PAW: Parallel-Wavelet Attention for Lightweight Asymmetric Semantic Image Communication in 6G Space Networks

An important challenge in semantic communication for 6G space networks is achieving reliable image reconstruction under low-SNR conditions with constrained onboard resources. Existing Swin-based schemes rely on symmetric transceivers and spatial-only attention, which are not well suited to satellite--cloud systems. To address these issues, we propose PAW, a Parallel-Wavelet Attention-based asymmetric architecture for semantic image communication. PAW integrates discrete wavelet transform (DWT) to enable high-fidelity channel and spatial attention through parallel frequency-aware recalibration paths. Lightweight MBConv and DSConv modules are further employed to reduce encoder-side computational complexity. Experiments show that PAW reduces encoder-side GPU runtime by up to 3.5 times compared to Swin-based baselines, while maintaining competitive reconstruction performance. These results suggest that joint channel & spatial attention is an effective design for future resource-constrained 6G space semantic communications.