Removing noise from hyperspectral images is an inevitable step to improve the quality of these types of images. Many methods have been proposed by researchers in this field. Most of these methods do not address simultaneous spatial-spectral similarities. When the noise More
Removing noise from hyperspectral images is an inevitable step to improve the quality of these types of images. Many methods have been proposed by researchers in this field. Most of these methods do not address simultaneous spatial-spectral similarities. When the noise removal method applies data globally without regard to spatial-spectral similarities, it usually has a negative effect on low-level pixels; when in the spectral data, a large number of pixels have little noise and a small number of pixels are destroyed by the high level of noise. In this paper, we first extract spatial-spectral similarities in images by defining cluster-based latent variables. In the following, a low-rank matrix factorization method based on these latent variables is proposed to eliminate the noise of hyperspectral images and to improve the resistance to noise (as compared to other methods). The performance of the proposed method is compared visually with six new methods on real noise-contaminated images. For quantitative comparison, the same experiments are done on clean images combined with six types of simulated noise. The simulation results show that by applying latent variables in the Bayesian inference framework, the performance of the noise removal method is improved and the proposed method performs better than the other methods.
Manuscript profile
