WaveCSI-Net: Adaptive Multi-Resolution Wavelet Network for Massive MIMO CSI Feedback

Abstract

Massive MIMO relies on accurate CSI feedback, yet existing deep learning methods often neglect frequency characteristics or rely on computationally expensive global attention. Observations across extensive experimental datasets indicate that CSI energy tends to concentrate in dense low-frequency sub-bands, while sparse high-frequency sub-bands capture essential structural details. To address this problem, we propose WaveCSI-Net, an adaptive multi-resolution network. It adopts a high-frequency-guided asymmetric architecture, where a deep branch extracts features from high-frequency components and a lightweight branch processes low-frequency ones to optimize computation. Furthermore, we introduce a Coordinate-Cross Dilated (CCD) module to efficiently capture cross-shaped correlations and long-range dependencies without heavy self-attention. Extensive experiments on the COST 2100 datasets demonstrate that WaveCSI-Net achieves leading NMSE performance while maintaining lightweight computational complexity.