to vague for the purpose of high-level analysıs 
such as matching and object recognition. 
Structure is used to represent the 
interrelationship of the boundaries, which 
provides the basis for many kinds of tasks in 
computer vision, specially for matching, object 
recognition purpose. 
3-D structural description of models. The 
representation scheme for 3d world is the 
typical topic in symbolic artificial intelligence 
and computer graphics. A structural 
description of an object consists of the 
descriptions of its parts and their 
interrelationships. The parts of an object can 
be primitive (nondecomposeable) or they may 
be further broken down into subparts. When 
the parts of an object are not primitives, the 
structural description of the object consists of 
one level of descriptions for each level of 
subparts. Such a multilevel description is 
called a hierarchic description and is useful 
for complex objects with many repetitions of 
parts and subparts. 
Operations on one laver 
By "operations on one level", we mean that input 
and output of the operation are in the same format, 
e.g., from raster image to raster image, from vector 
data to vector data, etc. During the processing, the 
content of representation may change. 
operations on the original raster images. These 
operations may include 1) calculating the 
characteristics of image (e.g. histogram 
transformation, etc); 2) image quality 
improvement (e.g. enhancement, noise 
suppression by filtering, etc). 
operations on the segmented images. Split- 
and-merge is the main mechanism in the 
segmentation procedure, which merges small 
regions into more meaningful big regions, or 
split the big regions into small regions if 
necessary. 
operations on the vector data. This refers to 
the line fitting or curve fitting algorithm, 
which reduces the data needed while keeping 
the result as closed as possible to the original 
data. To be useful for high-level analysis, 
these vector data must be approximated so as 
to overcome local noise, and be represented in 
a more manageable form. The more 
comprehensive that representation is, the 
better the performance of the analysis would 
be. 
Operations between the layers 
The operations under this category changes or 
602 
transfers the data from one representation to another 
representation, which are the essential parts of 
image analysis. 
operations between the original image and 
segmented image. These operations are 
generally called segmentation which usually 
is in two kinds of forms: a), edge detection 
and line following. This category of 
techniques study various of operators applied 
to raw images, which yield primitive edge 
elements, followed by a  concatenating 
procedure to make a coherent one 
dimensional feature from many local edge 
elements; b), Region-based methods, which 
depend on pixel statistics over localized areas 
of the image. Regions of an image 
segmentation should be uniform and 
homogeneous with respect to some 
characteristic such as grey tone or texture. 
Region interiors should be simple and without 
many small holes. Adjacent regions of a 
segmentation should have significantly 
different values with respect to the 
characteristic on which they are uniform. 
Boundary of each segment should be simple, 
not ragged, and must be spatially accurate 
[Haralick]. 
operations between the segmented image and 
vector data. This so-called vectorization 
procedure traces along the each region 
boundary to get the boundary position and 
the position is represented in chain code, 
which is used later by shape analysis. On the 
other hand, in order to integrate shape 
constraint into segmentation, there is another 
information flow which transfers the result of 
curve fitting into the region growing. The 
principle of encoding shape is described in 
section 3. 
operations between the vector data and 2d 
structural description. These operations build 
the structural description by performing a 
geometric analysis on the vector data. Vector- 
based perceptual grouping can be also 
included in this category, which organizes the 
fragmented  low-level descriptions into 
meaningful higher level descriptions by 
mimicking the human visual system in 
detecting geometric relationship such as 
collinearity, parallelism, connectivity, and 
repetitive patterns in an otherwise randomly 
distributed set of image elements, some 
relevant work can be referred to [Mohan]. 
operations between the 2d and 3d structural 
descriptions. Matching two or more than two 
images of the same scene from different 
viewing positions in order to recover the 
al