基于语义重要性的语义编码算法研究
Research on Semantic Coding Algorithm Based on Semantic Importance
林上豪 1刘芳芳 2郭彩丽 2仝顽杰1
作者信息
- 1. 北京邮电大学先进信息网络北京实验室,北京 100876
- 2. 北京邮电大学先进信息网络北京实验室,北京 100876;北京邮电大学网络体系构建与融合北京市重点实验室,北京 100876
- 折叠
摘要
为了满足智能通信实体之间信息高效传输的需求,语义通信研究面临语义信息的表示与度量、编码以及传输等基本问题.针对通信资源紧张、任务时延敏感的万物智联传输场景,提出了基于语义重要性的语义编码算法,并设计、搭建了语义通信实验平台,实现了语义信息传输.首先,结合无线通信的基本流程建立了面向智能任务的语义通信系统模型;其次,设计了语义重要性提取方法,通过聚合图像通道特征并自适应校准其特征响应以获得特征图对于任务结果的重要性权重;然后,设计了语义压缩方法,在实际传输时根据重要性权重进行特征图筛选,以实现语义数据压缩;最后,在搭建的语义通信平台上以工业互联网场景中表面缺陷分类任务为例对所提算法的性能进行实测验证.实验结果表明,所提算法具有传输数据量小、抗噪声能力强和任务处理时延短等优势.
Abstract
In order to meet the demand of efficient information transmission between intelligent communication entities,semantic communication researchfaces the basic problems of semantic information representation,encoding and transmission.In view of the communication resource shortage and task delay sensitive transmission scenario of everything intelligent link,a semantic coding algorithm based on semantic importance was proposed,and a semantic communication experiment platform was designed and built to realize semantic information transmission.Firstly,the intelligent task-oriented semantic communication system model was established based on the basic process of wireless communication.Secondly,a semantic importance extraction method was designed to obtain the importance weights of the feature maps to the task results by aggregating the channel features and adaptively calibrating the feature responses.Then,a semantic compression method was designed to filter the feature maps according to the importance weight in the actual transmissionfor semantic data compression.Finally,the performance of the proposed algorithm was verified on the semantic communication platform by taking the task of surface defect classification in the industrial Internet scene as an example.Experimental result shows that the proposed algorithm has the advantages of small data transmission,strong anti-noise ability and short task processing delay.
关键词
语义通信/语义编码/图像分类/实验平台Key words
semantic communication/semantic coding/image classification/experimental platform引用本文复制引用
基金项目
北京邮电大学校行动计划项目(2021XD-A01-3)
出版年
2024
NETL
NSTL
万方数据