Generative Joint Source-Channel Coding for Semantic Image Transmission

Recent works have shown that joint source-channel coding (JSCC) schemes using deep neural networks (DNNs), called DeepJSCC, provide promising results in wireless image transmission. However, these methods mostly focus on the distortion of reconstructed signals with respect to input image, rather than their perception by humans. focusing traditional metrics alone does not necessarily result high perceptual quality, especially extreme physical conditions, such as very low bandwidth compression ratio (BCR) and signal-to-noise (SNR) regimes. In this work, we propose two novel JSCC leverage quality generative models (DGMs) for transmission, namely InverseJSCC GenerativeJSCC. While former is an inverse problem approach latter end-to-end optimized scheme. both, optimize a weighted sum mean squared error (MSE) learned patch similarity (LPIPS) losses, which capture more semantic similarities other metrics. performs denoising distorted reconstructions DeepJSCC model solving optimization pre-trained style-based adversarial network (StyleGAN). Our simulation show significantly improves state-of-the-art terms edge cases. GenerativeJSCC, carry out training encoder StyleGAN-based decoder, GenerativeJSCC outperforms both quality.