100 
1) Areas e.g fields, lakes. 
These are homogeneous features extending over 
several or more pixels. 
2) Boundaries e.g field boundaries. 
These consist of low-level edge elements 
where these are defined to be local 
discontinuities in image features. 
(Pratt (1978)) . 
3) Thin lines e.g rivers, forest rides, 
roads, hedges. 
A line segment is defined by the u-shaped 
cross section of image features. 
These three types reflect the different 
approaches to image segmentation. The so 
called region growing strategy starts by 
growing an object in an area until a 
significant change in the measured features 
is noted. Another segmentation strategy aims 
to locate object boundaries thereby isolating 
the objects. 
The overall feature extraction process may 
consist of one or more of the following 
classes of operations. 
1) Preprocessing: This is primarily to remove 
the effect of image noise on the following 
stages. 
2) Feature Extraction: The actual process of 
transforming the image to a specific feature 
domain. 
3) Post Processing: The process by which 
detected features are cleaned i.e filtered of 
invalid or unwanted results. 
Knowledge of the features to be extracted 
can be used at any of the three stages. For 
example knowledge of the noise statistics 
(embodied into a noise model) may be used to 
optimally reduce noise in stage 1. In stage 2 
a knowledge of the feature type may be used 
to reduce the computational cost. World 
knowledge can also be used in stage 3 to test 
the validity of detected features. 
2 EDGE AND LINE FEATURE EXTRACTION 
Extensive surveys of the following classes of 
techniques can be found in Ballard(1982), 
Pratt(1978), Davis(1975), Carlotto(1984) , 
Duda & Hart(1972b). 
2.1 Local Operators 
Local operators use data from the 
neighbourhood of the edge candidate pixel. 
There are a wide variety of such local 
operators details of which can be found in 
Duda & Hart(1972b) and Pratt (1978). These 
simple operators often form the basis of 
commercial edge detection systems because of 
their computational simplicity and their 
potential parallel implementation. More 
sophisticated statistical operators have 
attempted to increase the detection 
performance in the presence of noise however 
Geiss(1984) found the Marr and 
Hildreth (1980) operator unsuccessful at 
locating edges in Synthetic Aperture Radar 
images and although the technique developed 
by Suk and Hong(1984) proved superior to 
simple operators it still, performed 
unacceptably in radar data. 
2.2 Image Modelling 
The following set of techniques all attempt 
to impart some level of knowledge into the 
feature extraction process. 
2.2.1 Parametric Modelling 
The basis of parametric modelling is to fit 
some parametric surface representing an ideal 
edge model to a local region of pixels from 
which the edge characteristics can then be 
derived. Again many researchers have proposed 
operators in this category which can be found 
in Rosenfeld and Kak(1982), Haralick (1980) 
and Chittineni(1983) . 
transform 
is the F 
high pas 
produces 
produced 
transform 
Unfortuns 
processe: 
structure 
2.2.2 Statistical Modelling 
2.4 Graph 
Statistical modelling forms the basis of a 
variety of methods which rely on statistical 
theory for edge detection and involves the 
combination of a noise and image model 
allowing calculation of probabilities of edge 
existence. Then statistical methods such as 
maximum likelihood estimation can be used to 
locate probable edges. Chen and 
Pavlidis(1980). Rosenfeld(1981). 
2.2.3 Template Matching 
In this method explicit knowledge of the 
desired feature shape is used to produce a 
template which is then matched to regions of 
the image. The feature will probably exist 
where the correlation between template and 
image data is greater than a set threshold. 
The method is successful for detecting very 
specific objects such as tanks and aircraft 
in military images where the template may be 
a section of an image containing the desired 
object. 
Simple template operators such as the 
Kirsch operator in Pratt(1978) attempt to 
generalise the process by detecting line and 
edge elements with variously orientated 
templates. 
Another method related to template matching 
by Stockman and Agrawala (1977) is the Hough 
Transform technique (Hough (1962)) later 
improved by Duda & Hart(1972a). For grey 
level images the transform produces peaks and 
troughs in Hough Space corresponding to light 
and dark straight lines in image space. These 
peaks or troughs can then be detected and the 
line parameters extracted, using traditional 
operators. In general the transform can be 
used to extract any feature shape which has a 
known parametric form such as the conic 
sections (Sloan & Ballard(1980)) and has been 
widely used because of its effectiveness in 
both clean and noisy images (see 
Shapiro(1978)). For radar images the 
technique performs successfully at detecting 
straight lines that are long relative to the 
scene size (Skingley & Rye(1985)). More 
recently a modified hough transform, the MUFF 
transform (Wallace(1985)) has been developed 
which uses a different parametric form for 
lines to enable the detection of short lines. 
The radon transform shown to be equivalent 
to the hough method by Deans (1981) is an 
invertible transform and so has been used for 
linear feature enhancement as shown by 
Murthy(1985) for radar images and for 
detection of straight lines in simulated 
noisy imagery as illustrated by 
Nasrabadi(1984). 
2.3 Transform Domain Processing 
This class of processing technique comprises 
a global two dimensional weighted transform 
followed by a spatial frequency filter in the 
A graph 
consists 
between r 
arc is a 
then seer 
path betw' 
graph. IJ 
known the 
solution 
demon st rai 
solution 
(Palay(19 
the usefu 
for nois 
guide the 
3 discuss: 
Local ope: 
are typi 
calculate 
approach 
is thus si 
noise. Ot 
susceptib: 
with edge 
effects 
sophistic, 
in low si 
modelling 
but param* 
disadvantc 
modelling 
accurate : 
they work 
known noi; 
algorithm 
specific i 
being lo < 
embody mo 
fully gi 
techniques 
and do nc 
noise. Gr< 
global or 
evaluation 
to noise, 
optimal 
Usually tl 
processes 
of knowle 
extensive 
exponenti 
practical 
initial b 
provided. 
(1984) has 
this class 
efficiency 
We may n 
thin line 
eharacteri 
human visu 
criteria. 
1) Globa!