Deep Learning-Based Robust Imaging Exploiting 2-D Array Compressive Measurement

This paper proposes a neural network-based robust imaging method from compressive measurements exploiting a two-dimensional (2-D) array. The practical implementation of a 2-D array becomes much more complicated as the number of antennas increases due to the requirement to allocate a different radio frequency front-end circuit to each antenna. An effective solution to this problem is to compress the received signal prior to digitization at the array. In this paper, we use the maximization of the mutual information between compressed measurement and the signal locations to determine the optimal compressive measurement. A neural network-based strategy for localizing sources using these compressed measurements is then proposed. We treat neural network training as a 2-D multilabel classification problem and design an appropriate loss function to train the network. Compared to the conventional approach, the proposed neural network-based approach provides more robust performance as it does not rely on any prior knowledge of received signals and the antenna configuration.

Automated summary

This paper proposes a neural network-based robust imaging method using compressive measurements from a two-dimensional array. The optimal compressive measurement is determined by maximizing the mutual information between the compressed measurement and the signal locations. A neural network-based strategy for localizing sources using these compressed measurements is then proposed. The proposed approach provides more robust performance compared to the conventional approach as it does not rely on any prior knowledge of received signals and the antenna configuration.