地下埋藏目标与分层粗糙面复合散射探地雷达回波特性研究

探地雷达（ground penetrating radar，GPR）是一种重要的无损检测工具，被广泛应用于道路缺陷检测、路面以下物体检测以及非结构化地形感知等领域. 针对基于深度学习的介电常数图反演方法特征提取不全、准确度不高的局限性，本文引入注意力机制, 提出了一种GPR图像智能反演算法，设计了适用于GPR图像反演的神经网络InvNet以完成从GPR图像到介电常数图的非线性映射. 该算法首先利用时间维度卷积和全局特征残差编码器提取GPR时间维度和全局特征，接着采用改进的注意力机制作为注意力语义提取模块提取特征图的全局语义信息，最终通过介电常数解码器完成GPR介电常数图重构. 在基于时域有限差分法构建的GPR反演数据集上，InvNet反演结果的结构相似性指数（structural similarity index measure，SSIM）达到了97.14%，均方误差（mean square error, MSE）降为0.003 0，平均绝对误差（mean absolute error，MAE）降为0.015 7；相比于最新GPR反演网络算法GPRInvNet和PINet，InvNet的SSIM分别提高了1.06%和0.86%，MSE分别降低了0.24%和0.25%，MAE分别降低了1.41%和1.02%. 实验表明本文方法能够根据GPR图像有效反演出对应的介电常数模型图，并且与现有方法的对比结果展示了其优越的反演性能.

Echo characteristics of GPR detecting of multiple buried targets and layered rough interface

Ground penetrating radar (GPR) is an important nondestructive testing tool, which is widely used in the fields of road defect detection, subsurface object detection and unstructured terrain perception. To overcome the limitations of the incomplete feature extraction and the low accuracy of deep learning based inversion method, an intelligent inversion algorithm with attention mechanism for GPR image is proposed, and a neural network named InvNet for GPR image inversion is designed to complete the task of nonlinear mapping from GPR image to permittivity image. Firstly, the time dimension convolution and the global feature residual encoder are used to obtain the time dimension and global feature of the GPR image. Then, an improved attention mechanism is used as an attention semantic extraction module to extract global semantic information from the feature map. Finally, the reconstruction of permittivity image is completed by the decoder of permittivity. The GPR inversion dataset is constructed based on the finite difference time domain method and the experiments on the dataset show that the structural similarity index measure (SSIM) of InvNet reached 97.14%, mean square error (MSE) decreased to 0.0030 and mean absolute error (MAE) decreased to 0.015 7. Compared with the latest GPR inversion network algorithms GPRInvNet and PINet, SSIM of InvNet is increased by 1.06% and 0.86%, MSE is reduced by 0.24%, 0.25% and MAE is reduced by 1.41% and 1.02% respectively. The results show that this method can effectively reverse the GPR image to the corresponding permittivity image and the comparison results with the existing methods show its superior inversion performance.