Part 1/4

minorbigarmSecurity

Nov 30, 2013 (3 years and 8 months ago)

64 views

PhD Thesis

Biometrics


Science studying measurements and statistics
of biological data


Most relevant application: id. recognition

2

Why Facial Biometrics ?


Most intuitive way of identification


Socially and culturally accepted worldwide


It may work without collaboration

2006

43.6 %

19.2 %

2001

3

Facial Biometrics


Challenges ahead


Less accurate than iris and fingerprint


Problems with uncontrolled
environments (illumination, viewpoint…)

Best system

Average

Fully automatic

4

Active Shape Models


Automatic training from
examples


User
-
defined template based
on landmarks


Model
-
based parametrization


Generative models

5

T.F.
Cootes
, C. J, Taylor,
D.H. Cooper, J. Graham (1995)

Computer Vision and Image Understanding, 61(1):38

59

This thesis…


Focus on 3 contributions to ASMs
on relevant aspects for facial feature
localization:


More accurate segmentation
invariant to in
-
plane rotations


Add robustness to out
-
of
-
plane
rotations


Estimate the Reliability of the
segmentation

1

2

3

4

6

ASM: Construction of the model


Face outlines based on landmarks


Shape statistics to learn spatial relations


Texture statistics for image search


Landmarked Training
Set

Local texture
statistics

Shape
statistics

PDM

IIMs

1

7

Point Distribution Model

1.
-

The input shapes are aligned to
remove scale, translation and
rotation effects.




1 1 2 2
,,,,...,1
T
L L
i i i i i i i
x y x y x y i N
  
u
i
v
i
u
i i i i
s
 
u R v t
Image Coordinates

Model Coordinates

1

1

8

Point Distribution Model

2.
-

Principal Component Analysis (PCA) on the aligned
shapes







i
i
j
j
i
i
T
i
j
N
i
T
i
i
N
i
i
diag
S
N
S
N
Φb
u
u
b
u
u
Φ
b
Λ
Λ
Φ
Φ
u
u
u
u
u
u



















)
(
1
1
1
1
1














(2
L)
-
space
representation

PCA
-
space
representation

1

1

9

Point Distribution Model (PDM)

i
i
j
j
i
Φb
u
u
b








Can determine
valid shapes



Can get closest
valid shape



Introduces a
representation
error

1

1

10

Point Distribution Model (PDM)

i
i
j
j
i
Φb
u
u
b






More specific

More general

1

1

11

PDM: Modes of variation

Variation from
1st Principal
Component

1

1

12

PDM: Modes of variation

Variation from
2nd Principal
Component

1

1

13

ASM: Local Texture Statistics (1)


First order derivatives of the pixel intensity


For each landmark


Sampled perpendicularly to the contour


1
:




k
k
j
j
i
g
Normalized
i
-
th

landmark

0
2
4
6
8
10
12
14
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
1

1

14

ASM: Local Texture Statistics (2)


Second order statistics for each landmark















N
s
i
s
i
T
i
s
i
i
N
s
s
i
i
N
N
1
,
,
1
,
1
1
1
g
g
g
g
G
g
g
i
-
th
landmark

0
2
4
6
8
10
12
14
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
0.3
0
2
4
6
8
10
12
14
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
1

1

15

ASM: Model Matching

1.
The average shape is placed on the image, roughly
matching the face position



1

1

16

2.
Displacement of each landmark to minimize
the
Mahalanobis

distance to the mean
profil







3.
Apply shape model restrictions



ASM: Model Matching

Steps 2 and 3 are repeated a fixed number of iterations

at different resolutions, increasing detail


1

1

17

ASM: Model Matching

1

1

18

t
t
j
j
t
th
t
T
t
iteration
t
Φb
u
u
b
u
u
Φ
b







1
),
(


ASM: Model Matching

1

1

19

ASM: Complex textures



Several factors modify facial appearance


beard, hair cut, glasses, teeth.


The distribution of the normalized gradient is
often non Gaussian nor
unimodal
.

1

1

20

ASM: Complex textures

1

1

21

Optimal Features ASM



Texture description based on Taylor series


Grids centered at the landmarks for local analysis


Non linear classifier (kNN) for inside
-
outside
labeling









0
0
0
0
)
(
)
(
!
)
(
)
(
x
x
x
x
n
x
f
x
f
n
n
inside

outside

1

2

22

B. van
Ginneken
, A.F. Frangi, J.J.
Staal
, B.M.
ter

Haar

Romeny
, and M.A.
Viergever

(2002)

IEEE Transactions on Medical Imaging, 21(8):924

933