Unsupervised Multiattention Domain Adaptive Decluttering Model for Metal Pipe Targets in GPR Images

The propagation of electromagnetic waves underground is subject to the heterogeneity of the medium and is exceedingly complex. In ground-penetrating radar (GPR) images with metal pipe targets, distinguishing target information clearly while decluttering interference under unsupervised conditions poses a significant challenge. Previous methods primarily leverage subspace method to decompose echo data into target and clutter information. However, due to the strong coupling of target and clutter signals in both time and frequency domains, effective separation remains difficult. To address this issue, this article proposes a unsupervised multiattention domain adaptation (UMDA-net) model capable of transforming targets from domains with complex clutter to domains with clean backgrounds. The overall structure consists of two U-net-based generators and a discriminator, guided by adversarial loss and consistency loss functions during training. Furthermore, multiple attention mechanisms, such as channel attention (CA), spatial attention (SA), and multihead self-attention (MSA), are integrated into the generators, enhancing semantic discrimination capability and suppressing clutter information similar to target echoes. Finally, both simulation and real-world experiments validate the effectiveness of the proposed model, and ablation studies demonstrate the contributions of each attention mechanism in clutter suppression. This research provides an effective solution for processing GPR images with metal pipe targets under unsupervised conditions.