In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
351 
forest, meadow, tilled sites, shrub bogs, abandoned 
mines, bushes with open forest, mire, river bank, 
floodplain, saline sites. 
All experiments proposed that if the first result of retrieval 
referred to the same class of GUMs as the query pattern, the 
result was correct. 
In experiments with IDB-1 the fixed number of images for the 
given patterns was retrieved. As for IDB-2, the retrieval was 
stopped as soon as all GUMs relating to the same class as the 
query pattern were received. 
Table 5 gives the results of the first queries to IDB-1. Though 
IDB-1 represents a relatively complex variant for visual 
retrieval due to ambiguity of objects division into classes, the 
retrieval results turned out to be rather encouraging. About 90% 
of queries show the correct result as the first choice of retrieval 
system. The first two results are correct for more than 65% of 
queries, and more than 40% of queries showed the first, second 
and the third retrieval results to be correct. 
2 nd retrieved 
pattern 
(class #) 
3 rd retrieved 
pattern 
(dass#^ 
query 
pattern 
(class #) 
1 retrieved 
pattern 
(class #) 
*!gS! S 1 
D35(20) 
D3(20) 
mk 
if 
% 'k' -Wus, 
D109(29) 
D107(30) 
.I?*»«* 
(*) 
D104 (3) 
D110(29) 
D94(16) 
D95(16) 
D96(15) 
D56(2) 
'iV.VfV 
V iV.S V 
!‘( ITITf 
D6(l) 
D76(14) 
D14(l) 
D25(15) 
D68(14) 
D85(23) 
D4(18) 
Table 5. Retrieved patterns for IDB-1. 
certain visual criteria, which was not accounted in the original 
classification. 
In a like manner, the system found new features of visual 
resemblance for pairs D109 and D107, D109 and D108, D94 
and D96, D6 and D21, D76 and D85 that were not included in 
original classification, but these features are correlated with 
perception. 
The experiments with the patterns taken from high-resolution 
images and represented in Table 6 turned out to be more 
optimistic when compared to the results with model patterns 
from IDB-1. In 50% of cases the system “chose” all possible 
patterns of the corresponding class, namely, bushes, bushes 
with light forest, resident sites, tilled sites, forest, and meadows 
as the first retrieval results. 
query pattern 1 st retrieved pattern 
bushes 
! bushes 
with light forest 
bushes 
with light forest 
dachas 
forest 
meadow 
meadow 
the next 
retrieved pattern 
2 nd _ 7 th bushes 
only 2 patterns 
in IDB-2 
2 nd -9 th dachas 
2 nd - 5 th forest 
2 nd - 6 th 
meadow 
Table 6. Retrieved patterns for IDB-2 
4. OUTLINING PIECEWISE-HOMOGENEOUS 
REGIONS 
Experiments were carried out with images received in 
panchromatic mode from Quickbird (0.7m) and Ikonos (lm) 
satellites. 
The retrieval system described was found capable of user 
training. Though some retrieval results (marked by (*) in Table 
5) show the incorrect result in accordance with the original 
classification in (Ma, 1996), in fact, they are visually similar to 
the query pattern by one visual criteria or other. Thus, 
according to the original classification, the pattern D 104(3) 
refers to the 3 rd class, but the retrieval demonstrated the visual 
resemblance of this pattern to D35(20) from the 20 th class. It 
means that both patterns can be integrated into one class by 
The results of the outlining of homogeneous regions according 
to model (1) as the first step of conceptual retrieval in IDB of 
high resolution are given in Figure 1, Figure 2. The outlining 
algorithm was realized by computational comparison of 
sketches of regions around neighbouring pixels.