DualFormer: A Joint Channel Group and Spatial Transformer Framework for Image Reconstruction in Semantic Communication

Semantic communication aims to improve wireless image transmission by focusing on task-relevant semantics rather than precise pixel recovery. However, existing deep learning methods often prioritize spatial feature modeling while overlooking semantic dependencies across channels. To address this limitation, we propose a two-stage end-to-end framework termed DualFormer, which combines a Channel Group Transformer (CGT) and a Spatial Transformer (ST) to jointly capture intra-patch semantic dependencies and spatial context. CGT employs grouped attention across channels, while ST integrates spatial cues via cascaded attention. Experiments on CIFAR-10 and Kodak under various channel conditions indicate that DualFormer tends to achieve better PSNR and MS-SSIM than representative baselines, especially in low-SNR and highcompression scenarios, showing promising robustness and adaptability.