An architecture based on class dependent neural networks for object-based classification

An architecture based on class dependent neuralnetworks for object-based classification

F. BachmannI. Niemeyer, P. Marpu

Institute for Mine-Surveying and GeodesyTechnische Universität Bergakademie Freiberg

8. April 2009

Object-base Image Analysis

Preprocessing Object Recognition Feature Extraction Classification

Feature Extraction

Intrinsic/sensorical object featuresI SpectrumI ShapeI Texture

Semantic object featuresI Intrinsic relationship between objectsI Relationships out of object organisation

⇒Multitude of features in feature space



Classification

I Assignment of meaning to spatial objectsI Creation of a conceptual scheme / OntologyI Classification (conceptual scheme ↔ classifier) → classify

How about automated classifiers?



Classification

I Assignment of meaning to spatial objectsI Creation of a conceptual scheme / OntologyI Classification (conceptual scheme ↔ classifier) → classify

How about automated classifiers?

Classification - Neural Networks

2-layered feed forward network

1 //WVUTPQRS0

''OOOOOOOOOOOOOOOOO

��@@@

@@@@

@@@@

@@@@

@@@@

@@@@

��555

5555

5555

5555

5555

5555

5555

555

Wh1

_

_��

1 //WVUTPQRS0

''OOOOOOOOOOOOOOOOO

��@@@

@@@@

@@@@

@@@@

@@@@

@@@@

��555

5555

5555

5555

5555

5555

5555

555

Wo

_

_��

g1(v) //WVUTPQRS1 //

''OOOOOOOOOOOOOOOOO

��@@@

@@@@

@@@@

@@@@

@@@@

@@@@

WVUTPQRS1 //

''OOOOOOOOOOOOOOOOO

��@@@

@@@@

@@@@

@@@@

@@@@

@@@@

WVUTPQRS1 // m1(v)

......

...

gi(v) //WVUTPQRSi //

77ooooooooooooooooo

''OOOOOOOOOOOOOOOOOWVUTPQRSj //

77ooooooooooooooooo

''OOOOOOOOOOOOOOOOOWVUTPQRSk // mk(v)

......

...

gN(v) // WVUTPQRSN //

77ooooooooooooooooo

??~~~~~~~~~~~~~~~~~~~~~~~ WVUTPQRSL //

77ooooooooooooooooo

??~~~~~~~~~~~~~~~~~~~~~~~ _^]\XYZ[K // mM (v)

Input1-th hidden

layeroutputlayer Output


Training of the classifier:I BackpropagationI Kalman Filter-trainingI Scaled-gradient Conjugated

Modification of network:I Class against classes →’is’/’is not’ class stateI Classification of ’is’/ ’is not’ class state


Training of the classifier:I BackpropagationI Kalman Filter-trainingI Scaled-gradient Conjugated

Modification of network:I Class against classes →’is’/’is not’ class stateI Classification of ’is’/ ’is not’ class state


’is’/’is not’ class feed forward network → Class discrimination

1 //WVUTPQRS0

((QQQQQQQQQQQQQQQQQQQQ

!!CCC

CCCC

CCCC

CCCC

CCCC

CCCC

CC

��777

7777

7777

7777

7777

7777

7777

7777

77Wh1

_

_��

1 //WVUTPQRS0


!!CCC

CCCC

CCCC

CCCC

CCCC

CCCC

CCWo

_

_��

g1(v) //WVUTPQRS1 //


!!CCC

CCCC

CCCC

CCCC

CCCC

CCCC

CCWVUTPQRS1 //

((QQQQQQQQQQQQQQQQQQQQWVUTPQRS1 // g(v) is class

......

...

gi(v) //WVUTPQRSi //

66mmmmmmmmmmmmmmmmmmmm

((QQQQQQQQQQQQQQQQQQQQWVUTPQRSj //

66mmmmmmmmmmmmmmmmmmmm WVUTPQRS2 // g(v) is notclass

......

gN(v) // WVUTPQRSN //

66mmmmmmmmmmmmmmmmmmm

=={{{{{{{{{{{{{{{{{{{{{{{{{ WVUTPQRSL

66mmmmmmmmmmmmmmmmmmmm

=={{{{{{{{{{{{{{{{{{{{{{{{{

Input1-th hidden

layeroutputlayer Output


Building a Class Dependent Network (CDN)

Neural Network Architecture describing a Semantic Ruleset for Object-Based Classification

Feature 1

class associative membership-function modeling

Class Ruleset mapping (omnivariante) decision-tree based Neural Network

generalized Feature Signal

Is class

Is not class

...

Feature 2Is class

Is not class

......

Feature NIs class

Is not class

...

select valuable Inputs according to separability

... ...

Is class

Is not class

...

...

Is class

Is not class... ...

Is class 1

Is not any class

Is class N

build Class Hierarchy Neural Network

...

?

...

Classification

Problems of controlled classifiersI Number of samples, adequate samplesI Feature reduction, separabilityI Non-unique result of classification (especial problem of FFNs),

error propagationI Intransparent classificationI Semantics of classes, still manual ’fuzzy’ rules

Sample Image (Aster/Juelich)

’is’ class (open pit): fuzzy values

’is not’ class (open pit): fuzzy values

Feed Forward Network

Class dependent Network

Classification

I Identification of good/bad classes by ’is’/ ’is not’ stateI Flexible with regard to class hierarchy and integration into the

usual classification processI Transferability of rulesets in time and space by retraining

existing networks.

Implementation

Available for Definiens Developer

http://tu-freiberg.de/fakult3/mage/geomonitoring/software/software.html

http://tu-freiberg.de/fakult3/mage/geomonitoring/software/software.html

ReferencesE. P. Baltsavias.Object extraction and revision by image analysis using existing geodata and knowledge: current status andsteps towards operational systems.ISPRS Journal of Photogrammetry and Remote Sensing, 58(3-4):129 – 151, 2004.Integration of Geodata and Imagery for Automated Refinement and Update of Spatial Databases.

T. Blaschke, S. Lang, E. Lorup, J. Strobl, and P. Zeil.Object-oriented image processing in an integrated GIS/remote sensing environment and perspectives forenvironmental applications.Environmental Information for Planning, Politics and the Public, 2:555–570, 2000.

GJ Hay and G. Castilla.Object-based image analysis: strengths, weaknesses, opportunities and threats (SWOT).International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences vol. XXXVI-4/C, 2006.

Nikhil R. Pal and Sankar K. Pal.A review on image segmentation techniques.Pattern Recognition, 26(9):1277 – 1294, 1993.

B. Tso and P.M. Mather.Classification Methods for Remotely Sensed Data.CRC Press, 2001.

A. Tzotsos.A SUPPORT VECTOR MACHINE APPROACH FOR OBJECT BASED IMAGE ANALYSIS.In 1 stInternational Conference on Object-based Image Analysis, http://www. commission4. isprs.org/obia06/index. html.

S. Ullman.High-Level Vision: Object Recognition and Visual Cognition.Mit Pr, 1996.

An architecture based on class dependent neural networks for object-based classification

Documents