the 
zion 
are 
1ing 
nce 
ame 
1 as 
the 
the 
long 
sing 
the 
nore 
tion 
become more unambiguous by using constraint 
propagation techniques. Sets of mutually consistent 
relations , which construct spatial decompositions of the 
scene, make a spatiali relation quadtree. 
It must be pointed out that the procedure described in 
the previous steps , which drives from a local ,low — 
level set of relation to a high — level , more global 
scene relation, is based on the discrimination 
graphs. Discrimination graphs (DGs) closely involve 
a categorization of relation class that belong to a 
particular image feature category. If this procedure is 
imperfect ,such as some relations is missed ,and others 
is interpreted incorrectly, we may repeat some 
processes to revise the errors. As more attributes and 
constraints are discovered in the image by using DGs 
furtherly , the relations shall be forced to become 
more specific. 
3. HIERARCHICAL QUADTREE 
In this section ,the problem of how to transform a 
primary image into a segmentation quadtree, and 
homogeneous region quadtree is described (see step 1 
— 2 in Figure 1). 
At first, the primary image is segmented by the 
judgement of the intraparallelism which involves the 
attributes stemming from the gray level, such as 
intensity, hue, saturation etc. To each node of the 
quadtree segment that maybe contains multiple 
objects, if the attributes are parallel, then give it a 
corresponding values , otherwise give it a question 
mark label which is a control label used to indicate 
where the segments must be continuously 
subdivided. As a result a quadtree composed of these 
values is obtained. It is called the segmentation 
quadtree. 
It has been shown in (V.S. Frost, 1985) that pixel 
intensities of neighboring pixels from the same object 
can be assumed to be incompletely correlated. For 
many objects more than one quadtree segment may be 
obtained. Thus these segments must be conbined into 
homogeneous regions on the basis of the knowledge of 
object that determine collections of segments, which 
559 
form “nutural” components of the scene. 
Homogeneous region quadtree formation consists of 
three facets, location code formation, attribute value 
description and region numbering . location code for a 
leaf representing a(2" X 2?) image consists of the m 
quadrant digits representing the recursive subdivision 
of the raster into quadrants. Attribute value refer to 
specific parts of segmentation quadtree and 
homogeneous region quadtree which satisfy a 
particular homogeity condition. With region number 
we associate two list , in the first list each node is 
represented by two fields, one for the quadrant 
location code ,one for the region number; in the other 
list the nodes that have same region number are 
assigned to same attribute value (see Figure 2). 
Thus the collection of these elements may be regarded 
as a binding which consists of location code, attribute 
value and region number. An example of a binding is 
shown in Figure 2. 
Location Region 
|code number 
  
  
  
  
  
  
  
  
  
  
determination 
  
  
[4 
| 
| Attribute value 
| 
| 
EA 
  
  
  
  
  
  
  
  
  
ER] ER) ER] ER 
Figure 2. An example of a binding for homogeneous region quadtree. 
Figure 3 illustrates the elements of segmentation 
quadtree and homogeneous region quadtree, and their 
transformation. The attribute and uniformity of 
segments, which are mainly based on statistical 
properties and sensor dependent; while the attribute 
and homogeneity of homogeneous regeions mainly 
refer to the knowledge of objects and theme , which 
include the geometric characteristics. For example, 
segments 30, 31, 32, 33 in Figure 3a indicate three 
rice lands with different water depth. Because all of 
them have the same attribute, regular polygon, they 
should be merged into a single homogeneous region by 
using form attribute and discrimination graphs.