Content-aware robust semantic transmission of images over wireless channels with GANs

Semantic Communication (SemCom) can significantly reduce the transmitted data volume and keep robustness. Task-oriented SemCom of images aims to convey the implicit meaning of source messages correctly, rather than achieving precise bit-by-bit reconstruction. Existing image SemCom systems directly perform semantic encoding and decoding on the entire image, which has not considered the correlation between image content and downstream tasks or the adaptability to channel noise. To this end, we propose a content-aware robust SemCom framework for image transmission based on Generative Adversarial Networks (GANs). Specifically, the accurate semantics of the image are extracted by the semantic encoder, and divided into two parts for different downstream tasks: Regions of Interest (ROI) and Regions of Non-Interest (RONI). By reducing the quantization accuracy of RONI, the amount of transmitted data volume is reduced significantly. During the transmission process of semantics, a Signal-to-Noise Ratio (SNR) is randomly initialized, enabling the model to learn the average noise distribution. The experimental results demonstrate that by reducing the quantization level of RONI, transmitted data volume is reduced up to 60.53% compared to using globally consistent quantization while maintaining comparable performance to existing methods in downstream semantic segmentation tasks. Moreover, our model exhibits increased robustness with variable SNRs.