CRissNet: An Efficient and Lightweight Network for CSI Feedback in Massive MIMO Systems

The attractive high spectrum and energy efficiency performance of massive multiple-input multiple-output (MIMO) highly relies on the availability of the channel state information (CSI). In frequency-division duplexing (FDD), the high dimensionality of the CSI matrices brings about the significant feedback overhead, while deep learning (DL)-based data compression has shown great promise with limited communication resources. In this paper, regarding the spatial-temporal correlation property of CSI matrices, we propose a low-parameter neural network named CRissNet, which utilizes the Criss-Cross attention to comprehensively extract the essence of CSI matrices. A new criss-cross correlation matrix (CCCM) is devised to measure the information importance in the angular-delay domain and provide additional explanatory insight for the applicability of the proposed attention scheme. Additionally, to enhance the weak correlation condition performance in the outdoors, we design an extended Criss-Cross attention algorithm, named Criss-Cross attention+, which enhances the self-attention with side-view observation of adjacent antennas. Some training methods of the real and imaginary part combination, variable sizes design, etc., are conducted, but not limited to the specific feedback neural networks. The simulations show the superiority of the proposed feedback scheme over the state-of-the-art methods for both indoor and outdoor scenes especially in high compress ratio condition. The open source codes are available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/CRissNet/CRissNet.git</uri>.