CNN and Attention-Based Joint Source Channel Coding for Semantic Communications in WSNs

Xinyue Liu, Zhen Huang, Yulu Zhang, Yunjian Jia, Wanli Wen2024 · DOI

TL;DR

This paper introduces an "Attention-based Adaptive Coding" (AAC) module for semantic communication in Wireless Sensor Networks (WSNs). This module dynamically adjusts semantic information based on channel conditions, enabling a single model to perform robustly across various Signal-to-Noise Ratio (SNR) environments, a significant improvement over existing methods that require retraining for different SNRs.

What Problem Does This Paper Address?

Wireless Sensor Networks (WSNs) are crucial for real-time applications, but their effectiveness is challenged by two main issues: robust communication in poor Signal-to-Noise Ratio (SNR) environments and alleviating bandwidth pressure from massive data transmission. Traditional communication methods, which rely on separate source and channel coding, are inefficient as they transmit raw data regardless of relevance and suffer from a "cliff effect" where performance degrades sharply in low SNR. While deep learning-based Joint Source Channel Coding (JSCC) offers a promising solution by transmitting semantic information, existing deep JSCC schemes often require retraining for different SNR levels or employ complex multi-decoder architectures, hindering their applicability in resource-constrained WSN devices. This paper aims to address the need for a single, adaptable semantic communication model that performs well across a range of SNRs.

Methods

The core innovation of this paper is the Attention-based Adaptive Coding (AAC) module. This module is designed to be flexible and computationally efficient, suitable for WSNs. It integrates two key attention mechanisms:

  • Window Attention Block (WAB): This mechanism focuses on prioritizing crucial regions within an image, mimicking biological vision by allocating more processing resources to important areas like edges and textures. This helps in extracting more relevant semantic information.
  • Channel Attention Mechanism: This component dynamically adjusts the semantic information based on the current channel state information (specifically, the SNR). It uses the Sum of Standard Deviation and Mean (SDM) to capture global information and then predicts weighting factors to fine-tune the output.

The AAC module is integrated into a novel end-to-end Joint Source Channel Coding (JSCC) scheme for image semantic communication. This scheme employs Convolutional Neural Networks (CNNs) for semantic encoding and decoding (S and S⁻¹) and the AAC module within the channel encoder and decoder (C and C⁻¹). The entire network is trained using the Adam optimizer to minimize the Mean Squared Error (MSE) between the original and reconstructed images.

Datasets

The paper utilizes the following datasets for training and evaluation:

  • CIFAR-10: Used for training and testing, with images of size 32x32.
  • Kodak: Used for evaluating performance on higher-resolution images (512x768).
  • ImageNet2012: Used as the training set for the Kodak evaluation, with images filtered to a patch size of 128x128.

Key Insights and Ideas

  • Adaptive Semantic Information is Key: The most significant insight is that dynamically adjusting the semantic information transmitted based on real-time channel conditions (SNR) is crucial for robust performance in WSNs. This moves beyond static encoding strategies.
  • Attention for Efficiency and Focus: The paper demonstrates that attention mechanisms, when carefully designed (like the window attention), can effectively prioritize important features without incurring prohibitive computational overhead, making them suitable for resource-constrained WSN devices.
  • Single Model Adaptability: The success of the AAC module in enabling a single model to adapt to a wide range of SNRs is a major breakthrough. This contrasts with previous approaches that required retraining or complex multi-model setups.
  • Improved Robustness in Low SNR: The proposed CA-Deep-JSCC scheme consistently outperforms baseline methods, especially in low SNR conditions, highlighting its enhanced robustness.

Key Results

  • Superior PSNR and SSIM: The proposed CA-Deep-JSCC scheme consistently achieved higher PSNR and SSIM values compared to baseline methods (Deep-JSCC and N-Deep-JSCC) across various SNR levels and bandwidth compression ratios on both CIFAR-10 and Kodak datasets.
  • Robustness Across SNRs: When trained on a range of SNRs (e.g., [0, 10] dB), the CA-Deep-JSCC scheme demonstrated outstanding performance across all tested SNR levels, verifying the effectiveness of the AAC module.
  • Visual Quality Improvement: Figure 6 shows that the CA-Deep-JSCC scheme produces visually superior reconstructed images, especially at lower SNRs, compared to the baseline models. For instance, at SNR = 1 dB and R = 1/12, CA-Deep-JSCC achieved a PSNR of 29.18 dB, while N-Deep-JSCC achieved 22.50 dB.

Limitations and Future Work

  • Limitations: The paper focuses on image semantic communication. While the AAC module is designed for efficiency, the computational cost of attention mechanisms, even window-based ones, might still be a consideration for extremely resource-constrained WSN nodes.
  • Future Work: The authors suggest extending the JSCC scheme's design to other modalities, including text, audio, and video. They also aim to explore semantic-level data transfer mechanisms and develop model lightening techniques for WSNs to implement a holistic semantic communication framework.

Summary updated

BibTeX
@misc{liu2024cnn,
  author = {Xinyue Liu and Zhen Huang and Yulu Zhang and Yunjian Jia and Wanli Wen},
  title = {{CNN and Attention-Based Joint Source Channel Coding for Semantic Communications in WSNs}},
  year = {2024},
  doi = {10.3390/s24030957},
}
Figure