Robust Semantic Communication System for Image Transmission
Likai Zhao, Ruofan Wang, Qi Cao, Guanchong Niu, Baoming Bai · 2024
This paper proposes a robust semantic communication system. Initially, the sender extracts the skeleton map from the transmitted image. The coordinates of the skeleton points are then subjected. In this process, the coordinates of important semantic information are transmitted with higher power, while less important coordinates are sent with lower power. The modulated coordinates are transmitted through a Gaussian white noise channel. At the receiver, a demodulator restores the skeleton map, and a conditional diffusion model is used to reconstruct the image. Compared to transmitting entire images, transmitting only the coordinates of skeleton points significantly saves band-width. The non-uniform modulation protects critical semantic information, enhancing robustness. Experimental results show that this method reduces the volume of the transmission data from 180kbits to just 5bits. The compression rate is around 10−6. In low SNR environments, the PSNR curve steadily increases as the SNR increases, demonstrating the robustness of the system. Our code is available at https://github.com/agougougoua/RSCT.