Germany. It includes
| property boundaries
duse type boundaries
ters and symbols de-
boundaries.
iding
| Vector Analysis
e analysis of runlength
resembles, in this first
ch, 1991). In this con-
al image lattice is split
nal signal vectors (see
r = (r1, T2,---, rk)"
lack and white pixels,
ige Row
example of Fig.4, for
columns have to be
tched patterns to cut
at purpose a so called
lenoting the ratio be-
width of the space be-
is rotational invariant
between similar trian-
lines having the same
To handle such cases,
: investigated, whereas
ould probably lead to
> elements of the run-
senting the minimum
atched pattern, are ex-
run by run.
#1... 5-1" (1)
black pixel runs (2)
- white pixel runs (3)
na 1996
or vice versa. The above described hatching factor reads as
= bi + ba (4)
w + w2
The determination of hatched areas takes into account the
following conditions:
|b, + b2| < Tp and [wi + w2| < Tw (5)
Ty, «h«T, (6)
with Te, T,, T; and T, being empirically fixed thresholds.
The latter condition allows little variations within the amount
of the hatching factor which are caused by non-parallelism,
non-equidistances and varying linewidths. These rules (5)
are not valid for areas with a fitting hatching factor but false
dimensions of black and white runs with respect to hatched
patterns.
The resulting images I, and I. of this analysis of the run-
length encoded image rows and columns (Fig. 6b and 6c)
show that all hatched areas have roughly been detected.
Those areas which have not been recognized by the row anal-
ysis, appear for that in the result of the column analysis. Both
kinds of hatched patterns have correctly been found and most
of the other cartographic objects like boundaries and symbols
have been eliminated. Of course, undesirable small areas of
disturbances appear and the borders of buildings show ugly
fringes. The approach fails, if the numbering of the houses
interrupts the regularity of the hatching lines. For these rea-
sons further improvement of the approach is required.
original binary map
runlength encoding of rows runlength encoding of columns
subvector analysis subvector analysis
erosion erosion
logical OR
elimination of disturbances
(opening, dilation)
foreground growing of blobs
| iteration
improvements
(closing, opening)
|
building areas in raster environment
Figure 5: Flow Chart of Building Detection Process
4.2 Elimination of Disturbances
Aiming at elimination of small disturbances which are mostly
false classifications due to hatching-like structures, the two
partial results I, and I. have been eroded with a disk shaped
structuring element Ss with bounding box of 5 x 5 pixels:
In = I, O 55 , Ice Ic Ss (7)
Afterwards the two images I,» and Ic2 have been overlayed
with a logical 'or' to get all hatched areas into a single image
I with
1 = la V Les - (8)
Fig.6d shows the result. Most of the disturbances are disap-
peared, but borders of the buildings do not appear in a satis-
fying quality yet, also the numbers of the houses are missing.
However, the structure of buildings grouped along the streets
can already be discerned. An opening with the same disk
shaped structuring element as described above was applied
to eliminate remaining small errors and smooth the borders
of the buildings.
I =10 5; = (I@ 55) @Ss (9)
Because the so called 'blobs' denoting the extracted shapes
of the buildings are still smaller than the original shapes, a
dilation is carried out:
IL=1®S:s (10)
Numbers of houses and borders are still missing, but little
gaps are closed and the blob shapes correspond better to the
original shapes of the houses. (see Fig. 6e).
4.3 Foreground Growing of Blobs
At that point, the blobs have to be expanded to the outlines
of the buildings. This is done based on the knowledge that
buildings have a closed border of black foreground pixels. For
that purpose the transition from the blobs to the grey back-
ground is examined using the original image. The analysis is
carried out in 4 different directions: western, eastern, north-
ern and southern. Moving from the inner part of the blob
to its border, the first following pixel of the grey background
is investigated. If this pixel is black in the original image,
the blob will be expanded to that pixel. This process is iter-
ated until the first white pixel in the original image has been
reached. This way the blobs are growing, but only in regard
to the black pixels of the original image (see resulting image
I; in Fig. 6f).
Most of the borders and even of the numbers of the houses
can be reconstructed by this method. With respect to the
numbers it has to be pointed out that it only works if the
numbers touch a hatching line. Furthermore not all gaps have
been closed, and the expansion of the blobs also happens in
undesired areas, such as boundaries.
4.4 Improvements
Because the precedent blob expansion only affects the black
pixels, gaps remain within the blobs. These blobs can easily
be closed by a closing operation with a slightly larger disk
shaped structuring element (with bounding box of 7 x 7
pixels):
Is = I3 9 S7 = (Is ® S7) © S7 (11)
Fig. 6g illustrates that almost all gaps are closed. After that
the disturbing fringes can be eliminated by an opening op-
eration with a relatively large structuring element (bounding
box of 13 x 13 pixels):
Is = 140 815 — (I4 O S13) ® S13 (12)
The result is to be seen in Fig. 6h. At the current stage
of the process the shapes of the blobs correspond well to
the real building shapes. Only one house number has been
lost. Repeating the foreground growing step up to this point
yields a slight improvement. Fig. 6i demonstrates the final
result. Critical areas are hatching-like structures which e.g.
appear at boundaries between neighbouring but not adjacent
buildings.
85
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
