Robust Semantic Feature Importance-Aware Communications for Wireless Image Transmission

Semantic communications (SemCom) have emerged as a key enabler for efficient and robust image transmission in sixth generation (6G) networks. However, existing SemCom methods mainly adopt a separate design of semantic importance classifiers and encoders, and few studies have explored joint end-to-end (E2E) optimization frameworks that simultaneously consider semantic importance and optimize for channel-robust transmission. In this letter, we propose a robust semantic feature importance-aware communications (SFI-SemCom) framework for image transmission. First, to distinguish the importance, an implicit-explicit semantic importance classifier is designed using segmentation and feature extractor. A progressive E2E optimization strategy is developed to optimize the SFI-SemCom, thereby achieving adaptive protection based on semantic importance. Furthermore, to preserve spatial correlation and enhance texture details in reconstructed images, we design a texture consistency loss function based on the differentiable gray level co-occurrence matrix (D-GLCM). Simulation results show that SFI-SemCom surpasses the state-of-the-art, reducing LPIPS by 70.6% under additive white Gaussian noise (AWGN) channels and by 37.2% under Rayleigh fading channels, while improving PSNR by 6.4% and 3.9%, respectively.