一种高光谱图像的双压缩感知模型

A double compressed sensing model of hyperspectral imagery

高光谱图像因其海量数据性,给存储、传输及后续分析处理带来了挑战。压缩感知理论提供了一种全新的信号采集框架。针对高光谱数据的三维特性,提出一种双压缩感知的采样与重构模型。该模型在采样阶段兼顾高光谱数据的空间和谱间稀疏特性,构造了能同时实现空间和谱间压缩采样的感知矩阵;重构阶段不同于传统的压缩感知重构方法直接重构高光谱数据,而是将高光谱数据分离成端元和丰度分别进行重构,然后利用重构的端元和丰度信息合成高光谱数据。实验结果表明,所提双压缩感知在低采样率下重构精度较三维压缩采样提高了10dB以上,更为显著的是运算速度提升了3个数量级,同时该方法还便于获得端元和丰度信息。

It is challenging for hyperspectral images to store, transport and subsequently analyze and process in terms of its huge amounts of data. Compressed sensing theory provides a new signal collection framework. A double compressive sampling and reconstruction model is proposed based on the three-dimensional characteristics of hyperspectral data. During the sampling stage, in terms of the sparsity of hyperspectral data between spatial and spectral, a sensing matrix is constructed to carry out spatial and spectral compressive sampling simultaneously. At the reconstruction stage, the proposed algorithm is different from the traditional reconstruction methods of compressed sensing, with which hyperspectral data are reconstructed directly; instead, with the proposed method, hyperspectral data are separated into endmembers and abundances to reconstruct respectively, then hyperspectral data are generated by reconstructed endmembers and abundances. Experimental results show that the reconstruction accuracy of double compressed sensing is improved by more than 10 dB under low sampling rate sampling, compared with three-dimensional compressive sampling, furthermore the computing speed is ascended by 3 orders of magnitude. Meanwhile, as a byproduct, endmembers and abundances can be obtained conveniently.