TL;DR
This paper introduces a novel digital semantic communication system (MDJCM) that overcomes the limitations of analog transmission and fixed modulation schemes. By developing a multi-order digital joint coding-modulation approach with a unique substitution training strategy, the system achieves superior performance and flexibility for wireless image transmission.
What Problem Does This Paper Address?
Modern communication systems are increasingly supporting intelligent applications, driving the need for semantic communication. While joint source-channel coding (JSCC) has shown promise, existing methods often rely on analog transmission or fixed digital modulation, which are incompatible with established digital systems and lack flexibility. Analog transmission faces challenges with power amplifiers, while fixed digital modulation requires model retraining for different orders. This paper addresses the need for a digital, flexible, and robust semantic communication system that can adapt to various modulation orders and channel conditions without requiring separate models for each.
Methods
The core of the proposed solution is the Multi-Order Digital Joint Coding-Modulation (MDJCM) scheme. Key methodological contributions include:
- Digital Semantic Communication Framework: Integrating a multi-order modulator/demodulator into a nonlinear transform source-channel coding (NTSCC) framework.
- Substitution Training Strategy: To handle the non-differentiable nature of modulation/demodulation, the authors propose treating it as a constrained quantization process. They introduce scaling operations and handcrafted noise to create a differentiable approximation for training. This is further validated by comparing probability distributions.
- Two-Phase Training: To mitigate the mismatch between the differentiable training approximation and the deterministic testing phase, a two-phase strategy is employed. The first phase uses the substitution training, and the second phase fine-tunes with the straight-through estimator (STE).
- Hierarchical Dimension Reduction: To dynamically control the transmission rate, a gradual dimension reduction strategy is implemented at the transmitter. This breaks down the complex feature selection into simpler, iterative steps, improving robustness.
- Conditional Network Design: A modulation conditional module (MCM) and an SNR conditional module (SCM) are designed to allow a single model to handle different modulation orders and channel conditions, respectively.
Datasets
- Training: ImageNet, DIV2K, CIFAR10.
- Evaluation: CIFAR10, Kodak, CLIC2021, CLIC2022.
Key Insights and Ideas
- Digital is the Future: The paper strongly advocates for digital semantic communication, highlighting its compatibility and scalability advantages over analog approaches.
- Approximating Non-Differentiability: The substitution training strategy is a clever way to enable end-to-end training of complex, non-differentiable operations like modulation.
- Two-Phase Training for Robustness: Combining a differentiable approximation with STE fine-tuning effectively bridges the gap between training and real-world deployment.
- Modulation Agnosticism: The conditional network design is crucial for creating a single, flexible model that can adapt to various modulation orders, a significant practical advantage.
- Hierarchical Reduction for Gradual Adaptation: Moving away from a single, drastic dimension reduction to a hierarchical approach improves the system's ability to handle information loss and adapt to varying rates.
Key Results
- The proposed MDJCM significantly outperforms existing analog JSCC techniques (DeepJSCC, NTSCC, NTSCC+) and digital methods (JCM, STE, BPG+LDPC) in terms of PSNR across various datasets, SNRs, and channel conditions (AWGN and Rayleigh fading).
- The two-phase training strategy consistently improves performance over single-phase training with STE, especially at lower SNRs and modulation orders.
- The multi-order modulator/demodulator shows only a slight performance degradation compared to models trained with fixed modulation orders, demonstrating its effectiveness and generality.
- The system achieves better visual quality and lower CBR costs compared to benchmarks, as evidenced by reconstructed image examples.
Limitations and Future Work
- Approximation Error: While the substitution training strategy is effective, a slight approximation error exists, particularly at edge constellation symbols. This error is more pronounced at lower modulation orders but is largely mitigated by the two-phase training strategy.
- Dataset Bias: The initial training on ImageNet led to a bias in symbol length distribution, which was addressed by incorporating the DIV2K dataset.
- Future Work: The paper implicitly suggests further exploration into more advanced conditional network designs and potentially more sophisticated hierarchical reduction strategies. The demonstrated compatibility with diverse channel conditions also opens avenues for more complex channel models.
Summary updated