336
synthetic data can be segmented and structured as original image
data. At image matching, synthetic data can be used independently,
combined, or merged with the original (or pre-processed) data. They
provide redundancy which can be helpful in difficult situations.
9. Image analyses
Analyses can serve for specifying (or.tuning) control parameters (to
be used in later stages), for establishing a powerful matching stra-
tegy, and for feature (or signature) extraction. Analyses can usual-
'ly be restricted to target segments.
A differentiation should be made between methods of analysis and
image properties to be analysed.
Methods can be statistical or structural. Examples of statistical
methods are clustering, principal component transformation, etc.
They do not imply external (a-priori) data, and have not been used
in photogrammetry.
Structural (or syntactic) methods are based on the context informa-
tion (in space or in time), and they can exploit external (a-priori)
information. Such methods can be based on different strategies,
1.e., implying sequential and/or parallel processing, and using
single or multiple algorithms.
Properties to be analysed are image intensity variation, its signa-
ture, and neighbourhood. Intensity variation is decisive for the
acceptance of a target segment; if not acceptable, target size
should be increased or it should be bypassed.
Image Signature refers to contrast, texture, directionality, etc.
Neighbourhood concerns properties such as connectivity and some
other ad jacency relationships.
10. Feature extraction
Feature or signature extraction implies image analysis. Techniques
vary from relatively simple to complex. The simplest variant is
extraction of image "primitives" (or tokens) such as edges, narrow
parallel bands, corner points, crossings, etc. More involved are
syntactic ("rule based") approaches, whereby primitives are assem-
bled into 'objects'. Automatic feature classification and self-lear-
ning systems, e.g., for pattern recognition, have been favourite
topics of research; nevertheless, success to date has been limited.
For a two-stage matching strategy (figure 2), however, extraction of
a few primitives (e.g. edges, corner points) suffices.
1l. Integration of external information
If properly selected and integrated, external (a-priori) information
contributes to accuracy and reliability of matching and thus of the
DTM. Such information also provides a reference for assessment of
performance.
External information refers to supplementary data (e.g., mean height
of trees or houses), geometric conditions, constraints and criteria,
and control and check data. Geometric conditions concern lines
(straight, curved, parallel), planes (horizontal, tilted), angles
337