Joint Source-Channel Coding for Image Semantic Communications: A Parallel Processing Approach

Semantic communication is the focus of sixth generation (6G) wireless communication systems, emphasizing the efficient extraction and transmission of semantic information. However, existing image semantic communication systems face challenges in robustly recovering semantic content and mitigating channel noise. To this end, we propose a module based on a parallel processing approach, called STC, which uses feature segmentation to feed semantic information into the CNN and Transformer respectively in parallel. Through this approach, the local information extraction capability of CNN and the non-local information extraction capability of Transformer are efficiently combined, and an improved channel attention mechanism is introduced to incorporate the channel state information into the model training to reduce the unfavorable impact of the channel. Based on STC module, we propose a novel joint source channel coding (JSCC) model. The designed JSCC model improves gradient flow and information exchange through the hierarchical integration of STC and residual blocks. Compared with existing state-of-the-art models and traditional separated source-channel coding schemes, the proposed JSCC model demonstrates superior performance across datasets of varying resolutions.