Quantized Compressed Sensing via Deep Neural Networks

Abstract

Compressed sensing (CS) is an efficient technique to acquire sparse signals in many wireless applications to, e.g., reduce the amount of data and save low-power sensors' batteries. This paper addresses efficient acquisition of sparse sources through quantized noisy compressive measurements where the encoder and decoder are realized by deep neural networks (DNNs). We devise a DNN based quantized compressed sensing (QCS) method aiming at minimizing the mean-square error of the signal reconstruction. Once trained offline, the proposed method enjoys extremely fast and low complexity decoding in the online communication phase. Simulation results demonstrate the superior rate-distortion performance of the proposed method compared to a polynomial-complexity QCS reconstruction scheme.