Digital Image Processing

using
Local Segmentation
Torsten Seemann
B. Sc (Hons)

School of Computer Science and Software Engineering

Faculty of Information Technology
Monash University
Australia.

Submission for the degree of Doctor of Philosophy

April 2002

Contents

1 Introduction

2 Notation and Terminology

2.1

Digital images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2

Image statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2.1

The histogram . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2.2

The mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.2.3

The variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10

2.2.4

The entropy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

Image algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

2.3.1

Image-scalar operations . . . . . . . . . . . . . . . . . . . . . . .

11

2.3.2

Image-image operations . . . . . . . . . . . . . . . . . . . . . . .

12

2.4

Image acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

2.5

Types of noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

14

2.5.1

Additive noise . . . . . . . . . . . . . . . . . . . . . . . . . . . .

14

2.5.2

Multiplicative noise . . . . . . . . . . . . . . . . . . . . . . . . . .

15

2.5.3

Impulse noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.3

iii

iv

CONTENTS
2.5.4

Quantization noise . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.5.5

The noise function . . . . . . . . . . . . . . . . . . . . . . . . . .

16

2.6

Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

17

2.7

Local windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

3 Local Segmentation in Image Processing

19

3.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

19

3.2

Global segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

21

3.2.1

Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

25

3.2.2

Spatial segmentation . . . . . . . . . . . . . . . . . . . . . . . . .

32

Local segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

37

3.3.1

The facet model . . . . . . . . . . . . . . . . . . . . . . . . . . . .

37

3.3.2

Block truncation coding . . . . . . . . . . . . . . . . . . . . . . .

38

3.3.3

SUSAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

3.4.1

Temporal filtering . . . . . . . . . . . . . . . . . . . . . . . . . . .

41

3.4.2

Spatial filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . .

42

3.4.3

Fixed linear filters . . . . . . . . . . . . . . . . . . . . . . . . . .

43

3.4.4

Rank filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

45

3.4.5

Locally adaptive filters . . . . . . . . . . . . . . . . . . . . . . . .

49

3.4.6

Filtering based on global segmentation . . . . . . . . . . . . . . .

59

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

60

3.3

3.4

3.5

CONTENTS

4 Denoising with Local Segmentation

63

4.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

63

4.2

Global image models . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

64

4.2.1

The facet model . . . . . . . . . . . . . . . . . . . . . . . . . . . .

65

4.2.2

Generalizing the facet model . . . . . . . . . . . . . . . . . . . . .

66

4.2.3

Image sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . .

67

4.2.4

Edges and lines . . . . . . . . . . . . . . . . . . . . . . . . . . . .

68

4.2.5

A useful noise model . . . . . . . . . . . . . . . . . . . . . . . . .

69

4.2.6

Pixel quantization . . . . . . . . . . . . . . . . . . . . . . . . . . .

71

4.2.7

A suitable image model . . . . . . . . . . . . . . . . . . . . . . .

72

4.2.8

Image margins . . . . . . . . . . . . . . . . . . . . . . . . . . . .

73

Measuring image similarity . . . . . . . . . . . . . . . . . . . . . . . . . .

76

4.3.1

Visual inspection . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

4.3.2

Traditional quantitative measures . . . . . . . . . . . . . . . . . .

78

4.3.3

Difference images . . . . . . . . . . . . . . . . . . . . . . . . . .

80

Test images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

82

4.4.1

Square . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

82

4.4.2

Lenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

82

4.4.3

Montage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

83

4.5

A one segment model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

84

4.6

A two segment model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

86

4.6.1

89

4.3

4.4

Application to denoising . . . . . . . . . . . . . . . . . . . . . . .

vi

CONTENTS
4.7

Discriminating between two models . . . . . . . . . . . . . . . . . . . . .

91

4.7.1

A simple model selection criterion . . . . . . . . . . . . . . . . . .

92

4.7.2

Students -test . . . . . . . . . . . . . . . . . . . . . . . . . . . .

96

4.8

Iterative reapplication of the filter . . . . . . . . . . . . . . . . . . . . . . 102

4.9

Rejecting poor local models . . . . . . . . . . . . . . . . . . . . . . . . . 107

4.10 Different local windows . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

4.11 Multiple overlapping local approximations . . . . . . . . . . . . . . . . . . 113
4.11.1 Choosing the weights . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.12 Larger numbers of segments . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.12.1 Extending the clustering algorithm . . . . . . . . . . . . . . . . . . 122
4.12.2 Extending the model order selection technique . . . . . . . . . . . 122
4.12.3 Choosing the initial means . . . . . . . . . . . . . . . . . . . . . . 123
4.12.4 Quantitative results for multiple classes . . . . . . . . . . . . . . . 126
4.13 Estimation of the noise variance . . . . . . . . . . . . . . . . . . . . . . . 132
4.14 Comparison to other denoising algorithms . . . . . . . . . . . . . . . . . . 138
4.14.1 The montage image . . . . . . . . . . . . . . . . . . . . . . . . . 139
4.14.2 The lenna image . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.14.3 The barb2 image . . . . . . . . . . . . . . . . . . . . . . . . . . 145
4.15 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
4.16 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151

CONTENTS

vii

5 Information Theoretic Local Segmentation

153

5.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

153

5.2

Statistical model selection . . . . . . . . . . . . . . . . . . . . . . . . . .

154

5.2.1

Maximum likelihood . . . . . . . . . . . . . . . . . . . . . . . . .

155

5.2.2

Interchangeability of probability and code length . . . . . . . . . .

157

5.2.3

Penalized likelihood . . . . . . . . . . . . . . . . . . . . . . . . .

158

5.2.4

Bayesianism . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

159

5.2.5

MDL : Minimum Description Length . . . . . . . . . . . . . . . .

160

5.2.6

MML : Minimum Message Length . . . . . . . . . . . . . . . . .

161

Describing segmentations . . . . . . . . . . . . . . . . . . . . . . . . . . .

163

5.3.1

Possible segmentations . . . . . . . . . . . . . . . . . . . . . . . .

164

5.3.2

Canonical segmentations . . . . . . . . . . . . . . . . . . . . . . .

164

5.3.3

Valid segmentations . . . . . . . . . . . . . . . . . . . . . . . . .

165

5.3.4

Segmentation parameters . . . . . . . . . . . . . . . . . . . . . . .

166

Using MML for local segmentation . . . . . . . . . . . . . . . . . . . . .

167

5.4.1

A message format . . . . . . . . . . . . . . . . . . . . . . . . . .

167

5.4.2

A uniform prior . . . . . . . . . . . . . . . . . . . . . . . . . . . .

169

5.4.3

A worked example . . . . . . . . . . . . . . . . . . . . . . . . . .

169

5.4.4

Posterior probability . . . . . . . . . . . . . . . . . . . . . . . . .

171

5.5

Application to denoising . . . . . . . . . . . . . . . . . . . . . . . . . . .

172

5.6

A better segment map prior . . . . . . . . . . . . . . . . . . . . . . . . . .

173

5.6.1

174

5.3

5.4

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

viii

CONTENTS
5.7

Optimal quantization of segment means . . . . . . . . . . . . . . . . . . . 174

5.7.1

5.8

Posterior blending of models . . . . . . . . . . . . . . . . . . . . . . . . . 179

5.8.1

5.9

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178

Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180

Incorporating Do No Harm . . . . . . . . . . . . . . . . . . . . . . . . . 181

5.9.1

The FUELS approach to DNH . . . . . . . . . . . . . . . . . . . . 181

5.9.2

Information based DNH . . . . . . . . . . . . . . . . . . . . . . . 182

5.10 Combining overlapping estimates . . . . . . . . . . . . . . . . . . . . . . 186

5.10.1 Equal weighting . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
5.10.2 Message length weightings . . . . . . . . . . . . . . . . . . . . . . 187
5.11 Estimating the noise level . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
5.12 Evolution of the message length . . . . . . . . . . . . . . . . . . . . . . . 191
5.13 Learning from the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
5.13.1 Learning a prior for the segment maps . . . . . . . . . . . . . . . . 194
5.13.2 Learning the prior for DNH models . . . . . . . . . . . . . . . . . 196
5.14 Incorporating more models . . . . . . . . . . . . . . . . . . . . . . . . . . 199
5.15 Improving the running time . . . . . . . . . . . . . . . . . . . . . . . . . . 200
5.16 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
5.16.1 The lenna image . . . . . . . . . . . . . . . . . . . . . . . . . . 204
5.16.2 The barb2 image . . . . . . . . . . . . . . . . . . . . . . . . . . 207
5.16.3 The camera image . . . . . . . . . . . . . . . . . . . . . . . . . 209
5.17 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
5.18 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

CONTENTS

ix

6 Extensions and Further Applications of Local Segmentation

215

6.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

215

6.2

Different noise models . . . . . . . . . . . . . . . . . . . . . . . . . . . .

216

6.2.1

Impulse noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

216

6.2.2

Multiplicative noise . . . . . . . . . . . . . . . . . . . . . . . . . .

222

6.2.3

Spatially varying noise . . . . . . . . . . . . . . . . . . . . . . . .

223

Different structural models . . . . . . . . . . . . . . . . . . . . . . . . . .

224

6.3.1

Multispectral pixels . . . . . . . . . . . . . . . . . . . . . . . . . .

224

6.3.2

Planar regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

224

6.3.3

Higher dimensional data . . . . . . . . . . . . . . . . . . . . . . .

226

6.3.4

Larger window sizes . . . . . . . . . . . . . . . . . . . . . . . . .

227

6.4

Pixel classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

228

6.5

Edge detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

230

6.5.1

Edge strength measurement . . . . . . . . . . . . . . . . . . . . .

231

6.5.2

Probabilistic edge detection . . . . . . . . . . . . . . . . . . . . .

234

6.6

Image enlargement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

237

6.7

Image compression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

241

6.7.1

Aiding lossy compression . . . . . . . . . . . . . . . . . . . . . .

242

6.7.2

Adaptive BTC . . . . . . . . . . . . . . . . . . . . . . . . . . . .

243

6.7.3

Lossless predictive coding . . . . . . . . . . . . . . . . . . . . . .

245

Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

247

6.3

6.8

7 Conclusions

249

List of Figures
2.1

A rectangular digital image of resolution
 . . . . . . . . . . . . . . .

2.2

A typical greyscale image of resolution 


 

. . . . . . . . . . . . . . .

2.3

The histogram for the greyscale image in Figure 2.2. . . . . . . . . . . . .

10

2.4

(a) low contrast image; (b) after enhancement. . . . . . . . . . . . . . . . .

12

2.5

Alpha-blending example: (a) first image; (b) second image; (c) blended image using   . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

2.6

Noise may be introduced at each step in the acquisition process. . . . . . .

13

2.7

Different types of noise: (a) original image; (b) additive noise; (c) multiplicative noise; (d) impulse noise. . . . . . . . . . . . . . . . . . . . . . .

15

2.8

(a) original image; (b) segmented into four segments. . . . . . . . . . . . .

17

2.9

Pixel connectedness: (a) 4-connected [Type I]; (b) 4-connected [Type II];
(c) 8-connected. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

18

2.10 Common local neighbourhood configurations: 5, 9, 13, 21, 25 pixels. . . . .

18

3.1

An image processing hierarchy. . . . . . . . . . . . . . . . . . . . . . . . .

20

3.2

A   sub-image taken from the centre of a  image with two segments.

22

3.3

(a) original image; (b) local  standard deviations, with black representing

3.4

0 and white 72. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

22

Histogram of local  standard deviations shown in Figure 3.3b. . . . . .

23

xi

xii

LIST OF FIGURES
3.5

A   sub-image taken from the right of the   image with two segments.

24

3.6

(a) the 8 bpp pellets image; (b) its histogram; (c) thresholded using  .

26

3.7

Determination of the threshold in mixture modeling. . . . . . . . . . . . .

27

3.8

(a) the mug image; (b) its histogram; (c) thresholded using 

. . . . .

29

3.9

(a)   image with 3 segments; (b) co-occurrence matrix for horizontal pairs
of pixels; (c) co-occurrence matrix for vertical pairs of pixels. . . . . . . . .

36

3.10 BTC example: (a) original    pixel block; (b) decomposed into a bit-plane
and two means; (c) reconstructed pixels. . . . . . . . . . . . . . . . . . . .

38

3.11 Comparison of the hard and soft cut-off functions for pixel assimilation in
the SUSAN algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

40

3.12 Application of the  box filter: (a) original; (b) noisy; (c) filtered. . . . .

42

3.13 Box filtering in the presence of edges: (a) original image; (b) filtered. . . .

43

3.14 (a)  box filter; (b) approximate   Gaussian filter. . . . . . . . . . . .

44

3.15 Application of the median to a homogeneous block of pixels. . . . . . . . .

46

3.16 Application of the median to a homogeneous block with impulse noise. . .

46

3.17 Application of the median to a noiseless edge block. . . . . . . . . . . . . .

47

3.18 The median (incorrectly) implicitly segmenting a noiseless corner block. . .

47

3.19 Using the centre weighted median ( !"# ) on a noiseless corner block. . . .

48

3.20 (a) original image; (b)   median filtered; (c)   weighted median, !"$ .

48

3.21 The Kuwahara filter considers nine  regions within a   window. . . .

50

3.22 Gradient inverse weighted smoothing (GIWS): (a) original pixels; (b) computed weights, unnormalized; (c) smoothed value when centre pixel included. 51
3.23 Pixel weighting function for the GIWS adaptive denoising algorithm. . . . .

52

3.24 Demonstration of the Sigma filter on a noiseless corner block. . . . . . . .

52

LIST OF FIGURES

xiii

3.25 Anisotropic diffusion % weighting functions when &'# . Each function has
been scaled for comparison purposes. . . . . . . . . . . . . . . . . . . . .

54

3.26 The SUSAN spatial weighting component when &' , ignoring that (*)+-,./0
 .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

57

3.27 The SUSAN intensity difference weighting component when #

1 . . . . .

58

3.28 SUSAN denoising: (a) original pixels; (b) spatial weights, unnormalized,
&23
; (c) intensity difference weights, unnormalized,

3
4 ; (d) denoised

value. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

58

4.1

Two images of the same scene: (a) light intensity; (b) range, darker is closer.

65

4.2

One dimensional polynomial approximation: (a) the original signal; (b) constant; (c) linear; (d) quadratic. . . . . . . . . . . . . . . . . . . . . . . . .

4.3

Discrete sampling of an object aligned to the pixel grid: (a) original scene;
(b) superimposed sampling grid; (c) digitized image. . . . . . . . . . . . .

4.4

66

67

Discrete sampling of an object mis-aligned with the pixel grid: (a) original
scene; (b) superimposed sampling grid; (c) digitized image. . . . . . . . . .

67

4.5

(a) step edge; (b) line; (c) ramp edge; (d) roof. . . . . . . . . . . . . . . . .

68

4.6

The standard Gaussian distribution: 57698:)+-,;
/

70

4.7

The effect of different Gaussian noise levels: (a) no noise; (b) added noise

. . . . . . . . . . . . . .

&'
 ; (c) added noise &2< . . . . . . . . . . . . . . . . . . . . . . . .

71

4.8

Effect of quantizing the signal intensity to 8 levels. . . . . . . . . . . . . .

72

4.9

Incomplete  neighbourhoods when processing pixels at the image margins. 74

4.10 Missing pixels are easily replaced with their nearest neighbours. . . . . . .

75

4.11 The pixels involved in border cases are shaded. . . . . . . . . . . . . . . .

76

4.12 Noisy equals original plus noise: (a) original image, = ; (b) additive noise, > ;
(c) noisy image, =@? == + > . . . . . . . . . . . . . . . . . . . . . . . . . . . .

77

xiv

LIST OF FIGURES
4.13 Visual assessment using ground truth: (a) original image; (b) denoised image. 78
4.14 Visual assessment without ground truth: (a) noisy image; (b) denoised image. 79
4.15 Qualitative measure of filter performance: (a) original image; (b) noisy &A
B image; (c) median filtered; (d) difference between original and denoised,

mid-grey representing zero. . . . . . . . . . . . . . . . . . . . . . . . . . .

81

4.16 Qualitative measure of filter performance: (a) noisy image; (b) median filtered; (c) difference between noisy and denoised, mid-grey representing zero. 82
4.17 The
B
C2 8 bpp square test image. . . . . . . . . . . . . . . . . . . . .

83

4.18 (a) the 


 

8 bpp lenna image; (b) histogram. . . . . . . . . . . . .

83

4.19 (a) the 


 

8 bpp montage image; (b) histogram. . . . . . . . . . . .

84

4.20 Equivalent naming conventions for pixels in the local window. . . . . . . .

85

4.21 (a) original; (b) noisy &'D
 version; (c) denoised using   averaging. . .

86

4.22 The mean is a poor threshold when the clusters have different populations. .

88

4.23 (a) noisy, &E

 ; (b) filtered with 1-segment model; (c) filtered with 2-

segment model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

90

4.24 Comparison of 1-segment and 2-segment models for denoising montage. .

90

4.25 [top to bottom, left to right] A mixture of two normal distributions with
common variance, with their means separated by (a) F& ; (b) F& ; (c) F& ; (d)
1& . 92
4.26 (a) noisy, &G
; ; (b)(f) simple model selection using HI
J;;. . . . . .

93

4.27 (a) noisy, &GI

1 ; (b)(f) simple model selection using HI
J;;. . . . . .

94

4.28 (a) noisy, &G#F ; (b)(f) simple model selection using HI
J;;. . . . . .

95

4.29 (a) noisy, &GB ; (b)(f) simple model selection using HI
J;;. . . . . .

95

4.30 Effect of H

on RMSE when denoising montage by switching between a

1-segment and 2-segment local model. . . . . . . . . . . . . . . . . . . . .

96

LIST OF FIGURES

xv

4.31 Comparison of  values for -test filtering of montage. . . . . . . . . . .

99

4.32 The effective H values used by the -test criterion: KLI and M$-NB . . .

99

4.33 Visual comparison of two model selection functions. Columns are: noisy,
HO$ filtered, and -test filtered images. Rows are: &2 10, 20, 30 and 40. .

101

4.34 RMSE comparison of model selection criteria for montage. . . . . . . . .

102

4.35 Oscillatory root images: (a) original image; (b) after one iteration; (c) after
two iterations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

103

4.36 Iterating 100 times: (a) noiseless square; (b)  local segmentation assuming &G# ; (c)  median filtered. . . . . . . . . . . . . . . . . . . .

104

4.37 Iterating 100 times: (a) noisy &DB square; (b)   local segmentation
assuming &'B ; (c)  median filtered. . . . . . . . . . . . . . . . . . .

105

4.38 Iterating 100 times: (a) noisy &O square; (b)   local segmentation
assuming &'$ ; (c)  median filtered. . . . . . . . . . . . . . . . . . .

105

4.39 Effect of iteration on RMSE for denoising montage. The same value of &
is used for each iteration. . . . . . . . . . . . . . . . . . . . . . . . . . . .

106

4.40 Bottom right hand corner of montage: (a) noisy &P:

F ; (b) after 1 iteration; (c) after 2 iterations; (d)(f) corresponding difference images relative
to the noiseless original. . . . . . . . . . . . . . . . . . . . . . . . . . . .

107

4.41 The effect of a two class model on a three class region: (a) original pixels;
(b) segmented assuming QL
. . . . . . . . . . . . . . . . . . . . . . . . .

108

4.42 Effect of do no harm (DNH) on RMSE when denoising montage. . . .

109

4.43 Effect of do no harm (DNH) on WCAE when denoising montage. . . .

110

4.44 DNH disabled: (a) original image; (b) QLD
, 64%; (c) QI
, 36%. . . . .

110

4.45 DNH enabled: (a) QRO
, 60%; (b) QL#

, 25%; (c) DNH, 15%. . . . . . .

111

4.46 Common windows: (a) 5; (b) 9; (c) 21; (d) 25; (e) 37 pixels. . . . . . . . .

112

xvi

LIST OF FIGURES
4.47 Effect of different windows for denoising montage, DNH disabled. . . . . 113
4.48 Effect of different windows for denoising montage, DNH enabled. . . . . 114
4.49 Each pixel participates in 9 local  windows. . . . . . . . . . . . . . . . 115
4.50 Weighting overlapping estimates: (a) any S ; (b) S*$ ; (c) ST#-4 ; (d) S*D
. 116
4.51 Effect of S on RMSE when denoising montage. . . . . . . . . . . . . . . 117
4.52 Effect of S on WCAE when denoising montage. . . . . . . . . . . . . . . 118
4.53 Denoising square: (a) noisy original, &DU
1 ; (b) output when S9V ;
(c) output when STD
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.54 Effect of confidence weighting on RMSE when denoising montage. . . . 119
4.55 Effect of activity weighting on RMSE when denoising montage. . . . . . 120
4.56 Comparison of three different methods for choosing the initial Q means, in
terms of RMSE denoising performance on montage. . . . . . . . . . . . . 125
4.57 Comparison of binary and multi-class local segmentation for denoising montage. DNH is disabled and ST . . . . . . . . . . . . . . . . . . . . . . . 127
4.58 Comparison of binary and multi-class local segmentation for denoising montage. DNH is disabled and STD
. . . . . . . . . . . . . . . . . . . . . . . 128
4.59 Comparison of  binary and multiclass local segmentation models on
montage. DNH is enabled and S*$ . . . . . . . . . . . . . . . . . . . . 129
4.60 Comparison of  binary and multiclass local segmentation models on
montage. DNH is enabled and S*O
. . . . . . . . . . . . . . . . . . . . 130
4.61 Binary local segmentation of noisy &'I montage. White denotes: (a) DNH;
(b) QLD
; (c) QLI
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4.62 Multi-class local segmentation of noisy &<W montage. White denotes:
(a) DNH; (b)(i) Q
to QL . . . . . . . . . . . . . . . . . . . . . . . 131
4.63 (a) noisy &'I montage; (b)  local standard deviations. . . . . . . . 133

LIST OF FIGURES

xvii

4.64 Histogram of local  standard deviations from Figure 4.63b. . . . . . . .

133

4.65 Smoothed version of main peak from Figure 4.64. . . . . . . . . . . . . . .

134

4.66 Filter masks used by Immerkrs noise variance estimation technique. . . .

134

4.67 (a) original lenna image; (b) after Immerkr structure suppression, with
mid grey representing zero. . . . . . . . . . . . . . . . . . . . . . . . . . .

135

4.68 Different noise estimation algorithms for montage. . . . . . . . . . . . .

136

4.69 Smoothed histogram of local standard deviations for lenna. . . . . . . . .

137

4.70 Different noise estimation algorithms for lenna, corrected for existing noise. 138
4.71 RMSE comparison for montage, true & supplied. . . . . . . . . . . . . .

140

4.72 RMSE comparison for montage, with & estimated. . . . . . . . . . . . .

141

4.73 (a) noisy &XY part of montage; (b) FUELS enhanced difference image,
RMSE=3.48; (c) SUSAN37 enhanced difference image, RMSE=4.11. . . .

142

4.74 WCAE comparison for montage image, with& estimated. . . . . . . . . .

143

4.75 (a) part of montage without added noise; (b) SUSAN37 output using


F
B1 ; (c) difference image. . . . . . . . . . . . . . . . . . . . . . . . . . .

143

4.76 RMSE comparison for lenna, with & estimated. . . . . . . . . . . . . . .

144

4.77 WCAE comparison for lenna image, with & estimated. . . . . . . . . . .

145

4.78 (a) the B

FABF barb2 image; (b) its histogram. . . . . . . . . . . . . .

146

4.79 RMSE comparison for barb2, with & estimated. . . . . . . . . . . . . . .

147

4.80 WCAE comparison for barb2, with & estimated. . . . . . . . . . . . . . .

148

4.81 For barb2 with no added noise: (a)(c) FUELS, SUSAN9 and SUSAN37
denoised output; (d)(f) corresponding enhanced difference images. . . . .

149

4.82 Partial histogram of FUELS and SUSAN denoising errors for barb2. . . .

150

xviii

LIST OF FIGURES

5.1

Scatter plot of 11 data points. . . . . . . . . . . . . . . . . . . . . . . . . . 156

5.2

Three different polynomial fits to the 11 data points. . . . . . . . . . . . . . 156

5.3

The Bayesian MAP estimate may not always coincide with the largest peaked
zone of probability mass. . . . . . . . . . . . . . . . . . . . . . . . . . . . 160

5.4

Local segmentation of a  window into two classes. . . . . . . . . . . . 163

5.5

Raster order convention for labels within a segment map. . . . . . . . . . . 163

5.6

The 16 possible
Z
binary segment maps. . . . . . . . . . . . . . . . . . 164

5.7

The 8 canonical
Z[
binary segment maps. . . . . . . . . . . . . . . . . . 164

5.8

The 7 valid 4-connected


binary segment maps. . . . . . . . . . . . . . 166

5.9

Noisy pixels and two candidate segment maps. . . . . . . . . . . . . . . . 166

5.10 The proposed local segmentation message format. . . . . . . . . . . . . . . 168

5.11 A   window of noisy pixels. . . . . . . . . . . . . . . . . . . . . . . . . 170
5.12 Candidate model 1 and its message length calculation. . . . . . . . . . . . 170
5.13 Candidate model 2 and its message length calculation. . . . . . . . . . . . 170
5.14 Comparing two message lengths. . . . . . . . . . . . . . . . . . . . . . . . 171
5.15 RMSE comparison for denoising montage, true & supplied. . . . . . . . . 172
5.16 (a) noisy &9\ middle section of montage; (b) FUELS model selection;
(c) Pseudo-MML model selection. Black denotes QLD
and white QR
. . 173
5.17 Non-uniform prior: RMSE comparison for denoising montage, true & supplied. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.18 Model selection comparison when &] , black denotes QP^
and white
QL#
: (a) FUELS; (b) Pseudo-MML; (c) Pseudo-MML with a non-uniform

prior. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.19 Pixels are encoded relative to their segment mean. . . . . . . . . . . . . . . 176

LIST OF FIGURES

xix

5.20 Different encodings for a segment means when &G# and _
BF . . . . .

177

5.21 RMSE comparison for denoising montage, true & supplied. . . . . . . . .

178

5.22 Model order selection when &Y` , black denotes Q

, white Q#a
:

(a) Pseudo-MML; (b) MML; (c) difference: white shows where MML chose
QLI
over Pseudo-MMLs QR
. . . . . . . . . . . . . . . . . . . . . . .

179

5.23 Posterior blending of the two example models from Section 5.4.3. . . . . .

180

5.24 RMSE comparison for denoising montage, with true & supplied. . . . . .

181

5.25 RMSE comparison for denoising montage, true & supplied. . . . . . . . .

182

5.26 Incorporating DNH: (a) null model encodes the data as is; (b) standard model. 183
5.27 FUELS-style DNH usage when denoising montage, true & supplied. . . .

184

5.28 RMSE comparison for denoising montage, true & supplied. . . . . . . . .

185

5.29 WCAE comparison for denoising montage, true & supplied. . . . . . . . .

186

5.30 How often DNH is invoked for montage, true & supplied. . . . . . . . . .

187

5.31 RMSE comparison for denoising montage, true & supplied. . . . . . . . .

188

5.32 RMSE comparison for denoising montage, true & supplied. . . . . . . . .

190

5.33 RMSE comparison for denoising montage, with & estimated. . . . . . . .

191

5.34 Evolution of the average two-part message length for the MML denoising
algorithm, with & estimated from the noisy montage image. . . . . . . . .

192

5.35 As the amount of data increases, the model part becomes less significant. . .

193

5.36 Pictorial representation of the iterative approach to image understanding. . .

194

5.37 Potential segment maps: (a) popular Qb^
; (bc) common edge patterns;
(d) would rarely occur. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

195

5.38 Learning c0d;)feg/ : RMSE comparison for montage, with & estimated. . . .

197

5.39 The 15 most popular (canonical) segment maps for montage when &'# .

197

xx

LIST OF FIGURES
5.40 Learning c0d;)fhjiGk/ : RMSE comparison for montage, with & estimated. . 198
5.41 Comparison of DNH invocation for montage, with & estimated. . . . . . . 199
5.42 More models: RMSE comparison for denoising montage, with & estimated. 201
5.43 Proportion of times a QL#
model was most probable in montage. . . . . 202
5.44 RMSE comparison for denoising montage, with & estimated. . . . . . . . 203
5.45 Proportion of times that QLI
was deemed best when denoising montage.

204

5.46 Comparison of algorithm running times for denoising montage. . . . . . . 205

5.47 RMSE comparison for denoising lenna, with & estimated. . . . . . . . . . 206
5.48 Enhanced difference images for lenna when &Y

B : (a) ground truth;

(b) FUELS; (c) MML-256/2; (d) MML-2; (e) SUSAN9. . . . . . . . . . . . 206

5.49 WCAE comparison for denoising lenna. . . . . . . . . . . . . . . . . . . 207
5.50 RMSE comparison for denoising barb2, with & estimated. . . . . . . . . . 208
5.51 Enhanced difference images for barb2 when &Y

B : (a) ground truth;

(b) FUELS; (c) MML-256/2; (d) MML-2; (e) SUSAN9. . . . . . . . . . . . 208

5.52 WCAE comparison for denoising barb2. . . . . . . . . . . . . . . . . . . 209
5.53 The
BF
B1 8 bit camera image and its histogram. . . . . . . . . . . . . 210
5.54 RMSE comparison for camera image, with & estimated. . . . . . . . . . . 210
5.55 Enhanced difference images for camera when &^
 : (a) ground truth;
(b) FUELS; (c) MML-256/2; (d) MML-2; (e) SUSAN9. . . . . . . . . . . . 211
6.1

Local segmentation decomposition as a core image processing component. . 216

6.2

(a) clean image; (b) with lAW 4 impulse noise; (c) multi-class FUELS,
MASS=2; (d) MASS=3; (e) median; (f) weighted median. . . . . . . . . . 219

6.3

(a) clean image; (b) with lX`-m
impulse noise; (c) multi-class FUELS,
MASS=2; (d) MASS=3; (e) median; (f) weighted median. . . . . . . . . . 221

LIST OF FIGURES

xxi

6.4

Computing the parameters for a homogeneous   planar model. . . . . .

225

6.5

An L22 volume image. . . . . . . . . . . . . . . . . . . . . . . . . .

226

6.6

Each voxel (dotted) has 6 neighbours in its immediate neighbourhood. . . .

227

6.7

(a) original lenna; (b) classification into smooth (black) and shaped (white)
feature points. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

6.8

(a) original lenna; (b) smooth points (white); (c) shaped points (white);
(d) textured points (white). . . . . . . . . . . . . . . . . . . . . . . . . . .

6.9

229

231

(a) original lenna, estimated &P:

4F ; (b) local segmentation; (c) Sobel;
(d) SUSAN using onp1&qr . . . . . . . . . . . . . . . . . . . . . . . .

233

6.10 The twelve inter-pixel boundaries for a   window, shown in bold. . . . .

234

6.11 Examples of segment maps, with boundaries shown in bold. . . . . . . . .

235

6.12 Example of probabilistic pixel boundary detection: (a) part of lenna; (b) s ;
(c) t ; (d) )fsvubtw/yxB
; (e) z )fs|{wuPt}{./~xB
; (f) R)fsJ,~t/ . . . . . . . . . . . .

236

6.13 Probabilistic edge strength for lenna using R)pt,.s}/ . . . . . . . . . . .

238

6.14 Doubling an images size means predicting all the ? pixels. . . . . . . . .

239

6.15 Pixel replication is the simplest expansion method. . . . . . . . . . . . . .

239

6.16 Using local segmentation for structure directed interpolation. . . . . . . . .

240

6.17 Comparison of doubling methods: (a) denoised image; (b) pixel replication;
(c) linear interpolation; (d) MML-256 based enlargement. . . . . . . . . . .

241

6.18 Lossy compression can be considered a two stage process. . . . . . . . . .

242

6.19 Standard BTC: (a) original image at 8 bpp; (b) reconstructed image at 2 bpp;
(c) bitmap. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

243

6.20 Adaptive BTC: (a) original image; (b) bitmap; (c) reconstructed image at
1.39 bpp, 41% homogeneous blocks using threshold of 8.5; (d) reconstructed
image at 2 bpp. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

244

6.21 The causal local neighbourhood consists of pixels known to both encoder
and decoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

245

List of Tables
4.1

Confidence intervals for normally distributed data. . . . . . . . . . . . . . .

70

4.2

Proportion of pixels at the margins for various mask and image sizes . . . .

77

4.3

Using hypothesis testing for model selection in local segmentation. . . . . .

97

4.4

Various values of

used in -testing. . . . . . . . . . . . . . . . . . . . .

98

4.5

Binary and multi-class model usage for noisy &'# montage. . . . . . .

131

5.1

Interchangeable use of probability and code length . . . . . . . . . . . . .

158

5.2

Number of possible segmentations for various window configurations. . . .

166

6.1

RMSE results for filtering impulse noise from montage. . . . . . . . . . .

220

xxiii

Listings
3.1

The k -means algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . .

30

3.2

The Q -means algorithm. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

31

4.1

An iterative method for choosing a binary threshold. . . . . . . . . . . . .

88

4.2

The Q -means algorithm for local segmentation. . . . . . . . . . . . . . . .

122

5.1

Algorithm to generate canonical segmentations only. . . . . . . . . . . . .

165

xxv

Abstract
A unifying philosophy for carrying out low level image processing called local segmentation is presented. Local segmentation provides a way to examine and understand existing
algorithms, as well as a paradigm for creating new ones. Local segmentation may be applied
to range of important image processing tasks. Using a traditional segmentation technique
in intensity thresholding and a simple model selection criterion, the new FUELS denoising
algorithm is shown to be highly competitive with state-of-the-art algorithms on a range of
images. In an effort to improve the local segmentation, the minimum message length information theoretic criterion for model selection (MML) is used to select between models
having different structure and complexity. This leads to further improvements in denoising
performance. Both FUELS and the MML variants thereof require no special user supplied
parameters, but instead learn from the image itself. It is believed that image processing in
general could benefit greatly from the application of the local segmentation methodology.

xxvii

Declaration
This thesis contains no material that has been accepted for the award of any other degree or
diploma in any university or other institution. Furthermore, to the best of my knowledge, this
thesis contains no material previously published or written by another person, except where
due reference is made in the text of the thesis.

Torsten Seemann
January 28, 2003

xxix

Acknowledgments
I love deadlines. I love the whooshing sound they make as they fly by.
Douglas Adams, 19522001
I wish to acknowledge the following people for their role in the completion of this thesis:

My supervisor, Peter Tischer, for his guidance, constructive proofreading, and many
fruitful discussions, some of which even stayed on topic. Thank you for helping me
produce a thesis I can be proud of.

David Powell, for many productive white-board exchanges over 6 years of sharing an
office, and for actually understanding most of my thesis after a single proofreading.

My parents, Sandra and Gunther Seemann, for always supporting my choices in life,
and only occasionally nagging me to finish up and get a real job that pays money.

My partner, Naomi Maguire, for her devoted love, support and encouragement. What
else could you ask for?

Naomis parents, Brenda and Bill Maguire, for accepting an unemployed doctoral student bum into their family.

Bernd Meyer, for a handful of valuable thesis related discussions we had when he
occasionally blessed our office with a visit.

The staff in the School of Computer Science and Software Engineering at Monash University. In particular, Trevor Dix, for regularly checking in on me, and commiserating
with the plight of the doctoral student.

This thesis was typeset using LATEX 2 [KD95]. Graphs were generated by gnuplot [WKL ],
and diagrams were constructed using XFig [SKYS].
xxxi

Chapter 1
Introduction
Image processing is a rapidly growing area of computer science. Its growth has been fueled
by technological advances in digital imaging, computer processors and mass storage devices.
Fields which traditionally used analog imaging are now switching to digital systems, for their
flexibility and affordability. Important examples are medicine, film and video production,
photography, remote sensing, and security monitoring. These and other sources produce
huge volumes of digital image data every day, more than could ever be examined manually.
Digital image processing is concerned primarily with extracting useful information from
images. Ideally, this is done by computers, with little or no human intervention. Image processing algorithms may be placed at three levels. At the lowest level are those techniques
which deal directly with the raw, possibly noisy pixel values, with denoising and edge detection being good examples. In the middle are algorithms which utilise low level results for
further means, such as segmentation and edge linking. At the highest level are those methods which attempt to extract semantic meaning from the information provided by the lower
levels, for example, handwriting recognition.
The literature abounds with algorithms for achieving various image processing tasks. However, there does not appear to be any unifying principle guiding many of them. Some are
one dimensional signal processing techniques which have been extended to two dimensions.
Others apply methods from alternative disciplines to image data in a somewhat inappropriate
manner. Many are the same basic algorithm with parameter values tweaked to suit the problem at hand. Alternatively, the parameters are optimized with respect to a suitable training
1

Chapter 1. Introduction

set, without thought on how to vary them for images with different properties. There do
exist well considered methods, but unfortunately a large proportion of new ideas have been
ad hoc, without any central guiding principle.
This thesis proposes a unified approach to low level image processing called local segmentation. The local segmentation principle states that the first step in processing a pixel should
be to segment explicitly the local region encompassing it. On a local scale, this has the effect
of making clear which pixels belong together, and which pixels do not. The segmentation
process results in a local approximation of the underlying image, effectively separating the
signal from the noise. Thus higher level algorithms can operate directly on the signal without risk of amplifying the noise. Local segmentation can be seen as providing a common
framework for constructing image processing algorithms.
Many existing image processing algorithms already make partial use of the local segmentation concept. It is possible to examine these algorithms with respect to the local segmentation
model they use. This helps to make their strengths and weaknesses more apparent. Even popular techniques, such as linear and rank filters, can be framed in terms of their application
of local segmentation. In most cases the segmentation is implicit rather than explicit. That
is, the choice of which pixels belong together is performed in a systematic, but sometimes
roundabout manner. The SUSAN image processing system, developed by Smith and Brady,
originally considered using explicit segmentation, but eventually chose to allow all pixels to
have partial membership in the centre pixels segment.
Image denoising is particularly suited to demonstrating the utility of local segmentation.
Denoising is the process of removing unwanted noise from an image. A denoised image is an
approximation to the underlying true image, before it was contaminated. A good denoising
algorithm must simultaneously preserve structure and remove noise. Obviously, to do this the
algorithm must be able to identify what structure is present. Local segmentation specifically
attempts to separate structure from noise on a local scale. Denoising would therefore be a
good application with which to test different approaches to local segmentation.
Local regions only contain a small number of pixels. It is unlikely that there would be more
than a few segments present at such a scale, so unconnected, homogeneous groups of pixels
are likely to part of the same global segment. Traditional threshold-based segmentation

3
techniques, such as Q -means, perform well with this sort of data. This technique is used
to develop FUELS, an algorithm for denoising greyscale images affected by additive noise.
FUELS requires only one parameter, the noise variance, which can be supplied by the user
or estimated automatically from the image. The noise variance and pixel values are used by
a model order selection criterion for deciding the optimal number of segments present in the
local region.
FUELS has two additional original features. If the optimal segmentation is deemed unfit,
the noisy pixels are passed through unmodified. This is called the do no harm principle.
Also, local segmentation produces multiple overlapping estimates for the value of each pixel
in the image. FUELS is able to combine these estimates to further improve its denoising
performance. It will be shown that FUELS is competitive with state-of-the-art denoising
algorithms over a range of noise levels.
Although FUELS uses a simple segmentation algorithm, it still achieves good results. It does,
however, have some limitations. The nature of FUELS model selection criterion makes it
unable to distinguish segments having a contrast difference less than a specified multiple of
the noise variance. The criterion is also not flexible enough to allow FUELS to compare
different models having the same number of segments. This is because the thresholding
algorithm does not incorporate any spatial information.
Local segmentation may be considered an inductive inference problem. We wish to infer
the underlying image structure given only the noisy pixel values. Bayesian and information
theoretic techniques are recognised as some of the most powerful for these types of problems.
One such technique is Wallaces minimum message length information theoretic criterion
(MML), which has been developed continually since 1968. MML is essentially a Bayesian
method, but provides a sound way to choose a point estimate from the posterior distribution
over models. It works well with small amounts of data, and is invariant under non-linear
transformations of the parameter space.
Under MML, a candidate model is evaluated using its message length. A message is an
efficient lossless encoding of the data, and consists of two parts. The first part encodes
the model and its parameters. The second part encodes the data given the model from the
first part. The candidate model with the shortest overall message length is deemed the best

Chapter 1. Introduction

model for the data. This is similar, but not identical, to the maximum a posteriori (MAP)
estimate used in Bayesian analysis. When the model contains continuous parameters, MML
optimally quantizes the prior density such that the MAP estimate is invariant under non-linear
transformations of the data and model parameters.
The model selection criterion used by FUELS is replaced by a more flexible MML one.
MML makes it straightforward to include extra models in the pool of candidate models
being considered. For a   window, there exists 256 unique ways to divide the 9 pixels
into one or two segments. The MML version of FUELS is modified to evaluate and consider
all of these. This allows spatial information to be exploited, and overcomes the minimum
contrast difference that FUELS requires between segments. The do no harm principle is
re-interpreted, and shown to fit naturally within the MML framework.
Using MML for model selection is shown to improve results relative to FUELS, and to outperform other good denoising algorithms. It produces better local approximations, especially
for very noisy images. Evaluating and comparing large numbers of models using MML,
however, is much more computationally intensive than FUELS. In terms of RMSE, the improvements are not large. This indicates that simpler techniques like FUELS and SUSAN are
already good trade-offs between efficiency and effectiveness. In critical applications where
the best possible modeling is required, the use of MML methods could be warranted.
The local segmentation paradigm is not limited to denoising applications. It is shown that
a variety of image processing tasks may be addressed using local segmentation. Breaking
an image into its low level structural components could be considered an essential first step,
from which many other tasks are derived. It is shown how local segmentation may be used
for edge detection, pixel classification, image enlargement and image compression. The
extension to different noise models, such as impulse noise, image models, such as planar
segments, and higher dimensional data, such as volume images, is also discussed.
The FUELS local segmentation algorithm has the desirable feature of being simple, while
still producing good results. This makes it well suited to robotic and computer vision applications. It can be implemented using low amounts of memory and processing power, ideal
for putting into hardware or embedded microcontrollers. Local segmentation is inherently
parallelizable, because each pixels local region is processed independently. Thus a highly

5
concurrent implementation would be possible. This could be useful in real time applications
where many images per second need to be analysed.
I believe that local segmentation provides a unifying philosophy for carrying out low level
image processing. It provides a way to examine and understand existing algorithms, as well
as a paradigm for creating new ones. A local segmentation analysis of an image can be
re-used by a wide range of image processing tasks. Using a traditional segmentation technique in intensity thresholding and a simple model selection criterion, the FUELS denoising
algorithm is shown to be highly competitive with state-of-the-art algorithms on a range of
images. In an effort to improve the local segmentation, MML is applied to select between
a larger set of models having different structure and complexity. This leads to further improvements in denoising performance. Both FUELS and the MML variants thereof require
no special user supplied parameters, but instead learn from the noisy image itself. I believe
that image processing in general could benefit greatly from the application of the techniques
proposed in this thesis.

Chapter 2
Notation and Terminology
2.1 Digital images
A digital image is a discrete two-dimensional function, w)p|,-/ , which has been quantized
over its domain and range [GN98]. Without loss of generality, it will be assumed that the
image is rectangular, consisting of rows and

columns. The resolution of such an image

is written as < . By convention, w)f ,./ is taken to be the top left corner of the image,
and w)pW
B,.I
/ the bottom right corner. This is summarized in Figure 2.1.

f(0,0)

f(X1,Y1)
y
Figure 2.1: A rectangular digital image of resolution rL .

Chapter 2. Notation and Terminology

Each distinct coordinate in an image is called a pixel, which is short for picture element.
The nature of the output of w)p|,-/ for each pixel is dependent on the type of image. Most
images are the result of measuring a specific physical phenomenon, such as light, heat, distance, or energy. The measurement could take any numerical form.
A greyscale image measures light intensity only. Each pixel is a scalar proportional to the
brightness. The minimum brightness is called black, and the maximum brightness is called
white. A typical example is given in Figure 2.2. A colour image measures the intensity and
chrominance of light. Each colour pixel is a vector of colour components. Common colour
spaces are RGB (red, green and blue), HSV (hue, saturation, value), and CMYK (cyan,
magenta, yellow, black), which is used in the printing industry [GW92]. Pixels in a range
image measure the depth of distance to an object in the scene. Range data is commonly used
in machine vision applications [KS00].

Figure 2.2: A typical greyscale image of resolution BB .

For storage purposes, pixel values need to be quantized. The brightness in greyscale images
is usually quantized to _

levels, so )p|,-/C<-,;
B,@,_$#
F . If _ has the form

B , the

image is referred to as having bits per pixel. Many common greyscale images use 8 bits
per pixel, giving 256 distinct grey levels. This is a rough bound on the number of different
intensities the human visual system is able to discern [Jah93]. For the same reasons, each
component in a colour pixel is usually stored using 8 bits.

2.2 Image statistics

Medical scans often use 1216 bits per pixel, because their accuracy could be critically
important. Those images to be processed predominantly by machine may often use higher
values of _

to avoid loss of accuracy throughout processing. Images not encoding visible

light intensity, such as range data, may also require a larger value of _

to store sufficient

distance information.
There are many other types of pixels. Some measure bands of the electromagnetic spectrum
such as infra-red or radio, or heat, in the case of thermal images. Volume images are actually
three-dimensional images, with each pixel being called a voxel. In some cases, volume
images may be treated as adjacent two-dimensional image slices. Although this thesis deals
with greyscale images, it is often straightforward to extend the methods to function with
different types of images.

2.2 Image statistics

2.2.1 The histogram
A histogram plots the relative frequency of each pixel value that occurs in a greyscale image.
Figure 2.3 shows the intensity histogram for the image from Figure 2.2. The histogram
provides a convenient summary of the intensities in an image, but is unable to convey any
information regarding spatial relationships between pixels. In this example, the image does
not contain many very low or very high intensity pixels. It is possible that peaks in the
histogram correspond to objects in the image, but it is difficult to be certain without visually
examining the image.

2.2.2 The mean

The image mean is the average pixel value of an image. For a greyscale image this is equal to
the average brightness or intensity. Let the image w)p|,~/ be referred to using the shorthand
= . The mean of this image, y= , may be calculated using Equation 2.1.

10

Chapter 2. Notation and Terminology

0.012

Relative frequency

0.01

0.008

0.006

0.004

0.002

50

100

150

200

250

Intensity

Figure 2.3: The histogram for the greyscale image in Figure 2.2.

y= 

w)p|,~/

(2.1)

2.2.3 The variance

The image variance, Fd;y= , gives an estimate of the spread of pixel values around the image
mean. It can be calculated using either Equation 2.2 or Equation 2.5. The latter has the
advantage of requiring only one pass through the image. The standard deviation is simply
z

Fd;y= .

Fd=|


=Py= ; {


C

(2.2)
)fw)p|,-/}Py= / {

(2.3)

2.3 Image algebra

11

1d;y= 


C

 

y= {
by= {




C

)p|,-/}

.4

w)|,-/ { O

(2.4)

w) ? , ? /

(2.5)
{
w)|,/

(2.6)

2.2.4 The entropy

The image histogram may be considered a probability distribution over pixel values. For
the case of a _ level greyscale image, the histogram entry for intensity 5 may be written as
c0d)+5/ . The entropy of an image, = , is given by Equation 2.7. The units of entropy are bits

when using logarithms to base 2.

H )+=/DA

cJd;)f5/-B

The entropy has a maximum value of B

{

c0d)+5/

(2.7)

when all intensities occur will equal frequency,

corresponding to a uniform histogram. It has a minimum value of 0 when all pixels have the
same intensity. The entropy is one measure of the information content of an image. Because
it is calculated from the histogram, it is unable to take spatial factors into consideration.

2.3 Image algebra

2.3.1 Image-scalar operations
Various useful arithmetical operations may be defined for images. Let represent the binary
operator for addition, subtraction, multiplication, or division. Equation 2.8 shows how to
combine a scalar, ! , and an image, , to produce a new image, = . This is a pixel-wise
operation each pixel in is operated on using with ! , and the result put in = .

12

Chapter 2. Notation and Terminology

=L9!

w)p|,-/

w)p|,-/w$)|,/Z!

