at
Fig. 4.2. segmentation of the cutout
Fig. 4.3 determined corners of several regions in
Fig. 4.2
5. DISCUSSION
With emphasis on the knowledge representation,
we have presented our idea to develop a
knowledge based system for scene analysis and
unterstanding in the above sections. Still we can't
expect that machine vision problem can be solved
without great difficulties, because what ES and AI
methods offer us are nothing else but programming
tools. Nevertheless, we can better modularize
knowledge and reasoning by using these methods,
so that they can be kept easy to read and modify,
and easy to take distributed and parallel
processing. The above object-oriented description
makes local and structural image analysis possible,
which is highly important because of limitations of
global and statistical analysis. Two crucial and
most difficult problems remain unsloved: a learning
mechanism for acquiring knowledge and
transforming it into internal representation; a
reasoning mechanism for finding out and correcting
segmentation errors and for recognizing objects.
224
In this context the goal of learning is to fill the KB
with suitably structured knowledge about objects
automatically or semiautomatically. A deductive
learning mechanism is already enabled by the
suggested object-oriented description: description
structures of individual objects are first analysed
and worked out, afterwards are generated
interactively as external descriptions, which are
finally converted into internal forms by the program
for external-internal transformation. The KB can
also be set up inductively (i.e.example-driven): the
membership hierarchy is first gained through an
available interface to GIS, with which the class
hierarchy is then richened and extended, e.g.
adding new object species. The same can be done
by using interpretation results from the DB or both
the results and the membership hierarchy. For
human examination and check an internal-external
transformation can be executed.
The reasoning for object recognition is realized in
three steps, which should be supported by a
number of algorithms for geometric and relational
object analysis in different levels:
- focusing: preclassify regions in the DB to find
candidates for certain objects in order to improve
the efficiency.
- backtracking through resegmentation: detect
segmentation errors and correct them. A complete
and perfect segmentation would greatly simplify the
followed HLP. Unfortunately for remotely sensed
data it has proved to be tremendously difficult and
even impossible owing to insufficient information
during segmentation and the difference between
human perception and mathematical algorithms.
For this problem there exist two solving methods.
One of them emphasizes that semantics should be
introduced into segmentation so as to guide and
improve it, while in the other segmentation errors
are expected to be found out in HLP and then
corrected by backtracking. The former integrates
the bottom-up and top-down strategies together and
is restricted to certain applications owing to the
early introduction of semantics; the latter uses first
the bottom-up and then the top-down strategy and
is adopted here.
- object recognition through reasoning: about
handling knowledge uncertainty there are two
reasoning strategies: inexact reasoning and
approximate reasoning which is based upon fuzzy-
set theory and mathematically sound. In many
model-based vision systems the problem is
indirectly handled by a more or less exact
reasoning, where uncertainty is implicitly contained
in the knowledge body and reasoning. The difficulty
of a explicit treating lies in determining a suitable
certainty measure for facts and rules. For efficient
verification of a fact backward chaining is very
helpful: a hypothesis is first generated with help of
available information and then verified. In addition,
the
the
bot
pro
WO
RE
Bat
Ler
Ext
Any
uno
Bro
Din
Ma:
19€
Cor
Inte
Ser
roc
Rer
Cor
Fu,
Inte
Gin
Lan
Dis
Dire
Car
Ma
Ace
