several applications based on the system we 
proposed. 
2. AN INTEGRATED ARCHITECTURE 
FOR IMAGE ANALYSIS 
In Fig.1 (next page), an integrated architecture for 
image analysis is proposed. In the following, the 
system is explanted in detail. 
Description of each layer 
original image. Raster images are the most 
common input for the image analysis, which 
can be in format of binary (2-valued), grey, or 
in multispectral forms. In our research, we 
only deal with 2-D images, not with 3-D 
images, such as range images and medical 
images. Each pixel on 2-D image is indexed 
from left to right and from top to bottom. 
segmented image. In order to interpret a 2-D 
image, the image is first partitioned into 
regions, and each region is uniform and 
homogeneous with respect to some criterions. 
For the purpose of incoperating the high level 
knowledge into the segmentation, an adequate 
data structure should be designed to represent 
the segmented image, which should fulfil the 
requirements: 1), it should be able to be used 
as the linkage between the original raster 
image and vector representation of objects 
implicitly contained on the image, which 
means that the data structure should 
represent each region directly and it should be 
easy to calculate the every kind of properties 
associated with regions such as the region 
boundary list, area of region, intensity mean 
of region, etc.; 2), the data structure should be 
in a hierarchic fashion. Such requirement is 
based on the observation that segmentation is 
an evolving procedure which usually starts 
from original raster image and gradually 
groups small regions into more meaningful 
regions. During such evolution, some 
grouping or decision making may go wrong 
due to a variety of reasons. Therefore it 
should be possible to return to more primitive 
status and make a new decision. Bearing these 
requirements in mind, a "N-node tree" has 
been developed (Fig.2). 
root node 
uH 
NU A 
E HmNENNENENENEHESMH 
lowest level 
Fig.2 
600 
One can regard the "N-node tree" as the 
extension of quad tree, and under such 
extension the number of children under a 
node is changeable. The whole tree consists of 
a number of levels, with each level 
representing segmentation results at different 
stages. Each node on one level describes a 
complete region which has no overlapping 
with other regions on the same level. For a 
node on one level, one can find out its 
associated original image pixels by tracing 
down the tree through its children until the 
lowest level is reached where each node 
represents the image pixel indexed from left 
to right and from top to bottom on the image. 
In practice, in order to facilitate the 
segmentation, the pixels corresponding to a 
region is stored as one of properties of a node. 
In addition, a label image, which has the same 
size as the original image, is created and 
valued by its corresponding label. By such 
strategy, it is easy to refer the label by image 
pixel or refer image pixels by a region node. 
vector representation. Vector representation is 
the critical step for shape analysis. Shape is a 
function of the position and direction of a 
simply connected curve defined within a two 
dimensional field. A simply connected curve 
is one in which any point on the curve has at 
most two neighbours which lie on the curve. 
The coding of shape may involve a 
description of a closed boundary or the pixels 
which lie with it. In order to incoperate the 
result from edge detection and point 
detection, we extent the concept of "regions" 
to include the lines and points by representing 
the region boundary using the edges between 
the region boundary pixels instead of pixels 
themselves (Fig.3). 
DILL 
Ag CICER 
Zn 0 
Cu 
Cirio 
[ JOE 0L 0L 0E OL 0E] 
[] image pixel 
c3 edge between image pixels 
Fig.3 
2-D structural description. The individual 
description on each object region is often not 
sufficient for the final goals of many 
applications, because such description may be