Design of Adaptive Thresholds For One-Bit Radar Imaging Based on Adversarial Samples

Abstract

In this paper, we present a new scheme of adaptive quantization thresholds for one-bit radar imaging based on adversarial samples. Radar imaging with one-bit compressive sensing (CS) is attractive due to the small storage burden and low requirements to the analog-to-digital converter. However, conventional one-bit quantization scheme with fixed thresholds does not use the magnitude information, possibly leading to difficulty in energy estimation and higher amplitude recovery error. Recently, adaptive thresholds methods have been developed to deal with the limitation of fixed thresholds scheme. Based on the adversarial training theory, the proposed new method embeds adversarial samples into the binary iterative hard thresholding (BIHT) algorithm and exploits an adaptive thresholds scheme based on the adversarial samples to improve the model robustness and imaging quality with one-bit coded data. Simulation results demonstrate that the proposed method outperforms the BIHT with fixed thresholds in one-bit radar imaging.