Evaluation of SVM classification of metallic objects based on a magnetic-dipole representation

Abstract

In the electromagnetic-induction (EMI) detection and discrimination of unexploded ordnance (UXO) it is important for inversion purposes to have an efficient forward model of the detector-target interaction. Here we revisit an attractively simple model for EMI response of a metallic object, namely a hypothetical anisotropic, infinitesimal magnetic dipole characterized by its magnetic polarizability tensor, and investigate the extent to which one can train a Support Vector Machine (SVM) to produce reliable gross characterization of objects based on the inferred tensor elements as discriminators. We obtain the frequency-dependent polarizability tensor elements for various object characteristics by using analytical solutions to the EMI equations. Then, using synthetic data and focusing on gross shape and especially size, we evaluate the classification success of different SVM formulations for different kinds of objects.