Resisting Quantization Noise in Semantic Image Communication with Adversarial Learning-enabled HARQ

Semantic communication exploits the inherent meaning embedded in data content and has become a key enabler in knowledge-driven image transmission frameworks. However, existing research predominantly focuses on computational methodologies, often overlooking the transmission mechanisms. In particular, quantization noise introduced during semantic compression and transmission severely impacts the fidelity and utility of the received data, which remains an unresolved issue. To address this challenge, we propose an Adversarial Learning-based Quantization Selective Hybrid Automatic Repeat reQuest (ALQS-HARQ) mechanism tailored for semantic image transmission in remote sensing satellite networks. The proposed framework dynamically configures the number of quantization bits at the semantic encoder to enhance robustness against quantization noise. Furthermore, we design a quantized bitmerging retransmission scheme, equipped with an intelligent decision-making module and a standardized packet header. A novel metric is introduced to evaluate the semantic recovery completeness, guiding efficient retransmission. Extensive experiments demonstrate that the proposed ALQS-HARQ mechanism achieves higher task success rates with reduced transmission overhead compared to conventional retransmission schemes, showcasing its superior efficiency and adaptability.