Detection and Classification of Land Mines from Ground Penetrating Radar Data Using Faster R-CNN

Abstract

A solution for automatic detection and classification of buried objects by implementing Faster Region Convolutional Neural Network (Faster R-CNN) with Ground Penetrating Radar (GPR) system is presented. Specifically, Faster R-CNN Inception-v2 was chosen, as a compromise between computational load and accuracy, compared with other Faster R-CNNs. Although the solution is general, in the sense that it can be retrained for arbitrary number of classes, here we focus on the discrimination between anti-tank (AT) mines signatures and standard hyperbolic signatures obtained from other objects, including anti-personnel (AP) mines. The image dataset used for training and testing the R-CNN network consists of GPR B-scans obtained both by gprMax based simulations and from real measured GPR data. The method performance is evaluated using Confusion matrices and ROC curves. Post processing approach based on object size and depth below ground surface enables discrimination of AP mines.