ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002
SEG(1,) contained in the partition, which were then used as
O.subs. Since this problem is sensitive to image rotation, we
used the minimum area MBR among rotated candidates, as
described in Figure 6 in determining component attributes.
Using MS(SEG(I,)), we find objects FO(SEG(I5)) in SEG(12)
and then attempt to find similar triangles in SEG(I,) and
SEG(I,) having an identical object arrangement on three
vertices. If this fails, 7; and 7, are judged not to match. In
fact, we used 3 similar triangle pairs in order to enhance the
matching accuracy, and if none of the three pairs are judged to
be similar, matching failure has occurred. Otherwise, we
determine the average translation (shift, scale, and rotation) of
I, from I; using found (at least two) similar triangles.
Finally, the two images are matched using the estimated
parameters.
dati
Figure 8. OR.ex2
Apartments in a real aerial image (Tama New Town in Tokyo) are
recognized. The small rectangle is the site used to define an NSR model.
Results and Considerations: Figures 9 (a and b) are a sample
pair of task images, which are generated by cutting two
overlapping images out of an aerial image, and applied an
artificial translation of (scaling, rotation) = (1.25, 60°) to the
image of Figure 9 (b). Figure 9 (c and d) are segmented versions.
The most similar triangle pair found is shown with their gravity
centers and the matching result is shown in Figure
9 (e). In this example, we obtained a very accurate estimation
of (scaling, rotation) = (1.24, 60° ) Ten experiments were
performed using different task images and the result of which
A - 392
are summarized in Table 1. The accuracy of scaling/rotation
parameter estimation was very high. The average error rate
was 0.51% / 1.99%.
Case Scale Rotation | S. Error | R. Error
1 0.51/0.5 45.8/45 2.0 1.8
2 0.80/0.8 51.1/50 0.0 22
3 1.60/1.6 5.4/5 0.0 8.0
4 1.25/1.25 | 310.1/310 0.8 0.0
5 2.00/2.0 311.0/315 0.0 1.3
6 0.63/0.63 | 354.9/355 0.2 0.0
7 1.98/2.0 43.4/45 1.0 3.6
8 1.24/1.25 60.0/60 0.8 0.0
9 0.63/0.63 15.4/15 0.3 2.7
10 0.50/0.5 316.0/315 0.0 0.3
Mean | ----- | ------ 0.51 1.99
TABLE 1. Accuracies Scale and Rotation
Estimated/Real data , rotations are in degrees, and errors are in 96.
4. CONCLUSIONS
We introduced the concept of the ill-configured object (ICO)
and proposed the concept of neighbor set representation (NSR)
of an object to represent the ICO. Several important properties
of NSR were clarified mathematically, especially the
possibility of characterizing an ICO (including non-ICO) as a
solution (fixed point) of a set theoretic equation of NSR of the
object. Using this property, we proposed an iterative algorithm
by which to find an ICO in an image. In addition, we reported
two applications of NSR. The first being ICO object
recognition in artificial and real images, and the second being
automatic matching of highly deviated landmark-less images.
In the former, we illustrated that ICO objects of varying
configurations can be recognized using only a small NSR
model set. In the latter, we illustrated that highly deviated
landmark-less images can be automatically matched with high
accuracy. This function provides a foundation of automatic
land cover change analysis using satellite/aerial images
obtained under different camera conditions. Future research
includes extension and applications of the NSR concept to a
wider range of media data.
REFERENCES
Watanabe, T. and Sugawa, K., and Sugihara, H., 2002.
A new pattern representation scheme using data compression.
IEEE trans. PAMI, 24(5), pp. 579-590.
Shapiro, L. G. and Haralick, R. M., 1981.
Structural descriptions and inexact matching.
IEEE trans. PAMI, 3(5), pp. 504-519.
Shapiro, L. G. and Haralick, R. M., 1982.
Organization of relational models for scene analysis.
IEEE trans. PAMI, 4(6), pp. 595-603.
Shapiro, L. G. and Haralick, R. M., 1985.
