压缩感知域高光谱数据高效压缩方法研究

Hyperspectral compressive imaging (HCI) technology based on the compressive sensing theory is becoming one of the hot research in hyperspectral remote sensing field during the recent years, because it breaks the limit of imaging resolution under the condition of existing technology,and meets the requirements of high-resolution applications in some cases. However the data obtained by HCI remains huge,and it can not meet the speed requirement of data transimission, so that we must compress the HCI data. Alougth the HCI data have much redundance, the existing compression methods for hyperspectral images are not suitable for the HCI data in CS domain, and therefore a novel CS data compression method is necessary. To deal with such problems, we will study on high-effcient HCI data compression based on spectral unmixing in this project, according to the property of HCI data. A linear unmixing model in CS domain and a fast optimization method for solving the model are stdied first, by which the HCI data are decomposed into the endmember spectral data and abundance image data. Then these endmember spectral data and abundance image data will be compressed losslessly through effective decorrelation and context modeling. With the analysis of the correlation between the estimate errors in CS domain and the classification accuracy, we perform an optimal quantization to the estimate errors. We believe that our method can achieve similar compression ratio with tranditional processing mode, and obtain better rate-distortion performance and higher classification accuracy.

以压缩感知(CS)为理论基础的高光谱压缩成像技术突破了现有技术对分辨率的限制，在某些应用场合下满足高分辨率应用要求，成为近年来高光谱遥感领域的研究热点之一。但是高光谱压缩成像的数据量仍然很大，不能满足数据传输速率的要求，必须进行压缩。尽管压缩成像数据存在较大冗余，但现有的高光谱图像压缩方法不适用于CS域的压缩成像数据，需要研究新的CS域数据压缩方法。基于此，本项目针对高光谱压缩成像数据的特性，研究基于混合光谱分解的高效压缩方法。通过研究CS域线性混合光谱分解模型及其快速优化求解方法，将高光谱成像数据分解为端元光谱数据和丰度图像数据；通过有效的去相关和上下文建模，对端元光谱数据和丰度图像数据进行无损压缩；通过分析CS域估计误差与重构图像失真和分类精度的相关性，对估计误差进行最优量化。该方法可以获得与传统方式相接近的压缩比，并具有更优的恢复图像质量，同时保持较高的分类精度。

本项目分析了高光谱压缩成像数据的特性，并设计了低复杂度的压缩感知成像方法；针对混合光谱分解问题，将压缩感知与线性解混相结合，提出了基于领域光谱加权的高光谱图像稀疏解混方法，并设计了快速优化求解方法，具有解混精度高、重构效果好、耗时短、效率高的优点；研究了高光谱图像数据的高效无损和有损压缩方法，重点解决了高效预测、分类谱间变换、低复杂度谱段配准、联合码率分配和码率控制、分布式编码等问题，有效改善了压缩性能；分析了高光谱图像有损压缩对光谱失真和地物分类应用的影响，为估计误差和分类精度的相关性模型的建立奠定基础，并为算法优化和性能改善提供依据。在本项目的资助下，出版英文书籍章节1章，在国内外期刊和国际会议上发表论文10篇，录用1篇，申请国家发明专利7项，其中授权4项。项目研究成果将在我国首次火星探测工程中得到应用。