(2.8)

This idea could be used to enhance an image which is too dark. Consider the image in
Figure 2.4a which uses 8 bits per pixel (256 levels), but only contains pixels with intensities
from 64 to 191. One may consider enhancing it to use the full intensity range. This can
be achieved using Equation 2.9, where nq denotes integer truncation, and floating point
precision is used for all pixels during the calculation. The result is given in Figure 2.4b.

=V

jbF


F


BF

(2.9)

Figure 2.4: (a) low contrast image; (b) after enhancement.

2.3.2 Image-image operations

The image-scalar operation may be extended to the combination of two images, and ,
having the same resolution. Instead of combining a scalar with each pixel, two pixels with the
same coordinate in different images are used instead. Equation 2.10 describes this process.

=09

w)|,/w)|,/$ )p|,-/9 { )p|,-/

(2.10)

2.4 Image acquisition

13

Imagine wanting to generate a blended version of two greyscale images of identical resolution. This could be achieved using Equation 2.11, where  determines the mixing proportion.
Alpha blending is a simple form of morphing, and is often used to dissolve between scenes
in film and television. A visual example for A#-4 is given in Figure 2.5.

(2.11)

=on+20u#)
"b/2@q

Figure 2.5: Alpha-blending example: (a) first image; (b) second image; (c) blended image
using vB .

2.4 Image acquisition

Image acquisition is the process of obtaining a digitized image from a real world source.
Each step in the acquisition process may introduce random changes into the values of pixels
in the image. These changes are called noise. Assume you want to send a photo of your
new house to a friend over the Internet. This may be achieved by taking a photograph with a
conventional camera, having the film made into a print, scanning the print into a computer,
and finally emailing it to your friend. Figure 2.6 shows the many potential sources of noise.

Actual house

Light

Atmospheric noise

House

Camera

Film grain noise

Film

Scanner

Digital image

CCD and stepper

motor noise

Email

Digital image

Network transmission
errors

Figure 2.6: Noise may be introduced at each step in the acquisition process.

14

Chapter 2. Notation and Terminology

The air between the photographer and the house may contain dust particles which interfere
with the light reaching the camera lens. The silver-halide crystals on the film vary in size and
are discontinuous, resulting in film grain noise in the printing process [MJ66]. Most scanners
use a CCD array to scan a row of the print, which may introduce photo-electronic noise. The
scanners CCD array is controlled by a fine stepper motor. This motor has some degree of
vibration and error in its movement, which may cause pixels to be mis-aligned. The scanner
also quantizes the CCD signal, introducing quantization noise [GN98]. Transmitting the
image over the Internet is nearly always a bit preserving operation thanks to error checking
in network protocols. However, an image transmitted to Earth from a remote space probe
launched in the 1970s is almost guaranteed to contain errors.

2.5 Types of noise

The previous example illustrated the manner in which an image may be affected by noise
during the acquisition process. The properties of the noise introduced at each capture step
are likely to vary. However, there are three standard noise models which model well the types
of noise encountered in most images: additive, multiplicative, and impulse noise. Figure 2.7
shows how these types of noise affect a typical greyscale image.

2.5.1 Additive noise

Let ?)p|,-/ be the noisy digitized version of the ideal image, w)p|,~/ , and 0)p|,~/ be a noise
function which returns random values coming from an arbitrary distribution. Then additive
noise can be described by Equation 2.12.

? )|,-/0#w)|,/uX0)|,-/

(2.12)

Additive noise is independent of the pixel values in the original image. Typically, w)p|,-/
is symmetric about zero. This has the effect of not altering the average brightness of the
image, or large parts thereof. Additive noise is a good model for the thermal noise within
photo-electronic sensors [Pit95].

2.5 Types of noise

15

Figure 2.7: Different types of noise: (a) original image; (b) additive noise; (c) multiplicative
noise; (d) impulse noise.

2.5.2 Multiplicative noise

Multiplicative noise, or speckle noise, is a signal dependent form of noise whose magnitude
is related to the value of the original pixel [KSSC87]. Equation 2.13 describes one simple
form it can take, but a more complex function of the original pixel value is also possible.
Multiplicative noise is an approximation to the noise encountered in images recorded on
film slides [Jai89] and from synthetic aperture radar [Lee81a, Cur91].

? )|,/0w)p|,~/ub0)|,/yw)p|,~/w#w)p|,-/@
uXw)p|,-/

(2.13)

16

Chapter 2. Notation and Terminology

2.5.3 Impulse noise

Impulse noise has the property of either leaving a pixel unmodified with probability
M ,
or replacing it altogether with probability . This is shown in Equation 2.14. Restricting
0)|,-/ to producing only the extreme intensities  or _AM
results in salt-pepper noise. The

source of impulse noise is usually the result of an error in transmission or an atmospheric or

man-made disturbance [PV90].

? )p|,-/0

0)|,-/

"mTdyBF-m

w)|,/

"mTdyBF-m

"2

(2.14)

2.5.4 Quantization noise

Quantization noise is due to the quantization of pixel values during the analog to digital
conversion. For example, imagine an analog image with brightness values ranging from 0
to 10. If it is quantized to accuracy -
, the digitized image will have 101 distinct grey
levels. A given intensity, 5 , could have originally been anywhere in the range 5"b-4B . This
uncertainty in the true value of 5 is called quantization noise [STM97].

2.5.5 The noise function

Usually the properties of the noise function, 0)|,-/ , do not vary with and . A spatially
invariant stochastic process is referred to as being stationary. The noise function could
theoretically take any form, but many standard probability distributions have been found
useful. For additive noise, the Gaussian and Laplacian distributions are often used [JRvdH93,
AG98]. The standard case of impulse noise uses a uniform distribution on -,_
.
The most common noise model used in this thesis is an additive zero-mean Gaussian of
unknown variance, independently and identically distributed for each pixel. The application
to alternative noise models is also considered. Some algorithms developed for additive noise
can be adapted to multiplicative noise by logarithmically transforming ?p)|,-/ , applying the
algorithm, and then applying the inverse transform [MPA00].

2.6 Segmentation

17

2.6 Segmentation
Segmentation involves partitioning an image into groups of pixels which are homogeneous
with respect to some criterion. Different groups must not intersect each other and adjacent
groups must be heterogeneous [PP93]. The groups are called segments. Figure 2.8a shows a
noisy image containing three objects on a background. The result of segmentation is given
in Figure 2.8b. Four segments were discovered, and are shown with a dashed outline.

Figure 2.8: (a) original image; (b) segmented into four segments.

The homogeneity criterion used for segmenting Figure 2.8 was based only on the similarity
of pixel intensities. For images containing large amounts of noise or fine structure, this
criterion may be insufficient for successful segmentation. In those cases, some information
regarding the spatial relationship between pixels is required. In particular, the assumption
that pixels belonging to the same segment are expected to be spatially connected is exploited.
A pixel only has eight immediate neighbours. If symmetry is required, there exists only
three forms of pixel connectedness which make sense, shown in Figure 2.9. There is one
type of 8-connectedness, and two types of 4-connectedness. Type I is more popular than
Type II due to the improved proximity of the four neighbouring pixels. In this thesis, only
8-connectedness and Type I 4-connectedness are considered.
Most digital images exist on a rectangular grid. This is primarily due to the arrangement of
image sensors on camera and scanning equipment. Research has been done on the superior
properties of hexagonal grids [Sta99], but they are unlikely to displace the square grid in the
near future. In this thesis we deal only with images sampled on a square grid.

18

Chapter 2. Notation and Terminology

Figure 2.9: Pixel connectedness: (a) 4-connected [Type I]; (b) 4-connected [Type II]; (c) 8connected.

2.7 Local windows

Most of the algorithms discussed in this thesis operate in an adaptive local manner, processing each pixel in turn. Only a small number of neighbouring pixels are included in any
calculation. Local processing can produce simpler algorithms with lower algorithmic time
and space complexity. They may be parallelizable, theoretically to one computer processor
per pixel. There is also evidence to suggest that the human visual system functions in a
parallel local manner, combining many local interpretations into an overall image [BBB00].
The union of the pixel being processed and its neighbouring pixels may be collectively referred to as a window, a mask, or the local region surrounding the pixel. Local windows
typically involve fewer than 50 pixels, on images with up to
 pixels. The only unbiased
window configuration is an isotropic one symmetrical and centered on the pixel to be processed. A circular window meets this requirement, but because pixels reside on a rectangular
grid, some pixels would cover an area only partially within the circle. Figure 2.10 shows five
local windows which are commonly used in image processing. Each is an approximation to
a circular window, with the square windows being simplest to implement.

Figure 2.10: Common local neighbourhood configurations: 5, 9, 13, 21, 25 pixels.

Chapter 3
Local Segmentation in Image Processing
3.1 Introduction
This thesis proposes the use of local segmentation as an effective way to achieve a variety of
low level image processing tasks. The local segmentation principle states that the first step in
processing a pixel should be to segment the local region encompassing that pixel. This provides a snapshot of the local structural features of the image, with the signal clearly separated
from the noise. It is hoped that the identified structural information could be used to implement many image processing tasks including, but not limited to, image denoising [Mas85],
pixel classification [Cho99], edge detection [HSSB98], and pixel interpolation [AW96].
Local segmentation can be seen to belong to a continuum of approaches to image understanding, as shown in Figure 3.1. At the lowest level is local segmentation which operates
in a purely local manner using only a small number of pixels. At a higher level is global
segmentation which attempts to group together related pixels from throughout the image.
The highest level is object recognition, whereby global segments are combined into logical
units representing real world objects of interest.
The fundamental component of the local segmentation approach is the segmentation algorithm itself. Most segmentation algorithms are designed to operate upon a whole image, or a
large portion thereof. Local segmentation can only utilise a small number of pixels belonging to fragments of larger segments. Thus a local segmentation algorithm differs in that it
19

20

Chapter 3. Local Segmentation in Image Processing

Object Recognition

Global Segmentation

Local Segmentation

Groups of segments

Metaglobal

Large groups of pixels / whole image

Small groups of pixels

Global

Local

Figure 3.1: An image processing hierarchy.

has less data and less context to work with. In Section 3.2 the suitability of applying global
segmentation algorithms to local segmentation will be examined.
Some image processing techniques can be seen or interpreted as exploiting the principle of
local segmentation in some way. In most cases the local segmentation principle is not stated
explicitly, nor used as a guiding principle for developing related algorithms. One example is the lossy image compression technique Block Truncation Coding, or BTC [FNK94].
BTC uses simple thresholding to segment    blocks of pixels into two classes, but it was
many years before alternative segmentation algorithms were considered. Further instances
of existing local segmentation-based algorithms will be explored in Section 3.3.
Those image processing tasks suited to local segmentation are often the first to encounter the
raw image data. This data is usually contaminated with one or more forms of noise. The
fundamental attribute of a local segmentation based algorithm should be to preserve as much
image structure (useful information) as possible, and to suppress or remove as much noise
(useless information) as possible. These goals are complementary and inseparable the
ability to identify structure implies the ability to identify noise, and vice versa.
Good image denoising algorithms specialize in extracting structure from noisy images. This
application is probably the most appropriate low level technique for demonstrating local
segmentation. In Section 3.4, the extent to which existing denoising algorithms utilise the
principles espoused in this thesis will be explored. It is shown that the trend in state-of-the-art
denoising algorithms has been toward a local segmentation perspective.
Upon reading this chapter, certain themes will become apparent. Image processing is an
enormous field consisting of many different algorithms. Within a specific field it is often

3.2 Global segmentation

21

difficult to compare results. This is due to the rarity of objective criteria for comparison, a
lack of standard test data, and simply the widely differing goals and needs of each system.
Some techniques are ad hoc in their approach, often having been inappropriately adapted
from other fields without thought to the validity of the underlying assumptions. Others have
one or more tunable parameters which, although data dependent, must be supplied by the
user, rather than learned automatically from the image itself. It is hoped that the work in this
thesis will go some way to improving this situation.

3.2 Global segmentation

Segmentation involves partitioning an image into groups of pixels which are homogeneous
with respect to some predicate. Each group is called a segment. Different segments must
not intersect and adjacent segments must be heterogeneous [PP93]. Pixels within a segment need not necessarily be spatially connected. Clustering is usually used to refer to
segmentation techniques using a homogeneity predicate which does not consider spatial information [HS85]. The segments produced by a clustering algorithm are sometimes called
clusters, but they are still legitimate segments.
Global segmentation is concerned with segmenting a whole image. Local segmentation deals
with segmenting sub-images which are small windows on a whole image. Although a subimage is still a valid image, it is also a fragment of a larger scene being processed in isolation.
A side-effect of this is shown in Figure 3.2. The image consists of two global segments: a
light cross on a dark background. Each of the segments is homogeneous and fully connected.
The  sub-image is also a cross on a background, but its background consists of four
spatially disjoint pixels. Without a larger context it is difficult to ascertain whether they
should be treated as one or four segments. A clustering algorithm, guided only by pixel
intensities, would group them into a single segment.
The number of pixels available to local segmentation is much lower than what most global
segmentation algorithms would expect. This has an effect on the typical number of distinct
segments expected to be encountered. Figure 3.3a is a typical greyscale image called lenna.
Figure 3.3b shows an image which plots, for each pixel from lenna, the standard deviation

22

Chapter 3. Local Segmentation in Image Processing

Figure 3.2: A } sub-image taken from the centre of a } image with two segments.

of the  sub-image centered on that pixel. The standard deviations range from 0 to 72,
and are represented here using intensities from black to white.

Figure 3.3: (a) original image; (b) local standard deviations, with black representing 0
and white 72.

Homogeneous regions in the original image produce dark areas in the standard deviation
image, while edges are responsible for lighter areas. The majority of Figure 3.3b is dark.
Figure 3.4 shows a histogram of the standard deviations, which is expected to be unimodal
and skewed to the right [RLU99]. The peak at 2.4 corresponds roughly to the natural level of
variation within homogeneous regions. The skew toward larger standard deviations is caused
by heterogeneous edge and texture regions varying above the natural level.

3.2 Global segmentation

23

0.08
0.07

Relative frequency

0.06
0.05
0.04
0.03
0.02
0.01
0

10

15

20

25

30

Local standard deviation

Figure 3.4: Histogram of local standard deviations shown in Figure 3.3b.

The area under the histogram near the natural level of variation is much higher than in the
skew. This suggests that it is highly likely that a randomly selected  sub-image will be
homogeneous, consisting of a single segment. This is in stark contrast to global segmentation, where the diagnosis or even consideration of a single segment result is rare. A local
segmentation algorithm must therefore be able to determine automatically the number of
segments present in the sub-image. Fortunately, this task is made easier because the number
of segments is likely to be small.
A small sub-image implies small segments. Global segmentation algorithms, in an attempt to
reduce over-segmentation, often disallow segments containing fewer than a specified number
of pixels. A local segmentation algorithm should expect to diagnose many small segments,
including those consisting of a single pixel. Figure 3.5 shows a alternate  sub-image
taken from a whole image. The one pixel segment may be considered noise by global segmentation, but local segmentation must be more lenient and allow for the fact that a lone
pixel may be part of a larger global segment.
Global segmentation deals mostly with segments consisting of a relatively large number of
pixels. This makes estimated parameter values for global segments naturally more robust.

24

Chapter 3. Local Segmentation in Image Processing

Figure 3.5: A sub-image taken from the right of the image with two segments.

Local segmentation must be frugal in its demands for pixel data. An important factor is the
processing of pixels from the image margins because they have incomplete neighbourhood
information. For a  window applied to an i i
is )fiP*
/yx@i { . For whole images ( iW
( i

image, the proportion of affected pixels

F ), this is usually insignificant, but for sub-images

 ) being locally segmented, this proportion can be very high. For the common  

window the proportion is x1 nearly every pixel in the block.

Although the underlying assumptions of local and global segmentation are different, any
good local segmentation algorithm will still owe a debt to its global counterparts. By 1994
over 1000 segmentation algorithms had been published [ZG94] and this number has likely
doubled since. The sheer variety of algorithms makes it difficult to neatly classify them, so
in this thesis they will be examined in terms of the following attributes:

whether or not spatial information is used

suitability to small sub-images

ability to detect the number of segments automatically

underlying segment and noise model

region growing (identify homogeneity) or edge following (identify heterogeneity)

attempt to optimize an objective criterion

time and space complexity

parallel or sequential execution [Ros81]

3.2 Global segmentation

25

In the following sections those global segmentation algorithms relevant to the development
of good local segmentation algorithms are examined. The discussion is broken into two main
parts. Section 3.2.1 deals only with clustering techniques (non-spatial segmentation), while
Section 3.2.2 covers methods which incorporate spatial information.

3.2.1 Clustering
Segmentation which does not use spatial information is sometimes called clustering. Clustering was used in numerical taxonomy and multivariate analysis long before electronic computers existed [Fis36]. Image processing has adapted clustering algorithms from other disciplines by treating pixel values as independent variables. For example, colour pixels have
3 attributes and each pixel can be considered a point in 3-dimensional space. Clustering
involves grouping pixels in this multivariate space.
For the case of greyscale intensity data, clustering into
of estimating

groups reduces to the simpler case


thresholds. Each threshold  to  is an intensity value which acts

as a dividing point between neighbouring clusters. Typically, each pixel in the thresholded
image, j)p|,-/ , is set to a unique intensity, 0 , associated with each of the

clusters.

Equation 3.1 expresses this formally.

v)p|,-/0

m

..
.

Z

..
.

)p|,-/`

)p|,-/`

 {

)p|,-/`

Z

(3.1)

)p|,-/

Thresholding assumes that pixels from different segments form separate populations based
solely on the disparity of their intensities. It is well suited to images containing relatively few
objects and low amounts of noise. If there is a large number of segments, or large variation
within segments, it is more likely that the segments pixel value distributions will overlap,
rendering valleys in the histogram less obvious or even non-existent [LCP90]. Thresholding
would be a good candidate for local segmentation, because on a small scale only a small

26

Chapter 3. Local Segmentation in Image Processing

number of segments are expected. Thresholding is suited also to a variety of homogeneity

criteria, but with one caveat: an intensity used by one cluster may not be used by another.

Binary thresholding
Thresholding is one of the oldest, simplest and most popular techniques used in image processing [PP93]. Most of the thresholding literature is concerned with classifying pixels into
object or background classes [Wes78, FM81, SSW88]. This is known as binary or bi-level
thresholding. Algorithms which deal with three or more classes are called multilevel thresholding techniques [RRK84, PG94]. On a local scale, most sub-images are expected to be
homogeneous. This implies that the next most likely situation is sub-images with two segments. Thus an examination of binary clustering will be useful.
If an image consists of two or more clear objects, the histogram should have a corresponding
number of peaks. The thresholds should be chosen at the valleys of the histogram. For bilevel thresholding, Prewitt et al [PM66] repeatedly smoothed the histogram until a single
minimum existed between two maxima, and chose 

as the intensity corresponding to the

minimum. An example of this method is given in Figure 3.6.

Figure 3.6: (a) the 8 bpp pellets image; (b) its histogram; (c) thresholded using .

Entropic methods treat the histogram as a probability distribution of _

symbols. Kapur et

al [KSW85] split the histogram into two parts and compute the entropy 1 of each distribution.
The optimal threshold,  , is chosen to maximize the sum of the entropies of the two parts.
1

The entropy measures the average information content of a set of symbols, assuming the probability of
each symbol is known. If the probability of symbol is
y , and there are  symbols, the entropy (in bits)
is given by the expression:  

~g~ .

3.2 Global segmentation

27

The aim is to retain as much information as possible in the binarized image by choosing
a split in which each of the two distributions are as uniform as possible. This method has
the advantage of not having to estimate any parameters, but does not help us decide on the
bimodality of the histogram.
The trend in thresholding has been toward mixture modeling, which treats an image histogram as a linear blend of parameterized statistical distributions [EH81, TSM85, MP00].
Kittlers minimum error method assumes the histogram is a mixture of Gaussians with separate means and variances [KI86]. The chosen objective criterion minimizes the classification
error. The resulting threshold corresponds to the intersection point of the two Gaussians,
shown in Figure 3.7. This point is also known as the Bayes minimum error threshold, and
Kittler provides both exhaustive and iterative search procedures for determining it [KI86].
Frequency

Minimum Error Threshold

Intensity
2

Figure 3.7: Determination of the threshold in mixture modeling.

The minimum error method replaced the previously popular method of Otsu [Ots79] and its
fast implementation [RRK84]. It was shown by Kurita et al [KOA92] that Otsus method is
actually equivalent to the Kittlers minimum error method if each distribution in the mixture
has the same form and the same variance. An adjustment to remove a mild bias of the variance estimates due to overlapping of the distributions was provided by Cho et al [CHY89].
In the minimum error paper, Kittler et al describe the problem of encountering a homogeneous image. In this case the histogram would be unimodal, causing the optimal threshold to
always be chosen at either the extreme left or extreme right of the histogram. To some extent
this condition can be used to distinguish between homogeneous and heterogeneous images.
This is important in terms of local segmentation as a large proportion of the sub-images are
expected to be homogeneous and should be treated as such.

28

Chapter 3. Local Segmentation in Image Processing

Methods to evaluate binary thresholds objectively do exist [WR78]. Sahoo et al [SSW88] use
uniformity and shape measures to compare eight different algorithms. The uniformity
measure is concerned only with minimizing the sum of intra-class variances, while the shape
measure incorporates spatial coherence of edge boundaries. Using real images, which were
not bimodal, they found the minimum error method and the entropic method to perform best.
They also point out that it would be trivial to design a new thresholding algorithm which
jointly optimizes itself with respect to the shape and uniformity measures. This severely
limits the usefulness of these objective criteria.
Lee et al [LCP90] examined five global thresholding algorithms. They used two test images
for which the correct binary segmentation was known. In addition to the shape and uniformity measures, they measured the probability of mis-classification relative to the correct
segmentation. They stated that no algorithm obviously stood out, and lamented on the difficulty of comparing algorithms across different types of data. Despite this, they decided that
Otsus method [Ots79] was the best overall. This suggests that the minimum error method
would have done as well if not better. Glasbey [Gla93] also performed a similar experiment
by generating artificial histograms from mixtures of Gaussian distributions with different
means and variances. He found the iterated form of the minimum error method to do best.
This is not surprising given the way the data was generated.

Local thresholding
Global thresholding methods use the same threshold for every pixel in an image. Difficulties
arise when the illumination of a scene varies across the image [GW92]. In Figure 3.8 the
mug has been thresholded at <D

, corresponding to the most prominent valley in its histogram. The resulting segmentation is poor, because the shadow and the darker background
area have been grouped with the mug. To counteract this, a different threshold,  )p|,-/ , can
be used for each pixel in (or region of) the image. This dynamic threshold could be based on
the histogram of an appropriately sized sub-image encompassing each pixel. This is known
as adaptive, variable or local thresholding [CK72].
Nakagawa et al [NR79] divide an image into non-overlapping windows. The histogram is
assumed to be a mixture of Gaussians, but the mixture parameters are estimated under the

3.2 Global segmentation

29

Figure 3.8: (a) the mug image; (b) its histogram; (c) thresholded using PF .

assumption of no overlap. A test for bimodality (two segments) versus unimodality (one segment) is then applied. The test involves three criteria and four user supplied numbers, and is
designed to ensure good separation and meaningful variances. If the window is bimodal, the
threshold is chosen to minimize the probability of mis-classification. For unimodal windows,
the threshold was chosen by linearly interpolating neighbouring thresholds. An extension to
windows of three segments is also described, where neighbouring threshold information aids
in choosing which of the two thresholds to choose.
Local thresholding is clearly exploiting the principle of local segmentation. Obviously, any
global thresholding technique could be used for determining each local threshold. Smaller
windows increase the chance of obtaining a unimodal histogram, corresponding to a homogeneous (or extremely noisy) region. If a method is unable to detect this situation, the
threshold it computes could be nonsensical. It is important to be able to determine the number of segments present in a local window.

Multilevel thresholding
Some binary thresholding techniques can be adapted to function with more than two clusters [RRK84, KI86]. The mixture modeling approach extends easily, because it models a
histogram as a blend of distributions, choosing thresholds at the valleys between the peaks
of distributions. Equation 3.2 gives the general form of a mixture model with Q classes. The


s are the mixing weights which must sum to unity, B)/ is the distribution function for


class  , and  holds the parameters of distribution  .

30

Chapter 3. Local Segmentation in Image Processing

)pZ/J




(3.2)

;1)p  /

Often each class is assumed normally distributed with unknown mean and variance. In this


situation )pZ/ has BQ
degrees of freedom. These parameters are usually estimated using
the Expectation-Maximization (E.M.) algorithm [DLR77]. The E.M. algorithm is an iterative
process which is guaranteed to increase, at each step, the statistical likelihood [Fis12] of
the data given the mixture model probability density function,

)pZ/ . Although the E.M.

algorithm always converges, it may do so only to a local optimum. Usually, multiple runs
at different starting conditions are required to overcome this. However, greyscale pixel data
is a simple one-dimensional case, and these problems usually only occur in multivariate
situations, such as colour pixels.
The general E.M. algorithm is computationally very expensive because it calculates Q integrals per datum per step. The integrals are used to determine the partial membership of
each datum to each distribution. For the normal distribution this must be done numerically
as no closed form exists. If the distributions are well separated, it is possible to assign each
pixel wholly to one class without biasing the parameter estimates. If, additionally, the distributions are assumed normal with a common variance, the maximum likelihood criterion
optimized by the E.M. algorithm reduces to minimizing the sum of squares of each datum to
its centroid [Alp98]. This criterion may be optimized using simpler techniques.
The Q -means algorithm really refers to two different algorithms, both designed to minimize
the sum of squared errors of each datum to its cluster centroid. The simpler form, referred to
as k -means by Hansen et al [HM01], is given in Listing 3.1. Hansen et al refers to the more
complex form as Q -means [McQ67, DH73, HW79], given in Listing 3.2.

1. Choose Q initial class means.

2. Assign each pixel to the class with the closest mean.
3. Recompute the class means using the new assignments.
4. If any pixels have switched class, go to Step 2.
Listing 3.1: The ! -means algorithm.

3.2 Global segmentation

31

1. Choose Q initial class means.

2. Assign each pixel to the class with the closest mean.
3. Reassign one pixel to a new class such that the move best reduces the overall
sum of square distances criterion.
4. Recompute the class means using the new assignments.
5. If the criterion was improved, go to Step 3.
Listing 3.2: The " -means algorithm.

In k -means, every pixel is reassigned per iteration, whereas Q -means only performs one optimal re-assignment. For one-dimensional data like greyscale pixels, both algorithms should
produce the same clustering if provided with sensible starting conditions. This is especially
true when Q is low, as it is expected to be in local segmentation.

The ISODATA [RC78] algorithm is similar to Q -means but introduces various heuristics for
improving the clustering result. These heuristics include forcing a bound on the minimum
and maximum number of classes, requiring a minimum class membership amount, and restricting the maximum variance within a class. Although these heuristics must be supplied as
parameters by the user, they do allow the ISODATA algorithm to go some way to estimating
the number of classes and ensuring they have sensible properties. They also allow the user
to incorporate a priori information into the clustering process. The natural level of variation
within the image, if known, could be used to restrict the maximum variance within a class.

Fuzzy ! -means [Bez81] is a variation on Q -means which allows each datum partial membership in each cluster, similar to a mixture model. Compared to the sudden jumps during re-assignment in Q -means, the cluster centroids in fuzzy ! -means move more smoothly.
Its objective function alss differs, claiming to better suited to non-spherically distributed
data [HB88]. Fuzzy ! -means must calculate partial memberships for each datum at each
step, which, from a computational perspective, may make it unsuited to local segmentation.

32

Chapter 3. Local Segmentation in Image Processing

Summary
Thresholding techniques produce segments having pixels with similar intensities. They can
handle any underlying image model as long as each intensity is associated with only one
segment. The disregard of spatial information means that pixels within a segment may not
be connected. In terms of local segmentation this may be an advantage, because the disconnected components of a local segment could actually be part of the same global segment.
Global thresholding will suffer when pixels from different segments overlap in their use of
intensities. If this is due to noise, a technique such as the minimum error method can estimate
the underlying cluster parameters and choose the thresholds to minimize the classification error. If the overlap is due to variation in illumination across the image, variable thresholding
could be used. This can be seen as a form of local segmentation.
The likelihood of encountering unimodal histograms is high when thresholding sub-images.
Within the local segmentation framework it is extremely important to be able to detect the
number of segments in the window. Most of the thresholding literature either ignores this
situation, or handles it with special parameters which must be provided by the user. It would
be better if these parameters were estimated from the image itself.
Thresholding has low space complexity. Only one pass through the image is required to
build a histogram. All further calculations are performed using only the histogram, as no
spatial information is required. For this reason its time complexity is also low. If the number
of pixels in the image is low compared to the number of possible grey levels (<T_ ),
the same benefits apply without using a histogram. Local segmentation must be applied
independently to each pixel, making thresholding an attractive option.

3.2.2 Spatial segmentation

Global segmentation algorithms which take spatial information into consideration usually
outperform their clustering-based counterparts. Spatial information is useful because most
segments corresponding to real world objects consist of pixels which are spatially connected.
Evaluating the quality of segmentation algorithms is an inherently difficult problem [PP93,
Zha97]. This section will discuss the main approaches to spatial segmentation and assess
their suitability to local segmentation.

3.2 Global segmentation

33

Region based segmentation

Region growing algorithms tackle segmentation by identifying homogeneous, spatially connected sets of pixels within an image. Splitting algorithms begin with large image regions
and break them up into smaller, more homogeneous ones. Merging algorithms compare
neighbouring regions (or pixels) and merge them if they have similar enough properties.
Some techniques use a combination of splitting and merging. What they all have in common
is the use of a measurable image property as a criterion for splitting and merging [YG01].
Region growing requires a seed to begin. Ideally the seed would be a region, but it could be
a single pixel. A new segment is grown from the seed by assimilating as many neighbouring pixels as possible that meet the homogeneity criterion. The resultant segment is then
removed from the process. A new seed is chosen from the remaining pixels. This continues
until all pixels have been allocated to a segment. As pixels are aggregated, the parameters
for each segment have to be updated. Thus the resulting segmentation could depend heavily
on the initial seed chosen and the order in which neighbouring pixels are examined.
Region splitting begins with a whole image and divides it up such that the parts are each
more homogeneous than the whole. The main difficulty is in deciding on where to make
a partition. Early approaches used a regular decomposition such as a quad tree [FKN80].
Splitting alone is insufficient for reasonable segmentation as it severely limits the shapes of
segments. Horowtiz et al [HP74] suggest a merging phase after splitting is complete. This
can be done efficiently using tree searching algorithms [Knu73], but the resulting regions
may have unnatural blocky boundaries.
A popular image model used today is the piece-wise constant segment model with additive
Gaussian noise. Under this model, pixels from the same segment are normally distributed
with mean
#

and variance & { . The variance provides a measure of homogeneity within

the segment. The sample mean and variance may be used to estimate
#

and & . Fishers

criterion [YG01] in Equation 3.3 is an example of a homogeneity predicate for splitting or

merging two classes, A and B.

(#*) {
& & { uX& ){

#'&

(3.3)

34

Chapter 3. Local Segmentation in Image Processing

$

Fishers is maximized by classes with widely separate means and low variance. Typically

would be compared to a threshold if it is lower, merging would proceed. The value of the
threshold impacts the number and size of segments accepted. Ideally, its value should depend
on the natural level of variation of pixel values in the image. From a local segmentation point
of view, Fishers criterion could useful for distinguishing between the existence of one or
more clusters in a sub-image.

Edge based segmentation

Edge based segmentation is the dual to region growing. It exploits spatial information by
detecting the edges in an image, which correspond to discontinuities in the homogeneity
criterion for segments. Edge detection is usually done with local linear gradient operators,
such as the Prewitt [PM66], Sobel [Sob70] and Laplacian [GW92] filters. These operators
work well for images with sharp edges and low amounts of noise. For noisy, busy images
they may produce false and missing edges. The detected boundaries may not necessarily
form a set of closed connected curves, so some edge linking may need to be required [Can86].
There are many edge based algorithms in the literature, particularly in the area of active
contours and snakes [KWT87]. However, the application of edge models to local regions
is unlikely to be successful as local regions contain too few pixels, of which most have
incomplete neighbour information.

Hybrid region-edge approaches

Watershed segmentation [MB90, HRA96] has become popular recently, assisted by the existence of an efficient implementation [VS91]. It is a hybrid approach which uses edge detection to determine seeds for growing regions, which are then merged. First an edge detector
is applied to the image, resulting in an elevation map which quantifies the edge magnitude
around each pixel. Seeds for region growing are taken from the points of lowest elevation, the
so-called basins. Region growing proceeds in a manner similar to how the basins would fill
if the elevation map was slowly and uniformly covered with water. When the immersion
process completes, the basins are merged into suitably sized global segments.

3.2 Global segmentation

35

The edge detection and region growing steps in watershed segmentation are only loosely coupled, because the elevation map can be provided beforehand. Other hybrid techniques tightly
integrate these steps, for example, Tabbs multi-scale image segmentation approach [TA97].
Tabb suggests that scale selection and edge and region detection can not be separated, and
that good segmentation algorithms effectively have to perform a Gestalt analysis. Although
this approach produces pleasing results on whole images, it is difficult to apply on a local
level where only a small number of pixels are available. His algorithm does have the admirable feature of automatically determining all required parameters from the image itself.
Relaxation labeling
Relaxation labeling can be applied to many areas of computer vision. In the past it has
been applied to scientific computing applications, particularly to the solution of simultaneous
nonlinear equations [RHZ76]. The basic elements of the relaxation labeling method are a
set of features belonging to an object and a set of labels. In the wider context of image
processing, these features are usually points, edges and surfaces. For the case of image
segmentation each pixel has a feature describing which segment it belongs to, and there is a
different label for each segment in the image.
The labeling schemes are usually probabilistic in that each pixel is partially assigned to all
segments. A pixels assignment takes the form of normalized probabilities which estimate
the likelihood of it belonging to each segment. Different techniques are used to maximize
(or minimize) the probabilities by iterative adjustment, taking into account the probabilities
associated with neighbouring features. Relaxation strategies do not necessarily guarantee
convergence, thus pixels may not end up with full membership in a single segment.
Two main strategies are used for optimizing the iteration process [YG01]. Simulated annealing, or stochastic relaxation, introduces a random element into the iteration scheme, to reduce
the chance of becoming stuck in secondary optima. Geman et al [GG84] applied this to a
local Markov random field model [Li01]. In contrast to stochastic relaxation, which makes
random changes, deterministic relaxation only makes changes which improve the configuration, and thus converges much faster. Besags Iterated Conditional Modes (ICM) [Bes86]
algorithm uses this approach for image segmentation. Due to its greedy nature, ICM does
not find a global minimum, and its results are dependent on the initial conditions chosen.

36

Chapter 3. Local Segmentation in Image Processing

Extending clustering methods

Haralick et al [HSD73] used the co-occurrence matrix to incorporate the spatial relationship
between pixels into the choice of threshold in texture analysis. This matrix is constructed
as a 2-D histogram from pairs of pixels at a fixed orientation and distance apart, similar to
a Markov model. Figure 3.9 gives an example of two co-occurrence matrices for a simple
 J image consisting of three noiseless segments. Pixels within segments should dominate

the diagonal entries of the matrix (similar grey levels) while the off-diagonal entries should
be made up of edge pixels (differing grey levels). By creating two separate histograms from
these groups and finding where the valley in the first matches a peak in the second, a threshold can be obtained. The interested reader is directed to the paper for further information.
This approach is not suited to small sub-images as there are too few pixels to generate a
meaningful co-occurrence matrix.

b
1

a 1

b
1

a 1

0
a b

a
b

Figure 3.9: (a) image with 3 segments; (b) co-occurrence matrix for horizontal pairs of
pixels; (c) co-occurrence matrix for vertical pairs of pixels.

Leung and Lam [LL96] introduce the concept of segmented-scene spatial entropy (SSE) to
extend the entropic thresholding methods to make use of spatial information. Thresholds are
chosen to maximize the amount of information in the spatial structure of the segmented
image. Their spatial model relates a pixel to its four immediate neighbours. This method is
too complicated to apply to small sets of pixels with little connectedness information.
An interesting approach to bi-level thresholding was taken by Lindarto [Lin96]. Traditional
entropic methods try to maximize the information needed to encode the pixel values once
the segmentation is determined. Lindarto, however, chooses a threshold which minimizes
the combined lossless encoding length (information content) of the binary segment map and

3.3 Local segmentation

37

the original pixels. The grey level pixels are encoded using a least-entropy linear predictor [Tis94] optimized for each segment, while the bitmap is efficiently stored using the JBIG
algorithm [PMGL88]. This approach uses Wallaces minimum message length (MML) inductive inference principle [WB68, OH94] to choose the model (threshold) which minimizes
the overall compressed length of the model plus the data given the model. The versatile MML
model selection technique will be applied to local segmentation in Chapter 5.

Summary
Segmentation algorithms which use spatial information can produce better results than those
which do not. This is especially true for noisy images, where pixel intensity alone is insufficient for distinguishing between segments. However, spatial segmentation algorithms would
struggle with the low amount of spatial information available in a small window. Local
segmentation must be applied to every pixel, and hence a simple and fast algorithm is also
desirable. Thresholding techniques are therefore an attractive option for local segmentation.

3.3 Local segmentation

Local segmentation is concerned with segmenting small sub-images in isolation. The main
difference between local segmentation and global segmentation is that the former has only
a small number of pixels available to it. This increases the chance that the sub-image is
homogeneous. If an edge is present, it may not be as obvious without the context of a whole
image. The concept of modeling and segmenting local regions is not new. Many algorithms
already use it in some form, but its application may be implicit and non-obvious.

3.3.1 The facet model

Most image processing applications assume that the pixel grid is a discrete approximation
to an underlying intensity surface. The facet model of Haralick et al [HW81] assumes that
the true surface is piece-wise continuous and that the image is a noisy sampled version of

38

Chapter 3. Local Segmentation in Image Processing

it. In particular, it assumes that each segment in the image is well modeled using a twodimensional polynomial function. The flat facet model assumes each piece has a constant
intensity. The sloped facet model allows the intensity to vary linearly in both directions, like
a plane. The idea extends easily to higher degree polynomials. It would be possible to fit
facet models to sub-images.
There are two main problems with the facet model. Firstly, it does not provide a criterion
for deciding which order polynomial best fits the local region. It is important to strike a
balance between over-fitting the noise and under-fitting the structure. Secondly, polynomial
functions are not well suited to modeling sharp edges between arbitrarily shaped segments.
Segment boundaries are visually very important and it is important to handle them correctly.

3.3.2 Block truncation coding

One of the oldest appearances of local segmentation is the lossy, fixed bit rate image compression technique block truncation coding, or BTC [DM79, FNK94]. In one form, BTC
divides an image into non-overlapping   blocks pixels. The pixels are segmented into
two clusters using thresholding at the block mean. The two cluster centroids, # and # , are
set average value of the pixels in each cluster. Figure 3.10 provides an example.

12

15

12

12

12

11

11

12

12

12

12

15

12

12

14

12

=2
0

= 12
1

Figure 3.10: BTC example: (a) original pixel block; (b) decomposed into a bit-plane and
two means; (c) reconstructed pixels.

From a local segmentation perspective, BTC assumes that segments consist of pixels of a
common grey level separated by step edges aligned to pixel boundaries. The assumption of
two segments implies the existence of a bimodal histogram, and the use of the mean as the
threshold implies that each segment contains approximately the same number of pixels.

3.3 Local segmentation

39

The original BTC algorithm has been modified repeatedly over the years [FNK94]. In standard BTC, the number of segments is always two. Some techniques allow the number of
segments to be varied on a block by block basis. For example, if the difference between the
two means,

%
#

+# % , is low enough, the block is considered homogeneous, so no bitmap and

only one mean are encoded. Conversely, if the block variance is very high then multilevel
thresholding can be used. This approach to automatically detecting the number of clusters is
rudimentary, but still effective. In Chapter 4, this idea will be extended to form the basis of
an effective local segmentation technique for removing noise from images.
Some BTC variants replace thresholding with spatial techniques designed to better preserve
image structure. Gradient based BTC [QS95] and correlation based BTC [Cai96] choose
a different threshold for each pixel in the block. Each threshold depends on neighbouring
gradient and pixel difference information. This itself is a crude form of relaxation labeling.
For most images the gain is minimal, suggesting that spatial information may be of limited
value when segmenting small regions.

3.3.3 SUSAN
SUSAN [Smi95, Smi96, SB97], much like the local segmentation framework of this thesis,
evolved to be a general approach to low level image processing. One of SUSANs design
constraints was efficiency, so that it could be run on real time data in a robotic vision system.
SUSAN was originally designed for edge and corner detection, but was also adapted for
structure preserving denoising. The denoising component differs a little from the SUSAN
framework presented here, and is treated separately in Section 3.4.5.
SUSAN processes greyscale images in a local manner using an approximately circular window containing 37 pixels. The centre pixel in the window is called the nucleus. Neighbouring
pixels similar in intensity to the nucleus are grouped into an USAN, or Univalue Segment Assimilating Nucleus. The USAN creation process may appear to be a form of region growing
using the nucleus as a seed, except that pixels in the USAN are not necessarily connected.
The formation of an USAN is more related to clustering than segmentation.
The pixel similarity function is controlled by a brightness threshold, . Smith first describes
the USAN as accepting only pixels with an intensity difference of units from the nucleus.

40

Chapter 3. Local Segmentation in Image Processing

This can be considered crisp clustering, where similar pixels receive full weight, and dissimilar pixels receive no weight. The size of the USAN is equal to the sum of weights given
to pixels in the window, which for crisp clustering, equals the number of pixels assimilated.
Smith, however, found a fuzzy membership function to give better feature detection. This
fuzzy function was optimally derived to have the form
,

.-0/*1325476

, where
8

is the intensity

difference from the nucleus. Figure 3.11 compares the crisp and fuzzy membership functions
as a function of
8

. When the fuzzy form is used, all pixels have partial contribution to the

size of the USAN. This is no longer an explicit segmentation of the local neighbourhood.

Cut-off function
Fuzzy / Soft
Crisp / Hard

USAN weight

0.8

0.6

0.4

0.2

10

15

20

25

30

35

40

Intensity difference from nucleus ()

Figure 3.11: Comparison of the hard and soft cut-off functions for pixel assimilation in the
SUSAN algorithm.

To perform feature detection, the USAN size for each pixel is plotted. On this surface homogeneous regions correspond to plateaus, edges to valleys, and corners to deeper valleys.
SUSAN processes this USAN surface to determine the directions and strengths of edges and
the positions of corners. Smith finds the SUSAN feature detector to give results on a par
with Cannys famous edge detection algorithm [Can86]. Smith states that SUSAN is suited
to images with segments of smoothly varying intensity and both step and ramp edges.

3.4 Denoising

41

The main difficulty with SUSAN is that its performance is dependent on a user supplied
brightness threshold, . Smith claims that SUSAN is relatively insensitive to the choice of ,
and advocates a default of 20. He states that may be used to vary SUSANs sensitivity to
features. The value of really controls the minimum feature contrast that can be detected.
My experiments show that should be proportional to the natural level of intensity variation
expected within image segments. This is the only way for the USAN to assimilate pixels in
homogeneous regions and not produce false edge output.

3.4 Denoising
Image denoising is the process of removing unwanted noise from an image [Mas85, BC90].
The noise can take a variety of forms and is introduced in differing amounts at each step
during the acquisition of the image. The literature abounds with denoising techniques, but
they may be classified broadly into temporal and spatial approaches.

3.4.1 Temporal filtering

Temporal filtering averages multiple sampled versions of the same scene [She90]. If the true
image is = , and i

samples, ; ;~:9 , are taken, the temporally filtered image, =? , can be

calculated using Equation 3.4.

= ? 
i

(3.4)

If each pixel in = was corrupted by the addition of symmetric zero-mean noise of variance
& { , then for any one pixel the expected noise variance, @)+=C/ { , is & { . However, the

expected noise for the ensemble average, =? , reduces to & { xi . If the noise was not additive
in nature, for example impulse noise, then a better averaging function, such as the median
(discussed later), should be used.
Temporal filtering is ideal if multiple version of the same image can be obtained. In practice
this does not usually happen, because the objects in the scene move, or the capturing equip-

42

Chapter 3. Local Segmentation in Image Processing

ment wobbles. Even slight variations can displace the pixels in each sampled image, causing
the ensemble average to become a blurred version of the original.
In this thesis it is assumed that only a single noisy version of an original image is available,
so temporal filtering of the form in Equation 3.4 can not be used. However, it will be shown
how overlapping local segmentation results for the same pixel coordinate can be combined
to suppress noise in a fashion similar to temporal filtering.

3.4.2 Spatial filtering

Where temporal filtering uses multiple versions of each pixel value at the same position
but different times, spatial filtering uses pixels at the same time but at different spatial
coordinates in the image. When temporally filtering a static scene, pixels from the same
position but different times are expected to have the same noiseless value. When those
pixels are not available, the best alternative is pixels near the current pixel. When only a
single noisy image is available, spatial filtering is the only option. This is the basis of nearly
all image denoising algorithms.
The simplest spatial filter is the averaging filter which replaces a pixel with the average of
itself and the pixels in the local neighbourhood. The most common averaging filter is the
box filter [McD81], which gives equal weight to the 9 pixels in a  window. This choice
of weights maximally reduces the noise variance when the noise is additive. Box filtering
is similar to temporal filtering except that samples from adjacent pixel positions are used as
approximations to multiple samples of the pixel at the same position. Figure 3.12 gives an
example of its application.

Figure 3.12: Application of the box filter: (a) original; (b) noisy; (c) filtered.

3.4 Denoising

43

Although an averaging filter performs well for additive noise in homogeneous regions, it
tends to blur edges and other image structure in heterogeneous regions. Figure 3.13 gives an
example of this undesirable behaviour. To combat this deficiency, most research is concerned
with structure preserving denoising algorithms.

Figure 3.13: Box filtering in the presence of edges: (a) original image; (b) filtered.

To preserve image structure while removing noise implies the ability to distinguish between
the two. It will be shown in later chapters that local segmentation is a good basis for modeling the relationship between structure and noise in an image. Segments correspond to
homogeneous regions and boundaries between segments are structure that needs to be preserved. Existing spatial denoising algorithms can be examined in terms of the way they use
local segmentation. Some use no local segmentation at all, some use it implicitly, and some
explicitly use a local segmentation model.

3.4.3 Fixed linear filters

The simplest denoising filters are linear ones. For a square window of side length
Fi

u#
,

the operation of a linear filter is described by Equation 3.5, where the weights, (T)/ , are
assumed to sum to one.

? )|,-/0

9
;=<

9 =; > 9

(T)p ? , ? /~)pTu@? ,ruA? /

The box filter described in Section 3.4.2 is a linear filter with i



(3.5)

and equal weights

(*)B? ,C? /J
x)f

FiuP
/ { . Its configuration is illustrated in Figure 3.14a. It shows the spatial

arrangements of the weights used by the linear filter. Interestingly, McDonnell [McD81]

44

Chapter 3. Local Segmentation in Image Processing

developed a recursive algorithm which can implement box filtering with only five operations
per pixel, independent of i .

1 1
1 1
1 1

1
1
1

BE

1 2
2 4
1 2

1
2
1

Figure 3.14: (a) box filter; (b) approximate } Gaussian filter.

Gaussian filters use normalized weights of the form in Equation 3.6. This filter varies the
weights such that those pixels further from the centre contribute less to the overall average.
The spatial extent of a linear Gaussian filter is determined by & , usually chosen to be around
)
1iPuj
/~xF , as to be mostly encapsulated by the bounding square. Figure 3.14b illustrates the

configuration of a   Gaussian filter in this case. Gaussian filters are sometimes preferred
over box filters because their effect on the Fourier spectrum of the image is better [GW92].

(*)B? ,F? /HGI,

JB<CK L JB>K

(3.6)

KM K

From a local segmentation point of view, each linear filter weight may be interpreted as
being proportional to the believed probability that the pixel is related to the centre pixel.
Relatedness really means in the same segment. The centre pixel usually has the highest
weight, as it should be the best estimator of its own original value when the noise is additive.
If the centre pixel is at coordinate )|,-/ , and another pixel is at )p7uN? ,u(? / , the implied
probability that they are in the same segment is given by Equation 3.7.

c0d)f ? )|,-/PORQ ? )uA? ,u@? /|O7FTS
S@BTSO/J

(T)U? ,C? /
(*)+-,./

(3.7)

For example, the box filter trusts all the pixels in the window equally, giving them probability
one of being from the same source. It has been stated that the correlation between pixels is
often on the order of 0.9 to 0.95 [NH88]. From this point of view, the coefficients of the
box filter are quite reasonable. For the approximate  Gaussian filter in Figure 3.14b, the
side pixels are considered
Bx1PV-4 correlated to the centre pixel, and the corner pixels

3.4 Denoising

45

x1RE 
B correlated. This may be considered a primitive form of local segmentation, but

one which is unable to adapt to changing image structure.

