Digital Semantic Communication System With a Learnable Channel Estimator for OFDM Transmission

This is the authors' abstract. We don't add key points for this paper.

Semantic communication (SemCom) based image transmission schemes, in contrast to conventional bit-based schemes, achieve a substantial reduction in transmission data volume while maintaining superior image quality under conditions of limited data due to their inherent robustness against channel noise. However, most existing SemCom systems are based on analog communication paradigms, allowing for the transmission of arbitrary complex values over the channel. The analog communication model may lead to incompatibility with digital commercial hardware and standardized protocols. In this paper, we propose an end-to-end digital SemCom system with a learnable channel estimator (DiSeLCE) aligning with the constraints of digital communication systems. Specifically, we first design a novel semantic encoder and decoder that integrate dense blocks to enhance feature extraction and significantly reduce the number of model parameters. Moreover, we employ a soft-to-hard quantization method to effectively mitigate the discontinuity in the model gradient due to quantization mappings. Furthermore, both the domain knowledge and residual network (ResNet) are leveraged to develop a learnable channel estimation model, further enhancing the robustness of the proposed system in frequency-selective fading channels for orthogonal frequency-division multiplexing (OFDM) transmission. Experimental results demonstrate that the proposed DiSeLCE performs well under frequency-selective fading channels, facilitating efficient image transmission even under low signal-to-noise ratios (SNR) conditions. Furthermore, DiSeLCE exhibits robust generalization capabilities across different types of datasets.