Abstract
Recent research on joint source channel coding (JSCC) for wireless communications has achieved great success owing to the employment of deep learning (DL). However, 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 scenarios with a broad range of SNRs, which is computationally inefficient and requires large storage. To overcome these drawbacks, this 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. Instead of manual allocation, ADJSCC uses channel-wise soft attention to scale features according to SNR conditions. Compared with existing methods, the proposed method takes less storage and is more robust in the presence of channel mismatch.
Introduction
In traditional wireless communication systems, image transmission is typically handled by separate source coding (compression) and channel coding (error protection) stages. Joint Source-Channel Coding (JSCC) aims to merge these processes to achieve higher efficiency. While Deep Learning (DL) has significantly advanced JSCC, current models suffer from "channel mismatch." A model trained for a high Signal-to-Noise Ratio (SNR) performs poorly when the signal weakens, and vice versa.
To maintain performance across varying conditions, engineers currently must store multiple SNR-specific models on a device, which is impractical for storage-constrained IoT devices. The proposed ADJSCC (Attention Deep Joint Source-Channel Coding) addresses this by using a single neural network that adapts to real-time signal strengths, matching the performance of multiple specialized models while reducing storage requirements by 90%.
Methodology: The ADJSCC Framework
The core innovation of ADJSCC is the introduction of Attention Feature (AF) modules into the standard JSCC architecture. This design mimics the dynamic resource allocation of traditional systems but does so within an end-to-end learned framework.
Architecture and AF Modules
The network consists of alternating Feature Learning (FL) modules, which use standard convolutions, and the new AF modules. The AF module acts as the adaptive "brain" of the system. It takes two inputs: the image features from the previous layer and the current channel SNR.
Channel-wise Soft Attention
The AF module employs "Global Average Pooling" to capture the global context of the feature maps. It then integrates the SNR value to predict a scaling factor for every feature channel. This process, known as channel-wise soft attention, allows the network to dynamically recalibrate its internal representations.
As illustrated in the figure above, when the SNR is low (noisy), the attention mechanism prioritizes robust, high-level semantic information while suppressing "fragile" details that are easily corrupted. Conversely, at high SNR, the module allows more detailed textures and edges to pass through, maximizing image reconstruction quality.
Experimental Setup
The researchers evaluated ADJSCC using three primary benchmarks:
- CIFAR-10: Used for initial training and validation on low-resolution images.
- ImageNet: Used to develop a versatile model capable of handling complex, large-scale imagery.
- Kodak Dataset: A set of 24 high-resolution images used to test generalization to real-world photography.
Key Results and Discussion
The experimental results demonstrate that ADJSCC provides a robust solution to the channel mismatch problem.
Storage and Efficiency
To cover an SNR range of 0 to 20 dB, a standard DL-JSCC approach would require approximately 10 separate models, totaling 408 MB. ADJSCC achieves comparable or superior performance with a single 41 MB model—a 90% reduction in storage.
Performance and Robustness
On the Kodak dataset, ADJSCC significantly outperformed traditional JSCC methods like SPIHT + RCPC, showing improvements of up to 13 dB in Peak Signal-to-Noise Ratio (PSNR) at high signal strengths. Furthermore, ADJSCC proved remarkably stable; even when the estimated SNR was slightly inaccurate, the attention mechanism ensured "soft" performance degradation rather than the catastrophic failure typical of fixed models.
Feature Sensitivity
Analysis revealed that low-level features (edges and textures) are more sensitive to noise. Consequently, attention modules in the earlier layers of the network exhibit more significant fluctuations in response to SNR changes compared to deeper layers, which focus on more stable semantic features.
Conclusion
ADJSCC represents a significant advancement in wireless image transmission by successfully integrating attention mechanisms to solve the channel mismatch problem. By dynamically reconfiguring feature importance based on SNR, it provides a storage-efficient and robust alternative to multiple specialized models. Future work will explore adapting this mechanism for more complex, non-differentiable channel interference and scaling the architecture for ultra-high-definition video transmission.
Summary updated