n-
te
lal
er
es
ly,
he
e~
li-
on
or
er
ur
. Gray-scale Spatial
Preprocessing Correction Filtering
Feature Statistical Interest Edge Texture,
Extraction Properties Points Extraction Segmentation
(rsage Patch
Analysis)
Matching
2 Method Cross LS Matching V — S (3D) Symbolic
Correlation Feature Based Matching
- Strategy Line Global Surface
Constraint Approximation|
Evaluation Results, Confidence
a Ts Analysis
|
images are not resampled but converted on the fly during
display operation or data processing. Usually, the operator
moves the floating mark quite slowly. Therefore, the im-
age patches of a current matching operation may have very
similar image characteristics like neighboring patches. This
means that certain parameters of the previous (neighbor-
ing) matching operation can immediately be used for the
processing strategy of the current patches. For example,
surface direction can be approximated at one side of the
patch, or texture based segment data and basic statistics
can reveal occlusion situations.
The patch size in our application is more an implementa-
tional than algorithmic issue. However, it is still important
since most matching methods are very timeconsuming and
our application needs a quasi real-time response (the pro-
cessing time should not take longer than what an operator
would need).
2.2 Structure of the Algorithm
Based on conditions imposed by our application the default
matching method is cross correlation (Ackermann, 1984). If
the current patch has enough texture information, the fore-
shortening is negligible and there is no occlusion or other
artifact, then correlation performs well. Since these condi-
tions are not always met, other methods must be used. A
first key issue is to find and parametrize the image charac-
teristics (called actual features in this paper) of the patches.
Figure 1.
Flowchart of the DOG system.
401
This problem itself is as complex as the matching, because
ideally it would address many high-level paradigms of scene
analysis and image interpretation. Because of the lack of a
robust, scene independent matching scheme, an iterative hi-
erarchical strategy is proposed. Figure 1 shows the flowchart
of the proposed DOG method. The suggested system has
five processing levels:
e Patch preprocessing
e Feature extraction - image patch analysis
e Matching procedure
e Matching strategy
Evaluation - result analysis
The patches may be subject to some image enhancement.
Scaling the pixel intensities or histogram equalization can
compensate for bad contrast. Spatial filtering, like median
or Gauss operators can remove noise or unneccessary details
which may be important for scale-space algorithms.
Feature extraction provides clues about the patch and may
guide the selection of the most appropriate matching method.
The basic statistical properties, like mean, median, mini-
mum and maximum intensities, standard deviation, auto-
correlation, etc. provide additional patch information.