, either in the spatial 
) be an inappropriate 
]e 1995]. Map sym- 
ted resolution of the 
ng rotation and scale 
symbols contain less 
; to filtering by image 
been made with a 
he computational re- 
process impractical. 
d successfully using 
son 1994, Kass 1987, 
:d to calculate poten- 
and dilatation are in- 
ecause of the limited 
ation from the ,origi- 
»d rule definition un- 
à À 
5c 
n point symbols 
  
iccording to the values 
classification methods 
features are often cor- 
a non-orthogonal n- 
appropriate for the re- 
ion reduction methods 
tion are not applicable 
y units. 
yy a set of rules based 
finition process is the 
reby the user iterative- 
)blem. 
3a 1996 
  
ID 63 SXx408 SY 605 
  
  
  
  
Area 12 Perimeter 9 Holes O 
Eccentricity 1.05 Circularity 1.86 
Elongation 0.12 Alpha 32.07 
Spreadness 0.02 M11 28 
M20 129 M02 10.3 
[. ID 64  SX410 SY 627 
Area 35 Perimeter 20 Holes 1 
Eccentricity -0.12 Circularity 1.10 
Elongation 0.03 Alpha 10.58 
Spreadness 0.01 M11 -1.9 
M20 165.5 M02 175.1 
  
ID 74 SX419 SY 488 
  
Area 47 Perimeter 30 Holes 1 
Eccentricity 1.16 Circularity 0.66 
Elongation 0.09 Alpha 6.72 
Spreadness 0.04 M11 11.2 
M20 493.3 M02 587.0 
Figure 3: image symbol database 
Show (every Record whose Cell "Holes" = 1 ^ 
and Cell "Area" » 100 and Cell "Area" « 200^ 
and Cell "Circularity" « 0.45 ^ 
and Cell "Spreadness" 201" 
and Cell "Compactness" » 30.0 ^ 
and Cell "FPO" » 2.8 and Cell "FPO" « 2.9 ^ 
and Cell "FP1" » 0.01 and Cell "FP1" « 0.2 ^ 
and Cell "FP2" » 0.9 and Cell "FP2" « 1.0 ^ 
and Cell "Absm11" « 200) 
FPO-FP2: Power spectrum values of 
fourier descriptors 
Absm!1 1:1 M, I where Mj; = X (x - x9 - yo) 
xp»Yg= center of gravity 
Table 1: Discrimination parameters for observation towers 
In an interactive training phase helped by an image symbol 
database (Figure 3), the symbol being sought is characterised 
by defining selection rules based on the identified shape de- 
scriptors. The user identifies parts belonging to a composed 
map symbol and setsup the different discrimination values for 
the symbol:In an interactive training phase helped by an 
image symbol database (Figure 3), the symbol being sought is 
characterised by defining selection rules based on the identi- 
fied shape descriptors. The user identifies parts belongingto a 
composed map symbol and sets up the different discrimination 
values for the symbol: 
In an interactive training phase helped by an image symbol 
database (Figure 3), the symbol being sought is characterised 
by defining selection rules based on the identified shape de- 
scriptors. The user identifies parts belonging to a composed 
map symbol and setsup the different discrimination values for 
the symbol: 
International 
The objects are then classified either as candidate objects or 
rejected according to specific characteristics. 
Depending on their complexity, certain symbols (e.g. triangu- 
lation points) can be detected in one pass, whereas aggregated 
symbols such as tree groups or avalanche obstructions need 
multiple passes. Fourier descriptors of the contour line have 
been proven to be very powerful [Lai 1994, Staib 1992, 
Udomkesmalee 1991]. For the multi-pass case, the rules do 
not need to produce a ,,perfect match because the matched 
objects represent only candidate symbols and the subsequent 
triangulation enables a better discrimination than at the single 
object level. 
After the shape discrimination, point or line symbols remain 
difficult to distinguish from similar background objects (see 
Figure 1b). To ensure to detect ,true" line symbols, a local 
Hough transformation will be applied for every line symbol 
candidate [Chang 1994, Palmer 1993]. Because only the en- 
closing boundary box of each symbol will be used for the 
Hough detection, potential performance problems are min- 
imised [Han 1994]. 
4. Triangulation 
To model the spatial distribution of complex map symbols, all 
candidate objects will be Delaunay-triangulated [Sedgewick 
1992] according to the minimum distance criteria and build 
the base triangle level (Figure 4). Starting from the seed trian- 
gle (typically the three nearest candidate points), each triangle 
side is the basis for the next possible triangle. The triangula- 
tion stops when no more points are found within a specified 
distance. Using the centre of gravity of the identified triangles, 
the next higher degree of triangulation are built until less than 
three centre points remain (Figures 4 and 5). The triangulation 
levels build a data structure called a tetra-tree (the full struc- 
ture is similar to a tetrahedron). The top triangle (level 3 in 
Figure 4) defines the generalised direction and centre of gravi- 
ty of the whole object. The convex hull of the base level trian- 
gles models the surrounding polygon. 
For avalanche obstacles (Figure 4), at least two levels of high- 
er aggregation must be reached, so that a potential obstacle 
can be defined as a recognised map symbol. 
In Figure 5, the triangulations for tree groups reflectthe gener- 
alisation effect of the higher triangulation levels. The triangles 
marked by ,,X“ in both figures show triangles which were re- 
jected because 
e atleast one point is too far away 
* the smallest angle is below the minimal angle value 
Triangulation can also be used to detect dry channels (see 
Figure 6), but instead of rejecting thin triangles, only extreme- 
ly elongated or even collinear triangles are admitted. In fact, 
we are looking for the reverse of the avalanche obstacles or 
tree groups. 
Within noisy background (see Figure 1b), the triangulation 
yields too many ,,proper" triangles and the hidden dry-channel 
symbol cannot be detected. 
Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996