WTSCNet: A CSI Feedback Network Based on Wavelet Transform Convolution and Attention Mechanism

To enhance the performance of massive MIMO systems, efficient downlink CSI compression and feedback are crucial in FDD mode. Deep learning (DL)-based methods surpass traditional compressed sensing but often rely on CNNs designed for image processing, neglecting essential channel and spatial information. This paper proposes WTSCNet, a novel CSI feedback network integrating wavelet transform convolution and attention mechanisms to balance network complexity and feature extraction. The encoder employs wavelet transform convolution for multi-resolution feature extraction, improving spatial information capture while reducing computational cost. The decoder incorporates the CARBlock module to enhance multi-scale and spatial-channel feature integration. Experimental results show that WTSCNet outperforms CNN-based methods like CRNet, achieving a 6.41 dB improvement in reconstruction accuracy at low compression ratios, while reducing complexity by 9.9M parameters compared to attention-based TransNet+ with a 0.3 dB accuracy gain. The proposed model offers a robust and efficient solution for CSI compression feedback.