Top Banner
Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti
14

Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Mar 31, 2015

Download

Documents

Ezequiel Poplin
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Variable Metric For Binary Vector Quantization

UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCEJOENSUU, FINLAND

Ismo Kärkkäinen and Pasi Fränti

Page 2: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Distance and distortion functions

Distance function:

Distortion function:

dK

k

d

jkikjid cxcxL

1

1

,

N

ipidN i

cxLCXE1

1 ,,

Page 3: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Distortion for binary data

Internal distortion for one variable:

djkjkdjkjkjk crcqD 1

qjk = the number of zeroes rjk = the number of onesCjk = the current centroid value for

variable k of group j.

Page 4: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Optimal centroid positionOptimal centroid position depends on the metric. Given:

1/1 dThe optimal position is:

jkjk

jkjk

rq

rc

Page 5: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Proposed algorithm

Create initial solution. Set d to dmax. While d > dmin, do: Perform one GLA iteration. Decrease d. Set d to dmin. Iterate GLA until solution has converged. Return final solution.

Page 6: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Example of centroid location

0.50

0.60

0.70

0.80

0.90

1.00

6.0 5.5 5.0 4.5 4.0 3.5 3.0 2.5 2.0 1.5 1.0

d

c

q=10,r=90

q=20,r=80

q=30,r=70

q=40,r=60

Page 7: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Second example

q = 7 5 r = 2 50 .0 1 .0

c = 0 .5

c = 0 .3 7

c = 0 .2 5

c = 0

d = ( = 0 )

d = 3 ( = 0 .5 )

d = 2 ( = 1 )

d = 1

Page 8: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Test images

Blockwise quantization of pixels into two levels according to the mean value of the 4x4 blocks.

44 pixel blocks.

Page 9: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Results for Bridge

Bridge1

.48

9

1.4

67

1.3

87

1.4

41

1.3

86

1.2

93

1.3

67

1.3

32

1.3

08

1.2

45

1.1

1.2

1.3

1.4

1.5

1.6

GLA SA Split PNN

Dis

tort

ion

Binary

Soft

Prop.

Page 10: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Results for Camera

Camera1

.77

1

1.7

65

1.6

74

1.6

93

1.6

46

1.5

45

1.6

59

1.6

01

1.5

56

1.4

96

1.4

1.5

1.6

1.7

1.8

GLA SA Split PNN

Dis

tort

ion

Binary

Soft

Prop.

Page 11: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Results for CCITT-5

CCITT-5

0.0

33

0.0

40

0.1

37

0.0

26

0.0

31

0.0

29

0.1

65

0.0

26

0.0

45

0.0

27

0.00

0.05

0.10

0.15

0.20

GLA SA Split PNN

Dis

tort

ion

Binary

Soft

Prop.

Page 12: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Results for DNA

DNA7

.05

6

6.6

22

6.5

74

7.2

30

6.7

57

6.6

62

6.5

96

6.5

70

6.6

27

6.5

70

6.2

6.4

6.6

6.8

7.0

7.2

7.4

GLA SA Split PNN

Dis

tort

ion

Binary

Soft

Prop.

Page 13: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Different parameter combinations (Bridge)

GLA SA

Linear Exponential Exponential

d0 0.05 0.1 0.25 0.95 0.99 0.95 0.99 2 1.325 1.335 1.359 1.318 1.314 1.278 1.254 3 1.319 1.327 1.346 1.312 1.310 1.269 1.245 4 1.318 1.325 1.344 1.310 1.309 1.268 1.246 5 1.317 1.324 1.344 1.311 1.309 1.270 1.249 6 1.318 1.323 1.342 1.311 1.307 1.271 1.252

Page 14: Variable Metric For Binary Vector Quantization UNIVERSITY OF JOENSUU DEPARTMENT OF COMPUTER SCIENCE JOENSUU, FINLAND Ismo Kärkkäinen and Pasi Fränti.

Different parameter combinations (DNA)

GLA SA Linear Exponential Exponential

d0 0.05 0.1 0.25 0.95 0.99 0.95 0.99 2 6.679 6.718 6.728 6.646 6.668 6.571 6.571 3 6.646 6.688 6.701 6.628 6.641 6.570 6.570 4 6.685 6.644 6.647 6.643 6.641 6.570 6.570 5 6.662 6.661 6.671 6.630 6.627 6.570 6.570 6 6.663 6.643 6.659 6.637 6.682 6.570 6.570