All fixed linear filters will defocus an image, blurring edges as in Figure 3.13, and attenuating narrow lines and other fine structure. The behaviour worsens as i

increases. A positive

feature is that smoothing is very good in homogeneous regions corrupted by additive noise,
reducing the noise variance by up to a factor of )
Fiu
/ { . What is required is a method to
distinguish between homogeneous and heterogeneous regions, to get the benefits of smoothing without the disadvantage of blurring edges.

3.4.4 Rank filters

Rank filters are designed to operate on a numerically ordered set of pixels. The pixels ;; V9
from the local neighbourhood are gathered and sorted into a new set -mU4 ; @

-9 4

, where

B - 4 - U 4 . For greyscale images, the ordering is determined by pixel intensity, but there

is no obvious analogue for higher dimensional data such as colour pixels. Manipulation of
the ordered set of data originated as a way to improve the statistical robustness of traditional
estimates (such as the mean) when the data was contaminated [MR74, Hub81]. Digital
images are often contaminated by noise. Thus, to some, it seemed natural to apply rank
techniques to image denoising.

The median
The simplest rank filter is the median filter. The median is the middle datum in a sorted
data sequence [Tuk71]. Just as the mean minimizes the average squared error, the median
minimizes the average absolute error. If the number of pixels is odd, the median is uniquely
defined. When it is even, there are an infinite number of solutions between the two middle
pixels. In this case, the average of the two middle values is usually taken as the median.
Equation 3.8 summarizes the calculation of the median, XW:Y , of

XW:YA

L \ 4
[- Z ]
_
K

- Z 4 uA - Z U4a`
{
K
K

^Q Q

ScbPSO

pixels.

(3.8)

46

Chapter 3. Local Segmentation in Image Processing

Figure 3.15 gives an example of the median applied to a  sub-image. The pixels values
are homogeneous with a small amount of additive noise. In this situation the median would
choose the filtered value to be 6. The mean, d , of these pixels is 5.78, which would usually be
rounded to 6. When the noise is additive and symmetric, the mean and median will usually
agree on the filtered value, as both estimate the centre of a distribution.
6 5
5 7
4 6

8
6
5

(455566678)

XW:YM

d I 4F

Figure 3.15: Application of the median to a homogeneous block of pixels.

Consider the situation in Figure 3.16, whereby the centre pixel from Figure 3.15 has been
contaminated by impulse noise. The pixel values are no longer homogeneous. A box filter
would produce a filtered estimate of 13.1, the outlying pixel value 73 having pulled the mean
away from the true centre. The breakdown point of the median is 50% because it is still able
to produce useful estimates when up to 50% of pixels are contaminated. In this case, the
median is 6 again a much better estimate than the mean.
6 5
5 73
4 6

8
6
5

( 4 5 5 5 6 6 6 8 73 )

XW:YM

d 
-m

Figure 3.16: Application of the median to a homogeneous block with impulse noise.

In the previous two examples the pixel blocks were meant to be homogeneous. When the
noise is additive, the mean and median behave similarly. In the presence of impulse noise
the median is superior to the mean, as it can cope with up to 50% contamination. A good
structure preserving filter must also function well in the presence of edges. Figure 3.17 shows
a  pixel block containing two noiseless segments. As expected, the mean averages across
the edge. The median correctly chooses 3 as its filtered estimate, making it appear to have
special structure preserving abilities.
Figure 3.18 gives a similar example, but now the centre pixel is on the corner of a noiseless
segment. The median in this case is 8 the value belonging to the alternate segment.
This illustrates that the median is implicitly performing primitive local segmentation. In the

3.4 Denoising

47

3 3
3 3
3 3

8
8
8

(333333888)

XWeYA

j
d $N

Figure 3.17: Application of the median to a noiseless edge block.

presence of two distinct segments, the median chooses a pixel from the segment containing
the majority of the pixels from the window. For a  window, this decision rule fails
whenever the intersection of the local window with the centre pixels segment consists of
fewer than five pixels. When the window contains more than two segments, the medians
behaviour is a more complicated function of the segment populations.
3 3
3 3
8 8

8
8
8

(333388888)

XWeYA

j
d # 4B

Figure 3.18: The median (incorrectly) implicitly segmenting a noiseless corner block.

The weighted median

The simple approach to local segmentation that the median takes causes it to fail when the
centre pixel belongs to a minority segment. Unfortunately, thin lines and other fine structure
correspond to this situation. Instead of moving toward a local segmentation analysis of the
problem, the concept of the median has been generalized to handle various special cases.
A popular variant is the centre weighted median [Bow84]. The philosophy of the centre
weighted median is to increase the size of the ordered set of pixels by adding ! extra duplicates of the centre pixel, before taking the median. The value of ! depends on the number of
pixels in the window, and the type of structures one wishes to preserve [KL91].
For  windows, ! 

could be used to preserve corner structures like the one in Fig-

ure 3.18. Figure 3.19 shows that when the centre weighted median is used, the corner pixel
is correctly filtered as 3. The duplication of the centre pixel simply improves the probability
that the centre pixel will belong to the majority segment. For the plain median filter, the centre pixel needs to be in a segment having at least f ixB
hg members, where f ig denotes round-

48

Chapter 3. Local Segmentation in Image Processing

ing to the nearest integer. The weighted median reduces this requirement to f )fi

P!/~xF
hg

pixels. This improvement comes at the cost of reducing its breakdown point.
3
3
8

3 8
3 8
8 8

(33333388888)

XW:YM

d I 4B

Figure 3.19: Using the centre weighted median (jb ) on a noiseless corner block.

Further variations on the median

Figure 3.20 illustrates the application of the median and centre weighted median filters to
a greyscale image. Despite their obvious failure on such a basic, albeit contrived, example,
these median based filters continue to be used as components in image processing systems.
Many extensions to the median have been suggested to overcome its defects [Tag96, GO96,
EM00]. These techniques suffer from not seriously attempting to diagnose the structural
features present within the local region [KL91], or require sets of parameters to be estimated
from a training set [CW00].

Figure 3.20: (a) original image; (b) } median filtered; (c) weighted median, jP .

The limitations of median based approaches are not due to the median operator itself, but to
its misapplication to heterogeneous data. The time spent on generalizing the median to two
dimensional image data may have been better spent concentrating on modeling the underlying image from a local segmentation point of view. Local segmentation would suggest that
pixels should first be appropriately segmented into homogeneous groups. The median could
then be applied just to the pixels belonging to the segment containing the centre pixel.

3.4 Denoising

49

3.4.5 Locally adaptive filters

An adaptive denoising algorithms adjusts the functionality of the filter for each pixel in
response to an images local properties. One would expect this to improve performance over
the static methods already discussed. Most adaptive algorithms are based on either or both
of the following two principles, whether they acknowledge it or not:

1. Pixels that are spatially close are more likely to be in the same segment.
2. Pixels that are similar in intensity are more likely to be in the same segment.

Adaptive denoising algorithms apply these principles in order to use only those pixels in
the same segment as the centre pixel for denoising. This grouping can be seen as a partial
step of local segmentation, whereby a homogeneous segment involving the centre pixel is
formed. Perhaps surprising is that the same two principles are also the basis for most global
image segmentation algorithms. This suggests that segmentation and denoising are highly
related tasks. In Chapters 4 and 5, this fact will be used to develop local segmentation based
denoising algorithms.

The Kuwahara filter

Given a  window, the Kuwahara filter [Kuw76] examines the nine unique  regions
which incorporate the centre pixel, shown with a bold outline in Figure 3.21. The mean of
the  window with the lowest variance is used as the denoised estimate. In terms of local
segmentation, utilizing the  window with lowest variance increases the chance that the
mean was computed from a homogeneous part of the image. If all global segments in the
image are at least  in size, then at least one of the nine candidate regions containing the
centre pixel will be truly homogeneous. Unfortunately this will not necessarily be the case.
There are many obvious cases where the Kuwahara filter is forced to choose a mean from
nine heterogeneous regions, resulting in blurring. For example, in Figure 3.21 the window
surrounded by a dotted line has the lowest variance. Unfortunately, that windows mean is
also highly biased to an intensity not equal to that of the centre pixel.

50

Chapter 3. Local Segmentation in Image Processing

Lowest Variance

Figure 3.21: The Kuwahara filter considers nine regions within a } window.

Despite its failings, the Kuwahara technique illustrates two important points. Firstly, the case
where the pixel being processed is at the centre of the window is not the only useful case;
each pixel participates in many overlapping windows. Secondly, the Kuwahara filter chooses
the best fitting model from a set of candidate models using an objective criterion. Both these
ideas are exploited by local segmentation denoising algorithms in Chapters 4 and 5.

Gradient inverse weighted smoothing

Gradient inverse weighted smoothing [WVL81], or GIWS, is an adaptive linear filter for
greyscale images. The weight for each pixel is inversely proportional to the intensity difference between it and the centre pixel. This is shown in Equation 3.9, where = ? is the noisy
image. The R operator prevents division by zero when two pixels have the same value.

(T)U? ,C? /HG

 
{

%

, ? )pTu@? ,ruA? /X ? )|,-/

(3.9)

There are two variants of the GIWS filter: the first includes the centre pixel in the weighted
average, while the second one does not. If the centre pixel is included it will always be given

3.4 Denoising

51

a weight of 2. If not included, GIWS becomes resistant to single pixel impulse noise [Hig91].
If the image is unaffected by impulse noise, the former variant should be used, as it will give
more smoothing in the presence of additive noise.
Figure 3.22 gives an example of GIWS applied to a noiseless   corner block. The resultant
weights are higher for those pixels in the centre pixels segment (intensity 3) than those in
the other segment (intensity 8). The smoothed estimate k is 3.56 when the centre pixel is
included in the linear combination, and 3.71 when omitted. Ideally the estimate should be
3, but in both cases it has been polluted with values from the other segment. The extent to
which this occurs may be controlled by modifying the

(a)

3 3
3 3
8 8

8
8
8

(b) (*)B? ,C? /HG

in the denominator of Equation 3.9.

2
2 0.2
2
2 0.2
0.2 0.2 0.2

k #-4F
(c) j

Figure 3.22: Gradient inverse weighted smoothing (GIWS): (a) original pixels; (b) computed
weights, unnormalized; (c) smoothed value when centre pixel included.

The failure of GIWS in Figure 3.22 may be attributed to its refusal to make absolute decisions about which pixels belong together. Its soft cut-off function, shown in Figure 3.23,
allows information from neighbouring segments to diffuse into the current segment. A local
segmentation approach would suggest the use of a hard cut-off to explicitly separate pixels
from different segments. The cut-off point would be related to the natural level of variation
within image segments. For the example in Figure 3.22, a hard-cut off function would have
filtered the block exactly, given any reasonable cut-off point.

Sigma filter
Lees Sigma filter [Lee83] is based on the alpha-trimmed mean estimate for the centroid of a
data set [RL87]. Firstly, the standard deviation, l , of the pixels in the window is calculated.
The filtered value is then set to the average of those pixels within
]l of centre pixel. If the
number of pixels taking part in the average is too small, Lee recommends taking the average
of all the non-centre pixels. Lee claims that the Sigma filter performs better than the median
and GIWS with the centre pixel included.

52

Chapter 3. Local Segmentation in Image Processing

2
1.8
1.6
1.4

Weight

1.2
1
0.8
0.6
0.4
0.2
0

10

15

20

Pixel difference

Figure 3.23: Pixel weighting function for the GIWS adaptive denoising algorithm.

Figure 3.24 demonstrates the Sigma filter on a corner block. The standard deviation of the
block is 2.5, resulting in the averaging of all pixels within 5 intensity units of the centre pixel.
Unfortunately, all the pixels are in this range, causing the Sigma filter to revert to a box filter.
This produces a blurred estimate of k IF .
3 3 8
3 3 8
8 8 8

&2I
F

j
k nmcoPp D q oPr  1

Figure 3.24: Demonstration of the Sigma filter on a noiseless corner block.

The Sigma filter tries to average only those pixels similar enough to the centre pixel. This
is an explicit form of local segmentation using a hard membership function. The similarity
measure uses the local variance to control the maximum intensity deviation a pixel may have
from the centre pixel while still being considered part of the same segment. In terms of
local segmentation, the local variance is an estimate of the natural level of variation within

3.4 Denoising

53

segments. In homogeneous regions it is a good estimate, but around edges it will be too high.
Thus the variation estimate is poor for those cases where its value is crucial, and acceptable
for those cases where it is not overly important.
The local variance is very sensitive to contamination from sources such as impulse noise. To
a lesser extent it also suffers from the small amount of pixel data used to calculate it. The
Sigma filter could be improved by replacing the local variance with a globally estimated or a
priori supplied natural level of variation. This is the approach taken by the local segmentation
based denoising algorithms of Chapters 4 and 5. The fact that the Sigma filter also falls back
to an overall average when too few pixels are available for averaging implies that it does not
allow segments to consist of a single pixel.

Anisotropic diffusion
Anisotropic diffusion [PM90, BSMH98] is a highly iterative process which slowly adjusts
pixel intensities to make them more like their neighbours. In its original form it has been
shown to be very similar to GIWS [Wei97]. Equation 3.10 describes the anisotropic diffusion
process: is the iteration number;
nearest ones;

is the set of neighbouring coordinates, usually the four

s

is the number of neighbours, usually four;

controls the rate of smoothing;

% is a user-supplied function; and & is a user supplied parameter to be described later.

)p|,~/0 2 )p|,~/u

%A)f 2 )p ? , ? /X 2 )p|,-/,& /

4atPu

(3.10)

The % function is the most important influence on behaviour of the diffusion process. Equation 3.11 lists the three most commonly used functions. Figure 3.25 plots these three functions with respect to the intensity difference, v , for &2 . Note that the functions have been
scaled to fit them together. Only the relative weights for a particular function are meaningful.
w
x

.PO QFdQ

Bd
SOFy
}~
S

%)Bv,&|/

Pv

%)Bv,&|/

{{z
{{ | K z K
{
v)
"<) z / { /

%)Bv,&|/

(3.11)
m
v

&

S
S

54

Chapter 3. Local Segmentation in Image Processing

160

function
Standard
Lorentz
Tukey

140

Relative weight

120
100
80
60
40
20
0

10

15

20

25

30

Intensity difference (d)

Figure 3.25: Anisotropic diffusion weighting functions when 

been scaled for comparison purposes.

. Each function has

The Standard formulation forces a pixel value toward the intensity of each neighbour, at a
rate directly proportional to its intensity difference from that neighbour. Equation 3.12 shows
that Standard diffusion can be interpreted as static linear filter. Equal averaging of all five
pixels in the local window occurs when

)p|,~/

#- .

N 2 )p ? ,~ ? /X 2 )|,/

- . 4tPu
$

2
)
"b
~/ )|,-/u
2 p) ? ,~ ? /
- 4 t]u

)p|,-/u


(3.12)
(3.13)

The problem with Standard diffusion is that, if left for many iterations, the denoised image
would become completely homogeneous. In a fashion similar to the GIWS in Section 3.4.5,
all neighbouring pixels receive some weight, no matter what intensity they have. Therefore,
information from unrelated segments can diffuse across segment boundaries.

3.4 Denoising

55

The Lorentz formulation attempts to reduce pixel leakage between segments by slowing the
diffusion process at some point, defined by the parameter & . This parameter provides a
sense of how different pixel values can be while still being in the same segment. The plot
of the Lorentz weighting function in Figure 3.25 reaches a maximum when G
v 
1& , and
decays thereafter. The decay is quite slow, so a large difference will still be given a significant weight. Thus pixels from different segments can still influence each other, eventually
producing a homogeneous image.
The Tukey formulation takes the Lorentz idea further and completely shuts down diffusion
when v $& , giving maximum weight when vI&|x  . A neighbouring pixel value deviating more than & intensity levels from the centre pixel will have no influence whatsoever. In
local segmentation, the & parameter may be considered proportional to the natural level of
variation of pixel values within a segment. For the simple case of constant intensity segments
with additive noise, this would relate to the standard deviation of the noise. The use of the
Tukey function ensures that the final image will never become a homogeneous image, with
the side-effect of being unable to preserve contrast differences less than & .
The

parameter controls the speed at which diffusion occurs. It is recommended that

be

made small and the number of iterations large, as to more closely approximate the theoretically continuous diffusion formulae. It must be noted that the diffusion process is very slow,
with thousands of iterations being common. This also usually requires that all intermediate
results be kept to floating point accuracy.

Adaptive smoothing
Adaptive smoothing [SMCM91] claims to be strongly related to anisotropic diffusion and
GIWS. It was designed to denoise both intensity and range images. Instead of using the
intensity difference to weight neighbouring pixels, a more thorough estimate of the gradient
is used, shown as the

operator in Equation 3.14. The parameter & is user-supplied, and

controls the scale at which the filter works.

(T)U? C, ? /G,

<
>
R L J < L J >
KM K

(3.14)

56

Chapter 3. Local Segmentation in Image Processing

From a local segmentation perspective, the gradient term discourages the incorporation of
pixels from different segments, as the gradient between them would cross a segment boundary and therefore be very high. Unfortunately, gradient functions are usually too coarse too
accurately control smoothing around fine image details [Smi95].

SUSAN
SUSAN was introduced in Section 3.3.3 in regard to its use of local segmentation principles
for edge and corner detection. SUSAN gathers the set of neighbouring pixels which are
most similar in intensity and closest in position to the centre pixel. This set is called the
USAN. The topographic pattern revealed by plotting the size of each USAN may be used to
determine the position of edges. Smith realized that a weighted average of the pixels within
an USAN form a denoised estimate of the centre pixel [Smi95, Smi96, SB97]. Unlike the
denoising algorithms described so far, SUSAN exploits both the similarity of pixel intensities
and their spatial proximity to the centre pixel. Equation 3.15 describes the resulting SUSAN
weighting function, where & is a spatial threshold and is a brightness threshold. The centre
pixel is excluded from the average.

(*)U? C, ? /GI,

<
>
< >
J C< K L J >K L J < L J >5  K

KM K

"m

(T)f-,.B/w

(3.15)

Due to its exponential-of-sums form, the weighting function is separable into its spatial and
intensity components. Ignoring that fact that the centre pixel receives no weight, the spatial
component is an isotropic 2-D Gaussian distribution with variance & { . Figure 3.26 plots
this function for &

V , the default value used by Smiths implementation 2. By default,

SUSAN uses a 37 pixel circular   mask (see Figure 4.46e). The choice of &G prevents
the Gaussian spatial weighting function from extending very far. Thus the weights for the
closest and furthest pixels differ by less than a factor of two when using a 37 pixel mask.
The intensity difference component is the positive half of a Gaussian function with variance
{ xB
. Figure 3.27 plots this function for


F , the value Smith claims works well for all

images [Smi95]. SUSANs Gaussian intensity weighting function effectively cuts off after
2

Available from http://www.fmrib.ox.ac.uk/ steve/susan/

3.4 Denoising

57

Relative spatial weight

1
0.9
0.8
0.7
0.6
0.5

-3

-2

-1
x

3 -3

-2

-1

2
y

Figure 3.26: The SUSAN spatial weighting component when L , ignoring that X]h .

three standard deviations. This corresponds to a intensity difference of  x

42 when

, which is

^
F . SUSANs intensity weighting is much closer to a hard cut-off than the

hyperbolic one used by GIWS, already seen in Figure 3.23.

Figure 3.28 gives an example of SUSAN denoising a  window. Due to the smaller mask,
& was reduced to 2. For this example

E
4 was chosen, resulting in a good filtered value

of k I-m
; . The SUSAN algorithm is very sensitive to the choice of . If the default values
had been used, the denoised estimate would have been about 6. If the slightly lower

Y

was used, it would have been 3.02. If is chosen poorly, it seems that SUSAN introduces
noise into its filtered estimate. My experiments have shown that should be proportional to
the natural level of variation within the image. For the case of additive zero-mean Gaussian
noise with variance l { , setting $l works well.
When denoising, SUSAN gives no weight to the centre pixel. This has the advantage of
helping to suppress impulse noise, but the disadvantage of reducing the amount of smoothing when the centre pixel is legitimate. SUSAN has an additional mode which has not been

58

Chapter 3. Local Segmentation in Image Processing

0.03

0.025

Weight

0.02

0.015

0.01

0.005

10

15

20

25

30

35

40

45

50

Intensity difference

Figure 3.27: The SUSAN intensity difference weighting component when b; .

3
3
8

3 8
3 8
8 8

0.78 0.88 0.78

0.88
0
0.88
0.78 0.88 0.78

1.00 1.00 0.02

1.00
0
0.02
0.02 0.02 0.02

j
k #-m
;

Figure 3.28: SUSAN denoising: (a) original pixels; (b) spatial weights, unnormalized, LP ;
(c) intensity difference weights, unnormalized, PF ; (d) denoised value.

mentioned yet. When the weights for all the neighbouring pixels are too small (corresponding to a weak USAN), SUSAN ignores them and uses the median of its eight immediate
neighbours instead. In terms of local segmentation, this mode is invoked when no pixels
appear to belong in the centre pixels segment.
SUSANs denoising outperforms the box filter, the median, GIWS, and the Saint-Marcs
adaptive smoothing method [Smi95]. It takes the intensity weighting idea from GIWS, but
modifies it to behave more like a hard cut-off function. Although the intensity weighting
function is continuous, it effectively cuts off to zero for differences greater than 

or so.

The introduction of a spatial weighting component, ignored by most denoising algorithms,

3.4 Denoising

59

also helps it to perform better. The main drawback is that the brightness threshold must
be supplied. Despite this, SUSAN is one of the best and fastest local denoising algorithms
available, in part due to its implicit application of the local segmentation principle.

3.4.6 Filtering based on global segmentation

It is possible to use a global segmentation algorithm to perform image denoising. Imagine a
piece-wise constant image corrupted by additive noise, and a segmentation algorithm which
can segment it correctly. The average of the pixels in each segment should be close to the
constant value that originally coloured each segment. By replacing each pixel with the mean
of the segment it belongs, the image is denoised. The problem with this scheme is simply
the difficulty of achieving perfect global segmentation.
Hussain et al [HR94] take a hybrid approach, using a combination of a Gibbs-Markov random field and relaxation labeling to isolate segment boundaries globally. A local filter is
then used to average the centre pixel with neighbours on the same side of a boundary. In
regions with no boundaries, all pixels are averaged and thus it behaves exactly as the box
filter in Section 3.4.3. Without providing numerical results, they claim to have better subjective results when compared to simple averaging. Unfortunately, when the high noise is high,
their algorithm suffers because the distinction between noise and structure is less obvious.
It is unclear why they did not just average all the pixels from global segments rather than
take local averages. I expect it is due to the fact that most large segments will exhibit some
spatial variation in intensity. In these cases a local average could be a better estimate of the
underlying pixel value when compared to a segment-wide average.
Watershed segmentation has also been adapted to denoising [HH95, HH99]. The image is
first segmented into catchment basins. These basins are smaller subsets of larger segments,
and are represented by the mean of the pixels they contain. For a given basin,

, those

neighbouring (spatially connected) basins with means considered close enough to s mean
are incorporated into . The pixels in are then set to the aggregated mean. The criteria for
closeness can vary. In the early paper [HH95], the closest 50% neighbouring basin means
are used, whereas later [HH99], only those means within some difference threshold,  , were
incorporated. They claim to outperform various filters, including anisotropic diffusion. The

60

Chapter 3. Local Segmentation in Image Processing

only test image used was a simple 3-class piece-wise constant synthetic image containing
very high contrast edges corrupted by additive zero-mean Gaussian noise with variance F { .
They admit their technique depends on the choice of  , which is related to the noise level.
The challenge is to estimate accurately the natural intensity variation one might expect to
observe within each segment.

3.5 Conclusions
Image segmentation algorithms attempt to partition an image into separate groups of related
pixels, called segments. Pixels are usually considered related if they have similar values,
or are located near each other. Clustering is a form of segmentation which ignores spatial
information. Segmentation is considered and essential component of image understanding
systems. Although segmentation of arbitrary images is inherently difficult, the principles
described in this chapter have been applied successfully to many tasks.
Local image processing algorithms assume that most of the information about a pixel can be
found within the small neighbourhood of pixels surrounding it. Local algorithms are often
used for low level tasks, such as denoising and edge detection. The local neighbourhood may
be considered a tiny image, but with different expected properties. It contains fewer pixels,
is made up of fewer distinct regions, and has a higher proportion of pixels in the margins.
Segmentation is a fundamental image processing task. Pixels within segments are homogeneous with respect to some predicate, and hence can be used to denoise the image while
preserving structure. The boundaries between segments correspond directly to discontinuities in the image, hence edge detection is achievable in the same framework. It would
seem obvious to apply these ideas directly to the sub-images used in local image processing.
Strangely, much of the literature seems to treat local techniques differently, preferring to use
ad hoc methods instead, which usually require image dependent parameters to be supplied,
or estimated from training sets.
Image denoising algorithms are a good basis for examining the state of the art for modeling
images on a local level. The SUSAN algorithm is currently one of the best local, one-pass
image denoising algorithms available. During its initial development, SUSAN used a form

3.5 Conclusions

61

of explicit local segmentation. Pixels similar enough to the centre pixel were assimilated
using a hard cut-off function, and then averaged. But rather than refine the local segmentation criterion used, SUSAN moved to an implicit local segmentation, admitting all the local
pixels to the local segment, but allowing their contribution to vary. This formulation made it
difficult to link the brightness threshold parameter to a specific image model.
This thesis advocates a progression toward an explicit local segmentation model, via the
suitable application of global segmentation techniques to local image data. The local segmentation principle states that the first step in processing a pixel should be to segment the
local region encompassing that pixel. The segmentation algorithm should automatically determine the number of segments present in an efficient manner, and any parameters should be
estimated from the image itself. The information obtained from local segmentation should
sufficiently separate the noise and structure within the region. The structural information
may then be used to achieve goals such as denoising and edge detection.

Chapter 4
Local Segmentation applied to
Structure Preserving Image Denoising
4.1 Introduction
The local segmentation principle may be used to develop a variety of low level image processing algorithms. In this chapter it will be applied to the specific problem of denoising
greyscale images contaminated by additive noise. The best image denoising techniques attempt to preserve image structure as well as remove noise. This problem domain is well
suited to demonstrating the utility of the local segmentation philosophy.
A multilevel thresholding technique will be used to segment the local region encompassing
each pixel. The number of segments will be determined automatically by ensuring that the
segment intensities are well separated. The separation criterion will adapt to the level of additive noise, which may be supplied by the user or estimated automatically by the algorithm.
The resulting segmentation provides a local approximation to the underlying pixel values,
which may be used to denoise the image.
The denoising algorithm presented is called FUELS, which stands for filtering using explicit
local segmentation. FUELS differs from existing local denoising methods in various ways.
The local segmentation process clearly decides which pixels belong together, and does so
democratically, without using the centre pixel as a reference value. If the computed local
63

64

Chapter 4. Denoising with Local Segmentation

approximation suggests changing a pixels value by too much, the approximation is ignored,
and the pixel is passed through unmodified. The fact that each local approximation overlaps
with its neighbour means that there are multiple estimates for the true value of each pixel.
By combining these overlapping estimates, denoising performance is further increased.
FUELS will be shown to outperform state-of-the-art algorithms on a variety of greyscale
images contaminated by additive noise. FUELS worst case error behaviour will be shown
to be proportional to the noise level, suggesting that it is quite adept at identifying structure
in the image. The denoised images produced by FUELS will be seen to preserve more image
structure than algorithms such as SUSAN and GIWS.

4.2 Global image models

An image model is a mathematical description of the processes affecting the final pixel values
in an image. As described in Section 2.5, these processes may include atmospheric effects in
the scene, noise in the capture device, and quantization of pixel values. It may be possible to
construct a model which accounts for all these steps, but it would probably consist of many
difficult to determine parameters. Often a simpler model can incorporate the main factors
and still achieve good results.
Modeling the steps from acquisition to a final digital image is sufficient, but not necessary,
for successful image processing. In many cases the full acquisition history may not even
be known. In any case, some assumptions or prior beliefs regarding the properties of the
original scene are also required. The characteristics of the light source and the object surfaces will influence greatly how we expect the image pixel values to be distributed spatially
and spectrally. The number, distance and size of objects in the scene will also affect the
proportion, scale and sharpness of edges in the image.
For example, consider the artificial images in Figure 4.11 . The first is greyscale light intensity
image of a simple object. The second image is also of the same object, except that the
intensity of each pixel represents the distance from the camera to each point in the scene.
1

These images are part of the Range Image Segmentation Comparison Project at the University of South
Florida. http://marathon.csee.usf.edu/range/seg-comp/images.html

4.2 Global image models

65

This is an example of a range image, for which the light source is actually a distance
measurement device. It has been shown that pixel intensities in a range image usually vary
in a spatially linear manner, due to the polyhedral nature of most objects [KS00]. However
this may not be an appropriate assumption to make when analyzing low resolution fingerprint
images in a criminal database. There one would expect many ridges and high contrast edges.
This thesis focuses on greyscale light intensity images. The generalization to other types of
images is considered in Chapter 6.

Figure 4.1: Two images of the same scene: (a) light intensity; (b) range, darker is closer.

4.2.1 The facet model

For greyscale image data, variations on the facet model [HW81] are most commonly used.
The facet model assumes that a digital image is a discrete approximation of a piece-wise
continuous intensity surface. Each piece is disjoint and is called a facet or segment. The
underlying pixel values in each facet are assumed to follow a two dimensional polynomial
function. This polynomial fit can be of any order. Figure 4.2 provides an example of a one
dimensional discrete signal approximated by increasingly higher order polynomial curves.
The simplest facet approximation is a zeroth order polynomial of the form w)|,/  ,
where

is a constant. Each pixel in a constant facet is assumed to have the same value.

For greyscale data this would be a scalar representing the intensity, but for colour data it
would be an RGB vector. An image containing constant facets referred to as being piece-

66

Chapter 4. Denoising with Local Segmentation

Figure 4.2: One dimensional polynomial approximation: (a) the original signal; (b) constant;
(c) linear; (d) quadratic.

wise constant. Piece-wise constant image models are commonly used in image processing.
They only have one parameter to estimate, and are simple to manipulate.
First order polynomial approximation in two dimensions has the mathematical form of a
plane, namely w)p|,~/JruLu! . An image containing facets of this type is piece-wise
planar. Pixel values in a planar facet are linearly dependent on the their position within
the facet. Planar facets are more flexible than constant facets, but at the expense of having
needing three parameters to be estimated for them. If and ! are small enough, and we are
concerned only with a small area within a larger planar facet, then a constant approximation
may be sufficiently accurate.

4.2.2 Generalizing the facet model

The facet model is the basis for the more general segmentation-based approach to image
modeling. The main difference is that the requirement for a polynomial approximation is
relaxed. Instead, the interior of a segment may be modeled using any mathematical description which defines a homogeneity criterion for the segment. For example, texture is often
modeled, not as a functional surface, but as a low order Markov model [Hua84]. A first order Markov model is probability distribution over pairs of pixel values. This model is quite
different from a polynomial one, but it can still be used to define a homogeneity criterion.
An important attribute of any homogeneity criterion used in image segmentation is whether
or not it takes spatial coherence into consideration. As discussed in Section 3.2.1, techniques
such as thresholding consider only pixel values, and not their relative spatial positions. This
can produce unconnected segments, which may or may not be desirable.

4.2 Global image models

67

4.2.3 Image sampling

Consider the simple undigitized monochrome scene in Figure 4.3a. It consists of a square
object of uniform high intensity on a dark background of uniform low intensity, and is perfectly piece-wise constant. Figure 4.3b illustrates the discretization of this scene into a T
B

digital image. The intensity of each pixel in the final image, shown in Figure 4.3c, is set to
the average light intensity of the area each pixel covers in the original scene. The resulting
image is also piece-wise constant, consisting of two distinct segments. This is because the
boundary between object and background coincided with the alignment of the pixel grid.

Figure 4.3: Discrete sampling of an object aligned to the pixel grid: (a) original scene;
(b) superimposed sampling grid; (c) digitized image.

Figure 4.4 shows what occurs when an object in a scene does not align exactly with the sampling grid. The sampling process has produced a range of pixel intensities in the digitized
image. Each pixel on the object boundary has received an intensity which is a blend between the intensities of the two original segments. In fact, there are now seven unique pixel
intensities compared to the original two.

Figure 4.4: Discrete sampling of an object mis-aligned with the pixel grid: (a) original scene;
(b) superimposed sampling grid; (c) digitized image.

68

Chapter 4. Denoising with Local Segmentation

Most observers would still assert the existence of only two segments, but would have some
difficulty assigning membership of each boundary pixel to a specific segment. One interpretation is that those pixels with intermediate intensities have partial membership to both
segments. If an application requires a pixel to belong only to one segment, that segment
in which it has maximum membership could be chosen. Alternatively, it could be claimed
that the image is still piece-wise constant, but now consists of ten segments. These differing
interpretations highlight the fact that an image model for a continuous scene may no longer
apply to its digitized counterpart.

4.2.4 Edges and lines

Edges are the boundaries between segments. Lines may be regarded as very narrow segments, having two edges situated very close together. Figure 4.5 gives a one dimensional
example of two types of edges and the two corresponding types of lines. The image models
used for edges and lines are not unrelated to the models used for the interior of segments.

Figure 4.5: (a) step edge; (b) line; (c) ramp edge; (d) roof.

The step edge defines a perfect transition from one segment to another. If segments are piecewise constant and pixels can only belong to one segment, then a step edge model is implicitly
being used. If a segment is very narrow, it necessarily has two edges in close proximity. This
arrangement is called a line. An arguably more realistic model for edges is the ramp edge. A
ramp allows for a smoother transition between segments. This may be useful for modeling
the blurred edges created from sampling a scene containing objects not aligned to the pixel
grid. Two nearby ramp edges result in a line structure called a roof.
Edge profiles may be modeled by any mathematical function desired, but steps and ramps
are by far the most commonly used. If a ramp transition occurs over a large number of pixels,

4.2 Global image models

69

it may be difficult to discriminate between it being an edge, or a planar facet. If pixels along
the ramp are assigned to a particular segment, the values of those pixels may be dissimilar to
the majority of pixels from inside the segment.

4.2.5 A useful noise model

As described in Section 2.4, different types of noise may be introduced at each step of the
image acquisition process. The general functional form for the noise component, 0)|,-/ , at
each pixel is given in Equation 4.1. Parameters and are the spatial position of the pixel in
question,




is a vector of fixed parameters determining some properties of the noise such as

its intensity and spread, and = is the original image which may be required if the noise term
is data dependent. This formulation has scope for a huge number of possible noise functions.

(4.1)

w)p|,0+ ,w=/

A simple, but still useful and versatile noise model is additive zero-mean Gaussian noise
which is independently and identically distributed (i.i.d.) for each pixel. Under this model
the noise adds to the original pixel value before digitization. The noise term may be written
like Equation 4.2, where 698:)=#w,& { / denotes a random sample from a normal distribution
of mean # and variance & { . Figure 4.6 plots the shape of this noise distribution when & { O
.

w)p|,T& { /0698

)f ,& { /

(4.2)

Because the noise is additive and symmetric about zero, it has the desirable effect, on average, of not altering the mean intensity of the image. It only has one parameter, the variance
& { , which determines the spread or strength of the noise. Although the work in this thesis

assumes that the noise variance is constant throughout the image, it would be possible to
vary it on a per pixel basis. This and other extensions are discussed in Chapter 6.
Consider a constant facet containing pixels with intensity 5 . After 8

)f-,~& { / noise is added,

it is expected that 99.7% of pixels will remain in the range 5T3F& . This is called the 1&

Chapter 4. Denoising with Local Segmentation

Pr(x)

70

-4

-3

-2

-1

Figure 4.6: The standard Gaussian distribution: ]

confidence interval for 5 [MR74]. Examination of Figure 4.6 shows that very little probability remains for values of 5 outside the confidence interval. Table 4.1 lists the number of
standard deviations from the mean for which a given proportion of a normally distributed
data is expected to lie.
Fraction of
data (%)
50.0
68.3
90.0
95.0
95.4
98.0
99.0
99.7

Number of standard
deviations from mean
0.674
1.000
1.645
1.960
2.000
2.326
2.576
3.000

Table 4.1: Confidence intervals for normally distributed data.

Figure 4.7 shows the effect of two different noise variances on the square image, which
has segments of intensity 50 and 200. When &

 , the two segments remain distinct.

When & is quadrupled, the square obtains some pixels clearly having intensities closer to the
original background intensity. After clamping, the F& confidence interval for the background
is FR

17E)f-,;
F/ , and
1BR

FCE)+B-,
FBB/ for the square. These limits have an overlap

4.2 Global image models

71

of 90, or approximately
1& . On average, this situation would result in about 5% of pixels
being closer to the opposing segment mean. For the
B
"R square image this would affect
about 5 pixels, roughly corresponding with what is observed.

Figure 4.7: The effect of different Gaussian noise levels: (a) no noise; (b) added noise
L. ; (c) added noise  .

For piece-wise constant segments, the noise standard deviation defines a natural level of
variation for the pixel values within that segment. The natural level of variation describes
the amount by which pixel values may vary while still belonging in same segment. For the
case of planar segments, the natural level of variation depends on both the noise level and
the coefficients of the fitted plane. If a global planar segment only has a mild slope, then
the variation due to the signal may be negligible for any local window onto that segment.
The natural level of variation in the window will be dominated by the noise rather than the
underlying image model.

4.2.6 Pixel quantization

Section 4.2.3 discussed the side-effects of sampling a continuous scene on a discrete image
grid. The sampled pixel intensity is computed as an average or integral over a small area of
the scene. This intensity must also be quantized if it is to be stored in digital form [GN98].
The number of bits per pixel (bpp), , is typically around 8 to 16. Most everyday images
use 8 bpp per band, but medical and scientific applications usually store data with higher
accuracy. The number of unique intensities that can be stored in bpp is _\

, which

comes to _<I
B1 for 8 bpp images.
Consider the one-dimensional signal drawn with a dotted line in Figure 4.8. Quantization
to 8 levels produces the signal plotted with the solid line. Due to the rounding processing,

72

Chapter 4. Denoising with Local Segmentation

a quantized value, 5 , could have originally had any value in the range 5 $-4 ,.5 uI-F/ .
This quantization noise may be approximated using the standard deviation of a uniform
distribution of width 1, which is z
x

, or about 0.29. Usually the other forms of noise are
at least an order of magnitude higher than this, so quantization noise can safely be ignored
by most applications.

Intensity
Original
7

Quantized

6
5
4
3
2
1
Position

0
Figure 4.8: Effect of quantizing the signal intensity to 8 levels.

A situation where one may not wish to ignore quantization noise is when magnifying an
image by interpolating pixel values. By default, the midpoint of the quantization region
would be used as the true value of the pixel, but it is possible to choose the true value from
anywhere within the original quantization interval. The most appropriate value could be
decided by optimization of a suitable cost function, for example, to maximize the sharpness
of the interpolated image.

4.2.7 A suitable image model

The aim of this chapter is to demonstrate a simple form of local segmentation applied to
denoising greyscale photographic-type images. Equation 4.3 mathematically describes the
process of going from an undigitized noise-free image, = , to a noisy quantized version, =1? .

? )p|,~/0+w)|,/uA8

)f-,& { / u9 h

(4.3)

4.2 Global image models

73

After quantization there is a final clamping step to ensure the pixel values fall within the
legal range. Equation 4.4 describes the clamping process, which takes the noisy quantized
image, =@?0 , and produces the final image, =;? , available to the image processor.

? )p|,-/0!=TG1)|,/Z

_

? )|,-/

? )|,-/

? )|,-/_

(4.4)

FRS@dy"S

To that end a global image model with the following properties will be assumed:

1. The image is composed of disjoint segments.

2. Segment boundaries coincide with pixel boundaries .
3. Each pixel may belong to one segment only.
4. A segment interior may be well approximated by a constant grey level, particularly at
a local level.
5. Each pixel is corrupted by additive zero-mean Gaussian noise.
6. The noise variance is common to all pixels (and hence segments).
7. Quantization noise may be ignored.
8. Saturation effects from pixel clamping are to be ignored.

An image model assuming piece-wise constant greyscale segments with additive noise is
probably the simplest, but still useful, image model to use. In Chapter 6, the extension of the
principles outlined in this chapter to more complex image models will be discussed.

4.2.8 Image margins

When processing a whole image with dimensions < in a local manner, each pixel )p|,-/
is processed in turn using only its immediate pixel neighbourhood. Difficulties arise when

74

Chapter 4. Denoising with Local Segmentation

processing pixels inhabiting the margins of the image, as these pixels have incomplete neighbourhoods. Figure 4.9 illustrates this situation for w)f-,/ when using a  window. The 5
pixels within the dotted window labelled ? are denoted as missing, and the remaining 4 are
denoted as available pixels.

?
?

Figure 4.9: Incomplete neighbourhoods when processing pixels at the image margins.

There are various approaches to handling this situation:

Constant fill In this scheme, missing pixels are replaced with a fixed constant value, most
commonly zero or mid-grey.
Wrap around The image is treated as a torus, as if the upper and lower, and left and right,
edges were spatially proximate. This is achieved by treating pixel coordinates in a
modulo fashion )pjuWL Q*',*u<3R^Q*/ . Note that vu

is used to handle

negative coordinates.
Nearest neighbour Missing pixels take on values equal to their nearest available neighbour.
Average fill The missing pixels are set to the arithmetic mean of the available pixels.
Specialization The algorithm is specifically modified to handle the special cases where the
neighbourhood contains fewer pixels than normal.
The constant fill and wrap around methods are clearly inferior because they invent pixel
values bearing little relevance to the available pixels. The specialization method is desirable, but even when using a  window there are eight special cases to handle (four sides
plus four corners), which can make it difficult to implement efficiently.

4.2 Global image models

75

Figure 4.10 pictorially describes the nearest neighbour approach. This technique is simple
to implement, and possesses the ability to preserve certain structure in the image. For example, if the available pixels near the margin form two separate populations, the extrapolated
neighbourhood would too. The disadvantage is that some pixel values, such as the corners,
are duplicated more than once, and hence over-represented in the extrapolated set. This can
bias the results, especially if the available pixels are very noisy.

Figure 4.10: Missing pixels are easily replaced with their nearest neighbours.

The average fill method chooses values for the missing pixels which are closest, in a least
squares sense, to the available pixels. This is desirable when the available pixels are homogeneous. However, when the available pixels are heterogeneous, this method would invent
a third class of pixels all having the same value. It also requires more computation that the
nearest neighbour approach.
In this thesis the average fill method will be used. The averaging process suggests it would
be more robust under noisy conditions when compared to nearest neighbour. It is hoped
that it should have little effect on the overall results compared to specialization, as the
margin pixels comprise only a small proportion of the total number of pixels in the image.
Consider a b

square window applied to an  image. In Figure 4.11 the shaded area

contains pixels which have an incomplete neighbourhood. The outer dotted line denotes the
effective increase in image resolution when extrapolating missing pixels.
Equation 4.5 gives an expression for

, the ratio of missing pixels to the total number of

pixels in the image. This is the proportion of pixels, which, when at the centre of the window,
are forced to generate missing pixel values using one of the techniques just described.

76

Chapter 4. Denoising with Local Segmentation

(m1)/2

(m1)/2

(m1)/2

(m1)/2

Y+m1

B%

X+m1

Figure 4.11: The pixels involved in border cases are shaded.



)=
/)p

u9I(Eu$
/
A

)=:wOT QRQ B~mbPSOgS@]Sd./

(4.5)

Table 4.2 lists the proportion of affected pixels for various common combinations of image
and neighbourhood dimensions. For the commonly used  window, about 1% of processing involves missing pixels. This is not large enough to be a serious concern, especially
given the fact that the objects of interest are often positioned near the centre of the image,
away from the margins.

4.3 Measuring image similarity

In most of the experiments performed in this thesis, an original image, = , has controlled
amounts of noise, > , added to it to produce a noisy version, = ? . An example is given in Figure 4.12. A denoising algorithm produces an estimate, = , of the original image. In structure
preserving filtering it is desired that only noise, and not image structure, be removed. Thus
the aim is to produce = as close as possible to = .

4.3 Measuring image similarity

77

3
3
3
5
5
5
7
7
7

320
512
1280
320
512
1280
320
512
1280

240
512
960
240
512
960
240
512
960

1.45%
0.78%
0.36%
2.90%
1.56%
0.73%
4.33%
2.33%
1.09%

Table 4.2: Proportion of pixels at the margins for various mask and image sizes

Figure 4.12: Noisy equals original plus noise: (a) original image, ; (b) additive noise, ;
(c) noisy image, ? = +

There are two different situations that may occur when measuring the quality of a denoised
image: when the original image is available, and when it is unknown. The first case usually
occurs in an experimental situation where a known image is artificially corrupted with noise.
The original image is ground truth, to which any denoised images may be compared directly.
The second case is the more realistic situation whereby a noisy image has been sourced, say
remotely from a space telescope, and one wishes to denoise it before further processing.
Here there is no ground truth with which to compare. In either case, there are subjective and
objective techniques for assessing the quality of denoised images.

4.3.1 Visual inspection

Visual inspection is a subjective process whereby a human viewer assesses the quality of an
image. In the case where ground truth is available, one or more assessors may perform a side
by side comparison of the denoised and original images. The comparison is usually rated us-

78

Chapter 4. Denoising with Local Segmentation

ing predefined quality classes, such as excellent, fine, passable, marginal, inferior
and unusable [GW92]. This type of experiment requires that the viewing conditions be
strictly controlled. This includes factors like ambient lighting levels, the display hardware,
and the position of the assessor relative to the displayed images.
Figure 4.13 shows samples of what an assessor may be asked to examine. For this example
a reasonable assessment for the denoised image may be passable, as there is blotchiness
in the foreground and background, and the top left corner of the bright square stands out as
being too dark. Obviously though, what is passable to one may be marginal to another,
and so on. However, most assessors will be consistent in their gradings, and a consensus
rating can usually be determined.

Figure 4.13: Visual assessment using ground truth: (a) original image; (b) denoised image.

When ground truth is unavailable, it is still possible to perform a visual comparison by inspecting the noisy and denoised images. Figure 4.14 shows this situation for the example.
Even without ground truth the assessor usually has some a priori beliefs on what features
are present in the original image, and the denoised image can be examined relative to these
beliefs. In effect, the assessor is determining the ground truth himself or herself. The human brain is quite good at identifying structure behind noise, so a skilled assessor could
use a noisy image as a guide to what is expected in the denoised output. The qualitative
descriptions described earlier can still be used.

4.3.2 Traditional quantitative measures

An objective measure for the quality of a denoised image would be very useful. One
traditional measure for the closeness of two data sets is the root mean squared error, or

4.3 Measuring image similarity

79

Figure 4.14: Visual assessment without ground truth: (a) noisy image; (b) denoised image.

RMSE [GW92]. When ground truth is available, the two data sets could be the denoised
image, = , and the original image, = . Equation 4.6 calculates the RMSE in this case.

{
wk )|,-/}bw)|,/ `

(4.6)

The RMSE is proportional to the disparity between two images. In the case of two equivalent
images, it is zero. For the case of additive zero-mean Gaussian noise, the RMSE between
the noisy and original images is exactly equal to the noise standard deviation.
The peak signal to noise ratio, or PSNR, is derived from the RMSE, and is measured in
decibels (dB) [RJ91]. This logarithmic measure is computed using Equation 4.7, where
_
is the maximum possible pixel intensity.

0i

#
F0B

_

I

(4.7)

Q.

RMSE and PSNR use the square of the pixel difference. This penalizes large errors very
heavily. For example, a single pixel in error by 100 will have the same contribution as
10,000 pixels in error by 1. An alternative measure, which aims to alleviate this potential
problem, is the mean absolute error, or MAE, calculated using Equation 4.8. The MAE
penalizes errors by their magnitude, and is less likely, compared to RMSE, to be biased by
occasional large errors.

