Deep Neural Networks for Landmines Images Classification

Abstract

This paper presents an efficient solution for automatic classification between the Anti-Tank (AT) landmines signatures and standard hyperbolic signatures obtained from other objects, including the Anti-personnel (AP) landmines based on pretrained deep Convolutional Neural Network (CNN). Specifically, two deep learning techniques have been tested and compared with another published landmine classification method. The first technique is based on VGG-16 pertained network for both features extraction and classification from the dataset of landmines images. While the second technique use Resnet-18 pretrained network for features extraction and Support Vector Machine (SVM) used for classification. The proposed algorithm has been tested using dataset of landmines images taken by the Laser Doppler Vibrometer based Acoustic to Seismic (LDV-A/S) landmine detection system. The results show that, the deep learning-based technique give higher classification accuracy than published landmine classification method. The Resnet-18 pretrained network-based and SVM classification gives better average accuracy than VGG-16 pertained network-based classification.