Partial Faces for Face Recognition

gaybayberryΤεχνίτη Νοημοσύνη και Ρομποτική

17 Νοε 2013 (πριν από 3 χρόνια και 6 μήνες)

60 εμφανίσεις

Page 1
Srinivas Gutta*, Vasanth Philomin*,
Miroslav Trajkovic*, Harry Wechsler**
*Philips Research –USA
Video and Display Processing Department
345 Scarborough Rd., Briarcliff Manor, NY
{Srinivas.Gutta, Vasanth.Philomin,Miroslav.Trajkovic}@philips.com
** George Mason University
Dept. of Computer Science
4400 University Dr., Fairfax, VA 22030
wechsler@cs.gmu.edu
Partial Faces for Face Recognition
Page 2
￿
Objective
￿
Problem Definition
￿
Difficulties
￿
Recognition Architecture
￿
Database Acquisition
￿
Experiments
￿
Discussion
Page 3
￿
To investigate if partial faces could be
used for recognition
￿
To establish how much of a partial face
is sufficient for recognition
Page 4
￿
MATCH / IDENTIFICATION-An image of an unknown
individual is collected (“ probe”) and the identity is
found searching a large set of images (gallery).
Matching becomes especially difficult when the probe
is a duplicate rather than same (counterpart) image
from the gallery
￿
SURVEILLANCE / VERIFICATION-The systems
checks if a given probe belongs to a relatively small
gallery labeled as a set of intruders. The system is
usually flooded with thousands of faces (e.g. airport
security)
Page 5
￿
Geometry of image formation
￿
Photometry -illumination
￿
Individual characteristics -gender, race
￿
Changes in facial expression and aging
￿
Scale up and robustness
Page 6
Face
Images
Face
Detection
Face
Normalization
RBF
Networks
Training
Testing
Learned
Model
Acceptance/
Rejection
Page 7
Srinivas Gutta and HarryWechsler(1997), Face
Recognition Using Hybrid Classifiers, Pattern
Recognition, Vol. 30, No. 4
Page 8
SELECT MAXIMUM
LINEAR WEIGHTS
UNIT WEIGHTS
OUTPUT NODES (j)
BASIS FUNCTION
NODES (i)
INPUT NODES (k)
G
G
G
G
Basis Function:
y
i

i
X−
µ
i
()
=exp

x
k

µ
ik
()
2h
σ
ik
2
o
2
k=1
D
￿
￿
￿
￿
￿
￿
￿
Activation of Output:
Zj
=wij
i
￿
y
i
+w
0j
￿
Radial Basis Function (RBF) Networks
Page 9
￿
Ensembles of Radial Basis Function (ERBF)
Networks
￿
Different network topologies
￿
Consensus
(‘averaging’)
and democracy
(‘majority’) concepts
￿
Training using both original and distorted
data
Page 10
C1C2C3
Original Images
Original + Gaussian Noise
Original + Geometric Transformation
RBF 22
RBF 23
RBF 21
RBF 11RBF 12
RBF 13
RBF 32RBF 33RBF 31
JUDGE
Input Images
Judge for ERBF1: Ifthe norm of the average of 5 of the
9 outputs is greater than threshold θthenaccept
(‘recognize’) elsereject.
Page 11
￿
3000 normalized frontal images
￿
150 subjects
￿
Resolution of images –640x480 encoded in
256 gray scale levels
￿
Size of faces: 64x72 or 32x72
Page 12
Page 13
￿
Experiment 1: Comparison of partial faces vs
full face recognition while varying the number of
images used for training
￿
Experiment 2: To asses recognition
performance depending upon how much of
partial face is used
Page 14
￿
Training and Testing strategy:k -fold Cross Validation
￿
Total number of CV cycles: 20
4.6795.338.6791.33
9 Images
per
Subject
7.6792.3315.3384.67
5 Images
per
Subject
20.6779.3322.3377.67
1 Image
per
subject
False
Positive
%
Rejected
(Correct)
%
False
Negative
%
Accepted
(correct)
%
Partial Face
Recognition
Using RBF
Network
Page 15
397496
9 Images
per
Subject
793991
5 Images
per
Subject
13871585
1 Image
per
subject
False
Positive
%
Rejected
(Correct)
%
False
Negative
%
Accepted
(correct)
%
Partial Face
Recognition
Using
Network
Ensemble
Page 16
496496
9 Images
per
Subject
595892
5 Images
per
Subject
9911189
1 Image
per
subject
False
Positive
%
Rejected
(Correct)
%
False
Negative
%
Accepted
(correct)
%
Full Face
Recognition
Using RBF
Network
Page 17
199397
9 Images
per
Subject
496595
5 Images
per
Subject
694892
1 Image
per
subject
False
Positive
%
Rejected
(Correct)
%
False
Negative
%
Accepted
(correct)
%
Full Face
Recognition
Using
Network
Ensemble
Page 18
Page 19
￿
Partial faces are sufficient for face
recognition as faces are primarily
symmetric in nature
￿
There is not much improvement in
performance as the amount of
information in the partial face is
increased