Deep-Learning Based Gaussian Mixture Likelihoods Joint Source-Channel Coding for Image Transmission.

Yong Fang,Qiqi He,Dandan Ding,Hangguan Shan,Lu Yu,Zhaoyang Zhang

DOI: https://doi.org/10.1109/WCSP58612.2023.10404462

2023-01-01

Abstract:Recent research on joint source channel coding (JSCC) for wireless image communications has achieved great success owing to the employment of deep learning (DL). However, most of the current DL-based JSCC methods improve the model distortion performance under different signal-to-noise ratios (SNRs) and neglect the joint optimization of the model's rate-distortion performance in the encoding and decoding process. Meanwhile, existing JSCC models typically use traditional channel coding for side information transmission, suffering from cliff-effect at low SNRs as well. To overcome these limitations, this paper proposes a new method called GMLJSCC, which utilizes Gaussian Mixture Likelihoods to parameterize the distributions of latent codes for code rate estimation, saving rate in the process of signal compression and transmission. The model also leverages DL-JSCC-based design to transmit side information, which provides accurate estimation for source statistical probability. The reconstruction performance of the model is also improved.