features are realized using the PHIGS (Programmer’s 
Hierarchical Interactive Graphics System) and C 
programming languages, while the automated 
algorithms are being realized in FORTRAN 77. The 
automated algorithms are currently being developed on 
a 486-based PC, but will eventually be implemented on 
the SPARCstation. 
The next section of this report discusses the types of 
automated classification algorithms that have been 
developed for eventual use within ROICATS. Section 
3 discusses the types of computer-based tools that are 
being considered. Section 4 discusses the operator- 
machine interface features that are being implemented. 
The paper concludes with a discussion of the future 
work required to complete the development of 
ROICATS. 
2.0 AUTOMATED SHIP CLASSIFICATION 
ALGORITHMS 
Over the past eight years, the Airborne Radar section 
has been researching and developing numerous 
algorithms that can be used to automatically classify 
high-resolution radar images of ships. Only a brief 
description of the most noteworthy of these algorithms 
is given. 
The various classification algorithms can be divided into 
two general catergories. The first category is based 
upon a robust technique that uses coarse image 
features to make a classification. Various coarse 
features are extracted from an input ship SAR image 
and are compared to a ship library containing the same 
features. There are basically six types of coarse 
features extracted: the number, level and location of 
major peaks in the image; the relative range location 
and height, in cross-range, of any vertical structures 
evident in the image; the estimate of the minimum ship 
length; the presence of the bow of the ship; the outlines 
of the sides of the ship; and the presence of a rotating 
antenna. After extracting values for these features, an 
Expert System (ES) methodology (i.e. a set of rules, a 
knowledge base, and a rule control strategy) is used to 
guide the comparison of features. Since the features 
are coarse and not necessarily persistent, there is a 
good chance that several ship classes will be declared 
to be valid matches by the ES. It is required that the 
ES output a list of candidate ships which is much 
shorter than the initial list of all possible ship classes. 
The list of candidate ship classes is presented to the 
Radar Operator and fed into the second category of 
classification algorithms. By having to consider a much 
smaller subset of the library of ships, the subsequent 
classification process should be faster and there should 
be less chance for error. Incorporated into the ES is a 
techique which automatically determines the quality of 
an input SAR image. This technique basically allows 
the ES to filter out those images which are considered 
too poor for classification. 
The second category of algorithms base their decisions 
on the actual pixel values in the input SAR image. 
There are three different techniques being developed in 
this category: a combinatorial optimization technique 
(sometimes referred to as simulated annealing), a 
decision theoretic technique, and neural networks. 
The combinatorial optimization technique takes a 
computer-based block model of a ship (an example of 
such a model is shown in Figure 1), and simulates its 
SAR image for various values of the following four 
parameters: the three angular velocities, roll, pitch, and 
yaw; and the ship heading with respect to the SAR line- 
of-sight. The simulated image is compared to the real 
input ship image using a two-dimensional correlation, 
and if the comparison is poor, the value of one of the 
parameters is modified and another image is simulated. 
The schedule, or method of selecting a parameter and 
its new value, is the heart of the combinatorial 
optimization technique. This technique finds the 
optimum set of parameter values so that the best match 
between the input and simulated images can be 
obtained. 
The decision theoretic method uses templates that 
represent incrementally learned simulated SAR images 
of ships. For each class of ship there are several 
templates, each representing a range of aspect angles 
corresponding to 45 degrees. Simulated SAR images of 
a ship are used to incrementally learn a template, also 
called a discriminant vector, that is representative of 
that ship for the range of aspect angles in question. 
Only after each vector is created can the decision 
theoretic method perform a classification. A 
classification decision is based upon computing the two- 
dimensional correlation between the input ship image 
and each discriminant vector. The vector which gives 
the highest correlation is considered to represent the 
class of the ship in the input SAR image. 
Neural Networks are also being evaluated as part of the 
second category of algorithms. These networks 
promise to provide quick recall of learned pattern 
associations with high recognition performance, even 
when noisy or incomplete input patterns are 
encountered. Research and development is being 
performed primarily with the BackPropagation Network 
and derivatives of it. 
Figure 2 shows a block diagram of the overall 
classification system, ROICATS. The operation of 
ROICATS is as follows. A real SPOT or RDP image 
326 
  
of a ship 
is libran 
potential 
the input 
detailed a 
Machine 
computer 
features 
sections. 
respective 
performs 
Since non 
expected 
performa 
the same. 
be coml 
probabilit 
enhanced 
since thes 
of the ini 
can be 
evidential 
made usi 
knowledg 
1986). T| 
adequate 
classificat 
intelligen 
final out 
Radar Oj 
Compute 
the mea 
informati 
informati 
classificat 
allows c 
ROICAT 
the Oper 
essential 
Also, on 
reasonab 
provided 
Operator 
eight gro 
listed in 
follows. 
The SAR 
will conta 
backscatt 
man-mad 
radar reti