数字语义通信中基于语义重要性的量化比特分配方法  

A Quantization Bit Allocation Method Based on Semantic Importance in Digital Semantic Communication

作　　者：朱翔本 郭彩丽[1] 杨洋 刘传宏 莫振扬 ZHU Xiangben;GUO Caili;YANG Yang;LIU Chuanhong;MO Zhenyang(School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876,China)

机构地区：[1]北京邮电大学信息与通信工程学院,北京100876

出　　处：《北京邮电大学学报》2024年第5期14-21,共8页Journal of Beijing University of Posts and Telecommunications

基　　金：国家自然科学基金项目(62371070);中央高校基本科研业务费专项资金项目(2021XD-A01-1)。

摘　　要：数字语义通信不仅保留了语义通信的优势,还能与现有通信系统兼容。而量化是实现数字语义通信的关键环节。在数字语义通信中,需要多个量化器对多维度的语义特征进行量化。由于量化器计算性能的限制,量化总比特数有限,所以迫切需要一种高效的比特分配方案。针对这一需求,提出了一种基于语义重要性的比特分配算法。首先,构建了基于语义重要性的量化比特分配问题模型,在最大比特数的限制下,综合考虑不同语义信息的重要性,使量化和传输过程中引入的失真最小化;接着,引入强化学习技术,将比特分配范围作为动作空间,将语义特征作为状态空间,提出了一种基于强化学习的量化比特分配算法;最后,通过仿真训练,获得了最优的比特分配策略。仿真结果表明,所提算法的收敛速度较快,在图像分类任务场景下,相较于基准算法,交叉熵下降达48.16%,分类准确率提高达12.65%。Digital semantic communication not only retains the advantages of semantic communication,but also is compatible with existing communication systems,with quantization being a key step in implementing digital semantic communication.Quantization in digital semantic communication requires multiple quantizers to quantify multi-dimensional semantic features.Due to limited hardware and constrained number of quantization bits,an efficient bit allocation scheme for quantizers is necessary.To solve this problem,a bit allocation algorithm based on semantic importance is proposed.Firstly,a quantized bit allocation problem based on semantic importance is constructed.Under the limitation of the maximum number of bits,the importance of different semantic information is considered to minimize the distortion caused by quantization and transmission.Then,a quantization bit allocation algorithm based on reinforcement learning is proposed with the bit allocation range as the action space and the semantic feature as the state space.Finally,the proposed algorithm is trained and the optimal bit allocation strategy is obtained.The simulation results show that the proposed algorithm converges quickly.In the image classification task scenario,the cross entropy of the proposed algorithm decreases by up to 48.16% and the classification accuracy increases by up to 12.65%,compared with the benchmark algorithm.

关 键 词：数字语义通信 语义重要性 量化比特分配 强化学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]