Está en la página 1de 20

Face recognition

Enee 752 – Knowledge


Engineering
Fiorella Haim
Problem definition
• Determine the identity of a face in an
image
• The image can be a frame from a video
• Processing needs to be fast
• Classification problem
• Need faces images for training
Challenges
• Variations between images of the same
face can be greater than those form
different faces
– Illumination
– Pose (angle)
– Accessories
– Expression
Motivation: Security
• Recognize criminals
– In public spaces (airports, shopping centers)
– In stores
• Verify identity to grant access in restricted
areas: non-invasive Biometrics
– Airports
– Office
– Risk: privacy rights
Motivation: Human-Machine
Interface
• ATMs
• Computers
• Cell phones (Omron)
• Intelligent buildings
– Music settings personalized
Learning
• Identify face features
• Train the system with a convenient database
ORL
DATABASE

• 40 persons
•10 images per
person
•Different times,
expressions,
lightning and
details
(glasses)
Yale Database
• 165 images of 15 individuals. There are 11
images per subject, one per different facial
expression or configuration: center-light,
w/glasses, happy, left-light, w/no glasses,
normal, right-light, sad, sleepy, surprised, and
wink.
Neural Networks: RBFNN
• Hock Koh et al approach
• Preprocessing:
– Detect the face
– Enhance the image
– Identify nose and boundaries
• Features: Radial grid centered on the
nose, value at each point: mean of region
Example of Radial Grid
Different area diameters - Radial grid size 6x12
6

-2

-4

-6
-6 -4 -2 0 2 4 6
Architecture
• Feed forward architecture, 3 layers
– Input layer NI nodes, fully connected to the hidden
layer
η×d2
– Hidden layer RBF σ =
2
− x −m 2

 j i 
g0 (x j ) = 1 g i ( x j ) = exp 
 η×d
2

 

– Output layer linear Nc nodes, one for each class


NH
y k ( x j ) = ∑ λik g i ( x j )
i =0
Architecture
INPUT LAYER HIDDEN LAYER OUTPUT LAYER

0
λ01

λ0C λ02

xj1 y1
1 1 1
m1

λ21
xj2 y2
λ22
2 2 2
m2
λ2C

yC
xjI NC
NI
NH
mH
Input feature Classification
vector size classes
Training and classification
• Xj training vector
1 _ if _ x j ∈ yk
• Output: yk ( x j ) = 
0 _ otherwise

• Calculate y and choose class k / yk is the


highest component of k
Results (author’s database)

Input Dimension NH = 1092 NH = 525 NH = 105 Average

8×8 91.2 90 88.4 89.9

16 × 16 97.1 96.4 94.2 95.9

24 × 24 97.6 97.5 92.7 95.9

32 × 32 97.7 97.4 93.9 96.3

Average 95.9 95.3 92.3


SVM: 2 different approaches
• Feature extraction:
– Guo et al: PCA analysis, linear SVM
– Bicego et al: Multilevel Spline coefficients,
2
x− y
radial SVM: −
K ( x, y ) = e
2σ 2

• Same multiclass approach


– One against one
– C(C-1)/2 trained machines
– Tennis tournament
Tennis Tournament
Example for 10 classes

C-1 comparisons
Running time SVM
• O(n2) for training
• O(n) for evaluating an input
• Bicego’s approach: for 100 classes, 10
minutes for training and 10 seconds for
classifying
– Matlab @ Intel Celeron 850 MHz
Comparison of different
approaches
Reference Feature # of Recognition Datbase Error Comment
Extraction features Method used (%)
Method
[2] Grid average 256 RBFNN own 2.9 Up to 25°
sampling variation in
direction
[3] PCA + Fisher 25 RBFNN ORL 1.92 Fast training
discriminant method
[1] PCA 10-100 SVM ORL 3

[4] Multilevel 80 SVM ORL 2.75 Fast


B- splines evaluation of
inputs
Conclusions
• Already in use mainly for security and
human-machine interface applications
• Machine learning techniques suitable for
this application domain, good results
– Trade-off: processing time – confidence
• Preprocessing and choice of features are
key issues -> domain expertise is needed
• Ethics? Privacy laws? Mistakes? Still to be
addressed.
References
• [1] Guodong Guo; Li, S.Z.; Kapluk Chan “Face recognition by
support vector machines”,
Automatic Face and Gesture Recognition, 2000. Proceedings.
Fourth IEEE International Conference on, Vol., Iss., 2000
Pages:196-201
• [2] Lian Hock Koh, Surendra Ranganath and Y. V. Venkatesh, “An
integrated automatic face detection and recognition system”, Pattern
Recognition, Volume 35, Issue 6, June 2002, Pages 1259-1273.
• [3] Face recognition with radial basis function (RBF) neural networks
Meng Joo Er; Shiqian Wu; Juwei Lu; Hock Lye Toh; Neural
Networks, IEEE Transactions on Volume 13, Issue 3, May 2002
Page(s):697 - 710
• [4] Face recognition with Multilevel B-Splines and Support Vector
Machines Manuele Bicego, Gianluca Iacono, Vittorio Murino
November 2003 Proceedings of the 2003 ACM SIGMM workshop
on Biometrics methods and applications

También podría gustarte