An application of nonlinear feature extraction to the classification of ISAR images

We present a scheme for performing nonlinear feature extraction on ISAR (inverse synthetic aperture radar) images of armoured vehicles. This allows a reduced dimensionality representation of the images that we demonstrate is effective at capturing structure present in the data. This is achieved by comparing the classification results obtained using a nearest neighbour classifier on the extracted features with results on the full dimensionality data. The dimensionality of the incoming data is 2401 and the technique presented here is used to reduce such images to a much lower dimensionality. This results in greatly decreased computing time when calculating the nearest match of a test point with a reference sample. Furthermore the transformation function can be implemented in hardware offering a very fast classifier for real-time applications.