methods is based on the using of "multilayer perceptron" (MLP) 
and on "back propagation" algorithm. The methods uses 
three-layer unidirectional network, where each neuron (node) 
from one layer is connected with all neurons of previous layer 
(see Fig. 1). 
layer 1 (input) layer 2 layer 3 (output) 
Fig. 1 . 
Any connections between two neurons i and j has a certain 
weight w-jj. In the course of data processing spectral features 
of processed image elements are assigned to individual neurons 
of the input layer. Neuron values in higher layers are 
computed from the expression 
x ; = S( JT w-5 j . x i) 
J 
where J is a set of neurons from previous layer, and S is a 
certain usually sigmoid function. 
Data processing by means of such neural networks runs in two 
steps: adaptation and evaluation. In the course of adaptation 
the weights of connections are changed until required results 
are obtained in the output layer. If the goal of this network 
is to perform data comprimation, it is necessary to require 
the equality of input and output values. If such configuration 
is obtained, then the values of middle layer (having lower 
capacity than input and output layers) can be considered as 
the effective comprimation of the original information. The® 
evaluation of image data consists in the introducing of all 
pixel values to the input of "instructed" network. 
The values obtained in the middle layer (having three neurons 
in the case of TM data processing) may be used as the 
components for color composite production. Assignement of 
individual components is chosen empirically. 
Histogram equalization 
Sometimes, the outputs of above mentioned transformation do 
not display sufficient image dynamics. The color composite 
would not be expressive enough. Therefore, it is necessary to 
225