LFSC: A Lite and Fidelity-Enhanced Semantic Communication Scheme

Semantic communication, focusing on the true meaning behind bits, has the potential to enable cross-lingual and cross-domain information exchange. However, deploying semantic communication systems on compute and memory-constrained Internet of Things (IoT) devices faces the challenge of low-fidelity semantics. To this end, we propose a lite and fidelity-enhanced semantic communication (LFSC) scheme, at the uplink transmitter side with limited resources. Considering the extraction error, we design a two-stage semantic extraction strategy, consisting of a lite vision Transformer (ViT) based convolution self-attention mechanism and a lite convolution neural network (CNN) subnet, to extract the global semantics and compensate the neglected detail semantics, respectively. At the uplink receiver side, which is an edge device with adequate computation resources, we utilize a moderate-size recycle attention-based semantic aggregation network as our semantic decoder. Simulation results show that the proposed LFSC can provide high-fidelity semantic extraction ability and significantly reduce the communication overhead with less than 1M parameters.