ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
component, as shown in Figure 6. The shape type is defined so
as to reflect the area and aspect ratio of the component. In order
to enhance the precision, we enclose the original component by
two MBRs that are rotated by 0° (non-rotation) and 45°, and
select the MBR having the lesser area (Figure 6: a, b). The area
and aspect ratio of the chosen MBR is used to define the
component shape type (Figure 6: c). By using this extended
attribute in component identification, we can extend the
applicability of MS to objects composed of components with
changed size. The problem of component deletion/addition is
addressed by introducing a likelihood measure between the
model O and the found object FO. More specifically, let
O.subs and FO.subs be their components, respectively,
classified by their attributes. Then, we introduce the following
likelihood function to measure their similarity:
L(FO, O) = (FO.subs NO.subs)/(FO.subs UO.subs).
Note that L(FO, O) 21.0 when FO has components identical to
those of O.subs. This function can be refined by adding a
measure of area similarity between O.area and FO.area. An
example refinement LL is (using | FO | to denote the area of
FO, etc):
LL(FO, O, FO.a.s, O.a.s)=
L(FO, O)x min (| FO |, | O |)/max (| FO |,| O |).
Note that ZZ(FO,O, FO. a. s, O. a. s)=1.0 when both FO and O
have identical components and sizes.
3. APPLICATIONS
3.1 Object Recognition
Purpose: ICO recognition in artificial and real images is
performed in order to verify that the NSR can find highly
varying ICOs using only a few models.
Method: We perform two experiments, OR.ex1, and OR.ex2. In
OR.exl, an artificial color map / contains three types of
ICOs: O,. attriv = stadium, O, . attriv = park, O5. attriv = apartment
each having four instances of varying configurations (Figure 7).
Image segmentation is performed using pixel color classification
to yield SEG(/). Uppermost objects are used to build the NSR
model set MS” = MSUrot(90°, MS), where MS is the topmost
three objects in the figure and ro:(90°, MS) is a rotated version
of MS So the model set MS” contains two models for each
ICO. Assuming that every configuration of components is
possible and O.area.shape varies, the
expanded MBR™ (c, O.subs) of the maximum extension is used
(see Figure 2 (h, i, and j) and Figure 4).
In OR.ex2, apartment recognition is performed using an actual
aerial image. The small window in Figure 8 shows the site used
to define MS. Including the rotated variant, MS" contains two
models (MS* = MS U rot(90°, MS)).
Results and Considerations: Figure 7 shows the result of
OR.exl. All 12 ICOs were successfully recognized using only
two models for each ICO. In addition, for OR.ex2, Figure 8
shows that although not completely free of recognition failure,
most of the visually recognizable apartments could be
recognized automatically using only two models.
A - 391
1 wir him For on
Figure 7. OR.ex1
Artificial map including 12 ICOs (stadiums, parks, and apartments).
Uppermost three ICOs are used to define NSR models. All 12 ICOs are
successfully recognized. For each ICO, the MBR covering O. subs is
shown as the estimated O.area. Components not in O.subs might
hang out of the MBR.
3.2 Automatic Matching of Highly Deviated Landmark-
less Images
Purpose: In a number of applications of aerial/satellite image
analysis, the image matching function is fundamental. Land
cover change monitoring, hazard map generation, and map
revision are only a few examples. Since adjustment of camera
conditions, such as the height and the direction is very difficult
when the images are taken on different occasions, we are given
two images of differing shift, scale, and rotation. Moreover,
landmarks that are usable in image matching are not usually
provided in normal images. Therefore, the problem of
automatic matching of two highly deviated landmark-less
images must be solved in order to automate these tasks. We
examine the applicability of NSR to this open problem.
Method Two images, I; and 75, can be automatically
matched using NSR. First, we automatically extract NSR
models MS(SEG(I;)) out of SEG(I,) and find them in
SEG(15). Notice that we must find varying (shifted, scaled,
and rotated) ICOs out of SEG(I5) using the model extracted
from SEG(I,). We then try to define similar triangles in
SEG(I,) and SEG(I,) using modeled and found objects,
respectively. In generating MS(SEG(I,), we introduce a
regular grid in SEG(/;) and generate one NSR object out of
each partition. We choose a fixed number of components (we
used three components having larger areas) out of
