SNR-aware semantic image transmission with deep learning-based channel estimation in fading channels

초록

Semantic communications (SCs) are envisioned as a key enabler for sixth-generation wireless systems, leveraging deep learning to achieve efficient end-to-end transmission. In this work, we propose Swin semantic image transmission (SwinSIT), a novel joint source-channel coding paradigm for image SCs, built upon the Swin transformer. SwinSIT employs the Swin transformer to construct both the semantic encoder and decoder for effective semantic extraction and reconstruction. To enhance robustness, we design a signal-to-noise-ratio (SNR)-aware module inspired by squeezing-and-excitation networks, which leverages SNR feedback for double-phase enhancement of the semantic map at both encoder and decoder. Furthermore, a convolutional neural network (CNN)-based channel estimator and compensator mitigates fading channel effects by repurposing an image-denoising CNN. To enable lightweight deployment on IoT devices, a joint pruning-quantization scheme compresses the SwinSIT model. Simulation results demonstrate that SwinSIT outperforms conventional benchmarks. Also, the compressed model preserves strong performance at significantly reduced complexity.

키워드