Source-linked AI summary
Deep Joint Source-Channel Coding for Wireless Image Transmission with Adaptive Rate Control
Mingyu Yang, Hun-Seok Kim
TL;DR
Wireless image transmission needs rate control under limited channel resources, while many deep JSCC methods rely on fixed-rate models. This paper uses a policy network with differentiable Gumbel-Softmax decisions to adapt feature transmission from SNR and image content. The single-model scheme reduces bandwidth in favorable conditions and achieves similar performance to fixed-rate models at corresponding operating conditions.
Problem
Existing deep JSCC methods are generally trained at fixed rates, requiring multiple models for multi-rate transmission under limited or expensive communication resources.
Method
A policy network dynamically masks image features according to channel SNR and image content, with Gumbel-Softmax enabling differentiable end-to-end training.
Results
The scheme reduces channel bandwidth utilization at high SNR or for simpler images while maintaining similar performance to specialized fixed-rate models.
Takeaways & Limitations
One trained model can support multiple rates and automatically allocate less bandwidth when channel conditions are favorable or image content is less informative.
Takeaways & Limitations
The evaluation assumes an AWGN channel with SNR known at both transmitter and receiver.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Deep JSCC addresses limitations of separate source and channel coding by adapting transmission rates to channel conditions and image content with one network model.
- Separate source and channel coding requires infinitely long codewords for optimality and breaks down with non-ergodic source or channel distributions.
- Deep JSCC can outperform separated coding, but existing methods generally use fixed rates and therefore require multiple trained models for multi-rate operation.
- The proposed scheme adapts rate to channel SNR and image contents using a single network model.
- A policy network dynamically selects transmission features, while Gumbel-Softmax makes the rate decisions differentiable for end-to-end training.
- Experiments show lower bandwidth utilization at high SNR or for simpler images while maintaining image quality comparable to specialized fixed-rate models.
2. ADATIVE JSCC METHOD
The proposed adaptive JSCC uses one model to select transmission features according to image content and SNR, providing multiple rates while balancing reconstruction quality and bandwidth usage.
- Overview: Multiple rates arise by activating different subsets of selective features through a thermometer-coded mask.The mask controls which selective feature groups are active, while non-selective groups remain active.
- Transmission and Rate: Active features are power-normalized and transmitted as complex-valued symbols over an AWGN channel, with SNR supplied to the encoder, decoder, and policy network.The transmission rate, measured in CPP, varies with the number of active groups.
- Policy Network: The policy network combines pooled image features with SNR to choose the number of active feature groups.It produces category probabilities, samples a decision, and converts that decision into the adaptive transmission mask.
- Gumbel Softmax: Gumbel-Softmax makes the discrete policy differentiable by using a continuous relaxation for the backward pass while retaining discrete sampling in the forward pass.This enables end-to-end optimization of the adaptive-rate JSCC model.
- Loss Function: The training loss combines image reconstruction accuracy with a channel-usage penalty weighted by α.The channel-usage term encourages fewer active feature groups and lower CPP.
3. EXPERIMENTS
Experiments show that the proposed method adapts transmission rate to SNR, rate-quality preferences, and image class, while maintaining performance comparable to fixed-rate models.
- At 0 dB SNR, the method selects its maximum possible rate, then lowers average rate as SNR increases.This demonstrates reduced channel-resource use under better channel conditions.
- Increasing α makes the policy prioritize rate more strongly, causing average rate to decrease faster as SNR rises.
- With α = 5 × 10−4, the method outperforms BPG+Capacity in the low-SNR region; with α = 1.5×10−3, it outperforms that baseline for all SNRs.
- The single adaptive model achieves similar performance to fixed-rate baseline models at each comparison point.The fixed-rate baselines use separately trained models with predetermined activation masking.
- Across CIFAR-10 classes, the policy assigns higher CPP to information-rich classes and lower CPP to simpler classes.Automobile and Truck receive higher CPP, while Ship and Airplane receive lower CPP.
- 0.613 PSNR standard deviation is achieved by adaptive rate control, versus 1.017, 0.991, 0.985, 0.947, and 0.923 for fixed-rate baselines.The comparison uses fixed CPP values of 0.25, 0.313, 0.375, 0.438, and 0.5, respectively.
4. CONCLUSION
The proposed deep JSCC scheme uses one model to support multiple rates, automatically adapting transmission to channel SNR and image content. Experiments show reduced bandwidth use in favorable or simpler conditions with negligible degradation relative to multiple single-rate models.
- The scheme supports multiple rates with a single model by dynamically masking image features according to channel SNR and image content.
- The Gumbel-Softmax trick makes the policy network differentiable for end-to-end training.
- The learned policy distributes less bandwidth when SNR is high or image content contains less information.
- The adaptive scheme experiences negligible performance degradation compared with multiple single-rate models at each operating condition.