Impact of Training Models on Deep Joint Source-Channel Coding Applicable to 5G Systems

Abstract

With the development of 5G technology and the proliferation of IoT devices, Deep Joint Source-Channel Coding (DeepJSCC) has attracted attention for efficiently transmitting video and image data. DeepJSCC can maintain a good peak signal-to-noise ratio (PSNR) of images even at a meager signal-to-noise ratio (SNR). In cellular communication systems, the compression ratio must adapt to channel fluctuations, requiring multiple training models at the base station. However, the optimal SNR and compression ratio combination during training has yet to be reported. This paper investigates the necessary number of training models by stepwise varying SNR and compression ratio during training.