In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
349 
similar in visual sense - so they represent GUMs of the same 
class - if they have the same distribution of signal pairs in 
cliques of the same type. 
Let us assume, the more frequent a combination of signals for a 
paiwise cliques r, the greater distance is observed between the 
pattern's marginal frequencies and the marginal frequencies of 
an independent random field (IRF). Therefore, texture sketch 
corresponding to pattern S=s can be defined as follows: 
Sketch (S=s \ w) = 
={r*ER: Dist(UMd\ s),MP(d))-TRESH Sketch ,rCR}, (1) 
structure is denominated by means of grey level, namely: the 
darker raster structure element, the higher its significance for 
visual content representation is. The outcomes of experiments 
are evidence of strong similarity in sketches of the same class 
patterns (class 1, Table 1) or even their full similarity (class 2, 
Table 1), if class patterns are similar in more than one visual 
criterion. 
To the contrary, if class patterns are similar in one visual 
criterion, significant elements of patterns demonstrate 
difference in positions of window w (class 3, Table 1). 
where R = ((m,n): m=0,...,M-l; n=0,...,N-l) is the finite 2-D 
lattice of the size |/?|=M*A( 
Dist(-,-) is the given type of distance between 
distributions, 
H^*(i/)s) - grey level difference histogram for 
clique r*, deD, 
MY(d)=(\Q\-abs(d))/\Q\~ are marginal frequencies of 
IRF, 
Q = {0,1,..., q max } is a finite set of grey levels q 
in lattice sites (m,n), 
D={-q m ax, •••,0,1,-.., qmax } - a set of grey level 
differences, 
w - the given proximity window, 
TRESHskeich - the threshold for types of the sketch 
cliques. 
The application of the model (1) also means the implicit use of 
perceptional data. It is evident that the more pronounced is the 
linearity, regularity, orientation, etc. of textural pattern, the 
greater will be the visual dissimilarity between the pattern and 
an IRF. 
Experiments with several textures allow to propose the 
following steps to get texture sketch (Gimel’farb, 1999): 
1. Compute grey level difference histograms for all 
cliques in the proximity window. 
2. Compute the distances between the grey level 
difference histograms and marginal frequencies of 
IRF. 
3. Find the clique family, which differs the least from 
the IRF (that is, corresponds to the least distance). 
4. Compute the distances between the grey level 
difference histograms and the clique family found in 
3. 
5. Compute the average distance AvDist and standard 
deviation STD of the distances in 4. 
6. Compute the threshold TRESH S ketch = AvDist+ STD. 
7. Choose the clique families whose distances exceed 
this threshold to represent the sketch. 
Experiments to obtain texture sketch were carried out using two 
types of model patterns: 
1. Dissimilarity patterns allowing instantaneous (not 
exceeding 200ms at a moment) separation from each 
other (Marr, 1982). 
2. Similarity patterns allowing to make separation only 
after thorough study. 
Table 1 represents sets of patterns according to the expert 
classification (Ma, 1996) and contains significant structures for 
diverse proximity; the importance of each element in the 
Table 1. Classes of similar patterns and corresponding sketches 
for 13x13 proximity window 
The size of a proximity window w depends on the textural 
pattern. The larger the window, the more accurate estimate of 
the texture sketch will be, but also the slow will be the search 
in SIDB. So the choice of the size of a proximity window