CLOSING: OPENING:
fb dad
[3511 3
DILATION EROSION
Y Y
|
EROSION DILATION
WITH STRUCTURING ELEMENT:
Figure 2: Mathematical Morphology Operators
useful to fill little gaps between border fringes (Fig. 2).
4 DEVELOPED APPROACH AND RESULTS
In this paper we concentrate on the first location step, the
detection of hatched areas of the map in the raster environ-
ment. For fulfilling the introduced pattern recognition task
it is important to choose rotational invariant features taking
into account all possible orientations of the hatched patterns.
For that reason simple template matching methods, as e.g.
(Stengele, 1995) uses for the recognition of characters and
symbols in a swiss topographic map 1:25000, have to be ruled
out because they would require the laborious rotation of the
templates in multiple directions.
An eyecatching feature of hatched patterns is their typical
periodical sequence of black and white pixels with an ap-
proximate constant ratio between the numbers of successive
black and white pixels. This striking feature is used in the
recognition process.
A subset of the scanned DGK5 sheet 'Karlsruhe Weststadt'
has been chosen (see Fig. 6a) to demonstrate the results of
the developed raster based approach on a real map example.
The subset was scanned with a resolution of 450 dpi and con-
tains approximately 200 x 300 pixels. A threshold operation
was applied to transform the scanned map data, originally
ranging of 255 grayvalues, into a binary image.
The subset represents a residential neighbourhood which is
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
typical for areas close to city centres in Germany. It includes
residential buildings, outbuildings, and property boundaries
surrounding courtyards, as well as landuse type boundaries
marking streets and pavement, characters and symbols de-
picting single trees and cross points of boundaries.
building
column
Figure 3: Directions of Runlength Vector Analysis
4.1 Detection of Hatched Areas
The developed approach is based on the analysis of runlength
encoded image rows and columns and resembles, in this first
step, an approach of (Shen/Ebi/Besslich, 1991). In this con-
nection, the analysis of the 2-dimensional image lattice is split
up into the analysis of two 1-dimensional signal vectors (see
Fig. 3). A runlength encoded vector r = (riT2, TR)
contains the number r; of successive black and white pixels,
respectively.
Figure 4: Part of an Image Row
The runlength encoded vector to the example of Fig.4, for
instance, reads as 7 = (3, 5, 3, 6, 4)".
These runlength encoded rows and columns have to be
searched for areas which contain hatched patterns to cut
them out of the original image. For that purpose a so called
hatching-factor h has been created denoting the ratio be-
tween width of the hatching line and width of the space be-
tween hatching lines. This factor h is rotational invariant
because of the well-known relations between similar trian-
gles. The only problem are hatching lines having the same
direction as the runlength-vector r . To handle such cases,
at least one second direction has to be investigated, whereas
more than one additional direction would probably lead to
more robust results.
Subvectors f containing 4 successive elements of the run-
length encoded vectors r and representing the minimum
number of hatching lines defining a hatched pattern, are ex-
amined in sequentially moving ahead run by run.
F = (ri, Ti-+1, Ti+2, ri43). ; Wm1,... ,k—1. (1)
With
(ri, ri+2) = (bi, ba)"
(rii, Ties) =" (wi, wa)"
black pixel runs (2)
white pixel runs (3)
84
or vice ver:
The deterr
following c
with Tp, 1
The latter
of the hat
non-equidi
are not val
dimension:
patterns.
The result
length enc
show that
Those are:
ysis, appe:
kinds of h:
of the oth:
have been
disturbance
fringes. T
interrupts
sons furth
runle
sul
Figur
4.2 Elir
Aiming a
false clas
partial re:
structurir
Afterwarc
with a loi
I with