E

A

w)p|,-/}

k )p|,-/

(4.8)

80

Chapter 4. Denoising with Local Segmentation

The RMSE and MAE are useful in that they provide a number which can be compared
objectively. Their drawback is that they do not take into account the spatial distribution of
the pixel differences. Many small differences may be more tolerable than fewer larger errors,
especially if those errors occur at busy locations in the image. In fact, this perceptual
masking effect [TH94] is exploited by lossy image and audio compression algorithms. Large
errors clumped together, or near the image margins, may be preferred the same errors spread
around the image. There is evidence to suggest that RMSE may be well correlated to human
observers subjective opinions [MM99]. This fact, combined with its simple formula, has
allowed RMSE to be used widely throughout the literature.
The worst case absolute error, or WCAE, is the magnitude of the single largest difference
between two images. It provides a measure of a denoising algorithms worst case performance. The calculation of the WCAE is given in Equation 4.9.

HXO

argmax
w)|,-/} w
k )|,/
- 4

(4.9)

In an experimental situation, artificial noise is added to an original image to create a noisy

version thereof. Difficulties can occur if the original image already contains some noise. A
good denoising algorithm is likely to remove both the original and artificial noise components, producing a denoised image which is less affected by noise than the original image
itself. In this situation one can not expect the denoised image to be the same as the already
noisy original. The original noise level is usually low compared to the artificial noise being
added, and can be safely ignored. Images generated by synthetic means, such ray-tracing
software [Ama87], can be made completely noiseless, not including quantization.

4.3.3 Difference images

A difference image has pixels representing the numerical difference between two images. If
-,_$
@ is the possible range of intensity values in the two images being compared, then

the range of the difference image will be _u#
B,_
@ . To view a difference image, the
pixel values must be mapped and clamped back to the legal range -,_
@ . Equation 4.10
computes the difference image, , between two equal-sized images, = and = { . If the two

4.3 Measuring image similarity

81

images are similar, the difference image mean should be 0, so the _xB
term centres 0 at the
middle grey level. The clamping process has the side-effect of not being able to distinguish
errors with a magnitude greater than _[xB
, but these should be rare.

A!=T

=[P= u

(4.10)

Figure 4.15 shows how a difference image can be used to compare a denoised image to
the original. The dark and light pixels in the difference image show where the denoising
algorithm did a poor job of estimating the original pixels. The mid-grey pixels around the
centre area represent where it did well. If the image were denoised perfectly, the difference
image would contain exactly the noise which was added.

Figure 4.15: Qualitative measure of filter performance: (a) original image; (b) noisy 
image; (c) median filtered; (d) difference between original and denoised, midgrey representing zero.

In the previous example ground truth was available, as Figure 4.15a was synthetically created. Usually, however, the noiseless image will be unavailable. The image processor must
use only the noisy image to recover the original pixel information. In this situation, only the
difference between the noisy and denoised version can be generated. Figure 4.16 gives an
example of this situation. Although there is less apparent structure in the difference image,

82

Chapter 4. Denoising with Local Segmentation

there are some clumps of very bright and dark pixels near the corners of the square. These
errors could correspond to where the filter performed erratically, or to where there were very
noisy pixels. Without the original image it is difficult to determine which it is.

Figure 4.16: Qualitative measure of filter performance: (a) noisy image; (b) median filtered;
(c) difference between noisy and denoised, mid-grey representing zero.

4.4 Test images

In this chapter, different denoising algorithms will be compared. It is laborious to provide
results for a large set of images at each stage of the discussion. For this reason, one or more
images from the small set introduced here will be used consistently throughout this chapter.
The use of a larger set of image test set will be deferred until the final results are considered.

4.4.1 Square
The square image in Figure 4.17 has already been encountered. It is an 8 bit per pixel
greyscale image of resolution
B
* . It consists of a 25 pixel square of intensity 200 atop
a 74 pixel background of intensity 50. It will be often used for illustrative purposes, and for
subjectively examining the effect of various techniques.

4.4.2 Lenna
The lenna image [len72] in Figure 4.18 has become a de facto standard test image throughout the image processing and image compression literature2 . Its usefulness lies in the fact
2

lenna is available from ftp://nic.funet.fi/pub/graphics/misc/test-images/

4.4 Test images

83

Figure 4.17: The "R 8 bpp square test image.

that it covers a wide range of image properties, such as flat regions, fine detail, varying edge
profiles, occasional scanning errors, and the fact that so many authors produce results using
it as a test image. One interesting feature is that the noise in lenna seems to be inversely
proportional to the brightness, perhaps a legacy of having been scanned from a negative.

Figure 4.18: (a) the B}B 8 bpp lenna image; (b) histogram.

4.4.3 Montage

Figure 4.19 shows a greyscale test image called montage, and its histogram. It has resolution 

 

and uses 8 bits per pixel. The image consists of four quadrants. The top left

84

Chapter 4. Denoising with Local Segmentation

is the middle
BF
BF section of a smoothed version of the lenna image3 . The bottom
right is a right hand fragment of a German village street scene, which contains some small
amounts of natural noise. The top right is a synthetically created image which is perfectly
piece-wise constant. It covers a range of segment boundary shapes and intensity differences
between adjacent segments. The bottom left is the same as the top right, except that the
segments are piece-wise planar, covering a range of horizontal and vertical gradients.

Figure 4.19: (a) the BB 8 bpp montage image; (b) histogram.

It is hoped that this image covers a wide range of image properties, while also being very
low in noise. The low noise level is important for experiments in which synthetic noise will
be added. The image has features such as constant, planar and textured regions, step and
ramp-like edges, fine details, and homogeneous regions. The montage image will be used
for measuring objectively the RMSE performance of various techniques.

4.5 A one segment model

Local segmentation is concerned with segmenting a small window on a much larger image.
The pixel being processed, )p|,-/ , is usually at the centre of the window. For the following
3

This image was provided by Bernd Meyer, the author of the TMW algorithm [MT97]. It is available for
download from http://www.csse.monash.edu.au/ torsten/phd/.

4.5 A one segment model

85

discussion a square  window containing

Y pixels will assumed, as shown in Fig-

ure 4.20. However, the techniques described may be applied to any window configuration
and size. For brevity we will refer to the set of pixels from the window as the vector . The
pixels are indexed in raster order, namely to . The centre pixel -

U41

{3 is the pixel

being processed, and may also be denoted simply , without the subscript.
w)pR
B,7
/
w)p
B,-/
)pR
B,ru<
/

)p|,7
/
)p|,-/
w)|,ru<
/

)pTu$
B,7
/
)pTu$
B,-/
w)*u<
B,u$
/

p
E

q
r

Figure 4.20: Equivalent naming conventions for pixels in the local window.

The simplest form of local segmentation is to do no segmentation at all, and to assume that
the pixels in

are homogeneous. Because the local region is small compared to the whole

image, a large proportion of local regions are expected to be homogeneous anyway. This is
the assumption made by the linear filters with fixed weights described in Section 3.4.2.
Under a piece-wise constant image model, pixels, , from the interior of a global segment all
have the same true value, denoted # . Under a 8

)f-,~& { / noise model, these pixels will have

noise, > , added to them. The noisy pixels, ? , have the properties shown in Equation 4.11.

? uX>

" SdS

#

PORQ

Z698

)f-,& { /

(4.11)

Equation 4.12 shows the noisy pixels to have the same expected value as they did prior to
noise being added. However, the uncertainty, or variance, in their values has increased from
0 to & { . The noisy pixels may be considered to be distributed 698

+ ?


p u9y

Fd;+?

{ ;1d;+-u$
{ ;Fd;yZ

)B#w,& { / .

#Lu9


#

;u<
"& {


& {

For this situation, the optimal least squares estimator, # k , for

#

(4.12)

is the sample mean [MR74],

computed using Equation 4.13. Assuming local homogeneity, the sample mean is also the
best estimate for every pixel within the window. The variance, or standard error of the

86

Chapter 4. Denoising with Local Segmentation

sample mean is & { x;

. Thus the more pixels from the same segment that are averaged, the

more accurate the estimate of their original value is expected to be.

j
k -k #'
k 

1dR k Z

& {

(4.13)

Figure 4.21 shows the effect of this filter on the square image from Section 4.17. There
is clear noise reduction in the centre of the square and in the background, but there is also
obvious blurring at the boundary between them. The assumption of only one segment existing in the local window is equivalent to the box filter described in Section 3.4.2, which gives
equal weight to each pixel in the window. As a result, it has all the same drawbacks as fixed
linear filters. If filtering were repeated, further blurring would occur, eventually producing
a completely homogeneous image. This common pixel value would be similar to the image
mean, which for square, is )
B
FB"u91F/yxFB#B .

Figure 4.21: (a) original; (b) noisy L. version; (c) denoised using averaging.

4.6 A two segment model

It may be thought that because the majority of pixels exist in homogeneous regions, that
treating all windows as such would not overly affect results. Unfortunately, although heterogeneous windows are in the minority, they do contain a lot of visual information in the
form of edges and texture. The human visual system is very sensitive to discontinuities, and
any blurring or alteration to sharp edges would not go unnoticed [Ede99]. It is therefore
important to preserve as much information in these regions as possible.

4.6 A two segment model

87

Reconsider the square image from Figure 4.21. Globally, it consists of only two segments,
thus any local window must contain, at most, two segments. One solution for preserving
edges is to first classify the pixels from the window into two segments. Only pixels belonging
to the same segment could be used to calculate the denoised pixel values. This would prevent
the mixing of pixels from different global segments, resulting in a sharper image.
The division of pixels into two segments is obviously a form of segmentation. The BTC
technique, discussed in Section 3.3.2, successfully uses a two-class model for encoding  J
pixel blocks. In most variants of BTC a single threshold,  , is used to divide the pixels into
two clusters [FNK94]. The representative values, # k and # k { , for each cluster are set equal to
the average of the pixels within each cluster. This is shown in Equation 4.14, where and

{ are the number of pixels in each cluster.

#k

#k

{ 

{ a

(4.14)

The threshold,  , may be chosen in many ways. Many techniques, including the popular
Absolute Moment BTC [LM84, CC94, MR95], simply use the mean of the pixels in the
window, calculated using Equation 4.15.

<

(4.15)

The mean measures the central tendency of the pixel values in the block. When the two
clusters have unequal numbers of pixels, the mean is biased toward the centroid of the larger
cluster. This pollutes the smaller cluster with pixels from the larger cluster, biasing both
clusters means. This effect is shown in Figure 4.22.
When the clusters are known to be normally distributed, it is better to choose  such that the
overall mean squared error (MSE) is minimized, as shown in Equation 4.16. The optimal 
can be found by exhaustive search of the

<

argmin
th- U4


possible binarizing thresholds [DM79].

) ? n# k / { u

) ? # k { / {

(4.16)

Chapter 4. Denoising with Local Segmentation

Frequency

88

Intensity
Biased
1

Overall
mean

Biased
2

Figure 4.22: The mean is a poor threshold when the clusters have different populations.

An iterative technique for determining the MSE threshold in Equation 4.16 was proposed
by Efrati et al [ELM91]. It uses the block mean as the initial threshold. The computed
cluster means are then themselves averaged to produce a new threshold, and new cluster
means computed. This continues until there is no significant change in the threshold. The
algorithm is outlined in Listing 4.1, where quantifies the required accuracy. Usually pixel
intensities have integer values, so the algorithm can terminate if the integer part of  is no
longer changing.

1. Let $

-? .

2. Compute # k and # k { as in Equation 4.14.

%

%
)# k u# k {
/  then exit.
{

4. Let the new $

) #k

u k# { / and go to Step 2.
{

3. If

Listing 4.1: An iterative method for choosing a binary threshold.

4.6 A two segment model

89

This algorithm produces means very similar to the levels produced by the Lloyd quantizer [Llo82], and is often referred to as Lloyds algorithm. In their BTC survey paper, Franti
et al [FNK94] found Lloyds algorithm to produce results identical to a true MSE optimizer
in nearly all situations, all while using only 1.71 iterations on average. It was not made clear
in which situations it failed perhaps it was due to the mean being a poor initial threshold,
resulting in convergence to a local, rather than global, optimum. Nevertheless, this method
will be used for selecting binary thresholds in this chapter.

4.6.1 Application to denoising

Equation 4.17 describes how binary clustering may be used to denoise the pixels which were
clustered. The denoised pixels, k , are set equal to the mean of the cluster in which they
belong. For this clustering algorithm, this corresponds to the cluster mean that each noisy
pixel is closest in intensity to.

k Z

#k
#k

? 

?

(4.17)

Figure 4.23 compares using 1-segment and 2-segment models for denoising square. Filtering under the assumption of two segments preserves the edges of the object much better. The
thresholding procedure appears to have correctly classified the pixels, even in the presence of
additive Gaussian noise. Within homogeneous regions the smoothing is a little worse, and is
especially noticeable in the centre of the light square. This occurs because the pixels in each
window are forcibly divided into two segments, even those which are homogeneous. Thus
the variance within each local cluster is reduced only by a factor of or { , compared to
 uA { when the window is treated homogeneously. The possibility of applying the
filter again to the denoised output for further smoothing is discussed in Section 4.8.
Figure 4.24 quantitatively compares the denoising performance of the 1-segment and 2segment local models for the montage image. For each value of & tested, artificial noise
6I8:)+-,& { / was added to montage, and the noisy pixels rounded to integers. The RMSE

was measured between the original montage and the denoised output of each algorithm.
The 2-segment model performs significantly better for all noise levels up to &ME
F . At the

90

Chapter 4. Denoising with Local Segmentation

Figure 4.23: (a) noisy, <. ; (b) filtered with 1-segment model; (c) filtered with 2-segment
model.

highest noise levels, the distinction between adjacent global segments becomes less clear,
especially if the contrast between them is low. When the noise swamps the signal, inappropriate thresholds are more likely to be chosen.

24
22

Local model
1-segment
2-segment

20
18

RMSE

16
14
12
10
8
6
4
2

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.24: Comparison of 1-segment and 2-segment models for denoising montage.

4.7 Discriminating between two models

91

4.7 Discriminating between two models

Modeling the local region as two segments helped to preserve edges and decrease the RMSE
between the original and denoised images. However, within homogeneous regions denoising
performance decreased. It would be desirable to be able to switch automatically between
one and two segment models as appropriate. The general form of this problem is called
model order selection, and many statistical and information-theoretic techniques have been
developed to tackle it [WF87, Ris87, Pre89, BS94].
If Q is the number of segments in the local window, the model selection problem can be
framed as choosing either Q<
of intensity
#

or Q<
. Consider an image with a dark background

comprising Bx1 of the total pixels, and a light foreground of intensity #

{

comprising the remaining
x1 of the pixels. If this image were corrupted by additive noise
698

)f-,& { / , its histogram would have the form of in Equation 4.18.




)B# ,& { / u



)B# { ,& { /
8

(4.18)

Figure 4.25a shows this histogram for #  , # {  and noise standard deviation &2O
. It
is clearly bimodal, because the two cluster means are F& apart. It was shown in Section 4.2.5
that 99.7% of normally distributed pixels will, on average, fall within F& of their mean.
Figures 4.25b-d show the histograms behaviour as the distance between the two means,
%
#

# , is decreased in multiples of the noise variance. At some point between

1& and
{

F& the histogram becomes unimodal, but not Gaussian, in nature. The exact point at which

this occurs would also depend on the number of pixels in each cluster. The more unequal the
blend, the more likely it is to appear unimodal.
From the previous observations it would be reasonable to suggest that a method for deciding
between Q<a
and Q<V
should be based on determining whether the two clusters are
well separated. If we restrict the criterion to use the difference between cluster means, the
threshold should depend on the noise variance and the number of pixels in each cluster.
Equation 4.19 gives a template for the proposed model order selection technique.

Q

#k

# k
{

F~ S@d"S

)p&,F , { /

(4.19)

92

Chapter 4. Denoising with Local Segmentation

10

11

12

10

11

12

10

11

12

10

11

12

Figure 4.25: [top to bottom, left to right] A mixture of two normal distributions with common
variance, with their means separated by (a)  ; (b) ; (c)  ; (d)  .

The two estimated cluster means, # k and # k { , are calculated using Equation 4.14. The hypothesis that there are two segments is rejected if the cluster means are too close together. The
measure of closeness is decided by comparing the inter-cluster distance with the output of
a threshold function, )+&, , { / . The threshold function may depend on the image noise
variance and the number of pixels in each cluster. This should make the thresholding process
adaptive to the image noise level.

4.7.1 A simple model selection criterion

Figure 4.25 showed that the two populations remained distinct up to a separation of around
B& or so, corresponding to about a 95% confidence interval. There is no exact answer, as

it varies for different cluster populations, but it is obvious that the magnitude of the noise
variance is an important factor. The simplest, reasonable form for the separation threshold is
given in Equation 4.20, where H is a positive constant.

4.7 Discriminating between two models

93

)p&,F ,F { /$HC&

(4.20)

This separation threshold is a form of Fishers criterion (Equation 3.3) under the assumption
of equal cluster variances. The two are related by the fact that H


if &PE&.&E&V) .

The problem is choosing an appropriate value for H .

Visual inspection
Consider the square image again, which has a background intensity of 50 and a foreground
intensity of 200. Figures 4.26a shows a noisy &G
; version of this image. Figures 4.26bf
show the denoised output values of H

equal to 1, 2, 3, 4 and 5 respectively. For each pixel,

if Q:
is chosen, the denoising filter uses the one segment model of Section 4.5. If QvD

is chosen, the two segment model of Section 4.6 is applied. In Figure 4.26, the foreground
and background are separated by
1& , so no difficulties are expected. The denoised images
for H$ are equivalent, having successfully smoothed nearly the whole image.

Figure 4.26: (a) noisy, L. ; (b)(f) simple model selection using 9..~ .

Figure 4.27 provides the same results as Figure 4.26, except that the noise has been increased
to &X
1 . The the two pixel populations are still well separated by  41& . Once again, for
H< , the maximum possible smoothing has occurred.

94

Chapter 4. Denoising with Local Segmentation

Figure 4.27: (a) noisy, LP; ; (b)(f) simple model selection using 9..y .

Figure 4.28 shows a much more visibly noisy image with &b:B , corresponding to only a
1& intensity difference between the two image segments. At this separation, there is a 2%

overlap of the two populations. The best smoothing occurs when H equals 3 or 4. When HO
 , some blurring of segment boundaries is observed, meaning Qj

was sometimes chosen

incorrectly. When the global segment means are only 1& apart, insisting on an equivalent
cluster separation at the local level will sometimes be incorrect.
Finally, Figure 4.29 presents the very noisy case of &PY . The true object intensities are
only -F1& apart, corresponding to a 10% overlap. Thus it is expected that about 10 of the 99
pixels will have intensities which could be mis-classified. The best denoising occurs when
H

equals 2 or 3, depending on what type of artifacts are preferred. As expected, for H# ,

the model selection criterion fails miserably, blurring all the edges around the square.
From this analysis it may be concluded that HO$ is a good all-round choice for the square
image when the noise has additive, zero-mean Gaussian characteristics. The disadvantage
of the H& threshold is that it does not consider the number of pixels in each cluster. There
is low confidence in the average of a cluster containing 1 pixel compared to its 8 pixel
counterpart, because the standard error of the former is 8 times higher. A further problem
with a fixed threshold is that two segments with an intensity difference of less than HC& will
probably never be detected.

4.7 Discriminating between two models

95

Figure 4.28: (a) noisy, LP; ; (b)(f) simple model selection using 9..~ .

Figure 4.29: (a) noisy, L ; (b)(f) simple model selection using 9..~ .

Quantitative results

Figure 4.30 shows the RMSE between the denoised and original montage image for 6
integer values of H . One would conclude that H


 performs best on this image over all

the noise levels. This supports the conclusion made in Section 4.7.1.

96

Chapter 4. Denoising with Local Segmentation

30

25

C=1
C=2
C=3
C=4
C=5
C=6

RMSE

20

15

10

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.30: Effect of on RMSE when denoising montage by switching between a 1segment and 2-segment local model.

4.7.2 Students -test

The simple separation criterion, )/ EHC& , does not take into consideration the number of
pixels in each cluster. As the number of pixels in a cluster decreases, we have less confidence
that the computed cluster mean is a good estimate of the true segment mean. For the  
case, the separation criterion for the situation where both clusters are about the same size,
say O and { I , should be different to the skewed case when 
and { 3 .
In the latter, there is a much greater risk that the single pixel cluster would be mis-classified,
as its mean was derived from a single noisy pixel.
The problem of deciding between QRE
and Q I
may be framed as traditional hypothesis
testing of the equality of two population means [Gro88]. The null hypothesis states that the
two class means are the same, and that only one class exists. The alternative hypothesis states
that the means are not equal, and represent independent sub-populations. The application to
model selection in local segmentation is shown in Table 4.3.

4.7 Discriminating between two models

One segment
Two segments

QRD

QR#

97

#
{
# X
 # {

Null hypothesis
Alternative hypothesis

Table 4.3: Using hypothesis testing for model selection in local segmentation.

If it is assumed a priori that the segments are piece-wise constant with additive Gaussian
noise 698:)+-,& { / , a statistical -test can be used to determine the more likely hypothesis [Gro88]. In Equation 4.21, is a random variable following Students distribution with
KL uA { X
degrees of freedom.

If

% %

is less than a critical value,

inferred to be homogeneous. If

= 2

% %

#k {

\ u

#k

&

(4.21)
K

, the null hypothesis is accepted, and the local region is

a 2

, the pixels are assumed to come from two popula-

tions. Equation 4.22 summarizes the -test model selection criterion for local segmentation.

Q

The value of

a 2

% %
% %

= 2

a 2

(4.22)

depends both on the degrees of freedom, K , of , and the percentage confi-

dence one wishes to have in the inference. Because the same window is used for each pixel,

a 2
K is fixed at u { 
A
#
. To achieve
;B)

F/ confidence,
should
F
equal { . This may be calculated using numerical integration, or from pre-computed

tables [PH66], an extract of which is given in Table 4.4. For the case of Kj , Students
distribution becomes a normal distribution.
For a window of

pixels, and a predetermined value for  , the mean separation threshold

function, )+&Z,F ,F { / , may be written as follows:

% %
%
#k n
#k {

& \ u

a 2

(4.23)

a 2

(4.24)

98

Chapter 4. Denoising with Local Segmentation

Degrees of
freedom K
2
7
11
19
23

.10
1.886
1.415
1.363
1.328
1.319
1.282

.025
4.303
2.365
2.201
2.093
2.069
1.960
F

Table 4.4: Various values of

%

#k

n# k %
{

.005
9.925
3.499
3.106
2.861
2.807
2.576

used in -testing.
uA
{

{

a 2
&(

a 2

)+&Z,F ,F { /^

&

(4.25)
(4.26)

Figure 4.31 compares four different values of  for denoising montage using the -test
criterion. The value of  is used to determine dynamically the mean separation threshold in
Equation 4.26. A   filter is used, so

# , and the noise variance is assumed known. At

very low noise levels, the value of  does not appear to affect results, but as & is increased,
the lowest value of A#-NB does best. This is the highest of the confidence levels tried.
In this thesis a significance level of  4BB , or 99% confidence, will be used. For 
windows, K'

 , so

a 2

1
{ { 


q

-NBB , or approximately 3.5. In this case,

the separation function reduces to Equation 4.27.

)+&Z,F ,F { /0

 &

{

(4.27)

The -test threshold function has a similar form to the simple one in Section 4.7.1, except
that H has been replaced by a term inversely related to the geometric mean of the two cluster
populations. Figure 4.32 plots versus the effective number of & separations required for
QI
to be declared. At the extremes it has values between 3 and 4, but at the middle where

the cluster populations are most symmetric is goes as low as 2.35. The -test approach allows
cluster means to be close together as long as the evidence, in the form of more accurate mean
estimates, is strong enough. When the cluster counts are skewed, the criterion insists on a
wider separation before accepting them.

4.7 Discriminating between two models

30

99

Value of
0.050
0.025
0.010
0.005

25

RMSE

20

15

10

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.31: Comparison of values for -test filtering of montage.

3.8
3.6

Effective value of C

3.4
3.2
3
2.8
2.6
2.4
2.2

3
4
5
6
Number of pixels in the first segment (m1 of 9)

Figure 4.32: The effective

values used by the -test criterion: CA and jPB .

100

Chapter 4. Denoising with Local Segmentation

In an implementation, Equation 4.26 would not need to be recomputed for each pixel. For a
given image, the values of

a 2

and & are constants. Because

degree of freedom within a window is

there are only )f

{ Y

, the only

. Figure 4.32 is also symmetric around xB

, so

I
/~xB
unique thresholds. These may be pre-computed and stored in a

look-up table. This feature would be very useful in a hardware implementation.

Visual inspection
Figure 4.33 provides a visual comparison of the -test and HO# model selection functions.
The four rows correspond to added noise standard deviations of 10, 20, 30 and 40. The three
columns contain the noisy image, the H3# denoised image, and the -test denoised image.
When &W

 , there is little to distinguish the two techniques. For &o

F , the light

squares have been filtered equally well, but the -test has performed slightly worse on the
background. For all cases other than 
or  , the effective value of H determined by the
-test is less than 3. Therefore, on average, it will choose Q 
more often than the simple
criterion. If there is a very noisy pixel present in a homogeneous area, the -test criterion is
more likely to place that pixel in a segment of its own.
Interestingly, increasing the noise to &2$B removes any observable differences between the
two methods. Normally, behaviour similar to that observed for &9

F is expected. But it

seems in this case that, by random chance, no extreme pixel values were present in the noisy
image. When & , the denoised outputs of both methods contain obvious artifacts. The
filtering errors in the H


 output image are milder, but a little more frequent, than those

observed in the -test output.

Both techniques suffer at similar locations in the image. Consider a local window straddling
the edge between the square and the background. If some of the noisy squares pixels are
darker than the original background, or some of the noisy background are lighter than the
original square, both algorithms will fail. The well-known fundamental drawback of thresholding algorithms is that they are unable to take spatial coherence into account. Unlike the
human visual system, the local segmentation method described thus far is not sophisticated
enough to recognize and extrapolate edge information.

4.7 Discriminating between two models

101

Figure 4.33: Visual comparison of two model selection functions. Columns are: noisy, XP
filtered, and -test filtered images. Rows are: L 10, 20, 30 and 40.

Quantitative results
Figure 4.34 shows that, for montage, the HED and -test criteria perform equivalently at
low noise levels. For values of &

, the simpler H

 criterion appears to do better.

So far, only a  window has been considered. It must be noted that the -test criterion
varies more as the window size increases. Although not shown, I have performed further

102

Chapter 4. Denoising with Local Segmentation

experiments which show that the -test criterion still performs worse when the window size is
increased to 21 and 25 pixels. This fact, in conjunction with the lack of compelling evidence
in the qualitative results, recommends the use of the simple HO# model selection criterion
over the more complicated -test criterion.

30

Force k=1
Force k=2
t-test =0.005
C=3

25

RMSE

20

15

10

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.34: RMSE comparison of model selection criteria for montage.

4.8 Iterative reapplication of the filter

It is possible to take the output of a denoising algorithm and feed it back as input to the same
algorithm, ad infinitum. Iteration has often been used as a way to improve the performance
of denoising algorithms [Mas85], particularly anisotropic diffusion [PM90]. Equation 4.28
describes this process formally, where = 2 is the denoised image after the

27

iteration. Ideally,

all pixel values are kept to floating point accuracy throughout. Rounding to integers after
each iteration introduces feedback of quantization noise, which could cause unusual artifacts.

4.8 Iterative reapplication of the filter

= 2
 FSd;){= 2 /

103

" S@d
S

= = ?

(4.28)

The effect on the final output as increases depends wholly on the properties of the denoising
algorithm. A stable algorithm is defined as one which, after a finite number of iterations ? ,
produces no further changes in the output. Thus = 2

={ , and is referred to as the root

image [FCG85]. An example of an unstable algorithm is the median filter. Imagine using a
local window of 5 pixels the centre pixels and its four nearest neighbours and handling
the image margins by taking the median of the available pixels. Applying this median filter to
the checkerboard image in Figure 4.35a results in the inverted checkerboard of Figure 4.35b.
Reapplying for a second iteration produces Figure 4.35c, which is identical to the original
image. In this case, the instability manifests itself as an oscillation between two images, the
difference of which has the maximum possible RMSE.

Figure 4.35: Oscillatory root images: (a) original image; (b) after one iteration; (c) after two
iterations.

Ideally, the root image should resemble the original but without the noise. Many existing
techniques produce a root image in which every pixel has the same value. For example, any
fixed linear filter with more than one non-zero weight will, if iterated, eventually produce
a homogeneous image. Even anisotropic diffusion [PM90], which is meant to be iterated,
behaves in this manner if the default stopping function is used. The final common intensity
will be roughly equal to the image mean. The exact value will vary with noise in the image,
and whether pixel values are rounded back to integers after each iteration.
One would expect the local segmentation denoising algorithm to be stable and, in general,
not to converge to homogeneity. This is because the explicit segmentation uses a hard cutoff to keep different groups of pixels separate. Pixels from different groups are unable to

104

Chapter 4. Denoising with Local Segmentation

influence each other. Of course, this only works if the two groups of pixels differ enough to
be detected by the local segmentation model selection criterion. This could be problematic if
the noise level was poorly estimated, or if the intensity difference between two neighbouring
segments was below the noise level.

Visual inspection
Figure 4.36 shows the result of iterating the local segmentation denoising algorithm of Section 4.7.1 100 times on the noiseless square. The simple model selection criterion with
HU^ is used, and &

was set to zero. The same value of & was used for each iteration.

The result of iterating a  median filter are also included for comparison purposes. It is
highly desirable for a denoising algorithm behave as an identity filter when the input image is
noiseless, and the local segmentation filter does that. The median filter truncates the corners,
but otherwise keeps the object boundaries sharp.

Figure 4.36: Iterating 100 times: (a) noiseless square; (b) local segmentation assuming R ; (c) } median filtered.

Figure 4.37 performs the same experiment, except that this time, noise 68

)f-,.F { / has

been added to square. Each iteration, the local segmentation filter assumed that &b:B .
The local segmentation filter proves to be both useful and stable. Its output only has two
pixel values: 47 for the background and 192 for the foreground. The reason these are not
equal to 50 and 200 as in the noiseless version could be due to three reasons: the noise
not being distributed symmetrically within each segment, clamping of pixels to -,.
BB , and
margin pixels being slightly over-represented due to their use in computing the extrapolated
values when the filter window goes outside the legal image coordinates, as discussed in
Section 4.2.8. Compared to Figure 4.28d, the result of only one iteration, the output is more

4.8 Iterative reapplication of the filter

105

visually pleasing. The median filter has also has a root image, albeit a disfigured one. The
noise has caused the median to behave unusually.

Figure 4.37: Iterating 100 times: (a) noisy # ; square; (b) local segmentation
assuming Rb; ; (c) median filtered.

As shown in Figure 4.38, the local segmentation filter fails when the noise is increased to
&MD . Its denoised output only contains pixels of intensity 101, although the mean of the

noiseless image is 88. This homogeneity is due to each of the original segments containing
pixel values more suited to the other segment. Unfortunately, the clustering algorithm can not
handle this situation. The result is intensity leakage between the segments. Given enough
iterations, this leakage will spread throughout the image. The median filter has reached a
better root image in this situation.

Figure 4.38: Iterating 100 times: (a) noisy #  square; (b) local segmentation
assuming R ; (c) median filtered.

Quantitative results
For montage, Figure 4.39 plots the RMSE between the original and denoised outputs as the
number of iterations is increased. The local segmentation filter used H3 and was supplied
with the true added noise level. In general, for a given level of added noise, the RMSE

106

Chapter 4. Denoising with Local Segmentation

worsens as more iterations are performed. This is a little surprising given the qualitative
results for square in Section 4.8. However, the assumption that the local region consists of
two piece-wise constant segments is clearly false for many positions within montage. Any
filtering errors after the first iteration would only be compounded by further iterations. Most
natural images would have properties in common with montage.

26

Number of iterations
20
10
5
2
1

24
22
20

RMSE

18
16
14
12
10
8
6
4

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.39: Effect of iteration on RMSE for denoising montage. The same value of is
used for each iteration.

The exception to the overall trend is that 2 iterations becomes slightly more beneficial than 1
when &3
; . This may just be the point at which the gain from further averaging outweighs
the loss from degrading image structure further. Figure 4.40 shows the effect of iterations
on a small portion of the noisy &O
F montage image, with the corresponding difference
images underneath. Both visual inspection and RMSE agree that 2 iterations has produced
a better output. The intensities of the text and background are more uniform and pleasing to
the eye. For montage, QjE
was chosen 80% of the time during the first iteration. This is
approaching a pure averaging filter, which, in terms of RMSE, becomes a good choice when
the noise level is high enough.

4.9 Rejecting poor local models

107

Figure 4.40: Bottom right hand corner of montage: (a) noisy 3; ; (b) after 1 iteration;
(c) after 2 iterations; (d)(f) corresponding difference images relative to the
noiseless original.

Conclusions
It is difficult to prove if a given denoising algorithm will be usefully stable in the general
case. The previous experiments showed that for some simple cases, the local segmentation
approach to denoising may be iterated successfully. When the noise is very high compared
to the contrast between objects, pixel leakage unfortunately occurs, and multiple iterations
could produce a homogeneous image. In Chapter 5, this problem will be tackled by using a
segmentation algorithm which takes spatial information into consideration.

4.9 Rejecting poor local models

The local image model used thus far has assumed that pixels from the window are well modeled using either one or two segments of constant intensity. For non-synthetic images this
assumption could be false for some windows. For example, the local region may consist of
more than two segments, or it might not be piece-wise constant in nature. Figure 4.41a gives
an example of a  block of pixels consisting of 3 segments with different constant intensities. If locally segmented under the assumption that QR
, the resulting local approximation
would have the pixel values given in Figure 4.41b.

108

Chapter 4. Denoising with Local Segmentation

(a)

20
20
40

40 60
40 60
40 40

assuming QL#

(b)

34
34
34

34 60
34 60
34 34

Figure 4.41: The effect of a two class model on a three class region: (a) original pixels;
(b) segmented assuming "CP .

The use of an incorrect model has forcibly merged two of the original segments. Its resulting
mean falls between the those of original segments. Thus 7 of the 9 pixels, including the centre
pixel, have a filtered value which is unreasonable given that the noise level is effectively zero.
If one believes that the noise is additive in nature and distributed 698:)+-,& { / , then confidence
interval analysis states that 99.7% of pixels are expected to fall within F& of their original
value. If a single filtered pixel value strays further than this, there is evidence to suggest that
the local segmentation model is inappropriate.
The ability to diagnose whether the best fitting model actually fits the data well is very useful.
All medical practitioners must take the Hippocratic Oath [HipBC], which has as one of its
cardinal principles the following: First of all, do no harm. This principle may be applied
to local segmentation when denoising images. If it appears that the best fitting model has
produced an unreasonable local approximation to the original image, dont use it. Instead,
use the original pixels as they were. It is probably better to leave them unmodified than to
give them new, possibly noisier, values. Equation 4.29 describes the do no harm (DNH)
principle. If the local approximation suggests changing any pixels value more than HC& , then
ignore it, and pass the pixels through unmodified.




R

k 2 ?

HC&

FRS@dy"S

For montage, Figure 4.42 shows the RMSE denoising performance of the  vH

(4.29)

`

local segmentation filter with and without DNH enabled. The algorithm was provided with
the value of & , as per all experiments so far. The DNH option provides an impressive improvement in RMSE for values of & up to 20, after which it does only slightly worse. When
&

is large, DNH is more likely to be invoked due to the dominating noise, rather than an

inability to model the underlying image structure. It must be remembered that for the results

4.9 Rejecting poor local models

109

presented here, the true noise variance was supplied to the local segmentation algorithm.
In a real situation & would have to be estimated from the noisy image. This issue will be
discussed in Section 4.13.

18

Default
With DNH

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.42: Effect of do no harm (DNH) on RMSE when denoising montage.

Figure 4.43 shows the effect that DNH has on the worst case absolute error (WCAE). As
expected the WCAE performance is much better. Enabling DNH has the effect of limiting the worst case behaviour of the filter for each pixel. This is most useful in low noise
environments, because it can help to avoid smoothing out fine image structure.
Figure 4.44a shows the top left quadrant of montage with added noise 6E8

)f-,. { / . Fig-

ures 4.44bc uses white to denote where QL
and QL#

are chosen when DNH is not used.
Figure 4.45 also uses white to show where Q2\
, Q2Y
and DNH are used when DNH is
enabled. DNH seems to be used mostly on the broad ramp edges and busy hat feather regions, where a two segment model is probably insufficient for modeling the image structure.
For this example, 73% of DNH invocations are due to rejecting a QLI
model. It may seem
unusual for a QL
model to be deemed better than a QLI
one, but then be rejected because

110

Chapter 4. Denoising with Local Segmentation

180

Default
With DNH

160
140

WCAE

120
100
80
60
40
20
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.43: Effect of do no harm (DNH) on WCAE when denoising montage.

of DNH. This occurs because binary thresholding sometimes produces poor segmentations,
and before DNH, QL
is a better overall choice.

Figure 4.44: DNH disabled: (a) original image; (b) "C , 64%; (c) "b , 36%.

The do no harm philosophy is not limited to local segmentation applications. It could be applied to any denoising algorithm to limit its worst case behaviour. All that is required is some
measure of the natural level of variation within segments around the pixel being processed.

4.10 Different local windows

111

Figure 4.45: DNH enabled: (a) "C , 60%; (b) "b , 25%; (c) DNH, 15%.

For example, the simple box filter of Section 3.4.3 could invoke the DNH rule when the local
average differs too much from any local pixel values. This filter would have two implicit local models: Q
and DNH. The result would be good smoothing in homogeneous regions,
and either blurring or no smoothing elsewhere.
Equation 4.29 is just one possible DNH rule. Instead of rejecting the local approximation if
just one pixel value would change too much, it could be relaxed to consider only the centre
pixel. Here DNH would be invoked less often because it is not testing the fitness of the whole
local segmentation model. For some windows the relaxed rule could improve denoising, but
for others, such as the one in Figure 4.41, it would do more damage than good.

4.10 Different local windows

The examples and experiments so far have used a  window containing 9 pixels. The
local segmentation approach is not restricted to this configuration. There are many other
commonly used windows, five of which are shown in Figure 4.46. The 5 pixel one is used
by anisotropic diffusion, while SUSAN utilizes the larger 37 pixel window.
As the number of pixels in the window increases, heterogeneous regions are less likely to be
well modeled by only two segments. Doing so could introduce large errors into the output,
regularly forcing DNH to reject filtering of the window. On the other hand, there would
be greater smoothing in homogeneous regions, because more pixels are being averaged.
Under the RMSE measure, a few large errors may have more negative influence than the

112

Chapter 4. Denoising with Local Segmentation

Figure 4.46: Common windows: (a) 5; (b) 9; (c) 21; (d) 25; (e) 37 pixels.

sum positive influence of many smaller errors due to better smoothing. The balance depends
on the features of the image being denoised.

Figure 4.47 gives results for denoising montage, first with DNH disabled. The general
trend suggests that smaller windows achieve better denoising with respect to RMSE. The 5
pixel window performs best until &U
 , at which point the next smallest (9 pixel) window takes over. This is an interesting result. For the   window, the 4 corner pixels are
roughly 41% further away from the centre pixel than the 4 side pixels. By argument of spatial coherence, the corner pixels are less likely to be related to the others. Segmenting fewer,
more related pixels means that QvY
should be chosen more often. A small window is less
likely to cross segment boundaries than a larger one. The advantage of choosing QjD
less
frequently is that homogeneous regions are more likely to be well modeled. For very noisy
images, Qv

is chosen more often anyway, so a smaller window will give less smoothing.

This is borne out by the 5 pixel windows RMSE increase as & gets higher.

When DNH is enabled, the results change quite dramatically. Figure 4.48 shows how DNH
limits the worst case behaviour of all the windows tested. The RMSE performances are
essentially equivalent until &E

 , whereafter the 9 pixel window takes over, just as in

Figure 4.47. It appears that the  window strikes a good balance between compact locality,
so that a two-segment model is sufficient, and having enough pixels for reasonable smoothing
in homogeneous regions.

4.11 Multiple overlapping local approximations

20

113

Window size (pixels)

5
9
21
25
37

18
16

RMSE

14
12
10
8
6
4

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.47: Effect of different windows for denoising montage, DNH disabled.

4.11 Multiple overlapping local approximations

The local segmentation principle states that the first step in processing a pixel should be to
segment the local neighbourhood of that pixel. This is a democratic process, treating every
pixel equally to infer the original values of every pixel in the neighbourhood, not just the one
being processed. This is unlike many other local denoising techniques, such as SUSAN and
GIWS, which use the centre pixel as a reference pixel for intensity comparisons.
Local segmentation provides a local approximation to the underlying image around each
pixel. Because each pixel is processed independently, the local approximations overlap to
some extent. If a window containing
the image participates in

pixels is used, it is easy to see that each pixel from

separate local segmentations, at each of the

possible positions

in the window. Figure 4.49 illustrates this graphically for of a  window, where
Thus, for each noisy pixel, there are

$ .

different estimates of its true value. The estimates

are not fully independent, as they share some pixel values in their derivation, so they can

114

Chapter 4. Denoising with Local Segmentation

25

Window size (pixels)

5
9
21
25
37

20

RMSE

15

10

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.48: Effect of different windows for denoising montage, DNH enabled.

be considered semi-independent. What would be desirable is to combine these multiple

overlapping estimates into a single compound estimate, rather than just using the centre
estimate as in Equation 4.17. This should have three main benefits: a larger effective window
size, further robustness, and improved denoising performance.
For square windows containing
)

pixels, the new effective window would contain exactly


/ { pixels, which is asymptotically a factor of four increase. For a  window

the factor is
BBxFC
1 . The calculation of the compound estimate is likely to require little
computational effort compared to performing local segmentation with a larger window. In
homogeneous areas, up to four times as many pixels will contribute to the calculation of the
mean, reducing its standard error by the same factor.
Let ;; be the

overlapping denoised estimates for the same pixel. How should these

be combined to produce a compound estimate k ? The simplest approach would be to take

a convex linear combination of the estimates, shown in Equation 4.30, where ( are the
weights. The denominator term ensures the weights sum to unity.

4.11 Multiple overlapping local approximations

115

Figure 4.49: Each pixel participates in 9 local windows.

v
k 

(H

(H

(4.30)

4.11.1 Choosing the weights

There are various factors which could be used to influence the weight given to a particular
estimate. Four possibilities are its position in the window from which it was computed, its
standard error, the number of segments diagnosed in its original window, and the measured
variance of its original segment.

Window position
Each denoised estimate is derived from a different local region. As shown in Figure 4.49,
the same shaped window is used for each region, but the position of the pixel within the

116

Chapter 4. Denoising with Local Segmentation

window differs per estimate. For a  window, the estimate appears in 9 different positions
at the four corners, the four sides, and once at the centre. At first consideration one may
guess that the centre estimate should be given more weight than those from the sides and
corners. However, the local segmentation process treats all pixels democratically, and the
thresholding technique takes no spatial information into consideration. There appears to be
little justification for treating them unequally.
This assertion will be tested by comparing a range of different weights. The weights should
be chosen in a manner such that the four corners are treated equally, and likewise for the four
edges. A simple model meeting these symmetry requirements is shown in Figure 4.50. Each
weight is calculated by raising a constant, S , to the power of its distance from the centre of
the window. In this case the Manhattan distance has been used. When SP] , the model
reverts to not producing a compound estimate at all. As S approaches 1, the estimates are
treated more and more equally. Do not confuse this S with that used in image processing to
measure the correlation between neighbouring pixels. Here the values being combined are
all estimates of the same pixel.

(a)

S {
S

S {
S

S {
S {

(b)

0
0
0

0 0
1 0
0 0

(c)
m

1 2
2 4
1 2

1
2
1

(d)

1 1
1 1
1 1

1
1
1

Figure 4.50: Weighting overlapping estimates: (a) any ; (b) ; (c) B ; (d) .

Figure 4.51 compares four values of S in terms of RMSE for denoising montage. To
observe the behaviour of overlapping averaging in the simplest possible environment, DNH
is disabled. The WCAE results under for the same conditions is plotted in Figure 4.52. It
is seen that any non-zero value of S improves both the RMSE and WCAE performance,
compared to not combining overlapping estimates at all. As argued earlier, the best results
are achieved when S*O
, which gives equal weight the overlapping estimates.
Figure 4.53 gives visual results for the square image after adding noise 68:)+-,
F { / to it.
The improved smoothing performance when S*O
is clearly visible. In both cases the edges
are reconstructed without any blurring.

4.11 Multiple overlapping local approximations

18

117

Value of
0
0.5
0.9
1

16
14

RMSE

12
10
8
6
4
2

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.51: Effect of on RMSE when denoising montage.

The standard error

A denoised estimate, ^ , takes its value as the mean of the cluster to which it belongs. It
was shown in Section 4.5 that the variance, or standard error, of an estimated cluster mean
is & { hx + , where

is the number of pixels in the cluster. This variance may be used to

measure the confidence in the estimate, and could contribute to its weight. A high standard
error corresponds to a low confidence. As

and & are assumed a priori known and constant,

the confidence component could be expressed by the relationship (G+ .

Figure 4.54 compares the RMSE of the confidence weighting, ( + , to equal weighting,
(H2

. The weights were normalized for each pixel using Equation 4.30. Once again,

DNH is disabled to limit the number of factors influencing the result. Somewhat surprisingly,
confidence weighting performs worse than equal weighting up to &2
F , at which point they
converge. In homogeneous areas of the image, both approaches should behave equivalently,
as + I

for every estimate. The variation must occur in heterogeneous areas of the image.

118

Chapter 4. Denoising with Local Segmentation

200

Value of
0
0.5
0.9
1

180
160

WCAE

140
120
100
80
60
40

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.52: Effect of on WCAE when denoising montage.

Figure 4.53: Denoising square: (a) noisy original, LP; ; (b) output when r ; (c) output
when .

In the vicinity of an edge, non-centre pixels from neighbouring homogeneous windows will
receive more weight than centre pixels from heterogeneous windows closer to the edge.
These estimates may be unreliable, so equal weighting (averaging) could be the best way to
minimize the average error, and hence RMSE. It was found that enabling DNH option did
not affect the pattern of performance observed in Figure 4.54.

4.11 Multiple overlapping local approximations

14

119

Confidence weighting
Equal weighting

12

RMSE

10

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.54: Effect of confidence weighting on RMSE when denoising montage.

Activity of the region

The activity of a window is considered low if it is homogeneous, and high if heterogeneous.

The number of segments, Q , determined to be present in the window may be used as a
measure of activity. A high activity region is diagnosed when the window truly does consist
of many segments, or when the assumed image model is unsuitable. One may wish to give
preference to estimates from less active regions, to minimize potential modeling errors.
If Q is the number of segments diagnosed for the window came from, the activity weight
could be expressed as (G
xBQ . The variables T and Q are strongly correlated, because
+

I


implies QLD
, and + r

implies QI
. To a large extent, the activity weighting

is already covered by the confidence weighting described previously, so it is unlikely to improve results. Numerical evidence to support this is given in Figure 4.55, where the activity
weighting provides no advantage over equal weighting.

120

Chapter 4. Denoising with Local Segmentation

14

Activity weighting
Equal weighting

12

RMSE

10

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.55: Effect of activity weighting on RMSE when denoising montage.

Measured class variance

Although the variance of each cluster is assumed to be & { , it is possible to also actually
calculate the sample variance, l { . If the image model is accurate,
l

and & are expected, on

average, to be very similar. A high value of l could be used to gauge reliability of a segment
mean. But due to the small number of pixels in most clusters, a reliable estimate of the
sample segment variance is difficult to obtain. In the case of "E
, the sample variance is
not even defined. For these reasons it will not be investigated further.

