition by mathematical analysis of the random patterns
Eyelids

Pupil Iris

3/32

Hunny Mehrotra

F-SIFT

Introduction

Why Iris?
Highly protected Externally visible Stable patterns Predictable shape

4/32

Hunny Mehrotra

F-SIFT

Introduction

Why Iris?

Table : Comparison of dierent biometric modalities based on their characteristics (H: High (10), M: Medium (6), and L: Low (3)) [1].
Modality Face Fingerprint Hand Geometry Keystrokes Hand Vein Iris Retinal Scan Signature Voice Print Face Thermogram Odor DNA Gait Ear Universality H M M M H H L M H H H M M Uniqueness L H M M H H L L H H H L M Permanence M H M L M H M L L L H H L H Collectability H M H M M M L H M H L L H M Performance L H M L M H H L L M L H L M Acceptability H M M M M L L H H H M L H H Circumvention L H M M H H H L L H L L M H Score 45 58 46 24 46 59 52 35 34 59 45 52 41 54

4/32

Hunny Mehrotra

F-SIFT

Introduction

Automated Iris Recognition System

Iris Database Acquisition Preprocessing Feature Extraction Identification


Figure : Dierent modules of automated iris recognition system.

5/32

Hunny Mehrotra

F-SIFT

Introduction

Automated Iris Recognition System


Input Image Enrollment Single Template Identity Required Preprocessing Feature Extraction

Database

N Templates

Probe Image

Preprocessing

Feature Extraction

Matching

Verification

Accept/Reject

Probe Image

Preprocessing

Feature Extraction

Matching

Identification

Top N Matches

Figure : Dierent modes of operation in a generic biometric system.


5/32 Hunny Mehrotra F-SIFT

Introduction

Performance Measures
Recognition
1 2 3 4 5 Equal Error Rate (EER) Genuine Acceptance Rate (GAR) Receiver Operating Characteristic (ROC) Curve

False Acceptance Rate (FAR) or False Match Rate (FMR) False Rejection Rate (FRR) or False Non Match Rate Stage (FNMR) 1
Data Acquisition
Stage 2 Stage 3

Stage 4

Preprocessing

Feature Extraction

Feature Mat and Decisio

10

Introduct

High Security Applications

False Rejection Rate

Equal Error Rate (Civilian Applications) Low Security Applications

False Acceptance Rate

(a) Error Rates


(b) ROC

Figure 1.6: Hypothetical receiver operating characteristics (ROC) curve

6/32

plot of FAR against FRR at various thresholds. It is a measure to eval Hunny Mehrotra F-SIFT

Introduction

Databases Used

Table : Publicly available iris databases.

Database BATH CASIA Iris Version 3.0

Subjects 50 249

Images 2000 2655

Resolution 320 240 320 280

Units Both Both

7/32

Hunny Mehrotra

F-SIFT

State-of-the-art

Feature Representation

Global Feature Extraction Approaches

Phase-based
Year 1993 2008 2009 First Author Daugman Miyazawa Krichen Approach 2-D Gabor wavelet 2-D Discrete Fourier Transform and POC Gabor phase-correlation Database Iris images from Ophthalmology Associates of Connecticut CASIA ICE ICE Results FR : 1 in 128000, FA: 1 in 151000 EER : 0.18% EER : 0.54% FRR : 2%

8/32

Hunny Mehrotra

F-SIFT

State-of-the-art

Feature Representation

Global Feature Extraction Approaches

Texture-based
Year 1997 2000 2002 2003 2004 2009 First Author Wildes Zhu Ma Ma Tan Sun Approach Laplacian of Gaussian lters Multichannel Gabor ltering Circular symmetric lters Bank of spatial lters Dyadic wavelet transform Multilobe dierential lters Perturbation-enhanced feature correlation lter Database Self captured (25 subjects) Self captured (109 subjects) CASIA CASIA BATH CASIAV3 ICE CASIAV3 CASIA-Iris-Thousand ICE Results Acc : 99.09% CRR : 99.85% CRR : 99.43% CRR : 100% FRR : 0.94% EER : 0.37% EER : 0.40% EER : 0.24%

