Current subjects in computer science

Seminar #2

PATTERN RECOGNITION

Seminar Description
Introduction to pattern recognition systems Bayesian decision theory Feature Extraction and Selection Methods Example: Face Recognition

FACE RECOGNITION

FACE RECOGNITION
The face is our primary focus of attention in social intercourse, playing a major role in conveying indentity and emotion. Humans can recognize thousands of faces.

FACE RECOGNITION

FACE DETECTION

FACE RECOGNITION BY APPEARANCE (I)

Appearance-based methods rely heavily on the concept of an image space. A twodimensional image I (x,y ) or any type of visual data (depth maps, flow fields, etc.) may be viewed as a vector (or point) in a very high dimensional space, often called the image space, where each coordinate of the space corresponds to one pixel value of the original image. In general, a grayscale image with r rows and c columns describes a vector x in a m -dimensional image space, where m = r x c . With this image representation, the image becomes a very high dimensional feature vector, and perhaps the simplest classification approach is a nearest neighbor classifier in the image space
7

FACE RECOGNITION BY APPEARANCE (II) 30 members

1 2 3 4 5 6 7 8 9 10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

12 255 100 K 12 13 234 34 K 34 15 34 1 K 23 2 K 56 23 34 24 33 123 K 46 22 5 21 K 76 23 34 213 K 12 11 23 23 K 23 M M M M M 12 0 0 K 255

8

Number of pixels

FACE RECOGNITION BY APPEARANCE (III)

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

Face recognition systems can not be directly implemented in such large spaces as the classifiers are slowed down by the great size of the input vector.
9

FACE RECOGNITION BY APPEARANCE (IV)

1.5

x 10

0.5

face1 face2 face3 face4 face5 face6 face7 face8

-0.5

-1

-1.5 -2

-1.5

-1

-0.5

0.5

1.5 x 10
4

10

DIMENSIONALITY REDUCTION

PCA 112x92=10304 2

Reducing the dimensionality of the data, we speed up the computations, without losing too much information. The most popular technique for dimensionality reduction purposes is Principal Component Analysis (PCA).
11

FACE RECOGNITION USING PCA TRAINING

1. Acquire an initial set of face images (the training set)

2. Calculate the PCA transform of the training set (compute the eigenvectors -eigenfaces-)
1.5 x 10
4

3. Project the training images onto the eigenspace

0.5

face1 face2 face3 face4 face5 face6 face7 face8

-0.5

-1

-1.5 -2

-1.5

-1

-0.5

0.5

1.5 x 10
4

12

FACE RECOGNITION USING PCA and Matlab TRAINING

%data matrix %each column is a face X=[x1 x2 x3 x4 x5 x6 x7 x8]; % PCA [V,D,Average] = pc_evectors(X,Nv); %projecting the data P=V'*Xm;
1.5 x 10
4

(matlab code)

0.5

face1 face2 face3 face4 face5 face6 face7 face8

-0.5

-1

-1.5 -2

-1.5

-1

-0.5

0.5

1.5 x 10
4

13

FACE RECOGNITION USING PCA and Matlab (II)

TEST

(TO RECOGNIZE NEW FACES)

1. Project the new image onto the eigenspace 2. Use a classifier to see which class is nearest to the new face

classifier.m

14

PRACTICE EXERCISES
PRACTICE 1: Study of PCA with bidimensional points and with images Face recognition system based on PCA with the face database ORL. PRACTICE 2: Generation of a face database with Stralsund students faces PRACTICE 3: Use our face database to test the face recognition system based on PCA.

THIS AFTERNOON!!!
15