Semantic Communication with Pre-trained Deep Learning Models: A Feasibility Study

Recently, semantic communication (SemCom) has made great progress due to the development of deep learning (DL), which has achieved great success in computer vision (CV) and natural language processing (NLP). Although existing semantic communication frameworks can significantly improve transmission efficiency and reliability, most of these architectures did not make full use of existing powerful DL models in the CV field. This motivates us to develop a semantic communication system using existing well-designed and trained models. In this paper, we utilize a pre-trained transformer-based model called IPT, which was proposed for the CV task and well-trained, to establish our SemCom system for image transmission, denoted as IPT. Experimental results show that IPT can recover images with high quality under a variety of signal-to-noise ratio (SNR) conditions and exhibits graceful performance degradation of the reconstruction quality with channel SNR. The results demonstrate that building a semantic communication system with a powerful pre-trained deep learning model is feasible.