1. The Nose on Your Face May
Not be so Plain: Using the
Nose as a Biometric
Adrian Moorhouse and Adrian Evans
Department of Electronic & Electrical Engineering, University of Bath
Gary Atkinson, Jiuai Sun and Melvyn Smith
Machine Vision Laboratory, University of the West of England
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
2. Overview
• Introduction
• Photometric stereo image acquisition
• Nose feature extraction
– Nose region segmentation
– Curvature-based landmark extraction
– Geometric ratio and ridge features
• Evaluation of classification performance
• Discussion and conclusions
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
3. Introduction
• A number of face-based biometrics have been
proposed e.g. iris, ear and retina
B. Griaule, “Understanding Biometrics”, Online, 2008.
• The nose is hard to conceal and relatively
invariant to expression
• Provides fixation points for face recognition
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
4. Introduction
• Noses can be divided into 6 types
Nasion
• Full 3D nose matching is computationally
expensive
• Shape of the ridge important
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
5. Photometric stereo image
capture
• Capture made using PhotoFace system (UWE)
– 4 flashguns and 1 200 fps camera
– 4 fames captured in ~20 ms
Input
images
Bump
map
Surface normals
• Albedo image unaffected by lighting
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
6. Nose feature extraction
• Nose region segmentation
– Multistage classifier for face recognition
– Skin colour used to remove non-face skin pixels
– Nose tip is the closest object to camera
– Nose region proportional to output of face recognition
stage
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
7. Nose feature extraction
• Curvature-based landmark detection
– Applied to the surface normals image
– Robust mechanism for identifying nasal landmarks
– Principle curvatures κmin and κmax found via mean
(H) and Gaussian (K) curvatures:
κ min = H − H 2 − K κ max = H + H 2 − K
Surface shape classes
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
8. Nose feature extraction
• Curvature-based landmark detection
– Binary convex and concave images filtered using
opening by reconstruction
Before After
– Nose tip is largest convex region
– Nasion is largest concave region above tip
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
9. Nose feature extraction
• Geometric nasal proportions
– 2 ratios defined:
Saddle width
Saddle ratio =
Ridge length
Nose tip width
Nose tip ratio =
Ridge length
– Combined in a 2 element feature vector
– Width at centroid
more robust
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
10. Nose feature extraction
• Nose ridge profile
– Defined between nasion and tip
• Robust to variations in pose
– 3D shape captured by extracting
ridge points from range image
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
11. Nose feature extraction
• Nose ridge profile represented using Fourier
descriptors
• Ridge reflected to make closed contour
• Coefficients adjusted to make invariant to scale
and rotation
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
12. Evaluation of classification
performance
• Photoface database used
– 36 single captures
– 4 multiple captures (data-sets A,B,C and D)
Dataset C
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
13. Evaluation of classification
performance
• Training set of 44 images
– 36 single images
– 2 neutral images from data-sets A, B, C and D
• Test set, remaining n-2 images from data-
sets A, B, C and D
• Euclidean distance used to find closest
match in training set for each test image
• Random change of correct recognition is
2/44 = 4.54%
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
14. Evaluation of classification
performance
• Geometric ratios results
– Poor rank 1 performance
– Rank 10 performance better
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
15. Evaluation of classification
performance
• Geometric ratios results
Dendrogram of distances between data-sets A (features 1-12), B (features 13-22), C
(features 23-28) and D (features 29-34) for the geometric ratios.
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
16. Evaluation of classification
performance
• Geometric ratios results
Input image from data-set D and closest matches from training set
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
17. Evaluation of classification
performance
• Nose ridge profile
– Rank 1 performance much improved
– Rank 10 performance slightly worse
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
18. Evaluation of classification
performance
• Nose ridge profile
Dendrogram showing the distances between data-sets C (features 1-6) and D
(features 11-12), for ridge FD features.
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
19. Evaluation of classification
performance
• Nose ridge profile
Input image from data-set A and closest matches from training set
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
20. Evaluation of classification
performance
• Cumulative Match Characteristic plot for
combined features
1 The Eigennose
0.9 technique applying
0.8 the Eigenface
method to the nose
Probability of recognition
0.7
0.6
region of the face
0.5
0.4
0.3
Eigennose
0.2
Geometric Ratios
0.1 Ridge FD
Combined Geometric Ratios and Ridge FD
0
1 3 5 7 9 11 13 15 17 19 21 23
Rank
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
21. Evaluation of classification
performance
• Cumulative Match Characteristic plot for
combined features The combined
1
Geometric Ratios
0.9
(GR) and Ridge FD
0.8
technique uses the
Probability of recognition
0.7
GR to select the 12
0.6
closest faces and
0.5 applies the ridge FD
0.4 to this subset
0.3
Eigennose
0.2
Geometric Ratios
0.1 Ridge FD
Combined Geometric Ratios and Ridge FD
0
1 3 5 7 9 11 13 15 17 19 21 23
Rank
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009
22. Discussion and conclusions
• The nose’s biometric potential is largely
unexplored
• Curvature provides robust method for identifying
landmarks in PS images
• Geometric ratios and nose ridge shape both
show the nose’s biometric potential
• Recognition currently far lower than other
biometrics
• Evaluation over larger database and in
conjunction with other recognition techniques
ongoing
IET 3rd International Conference on Imaging for Crime Detection and
Prevention (ICDP-09), Kingston, UK, 3rd Dec 2009