Global-Local Knowledge-Assisted Joint Source–Channel Coding

This paper presents a novel knowledge-assisted joint source-channel coding (JSCC) scheme that integrates implicit neural compression (INC) with analog residual transmission for high-quality and efficient image delivery. To improve compression efficiency and decoding delay, we introduce a knowledge base architecture that separates learning and inference: the global knowledge, implemented on the cloud, compresses image-specific latent representations and lightweight network parameters; the local knowledge, deployed at the edge, reconstructs side information using lightweight inference from the received bitstream. Although the side information enables coarse reconstruction, it often lacks the fine-grained visual details, i.e., high-frequency components. To supplement the missing high-frequency details, the residuals between the original and side information are transmitted via pseudo-analog modulation and reconstructed at the receiver. The residual transmission not only enables high-quality reconstruction of high-frequency components but also provides graceful improvement under instantaneous channel conditions. Experimental results on the Kodak and CLIC2020 datasets demonstrate that the proposed scheme significantly outperforms conventional and learned image codecs in terms of reconstruction quality and robustness to unreliable channel conditions.