y
és,
PICTURE CLASSIFICATION AND SEGMENTATION BY FRATURE COMBINATION
Ta IN MULTISPECTRAL DATA
TE. Mauer, R. Schürf
Forschungsinstitut für Informationsverarbeitung
und Mustererkennung
Breslauer Straße 48, D-7500 Karlsruhe 1
ABSTRACT
The principal methods for image classification and image
segmentation, which have been developped until now, work with
a separate use of multispectral and textural properties. However
those methods leads to satisfactory results only in some specific
domains or in image which are mainly composed by uniform types
of either spectral or textural objects.
Invenstigations show that some methods have complementary
properties and are able to support one another. Considering the
performance of classification in more general applications, the
combination of such complementary methods promises an essential
improvement in picture processing.
As a result of the comparative investigations between the
different methods a proposal of a processing system is described.
This segmentation and classification system processes real images
without a priori knowledge combining the complementary methods.
A possible application is the production of thematic maps to
planning and ecologic aims, for example by evaluation of image
data from satellites. :
INTRODUCTION
Most of the developped methods for image analysis, segmentation
and automatic classification evaluate either multispectral
features or textural properties, but not simultaneously. The
separate use of one of these methods is successful at special
problems of picture processing and for images which are composed
of uniform types of objects. Most of the real images contain both
types of objects together. Considering the performance of
classification in more general applications, the results of the
separate use of oniy one of these methods are not satisfactory.
Therefore the combination of both methods promises an essential
improvement in picture processing.
This paper describes the advantages and shortcommings of the
multispectral analysis and the texture analysis. It will be
demonstrated that the multispectral image processing is able to
support the texture analysis, and a combination of these methods
is proposed. In order to simplify such a combination, the train-
35
A MENU Le ll a ln MERERI