Summary
Of the four influences considered for weighting overlapping estimates, equal weighting, or
simple averaging, was found to perform best in terms of RMSE. If overlapping is enabled,
the compound estimate, k , will be computed using Equation 4.31, where 

are the

4.12 Larger numbers of segments

121

estimates being combined. Using a linear combination for calculating the compound estimate is only one possibility. Non-linear functions, like the median, could also be used, but
further investigation of these possibilities is beyond the scope of this chapter.

j
k 

(4.31)

When DNH is triggered, the filtered pixel value for the window are set equal to the original
pixel values it is assumed that the best estimate of a pixels true value is itself. In this
thesis, an estimate, , from a DNH model is included in the formation of the compound
estimate. An alternative would be to consider DNH estimates as unreliable, and assign them
zero weight. If the compound estimate could not be formed due to all estimates being unreliable, the original noisy pixel value could used instead.

4.12 Larger numbers of segments

It has been assumed so far that the local window consists of pixels from, at most, two segments. For most real images there will be occasions where this is not the case. This could be
due to the presence of three or more segments, or two segments not being aligned to the pixel
grid, as described in Section 4.2.3. In any case, results could improve if a larger number of
segments were considered. If segments are allowed to be as small as one pixel, it is possible
for

separate segments to occupy an

pixel window.

The obvious progression is to extend the thresholding method from Section 4.7 to work for
more than two clusters. Local segmentation can be considered to have two main components.
The first is the segmentation algorithm, which provides different candidate segmentations of
the local region. In our case, each segmentation differs only in the number of clusters, Q , but
it could be possible to incorporate alternative segmentations. The second component is the
model selection technique, which decides which of the candidate segmentations is the most
appropriate. In our case, this would choose the appropriate value of Q for the window.

122

Chapter 4. Denoising with Local Segmentation

4.12.1 Extending the clustering algorithm

In Section 4.6, Lloyds iterative threshold selection method was used for binary thresholding. What is required is an equivalently good technique which works for up to

clusters.

A generalized form of Lloyds method is the Q -means or k -means algorithm, referred to

here as Q -means. This algorithm was discussed in Section 3.2.1, and is repeated below in
Listing 4.2. In Step 2, closest is usually taken to be the Euclidean norm between the two
values. For greyscale pixels this simplifies to be the absolute intensity difference, but for
higher dimensional data such as colour pixels, a full calculation would be necessary.
1. Choose Q initial means (to be discussed in Section 4.12.3).
2. Assign each datum to the cluster with mean closest to its own value.
3. Recompute the cluster means using the new assignments.
4. If any pixels have switched clusters, go to Step 2.
Listing 4.2: The " -means algorithm for local segmentation.

The objective function of Q -means is to minimize the sum of squared deviations of pixels
from their cluster means. Each iteration of the Q -means algorithm is guaranteed to improve
the objective function, and the algorithm will always terminate. Unfortunately, the algorithm
will stop at the nearest local minimum, which is not necessarily the global optimum [GG91,
BB95]. The minimum reached depends wholly on the choice of initial cluster means [BF98].
Three techniques for choosing the initial means will be investigated in Section 4.12.3.

4.12.2 Extending the model order selection technique

So far, the decision between QRO
and QI
has been decided by examining the separation
of the two cluster means. The Q -means algorithm can be used to segment pixels into an
arbitrary number of clusters. The mean separation criterion can easily be extended to work
with more than two clusters each cluster mean is required to be far enough away from
every other cluster mean. The highest value of Q for which all the clusters are mutually well

4.12 Larger numbers of segments

123

separated is deemed the optimal clustering. Equation 4.32 describes this formally, where Q
is the optimal number of clusters and # k is the estimated mean for the 

argmax
ie ~\a,[


 a\

#
k n# k

)+&Z,F

27

cluster.

,F /
{

(4.32)

The Q -means algorithm can easily be configured to generate cluster means in ascending numerical order, so only adjacent cluster means need to be tested for separation. Equation 4.33
shows a simplified version of Equation 4.32, where # k - 4 is the 

e
argmax



# k - 4 # k -

U 4

<)p&,F

27

sorted cluster mean.

, /
{

(4.33)

4.12.3 Choosing the initial means

Inappropriate starting conditions can cause the Q -means algorithm to converge to a local
minimum. This problem can mostly be avoided in the one dimensional case (greyscale pixel
values) by judicious choice of the initial cluster means.

Measurement space
One approach is to spread the Q initial means evenly throughout the measurement space.
Equation 4.34 describes this, where ? 0 and ? are the lowest and highest pixel values
being clustered, and d# is the initial mean for cluster  .


d# X ? 0 u#)B 
/

? '? 0
QT

"RS@dS

}
J;;~Q

(4.34)

This method assumes that the pixel values are uniformly distributed in intensity. If the pixel
value distribution is highly skewed, some clusters may never be used. This is known as the
empty cluster problem [HM01].

124

Chapter 4. Denoising with Local Segmentation

Rank space
An alternative assignment of initial means is given in Equation 4.35, where ? - 4 is the


27

sorted pixel value. Instead of spreading the initial means uniformly by intensity, they are
chosen uniformly by rank, where a pixels rank is its sorted position. For example, in Equation 4.34 the pixels ?  and ? are the same as ? -mU4 and ? -


d# b ?

)n

\ \ Z

" S@d
S

q/

respectively.

(4.35)

}
J;;~Q

The rank technique improves the chance that the initial clusters will have at least one member
each, as the initial means are equal to existing pixel values. If the pixel data lacks variety,
adjacent clusters may still be assigned the same initial mean. This would result in one cluster
arbitrarily being chosen over another, leaving one empty after the first iteration.

Unique rank space

A simple modification can be made to the rank space method to avoid the problem of clusters
sharing an initial mean. The

sorted pixels are first scanned to remove duplicates, leaving

? pixels. The rank space approach can then be applied to the

? unique pixels. This is

described in Equation 4.36.


d# b ? 
\ Z

\ 

If $?

" S@d
S

}
J;;~Q

O Q

Qv$

? $

(4.36)

Q , it is impossible to cluster the data into Q distinct groups. Section 4.12.3 will

discuss potential bounds on Q in more detail.

Comparison of methods for choosing initial means

Figure 4.56 plots the RMSE denoising performance of the Q -means-based denoising algorithm for the three different Q -means initialization techniques. Enhanced features such as
DNH and overlapping averaging are disabled. At low noise levels the unique rank space
approach does best. From &' onwards there is little difference between the three methods.

4.12 Larger numbers of segments

18

125

Method for initial means

Measurement space
Rank space
Unique rank space

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.56: Comparison of three different methods for choosing the initial " means, in terms
of RMSE denoising performance on montage.

Forcing an upper bound on Q

Let 

be the highest value of Q to be tested. Higher values of Q require more computation

because there are more parameters to estimate and distributions to compare. Any heuristics
or computable bounds on 

for each pixel would reduce the amount of computation required

by local segmentation.

Maximum upper bound If each of the

pixels has a distinct value distant enough from

every other pixels value, then the algorithm will infer

upper bound on 

is therefore

unique clusters. An absolute

Block variance If the window variance, Fd@
?. , is less than the estimate of the noise vari
to produce clusters far enough apart. In this
ance, & { , then it is not possible for Q I

situation O
.

126

Chapter 4. Denoising with Local Segmentation

Maximum possible fit The range of the pixel values, ) ? -? 0 / , also limits the maximum value of Q achievable. Equation 4.37 calculates how many HC& -wide normal
distributions could fit within the pixel range. If using the simple mean separation
criterion, the same value of H should be used for each window. If using the -test separation criterion, the maximum possible effective value of H must be chosen to ensure
correctness in the extreme case.

 

? '? 0
u<


H&

(4.37)

Unique pixel values It is not possible to have more clusters than unique pixel values, because pixels with the same value are necessarily clustered together. If the pixels are
already sorted, as required by the initial means method of Equation 4.36, the number
of unique values may be obtained with a simple one-pass  )fI/ algorithm.
Limited by the user The user may wish to place an upper limit on  . They may have
prior expectations regarding the number of segments in each region, or simply wish
the implementation to run more quickly. This is particularly suited to large windows,
because one would not reasonably expect to find 49 clusters in a   window.

4.12.4 Quantitative results for multiple classes

The local segmentation denoising algorithm could easily utilise the multi-class Q -means segmentation algorithm in Listing 4.2, and the model selection criterion of Equation 4.33. The
montage image contains many local areas having more than two segments, for example, the
intersection of the four quadrants. One would expect the multi-class mode to better denoise
these regions. The local segmentation denoising algorithm now has three main attributes:

1. Support for binary or multi-class local segment models.

2. An ability to average overlapping estimates or just use the central estimate.
3. A DNH option to minimize worst case behaviour filtering behaviour.

4.12 Larger numbers of segments

127

Figure 4.57 compares binary and multi-class thresholding in terms of RMSE for denoising
montage. For the moment, both overlapping averaging and DNH are disabled. The RMSE
is much better for the multi-class model, especially at lower noise levels.

18

DNH is disabled and =0

Binary
Multi-class

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.57: Comparison of binary and multi-class local segmentation for denoising montage. DNH is disabled and r .

Figure 4.58 provides results for when overlapping averaging is enabled. Although the RMSE
improves for both techniques, binary thresholding benefits the most. The gap between the
two methods is only significant when &Y
 . This suggests that averaging estimates from
9 possibly incorrect binary models may produce a compound estimate which better reflects
the underlying image than a single estimate alone.
In Section 4.9, do no harm was introduced as a way to limit the worst case performance
of a denoising filter, and was found to dramatically improve RMSE performance at lower
noise levels. Figure 4.59 compares the same binary and multi-class algorithms when DNH
is enabled, but overlapping averaging is disabled. Surprisingly, binary thresholding does
marginally better when &M$ , while multi-class thresholding is better at high noise levels.

128

Chapter 4. Denoising with Local Segmentation

14

DNH is disabled and =1

Binary
Multi-class

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.58: Comparison of binary and multi-class local segmentation for denoising montage. DNH is disabled and r .

When the noise level is low, intricate image structure is more easily preserved. The multiclass algorithm will do its best to choose a high value of Q to fit this structure, but will
probably not capture all the detail. The binary model will also do its best, but with only two
clusters it is unlikely to do as well as the multi-class fit. Thus the binary method is forced
to invoke DNH, allowing fine detail to pass through unmodified. Retaining fine detail with a
small amount of noise is more beneficial than removing both noise and some image structure.
Figure 4.60 again compares the two methods when both DNH and overlapping averaging
are enabled. The RMSE metric is unable to distinguish the two methods to any significant
level. Although not shown, my experiments show this behaviour is duplicated when larger
windows, such as   and   , are used instead.
These results show, at least for montage, that more complex multi-class modeling does
not improve overall RMSE results when DNH and overlapping averaging are utilized. This
behaviour could be exploited if the algorithm was implemented in hardware, or as embedded

4.12 Larger numbers of segments

18

129

DNH is enabled and =0

Binary
Multi-class

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.59: Comparison of binary and multiclass local segmentation models on montage. DNH is enabled and r .

software running on a system with relatively low processing and memory capabilities. The
results also suggest that a large proportion of local windows are well modeled by two or
fewer segments, or that DNH models complex regions better than a piece-wise constant
multiclass model can. When a large number of segments are present, there are too few pixels
with which to estimate the class parameters.
Consider a noisy &

 version of montage. Figure 4.61 shows, using white, where the

binary thresholding method used DNH, Q

, and Q'

. Figure 4.62 shows the same for

the multi-class method for values of Q up to 8. The image for QL is not included, as it was
only used twice for the whole image. Table 4.5 lists the frequencies of model usage, DNH
included, for the binary and multi-class methods for the same image.
The distribution of Q for the multiclass method is monotonic decreasing. It did not use values
of Q

 very often in montage. Interestingly, QjI was chosen twice. This is equivalent

to using DNH, because each pixel is given its own cluster. DNH is not used very often in the

130

Chapter 4. Denoising with Local Segmentation

14

DNH is enabled and =1

Binary
Multi-class

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.60: Comparison of binary and multiclass local segmentation models on montage. DNH is enabled and .

Figure 4.61: Binary local segmentation of noisy G montage. White denotes: (a) DNH;
(b) "C ; (c) "b .

synthetic quadrants of montage, its invocation being more beneficial in the natural areas.
When restricted to binary thresholding, the number of homogeneous regions remains the
 need to be reclassified. It seems
same. The 16% of windows previously diagnosed as Q #

that seems that 11% were allocated to DNH, while 5% were adequately handled by QR
.

4.12 Larger numbers of segments

131

Figure 4.62: Multi-class local segmentation of noisy :` montage. White denotes:

(a) DNH; (b)(i) " to "7 .

Model Multiclass Binary

used 


DNH 4%
15%
Q

52%
52%
QI

28%
33%
Q#
10%

Q
4%

QI
1.1%

Q#
0.2%

QI
0.03%

Q#
11 pixels

2 pixels

Q#
Table 4.5: Binary and multi-class model usage for noisy LP montage.

132

Chapter 4. Denoising with Local Segmentation

4.13 Estimation of the noise variance

In the experiments performed in this chapter, known amounts of synthetic noise are added to
ground truth images. The various model selection criteria explored require that an estimate of
the image noise variance, & { , be known a priori. Local segmentation uses the supplied noise
variance to measure the natural level of variation of pixel values within segments. This helps
discriminate between those features attributed to structure, and those attributed to noise.
Usually, the exact noise variance is unknown. It must be guessed by the user, or estimated
from the noisy data itself. There are various algorithms in the literature for estimating the
variance of additive noise in an image [Ols93]. The algorithms fall into two main categories:
1. Those which filter the noisy image first to remove structure and then estimate the
variance from the residuals [Imm96, RLU99]
2. Those which ignore heterogeneous regions and use the remaining pixels to estimate
the variance [Lee81b, Mas85, MJR90].
Those in the first category tend to overestimate the variance, as they are unable to completely
remove structure. The second category tends to underestimate the variance, as avoiding
heterogeneous regions would naturally bias them to less noisy regions.
Figure 4.63a shows montage with &

 noise added to it. Figure 4.63b is the same

image except that each pixel has been replaced by the unbiased standard deviation of its
 local neighbourhood, itself included. As expected, the standard deviation is lowest in

homogeneous regions, and highest in textured and edge regions. It is possible to use the
distribution of local standard deviations to estimate the global noise variance, & { .
Figure 4.64 shows a histogram of all local standard deviations in montage having values
from 0 to 100. If montage was completely homogeneous, the histogram would be normal
with mean & and variance & { 1x  . However, the existence of  heterogeneous regions in
montage places many high variance entries in the right hand tail of the histogram [RLU99].
Figure 4.65 plots a smoothed version of the histogram for values up to 20. The smoothing
was performed using a sliding window average of width 3. The smoothed histogram has
a clear peak at around 5, which is equal to & , the standard deviation of the added noise.

4.13 Estimation of the noise variance

133

Figure 4.63: (a) noisy RP montage; (b) local standard deviations.

0.008

0.007

Relative frequency

0.006

0.005

0.004

0.003

0.002

0.001

20

40

60

80

100

Local standard deviations

Figure 4.64: Histogram of local standard deviations from Figure 4.63b.

Thus the mode of the local standard deviation distribution could be used as an estimate for
& [BS85]. This method belongs to the second category of noise estimation algorithms.

134

Chapter 4. Denoising with Local Segmentation

0.008

0.007

Relative frequency

0.006

0.005

0.004

0.003

0.002

0.001

10

15

20

Local standard deviations

Figure 4.65: Smoothed version of main peak from Figure 4.64.

A noise estimation technique from the first category is Immerkrs method [Imm96], designed specifically for the case of zero mean additive Gaussian noise. To remove image
structure it applies the   linear filter, i , in Equation 4.66, to the noisy image. This filter
is the weighted difference of two Laplacian [GW92] filters, and { , which estimate the
second derivative of the image signal. The effect of i is to reduce constant, planar and
quadratic  facets to zero plus a linear combination of the noise. The effect of this filter
on lenna is given in Figure 4.67.

0 1
1 -4
0 1

0
1
0

{ 

1 0
0 -4
1 0

1
0
1

iW# { X
F 

1
-2
1

-2 1
4 -2
-2 1

Figure 4.66: Filter masks used by Immerkrs noise variance estimation technique.

Once the image has been filtered to remove structure, the filtered pixel values can be used to
compute the estimated noise variance, denoted & k { . Equation 4.38 describes the calculation,

4.13 Estimation of the noise variance

135

Figure 4.67: (a) original lenna image; (b) after Immerkr structure suppression, with mid
grey representing zero.

where w)|,-/ i

denotes the result of applying filter i

at pixel )p|,-/ . The formula does

not use margin pixels because they do not have a full   neighbourhood.

&k { 


{ {

B)Wb
B/@)+IX
B/ ~

)+w)|,/ i'/ {

(4.38)
%

By using the fact that, for a zero mean Gaussian random variable , { [ 
{

Immerkr describes an alternative method for computing the estimated noise standard deviation. Rather than the sum of squared residuals, Equation 4.39 uses the scaled sum of absolute
deviations. This formulation has two advantages: the summation requires no multiplications,
and the absolute deviation is more robust to the presence of outliers [Hub81]. It would also
be possible to use alternative robust methods, such as the median absolute deviation from the
median [RL87], but these will not be investigated here.

&'
k 


{ { %

)pWb
B/)fIX
B/ ~

w)|,-/ i

(4.39)

Figure 4.68 compares the three noise estimation techniques for estimating the standard deviation of synthetic noise added to montage. As hypothesized earlier, the two Immerkr

136

Chapter 4. Denoising with Local Segmentation

methods consistently overestimate & , because they can not remove all the image structure.
The mode method does very well, but slightly underestimates high values of & .
Although montage is assumed noiseless, it was mentioned in Section 4.4 that the lower
right quadrant does contain a small amount of noise, hence the robust Immerkr estimate
may have some truth to it. If the noise level varies across an image, any global estimate will
fall somewhere between the lowest and highest levels.

35

Estimation Method
Actual
Mode of local variances
Immerkaer (standard)
Immerkaer (robust)

Estimated standard deviation (

)

30
25
20
15
10
5
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.68: Different noise estimation algorithms for montage.

In real situations, the noise variance must be estimated from the noisy image without reference to any ground truth. Consider the familiar lenna image from Figure 4.18, which
already contains noise. Let us assume that it is additive Gaussian noise with variance l { .
If synthetic noise of variance & { were added to lenna, the actual noise variance would be
a combination of the original and the synthetic, namely & { u
l

{ . We can compare noise

estimation techniques in the same manner as before on images already containing noise by
correcting the actual standard deviation to include the pre-existing noise. Instead of comparing estimated noise levels to & , the curve

& { ul { can be used instead.

4.13 Estimation of the noise variance

137

0.014

Relative Frequency

0.012
0.01
0.008
0.006
0.004
0.002
0

2.5

7.5

10

Local Standard Deviation

Figure 4.69: Smoothed histogram of local standard deviations for lenna.

A method for estimating the base noise level, l , is needed. Figure 4.69 plots a smoothed
histogram of lennas local   standard deviations. The mode of this histogram estil Y
NB , while the robust Immerkr method estimates R
l Y
NB . Figure 4.70 commates 

pares noise estimation algorithms for lenna, using

l

)
NBu
4F/~xB
*
4B
to correct

the curves. All the methods now mostly follow the ground truth. The standard Immerkr
method slightly overestimates at the lower end, but this is somewhat due to its estimate not
contributing to the base noise level estimate, l . The mode method slightly underestimates
at the higher end, just as it did for montage. At higher noise levels the assumption that all
the heterogeneous variances exist in the far right of the histogram breaks down, causing the
mode to be less accurate. It may be possible to estimate the mode by first fitting a spline to
the data, and then analytically determining the peak. Overall, the deviations are minor, and
any of the estimates would probably produce similar denoising results.
The specific noise estimation procedure used is only important with respect to the quality of
the variance estimates it produces. The local segmentation process does not require modification if the noise estimation technique is varied. The noise estimator is a separate module,
which may be replaced with a better technique if one becomes available.

138

Chapter 4. Denoising with Local Segmentation

35

Estimation Method
Actual
Mode of local variances
Immerkaer (standard)
Immerkaer (robust)

Estimated Standard Deviation

30
25
20
15
10
5
0

10

15

20

25

30

Added Gaussian Noise Standard Deviation

Figure 4.70: Different noise estimation algorithms for lenna, corrected for existing noise.

4.14 Comparison to other denoising algorithms

This chapter has developed a family of denoising algorithms based on the local segmentation principle. Various modifications to the basic technique were explored, many of which
improved results both visually and in terms of RMSE. The best trade-off between efficiency
and performance over a range of noise levels is determined to have the following attributes:

Use of a local  window (Section 4.10)

Model local regions as consisting of either 1 or 2 segments (Section 4.12)

Use of Lloyds algorithm for threshold selection local segmentation (Section 4.6)

Simple model selection criterion using HO (Section 4.7)

Use of the do no harm (DNH) option to limit worst case behaviour (Section 4.9)

Equal averaging of overlapping estimates, after DNH has been applied (Section 4.11)

One iteration only (Section 4.8)

Estimate noise variance with Immerkrs robust method (Section 4.13)

4.14 Comparison to other denoising algorithms

139

This variant shall hereafter be referred to as FUELS filtering using explicit local segmentation. FUELS will be compared to the four denoising algorithms described below, each of
which was described in detail in Section 3.4.2. These algorithms were chosen because, like
FUELS, they operate locally, are relatively efficient, and use similar image models.
WMED (Section 3.4.4)
WMED is the   centre weighted median with the centre pixel included three times.
Although not as efficient as the mean in homogeneous regions, it can preserve fine
lines and some edges. WMED may be considered a minimum level of performance
that any structure preserving denoising algorithm should reach. It is probably better
known for its resistance to impulse noise.
GIWS (Section 3.4.5)
Gradient inverse weighting smoothing, centre pixel not included, will be used. GIWS
can be considered an accelerated single iteration of anisotropic diffusion.
SUSAN37 (Section 3.4.5)
The standard 37 pixel SUSAN filter is one of the best local denoising algorithms in the
literature, and has a very efficient implementation. It requires a brightness threshold
, and Smith claims that

:
F works well over all image types. I found that setting

#& k gave better results, where & k is the same estimated noise level that FUELS uses

when it is not provided with one.

SUSAN9 (Section 3.4.5)
The SUSAN implementation also has a 9 pixel mode, which uses a  window.
This was also included as it is closer in size to the window that FUELS uses. When
something affects both SUSAN9 and SUSAN37, the generic term SUSAN is used.

4.14.1 The montage image

Supplied with noise level
Figure 4.71 plots the RMSE performance of FUELS, WMED, GIWS, SUSAN37 and SUSAN9 for montage. Just for this time, FUELS and SUSAN were provided with the true

140

Chapter 4. Denoising with Local Segmentation

value of & . FUELS clearly outperforms the others at all noise levels. SUSAN37 also did well
until &E

, after which point SUSAN9 took over. SUSAN9 performed badly for very low
noise levels. It is unable to assimilate enough pixels, and switches to being a  median
filter. GIWS consistently tracks the SUSAN37 method, but 2 RMSE units higher.

20

FUELS
GIWS
18 SUSAN37
SUSAN9
WMED
16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.71: RMSE comparison for montage, true supplied.

Estimated noise level

Figure 4.72 repeats the experiment, except that this time FUELS and SUSAN utilise the
estimated noise variance. Immerkrs noise estimation method tends to overestimate & for
montage. This is due to the large number of strong step edges, which are somewhat atypical
for photographic type data. The results once again show FUELS to perform best, although
SUSAN is within one RMSE unit at most points. It is not clear why SUSAN37 does better at
low noise levels, and SUSAN9 at high levels. I would expect a larger mask to provide more
smoothing at higher noise levels than a smaller one.

4.14 Comparison to other denoising algorithms

141

20

FUELS
GIWS
18 SUSAN37
SUSAN9
WMED
16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.72: RMSE comparison for montage, with estimated.

The use of the estimated noise variance has narrowed the gap between those algorithms
which exploit its knowledge, namely FUELS and SUSAN, and those which do not, namely
GIWS and WMED. The poor performance of SUSAN9 at low noise levels has also gone.
This must be due to the estimated noise variance being higher than the true variance.
The resulting brightness threshold supplied to SUSAN is therefore higher, allowing it to
assimilate more pixels.

The denoised output

Figure 4.73 compares SUSAN and FUELS in terms of structure preservation. The F F
image examined is from the centre of montage when &23 . Ideally the difference images
should contain no structure. There are no obvious differences in the top left quadrant. In the
top right quadrant both had difficulty with the oval shape, because it has a similar intensity
to its background. SUSAN37 had trouble with the cross above it, whereas FUELS left little
trace of it in the difference image. In the lower left quadrant FUELS seems to have smaller

142

Chapter 4. Denoising with Local Segmentation

errors at segment boundaries, but SUSAN37 seems to have removed more noise within segments. This could be due it using a 37 pixel window, compared to FUELS 9 pixels. The
bottom right quadrant contains a lot of fine detail, and overall FUELS appears to have done
better than SUSAN37 there. The roof and shutters are much more obvious in the SUSAN37
filtered difference image.

Figure 4.73: (a) noisy I part of montage; (b) FUELS enhanced difference image,
RMSE=3.48; (c) SUSAN37 enhanced difference image, RMSE=4.11.

Figure 4.74 plots the WCAE for montage. Both WMED and GIWS have a near constant
WCAE, suggesting there is a pixel pattern in montage which they both consistently get
wrong. FUELS WCAE seems directly proportional to & . This behaviour is desirable because it indicates that its mistakes are due to the random nature of the noise, rather than
difficulty preserving any particular image structure. Overlapping averaging and DNH are
mostly responsible for this positive feature of FUELS.

Both SUSANs have poor worst case performance at lower noise levels. Figure 4.75a shows
a vF part of montage without added noise, for which SUSAN is given a brightness
threshold


B
B4F . Figures 4.75bc show SUSAN37s output and the corresponding dif-

ference image. The two white spots in the difference image are large filtering errors, the
worst of which is out by 150 intensity units. This behaviour can be explained by the fact
that SUSAN switches into a special mode when all the neighbouring weights are too small.
In this mode, SUSAN assumes the centre pixel is impulse noise, and instead uses the median of its eight immediate neighbours as its denoised estimate. The region of montage in
Figure 4.75a is very busy, and the median obviously chose inappropriately.

4.14 Comparison to other denoising algorithms

143

250

200

WCAE

150

100

50

FUELS
GIWS
SUSAN37
SUSAN9
WMED
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.74: WCAE comparison for montage image, with estimated.

Figure 4.75: (a) part of montage without added noise; (b) SUSAN37 output using ;i; ;
(c) difference image.

4.14.2 The lenna image

The lenna image [len72] in Figure 4.18 is the single most popular test image used by the
image processing community. Because lenna is not synthetic, it already contains some
noise of unknown distribution. In Section 4.13, under the assumption of additive Gaussian
noise, the natural noise standard deviation, l , was estimated as being somewhere between 2
and 3. Using the original lenna as ground truth is difficult, because any good denoising

144

Chapter 4. Denoising with Local Segmentation

algorithm will remove both the original and synthetic noise. As the added noise level, & ,
increases, the original noise is swamped and hence can mostly be ignored. However, when
&$\
]l or so, the original noise level will impact the RMSE results. In these cases visual

inspection of the denoised and difference images would be necessary to fairly compare the
quality of smoothing and structure preservation.
Figure 4.76 plots RMSE results for lenna. Although FUELS performs best at all noise
levels, the two SUSAN variants are very close. In fact, when &E , FUELS, SUSAN and
GIWS perform within one RMSE unit of one another. Perhaps the common use of lenna
as a test image has mildly biased algorithmic development toward those techniques which
do well on lenna and other images with similar properties.

20

FUELS
GIWS
18 SUSAN37
SUSAN9
WMED
16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.76: RMSE comparison for lenna, with estimated.

Figure 4.77 shows the WCAE when denoising lenna. All except FUELS have a nearconstant WCAE. Interestingly, SUSAN37 improves its WCAE as the added noise increases
from none to low amounts. This unusual behaviour is once again likely to be due to its
fall-back median filter. Although the monotonic increase of FUELS WCAE is desirable, it

4.14 Comparison to other denoising algorithms

145

reaches levels higher than the other algorithms when &

 . FUELS has some large errors

in the feathers of lennas hat, which is notoriously difficult to process. Modeling them as
piece-wise constant segments is likely to be problematic, and invoking DNH in high noise
environments will create errors proportional to the noise level.

250

200

WCAE

150

100

50

FUELS
GIWS
SUSAN37
SUSAN9
WMED
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.77: WCAE comparison for lenna image, with estimated.

4.14.3 The barb2 image

Figure 4.78a shows barb24 from the JPEG test set [PM93]. It is larger than lenna, having
resolution B
FRB1 . A feature of images from the JPEG test set is that they only contain
intensities in the range
;-,
F . The histogram of barb2 in Figure 4.78b clearly shows
the saturation effect caused by pixels being clamped to the reduced range. The image is
interesting in that it contains large areas of straight edge patterns in the chair, clothes and
books, but also contains clearly homogeneous background regions, including a dark, noisy
4

Available from http://www.csse.monash.edu.au/ torsten/phd/

146

Chapter 4. Denoising with Local Segmentation

margin on the right hand side. This variation in features make barb2 useful for examining
the structure preservation capabilities of denoising algorithms.

Figure 4.78: (a) the ;;CL; barb2 image; (b) its histogram.

Figure 4.79 compares the denoising algorithms in terms of RMSE on barb2. Compared to
lenna, the results are more diverse, but FUELS still does best for all values of & . Another
interesting fact is that SUSAN37 achieves lower RMSE results than SUSAN9 up until &2 ,
after which they swap. Intuition suggests to me that this should be the other way around, as
a larger mask should allow better smoothing in the presence of more noise. The threshold
chosen for SUSAN may not be appropriate for all image and noise level combinations.
When denoising a highly patterned image like barb2, one would expect most algorithms
to produce some large errors. The WCAE graph in Figure 4.80 supports this hypothesis.
SUSAN9 had particular trouble with the large number of edges in the image, again due to its
fall-back median filter. FUELS has done very well for &3
 , assisted by its DNH feature.
Figure 4.81 compares the denoised outputs of FUELS and SUSAN for a sub-image of
barb2 without added noise. The estimated noise standard deviation was 4.2. The corresponding difference images are also included. These were enhanced by linearly stretching
differences in the range C

,;

 to  ,
BB . Large differences were clamped to 

.
FUELS has left very little structure in the difference image. Its DNH option has obviously
been used in the vest area, because most of the differences there are zero. SUSAN9s difference image exhibits little image structure too. The large differences occur in clumps, rather
than spread evenly across the image. They are particularly noticeable on the vest. This could

4.15 Conclusions

147

20

FUELS
GIWS
18 SUSAN37
SUSAN9
WMED
16

RMSE

14
12
10
8
6
4
2

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.79: RMSE comparison for barb2, with estimated.

be due to the fall-back median filter, which is a poor choice for that type of image structure.
The face and wicker chair from barb2 are clearly noticeable in the SUSAN37 difference
image. Although SUSAN37 produces smaller differences on average than SUSAN9, they
are correlated with edges in barb2. This is probably due to the 37 pixel mask being more
likely to encompass segment boundaries, and hence blur them slightly.
A histogram of the three difference images is given in Figure 4.82. FUELS errors all occur
in a tight distribution around zero. SUSAN37 has a slightly wider band, and one or two
outlier errors at 50. SUSAN9s predisposition to larger errors is clearly indicated by the fat
tails of its distribution.

4.15 Conclusions
It has been shown that the principles of local segmentation can be used to develop effective
denoising algorithms. After many analyses, the FUELS algorithm for denoising greyscale

148

Chapter 4. Denoising with Local Segmentation

140
120

WCAE

100
80
60
40
FUELS
GIWS
SUSAN37
SUSAN9
WMED

20
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 4.80: WCAE comparison for barb2, with estimated.

images contaminated by additive noise was presented. FUELS has an efficient implementation, and only requires one parameter, the level of noise in the image. This can be supplied
by the user, or FUELS can determine it automatically. FUELS was shown to outperform
existing methods, like SUSAN and GIWS, for a variety of images and noise levels.
Both quantitative and qualitative methods were used to compare FUELS to other methods.
The RMSE was used to measure objectively the closeness of denoised images to the originals. FUELS consistently produced lower RMSE results than SUSAN, the next best performer. The WCAE was used to gauge the worst case performance of each algorithm. FUELS had the desirable attribute of having a WCAE proportional to the noise level in the
image. The others, SUSAN included, tended to have constant or erratic WCAEs. To assess
the structure preserving ability of each algorithm, difference images were used to highlight
those areas of the image in which larger errors occurred. Although structure was apparent in
all the difference images, FUELS tended to contain the least.
The FUELS algorithm has various attributes which are responsible for its good performance.

4.15 Conclusions

149

Figure 4.81: For barb2 with no added noise: (a)(c) FUELS, SUSAN9 and SUSAN37
denoised output; (d)(f) corresponding enhanced difference images.

Like SUSAN and GIWS, FUELS attempts to average only those pixels which, in some sense,
belong together. FUELS achieves this by explicitly segmenting the whole local region, insisting that each pixel belong wholly to one segment. This contrasts with SUSAN and GIWS,
which advocate a soft cut-off. A hard-cut off ensures that pixel values can not diffuse across
segment boundaries to influence other, unrelated, pixels. It has the advantage of providing
a local approximation to the underlying image, which is arrived at democratically, because
each pixel contributes equally to the local segmentation. This is unlike SUSAN and GIWS,
which assimilate pixels based on their relationship to the centre pixel in the window only.
FUELS acknowledges the requirement for denoising algorithms to have magic parameters
to allow adaptation to different images. The most common type of parameter controls the
distinction between noise and structure. GIWS implicitly has one in the
{

term in the de-

nominator of its weighting function, while SUSAN needs a brightness threshold to control

150

Chapter 4. Denoising with Local Segmentation

30

SUSAN9
SUSAN37
FUELS

25

Frequency

20
15
10
5
0
-100

-50

0
Error

50

100

Figure 4.82: Partial histogram of FUELS and SUSAN denoising errors for barb2.

its pixel assimilation. It is not always clear how to choose these parameters appropriately.
The use of local segmentation naturally links the image model to the segmentation algorithm used. FUELS was developed under the assumption of piece-wise constant segments
corrupted by additive zero-mean Gaussian noise. This obviously suggests the noise standard
deviation as a measure of the natural level of pixel variation within segments. FUELS has
the advantage of automatically determining this parameter if it is unknown to the user.
Determining a local approximation to the underlying image at each pixel position has two
advantages. Firstly, because each pixel participates in multiple overlapping local approximations, there exist multiple estimates of that pixels true value. FUELS exploits this by
averaging the overlapping estimates. This has the effect of increasing the effective window
size by a factor of 2.8, without the overhead of segmenting more pixels. Secondly, the local approximation can be assessed, and rejected if necessary. Just because an algorithm has
determined its best estimate for a pixels value does not mean that estimate is of high quality. If any of the locally approximated pixel values differ too much from their original noisy

4.16 Related work

151

values, FUELS refuses to accept them. Instead, the local approximation is set equal to the
unfiltered pixel values. The acceptable intensity difference used is strongly related to the
estimated image noise level. This do no harm philosophy significantly improves results at
lower noise levels, and may be applied to any denoising algorithm, not just FUELS.
Examination of images in terms of local segmentation has lead to a better understanding of
image processing on a small scale, particularly for the commonly used  configuration.
At conversations and seminars, I have often heard the off-hand comment that only 10% to
20% of images are edges. Analysis of the distribution of Q values chosen for images in the
chapter suggest that only around 5060% of pixels are locally homogeneous, 2030% consist
of two segments, and 1020% tend to be difficult to model well. Perhaps the speakers were
confusing edges with those pixels which are difficult to predict or model. The success of
FUELS suggests that attempts to model these difficult blocks does not significantly improve
denoising performance.

4.16 Related work

An early form of the work in Section 4.7.1 appeared in Proceedings IAPR Conference on

Pattern Recognition 1998 [ST98]. That paper outlined a structure preserving denoising algorithm which decided between QG

and QE
, with Q

being chosen if the range of

the pixels in the  neighbourhood was greater than F& . The noise level & was estimated as
an average of the local variances of homogeneous regions, with homogeneity being declared
if the Sobel edge strength [Sob70] was less than 16. The QR
model was thresholded using
the dynamic mean [CC94], and not iteratively determined. The concept of overlapping
averaging was also introduced, albeit only with equal weights.

Chapter 5
Information Theoretic Local
Segmentation
5.1 Introduction
In Chapter 4, the principle of local segmentation was introduced and applied to the problem
of image denoising. Initially, two candidate segmentations were considered for modeling the
local neighbourhood of each pixel being processed. The first candidate assumed the local
region was homogeneous. The second candidate was generated by thresholding the pixels
into two classes. A simple model selection criterion was then used to decide which of the
two candidates best modeled the underlying image. The selection criterion was based on
the level of noise in the image. It tried to choose a two-segment model only if the segment
means were separated enough. Later, the concept was extended to allow for more than two
candidate segmentations by using the Q -means algorithm for multi-level thresholding. Up to

candidates were considered, where

was the number of pixels in the window.

Surprisingly, allowing the model selection process to consider models with more than two
segments did not improve results significantly. There could be two main reasons for this.
Firstly, it may be true, that, on a small scale, one is unlikely to encounter junctions between
more than two segments. Secondly, the use of thresholding to group small amounts of data
into a relatively large number of clusters may be inappropriate. Thresholding relies solely
on pixel intensity for guiding segment membership. As the image noise level increases,
153

154

Chapter 5. Information Theoretic Local Segmentation

spatial information may be required for successful segmentation. This could be used to
better distinguish whether pixel variation is due to noise, or to underlying image structure.
Local segmentation is fundamentally concerned with choosing the best underlying local image model, from a set of candidate models. This is an example of the quite general problem of
inductive inference, or model selection. Over the last few decades Bayesian techniques have
proved to be the most reliable for model selection [RR98, BS94]. In this chapter, Wallaces
Minimum Message Length (MML) criterion [WB68, WF87, WD00] for model selection is
used to objectively compare different local segmentations. MML is an information theoretic criterion related to traditional Bayesianism, but extended to function better with models
containing continuous parameters. MML evaluates models using their message length. A
message is an efficient, unambiguous joint encoding of a model and the data. The model
associated with the message of shortest overall length is deemed the best model for the data.
Image denoising will again be used as a test bed for exploring the potential of an information
theoretic approach to local segmentation. It will be shown that the MML model selection
criterion leads to better RMSE performance, especially at higher noise levels, by removing the minimum contrast difference that FUELS requires. Instead of being treated as a
post-processing step, the do no harm heuristic will be shown to fit naturally into MMLs
information theoretic framework. A much larger set of candidate segmentations are considered, allowing spatial information to be exploited. With FUELS it unclear how to compare
two different Q'

models, but the MML criterion makes this straightforward. The MML

denoising algorithm also learns all of its required parameters from the noisy image itself.

5.2 Statistical model selection

Imagine we have some set of measurements, h , from the real world, and a set of models,


:

,  ;, m , with which we attempt to explain these measurements. How do we assess

{


k

the quality of and choose the best model from ? Consider the familiar problem of polynomial regression. The measurement data consists of i
Equation 5.1 describes an

27

coordinate pairs, )p ,1/;):9,P9 / .

order polynomial regression model, with a positive integer.

5.2 Statistical model selection

)Z/

155

 u@ *u@ { { u$uA]1

(5.1)

A particular model, , is fully determined by its polynomial order , and its corresponding
polynomial coefficients, to . Because the coefficients are continuous variables, there is
an infinite number of possible models, irrespective of the value of . Because it is impossible
to evaluate an unbounded number of models, the model space must be reduced intelligently.
The residuals,  @;9 , are the differences between the regression model and the actual measurements. Typically, the residuals are assumed normally distributed, namely 698

)f ,& { / .

For a given model order , the least-sum-of-least-squares algorithm [Nie94] can be used to
estimate the optimal polynomial coefficients by minimizing


 { . This reduces the set of

models, , to the family of least-squares polynomials parameterized by . A model selection

criterion only needs to choose the most appropriate value of for the data.
Figure 5.1 gives an example of 11 data pairs. These pairs exhibit some positive correlation.
Figure 5.2 shows three polynomial fits to these data pairs: 2I constant, E
linear and
G#
quadratic. The constant model does not fit the data very well. The linear and quadratic

models are more difficult to judge. The linear model has the virtue of being simpler, but the
quadratic appears to fit the end points more closely. An objective criterion is required to
choose the best model automatically.

5.2.1 Maximum likelihood

The maximum likelihood (ML) approach is an objective model selection criterion. ML
chooses the model which maximizes the probability of the data given the model the socalled likelihood of the data. Equation 5.2 describes the ML criterion.


k

argmax
 t

c0d)+h

%

(5.2)

For the regression example, a model consists of the value of , and u#
polynomial coefficients. The data consists of i -coordinates in the form of residuals from the fitted model,
namely @D)p/ . As mentioned in Section 5.2, the residuals are assumed normally

156

Chapter 5. Information Theoretic Local Segmentation

12

10

8
6
4
2
0

10

10

Figure 5.1: Scatter plot of 11 data points.

12

D
f(x)=a0
f(x)=b0+b1x
2
f(x)=c0+c1x+c2x

10

8
6
4
2
0

6
x

Figure 5.2: Three different polynomial fits to the 11 data points.

5.2 Statistical model selection

157

distributed. The variance of this distribution is calculated from the residuals [Nie94], and
is assumed true and known, just like the coordinates. The likelihood then is equal to the
joint probability of the residuals, given the model. Traditionally the residuals are assumed
independent, hence the joint probability becomes a simple product, shown in Equation 5.3.

cJd;)fh

%

%
/0#c0d;) ,  { ,;;;,9 )/,&|/0

cJd)@ )/,& /

(5.3)

Which model would the ML approach choose from the three in Figure 5.2? The higher the
polynomial order, the better the fit to the data, and the higher the likelihood. Thus ML would
choose the most complex model from

, which for Figure 5.2 is the quadratic polynomial

(3`
). Imagine a polynomial with the same number of coefficients as there were data
points. This curve would pass through each and every point exactly, causing all residuals
to vanish. In this situation the likelihood is unity, and ML would consider it the best fit for
the data. This is despite the fact that the fitted polynomial would probably be contorted and
non-intuitive, especially for values of outside the range of the original data. Moreover, if
polynomials with degree i

are allowed, the ML fit is not even unique. ML has a strong

tendency to over-fit, as the complexity of the model is not taken into consideration.

5.2.2 Interchangeability of probability and code length

Often it is inconvenient to work with probabilities. By taking the negative logarithm of a
probability, we obtain what is called a code length in Shannons information theory [Sha48].
The MML criterion for model selection introduced in Section 5.1 constructs codes, or messages, for comparing different models. Code lengths are measured in bits when the logarithm
is to base 2, and nits when natural logarithms are used. Where probabilities are multiplied,
code lengths are added, and where probabilities are maximized, code lengths are minimized.
Probabilities and code lengths are interchangeable using the rules in Table 5.1.
For example, reconsider the maximum likelihood formula in Equation 5.2. When phrased in
terms of code lengths, ML chooses the model which minimizes the code length of the data,
called the negative log-likelihood. This arrangement is shown in Equations 5.4 and 5.5.

158

Chapter 5. Information Theoretic Local Segmentation

Probability

c0d )pZ
/
- 4

Code length (bits)

)Z/
2B

maximize

c0d)Z/

minimize

c0d)Z/c0d)+-/
cJd)pZ/[b)f ,;
/

)Z/ u)p/
)Z/b)f , /

Table 5.1: Interchangeable use of probability and code length


k

argmin
 t 

'BJc0d)+h

argmin
 t 

r)fh

%

%

(5.4)
(5.5)

5.2.3 Penalized likelihood

The maximum likelihood approach to model selection tends to over-fit the data, as it does
not consider the model complexity. Human minds and science tend to prefer the simplest
explanation of a data set which still has some useful predictive power [Tho18]. To incorporate this desire, many techniques have been devised which penalize the likelihood function to
discourage over-fitting. Penalized likelihood criteria typically have the form of Equation 5.6,
where

 )/ is a penalty term. The penalty term may depend on features of the data and

model. For example, the amount of data, or the number of model parameters.
"!
k

argmin
 t

)fh

%

/u

 ) ,.hv/

(5.6)

Over the years many penalized likelihood techniques have been developed for particular
domains [Kuh00]. The most enduring approach has been the Akaike Information Criterion, or AIC [Aka92, Boz83]. AIC is approximately equivalent to minimizing the expected
Kullback-Leibler distance [KL51, DDF 98], and has been applied to various problems such
as Markov chains and segmenting time series data. Penalized likelihoods methods have been
mostly superseded by other techniques, discussed later.
The local segmentation model selection criteria of Chapter 4 may be interpreted as a penalized likelihood technique. The Q -means clustering algorithm attempts to find the maximum

5.2 Statistical model selection

159

likelihood clustering for the pixel data, under the assumption of Gaussian clusters. The mean
separation criterion may be considered the penalty term. When the means are not separated
enough, the penalty term becomes infinite, and hence that model can not be chosen.

5.2.4 Bayesianism
Equation 5.7 shows Bayes formula for the probability of a model given the data, usually
referred to as the posterior probability of the model. As c0d)+hv/ is constant with respect to
different models for the same data, it may be disregarded. The resulting posterior probability
is proportional to the product of the model probability and the probability of the data with
respect to that model, the latter being the familiar likelihood.

c0d)

 %

hj/0

c0d)

$#

cJd;)fhv/

hv/


c0d) / 2c0d)+h
cJd;)fhv/

%

GIcJd;)

/[cJd;)fh

%

(5.7)

The model probability, c0d) / , is usually referred to as the prior probability of the model.
The prior defines a probability distribution over all possible models and their parameters, and
must be supplied by the user. It allows the incorporation of relevant background knowledge
before the data is observed. Ignorance can be expressed using a uniform prior, which gives
equal probability to all models. Thus the prior term may be ignored, and using Bayes rule
becomes equivalent to using the maximum-likelihood method of Section 5.2.1.
The resulting posterior is a probability distribution over all models. If any of the model parameters are continuous, the posterior will be a density, and any specific model will have zero
posterior probability attached to it [BO95]. Some Bayesians will insist that, for inference
problems, the posterior distribution is a sufficient final result they believe it is unsound to
select a specific model from it [OB94]. However, some applications require either a single
best model, or at most a few likely ones.
Collapsing the posterior distribution results in a single best model, called a point estimate.
Common point estimates are the mode, mean, or median of the posterior distribution, each
corresponding to different loss functions [OB94]. The mode of the posterior is most commonly used, and is called the maximum a posteriori (MAP) estimate. The MAP estimate
picks the model which makes the data most likely, and is shown in Equation 5.8.

160

Chapter 5. Information Theoretic Local Segmentation

kX
&

argmax
 t%

c0d) /c0d)fh

%

(5.8)

The main problem with the MAP estimate is that, although it picks the peak of the posterior
distribution, this peak may not have much probability mass associated with it. An alternate
lower peak may have much more posterior probability mass surrounding it, and could be a
better overall estimate. Figure 5.3 gives an example of this situation for the case of a model
with a single continuous parameter. When the model consists only of discrete parameters,
the MAP estimate does not suffer from the same difficulties, because neighbouring (discrete)
points in model space are not necessarily related.
Posterior
probability

Model
MAP estimate

Alternate estimate

Figure 5.3: The Bayesian MAP estimate may not always coincide with the largest peaked
zone of probability mass.

5.2.5 MDL : Minimum Description Length

Model selection has moved toward information-based interpretations of Bayess rule. Rissanens original Minimum Description Length (MDL) criterion [Ris78] is given in Equation 5.9, where )+h

%

/ is the negative log-likelihood of the data, i  is the number of model

parameters, and iY is the number of data items. For the regression example in Section 5.2,
i   uA
, the number of polynomial coefficients, and i+Y#i , the number of data points.

5.2 Statistical model selection

k
Y

161

argmin
 t

F'& i  u
{

i 

B
{

iYu()fh

%

(5.9)
/

The B & function is pronounced log star. It estimates the length in bits for encoding an
{

. Its computation is outlined in Equation 5.10, where 2.865 is

arbitrary integer, D
F, 9

a normalization constant which ensures decodability. The first B term encodes the value
of the integer. To decode the value, the decoder needs to know how many bits were used to
encode it. This is provided by the double-B term. The triple-B term encodes the number
of bits used for the double-B term, and so on. Eventually a point is reached where encoder
and decoder have mutual understanding. The log star term is usually small compared to the
overall code length, so is sometimes ignored when applying Equation 5.9 [BO95].

