Target classification using neural and classical techniques

This paper describes the results of a limited study to investigate the relative performance of a number of classical classification methods when compared to a multilayer perceptron (MLP) neural network. The comparison uses feature data extracted from segmentations of infrared images of real scenes. For the purposes of this investigation the objects extracted from these images were grouped into two categories, target and nontarget. The performance of the classifiers was then determined using a performance measure that penalised false alarms. The results of this investigation suggest, for this type of classification problem, that there is little to choose between the classification performance of both a k-nearest neighbour approach and the MLP. Furthermore, the results show that both these methods are not unduly affected by the type of pre-processing applied to normalise the classifier input data. This is in contrast to the other classification methods investigated which are shown to be particularly sensitive to the form of pre-processing used