移动通信2025，Vol.49Issue(7)：61-67,7.DOI:10.3969/j.issn.1006-1010.20250524-0003
基于生成对抗网络的图像语义HARQ重传技术
Image Semantic HARQ Retransmission Based on Generative Adversarial Networks
摘要
Abstract
The vision of 6G leads to an explosive growth in image transmission demands due to the massive scale of intelligent connections.However,traditional pixel-level or symbol-level image transmission consumes significant communication resources,and there is substantial redundant information in the image data that is irrelevant to the task,resulting in low communication efficiency.To address this issue,semantic communication,as an emerging communication paradigm,enhances communication efficiency by extracting and transmitting key semantic features.However,the presence of noise and the time-varying nature of the communication link cause semantic errors,which impair the reliability of semantic communication and further impact task execution.Unlike traditional algebraic theory-based channel coding,semantic communication lacks a well-established theory for error detection and correction,as well as effective error control mechanisms.To achieve semantic-level error recovery in communication,we propose using Generative Adversarial Networks(GANs)to implement a semantic error corrector for image semantics,and a semantic error detector through supervised learning.When errors cannot be corrected,Hybrid Automatic Repeat reQuest(HARQ)technology is introduced to enhance the reliability of image semantic transmission.Specifically,a data-level fusion strategy is employed to achieve equal-gain combining of image semantic information from multiple retransmissions,thereby improving the quality of received image semantics.Finally,experimental simulation results show that the proposed image semantic HARQ scheme significantly improves the multi-scale structural similarity(MS-SSIM)and neural perceptual index(LPIPS)of the received images.Notably,under low signal-to-noise ratio(SNR)conditions,the proposed scheme demonstrates significant improvements in peak signal-to-noise ratio,MS-SSIM,and LPIPS compared to strategies without retransmission.Additionally,simulation results also verify that when the SNR exceeds 2 dB,the average number of retransmissions with the error correction scheme reduces by approximately one compared to the method without an error correction scheme.关键词
混合自动重传请求/生成对抗网络/语义通信/图像传输/差错控制Key words
hybrid automatic repeat request/generative adversarial network/semantic communication/image transmission/error control分类
信息技术与安全科学引用本文复制引用
符颢议,王旭,李永康,施政,杨光华..基于生成对抗网络的图像语义HARQ重传技术[J].移动通信,2025,49(7):61-67,7.基金项目
国家自然科学基金面上项目"面向超可靠低延时通信的叉包HARQ传输理论与方法"(62171200) （62171200）
国家自然科学基金国际(地区)合作与交流项目"AI驱动的6G无线智能空口传输理论与关键技术"(62261160650) （地区）
国家自然科学基金面上项目"智能反射面辅助的安全通信理论与关键技术研究"(62171201) （62171201）
广东省自然科学基金面上项目"叉包HARQ辅助太赫兹通信的传输理论和方法研究"(2023A1515010900) （2023A1515010900）
