×

You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Title
Fusion of sensor data, knowledge sources and algorithms for extraction and classification of topographic objects
Author
Baltsavias, Emmanuel P.

International Archives of Photogrammetry and Remote Sensing,Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999
SCALE CHARACTERISTICS OF LOCAL AUTOCOVARIANCES FOR TEXTURE SEGMENTATION
Annett Faber, Wolfgang Forstner
Institute of Photogrammetry, University Bonn, NuBallee 15, D-53115 Bonn, Germany, email: (annett,wf)@ipb.uni-bonn.de
KEY WORDS: texture representation, Laplacian pyramid, spectral decomposition, eigenvalues, urban structure.
ABSTRACT
This paper describes research on the extraction of urban structures from aerial images or high resolution satellite scenes
like the German MOMS02 sensor. We aim at a separation of neighboring textured regions important for describing different
urban structures. It has been shown, that grey level segmentation alone is not sufficient to solve this problem. If we want
to use texture additionally, the development of a suitable representation of texture images is required. We use the scale
characteristics of the local autocovariance function, called SCAF, of the possibly multiband image function. The final result
of the process are texture edges.
1 INTRODUCTION
Texture is one of the most fundamental and at the same
time most interesting characteristics of visible surfaces in
the human perception process. Therefore, in pattern recog
nition the analysis of textures is very important. Numerous
techniques for texture analysis have been proposed. They
can be mainly categorized as (Haralick and Shapiro, 1992):
1. Texture classification: For a given textured region, de
cide, to which one out of a finite number of classes the
region belongs.
2. Texture synthesis: For a given textured region, deter
mine a description or a model.
3. Texture segmentation: In a given image, which con
tains many textured regions, determine the boundaries
between these regions.
The procedures developed for the solution of these prob
lems can be subdivided into three categories: structural ap
proaches, statistical approaches and filterbased approaches
(J.-R de Beuaville and Langlais, 1994, Reed and du Buf,
1993, Shao and Forstner, 1994).
Structural approaches assume that textures contain de
tectable primitive elements which generate the texture
in a regular or irregular manner, following certain con
nection rules. Such approaches use e. g. so-called
texture grammars (Carlucci, 1976) or texture elements
(Julez and Bergen, 1983).
Due to the almost unlimited number of possible unit
patterns and the complexity of the rules, these proce
dures have shown lower success in texture analysis
than the following two approaches.
Statistical approaches use statistical characteristics, deriv
able from the images. These characteristics are often
sufficient for texture classification and segmentation,
without needing generation rules for the textures. They
can further be separated into approaches which use
• descriptions derivable directly from the images
such as variance, entropy or other values ob
tained from the local pixel neighborhood, as e. g.
co-occurrences (Haralick, 1979) and
• model based descriptions, such as autoregres
sive models, Markov or Gibbs random fields (An-
drey and Tarroux, 1996, Derin and Cole, 1986).
Statistical approaches usually refer to the image grid,
thus have difficulties in handling scale space proper
ties.
Filter based approaches assume that the image function
can be described locally by its amplitude spectra. Ga
bor wavelets have been the first choice together with
linear and nonlinear post processing steps to achieve
multi scale features in the different channels (A. C. Bovik
and Geisler, 1990, Bigun and du Buf, 1992, Malik and
Perona, 1990). This class of approaches is motivated
by their similarity to the human visual system.
The advantage is that the filter responses for basic ge
ometrical transformations are predictable and the fil
ters work equally for natural scenes of different texture
types. However, there is no general approach for the
selection of a suitable filter bank and for the linkage of
different image channels.
2 MOTIVATION
We are interested in analyzing satellite and aerial images
especially for extracting urban structures. Therefore, we
need powerful techniques for image segmentation in the
presence of natural textures. Grey level or color segmen
tation results often suffer from over or under segmentation.
Texture segmentation can reduce the amount of over seg
mentation, mostly without the risk of under-segmentation.
Our first experiences with Gabor wavelets based on (Ma
lik and Perona, 1990) demonstrated the feasibility of a fil
ter based approach for texture edge extraction (Shao and
Forstner, 1994). Our approach used the edge detection
scheme in ourfeature extraction program FEX (Fuchs, 1998),
cf. Figure 1c). It exploits its ability to handle multichannel
images by taking the filter responses as a multichannel im
age as input.
Due to the heavy algorithmic complexity, we develop a new
filter based scheme for deriving texture edges again us
ing the edge extraction scheme of our feature extraction