PaperlayerView exact revision on arXiv

Source-linked AI summary

Wireless Image Transmission Using Deep Source Channel Coding With Attention Modules

Jialong Xu, Bo Ai, Wei Chen, Ang Yang, Peng Sun, Miguel Rodrigues

arXiv:2012.00533v3cs.IT

TL;DR

DL-based JSCC can suffer when deployment SNR differs from the training SNR, while covering many SNRs with separate networks increases computation and storage. The paper proposes ADJSCC, which uses channel-wise soft attention to scale features according to SNR, and reports adaptability, robustness, and lower resource requirements than comparison approaches.

  • Problem

    DL-based JSCC commonly assumes training and deployment channel SNRs match, while broad SNR coverage with multiple networks raises training computation and storage demands.

  • Method

    ADJSCC uses attention-based feature scaling, combining channel SNR with learned features to adapt joint source-channel coding within a single network.

  • Results

    ADJSCC demonstrates adaptability across channel conditions, robustness to mismatch, and comparable performance using 10.06% of BDJSCC-10’s storage and 10.36% of its training time at the same PSNR.

  • Takeaways & Limitations

    ADJSCC is presented as a practical alternative to storing and switching among multiple SNR-specific JSCC networks.

Abstract

from arXiv · show

Recent research on joint source channel coding (JSCC) for wireless communications has achieved great success owing to the employment of deep learning (DL). However, the existing work on DL based JSCC usually trains the designed network to operate under a specific signal-to-noise ratio (SNR) regime, without taking into account that the SNR level during the deployment stage may differ from that during the training stage. A number of networks are required to cover the scenario with a broad range of SNRs, which is computational inefficiency (in the training stage) and requires large storage. To overcome these drawbacks our paper proposes a novel method called Attention DL based JSCC (ADJSCC) that can successfully operate with different SNR levels during transmission. This design is inspired by the resource assignment strategy in traditional JSCC, which dynamically adjusts the compression ratio in source coding and the channel coding rate according to the channel SNR. This is achieved by resorting to attention mechanisms because these are able to allocate computing resources to more critical tasks. Instead of applying the resource allocation strategy in traditional JSCC, the ADJSCC uses the channel-wise soft attention to scaling features according to SNR conditions. We compare the ADJSCC method with the state-of-the-art DL based JSCC method through extensive experiments to demonstrate its adaptability, robustness and versatility. Compared with the existing methods, the proposed method takes less storage and is more robust in the presence of channel mismatch.

I. INTRODUCTION

Traditional modular communication separates source and channel coding, but finite-blocklength wireless systems motivate JSCC; deep learning has enabled strong source, channel, and joint coding results, while SNR mismatch remains a central problem.

  • I. INTRODUCTION: Shannon separation is optimal under certain conditions but assumes arbitrarily long codewords, conflicting with latency constraints in wireless applications.The paper highlights autonomous driving, smart manufacturing, and telemedicine as latency-sensitive settings.
  • I. INTRODUCTION: Deep learning has improved image compression and channel coding, motivating its use for joint source-channel coding.Prior work reports improvements over JPEG, JPEG2000, BPG, and conventional channel-coding approaches.
  • I. INTRODUCTION: Traditional JSCC allocates more channel-coding bits and fewer source-coding bits at low SNR, reversing this allocation at high SNR.This SNR-dependent resource assignment motivates the proposed approach.
  • I. INTRODUCTION: Existing DL-based JSCC methods generally optimize networks for specific SNR conditions, and deployment mismatch can seriously degrade performance.Covering a broad SNR range by training multiple networks increases training computation and storage requirements.
  • I. INTRODUCTION: ADJSCC uses attention-based resource allocation to scale intermediate features according to channel SNR and reports greater channel-mismatch robustness than prior DL-based JSCC.The method also feeds back channel SNR rather than the channel output and targets a single adaptive network.

II. SYSTEM MODEL

The system is a point-to-point image transmission model with SNR feedback, in which neural encoder and decoder functions jointly map images to noisy channel symbols and reconstructed images while minimizing distortion.

  • II. SYSTEM MODEL: The transmitter and receiver both know the channel SNR, which is fed back to the joint source-channel encoder.The image is represented as x ∈ R^n, with n = H×W×C.
  • II. SYSTEM MODEL: The encoder maps an n-dimensional real image vector to k complex-valued channel input symbols under an average power constraint.The encoder is represented by f_θ: R^n × R → C^k, and the constraint is 1/k E(zz*) ≤ 1.
  • II. SYSTEM MODEL: The encoded symbols pass through a noisy channel, after which the decoder maps the received symbols and SNR to an n-dimensional image estimate.The decoder is represented by g_φ: C^k × R → R^n.
  • II. SYSTEM MODEL: The model includes AWGN and differentiable fading-channel formulations, but non-differentiable channels are outside the paper’s scope.For fading channels, receiver equalization yields an equivalent representation with a different noise distribution.
  • II. SYSTEM MODEL: The bandwidth ratio R = k/n relates channel input size to image size, while training seeks encoder and decoder parameters that minimize expected reconstruction distortion.The expectation is taken over the joint distribution of original and reconstructed images and the SNR distribution.

