Universal Image Semantic Communication for Edge Network
Chenxin Yuan, Haojiang Ye, Yiming Miao · 2024
The increasing volume of data demands efficient communication systems for transmitting visual information across diverse application scenarios. Traditional methods based on Shannon's information theory overlook semantic content, leading to inefficient transmission of redundant data. Recent proposed deep learning-based semantic communication systems are often computationally intensive and domain-specific, limiting their practicality on edge devices and across various image types. We propose a universal image semantic communication (UISC) system - a lightweight, deep learning-based semantic communication framework for general-purpose image transmission. By processing low-frequency and high-frequency image components separately, our system reduces network complexity and enhances robustness to domain gaps. Unlike previous methods, UISC is versatile enough to handle RGB natural images, cartoons, text-embedded images, and grayscale medical images. Experimental results demonstrate that our method outperforms existing methods in both transmission accuracy and latency while consuming significantly less computational resources.