F & )O
NB uXB ubB )+B Z/ uXB )+B )pB Z/y/|u$
{
{
{
{
{
{
{

(5.10)

This MDL criterion may be interpreted as a penalized likelihood or Bayesian method, with
the model order costing B & i  bits, and each model parameter costing F
{

iY

bits.

The model parameters are transmitted more accurately only if justified by more data. MDL
has been developed over the years [Ris87, Ris00], but each formulation has one or more
of the following drawbacks: not being invariant under transformation of the data or model
parameters, poor performance with small amounts of data, an inability to specify useful prior
knowledge about the data, and a focus on selecting a model class rather than a particular
model [WF87, BO95]. In local segmentation, a model class would be Q2

or Q2:
, but

the model class does not specify any particular parameter values, such as which segment
each pixel belongs to, or the segment means.

5.2.6 MML : Minimum Message Length

Wallaces MML criterion [WB68, BW72, WF87, WD00] is an information theoretic criterion for model selection whose origins even pre-date AIC and MDL. Good introductions to
MML are found in Oliver et al [OH94, OB94, BO95]. MML proposes that, for each model
being considered, a two part message containing a description of the model and the data

162

Chapter 5. Information Theoretic Local Segmentation

given that model be constructed. The message having the shortest overall message length is
considered to contain the best model for the data. This is shown in Equation 5.11.


k

argmin
 t

) / u()fh

%

(5.11)
/

Each message must be constructed in such a way that it can be decoded by a receiver given
only the same prior knowledge. That is, the message must be a lossless encoding of the data.
If the data and model parameters are all discrete, MML is equivalent to the Bayesian MAP
estimate in Equation 5.8, due to the equivalence of probability and code length.
The fundamental difference to Bayes rule occurs when the model consists of one or more
continuous parameters. To transmit a continuous value with a finite number of bits, it must
be quantized. MML provides a framework to optimally quantize the prior density. The
quantization may be interpreted as modifying the shape of the prior density such that peaked
regions with little probability mass are flattened out, and high mass regions are somewhat
boosted [Far99]. The posterior mode in the MML situation may therefore be different to the
Bayesian MAP estimate when the models have continuous parameters.
Quantizing model parameters introduces two complications. Firstly, the decoder does not
usually know which quantization levels were used. Secondly, data usually needs to be encoded with respect to specific parameter values, not parameter ranges. The first issue is resolved by including some bits for encoding the quantization bin sizes, the so-called accuracy
of parameter values (AOPVs) [OH94]. The second problem is handled by computing an expected message length rather than an absolute message length. The expectation is computed
by averaging all the message lengths that would result from using all possible parameter
values within the quantization regions.
MML is essentially a Bayesian method at heart. By converting Bayes formula in Equation 5.7 to code lengths, the MML formula in Equation 5.11 is arrived at. MML and


Bayesianism both advocate the incorporation of prior beliefs via cJd;) / . This is unlike MDL,
which attempts to do away with priors altogether [BO95]. An inference produced by MML
has the advantage of being invariant under linear and non-linear transforms of the parameter
space. MML also works well for small amounts of data [BO95, WD00]. When the amount

5.3 Describing segmentations

163

of data is very large, the length of the second part of the message dominates, and MML
gracefully reverts to maximum likelihood, just as Bayes rule and MDL do.

5.3 Describing segmentations

Segmenting pixels into a fixed number of classes results in two sets of information: class
labels denoting which class each pixel belongs to, and one or more parameters describing
the properties of each class . Consider segmentation of the noisy pixels, 0? , in Figure 5.4. If
a threshold of 60 is used, the first class has

\ members averaging # 

B
, and the

second class { I members averaging # { IF . If the classes are numbered from 1, then
e contains the class labels for each pixel. The resulting local approximation, , is computed
by replacing each pixel with the mean of the segment it belongs to.

20
? = 24

23

22 84
21 81
22 87

I
B

#
{ #F

$

{ $

1 1
e = 1 1
1 1

2
2
2

22 22 84
= 22 22 84
22 22 84

Figure 5.4: Local segmentation of a window into two classes.

The segment map, e , consists of

tation. If 

labels one for each pixel taking part in the segmen-

is the maximum number of segments allowed, then each label could take on a

value from 1 to  . By imposing a canonical ordering on the elements, the segment map, e ,
can be treated as a 1-D vector, with elements indexed as !@ . Figure 5.5 illustrates the use of
a canonical raster ordering for labels within a segment map.

!
e = !
!

! {

p
! E

!
r

!g
B,.
,;;;,

Figure 5.5: Raster order convention for labels within a segment map.

164

Chapter 5. Information Theoretic Local Segmentation

5.3.1 Possible segmentations

If there are

pixels, each of which may belong to one of 

segments, then there are 

distinct ways to assign pixels to segments. That is, there are 

possible values for e .

Consider the local segmentation of a


window into 1 or 2 segments. In this case
and I
, so there are

#


 possible segment maps. Figure 5.6 shows all of them, with

light and dark squares denoting !O
and !.#

respectively.

Figure 5.6: The 16 possible binary segment maps.

5.3.2 Canonical segmentations

Figure 5.6 has the feature that the second row is an inversion of the first row, the role of labels
having been interchanged. The labels themselves are not crucial what is important is the
uniqueness of the segment map pattern. Figure 5.7 shows the 8 canonical segmentations after
the inverted duplicates are removed. When a
is always

, the number of canonical segmentations

. Note also that the bottom right hand pixel is always in the same dark

segment. Because the labels are interchangeable, one degree of freedom is removed, so only


labels need to be determined relative to a predetermined one.

Figure 5.7: The 8 canonical binary segment maps.

Listing 5.1 describes a recursive algorithm which, given

and  , generates only canonical

segmentations. The algorithm fixes the class label for the first pixel as 1. It then does a

5.3 Describing segmentations

165

/ / Global constants
/ / M : number of pixels
/ / K : maximum class label ie. number of segments
//
//
//
//

Parameters to generate function

[c] : current segment map vector so far
m : number of labels in [c] so far
k : largest label used so far

/ / Return parameter of the generate function

/ / list : list of canonical segmentations ([c1],[c2],[c3],...)
proc g e n e r a t e ( [ c ] , m , k , l i s t )
i f ( m M) t h e n
l i s t . insert ( [ c ] )
/ / we are done with this vector
else
f o r i : = 1 t o min ( k + 1 , K)
g e n e r a t e ( [ c ] . prepend ( i ) , m+ 1 , max ( i , k ) , l i s t )
endfor
endif
endproc
proc main
list = ()
generate ( [ 1 ] , 1 , 1 , l i s t )
endproc

/ / empty list
/ / generate all canonical K-class M-tuples

Listing 5.1: Algorithm to generate canonical segmentations only.

depth first traversal to determine the next label in e . The only labels considered at each step
are those up to and including the largest label used so far, plus the next one along, but always
restricted to the range
B,' . This ensures that no duplicates are generated.

5.3.3 Valid segmentations

For the binary segment maps generated so far, no specific requirements on spatial structure
have been imposed. One may wish to insist that the labels for a given segment be spatially
connected. In Section 2.6 the concept of pixel connectedness was introduced. For the example of
Z[
windows, one could insist that the pixels within a segment be 4-connected.
Figure 5.8 shows all the valid binary segment maps having, at most, two sets of labels.
Valid segmentations may be generated by generating all canonical segmentations with List-

166

Chapter 5. Information Theoretic Local Segmentation

Figure 5.8: The 7 valid 4-connected binary segment maps.

ing 5.1, and removing those which fail a spatial-connectedness test. The algorithm for this
test is relatively simple, and is not included in this discussion. Table 5.2 shows that as
and 

increase, the proportion of canonical segmentations among all possible segmentations

decreases, as does the proportion of valid segmentations within the canonical set.
Pixels in
window ( * )
4
9
9
9

Max. number
segments (+ )
2
2
3
4

Number of segment maps

Possible Canonical Valid
16
8
7
512
256
133
19,683
3,281
915
262,144
11,051
2,411

Table 5.2: Number of possible segmentations for various window configurations.

5.3.4 Segmentation parameters

Given the noisy data,

? , a segment map, e , and  , it is a simple matter to compute the

segment parameters. For the case of additive zero-mean Gaussian noise, the mean is optimal,
but this could be modified for different types of noise. Figure 5.9 gives a sample  block
of pixels and two candidate segment maps: the homogeneous e , and the heterogeneous e { .

20 22
? = 24 21
23 22

84
81
87

1 1 1
1 1
1 1 1

e = 1

1 1 2
e { = 1 1 2
1 1 2

]I

Figure 5.9: Noisy pixels and two candidate segment maps.

The number of segments The number of segments is equal to the number of different labels used in the segment map, e , and shall be referred to as b)feg/ . Because e is
homogeneous, b)fe /JD
. For the heterogeneous segment map, b)fe { /0I
.

5.4 Using MML for local segmentation

167

The segment populations The population,

, for segment  is equal to the frequency of

pixels in the segment map having label  . Segment map e only has one population, so

I

# , whereas e { has two segments, so

 and
{ # .

Representative values for each segment In this thesis, the noise is assumed to be additive
zero-mean Gaussian, for which the arithmetic mean is an efficient estimator of the true
population mean. Each segment has one mean, calculated using Equation 5.12. For
e , #'$g
4 , while for e { , # #

B
and # { $F .

#.

+ ,

(5.12)

5.4 Using MML for local segmentation

The FUELS algorithm developed in Chapter 4 has two main characteristics which could
be improved. The first is that only two candidate clusterings are considered for each local
region. This limitation will be tackled later in Section 5.14. The second drawback is that
FUELS model selection criterion depends solely on the numerical intensity difference between segment means. Specifically, they must be at least H& units apart, where & is the
image noise level, and H

3 . Thus FUELS is unable to distinguish two segments having a

contrast difference below F& grey levels.

A different model selection criterion is required to handle adjacent segments of low contrast.
In Section 5.2.6, the MML approach to model selection was described. MML evaluates a
model using the length of a two part message which describes jointly the model and data.
The first part of the message describes the model and its parameters. The second part of the
message describes the data with respect to the prefixed model. Each message must be an
unambiguous and efficient encoding of the data it describes. The model associated with the
message of overall shortest length, measured in bits, is the preferred model for the data.

5.4.1 A message format

Figure 5.10 proposes a message format for encoding noisy pixels under a local segmentation


model. The first part of the message is the model, , which contains all the information

168

Chapter 5. Information Theoretic Local Segmentation

required to construct a local approximation to the underlying pixel values. The segment
map, e , states which segment each pixel belongs to. The segment map is followed by the
means for each segment. The number of means depends solely on the number of unique
segments in e . This value is denoted b)+eg/ , and may be derived from the segment map, e .

Model

Segment map

Data

.....

p | c , 1 , 2 . . . . .

K(c) segment means

Noisy pixel data, given the model

Figure 5.10: The proposed local segmentation message format.

Equation 5.13 states that the message length of the model part is equal to the sum of the
lengths of each component it contains. This is only true if each component of the message is assumed independent. The details of encoding each component depend on our prior
expectations for the pixel values. This will be discussed soon in Section 5.4.2.

) /0-)+eg/u/.

-104

)B#Z

eg/

(5.13)

The second part of the message encodes the data with respect to the model preceding it.
In our case the data consists of noisy pixels, ?  ? . Pixels from a greyscale image are
usually quantized to one of _

integers. If encoded independently under the assumption of

uniformity, each pixel value requires at most F

{

bits, but relative to a model (segment

map and means), a more efficient encoding is possible.

It was shown in Section 4.2.6 that a quantized pixel value, 5 , could be considered to have had
a true value anywhere in the range 5 '-,.5u-4B/ . In Chapter 4, it was assumed that clean
images were corrupted by additive Gaussian noise 6W8

)f-,~& { / , with &

assumed constant

across the image. Thus it is assumed that pixels within a segment are normally distributed
around the segment mean. Equation 5.14 calculates the length of the data component (the
negative log-likelihood) given the model parameters.

5.4 Using MML for local segmentation

)fh

% 

/0D

169
] 
2 q

< 4
36587

K

KM K

,
&


] 

vB

(5.14)

5.4.2 A uniform prior

Our prior expectation regarding the types of local segmentation structure present in an image
directly impacts on the encoding of the model parameters. Like FUELS, the model domain
is restricted to  windows (

O ) consisting of a maximum of two segments ( `D

).

canonical binary segment maps. For
In Section 5.3.2, it was shown that there are only

the moment, let us assume that each segment map is equally likely. Equation 5.15 shows that
e may therefore be encoded with

3
bits. Once e is known, the number of segments,

b)feg/ , and the number of pixels belonging to each segment are implicitly known.

)feg/0<B

I



(5.15)

How many bits should be used for each segment mean? An optimal number will be discussed
in Section 5.7. For the moment a useful approximation is considered. We already know that
a pixel value can be encoded with at most B

_
{

bits. We also know that a denoised pixel is

usually rounded back to an integer at the final stage of processing. Thus a reasonable number
of bits for a segment mean could also be B
{

bits. Equation 5.16 gives an expression for

the overall length of the model component of the message under this assumption. Note that
because

is constant for all models, it could be ignored.

r) /0#

ub)feg/-F
{

(5.16)

5.4.3 A worked example

In this section an example of the MML model selection criterion will be given. Figure 5.11
shows the noisy   window which will be used. The global noise standard deviation is

170

Chapter 5. Information Theoretic Local Segmentation

15 14 15
? = 13 17 12
16 11 10

&'D
B4

(assumed)

Figure 5.11: A window of noisy pixels.

assumed to be 1.5. Two candidate models will be considered, coinciding with those which
would have been generated by FUELS.
Figure 5.12 shows the first candidate model. It is the simplest one possible a homogeneous
segment map and a single mean. This model is 16 bits long 8 bits for the segment map
and 8 bits for the average pixel value.
1
e = 1
1

1 1
1 1
1 1

#'D
-4F

\

 ) /Ju$
C2B
BFCD
; bits

% 
{
)+h
/J#
1NB bits
9#
)
hv/0 4B bits

Figure 5.12: Candidate model 1 and its message length calculation.

Figure 5.12 shows the second candidate model. It is the result of iteratively thresholding the
noisy pixel values, just as FUELS does. The resulting heterogeneous segment map fits the
pixels well. Its model length is 8 bits longer than the first candidates, as two means must be
encoded. However the data encoding length is much shorter. This is due to the pixels being
closer in value to their segment means.
1
e = 1
1

1 1
1 2
2 2

# 

#
{ 
F

$

{ $

r) /J$u9
*2B
B17
1 bits
% 
{
 )fh
r
/J
;-
F bits
9#
r)
hv/0B-4
F bits

Figure 5.13: Candidate model 2 and its message length calculation.

Under the MML criterion, two candidate models are compared using their overall two-part
message lengths. Figure 5.14 illustrates the comparison using the two example models just
described. Candidate 2 has the shorter overall message length, and therefore Figure 5.13 is
the preferred model for the noisy pixels in Figure 5.11.
Using FUELS model selection criterion, Candidate 2 would be rejected because its two
%

means are too similar:
A
B

$
F . Unlike FUELS, the MML criterion is not limited

5.4 Using MML for local segmentation

Candidate 1

171

: 16 bits


Candidate 2

h : 24.93 bits

: 24 bits

h : 14.23 bits

Figure 5.14: Comparing two message lengths.

to a fixed minimum contrast difference. In this example, MML made the better inference. It
should be noted that the -test threshold from Equation 4.27, not used by FUELS, is equal to
3.71 for this example. In this case the -test and MML would concur.

5.4.4 Posterior probability

The overall length of a message in MML is related to the Bayesian posterior probability. The
interchangeability of code length and probability ensure this fact. Equation 5.17 describes


how to compute the posterior probability of a model, By) k / . The denominator normalizes
the probabilities over the set of models

By) k /J

considered.

^-;=
: < Y 4


 - < Y 4
 t4

(5.17)

For the example in Section 5.4.3, Candidates 1 and 2 have posterior probabilities -m
 and
-N respectively. Here the best model is over six times as likely as its nearest competitor.

This is good evidence for the existence of two local segments, but not compelling evidence.
These posterior probabilities may be interpreted as saying that, in 13% of cases like this, the
less likely candidate may actually be the correct one.
A corollary of Equation 5.17 is that the difference in message length of any two models
may be used to measure the relative posterior probability of the two models. For example,
Candidate 2 had a message length 2.7 bits shorter than Candidate 1. In terms of posterior

probability, the Candidate 2 is

{ O-4 times as probable as Candidate 2. This is a useful
shorthand for assessing the relative merit of two competing models.

172

Chapter 5. Information Theoretic Local Segmentation

5.5 Application to denoising

The FUELS denoising algorithm generates two candidate local segmentation models. It
chooses what it considers to be the better one, and uses the resulting local approximation to
estimate the underlying noiseless pixel values. It is a simple matter to replace the existing
FUELS model selection criterion with the MML one described in Section 5.4. Let the resulting denoising algorithm be called Pseudo-MML. This name has been chosen to avoid
confusion when MML is applied differently later.
Figure 5.15 compares the RMSE performance of FUELS and Pseudo-MML for denoising
montage. The true value of & was given to both algorithms. Overlapping averaging and
DNH were disabled, as the primary goal is to observe the effect of changing the model
selection criterion. Pseudo-MML only considered the same two candidate models as FUELS.

20

FUELS =0, DNH disabled

Pseudo-MML

18
16

RMSE

14
12
10
8
6
4

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.15: RMSE comparison for denoising montage, true supplied.

The results show FUELS to outperform Pseudo-MML at all noise levels, with the discrepancy increasing along with the noise level. Figure 5.16 shows where the two algorithms

5.6 A better segment map prior

173

differed in their model selections. Pseudo-MML is much more likely to choose QGE
than
FUELS. One advantage of this is that it was able to identify the low contrast oval edge around
the h. However, it also chose QLI
in obviously homogeneous regions.

Figure 5.16: (a) noisy U middle section of montage; (b) FUELS model selection;
(c) Pseudo-MML model selection. Black denotes " and white "Cb .

5.6 A better segment map prior

In Section 5.5, it was shown that the Pseudo-MML model selection criterion has a tendency
to over-fit. It chose Q

more often than it should have, resulting in less smoothing and

poorer RMSE results compared to FUELS. Although it only considers two models, PseudoMML uses the same number of bits for encoding any segment map encountered. This corresponds to a uniform prior probability distribution for e .
The only way to modify Pseudo-MMLs behaviour is to change the prior. The analysis of
FUELS in Chapter 4 showed that, for typical images, around 50% of  regions could be
considered homogeneous. Thus it would be more realistic to assign a higher prior probability
to the single homogeneous segment map, and share the remaining probability equally among
the remaining heterogeneous segment maps. This arrangement is shown in Equation 5.18.
Note that 0.5 is an arbitrary choice, and its optimization will be considered in later sections.

c0d)+eg/J

-4
{


Z q\

b)feg/0D

b)feg/0#

(5.18)

174

Chapter 5. Information Theoretic Local Segmentation

Equation 5.19 gives a new expression for the overall message length using this prior. In
terms of encoding e , the new prior corresponds to using 1 bit for homogeneous cases, and
nearly 9 bits for each heterogeneous case. Although not implemented as such, this could
be interpreted as a two step process. First, a single bit states whether QP
or Qo
. If
QLE
, nothing else needs to be sent. If Q #
, another B
{

FB7# 4F bits are used to state

the which of the QjO

segment maps is relevant. It should be noted that among the various
heterogeneous segment maps, some are probably more likely to occur than others. This issue
will be discussed later in Section 5.13.

)

9#

hv/02B
{

c0d)+eg/ u/b)feg/-B
{

_bu)+h

% 

(5.19)

5.6.1 Results
Figure 5.17 is the same as Figure 5.15 except that there is an extra entry for Pseudo-MML using the non-uniform prior of Equation 5.18. The more suitable prior has resulted in PseudoMMLs performance reaching that of FUELS.
Figure 5.18 shows where the three algorithms graphed in Figure 5.17 differed in their model
selections for the middle section of montage. The non-uniform prior improved PseudoMMLs model order selections, because most of the spurious Q

decisions are gone. It

was still able to discern most of the low contrast oval edge around the h, along with the
horizontal boundary below the oval. This example illustrates the importance of choosing a
suitable prior. It is especially important when the window is small, because the model part
of the message contributes significantly to the overall message length.

5.7 Optimal quantization of segment means

Under the local segmentation framework, each segment mean is used as the location parameter for a Gaussian distribution. The spread parameter for the distributions is assumed common and equal to the global noise standard deviation, & . The noisy pixel values from each
segment are encoded with respect to the appropriate distribution, illustrated in Figure 5.19.

5.7 Optimal quantization of segment means

20

175

FUELS =0 and DNH disabled

Pseudo-MML (uniform prior)
Pseudo-MML (non-uniform prior)

18
16

RMSE

14
12
10
8
6
4

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.17: Non-uniform prior: RMSE comparison for denoising montage, true supplied.

Figure 5.18: Model selection comparison when LP , black denotes " and white "P :
(a) FUELS; (b) Pseudo-MML; (c) Pseudo-MML with a non-uniform prior.

In Section 5.4.2 it was arbitrarily decided that B

{

bits was a reasonable accuracy for

encoding a segment mean. If the segment means were encoded using fewer bits, the model
part of the two part message would be shorter. Correspondingly, the data part should expand,
as the segment means are less precise. There should exist a quantization level for each
segment mean which optimally trades off the length decrease of the model part with the

176

Chapter 5. Information Theoretic Local Segmentation

Figure 5.19: Pixels are encoded relative to their segment mean.

length increase of the data part.

Much of the MML literature is concerned with optimally quantizing the parameter space for
various standard statistical models [WD00]. Under certain uniformity assumptions [OH94],
the optimal bin size, $>
tion 5.20, where

@? A

, for the sample mean of a normal distribution is given in Equa-

is the number of data points used to compute the sample mean.

$>

@?BA

&

(5.20)

As the variance increases, the accuracy with which we are willing to encode the mean decreases. This is because the distribution has a wider spread, so its exact location is not as
important. As the amount of data increases, a higher accuracy is warranted. Equation 5.21
describes the number of bits required to encode an optimally quantized segment mean, where
_

is the intensity range of the image and & is the standard deviation of the noise.

)=#|/0B

_
{ $>

@?BA

<B
{

_
&

(5.21)

A segment mean encoded with this number of bits is still decodable. Both & and _
assumed to be a priori known, and

are

may be determined from e , the first model parameter

encoded. Figure 5.20 compares the number of bits needed to encode a segment mean using
the two methods described. The original method uses exactly 8 bits, regardless of the noise
level. The optimal quantization method uses much fewer bits, and the function grows slowly.
Quantization of segment means causes one difficulty. Each pixel is no longer encoded with
respect to a fully accurate segment mean. Rather, a quantized range of possible segment

5.7 Optimal quantization of segment means

10

177

Fixed
Optimally quantized

Accuracy of mean (bits)

Number of pixels in segment (m)

Figure 5.20: Different encodings for a segment means when Rb and CP .

means has been communicated. Which mean from the range should be used? Rather than
taking an arbitrary point estimate, such as the midpoint, MML prefers to compute the expected message length. That is, the average of all message lengths that would result if every
possible point estimate in the quantization range was tested. It is often assumed that each
value in the quantization range is equally likely, corresponding to a locally flat prior density
in the range. A convenient closed form approximation to the expected message length can
sometimes be derived [WF87].
For the purpose of this thesis, a point estimate equal to the sample mean will be used. This
simplifies the computation and reduces the running time, without hopefully causing too much
variation in message length. The resulting expression for the approximate expected message
length is in Equation 5.22.

r)

9#

hv/ %E2B
{

c0d;)feg/ u
.

-104

F
{

_
&

u()fh

% 

(5.22)

178

Chapter 5. Information Theoretic Local Segmentation

5.7.1 Results
Figure 5.21 provides RMSE results for denoising montage. The algorithm using quantized
means is referred to as MML to distinguish it from the earlier Pseudo-MML. Both algorithms use the non-uniform prior of Equation 5.18 to encode the segment map. DNH and
overlapping are still disabled, and the true value of & is provided.

18

FUELS =0 and DNH disabled

MML
Pseudo-MML

16

RMSE

14
12
10
8
6
4

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.21: RMSE comparison for denoising montage, true supplied.

The RMSE performance of the three algorithms is mostly indistinguishable. FUELS does
marginally better at low noise levels, while MML does slightly better when the image is
very noisy. Figure 5.22 compares the value of Q Pseudo-MML and MML chose for each
pixel when &\ . Their model selections are very similar, so Figure 5.22c highlights the
differences. Mid-grey indicates that both algorithms chose the same value of Q , white that
MML chose a higher order model, and black that MML chose a lower order model.
Interestingly, if there was a difference, it was always that MML chose Q\
compared to
Pseudo-MMLs Q]
. The quantization of segment means has reduced the cost of Q\

5.8 Posterior blending of models

179

Figure 5.22: Model order selection when a , black denotes " , white "Oa :
(a) Pseudo-MML; (b) MML; (c) difference: white shows where MML chose "7P
over Pseudo-MMLs " .

models enough for them to be selected more often. In fact, the message lengths calculated
by the MML denoising algorithm are always shorter than those of Pseudo-MML. For a given
segmentation, both algorithms use the same number of bits for the segment map, but MML
uses fewer bits for the means. Because both use the true mean as the point estimate, the pixels
are encoded in the same number of bits too. Thus MMLs messages are always shorter.

5.8 Posterior blending of models

In Section 5.4.4 it was described how each models posterior probability could be calculated
from its message length. The posterior probability may be interpreted as a confidence in each
models ability to approximate the underlying image. This type of information is not really
available when using FUELs model order selection criterion.
The MML criterion chooses the model with the overall shortest message length, corresponding to the highest posterior probability. This is a good thing to do when one wishes to infer
the best complete model for the window. If one is only interested in the true pixel values,
and not the most appropriate segment map and segment means, then the best thing to do is
to blend over all models weighted by their posterior probability. Parameters of non-interest
should be integrated out [BS94].
In the MML and Pseudo-MML algorithms, the local approximation suggested by the most
probable model alone was used for denoising. It is possible to produce a blended local ap k DGFHB , which is a linear combination of the local approximations from each
proximation, E
z

180

Chapter 5. Information Theoretic Local Segmentation

model considered. The weight given to each model is exactly equal to its posterior proba

bility. Equation 5.23 describes its calculation, where ZB) / is the posterior probability of


model , and k  is the local approximation associated with model .

k DGFHB
E


z

 t%


B) / k 

(5.23)

The blended local approximation may alternatively be considered those pixel values inferred


by a composite model, DGFIHU z . This is shown in Equation 5.24. Combining the predictions
from various experts is a popular technique in computer science, particularly in image
compression [STM97, MT97].

DGFIHU


z

 t

ZB) /

k DGFIHU
J
z

k LKNMPORQS

(5.24)

Recall the two example candidate models from Section 5.4.3. Figure 5.23 applies Equation 5.23 to blend the local approximations associated with the two candidates. In this case,
rounding back to integers would cause single best local approximation to equal the posterior
blended one. However, the denoised pixels could be retained at a higher accuracy if further
processing was to be done.

k DGFHB
E
$-m
T
z

13.8 13.8 13.8

13.8 13.8 13.8 uX 4T
13.8 13.8 13.8

15.0 15.0 15.0

15.0 15.0 11.0 
15.0 11.0 11.0

14.8 14.8 14.8

14.8 14.8 11.4
14.8 11.4 11.4

Figure 5.23: Posterior blending of the two example models from Section 5.4.3.

5.8.1 Results
Figure 5.24 compares the RMSE performance of the posterior blended version of MML to
the MML algorithm which simply uses the most probable model. For &

, posterior

blending improves the RMSE performance of the MML algorithm. It may also be possible
to blend models in the FUELS algorithm, but a metric would have to be invented to quantify
the suitability of each candidate model, whereas this metric comes naturally for MML.

5.9 Incorporating Do No Harm

18

181

FUELS =0 and DNH disabled

MML
MML with posterior blending

16

RMSE

14
12
10
8
6
4

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.24: RMSE comparison for denoising montage, with true supplied.

The ability to blend over multiple models is very useful when there is no clear mode (peak) in
the posterior distribution. In this case, the posterior mean has been used. This is related to the
use of a squared error loss function [OB94], which also happens to be what the RMSE metric
uses. Combining multiple local approximations together tends to produce a better overall
local approximation. In the case where one model has nearly all the posterior probability
associated with it, the algorithm behaves just like the non-blended version.

5.9 Incorporating Do No Harm

5.9.1 The FUELS approach to DNH
In Section 4.9 the do no harm (DNH) idea was applied to FUELS. DNH has the effect of
limiting the worst case behaviour of a denoising algorithm. In FUELS, if any pixel value
in the optimal local approximation differs from its noisy value by more than F& , the local

182

Chapter 5. Information Theoretic Local Segmentation

approximation is rejected, and the noisy pixels are passed through unmodified. It is a simple
matter to incorporate this post processing to the posterior blended pixel values produced by
the MML denoising algorithm.
The results so far have shown FUELS and MML to perform similarly. Figure 5.25 compares
their RMSE performance when both utilise FUELS-style DNH. As expected, their performances are much better at lower noise levels when DNH is enabled. When &I , MML
consistently beats FUELS, but the gap is not significant. Although not shown, their WCAEs
are identical at all points.

18

FUELS =0 and DNH enabled

MML + FUELS-style DNH

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.25: RMSE comparison for denoising montage, true supplied.

5.9.2 Information based DNH

In FUELS, the HC& threshold was common to both the DNH and model selection components. It may be beneficial for the MML algorithm to use an information theoretic approach
to DNH which fits in more appropriately with its model selection criterion and posterior

5.9 Incorporating Do No Harm

183

blending. When DNH is invoked by FUELS, the noisy pixel values are passed through unaltered. Under MML, this corresponds to encoding pixels as they are, without respect to any
particular model. A raw encoding of

noisy pixel values requires, at most,

B
{

bits.

For example, a raw encoding of a   window of 8 bpp greyscale pixels would need 72 bits.
The raw encoding may be interpreted as a null model. The null model could be placed into
the existing pool of segment-based candidate models, and be judged alongside them. This
is similar to the -test method in Section 4.7.2, where a null and alternative hypothesis are
compared. The null model could actually have the shortest two-part message length, causing
it to be selected as the best model. This is like FUELS where DNH is invoked when all
other models appear unreliable. When posterior blending is used, the null model becomes
part of blend, with its local approximation being equal to the original pixel values.
For this idea to work correctly, the null model and the standard segment-based models must
be compared fairly. Each model needs to be prefixed by a binary event stating whether a
standard, or null, model is to follow. Figure 5.26 illustrates this arrangement, where DNH
denotes the use of the null model.
(a)
(b)

DNH
Not DNH

h
h

% 

Figure 5.26: Incorporating DNH: (a) null model encodes the data as is; (b) standard model.

Let c0d)+hji'k2/ denote the prior probability associated with the prefix for the null model. For
example, if c0d)+hji'k/r-
the null model would have a prefix of length )'B
-N
bits. The prefix for a standard, segment-based model would be only )2B
{

-m
/r
{
-N/#-m

bits, because according to the prior, they will be used more often.
Figure 5.27 plots the proportion of pixels for which FUELS-style DNH is invoked when
denoising montage. The two curves are effectively plotting suitable values of cJd)fhjiGk/ to
use at different noise levels. Both curves have a hyperbolic shape, approximately following
Equation 5.25. When &G# , the prefix has zero length for the null model and infinite length
for any standard model. If this formulation was used, DNH would be invoked for every pixel

184

Chapter 5. Information Theoretic Local Segmentation

FUELS =0 and DNH enabled

MML + FUELS-style DNH
Curve f()=1/(1+)

Proportion of times DNH was invoked

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.27: FUELS-style DNH usage when denoising montage, true supplied.

when &2$ . This makes sense because if the noise level is zero, then all pixels are noiseless
and should remain unaltered.

c0d)+hji'k/J

ub&
T

)fhjiGk/0B

)
ub&|/

(5.25)

Figure 5.28 compares three denoising algorithms: FUELS using its DNH, MML using
FUELS-style DNH, and MML using the new information based DNH, with c0d)fh i'k/C

x)
[u& / . For &

 , MML with its information based DNH consistently outperforms the

others in terms of RMSE when denoising montage. Although not shown, similar behaviour
was observed for other images.
The worst case absolute error (WCAE) results are provided in Figure 5.29. FUELS and
MML with FUELS-style DNH have the same WCAE for all noise levels tested. When
&

3


, the information-based DNH approach consistently has a lower WCAE. This is inter-

esting because the MML DNH approach makes no explicit attempt to restrict the maximum

5.9 Incorporating Do No Harm

18

185

FUELS =0 and DNH enabled

MML + FUELS-style DNH
MML + information based DNH

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.28: RMSE comparison for denoising montage, true supplied.

allowed change in pixel intensity during the denoising step. Of course, large alterations in
intensity are unlikely to fit within the Gaussian segment models well, causing those models
to have low posterior probability.
DNH ensures that pixels do not change by more than 1& relative to the noisy image only.
However, the WCAE is taken between the denoised image and the ground truth image. Consider a noiseless image corrupted by additive noise with standard deviation & . There will
always be a small fraction of pixels whose values change by more than F& . In the worst case,
a pixel could change by _
intensity units. DNH operates relative to the noisy image, so
it is still possible for the WCAE to be very large.
Figure 5.30 plots the actual relative frequencies of DNH usage for the three algorithms tested.
MML invokes DNH less often when it uses the information based DNH than it does with
FUELS-style DNH. There is a trade-off between invoking DNH too often, whereby no denoising occurs, or not invoking it, whereby there is a risk of producing a poor local approximation which detrimentally affects the RMSE. It seems that the information based DNH is

186

Chapter 5. Information Theoretic Local Segmentation

200

MML + FUELS-style DNH

FUELS =0 and DNH enabled
MML + information based DNH

180
160

There are two identical curves here

140

WCAE

120
100
80
60
40
20
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.29: WCAE comparison for denoising montage, true supplied.

better at making this trade-off. Although it was decided in a sensible manner, the 1& threshold used by FUELS-style DNH may not be optimal. It is possible that a higher value, such
as B& , could improve FUELS results.

5.10 Combining overlapping estimates

In Section 4.11 the idea of combining estimates from overlapping windows was introduced.
For FUELS, various schemes for linearly combining estimates were tested. It was found that
an equal weighting (SR

) gave the best RMSE results. This same concept may be applied

to the MML algorithm with ease.

5.10 Combining overlapping estimates

FUELS =0 and DNH enabled

MML + FUELS-style DNH
MML + information based DNH

0.9
Proportion of times DNH was invoked

187

0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.30: How often DNH is invoked for montage, true supplied.

5.10.1 Equal weighting

Figure 5.31 compares the RMSE performance of four algorithms on the montage image:
FUELS with SM

 and SM^
, and MML with SM

 and S

. The appropriate DNH

option is used for all instances. As expected, overlapping averaging significantly improves
the RMSE results for both FUELS and MML at all noise levels. When S

, there is

little to no difference between the two. It appears FUELS is able to benefit more from the
overlapping averaging than MML. This may be due to MML having already an advantage
through its application of posterior blending of models at the same pixel position.

5.10.2 Message length weightings

Overlapping averaging combines estimates of the same pixel from different models. For FUELS in Section 4.11, the combination was restricted to a linear one. Equation 5.26 describes

188

Chapter 5. Information Theoretic Local Segmentation

18

All with DNH enabled

FUELS =0
MML =0
FUELS =1
MML =1

16
14

RMSE

12
10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.31: RMSE comparison for denoising montage, true supplied.

this again, where ;;^

of the

are the denoised estimates for the same pixel position from each

overlapping models, k is the final denoised value, and ( are the weights.

k 

( 
H

(H

(5.26)

For FUELS, various schemes for choosing the weights were assessed. Each scheme tried to
use some attribute of each model to modify the weights. The attributes attempted to capture
the quality, or goodness of fit, of a model.

Model and data posterior

In an MML framework, the message length of a model is a natural measure of its goodness of
fit. Equation 5.27 bases the weights on the unnormalized posterior probability of the model

5.10 Combining overlapping estimates

189

and data. This is done using the message length associated with the model and data from
which the estimate originally came. This attempts to give more weight to those estimates
which came from models which perhaps better explained their local region.

(H[I

 -  4^-VU XW  4

"~

(5.27)

Model and pixel posterior

An alternative to using the full message length is to use only the length of the model and pixel
in question. It could be argued that we are only interested in how well the local approximation modeled the particular pixel being combined. Equation 5.28 describes this formulation.

(H[I

 -  4^-L Y W  4

(5.28)

Pixel posterior
One final potential weighting scheme is given in Equation 5.29. The weight is proportional
to the posterior probability of the pixel alone. This attempts to measure how well the noisy
pixel fitted into the model it came from. That model was either the MAP or posterior blended
model for the window the pixel originated from. It is expected that this weight would behave

similarly to a weight based on the squared error: (To) ? k / { . This is because a pixel
message length is based on a Gaussian distribution.

(H 

 -L Y W  4

(5.29)

Results
Figure 5.32 compares the RMSE performance of four weighting schemes for montage:
SD

, and the three probabilistic ones just described. None of the variable weighting

schemes perform better than equal weighting for any values of & . The same results were
found to occur for the lenna image. For these reasons, schemes other than equal weighting
will not be explored further.

190

Chapter 5. Information Theoretic Local Segmentation

16

Model + Data
Model + Pixel
Pixel
Equal (=1)

14
12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.32: RMSE comparison for denoising montage, true supplied.

5.11 Estimating the noise level

The analyses performed so far in this chapter have assumed that the true value of & , the standard deviation of the synthetic noise, is known. In real situations, the noise is not synthetic
and its level is unknown. It typically has to be estimated from the data. In Section 4.13,
the FUELS algorithm was adapted to use the robust Immerkr noise variance estimation
algorithm [Imm96]. MML could use this estimate too.

Figure 5.33 compares the FUELS and MML algorithms when both use Immerkrs estimated value for & , rather than the true value. The main variation occurs for low & , where
the noise level is over-estimated. This causes DNH to be invoked less often, and the RMSE
results to be higher. Either way, there is very little difference between FUELS and MML.

5.12 Evolution of the message length

14

191

FUELS
MML

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.33: RMSE comparison for denoising montage, with estimated.

5.12 Evolution of the message length

Each modification to the MML denoising algorithm has so far improved its RMSE performance. This suggests that, on average, the models generated are becoming more reliable.
Under the MML framework, models which have shorter two-part messages are assumed to
better explain the data they are modeling. Figure 5.34 plots the average message length over
the whole montage image for the various MML-based algorithms explored so far. The
estimated noise level is used in all cases.
As expected, the average message length is mostly an increasing function of the noise level.
For Pseudo-MML, switching to a more suitable segment map prior reduces the average message length by about 4 bits. The move from Pseudo-MML to MML via the use of quantized
segment means provides a similar gain. The inclusion of a DNH candidate model again
improves the average message length for the MML algorithm at low noise levels.
For the algorithms without DNH, the average message length is observed to slightly increase,

192

Chapter 5. Information Theoretic Local Segmentation

Average message length per 3x3 block (bits)

75
70
65
60
55
50
45
40

From worst to best:

Pseudo-MML
Pseudo-MML + non-uniform prior
MML
MML + DNH

35
30

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.34: Evolution of the average two-part message length for the MML denoising algorithm, with estimated from the noisy montage image.

rather than decrease, when &

`
. Immerkrs method overestimates &

when it is very

low. This could increase the relative cost of two segment models, causing fine structure
to be corrupted. Usually this structure is swamped by noise, but when there is no noise
it becomes significant. The expression in Equation 5.21 for the optimal quantization of
sample means begins to break down when &|_ . This is due to an assumption of uniformity
within quantization regions [OH94]. Also, the implementation of numerical integration for
calculating the likelihood may not be accurate for very small intervals 1 .

5.13 Learning from the data

In this chapter, a new approach to local segmentation model selection has been developed
and applied to the problem of image denoising. Using FUELS as a basis, each component
1

The standard erf() and erfc() functions from the GNU C maths library v2.2 were used.

5.13 Learning from the data

193

was slowly modified or replaced to use ideas and techniques from information theory. This
helped to remove arbitrary constants from the algorithm, and allow a wider range of image
structure to be potentially discovered. The results for the new MML denoising algorithm
are not significantly better than FUELS. Using MML improves the WCAE at higher noise
levels, but only mildly improves RMSE. It is possible that only slight improvements are
possible because FUELS is already quite good Chapter 4 has already shown that FUELS
outperforms rival denoising algorithms.
The performance of MML is dependent on the prior used. When denoising, the prior consists
of & , c0d)+hji'k2/ , and c0d)+eg/ . The prior determines the length of the model part of each
message being compared. The prior is determined before processing of the image begins,
and is not modified during processing at all. It could be possible to vary the priors on a per
block basis. For example, the prior probability of each segment map could be dependent on
the posterior probability of segment maps from nearby pixels already processed. The aim
would be to exploit edge coherence. However, in this chapter the same priors will be re-used
for processing each pixel in the image.
Under the MML framework, models are assessed using their respective message lengths. The
overall message length (the posterior) is computed as the length of the model (determined
by the prior) plus the length of the data given the model (the negative log-likelihood). If the
amount of data is large, the data part will dominate and eventually swamp the effect of the
prior. If the amount of data is small, the prior exerts a large influence on the posterior [Pre89],
as illustrated in Figure 5.35. In fact, once the amount of data becomes large enough, the
model component may be ignored, and MML reduces to maximum likelihood.



% 

hI;;

% 

h3;;;;;;

% 

Figure 5.35: As the amount of data increases, the model part becomes less significant.

Because local image processing algorithms only use a small number of pixels, and different images vary in their properties, a technique for dynamically learning the prior from

194

Chapter 5. Information Theoretic Local Segmentation

the noisy image is desirable. This may be achieved by iterating the algorithm. After processing each pixel using the global prior (common to each local window), a local posterior
is obtained. The local posteriors can be accumulated to form a global posterior when the
whole image has been processed. The global posterior is simply a probability distribution
over possible models, formed from observation of the image at all its localities. It captures
the particular essence of an image, briefly summarizing its local segmentation features. This
makes it ideal for use as a new global prior for re-processing the same image from scratch.
Iteration should continue until the priors converge.
The iteration process has an analogy with the human visual system [FH96, Ede99]. Imagine
visiting a new friends house and pulling a photo album from their bookshelf. Before you
open it, you have no idea of what the first photo will contain. This could be considered
a state of total ignorance, expressed by the uniform priors already used in Section 5.4.2.
After opening the album to the first picture, your visual system scans it to form an initial
impression. The information gained from this first scan may be considered a posterior, which
is then used as a revised prior for a second scan. This process continues until we have
recognized or made sense of the picture, as shown in Figure 5.36.

?
Figure 5.36: Pictorial representation of the iterative approach to image understanding.

A data driven prior incorporates image dependent information which is shared by all the
candidate local models. Because this information is common to all models, it does not need
to be incorporated into the message length. Its code length is the same for every model.

5.13.1 Learning a prior for the segment maps

The prior probability distribution for the segment maps, cJd;)feg/ , is important, because low
level image structure is the heart of local segmentation. Two priors have been considered

5.13 Learning from the data

195

so far. The uniform prior assumed all segment maps to be equally likely. In Section 5.6,
a non-uniform prior biased toward Q
was introduced. This prior improved the results
significantly, because it coincided more accurately with the expectation about which segment
maps would be common in montage.
Figure 5.37 shows four segment maps. Currently, the prior probability for the first homogeneous block is 0.5. In fact, that is expected to vary for different images. The other three
segment maps have the same prior probability -4Bx)f

O
/Z-NB-
F . However, the

255 heterogeneous segment maps are not expected to occur with equal frequency. Common
edge patterns would probably occur more often than, say, the random looking segment map
in Figure 5.37d. A more informative segment map prior should improve denoising results.

Figure 5.37: Potential segment maps: (a) popular "b ; (bc) common edge patterns;
(d) would rarely occur.

The segment map prior may be considered a multistate probability distribution, c0d)+eg/ , which
sums to unity over all possible segment maps. This is a global prior which is re-used for
each window processed. After processing each pixel )|,/ , a local posterior probability


distribution over models considered is obtained. Let it be denoted ZFy)|,, / , where
refers to a specific model from

, the set of candidate models.

Each local posterior may be accumulated to form a global posterior, denoted Post )feg/ , using
Equation 5.30. The global posterior probability for a specific segment map is the normalized
sum of local posterior probabilities for models using that particular segment map. For example, imagine an image with four pixels, and that only one candidate model per pixel used a
homogeneous segment map. If that model had local posterior probabilities 0.9, 0.7, 0.1, and
0.3 for each pixel respectively, then the global posterior for the homogeneous segment map
would be )+-4"uX-4uX-m
uX-N/~x17$- .

Post )feg/0

 t4

B)p|,~Z, /

where

e

(5.30)

196

Chapter 5. Information Theoretic Local Segmentation

Equation 5.31 describes how the global posterior may be used as a new global prior for
the image. This iterative process begins with the same noisy image each time, but uses an
improved, data driven prior. This is applicable to any components of the prior distribution.

c0d)+eg/ 2 

Post )feg/
2

PO Q

c0d)+eg/

J[.-BScOL
SOym

(5.31)

This iterative process is equivalent to using the E.M. algorithm discussed in Section 3.2.1.
The E.M. algorithm is guaranteed to converge, but not necessarily to a global optimum. The
best way to encourage a good optimum is to use reasonable starting conditions. The two
priors for e used so far are both reasonable, but the uniform prior would probably be a better
choice when applying the algorithm to images with unknown properties, as it has no bias.

Results
Figure 5.38 compares the MML denoising algorithm using 1 iteration and a non-uniform
segment map prior, to that using 6 iterations to learn a segment map prior. For montage, a
small improvement in RMSE is observed for &

 . This suggests that iteration has made

a useful difference to the segment map priors. This could allow the algorithm to identify
structure behind the noise better, because common structural segment maps are cheaper to
encode. The original prior set all QI
segment maps to be equally likely, making it difficult
to give preference to one over the other when the noise level was high. Another possibility is
that the algorithm has learned to choose QRE
more often, which at higher noise levels will
probably give more smoothing without the concern of damaging high contrast edges.
Figure 5.39 shows the 15 most probable segment maps that the new MML algorithm learned
from montage. Note the distinct lack of random-looking patterns. The top row coincides
with the many boundaries between squares in the two artificial quadrants in montage.

5.13.2 Learning the prior for DNH models

The prior probability c0d)+hji'k2/ is used to encode the prefix which distinguishes between
the null model and standard models. In Section 5.9, the quite reasonable prior, cJd)fhjiGk/0

5.13 Learning from the data

14

197

MML learning Pr(c)

1 iteration
6 iterations

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.38: Learning \]R^ : RMSE comparison for montage, with estimated.

Figure 5.39: The 15 most popular (canonical) segment maps for montage when R .

x)
[u& / , was used. However, it should be possible to use the data driven prior method to

learn automatically a probability suited to the noisy image being processed. This could be
done in the same way as the segment maps were learned. Instead of accumulating the local

198

Chapter 5. Information Theoretic Local Segmentation

posterior probabilities from models using a particular segment map, those posteriors for the
null models are accumulated. This global posterior probability is exactly equal to the best
value of c0d;)fhjiGk/ to use for the next iteration.

Results
Figure 5.40 compares the RMSE performance of the non-iterated, one-pass MML algorithm
to one which attempts to learn only c0d)+hji'k/ over 6 iterations. The prior for the segment
map was not learned. The results indicate that learning c0d)fh i'k/ makes little difference to
the algorithms performance. The same result occurred when an initial uniform prior, namely
c0d)+hji'k/0- , was used. This suggests that the existing static prior is a good one.

14

MML learning Pr(DNH)

1 iteration
6 iterations

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.40: Learning \]R_@`! : RMSE comparison for montage, with estimated.

Figure 5.41 shows proportion of times that DNH was chosen as the best model by the two
algorithms. Iterated MML learns to use DNH less often than its static counterpart, but this
makes little difference to the RMSE results. A possible explanation is that Figure 5.41 only

5.14 Incorporating more models

199

plots how often the null model was the best model. Because there are only three candidate
models, the best model could have posterior probability as low as 0.34 or so. After posterior blending, the influence of the best model may actually be quite low compared to the
sum influence of other candidates.

