基于字典学习及压缩感知的干涉高光谱数据压缩重建算法研究

Interference hyperspectral imaging is a very practical technology in the fields of remote sensing, meteorology, military, environmental monitoring and geology. In recent years the theorys of dictionary learning and compressed sensing put forward for the development of traditional hyperspectral imaging technology, but due to the special imaging principle, interference hyperspectral data has special characteristics, its fixed-position interference fringes and directional shift background lead to the invalid result of the application of traditional dictionary learning, compressed sensing and matching pursuit algorithm. According to the special interference hyperspectral imaging principle,the project first change the structure of interference hyperspectral data in order to get a data model suitable for sparse representation. Secondly, according to the correlation of spacial and spectrum region we want to find a better way of dictionary learning for interference hyperspectral data. Finally,according to the special characteristics of interference hyperspectral data, we propose the best inter-frames prediction model, inter-frames adaptive sampling rate of compressed sensing, adaptive sparsity matching pursuit algorithm, which will give the theoretical and technical base of the application for interference hyperspectral data.

干涉高光谱技术在航天遥感、气象、军事、环境监测和地质等领域均有实际应用价值，近年来字典学习和压缩感知理论的提出对传统高光谱成像技术的发展具有重要的推动作用，然而干涉高光谱成像仪的成像原理为光的干涉，所生成的干涉高光谱数据具有区别于普通高光谱数据的特殊性，其固定位置且高幅值的干涉条纹和方向性移位的背景导致直接应用传统的字典学习方法、压缩感知的固定采样率以及固定稀疏度的匹配追踪算法时无法得到理想的压缩重建效果。本项目针对干涉高光谱技术的特殊成像原理，首先改变干涉高光谱图像的数据结构，建立适合压缩感知稀疏表示的数据模型，其次，根据干涉高光谱数据特殊的空间与谱间相关性找到最佳的字典学习方法，最后，针对干涉高光谱图像数据区别于普通高光谱数据的特点，提出最佳的帧间预测模型、帧间压缩感知自适应采样率和自适应稀疏度匹配追踪算法，旨在为字典学习及压缩感知方法在干涉高光谱数据的应用奠定理论与技术基础。

干涉高光谱技术在航天遥感、气象、军事、环境监测和地质等领域均有实际应用价值，本项目针对干涉高光谱技术的特殊成像原理，首先改变干涉高光谱图像的数据结构，建立适合压缩感知稀疏表示的数据模型（背景层与干涉条纹层分开的图层分解模型）。其次，根据干涉高光谱数据特殊的空间与谱间相关性，结合字典学习方法训练出LASIS与LSMIS干涉高光谱数据的冗余字典，完成对干涉高光谱数据的最佳稀疏表示。最后，针对干涉高光谱图像数据区别于普通高光谱数据的特点，提出改进的正则正交匹配追踪算法，旨在为字典学习及压缩感知方法在干涉高光谱数据的应用奠定理论与技术基础。