通信学报 ›› 2023, Vol. 44 ›› Issue (3): 198-208.doi: 10.11959/j.issn.1000-436x.2023050

• 学术通信 • 上一篇    下一篇

面向6G的深度图像语义通信模型

江沸菠1, 彭于波1, 董莉2,3   

  1. 1 湖南师范大学信息科学与工程学院,湖南 长沙 410081
    2 湖南工商大学长沙人工智能社会实验室,湖南 长沙 410205
    3 湘江实验室,湖南 长沙 410205
  • 修回日期:2023-02-09 出版日期:2023-03-25 发布日期:2023-03-01
  • 作者简介:江沸菠(1982− ),男,湖南株洲人,博士,湖南师范大学副教授、硕士生导师,主要研究方向为深度学习与物联网等
    彭于波(1996− ),男,重庆人,湖南师范大学硕士生,主要研究方向为语义通信和联邦学习
    董莉(1982− ),女,湖南长沙人,博士,湖南工商大学讲师、硕士生导师,主要研究方向为深度学习与物联网等
  • 基金资助:
    国家自然科学基金资助项目(41904127);国家自然科学基金资助项目(41604117);湘江实验室开放基金资助项目(6109408DL001);湖南省教育厅科学研究优秀青年基金资助项目(7103408DL001);湖南省教育厅资助科研项目(21A0372)

Deep image semantic communication model for 6G

Feibo JIANG1, Yubo PENG1, Li DONG2,3   

  1. 1 School of Information Science and Engineering, Hunan Normal University, Changsha 410081, China
    2 Changsha Social Laboratory of Artificial Intelligence, Hunan University of Technology and Business, Changsha 410205, China
    3 Xiangjiang Laboratory, Changsha 410205, China
  • Revised:2023-02-09 Online:2023-03-25 Published:2023-03-01
  • Supported by:
    The National Natural Science Foundation of China(41904127);The National Natural Science Foundation of China(41604117);Open Project of Xiangjiang Laboratory(6109408DL001);Project of Outstanding Youth in Scientific Research of Hunan Provincial Department of Education(7103408DL001);Scientific Research Fund of Hunan Provincial Education Department(21A0372)

摘要:

目前的语义通信模型在处理图像数据方面仍有可改善的部分,包括有效的图像语义编解码、高效的语义模型训练和精准的图像语义评估。为此,提出了一种深度图像语义通信(DeepISC)模型。首先采用基于 vision transformer的自编码器(ViTA)网络实现高质量的图像语义编解码;接着采用自编码器实现信道编解码,保证语义在信道上的传输;然后利用判别器网络(DSN)和ViTA的双网络架构协同训练,提高重建图像的语义精度;最后针对不同的下游视觉任务提出不同的图像语义评估指标。仿真结果表明,相较于其他方案,DeepISC可以更有效地还原传输图像的语义特征,使重建图像在各个下游任务中都展现出与原图像相同或相近的语义结果。

关键词: 人工智能, 6G, 语义通信, 图像识别, 特征提取

Abstract:

Current semantic communication models still have some parts that can be improved in processing image data, including effective image semantic codec, efficient semantic model training, and accurate image semantic evaluation.Hence, a deep image semantic communication (DeepISC) model was proposed.The vision transformer-based autoencoder (ViTA) network was used to achieve high-quality image semantic encoding and decoding.Then, an autoencoder realized channel codec to ensure the transmission of semantics on the channel.Furthermore, the discriminator network (DSN) and ViTA’s dual network architecture were used to jointly train, thus improving the semantic accuracy of the reconstructed image.Finally, for different downstream vision tasks, different evaluation indicators of image semantics were presented.Simulation results show that compared with other schemes, DeepISC can more effectively restore the semantic features of the transmitted image, so that the reconstructed image can show the same or similar semantic results as the original image in various downstream tasks.

Key words: artificial intelligence, 6G, semantic communication, image recognition, feature extraction

中图分类号: 

No Suggested Reading articles found!