SEER: Knowledge-driven semantic image restoration with vision-language diffusion alignment

Abstract

Semantic communication is an emerging paradigm to enhance network efficiency and perceptual quality, particularly demonstrating strong potential in image generation tasks. However, existing deep learning (DL)-based single-modal reconstruction approaches often suffer from semantic distortion and image blurring under bandwidth-limited and highly noisy channel conditions, limiting their suitability in task-oriented perception scenarios. Although generative AI-based semantic communication can significantly reduce data transmission volume, its high sensitivity to channel noise and lack of dynamic adaptation mechanisms limit the stability of reconstruction. To address these challenges, this paper proposes a multi-modal semantic communication framework named SEER, designed for resource-constrained intelligent sensing terminals. Built upon a pretrained language model, SEER incorporates a channel-aware prompt control strategy, a dual-modal integrative semantic restoration mechanism (DISR), and a single-pass sequential cross-modal reconstruction pathway to achieve collaborative semantic representation and robust structural recovery between images and text. Experimental results demonstrate that SEER achieves approximately $2.08\%$ bandwidth compression, while outperforming existing methods under extreme channel conditions by $33.92\%$ in structural fidelity and $12.64\%$ in perceptual consistency, highlighting its strong engineering deployability.