LECA-Net: A Context-Aware Local Enhancement Network for Massive MIMO CSI Feedback

Acquiring accurate downlink channel state information (CSI) at the base station (BS) is crucial for massive multiple-input multiple-output (MIMO) system. This is particularly challenging in frequency division duplexing mode, where user equipment needs feedback CSI to the BS. In addition, the large number of BS antennas further exacerbates CSI feedback overhead. Recently, deep learning has demonstrated better performance and efficiency for massive MIMO CSI feedback. In this paper, we propose a novel context-aware local enhancement network, termed as LECA-Net, where the network incorporates a context-aware local enhancement block to extract high-frequency and low-frequency spatial information from CSI image. We also design a multi-scale feature extraction block to extract CSI features with different receptive field sizes. Additionally, gated recurrent unit is utilized to capture the long-term dependencies among CSI data. To reduce computational complexity, we replace standard convolution with blueprint separable convolution. Simulation results show that the proposed LECA-Net achieves the best feedback performance with lower computational complexity compared to state-of-the-art methods. Additionally, compared to the existing TransNet method, our proposed LECA-Net achieves average gains of 8.71% and 29.19%, respectively in indoor and outdoor scenarios, as well as reduces computational complexity by 20.44%.