Spatial-spectral weighted nuclear norm minimization for hyperspectral image denoising

Abstract

Hyperspectral images (HSIs) are inevitably corrupted by various kinds of noise due to the instrumental and environmental factors. This degradation of HSI data affects the subsequent applications of these images. Despite the extensive research conducted into HSI denoising, satisfactory results under heavy noise levels have not yet been obtained. Assuming that the latent clean HSI holds the low-rank (LR) property while the noisy component does not, we propose a two-step Spatial-Spectral Weighted Nuclear Norm Minimization (SSWNNM) algorithm for HSI Denoising. Considering the LR property across the spectra, a Weighted Nuclear Norm Minimization (WNNM) algorithm is conducted to recover the spectral LR structure. In the spatial domain, nonlocal similar cubic patches are found and stacked into an LR matrix, which contains the local detailed spatial texture information. We further utilize Multi-channel Weighted Nuclear Norm Minimization (MCWNNM) to recover this spatial LR matrix. Experiments conducted on simulated and real HSI data demonstrate that the proposed denoising method outperforms state-of-the-art methods, both in terms of visual quality and several quantitative assessment indices.