A GPR Inversion Network Fusing Multi-Scale Convolution and Hybrid Attention

To address the degradation of target inversion accuracy and algorithmic stability caused by complex background clutter interference in GPR data, this paper proposes MSA-UNet, an enhanced inversion network fusing multi-scale convolution and hybrid attention mechanisms. Built upon the standard encoder-decoder architecture, the proposed network first utilizes a multi-scale convolution module to capture target features at varying scales in parallel, then employs attention gates to spatially suppress background clutter while focusing on authentic target regions, and finally integrates a channel attention module (Squeeze-and-Excitation Block) at the bottleneck layer to adaptively recalibrate feature channel weights for information enhancement. Comparative results demonstrate that MSA-UNet achieves superior inversion accuracy and robustness over Pix2Pix and standard U-Net when processing clutter-laden GPR B-scan images.