Clutter Removal in Ground-Penetrating Radar Images Using Deep Neural Networks

Abstract

The clutter in ground-penetrating radar (GPR) images obscures and disguises subsurface target reflections, which greatly challenges the accurate target identification. Conventional clutter removal methods suffer from limited clutter removal capability. They either leave residual clutter or deteriorate target reflections. To address the challenges in suppressing clutter in GPR radargrams, we present a deep learning-based method that leverages the powerful learning capability of the deep neural network to remove clutter in diverse real-world scenarios. The network takes the raw GPR radargram as the input, preserves the information related to target reflections and eliminates unwanted clutter features in an encoder-decoder manner, and finally reconstructs the clutter-free radargram. Experimental results demonstrate that the well-trained network successfully removes clutter and restores target reflections with consistent high performance in various real-world scenarios.