面向安全性的视频分布式压缩感知关键技术研究

With limited computational power and resources, effective representation and security processing techniques are crucial for digital information acquisition from videos. As an emerging technique for digital signal processing, distributed compressive sensing (DCS) has been attracting increasingly wider attention for its small amount of data, joint sparsity and high probability reconstruction capacity. This project aims to apply DCS to video processing, mainly targeting the challenges in terms of spatiotemporal feature extraction, data volume and security measurement, side information and joint sparse reconstruction. The major contents include: i) Based on the feature distribution and sparse representation in spatial, temporal and transform domains, investigate DCS domain super-wavelet transform for over-complete dictionary design and optimization as well as new data generation and representation models; ii) Based on statistical characteristics of the correlation and differentiation of signals in sub-space models, investigate the relationship between the DCS data structure and measurement rate and design collaborative reconstruction method for the measurement matrix under secure keys; iii) Under noise and attack conditions, based on side information and support set in prior, investigate joint sparse reconstruction methods for highly accurate data matching and video restoration models. The proposed project is featured in terms of data security, spatiotemporal association and high restoration accuracy, and relevant outcomes will significantly and profoundly facilitate cloud computing, Internet+ and big data applications.

视频的有效表示及其安全性处理是数字信息获取的关键。信号处理引入的分布式压缩感知（Distributed Compressive Sensing, DCS）理论，以其数据量小、联合稀疏性及其高概率重构的特色，受到了信息领域的广泛关注。项目将DCS用于视频处理的研究中，主要解决时空关联与特征提取、数据量与安全性测量以及联合稀疏问题，内容包括：（1）基于视频信号的空域、时域、变换域条件分布及其稀疏表示，研究DCS超小波变换，设计优化的过完备字典,建立新的数据生成和表示模型。（2）在子空间模型下,根据信号相关性和差异性特征，研究感知数据结构与测量率的关系，设计密钥条件下测量矩阵的协同构造方法。（3）在噪声和攻击下，基于边信息及支撑集的先验知识，研究联合稀疏重构算法，实现高精度数据匹配与视频恢复模型。项目具有数据安全性和高质量恢复性，其研究成果将对云计算、互联网+及大数据的应用具有重要的意义。

针对视频信号的安全性预处理研究，项目基于DCS研究了视频信号的关键帧/非关键帧的统计特性及其特征提取，实现了基于过完备字典及其冗余度优化的联合稀疏表示方法；针对视频信号的安全性测量问题，研究了视频信号不同帧的相关性、稀疏度与自适应测量率的关系，并通过密钥设计确定性与结构化测量矩阵，建立了多维信号的子空间测量模型；结合噪声和攻击的应用环境，研究了关键帧的边信息、数据支撑集以及稀疏先验知识，建立了安全的视频DCS信号重构与恢复实现模型。项目结合深度学习理论进行了研究，丰富了多模态感知与数据融合的云安全关键技术，研究成果在IEEE Trans. Geoscience and Remote Sensing，Pattern Recognition等发表高水平论文17篇，授权发明专利7项，对智能科学领域发展具有一定的意义。