In: Wagner W., SzSkely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
350
should be based upon a reasonable compromise between
computational power and representativeness of the sketch.
In particular, for 220x110-pattems a 9x9-window gives already
acceptable results. It appeared to be possible to increase the
size of the window even to 13x13 (Table 2).
Table 2. Sketches for proximity windows of different sizes.
In addition, one significant thing was revealed, i.e. spatial
position and quantitative composition of raster elements
essential for visual representation of a homogeneous region
remain the same even at a big change of visual detail(Table 3).
Samples
manufactur
ed forest
Sketches
Samples
deciduous
forest
Sketches
rare forest
with sand
■
glade with
trees
resident
sites
dachas
Samples of
pine forest
Satellite,
resolution,
band
Sketches
QuickBird
0,7 M
0,445-0,90
Table 3. Pine forest samples and sketches for different
resolutions.
It turns out that the property to seize visual information, if
patterns are similar in more than one criterion, is sufficient for
making comparison of natural objects:
1. Referred to the same class and received with different
resolution (Table 3).
2. Referred to different classes and received with the
same resolution (Table 4).
Table 4. Quickbird-samples and corresponding sketches.
3. CONCEPTUAL QUERY IN IMAGE DATA BASE
The sketch (1) of textural pattern is the characteristic of the
pattern that allows us to discern different textures. The measure
of visual dissimilarity can be used for the arrangement of all
images in image data base (IDB) in order of increasing
similarity with the query pattern as follows:
1. Enter the query spatial-homogenous pattern s.
2. Retrieve the pattern sketch in accord with the model
(1).
3. Calculate measures of dissimilarity Dist(s„s) between
the query pattern and images [s,: t=l,2,...] from IDB.
4. Choose image s t * with the least value of dissimilarity
measure as the first retrieval result:
Distfs^ ,s) = min\Dist(s,, s) 1
t S ( eIDB' t •
Choose the subsequent retrieval results by ranking the values
Dist{s„s) in increasing order.
The experiments were carried out with two IDBs:
1. IDB-1 that contains the patterns of Brodatz textures
(Brodatz, 1966). The pattern classification for 32
classes was taken as a basis for visual comparison of
results of IDB-1 queries (Ma, 1996). The
classification of various groups of experts can be
different; therefore, the experiment was followed by
certain changes in the set of classes as a result of the
self-training of retrieval system.
2. IDB-2 that contains 64 patterns of high resolution
image (Quickbird , 0.7m) and grouped together into
12 classes of GUMs as follows: bushes, resident sites,