空谱自适应张量表示的高光谱图像目标检测方法研究

Hyperspectral image target detection is a hot topic in remote signal processing research. Although the sparse-representation based methods and low-rank matrix discriminant analysis based methods have achieved promising detection results, these methods convert the hyperspectral images to their matrix or vector formats which break down the internal structures of hyperspectral images. Besides, they are lack in spatial-spectral features’ joint description. In fact, hyperspectral images are third-order tensors. Thus tensor based analysis of hyperspectral images is necessary. This project focuses on spatial-spectral adaptive hyperspectral images tensor representation theory and target detection methods. The main contributions will be made as follows: we will exploit the spatial-spectral adaptive hyperspectral image tensor mode and propose sparse and low-rank priors metrics for hyperspectral tensors. Then, we design the tensor dictionary learning method according to the max-discriminatory-margin criterion, propose the tensor sparse representation based and multi-component tensor decomposition based methods for hyperspectral tartget detection. Further, we will combine the non-local, statistics and geometry information to build the non-local tensor discriminatory-information–driven hyperspectral target detection variation model. Finally, effective and efficient algorithms will be designed to solve the proposed problem based on the alternating direction method of multipliers and splitting operators methods. The project would not only establish the solid fundamentals for hyperspectral target detection, but also for the super-resolution, fusion based classification and feature extraction, which will have a wide application perspectives in the fields of earth observation and deep space exploration.

高光谱目标检测是遥感信息处理研究领域的热点问题。虽然现有的基于稀疏表示与低秩矩阵鉴别分析方法已取得了较好的检测结果，但是这些方法对高光谱图像进行了矩阵化或向量化处理，破坏了高光谱图像的内在结构，缺乏对空谱联合特征的刻画。高光谱图像实质是一个三阶张量，因此需要对高光谱图像进行张量建模分析。本项目研究空谱自适应张量表示的高光谱图像目标检测理论和算法，主要内容包括：利用高光谱图像空谱自适应张量表示模式，提出高光谱张量的稀疏与低秩先验度量；根据最大鉴别距离准则设计张量字典学习方法，提出基于张量稀疏表示和多成分张量分解的目标检测方法；融合图像中的非局部、统计和几何结构信息，建立非局部张量鉴别信息驱动的高光谱目标检测变分模型；利用交替迭代、算子分裂等设计模型求解算法。项目不仅对目标检测，同时也对高光谱图像超分辨率，融合分类，特征提取提供强有力的理论支撑，在对地观测、深空探测具有广阔的应用前景。

高光谱图像目标检测是遥感图像处理的热点问题。本项目构建高光谱图像的结构化表示方法，挖掘高光谱图像先验信息，设计正则项，研究了基于张量表示模型的高光谱图像融合与目标检测方法。（1）完成了基于非局部张量CP分解的高光谱与多光谱图像融合：提取高光谱图像的非局部相似块构成张量，利用张量CP分解刻画其空间、光谱、非局部结构，同时建立高光谱与多光谱图像的退化关系，引入多光谱图像的空间信息，实现高光谱图像的高精度重构。（2）完成了基于Tucker3张量分解的高光谱图像计算成像算法：利用四阶张量对高光谱图像中的相似块建模，设计张量Tucker3分解对四阶张量建模，刻画其低秩结构，耦合压缩感知采样信号与全色图像，建立了高光谱图像融合计算成像的变分模型。（3）完成了基于高光谱联合压缩感知与目标检测模型：将高光谱分解为背景与异常部分，引入马氏距离正则项刻画背景结构，建立基于张量鲁棒PCA的联合压缩感知与目标检测模型。（4）完成了基于低秩分解的高光谱视频气体检测方法：针对高光谱视频中气体的时、空、谱连续结构，构建基于3DTV约束的低秩分解模型，提取高光谱视频中的气体结构。本项目的研究对于推动高光谱图像的应用具有重要的理论与实际意义。