ITERATIVE ALGORITHM FOR MULTIDIMENSIONAL IMAGE ANALYSIS 
V.M.Lisitsyn, V.A.Stephanov, N.N.Pasechny, K.V.Obrosov 
The Research Institute of Aviation Systems, Moscow, Russia, ISPRS Comission III 
ABSTRACT 
The paper considers 
multidimensional speckled images; 
the presence of some statistical dependence between the value of one 
assessment 
provides a common approach to segmentation procedure 
of image generation 
and the validity of value 
processes feature 
the method of se 
he 
entation and analysis of one class of 
specific feature of these images is 
component 
of the 
other. The proposed method 
irrespective of physical 
under stated conditions. For normal 
. v. by . . . 
parameter assessment distribution based on a maximum a posteriori probability 
criterion. 
Each region being segmented is 
assigned an index. At the same time 
algorithm allows to carry out analysis of the reference scene, since its output 
is not only a se 
KEY WORDS: Algorithm, Image Analysis, Image Processing, Remote 
INTRODUCTION 
In the automatic analysis of scenes, the 
main problem is the conversion of 
information, the image of a scene 
represented as a two-dimensional 
function, into some description of this 
image. This description can be stored in 
a memory section which is thousands times 
smaller than that required for storing 
the image. At the same time, the 
information contained in the original 
image and essential for the above 
analysis is retained and converted into a 
processable form (Duda,1973). 
An indispensable stage of image 
processing aimed at preparing iis 
desoription is segmentation which 
consists in fragmenting the image into 
regions which are ooherent by some 
atiribute. Segmentation can be aimed at 
discerni scene objects and suppres- 
si insignificant details (Bopmeenxo, 
1987). 
We shall have a look at 
image types - a class 
images which feature some statistical 
correlation between the observed value of 
one component and the degree of certainty 
one of numerous 
of 4-dimensional 
of the estimated value of another  compo- 
nent. This correlation is characteristic 
of multidimensional speckle images 
generated by remote sensing facilities, 
such as radars, radio-optical ranging and 
deteotion systems and laser radars. In 
the course of the generation of these 
images, each pixel (7,ÿ) is assigned a 
a 
detector-outputted signal/noise ratio and 
some parameter of the reflected signal is 
estimated. The fourth, estimated 
component of the image oan be represented 
by such parameters as the velocity, 
range, radiation polarization factor, 
ete, or their combination, for which the 
degree of estimate certainty is 
probabilistically related to the 
signal/noise ratio. 
obtained by means o 
| feature a specific e 
884 
ented image but also characteristics of each of regions being 
extracted (location and parameter vector). 
Sensing Application 
which is received is a 
superposition of functions of scattering 
by a set of elementary =—reflecti 
surfaces, characterized by differen 
phases and amplitudes. Interference 
results in that in each pixel the 
intensity of the reflected signal takes 
random, Weibull-distributed value, which 
generates the so-called speckle 
structure. The speckle structure of the 
image has a considerable effect on the 
accuracy of estimates of the component 
being evaluated, which, in the first 
approximation, is specified by the ratio 
(Dansac,1985): 
reflected signal 
an — 1m, (1) 
[A7 A, 
where A is the intensity of the 
received, A is 
noises. Therefore values 
elements of the speckle image 
being estimated can have 
different estimate variances 
from occasional fluctuations. 
signal 
the power of internal 
of different 
component 
essentially 
resulting 
(Lisitsyn, 
proposes an 
For the above class of images, 
1990; JmcmuwH, 1990) 
iterative algorithm for segmenting 
Doppler laser radar images. This 
algorithm formed a binary image whose 
pixels belonging to patterns which 
represented moving objeots had a value of 
"|" and the remaining pixels had a value 
of "O". Drawbaoks of the algorithm are 
instable operation when patterns 
representing different objects touched 
each other or scene objects partially 
shade each other. Moreover, that the 
segmented image is binary makes it 
difficult to resolve discerned patterns. 
The algorithm which will be proposed 
below oan be considered as a 
generalization of the algorithm for 
binary segmentation of Doppler laser 
radar images, to the oase of nonbinary 
segmentation and its extension to other 
types of images belonging to the same