Semantic Communication for Image Transmission using Pre-trained Transformer Models

Recently, significant progress has been made in semantic communication (SemCom) thanks to the advancements in deep learning (DL) techniques, particularly in computer vision (CV) and natural language processing (NLP). However, most existing SemCom frameworks fail to fully leverage the power of DL models in the field of CV. This has inspired us to develop a SemCom system that utilizes a well-designed and trained DL model.

In this paper, we propose to use a pre-trained transformer-based model called IPT, which was initially designed and trained for CV tasks, to establish our SemCom system for image transmission. We refer to this system as 'ipt'. Our experimental results demonstrate that 'ipt' can effectively restore images with high quality under various signal-to-noise ratio (SNR) conditions. Moreover, the reconstruction quality of 'ipt' gracefully degrades with decreasing channel SNR. These results clearly indicate the feasibility of building a semantic communication system using powerful pre-trained DL models.