III. PROPOSED METHOD

ADJSCC addresses SNR mismatch with one adaptive JSCC network that uses attention to reweight learned features according to channel conditions, rather than maintaining separate SNR-specific networks.

  • III. PROPOSED METHOD: Existing range-adaptive DL-based JSCC methods train multiple SNR-specific networks, increasing training computation and testing-stage storage demands.The proposed method instead targets a single network that adapts across a wide SNR range.
  • III. PROPOSED METHOD: ADJSCC applies attention mechanisms to implement SNR-dependent resource allocation within joint source-channel coding.Its motivation is the traditional JSCC strategy of adjusting source compression and channel-coding rate with SNR.
  • III. PROPOSED METHOD: The encoder and decoder alternate feature-learning modules with attention-feature modules that use SNR and learned features to produce scaling parameters.The scaling parameters recalibrate the channel-wise features before subsequent processing.
  • III. PROPOSED METHOD: Soft attention is used instead of hard attention because hard attention is non-differentiable and obstructs backpropagation during training.The method generates different scaling parameters for different feature channels.
  • III. PROPOSED METHOD: The attention-feature module globally pools learned features, combines the resulting feature information with SNR, and multiplies the features by predicted scaling factors.This channel-wise recalibration combines image features with wireless channel information.

1) Context Extraction:

The context-extraction stage summarizes learned feature maps globally and combines those summaries with the channel SNR to form the context used by attention.

  • 1) Context Extraction:: Global average pooling extracts global information from feature maps whose convolutional receptive fields are local.For each channel, the pooled value is obtained by averaging its spatial elements.
  • 1) Context Extraction:: The pooled feature information is concatenated with the channel SNR μ to produce the attention context vector.The resulting context contains the SNR and one global feature descriptor for each channel.

2) Factor Prediction:

The factor prediction network maps context information to channel-wise scaling factors for attention-based feature processing. Its compact two-layer design uses ReLU and sigmoid activations, while “channel-wise” refers to feature channels.

  • 2) Factor Prediction:: The factor prediction network Pω predicts scaling factor S from context information I.The network is used to generate attention scaling factors.
  • 2) Factor Prediction:: Pω contains two fully connected layers, with ReLU in the first and sigmoid in the last.The sigmoid constrains the output range to (0,1).
  • 2) Factor Prediction:: In channel-wise soft attention, “channel” denotes a feature channel rather than the communications channel.
  • 2) Factor Prediction:: The AF module comprises context extraction, factor prediction, and feature recalibration.

3) Feature Recalibration:

Feature recalibration converts predicted scaling factors into channel-wise attention features by scaling the feature map. The operation combines image-derived features with SNR information to implement channel-wise attention.

  • 3) Feature Recalibration:: Feature recalibration produces feature map F_A from feature map F_G using the predicted scaling factors.
  • 3) Feature Recalibration:: The AF module applies channel-wise scaling to the feature map using each factor S_i.The resulting recalibrated features form the attention feature map.
  • 3) Feature Recalibration:: ADJSCC combines SNR with inherent image features to compute channel-wise attention.This design reflects the dependence of source- and channel-coding bit allocation on channel SNR.

IV. SIMULATION RESULTS

The simulations implement and compare BDJSCC and ADJSCC architectures using CIFAR-10, average MSE, and average PSNR. ADJSCC is trained across [0, 20] dB, whereas BDJSCC is trained at a specific SNR.

  • ADJSCC architecture: The AF algorithm extracts global-average-pooled features, concatenates them with SNR, predicts scaling factors, and constructs channel-wise attention features.
  • BDJSCC architecture: BDJSCC uses five encoder and five decoder modules built primarily from convolutional or transposed convolutional, GDN, and activation layers.The fifth module differs by omitting the PReLU or sigmoid activation described for the other modules.
  • ADJSCC architecture: ADJSCC inserts AF modules after each FL module except the last encoder and decoder FL modules.The AF modules receive preceding FL outputs and channel SNR information.
  • ADJSCC architecture: The encoder’s final convolution output channel size controls the attainable bandwidth ratio.
  • Evaluation metrics: The comparison uses average MSE and average PSNR, with PSNR computed per image and then averaged over the tested images.Average MSE is defined over N transmitted images.
  • Simulation setup: ADJSCC is trained uniformly over [0, 20] dB, while BDJSCC is trained at a specific SNR; CIFAR-10 and 10 transmissions per test image are used.The models use Adam with learning rate 10^-4, batch size 128, and 1280 training epochs.

A. ADJSCC Adaptability Experiments

ADJSCC adapts to a range of SNR conditions using attention-based feature scaling, outperforming specific-SNR BDJSCC models especially at low bandwidth ratios and mismatched test conditions.