Proportion of times DNH was invoked

0.12

MML learning Pr(DNH)

1 iteration
6 iterations

0.1

0.08

0.06

0.04

0.02

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.41: Comparison of DNH invocation for montage, with estimated.

5.14 Incorporating more models

Learning a non-uniform prior for the segment maps allows some spatial information to be
incorporated into the MML denoising algorithm. This typically manifests itself at higher
noise levels. In most circumstances, when & is high, FUELS has little choice but to choose
Q<

. The MML algorithm may choose the alternative Q<`

thresholded model if its

message length is short enough. However, it is still hampered from having to choose from a
pool of only 3 candidate models: QL
, one thresholded QR
, and DNH.

200

Chapter 5. Information Theoretic Local Segmentation

Under the MML framework it is a simple matter to incorporate more candidate models. For
example, all valid segmentations (Section 5.3.3) could be tested, or a few extra QY
segmentations produced using different thresholds could be added. This thesis will take the
simplest route and consider all 256 canonical   binary segment maps. The examination
of all possible segment maps increases the model pool from 3 to 257 candidates, including DNH. The posterior distribution should be more varied, and the data driven segment
map prior should more accurately reflect the local image structures. However, this is at the
expense of nearly a one hundred fold increase in computation.
Figure 5.42 compares FUELS with two versions of the MML algorithm. Let MML-2 be
the familiar version which uses 2 candidate local segmentations plus DNH, and let MML256 be the new version which includes all 256 models with unique segment maps. Both
algorithms are run for 6 iterations to learn cJd;)fhjiGk/ and c0d)+eg/ . When &E , there is no
noticeable difference between MML-2 and MML-256. For images with little noise, thresholding adequately segments the window. As the noise increases, the RMSE performance of
MML-256 improves, until &'#B , where it worsens again.
Figure 5.43 plots the proportion of times that the candidate with the shortest message length
was Q':
. It is obvious from the graph that considering more 2-segment models per local
region significantly increases the chance of choosing a specific 2-segment model as the best
model. This supports the argument that thresholding is insufficient when the noise level is
high. Although not shown, the increase in Q
is mostly at the expense of Q

. DNH

usage of the MML algorithms remained relatively unchanged.

5.15 Improving the running time

The MML-256 algorithm has a superior ability to discern the local structure of a noisy image.
This power comes at the cost of higher computational requirements. On a serial computer,
the running time per pixel is proportional to the number of candidate models considered.
The parameter estimation and message length calculations for each model are independent
of each other a parallel implementation could compute them simultaneously. The comparison of the resulting message lengths could be done serially or using a recursive binary
algorithm. This is probably insignificant compared to the model calculation anyway.

5.15 Improving the running time

14

201

FUELS
MML-2
MML-256

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.42: More models: RMSE comparison for denoising montage, with estimated.

If restricted to a serial computer, there is still a simple modification that can reduce the
number of models which need to be considered for each window. Let & { be the assumed
 Fd@
? the sample variance of the pixels in the window. If
image noise variance, and l { #
l

& , the window is likely to be homogeneous, as the local variance is lower than the global

noise variance. If it happens not to be homogeneous, a thresholded model would probably

be sufficient. Thus if l & , MML-256 can change its mode of operation to that of MML-2.
Only 3, rather than 257, candidate models need to be considered. The unconsidered models
are assigned zero posterior probability. Relative to the remaining 3, the other 254 should
have long message lengths anyway. Let this variant of MML-256 be called MML-256/2.
Figure 5.44 compares the RMSE performance of the 3 MML variants, each run for 6 iterations. For &

V
 , MML-256 and MML-256/2 perform equivalently. As the noise in

montage increases, the faster MML-256/2 implementation actually outperforms MML256. Similar behaviour was observed for other images. Although not shown, the DNH usage
of MML-256 and MML-256/2 is nearly identical.

202

Chapter 5. Information Theoretic Local Segmentation

Proportion of times a k=2 model was most probable

0.4

FUELS
MML-2
MML-256

0.35

0.3

0.25

0.2

0.15

0.1

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.43: Proportion of times a "P model was most probable in montage.

Figure 5.45 show the proportion of two-segment regions diagnosed by the algorithms. MML256/2 is less likely than MML-256 to choose higher order models, especially as & increases.
Switching to MML-2 mode forces MML-256/2 to choose QR
more often. When the noise
is very high, averaging all pixels in the window pays off better in terms of RMSE, even
if it removes slightly more image structure. This does suggest, however, that the posterior
probabilities for the various two-segment models are overstated. This could be true as the
message length calculations do not strictly account for the quantization of segment means,
giving slightly shorter messages than warranted by the data.
The reason for introducing MML-256/2 was to decrease the running time. Figure 5.46 compares the average processing time per pixel used by each algorithm 2 . FUELS is very fast
relative to the MML variants, and barely registers on the graph. MML-2 is about 60 times
slower than FUELS. Somewhat surprising is that MML-256 is only about 43 times slower
than MML-2, despite having to evaluate 86 times as many models. The positive result is
2

The experiments were run on a 1.2GHz AMD Athlon with 512MB RAM running Linux kernel 2.4.16.

5.16 Results

14

203

MML-2
MML-256
MML-256/2

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.44: RMSE comparison for denoising montage, with estimated.

that MML-256/2 requires, on average, 60% of the computation that MML-256 does, and
achieves better RMSE results at higher noise levels.

5.16 Results
This chapter has so far shown that an MML-based local segmentation criterion improves the
denoising performance of the FUELS algorithm. This is a direct benefit of improved modeling of images at a local scale. Two useful MML variants have been identified. MML-2
considers 3 models in total, and is a logical extension to FUELS. MML-256/2 increases the
pool of candidate models considered to 257, in an effort to push the structure recognition
capabilities a little further. Together with FUELS, they represent the common trade-off between denoising ability and processing speed. In this section MML-2, MML-2/256, FUELS
and SUSAN will be compared using three test images.

204

Chapter 5. Information Theoretic Local Segmentation

Proportion of windows for which k=2 was chosen

0.4

MML-256
MML-256/2
MML-2

0.35

0.3

0.25

0.2

0.15

0.1

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.45: Proportion of times that "P was deemed best when denoising montage.

5.16.1 The lenna image

Figure 5.47 provides RMSE results for denoising the original lenna image 3 . The MML
algorithms used 6 iterations each to learn the segment map and DNH prior probabilities.
Both FUELS and MML used overlapping averaging. SUSAN9 is the   variant of the
SUSAN algorithm, chosen over SUSAN37 based on the results in Chapter 4. All of the
algorithms utilized the estimated noise variance as described in Section 4.14. The RMSE
was calculated using the original lenna as ground truth, despite it already containing noise.
SUSAN9 appears to perform worst for lenna in terms of RMSE. When &

 , the two

MML variants are beaten by FUELS, but as the noise increases, they outperform FUELS. The
poor performance of MML relative to FUELS at low noise levels is an interesting problem.
That fact is that the ground truth lenna is already noisy. The MML algorithms could
actually be producing better noiseless estimates of the true, but unknown, lenna image.
3

Available from ftp://nic.funet.fi/pub/graphics/misc/test-images/

5.16 Results

205

14

Running time per pixel (msec)

12
10
8
MML-256
MML-256/2
MML-2
FUELS

6
4
2
0
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.46: Comparison of algorithm running times for denoising montage.

Interestingly, MML-2 beats MML-256/2 when & $

F . Although not shown, the same result
occurs using the mean absolute error (MAE) metric. This unexpected behaviour was not
observed with the montage test image in Section 5.15.
Figure 5.48 compares the enhanced difference images, relative to the original lenna, for
the four algorithms on a F  F sub-image taken from the centre of lenna. FUELS and
MML-256/2 have similar difference images, with some larger errors scattered under the
eye. MML-2s output does not exhibit these same errors. This is because MML-2 inferred
Q 
at those points, whereas the others chose DNH or QE
(not shown). The SUSAN9

difference image does have less apparent structure in it, but its homogeneous regions are
more blotchy. This means it had difficulty assimilating large numbers of pixels, resulting in
poor smoothing and a higher RMSE result.
Figure 5.49 shows the worst case absolute error (WCAE) profile of the four denoising algorithms. As expected, SUSAN9 suffers at low noise levels, as it has no DNH-like capabilities.
The most interesting feature is that the WCAE of MML-256/2 nearly doubles after &
F ,

206

Chapter 5. Information Theoretic Local Segmentation

12

SUSAN9
FUELS
MML-256/2
MML-2

10

RMSE

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.47: RMSE comparison for denoising lenna, with estimated.

Figure 5.48: Enhanced difference images for lenna when X : (a) ground truth; (b) FUELS; (c) MML-256/2; (d) MML-2; (e) SUSAN9.

whereas MML-2 increases gracefully. This also helps to explain the earlier RMSE anomaly,
because large errors contribute significantly to the RMSE value. This suggests that some
two-segment models were made inappropriately probable by MML-256/2 when the noise
level was high. If an images properties vary spatially, a single global segment prior could
produce this type of behaviour.

5.16 Results

160

207

FUELS
SUSAN9
MML-2
MML-256/2

140
120

WCAE

100
80
60
40
20
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.49: WCAE comparison for denoising lenna.

5.16.2 The barb2 image

Figure 5.50 plots the RMSE for denoising the larger and more textured barb2 image 4 ,
originally introduced in Section 4.14.3. As usual, FUELS beats MML-2 and MML-256/2
at lower noise levels, but as & increases, the disparity between the three is more apparent.
MML-256/2 proves that a higher level of image modeling can improve RMSE results to a
significant level. When &3
 , it beats FUELS and SUSAN by up to 1.8 RMSE units.
The enhanced difference images when &$
B are given in Figure 5.51. They are from a


sub-image from the centre of barb2. These difference images have similar proper-

ties to those in Figure 5.48 for lenna. The errors in MML-2s difference image are lower
in magnitude, particularly the small circular cluster on the left hand side. This cluster has
occurred because the noisy image (not shown) had some very noisy pixels clumped together.
The WCAE results for barb2 are given in Figure 5.52. The deficiency of SUSAN9 at low
4

Available from http://www.csse.monash.edu.au/ torsten/phd/images/

208

Chapter 5. Information Theoretic Local Segmentation

16

SUSAN9
FUELS
MML-2
MML-256/2

14
12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.50: RMSE comparison for denoising barb2, with estimated.

Figure 5.51: Enhanced difference images for barb2 when X : (a) ground truth; (b) FUELS; (c) MML-256/2; (d) MML-2; (e) SUSAN9.

noise levels is again obvious, but at higher noise levels it does better than the other techniques. SUSAN9s behaviour, partially dependent on the threshold, seems to be ambitious
for low & , but successful for high & . MML-256/2 has a lower WCAE for most high values of
& , which is much better than its behaviour observed for lenna. This is perhaps due to the

distribution of segment maps being more uniform across the image, in particular the large
number of crisp edges on the wicker chair and clothing.

5.16 Results

209

140
120

WCAE

100
80
60
40
SUSAN9
FUELS
MML-2
MML-256/2

20
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.52: WCAE comparison for denoising barb2.

5.16.3 The camera image

The camera image5 is shown in Figure 5.53. It has dimensions
BF
B1 , which is smaller
than most test images used so far. It consists of a large number of homogeneous regions.
Although the mans coat appears uniformly dark, it hides a lot of low contrast structure.
Figure 5.54 compares the RMSE denoising performance on camera. Unlike for lenna
and barb2, FUELS and the MML variants have very similar results, while SUSAN9 did
relatively worse. MML-2 and MML-256/2 are indistinguishable. The large number of homogeneous regions makes it difficult for MML-256/2 to stand out, as it gains only from
considering more two-segment models. The fact that camera contains fewer pixels also
means that the MML algorithms have less data with which to re-estimate the priors.
The difference images for a homogeneous, possibly mildly planar, 

sub-image of
camera are given in Figure 5.55. FUELS limited model selection criterion has obviously
5

Available from ftp://nic.funet.fi/pub/graphics/misc/test-images/

210

Chapter 5. Information Theoretic Local Segmentation

Figure 5.53: The } 8 bit camera image and its histogram.

14

SUSAN9
FUELS
MML
MML-256/2

12

RMSE

10
8
6
4
2
0

10

15

20

25

30

Added Gaussian noise standard deviation ()

Figure 5.54: RMSE comparison for camera image, with estimated.

failed on a few pixels, but this is understandable given that &o
 . MML-256/2 also had
a little trouble with the same points, suggesting that it may be incorrectly favouring two-

5.17 Conclusions

211

segment models. Recall from Section 5.7 that the approximate expected message length is
calculated for each candidate. This approximation may be mildly biased toward 2-segment
models. MML-2 had no difficulty ignoring the correctly diagnosed glitch in the original
image. It is less likely to find an (un)suitable two-segment model, as it only considers one.

Figure 5.55: Enhanced difference images for camera when R : (a) ground truth; (b) FUELS; (c) MML-256/2; (d) MML-2; (e) SUSAN9.

5.17 Conclusions
In this chapter, the MML methodology was applied to local segmentation. The simple mean
separation criterion used by FUELS was replaced by a message length comparison between
candidate models. Each message was a concise, decodable encoding of the noisy pixels.
The message with the shortest length was deemed to contain the most appropriate model
for the local region. By using MML, the arbitrary constant, H , that FUELS required, was
eliminated. This introduced the possibility of diagnosing the presence of two segments with
close, but distinct intensities. FUELS was incapable of this in very noisy images.
MML made it straightforward to incorporate alternative models into the pool of candidates
being considered. All that was required was that the models messages be decodable. The
do no harm (DNH) concept was first introduced by FUELS to reduce its worst case denoising performance. It was a simple matter to incorporate a DNH model into the MML
framework. The DNH candidate simply encoded the pixel values as they were, without respect to any segment map or means. This information theoretic approach to DNH was found
to improve RMSE results compared to FUELS simpler method.
MML chooses a single best model from the pool of candidates, the so-called MAP model.
The message length of a candidate model may be interpreted as a posterior probability for

212

Chapter 5. Information Theoretic Local Segmentation

that model. Posterior probabilities provide an objective measure of the confidence in each
model. It was shown that if one is only interested in the resulting local approximation,
and not the details of the model, then blending over all models weighted by their posterior
probability was a good thing to do. This alters results when other candidate models have
message lengths close to, but just short of, that of the MAP model. Posterior blending was
found to improve RMSE performance, especially for very noisy images.
The ease with which a DNH model was incorporated paved the way for the addition of
more segment based models. For a  window, there exists only 256 canonical binary
segment maps, including the homogeneous one. It was a simple matter to enumerate all these
models and add them to the pool. Along with DNH, this brought the total number of models
considered per pixel to 257. Doing this effectively incorporated spatial information into
the local segmentation, as segment membership was no longer restricted by pixel intensity.
The RMSE results improved, but at the expense of a large increase in running time. The
running time was halved by switching to only 3 candidate models if the local variance was
low enough to warrant a more limited search.
Using a large pool of candidate models in conjunction with posterior blending may lead to
a more accurate local approximation. Small improvements in the accuracy of the local approximation may not be noticeable because the final approximated pixel values are rounded
to their nearest integers. There may be applications where one is willing to pay for a more
accurate local approximation. For example, examination of small features where there is
little contrast may be important in medical imaging applications.
Although an image may use the full range of intensities, a sub-image may only use a small
range. If the pixel values in the sub-image were known to higher accuracy, an enhanced
version may be more accurate than if the original integer pixels were used. There was not
enough time left to pursue this line of investigation. In future work, a photo-realistic ground
truth image could be generated synthetically. The pixel depth of this image could be far
greater than 8 bits. Varying amounts of noise could be added to it, and the pixel values
of the resulting image rounded to 8 bit precision. A denoising algorithm which produces
floating point pixel estimates, such as FUELS or MML-256/2, could then be compared to
the more accurate ground truth image. This would determine whether more accurate local
approximations contain any further meaningful information.

5.18 Related work

213

When segmenting  regions, only a small amount of pixel data is available, and the model
component of a two part MML message becomes significant. This places demands on efficiently encoding the model, demands which may be less important for larger windows. The
length of the model part is directly related to the prior probability distributions used. The
prior for the segment maps could be directly controlled. It was shown that changing it to a
more meaningful non-uniform distribution improved results. The logical extension to this
was to learn the prior from the noisy image itself. The idea of data-driven priors was applied
to the MML denoising algorithm, something FUELS was incapable of doing. It was found
that the priors converged in about 6 iterations. The RMSE results were further improved
using this technique.
The result of this chapter is two useful techniques for local segmentation. MML considers
3 candidate models, and is the logical extension to FUELS. MML-256/2 extends this even
further to consider 257 models in most cases, but only 3 models when a full search is deemed
unnecessary. The subjective analysis also showed the MML methods to produce better local
approximations, on average, than FUELS and SUSAN. However, this small improvement
only came with a lot of computational and developmental effort, suggesting that FUELS and
SUSAN are already very good. The MML-256/2 algorithm could potentially be improved
by deriving a more accurate approximation for the expected message length. In general, the
use of MML-256/2 is only warranted when a proper analysis is vital, the noise level is very
high, or the best possible local approximation is required.

5.18 Related work

An early form of the work in this chapter appeared in Proceedings of the International Picture

Coding Symposium 1997 [ST97]. In that paper, 11 candidate segmentations of the 
window were considered: 3 homogeneous models using the mean, median and midpoint
as the reconstructed value, and 8 binary models using the 8 possible thresholds to generate
segment means. These models were constructed differently to those described in Section 5.4.
The first event stated whether QL
or QL#
, with both events assumed equally likely. If QR

the representative value for the block was assumed uniform on -,._G'
@ . If QR
, the next

214

Chapter 5. Information Theoretic Local Segmentation

event was how many pixels,

, were in the cluster with the lower mean, assumed uniform

on
B,. . This was followed by a segment map describing the spatial arrangement of the
cluster labels. Because

was already known, there were only a c

b possible bitmaps, which

were assumed equally likely. The first cluster mean, # , was assumed uniform on -,_bb
@ .
Because the second mean was known to be higher than the first, a uniform distribution on
# uD
B,._#I
@ could be used. Finally, the residuals between the local approximation and

the noisy data were encoded. These were treated differently than in the MML method of this
chapter. Two discrete residual distributions were kept, one for use by QRO
models, and one
by Qb]
models. No numerical integration was required as all segment parameters were
quantized to their nearest integers.
Only the centre pixel of the most probable resulting local approximation was used for denoising. The ideas of overlapping estimates and do no harm had not been developed yet. The
algorithm was also iterated such that the posterior probabilities were fed back to be priors,
just as MML does in Section 5.13. The idea of posterior blending was not yet developed, and
the global posteriors were only updated from the most probable model at each point, rather
than a weighted contribution from all models.

Chapter 6
Extensions and Further Applications of
Local Segmentation
6.1 Introduction
In Chapters 4 and 5 the principle of local segmentation was applied successfully to removing additive noise from greyscale images. The relative performance of different ideas and
parameter settings were evaluated and compared using the RMSE criterion and difference
images. Image denoising was naturally suited to demonstrating some advantages of examining images from a local segmentation perspective. However, one should not be led into
thinking that its usefulness ends there.
Local segmentation may be considered a core component of many other image processing
applications. A local segmentation analysis provides a way to split an image into its signal
and noise components. The signal may then be enhanced further without interference from
the noise. It provides a snapshot of the structural features at each point in an image, identifying which pixels belong together and where boundaries occur. Figure 6.1 espouses the
application of local segmentation as a first step in low level image processing, from which
many other tasks may be derived.
This chapter will illustrate how local segmentation could be applied to pixel classification,
edge detection, pixel interpolation and image compression. It will also examine alternative
215

216

Chapter 6. Extensions and Further Applications of Local Segmentation

Structure
Denoising
Edge detection
Compression
Interpolation

Raw image

Etc. . . . . . . . .

Noise
Local Segmentation Decomposition

Figure 6.1: Local segmentation decomposition as a core image processing component.

image models, noise models, and segmentation algorithms for use in local segmentation.
The extension to different data types will also be considered. Most of the ideas will only be
presented in brief, as a full study is beyond the scope of this thesis. However, some topics
do include preliminary implementations and results.

6.2 Different noise models

The FUELS and MML techniques described in this thesis have focused on the removal of additive zero-mean Gaussian noise, independently and identically distributed for each pixel. In
this section it will be shown how different noise models, such as impulse and multiplicative
noise, may be incorporated into the local segmentation framework.

6.2.1 Impulse noise

Recall Section 2.5, which described the nature of ergodic impulse noise. Pixels in an impulse
affected image have probability l of being corrupted. This corruption takes the form of a

6.2 Different noise models

217

pixel being completely replaced by a random pixel value. This differs from additive noise
where some information about the original pixel value is retained. The random replacement
pixel is usually assumed to be drawn uniformly from the full intensity range -,_
@ . A
pixel which has been affected by impulse noise is known as a corrupt pixel.
There is a small chance that a corrupt pixel will have an intensity very similar, or equal, to its
original value. However, most of the time its value will stand out from its neighbours. The
median filter in Section 3.4.4 is sometimes well suited for filtering impulse noise. In homogeneous regions, the median filter can cope with up to 50% of the pixels being corrupt and
still produce a reasonable denoised value. In heterogeneous regions it is less well behaved.
In terms of local segmentation, a corrupt pixel may be considered a very small segment having an intensity sufficiently different from its neighbours. In the simplest case, this segment
would consist of one (corrupt) pixel and be surrounded by other larger segments. It is possible for corrupt pixels to be adjacent, resulting in small, spatially connected groups of corrupt
pixels. If an adjacent pair of corrupt pixels is similar enough in intensity, it could even be
(mis)interpreted as a valid two pixel segment.
Let the minimum acceptable segment size (MASS) be the smallest sized local segment that
should be interpreted as structure rather than noise. Recall the   median filter from Section 3.4.4, which was unable to filter the centre pixel correctly when the intersection of its
segment and the windows contained fewer than 5 pixels. It could be considered as having
a MASS of 5. The centre weighted median from Section 3.4.4 is able to adjust its effective
MASS by varying the centre weight, ! . When ! I
, the MASS reduced to 4.
Ideally, the MASS should depend on l , the fraction of corrupted pixels in the image. The
MASS could be a global parameter to the local segmentation process, in the same way that
&

is for additive noise. It could be supplied by the user, or somehow estimated from the

image itself. Equation 6.1 gives an expression for the probability that a given pixel, , will
be surrounded (in an 8-connected sense) by exactly corrupt pixels, where l is the per pixel
probability of impulse noise.

c0d) has exactly corrupt neighbours /J

)
"lF/ r

(6.1)

218

Chapter 6. Extensions and Further Applications of Local Segmentation

The probability of a given pixel having no corrupt neighbours is )
rlF/ r . The probability
of that pixel being itself corrupt is l . Therefore the probability of a randomly selected pixel
being isolated and corrupt is l-)
J2lF/ r . When l
; , this is only 4.3% of pixels, meaning
5.7% of corrupt pixels should have one or more corrupt neighbours. This corresponds to
the majority of impulse noise occurring in clumps rather than isolation, with pairs being
most common. Fortunately, these pairs are not necessarily similar in intensity and could be
diagnosed as two separate single pixel segments.
The multi-class version of FUELS from Section 4.12 could be adapted to additionally remove
impulse noise. For each pixel, the optimal local segmentation could be determined, using the
estimated additive noise level, & , as before. If the centre pixels segment consists of fewer
than MASS pixels, it could be considered an impulse segment. The pixels from an impulse
segment need to be assigned new values. The mean of the largest segment not containing the
centre pixel could be used. For a  window, that segment will always contain some pixels
neighbouring the centre pixel. However, for larger windows this would not necessarily be the
case. A more advanced version of this idea could take that into consideration when choosing
replacement pixel values.

Results
Figure 6.2 illustrates the application of this idea to the middle section of montage contaminated by 5% impulse noise. The multi-class FUELS algorithm was modified to handle
both additive and impulse noise as described earlier. Both DNH and overlapping averaging
were disabled because they would only complicate the implementation. Results for MASS=2
and MASS=3 are included. The outputs of the   median (effective MASS=5) and centre
3-weighted median filters (MASS=4) are also present for comparison. The image margins
should be ignored, as the method used for handling missing pixels is not entirely appropriate when impulse noise is present in the image.
When MASS=2, single pixel segments will be obliterated, while larger segments will remain
intact. In Figure 6.2 it can be seen that the remaining noise mainly consists of pairs of pixels
which are similar in intensity. But this means two-pixel segments, such as the dot on the
i, remain intact. When MASS=3, the modified FUELS manages to remove most of the

6.2 Different noise models

219

Figure 6.2: (a) clean image; (b) with d# B impulse noise; (c) multi-class FUELS,
MASS=2; (d) MASS=3; (e) median; (f) weighted median.

220

Chapter 6. Extensions and Further Applications of Local Segmentation

noise that remained when MASS=2. But this comes at the expense of structure loss on the
window shutters, the dots on j and ?, and the hole and tip of the a. At first glance the
median filtered image looks very clean, but much of the fine detail and the letters have been
seriously damaged. The weighted median appears to have done a better job overall, but also
suffers from corruption of the letters. FUELS with MASS=2 did best of all on the letters.
Figure 6.3 performs the same experiment, but with 10% (lR

-
) corruption. In this envi-

ronment, single pixel impulse segments are less likely to occur than when l2]-N . The
MASS=2 filter still does reasonably well under these conditions the letters are all legible
but much noise still remains. When MASS=3, less noise remains, but some of the letters are unrecognizable. The two median filters performed similarly, but the centre weighted
median did manage to retain more of the letter structure.
Table 6.1 contains RMSE results for the filtered outputs demonstrated in Figures 6.2 and
6.3. When l

-4B , the weighted median and MASS=2 do best. This correlates with

the subjective evaluation. When lj

 
, the weighted median again does best in terms of

RMSE. MASS=2 does worst, as it can only remove single pixel noise segments, which are in
the minority for 10% corruption. MASS=3 and median perform similarly. Once again, this
does not differ substantially from the conclusions drawn from the earlier visual examination.
Method
Median
W. Median
MASS=2
MASS=3

l#-N

l#-m

16.4
11.1
12.7
15.7

17.3
13.5
19.2
17.9

Table 6.1: RMSE results for filtering impulse noise from montage.

Conclusions
One method for adapting the multi-class FUELS algorithm to remove impulse noise was
presented. It functioned by treating all local segments containing fewer than a specified
number of pixels, the MASS, as noise. Corrupt pixels were replaced by the denoised mean
of their largest neighbouring segment. Additive noise was simultaneously removed for those
segments not diagnosed as impulse noise.

6.2 Different noise models

221

Figure 6.3: (a) clean image; (b) with d PB impulse noise; (c) multi-class FUELS, MASS=2;
(d) MASS=3; (e) median; (f) weighted median.

222

Chapter 6. Extensions and Further Applications of Local Segmentation

The MASS=2 implementation works well for 5% corruption. It tended to preserve more
structure than the median varieties, but could not cope with the occasional pairs of corrupt
pixels with similar intensities. As the impulse noise level was increased, higher values for
the MASS are required to cope. In those situations it tends to leave more noise behind,
but preserve more structure in the areas where it took it away. This should make it more
amenable to multiple passes through the data.
The best MASS to use depends on the amount of impulse noise present. It is preferred that
the algorithm determine the most suitable MASS for a particular image automatically. One
possibility is to apply filters with known MASS and examine the number of pixels which
suffer gross changes. From that number it may be possible to estimate the percentage noise
corruption, l , and hence a suitable MASS.

6.2.2 Multiplicative noise

Multiplicative noise was first introduced in Section 2.5. It may be considered additive noise,
with the noise term being proportional to the intensity of the original pixel. Equation 6.2
illustrates the situation, where w)p|,-/ is a random value from some unspecified distribution.

? )p|,~/0I)p|,-/uXw)p|,-/~)p|,-/

(6.2)

It is common to assume that the multiplicative noise function 0)p|,~/ is normally distributed
as 8:)+-,& { / , with & constant over the image. Multiplying a variable 698:)+-,& { / by a scalar,

, results in a variable 68

)f ,C { & { / . In Equation 6.2, would be the intensity of the original

pixel. Thus multiplicative noise could be considered additive noise where the variance is
dependent on the original pixel value. There are various ways in which this behaviour could
be exploited by the local segmentation algorithms developed in this thesis.
Consider a small block of pixels corrupted by the multiplicative noise model just described.
Under the assumption of homogeneity, the mean, # , of the block would be a reasonable
estimate of the original intensities. FUELS would assume the block variance to be & { . Under
a multiplicative noise model it would be approximately equal to )B#&|/ { . Similar arguments
would apply to the clusters determined under a heterogeneous model. A modified FUELS

6.2 Different noise models

223

model selection criterion would use the per cluster variance to determine suitability, rather
than a common noise variance. A simpler alternative is to simply use the average brightness
of the block to determine a local noise variance to be used by all derived clusters.
The treatment just described is not ideal. The mean is not necessarily the best estimate to
use, as bright pixels are less reliable than dark ones. The binary clustering algorithm used
by FUELS assumes the clusters have a common variance. To properly handle multiplicative
noise a more complex clustering technique should be used. However, it has been shown that
with some modifications, local segmentation could be used for more than just additive noise.

6.2.3 Spatially varying noise

It is possible for the noise to vary from location to location within an image. Consider an
additive zero-mean noise component in which the variance depends on the spatial location
within the image. This is shown in Equation 6.3, where & is a function of the pixel position.

? )p|,-/wI)p|,-/uA8

)f ,&J)p|,~/ { /

(6.3)

Under this model the variance could vary erratically from point to point. It would be difficult
to determine an appropriate values of & for each pixel. A more likely situation is for the
variance to vary smoothly across the image, or to be constant within global segments but
discontinuous between segments. In these cases, the variance will be mostly consistent on
a local scale. The technique used by FUELS and MML to estimate the global noise variance could be applied to small sub-images centered on each pixel. Under the assumption
of smoothness, the variance could be measured in a regular fashion for a proportion of all
pixels. The variance for the remaining pixels could be interpolated from the estimated ones.
The Q -means segmentation algorithm used by FUELS could not be applied unmodified to an
image containing known, spatially varying noise. Although it does not take spatial information into account, and the value of a noisy pixel is still the best estimate of its true value, the
equal averaging of pixels to estimate the segment mean is inappropriate. Instead a weighted
mean, based on the noise variance of each pixel value, should be used. The model selection
criterion would also need to be modified to handle the different variance of each pixel, and
therefore of each segment.

224

Chapter 6. Extensions and Further Applications of Local Segmentation

6.3 Different structural models

Image models have two components: the structural model for describing the underlying
noiseless signal, and the noise model for describing the way in which the signal was corrupted. It has already been shown how different noise models may be incorporated into local
segmentation. In this section, extensions to the structural component will be briefly explored.

6.3.1 Multispectral pixels

The work in this thesis has been derived under the assumption of greyscale pixel data, but
local segmentation is not restricted to this. A pixel from an alternative colour space may
be interpreted as a vector, rather than scalar. Consider the RGB triplet sometimes used
for colour pixels. If the noise is considered additive zero-mean Gaussian with the same
variance in each band, FUELS and MML may be used with little modification. The Q means clustering algorithm extends naturally to multi-dimensional data, although it is more
sensitive to starting conditions. The mean separation criterion would have to be modified
to be a distance in RGB space, rather than a simple intensity difference. Under the MML
framework, each segment mean and residual would be encoded as a vector. Each colour
component in the vector could be considered 3 independent scalars. These ideas also apply
to multispectral data, which may have any number of bands.

6.3.2 Planar regions

The structural model used so far has assumed piece-wise constant segments. Piece-wise constant segments are the lowest level in the facet model hierarchy, described in Section 4.2.1.
It was argued that, on a local scale, constant facets would be sufficient to model most image behaviour. The success of the FUELS and MML algorithms have shown that a simple
model can be taken a long way, but not all image data is well modeled by constant facets. In
Section 4.2, it was stated that range images contain segments which vary in a linear, or planar, fashion. Even photographic images like lenna and barb2 have regions which exhibit
some planar-like behaviour. It would be useful to incorporate planar segments into the local
segmentation framework.

6.3 Different structural models

225

A constant segment has only one parameter, # , to describe its average intensity. A planar
segment has three parameters. The intensity profile for a planar segment may be written
like Equation 6.4, where
#

and 8 are the horizontal and

w)p|,~/w# uP:8 ubR8

(6.4)

is the average brightness, and

vertical gradient terms respectively.

If the noise is assumed distributed as 8:)+-,& { / for each pixel, the optimal values for the three
parameters may be determined using the familiar least-sum-of-squares algorithm [Nie94].
A simpler solution for pixels from a regular square region, like the popular   window, is
given in Figure 6.4. The gradient terms are actually equivalent to the outputs of the horizontal
and vertical Prewitt edge detector masks [PM66].

p
E

#2

D
8




2 / 
u ) E '
/ u) D 2
/

D
m) ' / u#) 2 { / #
u ) G /

r
p
p
p

m)

Figure 6.4: Computing the parameters for a homogeneous } planar model.

It has been shown how to derive a planar model for a homogeneous region. Under the
MML framework, this model could be incorporated into the pool of candidates, alongside
the homogeneous constant model and two-segment constant models. The model part of its
message would be longer, as three parameters need to be encoded to describe the fitted plane.
It should be possible to determine the optimal quantization levels for
8

and 8 . The data

part is the residuals between the fitted plane and the noisy pixel values. MML naturally takes
into consideration the complexity of each model via its overall message length. The planar
model would only be chosen over the constant model as warranted by the data.
Fitting planes to each segment in a heterogeneous window is more complicated. Obviously,
when there is more than one segment present, there are fewer pixels in each segment. It is
difficult to estimate reliably the gradient terms for the fitted planes, due to the lack of data.
For an MML message, the small number of pixels would make it difficult to save enough bits
in the data part to make up for the extra bits used for encoding the parameters of the plane.

226

Chapter 6. Extensions and Further Applications of Local Segmentation

FUELS and MML allow segments to be unconnected, further complicating the modeling.
For a  window, fitting planes to two or more segments would not probably be justified.
A homogeneous planar model can already model ramp edges very well. This phenomenon
is what two segment regions were previously forced to approximate.

6.3.3 Higher dimensional data

The digital images processed in this thesis were two dimensional in nature, with each coordinate hosting a pixel. Today, medical imaging techniques are able to generate three dimensional (3D) volume images. Figure 6.5 shows a small vjM volume image, which has
160 elements in all. Volume images are sometimes treated as a series of 2D image slices.
Each slice may also be temporally shifted due to time spent in the acquisition process.

z
x
y

Figure 6.5: An Lr* volume image.

Each element in a volume image is called a voxel, which is shorthand for volume element.
In the simplest case, a voxel stores a scalar measurement, such as a bone density or radioactive dye intensity. Figure 6.6 shows the local neighbourhood of a voxel. Translating the 2D
concept of 4-connectivity to 3D results in a neighbourhood of 6 voxels.
Local segmentation is a philosophy, not a particular algorithm or implementation. Its principles apply equally well to 2D greyscale images as they do to 3D volume images. A local
neighbourhood for each voxel can be defined. This neighbourhood may then be segmented
using any suitable technique. If spatial information is not important, then FUELS or MML2 could be applied without any major changes. The principles of DNH and overlapping
averaging would still apply to volume data.

6.3 Different structural models

227

Figure 6.6: Each voxel (dotted) has 6 neighbours in its immediate neighbourhood.

The extension to 3D is only the beginning. Medical scanners are able to take temporal sequences of volume images. For example, to examine blood flow through the body over a
period of time. This could be treated as 4D information. The local neighbourhood would
consist of voxels around the current voxel in the current frame, along with voxels around
the same position in adjacent frames. A possible complication is that the time dimension is inherently different from the spatial ones, whereas FUELS and MML assume certain
isotropic behaviour. Planar models might be useful for handling this situation.
These ideas also apply to digital video, which is a time series of 2D images. Local segmentation would be able to detect a change in the number of local segments around each pixel. This
change would probably be associated with moving objects in the recorded scenes. Hence,
local segmentation could be used as a low level step in object tracking systems.

6.3.4 Larger window sizes

In Section 4.10, FUELS was tested on different windows sizes, the largest of which contained
37 pixels. It was found that smaller windows produced better lower RMSE when denoising.
This may not be true if a more sophisticated image model was used. FUELS and MML-2
extend naturally to larger mask sizes, as they only considered a fixed number of candidate
models based on thresholding.
MML-256 took the path of considering all possible segment maps. For a   window,
this came to 256 distinct segmentations. This exhaustive approach does not scale well to
larger windows. For each extra pixel included in the window, the number of binary segment

228

Chapter 6. Extensions and Further Applications of Local Segmentation

maps doubles. As the window size increases, the average number of segments present in the
underlying image is also expected to rise. This also increases the total number of candidate
segmentations to be potentially considered.
Obviously the search space needs to be restricted. The homogeneous and optimally thresholded models should always be included. A segmentation algorithm which exploits spatial
information could be used to generate extra candidates. The thresholded segmentation could
be used a starting point for the spatial algorithm. Most spatial segmentation algorithms require one or more parameters to control their operation. A few different values for these
parameters could be tried, and each resulting segmentation added to the pool.
Reducing the search space is only part of the problem. Each segment map needs to be costed
in bits. An explicit prior distribution over all possible segment maps is no longer suitable, due
to the large number of possibilities. Instead, a method for efficiently coding segment maps
is needed, which implicitly gives a prior to all segment maps. The field of lossless image
compression would be a good place to find a method for doing this. I have already done
some preliminary investigations into the use of a low order Markov model for compressing
binary segment maps, much like the JBIG algorithm [GGLR81, PMGL88]. The Markov
model parameters are dynamically learned using the data-driven prior approach. The details
of this work are outside the scope of this thesis.

6.4 Pixel classification

Pixel classification is the process of classifying each pixel in an image as a smooth, shaped,
or textured feature point [Cho99]. Smooth points correspond to homogeneous regions, such
as those positioned in the interior of global segments with a low natural level of variation.
Shaped regions, or edge regions, are those occurring near the boundaries and junctions between global segments. There is no universally accepted definition for texture. Caelli and
Reye [CR93] propose that textured regions and edge regions are the same features, but at
different scales. They suggest that texture be considered a high resolution area containing
many edges in close proximity.
Local segmentation necessarily classifies the region surrounding each pixel. The most important attribute determined by the local segmentation process is Q , the number of segments

6.4 Pixel classification

229

present. This Q may be used for pixel classification. Ignoring DNH for the moment, the
multi-class  FUELS algorithm estimates Q to be between 1 and 9. When Q
, the
window is smooth. When Q

, the window contains segment boundaries, so a shaped

classification is more appropriate. In summary, we have:

Smooth
Shaped

QLD

Q $

Figure 6.7 shows lenna and its corresponding classification map. Black pixels denote
smooth points and white pixels denote shaped points. The classification is based on values
of Q as determined by the multi-class FUELS algorithm with DNH disabled. The noise level
was estimated to be &'#
4BF .

Figure 6.7: (a) original lenna; (b) classification into smooth (black) and shaped (white)
feature points.

Of the pixels in Figure 6.7b, 59% are smooth and 41% are shaped. The smooth points do
indeed correlate with the large homogeneous regions in lenna, and the shaped points to
the edges and texture. Pixel classification of this type could be used in various ways. For
example, a lossy image compression algorithm could allocate more bits to the shaped areas
to preserve sharpness. A global segmentation algorithm could attempt to cluster only smooth
points first, and then apply region growing to annex shaped points accordingly.

230

Chapter 6. Extensions and Further Applications of Local Segmentation

In the example of Figure 6.7, shaped and textured feature points were not distinguished. Windows which are not well modeled by one or more segments could be considered textured
regions. The multi-class FUELS algorithm uses its DNH mode to identify such windows.
In Section 4.12.4 it was also shown that windows where Qo

, where

is the number

of pixels in the window, are equivalent to DNH. Thus local segmentation could be used to
classify pixels into three groups as follows:
Smooth

Q

Shaped

Textured

Qj

QI

or DNH

Figure 6.8 applies these classification rules to lenna. Only 5% of pixels are classified as
textured. About a third occur near edges, possibly ramps. These are difficult to model with
piece-wise constant segments. Another third seem to be isolated patches, perhaps planar regions for which a segment-based model was inappropriate. The last third do actually appear
in textured regions of the image, particularly the feathers in lennas boa. This result is less
encouraging than when only smooth and shaped points were considered.
Pixel classification could possibly be improved by examining the image at different scales.
Chou [Cho99] performs edge strength measurements on the original image and two downsampled versions thereof. A set of rules relating edge strength at each position over the
various scales is then used to make the final decision. For example, a low edge response at
all scales gives strong evidence for the point being smooth, and correspondingly, consistently
high edge responses suggest a shaped point. If the edge strength varies relatively over scales
then it could be considered a textured point. It would be possible to apply these types of
rules using the values of Q determined at different scales.

6.5 Edge detection

Edge detection describes the general problem of determining the location, magnitude and
orientation of (segment) boundaries in an image [HSSB98]. The precision and accuracy to
which these edge attributes may be determined is what discriminates between algorithms.

6.5 Edge detection

231

Figure 6.8: (a) original lenna; (b) smooth points (white); (c) shaped points (white); (d) textured points (white).

The shaped pixel classification in Section 6.4 may be considered a primitive edge detector.
It provides a rough location, but no magnitude or orientation. This section will explore some
more advanced methods for edge detection from a local segmentation perspective.

6.5.1 Edge strength measurement

Edge strength measurement algorithms are used to estimate the magnitude of an edge transition around each point in the image. They do not attempt to determine the exact position
or direction of an edge. Edge strength may be used as a simple edge detector. It should

232

Chapter 6. Extensions and Further Applications of Local Segmentation

give a low response in homogeneous regions, and a high response in the vicinity of segment
boundaries. The response usually increases with both the sharpness and contrast of the edge.
The output of an edge strength algorithm could be used by segmentation algorithms. For
example, those pixels furthest from high edge activity may be used as seeds for a region
growing algorithm. A map of edge strengths could be used as the elevation map for the
watershed segmentation algorithm [HH94]. Alternatively, the identified edge pixels may be
used to initiate a boundary tracking process.
The commonly used Sobel  linear convolution masks [Sob70, GW92] treat the image
as a functional surface, and attempt to estimate the gradient magnitude in two orthogonal
directions. Equation 6.5 shows the horizontal mask, , and the vertical mask, . The edge
strength measurement ,fefgV)J)p|,-/ at each point is the Pythagorean combination of the result
of convolving the two directional masks with each pixel w)p|,~/ .

0 -1

0 -2

0 -1

-1

-2 -1

,4ehgV)J)|,/

N)p|,-/ {wuNw)|,-/ {

(6.5)
Local segmentation may be applied to edge strength measurement. Consider FUELS thresholding from Section 4.14, which diagnoses the window as consisting of one or two segments.
For the two segment case, the numerical difference between the two cluster means could be
considered a measure of edge strength. Equation 6.6 gives an expression for computing
an edge strength image, i

e . Because only regions with a contrast difference greater than

H& are diagnosed as consisting of two segments, that amount is subtracted from the edge

strength. Homogeneous blocks receive an edge strength of zero.

e)p|,~/0R7,

%
#

)p|,~/N# )p|,-/ % PH&'

{

(6.6)

This measure of edge strength is only one possibility. An alternative would be to consider the
intensity difference between the centre pixel and the neighbouring pixel nearest in intensity,
but from the other segment.

6.5 Edge detection

233

Results
Figure 6.9 provides edge strength results for lenna. Equation 6.6 was used by the local segmentation based method. The Sobel and local segmentation outputs were linearly stretched
to the full intensity range, to improve visibility of the edge strengths. SUSANs edge detector
output is also included1 . It used the same value of as used for denoising in Chapter 4.

Figure 6.9: (a) original lenna, estimated LPF ; (b) local segmentation; (c) Sobel; (d) SUSAN using |kjBl[b .

The local segmentation edge strength measurement algorithm presented here is very simplistic. Its performance could be said to fall somewhere between Sobel and SUSAN. It appears
1

SUSAN source code: http://www.fmrib.ox.ac.uk/ steve/susan/

234

Chapter 6. Extensions and Further Applications of Local Segmentation

to be less sensitive to noise than SUSAN, but picks up more potential detail than Sobel. It
would be interesting to apply thinning and edge linking to the local segmentation output, but
that is outside the scope of this chapter. The results are useful in that they essentially come
for free from the local segmentation decomposition, and do not use any spatial information.
A more advanced implementation could consider the actual pattern of cluster assignments
for each window, and accordingly determine edge direction, or disregard incoherent patterns.

6.5.2 Probabilistic edge detection

The MML-256 local segmentation algorithm from Chapter 5 considers all 256 possible  
segment maps. This results in a posterior probability distribution over possible binary segment maps for the local   window. The segment model assumes that pixels belong wholly
to one segment, and that segment boundaries occur between pixels. Thus a segmentation implicitly defines the existence of segment boundaries.
Consider the   block of pixels of Figure 6.10. A pixel interface is the junction line between two adjacent pixels. There are 12 pixel interfaces in a  block, 6 of them horizontal
and 6 of them vertical. A edge, or boundary, is said to exist at a pixel interface if the segment
labels for the two pixels at the interface are different. At first glance it may appear that there

are
{ boundary patterns, but there are really much fewer, because each pattern must form
one or more logical segments.

Figure 6.10: The twelve inter-pixel boundaries for a window, shown in bold.

Figure 6.11 illustrates three potential segment maps. The homogeneous segment map asserts
that no boundaries are present in the window. The window with a vertical edge asserts the
existence of a boundary at 3 of the 12 possible interfaces. The diagonal line in the third
window is even more complex, insisting that there is a boundary present at 6 interfaces.

6.5 Edge detection

235

Figure 6.11: Examples of segment maps, with boundaries shown in bold.

The posterior information gleaned from the MML-256 local segmentation process may be
used for edge detection. For an <

image there are )]
/1)fI
/ horizontal pixel

boundaries, and the same number of vertical ones. Each processed pixel produces a probability distribution over possible segment maps. If the segment labels on both sides of an
interface are different, then a boundary is present along that interface. For a given segment
map, the probability of each interfaces boundary existing in the image can be assumed to
have the same probability the posterior probability for the segment map in question.
Thus a probability for the existence of an edge at each pixel interface can be determined.
Each interface occurs in 6 overlapping windows, so probabilities can be accumulated. The
horizontal and vertical probabilities can be stored in two images, called s and t , each of
resolution )pL
/)+'R
/ . Each image may be visualized alone, or they could be combined
into a single image, i . Three different methods for combining the horizontal and vertical
edge probability images are considered:

1. Averaging: iT
{

)fs uPtw/ .

2. Euclidean magnitude, or normalized { norm: i

m K on K .
{

3. The maximum, or norm: i*$R)fsJ,~t/ .

Results
Figure 6.12 gives results for part of lenna. Black represents probability zero, and white
probability one. The MML-256 algorithm was iterated 6 times to learn priors for the segment
maps. The DNH model was treated like a homogeneous window, in that it did not contribute

236

Chapter 6. Extensions and Further Applications of Local Segmentation

Figure 6.12: Example of probabilistic pixel boundary detection: (a) part of lenna; (b) p ;
(c) q ; (d) Gpsrtq ;u; ; (e) v Gp { rtq { ;u; ; (f) wyxz.Gp;q .

any posterior probability to any pixel interfaces. An alternative would have been to treat all
pixel interfaces as being active.
This technique seems to work extremely well. The vertical image clearly picks up the hanging feather at the top of the image. This feather does not appear at all in the horizontal
image, as desired. The diagonal brim of the hat is clearly detected in both s and t , because

6.6 Image enlargement

237

diagonal edges are a blend of horizontal and vertical changes. For visualization purposes,
the  )t,.s/ function appears to give an output image with more contrast. The R operation clearly states whether an edge is present or not, regardless of its orientation. The other
two methods for combining t and s respond more to diagonal edges than to vertical and
horizontal ones.
Figure 6.13 shows the combined probabilistic edge strength image for the complete lenna
image, using the R)+s,~t/ method. It was produced under the same conditions as Figure 6.12f. The results are quite remarkable. All the important boundaries are clearly visible.
Most interesting is the amount of detail with respect to individual feathers on the boa hanging from the hat. It has also managed to detect a large proportion of the fine diagonal lines
present above the band on the hat.

