1. Neural Networks Ensembles For Face Recognition
B.Mohabbati , R.Ebrahimpour , S.Kasaei , E.Kabir
Department of Mathematics and Computer Science
Amirkabir University of Technology
School of Cognitive Sciences, Institute for Studies
in Theoretical Physics and Mathematics (IPM)
Computer Engineering Department, Sharif University of Technology
Department of Electrical and Computer Engineering, Tarbiat Modarres University
Tehran, Iran

2. Outline Face Recognition What is face recognition? Its applications
Different approaches
Neural Networks approach
Combining Classifiers
Experimental results
Conclusions

3. What is Face Recognition?
It works with the most obvious individual identifier – the human face.
A set of two task:
Face Identification: Given a face image that belongs to a person in a database, tell whose image it is.
Face Verification: Given a face image that might not belong to the database, verify whether it is from the person it is claimed to be in the database.
Biometrics : digital analysis using cameras or scanners of biological characteristics such as
facial structure, fingerprints and iris patterns to match profiles to databases of
people

4. Face recognition system
Applications
Applications in security :
Authentication
Identification
Multimedia Management
Security
Smart Cards
Surveillance
Airports and railway stations Stadiums Public transportation Government offices Businesses of all kinds
Quick way to discover criminals
face information
for each person
in a fixed domain
a person whose face is
in the input image
Face recognition system

5. Feature-Based Approaches:
Different Approaches
Feature-Based Approaches:
Holistic Approaches :
Geometric Features
Euclidian Distance
Graph Matching
Template Matching
Transform-Markov Model
Hidden-Markov Model …
Neural computing provides technical information
processing methods that are similar to the way
information is processed in biological systems,
such as the human brain
Neural Networks
Eigenfaces “Principal Component Analysis(PCA) “
Fisherfaces “Linear Discriminant Analysis(LDA)”

6. Biological inspirations
Some numbers…
The human brain contains about 10 billion nerve cells (neurons)
Each neuron is connected to the others through synapses
Properties of the brain
It can learn, reorganize itself from experience
It adapts to the environment
It is robust and fault tolerant

7. Biological neuron A neuron has A branching input (dendrites)
A branching output (the axon)
The information circulates from the dendrites to the axon via the cell body
Axon connects to dendrites via synapses
Synapses vary in strength
Synapses may be excitatory or inhibitory

8. What is an artificial neuron ?
Definition : Non linear, parameterized function with restricted output range
The perceptron algorithm converges if examples are linearly separable + y
Activation functions w0 w1 w2 wn
Activation functions
Linear
Logistic
Perceptron
Rosenblatt (1962)
Linear separation
Inputs :Vector of real values
Outputs :1 or -1 x1 x2 xn

9. Learning
The procedure that consists in estimating the parameters of neurons so that the whole network can perform a specific task
2 types of learning
The supervised learning
The unsupervised learning
The Learning process (supervised)
Present the network a number of inputs and their corresponding outputs
See how closely the actual outputs match the desired ones
Modify the parameters to better approximate the desired outputs
The desired response of the neural network in function of particular inputs is well known.
A “Professor” may provide examples and teach the neural network how to fulfill a certain task
Idea : group typical input data in function of resemblance criteria un-known a priori
Data clustering
No need of a professor
The network finds itself the correlations between the data
Examples of such networks :
Kohonen feature maps

10. Multi-Layer Perceptron (MLP)
Class1 Class2
Output layer
One or more hidden layers
Sigmoid activations functions
2nd hidden layer
1st hidden layer
Input data

11. Error Back Propagation
MLP Learning method
Error Back Propagation
-Used to train the MLP
-Uses gradient descent to minimise the
squared error between actual and desired outputs
-The error is summed over all inputs
-Error space can have local minima,
which traps gradient descent
Back-propagation algorithm
I f the jth node is an output unit
Credit assignment
Learning Rate : Momentum term to smooth
The weight changes over time