  • Training and evaluation setup: ADJSCC is trained over a uniform SNR range from 0 dB to 20 dB, whereas each BDJSCC model is trained at a specific SNR.Both methods are evaluated at specific test SNR values.
  • Performance adaptability: At R = 1/12, ADJSCC outperforms every BDJSCC model trained at a specific SNR.Its advantage increases as SNRtest deviates from SNRtrain.
  • Performance adaptability: 6dB: ADJSCC’s maximum margin over BDJSCC trained at SNRtrain = 1 dB at increasing SNRtest.
  • Bandwidth-ratio dependence: At R = 1/6, ADJSCC shows similar results, but its high-SNR advantage over BDJSCC trained at 13 dB or 19 dB nearly disappears.
  • Attention mechanism: ADJSCC uses attention mechanisms to accommodate a range of SNR conditions, with AF modules generating scaling factors from SNR and learned features.The scaling factors filter feature-module outputs.
  • Attention mechanism: The 23rd feature’s heatmaps at 1 dB and 19 dB illustrate SNR-dependent feature emphasis in the first encoder attention module.The passage associates high-SNR operation with enhancing cap information and poorer channels with protecting more robust information.

B. ADJSCC Robustness Experiments

ADJSCC retains an advantage over BDJSCC under AWGN channel mismatch, although both ADJSCC models experience some performance loss relative to matched conditions.

  • Mismatch performance: Under AWGN channel mismatch at R = 1/6, ADJSCC still outperforms BDJSCC.
  • Mismatch performance: Both ADJSCC models incur some performance loss under channel mismatch compared with the no-mismatch case.
  • Mismatch performance: ADJSCC is especially more robust than BDJSCC in high SNRfb regimes, where its gain is higher when feedback SNR is large.

C. ADJSCC Versatility

ADJSCC generalizes across datasets and image resolutions while maintaining performance across SNR conditions. Its performance closely approaches an ensemble of SNR-specific BDJSCC models, with small differences at high SNR.

  • ADJSCC trained on ImageNet and tested on Kodak is compared with BDJSCC models trained at specific SNRs.The ImageNet-trained ADJSCC uses a uniform 0–20 dB SNR distribution and bandwidth ratio R=1/6.
  • ADJSCC approaches the performance of the BDJSCC ensemble when SNRtest ≤17dB, with negligible difference.
  • 0.3dB lower performance occurs for ADJSCC than the BDJSCC ensemble when SNRtest >17dB.This regime corresponds to images with PSNR >35dB, whose quality is described as virtually indistinguishable to human eyes.
  • ADJSCC and BDJSCC reconstructions vary across transmissions at the same SNRtest because channel noise is random.The visual comparison uses SSIM to evaluate image quality.
  • ImageNet-trained ADJSCC suffers only 1dB to 3dB loss when tested on CIFAR-10, whereas CIFAR-10-trained ADJSCC performs much worse on Kodak.The results indicate limited image-size mismatch loss when training uses many higher-resolution images, but much larger loss after training on lower-resolution images.
  • Compared with TJSCC at twice the bandwidth ratio, ADJSCC is evaluated on Kodak using R=1/6 while TJSCC uses Rt=1/3.TJSCC separates Kodak color images into three channels because it supports only grayscale images.

V. STORAGE OVERHEAD AND COMPUTATIONAL COMPLEXITY

ADJSCC adds little model overhead but avoids the substantial storage and computational burden of maintaining multiple SNR-specific BDJSCC networks. At matched PSNR, it requires only a small fraction of BDJSCC-10 resources.

  • ADJSCC uses 10,758,191 parameters and approximately 41.04 MB, versus 10,690,351 parameters and approximately 40.78 MB for BDJSCC.The ADJSCC model has 0.6% more parameters than BDJSCC.
  • BDJSCC strategies use one, two, five, or ten SNR-specific models, with BDJSCC-10 requiring ten stored networks selected according to SNR conditions.BDJSCC-1 has slightly lower storage than ADJSCC but considerably underperforms.
  • 114 ms versus 110 ms is the mean training time per batch for ADJSCC and BDJSCC, while inference takes 53 ms versus 49 ms, respectively.These single-model timings do not account for the multiple BDJSCC models needed for comparable performance.
  • ADJSCC requires much lower overall computational and storage complexity than BDJSCC when multiple BDJSCC models are needed for comparable performance.

VI. CONCLUSION

The paper proposes ADJSCC, an attention-based JSCC method that adapts to varying channel conditions. Experiments demonstrate adaptability, robustness, versatility, and substantially lower resource requirements than multi-model BDJSCC strategies.

  • ADJSCC uses attention mechanisms to adapt automatically to various channel conditions.
  • ADJSCC combines feature-learning and attention-feature modules, using channel-wise soft attention to generate scaling factors and recalibrate features.
  • At equal PSNR, ADJSCC needs 10.06% of BDJSCC-10’s storage and 10.36% of its training time.
  • Future work includes extending ADJSCC to high-definition images and real wireless channels.
Loading 2012.00533v3…