Conclusions
The results of this section clearly show the power of the MML-256 local segmentation framework. The horizontal and vertical edge strength decomposition derived naturally from the
local segmentation models used. It would be possible to transform this orthogonal decomposition into another coordinate space. For example, polar coordinates would describe the edge
magnitude and the edge orientation at each position. The { norm plotted in Figure 6.12e
is effectively the magnitude. The angle of orientation,  , may additionally be derived using

- 4
the trigonometric expression A#Fd[PO|{ - 4 .

6.6 Image enlargement

Image enlargement increases the resolution of an image. It is also called image expansion,
zooming, and interpolation. Consider the simplest problem of doubling the size of an image
from <

to
<
1 . The doubled image has four times as many pixels as the original.

Exactly 75% of the pixels need new values determined for them. This situation is shown in
Figure 6.14. The problem is to predict suitable values for the unknown pixels, such that the
features of the original image remain intact.

238

Chapter 6. Extensions and Further Applications of Local Segmentation

Figure 6.13: Probabilistic edge strength for lenna using w}xzV~qp .

The simplest technique for image enlargement is pixel replication. It makes the assumption
that the known pixel values were obtained from a

down-sampling of an originally larger
image. Thus, the best replacement pixels have values equal to the known pixel. An example is given in Figure 6.15. The main disadvantage with pixel replication is that, for natural
images, the results are very blocky. The next simplest idea is to interpolate the pixel intensities linearly. This often looks better than pixel replication, but has a tendency to produce
blurred edges, much like the box filter for denoising.

6.6 Image enlargement

239

?
?

?
?

?
?

?
?

?
?

?
?

?
?

?
?

?
?

Figure 6.14: Doubling an images size means predicting all the ? pixels.

?
?

?
?

?
?

?
?

Figure 6.15: Pixel replication is the simplest expansion method.

Structure directed enlargement algorithms attempt to maintain edge coherence when estimating values for the unknown pixels. It is possible to use the MML-256 local segmentation
method to aid image enlargement. Consider the Z window of the enlarged image in Figure 6.16a. It has 9 known pixels, which are numbered. In the original image, assume the best
model used the binary segment map denoted by the white and grey pixels. Segment 1 had 4
pixels in it, and segment 2 had 5 pixels in it.
Now consider the   sub-window in Figure 6.16b. It has 4 known pixels, and 5 unknown
pixels. The segment memberships that the 4 known pixels had in the original segment map
can be assumed consistent. All that remains is to determine which segment each of the
unknown pixels belong to. When that is achieved, each unknown pixel can be assigned a
value equal to its segments mean. Because only 5 pixels dont belong to a segment, there
are only
q[#B
possible segment maps that could apply.
The MML-256 algorithm can be used to calculate the message lengths for these 32 possible
models. The same prior used for the original image could be re-used, ensuring that the

240

(a)

Chapter 6. Extensions and Further Applications of Local Segmentation

Segment 1

(b)

(c)

Segment 2
32 possibilities

Figure 6.16: Using local segmentation for structure directed interpolation.

original image structure is replicated in the enlarged image. The segment means in each
message would be equal to the segment means calculated from the original 9 known pixels
in the  window. The data part of the message would only consist of residuals for the
known pixels. Of the 32 possible segment maps, the one with the shortest message length is
used to fill in values for the unknown pixels. It is straightforward to incorporate overlapping
average and posterior blending into image enlargement too.

Figure 6.17 compares three image enlargement methods for doubling montage (only the
middle shown). As expected, pixel replication produces blocky output, and linear interpolation produced blurry output. In comparison, the MML-256 result is good. The edges of
the artificial shapes and letters look more continuous. MML-256 based zooming may be
considered a fractal type approach, because the same prior is used at two different scales,
ensuring the image is self-similar at different resolutions.

Digital zooming has many applications. Most digital cameras on the market only support
optical zooming up to a point, after which digital zooming takes over. It is important that
the invented zoomed data appear realistic. Deinterlacing of digitized television signals can
be interpreted as taking an interlaced frame of size
 , and producing a full frame of
size
<
F . Predicting pixel values for the missing rows could be done by adapting the
MML-256 technique just described.

6.7 Image compression

241

Figure 6.17: Comparison of doubling methods: (a) denoised image; (b) pixel replication;
(c) linear interpolation; (d) MML-256 based enlargement.

6.7 Image compression

Image compression is concerned with efficiently encoding images for storage and transmission [RJ91]. Lossless image compression encodes the data exactly, the decoded image being
identical to the original. Lossy image compression encodes an approximation of the original
image in order to reduce the encoded length. Lossy algorithms must trade-off the level of
compression and the amount of distortion in the reconstructed image.
In the past, image processing and image compression have been treated as separate disciplines. The use of information theoretic techniques for analysing data, such as MML and

242

Chapter 6. Extensions and Further Applications of Local Segmentation

MDL, has proffered a merging of the two disciplines. Both image compression and MML
are concerned with optimal lossless encodings of the data. The data part of a two part MML
message may be considered the noise, while the model part is the structure. This could be
useful to a lossy compression algorithm for deciding which image information is important, in addition to any heuristics based on the human visual system. Better compression is
intimately linked with better understanding of data.

6.7.1 Aiding lossy compression

Lossy image compression may be considered to consist of two stages, illustrated in Figure 6.18. Firstly, a suitable approximation to the original image is determined. Secondly, the
approximate image is losslessly encoded. Although all lossy image compression algorithms
effectively do this, it is not always clear how they choose which information to discard.

0110010110001
0010100001110
1001100101100
1010101100011
1011001001011
Original Image

Lossy Approximation

Lossless encoding

Figure 6.18: Lossy compression can be considered a two stage process.

An alternative to allowing the lossy compression algorithm to distinguish between structure

and noise is to preprocess the image to remove the noise first. The parameters to the lossy
algorithm can then be chosen to preserve more of the information they are given. That is, the
discarding of visually unimportant information (first stage) can be replaced by a denoising
algorithm, leaving the encoding (second stage) to the compression algorithm.
In Chapters 4 and 5, local segmentation was shown to produce good denoising algorithms.
The denoised output of FUELS could be used as the input to a compression program. Having
the noise removed would assist in decreasing the size of the compressed file. An alternative
approach would be for a local segmentation technique to determine where bits should be
spent. For example, QO

regions have low information content, and hence fewer bits

6.7 Image compression

243

should be used for encoding those regions. Important structural details, such as edges, occur
in Qv
regions, and more bits should be used to ensure that these features are represented
accurately in the lossy compressed image.

6.7.2 Adaptive BTC

Block truncation coding (BTC) is a simple and fast technique for image compression, and
was discussed in Section 3.3.2. In its original form, BTC processes  blocks of pixels
at a time, and outputs data with a fixed bit rate. For 8 bit per pixel greyscale image data,
the resulting compression ratio is 4:1, or 2 bits per pixel. An example of its operation on a
sub-image of lenna is given in Figure 6.19.

Figure 6.19: Standard BTC: (a) original image at 8 bpp; (b) reconstructed image at 2 bpp;
(c) bitmap.

Local segmentation and BTC are closely related. Both techniques involve segmenting a
small group of connected pixels. The segmentation includes a bitmap for describing pixel
assignments to segments, and representative intensities for each segment. The original BTC
method always used two classes, which is similar to FUELS. However, BTC used the block
mean as a threshold, and chose the segment means to preserve the variance of the overall
block. This is in contrast to FUELS use of the adaptive Lloyd quantizer.
When a block is homogeneous, there is no need to transmit a bitmap and two means
a single mean is sufficient. Some variants of BTC attempt to exploit this property to save
bits and improve compression. To make a decodable bitstream, each blocks encoding is
preceded by one bit to distinguish between homogeneous and standard blocks. An alternative
is to send the two means first, and if they are the same, not to send a bitmap. Mitchell et

244

Chapter 6. Extensions and Further Applications of Local Segmentation

al [MD80] assume the region is homogeneous if the block variance is less than one, and
Nasiopoulos et al [NWM91] do the same if the range of the block is less than some threshold.
FUELS already provides a way to estimate a suitable threshold for distinguishing between
homogeneous and heterogeneous blocks. Figure 6.20 shows how Adaptive BTC can improve
the bit rate with little loss in image quality compared to standard BTC. The threshold used is
F& , where & is estimated as 2.84. This results in 41% of blocks being declared homogeneous,

reducing the overall bit rate to 1.39 bpp. The bitmap image clearly shows how two segments
were only used in high activity regions of the image.

Figure 6.20: Adaptive BTC: (a) original image; (b) bitmap; (c) reconstructed image at 1.39
bpp, 41% homogeneous blocks using threshold of 8.5; (d) reconstructed image
at 2 bpp.

Prefixing each block with a bit describing how many segments are in the block implicitly
places a prior probability distribution over segmentations. Specifically, it gives probability
0.5 to QAW
, and shares the remaining 0.5 between all possible  J clusterings, of which

there are

q #
FFB . This is obviously an inefficient coding, because Section 5.14 showed
that not all binary segment maps are equally likely. The MML methodology could be applied
to first determine these probabilities from the image. These probabilities could then be sent
up front, and the rest of the image encoded using them. If the cost of the upfront information
is small enough, a net gain in bit rate would be possible.

6.7 Image compression

245

6.7.3 Lossless predictive coding

The great majority of lossless compression algorithms are based on predicting the value
of the next pixel to be encoded. Examples include CALIC [WM96], LOCO [WSS00],
HBB [STM97], TMW [MT97] and Glicbawls [MT01]. The prediction must be formed using
only those pixels which have already been encoded the causal neighbourhood shown in
Figure 6.21. This is because the decoder does not yet know about future pixels.

NN
NW

WW

NE

Figure 6.21: The causal local neighbourhood consists of pixels known to both encoder and
decoder.

Typically, only pixels in the local neighbourhood are used for prediction. In Figure 6.21 they
have been labelled with the standard compass point notation [STM97], where N denotes
north, W denotes west, and so on. The pixel to be encoded, ?, is usually referred to as the
current pixel. The predicted value for the current pixel is often taken to be a linear combination of the pixels in the causal neighbourhood. For example, the Pirsch predictor [Pir80] is
calculated as 0.5W+0.25N+0.25NE.
Local segmentation could be used to improve the quality of predicted values. The amount
of noise in a predicted pixel value is a linear combination of the noise in each of the pixels
used for the prediction. It is desirable that this noise be minimized. Instead of using the
decoded (original) pixel values for prediction, denoised versions of them could be used.
For example, FUELS could be used to filter the image as it is decoded. The overlapping
estimate technique would even provide reasonable denoised values for those pixels without
a complete two-sided neighbourhood.

246

Chapter 6. Extensions and Further Applications of Local Segmentation

When encoding each pixel, it must be with respect to some distribution over possible pixel
values. The most commonly used distributions are Gaussian and Laplacian, due to their
symmetry and unimodality. The predicted value is used to estimate the location parameter
of the distribution. The spread parameter captures the level of confidence we have in the predicted value. Besides the prediction formula, the method used for determining an appropriate
spread value is what distinguishes most algorithms.
Some techniques implicitly use a single spread parameter for the whole image, such as lossless JPEG [PM93]. Some estimate it directly from the causal neighbourhood, like Glicbawls.
CALIC and LOCO use the idea of context coding, whereby the local causal neighbourhood is
classified into a context. Only a small number of contexts are used, with each one effectively
providing a spread estimate to use for encoding the current pixel.
The FUELS approach could be used for determining contexts. Let us assume the decoder has
been transmitted an estimate of the global noise variance, & . The causal neighbourhood could
be locally segmented. The optimally determined segment map could be used as a context
number. The number of contexts could be controlled through the number of neighbours
taking part in the segmentation. Most useful is that local segmentation could distinguish
between homogeneous and heterogeneous contexts. The spread parameter for each context
could be adaptively learned in a fashion similar to CALIC and LOCO.
A more general approach is to use local segmentation for both the prediction and the context
selection. The causal region could be optimally segmented. If Q

, the predicted value

will be in the same segment, or another new segment. The predicted value should probably
be equal to the mean of the causal neighbourhood. If Q

, the current pixel could belong

to either of the causal segments, or be from a new segment. The encoder could examine the
current pixel and determine which of the two segments it belongs to. A single binary event
could be used to state which of those two segments it is. The mean of that segment could be
used as the predicted value. This approach to prediction should be good at predicting large
changes in the vicinity of edges.
A more complex version of the previous idea could be based on the MML-256 algorithm. Let
us assume the encoder and decoder agree on a prior over segment maps. The causal neighbourhood could be segmented in all possible ways, not including the current pixel. For each

6.8 Conclusions

247

segment map considered, there are two new segmentations containing an extra pixel which
can be derived by assuming that the current pixel is in each of the two possible segments
(assuming QL#
). The prior could be used to assign weights to all these segmentations. The
predicted values could be blended together based on the posterior probability of each of the
segmentations. This is similar to the way TMW blends prediction distributions.

6.8 Conclusions
It has been shown how to apply local segmentation to a variety of image processing tasks
besides denoising, including edge detection, image zooming and image compression. Also
discussed was how to incorporate different image models and noise models into the local segmentation framework. These followed on from the constant facet and additive noise models
used by FUELS and the MML denoising algorithms in Chapters 4 and 5. The extension of
the ideas in this thesis from 2D images to 2D video sequences, 3D volume images, and 4D
volume image sequences was also considered.
The local segmentation principle is a simple one. It states that whenever a pixel is processed,
only those neighbouring pixels in the same segment should take part in the processing. This
chapter illustrated how this idea provides a consistent and unified way of examining images
in a low level manner. The simplicity of local segmentation allows it to be applied to any
algorithms where pixels are processed by looking at the values of neighbouring pixels. Its
power and flexibility lies in its simplicity.
Most of the ideas in this chapter could only be explored in brief, as the work required is outside the scope of this thesis. The most promising results are for probabilistic edge detection
and impulse noise removal. These ideas in particular would be deserving of further research.
The FUELS local segmentation methodology has a simple implementation, requiring little
memory and processing power. It could be crafted as a generic image processing tool which
could function in real time, potentially embedded in hardware.

Chapter 7
Conclusions
Local segmentation is already used, in some form, by a number of image processing algorithms. Sometimes its use is implicit, like with median filters, or more explicit, such as for
SUSAN. The properties of a particular algorithm can be examined by isolating the local segmentation model it uses. This can help to reveal its strengths and weaknesses. The more
situations an image processing techniques local segmentation criterion diagnoses correctly,
the better its performance. Image denoising is well suited to demonstrating the utility of local
segmentation. Good image denoising algorithms attempt to identify and preserve structure,
while simultaneously removing noise.
A denoising algorithm called FUELS was developed. Although binary thresholding is a very
simple type of segmentation, FUELS showed that it was good enough to outperform existing
state-of-the-art denoising algorithms. Unconnected segments are often unavoidable, because
the intersection of a window with a global segment may disguise their actual connectedness
via pixels outside the window. Thus the main limitation of thresholding may actually be an
advantage when applied locally. The windows used in local processing are usually too small
to incorporate much spatial information anyway.
The move from assuming homogeneity within the local region to allowing for the presence
of two segments lead to a big improvement in performance. This gain was much larger than
that which came from also allowing more than two segments. Although this was true for
 windows, it may not be true for larger ones. The popularity of the  window, along

249

250

Chapter 7. Conclusions

with results for FUELS, suggest that it represents a good trade-off between compact locality
and the need for more detailed models.
The work on FUELS introduced the idea of do no harm (DNH). FUELS generates two
candidate models, and chooses the one it thinks is better. It is possible that both candidates
are poor ones for the local region, so using one could do more damage than good. If this
situation was identified by the DNH mode, the pixels were left unmodified. This is similar
to defaulting to a null model if there is insufficient evidence for any of the alternatives. The
DNH idea is not particular to FUELS or even local segmentation it could be applied to any
denoising algorithm to improve its worst case performance, especially at low noise levels.
Most denoising algorithms use the centre pixel of the window as a reference pixel to compare with each other pixel in turn, to produce a denoised estimate of the centre pixel only.
Local segmentation differs from this in that it treats all pixels in the window democratically,
producing denoised values for them all. Because windows overlap, there are multiple estimates for each pixel. With little extra work, FUELS was able to average these estimates to
further improve performance. This illustrates the advantages of combining predictions from
different experts, which is harder to do when the centre pixel receives special treatment.
Moving to an MML framework lead to better local segmentation, which itself lead to further
improvements in denoising performance. MML made it straightforward to consider a larger
set of models. It provides a robust method for comparing the relative merit of different
models, but can not rate them in an absolute sense. The DNH principle suggested the addition
of a null model, which left the pixels unmodified, to serve as a benchmark. By comparing
against this model it was possible to establish whether the other models were useful or not.
The MML denoising algorithms performed best of all when the noise level was high. This
showed that thresholding, as a technique for local segmentation, was insufficient for very
noisy images. MML provided a way to incorporate spatial information through its use of a
data-driven prior for the segment maps. It was able to learn the general characteristics of an
image, which guided it in segmenting each local region. It found that for most images, some
segment maps are more likely to occur than others. Homogeneous windows were the most
common situation diagnosed, which is why simplistic denoising techniques, like box filters,
do well for most parts of an image.

251
Local segmentation could be considered for use in any situation where the current pixel is
processed by making reference to the values of its neighbouring pixels. Algorithms that deal
with these situations make up a large proportion of algorithms used in image processing. It
was shown that local segmentation is applicable to a host of image processing tasks besides
denoising. This is because it makes obvious the connection between higher level analysis,
which typically involves segmentation, and low level analysis, namely local segmentation.
The principle of local segmentation is a simple one, but has not really been explicitly stated
in the literature. The success of FUELS showed that even a simplistic implementation of local segmentation was able to produce high quality results. The application of MML to local
segmentation was original, and lead to further small improvements in local image modeling.
Because FUELS and SUSAN already represent an excellent trade-off between efficiency
and effectiveness, many applications may not be able to exploit MMLs better local approximations anyway. The success of local segmentation in denoising bodes well for its wider
application in image processing. The idea of data driven priors, using a null model as a reference, and mixing expert predictions can be applied not just to images, but to nearly any
data processing domain.

Bibliography
[AG98]

C. Ambroise and G. Govaert. Convergence of an EM-type algorithm for spatial

clustering. Pattern Recognition Letters, 19:919927, 1998.

[Aka92] H. Akaike. Information theory and an extension of the maximum likelihood principle. In Samuel Kotz and Norman L. Johnson, editors, Breakthroughs in Statistics, volume I, pages 599624. Springer-Verlag., New York, 1992. with an introduction by J. deLeeuw.
[Alp98]

Ethem Alpaydin. Soft vector quantization and the EM algorithm. Neural Networks, 11:467477, 1998.

[Ama87] John Amanatides. Realism in computer graphics: A survey. IEEE Computer

Graphics and Applications, 7(1):4456, January 1987.
[AW96] Jan Allebach and Ping Wah Wong. Edge-directed interpolation. In International
Conference on Image Processing, pages 707710. IEEE, 1996.
[BB95]

Leon Bottou and Yoshua Bengio. Convergence properties of the k-means algorithm. In Advances in Neural Information Processing Systems, volume 7, Denver,
1995. MIT Press.

[BBB00] Kamel Belkacem-Boussaid and Azeddine Beghdadi. A new image smoothing

method based on a simple model of spatial processing in the early stages of human
vision. IEEE Transactions on Image Processing, 9(2):220226, February 2000.
[BC90]

J. M. H. Du Buf and T. G. Campbell. A quantitative comparison of edge preserving

smoothing techniques. Signal Processing, 21(4):289301, 1990.
253

254
[Bes86]

BIBLIOGRAPHY
J. Besag. On the statistical analysis of dirty pictures. Journal of the Royal Statistical Society, Series B, 48(3):259302, 1986.

[Bez81] J. C. Bezdek, editor. Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum, NY, 1981.
[BF98]

Paul S. Bradley and Usama M. Fayyad. Refining initial points for K-Means clustering. In Proc. 15th International Conf. on Machine Learning, pages 9199.
Morgan Kaufmann, San Francisco, CA, 1998.

[BO95]

Rohan. A. Baxter and Jonathan J. Oliver. MDL and MML: Similarities and differences (Introduction to minimum encoding inference - Part III). Technical Report
94/207, Department of Computer Science, Monash University, January 1995.

[Bow84] D. R. K. Bownrigg. The weighted median filter. Communications of the ACM,

27:807818, 1984.
[Boz83] H. Bozdogan. Determining the number of component clusters in the standard
multivariate normal mixture model using model selection criteria. Technical Report TR UIC/DQM/A83-1, Quantitative Methods Dept., University of Illinois,
Chicago, Illinois, 60680, 16 June 1983.
[BS85]

R. Bracho and A.C. Sanderson. Segmentation of images based on intensity gradient information. In Proceedings of CVPR-85 Conference on Computer Vision and
Pattern Recognition, San Francisco, pages 341347, 1985.

[BS94]

J. M. Bernado and A. F. M. Smith. Bayesian Theory. Wiley, New York, 1994.

[BSMH98] Michael J. Black, Guillermo Sapiro, David H. Marimont, and David Heeger.
Robust anisotropic diffusion. IEEE Transactions on Image Processing, 7(3):421
432, March 1998.
[BW72] D. M. Boulton and C. S. Wallace. An information measure for hierarchic classification. Computer Journal, 16(3):254238, 1972.
[Cai96]

Z. Q. Cai.

Correlation-based block truncation coding.

32(25):23052306, December 1996.

Electronics Letters,

BIBLIOGRAPHY

255

[Can86] John Canny. A computational approach to edge detection. IEEE Trans. Pattern
Analysis and Machine Intelligence, PAMI-8(6):679698, November 1986.
[CC94]

K. W. Chan and K. L. Chan. Optimized adaptive AMBTC. In International

Conference on Image Processing, pages 371374. IEEE, 1994.

[Cho99] Wen-Shou Chou. Classifying image pixels into shaped, smooth, and textured
points. Pattern Recognition, 32:16971706, 1999.
[CHY89] Sungzoon Cho, Robert Haralick, and Seungku Yi. Improvement of Kittler and
Illingworths minimum error thresholding. Pattern Recognition, 22(5):609517,
1989.
[CK72]

C. K. Chow and T. Kaneko. Automatic boundary detection of the left ventricle

from cineangiograms. Comput. Biomed. Res., 5:388410, 1972.

[CR93]

T. Caelli and D. Reye. On the classification of image regions by color, texture and
shape. Pattern Recognition, 26:461470, 1993.

[Cur91]

A. Curri. Synthetic aperture radar. Electronics and Communication Engineering

Journal, 3(4):159170, 1991.

[CW00] T. Chen and H. R. Wu. Impulse noise removal by multi-state median filtering. In
Proc. IEEE Int. Conf. on Acoustics, Speech, and Signal Processing, volume IV,
pages 21832186, June 2000.
[DDF 98] D. L. Dowe, M. Doran, G. E. Farr, A. J. Hurst, D. R. Powell, and T. Seemann.
Kullback-Leibler distance, probability and football prediction. In W. Robb, editor,
Proceedings of the Fourteenth Biennial Australian Statistical Conference (ASC14), page 80, July 1998.
[DH73]

R. O. Duda and P. E. Hart. Pattern Classification and scene analysis. Wiley, New
York, 1973.

[DLR77] A. P. Dempster, N. M. Laird, and D. B. Rubin. Maximum likelihood from incomplete data via the EM algorithm (with discussion). Journal of the Royal Statistical
Society (Series B), 39(1):138, 1977.

256

BIBLIOGRAPHY

[DM79] E. J. Delp and O. R. Mitchell. Image coding using block truncation coding. IEEE
Trans. Commun., 27:13351342, 1979.
[Ede99] Shimon Edelman. Representation and recognition in vision. MIT Press, Cambridge, Mass., 1999.
[EH81]

B. S. Everitt and D. J. Hand. Finite Mixture Distributions. Chapman and Hall Ltd,
1981.

[ELM91] N. Efrati, H. Liciztin, and H. B. Mitchell. Classified block truncation codingvector quantization: an edge sensitive image compression algorithm. Signal Process. Image Commun., 3:275283, 1991.
[EM00]

How-Lung Eng and Kai-Kuang Ma. Noise adaptive soft-switching median filter
for image denoising. In Proc. IEEE Int. Conf. on Acoustics, Speech, and Signal
Processing, volume IV, pages 21752178, June 2000.

[Far99]

G. Farr. Information Theory and MML Inference (Lecture Notes). School of

Computer Science and Software Engineering, Monash University, 1999.

[FCG85] J.P. Fitch, E.J. Coyle, and N.C. Gallagher, Jr. Root properties and convergence
rates of median filters. IEEE Trans. Acoustics, Speech, and Signal Processing,
33:230240, 1985.
[FH96]

Winfried A. Fesslenz and Georg Hartmann. Preattentive grouping and attentive

selection for early visual computation. In Proc. 13th International Conference
on Pattern Recognition, volume IV: Parallel and Connectionist Systems, pages
340345, Los Alamitos, CA., 1996. IEEE Computer Science.

[Fis12]

R. A. Fisher. Messenger Math., 41:155160, 1912.

[Fis36]

R. A. Fisher. Annals of Eugenics, 7:376386, 1936.

[FKN80] Henry Fuchs, Zvi M. Kedem, and Bruce F. Naylor. On visible surface generation
by a priori tree structures. Computer Graphics (SIGGRAPH 80 Proceedings),
14(3):124133, July 1980.

BIBLIOGRAPHY
[FM81]

257

K. S. Fu and J. K. Mui. A survey on image segmentation. Pattern Recognition,

13:316, 1981.

[FNK94] Pasi Franti, Olli Nevalainen, and Timo Kaukoranta. Compression of digital images by block truncation coding: A survey. The Computer Journal, 37(4):308
332, 1994.
[GG84]

S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions and the

Bayesian restoration of images. IEEE Trans. Patt. Anal. Machine Intell., 6(6):712
741, November 1984.

[GG91]

A. Gersho and R. Gray. Vector Quantization and Signal Compression. Kluwer

Academic Publishers, 1991.

[GGLR81] Jr. Glen G. Langdon and Jorma Rissanen. Compression of black-white images
with arithmetic coding. IEEE Transactions on Communications, 29(6):858867,
June 1981.
[Gla93]

C. A. Glasbey. An analysis of histogram-based thresholding algorithms. CVGIP:

Graphical Models and Image Processing, 55(6):532537, November 1993.

[GN98]

Robert M. Gray and David L. Neuhoff. Quantization. IEEE Transactions on

Information Theory, 44(6):163, October 1998.

[GO96]

Mehmet I. Gurelli and Levent Onural. A class of adaptive directional image

smoothing filters. Pattern Recognition, 29(12):19952004, 1996.

[Gro88] Richard A. Groeneveld. Introductory Statistical Methods : An Integrated Approach Using Minitab. PWS-KENT Publishing Company, 1988.
[GW92] Rafael C. Gonzalez and Richard E. Woods. Digital Image Processing. AddisonWesley, 3rd edition, 1992.
[HB88]

R. J. Hathaway and J. C. Bezdek. Recent convergence results for the fuzzy cmeans clustering algorithms. J. Classification, 5(2):237247, 1988.

[HH94]

Michael W. Hansen and William E. Higgins. Watershed-driven relaxation labeling

for image segmentation. In International Conference on Image Processing, pages
460463. IEEE, 1994.

258
[HH95]

BIBLIOGRAPHY
Michael W. Hansen and William E. Higgins.

Image enhancement using

watershed-based maximum homogeneity filtering. In IEEE 2nd International

Conference on Image Processing, volume 2, pages 482485. IEEE, 1995.
[HH99]

Michael W. Hansen and William E. Higgins.

Watershed-based maximum-

homogeneity filtering. IEEE Trans. Image Processing, 8(7):982988, July 1999.

[Hig91]

W. E. Higgins. A flexible implementation of maximum-homogeneity filtering for

2-D and 3-D images. IEEE Trans. Signal Processing, 39(10):23252331, October
1991.

[HipBC] Hippocrates.

The oath.

400 B.C.

Translated by Francis Adams.

http://classics.mit.edu/Hippocrates/hippooath.html.
[HM01] Pierre Hansen and Nenad Mladenovic. J-MEANS: a new local search heuristic
for minimum sums of squares clustering. Pattern Recognition, 34:405413, 2001.
[HP74]

S. Horowitz and T. Pavlidis. Picture segmentation by a directed split-and-merge

procedure. In Proc. 2nd Int. Joint Conf. Pattern Recognition (ICPR74), pages
424433, 1974.

[HR94]

Iftekhar Hussain and Todd R. Reed. Segmentation-based nonlinear image smoothing. In International Conference on Image Processing, pages 507511. IEEE,
1994.

[HRA96] D. Hagyard, M. Razaz, and P. Atkin. Analysis of watershed algorithms for

greyscale images. In International Conference on Image Processing, pages 4144.
IEEE, 1996.
[HS85]

Robert M. Haralick and Linda G. Shapiro. Image segmentation techniques. Computer Vision, Graphics, and Image Processing, 29:100132, 1985.

[HSD73] R. M. Haralick, K. Shanmugam, and I. Dinstein. Texture features for image classification. IEEE Trans. Systems, Mans and Cybernetics, SMC-3:610621, 1973.
[HSSB98] Mike Heath, Sudeep Sarkar, Thomas Sanocki, and Kevin Bowyer. Comparision
of edge detectors: A methodology and initial study. Computer Vision and Image
Understanding, 69(1):3854, January 1998.

BIBLIOGRAPHY

259

[Hua84] N. Huang. Markov model for image segmentation. In 22nd Allerton Conference
on Communication, Control, and Computing, pages 775781, October 35 1984.
[Hub81] Peter J. Huber, editor. Robust Statistics. Wiley series in probability and mathematical statistics. John Wiley & Sons, Inc., 1981.
[HW79] J. A. Hartigan and M. A. Wong. A K-means clustering algorithm. Applied Statistics, 28:100108, 1979.
[HW81] R. M. Haralick and L. Watson. A facet model for image data. Computer Graphics
and Image Processing, 15:113129, 1981.
[Imm96] John Immerkr. Fast noise variance estimation. Computer Vision and Image
Understanding, 64(2):300302, September 1996.
[Jah93]

Bernd Jahne. Digital Image Processing - Concepts, Algorithms and Scientific

Applications. Springer-Verlag, second edition, 1993.

[Jai89]

Anil K. Jain. Fundamentals of Digital Image Processing. Prentice Hall, 1989.

[JRvdH93] R. C. Jansen, K. Reinink, and G. W. A M. van der Heijden. Analysis of grey

level histograms by using statistical methods for mixtures of distributions. Pattern
Recognition Letters, 14:585590, July 1993.
[KD95]

Helmut Kopka and Patrick W. Daly. A Guide to LATEX2e - Document Preparation

for Beginners and Advanced Users. Addison-Wesley, second edition, 1995.

[KI86]

J. Kittler and J. Illingworth. Minimum error thresholding. Pattern Recognition,

19(1):4147, 1986.

[KL51]

S. Kullback and R. A. Leibler. On information and sufficiency. Annals of Mathematical Statistics, 22(1):7986, 1951.

[KL91]

S. Ko and Y. Lee. Center weighted median filters and their applications to image
enhancement. IEEE Trans. on Circuits and Systems, 38(9):984993, September
1991.

[Knu73] D. E. Knuth. Sorting and Searching, volume 3 of The Art of Computer Programming. Addison-Wesley, 2nd edition, 1973.

260

BIBLIOGRAPHY

[KOA92] T. Kurita, N. Otsu, and N. Abdelmalek. Maximum likelihood thresholding based

on population mixture models. Pattern Recognition, 25(10):12311240, 1992.
[KS00]

Klaus Koster and Michael Spann. MIR: An approach to robust clustering

application to range segmentation. IEEE Trans. on Pattern Analysis and Machine
Intelligence, 22(5):430444, May 200.

[KSSC87] Darwin T. Kuan, Alexander A. Sawchuk, Timothy C. Strand, and Pierre Chavel.
Adaptive restoration of images with speckle. IEEE Transactions on Acoustics,
Speech, and Signal Processing, ASSP-35(3):373383, March 1987.
[KSW85] J. N. Kapur, P. K. Sahoo, and A. K. C. Wong. A new method for gray-level picture
thresholding using the entropy of the histogram. Computer Vision, Graphics, and
Image Processing, 29:273285, 1985.
[Kuh00] J. Kuha. Model assessment and model choice: an annotated bibliography. Technical report, Department of Statistics, Pennsylvania State University, April 2000.
http://www.stat.psu.edu/jkuha/msbib/biblio.html.
[Kuw76] M. Kuwahara. Processing of RI-angio-cardiographic images. In K. Preston
and M. Onoe, editors, Digital Processing of Biomedical Images, pages 187203.
Plenum Press, New York, NY, 1976.
[KWT87] M. Kass, A.P. Witkin, and D. Terzopoulos. Snakes: Active contour models. In
International Conference on Computer Vision (ICCV), pages 259268, London,
1987. IEEE.
[LCP90] Sang Uk Lee, Seok Yoon Chung, and Rae Hong Park. A comparitive performance
study of several global thresholding techniques for segmentation. Computer Vision, Graphics, and Image Processing, 52:171190, 1990.
[Lee81a] J. S. Lee. Speckle analysis and smoothing of synthetic aperture radar images.
Computer Graphics and Image Processing, 17:2432, 1981.
[Lee81b] Jong-Sen Lee. Refined filtering of image noise using local statistics. Computer
Graphics and Image Processing, 15:380389, 1981.

BIBLIOGRAPHY
[Lee83]

261

J-S. Lee. Digital image smoothing and the sigma filter. Computer Vision, Graphics, and Image Processing, 24:255269, 1983.

[len72]

Playmate of the Month: Lenna Sjoo blom. Playboy: Entertainment for Men, pages
134141, November 1972.

[Li01]

S. Z. Li. Markov random field modeling in image analysis. Springer, 2001.

[Lin96]

Tetra Lindarto. MML segmentation-based image coding. Masters thesis, Department of Computer Science, Monash University, May 1996.

[LL96]

C. K. Leung and F. K. Lam. Maximum segmented-scene spatial entropy thresholding. In International Conference on Image Processing, pages 963966. IEEE,
1996.

[Llo82]

S. P. Lloyd. Least squares quantization in PCM. IEEE Trans. Inf. Theory, 28:129
137, 1982.

[LM84]

M. D. Lema and O. R. Mitchell. Absolute moment block truncation coding and

its application to color images. IEEE Trans. Commun., 32:11481157, 1984.

[Mas85] G. A. Mastin. Adaptive filters for digital image noise smoothing: An evaluation.
Computer Vision, Graphics, and Image Processing, 31:103121, 1985.
[MB90] F. Meyer and S. Beucher. Morphological segmentation. J. Vis. Commun. Image
Represent., 1(1):2146, 1990.
[McD81] M. J. McDonnell. Box-filtering techniques. Computer Graphics and Image Processing, 17:6570, 1981.
[McQ67] J. B. McQueen. Some methods of classification and analysis of multivariate observations. In L.M. Le Cam and J. Neyman, editors, Proceedings of Fifth Berkeley
Symposium on Mathematical Statistics and Probability, pages 281297, 1967.
[MD80] O. R. Mitchell and E. J. Delp. Multilevel graphics representation using block
truncation coding. Proc. IEEE, 68:868873, 1980.
[MJ66]

C. E. K. Mees and T. H. James. The Theory of the Photographics Process.

Macmillan, 1966.

262

BIBLIOGRAPHY

[MJR90] P. Meer, J. Jolion, and A. Rosenfeld. A fast parallel algorithm for blind estimation
of noise variance. IEEE Trans. Pattern Anal. Mach. Intelligence, 12(2):216223,
1990.
[MM99] Jean-Bernard Martens and Lydia Meesters. The role of image dissimilarity in
image quality models. In B. E. Rogowitz and T. N. Pappas, editors, Proceedings
of SPIE: Human Vision and Electronic Imaging IV, volume 3644, pages 258269,
January 1999.
[MP00]

Geoffrey McLachlan and David Peel. Finite Mixture Models. Wiley series in
probability and statistics. John Wiley & Sons, Inc., 2000.

[MPA00] Volker Metzler, Marc Puls, and Til Aach. Restoration of ultrasound images by
nonlinear scale-space filtering. In Edward R. Dougherty and Jaakko T. Astola,
editors, SPIE Proceedings: Nonlinear Image Processing XI, volume 3961, pages
6980. SPIE, 2000.
[MR74] Frederick Mosteller and Robert E. K. Rourke, editors. Sturdy Statistics: Nonparametric and Order Statistics. Addison-Wesley, 1974.
[MR95] Kai-Kuang Ma and Sarah A. Rajala. New properties of AMBTC. IEEE Signal
Processing Letters, 2(2):3436, February 1995.
[MT97]

Bernd Meyer and Peter E. Tischer. TMW a new method for lossless image
compression. In Picture Coding Symposium (PCS97), pages 533538, Berlin,
Germany, 1997. VDE-Verlag GMBH.

[MT01]

Bernd Meyer and Peter Tischer. Glicbawls grey level image compression by
adaptive weighted least squares. In J. A. Storer and M. Cohn, editors, Proceedings
of the Data Compression Conference (DCC2001), Piscataway, NJ, USA, March
2001. IEEE Service Center.

[NH88]

A. Netravali and B. Haskell. Digital Pictures, Representation and Compression.

Plenum, New York, 1988.

[Nie94]

Yves Nievergelt. Total least squares: State-of-the-art regression in numerical analysis. SIAM Review, 36(2):258264, June 1994.

BIBLIOGRAPHY
[NR79]

263

Yasuo Nakagawa and Azriel Rosenfeld. Some experiments on variable thresholding. Pattern Recognition, 11:191204, 1979.

[NWM91] P. Nasiopoulos, R. Ward, and D. Morse. Adaptive compression coding. IEEE

Trans. Commun., 39:12451254, 1991.
[OB94]

Jonathan J. Oliver and Rohan A. Baxter. MML and Bayesianism: Similarities and
differences. Technical Report 206, Department of Computer Science, Monash
University, December 1994.

[OH94]

Jonathan J. Oliver and David Hand. Introduction to minimum encoding inference. Technical Report 205, Department of Computer Science, Monash University, November 1994.

[Ols93]

S. I. Olsen. Estimation of noise in images: An evaluation. CVGIP: Graphical

Models and Image Processing, 55(4):319323, July 1993.

[Ots79]

N. Otsu. A threshold selection method from gray level histograms. IEEE Trans.
Systems, Man and Cybernetics, 9:6266, March 1979.

[PG94]

N. Papamarkos and B. Gatos. A new approach for multilevel threshold selection. CVGIP: Graphical Models and Image Processing, 56(5):357370, September 1994.

[PH66]

E. S. Pearson and H. O. Heartly, editors. Biometrika Tables for Statisticians,

volume 1. Cambridge University Press, 3rd edition, 1966.

[Pir80]

P. Pirsch. A new predictor design for DPCM coding of TV signals. In ICC Conference Report (International Conf. on Communications), pages 31.2.131.2.5,
Seattle, WA, 1980.

[Pit95]

Ioannis Pitas. Digital Image Processing Algorithms. Prentice Hall International

Series in Acoustics, Speech and Signal Processing. Prentice Hall, 1995.

[PM66]

J. M. Prewitt and M. L. Mendelsohn. The analysis of cell images. Ann. New York
Acad. Sci.,, 128:10251053, 1966.

264
[PM90]

BIBLIOGRAPHY
P. Perona and J. Malik. Scale-space and edge detection using anisotropic diffusion.
IEEE Trans. Pattern Anal. Machine Intell., 12(7):629639, July 1990.

[PM93]

W. Pennebaker and J. Mitchell. JPEG: Still Image Compression Standard. Van

Nostrand Reinhold, New York, 1993.

[PMGL88] W.B. Pennebaker, J.L. Mitchell, and R.B. Arps G.G. Langdon. An overview of
the basic principles of the Q-coder adaptive binary arithmetic coder. IBM Journal
of research and development, 32(6):771726, November 1988.
[PP93]

Nikhil R. Pal and Sankar K. Pal. A review on image segmentation techniques.

Pattern Recognition, 26(9):12771294, 1993.

[Pre89]

S. J. Press. Bayesian Statistics. Wiley, 1989.

[PV90]

I. Pitas and A. N. Venetsanopoulos. Nonlinear Digital Filters: Principles and

Applications. Kluwer Academic, 1990.

[QS95]

M. K. Quweider and E. Salari. Gradient-based block truncation coding. Electronics Letters, 31(5):353354, March 1995.

[RC78]

T.W. Ridler and S. Calvard. Picture thresholding using an iterative selection

method. IEEE Transactions on Systems, Man and Cybernetics, SMC-8(8):630
632, August 1978.

[RHZ76] A. Rosenfeld, R. A. Hummel, and S. W. Zucker. Scene labeling by relaxation

operations. IEEE Trans. on Systems, Man and Cybernetics, 6:420433, 1976.
[Ris78]

J. Rissanen. Modeling by shortest data description. Automatica, 14:465471,

1978.

[Ris87]

Jorma Rissanen. Stochastic complexity. J. R. Statist. Soc. B, 49(3):223239, 1987.

[Ris00]

Jorma Rissanen.

MDL denoising.

IEEE Trans. on Information Theory,

46(7):25372543, 2000.
[RJ91]

Majid Rabbani and Paul W. Jones. Digital Image Compression Techniques. SPIE
Optical Engineering Press, 1991.

BIBLIOGRAPHY
[RL87]

265

Peter J. Rousseeuw and Annick. M. Leroy, editors. Robust regression and outlier detection. Wiley series in probability and mathematical statistics: Applied
probability and statistics. John Wiley & Sons, Inc., 1987.

[RLU99] K. Rank, M. Lendl, and R. Unbehauen. Estimation of image noise variance. IEE
Proc.-Vis. Image Signal Process., 146(2):8084, April 1999.
[Ros81] A. Rosenfeld. Image pattern recognition. Proc. IEEE, 69:596605, 1981.
[RR98]

Stephen J. Roberts and Iead Rezek. Bayesian approaches to Gaussian mixture

modeling. IEEE Trans. Pattern Analysis and Machine Intelligence, 20(11):1133
1142, November 1998.

[RRK84] S. S. Reddi, S. F. Rudin, and H. R. Keshavan. An optimal multiple threshold

scheme for image segmentation. IEEE Trans. Systems, Man, and Cybernetics,
14(4):661665, July/August 1984.
[SB97]

S. M. Smith and J. M. Brady. SUSAN - a new approach to low level image

processing. International Journal of Computer Vision, 23(1):4578, May 1997.

[Sha48]

C. E. Shannon. A mathematical theory of communication. Bell Systems Technical

Journal, 27:379423, 1948.

[She90]

Q. Shen. Fuzzy image smoothing. In proc. Int. Conf. on Pattern Recognition,

pages 7478, 1990.

[SKYS] Brian V. Smith, Paul King, Ken Yap, and Supoj Sutanthavibul. XFig: vector
graphics drawing software. http://www.xfig.org/.
[SMCM91] Philippe Saint-Marc, Jer-Sen Chen, and Gerard Medioni. Adaptive smoothing: A general tool for early vision. IEEE Trans. Pattern Analysis and Machine
Intelligence, 13(6):514529, June 1991.
[Smi95] S. M. Smith. SUSAN - a new approach to low level image processing. Technical
Report TR95SMS1, Defence Research Agency, Chobham Lane, Chertsey, Surrey,
UK, 1995. http:/www.fmrib.ox.ac.uk/steve/.

266

BIBLIOGRAPHY

[Smi96] S. M. Smith. Flexible filter neighbourhood designation. In Proc. 13th Int. Conf.
on Pattern Recognition, volume 1, pages 206212, 1996.
[Sob70] I. E. Sobel. Camera Models and Machine Perception. PhD thesis, Electrical
Engineering Department, Stanford University, Stanford, CA., 1970.
[SSW88] P. K. Sahoo, S. Soltani, and A. K. C. Wong. A survey of thresholding techniques.
Computer Vision, Graphics and Image Processing, 41:233260, 1988.
[ST97]

Torsten Seemann and Peter E. Tischer. An MML approach to noise cleaning. In

Picture Coding Symposium (PCS97), pages 3943, Berlin, Germany, September
1997. VDE-Verlag GMBH.

[ST98]

Torsten Seemann and Peter E. Tischer. Structure preserving noise filtering using explicit local segmentation. In Proc. IAPR Conf. on Pattern Recognition
(ICPR98), volume II, pages 16101612, August 1998.

[Sta99]

R. C. Staunton. Hexagonal sampling in image processing. Advances in Imaging

and Electron Physics, 107:231307, April 1999.

[STM97] Torsten Seemann, Peter E. Tischer, and Bernd Meyer. History based blending of
image sub-predictors. In Picture Coding Symposium (PCS97), pages 147151,
Berlin, Germany, September 1997. VDE-Verlag GMBH.
[TA97]

Mark Tabb and Narendra Ahuja. Multiscale image segmentation by integrated

edge and region detection. IEEE Trans. Image Processing, 6(5):642655, May
1997.

[Tag96]

Akira Taguchi. A design method of fuzzy weighted median filters. In International

Conference on Image Processing, pages 423426. IEEE, 1996.

[TH94]

Patrick Teo and David Heeger. Perceptual image distortion. In First IEEE International Conference on Image Processing, volume 2, pages 982986, November
1994.

[Tho18] W. M. Thorburn. The myth of Occams razor. Mind, 27:345353, 1918.

BIBLIOGRAPHY
[Tis94]

267

Peter E. Tischer. Optimal predictors for image compression. Technical Report

189, Department of Computer Science, Monash University, 1994.

[TSM85] D. M. Titterington, A. F. M. Smith, and U. E. Makov. Statistical Analysis of Finite

Mixture Distributions. John Wiley & Sons Ltd., 1985.
[Tuk71] J. W. Tukey. Exploratory Data Analysis. Addison-Wesley, 1971.
[VS91]

Luc Vincent and Pierre Soille. Watersheds in digital spaces: An efficient algorithm
based on immersion simulations. IEEE Trans. Pattern Analysis and Machine Intelligence, 13(6):583598, 1991.

[WB68] C. S. Wallace and D. M. Boulton. An information measure for classification.

Computer Journal, 11(2):185194, 1968.
[WD00] C. S. Wallace and D. L. Dowe. MML clustering of multi-state, Poisson, von
Mises circular and Gaussian distributions. Journal of Statistics and Computing,
10(1):7383, January 2000.
[Wei97] Joachim Weickert. A review of nonlinear diffusion filtering. In B. ter Haar Romney, L. Florack, J. Koenderink, and M. Viergever, editors, Scale-space theory in
computer vision, volume 1252, pages 328. Springer-Verlag, 1997.
[Wes78] Joan S. Weszka. A survey of threshold selection techniques. Computer Graphics
and Image Processing, 7:259265, 1978.
[WF87]

C. S. Wallace and P. R. Freeman. Estimation and inference by compact coding. J.

R. Statist. Soc. B, 49(3):240265, 1987.

[WKL ] Thomas Williams, Colin Kelley, Russell Lang, Dave Kotz, John Campbell,
Gershon Elber, and Alexander Woo.

gnuplot:

graph plotting software.

http://www.gnuplot.info/.
[WM96] Xiaolin Wu and Nasir Memon. CALIC - a context based adaptive lossless image
codec. IEEE ASSP, 4:18901893, 1996.
[WR78] Joan S. Weszka and Azriel Rosenfeld. Threshold evaluation techniques. IEEE
Trans. Systems, Man, and Cybernetics, 8(8):622629, August 1978.

268

BIBLIOGRAPHY

[WSS00] Marcelo J. Weinberger, Gadiel Seroussi, and Guillermo Sapiro. The LOCO-I
lossless image compression algorithm: Principles and standardization into JPEGLS. volume 9, pages 13091324, August 2000.
[WVL81] D. C. C. Wang, A. H. Vagnucci, and C. C. Li. Gradient inverse weighted smoothing scheme and the evaluation of its performance. CGIP, 15(2):167181, February
1981.
[YG01]

G. Z. Yang and D. F. Gillies. Computer Vision Lecture Notes. Department of Computing, Imperial College, UK, 2001. http://www.doc.ic.ac.uk/gzy.

[ZG94]

Y. J. Zhang and J. J. Gerbrands. Objective and quantitative segmentation evaluation and comparison. Signal Processing, 39:4354, 1994.

[Zha97] Y. J. Zhang. Evaluation and comparison of different segmentation algorithms.

Pattern Recognition Letters, 18:963974, 1997.