基于张量低秩和深度先验的高光谱与多光谱图像融合理论与方法

Hyperspectral and Multi-spectral Image (HS+MS) fusion is a hot research topic in the hyperspectal remote sensing community. As an inverse problem of multisource incomplete observed multichannel images reconstruction, the high resolution fused image is obtained by fusing low spatial resolution hyperspectral (LRHS) images and high spatial resolution multispectral (HRMS) images to compute high spatial resolution hyperspectral images (HRHS). The challenges lie in the ill-posed, the need to achieve an optimal balance between maximizing the spatial geometrical details injection and preserve the fine spectral information, as well as the problem of massive data computing. Thus, the compacted and structured image priors are the key issues of the expressive power of the representation based learning methods. The main issue of this project focuses on developing a new theory and methods for HS+MS fusion with tensor low rank and deep priors.. Main contributions will be made as follows. Firstly, by exploiting the spectral-spatial correlations among hyperspectral images, we will establish a matrix-variate hierarchical Bayesian fusion framework and transfer learning based fusion mechanism with deep restricted Boltzmann machine. Secondly, we will propose a package of HS+MS fusion models such as compound regularization models with structured low rank tensor and geometrical priors, low rank constrained lightweight tensorizing network for deep priors learning models. In this end, we will propose jointly optimizing images fusion models with tensor based low rank and deep priors. Finally, we will develop high-performance "Plug-and-Play" algorithms for HS+MS deep fusion by using the proximal splitting based convex optimization and stochastic non-convex optimization methods. The research will provide a new theory and methodology for multi-source remote sensing image fusion, and can be widely used in application fields such as ground observation, environmental monitoring and military reconnaissance. Furthermore, the research has significance in extending the theory of data analysis, information fusion and pattern recognition for high-dimensional and multi-channel Images.

高光谱与多光谱图像融合是高光谱遥感重要热点问题。作为多源不完全观测多通道图像重建高分辨图像的反问题，其难点是需解决病态性、多光谱图像几何细节注入和高光谱图像精细光谱保持之间的优化平衡及计算数据量大的问题。项目以提高先验模型“紧致性”和“结构化”的图像表征能力为核心，探索基于张量低秩与深度先验的高光谱与多光谱图像融合理论与方法。创新点为：充分挖掘高光谱图像空-谱联合相关性, 建立矩阵随机变量分层贝叶斯融合和受限深度玻尔兹曼基迁移学习融合机理，提出张量结构化低秩与几何先验复合正则化模型和轻量型低秩约束张量化网络的深度先验学习模型；最终提出张量结构化低秩与深度先验协同优化融合模型；设计邻近算子分裂凸优化和随机非凸优化的高性能“即插-即用”融合算法。项目为多源遥感图像融合提供新理论和方法，可广泛应用于对地观测、环境监测和军事侦察等，对推动高维多通道图像分析和融合识别应用具有重要理论意义和应用价值。

高光谱与多光谱图像融合是高光谱遥感重要热点问题。作为多源不完全观测多通道图像重建高分辨图像的反问题，其难点是需解决病态性、多光谱图像几何细节注入和高光谱图像精细光谱保持之间的优化平衡及计算数据量大的问题。项目以提高先验模型“紧致性”和“结构化”的图像表征能力为核心，探索和发展张量低秩与深度先验学习的高光谱与多光谱图像融合理论、方法与应用技术。.取得的主要创新性成果为：. 1)研究和建立高光谱与多光谱图像融合的矩阵变量分层贝叶斯与受限玻尔兹曼基建模机理，发展了变分正则化计算融合、张量结构化低秩与几何先验复合正则化空-谱融合模型、优化机理和表示学习方法，探索了低秩空-谱视频稳定化模型与高效算法等，为遥感图像质量完善奠定了模型和算法基础； . 2)提出和发展了高光谱与多光谱图像融合的轻量型张量化深度递归残差网络先验学习、模型启发和注意力机制的空-谱遥感图像融合新方法，并发展联合张量结构化低秩与深度先验的高光谱与多光谱图像融合优化模型，在邻近算子分裂框架提出深度展开网络的高性能融合算法等，提高了融合网络的效能和可解释性；. 3)探索和发展了高光谱图像稀疏与低秩子空间分析的高效聚类方法、结构化内容感知卷积与图神经网络的高光谱图像小样本地物精细分类的新方法，为多源遥感融合协同分类奠定技术基础；. 4)深入研究了先验学习和模型驱动的高光谱图像融合超分辨成像方法，建立模型迁移和可学习子空间投影计算成像模型，大幅度降低了双相机融合容易引起的光谱失真。研发了双相机融合计算高光谱成像系统。. 项目代表性工作发表在IEEE TIP, IEEE TGRS, IEEE TNNLS，IEEE TCSVT，IEEE TMM，IEEE JSTARS等国际权威刊物；形成了系列专利技术，研制了小型机载高光谱成像系统、双相机感算融合成像和智能解译系统等，超额完成各类任务指标；为高光谱遥感目标探测、融合分类、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。