The document discusses a face recognition system developed by Dong Hyun Roh using neural networks, outlining various related research methodologies such as template matching and PCA for feature extraction. It presents experiments conducted with a dataset of 15 individuals and analyzes the performance of the compression and recognition networks, showing varying recognition rates across different experimental setups. The conclusion suggests further improvements to enhance the performance of the face recognition system.

1. Face Recognition
using
Neural Network
by Dong Hyun Roh
Computer Science Department
KAIST

2. content
 face recognition
 related researches
 my system
 experiments
 analysis
 future work
 references

3. face recognition
 What is face recognition?
face recognition system
face information
for each person
in a fixed domain
a person whose face is
in the input image

4. related researches
 Template matching
very simple, very sensitive to noise
 PCA(associative memory)
use PCA to extract important features and to reduce the dimension
of the input data
 Image compression(back propagation network)
use compression network to extract important features and to
reduce the dimension of the input data
this is my approach
 Others
extract visual feature like edges, shape of each components, etc.

5. My system
 Object
face recognition for the face image domain in which there are some
small variations
 face Image domain
15 persons
12 images per person
variation such as expression, direction of light, noise, glasses
cen gla hap lef nog noi
nor rig sad sle sur win

6. My system
 Procedure
Normalization
Hidden nodes
of
Compression
network
Recognition
network
Person 1

7. My system
 Normalization
original image : 320*243
normalized image : 32*24
block size : 10*10
averaging the intensity of pixels in block
 compression network
input = raw data of image
output = raw data of image ( same as input )
32*24 - 40 - 32*24 structure
backpropagation learning algorithm

8. My system
 recognition network
input = values of hidden layer of compression network
output = a person whose face is in the image
40 - 10 - 15 structure
backpropagation learning algorithm

9. Experiment 1
 Compression network
training data
nor images of 15 persons
2000 trainings, learning rate = 0.005
result(sample)
cen gla hap lef nog
noi
nor
rig sad sle sur win
Training
image

10. Experiment 2
 Recognition network
use the result of experiment 1
training data
nor images of 15 persons
2000 trainings, learning rate = 0.05
result
total face image 180 ( including the training data)
the rate of correct recognition = 133/180*100 = 73.9%
the distribution of errors
nor cen gla hap lef nog noi rig sad sle sur win
0 6 3 3 12 2 0 12 2 2 2 3

11. Experiment 3
 Recognition network
not use the result of experiment 1
assign random 40 dimensional key to each image
training data
 nor image of 15 persons
2000 trainings, learning rate = 0.05
result
total face image 180 ( including the training data)
the rate of correct recognition =12/180*100 = 6.7%
the distribution of errors
nor cen gla hap lef nog noi rig sad sle sur win
14 14 14 14 14 14 14 14 14 14 14 14

12. Experiment 4
 Recognition network
not use compression network
use raw image(32*24) for training and test
training data
nor image of 15 persons
2000 trainings, learning rate = 0.05
32*24 - 30-15 structure
result
total 180 images, the rate of correct recognition = 6.1%
the distribution of errors
nor cen gla hap lef nog noi rig sad sle sur win
14 14 14 15 14 14 13 15 14 14 14 14

13. Analysis
 The hidden of compression network
encode the inputs in a smaller dimensional subspace that retains
most of the important information
if the hidden units are linear, the best solution to this problem is the
least squares solution(i.e. to have the hidden units span the L
principle components with the highest eigenvalues)
Cottrell et al. Found that the weights of 16 hidden units span the
space of the 13 first engenvectors of he covariance matrix of the
inputs.
when transformed to gray scale and graphically displayed, the
hidden unit receptive and projective fields looked “face-like” and
showed some similarity to the eigenvectors or eigenfaces
The hidden value of compression network is very
useful in face recognition!!

14. Further work
 Improvement of the performance of face recognition
improvement of the compression network
 If the compression network is trained with all face images,
what will be different from experiment 2 ?
The performance of compression network
The performance of recognition network

15. Reference
 Dominque Valentin, Herve Abdi, Alice J. O’Toole, Garrison
W. Cottrell, “Connectionist Models of Face Processing: A
survey”, 1994