用神经网络校准雷达数据，精准重建物体介电属性。
Data-Driven Calibration Technique for Quantitative Radar Imaging
- 构建神经网络预测散射场，反推校准因子。
- 在Fresnel数据集上实现介电常数重建误差<5%。
- 适合雷达成像、无损检测等需要高精度电磁参数的应用。
定量反演算法可重建场景中每一点的电学属性（如介电常数和电导率），但在实测数据上应用困难，因无法获知场景中的入射波场。这一信息通常未知，受天线特性、路径损耗、波形等因素影响。本文引入一个标量校准因子以补偿这些因素。为求解该因子，我们通过训练一个简单的前馈全连接神经网络，学习从介电常数分布到雷达处散射场的映射关系，并将正向问题嵌入反演流程中。通过最小化实测与仿真散射场之间的差异，优化每个发射源的校准因子。我们在Fresnel Institute数据集上验证了该方法的有效性。
原文摘要 · Abstract (English)
Quantitative inversion algorithms allow for the reconstruction of electrical properties (such as permittivity, and conductivity) for every point in a scene. However, they are challenging to use on measured datasets due to the need to know the incident wave field in the scene. In general, this is unknown due to factors such as antenna characteristics, path loss, waveform factors, etc. In this paper, we introduce a scalar calibration factor to account for these factors. To solve for the calibration factor, we augment the inversion procedure by including the forward problem, which we solve by training a simple feed-forward fully connected neural network to learn a mapping between the underlying permittivity distribution and the scattered field at the radar. We then minimize the mismatch between the measured and simulated fields to optimize the scalar calibration factor for each transmitter. We use the Fresnel Institute dataset to test our algorithm.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。