Noise-Resilient Semantic Communication via Frequency-Decoupled Quantization

Semantic communication has emerged as a promising paradigm in next-generation communication systems, leveraging advanced artificial intelligence (AI) models to extract and transmit semantic representations for efficient information exchange. However, the reliability of received information is often compromised by unpredictable semantic noise, such as corruptions or distortions in the transmitted representations. Traditional methods typically rely on adversarial training with artificially injected noise to improve robustness. Yet, these approaches suffer from limited adaptability to varying noise conditions and incur considerable computational overhead during training. To address these challenges, this paper introduces Semantic communication with High-and-Low Frequency Decomposition (Se-HiLo), a novel noise-resilient scheme designed for image transmission. Se-HiLo integrates a Finite Scalar Quantization (FSQ) module that enhances robustness by constraining encoded representations within predefined discrete spaces, thereby eliminating the need for adversarial training. While FSQ strengthens resistance to noise, it inherently limits representational expressiveness. To mitigate this trade-off, Se-HiLo further incorporates a transformer-based high-and-low frequency decomposition module that separates image representations into distinct frequency components and encodes them into independent FSQ spaces, thus preserving semantic diversity and expressiveness. Extensive experiments validate that Se-HiLo significantly improves noise robustness and maintains accurate semantic communication across a wide range of noise environments.