基于学习的无线通信编译码方法研究

The existing source coding depends on some simplified models and has poor adaptability to the fluctuation of network bandwidth. The existing channel coding depends on ideal channel model which is hard to be guaranteed in practical system. From the network information theory standpoint, the existing framework of joint source channel coding needs to be optimized. In this proposal, we propose to investigate the generative network based image/video coding schemes, and the learning based source coding schemes are designed to realize a theoretical model between cost and distortion in lossy compression. The deep network based deconvolution confidence propagation decoder and whitening the colored noise are studied. We investigate the mechanism analysis of learning-based coding via EXIT chart. This paper studies scheme based on learning-based source channel coding and the new receiver adaptive mechanism, establishes, and explores the migration learning problem from the channel model to the actual channel. Furthermore, the application demonstration system is built based on the above technologies, and a new method of deep crossing between machine learning and source and channel coding is developed to support the development of the new generation of multimedia and communication technologies. The learning based joint source channel coding, the adaptive video transmission and the auto-encoder based end-to-end video transmission are designed to explore the problem of transfer learning from the model to the actual channel. Combining them together, we build a learning based mobile internet video streaming platform, to support and promote the new generation of mobile internet.

现有信源编码依赖于各种简化模型，对网络带宽变化适应性差；现有信道编码依赖理想的信道模型，对实际信道变化适配性差；现有信源信道联合编码框架从网络信息论的观点尚待优化。针对以上问题，本项目研究基于生成网络的图像压缩编码和视频压缩编码，建立新的基于学习的信源编码框架，有损压缩的代价和失真之间的理论模型指导基于学习的信源编码；研究基于深度网络的解卷积置信传播译码器和采用深度学习网络进行非高斯噪声信道的白化，建立新的基于学习的信道编码框架，提出利用互信息转移图分析基于学习的编码并扩展到深度学习的机理分析；研究采用基于学习的信源信道编码并结合新的接收端自动适配机制自动适配视频传输方案，建立基于自编码器的端到端视频传输框架并探索从信道模型到实际信道的迁移学习问题。进一步采用上述技术搭建应用示范系统，发展机器学习与信源信道编码的新方法，支撑新一代多媒体和通信技术的发展。

随着移动智能终端的普及和移动互联网的发展，越来越多的用户通过移动智能终端使用互联网服务，促使移动数据量呈现出爆炸式增长。最新的思科 Visual Networking Index（VNI）报告显示，无线业务流量以每年接近 100% 的幅度增长。无线网络使用量的加速增长使得无线需求超出现有网络设施的容量，将造成网络的严重拥塞。从二代（2G）一直到五代（5G）的移动通信标准和 Wi-Fi 等无线局域网标准，每一代都以在带宽和能量受限条件下提高频谱效率和传输速率为目标。现有的传输系统针对理想的简化的信道模型优化，在实际传输过程中达不到最优效果。基于人工智能（Artificial Intelligence-AI）的无线通信技术为解决信源信道编码问题提供了很好的思路和基础。.本项目针对现有信源编码依赖于各种简化模型，对网络带宽变化适应性差的问题，研究基于生成网络的视频压缩编码进一步压缩时间冗余；针对现有信源信道编码框架有待优化的问题，研究采用基于学习的信源信道编码构成的自动适配图像视频传输方案，进一步利用深度学习作为工具，挖掘相近图像间的规律，改进传统编码框架，减小传输的代价和重建图像的视觉质量；针对现有网络架构难以满足对传输延迟和吞吐率的更高要求，提出了移动边缘计算网络中基于学习的推送和缓存联合管理策略来最大化网络带宽的利用率，利用强化学习来获取文件流行度，文件之间的相关性以及用户的使用习惯，从而实现更有效、更智能的缓存和网络资源管理。为支持下一代6G无线通信系统的发展，本项目将人工智能深度学习全面渗透至信源压缩编码、无线传输与编译码和边缘计算。本项目成果将进一步丰富和发展现有的机器学习和无线通信中信号处理等基础理论，为建立适用于6G无线通信传输系统提供理论支撑和技术方法。因此，本项目具有重要的理论意义和实际应用价值。