高光谱图像联合超分辨与分类的结构化稀疏学习理论与算法

Hyper-spectral image classification at finer scale is a hot topic in the field of quantitative remote sensing. Due to the low spatial resolution, various degradations such as blurring and noisy effect, mixed pixels, and lack of labeled samples, most previous image restoration and super-resolution approaches are aimed at reconstruction, with pattern classification only as an after-thought. Thus, it is difficult to achieve a power hyper-spectral image classification with high precision and good generalization performance because the non-jointly framework are lack of coupled optimization and feedback mechanism between the bottom to up and up to down. To overcome these problems, the main issue of this project focuses on the multi-task structured sparse learning theory and related algorithm of simultaneous super-resolution and classification for hyper-spectral image. The main contributions will be made as follows. Firstly, we will establish a novel group-graph sparsity measure by effectively exploiting both spatial and spectral correlation and contextual information in hyper-spectral image, and propose a spatial-spectral jointly structured sparsity based regularization method. Secondly, by incorporating a ‘high resolution reconstruction error ’(which aim at deblurring, denoising and resolution enhancement ),‘strong pattern discriminative’ criterion into the optimizing objective function, we will propose a new task-driven structured tensor dictionary learning method. Thirdly, we will present a structured sparse learning model of jointly super-resolution and classification for hyper-spectral image. Finally, we will develop several efficient and effective simultaneous super-resolution and classification algorithm with increment learning capacity for hyper-spectral image. The research of the proposed multi-task jointly optimizing model and algorithms will not only establish the solid fundamentals for fusion based classification, target detection, understanding, but also has theoretical significance and application perspectives for high dimensional signal processing, understanding and pattern recognition problems.

高光谱图像精细分类是定量遥感领域的热点问题。然而高光谱图像空间分辨率低、存在模糊、噪声和混合像元效应，图像恢复、超分辨等预处理和模式分类任务的分离处理缺乏底层与高层的双向优化与反馈机制，少量样本下分类精度和泛化性很难保证。项目研究高光谱图像联合超分辨与分类协同优化的结构化稀疏学习理论与算法。主要创新内容为：建立高光谱图像空谱联合相关性和上下文特性的概率图描述，提出组群和图结构化稀疏性新度量与空谱联合结构化稀疏正则化方法；以“高分辨重建性”（蕴含去模糊、去噪和分辨率提升）与“强模式鉴别性”为优化目标，提出多任务驱动的结构化张量字典学习方法；建立高光谱图像联合超分辨与分类协同优化的结构化稀疏学习模型；结合算子分裂、随机优化和GPU架构，设计具有增量学习能力的联合超分辨和亚像元分类的高性能算法。项目为高光谱遥感融合分类、目标探测、解译等提供有力支撑，对高维信息感知和模式识别具有理论和应用价值。

高光谱图像空间分辨率低、存在模糊、噪声和混合像元效应，图像恢复、超分辨等预处理和模式分类任务的分离处理缺乏底层与高层的双向优化与反馈机制，少量样本下分类精度和泛化性很难保证，项目聚焦定量遥感领域的高光谱图像超分辨与精细分类问题，研究结构化稀疏学习理论与应用算法。. 项目建立高光谱图像空谱联合相关性和上下文特性的概率图描述，研究组群和图结构化稀疏性新度量与空谱联合结构化稀疏正则化方法；以“高分辨重建性”（蕴含去模糊、去噪和分辨率提升）与“强模式鉴别性”为优化目标，研究多任务驱动的结构化字典学习方法,最终建立了一套高光谱图像联合超分辨与分类协同优化的结构化稀疏学习理论与算法。主要创新性工作为：. 1) 提出了多/高光谱图像全色融合超分辨的高阶及分数阶正则化模型与方法；. 2) 建立了基于先验建模的空-谱遥感图像融合与超分辨理论与算法；. 3) 提出了基于结构化稀疏表示学习的高光谱图像监督分类模型与算法；. 4) 提出非局部结构化低秩表示及其监督和无监督分类模型与算法；. 5) 提出空-谱内容自适应卷积网络的高光谱图像分类深度学习方法；. 6) 建立高光谱图像联合超分辨与分类协同优化的结构化稀疏学习模型；. 7) 结合算子分裂、随机优化和GPU架构，设计具有增量学习能力的联合超分辨和亚像元分类的高性能算法。..项目研发了高光谱图像处理与智能分析系统，形成了系列专利技术，为高光谱遥感融合分类、目标探测、解译等提供有力支撑，对高维信息感知和模式识别具有理论和应用价值。