12. Back Propagation Algorithm
Input Pattern 1
Target Output 1 =
Input Pattern 2
Target Output 2
1000 “ Binary digit”
0100
Input Pattern n
Initialise weights
Present input and target
Calculate the actual output
Given this input
From last layer, work backwards
Updating the weights

13. Properties of Neural Networks
Supervised networks are universal approximators
Theorem : A Multi Layer Perceptron (MLP) with only one hiddenlayer has the capability to act as a universal approximator, (Hornik et al., 1989)
Type of Approximators
Linear approximators : for a given precision, the number of parameters grows exponentially with the number of variables (polynomials)
Non-linear approximators (NN), the number of parameters grows linearly with the number of variables
Three-Layer
Arbitrary
(Complexity
Limited by No.
of Nodes) A B

14. Face recognition using Neural Network
Problems
Work done on a very selective set of face images, mostly:
In upright position
Lighting and background controlled
Either in frontal or profile view
Have no occlusions, facial hair
Requirements
Accurate
Efficient
Light invariant
Rotation invariant and …
Image Data
Train-Test
Domain Transformation
(Preprocessing)
Neural Networks model
(Classifier)
Classified face result
Face recognition
Standard Face Recognition Architecture

15. Preprocessing methods
Why Preprocessing ?
The curse of Dimensionality
The quantity of training data grows exponentially with the dimension of the input space
In practice, we only have limited quantity of input data
Increasing the dimensionality of the problem leads to give a poor representation of the mapping
The preprocessing has a huge impact on performances of neural networks
Preprocessing methods
Normalization
-Inputs of the neural net are often of different types with different orders of magnitude
(E.g. Pressure, Temperature, etc.)
- It is necessary to normalize the data so that they have the same impact on the model
- Translate input values so that they can be exploitable by the neural network
Component reduction
Sometimes, the number of inputs is too large to be exploited
The reduction of the input number simplifies the construction of the model
Goal : Better representation of the data in order to get a more synthetic view without losing relevant information
Reduction methods (PCA, CCA(Curvilinear Components Analysis), etc.)

16. Principal Components Analysis (PCA)
Principle
Linear projection method to reduce the number of parameters
Transfer a set of correlated variables into a new set of uncorrelated variables
Map the data into a space of lower dimensionality
Form of unsupervised learning
Properties
It can be viewed as a rotation of the existing axes to new positions in the space defined by original variables
New axes are orthogonal and represent the directions with maximum variability
Faces viewed as vectors
Supposing a data set X with N data points, each of dimensions P

17. The PCA - Eigenface
Here we will summarize the concept and method to finding principal factors:
XM = X -
For training set X1, X2, …… XN
Average face
Difference vector (Zero Mean)
Covariance matrix
Choose i largest Eigenvalues
“ i is the inherent dimensionality of the subspace
governing the original image “
First principal components

18. The PCA - Eigenface = 0.9571 * -0.1945 * +0.0461 * +0.586 * M-PCA PCA
= * * * *
Training
image set
Adjusted Training
image set
M-PCA
Image mean
PCA
N -Eigenface

19. Training the classifiers
image set
Adjusted Training
Image mean
PCA
N -Eigenface
M-PCA
M-PCA
Combining
Classifiers
Training the classifiers

20. Neural Networks Ensembles
Two main approach to committee machines (Combining classifiers, classifiers fusion):
1-Static committee machines (Ensemble) 2- Dynamic committee machines (modular)
Static committee machines (Ensemble):
Static committee machines are ones where the responses of experts are combined without the mechanism seeing the input.
Ensembles of classifiers are trained on different or similar data and using different or similar features. The classifiers are run simultaneously and their outputs are merged into one compound classification.
2 main methods : “Ensemble averaging and Boosting” …
y2(n)
yL(n)
Input x(n)
Expert L
Expert1
Expert2
Combiner
output
y1(n)