2012

Zhang

9/32

Hunny Mehrotra

F-SIFT

State-of-the-art

Feature Representation

Global Feature Extraction Approaches

Zero-Crossing
Year 1998 2002 2005 2007 2012 First Author Boles Sanchez-Avila Sanchez-Avila Monro Ahamed Approach Wavelet transform Dyadic wavelet transform Dyadic wavelet transform on an annular region Discrete Cosine Transform Curvelet transform Database Self acquired (20 subjects) Self captured (50 subjects) CASIA BATH CASIA UBIRIS.vl UPOL Results Acc : 97.9% Acc : 99.6% CRR : 100% CRR : 99.3%

10/32

Hunny Mehrotra

F-SIFT

State-of-the-art

Feature Representation

Global Feature Extraction Approaches

Zero-Crossing
Year 1998 2002 2005 2007 2012 First Author Boles Sanchez-Avila Sanchez-Avila Monro Ahamed Approach Wavelet transform Dyadic wavelet transform Dyadic wavelet transform on an annular region Discrete Cosine Transform Curvelet transform Database Self acquired (20 subjects) Self captured (50 subjects) CASIA BATH CASIA UBIRIS.vl UPOL Results Acc : 97.9% Acc : 99.6% CRR : 100% CRR : 99.3%

Global features fail to possess invariance to transformations, illumination, and occlusion.

10/32

Hunny Mehrotra

F-SIFT

State-of-the-art

Feature Representation

Local Feature Extraction Approaches

Keypoint Descriptor
Year 2009 2009 2009 2010 2011 2013 First Author Belcher Mehrotra Mehrotra Du Zhang Sun Approach Region-based SIFT Harris Corner and Entropy SURF Gabor and SIFT Deformable DAISY Matcher SIFT Database ICE WVU BATH CASIAV3 BATH CASIAV3 ICE CASIA-Iris-Thousand CASIAV3 Self acquired (18 subjects) Results EER : 5.57% EER : 8.28% Acc : 87.42% Acc : 92.78% Acc : 95.48% Acc : 95.77% EER : 0.026% EER : 0.21% EER : 0.59% Acc : 98.15%

11/32

Hunny Mehrotra

F-SIFT

State-of-the-art

Feature Representation

Local Feature Extraction Approaches

Keypoint Descriptor
Year 2009 2009 2009 2010 2011 2013 First Author Belcher Mehrotra Mehrotra Du Zhang Sun Approach Region-based SIFT Harris Corner and Entropy SURF Gabor and SIFT Deformable DAISY Matcher SIFT Database ICE WVU BATH CASIAV3 BATH CASIAV3 ICE CASIA-Iris-Thousand CASIAV3 Self acquired (18 subjects) Results EER : 5.57% EER : 8.28% Acc : 87.42% Acc : 92.78% Acc : 95.48% Acc : 95.77% EER : 0.026% EER : 0.21% EER : 0.59% Acc : 98.15%

SIFT wrongly pairs two dierent regions of iris.

11/32

Hunny Mehrotra

F-SIFT

Iris Localization

Iris Localization

Finding the pupil and iris boundary from an eye Traditional approaches uses Circular Hough Transform Proposed approach uses spectrum image based scheme Steps involved are
1

Thresholding
Static Adaptive

2 3 4

Pupil detection Iris detection Removal of eyelids (also referred as noise)

12/32

Hunny Mehrotra

F-SIFT

Iris Localization

Static Thresholding

Figure : Thresholding using static value

13/32

Hunny Mehrotra

F-SIFT

Iris Localization

Static Thresholding

Figure : Thresholding using static value

Value of static threshold needs to be adjusted according to the database

13/32

Hunny Mehrotra

F-SIFT

Iris Localization

Adaptive Thresholding

: 0.10; : 0

= 0.15; : 0

= 0.20; : 1

= 0.25; : 1

= 0.30; : 32

= 0.35; : 23

= 0.40; : 18
14/32

= 0.45; : 23
F-SIFT

= 0.50; : 30

Hunny Mehrotra

Iris Localization

Localization

Pupil Detection

15/32

Hunny Mehrotra

F-SIFT

Iris Localization

Localization

Pupil Detection

Iris Detection

15/32

Hunny Mehrotra

F-SIFT

Iris Localization

Localization

Pupil Detection

Iris Detection

Annular Iris
15/32 Hunny Mehrotra F-SIFT

Iris Localization

Localization

Pupil Detection Cartesian to polar conversion introduces aliasing artifacts. Idea is to directly use annular iris image for feature extraction. Iris Detection

