Abstract:

Aiming at the problem of low spatial resolution of hyperspectral images, this paper analyzes the resolution enhancement methods of hyperspectral images and proposes a super resolution method. The proposed method uses non-parametric Bayesian sparse representation to fuse high-resolution images with low-spatial-resolution hyperspectral images. Firstly, the probability distribution and a set of Bernoulli distribution of material reflectance spectrum in the image are obtained from the hyperspectral image. Secondly, the dictionary is obtained through Bayesian dictionary learning, and the dictionary is transformed according to the spectral quantification of high-resolution image. Thirdly, the transformed dictionary is used to calculate the sparse coding matrix of the high resolution image. Finally, the dictionary and sparse coding matrix are combined to reconstruct the high-resolution hyperspectral image. The experimental results show that the proposed method is superior to the traditional methods in both subjective visual reconstruction of detail information, root mean square error and peak signal to noise ratio. Compared with the similar sparse representation method, the matrix factorization method and the coupled spectral unmixing method, the proposed method has the enhancement of reconstruction effect and its effectiveness is validated.

Key words: super resolution, hyperspectral image, Bayesian sparse representation, dictionary learning, sparse coding

摘要：

针对获取的高光谱图像空间分辨率较低的问题，对高光谱图像的分辨率增强方法进行分析研究，提出一种超分辨率方法。该方法使用非参数贝叶斯稀疏表示方法，将高分辨率图像与低空间分辨率的高光谱图像融合。首先，从高光谱图像中推测出材料反射光谱的概率分布以及一组伯努利分布；其次，通过贝叶斯字典学习得到光谱字典，并根据高分辨率图像的频谱量化进行字典变换；然后，利用变换后的字典计算高分辨率图像的稀疏编码矩阵；最后，将学习的字典与稀疏编码矩阵联合重建高分辨率的高光谱图像。实验结果表明，无论是主观视觉上的细节信息重建，还是客观指标的均方根误差以及峰值信噪比等，该方法均优于传统方法，与相似的稀疏表示方法、矩阵分解方法以及耦合光谱解混合方法相比，重建效果也有所提升，验证了有效性。

关键词: 超分辨率, 高光谱图像, 贝叶斯稀疏表示, 字典学习, 稀疏编码