基于压缩感知的特定信号盲源提取问题研究

The observed signals in real world can be regarded as linear or nonlinear mixtures of source signals. When the source signals are independent or at least incorrelated from each other, the methods based on blind source seperation (BSS) and blind source extraction (BSE) can work well in extracting the desired source signal from the observations. However, when the source signals are dependent or correlated,the BSS and BSE methods may work poorly. Under these circumstances, priori information on the desired source signal plays a key role. The priori information generally comes from how human beings recognize the world and acquire information from the world. The results from biologist research shows that: (1) unsupervised, or self-organized learning algorithm is more suitable for human visual system; (2) the visual characteristics of the signal had better be expressed as invariant and sparse. BSE algorithm can reflect the former feature, and the latter feature can be expressed as compressed sensing.In this way, the the desired signal extraction problem is transformed into desired blind source extraction based on compressed sensing. The research on this scientific issue not only helps understand how human visual system works, but also contributes to the solution of practical problems in intelligent video surveillance and hyperspectral image processing.

现实世界中观测到的信号往往可以看作、或近似看作各种形式源信号的线性或非线性混合。当源信号相互独立、或至少不相关的时候，基于盲源分离或盲源提取的算法能够提取出特定源信号。但源信号在更多时候是互相关、非独立的，这时候，一般的信号提取算法难以奏效，往往需要一些先验知识的帮忙。关于先验知识，其中大部分来自于对生物的理解，信息获取和感知，即研究生物视觉系统如何识别物体。生物学家的研究表明：(1)无监督的、或自组织的学习算法更适合人类视觉系统；(2) 视觉特征信号的编码最好是不变的和稀疏的。前者可以用盲源提取算法来体现，后者可以用压缩感知或稀疏的方式进行表达。这样，相关信号的提取问题就转化为基于压缩感知（先验特征）的特定信号的盲源（源的个数未知）提取的科学问题。这个科学问题的研究将不但对于理解人类视觉的信息获取与识别具有科学价值，还助于智能视频监控、高光谱图像处理等实际问题的解决。

本项目从人类视觉系统对图像中的选择性响应机理出发，利用现代信号处理技术研究如何从观测信号中提取特定信号。经过四年的研究，我们建立了特定信号盲源提取方法体系，提出了基于压缩感知约束的特定信号检测与识别的方法。发表/录用论文15篇，其中中科院JCR二区2篇，国际会议6篇，另外投稿8篇；已授权专利5个，已受理专利3个；培养博士研究生3名，毕业两名；培养硕士研究生10名，已毕业3名。参加学术讨论8人次。