Annular Iris
15/32 Hunny Mehrotra F-SIFT

Scale Invariant Feature Transform

Local Feature Extraction


Global features Global features fail to possess invariance to pose, transformation and occlusion. Could not extract features from annular iris

16/32

Hunny Mehrotra

F-SIFT

Scale Invariant Feature Transform

Local Feature Extraction


Global features Global features fail to possess invariance to pose, transformation and occlusion. Could not extract features from annular iris Scale Invariant Feature Transform

Figure : Steps involved in SIFT

16/32

Hunny Mehrotra

F-SIFT

Scale Invariant Feature Transform

Scale Space Construction


1

Convolve input iris (I ) with variable scale Gaussians to generate LOG images. Dierence of Gaussian (DOG) is obtained from dierence of two nearby scales

17/32

Hunny Mehrotra

F-SIFT

Scale Invariant Feature Transform

Keypoint Detection
3

The maxima or minima of DOG images is taken as the location of candidate keypoint.

18/32

Change in scale
Hunny Mehrotra

F-SIFT

Scale Invariant Feature Transform

Keypoint Orientation
4 5 6

Orientation is assigned to each keypoint to achieve invariance to rotation. For each sample at L(x , y ), magnitude (m) and orientation () is obtained. Histogram of is formed for w w window around the keypoint; weighted by m Peak in the histogram denotes orientation

(a) Detected keypoints


19/32 Hunny Mehrotra

(b) Oriented window


F-SIFT

Scale Invariant Feature Transform

Keypoint Descriptor
8 9 10

16 16 window is taken around each keypoint relative to Total descriptor size is 4 4 8

Orientation histogram is formed for each 4 4 window of 8 bins each

20/32

Hunny Mehrotra

F-SIFT

F-SIFT

SIFT Pairing

21/32

Hunny Mehrotra

F-SIFT

F-SIFT

SIFT Pairing

Drawbacks. . . Uses gradient information which is unsuitable for iris Fails for local pattern deformation

21/32

Hunny Mehrotra

F-SIFT

F-SIFT

F-SIFT

Need features suitable to characterize iris texture

22/32

Hunny Mehrotra

F-SIFT

F-SIFT

F-SIFT

Need features suitable to characterize iris texture Fourier Fourier has been widely used to represent texture Has property of describing periodic function that contains repetitive patterns

22/32

Hunny Mehrotra

F-SIFT

F-SIFT

F-SIFT
F-SIFT Combines advantages of Fourier and Scale Invariant Feature Transform (SIFT) Provides an improvisation over SIFT Contributions 1 Develop novel feature descriptor named F-SIFT
2

Keypoint pairing using phase information The steps involved in F-SIFT are:
1 2 3

Keypoint detection using SIFT Keypoint descriptor using Fourier transform Keypoint pairing using Phase Only Correlation

22/32

Hunny Mehrotra

F-SIFT

F-SIFT

P (n ,n )
ij 11 2

FFT: i th keyblock SIFT: Gallery ith keyblock

P (n ,n ) = 0.9946
ij 1 2

0.8 0.6 0.4 0.2 0 40 40 50

n2

20 0 10 0 10

20

30

n1

Phase-Only Correlation (Pij ) FFT: jth keyblock SIFT: Probe jth keyblock

Figure : Block diagram of proposed F-SIFT

23/32

Hunny Mehrotra

F-SIFT

F-SIFT

Keypoint Descriptor

Keypoint descriptor using Fourier

For each keypoint a window of size W W is taken and descriptor vector is obtained using 1 W2
) (x + W 2 ) (y + W 2

K i (u , v ) =

I (n1 , n2 )e i 2( W
n1 =(x W ) n2 =(y W ) 2 2

n1 u

n2 v W

= A(u , v )e i (u,v )

24/32

Hunny Mehrotra

F-SIFT

F-SIFT

Keypoint Pairing

Keypoint Pairing: Phase-Only Correlation

For two blocks Ki and Kj to be paired


1

Cross phase spectrum is dened by Rij (u , v ) = e i {i (u,v )j (u,v )}

POC function is inverse FFT of Rij 1 Pij (n1 , n2 ) = W2


(x + W 2 ) (y + W 2 )

