Deep joint source-channel coding for wireless image transmission with auxiliary semantics

Semantic communication, as an emerging paradigm, has achieved significant success by combining deep learning (DL) with joint source-channel coding (DeepJSCC). However, Current methods struggle to fully adapt to the dynamic complexity of semantic information, often suffering from loss of key details due to feature homogeneity and insufficient noise robustness during transmission, especially in environments with low signal-to-noise ratio (SNR) and high compression scenarios (CR). Therefore, we propose a novel dual-semantic framework for semantic communication, the deep joint source-channel coding for wireless image transmission with auxiliary semantics (SAJSCC), which employs a guidance vector generation network (GVNet) to extract image-specific guidance vectors, and explicitly constructs complementary spaces for the primary feature stream and auxiliary semantics. The decoder integrates a hierarchical feature guidance mechanism, adopting dual decoding strategies to dynamically balance computational efficiency and reconstruction fidelity. Furthermore, we design a concurrent channel-and-spatial attention module (CCSAM) that suppresses noise interference through joint channel–spatial re-weighting. Experimental evaluations show that the proposed SAJSCC framework significantly surpasses existing approaches, delivering a peak signal-to-noise ratio (PSNR) gain of 5.02 dB and a structural similarity index measure (SSIM) improvement of 0.15 under 0 dB SNR conditions. Our framework significantly enhances the recovery of complex textures, establishing a new paradigm for highly robust semantic communication.