three-component pixel value RGB. Thus, the 
dimension of element must be considered as a third 
co-ordinate of data types. Analogously, the 
dimension of structure must be the fourth data type 
co-ordinate. However, nobody thinks that the 
images of different size are the different data types. 
So, let's adopt the following agreement: the 
dimension will be the binary characteristic taking its 
value on the set of (one, any). It means that any 
element is presumed to be either the one number 
(symbol) or some list (vector) of arbitrary length. 
This agreement makes the software implementation 
of our framework much more easy [1]. 
2.2. Procedure types 
All the necessary procedure types have 
been discussed in [1]. That's why we'll only mark 
the principle moments. As mentioned above, the 
processing frames have some  input/output 
information links. So, some input set and output set 
must be determined for each principal processing 
frame type. The input set description includes the 
number of input data with their types. The output set 
description includes the number of output data with 
their types. In this section we shall consider the 
structure dimension as a fourth co-ordinate of data 
type. 
Generally, it is possible to define the 
mapping of any input set onto any output set. 
However, it leads to the set of procedures of 
unobservable size. So, we need to introduce some 
additional constraints. 
The first important constraint concerns the possible 
content of input and output sets: 
any input(output) set can contain only data of the 
same semantic level and same organization level. 
The second important constraint we refer as an 
Exclusive Modularity Principle: 
any mapping of any input data set onto any output 
data set can be represented through a combination 
of on-level and exclusive inter-level procedures only. 
The third constraint declares that: 
any inter-level procedure satisfies the condition of 
Nin=1 and Ng, t=1 (where Nj, Ngyt - number of 
inputs and outputs, correspondently). 
In particular, it means that the fusion procedures 
change neither organization nor semantic level of 
data representation. 
The fourth important constraint we adopt in the form 
of the Cumulative Fusion Principle: 
any structure fusion procedure can be represented 
as a combination of number of the pair-wise fusion 
procedures. 
386 
It is very powerful assumption that is not 
true for some usable fusion approaches. However, 
the cumulative fusion proposal makes it possible to 
design the required fusion scheme using the unified 
fusion frames (with two input and one output links) 
for any possible set of sensors. 
The last constraint of our framework closes the set 
of on-level procedures: 
only two types of on-level procedures are available: 
filters and pair-wise fusion procedures. 
Now, after this preliminary discussion, we 
are ready to outline the developed set of procedure 
types. 
There are two types of the exclusive inter- 
level procedures: procedures that transform data 
with level increasing and procedures that transforms 
data with level decreasing. When the level of data 
abstraction increases, the extraction of information 
occurs. This case takes place immediately during 
data processing and fusion. The converse case 
corresponds to the data modeling process. We shall 
not consider any modeling here. 
A. Semantic inter-level procedures. Procedures 
of this class preserve the structure while updating 
the elements. The most important procedures are: 
e feature extraction - calculates some feature for 
each element using measurements; there must 
be a set of such procedures to obtain the 
required feature vectors; 
e feature-based classification - assigns the 
symbolic (class) labels for each element using 
its features; may be of soft or hard type, may 
the Bayesian classification, cluster analysis and 
SO On; 
e object detection - assigns the symbolic (class) 
labels for each element using the matching 
techniques at the measurement level; may be 
of soft or hard type; may use the Bayesian 
classification, correlation, model-based 
methods and so on; there must be a set of such 
procedures if a set of known objects is given; 
e decision making (recognition) - assigns the 
unique symbolic (class) label for each element 
using the soft (probabilistic, fuzzy) or hard 
evidences from different sources. 
B. Structural inter-level procedures. Procedures 
of this class preserve the element types while 
updating the structural organization. The most 
important procedures are: 
e segmentation - transforms the raster data into 
regions and stores them in some structure; may 
use the contour-based, region-based, texture- 
based, relaxation and other techniques; 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996 
  
e scene 
basec 
basec 
spatie 
e oObjec 
presu 
move 
previ 
image 
filterii 
extra 
2.2.2. Pai 
P 
structures 
of the sa 
element f 
A. Elen 
important 
e elem 
struct 
dimei 
struct 
prese 
use i 
levels 
meas 
e SOft ( 
the s 
both 
dime 
Scha 
e evide 
hard 
struct 
may | 
B. Stru 
important 
e  Síruci 
struct 
use ir 
e struct 
of t 
dimei 
data: 
* rang 
struct 
evide 
elem 
(thres 
of usi 
1 
each of 
System © 
Structure