Rij (u , v )e i 2( W
W u =(x W 2 ) v =(y 2 )

n1 u

n2 v W

25/32

Hunny Mehrotra

F-SIFT

F-SIFT

Keypoint Pairing

1 0.8 Pij(n1,n2) 0.6 0.4 0.2 0 30 20 10 n2 0 0 10 n1 30 20

Pij(n1,n2)=0.9924
Pij(n1,n2)

0.8 0.6 0.4 0.2 0 30 20 10

Pij(n1,n2) = 0.4541

n2

10

15

20

25

n1

Intra region

Inter region

Figure : POC of two regions of iris

26/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

Experimental Results

Table : Recognition performance (in %) for SIFT, SURF, and F-SIFT.


Database Approach SIFT [2] SURF [3] F-SIFT ACC 92.08 98.24 97.13 BATH FAR 7.01 1.44 0.72 FRR 8.82 2.06 5.00 CASIAV3 ACC FAR FRR 88.76 8.80 13.67 96.52 2.16 4.78 97.95 4.09 0.00

Table : GAR at 0.01% FMR using dierent local features.


Approach SIFT SURF F-SIFT BATH 34.83 50.72 43.82 CASIAV3 0.00 0.43 12.23

27/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

BATH Database
100 90
Genuine Acceptance Rate

80 70 60 50 40 30 2 10
1 0 1

SIFT SURF FSIFT 10


False Acceptance Rate

10

10

10

Figure : ROC curves comparing SIFT, SURF, and F-SIFT approaches.

28/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

CASIAV3 Database
100

Genuine Acceptance Rate

80

60

40

20

0 2 10

SIFT SURF FSIFT 10


1

10 False Acceptance Rate

10

10

Figure : ROC curves comparing SIFT, SURF, and F-SIFT approaches.

28/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

SIFT: BATH Database


10
5

Imposter Genuine

10

Frequency

10

10

10

10

10

20

30

40 Scores

50

60

70

80

Figure : Distribution of genuine and imposter scores.

29/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

SURF: BATH Database


10 10 10 Frequency 10 10 10 10
6

Imposter Genuine

20

40

60 Scores

80

100

120

Figure : Distribution of genuine and imposter scores.

29/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

F-SIFT: BATH Database


10
5

Imposter Genuine

10

Frequency

10

10

10

10

10

Scores

15

20

25

Figure : Distribution of genuine and imposter scores.

29/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

SIFT: CASIAV3 Database


10
5

Imposter Genuine

10

Frequency

10

10

10

10

10

20

30

40 Scores

50

60

70

80

Figure : Distribution of genuine and imposter scores.

30/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

SURF: CASIAV3 Database


10
5

Imposter Genuine

10

Frequency

10

10

10

10

50

Scores

100

150

Figure : Distribution of genuine and imposter scores.

30/32

Hunny Mehrotra

F-SIFT

F-SIFT

Performance

F-SIFT: CASIAV3 Database


10
4

Imposter Genuine

10

Frequency

10

10

10

10

20

30

40

Scores

50

60

70

80

90

Figure : Distribution of genuine and imposter scores.

30/32

Hunny Mehrotra

F-SIFT

Conclusions

Conclusions

F-SIFT reduces wrong matches by conventional SIFT approach. Eciently characterizes iris texture. The proposed descriptor, performs comparable to SURF for BATH database. Outperforms other descriptors for CASIAV3 database.

31/32

Hunny Mehrotra

F-SIFT

Conclusions

Bibliography I

A.K. Jain, P. Flynn, and A. A. Ross. Handbook of Biometrics. Springer-Verlag New York, Inc., 2007. H. Mehrotra, B. Majhi, and P. Gupta. Robust iris indexing scheme using geometric hashing of SIFT keypoints. Journal of Network and Computer Applications, 33(3):300313, 2010. H. Mehrotra, G. S. Badrinath, B. Majhi, and P. Gupta. An ecient iris recognition using local feature descriptor. In IEEE International Conference on Image Processing, pages 19571960, 2009. D. G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91110, 2004. H. Bay, A. Ess, T. Tuytelaars, and L. Van Gool. Speeded-up robust features (SURF). Computer Vision and Image Understanding, 110(3):346359, 2008.

32/32

Hunny Mehrotra

F-SIFT

You might also like