This document presents a paper on iris recognition using key point-based relative distances as features. It introduces iris image preprocessing such as segmentation, normalization, and Gabor filtering. Features are extracted by detecting key points within filtered image blocks and calculating relative distances between points. The method is tested on two databases for verification and identification, achieving error rates of 0.008% and 0.0015%. While the method has advantages like rotation invariance and compact features, the authors note it may be more suitable for medium security applications than high-security ones.

1. Paper Presentation:
“The relative distance
of key point based
iris recognition” Li Yu
David Zhang
Kuanquan Wang
‘Grandma, do you mind if I
do an iris recognition scan?’
Rueshyna ● Jaiyalas
May 23 2010

2. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

3. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

4. IMAGE PREPROCESSING (1/6)

5. IMAGE PREPROCESSING (2/6)
Longest Chord
(x,y) (xp,yp)

6. IMAGE PREPROCESSING (3/6)
Inner
Boundary
Outer
Boundary

7. IMAGE PREPROCESSING (4/6)

8. IMAGE PREPROCESSING (5/6)
r
1 0
0 π
Φ

9. IMAGE PREPROCESSING (6/6)
Φ
0 π
r 64 pixels
1
256 pixels

10. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

11. GABOR FILTER (1/7)
Original Image Gabor Filter
Result Image

12. GABOR FILTER (2/7)

13. GABOR FILTER (3/7)
For Multi-channels:
θ = 0, 45, 90, 135
T = α = β = 4, 8, 16, 32

14. GABOR FILTER (5/7)
even-symmetric (real part)
φ = 0, 90 odd-symmetric (imaginary part)
θ = 0, 45, 90, 135 Unsupervised Texture Segmentation Using Gabor Filters
Anil K. Jain
Farshid Farrokhnia
T = α = β = 4, 8, 16, 32 Multichannel Texture Analysis Using Localized Spatial Filters
Alan Conrad Bovik
Marianna Clark
Wilson S. Geisler

15. GABOR FILTER (4/7)
φ = 0, 90
θ = 0, 45, 90, 135 Unsupervised Texture Segmentation Using Gabor Filters
Anil K. Jain
Farshid Farrokhnia
T = α = β = 4, 8, 16, 32 Multichannel Texture Analysis Using Localized Spatial Filters
Alan Conrad Bovik
Marianna Clark
Wilson S. Geisler

16. GABOR FILTER (6/7)
φ = 0, 90 2 Parts
θ = 0, 45, 90, 135
16 Channels
T = α = β = 4, 8, 16, 32

17. GABOR FILTER (7/7)
φ = 0, 90 2 Parts
θ = 0, 45, 90, 135 32 Filters
16 Channels
T = α = β = 4, 8, 16, 32

18. FILTER EXAMPLE (1/3)
T = 4 ;θ = 0
T = 8 ; θ = 45
4 Channels
T = 16 ; θ = 90
T = 32 ; θ = 135

19. FILTER EXAMPLE (2/3)
T = 4 ;θ = 0
T = 8 ; θ = 45
4 Even-Symmetric Filters
T = 16 ; θ = 90
with φ = 0
T = 32 ; θ = 135

20. FILTER EXAMPLE (3/3)

21. KEY POINTS (1/4)
256 pixels
64 pixels
32 pixels
32 pixels
To divide filtered image into 16 blocks

22. KEY POINTS (2/4)
256 pixels
64 pixels
→ obtain a key point: by
here, m = 64

23. KEY POINTS (3/4)
for each filtered image
16 key points

24. KEY POINTS (4/4)
32 filtered
images
exist 32 filters
16×32 = 512 key points

25. 32
blo RELATIVE DISTANCE (1/3)
ck
s
32 filtered
images
To obtain the center of key points in the jth blocks by:

26. RELATIVE DISTANCE (2/3)
for some block j:
Oj
KPn
Dj(n)

27. RELATIVE DISTANCE (3/3)
32 filtered
images
There are 32 D in 1st blocks
There are 32 D in 2nd blocks 16 x 32
.
. = 512 Distances
. = 512 Features!!
There are 32 D in 16th blocks

28. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

29. EXPERIMENTS
Database Experiments Modes
CASIA Verification
108×7 = 756 Identification
320×280 pixels
private database
254×4 = 1016
768×568 pixels

30. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

31. NUMBER OF COMPARISONS
(1/2)
CASIA private database
570,780 total 1,031,240 total
2,268 intra-class 1524 intra-class
568,512 inter-class 1,029,716 inter-class

32. NUMBER OF COMPARISONS
(2/2)
Small overlaps
CASIA private database

33. ACCURACY
ROC curve
0.008% 0.0015%
CASIA private database

34. PARAMETER M
control tradeoff
accuracy
speed
small m
lose feature
noise sensitive
large m
many redundant features

35. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

36. IDENTIFICATION
CASIA : 108×(3+4)
private : 254×(3+1)
Testing
Training

37. OUTLINE
The Relative Distance of Key Point Based
Iris Recognition
Iris Image Preprocessing
Features Extracting
Experimental Results
Verification
Identification
Conclusion

38. MISMATCHING
Lower resolution!!
Eyelids obscuring!!
Darkness!!

39. FEATURES (1/5)
Good recognition rate (with a better choice of m)
point number in a block

40. FEATURES (2/5)
Can avoid influence of rotation transform

41. FEATURES (3/5)
Feature dimensions is only 512

42. FEATURES (4/5)
Compared with Daugman’s and Ma’s methods
integrating location and modality info.
more powerful when:
FAR > 0.008% (in CASIA)
FAR > 0.0015% (in private database)

43. FEATURES (5/5)
0.008% 0.0015%
CASIA private database

44. FINALLY
This proposed method is more suitable for
medium security such as those for civil
access control (require a high match rate)
Daugmen and Ma’s methods are more
suitable for high security system such as
military departments (require a low FAR)