基于双字典的高光谱与多光谱图像融合

在遥感领域, 高分辨率高光谱图像(high-resolution hyperspectral images, HR-HSI)的获取极具挑战, 通过融合低分辨率高光谱图像(low-resolution hyperspectral images, LR-HSI)与高分辨率多光谱图像(high-resolution multispectral images, HR-MSI)获得 HR-HSI 是较为经济的方式。基于光谱字典的传统融合算法在保持光谱连续性上效果优异, 但空间信息的表现力仍有增强的潜力。为此, 文章提出一种基于双字典的图像融合算法。首先利用 LR-HSI 和 HR-MSI 分别训练出光谱字典和空间字典, 然后基于光谱字典和空间字典分别在光谱域和空间域得到光谱型高分辨率高光谱图像(spectral high-resolution hyperspectral images, SPC-HR-HSI)和空间型高分辨率高光谱图像(spatial high-resolution hyperspectral images, Spa-HR-HSI), 并利用双域图像在迭代更新中相互约束, 彼此促进, 直至收敛, 最终融合出 HR-HSI。由于高光谱图像本身存在较强的低秩特性, 该文利用局部低秩与非局部低秩约束, 进一步增强目标图像的融合质量。实验结果表明, 该文提出的算法融合结果优于其他对比算法。

In the field of remote sensing, the acquisition of high-resolution hyperspectral images (HR-HSI) is extremely challenging. It is more economical to obtain HR-HSI by fusing low-resolution hyperspectral images (LR-HSI) with high-resolution multispectral images (HR-MSI). The traditional fusion algorithm based on spectral dictionary is excellent in spectral continuity, but the expressiveness of spatial information is still too high. Therefore, an image fusion algorithm based on double dictionary is proposed. Specifically, the spectral dictionary and the spatial dictionary are first trained by LR-HSI and HR-MSI, and then the spectral high-resolution hyperspectral images (Spc-HR-HSI) and spatial high-resolution hyperspectral images (Spa-HR-HSI) are obtained in the spectral domain and the spatial domain based on the spectral dictionary and the spatial dictionary respectively, and dual-domain images are constrained and promoted with each other in iterative update until convergence, so as to fuse out HR-HSI. In view of the strong low-rank characteristics of the HSI, the local low-rank and non-local low-rank constraints are used to further enhance the fusion quality of the target image. The experimental results show that the fusion results of the proposed algorithm are better than those of other comparison algorithms.