Blind city classification using aggregation of clusterings

savagelizardΤεχνίτη Νοημοσύνη και Ρομποτική

25 Νοε 2013 (πριν από 3 χρόνια και 8 μήνες)

58 εμφανίσεις

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


1

Blind city classification using aggregation
of clusterings

Ivan Kyrgyzov, Henri Maître


T
él
écom

Paris

7
th

CNES/DLR Workshop on Information Extraction and Scene
Understanding for Meter Resolution Images

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


2

Plan of presentation

1.
Problem statement


2.
Unsupervised clustering algorithms


3.
The optimal number of clusters (MDL criterion)


4.
Usupervised clustering combination


5.
General schema of image clustering


6.
Image data


7.
Results

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


3

1. Problem statement

Satellite Image Analysis as a Pattern Recognition Problem

Questions
:

Problem
: different algorithms give different clusterings




different views of data representations



some clusters are common, some
clusters
are unique

Method
:
unsupervised
clustering algorithms



1) Do not search classes manually.



2) Descriptors form clusters.

Solution
: combine results of clustering algorithms


and find a consensus



Which patterns?



How many pattern prototypes?



How to discriminate them?

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


4

2. Unsupervised

clustering algorithms


EM
-
algorithm for GMM estimation (Autoclass, Cheeseman, 1996)


Kernel K
-
means algorithm (
Taylor & Cristianini 2004)



Spectral K
-
means algorithm (
A.Ng, 2002
)


Ward hierarchical algorithm (Ward, 1963)



Combining different

clusterings

using

a

co
-
association

matrix

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


5

3.
The optimal number of clusters (MDL)

Minimum Description Length (Rissanen, 1978)

)
log(
2
1
))
|
(
log(
min
,
M
P









(1)

(2)

P(X|
Θ
)

of Gaussians and a “hard clustering” (1) is:


















i
i
i
i
const
M
n
n
log
2
1
log

i
n
= number of samples in i
th

cluster,

i

= determinant of covariance matrix

i

calculated both for spectral and kernel clustering algorithms

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


6

4. Unsupervised clustering combination


2
min
arg
T
XX
C
X




1
,
0
,


X
I
X
X
to
subject
T
Co
-
association

matrix

s
clustering
of
number
,
1
1
N
D
N
C
N
n
n







otherwise.
0,
;
clustering

n
for
cluster

same

in the

are

j

and

i

if

,
1
th
n
ij
D
Fred & Jain

(PAMI 2005)

Final

c
ombination

= , such that:

X
(3)


otherwise.
0,
Q;
1,...,
q

i,


q,


i

if

,
1







iq
X
Combination algorithm:




single
-
link based clustering (1974, Hubert) to minimize (3)

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


7

5. General schema of image clustering

SPOT5
images

Samples

Image data

Models



Model 1

Model 2

Selected
Models


SModel 1

SModel 2

Clustering 1

Clustering 2

Sample
extraction

Feature
extraction

Unsupervised

feature

selection

Unsupervised

clustering

Unsupervised

combination

2
min
T
XX
C

Consensus

clustering

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


8

6. Image data SPOT5, 5 m/pixel


(64x64 pixels)

529 samples per image

Sample extraction

©
CNES

Original images

Database of samples

Sample features

Haralick

Geometrical

QMF

Gabor

Sample 1

. . .

. . .

. . .

. . .

Sample 2

. . .

. . .

. . .

. . .

Feature extraction

and selection

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


9

6.1. Image data SPOT5, 5 m/pixel

Barcelona

Los Angeles

Istanbul

Madrid

Image examples of world cities

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


10

7. Results. Clustering by Autoclass

Image examples of world cities

Estimated number of clusters: 14

Barcelona

Los Angeles

Istanbul

Madrid

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


11

7.1. Clustering by Kernel K
-
means

Optimal number of clusters: 7

2
4
6
7
8
10
12
14
16
-5.05
-5
-4.95
-4.9
-4.85
-4.8
-4.75
-4.7
-4.65
-4.6
x 10
4
The number of clusters
MDL criterion
Kernel K-means
Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


12

7.2. Clustering by Kernel K
-
means

Image examples of world cities

Optimal number of clusters: 7

Barcelona

Los Angeles

Istanbul

Madrid

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


13

7.3. Clustering by Spectral K
-
means

Optimal number of clusters: 9

2
4
6
8
9
10
12
14
16
18
20
-2.2
-2
-1.8
-1.6
-1.4
-1.2
-1
-0.8
x 10
4
The number of clusters
MDL criterion
Spectral K-means
Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


14

7.4. Clustering by Spectral K
-
means

Image examples of world cities

Optimal number of clusters: 9

Barcelona

Los Angeles

Istanbul

Madrid

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


15

7.5. Clustering by Ward algorithm

Optimal number of clusters: 4

2
4
6
8
10
12
14
16
-3.4
-3.3
-3.2
-3.1
-3
-2.9
-2.8
x 10
4
The number of clusters
MDL criterion
Ward clustering
Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


16

7.6.
Clustering by Ward algorithm

Image examples of world cities

Optimal number of clusters: 4

Barcelona

Los Angeles

Istanbul

Madrid

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


17

7.7. Unsupervised combination

Optimal number of clusters: 12

0
5
10
12
15
20
25
30
35
40
10
5
10
6
10
7
Number of clusters
Min







2
log
T
XX
C
Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


18

7.8. Unsupervised combination

Image examples of world cities

Optimal number of clusters: 12

Barcelona

Los Angeles

Istanbul

Madrid

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


19


Conclusions


consensus solution which reduces redundant
information in clusterings


relevant unsupervised approach


interpretable clusters and their relations


generalizes

information and helps to understand an
image context



Combination of different optimal clusterings:

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


20

Perspectives

-
Unsupervised combination can be applied to data with different


clustering algorithms and metrcis, groups of
features,


classifications, segmentations,
maps,
labellings,


temporal images, etc., …


-

Using clusters for semi
-
supervised tasks




Construction of a satellite image semantic



Propose to the user images from different clusters which will be



associated with semantic terms by human interpretation.

Competence Centre on Information Extraction

and Image Understanding for Earth Observation

29/03/07

Blind city classification using aggregation of clusterings


21

Thank you for your attention!

Questions ?