Transfer Learning for Classification of GPR B-scan Images from Clutter using Vision Transformer and VGG-16

Large areas of land are contaminated with landmines as a result of ongoing wars and previous conflicts, making accurate detection of these landmines crucial to community safety and humanitarian efforts. Deep Learning models have shown to be incredible at detection and classification tasks, but also require huge amounts of training data. It is challenging to obtain large amounts of domain-specific data for deep learning, so to overcome this challenge, transfer learning can be used. The Transfer Learning pipeline for buried landmine identification that we have presented in this study is based on a pre-trained convolutional neural network (CNN), called Visual Geometry Group (VGG16), and a pre-trained Vision-Transformer (ViT) model that is applied to ground penetrating radar (GPR) pictures. The proposed models have been fine-tuned on a custom GPR dataset generated using gprMax by simulation. The findings demonstrate the suggested algorithm’s ability to correctly identify buried objects as either landmines or uninteresting things (clutter). The two models - VGG16 and ViT were capable of achieving accuracies of.96 and.94 respectively.