基于生成对抗网络的两阶段探地雷达图像反演方法

      Two-Stage GPR image inversion method based on Generative Adversarial Network

      • 摘要: 在探地雷达(ground penetrating radar, GPR)应用中,反演成像是解译GPR数据信息的关键技术。现有基于深度学习的GPR图像反演技术大多应用于地下均匀介质的理想环境,然而真实环境中采集到的数据通常包含复杂的噪声与杂波信号,对反演精度有很大影响。针对这一问题,本文提出了一种基于生成对抗网络(generative adversarial network, GAN)的两阶段GPR图像反演网络TSInvNet,以重构真实环境中地下目标的位置分布。该方法先将GPR B-scan图像使用改进的空间自适应归一化(spatially-adaptive normalization, SPADE)生成器的去噪网络TSInvNet1进行处理后,接着送入引入置换注意力(shuffle attention, SA)模型的反演网络TSInvNet2进行反演。在模拟数据与真实数据上的实验结果表明,TSInvNet能够根据GPR B-scan图像准确反演出地下目标的位置,在具有复杂噪声与多目标情况下的反演应用中具有强鲁棒性和精确反演性能。

         

        Abstract: In the application of ground penetrating radar (GPR), the inversion imaging is the key technology for interpreting the data information of GPR. Existing GPR image inversion techniques based on deep learning are mostly applied to the ideal environment of underground homogeneous media. However, the data collected in real environments usually contain complex noise and clutter signals, which greatly affect the accuracy of the inversion. To address this issue, this paper proposes a two-stage GPR image inversion network based on Generative Adversarial Network (GAN), named TSInvNet, to reconstruct the spatial distribution of underground targets in real environments. This method processes the GPR B-scan images through a denoising network, TSInvNet1, using an improved Spatially-Adaptive Normalization (SPADE) generator, and then inputs the processed images into an inversion network, TSInvNet2, which introduces a Shuffle Attention (SA) model for inversion. Experimental results on simulated and real data demonstrate that TSInvNet can accurately invert the underground target positions based on GPR B-scan images, showing strong robustness and precise inversion performance in applications involving complex noise and multiple targets.

         

      /

      返回文章
      返回