idence that
symbol or
s a floating
1s to 1, the
/mbol.
backpropa-
1990 and
ın adaptive
n modifica-
it networks
As it con-
lues (learn-
activation
>n neurons,
', the error
to decrease
eurons, and
1elps in not
function’s
more itera-
>w symbol
uild a train-
the system.
bd, and the
ot. Second,
symbol is
nd its false
ed in recall
i networks
entation. It
ouching or
to the same
ietworks.
Iter. Not all
nbol's neu-
se enough.
rents accu-
per symbol
ecognition.
on others is
d indepen-
at the same
1 symbol is
' ones with
n the sym-
. However,
before removing a symbol, we compute the number of
remaining black pixels it overlaps. If that number still
exceeds the user threshold, we remove it and add it to a
list of recognized symbols. Otherwise, the symbol is dis-
carded. When two symbols match at the same (or close)
location, this method selects the symbol matching the
most pixels first. If there are enough remaining pixels (a
rare case), then the other symbol can also be selected.
5. EXPERIMENT RESULTS
We used a 7000x5000 pixel map for training and for
adjusting the thresholds. Another map of similar size was
used for computing the results. It takes around one minute
to process one symbol in one orientation. Table 1 presents
results using only the Hausdorff distance. Unfortunately,
some symbols were missed. Most misses are caused by
very badly drawn symbols. Most come from regions
where symbols are hand drawn in very crowded spaces.
The problem symbols are usually very small and very
badly written. Such symbols were not present in the train-
ing map.
Table 2 shows the recognition results when trained
neural networks are also used, with a measure of confi-
dence of 0.85. When the output of the neural network is
greater than 0.85, the candidate is accepted. This very
high measure of confidence is better used when the goal is
to accept only the most probable symbols. Some might be
missed but no candidates should be accepted by error as is
reflected by our results. Only the filled circle is not cor-
rectly recognized. This is due to the fact that the test
image contains a lot of false positive instances that were
not present in the training image. For this symbol, more
training would thus be necessary.
Table 3 shows the result when using trained neural
networks with a measure of confidence of ().5, where can-
didates will be accepted when there is more confidence
that they are symbols than false positives. Here only a few
symbols are missed, and only a few false positives remain.
Table 4 shows the result when using trained neural
networks with a measure of confidence of 0.15, where
candidates are accepted when there is no compelling evi-
dence that the symbol should be rejected. When compar-
ing to the symbols missed by the Hausdorff distance,
almost no misses are caused by the neural networks. But
more false positives are present. For our application, this
is the best threshold to use, because it is easier for a user
to remove false positives than to search for missed sym-
bols. Again, these results confirm that more training is
necessary for the filled circle. The six extra misses for the
filled triangle all come from larger triangles not present in
the training set. Again, more training would fix this prob-
lem.
Figure 1 shows a 1200x400 part of the test map. Fig-
ure 2 shows the recognition of ellipses in Figure 1, as pro-
duced by the Hausdorff distance. All of them are recog-
nized, but one false positive was produced near the lower
683
right corner. The false positive is later correctly removed
by the symbol's neural network.
As expected, the Hausdorff distance produces a lot of
false positives. But after neural network filtering, the rec-
ognition results are acceptable for user corrections. We
expect even better results when training is performed on
more than one map (and thresholds are adjusted not to
miss any symbol) making user corrections simpler. For
sets of maps that are better drawn, over 98% recognition
rates have been accomplished.
For the image shown in Figure 1, we get perfect rec-
ognition. It contains a lot of different symbols, in different
orientation and scale, touching and overlapping each
other. All 83 instances of the 6 symbols shown in the
tables are correctly recognized in around 30 seconds.
6. CONCLUSION
Our strategy for achieving near perfect recognition is to
first generate results that have near zero miss and then:
reduce the (numerous) false positives to an amount that
can be quickly handled by a human operator. Symbol rec-
ognition based on Hausdorff distance combined with indi-
vidually trained neural networks results in accurate
recognition.
Our objectives for the near future are as follows.
First, we would augment the functionality of the system
with various additions such as facilities for the user to
specify constraints on the symbols and their surroundings.
Second, we will study maps for new application areas
(such as navigation and terrain modeling) for which a
knowledge-based component of the system will become
very important since information needs to be inferred by
the system from the extracted symbol information.
Finally, we will extend our work to use color images as
input instead of only black and white ones.
Table 1: Hausdorff
Symbol | Correct Miss sd Te
= |5 0 39
= | 17 1 34
RA | 30 2 145
0) 49 3 129
d 127 26 241
ie 29 1 178
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
