Robust End-to-End Image Transmission with Residual Learning

TL;DR

This work proposes a layered image transmission scheme at the application layer (AL) that is robust to end-to-end (E2E) channel errors and demonstrates high robustness to E2E channel errors.

Abstract

Recently, deep learning (DL) based image transmission at the physical layer (PL) has become a rising trend due to its ability to significantly outperform conventional separation-based digital transmissions. However, implementing solutions at the PL requires a major shift in established standards, such as those in cellular communications. Application layer (AL) solutions present a more feasible and standards-compliant alternative. In this work, we propose a layered image transmission scheme at the AL that is robust to end-to-end (E2E) channel errors. The base layer transmits a coarse image, while the enhancement layer transmits the residual between the original and coarse images. By mapping the residual image into a latent representation that aligns with the structure of the E2E channel, our proposed solution demonstrates high robustness to E2E channel errors.

Figures (11)

• Figure 1: $\text{DS}^{\text{2}}\text{C}^{\text{2}}$ and the inherent RGAN architectures.
• Figure 2: Binarization, interleaving, channel error modelling, and deinterleaving operations used in $\text{DS}^{\text{2}}\text{C}^{\text{2}}$.
• Figure 3: The details of BResNet encoder and decoder in the $\text{DS}^{\text{2}}\text{C}^{\text{2}}$ architecture.
• Figure 4: Similar structures observed in the latent representations of $r_{\text{i}}$ (left) and $r_{\text{i}}'$ (right) for Cityscapes and Kodak datasets.
• Figure 5: Structures observed in the latent representations of original image $x_{\text{i}}$ (left) and residual image $r_{\text{i}}$ (right) for ADE20K dataset.