Hyperspectral Image Denoising by Jointly Using Variational Bayes Matrix Factorization and Deep Image Prior

This paper delves into the dynamic field of hyperspectral imaging, focusing on the challenging task of denoising hyperspectral images given the presence of diverse noise types in the data. Within this context, we propose an innovative blind model-data-driven learning algorithm that combines a Bayesian low-rank matrix factorization-based scheme with a deep image prior convolutional neural network. Significantly, the algorithm estimates the spectrum-related image component via the Bayesian algorithm and the spatial information-related component by the network, resulting in improved denoising performance. Simulation results demonstrate the algorithm’s efficacy, surpassing contemporary unsupervised model-driven and model-data-driven HSI denoising algorithms. This work contributes to advancing the field by integrating model-driven principles with deep learning networks for enhanced hyperspectral image denoising.