A iS
i_polygo0168 ||i_polygo112 || i-Polvgo224 |
art art art
i_line0002
i_line0013
i_line0020
i_line0027
i_line0118 i_lLine0034 i_line0174
i_line0125 i_line0041 i_line0181
i_line0132 i_line0048 i_line0188
i_line0139 i_line0055 i_line0195
i_line0146 i_line0062 i_line0202
i_line0153 i_line0069 i_line0209
i lineoieo i_line0076 i_line0216
i_line0167 i_line0083 i_line0223
i_line0090
i_line0097
i_line0104
i lineO1i1i1i
2
i segment
Figure 2: Detail of the part-of hierarchy of the specific model
2.2 Model building
In the second phase, the scene description obtained after the
map analysis is combined with the generic model in the image
domain and the specific model in the image domain is built.
A detail of this specific model representing building 0235 and
its parts as far as they are given in the map, is shown in Fig. 2.
For each node (instance) in the scene description we create a
new node (concept) in the specific model. This new concept
is a specialization of the corresponding concept in the generic
model in the image domain and thus inherits its declarative
and procedural knowledge. The values of the attributes in
the scene description after map analysis are stored after a
transformation as restrictions for the corresponding attributes
of the newly created concepts. They serve as initial estimates
for the calculation of the attribute values out of the image
data.
The relations between the instances in the scene description
are transfered accordingly into relations between the new con-
cepts. Whilst the generic model in the image domain de-
scribes in a general form the representation of an arbitrary
scene in an aerial image, the specific model in the image do-
main describes in a detailed manner that part of the world,
which is subject to the current analysis. The grade of detail
depends of course from the contents of the map.
2.3 Image primitives
Prior to the model based image analysis primitives are ex-
tracted from the image data. We work with large scale color
aerial images, which after digitization have a pixel size of 30
cm x 30 cm on the ground. Line segments and regions serve
as primitives. The line segments are extracted with a gradi-
ent based procedure (Quint and Bahr, 1994). The regions
are gained by segmenting the aerial image using a Bayesian
homogeneity predicate (Quint and Landes, 1996).
The regions and the line segments are combined in an at-
670
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
tributed undirected graph. The nodes of the graph are at-
tributed with the regions. Nodes corresponding to neigh-
bouring regions are connected with links. A link between two
nodes is attributed with the line segment(s) which compose
the border between the corresponding regions. This feature
graph is the database on which the model based image anal-
ysis operates.
2.4 Image analysis
In the third phase, the specific model in the image domain
is used to perform the actual image analysis. The aim of
this phase is to verify in the image the objects found after
the map analysis and to detect and describe other objects
of the scene which are not represented in the map. For the
later, the context gained through the verification of the map
objects will be helpful.
The strategy followed in the analysis process is a general,
problem independent strategy provided by the shell ERNEST.
The analysis starts by creating a modified concept for the goal
concept (expansion step). A modified concept is a preliminary
result and it reflects constraints for the concept that have
been determinated out of the context of the current analysis
state.
Following top-down the hierarchy in the semantic network,
stepwise the concepts on lower hierarchical levels are ex-
panded until a concept on the lowest level is reached. Since
this concept does not depend from other concepts, its corre-
spondence with a primitive in the database can be established
and its attributes can be calculated. This is called instantia-
tion.
Analysis now moves bottom-up to the concept at the next
higher hierarchical level. If instances have been found for all
parts of this concept, the concept itself can be instantiated.
Otherwise the analysis continues with the next not yet instan-
tiated concept on a lower level. After an instantiation, the
acquired knowledge is propagated bottom-up and top-down
to impose constraints and restrict the search space. Thus, in
the analysis process top-down and bottom-up processing al-
ternate. As well, expansion and instantiation alternate during
the analysis.
Generally, while performing an instantiation it is possible
to establish several correspondences between a concept and
primitives in the data base. However, only one of these cor-
respondences leads to the correct interpretation. Since it
usually is not possible to ultimately decide at the lower levels
which correspondence is correct, all possible correspondences
have to be accounted for.
Thus, the image analysis is a search process, which can be
graphically represented by a tree. Each node of the tree repre-
sents a state of the analysis process. If in a given state several
correspondences are possible, the search tree is splitted: for
each hypothesis a new node as successor of the current node
is created.
The analysis process continues with that leaf node of the
search tree which is considered to be the best according to a
problem dependent evaluation. It is know that the problem
of finding an optimal path in a search tree can be solved by
the A*-algorithm (Nilsson, 1982). Its application is possible
if one can evaluate the path from the root node to the current
node and if one can give an estimate for the valuation of the
path from the current node to the (not yet known) terminal
node containing the solution.
Fi
The f
very i
efficie
succe:
of the
the gi
is calc
modif
valuat
be cre
node.
When
of ma
ation
can n
is per
uncer
level
choto
primit
o
to the
The v.
cepts
jective
0 and
frame
1976)
tive b
ods d
combi
netwo
We &
comp:
ates a
groupi
radion
categc
line se
of an
polyge
polyg:
Then
image