21. S Dynamic committee machines: … Gating network Input x(n) output y1(n)
Dynamic committee machines: input signal is directly involved in combining ouputs
Eg Mixtures of experts and hierarchical mixtures of experts
Gating network decides the weighting of each network
y1(n) …
Gating network
output
y2(n)
yL(n)
Input x(n) S
g1(n)
g2(n)
gL(n)
ExpertL
Expert1
Expert2

22. Some of Ensemble Methods:
Majority voting : In majority voting every voter has one vote that can be cast for any one candidate. The candidate that received the majority (i.e. more than half) of the votes, wins the election.
Average voting : Average of vote and the candidate with the highest average wins the election.
Product rule voting :Eeach voter gives a confidence value for each candidate. Then all confidence values are multiplied per candidate. The candidate with the highest confidence product wins
Classifiers
Support for W1
Support for W2
Decision C1
0.8
0.2 W1 C2
0.4
0.6 W2 C3
0.3
0.7 C4 C5
MAJ -
AVR
0.48
0.52
PRO
0.2352 C1 C2 C3 C4 C5
Combiner W1 W2

23. Purposed method:
Consider the Ensemble network system shown in this Fig. Here we see a set of M “level-0” networks . N1 to Nm whose outputs are combined using a “level-1” Network N* .
Training Rule: The idea is to train the “level-0” networks first and then examine their behavior when generalization. This provides a new training set which is used to the train “level-1” network. … y1 y2 ym y
M-PCA N1 N2 Nm N*
level-0
level-1 …

24. Experimental results and discussion
Face Database
We have used the ORL database which contains a set of faces taken between April 1992 and April 1994 at the Olivetti Research Laboratory in Cambridge, UK.
There are 10 different images of 40 distinct subjects.
There are variations in facial expression (open/closed eyes, smiling/non-smiling), and facial details (glasses/no glasses). All the images were taken against a dark homogeneous
background with the subjects in an up-right, frontal position, with tolerance for some tilting and rotation of up to about 20 degrees. There is some variation in scale of up to about 10.
The images are grayscale with a resolution of 92*112.
The set of 10 images for one subject, considerable variation can be seen

25. Experimental results and discussion
Training
image set
Adjusted Training
Image mean
PCA
N -Eigenface
Step 1
M-PCA
Combining
Classifiers
Step 2
Testing
image set
Zero mean
Testing images
M-PCA
Combining
Classifiers
Step 3
Classified faced

26. Different net topologies for the same input data
Percentage correct classification on test set (200 faces – 5 times repetition)
Different net topologies for the same input data
Principal Components
Net Topology
Best (%)
Average(%)
1-25
25:40:40 89
87.7
1-30
30:80:40
88.5
87.9
1-35
35:80:40 90
89.7
1-40
40:40:40
89.5
88.9
1-50
50:40:40
86.6
1-100
100:60:40 86
84.1
1-200
200:80:40 84
82.2
Principal Components
Net Topology
Best
Average
1-40
40:40:40
89.5
88.9
40:20:40
83.8
40:80:40
90.5
89.1
40:100:40 89
87.6

27. Correct rate Percentage of Some manually selected committees
Averaging
Majority voting
Purposed method
40:20:40 (5 nets)
90.5
88.15
93.35
40:40:40 (5 nets)
88.35
88.5 95
40:80:40 (5 nets) 92 90
97.15 … y1 y2 y5 y
M-PCA
MLP1
MLP2
MLP5
MLP*
level-0
level-1

28. Three main strategies for making independent individual classifiers:
Conclusion
Three main strategies for making independent individual classifiers:
Using different procedure, for example different kind of classifiers or different parameters in an identical procedure. For example different MLP can be trained based on different initial weights or learning parameter or number of nodes.
Using different representation of input (or feature set), although this approach is practically effective in improving generalization, managing the procedure and analyzing is more complicated.(K.M. Ali and M.J. Pazzini,1995)
Using identical representation but different training sets

29. Laboratory for develop and maintaining the ORL
Acknowledgments
I Would like to thank prof. Shahshahani for helpful comments and Olivertti Reseaerch
Laboratory for develop and maintaining the ORL
database .

30. Tanks for your attention