基于视频信号空时稀疏的压缩感知重构方法

Compressed sensing (CS) is an emerging technology in signal processing and enables to efficiently capture and compress video signals with high resolution in the resource constrained environment. It has wide potential applications in wireless video communication, due to the great reduction of cost and computational complexity at encoder, and the robustness of recovery at decoder. In this project, we mainly focus on the CS reconstruction algorithms based on video spatial/temporal sparsity. The main content of the study is as follows. 1) Research and implementation of the spatial/temporal sparsity model (STSM) for video signals: Based on the relation between sparsity and spatial/temporal correlation existed in video, STSM is proposed to fully explore the video structural characteristics for CS reconstruction. 2) Research of the STSM-based reconstruction algorithms for compressed video sensing: Under the circumstance with measurement noises and channel errors, the sparse decomposition-, message passing- and analysis- based reconstruction algorithms are develped in oreder to reduce the required number of measurements and ensure the recovered video quality. 3) The joint optimization of sparse representation and reconstruction for further improvements in performance. With these key technologies, it is possible to substantially improve the compression efficiency and robustness of compressed video sensing systems, and finally lay the theoretical foundation for the practical CS video applications.

压缩感知作为一种新兴的信号处理技术，可以在资源受限的信道环境中实现大分辨率视频信号的采集压缩。其采样成本低，编码端简单和解码重构的鲁棒性都很适合于无线视频通信的应用场景。本项目重点研究基于视频信号空时稀疏的压缩感知重构方法，主要研究内容包括：第一，利用视频信号空时相关性和稀疏性的关系，建立视频空时稀疏模型，最大程度地挖掘视频信号的结构特征，用于视频压缩感知重构；第二，在含有测量噪声和信道误码的情况下，利用视频信号空时稀疏模型，研究能够降低重构所需测量值数，且保证恢复视频质量的重构算法，包括稀疏重构、消息传递重构和直接"分析"求解重构等；最后，进一步研究视频信号稀疏表示和重构的联合优化。将这些关键技术应用于视频压缩感知系统，可以大幅提高整个系统的压缩效率和鲁棒性，研究成果能够为视频压缩感知的实用化奠定理论基础。

由于整个视频压缩感知系统的性能在很大程度上将依赖于解码端的重构，而现有压缩感知重构方法未能全面考虑视频信号空时稀疏性的结构特征，往往不能取得理想的效果。因此，本项目重点研究了基于视频信号空时稀疏模型的压缩感知重构。主要内容包括：第一，利用视频信号空时相关性和稀疏性的关系，建立了视频空时稀疏模型，最大程度地挖掘了视频信号的结构特征，基于此对视频进行重构，重构质量显著提升；第二，在含有测量噪声和信道误码的情况下，利用视频信号空时稀疏模型，研究了能够降低重构所需测量值数，且保证恢复视频质量的重构算法；最后，研究了视频信号稀疏表示和重构的联合优化。基于这些关键技术，本项目大幅提高了整个视频压缩感知系统的压缩效率和鲁棒性，研究成果为视频压缩感知的实用化奠定了理论基础。在本项目资助下，已发表或录用学术论文共23篇(SCI：18篇，EI：4篇)，申请国家技术发明专利22件，其中已授权专利10件；有14名博士/硕士研究生参研本项目，其中6名学生已毕业，在研学生8名；参加国际会议或研讨会8次；已完成译著一本《认知网络测量与大数据》（ISBN:978-7-121-27551-7）。目前取得的研究成果已经达到项目申请时的预期成果要求，并将在今后三年内继续整理输出相关研究成果。