Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-hop Deep Joint Source-Channel Coding with Deep Hash Distillation for Semantically Aligned Image Retrieval

Didrik Bergström, Deniz Gündüz, Onur Günlü

TL;DR

This paper addresses image transmission over multi-hop AWGN channels with DeepJSCC, where noise accumulation degrades reconstruction. It proposes a semantic alignment framework by freezing a Deep Hash Distillation module ${\mathcal{H}}^{\star}$ and training the DeepJSCC encoder–decoder to minimize ${\mathcal{L}}_{\text{MSE}}({\mathbf S}, \widehat{{\mathbf S}})$ and ${\mathcal{L}}_{\text{SdH}}({\mathbf h}, \widehat{{\mathbf h}})$, aligning hashes ${\mathbf h}=\mathcal{H}(\mathbf S)$ and $\widehat{{\mathbf h}}=\mathcal{H}(\widehat{{\mathbf S}})$. The evaluation shows improved perceptual quality via ${\text{LPIPS}}$ and robustness to quantization in both decode-and-forward and quantize-and-forward relaying, with a measured trade-off in ${\text{PSNR}}$ performance. The approach demonstrates how semantic clustering via deep hashing can mitigate noise accumulation and enable security-oriented, semantically faithful transmission in practical multi-hop networks. Overall, it advances semantic communications by integrating hash-based semantic alignment with end-to-end deep JSCC.

Abstract

We consider image transmission via deep joint source-channel coding (DeepJSCC) over multi-hop additive white Gaussian noise (AWGN) channels by training a DeepJSCC encoder-decoder pair with a pre-trained deep hash distillation (DHD) module to semantically cluster images, facilitating security-oriented applications through enhanced semantic consistency and improving the perceptual reconstruction quality. We train the DeepJSCC module to both reduce mean square error (MSE) and minimize cosine distance between DHD hashes of source and reconstructed images. Significantly improved perceptual quality as a result of semantic alignment is illustrated for different multi-hop settings, for which classical DeepJSCC may suffer from noise accumulation, measured by the learned perceptual image patch similarity (LPIPS) metric.

Multi-hop Deep Joint Source-Channel Coding with Deep Hash Distillation for Semantically Aligned Image Retrieval

TL;DR

This paper addresses image transmission over multi-hop AWGN channels with DeepJSCC, where noise accumulation degrades reconstruction. It proposes a semantic alignment framework by freezing a Deep Hash Distillation module and training the DeepJSCC encoder–decoder to minimize and , aligning hashes and . The evaluation shows improved perceptual quality via and robustness to quantization in both decode-and-forward and quantize-and-forward relaying, with a measured trade-off in performance. The approach demonstrates how semantic clustering via deep hashing can mitigate noise accumulation and enable security-oriented, semantically faithful transmission in practical multi-hop networks. Overall, it advances semantic communications by integrating hash-based semantic alignment with end-to-end deep JSCC.

Abstract

We consider image transmission via deep joint source-channel coding (DeepJSCC) over multi-hop additive white Gaussian noise (AWGN) channels by training a DeepJSCC encoder-decoder pair with a pre-trained deep hash distillation (DHD) module to semantically cluster images, facilitating security-oriented applications through enhanced semantic consistency and improving the perceptual reconstruction quality. We train the DeepJSCC module to both reduce mean square error (MSE) and minimize cosine distance between DHD hashes of source and reconstructed images. Significantly improved perceptual quality as a result of semantic alignment is illustrated for different multi-hop settings, for which classical DeepJSCC may suffer from noise accumulation, measured by the learned perceptual image patch similarity (LPIPS) metric.

Paper Structure

This paper contains 13 sections, 9 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Decode-and-Forward (DF) Protocol
  4. Quantize-and-Forward (QF) Protocol
  5. Proposed Scheme
  6. Deep Hash Distillation (DHD)
  7. Training of DeepJSCC with DHD
  8. Experimental Setup
  9. Experimental Results and Discussion
  10. DF Relaying
  11. QF Relaying
  12. Conclusion
  13. Acknowledgement

Figures (3)

  • Figure 1: System design for DeepJSCC-DHD with frozen ${\mathcal{H}}^{\star}$.
  • Figure 2: DF multi-hop relay performance measured in LPIPS and PSNR. The line styles {dashed, dotted dash-dotted} belong to the SNRs $\{-5,-10,-15\}$ dB, respectively.
  • Figure 3: QF multi-hop relay performance measured in LPIPS and PSNR. The line styles {dashed, dotted, dash-dotted} belong to the SNRs $\{-5,-10,-15\}$ dB, respectively.