[Paper Review] Deep Joint Source-Channel Coding for Wireless Image Transmission with Adaptive Rate Control

Mingyu Yang, Hun-Seok Kim|arXiv (Cornell University)|Oct 9, 2021

Advanced Data Compression Techniques7 citations

TL;DR

This paper proposes a single deep neural network (DNN)-based joint source-channel coding (JSCC) scheme that adaptively controls transmission rate based on channel SNR and image content using a differentiable policy network with the Gumbel-Softmax trick. The method achieves performance comparable to fixed-rate models across multiple SNR and image complexity conditions while dynamically reducing bandwidth usage in high-SNR or low-complexity scenarios.

ABSTRACT

We present a novel adaptive deep joint source-channel coding (JSCC) scheme for wireless image transmission. The proposed scheme supports multiple rates using a single deep neural network (DNN) model and learns to dynamically control the rate based on the channel condition and image contents. Specifically, a policy network is introduced to exploit the tradeoff space between the rate and signal quality. To train the policy network, the Gumbel-Softmax trick is adopted to make the policy network differentiable and hence the whole JSCC scheme can be trained end-to-end. To the best of our knowledge, this is the first deep JSCC scheme that can automatically adjust its rate using a single network model. Experiments show that our scheme successfully learns a reasonable policy that decreases channel bandwidth utilization for high SNR scenarios or simple image contents. For an arbitrary target rate, our rate-adaptive scheme using a single model achieves similar performance compared to an optimized model specifically trained for that fixed target rate. To reproduce our results, we make the source code publicly available at https://github.com/mingyuyng/Dynamic_JSCC.

Motivation & Objective

• To address the inefficiency of fixed-rate deep JSCC models in dynamic wireless environments with varying channel conditions and image content.
• To enable a single DNN model to support multiple transmission rates adaptively without retraining for each rate.
• To learn a dynamic rate control policy that reduces channel bandwidth usage when SNR is high or image content is simple.
• To maintain image quality comparable to specialized fixed-rate models while using a single unified architecture.

Proposed method

• A policy network generates a differentiable binary mask using the Gumbel-Softmax trick to select active feature groups for transmission.
• The channel encoder and decoder incorporate SNR-adaptive modules that modulate features based on the channel's signal-to-noise ratio (SNR).
• Transmission rate is quantified as channel usage per pixel (CPP), with values ranging from 0.25 to 0.5 based on the number of active feature groups.
• The model is trained end-to-end with a loss function that balances image quality (PSNR) and rate efficiency (via hyperparameter α).
• Feature groups are split into selective (G_s = 4) and non-selective (G_n = 4) types, where only selective groups are masked by the policy network.
• Complex-valued transmission symbols are formed by splitting real and imaginary parts of selected features, ensuring unit average power.

Experimental results

Research questions

• RQ1 Can a single deep JSCC model dynamically adjust its transmission rate based on both channel SNR and image content?
• RQ2 How does the proposed adaptive rate control policy compare in performance to fixed-rate models trained specifically for each rate?
• RQ3 Does the use of the Gumbel-Softmax trick enable effective end-to-end training of a differentiable rate selection mechanism?
• RQ4 To what extent does the policy network reduce bandwidth usage in high-SNR or low-complexity image scenarios?
• RQ5 How does adaptive rate control affect the uniformity of reconstructed image quality across different image classes?

Key findings

• The proposed method learns a policy that reduces channel bandwidth utilization (lower CPP) in high-SNR conditions, with average rate dropping from 0.5 at 0 dB to lower values as SNR increases.
• For a given SNR and image class, the policy assigns higher CPP to complex images (e.g., Automobile, Truck) and lower CPP to simpler ones (e.g., Ship, Airplane), improving quality uniformity.
• The standard deviation of PSNR across CIFAR-10 classes is 0.613 with adaptive rate control, significantly lower than fixed-rate baselines (ranging from 0.923 to 1.017).
• At α = 1.5×10⁻³, the method outperforms BPG+Shannon capacity baseline across all SNR levels, demonstrating robustness and efficiency.
• The adaptive model achieves PSNR performance equivalent to fixed-rate models trained for each specific rate, despite using a single unified architecture.
• The method maintains near-identical image quality to optimized fixed-rate models while reducing the need for multiple model deployments.