Digital semantic communication has garnered significant research attention due to its potential to surpass traditional transmission methods. However, current VQ-based approaches are constrained by fixed codebooks, limiting rate flexibility and failing...
Digital semantic communication has garnered significant research attention due to its potential to surpass traditional transmission methods. However, current VQ-based approaches are constrained by fixed codebooks, limiting rate flexibility and failing to address semantic importance during transmission. To address these dual challenges of structural rigidity and transmission inefficiency, we introduce ResUME, a novel and efficient end-to-end digital semantic communication framework.
ResUME integrates Residual Vector Quantization (RVQ) with Unequal Modulation Encoding (UME). By leveraging the hierarchical structure of RVQ, the framework mitigates semantic degradation even at high compression ratios. By aligning modulation orders with the hierarchical importance of residuals, UME optimizes transmission efficiency and reconstruction performance.
To enhance training reliability and ensure robust transmission, we propose a Channel-Aware Noise Substituting VQ (CA-NSVQ) technique. CA-NSVQ stabilizes training by modeling combined distortion as a single Gaussian, while the loss function optimizes the Evidence Lower Bound (ELBO) within a variational framework.
Our experimental evaluations confirm that ResUME, particularly when combined with UME, consistently outperforms existing VQ-based frameworks in reconstruction quality. Furthermore, it demonstrates flexible adaptation to diverse SNR conditions via modulation optimization, enabling effective rate control within a single, fixed backbone network even under strict channel bandwidth ratio (CBR) constraints.