Joint Deep Learning for Simultaneous Clutter Removal and Buried Object Detection in GPR

Abstract

Ground Penetrating Radar (GPR) data presents a challenging problem for detecting subsurface targets due to surface reflections and the complex clutter caused by heterogeneous soil structures. While traditional methods treat clutter removal and target detection as separate processes, this study presents an integrated deep learning approach that simultaneously optimizes both tasks. In the first phase of the study, the first proposed model, Dec-YOLO (Model I), demonstrated that 'joint training' of clutter removal networks (UNet, CR-Net, DC-ViT) from the literature with a detection network improves detection performance compared to sequential methods. Building on this finding, the second phase proposes the original RAFDeC-YOLO (Model II) architecture. This architecture features a specialized denoising block (decoder) that branches off from the standard YOLOv12 backbone. The fundamental innovation of this branch is that it feeds back the cleaned and enriched feature maps it produces back to the relevant neck layers of the YOLO architecture via the proposed Residual Adapter Fusion mechanism. This strategic feature transfer maximizes discriminative power, particularly in challenging scenarios such as weak dielectric targets and asphalt-covered surfaces, by enabling the detection network to access both noisy raw data and cleaned spatial details. Experimental results demonstrate that the proposed framework outperforms state-of-the-art methods, achieving over 25.8% higher localization accuracy on hybrid datasets and an 87.5% improvement in challenging real-world scenarios, while reducing computational complexity by approximately 43% for efficient deployment. © 2013 IEEE.