Target classification of ISAR images based on feature space optimisation of local non-negative matrix factorisation

The problem of target classification using inverse synthetic aperture radar (ISAR) images is studied under conditions of mass data processing, sparse scattering centre distribution, image deterioration and variation with the radar imaging view, all of which make target classification difficult. In this study, the authors propose a novel method based on combination of the feature space and the visual perception theory to achieve an accurate and robust classification of ISAR images. In order to make full use of local spatial structure information for classification, the local non-negative matrix factorisation (LNMF) is employed to construct an initial feature space, which is then optimised to calculate more discriminable feature projection vectors of each target. The approaches including speckle noise and stripes suppression, centroid and scale normalisation, LNMF, feature space optimisation with the maximum intersubject variation and minimum intrasubject variation and feature projection vectors calculation are detailed. Finally, the classification is performed with a k neighbours classifier. ISAR images used are obtained by range-Doppler imaging method with radar echoes of aircraft models generated by RadBase. Simulation results show a significant improvement on recognition accuracy and robustness of the proposed method.