Deep CSI Compression for Massive MIMO: A Self-Information Model-Driven Neural Network

In order to fully exploit the advantages of massive multiple-input multiple-output (mMIMO), it is critical for transmitter accurately acquire channel state information (CSI). Deep learning (DL)-based methods have been proposed CSI compression and feedback transmitter. Although most existing DL-based consider matrix as an image, structural features image are rarely exploited in neural network design. As such, we propose a model self-information that dynamically measures amount contained each patch from perspective features. Then, by applying model, model-and-data-driven feedback, namely IdasNet. The IdasNet includes design module deletion selection (IDAS), encoder informative feature (IFC), decoder recovery (IFR). particular, model-driven IDAS pre-compresses removing redundancy terms self-information. IFC then conducts pre-compressed generates codeword which contains two components, i.e., values position indices values. Subsequently, IFR decouples well recover image. Experimental results verify noticeably outperforms networks under various ratios while has number parameters reduced orders-of-magnitude compared with methods.