ng 
ri 
he 
e- 
ma 
  
  
  
  
  
  
Figure 1: Two pictures containing a set of dots 
a feature extraction system based on this bottom- 
up data-directed organization of interesting percep- 
tual events: namely the function of grouping pixels 
or image events with weak evidence into new image 
events which would be better interpretable and the 
function of describing the new image events effecti- 
vely to provide more information about their featu- 
res. How to realize these two functions in a system is 
the objective which we are going to discuss. 
In this paper, a two-stage paradigm for feature extrac- 
tion from images is proposed. It includes two procedu- 
res of neural network based feature grouping and mo- 
del driven feature describing and it takes into account 
the above mentioned troubles for feature extraction. 
In order to demonstrate this paradigm, we just look at 
the problem of extracting straight lines from image ar- 
rays. Based on a novel neural network model, we pre- 
sent a high-quality line finder which gives a complete 
description about geometric and photometric proper- 
ties of extracted lines. The approach can be extended 
to find other more complex image events, including 
arcs, curves, polylines, and polygons. 
2 A Two-Stage Paradigm 
Image Understanding can be thought of as an infe- 
rence process in which a description of the outside 
world is inferred from images of the world, having 
assistance of our a priori knowledge and experiences 
(ZHENG, 1992). Drawing inference from image data 
requires, first, the ability of discovering image events 
and representing their features in a relevant way, as 
mentioned earlier. Before we realize this ability tech- 
nically, we should first know how an image event is 
perceived. 
An image is a distribution of the luminance inter- 
cepted by the camera lens. Many factors, including 
the surface material, the atmospheric conditions, the 
light source, the ambientlight, the camera angle and 
characteristics etc., are confounded in the image and 
contribute to a single measurement, say the intensity 
of a pixel. The various factors cannot be separated, 
865 
as long as they are not measured. So a single pixel 
can support one hypothesis only with very weak evi- 
dence and it can present some visual impressions only 
in combination with many other pixels. This can be 
made explicit by examining the two pictures illustra- 
ted in Figure 1. Here both pictures contain a set of 
dots as stimuli. However, only the left picture present 
a visual impression of a straight line. This suggests 
that image features are represented by a group of pi- 
xels and the first step in feature extracting is grouping 
such pixels into so called feature-support regions, ba- 
sed on our knowledge of what we want to extract. 
For digital images, the situation is more complex, as 
they are corrupted by both discrete spatial sampling 
and intensity quantization, and there is stochastic 
component in image data. In this case, the evidence 
which can be given by a pixel is not only weak but 
also erroneous and unreliable. So, the second step in 
feature extraction is describing a group of noisy pi- 
xels obtained through grouping. This is an ill-posed 
problem as one can hypothesize an infinite number 
of different underlying descriptions (ZHENG, 1990). 
Of course, many of these descriptions are senseless. 
But the main question is how to find the description 
which we think of as the best one according to our a 
priori knowledge about what the group of noisy pixels 
should present. This means that a feature extraction 
system should be able to verify and describe feature- 
support regions using models and to give more com- 
prehensive information about their features for sub- 
sequent inference and reasoning. 
Obviously, features in images may have different 
degrees of abstraction and complexity. So, feature 
extraction requires a bottom-up hierarchy of the 
grouping-describing process, from low to higher ab- 
straction. 
3 Perceptual Grouping 
The human vision is the only example for developing 
an artificial system to solve visual problems like fea- 
ture extraction. However, our knowledge about the 
human visual system is very limited, no matter from 
physiological or psychological point of view. 
One of the most obvious and interesting facts of hu- 
man visual perception is the ability of the so called 
perceptual grouping which is based on the researches 
of the Gestalt psychologists. They argue that humans 
must be able to partion a scene into coherent, organi- 
zed and independently recognizable entities or groups 
by using a set of generic criteria and Gestalt laws and 
the human visual system is very good at detecting