Deep Learning based Synthetic Aperture Imaging in the Presence of Phase Errors via Decoding Priors
Deep Learning based Synthetic Aperture Imaging in the Presence of Phase Errors via Decoding Priors
复制标题
DOI：
10.1109/radarconf2351548.2023.10149740
复制
发表时间：
2023-05
期刊：
影响因子：
--
通讯作者：
Samia Kazemi;Bariscan Yonel;B. Yazıcı
中科院分区：
文献类型：
--
作者：
Samia Kazemi;Bariscan Yonel;B. Yazıcı
在本文中，我们设计了一种基于深度学习（DL）的方法，用于存在相位误差的合成孔径成像。由于不可预见的环境变化、传感器位置的波动以及背景介质中的多重散射效应而导致的传输介质中的随机变化通常会导致假定数据模型中的不确定性。依赖于反投影估计的成像算法在这些情况下容易受到估计误差的影响。此外，在介质的动态性质下，在相同的操作条件下收集大量的测量结果可能变得具有挑战性。为此，我们的成像网络在三个主要步骤中结合了DL：首先，我们实现了深度网络（DN），用于预处理错误的测量;其次，我们通过恢复与散射介质相关联的反射率矢量的编码版本来实现基于DL的解码先验，以降低样本复杂度，然后通过解码DN将其映射到图像估计;最后，我们考虑一个固定的步骤实现的迭代算法的形式的递归神经网络（RNN）通过使用展开技术，导致基于模型的成像算子。所有三个DN的参数以监督的方式同时学习。我们使用模拟高保真合成孔径测量验证了我们的方法的可行性。
In this paper, we designed a deep learning (DL) based method for synthetic aperture imaging in the presence of phase errors. Random variations in the transmission medium resulting from unforeseen environmental changes, fluctuations in sensor locations, and multiple scattering effects in the background medium often amount to uncertainties in the assumed data models. Imaging algorithms that rely on back-projected estimates are susceptible to estimation errors under these circumstances. Moreover, under dynamic nature of the medium, collecting high volume of measurements under the same operating conditions may become challenging. Towards this end, our imaging network incorporates DL in three major steps: first, we implement a deep network (DN) for pre-processing the erroneous measurements; second, we implement a DL-based decoding prior by recovering an encoded version of the reflectivity vector associated with the scattering media to reduce sample complexity, which is then mapped to an image estimate by a decoding DN; finally, we consider a fixed step implementation of an iterative algorithm in the form of a recurrent neural network (RNN) by using the unrolling technique that leads to a model-based imaging operator. The parameters of all three DNs are learned simultaneously in a supervised manner. We verified the feasibility of our approach using simulated high fidelity synthetic aperture measurements.