Support Vector Machine Pattern. Recognition Approach ... Automated classifier. Computer classification. 0 loops. ICCP'09, Cluj Napoca, August 27-29, 2009.
ICCP’09, Cluj Napoca, August 27-29, 2009
Carpet Wear Classification based on Support Vector Machine Pattern Recognition Approach 1,2
C. Copot, S. Syafiie, S. Vargas, R. de Keyser, L. van Langenhove, C. Lazar 2
4
3
2
1
Department of Automatic Control and Applied Informatics, “Gheorghe Asachi” Tehnical University of Iasi 2 Department of Electrical energy, Systems and Automation, Ghent University 3 Department of Telecommunication, Ghent University 4 Department of Textile, Ghent University
1
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009
Outline
● ● ● ● ● ●
Introduction Features extraction PCA (Principal Component Analysis) Features classification Results Conclusion
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Introduction
Objective Objective 0 loops
Carpet sample
Mechanical wear
Visual assessment by human expert (compared with standard)
5000 loops 10000 loops 15000 loops 20000 loops 25000 loops
Label 5 4.5 4 . . . 1
Automated classifier
Human classifier 3D Laser Scanner
Co-occurrence matrix
Computer classification
Computer LABEL
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009
Data Acquisition
Fig 1. Digital image
Introduction
Fig 2. Laser image
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Features extraction
Co-occurrence matrix ¾ Co-occurrence matrix composed 1
3
4
2
7
1
1
2
5
6
3
6
7
1
2
1
1
2
7
5
Gray-scale matrix
3
4
5
6
7
2
3
1
0
0
0
0
0
0
0
0
1
0
2
0
0
0
1
0
1
0
4
0
1
0
0
0
0
0
5
0
0
0
0
0
1
0
6
0
0
0
0
0
0
1
7
1
0
0
0
1
0
0
1
1
2
2 3
Co-occurrence matrix ( 0 o & d = 1)
Cd ( g (i, j ), g (i + k , j + l )) = Cd ( g (i, j ), g (i + k , j + l )) + 1 Where:
d = k2 + l2
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009
Haralick’s Features 9 9 9 9 9 9 9 9 9 9 9 9 9
Features extraction
Energy Sum Entropy Entropy Sum Variance Inertia Difference Variance Homogeneity Difference Entropy Contrast Informational Measures of Correlation Correlation Maximal Correlation Coefficient Sum Average Theoretical representation
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009
Features Validation
Features extraction
Name of carpet
Human expert classification
20 kl 517
BMW
20 kl 517 – 2
2.5
20 kl 517 – 4
2.5
20 kl 517 – 6
2.5
20 kl 517 – 8
2
20 kl 517 – 10
2
20 kl 517 – 12
2
Table of human expert Haralick feature - Energy
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 PCA
What and why PCA?
¾ Principal Component Analysis is a useful technique used to reduce the dimensionality of large data sets, such as those from micro array analysis; ¾ When there are more than three variables, it is more difficult to visualize graphically their relationships. Principal Components 0%
2% 4%
0%
PC1 = 94%
PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC 7 PC 8 PC 9 PC 10 PC 11 PC 12 PC 13 PC 14
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 PCA
PCA Results carpet KL-517
class 2 class 2.5 class 5
carpet KL-517
60
class 2 class 2.5 class 5
40
40 30 20
20
PCA 3
10
PCA 2
0
0 -10 -20
-20
-30 -40
-40
-50 100 50
-60
100 50
0 -80 -150
-100
-50
0
PCA 1
50
100
PCA 2
2D Plotting of first tow PC
0 -50
-50
-100 -100
-150
PCA 1
3D Plotting of first third PC
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Features classification
Binary SVM
¾ A separated hyperplane:
wT x + b = 0 wT x + b > 0 if
y =1
wT x + b < 0 if
y = −1 Where
⎡ + 1⎤ w T x + b = ⎢⎢ 0 ⎥⎥ ⎢⎣ − 1 ⎥⎦
- x is training data - y is indicator vector
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Features classification
Nonlinear Nonlinear -- SVM SVM
Input space
k ( xi , x j ) = ( xi x j + 1) p
k ( xi , x j ) = e
−|| x − y|| / 2σ 2
The separable plane will be:
Feature space
- polynomial kernel - Gaussian kernel
k ( xi , x j ) wT + b = 0
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Features classification
Multi – class SVM ¾ One - Against - All decomposition 8
c las s c las s c las s c las s c las s
6
1 2 3 4 5
PC 2
4
2
0
-2
-4
-6 -4
-2
0
2
PC 1
4
6
8
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Results
Testing Results ¾For every carpet are used: - 4 points for training - 1 point for testing
Type of carpet
Polynomial kernel %
Gaussian kernel %
A8 – 501
84,4
84,4
A8 – 701
84,4
100
Big 4
100
100
20 KL 803 beige
84,4
84,4
LA 7
84,4
100
Pr 84
100
100
20 KL 517
100
100
Big 8
86
86
LA 9
72
100
20 KL 803
100
100
Over all
89,56
95,48
Table of percentage classification
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009 Results
SVM Classification carpet KL-517 class 2 class 2.5 class 5 SVM 40 20
PCA 3
0 -20 -40 -60 60 40 20 0
100
-20
50 0
-40 -50
-60 PCA 2
-80
-100 -150
PCA 1
Classification for carpet KL-517
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009
Conclusion ¾ Automated carpet classification can be achieved using co-occurrence matrix, PCA and SVM ¾ Polynomial kernel gives 88,32% over all classification ¾ Gaussian kernel gives 94,24% over all classification
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
ICCP’09, Cluj Napoca, August 27-29, 2009
THANK YOU!
Cosmin COPOŢ – “Gheorghe Asachi” Tehnical University of Iasi
