High-Accuracy CSI Feedback With Super-Resolution Network for Massive MIMO Systems

Abstract

Acquiring accurate channel state information (CSI) is critical for downlink precoding in frequency division duplexity (FDD) massive multiple-input multiple-output (MIMO) systems. In contrast to the traditional compressive sensing (CS) based methods, whose performance is hindered by excessive feedback overhead, this letter proposes a super-resolution network (SRNet) to compress and reconstruct the CSI. Specifically, the SRNet consists of encoder and decoder, where the encoder can transform channel matrices into codewords, and the decoder can restore different levels of spatial frequency features of CSI image based on a modified embedded block residual network (EBRN+). In addition, a principal component mark (PCM) method is proposed before encoding to lighten the encoder at UE. The experiment results show that our proposed model can achieve better performance than the state-of-the-art models with less training parameters and lower computational complexity at UE. Moreover, the superiority of our proposed model becomes much more significant especially under high compression ratio scenarios.