:AKLINES 
hm, empha- 
e number of 
in idea was 
oy detecting 
g breaklines 
sparser net- 
| or squares 
e studied by 
gorithms for 
s. 
jes 
the mathe- 
without con- 
ts are used 
.. Detection 
g difference 
/., 1993) for 
omatic digi- 
this kind of 
ioimage per 
?e input im- 
e radiomet- 
. Large dif- 
> be caused 
(Figure 3) 
. The scale 
1:15 000 
area varied 
| operators, 
1 operators 
size (ISEF) 
eakline de- 
els. Due to 
efore using 
-F-operator 
med at the 
e results of 
-F-operator 
klines were 
t breakline 
fthe image 
nages than 
on the right 
significant 
om the left 
ding of the 
ed. 
elling 
omplicated 
s handled. 
one) break- 
e mesh and 
s of break- 
aklines can 
the Hough 
s, the suit- 
dge pixels 
inking step 
  
Figure 3: Above: left, difference and right orthoimage. Below: breakline detection from images by using ISEF-edge detector. 
e whether the existence of a breakline within a single 
mesh can be determined by using the Hough trans- 
form, 
e whether a possible breakline within a mesh can be 
determined approximately, and 
e whether possible breaklines in neighbouring meshes 
could be modelled by using a single continuous line 
and triangles, or would it be easier to model break- 
lines by densifying the network with squares. 
The Hough transform was computed for each mesh sepa- 
rately, for example, 16 x 16 Hough transforms were com- 
puted at pyramid level 2 for areas each consisting of 22 x 22 
pixels. If the count of the global maximum cell in the accu- 
mulator array exceeded a specific threshold, it indicated that 
there was a breakline within that mesh. The threshold was 
used in order to filter out weak, insignificant breaklines and 
noise. The results of the Hough transform showed that the 
transform was sensitive to mesh partition. Due to noisy test 
images gaps appeared at breakline locations in which there 
were only a few edge pixels left within a single mesh, when 
a high threshold was used (c.f. the gray meshes between 
the white meshes in the left hand side of figure 4.a), while 
by using a lower threshold a lot of insignificant breaklines 
were found and the requirement to keep the number of new 
meshes low could not be fulfilled. To handle this problem 
a new algorithm for using two thresholds was developed. 
A higher threshold was used for edge points correspond- 
ing to significant breaklines and another lower threshold for 
potential, weak breaklines. The main idea was to get more 
global information about locations of breaklines by compar- 
ing the results of neighbouring meshes. If a chain of signifi- 
cant breakline meshes was found by using the higherthresh- 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
old, it was assumed to indicate that neighbouring meshes 
corresponding to potential breaklines were also significant 
breakline meshes. 
The results obtained by using the new algorithm seemed 
to be promising when the available test images were used. 
Almost all meshes corresponding to the most significant 
breakline were found and rough locations of the two most 
significant breaklines could be found, c.f. locations A and B 
in figure 4.c. Only a few insignificant breaklines, such as 
locations C, F and G were found. However, local distur- 
bances, such as dust at locations D and E, could not be sep- 
arated from breaklines when two orthoimages were used. 
When more images are used this should not be a problem. 
On the basis of these test results, the new algorithm will be 
implemented to applicable parts to the program package for 
airborne video imagery being developed. By means of the 
algorithm it could be determined whether there was a break- 
line or not within a single mesh — at least in the most cases 
of the available test images — but finding the approximate 
locations of breaklines within single meshes seemed to be 
a more complicated problem. If the breaklines were mod- 
elled by using a continuous line that consists of line seg- 
ments in adjacent meshes, at least two problems should be 
solved: connecting successive disconnected line segments, 
when different edges corresponding to the same breakline 
are found, and handling of the breakpoints or discontinuities 
that did notform a breakline, but rather form an area consist- 
ing of single breakpoints and causing occasional direction 
for corresponding line segment. Due to these problems a 
variable sized grid network utilizing rough information about 
the locations of the breakline meshes was found to be more 
335