VAE-GAN-Based Semantic Communication for High-Quality Image Transmission
Naoki Omi, Shun Kojima, Chang-Jun Ahn · IEEE Transactions on Wireless Communications · 2025
In semantic communication systems, deep learning-based joint source-channel coding (DeepJSCC) has demonstrated superior performance, particularly in low signal-to-noise ratio (SNR) scenarios and under limited bandwidth conditions, compared to traditional communication technologies. However, most existing studies rely on communication models based on autoencoders (AE), where the distribution of transmission symbols is treated as a complete black box, causing inefficient use of limited transmission symbols. Moreover, many existing methods for image transmission focus on optimizing pixel-wise metrics, without considering the semantic information of the images. This pixel-level optimization often compromises the semantic fidelity and perceptual quality of the reconstructed images. To address these problems, we propose a novel semantic communication system that combines a variational autoencoder (VAE) and a generative adversarial network (GAN). Specifically, a VAE is used to arbitrarily control the distribution of transmission symbols according to channel conditions, while a GAN is used to maximize the similarity of semantic information. Simulation results demonstrate that the proposed method improves the perceptual quality of the reconstructed images compared to conventional approaches.