基于Swin Transformer的图像语义通信系统  

Image semantic communication system based on swin transformer

作　　者：孙洋舟 严天峰[1,2,3] 孙文灏 汤春阳 王映植 Sun Yangzhou;Yan Tianfeng;Sun Wenhao;Tang Chunyang;Wang Yingzhi(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Gansu Province Radio Monitoring and Positioning Industry Technology Center,Lanzhou 730070,China;Silk Road Brahma Communication Technology,Lanzhou 730030,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070 [2]甘肃省无线电监测及定位行业技术中心,兰州730070 [3]丝路梵天(甘肃)通信技术有限公司,兰州730030

出　　处：《电子测量技术》2024年第24期85-92,共8页Electronic Measurement Technology

基　　金：甘肃省科技重大专项项目(22ZD6GA041);甘肃省拔尖人才项目(6660030102)资助。

摘　　要：语义通信是一种旨在传递语义信息的通信方式,其通过可以有效的减少冗余和传输数据量等特点。目前语义通信的研究仅处于起步阶段,更多的理论研究有助于推动语义通信系统的真正实施。实现语义通信的核心技术端到端信源信道联合编码在过去几年中取得了长足的进步,语义图像也得到了发展。为解决计算效率过低、语义特征提取不足等问题,本文设计了一款新的神经网络JSCC。具体而言,受Swin Transformer在视觉任务中的优异表现的启发,首次将Swin-Transformer模块与残差网络相结合,设计出基于Swin Transformer的图像语义通信系统。为了解决传统的CNN对图像特征提取效率欠佳等问题,引入注意力残差网络模块初步提取图像语义特征,然后通过Swin Transformer进一步对图像语义特征进行提取。通过实验的结果的验证,与已有方案相比,本文所提出的方案在PSNR取得了高于2 dB的性能提升,在MS-SSIM性能上取得了5%以上的性能提升。Semantic communication is a type of communication designed to convey semantic information,which is characterized by the fact that it can effectively reduce redundancy and the amount of transmitted data.Currently the research on semantic communication is only in its infancy,and more theoretical research can help to promote the real implementation of semantic communication systems.The core technology for realizing semantic communication,end-to-end joint source channel coding,has made great progress in the past few years,and semantic images have also been developed.In order to solve the problems of computational inefficiency and insufficient semantic feature extraction,a new neural network JSCC is designed in this paper.Specifically,inspired by the excellent performance of Swin Transformer in visual tasks,the Swin-Transformer module is combined with residual networks for the first time,and a Swin Transformer-based image semantic communication system.In order to solve the problems such as the poor efficiency of traditional CNN for image feature extraction,the attention residual network module is introduced to extract the image semantic features initially,and then the image semantic features are further extracted by Swin Transformer.Through the verification of the experimental results,compared with the existing schemes,the proposed scheme in this paper achieves higher than 2 dB performance improvement in PSNR and more than 5%performance improvement in MS-SSIM performance.

关 键 词：语义通信 Swin Transformer 信源信道联合编码 

分 类 号：TN914[电子电信—通信与信息系统]