Landmine Identification From Pulse Induction Metal Detector Data Using Machine Learning
Marko Šimić, Davorin Ambruš, Vedran Bilas
Abstract
The significant presence of metallic clutter in minefields results in a high false alarm rate of the conventional metal detectors due to their inability to discriminate between metallic parts of a landmine and nonhazardous clutter. In this letter, we present electromagnetic induction (EMI) based system for identification of small hidden metallic objects. A 1-D convolutional neural network is employed to infer an object class from time-domain magnetic polarizability tensor features. The proposed approach is experimentally validated under laboratory conditions on the dataset, including two types of landmines and five types of metallic clutter. The model trained on simulations and tested on measurements achieves 98% accuracy (with zero false negatives) for both multiclass and binary “threat or nonthreat” classification problems.