other 
ts of 
and 
are 
ually 
ction 
plied 
ome 
tural 
rent 
the 
three-dimensional geometry of the scene, is 
the correspondence (stereo matching) 
problem. According to the space where the 
matching takes place, the existing techniques 
for solving the correspondence problem 
roughly fall into two categories: image space 
based and object space based. In the image 
space based matching, the primitives of one 
image are compared with ones on the another 
image. Many solutions to the matching have 
been proposed in the image space. The 
methods vary with different choice of 
primitives: area-based (intensity-based), 
feature-based and structure-based (relational 
matching). Recently, several articles are 
devoted to the object space based matching. 
This method emerged originally from the task 
of reconstructing digital terrain model from a 
pair of digital images, independently 
developed by Wrobel [Wrobel] and Helava 
[Helava], etc. Helava used the concept of 
"groundel" as a unit in object space similar to 
the "pixel" in the image space. The image 
intensities corresponding to each groundel can 
be analytically computed, if all pertinent 
geometric and  radiometric parameters 
(including groundel reflectance, etc.) are 
known. A least square method is adopted to 
determine a set of unknown quantities or 
improvements to their approximate values 
used in the analytical prediction process. 
The mapping of 3-D structure into 2-D 
description is formulated by the concept 
"aspect graph" which was introduced by 
Koenderink and van Doorn [Koenderink] 
[Gigus], based on the idea of using the aspect 
graph of topologically distinct views of an 
object to represent its shape. Informally, at 
each vertex of the aspect graph there is a view 
- an aspect - that is representative of the 
projections of the object from a connected set 
of viewpoints from which the object appears 
qualitatively similar. Two aspects are adjacent 
in the graph if the corresponding sets of 
viewpoints are adjacent. A visual event is said 
to occur when the view changes as the 
observer moves between adjacent sets. 
Control mechanism 
For a complex system, a control mechanism is 
always required to control the reactions on the 
components as well as the ones between the 
components inside the system. In Fig.1, such control 
mechanism is monitored by a database which has 
the description of two kinds of knowledge, that is, 
knowledge about objects and about analysis tools 
(Le. image analysis techniques). The control 
structure concerns the model to integrate diverse 
levels and control the proper information flow and 
scheme. The characteristics of image and 
specification and requirement from the application 
603 
can also be integrated. This is sometimes regarded 
as "knowledge-based image processing system" 
[Matsuyama,87,89] [Nicolin]. 
3. A UNIFIED MEASUREMENT BASED ON 
MDL PRINCIPLE 
In the last section, we have proposed a architecture 
for an integrated image analysis. We want to 
emphasis on the interactive reactions between 
different levels or layers. The operations from low 
layer to high layers have been addressed in a lot of 
literature. In our research, we are interested to 
integrate the information from high layers into low 
layers' operations. In order to do so, we must 
answer the following theoretic questions: 
1) What kind of knowledge can be integrated 
in low and middle-level processing. 
2) What is the proper language which can 
describe the information from different layers 
on the common ground. 
We here examine the several tools or criterion 
provided by probability theory and information 
theory. 
MAP, BE, ML, LS 
The Maximum A Posteriori (MAP) criterion selects 
the best solution on the model X that maximizes the 
conditional probability of the model given the data 
Y: P(X/Y) The MAP criterion leads to three 
important estimation methods, namely Bayes 
Estimation (BE), Maximum Likehood (ML), and 
Least Square (LS). 
applying Bayes' theorem gives 
P(X/Y) = P(Y/X)P(X)/P(Y) 
where P(Y/X) is the conditional probability of 
getting data Y given the model X and P(X) the priori 
probability of the model X. If we assume P(Y) is 
constant, maximizing P(X/Y) is equivalent to 
maximize the 
P(Y/X)P(X), 
which is the principle of Bayesian Estimation. 
Further, under the specification that the priori 
probability P(X) are all the same (constant), the MAP 
criterion leads to the simpler maximum likehood 
principle of maximizing 
P(Y/X) 
If the random variables to which the data Y refer are 
normally distributed, the maximum  likehood 
method will give the same results as the least