Deep Joint Source-Channel Coding Based on Signal Space Diversity and Correlation Interleaving

Semantic communication (SemCom) has received much attention as a cutting-edge paradigm. However, in Sem-Com, encoded semantic symbols exhibit strong correlations. When deep fading occurs, these symbols are susceptible to consecutive errors that impair their inter-symbol correlation, consequently degrading decoding performance at the receiver. To address it, a deep joint source-channel coding framework based on signal space diversity and correlation interleaving for SemCom is proposed. For various transmission tasks, the semantic coder extracts distinct features of semantic information, resulting in constellation points with non-uniform probability distributions. Leveraging the neural networks, it performs adaptive angle rotation by capturing the intrinsic characteristics of these points. Subsequently, the correlation in the semantic information is mined utilizing the multi-head attention mechanism, and then it is integrated with imperfect channel state information for component interleaving, thus avoiding the simultaneous fading of strongly correlated symbols. Experiments show that the framework improves the classification accuracy and image reconstruction quality over fading channels, verifying its anti-fading and robustness advantages.