PhD thesis as published 2007
This thesis documents research into the use of feature extraction for achieving robust automatic target recognition (ATR). Classification experiments were performed on inverse synthetic aperture radar (ISAR) images of military vehicles. The issue of robustness was considered over differences between vehicles used for training and those used for testing a classifier. These differences were broken down into two categories: the vehicle used for testing was represented during training, but its configuration was different; and the vehicle used for testing was not represented during training. The specific approach adopted for performing the feature extraction was a flexible radial basis function network. The merit of a given set of extracted features was measured through the error rate obtained using the nearest-neighbour classification rule and compared with linear features. Comparison was also made with results from a support vector machine. A major result of this work has been to perform classification on an extensive and varied real-world data set to show that configuration changes to vehicles are not the major challenge previously thought. Instead it is recommended that future research tackle the challenge of training a classifier on simulated or ISAR data and testing on data from an operationally-representative sensor.