HRR Signature Classification using Syntactic Pattern Recognition Michael A. Turnbaugh Kenneth W. Bauer, Jr. Mark E. Oxley J.O. Miller 2950 Hobson Way Wright-Patterson AFB, OH 45433-7765 937-255-6565 michael.turnbaughaafit.edu kenneth.bauer@afit.edu mark.oxleyWafit.edu john.miller(afit.edu Abstract An Automatic Target Classification system contains a classifier that maps a vector of real numbered features characteristic to a specific target onto a class label. Other features can be a string of symbols or alphabets that may not involve real numbers at all. There are certain orderings of the symbols in the strings governed by syntax rules, thus, generating a language, (that is, a collection of strings). Thus, a classifier would map a string to a class label. Such a classifier is called a syntactical classifier and varies greatly from its vector space counter part. This paper will give an overview of the construction of a grammar that generates a language then shows how they fit into a syntactical classification system. The performances of two syntactical classification systems with two and ten labels respectively are presented via confusion matrices. Experiments performed on public release DCS database indicate this approach has sufficient power to perform target detection using HRR signatures. 12 TABLE OF CONTENTS 1. INTRODUCTION .................................1 2. SUMMARY OF RELATED RESEARCH ......................2 3. SYNTACTICAL CLASSIFICATION SYSTEM ..............3 4. HRR PROFILE CLASSIFICATION TESTS ...............5 5. CONCLUSIONS AND FUTURE RESEARCH ............7 REFERENCES................................ 8 BIOGRAPHY................................ 9 1. INTRODUCTION Textbooks such as [1] describe pattern recognition as the categorization of data into identifiable classes by extracting features or attributes of the data from a background of irrelevant detail. The methodology to perform such a task can be broken up into two distinct schools of thought. 1 1U.S. Government work not protected by U.S. copyright. 2IEEEAC paper #1 193, Version 6, Updated December 17, 2007 The first is the decision theoretic or statistical approach. The statistical approach is generally based on using numerical features as a means to distinguish one class of objects from any other class. These features are then used to classify a new object based on its features being more like the features of a template of prototypes of a given class than the prototypes of any other class. Measures such as Mahalanobis distance or city block distance are examples of methods for computing this likeness. The second approach is the structural approach. This method attempts to use representations of the class's structure as a means of distinguishing the objects of a class from other classes. In this context, structure refers to the way the components of a pattern are related to one another. A specific technique which uses the structural approach is syntactic pattern recognition. Syntactic techniques use formal language theory, which we will now discuss further. Formal language theory can be traced back to the work of Chomsky, among others, who is often cited as a pioneer in the area. In [2], Chomsky introduced the hierarchy of grammars, which is a basis for syntactic pattern recognition. The clear analogy between the Chomsky hierarchy and the grammatical decomposition of an English sentence gives a fairly straightforward means of illustrating this idea. Another pioneer in the area of syntactic pattern recognition is King-Sun Fu, whose work in grammatical inference through the 1970s and 1980s focused on developing formal grammars whose rules were learned through the training process. In [3], Fu describes the theoretical basis and gives several application areas of syntactic pattern recognition. His chapter on grammatical inference serves as a framework for the development of grammars useful in pattern recognition applications such as the grammar we develop in our research. In our application, we develop a grammar to represent the structure of high range resolution (HRR) signatures generated from synthetic aperture radar (SAR). The grammar describes the structure of the HRR signature using peaks extracted from the HRR profile while suppressing 1