Image Semantic Communication over Fading Channel: A Learned Broadcast Approach
Kangning Ma, Shuo Shao, Meixia Tao · 2023
We consider the image semantic communication in a point-to-point block fading Gaussian channel, with a finite number of channel states. A deep learning-aided broadcast approach is proposed to complete the semantic transmission of different qualities according to the channel state. An autoencoder (AE) is deployed to guide the encoding and decoding of the superposition code within this approach. We introduce some sub-constellations to designate the layer mapping in this superposition coding and generate a super-constellation through the symbol-by-symbol superposition. This super-constellation will be designed for allowing recovering of the semantic message with an adaptive decoding rate at the receiver side. We optimize its geometry and power allocation during training, to minimize the average communication error across all states under a total power constraint. The example exhibits our proposed scheme can adaptively adjust the coding in response to the noise environment and achieve a decent error performance in each channel state.